RFC optional rte optional stdatomics API
Subsystems (2)
cover RFC optional rte optional stdatomics API
This series introduces API additions prefixed in the rte namespace that allow the optional use of stdatomics.h from C11 using enable_stdatomics=true for targets where enable_stdatomics=false no functional change is intended.
Be aware this does not contain all changes to use stdatomics across the DPDK tree it only introduces the minimum to allow the option to be used which is a pre-requisite for a clean CI (probably using clang) that can be run with enable_stdatomics=true enabled.
It is planned that subsequent series will be introduced per lib/driver as appropriate to further enable stdatomics use when enable_stdatomics=true.
Notes:
* additional libraries beyond EAL make visible atomics use across the API/ABI surface they will be converted in the next series.
* the eal: add rte atomic qualifier with casts patch needs some discussion as to whether or not the legacy rte_atomic APIs should be converted to work with enable_stdatomic=true right now some implementation dependent casts are used to prevent cascading / having to convert too much in the intial series.
* windows will obviously need complete conversion of libraries including atomics that are not crossing API/ABI boundaries. those conversions will introduced in separate series as new along side the existing msvc series.
Please keep in mind we would like to prioritize the review / acceptance of this patch since it needs to be completed in the 23.11 merge window.
Thank you all for the discussion that lead to the formation of this series.
Tyler Retzlaff (6):
eal: provide rte stdatomics optional atomics API
eal: adapt EAL to present rte optional atomics API
eal: add rte atomic qualifier with casts
distributor: adapt for EAL optional atomics API changes
bpf: adapt for EAL optional atomics API changes
devtools: forbid new direct use of GCC atomic builtins
app/test/test_mcslock.c | 6 +- config/meson.build | 1 + config/rte_config.h | 1 + devtools/checkpatches.sh | 8 ++ lib/bpf/bpf_pkt.c | 6 +- lib/distributor/distributor_private.h | 2 +- lib/distributor/rte_distributor_single.c | 44 ++++----- lib/eal/arm/include/rte_atomic_64.h | 32 +++--- lib/eal/arm/include/rte_pause_64.h | 26 ++--- lib/eal/arm/rte_power_intrinsics.c | 8 +- lib/eal/common/eal_common_trace.c | 16 +-- lib/eal/include/generic/rte_atomic.h | 66 ++++++++----- lib/eal/include/generic/rte_pause.h | 41 ++++---- lib/eal/include/generic/rte_rwlock.h | 47 ++++----- lib/eal/include/generic/rte_spinlock.h | 19 ++-- lib/eal/include/meson.build | 1 + lib/eal/include/rte_mcslock.h | 50 +++++----- lib/eal/include/rte_pflock.h | 24 ++--- lib/eal/include/rte_seqcount.h | 18 ++-- lib/eal/include/rte_stdatomic.h | 162 +++++++++++++++++++++++++++++++ lib/eal/include/rte_ticketlock.h | 42 ++++---- lib/eal/include/rte_trace_point.h | 4 +- lib/eal/ppc/include/rte_atomic.h | 50 +++++----- lib/eal/x86/include/rte_atomic.h | 4 +- lib/eal/x86/include/rte_spinlock.h | 2 +- lib/eal/x86/rte_power_intrinsics.c | 7 +- meson_options.txt | 1 + 27 files changed, 445 insertions(+), 243 deletions(-) create mode 100644 lib/eal/include/rte_stdatomic.h
Signature · 1 line
1.8.3.1
PATCH 1/6 eal: provide rte stdatomics optional atomics API
Provide API for atomic operations in the rte namespace that may optionally be configured to use C11 atomics with meson option enable_stdatomics=true
--- config/meson.build | 1 + config/rte_config.h | 1 + lib/eal/include/meson.build | 1 + lib/eal/include/rte_stdatomic.h | 162 ++++++++++++++++++++++++++++++++++++++++ meson_options.txt | 1 + 5 files changed, 166 insertions(+) create mode 100644 lib/eal/include/rte_stdatomic.h
config/meson.build +1 -0
@@ -303,6 +303,7 @@ endforeach # set other values pulled from the build options dpdk_conf.set('RTE_MAX_ETHPORTS', get_option('max_ethports')) dpdk_conf.set('RTE_LIBEAL_USE_HPET', get_option('use_hpet')) +dpdk_conf.set('RTE_ENABLE_STDATOMIC', get_option('enable_stdatomic')) dpdk_conf.set('RTE_ENABLE_TRACE_FP', get_option('enable_trace_fp')) # values which have defaults which may be overridden dpdk_conf.set('RTE_MAX_VFIO_GROUPS', 64)
config/rte_config.h +1 -0
@@ -13,6 +13,7 @@#define _RTE_CONFIG_H_#include <rte_build_config.h> +#include <rte_stdatomic.h> /* legacy defines */ #ifdef RTE_EXEC_ENV_LINUX
lib/eal/include/meson.build +1 -0
@@ -43,6 +43,7 @@ headers += files( 'rte_seqlock.h', 'rte_service.h', 'rte_service_component.h', + 'rte_stdatomic.h', 'rte_string_fns.h', 'rte_tailq.h', 'rte_thread.h',
lib/eal/include/rte_stdatomic.h new file +162 -0
162 lines not on this page · raw message
meson_options.txt +1 -0
@@ -46,6 +46,7 @@ option('mbuf_refcnt_atomic', type: 'boolean', value: true, description: 'Atomically access the mbuf refcnt.') option('platform', type: 'string', value: 'native', description: 'Platform to build, either "native", "generic" or a SoC. Please refer to the Linux build guide for more information.') +option('enable_stdatomic', type: 'boolean', value: false, description: 'enable use of C11 stdatomic') option('enable_trace_fp', type: 'boolean', value: false, description: 'enable fast path trace points.') option('tests', type: 'boolean', value: true, description:
Signature · 1 line
1.8.3.1
Re: [PATCH 1/6] eal: provide rte stdatomics optional atomics API
Quoted 14 lines · Tyler Retzlaff
Provide API for atomic operations in the rte namespace that may optionally be configured to use C11 atomics with meson option enable_stdatomics=true
Signed-off-by Tyler Retzlaff <[email protected]>--- config/meson.build | 1 + config/rte_config.h | 1 + lib/eal/include/meson.build | 1 + lib/eal/include/rte_stdatomic.h | 162 ++++++++++++++++++++++++++++++++++++++++ meson_options.txt | 1 + 5 files changed, 166 insertions(+) create mode 100644 lib/eal/include/rte_stdatomic.h
<snip>
meson_options.txt option('mbuf_refcnt_atomic', type: 'boolean', value: true, description: · line 49
Quoted 9 lines
meson_options.txt +1
-0@@ -46,6 +46,7 @@ option('mbuf_refcnt_atomic', type: 'boolean', value: true, description: 'Atomically access the mbuf refcnt.') option('platform', type: 'string', value: 'native', description: 'Platform to build, either "native", "generic" or a SoC. Please refer to the Linux build guide for more information.') +option('enable_stdatomic', type: 'boolean', value: false, description: 'enable use of C11 stdatomic')
Minor nit - all other options in this file put the description on it's own line. For consistency I think we should do the same here.
Quoted 6 lines
meson_options.txt +0
-0option('enable_trace_fp', type: 'boolean', value: false, description: 'enable fast path trace points.') option('tests', type: 'boolean', value: true, description:Signature · 1 line
1.8.3.1
RE: [PATCH 1/6] eal: provide rte stdatomics optional atomics API
In reply to Tyler RetzlaffQuoted 63 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Friday, 11 August 2023 03.32Provide API for atomic operations in the rte namespace that may optionally be configured to use C11 atomics with meson option enable_stdatomics=true
Signed-off-by Tyler Retzlaff <[email protected]>--- config/meson.build | 1 + config/rte_config.h | 1 + lib/eal/include/meson.build | 1 + lib/eal/include/rte_stdatomic.h | 162 ++++++++++++++++++++++++++++++++++++++++ meson_options.txt | 1 + 5 files changed, 166 insertions(+) create mode 100644 lib/eal/include/rte_stdatomic.hconfig/meson.build +1
-0@@ -303,6 +303,7 @@ endforeach # set other values pulled from the build options dpdk_conf.set('RTE_MAX_ETHPORTS', get_option('max_ethports')) dpdk_conf.set('RTE_LIBEAL_USE_HPET', get_option('use_hpet')) +dpdk_conf.set('RTE_ENABLE_STDATOMIC', get_option('enable_stdatomic')) dpdk_conf.set('RTE_ENABLE_TRACE_FP', get_option('enable_trace_fp')) # values which have defaults which may be overridden dpdk_conf.set('RTE_MAX_VFIO_GROUPS', 64)config/rte_config.h +1
-0@@ -13,6 +13,7 @@#define _RTE_CONFIG_H_#include <rte_build_config.h> +#include <rte_stdatomic.h> /* legacy defines */ #ifdef RTE_EXEC_ENV_LINUX lib/eal/include/meson.build +1
-0@@ -43,6 +43,7 @@ headers += files( 'rte_seqlock.h', 'rte_service.h', 'rte_service_component.h', + 'rte_stdatomic.h', 'rte_string_fns.h', 'rte_tailq.h', 'rte_thread.h',lib/eal/include/rte_stdatomic.h new file +2
-0@@ -0,0 +1,162 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2023 Microsoft Corporation
lib/eal/include/rte_stdatomic.h line 12
Quoted 10 lines
lib/eal/include/rte_stdatomic.h new file +10
-0+*/ + +#ifndef _RTE_STDATOMIC_H_ +#define _RTE_STDATOMIC_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +typedef int rte_memory_order;
In C11 memory_order is an enumerated type, and in GCC built-ins it is an int. If possible, rte_memory_order should be too; i.e. remove the typedef here, and make two variants of it instead.
lib/eal/include/rte_stdatomic.h line 21
Quoted 9 lines
lib/eal/include/rte_stdatomic.h new file +9
-0+ +#ifdef RTE_ENABLE_STDATOMIC +#ifdef __STDC_NO_ATOMICS__ +#error enable_stdatomics=true but atomics not supported by toolchain +#endif + +#include <stdatomic.h> + +#define __rte_atomic _Atomic
Move the (changed) C11 memory order type definition here:
/* The memory order is an enumerated type in C11. */ #define memory_order rte_memory_order
Quoted 26 lines
lib/eal/include/rte_stdatomic.h new file +26
-0+ +#define rte_memory_order_relaxed memory_order_relaxed +#ifdef __ATOMIC_RELAXED +_Static_assert (rte_memory_order_relaxed == __ATOMIC_RELAXED, +"rte_memory_order_relaxed == __ATOMIC_RELAXED" ); +#endif + +#define rte_memory_order_consume memory_order_consume +#ifdef __ATOMIC_CONSUME +_Static_assert (rte_memory_order_consume == __ATOMIC_CONSUME, +"rte_memory_order_consume == __ATOMIC_CONSUME" ); +#endif + +#define rte_memory_order_acquire memory_order_acquire +#ifdef __ATOMIC_ACQUIRE +_Static_assert (rte_memory_order_acquire == __ATOMIC_ACQUIRE, +"rte_memory_order_acquire == __ATOMIC_ACQUIRE" ); +#endif + +#define rte_memory_order_release memory_order_release +#ifdef __ATOMIC_RELEASE +_Static_assert (rte_memory_order_release == __ATOMIC_RELEASE, +"rte_memory_order_release == __ATOMIC_RELEASE" ); +#endif + +#define rte_memory_order_acq_rel memory_order_acq_rel
lib/eal/include/rte_stdatomic.h line 57
Quoted 10 lines
lib/eal/include/rte_stdatomic.h new file +10
-0+#ifdef __ATOMIC_ACQ_REL +_Static_assert (rte_memory_order_acq_rel == __ATOMIC_ACQ_REL, +"rte_memory_order_acq_rel == __ATOMIC_ACQ_REL" ); +#endif + +#define rte_memory_order_seq_cst memory_order_seq_cst +#ifdef __ATOMIC_SEQ_CST +_Static_assert (rte_memory_order_seq_cst == __ATOMIC_SEQ_CST, +"rte_memory_order_seq_cst == __ATOMIC_SEQ_CST" ); +#endif
Excellent idea adding these _Static_asserts!
Have you tested (with the toolchain you are targeting with this _Static_assert) that e.g. __ATOMIC_RELAXED is actually #defined, so the preprocessor can see it? (I guess that being a built-it, it might not be a #define, it might be a magic value known by the compiler only.)
Quoted 37 lines
lib/eal/include/rte_stdatomic.h new file +37
-0+ +#define rte_atomic_load_explicit (ptr, memorder) \ + atomic_load_explicit(ptr, memorder) + +#define rte_atomic_store_explicit (ptr, val, memorder) \ + atomic_store_explicit(ptr, val, memorder) + +#define rte_atomic_exchange_explicit (ptr, val, memorder) \ + atomic_exchange_explicit(ptr, val, memorder) + +#define rte_atomic_compare_exchange_strong_explicit ( \ + ptr, expected, desired, succ_memorder, fail_memorder) \ + atomic_compare_exchange_strong_explicit( \ + ptr, expected, desired, succ_memorder, fail_memorder) + +#define rte_atomic_compare_exchange_weak_explicit ( \ + ptr, expected, desired, succ_memorder, fail_memorder) \ + atomic_compare_exchange_strong_explicit( \ + ptr, expected, desired, succ_memorder, fail_memorder) + +#define rte_atomic_fetch_add_explicit (ptr, val, memorder) \ + atomic_fetch_add_explicit(ptr, val, memorder) + +#define rte_atomic_fetch_sub_explicit (ptr, val, memorder) \ + atomic_fetch_sub_explicit(ptr, val, memorder) + +#define rte_atomic_fetch_and_explicit (ptr, val, memorder) \ + atomic_fetch_and_explicit(ptr, val, memorder) + +#define rte_atomic_fetch_xor_explicit (ptr, val, memorder) \ + atomic_fetch_xor_explicit(ptr, val, memorder) + +#define rte_atomic_fetch_or_explicit (ptr, val, memorder) \ + atomic_fetch_or_explicit(ptr, val, memorder) + +#define rte_atomic_fetch_nand_explicit (ptr, val, memorder) \ + atomic_fetch_nand_explicit(ptr, val, memorder)
lib/eal/include/rte_stdatomic.h line 104
Quoted 10 lines
lib/eal/include/rte_stdatomic.h new file +10
-0+ +#define rte_atomic_flag_test_and_set_explict (ptr, memorder) \ + atomic_flag_test_and_set_explicit(ptr, memorder) + +#define rte_atomic_flag_clear_explicit (ptr, memorder) \ + atomic_flag_clear(ptr, memorder) + +#else + +#define __rte_atomic
Move the built-ins memory order type definition here:
/* The memory order is an integer type in GCC built-ins,
* not an enumerated type like in C11.
*/
typedef int rte_memory_order;
lib/eal/include/rte_stdatomic.h line 111
Quoted 7 lines
lib/eal/include/rte_stdatomic.h new file +7
-0+ +#define rte_memory_order_relaxed __ATOMIC_RELAXED +#define rte_memory_order_consume __ATOMIC_CONSUME +#define rte_memory_order_acquire __ATOMIC_ACQUIRE +#define rte_memory_order_release __ATOMIC_RELEASE +#define rte_memory_order_acq_rel __ATOMIC_ACQ_REL +#define rte_memory_order_seq_cst __ATOMIC_SEQ_CST
Agree; the memorder type is int, so no enum here.
Quoted 68 lines
lib/eal/include/rte_stdatomic.h new file +51
-0+ +#define rte_atomic_load_explicit (ptr, memorder) \ + __atomic_load_n(ptr, memorder) + +#define rte_atomic_store_explicit (ptr, val, memorder) \ + __atomic_store_n(ptr, val, memorder) + +#define rte_atomic_exchange_explicit (ptr, val, memorder) \ + __atomic_exchange_n(ptr, val, memorder) + +#define rte_atomic_compare_exchange_strong_explicit ( \ + ptr, expected, desired, succ_memorder, fail_memorder) \ + __atomic_compare_exchange_n( \ + ptr, expected, desired, 0, succ_memorder, fail_memorder) + +#define rte_atomic_compare_exchange_weak_explicit ( \ + ptr, expected, desired, succ_memorder, fail_memorder) \ + __atomic_compare_exchange_n( \ + ptr, expected, desired, 1, succ_memorder, fail_memorder) + +#define rte_atomic_fetch_add_explicit (ptr, val, memorder) \ + __atomic_fetch_add(ptr, val, memorder) + +#define rte_atomic_fetch_sub_explicit (ptr, val, memorder) \ + __atomic_fetch_sub(ptr, val, memorder) + +#define rte_atomic_fetch_and_explicit (ptr, val, memorder) \ + __atomic_fetch_and(ptr, val, memorder) + +#define rte_atomic_fetch_xor_explicit (ptr, val, memorder) \ + __atomic_fetch_xor(ptr, val, memorder) + +#define rte_atomic_fetch_or_explicit (ptr, val, memorder) \ + __atomic_fetch_or(ptr, val, memorder) + +#define rte_atomic_fetch_nand_explicit (ptr, val, memorder) \ + __atomic_fetch_nand(ptr, val, memorder) + +#define rte_atomic_flag_test_and_set_explicit (ptr, memorder) \ + __atomic_test_and_set(ptr, memorder) + +#define rte_atomic_flag_clear_explicit (ptr, memorder) \ + __atomic_clear(ptr, memorder) + +#endif + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_STDATOMIC_H_ */ meson_options.txt +0
-0@@ -46,6 +46,7 @@ option('mbuf_refcnt_atomic', type: 'boolean', value:true, description:
'Atomically access the mbuf refcnt.')
option('platform', type: 'string', value: 'native', description:
'Platform to build, either "native", "generic" or a SoC. Please
refer to the Linux build guide for more information.') +option('enable_stdatomic', type: 'boolean', value: false, description:
'enable use of C11 stdatomic')
option('enable_trace_fp', type: 'boolean', value: false, description:
'enable fast path trace points.')
option('tests', type: 'boolean', value: true, description:
--
1.8.3.1
Re: [PATCH 1/6] eal: provide rte stdatomics optional atomics API
Quoted 67 lines · Morten Brørup
Quoted 67 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Friday, 11 August 2023 03.32Provide API for atomic operations in the rte namespace that may optionally be configured to use C11 atomics with meson option enable_stdatomics=true
Signed-off-by Tyler Retzlaff <[email protected]>--- config/meson.build | 1 + config/rte_config.h | 1 + lib/eal/include/meson.build | 1 + lib/eal/include/rte_stdatomic.h | 162 ++++++++++++++++++++++++++++++++++++++++ meson_options.txt | 1 + 5 files changed, 166 insertions(+) create mode 100644 lib/eal/include/rte_stdatomic.hconfig/meson.build +1
-0@@ -303,6 +303,7 @@ endforeach # set other values pulled from the build options dpdk_conf.set('RTE_MAX_ETHPORTS', get_option('max_ethports')) dpdk_conf.set('RTE_LIBEAL_USE_HPET', get_option('use_hpet')) +dpdk_conf.set('RTE_ENABLE_STDATOMIC', get_option('enable_stdatomic')) dpdk_conf.set('RTE_ENABLE_TRACE_FP', get_option('enable_trace_fp')) # values which have defaults which may be overridden dpdk_conf.set('RTE_MAX_VFIO_GROUPS', 64)config/rte_config.h +1
-0@@ -13,6 +13,7 @@#define _RTE_CONFIG_H_#include <rte_build_config.h> +#include <rte_stdatomic.h> /* legacy defines */ #ifdef RTE_EXEC_ENV_LINUX lib/eal/include/meson.build +1
-0@@ -43,6 +43,7 @@ headers += files( 'rte_seqlock.h', 'rte_service.h', 'rte_service_component.h', + 'rte_stdatomic.h', 'rte_string_fns.h', 'rte_tailq.h', 'rte_thread.h',lib/eal/include/rte_stdatomic.h new file +6
-0@@ -0,0 +1,162 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2023 Microsoft Corporation +*/ + +#ifndef _RTE_STDATOMIC_H_ +#define _RTE_STDATOMIC_H_
Quoted 8 lines
lib/eal/include/rte_stdatomic.h line 12
Quoted 6 lines
lib/eal/include/rte_stdatomic.h new file +6
-0+ +#ifdef __cplusplus +extern "C" { +#endif + +typedef int rte_memory_order ;In C11
memory_orderis an enumerated type, and in GCC built-ins it is an int. If possible,rte_memory_ordershould be too; i.e. remove the typedef here, and make two variants of it instead.
will be in v2
Quoted 55 lines
lib/eal/include/rte_stdatomic.h line 21
Quoted 10 lines
lib/eal/include/rte_stdatomic.h new file +9
-0+ +#ifdef RTE_ENABLE_STDATOMIC +#ifdef __STDC_NO_ATOMICS__ +#error enable_stdatomics=true but atomics not supported by toolchain +#endif + +#include <stdatomic.h> + +#define __rte_atomic _AtomicMove the (changed) C11 memory order type definition here:
/* The memory order is an enumerated type in C11. */ #define
memory_orderrte_memory_orderQuoted 26 lines
lib/eal/include/rte_stdatomic.h new file +26
-0+ +#define rte_memory_order_relaxed memory_order_relaxed +#ifdef __ATOMIC_RELAXED +_Static_assert (rte_memory_order_relaxed == __ATOMIC_RELAXED, +"rte_memory_order_relaxed == __ATOMIC_RELAXED" ); +#endif + +#define rte_memory_order_consume memory_order_consume +#ifdef __ATOMIC_CONSUME +_Static_assert (rte_memory_order_consume == __ATOMIC_CONSUME, +"rte_memory_order_consume == __ATOMIC_CONSUME" ); +#endif + +#define rte_memory_order_acquire memory_order_acquire +#ifdef __ATOMIC_ACQUIRE +_Static_assert (rte_memory_order_acquire == __ATOMIC_ACQUIRE, +"rte_memory_order_acquire == __ATOMIC_ACQUIRE" ); +#endif + +#define rte_memory_order_release memory_order_release +#ifdef __ATOMIC_RELEASE +_Static_assert (rte_memory_order_release == __ATOMIC_RELEASE, +"rte_memory_order_release == __ATOMIC_RELEASE" ); +#endif + +#define rte_memory_order_acq_rel memory_order_acq_rellib/eal/include/rte_stdatomic.h line 57
Quoted 10 lines
lib/eal/include/rte_stdatomic.h new file +10
-0+#ifdef __ATOMIC_ACQ_REL +_Static_assert (rte_memory_order_acq_rel == __ATOMIC_ACQ_REL, +"rte_memory_order_acq_rel == __ATOMIC_ACQ_REL" ); +#endif + +#define rte_memory_order_seq_cst memory_order_seq_cst +#ifdef __ATOMIC_SEQ_CST +_Static_assert (rte_memory_order_seq_cst == __ATOMIC_SEQ_CST, +"rte_memory_order_seq_cst == __ATOMIC_SEQ_CST" ); +#endif Excellent idea adding these
_Static_asserts!
Quoted 1 line
Have you tested (with the toolchain you are targeting with this
_Static_assert) that e.g.__ATOMIC_RELAXEDis actually #defined, so the preprocessor can see it? (I guess that being a built-it, it might not be a #define, it might be a magic value known by the compiler only.)
* llvm and gcc both expose it as a built-in #define for test builds i have run. worst case the assert is lost if it isn't.
* since i have to handle non-{clang,gcc} too i still guard with ifdef
* i do need to switch to using assert.h static_assert macro to inter-operate with c++ in v2
Quoted 135 lines
Quoted 38 lines
lib/eal/include/rte_stdatomic.h new file +37
-0+ +#define rte_atomic_load_explicit (ptr, memorder) \ + atomic_load_explicit(ptr, memorder) + +#define rte_atomic_store_explicit (ptr, val, memorder) \ + atomic_store_explicit(ptr, val, memorder) + +#define rte_atomic_exchange_explicit (ptr, val, memorder) \ + atomic_exchange_explicit(ptr, val, memorder) + +#define rte_atomic_compare_exchange_strong_explicit ( \ + ptr, expected, desired, succ_memorder, fail_memorder) \ + atomic_compare_exchange_strong_explicit( \ + ptr, expected, desired, succ_memorder, fail_memorder) + +#define rte_atomic_compare_exchange_weak_explicit ( \ + ptr, expected, desired, succ_memorder, fail_memorder) \ + atomic_compare_exchange_strong_explicit( \ + ptr, expected, desired, succ_memorder, fail_memorder) + +#define rte_atomic_fetch_add_explicit (ptr, val, memorder) \ + atomic_fetch_add_explicit(ptr, val, memorder) + +#define rte_atomic_fetch_sub_explicit (ptr, val, memorder) \ + atomic_fetch_sub_explicit(ptr, val, memorder) + +#define rte_atomic_fetch_and_explicit (ptr, val, memorder) \ + atomic_fetch_and_explicit(ptr, val, memorder) + +#define rte_atomic_fetch_xor_explicit (ptr, val, memorder) \ + atomic_fetch_xor_explicit(ptr, val, memorder) + +#define rte_atomic_fetch_or_explicit (ptr, val, memorder) \ + atomic_fetch_or_explicit(ptr, val, memorder) + +#define rte_atomic_fetch_nand_explicit (ptr, val, memorder) \ + atomic_fetch_nand_explicit(ptr, val, memorder)lib/eal/include/rte_stdatomic.h line 104
Quoted 10 lines
lib/eal/include/rte_stdatomic.h new file +10
-0+ +#define rte_atomic_flag_test_and_set_explict (ptr, memorder) \ + atomic_flag_test_and_set_explicit(ptr, memorder) + +#define rte_atomic_flag_clear_explicit (ptr, memorder) \ + atomic_flag_clear(ptr, memorder) + +#else + +#define __rte_atomicMove the built-ins memory order type definition here:
/* The memory order is an integer type in GCC built-ins,
* not an enumerated type like in C11.
*/
typedef intrte_memory_order;lib/eal/include/rte_stdatomic.h line 111
Quoted 7 lines
lib/eal/include/rte_stdatomic.h new file +7
-0+ +#define rte_memory_order_relaxed __ATOMIC_RELAXED +#define rte_memory_order_consume __ATOMIC_CONSUME +#define rte_memory_order_acquire __ATOMIC_ACQUIRE +#define rte_memory_order_release __ATOMIC_RELEASE +#define rte_memory_order_acq_rel __ATOMIC_ACQ_REL +#define rte_memory_order_seq_cst __ATOMIC_SEQ_CSTAgree; the memorder type is int, so no enum here.
Quoted 68 lines
lib/eal/include/rte_stdatomic.h new file +51
-0+ +#define rte_atomic_load_explicit (ptr, memorder) \ + __atomic_load_n(ptr, memorder) + +#define rte_atomic_store_explicit (ptr, val, memorder) \ + __atomic_store_n(ptr, val, memorder) + +#define rte_atomic_exchange_explicit (ptr, val, memorder) \ + __atomic_exchange_n(ptr, val, memorder) + +#define rte_atomic_compare_exchange_strong_explicit ( \ + ptr, expected, desired, succ_memorder, fail_memorder) \ + __atomic_compare_exchange_n( \ + ptr, expected, desired, 0, succ_memorder, fail_memorder) + +#define rte_atomic_compare_exchange_weak_explicit ( \ + ptr, expected, desired, succ_memorder, fail_memorder) \ + __atomic_compare_exchange_n( \ + ptr, expected, desired, 1, succ_memorder, fail_memorder) + +#define rte_atomic_fetch_add_explicit (ptr, val, memorder) \ + __atomic_fetch_add(ptr, val, memorder) + +#define rte_atomic_fetch_sub_explicit (ptr, val, memorder) \ + __atomic_fetch_sub(ptr, val, memorder) + +#define rte_atomic_fetch_and_explicit (ptr, val, memorder) \ + __atomic_fetch_and(ptr, val, memorder) + +#define rte_atomic_fetch_xor_explicit (ptr, val, memorder) \ + __atomic_fetch_xor(ptr, val, memorder) + +#define rte_atomic_fetch_or_explicit (ptr, val, memorder) \ + __atomic_fetch_or(ptr, val, memorder) + +#define rte_atomic_fetch_nand_explicit (ptr, val, memorder) \ + __atomic_fetch_nand(ptr, val, memorder) + +#define rte_atomic_flag_test_and_set_explicit (ptr, memorder) \ + __atomic_test_and_set(ptr, memorder) + +#define rte_atomic_flag_clear_explicit (ptr, memorder) \ + __atomic_clear(ptr, memorder) + +#endif + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_STDATOMIC_H_ */ meson_options.txt +0
-0@@ -46,6 +46,7 @@ option('mbuf_refcnt_atomic', type: 'boolean', value:true, description:
'Atomically access the mbuf refcnt.')
option('platform', type: 'string', value: 'native', description:
'Platform to build, either "native", "generic" or a SoC. Please
refer to the Linux build guide for more information.') +option('enable_stdatomic', type: 'boolean', value: false, description:
'enable use of C11 stdatomic')
option('enable_trace_fp', type: 'boolean', value: false, description:
'enable fast path trace points.')
option('tests', type: 'boolean', value: true, description:
--
1.8.3.1
RE: [PATCH 1/6] eal: provide rte stdatomics optional atomics API
Quoted 13 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Friday, 11 August 2023 17.55Quoted 9 lines · Morten Brørup
Quoted 9 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Friday, 11 August 2023 03.32Provide API for atomic operations in the rte namespace that may optionally be configured to use C11 atomics with meson option enable_stdatomics=true
Signed-off-by Tyler Retzlaff <[email protected]>---
[...]
Quoted 5 lines
Quoted 5 lines
Move the (changed) C11 memory order type definition here:
/* The memory order is an enumerated type in C11. */ #define
memory_orderrte_memory_order
No objections to your typedef in v2.
[...]
Quoted 7 lines
Quoted 7 lines
lib/eal/include/rte_stdatomic.h line 57
Quoted 4 lines
lib/eal/include/rte_stdatomic.h new file +4
-0+#ifdef __ATOMIC_SEQ_CST +_Static_assert (rte_memory_order_seq_cst == __ATOMIC_SEQ_CST, +"rte_memory_order_seq_cst == __ATOMIC_SEQ_CST" ); +#endif Excellent idea adding these
_Static_asserts!
Quoted 7 lines
Quoted 1 line
Have you tested (with the toolchain you are targeting with this
_Static_assert) that e.g.__ATOMIC_RELAXEDis actually #defined, so the preprocessor can see it? (I guess that being a built-it, it might not be a #define, it might be a magic value known by the compiler only.)* llvm and gcc both expose it as a built-in #define for test builds i have run. worst case the assert is lost if it isn't.
I only wanted to check that we didn't always hit the "worst case" where the static_assert is lost, so thank you for the confirmation regarding GCC/clang.
Quoted 1 line
* since i have to handle non-{clang,gcc} too i still guard with ifdef
Agree.
Quoted 2 lines
* i do need to switch to using
assert.hstatic_assertmacro to inter-operate with c++ in v2
OK. I reviewed that in v2 too.
-Morten
PATCH 2/6 eal: adapt EAL to present rte optional atomics API
In reply to Tyler RetzlaffAdapt the EAL public headers to use rte optional atomics API instead of directly using and exposing toolchain specific atomic builtin intrinsics.
--- app/test/test_mcslock.c | 6 ++-- lib/eal/arm/include/rte_atomic_64.h | 32 +++++++++++----------- lib/eal/arm/include/rte_pause_64.h | 26 +++++++++--------- lib/eal/arm/rte_power_intrinsics.c | 8 +++--- lib/eal/common/eal_common_trace.c | 16 ++++++----- lib/eal/include/generic/rte_atomic.h | 50 +++++++++++++++++----------------- lib/eal/include/generic/rte_pause.h | 38 +++++++++++++------------- lib/eal/include/generic/rte_rwlock.h | 47 +++++++++++++++++--------------- lib/eal/include/generic/rte_spinlock.h | 19 ++++++------- lib/eal/include/rte_mcslock.h | 50 +++++++++++++++++----------------- lib/eal/include/rte_pflock.h | 24 ++++++++-------- lib/eal/include/rte_seqcount.h | 18 ++++++------ lib/eal/include/rte_ticketlock.h | 42 ++++++++++++++-------------- lib/eal/include/rte_trace_point.h | 4 +-- lib/eal/ppc/include/rte_atomic.h | 50 +++++++++++++++++----------------- lib/eal/x86/include/rte_atomic.h | 4 +-- lib/eal/x86/include/rte_spinlock.h | 2 +- lib/eal/x86/rte_power_intrinsics.c | 6 ++-- 18 files changed, 225 insertions(+), 217 deletions(-)
app/test/test_mcslock.c +3 -3
@@ -36,9 +36,9 @@* lock multiple times. */ -rte_mcslock_t *p_ml; -rte_mcslock_t *p_ml_try; -rte_mcslock_t *p_ml_perf; +rte_mcslock_t * __rte_atomic p_ml; +rte_mcslock_t * __rte_atomic p_ml_try; +rte_mcslock_t * __rte_atomic p_ml_perf;static unsigned int count;
lib/eal/arm/include/rte_atomic_64.h +16 -16
@@ -107,33 +107,33 @@*/ RTE_SET_USED (failure);/* Find invalid memory order */ -RTE_ASSERT (success == __ATOMIC_RELAXED || - success == __ATOMIC_ACQUIRE || - success == __ATOMIC_RELEASE || - success == __ATOMIC_ACQ_REL || - success == __ATOMIC_SEQ_CST); +RTE_ASSERT (success == rte_memory_order_relaxed || + success == rte_memory_order_acquire || + success == rte_memory_order_release || + success == rte_memory_order_acq_rel || + success == rte_memory_order_seq_cst);rte_int128_t expected = *exp;rte_int128_t desired = *src;rte_int128_t old;#if defined(__ARM_FEATURE_ATOMICS) || defined(RTE_ARM_FEATURE_ATOMICS ) -if (success == __ATOMIC_RELAXED) +if (success == rte_memory_order_relaxed)__cas_128_relaxed (dst, exp, desired); -else if (success == __ATOMIC_ACQUIRE) +else if (success == rte_memory_order_acquire)__cas_128_acquire (dst, exp, desired); -else if (success == __ATOMIC_RELEASE) +else if (success == rte_memory_order_release)__cas_128_release (dst, exp, desired);else __cas_128_acq_rel (dst, exp, desired); old = *exp;#else -#define __HAS_ACQ (mo) ((mo) != __ATOMIC_RELAXED && (mo) != __ATOMIC_RELEASE) -#define __HAS_RLS (mo) ((mo) == __ATOMIC_RELEASE || (mo) == __ATOMIC_ACQ_REL || \ - (mo) == __ATOMIC_SEQ_CST) +#define __HAS_ACQ (mo) ((mo) != rte_memory_order_relaxed && (mo) != rte_memory_order_release) +#define __HAS_RLS (mo) ((mo) == rte_memory_order_release || (mo) == rte_memory_order_acq_rel || \ + (mo) == rte_memory_order_seq_cst) -int ldx_mo =__HAS_ACQ (success) ? __ATOMIC_ACQUIRE : __ATOMIC_RELAXED; -int stx_mo =__HAS_RLS (success) ? __ATOMIC_RELEASE : __ATOMIC_RELAXED; +int ldx_mo =__HAS_ACQ (success) ? rte_memory_order_acquire : rte_memory_order_relaxed; +int stx_mo =__HAS_RLS (success) ? rte_memory_order_release : rte_memory_order_relaxed;#undef __HAS_ACQ#undef __HAS_RLS
@@ -153,7 +153,7 @@ : "Q" (src->val[0]) \ : "memory"); } -if (ldx_mo == __ATOMIC_RELAXED) +if (ldx_mo == rte_memory_order_relaxed) __LOAD_128("ldxp", dst, old) else__LOAD_128 ("ldaxp" , dst, old)
@@ -170,7 +170,7 @@ : "memory"); }if (likely (old.int128 == expected.int128)) { -if (stx_mo == __ATOMIC_RELAXED) +if (stx_mo == rte_memory_order_relaxed) __STORE_128("stxp", dst, desired, ret) else__STORE_128 ("stlxp" , dst, desired, ret)
@@ -181,7 +181,7 @@* needs to be stored back to ensure it was read * atomically. */ -if (stx_mo == __ATOMIC_RELAXED) +if (stx_mo == rte_memory_order_relaxed) __STORE_128("stxp", dst, old, ret) else__STORE_128 ("stlxp" , dst, old, ret)
lib/eal/arm/include/rte_pause_64.h +13 -13
@@ -41,7 +41,7 @@ static inline void rte_pause(void)* implicitly to exit WFE. */ #define __RTE_ARM_LOAD_EXC_8 (src, dst, memorder) { \ - if (memorder == __ATOMIC_RELAXED) { \ + if (memorder == rte_memory_order_relaxed) { \ asm volatile("ldxrb %w[tmp], [%x[addr]]" \ : [tmp] "=&r" (dst) \ : [addr] "r" (src) \
@@ -60,7 +60,7 @@ static inline void rte_pause(void)* implicitly to exit WFE. */ #define __RTE_ARM_LOAD_EXC_16 (src, dst, memorder) { \ - if (memorder == __ATOMIC_RELAXED) { \ + if (memorder == rte_memory_order_relaxed) { \ asm volatile("ldxrh %w[tmp], [%x[addr]]" \ : [tmp] "=&r" (dst) \ : [addr] "r" (src) \
@@ -79,7 +79,7 @@ static inline void rte_pause(void)* implicitly to exit WFE. */ #define __RTE_ARM_LOAD_EXC_32 (src, dst, memorder) { \ - if (memorder == __ATOMIC_RELAXED) { \ + if (memorder == rte_memory_order_relaxed) { \ asm volatile("ldxr %w[tmp], [%x[addr]]" \ : [tmp] "=&r" (dst) \ : [addr] "r" (src) \
@@ -98,7 +98,7 @@ static inline void rte_pause(void)* implicitly to exit WFE. */ #define __RTE_ARM_LOAD_EXC_64 (src, dst, memorder) { \ - if (memorder == __ATOMIC_RELAXED) { \ + if (memorder == rte_memory_order_relaxed) { \ asm volatile("ldxr %x[tmp], [%x[addr]]" \ : [tmp] "=&r" (dst) \ : [addr] "r" (src) \
@@ -118,7 +118,7 @@ static inline void rte_pause(void)*/ #define __RTE_ARM_LOAD_EXC_128 (src, dst, memorder) { \ volatile rte_int128_t *dst_128 = (volatile rte_int128_t *)&dst; \ - if (memorder == __ATOMIC_RELAXED) { \ + if (memorder == rte_memory_order_relaxed) { \ asm volatile("ldxp %x[tmp0], %x[tmp1], [%x[addr]]" \ : [tmp0] "=&r" (dst_128->val[0]), \ [tmp1] "=&r" (dst_128->val[1]) \
@@ -153,8 +153,8 @@ static inline void rte_pause(void) {uint16_t value; -RTE_BUILD_BUG_ON (memorder != __ATOMIC_ACQUIRE && - memorder != __ATOMIC_RELAXED); +RTE_BUILD_BUG_ON (memorder != rte_memory_order_acquire && + memorder != rte_memory_order_relaxed); __RTE_ARM_LOAD_EXC_16(addr, value, memorder)if (value != expected) {
@@ -172,8 +172,8 @@ static inline void rte_pause(void) {uint32_t value; -RTE_BUILD_BUG_ON (memorder != __ATOMIC_ACQUIRE && - memorder != __ATOMIC_RELAXED); +RTE_BUILD_BUG_ON (memorder != rte_memory_order_acquire && + memorder != rte_memory_order_relaxed); __RTE_ARM_LOAD_EXC_32(addr, value, memorder)if (value != expected) {
@@ -191,8 +191,8 @@ static inline void rte_pause(void) {uint64_t value; -RTE_BUILD_BUG_ON (memorder != __ATOMIC_ACQUIRE && - memorder != __ATOMIC_RELAXED); +RTE_BUILD_BUG_ON (memorder != rte_memory_order_acquire && + memorder != rte_memory_order_relaxed); __RTE_ARM_LOAD_EXC_64(addr, value, memorder)if (value != expected) {
@@ -206,8 +206,8 @@ static inline void rte_pause(void)#define RTE_WAIT_UNTIL_MASKED (addr, mask, cond, expected, memorder) do { \ RTE_BUILD_BUG_ON(!__builtin_constant_p(memorder)); \ - RTE_BUILD_BUG_ON(memorder != __ATOMIC_ACQUIRE && \ - memorder != __ATOMIC_RELAXED); \ + RTE_BUILD_BUG_ON(memorder != rte_memory_order_acquire && \ + memorder != rte_memory_order_relaxed); \ const uint32_t size = sizeof(*(addr)) << 3; \ typeof(*(addr)) expected_value = (expected); \ typeof(*(addr)) value; \
lib/eal/arm/rte_power_intrinsics.c +4 -4
@@ -33,19 +33,19 @@switch (pmc->size) {case sizeof (uint8_t ): - __RTE_ARM_LOAD_EXC_8(pmc->addr, cur_value, __ATOMIC_RELAXED) + __RTE_ARM_LOAD_EXC_8(pmc->addr, cur_value, rte_memory_order_relaxed)__RTE_ARM_WFE () break;case sizeof (uint16_t ): - __RTE_ARM_LOAD_EXC_16(pmc->addr, cur_value, __ATOMIC_RELAXED) + __RTE_ARM_LOAD_EXC_16(pmc->addr, cur_value, rte_memory_order_relaxed)__RTE_ARM_WFE () break;case sizeof (uint32_t ): - __RTE_ARM_LOAD_EXC_32(pmc->addr, cur_value, __ATOMIC_RELAXED) + __RTE_ARM_LOAD_EXC_32(pmc->addr, cur_value, rte_memory_order_relaxed)__RTE_ARM_WFE () break;case sizeof (uint64_t ): - __RTE_ARM_LOAD_EXC_64(pmc->addr, cur_value, __ATOMIC_RELAXED) + __RTE_ARM_LOAD_EXC_64(pmc->addr, cur_value, rte_memory_order_relaxed)__RTE_ARM_WFE () break;default :
lib/eal/common/eal_common_trace.c +9 -7
@@ -103,11 +103,11 @@ struct trace_point_head * trace_mode_set(rte_trace_point_t *t, enum rte_trace_mode mode) {if (mode ==RTE_TRACE_MODE_OVERWRITE ) -__atomic_fetch_and (t, ~__RTE_TRACE_FIELD_ENABLE_DISCARD, - __ATOMIC_RELEASE); +rte_atomic_fetch_and_explicit (t, ~__RTE_TRACE_FIELD_ENABLE_DISCARD, + rte_memory_order_release);else -__atomic_fetch_or (t, __RTE_TRACE_FIELD_ENABLE_DISCARD, - __ATOMIC_RELEASE); +rte_atomic_fetch_or_explicit (t, __RTE_TRACE_FIELD_ENABLE_DISCARD, + rte_memory_order_release); }void
@@ -141,7 +141,7 @@ rte_trace_mode rte_trace_mode_get(void)if (trace_point_is_invalid (t))return false; - val =__atomic_load_n (t, __ATOMIC_ACQUIRE); + val =rte_atomic_load_explicit (t, rte_memory_order_acquire);return (val & __RTE_TRACE_FIELD_ENABLE_MASK) !=0 ; }
@@ -153,7 +153,8 @@ rte_trace_mode rte_trace_mode_get(void)if (trace_point_is_invalid (t))return -ERANGE ; - prev =__atomic_fetch_or (t, __RTE_TRACE_FIELD_ENABLE_MASK, __ATOMIC_RELEASE); + prev =rte_atomic_fetch_or_explicit (t, __RTE_TRACE_FIELD_ENABLE_MASK, + rte_memory_order_release);if ((prev & __RTE_TRACE_FIELD_ENABLE_MASK) ==0 )__atomic_fetch_add (&trace.status,1 , __ATOMIC_RELEASE);return 0 ;
@@ -167,7 +168,8 @@ rte_trace_mode rte_trace_mode_get(void)if (trace_point_is_invalid (t))return -ERANGE ; - prev =__atomic_fetch_and (t, ~__RTE_TRACE_FIELD_ENABLE_MASK, __ATOMIC_RELEASE); + prev =rte_atomic_fetch_and_explicit (t, ~__RTE_TRACE_FIELD_ENABLE_MASK, + rte_memory_order_release);if ((prev & __RTE_TRACE_FIELD_ENABLE_MASK) !=0 )__atomic_fetch_sub (&trace.status,1 , __ATOMIC_RELEASE);return 0 ;
lib/eal/include/generic/rte_atomic.h +25 -25
194 lines not on this page · raw message
lib/eal/include/generic/rte_pause.h +19 -19
@@ -35,13 +35,13 @@* A 16-bit expected value to be in the memory location. * @param memorder * Two different memory orders that can be specified: -* __ATOMIC_ACQUIRE and __ATOMIC_RELAXED. These map to +* rte_memory_order_acquire and rte_memory_order_relaxed. These map to * C++11 memory orders with the same names, see the C++11 standard or * the GCC wiki on atomic synchronization for detailed definition. */ static __rte_always_inline voidrte_wait_until_equal_16 (volatile uint16_t *addr,uint16_t expected, -int memorder); +rte_memory_order memorder);/** * Wait for *addr to be updated with a 32-bit expected value, with a relaxed
@@ -53,13 +53,13 @@* A 32-bit expected value to be in the memory location. * @param memorder * Two different memory orders that can be specified: -* __ATOMIC_ACQUIRE and __ATOMIC_RELAXED. These map to +* rte_memory_order_acquire and rte_memory_order_relaxed. These map to * C++11 memory orders with the same names, see the C++11 standard or * the GCC wiki on atomic synchronization for detailed definition. */ static __rte_always_inline voidrte_wait_until_equal_32 (volatile uint32_t *addr,uint32_t expected, -int memorder); +rte_memory_order memorder);/** * Wait for *addr to be updated with a 64-bit expected value, with a relaxed
@@ -71,42 +71,42 @@* A 64-bit expected value to be in the memory location. * @param memorder * Two different memory orders that can be specified: -* __ATOMIC_ACQUIRE and __ATOMIC_RELAXED. These map to +* rte_memory_order_acquire and rte_memory_order_relaxed. These map to * C++11 memory orders with the same names, see the C++11 standard or * the GCC wiki on atomic synchronization for detailed definition. */ static __rte_always_inline voidrte_wait_until_equal_64 (volatile uint64_t *addr,uint64_t expected, -int memorder); +rte_memory_order memorder);#ifndef RTE_WAIT_UNTIL_EQUAL_ARCH_DEFINED static __rte_always_inline voidrte_wait_until_equal_16 (volatile uint16_t *addr,uint16_t expected, -int memorder) +rte_memory_order memorder) { -assert (memorder == __ATOMIC_ACQUIRE || memorder == __ATOMIC_RELAXED); +assert (memorder == rte_memory_order_acquire || memorder == rte_memory_order_relaxed); -while (__atomic_load_n (addr, memorder) != expected) +while (rte_atomic_load_explicit (addr, memorder) != expected)rte_pause (); } static __rte_always_inline voidrte_wait_until_equal_32 (volatile uint32_t *addr,uint32_t expected, -int memorder) +rte_memory_order memorder) { -assert (memorder == __ATOMIC_ACQUIRE || memorder == __ATOMIC_RELAXED); +assert (memorder == rte_memory_order_acquire || memorder == rte_memory_order_relaxed); -while (__atomic_load_n (addr, memorder) != expected) +while (rte_atomic_load_explicit (addr, memorder) != expected)rte_pause (); } static __rte_always_inline voidrte_wait_until_equal_64 (volatile uint64_t *addr,uint64_t expected, -int memorder) +rte_memory_order memorder) { -assert (memorder == __ATOMIC_ACQUIRE || memorder == __ATOMIC_RELAXED); +assert (memorder == rte_memory_order_acquire || memorder == rte_memory_order_relaxed); -while (__atomic_load_n (addr, memorder) != expected) +while (rte_atomic_load_explicit (addr, memorder) != expected)rte_pause (); }
@@ -124,16 +124,16 @@* An expected value to be in the memory location. * @param memorder * Two different memory orders that can be specified: -* __ATOMIC_ACQUIRE and __ATOMIC_RELAXED. These map to +* rte_memory_order_acquire and rte_memory_order_relaxed. These map to * C++11 memory orders with the same names, see the C++11 standard or * the GCC wiki on atomic synchronization for detailed definition. */ #define RTE_WAIT_UNTIL_MASKED (addr, mask, cond, expected, memorder) do { \ RTE_BUILD_BUG_ON(!__builtin_constant_p(memorder)); \ - RTE_BUILD_BUG_ON(memorder != __ATOMIC_ACQUIRE && \ - memorder != __ATOMIC_RELAXED); \ + RTE_BUILD_BUG_ON((memorder) != rte_memory_order_acquire && \ + (memorder) != rte_memory_order_relaxed); \ typeof(*(addr)) expected_value = (expected); \ - while (!((__atomic_load_n((addr), (memorder)) & (mask)) cond \ + while (!((rte_atomic_load_explicit((addr), (memorder)) & (mask)) cond \ expected_value)) \ rte_pause(); \ } while (0)
lib/eal/include/generic/rte_rwlock.h +25 -22
126 lines not on this page · raw message
lib/eal/include/generic/rte_spinlock.h +9 -10
@@ -28,7 +28,7 @@* The rte_spinlock_t type. */ typedef struct __rte_lockable { -volatile int locked;/**< lock status 0 = unlocked, 1 = locked */ + volatile int __rte_atomic locked; /**< lock status 0 = unlocked, 1 = locked */ }rte_spinlock_t ;/**
@@ -65,10 +65,10 @@ {int exp =0 ; -while (!__atomic_compare_exchange_n (&sl->locked, &exp,1 ,0 , - __ATOMIC_ACQUIRE, __ATOMIC_RELAXED)) { -rte_wait_until_equal_32 ((volatile uint32_t *)&sl->locked, -0 , __ATOMIC_RELAXED); +while (!rte_atomic_compare_exchange_strong_explicit (&sl->locked, &exp,1 , + rte_memory_order_acquire, rte_memory_order_relaxed)) { +rte_wait_until_equal_32 ((volatile uint32_t *)(uintptr_t )&sl->locked, +0 , rte_memory_order_relaxed); exp =0 ; } }
@@ -89,7 +89,7 @@ rte_spinlock_unlock(rte_spinlock_t *sl) __rte_no_thread_safety_analysis { -__atomic_store_n (&sl->locked,0 , __ATOMIC_RELEASE); +rte_atomic_store_explicit (&sl->locked,0 , rte_memory_order_release); }#endif
@@ -112,9 +112,8 @@ __rte_no_thread_safety_analysis {int exp =0 ; -return __atomic_compare_exchange_n (&sl->locked, &exp,1 , -0 ,/* disallow spurious failure */ - __ATOMIC_ACQUIRE, __ATOMIC_RELAXED); +return rte_atomic_compare_exchange_strong_explicit (&sl->locked, &exp,1 , + rte_memory_order_acquire, rte_memory_order_relaxed); }#endif
@@ -128,7 +127,7 @@*/ static inline int rte_spinlock_is_locked (rte_spinlock_t *sl) { -return __atomic_load_n (&sl->locked, __ATOMIC_ACQUIRE); +return rte_atomic_load_explicit (&sl->locked, rte_memory_order_acquire); }/**
lib/eal/include/rte_mcslock.h +25 -25
133 lines not on this page · raw message
lib/eal/include/rte_pflock.h +13 -11
@@ -40,8 +40,8 @@*/ struct rte_pflock {struct { -uint16_t in; -uint16_t out; + uint16_t __rte_atomic in; + uint16_t __rte_atomic out; } rd, wr; };typedef struct rte_pflock rte_pflock_t ;
@@ -116,14 +116,14 @@ struct rte_pflock {* If no writer is present, then the operation has completed * successfully. */ - w =__atomic_fetch_add (&pf->rd.in,RTE_PFLOCK_RINC , __ATOMIC_ACQUIRE) + w =rte_atomic_fetch_add_explicit (&pf->rd.in,RTE_PFLOCK_RINC , rte_memory_order_acquire) &RTE_PFLOCK_WBITS ;if (w ==0 )return ;/* Wait for current write phase to complete. */ RTE_WAIT_UNTIL_MASKED(&pf->rd.in, RTE_PFLOCK_WBITS, !=, w, - __ATOMIC_ACQUIRE); + rte_memory_order_acquire); }/**
@@ -139,7 +139,7 @@ struct rte_pflock {static inline void rte_pflock_read_unlock (rte_pflock_t *pf) { -__atomic_fetch_add (&pf->rd.out,RTE_PFLOCK_RINC , __ATOMIC_RELEASE); +rte_atomic_fetch_add_explicit (&pf->rd.out,RTE_PFLOCK_RINC , rte_memory_order_release); }/**
@@ -160,8 +160,9 @@ struct rte_pflock {/* Acquire ownership of write-phase. * This is same as rte_ticketlock_lock(). */ - ticket =__atomic_fetch_add (&pf->wr.in,1 , __ATOMIC_RELAXED); -rte_wait_until_equal_16 (&pf->wr.out, ticket, __ATOMIC_ACQUIRE); + ticket =rte_atomic_fetch_add_explicit (&pf->wr.in,1 , rte_memory_order_relaxed); +rte_wait_until_equal_16 ((uint16_t *)(uintptr_t )&pf->wr.out, ticket, + rte_memory_order_acquire);/* * Acquire ticket on read-side in order to allow them
@@ -172,10 +173,11 @@ struct rte_pflock {* speculatively. */ w =RTE_PFLOCK_PRES | (ticket &RTE_PFLOCK_PHID ); - ticket =__atomic_fetch_add (&pf->rd.in, w, __ATOMIC_RELAXED); + ticket =rte_atomic_fetch_add_explicit (&pf->rd.in, w, rte_memory_order_relaxed);/* Wait for any pending readers to flush. */ -rte_wait_until_equal_16 (&pf->rd.out, ticket, __ATOMIC_ACQUIRE); +rte_wait_until_equal_16 ((uint16_t *)(uintptr_t )&pf->rd.out, ticket, + rte_memory_order_acquire); }/**
@@ -192,10 +194,10 @@ struct rte_pflock { rte_pflock_write_unlock(rte_pflock_t *pf) {/* Migrate from write phase to read phase. */ -__atomic_fetch_and (&pf->rd.in,RTE_PFLOCK_LSB , __ATOMIC_RELEASE); +rte_atomic_fetch_and_explicit (&pf->rd.in,RTE_PFLOCK_LSB , rte_memory_order_release);/* Allow other writers to continue. */ -__atomic_fetch_add (&pf->wr.out,1 , __ATOMIC_RELEASE); +rte_atomic_fetch_add_explicit (&pf->wr.out,1 , rte_memory_order_release); }#ifdef __cplusplus
lib/eal/include/rte_seqcount.h +9 -9
@@ -31,7 +31,7 @@* The RTE seqcount type. */ typedef struct { -uint32_t sn;/**< A sequence number for the protected data. */ + uint32_t __rte_atomic sn; /**< A sequence number for the protected data. */ }rte_seqcount_t ;/**
@@ -105,11 +105,11 @@static inline uint32_t rte_seqcount_read_begin (const rte_seqcount_t *seqcount) { -/* __ATOMIC_ACQUIRE to prevent loads after (in program order) +/* rte_memory_order_acquire to prevent loads after (in program order) * from happening before the sn load. Synchronizes-with the * store release in rte_seqcount_write_end(). */ -return __atomic_load_n (&seqcount->sn, __ATOMIC_ACQUIRE); +return rte_atomic_load_explicit (&seqcount->sn, rte_memory_order_acquire); }/**
@@ -160,9 +160,9 @@return true;/* make sure the data loads happens before the sn load */ -rte_atomic_thread_fence (__ATOMIC_ACQUIRE); +rte_atomic_thread_fence (rte_memory_order_acquire); - end_sn =__atomic_load_n (&seqcount->sn, __ATOMIC_RELAXED); + end_sn =rte_atomic_load_explicit (&seqcount->sn, rte_memory_order_relaxed);/* A writer incremented the sequence number during this read * critical section.
@@ -204,12 +204,12 @@ sn = seqcount->sn +1 ; -__atomic_store_n (&seqcount->sn, sn, __ATOMIC_RELAXED); +rte_atomic_store_explicit (&seqcount->sn, sn, rte_memory_order_relaxed); -/* __ATOMIC_RELEASE to prevent stores after (in program order) +/* rte_memory_order_release to prevent stores after (in program order) * from happening before the sn store. */ -rte_atomic_thread_fence (__ATOMIC_RELEASE); +rte_atomic_thread_fence (rte_memory_order_release); }/**
@@ -236,7 +236,7 @@ sn = seqcount->sn +1 ;/* Synchronizes-with the load acquire in rte_seqcount_read_begin(). */ -__atomic_store_n (&seqcount->sn, sn, __ATOMIC_RELEASE); +rte_atomic_store_explicit (&seqcount->sn, sn, rte_memory_order_release); }#ifdef __cplusplus
lib/eal/include/rte_ticketlock.h +22 -20
115 lines not on this page · raw message
lib/eal/include/rte_trace_point.h +2 -2
@@ -32,7 +32,7 @@#include <rte_uuid.h> /** The tracepoint object. */ -typedef uint64_t rte_trace_point_t ; +typedef uint64_t __rte_atomic rte_trace_point_t;/** * Macro to define the tracepoint arguments in RTE_TRACE_POINT macro.
@@ -358,7 +358,7 @@ struct __rte_trace_header {#define __rte_trace_point_emit_header_generic (t) \ void *mem; \ do { \ - const uint64_t val = __atomic_load_n(t, __ATOMIC_ACQUIRE); \ + const uint64_t val = rte_atomic_load_explicit(t, rte_memory_order_acquire); \ if (likely(!(val & __RTE_TRACE_FIELD_ENABLE_MASK))) \ return; \ mem = __rte_trace_mem_get(val); \
lib/eal/ppc/include/rte_atomic.h +25 -25
161 lines not on this page · raw message
lib/eal/x86/include/rte_atomic.h +2 -2
@@ -82,14 +82,14 @@/** * Synchronization fence between threads based on the specified memory order. * -* On x86 the __atomic_thread_fence(__ATOMIC_SEQ_CST) generates full 'mfence' +* On x86 the __atomic_thread_fence(rte_memory_order_seq_cst) generates full 'mfence' * which is quite expensive. The optimized implementation of rte_smp_mb is * used instead. */ static __rte_always_inline voidrte_atomic_thread_fence (int memorder) { -if (memorder == __ATOMIC_SEQ_CST) +if (memorder == rte_memory_order_seq_cst)rte_smp_mb ();else __atomic_thread_fence (memorder);
lib/eal/x86/include/rte_spinlock.h +1 -1
@@ -78,7 +78,7 @@ static inline int rte_tm_supported(void) }static inline int -rte_try_tm (volatile int *lock) +rte_try_tm(volatile int __rte_atomic *lock) {int i, retries;
lib/eal/x86/rte_power_intrinsics.c +3 -3
@@ -23,9 +23,9 @@uint64_t val;/* trigger a write but don't change the value */ - val =__atomic_load_n ((volatile uint64_t *)addr, __ATOMIC_RELAXED); -__atomic_compare_exchange_n ((volatile uint64_t *)addr, &val, val,0 , - __ATOMIC_RELAXED, __ATOMIC_RELAXED); + val =rte_atomic_load_explicit ((volatile uint64_t *)addr, rte_memory_order_relaxed); +rte_atomic_compare_exchange_strong_explicit ((volatile uint64_t *)addr, &val, val, + rte_memory_order_relaxed, rte_memory_order_relaxed); }static bool wait_supported;
Signature · 1 line
1.8.3.1
PATCH 3/6 eal: add rte atomic qualifier with casts
In reply to Tyler RetzlaffIntroduce __rte_atomic qualifying casts in rte_optional atomics inline functions to prevent cascading the need to pass __rte_atomic qualified arguments.
Warning, this is really implementation dependent and being done temporarily to avoid having to convert more of the libraries and tests in DPDK in the initial series that introduces the API. The consequence of the assumption of the ABI of the types in question not being ``the same'' is only a risk that may be realized when enable_stdatomic=true.
--- lib/eal/include/generic/rte_atomic.h | 48 ++++++++++++++++++++++++------------ lib/eal/include/generic/rte_pause.h | 9 ++++--- lib/eal/x86/rte_power_intrinsics.c | 7 +++--- 3 files changed, 42 insertions(+), 22 deletions(-)
lib/eal/include/generic/rte_atomic.h +32 -16
@@ -273,7 +273,8 @@static inline void rte_atomic16_add (rte_atomic16_t *v,int16_t inc) { -rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst); + rte_atomic_fetch_add_explicit((volatile int16_t __rte_atomic *)&v->cnt, inc, + rte_memory_order_seq_cst); }/**
@@ -287,7 +288,8 @@static inline void rte_atomic16_sub (rte_atomic16_t *v,int16_t dec) { -rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst); + rte_atomic_fetch_sub_explicit((volatile int16_t __rte_atomic *)&v->cnt, dec, + rte_memory_order_seq_cst); }/**
@@ -340,7 +342,8 @@static inline int16_t rte_atomic16_add_return (rte_atomic16_t *v,int16_t inc) { -return rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst) + inc; + return rte_atomic_fetch_add_explicit((volatile int16_t __rte_atomic *)&v->cnt, inc, + rte_memory_order_seq_cst) + inc; }/**
@@ -360,7 +363,8 @@static inline int16_t rte_atomic16_sub_return (rte_atomic16_t *v,int16_t dec) { -return rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst) - dec; + return rte_atomic_fetch_sub_explicit((volatile int16_t __rte_atomic *)&v->cnt, dec, + rte_memory_order_seq_cst) - dec; }/**
@@ -379,7 +383,8 @@#ifdef RTE_FORCE_INTRINSICS static inline int rte_atomic16_inc_and_test (rte_atomic16_t *v) { -return rte_atomic_fetch_add_explicit (&v->cnt,1 , rte_memory_order_seq_cst) +1 ==0 ; + return rte_atomic_fetch_add_explicit((volatile int16_t __rte_atomic *)&v->cnt, 1, + rte_memory_order_seq_cst) +1 ==0 ; }#endif
@@ -399,7 +404,8 @@ static inline int rte_atomic16_inc_and_test(rte_atomic16_t *v)#ifdef RTE_FORCE_INTRINSICS static inline int rte_atomic16_dec_and_test (rte_atomic16_t *v) { -return rte_atomic_fetch_sub_explicit (&v->cnt,1 , rte_memory_order_seq_cst) -1 ==0 ; + return rte_atomic_fetch_sub_explicit((volatile int16_t __rte_atomic *)&v->cnt, 1, + rte_memory_order_seq_cst) -1 ==0 ; }#endif
@@ -552,7 +558,8 @@ static inline void rte_atomic16_clear(rte_atomic16_t *v)static inline void rte_atomic32_add (rte_atomic32_t *v,int32_t inc) { -rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst); + rte_atomic_fetch_add_explicit((volatile int32_t __rte_atomic *)&v->cnt, inc, + rte_memory_order_seq_cst); }/**
@@ -566,7 +573,8 @@ static inline void rte_atomic16_clear(rte_atomic16_t *v)static inline void rte_atomic32_sub (rte_atomic32_t *v,int32_t dec) { -rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst); + rte_atomic_fetch_sub_explicit((volatile int32_t __rte_atomic *)&v->cnt, dec, + rte_memory_order_seq_cst); }/**
@@ -619,7 +627,8 @@ static inline void rte_atomic16_clear(rte_atomic16_t *v)static inline int32_t rte_atomic32_add_return (rte_atomic32_t *v,int32_t inc) { -return rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst) + inc; + return rte_atomic_fetch_add_explicit((volatile int32_t __rte_atomic *)&v->cnt, inc, + rte_memory_order_seq_cst) + inc; }/**
@@ -639,7 +648,8 @@ static inline void rte_atomic16_clear(rte_atomic16_t *v)static inline int32_t rte_atomic32_sub_return (rte_atomic32_t *v,int32_t dec) { -return rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst) - dec; + return rte_atomic_fetch_sub_explicit((volatile int32_t __rte_atomic *)&v->cnt, dec, + rte_memory_order_seq_cst) - dec; }/**
@@ -658,7 +668,8 @@ static inline void rte_atomic16_clear(rte_atomic16_t *v)#ifdef RTE_FORCE_INTRINSICS static inline int rte_atomic32_inc_and_test (rte_atomic32_t *v) { -return rte_atomic_fetch_add_explicit (&v->cnt,1 , rte_memory_order_seq_cst) +1 ==0 ; + return rte_atomic_fetch_add_explicit((volatile int32_t __rte_atomic *)&v->cnt, 1, + rte_memory_order_seq_cst) +1 ==0 ; }#endif
@@ -678,7 +689,8 @@ static inline int rte_atomic32_inc_and_test(rte_atomic32_t *v)#ifdef RTE_FORCE_INTRINSICS static inline int rte_atomic32_dec_and_test (rte_atomic32_t *v) { -return rte_atomic_fetch_sub_explicit (&v->cnt,1 , rte_memory_order_seq_cst) -1 ==0 ; + return rte_atomic_fetch_sub_explicit((volatile int32_t __rte_atomic *)&v->cnt, 1, + rte_memory_order_seq_cst) -1 ==0 ; }#endif
@@ -884,7 +896,8 @@ static inline void rte_atomic32_clear(rte_atomic32_t *v)static inline void rte_atomic64_add (rte_atomic64_t *v,int64_t inc) { -rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst); + rte_atomic_fetch_add_explicit((volatile int64_t __rte_atomic *)&v->cnt, inc, + rte_memory_order_seq_cst); }#endif
@@ -903,7 +916,8 @@ static inline void rte_atomic32_clear(rte_atomic32_t *v)static inline void rte_atomic64_sub (rte_atomic64_t *v,int64_t dec) { -rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst); + rte_atomic_fetch_sub_explicit((volatile int64_t __rte_atomic *)&v->cnt, dec, + rte_memory_order_seq_cst); }#endif
@@ -961,7 +975,8 @@ static inline void rte_atomic32_clear(rte_atomic32_t *v)static inline int64_t rte_atomic64_add_return (rte_atomic64_t *v,int64_t inc) { -return rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst) + inc; + return rte_atomic_fetch_add_explicit((volatile int64_t __rte_atomic *)&v->cnt, inc, + rte_memory_order_seq_cst) + inc; }#endif
@@ -985,7 +1000,8 @@ static inline void rte_atomic32_clear(rte_atomic32_t *v)static inline int64_t rte_atomic64_sub_return (rte_atomic64_t *v,int64_t dec) { -return rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst) - dec; + return rte_atomic_fetch_sub_explicit((volatile int64_t __rte_atomic *)&v->cnt, dec, + rte_memory_order_seq_cst) - dec; }#endif
lib/eal/include/generic/rte_pause.h +6 -3
@@ -86,7 +86,8 @@ {assert (memorder == rte_memory_order_acquire || memorder == rte_memory_order_relaxed); -while (rte_atomic_load_explicit (addr, memorder) != expected) + while (rte_atomic_load_explicit((volatile uint16_t __rte_atomic *)addr, memorder) + != expected)rte_pause (); }
@@ -96,7 +97,8 @@ {assert (memorder == rte_memory_order_acquire || memorder == rte_memory_order_relaxed); -while (rte_atomic_load_explicit (addr, memorder) != expected) + while (rte_atomic_load_explicit((volatile uint32_t __rte_atomic *)addr, memorder) + != expected)rte_pause (); }
@@ -106,7 +108,8 @@ {assert (memorder == rte_memory_order_acquire || memorder == rte_memory_order_relaxed); -while (rte_atomic_load_explicit (addr, memorder) != expected) + while (rte_atomic_load_explicit((volatile uint64_t __rte_atomic *)addr, memorder) + != expected)rte_pause (); }
lib/eal/x86/rte_power_intrinsics.c +4 -3
@@ -23,9 +23,10 @@uint64_t val;/* trigger a write but don't change the value */ - val =rte_atomic_load_explicit ((volatile uint64_t *)addr, rte_memory_order_relaxed); -rte_atomic_compare_exchange_strong_explicit ((volatile uint64_t *)addr, &val, val, - rte_memory_order_relaxed, rte_memory_order_relaxed); + val = rte_atomic_load_explicit((volatile uint64_t __rte_atomic *)addr, + rte_memory_order_relaxed); + rte_atomic_compare_exchange_strong_explicit((volatile uint64_t __rte_atomic *)addr, + &val, val, rte_memory_order_relaxed, rte_memory_order_relaxed); }static bool wait_supported;
Signature · 1 line
1.8.3.1
PATCH 4/6 distributor: adapt for EAL optional atomics API changes
In reply to Tyler RetzlaffAdapt distributor for EAL optional atomics API changes
--- lib/distributor/distributor_private.h | 2 +- lib/distributor/rte_distributor_single.c | 44 ++++++++++++++++---------------- 2 files changed, 23 insertions(+), 23 deletions(-)
lib/distributor/distributor_private.h +1 -1
@@ -52,7 +52,7 @@* Only 64-bits of the memory is actually used though. */ union rte_distributor_buffer_single { -volatile int64_t bufptr64; + volatile int64_t __rte_atomic bufptr64;char pad[RTE_CACHE_LINE_SIZE *3 ]; } __rte_cache_aligned;
lib/distributor/rte_distributor_single.c +22 -22
117 lines not on this page · raw message
Signature · 1 line
1.8.3.1
PATCH 5/6 bpf: adapt for EAL optional atomics API changes
In reply to Tyler RetzlaffAdapt bpf for EAL optional atomics API changes
--- lib/bpf/bpf_pkt.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)
lib/bpf/bpf_pkt.c +3 -3
@@ -25,7 +25,7 @@struct bpf_eth_cbi {/* used by both data & control path */ -uint32_t use;/*usage counter */ + uint32_t __rte_atomic use; /*usage counter */const struct rte_eth_rxtx_callback *cb;/* callback handle */ struct rte_bpf *bpf;struct rte_bpf_jit jit;
@@ -110,8 +110,8 @@ struct bpf_eth_cbh {/* in use, busy wait till current RX/TX iteration is finished */ if ((puse &BPF_ETH_CBI_INUSE ) !=0 ) { -RTE_WAIT_UNTIL_MASKED ((uint32_t *)(uintptr_t )&cbi->use, - UINT32_MAX, !=, puse, __ATOMIC_RELAXED); + RTE_WAIT_UNTIL_MASKED((uint32_t __rte_atomic *)(uintptr_t)&cbi->use, + UINT32_MAX, !=, puse, rte_memory_order_relaxed); } }
Signature · 1 line
1.8.3.1
PATCH 6/6 devtools: forbid new direct use of GCC atomic builtins
In reply to Tyler RetzlaffRefrain from using compiler __atomic_xxx builtins DPDK now requires the use of rte_atomic_<op>_explicit macros when operating on DPDK atomic variables.
--- devtools/checkpatches.sh | 8 ++++++++ 1 file changed, 8 insertions(+)
devtools/checkpatches.sh +8 -0
@@ -102,6 +102,14 @@ check_forbidden_additions() { # <patch> -f $(dirname $(readlink -f $0))/check-forbidden-tokens.awk \ "$1" || res=1 + # refrain from using compiler __atomic_xxx builtins + awk -v FOLDERS="lib drivers app examples" \ + -v EXPRESSIONS="__atomic_.*\\\(" \ + -v RET_ON_FAIL=1 \ + -v MESSAGE='Using __atomic_xxx builtins' \ + -f $(dirname $(readlink -f $0))/check-forbidden-tokens.awk \ + "$1" || res=1 + # refrain from using compiler __atomic_thread_fence() # It should be avoided on x86 for SMP case. awk -v FOLDERS="lib drivers app examples" \
Signature · 1 line
1.8.3.1
Re: [PATCH 6/6] devtools: forbid new direct use of GCC atomic builtins
Quoted 7 lines · Tyler Retzlaff
Refrain from using compiler
__atomic_xxxbuiltins DPDK now requires the use of rte_atomic_<op>_explicit macros when operating on DPDK atomic variables.Signed-off-by Tyler Retzlaff <[email protected]>Acked-by Morten Brørup <[email protected]>---
RE: [PATCH 6/6] devtools: forbid new direct use of GCC atomic builtins
In reply to Tyler RetzlaffQuoted 9 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Friday, 11 August 2023 03.32Refrain from using compiler
__atomic_xxxbuiltins DPDK now requires the use of rte_atomic_<op>_explicit macros when operating on DPDK atomic variables.Signed-off-by Tyler Retzlaff <[email protected]>Acked-by Morten Brørup <[email protected]>
The Acked-by should have been:
Quoted 9 lines
--- devtools/checkpatches.sh | 8 ++++++++ 1 file changed, 8 insertions(+)devtools/checkpatches.sh +0
-0
devtools/checkpatches.sh check_forbidden_additions() { # <patch> · line 108
Quoted 7 lines
devtools/checkpatches.sh +3
-0@@ -102,6 +102,14 @@ check_forbidden_additions() { # <patch> -f $(dirname $(readlink -f $0))/check-forbidden-tokens.awk \ "$1" || res=1 + # refrain from using compiler __atomic_xxx builtins + awk -v FOLDERS="lib drivers app examples" \ + -v EXPRESSIONS="__atomic_.*\\\(" \
This expression is a superset of other expressions in checkpatches (search for "__atomic" in the checkpatches, and you'll find them). Perhaps they can be removed?
Quoted 11 lines
devtools/checkpatches.sh +5
-0+ -v RET_ON_FAIL=1 \ + -v MESSAGE='Using __atomic_xxx builtins' \ + -f $(dirname $(readlink -f $0))/check-forbidden-tokens.awk \ + "$1" || res=1 + # refrain from using compiler __atomic_thread_fence()devtools/checkpatches.sh +0
-1# It should be avoided on x86 for SMP case. awk -v FOLDERS="lib drivers app examples" \ --1.8.3.1
Re: [PATCH 6/6] devtools: forbid new direct use of GCC atomic builtins
Quoted 7 lines · Morten Brørup
Quoted 7 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Friday, 11 August 2023 03.32Refrain from using compiler
__atomic_xxxbuiltins DPDK now requires the use of rte_atomic_<op>_explicit macros when operating on DPDK atomic variables.
Quoted 5 lines
Quoted 2 lines
Signed-off-by Tyler Retzlaff <[email protected]>Acked-by Morten Brørup <[email protected]>The Acked-by should have been:
Suggested-by Morten Brørup <[email protected]>
ooh, did i make a mistake? i was carrying the ack from my abandoned series (or i thought you had acked this patch on that series sorry).
i'll change it to suggested-by.
thanks!
Quoted 14 lines
Quoted 14 lines
--- devtools/checkpatches.sh | 8 ++++++++ 1 file changed, 8 insertions(+)devtools/checkpatches.sh +0
-0@@ -102,6 +102,14 @@ check_forbidden_additions() { # <patch> -f $(dirname $(readlink -f $0))/check-forbidden-tokens.awk \ "$1" || res=1
Quoted 5 lines
devtools/checkpatches.sh check_forbidden_additions() { # <patch> · line 107
Quoted 3 lines
devtools/checkpatches.sh +3
-0+ # refrain from using compiler __atomic_xxx builtins + awk -v FOLDERS="lib drivers app examples" \ + -v EXPRESSIONS="__atomic_.*\\\(" \This expression is a superset of other expressions in checkpatches (search for "
__atomic" in the checkpatches, and you'll find them). Perhaps they can be removed?
yes, seems like a good idea.
v2
Quoted 13 lines
Quoted 12 lines
devtools/checkpatches.sh +5
-0+ -v RET_ON_FAIL=1 \ + -v MESSAGE='Using __atomic_xxx builtins' \ + -f $(dirname $(readlink -f $0))/check-forbidden-tokens.awk \ + "$1" || res=1 + # refrain from using compiler __atomic_thread_fence()devtools/checkpatches.sh +0
-1# It should be avoided on x86 for SMP case. awk -v FOLDERS="lib drivers app examples" \ --1.8.3.1
RE: [PATCH 6/6] devtools: forbid new direct use of GCC atomic builtins
Quoted 17 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Friday, 11 August 2023 17.57Quoted 7 lines · Morten Brørup
Quoted 7 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Friday, 11 August 2023 03.32Refrain from using compiler
__atomic_xxxbuiltins DPDK now requires the use of rte_atomic_<op>_explicit macros when operating on DPDK atomic variables.Quoted 5 lines
Quoted 2 lines
Signed-off-by Tyler Retzlaff <[email protected]>Acked-by Morten Brørup <[email protected]>The Acked-by should have been:
Suggested-by Morten Brørup <[email protected]>
Quoted 6 lines
ooh, did i make a mistake? i was carrying the ack from my abandoned series (or i thought you had acked this patch on that series sorry).
i'll change it to suggested-by.
thanks!
No problem. Both tags mean that I approve of the concept anyway.
Minor mistakes are bound to come with big sets like this. Better in the comments than in the code. :-)
Quoted 40 lines
Quoted 15 lines
Quoted 15 lines
--- devtools/checkpatches.sh | 8 ++++++++ 1 file changed, 8 insertions(+)devtools/checkpatches.sh +0
-0@@ -102,6 +102,14 @@ check_forbidden_additions() { # <patch> -f $(dirname $(readlink -f $0))/check-forbidden-tokens.awk \ "$1" || res=1Quoted 5 lines
devtools/checkpatches.sh check_forbidden_additions() { # <patch> · line 107
Quoted 3 lines
devtools/checkpatches.sh +3
-0+ # refrain from using compiler __atomic_xxx builtins + awk -v FOLDERS="lib drivers app examples" \ + -v EXPRESSIONS="__atomic_.*\\\(" \This expression is a superset of other expressions in checkpatches (search
for "
__atomic" in the checkpatches, and you'll find them). Perhaps they can be removed?yes, seems like a good idea.
v2
Quoted 13 lines
Quoted 12 lines
devtools/checkpatches.sh +5
-0+ -v RET_ON_FAIL=1 \ + -v MESSAGE='Using __atomic_xxx builtins' \ + -f $(dirname $(readlink -f $0))/check-forbidden-tokens.awk \ + "$1" || res=1 + # refrain from using compiler __atomic_thread_fence()devtools/checkpatches.sh +0
-1# It should be avoided on x86 for SMP case. awk -v FOLDERS="lib drivers app examples" \ --1.8.3.1
v2 cover RFC optional rte optional stdatomics API
In reply to Tyler RetzlaffThis series introduces API additions prefixed in the rte namespace that allow the optional use of stdatomics.h from C11 using enable_stdatomics=true for targets where enable_stdatomics=false no functional change is intended.
Be aware this does not contain all changes to use stdatomics across the DPDK tree it only introduces the minimum to allow the option to be used which is a pre-requisite for a clean CI (probably using clang) that can be run with enable_stdatomics=true enabled.
It is planned that subsequent series will be introduced per lib/driver as appropriate to further enable stdatomics use when enable_stdatomics=true.
Notes:
* Additional libraries beyond EAL make visible atomics use across the API/ABI surface they will be converted in the subsequent series.
* The eal: add rte atomic qualifier with casts patch needs some discussion as to whether or not the legacy rte_atomic APIs should be converted to work with enable_stdatomic=true right now some implementation dependent casts are used to prevent cascading / having to convert too much in the intial series.
* Windows will obviously need complete conversion of libraries including atomics that are not crossing API/ABI boundaries. those conversions will introduced in separate series as new along side the existing msvc series.
Please keep in mind we would like to prioritize the review / acceptance of this patch since it needs to be completed in the 23.11 merge window.
Thank you all for the discussion that lead to the formation of this series.
v2:
* Wrap meson_options.txt option description to newline and indent to be consistent with other options.
* Provide separate typedef of rte_memory_order for enable_stdatomic=true VS enable_stdatomic=false instead of a single typedef to int
note: slight tweak to reviewers feedback i've chosen to use a typedef
for both enable_stdatomic={true,false} (just seemed more consistent)
* Bring in assert.h and use static_assert macro instead of _Static_assert keyword to better interoperate with c/c++
* Directly include rte_stdatomic.h where into other places it is consumed instead of hacking it globally into rte_config.h
* Provide and use __rte_atomic_thread_fence to allow conditional expansion within the body of existing rte_atomic_thread_fence inline function to maintain per-arch optimizations when enable_stdatomic=false
Tyler Retzlaff (6):
eal: provide rte stdatomics optional atomics API
eal: adapt EAL to present rte optional atomics API
eal: add rte atomic qualifier with casts
distributor: adapt for EAL optional atomics API changes
bpf: adapt for EAL optional atomics API changes
devtools: forbid new direct use of GCC atomic builtins
app/test/test_mcslock.c | 6 +- config/meson.build | 1 + devtools/checkpatches.sh | 6 +- lib/bpf/bpf_pkt.c | 6 +- lib/distributor/distributor_private.h | 2 +- lib/distributor/rte_distributor_single.c | 44 ++++---- lib/eal/arm/include/rte_atomic_32.h | 4 +- lib/eal/arm/include/rte_atomic_64.h | 36 +++--- lib/eal/arm/include/rte_pause_64.h | 26 ++--- lib/eal/arm/rte_power_intrinsics.c | 8 +- lib/eal/common/eal_common_trace.c | 16 +-- lib/eal/include/generic/rte_atomic.h | 67 +++++++----- lib/eal/include/generic/rte_pause.h | 42 +++---- lib/eal/include/generic/rte_rwlock.h | 48 ++++---- lib/eal/include/generic/rte_spinlock.h | 20 ++-- lib/eal/include/meson.build | 1 + lib/eal/include/rte_mcslock.h | 51 ++++----- lib/eal/include/rte_pflock.h | 25 +++-- lib/eal/include/rte_seqcount.h | 19 ++-- lib/eal/include/rte_stdatomic.h | 182 +++++++++++++++++++++++++++++++ lib/eal/include/rte_ticketlock.h | 43 ++++---- lib/eal/include/rte_trace_point.h | 5 +- lib/eal/loongarch/include/rte_atomic.h | 4 +- lib/eal/ppc/include/rte_atomic.h | 54 ++++----- lib/eal/riscv/include/rte_atomic.h | 4 +- lib/eal/x86/include/rte_atomic.h | 8 +- lib/eal/x86/include/rte_spinlock.h | 2 +- lib/eal/x86/rte_power_intrinsics.c | 7 +- meson_options.txt | 2 + 29 files changed, 481 insertions(+), 258 deletions(-) create mode 100644 lib/eal/include/rte_stdatomic.h
Signature · 1 line
1.8.3.1
PATCH v2 1/6 eal: provide rte stdatomics optional atomics API
Provide API for atomic operations in the rte namespace that may optionally be configured to use C11 atomics with meson option enable_stdatomics=true
--- config/meson.build | 1 + lib/eal/include/generic/rte_atomic.h | 1 + lib/eal/include/generic/rte_pause.h | 1 + lib/eal/include/generic/rte_rwlock.h | 1 + lib/eal/include/generic/rte_spinlock.h | 1 + lib/eal/include/meson.build | 1 + lib/eal/include/rte_mcslock.h | 1 + lib/eal/include/rte_pflock.h | 1 + lib/eal/include/rte_seqcount.h | 1 + lib/eal/include/rte_stdatomic.h | 182 +++++++++++++++++++++++++++++++++ lib/eal/include/rte_ticketlock.h | 1 + lib/eal/include/rte_trace_point.h | 1 + meson_options.txt | 2 + 13 files changed, 195 insertions(+) create mode 100644 lib/eal/include/rte_stdatomic.h
config/meson.build +1 -0
@@ -303,6 +303,7 @@ endforeach # set other values pulled from the build options dpdk_conf.set('RTE_MAX_ETHPORTS', get_option('max_ethports')) dpdk_conf.set('RTE_LIBEAL_USE_HPET', get_option('use_hpet')) +dpdk_conf.set('RTE_ENABLE_STDATOMIC', get_option('enable_stdatomic')) dpdk_conf.set('RTE_ENABLE_TRACE_FP', get_option('enable_trace_fp')) # values which have defaults which may be overridden dpdk_conf.set('RTE_MAX_VFIO_GROUPS', 64)
lib/eal/include/generic/rte_atomic.h +1 -0
@@ -15,6 +15,7 @@#include <stdint.h> #include <rte_compat.h> #include <rte_common.h> +#include <rte_stdatomic.h> #ifdef __DOXYGEN__
lib/eal/include/generic/rte_pause.h +1 -0
@@ -16,6 +16,7 @@#include <assert.h> #include <rte_common.h> #include <rte_atomic.h> +#include <rte_stdatomic.h> /** * Pause CPU execution for a short while
lib/eal/include/generic/rte_rwlock.h +1 -0
@@ -32,6 +32,7 @@#include <rte_common.h> #include <rte_lock_annotations.h> #include <rte_pause.h> +#include <rte_stdatomic.h> /** * The rte_rwlock_t type.
lib/eal/include/generic/rte_spinlock.h +1 -0
@@ -23,6 +23,7 @@#endif #include <rte_lock_annotations.h> #include <rte_pause.h> +#include <rte_stdatomic.h> /** * The rte_spinlock_t type.
lib/eal/include/meson.build +1 -0
@@ -42,6 +42,7 @@ headers += files( 'rte_seqlock.h', 'rte_service.h', 'rte_service_component.h', + 'rte_stdatomic.h', 'rte_string_fns.h', 'rte_tailq.h', 'rte_thread.h',
lib/eal/include/rte_mcslock.h +1 -0
@@ -27,6 +27,7 @@#include <rte_common.h> #include <rte_pause.h> #include <rte_branch_prediction.h> +#include <rte_stdatomic.h> /** * The rte_mcslock_t type.
lib/eal/include/rte_pflock.h +1 -0
@@ -34,6 +34,7 @@#include <rte_compat.h> #include <rte_common.h> #include <rte_pause.h> +#include <rte_stdatomic.h> /** * The rte_pflock_t type.
lib/eal/include/rte_seqcount.h +1 -0
@@ -26,6 +26,7 @@#include <rte_atomic.h> #include <rte_branch_prediction.h> #include <rte_compat.h> +#include <rte_stdatomic.h> /** * The RTE seqcount type.
lib/eal/include/rte_stdatomic.h new file +182 -0
182 lines not on this page · raw message
lib/eal/include/rte_ticketlock.h +1 -0
@@ -24,6 +24,7 @@#include <rte_common.h> #include <rte_lcore.h> #include <rte_pause.h> +#include <rte_stdatomic.h> /** * The rte_ticketlock_t type.
lib/eal/include/rte_trace_point.h +1 -0
@@ -30,6 +30,7 @@#include <rte_per_lcore.h> #include <rte_string_fns.h> #include <rte_uuid.h> +#include <rte_stdatomic.h> /** The tracepoint object. */ typedef uint64_t rte_trace_point_t ;
meson_options.txt +2 -0
@@ -46,6 +46,8 @@ option('mbuf_refcnt_atomic', type: 'boolean', value: true, description: 'Atomically access the mbuf refcnt.') option('platform', type: 'string', value: 'native', description: 'Platform to build, either "native", "generic" or a SoC. Please refer to the Linux build guide for more information.') +option('enable_stdatomic', type: 'boolean', value: false, description: + 'enable use of C11 stdatomic') option('enable_trace_fp', type: 'boolean', value: false, description: 'enable fast path trace points.') option('tests', type: 'boolean', value: true, description:
Signature · 1 line
1.8.3.1
RE: [PATCH v2 1/6] eal: provide rte stdatomics optional atomics API
Quoted 9 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Friday, 11 August 2023 19.32Provide API for atomic operations in the rte namespace that may optionally be configured to use C11 atomics with meson option enable_stdatomics=true
Signed-off-by Tyler Retzlaff <[email protected]>---
PATCH v2 2/6 eal: adapt EAL to present rte optional atomics API
In reply to Tyler RetzlaffAdapt the EAL public headers to use rte optional atomics API instead of directly using and exposing toolchain specific atomic builtin intrinsics.
--- app/test/test_mcslock.c | 6 ++-- lib/eal/arm/include/rte_atomic_32.h | 4 +-- lib/eal/arm/include/rte_atomic_64.h | 36 +++++++++++------------ lib/eal/arm/include/rte_pause_64.h | 26 ++++++++-------- lib/eal/arm/rte_power_intrinsics.c | 8 ++--- lib/eal/common/eal_common_trace.c | 16 +++++----- lib/eal/include/generic/rte_atomic.h | 50 +++++++++++++++---------------- lib/eal/include/generic/rte_pause.h | 38 ++++++++++++------------ lib/eal/include/generic/rte_rwlock.h | 47 +++++++++++++++-------------- lib/eal/include/generic/rte_spinlock.h | 19 ++++++------ lib/eal/include/rte_mcslock.h | 50 +++++++++++++++---------------- lib/eal/include/rte_pflock.h | 24 ++++++++------- lib/eal/include/rte_seqcount.h | 18 ++++++------ lib/eal/include/rte_ticketlock.h | 42 +++++++++++++------------- lib/eal/include/rte_trace_point.h | 4 +-- lib/eal/loongarch/include/rte_atomic.h | 4 +-- lib/eal/ppc/include/rte_atomic.h | 54 +++++++++++++++++----------------- lib/eal/riscv/include/rte_atomic.h | 4 +-- lib/eal/x86/include/rte_atomic.h | 8 ++--- lib/eal/x86/include/rte_spinlock.h | 2 +- lib/eal/x86/rte_power_intrinsics.c | 6 ++-- 21 files changed, 237 insertions(+), 229 deletions(-)
app/test/test_mcslock.c +3 -3
@@ -36,9 +36,9 @@* lock multiple times. */ -rte_mcslock_t *p_ml; -rte_mcslock_t *p_ml_try; -rte_mcslock_t *p_ml_perf; +rte_mcslock_t * __rte_atomic p_ml; +rte_mcslock_t * __rte_atomic p_ml_try; +rte_mcslock_t * __rte_atomic p_ml_perf;static unsigned int count;
lib/eal/arm/include/rte_atomic_32.h +2 -2
@@ -34,9 +34,9 @@#define rte_io_rmb () rte_rmb() static __rte_always_inline void -rte_atomic_thread_fence (int memorder) +rte_atomic_thread_fence (rte_memory_order memorder) { -__atomic_thread_fence (memorder); +__rte_atomic_thread_fence (memorder); }#ifdef __cplusplus
lib/eal/arm/include/rte_atomic_64.h +18 -18
@@ -38,9 +38,9 @@#define rte_io_rmb () rte_rmb() static __rte_always_inline void -rte_atomic_thread_fence (int memorder) +rte_atomic_thread_fence (rte_memory_order memorder) { -__atomic_thread_fence (memorder); +__rte_atomic_thread_fence (memorder); }/*------------------------ 128 bit atomic operations -------------------------*/
@@ -107,33 +107,33 @@*/ RTE_SET_USED (failure);/* Find invalid memory order */ -RTE_ASSERT (success == __ATOMIC_RELAXED || - success == __ATOMIC_ACQUIRE || - success == __ATOMIC_RELEASE || - success == __ATOMIC_ACQ_REL || - success == __ATOMIC_SEQ_CST); +RTE_ASSERT (success == rte_memory_order_relaxed || + success == rte_memory_order_acquire || + success == rte_memory_order_release || + success == rte_memory_order_acq_rel || + success == rte_memory_order_seq_cst);rte_int128_t expected = *exp;rte_int128_t desired = *src;rte_int128_t old;#if defined(__ARM_FEATURE_ATOMICS) || defined(RTE_ARM_FEATURE_ATOMICS ) -if (success == __ATOMIC_RELAXED) +if (success == rte_memory_order_relaxed)__cas_128_relaxed (dst, exp, desired); -else if (success == __ATOMIC_ACQUIRE) +else if (success == rte_memory_order_acquire)__cas_128_acquire (dst, exp, desired); -else if (success == __ATOMIC_RELEASE) +else if (success == rte_memory_order_release)__cas_128_release (dst, exp, desired);else __cas_128_acq_rel (dst, exp, desired); old = *exp;#else -#define __HAS_ACQ (mo) ((mo) != __ATOMIC_RELAXED && (mo) != __ATOMIC_RELEASE) -#define __HAS_RLS (mo) ((mo) == __ATOMIC_RELEASE || (mo) == __ATOMIC_ACQ_REL || \ - (mo) == __ATOMIC_SEQ_CST) +#define __HAS_ACQ (mo) ((mo) != rte_memory_order_relaxed && (mo) != rte_memory_order_release) +#define __HAS_RLS (mo) ((mo) == rte_memory_order_release || (mo) == rte_memory_order_acq_rel || \ + (mo) == rte_memory_order_seq_cst) -int ldx_mo =__HAS_ACQ (success) ? __ATOMIC_ACQUIRE : __ATOMIC_RELAXED; -int stx_mo =__HAS_RLS (success) ? __ATOMIC_RELEASE : __ATOMIC_RELAXED; +int ldx_mo =__HAS_ACQ (success) ? rte_memory_order_acquire : rte_memory_order_relaxed; +int stx_mo =__HAS_RLS (success) ? rte_memory_order_release : rte_memory_order_relaxed;#undef __HAS_ACQ#undef __HAS_RLS
@@ -153,7 +153,7 @@ : "Q" (src->val[0]) \ : "memory"); } -if (ldx_mo == __ATOMIC_RELAXED) +if (ldx_mo == rte_memory_order_relaxed) __LOAD_128("ldxp", dst, old) else__LOAD_128 ("ldaxp" , dst, old)
@@ -170,7 +170,7 @@ : "memory"); }if (likely (old.int128 == expected.int128)) { -if (stx_mo == __ATOMIC_RELAXED) +if (stx_mo == rte_memory_order_relaxed) __STORE_128("stxp", dst, desired, ret) else__STORE_128 ("stlxp" , dst, desired, ret)
@@ -181,7 +181,7 @@* needs to be stored back to ensure it was read * atomically. */ -if (stx_mo == __ATOMIC_RELAXED) +if (stx_mo == rte_memory_order_relaxed) __STORE_128("stxp", dst, old, ret) else__STORE_128 ("stlxp" , dst, old, ret)
lib/eal/arm/include/rte_pause_64.h +13 -13
@@ -41,7 +41,7 @@ static inline void rte_pause(void)* implicitly to exit WFE. */ #define __RTE_ARM_LOAD_EXC_8 (src, dst, memorder) { \ - if (memorder == __ATOMIC_RELAXED) { \ + if (memorder == rte_memory_order_relaxed) { \ asm volatile("ldxrb %w[tmp], [%x[addr]]" \ : [tmp] "=&r" (dst) \ : [addr] "r" (src) \
@@ -60,7 +60,7 @@ static inline void rte_pause(void)* implicitly to exit WFE. */ #define __RTE_ARM_LOAD_EXC_16 (src, dst, memorder) { \ - if (memorder == __ATOMIC_RELAXED) { \ + if (memorder == rte_memory_order_relaxed) { \ asm volatile("ldxrh %w[tmp], [%x[addr]]" \ : [tmp] "=&r" (dst) \ : [addr] "r" (src) \
@@ -79,7 +79,7 @@ static inline void rte_pause(void)* implicitly to exit WFE. */ #define __RTE_ARM_LOAD_EXC_32 (src, dst, memorder) { \ - if (memorder == __ATOMIC_RELAXED) { \ + if (memorder == rte_memory_order_relaxed) { \ asm volatile("ldxr %w[tmp], [%x[addr]]" \ : [tmp] "=&r" (dst) \ : [addr] "r" (src) \
@@ -98,7 +98,7 @@ static inline void rte_pause(void)* implicitly to exit WFE. */ #define __RTE_ARM_LOAD_EXC_64 (src, dst, memorder) { \ - if (memorder == __ATOMIC_RELAXED) { \ + if (memorder == rte_memory_order_relaxed) { \ asm volatile("ldxr %x[tmp], [%x[addr]]" \ : [tmp] "=&r" (dst) \ : [addr] "r" (src) \
@@ -118,7 +118,7 @@ static inline void rte_pause(void)*/ #define __RTE_ARM_LOAD_EXC_128 (src, dst, memorder) { \ volatile rte_int128_t *dst_128 = (volatile rte_int128_t *)&dst; \ - if (memorder == __ATOMIC_RELAXED) { \ + if (memorder == rte_memory_order_relaxed) { \ asm volatile("ldxp %x[tmp0], %x[tmp1], [%x[addr]]" \ : [tmp0] "=&r" (dst_128->val[0]), \ [tmp1] "=&r" (dst_128->val[1]) \
@@ -153,8 +153,8 @@ static inline void rte_pause(void) {uint16_t value; -RTE_BUILD_BUG_ON (memorder != __ATOMIC_ACQUIRE && - memorder != __ATOMIC_RELAXED); +RTE_BUILD_BUG_ON (memorder != rte_memory_order_acquire && + memorder != rte_memory_order_relaxed); __RTE_ARM_LOAD_EXC_16(addr, value, memorder)if (value != expected) {
@@ -172,8 +172,8 @@ static inline void rte_pause(void) {uint32_t value; -RTE_BUILD_BUG_ON (memorder != __ATOMIC_ACQUIRE && - memorder != __ATOMIC_RELAXED); +RTE_BUILD_BUG_ON (memorder != rte_memory_order_acquire && + memorder != rte_memory_order_relaxed); __RTE_ARM_LOAD_EXC_32(addr, value, memorder)if (value != expected) {
@@ -191,8 +191,8 @@ static inline void rte_pause(void) {uint64_t value; -RTE_BUILD_BUG_ON (memorder != __ATOMIC_ACQUIRE && - memorder != __ATOMIC_RELAXED); +RTE_BUILD_BUG_ON (memorder != rte_memory_order_acquire && + memorder != rte_memory_order_relaxed); __RTE_ARM_LOAD_EXC_64(addr, value, memorder)if (value != expected) {
@@ -206,8 +206,8 @@ static inline void rte_pause(void)#define RTE_WAIT_UNTIL_MASKED (addr, mask, cond, expected, memorder) do { \ RTE_BUILD_BUG_ON(!__builtin_constant_p(memorder)); \ - RTE_BUILD_BUG_ON(memorder != __ATOMIC_ACQUIRE && \ - memorder != __ATOMIC_RELAXED); \ + RTE_BUILD_BUG_ON(memorder != rte_memory_order_acquire && \ + memorder != rte_memory_order_relaxed); \ const uint32_t size = sizeof(*(addr)) << 3; \ typeof(*(addr)) expected_value = (expected); \ typeof(*(addr)) value; \
lib/eal/arm/rte_power_intrinsics.c +4 -4
@@ -33,19 +33,19 @@switch (pmc->size) {case sizeof (uint8_t ): - __RTE_ARM_LOAD_EXC_8(pmc->addr, cur_value, __ATOMIC_RELAXED) + __RTE_ARM_LOAD_EXC_8(pmc->addr, cur_value, rte_memory_order_relaxed)__RTE_ARM_WFE () break;case sizeof (uint16_t ): - __RTE_ARM_LOAD_EXC_16(pmc->addr, cur_value, __ATOMIC_RELAXED) + __RTE_ARM_LOAD_EXC_16(pmc->addr, cur_value, rte_memory_order_relaxed)__RTE_ARM_WFE () break;case sizeof (uint32_t ): - __RTE_ARM_LOAD_EXC_32(pmc->addr, cur_value, __ATOMIC_RELAXED) + __RTE_ARM_LOAD_EXC_32(pmc->addr, cur_value, rte_memory_order_relaxed)__RTE_ARM_WFE () break;case sizeof (uint64_t ): - __RTE_ARM_LOAD_EXC_64(pmc->addr, cur_value, __ATOMIC_RELAXED) + __RTE_ARM_LOAD_EXC_64(pmc->addr, cur_value, rte_memory_order_relaxed)__RTE_ARM_WFE () break;default :
lib/eal/common/eal_common_trace.c +9 -7
@@ -103,11 +103,11 @@ struct trace_point_head * trace_mode_set(rte_trace_point_t *t, enum rte_trace_mode mode) {if (mode ==RTE_TRACE_MODE_OVERWRITE ) -__atomic_fetch_and (t, ~__RTE_TRACE_FIELD_ENABLE_DISCARD, - __ATOMIC_RELEASE); +rte_atomic_fetch_and_explicit (t, ~__RTE_TRACE_FIELD_ENABLE_DISCARD, + rte_memory_order_release);else -__atomic_fetch_or (t, __RTE_TRACE_FIELD_ENABLE_DISCARD, - __ATOMIC_RELEASE); +rte_atomic_fetch_or_explicit (t, __RTE_TRACE_FIELD_ENABLE_DISCARD, + rte_memory_order_release); }void
@@ -141,7 +141,7 @@ rte_trace_mode rte_trace_mode_get(void)if (trace_point_is_invalid (t))return false; - val =__atomic_load_n (t, __ATOMIC_ACQUIRE); + val =rte_atomic_load_explicit (t, rte_memory_order_acquire);return (val & __RTE_TRACE_FIELD_ENABLE_MASK) !=0 ; }
@@ -153,7 +153,8 @@ rte_trace_mode rte_trace_mode_get(void)if (trace_point_is_invalid (t))return -ERANGE ; - prev =__atomic_fetch_or (t, __RTE_TRACE_FIELD_ENABLE_MASK, __ATOMIC_RELEASE); + prev =rte_atomic_fetch_or_explicit (t, __RTE_TRACE_FIELD_ENABLE_MASK, + rte_memory_order_release);if ((prev & __RTE_TRACE_FIELD_ENABLE_MASK) ==0 )__atomic_fetch_add (&trace.status,1 , __ATOMIC_RELEASE);return 0 ;
@@ -167,7 +168,8 @@ rte_trace_mode rte_trace_mode_get(void)if (trace_point_is_invalid (t))return -ERANGE ; - prev =__atomic_fetch_and (t, ~__RTE_TRACE_FIELD_ENABLE_MASK, __ATOMIC_RELEASE); + prev =rte_atomic_fetch_and_explicit (t, ~__RTE_TRACE_FIELD_ENABLE_MASK, + rte_memory_order_release);if ((prev & __RTE_TRACE_FIELD_ENABLE_MASK) !=0 )__atomic_fetch_sub (&trace.status,1 , __ATOMIC_RELEASE);return 0 ;
lib/eal/include/generic/rte_atomic.h +25 -25
194 lines not on this page · raw message
lib/eal/include/generic/rte_pause.h +19 -19
@@ -36,13 +36,13 @@* A 16-bit expected value to be in the memory location. * @param memorder * Two different memory orders that can be specified: -* __ATOMIC_ACQUIRE and __ATOMIC_RELAXED. These map to +* rte_memory_order_acquire and rte_memory_order_relaxed. These map to * C++11 memory orders with the same names, see the C++11 standard or * the GCC wiki on atomic synchronization for detailed definition. */ static __rte_always_inline voidrte_wait_until_equal_16 (volatile uint16_t *addr,uint16_t expected, -int memorder); +rte_memory_order memorder);/** * Wait for *addr to be updated with a 32-bit expected value, with a relaxed
@@ -54,13 +54,13 @@* A 32-bit expected value to be in the memory location. * @param memorder * Two different memory orders that can be specified: -* __ATOMIC_ACQUIRE and __ATOMIC_RELAXED. These map to +* rte_memory_order_acquire and rte_memory_order_relaxed. These map to * C++11 memory orders with the same names, see the C++11 standard or * the GCC wiki on atomic synchronization for detailed definition. */ static __rte_always_inline voidrte_wait_until_equal_32 (volatile uint32_t *addr,uint32_t expected, -int memorder); +rte_memory_order memorder);/** * Wait for *addr to be updated with a 64-bit expected value, with a relaxed
@@ -72,42 +72,42 @@* A 64-bit expected value to be in the memory location. * @param memorder * Two different memory orders that can be specified: -* __ATOMIC_ACQUIRE and __ATOMIC_RELAXED. These map to +* rte_memory_order_acquire and rte_memory_order_relaxed. These map to * C++11 memory orders with the same names, see the C++11 standard or * the GCC wiki on atomic synchronization for detailed definition. */ static __rte_always_inline voidrte_wait_until_equal_64 (volatile uint64_t *addr,uint64_t expected, -int memorder); +rte_memory_order memorder);#ifndef RTE_WAIT_UNTIL_EQUAL_ARCH_DEFINED static __rte_always_inline voidrte_wait_until_equal_16 (volatile uint16_t *addr,uint16_t expected, -int memorder) +rte_memory_order memorder) { -assert (memorder == __ATOMIC_ACQUIRE || memorder == __ATOMIC_RELAXED); +assert (memorder == rte_memory_order_acquire || memorder == rte_memory_order_relaxed); -while (__atomic_load_n (addr, memorder) != expected) +while (rte_atomic_load_explicit (addr, memorder) != expected)rte_pause (); } static __rte_always_inline voidrte_wait_until_equal_32 (volatile uint32_t *addr,uint32_t expected, -int memorder) +rte_memory_order memorder) { -assert (memorder == __ATOMIC_ACQUIRE || memorder == __ATOMIC_RELAXED); +assert (memorder == rte_memory_order_acquire || memorder == rte_memory_order_relaxed); -while (__atomic_load_n (addr, memorder) != expected) +while (rte_atomic_load_explicit (addr, memorder) != expected)rte_pause (); } static __rte_always_inline voidrte_wait_until_equal_64 (volatile uint64_t *addr,uint64_t expected, -int memorder) +rte_memory_order memorder) { -assert (memorder == __ATOMIC_ACQUIRE || memorder == __ATOMIC_RELAXED); +assert (memorder == rte_memory_order_acquire || memorder == rte_memory_order_relaxed); -while (__atomic_load_n (addr, memorder) != expected) +while (rte_atomic_load_explicit (addr, memorder) != expected)rte_pause (); }
@@ -125,16 +125,16 @@* An expected value to be in the memory location. * @param memorder * Two different memory orders that can be specified: -* __ATOMIC_ACQUIRE and __ATOMIC_RELAXED. These map to +* rte_memory_order_acquire and rte_memory_order_relaxed. These map to * C++11 memory orders with the same names, see the C++11 standard or * the GCC wiki on atomic synchronization for detailed definition. */ #define RTE_WAIT_UNTIL_MASKED (addr, mask, cond, expected, memorder) do { \ RTE_BUILD_BUG_ON(!__builtin_constant_p(memorder)); \ - RTE_BUILD_BUG_ON(memorder != __ATOMIC_ACQUIRE && \ - memorder != __ATOMIC_RELAXED); \ + RTE_BUILD_BUG_ON((memorder) != rte_memory_order_acquire && \ + (memorder) != rte_memory_order_relaxed); \ typeof(*(addr)) expected_value = (expected); \ - while (!((__atomic_load_n((addr), (memorder)) & (mask)) cond \ + while (!((rte_atomic_load_explicit((addr), (memorder)) & (mask)) cond \ expected_value)) \ rte_pause(); \ } while (0)
lib/eal/include/generic/rte_rwlock.h +25 -22
126 lines not on this page · raw message
lib/eal/include/generic/rte_spinlock.h +9 -10
@@ -29,7 +29,7 @@* The rte_spinlock_t type. */ typedef struct __rte_lockable { -volatile int locked;/**< lock status 0 = unlocked, 1 = locked */ + volatile int __rte_atomic locked; /**< lock status 0 = unlocked, 1 = locked */ }rte_spinlock_t ;/**
@@ -66,10 +66,10 @@ {int exp =0 ; -while (!__atomic_compare_exchange_n (&sl->locked, &exp,1 ,0 , - __ATOMIC_ACQUIRE, __ATOMIC_RELAXED)) { -rte_wait_until_equal_32 ((volatile uint32_t *)&sl->locked, -0 , __ATOMIC_RELAXED); +while (!rte_atomic_compare_exchange_strong_explicit (&sl->locked, &exp,1 , + rte_memory_order_acquire, rte_memory_order_relaxed)) { +rte_wait_until_equal_32 ((volatile uint32_t *)(uintptr_t )&sl->locked, +0 , rte_memory_order_relaxed); exp =0 ; } }
@@ -90,7 +90,7 @@ rte_spinlock_unlock(rte_spinlock_t *sl) __rte_no_thread_safety_analysis { -__atomic_store_n (&sl->locked,0 , __ATOMIC_RELEASE); +rte_atomic_store_explicit (&sl->locked,0 , rte_memory_order_release); }#endif
@@ -113,9 +113,8 @@ __rte_no_thread_safety_analysis {int exp =0 ; -return __atomic_compare_exchange_n (&sl->locked, &exp,1 , -0 ,/* disallow spurious failure */ - __ATOMIC_ACQUIRE, __ATOMIC_RELAXED); +return rte_atomic_compare_exchange_strong_explicit (&sl->locked, &exp,1 , + rte_memory_order_acquire, rte_memory_order_relaxed); }#endif
@@ -129,7 +128,7 @@*/ static inline int rte_spinlock_is_locked (rte_spinlock_t *sl) { -return __atomic_load_n (&sl->locked, __ATOMIC_ACQUIRE); +return rte_atomic_load_explicit (&sl->locked, rte_memory_order_acquire); }/**
lib/eal/include/rte_mcslock.h +25 -25
133 lines not on this page · raw message
lib/eal/include/rte_pflock.h +13 -11
@@ -41,8 +41,8 @@*/ struct rte_pflock {struct { -uint16_t in; -uint16_t out; + uint16_t __rte_atomic in; + uint16_t __rte_atomic out; } rd, wr; };typedef struct rte_pflock rte_pflock_t ;
@@ -117,14 +117,14 @@ struct rte_pflock {* If no writer is present, then the operation has completed * successfully. */ - w =__atomic_fetch_add (&pf->rd.in,RTE_PFLOCK_RINC , __ATOMIC_ACQUIRE) + w =rte_atomic_fetch_add_explicit (&pf->rd.in,RTE_PFLOCK_RINC , rte_memory_order_acquire) &RTE_PFLOCK_WBITS ;if (w ==0 )return ;/* Wait for current write phase to complete. */ RTE_WAIT_UNTIL_MASKED(&pf->rd.in, RTE_PFLOCK_WBITS, !=, w, - __ATOMIC_ACQUIRE); + rte_memory_order_acquire); }/**
@@ -140,7 +140,7 @@ struct rte_pflock {static inline void rte_pflock_read_unlock (rte_pflock_t *pf) { -__atomic_fetch_add (&pf->rd.out,RTE_PFLOCK_RINC , __ATOMIC_RELEASE); +rte_atomic_fetch_add_explicit (&pf->rd.out,RTE_PFLOCK_RINC , rte_memory_order_release); }/**
@@ -161,8 +161,9 @@ struct rte_pflock {/* Acquire ownership of write-phase. * This is same as rte_ticketlock_lock(). */ - ticket =__atomic_fetch_add (&pf->wr.in,1 , __ATOMIC_RELAXED); -rte_wait_until_equal_16 (&pf->wr.out, ticket, __ATOMIC_ACQUIRE); + ticket =rte_atomic_fetch_add_explicit (&pf->wr.in,1 , rte_memory_order_relaxed); +rte_wait_until_equal_16 ((uint16_t *)(uintptr_t )&pf->wr.out, ticket, + rte_memory_order_acquire);/* * Acquire ticket on read-side in order to allow them
@@ -173,10 +174,11 @@ struct rte_pflock {* speculatively. */ w =RTE_PFLOCK_PRES | (ticket &RTE_PFLOCK_PHID ); - ticket =__atomic_fetch_add (&pf->rd.in, w, __ATOMIC_RELAXED); + ticket =rte_atomic_fetch_add_explicit (&pf->rd.in, w, rte_memory_order_relaxed);/* Wait for any pending readers to flush. */ -rte_wait_until_equal_16 (&pf->rd.out, ticket, __ATOMIC_ACQUIRE); +rte_wait_until_equal_16 ((uint16_t *)(uintptr_t )&pf->rd.out, ticket, + rte_memory_order_acquire); }/**
@@ -193,10 +195,10 @@ struct rte_pflock { rte_pflock_write_unlock(rte_pflock_t *pf) {/* Migrate from write phase to read phase. */ -__atomic_fetch_and (&pf->rd.in,RTE_PFLOCK_LSB , __ATOMIC_RELEASE); +rte_atomic_fetch_and_explicit (&pf->rd.in,RTE_PFLOCK_LSB , rte_memory_order_release);/* Allow other writers to continue. */ -__atomic_fetch_add (&pf->wr.out,1 , __ATOMIC_RELEASE); +rte_atomic_fetch_add_explicit (&pf->wr.out,1 , rte_memory_order_release); }#ifdef __cplusplus
lib/eal/include/rte_seqcount.h +9 -9
@@ -32,7 +32,7 @@* The RTE seqcount type. */ typedef struct { -uint32_t sn;/**< A sequence number for the protected data. */ + uint32_t __rte_atomic sn; /**< A sequence number for the protected data. */ }rte_seqcount_t ;/**
@@ -106,11 +106,11 @@static inline uint32_t rte_seqcount_read_begin (const rte_seqcount_t *seqcount) { -/* __ATOMIC_ACQUIRE to prevent loads after (in program order) +/* rte_memory_order_acquire to prevent loads after (in program order) * from happening before the sn load. Synchronizes-with the * store release in rte_seqcount_write_end(). */ -return __atomic_load_n (&seqcount->sn, __ATOMIC_ACQUIRE); +return rte_atomic_load_explicit (&seqcount->sn, rte_memory_order_acquire); }/**
@@ -161,9 +161,9 @@return true;/* make sure the data loads happens before the sn load */ -rte_atomic_thread_fence (__ATOMIC_ACQUIRE); +rte_atomic_thread_fence (rte_memory_order_acquire); - end_sn =__atomic_load_n (&seqcount->sn, __ATOMIC_RELAXED); + end_sn =rte_atomic_load_explicit (&seqcount->sn, rte_memory_order_relaxed);/* A writer incremented the sequence number during this read * critical section.
@@ -205,12 +205,12 @@ sn = seqcount->sn +1 ; -__atomic_store_n (&seqcount->sn, sn, __ATOMIC_RELAXED); +rte_atomic_store_explicit (&seqcount->sn, sn, rte_memory_order_relaxed); -/* __ATOMIC_RELEASE to prevent stores after (in program order) +/* rte_memory_order_release to prevent stores after (in program order) * from happening before the sn store. */ -rte_atomic_thread_fence (__ATOMIC_RELEASE); +rte_atomic_thread_fence (rte_memory_order_release); }/**
@@ -237,7 +237,7 @@ sn = seqcount->sn +1 ;/* Synchronizes-with the load acquire in rte_seqcount_read_begin(). */ -__atomic_store_n (&seqcount->sn, sn, __ATOMIC_RELEASE); +rte_atomic_store_explicit (&seqcount->sn, sn, rte_memory_order_release); }#ifdef __cplusplus
lib/eal/include/rte_ticketlock.h +22 -20
115 lines not on this page · raw message
lib/eal/include/rte_trace_point.h +2 -2
@@ -33,7 +33,7 @@#include <rte_stdatomic.h> /** The tracepoint object. */ -typedef uint64_t rte_trace_point_t ; +typedef uint64_t __rte_atomic rte_trace_point_t;/** * Macro to define the tracepoint arguments in RTE_TRACE_POINT macro.
@@ -359,7 +359,7 @@ struct __rte_trace_header {#define __rte_trace_point_emit_header_generic (t) \ void *mem; \ do { \ - const uint64_t val = __atomic_load_n(t, __ATOMIC_ACQUIRE); \ + const uint64_t val = rte_atomic_load_explicit(t, rte_memory_order_acquire); \ if (likely(!(val & __RTE_TRACE_FIELD_ENABLE_MASK))) \ return; \ mem = __rte_trace_mem_get(val); \
lib/eal/loongarch/include/rte_atomic.h +2 -2
@@ -35,9 +35,9 @@#define rte_io_rmb () rte_mb() static __rte_always_inline void -rte_atomic_thread_fence (int memorder) +rte_atomic_thread_fence (rte_memory_order memorder) { -__atomic_thread_fence (memorder); +__rte_atomic_thread_fence (memorder); }#ifdef __cplusplus
lib/eal/ppc/include/rte_atomic.h +27 -27
172 lines not on this page · raw message
lib/eal/riscv/include/rte_atomic.h +2 -2
@@ -40,9 +40,9 @@#define rte_io_rmb () asm volatile("fence ir, ir" : : : "memory") static __rte_always_inline void -rte_atomic_thread_fence (int memorder) +rte_atomic_thread_fence (rte_memory_order memorder) { -__atomic_thread_fence (memorder); +__rte_atomic_thread_fence (memorder); }#ifdef __cplusplus
lib/eal/x86/include/rte_atomic.h +4 -4
@@ -82,17 +82,17 @@/** * Synchronization fence between threads based on the specified memory order. * -* On x86 the __atomic_thread_fence(__ATOMIC_SEQ_CST) generates full 'mfence' +* On x86 the __rte_atomic_thread_fence(rte_memory_order_seq_cst) generates full 'mfence' * which is quite expensive. The optimized implementation of rte_smp_mb is * used instead. */ static __rte_always_inline void -rte_atomic_thread_fence (int memorder) +rte_atomic_thread_fence (rte_memory_order memorder) { -if (memorder == __ATOMIC_SEQ_CST) +if (memorder == rte_memory_order_seq_cst)rte_smp_mb ();else -__atomic_thread_fence (memorder); +__rte_atomic_thread_fence (memorder); }/*------------------------- 16 bit atomic operations -------------------------*/
lib/eal/x86/include/rte_spinlock.h +1 -1
@@ -78,7 +78,7 @@ static inline int rte_tm_supported(void) }static inline int -rte_try_tm (volatile int *lock) +rte_try_tm(volatile int __rte_atomic *lock) {int i, retries;
lib/eal/x86/rte_power_intrinsics.c +3 -3
@@ -23,9 +23,9 @@uint64_t val;/* trigger a write but don't change the value */ - val =__atomic_load_n ((volatile uint64_t *)addr, __ATOMIC_RELAXED); -__atomic_compare_exchange_n ((volatile uint64_t *)addr, &val, val,0 , - __ATOMIC_RELAXED, __ATOMIC_RELAXED); + val =rte_atomic_load_explicit ((volatile uint64_t *)addr, rte_memory_order_relaxed); +rte_atomic_compare_exchange_strong_explicit ((volatile uint64_t *)addr, &val, val, + rte_memory_order_relaxed, rte_memory_order_relaxed); }static bool wait_supported;
Signature · 1 line
1.8.3.1
RE: [PATCH v2 2/6] eal: adapt EAL to present rte optional atomics API
Quoted 8 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Friday, 11 August 2023 19.32Adapt the EAL public headers to use rte optional atomics API instead of directly using and exposing toolchain specific atomic builtin intrinsics.
Signed-off-by Tyler Retzlaff <[email protected]>---
[...]
app/test/test_mcslock.c line 41
Quoted 12 lines
app/test/test_mcslock.c +3
-3@@ -36,9 +36,9 @@* lock multiple times. */ -rte_mcslock_t *p_ml; -rte_mcslock_t *p_ml_try; -rte_mcslock_t *p_ml_perf; +rte_mcslock_t * __rte_atomic p_ml; +rte_mcslock_t * __rte_atomic p_ml_try; +rte_mcslock_t * __rte_atomic p_ml_perf;
Although this looks weird, it is pointers themselves, not the structures, that are used atomically. So it is correct.
lib/eal/include/generic/rte_pause.h line 41
Quoted 13 lines
lib/eal/include/generic/rte_pause.h +1
-1@@ -36,13 +36,13 @@* A 16-bit expected value to be in the memory location. * @param memorder * Two different memory orders that can be specified: -* __ATOMIC_ACQUIRE and __ATOMIC_RELAXED. These map to +* rte_memory_order_acquire and rte_memory_order_relaxed. These map to * C++11 memory orders with the same names, see the C++11 standard or * the GCC wiki on atomic synchronization for detailed definition.
Delete the last part of the description, starting at "These map to...".
Quoted 9 lines
lib/eal/include/generic/rte_pause.h +1
-1*/ static __rte_always_inline voidrte_wait_until_equal_16 (volatile uint16_t *addr,uint16_t expected, -int memorder); +rte_memory_order memorder);/** * Wait for *addr to be updated with a 32-bit expected value, with a relaxed
lib/eal/include/generic/rte_pause.h line 59
Quoted 7 lines
lib/eal/include/generic/rte_pause.h +1
-1@@ -54,13 +54,13 @@* A 32-bit expected value to be in the memory location. * @param memorder * Two different memory orders that can be specified: -* __ATOMIC_ACQUIRE and __ATOMIC_RELAXED. These map to +* rte_memory_order_acquire and rte_memory_order_relaxed. These map to * C++11 memory orders with the same names, see the C++11 standard or * the GCC wiki on atomic synchronization for detailed definition.
Delete the last part of the description, starting at "These map to...".
Quoted 9 lines
lib/eal/include/generic/rte_pause.h +1
-1*/ static __rte_always_inline voidrte_wait_until_equal_32 (volatile uint32_t *addr,uint32_t expected, -int memorder); +rte_memory_order memorder);/** * Wait for *addr to be updated with a 64-bit expected value, with a relaxed
lib/eal/include/generic/rte_pause.h line 77
Quoted 7 lines
lib/eal/include/generic/rte_pause.h +1
-1@@ -72,42 +72,42 @@* A 64-bit expected value to be in the memory location. * @param memorder * Two different memory orders that can be specified: -* __ATOMIC_ACQUIRE and __ATOMIC_RELAXED. These map to +* rte_memory_order_acquire and rte_memory_order_relaxed. These map to * C++11 memory orders with the same names, see the C++11 standard or * the GCC wiki on atomic synchronization for detailed definition.
Delete the last part of the description, starting at "These map to...".
Quoted 5 lines
lib/eal/include/generic/rte_pause.h +1
-1*/ static __rte_always_inline voidrte_wait_until_equal_64 (volatile uint64_t *addr,uint64_t expected, -int memorder); +rte_memory_order memorder);
[...]
lib/eal/include/generic/rte_pause.h line 130
Quoted 8 lines
lib/eal/include/generic/rte_pause.h +1
-1@@ -125,16 +125,16 @@* An expected value to be in the memory location. * @param memorder * Two different memory orders that can be specified: -* __ATOMIC_ACQUIRE and __ATOMIC_RELAXED. These map to +* rte_memory_order_acquire and rte_memory_order_relaxed. These map to * C++11 memory orders with the same names, see the C++11 standard or * the GCC wiki on atomic synchronization for detailed definition.
Delete the last part of the description, starting at "These map to...".
There might be more similar comments that need removal; I haven't tried searching.
Quoted 2 lines
lib/eal/include/generic/rte_pause.h +0
-0*/ #define RTE_WAIT_UNTIL_MASKED (addr, mask, cond, expected, memorder) do { \
[...]
Quoted 9 lines
lib/eal/include/generic/rte_spinlock.h +1
-1@@ -29,7 +29,7 @@* The rte_spinlock_t type. */ typedef struct __rte_lockable { - volatile int locked; /**< lock status 0 = unlocked, 1 = locked */ + volatile int __rte_atomic locked; /**< lock status 0 = unlocked, 1 =locked */
I think __rte_atomic should be before the type:
volatile __rte_atomic int locked; /**< lock status [...]
Alternatively (just mentioning it, I know we don't use this form):
volatile __rte_atomic(int) locked; /**< lock status [...]
Thinking of where you would put "const" might help.
Maybe your order is also correct, so it is a matter of preference.
The DPDK coding style guidelines doesn't mention where to place "const", but looking at the code, it seems to use "const unsigned int" and "const char *".
Quoted 3 lines
}
rte_spinlock_t;/**
[...]
lib/eal/include/rte_mcslock.h line 36
Quoted 9 lines
lib/eal/include/rte_mcslock.h +1
-2@@ -33,8 +33,8 @@* The rte_mcslock_t type. */ typedef struct rte_mcslock { - struct rte_mcslock *next; - int locked; /* 1 if the queue locked, 0 otherwise */ + struct rte_mcslock * __rte_atomic next;
Correct, the pointer is atomic, not the struct.
lib/eal/include/rte_mcslock.h line 37
Quoted 1 line
lib/eal/include/rte_mcslock.h +1
-0+ int __rte_atomic locked; /* 1 if the queue locked, 0 otherwise */
Again, I think __rte_atomic should be before the type:
__rte_atomic int locked; /* 1 if the queue locked, 0 otherwise */
Quoted 2 lines
lib/eal/include/rte_mcslock.h +0
-0} rte_mcslock_t;
[...]
Quoted 29 lines
lib/eal/include/rte_mcslock.h +5
-5@@ -101,34 +101,34 @@* A pointer to the node of MCS lock passed in rte_mcslock_lock. */ static inline void -rte_mcslock_unlock (rte_mcslock_t **msl,rte_mcslock_t *me) +rte_mcslock_unlock(rte_mcslock_t * __rte_atomic *msl, rte_mcslock_t * __rte_atomic me) {/* Check if there are more nodes in the queue. */ -if (likely (__atomic_load_n (&me->next, __ATOMIC_RELAXED) ==NULL )) { +if (likely (rte_atomic_load_explicit (&me->next, rte_memory_order_relaxed) ==NULL )) {/* No, last member in the queue. */ -rte_mcslock_t *save_me =__atomic_load_n (&me, __ATOMIC_RELAXED); +rte_mcslock_t *save_me =rte_atomic_load_explicit (&me, rte_memory_order_relaxed);/* Release the lock by setting it to NULL */ -if (likely (__atomic_compare_exchange_n (msl, &save_me,NULL ,0 , - __ATOMIC_RELEASE, __ATOMIC_RELAXED))) +if (likely (rte_atomic_compare_exchange_strong_explicit (msl, &save_me,NULL , + rte_memory_order_release, rte_memory_order_relaxed)))return ;/* Speculative execution would be allowed to read in the * while-loop first. This has the potential to cause a * deadlock. Need a load barrier.
lib/eal/include/rte_mcslock.h line 130
Quoted 10 lines
lib/eal/include/rte_mcslock.h +3
-3*/ -__atomic_thread_fence (__ATOMIC_ACQUIRE); +__rte_atomic_thread_fence (rte_memory_order_acquire);/* More nodes added to the queue by other CPUs. * Wait until the next pointer is set. */ -uintptr_t *next; - next = (uintptr_t *)&me->next; + uintptr_t __rte_atomic *next; + next = (uintptr_t __rte_atomic *)&me->next;
This way around, I think:
__rte_atomic uintptr_t *next; next = (__rte_atomic uintptr_t *)&me->next;
[...]
lib/eal/include/rte_pflock.h line 45
Quoted 10 lines
lib/eal/include/rte_pflock.h +2
-2@@ -41,8 +41,8 @@*/ struct rte_pflock {struct { -uint16_t in; -uint16_t out; + uint16_t __rte_atomic in; + uint16_t __rte_atomic out;
Again, I think __rte_atomic should be before the type:
__rte_atomic uint16_t in; __rte_atomic uint16_t out;
Quoted 2 lines
lib/eal/include/rte_pflock.h +0
-0} rd, wr; };
[...]
lib/eal/include/rte_seqcount.h line 36
Quoted 9 lines
lib/eal/include/rte_seqcount.h +1
-1@@ -32,7 +32,7 @@* The RTE seqcount type. */ typedef struct { - uint32_t sn; /**< A sequence number for the protected data. */ + uint32_t __rte_atomic sn; /**< A sequence number for the protected data. */
Again, I think __rte_atomic should be before the type:
__rte_atomic uint32_t sn; /**< A sequence [...]
Quoted 1 line
lib/eal/include/rte_seqcount.h +0
-0} rte_seqcount_t;
[...]
lib/eal/include/rte_ticketlock.h line 36
Quoted 13 lines
lib/eal/include/rte_ticketlock.h +3
-3@@ -30,10 +30,10 @@* The rte_ticketlock_t type. */ typedef union { - uint32_t tickets; + uint32_t __rte_atomic tickets; struct { - uint16_t current; - uint16_t next; + uint16_t __rte_atomic current; + uint16_t __rte_atomic next;
Again, I think __rte_atomic should be before the type:
__rte_atomic uint16_t current; __rte_atomic uint16_t next;
Quoted 11 lines
lib/eal/include/rte_ticketlock.h +1
-1} s; } rte_ticketlock_t;@@ -127,7 +129,7 @@ typedef struct { rte_ticketlock_t tl; /**< the actual ticketlock */ - int user; /**< core id using lock, TICKET_LOCK_INVALID_ID for unused */ + int __rte_atomic user; /**< core id using lock, TICKET_LOCK_INVALID_ID
Quoted 1 line
for unused */
Again, I think __rte_atomic should be before the type:
__rte_atomic int user; /**< core id [...]
Quoted 2 lines
unsigned int count; /**< count of time this lock has been called */
}rte_ticketlock_recursive_t;
[...]
lib/eal/include/rte_trace_point.h line 36
Quoted 8 lines
lib/eal/include/rte_trace_point.h +1
-1@@ -33,7 +33,7 @@#include <rte_stdatomic.h> /** The tracepoint object. */ -typedef uint64_t rte_trace_point_t ; +typedef uint64_t __rte_atomic rte_trace_point_t;
Again, I think __rte_atomic should be before the type:
typedef __rte_atomic uint64_t rte_trace_point_t;
[...]
At the risk of having gone "speed blind" by all the search-replaces along the way...
Re: [PATCH v2 2/6] eal: adapt EAL to present rte optional atomics API
Quoted 11 lines · Morten Brørup
Quoted 8 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Friday, 11 August 2023 19.32Adapt the EAL public headers to use rte optional atomics API instead of directly using and exposing toolchain specific atomic builtin intrinsics.
Signed-off-by Tyler Retzlaff <[email protected]>---
[...]
will fix the comments identified.
Quoted 16 lines
[...]
Quoted 9 lines
lib/eal/include/generic/rte_spinlock.h +1
-1@@ -29,7 +29,7 @@* The rte_spinlock_t type. */ typedef struct __rte_lockable { - volatile int locked; /**< lock status 0 = unlocked, 1 = locked */ + volatile int __rte_atomic locked; /**< lock status 0 = unlocked, 1 =locked */
I think
__rte_atomicshould be before the type:
volatile__rte_atomicint locked; /**< lock status [...]
Alternatively (just mentioning it, I know we don't use this form):
Quoted 5 lines
volatile __rte_atomic(int) locked; /**< lock status [...]Thinking of where you would put "const" might help.
Maybe your order is also correct, so it is a matter of preference.
so for me what you suggest is the canonical convention for c and i did initially try to make the change with this convention but ran into trouble when using the keyword in a context used as a type specifier and the type was incomplete.
the rte_mcslock is a good example for illustration.
// original struct
typedef struct rte_mcslock {
struct rte_mcslock *next;
...
};it simply doesn't work / won't compile (at least with clang) which is what drove me to use the less-often used syntax.
typedef struct rte_mcslock {
_Atomic struct rte_mcslock *next;
...
}; In file included from ../app/test/test_mcslock.c:19:
..\lib\eal\include\rte_mcslock.h:36:2: error: _Atomic cannot be applied to incomplete type 'struct rte_mcslock'
_Atomic struct rte_mcslock *next;
^
..\lib\eal\include\rte_mcslock.h:35:16: note: definition of 'struct
rte_mcslock' is not complete until the closing '}'
typedef struct rte_mcslock {
^1 error generated.
so i ended up choosing to use a single syntax by convention consistently rather than using one for the exceptional case and one everywhere else.
i think (based on our other thread of discussion) i would recommend we use adopt and require the use of the _Atomic(T) macro to disambiguate it also has the advantage of not being churned later when we can do c++23.
// using macro
typedef struct rte_mcslock {
_Atomic(struct rte_mcslock *) next;
...
};this is much easier at a glance to know when the specified type is the T or the T * similarly in parameter lists it becomes more clear too.
e.g.
void foo(int *v)
that it is either void foo(_Atomic(int) *v) or void foo(_Atomic(int *) v) becomes much clearer without having to do mental gymnastics.
so i propose we retain
#define __rte_atomic _Atomic
allow it to be used in contexts where we need a type-qualifier. note:
most of the cases where _Atomic is used as a type-qualifier it is a red flag that we are sensitive to an implementation detail of the compiler. in time i hope most of these will go away as we remove deprecated rte_atomic_xx apis.
but also introduce the following macro
#define RTE_ATOMIC(type) _Atomic(type)
require it be used in the contexts that we are using it as a type-specifier.
if folks agree with this please reply back positively and i'll update the series. feel free to propose alternate names or whatever, but sooner than later so i don't have to churn things too much :)
thanks!
Quoted 2 lines
The DPDK coding style guidelines doesn't mention where to place "const", but looking at the code, it seems to use "const unsigned int" and "const char *".
we probably should document it as a convention and most likely we should adopt what is already in use more commonly.
Quoted 169 lines
Quoted 4 lines
}
rte_spinlock_t;/**
[...]
lib/eal/include/rte_mcslock.h line 36
Quoted 9 lines
lib/eal/include/rte_mcslock.h +1
-2@@ -33,8 +33,8 @@* The rte_mcslock_t type. */ typedef struct rte_mcslock { - struct rte_mcslock *next; - int locked; /* 1 if the queue locked, 0 otherwise */ + struct rte_mcslock * __rte_atomic next;Correct, the pointer is atomic, not the struct.
lib/eal/include/rte_mcslock.h line 37
Quoted 1 line
lib/eal/include/rte_mcslock.h +1
-0+ int __rte_atomic locked; /* 1 if the queue locked, 0 otherwise */Again, I think
__rte_atomicshould be before the type:
__rte_atomicint locked; /* 1 if the queue locked, 0 otherwise */Quoted 2 lines
lib/eal/include/rte_mcslock.h +0
-0} rte_mcslock_t;[...]
Quoted 29 lines
lib/eal/include/rte_mcslock.h +5
-5@@ -101,34 +101,34 @@* A pointer to the node of MCS lock passed in rte_mcslock_lock. */ static inline void -rte_mcslock_unlock (rte_mcslock_t **msl,rte_mcslock_t *me) +rte_mcslock_unlock(rte_mcslock_t * __rte_atomic *msl, rte_mcslock_t * __rte_atomic me) {/* Check if there are more nodes in the queue. */ -if (likely (__atomic_load_n (&me->next, __ATOMIC_RELAXED) ==NULL )) { +if (likely (rte_atomic_load_explicit (&me->next, rte_memory_order_relaxed) ==NULL )) {/* No, last member in the queue. */ -rte_mcslock_t *save_me =__atomic_load_n (&me, __ATOMIC_RELAXED); +rte_mcslock_t *save_me =rte_atomic_load_explicit (&me, rte_memory_order_relaxed);/* Release the lock by setting it to NULL */ -if (likely (__atomic_compare_exchange_n (msl, &save_me,NULL ,0 , - __ATOMIC_RELEASE, __ATOMIC_RELAXED))) +if (likely (rte_atomic_compare_exchange_strong_explicit (msl, &save_me,NULL , + rte_memory_order_release, rte_memory_order_relaxed)))return ;/* Speculative execution would be allowed to read in the * while-loop first. This has the potential to cause a * deadlock. Need a load barrier. lib/eal/include/rte_mcslock.h line 130
Quoted 10 lines
lib/eal/include/rte_mcslock.h +3
-3*/ -__atomic_thread_fence (__ATOMIC_ACQUIRE); +__rte_atomic_thread_fence (rte_memory_order_acquire);/* More nodes added to the queue by other CPUs. * Wait until the next pointer is set. */ -uintptr_t *next; - next = (uintptr_t *)&me->next; + uintptr_t __rte_atomic *next; + next = (uintptr_t __rte_atomic *)&me->next;This way around, I think:
__rte_atomic uintptr_t *next; next = (__rte_atomic uintptr_t *)&me->next;[...]
lib/eal/include/rte_pflock.h line 45
Quoted 10 lines
lib/eal/include/rte_pflock.h +2
-2@@ -41,8 +41,8 @@*/ struct rte_pflock {struct { -uint16_t in; -uint16_t out; + uint16_t __rte_atomic in; + uint16_t __rte_atomic out;Again, I think
__rte_atomicshould be before the type:__rte_atomic uint16_t in; __rte_atomic uint16_t out;Quoted 2 lines
lib/eal/include/rte_pflock.h +0
-0} rd, wr; };[...]
lib/eal/include/rte_seqcount.h line 36
Quoted 9 lines
lib/eal/include/rte_seqcount.h +1
-1@@ -32,7 +32,7 @@* The RTE seqcount type. */ typedef struct { - uint32_t sn; /**< A sequence number for the protected data. */ + uint32_t __rte_atomic sn; /**< A sequence number for the protected data. */Again, I think
__rte_atomicshould be before the type:__rte_atomic uint32_t sn; /**< A sequence [...]Quoted 1 line
lib/eal/include/rte_seqcount.h +0
-0} rte_seqcount_t;[...]
lib/eal/include/rte_ticketlock.h line 36
Quoted 13 lines
lib/eal/include/rte_ticketlock.h +3
-3@@ -30,10 +30,10 @@* The rte_ticketlock_t type. */ typedef union { - uint32_t tickets; + uint32_t __rte_atomic tickets; struct { - uint16_t current; - uint16_t next; + uint16_t __rte_atomic current; + uint16_t __rte_atomic next;Again, I think
__rte_atomicshould be before the type:__rte_atomic uint16_t current; __rte_atomic uint16_t next;Quoted 11 lines
lib/eal/include/rte_ticketlock.h +1
-1} s; } rte_ticketlock_t;@@ -127,7 +129,7 @@ typedef struct { rte_ticketlock_t tl; /**< the actual ticketlock */ - int user; /**< core id using lock, TICKET_LOCK_INVALID_ID for unused */ + int __rte_atomic user; /**< core id using lock, TICKET_LOCK_INVALID_IDQuoted 1 line
for unused */
Again, I think
__rte_atomicshould be before the type:__rte_atomic int user; /**< core id [...]Quoted 2 lines
unsigned int count; /**< count of time this lock has been called */
}rte_ticketlock_recursive_t;[...]
lib/eal/include/rte_trace_point.h line 36
Quoted 8 lines
lib/eal/include/rte_trace_point.h +1
-1@@ -33,7 +33,7 @@#include <rte_stdatomic.h> /** The tracepoint object. */ -typedef uint64_t rte_trace_point_t ; +typedef uint64_t __rte_atomic rte_trace_point_t;Again, I think
__rte_atomicshould be before the type:
typedef__rte_atomicuint64_trte_trace_point_t;[...]
At the risk of having gone "speed blind" by all the search-replaces along the way...
Reviewed-by Morten Brørup <[email protected]>
RE: [PATCH v2 2/6] eal: adapt EAL to present rte optional atomics API
Quoted 23 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Monday, 14 August 2023 19.47Quoted 4 lines · Morten Brørup
Quoted 4 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Friday, 11 August 2023 19.32Adapt the EAL public headers to use rte optional atomics API instead
of
Quoted 1 line
Quoted 1 line
directly using and exposing toolchain specific atomic builtin
intrinsics.
Quoted 6 lines
Quoted 3 lines
Signed-off-by Tyler Retzlaff <[email protected]>---
[...]
will fix the comments identified.
Quoted 3 lines
[...]
Quoted 17 lines
Quoted 8 lines
lib/eal/include/generic/rte_spinlock.h line 32
Quoted 8 lines
lib/eal/include/generic/rte_spinlock.h +1
-1@@ -29,7 +29,7 @@* The rte_spinlock_t type. */ typedef struct __rte_lockable { - volatile int locked; /**< lock status 0 = unlocked, 1 = locked */ + volatile int __rte_atomic locked; /**< lock status 0 = unlocked, 1=
Quoted 8 lines
Quoted 1 line
locked */
I think
__rte_atomicshould be before the type:
volatile__rte_atomicint locked; /**< lock status [...]
Alternatively (just mentioning it, I know we don't use this form):volatile __rte_atomic(int) locked; /**< lock status [...]Thinking of where you would put "const" might help.
Regarding "const", I use the mental trick of reading from right-to-left when pointers are involved, e.g.:
const int * * const x;
----5---- 4 3 --2-- 1
x(1) is a const(2) pointer(3) to a pointer(4) to a const int(5).
And yes, treating "const int" as one word is cheating... formally it should be "int" "const", i.e. the reverse order; but that is not the convention, so I have learned to accept it.
Quoted 39 lines
Quoted 2 lines
Maybe your order is also correct, so it is a matter of preference.
so for me what you suggest is the canonical convention for c and i did initially try to make the change with this convention but ran into trouble when using the keyword in a context used as a type specifier and the type was incomplete.
the
rte_mcslockis a good example for illustration.// original struct typedef struct rte_mcslock { struct rte_mcslock *next; ... };it simply doesn't work / won't compile (at least with clang) which is what drove me to use the less-often used syntax.
typedef struct rte_mcslock { _Atomic struct rte_mcslock *next; ... };In file included from
../app/test/test_mcslock.c:19:
..\lib\eal\include\rte_mcslock.h:36:2: error:_Atomiccannot be
applied
to incomplete type 'structrte_mcslock'_Atomic struct rte_mcslock *next; ^ ..\lib\eal\include\rte_mcslock.h:35:16: note: definition of 'struct rte_mcslock' is not complete until the closing '}' typedef struct rte_mcslock { ^1 error generated.
so i ended up choosing to use a single syntax by convention consistently rather than using one for the exceptional case and one everywhere else.
Quoted 7 lines
i think (based on our other thread of discussion) i would recommend we use adopt and require the use of the
_Atomic(T)macro to disambiguate it also has the advantage of not being churned later when we can do c++23.// using macro typedef struct rte_mcslock { _Atomic(struct rte_mcslock *) next;
This makes it an atomic pointer. Your example above tried making the struct rts_mcslock atomic. Probably what you wanted was:
typedef struct rte_mcslock {
struct rte_mcslock * _Atomic next;
...};
Like "const", the convention should be putting it before any type, but after the "*" for pointers.
I suppose clang doesn't accept applying _Atomic to incomplete types, regardless where you put it... I.e. this should also fail, I guess:
typedef struct rte_mcslock {
struct rte_mcslock _Atomic * next;
...};
Quoted 6 lines
...};
this is much easier at a glance to know when the specified type is the T or the T * similarly in parameter lists it becomes more clear too.
Quoted 6 lines
e.g.
voidfoo(int *v)that it is either void
foo(_Atomic(int) *v)or void foo(_Atomic(int *)
v) becomes
much clearer without having to do mental gymnastics.
The same could be said about making "const" clearer:
void foo(const(int) * v) instead of void foo(const int * v), and void foo(const(int *) v) instead of void foo(int * const v).
Luckily, we don't need toolchain specific handling of "const", so let's just leave that the way it is. :-)
Quoted 14 lines
so i propose we retain
#define __rte_atomic _Atomicallow it to be used in contexts where we need a type-qualifier. note:
most of the cases where_Atomicis used as a type-qualifier it is a red flag that we are sensitive to an implementation detail of the compiler. in time i hope most of these will go away as we remove deprecatedrte_atomic_xxapis.but also introduce the following macro
Quoted 7 lines
#define
RTE_ATOMIC(type)_Atomic(type)
require it be used in the contexts that we are using it as a type-
specifier.if folks agree with this please reply back positively and i'll update the series. feel free to propose alternate names or whatever, but sooner than later so i don't have to churn things too much :)
+1 to Tyler's updated proposal, with macro names as suggested.
If anyone disagrees, please speak up soon!
If in doubt, please read https://en.cppreference.com/w/c/language/atomic carefully. It says:
(1) _Atomic(type-name) (since C11): Use as a type specifier; this designates a new atomic type.
(2) _Atomic type-name (since C11): Use as a type qualifier; this designates the atomic version of type-name. In this role, it may be mixed with const, volatile, and restrict, although unlike other qualifiers, the atomic version of type-name may have a different size, alignment, and object representation.
NB: I hadn't noticed this before, otherwise I had probably suggested using _Atomic(T) earlier on. We learn something new every day. :-)
Quoted 2 lines
thanks!
Sorry about the late response, Tyler. Other work prevented me from setting aside coherent time to review your updated proposal.
Quoted 8 lines
Quoted 3 lines
The DPDK coding style guidelines doesn't mention where to place
"const", but looking at the code, it seems to use "const unsigned int" and "const char *".
we probably should document it as a convention and most likely we should adopt what is already in use more commonly.
+1, but not as part of this series. :-)
Re: [PATCH v2 2/6] eal: adapt EAL to present rte optional atomics API
Quoted 39 lines · Morten Brørup
Quoted 39 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Monday, 14 August 2023 19.47Quoted 4 lines · Morten Brørup
Quoted 4 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Friday, 11 August 2023 19.32Adapt the EAL public headers to use rte optional atomics API instead
of
Quoted 1 line
Quoted 1 line
directly using and exposing toolchain specific atomic builtin
intrinsics.
Quoted 6 lines
Quoted 3 lines
Signed-off-by Tyler Retzlaff <[email protected]>---
[...]
will fix the comments identified.
Quoted 11 lines
[...]
lib/eal/include/generic/rte_spinlock.h line 32
Quoted 8 lines
lib/eal/include/generic/rte_spinlock.h +1
-1@@ -29,7 +29,7 @@* The rte_spinlock_t type. */ typedef struct __rte_lockable { - volatile int locked; /**< lock status 0 = unlocked, 1 = locked */ + volatile int __rte_atomic locked; /**< lock status 0 = unlocked, 1=
Quoted 7 lines
Quoted 1 line
locked */
I think
__rte_atomicshould be before the type:
volatile__rte_atomicint locked; /**< lock status [...]
Alternatively (just mentioning it, I know we don't use this form):volatile __rte_atomic(int) locked; /**< lock status [...]
Quoted 6 lines
Quoted 1 line
Quoted 1 line
Thinking of where you would put "const" might help.
Regarding "const", I use the mental trick of reading from right-to-left when pointers are involved, e.g.:
const int * * const x;
----5---- 4 3 --2-- 1
yes, i'm very familiar with where it can appear in the syntax and applied. but it's always good to have someone summarize it like this for the discussion.
Quoted 4 lines
x(1) is a const(2) pointer(3) to a pointer(4) to a const int(5).
And yes, treating "const int" as one word is cheating... formally it should be "int" "const", i.e. the reverse order; but that is not the convention, so I have learned to accept it.
it more often is the convention in c++, but i agree in c conventionally people put the const first.
Quoted 48 lines
Quoted 40 lines
Quoted 3 lines
Maybe your order is also correct, so it is a matter of preference.
so for me what you suggest is the canonical convention for c and i did initially try to make the change with this convention but ran into trouble when using the keyword in a context used as a type specifier and the type was incomplete.
the
rte_mcslockis a good example for illustration.// original struct typedef struct rte_mcslock { struct rte_mcslock *next; ... };it simply doesn't work / won't compile (at least with clang) which is what drove me to use the less-often used syntax.
typedef struct rte_mcslock { _Atomic struct rte_mcslock *next; ... };In file included from
../app/test/test_mcslock.c:19:
..\lib\eal\include\rte_mcslock.h:36:2: error:_Atomiccannot be
applied
to incomplete type 'structrte_mcslock'_Atomic struct rte_mcslock *next; ^ ..\lib\eal\include\rte_mcslock.h:35:16: note: definition of 'struct rte_mcslock' is not complete until the closing '}' typedef struct rte_mcslock { ^1 error generated.
so i ended up choosing to use a single syntax by convention consistently rather than using one for the exceptional case and one everywhere else.
Quoted 7 lines
i think (based on our other thread of discussion) i would recommend we use adopt and require the use of the
_Atomic(T)macro to disambiguate it also has the advantage of not being churned later when we can do c++23.// using macro typedef struct rte_mcslock { _Atomic(struct rte_mcslock *) next;
Quoted 5 lines
This makes it an atomic pointer. Your example above tried making the struct
rts_mcslockatomic. Probably what you wanted was:typedef struct rte_mcslock { struct rte_mcslock * _Atomic next; ...};
this is what my v2 in the patch had. but following your const example you indicated you preferred the equivalent of const T' over T const` i was trying to illustrate that if you replace T = struct foo * the compiler can't disambiguate between type and pointer to type and produces an error.
Quoted 2 lines
Like "const", the convention should be putting it before any type, but after the "*" for pointers.
i see, thank you for this clarification. I had not understood that you were suggesting that for pointer types specifically i should use one placement and for non-pointer types i should use another.
Quoted 6 lines
I suppose clang doesn't accept applying
_Atomicto incomplete types, regardless where you put it... I.e. this should also fail, I guess:typedef struct rte_mcslock { struct rte_mcslock _Atomic * next; ...};
actually I think for C11 atomics i think you can actually do this because you can declare an entire struct object to be atomic. However, since we need to intersect with what non-C11 gcc builtin atomics do we would not be able to make struct objects atomic as gcc only let's you do atomic things with integer and pointer types.
Quoted 38 lines
Quoted 7 lines
...};
this is much easier at a glance to know when the specified type is the T or the T * similarly in parameter lists it becomes more clear too.
Quoted 6 lines
e.g.
voidfoo(int *v)that it is either void
foo(_Atomic(int) *v)or void foo(_Atomic(int *)
v) becomes
much clearer without having to do mental gymnastics.The same could be said about making "const" clearer:
voidfoo(const(int) * v)instead of voidfoo(const int * v), and voidfoo(const(int *) v)instead of voidfoo(int * const v).Luckily, we don't need toolchain specific handling of "const", so let's just leave that the way it is. :-)
Quoted 18 lines
so i propose we retain
#define __rte_atomic _Atomicallow it to be used in contexts where we need a type-qualifier. note:
most of the cases where_Atomicis used as a type-qualifier it is a red flag that we are sensitive to an implementation detail of the compiler. in time i hope most of these will go away as we remove deprecatedrte_atomic_xxapis.but also introduce the following macro
#define
RTE_ATOMIC(type)_Atomic(type)
require it be used in the contexts that we are using it as a type-
specifier.
Quoted 5 lines
Quoted 3 lines
if folks agree with this please reply back positively and i'll update the series. feel free to propose alternate names or whatever, but sooner than later so i don't have to churn things too much :)
+1 to Tyler's updated proposal, with macro names as suggested.
yeah, I think it really helps clarify the pointer vs regular type specification by whacking the ( ) around what we are talking about instead of using positioning of _Atomic in two different places.
Quoted 8 lines
If anyone disagrees, please speak up soon!
If in doubt, please read https://en.cppreference.com/w/c/language/atomic carefully. It says:
(1)_Atomic(type-name)(since C11): Use as a type specifier; this designates a new atomic type.
(2)_Atomictype-name (since C11): Use as a type qualifier; this designates the atomic version of type-name. In this role, it may be mixed with const, volatile, and restrict, although unlike other qualifiers, the atomic version of type-name may have a different size, alignment, and object representation.NB: I hadn't noticed this before, otherwise I had probably suggested using
_Atomic(T)earlier on. We learn something new every day. :-)
yeah, i knew about this which is why i was being really careful about 'qualification' vs 'specification' in my mails.
Quoted 5 lines
Quoted 3 lines
thanks!
Sorry about the late response, Tyler. Other work prevented me from setting aside coherent time to review your updated proposal.
meh it's okay, based on the other thread i kind of guessed you might agree with using _Atomic(T) so i just submitted a new version an hour ago with the changes. i hope it meets your approval, one thing i'm kind of edgy about is the actual macro name itself RTE_ATOMIC(type) it seems kinda ugly, so if someone has an opinion there i'm open to it.
Quoted 11 lines
Quoted 9 lines
Quoted 4 lines
The DPDK coding style guidelines doesn't mention where to place
"const", but looking at the code, it seems to use "const unsigned int" and "const char *".
we probably should document it as a convention and most likely we should adopt what is already in use more commonly.
+1, but not as part of this series. :-)
i'll look into doing it once we get this series merged.
thanks!
PATCH v2 3/6 eal: add rte atomic qualifier with casts
In reply to Tyler RetzlaffIntroduce __rte_atomic qualifying casts in rte_optional atomics inline functions to prevent cascading the need to pass __rte_atomic qualified arguments.
Warning, this is really implementation dependent and being done temporarily to avoid having to convert more of the libraries and tests in DPDK in the initial series that introduces the API. The consequence of the assumption of the ABI of the types in question not being ``the same'' is only a risk that may be realized when enable_stdatomic=true.
--- lib/eal/include/generic/rte_atomic.h | 48 ++++++++++++++++++++++++------------ lib/eal/include/generic/rte_pause.h | 9 ++++--- lib/eal/x86/rte_power_intrinsics.c | 7 +++--- 3 files changed, 42 insertions(+), 22 deletions(-)
lib/eal/include/generic/rte_atomic.h +32 -16
@@ -274,7 +274,8 @@static inline void rte_atomic16_add (rte_atomic16_t *v,int16_t inc) { -rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst); + rte_atomic_fetch_add_explicit((volatile int16_t __rte_atomic *)&v->cnt, inc, + rte_memory_order_seq_cst); }/**
@@ -288,7 +289,8 @@static inline void rte_atomic16_sub (rte_atomic16_t *v,int16_t dec) { -rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst); + rte_atomic_fetch_sub_explicit((volatile int16_t __rte_atomic *)&v->cnt, dec, + rte_memory_order_seq_cst); }/**
@@ -341,7 +343,8 @@static inline int16_t rte_atomic16_add_return (rte_atomic16_t *v,int16_t inc) { -return rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst) + inc; + return rte_atomic_fetch_add_explicit((volatile int16_t __rte_atomic *)&v->cnt, inc, + rte_memory_order_seq_cst) + inc; }/**
@@ -361,7 +364,8 @@static inline int16_t rte_atomic16_sub_return (rte_atomic16_t *v,int16_t dec) { -return rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst) - dec; + return rte_atomic_fetch_sub_explicit((volatile int16_t __rte_atomic *)&v->cnt, dec, + rte_memory_order_seq_cst) - dec; }/**
@@ -380,7 +384,8 @@#ifdef RTE_FORCE_INTRINSICS static inline int rte_atomic16_inc_and_test (rte_atomic16_t *v) { -return rte_atomic_fetch_add_explicit (&v->cnt,1 , rte_memory_order_seq_cst) +1 ==0 ; + return rte_atomic_fetch_add_explicit((volatile int16_t __rte_atomic *)&v->cnt, 1, + rte_memory_order_seq_cst) +1 ==0 ; }#endif
@@ -400,7 +405,8 @@ static inline int rte_atomic16_inc_and_test(rte_atomic16_t *v)#ifdef RTE_FORCE_INTRINSICS static inline int rte_atomic16_dec_and_test (rte_atomic16_t *v) { -return rte_atomic_fetch_sub_explicit (&v->cnt,1 , rte_memory_order_seq_cst) -1 ==0 ; + return rte_atomic_fetch_sub_explicit((volatile int16_t __rte_atomic *)&v->cnt, 1, + rte_memory_order_seq_cst) -1 ==0 ; }#endif
@@ -553,7 +559,8 @@ static inline void rte_atomic16_clear(rte_atomic16_t *v)static inline void rte_atomic32_add (rte_atomic32_t *v,int32_t inc) { -rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst); + rte_atomic_fetch_add_explicit((volatile int32_t __rte_atomic *)&v->cnt, inc, + rte_memory_order_seq_cst); }/**
@@ -567,7 +574,8 @@ static inline void rte_atomic16_clear(rte_atomic16_t *v)static inline void rte_atomic32_sub (rte_atomic32_t *v,int32_t dec) { -rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst); + rte_atomic_fetch_sub_explicit((volatile int32_t __rte_atomic *)&v->cnt, dec, + rte_memory_order_seq_cst); }/**
@@ -620,7 +628,8 @@ static inline void rte_atomic16_clear(rte_atomic16_t *v)static inline int32_t rte_atomic32_add_return (rte_atomic32_t *v,int32_t inc) { -return rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst) + inc; + return rte_atomic_fetch_add_explicit((volatile int32_t __rte_atomic *)&v->cnt, inc, + rte_memory_order_seq_cst) + inc; }/**
@@ -640,7 +649,8 @@ static inline void rte_atomic16_clear(rte_atomic16_t *v)static inline int32_t rte_atomic32_sub_return (rte_atomic32_t *v,int32_t dec) { -return rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst) - dec; + return rte_atomic_fetch_sub_explicit((volatile int32_t __rte_atomic *)&v->cnt, dec, + rte_memory_order_seq_cst) - dec; }/**
@@ -659,7 +669,8 @@ static inline void rte_atomic16_clear(rte_atomic16_t *v)#ifdef RTE_FORCE_INTRINSICS static inline int rte_atomic32_inc_and_test (rte_atomic32_t *v) { -return rte_atomic_fetch_add_explicit (&v->cnt,1 , rte_memory_order_seq_cst) +1 ==0 ; + return rte_atomic_fetch_add_explicit((volatile int32_t __rte_atomic *)&v->cnt, 1, + rte_memory_order_seq_cst) +1 ==0 ; }#endif
@@ -679,7 +690,8 @@ static inline int rte_atomic32_inc_and_test(rte_atomic32_t *v)#ifdef RTE_FORCE_INTRINSICS static inline int rte_atomic32_dec_and_test (rte_atomic32_t *v) { -return rte_atomic_fetch_sub_explicit (&v->cnt,1 , rte_memory_order_seq_cst) -1 ==0 ; + return rte_atomic_fetch_sub_explicit((volatile int32_t __rte_atomic *)&v->cnt, 1, + rte_memory_order_seq_cst) -1 ==0 ; }#endif
@@ -885,7 +897,8 @@ static inline void rte_atomic32_clear(rte_atomic32_t *v)static inline void rte_atomic64_add (rte_atomic64_t *v,int64_t inc) { -rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst); + rte_atomic_fetch_add_explicit((volatile int64_t __rte_atomic *)&v->cnt, inc, + rte_memory_order_seq_cst); }#endif
@@ -904,7 +917,8 @@ static inline void rte_atomic32_clear(rte_atomic32_t *v)static inline void rte_atomic64_sub (rte_atomic64_t *v,int64_t dec) { -rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst); + rte_atomic_fetch_sub_explicit((volatile int64_t __rte_atomic *)&v->cnt, dec, + rte_memory_order_seq_cst); }#endif
@@ -962,7 +976,8 @@ static inline void rte_atomic32_clear(rte_atomic32_t *v)static inline int64_t rte_atomic64_add_return (rte_atomic64_t *v,int64_t inc) { -return rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst) + inc; + return rte_atomic_fetch_add_explicit((volatile int64_t __rte_atomic *)&v->cnt, inc, + rte_memory_order_seq_cst) + inc; }#endif
@@ -986,7 +1001,8 @@ static inline void rte_atomic32_clear(rte_atomic32_t *v)static inline int64_t rte_atomic64_sub_return (rte_atomic64_t *v,int64_t dec) { -return rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst) - dec; + return rte_atomic_fetch_sub_explicit((volatile int64_t __rte_atomic *)&v->cnt, dec, + rte_memory_order_seq_cst) - dec; }#endif
lib/eal/include/generic/rte_pause.h +6 -3
@@ -87,7 +87,8 @@ {assert (memorder == rte_memory_order_acquire || memorder == rte_memory_order_relaxed); -while (rte_atomic_load_explicit (addr, memorder) != expected) + while (rte_atomic_load_explicit((volatile uint16_t __rte_atomic *)addr, memorder) + != expected)rte_pause (); }
@@ -97,7 +98,8 @@ {assert (memorder == rte_memory_order_acquire || memorder == rte_memory_order_relaxed); -while (rte_atomic_load_explicit (addr, memorder) != expected) + while (rte_atomic_load_explicit((volatile uint32_t __rte_atomic *)addr, memorder) + != expected)rte_pause (); }
@@ -107,7 +109,8 @@ {assert (memorder == rte_memory_order_acquire || memorder == rte_memory_order_relaxed); -while (rte_atomic_load_explicit (addr, memorder) != expected) + while (rte_atomic_load_explicit((volatile uint64_t __rte_atomic *)addr, memorder) + != expected)rte_pause (); }
lib/eal/x86/rte_power_intrinsics.c +4 -3
@@ -23,9 +23,10 @@uint64_t val;/* trigger a write but don't change the value */ - val =rte_atomic_load_explicit ((volatile uint64_t *)addr, rte_memory_order_relaxed); -rte_atomic_compare_exchange_strong_explicit ((volatile uint64_t *)addr, &val, val, - rte_memory_order_relaxed, rte_memory_order_relaxed); + val = rte_atomic_load_explicit((volatile uint64_t __rte_atomic *)addr, + rte_memory_order_relaxed); + rte_atomic_compare_exchange_strong_explicit((volatile uint64_t __rte_atomic *)addr, + &val, val, rte_memory_order_relaxed, rte_memory_order_relaxed); }static bool wait_supported;
Signature · 1 line
1.8.3.1
RE: [PATCH v2 3/6] eal: add rte atomic qualifier with casts
Quoted 27 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Friday, 11 August 2023 19.32Introduce
__rte_atomicqualifying casts inrte_optionalatomics inline functions to prevent cascading the need to pass__rte_atomicqualified arguments.Warning, this is really implementation dependent and being done temporarily to avoid having to convert more of the libraries and tests in DPDK in the initial series that introduces the API. The consequence of the assumption of the ABI of the types in question not being ``the same'' is only a risk that may be realized when enable_stdatomic=true.
Signed-off-by Tyler Retzlaff <[email protected]>--- lib/eal/include/generic/rte_atomic.h | 48 ++++++++++++++++++++++++-----------(quoted diff) +0
-1- lib/eal/include/generic/rte_pause.h | 9 ++++--- lib/eal/x86/rte_power_intrinsics.c | 7 +++--- 3 files changed, 42 insertions(+), 22 deletions(-)lib/eal/include/generic/rte_atomic.h +0
-0
lib/eal/include/generic/rte_atomic.h line 279
Quoted 7 lines
lib/eal/include/generic/rte_atomic.h +2
-1@@ -274,7 +274,8 @@static inline void rte_atomic16_add (rte_atomic16_t *v,int16_t inc) { -rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst); + rte_atomic_fetch_add_explicit((volatile int16_t __rte_atomic *)&v->cnt, inc, + rte_memory_order_seq_cst);
As mentioned in my review to the 2/6 patch, I think __rte_atomic should come before the type, like this:
(volatile __rte_atomic int16_t *)
Same with all the changes.
Otherwise good.
PATCH v2 4/6 distributor: adapt for EAL optional atomics API changes
In reply to Tyler RetzlaffAdapt distributor for EAL optional atomics API changes
--- lib/distributor/distributor_private.h | 2 +- lib/distributor/rte_distributor_single.c | 44 ++++++++++++++++---------------- 2 files changed, 23 insertions(+), 23 deletions(-)
lib/distributor/distributor_private.h +1 -1
@@ -52,7 +52,7 @@* Only 64-bits of the memory is actually used though. */ union rte_distributor_buffer_single { -volatile int64_t bufptr64; + volatile int64_t __rte_atomic bufptr64;char pad[RTE_CACHE_LINE_SIZE *3 ]; } __rte_cache_aligned;
lib/distributor/rte_distributor_single.c +22 -22
117 lines not on this page · raw message
Signature · 1 line
1.8.3.1
RE: [PATCH v2 4/6] distributor: adapt for EAL optional atomics API changes
Quoted 18 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Friday, 11 August 2023 19.32Adapt distributor for EAL optional atomics API changes
Signed-off-by Tyler Retzlaff <[email protected]>--- lib/distributor/distributor_private.h | 2 +- lib/distributor/rte_distributor_single.c | 44 ++++++++++++++++---------------(quoted diff) +0
-1- 2 files changed, 23 insertions(+), 23 deletions(-)lib/distributor/distributor_private.h +0
-0
lib/distributor/distributor_private.h line 55
Quoted 5 lines
lib/distributor/distributor_private.h +1
-1@@ -52,7 +52,7 @@* Only 64-bits of the memory is actually used though. */ union rte_distributor_buffer_single { -volatile int64_t bufptr64; + volatile int64_t __rte_atomic bufptr64;
As mentioned in my review to the 2/6 patch, I think __rte_atomic should come before the type, like this:
lib/distributor/distributor_private.h line 57
Quoted 4 lines
lib/distributor/distributor_private.h +1
-0+ volatile __rte_atomic int64_t bufptr64;char pad[RTE_CACHE_LINE_SIZE *3 ]; } __rte_cache_aligned;
The rest is simple search-replace; easy to review, so...
PATCH v2 5/6 bpf: adapt for EAL optional atomics API changes
In reply to Tyler RetzlaffAdapt bpf for EAL optional atomics API changes
--- lib/bpf/bpf_pkt.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)
lib/bpf/bpf_pkt.c +3 -3
@@ -25,7 +25,7 @@struct bpf_eth_cbi {/* used by both data & control path */ -uint32_t use;/*usage counter */ + uint32_t __rte_atomic use; /*usage counter */const struct rte_eth_rxtx_callback *cb;/* callback handle */ struct rte_bpf *bpf;struct rte_bpf_jit jit;
@@ -110,8 +110,8 @@ struct bpf_eth_cbh {/* in use, busy wait till current RX/TX iteration is finished */ if ((puse &BPF_ETH_CBI_INUSE ) !=0 ) { -RTE_WAIT_UNTIL_MASKED ((uint32_t *)(uintptr_t )&cbi->use, - UINT32_MAX, !=, puse, __ATOMIC_RELAXED); + RTE_WAIT_UNTIL_MASKED((uint32_t __rte_atomic *)(uintptr_t)&cbi->use, + UINT32_MAX, !=, puse, rte_memory_order_relaxed); } }
Signature · 1 line
1.8.3.1
RE: [PATCH v2 5/6] bpf: adapt for EAL optional atomics API changes
Quoted 15 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Friday, 11 August 2023 19.32Adapt bpf for EAL optional atomics API changes
Signed-off-by Tyler Retzlaff <[email protected]>--- lib/bpf/bpf_pkt.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)lib/bpf/bpf_pkt.c +0
-0
lib/bpf/bpf_pkt.c line 28
Quoted 5 lines
lib/bpf/bpf_pkt.c +1
-1@@ -25,7 +25,7 @@struct bpf_eth_cbi {/* used by both data & control path */ -uint32_t use;/*usage counter */ + uint32_t __rte_atomic use; /*usage counter */
As mentioned in my review to the 2/6 patch, I think __rte_atomic should come before the type, like this:
__rte_atomic uint32_t use; /*usage counter */
lib/bpf/bpf_pkt.c struct bpf_eth_cbh { · line 113
Quoted 10 lines
lib/bpf/bpf_pkt.c +1
-2const struct rte_eth_rxtx_callback *cb;/* callback handle */ struct rte_bpf *bpf;struct rte_bpf_jit jit;@@ -110,8 +110,8 @@ struct bpf_eth_cbh {/* in use, busy wait till current RX/TX iteration is finished */ if ((puse &BPF_ETH_CBI_INUSE ) !=0 ) { -RTE_WAIT_UNTIL_MASKED ((uint32_t *)(uintptr_t )&cbi->use, - UINT32_MAX, !=, puse, __ATOMIC_RELAXED); + RTE_WAIT_UNTIL_MASKED((uint32_t __rte_atomic *)(uintptr_t)&cbi->use,
And here:
RTE_WAIT_UNTIL_MASKED((__rte_atomic uint32_t *) [...]
Quoted 6 lines
lib/bpf/bpf_pkt.c +1
-1+ UINT32_MAX, !=, puse, rte_memory_order_relaxed); } } --1.8.3.1
PATCH v2 6/6 devtools: forbid new direct use of GCC atomic builtins
In reply to Tyler RetzlaffRefrain from using compiler __atomic_xxx builtins DPDK now requires the use of rte_atomic_<op>_explicit macros when operating on DPDK atomic variables.
--- devtools/checkpatches.sh | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)
devtools/checkpatches.sh +3 -3
@@ -111,11 +111,11 @@ check_forbidden_additions() { # <patch> -f $(dirname $(readlink -f $0))/check-forbidden-tokens.awk \ "$1" || res=1 - # refrain from using compiler __atomic_{add,and,nand,or,sub,xor}_fetch() + # refrain from using compiler __atomic_xxx builtins awk -v FOLDERS="lib drivers app examples" \ - -v EXPRESSIONS="__atomic_(add|and|nand|or|sub|xor)_fetch\\\(" \ + -v EXPRESSIONS="__atomic_.*\\\(" \ -v RET_ON_FAIL=1 \ - -v MESSAGE='Using __atomic_op_fetch, prefer __atomic_fetch_op' \ + -v MESSAGE='Using __atomic_xxx builtins' \ -f $(dirname $(readlink -f $0))/check-forbidden-tokens.awk \ "$1" || res=1
Signature · 1 line
1.8.3.1
RE: [PATCH v2 6/6] devtools: forbid new direct use of GCC atomic builtins
Quoted 11 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Friday, 11 August 2023 19.32Refrain from using compiler
__atomic_xxxbuiltins DPDK now requires the use of rte_atomic_<op>_explicit macros when operating on DPDK atomic variables.Signed-off-by Tyler Retzlaff <[email protected]>Suggested-by Morten Brørup <[email protected]>Acked-by Bruce Richardson <[email protected]>---
v3 cover RFC optional rte optional stdatomics API
In reply to Tyler RetzlaffThis series introduces API additions prefixed in the rte namespace that allow the optional use of stdatomics.h from C11 using enable_stdatomics=true for targets where enable_stdatomics=false no functional change is intended.
Be aware this does not contain all changes to use stdatomics across the DPDK tree it only introduces the minimum to allow the option to be used which is a pre-requisite for a clean CI (probably using clang) that can be run with enable_stdatomics=true enabled.
It is planned that subsequent series will be introduced per lib/driver as appropriate to further enable stdatomics use when enable_stdatomics=true.
Notes:
* Additional libraries beyond EAL make visible atomics use across the API/ABI surface they will be converted in the subsequent series.
* The eal: add rte atomic qualifier with casts patch needs some discussion as to whether or not the legacy rte_atomic APIs should be converted to work with enable_stdatomic=true right now some implementation dependent casts are used to prevent cascading / having to convert too much in the intial series.
* Windows will obviously need complete conversion of libraries including atomics that are not crossing API/ABI boundaries. those conversions will introduced in separate series as new along side the existing msvc series.
Please keep in mind we would like to prioritize the review / acceptance of this patch since it needs to be completed in the 23.11 merge window.
Thank you all for the discussion that lead to the formation of this series.
v3:
* Remove comments from APIs mentioning the mapping to C++ memory model memory orders
* Introduce and use new macro RTE_ATOMIC(type) to be used in contexts where _Atomic is used as a type specifier to declare variables. The macro allows more clarity about what the atomic type being specified is. e.g. _Atomic(T *) vs _Atomic(T) it is easier to understand that the former is an atomic pointer type and the latter is an atomic type. it also has the benefit of (in the future) being interoperable with c++23 syntactically
note: Morten i have retained your 'reviewed-by' tags if you disagree given the changes in the above version please indicate as such but i believe the changes are in the spirit of the feedback you provided
v2:
* Wrap meson_options.txt option description to newline and indent to be consistent with other options.
* Provide separate typedef of rte_memory_order for enable_stdatomic=true VS enable_stdatomic=false instead of a single typedef to int
note: slight tweak to reviewers feedback i've chosen to use a typedef
for both enable_stdatomic={true,false} (just seemed more consistent)
* Bring in assert.h and use static_assert macro instead of _Static_assert keyword to better interoperate with c/c++
* Directly include rte_stdatomic.h where into other places it is consumed instead of hacking it globally into rte_config.h
* Provide and use __rte_atomic_thread_fence to allow conditional expansion within the body of existing rte_atomic_thread_fence inline function to maintain per-arch optimizations when enable_stdatomic=false
Tyler Retzlaff (6):
eal: provide rte stdatomics optional atomics API
eal: adapt EAL to present rte optional atomics API
eal: add rte atomic qualifier with casts
distributor: adapt for EAL optional atomics API changes
bpf: adapt for EAL optional atomics API changes
devtools: forbid new direct use of GCC atomic builtins
app/test/test_mcslock.c | 6 +- config/meson.build | 1 + devtools/checkpatches.sh | 6 +- lib/bpf/bpf_pkt.c | 6 +- lib/distributor/distributor_private.h | 2 +- lib/distributor/rte_distributor_single.c | 44 ++++---- lib/eal/arm/include/rte_atomic_32.h | 4 +- lib/eal/arm/include/rte_atomic_64.h | 36 +++--- lib/eal/arm/include/rte_pause_64.h | 26 ++--- lib/eal/arm/rte_power_intrinsics.c | 8 +- lib/eal/common/eal_common_trace.c | 16 +-- lib/eal/include/generic/rte_atomic.h | 67 ++++++----- lib/eal/include/generic/rte_pause.h | 50 ++++----- lib/eal/include/generic/rte_rwlock.h | 48 ++++---- lib/eal/include/generic/rte_spinlock.h | 20 ++-- lib/eal/include/meson.build | 1 + lib/eal/include/rte_mcslock.h | 51 ++++----- lib/eal/include/rte_pflock.h | 25 +++-- lib/eal/include/rte_seqcount.h | 19 ++-- lib/eal/include/rte_stdatomic.h | 184 +++++++++++++++++++++++++++++++ lib/eal/include/rte_ticketlock.h | 43 ++++---- lib/eal/include/rte_trace_point.h | 5 +- lib/eal/loongarch/include/rte_atomic.h | 4 +- lib/eal/ppc/include/rte_atomic.h | 54 ++++----- lib/eal/riscv/include/rte_atomic.h | 4 +- lib/eal/x86/include/rte_atomic.h | 8 +- lib/eal/x86/include/rte_spinlock.h | 2 +- lib/eal/x86/rte_power_intrinsics.c | 7 +- meson_options.txt | 2 + 29 files changed, 483 insertions(+), 266 deletions(-) create mode 100644 lib/eal/include/rte_stdatomic.h
Signature · 1 line
1.8.3.1
PATCH v3 1/6 eal: provide rte stdatomics optional atomics API
Provide API for atomic operations in the rte namespace that may optionally be configured to use C11 atomics with meson option enable_stdatomics=true
--- config/meson.build | 1 + lib/eal/include/generic/rte_atomic.h | 1 + lib/eal/include/generic/rte_pause.h | 1 + lib/eal/include/generic/rte_rwlock.h | 1 + lib/eal/include/generic/rte_spinlock.h | 1 + lib/eal/include/meson.build | 1 + lib/eal/include/rte_mcslock.h | 1 + lib/eal/include/rte_pflock.h | 1 + lib/eal/include/rte_seqcount.h | 1 + lib/eal/include/rte_stdatomic.h | 182 +++++++++++++++++++++++++++++++++ lib/eal/include/rte_ticketlock.h | 1 + lib/eal/include/rte_trace_point.h | 1 + meson_options.txt | 2 + 13 files changed, 195 insertions(+) create mode 100644 lib/eal/include/rte_stdatomic.h
config/meson.build +1 -0
@@ -303,6 +303,7 @@ endforeach # set other values pulled from the build options dpdk_conf.set('RTE_MAX_ETHPORTS', get_option('max_ethports')) dpdk_conf.set('RTE_LIBEAL_USE_HPET', get_option('use_hpet')) +dpdk_conf.set('RTE_ENABLE_STDATOMIC', get_option('enable_stdatomic')) dpdk_conf.set('RTE_ENABLE_TRACE_FP', get_option('enable_trace_fp')) # values which have defaults which may be overridden dpdk_conf.set('RTE_MAX_VFIO_GROUPS', 64)
lib/eal/include/generic/rte_atomic.h +1 -0
@@ -15,6 +15,7 @@#include <stdint.h> #include <rte_compat.h> #include <rte_common.h> +#include <rte_stdatomic.h> #ifdef __DOXYGEN__
lib/eal/include/generic/rte_pause.h +1 -0
@@ -16,6 +16,7 @@#include <assert.h> #include <rte_common.h> #include <rte_atomic.h> +#include <rte_stdatomic.h> /** * Pause CPU execution for a short while
lib/eal/include/generic/rte_rwlock.h +1 -0
@@ -32,6 +32,7 @@#include <rte_common.h> #include <rte_lock_annotations.h> #include <rte_pause.h> +#include <rte_stdatomic.h> /** * The rte_rwlock_t type.
lib/eal/include/generic/rte_spinlock.h +1 -0
@@ -23,6 +23,7 @@#endif #include <rte_lock_annotations.h> #include <rte_pause.h> +#include <rte_stdatomic.h> /** * The rte_spinlock_t type.
lib/eal/include/meson.build +1 -0
@@ -42,6 +42,7 @@ headers += files( 'rte_seqlock.h', 'rte_service.h', 'rte_service_component.h', + 'rte_stdatomic.h', 'rte_string_fns.h', 'rte_tailq.h', 'rte_thread.h',
lib/eal/include/rte_mcslock.h +1 -0
@@ -27,6 +27,7 @@#include <rte_common.h> #include <rte_pause.h> #include <rte_branch_prediction.h> +#include <rte_stdatomic.h> /** * The rte_mcslock_t type.
lib/eal/include/rte_pflock.h +1 -0
@@ -34,6 +34,7 @@#include <rte_compat.h> #include <rte_common.h> #include <rte_pause.h> +#include <rte_stdatomic.h> /** * The rte_pflock_t type.
lib/eal/include/rte_seqcount.h +1 -0
@@ -26,6 +26,7 @@#include <rte_atomic.h> #include <rte_branch_prediction.h> #include <rte_compat.h> +#include <rte_stdatomic.h> /** * The RTE seqcount type.
lib/eal/include/rte_stdatomic.h new file +182 -0
182 lines not on this page · raw message
lib/eal/include/rte_ticketlock.h +1 -0
@@ -24,6 +24,7 @@#include <rte_common.h> #include <rte_lcore.h> #include <rte_pause.h> +#include <rte_stdatomic.h> /** * The rte_ticketlock_t type.
lib/eal/include/rte_trace_point.h +1 -0
@@ -30,6 +30,7 @@#include <rte_per_lcore.h> #include <rte_string_fns.h> #include <rte_uuid.h> +#include <rte_stdatomic.h> /** The tracepoint object. */ typedef uint64_t rte_trace_point_t ;
meson_options.txt +2 -0
@@ -46,6 +46,8 @@ option('mbuf_refcnt_atomic', type: 'boolean', value: true, description: 'Atomically access the mbuf refcnt.') option('platform', type: 'string', value: 'native', description: 'Platform to build, either "native", "generic" or a SoC. Please refer to the Linux build guide for more information.') +option('enable_stdatomic', type: 'boolean', value: false, description: + 'enable use of C11 stdatomic') option('enable_trace_fp', type: 'boolean', value: false, description: 'enable fast path trace points.') option('tests', type: 'boolean', value: true, description:
Signature · 1 line
1.8.3.1
RE: [PATCH v3 1/6] eal: provide rte stdatomics optional atomics API
Quoted 10 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Wednesday, 16 August 2023 21.19Provide API for atomic operations in the rte namespace that may optionally be configured to use C11 atomics with meson option enable_stdatomics=true
Signed-off-by Tyler Retzlaff <[email protected]>Reviewed-by Morten Brørup <[email protected]>---
"#define RTE_ATOMIC(type) [...]" is missing in lib/eal/include/rte_stdatomic.h, both with and without RTE_ENABLE_STDATOMIC.
I suggest you keep it together with "#define __rte_atomic [...]".
Please also add descriptions as comments to both (type-qualifier and -specifier) in both locations. Your descriptions from the mailing list discussion are fine.
Re: [PATCH v3 1/6] eal: provide rte stdatomics optional atomics API
Quoted 7 lines · Morten Brørup
Quoted 7 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Wednesday, 16 August 2023 21.19Provide API for atomic operations in the rte namespace that may optionally be configured to use C11 atomics with meson option enable_stdatomics=true
Quoted 7 lines
Quoted 3 lines
Signed-off-by Tyler Retzlaff <[email protected]>Reviewed-by Morten Brørup <[email protected]>---
"#define
RTE_ATOMIC(type)[...]" is missing inlib/eal/include/rte_stdatomic.h, both with and withoutRTE_ENABLE_STDATOMIC.I suggest you keep it together with "#define
__rte_atomic[...]".
this seems to be an error i made in patch submission, somehow i managed to drop the changes i made from v2-v3 from the series i submitted i'm confused how i managed to fat finger it.
i'll submit a new series with the correct changes to rte_stdatomic.h
Quoted 3 lines
Please also add descriptions as comments to both (type-qualifier and -specifier) in both locations. Your descriptions from the mailing list discussion are fine.
i'll do this in the v4 commit message
thanks
RE: [PATCH v3 1/6] eal: provide rte stdatomics optional atomics API
Quoted 16 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Wednesday, 16 August 2023 23.04Quoted 7 lines · Morten Brørup
Quoted 7 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Wednesday, 16 August 2023 21.19Provide API for atomic operations in the rte namespace that may optionally be configured to use C11 atomics with meson option enable_stdatomics=true
Quoted 5 lines
Quoted 3 lines
Signed-off-by Tyler Retzlaff <[email protected]>Reviewed-by Morten Brørup <[email protected]>---
"#define
RTE_ATOMIC(type)[...]" is missing in
Quoted 8 lines
lib/eal/include/rte_stdatomic.h, both with and withoutRTE_ENABLE_STDATOMIC.Quoted 2 lines
I suggest you keep it together with "#define
__rte_atomic[...]".this seems to be an error i made in patch submission, somehow i managed to drop the changes i made from v2-v3 from the series i submitted i'm confused how i managed to fat finger it.
Happens to me all the time - too often, if you look at my patch submission track record. ;-)
Quoted 9 lines
i'll submit a new series with the correct changes to
rte_stdatomic.hQuoted 2 lines
Please also add descriptions as comments to both (type-qualifier and -
specifier) in both locations. Your descriptions from the mailing list discussion are fine.
Quoted 1 line
i'll do this in the v4 commit message
Please put the descriptions about their usage in the code as documentation comments, not in the commit message.
Quoted 2 lines
thanks
Re: [PATCH v3 1/6] eal: provide rte stdatomics optional atomics API
In reply to Tyler RetzlaffQuoted 12 lines · Tyler Retzlaff
Quoted 11 lines · Morten Brørup
Quoted 10 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Wednesday, 16 August 2023 21.19Provide API for atomic operations in the rte namespace that may optionally be configured to use C11 atomics with meson option enable_stdatomics=true
Signed-off-by Tyler Retzlaff <[email protected]>Reviewed-by Morten Brørup <[email protected]>---
Quoted 7 lines
Quoted 3 lines
"#define
RTE_ATOMIC(type)[...]" is missing inlib/eal/include/rte_stdatomic.h, both with and withoutRTE_ENABLE_STDATOMIC.I suggest you keep it together with "#define
__rte_atomic[...]".this seems to be an error i made in patch submission, somehow i managed to drop the changes i made from v2-v3 from the series i submitted i'm confused how i managed to fat finger it.
okay! i know what i did. somehow the additions of the macros were placed into patch 2 instead of patch 1 where they belong. i'm glad there's a simple explanation :)
v4 coming up with the it moved into the patch it belongs in.
Quoted 9 lines
i'll submit a new series with the correct changes to
rte_stdatomic.hQuoted 3 lines
Please also add descriptions as comments to both (type-qualifier and -specifier) in both locations. Your descriptions from the mailing list discussion are fine.
i'll do this in the v4 commit message
thanks
PATCH v3 2/6 eal: adapt EAL to present rte optional atomics API
In reply to Tyler RetzlaffAdapt the EAL public headers to use rte optional atomics API instead of directly using and exposing toolchain specific atomic builtin intrinsics.
--- app/test/test_mcslock.c | 6 ++-- lib/eal/arm/include/rte_atomic_32.h | 4 +-- lib/eal/arm/include/rte_atomic_64.h | 36 +++++++++++------------ lib/eal/arm/include/rte_pause_64.h | 26 ++++++++-------- lib/eal/arm/rte_power_intrinsics.c | 8 ++--- lib/eal/common/eal_common_trace.c | 16 +++++----- lib/eal/include/generic/rte_atomic.h | 50 +++++++++++++++---------------- lib/eal/include/generic/rte_pause.h | 46 ++++++++++++----------------- lib/eal/include/generic/rte_rwlock.h | 47 +++++++++++++++-------------- lib/eal/include/generic/rte_spinlock.h | 19 ++++++------ lib/eal/include/rte_mcslock.h | 50 +++++++++++++++---------------- lib/eal/include/rte_pflock.h | 24 ++++++++------- lib/eal/include/rte_seqcount.h | 18 ++++++------ lib/eal/include/rte_stdatomic.h | 2 ++ lib/eal/include/rte_ticketlock.h | 42 +++++++++++++------------- lib/eal/include/rte_trace_point.h | 4 +-- lib/eal/loongarch/include/rte_atomic.h | 4 +-- lib/eal/ppc/include/rte_atomic.h | 54 +++++++++++++++++----------------- lib/eal/riscv/include/rte_atomic.h | 4 +-- lib/eal/x86/include/rte_atomic.h | 8 ++--- lib/eal/x86/include/rte_spinlock.h | 2 +- lib/eal/x86/rte_power_intrinsics.c | 6 ++-- 22 files changed, 239 insertions(+), 237 deletions(-)
app/test/test_mcslock.c +3 -3
@@ -36,9 +36,9 @@* lock multiple times. */ -rte_mcslock_t *p_ml; -rte_mcslock_t *p_ml_try; -rte_mcslock_t *p_ml_perf; +RTE_ATOMIC (rte_mcslock_t *) p_ml; +RTE_ATOMIC (rte_mcslock_t *) p_ml_try; +RTE_ATOMIC (rte_mcslock_t *) p_ml_perf;static unsigned int count;
lib/eal/arm/include/rte_atomic_32.h +2 -2
@@ -34,9 +34,9 @@#define rte_io_rmb () rte_rmb() static __rte_always_inline void -rte_atomic_thread_fence (int memorder) +rte_atomic_thread_fence (rte_memory_order memorder) { -__atomic_thread_fence (memorder); +__rte_atomic_thread_fence (memorder); }#ifdef __cplusplus
lib/eal/arm/include/rte_atomic_64.h +18 -18
@@ -38,9 +38,9 @@#define rte_io_rmb () rte_rmb() static __rte_always_inline void -rte_atomic_thread_fence (int memorder) +rte_atomic_thread_fence (rte_memory_order memorder) { -__atomic_thread_fence (memorder); +__rte_atomic_thread_fence (memorder); }/*------------------------ 128 bit atomic operations -------------------------*/
@@ -107,33 +107,33 @@*/ RTE_SET_USED (failure);/* Find invalid memory order */ -RTE_ASSERT (success == __ATOMIC_RELAXED || - success == __ATOMIC_ACQUIRE || - success == __ATOMIC_RELEASE || - success == __ATOMIC_ACQ_REL || - success == __ATOMIC_SEQ_CST); +RTE_ASSERT (success == rte_memory_order_relaxed || + success == rte_memory_order_acquire || + success == rte_memory_order_release || + success == rte_memory_order_acq_rel || + success == rte_memory_order_seq_cst);rte_int128_t expected = *exp;rte_int128_t desired = *src;rte_int128_t old;#if defined(__ARM_FEATURE_ATOMICS) || defined(RTE_ARM_FEATURE_ATOMICS ) -if (success == __ATOMIC_RELAXED) +if (success == rte_memory_order_relaxed)__cas_128_relaxed (dst, exp, desired); -else if (success == __ATOMIC_ACQUIRE) +else if (success == rte_memory_order_acquire)__cas_128_acquire (dst, exp, desired); -else if (success == __ATOMIC_RELEASE) +else if (success == rte_memory_order_release)__cas_128_release (dst, exp, desired);else __cas_128_acq_rel (dst, exp, desired); old = *exp;#else -#define __HAS_ACQ (mo) ((mo) != __ATOMIC_RELAXED && (mo) != __ATOMIC_RELEASE) -#define __HAS_RLS (mo) ((mo) == __ATOMIC_RELEASE || (mo) == __ATOMIC_ACQ_REL || \ - (mo) == __ATOMIC_SEQ_CST) +#define __HAS_ACQ (mo) ((mo) != rte_memory_order_relaxed && (mo) != rte_memory_order_release) +#define __HAS_RLS (mo) ((mo) == rte_memory_order_release || (mo) == rte_memory_order_acq_rel || \ + (mo) == rte_memory_order_seq_cst) -int ldx_mo =__HAS_ACQ (success) ? __ATOMIC_ACQUIRE : __ATOMIC_RELAXED; -int stx_mo =__HAS_RLS (success) ? __ATOMIC_RELEASE : __ATOMIC_RELAXED; +int ldx_mo =__HAS_ACQ (success) ? rte_memory_order_acquire : rte_memory_order_relaxed; +int stx_mo =__HAS_RLS (success) ? rte_memory_order_release : rte_memory_order_relaxed;#undef __HAS_ACQ#undef __HAS_RLS
@@ -153,7 +153,7 @@ : "Q" (src->val[0]) \ : "memory"); } -if (ldx_mo == __ATOMIC_RELAXED) +if (ldx_mo == rte_memory_order_relaxed) __LOAD_128("ldxp", dst, old) else__LOAD_128 ("ldaxp" , dst, old)
@@ -170,7 +170,7 @@ : "memory"); }if (likely (old.int128 == expected.int128)) { -if (stx_mo == __ATOMIC_RELAXED) +if (stx_mo == rte_memory_order_relaxed) __STORE_128("stxp", dst, desired, ret) else__STORE_128 ("stlxp" , dst, desired, ret)
@@ -181,7 +181,7 @@* needs to be stored back to ensure it was read * atomically. */ -if (stx_mo == __ATOMIC_RELAXED) +if (stx_mo == rte_memory_order_relaxed) __STORE_128("stxp", dst, old, ret) else__STORE_128 ("stlxp" , dst, old, ret)
lib/eal/arm/include/rte_pause_64.h +13 -13
@@ -41,7 +41,7 @@ static inline void rte_pause(void)* implicitly to exit WFE. */ #define __RTE_ARM_LOAD_EXC_8 (src, dst, memorder) { \ - if (memorder == __ATOMIC_RELAXED) { \ + if (memorder == rte_memory_order_relaxed) { \ asm volatile("ldxrb %w[tmp], [%x[addr]]" \ : [tmp] "=&r" (dst) \ : [addr] "r" (src) \
@@ -60,7 +60,7 @@ static inline void rte_pause(void)* implicitly to exit WFE. */ #define __RTE_ARM_LOAD_EXC_16 (src, dst, memorder) { \ - if (memorder == __ATOMIC_RELAXED) { \ + if (memorder == rte_memory_order_relaxed) { \ asm volatile("ldxrh %w[tmp], [%x[addr]]" \ : [tmp] "=&r" (dst) \ : [addr] "r" (src) \
@@ -79,7 +79,7 @@ static inline void rte_pause(void)* implicitly to exit WFE. */ #define __RTE_ARM_LOAD_EXC_32 (src, dst, memorder) { \ - if (memorder == __ATOMIC_RELAXED) { \ + if (memorder == rte_memory_order_relaxed) { \ asm volatile("ldxr %w[tmp], [%x[addr]]" \ : [tmp] "=&r" (dst) \ : [addr] "r" (src) \
@@ -98,7 +98,7 @@ static inline void rte_pause(void)* implicitly to exit WFE. */ #define __RTE_ARM_LOAD_EXC_64 (src, dst, memorder) { \ - if (memorder == __ATOMIC_RELAXED) { \ + if (memorder == rte_memory_order_relaxed) { \ asm volatile("ldxr %x[tmp], [%x[addr]]" \ : [tmp] "=&r" (dst) \ : [addr] "r" (src) \
@@ -118,7 +118,7 @@ static inline void rte_pause(void)*/ #define __RTE_ARM_LOAD_EXC_128 (src, dst, memorder) { \ volatile rte_int128_t *dst_128 = (volatile rte_int128_t *)&dst; \ - if (memorder == __ATOMIC_RELAXED) { \ + if (memorder == rte_memory_order_relaxed) { \ asm volatile("ldxp %x[tmp0], %x[tmp1], [%x[addr]]" \ : [tmp0] "=&r" (dst_128->val[0]), \ [tmp1] "=&r" (dst_128->val[1]) \
@@ -153,8 +153,8 @@ static inline void rte_pause(void) {uint16_t value; -RTE_BUILD_BUG_ON (memorder != __ATOMIC_ACQUIRE && - memorder != __ATOMIC_RELAXED); +RTE_BUILD_BUG_ON (memorder != rte_memory_order_acquire && + memorder != rte_memory_order_relaxed); __RTE_ARM_LOAD_EXC_16(addr, value, memorder)if (value != expected) {
@@ -172,8 +172,8 @@ static inline void rte_pause(void) {uint32_t value; -RTE_BUILD_BUG_ON (memorder != __ATOMIC_ACQUIRE && - memorder != __ATOMIC_RELAXED); +RTE_BUILD_BUG_ON (memorder != rte_memory_order_acquire && + memorder != rte_memory_order_relaxed); __RTE_ARM_LOAD_EXC_32(addr, value, memorder)if (value != expected) {
@@ -191,8 +191,8 @@ static inline void rte_pause(void) {uint64_t value; -RTE_BUILD_BUG_ON (memorder != __ATOMIC_ACQUIRE && - memorder != __ATOMIC_RELAXED); +RTE_BUILD_BUG_ON (memorder != rte_memory_order_acquire && + memorder != rte_memory_order_relaxed); __RTE_ARM_LOAD_EXC_64(addr, value, memorder)if (value != expected) {
@@ -206,8 +206,8 @@ static inline void rte_pause(void)#define RTE_WAIT_UNTIL_MASKED (addr, mask, cond, expected, memorder) do { \ RTE_BUILD_BUG_ON(!__builtin_constant_p(memorder)); \ - RTE_BUILD_BUG_ON(memorder != __ATOMIC_ACQUIRE && \ - memorder != __ATOMIC_RELAXED); \ + RTE_BUILD_BUG_ON(memorder != rte_memory_order_acquire && \ + memorder != rte_memory_order_relaxed); \ const uint32_t size = sizeof(*(addr)) << 3; \ typeof(*(addr)) expected_value = (expected); \ typeof(*(addr)) value; \
lib/eal/arm/rte_power_intrinsics.c +4 -4
@@ -33,19 +33,19 @@switch (pmc->size) {case sizeof (uint8_t ): - __RTE_ARM_LOAD_EXC_8(pmc->addr, cur_value, __ATOMIC_RELAXED) + __RTE_ARM_LOAD_EXC_8(pmc->addr, cur_value, rte_memory_order_relaxed)__RTE_ARM_WFE () break;case sizeof (uint16_t ): - __RTE_ARM_LOAD_EXC_16(pmc->addr, cur_value, __ATOMIC_RELAXED) + __RTE_ARM_LOAD_EXC_16(pmc->addr, cur_value, rte_memory_order_relaxed)__RTE_ARM_WFE () break;case sizeof (uint32_t ): - __RTE_ARM_LOAD_EXC_32(pmc->addr, cur_value, __ATOMIC_RELAXED) + __RTE_ARM_LOAD_EXC_32(pmc->addr, cur_value, rte_memory_order_relaxed)__RTE_ARM_WFE () break;case sizeof (uint64_t ): - __RTE_ARM_LOAD_EXC_64(pmc->addr, cur_value, __ATOMIC_RELAXED) + __RTE_ARM_LOAD_EXC_64(pmc->addr, cur_value, rte_memory_order_relaxed)__RTE_ARM_WFE () break;default :
lib/eal/common/eal_common_trace.c +9 -7
@@ -103,11 +103,11 @@ struct trace_point_head * trace_mode_set(rte_trace_point_t *t, enum rte_trace_mode mode) {if (mode ==RTE_TRACE_MODE_OVERWRITE ) -__atomic_fetch_and (t, ~__RTE_TRACE_FIELD_ENABLE_DISCARD, - __ATOMIC_RELEASE); +rte_atomic_fetch_and_explicit (t, ~__RTE_TRACE_FIELD_ENABLE_DISCARD, + rte_memory_order_release);else -__atomic_fetch_or (t, __RTE_TRACE_FIELD_ENABLE_DISCARD, - __ATOMIC_RELEASE); +rte_atomic_fetch_or_explicit (t, __RTE_TRACE_FIELD_ENABLE_DISCARD, + rte_memory_order_release); }void
@@ -141,7 +141,7 @@ rte_trace_mode rte_trace_mode_get(void)if (trace_point_is_invalid (t))return false; - val =__atomic_load_n (t, __ATOMIC_ACQUIRE); + val =rte_atomic_load_explicit (t, rte_memory_order_acquire);return (val & __RTE_TRACE_FIELD_ENABLE_MASK) !=0 ; }
@@ -153,7 +153,8 @@ rte_trace_mode rte_trace_mode_get(void)if (trace_point_is_invalid (t))return -ERANGE ; - prev =__atomic_fetch_or (t, __RTE_TRACE_FIELD_ENABLE_MASK, __ATOMIC_RELEASE); + prev =rte_atomic_fetch_or_explicit (t, __RTE_TRACE_FIELD_ENABLE_MASK, + rte_memory_order_release);if ((prev & __RTE_TRACE_FIELD_ENABLE_MASK) ==0 )__atomic_fetch_add (&trace.status,1 , __ATOMIC_RELEASE);return 0 ;
@@ -167,7 +168,8 @@ rte_trace_mode rte_trace_mode_get(void)if (trace_point_is_invalid (t))return -ERANGE ; - prev =__atomic_fetch_and (t, ~__RTE_TRACE_FIELD_ENABLE_MASK, __ATOMIC_RELEASE); + prev =rte_atomic_fetch_and_explicit (t, ~__RTE_TRACE_FIELD_ENABLE_MASK, + rte_memory_order_release);if ((prev & __RTE_TRACE_FIELD_ENABLE_MASK) !=0 )__atomic_fetch_sub (&trace.status,1 , __ATOMIC_RELEASE);return 0 ;
lib/eal/include/generic/rte_atomic.h +25 -25
194 lines not on this page · raw message
lib/eal/include/generic/rte_pause.h +19 -27
103 lines not on this page · raw message
lib/eal/include/generic/rte_rwlock.h +25 -22
126 lines not on this page · raw message
lib/eal/include/generic/rte_spinlock.h +9 -10
@@ -29,7 +29,7 @@* The rte_spinlock_t type. */ typedef struct __rte_lockable { -volatile int locked;/**< lock status 0 = unlocked, 1 = locked */ +volatile RTE_ATOMIC (int ) locked;/**< lock status 0 = unlocked, 1 = locked */ }rte_spinlock_t ;/**
@@ -66,10 +66,10 @@ {int exp =0 ; -while (!__atomic_compare_exchange_n (&sl->locked, &exp,1 ,0 , - __ATOMIC_ACQUIRE, __ATOMIC_RELAXED)) { -rte_wait_until_equal_32 ((volatile uint32_t *)&sl->locked, -0 , __ATOMIC_RELAXED); +while (!rte_atomic_compare_exchange_strong_explicit (&sl->locked, &exp,1 , + rte_memory_order_acquire, rte_memory_order_relaxed)) { +rte_wait_until_equal_32 ((volatile uint32_t *)(uintptr_t )&sl->locked, +0 , rte_memory_order_relaxed); exp =0 ; } }
@@ -90,7 +90,7 @@ rte_spinlock_unlock(rte_spinlock_t *sl) __rte_no_thread_safety_analysis { -__atomic_store_n (&sl->locked,0 , __ATOMIC_RELEASE); +rte_atomic_store_explicit (&sl->locked,0 , rte_memory_order_release); }#endif
@@ -113,9 +113,8 @@ __rte_no_thread_safety_analysis {int exp =0 ; -return __atomic_compare_exchange_n (&sl->locked, &exp,1 , -0 ,/* disallow spurious failure */ - __ATOMIC_ACQUIRE, __ATOMIC_RELAXED); +return rte_atomic_compare_exchange_strong_explicit (&sl->locked, &exp,1 , + rte_memory_order_acquire, rte_memory_order_relaxed); }#endif
@@ -129,7 +128,7 @@*/ static inline int rte_spinlock_is_locked (rte_spinlock_t *sl) { -return __atomic_load_n (&sl->locked, __ATOMIC_ACQUIRE); +return rte_atomic_load_explicit (&sl->locked, rte_memory_order_acquire); }/**
lib/eal/include/rte_mcslock.h +25 -25
133 lines not on this page · raw message
lib/eal/include/rte_pflock.h +13 -11
@@ -41,8 +41,8 @@*/ struct rte_pflock {struct { -uint16_t in; -uint16_t out; +RTE_ATOMIC (uint16_t ) in; +RTE_ATOMIC (uint16_t ) out; } rd, wr; };typedef struct rte_pflock rte_pflock_t ;
@@ -117,14 +117,14 @@ struct rte_pflock {* If no writer is present, then the operation has completed * successfully. */ - w =__atomic_fetch_add (&pf->rd.in,RTE_PFLOCK_RINC , __ATOMIC_ACQUIRE) + w =rte_atomic_fetch_add_explicit (&pf->rd.in,RTE_PFLOCK_RINC , rte_memory_order_acquire) &RTE_PFLOCK_WBITS ;if (w ==0 )return ;/* Wait for current write phase to complete. */ RTE_WAIT_UNTIL_MASKED(&pf->rd.in, RTE_PFLOCK_WBITS, !=, w, - __ATOMIC_ACQUIRE); + rte_memory_order_acquire); }/**
@@ -140,7 +140,7 @@ struct rte_pflock {static inline void rte_pflock_read_unlock (rte_pflock_t *pf) { -__atomic_fetch_add (&pf->rd.out,RTE_PFLOCK_RINC , __ATOMIC_RELEASE); +rte_atomic_fetch_add_explicit (&pf->rd.out,RTE_PFLOCK_RINC , rte_memory_order_release); }/**
@@ -161,8 +161,9 @@ struct rte_pflock {/* Acquire ownership of write-phase. * This is same as rte_ticketlock_lock(). */ - ticket =__atomic_fetch_add (&pf->wr.in,1 , __ATOMIC_RELAXED); -rte_wait_until_equal_16 (&pf->wr.out, ticket, __ATOMIC_ACQUIRE); + ticket =rte_atomic_fetch_add_explicit (&pf->wr.in,1 , rte_memory_order_relaxed); +rte_wait_until_equal_16 ((uint16_t *)(uintptr_t )&pf->wr.out, ticket, + rte_memory_order_acquire);/* * Acquire ticket on read-side in order to allow them
@@ -173,10 +174,11 @@ struct rte_pflock {* speculatively. */ w =RTE_PFLOCK_PRES | (ticket &RTE_PFLOCK_PHID ); - ticket =__atomic_fetch_add (&pf->rd.in, w, __ATOMIC_RELAXED); + ticket =rte_atomic_fetch_add_explicit (&pf->rd.in, w, rte_memory_order_relaxed);/* Wait for any pending readers to flush. */ -rte_wait_until_equal_16 (&pf->rd.out, ticket, __ATOMIC_ACQUIRE); +rte_wait_until_equal_16 ((uint16_t *)(uintptr_t )&pf->rd.out, ticket, + rte_memory_order_acquire); }/**
@@ -193,10 +195,10 @@ struct rte_pflock { rte_pflock_write_unlock(rte_pflock_t *pf) {/* Migrate from write phase to read phase. */ -__atomic_fetch_and (&pf->rd.in,RTE_PFLOCK_LSB , __ATOMIC_RELEASE); +rte_atomic_fetch_and_explicit (&pf->rd.in,RTE_PFLOCK_LSB , rte_memory_order_release);/* Allow other writers to continue. */ -__atomic_fetch_add (&pf->wr.out,1 , __ATOMIC_RELEASE); +rte_atomic_fetch_add_explicit (&pf->wr.out,1 , rte_memory_order_release); }#ifdef __cplusplus
lib/eal/include/rte_seqcount.h +9 -9
@@ -32,7 +32,7 @@* The RTE seqcount type. */ typedef struct { -uint32_t sn;/**< A sequence number for the protected data. */ +RTE_ATOMIC (uint32_t ) sn;/**< A sequence number for the protected data. */ }rte_seqcount_t ;/**
@@ -106,11 +106,11 @@static inline uint32_t rte_seqcount_read_begin (const rte_seqcount_t *seqcount) { -/* __ATOMIC_ACQUIRE to prevent loads after (in program order) +/* rte_memory_order_acquire to prevent loads after (in program order) * from happening before the sn load. Synchronizes-with the * store release in rte_seqcount_write_end(). */ -return __atomic_load_n (&seqcount->sn, __ATOMIC_ACQUIRE); +return rte_atomic_load_explicit (&seqcount->sn, rte_memory_order_acquire); }/**
@@ -161,9 +161,9 @@return true;/* make sure the data loads happens before the sn load */ -rte_atomic_thread_fence (__ATOMIC_ACQUIRE); +rte_atomic_thread_fence (rte_memory_order_acquire); - end_sn =__atomic_load_n (&seqcount->sn, __ATOMIC_RELAXED); + end_sn =rte_atomic_load_explicit (&seqcount->sn, rte_memory_order_relaxed);/* A writer incremented the sequence number during this read * critical section.
@@ -205,12 +205,12 @@ sn = seqcount->sn +1 ; -__atomic_store_n (&seqcount->sn, sn, __ATOMIC_RELAXED); +rte_atomic_store_explicit (&seqcount->sn, sn, rte_memory_order_relaxed); -/* __ATOMIC_RELEASE to prevent stores after (in program order) +/* rte_memory_order_release to prevent stores after (in program order) * from happening before the sn store. */ -rte_atomic_thread_fence (__ATOMIC_RELEASE); +rte_atomic_thread_fence (rte_memory_order_release); }/**
@@ -237,7 +237,7 @@ sn = seqcount->sn +1 ;/* Synchronizes-with the load acquire in rte_seqcount_read_begin(). */ -__atomic_store_n (&seqcount->sn, sn, __ATOMIC_RELEASE); +rte_atomic_store_explicit (&seqcount->sn, sn, rte_memory_order_release); }#ifdef __cplusplus
lib/eal/include/rte_stdatomic.h +2 -0
@@ -18,6 +18,7 @@#include <stdatomic.h> +#define RTE_ATOMIC (type) _Atomic(type)#define __rte_atomic _Atomic/* The memory order is an enumerated type in C11. */
@@ -110,6 +111,7 @@#else +#define RTE_ATOMIC (type) type#define __rte_atomic/* The memory order is an integer type in GCC built-ins,
lib/eal/include/rte_ticketlock.h +22 -20
115 lines not on this page · raw message
lib/eal/include/rte_trace_point.h +2 -2
@@ -33,7 +33,7 @@#include <rte_stdatomic.h> /** The tracepoint object. */ -typedef uint64_t rte_trace_point_t ; +typedef RTE_ATOMIC (uint64_t )rte_trace_point_t ;/** * Macro to define the tracepoint arguments in RTE_TRACE_POINT macro.
@@ -359,7 +359,7 @@ struct __rte_trace_header {#define __rte_trace_point_emit_header_generic (t) \ void *mem; \ do { \ - const uint64_t val = __atomic_load_n(t, __ATOMIC_ACQUIRE); \ + const uint64_t val = rte_atomic_load_explicit(t, rte_memory_order_acquire); \ if (likely(!(val & __RTE_TRACE_FIELD_ENABLE_MASK))) \ return; \ mem = __rte_trace_mem_get(val); \
lib/eal/loongarch/include/rte_atomic.h +2 -2
@@ -35,9 +35,9 @@#define rte_io_rmb () rte_mb() static __rte_always_inline void -rte_atomic_thread_fence (int memorder) +rte_atomic_thread_fence (rte_memory_order memorder) { -__atomic_thread_fence (memorder); +__rte_atomic_thread_fence (memorder); }#ifdef __cplusplus
lib/eal/ppc/include/rte_atomic.h +27 -27
172 lines not on this page · raw message
lib/eal/riscv/include/rte_atomic.h +2 -2
@@ -40,9 +40,9 @@#define rte_io_rmb () asm volatile("fence ir, ir" : : : "memory") static __rte_always_inline void -rte_atomic_thread_fence (int memorder) +rte_atomic_thread_fence (rte_memory_order memorder) { -__atomic_thread_fence (memorder); +__rte_atomic_thread_fence (memorder); }#ifdef __cplusplus
lib/eal/x86/include/rte_atomic.h +4 -4
@@ -82,17 +82,17 @@/** * Synchronization fence between threads based on the specified memory order. * -* On x86 the __atomic_thread_fence(__ATOMIC_SEQ_CST) generates full 'mfence' +* On x86 the __rte_atomic_thread_fence(rte_memory_order_seq_cst) generates full 'mfence' * which is quite expensive. The optimized implementation of rte_smp_mb is * used instead. */ static __rte_always_inline void -rte_atomic_thread_fence (int memorder) +rte_atomic_thread_fence (rte_memory_order memorder) { -if (memorder == __ATOMIC_SEQ_CST) +if (memorder == rte_memory_order_seq_cst)rte_smp_mb ();else -__atomic_thread_fence (memorder); +__rte_atomic_thread_fence (memorder); }/*------------------------- 16 bit atomic operations -------------------------*/
lib/eal/x86/include/rte_spinlock.h +1 -1
@@ -78,7 +78,7 @@ static inline int rte_tm_supported(void) }static inline int -rte_try_tm (volatile int *lock) +rte_try_tm (volatile RTE_ATOMIC (int ) *lock) {int i, retries;
lib/eal/x86/rte_power_intrinsics.c +3 -3
@@ -23,9 +23,9 @@uint64_t val;/* trigger a write but don't change the value */ - val =__atomic_load_n ((volatile uint64_t *)addr, __ATOMIC_RELAXED); -__atomic_compare_exchange_n ((volatile uint64_t *)addr, &val, val,0 , - __ATOMIC_RELAXED, __ATOMIC_RELAXED); + val =rte_atomic_load_explicit ((volatile uint64_t *)addr, rte_memory_order_relaxed); +rte_atomic_compare_exchange_strong_explicit ((volatile uint64_t *)addr, &val, val, + rte_memory_order_relaxed, rte_memory_order_relaxed); }static bool wait_supported;
Signature · 1 line
1.8.3.1
PATCH v3 3/6 eal: add rte atomic qualifier with casts
In reply to Tyler RetzlaffIntroduce __rte_atomic qualifying casts in rte_optional atomics inline functions to prevent cascading the need to pass __rte_atomic qualified arguments.
Warning, this is really implementation dependent and being done temporarily to avoid having to convert more of the libraries and tests in DPDK in the initial series that introduces the API. The consequence of the assumption of the ABI of the types in question not being ``the same'' is only a risk that may be realized when enable_stdatomic=true.
--- lib/eal/include/generic/rte_atomic.h | 48 ++++++++++++++++++++++++------------ lib/eal/include/generic/rte_pause.h | 9 ++++--- lib/eal/x86/rte_power_intrinsics.c | 7 +++--- 3 files changed, 42 insertions(+), 22 deletions(-)
lib/eal/include/generic/rte_atomic.h +32 -16
@@ -274,7 +274,8 @@static inline void rte_atomic16_add (rte_atomic16_t *v,int16_t inc) { -rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst); + rte_atomic_fetch_add_explicit((volatile __rte_atomic int16_t *)&v->cnt, inc, + rte_memory_order_seq_cst); }/**
@@ -288,7 +289,8 @@static inline void rte_atomic16_sub (rte_atomic16_t *v,int16_t dec) { -rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst); + rte_atomic_fetch_sub_explicit((volatile __rte_atomic int16_t *)&v->cnt, dec, + rte_memory_order_seq_cst); }/**
@@ -341,7 +343,8 @@static inline int16_t rte_atomic16_add_return (rte_atomic16_t *v,int16_t inc) { -return rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst) + inc; + return rte_atomic_fetch_add_explicit((volatile __rte_atomic int16_t *)&v->cnt, inc, + rte_memory_order_seq_cst) + inc; }/**
@@ -361,7 +364,8 @@static inline int16_t rte_atomic16_sub_return (rte_atomic16_t *v,int16_t dec) { -return rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst) - dec; + return rte_atomic_fetch_sub_explicit((volatile __rte_atomic int16_t *)&v->cnt, dec, + rte_memory_order_seq_cst) - dec; }/**
@@ -380,7 +384,8 @@#ifdef RTE_FORCE_INTRINSICS static inline int rte_atomic16_inc_and_test (rte_atomic16_t *v) { -return rte_atomic_fetch_add_explicit (&v->cnt,1 , rte_memory_order_seq_cst) +1 ==0 ; + return rte_atomic_fetch_add_explicit((volatile __rte_atomic int16_t *)&v->cnt, 1, + rte_memory_order_seq_cst) +1 ==0 ; }#endif
@@ -400,7 +405,8 @@ static inline int rte_atomic16_inc_and_test(rte_atomic16_t *v)#ifdef RTE_FORCE_INTRINSICS static inline int rte_atomic16_dec_and_test (rte_atomic16_t *v) { -return rte_atomic_fetch_sub_explicit (&v->cnt,1 , rte_memory_order_seq_cst) -1 ==0 ; + return rte_atomic_fetch_sub_explicit((volatile __rte_atomic int16_t *)&v->cnt, 1, + rte_memory_order_seq_cst) -1 ==0 ; }#endif
@@ -553,7 +559,8 @@ static inline void rte_atomic16_clear(rte_atomic16_t *v)static inline void rte_atomic32_add (rte_atomic32_t *v,int32_t inc) { -rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst); + rte_atomic_fetch_add_explicit((volatile __rte_atomic int32_t *)&v->cnt, inc, + rte_memory_order_seq_cst); }/**
@@ -567,7 +574,8 @@ static inline void rte_atomic16_clear(rte_atomic16_t *v)static inline void rte_atomic32_sub (rte_atomic32_t *v,int32_t dec) { -rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst); + rte_atomic_fetch_sub_explicit((volatile __rte_atomic int32_t *)&v->cnt, dec, + rte_memory_order_seq_cst); }/**
@@ -620,7 +628,8 @@ static inline void rte_atomic16_clear(rte_atomic16_t *v)static inline int32_t rte_atomic32_add_return (rte_atomic32_t *v,int32_t inc) { -return rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst) + inc; + return rte_atomic_fetch_add_explicit((volatile __rte_atomic int32_t *)&v->cnt, inc, + rte_memory_order_seq_cst) + inc; }/**
@@ -640,7 +649,8 @@ static inline void rte_atomic16_clear(rte_atomic16_t *v)static inline int32_t rte_atomic32_sub_return (rte_atomic32_t *v,int32_t dec) { -return rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst) - dec; + return rte_atomic_fetch_sub_explicit((volatile __rte_atomic int32_t *)&v->cnt, dec, + rte_memory_order_seq_cst) - dec; }/**
@@ -659,7 +669,8 @@ static inline void rte_atomic16_clear(rte_atomic16_t *v)#ifdef RTE_FORCE_INTRINSICS static inline int rte_atomic32_inc_and_test (rte_atomic32_t *v) { -return rte_atomic_fetch_add_explicit (&v->cnt,1 , rte_memory_order_seq_cst) +1 ==0 ; + return rte_atomic_fetch_add_explicit((volatile __rte_atomic int32_t *)&v->cnt, 1, + rte_memory_order_seq_cst) +1 ==0 ; }#endif
@@ -679,7 +690,8 @@ static inline int rte_atomic32_inc_and_test(rte_atomic32_t *v)#ifdef RTE_FORCE_INTRINSICS static inline int rte_atomic32_dec_and_test (rte_atomic32_t *v) { -return rte_atomic_fetch_sub_explicit (&v->cnt,1 , rte_memory_order_seq_cst) -1 ==0 ; + return rte_atomic_fetch_sub_explicit((volatile __rte_atomic int32_t *)&v->cnt, 1, + rte_memory_order_seq_cst) -1 ==0 ; }#endif
@@ -885,7 +897,8 @@ static inline void rte_atomic32_clear(rte_atomic32_t *v)static inline void rte_atomic64_add (rte_atomic64_t *v,int64_t inc) { -rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst); + rte_atomic_fetch_add_explicit((volatile __rte_atomic int64_t *)&v->cnt, inc, + rte_memory_order_seq_cst); }#endif
@@ -904,7 +917,8 @@ static inline void rte_atomic32_clear(rte_atomic32_t *v)static inline void rte_atomic64_sub (rte_atomic64_t *v,int64_t dec) { -rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst); + rte_atomic_fetch_sub_explicit((volatile __rte_atomic int64_t *)&v->cnt, dec, + rte_memory_order_seq_cst); }#endif
@@ -962,7 +976,8 @@ static inline void rte_atomic32_clear(rte_atomic32_t *v)static inline int64_t rte_atomic64_add_return (rte_atomic64_t *v,int64_t inc) { -return rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst) + inc; + return rte_atomic_fetch_add_explicit((volatile __rte_atomic int64_t *)&v->cnt, inc, + rte_memory_order_seq_cst) + inc; }#endif
@@ -986,7 +1001,8 @@ static inline void rte_atomic32_clear(rte_atomic32_t *v)static inline int64_t rte_atomic64_sub_return (rte_atomic64_t *v,int64_t dec) { -return rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst) - dec; + return rte_atomic_fetch_sub_explicit((volatile __rte_atomic int64_t *)&v->cnt, dec, + rte_memory_order_seq_cst) - dec; }#endif
lib/eal/include/generic/rte_pause.h +6 -3
@@ -81,7 +81,8 @@ {assert (memorder == rte_memory_order_acquire || memorder == rte_memory_order_relaxed); -while (rte_atomic_load_explicit (addr, memorder) != expected) + while (rte_atomic_load_explicit((volatile __rte_atomic uint16_t *)addr, memorder) + != expected)rte_pause (); }
@@ -91,7 +92,8 @@ {assert (memorder == rte_memory_order_acquire || memorder == rte_memory_order_relaxed); -while (rte_atomic_load_explicit (addr, memorder) != expected) + while (rte_atomic_load_explicit((volatile __rte_atomic uint32_t *)addr, memorder) + != expected)rte_pause (); }
@@ -101,7 +103,8 @@ {assert (memorder == rte_memory_order_acquire || memorder == rte_memory_order_relaxed); -while (rte_atomic_load_explicit (addr, memorder) != expected) + while (rte_atomic_load_explicit((volatile __rte_atomic uint64_t *)addr, memorder) + != expected)rte_pause (); }
lib/eal/x86/rte_power_intrinsics.c +4 -3
@@ -23,9 +23,10 @@uint64_t val;/* trigger a write but don't change the value */ - val =rte_atomic_load_explicit ((volatile uint64_t *)addr, rte_memory_order_relaxed); -rte_atomic_compare_exchange_strong_explicit ((volatile uint64_t *)addr, &val, val, - rte_memory_order_relaxed, rte_memory_order_relaxed); + val = rte_atomic_load_explicit((volatile __rte_atomic uint64_t *)addr, + rte_memory_order_relaxed); + rte_atomic_compare_exchange_strong_explicit((volatile __rte_atomic uint64_t *)addr, + &val, val, rte_memory_order_relaxed, rte_memory_order_relaxed); }static bool wait_supported;
Signature · 1 line
1.8.3.1
PATCH v3 4/6 distributor: adapt for EAL optional atomics API changes
In reply to Tyler RetzlaffAdapt distributor for EAL optional atomics API changes
--- lib/distributor/distributor_private.h | 2 +- lib/distributor/rte_distributor_single.c | 44 ++++++++++++++++---------------- 2 files changed, 23 insertions(+), 23 deletions(-)
lib/distributor/distributor_private.h +1 -1
@@ -52,7 +52,7 @@* Only 64-bits of the memory is actually used though. */ union rte_distributor_buffer_single { -volatile int64_t bufptr64; +volatile RTE_ATOMIC (int64_t ) bufptr64;char pad[RTE_CACHE_LINE_SIZE *3 ]; } __rte_cache_aligned;
lib/distributor/rte_distributor_single.c +22 -22
117 lines not on this page · raw message
Signature · 1 line
1.8.3.1
PATCH v3 5/6 bpf: adapt for EAL optional atomics API changes
In reply to Tyler RetzlaffAdapt bpf for EAL optional atomics API changes
--- lib/bpf/bpf_pkt.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)
lib/bpf/bpf_pkt.c +3 -3
@@ -25,7 +25,7 @@struct bpf_eth_cbi {/* used by both data & control path */ -uint32_t use;/*usage counter */ +RTE_ATOMIC (uint32_t ) use;/*usage counter */ const struct rte_eth_rxtx_callback *cb;/* callback handle */ struct rte_bpf *bpf;struct rte_bpf_jit jit;
@@ -110,8 +110,8 @@ struct bpf_eth_cbh {/* in use, busy wait till current RX/TX iteration is finished */ if ((puse &BPF_ETH_CBI_INUSE ) !=0 ) { -RTE_WAIT_UNTIL_MASKED ((uint32_t *)(uintptr_t )&cbi->use, - UINT32_MAX, !=, puse, __ATOMIC_RELAXED); + RTE_WAIT_UNTIL_MASKED((__rte_atomic uint32_t *)(uintptr_t)&cbi->use, + UINT32_MAX, !=, puse, rte_memory_order_relaxed); } }
Signature · 1 line
1.8.3.1
PATCH v3 6/6 devtools: forbid new direct use of GCC atomic builtins
In reply to Tyler RetzlaffRefrain from using compiler __atomic_xxx builtins DPDK now requires the use of rte_atomic_<op>_explicit macros when operating on DPDK atomic variables.
--- devtools/checkpatches.sh | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)
devtools/checkpatches.sh +3 -3
@@ -111,11 +111,11 @@ check_forbidden_additions() { # <patch> -f $(dirname $(readlink -f $0))/check-forbidden-tokens.awk \ "$1" || res=1 - # refrain from using compiler __atomic_{add,and,nand,or,sub,xor}_fetch() + # refrain from using compiler __atomic_xxx builtins awk -v FOLDERS="lib drivers app examples" \ - -v EXPRESSIONS="__atomic_(add|and|nand|or|sub|xor)_fetch\\\(" \ + -v EXPRESSIONS="__atomic_.*\\\(" \ -v RET_ON_FAIL=1 \ - -v MESSAGE='Using __atomic_op_fetch, prefer __atomic_fetch_op' \ + -v MESSAGE='Using __atomic_xxx builtins' \ -f $(dirname $(readlink -f $0))/check-forbidden-tokens.awk \ "$1" || res=1
Signature · 1 line
1.8.3.1
v4 cover RFC optional rte optional stdatomics API
In reply to Tyler RetzlaffThis series introduces API additions prefixed in the rte namespace that allow the optional use of stdatomics.h from C11 using enable_stdatomics=true for targets where enable_stdatomics=false no functional change is intended.
Be aware this does not contain all changes to use stdatomics across the DPDK tree it only introduces the minimum to allow the option to be used which is a pre-requisite for a clean CI (probably using clang) that can be run with enable_stdatomics=true enabled.
It is planned that subsequent series will be introduced per lib/driver as appropriate to further enable stdatomics use when enable_stdatomics=true.
Notes:
* Additional libraries beyond EAL make visible atomics use across the API/ABI surface they will be converted in the subsequent series.
* The eal: add rte atomic qualifier with casts patch needs some discussion as to whether or not the legacy rte_atomic APIs should be converted to work with enable_stdatomic=true right now some implementation dependent casts are used to prevent cascading / having to convert too much in the intial series.
* Windows will obviously need complete conversion of libraries including atomics that are not crossing API/ABI boundaries. those conversions will introduced in separate series as new along side the existing msvc series.
Please keep in mind we would like to prioritize the review / acceptance of this patch since it needs to be completed in the 23.11 merge window.
Thank you all for the discussion that lead to the formation of this series.
v4:
* Move the definition of #define RTE_ATOMIC(type) to patch 1 where it belongs (a mistake in v3)
* Provide comments for both RTE_ATOMIC and __rte_atomic macros indicating their use as specified or qualified contexts.
v3:
* Remove comments from APIs mentioning the mapping to C++ memory model memory orders
* Introduce and use new macro RTE_ATOMIC(type) to be used in contexts where _Atomic is used as a type specifier to declare variables. The macro allows more clarity about what the atomic type being specified is. e.g. _Atomic(T *) vs _Atomic(T) it is easier to understand that the former is an atomic pointer type and the latter is an atomic type. it also has the benefit of (in the future) being interoperable with c++23 syntactically
note: Morten i have retained your 'reviewed-by' tags if you disagree given the changes in the above version please indicate as such but i believe the changes are in the spirit of the feedback you provided
v2:
* Wrap meson_options.txt option description to newline and indent to be consistent with other options.
* Provide separate typedef of rte_memory_order for enable_stdatomic=true VS enable_stdatomic=false instead of a single typedef to int
note: slight tweak to reviewers feedback i've chosen to use a typedef
for both enable_stdatomic={true,false} (just seemed more consistent)
* Bring in assert.h and use static_assert macro instead of _Static_assert keyword to better interoperate with c/c++
* Directly include rte_stdatomic.h where into other places it is consumed instead of hacking it globally into rte_config.h
* Provide and use __rte_atomic_thread_fence to allow conditional expansion within the body of existing rte_atomic_thread_fence inline function to maintain per-arch optimizations when enable_stdatomic=false
Tyler Retzlaff (6):
eal: provide rte stdatomics optional atomics API
eal: adapt EAL to present rte optional atomics API
eal: add rte atomic qualifier with casts
distributor: adapt for EAL optional atomics API changes
bpf: adapt for EAL optional atomics API changes
devtools: forbid new direct use of GCC atomic builtins
app/test/test_mcslock.c | 6 +- config/meson.build | 1 + devtools/checkpatches.sh | 6 +- lib/bpf/bpf_pkt.c | 6 +- lib/distributor/distributor_private.h | 2 +- lib/distributor/rte_distributor_single.c | 44 +++---- lib/eal/arm/include/rte_atomic_32.h | 4 +- lib/eal/arm/include/rte_atomic_64.h | 36 +++--- lib/eal/arm/include/rte_pause_64.h | 26 ++-- lib/eal/arm/rte_power_intrinsics.c | 8 +- lib/eal/common/eal_common_trace.c | 16 +-- lib/eal/include/generic/rte_atomic.h | 67 +++++++---- lib/eal/include/generic/rte_pause.h | 50 ++++---- lib/eal/include/generic/rte_rwlock.h | 48 ++++---- lib/eal/include/generic/rte_spinlock.h | 20 ++-- lib/eal/include/meson.build | 1 + lib/eal/include/rte_mcslock.h | 51 ++++---- lib/eal/include/rte_pflock.h | 25 ++-- lib/eal/include/rte_seqcount.h | 19 +-- lib/eal/include/rte_stdatomic.h | 198 +++++++++++++++++++++++++++++++ lib/eal/include/rte_ticketlock.h | 43 +++---- lib/eal/include/rte_trace_point.h | 5 +- lib/eal/loongarch/include/rte_atomic.h | 4 +- lib/eal/ppc/include/rte_atomic.h | 54 ++++----- lib/eal/riscv/include/rte_atomic.h | 4 +- lib/eal/x86/include/rte_atomic.h | 8 +- lib/eal/x86/include/rte_spinlock.h | 2 +- lib/eal/x86/rte_power_intrinsics.c | 7 +- meson_options.txt | 2 + 29 files changed, 497 insertions(+), 266 deletions(-) create mode 100644 lib/eal/include/rte_stdatomic.h
Signature · 1 line
1.8.3.1
PATCH v4 1/6 eal: provide rte stdatomics optional atomics API
Provide API for atomic operations in the rte namespace that may optionally be configured to use C11 atomics with meson option enable_stdatomics=true
--- config/meson.build | 1 + lib/eal/include/generic/rte_atomic.h | 1 + lib/eal/include/generic/rte_pause.h | 1 + lib/eal/include/generic/rte_rwlock.h | 1 + lib/eal/include/generic/rte_spinlock.h | 1 + lib/eal/include/meson.build | 1 + lib/eal/include/rte_mcslock.h | 1 + lib/eal/include/rte_pflock.h | 1 + lib/eal/include/rte_seqcount.h | 1 + lib/eal/include/rte_stdatomic.h | 198 +++++++++++++++++++++++++++++++++ lib/eal/include/rte_ticketlock.h | 1 + lib/eal/include/rte_trace_point.h | 1 + meson_options.txt | 2 + 13 files changed, 211 insertions(+) create mode 100644 lib/eal/include/rte_stdatomic.h
config/meson.build +1 -0
@@ -303,6 +303,7 @@ endforeach # set other values pulled from the build options dpdk_conf.set('RTE_MAX_ETHPORTS', get_option('max_ethports')) dpdk_conf.set('RTE_LIBEAL_USE_HPET', get_option('use_hpet')) +dpdk_conf.set('RTE_ENABLE_STDATOMIC', get_option('enable_stdatomic')) dpdk_conf.set('RTE_ENABLE_TRACE_FP', get_option('enable_trace_fp')) # values which have defaults which may be overridden dpdk_conf.set('RTE_MAX_VFIO_GROUPS', 64)
lib/eal/include/generic/rte_atomic.h +1 -0
@@ -15,6 +15,7 @@#include <stdint.h> #include <rte_compat.h> #include <rte_common.h> +#include <rte_stdatomic.h> #ifdef __DOXYGEN__
lib/eal/include/generic/rte_pause.h +1 -0
@@ -16,6 +16,7 @@#include <assert.h> #include <rte_common.h> #include <rte_atomic.h> +#include <rte_stdatomic.h> /** * Pause CPU execution for a short while
lib/eal/include/generic/rte_rwlock.h +1 -0
@@ -32,6 +32,7 @@#include <rte_common.h> #include <rte_lock_annotations.h> #include <rte_pause.h> +#include <rte_stdatomic.h> /** * The rte_rwlock_t type.
lib/eal/include/generic/rte_spinlock.h +1 -0
@@ -23,6 +23,7 @@#endif #include <rte_lock_annotations.h> #include <rte_pause.h> +#include <rte_stdatomic.h> /** * The rte_spinlock_t type.
lib/eal/include/meson.build +1 -0
@@ -42,6 +42,7 @@ headers += files( 'rte_seqlock.h', 'rte_service.h', 'rte_service_component.h', + 'rte_stdatomic.h', 'rte_string_fns.h', 'rte_tailq.h', 'rte_thread.h',
lib/eal/include/rte_mcslock.h +1 -0
@@ -27,6 +27,7 @@#include <rte_common.h> #include <rte_pause.h> #include <rte_branch_prediction.h> +#include <rte_stdatomic.h> /** * The rte_mcslock_t type.
lib/eal/include/rte_pflock.h +1 -0
@@ -34,6 +34,7 @@#include <rte_compat.h> #include <rte_common.h> #include <rte_pause.h> +#include <rte_stdatomic.h> /** * The rte_pflock_t type.
lib/eal/include/rte_seqcount.h +1 -0
@@ -26,6 +26,7 @@#include <rte_atomic.h> #include <rte_branch_prediction.h> #include <rte_compat.h> +#include <rte_stdatomic.h> /** * The RTE seqcount type.
lib/eal/include/rte_stdatomic.h new file +198 -0
198 lines not on this page · raw message
lib/eal/include/rte_ticketlock.h +1 -0
@@ -24,6 +24,7 @@#include <rte_common.h> #include <rte_lcore.h> #include <rte_pause.h> +#include <rte_stdatomic.h> /** * The rte_ticketlock_t type.
lib/eal/include/rte_trace_point.h +1 -0
@@ -30,6 +30,7 @@#include <rte_per_lcore.h> #include <rte_string_fns.h> #include <rte_uuid.h> +#include <rte_stdatomic.h> /** The tracepoint object. */ typedef uint64_t rte_trace_point_t ;
meson_options.txt +2 -0
@@ -46,6 +46,8 @@ option('mbuf_refcnt_atomic', type: 'boolean', value: true, description: 'Atomically access the mbuf refcnt.') option('platform', type: 'string', value: 'native', description: 'Platform to build, either "native", "generic" or a SoC. Please refer to the Linux build guide for more information.') +option('enable_stdatomic', type: 'boolean', value: false, description: + 'enable use of C11 stdatomic') option('enable_trace_fp', type: 'boolean', value: false, description: 'enable fast path trace points.') option('tests', type: 'boolean', value: true, description:
Signature · 1 line
1.8.3.1
RE: [PATCH v4 1/6] eal: provide rte stdatomics optional atomics API
Quoted 10 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Wednesday, 16 August 2023 23.39Provide API for atomic operations in the rte namespace that may optionally be configured to use C11 atomics with meson option enable_stdatomics=true
Signed-off-by Tyler Retzlaff <[email protected]>Reviewed-by Morten Brørup <[email protected]>---
Speed blindness during my review... I have now spotted a couple of copy-paste typos:
lib/eal/include/rte_stdatomic.h line 86
Quoted 3 lines
lib/eal/include/rte_stdatomic.h new file +3
-0+#define rte_atomic_compare_exchange_weak_explicit ( \ + ptr, expected, desired, succ_memorder, fail_memorder) \ + atomic_compare_exchange_strong_explicit( \
atomic_compare_exchange_weak_explicit, not strong.
Quoted 2 lines
(quoted diff) +2
-0+ ptr, expected, desired, succ_memorder, fail_memorder) +
[...]
lib/eal/include/rte_stdatomic.h line 111
Quoted 2 lines
lib/eal/include/rte_stdatomic.h new file +2
-0+#define rte_atomic_flag_clear_explicit (ptr, memorder) \ + atomic_flag_clear(ptr, memorder)
atomic_flag_clear_explicit(ptr, memorder), missing _explicit.
Re: [PATCH v4 1/6] eal: provide rte stdatomics optional atomics API
Quoted 11 lines · Morten Brørup
Quoted 10 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Wednesday, 16 August 2023 23.39Provide API for atomic operations in the rte namespace that may optionally be configured to use C11 atomics with meson option enable_stdatomics=true
Signed-off-by Tyler Retzlaff <[email protected]>Reviewed-by Morten Brørup <[email protected]>---
Quoted 7 lines
Speed blindness during my review... I have now spotted a couple of copy-paste typos:
lib/eal/include/rte_stdatomic.h line 86
Quoted 3 lines
lib/eal/include/rte_stdatomic.h new file +3
-0+#define rte_atomic_compare_exchange_weak_explicit ( \ + ptr, expected, desired, succ_memorder, fail_memorder) \ + atomic_compare_exchange_strong_explicit( \
atomic_compare_exchange_weak_explicit, not strong.
yikes, thanks for catching that cut & past error
Quoted 10 lines
Quoted 3 lines
(quoted diff) +2
-0+ ptr, expected, desired, succ_memorder, fail_memorder) +[...]
lib/eal/include/rte_stdatomic.h line 111
Quoted 2 lines
lib/eal/include/rte_stdatomic.h new file +2
-0+#define rte_atomic_flag_clear_explicit (ptr, memorder) \ + atomic_flag_clear(ptr, memorder)
atomic_flag_clear_explicit(ptr, memorder), missing_explicit.
yes, currently unused otherwise it would have failed to compie
i'll correct this too.
thank you for the careful review i look at the diffs over and over and still it's hard to spot subtle swaps/exchanges of things.
Quoted 1 line
RE: [PATCH v4 1/6] eal: provide rte stdatomics optional atomics API
Quoted 32 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Thursday, 17 August 2023 21.10Quoted 7 lines · Morten Brørup
Quoted 7 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Wednesday, 16 August 2023 23.39Provide API for atomic operations in the rte namespace that may optionally be configured to use C11 atomics with meson option enable_stdatomics=true
Quoted 5 lines
Quoted 3 lines
Signed-off-by Tyler Retzlaff <[email protected]>Reviewed-by Morten Brørup <[email protected]>---
Speed blindness during my review... I have now spotted a couple of copy-
paste typos:
Quoted 6 lines
lib/eal/include/rte_stdatomic.h line 86
Quoted 4 lines
lib/eal/include/rte_stdatomic.h new file +3
-0+#define rte_atomic_compare_exchange_weak_explicit ( \ + ptr, expected, desired, succ_memorder, fail_memorder) \ + atomic_compare_exchange_strong_explicit( \
atomic_compare_exchange_weak_explicit, not strong.yikes, thanks for catching that cut & past error
Quoted 6 lines
Quoted 3 lines
(quoted diff) +2
-0+ ptr, expected, desired, succ_memorder, fail_memorder) +[...]
Quoted 6 lines
Quoted 4 lines
lib/eal/include/rte_stdatomic.h line 111
Quoted 2 lines
lib/eal/include/rte_stdatomic.h new file +2
-0+#define rte_atomic_flag_clear_explicit (ptr, memorder) \ + atomic_flag_clear(ptr, memorder)
atomic_flag_clear_explicit(ptr, memorder), missing_explicit.yes, currently unused otherwise it would have failed to compie
Yes, I guessed something similar, when I spotted this... two parameters being passed to a single-parameter function.
Quoted 5 lines
i'll correct this too.
thank you for the careful review i look at the diffs over and over and still it's hard to spot subtle swaps/exchanges of things.
Yes, when reviewing many similar lines of code, the probability of overlooking something increases rapidly, even for external reviewers.
About 20 years ago, a Danish consulting company realized this, and turned it to something positive. They set up a highly specialized organization, Specialisterne (https://specialisterne.com/), to offer autistic consultants for tasks like this. These autistics enjoy repetitive tasks, and are excellent at spotting subtle differences normal people would likely overlook.
PATCH v4 2/6 eal: adapt EAL to present rte optional atomics API
In reply to Tyler RetzlaffAdapt the EAL public headers to use rte optional atomics API instead of directly using and exposing toolchain specific atomic builtin intrinsics.
--- app/test/test_mcslock.c | 6 ++-- lib/eal/arm/include/rte_atomic_32.h | 4 +-- lib/eal/arm/include/rte_atomic_64.h | 36 +++++++++++------------ lib/eal/arm/include/rte_pause_64.h | 26 ++++++++-------- lib/eal/arm/rte_power_intrinsics.c | 8 ++--- lib/eal/common/eal_common_trace.c | 16 +++++----- lib/eal/include/generic/rte_atomic.h | 50 +++++++++++++++---------------- lib/eal/include/generic/rte_pause.h | 46 ++++++++++++----------------- lib/eal/include/generic/rte_rwlock.h | 47 +++++++++++++++-------------- lib/eal/include/generic/rte_spinlock.h | 19 ++++++------ lib/eal/include/rte_mcslock.h | 50 +++++++++++++++---------------- lib/eal/include/rte_pflock.h | 24 ++++++++------- lib/eal/include/rte_seqcount.h | 18 ++++++------ lib/eal/include/rte_ticketlock.h | 42 +++++++++++++------------- lib/eal/include/rte_trace_point.h | 4 +-- lib/eal/loongarch/include/rte_atomic.h | 4 +-- lib/eal/ppc/include/rte_atomic.h | 54 +++++++++++++++++----------------- lib/eal/riscv/include/rte_atomic.h | 4 +-- lib/eal/x86/include/rte_atomic.h | 8 ++--- lib/eal/x86/include/rte_spinlock.h | 2 +- lib/eal/x86/rte_power_intrinsics.c | 6 ++-- 21 files changed, 237 insertions(+), 237 deletions(-)
app/test/test_mcslock.c +3 -3
@@ -36,9 +36,9 @@* lock multiple times. */ -rte_mcslock_t *p_ml; -rte_mcslock_t *p_ml_try; -rte_mcslock_t *p_ml_perf; +RTE_ATOMIC (rte_mcslock_t *) p_ml; +RTE_ATOMIC (rte_mcslock_t *) p_ml_try; +RTE_ATOMIC (rte_mcslock_t *) p_ml_perf;static unsigned int count;
lib/eal/arm/include/rte_atomic_32.h +2 -2
@@ -34,9 +34,9 @@#define rte_io_rmb () rte_rmb() static __rte_always_inline void -rte_atomic_thread_fence (int memorder) +rte_atomic_thread_fence (rte_memory_order memorder) { -__atomic_thread_fence (memorder); +__rte_atomic_thread_fence (memorder); }#ifdef __cplusplus
lib/eal/arm/include/rte_atomic_64.h +18 -18
@@ -38,9 +38,9 @@#define rte_io_rmb () rte_rmb() static __rte_always_inline void -rte_atomic_thread_fence (int memorder) +rte_atomic_thread_fence (rte_memory_order memorder) { -__atomic_thread_fence (memorder); +__rte_atomic_thread_fence (memorder); }/*------------------------ 128 bit atomic operations -------------------------*/
@@ -107,33 +107,33 @@*/ RTE_SET_USED (failure);/* Find invalid memory order */ -RTE_ASSERT (success == __ATOMIC_RELAXED || - success == __ATOMIC_ACQUIRE || - success == __ATOMIC_RELEASE || - success == __ATOMIC_ACQ_REL || - success == __ATOMIC_SEQ_CST); +RTE_ASSERT (success == rte_memory_order_relaxed || + success == rte_memory_order_acquire || + success == rte_memory_order_release || + success == rte_memory_order_acq_rel || + success == rte_memory_order_seq_cst);rte_int128_t expected = *exp;rte_int128_t desired = *src;rte_int128_t old;#if defined(__ARM_FEATURE_ATOMICS) || defined(RTE_ARM_FEATURE_ATOMICS ) -if (success == __ATOMIC_RELAXED) +if (success == rte_memory_order_relaxed)__cas_128_relaxed (dst, exp, desired); -else if (success == __ATOMIC_ACQUIRE) +else if (success == rte_memory_order_acquire)__cas_128_acquire (dst, exp, desired); -else if (success == __ATOMIC_RELEASE) +else if (success == rte_memory_order_release)__cas_128_release (dst, exp, desired);else __cas_128_acq_rel (dst, exp, desired); old = *exp;#else -#define __HAS_ACQ (mo) ((mo) != __ATOMIC_RELAXED && (mo) != __ATOMIC_RELEASE) -#define __HAS_RLS (mo) ((mo) == __ATOMIC_RELEASE || (mo) == __ATOMIC_ACQ_REL || \ - (mo) == __ATOMIC_SEQ_CST) +#define __HAS_ACQ (mo) ((mo) != rte_memory_order_relaxed && (mo) != rte_memory_order_release) +#define __HAS_RLS (mo) ((mo) == rte_memory_order_release || (mo) == rte_memory_order_acq_rel || \ + (mo) == rte_memory_order_seq_cst) -int ldx_mo =__HAS_ACQ (success) ? __ATOMIC_ACQUIRE : __ATOMIC_RELAXED; -int stx_mo =__HAS_RLS (success) ? __ATOMIC_RELEASE : __ATOMIC_RELAXED; +int ldx_mo =__HAS_ACQ (success) ? rte_memory_order_acquire : rte_memory_order_relaxed; +int stx_mo =__HAS_RLS (success) ? rte_memory_order_release : rte_memory_order_relaxed;#undef __HAS_ACQ#undef __HAS_RLS
@@ -153,7 +153,7 @@ : "Q" (src->val[0]) \ : "memory"); } -if (ldx_mo == __ATOMIC_RELAXED) +if (ldx_mo == rte_memory_order_relaxed) __LOAD_128("ldxp", dst, old) else__LOAD_128 ("ldaxp" , dst, old)
@@ -170,7 +170,7 @@ : "memory"); }if (likely (old.int128 == expected.int128)) { -if (stx_mo == __ATOMIC_RELAXED) +if (stx_mo == rte_memory_order_relaxed) __STORE_128("stxp", dst, desired, ret) else__STORE_128 ("stlxp" , dst, desired, ret)
@@ -181,7 +181,7 @@* needs to be stored back to ensure it was read * atomically. */ -if (stx_mo == __ATOMIC_RELAXED) +if (stx_mo == rte_memory_order_relaxed) __STORE_128("stxp", dst, old, ret) else__STORE_128 ("stlxp" , dst, old, ret)
lib/eal/arm/include/rte_pause_64.h +13 -13
@@ -41,7 +41,7 @@ static inline void rte_pause(void)* implicitly to exit WFE. */ #define __RTE_ARM_LOAD_EXC_8 (src, dst, memorder) { \ - if (memorder == __ATOMIC_RELAXED) { \ + if (memorder == rte_memory_order_relaxed) { \ asm volatile("ldxrb %w[tmp], [%x[addr]]" \ : [tmp] "=&r" (dst) \ : [addr] "r" (src) \
@@ -60,7 +60,7 @@ static inline void rte_pause(void)* implicitly to exit WFE. */ #define __RTE_ARM_LOAD_EXC_16 (src, dst, memorder) { \ - if (memorder == __ATOMIC_RELAXED) { \ + if (memorder == rte_memory_order_relaxed) { \ asm volatile("ldxrh %w[tmp], [%x[addr]]" \ : [tmp] "=&r" (dst) \ : [addr] "r" (src) \
@@ -79,7 +79,7 @@ static inline void rte_pause(void)* implicitly to exit WFE. */ #define __RTE_ARM_LOAD_EXC_32 (src, dst, memorder) { \ - if (memorder == __ATOMIC_RELAXED) { \ + if (memorder == rte_memory_order_relaxed) { \ asm volatile("ldxr %w[tmp], [%x[addr]]" \ : [tmp] "=&r" (dst) \ : [addr] "r" (src) \
@@ -98,7 +98,7 @@ static inline void rte_pause(void)* implicitly to exit WFE. */ #define __RTE_ARM_LOAD_EXC_64 (src, dst, memorder) { \ - if (memorder == __ATOMIC_RELAXED) { \ + if (memorder == rte_memory_order_relaxed) { \ asm volatile("ldxr %x[tmp], [%x[addr]]" \ : [tmp] "=&r" (dst) \ : [addr] "r" (src) \
@@ -118,7 +118,7 @@ static inline void rte_pause(void)*/ #define __RTE_ARM_LOAD_EXC_128 (src, dst, memorder) { \ volatile rte_int128_t *dst_128 = (volatile rte_int128_t *)&dst; \ - if (memorder == __ATOMIC_RELAXED) { \ + if (memorder == rte_memory_order_relaxed) { \ asm volatile("ldxp %x[tmp0], %x[tmp1], [%x[addr]]" \ : [tmp0] "=&r" (dst_128->val[0]), \ [tmp1] "=&r" (dst_128->val[1]) \
@@ -153,8 +153,8 @@ static inline void rte_pause(void) {uint16_t value; -RTE_BUILD_BUG_ON (memorder != __ATOMIC_ACQUIRE && - memorder != __ATOMIC_RELAXED); +RTE_BUILD_BUG_ON (memorder != rte_memory_order_acquire && + memorder != rte_memory_order_relaxed); __RTE_ARM_LOAD_EXC_16(addr, value, memorder)if (value != expected) {
@@ -172,8 +172,8 @@ static inline void rte_pause(void) {uint32_t value; -RTE_BUILD_BUG_ON (memorder != __ATOMIC_ACQUIRE && - memorder != __ATOMIC_RELAXED); +RTE_BUILD_BUG_ON (memorder != rte_memory_order_acquire && + memorder != rte_memory_order_relaxed); __RTE_ARM_LOAD_EXC_32(addr, value, memorder)if (value != expected) {
@@ -191,8 +191,8 @@ static inline void rte_pause(void) {uint64_t value; -RTE_BUILD_BUG_ON (memorder != __ATOMIC_ACQUIRE && - memorder != __ATOMIC_RELAXED); +RTE_BUILD_BUG_ON (memorder != rte_memory_order_acquire && + memorder != rte_memory_order_relaxed); __RTE_ARM_LOAD_EXC_64(addr, value, memorder)if (value != expected) {
@@ -206,8 +206,8 @@ static inline void rte_pause(void)#define RTE_WAIT_UNTIL_MASKED (addr, mask, cond, expected, memorder) do { \ RTE_BUILD_BUG_ON(!__builtin_constant_p(memorder)); \ - RTE_BUILD_BUG_ON(memorder != __ATOMIC_ACQUIRE && \ - memorder != __ATOMIC_RELAXED); \ + RTE_BUILD_BUG_ON(memorder != rte_memory_order_acquire && \ + memorder != rte_memory_order_relaxed); \ const uint32_t size = sizeof(*(addr)) << 3; \ typeof(*(addr)) expected_value = (expected); \ typeof(*(addr)) value; \
lib/eal/arm/rte_power_intrinsics.c +4 -4
@@ -33,19 +33,19 @@switch (pmc->size) {case sizeof (uint8_t ): - __RTE_ARM_LOAD_EXC_8(pmc->addr, cur_value, __ATOMIC_RELAXED) + __RTE_ARM_LOAD_EXC_8(pmc->addr, cur_value, rte_memory_order_relaxed)__RTE_ARM_WFE () break;case sizeof (uint16_t ): - __RTE_ARM_LOAD_EXC_16(pmc->addr, cur_value, __ATOMIC_RELAXED) + __RTE_ARM_LOAD_EXC_16(pmc->addr, cur_value, rte_memory_order_relaxed)__RTE_ARM_WFE () break;case sizeof (uint32_t ): - __RTE_ARM_LOAD_EXC_32(pmc->addr, cur_value, __ATOMIC_RELAXED) + __RTE_ARM_LOAD_EXC_32(pmc->addr, cur_value, rte_memory_order_relaxed)__RTE_ARM_WFE () break;case sizeof (uint64_t ): - __RTE_ARM_LOAD_EXC_64(pmc->addr, cur_value, __ATOMIC_RELAXED) + __RTE_ARM_LOAD_EXC_64(pmc->addr, cur_value, rte_memory_order_relaxed)__RTE_ARM_WFE () break;default :
lib/eal/common/eal_common_trace.c +9 -7
@@ -103,11 +103,11 @@ struct trace_point_head * trace_mode_set(rte_trace_point_t *t, enum rte_trace_mode mode) {if (mode ==RTE_TRACE_MODE_OVERWRITE ) -__atomic_fetch_and (t, ~__RTE_TRACE_FIELD_ENABLE_DISCARD, - __ATOMIC_RELEASE); +rte_atomic_fetch_and_explicit (t, ~__RTE_TRACE_FIELD_ENABLE_DISCARD, + rte_memory_order_release);else -__atomic_fetch_or (t, __RTE_TRACE_FIELD_ENABLE_DISCARD, - __ATOMIC_RELEASE); +rte_atomic_fetch_or_explicit (t, __RTE_TRACE_FIELD_ENABLE_DISCARD, + rte_memory_order_release); }void
@@ -141,7 +141,7 @@ rte_trace_mode rte_trace_mode_get(void)if (trace_point_is_invalid (t))return false; - val =__atomic_load_n (t, __ATOMIC_ACQUIRE); + val =rte_atomic_load_explicit (t, rte_memory_order_acquire);return (val & __RTE_TRACE_FIELD_ENABLE_MASK) !=0 ; }
@@ -153,7 +153,8 @@ rte_trace_mode rte_trace_mode_get(void)if (trace_point_is_invalid (t))return -ERANGE ; - prev =__atomic_fetch_or (t, __RTE_TRACE_FIELD_ENABLE_MASK, __ATOMIC_RELEASE); + prev =rte_atomic_fetch_or_explicit (t, __RTE_TRACE_FIELD_ENABLE_MASK, + rte_memory_order_release);if ((prev & __RTE_TRACE_FIELD_ENABLE_MASK) ==0 )__atomic_fetch_add (&trace.status,1 , __ATOMIC_RELEASE);return 0 ;
@@ -167,7 +168,8 @@ rte_trace_mode rte_trace_mode_get(void)if (trace_point_is_invalid (t))return -ERANGE ; - prev =__atomic_fetch_and (t, ~__RTE_TRACE_FIELD_ENABLE_MASK, __ATOMIC_RELEASE); + prev =rte_atomic_fetch_and_explicit (t, ~__RTE_TRACE_FIELD_ENABLE_MASK, + rte_memory_order_release);if ((prev & __RTE_TRACE_FIELD_ENABLE_MASK) !=0 )__atomic_fetch_sub (&trace.status,1 , __ATOMIC_RELEASE);return 0 ;
lib/eal/include/generic/rte_atomic.h +25 -25
194 lines not on this page · raw message
lib/eal/include/generic/rte_pause.h +19 -27
103 lines not on this page · raw message
lib/eal/include/generic/rte_rwlock.h +25 -22
126 lines not on this page · raw message
lib/eal/include/generic/rte_spinlock.h +9 -10
@@ -29,7 +29,7 @@* The rte_spinlock_t type. */ typedef struct __rte_lockable { -volatile int locked;/**< lock status 0 = unlocked, 1 = locked */ +volatile RTE_ATOMIC (int ) locked;/**< lock status 0 = unlocked, 1 = locked */ }rte_spinlock_t ;/**
@@ -66,10 +66,10 @@ {int exp =0 ; -while (!__atomic_compare_exchange_n (&sl->locked, &exp,1 ,0 , - __ATOMIC_ACQUIRE, __ATOMIC_RELAXED)) { -rte_wait_until_equal_32 ((volatile uint32_t *)&sl->locked, -0 , __ATOMIC_RELAXED); +while (!rte_atomic_compare_exchange_strong_explicit (&sl->locked, &exp,1 , + rte_memory_order_acquire, rte_memory_order_relaxed)) { +rte_wait_until_equal_32 ((volatile uint32_t *)(uintptr_t )&sl->locked, +0 , rte_memory_order_relaxed); exp =0 ; } }
@@ -90,7 +90,7 @@ rte_spinlock_unlock(rte_spinlock_t *sl) __rte_no_thread_safety_analysis { -__atomic_store_n (&sl->locked,0 , __ATOMIC_RELEASE); +rte_atomic_store_explicit (&sl->locked,0 , rte_memory_order_release); }#endif
@@ -113,9 +113,8 @@ __rte_no_thread_safety_analysis {int exp =0 ; -return __atomic_compare_exchange_n (&sl->locked, &exp,1 , -0 ,/* disallow spurious failure */ - __ATOMIC_ACQUIRE, __ATOMIC_RELAXED); +return rte_atomic_compare_exchange_strong_explicit (&sl->locked, &exp,1 , + rte_memory_order_acquire, rte_memory_order_relaxed); }#endif
@@ -129,7 +128,7 @@*/ static inline int rte_spinlock_is_locked (rte_spinlock_t *sl) { -return __atomic_load_n (&sl->locked, __ATOMIC_ACQUIRE); +return rte_atomic_load_explicit (&sl->locked, rte_memory_order_acquire); }/**
lib/eal/include/rte_mcslock.h +25 -25
133 lines not on this page · raw message
lib/eal/include/rte_pflock.h +13 -11
@@ -41,8 +41,8 @@*/ struct rte_pflock {struct { -uint16_t in; -uint16_t out; +RTE_ATOMIC (uint16_t ) in; +RTE_ATOMIC (uint16_t ) out; } rd, wr; };typedef struct rte_pflock rte_pflock_t ;
@@ -117,14 +117,14 @@ struct rte_pflock {* If no writer is present, then the operation has completed * successfully. */ - w =__atomic_fetch_add (&pf->rd.in,RTE_PFLOCK_RINC , __ATOMIC_ACQUIRE) + w =rte_atomic_fetch_add_explicit (&pf->rd.in,RTE_PFLOCK_RINC , rte_memory_order_acquire) &RTE_PFLOCK_WBITS ;if (w ==0 )return ;/* Wait for current write phase to complete. */ RTE_WAIT_UNTIL_MASKED(&pf->rd.in, RTE_PFLOCK_WBITS, !=, w, - __ATOMIC_ACQUIRE); + rte_memory_order_acquire); }/**
@@ -140,7 +140,7 @@ struct rte_pflock {static inline void rte_pflock_read_unlock (rte_pflock_t *pf) { -__atomic_fetch_add (&pf->rd.out,RTE_PFLOCK_RINC , __ATOMIC_RELEASE); +rte_atomic_fetch_add_explicit (&pf->rd.out,RTE_PFLOCK_RINC , rte_memory_order_release); }/**
@@ -161,8 +161,9 @@ struct rte_pflock {/* Acquire ownership of write-phase. * This is same as rte_ticketlock_lock(). */ - ticket =__atomic_fetch_add (&pf->wr.in,1 , __ATOMIC_RELAXED); -rte_wait_until_equal_16 (&pf->wr.out, ticket, __ATOMIC_ACQUIRE); + ticket =rte_atomic_fetch_add_explicit (&pf->wr.in,1 , rte_memory_order_relaxed); +rte_wait_until_equal_16 ((uint16_t *)(uintptr_t )&pf->wr.out, ticket, + rte_memory_order_acquire);/* * Acquire ticket on read-side in order to allow them
@@ -173,10 +174,11 @@ struct rte_pflock {* speculatively. */ w =RTE_PFLOCK_PRES | (ticket &RTE_PFLOCK_PHID ); - ticket =__atomic_fetch_add (&pf->rd.in, w, __ATOMIC_RELAXED); + ticket =rte_atomic_fetch_add_explicit (&pf->rd.in, w, rte_memory_order_relaxed);/* Wait for any pending readers to flush. */ -rte_wait_until_equal_16 (&pf->rd.out, ticket, __ATOMIC_ACQUIRE); +rte_wait_until_equal_16 ((uint16_t *)(uintptr_t )&pf->rd.out, ticket, + rte_memory_order_acquire); }/**
@@ -193,10 +195,10 @@ struct rte_pflock { rte_pflock_write_unlock(rte_pflock_t *pf) {/* Migrate from write phase to read phase. */ -__atomic_fetch_and (&pf->rd.in,RTE_PFLOCK_LSB , __ATOMIC_RELEASE); +rte_atomic_fetch_and_explicit (&pf->rd.in,RTE_PFLOCK_LSB , rte_memory_order_release);/* Allow other writers to continue. */ -__atomic_fetch_add (&pf->wr.out,1 , __ATOMIC_RELEASE); +rte_atomic_fetch_add_explicit (&pf->wr.out,1 , rte_memory_order_release); }#ifdef __cplusplus
lib/eal/include/rte_seqcount.h +9 -9
@@ -32,7 +32,7 @@* The RTE seqcount type. */ typedef struct { -uint32_t sn;/**< A sequence number for the protected data. */ +RTE_ATOMIC (uint32_t ) sn;/**< A sequence number for the protected data. */ }rte_seqcount_t ;/**
@@ -106,11 +106,11 @@static inline uint32_t rte_seqcount_read_begin (const rte_seqcount_t *seqcount) { -/* __ATOMIC_ACQUIRE to prevent loads after (in program order) +/* rte_memory_order_acquire to prevent loads after (in program order) * from happening before the sn load. Synchronizes-with the * store release in rte_seqcount_write_end(). */ -return __atomic_load_n (&seqcount->sn, __ATOMIC_ACQUIRE); +return rte_atomic_load_explicit (&seqcount->sn, rte_memory_order_acquire); }/**
@@ -161,9 +161,9 @@return true;/* make sure the data loads happens before the sn load */ -rte_atomic_thread_fence (__ATOMIC_ACQUIRE); +rte_atomic_thread_fence (rte_memory_order_acquire); - end_sn =__atomic_load_n (&seqcount->sn, __ATOMIC_RELAXED); + end_sn =rte_atomic_load_explicit (&seqcount->sn, rte_memory_order_relaxed);/* A writer incremented the sequence number during this read * critical section.
@@ -205,12 +205,12 @@ sn = seqcount->sn +1 ; -__atomic_store_n (&seqcount->sn, sn, __ATOMIC_RELAXED); +rte_atomic_store_explicit (&seqcount->sn, sn, rte_memory_order_relaxed); -/* __ATOMIC_RELEASE to prevent stores after (in program order) +/* rte_memory_order_release to prevent stores after (in program order) * from happening before the sn store. */ -rte_atomic_thread_fence (__ATOMIC_RELEASE); +rte_atomic_thread_fence (rte_memory_order_release); }/**
@@ -237,7 +237,7 @@ sn = seqcount->sn +1 ;/* Synchronizes-with the load acquire in rte_seqcount_read_begin(). */ -__atomic_store_n (&seqcount->sn, sn, __ATOMIC_RELEASE); +rte_atomic_store_explicit (&seqcount->sn, sn, rte_memory_order_release); }#ifdef __cplusplus
lib/eal/include/rte_ticketlock.h +22 -20
115 lines not on this page · raw message
lib/eal/include/rte_trace_point.h +2 -2
@@ -33,7 +33,7 @@#include <rte_stdatomic.h> /** The tracepoint object. */ -typedef uint64_t rte_trace_point_t ; +typedef RTE_ATOMIC (uint64_t )rte_trace_point_t ;/** * Macro to define the tracepoint arguments in RTE_TRACE_POINT macro.
@@ -359,7 +359,7 @@ struct __rte_trace_header {#define __rte_trace_point_emit_header_generic (t) \ void *mem; \ do { \ - const uint64_t val = __atomic_load_n(t, __ATOMIC_ACQUIRE); \ + const uint64_t val = rte_atomic_load_explicit(t, rte_memory_order_acquire); \ if (likely(!(val & __RTE_TRACE_FIELD_ENABLE_MASK))) \ return; \ mem = __rte_trace_mem_get(val); \
lib/eal/loongarch/include/rte_atomic.h +2 -2
@@ -35,9 +35,9 @@#define rte_io_rmb () rte_mb() static __rte_always_inline void -rte_atomic_thread_fence (int memorder) +rte_atomic_thread_fence (rte_memory_order memorder) { -__atomic_thread_fence (memorder); +__rte_atomic_thread_fence (memorder); }#ifdef __cplusplus
lib/eal/ppc/include/rte_atomic.h +27 -27
172 lines not on this page · raw message
lib/eal/riscv/include/rte_atomic.h +2 -2
@@ -40,9 +40,9 @@#define rte_io_rmb () asm volatile("fence ir, ir" : : : "memory") static __rte_always_inline void -rte_atomic_thread_fence (int memorder) +rte_atomic_thread_fence (rte_memory_order memorder) { -__atomic_thread_fence (memorder); +__rte_atomic_thread_fence (memorder); }#ifdef __cplusplus
lib/eal/x86/include/rte_atomic.h +4 -4
@@ -82,17 +82,17 @@/** * Synchronization fence between threads based on the specified memory order. * -* On x86 the __atomic_thread_fence(__ATOMIC_SEQ_CST) generates full 'mfence' +* On x86 the __rte_atomic_thread_fence(rte_memory_order_seq_cst) generates full 'mfence' * which is quite expensive. The optimized implementation of rte_smp_mb is * used instead. */ static __rte_always_inline void -rte_atomic_thread_fence (int memorder) +rte_atomic_thread_fence (rte_memory_order memorder) { -if (memorder == __ATOMIC_SEQ_CST) +if (memorder == rte_memory_order_seq_cst)rte_smp_mb ();else -__atomic_thread_fence (memorder); +__rte_atomic_thread_fence (memorder); }/*------------------------- 16 bit atomic operations -------------------------*/
lib/eal/x86/include/rte_spinlock.h +1 -1
@@ -78,7 +78,7 @@ static inline int rte_tm_supported(void) }static inline int -rte_try_tm (volatile int *lock) +rte_try_tm (volatile RTE_ATOMIC (int ) *lock) {int i, retries;
lib/eal/x86/rte_power_intrinsics.c +3 -3
@@ -23,9 +23,9 @@uint64_t val;/* trigger a write but don't change the value */ - val =__atomic_load_n ((volatile uint64_t *)addr, __ATOMIC_RELAXED); -__atomic_compare_exchange_n ((volatile uint64_t *)addr, &val, val,0 , - __ATOMIC_RELAXED, __ATOMIC_RELAXED); + val =rte_atomic_load_explicit ((volatile uint64_t *)addr, rte_memory_order_relaxed); +rte_atomic_compare_exchange_strong_explicit ((volatile uint64_t *)addr, &val, val, + rte_memory_order_relaxed, rte_memory_order_relaxed); }static bool wait_supported;
Signature · 1 line
1.8.3.1
PATCH v4 3/6 eal: add rte atomic qualifier with casts
In reply to Tyler RetzlaffIntroduce __rte_atomic qualifying casts in rte_optional atomics inline functions to prevent cascading the need to pass __rte_atomic qualified arguments.
Warning, this is really implementation dependent and being done temporarily to avoid having to convert more of the libraries and tests in DPDK in the initial series that introduces the API. The consequence of the assumption of the ABI of the types in question not being ``the same'' is only a risk that may be realized when enable_stdatomic=true.
--- lib/eal/include/generic/rte_atomic.h | 48 ++++++++++++++++++++++++------------ lib/eal/include/generic/rte_pause.h | 9 ++++--- lib/eal/x86/rte_power_intrinsics.c | 7 +++--- 3 files changed, 42 insertions(+), 22 deletions(-)
lib/eal/include/generic/rte_atomic.h +32 -16
@@ -274,7 +274,8 @@static inline void rte_atomic16_add (rte_atomic16_t *v,int16_t inc) { -rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst); + rte_atomic_fetch_add_explicit((volatile __rte_atomic int16_t *)&v->cnt, inc, + rte_memory_order_seq_cst); }/**
@@ -288,7 +289,8 @@static inline void rte_atomic16_sub (rte_atomic16_t *v,int16_t dec) { -rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst); + rte_atomic_fetch_sub_explicit((volatile __rte_atomic int16_t *)&v->cnt, dec, + rte_memory_order_seq_cst); }/**
@@ -341,7 +343,8 @@static inline int16_t rte_atomic16_add_return (rte_atomic16_t *v,int16_t inc) { -return rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst) + inc; + return rte_atomic_fetch_add_explicit((volatile __rte_atomic int16_t *)&v->cnt, inc, + rte_memory_order_seq_cst) + inc; }/**
@@ -361,7 +364,8 @@static inline int16_t rte_atomic16_sub_return (rte_atomic16_t *v,int16_t dec) { -return rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst) - dec; + return rte_atomic_fetch_sub_explicit((volatile __rte_atomic int16_t *)&v->cnt, dec, + rte_memory_order_seq_cst) - dec; }/**
@@ -380,7 +384,8 @@#ifdef RTE_FORCE_INTRINSICS static inline int rte_atomic16_inc_and_test (rte_atomic16_t *v) { -return rte_atomic_fetch_add_explicit (&v->cnt,1 , rte_memory_order_seq_cst) +1 ==0 ; + return rte_atomic_fetch_add_explicit((volatile __rte_atomic int16_t *)&v->cnt, 1, + rte_memory_order_seq_cst) +1 ==0 ; }#endif
@@ -400,7 +405,8 @@ static inline int rte_atomic16_inc_and_test(rte_atomic16_t *v)#ifdef RTE_FORCE_INTRINSICS static inline int rte_atomic16_dec_and_test (rte_atomic16_t *v) { -return rte_atomic_fetch_sub_explicit (&v->cnt,1 , rte_memory_order_seq_cst) -1 ==0 ; + return rte_atomic_fetch_sub_explicit((volatile __rte_atomic int16_t *)&v->cnt, 1, + rte_memory_order_seq_cst) -1 ==0 ; }#endif
@@ -553,7 +559,8 @@ static inline void rte_atomic16_clear(rte_atomic16_t *v)static inline void rte_atomic32_add (rte_atomic32_t *v,int32_t inc) { -rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst); + rte_atomic_fetch_add_explicit((volatile __rte_atomic int32_t *)&v->cnt, inc, + rte_memory_order_seq_cst); }/**
@@ -567,7 +574,8 @@ static inline void rte_atomic16_clear(rte_atomic16_t *v)static inline void rte_atomic32_sub (rte_atomic32_t *v,int32_t dec) { -rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst); + rte_atomic_fetch_sub_explicit((volatile __rte_atomic int32_t *)&v->cnt, dec, + rte_memory_order_seq_cst); }/**
@@ -620,7 +628,8 @@ static inline void rte_atomic16_clear(rte_atomic16_t *v)static inline int32_t rte_atomic32_add_return (rte_atomic32_t *v,int32_t inc) { -return rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst) + inc; + return rte_atomic_fetch_add_explicit((volatile __rte_atomic int32_t *)&v->cnt, inc, + rte_memory_order_seq_cst) + inc; }/**
@@ -640,7 +649,8 @@ static inline void rte_atomic16_clear(rte_atomic16_t *v)static inline int32_t rte_atomic32_sub_return (rte_atomic32_t *v,int32_t dec) { -return rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst) - dec; + return rte_atomic_fetch_sub_explicit((volatile __rte_atomic int32_t *)&v->cnt, dec, + rte_memory_order_seq_cst) - dec; }/**
@@ -659,7 +669,8 @@ static inline void rte_atomic16_clear(rte_atomic16_t *v)#ifdef RTE_FORCE_INTRINSICS static inline int rte_atomic32_inc_and_test (rte_atomic32_t *v) { -return rte_atomic_fetch_add_explicit (&v->cnt,1 , rte_memory_order_seq_cst) +1 ==0 ; + return rte_atomic_fetch_add_explicit((volatile __rte_atomic int32_t *)&v->cnt, 1, + rte_memory_order_seq_cst) +1 ==0 ; }#endif
@@ -679,7 +690,8 @@ static inline int rte_atomic32_inc_and_test(rte_atomic32_t *v)#ifdef RTE_FORCE_INTRINSICS static inline int rte_atomic32_dec_and_test (rte_atomic32_t *v) { -return rte_atomic_fetch_sub_explicit (&v->cnt,1 , rte_memory_order_seq_cst) -1 ==0 ; + return rte_atomic_fetch_sub_explicit((volatile __rte_atomic int32_t *)&v->cnt, 1, + rte_memory_order_seq_cst) -1 ==0 ; }#endif
@@ -885,7 +897,8 @@ static inline void rte_atomic32_clear(rte_atomic32_t *v)static inline void rte_atomic64_add (rte_atomic64_t *v,int64_t inc) { -rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst); + rte_atomic_fetch_add_explicit((volatile __rte_atomic int64_t *)&v->cnt, inc, + rte_memory_order_seq_cst); }#endif
@@ -904,7 +917,8 @@ static inline void rte_atomic32_clear(rte_atomic32_t *v)static inline void rte_atomic64_sub (rte_atomic64_t *v,int64_t dec) { -rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst); + rte_atomic_fetch_sub_explicit((volatile __rte_atomic int64_t *)&v->cnt, dec, + rte_memory_order_seq_cst); }#endif
@@ -962,7 +976,8 @@ static inline void rte_atomic32_clear(rte_atomic32_t *v)static inline int64_t rte_atomic64_add_return (rte_atomic64_t *v,int64_t inc) { -return rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst) + inc; + return rte_atomic_fetch_add_explicit((volatile __rte_atomic int64_t *)&v->cnt, inc, + rte_memory_order_seq_cst) + inc; }#endif
@@ -986,7 +1001,8 @@ static inline void rte_atomic32_clear(rte_atomic32_t *v)static inline int64_t rte_atomic64_sub_return (rte_atomic64_t *v,int64_t dec) { -return rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst) - dec; + return rte_atomic_fetch_sub_explicit((volatile __rte_atomic int64_t *)&v->cnt, dec, + rte_memory_order_seq_cst) - dec; }#endif
lib/eal/include/generic/rte_pause.h +6 -3
@@ -81,7 +81,8 @@ {assert (memorder == rte_memory_order_acquire || memorder == rte_memory_order_relaxed); -while (rte_atomic_load_explicit (addr, memorder) != expected) + while (rte_atomic_load_explicit((volatile __rte_atomic uint16_t *)addr, memorder) + != expected)rte_pause (); }
@@ -91,7 +92,8 @@ {assert (memorder == rte_memory_order_acquire || memorder == rte_memory_order_relaxed); -while (rte_atomic_load_explicit (addr, memorder) != expected) + while (rte_atomic_load_explicit((volatile __rte_atomic uint32_t *)addr, memorder) + != expected)rte_pause (); }
@@ -101,7 +103,8 @@ {assert (memorder == rte_memory_order_acquire || memorder == rte_memory_order_relaxed); -while (rte_atomic_load_explicit (addr, memorder) != expected) + while (rte_atomic_load_explicit((volatile __rte_atomic uint64_t *)addr, memorder) + != expected)rte_pause (); }
lib/eal/x86/rte_power_intrinsics.c +4 -3
@@ -23,9 +23,10 @@uint64_t val;/* trigger a write but don't change the value */ - val =rte_atomic_load_explicit ((volatile uint64_t *)addr, rte_memory_order_relaxed); -rte_atomic_compare_exchange_strong_explicit ((volatile uint64_t *)addr, &val, val, - rte_memory_order_relaxed, rte_memory_order_relaxed); + val = rte_atomic_load_explicit((volatile __rte_atomic uint64_t *)addr, + rte_memory_order_relaxed); + rte_atomic_compare_exchange_strong_explicit((volatile __rte_atomic uint64_t *)addr, + &val, val, rte_memory_order_relaxed, rte_memory_order_relaxed); }static bool wait_supported;
Signature · 1 line
1.8.3.1
PATCH v4 4/6 distributor: adapt for EAL optional atomics API changes
In reply to Tyler RetzlaffAdapt distributor for EAL optional atomics API changes
--- lib/distributor/distributor_private.h | 2 +- lib/distributor/rte_distributor_single.c | 44 ++++++++++++++++---------------- 2 files changed, 23 insertions(+), 23 deletions(-)
lib/distributor/distributor_private.h +1 -1
@@ -52,7 +52,7 @@* Only 64-bits of the memory is actually used though. */ union rte_distributor_buffer_single { -volatile int64_t bufptr64; +volatile RTE_ATOMIC (int64_t ) bufptr64;char pad[RTE_CACHE_LINE_SIZE *3 ]; } __rte_cache_aligned;
lib/distributor/rte_distributor_single.c +22 -22
117 lines not on this page · raw message
Signature · 1 line
1.8.3.1
PATCH v4 5/6 bpf: adapt for EAL optional atomics API changes
In reply to Tyler RetzlaffAdapt bpf for EAL optional atomics API changes
--- lib/bpf/bpf_pkt.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)
lib/bpf/bpf_pkt.c +3 -3
@@ -25,7 +25,7 @@struct bpf_eth_cbi {/* used by both data & control path */ -uint32_t use;/*usage counter */ +RTE_ATOMIC (uint32_t ) use;/*usage counter */ const struct rte_eth_rxtx_callback *cb;/* callback handle */ struct rte_bpf *bpf;struct rte_bpf_jit jit;
@@ -110,8 +110,8 @@ struct bpf_eth_cbh {/* in use, busy wait till current RX/TX iteration is finished */ if ((puse &BPF_ETH_CBI_INUSE ) !=0 ) { -RTE_WAIT_UNTIL_MASKED ((uint32_t *)(uintptr_t )&cbi->use, - UINT32_MAX, !=, puse, __ATOMIC_RELAXED); + RTE_WAIT_UNTIL_MASKED((__rte_atomic uint32_t *)(uintptr_t)&cbi->use, + UINT32_MAX, !=, puse, rte_memory_order_relaxed); } }
Signature · 1 line
1.8.3.1
PATCH v4 6/6 devtools: forbid new direct use of GCC atomic builtins
In reply to Tyler RetzlaffRefrain from using compiler __atomic_xxx builtins DPDK now requires the use of rte_atomic_<op>_explicit macros when operating on DPDK atomic variables.
--- devtools/checkpatches.sh | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)
devtools/checkpatches.sh +3 -3
@@ -111,11 +111,11 @@ check_forbidden_additions() { # <patch> -f $(dirname $(readlink -f $0))/check-forbidden-tokens.awk \ "$1" || res=1 - # refrain from using compiler __atomic_{add,and,nand,or,sub,xor}_fetch() + # refrain from using compiler __atomic_xxx builtins awk -v FOLDERS="lib drivers app examples" \ - -v EXPRESSIONS="__atomic_(add|and|nand|or|sub|xor)_fetch\\\(" \ + -v EXPRESSIONS="__atomic_.*\\\(" \ -v RET_ON_FAIL=1 \ - -v MESSAGE='Using __atomic_op_fetch, prefer __atomic_fetch_op' \ + -v MESSAGE='Using __atomic_xxx builtins' \ -f $(dirname $(readlink -f $0))/check-forbidden-tokens.awk \ "$1" || res=1
Signature · 1 line
1.8.3.1
RE: [PATCH v4 6/6] devtools: forbid new direct use of GCC atomic builtins
Quoted 6 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Wednesday, 16 August 2023 23.39Refrain from using compiler
__atomic_xxxbuiltins DPDK now requires the use of rte_atomic_<op>_explicit macros when operating on DPDK atomic variables.
There is probably no end to how much can be added to checkpatches.
You got the important stuff, so below are only further suggestions!
[...]
devtools/checkpatches.sh check_forbidden_additions() { # <patch> · line 118
Quoted 8 lines
devtools/checkpatches.sh +3
-3- # refrain from using compiler __atomic_{add,and,nand,or,sub,xor}_fetch() + # refrain from using compiler __atomic_xxx builtins awk -v FOLDERS="lib drivers app examples" \ - -v EXPRESSIONS="__atomic_(add|and|nand|or|sub|xor)_fetch\\\(" \ + -v EXPRESSIONS="__atomic_.*\\\(" \ -v RET_ON_FAIL=1 \ - -v MESSAGE='Using __atomic_op_fetch, prefer __atomic_fetch_op' \ + -v MESSAGE='Using __atomic_xxx builtins' \
Alternatively:
-v MESSAGE='Using __atomic_xxx built-ins, prefer rte_atomic_xxx' \
Quoted 5 lines
devtools/checkpatches.sh +0
-1-f $(dirname $(readlink -f $0))/check-forbidden-tokens.awk \ "$1" || res=1 --1.8.3.1
This could be updated too:
# refrain from using compiler __atomic_thread_fence()
# It should be avoided on x86 for SMP case.
awk -v FOLDERS="lib drivers app examples" \
-v EXPRESSIONS="__atomic_thread_fence\\\(" \
-v RET_ON_FAIL=1 \
- -v MESSAGE='Using __atomic_thread_fence' \
+ -v MESSAGE='Using __atomic_thread_fence built-in, prefer __rte_atomic_thread_fence' \
-f $(dirname $(readlink -f $0))/check-forbidden-tokens.awk \
"$1" || res=1You could also add C11 variants of these tests... atomic_(load|store|exchange|compare_exchange_(strong|weak)|fetch_(add|sub|and|xor|or|nand)|flag_(test_and_set|clear))[_explicit], and atomic_thread_fence.
And a test for using "_Atomic".
-Morten
Re: [PATCH v4 6/6] devtools: forbid new direct use of GCC atomic builtins
Quoted 23 lines · Morten Brørup
Quoted 6 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Wednesday, 16 August 2023 23.39Refrain from using compiler
__atomic_xxxbuiltins DPDK now requires the use of rte_atomic_<op>_explicit macros when operating on DPDK atomic variables.There is probably no end to how much can be added to checkpatches.
You got the important stuff, so below are only further suggestions!
[...]
devtools/checkpatches.sh check_forbidden_additions() { # <patch> · line 118
Quoted 8 lines
devtools/checkpatches.sh +3
-3- # refrain from using compiler __atomic_{add,and,nand,or,sub,xor}_fetch() + # refrain from using compiler __atomic_xxx builtins awk -v FOLDERS="lib drivers app examples" \ - -v EXPRESSIONS="__atomic_(add|and|nand|or|sub|xor)_fetch\\\(" \ + -v EXPRESSIONS="__atomic_.*\\\(" \ -v RET_ON_FAIL=1 \ - -v MESSAGE='Using __atomic_op_fetch, prefer __atomic_fetch_op' \ + -v MESSAGE='Using __atomic_xxx builtins' \Alternatively:
Quoted 1 line
-v MESSAGE='Using __atomic_xxx built-ins, prefer rte_atomic_xxx' \
i can adjust the wording as you suggest, no problem
Quoted 16 lines
Quoted 6 lines
devtools/checkpatches.sh +0
-1-f $(dirname $(readlink -f $0))/check-forbidden-tokens.awk \ "$1" || res=1 --1.8.3.1
This could be updated too:
# refrain from using compiler __atomic_thread_fence() # It should be avoided on x86 for SMP case. awk -v FOLDERS="lib drivers app examples" \ -v EXPRESSIONS="__atomic_thread_fence\\\(" \ -v RET_ON_FAIL=1 \(quoted diff) +1
-1- -v MESSAGE='Using __atomic_thread_fence' \ + -v MESSAGE='Using __atomic_thread_fence built-in, prefer __rte_atomic_thread_fence' \
Quoted 1 line
-f $(dirname $(readlink -f $0))/check-forbidden-tokens.awk \
yeah, i left this one separate i think the advice is actually use rte_atomic_thread_fence which may be an inline function that uses __rte_atomic_thread_fence
Quoted 7 lines
"$1" || res=1You could also add C11 variants of these tests... atomic_(load|store|exchange|compare_exchange_(strong|weak)|fetch_(add|sub|and|xor|or|nand)|flag_(test_and_set|clear))[_explicit], and
atomic_thread_fence.And a test for using "
_Atomic".
direct use would cause early compilation in the CI so it would be caught fairly early. i'm not sure i want to get into the business of trying to add redundant (albiet cheaper) earlier checks.
though if there is a general call for this from the reviewers i'll add them.
Quoted 2 lines
-Morten
RE: [PATCH v4 6/6] devtools: forbid new direct use of GCC atomic builtins
Quoted 49 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Thursday, 17 August 2023 21.14Quoted 13 lines · Morten Brørup
Quoted 6 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Wednesday, 16 August 2023 23.39Refrain from using compiler
__atomic_xxxbuiltins DPDK now requires the use of rte_atomic_<op>_explicit macros when operating on DPDK atomic variables.There is probably no end to how much can be added to checkpatches.
You got the important stuff, so below are only further suggestions!
[...]
Quoted 11 lines
devtools/checkpatches.sh check_forbidden_additions() { # <patch> · line 118
Quoted 8 lines
devtools/checkpatches.sh +3
-3- # refrain from using compiler __atomic_{add,and,nand,or,sub,xor}_fetch() + # refrain from using compiler __atomic_xxx builtins awk -v FOLDERS="lib drivers app examples" \ - -v EXPRESSIONS="__atomic_(add|and|nand|or|sub|xor)_fetch\\\(" \ + -v EXPRESSIONS="__atomic_.*\\\(" \ -v RET_ON_FAIL=1 \ - -v MESSAGE='Using __atomic_op_fetch, prefer __atomic_fetch_op' \ + -v MESSAGE='Using __atomic_xxx builtins' \Alternatively:
-v MESSAGE='Using__atomic_xxxbuilt-ins, preferrte_atomic_xxx'\
i can adjust the wording as you suggest, no problem
Quoted 14 lines
Quoted 6 lines
devtools/checkpatches.sh +0
-1-f $(dirname $(readlink -f $0))/check-forbidden-tokens.awk \ "$1" || res=1 --1.8.3.1
This could be updated too:
# refrain from using compiler __atomic_thread_fence() # It should be avoided on x86 for SMP case. awk -v FOLDERS="lib drivers app examples" \ -v EXPRESSIONS="__atomic_thread_fence\\\(" \ -v RET_ON_FAIL=1 \Review
Quoted 2 lines
(quoted diff) +1
-1- -v MESSAGE='Using __atomic_thread_fence' \ + -v MESSAGE='Using __atomic_thread_fence built-in, prefer
__rte_atomic_thread_fence' \
Quoted 5 lines
Quoted 1 line
-f $(dirname $(readlink -f $0))/check-forbidden-tokens.awk \yeah, i left this one separate i think the advice is actually use
rte_atomic_thread_fencewhich may be an inline function that uses__rte_atomic_thread_fence
I now noticed that the comment to this says "# [...] should be avoided on x86 for SMP case." Wouldn't that apply to __rte_atomic_thread_fence too? So would we want a similar warning for __rte_atomic_thread_fence in checkpatches; i.e. warnings for all variants of [__[rte_]]atomic_thread_fence?
If the use of [__[rte_]]atomic_thread_fence is only conditionally prohibited, I think that we should move the warning to the definition of __rte_atomic_thread_fence itself, gated by x64 and SMP being defined. The CI would catch its use in DPDK, but still allow application developers to use it (for targets not being x86 SMP). What do others think?
Quoted 17 lines
Quoted 5 lines
"$1" || res=1You could also add C11 variants of these tests...
atomic_(load|store|exchange|compare_exchange_(strong|weak)|fetch_(add|sub|and| xor|or|nand)|flag_(test_and_set|clear))[_explicit], and
Quoted 3 lines
atomic_thread_fence.And a test for using "
_Atomic".direct use would cause early compilation in the CI so it would be caught fairly early. i'm not sure i want to get into the business of trying to add redundant (albiet cheaper) earlier checks.
though if there is a general call for this from the reviewers i'll add them.
Re: [PATCH v4 6/6] devtools: forbid new direct use of GCC atomic builtins
Quoted 49 lines · Morten Brørup
Quoted 49 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Thursday, 17 August 2023 21.14Quoted 13 lines · Morten Brørup
Quoted 6 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Wednesday, 16 August 2023 23.39Refrain from using compiler
__atomic_xxxbuiltins DPDK now requires the use of rte_atomic_<op>_explicit macros when operating on DPDK atomic variables.There is probably no end to how much can be added to checkpatches.
You got the important stuff, so below are only further suggestions!
[...]
Quoted 11 lines
devtools/checkpatches.sh check_forbidden_additions() { # <patch> · line 118
Quoted 8 lines
devtools/checkpatches.sh +3
-3- # refrain from using compiler __atomic_{add,and,nand,or,sub,xor}_fetch() + # refrain from using compiler __atomic_xxx builtins awk -v FOLDERS="lib drivers app examples" \ - -v EXPRESSIONS="__atomic_(add|and|nand|or|sub|xor)_fetch\\\(" \ + -v EXPRESSIONS="__atomic_.*\\\(" \ -v RET_ON_FAIL=1 \ - -v MESSAGE='Using __atomic_op_fetch, prefer __atomic_fetch_op' \ + -v MESSAGE='Using __atomic_xxx builtins' \Alternatively:
-v MESSAGE='Using__atomic_xxxbuilt-ins, preferrte_atomic_xxx'\
i can adjust the wording as you suggest, no problem
Quoted 14 lines
Quoted 6 lines
devtools/checkpatches.sh +0
-1-f $(dirname $(readlink -f $0))/check-forbidden-tokens.awk \ "$1" || res=1 --1.8.3.1
This could be updated too:
# refrain from using compiler __atomic_thread_fence() # It should be avoided on x86 for SMP case. awk -v FOLDERS="lib drivers app examples" \ -v EXPRESSIONS="__atomic_thread_fence\\\(" \ -v RET_ON_FAIL=1 \Review
Quoted 2 lines
(quoted diff) +1
-1- -v MESSAGE='Using __atomic_thread_fence' \ + -v MESSAGE='Using __atomic_thread_fence built-in, prefer
__rte_atomic_thread_fence' \
Quoted 7 lines
Quoted 5 lines
Quoted 1 line
-f $(dirname $(readlink -f $0))/check-forbidden-tokens.awk \yeah, i left this one separate i think the advice is actually use
rte_atomic_thread_fencewhich may be an inline function that uses__rte_atomic_thread_fenceI now noticed that the comment to this says "# [...] should be avoided on x86 for SMP case." Wouldn't that apply to
__rte_atomic_thread_fencetoo? So would we want a similar warning for__rte_atomic_thread_fencein checkpatches; i.e. warnings for all variants of [__[rte_]]atomic_thread_fence?
to understand how this applies we need to look atx86/include/rte_atomic.h
static __rte_always_inline void rte_atomic_thread_fence(rte_memory_order memorder)
{
if (memorder == rte_memory_order_seq_cst)
rte_smp_mb();
else
__rte_atomic_thread_fence(memorder);
}So what i've done is this
You *should* always use rte_atomic_thread_fence() because it does the dance to give you what you are supposed to on "x86 for the SMP case" right?
Quoted 2 lines
If the use of [__[rte_]]atomic_thread_fence is only conditionally prohibited, I think that we should move the warning to the definition of
__rte_atomic_thread_fenceitself, gated by x64 and SMP being defined. The CI would catch its use in DPDK, but still allow application developers to use it (for targets not being x86 SMP). What do others think?
I'll tweak the patch to warn about __rte_atomic and __atomic since the correct usage is always using rte_atomic_xxx
Quoted 18 lines
Quoted 18 lines
Quoted 6 lines
"$1" || res=1You could also add C11 variants of these tests...
atomic_(load|store|exchange|compare_exchange_(strong|weak)|fetch_(add|sub|and| xor|or|nand)|flag_(test_and_set|clear))[_explicit], and
Quoted 3 lines
atomic_thread_fence.And a test for using "
_Atomic".direct use would cause early compilation in the CI so it would be caught fairly early. i'm not sure i want to get into the business of trying to add redundant (albiet cheaper) earlier checks.
though if there is a general call for this from the reviewers i'll add them.
v5 cover optional rte optional stdatomics API
In reply to Tyler RetzlaffThis series introduces API additions prefixed in the rte namespace that allow the optional use of stdatomics.h from C11 using enable_stdatomics=true for targets where enable_stdatomics=false no functional change is intended.
Be aware this does not contain all changes to use stdatomics across the DPDK tree it only introduces the minimum to allow the option to be used which is a pre-requisite for a clean CI (probably using clang) that can be run with enable_stdatomics=true enabled.
It is planned that subsequent series will be introduced per lib/driver as appropriate to further enable stdatomics use when enable_stdatomics=true.
Notes:
* Additional libraries beyond EAL make visible atomics use across the API/ABI surface they will be converted in the subsequent series.
* The eal: add rte atomic qualifier with casts patch needs some discussion as to whether or not the legacy rte_atomic APIs should be converted to work with enable_stdatomic=true right now some implementation dependent casts are used to prevent cascading / having to convert too much in the intial series.
* Windows will obviously need complete conversion of libraries including atomics that are not crossing API/ABI boundaries. those conversions will introduced in separate series as new along side the existing msvc series.
Please keep in mind we would like to prioritize the review / acceptance of this patch since it needs to be completed in the 23.11 merge window.
Thank you all for the discussion that lead to the formation of this series.
v5:
* Add RTE_ATOMIC to doxygen configuration PREDEFINED macros list to fix documentation generation failure
* Fix two typos in expansion of C11 atomics macros strong -> weak and add missing _explicit
* Adjust devtools/checkpatches messages based on feedback. i have chosen not to try and catch use of C11 atomics or _Atomic since using those directly will be picked up by existing CI passes where by compilation error where enable_stdatomic=false (the default for most platforms)
v4:
* Move the definition of #define RTE_ATOMIC(type) to patch 1 where it belongs (a mistake in v3)
* Provide comments for both RTE_ATOMIC and __rte_atomic macros indicating their use as specified or qualified contexts.
v3:
* Remove comments from APIs mentioning the mapping to C++ memory model memory orders
* Introduce and use new macro RTE_ATOMIC(type) to be used in contexts where _Atomic is used as a type specifier to declare variables. The macro allows more clarity about what the atomic type being specified is. e.g. _Atomic(T *) vs _Atomic(T) it is easier to understand that the former is an atomic pointer type and the latter is an atomic type. it also has the benefit of (in the future) being interoperable with c++23 syntactically
note: Morten i have retained your 'reviewed-by' tags if you disagree given the changes in the above version please indicate as such but i believe the changes are in the spirit of the feedback you provided
v2:
* Wrap meson_options.txt option description to newline and indent to be consistent with other options.
* Provide separate typedef of rte_memory_order for enable_stdatomic=true VS enable_stdatomic=false instead of a single typedef to int
note: slight tweak to reviewers feedback i've chosen to use a typedef
for both enable_stdatomic={true,false} (just seemed more consistent)
* Bring in assert.h and use static_assert macro instead of _Static_assert keyword to better interoperate with c/c++
* Directly include rte_stdatomic.h where into other places it is consumed instead of hacking it globally into rte_config.h
* Provide and use __rte_atomic_thread_fence to allow conditional expansion within the body of existing rte_atomic_thread_fence inline function to maintain per-arch optimizations when enable_stdatomic=false
Tyler Retzlaff (6):
eal: provide rte stdatomics optional atomics API
eal: adapt EAL to present rte optional atomics API
eal: add rte atomic qualifier with casts
distributor: adapt for EAL optional atomics API changes
bpf: adapt for EAL optional atomics API changes
devtools: forbid new direct use of GCC atomic builtins
app/test/test_mcslock.c | 6 +- config/meson.build | 1 + devtools/checkpatches.sh | 8 +- doc/api/doxy-api.conf.in | 1 + lib/bpf/bpf_pkt.c | 6 +- lib/distributor/distributor_private.h | 2 +- lib/distributor/rte_distributor_single.c | 44 +++---- lib/eal/arm/include/rte_atomic_32.h | 4 +- lib/eal/arm/include/rte_atomic_64.h | 36 +++--- lib/eal/arm/include/rte_pause_64.h | 26 ++-- lib/eal/arm/rte_power_intrinsics.c | 8 +- lib/eal/common/eal_common_trace.c | 16 +-- lib/eal/include/generic/rte_atomic.h | 67 +++++++---- lib/eal/include/generic/rte_pause.h | 50 ++++---- lib/eal/include/generic/rte_rwlock.h | 48 ++++---- lib/eal/include/generic/rte_spinlock.h | 20 ++-- lib/eal/include/meson.build | 1 + lib/eal/include/rte_mcslock.h | 51 ++++---- lib/eal/include/rte_pflock.h | 25 ++-- lib/eal/include/rte_seqcount.h | 19 +-- lib/eal/include/rte_stdatomic.h | 198 +++++++++++++++++++++++++++++++ lib/eal/include/rte_ticketlock.h | 43 +++---- lib/eal/include/rte_trace_point.h | 5 +- lib/eal/loongarch/include/rte_atomic.h | 4 +- lib/eal/ppc/include/rte_atomic.h | 54 ++++----- lib/eal/riscv/include/rte_atomic.h | 4 +- lib/eal/x86/include/rte_atomic.h | 8 +- lib/eal/x86/include/rte_spinlock.h | 2 +- lib/eal/x86/rte_power_intrinsics.c | 7 +- meson_options.txt | 2 + 30 files changed, 499 insertions(+), 267 deletions(-) create mode 100644 lib/eal/include/rte_stdatomic.h
Signature · 1 line
1.8.3.1
PATCH v5 1/6 eal: provide rte stdatomics optional atomics API
Provide API for atomic operations in the rte namespace that may optionally be configured to use C11 atomics with meson option enable_stdatomics=true
--- config/meson.build | 1 + doc/api/doxy-api.conf.in | 1 + lib/eal/include/generic/rte_atomic.h | 1 + lib/eal/include/generic/rte_pause.h | 1 + lib/eal/include/generic/rte_rwlock.h | 1 + lib/eal/include/generic/rte_spinlock.h | 1 + lib/eal/include/meson.build | 1 + lib/eal/include/rte_mcslock.h | 1 + lib/eal/include/rte_pflock.h | 1 + lib/eal/include/rte_seqcount.h | 1 + lib/eal/include/rte_stdatomic.h | 198 +++++++++++++++++++++++++++++++++ lib/eal/include/rte_ticketlock.h | 1 + lib/eal/include/rte_trace_point.h | 1 + meson_options.txt | 2 + 14 files changed, 212 insertions(+) create mode 100644 lib/eal/include/rte_stdatomic.h
config/meson.build +1 -0
@@ -303,6 +303,7 @@ endforeach # set other values pulled from the build options dpdk_conf.set('RTE_MAX_ETHPORTS', get_option('max_ethports')) dpdk_conf.set('RTE_LIBEAL_USE_HPET', get_option('use_hpet')) +dpdk_conf.set('RTE_ENABLE_STDATOMIC', get_option('enable_stdatomic')) dpdk_conf.set('RTE_ENABLE_TRACE_FP', get_option('enable_trace_fp')) # values which have defaults which may be overridden dpdk_conf.set('RTE_MAX_VFIO_GROUPS', 64)
doc/api/doxy-api.conf.in +1 -0
@@ -84,6 +84,7 @@ INPUT += @API_EXAMPLES@ FILE_PATTERNS = rte_*.h \ cmdline.h PREDEFINED = __DOXYGEN__ \ + RTE_ATOMIC \ RTE_HAS_CPUSET \ VFIO_PRESENT \ __rte_lockable= \
lib/eal/include/generic/rte_atomic.h +1 -0
@@ -15,6 +15,7 @@#include <stdint.h> #include <rte_compat.h> #include <rte_common.h> +#include <rte_stdatomic.h> #ifdef __DOXYGEN__
lib/eal/include/generic/rte_pause.h +1 -0
@@ -16,6 +16,7 @@#include <assert.h> #include <rte_common.h> #include <rte_atomic.h> +#include <rte_stdatomic.h> /** * Pause CPU execution for a short while
lib/eal/include/generic/rte_rwlock.h +1 -0
@@ -32,6 +32,7 @@#include <rte_common.h> #include <rte_lock_annotations.h> #include <rte_pause.h> +#include <rte_stdatomic.h> /** * The rte_rwlock_t type.
lib/eal/include/generic/rte_spinlock.h +1 -0
@@ -23,6 +23,7 @@#endif #include <rte_lock_annotations.h> #include <rte_pause.h> +#include <rte_stdatomic.h> /** * The rte_spinlock_t type.
lib/eal/include/meson.build +1 -0
@@ -42,6 +42,7 @@ headers += files( 'rte_seqlock.h', 'rte_service.h', 'rte_service_component.h', + 'rte_stdatomic.h', 'rte_string_fns.h', 'rte_tailq.h', 'rte_thread.h',
lib/eal/include/rte_mcslock.h +1 -0
@@ -27,6 +27,7 @@#include <rte_common.h> #include <rte_pause.h> #include <rte_branch_prediction.h> +#include <rte_stdatomic.h> /** * The rte_mcslock_t type.
lib/eal/include/rte_pflock.h +1 -0
@@ -34,6 +34,7 @@#include <rte_compat.h> #include <rte_common.h> #include <rte_pause.h> +#include <rte_stdatomic.h> /** * The rte_pflock_t type.
lib/eal/include/rte_seqcount.h +1 -0
@@ -26,6 +26,7 @@#include <rte_atomic.h> #include <rte_branch_prediction.h> #include <rte_compat.h> +#include <rte_stdatomic.h> /** * The RTE seqcount type.
lib/eal/include/rte_stdatomic.h new file +198 -0
198 lines not on this page · raw message
lib/eal/include/rte_ticketlock.h +1 -0
@@ -24,6 +24,7 @@#include <rte_common.h> #include <rte_lcore.h> #include <rte_pause.h> +#include <rte_stdatomic.h> /** * The rte_ticketlock_t type.
lib/eal/include/rte_trace_point.h +1 -0
@@ -30,6 +30,7 @@#include <rte_per_lcore.h> #include <rte_string_fns.h> #include <rte_uuid.h> +#include <rte_stdatomic.h> /** The tracepoint object. */ typedef uint64_t rte_trace_point_t ;
meson_options.txt +2 -0
@@ -46,6 +46,8 @@ option('mbuf_refcnt_atomic', type: 'boolean', value: true, description: 'Atomically access the mbuf refcnt.') option('platform', type: 'string', value: 'native', description: 'Platform to build, either "native", "generic" or a SoC. Please refer to the Linux build guide for more information.') +option('enable_stdatomic', type: 'boolean', value: false, description: + 'enable use of C11 stdatomic') option('enable_trace_fp', type: 'boolean', value: false, description: 'enable fast path trace points.') option('tests', type: 'boolean', value: true, description:
Signature · 1 line
1.8.3.1
PATCH v5 2/6 eal: adapt EAL to present rte optional atomics API
In reply to Tyler RetzlaffAdapt the EAL public headers to use rte optional atomics API instead of directly using and exposing toolchain specific atomic builtin intrinsics.
--- app/test/test_mcslock.c | 6 ++-- lib/eal/arm/include/rte_atomic_32.h | 4 +-- lib/eal/arm/include/rte_atomic_64.h | 36 +++++++++++------------ lib/eal/arm/include/rte_pause_64.h | 26 ++++++++-------- lib/eal/arm/rte_power_intrinsics.c | 8 ++--- lib/eal/common/eal_common_trace.c | 16 +++++----- lib/eal/include/generic/rte_atomic.h | 50 +++++++++++++++---------------- lib/eal/include/generic/rte_pause.h | 46 ++++++++++++----------------- lib/eal/include/generic/rte_rwlock.h | 47 +++++++++++++++-------------- lib/eal/include/generic/rte_spinlock.h | 19 ++++++------ lib/eal/include/rte_mcslock.h | 50 +++++++++++++++---------------- lib/eal/include/rte_pflock.h | 24 ++++++++------- lib/eal/include/rte_seqcount.h | 18 ++++++------ lib/eal/include/rte_ticketlock.h | 42 +++++++++++++------------- lib/eal/include/rte_trace_point.h | 4 +-- lib/eal/loongarch/include/rte_atomic.h | 4 +-- lib/eal/ppc/include/rte_atomic.h | 54 +++++++++++++++++----------------- lib/eal/riscv/include/rte_atomic.h | 4 +-- lib/eal/x86/include/rte_atomic.h | 8 ++--- lib/eal/x86/include/rte_spinlock.h | 2 +- lib/eal/x86/rte_power_intrinsics.c | 6 ++-- 21 files changed, 237 insertions(+), 237 deletions(-)
app/test/test_mcslock.c +3 -3
@@ -36,9 +36,9 @@* lock multiple times. */ -rte_mcslock_t *p_ml; -rte_mcslock_t *p_ml_try; -rte_mcslock_t *p_ml_perf; +RTE_ATOMIC (rte_mcslock_t *) p_ml; +RTE_ATOMIC (rte_mcslock_t *) p_ml_try; +RTE_ATOMIC (rte_mcslock_t *) p_ml_perf;static unsigned int count;
lib/eal/arm/include/rte_atomic_32.h +2 -2
@@ -34,9 +34,9 @@#define rte_io_rmb () rte_rmb() static __rte_always_inline void -rte_atomic_thread_fence (int memorder) +rte_atomic_thread_fence (rte_memory_order memorder) { -__atomic_thread_fence (memorder); +__rte_atomic_thread_fence (memorder); }#ifdef __cplusplus
lib/eal/arm/include/rte_atomic_64.h +18 -18
@@ -38,9 +38,9 @@#define rte_io_rmb () rte_rmb() static __rte_always_inline void -rte_atomic_thread_fence (int memorder) +rte_atomic_thread_fence (rte_memory_order memorder) { -__atomic_thread_fence (memorder); +__rte_atomic_thread_fence (memorder); }/*------------------------ 128 bit atomic operations -------------------------*/
@@ -107,33 +107,33 @@*/ RTE_SET_USED (failure);/* Find invalid memory order */ -RTE_ASSERT (success == __ATOMIC_RELAXED || - success == __ATOMIC_ACQUIRE || - success == __ATOMIC_RELEASE || - success == __ATOMIC_ACQ_REL || - success == __ATOMIC_SEQ_CST); +RTE_ASSERT (success == rte_memory_order_relaxed || + success == rte_memory_order_acquire || + success == rte_memory_order_release || + success == rte_memory_order_acq_rel || + success == rte_memory_order_seq_cst);rte_int128_t expected = *exp;rte_int128_t desired = *src;rte_int128_t old;#if defined(__ARM_FEATURE_ATOMICS) || defined(RTE_ARM_FEATURE_ATOMICS ) -if (success == __ATOMIC_RELAXED) +if (success == rte_memory_order_relaxed)__cas_128_relaxed (dst, exp, desired); -else if (success == __ATOMIC_ACQUIRE) +else if (success == rte_memory_order_acquire)__cas_128_acquire (dst, exp, desired); -else if (success == __ATOMIC_RELEASE) +else if (success == rte_memory_order_release)__cas_128_release (dst, exp, desired);else __cas_128_acq_rel (dst, exp, desired); old = *exp;#else -#define __HAS_ACQ (mo) ((mo) != __ATOMIC_RELAXED && (mo) != __ATOMIC_RELEASE) -#define __HAS_RLS (mo) ((mo) == __ATOMIC_RELEASE || (mo) == __ATOMIC_ACQ_REL || \ - (mo) == __ATOMIC_SEQ_CST) +#define __HAS_ACQ (mo) ((mo) != rte_memory_order_relaxed && (mo) != rte_memory_order_release) +#define __HAS_RLS (mo) ((mo) == rte_memory_order_release || (mo) == rte_memory_order_acq_rel || \ + (mo) == rte_memory_order_seq_cst) -int ldx_mo =__HAS_ACQ (success) ? __ATOMIC_ACQUIRE : __ATOMIC_RELAXED; -int stx_mo =__HAS_RLS (success) ? __ATOMIC_RELEASE : __ATOMIC_RELAXED; +int ldx_mo =__HAS_ACQ (success) ? rte_memory_order_acquire : rte_memory_order_relaxed; +int stx_mo =__HAS_RLS (success) ? rte_memory_order_release : rte_memory_order_relaxed;#undef __HAS_ACQ#undef __HAS_RLS
@@ -153,7 +153,7 @@ : "Q" (src->val[0]) \ : "memory"); } -if (ldx_mo == __ATOMIC_RELAXED) +if (ldx_mo == rte_memory_order_relaxed) __LOAD_128("ldxp", dst, old) else__LOAD_128 ("ldaxp" , dst, old)
@@ -170,7 +170,7 @@ : "memory"); }if (likely (old.int128 == expected.int128)) { -if (stx_mo == __ATOMIC_RELAXED) +if (stx_mo == rte_memory_order_relaxed) __STORE_128("stxp", dst, desired, ret) else__STORE_128 ("stlxp" , dst, desired, ret)
@@ -181,7 +181,7 @@* needs to be stored back to ensure it was read * atomically. */ -if (stx_mo == __ATOMIC_RELAXED) +if (stx_mo == rte_memory_order_relaxed) __STORE_128("stxp", dst, old, ret) else__STORE_128 ("stlxp" , dst, old, ret)
lib/eal/arm/include/rte_pause_64.h +13 -13
@@ -41,7 +41,7 @@ static inline void rte_pause(void)* implicitly to exit WFE. */ #define __RTE_ARM_LOAD_EXC_8 (src, dst, memorder) { \ - if (memorder == __ATOMIC_RELAXED) { \ + if (memorder == rte_memory_order_relaxed) { \ asm volatile("ldxrb %w[tmp], [%x[addr]]" \ : [tmp] "=&r" (dst) \ : [addr] "r" (src) \
@@ -60,7 +60,7 @@ static inline void rte_pause(void)* implicitly to exit WFE. */ #define __RTE_ARM_LOAD_EXC_16 (src, dst, memorder) { \ - if (memorder == __ATOMIC_RELAXED) { \ + if (memorder == rte_memory_order_relaxed) { \ asm volatile("ldxrh %w[tmp], [%x[addr]]" \ : [tmp] "=&r" (dst) \ : [addr] "r" (src) \
@@ -79,7 +79,7 @@ static inline void rte_pause(void)* implicitly to exit WFE. */ #define __RTE_ARM_LOAD_EXC_32 (src, dst, memorder) { \ - if (memorder == __ATOMIC_RELAXED) { \ + if (memorder == rte_memory_order_relaxed) { \ asm volatile("ldxr %w[tmp], [%x[addr]]" \ : [tmp] "=&r" (dst) \ : [addr] "r" (src) \
@@ -98,7 +98,7 @@ static inline void rte_pause(void)* implicitly to exit WFE. */ #define __RTE_ARM_LOAD_EXC_64 (src, dst, memorder) { \ - if (memorder == __ATOMIC_RELAXED) { \ + if (memorder == rte_memory_order_relaxed) { \ asm volatile("ldxr %x[tmp], [%x[addr]]" \ : [tmp] "=&r" (dst) \ : [addr] "r" (src) \
@@ -118,7 +118,7 @@ static inline void rte_pause(void)*/ #define __RTE_ARM_LOAD_EXC_128 (src, dst, memorder) { \ volatile rte_int128_t *dst_128 = (volatile rte_int128_t *)&dst; \ - if (memorder == __ATOMIC_RELAXED) { \ + if (memorder == rte_memory_order_relaxed) { \ asm volatile("ldxp %x[tmp0], %x[tmp1], [%x[addr]]" \ : [tmp0] "=&r" (dst_128->val[0]), \ [tmp1] "=&r" (dst_128->val[1]) \
@@ -153,8 +153,8 @@ static inline void rte_pause(void) {uint16_t value; -RTE_BUILD_BUG_ON (memorder != __ATOMIC_ACQUIRE && - memorder != __ATOMIC_RELAXED); +RTE_BUILD_BUG_ON (memorder != rte_memory_order_acquire && + memorder != rte_memory_order_relaxed); __RTE_ARM_LOAD_EXC_16(addr, value, memorder)if (value != expected) {
@@ -172,8 +172,8 @@ static inline void rte_pause(void) {uint32_t value; -RTE_BUILD_BUG_ON (memorder != __ATOMIC_ACQUIRE && - memorder != __ATOMIC_RELAXED); +RTE_BUILD_BUG_ON (memorder != rte_memory_order_acquire && + memorder != rte_memory_order_relaxed); __RTE_ARM_LOAD_EXC_32(addr, value, memorder)if (value != expected) {
@@ -191,8 +191,8 @@ static inline void rte_pause(void) {uint64_t value; -RTE_BUILD_BUG_ON (memorder != __ATOMIC_ACQUIRE && - memorder != __ATOMIC_RELAXED); +RTE_BUILD_BUG_ON (memorder != rte_memory_order_acquire && + memorder != rte_memory_order_relaxed); __RTE_ARM_LOAD_EXC_64(addr, value, memorder)if (value != expected) {
@@ -206,8 +206,8 @@ static inline void rte_pause(void)#define RTE_WAIT_UNTIL_MASKED (addr, mask, cond, expected, memorder) do { \ RTE_BUILD_BUG_ON(!__builtin_constant_p(memorder)); \ - RTE_BUILD_BUG_ON(memorder != __ATOMIC_ACQUIRE && \ - memorder != __ATOMIC_RELAXED); \ + RTE_BUILD_BUG_ON(memorder != rte_memory_order_acquire && \ + memorder != rte_memory_order_relaxed); \ const uint32_t size = sizeof(*(addr)) << 3; \ typeof(*(addr)) expected_value = (expected); \ typeof(*(addr)) value; \
lib/eal/arm/rte_power_intrinsics.c +4 -4
@@ -33,19 +33,19 @@switch (pmc->size) {case sizeof (uint8_t ): - __RTE_ARM_LOAD_EXC_8(pmc->addr, cur_value, __ATOMIC_RELAXED) + __RTE_ARM_LOAD_EXC_8(pmc->addr, cur_value, rte_memory_order_relaxed)__RTE_ARM_WFE () break;case sizeof (uint16_t ): - __RTE_ARM_LOAD_EXC_16(pmc->addr, cur_value, __ATOMIC_RELAXED) + __RTE_ARM_LOAD_EXC_16(pmc->addr, cur_value, rte_memory_order_relaxed)__RTE_ARM_WFE () break;case sizeof (uint32_t ): - __RTE_ARM_LOAD_EXC_32(pmc->addr, cur_value, __ATOMIC_RELAXED) + __RTE_ARM_LOAD_EXC_32(pmc->addr, cur_value, rte_memory_order_relaxed)__RTE_ARM_WFE () break;case sizeof (uint64_t ): - __RTE_ARM_LOAD_EXC_64(pmc->addr, cur_value, __ATOMIC_RELAXED) + __RTE_ARM_LOAD_EXC_64(pmc->addr, cur_value, rte_memory_order_relaxed)__RTE_ARM_WFE () break;default :
lib/eal/common/eal_common_trace.c +9 -7
@@ -103,11 +103,11 @@ struct trace_point_head * trace_mode_set(rte_trace_point_t *t, enum rte_trace_mode mode) {if (mode ==RTE_TRACE_MODE_OVERWRITE ) -__atomic_fetch_and (t, ~__RTE_TRACE_FIELD_ENABLE_DISCARD, - __ATOMIC_RELEASE); +rte_atomic_fetch_and_explicit (t, ~__RTE_TRACE_FIELD_ENABLE_DISCARD, + rte_memory_order_release);else -__atomic_fetch_or (t, __RTE_TRACE_FIELD_ENABLE_DISCARD, - __ATOMIC_RELEASE); +rte_atomic_fetch_or_explicit (t, __RTE_TRACE_FIELD_ENABLE_DISCARD, + rte_memory_order_release); }void
@@ -141,7 +141,7 @@ rte_trace_mode rte_trace_mode_get(void)if (trace_point_is_invalid (t))return false; - val =__atomic_load_n (t, __ATOMIC_ACQUIRE); + val =rte_atomic_load_explicit (t, rte_memory_order_acquire);return (val & __RTE_TRACE_FIELD_ENABLE_MASK) !=0 ; }
@@ -153,7 +153,8 @@ rte_trace_mode rte_trace_mode_get(void)if (trace_point_is_invalid (t))return -ERANGE ; - prev =__atomic_fetch_or (t, __RTE_TRACE_FIELD_ENABLE_MASK, __ATOMIC_RELEASE); + prev =rte_atomic_fetch_or_explicit (t, __RTE_TRACE_FIELD_ENABLE_MASK, + rte_memory_order_release);if ((prev & __RTE_TRACE_FIELD_ENABLE_MASK) ==0 )__atomic_fetch_add (&trace.status,1 , __ATOMIC_RELEASE);return 0 ;
@@ -167,7 +168,8 @@ rte_trace_mode rte_trace_mode_get(void)if (trace_point_is_invalid (t))return -ERANGE ; - prev =__atomic_fetch_and (t, ~__RTE_TRACE_FIELD_ENABLE_MASK, __ATOMIC_RELEASE); + prev =rte_atomic_fetch_and_explicit (t, ~__RTE_TRACE_FIELD_ENABLE_MASK, + rte_memory_order_release);if ((prev & __RTE_TRACE_FIELD_ENABLE_MASK) !=0 )__atomic_fetch_sub (&trace.status,1 , __ATOMIC_RELEASE);return 0 ;
lib/eal/include/generic/rte_atomic.h +25 -25
194 lines not on this page · raw message
lib/eal/include/generic/rte_pause.h +19 -27
103 lines not on this page · raw message
lib/eal/include/generic/rte_rwlock.h +25 -22
126 lines not on this page · raw message
lib/eal/include/generic/rte_spinlock.h +9 -10
@@ -29,7 +29,7 @@* The rte_spinlock_t type. */ typedef struct __rte_lockable { -volatile int locked;/**< lock status 0 = unlocked, 1 = locked */ +volatile RTE_ATOMIC (int ) locked;/**< lock status 0 = unlocked, 1 = locked */ }rte_spinlock_t ;/**
@@ -66,10 +66,10 @@ {int exp =0 ; -while (!__atomic_compare_exchange_n (&sl->locked, &exp,1 ,0 , - __ATOMIC_ACQUIRE, __ATOMIC_RELAXED)) { -rte_wait_until_equal_32 ((volatile uint32_t *)&sl->locked, -0 , __ATOMIC_RELAXED); +while (!rte_atomic_compare_exchange_strong_explicit (&sl->locked, &exp,1 , + rte_memory_order_acquire, rte_memory_order_relaxed)) { +rte_wait_until_equal_32 ((volatile uint32_t *)(uintptr_t )&sl->locked, +0 , rte_memory_order_relaxed); exp =0 ; } }
@@ -90,7 +90,7 @@ rte_spinlock_unlock(rte_spinlock_t *sl) __rte_no_thread_safety_analysis { -__atomic_store_n (&sl->locked,0 , __ATOMIC_RELEASE); +rte_atomic_store_explicit (&sl->locked,0 , rte_memory_order_release); }#endif
@@ -113,9 +113,8 @@ __rte_no_thread_safety_analysis {int exp =0 ; -return __atomic_compare_exchange_n (&sl->locked, &exp,1 , -0 ,/* disallow spurious failure */ - __ATOMIC_ACQUIRE, __ATOMIC_RELAXED); +return rte_atomic_compare_exchange_strong_explicit (&sl->locked, &exp,1 , + rte_memory_order_acquire, rte_memory_order_relaxed); }#endif
@@ -129,7 +128,7 @@*/ static inline int rte_spinlock_is_locked (rte_spinlock_t *sl) { -return __atomic_load_n (&sl->locked, __ATOMIC_ACQUIRE); +return rte_atomic_load_explicit (&sl->locked, rte_memory_order_acquire); }/**
lib/eal/include/rte_mcslock.h +25 -25
133 lines not on this page · raw message
lib/eal/include/rte_pflock.h +13 -11
@@ -41,8 +41,8 @@*/ struct rte_pflock {struct { -uint16_t in; -uint16_t out; +RTE_ATOMIC (uint16_t ) in; +RTE_ATOMIC (uint16_t ) out; } rd, wr; };typedef struct rte_pflock rte_pflock_t ;
@@ -117,14 +117,14 @@ struct rte_pflock {* If no writer is present, then the operation has completed * successfully. */ - w =__atomic_fetch_add (&pf->rd.in,RTE_PFLOCK_RINC , __ATOMIC_ACQUIRE) + w =rte_atomic_fetch_add_explicit (&pf->rd.in,RTE_PFLOCK_RINC , rte_memory_order_acquire) &RTE_PFLOCK_WBITS ;if (w ==0 )return ;/* Wait for current write phase to complete. */ RTE_WAIT_UNTIL_MASKED(&pf->rd.in, RTE_PFLOCK_WBITS, !=, w, - __ATOMIC_ACQUIRE); + rte_memory_order_acquire); }/**
@@ -140,7 +140,7 @@ struct rte_pflock {static inline void rte_pflock_read_unlock (rte_pflock_t *pf) { -__atomic_fetch_add (&pf->rd.out,RTE_PFLOCK_RINC , __ATOMIC_RELEASE); +rte_atomic_fetch_add_explicit (&pf->rd.out,RTE_PFLOCK_RINC , rte_memory_order_release); }/**
@@ -161,8 +161,9 @@ struct rte_pflock {/* Acquire ownership of write-phase. * This is same as rte_ticketlock_lock(). */ - ticket =__atomic_fetch_add (&pf->wr.in,1 , __ATOMIC_RELAXED); -rte_wait_until_equal_16 (&pf->wr.out, ticket, __ATOMIC_ACQUIRE); + ticket =rte_atomic_fetch_add_explicit (&pf->wr.in,1 , rte_memory_order_relaxed); +rte_wait_until_equal_16 ((uint16_t *)(uintptr_t )&pf->wr.out, ticket, + rte_memory_order_acquire);/* * Acquire ticket on read-side in order to allow them
@@ -173,10 +174,11 @@ struct rte_pflock {* speculatively. */ w =RTE_PFLOCK_PRES | (ticket &RTE_PFLOCK_PHID ); - ticket =__atomic_fetch_add (&pf->rd.in, w, __ATOMIC_RELAXED); + ticket =rte_atomic_fetch_add_explicit (&pf->rd.in, w, rte_memory_order_relaxed);/* Wait for any pending readers to flush. */ -rte_wait_until_equal_16 (&pf->rd.out, ticket, __ATOMIC_ACQUIRE); +rte_wait_until_equal_16 ((uint16_t *)(uintptr_t )&pf->rd.out, ticket, + rte_memory_order_acquire); }/**
@@ -193,10 +195,10 @@ struct rte_pflock { rte_pflock_write_unlock(rte_pflock_t *pf) {/* Migrate from write phase to read phase. */ -__atomic_fetch_and (&pf->rd.in,RTE_PFLOCK_LSB , __ATOMIC_RELEASE); +rte_atomic_fetch_and_explicit (&pf->rd.in,RTE_PFLOCK_LSB , rte_memory_order_release);/* Allow other writers to continue. */ -__atomic_fetch_add (&pf->wr.out,1 , __ATOMIC_RELEASE); +rte_atomic_fetch_add_explicit (&pf->wr.out,1 , rte_memory_order_release); }#ifdef __cplusplus
lib/eal/include/rte_seqcount.h +9 -9
@@ -32,7 +32,7 @@* The RTE seqcount type. */ typedef struct { -uint32_t sn;/**< A sequence number for the protected data. */ +RTE_ATOMIC (uint32_t ) sn;/**< A sequence number for the protected data. */ }rte_seqcount_t ;/**
@@ -106,11 +106,11 @@static inline uint32_t rte_seqcount_read_begin (const rte_seqcount_t *seqcount) { -/* __ATOMIC_ACQUIRE to prevent loads after (in program order) +/* rte_memory_order_acquire to prevent loads after (in program order) * from happening before the sn load. Synchronizes-with the * store release in rte_seqcount_write_end(). */ -return __atomic_load_n (&seqcount->sn, __ATOMIC_ACQUIRE); +return rte_atomic_load_explicit (&seqcount->sn, rte_memory_order_acquire); }/**
@@ -161,9 +161,9 @@return true;/* make sure the data loads happens before the sn load */ -rte_atomic_thread_fence (__ATOMIC_ACQUIRE); +rte_atomic_thread_fence (rte_memory_order_acquire); - end_sn =__atomic_load_n (&seqcount->sn, __ATOMIC_RELAXED); + end_sn =rte_atomic_load_explicit (&seqcount->sn, rte_memory_order_relaxed);/* A writer incremented the sequence number during this read * critical section.
@@ -205,12 +205,12 @@ sn = seqcount->sn +1 ; -__atomic_store_n (&seqcount->sn, sn, __ATOMIC_RELAXED); +rte_atomic_store_explicit (&seqcount->sn, sn, rte_memory_order_relaxed); -/* __ATOMIC_RELEASE to prevent stores after (in program order) +/* rte_memory_order_release to prevent stores after (in program order) * from happening before the sn store. */ -rte_atomic_thread_fence (__ATOMIC_RELEASE); +rte_atomic_thread_fence (rte_memory_order_release); }/**
@@ -237,7 +237,7 @@ sn = seqcount->sn +1 ;/* Synchronizes-with the load acquire in rte_seqcount_read_begin(). */ -__atomic_store_n (&seqcount->sn, sn, __ATOMIC_RELEASE); +rte_atomic_store_explicit (&seqcount->sn, sn, rte_memory_order_release); }#ifdef __cplusplus
lib/eal/include/rte_ticketlock.h +22 -20
115 lines not on this page · raw message
lib/eal/include/rte_trace_point.h +2 -2
@@ -33,7 +33,7 @@#include <rte_stdatomic.h> /** The tracepoint object. */ -typedef uint64_t rte_trace_point_t ; +typedef RTE_ATOMIC (uint64_t )rte_trace_point_t ;/** * Macro to define the tracepoint arguments in RTE_TRACE_POINT macro.
@@ -359,7 +359,7 @@ struct __rte_trace_header {#define __rte_trace_point_emit_header_generic (t) \ void *mem; \ do { \ - const uint64_t val = __atomic_load_n(t, __ATOMIC_ACQUIRE); \ + const uint64_t val = rte_atomic_load_explicit(t, rte_memory_order_acquire); \ if (likely(!(val & __RTE_TRACE_FIELD_ENABLE_MASK))) \ return; \ mem = __rte_trace_mem_get(val); \
lib/eal/loongarch/include/rte_atomic.h +2 -2
@@ -35,9 +35,9 @@#define rte_io_rmb () rte_mb() static __rte_always_inline void -rte_atomic_thread_fence (int memorder) +rte_atomic_thread_fence (rte_memory_order memorder) { -__atomic_thread_fence (memorder); +__rte_atomic_thread_fence (memorder); }#ifdef __cplusplus
lib/eal/ppc/include/rte_atomic.h +27 -27
172 lines not on this page · raw message
lib/eal/riscv/include/rte_atomic.h +2 -2
@@ -40,9 +40,9 @@#define rte_io_rmb () asm volatile("fence ir, ir" : : : "memory") static __rte_always_inline void -rte_atomic_thread_fence (int memorder) +rte_atomic_thread_fence (rte_memory_order memorder) { -__atomic_thread_fence (memorder); +__rte_atomic_thread_fence (memorder); }#ifdef __cplusplus
lib/eal/x86/include/rte_atomic.h +4 -4
@@ -82,17 +82,17 @@/** * Synchronization fence between threads based on the specified memory order. * -* On x86 the __atomic_thread_fence(__ATOMIC_SEQ_CST) generates full 'mfence' +* On x86 the __rte_atomic_thread_fence(rte_memory_order_seq_cst) generates full 'mfence' * which is quite expensive. The optimized implementation of rte_smp_mb is * used instead. */ static __rte_always_inline void -rte_atomic_thread_fence (int memorder) +rte_atomic_thread_fence (rte_memory_order memorder) { -if (memorder == __ATOMIC_SEQ_CST) +if (memorder == rte_memory_order_seq_cst)rte_smp_mb ();else -__atomic_thread_fence (memorder); +__rte_atomic_thread_fence (memorder); }/*------------------------- 16 bit atomic operations -------------------------*/
lib/eal/x86/include/rte_spinlock.h +1 -1
@@ -78,7 +78,7 @@ static inline int rte_tm_supported(void) }static inline int -rte_try_tm (volatile int *lock) +rte_try_tm (volatile RTE_ATOMIC (int ) *lock) {int i, retries;
lib/eal/x86/rte_power_intrinsics.c +3 -3
@@ -23,9 +23,9 @@uint64_t val;/* trigger a write but don't change the value */ - val =__atomic_load_n ((volatile uint64_t *)addr, __ATOMIC_RELAXED); -__atomic_compare_exchange_n ((volatile uint64_t *)addr, &val, val,0 , - __ATOMIC_RELAXED, __ATOMIC_RELAXED); + val =rte_atomic_load_explicit ((volatile uint64_t *)addr, rte_memory_order_relaxed); +rte_atomic_compare_exchange_strong_explicit ((volatile uint64_t *)addr, &val, val, + rte_memory_order_relaxed, rte_memory_order_relaxed); }static bool wait_supported;
Signature · 1 line
1.8.3.1
PATCH v5 3/6 eal: add rte atomic qualifier with casts
In reply to Tyler RetzlaffIntroduce __rte_atomic qualifying casts in rte_optional atomics inline functions to prevent cascading the need to pass __rte_atomic qualified arguments.
Warning, this is really implementation dependent and being done temporarily to avoid having to convert more of the libraries and tests in DPDK in the initial series that introduces the API. The consequence of the assumption of the ABI of the types in question not being ``the same'' is only a risk that may be realized when enable_stdatomic=true.
--- lib/eal/include/generic/rte_atomic.h | 48 ++++++++++++++++++++++++------------ lib/eal/include/generic/rte_pause.h | 9 ++++--- lib/eal/x86/rte_power_intrinsics.c | 7 +++--- 3 files changed, 42 insertions(+), 22 deletions(-)
lib/eal/include/generic/rte_atomic.h +32 -16
@@ -274,7 +274,8 @@static inline void rte_atomic16_add (rte_atomic16_t *v,int16_t inc) { -rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst); + rte_atomic_fetch_add_explicit((volatile __rte_atomic int16_t *)&v->cnt, inc, + rte_memory_order_seq_cst); }/**
@@ -288,7 +289,8 @@static inline void rte_atomic16_sub (rte_atomic16_t *v,int16_t dec) { -rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst); + rte_atomic_fetch_sub_explicit((volatile __rte_atomic int16_t *)&v->cnt, dec, + rte_memory_order_seq_cst); }/**
@@ -341,7 +343,8 @@static inline int16_t rte_atomic16_add_return (rte_atomic16_t *v,int16_t inc) { -return rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst) + inc; + return rte_atomic_fetch_add_explicit((volatile __rte_atomic int16_t *)&v->cnt, inc, + rte_memory_order_seq_cst) + inc; }/**
@@ -361,7 +364,8 @@static inline int16_t rte_atomic16_sub_return (rte_atomic16_t *v,int16_t dec) { -return rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst) - dec; + return rte_atomic_fetch_sub_explicit((volatile __rte_atomic int16_t *)&v->cnt, dec, + rte_memory_order_seq_cst) - dec; }/**
@@ -380,7 +384,8 @@#ifdef RTE_FORCE_INTRINSICS static inline int rte_atomic16_inc_and_test (rte_atomic16_t *v) { -return rte_atomic_fetch_add_explicit (&v->cnt,1 , rte_memory_order_seq_cst) +1 ==0 ; + return rte_atomic_fetch_add_explicit((volatile __rte_atomic int16_t *)&v->cnt, 1, + rte_memory_order_seq_cst) +1 ==0 ; }#endif
@@ -400,7 +405,8 @@ static inline int rte_atomic16_inc_and_test(rte_atomic16_t *v)#ifdef RTE_FORCE_INTRINSICS static inline int rte_atomic16_dec_and_test (rte_atomic16_t *v) { -return rte_atomic_fetch_sub_explicit (&v->cnt,1 , rte_memory_order_seq_cst) -1 ==0 ; + return rte_atomic_fetch_sub_explicit((volatile __rte_atomic int16_t *)&v->cnt, 1, + rte_memory_order_seq_cst) -1 ==0 ; }#endif
@@ -553,7 +559,8 @@ static inline void rte_atomic16_clear(rte_atomic16_t *v)static inline void rte_atomic32_add (rte_atomic32_t *v,int32_t inc) { -rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst); + rte_atomic_fetch_add_explicit((volatile __rte_atomic int32_t *)&v->cnt, inc, + rte_memory_order_seq_cst); }/**
@@ -567,7 +574,8 @@ static inline void rte_atomic16_clear(rte_atomic16_t *v)static inline void rte_atomic32_sub (rte_atomic32_t *v,int32_t dec) { -rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst); + rte_atomic_fetch_sub_explicit((volatile __rte_atomic int32_t *)&v->cnt, dec, + rte_memory_order_seq_cst); }/**
@@ -620,7 +628,8 @@ static inline void rte_atomic16_clear(rte_atomic16_t *v)static inline int32_t rte_atomic32_add_return (rte_atomic32_t *v,int32_t inc) { -return rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst) + inc; + return rte_atomic_fetch_add_explicit((volatile __rte_atomic int32_t *)&v->cnt, inc, + rte_memory_order_seq_cst) + inc; }/**
@@ -640,7 +649,8 @@ static inline void rte_atomic16_clear(rte_atomic16_t *v)static inline int32_t rte_atomic32_sub_return (rte_atomic32_t *v,int32_t dec) { -return rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst) - dec; + return rte_atomic_fetch_sub_explicit((volatile __rte_atomic int32_t *)&v->cnt, dec, + rte_memory_order_seq_cst) - dec; }/**
@@ -659,7 +669,8 @@ static inline void rte_atomic16_clear(rte_atomic16_t *v)#ifdef RTE_FORCE_INTRINSICS static inline int rte_atomic32_inc_and_test (rte_atomic32_t *v) { -return rte_atomic_fetch_add_explicit (&v->cnt,1 , rte_memory_order_seq_cst) +1 ==0 ; + return rte_atomic_fetch_add_explicit((volatile __rte_atomic int32_t *)&v->cnt, 1, + rte_memory_order_seq_cst) +1 ==0 ; }#endif
@@ -679,7 +690,8 @@ static inline int rte_atomic32_inc_and_test(rte_atomic32_t *v)#ifdef RTE_FORCE_INTRINSICS static inline int rte_atomic32_dec_and_test (rte_atomic32_t *v) { -return rte_atomic_fetch_sub_explicit (&v->cnt,1 , rte_memory_order_seq_cst) -1 ==0 ; + return rte_atomic_fetch_sub_explicit((volatile __rte_atomic int32_t *)&v->cnt, 1, + rte_memory_order_seq_cst) -1 ==0 ; }#endif
@@ -885,7 +897,8 @@ static inline void rte_atomic32_clear(rte_atomic32_t *v)static inline void rte_atomic64_add (rte_atomic64_t *v,int64_t inc) { -rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst); + rte_atomic_fetch_add_explicit((volatile __rte_atomic int64_t *)&v->cnt, inc, + rte_memory_order_seq_cst); }#endif
@@ -904,7 +917,8 @@ static inline void rte_atomic32_clear(rte_atomic32_t *v)static inline void rte_atomic64_sub (rte_atomic64_t *v,int64_t dec) { -rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst); + rte_atomic_fetch_sub_explicit((volatile __rte_atomic int64_t *)&v->cnt, dec, + rte_memory_order_seq_cst); }#endif
@@ -962,7 +976,8 @@ static inline void rte_atomic32_clear(rte_atomic32_t *v)static inline int64_t rte_atomic64_add_return (rte_atomic64_t *v,int64_t inc) { -return rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst) + inc; + return rte_atomic_fetch_add_explicit((volatile __rte_atomic int64_t *)&v->cnt, inc, + rte_memory_order_seq_cst) + inc; }#endif
@@ -986,7 +1001,8 @@ static inline void rte_atomic32_clear(rte_atomic32_t *v)static inline int64_t rte_atomic64_sub_return (rte_atomic64_t *v,int64_t dec) { -return rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst) - dec; + return rte_atomic_fetch_sub_explicit((volatile __rte_atomic int64_t *)&v->cnt, dec, + rte_memory_order_seq_cst) - dec; }#endif
lib/eal/include/generic/rte_pause.h +6 -3
@@ -81,7 +81,8 @@ {assert (memorder == rte_memory_order_acquire || memorder == rte_memory_order_relaxed); -while (rte_atomic_load_explicit (addr, memorder) != expected) + while (rte_atomic_load_explicit((volatile __rte_atomic uint16_t *)addr, memorder) + != expected)rte_pause (); }
@@ -91,7 +92,8 @@ {assert (memorder == rte_memory_order_acquire || memorder == rte_memory_order_relaxed); -while (rte_atomic_load_explicit (addr, memorder) != expected) + while (rte_atomic_load_explicit((volatile __rte_atomic uint32_t *)addr, memorder) + != expected)rte_pause (); }
@@ -101,7 +103,8 @@ {assert (memorder == rte_memory_order_acquire || memorder == rte_memory_order_relaxed); -while (rte_atomic_load_explicit (addr, memorder) != expected) + while (rte_atomic_load_explicit((volatile __rte_atomic uint64_t *)addr, memorder) + != expected)rte_pause (); }
lib/eal/x86/rte_power_intrinsics.c +4 -3
@@ -23,9 +23,10 @@uint64_t val;/* trigger a write but don't change the value */ - val =rte_atomic_load_explicit ((volatile uint64_t *)addr, rte_memory_order_relaxed); -rte_atomic_compare_exchange_strong_explicit ((volatile uint64_t *)addr, &val, val, - rte_memory_order_relaxed, rte_memory_order_relaxed); + val = rte_atomic_load_explicit((volatile __rte_atomic uint64_t *)addr, + rte_memory_order_relaxed); + rte_atomic_compare_exchange_strong_explicit((volatile __rte_atomic uint64_t *)addr, + &val, val, rte_memory_order_relaxed, rte_memory_order_relaxed); }static bool wait_supported;
Signature · 1 line
1.8.3.1
PATCH v5 4/6 distributor: adapt for EAL optional atomics API changes
In reply to Tyler RetzlaffAdapt distributor for EAL optional atomics API changes
--- lib/distributor/distributor_private.h | 2 +- lib/distributor/rte_distributor_single.c | 44 ++++++++++++++++---------------- 2 files changed, 23 insertions(+), 23 deletions(-)
lib/distributor/distributor_private.h +1 -1
@@ -52,7 +52,7 @@* Only 64-bits of the memory is actually used though. */ union rte_distributor_buffer_single { -volatile int64_t bufptr64; +volatile RTE_ATOMIC (int64_t ) bufptr64;char pad[RTE_CACHE_LINE_SIZE *3 ]; } __rte_cache_aligned;
lib/distributor/rte_distributor_single.c +22 -22
117 lines not on this page · raw message
Signature · 1 line
1.8.3.1
PATCH v5 5/6 bpf: adapt for EAL optional atomics API changes
In reply to Tyler RetzlaffAdapt bpf for EAL optional atomics API changes
--- lib/bpf/bpf_pkt.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)
lib/bpf/bpf_pkt.c +3 -3
@@ -25,7 +25,7 @@struct bpf_eth_cbi {/* used by both data & control path */ -uint32_t use;/*usage counter */ +RTE_ATOMIC (uint32_t ) use;/*usage counter */ const struct rte_eth_rxtx_callback *cb;/* callback handle */ struct rte_bpf *bpf;struct rte_bpf_jit jit;
@@ -110,8 +110,8 @@ struct bpf_eth_cbh {/* in use, busy wait till current RX/TX iteration is finished */ if ((puse &BPF_ETH_CBI_INUSE ) !=0 ) { -RTE_WAIT_UNTIL_MASKED ((uint32_t *)(uintptr_t )&cbi->use, - UINT32_MAX, !=, puse, __ATOMIC_RELAXED); + RTE_WAIT_UNTIL_MASKED((__rte_atomic uint32_t *)(uintptr_t)&cbi->use, + UINT32_MAX, !=, puse, rte_memory_order_relaxed); } }
Signature · 1 line
1.8.3.1
PATCH v5 6/6 devtools: forbid new direct use of GCC atomic builtins
In reply to Tyler RetzlaffRefrain from using compiler __atomic_xxx builtins DPDK now requires the use of rte_atomic_<op>_explicit macros when operating on DPDK atomic variables.
--- devtools/checkpatches.sh | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-)
devtools/checkpatches.sh +4 -4
@@ -107,15 +107,15 @@ check_forbidden_additions() { # <patch> awk -v FOLDERS="lib drivers app examples" \ -v EXPRESSIONS="__atomic_thread_fence\\\(" \ -v RET_ON_FAIL=1 \ - -v MESSAGE='Using __atomic_thread_fence' \ + -v MESSAGE='Using __atomic_thread_fence built-in, prefer rte_atomic_thread_fence' \ -f $(dirname $(readlink -f $0))/check-forbidden-tokens.awk \ "$1" || res=1 - # refrain from using compiler __atomic_{add,and,nand,or,sub,xor}_fetch() + # refrain from using compiler __atomic_xxx builtins awk -v FOLDERS="lib drivers app examples" \ - -v EXPRESSIONS="__atomic_(add|and|nand|or|sub|xor)_fetch\\\(" \ + -v EXPRESSIONS="__atomic_.*\\\(" \ -v RET_ON_FAIL=1 \ - -v MESSAGE='Using __atomic_op_fetch, prefer __atomic_fetch_op' \ + -v MESSAGE='Using __atomic_xxx built-ins, prefer rte_atomic_xxx' \ -f $(dirname $(readlink -f $0))/check-forbidden-tokens.awk \ "$1" || res=1
Signature · 1 line
1.8.3.1
Re: [PATCH v5 0/6] optional rte optional stdatomics API
In reply to Tyler RetzlaffQuoted 117 lines
This series introduces API additions prefixed in the rte namespace that allow the optional use of
stdatomics.hfrom C11 using enable_stdatomics=true for targets where enable_stdatomics=false no functional change is intended.Be aware this does not contain all changes to use stdatomics across the DPDK tree it only introduces the minimum to allow the option to be used which is a pre-requisite for a clean CI (probably using clang) that can be run with enable_stdatomics=true enabled.
It is planned that subsequent series will be introduced per lib/driver as appropriate to further enable stdatomics use when enable_stdatomics=true.
Notes:
* Additional libraries beyond EAL make visible atomics use across the API/ABI surface they will be converted in the subsequent series.
* The eal: add rte atomic qualifier with casts patch needs some discussion as to whether or not the legacy
rte_atomicAPIs should be converted to work with enable_stdatomic=true right now some implementation dependent casts are used to prevent cascading / having to convert too much in the intial series.* Windows will obviously need complete conversion of libraries including atomics that are not crossing API/ABI boundaries. those conversions will introduced in separate series as new along side the existing msvc series.
Please keep in mind we would like to prioritize the review / acceptance of this patch since it needs to be completed in the 23.11 merge window.
Thank you all for the discussion that lead to the formation of this series.
v5:
* AddRTE_ATOMICto doxygen configuration PREDEFINED macros list to fix documentation generation failure
* Fix two typos in expansion of C11 atomics macros strong -> weak and add missing_explicit
* Adjust devtools/checkpatches messages based on feedback. i have chosen not to try and catch use of C11 atomics or_Atomicsince using those directly will be picked up by existing CI passes where by compilation error where enable_stdatomic=false (the default for most platforms)v4:
* Move the definition of #defineRTE_ATOMIC(type)to patch 1 where it belongs (a mistake in v3)
* Provide comments for bothRTE_ATOMICand__rte_atomicmacros indicating their use as specified or qualified contexts.v3:
* Remove comments from APIs mentioning the mapping to C++ memory model memory orders
* Introduce and use new macroRTE_ATOMIC(type)to be used in contexts where_Atomicis used as a type specifier to declare variables. The macro allows more clarity about what the atomic type being specified is. e.g._Atomic(T *)vs_Atomic(T)it is easier to understand that the former is an atomic pointer type and the latter is an atomic type. it also has the benefit of (in the future) being interoperable with c++23 syntactically
note: Morten i have retained your 'reviewed-by' tags if you disagree given the changes in the above version please indicate as such but i believe the changes are in the spirit of the feedback you providedv2:
* Wrap meson_options.txt option description to newline and indent to be consistent with other options.
* Provide separate typedef ofrte_memory_orderfor enable_stdatomic=true VS enable_stdatomic=false instead of a single typedef to int
note: slight tweak to reviewers feedback i've chosen to use a typedef
for both enable_stdatomic={true,false} (just seemed more consistent)
* Bring inassert.hand usestatic_assertmacro instead of_Static_assertkeyword to better interoperate with c/c++
* Directly includerte_stdatomic.hwhere into other places it is consumed instead of hacking it globally intorte_config.h
* Provide and use__rte_atomic_thread_fenceto allow conditional expansion within the body of existingrte_atomic_thread_fenceinline function to maintain per-arch optimizations when enable_stdatomic=falseTyler Retzlaff (6):
eal: provide rte stdatomics optional atomics API
eal: adapt EAL to present rte optional atomics API
eal: add rte atomic qualifier with casts
distributor: adapt for EAL optional atomics API changes
bpf: adapt for EAL optional atomics API changes
devtools: forbid new direct use of GCC atomic builtinsapp/test/test_mcslock.c | 6 +- config/meson.build | 1 + devtools/checkpatches.sh | 8 +- doc/api/doxy-api.conf.in | 1 + lib/bpf/bpf_pkt.c | 6 +- lib/distributor/distributor_private.h | 2 +- lib/distributor/rte_distributor_single.c | 44 +++---- lib/eal/arm/include/rte_atomic_32.h | 4 +- lib/eal/arm/include/rte_atomic_64.h | 36 +++--- lib/eal/arm/include/rte_pause_64.h | 26 ++-- lib/eal/arm/rte_power_intrinsics.c | 8 +- lib/eal/common/eal_common_trace.c | 16 +-- lib/eal/include/generic/rte_atomic.h | 67 +++++++---- lib/eal/include/generic/rte_pause.h | 50 ++++---- lib/eal/include/generic/rte_rwlock.h | 48 ++++---- lib/eal/include/generic/rte_spinlock.h | 20 ++-- lib/eal/include/meson.build | 1 + lib/eal/include/rte_mcslock.h | 51 ++++---- lib/eal/include/rte_pflock.h | 25 ++-- lib/eal/include/rte_seqcount.h | 19 +-- lib/eal/include/rte_stdatomic.h | 198 +++++++++++++++++++++++++++++++ lib/eal/include/rte_ticketlock.h | 43 +++---- lib/eal/include/rte_trace_point.h | 5 +- lib/eal/loongarch/include/rte_atomic.h | 4 +- lib/eal/ppc/include/rte_atomic.h | 54 ++++----- lib/eal/riscv/include/rte_atomic.h | 4 +- lib/eal/x86/include/rte_atomic.h | 8 +- lib/eal/x86/include/rte_spinlock.h | 2 +- lib/eal/x86/rte_power_intrinsics.c | 7 +- meson_options.txt | 2 + 30 files changed, 499 insertions(+), 267 deletions(-)
Quoted 2 lines
create mode 100644
lib/eal/include/rte_stdatomic.h
v6 cover rte atomics API for optional stdatomic
In reply to Tyler RetzlaffThis series introduces API additions prefixed in the rte namespace that allow the optional use of stdatomics.h from C11 using enable_stdatomics=true for targets where enable_stdatomics=false no functional change is intended.
Be aware this does not contain all changes to use stdatomics across the DPDK tree it only introduces the minimum to allow the option to be used which is a pre-requisite for a clean CI (probably using clang) that can be run with enable_stdatomics=true enabled.
It is planned that subsequent series will be introduced per lib/driver as appropriate to further enable stdatomics use when enable_stdatomics=true.
Notes:
* Additional libraries beyond EAL make visible atomics use across the API/ABI surface they will be converted in the subsequent series.
* The eal: add rte atomic qualifier with casts patch needs some discussion as to whether or not the legacy rte_atomic APIs should be converted to work with enable_stdatomic=true right now some implementation dependent casts are used to prevent cascading / having to convert too much in the intial series.
* Windows will obviously need complete conversion of libraries including atomics that are not crossing API/ABI boundaries. those conversions will introduced in separate series as new along side the existing msvc series.
Please keep in mind we would like to prioritize the review / acceptance of this patch since it needs to be completed in the 23.11 merge window.
Thank you all for the discussion that lead to the formation of this series.
v6:
* Adjust checkpatches to warn about use of __rte_atomic_thread_fence and suggest use of rte_atomic_thread_fence. Use the existing check more generic check for __atomic_xxx to catch use of __atomic_thread_fence and recommend rte_atomic_xxx.
v5:
* Add RTE_ATOMIC to doxygen configuration PREDEFINED macros list to fix documentation generation failure
* Fix two typos in expansion of C11 atomics macros strong -> weak and add missing _explicit
* Adjust devtools/checkpatches messages based on feedback. i have chosen not to try and catch use of C11 atomics or _Atomic since using those directly will be picked up by existing CI passes where by compilation error where enable_stdatomic=false (the default for most platforms)
v4:
* Move the definition of #define RTE_ATOMIC(type) to patch 1 where it belongs (a mistake in v3)
* Provide comments for both RTE_ATOMIC and __rte_atomic macros indicating their use as specified or qualified contexts.
v3:
* Remove comments from APIs mentioning the mapping to C++ memory model memory orders
* Introduce and use new macro RTE_ATOMIC(type) to be used in contexts where _Atomic is used as a type specifier to declare variables. The macro allows more clarity about what the atomic type being specified is. e.g. _Atomic(T *) vs _Atomic(T) it is easier to understand that the former is an atomic pointer type and the latter is an atomic type. it also has the benefit of (in the future) being interoperable with c++23 syntactically
note: Morten i have retained your 'reviewed-by' tags if you disagree given the changes in the above version please indicate as such but i believe the changes are in the spirit of the feedback you provided
v2:
* Wrap meson_options.txt option description to newline and indent to be consistent with other options.
* Provide separate typedef of rte_memory_order for enable_stdatomic=true VS enable_stdatomic=false instead of a single typedef to int
note: slight tweak to reviewers feedback i've chosen to use a typedef
for both enable_stdatomic={true,false} (just seemed more consistent)
* Bring in assert.h and use static_assert macro instead of _Static_assert keyword to better interoperate with c/c++
* Directly include rte_stdatomic.h where into other places it is consumed instead of hacking it globally into rte_config.h
* Provide and use __rte_atomic_thread_fence to allow conditional expansion within the body of existing rte_atomic_thread_fence inline function to maintain per-arch optimizations when enable_stdatomic=false
Tyler Retzlaff (6):
eal: provide rte stdatomics optional atomics API
eal: adapt EAL to present rte optional atomics API
eal: add rte atomic qualifier with casts
distributor: adapt for EAL optional atomics API changes
bpf: adapt for EAL optional atomics API changes
devtools: forbid new direct use of GCC atomic builtins
app/test/test_mcslock.c | 6 +- config/meson.build | 1 + devtools/checkpatches.sh | 12 +- doc/api/doxy-api.conf.in | 1 + lib/bpf/bpf_pkt.c | 6 +- lib/distributor/distributor_private.h | 2 +- lib/distributor/rte_distributor_single.c | 44 +++---- lib/eal/arm/include/rte_atomic_32.h | 4 +- lib/eal/arm/include/rte_atomic_64.h | 36 +++--- lib/eal/arm/include/rte_pause_64.h | 26 ++-- lib/eal/arm/rte_power_intrinsics.c | 8 +- lib/eal/common/eal_common_trace.c | 16 +-- lib/eal/include/generic/rte_atomic.h | 67 +++++++---- lib/eal/include/generic/rte_pause.h | 50 ++++---- lib/eal/include/generic/rte_rwlock.h | 48 ++++---- lib/eal/include/generic/rte_spinlock.h | 20 ++-- lib/eal/include/meson.build | 1 + lib/eal/include/rte_mcslock.h | 51 ++++---- lib/eal/include/rte_pflock.h | 25 ++-- lib/eal/include/rte_seqcount.h | 19 +-- lib/eal/include/rte_stdatomic.h | 198 +++++++++++++++++++++++++++++++ lib/eal/include/rte_ticketlock.h | 43 +++---- lib/eal/include/rte_trace_point.h | 5 +- lib/eal/loongarch/include/rte_atomic.h | 4 +- lib/eal/ppc/include/rte_atomic.h | 54 ++++----- lib/eal/riscv/include/rte_atomic.h | 4 +- lib/eal/x86/include/rte_atomic.h | 8 +- lib/eal/x86/include/rte_spinlock.h | 2 +- lib/eal/x86/rte_power_intrinsics.c | 7 +- meson_options.txt | 2 + 30 files changed, 501 insertions(+), 269 deletions(-) create mode 100644 lib/eal/include/rte_stdatomic.h
Signature · 1 line
1.8.3.1
PATCH v6 1/6 eal: provide rte stdatomics optional atomics API
Provide API for atomic operations in the rte namespace that may optionally be configured to use C11 atomics with meson option enable_stdatomics=true
--- config/meson.build | 1 + doc/api/doxy-api.conf.in | 1 + lib/eal/include/generic/rte_atomic.h | 1 + lib/eal/include/generic/rte_pause.h | 1 + lib/eal/include/generic/rte_rwlock.h | 1 + lib/eal/include/generic/rte_spinlock.h | 1 + lib/eal/include/meson.build | 1 + lib/eal/include/rte_mcslock.h | 1 + lib/eal/include/rte_pflock.h | 1 + lib/eal/include/rte_seqcount.h | 1 + lib/eal/include/rte_stdatomic.h | 198 +++++++++++++++++++++++++++++++++ lib/eal/include/rte_ticketlock.h | 1 + lib/eal/include/rte_trace_point.h | 1 + meson_options.txt | 2 + 14 files changed, 212 insertions(+) create mode 100644 lib/eal/include/rte_stdatomic.h
config/meson.build +1 -0
@@ -303,6 +303,7 @@ endforeach # set other values pulled from the build options dpdk_conf.set('RTE_MAX_ETHPORTS', get_option('max_ethports')) dpdk_conf.set('RTE_LIBEAL_USE_HPET', get_option('use_hpet')) +dpdk_conf.set('RTE_ENABLE_STDATOMIC', get_option('enable_stdatomic')) dpdk_conf.set('RTE_ENABLE_TRACE_FP', get_option('enable_trace_fp')) # values which have defaults which may be overridden dpdk_conf.set('RTE_MAX_VFIO_GROUPS', 64)
doc/api/doxy-api.conf.in +1 -0
@@ -84,6 +84,7 @@ INPUT += @API_EXAMPLES@ FILE_PATTERNS = rte_*.h \ cmdline.h PREDEFINED = __DOXYGEN__ \ + RTE_ATOMIC \ RTE_HAS_CPUSET \ VFIO_PRESENT \ __rte_lockable= \
lib/eal/include/generic/rte_atomic.h +1 -0
@@ -15,6 +15,7 @@#include <stdint.h> #include <rte_compat.h> #include <rte_common.h> +#include <rte_stdatomic.h> #ifdef __DOXYGEN__
lib/eal/include/generic/rte_pause.h +1 -0
@@ -16,6 +16,7 @@#include <assert.h> #include <rte_common.h> #include <rte_atomic.h> +#include <rte_stdatomic.h> /** * Pause CPU execution for a short while
lib/eal/include/generic/rte_rwlock.h +1 -0
@@ -32,6 +32,7 @@#include <rte_common.h> #include <rte_lock_annotations.h> #include <rte_pause.h> +#include <rte_stdatomic.h> /** * The rte_rwlock_t type.
lib/eal/include/generic/rte_spinlock.h +1 -0
@@ -23,6 +23,7 @@#endif #include <rte_lock_annotations.h> #include <rte_pause.h> +#include <rte_stdatomic.h> /** * The rte_spinlock_t type.
lib/eal/include/meson.build +1 -0
@@ -42,6 +42,7 @@ headers += files( 'rte_seqlock.h', 'rte_service.h', 'rte_service_component.h', + 'rte_stdatomic.h', 'rte_string_fns.h', 'rte_tailq.h', 'rte_thread.h',
lib/eal/include/rte_mcslock.h +1 -0
@@ -27,6 +27,7 @@#include <rte_common.h> #include <rte_pause.h> #include <rte_branch_prediction.h> +#include <rte_stdatomic.h> /** * The rte_mcslock_t type.
lib/eal/include/rte_pflock.h +1 -0
@@ -34,6 +34,7 @@#include <rte_compat.h> #include <rte_common.h> #include <rte_pause.h> +#include <rte_stdatomic.h> /** * The rte_pflock_t type.
lib/eal/include/rte_seqcount.h +1 -0
@@ -26,6 +26,7 @@#include <rte_atomic.h> #include <rte_branch_prediction.h> #include <rte_compat.h> +#include <rte_stdatomic.h> /** * The RTE seqcount type.
lib/eal/include/rte_stdatomic.h new file +198 -0
198 lines not on this page · raw message
lib/eal/include/rte_ticketlock.h +1 -0
@@ -24,6 +24,7 @@#include <rte_common.h> #include <rte_lcore.h> #include <rte_pause.h> +#include <rte_stdatomic.h> /** * The rte_ticketlock_t type.
lib/eal/include/rte_trace_point.h +1 -0
@@ -30,6 +30,7 @@#include <rte_per_lcore.h> #include <rte_string_fns.h> #include <rte_uuid.h> +#include <rte_stdatomic.h> /** The tracepoint object. */ typedef uint64_t rte_trace_point_t ;
meson_options.txt +2 -0
@@ -46,6 +46,8 @@ option('mbuf_refcnt_atomic', type: 'boolean', value: true, description: 'Atomically access the mbuf refcnt.') option('platform', type: 'string', value: 'native', description: 'Platform to build, either "native", "generic" or a SoC. Please refer to the Linux build guide for more information.') +option('enable_stdatomic', type: 'boolean', value: false, description: + 'enable use of C11 stdatomic') option('enable_trace_fp', type: 'boolean', value: false, description: 'enable fast path trace points.') option('tests', type: 'boolean', value: true, description:
Signature · 1 line
1.8.3.1
Re: [PATCH v6 1/6] eal: provide rte stdatomics optional atomics API
22/08/2023 23:00, Tyler Retzlaff:
lib/eal/include/generic/rte_rwlock.h line 35
Quoted 7 lines
lib/eal/include/generic/rte_rwlock.h +1
-0@@ -32,6 +32,7 @@#include <rte_common.h> #include <rte_lock_annotations.h> #include <rte_pause.h> +#include <rte_stdatomic.h>
I'm not sure about adding the include in patch 1 if it is not used here.
Quoted 20 lines
lib/eal/include/rte_stdatomic.h new file +17
-0@@ -0,0 +1,198 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2023 Microsoft Corporation +*/ + +#ifndef _RTE_STDATOMIC_H_ +#define _RTE_STDATOMIC_H_ + +#include <assert.h> + +#ifdef __cplusplus +extern "C" { +#endif + +#ifdef RTE_ENABLE_STDATOMIC +#ifdef __STDC_NO_ATOMICS__ +#error enable_stdatomics=true but atomics not supported by toolchain +#endif
lib/eal/include/rte_stdatomic.h line 27
Quoted 10 lines
lib/eal/include/rte_stdatomic.h new file +10
-0+ +#include <stdatomic.h> + +/* RTE_ATOMIC(type) is provided for use as a type specifier +* permitting designation of an rte atomic type. +*/ +#define RTE_ATOMIC (type) _Atomic(type) + +/* __rte_atomic is provided for type qualification permitting +* designation of an rte atomic qualified type-name.
Sorry I don't understand this comment.
lib/eal/include/rte_stdatomic.h line 37
Quoted 10 lines
lib/eal/include/rte_stdatomic.h new file +10
-0+*/ +#define __rte_atomic _Atomic + +/* The memory order is an enumerated type in C11. */ +typedef memory_order rte_memory_order ; + +#define rte_memory_order_relaxed memory_order_relaxed +#ifdef __ATOMIC_RELAXED +static_assert (rte_memory_order_relaxed == __ATOMIC_RELAXED, +"rte_memory_order_relaxed == __ATOMIC_RELAXED" );
Not sure about using static_assert or RTE_BUILD_BUG_ON
Quoted 72 lines
lib/eal/include/rte_stdatomic.h new file +72
-0+#endif + +#define rte_memory_order_consume memory_order_consume +#ifdef __ATOMIC_CONSUME +static_assert (rte_memory_order_consume == __ATOMIC_CONSUME, +"rte_memory_order_consume == __ATOMIC_CONSUME" ); +#endif + +#define rte_memory_order_acquire memory_order_acquire +#ifdef __ATOMIC_ACQUIRE +static_assert (rte_memory_order_acquire == __ATOMIC_ACQUIRE, +"rte_memory_order_acquire == __ATOMIC_ACQUIRE" ); +#endif + +#define rte_memory_order_release memory_order_release +#ifdef __ATOMIC_RELEASE +static_assert (rte_memory_order_release == __ATOMIC_RELEASE, +"rte_memory_order_release == __ATOMIC_RELEASE" ); +#endif + +#define rte_memory_order_acq_rel memory_order_acq_rel +#ifdef __ATOMIC_ACQ_REL +static_assert (rte_memory_order_acq_rel == __ATOMIC_ACQ_REL, +"rte_memory_order_acq_rel == __ATOMIC_ACQ_REL" ); +#endif + +#define rte_memory_order_seq_cst memory_order_seq_cst +#ifdef __ATOMIC_SEQ_CST +static_assert (rte_memory_order_seq_cst == __ATOMIC_SEQ_CST, +"rte_memory_order_seq_cst == __ATOMIC_SEQ_CST" ); +#endif + +#define rte_atomic_load_explicit (ptr, memorder) \ + atomic_load_explicit(ptr, memorder) + +#define rte_atomic_store_explicit (ptr, val, memorder) \ + atomic_store_explicit(ptr, val, memorder) + +#define rte_atomic_exchange_explicit (ptr, val, memorder) \ + atomic_exchange_explicit(ptr, val, memorder) + +#define rte_atomic_compare_exchange_strong_explicit ( \ + ptr, expected, desired, succ_memorder, fail_memorder) \ + atomic_compare_exchange_strong_explicit( \ + ptr, expected, desired, succ_memorder, fail_memorder) + +#define rte_atomic_compare_exchange_weak_explicit ( \ + ptr, expected, desired, succ_memorder, fail_memorder) \ + atomic_compare_exchange_weak_explicit( \ + ptr, expected, desired, succ_memorder, fail_memorder) + +#define rte_atomic_fetch_add_explicit (ptr, val, memorder) \ + atomic_fetch_add_explicit(ptr, val, memorder) + +#define rte_atomic_fetch_sub_explicit (ptr, val, memorder) \ + atomic_fetch_sub_explicit(ptr, val, memorder) + +#define rte_atomic_fetch_and_explicit (ptr, val, memorder) \ + atomic_fetch_and_explicit(ptr, val, memorder) + +#define rte_atomic_fetch_xor_explicit (ptr, val, memorder) \ + atomic_fetch_xor_explicit(ptr, val, memorder) + +#define rte_atomic_fetch_or_explicit (ptr, val, memorder) \ + atomic_fetch_or_explicit(ptr, val, memorder) + +#define rte_atomic_fetch_nand_explicit (ptr, val, memorder) \ + atomic_fetch_nand_explicit(ptr, val, memorder) + +#define rte_atomic_flag_test_and_set_explicit (ptr, memorder) \ + atomic_flag_test_and_set_explicit(ptr, memorder) +
lib/eal/include/rte_stdatomic.h line 119
Quoted 10 lines
lib/eal/include/rte_stdatomic.h new file +10
-0+#define rte_atomic_flag_clear_explicit (ptr, memorder) \ + atomic_flag_clear_explicit(ptr, memorder) + +/* We provide internal macro here to allow conditional expansion +* in the body of the per-arch rte_atomic_thread_fence inline functions. +*/ +#define __rte_atomic_thread_fence (memorder) \ + atomic_thread_fence(memorder) + +#else
Better to add some context in comment of this "else": /* !RTE_ENABLE_STDATOMIC */
lib/eal/include/rte_stdatomic.h line 23
Quoted 4 lines
lib/eal/include/rte_stdatomic.h new file +4
-0+ +/* RTE_ATOMIC(type) is provided for use as a type specifier +* permitting designation of an rte atomic type. +*/
The comment should say it has no effect.
Or no comment at all for this part.
Quoted 75 lines
lib/eal/include/rte_stdatomic.h new file +75
-0+#define RTE_ATOMIC (type) type + +/* __rte_atomic is provided for type qualification permitting +* designation of an rte atomic qualified type-name. +*/ +#define __rte_atomic + +/* The memory order is an integer type in GCC built-ins, +* not an enumerated type like in C11. +*/ +typedef int rte_memory_order ; + +#define rte_memory_order_relaxed __ATOMIC_RELAXED +#define rte_memory_order_consume __ATOMIC_CONSUME +#define rte_memory_order_acquire __ATOMIC_ACQUIRE +#define rte_memory_order_release __ATOMIC_RELEASE +#define rte_memory_order_acq_rel __ATOMIC_ACQ_REL +#define rte_memory_order_seq_cst __ATOMIC_SEQ_CST + +#define rte_atomic_load_explicit (ptr, memorder) \ + __atomic_load_n(ptr, memorder) + +#define rte_atomic_store_explicit (ptr, val, memorder) \ + __atomic_store_n(ptr, val, memorder) + +#define rte_atomic_exchange_explicit (ptr, val, memorder) \ + __atomic_exchange_n(ptr, val, memorder) + +#define rte_atomic_compare_exchange_strong_explicit ( \ + ptr, expected, desired, succ_memorder, fail_memorder) \ + __atomic_compare_exchange_n( \ + ptr, expected, desired, 0, succ_memorder, fail_memorder) + +#define rte_atomic_compare_exchange_weak_explicit ( \ + ptr, expected, desired, succ_memorder, fail_memorder) \ + __atomic_compare_exchange_n( \ + ptr, expected, desired, 1, succ_memorder, fail_memorder) + +#define rte_atomic_fetch_add_explicit (ptr, val, memorder) \ + __atomic_fetch_add(ptr, val, memorder) + +#define rte_atomic_fetch_sub_explicit (ptr, val, memorder) \ + __atomic_fetch_sub(ptr, val, memorder) + +#define rte_atomic_fetch_and_explicit (ptr, val, memorder) \ + __atomic_fetch_and(ptr, val, memorder) + +#define rte_atomic_fetch_xor_explicit (ptr, val, memorder) \ + __atomic_fetch_xor(ptr, val, memorder) + +#define rte_atomic_fetch_or_explicit (ptr, val, memorder) \ + __atomic_fetch_or(ptr, val, memorder) + +#define rte_atomic_fetch_nand_explicit (ptr, val, memorder) \ + __atomic_fetch_nand(ptr, val, memorder) + +#define rte_atomic_flag_test_and_set_explicit (ptr, memorder) \ + __atomic_test_and_set(ptr, memorder) + +#define rte_atomic_flag_clear_explicit (ptr, memorder) \ + __atomic_clear(ptr, memorder) + +/* We provide internal macro here to allow conditional expansion +* in the body of the per-arch rte_atomic_thread_fence inline functions. +*/ +#define __rte_atomic_thread_fence (memorder) \ + __atomic_thread_fence(memorder) + +#endif + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_STDATOMIC_H_ */
Re: [PATCH v6 1/6] eal: provide rte stdatomics optional atomics API
Quoted 11 lines · Thomas Monjalon
22/08/2023 23:00, Tyler Retzlaff:
lib/eal/include/generic/rte_rwlock.h line 35
Quoted 7 lines
lib/eal/include/generic/rte_rwlock.h +1
-0@@ -32,6 +32,7 @@#include <rte_common.h> #include <rte_lock_annotations.h> #include <rte_pause.h> +#include <rte_stdatomic.h> I'm not sure about adding the include in patch 1 if it is not used here.
Yes, this is something I had already fixed locally.
Quoted 25 lines
Quoted 25 lines
lib/eal/include/rte_stdatomic.h new file +21
-0@@ -0,0 +1,198 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2023 Microsoft Corporation +*/ + +#ifndef _RTE_STDATOMIC_H_ +#define _RTE_STDATOMIC_H_ + +#include <assert.h> + +#ifdef __cplusplus +extern "C" { +#endif + +#ifdef RTE_ENABLE_STDATOMIC +#ifdef __STDC_NO_ATOMICS__ +#error enable_stdatomics=true but atomics not supported by toolchain +#endif + +#include <stdatomic.h> + +/* RTE_ATOMIC(type) is provided for use as a type specifier
Quoted 8 lines
lib/eal/include/rte_stdatomic.h line 27
Quoted 6 lines
lib/eal/include/rte_stdatomic.h new file +6
-0+* permitting designation of an rte atomic type. +*/ +#define RTE_ATOMIC (type) _Atomic(type) + +/* __rte_atomic is provided for type qualification permitting +* designation of an rte atomic qualified type-name. Sorry I don't understand this comment.
The difference between atomic qualifier and atomic specifier and the need for exposing those two notions are not obvious to me.
One clue I have is with one use later in the series:
+rte_mcslock_lock(RTE_ATOMIC(rte_mcslock_t *) *msl, rte_mcslock_t *me)
...
+ prev = rte_atomic_exchange_explicit(msl, me, rte_memory_order_acq_rel);
So at least RTE_ATOMIC() seems necessary.
Quoted 13 lines
lib/eal/include/rte_stdatomic.h line 37
Quoted 11 lines
lib/eal/include/rte_stdatomic.h new file +10
-0+*/ +#define __rte_atomic _Atomic + +/* The memory order is an enumerated type in C11. */ +typedef memory_order rte_memory_order ; + +#define rte_memory_order_relaxed memory_order_relaxed +#ifdef __ATOMIC_RELAXED +static_assert (rte_memory_order_relaxed == __ATOMIC_RELAXED, +"rte_memory_order_relaxed == __ATOMIC_RELAXED" );Not sure about using
static_assertorRTE_BUILD_BUG_ON
Do you mean you want no check at all in a public facing header?
Or is it that we have RTE_BUILD_BUG_ON and we should keep using it instead of static_assert?
I remember some problems with RTE_BUILD_BUG_ON where the compiler would silently drop the whole expression and reported no problem as it could not evaluate the expression.
At least, with static_assert (iirc, it is new to C11) the compiler complains with a clear "error: expression in static assertion is not constant".
We could fix RTE_BUILD_BUG_ON, but I guess the fix would be equivalent to map it to static_assert(!condition).
Using language standard constructs seems a better choice.
Signature · 1 line
David Marchand
RE: [PATCH v6 1/6] eal: provide rte stdatomics optional atomics API
Quoted 6 lines
From: David Marchand [mailto:[email protected]]
Sent: Friday, 29 September 2023 10.04Quoted 2 lines · Thomas Monjalon
22/08/2023 23:00, Tyler Retzlaff:
[...]
Quoted 15 lines
Quoted 9 lines
lib/eal/include/rte_stdatomic.h line 37
Quoted 7 lines
lib/eal/include/rte_stdatomic.h new file +7
-0+/* The memory order is an enumerated type in C11. */ +typedef memory_order rte_memory_order ; + +#define rte_memory_order_relaxed memory_order_relaxed +#ifdef __ATOMIC_RELAXED +static_assert (rte_memory_order_relaxed == __ATOMIC_RELAXED, +"rte_memory_order_relaxed == __ATOMIC_RELAXED" );Not sure about using
static_assertorRTE_BUILD_BUG_ONDo you mean you want no check at all in a public facing header?
Or is it that we have
RTE_BUILD_BUG_ONand we should keep using it instead ofstatic_assert?
Quoted 9 lines
I remember some problems with
RTE_BUILD_BUG_ONwhere the compiler would silently drop the whole expression and reported no problem as it could not evaluate the expression.
At least, withstatic_assert(iirc, it is new to C11) the compiler complains with a clear "error: expression in static assertion is not constant".
We could fixRTE_BUILD_BUG_ON, but I guess the fix would be equivalent to map it tostatic_assert(!condition).
Using language standard constructs seems a better choice.
+1 to using language standard constructs.
static_assert became standard in C11. (Formally, _Static_assert is standard C11, and static_assert is available through a convenience macro in C11 [1].)
In my opinion, our move to C11 thus makes RTE_BUILD_BUG_ON obsolete.
We should mark RTE_BUILD_BUG_ON as deprecated, and disallow RTE_BUILD_BUG_ON in new code. Perhaps checkpatches could catch this?
[1]: https://en.cppreference.com/w/c/language/_Static_assert
PS: static_assert also has the advantage that it can be used directly in header files. RTE_BUILD_BUG_ON can only be used in functions, and thus needs to be wrapped in a dummy (static inline) function when used in a header file.
Quoted 4 lines
--
David Marchand
Re: [PATCH v6 1/6] eal: provide rte stdatomics optional atomics API
Quoted 1 line · Morten Brørup
In my opinion, our move to C11 thus makes
RTE_BUILD_BUG_ONobsolete.
That's my thought too.
Quoted 2 lines
We should mark
RTE_BUILD_BUG_ONas deprecated, and disallowRTE_BUILD_BUG_ONin new code. Perhaps checkpatches could catch this?
For a clear deprecation of a part of DPDK API, I don't see a need to add something in checkpatch.
Putting a RTE_DEPRECATED in RTE_BUILD_BUG_ON directly triggers a build warning (caught by CI since we run with Werror).
lib/eal/include/rte_common.h +0 -0
@@ -495,7 +495,7 @@ rte_is_aligned(const void * const __rte_restrict
ptr, const unsigned int align)
/**
* Triggers an error at compilation time if the condition is true. */
-#define RTE_BUILD_BUG_ON(condition) ((void)sizeof(char[1 - 2*!!(condition)])) +#define RTE_BUILD_BUG_ON(condition) RTE_DEPRECATED(RTE_BUILD_BUG_ON) ((void)sizeof(char[1 - 2*!!(condition)]))
/*********** Cache line related macros ********/
$ ninja -C build-mini
...
[18/333] Compiling C object lib/librte_eal.a.p/eal_common_eal_common_trace.c.o
../lib/eal/common/eal_common_trace.c: In function ‘eal_trace_init’:
../lib/eal/common/eal_common_trace.c:44:20: warning:
"RTE_BUILD_BUG_ON" is deprecated
44 | RTE_BUILD_BUG_ON((offsetof(struct __rte_trace_header,
mem) % 8) != 0);
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
[38/333] Compiling C object lib/librte_eal.a.p/eal_common_malloc_heap.c.o
../lib/eal/common/malloc_heap.c: In function ‘malloc_heap_destroy’:
../lib/eal/common/malloc_heap.c:1398:20: warning: "RTE_BUILD_BUG_ON"
is deprecated
1398 | RTE_BUILD_BUG_ON(offsetof(struct malloc_heap, lock) != 0);
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
[50/333] Compiling C object lib/librte_eal.a.p/eal_unix_rte_thread.c.o
../lib/eal/unix/rte_thread.c: In function ‘rte_thread_self’:
../lib/eal/unix/rte_thread.c:239:20: warning: "RTE_BUILD_BUG_ON" is deprecated
239 | RTE_BUILD_BUG_ON(sizeof(pthread_t) > sizeof(uintptr_t));
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~Signature · 1 line
David Marchand
Re: [PATCH v6 1/6] eal: provide rte stdatomics optional atomics API
Quoted 6 lines · David Marchand
Quoted 1 line · Morten Brørup
In my opinion, our move to C11 thus makes
RTE_BUILD_BUG_ONobsolete.That's my thought too.
Quoted 1 line
Quoted 7 lines
Quoted 1 line
We should mark
RTE_BUILD_BUG_ONas deprecated, and disallowRTE_BUILD_BUG_ONin new code. Perhaps checkpatches could catch this?For a clear deprecation of a part of DPDK API, I don't see a need to add something in checkpatch.
Putting aRTE_DEPRECATEDinRTE_BUILD_BUG_ONdirectly triggers a build warning (caught by CI since we run with Werror).
Would it not be sufficient to just make it an alias for the C11 static assertions? It's not like its a lot of code to maintain, and if app users have it in their code I'm not sure we get massive benefit from forcing them to edit their code. I'd rather see it kept as a one-line macro purely from a backward compatibility viewpoint. We can replace internal usages, though
- which can be checked by checkpatch.
/Bruce
Re: [PATCH v6 1/6] eal: provide rte stdatomics optional atomics API
Quoted 16 lines · Bruce Richardson
Quoted 6 lines · David Marchand
Quoted 1 line · Morten Brørup
In my opinion, our move to C11 thus makes
RTE_BUILD_BUG_ONobsolete.That's my thought too.
Quoted 1 line
Quoted 7 lines
Quoted 1 line
We should mark
RTE_BUILD_BUG_ONas deprecated, and disallowRTE_BUILD_BUG_ONin new code. Perhaps checkpatches could catch this?For a clear deprecation of a part of DPDK API, I don't see a need to add something in checkpatch.
Putting aRTE_DEPRECATEDinRTE_BUILD_BUG_ONdirectly triggers a build warning (caught by CI since we run with Werror).
Quoted 6 lines
Would it not be sufficient to just make it an alias for the C11 static assertions? It's not like its a lot of code to maintain, and if app users have it in their code I'm not sure we get massive benefit from forcing them to edit their code. I'd rather see it kept as a one-line macro purely from a backward compatibility viewpoint. We can replace internal usages, though
- which can be checked by checkpatch.
No, there is no massive benefit, just trying to reduce our ever growing API surface.
Note, this macro should have been kept internal but it was introduced at a time such matter was not considered...
Signature · 1 line
David Marchand
Re: [PATCH v6 1/6] eal: provide rte stdatomics optional atomics API
29/09/2023 11:34, David Marchand:
Quoted 25 lines
Quoted 16 lines · Bruce Richardson
Quoted 6 lines · David Marchand
Quoted 1 line · Morten Brørup
In my opinion, our move to C11 thus makes
RTE_BUILD_BUG_ONobsolete.That's my thought too.
Quoted 1 line
Quoted 7 lines
Quoted 1 line
We should mark
RTE_BUILD_BUG_ONas deprecated, and disallowRTE_BUILD_BUG_ONin new code. Perhaps checkpatches could catch this?For a clear deprecation of a part of DPDK API, I don't see a need to add something in checkpatch.
Putting aRTE_DEPRECATEDinRTE_BUILD_BUG_ONdirectly triggers a build warning (caught by CI since we run with Werror).Quoted 6 lines
Would it not be sufficient to just make it an alias for the C11 static assertions? It's not like its a lot of code to maintain, and if app users have it in their code I'm not sure we get massive benefit from forcing them to edit their code. I'd rather see it kept as a one-line macro purely from a backward compatibility viewpoint. We can replace internal usages, though
- which can be checked by checkpatch.
Quoted 5 lines
No, there is no massive benefit, just trying to reduce our ever growing API surface.
Note, this macro should have been kept internal but it was introduced at a time such matter was not considered...
I agree with all.
Now taking techboard hat, we agreed to avoid breaking API if possible. So we should keep RTE_BUILD_BUG_ON forever even if not used. Internally we can replace its usages.
Re: [PATCH v6 1/6] eal: provide rte stdatomics optional atomics API
Quoted 30 lines · Thomas Monjalon
29/09/2023 11:34, David Marchand:
Quoted 28 lines
Quoted 16 lines · Bruce Richardson
Quoted 6 lines · David Marchand
Quoted 1 line · Morten Brørup
In my opinion, our move to C11 thus makes
RTE_BUILD_BUG_ONobsolete.That's my thought too.
Quoted 1 line
Quoted 7 lines
Quoted 1 line
We should mark
RTE_BUILD_BUG_ONas deprecated, and disallowRTE_BUILD_BUG_ONin new code. Perhaps checkpatches could catch this?For a clear deprecation of a part of DPDK API, I don't see a need to add something in checkpatch.
Putting aRTE_DEPRECATEDinRTE_BUILD_BUG_ONdirectly triggers a build warning (caught by CI since we run with Werror).Quoted 6 lines
Would it not be sufficient to just make it an alias for the C11 static assertions? It's not like its a lot of code to maintain, and if app users have it in their code I'm not sure we get massive benefit from forcing them to edit their code. I'd rather see it kept as a one-line macro purely from a backward compatibility viewpoint. We can replace internal usages, though
- which can be checked by checkpatch.No, there is no massive benefit, just trying to reduce our ever growing API surface.
Quoted 7 lines
Quoted 2 lines
Note, this macro should have been kept internal but it was introduced at a time such matter was not considered...
I agree with all.
Now taking techboard hat, we agreed to avoid breaking API if possible. So we should keepRTE_BUILD_BUG_ONforever even if not used. Internally we can replace its usages.
So back to the original topic, I get that static_assert is ok for this patch.
Signature · 1 line
David Marchand
Re: [PATCH v6 1/6] eal: provide rte stdatomics optional atomics API
29/09/2023 13:38, David Marchand:
Quoted 34 lines · Thomas Monjalon
Quoted 33 lines
29/09/2023 11:34, David Marchand:
Quoted 30 lines
Quoted 16 lines · Bruce Richardson
Quoted 6 lines · David Marchand
Quoted 1 line · Morten Brørup
In my opinion, our move to C11 thus makes
RTE_BUILD_BUG_ONobsolete.That's my thought too.
Quoted 1 line
Quoted 7 lines
Quoted 1 line
We should mark
RTE_BUILD_BUG_ONas deprecated, and disallowRTE_BUILD_BUG_ONin new code. Perhaps checkpatches could catch this?For a clear deprecation of a part of DPDK API, I don't see a need to add something in checkpatch.
Putting aRTE_DEPRECATEDinRTE_BUILD_BUG_ONdirectly triggers a build warning (caught by CI since we run with Werror).Quoted 6 lines
Would it not be sufficient to just make it an alias for the C11 static assertions? It's not like its a lot of code to maintain, and if app users have it in their code I'm not sure we get massive benefit from forcing them to edit their code. I'd rather see it kept as a one-line macro purely from a backward compatibility viewpoint. We can replace internal usages, though
- which can be checked by checkpatch.No, there is no massive benefit, just trying to reduce our ever growing API surface.
Note, this macro should have been kept internal but it was introduced at a time such matter was not considered...
Quoted 6 lines
Quoted 4 lines
I agree with all.
Now taking techboard hat, we agreed to avoid breaking API if possible. So we should keepRTE_BUILD_BUG_ONforever even if not used. Internally we can replace its usages.So back to the original topic, I get that
static_assertis ok for this patch.
Yes we can use static_assert.
PATCH v6 2/6 eal: adapt EAL to present rte optional atomics API
In reply to Tyler RetzlaffAdapt the EAL public headers to use rte optional atomics API instead of directly using and exposing toolchain specific atomic builtin intrinsics.
--- app/test/test_mcslock.c | 6 ++-- lib/eal/arm/include/rte_atomic_32.h | 4 +-- lib/eal/arm/include/rte_atomic_64.h | 36 +++++++++++------------ lib/eal/arm/include/rte_pause_64.h | 26 ++++++++-------- lib/eal/arm/rte_power_intrinsics.c | 8 ++--- lib/eal/common/eal_common_trace.c | 16 +++++----- lib/eal/include/generic/rte_atomic.h | 50 +++++++++++++++---------------- lib/eal/include/generic/rte_pause.h | 46 ++++++++++++----------------- lib/eal/include/generic/rte_rwlock.h | 47 +++++++++++++++-------------- lib/eal/include/generic/rte_spinlock.h | 19 ++++++------ lib/eal/include/rte_mcslock.h | 50 +++++++++++++++---------------- lib/eal/include/rte_pflock.h | 24 ++++++++------- lib/eal/include/rte_seqcount.h | 18 ++++++------ lib/eal/include/rte_ticketlock.h | 42 +++++++++++++------------- lib/eal/include/rte_trace_point.h | 4 +-- lib/eal/loongarch/include/rte_atomic.h | 4 +-- lib/eal/ppc/include/rte_atomic.h | 54 +++++++++++++++++----------------- lib/eal/riscv/include/rte_atomic.h | 4 +-- lib/eal/x86/include/rte_atomic.h | 8 ++--- lib/eal/x86/include/rte_spinlock.h | 2 +- lib/eal/x86/rte_power_intrinsics.c | 6 ++-- 21 files changed, 237 insertions(+), 237 deletions(-)
app/test/test_mcslock.c +3 -3
@@ -36,9 +36,9 @@* lock multiple times. */ -rte_mcslock_t *p_ml; -rte_mcslock_t *p_ml_try; -rte_mcslock_t *p_ml_perf; +RTE_ATOMIC (rte_mcslock_t *) p_ml; +RTE_ATOMIC (rte_mcslock_t *) p_ml_try; +RTE_ATOMIC (rte_mcslock_t *) p_ml_perf;static unsigned int count;
lib/eal/arm/include/rte_atomic_32.h +2 -2
@@ -34,9 +34,9 @@#define rte_io_rmb () rte_rmb() static __rte_always_inline void -rte_atomic_thread_fence (int memorder) +rte_atomic_thread_fence (rte_memory_order memorder) { -__atomic_thread_fence (memorder); +__rte_atomic_thread_fence (memorder); }#ifdef __cplusplus
lib/eal/arm/include/rte_atomic_64.h +18 -18
@@ -38,9 +38,9 @@#define rte_io_rmb () rte_rmb() static __rte_always_inline void -rte_atomic_thread_fence (int memorder) +rte_atomic_thread_fence (rte_memory_order memorder) { -__atomic_thread_fence (memorder); +__rte_atomic_thread_fence (memorder); }/*------------------------ 128 bit atomic operations -------------------------*/
@@ -107,33 +107,33 @@*/ RTE_SET_USED (failure);/* Find invalid memory order */ -RTE_ASSERT (success == __ATOMIC_RELAXED || - success == __ATOMIC_ACQUIRE || - success == __ATOMIC_RELEASE || - success == __ATOMIC_ACQ_REL || - success == __ATOMIC_SEQ_CST); +RTE_ASSERT (success == rte_memory_order_relaxed || + success == rte_memory_order_acquire || + success == rte_memory_order_release || + success == rte_memory_order_acq_rel || + success == rte_memory_order_seq_cst);rte_int128_t expected = *exp;rte_int128_t desired = *src;rte_int128_t old;#if defined(__ARM_FEATURE_ATOMICS) || defined(RTE_ARM_FEATURE_ATOMICS ) -if (success == __ATOMIC_RELAXED) +if (success == rte_memory_order_relaxed)__cas_128_relaxed (dst, exp, desired); -else if (success == __ATOMIC_ACQUIRE) +else if (success == rte_memory_order_acquire)__cas_128_acquire (dst, exp, desired); -else if (success == __ATOMIC_RELEASE) +else if (success == rte_memory_order_release)__cas_128_release (dst, exp, desired);else __cas_128_acq_rel (dst, exp, desired); old = *exp;#else -#define __HAS_ACQ (mo) ((mo) != __ATOMIC_RELAXED && (mo) != __ATOMIC_RELEASE) -#define __HAS_RLS (mo) ((mo) == __ATOMIC_RELEASE || (mo) == __ATOMIC_ACQ_REL || \ - (mo) == __ATOMIC_SEQ_CST) +#define __HAS_ACQ (mo) ((mo) != rte_memory_order_relaxed && (mo) != rte_memory_order_release) +#define __HAS_RLS (mo) ((mo) == rte_memory_order_release || (mo) == rte_memory_order_acq_rel || \ + (mo) == rte_memory_order_seq_cst) -int ldx_mo =__HAS_ACQ (success) ? __ATOMIC_ACQUIRE : __ATOMIC_RELAXED; -int stx_mo =__HAS_RLS (success) ? __ATOMIC_RELEASE : __ATOMIC_RELAXED; +int ldx_mo =__HAS_ACQ (success) ? rte_memory_order_acquire : rte_memory_order_relaxed; +int stx_mo =__HAS_RLS (success) ? rte_memory_order_release : rte_memory_order_relaxed;#undef __HAS_ACQ#undef __HAS_RLS
@@ -153,7 +153,7 @@ : "Q" (src->val[0]) \ : "memory"); } -if (ldx_mo == __ATOMIC_RELAXED) +if (ldx_mo == rte_memory_order_relaxed) __LOAD_128("ldxp", dst, old) else__LOAD_128 ("ldaxp" , dst, old)
@@ -170,7 +170,7 @@ : "memory"); }if (likely (old.int128 == expected.int128)) { -if (stx_mo == __ATOMIC_RELAXED) +if (stx_mo == rte_memory_order_relaxed) __STORE_128("stxp", dst, desired, ret) else__STORE_128 ("stlxp" , dst, desired, ret)
@@ -181,7 +181,7 @@* needs to be stored back to ensure it was read * atomically. */ -if (stx_mo == __ATOMIC_RELAXED) +if (stx_mo == rte_memory_order_relaxed) __STORE_128("stxp", dst, old, ret) else__STORE_128 ("stlxp" , dst, old, ret)
lib/eal/arm/include/rte_pause_64.h +13 -13
@@ -41,7 +41,7 @@ static inline void rte_pause(void)* implicitly to exit WFE. */ #define __RTE_ARM_LOAD_EXC_8 (src, dst, memorder) { \ - if (memorder == __ATOMIC_RELAXED) { \ + if (memorder == rte_memory_order_relaxed) { \ asm volatile("ldxrb %w[tmp], [%x[addr]]" \ : [tmp] "=&r" (dst) \ : [addr] "r" (src) \
@@ -60,7 +60,7 @@ static inline void rte_pause(void)* implicitly to exit WFE. */ #define __RTE_ARM_LOAD_EXC_16 (src, dst, memorder) { \ - if (memorder == __ATOMIC_RELAXED) { \ + if (memorder == rte_memory_order_relaxed) { \ asm volatile("ldxrh %w[tmp], [%x[addr]]" \ : [tmp] "=&r" (dst) \ : [addr] "r" (src) \
@@ -79,7 +79,7 @@ static inline void rte_pause(void)* implicitly to exit WFE. */ #define __RTE_ARM_LOAD_EXC_32 (src, dst, memorder) { \ - if (memorder == __ATOMIC_RELAXED) { \ + if (memorder == rte_memory_order_relaxed) { \ asm volatile("ldxr %w[tmp], [%x[addr]]" \ : [tmp] "=&r" (dst) \ : [addr] "r" (src) \
@@ -98,7 +98,7 @@ static inline void rte_pause(void)* implicitly to exit WFE. */ #define __RTE_ARM_LOAD_EXC_64 (src, dst, memorder) { \ - if (memorder == __ATOMIC_RELAXED) { \ + if (memorder == rte_memory_order_relaxed) { \ asm volatile("ldxr %x[tmp], [%x[addr]]" \ : [tmp] "=&r" (dst) \ : [addr] "r" (src) \
@@ -118,7 +118,7 @@ static inline void rte_pause(void)*/ #define __RTE_ARM_LOAD_EXC_128 (src, dst, memorder) { \ volatile rte_int128_t *dst_128 = (volatile rte_int128_t *)&dst; \ - if (memorder == __ATOMIC_RELAXED) { \ + if (memorder == rte_memory_order_relaxed) { \ asm volatile("ldxp %x[tmp0], %x[tmp1], [%x[addr]]" \ : [tmp0] "=&r" (dst_128->val[0]), \ [tmp1] "=&r" (dst_128->val[1]) \
@@ -153,8 +153,8 @@ static inline void rte_pause(void) {uint16_t value; -RTE_BUILD_BUG_ON (memorder != __ATOMIC_ACQUIRE && - memorder != __ATOMIC_RELAXED); +RTE_BUILD_BUG_ON (memorder != rte_memory_order_acquire && + memorder != rte_memory_order_relaxed); __RTE_ARM_LOAD_EXC_16(addr, value, memorder)if (value != expected) {
@@ -172,8 +172,8 @@ static inline void rte_pause(void) {uint32_t value; -RTE_BUILD_BUG_ON (memorder != __ATOMIC_ACQUIRE && - memorder != __ATOMIC_RELAXED); +RTE_BUILD_BUG_ON (memorder != rte_memory_order_acquire && + memorder != rte_memory_order_relaxed); __RTE_ARM_LOAD_EXC_32(addr, value, memorder)if (value != expected) {
@@ -191,8 +191,8 @@ static inline void rte_pause(void) {uint64_t value; -RTE_BUILD_BUG_ON (memorder != __ATOMIC_ACQUIRE && - memorder != __ATOMIC_RELAXED); +RTE_BUILD_BUG_ON (memorder != rte_memory_order_acquire && + memorder != rte_memory_order_relaxed); __RTE_ARM_LOAD_EXC_64(addr, value, memorder)if (value != expected) {
@@ -206,8 +206,8 @@ static inline void rte_pause(void)#define RTE_WAIT_UNTIL_MASKED (addr, mask, cond, expected, memorder) do { \ RTE_BUILD_BUG_ON(!__builtin_constant_p(memorder)); \ - RTE_BUILD_BUG_ON(memorder != __ATOMIC_ACQUIRE && \ - memorder != __ATOMIC_RELAXED); \ + RTE_BUILD_BUG_ON(memorder != rte_memory_order_acquire && \ + memorder != rte_memory_order_relaxed); \ const uint32_t size = sizeof(*(addr)) << 3; \ typeof(*(addr)) expected_value = (expected); \ typeof(*(addr)) value; \
lib/eal/arm/rte_power_intrinsics.c +4 -4
@@ -33,19 +33,19 @@switch (pmc->size) {case sizeof (uint8_t ): - __RTE_ARM_LOAD_EXC_8(pmc->addr, cur_value, __ATOMIC_RELAXED) + __RTE_ARM_LOAD_EXC_8(pmc->addr, cur_value, rte_memory_order_relaxed)__RTE_ARM_WFE () break;case sizeof (uint16_t ): - __RTE_ARM_LOAD_EXC_16(pmc->addr, cur_value, __ATOMIC_RELAXED) + __RTE_ARM_LOAD_EXC_16(pmc->addr, cur_value, rte_memory_order_relaxed)__RTE_ARM_WFE () break;case sizeof (uint32_t ): - __RTE_ARM_LOAD_EXC_32(pmc->addr, cur_value, __ATOMIC_RELAXED) + __RTE_ARM_LOAD_EXC_32(pmc->addr, cur_value, rte_memory_order_relaxed)__RTE_ARM_WFE () break;case sizeof (uint64_t ): - __RTE_ARM_LOAD_EXC_64(pmc->addr, cur_value, __ATOMIC_RELAXED) + __RTE_ARM_LOAD_EXC_64(pmc->addr, cur_value, rte_memory_order_relaxed)__RTE_ARM_WFE () break;default :
lib/eal/common/eal_common_trace.c +9 -7
@@ -103,11 +103,11 @@ struct trace_point_head * trace_mode_set(rte_trace_point_t *t, enum rte_trace_mode mode) {if (mode ==RTE_TRACE_MODE_OVERWRITE ) -__atomic_fetch_and (t, ~__RTE_TRACE_FIELD_ENABLE_DISCARD, - __ATOMIC_RELEASE); +rte_atomic_fetch_and_explicit (t, ~__RTE_TRACE_FIELD_ENABLE_DISCARD, + rte_memory_order_release);else -__atomic_fetch_or (t, __RTE_TRACE_FIELD_ENABLE_DISCARD, - __ATOMIC_RELEASE); +rte_atomic_fetch_or_explicit (t, __RTE_TRACE_FIELD_ENABLE_DISCARD, + rte_memory_order_release); }void
@@ -141,7 +141,7 @@ rte_trace_mode rte_trace_mode_get(void)if (trace_point_is_invalid (t))return false; - val =__atomic_load_n (t, __ATOMIC_ACQUIRE); + val =rte_atomic_load_explicit (t, rte_memory_order_acquire);return (val & __RTE_TRACE_FIELD_ENABLE_MASK) !=0 ; }
@@ -153,7 +153,8 @@ rte_trace_mode rte_trace_mode_get(void)if (trace_point_is_invalid (t))return -ERANGE ; - prev =__atomic_fetch_or (t, __RTE_TRACE_FIELD_ENABLE_MASK, __ATOMIC_RELEASE); + prev =rte_atomic_fetch_or_explicit (t, __RTE_TRACE_FIELD_ENABLE_MASK, + rte_memory_order_release);if ((prev & __RTE_TRACE_FIELD_ENABLE_MASK) ==0 )__atomic_fetch_add (&trace.status,1 , __ATOMIC_RELEASE);return 0 ;
@@ -167,7 +168,8 @@ rte_trace_mode rte_trace_mode_get(void)if (trace_point_is_invalid (t))return -ERANGE ; - prev =__atomic_fetch_and (t, ~__RTE_TRACE_FIELD_ENABLE_MASK, __ATOMIC_RELEASE); + prev =rte_atomic_fetch_and_explicit (t, ~__RTE_TRACE_FIELD_ENABLE_MASK, + rte_memory_order_release);if ((prev & __RTE_TRACE_FIELD_ENABLE_MASK) !=0 )__atomic_fetch_sub (&trace.status,1 , __ATOMIC_RELEASE);return 0 ;
lib/eal/include/generic/rte_atomic.h +25 -25
194 lines not on this page · raw message
lib/eal/include/generic/rte_pause.h +19 -27
103 lines not on this page · raw message
lib/eal/include/generic/rte_rwlock.h +25 -22
126 lines not on this page · raw message
lib/eal/include/generic/rte_spinlock.h +9 -10
@@ -29,7 +29,7 @@* The rte_spinlock_t type. */ typedef struct __rte_lockable { -volatile int locked;/**< lock status 0 = unlocked, 1 = locked */ +volatile RTE_ATOMIC (int ) locked;/**< lock status 0 = unlocked, 1 = locked */ }rte_spinlock_t ;/**
@@ -66,10 +66,10 @@ {int exp =0 ; -while (!__atomic_compare_exchange_n (&sl->locked, &exp,1 ,0 , - __ATOMIC_ACQUIRE, __ATOMIC_RELAXED)) { -rte_wait_until_equal_32 ((volatile uint32_t *)&sl->locked, -0 , __ATOMIC_RELAXED); +while (!rte_atomic_compare_exchange_strong_explicit (&sl->locked, &exp,1 , + rte_memory_order_acquire, rte_memory_order_relaxed)) { +rte_wait_until_equal_32 ((volatile uint32_t *)(uintptr_t )&sl->locked, +0 , rte_memory_order_relaxed); exp =0 ; } }
@@ -90,7 +90,7 @@ rte_spinlock_unlock(rte_spinlock_t *sl) __rte_no_thread_safety_analysis { -__atomic_store_n (&sl->locked,0 , __ATOMIC_RELEASE); +rte_atomic_store_explicit (&sl->locked,0 , rte_memory_order_release); }#endif
@@ -113,9 +113,8 @@ __rte_no_thread_safety_analysis {int exp =0 ; -return __atomic_compare_exchange_n (&sl->locked, &exp,1 , -0 ,/* disallow spurious failure */ - __ATOMIC_ACQUIRE, __ATOMIC_RELAXED); +return rte_atomic_compare_exchange_strong_explicit (&sl->locked, &exp,1 , + rte_memory_order_acquire, rte_memory_order_relaxed); }#endif
@@ -129,7 +128,7 @@*/ static inline int rte_spinlock_is_locked (rte_spinlock_t *sl) { -return __atomic_load_n (&sl->locked, __ATOMIC_ACQUIRE); +return rte_atomic_load_explicit (&sl->locked, rte_memory_order_acquire); }/**
lib/eal/include/rte_mcslock.h +25 -25
133 lines not on this page · raw message
lib/eal/include/rte_pflock.h +13 -11
@@ -41,8 +41,8 @@*/ struct rte_pflock {struct { -uint16_t in; -uint16_t out; +RTE_ATOMIC (uint16_t ) in; +RTE_ATOMIC (uint16_t ) out; } rd, wr; };typedef struct rte_pflock rte_pflock_t ;
@@ -117,14 +117,14 @@ struct rte_pflock {* If no writer is present, then the operation has completed * successfully. */ - w =__atomic_fetch_add (&pf->rd.in,RTE_PFLOCK_RINC , __ATOMIC_ACQUIRE) + w =rte_atomic_fetch_add_explicit (&pf->rd.in,RTE_PFLOCK_RINC , rte_memory_order_acquire) &RTE_PFLOCK_WBITS ;if (w ==0 )return ;/* Wait for current write phase to complete. */ RTE_WAIT_UNTIL_MASKED(&pf->rd.in, RTE_PFLOCK_WBITS, !=, w, - __ATOMIC_ACQUIRE); + rte_memory_order_acquire); }/**
@@ -140,7 +140,7 @@ struct rte_pflock {static inline void rte_pflock_read_unlock (rte_pflock_t *pf) { -__atomic_fetch_add (&pf->rd.out,RTE_PFLOCK_RINC , __ATOMIC_RELEASE); +rte_atomic_fetch_add_explicit (&pf->rd.out,RTE_PFLOCK_RINC , rte_memory_order_release); }/**
@@ -161,8 +161,9 @@ struct rte_pflock {/* Acquire ownership of write-phase. * This is same as rte_ticketlock_lock(). */ - ticket =__atomic_fetch_add (&pf->wr.in,1 , __ATOMIC_RELAXED); -rte_wait_until_equal_16 (&pf->wr.out, ticket, __ATOMIC_ACQUIRE); + ticket =rte_atomic_fetch_add_explicit (&pf->wr.in,1 , rte_memory_order_relaxed); +rte_wait_until_equal_16 ((uint16_t *)(uintptr_t )&pf->wr.out, ticket, + rte_memory_order_acquire);/* * Acquire ticket on read-side in order to allow them
@@ -173,10 +174,11 @@ struct rte_pflock {* speculatively. */ w =RTE_PFLOCK_PRES | (ticket &RTE_PFLOCK_PHID ); - ticket =__atomic_fetch_add (&pf->rd.in, w, __ATOMIC_RELAXED); + ticket =rte_atomic_fetch_add_explicit (&pf->rd.in, w, rte_memory_order_relaxed);/* Wait for any pending readers to flush. */ -rte_wait_until_equal_16 (&pf->rd.out, ticket, __ATOMIC_ACQUIRE); +rte_wait_until_equal_16 ((uint16_t *)(uintptr_t )&pf->rd.out, ticket, + rte_memory_order_acquire); }/**
@@ -193,10 +195,10 @@ struct rte_pflock { rte_pflock_write_unlock(rte_pflock_t *pf) {/* Migrate from write phase to read phase. */ -__atomic_fetch_and (&pf->rd.in,RTE_PFLOCK_LSB , __ATOMIC_RELEASE); +rte_atomic_fetch_and_explicit (&pf->rd.in,RTE_PFLOCK_LSB , rte_memory_order_release);/* Allow other writers to continue. */ -__atomic_fetch_add (&pf->wr.out,1 , __ATOMIC_RELEASE); +rte_atomic_fetch_add_explicit (&pf->wr.out,1 , rte_memory_order_release); }#ifdef __cplusplus
lib/eal/include/rte_seqcount.h +9 -9
@@ -32,7 +32,7 @@* The RTE seqcount type. */ typedef struct { -uint32_t sn;/**< A sequence number for the protected data. */ +RTE_ATOMIC (uint32_t ) sn;/**< A sequence number for the protected data. */ }rte_seqcount_t ;/**
@@ -106,11 +106,11 @@static inline uint32_t rte_seqcount_read_begin (const rte_seqcount_t *seqcount) { -/* __ATOMIC_ACQUIRE to prevent loads after (in program order) +/* rte_memory_order_acquire to prevent loads after (in program order) * from happening before the sn load. Synchronizes-with the * store release in rte_seqcount_write_end(). */ -return __atomic_load_n (&seqcount->sn, __ATOMIC_ACQUIRE); +return rte_atomic_load_explicit (&seqcount->sn, rte_memory_order_acquire); }/**
@@ -161,9 +161,9 @@return true;/* make sure the data loads happens before the sn load */ -rte_atomic_thread_fence (__ATOMIC_ACQUIRE); +rte_atomic_thread_fence (rte_memory_order_acquire); - end_sn =__atomic_load_n (&seqcount->sn, __ATOMIC_RELAXED); + end_sn =rte_atomic_load_explicit (&seqcount->sn, rte_memory_order_relaxed);/* A writer incremented the sequence number during this read * critical section.
@@ -205,12 +205,12 @@ sn = seqcount->sn +1 ; -__atomic_store_n (&seqcount->sn, sn, __ATOMIC_RELAXED); +rte_atomic_store_explicit (&seqcount->sn, sn, rte_memory_order_relaxed); -/* __ATOMIC_RELEASE to prevent stores after (in program order) +/* rte_memory_order_release to prevent stores after (in program order) * from happening before the sn store. */ -rte_atomic_thread_fence (__ATOMIC_RELEASE); +rte_atomic_thread_fence (rte_memory_order_release); }/**
@@ -237,7 +237,7 @@ sn = seqcount->sn +1 ;/* Synchronizes-with the load acquire in rte_seqcount_read_begin(). */ -__atomic_store_n (&seqcount->sn, sn, __ATOMIC_RELEASE); +rte_atomic_store_explicit (&seqcount->sn, sn, rte_memory_order_release); }#ifdef __cplusplus
lib/eal/include/rte_ticketlock.h +22 -20
115 lines not on this page · raw message
lib/eal/include/rte_trace_point.h +2 -2
@@ -33,7 +33,7 @@#include <rte_stdatomic.h> /** The tracepoint object. */ -typedef uint64_t rte_trace_point_t ; +typedef RTE_ATOMIC (uint64_t )rte_trace_point_t ;/** * Macro to define the tracepoint arguments in RTE_TRACE_POINT macro.
@@ -359,7 +359,7 @@ struct __rte_trace_header {#define __rte_trace_point_emit_header_generic (t) \ void *mem; \ do { \ - const uint64_t val = __atomic_load_n(t, __ATOMIC_ACQUIRE); \ + const uint64_t val = rte_atomic_load_explicit(t, rte_memory_order_acquire); \ if (likely(!(val & __RTE_TRACE_FIELD_ENABLE_MASK))) \ return; \ mem = __rte_trace_mem_get(val); \
lib/eal/loongarch/include/rte_atomic.h +2 -2
@@ -35,9 +35,9 @@#define rte_io_rmb () rte_mb() static __rte_always_inline void -rte_atomic_thread_fence (int memorder) +rte_atomic_thread_fence (rte_memory_order memorder) { -__atomic_thread_fence (memorder); +__rte_atomic_thread_fence (memorder); }#ifdef __cplusplus
lib/eal/ppc/include/rte_atomic.h +27 -27
172 lines not on this page · raw message
lib/eal/riscv/include/rte_atomic.h +2 -2
@@ -40,9 +40,9 @@#define rte_io_rmb () asm volatile("fence ir, ir" : : : "memory") static __rte_always_inline void -rte_atomic_thread_fence (int memorder) +rte_atomic_thread_fence (rte_memory_order memorder) { -__atomic_thread_fence (memorder); +__rte_atomic_thread_fence (memorder); }#ifdef __cplusplus
lib/eal/x86/include/rte_atomic.h +4 -4
@@ -82,17 +82,17 @@/** * Synchronization fence between threads based on the specified memory order. * -* On x86 the __atomic_thread_fence(__ATOMIC_SEQ_CST) generates full 'mfence' +* On x86 the __rte_atomic_thread_fence(rte_memory_order_seq_cst) generates full 'mfence' * which is quite expensive. The optimized implementation of rte_smp_mb is * used instead. */ static __rte_always_inline void -rte_atomic_thread_fence (int memorder) +rte_atomic_thread_fence (rte_memory_order memorder) { -if (memorder == __ATOMIC_SEQ_CST) +if (memorder == rte_memory_order_seq_cst)rte_smp_mb ();else -__atomic_thread_fence (memorder); +__rte_atomic_thread_fence (memorder); }/*------------------------- 16 bit atomic operations -------------------------*/
lib/eal/x86/include/rte_spinlock.h +1 -1
@@ -78,7 +78,7 @@ static inline int rte_tm_supported(void) }static inline int -rte_try_tm (volatile int *lock) +rte_try_tm (volatile RTE_ATOMIC (int ) *lock) {int i, retries;
lib/eal/x86/rte_power_intrinsics.c +3 -3
@@ -23,9 +23,9 @@uint64_t val;/* trigger a write but don't change the value */ - val =__atomic_load_n ((volatile uint64_t *)addr, __ATOMIC_RELAXED); -__atomic_compare_exchange_n ((volatile uint64_t *)addr, &val, val,0 , - __ATOMIC_RELAXED, __ATOMIC_RELAXED); + val =rte_atomic_load_explicit ((volatile uint64_t *)addr, rte_memory_order_relaxed); +rte_atomic_compare_exchange_strong_explicit ((volatile uint64_t *)addr, &val, val, + rte_memory_order_relaxed, rte_memory_order_relaxed); }static bool wait_supported;
Signature · 1 line
1.8.3.1
PATCH v6 3/6 eal: add rte atomic qualifier with casts
In reply to Tyler RetzlaffIntroduce __rte_atomic qualifying casts in rte_optional atomics inline functions to prevent cascading the need to pass __rte_atomic qualified arguments.
Warning, this is really implementation dependent and being done temporarily to avoid having to convert more of the libraries and tests in DPDK in the initial series that introduces the API. The consequence of the assumption of the ABI of the types in question not being ``the same'' is only a risk that may be realized when enable_stdatomic=true.
--- lib/eal/include/generic/rte_atomic.h | 48 ++++++++++++++++++++++++------------ lib/eal/include/generic/rte_pause.h | 9 ++++--- lib/eal/x86/rte_power_intrinsics.c | 7 +++--- 3 files changed, 42 insertions(+), 22 deletions(-)
lib/eal/include/generic/rte_atomic.h +32 -16
@@ -274,7 +274,8 @@static inline void rte_atomic16_add (rte_atomic16_t *v,int16_t inc) { -rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst); + rte_atomic_fetch_add_explicit((volatile __rte_atomic int16_t *)&v->cnt, inc, + rte_memory_order_seq_cst); }/**
@@ -288,7 +289,8 @@static inline void rte_atomic16_sub (rte_atomic16_t *v,int16_t dec) { -rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst); + rte_atomic_fetch_sub_explicit((volatile __rte_atomic int16_t *)&v->cnt, dec, + rte_memory_order_seq_cst); }/**
@@ -341,7 +343,8 @@static inline int16_t rte_atomic16_add_return (rte_atomic16_t *v,int16_t inc) { -return rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst) + inc; + return rte_atomic_fetch_add_explicit((volatile __rte_atomic int16_t *)&v->cnt, inc, + rte_memory_order_seq_cst) + inc; }/**
@@ -361,7 +364,8 @@static inline int16_t rte_atomic16_sub_return (rte_atomic16_t *v,int16_t dec) { -return rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst) - dec; + return rte_atomic_fetch_sub_explicit((volatile __rte_atomic int16_t *)&v->cnt, dec, + rte_memory_order_seq_cst) - dec; }/**
@@ -380,7 +384,8 @@#ifdef RTE_FORCE_INTRINSICS static inline int rte_atomic16_inc_and_test (rte_atomic16_t *v) { -return rte_atomic_fetch_add_explicit (&v->cnt,1 , rte_memory_order_seq_cst) +1 ==0 ; + return rte_atomic_fetch_add_explicit((volatile __rte_atomic int16_t *)&v->cnt, 1, + rte_memory_order_seq_cst) +1 ==0 ; }#endif
@@ -400,7 +405,8 @@ static inline int rte_atomic16_inc_and_test(rte_atomic16_t *v)#ifdef RTE_FORCE_INTRINSICS static inline int rte_atomic16_dec_and_test (rte_atomic16_t *v) { -return rte_atomic_fetch_sub_explicit (&v->cnt,1 , rte_memory_order_seq_cst) -1 ==0 ; + return rte_atomic_fetch_sub_explicit((volatile __rte_atomic int16_t *)&v->cnt, 1, + rte_memory_order_seq_cst) -1 ==0 ; }#endif
@@ -553,7 +559,8 @@ static inline void rte_atomic16_clear(rte_atomic16_t *v)static inline void rte_atomic32_add (rte_atomic32_t *v,int32_t inc) { -rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst); + rte_atomic_fetch_add_explicit((volatile __rte_atomic int32_t *)&v->cnt, inc, + rte_memory_order_seq_cst); }/**
@@ -567,7 +574,8 @@ static inline void rte_atomic16_clear(rte_atomic16_t *v)static inline void rte_atomic32_sub (rte_atomic32_t *v,int32_t dec) { -rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst); + rte_atomic_fetch_sub_explicit((volatile __rte_atomic int32_t *)&v->cnt, dec, + rte_memory_order_seq_cst); }/**
@@ -620,7 +628,8 @@ static inline void rte_atomic16_clear(rte_atomic16_t *v)static inline int32_t rte_atomic32_add_return (rte_atomic32_t *v,int32_t inc) { -return rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst) + inc; + return rte_atomic_fetch_add_explicit((volatile __rte_atomic int32_t *)&v->cnt, inc, + rte_memory_order_seq_cst) + inc; }/**
@@ -640,7 +649,8 @@ static inline void rte_atomic16_clear(rte_atomic16_t *v)static inline int32_t rte_atomic32_sub_return (rte_atomic32_t *v,int32_t dec) { -return rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst) - dec; + return rte_atomic_fetch_sub_explicit((volatile __rte_atomic int32_t *)&v->cnt, dec, + rte_memory_order_seq_cst) - dec; }/**
@@ -659,7 +669,8 @@ static inline void rte_atomic16_clear(rte_atomic16_t *v)#ifdef RTE_FORCE_INTRINSICS static inline int rte_atomic32_inc_and_test (rte_atomic32_t *v) { -return rte_atomic_fetch_add_explicit (&v->cnt,1 , rte_memory_order_seq_cst) +1 ==0 ; + return rte_atomic_fetch_add_explicit((volatile __rte_atomic int32_t *)&v->cnt, 1, + rte_memory_order_seq_cst) +1 ==0 ; }#endif
@@ -679,7 +690,8 @@ static inline int rte_atomic32_inc_and_test(rte_atomic32_t *v)#ifdef RTE_FORCE_INTRINSICS static inline int rte_atomic32_dec_and_test (rte_atomic32_t *v) { -return rte_atomic_fetch_sub_explicit (&v->cnt,1 , rte_memory_order_seq_cst) -1 ==0 ; + return rte_atomic_fetch_sub_explicit((volatile __rte_atomic int32_t *)&v->cnt, 1, + rte_memory_order_seq_cst) -1 ==0 ; }#endif
@@ -885,7 +897,8 @@ static inline void rte_atomic32_clear(rte_atomic32_t *v)static inline void rte_atomic64_add (rte_atomic64_t *v,int64_t inc) { -rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst); + rte_atomic_fetch_add_explicit((volatile __rte_atomic int64_t *)&v->cnt, inc, + rte_memory_order_seq_cst); }#endif
@@ -904,7 +917,8 @@ static inline void rte_atomic32_clear(rte_atomic32_t *v)static inline void rte_atomic64_sub (rte_atomic64_t *v,int64_t dec) { -rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst); + rte_atomic_fetch_sub_explicit((volatile __rte_atomic int64_t *)&v->cnt, dec, + rte_memory_order_seq_cst); }#endif
@@ -962,7 +976,8 @@ static inline void rte_atomic32_clear(rte_atomic32_t *v)static inline int64_t rte_atomic64_add_return (rte_atomic64_t *v,int64_t inc) { -return rte_atomic_fetch_add_explicit (&v->cnt, inc, rte_memory_order_seq_cst) + inc; + return rte_atomic_fetch_add_explicit((volatile __rte_atomic int64_t *)&v->cnt, inc, + rte_memory_order_seq_cst) + inc; }#endif
@@ -986,7 +1001,8 @@ static inline void rte_atomic32_clear(rte_atomic32_t *v)static inline int64_t rte_atomic64_sub_return (rte_atomic64_t *v,int64_t dec) { -return rte_atomic_fetch_sub_explicit (&v->cnt, dec, rte_memory_order_seq_cst) - dec; + return rte_atomic_fetch_sub_explicit((volatile __rte_atomic int64_t *)&v->cnt, dec, + rte_memory_order_seq_cst) - dec; }#endif
lib/eal/include/generic/rte_pause.h +6 -3
@@ -81,7 +81,8 @@ {assert (memorder == rte_memory_order_acquire || memorder == rte_memory_order_relaxed); -while (rte_atomic_load_explicit (addr, memorder) != expected) + while (rte_atomic_load_explicit((volatile __rte_atomic uint16_t *)addr, memorder) + != expected)rte_pause (); }
@@ -91,7 +92,8 @@ {assert (memorder == rte_memory_order_acquire || memorder == rte_memory_order_relaxed); -while (rte_atomic_load_explicit (addr, memorder) != expected) + while (rte_atomic_load_explicit((volatile __rte_atomic uint32_t *)addr, memorder) + != expected)rte_pause (); }
@@ -101,7 +103,8 @@ {assert (memorder == rte_memory_order_acquire || memorder == rte_memory_order_relaxed); -while (rte_atomic_load_explicit (addr, memorder) != expected) + while (rte_atomic_load_explicit((volatile __rte_atomic uint64_t *)addr, memorder) + != expected)rte_pause (); }
lib/eal/x86/rte_power_intrinsics.c +4 -3
@@ -23,9 +23,10 @@uint64_t val;/* trigger a write but don't change the value */ - val =rte_atomic_load_explicit ((volatile uint64_t *)addr, rte_memory_order_relaxed); -rte_atomic_compare_exchange_strong_explicit ((volatile uint64_t *)addr, &val, val, - rte_memory_order_relaxed, rte_memory_order_relaxed); + val = rte_atomic_load_explicit((volatile __rte_atomic uint64_t *)addr, + rte_memory_order_relaxed); + rte_atomic_compare_exchange_strong_explicit((volatile __rte_atomic uint64_t *)addr, + &val, val, rte_memory_order_relaxed, rte_memory_order_relaxed); }static bool wait_supported;
Signature · 1 line
1.8.3.1
PATCH v6 4/6 distributor: adapt for EAL optional atomics API changes
In reply to Tyler RetzlaffAdapt distributor for EAL optional atomics API changes
--- lib/distributor/distributor_private.h | 2 +- lib/distributor/rte_distributor_single.c | 44 ++++++++++++++++---------------- 2 files changed, 23 insertions(+), 23 deletions(-)
lib/distributor/distributor_private.h +1 -1
@@ -52,7 +52,7 @@* Only 64-bits of the memory is actually used though. */ union rte_distributor_buffer_single { -volatile int64_t bufptr64; +volatile RTE_ATOMIC (int64_t ) bufptr64;char pad[RTE_CACHE_LINE_SIZE *3 ]; } __rte_cache_aligned;
lib/distributor/rte_distributor_single.c +22 -22
117 lines not on this page · raw message
Signature · 1 line
1.8.3.1
PATCH v6 5/6 bpf: adapt for EAL optional atomics API changes
In reply to Tyler RetzlaffAdapt bpf for EAL optional atomics API changes
--- lib/bpf/bpf_pkt.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)
lib/bpf/bpf_pkt.c +3 -3
@@ -25,7 +25,7 @@struct bpf_eth_cbi {/* used by both data & control path */ -uint32_t use;/*usage counter */ +RTE_ATOMIC (uint32_t ) use;/*usage counter */ const struct rte_eth_rxtx_callback *cb;/* callback handle */ struct rte_bpf *bpf;struct rte_bpf_jit jit;
@@ -110,8 +110,8 @@ struct bpf_eth_cbh {/* in use, busy wait till current RX/TX iteration is finished */ if ((puse &BPF_ETH_CBI_INUSE ) !=0 ) { -RTE_WAIT_UNTIL_MASKED ((uint32_t *)(uintptr_t )&cbi->use, - UINT32_MAX, !=, puse, __ATOMIC_RELAXED); + RTE_WAIT_UNTIL_MASKED((__rte_atomic uint32_t *)(uintptr_t)&cbi->use, + UINT32_MAX, !=, puse, rte_memory_order_relaxed); } }
Signature · 1 line
1.8.3.1
PATCH v6 6/6 devtools: forbid new direct use of GCC atomic builtins
In reply to Tyler RetzlaffRefrain from using compiler __atomic_xxx builtins DPDK now requires the use of rte_atomic_<op>_explicit macros when operating on DPDK atomic variables.
--- devtools/checkpatches.sh | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-)
devtools/checkpatches.sh +6 -6
@@ -102,20 +102,20 @@ check_forbidden_additions() { # <patch> -f $(dirname $(readlink -f $0))/check-forbidden-tokens.awk \ "$1" || res=1 - # refrain from using compiler __atomic_thread_fence() + # refrain from using compiler __rte_atomic_thread_fence() # It should be avoided on x86 for SMP case. awk -v FOLDERS="lib drivers app examples" \ - -v EXPRESSIONS="__atomic_thread_fence\\\(" \ + -v EXPRESSIONS="__rte_atomic_thread_fence\\\(" \ -v RET_ON_FAIL=1 \ - -v MESSAGE='Using __atomic_thread_fence' \ + -v MESSAGE='Using __rte_atomic_thread_fence, prefer rte_atomic_thread_fence' \ -f $(dirname $(readlink -f $0))/check-forbidden-tokens.awk \ "$1" || res=1 - # refrain from using compiler __atomic_{add,and,nand,or,sub,xor}_fetch() + # refrain from using compiler __atomic_xxx builtins awk -v FOLDERS="lib drivers app examples" \ - -v EXPRESSIONS="__atomic_(add|and|nand|or|sub|xor)_fetch\\\(" \ + -v EXPRESSIONS="__atomic_.*\\\(" \ -v RET_ON_FAIL=1 \ - -v MESSAGE='Using __atomic_op_fetch, prefer __atomic_fetch_op' \ + -v MESSAGE='Using __atomic_xxx built-ins, prefer rte_atomic_xxx' \ -f $(dirname $(readlink -f $0))/check-forbidden-tokens.awk \ "$1" || res=1
Signature · 1 line
1.8.3.1
Re: [PATCH v6 0/6] rte atomics API for optional stdatomic
In reply to Tyler Retzlaffping for additional reviewers.
thanks!
Quoted 126 lines · Tyler Retzlaff
This series introduces API additions prefixed in the rte namespace that allow the optional use of
stdatomics.hfrom C11 using enable_stdatomics=true for targets where enable_stdatomics=false no functional change is intended.Be aware this does not contain all changes to use stdatomics across the DPDK tree it only introduces the minimum to allow the option to be used which is a pre-requisite for a clean CI (probably using clang) that can be run with enable_stdatomics=true enabled.
It is planned that subsequent series will be introduced per lib/driver as appropriate to further enable stdatomics use when enable_stdatomics=true.
Notes:
* Additional libraries beyond EAL make visible atomics use across the API/ABI surface they will be converted in the subsequent series.
* The eal: add rte atomic qualifier with casts patch needs some discussion as to whether or not the legacy
rte_atomicAPIs should be converted to work with enable_stdatomic=true right now some implementation dependent casts are used to prevent cascading / having to convert too much in the intial series.* Windows will obviously need complete conversion of libraries including atomics that are not crossing API/ABI boundaries. those conversions will introduced in separate series as new along side the existing msvc series.
Please keep in mind we would like to prioritize the review / acceptance of this patch since it needs to be completed in the 23.11 merge window.
Thank you all for the discussion that lead to the formation of this series.
v6:
* Adjust checkpatches to warn about use of__rte_atomic_thread_fenceand suggest use ofrte_atomic_thread_fence. Use the existing check more generic check for__atomic_xxxto catch use of__atomic_thread_fenceand recommendrte_atomic_xxx.v5:
* AddRTE_ATOMICto doxygen configuration PREDEFINED macros list to fix documentation generation failure
* Fix two typos in expansion of C11 atomics macros strong -> weak and add missing_explicit
* Adjust devtools/checkpatches messages based on feedback. i have chosen not to try and catch use of C11 atomics or_Atomicsince using those directly will be picked up by existing CI passes where by compilation error where enable_stdatomic=false (the default for most platforms)v4:
* Move the definition of #defineRTE_ATOMIC(type)to patch 1 where it belongs (a mistake in v3)
* Provide comments for bothRTE_ATOMICand__rte_atomicmacros indicating their use as specified or qualified contexts.v3:
* Remove comments from APIs mentioning the mapping to C++ memory model memory orders
* Introduce and use new macroRTE_ATOMIC(type)to be used in contexts where_Atomicis used as a type specifier to declare variables. The macro allows more clarity about what the atomic type being specified is. e.g._Atomic(T *)vs_Atomic(T)it is easier to understand that the former is an atomic pointer type and the latter is an atomic type. it also has the benefit of (in the future) being interoperable with c++23 syntactically
note: Morten i have retained your 'reviewed-by' tags if you disagree given the changes in the above version please indicate as such but i believe the changes are in the spirit of the feedback you providedv2:
* Wrap meson_options.txt option description to newline and indent to be consistent with other options.
* Provide separate typedef ofrte_memory_orderfor enable_stdatomic=true VS enable_stdatomic=false instead of a single typedef to int
note: slight tweak to reviewers feedback i've chosen to use a typedef
for both enable_stdatomic={true,false} (just seemed more consistent)
* Bring inassert.hand usestatic_assertmacro instead of_Static_assertkeyword to better interoperate with c/c++
* Directly includerte_stdatomic.hwhere into other places it is consumed instead of hacking it globally intorte_config.h
* Provide and use__rte_atomic_thread_fenceto allow conditional expansion within the body of existingrte_atomic_thread_fenceinline function to maintain per-arch optimizations when enable_stdatomic=falseTyler Retzlaff (6):
eal: provide rte stdatomics optional atomics API
eal: adapt EAL to present rte optional atomics API
eal: add rte atomic qualifier with casts
distributor: adapt for EAL optional atomics API changes
bpf: adapt for EAL optional atomics API changes
devtools: forbid new direct use of GCC atomic builtinsapp/test/test_mcslock.c | 6 +- config/meson.build | 1 + devtools/checkpatches.sh | 12 +- doc/api/doxy-api.conf.in | 1 + lib/bpf/bpf_pkt.c | 6 +- lib/distributor/distributor_private.h | 2 +- lib/distributor/rte_distributor_single.c | 44 +++---- lib/eal/arm/include/rte_atomic_32.h | 4 +- lib/eal/arm/include/rte_atomic_64.h | 36 +++--- lib/eal/arm/include/rte_pause_64.h | 26 ++-- lib/eal/arm/rte_power_intrinsics.c | 8 +- lib/eal/common/eal_common_trace.c | 16 +-- lib/eal/include/generic/rte_atomic.h | 67 +++++++---- lib/eal/include/generic/rte_pause.h | 50 ++++---- lib/eal/include/generic/rte_rwlock.h | 48 ++++---- lib/eal/include/generic/rte_spinlock.h | 20 ++-- lib/eal/include/meson.build | 1 + lib/eal/include/rte_mcslock.h | 51 ++++---- lib/eal/include/rte_pflock.h | 25 ++-- lib/eal/include/rte_seqcount.h | 19 +-- lib/eal/include/rte_stdatomic.h | 198 +++++++++++++++++++++++++++++++ lib/eal/include/rte_ticketlock.h | 43 +++---- lib/eal/include/rte_trace_point.h | 5 +- lib/eal/loongarch/include/rte_atomic.h | 4 +- lib/eal/ppc/include/rte_atomic.h | 54 ++++----- lib/eal/riscv/include/rte_atomic.h | 4 +- lib/eal/x86/include/rte_atomic.h | 8 +- lib/eal/x86/include/rte_spinlock.h | 2 +- lib/eal/x86/rte_power_intrinsics.c | 7 +- meson_options.txt | 2 + 30 files changed, 501 insertions(+), 269 deletions(-) create mode 100644 lib/eal/include/rte_stdatomic.hSignature · 1 line
1.8.3.1
Re: [PATCH v6 0/6] rte atomics API for optional stdatomic
In reply to Tyler RetzlaffHello,
Quoted 24 lines · Tyler Retzlaff
This series introduces API additions prefixed in the rte namespace that allow the optional use of
stdatomics.hfrom C11 using enable_stdatomics=true for targets where enable_stdatomics=false no functional change is intended.Be aware this does not contain all changes to use stdatomics across the DPDK tree it only introduces the minimum to allow the option to be used which is a pre-requisite for a clean CI (probably using clang) that can be run with enable_stdatomics=true enabled.
It is planned that subsequent series will be introduced per lib/driver as appropriate to further enable stdatomics use when enable_stdatomics=true.
Notes:
* Additional libraries beyond EAL make visible atomics use across the API/ABI surface they will be converted in the subsequent series.
* The eal: add rte atomic qualifier with casts patch needs some discussion as to whether or not the legacy
rte_atomicAPIs should be converted to work with enable_stdatomic=true right now some implementation dependent casts are used to prevent cascading / having to convert too much in the intial series.
Quoted 8 lines
* Windows will obviously need complete conversion of libraries including atomics that are not crossing API/ABI boundaries. those conversions will introduced in separate series as new along side the existing msvc series.
Please keep in mind we would like to prioritize the review / acceptance of this patch since it needs to be completed in the 23.11 merge window.
Thank you all for the discussion that lead to the formation of this series.
I did a number of updates on this v6:
- moved rte_stdatomic.h from patch 1 to later patches where needed,
- added a RN entry,
- tried to consistently/adjusted indent,
- fixed mentions of stdatomic*s* to simple atomic, like in the build option name,
- removed unneded comments (Thomas review on patch 1),
Series applied, thanks Tyler.
Two things are missing:
- add doxygen tags in the new API (this can be fixed later in this release, can you look at it?),
- add compilation tests for enable_stdatomic (I'll send a patch soon for devtools and GHA),
Signature · 1 line
David Marchand