replace rte atomics with GCC builtin atomics
Subsystems (3)
cover replace rte atomics with GCC builtin atomics
Replace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
This series covers the libraries and drivers that are built on Windows.
The code has be converted to use the __atomic builtins but there are additional during conversion i notice that there may be some issues that need to be addressed.
I'll comment in the patches where my concerns are so the maintainers may comment.
Tyler Retzlaff (7):
ring: replace rte atomics with GCC builtin atomics
stack: replace rte atomics with GCC builtin atomics dma/idxd: replace rte atomics with GCC builtin atomics net/ice: replace rte atomics with GCC builtin atomics net/ixgbe: replace rte atomics with GCC builtin atomics net/null: replace rte atomics with GCC builtin atomics net/ring: replace rte atomics with GCC builtin atomics
drivers/dma/idxd/idxd_internal.h | 3 +-- drivers/dma/idxd/idxd_pci.c | 6 +++--- drivers/net/ice/ice_dcf.c | 1 - drivers/net/ice/ice_dcf_ethdev.c | 1 - drivers/net/ice/ice_ethdev.c | 10 ++++++---- drivers/net/ixgbe/ixgbe_bypass.c | 1 - drivers/net/ixgbe/ixgbe_ethdev.c | 12 ++++++------ drivers/net/ixgbe/ixgbe_ethdev.h | 3 ++- drivers/net/ixgbe/ixgbe_flow.c | 1 - drivers/net/ixgbe/ixgbe_rxtx.c | 1 - drivers/net/null/rte_eth_null.c | 20 ++++++++++---------- drivers/net/ring/rte_eth_ring.c | 20 ++++++++++---------- lib/ring/rte_ring_core.h | 1 - lib/ring/rte_ring_generic_pvt.h | 10 ++++++---- lib/stack/rte_stack_lf_generic.h | 11 +++++------ 15 files changed, 49 insertions(+), 52 deletions(-)
Signature · 1 line
1.8.3.1
PATCH 1/7 ring: replace rte atomics with GCC builtin atomics
Replace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
--- lib/ring/rte_ring_core.h | 1 - lib/ring/rte_ring_generic_pvt.h | 10 ++++++---- 2 files changed, 6 insertions(+), 5 deletions(-)
lib/ring/rte_ring_core.h +0 -1
@@ -31,7 +31,6 @@#include <rte_config.h> #include <rte_memory.h> #include <rte_lcore.h> -#include <rte_atomic.h> #include <rte_branch_prediction.h> #include <rte_memzone.h> #include <rte_pause.h>
lib/ring/rte_ring_generic_pvt.h +6 -4
@@ -92,8 +92,9 @@if (is_sp) r->prod.head = *new_head, success =1 ;else - success =rte_atomic32_cmpset (&r->prod.head, - *old_head, *new_head); + success =__atomic_compare_exchange_n (&r->prod.head, + old_head, *new_head,0 , + __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST); }while (unlikely (success ==0 ));return n; }
@@ -162,8 +163,9 @@rte_smp_rmb (); success =1 ; }else { - success =rte_atomic32_cmpset (&r->cons.head, *old_head, - *new_head); + success =__atomic_compare_exchange_n (&r->cons.head, + old_head, *new_head,0 , + __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST); } }while (unlikely (success ==0 ));return n;
Signature · 1 line
1.8.3.1
PATCH 2/7 stack: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffReplace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
--- lib/stack/rte_stack_lf_generic.h | 11 +++++------ 1 file changed, 5 insertions(+), 6 deletions(-)
lib/stack/rte_stack_lf_generic.h +5 -6
@@ -26,8 +26,7 @@* elements. If the mempool is near-empty to the point that this is a * concern, the user should consider increasing the mempool size. */ -return (unsigned int )rte_atomic64_read ((rte_atomic64_t *) - &s->stack_lf.used.len); +return __atomic_load_n (&s->stack_lf.used.len, __ATOMIC_SEQ_CST); }static __rte_always_inline void
@@ -68,7 +67,7 @@ __ATOMIC_RELAXED); }while (success ==0 ); -rte_atomic64_add ((rte_atomic64_t *)&list->len, num); +__atomic_fetch_add (&list->len, num, __ATOMIC_SEQ_CST); } static __rte_always_inline struct rte_stack_lf_elem *
@@ -82,14 +81,14 @@/* Reserve num elements, if available */ while (1 ) { -uint64_t len =rte_atomic64_read ((rte_atomic64_t *)&list->len); +uint64_t len =__atomic_load_n (&list->len, __ATOMIC_SEQ_CST);/* Does the list contain enough elements? */ if (unlikely (len < num))return NULL ; -if (rte_atomic64_cmpset ((volatile uint64_t *)&list->len, - len, len - num)) +if (__atomic_compare_exchange_n (&list->len, &len, len - num, +0 , __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))break ; }
Signature · 1 line
1.8.3.1
PATCH 3/7 dma/idxd: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffReplace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
--- drivers/dma/idxd/idxd_internal.h | 3 +-- drivers/dma/idxd/idxd_pci.c | 6 +++--- 2 files changed, 4 insertions(+), 5 deletions(-)
drivers/dma/idxd/idxd_internal.h +1 -2
@@ -7,7 +7,6 @@#include <rte_dmadev_pmd.h> #include <rte_spinlock.h> -#include <rte_atomic.h> #include "idxd_hw_defs.h"
@@ -34,7 +33,7 @@ struct idxd_pci_common {rte_spinlock_t lk;uint8_t wq_cfg_sz; -rte_atomic16_t ref_count; +int16_t ref_count;volatile struct rte_idxd_bar0 *regs;volatile uint32_t *wq_regs_base;volatile struct rte_idxd_grpcfg *grp_regs;
drivers/dma/idxd/idxd_pci.c +3 -3
@@ -6,7 +6,6 @@#include <rte_devargs.h> #include <rte_dmadev_pmd.h> #include <rte_malloc.h> -#include <rte_atomic.h> #include "idxd_internal.h"
@@ -136,7 +135,8 @@/* if this is the last WQ on the device, disable the device and free * the PCI struct */ - is_last_wq =rte_atomic16_dec_and_test (&idxd->u.pci->ref_count); + is_last_wq =__atomic_fetch_sub (&idxd->u.pci->ref_count,1 , + __ATOMIC_SEQ_CST) -1 ==0 ;if (is_last_wq) {/* disable the device */ err_code =idxd_pci_dev_command (idxd, idxd_disable_dev);
@@ -350,7 +350,7 @@free (idxd.u.pci);return ret; } -rte_atomic16_inc (&idxd.u.pci->ref_count); +__atomic_fetch_add (&idxd.u.pci->ref_count,1 , __ATOMIC_SEQ_CST); }return 0 ;
Signature · 1 line
1.8.3.1
PATCH 4/7 net/ice: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffReplace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
--- drivers/net/ice/ice_dcf.c | 1 - drivers/net/ice/ice_dcf_ethdev.c | 1 - drivers/net/ice/ice_ethdev.c | 10 ++++++---- 3 files changed, 6 insertions(+), 6 deletions(-)
drivers/net/ice/ice_dcf.c +0 -1
@@ -14,7 +14,6 @@#include <rte_common.h> #include <rte_pci.h> -#include <rte_atomic.h> #include <rte_eal.h> #include <rte_ether.h> #include <ethdev_driver.h>
drivers/net/ice/ice_dcf_ethdev.c +0 -1
@@ -11,7 +11,6 @@#include <rte_interrupts.h> #include <rte_debug.h> #include <rte_pci.h> -#include <rte_atomic.h> #include <rte_eal.h> #include <rte_ether.h> #include <ethdev_pci.h>
drivers/net/ice/ice_ethdev.c +6 -4
@@ -3927,8 +3927,9 @@ static int ice_init_rss(struct ice_pf *pf)struct rte_eth_link *dst = link;struct rte_eth_link *src = &dev->data->dev_link; -if (rte_atomic64_cmpset ((uint64_t *)dst, *(uint64_t *)dst, - *(uint64_t *)src) ==0 ) +if (!__atomic_compare_exchange_n ((uint64_t *)dst, + (uint64_t *)dst, *(uint64_t *)src,0 , + __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))return -1 ;return 0 ;
@@ -3941,8 +3942,9 @@ static int ice_init_rss(struct ice_pf *pf)struct rte_eth_link *dst = &dev->data->dev_link;struct rte_eth_link *src = link; -if (rte_atomic64_cmpset ((uint64_t *)dst, *(uint64_t *)dst, - *(uint64_t *)src) ==0 ) +if (!__atomic_compare_exchange_n ((uint64_t *)dst, + (uint64_t *)dst, *(uint64_t *)src,0 , + __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))return -1 ;return 0 ;
Signature · 1 line
1.8.3.1
PATCH 6/7 net/null: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffReplace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
--- drivers/net/null/rte_eth_null.c | 20 ++++++++++---------- 1 file changed, 10 insertions(+), 10 deletions(-)
drivers/net/null/rte_eth_null.c +10 -10
@@ -37,8 +37,8 @@ struct null_queue {struct rte_mempool *mb_pool;struct rte_mbuf *dummy_packet; -rte_atomic64_t rx_pkts; -rte_atomic64_t tx_pkts; +int64_t rx_pkts; +int64_t tx_pkts; };struct pmd_options {
@@ -101,7 +101,7 @@ struct pmd_internals { bufs[i]->port = h->internals->port_id; } -rte_atomic64_add (&(h->rx_pkts), i); +__atomic_fetch_add (&h->rx_pkts, i, __ATOMIC_SEQ_CST);return i; }
@@ -128,7 +128,7 @@ struct pmd_internals { bufs[i]->port = h->internals->port_id; } -rte_atomic64_add (&(h->rx_pkts), i); +__atomic_fetch_add (&h->rx_pkts, i, __ATOMIC_SEQ_CST);return i; }
@@ -152,7 +152,7 @@ struct pmd_internals {for (i =0 ; i < nb_bufs; i++)rte_pktmbuf_free (bufs[i]); -rte_atomic64_add (&(h->tx_pkts), i); +__atomic_fetch_add (&h->tx_pkts, i, __ATOMIC_SEQ_CST);return i; }
@@ -174,7 +174,7 @@ struct pmd_internals {rte_pktmbuf_free (bufs[i]); } -rte_atomic64_add (&(h->tx_pkts), i); +__atomic_fetch_add (&h->tx_pkts, i, __ATOMIC_SEQ_CST);return i; }
@@ -317,7 +317,7 @@ struct pmd_internals { RTE_DIM(internal->rx_null_queues)));for (i =0 ; i < num_stats; i++) { igb_stats->q_ipackets[i] = - internal->rx_null_queues[i].rx_pkts.cnt; + internal->rx_null_queues[i].rx_pkts; rx_total += igb_stats->q_ipackets[i]; }
@@ -326,7 +326,7 @@ struct pmd_internals { RTE_DIM(internal->tx_null_queues)));for (i =0 ; i < num_stats; i++) { igb_stats->q_opackets[i] = - internal->tx_null_queues[i].tx_pkts.cnt; + internal->tx_null_queues[i].tx_pkts; tx_total += igb_stats->q_opackets[i]; }
@@ -347,9 +347,9 @@ struct pmd_internals { internal = dev->data->dev_private;for (i =0 ; i <RTE_DIM (internal->rx_null_queues); i++) - internal->rx_null_queues[i].rx_pkts.cnt =0 ; + internal->rx_null_queues[i].rx_pkts =0 ;for (i =0 ; i <RTE_DIM (internal->tx_null_queues); i++) - internal->tx_null_queues[i].tx_pkts.cnt =0 ; + internal->tx_null_queues[i].tx_pkts =0 ;return 0 ; }
Signature · 1 line
1.8.3.1
PATCH 7/7 net/ring: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffReplace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
--- drivers/net/ring/rte_eth_ring.c | 20 ++++++++++---------- 1 file changed, 10 insertions(+), 10 deletions(-)
drivers/net/ring/rte_eth_ring.c +10 -10
@@ -44,8 +44,8 @@ enum dev_action {struct ring_queue {struct rte_ring *rng; -rte_atomic64_t rx_pkts; -rte_atomic64_t tx_pkts; +int64_t rx_pkts; +int64_t tx_pkts; };struct pmd_internals {
@@ -80,9 +80,9 @@ struct pmd_internals {const uint16_t nb_rx = (uint16_t )rte_ring_dequeue_burst (r->rng, ptrs, nb_bufs,NULL );if (r->rng->flags &RING_F_SC_DEQ ) - r->rx_pkts.cnt += nb_rx; + r->rx_pkts += nb_rx;else -rte_atomic64_add (&(r->rx_pkts), nb_rx); +__atomic_fetch_add (&r->rx_pkts, nb_rx, __ATOMIC_SEQ_CST);return nb_rx; }
@@ -94,9 +94,9 @@ struct pmd_internals {const uint16_t nb_tx = (uint16_t )rte_ring_enqueue_burst (r->rng, ptrs, nb_bufs,NULL );if (r->rng->flags &RING_F_SP_ENQ ) - r->tx_pkts.cnt += nb_tx; + r->tx_pkts += nb_tx;else -rte_atomic64_add (&(r->tx_pkts), nb_tx); +__atomic_fetch_add (&r->tx_pkts, nb_tx, __ATOMIC_SEQ_CST);return nb_tx; }
@@ -184,13 +184,13 @@ struct pmd_internals {for (i =0 ; i <RTE_ETHDEV_QUEUE_STAT_CNTRS && i < dev->data->nb_rx_queues; i++) { - stats->q_ipackets[i] = internal->rx_ring_queues[i].rx_pkts.cnt; + stats->q_ipackets[i] = internal->rx_ring_queues[i].rx_pkts; rx_total += stats->q_ipackets[i]; }for (i =0 ; i <RTE_ETHDEV_QUEUE_STAT_CNTRS && i < dev->data->nb_tx_queues; i++) { - stats->q_opackets[i] = internal->tx_ring_queues[i].tx_pkts.cnt; + stats->q_opackets[i] = internal->tx_ring_queues[i].tx_pkts; tx_total += stats->q_opackets[i]; }
@@ -207,9 +207,9 @@ struct pmd_internals {struct pmd_internals *internal = dev->data->dev_private;for (i =0 ; i < dev->data->nb_rx_queues; i++) - internal->rx_ring_queues[i].rx_pkts.cnt =0 ; + internal->rx_ring_queues[i].rx_pkts =0 ;for (i =0 ; i < dev->data->nb_tx_queues; i++) - internal->tx_ring_queues[i].tx_pkts.cnt =0 ; + internal->tx_ring_queues[i].tx_pkts =0 ;return 0 ; }
Signature · 1 line
1.8.3.1
PATCH 5/7 net/ixgbe: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffReplace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
--- drivers/net/ixgbe/ixgbe_bypass.c | 1 - drivers/net/ixgbe/ixgbe_ethdev.c | 12 ++++++------ drivers/net/ixgbe/ixgbe_ethdev.h | 3 ++- drivers/net/ixgbe/ixgbe_flow.c | 1 - drivers/net/ixgbe/ixgbe_rxtx.c | 1 - 5 files changed, 8 insertions(+), 10 deletions(-)
drivers/net/ixgbe/ixgbe_bypass.c +0 -1
@@ -3,7 +3,6 @@*/ #include <time.h> -#include <rte_atomic.h> #include <ethdev_driver.h> #include "ixgbe_ethdev.h" #include "ixgbe_bypass_api.h"
drivers/net/ixgbe/ixgbe_ethdev.c +6 -6
@@ -1127,7 +1127,7 @@ struct rte_ixgbe_xstats_name_off {return 0 ; } -rte_atomic32_clear (&ad->link_thread_running); +__atomic_clear (&ad->link_thread_running, __ATOMIC_SEQ_CST);ixgbe_parse_devargs (eth_dev->data->dev_private, pci_dev->device.devargs);rte_eth_copy_pci_info (eth_dev, pci_dev);
@@ -1625,7 +1625,7 @@ static int ixgbe_l2_tn_filter_init(struct rte_eth_dev *eth_dev)return 0 ; } -rte_atomic32_clear (&ad->link_thread_running); +__atomic_clear (&ad->link_thread_running, __ATOMIC_SEQ_CST);ixgbevf_parse_devargs (eth_dev->data->dev_private, pci_dev->device.devargs);
@@ -4186,7 +4186,7 @@ static int ixgbevf_dev_xstats_get_names(__rte_unused struct rte_eth_dev *dev,struct ixgbe_adapter *ad = dev->data->dev_private;uint32_t timeout = timeout_ms ? timeout_ms :WARNING_TIMEOUT ; -while (rte_atomic32_read (&ad->link_thread_running)) { +while (__atomic_load_n (&ad->link_thread_running, __ATOMIC_SEQ_CST)) {msec_delay (1 ); timeout--;
@@ -4222,7 +4222,7 @@ static int ixgbevf_dev_xstats_get_names(__rte_unused struct rte_eth_dev *dev,ixgbe_setup_link (hw, speed, true); intr->flags &= ~IXGBE_FLAG_NEED_LINK_CONFIG ; -rte_atomic32_clear (&ad->link_thread_running); +__atomic_clear (&ad->link_thread_running, __ATOMIC_SEQ_CST);return NULL ; }
@@ -4317,7 +4317,7 @@ static int ixgbevf_dev_xstats_get_names(__rte_unused struct rte_eth_dev *dev,if (link_up ==0 ) {if (ixgbe_get_media_type (hw) == ixgbe_media_type_fiber) {ixgbe_dev_wait_setup_link_complete (dev,0 ); -if (rte_atomic32_test_and_set (&ad->link_thread_running)) { +if (__atomic_test_and_set (&ad->link_thread_running, __ATOMIC_SEQ_CST)) {/* To avoid race condition between threads, set * the IXGBE_FLAG_NEED_LINK_CONFIG flag only * when there is no link thread running.
@@ -4330,7 +4330,7 @@ static int ixgbevf_dev_xstats_get_names(__rte_unused struct rte_eth_dev *dev, dev) < 0) {PMD_DRV_LOG (ERR ,"Create link thread failed!" ); -rte_atomic32_clear (&ad->link_thread_running); +__atomic_clear (&ad->link_thread_running, __ATOMIC_SEQ_CST); } }else { PMD_DRV_LOG(ERR,
drivers/net/ixgbe/ixgbe_ethdev.h +2 -1
@@ -6,6 +6,7 @@#define _IXGBE_ETHDEV_H_#include <stdint.h> +#include <stdbool.h> #include <sys/queue.h> #include "base/ixgbe_type.h"
@@ -510,7 +511,7 @@ struct ixgbe_adapter {*/ uint8_t pflink_fullchk;uint8_t mac_ctrl_frame_fwd; -rte_atomic32_t link_thread_running; +bool link_thread_running;pthread_t link_thread_tid; };
drivers/net/ixgbe/ixgbe_flow.c +0 -1
@@ -18,7 +18,6 @@#include <rte_log.h> #include <rte_debug.h> #include <rte_pci.h> -#include <rte_atomic.h> #include <rte_branch_prediction.h> #include <rte_memory.h> #include <rte_eal.h>
drivers/net/ixgbe/ixgbe_rxtx.c +0 -1
@@ -27,7 +27,6 @@#include <rte_eal.h> #include <rte_per_lcore.h> #include <rte_lcore.h> -#include <rte_atomic.h> #include <rte_branch_prediction.h> #include <rte_mempool.h> #include <rte_malloc.h>
Signature · 1 line
1.8.3.1
Re: [PATCH 1/7] ring: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffQuoted 39 lines · Tyler Retzlaff
Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.Signed-off-by Tyler Retzlaff <[email protected]>--- lib/ring/rte_ring_core.h | 1 - lib/ring/rte_ring_generic_pvt.h | 10 ++++++---- 2 files changed, 6 insertions(+), 5 deletions(-)lib/ring/rte_ring_core.h +0
-1@@ -31,7 +31,6 @@#include <rte_config.h> #include <rte_memory.h> #include <rte_lcore.h> -#include <rte_atomic.h> #include <rte_branch_prediction.h> #include <rte_memzone.h> #include <rte_pause.h> lib/ring/rte_ring_generic_pvt.h +3
-2@@ -92,8 +92,9 @@if (is_sp) r->prod.head = *new_head, success =1 ;else - success =rte_atomic32_cmpset (&r->prod.head, - *old_head, *new_head); + success =__atomic_compare_exchange_n (&r->prod.head, + old_head, *new_head,0 , + __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST); }while (unlikely (success ==0 ));return n; }@@ -162,8 +163,9 @@rte_smp_rmb ();
lib/ring/rte_ring_generic_pvt.h line 171
Quoted 10 lines
lib/ring/rte_ring_generic_pvt.h +3
-2success =1 ; }else { - success =rte_atomic32_cmpset (&r->cons.head, *old_head, - *new_head); + success =__atomic_compare_exchange_n (&r->cons.head, + old_head, *new_head,0 , + __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST); } }while (unlikely (success ==0 ));return n;
just something i noticed and not related to this change.
i note that old_head for both __rte_ring_move_prod_head and __rte_ring_move_con_head are performing a non-atomic load to initialize *old_head probably not the best idea.
Re: [PATCH 4/7] net/ice: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffQuoted 51 lines · Tyler Retzlaff
Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.Signed-off-by Tyler Retzlaff <[email protected]>--- drivers/net/ice/ice_dcf.c | 1 - drivers/net/ice/ice_dcf_ethdev.c | 1 - drivers/net/ice/ice_ethdev.c | 10 ++++++---- 3 files changed, 6 insertions(+), 6 deletions(-)drivers/net/ice/ice_dcf.c +0
-1@@ -14,7 +14,6 @@#include <rte_common.h> #include <rte_pci.h> -#include <rte_atomic.h> #include <rte_eal.h> #include <rte_ether.h> #include <ethdev_driver.h> drivers/net/ice/ice_dcf_ethdev.c +0
-1@@ -11,7 +11,6 @@#include <rte_interrupts.h> #include <rte_debug.h> #include <rte_pci.h> -#include <rte_atomic.h> #include <rte_eal.h> #include <rte_ether.h> #include <ethdev_pci.h> drivers/net/ice/ice_ethdev.c +3
-2@@ -3927,8 +3927,9 @@ static int ice_init_rss(struct ice_pf *pf)struct rte_eth_link *dst = link;struct rte_eth_link *src = &dev->data->dev_link; -if (rte_atomic64_cmpset ((uint64_t *)dst, *(uint64_t *)dst, - *(uint64_t *)src) ==0 ) +if (!__atomic_compare_exchange_n ((uint64_t *)dst, + (uint64_t *)dst, *(uint64_t *)src,0 , + __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))return -1 ;return 0 ;
drivers/net/ice/ice_ethdev.c static int ice_init_rss(struct ice_pf *pf) · line 3949
Quoted 10 lines
drivers/net/ice/ice_ethdev.c +3
-2@@ -3941,8 +3942,9 @@ static int ice_init_rss(struct ice_pf *pf)struct rte_eth_link *dst = &dev->data->dev_link;struct rte_eth_link *src = link; -if (rte_atomic64_cmpset ((uint64_t *)dst, *(uint64_t *)dst, - *(uint64_t *)src) ==0 ) +if (!__atomic_compare_exchange_n ((uint64_t *)dst, + (uint64_t *)dst, *(uint64_t *)src,0 , + __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))return -1 ;
*(uint64_t *)dst for the second parameter look like a bug to me, a non-atomic load will be generated.
probably this code should be corrected by performing __atomic_load_n(dst, ...) to a stack variable and then performing the cmpset/compare_exchange.
Re: [PATCH 6/7] net/null: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffQuoted 69 lines · Tyler Retzlaff
Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.Signed-off-by Tyler Retzlaff <[email protected]>--- drivers/net/null/rte_eth_null.c | 20 ++++++++++---------- 1 file changed, 10 insertions(+), 10 deletions(-)drivers/net/null/rte_eth_null.c +7
-7@@ -37,8 +37,8 @@ struct null_queue {struct rte_mempool *mb_pool;struct rte_mbuf *dummy_packet; -rte_atomic64_t rx_pkts; -rte_atomic64_t tx_pkts; +int64_t rx_pkts; +int64_t tx_pkts; };struct pmd_options {@@ -101,7 +101,7 @@ struct pmd_internals { bufs[i]->port = h->internals->port_id; } -rte_atomic64_add (&(h->rx_pkts), i); +__atomic_fetch_add (&h->rx_pkts, i, __ATOMIC_SEQ_CST);return i; }@@ -128,7 +128,7 @@ struct pmd_internals { bufs[i]->port = h->internals->port_id; } -rte_atomic64_add (&(h->rx_pkts), i); +__atomic_fetch_add (&h->rx_pkts, i, __ATOMIC_SEQ_CST);return i; }@@ -152,7 +152,7 @@ struct pmd_internals {for (i =0 ; i < nb_bufs; i++)rte_pktmbuf_free (bufs[i]); -rte_atomic64_add (&(h->tx_pkts), i); +__atomic_fetch_add (&h->tx_pkts, i, __ATOMIC_SEQ_CST);return i; }@@ -174,7 +174,7 @@ struct pmd_internals {rte_pktmbuf_free (bufs[i]); } -rte_atomic64_add (&(h->tx_pkts), i); +__atomic_fetch_add (&h->tx_pkts, i, __ATOMIC_SEQ_CST);return i; }@@ -317,7 +317,7 @@ struct pmd_internals { RTE_DIM(internal->rx_null_queues)));for (i =0 ; i < num_stats; i++) { igb_stats->q_ipackets[i] = - internal->rx_null_queues[i].rx_pkts.cnt; + internal->rx_null_queues[i].rx_pkts; rx_total += igb_stats->q_ipackets[i]; }
drivers/net/null/rte_eth_null.c struct pmd_internals { · line 332
Quoted 8 lines
drivers/net/null/rte_eth_null.c +1
-1@@ -326,7 +326,7 @@ struct pmd_internals { RTE_DIM(internal->tx_null_queues)));for (i =0 ; i < num_stats; i++) { igb_stats->q_opackets[i] = - internal->tx_null_queues[i].tx_pkts.cnt; + internal->tx_null_queues[i].tx_pkts; tx_total += igb_stats->q_opackets[i]; }
these variables are operated on with atomic builtins in other places yet here they are being non-atomically loaded. should probably be using _atomic_load_n(...)
drivers/net/null/rte_eth_null.c struct pmd_internals { · line 352
Quoted 9 lines
drivers/net/null/rte_eth_null.c +2
-2@@ -347,9 +347,9 @@ struct pmd_internals { internal = dev->data->dev_private;for (i =0 ; i <RTE_DIM (internal->rx_null_queues); i++) - internal->rx_null_queues[i].rx_pkts.cnt =0 ; + internal->rx_null_queues[i].rx_pkts =0 ;for (i =0 ; i <RTE_DIM (internal->tx_null_queues); i++) - internal->tx_null_queues[i].tx_pkts.cnt =0 ; + internal->tx_null_queues[i].tx_pkts =0 ;
same thing, these should probably be __atomic_store_n(...)
Quoted 8 lines · On Fri, 17 Mar 2023 13:19:41 -0700 Tyler Retzlaff…
Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.This series covers the libraries and drivers that are built on Windows.
The code has be converted to use the
__atomicbuiltins but there are additional during conversion i notice that there may be some issues that need to be addressed.
I don't think all these cmpset need to use SEQ_CST. Especially for the places where it is used a loop, might be more efficient with some of the other memory models.
Quoted 8 lines · Stephen Hemminger
Quoted 5 lines · On Fri, 17 Mar 2023 13:19:41 -0700 Tyler Retzlaff…
Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.This series covers the libraries and drivers that are built on Windows.
Quoted 7 lines
Quoted 3 lines
The code has be converted to use the
__atomicbuiltins but there are additional during conversion i notice that there may be some issues that need to be addressed.I don't think all these cmpset need to use
SEQ_CST. Especially for the places where it is used a loop, might be more efficient with some of the other memory models.
i agree.
however, i'm not trying to improve the code with this change, just decouple it from rte_atomics.h so trying my best to avoid any unnecessary semantic change.
certainly if the maintainers of this code wish to weaken the ordering where appropriate after the change is merged they can do so and handily this change has enabled them to do so easily allowing them to test just their change in isolation.
Quoted 22 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Friday, 17 March 2023 22.49Quoted 8 lines · Stephen Hemminger
Quoted 5 lines · On Fri, 17 Mar 2023 13:19:41 -0700 Tyler Retzlaff…
Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.This series covers the libraries and drivers that are built on Windows.
Quoted 7 lines
Quoted 3 lines
The code has be converted to use the
__atomicbuiltins but there are additional during conversion i notice that there may be some issues that need to be addressed.I don't think all these cmpset need to use
SEQ_CST. Especially for the places where it is used a loop, might be more efficient with some of the other memory models.i agree.
Quoted 8 lines
however, i'm not trying to improve the code with this change, just decouple it from
rte_atomics.hso trying my best to avoid any unnecessary semantic change.certainly if the maintainers of this code wish to weaken the ordering where appropriate after the change is merged they can do so and handily this change has enabled them to do so easily allowing them to test just their change in isolation.
I agree with the two-step approach, where this first step is a simple search-and-replacement; but I insist that you add a FIXME or similar note where you have blindly used SEQ_CST, indicating that the memory order needs to be reviewed and potentially corrected.
Also, in a couple of the drivers, you are using int64_t for packet counters. These cannot be negative and should be uint64_t. And AFAIK, such counters can use RELAXED memory order.
Quoted 26 lines · Morten Brørup
Quoted 26 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Friday, 17 March 2023 22.49Quoted 8 lines · Stephen Hemminger
Quoted 5 lines · On Fri, 17 Mar 2023 13:19:41 -0700 Tyler Retzlaff…
Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.This series covers the libraries and drivers that are built on Windows.
Quoted 7 lines
Quoted 3 lines
The code has be converted to use the
__atomicbuiltins but there are additional during conversion i notice that there may be some issues that need to be addressed.I don't think all these cmpset need to use
SEQ_CST. Especially for the places where it is used a loop, might be more efficient with some of the other memory models.i agree.
however, i'm not trying to improve the code with this change, just decouple it from
rte_atomics.hso trying my best to avoid any unnecessary semantic change.
Quoted 6 lines
Quoted 4 lines
certainly if the maintainers of this code wish to weaken the ordering where appropriate after the change is merged they can do so and handily this change has enabled them to do so easily allowing them to test just their change in isolation.
I agree with the two-step approach, where this first step is a simple search-and-replacement; but I insist that you add a FIXME or similar note where you have blindly used
SEQ_CST, indicating that the memory order needs to be reviewed and potentially corrected.
i think the maintainers need to take some responsibility, if they see optimizations they missed when previously writing the code they need to follow up with a patch themselves. i can't do everything for them and marking things i'm not sure about will only lead to me having to churn patch series to remove the unwanted comments later.
keep in mind i have to touch each of these again when converting to standard so that's a better time to review ~everything in more detail because when converting to standard that's when suddenly you get a bunch of code generation that is "fallback" to seq_cst that isn't happening now.
the series that converts to standard needs to be up for review as soon as possible to maximize available time for feedback before 23.11 so it would be better to get the simpler cut & paste normalizing the code out of the way to unblock that series submission.
Quoted 2 lines
Also, in a couple of the drivers, you are using
int64_tfor packet counters. These cannot be negative and should beuint64_t. And AFAIK, such counters can use RELAXED memory order.
i know you don't mean to say i selected the types and rather that the types that were selected are not quite correct for their usage. again on the review that actually adopts std atomics is a better place to make any potential type changes since we are "breaking" the API for 23.11 anyway. further, the std atomics series technically changes all the types so it's probably better to make one type change then rather than one now and one later.
i think it would be best to get these validated and merged asap so we can get to the std atomics review. when that series is up let's discuss further how i can mark areas of concern, with that series i expect there will have to be some changes in order to avoid minor regressions.
thanks!
Quoted 41 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Wednesday, 22 March 2023 15.22Quoted 11 lines · Morten Brørup
Quoted 11 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Friday, 17 March 2023 22.49Quoted 7 lines · Stephen Hemminger
Quoted 4 lines · On Fri, 17 Mar 2023 13:19:41 -0700 Tyler Retzlaff…
Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.This series covers the libraries and drivers that are built on
Windows.
Quoted 15 lines
Quoted 15 lines
Quoted 8 lines
Quoted 4 lines
The code has be converted to use the
__atomicbuiltins but there are additional during conversion i notice that there may be some issues that need to be addressed.I don't think all these cmpset need to use
SEQ_CST. Especially for the places where it is used a loop, might be more efficient with some of the other memory models.i agree.
however, i'm not trying to improve the code with this change, just decouple it from
rte_atomics.hso trying my best to avoid any unnecessary semantic change.Quoted 6 lines
Quoted 4 lines
certainly if the maintainers of this code wish to weaken the ordering where appropriate after the change is merged they can do so and handily this change has enabled them to do so easily allowing them to test just their change in isolation.
I agree with the two-step approach, where this first step is a simple
search-and-replacement; but I insist that you add a FIXME or similar note where you have blindly used
SEQ_CST, indicating that the memory order needs to be reviewed and potentially corrected.
Quoted 5 lines
i think the maintainers need to take some responsibility, if they see optimizations they missed when previously writing the code they need to follow up with a patch themselves. i can't do everything for them and marking things i'm not sure about will only lead to me having to churn patch series to remove the unwanted comments later.
The previous atomic functions didn't have the "memory order" parameter, so the maintainers didn't have to think about it - and thus they didn't miss any optimizations when accepting the code.
I also agree 100 % that it is not your responsibility to consider or determine which memory order is appropriate!
But I think you should mark the locations where you are changing from the old rte_atomic functions (where no memory order optimization was available) to the new functions - to highlight where the option of memory ordering has been introduced and knowingly ignored (by you).
Quoted 5 lines
keep in mind i have to touch each of these again when converting to standard so that's a better time to review ~everything in more detail because when converting to standard that's when suddenly you get a bunch of code generation that is "fallback" to
seq_cstthat isn't happening now.
I think you should to do it when replacing the rte_atomic functions with the __atomic functions. It will make it easier to see where the memory order was knowingly ignored, and should be reviewed for optimization.
Quoted 7 lines
the series that converts to standard needs to be up for review as soon as possible to maximize available time for feedback before 23.11 so it would be better to get the simpler cut & paste normalizing the code out of the way to unblock that series submission.
Quoted 1 line
Quoted 6 lines
Quoted 1 line
Also, in a couple of the drivers, you are using
int64_tfor packet counters.These cannot be negative and should be
uint64_t. And AFAIK, such counters can use RELAXED memory order.i know you don't mean to say i selected the types and rather that the types that were selected are not quite correct for their usage.
Yes; the previous types were also signed, and you didn't change that.
Quoted 7 lines
again
on the review that actually adopts std atomics is a better place to make any potential type changes since we are "breaking" the API for 23.11 anyway. further, the std atomics series technically changes all the types so it's probably better to make one type change then rather than one now and one later.
Quoted 6 lines
i think it would be best to get these validated and merged asap so we can get to the std atomics review. when that series is up let's discuss further how i can mark areas of concern, with that series i expect there will have to be some changes in order to avoid minor regressions.
thanks!
I thought it would be better to catch these details (i.e. memory ordering and signedness) early on, but I now understand that you planned to do it in a later step. So I'll let you proceed as you have planned.
Thanks for all your work on this, Tyler. It is much appreciated!
-Morten
Quoted 47 lines · Morten Brørup
Quoted 41 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Wednesday, 22 March 2023 15.22Quoted 11 lines · Morten Brørup
Quoted 11 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Friday, 17 March 2023 22.49Quoted 7 lines · Stephen Hemminger
Quoted 4 lines · On Fri, 17 Mar 2023 13:19:41 -0700 Tyler Retzlaff…
Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.This series covers the libraries and drivers that are built on
Windows.
Quoted 15 lines
Quoted 15 lines
Quoted 8 lines
Quoted 4 lines
The code has be converted to use the
__atomicbuiltins but there are additional during conversion i notice that there may be some issues that need to be addressed.I don't think all these cmpset need to use
SEQ_CST. Especially for the places where it is used a loop, might be more efficient with some of the other memory models.i agree.
however, i'm not trying to improve the code with this change, just decouple it from
rte_atomics.hso trying my best to avoid any unnecessary semantic change.Quoted 6 lines
Quoted 4 lines
certainly if the maintainers of this code wish to weaken the ordering where appropriate after the change is merged they can do so and handily this change has enabled them to do so easily allowing them to test just their change in isolation.
I agree with the two-step approach, where this first step is a simple
search-and-replacement; but I insist that you add a FIXME or similar note where you have blindly used
SEQ_CST, indicating that the memory order needs to be reviewed and potentially corrected.Quoted 5 lines
i think the maintainers need to take some responsibility, if they see optimizations they missed when previously writing the code they need to follow up with a patch themselves. i can't do everything for them and marking things i'm not sure about will only lead to me having to churn patch series to remove the unwanted comments later.
Quoted 6 lines
The previous atomic functions didn't have the "memory order" parameter, so the maintainers didn't have to think about it - and thus they didn't miss any optimizations when accepting the code.
I also agree 100 % that it is not your responsibility to consider or determine which memory order is appropriate!
But I think you should mark the locations where you are changing from the old
rte_atomicfunctions (where no memory order optimization was available) to the new functions - to highlight where the option of memory ordering has been introduced and knowingly ignored (by you).
first, i have to apologize i confused myself about which of the many patch series i have up right now that you were commenting on.
let me ask for clarification in relation to this series.
isn't that every single usage of the rte_atomic APIs? i mean are you literally asking for the entire patch series to look like the following patch snippet with the expectation that maintainers will come along and clean up/review after this series is merged?
-rte_atomic_add32(&o, v);
+//FIXME: opportunity for relaxing ordering constraint, please review +__atomic_fetch_add(&o, v, order);
this would just be a mechanical addition to this series so i can certainly accomodate that, i thought something more complicated was being asked for. if this is all, then sure no problem.
Quoted 35 lines
Quoted 4 lines
keep in mind i have to touch each of these again when converting to standard so that's a better time to review ~everything in more detail because when converting to standard that's when suddenly you get a bunch of code generation that is "fallback" to
seq_cstthat isn't happening now.I think you should to do it when replacing the
rte_atomicfunctions with the__atomicfunctions. It will make it easier to see where the memory order was knowingly ignored, and should be reviewed for optimization.Quoted 7 lines
the series that converts to standard needs to be up for review as soon as possible to maximize available time for feedback before 23.11 so it would be better to get the simpler cut & paste normalizing the code out of the way to unblock that series submission.
Quoted 1 line
Quoted 6 lines
Quoted 1 line
Also, in a couple of the drivers, you are using
int64_tfor packet counters.These cannot be negative and should be
uint64_t. And AFAIK, such counters can use RELAXED memory order.i know you don't mean to say i selected the types and rather that the types that were selected are not quite correct for their usage.
Yes; the previous types were also signed, and you didn't change that.
Quoted 12 lines
again
on the review that actually adopts std atomics is a better place to make any potential type changes since we are "breaking" the API for 23.11 anyway. further, the std atomics series technically changes all the types so it's probably better to make one type change then rather than one now and one later.i think it would be best to get these validated and merged asap so we can get to the std atomics review. when that series is up let's discuss further how i can mark areas of concern, with that series i expect there will have to be some changes in order to avoid minor regressions.
Quoted 5 lines
Quoted 1 line
thanks!
I thought it would be better to catch these details (i.e. memory ordering and signedness) early on, but I now understand that you planned to do it in a later step. So I'll let you proceed as you have planned.
Thanks for all your work on this, Tyler. It is much appreciated!
again, sorry for the confusion the sooner i can get some of these merged the easier it will be for me to manage the final series. i hope david/thomas can merge the simple normalization patches as soon as 23.03 cycle is complete.
Quoted 2 lines
-Morten
Quoted 64 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Wednesday, 22 March 2023 16.30Quoted 4 lines · Morten Brørup
Quoted 4 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Wednesday, 22 March 2023 15.22On Wed, Mar 22, 2023 at 12:28:44PM +0100, Morten Brørup wrote:
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Friday, 17 March 2023 22.49Quoted 4 lines · Stephen Hemminger
Quoted 1 line · On Fri, 17 Mar 2023 13:19:41 -0700 Tyler Retzlaff…
Replace the use of
rte_atomic.htypes and functions, instead useGCC
Quoted 6 lines
Quoted 6 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
supplied C++11 memory model builtins.
This series covers the libraries and drivers that are built on
Windows.
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
The code has be converted to use the
__atomicbuiltins but thereare
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
additional during conversion i notice that there may be some
issues
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 5 lines
Quoted 1 line
that need to be addressed.
I don't think all these cmpset need to use
SEQ_CST. Especially for the places where it is used a loop, might be more efficient with some of the other memory models.Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
i agree.
however, i'm not trying to improve the code with this change, just decouple it from
rte_atomics.hso trying my best to avoid any unnecessary semantic change.certainly if the maintainers of this code wish to weaken the ordering where appropriate after the change is merged they can do so and
handily
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
this change has enabled them to do so easily allowing them to test
just
Quoted 5 lines
Quoted 5 lines
Quoted 3 lines
Quoted 1 line
their change in isolation.
I agree with the two-step approach, where this first step is a simple
search-and-replacement; but I insist that you add a FIXME or similar note where you have blindly used
SEQ_CST, indicating that the memory orderneeds to
Quoted 9 lines
Quoted 7 lines
be reviewed and potentially corrected.
i think the maintainers need to take some responsibility, if they see optimizations they missed when previously writing the code they need to follow up with a patch themselves. i can't do everything for them and marking things i'm not sure about will only lead to me having to churn patch series to remove the unwanted comments later.
The previous atomic functions didn't have the "memory order" parameter, so
the maintainers didn't have to think about it - and thus they didn't miss any optimizations when accepting the code.
Quoted 2 lines
I also agree 100 % that it is not your responsibility to consider or
determine which memory order is appropriate!
Quoted 1 line
Quoted 8 lines
Quoted 1 line
But I think you should mark the locations where you are changing from the
old
rte_atomicfunctions (where no memory order optimization was available) to the new functions - to highlight where the option of memory ordering has been introduced and knowingly ignored (by you).Quoted 1 line
first, i have to apologize i confused myself about which of the many patch series i have up right now that you were commenting on.
No worries... you are rushing through quite an effort for this, so a little confusion is perfectly understandable. Especially when I'm replying to an ageing email. :-)
Quoted 4 lines
let me ask for clarification in relation to this series.
isn't that every single usage of the
rte_atomicAPIs?
Probably, yes.
Review
Quoted 8 lines
i mean are you
literally asking for the entire patch series to look like the following patch snippet with the expectation that maintainers will come along and clean up/review after this series is merged?(quoted diff) +2
-1-rte_atomic_add32(&o, v); +//FIXME: opportunity for relaxing ordering constraint, please review +__atomic_fetch_add(&o, v, order);
Exactly. And something similar for the rte_atomicXX_t variables changed to intXX_t, such as the packet counters.
Realistically, I don't expect the maintainers to clean them up anytime soon. The purpose is to make the FIXMEs stick until someone eventually cleans them up, so they are not forgotten as time passes.
Quoted 4 lines
this would just be a mechanical addition to this series so i can certainly accomodate that, i thought something more complicated was being asked for. if this is all, then sure no problem.
Great.
Quoted 46 lines
Quoted 7 lines
Quoted 5 lines
keep in mind i have to touch each of these again when converting to standard so that's a better time to review ~everything in more detail because when converting to standard that's when suddenly you get a bunch of code generation that is "fallback" to
seq_cstthat isn't happening now.I think you should to do it when replacing the
rte_atomicfunctions with the
__atomicfunctions. It will make it easier to see where the memory order was knowingly ignored, and should be reviewed for optimization.Quoted 9 lines
Quoted 9 lines
the series that converts to standard needs to be up for review as soon as possible to maximize available time for feedback before 23.11 so it would be better to get the simpler cut & paste normalizing the code out of the way to unblock that series submission.
Quoted 2 lines
Also, in a couple of the drivers, you are using
int64_tfor packetcounters.
Quoted 1 line
Quoted 1 line
These cannot be negative and should be
uint64_t. And AFAIK, such counterscan
Quoted 14 lines
Quoted 4 lines
use RELAXED memory order.
i know you don't mean to say i selected the types and rather that the types that were selected are not quite correct for their usage.
Yes; the previous types were also signed, and you didn't change that.
Quoted 7 lines
again
on the review that actually adopts std atomics is a better place to make any potential type changes since we are "breaking" the API for 23.11 anyway. further, the std atomics series technically changes all the types so it's probably better to make one type change then rather than one now and one later.Quoted 8 lines
Quoted 6 lines
i think it would be best to get these validated and merged asap so we can get to the std atomics review. when that series is up let's discuss further how i can mark areas of concern, with that series i expect there will have to be some changes in order to avoid minor regressions.
thanks!
I thought it would be better to catch these details (i.e. memory ordering
and signedness) early on, but I now understand that you planned to do it in a later step. So I'll let you proceed as you have planned.
Quoted 1 line
Quoted 6 lines
Quoted 1 line
Thanks for all your work on this, Tyler. It is much appreciated!
again, sorry for the confusion the sooner i can get some of these merged the easier it will be for me to manage the final series. i hope david/thomas can merge the simple normalization patches as soon as 23.03 cycle is complete.
Yes. An early merge would also provide more time for reviewing and optimizing the memory order of the most important atomic operations.
Quoted 119 lines
-----Original Message-----
From: Morten Brørup <[email protected]>
Sent: Wednesday, March 22, 2023 11:14 AM
To: Tyler Retzlaff <[email protected]>
Cc: Stephen Hemminger <[email protected]>; [email protected]; Honnappa Nagarahalli <[email protected]>; Ruifeng Wang <[email protected]>; [email protected]
Subject: RE: [PATCH 0/7] replace rte atomics with GCC builtin atomicsQuoted 12 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Wednesday, 22 March 2023 16.30Quoted 8 lines · Morten Brørup
Quoted 8 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Wednesday, 22 March 2023 15.22Quoted 4 lines · Morten Brørup
Quoted 4 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Friday, 17 March 2023 22.49On Fri, Mar 17, 2023 at 02:42:26PM -0700, Stephen Hemminger
Quoted 37 lines · wrote:
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 2 lines · On Fri, 17 Mar 2023 13:19:41 -0700 Tyler Retzlaff…
Replace the use of
rte_atomic.htypes and functions, instead useGCC
Quoted 8 lines
Quoted 8 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
supplied C++11 memory model builtins.
This series covers the libraries and drivers that are built on
Windows.
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
The code has be converted to use the
__atomicbuiltins but thereare
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
additional during conversion i notice that there may be some
issues
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 5 lines
Quoted 1 line
that need to be addressed.
I don't think all these cmpset need to use
SEQ_CST. Especially for the places where it is used a loop, might be more efficient with some of the other memory models.i agree.
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
however, i'm not trying to improve the code with this change, just decouple it from
rte_atomics.hso trying my best to avoid any unnecessary semantic change.certainly if the maintainers of this code wish to weaken the ordering where appropriate after the change is merged they can do so and
handily
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
this change has enabled them to do so easily allowing them to test
just
Quoted 15 lines
Quoted 7 lines
Quoted 7 lines
Quoted 4 lines
Quoted 1 line
their change in isolation.
I agree with the two-step approach, where this first step is a simple
search-and-replacement; but I insist that you add a FIXME or similar note where you have blindly used
SEQ_CST, indicating that the memory orderneeds to
Quoted 7 lines
Quoted 7 lines
be reviewed and potentially corrected.
i think the maintainers need to take some responsibility, if they see optimizations they missed when previously writing the code they need to follow up with a patch themselves. i can't do everything for them and marking things i'm not sure about will only lead to me having to churn patch series to remove the unwanted
comments later.
Quoted 11 lines
Quoted 3 lines
The previous atomic functions didn't have the "memory order" parameter, so
the maintainers didn't have to think about it - and thus they didn't miss any optimizations when accepting the code.
Quoted 2 lines
I also agree 100 % that it is not your responsibility to consider or
determine which memory order is appropriate!
Quoted 3 lines
But I think you should mark the locations where you are changing from the
Quoted 7 lines
old
rte_atomicfunctions (where no memory order optimization was available) to the new functions - to highlight where the option of memory ordering has been introduced and knowingly ignored (by you).Quoted 1 line
first, i have to apologize i confused myself about which of the many patch series i have up right now that you were commenting on.
No worries... you are rushing through quite an effort for this, so a little confusion is perfectly understandable. Especially when I'm replying to an ageing email. :-)
Quoted 4 lines
let me ask for clarification in relation to this series.
isn't that every single usage of the
rte_atomicAPIs?Probably, yes.
Review
Quoted 8 lines
i mean are you
literally asking for the entire patch series to look like the following patch snippet with the expectation that maintainers will come along and clean up/review after this series is merged?(quoted diff) +2
-1-rte_atomic_add32(&o, v); +//FIXME: opportunity for relaxing ordering constraint, please review +__atomic_fetch_add(&o, v, order);Exactly. And something similar for the
rte_atomicXX_tvariables changed tointXX_t, such as the packet counters.
Quoted 3 lines
Realistically, I don't expect the maintainers to clean them up anytime soon. The purpose is to make the FIXMEs stick until someone eventually cleans them up, so they are not forgotten as time passes.
Cleaning up the rte_atomic APIs is a different effort. There is already lot of effort that has gone into this and there is more effort happening (rte_ring being a painful one)
Instead of having FIXME, why not just send a separate patch with SEQ_CST (still a search and replace)? We can leave the tougher ones like rte_ring as they are being worked on.
Quoted 69 lines
Quoted 5 lines
this would just be a mechanical addition to this series so i can certainly accomodate that, i thought something more complicated was being asked for. if this is all, then sure no problem.
Great.
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
keep in mind i have to touch each of these again when converting to standard so that's a better time to review ~everything in more detail because when converting to standard that's when suddenly you get a bunch of code generation that is "fallback" to
seq_cstthat isn'thappening now.
Quoted 13 lines
Quoted 3 lines
I think you should to do it when replacing the
rte_atomicfunctions with the
__atomicfunctions. It will make it easier to see where the memory order was knowingly ignored, and should be reviewed for optimization.Quoted 8 lines
Quoted 8 lines
the series that converts to standard needs to be up for review as soon as possible to maximize available time for feedback before 23.11 so it would be better to get the simpler cut & paste normalizing the code out of the way to unblock that series submission.
Quoted 1 line
Quoted 24 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Also, in a couple of the drivers, you are using
int64_tfor packetcounters.
Quoted 2 lines
Quoted 2 lines
These cannot be negative and should be
uint64_t. And AFAIK, such counterscan
Quoted 18 lines
Quoted 4 lines
use RELAXED memory order.
i know you don't mean to say i selected the types and rather that the types that were selected are not quite correct for their usage.
Yes; the previous types were also signed, and you didn't change that.
Quoted 11 lines
again
on the review that actually adopts std atomics is a better place to make any potential type changes since we are "breaking" the API for 23.11 anyway. further, the std atomics series technically changes all the types so it's probably better to make one type change then rather than one now and one later.i think it would be best to get these validated and merged asap so we can get to the std atomics review. when that series is up let's discuss further how i can mark areas of concern, with that series i expect there will have to be some changes in order to avoid minor
regressions.
Quoted 8 lines
Quoted 5 lines
Quoted 2 lines
thanks!
I thought it would be better to catch these details (i.e. memory ordering
and signedness) early on, but I now understand that you planned to do it in a later step. So I'll let you proceed as you have planned.
Quoted 1 line
Quoted 6 lines
Quoted 1 line
Thanks for all your work on this, Tyler. It is much appreciated!
again, sorry for the confusion the sooner i can get some of these merged the easier it will be for me to manage the final series. i hope david/thomas can merge the simple normalization patches as soon as 23.03 cycle is complete.
Yes. An early merge would also provide more time for reviewing and optimizing the memory order of the most important atomic operations.
Quoted 132 lines
From: Honnappa Nagarahalli [mailto:[email protected]]
Sent: Wednesday, 22 March 2023 17.40Quoted 10 lines
From: Morten Brørup <[email protected]>
Sent: Wednesday, March 22, 2023 11:14 AMQuoted 7 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Wednesday, 22 March 2023 16.30Quoted 3 lines · Morten Brørup
Quoted 3 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Wednesday, 22 March 2023 15.22Quoted 19 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines · Morten Brørup
Quoted 4 lines
Quoted 4 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Friday, 17 March 2023 22.49On Fri, Mar 17, 2023 at 02:42:26PM -0700, Stephen Hemminger
Quoted 13 lines · wrote:
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 2 lines · On Fri, 17 Mar 2023 13:19:41 -0700 Tyler Retzlaff…
Replace the use of
rte_atomic.htypes and functions, instead useGCC
Quoted 7 lines
Quoted 7 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
supplied C++11 memory model builtins.
This series covers the libraries and drivers that are built on
Windows.
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
The code has be converted to use the
__atomicbuiltinsbut
Quoted 9 lines
Quoted 9 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
there
are
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
additional during conversion i notice that there may be some
issues
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 1 line
that need to be addressed.
I don't think all these cmpset need to use
SEQ_CST. Especially for the places where it is used a loop, mightbe
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 1 line
more efficient with some of the other memory models.
i agree.
however, i'm not trying to improve the code with this
change,
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
just decouple it from
rte_atomics.hso trying my best toavoid
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
any unnecessary semantic change.
certainly if the maintainers of this code wish to weaken the ordering where appropriate after the change is merged they
can
Quoted 3 lines
Quoted 3 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
do so and
handily
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
this change has enabled them to do so easily allowing them
to
Quoted 8 lines
Quoted 8 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
test
just
Quoted 6 lines
Quoted 6 lines
Quoted 4 lines
Quoted 1 line
their change in isolation.
I agree with the two-step approach, where this first step is a simple
search-and-replacement; but I insist that you add a FIXME or similar note where you have blindly used
SEQ_CST, indicatingthat
Quoted 5 lines
Quoted 5 lines
Quoted 1 line
Quoted 1 line
the memory order
needs to
Quoted 3 lines
Quoted 3 lines
be reviewed and potentially corrected.
i think the maintainers need to take some responsibility, if
they
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
see optimizations they missed when previously writing the code they need to follow up with a patch themselves. i can't do everything for them and marking things i'm not sure about will only lead to me having to churn patch series to remove the
unwanted
Quoted 8 lines
comments later.
Quoted 7 lines
Quoted 3 lines
The previous atomic functions didn't have the "memory order" parameter, so
the maintainers didn't have to think about it - and thus they didn't miss any optimizations when accepting the code.
Quoted 2 lines
I also agree 100 % that it is not your responsibility to consider
or
Quoted 9 lines
Quoted 9 lines
determine which memory order is appropriate!
Quoted 3 lines
But I think you should mark the locations where you are changing from the
old
rte_atomicfunctions (where no memory order optimization was available) to the new functions - to highlight where the option of memory ordering has been introduced and knowingly ignored (by you).Quoted 1 line
Quoted 4 lines
Quoted 2 lines
first, i have to apologize i confused myself about which of the many patch series i have up right now that you were commenting on.
No worries... you are rushing through quite an effort for this, so a
little
Quoted 1 line
confusion is perfectly understandable. Especially when I'm replying to
an ageing
Quoted 9 lines
email. :-)
Quoted 4 lines
let me ask for clarification in relation to this series.
isn't that every single usage of the
rte_atomicAPIs?Probably, yes.
Quoted 7 lines
Review
Quoted 7 lines
i mean are you
literally asking for the entire patch series to look like the following patch snippet with the expectation that maintainers will come along and clean up/review after this series is merged?(quoted diff) +1
-1-rte_atomic_add32(&o, v); +//FIXME: opportunity for relaxing ordering constraint, pleasereview
Quoted 3 lines
Quoted 1 line
+__atomic_fetch_add(&o, v, order);
Exactly. And something similar for the
rte_atomicXX_tvariableschanged to
Quoted 3 lines
intXX_t, such as the packet counters.Realistically, I don't expect the maintainers to clean them up anytime
soon. The
Quoted 1 line
purpose is to make the FIXMEs stick until someone eventually cleans
them up, so
Quoted 8 lines
Quoted 1 line
they are not forgotten as time passes.
Cleaning up the
rte_atomicAPIs is a different effort. There is already lot of effort that has gone into this and there is more effort happening (rte_ringbeing a painful one)Instead of having FIXME, why not just send a separate patch with
SEQ_CST(still a search and replace)? We can leave the tougher ones likerte_ringas they are being worked on.
The FIXME makes it possible in the future to differentiate between the instances that still need review and the instances that have been reviewed where SEQ_CST was the correct choice. (Similarly for the choice of type for variables previously rte_atomicNN_t.)
Quoted 88 lines
Quoted 12 lines
Quoted 6 lines
this would just be a mechanical addition to this series so i can certainly accomodate that, i thought something more complicated was being asked for. if this is all, then sure no problem.
Great.
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
keep in mind i have to touch each of these again when converting to standard so that's a better time to review ~everything in
more
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
detail because when converting to standard that's when suddenly you get a bunch of code generation that is "fallback" to
seq_cstthat isn't
Quoted 3 lines
happening now.
Quoted 2 lines
Quoted 2 lines
I think you should to do it when replacing the
rte_atomicfunctions
Quoted 3 lines
Quoted 3 lines
Quoted 1 line
with the
__atomicfunctions. It will make it easier to see where the memory order was knowingly ignored, and should be reviewed foroptimization.
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
the series that converts to standard needs to be up for review
as
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
soon as possible to maximize available time for feedback before 23.11 so it would be better to get the simpler cut & paste normalizing the code out of the way to unblock that series
submission.
Quoted 11 lines
Quoted 11 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Also, in a couple of the drivers, you are using
int64_tfor packetcounters.
Quoted 2 lines
Quoted 2 lines
These cannot be negative and should be
uint64_t. And AFAIK, such counterscan
Quoted 3 lines
Quoted 3 lines
use RELAXED memory order.
i know you don't mean to say i selected the types and rather
that
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
the types that were selected are not quite correct for their
usage.
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Yes; the previous types were also signed, and you didn't change
that.
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
again
on the review that actually adopts std atomics is a better place to make any potential type changes since we are "breaking" theAPI
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
for 23.11 anyway. further, the std atomics series technically changes all the types so it's probably better to make one type change then rather than one now and one later.
i think it would be best to get these validated and merged asap
so
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
we can get to the std atomics review. when that series is up
let's
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
discuss further how i can mark areas of concern, with that
series
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
i expect there will have to be some changes in order to avoid
minor
Quoted 7 lines
regressions.
Quoted 6 lines
Quoted 5 lines
Quoted 2 lines
thanks!
I thought it would be better to catch these details (i.e. memory ordering
and signedness) early on, but I now understand that you planned to
do
Quoted 6 lines
Quoted 6 lines
it in a later step. So I'll let you proceed as you have planned.
Quoted 2 lines
Thanks for all your work on this, Tyler. It is much appreciated!
again, sorry for the confusion the sooner i can get some of these merged the easier it will be for me to manage the final series. i
hope
Quoted 4 lines
Quoted 2 lines
david/thomas can merge the simple normalization patches as soon as 23.03 cycle is complete.
Yes. An early merge would also provide more time for reviewing and
optimizing
Quoted 1 line
the memory order of the most important atomic operations.
Quoted 154 lines
-----Original Message-----
From: Morten Brørup <[email protected]>
Sent: Wednesday, March 22, 2023 12:08 PM
To: Honnappa Nagarahalli <[email protected]>; Tyler Retzlaff <[email protected]>
Cc: Stephen Hemminger <[email protected]>; [email protected]; Ruifeng Wang <[email protected]>; [email protected]; nd <[email protected]>; nd <[email protected]>
Subject: RE: [PATCH 0/7] replace rte atomics with GCC builtin atomicsQuoted 142 lines
From: Honnappa Nagarahalli [mailto:[email protected]]
Sent: Wednesday, 22 March 2023 17.40Quoted 10 lines
From: Morten Brørup <[email protected]>
Sent: Wednesday, March 22, 2023 11:14 AMQuoted 7 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Wednesday, 22 March 2023 16.30Quoted 3 lines · Morten Brørup
Quoted 3 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Wednesday, 22 March 2023 15.22Quoted 20 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines · Morten Brørup
Quoted 5 lines
Quoted 5 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Friday, 17 March 2023 22.49On Fri, Mar 17, 2023 at 02:42:26PM -0700, Stephen
HemmingerQuoted 13 lines · wrote:
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 2 lines · On Fri, 17 Mar 2023 13:19:41 -0700 Tyler Retzlaff…
Replace the use of
rte_atomic.htypes and functions, instead useGCC
Quoted 7 lines
Quoted 7 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
supplied C++11 memory model builtins.
This series covers the libraries and drivers that are built on
Windows.
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
The code has be converted to use the
__atomicbuiltinsbut
Quoted 9 lines
Quoted 9 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
there
are
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
additional during conversion i notice that there may be some
issues
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 1 line
that need to be addressed.
I don't think all these cmpset need to use
SEQ_CST. Especially for the places where it is used a loop, mightbe
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 1 line
more efficient with some of the other memory models.
i agree.
however, i'm not trying to improve the code with this
change,
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
just decouple it from
rte_atomics.hso trying my best toavoid
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
any unnecessary semantic change.
certainly if the maintainers of this code wish to weaken the ordering where appropriate after the change is merged they
can
Quoted 3 lines
Quoted 3 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
do so and
handily
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
this change has enabled them to do so easily allowing them
to
Quoted 8 lines
Quoted 8 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
test
just
Quoted 6 lines
Quoted 6 lines
Quoted 4 lines
Quoted 1 line
their change in isolation.
I agree with the two-step approach, where this first step is a simple
search-and-replacement; but I insist that you add a FIXME or similar note where you have blindly used
SEQ_CST, indicatingthat
Quoted 5 lines
Quoted 5 lines
Quoted 1 line
Quoted 1 line
the memory order
needs to
Quoted 3 lines
Quoted 3 lines
be reviewed and potentially corrected.
i think the maintainers need to take some responsibility, if
they
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
see optimizations they missed when previously writing the code they need to follow up with a patch themselves. i can't do everything for them and marking things i'm not sure about will only lead to me having to churn patch series to remove the
unwanted
Quoted 9 lines
comments later.
Quoted 8 lines
Quoted 3 lines
The previous atomic functions didn't have the "memory order" parameter, so
the maintainers didn't have to think about it - and thus they didn't miss any optimizations when accepting the code.
Quoted 3 lines
I also agree 100 % that it is not your responsibility to consider
or
Quoted 9 lines
Quoted 9 lines
determine which memory order is appropriate!
Quoted 3 lines
But I think you should mark the locations where you are changing from the
old
rte_atomicfunctions (where no memory order optimization was available) to the new functions - to highlight where the option of memory ordering has been introduced and knowingly ignored (by you).Quoted 1 line
Quoted 4 lines
Quoted 2 lines
first, i have to apologize i confused myself about which of the many patch series i have up right now that you were commenting on.
No worries... you are rushing through quite an effort for this, so a
little
Quoted 2 lines
confusion is perfectly understandable. Especially when I'm replying to
an ageing
Quoted 9 lines
email. :-)
Quoted 4 lines
let me ask for clarification in relation to this series.
isn't that every single usage of the
rte_atomicAPIs?Probably, yes.
Quoted 7 lines
Review
Quoted 7 lines
i mean are you
literally asking for the entire patch series to look like the following patch snippet with the expectation that maintainers will come along and clean up/review after this series is merged?(quoted diff) +1
-1-rte_atomic_add32(&o, v); +//FIXME: opportunity for relaxing ordering constraint, pleasereview
Quoted 3 lines
Quoted 1 line
+__atomic_fetch_add(&o, v, order);
Exactly. And something similar for the
rte_atomicXX_tvariableschanged to
Quoted 4 lines
intXX_t, such as the packet counters.Realistically, I don't expect the maintainers to clean them up anytime
soon. The
Quoted 1 line
purpose is to make the FIXMEs stick until someone eventually cleans
them up, so
Quoted 1 line
they are not forgotten as time passes.
Cleaning up the
rte_atomicAPIs is a different effort. There is already lot of effort that has gone into this and there is more effort happening (rte_ringbeing a painful one)
Quoted 8 lines
Quoted 3 lines
Instead of having FIXME, why not just send a separate patch with
SEQ_CST(still a search and replace)? We can leave the tougher ones likerte_ringas they are being worked on.The FIXME makes it possible in the future to differentiate between the instances that still need review and the instances that have been reviewed where
SEQ_CSTwas the correct choice. (Similarly for the choice of type for variables previouslyrte_atomicNN_t.)
Apologies, relooked at the heading of this patch, got confused with other patches.
The changes Arm had done for rte_atomic_ to __atomic_xxx were not direct replacements. The algorithms were studied, relaxed where required, race conditions fixed, performance benchmarked. IMO, we need to go through the same steps here.
I looked at the series, we should just review the patch and make suggested changes. Are we constrained by any deadlines for this work?
I would suggest to drop 1/7. Arm is working on removing the non-C11 algorithm for rte_ring (not sure if we will be successful). I think it is better to explore this approach rather than the changes in patch 1/7.
Quoted 94 lines
Quoted 94 lines
Quoted 4 lines
Quoted 4 lines
Quoted 10 lines
Quoted 3 lines
this would just be a mechanical addition to this series so i can certainly accomodate that, i thought something more complicated was being asked for. if this is all, then sure no problem.
Great.
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
keep in mind i have to touch each of these again when converting to standard so that's a better time to review ~everything in
more
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
detail because when converting to standard that's when suddenly you get a bunch of code generation that is "fallback" to
seq_cstthat isn't
Quoted 3 lines
happening now.
Quoted 2 lines
Quoted 2 lines
I think you should to do it when replacing the
rte_atomicfunctions
Quoted 3 lines
Quoted 3 lines
Quoted 1 line
with the
__atomicfunctions. It will make it easier to see where the memory order was knowingly ignored, and should be reviewed foroptimization.
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
the series that converts to standard needs to be up for review
as
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
soon as possible to maximize available time for feedback before
23.11 so it would be better to get the simpler cut & paste normalizing the code out of the way to unblock that seriessubmission.
Quoted 11 lines
Quoted 11 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Also, in a couple of the drivers, you are using
int64_tfor packetcounters.
Quoted 2 lines
Quoted 2 lines
These cannot be negative and should be
uint64_t. And AFAIK, such counterscan
Quoted 3 lines
Quoted 3 lines
use RELAXED memory order.
i know you don't mean to say i selected the types and rather
that
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
the types that were selected are not quite correct for their
usage.
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Yes; the previous types were also signed, and you didn't change
that.
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
again
on the review that actually adopts std atomics is a better place to make any potential type changes since we are "breaking" theAPI
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
for 23.11 anyway. further, the std atomics series technically changes all the types so it's probably better to make one type change then rather than one now and one later.
i think it would be best to get these validated and merged asap
so
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
we can get to the std atomics review. when that series is up
let's
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
discuss further how i can mark areas of concern, with that
series
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
i expect there will have to be some changes in order to avoid
minor
Quoted 7 lines
regressions.
Quoted 6 lines
Quoted 5 lines
Quoted 2 lines
thanks!
I thought it would be better to catch these details (i.e. memory ordering
and signedness) early on, but I now understand that you planned to
do
Quoted 6 lines
Quoted 6 lines
it in a later step. So I'll let you proceed as you have planned.
Quoted 2 lines
Thanks for all your work on this, Tyler. It is much appreciated!
again, sorry for the confusion the sooner i can get some of these merged the easier it will be for me to manage the final series. i
hope
Quoted 4 lines
Quoted 2 lines
david/thomas can merge the simple normalization patches as soon as 23.03 cycle is complete.
Yes. An early merge would also provide more time for reviewing and
optimizing
Quoted 1 line
the memory order of the most important atomic operations.
Quoted 158 lines · Honnappa Nagarahalli
Quoted 158 lines
-----Original Message-----
From: Morten Brørup <[email protected]>
Sent: Wednesday, March 22, 2023 12:08 PM
To: Honnappa Nagarahalli <[email protected]>; Tyler Retzlaff <[email protected]>
Cc: Stephen Hemminger <[email protected]>; [email protected]; Ruifeng Wang <[email protected]>; [email protected]; nd <[email protected]>; nd <[email protected]>
Subject: RE: [PATCH 0/7] replace rte atomics with GCC builtin atomicsQuoted 145 lines
From: Honnappa Nagarahalli [mailto:[email protected]]
Sent: Wednesday, 22 March 2023 17.40Quoted 10 lines
From: Morten Brørup <[email protected]>
Sent: Wednesday, March 22, 2023 11:14 AMQuoted 7 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Wednesday, 22 March 2023 16.30Quoted 3 lines · Morten Brørup
Quoted 3 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Wednesday, 22 March 2023 15.22Quoted 20 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines · Morten Brørup
Quoted 5 lines
Quoted 5 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Friday, 17 March 2023 22.49On Fri, Mar 17, 2023 at 02:42:26PM -0700, Stephen
HemmingerQuoted 13 lines · wrote:
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 2 lines · On Fri, 17 Mar 2023 13:19:41 -0700 Tyler Retzlaff…
Replace the use of
rte_atomic.htypes and functions, instead useGCC
Quoted 7 lines
Quoted 7 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
supplied C++11 memory model builtins.
This series covers the libraries and drivers that are built on
Windows.
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
The code has be converted to use the
__atomicbuiltinsbut
Quoted 9 lines
Quoted 9 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
there
are
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
additional during conversion i notice that there may be some
issues
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 1 line
that need to be addressed.
I don't think all these cmpset need to use
SEQ_CST. Especially for the places where it is used a loop, mightbe
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 1 line
more efficient with some of the other memory models.
i agree.
however, i'm not trying to improve the code with this
change,
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
just decouple it from
rte_atomics.hso trying my best toavoid
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
any unnecessary semantic change.
certainly if the maintainers of this code wish to weaken the ordering where appropriate after the change is merged they
can
Quoted 3 lines
Quoted 3 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
do so and
handily
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
this change has enabled them to do so easily allowing them
to
Quoted 8 lines
Quoted 8 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
test
just
Quoted 6 lines
Quoted 6 lines
Quoted 4 lines
Quoted 1 line
their change in isolation.
I agree with the two-step approach, where this first step is a simple
search-and-replacement; but I insist that you add a FIXME or similar note where you have blindly used
SEQ_CST, indicatingthat
Quoted 5 lines
Quoted 5 lines
Quoted 1 line
Quoted 1 line
the memory order
needs to
Quoted 3 lines
Quoted 3 lines
be reviewed and potentially corrected.
i think the maintainers need to take some responsibility, if
they
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
see optimizations they missed when previously writing the code they need to follow up with a patch themselves. i can't do everything for them and marking things i'm not sure about will only lead to me having to churn patch series to remove the
unwanted
Quoted 9 lines
comments later.
Quoted 8 lines
Quoted 3 lines
The previous atomic functions didn't have the "memory order" parameter, so
the maintainers didn't have to think about it - and thus they didn't miss any optimizations when accepting the code.
Quoted 3 lines
I also agree 100 % that it is not your responsibility to consider
or
Quoted 9 lines
Quoted 9 lines
determine which memory order is appropriate!
Quoted 3 lines
But I think you should mark the locations where you are changing from the
old
rte_atomicfunctions (where no memory order optimization was available) to the new functions - to highlight where the option of memory ordering has been introduced and knowingly ignored (by you).Quoted 1 line
Quoted 4 lines
Quoted 2 lines
first, i have to apologize i confused myself about which of the many patch series i have up right now that you were commenting on.
No worries... you are rushing through quite an effort for this, so a
little
Quoted 2 lines
confusion is perfectly understandable. Especially when I'm replying to
an ageing
Quoted 9 lines
email. :-)
Quoted 4 lines
let me ask for clarification in relation to this series.
isn't that every single usage of the
rte_atomicAPIs?Probably, yes.
Quoted 7 lines
Review
Quoted 7 lines
i mean are you
literally asking for the entire patch series to look like the following patch snippet with the expectation that maintainers will come along and clean up/review after this series is merged?(quoted diff) +1
-1-rte_atomic_add32(&o, v); +//FIXME: opportunity for relaxing ordering constraint, pleasereview
Quoted 3 lines
Quoted 1 line
+__atomic_fetch_add(&o, v, order);
Exactly. And something similar for the
rte_atomicXX_tvariableschanged to
Quoted 4 lines
intXX_t, such as the packet counters.Realistically, I don't expect the maintainers to clean them up anytime
soon. The
Quoted 1 line
purpose is to make the FIXMEs stick until someone eventually cleans
them up, so
Quoted 1 line
they are not forgotten as time passes.
Cleaning up the
rte_atomicAPIs is a different effort. There is already lot of effort that has gone into this and there is more effort happening (rte_ringbeing a painful one)Instead of having FIXME, why not just send a separate patch with
SEQ_CST(still a search and replace)? We can leave the tougher ones likerte_ringas they are being worked on.
Quoted 5 lines
Quoted 4 lines
The FIXME makes it possible in the future to differentiate between the instances that still need review and the instances that have been reviewed where
SEQ_CSTwas the correct choice. (Similarly for the choice of type for variables previouslyrte_atomicNN_t.)Apologies, relooked at the heading of this patch, got confused with other patches.
yeah, i did the same thing this morning :)
Quoted 4 lines
The changes Arm had done for
rte_atomic_to__atomic_xxxwere not direct replacements. The algorithms were studied, relaxed where required, race conditions fixed, performance benchmarked. IMO, we need to go through the same steps here.I looked at the series, we should just review the patch and make suggested changes. Are we constrained by any deadlines for this work?
i'm going to say yes but i'll qualify. the use of the rte_atomic_xxx APIs drags in extra work when creating a series that performs the actual conversions to the standard atomics.
if i don't decouple ring from rte_atomic_xxx that means i have to go convert all the rte_atomic.h to standard atomics and working around some of the implementation detail to do it is very time consuming. which then has further flow on effects because then i have to go fix every single driver that is still using rte_atomic.h.
incidentally i have a work in progress to decouple everything from rte_atomic.h (including all drivers) but it would really negatively impact getting standard atomics introduced if we had to serialize the introduction behind a total removal of rte_atomic or had to make changes to every consumer of the old rte_atomic APIs.
if we can get by with a comment on the rte_atomic_xxx lines in this series it would be helpful. when we bring the next series for standard atomics i'm not adverse to introducing changes to the ordering in that series if requested so long as i can get the series up 'soon' so there is lots of review time runway for 23.11.
Quoted 2 lines
I would suggest to drop 1/7. Arm is working on removing the non-C11 algorithm for
rte_ring(not sure if we will be successful). I think it is better to explore this approach rather than the changes in patch 1/7.
i think my answer here is timing. i'd rather take the work from arm but if it isn't coming for a while then it becomes a blocker.
we're waiting for the 23.07 start before this series can be merged. how about we re-evaluate where arm is at when the merge window opens. we can then decide to drop 1/7 or not at that time?
ty
PATCH v2 1/7 ring: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffReplace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
--- lib/ring/rte_ring_core.h | 1 - lib/ring/rte_ring_generic_pvt.h | 12 ++++++++---- 2 files changed, 8 insertions(+), 5 deletions(-)
lib/ring/rte_ring_core.h +0 -1
@@ -31,7 +31,6 @@#include <rte_config.h> #include <rte_memory.h> #include <rte_lcore.h> -#include <rte_atomic.h> #include <rte_branch_prediction.h> #include <rte_memzone.h> #include <rte_pause.h>
lib/ring/rte_ring_generic_pvt.h +8 -4
@@ -92,8 +92,10 @@if (is_sp) r->prod.head = *new_head, success =1 ;else - success =rte_atomic32_cmpset (&r->prod.head, - *old_head, *new_head); +// NOTE: review for potential ordering optimization + success =__atomic_compare_exchange_n (&r->prod.head, + old_head, *new_head,0 , + __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST); }while (unlikely (success ==0 ));return n; }
@@ -162,8 +164,10 @@rte_smp_rmb (); success =1 ; }else { - success =rte_atomic32_cmpset (&r->cons.head, *old_head, - *new_head); +// NOTE: review for potential ordering optimization + success =__atomic_compare_exchange_n (&r->cons.head, + old_head, *new_head,0 , + __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST); } }while (unlikely (success ==0 ));return n;
Signature · 1 line
1.8.3.1
v2 cover replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffReplace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
This series covers the libraries and drivers that are built on Windows.
The code has be converted to use the __atomic builtins but there are additional during conversion i notice that there may be some issues that need to be addressed.
I'll comment in the patches where my concerns are so the maintainers may comment.
v2:
* comment code where optimizations may be possible now that memory order can be specified.
* comment code where operations should potentially be atomic so that maintainers can review.
* change a couple of variables labeled as counters to be unsigned.
Tyler Retzlaff (7):
ring: replace rte atomics with GCC builtin atomics
stack: replace rte atomics with GCC builtin atomics dma/idxd: replace rte atomics with GCC builtin atomics net/ice: replace rte atomics with GCC builtin atomics net/ixgbe: replace rte atomics with GCC builtin atomics net/null: replace rte atomics with GCC builtin atomics net/ring: replace rte atomics with GCC builtin atomics
drivers/dma/idxd/idxd_internal.h | 3 +-- drivers/dma/idxd/idxd_pci.c | 8 +++++--- drivers/net/ice/ice_dcf.c | 1 - drivers/net/ice/ice_dcf_ethdev.c | 1 - drivers/net/ice/ice_ethdev.c | 12 ++++++++---- drivers/net/ixgbe/ixgbe_bypass.c | 1 - drivers/net/ixgbe/ixgbe_ethdev.c | 18 ++++++++++++------ drivers/net/ixgbe/ixgbe_ethdev.h | 3 ++- drivers/net/ixgbe/ixgbe_flow.c | 1 - drivers/net/ixgbe/ixgbe_rxtx.c | 1 - drivers/net/null/rte_eth_null.c | 28 ++++++++++++++++++---------- drivers/net/ring/rte_eth_ring.c | 26 ++++++++++++++++---------- lib/ring/rte_ring_core.h | 1 - lib/ring/rte_ring_generic_pvt.h | 12 ++++++++---- lib/stack/rte_stack_lf_generic.h | 16 +++++++++------- 15 files changed, 79 insertions(+), 53 deletions(-)
Signature · 1 line
1.8.3.1
PATCH v2 2/7 stack: replace rte atomics with GCC builtin atomics
Replace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
--- lib/stack/rte_stack_lf_generic.h | 16 +++++++++------- 1 file changed, 9 insertions(+), 7 deletions(-)
lib/stack/rte_stack_lf_generic.h +9 -7
@@ -26,8 +26,8 @@* elements. If the mempool is near-empty to the point that this is a * concern, the user should consider increasing the mempool size. */ -return (unsigned int )rte_atomic64_read ((rte_atomic64_t *) - &s->stack_lf.used.len); +// NOTE: review for potential ordering optimization +return __atomic_load_n (&s->stack_lf.used.len, __ATOMIC_SEQ_CST); }static __rte_always_inline void
@@ -67,8 +67,8 @@1 , __ATOMIC_RELEASE, __ATOMIC_RELAXED); }while (success ==0 ); - -rte_atomic64_add ((rte_atomic64_t *)&list->len, num); +// NOTE: review for potential ordering optimization +__atomic_fetch_add (&list->len, num, __ATOMIC_SEQ_CST); } static __rte_always_inline struct rte_stack_lf_elem *
@@ -82,14 +82,16 @@/* Reserve num elements, if available */ while (1 ) { -uint64_t len =rte_atomic64_read ((rte_atomic64_t *)&list->len); +// NOTE: review for potential ordering optimization +uint64_t len =__atomic_load_n (&list->len, __ATOMIC_SEQ_CST);/* Does the list contain enough elements? */ if (unlikely (len < num))return NULL ; -if (rte_atomic64_cmpset ((volatile uint64_t *)&list->len, - len, len - num)) +// NOTE: review for potential ordering optimization +if (__atomic_compare_exchange_n (&list->len, &len, len - num, +0 , __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))break ; }
Signature · 1 line
1.8.3.1
PATCH v2 3/7 dma/idxd: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffReplace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
--- drivers/dma/idxd/idxd_internal.h | 3 +-- drivers/dma/idxd/idxd_pci.c | 8 +++++--- 2 files changed, 6 insertions(+), 5 deletions(-)
drivers/dma/idxd/idxd_internal.h +1 -2
@@ -7,7 +7,6 @@#include <rte_dmadev_pmd.h> #include <rte_spinlock.h> -#include <rte_atomic.h> #include "idxd_hw_defs.h"
@@ -34,7 +33,7 @@ struct idxd_pci_common {rte_spinlock_t lk;uint8_t wq_cfg_sz; -rte_atomic16_t ref_count; +uint16_t ref_count;volatile struct rte_idxd_bar0 *regs;volatile uint32_t *wq_regs_base;volatile struct rte_idxd_grpcfg *grp_regs;
drivers/dma/idxd/idxd_pci.c +5 -3
@@ -6,7 +6,6 @@#include <rte_devargs.h> #include <rte_dmadev_pmd.h> #include <rte_malloc.h> -#include <rte_atomic.h> #include "idxd_internal.h"
@@ -136,7 +135,9 @@/* if this is the last WQ on the device, disable the device and free * the PCI struct */ - is_last_wq =rte_atomic16_dec_and_test (&idxd->u.pci->ref_count); +// NOTE: review for potential ordering optimization + is_last_wq =__atomic_fetch_sub (&idxd->u.pci->ref_count,1 , + __ATOMIC_SEQ_CST) -1 ==0 ;if (is_last_wq) {/* disable the device */ err_code =idxd_pci_dev_command (idxd, idxd_disable_dev);
@@ -350,7 +351,8 @@free (idxd.u.pci);return ret; } -rte_atomic16_inc (&idxd.u.pci->ref_count); +// NOTE: review for potential ordering optimization +__atomic_fetch_add (&idxd.u.pci->ref_count,1 , __ATOMIC_SEQ_CST); }return 0 ;
Signature · 1 line
1.8.3.1
PATCH v2 4/7 net/ice: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffReplace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
--- drivers/net/ice/ice_dcf.c | 1 - drivers/net/ice/ice_dcf_ethdev.c | 1 - drivers/net/ice/ice_ethdev.c | 12 ++++++++---- 3 files changed, 8 insertions(+), 6 deletions(-)
drivers/net/ice/ice_dcf.c +0 -1
@@ -14,7 +14,6 @@#include <rte_common.h> #include <rte_pci.h> -#include <rte_atomic.h> #include <rte_eal.h> #include <rte_ether.h> #include <ethdev_driver.h>
drivers/net/ice/ice_dcf_ethdev.c +0 -1
@@ -11,7 +11,6 @@#include <rte_interrupts.h> #include <rte_debug.h> #include <rte_pci.h> -#include <rte_atomic.h> #include <rte_eal.h> #include <rte_ether.h> #include <ethdev_pci.h>
drivers/net/ice/ice_ethdev.c +8 -4
@@ -3927,8 +3927,10 @@ static int ice_init_rss(struct ice_pf *pf)struct rte_eth_link *dst = link;struct rte_eth_link *src = &dev->data->dev_link; -if (rte_atomic64_cmpset ((uint64_t *)dst, *(uint64_t *)dst, - *(uint64_t *)src) ==0 ) +// NOTE: review for potential ordering optimization +if (!__atomic_compare_exchange_n ((uint64_t *)dst, + (uint64_t *)dst, *(uint64_t *)src,0 , + __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))return -1 ;return 0 ;
@@ -3941,8 +3943,10 @@ static int ice_init_rss(struct ice_pf *pf)struct rte_eth_link *dst = &dev->data->dev_link;struct rte_eth_link *src = link; -if (rte_atomic64_cmpset ((uint64_t *)dst, *(uint64_t *)dst, - *(uint64_t *)src) ==0 ) +// NOTE: review for potential ordering optimization +if (!__atomic_compare_exchange_n ((uint64_t *)dst, + (uint64_t *)dst, *(uint64_t *)src,0 , + __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))return -1 ;return 0 ;
Signature · 1 line
1.8.3.1
PATCH v2 5/7 net/ixgbe: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffReplace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
--- drivers/net/ixgbe/ixgbe_bypass.c | 1 - drivers/net/ixgbe/ixgbe_ethdev.c | 18 ++++++++++++------ drivers/net/ixgbe/ixgbe_ethdev.h | 3 ++- drivers/net/ixgbe/ixgbe_flow.c | 1 - drivers/net/ixgbe/ixgbe_rxtx.c | 1 - 5 files changed, 14 insertions(+), 10 deletions(-)
drivers/net/ixgbe/ixgbe_bypass.c +0 -1
@@ -3,7 +3,6 @@*/ #include <time.h> -#include <rte_atomic.h> #include <ethdev_driver.h> #include "ixgbe_ethdev.h" #include "ixgbe_bypass_api.h"
drivers/net/ixgbe/ixgbe_ethdev.c +12 -6
@@ -1127,7 +1127,8 @@ struct rte_ixgbe_xstats_name_off {return 0 ; } -rte_atomic32_clear (&ad->link_thread_running); +// NOTE: review for potential ordering optimization +__atomic_clear (&ad->link_thread_running, __ATOMIC_SEQ_CST);ixgbe_parse_devargs (eth_dev->data->dev_private, pci_dev->device.devargs);rte_eth_copy_pci_info (eth_dev, pci_dev);
@@ -1625,7 +1626,8 @@ static int ixgbe_l2_tn_filter_init(struct rte_eth_dev *eth_dev)return 0 ; } -rte_atomic32_clear (&ad->link_thread_running); +// NOTE: review for potential ordering optimization +__atomic_clear (&ad->link_thread_running, __ATOMIC_SEQ_CST);ixgbevf_parse_devargs (eth_dev->data->dev_private, pci_dev->device.devargs);
@@ -4186,7 +4188,8 @@ static int ixgbevf_dev_xstats_get_names(__rte_unused struct rte_eth_dev *dev,struct ixgbe_adapter *ad = dev->data->dev_private;uint32_t timeout = timeout_ms ? timeout_ms :WARNING_TIMEOUT ; -while (rte_atomic32_read (&ad->link_thread_running)) { +// NOTE: review for potential ordering optimization +while (__atomic_load_n (&ad->link_thread_running, __ATOMIC_SEQ_CST)) {msec_delay (1 ); timeout--;
@@ -4222,7 +4225,8 @@ static int ixgbevf_dev_xstats_get_names(__rte_unused struct rte_eth_dev *dev,ixgbe_setup_link (hw, speed, true); intr->flags &= ~IXGBE_FLAG_NEED_LINK_CONFIG ; -rte_atomic32_clear (&ad->link_thread_running); +// NOTE: review for potential ordering optimization +__atomic_clear (&ad->link_thread_running, __ATOMIC_SEQ_CST);return NULL ; }
@@ -4317,7 +4321,8 @@ static int ixgbevf_dev_xstats_get_names(__rte_unused struct rte_eth_dev *dev,if (link_up ==0 ) {if (ixgbe_get_media_type (hw) == ixgbe_media_type_fiber) {ixgbe_dev_wait_setup_link_complete (dev,0 ); -if (rte_atomic32_test_and_set (&ad->link_thread_running)) { +// NOTE: review for potential ordering optimization +if (__atomic_test_and_set (&ad->link_thread_running, __ATOMIC_SEQ_CST)) {/* To avoid race condition between threads, set * the IXGBE_FLAG_NEED_LINK_CONFIG flag only * when there is no link thread running.
@@ -4330,7 +4335,8 @@ static int ixgbevf_dev_xstats_get_names(__rte_unused struct rte_eth_dev *dev, dev) < 0) {PMD_DRV_LOG (ERR ,"Create link thread failed!" ); -rte_atomic32_clear (&ad->link_thread_running); +// NOTE: review for potential ordering optimization +__atomic_clear (&ad->link_thread_running, __ATOMIC_SEQ_CST); } }else { PMD_DRV_LOG(ERR,
drivers/net/ixgbe/ixgbe_ethdev.h +2 -1
@@ -6,6 +6,7 @@#define _IXGBE_ETHDEV_H_#include <stdint.h> +#include <stdbool.h> #include <sys/queue.h> #include "base/ixgbe_type.h"
@@ -510,7 +511,7 @@ struct ixgbe_adapter {*/ uint8_t pflink_fullchk;uint8_t mac_ctrl_frame_fwd; -rte_atomic32_t link_thread_running; +bool link_thread_running;pthread_t link_thread_tid; };
drivers/net/ixgbe/ixgbe_flow.c +0 -1
@@ -18,7 +18,6 @@#include <rte_log.h> #include <rte_debug.h> #include <rte_pci.h> -#include <rte_atomic.h> #include <rte_branch_prediction.h> #include <rte_memory.h> #include <rte_eal.h>
drivers/net/ixgbe/ixgbe_rxtx.c +0 -1
@@ -27,7 +27,6 @@#include <rte_eal.h> #include <rte_per_lcore.h> #include <rte_lcore.h> -#include <rte_atomic.h> #include <rte_branch_prediction.h> #include <rte_mempool.h> #include <rte_malloc.h>
Signature · 1 line
1.8.3.1
PATCH v2 7/7 net/ring: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffReplace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
--- drivers/net/ring/rte_eth_ring.c | 26 ++++++++++++++++---------- 1 file changed, 16 insertions(+), 10 deletions(-)
drivers/net/ring/rte_eth_ring.c +16 -10
@@ -44,8 +44,8 @@ enum dev_action {struct ring_queue {struct rte_ring *rng; -rte_atomic64_t rx_pkts; -rte_atomic64_t tx_pkts; +uint64_t rx_pkts; +uint64_t tx_pkts; };struct pmd_internals {
@@ -80,9 +80,10 @@ struct pmd_internals {const uint16_t nb_rx = (uint16_t )rte_ring_dequeue_burst (r->rng, ptrs, nb_bufs,NULL );if (r->rng->flags &RING_F_SC_DEQ ) - r->rx_pkts.cnt += nb_rx; + r->rx_pkts += nb_rx;else -rte_atomic64_add (&(r->rx_pkts), nb_rx); +// NOTE: review for potential ordering optimization +__atomic_fetch_add (&r->rx_pkts, nb_rx, __ATOMIC_SEQ_CST);return nb_rx; }
@@ -94,9 +95,10 @@ struct pmd_internals {const uint16_t nb_tx = (uint16_t )rte_ring_enqueue_burst (r->rng, ptrs, nb_bufs,NULL );if (r->rng->flags &RING_F_SP_ENQ ) - r->tx_pkts.cnt += nb_tx; + r->tx_pkts += nb_tx;else -rte_atomic64_add (&(r->tx_pkts), nb_tx); +// NOTE: review for potential ordering optimization +__atomic_fetch_add (&r->tx_pkts, nb_tx, __ATOMIC_SEQ_CST);return nb_tx; }
@@ -184,13 +186,15 @@ struct pmd_internals {for (i =0 ; i <RTE_ETHDEV_QUEUE_STAT_CNTRS && i < dev->data->nb_rx_queues; i++) { - stats->q_ipackets[i] = internal->rx_ring_queues[i].rx_pkts.cnt; +// NOTE: review for atomic access + stats->q_ipackets[i] = internal->rx_ring_queues[i].rx_pkts; rx_total += stats->q_ipackets[i]; }for (i =0 ; i <RTE_ETHDEV_QUEUE_STAT_CNTRS && i < dev->data->nb_tx_queues; i++) { - stats->q_opackets[i] = internal->tx_ring_queues[i].tx_pkts.cnt; +// NOTE: review for atomic access + stats->q_opackets[i] = internal->tx_ring_queues[i].tx_pkts; tx_total += stats->q_opackets[i]; }
@@ -207,9 +211,11 @@ struct pmd_internals {struct pmd_internals *internal = dev->data->dev_private;for (i =0 ; i < dev->data->nb_rx_queues; i++) - internal->rx_ring_queues[i].rx_pkts.cnt =0 ; +// NOTE: review for atomic access + internal->rx_ring_queues[i].rx_pkts =0 ;for (i =0 ; i < dev->data->nb_tx_queues; i++) - internal->tx_ring_queues[i].tx_pkts.cnt =0 ; +// NOTE: review for atomic access + internal->tx_ring_queues[i].tx_pkts =0 ;return 0 ; }
Signature · 1 line
1.8.3.1
PATCH v2 6/7 net/null: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffReplace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
--- drivers/net/null/rte_eth_null.c | 28 ++++++++++++++++++---------- 1 file changed, 18 insertions(+), 10 deletions(-)
drivers/net/null/rte_eth_null.c +18 -10
@@ -37,8 +37,8 @@ struct null_queue {struct rte_mempool *mb_pool;struct rte_mbuf *dummy_packet; -rte_atomic64_t rx_pkts; -rte_atomic64_t tx_pkts; +uint64_t rx_pkts; +uint64_t tx_pkts; };struct pmd_options {
@@ -101,7 +101,8 @@ struct pmd_internals { bufs[i]->port = h->internals->port_id; } -rte_atomic64_add (&(h->rx_pkts), i); +// NOTE: review for potential ordering optimization +__atomic_fetch_add (&h->rx_pkts, i, __ATOMIC_SEQ_CST);return i; }
@@ -128,7 +129,8 @@ struct pmd_internals { bufs[i]->port = h->internals->port_id; } -rte_atomic64_add (&(h->rx_pkts), i); +// NOTE: review for potential ordering optimization +__atomic_fetch_add (&h->rx_pkts, i, __ATOMIC_SEQ_CST);return i; }
@@ -152,7 +154,8 @@ struct pmd_internals {for (i =0 ; i < nb_bufs; i++)rte_pktmbuf_free (bufs[i]); -rte_atomic64_add (&(h->tx_pkts), i); +// NOTE: review for potential ordering optimization +__atomic_fetch_add (&h->tx_pkts, i, __ATOMIC_SEQ_CST);return i; }
@@ -174,7 +177,8 @@ struct pmd_internals {rte_pktmbuf_free (bufs[i]); } -rte_atomic64_add (&(h->tx_pkts), i); +// NOTE: review for potential ordering optimization +__atomic_fetch_add (&h->tx_pkts, i, __ATOMIC_SEQ_CST);return i; }
@@ -316,8 +320,9 @@ struct pmd_internals {RTE_MIN (dev->data->nb_rx_queues, RTE_DIM(internal->rx_null_queues)));for (i =0 ; i < num_stats; i++) { +// NOTE: review for atomic access igb_stats->q_ipackets[i] = - internal->rx_null_queues[i].rx_pkts.cnt; + internal->rx_null_queues[i].rx_pkts; rx_total += igb_stats->q_ipackets[i]; }
@@ -325,8 +330,9 @@ struct pmd_internals {RTE_MIN (dev->data->nb_tx_queues, RTE_DIM(internal->tx_null_queues)));for (i =0 ; i < num_stats; i++) { +// NOTE: review for atomic access igb_stats->q_opackets[i] = - internal->tx_null_queues[i].tx_pkts.cnt; + internal->tx_null_queues[i].tx_pkts; tx_total += igb_stats->q_opackets[i]; }
@@ -347,9 +353,11 @@ struct pmd_internals { internal = dev->data->dev_private;for (i =0 ; i <RTE_DIM (internal->rx_null_queues); i++) - internal->rx_null_queues[i].rx_pkts.cnt =0 ; +// NOTE: review for atomic access + internal->rx_null_queues[i].rx_pkts =0 ;for (i =0 ; i <RTE_DIM (internal->tx_null_queues); i++) - internal->tx_null_queues[i].tx_pkts.cnt =0 ; +// NOTE: review for atomic access + internal->tx_null_queues[i].tx_pkts =0 ;return 0 ; }
Signature · 1 line
1.8.3.1
v3 cover replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffReplace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
This series covers the libraries and drivers that are built on Windows.
The code has be converted to use the __atomic builtins but there are additional during conversion i notice that there may be some issues that need to be addressed.
I'll comment in the patches where my concerns are so the maintainers may comment.
v3:
* style, don't use c99 comments
v2:
* comment code where optimizations may be possible now that memory order can be specified.
* comment code where operations should potentially be atomic so that maintainers can review.
* change a couple of variables labeled as counters to be unsigned.
Tyler Retzlaff (7):
ring: replace rte atomics with GCC builtin atomics
stack: replace rte atomics with GCC builtin atomics dma/idxd: replace rte atomics with GCC builtin atomics net/ice: replace rte atomics with GCC builtin atomics net/ixgbe: replace rte atomics with GCC builtin atomics net/null: replace rte atomics with GCC builtin atomics net/ring: replace rte atomics with GCC builtin atomics
drivers/dma/idxd/idxd_internal.h | 3 +-- drivers/dma/idxd/idxd_pci.c | 8 +++++--- drivers/net/ice/ice_dcf.c | 1 - drivers/net/ice/ice_dcf_ethdev.c | 1 - drivers/net/ice/ice_ethdev.c | 12 ++++++++---- drivers/net/ixgbe/ixgbe_bypass.c | 1 - drivers/net/ixgbe/ixgbe_ethdev.c | 18 ++++++++++++------ drivers/net/ixgbe/ixgbe_ethdev.h | 3 ++- drivers/net/ixgbe/ixgbe_flow.c | 1 - drivers/net/ixgbe/ixgbe_rxtx.c | 1 - drivers/net/null/rte_eth_null.c | 28 ++++++++++++++++++---------- drivers/net/ring/rte_eth_ring.c | 26 ++++++++++++++++---------- lib/ring/rte_ring_core.h | 1 - lib/ring/rte_ring_generic_pvt.h | 12 ++++++++---- lib/stack/rte_stack_lf_generic.h | 16 +++++++++------- 15 files changed, 79 insertions(+), 53 deletions(-)
Signature · 1 line
1.8.3.1
PATCH v3 2/7 stack: replace rte atomics with GCC builtin atomics
Replace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
--- lib/stack/rte_stack_lf_generic.h | 16 +++++++++------- 1 file changed, 9 insertions(+), 7 deletions(-)
lib/stack/rte_stack_lf_generic.h +9 -7
@@ -26,8 +26,8 @@* elements. If the mempool is near-empty to the point that this is a * concern, the user should consider increasing the mempool size. */ -return (unsigned int )rte_atomic64_read ((rte_atomic64_t *) - &s->stack_lf.used.len); +/* NOTE: review for potential ordering optimization */ +return __atomic_load_n (&s->stack_lf.used.len, __ATOMIC_SEQ_CST); }static __rte_always_inline void
@@ -67,8 +67,8 @@1 , __ATOMIC_RELEASE, __ATOMIC_RELAXED); }while (success ==0 ); - -rte_atomic64_add ((rte_atomic64_t *)&list->len, num); +/* NOTE: review for potential ordering optimization */ +__atomic_fetch_add (&list->len, num, __ATOMIC_SEQ_CST); } static __rte_always_inline struct rte_stack_lf_elem *
@@ -82,14 +82,16 @@/* Reserve num elements, if available */ while (1 ) { -uint64_t len =rte_atomic64_read ((rte_atomic64_t *)&list->len); +/* NOTE: review for potential ordering optimization */ +uint64_t len =__atomic_load_n (&list->len, __ATOMIC_SEQ_CST);/* Does the list contain enough elements? */ if (unlikely (len < num))return NULL ; -if (rte_atomic64_cmpset ((volatile uint64_t *)&list->len, - len, len - num)) +/* NOTE: review for potential ordering optimization */ +if (__atomic_compare_exchange_n (&list->len, &len, len - num, +0 , __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))break ; }
Signature · 1 line
1.8.3.1
PATCH v3 3/7 dma/idxd: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffReplace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
--- drivers/dma/idxd/idxd_internal.h | 3 +-- drivers/dma/idxd/idxd_pci.c | 8 +++++--- 2 files changed, 6 insertions(+), 5 deletions(-)
drivers/dma/idxd/idxd_internal.h +1 -2
@@ -7,7 +7,6 @@#include <rte_dmadev_pmd.h> #include <rte_spinlock.h> -#include <rte_atomic.h> #include "idxd_hw_defs.h"
@@ -34,7 +33,7 @@ struct idxd_pci_common {rte_spinlock_t lk;uint8_t wq_cfg_sz; -rte_atomic16_t ref_count; +uint16_t ref_count;volatile struct rte_idxd_bar0 *regs;volatile uint32_t *wq_regs_base;volatile struct rte_idxd_grpcfg *grp_regs;
drivers/dma/idxd/idxd_pci.c +5 -3
@@ -6,7 +6,6 @@#include <rte_devargs.h> #include <rte_dmadev_pmd.h> #include <rte_malloc.h> -#include <rte_atomic.h> #include "idxd_internal.h"
@@ -136,7 +135,9 @@/* if this is the last WQ on the device, disable the device and free * the PCI struct */ - is_last_wq =rte_atomic16_dec_and_test (&idxd->u.pci->ref_count); +/* NOTE: review for potential ordering optimization */ + is_last_wq =__atomic_fetch_sub (&idxd->u.pci->ref_count,1 , + __ATOMIC_SEQ_CST) -1 ==0 ;if (is_last_wq) {/* disable the device */ err_code =idxd_pci_dev_command (idxd, idxd_disable_dev);
@@ -350,7 +351,8 @@free (idxd.u.pci);return ret; } -rte_atomic16_inc (&idxd.u.pci->ref_count); +/* NOTE: review for potential ordering optimization */ +__atomic_fetch_add (&idxd.u.pci->ref_count,1 , __ATOMIC_SEQ_CST); }return 0 ;
Signature · 1 line
1.8.3.1
PATCH v3 4/7 net/ice: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffReplace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
--- drivers/net/ice/ice_dcf.c | 1 - drivers/net/ice/ice_dcf_ethdev.c | 1 - drivers/net/ice/ice_ethdev.c | 12 ++++++++---- 3 files changed, 8 insertions(+), 6 deletions(-)
drivers/net/ice/ice_dcf.c +0 -1
@@ -14,7 +14,6 @@#include <rte_common.h> #include <rte_pci.h> -#include <rte_atomic.h> #include <rte_eal.h> #include <rte_ether.h> #include <ethdev_driver.h>
drivers/net/ice/ice_dcf_ethdev.c +0 -1
@@ -11,7 +11,6 @@#include <rte_interrupts.h> #include <rte_debug.h> #include <rte_pci.h> -#include <rte_atomic.h> #include <rte_eal.h> #include <rte_ether.h> #include <ethdev_pci.h>
drivers/net/ice/ice_ethdev.c +8 -4
@@ -3927,8 +3927,10 @@ static int ice_init_rss(struct ice_pf *pf)struct rte_eth_link *dst = link;struct rte_eth_link *src = &dev->data->dev_link; -if (rte_atomic64_cmpset ((uint64_t *)dst, *(uint64_t *)dst, - *(uint64_t *)src) ==0 ) +/* NOTE: review for potential ordering optimization */ +if (!__atomic_compare_exchange_n ((uint64_t *)dst, + (uint64_t *)dst, *(uint64_t *)src,0 , + __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))return -1 ;return 0 ;
@@ -3941,8 +3943,10 @@ static int ice_init_rss(struct ice_pf *pf)struct rte_eth_link *dst = &dev->data->dev_link;struct rte_eth_link *src = link; -if (rte_atomic64_cmpset ((uint64_t *)dst, *(uint64_t *)dst, - *(uint64_t *)src) ==0 ) +/* NOTE: review for potential ordering optimization */ +if (!__atomic_compare_exchange_n ((uint64_t *)dst, + (uint64_t *)dst, *(uint64_t *)src,0 , + __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))return -1 ;return 0 ;
Signature · 1 line
1.8.3.1
PATCH v3 1/7 ring: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffReplace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
--- lib/ring/rte_ring_core.h | 1 - lib/ring/rte_ring_generic_pvt.h | 12 ++++++++---- 2 files changed, 8 insertions(+), 5 deletions(-)
lib/ring/rte_ring_core.h +0 -1
@@ -31,7 +31,6 @@#include <rte_config.h> #include <rte_memory.h> #include <rte_lcore.h> -#include <rte_atomic.h> #include <rte_branch_prediction.h> #include <rte_memzone.h> #include <rte_pause.h>
lib/ring/rte_ring_generic_pvt.h +8 -4
@@ -92,8 +92,10 @@if (is_sp) r->prod.head = *new_head, success =1 ;else - success =rte_atomic32_cmpset (&r->prod.head, - *old_head, *new_head); +/* NOTE: review for potential ordering optimization */ + success =__atomic_compare_exchange_n (&r->prod.head, + old_head, *new_head,0 , + __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST); }while (unlikely (success ==0 ));return n; }
@@ -162,8 +164,10 @@rte_smp_rmb (); success =1 ; }else { - success =rte_atomic32_cmpset (&r->cons.head, *old_head, - *new_head); +/* NOTE: review for potential ordering optimization */ + success =__atomic_compare_exchange_n (&r->cons.head, + old_head, *new_head,0 , + __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST); } }while (unlikely (success ==0 ));return n;
Signature · 1 line
1.8.3.1
PATCH v3 7/7 net/ring: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffReplace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
--- drivers/net/ring/rte_eth_ring.c | 26 ++++++++++++++++---------- 1 file changed, 16 insertions(+), 10 deletions(-)
drivers/net/ring/rte_eth_ring.c +16 -10
@@ -44,8 +44,8 @@ enum dev_action {struct ring_queue {struct rte_ring *rng; -rte_atomic64_t rx_pkts; -rte_atomic64_t tx_pkts; +uint64_t rx_pkts; +uint64_t tx_pkts; };struct pmd_internals {
@@ -80,9 +80,10 @@ struct pmd_internals {const uint16_t nb_rx = (uint16_t )rte_ring_dequeue_burst (r->rng, ptrs, nb_bufs,NULL );if (r->rng->flags &RING_F_SC_DEQ ) - r->rx_pkts.cnt += nb_rx; + r->rx_pkts += nb_rx;else -rte_atomic64_add (&(r->rx_pkts), nb_rx); +/* NOTE: review for potential ordering optimization */ +__atomic_fetch_add (&r->rx_pkts, nb_rx, __ATOMIC_SEQ_CST);return nb_rx; }
@@ -94,9 +95,10 @@ struct pmd_internals {const uint16_t nb_tx = (uint16_t )rte_ring_enqueue_burst (r->rng, ptrs, nb_bufs,NULL );if (r->rng->flags &RING_F_SP_ENQ ) - r->tx_pkts.cnt += nb_tx; + r->tx_pkts += nb_tx;else -rte_atomic64_add (&(r->tx_pkts), nb_tx); +/* NOTE: review for potential ordering optimization */ +__atomic_fetch_add (&r->tx_pkts, nb_tx, __ATOMIC_SEQ_CST);return nb_tx; }
@@ -184,13 +186,15 @@ struct pmd_internals {for (i =0 ; i <RTE_ETHDEV_QUEUE_STAT_CNTRS && i < dev->data->nb_rx_queues; i++) { - stats->q_ipackets[i] = internal->rx_ring_queues[i].rx_pkts.cnt; +/* NOTE: review for atomic access */ + stats->q_ipackets[i] = internal->rx_ring_queues[i].rx_pkts; rx_total += stats->q_ipackets[i]; }for (i =0 ; i <RTE_ETHDEV_QUEUE_STAT_CNTRS && i < dev->data->nb_tx_queues; i++) { - stats->q_opackets[i] = internal->tx_ring_queues[i].tx_pkts.cnt; +/* NOTE: review for atomic access */ + stats->q_opackets[i] = internal->tx_ring_queues[i].tx_pkts; tx_total += stats->q_opackets[i]; }
@@ -207,9 +211,11 @@ struct pmd_internals {struct pmd_internals *internal = dev->data->dev_private;for (i =0 ; i < dev->data->nb_rx_queues; i++) - internal->rx_ring_queues[i].rx_pkts.cnt =0 ; +/* NOTE: review for atomic access */ + internal->rx_ring_queues[i].rx_pkts =0 ;for (i =0 ; i < dev->data->nb_tx_queues; i++) - internal->tx_ring_queues[i].tx_pkts.cnt =0 ; +/* NOTE: review for atomic access */ + internal->tx_ring_queues[i].tx_pkts =0 ;return 0 ; }
Signature · 1 line
1.8.3.1
PATCH v3 6/7 net/null: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffReplace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
--- drivers/net/null/rte_eth_null.c | 28 ++++++++++++++++++---------- 1 file changed, 18 insertions(+), 10 deletions(-)
drivers/net/null/rte_eth_null.c +18 -10
@@ -37,8 +37,8 @@ struct null_queue {struct rte_mempool *mb_pool;struct rte_mbuf *dummy_packet; -rte_atomic64_t rx_pkts; -rte_atomic64_t tx_pkts; +uint64_t rx_pkts; +uint64_t tx_pkts; };struct pmd_options {
@@ -101,7 +101,8 @@ struct pmd_internals { bufs[i]->port = h->internals->port_id; } -rte_atomic64_add (&(h->rx_pkts), i); +/* NOTE: review for potential ordering optimization */ +__atomic_fetch_add (&h->rx_pkts, i, __ATOMIC_SEQ_CST);return i; }
@@ -128,7 +129,8 @@ struct pmd_internals { bufs[i]->port = h->internals->port_id; } -rte_atomic64_add (&(h->rx_pkts), i); +/* NOTE: review for potential ordering optimization */ +__atomic_fetch_add (&h->rx_pkts, i, __ATOMIC_SEQ_CST);return i; }
@@ -152,7 +154,8 @@ struct pmd_internals {for (i =0 ; i < nb_bufs; i++)rte_pktmbuf_free (bufs[i]); -rte_atomic64_add (&(h->tx_pkts), i); +/* NOTE: review for potential ordering optimization */ +__atomic_fetch_add (&h->tx_pkts, i, __ATOMIC_SEQ_CST);return i; }
@@ -174,7 +177,8 @@ struct pmd_internals {rte_pktmbuf_free (bufs[i]); } -rte_atomic64_add (&(h->tx_pkts), i); +/* NOTE: review for potential ordering optimization */ +__atomic_fetch_add (&h->tx_pkts, i, __ATOMIC_SEQ_CST);return i; }
@@ -316,8 +320,9 @@ struct pmd_internals {RTE_MIN (dev->data->nb_rx_queues, RTE_DIM(internal->rx_null_queues)));for (i =0 ; i < num_stats; i++) { +/* NOTE: review for atomic access */ igb_stats->q_ipackets[i] = - internal->rx_null_queues[i].rx_pkts.cnt; + internal->rx_null_queues[i].rx_pkts; rx_total += igb_stats->q_ipackets[i]; }
@@ -325,8 +330,9 @@ struct pmd_internals {RTE_MIN (dev->data->nb_tx_queues, RTE_DIM(internal->tx_null_queues)));for (i =0 ; i < num_stats; i++) { +/* NOTE: review for atomic access */ igb_stats->q_opackets[i] = - internal->tx_null_queues[i].tx_pkts.cnt; + internal->tx_null_queues[i].tx_pkts; tx_total += igb_stats->q_opackets[i]; }
@@ -347,9 +353,11 @@ struct pmd_internals { internal = dev->data->dev_private;for (i =0 ; i <RTE_DIM (internal->rx_null_queues); i++) - internal->rx_null_queues[i].rx_pkts.cnt =0 ; +/* NOTE: review for atomic access */ + internal->rx_null_queues[i].rx_pkts =0 ;for (i =0 ; i <RTE_DIM (internal->tx_null_queues); i++) - internal->tx_null_queues[i].tx_pkts.cnt =0 ; +/* NOTE: review for atomic access */ + internal->tx_null_queues[i].tx_pkts =0 ;return 0 ; }
Signature · 1 line
1.8.3.1
PATCH v3 5/7 net/ixgbe: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffReplace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
--- drivers/net/ixgbe/ixgbe_bypass.c | 1 - drivers/net/ixgbe/ixgbe_ethdev.c | 18 ++++++++++++------ drivers/net/ixgbe/ixgbe_ethdev.h | 3 ++- drivers/net/ixgbe/ixgbe_flow.c | 1 - drivers/net/ixgbe/ixgbe_rxtx.c | 1 - 5 files changed, 14 insertions(+), 10 deletions(-)
drivers/net/ixgbe/ixgbe_bypass.c +0 -1
@@ -3,7 +3,6 @@*/ #include <time.h> -#include <rte_atomic.h> #include <ethdev_driver.h> #include "ixgbe_ethdev.h" #include "ixgbe_bypass_api.h"
drivers/net/ixgbe/ixgbe_ethdev.c +12 -6
@@ -1127,7 +1127,8 @@ struct rte_ixgbe_xstats_name_off {return 0 ; } -rte_atomic32_clear (&ad->link_thread_running); +/* NOTE: review for potential ordering optimization */ +__atomic_clear (&ad->link_thread_running, __ATOMIC_SEQ_CST);ixgbe_parse_devargs (eth_dev->data->dev_private, pci_dev->device.devargs);rte_eth_copy_pci_info (eth_dev, pci_dev);
@@ -1625,7 +1626,8 @@ static int ixgbe_l2_tn_filter_init(struct rte_eth_dev *eth_dev)return 0 ; } -rte_atomic32_clear (&ad->link_thread_running); +/* NOTE: review for potential ordering optimization */ +__atomic_clear (&ad->link_thread_running, __ATOMIC_SEQ_CST);ixgbevf_parse_devargs (eth_dev->data->dev_private, pci_dev->device.devargs);
@@ -4186,7 +4188,8 @@ static int ixgbevf_dev_xstats_get_names(__rte_unused struct rte_eth_dev *dev,struct ixgbe_adapter *ad = dev->data->dev_private;uint32_t timeout = timeout_ms ? timeout_ms :WARNING_TIMEOUT ; -while (rte_atomic32_read (&ad->link_thread_running)) { +/* NOTE: review for potential ordering optimization */ +while (__atomic_load_n (&ad->link_thread_running, __ATOMIC_SEQ_CST)) {msec_delay (1 ); timeout--;
@@ -4222,7 +4225,8 @@ static int ixgbevf_dev_xstats_get_names(__rte_unused struct rte_eth_dev *dev,ixgbe_setup_link (hw, speed, true); intr->flags &= ~IXGBE_FLAG_NEED_LINK_CONFIG ; -rte_atomic32_clear (&ad->link_thread_running); +/* NOTE: review for potential ordering optimization */ +__atomic_clear (&ad->link_thread_running, __ATOMIC_SEQ_CST);return NULL ; }
@@ -4317,7 +4321,8 @@ static int ixgbevf_dev_xstats_get_names(__rte_unused struct rte_eth_dev *dev,if (link_up ==0 ) {if (ixgbe_get_media_type (hw) == ixgbe_media_type_fiber) {ixgbe_dev_wait_setup_link_complete (dev,0 ); -if (rte_atomic32_test_and_set (&ad->link_thread_running)) { +/* NOTE: review for potential ordering optimization */ +if (__atomic_test_and_set (&ad->link_thread_running, __ATOMIC_SEQ_CST)) {/* To avoid race condition between threads, set * the IXGBE_FLAG_NEED_LINK_CONFIG flag only * when there is no link thread running.
@@ -4330,7 +4335,8 @@ static int ixgbevf_dev_xstats_get_names(__rte_unused struct rte_eth_dev *dev, dev) < 0) {PMD_DRV_LOG (ERR ,"Create link thread failed!" ); -rte_atomic32_clear (&ad->link_thread_running); +/* NOTE: review for potential ordering optimization */ +__atomic_clear (&ad->link_thread_running, __ATOMIC_SEQ_CST); } }else { PMD_DRV_LOG(ERR,
drivers/net/ixgbe/ixgbe_ethdev.h +2 -1
@@ -6,6 +6,7 @@#define _IXGBE_ETHDEV_H_#include <stdint.h> +#include <stdbool.h> #include <sys/queue.h> #include "base/ixgbe_type.h"
@@ -510,7 +511,7 @@ struct ixgbe_adapter {*/ uint8_t pflink_fullchk;uint8_t mac_ctrl_frame_fwd; -rte_atomic32_t link_thread_running; +bool link_thread_running;pthread_t link_thread_tid; };
drivers/net/ixgbe/ixgbe_flow.c +0 -1
@@ -18,7 +18,6 @@#include <rte_log.h> #include <rte_debug.h> #include <rte_pci.h> -#include <rte_atomic.h> #include <rte_branch_prediction.h> #include <rte_memory.h> #include <rte_eal.h>
drivers/net/ixgbe/ixgbe_rxtx.c +0 -1
@@ -27,7 +27,6 @@#include <rte_eal.h> #include <rte_per_lcore.h> #include <rte_lcore.h> -#include <rte_atomic.h> #include <rte_branch_prediction.h> #include <rte_mempool.h> #include <rte_malloc.h>
Signature · 1 line
1.8.3.1
Quoted 15 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Thursday, 23 March 2023 23.35Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.This series covers the libraries and drivers that are built on Windows.
The code has be converted to use the
__atomicbuiltins but there are additional during conversion i notice that there may be some issues that need to be addressed.I'll comment in the patches where my concerns are so the maintainers may comment.
Quoted 7 lines
v2:
* comment code where optimizations may be possible now that memory order can be specified.
* comment code where operations should potentially be atomic so that maintainers can review.
* change a couple of variables labeled as counters to be unsigned.
All good.
Quoted 18 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Thursday, 23 March 2023 23.54Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.This series covers the libraries and drivers that are built on Windows.
The code has be converted to use the
__atomicbuiltins but there are additional during conversion i notice that there may be some issues that need to be addressed.I'll comment in the patches where my concerns are so the maintainers may comment.
v3:
* style, don't use c99 comments
Quoted 7 lines
v2:
* comment code where optimizations may be possible now that memory order can be specified.
* comment code where operations should potentially be atomic so that maintainers can review.
* change a couple of variables labeled as counters to be unsigned.
I didn't see the v3 when ack'ing the v2, so in case v2 is quickly skipped by maintainers...
Quoted 26 lines · Morten Brørup
Quoted 18 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Thursday, 23 March 2023 23.54Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.This series covers the libraries and drivers that are built on Windows.
The code has be converted to use the
__atomicbuiltins but there are additional during conversion i notice that there may be some issues that need to be addressed.I'll comment in the patches where my concerns are so the maintainers may comment.
v3:
* style, don't use c99 commentsQuoted 7 lines
v2:
* comment code where optimizations may be possible now that memory order can be specified.
* comment code where operations should potentially be atomic so that maintainers can review.
* change a couple of variables labeled as counters to be unsigned.
Quoted 1 line
I didn't see the v3 when ack'ing the v2, so in case v2 is quickly skipped by maintainers...
yeah, my fault. i hammed up the comment style used and needed to quickly submit v3 to satisfy checkpatches.
thanks!
Quoted 2 lines
Series-acked-by Morten Brørup <[email protected]>
Quoted 195 lines · Tyler Retzlaff
Quoted 158 lines · Honnappa Nagarahalli
Quoted 158 lines
-----Original Message-----
From: Morten Brørup <[email protected]>
Sent: Wednesday, March 22, 2023 12:08 PM
To: Honnappa Nagarahalli <[email protected]>; Tyler Retzlaff <[email protected]>
Cc: Stephen Hemminger <[email protected]>; [email protected]; Ruifeng Wang <[email protected]>; [email protected]; nd <[email protected]>; nd <[email protected]>
Subject: RE: [PATCH 0/7] replace rte atomics with GCC builtin atomicsQuoted 145 lines
From: Honnappa Nagarahalli [mailto:[email protected]]
Sent: Wednesday, 22 March 2023 17.40Quoted 10 lines
From: Morten Brørup <[email protected]>
Sent: Wednesday, March 22, 2023 11:14 AMQuoted 7 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Wednesday, 22 March 2023 16.30Quoted 3 lines · Morten Brørup
Quoted 3 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Wednesday, 22 March 2023 15.22Quoted 20 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines · Morten Brørup
Quoted 5 lines
Quoted 5 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Friday, 17 March 2023 22.49On Fri, Mar 17, 2023 at 02:42:26PM -0700, Stephen
HemmingerQuoted 13 lines · wrote:
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 2 lines · On Fri, 17 Mar 2023 13:19:41 -0700 Tyler Retzlaff…
Replace the use of
rte_atomic.htypes and functions, instead useGCC
Quoted 7 lines
Quoted 7 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
supplied C++11 memory model builtins.
This series covers the libraries and drivers that are built on
Windows.
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
The code has be converted to use the
__atomicbuiltinsbut
Quoted 9 lines
Quoted 9 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
there
are
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
additional during conversion i notice that there may be some
issues
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 1 line
that need to be addressed.
I don't think all these cmpset need to use
SEQ_CST. Especially for the places where it is used a loop, mightbe
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 1 line
more efficient with some of the other memory models.
i agree.
however, i'm not trying to improve the code with this
change,
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
just decouple it from
rte_atomics.hso trying my best toavoid
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
any unnecessary semantic change.
certainly if the maintainers of this code wish to weaken the ordering where appropriate after the change is merged they
can
Quoted 3 lines
Quoted 3 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
do so and
handily
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
this change has enabled them to do so easily allowing them
to
Quoted 8 lines
Quoted 8 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
test
just
Quoted 6 lines
Quoted 6 lines
Quoted 4 lines
Quoted 1 line
their change in isolation.
I agree with the two-step approach, where this first step is a simple
search-and-replacement; but I insist that you add a FIXME or similar note where you have blindly used
SEQ_CST, indicatingthat
Quoted 5 lines
Quoted 5 lines
Quoted 1 line
Quoted 1 line
the memory order
needs to
Quoted 3 lines
Quoted 3 lines
be reviewed and potentially corrected.
i think the maintainers need to take some responsibility, if
they
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
see optimizations they missed when previously writing the code they need to follow up with a patch themselves. i can't do everything for them and marking things i'm not sure about will only lead to me having to churn patch series to remove the
unwanted
Quoted 9 lines
comments later.
Quoted 8 lines
Quoted 3 lines
The previous atomic functions didn't have the "memory order" parameter, so
the maintainers didn't have to think about it - and thus they didn't miss any optimizations when accepting the code.
Quoted 3 lines
I also agree 100 % that it is not your responsibility to consider
or
Quoted 9 lines
Quoted 9 lines
determine which memory order is appropriate!
Quoted 3 lines
But I think you should mark the locations where you are changing from the
old
rte_atomicfunctions (where no memory order optimization was available) to the new functions - to highlight where the option of memory ordering has been introduced and knowingly ignored (by you).Quoted 1 line
Quoted 4 lines
Quoted 2 lines
first, i have to apologize i confused myself about which of the many patch series i have up right now that you were commenting on.
No worries... you are rushing through quite an effort for this, so a
little
Quoted 2 lines
confusion is perfectly understandable. Especially when I'm replying to
an ageing
Quoted 9 lines
email. :-)
Quoted 4 lines
let me ask for clarification in relation to this series.
isn't that every single usage of the
rte_atomicAPIs?Probably, yes.
Quoted 7 lines
Review
Quoted 7 lines
i mean are you
literally asking for the entire patch series to look like the following patch snippet with the expectation that maintainers will come along and clean up/review after this series is merged?(quoted diff) +1
-1-rte_atomic_add32(&o, v); +//FIXME: opportunity for relaxing ordering constraint, pleasereview
Quoted 3 lines
Quoted 1 line
+__atomic_fetch_add(&o, v, order);
Exactly. And something similar for the
rte_atomicXX_tvariableschanged to
Quoted 4 lines
intXX_t, such as the packet counters.Realistically, I don't expect the maintainers to clean them up anytime
soon. The
Quoted 1 line
purpose is to make the FIXMEs stick until someone eventually cleans
them up, so
Quoted 1 line
they are not forgotten as time passes.
Cleaning up the
rte_atomicAPIs is a different effort. There is already lot of effort that has gone into this and there is more effort happening (rte_ringbeing a painful one)Instead of having FIXME, why not just send a separate patch with
SEQ_CST(still a search and replace)? We can leave the tougher ones likerte_ringas they are being worked on.Quoted 5 lines
Quoted 4 lines
The FIXME makes it possible in the future to differentiate between the instances that still need review and the instances that have been reviewed where
SEQ_CSTwas the correct choice. (Similarly for the choice of type for variables previouslyrte_atomicNN_t.)Apologies, relooked at the heading of this patch, got confused with other patches.
yeah, i did the same thing this morning :)
Quoted 4 lines
The changes Arm had done for
rte_atomic_to__atomic_xxxwere not direct replacements. The algorithms were studied, relaxed where required, race conditions fixed, performance benchmarked. IMO, we need to go through the same steps here.I looked at the series, we should just review the patch and make suggested changes. Are we constrained by any deadlines for this work?
i'm going to say yes but i'll qualify. the use of the
rte_atomic_xxxAPIs drags in extra work when creating a series that performs the actual conversions to the standard atomics.if i don't decouple ring from
rte_atomic_xxxthat means i have to go convert all therte_atomic.hto standard atomics and working around some of the implementation detail to do it is very time consuming. which then has further flow on effects because then i have to go fix every single driver that is still usingrte_atomic.h.incidentally i have a work in progress to decouple everything from
rte_atomic.h(including all drivers) but it would really negatively impact getting standard atomics introduced if we had to serialize the introduction behind a total removal ofrte_atomicor had to make changes to every consumer of the oldrte_atomicAPIs.if we can get by with a comment on the
rte_atomic_xxxlines in this series it would be helpful. when we bring the next series for standard atomics i'm not adverse to introducing changes to the ordering in that series if requested so long as i can get the series up 'soon' so there is lots of review time runway for 23.11.Quoted 1 line
Quoted 8 lines
Quoted 1 line
I would suggest to drop 1/7. Arm is working on removing the non-C11 algorithm for
rte_ring(not sure if we will be successful). I think it is better to explore this approach rather than the changes in patch 1/7.i think my answer here is timing. i'd rather take the work from arm but if it isn't coming for a while then it becomes a blocker.
we're waiting for the 23.07 start before this series can be merged. how about we re-evaluate where arm is at when the merge window opens. we can then decide to drop 1/7 or not at that time?
ping?
any update if there is going to be a series from arm as an acceptable replacement for patch 1/7? otherwise i think we should take the patch as is. it isn't altering the semantics of the code and is fairly low line count change so shouldn't distrupt any out of tree work as a result of the churn.
please update asap, this is one of the two series that is preventing submission of the first series converting to standard atomics for review.
thanks!
Quoted 2 lines
ty
Quoted 246 lines
-----Original Message-----
From: Tyler Retzlaff <[email protected]>
Sent: Monday, May 1, 2023 10:38 PM
To: Honnappa Nagarahalli <[email protected]>
Cc: Morten Brørup <[email protected]>; Stephen Hemminger <[email protected]>; [email protected]; Ruifeng Wang <[email protected]>; [email protected]; nd <[email protected]>
Subject: Re: [PATCH 0/7] replace rte atomics with GCC builtin atomicsQuoted 9 lines · Tyler Retzlaff
Quoted 8 lines · Honnappa Nagarahalli
Quoted 8 lines
-----Original Message-----
From: Morten Br�rup <[email protected]>
Sent: Wednesday, March 22, 2023 12:08 PM
To: Honnappa Nagarahalli <[email protected]>; Tyler Retzlaff <[email protected]>
Cc: Stephen Hemminger <[email protected]>;Quoted 15 lines
Quoted 15 lines
Quoted 15 lines
Ruifeng Wang <[email protected]>; [email protected]; nd <[email protected]>; nd <[email protected]>
Subject: RE: [PATCH 0/7] replace rte atomics with GCC builtin atomicsQuoted 10 lines
From: Honnappa Nagarahalli [mailto:[email protected]]
Sent: Wednesday, 22 March 2023 17.40Quoted 7 lines
From: Morten Br�rup <[email protected]>
Sent: Wednesday, March 22, 2023 11:14 AMQuoted 4 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Wednesday, 22 March 2023 16.30On Wed, Mar 22, 2023 at 03:58:07PM +0100, Morten Br�rup
Quoted 5 lines · wrote:
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
From: Tyler Retzlaff
[mailto:[email protected]]
Sent: Wednesday, 22 March 2023 15.22On Wed, Mar 22, 2023 at 12:28:44PM +0100, Morten Br�rup
Quoted 87 lines · wrote:
Quoted 87 lines
Quoted 87 lines
Quoted 87 lines
Quoted 13 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
From: Tyler Retzlaff
[mailto:[email protected]]
Sent: Friday, 17 March 2023 22.49On Fri, Mar 17, 2023 at 02:42:26PM -0700, Stephen
HemmingerQuoted 6 lines · wrote:
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 2 lines · On Fri, 17 Mar 2023 13:19:41 -0700 Tyler Retzlaff…
Replace the use of rte_atomic.h types and
functions, instead useGCC
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
supplied C++11 memory model builtins.
This series covers the libraries and drivers
that are built onWindows.
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
The code has be converted to use the __atomic
builtinsbut
Quoted 10 lines
Quoted 10 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
there
are
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
additional during conversion i notice that there
may be someissues
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 1 line
that need to be addressed.
I don't think all these cmpset need to use
SEQ_CST. Especially for the places where it is used a loop, mightbe
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 1 line
more efficient with some of the other memory models.
i agree.
however, i'm not trying to improve the code with
thischange,
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
just decouple it from
rte_atomics.hso trying my
best toavoid
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
any unnecessary semantic change.
certainly if the maintainers of this code wish to
weaken the ordering where appropriate after the
change is merged theycan
Quoted 4 lines
Quoted 4 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
do so and
handily
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
this change has enabled them to do so easily
allowing themto
Quoted 9 lines
Quoted 9 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
test
just
Quoted 7 lines
Quoted 7 lines
Quoted 4 lines
Quoted 1 line
their change in isolation.
I agree with the two-step approach, where this first step is a simple
search-and-replacement; but I insist that you add a FIXME or similar note where you have blindly used
SEQ_CST, indicatingthat
Quoted 6 lines
Quoted 6 lines
Quoted 1 line
Quoted 1 line
the memory order
needs to
Quoted 4 lines
Quoted 4 lines
be reviewed and potentially corrected.
i think the maintainers need to take some
responsibility, ifthey
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
see optimizations they missed when previously writing the code they need to follow up with a patch themselves. i can't do everything for them and marking things i'm not sure about will only lead to me having to churn patch series to remove the
unwanted
Quoted 9 lines
comments later.
Quoted 8 lines
Quoted 3 lines
The previous atomic functions didn't have the "memory order" parameter, so
the maintainers didn't have to think about it - and thus they didn't miss any optimizations when accepting the code.
Quoted 3 lines
I also agree 100 % that it is not your responsibility to consider
or
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
determine which memory order is appropriate!
Quoted 3 lines
But I think you should mark the locations where you are changing from the
old
rte_atomicfunctions (where no memory order optimization was
available) to the new functions - to highlight where the option of memory ordering has been introduced and knowinglyignored (by you).
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 1 line
first, i have to apologize i confused myself about which of the many patch series i have up right now that you were commenting
on.
Quoted 12 lines
Quoted 12 lines
Quoted 12 lines
Quoted 12 lines
Quoted 3 lines
No worries... you are rushing through quite an effort for this, so a
little
Quoted 2 lines
confusion is perfectly understandable. Especially when I'm replying to
an ageing
Quoted 5 lines
email. :-)
Quoted 3 lines
let me ask for clarification in relation to this series.
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 1 line
isn't that every single usage of the
rte_atomicAPIs?Probably, yes.
Quoted 4 lines
i mean are you
literally asking for the entire patch series to look like the following patch snippet with the expectation that maintainers will come along and clean up/review after this series ismerged?
Quoted 27 lines
Quoted 27 lines
Quoted 27 lines
Quoted 26 lines
Quoted 4 lines
Review
Quoted 4 lines
(quoted diff) +2
-1-rte_atomic_add32(&o, v); +//FIXME: opportunity for relaxing ordering constraint, +pleasereview
Quoted 4 lines
Quoted 1 line
+__atomic_fetch_add(&o, v, order);
Exactly. And something similar for the
rte_atomicXX_tvariableschanged to
Quoted 4 lines
intXX_t, such as the packet counters.Realistically, I don't expect the maintainers to clean them up anytime
soon. The
Quoted 2 lines
purpose is to make the FIXMEs stick until someone eventually cleans
them up, so
Quoted 1 line
they are not forgotten as time passes.
Cleaning up the
rte_atomicAPIs is a different effort. There is already lot of effort that has gone into this and there is more effort happening (rte_ringbeing a painful one)Instead of having FIXME, why not just send a separate patch with
SEQ_CST(still a search and replace)? We can leave the tougher ones likerte_ringas they are being worked on.Quoted 5 lines
Quoted 5 lines
Quoted 4 lines
The FIXME makes it possible in the future to differentiate between the instances that still need review and the instances that have been reviewed where
SEQ_CSTwas the correct choice. (Similarly for the choice of type for variables previouslyrte_atomicNN_t.)Apologies, relooked at the heading of this patch, got confused with other
patches.
Quoted 5 lines
yeah, i did the same thing this morning :)
Quoted 2 lines
The changes Arm had done for
rte_atomic_to__atomic_xxxwere not directreplacements. The algorithms were studied, relaxed where required, race conditions fixed, performance benchmarked. IMO, we need to go through the same steps here.
Quoted 2 lines
Quoted 2 lines
I looked at the series, we should just review the patch and make suggested
changes. Are we constrained by any deadlines for this work?
Quoted 17 lines
i'm going to say yes but i'll qualify. the use of the
rte_atomic_xxxAPIs drags in extra work when creating a series that performs the actual conversions to the standard atomics.if i don't decouple ring from
rte_atomic_xxxthat means i have to go convert all therte_atomic.hto standard atomics and working around some of the implementation detail to do it is very time consuming. which then has further flow on effects because then i have to go fix every single driver that is still usingrte_atomic.h.incidentally i have a work in progress to decouple everything from
rte_atomic.h(including all drivers) but it would really negatively impact getting standard atomics introduced if we had to serialize the introduction behind a total removal ofrte_atomicor had to make changes to every consumer of the oldrte_atomicAPIs.Quoted 8 lines
if we can get by with a comment on the
rte_atomic_xxxlines in this series it would be helpful. when we bring the next series for standard atomics i'm not adverse to introducing changes to the ordering in that series if requested so long as i can get the series up 'soon' so there is lots of review time runway for 23.11.Quoted 2 lines
I would suggest to drop 1/7. Arm is working on removing the non-C11
algorithm for
rte_ring(not sure if we will be successful). I think it is better to explore this approach rather than the changes in patch 1/7.Quoted 7 lines
i think my answer here is timing. i'd rather take the work from arm but if it isn't coming for a while then it becomes a blocker.
we're waiting for the 23.07 start before this series can be merged. how about we re-evaluate where arm is at when the merge window opens. we can then decide to drop 1/7 or not at that time?
Quoted 6 lines
ping?
any update if there is going to be a series from arm as an acceptable replacement for patch 1/7? otherwise i think we should take the patch as is. it isn't altering the semantics of the code and is fairly low line count change so shouldn't distrupt any out of tree work as a result of the churn.
Yes, we are working on a patch. There is a RFC [1], but we are still working on proving if the algorithm is correct. But, the plan is to find a solution (or present alternatives if there are no solutions) in 23.07 release.
[1] https://patchwork.dpdk.org/project/dpdk/patch/[email protected]/
Quoted 8 lines
please update asap, this is one of the two series that is preventing submission of the first series converting to standard atomics for review.
thanks!
Quoted 2 lines
ty
Hello Tyler,
Quoted 32 lines · Tyler Retzlaff
Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.This series covers the libraries and drivers that are built on Windows.
The code has be converted to use the
__atomicbuiltins but there are additional during conversion i notice that there may be some issues that need to be addressed.I'll comment in the patches where my concerns are so the maintainers may comment.
v3:
* style, don't use c99 commentsv2:
* comment code where optimizations may be possible now that memory order can be specified.
* comment code where operations should potentially be atomic so that maintainers can review.
* change a couple of variables labeled as counters to be unsigned.Tyler Retzlaff (7):
ring: replace rte atomics with GCC builtin atomics
stack: replace rte atomics with GCC builtin atomics dma/idxd: replace rte atomics with GCC builtin atomics net/ice: replace rte atomics with GCC builtin atomics net/ixgbe: replace rte atomics with GCC builtin atomics net/null: replace rte atomics with GCC builtin atomics net/ring: replace rte atomics with GCC builtin atomics
Quoted 17 lines
drivers/dma/idxd/idxd_internal.h | 3 +-- drivers/dma/idxd/idxd_pci.c | 8 +++++--- drivers/net/ice/ice_dcf.c | 1 - drivers/net/ice/ice_dcf_ethdev.c | 1 - drivers/net/ice/ice_ethdev.c | 12 ++++++++---- drivers/net/ixgbe/ixgbe_bypass.c | 1 - drivers/net/ixgbe/ixgbe_ethdev.c | 18 ++++++++++++------ drivers/net/ixgbe/ixgbe_ethdev.h | 3 ++- drivers/net/ixgbe/ixgbe_flow.c | 1 - drivers/net/ixgbe/ixgbe_rxtx.c | 1 - drivers/net/null/rte_eth_null.c | 28 ++++++++++++++++++---------- drivers/net/ring/rte_eth_ring.c | 26 ++++++++++++++++---------- lib/ring/rte_ring_core.h | 1 - lib/ring/rte_ring_generic_pvt.h | 12 ++++++++---- lib/stack/rte_stack_lf_generic.h | 16 +++++++++------- 15 files changed, 79 insertions(+), 53 deletions(-)
There is still some code using the DPDK "legacy" atomic API, but I guess this will be converted later.
As you proposed, I dropped patch 1 on the ring library (waiting for ARM to provide an alternative) and applied this series, thanks.
Note: Thomas, Ferruh, we will have to be careful when merging subtrees to make sure we are not reintroducing those again (like for example net/ice).
Signature · 1 line
David Marchand
Quoted 54 lines · David Marchand
Hello Tyler,
Quoted 32 lines · Tyler Retzlaff
Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.This series covers the libraries and drivers that are built on Windows.
The code has be converted to use the
__atomicbuiltins but there are additional during conversion i notice that there may be some issues that need to be addressed.I'll comment in the patches where my concerns are so the maintainers may comment.
v3:
* style, don't use c99 commentsv2:
* comment code where optimizations may be possible now that memory order can be specified.
* comment code where operations should potentially be atomic so that maintainers can review.
* change a couple of variables labeled as counters to be unsigned.Tyler Retzlaff (7):
ring: replace rte atomics with GCC builtin atomics
stack: replace rte atomics with GCC builtin atomics dma/idxd: replace rte atomics with GCC builtin atomics net/ice: replace rte atomics with GCC builtin atomics net/ixgbe: replace rte atomics with GCC builtin atomics net/null: replace rte atomics with GCC builtin atomics net/ring: replace rte atomics with GCC builtin atomicsQuoted 17 lines
drivers/dma/idxd/idxd_internal.h | 3 +-- drivers/dma/idxd/idxd_pci.c | 8 +++++--- drivers/net/ice/ice_dcf.c | 1 - drivers/net/ice/ice_dcf_ethdev.c | 1 - drivers/net/ice/ice_ethdev.c | 12 ++++++++---- drivers/net/ixgbe/ixgbe_bypass.c | 1 - drivers/net/ixgbe/ixgbe_ethdev.c | 18 ++++++++++++------ drivers/net/ixgbe/ixgbe_ethdev.h | 3 ++- drivers/net/ixgbe/ixgbe_flow.c | 1 - drivers/net/ixgbe/ixgbe_rxtx.c | 1 - drivers/net/null/rte_eth_null.c | 28 ++++++++++++++++++---------- drivers/net/ring/rte_eth_ring.c | 26 ++++++++++++++++---------- lib/ring/rte_ring_core.h | 1 - lib/ring/rte_ring_generic_pvt.h | 12 ++++++++---- lib/stack/rte_stack_lf_generic.h | 16 +++++++++------- 15 files changed, 79 insertions(+), 53 deletions(-)
Quoted 2 lines
There is still some code using the DPDK "legacy" atomic API, but I guess this will be converted later.
Yes, it will be converted later.
If I did it correctly... the series was an attempt to move away from the legacy API where there was a dependency on EAL that would change when moving to stdatomic. I'm hoping that the remaining use of the legacy API are not sensitive to the theoretical ABI surface changing when that move is complete.
Quoted 9 lines
As you proposed, I dropped patch 1 on the ring library (waiting for ARM to provide an alternative) and applied this series, thanks.
Note: Thomas, Ferruh, we will have to be careful when merging subtrees to make sure we are not reintroducing those again (like for example net/ice).
Signature · 1 line
David Marchand
Quoted 58 lines · Tyler Retzlaff
Quoted 54 lines · David Marchand
Hello Tyler,
Quoted 32 lines · Tyler Retzlaff
Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.This series covers the libraries and drivers that are built on Windows.
The code has be converted to use the
__atomicbuiltins but there are additional during conversion i notice that there may be some issues that need to be addressed.I'll comment in the patches where my concerns are so the maintainers may comment.
v3:
* style, don't use c99 commentsv2:
* comment code where optimizations may be possible now that memory order can be specified.
* comment code where operations should potentially be atomic so that maintainers can review.
* change a couple of variables labeled as counters to be unsigned.Tyler Retzlaff (7):
ring: replace rte atomics with GCC builtin atomics
stack: replace rte atomics with GCC builtin atomics dma/idxd: replace rte atomics with GCC builtin atomics net/ice: replace rte atomics with GCC builtin atomics net/ixgbe: replace rte atomics with GCC builtin atomics net/null: replace rte atomics with GCC builtin atomics net/ring: replace rte atomics with GCC builtin atomicsQuoted 17 lines
drivers/dma/idxd/idxd_internal.h | 3 +-- drivers/dma/idxd/idxd_pci.c | 8 +++++--- drivers/net/ice/ice_dcf.c | 1 - drivers/net/ice/ice_dcf_ethdev.c | 1 - drivers/net/ice/ice_ethdev.c | 12 ++++++++---- drivers/net/ixgbe/ixgbe_bypass.c | 1 - drivers/net/ixgbe/ixgbe_ethdev.c | 18 ++++++++++++------ drivers/net/ixgbe/ixgbe_ethdev.h | 3 ++- drivers/net/ixgbe/ixgbe_flow.c | 1 - drivers/net/ixgbe/ixgbe_rxtx.c | 1 - drivers/net/null/rte_eth_null.c | 28 ++++++++++++++++++---------- drivers/net/ring/rte_eth_ring.c | 26 ++++++++++++++++---------- lib/ring/rte_ring_core.h | 1 - lib/ring/rte_ring_generic_pvt.h | 12 ++++++++---- lib/stack/rte_stack_lf_generic.h | 16 +++++++++------- 15 files changed, 79 insertions(+), 53 deletions(-)Quoted 2 lines
There is still some code using the DPDK "legacy" atomic API, but I guess this will be converted later.
Quoted 7 lines
Yes, it will be converted later.
If I did it correctly... the series was an attempt to move away from the legacy API where there was a dependency on EAL that would change when moving to stdatomic. I'm hoping that the remaining use of the legacy API are not sensitive to the theoretical ABI surface changing when that move is complete.
Ok.
Quoted 6 lines
Quoted 6 lines
As you proposed, I dropped patch 1 on the ring library (waiting for ARM to provide an alternative) and applied this series, thanks.
Note: Thomas, Ferruh, we will have to be careful when merging subtrees to make sure we are not reintroducing those again (like for example net/ice).
Well, I have some second thought about this series so I did not push it to dpdk.org yet.
Drivers maintainers were not copied so I would like another pair of eyes on the series: ideally no /* Note: */ should be left when merging those patches.
I'll reply individually on the patches.
Signature · 1 line
David Marchand
Re: [PATCH v3 2/7] stack: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffHello Olivier,
Review please.
Quoted 58 lines · Tyler Retzlaff
Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.Signed-off-by Tyler Retzlaff <[email protected]>--- lib/stack/rte_stack_lf_generic.h | 16 +++++++++------- 1 file changed, 9 insertions(+), 7 deletions(-)lib/stack/rte_stack_lf_generic.h +9
-8@@ -26,8 +26,8 @@* elements. If the mempool is near-empty to the point that this is a * concern, the user should consider increasing the mempool size. */ -return (unsigned int )rte_atomic64_read ((rte_atomic64_t *) - &s->stack_lf.used.len); +/* NOTE: review for potential ordering optimization */ +return __atomic_load_n (&s->stack_lf.used.len, __ATOMIC_SEQ_CST); }static __rte_always_inline void@@ -67,8 +67,8 @@1 , __ATOMIC_RELEASE, __ATOMIC_RELAXED); }while (success ==0 ); - -rte_atomic64_add ((rte_atomic64_t *)&list->len, num); +/* NOTE: review for potential ordering optimization */ +__atomic_fetch_add (&list->len, num, __ATOMIC_SEQ_CST); } static __rte_always_inline struct rte_stack_lf_elem *@@ -82,14 +82,16 @@/* Reserve num elements, if available */ while (1 ) { -uint64_t len =rte_atomic64_read ((rte_atomic64_t *)&list->len); +/* NOTE: review for potential ordering optimization */ +uint64_t len =__atomic_load_n (&list->len, __ATOMIC_SEQ_CST);/* Does the list contain enough elements? */ if (unlikely (len < num))return NULL ; -if (rte_atomic64_cmpset ((volatile uint64_t *)&list->len, - len, len - num)) +/* NOTE: review for potential ordering optimization */ +if (__atomic_compare_exchange_n (&list->len, &len, len - num, +0 , __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))break ; } --1.8.3.1
Signature · 1 line
David Marchand
Re: [PATCH v3 3/7] dma/idxd: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffHello Bruce, Kevin,
Review please.
Quoted 67 lines · Tyler Retzlaff
Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.Signed-off-by Tyler Retzlaff <[email protected]>--- drivers/dma/idxd/idxd_internal.h | 3 +-- drivers/dma/idxd/idxd_pci.c | 8 +++++--- 2 files changed, 6 insertions(+), 5 deletions(-)drivers/dma/idxd/idxd_internal.h +1
-2@@ -7,7 +7,6 @@#include <rte_dmadev_pmd.h> #include <rte_spinlock.h> -#include <rte_atomic.h> #include "idxd_hw_defs.h" @@ -34,7 +33,7 @@ struct idxd_pci_common {rte_spinlock_t lk;uint8_t wq_cfg_sz; -rte_atomic16_t ref_count; +uint16_t ref_count;volatile struct rte_idxd_bar0 *regs;volatile uint32_t *wq_regs_base;volatile struct rte_idxd_grpcfg *grp_regs;drivers/dma/idxd/idxd_pci.c +5
-4@@ -6,7 +6,6 @@#include <rte_devargs.h> #include <rte_dmadev_pmd.h> #include <rte_malloc.h> -#include <rte_atomic.h> #include "idxd_internal.h" @@ -136,7 +135,9 @@/* if this is the last WQ on the device, disable the device and free * the PCI struct */ - is_last_wq =rte_atomic16_dec_and_test (&idxd->u.pci->ref_count); +/* NOTE: review for potential ordering optimization */ + is_last_wq =__atomic_fetch_sub (&idxd->u.pci->ref_count,1 , + __ATOMIC_SEQ_CST) -1 ==0 ;if (is_last_wq) {/* disable the device */ err_code =idxd_pci_dev_command (idxd, idxd_disable_dev);@@ -350,7 +351,8 @@free (idxd.u.pci);return ret; } -rte_atomic16_inc (&idxd.u.pci->ref_count); +/* NOTE: review for potential ordering optimization */ +__atomic_fetch_add (&idxd.u.pci->ref_count,1 , __ATOMIC_SEQ_CST); }return 0 ; --1.8.3.1
Signature · 1 line
David Marchand
Re: [PATCH v3 4/7] net/ice: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffHello Qiming, Qi,
Review please.
Quoted 68 lines · Tyler Retzlaff
Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.Signed-off-by Tyler Retzlaff <[email protected]>--- drivers/net/ice/ice_dcf.c | 1 - drivers/net/ice/ice_dcf_ethdev.c | 1 - drivers/net/ice/ice_ethdev.c | 12 ++++++++---- 3 files changed, 8 insertions(+), 6 deletions(-)drivers/net/ice/ice_dcf.c +0
-1@@ -14,7 +14,6 @@#include <rte_common.h> #include <rte_pci.h> -#include <rte_atomic.h> #include <rte_eal.h> #include <rte_ether.h> #include <ethdev_driver.h> drivers/net/ice/ice_dcf_ethdev.c +0
-1@@ -11,7 +11,6 @@#include <rte_interrupts.h> #include <rte_debug.h> #include <rte_pci.h> -#include <rte_atomic.h> #include <rte_eal.h> #include <rte_ether.h> #include <ethdev_pci.h> drivers/net/ice/ice_ethdev.c +8
-5@@ -3927,8 +3927,10 @@ static int ice_init_rss(struct ice_pf *pf)struct rte_eth_link *dst = link;struct rte_eth_link *src = &dev->data->dev_link; -if (rte_atomic64_cmpset ((uint64_t *)dst, *(uint64_t *)dst, - *(uint64_t *)src) ==0 ) +/* NOTE: review for potential ordering optimization */ +if (!__atomic_compare_exchange_n ((uint64_t *)dst, + (uint64_t *)dst, *(uint64_t *)src,0 , + __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))return -1 ;return 0 ;@@ -3941,8 +3943,10 @@ static int ice_init_rss(struct ice_pf *pf)struct rte_eth_link *dst = &dev->data->dev_link;struct rte_eth_link *src = link; -if (rte_atomic64_cmpset ((uint64_t *)dst, *(uint64_t *)dst, - *(uint64_t *)src) ==0 ) +/* NOTE: review for potential ordering optimization */ +if (!__atomic_compare_exchange_n ((uint64_t *)dst, + (uint64_t *)dst, *(uint64_t *)src,0 , + __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))return -1 ;return 0 ; --1.8.3.1
Signature · 1 line
David Marchand
Re: [PATCH v3 5/7] net/ixgbe: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffHello Qiming, Wenjun,
Review please.
Quoted 137 lines · Tyler Retzlaff
Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.Signed-off-by Tyler Retzlaff <[email protected]>--- drivers/net/ixgbe/ixgbe_bypass.c | 1 - drivers/net/ixgbe/ixgbe_ethdev.c | 18 ++++++++++++------ drivers/net/ixgbe/ixgbe_ethdev.h | 3 ++- drivers/net/ixgbe/ixgbe_flow.c | 1 - drivers/net/ixgbe/ixgbe_rxtx.c | 1 - 5 files changed, 14 insertions(+), 10 deletions(-)drivers/net/ixgbe/ixgbe_bypass.c +0
-1@@ -3,7 +3,6 @@*/ #include <time.h> -#include <rte_atomic.h> #include <ethdev_driver.h> #include "ixgbe_ethdev.h" #include "ixgbe_bypass_api.h" drivers/net/ixgbe/ixgbe_ethdev.c +12
-6@@ -1127,7 +1127,8 @@ struct rte_ixgbe_xstats_name_off {return 0 ; } -rte_atomic32_clear (&ad->link_thread_running); +/* NOTE: review for potential ordering optimization */ +__atomic_clear (&ad->link_thread_running, __ATOMIC_SEQ_CST);ixgbe_parse_devargs (eth_dev->data->dev_private, pci_dev->device.devargs);rte_eth_copy_pci_info (eth_dev, pci_dev);@@ -1625,7 +1626,8 @@ static int ixgbe_l2_tn_filter_init(struct rte_eth_dev *eth_dev)return 0 ; } -rte_atomic32_clear (&ad->link_thread_running); +/* NOTE: review for potential ordering optimization */ +__atomic_clear (&ad->link_thread_running, __ATOMIC_SEQ_CST);ixgbevf_parse_devargs (eth_dev->data->dev_private, pci_dev->device.devargs);@@ -4186,7 +4188,8 @@ static int ixgbevf_dev_xstats_get_names(__rte_unused struct rte_eth_dev *dev,struct ixgbe_adapter *ad = dev->data->dev_private;uint32_t timeout = timeout_ms ? timeout_ms :WARNING_TIMEOUT ; -while (rte_atomic32_read (&ad->link_thread_running)) { +/* NOTE: review for potential ordering optimization */ +while (__atomic_load_n (&ad->link_thread_running, __ATOMIC_SEQ_CST)) {msec_delay (1 ); timeout--;@@ -4222,7 +4225,8 @@ static int ixgbevf_dev_xstats_get_names(__rte_unused struct rte_eth_dev *dev,ixgbe_setup_link (hw, speed, true); intr->flags &= ~IXGBE_FLAG_NEED_LINK_CONFIG ; -rte_atomic32_clear (&ad->link_thread_running); +/* NOTE: review for potential ordering optimization */ +__atomic_clear (&ad->link_thread_running, __ATOMIC_SEQ_CST);return NULL ; }@@ -4317,7 +4321,8 @@ static int ixgbevf_dev_xstats_get_names(__rte_unused struct rte_eth_dev *dev,if (link_up ==0 ) {if (ixgbe_get_media_type (hw) == ixgbe_media_type_fiber) {ixgbe_dev_wait_setup_link_complete (dev,0 ); -if (rte_atomic32_test_and_set (&ad->link_thread_running)) { +/* NOTE: review for potential ordering optimization */ +if (__atomic_test_and_set (&ad->link_thread_running, __ATOMIC_SEQ_CST)) {/* To avoid race condition between threads, set * the IXGBE_FLAG_NEED_LINK_CONFIG flag only * when there is no link thread running. @@ -4330,7 +4335,8 @@ static int ixgbevf_dev_xstats_get_names(__rte_unused struct rte_eth_dev *dev, dev) < 0) {PMD_DRV_LOG (ERR ,"Create link thread failed!" ); -rte_atomic32_clear (&ad->link_thread_running); +/* NOTE: review for potential ordering optimization */ +__atomic_clear (&ad->link_thread_running, __ATOMIC_SEQ_CST); } }else { PMD_DRV_LOG(ERR,drivers/net/ixgbe/ixgbe_ethdev.h +2
-1@@ -6,6 +6,7 @@#define _IXGBE_ETHDEV_H_#include <stdint.h> +#include <stdbool.h> #include <sys/queue.h> #include "base/ixgbe_type.h" @@ -510,7 +511,7 @@ struct ixgbe_adapter {*/ uint8_t pflink_fullchk;uint8_t mac_ctrl_frame_fwd; -rte_atomic32_t link_thread_running; +bool link_thread_running;pthread_t link_thread_tid; };drivers/net/ixgbe/ixgbe_flow.c +0
-1@@ -18,7 +18,6 @@#include <rte_log.h> #include <rte_debug.h> #include <rte_pci.h> -#include <rte_atomic.h> #include <rte_branch_prediction.h> #include <rte_memory.h> #include <rte_eal.h> drivers/net/ixgbe/ixgbe_rxtx.c +0
-2@@ -27,7 +27,6 @@#include <rte_eal.h> #include <rte_per_lcore.h> #include <rte_lcore.h> -#include <rte_atomic.h> #include <rte_branch_prediction.h> #include <rte_mempool.h> #include <rte_malloc.h> --1.8.3.1
Signature · 1 line
David Marchand
Re: [PATCH v3 7/7] net/ring: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffHello Bruce,
For you again, review please.
Quoted 85 lines · Tyler Retzlaff
Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.Signed-off-by Tyler Retzlaff <[email protected]>--- drivers/net/ring/rte_eth_ring.c | 26 ++++++++++++++++---------- 1 file changed, 16 insertions(+), 10 deletions(-)drivers/net/ring/rte_eth_ring.c +16
-11@@ -44,8 +44,8 @@ enum dev_action {struct ring_queue {struct rte_ring *rng; -rte_atomic64_t rx_pkts; -rte_atomic64_t tx_pkts; +uint64_t rx_pkts; +uint64_t tx_pkts; };struct pmd_internals {@@ -80,9 +80,10 @@ struct pmd_internals {const uint16_t nb_rx = (uint16_t )rte_ring_dequeue_burst (r->rng, ptrs, nb_bufs,NULL );if (r->rng->flags &RING_F_SC_DEQ ) - r->rx_pkts.cnt += nb_rx; + r->rx_pkts += nb_rx;else -rte_atomic64_add (&(r->rx_pkts), nb_rx); +/* NOTE: review for potential ordering optimization */ +__atomic_fetch_add (&r->rx_pkts, nb_rx, __ATOMIC_SEQ_CST);return nb_rx; }@@ -94,9 +95,10 @@ struct pmd_internals {const uint16_t nb_tx = (uint16_t )rte_ring_enqueue_burst (r->rng, ptrs, nb_bufs,NULL );if (r->rng->flags &RING_F_SP_ENQ ) - r->tx_pkts.cnt += nb_tx; + r->tx_pkts += nb_tx;else -rte_atomic64_add (&(r->tx_pkts), nb_tx); +/* NOTE: review for potential ordering optimization */ +__atomic_fetch_add (&r->tx_pkts, nb_tx, __ATOMIC_SEQ_CST);return nb_tx; }@@ -184,13 +186,15 @@ struct pmd_internals {for (i =0 ; i <RTE_ETHDEV_QUEUE_STAT_CNTRS && i < dev->data->nb_rx_queues; i++) { - stats->q_ipackets[i] = internal->rx_ring_queues[i].rx_pkts.cnt; +/* NOTE: review for atomic access */ + stats->q_ipackets[i] = internal->rx_ring_queues[i].rx_pkts; rx_total += stats->q_ipackets[i]; }for (i =0 ; i <RTE_ETHDEV_QUEUE_STAT_CNTRS && i < dev->data->nb_tx_queues; i++) { - stats->q_opackets[i] = internal->tx_ring_queues[i].tx_pkts.cnt; +/* NOTE: review for atomic access */ + stats->q_opackets[i] = internal->tx_ring_queues[i].tx_pkts; tx_total += stats->q_opackets[i]; }@@ -207,9 +211,11 @@ struct pmd_internals {struct pmd_internals *internal = dev->data->dev_private;for (i =0 ; i < dev->data->nb_rx_queues; i++) - internal->rx_ring_queues[i].rx_pkts.cnt =0 ; +/* NOTE: review for atomic access */ + internal->rx_ring_queues[i].rx_pkts =0 ;for (i =0 ; i < dev->data->nb_tx_queues; i++) - internal->tx_ring_queues[i].tx_pkts.cnt =0 ; +/* NOTE: review for atomic access */ + internal->tx_ring_queues[i].tx_pkts =0 ;return 0 ; } --1.8.3.1
Signature · 1 line
David Marchand
Re: [PATCH v3 6/7] net/null: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffHello Tetsuya,
Review please.
Quoted 103 lines · Tyler Retzlaff
Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.Signed-off-by Tyler Retzlaff <[email protected]>--- drivers/net/null/rte_eth_null.c | 28 ++++++++++++++++++---------- 1 file changed, 18 insertions(+), 10 deletions(-)drivers/net/null/rte_eth_null.c +18
-11@@ -37,8 +37,8 @@ struct null_queue {struct rte_mempool *mb_pool;struct rte_mbuf *dummy_packet; -rte_atomic64_t rx_pkts; -rte_atomic64_t tx_pkts; +uint64_t rx_pkts; +uint64_t tx_pkts; };struct pmd_options {@@ -101,7 +101,8 @@ struct pmd_internals { bufs[i]->port = h->internals->port_id; } -rte_atomic64_add (&(h->rx_pkts), i); +/* NOTE: review for potential ordering optimization */ +__atomic_fetch_add (&h->rx_pkts, i, __ATOMIC_SEQ_CST);return i; }@@ -128,7 +129,8 @@ struct pmd_internals { bufs[i]->port = h->internals->port_id; } -rte_atomic64_add (&(h->rx_pkts), i); +/* NOTE: review for potential ordering optimization */ +__atomic_fetch_add (&h->rx_pkts, i, __ATOMIC_SEQ_CST);return i; }@@ -152,7 +154,8 @@ struct pmd_internals {for (i =0 ; i < nb_bufs; i++)rte_pktmbuf_free (bufs[i]); -rte_atomic64_add (&(h->tx_pkts), i); +/* NOTE: review for potential ordering optimization */ +__atomic_fetch_add (&h->tx_pkts, i, __ATOMIC_SEQ_CST);return i; }@@ -174,7 +177,8 @@ struct pmd_internals {rte_pktmbuf_free (bufs[i]); } -rte_atomic64_add (&(h->tx_pkts), i); +/* NOTE: review for potential ordering optimization */ +__atomic_fetch_add (&h->tx_pkts, i, __ATOMIC_SEQ_CST);return i; }@@ -316,8 +320,9 @@ struct pmd_internals {RTE_MIN (dev->data->nb_rx_queues, RTE_DIM(internal->rx_null_queues)));for (i =0 ; i < num_stats; i++) { +/* NOTE: review for atomic access */ igb_stats->q_ipackets[i] = - internal->rx_null_queues[i].rx_pkts.cnt; + internal->rx_null_queues[i].rx_pkts; rx_total += igb_stats->q_ipackets[i]; }@@ -325,8 +330,9 @@ struct pmd_internals {RTE_MIN (dev->data->nb_tx_queues, RTE_DIM(internal->tx_null_queues)));for (i =0 ; i < num_stats; i++) { +/* NOTE: review for atomic access */ igb_stats->q_opackets[i] = - internal->tx_null_queues[i].tx_pkts.cnt; + internal->tx_null_queues[i].tx_pkts; tx_total += igb_stats->q_opackets[i]; }@@ -347,9 +353,11 @@ struct pmd_internals { internal = dev->data->dev_private;for (i =0 ; i <RTE_DIM (internal->rx_null_queues); i++) - internal->rx_null_queues[i].rx_pkts.cnt =0 ; +/* NOTE: review for atomic access */ + internal->rx_null_queues[i].rx_pkts =0 ;for (i =0 ; i <RTE_DIM (internal->tx_null_queues); i++) - internal->tx_null_queues[i].tx_pkts.cnt =0 ; +/* NOTE: review for atomic access */ + internal->tx_null_queues[i].tx_pkts =0 ;return 0 ; } --1.8.3.1
Signature · 1 line
David Marchand
Quoted 74 lines · David Marchand
Quoted 58 lines · Tyler Retzlaff
Quoted 54 lines · David Marchand
Hello Tyler,
Quoted 32 lines · Tyler Retzlaff
Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.This series covers the libraries and drivers that are built on Windows.
The code has be converted to use the
__atomicbuiltins but there are additional during conversion i notice that there may be some issues that need to be addressed.I'll comment in the patches where my concerns are so the maintainers may comment.
v3:
* style, don't use c99 commentsv2:
* comment code where optimizations may be possible now that memory order can be specified.
* comment code where operations should potentially be atomic so that maintainers can review.
* change a couple of variables labeled as counters to be unsigned.Tyler Retzlaff (7):
ring: replace rte atomics with GCC builtin atomics
stack: replace rte atomics with GCC builtin atomics dma/idxd: replace rte atomics with GCC builtin atomics net/ice: replace rte atomics with GCC builtin atomics net/ixgbe: replace rte atomics with GCC builtin atomics net/null: replace rte atomics with GCC builtin atomics net/ring: replace rte atomics with GCC builtin atomicsQuoted 17 lines
drivers/dma/idxd/idxd_internal.h | 3 +-- drivers/dma/idxd/idxd_pci.c | 8 +++++--- drivers/net/ice/ice_dcf.c | 1 - drivers/net/ice/ice_dcf_ethdev.c | 1 - drivers/net/ice/ice_ethdev.c | 12 ++++++++---- drivers/net/ixgbe/ixgbe_bypass.c | 1 - drivers/net/ixgbe/ixgbe_ethdev.c | 18 ++++++++++++------ drivers/net/ixgbe/ixgbe_ethdev.h | 3 ++- drivers/net/ixgbe/ixgbe_flow.c | 1 - drivers/net/ixgbe/ixgbe_rxtx.c | 1 - drivers/net/null/rte_eth_null.c | 28 ++++++++++++++++++---------- drivers/net/ring/rte_eth_ring.c | 26 ++++++++++++++++---------- lib/ring/rte_ring_core.h | 1 - lib/ring/rte_ring_generic_pvt.h | 12 ++++++++---- lib/stack/rte_stack_lf_generic.h | 16 +++++++++------- 15 files changed, 79 insertions(+), 53 deletions(-)Quoted 2 lines
There is still some code using the DPDK "legacy" atomic API, but I guess this will be converted later.
Quoted 7 lines
Yes, it will be converted later.
If I did it correctly... the series was an attempt to move away from the legacy API where there was a dependency on EAL that would change when moving to stdatomic. I'm hoping that the remaining use of the legacy API are not sensitive to the theoretical ABI surface changing when that move is complete.
Ok.
Quoted 3 lines
Quoted 3 lines
As you proposed, I dropped patch 1 on the ring library (waiting for ARM to provide an alternative) and applied this series, thanks.
Quoted 6 lines
Quoted 3 lines
Quoted 3 lines
Note: Thomas, Ferruh, we will have to be careful when merging subtrees to make sure we are not reintroducing those again (like for example net/ice).
Well, I have some second thought about this series so I did not push it to dpdk.org yet.
Understood. It's very important to have these reviewed well so no objection just hope we can get them reviewed properly soon.
Quoted 3 lines
Drivers maintainers were not copied so I would like another pair of eyes on the series: ideally no /* Note: */ should be left when merging those patches.
The /* Note: */ was explicitly requested by other reviewers as they were concerned we would lose track of opportunities to weaken ordering after switching from __sync to __atomic.
Is your request that the comments now be removed?
Thanks!
Quoted 5 lines
I'll reply individually on the patches.
Signature · 1 line
David Marchand
Quoted 99 lines
-----Original Message-----
From: Tyler Retzlaff <[email protected]>
Sent: Wednesday, May 24, 2023 5:51 PM
To: David Marchand <[email protected]>
Cc: [email protected]; Honnappa Nagarahalli <[email protected]>; Ruifeng Wang <[email protected]>; [email protected]; [email protected]; [email protected]; Ferruh Yigit <[email protected]>
Subject: Re: [PATCH v3 0/7] replace rte atomics with GCC builtin atomicsQuoted 17 lines · David Marchand
Quoted 15 lines · Tyler Retzlaff
Quoted 14 lines · David Marchand
Hello Tyler,
Quoted 10 lines · Tyler Retzlaff
Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.This series covers the libraries and drivers that are built on Windows.
The code has be converted to use the
__atomicbuiltins but there are additional during conversion i notice that there may be some issues that need to be addressed.Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
I'll comment in the patches where my concerns are so the maintainers may comment.
v3:
* style, don't use c99 commentsv2:
* comment code where optimizations may be possible now thatmemory
Quoted 14 lines
Quoted 14 lines
Quoted 14 lines
Quoted 14 lines
order can be specified.
* comment code where operations should potentially be atomic so that maintainers can review.
* change a couple of variables labeled as counters to be unsigned.Tyler Retzlaff (7):
ring: replace rte atomics with GCC builtin atomics
stack: replace rte atomics with GCC builtin atomics dma/idxd: replace rte atomics with GCC builtin atomics net/ice: replace rte atomics with GCC builtin atomics net/ixgbe: replace rte atomics with GCC builtin atomics net/null: replace rte atomics with GCC builtin atomics net/ring: replace rte atomics with GCC builtin atomicsQuoted 12 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
drivers/dma/idxd/idxd_internal.h | 3 +-- drivers/dma/idxd/idxd_pci.c | 8 +++++--- drivers/net/ice/ice_dcf.c | 1 - drivers/net/ice/ice_dcf_ethdev.c | 1 -Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
drivers/net/ice/ice_ethdev.c | 12 ++++++++---- drivers/net/ixgbe/ixgbe_bypass.c | 1 - drivers/net/ixgbe/ixgbe_ethdev.c | 18 ++++++++++++------ drivers/net/ixgbe/ixgbe_ethdev.h | 3 ++- drivers/net/ixgbe/ixgbe_flow.c | 1 - drivers/net/ixgbe/ixgbe_rxtx.c | 1 - drivers/net/null/rte_eth_null.c | 28++++++++++++++++++----------
drivers/net/ring/rte_eth_ring.c| 26++++++++++++++++----------
Quoted 23 lines
Quoted 9 lines
Quoted 8 lines
Quoted 5 lines
lib/ring/rte_ring_core.h | 1 - lib/ring/rte_ring_generic_pvt.h | 12 ++++++++---- lib/stack/rte_stack_lf_generic.h | 16 +++++++++------- 15 files changed, 79 insertions(+), 53 deletions(-)There is still some code using the DPDK "legacy" atomic API, but I guess this will be converted later.
Quoted 7 lines
Yes, it will be converted later.
If I did it correctly... the series was an attempt to move away from the legacy API where there was a dependency on EAL that would change when moving to stdatomic. I'm hoping that the remaining use of the legacy API are not sensitive to the theoretical ABI surface changing when that move is complete.
Ok.
Quoted 3 lines
Quoted 3 lines
As you proposed, I dropped patch 1 on the ring library (waiting for ARM to provide an alternative) and applied this series, thanks.
Quoted 6 lines
Quoted 3 lines
Quoted 3 lines
Note: Thomas, Ferruh, we will have to be careful when merging subtrees to make sure we are not reintroducing those again (like for example net/ice).
Well, I have some second thought about this series so I did not push it to dpdk.org yet.
Understood. It's very important to have these reviewed well so no objection just hope we can get them reviewed properly soon.
Quoted 7 lines
Quoted 3 lines
Drivers maintainers were not copied so I would like another pair of eyes on the series: ideally no /* Note: */ should be left when merging those patches.
The /* Note: */ was explicitly requested by other reviewers as they were concerned we would lose track of opportunities to weaken ordering after switching from
__syncto__atomic.
Note that some of the changes that I checked are in control plane. While it is good to optimize those, but the benefits might not be much. The presence of SEQ_CST also can act as a note.
Quoted 10 lines
Is your request that the comments now be removed?
Thanks!
Quoted 5 lines
I'll reply individually on the patches.
--
David Marchand
Morten,
David and Honnappa are discussing the /* NOTE: */ comments that were added. If the three of you could come to conclusion about keeping or removing them it would be appreciated.
Thanks!
Quoted 118 lines · Honnappa Nagarahalli
Quoted 99 lines
-----Original Message-----
From: Tyler Retzlaff <[email protected]>
Sent: Wednesday, May 24, 2023 5:51 PM
To: David Marchand <[email protected]>
Cc: [email protected]; Honnappa Nagarahalli <[email protected]>; Ruifeng Wang <[email protected]>; [email protected]; [email protected]; [email protected]; Ferruh Yigit <[email protected]>
Subject: Re: [PATCH v3 0/7] replace rte atomics with GCC builtin atomicsQuoted 17 lines · David Marchand
Quoted 15 lines · Tyler Retzlaff
Quoted 14 lines · David Marchand
Hello Tyler,
Quoted 10 lines · Tyler Retzlaff
Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.This series covers the libraries and drivers that are built on Windows.
The code has be converted to use the
__atomicbuiltins but there are additional during conversion i notice that there may be some issues that need to be addressed.Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
I'll comment in the patches where my concerns are so the maintainers may comment.
v3:
* style, don't use c99 commentsv2:
* comment code where optimizations may be possible now thatmemory
Quoted 14 lines
Quoted 14 lines
Quoted 14 lines
Quoted 14 lines
order can be specified.
* comment code where operations should potentially be atomic so that maintainers can review.
* change a couple of variables labeled as counters to be unsigned.Tyler Retzlaff (7):
ring: replace rte atomics with GCC builtin atomics
stack: replace rte atomics with GCC builtin atomics dma/idxd: replace rte atomics with GCC builtin atomics net/ice: replace rte atomics with GCC builtin atomics net/ixgbe: replace rte atomics with GCC builtin atomics net/null: replace rte atomics with GCC builtin atomics net/ring: replace rte atomics with GCC builtin atomicsQuoted 12 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
drivers/dma/idxd/idxd_internal.h | 3 +-- drivers/dma/idxd/idxd_pci.c | 8 +++++--- drivers/net/ice/ice_dcf.c | 1 - drivers/net/ice/ice_dcf_ethdev.c | 1 -Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
drivers/net/ice/ice_ethdev.c | 12 ++++++++---- drivers/net/ixgbe/ixgbe_bypass.c | 1 - drivers/net/ixgbe/ixgbe_ethdev.c | 18 ++++++++++++------ drivers/net/ixgbe/ixgbe_ethdev.h | 3 ++- drivers/net/ixgbe/ixgbe_flow.c | 1 - drivers/net/ixgbe/ixgbe_rxtx.c | 1 - drivers/net/null/rte_eth_null.c | 28++++++++++++++++++----------
drivers/net/ring/rte_eth_ring.c| 26++++++++++++++++----------
Quoted 23 lines
Quoted 9 lines
Quoted 8 lines
Quoted 5 lines
lib/ring/rte_ring_core.h | 1 - lib/ring/rte_ring_generic_pvt.h | 12 ++++++++---- lib/stack/rte_stack_lf_generic.h | 16 +++++++++------- 15 files changed, 79 insertions(+), 53 deletions(-)There is still some code using the DPDK "legacy" atomic API, but I guess this will be converted later.
Quoted 7 lines
Yes, it will be converted later.
If I did it correctly... the series was an attempt to move away from the legacy API where there was a dependency on EAL that would change when moving to stdatomic. I'm hoping that the remaining use of the legacy API are not sensitive to the theoretical ABI surface changing when that move is complete.
Ok.
Quoted 3 lines
Quoted 3 lines
As you proposed, I dropped patch 1 on the ring library (waiting for ARM to provide an alternative) and applied this series, thanks.
Quoted 6 lines
Quoted 3 lines
Quoted 3 lines
Note: Thomas, Ferruh, we will have to be careful when merging subtrees to make sure we are not reintroducing those again (like for example net/ice).
Well, I have some second thought about this series so I did not push it to dpdk.org yet.
Understood. It's very important to have these reviewed well so no objection just hope we can get them reviewed properly soon.
Quoted 7 lines
Quoted 3 lines
Drivers maintainers were not copied so I would like another pair of eyes on the series: ideally no /* Note: */ should be left when merging those patches.
The /* Note: */ was explicitly requested by other reviewers as they were concerned we would lose track of opportunities to weaken ordering after switching from
__syncto__atomic.Note that some of the changes that I checked are in control plane. While it is good to optimize those, but the benefits might not be much. The presence of
SEQ_CSTalso can act as a note.Quoted 10 lines
Is your request that the comments now be removed?
Thanks!
Quoted 5 lines
I'll reply individually on the patches.
--
David Marchand
Quoted 34 lines
From: Tyler Retzlaff [mailto:[email protected]]
Sent: Thursday, 25 May 2023 02.03Morten,
David and Honnappa are discussing the /* NOTE: */ comments that were added. If the three of you could come to conclusion about keeping or removing them it would be appreciated.
Thanks!
Quoted 10 lines · Honnappa Nagarahalli
Quoted 10 lines
From: Tyler Retzlaff <[email protected]>
Sent: Wednesday, May 24, 2023 5:51 PMQuoted 5 lines · David Marchand
Quoted 3 lines · Tyler Retzlaff
Quoted 2 lines · David Marchand
Hello Tyler,
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 5 lines · Tyler Retzlaff
Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.This series covers the libraries and drivers that are built on
Windows.
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
The code has be converted to use the
__atomicbuiltins but there are additional during conversion i notice that there may be some issues that need to be addressed.
[...]
Quoted 14 lines
Quoted 4 lines
Quoted 4 lines
Quoted 2 lines
Well, I have some second thought about this series so I did not push it to dpdk.org yet.
Understood. It's very important to have these reviewed well so no
objection just
Quoted 9 lines
Quoted 9 lines
hope we can get them reviewed properly soon.
Quoted 3 lines
Drivers maintainers were not copied so I would like another pair of eyes on the series: ideally no /* Note: */ should be left when merging those patches.
The /* Note: */ was explicitly requested by other reviewers as they were concerned we would lose track of opportunities to weaken ordering after switching from
__syncto__atomic.
This patch series is an important step towards the more flexible C11 atomics, and I consider further optimization "nice to have", not "must have".
So I don't think we should hold back these patches and require of the maintainers to optimize the atomic accesses before. I would rather leave the notes in the code, so they can be optimized by anyone with the required skills and/or testing facilities at a later time.
I agree that it would be ideal if anyone (e.g. the maintainers) can make optimize the affected libraries/drivers in time for the coming release, but they can be separate patches after this series.
Quoted 3 lines
Quoted 1 line
Note that some of the changes that I checked are in control plane. While it
is good to optimize those, but the benefits might not be much. The presence of
SEQ_CSTalso can act as a note.
I vote against using SEQ_CST as a note. SEQ_CST might be the correct memory order in some locations, so it would require a note in those locations that SEQ_CST has been reviewed and is the optimal memory order there. I would rather have notes where we know that further consideration for optimization is warranted.
If atomics are used in the control plane, the memory order still need to be correct (i.e. not causing failure, which SEQ_CST should assure). So the note should remain, if not reviewed for optimization. A reviewer can add to the note that this is control plane only, so optimization is not important. Alternatively, if those control plane variables don't need to be atomics, they can be replaced by non-atomic types and accesses - such a modification can also be considered an optimization.
PS: If someone spotted an opportunity for optimization anywhere in DPDK, but was unable to implement and/or test it himself, adding a note about it in the source code could be an alternative. On the other hand, such ideas might belong in Bugzilla instead... (Just arguing for keeping the notes. Not trying to broaden the discussion!)
Bruce Richardson
Acked-byRe: [PATCH v3 3/7] dma/idxd: replace rte atomics with GCC builtin atomics
In reply to David MarchandQuoted 12 lines · David Marchand
Hello Bruce, Kevin,
Review please.
Quoted 5 lines · Tyler Retzlaff
Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.Signed-off-by Tyler Retzlaff <[email protected]>
Two small comments inline below.
Quoted 38 lines
Quoted 38 lines
--- drivers/dma/idxd/idxd_internal.h | 3 +-- drivers/dma/idxd/idxd_pci.c | 8 +++++--- 2 files changed, 6 insertions(+), 5 deletions(-)drivers/dma/idxd/idxd_internal.h +1
-2@@ -7,7 +7,6 @@#include <rte_dmadev_pmd.h> #include <rte_spinlock.h> -#include <rte_atomic.h> #include "idxd_hw_defs.h" @@ -34,7 +33,7 @@ struct idxd_pci_common {rte_spinlock_t lk;uint8_t wq_cfg_sz; -rte_atomic16_t ref_count; +uint16_t ref_count;volatile struct rte_idxd_bar0 *regs;volatile uint32_t *wq_regs_base;volatile struct rte_idxd_grpcfg *grp_regs;drivers/dma/idxd/idxd_pci.c +0
-1@@ -6,7 +6,6 @@#include <rte_devargs.h> #include <rte_dmadev_pmd.h> #include <rte_malloc.h> -#include <rte_atomic.h> #include "idxd_internal.h"
Quoted 8 lines
drivers/dma/idxd/idxd_pci.c line 140
Quoted 8 lines
drivers/dma/idxd/idxd_pci.c +3
-1@@ -136,7 +135,9 @@/* if this is the last WQ on the device, disable the device and free * the PCI struct */ - is_last_wq =rte_atomic16_dec_and_test (&idxd->u.pci->ref_count); +/* NOTE: review for potential ordering optimization */ + is_last_wq =__atomic_fetch_sub (&idxd->u.pci->ref_count,1 , + __ATOMIC_SEQ_CST) -1 ==0 ;
Rather than "__atomic_fetch_sub(...) - 1 == 0", I think just comparing "== 1" is simpler and better. I would also bracket the comparison for clarity.
Quoted 9 lines
drivers/dma/idxd/idxd_pci.c line 354
Quoted 9 lines
drivers/dma/idxd/idxd_pci.c +1
-1if (is_last_wq) {/* disable the device */ err_code =idxd_pci_dev_command (idxd, idxd_disable_dev);@@ -350,7 +351,8 @@free (idxd.u.pci);return ret; } -rte_atomic16_inc (&idxd.u.pci->ref_count); +/* NOTE: review for potential ordering optimization */
I think we can drop the note. Since this is not datapath code the perf is not that important.
Quoted 11 lines
Quoted 7 lines
drivers/dma/idxd/idxd_pci.c +1
-1+__atomic_fetch_add (&idxd.u.pci->ref_count,1 , __ATOMIC_SEQ_CST); }return 0 ; --1.8.3.1
Signature · 1 line
David Marchand
Bruce Richardson
Acked-byRe: [PATCH v3 7/7] net/ring: replace rte atomics with GCC builtin atomics
In reply to David MarchandQuoted 13 lines · David Marchand
Hello Bruce,
For you again, review please.
Quoted 6 lines · Tyler Retzlaff
Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.Signed-off-by Tyler Retzlaff <[email protected]>---
Re: [PATCH v3 3/7] dma/idxd: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffQuoted 9 lines · Tyler Retzlaff
Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.Signed-off-by Tyler Retzlaff <[email protected]>--- drivers/dma/idxd/idxd_internal.h | 3 +-- drivers/dma/idxd/idxd_pci.c | 8 +++++--- 2 files changed, 6 insertions(+), 5 deletions(-)
RE: [PATCH v3 3/7] dma/idxd: replace rte atomics with GCC builtin atomics
In reply to Bruce RichardsonQuoted 74 lines
From: Bruce Richardson [mailto:[email protected]]
Sent: Thursday, 25 May 2023 10.42Quoted 12 lines · David Marchand
Hello Bruce, Kevin,
Review please.
Quoted 5 lines · Tyler Retzlaff
Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.Signed-off-by Tyler Retzlaff <[email protected]>Two small comments inline below.
Acked-by Bruce Richardson <[email protected]>Quoted 6 lines
Review
Quoted 6 lines
--- drivers/dma/idxd/idxd_internal.h | 3 +-- drivers/dma/idxd/idxd_pci.c | 8 +++++--- 2 files changed, 6 insertions(+), 5 deletions(-)(quoted diff) +0
-0
b/drivers/dma/idxd/idxd_internal.hQuoted 30 lines
Quoted 30 lines
index 180a858..cd41777 100644
drivers/dma/idxd/idxd_internal.h +1
-2@@ -7,7 +7,6 @@#include <rte_dmadev_pmd.h> #include <rte_spinlock.h> -#include <rte_atomic.h> #include "idxd_hw_defs.h" @@ -34,7 +33,7 @@ struct idxd_pci_common {rte_spinlock_t lk;uint8_t wq_cfg_sz; -rte_atomic16_t ref_count; +uint16_t ref_count;volatile struct rte_idxd_bar0 *regs;volatile uint32_t *wq_regs_base;volatile struct rte_idxd_grpcfg *grp_regs;drivers/dma/idxd/idxd_pci.c +0
-1@@ -6,7 +6,6 @@#include <rte_devargs.h> #include <rte_dmadev_pmd.h> #include <rte_malloc.h> -#include <rte_atomic.h> Quoted 4 lines
drivers/dma/idxd/idxd_pci.c line 135
Quoted 4 lines
drivers/dma/idxd/idxd_pci.c +0
-0#include "idxd_internal.h" @@ -136,7 +135,9 @@/* if this is the last WQ on the device, disable the device and free
Quoted 6 lines
drivers/dma/idxd/idxd_pci.c line 140
Quoted 6 lines
drivers/dma/idxd/idxd_pci.c +3
-1* the PCI struct */ - is_last_wq =rte_atomic16_dec_and_test (&idxd->u.pci->ref_count); +/* NOTE: review for potential ordering optimization */ + is_last_wq =__atomic_fetch_sub (&idxd->u.pci->ref_count,1 , + __ATOMIC_SEQ_CST) -1 ==0 ;Rather than "
__atomic_fetch_sub(...)- 1 == 0", I think just comparing "== 1" is simpler and better. I would also bracket the comparison for clarity.
Quoted 12 lines
Quoted 9 lines
drivers/dma/idxd/idxd_pci.c line 354
Quoted 9 lines
drivers/dma/idxd/idxd_pci.c +1
-1if (is_last_wq) {/* disable the device */ err_code =idxd_pci_dev_command (idxd, idxd_disable_dev);@@ -350,7 +351,8 @@free (idxd.u.pci);return ret; } -rte_atomic16_inc (&idxd.u.pci->ref_count); +/* NOTE: review for potential ordering optimization */ I think we can drop the note. Since this is not datapath code the perf is not that important.
Following up on my previous input to the discussion about these notes...
I agree with Bruce on this location. Here it is purely used in the control plane, and atomicity is required, but optimization of this would be a waste of brain power, so we can drop the notes in such situations. Perhaps Honnappa was referring to something similar - and then I agree with Honnappa too. ;-)
In principle: This specific note has been actively considered for optimization, and the conclusion was that further optimization is not required, and thus SEQ_CST is the correct choice here. Ideal to change now, but could be changed with a later (separate) patch as well.
Quoted 13 lines
Quoted 2 lines
drivers/dma/idxd/idxd_pci.c line 355
Quoted 2 lines
drivers/dma/idxd/idxd_pci.c +1
-0+ __atomic_fetch_add(&idxd.u.pci->ref_count, 1,
__ATOMIC_SEQ_CST);Quoted 10 lines
Quoted 6 lines
drivers/dma/idxd/idxd_pci.c +0
-1}return 0 ; --1.8.3.1
--
David Marchand
PATCH v4 1/6 stack: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffReplace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
--- lib/stack/rte_stack_lf_generic.h | 16 +++++++++------- 1 file changed, 9 insertions(+), 7 deletions(-)
lib/stack/rte_stack_lf_generic.h +9 -7
@@ -26,8 +26,8 @@* elements. If the mempool is near-empty to the point that this is a * concern, the user should consider increasing the mempool size. */ -return (unsigned int )rte_atomic64_read ((rte_atomic64_t *) - &s->stack_lf.used.len); +/* NOTE: review for potential ordering optimization */ +return __atomic_load_n (&s->stack_lf.used.len, __ATOMIC_SEQ_CST); }static __rte_always_inline void
@@ -67,8 +67,8 @@1 , __ATOMIC_RELEASE, __ATOMIC_RELAXED); }while (success ==0 ); - -rte_atomic64_add ((rte_atomic64_t *)&list->len, num); +/* NOTE: review for potential ordering optimization */ +__atomic_fetch_add (&list->len, num, __ATOMIC_SEQ_CST); } static __rte_always_inline struct rte_stack_lf_elem *
@@ -82,14 +82,16 @@/* Reserve num elements, if available */ while (1 ) { -uint64_t len =rte_atomic64_read ((rte_atomic64_t *)&list->len); +/* NOTE: review for potential ordering optimization */ +uint64_t len =__atomic_load_n (&list->len, __ATOMIC_SEQ_CST);/* Does the list contain enough elements? */ if (unlikely (len < num))return NULL ; -if (rte_atomic64_cmpset ((volatile uint64_t *)&list->len, - len, len - num)) +/* NOTE: review for potential ordering optimization */ +if (__atomic_compare_exchange_n (&list->len, &len, len - num, +0 , __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))break ; }
Signature · 1 line
1.8.3.1
PATCH v4 3/6 net/ice: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffReplace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
--- drivers/net/ice/ice_dcf.c | 1 - drivers/net/ice/ice_dcf_ethdev.c | 1 - drivers/net/ice/ice_ethdev.c | 12 ++++++++---- 3 files changed, 8 insertions(+), 6 deletions(-)
drivers/net/ice/ice_dcf.c +0 -1
@@ -14,7 +14,6 @@#include <rte_common.h> #include <rte_pci.h> -#include <rte_atomic.h> #include <rte_eal.h> #include <rte_ether.h> #include <ethdev_driver.h>
drivers/net/ice/ice_dcf_ethdev.c +0 -1
@@ -11,7 +11,6 @@#include <rte_interrupts.h> #include <rte_debug.h> #include <rte_pci.h> -#include <rte_atomic.h> #include <rte_eal.h> #include <rte_ether.h> #include <ethdev_pci.h>
drivers/net/ice/ice_ethdev.c +8 -4
@@ -3927,8 +3927,10 @@ static int ice_init_rss(struct ice_pf *pf)struct rte_eth_link *dst = link;struct rte_eth_link *src = &dev->data->dev_link; -if (rte_atomic64_cmpset ((uint64_t *)dst, *(uint64_t *)dst, - *(uint64_t *)src) ==0 ) +/* NOTE: review for potential ordering optimization */ +if (!__atomic_compare_exchange_n ((uint64_t *)dst, + (uint64_t *)dst, *(uint64_t *)src,0 , + __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))return -1 ;return 0 ;
@@ -3941,8 +3943,10 @@ static int ice_init_rss(struct ice_pf *pf)struct rte_eth_link *dst = &dev->data->dev_link;struct rte_eth_link *src = link; -if (rte_atomic64_cmpset ((uint64_t *)dst, *(uint64_t *)dst, - *(uint64_t *)src) ==0 ) +/* NOTE: review for potential ordering optimization */ +if (!__atomic_compare_exchange_n ((uint64_t *)dst, + (uint64_t *)dst, *(uint64_t *)src,0 , + __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))return -1 ;return 0 ;
Signature · 1 line
1.8.3.1
v4 cover replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffReplace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
This series covers the libraries and drivers that are built on Windows.
The code has be converted to use the __atomic builtins but there are additional during conversion I notice that there may be some issues that need to be addressed.
I'll comment in the patches where my concerns are so the maintainers may comment.
v4:
* drop patch for lib/ring it will be provided by ARM / Honnappa
* rebase for changes in dma/idxd merge
* adapt __atomic_fetch_sub(...) - 1 == 0 to be (__atomic_fetch_sub(...) == 1) as per feedback.
* drop one /* NOTE: review for potential ordering optimization */ since the note reference non-critical to perf control path.
note:
Remainder of the NOTE comments have been retained since there seems to be no consensus but stronger opinion/argument to keep expressed. while I generally agree that changes should not include ``TODO'' style comments I also agree that without these comments in your face people are very unlikely to feel compelled to make the review they are trying to solicit without them. if it is absolute that the series won't be merged with them then I will remove them, but please be explicit soon.
v3:
* style, don't use c99 comments
v2:
* comment code where optimizations may be possible now that memory order can be specified.
* comment code where operations should potentially be atomic so that maintainers can review.
* change a couple of variables labeled as counters to be unsigned.
Tyler Retzlaff (6):
stack: replace rte atomics with GCC builtin atomics dma/idxd: replace rte atomics with GCC builtin atomics net/ice: replace rte atomics with GCC builtin atomics net/ixgbe: replace rte atomics with GCC builtin atomics net/null: replace rte atomics with GCC builtin atomics net/ring: replace rte atomics with GCC builtin atomics
drivers/dma/idxd/idxd_internal.h | 3 +-- drivers/dma/idxd/idxd_pci.c | 11 ++++++----- drivers/net/ice/ice_dcf.c | 1 - drivers/net/ice/ice_dcf_ethdev.c | 1 - drivers/net/ice/ice_ethdev.c | 12 ++++++++---- drivers/net/ixgbe/ixgbe_bypass.c | 1 - drivers/net/ixgbe/ixgbe_ethdev.c | 18 ++++++++++++------ drivers/net/ixgbe/ixgbe_ethdev.h | 3 ++- drivers/net/ixgbe/ixgbe_flow.c | 1 - drivers/net/ixgbe/ixgbe_rxtx.c | 1 - drivers/net/null/rte_eth_null.c | 28 ++++++++++++++++++---------- drivers/net/ring/rte_eth_ring.c | 26 ++++++++++++++++---------- lib/stack/rte_stack_lf_generic.h | 16 +++++++++------- 13 files changed, 72 insertions(+), 50 deletions(-)
Signature · 1 line
1.8.3.1
PATCH v4 2/6 dma/idxd: replace rte atomics with GCC builtin atomics
Replace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
--- drivers/dma/idxd/idxd_internal.h | 3 +-- drivers/dma/idxd/idxd_pci.c | 11 ++++++----- 2 files changed, 7 insertions(+), 7 deletions(-)
drivers/dma/idxd/idxd_internal.h +1 -2
@@ -7,7 +7,6 @@#include <rte_dmadev_pmd.h> #include <rte_spinlock.h> -#include <rte_atomic.h> #include "idxd_hw_defs.h"
@@ -34,7 +33,7 @@ struct idxd_pci_common {rte_spinlock_t lk;uint8_t wq_cfg_sz; -rte_atomic16_t ref_count; +uint16_t ref_count;volatile struct rte_idxd_bar0 *regs;volatile uint32_t *wq_regs_base;volatile struct rte_idxd_grpcfg *grp_regs;
drivers/dma/idxd/idxd_pci.c +6 -5
@@ -6,7 +6,6 @@#include <rte_devargs.h> #include <rte_dmadev_pmd.h> #include <rte_malloc.h> -#include <rte_atomic.h> #include "idxd_internal.h"
@@ -136,7 +135,8 @@/* if this is the last WQ on the device, disable the device and free * the PCI struct */ - is_last_wq =rte_atomic16_dec_and_test (&idxd->u.pci->ref_count); +/* NOTE: review for potential ordering optimization */ + is_last_wq = (__atomic_fetch_sub (&idxd->u.pci->ref_count,1 , __ATOMIC_SEQ_CST) ==1 );if (is_last_wq) {/* disable the device */ err_code =idxd_pci_dev_command (idxd, idxd_disable_dev);
@@ -322,8 +322,9 @@return ret; } qid =rte_dma_get_dev_id_by_name (qname); - max_qid =rte_atomic16_read ( - &((struct idxd_dmadev *)rte_dma_fp_objs[qid].dev_private)->u.pci->ref_count); + max_qid =__atomic_load_n ( + &((struct idxd_dmadev *)rte_dma_fp_objs[qid].dev_private)->u.pci->ref_count, + __ATOMIC_SEQ_CST);/* we have queue 0 done, now configure the rest of the queues */ for (qid =1 ; qid < max_qid; qid++) {
@@ -380,7 +381,7 @@free (idxd.u.pci);return ret; } -rte_atomic16_inc (&idxd.u.pci->ref_count); +__atomic_fetch_add (&idxd.u.pci->ref_count,1 , __ATOMIC_SEQ_CST); }return 0 ;
Signature · 1 line
1.8.3.1
PATCH v4 4/6 net/ixgbe: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffReplace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
--- drivers/net/ixgbe/ixgbe_bypass.c | 1 - drivers/net/ixgbe/ixgbe_ethdev.c | 18 ++++++++++++------ drivers/net/ixgbe/ixgbe_ethdev.h | 3 ++- drivers/net/ixgbe/ixgbe_flow.c | 1 - drivers/net/ixgbe/ixgbe_rxtx.c | 1 - 5 files changed, 14 insertions(+), 10 deletions(-)
drivers/net/ixgbe/ixgbe_bypass.c +0 -1
@@ -3,7 +3,6 @@*/ #include <time.h> -#include <rte_atomic.h> #include <ethdev_driver.h> #include "ixgbe_ethdev.h" #include "ixgbe_bypass_api.h"
drivers/net/ixgbe/ixgbe_ethdev.c +12 -6
@@ -1127,7 +1127,8 @@ struct rte_ixgbe_xstats_name_off {return 0 ; } -rte_atomic32_clear (&ad->link_thread_running); +/* NOTE: review for potential ordering optimization */ +__atomic_clear (&ad->link_thread_running, __ATOMIC_SEQ_CST);ixgbe_parse_devargs (eth_dev->data->dev_private, pci_dev->device.devargs);rte_eth_copy_pci_info (eth_dev, pci_dev);
@@ -1625,7 +1626,8 @@ static int ixgbe_l2_tn_filter_init(struct rte_eth_dev *eth_dev)return 0 ; } -rte_atomic32_clear (&ad->link_thread_running); +/* NOTE: review for potential ordering optimization */ +__atomic_clear (&ad->link_thread_running, __ATOMIC_SEQ_CST);ixgbevf_parse_devargs (eth_dev->data->dev_private, pci_dev->device.devargs);
@@ -4186,7 +4188,8 @@ static int ixgbevf_dev_xstats_get_names(__rte_unused struct rte_eth_dev *dev,struct ixgbe_adapter *ad = dev->data->dev_private;uint32_t timeout = timeout_ms ? timeout_ms :WARNING_TIMEOUT ; -while (rte_atomic32_read (&ad->link_thread_running)) { +/* NOTE: review for potential ordering optimization */ +while (__atomic_load_n (&ad->link_thread_running, __ATOMIC_SEQ_CST)) {msec_delay (1 ); timeout--;
@@ -4222,7 +4225,8 @@ static int ixgbevf_dev_xstats_get_names(__rte_unused struct rte_eth_dev *dev,ixgbe_setup_link (hw, speed, true); intr->flags &= ~IXGBE_FLAG_NEED_LINK_CONFIG ; -rte_atomic32_clear (&ad->link_thread_running); +/* NOTE: review for potential ordering optimization */ +__atomic_clear (&ad->link_thread_running, __ATOMIC_SEQ_CST);return NULL ; }
@@ -4317,7 +4321,8 @@ static int ixgbevf_dev_xstats_get_names(__rte_unused struct rte_eth_dev *dev,if (link_up ==0 ) {if (ixgbe_get_media_type (hw) == ixgbe_media_type_fiber) {ixgbe_dev_wait_setup_link_complete (dev,0 ); -if (rte_atomic32_test_and_set (&ad->link_thread_running)) { +/* NOTE: review for potential ordering optimization */ +if (__atomic_test_and_set (&ad->link_thread_running, __ATOMIC_SEQ_CST)) {/* To avoid race condition between threads, set * the IXGBE_FLAG_NEED_LINK_CONFIG flag only * when there is no link thread running.
@@ -4330,7 +4335,8 @@ static int ixgbevf_dev_xstats_get_names(__rte_unused struct rte_eth_dev *dev, dev) < 0) {PMD_DRV_LOG (ERR ,"Create link thread failed!" ); -rte_atomic32_clear (&ad->link_thread_running); +/* NOTE: review for potential ordering optimization */ +__atomic_clear (&ad->link_thread_running, __ATOMIC_SEQ_CST); } }else { PMD_DRV_LOG(ERR,
drivers/net/ixgbe/ixgbe_ethdev.h +2 -1
@@ -6,6 +6,7 @@#define _IXGBE_ETHDEV_H_#include <stdint.h> +#include <stdbool.h> #include <sys/queue.h> #include "base/ixgbe_type.h"
@@ -510,7 +511,7 @@ struct ixgbe_adapter {*/ uint8_t pflink_fullchk;uint8_t mac_ctrl_frame_fwd; -rte_atomic32_t link_thread_running; +bool link_thread_running;pthread_t link_thread_tid; };
drivers/net/ixgbe/ixgbe_flow.c +0 -1
@@ -18,7 +18,6 @@#include <rte_log.h> #include <rte_debug.h> #include <rte_pci.h> -#include <rte_atomic.h> #include <rte_branch_prediction.h> #include <rte_memory.h> #include <rte_eal.h>
drivers/net/ixgbe/ixgbe_rxtx.c +0 -1
@@ -27,7 +27,6 @@#include <rte_eal.h> #include <rte_per_lcore.h> #include <rte_lcore.h> -#include <rte_atomic.h> #include <rte_branch_prediction.h> #include <rte_mempool.h> #include <rte_malloc.h>
Signature · 1 line
1.8.3.1
PATCH v4 5/6 net/null: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffReplace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
--- drivers/net/null/rte_eth_null.c | 28 ++++++++++++++++++---------- 1 file changed, 18 insertions(+), 10 deletions(-)
drivers/net/null/rte_eth_null.c +18 -10
@@ -37,8 +37,8 @@ struct null_queue {struct rte_mempool *mb_pool;struct rte_mbuf *dummy_packet; -rte_atomic64_t rx_pkts; -rte_atomic64_t tx_pkts; +uint64_t rx_pkts; +uint64_t tx_pkts; };struct pmd_options {
@@ -101,7 +101,8 @@ struct pmd_internals { bufs[i]->port = h->internals->port_id; } -rte_atomic64_add (&(h->rx_pkts), i); +/* NOTE: review for potential ordering optimization */ +__atomic_fetch_add (&h->rx_pkts, i, __ATOMIC_SEQ_CST);return i; }
@@ -128,7 +129,8 @@ struct pmd_internals { bufs[i]->port = h->internals->port_id; } -rte_atomic64_add (&(h->rx_pkts), i); +/* NOTE: review for potential ordering optimization */ +__atomic_fetch_add (&h->rx_pkts, i, __ATOMIC_SEQ_CST);return i; }
@@ -152,7 +154,8 @@ struct pmd_internals {for (i =0 ; i < nb_bufs; i++)rte_pktmbuf_free (bufs[i]); -rte_atomic64_add (&(h->tx_pkts), i); +/* NOTE: review for potential ordering optimization */ +__atomic_fetch_add (&h->tx_pkts, i, __ATOMIC_SEQ_CST);return i; }
@@ -174,7 +177,8 @@ struct pmd_internals {rte_pktmbuf_free (bufs[i]); } -rte_atomic64_add (&(h->tx_pkts), i); +/* NOTE: review for potential ordering optimization */ +__atomic_fetch_add (&h->tx_pkts, i, __ATOMIC_SEQ_CST);return i; }
@@ -316,8 +320,9 @@ struct pmd_internals {RTE_MIN (dev->data->nb_rx_queues, RTE_DIM(internal->rx_null_queues)));for (i =0 ; i < num_stats; i++) { +/* NOTE: review for atomic access */ igb_stats->q_ipackets[i] = - internal->rx_null_queues[i].rx_pkts.cnt; + internal->rx_null_queues[i].rx_pkts; rx_total += igb_stats->q_ipackets[i]; }
@@ -325,8 +330,9 @@ struct pmd_internals {RTE_MIN (dev->data->nb_tx_queues, RTE_DIM(internal->tx_null_queues)));for (i =0 ; i < num_stats; i++) { +/* NOTE: review for atomic access */ igb_stats->q_opackets[i] = - internal->tx_null_queues[i].tx_pkts.cnt; + internal->tx_null_queues[i].tx_pkts; tx_total += igb_stats->q_opackets[i]; }
@@ -347,9 +353,11 @@ struct pmd_internals { internal = dev->data->dev_private;for (i =0 ; i <RTE_DIM (internal->rx_null_queues); i++) - internal->rx_null_queues[i].rx_pkts.cnt =0 ; +/* NOTE: review for atomic access */ + internal->rx_null_queues[i].rx_pkts =0 ;for (i =0 ; i <RTE_DIM (internal->tx_null_queues); i++) - internal->tx_null_queues[i].tx_pkts.cnt =0 ; +/* NOTE: review for atomic access */ + internal->tx_null_queues[i].tx_pkts =0 ;return 0 ; }
Signature · 1 line
1.8.3.1
PATCH v4 6/6 net/ring: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffReplace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
--- drivers/net/ring/rte_eth_ring.c | 26 ++++++++++++++++---------- 1 file changed, 16 insertions(+), 10 deletions(-)
drivers/net/ring/rte_eth_ring.c +16 -10
@@ -44,8 +44,8 @@ enum dev_action {struct ring_queue {struct rte_ring *rng; -rte_atomic64_t rx_pkts; -rte_atomic64_t tx_pkts; +uint64_t rx_pkts; +uint64_t tx_pkts; };struct pmd_internals {
@@ -80,9 +80,10 @@ struct pmd_internals {const uint16_t nb_rx = (uint16_t )rte_ring_dequeue_burst (r->rng, ptrs, nb_bufs,NULL );if (r->rng->flags &RING_F_SC_DEQ ) - r->rx_pkts.cnt += nb_rx; + r->rx_pkts += nb_rx;else -rte_atomic64_add (&(r->rx_pkts), nb_rx); +/* NOTE: review for potential ordering optimization */ +__atomic_fetch_add (&r->rx_pkts, nb_rx, __ATOMIC_SEQ_CST);return nb_rx; }
@@ -94,9 +95,10 @@ struct pmd_internals {const uint16_t nb_tx = (uint16_t )rte_ring_enqueue_burst (r->rng, ptrs, nb_bufs,NULL );if (r->rng->flags &RING_F_SP_ENQ ) - r->tx_pkts.cnt += nb_tx; + r->tx_pkts += nb_tx;else -rte_atomic64_add (&(r->tx_pkts), nb_tx); +/* NOTE: review for potential ordering optimization */ +__atomic_fetch_add (&r->tx_pkts, nb_tx, __ATOMIC_SEQ_CST);return nb_tx; }
@@ -184,13 +186,15 @@ struct pmd_internals {for (i =0 ; i <RTE_ETHDEV_QUEUE_STAT_CNTRS && i < dev->data->nb_rx_queues; i++) { - stats->q_ipackets[i] = internal->rx_ring_queues[i].rx_pkts.cnt; +/* NOTE: review for atomic access */ + stats->q_ipackets[i] = internal->rx_ring_queues[i].rx_pkts; rx_total += stats->q_ipackets[i]; }for (i =0 ; i <RTE_ETHDEV_QUEUE_STAT_CNTRS && i < dev->data->nb_tx_queues; i++) { - stats->q_opackets[i] = internal->tx_ring_queues[i].tx_pkts.cnt; +/* NOTE: review for atomic access */ + stats->q_opackets[i] = internal->tx_ring_queues[i].tx_pkts; tx_total += stats->q_opackets[i]; }
@@ -207,9 +211,11 @@ struct pmd_internals {struct pmd_internals *internal = dev->data->dev_private;for (i =0 ; i < dev->data->nb_rx_queues; i++) - internal->rx_ring_queues[i].rx_pkts.cnt =0 ; +/* NOTE: review for atomic access */ + internal->rx_ring_queues[i].rx_pkts =0 ;for (i =0 ; i < dev->data->nb_tx_queues; i++) - internal->tx_ring_queues[i].tx_pkts.cnt =0 ; +/* NOTE: review for atomic access */ + internal->tx_ring_queues[i].tx_pkts =0 ;return 0 ; }
Signature · 1 line
1.8.3.1
Re: [PATCH v4 6/6] net/ring: replace rte atomics with GCC builtin atomics
Hi Tyler,
Few comments below.
Quoted 26 lines · Tyler Retzlaff
Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.Signed-off-by Tyler Retzlaff <[email protected]>Acked-by Morten Brørup <[email protected]>Acked-by Bruce Richardson <[email protected]>--- drivers/net/ring/rte_eth_ring.c | 26 ++++++++++++++++---------- 1 file changed, 16 insertions(+), 10 deletions(-)drivers/net/ring/rte_eth_ring.c +2
-2@@ -44,8 +44,8 @@ enum dev_action {struct ring_queue {struct rte_ring *rng; -rte_atomic64_t rx_pkts; -rte_atomic64_t tx_pkts; +uint64_t rx_pkts; +uint64_t tx_pkts; };struct pmd_internals {
drivers/net/ring/rte_eth_ring.c struct pmd_internals { · line 86
Quoted 9 lines
drivers/net/ring/rte_eth_ring.c +3
-2@@ -80,9 +80,10 @@ struct pmd_internals {const uint16_t nb_rx = (uint16_t )rte_ring_dequeue_burst (r->rng, ptrs, nb_bufs,NULL );if (r->rng->flags &RING_F_SC_DEQ ) - r->rx_pkts.cnt += nb_rx; + r->rx_pkts += nb_rx;else -rte_atomic64_add (&(r->rx_pkts), nb_rx); +/* NOTE: review for potential ordering optimization */ +__atomic_fetch_add (&r->rx_pkts, nb_rx, __ATOMIC_SEQ_CST);
We can use __ATOMIC_RELAXED here (and below too), since there is no ordering constraint. We only want statistics to be correct.
You can remove the other NOTEs from the patch.
Quoted 51 lines
drivers/net/ring/rte_eth_ring.c +11
-6return nb_rx; }@@ -94,9 +95,10 @@ struct pmd_internals {const uint16_t nb_tx = (uint16_t )rte_ring_enqueue_burst (r->rng, ptrs, nb_bufs,NULL );if (r->rng->flags &RING_F_SP_ENQ ) - r->tx_pkts.cnt += nb_tx; + r->tx_pkts += nb_tx;else -rte_atomic64_add (&(r->tx_pkts), nb_tx); +/* NOTE: review for potential ordering optimization */ +__atomic_fetch_add (&r->tx_pkts, nb_tx, __ATOMIC_SEQ_CST);return nb_tx; }@@ -184,13 +186,15 @@ struct pmd_internals {for (i =0 ; i <RTE_ETHDEV_QUEUE_STAT_CNTRS && i < dev->data->nb_rx_queues; i++) { - stats->q_ipackets[i] = internal->rx_ring_queues[i].rx_pkts.cnt; +/* NOTE: review for atomic access */ + stats->q_ipackets[i] = internal->rx_ring_queues[i].rx_pkts; rx_total += stats->q_ipackets[i]; }for (i =0 ; i <RTE_ETHDEV_QUEUE_STAT_CNTRS && i < dev->data->nb_tx_queues; i++) { - stats->q_opackets[i] = internal->tx_ring_queues[i].tx_pkts.cnt; +/* NOTE: review for atomic access */ + stats->q_opackets[i] = internal->tx_ring_queues[i].tx_pkts; tx_total += stats->q_opackets[i]; }@@ -207,9 +211,11 @@ struct pmd_internals {struct pmd_internals *internal = dev->data->dev_private;for (i =0 ; i < dev->data->nb_rx_queues; i++) - internal->rx_ring_queues[i].rx_pkts.cnt =0 ; +/* NOTE: review for atomic access */ + internal->rx_ring_queues[i].rx_pkts =0 ;for (i =0 ; i < dev->data->nb_tx_queues; i++) - internal->tx_ring_queues[i].tx_pkts.cnt =0 ; +/* NOTE: review for atomic access */ + internal->tx_ring_queues[i].tx_pkts =0 ;return 0 ; }Signature · 1 line
1.8.3.1
PATCH v5 1/6 stack: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffReplace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
--- lib/stack/rte_stack_lf_generic.h | 16 +++++++++------- 1 file changed, 9 insertions(+), 7 deletions(-)
lib/stack/rte_stack_lf_generic.h +9 -7
@@ -26,8 +26,8 @@* elements. If the mempool is near-empty to the point that this is a * concern, the user should consider increasing the mempool size. */ -return (unsigned int )rte_atomic64_read ((rte_atomic64_t *) - &s->stack_lf.used.len); +/* NOTE: review for potential ordering optimization */ +return __atomic_load_n (&s->stack_lf.used.len, __ATOMIC_SEQ_CST); }static __rte_always_inline void
@@ -67,8 +67,8 @@1 , __ATOMIC_RELEASE, __ATOMIC_RELAXED); }while (success ==0 ); - -rte_atomic64_add ((rte_atomic64_t *)&list->len, num); +/* NOTE: review for potential ordering optimization */ +__atomic_fetch_add (&list->len, num, __ATOMIC_SEQ_CST); } static __rte_always_inline struct rte_stack_lf_elem *
@@ -82,14 +82,16 @@/* Reserve num elements, if available */ while (1 ) { -uint64_t len =rte_atomic64_read ((rte_atomic64_t *)&list->len); +/* NOTE: review for potential ordering optimization */ +uint64_t len =__atomic_load_n (&list->len, __ATOMIC_SEQ_CST);/* Does the list contain enough elements? */ if (unlikely (len < num))return NULL ; -if (rte_atomic64_cmpset ((volatile uint64_t *)&list->len, - len, len - num)) +/* NOTE: review for potential ordering optimization */ +if (__atomic_compare_exchange_n (&list->len, &len, len - num, +0 , __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))break ; }
Signature · 1 line
1.8.3.1
PATCH v5 3/6 net/ice: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffReplace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
--- drivers/net/ice/ice_dcf.c | 1 - drivers/net/ice/ice_dcf_ethdev.c | 1 - drivers/net/ice/ice_ethdev.c | 12 ++++++++---- 3 files changed, 8 insertions(+), 6 deletions(-)
drivers/net/ice/ice_dcf.c +0 -1
@@ -14,7 +14,6 @@#include <rte_common.h> #include <rte_pci.h> -#include <rte_atomic.h> #include <rte_eal.h> #include <rte_ether.h> #include <ethdev_driver.h>
drivers/net/ice/ice_dcf_ethdev.c +0 -1
@@ -11,7 +11,6 @@#include <rte_interrupts.h> #include <rte_debug.h> #include <rte_pci.h> -#include <rte_atomic.h> #include <rte_eal.h> #include <rte_ether.h> #include <ethdev_pci.h>
drivers/net/ice/ice_ethdev.c +8 -4
@@ -3927,8 +3927,10 @@ static int ice_init_rss(struct ice_pf *pf)struct rte_eth_link *dst = link;struct rte_eth_link *src = &dev->data->dev_link; -if (rte_atomic64_cmpset ((uint64_t *)dst, *(uint64_t *)dst, - *(uint64_t *)src) ==0 ) +/* NOTE: review for potential ordering optimization */ +if (!__atomic_compare_exchange_n ((uint64_t *)dst, + (uint64_t *)dst, *(uint64_t *)src,0 , + __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))return -1 ;return 0 ;
@@ -3941,8 +3943,10 @@ static int ice_init_rss(struct ice_pf *pf)struct rte_eth_link *dst = &dev->data->dev_link;struct rte_eth_link *src = link; -if (rte_atomic64_cmpset ((uint64_t *)dst, *(uint64_t *)dst, - *(uint64_t *)src) ==0 ) +/* NOTE: review for potential ordering optimization */ +if (!__atomic_compare_exchange_n ((uint64_t *)dst, + (uint64_t *)dst, *(uint64_t *)src,0 , + __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))return -1 ;return 0 ;
Signature · 1 line
1.8.3.1
v5 cover replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffReplace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
This series covers the libraries and drivers that are built on Windows.
The code has be converted to use the __atomic builtins but there are additional during conversion I notice that there may be some issues that need to be addressed.
I'll comment in the patches where my concerns are so the maintainers may comment.
v5:
* use relaxed ordering for counter increments in net/ring patch
* remove note comments from net/ring patch
v4:
* drop patch for lib/ring it will be provided by ARM / Honnappa
* rebase for changes in dma/idxd merge
* adapt __atomic_fetch_sub(...) - 1 == 0 to be (__atomic_fetch_sub(...) == 1) as per feedback.
* drop one /* NOTE: review for potential ordering optimization */ since the note reference non-critical to perf control path.
note:
Remainder of the NOTE comments have been retained since there seems to be no consensus but stronger opinion/argument to keep expressed. while I generally agree that changes should not include ``TODO'' style comments I also agree that without these comments in your face people are very unlikely to feel compelled to make the review they are trying to solicit without them. if it is absolute that the series won't be merged with them then I will remove them, but please be explicit soon.
v3:
* style, don't use c99 comments
v2:
* comment code where optimizations may be possible now that memory order can be specified.
* comment code where operations should potentially be atomic so that maintainers can review.
* change a couple of variables labeled as counters to be unsigned.
Tyler Retzlaff (6):
stack: replace rte atomics with GCC builtin atomics dma/idxd: replace rte atomics with GCC builtin atomics net/ice: replace rte atomics with GCC builtin atomics net/ixgbe: replace rte atomics with GCC builtin atomics net/null: replace rte atomics with GCC builtin atomics net/ring: replace rte atomics with GCC builtin atomics
drivers/dma/idxd/idxd_internal.h | 3 +-- drivers/dma/idxd/idxd_pci.c | 11 ++++++----- drivers/net/ice/ice_dcf.c | 1 - drivers/net/ice/ice_dcf_ethdev.c | 1 - drivers/net/ice/ice_ethdev.c | 12 ++++++++---- drivers/net/ixgbe/ixgbe_bypass.c | 1 - drivers/net/ixgbe/ixgbe_ethdev.c | 18 ++++++++++++------ drivers/net/ixgbe/ixgbe_ethdev.h | 3 ++- drivers/net/ixgbe/ixgbe_flow.c | 1 - drivers/net/ixgbe/ixgbe_rxtx.c | 1 - drivers/net/null/rte_eth_null.c | 28 ++++++++++++++++++---------- drivers/net/ring/rte_eth_ring.c | 20 ++++++++++---------- lib/stack/rte_stack_lf_generic.h | 16 +++++++++------- 13 files changed, 66 insertions(+), 50 deletions(-)
Signature · 1 line
1.8.3.1
PATCH v5 4/6 net/ixgbe: replace rte atomics with GCC builtin atomics
Replace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
--- drivers/net/ixgbe/ixgbe_bypass.c | 1 - drivers/net/ixgbe/ixgbe_ethdev.c | 18 ++++++++++++------ drivers/net/ixgbe/ixgbe_ethdev.h | 3 ++- drivers/net/ixgbe/ixgbe_flow.c | 1 - drivers/net/ixgbe/ixgbe_rxtx.c | 1 - 5 files changed, 14 insertions(+), 10 deletions(-)
drivers/net/ixgbe/ixgbe_bypass.c +0 -1
@@ -3,7 +3,6 @@*/ #include <time.h> -#include <rte_atomic.h> #include <ethdev_driver.h> #include "ixgbe_ethdev.h" #include "ixgbe_bypass_api.h"
drivers/net/ixgbe/ixgbe_ethdev.c +12 -6
@@ -1127,7 +1127,8 @@ struct rte_ixgbe_xstats_name_off {return 0 ; } -rte_atomic32_clear (&ad->link_thread_running); +/* NOTE: review for potential ordering optimization */ +__atomic_clear (&ad->link_thread_running, __ATOMIC_SEQ_CST);ixgbe_parse_devargs (eth_dev->data->dev_private, pci_dev->device.devargs);rte_eth_copy_pci_info (eth_dev, pci_dev);
@@ -1625,7 +1626,8 @@ static int ixgbe_l2_tn_filter_init(struct rte_eth_dev *eth_dev)return 0 ; } -rte_atomic32_clear (&ad->link_thread_running); +/* NOTE: review for potential ordering optimization */ +__atomic_clear (&ad->link_thread_running, __ATOMIC_SEQ_CST);ixgbevf_parse_devargs (eth_dev->data->dev_private, pci_dev->device.devargs);
@@ -4186,7 +4188,8 @@ static int ixgbevf_dev_xstats_get_names(__rte_unused struct rte_eth_dev *dev,struct ixgbe_adapter *ad = dev->data->dev_private;uint32_t timeout = timeout_ms ? timeout_ms :WARNING_TIMEOUT ; -while (rte_atomic32_read (&ad->link_thread_running)) { +/* NOTE: review for potential ordering optimization */ +while (__atomic_load_n (&ad->link_thread_running, __ATOMIC_SEQ_CST)) {msec_delay (1 ); timeout--;
@@ -4222,7 +4225,8 @@ static int ixgbevf_dev_xstats_get_names(__rte_unused struct rte_eth_dev *dev,ixgbe_setup_link (hw, speed, true); intr->flags &= ~IXGBE_FLAG_NEED_LINK_CONFIG ; -rte_atomic32_clear (&ad->link_thread_running); +/* NOTE: review for potential ordering optimization */ +__atomic_clear (&ad->link_thread_running, __ATOMIC_SEQ_CST);return NULL ; }
@@ -4317,7 +4321,8 @@ static int ixgbevf_dev_xstats_get_names(__rte_unused struct rte_eth_dev *dev,if (link_up ==0 ) {if (ixgbe_get_media_type (hw) == ixgbe_media_type_fiber) {ixgbe_dev_wait_setup_link_complete (dev,0 ); -if (rte_atomic32_test_and_set (&ad->link_thread_running)) { +/* NOTE: review for potential ordering optimization */ +if (__atomic_test_and_set (&ad->link_thread_running, __ATOMIC_SEQ_CST)) {/* To avoid race condition between threads, set * the IXGBE_FLAG_NEED_LINK_CONFIG flag only * when there is no link thread running.
@@ -4330,7 +4335,8 @@ static int ixgbevf_dev_xstats_get_names(__rte_unused struct rte_eth_dev *dev, dev) < 0) {PMD_DRV_LOG (ERR ,"Create link thread failed!" ); -rte_atomic32_clear (&ad->link_thread_running); +/* NOTE: review for potential ordering optimization */ +__atomic_clear (&ad->link_thread_running, __ATOMIC_SEQ_CST); } }else { PMD_DRV_LOG(ERR,
drivers/net/ixgbe/ixgbe_ethdev.h +2 -1
@@ -6,6 +6,7 @@#define _IXGBE_ETHDEV_H_#include <stdint.h> +#include <stdbool.h> #include <sys/queue.h> #include "base/ixgbe_type.h"
@@ -510,7 +511,7 @@ struct ixgbe_adapter {*/ uint8_t pflink_fullchk;uint8_t mac_ctrl_frame_fwd; -rte_atomic32_t link_thread_running; +bool link_thread_running;pthread_t link_thread_tid; };
drivers/net/ixgbe/ixgbe_flow.c +0 -1
@@ -18,7 +18,6 @@#include <rte_log.h> #include <rte_debug.h> #include <rte_pci.h> -#include <rte_atomic.h> #include <rte_branch_prediction.h> #include <rte_memory.h> #include <rte_eal.h>
drivers/net/ixgbe/ixgbe_rxtx.c +0 -1
@@ -27,7 +27,6 @@#include <rte_eal.h> #include <rte_per_lcore.h> #include <rte_lcore.h> -#include <rte_atomic.h> #include <rte_branch_prediction.h> #include <rte_mempool.h> #include <rte_malloc.h>
Signature · 1 line
1.8.3.1
PATCH v5 2/6 dma/idxd: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffReplace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
--- drivers/dma/idxd/idxd_internal.h | 3 +-- drivers/dma/idxd/idxd_pci.c | 11 ++++++----- 2 files changed, 7 insertions(+), 7 deletions(-)
drivers/dma/idxd/idxd_internal.h +1 -2
@@ -7,7 +7,6 @@#include <rte_dmadev_pmd.h> #include <rte_spinlock.h> -#include <rte_atomic.h> #include "idxd_hw_defs.h"
@@ -34,7 +33,7 @@ struct idxd_pci_common {rte_spinlock_t lk;uint8_t wq_cfg_sz; -rte_atomic16_t ref_count; +uint16_t ref_count;volatile struct rte_idxd_bar0 *regs;volatile uint32_t *wq_regs_base;volatile struct rte_idxd_grpcfg *grp_regs;
drivers/dma/idxd/idxd_pci.c +6 -5
@@ -6,7 +6,6 @@#include <rte_devargs.h> #include <rte_dmadev_pmd.h> #include <rte_malloc.h> -#include <rte_atomic.h> #include "idxd_internal.h"
@@ -136,7 +135,8 @@/* if this is the last WQ on the device, disable the device and free * the PCI struct */ - is_last_wq =rte_atomic16_dec_and_test (&idxd->u.pci->ref_count); +/* NOTE: review for potential ordering optimization */ + is_last_wq = (__atomic_fetch_sub (&idxd->u.pci->ref_count,1 , __ATOMIC_SEQ_CST) ==1 );if (is_last_wq) {/* disable the device */ err_code =idxd_pci_dev_command (idxd, idxd_disable_dev);
@@ -322,8 +322,9 @@return ret; } qid =rte_dma_get_dev_id_by_name (qname); - max_qid =rte_atomic16_read ( - &((struct idxd_dmadev *)rte_dma_fp_objs[qid].dev_private)->u.pci->ref_count); + max_qid =__atomic_load_n ( + &((struct idxd_dmadev *)rte_dma_fp_objs[qid].dev_private)->u.pci->ref_count, + __ATOMIC_SEQ_CST);/* we have queue 0 done, now configure the rest of the queues */ for (qid =1 ; qid < max_qid; qid++) {
@@ -380,7 +381,7 @@free (idxd.u.pci);return ret; } -rte_atomic16_inc (&idxd.u.pci->ref_count); +__atomic_fetch_add (&idxd.u.pci->ref_count,1 , __ATOMIC_SEQ_CST); }return 0 ;
Signature · 1 line
1.8.3.1
PATCH v5 5/6 net/null: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffReplace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
--- drivers/net/null/rte_eth_null.c | 28 ++++++++++++++++++---------- 1 file changed, 18 insertions(+), 10 deletions(-)
drivers/net/null/rte_eth_null.c +18 -10
@@ -37,8 +37,8 @@ struct null_queue {struct rte_mempool *mb_pool;struct rte_mbuf *dummy_packet; -rte_atomic64_t rx_pkts; -rte_atomic64_t tx_pkts; +uint64_t rx_pkts; +uint64_t tx_pkts; };struct pmd_options {
@@ -101,7 +101,8 @@ struct pmd_internals { bufs[i]->port = h->internals->port_id; } -rte_atomic64_add (&(h->rx_pkts), i); +/* NOTE: review for potential ordering optimization */ +__atomic_fetch_add (&h->rx_pkts, i, __ATOMIC_SEQ_CST);return i; }
@@ -128,7 +129,8 @@ struct pmd_internals { bufs[i]->port = h->internals->port_id; } -rte_atomic64_add (&(h->rx_pkts), i); +/* NOTE: review for potential ordering optimization */ +__atomic_fetch_add (&h->rx_pkts, i, __ATOMIC_SEQ_CST);return i; }
@@ -152,7 +154,8 @@ struct pmd_internals {for (i =0 ; i < nb_bufs; i++)rte_pktmbuf_free (bufs[i]); -rte_atomic64_add (&(h->tx_pkts), i); +/* NOTE: review for potential ordering optimization */ +__atomic_fetch_add (&h->tx_pkts, i, __ATOMIC_SEQ_CST);return i; }
@@ -174,7 +177,8 @@ struct pmd_internals {rte_pktmbuf_free (bufs[i]); } -rte_atomic64_add (&(h->tx_pkts), i); +/* NOTE: review for potential ordering optimization */ +__atomic_fetch_add (&h->tx_pkts, i, __ATOMIC_SEQ_CST);return i; }
@@ -316,8 +320,9 @@ struct pmd_internals {RTE_MIN (dev->data->nb_rx_queues, RTE_DIM(internal->rx_null_queues)));for (i =0 ; i < num_stats; i++) { +/* NOTE: review for atomic access */ igb_stats->q_ipackets[i] = - internal->rx_null_queues[i].rx_pkts.cnt; + internal->rx_null_queues[i].rx_pkts; rx_total += igb_stats->q_ipackets[i]; }
@@ -325,8 +330,9 @@ struct pmd_internals {RTE_MIN (dev->data->nb_tx_queues, RTE_DIM(internal->tx_null_queues)));for (i =0 ; i < num_stats; i++) { +/* NOTE: review for atomic access */ igb_stats->q_opackets[i] = - internal->tx_null_queues[i].tx_pkts.cnt; + internal->tx_null_queues[i].tx_pkts; tx_total += igb_stats->q_opackets[i]; }
@@ -347,9 +353,11 @@ struct pmd_internals { internal = dev->data->dev_private;for (i =0 ; i <RTE_DIM (internal->rx_null_queues); i++) - internal->rx_null_queues[i].rx_pkts.cnt =0 ; +/* NOTE: review for atomic access */ + internal->rx_null_queues[i].rx_pkts =0 ;for (i =0 ; i <RTE_DIM (internal->tx_null_queues); i++) - internal->tx_null_queues[i].tx_pkts.cnt =0 ; +/* NOTE: review for atomic access */ + internal->tx_null_queues[i].tx_pkts =0 ;return 0 ; }
Signature · 1 line
1.8.3.1
PATCH v5 6/6 net/ring: replace rte atomics with GCC builtin atomics
In reply to Tyler RetzlaffReplace the use of rte_atomic.h types and functions, instead use GCC supplied C++11 memory model builtins.
--- drivers/net/ring/rte_eth_ring.c | 20 ++++++++++---------- 1 file changed, 10 insertions(+), 10 deletions(-)
drivers/net/ring/rte_eth_ring.c +10 -10
@@ -44,8 +44,8 @@ enum dev_action {struct ring_queue {struct rte_ring *rng; -rte_atomic64_t rx_pkts; -rte_atomic64_t tx_pkts; +uint64_t rx_pkts; +uint64_t tx_pkts; };struct pmd_internals {
@@ -80,9 +80,9 @@ struct pmd_internals {const uint16_t nb_rx = (uint16_t )rte_ring_dequeue_burst (r->rng, ptrs, nb_bufs,NULL );if (r->rng->flags &RING_F_SC_DEQ ) - r->rx_pkts.cnt += nb_rx; + r->rx_pkts += nb_rx;else -rte_atomic64_add (&(r->rx_pkts), nb_rx); +__atomic_fetch_add (&r->rx_pkts, nb_rx, __ATOMIC_RELAXED);return nb_rx; }
@@ -94,9 +94,9 @@ struct pmd_internals {const uint16_t nb_tx = (uint16_t )rte_ring_enqueue_burst (r->rng, ptrs, nb_bufs,NULL );if (r->rng->flags &RING_F_SP_ENQ ) - r->tx_pkts.cnt += nb_tx; + r->tx_pkts += nb_tx;else -rte_atomic64_add (&(r->tx_pkts), nb_tx); +__atomic_fetch_add (&r->tx_pkts, nb_tx, __ATOMIC_RELAXED);return nb_tx; }
@@ -184,13 +184,13 @@ struct pmd_internals {for (i =0 ; i <RTE_ETHDEV_QUEUE_STAT_CNTRS && i < dev->data->nb_rx_queues; i++) { - stats->q_ipackets[i] = internal->rx_ring_queues[i].rx_pkts.cnt; + stats->q_ipackets[i] = internal->rx_ring_queues[i].rx_pkts; rx_total += stats->q_ipackets[i]; }for (i =0 ; i <RTE_ETHDEV_QUEUE_STAT_CNTRS && i < dev->data->nb_tx_queues; i++) { - stats->q_opackets[i] = internal->tx_ring_queues[i].tx_pkts.cnt; + stats->q_opackets[i] = internal->tx_ring_queues[i].tx_pkts; tx_total += stats->q_opackets[i]; }
@@ -207,9 +207,9 @@ struct pmd_internals {struct pmd_internals *internal = dev->data->dev_private;for (i =0 ; i < dev->data->nb_rx_queues; i++) - internal->rx_ring_queues[i].rx_pkts.cnt =0 ; + internal->rx_ring_queues[i].rx_pkts =0 ;for (i =0 ; i < dev->data->nb_tx_queues; i++) - internal->tx_ring_queues[i].tx_pkts.cnt =0 ; + internal->tx_ring_queues[i].tx_pkts =0 ;return 0 ; }
Signature · 1 line
1.8.3.1
Quoted 55 lines · Tyler Retzlaff
Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.This series covers the libraries and drivers that are built on Windows.
The code has be converted to use the
__atomicbuiltins but there are additional during conversion I notice that there may be some issues that need to be addressed.I'll comment in the patches where my concerns are so the maintainers may comment.
v5:
* use relaxed ordering for counter increments in net/ring patch
* remove note comments from net/ring patchv4:
* drop patch for lib/ring it will be provided by ARM / Honnappa
* rebase for changes in dma/idxd merge
* adapt__atomic_fetch_sub(...)- 1 == 0 to be (__atomic_fetch_sub(...)== 1) as per feedback.
* drop one /* NOTE: review for potential ordering optimization */ since the note reference non-critical to perf control path.note:
Remainder of the NOTE comments have been retained since there seems to be no consensus but stronger opinion/argument to keep expressed. while I generally agree that changes should not include ``TODO'' style comments I also agree that without these comments in your face people are very unlikely to feel compelled to make the review they are trying to solicit without them. if it is absolute that the series won't be merged with them then I will remove them, but please be explicit soon.
v3:
* style, don't use c99 commentsv2:
* comment code where optimizations may be possible now that memory order can be specified.
* comment code where operations should potentially be atomic so that maintainers can review.
* change a couple of variables labeled as counters to be unsigned.Tyler Retzlaff (6):
stack: replace rte atomics with GCC builtin atomics dma/idxd: replace rte atomics with GCC builtin atomics net/ice: replace rte atomics with GCC builtin atomics net/ixgbe: replace rte atomics with GCC builtin atomics net/null: replace rte atomics with GCC builtin atomics net/ring: replace rte atomics with GCC builtin atomics
Quoted 14 lines
drivers/dma/idxd/idxd_internal.h | 3 +-- drivers/dma/idxd/idxd_pci.c | 11 ++++++----- drivers/net/ice/ice_dcf.c | 1 - drivers/net/ice/ice_dcf_ethdev.c | 1 - drivers/net/ice/ice_ethdev.c | 12 ++++++++---- drivers/net/ixgbe/ixgbe_bypass.c | 1 - drivers/net/ixgbe/ixgbe_ethdev.c | 18 ++++++++++++------ drivers/net/ixgbe/ixgbe_ethdev.h | 3 ++- drivers/net/ixgbe/ixgbe_flow.c | 1 - drivers/net/ixgbe/ixgbe_rxtx.c | 1 - drivers/net/null/rte_eth_null.c | 28 ++++++++++++++++++---------- drivers/net/ring/rte_eth_ring.c | 20 ++++++++++---------- lib/stack/rte_stack_lf_generic.h | 16 +++++++++------- 13 files changed, 66 insertions(+), 50 deletions(-)
I am not really enthousiastic about those NOTE:.
I would prefer we get an explicit go/nogo from each maintainers, but this did not happen.
I think that this indicates that those NOTE: will rot in the code now.
Thomas proposed to track those NOTE: in the release announce mail and that we ping maintainers regularly.
Let's see how it goes.
I am merging this series so we can progress on the $SUBJECT. Series applied, thanks.
Tyler, about the patch on the ring library that was dropped by got no viable alternative, I'll wait for a decision from ARM and you.
Signature · 1 line
David Marchand
Quoted 72 lines · David Marchand
Quoted 55 lines · Tyler Retzlaff
Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.This series covers the libraries and drivers that are built on Windows.
The code has be converted to use the
__atomicbuiltins but there are additional during conversion I notice that there may be some issues that need to be addressed.I'll comment in the patches where my concerns are so the maintainers may comment.
v5:
* use relaxed ordering for counter increments in net/ring patch
* remove note comments from net/ring patchv4:
* drop patch for lib/ring it will be provided by ARM / Honnappa
* rebase for changes in dma/idxd merge
* adapt__atomic_fetch_sub(...)- 1 == 0 to be (__atomic_fetch_sub(...)== 1) as per feedback.
* drop one /* NOTE: review for potential ordering optimization */ since the note reference non-critical to perf control path.note:
Remainder of the NOTE comments have been retained since there seems to be no consensus but stronger opinion/argument to keep expressed. while I generally agree that changes should not include ``TODO'' style comments I also agree that without these comments in your face people are very unlikely to feel compelled to make the review they are trying to solicit without them. if it is absolute that the series won't be merged with them then I will remove them, but please be explicit soon.
v3:
* style, don't use c99 commentsv2:
* comment code where optimizations may be possible now that memory order can be specified.
* comment code where operations should potentially be atomic so that maintainers can review.
* change a couple of variables labeled as counters to be unsigned.Tyler Retzlaff (6):
stack: replace rte atomics with GCC builtin atomics dma/idxd: replace rte atomics with GCC builtin atomics net/ice: replace rte atomics with GCC builtin atomics net/ixgbe: replace rte atomics with GCC builtin atomics net/null: replace rte atomics with GCC builtin atomics net/ring: replace rte atomics with GCC builtin atomicsQuoted 14 lines
drivers/dma/idxd/idxd_internal.h | 3 +-- drivers/dma/idxd/idxd_pci.c | 11 ++++++----- drivers/net/ice/ice_dcf.c | 1 - drivers/net/ice/ice_dcf_ethdev.c | 1 - drivers/net/ice/ice_ethdev.c | 12 ++++++++---- drivers/net/ixgbe/ixgbe_bypass.c | 1 - drivers/net/ixgbe/ixgbe_ethdev.c | 18 ++++++++++++------ drivers/net/ixgbe/ixgbe_ethdev.h | 3 ++- drivers/net/ixgbe/ixgbe_flow.c | 1 - drivers/net/ixgbe/ixgbe_rxtx.c | 1 - drivers/net/null/rte_eth_null.c | 28 ++++++++++++++++++---------- drivers/net/ring/rte_eth_ring.c | 20 ++++++++++---------- lib/stack/rte_stack_lf_generic.h | 16 +++++++++------- 13 files changed, 66 insertions(+), 50 deletions(-)
Quoted 8 lines
I am not really enthousiastic about those NOTE:.
I would prefer we get an explicit go/nogo from each maintainers, but this did not happen.
I think that this indicates that those NOTE: will rot in the code now.Thomas proposed to track those NOTE: in the release announce mail and that we ping maintainers regularly.
Let's see how it goes.
Let's leave it for one release cycle, if with the the announce mail maintainers take no action within that time I'll commit to going through and cleaning them out before 23.11 rc1.
Quoted 3 lines
I am merging this series so we can progress on the $SUBJECT. Series applied, thanks.
Thanks David, this will allow forward progress.
Quoted 4 lines
Tyler, about the patch on the ring library that was dropped by got no viable alternative, I'll wait for a decision from ARM and you.
I'll wait for Honnappa to follow up and we'll decide what to do when he does.
Quoted 3 lines
Signature · 2 lines
-- David Marchand
I want to report a possible regression from this patch series seen from CI testing on our Intel 82599ES 10G NIC, which we failed to report to patchwork when this initially went under CI due to a bug in our Jenkins reporting scripts. Use of the ixgbe driver appears to be affected. Tyler I apologize for the issues seen with reporting. We've made some temporary changes to avoid this happening again, and are currently reworking our reporting process entirely to provide greater reliability.
Here is a DTS snippet showing the issue, and the full log for the failing virtio_smoke test can be downloaded here:
https://dpdkdashboard.iol.unh.edu/results/dashboard/patchsets/26560/
06/06/2023 18:22:58 TestVirtioSmoke: Start send packets and
verify
06/06/2023 18:22:58 tester: ifconfig enp134s0f0 mtu
9000
06/06/2023 18:22:58 tester:
06/06/2023 18:42:59 TestVirtioSmoke: Test Casetest_virtio_pvp Result FAILED: TIMEOUT on port start 0
06/06/2023 18:42:59 TestVirtioSmoke: port start 0
ixgbe_dev_wait_setup_link_complete(): IXGBE link thread not complete too long time!ixgbe_dev_wait_setup_link_complete(): IXGBE link thread not complete too long time!ixgbe_dev_wait_setup_link_complete(): IXGBE link thread not complete too long time!
We initially took this Intel10G testing offline to investigate as we thought it was a lab infra failure. Obviously that wasn't the case, so ideally we will bring this back online when appropriate. But, I don't want to do so right now and start failing everyone's patchseries which are obviously unrelated to this. Comments on this are welcome, otherwise of course I will just return this test coverage to our CI when the state of the git tree allows for it.
Apologies for the missing report and the timeline on this. We are taking action to deliver results more reliably going forward.
Quoted 106 lines · Tyler Retzlaff
Quoted 19 lines · David Marchand
Quoted 17 lines · Tyler Retzlaff
Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.This series covers the libraries and drivers that are built on Windows.
The code has be converted to use the
__atomicbuiltins but there are additional during conversion I notice that there may be some issues that need to be addressed.I'll comment in the patches where my concerns are so the maintainers may comment.
v5:
* use relaxed ordering for counter increments in net/ring patch
* remove note comments from net/ring patchQuoted 5 lines
Quoted 5 lines
v4:
* drop patch for lib/ring it will be provided by ARM / Honnappa
* rebase for changes in dma/idxd merge
* adapt__atomic_fetch_sub(...)- 1 == 0 to be(
__atomic_fetch_sub(...)== 1)Quoted 2 lines
Quoted 2 lines
as per feedback.
* drop one /* NOTE: review for potential ordering optimization */since
Quoted 46 lines
Quoted 31 lines
the note reference non-critical to perf control path.
note:
Remainder of the NOTE comments have been retained since there seems to be no consensus but stronger opinion/argument to keep expressed. while I generally agree that changes should not include ``TODO'' style comments I also agree that without these comments in your face people are very unlikely to feel compelled to make the review they are trying to solicit without them. if it is absolute that the series won't be merged with them then I will remove them, but please be explicit soon.
v3:
* style, don't use c99 commentsv2:
* comment code where optimizations may be possible now that memory order can be specified.
* comment code where operations should potentially be atomic so that maintainers can review.
* change a couple of variables labeled as counters to be unsigned.Tyler Retzlaff (6):
stack: replace rte atomics with GCC builtin atomics dma/idxd: replace rte atomics with GCC builtin atomics net/ice: replace rte atomics with GCC builtin atomics net/ixgbe: replace rte atomics with GCC builtin atomics net/null: replace rte atomics with GCC builtin atomics net/ring: replace rte atomics with GCC builtin atomicsQuoted 14 lines
drivers/dma/idxd/idxd_internal.h | 3 +-- drivers/dma/idxd/idxd_pci.c | 11 ++++++----- drivers/net/ice/ice_dcf.c | 1 - drivers/net/ice/ice_dcf_ethdev.c | 1 - drivers/net/ice/ice_ethdev.c | 12 ++++++++---- drivers/net/ixgbe/ixgbe_bypass.c | 1 - drivers/net/ixgbe/ixgbe_ethdev.c | 18 ++++++++++++------ drivers/net/ixgbe/ixgbe_ethdev.h | 3 ++- drivers/net/ixgbe/ixgbe_flow.c | 1 - drivers/net/ixgbe/ixgbe_rxtx.c | 1 - drivers/net/null/rte_eth_null.c | 28 ++++++++++++++++++---------- drivers/net/ring/rte_eth_ring.c | 20 ++++++++++---------- lib/stack/rte_stack_lf_generic.h | 16 +++++++++------- 13 files changed, 66 insertions(+), 50 deletions(-)Quoted 8 lines
I am not really enthousiastic about those NOTE:.
I would prefer we get an explicit go/nogo from each maintainers, but this did not happen.
I think that this indicates that those NOTE: will rot in the code now.Thomas proposed to track those NOTE: in the release announce mail and that we ping maintainers regularly.
Let's see how it goes.Let's leave it for one release cycle, if with the the announce mail maintainers take no action within that time I'll commit to going through and cleaning them out before 23.11 rc1.
Quoted 3 lines
I am merging this series so we can progress on the $SUBJECT. Series applied, thanks.
Thanks David, this will allow forward progress.
Quoted 4 lines
Tyler, about the patch on the ring library that was dropped by got no viable alternative, I'll wait for a decision from ARM and you.
I'll wait for Honnappa to follow up and we'll decide what to do when he does.
Quoted 3 lines
--
David Marchand
Signature · 9 lines
Patrick Robb Technical Service Manager UNH InterOperability Laboratory 21 Madbury Rd, Suite 100, Durham, NH 03824 www.iol.unh.edu
Hello Patrick,
Quoted 11 lines · Patrick Robb
I want to report a possible regression from this patch series seen from CI testing on our Intel 82599ES 10G NIC, which we failed to report to patchwork when this initially went under CI due to a bug in our Jenkins reporting scripts. Use of the ixgbe driver appears to be affected. Tyler I apologize for the issues seen with reporting. We've made some temporary changes to avoid this happening again, and are currently reworking our reporting process entirely to provide greater reliability.
Here is a DTS snippet showing the issue, and the full log for the failing
virtio_smoketest can be downloaded here: https://dpdkdashboard.iol.unh.edu/results/dashboard/patchsets/26560/06/06/2023 18:22:58 TestVirtioSmoke: Start send packets and verify
06/06/2023 18:22:58 tester: ifconfig enp134s0f0 mtu 9000
06/06/2023 18:22:58 tester:
06/06/2023 18:42:59 TestVirtioSmoke: Test Casetest_virtio_pvpResult FAILED: TIMEOUT on port start 0
06/06/2023 18:42:59 TestVirtioSmoke: port start 0
Quoted 7 lines
ixgbe_dev_wait_setup_link_complete(): IXGBE link thread not complete too long time!ixgbe_dev_wait_setup_link_complete(): IXGBE link thread not complete too long time!ixgbe_dev_wait_setup_link_complete(): IXGBE link thread not complete too long time!We initially took this Intel10G testing offline to investigate as we thought it was a lab infra failure. Obviously that wasn't the case, so ideally we will bring this back online when appropriate. But, I don't want to do so right now and start failing everyone's patchseries which are obviously unrelated to this. Comments on this are welcome, otherwise of course I will just return this test coverage to our CI when the state of the git tree allows for it.
Apologies for the missing report and the timeline on this. We are taking action to deliver results more reliably going forward.
(reduced the cc list a bit)
This is probably the same issue than what was reported by Intel
validation: https://bugs.dpdk.org/show_bug.cgi?id=1249
A fix has been merged in next-net-intel, it will reach the main repo soon.
https://git.dpdk.org/next/dpdk-next-net-intel/commit/?id=fe4ce0aee766969a0e27fe28ced8ee7c761a2c4e
Signature · 1 line
David Marchand
Hi Patrick,
I will take a look at this as a priority asap.
Thanks for bringing it to my attention.
Quoted 162 lines · Patrick Robb
I want to report a possible regression from this patch series seen from CI testing on our Intel 82599ES 10G NIC, which we failed to report to patchwork when this initially went under CI due to a bug in our Jenkins reporting scripts. Use of the ixgbe driver appears to be affected. Tyler I apologize for the issues seen with reporting. We've made some temporary changes to avoid this happening again, and are currently reworking our reporting process entirely to provide greater reliability.
Here is a DTS snippet showing the issue, and the full log for the failing
virtio_smoketest can be downloaded here:
https://dpdkdashboard.iol.unh.edu/results/dashboard/patchsets/26560/06/06/2023 18:22:58 TestVirtioSmoke: Start send packets and
verify
06/06/2023 18:22:58 tester: ifconfig enp134s0f0 mtu
9000
06/06/2023 18:22:58 tester:
06/06/2023 18:42:59 TestVirtioSmoke: Test Casetest_virtio_pvpResult FAILED: TIMEOUT on port start 0
06/06/2023 18:42:59 TestVirtioSmoke: port start 0
ixgbe_dev_wait_setup_link_complete(): IXGBE link thread not complete too long time!ixgbe_dev_wait_setup_link_complete(): IXGBE link thread not complete too long time!ixgbe_dev_wait_setup_link_complete(): IXGBE link thread not complete too long time!We initially took this Intel10G testing offline to investigate as we thought it was a lab infra failure. Obviously that wasn't the case, so ideally we will bring this back online when appropriate. But, I don't want to do so right now and start failing everyone's patchseries which are obviously unrelated to this. Comments on this are welcome, otherwise of course I will just return this test coverage to our CI when the state of the git tree allows for it.
Apologies for the missing report and the timeline on this. We are taking action to deliver results more reliably going forward.
Quoted 106 lines · Tyler Retzlaff
Quoted 19 lines · David Marchand
Quoted 17 lines · Tyler Retzlaff
Replace the use of
rte_atomic.htypes and functions, instead use GCC supplied C++11 memory model builtins.This series covers the libraries and drivers that are built on Windows.
The code has be converted to use the
__atomicbuiltins but there are additional during conversion I notice that there may be some issues that need to be addressed.I'll comment in the patches where my concerns are so the maintainers may comment.
v5:
* use relaxed ordering for counter increments in net/ring patch
* remove note comments from net/ring patchQuoted 5 lines
Quoted 5 lines
v4:
* drop patch for lib/ring it will be provided by ARM / Honnappa
* rebase for changes in dma/idxd merge
* adapt__atomic_fetch_sub(...)- 1 == 0 to be(
__atomic_fetch_sub(...)== 1)Quoted 2 lines
Quoted 2 lines
as per feedback.
* drop one /* NOTE: review for potential ordering optimization */since
Quoted 46 lines
Quoted 31 lines
the note reference non-critical to perf control path.
note:
Remainder of the NOTE comments have been retained since there seems to be no consensus but stronger opinion/argument to keep expressed. while I generally agree that changes should not include ``TODO'' style comments I also agree that without these comments in your face people are very unlikely to feel compelled to make the review they are trying to solicit without them. if it is absolute that the series won't be merged with them then I will remove them, but please be explicit soon.
v3:
* style, don't use c99 commentsv2:
* comment code where optimizations may be possible now that memory order can be specified.
* comment code where operations should potentially be atomic so that maintainers can review.
* change a couple of variables labeled as counters to be unsigned.Tyler Retzlaff (6):
stack: replace rte atomics with GCC builtin atomics dma/idxd: replace rte atomics with GCC builtin atomics net/ice: replace rte atomics with GCC builtin atomics net/ixgbe: replace rte atomics with GCC builtin atomics net/null: replace rte atomics with GCC builtin atomics net/ring: replace rte atomics with GCC builtin atomicsQuoted 14 lines
drivers/dma/idxd/idxd_internal.h | 3 +-- drivers/dma/idxd/idxd_pci.c | 11 ++++++----- drivers/net/ice/ice_dcf.c | 1 - drivers/net/ice/ice_dcf_ethdev.c | 1 - drivers/net/ice/ice_ethdev.c | 12 ++++++++---- drivers/net/ixgbe/ixgbe_bypass.c | 1 - drivers/net/ixgbe/ixgbe_ethdev.c | 18 ++++++++++++------ drivers/net/ixgbe/ixgbe_ethdev.h | 3 ++- drivers/net/ixgbe/ixgbe_flow.c | 1 - drivers/net/ixgbe/ixgbe_rxtx.c | 1 - drivers/net/null/rte_eth_null.c | 28 ++++++++++++++++++---------- drivers/net/ring/rte_eth_ring.c | 20 ++++++++++---------- lib/stack/rte_stack_lf_generic.h | 16 +++++++++------- 13 files changed, 66 insertions(+), 50 deletions(-)Quoted 8 lines
I am not really enthousiastic about those NOTE:.
I would prefer we get an explicit go/nogo from each maintainers, but this did not happen.
I think that this indicates that those NOTE: will rot in the code now.Thomas proposed to track those NOTE: in the release announce mail and that we ping maintainers regularly.
Let's see how it goes.Let's leave it for one release cycle, if with the the announce mail maintainers take no action within that time I'll commit to going through and cleaning them out before 23.11 rc1.
Quoted 3 lines
I am merging this series so we can progress on the $SUBJECT. Series applied, thanks.
Thanks David, this will allow forward progress.
Quoted 4 lines
Tyler, about the patch on the ring library that was dropped by got no viable alternative, I'll wait for a decision from ARM and you.
I'll wait for Honnappa to follow up and we'll decide what to do when he does.
Quoted 3 lines
--
David MarchandSignature · 9 lines
Patrick Robb Technical Service Manager UNH InterOperability Laboratory 21 Madbury Rd, Suite 100, Durham, NH 03824 www.iol.unh.edu
Quoted 18 lines · David Marchand
Hello Patrick,
Quoted 15 lines · Patrick Robb
I want to report a possible regression from this patch series seen from CI testing on our Intel 82599ES 10G NIC, which we failed to report to patchwork when this initially went under CI due to a bug in our Jenkins reporting scripts. Use of the ixgbe driver appears to be affected. Tyler I apologize for the issues seen with reporting. We've made some temporary changes to avoid this happening again, and are currently reworking our reporting process entirely to provide greater reliability.
Here is a DTS snippet showing the issue, and the full log for the failing
virtio_smoketest can be downloaded here: https://dpdkdashboard.iol.unh.edu/results/dashboard/patchsets/26560/06/06/2023 18:22:58 TestVirtioSmoke: Start send packets and verify
06/06/2023 18:22:58 tester: ifconfig enp134s0f0 mtu 9000
06/06/2023 18:22:58 tester:
06/06/2023 18:42:59 TestVirtioSmoke: Test Casetest_virtio_pvpResult FAILED: TIMEOUT on port start 0
06/06/2023 18:42:59 TestVirtioSmoke: port start 0
ixgbe_dev_wait_setup_link_complete(): IXGBE link thread not complete too long time!ixgbe_dev_wait_setup_link_complete(): IXGBE link thread not complete too long time!ixgbe_dev_wait_setup_link_complete(): IXGBE link thread not complete too long time!
Quoted 9 lines
Quoted 3 lines
We initially took this Intel10G testing offline to investigate as we thought it was a lab infra failure. Obviously that wasn't the case, so ideally we will bring this back online when appropriate. But, I don't want to do so right now and start failing everyone's patchseries which are obviously unrelated to this. Comments on this are welcome, otherwise of course I will just return this test coverage to our CI when the state of the git tree allows for it.
Apologies for the missing report and the timeline on this. We are taking action to deliver results more reliably going forward.
(reduced the cc list a bit)
This is probably the same issue than what was reported by Intel
validation: https://bugs.dpdk.org/show_bug.cgi?id=1249
Thanks David
I should have read the next thread in the mail chain before replying.
Quoted 2 lines
A fix has been merged in next-net-intel, it will reach the main repo soon.
https://git.dpdk.org/next/dpdk-next-net-intel/commit/?id=fe4ce0aee766969a0e27fe28ced8ee7c761a2c4e
Patrick please let me know if after this integration I still need to investigate further.
Thanks
Quoted 4 lines
Signature · 2 lines
-- David Marchand
Thanks David, Tyler,
I ran the next-net-intel branch through DTS with the nic utilizing the ixgbe driver, and everything is passing now. When this reaches the main repo I will return the nic in question to UNH CI testing.
Best,
Patrick
Quoted 69 lines · Tyler Retzlaff
Quoted 5 lines · David Marchand
Hello Patrick,
Quoted 2 lines · Patrick Robb
I want to report a possible regression from this patch series seen
from CI testing on our Intel 82599ES 10G NIC, which we failed to report to patchwork when this initially went under CI due to a bug in our Jenkins reporting scripts. Use of the ixgbe driver appears to be affected. Tyler I apologize for the issues seen with reporting. We've made some temporary changes to avoid this happening again, and are currently reworking our reporting process entirely to provide greater reliability.
Quoted 2 lines
Quoted 2 lines
Here is a DTS snippet showing the issue, and the full log for the
failing
virtio_smoketest can be downloaded here:
https://dpdkdashboard.iol.unh.edu/results/dashboard/patchsets/26560/Quoted 2 lines
Quoted 2 lines
06/06/2023 18:22:58 TestVirtioSmoke: Start send packets
and verify
Quoted 1 line
Quoted 1 line
06/06/2023 18:22:58 tester: ifconfig
enp134s0f0 mtu 9000
Quoted 2 lines
Quoted 2 lines
06/06/2023 18:22:58 tester:
06/06/2023 18:42:59 TestVirtioSmoke: Test Case
test_virtio_pvpResult FAILED: TIMEOUT on port start 0Quoted 3 lines
Quoted 3 lines
06/06/2023 18:42:59 TestVirtioSmoke: port start 0
ixgbe_dev_wait_setup_link_complete(): IXGBE link thread not completetoo long time!
Quoted 1 line
Quoted 1 line
ixgbe_dev_wait_setup_link_complete(): IXGBE link thread not completetoo long time!
Quoted 1 line
Quoted 1 line
ixgbe_dev_wait_setup_link_complete(): IXGBE link thread not completetoo long time!
Quoted 2 lines
Quoted 2 lines
We initially took this Intel10G testing offline to investigate as we
thought it was a lab infra failure. Obviously that wasn't the case, so ideally we will bring this back online when appropriate. But, I don't want to do so right now and start failing everyone's patchseries which are obviously unrelated to this. Comments on this are welcome, otherwise of course I will just return this test coverage to our CI when the state of the git tree allows for it.
Quoted 2 lines
Quoted 2 lines
Apologies for the missing report and the timeline on this. We are
taking action to deliver results more reliably going forward.
Quoted 6 lines
(reduced the cc list a bit)
This is probably the same issue than what was reported by Intel
validation: https://bugs.dpdk.org/show_bug.cgi?id=1249Thanks David
I should have read the next thread in the mail chain before replying.
Quoted 1 line
A fix has been merged in next-net-intel, it will reach the main repo
soon.
Quoted 1 line
https://git.dpdk.org/next/dpdk-next-net-intel/commit/?id=fe4ce0aee766969a0e27fe28ced8ee7c761a2c4e
Patrick please let me know if after this integration I still need to investigate further.
Thanks
Quoted 4 lines
--
David Marchand
Signature · 9 lines
Patrick Robb Technical Service Manager UNH InterOperability Laboratory 21 Madbury Rd, Suite 100, Durham, NH 03824 www.iol.unh.edu
Quoted 4 lines · Patrick Robb
Thanks David, Tyler,
I ran the next-net-intel branch through DTS with the nic utilizing the ixgbe driver, and everything is passing now. When this reaches the main repo I will return the nic in question to UNH CI testing.
Thomas pulled the fix in main and tagged v23.07-rc2. At the time I write this mail, main, next-baseband, next-net, next-net-intel and next-virtio branches are fine.
Other subtree repositories will get updated soon. Ping Ajit, Jerin, Raslan, Akhil for their respective subtrees.
Thanks.
Signature · 1 line
David Marchand