Support smp.clusters for x86 in QEMU
v7 cover Support smp.clusters for x86 in QEMU
From: Zhao Liu <[email protected]>
Hi list,
This is the our v7 patch series, rebased on the master branch at the commit d328fef93ae7 ("Merge tag 'pull-20231230' of
https://gitlab.com/rth7680/qemu into staging").
No more change since v6 [1] exclude the comment nit update.
Welcome your comments!
PS: Since v5, we have dropped "x-l2-cache-topo" option and now are working on porting the original x-l2-cache-topo option to smp [2]. Just like:
-smp cpus=4,sockets=2,cores=2,threads=1, \
l3-cache=socket,l2-cache=core,l1-i-cache=core,l1-d-cache=coreThe cache topology enhancement in this patch set is the preparation for supporting future user-configurable cache topology (via generic cli interface).
---
# Introduction
This series adds the cluster support for x86 PC machine, which allows x86 can use smp.clusters to configure the module level CPU topology of x86.
This series also is the preparation to help x86 to define the more flexible cache topology, such as having multiple cores share the same L2 cache at cluster level. (That was what x-l2-cache-topo did, and we will explore a generic way.)
About why we don't share L2 cache at cluster and need a configuration way, pls see section: ## Why not share L2 cache in cluster directly.
# Background
The "clusters" parameter in "smp" is introduced by ARM [3], but x86 hasn't supported it.
At present, x86 defaults L2 cache is shared in one core, but this is not enough. There're some platforms that multiple cores share the same L2 cache, e.g., Alder Lake-P shares L2 cache for one module of Atom cores [4], that is, every four Atom cores shares one L2 cache. Therefore, we need the new CPU topology level (cluster/module).
Another reason is for hybrid architecture. cluster support not only provides another level of topology definition in x86, but would also provide required code change for future our hybrid topology support.
# Overview
## Introduction of module level for x86
"cluster" in smp is the CPU topology level which is between "core" and die.
For x86, the "cluster" in smp is corresponding to the module level [4], which is above the core level. So use the "module" other than "cluster" in x86 code.
And please note that x86 already has a cpu topology level also named "cluster" [5], this level is at the upper level of the package. Here, the cluster in x86 cpu topology is completely different from the "clusters" as the smp parameter. After the module level is introduced, the cluster as the smp parameter will actually refer to the module level of x86.
## Why not share L2 cache in cluster directly
Though "clusters" was introduced to help define L2 cache topology [3], using cluster to define x86's L2 cache topology will cause the compatibility problem:
Currently, x86 defaults that the L2 cache is shared in one core, which actually implies a default setting "cores per L2 cache is 1" and therefore implicitly defaults to having as many L2 caches as cores.
For example (i386 PC machine):
-smp 16,sockets=2,dies=2,cores=2,threads=2,maxcpus=16 (*)
Considering the topology of the L2 cache, this (*) implicitly means "1 core per L2 cache" and "2 L2 caches per die".
If we use cluster to configure L2 cache topology with the new default setting "clusters per L2 cache is 1", the above semantics will change to "2 cores per cluster" and "1 cluster per L2 cache", that is, "2 cores per L2 cache".
So the same command (*) will cause changes in the L2 cache topology, further affecting the performance of the virtual machine.
Therefore, x86 should only treat cluster as a cpu topology level and avoid using it to change L2 cache by default for compatibility.
## module level in CPUID
Linux kernel (from v6.4, with commit edc0a2b595765 ("x86/topology: Fix erroneous smp_num_siblings on Intel Hybrid platforms") is able to handle platforms with Module level enumerated via CPUID.1F.
Expose the module level in CPUID[0x1F] (for Intel CPUs) if the machine has more than 1 modules since v3.
## New cache topology info in CPUCacheInfo
(This is in preparation for users being able to configure cache topology from the cli later on.)
Currently, by default, the cache topology is encoded as:
1. i/d cache is shared in one core.
2. L2 cache is shared in one core.
3. L3 cache is shared in one die.
This default general setting has caused a misunderstanding, that is, the cache topology is completely equated with a specific cpu topology, such as the connection between L2 cache and core level, and the connection between L3 cache and die level.
In fact, the settings of these topologies depend on the specific platform and are not static. For example, on Alder Lake-P, every four Atom cores share the same L2 cache [3].
Thus, in this patch set, we explicitly define the corresponding cache topology for different cpu models and this has two benefits:
1. Easy to expand to new CPU models in the future, which has different cache topology.
2. It can easily support custom cache topology by some command.
# Patch description
patch 1 Fixes about x86 topology and Intel l1 cache topology.
patch 2-3 Cleanups about topology related CPUID encoding and QEMU
topology variables.
patch 4-5 Refactor CPUID[0x1F] encoding to prepare to introduce module
level.
patch 6-12 Add the module as the new CPU topology level in x86, and it
is corresponding to the cluster level in generic code.
patch 13,14,16 Add cache topology information in cache models.
patch 15 Update AMD CPUs' cache topology encoding.
[1]: https://lore.kernel.org/qemu-devel/[email protected]/
[2]: https://lists.gnu.org/archive/html/qemu-devel/2023-10/msg01954.html
[3]: https://patchew.org/QEMU/[email protected]/
[4]: https://www.intel.com/content/www/us/en/products/platforms/details/alder-lake-p.html
[5]: SDM, vol.3, ch.9, 9.9.1 Hierarchical Mapping of Shared Resources.
Best Regards,
Zhao
---
Changelog:
Changes since v6:
* Update the comment when check cluster-id. Since there's no v8.2, the cluster-id support should at least start from v9.0.
* Rebase on commit d328fef93ae7 ("Merge tag 'pull-20231230' of
https://gitlab.com/rth7680/qemu into staging").
Changes since v5:
* The first four patches of v5 [1] have been merged, v6 contains the remaining patches.
* Reabse on the latest master.
* Update the comment when check cluster-id. Since current QEMU is v8.2, the cluster-id support should at least start from v8.3.
Changes since v4:
* Drop the "x-l2-cache-topo" option. (Michael)
* Add A/R/T tags.
Changes since v3 (main changes):
* Expose module level in CPUID[0x1F].
* Fix compile warnings. (Babu)
* Fixes cache topology uninitialization bugs for some AMD CPUs. (Babu)
Changes since v2:
* Add "Tested-by", "Reviewed-by" and "ACKed-by" tags.
* Use newly added wrapped helper to get cores per socket in qemu_init_vcpu().
Changes since v1:
* Reordered patches. (Yanan)
* Deprecated the patch to fix comment of machine_parse_smp_config(). (Yanan)
* Rename test-x86-cpuid.c to test-x86-topo.c. (Yanan)
* Split the intel's l1 cache topology fix into a new separate patch. (Yanan)
* Combined module_id and APIC ID for module level support into one patch. (Yanan)
* Make cache_into_passthrough case of cpuid 0x04 leaf in
* cpu_x86_cpuid() use max_processor_ids_for_cache() and max_core_ids_in_package() to encode CPUID[4]. (Yanan)
* Add the prefix "CPU_TOPO_LEVEL_*" for CPU topology level names. (Yanan)
---
Zhao Liu (10):
i386/cpu: Fix i/d-cache topology to core level for Intel CPU i386/cpu: Use APIC ID offset to encode cache topo in CPUID[4] i386/cpu: Consolidate the use of topo_info in cpu_x86_cpuid()
i386: Split topology types of CPUID[0x1F] from the definitions of
CPUID[0xB]
i386: Decouple CPUID[0x1F] subleaf with specific topology level
i386: Expose module level in CPUID[0x1F]
i386: Add cache topology info in CPUCacheInfo
i386: Use CPUCacheInfo.share_level to encode CPUID[4]
i386: Use offsets get NumSharingCache for CPUID[0x8000001D].EAX[bits
25:14]
i386: Use CPUCacheInfo.share_level to encode
CPUID[0x8000001D].EAX[bits 25:14]Zhuocheng Ding (6):
i386: Introduce module-level cpu topology to CPUX86State
i386: Support modules_per_die in X86CPUTopoInfo
i386: Support module_id in X86CPUTopoIDs i386/cpu: Introduce cluster-id to X86CPU
tests: Add test case of APIC ID for module level parsing hw/i386/pc: Support smp.clusters for x86 PC machine
hw/i386/pc.c | 1 + hw/i386/x86.c | 49 ++++++- include/hw/i386/topology.h | 35 ++++- qemu-options.hx | 10 +- target/i386/cpu.c | 289 +++++++++++++++++++++++++++++-------- target/i386/cpu.h | 43 +++++- target/i386/kvm/kvm.c | 2 +- tests/unit/test-x86-topo.c | 56 ++++--- 8 files changed, 379 insertions(+), 106 deletions(-)
Signature · 1 line
2.34.1
PATCH v7 1/16 i386/cpu: Fix i/d-cache topology to core level for Intel CPU
From: Zhao Liu <[email protected]>
For i-cache and d-cache, current QEMU hardcodes the maximum IDs for CPUs sharing cache (CPUID.04H.00H:EAX[bits 25:14] and CPUID.04H.01H:EAX[bits 25:14]) to 0, and this means i-cache and d-cache are shared in the SMT level.
This is correct if there's single thread per core, but is wrong for the hyper threading case (one core contains multiple threads) since the i-cache and d-cache are shared in the core level other than SMT level.
For AMD CPU, commit 8f4202fb1080 ("i386: Populate AMD Processor Cache Information for cpuid 0x8000001D") has already introduced i/d cache topology as core level by default.
Therefore, in order to be compatible with both multi-threaded and single-threaded situations, we should set i-cache and d-cache be shared at the core level by default.
This fix changes the default i/d cache topology from per-thread to per-core. Potentially, this change in L1 cache topology may affect the performance of the VM if the user does not specifically specify the topology or bind the vCPU. However, the way to achieve optimal performance should be to create a reasonable topology and set the appropriate vCPU affinity without relying on QEMU's default topology structure.
---
Changes since v3:
* Change the description of current i/d cache encoding status to avoid misleading to "architectural rules". (Xiaoyao)
Changes since v1:
* Split this fix from the patch named "i386/cpu: Fix number of addressable IDs in CPUID.04H".
* Add the explanation of the impact on performance. (Xiaoyao)
--- target/i386/cpu.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
target/i386/cpu.c +2 -2
@@ -6113,12 +6113,12 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,switch (count) {case 0 :/* L1 dcache info */ encode_cache_cpuid4 (env->cache_info_cpuid4.l1d_cache, -1 , cs->nr_cores, + cs->nr_threads, cs->nr_cores, eax, ebx, ecx, edx);break ;case 1 :/* L1 icache info */ encode_cache_cpuid4 (env->cache_info_cpuid4.l1i_cache, -1 , cs->nr_cores, + cs->nr_threads, cs->nr_cores, eax, ebx, ecx, edx);break ;case 2 :/* L2 cache info */
Signature · 1 line
2.34.1
PATCH v7 2/16 i386/cpu: Use APIC ID offset to encode cache topo in CPUID[4]
In reply to Zhao LiuFrom: Zhao Liu <[email protected]>
Refer to the fixes of cache_info_passthrough ([1], [2]) and SDM, the CPUID.04H:EAX[bits 25:14] and CPUID.04H:EAX[bits 31:26] should use the nearest power-of-2 integer.
The nearest power-of-2 integer can be calculated by pow2ceil() or by using APIC ID offset (like L3 topology using 1 << die_offset [3]).
But in fact, CPUID.04H:EAX[bits 25:14] and CPUID.04H:EAX[bits 31:26] are associated with APIC ID. For example, in linux kernel, the field "num_threads_sharing" (Bits 25 - 14) is parsed with APIC ID. And for another example, on Alder Lake P, the CPUID.04H:EAX[bits 31:26] is not matched with actual core numbers and it's calculated by:
"(1 << (pkg_offset - core_offset)) - 1".
Therefore the offset of APIC ID should be preferred to calculate nearest power-of-2 integer for CPUID.04H:EAX[bits 25:14] and CPUID.04H:EAX[bits 31:26]:
1. d/i cache is shared in a core, 1 << core_offset should be used instand of "cs->nr_threads" in encode_cache_cpuid4() for CPUID.04H.00H:EAX[bits 25:14] and CPUID.04H.01H:EAX[bits 25:14].
2. L2 cache is supposed to be shared in a core as for now, thereby 1 << core_offset should also be used instand of "cs->nr_threads" in encode_cache_cpuid4() for CPUID.04H.02H:EAX[bits 25:14].
3. Similarly, the value for CPUID.04H:EAX[bits 31:26] should also be calculated with the bit width between the Package and SMT levels in the APIC ID (1 << (pkg_offset - core_offset) - 1).
In addition, use APIC ID offset to replace "pow2ceil()" for cache_info_passthrough case.
[1]: efb3934adf9e ("x86: cpu: make sure number of addressable IDs for processor cores meets the spec") [2]: d7caf13b5fcf ("x86: cpu: fixup number of addressable IDs for logical processors sharing cache") [3]: d65af288a84d ("i386: Update new x86_apicid parsing rules with die_offset support")
---
Changes since v3:
* Fix compile warnings. (Babu)
* Fix spelling typo.
Changes since v1:
* Use APIC ID offset to replace "pow2ceil()" for cache_info_passthrough case. (Yanan)
* Split the L1 cache fix into a separate patch.
* Rename the title of this patch (the original is "i386/cpu: Fix number of addressable IDs in CPUID.04H").
--- target/i386/cpu.c | 30 +++++++++++++++++++++++------- 1 file changed, 23 insertions(+), 7 deletions(-)
target/i386/cpu.c +23 -7
@@ -6014,7 +6014,6 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count, {X86CPU *cpu =env_archcpu (env);CPUState *cs =env_cpu (env); -uint32_t die_offset;uint32_t limit;uint32_t signature[3 ];X86CPUTopoInfo topo_info;
@@ -6098,39 +6097,56 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,int host_vcpus_per_cache =1 + ((*eax &0x3FFC000 ) >>14 );int vcpus_per_socket = cs->nr_cores * cs->nr_threads;if (cs->nr_cores >1 ) { +int addressable_cores_offset = +apicid_pkg_offset (&topo_info) - +apicid_core_offset (&topo_info); + *eax &= ~0xFC000000 ; - *eax |= (pow2ceil (cs->nr_cores) -1 ) <<26 ; + *eax |= (1 << (addressable_cores_offset -1 )) <<26 ; }if (host_vcpus_per_cache > vcpus_per_socket) { +int pkg_offset =apicid_pkg_offset (&topo_info); + *eax &= ~0x3FFC000 ; - *eax |= (pow2ceil (vcpus_per_socket) -1 ) <<14 ; + *eax |= (1 << (pkg_offset -1 )) <<14 ; } } }else if (cpu->vendor_cpuid_only &&IS_AMD_CPU (env)) { *eax = *ebx = *ecx = *edx =0 ; }else { *eax =0 ; +int addressable_cores_offset =apicid_pkg_offset (&topo_info) - +apicid_core_offset (&topo_info); +int core_offset, die_offset; +switch (count) {case 0 :/* L1 dcache info */ + core_offset =apicid_core_offset (&topo_info);encode_cache_cpuid4 (env->cache_info_cpuid4.l1d_cache, - cs->nr_threads, cs->nr_cores, + (1 << core_offset), + (1 << addressable_cores_offset), eax, ebx, ecx, edx);break ;case 1 :/* L1 icache info */ + core_offset =apicid_core_offset (&topo_info);encode_cache_cpuid4 (env->cache_info_cpuid4.l1i_cache, - cs->nr_threads, cs->nr_cores, + (1 << core_offset), + (1 << addressable_cores_offset), eax, ebx, ecx, edx);break ;case 2 :/* L2 cache info */ + core_offset =apicid_core_offset (&topo_info);encode_cache_cpuid4 (env->cache_info_cpuid4.l2_cache, - cs->nr_threads, cs->nr_cores, + (1 << core_offset), + (1 << addressable_cores_offset), eax, ebx, ecx, edx);break ;case 3 :/* L3 cache info */ die_offset =apicid_die_offset (&topo_info);if (cpu->enable_l3_cache) {encode_cache_cpuid4 (env->cache_info_cpuid4.l3_cache, - (1 << die_offset), cs->nr_cores, + (1 << die_offset), + (1 << addressable_cores_offset), eax, ebx, ecx, edx);break ; }
Signature · 1 line
2.34.1
Re: [PATCH v7 02/16] i386/cpu: Use APIC ID offset to encode cache topo in CPUID[4]
Quoted 6 lines · Zhao Liu
From: Zhao Liu <[email protected]>
Refer to the fixes of
cache_info_passthrough([1], [2]) and SDM, the CPUID.04H:EAX[bits 25:14] and CPUID.04H:EAX[bits 31:26] should use the nearest power-of-2 integer.
Quoted 6 lines
The nearest power-of-2 integer can be calculated by
pow2ceil()or by using APIC ID offset (like L3 topology using 1 <<die_offset[3]).But in fact, CPUID.04H:EAX[bits 25:14] and CPUID.04H:EAX[bits 31:26] are associated with APIC ID. For example, in linux kernel, the field "
num_threads_sharing" (Bits 25 - 14) is parsed with APIC ID.
And for
Quoted 3 lines
another example, on Alder Lake P, the CPUID.04H:EAX[bits 31:26] is not matched with actual core numbers and it's calculated by:
"(1 << (pkg_offset-core_offset)) - 1".
could you elaborate it more? what is the value of actual core numbers on Alder lake P? and what is the pkg_offset and core_offset?
Quoted 5 lines
Therefore the offset of APIC ID should be preferred to calculate nearest power-of-2 integer for CPUID.04H:EAX[bits 25:14] and CPUID.04H:EAX[bits 31:26]:
1. d/i cache is shared in a core, 1 <<core_offsetshould be usedinstand of "cs->nr_threads" in encode_cache_cpuid4() for
/s/instand/instead
Quoted 3 lines
CPUID.04H.00H:EAX[bits 25:14] and CPUID.04H.01H:EAX[bits 25:14].
2. L2 cache is supposed to be shared in a core as for now, thereby
1 << core_offset should also be used instand of "cs->nr_threads" inditto
Quoted 46 lines
encode_cache_cpuid4() for CPUID.04H.02H:EAX[bits 25:14].3. Similarly, the value for CPUID.04H:EAX[bits 31:26] should also be calculated with the bit width between the Package and SMT levels in the APIC ID (1 << (
pkg_offset-core_offset) - 1).In addition, use APIC ID offset to replace "
pow2ceil()" forcache_info_passthroughcase.[1]:
efb3934adf9e("x86: cpu: make sure number of addressable IDs for processor cores meets the spec") [2]:d7caf13b5fcf("x86: cpu: fixup number of addressable IDs for logical processors sharing cache") [3]:d65af288a84d("i386: Update newx86_apicidparsing rules withdie_offsetsupport")Fixes 7e3482f82480 ("i386: Helpers to encode cache information consistently")Suggested-by Robert Hoo <[email protected]>Signed-off-by Zhao Liu <[email protected]>Tested-by Babu Moger <[email protected]>Tested-by Yongwei Ma <[email protected]>---
Changes since v3:
* Fix compile warnings. (Babu)
* Fix spelling typo.Changes since v1:
* Use APIC ID offset to replace "pow2ceil()" forcache_info_passthroughcase. (Yanan)
* Split the L1 cache fix into a separate patch.
* Rename the title of this patch (the original is "i386/cpu: Fix number of addressable IDs in CPUID.04H").--- target/i386/cpu.c | 30 +++++++++++++++++++++++------- 1 file changed, 23 insertions(+), 7 deletions(-)target/i386/cpu.c +0
-1@@ -6014,7 +6014,6 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count, {X86CPU *cpu =env_archcpu (env);CPUState *cs =env_cpu (env); -uint32_t die_offset;uint32_t limit;uint32_t signature[3 ];X86CPUTopoInfo topo_info;
target/i386/cpu.c void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count, · line 6105
Quoted 10 lines
target/i386/cpu.c +5
-1@@ -6098,39 +6097,56 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,int host_vcpus_per_cache =1 + ((*eax &0x3FFC000 ) >>14 );int vcpus_per_socket = cs->nr_cores * cs->nr_threads;if (cs->nr_cores >1 ) { +int addressable_cores_offset = +apicid_pkg_offset (&topo_info) - +apicid_core_offset (&topo_info); + *eax &= ~0xFC000000 ; - *eax |= (pow2ceil (cs->nr_cores) -1 ) <<26 ; + *eax |= (1 << (addressable_cores_offset -1 )) <<26 ;
it should be ((1 << addressable_cores_offset) - 1) << 26
I think naming it addressable_cores_width is better than addressable_cores_offset. It's not offset because offset means the bit position from bit 0.
And we can get the width by another algorithm:
int addressable_cores_width = apicid_core_width(&topo_info) + apicid_die_width(&topo_info); *eax |= ((1 << addressable_cores_width) - 1)) << 26;
target/i386/cpu.c void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count, · line 6111
Quoted 7 lines
target/i386/cpu.c +3
-1}if (host_vcpus_per_cache > vcpus_per_socket) { +int pkg_offset =apicid_pkg_offset (&topo_info); + *eax &= ~0x3FFC000 ; - *eax |= (pow2ceil (vcpus_per_socket) -1 ) <<14 ; + *eax |= (1 << (pkg_offset -1 )) <<14 ;
Ditto, ((1 << pkg_offset) - 1) << 14
For this one, I think pow2ceil(vcpus_per_socket) is better. Because it's intuitive that when host_vcpus_per_cache > vcpus_per_socket, we expose vcpus_per_cache (configured by users) to VM.
Quoted 45 lines
target/i386/cpu.c +15
-4} } }else if (cpu->vendor_cpuid_only &&IS_AMD_CPU (env)) { *eax = *ebx = *ecx = *edx =0 ; }else { *eax =0 ; +int addressable_cores_offset =apicid_pkg_offset (&topo_info) - +apicid_core_offset (&topo_info); +int core_offset, die_offset; +switch (count) {case 0 :/* L1 dcache info */ + core_offset =apicid_core_offset (&topo_info);encode_cache_cpuid4 (env->cache_info_cpuid4.l1d_cache, - cs->nr_threads, cs->nr_cores, + (1 << core_offset), + (1 << addressable_cores_offset), eax, ebx, ecx, edx);break ;case 1 :/* L1 icache info */ + core_offset =apicid_core_offset (&topo_info);encode_cache_cpuid4 (env->cache_info_cpuid4.l1i_cache, - cs->nr_threads, cs->nr_cores, + (1 << core_offset), + (1 << addressable_cores_offset), eax, ebx, ecx, edx);break ;case 2 :/* L2 cache info */ + core_offset =apicid_core_offset (&topo_info);encode_cache_cpuid4 (env->cache_info_cpuid4.l2_cache, - cs->nr_threads, cs->nr_cores, + (1 << core_offset), + (1 << addressable_cores_offset), eax, ebx, ecx, edx);break ;case 3 :/* L3 cache info */ die_offset =apicid_die_offset (&topo_info);if (cpu->enable_l3_cache) {encode_cache_cpuid4 (env->cache_info_cpuid4.l3_cache, - (1 << die_offset), cs->nr_cores, + (1 << die_offset), + (1 << addressable_cores_offset), eax, ebx, ecx, edx);break ; }
Re: [PATCH v7 02/16] i386/cpu: Use APIC ID offset to encode cache topo in CPUID[4]
Hi Xiaoyao,
Quoted 19 lines · Xiaoyao Li
Date: Wed, 10 Jan 2024 17:31:28 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 02/16] i386/cpu: Use APIC ID offset to encode cache
topo in CPUID[4]Quoted 6 lines · Zhao Liu
From: Zhao Liu <[email protected]>
Refer to the fixes of
cache_info_passthrough([1], [2]) and SDM, the CPUID.04H:EAX[bits 25:14] and CPUID.04H:EAX[bits 31:26] should use the nearest power-of-2 integer.Quoted 6 lines
The nearest power-of-2 integer can be calculated by
pow2ceil()or by using APIC ID offset (like L3 topology using 1 <<die_offset[3]).But in fact, CPUID.04H:EAX[bits 25:14] and CPUID.04H:EAX[bits 31:26] are associated with APIC ID. For example, in linux kernel, the field "
num_threads_sharing" (Bits 25 - 14) is parsed with APIC ID.
Quoted 7 lines
And for
Quoted 3 lines
another example, on Alder Lake P, the CPUID.04H:EAX[bits 31:26] is not matched with actual core numbers and it's calculated by:
"(1 << (pkg_offset-core_offset)) - 1".could you elaborate it more? what is the value of actual core numbers on Alder lake P? and what is the
pkg_offsetandcore_offset?
For example, the following's the CPUID dump of an ADL-S machine:
CPUID.04H:
0x00000004 0x00: eax=0xfc004121 ebx=0x01c0003f ecx=0x0000003f edx=0x00000000 0x00000004 0x01: eax=0xfc004122 ebx=0x01c0003f ecx=0x0000007f edx=0x00000000 0x00000004 0x02: eax=0xfc01c143 ebx=0x03c0003f ecx=0x000007ff edx=0x00000000 0x00000004 0x03: eax=0xfc1fc163 ebx=0x0240003f ecx=0x00009fff edx=0x00000004 0x00000004 0x04: eax=0x00000000 ebx=0x00000000 ecx=0x00000000 edx=0x00000000
CPUID.1FH:
0x0000001f 0x00: eax=0x00000001 ebx=0x00000001 ecx=0x00000100 edx=0x0000004c 0x0000001f 0x01: eax=0x00000007 ebx=0x00000014 ecx=0x00000201 edx=0x0000004c 0x0000001f 0x02: eax=0x00000000 ebx=0x00000000 ecx=0x00000002 edx=0x0000004c
The CPUID.04H:EAX[bits 31:26] is 63.
From CPUID.1FH.00H:EAX[bits 04:00], the core_offset is 1, and from CPUID.1FH.01H:EAX[bits 04:00], the pkg_offset is 7.
Thus we can verify that the above equation as:
1 << (0x7 - 0x1) - 1 = 63.
"Maximum number of addressable IDs" refers to the maximum number of IDs that can be enumerated in the APIC ID's topology layout, which does not necessarily correspond to the actual number of topology domains.
Quoted 8 lines
Quoted 6 lines
Therefore the offset of APIC ID should be preferred to calculate nearest power-of-2 integer for CPUID.04H:EAX[bits 25:14] and CPUID.04H:EAX[bits 31:26]:
1. d/i cache is shared in a core, 1 <<core_offsetshould be usedinstand of "cs->nr_threads" in encode_cache_cpuid4() for/s/instand/instead
Thanks!
Quoted 6 lines
Quoted 4 lines
CPUID.04H.00H:EAX[bits 25:14] and CPUID.04H.01H:EAX[bits 25:14]. 2. L2 cache is supposed to be shared in a core as for now, thereby 1 << core_offset should also be used instand of "cs->nr_threads" inditto
Okay.
Quoted 51 lines
Quoted 51 lines
encode_cache_cpuid4() for CPUID.04H.02H:EAX[bits 25:14].3. Similarly, the value for CPUID.04H:EAX[bits 31:26] should also be calculated with the bit width between the Package and SMT levels in the APIC ID (1 << (
pkg_offset-core_offset) - 1).In addition, use APIC ID offset to replace "
pow2ceil()" forcache_info_passthroughcase.[1]:
efb3934adf9e("x86: cpu: make sure number of addressable IDs for processor cores meets the spec") [2]:d7caf13b5fcf("x86: cpu: fixup number of addressable IDs for logical processors sharing cache") [3]:d65af288a84d("i386: Update newx86_apicidparsing rules withdie_offsetsupport")Fixes 7e3482f82480 ("i386: Helpers to encode cache information consistently")Suggested-by Robert Hoo <[email protected]>Signed-off-by Zhao Liu <[email protected]>Tested-by Babu Moger <[email protected]>Tested-by Yongwei Ma <[email protected]>---
Changes since v3:
* Fix compile warnings. (Babu)
* Fix spelling typo.Changes since v1:
* Use APIC ID offset to replace "pow2ceil()" forcache_info_passthroughcase. (Yanan)
* Split the L1 cache fix into a separate patch.
* Rename the title of this patch (the original is "i386/cpu: Fix number of addressable IDs in CPUID.04H").--- target/i386/cpu.c | 30 +++++++++++++++++++++++------- 1 file changed, 23 insertions(+), 7 deletions(-)target/i386/cpu.c +1
-1@@ -6014,7 +6014,6 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count, {X86CPU *cpu =env_archcpu (env);CPUState *cs =env_cpu (env); -uint32_t die_offset;uint32_t limit;uint32_t signature[3 ];X86CPUTopoInfo topo_info;@@ -6098,39 +6097,56 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,int host_vcpus_per_cache =1 + ((*eax &0x3FFC000 ) >>14 );int vcpus_per_socket = cs->nr_cores * cs->nr_threads;if (cs->nr_cores >1 ) { + int addressable_cores_offset =
Quoted 8 lines
target/i386/cpu.c void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count, · line 6105
Quoted 6 lines
target/i386/cpu.c +4
-1+apicid_pkg_offset (&topo_info) - +apicid_core_offset (&topo_info); + *eax &= ~0xFC000000 ; - *eax |= (pow2ceil (cs->nr_cores) -1 ) <<26 ; + *eax |= (1 << (addressable_cores_offset -1 )) <<26 ;it should be ((1 <<
addressable_cores_offset) - 1) << 26
Good catch! The helper wrapped in a subsequent patch masks the error here.
Quoted 4 lines
I think naming it
addressable_cores_widthis better thanaddressable_cores_offset. It's not offset because offset means the bit position from bit 0.
I agree, "width" is better.
Quoted 6 lines
And we can get the width by another algorithm:
int
addressable_cores_width=apicid_core_width(&topo_info)+apicid_die_width(&topo_info);
*eax |= ((1 <<addressable_cores_width) - 1)) << 26;
This algorithm lacks flexibility because there will be more topology levels between package and core, such as the cluster being introduced...
Using "addressable_cores_width" is clear enough.
Quoted 10 lines
target/i386/cpu.c void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count, · line 6111
Quoted 8 lines
target/i386/cpu.c +3
-1}if (host_vcpus_per_cache > vcpus_per_socket) { +int pkg_offset =apicid_pkg_offset (&topo_info); + *eax &= ~0x3FFC000 ; - *eax |= (pow2ceil (vcpus_per_socket) -1 ) <<14 ; + *eax |= (1 << (pkg_offset -1 )) <<14 ;Ditto, ((1 <<
pkg_offset) - 1) << 14
Thanks!
Quoted 4 lines
For this one, I think
pow2ceil(vcpus_per_socket)is better. Because it's intuitive that whenhost_vcpus_per_cache>vcpus_per_socket, we exposevcpus_per_cache(configured by users) to VM.
I tend to use a uniform calculation that is less confusing and easier to maintain. Since this field encodes "Maximum number of addressable IDs", OS can't get the exact number of CPUs/vCPUs sharing L3 from here, it can only know that L3 is shared at the package level.
Thanks,
Zhao
Re: [PATCH v7 02/16] i386/cpu: Use APIC ID offset to encode cache topo in CPUID[4]
Quoted 174 lines · Zhao Liu
Hi Xiaoyao,
Quoted 19 lines · Xiaoyao Li
Date: Wed, 10 Jan 2024 17:31:28 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 02/16] i386/cpu: Use APIC ID offset to encode cache
topo in CPUID[4]Quoted 6 lines · Zhao Liu
From: Zhao Liu <[email protected]>
Refer to the fixes of
cache_info_passthrough([1], [2]) and SDM, the CPUID.04H:EAX[bits 25:14] and CPUID.04H:EAX[bits 31:26] should use the nearest power-of-2 integer.Quoted 6 lines
The nearest power-of-2 integer can be calculated by
pow2ceil()or by using APIC ID offset (like L3 topology using 1 <<die_offset[3]).But in fact, CPUID.04H:EAX[bits 25:14] and CPUID.04H:EAX[bits 31:26] are associated with APIC ID. For example, in linux kernel, the field "
num_threads_sharing" (Bits 25 - 14) is parsed with APIC ID.Quoted 7 lines
And for
Quoted 3 lines
another example, on Alder Lake P, the CPUID.04H:EAX[bits 31:26] is not matched with actual core numbers and it's calculated by:
"(1 << (pkg_offset-core_offset)) - 1".could you elaborate it more? what is the value of actual core numbers on Alder lake P? and what is the
pkg_offsetandcore_offset?For example, the following's the CPUID dump of an ADL-S machine:
CPUID.04H:
0x000000040x00: eax=0xfc004121 ebx=0x01c0003f ecx=0x0000003f edx=0x000000000x000000040x01: eax=0xfc004122 ebx=0x01c0003f ecx=0x0000007f edx=0x000000000x000000040x02: eax=0xfc01c143 ebx=0x03c0003f ecx=0x000007ff edx=0x000000000x000000040x03: eax=0xfc1fc163 ebx=0x0240003f ecx=0x00009fff edx=0x000000040x000000040x04: eax=0x00000000 ebx=0x00000000 ecx=0x00000000 edx=0x00000000CPUID.1FH:
0x0000001f0x00: eax=0x00000001 ebx=0x00000001 ecx=0x00000100 edx=0x0000004c0x0000001f0x01: eax=0x00000007 ebx=0x00000014 ecx=0x00000201 edx=0x0000004c0x0000001f0x02: eax=0x00000000 ebx=0x00000000 ecx=0x00000002 edx=0x0000004cThe CPUID.04H:EAX[bits 31:26] is 63.
From CPUID.1FH.00H:EAX[bits 04:00], thecore_offsetis 1, and from
CPUID.1FH.01H:EAX[bits 04:00], thepkg_offsetis 7.Thus we can verify that the above equation as:
1 << (
0x7-0x1) - 1 = 63."Maximum number of addressable IDs" refers to the maximum number of IDs that can be enumerated in the APIC ID's topology layout, which does not necessarily correspond to the actual number of topology domains.
Quoted 8 lines
Quoted 6 lines
Therefore the offset of APIC ID should be preferred to calculate nearest power-of-2 integer for CPUID.04H:EAX[bits 25:14] and CPUID.04H:EAX[bits 31:26]:
1. d/i cache is shared in a core, 1 <<core_offsetshould be usedinstand of "cs->nr_threads" in encode_cache_cpuid4() for/s/instand/instead
Thanks!
Quoted 6 lines
Quoted 4 lines
CPUID.04H.00H:EAX[bits 25:14] and CPUID.04H.01H:EAX[bits 25:14]. 2. L2 cache is supposed to be shared in a core as for now, thereby 1 << core_offset should also be used instand of "cs->nr_threads" inditto
Okay.
Quoted 51 lines
Quoted 51 lines
encode_cache_cpuid4() for CPUID.04H.02H:EAX[bits 25:14].3. Similarly, the value for CPUID.04H:EAX[bits 31:26] should also be calculated with the bit width between the Package and SMT levels in the APIC ID (1 << (
pkg_offset-core_offset) - 1).In addition, use APIC ID offset to replace "
pow2ceil()" forcache_info_passthroughcase.[1]:
efb3934adf9e("x86: cpu: make sure number of addressable IDs for processor cores meets the spec") [2]:d7caf13b5fcf("x86: cpu: fixup number of addressable IDs for logical processors sharing cache") [3]:d65af288a84d("i386: Update newx86_apicidparsing rules withdie_offsetsupport")Fixes 7e3482f82480 ("i386: Helpers to encode cache information consistently")Suggested-by Robert Hoo <[email protected]>Signed-off-by Zhao Liu <[email protected]>Tested-by Babu Moger <[email protected]>Tested-by Yongwei Ma <[email protected]>---
Changes since v3:
* Fix compile warnings. (Babu)
* Fix spelling typo.Changes since v1:
* Use APIC ID offset to replace "pow2ceil()" forcache_info_passthroughcase. (Yanan)
* Split the L1 cache fix into a separate patch.
* Rename the title of this patch (the original is "i386/cpu: Fix number of addressable IDs in CPUID.04H").--- target/i386/cpu.c | 30 +++++++++++++++++++++++------- 1 file changed, 23 insertions(+), 7 deletions(-)target/i386/cpu.c +1
-1@@ -6014,7 +6014,6 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count, {X86CPU *cpu =env_archcpu (env);CPUState *cs =env_cpu (env); -uint32_t die_offset;uint32_t limit;uint32_t signature[3 ];X86CPUTopoInfo topo_info;@@ -6098,39 +6097,56 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,int host_vcpus_per_cache =1 + ((*eax &0x3FFC000 ) >>14 );int vcpus_per_socket = cs->nr_cores * cs->nr_threads;if (cs->nr_cores >1 ) { + int addressable_cores_offset =Quoted 8 lines
target/i386/cpu.c void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count, · line 6105
Quoted 6 lines
target/i386/cpu.c +4
-1+apicid_pkg_offset (&topo_info) - +apicid_core_offset (&topo_info); + *eax &= ~0xFC000000 ; - *eax |= (pow2ceil (cs->nr_cores) -1 ) <<26 ; + *eax |= (1 << (addressable_cores_offset -1 )) <<26 ;it should be ((1 <<
addressable_cores_offset) - 1) << 26Good catch! The helper wrapped in a subsequent patch masks the error here.
Quoted 4 lines
I think naming it
addressable_cores_widthis better thanaddressable_cores_offset. It's not offset because offset means the bit position from bit 0.I agree, "width" is better.
Quoted 6 lines
And we can get the width by another algorithm:
int
addressable_cores_width=apicid_core_width(&topo_info)+apicid_die_width(&topo_info);
*eax |= ((1 <<addressable_cores_width) - 1)) << 26;This algorithm lacks flexibility because there will be more topology levels between package and core, such as the cluster being introduced...
Using "
addressable_cores_width" is clear enough.Quoted 10 lines
target/i386/cpu.c void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count, · line 6111
Quoted 8 lines
target/i386/cpu.c +3
-1}if (host_vcpus_per_cache > vcpus_per_socket) { +int pkg_offset =apicid_pkg_offset (&topo_info); + *eax &= ~0x3FFC000 ; - *eax |= (pow2ceil (vcpus_per_socket) -1 ) <<14 ; + *eax |= (1 << (pkg_offset -1 )) <<14 ;Ditto, ((1 <<
pkg_offset) - 1) << 14Thanks!
Quoted 1 line
Quoted 6 lines
Quoted 3 lines
For this one, I think
pow2ceil(vcpus_per_socket)is better. Because it's intuitive that whenhost_vcpus_per_cache>vcpus_per_socket, we exposevcpus_per_cache(configured by users) to VM.I tend to use a uniform calculation that is less confusing and easier to maintain.
less confusing?
the original code is
if (host_vcpus_per_cache > vcpus_per_socket) {
*eax |= (pow2ceil(vcpus_per_socket) - 1) << 14;
}and this patch is going to change it to
if (host_vcpus_per_cache > vcpus_per_socket) {
int pkg_offset = apicid_pkg_offset(&topo_info);
*eax |= (1 << pkg_offset - 1)) << 14;
}Apparently, the former is clearer that everyone knows what is wants to do is "when guest's total vcpus_per_socket is even smaller than host's vcpu_per_cache, using guest's configuration". While the latter is more confusing.
Quoted 3 lines
Since this field encodes "Maximum number of addressable IDs", OS can't get the exact number of CPUs/vCPUs sharing L3 from here, it can only know that L3 is shared at the package level.
It doesn't matter with L3. What the code want to fulfill is that,
host_vcpus_per_cache is the actual number of LPs that share this level of cache. While vcpus_per_socket is the maximum numbere of LPs that can share a cache (at any level) in guest. When guest's maximum number is even smaller than host's, use guest's value.
Quoted 3 lines
Thanks,
Zhao
Re: [PATCH v7 02/16] i386/cpu: Use APIC ID offset to encode cache topo in CPUID[4]
Hi Xiaoyao,
Quoted 202 lines · Xiaoyao Li
Date: Sun, 14 Jan 2024 22:11:59 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 02/16] i386/cpu: Use APIC ID offset to encode cache
topo in CPUID[4]Quoted 174 lines · Zhao Liu
Hi Xiaoyao,
Quoted 19 lines · Xiaoyao Li
Date: Wed, 10 Jan 2024 17:31:28 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 02/16] i386/cpu: Use APIC ID offset to encode cache
topo in CPUID[4]Quoted 6 lines · Zhao Liu
From: Zhao Liu <[email protected]>
Refer to the fixes of
cache_info_passthrough([1], [2]) and SDM, the CPUID.04H:EAX[bits 25:14] and CPUID.04H:EAX[bits 31:26] should use the nearest power-of-2 integer.Quoted 6 lines
The nearest power-of-2 integer can be calculated by
pow2ceil()or by using APIC ID offset (like L3 topology using 1 <<die_offset[3]).But in fact, CPUID.04H:EAX[bits 25:14] and CPUID.04H:EAX[bits 31:26] are associated with APIC ID. For example, in linux kernel, the field "
num_threads_sharing" (Bits 25 - 14) is parsed with APIC ID.Quoted 7 lines
And for
Quoted 3 lines
another example, on Alder Lake P, the CPUID.04H:EAX[bits 31:26] is not matched with actual core numbers and it's calculated by:
"(1 << (pkg_offset-core_offset)) - 1".could you elaborate it more? what is the value of actual core numbers on Alder lake P? and what is the
pkg_offsetandcore_offset?For example, the following's the CPUID dump of an ADL-S machine:
CPUID.04H:
0x000000040x00: eax=0xfc004121 ebx=0x01c0003f ecx=0x0000003f edx=0x000000000x000000040x01: eax=0xfc004122 ebx=0x01c0003f ecx=0x0000007f edx=0x000000000x000000040x02: eax=0xfc01c143 ebx=0x03c0003f ecx=0x000007ff edx=0x000000000x000000040x03: eax=0xfc1fc163 ebx=0x0240003f ecx=0x00009fff edx=0x000000040x000000040x04: eax=0x00000000 ebx=0x00000000 ecx=0x00000000 edx=0x00000000CPUID.1FH:
0x0000001f0x00: eax=0x00000001 ebx=0x00000001 ecx=0x00000100 edx=0x0000004c0x0000001f0x01: eax=0x00000007 ebx=0x00000014 ecx=0x00000201 edx=0x0000004c0x0000001f0x02: eax=0x00000000 ebx=0x00000000 ecx=0x00000002 edx=0x0000004cThe CPUID.04H:EAX[bits 31:26] is 63.
From CPUID.1FH.00H:EAX[bits 04:00], thecore_offsetis 1, and from
CPUID.1FH.01H:EAX[bits 04:00], thepkg_offsetis 7.Thus we can verify that the above equation as:
1 << (
0x7-0x1) - 1 = 63."Maximum number of addressable IDs" refers to the maximum number of IDs that can be enumerated in the APIC ID's topology layout, which does not necessarily correspond to the actual number of topology domains.
Quoted 8 lines
Quoted 6 lines
Therefore the offset of APIC ID should be preferred to calculate nearest power-of-2 integer for CPUID.04H:EAX[bits 25:14] and CPUID.04H:EAX[bits 31:26]:
1. d/i cache is shared in a core, 1 <<core_offsetshould be usedinstand of "cs->nr_threads" in encode_cache_cpuid4() for/s/instand/instead
Thanks!
Quoted 6 lines
Quoted 4 lines
CPUID.04H.00H:EAX[bits 25:14] and CPUID.04H.01H:EAX[bits 25:14]. 2. L2 cache is supposed to be shared in a core as for now, thereby 1 << core_offset should also be used instand of "cs->nr_threads" inditto
Okay.
Quoted 51 lines
Quoted 51 lines
encode_cache_cpuid4() for CPUID.04H.02H:EAX[bits 25:14].3. Similarly, the value for CPUID.04H:EAX[bits 31:26] should also be calculated with the bit width between the Package and SMT levels in the APIC ID (1 << (
pkg_offset-core_offset) - 1).In addition, use APIC ID offset to replace "
pow2ceil()" forcache_info_passthroughcase.[1]:
efb3934adf9e("x86: cpu: make sure number of addressable IDs for processor cores meets the spec") [2]:d7caf13b5fcf("x86: cpu: fixup number of addressable IDs for logical processors sharing cache") [3]:d65af288a84d("i386: Update newx86_apicidparsing rules withdie_offsetsupport")Fixes 7e3482f82480 ("i386: Helpers to encode cache information consistently")Suggested-by Robert Hoo <[email protected]>Signed-off-by Zhao Liu <[email protected]>Tested-by Babu Moger <[email protected]>Tested-by Yongwei Ma <[email protected]>---
Changes since v3:
* Fix compile warnings. (Babu)
* Fix spelling typo.Changes since v1:
* Use APIC ID offset to replace "pow2ceil()" forcache_info_passthroughcase. (Yanan)
* Split the L1 cache fix into a separate patch.
* Rename the title of this patch (the original is "i386/cpu: Fix number of addressable IDs in CPUID.04H").--- target/i386/cpu.c | 30 +++++++++++++++++++++++------- 1 file changed, 23 insertions(+), 7 deletions(-)target/i386/cpu.c +1
-1@@ -6014,7 +6014,6 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count, {X86CPU *cpu =env_archcpu (env);CPUState *cs =env_cpu (env); -uint32_t die_offset;uint32_t limit;uint32_t signature[3 ];X86CPUTopoInfo topo_info;@@ -6098,39 +6097,56 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,int host_vcpus_per_cache =1 + ((*eax &0x3FFC000 ) >>14 );int vcpus_per_socket = cs->nr_cores * cs->nr_threads;if (cs->nr_cores >1 ) { + int addressable_cores_offset =Quoted 8 lines
target/i386/cpu.c void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count, · line 6105
Quoted 6 lines
target/i386/cpu.c +4
-1+apicid_pkg_offset (&topo_info) - +apicid_core_offset (&topo_info); + *eax &= ~0xFC000000 ; - *eax |= (pow2ceil (cs->nr_cores) -1 ) <<26 ; + *eax |= (1 << (addressable_cores_offset -1 )) <<26 ;it should be ((1 <<
addressable_cores_offset) - 1) << 26Good catch! The helper wrapped in a subsequent patch masks the error here.
Quoted 4 lines
I think naming it
addressable_cores_widthis better thanaddressable_cores_offset. It's not offset because offset means the bit position from bit 0.I agree, "width" is better.
Quoted 6 lines
And we can get the width by another algorithm:
int
addressable_cores_width=apicid_core_width(&topo_info)+apicid_die_width(&topo_info);
*eax |= ((1 <<addressable_cores_width) - 1)) << 26;This algorithm lacks flexibility because there will be more topology levels between package and core, such as the cluster being introduced...
Using "
addressable_cores_width" is clear enough.Quoted 10 lines
target/i386/cpu.c void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count, · line 6111
Quoted 8 lines
target/i386/cpu.c +3
-1}if (host_vcpus_per_cache > vcpus_per_socket) { +int pkg_offset =apicid_pkg_offset (&topo_info); + *eax &= ~0x3FFC000 ; - *eax |= (pow2ceil (vcpus_per_socket) -1 ) <<14 ; + *eax |= (1 << (pkg_offset -1 )) <<14 ;Ditto, ((1 <<
pkg_offset) - 1) << 14Thanks!
Quoted 1 line
Quoted 6 lines
Quoted 3 lines
For this one, I think
pow2ceil(vcpus_per_socket)is better. Because it's intuitive that whenhost_vcpus_per_cache>vcpus_per_socket, we exposevcpus_per_cache(configured by users) to VM.I tend to use a uniform calculation that is less confusing and easier to maintain.
less confusing?
the original code is
if (host_vcpus_per_cache > vcpus_per_socket) { *eax |= (pow2ceil(vcpus_per_socket) - 1) << 14; }and this patch is going to change it to
if (host_vcpus_per_cache > vcpus_per_socket) { int pkg_offset = apicid_pkg_offset(&topo_info); *eax |= (1 << pkg_offset - 1)) << 14; }
Quoted 4 lines
Apparently, the former is clearer that everyone knows what is wants to do is "when guest's total
vcpus_per_socketis even smaller than host'svcpu_per_cache, using guest's configuration". While the latter is more confusing.
IMO, the only differences are the variable naming and the way the details are encoded, what is actually trying to be expressed is the same - both set the cache topology at the package level.
There is no reason to use two encoding ways for the same field, and it'll be a code maintenance disaster.
I can add comment here to allay your concern.
Quoted 6 lines
Quoted 4 lines
Since this field encodes "Maximum number of addressable IDs", OS can't get the exact number of CPUs/vCPUs sharing L3 from here, it can only know that L3 is shared at the package level.
It doesn't matter with L3. What the code want to fulfill is that,
Yes, I misremembered here.
Quoted 6 lines
host_vcpus_per_cacheis the actual number of LPs that share this level of cache. Whilevcpus_per_socketis the maximum numbere of LPs that can share a cache (at any level) in guest. When guest's maximum number is even smaller than host's, use guest's value.
From the Guest's view, the cache is shared at package level. In hardware, this one field only reflects the topology level and does not accurately reflect the number of sharing CPUs.
So, we just need to make it clear that in this case the Guest cache topology level is package.
Thanks,
Zhao
Re: [PATCH v7 02/16] i386/cpu: Use APIC ID offset to encode cache topo in CPUID[4]
In reply to Zhao LiuQuoted 51 lines · Zhao Liu
Hi Xiaoyao,
Quoted 19 lines · Xiaoyao Li
Date: Wed, 10 Jan 2024 17:31:28 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 02/16] i386/cpu: Use APIC ID offset to encode cache
topo in CPUID[4]Quoted 6 lines · Zhao Liu
From: Zhao Liu <[email protected]>
Refer to the fixes of
cache_info_passthrough([1], [2]) and SDM, the CPUID.04H:EAX[bits 25:14] and CPUID.04H:EAX[bits 31:26] should use the nearest power-of-2 integer.Quoted 6 lines
The nearest power-of-2 integer can be calculated by
pow2ceil()or by using APIC ID offset (like L3 topology using 1 <<die_offset[3]).But in fact, CPUID.04H:EAX[bits 25:14] and CPUID.04H:EAX[bits 31:26] are associated with APIC ID. For example, in linux kernel, the field "
num_threads_sharing" (Bits 25 - 14) is parsed with APIC ID.Quoted 7 lines
And for
Quoted 3 lines
another example, on Alder Lake P, the CPUID.04H:EAX[bits 31:26] is not matched with actual core numbers and it's calculated by:
"(1 << (pkg_offset-core_offset)) - 1".could you elaborate it more? what is the value of actual core numbers on Alder lake P? and what is the
pkg_offsetandcore_offset?For example, the following's the CPUID dump of an ADL-S machine:
CPUID.04H:
0x000000040x00: eax=0xfc004121 ebx=0x01c0003f ecx=0x0000003f edx=0x000000000x000000040x01: eax=0xfc004122 ebx=0x01c0003f ecx=0x0000007f edx=0x000000000x000000040x02: eax=0xfc01c143 ebx=0x03c0003f ecx=0x000007ff edx=0x000000000x000000040x03: eax=0xfc1fc163 ebx=0x0240003f ecx=0x00009fff edx=0x000000040x000000040x04: eax=0x00000000 ebx=0x00000000 ecx=0x00000000 edx=0x00000000CPUID.1FH:
0x0000001f0x00: eax=0x00000001 ebx=0x00000001 ecx=0x00000100 edx=0x0000004c0x0000001f0x01: eax=0x00000007 ebx=0x00000014 ecx=0x00000201 edx=0x0000004c0x0000001f0x02: eax=0x00000000 ebx=0x00000000 ecx=0x00000002 edx=0x0000004cThe CPUID.04H:EAX[bits 31:26] is 63.
From CPUID.1FH.00H:EAX[bits 04:00], thecore_offsetis 1, and from
CPUID.1FH.01H:EAX[bits 04:00], thepkg_offsetis 7.
Quoted 8 lines
Thus we can verify that the above equation as:
1 << (
0x7-0x1) - 1 = 63."Maximum number of addressable IDs" refers to the maximum number of IDs that can be enumerated in the APIC ID's topology layout, which does not necessarily correspond to the actual number of topology domains.
you still don't tell how many core numbers on Alder lake P.
I guess the number is far smaller than 64, which is not matched with (63 + 1)
Re: [PATCH v7 02/16] i386/cpu: Use APIC ID offset to encode cache topo in CPUID[4]
Hi Xiaoyao,
Quoted 66 lines · Xiaoyao Li
Date: Mon, 15 Jan 2024 11:51:05 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 02/16] i386/cpu: Use APIC ID offset to encode cache
topo in CPUID[4]Quoted 51 lines · Zhao Liu
Hi Xiaoyao,
Quoted 19 lines · Xiaoyao Li
Date: Wed, 10 Jan 2024 17:31:28 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 02/16] i386/cpu: Use APIC ID offset to encode cache
topo in CPUID[4]Quoted 6 lines · Zhao Liu
From: Zhao Liu <[email protected]>
Refer to the fixes of
cache_info_passthrough([1], [2]) and SDM, the CPUID.04H:EAX[bits 25:14] and CPUID.04H:EAX[bits 31:26] should use the nearest power-of-2 integer.Quoted 6 lines
The nearest power-of-2 integer can be calculated by
pow2ceil()or by using APIC ID offset (like L3 topology using 1 <<die_offset[3]).But in fact, CPUID.04H:EAX[bits 25:14] and CPUID.04H:EAX[bits 31:26] are associated with APIC ID. For example, in linux kernel, the field "
num_threads_sharing" (Bits 25 - 14) is parsed with APIC ID.Quoted 7 lines
And for
Quoted 3 lines
another example, on Alder Lake P, the CPUID.04H:EAX[bits 31:26] is not matched with actual core numbers and it's calculated by:
"(1 << (pkg_offset-core_offset)) - 1".could you elaborate it more? what is the value of actual core numbers on Alder lake P? and what is the
pkg_offsetandcore_offset?For example, the following's the CPUID dump of an ADL-S machine:
CPUID.04H:
0x000000040x00: eax=0xfc004121 ebx=0x01c0003f ecx=0x0000003f edx=0x000000000x000000040x01: eax=0xfc004122 ebx=0x01c0003f ecx=0x0000007f edx=0x000000000x000000040x02: eax=0xfc01c143 ebx=0x03c0003f ecx=0x000007ff edx=0x000000000x000000040x03: eax=0xfc1fc163 ebx=0x0240003f ecx=0x00009fff edx=0x000000040x000000040x04: eax=0x00000000 ebx=0x00000000 ecx=0x00000000 edx=0x00000000CPUID.1FH:
0x0000001f0x00: eax=0x00000001 ebx=0x00000001 ecx=0x00000100 edx=0x0000004c0x0000001f0x01: eax=0x00000007 ebx=0x00000014 ecx=0x00000201 edx=0x0000004c0x0000001f0x02: eax=0x00000000 ebx=0x00000000 ecx=0x00000002 edx=0x0000004cThe CPUID.04H:EAX[bits 31:26] is 63.
From CPUID.1FH.00H:EAX[bits 04:00], thecore_offsetis 1, and from
CPUID.1FH.01H:EAX[bits 04:00], thepkg_offsetis 7.Quoted 8 lines
Thus we can verify that the above equation as:
1 << (
0x7-0x1) - 1 = 63."Maximum number of addressable IDs" refers to the maximum number of IDs that can be enumerated in the APIC ID's topology layout, which does not necessarily correspond to the actual number of topology domains.
Quoted 5 lines
you still don't tell how many core numbers on Alder lake P.
I guess the number is far smaller than 64, which is not matched with (63 +
1)
There're 8 P cores (with 2 threads per P core) + 4 E cores (with 1 thread per E core) for this machine (ADL-S).
Thus this field only shows the theoretical size of the id space and does not reflect the actual cores numbers.
Thanks,
Zhao
PATCH v7 3/16 i386/cpu: Consolidate the use of topo_info in cpu_x86_cpuid()
In reply to Zhao LiuFrom: Zhao Liu <[email protected]>
In cpu_x86_cpuid(), there are many variables in representing the cpu topology, e.g., topo_info, cs->nr_cores/cs->nr_threads.
Since the names of cs->nr_cores/cs->nr_threads does not accurately represent its meaning, the use of cs->nr_cores/cs->nr_threads is prone to confusion and mistakes.
And the structure X86CPUTopoInfo names its members clearly, thus the variable "topo_info" should be preferred.
In addition, in cpu_x86_cpuid(), to uniformly use the topology variable, replace env->dies with topo_info.dies_per_pkg as well.
---
Changes since v3:
* Fix typo. (Babu)
Changes since v1:
* Extract cores_per_socket from the code block and use it as a local variable for cpu_x86_cpuid(). (Yanan)
* Remove vcpus_per_socket variable and use cpus_per_pkg directly. (Yanan)
* Replace env->dies with topo_info.dies_per_pkg in cpu_x86_cpuid().
--- target/i386/cpu.c | 31 ++++++++++++++++++------------- 1 file changed, 18 insertions(+), 13 deletions(-)
target/i386/cpu.c +18 -13
@@ -6017,11 +6017,16 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,uint32_t limit;uint32_t signature[3 ];X86CPUTopoInfo topo_info; +uint32_t cores_per_pkg; +uint32_t cpus_per_pkg; topo_info.dies_per_pkg = env->nr_dies; topo_info.cores_per_die = cs->nr_cores / env->nr_dies; topo_info.threads_per_core = cs->nr_threads; + cores_per_pkg = topo_info.cores_per_die * topo_info.dies_per_pkg; + cpus_per_pkg = cores_per_pkg * topo_info.threads_per_core; +/* Calculate & apply limits for different index ranges */ if (index >=0xC0000000 ) { limit = env->cpuid_xlevel2;
@@ -6057,8 +6062,8 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count, *ecx |=CPUID_EXT_OSXSAVE ; } *edx = env->features[FEAT_1_EDX ]; -if (cs->nr_cores * cs->nr_threads >1 ) { - *ebx |= (cs->nr_cores * cs->nr_threads) <<16 ; +if (cpus_per_pkg >1 ) { + *ebx |= cpus_per_pkg <<16 ; *edx |=CPUID_HT ; }if (!cpu->enable_pmu) {
@@ -6095,8 +6100,8 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,*/ if (*eax &31 ) {int host_vcpus_per_cache =1 + ((*eax &0x3FFC000 ) >>14 ); -int vcpus_per_socket = cs->nr_cores * cs->nr_threads; -if (cs->nr_cores >1 ) { + +if (cores_per_pkg >1 ) {int addressable_cores_offset =apicid_pkg_offset (&topo_info) -apicid_core_offset (&topo_info);
@@ -6104,7 +6109,7 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count, *eax &= ~0xFC000000 ; *eax |= (1 << (addressable_cores_offset -1 )) <<26 ; } -if (host_vcpus_per_cache > vcpus_per_socket) { +if (host_vcpus_per_cache > cpus_per_pkg) {int pkg_offset =apicid_pkg_offset (&topo_info); *eax &= ~0x3FFC000 ;
@@ -6249,12 +6254,12 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,switch (count) {case 0 : *eax =apicid_core_offset (&topo_info); - *ebx = cs->nr_threads; + *ebx = topo_info.threads_per_core; *ecx |=CPUID_TOPOLOGY_LEVEL_SMT ;break ;case 1 : *eax =apicid_pkg_offset (&topo_info); - *ebx = cs->nr_cores * cs->nr_threads; + *ebx = cpus_per_pkg; *ecx |=CPUID_TOPOLOGY_LEVEL_CORE ;break ;default :
@@ -6274,7 +6279,7 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,break ;case 0x1F :/* V2 Extended Topology Enumeration Leaf */ -if (env->nr_dies <2 ) { +if (topo_info.dies_per_pkg <2 ) { *eax = *ebx = *ecx = *edx =0 ;break ; }
@@ -6284,7 +6289,7 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,switch (count) {case 0 : *eax =apicid_core_offset (&topo_info); - *ebx = cs->nr_threads; + *ebx = topo_info.threads_per_core; *ecx |=CPUID_TOPOLOGY_LEVEL_SMT ;break ;case 1 :
@@ -6294,7 +6299,7 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,break ;case 2 : *eax =apicid_pkg_offset (&topo_info); - *ebx = cs->nr_cores * cs->nr_threads; + *ebx = cpus_per_pkg; *ecx |=CPUID_TOPOLOGY_LEVEL_DIE ;break ;default :
@@ -6518,7 +6523,7 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,* discards multiple thread information if it is set. * So don't set it here for Intel to make Linux guests happy. */ -if (cs->nr_cores * cs->nr_threads >1 ) { +if (cpus_per_pkg >1 ) {if (env->cpuid_vendor1 !=CPUID_VENDOR_INTEL_1 || env->cpuid_vendor2 !=CPUID_VENDOR_INTEL_2 || env->cpuid_vendor3 !=CPUID_VENDOR_INTEL_3 ) {
@@ -6584,7 +6589,7 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count, *eax |= (cpu_x86_virtual_addr_width (env) <<8 ); } *ebx = env->features[FEAT_8000_0008_EBX ]; -if (cs->nr_cores * cs->nr_threads >1 ) { +if (cpus_per_pkg >1 ) {/* * Bits 15:12 is "The number of bits in the initial * Core::X86::Apic::ApicId[ApicId] value that indicate
@@ -6592,7 +6597,7 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,* Bits 7:0 is "The number of threads in the package is NC+1" */ *ecx = (apicid_pkg_offset (&topo_info) <<12 ) | - ((cs->nr_cores * cs->nr_threads) -1 ); + (cpus_per_pkg -1 ); }else { *ecx =0 ; }
Signature · 1 line
2.34.1
Re: [PATCH v7 03/16] i386/cpu: Consolidate the use of topo_info in cpu_x86_cpuid()
Quoted 4 lines · Zhao Liu
From: Zhao Liu <[email protected]>
In
cpu_x86_cpuid(), there are many variables in representing the cpu topology, e.g.,topo_info, cs->nr_cores/cs->nr_threads.
Please use comma instead of slash. cs->nr_cores/cs->nr_threads looks like one variable.
Quoted 34 lines
Since the names of cs->nr_cores/cs->nr_threads does not accurately represent its meaning, the use of cs->nr_cores/cs->nr_threads is prone to confusion and mistakes.
And the structure X86CPUTopoInfo names its members clearly, thus the variable "
topo_info" should be preferred.In addition, in
cpu_x86_cpuid(), to uniformly use the topology variable, replaceenv->dieswith topo_info.dies_per_pkg as well.Suggested-by Robert Hoo <[email protected]>Signed-off-by Zhao Liu <[email protected]>Tested-by Babu Moger <[email protected]>Tested-by Yongwei Ma <[email protected]>---
Changes since v3:
* Fix typo. (Babu)Changes since v1:
* Extractcores_per_socketfrom the code block and use it as a local variable forcpu_x86_cpuid(). (Yanan)
* Removevcpus_per_socketvariable and usecpus_per_pkgdirectly.(Yanan) * Replace env->dies with topo_info.dies_per_pkg in cpu_x86_cpuid(). --- target/i386/cpu.c | 31 ++++++++++++++++++------------- 1 file changed, 18 insertions(+), 13 deletions(-)target/i386/cpu.c +0
-0
target/i386/cpu.c void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count, · line 6021
Quoted 5 lines
target/i386/cpu.c +2
-0@@ -6017,11 +6017,16 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,uint32_t limit;uint32_t signature[3 ];X86CPUTopoInfo topo_info; +uint32_t cores_per_pkg; +uint32_t cpus_per_pkg;
I prefer to lps_per_pkg or threads_per_pkg.
Other than it,
Quoted 111 lines
target/i386/cpu.c +16
-13topo_info.dies_per_pkg = env->nr_dies; topo_info.cores_per_die = cs->nr_cores / env->nr_dies; topo_info.threads_per_core = cs->nr_threads; + cores_per_pkg = topo_info.cores_per_die * topo_info.dies_per_pkg; + cpus_per_pkg = cores_per_pkg * topo_info.threads_per_core; +/* Calculate & apply limits for different index ranges */ if (index >=0xC0000000 ) { limit = env->cpuid_xlevel2;@@ -6057,8 +6062,8 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count, *ecx |=CPUID_EXT_OSXSAVE ; } *edx = env->features[FEAT_1_EDX ]; -if (cs->nr_cores * cs->nr_threads >1 ) { - *ebx |= (cs->nr_cores * cs->nr_threads) <<16 ; +if (cpus_per_pkg >1 ) { + *ebx |= cpus_per_pkg <<16 ; *edx |=CPUID_HT ; }if (!cpu->enable_pmu) {@@ -6095,8 +6100,8 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,*/ if (*eax &31 ) {int host_vcpus_per_cache =1 + ((*eax &0x3FFC000 ) >>14 ); -int vcpus_per_socket = cs->nr_cores * cs->nr_threads; -if (cs->nr_cores >1 ) { + +if (cores_per_pkg >1 ) {int addressable_cores_offset =apicid_pkg_offset (&topo_info) -apicid_core_offset (&topo_info);@@ -6104,7 +6109,7 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count, *eax &= ~0xFC000000 ; *eax |= (1 << (addressable_cores_offset -1 )) <<26 ; } -if (host_vcpus_per_cache > vcpus_per_socket) { +if (host_vcpus_per_cache > cpus_per_pkg) {int pkg_offset =apicid_pkg_offset (&topo_info); *eax &= ~0x3FFC000 ;@@ -6249,12 +6254,12 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,switch (count) {case 0 : *eax =apicid_core_offset (&topo_info); - *ebx = cs->nr_threads; + *ebx = topo_info.threads_per_core; *ecx |=CPUID_TOPOLOGY_LEVEL_SMT ;break ;case 1 : *eax =apicid_pkg_offset (&topo_info); - *ebx = cs->nr_cores * cs->nr_threads; + *ebx = cpus_per_pkg; *ecx |=CPUID_TOPOLOGY_LEVEL_CORE ;break ;default :@@ -6274,7 +6279,7 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,break ;case 0x1F :/* V2 Extended Topology Enumeration Leaf */ -if (env->nr_dies <2 ) { +if (topo_info.dies_per_pkg <2 ) { *eax = *ebx = *ecx = *edx =0 ;break ; }@@ -6284,7 +6289,7 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,switch (count) {case 0 : *eax =apicid_core_offset (&topo_info); - *ebx = cs->nr_threads; + *ebx = topo_info.threads_per_core; *ecx |=CPUID_TOPOLOGY_LEVEL_SMT ;break ;case 1 :@@ -6294,7 +6299,7 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,break ;case 2 : *eax =apicid_pkg_offset (&topo_info); - *ebx = cs->nr_cores * cs->nr_threads; + *ebx = cpus_per_pkg; *ecx |=CPUID_TOPOLOGY_LEVEL_DIE ;break ;default :@@ -6518,7 +6523,7 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,* discards multiple thread information if it is set. * So don't set it here for Intel to make Linux guests happy. */ -if (cs->nr_cores * cs->nr_threads >1 ) { +if (cpus_per_pkg >1 ) {if (env->cpuid_vendor1 !=CPUID_VENDOR_INTEL_1 || env->cpuid_vendor2 !=CPUID_VENDOR_INTEL_2 || env->cpuid_vendor3 !=CPUID_VENDOR_INTEL_3 ) {@@ -6584,7 +6589,7 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count, *eax |= (cpu_x86_virtual_addr_width (env) <<8 ); } *ebx = env->features[FEAT_8000_0008_EBX ]; -if (cs->nr_cores * cs->nr_threads >1 ) { +if (cpus_per_pkg >1 ) {/* * Bits 15:12 is "The number of bits in the initial * Core::X86::Apic::ApicId[ApicId] value that indicate @@ -6592,7 +6597,7 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,* Bits 7:0 is "The number of threads in the package is NC+1" */ *ecx = (apicid_pkg_offset (&topo_info) <<12 ) | - ((cs->nr_cores * cs->nr_threads) -1 ); + (cpus_per_pkg -1 ); }else { *ecx =0 ; }
Re: [PATCH v7 03/16] i386/cpu: Consolidate the use of topo_info in cpu_x86_cpuid()
Hi Xiaoyao,
Quoted 5 lines · Xiaoyao Li
Date: Wed, 10 Jan 2024 19:52:38 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 03/16] i386/cpu: Consolidate the use oftopo_infoin
cpu_x86_cpuid()
Quoted 8 lines
Quoted 4 lines · Zhao Liu
From: Zhao Liu <[email protected]>
In
cpu_x86_cpuid(), there are many variables in representing the cpu topology, e.g.,topo_info, cs->nr_cores/cs->nr_threads.Please use comma instead of slash. cs->nr_cores/cs->nr_threads looks like one variable.
Okay.
Quoted 30 lines
Quoted 30 lines
Since the names of cs->nr_cores/cs->nr_threads does not accurately represent its meaning, the use of cs->nr_cores/cs->nr_threads is prone to confusion and mistakes.
And the structure X86CPUTopoInfo names its members clearly, thus the variable "
topo_info" should be preferred.In addition, in
cpu_x86_cpuid(), to uniformly use the topology variable, replaceenv->dieswith topo_info.dies_per_pkg as well.Suggested-by Robert Hoo <[email protected]>Signed-off-by Zhao Liu <[email protected]>Tested-by Babu Moger <[email protected]>Tested-by Yongwei Ma <[email protected]>---
Changes since v3:
* Fix typo. (Babu)Changes since v1:
* Extractcores_per_socketfrom the code block and use it as a local variable forcpu_x86_cpuid(). (Yanan)
* Removevcpus_per_socketvariable and usecpus_per_pkgdirectly.(Yanan) * Replace env->dies with topo_info.dies_per_pkg in cpu_x86_cpuid(). --- target/i386/cpu.c | 31 ++++++++++++++++++------------- 1 file changed, 18 insertions(+), 13 deletions(-)
Quoted 12 lines
target/i386/cpu.c void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count, · line 6021
Quoted 10 lines
target/i386/cpu.c +2
-0@@ -6017,11 +6017,16 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,uint32_t limit;uint32_t signature[3 ];X86CPUTopoInfo topo_info; +uint32_t cores_per_pkg; +uint32_t cpus_per_pkg;I prefer to
lps_per_pkgorthreads_per_pkg.
Okay, lp is not common in QEMU code, so I would change this to threads_per_pkg.
Quoted 4 lines
Other than it,
Reviewed-by Xiaoyao Li <[email protected]>
Thanks!
-Zhao
PATCH v7 4/16 i386: Split topology types of CPUID[0x1F] from the definitions of CPUID[0xB]
In reply to Zhao LiuFrom: Zhao Liu <[email protected]>
CPUID[0xB] defines SMT, Core and Invalid types, and this leaf is shared by Intel and AMD CPUs.
But for extended topology levels, Intel CPU (in CPUID[0x1F]) and AMD CPU (in CPUID[0x80000026]) have the different definitions with different enumeration values.
Though CPUID[0x80000026] hasn't been implemented in QEMU, to avoid possible misunderstanding, split topology types of CPUID[0x1F] from the definitions of CPUID[0xB] and introduce CPUID[0x1F]-specific topology types.
---
Changes since v3:
* New commit to prepare to refactor CPUID[0x1F] encoding.
--- target/i386/cpu.c | 14 +++++++------- target/i386/cpu.h | 13 +++++++++---- 2 files changed, 16 insertions(+), 11 deletions(-)
target/i386/cpu.c +7 -7
@@ -6255,17 +6255,17 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,case 0 : *eax =apicid_core_offset (&topo_info); *ebx = topo_info.threads_per_core; - *ecx |=CPUID_TOPOLOGY_LEVEL_SMT ; + *ecx |=CPUID_B_ECX_TOPO_LEVEL_SMT <<8 ;break ;case 1 : *eax =apicid_pkg_offset (&topo_info); *ebx = cpus_per_pkg; - *ecx |=CPUID_TOPOLOGY_LEVEL_CORE ; + *ecx |=CPUID_B_ECX_TOPO_LEVEL_CORE <<8 ;break ;default : *eax =0 ; *ebx =0 ; - *ecx |=CPUID_TOPOLOGY_LEVEL_INVALID ; + *ecx |=CPUID_B_ECX_TOPO_LEVEL_INVALID <<8 ; }assert (!(*eax & ~0x1f ));
@@ -6290,22 +6290,22 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,case 0 : *eax =apicid_core_offset (&topo_info); *ebx = topo_info.threads_per_core; - *ecx |=CPUID_TOPOLOGY_LEVEL_SMT ; + *ecx |=CPUID_1F_ECX_TOPO_LEVEL_SMT <<8 ;break ;case 1 : *eax =apicid_die_offset (&topo_info); *ebx = topo_info.cores_per_die * topo_info.threads_per_core; - *ecx |=CPUID_TOPOLOGY_LEVEL_CORE ; + *ecx |=CPUID_1F_ECX_TOPO_LEVEL_CORE <<8 ;break ;case 2 : *eax =apicid_pkg_offset (&topo_info); *ebx = cpus_per_pkg; - *ecx |=CPUID_TOPOLOGY_LEVEL_DIE ; + *ecx |=CPUID_1F_ECX_TOPO_LEVEL_DIE <<8 ;break ;default : *eax =0 ; *ebx =0 ; - *ecx |=CPUID_TOPOLOGY_LEVEL_INVALID ; + *ecx |=CPUID_1F_ECX_TOPO_LEVEL_INVALID <<8 ; }assert (!(*eax & ~0x1f )); *ebx &=0xffff ;/* The count doesn't need to be reliable. */
target/i386/cpu.h +9 -4
@@ -1009,10 +1009,15 @@ uint64_t x86_cpu_get_supported_feature_word(FeatureWord w,#define CPUID_MWAIT_EMX (1U << 0)/* enumeration supported */ /* CPUID[0xB].ECX level types */ -#define CPUID_TOPOLOGY_LEVEL_INVALID (0U << 8) -#define CPUID_TOPOLOGY_LEVEL_SMT (1U << 8) -#define CPUID_TOPOLOGY_LEVEL_CORE (2U << 8) -#define CPUID_TOPOLOGY_LEVEL_DIE (5U << 8) +#define CPUID_B_ECX_TOPO_LEVEL_INVALID 0 +#define CPUID_B_ECX_TOPO_LEVEL_SMT 1 +#define CPUID_B_ECX_TOPO_LEVEL_CORE 2 + +/* COUID[0x1F].ECX level types */ +#define CPUID_1F_ECX_TOPO_LEVEL_INVALID CPUID_B_ECX_TOPO_LEVEL_INVALID +#define CPUID_1F_ECX_TOPO_LEVEL_SMT CPUID_B_ECX_TOPO_LEVEL_SMT +#define CPUID_1F_ECX_TOPO_LEVEL_CORE CPUID_B_ECX_TOPO_LEVEL_CORE +#define CPUID_1F_ECX_TOPO_LEVEL_DIE 5/* MSR Feature Bits */ #define MSR_ARCH_CAP_RDCL_NO (1U << 0)
Signature · 1 line
2.34.1
PATCH v7 5/16 i386: Decouple CPUID[0x1F] subleaf with specific topology level
In reply to Zhao LiuFrom: Zhao Liu <[email protected]>
At present, the subleaf 0x02 of CPUID[0x1F] is bound to the "die" level.
In fact, the specific topology level exposed in 0x1F depends on the platform's support for extension levels (module, tile and die).
To help expose "module" level in 0x1F, decouple CPUID[0x1F] subleaf with specific topology level.
---
Changes since v3:
* New patch to prepare to expose module level in 0x1F.
* Move the CPUTopoLevel enumeration definition from "i386: Add cache topology info in CPUCacheInfo" to this patch. Note, to align with topology types in SDM, revert the name of CPU_TOPO_LEVEL_UNKNOW to CPU_TOPO_LEVEL_INVALID.
--- target/i386/cpu.c | 136 +++++++++++++++++++++++++++++++++++++--------- target/i386/cpu.h | 15 +++++ 2 files changed, 126 insertions(+), 25 deletions(-)
target/i386/cpu.c +111 -25
@@ -269,6 +269,116 @@ static void encode_cache_cpuid4(CPUCacheInfo *cache, (cache->complex_indexing ?CACHE_COMPLEX_IDX :0 ); } +static uint32_t num_cpus_by_topo_level (X86CPUTopoInfo *topo_info, +enum CPUTopoLevel topo_level) +{ +switch (topo_level) { +case CPU_TOPO_LEVEL_SMT : +return 1 ; +case CPU_TOPO_LEVEL_CORE : +return topo_info->threads_per_core; +case CPU_TOPO_LEVEL_DIE : +return topo_info->threads_per_core * topo_info->cores_per_die; +case CPU_TOPO_LEVEL_PACKAGE : +return topo_info->threads_per_core * topo_info->cores_per_die * + topo_info->dies_per_pkg; +default : +g_assert_not_reached (); + } +return 0 ; +} + +static uint32_t apicid_offset_by_topo_level (X86CPUTopoInfo *topo_info, +enum CPUTopoLevel topo_level) +{ +switch (topo_level) { +case CPU_TOPO_LEVEL_SMT : +return 0 ; +case CPU_TOPO_LEVEL_CORE : +return apicid_core_offset (topo_info); +case CPU_TOPO_LEVEL_DIE : +return apicid_die_offset (topo_info); +case CPU_TOPO_LEVEL_PACKAGE : +return apicid_pkg_offset (topo_info); +default : +g_assert_not_reached (); + } +return 0 ; +} + +static uint32_t cpuid1f_topo_type (enum CPUTopoLevel topo_level) +{ +switch (topo_level) { +case CPU_TOPO_LEVEL_INVALID : +return CPUID_1F_ECX_TOPO_LEVEL_INVALID ; +case CPU_TOPO_LEVEL_SMT : +return CPUID_1F_ECX_TOPO_LEVEL_SMT ; +case CPU_TOPO_LEVEL_CORE : +return CPUID_1F_ECX_TOPO_LEVEL_CORE ; +case CPU_TOPO_LEVEL_DIE : +return CPUID_1F_ECX_TOPO_LEVEL_DIE ; +default : +/* Other types are not supported in QEMU. */ +g_assert_not_reached (); + } +return 0 ; +} + +static void encode_topo_cpuid1f (CPUX86State *env,uint32_t count, +X86CPUTopoInfo *topo_info, +uint32_t *eax,uint32_t *ebx, +uint32_t *ecx,uint32_t *edx) +{ + static DECLARE_BITMAP(topo_bitmap, CPU_TOPO_LEVEL_MAX); +X86CPU *cpu =env_archcpu (env); +unsigned long level, next_level; +uint32_t num_cpus_next_level, offset_next_level; + +/* +* Initialize the bitmap to decide which levels should be +* encoded in 0x1f. +*/ +if (!count) { +/* SMT and core levels are exposed in 0x1f leaf by default. */ +set_bit (CPU_TOPO_LEVEL_SMT , topo_bitmap); +set_bit (CPU_TOPO_LEVEL_CORE , topo_bitmap); + +if (env->nr_dies >1 ) { +set_bit (CPU_TOPO_LEVEL_DIE , topo_bitmap); + } + } + + *ecx = count &0xff ; + *edx = cpu->apic_id; + + level =find_first_bit (topo_bitmap,CPU_TOPO_LEVEL_MAX ); +if (level ==CPU_TOPO_LEVEL_MAX ) { + num_cpus_next_level =0 ; + offset_next_level =0 ; + +/* Encode CPU_TOPO_LEVEL_INVALID into the last subleaf of 0x1f. */ + level =CPU_TOPO_LEVEL_INVALID ; + }else { + next_level =find_next_bit (topo_bitmap,CPU_TOPO_LEVEL_MAX , level +1 ); +if (next_level ==CPU_TOPO_LEVEL_MAX ) { + next_level =CPU_TOPO_LEVEL_PACKAGE ; + } + + num_cpus_next_level =num_cpus_by_topo_level (topo_info, next_level); + offset_next_level =apicid_offset_by_topo_level (topo_info, next_level); + } + + *eax = offset_next_level; + *ebx = num_cpus_next_level; + *ecx |=cpuid1f_topo_type (level) <<8 ; + +assert (!(*eax & ~0x1f )); + *ebx &=0xffff ;/* The count doesn't need to be reliable. */ +if (level !=CPU_TOPO_LEVEL_MAX ) { +clear_bit (level, topo_bitmap); + } +} +/* Encode cache info for CPUID[0x80000005].ECX or CPUID[0x80000005].EDX */ static uint32_t encode_cache_cpuid80000005 (CPUCacheInfo *cache) {
@@ -6284,31 +6394,7 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,break ; } - *ecx = count &0xff ; - *edx = cpu->apic_id; -switch (count) { -case 0 : - *eax =apicid_core_offset (&topo_info); - *ebx = topo_info.threads_per_core; - *ecx |=CPUID_1F_ECX_TOPO_LEVEL_SMT <<8 ; -break ; -case 1 : - *eax =apicid_die_offset (&topo_info); - *ebx = topo_info.cores_per_die * topo_info.threads_per_core; - *ecx |=CPUID_1F_ECX_TOPO_LEVEL_CORE <<8 ; -break ; -case 2 : - *eax =apicid_pkg_offset (&topo_info); - *ebx = cpus_per_pkg; - *ecx |=CPUID_1F_ECX_TOPO_LEVEL_DIE <<8 ; -break ; -default : - *eax =0 ; - *ebx =0 ; - *ecx |=CPUID_1F_ECX_TOPO_LEVEL_INVALID <<8 ; - } -assert (!(*eax & ~0x1f )); - *ebx &=0xffff ;/* The count doesn't need to be reliable. */ +encode_topo_cpuid1f (env, count, &topo_info, eax, ebx, ecx, edx);break ;case 0xD : {/* Processor Extended State */
target/i386/cpu.h +15 -0
@@ -1008,6 +1008,21 @@ uint64_t x86_cpu_get_supported_feature_word(FeatureWord w,#define CPUID_MWAIT_IBE (1U << 1)/* Interrupts can exit capability */ #define CPUID_MWAIT_EMX (1U << 0)/* enumeration supported */ +/* +* CPUTopoLevel is the general i386 topology hierarchical representation, +* ordered by increasing hierarchical relationship. +* Its enumeration value is not bound to the type value of Intel (CPUID[0x1F]) +* or AMD (CPUID[0x80000026]). +*/ +enum CPUTopoLevel { +CPU_TOPO_LEVEL_INVALID , +CPU_TOPO_LEVEL_SMT , +CPU_TOPO_LEVEL_CORE , +CPU_TOPO_LEVEL_DIE , +CPU_TOPO_LEVEL_PACKAGE , +CPU_TOPO_LEVEL_MAX , +}; +/* CPUID[0xB].ECX level types */ #define CPUID_B_ECX_TOPO_LEVEL_INVALID 0#define CPUID_B_ECX_TOPO_LEVEL_SMT 1
Signature · 1 line
2.34.1
Re: [PATCH v7 05/16] i386: Decouple CPUID[0x1F] subleaf with specific topology level
Quoted 88 lines · Zhao Liu
From: Zhao Liu <[email protected]>
At present, the subleaf
0x02of CPUID[0x1F] is bound to the "die" level.In fact, the specific topology level exposed in
0x1Fdepends on the platform's support for extension levels (module, tile and die).To help expose "module" level in
0x1F, decouple CPUID[0x1F] subleaf with specific topology level.Signed-off-by Zhao Liu <[email protected]>Tested-by Babu Moger <[email protected]>Tested-by Yongwei Ma <[email protected]>---
Changes since v3:
* New patch to prepare to expose module level in0x1F.
* Move the CPUTopoLevel enumeration definition from "i386: Add cache topology info in CPUCacheInfo" to this patch. Note, to align with topology types in SDM, revert the name ofCPU_TOPO_LEVEL_UNKNOWtoCPU_TOPO_LEVEL_INVALID.--- target/i386/cpu.c | 136 +++++++++++++++++++++++++++++++++++++--------- target/i386/cpu.h | 15 +++++ 2 files changed, 126 insertions(+), 25 deletions(-)target/i386/cpu.c +54
-0@@ -269,6 +269,116 @@ static void encode_cache_cpuid4(CPUCacheInfo *cache, (cache->complex_indexing ?CACHE_COMPLEX_IDX :0 ); } +static uint32_t num_cpus_by_topo_level (X86CPUTopoInfo *topo_info, +enum CPUTopoLevel topo_level) +{ +switch (topo_level) { +case CPU_TOPO_LEVEL_SMT : +return 1 ; +case CPU_TOPO_LEVEL_CORE : +return topo_info->threads_per_core; +case CPU_TOPO_LEVEL_DIE : +return topo_info->threads_per_core * topo_info->cores_per_die; +case CPU_TOPO_LEVEL_PACKAGE : +return topo_info->threads_per_core * topo_info->cores_per_die * + topo_info->dies_per_pkg; +default : +g_assert_not_reached (); + } +return 0 ; +} + +static uint32_t apicid_offset_by_topo_level (X86CPUTopoInfo *topo_info, +enum CPUTopoLevel topo_level) +{ +switch (topo_level) { +case CPU_TOPO_LEVEL_SMT : +return 0 ; +case CPU_TOPO_LEVEL_CORE : +return apicid_core_offset (topo_info); +case CPU_TOPO_LEVEL_DIE : +return apicid_die_offset (topo_info); +case CPU_TOPO_LEVEL_PACKAGE : +return apicid_pkg_offset (topo_info); +default : +g_assert_not_reached (); + } +return 0 ; +} + +static uint32_t cpuid1f_topo_type (enum CPUTopoLevel topo_level) +{ +switch (topo_level) { +case CPU_TOPO_LEVEL_INVALID : +return CPUID_1F_ECX_TOPO_LEVEL_INVALID ; +case CPU_TOPO_LEVEL_SMT : +return CPUID_1F_ECX_TOPO_LEVEL_SMT ; +case CPU_TOPO_LEVEL_CORE : +return CPUID_1F_ECX_TOPO_LEVEL_CORE ; +case CPU_TOPO_LEVEL_DIE : +return CPUID_1F_ECX_TOPO_LEVEL_DIE ; +default : +/* Other types are not supported in QEMU. */ +g_assert_not_reached (); + } +return 0 ; +}
target/i386/cpu.c static void encode_cache_cpuid4(CPUCacheInfo *cache, · line 335
Quoted 10 lines
target/i386/cpu.c +10
-0+ +static void encode_topo_cpuid1f (CPUX86State *env,uint32_t count, +X86CPUTopoInfo *topo_info, +uint32_t *eax,uint32_t *ebx, +uint32_t *ecx,uint32_t *edx) +{ + static DECLARE_BITMAP(topo_bitmap, CPU_TOPO_LEVEL_MAX); +X86CPU *cpu =env_archcpu (env); +unsigned long level, next_level; +uint32_t num_cpus_next_level, offset_next_level;
again, I dislike the name of cpus to represent the logical process or thread. we can call it, num_lps_next_level, or num_threads_next_level;
target/i386/cpu.c static void encode_cache_cpuid4(CPUCacheInfo *cache, · line 341
Quoted 6 lines
target/i386/cpu.c +6
-0+ +/* +* Initialize the bitmap to decide which levels should be +* encoded in 0x1f. +*/ +if (!count) {
using static bitmap and initialize the bitmap on (count == 0), looks bad to me. It highly relies on the order of how encode_topo_cpuid1f() is called, and fragile.
Instead, we can maintain an array in CPUX86State, e.g.,
target/i386/cpu.h +2 -0
@@ -1904,6 +1904,8 @@ typedef struct CPUArchState {/* Number of dies within this CPU package. */ unsigned nr_dies; + +unint8_t valid_cpu_topo[CPU_TOPO_LEVEL_MAX ]; } CPUX86State;
and initialize it as below, when initializing the env
env->valid_cpu_topo[0] = CPU_TOPO_LEVEL_SMT;
env->valid_cpu_topo[1] = CPU_TOPO_LEVEL_CORE;
if (env->nr_dies > 1) {
env->valid_cpu_topo[2] = CPU_TOPO_LEVEL_DIE;
}then in encode_topo_cpuid1f(), we can get level and next_level as
level = env->valid_cpu_topo[count];next_level = env->valid_cpu_topo[count + 1];
Quoted 102 lines
target/i386/cpu.c +41
-25+/* SMT and core levels are exposed in 0x1f leaf by default. */ +set_bit (CPU_TOPO_LEVEL_SMT , topo_bitmap); +set_bit (CPU_TOPO_LEVEL_CORE , topo_bitmap); + +if (env->nr_dies >1 ) { +set_bit (CPU_TOPO_LEVEL_DIE , topo_bitmap); + } + } + + *ecx = count &0xff ; + *edx = cpu->apic_id; + + level =find_first_bit (topo_bitmap,CPU_TOPO_LEVEL_MAX ); +if (level ==CPU_TOPO_LEVEL_MAX ) { + num_cpus_next_level =0 ; + offset_next_level =0 ; + +/* Encode CPU_TOPO_LEVEL_INVALID into the last subleaf of 0x1f. */ + level =CPU_TOPO_LEVEL_INVALID ; + }else { + next_level =find_next_bit (topo_bitmap,CPU_TOPO_LEVEL_MAX , level +1 ); +if (next_level ==CPU_TOPO_LEVEL_MAX ) { + next_level =CPU_TOPO_LEVEL_PACKAGE ; + } + + num_cpus_next_level =num_cpus_by_topo_level (topo_info, next_level); + offset_next_level =apicid_offset_by_topo_level (topo_info, next_level); + } + + *eax = offset_next_level; + *ebx = num_cpus_next_level; + *ecx |=cpuid1f_topo_type (level) <<8 ; + +assert (!(*eax & ~0x1f )); + *ebx &=0xffff ;/* The count doesn't need to be reliable. */ +if (level !=CPU_TOPO_LEVEL_MAX ) { +clear_bit (level, topo_bitmap); + } +} +/* Encode cache info for CPUID[0x80000005].ECX or CPUID[0x80000005].EDX */ static uint32_t encode_cache_cpuid80000005 (CPUCacheInfo *cache) {@@ -6284,31 +6394,7 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,break ; } - *ecx = count &0xff ; - *edx = cpu->apic_id; -switch (count) { -case 0 : - *eax =apicid_core_offset (&topo_info); - *ebx = topo_info.threads_per_core; - *ecx |=CPUID_1F_ECX_TOPO_LEVEL_SMT <<8 ; -break ; -case 1 : - *eax =apicid_die_offset (&topo_info); - *ebx = topo_info.cores_per_die * topo_info.threads_per_core; - *ecx |=CPUID_1F_ECX_TOPO_LEVEL_CORE <<8 ; -break ; -case 2 : - *eax =apicid_pkg_offset (&topo_info); - *ebx = cpus_per_pkg; - *ecx |=CPUID_1F_ECX_TOPO_LEVEL_DIE <<8 ; -break ; -default : - *eax =0 ; - *ebx =0 ; - *ecx |=CPUID_1F_ECX_TOPO_LEVEL_INVALID <<8 ; - } -assert (!(*eax & ~0x1f )); - *ebx &=0xffff ;/* The count doesn't need to be reliable. */ +encode_topo_cpuid1f (env, count, &topo_info, eax, ebx, ecx, edx);break ;case 0xD : {/* Processor Extended State */ target/i386/cpu.h +15
-0@@ -1008,6 +1008,21 @@ uint64_t x86_cpu_get_supported_feature_word(FeatureWord w,#define CPUID_MWAIT_IBE (1U << 1)/* Interrupts can exit capability */ #define CPUID_MWAIT_EMX (1U << 0)/* enumeration supported */ +/* +* CPUTopoLevel is the general i386 topology hierarchical representation, +* ordered by increasing hierarchical relationship. +* Its enumeration value is not bound to the type value of Intel (CPUID[0x1F]) +* or AMD (CPUID[0x80000026]). +*/ +enum CPUTopoLevel { +CPU_TOPO_LEVEL_INVALID , +CPU_TOPO_LEVEL_SMT , +CPU_TOPO_LEVEL_CORE , +CPU_TOPO_LEVEL_DIE , +CPU_TOPO_LEVEL_PACKAGE , +CPU_TOPO_LEVEL_MAX , +}; +/* CPUID[0xB].ECX level types */ #define CPUID_B_ECX_TOPO_LEVEL_INVALID 0#define CPUID_B_ECX_TOPO_LEVEL_SMT 1
Re: [PATCH v7 05/16] i386: Decouple CPUID[0x1F] subleaf with specific topology level
Hi Xiaoyao,
Quoted 98 lines · Xiaoyao Li
Date: Thu, 11 Jan 2024 11:19:34 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 05/16] i386: Decouple CPUID[0x1F] subleaf with
specific topology levelQuoted 92 lines · Zhao Liu
From: Zhao Liu <[email protected]>
At present, the subleaf
0x02of CPUID[0x1F] is bound to the "die" level.In fact, the specific topology level exposed in
0x1Fdepends on the platform's support for extension levels (module, tile and die).To help expose "module" level in
0x1F, decouple CPUID[0x1F] subleaf with specific topology level.Signed-off-by Zhao Liu <[email protected]>Tested-by Babu Moger <[email protected]>Tested-by Yongwei Ma <[email protected]>---
Changes since v3:
* New patch to prepare to expose module level in0x1F.
* Move the CPUTopoLevel enumeration definition from "i386: Add cache topology info in CPUCacheInfo" to this patch. Note, to align with topology types in SDM, revert the name ofCPU_TOPO_LEVEL_UNKNOWtoCPU_TOPO_LEVEL_INVALID.--- target/i386/cpu.c | 136 +++++++++++++++++++++++++++++++++++++--------- target/i386/cpu.h | 15 +++++ 2 files changed, 126 insertions(+), 25 deletions(-)target/i386/cpu.c +59
-0@@ -269,6 +269,116 @@ static void encode_cache_cpuid4(CPUCacheInfo *cache, (cache->complex_indexing ?CACHE_COMPLEX_IDX :0 ); } +static uint32_t num_cpus_by_topo_level (X86CPUTopoInfo *topo_info, +enum CPUTopoLevel topo_level) +{ +switch (topo_level) { +case CPU_TOPO_LEVEL_SMT : +return 1 ; +case CPU_TOPO_LEVEL_CORE : +return topo_info->threads_per_core; +case CPU_TOPO_LEVEL_DIE : +return topo_info->threads_per_core * topo_info->cores_per_die; +case CPU_TOPO_LEVEL_PACKAGE : +return topo_info->threads_per_core * topo_info->cores_per_die * + topo_info->dies_per_pkg; +default : +g_assert_not_reached (); + } +return 0 ; +} + +static uint32_t apicid_offset_by_topo_level (X86CPUTopoInfo *topo_info, +enum CPUTopoLevel topo_level) +{ +switch (topo_level) { +case CPU_TOPO_LEVEL_SMT : +return 0 ; +case CPU_TOPO_LEVEL_CORE : +return apicid_core_offset (topo_info); +case CPU_TOPO_LEVEL_DIE : +return apicid_die_offset (topo_info); +case CPU_TOPO_LEVEL_PACKAGE : +return apicid_pkg_offset (topo_info); +default : +g_assert_not_reached (); + } +return 0 ; +} + +static uint32_t cpuid1f_topo_type (enum CPUTopoLevel topo_level) +{ +switch (topo_level) { +case CPU_TOPO_LEVEL_INVALID : +return CPUID_1F_ECX_TOPO_LEVEL_INVALID ; +case CPU_TOPO_LEVEL_SMT : +return CPUID_1F_ECX_TOPO_LEVEL_SMT ; +case CPU_TOPO_LEVEL_CORE : +return CPUID_1F_ECX_TOPO_LEVEL_CORE ; +case CPU_TOPO_LEVEL_DIE : +return CPUID_1F_ECX_TOPO_LEVEL_DIE ; +default : +/* Other types are not supported in QEMU. */ +g_assert_not_reached (); + } +return 0 ; +} + +static void encode_topo_cpuid1f (CPUX86State *env,uint32_t count, +X86CPUTopoInfo *topo_info, +uint32_t *eax,uint32_t *ebx, +uint32_t *ecx,uint32_t *edx)
Quoted 8 lines
target/i386/cpu.c static void encode_cache_cpuid4(CPUCacheInfo *cache, · line 334
Quoted 5 lines
target/i386/cpu.c +5
-0+{ + static DECLARE_BITMAP(topo_bitmap, CPU_TOPO_LEVEL_MAX); +X86CPU *cpu =env_archcpu (env); +unsigned long level, next_level; +uint32_t num_cpus_next_level, offset_next_level;again, I dislike the name of cpus to represent the logical process or thread. we can call it,
num_lps_next_level, ornum_threads_next_level;
Okay, will use num_threads_next_level ;-)
Quoted 29 lines
target/i386/cpu.c static void encode_cache_cpuid4(CPUCacheInfo *cache, · line 340
Quoted 7 lines
target/i386/cpu.c +6
-0+ +/* +* Initialize the bitmap to decide which levels should be +* encoded in 0x1f. +*/ +if (!count) {using static bitmap and initialize the bitmap on (count == 0), looks bad to me. It highly relies on the order of how
encode_topo_cpuid1f()is called, and fragile.Instead, we can maintain an array in CPUX86State, e.g.,
target/i386/cpu.h +2
-0@@ -1904,6 +1904,8 @@ typedef struct CPUArchState {/* Number of dies within this CPU package. */ unsigned nr_dies; + +unint8_t valid_cpu_topo[CPU_TOPO_LEVEL_MAX ]; } CPUX86State;and initialize it as below, when initializing the env
env->valid_cpu_topo[0] =
CPU_TOPO_LEVEL_SMT;
env->valid_cpu_topo[1] =CPU_TOPO_LEVEL_CORE;
Quoted 9 lines
if (env->nr_dies > 1) { env->valid_cpu_topo[2] = CPU_TOPO_LEVEL_DIE; }then in
encode_topo_cpuid1f(), we can get level andnext_levelaslevel = env->valid_cpu_topo[count];
next_level= env->valid_cpu_topo[count + 1];
Good idea, let me try this way.
Thanks,
Zhao
Re: [PATCH v7 05/16] i386: Decouple CPUID[0x1F] subleaf with specific topology level
In reply to Xiaoyao LiHi Xiaoyao,
Quoted 114 lines · Xiaoyao Li
Date: Thu, 11 Jan 2024 11:19:34 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 05/16] i386: Decouple CPUID[0x1F] subleaf with
specific topology levelQuoted 87 lines · Zhao Liu
From: Zhao Liu <[email protected]>
At present, the subleaf
0x02of CPUID[0x1F] is bound to the "die" level.In fact, the specific topology level exposed in
0x1Fdepends on the platform's support for extension levels (module, tile and die).To help expose "module" level in
0x1F, decouple CPUID[0x1F] subleaf with specific topology level.Signed-off-by Zhao Liu <[email protected]>Tested-by Babu Moger <[email protected]>Tested-by Yongwei Ma <[email protected]>---
Changes since v3:
* New patch to prepare to expose module level in0x1F.
* Move the CPUTopoLevel enumeration definition from "i386: Add cache topology info in CPUCacheInfo" to this patch. Note, to align with topology types in SDM, revert the name ofCPU_TOPO_LEVEL_UNKNOWtoCPU_TOPO_LEVEL_INVALID.--- target/i386/cpu.c | 136 +++++++++++++++++++++++++++++++++++++--------- target/i386/cpu.h | 15 +++++ 2 files changed, 126 insertions(+), 25 deletions(-)target/i386/cpu.c +54
-0@@ -269,6 +269,116 @@ static void encode_cache_cpuid4(CPUCacheInfo *cache, (cache->complex_indexing ?CACHE_COMPLEX_IDX :0 ); } +static uint32_t num_cpus_by_topo_level (X86CPUTopoInfo *topo_info, +enum CPUTopoLevel topo_level) +{ +switch (topo_level) { +case CPU_TOPO_LEVEL_SMT : +return 1 ; +case CPU_TOPO_LEVEL_CORE : +return topo_info->threads_per_core; +case CPU_TOPO_LEVEL_DIE : +return topo_info->threads_per_core * topo_info->cores_per_die; +case CPU_TOPO_LEVEL_PACKAGE : +return topo_info->threads_per_core * topo_info->cores_per_die * + topo_info->dies_per_pkg; +default : +g_assert_not_reached (); + } +return 0 ; +} + +static uint32_t apicid_offset_by_topo_level (X86CPUTopoInfo *topo_info, +enum CPUTopoLevel topo_level) +{ +switch (topo_level) { +case CPU_TOPO_LEVEL_SMT : +return 0 ; +case CPU_TOPO_LEVEL_CORE : +return apicid_core_offset (topo_info); +case CPU_TOPO_LEVEL_DIE : +return apicid_die_offset (topo_info); +case CPU_TOPO_LEVEL_PACKAGE : +return apicid_pkg_offset (topo_info); +default : +g_assert_not_reached (); + } +return 0 ; +} + +static uint32_t cpuid1f_topo_type (enum CPUTopoLevel topo_level) +{ +switch (topo_level) { +case CPU_TOPO_LEVEL_INVALID : +return CPUID_1F_ECX_TOPO_LEVEL_INVALID ; +case CPU_TOPO_LEVEL_SMT : +return CPUID_1F_ECX_TOPO_LEVEL_SMT ; +case CPU_TOPO_LEVEL_CORE : +return CPUID_1F_ECX_TOPO_LEVEL_CORE ; +case CPU_TOPO_LEVEL_DIE : +return CPUID_1F_ECX_TOPO_LEVEL_DIE ; +default : +/* Other types are not supported in QEMU. */ +g_assert_not_reached (); + } +return 0 ; +}target/i386/cpu.c static void encode_cache_cpuid4(CPUCacheInfo *cache, · line 334
Quoted 10 lines
target/i386/cpu.c +10
-0+ +static void encode_topo_cpuid1f (CPUX86State *env,uint32_t count, +X86CPUTopoInfo *topo_info, +uint32_t *eax,uint32_t *ebx, +uint32_t *ecx,uint32_t *edx) +{ + static DECLARE_BITMAP(topo_bitmap, CPU_TOPO_LEVEL_MAX); +X86CPU *cpu =env_archcpu (env); +unsigned long level, next_level; +uint32_t num_cpus_next_level, offset_next_level;again, I dislike the name of cpus to represent the logical process or thread. we can call it,
num_lps_next_level, ornum_threads_next_level;target/i386/cpu.c static void encode_cache_cpuid4(CPUCacheInfo *cache, · line 340
Quoted 6 lines
target/i386/cpu.c +6
-0+ +/* +* Initialize the bitmap to decide which levels should be +* encoded in 0x1f. +*/ +if (!count) {
Quoted 5 lines
using static bitmap and initialize the bitmap on (count == 0), looks bad to me. It highly relies on the order of how
encode_topo_cpuid1f()is called, and fragile.Instead, we can maintain an array in CPUX86State, e.g.,
In my practice, I have found this way to be rather tricky since...
target/i386/cpu.h typedef struct CPUArchState { · line 1910
Quoted 11 lines
target/i386/cpu.h +2
-0@@ -1904,6 +1904,8 @@ typedef struct CPUArchState {/* Number of dies within this CPU package. */ unsigned nr_dies; + +unint8_t valid_cpu_topo[CPU_TOPO_LEVEL_MAX ]; } CPUX86State;
this array actually pre-binds the 0x1f subleaf to the topology level, so this way brings difficulties to the array initialization stage...
Quoted 8 lines
and initialize it as below, when initializing the env
env->valid_cpu_topo[0] =
CPU_TOPO_LEVEL_SMT;
env->valid_cpu_topo[1] =CPU_TOPO_LEVEL_CORE;if (env->nr_dies > 1) { env->valid_cpu_topo[2] = CPU_TOPO_LEVEL_DIE; }
... as here.
Based on 1f encoding rule, with module, we may need this logic like this:
// If there's module, encode the module level at ECX=2.
if (env->nr_modules > 1) {
env->valid_cpu_topo[2] = CPU_TOPO_LEVEL_MODULE;
if (env->nr_dies > 1) {
env->valid_cpu_topo[3] = CPU_TOPO_LEVEL_DIE;
}
} else if (env->nr_dies > 1) { // Otherwise, encode die directly.
env->valid_cpu_topo[2] = CPU_TOPO_LEVEL_DIE;
}Such case-by-case checking lacks scalability, and if more levels are supported in the future, such as tiles, the whole checking becomes unclean. Am I understanding you correctly?
About the static bitmap, declaring it as static is an optimization. Because the count (ECX, e.g., ECX=N) means the Nth topology levels, if we didn't use static virable, we would need to iterate each time to find the Nth level.
Since we know that the subleaf of 0x1f must be sequentially encoded, the logic of this static code is always in effect.
What do you think?
Thanks,
Zhao
Quoted 108 lines
then in
encode_topo_cpuid1f(), we can get level andnext_levelaslevel = env->valid_cpu_topo[count];
next_level= env->valid_cpu_topo[count + 1];Quoted 100 lines
target/i386/cpu.c +41
-25+/* SMT and core levels are exposed in 0x1f leaf by default. */ +set_bit (CPU_TOPO_LEVEL_SMT , topo_bitmap); +set_bit (CPU_TOPO_LEVEL_CORE , topo_bitmap); + +if (env->nr_dies >1 ) { +set_bit (CPU_TOPO_LEVEL_DIE , topo_bitmap); + } + } + + *ecx = count &0xff ; + *edx = cpu->apic_id; + + level =find_first_bit (topo_bitmap,CPU_TOPO_LEVEL_MAX ); +if (level ==CPU_TOPO_LEVEL_MAX ) { + num_cpus_next_level =0 ; + offset_next_level =0 ; + +/* Encode CPU_TOPO_LEVEL_INVALID into the last subleaf of 0x1f. */ + level =CPU_TOPO_LEVEL_INVALID ; + }else { + next_level =find_next_bit (topo_bitmap,CPU_TOPO_LEVEL_MAX , level +1 ); +if (next_level ==CPU_TOPO_LEVEL_MAX ) { + next_level =CPU_TOPO_LEVEL_PACKAGE ; + } + + num_cpus_next_level =num_cpus_by_topo_level (topo_info, next_level); + offset_next_level =apicid_offset_by_topo_level (topo_info, next_level); + } + + *eax = offset_next_level; + *ebx = num_cpus_next_level; + *ecx |=cpuid1f_topo_type (level) <<8 ; + +assert (!(*eax & ~0x1f )); + *ebx &=0xffff ;/* The count doesn't need to be reliable. */ +if (level !=CPU_TOPO_LEVEL_MAX ) { +clear_bit (level, topo_bitmap); + } +} +/* Encode cache info for CPUID[0x80000005].ECX or CPUID[0x80000005].EDX */ static uint32_t encode_cache_cpuid80000005 (CPUCacheInfo *cache) {@@ -6284,31 +6394,7 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,break ; } - *ecx = count &0xff ; - *edx = cpu->apic_id; -switch (count) { -case 0 : - *eax =apicid_core_offset (&topo_info); - *ebx = topo_info.threads_per_core; - *ecx |=CPUID_1F_ECX_TOPO_LEVEL_SMT <<8 ; -break ; -case 1 : - *eax =apicid_die_offset (&topo_info); - *ebx = topo_info.cores_per_die * topo_info.threads_per_core; - *ecx |=CPUID_1F_ECX_TOPO_LEVEL_CORE <<8 ; -break ; -case 2 : - *eax =apicid_pkg_offset (&topo_info); - *ebx = cpus_per_pkg; - *ecx |=CPUID_1F_ECX_TOPO_LEVEL_DIE <<8 ; -break ; -default : - *eax =0 ; - *ebx =0 ; - *ecx |=CPUID_1F_ECX_TOPO_LEVEL_INVALID <<8 ; - } -assert (!(*eax & ~0x1f )); - *ebx &=0xffff ;/* The count doesn't need to be reliable. */ +encode_topo_cpuid1f (env, count, &topo_info, eax, ebx, ecx, edx);break ;case 0xD : {/* Processor Extended State */ target/i386/cpu.h +15
-0@@ -1008,6 +1008,21 @@ uint64_t x86_cpu_get_supported_feature_word(FeatureWord w,#define CPUID_MWAIT_IBE (1U << 1)/* Interrupts can exit capability */ #define CPUID_MWAIT_EMX (1U << 0)/* enumeration supported */ +/* +* CPUTopoLevel is the general i386 topology hierarchical representation, +* ordered by increasing hierarchical relationship. +* Its enumeration value is not bound to the type value of Intel (CPUID[0x1F]) +* or AMD (CPUID[0x80000026]). +*/ +enum CPUTopoLevel { +CPU_TOPO_LEVEL_INVALID , +CPU_TOPO_LEVEL_SMT , +CPU_TOPO_LEVEL_CORE , +CPU_TOPO_LEVEL_DIE , +CPU_TOPO_LEVEL_PACKAGE , +CPU_TOPO_LEVEL_MAX , +}; +/* CPUID[0xB].ECX level types */ #define CPUID_B_ECX_TOPO_LEVEL_INVALID 0#define CPUID_B_ECX_TOPO_LEVEL_SMT 1
PATCH v7 6/16 i386: Introduce module-level cpu topology to CPUX86State
In reply to Zhao LiuFrom: Zhuocheng Ding <[email protected]>
smp command has the "clusters" parameter but x86 hasn't supported that level. "cluster" is a CPU topology level concept above cores, in which the cores may share some resources (L2 cache or some others like L3 cache tags, depending on the Archs) [1][2]. For x86, the resource shared by cores at the cluster level is mainly the L2 cache.
However, using cluster to define x86's L2 cache topology will cause the compatibility problem:
Currently, x86 defaults that the L2 cache is shared in one core, which actually implies a default setting "cores per L2 cache is 1" and therefore implicitly defaults to having as many L2 caches as cores.
For example (i386 PC machine):
-smp 16,sockets=2,dies=2,cores=2,threads=2,maxcpus=16 (*)
Considering the topology of the L2 cache, this (*) implicitly means "1 core per L2 cache" and "2 L2 caches per die".
If we use cluster to configure L2 cache topology with the new default setting "clusters per L2 cache is 1", the above semantics will change to "2 cores per cluster" and "1 cluster per L2 cache", that is, "2 cores per L2 cache".
So the same command (*) will cause changes in the L2 cache topology, further affecting the performance of the virtual machine.
Therefore, x86 should only treat cluster as a cpu topology level and avoid using it to change L2 cache by default for compatibility.
"cluster" in smp is the CPU topology level which is between "core" and die.
For x86, the "cluster" in smp is corresponding to the module level [2], which is above the core level. So use the "module" other than "cluster" in i386 code.
And please note that x86 already has a cpu topology level also named "cluster" [3], this level is at the upper level of the package. Here, the cluster in x86 cpu topology is completely different from the "clusters" as the smp parameter. After the module level is introduced, the cluster as the smp parameter will actually refer to the module level of x86.
[1]: 864c3b5c32f0 ("hw/core/machine: Introduce CPU cluster topology support") [2]: Yanan's comment about "cluster",
https://lists.gnu.org/archive/html/qemu-devel/2023-02/msg04051.html
[3]: SDM, vol.3, ch.9, 9.9.1 Hierarchical Mapping of Shared Resources.
---
Changes since v1:
* The background of the introduction of the "cluster" parameter and its exact meaning were revised according to Yanan's explanation. (Yanan)
--- hw/i386/x86.c | 1 + target/i386/cpu.c | 1 + target/i386/cpu.h | 5 +++++ 3 files changed, 7 insertions(+)
hw/i386/x86.c +1 -0
@@ -310,6 +310,7 @@ void x86_cpu_pre_plug(HotplugHandler *hotplug_dev,init_topo_info (&topo_info, x86ms); env->nr_dies = ms->smp.dies; + env->nr_modules = ms->smp.clusters;/* * If APIC ID is not set,
target/i386/cpu.c +1 -0
@@ -7699,6 +7699,7 @@ static void x86_cpu_initfn(Object *obj)CPUX86State *env = &cpu->env; env->nr_dies =1 ; + env->nr_modules =1 ; object_property_add(obj, "feature-words", "X86CPUFeatureWordInfo", x86_cpu_get_feature_words,
target/i386/cpu.h +5 -0
@@ -1904,6 +1904,11 @@ typedef struct CPUArchState {/* Number of dies within this CPU package. */ unsigned nr_dies; +/* +* Number of modules within this CPU package. +* Module level in x86 cpu topology is corresponding to smp.clusters. +*/ +unsigned nr_modules; } CPUX86State;struct kvm_msrs ;
Signature · 1 line
2.34.1
PATCH v7 7/16 i386: Support modules_per_die in X86CPUTopoInfo
In reply to Zhao LiuFrom: Zhuocheng Ding <[email protected]>
Support module level in i386 cpu topology structure "X86CPUTopoInfo".
Since x86 does not yet support the "clusters" parameter in "-smp", X86CPUTopoInfo.modules_per_die is currently always 1. Therefore, the module level width in APIC ID, which can be calculated by "apicid_bitwidth_for_count(topo_info->modules_per_die)", is always 0 for now, so we can directly add APIC ID related helpers to support module level parsing.
In addition, update topology structure in test-x86-topo.c.
---
Changes since v3:
* Drop the description about not exposing module level in commit message.
* Update topology related calculation in newly added helpers:
num_cpus_by_topo_level() and apicid_offset_by_topo_level().
Changes since v1:
* Include module level related helpers (apicid_module_width() and apicid_module_offset()) in this patch. (Yanan)
--- hw/i386/x86.c | 3 ++- include/hw/i386/topology.h | 22 +++++++++++++++---- target/i386/cpu.c | 17 +++++++++----- tests/unit/test-x86-topo.c | 45 ++++++++++++++++++++------------------ 4 files changed, 55 insertions(+), 32 deletions(-)
hw/i386/x86.c +2 -1
@@ -72,7 +72,8 @@ static void init_topo_info(X86CPUTopoInfo *topo_info,MachineState *ms =MACHINE (x86ms); topo_info->dies_per_pkg = ms->smp.dies; - topo_info->cores_per_die = ms->smp.cores; + topo_info->modules_per_die = ms->smp.clusters; + topo_info->cores_per_module = ms->smp.cores; topo_info->threads_per_core = ms->smp.threads; }
include/hw/i386/topology.h +18 -4
@@ -56,7 +56,8 @@ typedef struct X86CPUTopoIDs {typedef struct X86CPUTopoInfo {unsigned dies_per_pkg; -unsigned cores_per_die; +unsigned modules_per_die; +unsigned cores_per_module;unsigned threads_per_core; }X86CPUTopoInfo ;
@@ -77,7 +78,13 @@ static inline unsigned apicid_smt_width(X86CPUTopoInfo *topo_info)/* Bit width of the Core_ID field */ static inline unsigned apicid_core_width (X86CPUTopoInfo *topo_info) { -return apicid_bitwidth_for_count (topo_info->cores_per_die); +return apicid_bitwidth_for_count (topo_info->cores_per_module); +} + +/* Bit width of the Module_ID (cluster ID) field */ +static inline unsigned apicid_module_width (X86CPUTopoInfo *topo_info) +{ +return apicid_bitwidth_for_count (topo_info->modules_per_die); }/* Bit width of the Die_ID field */
@@ -92,10 +99,16 @@ static inline unsigned apicid_core_offset(X86CPUTopoInfo *topo_info)return apicid_smt_width (topo_info); } +/* Bit offset of the Module_ID (cluster ID) field */ +static inline unsigned apicid_module_offset (X86CPUTopoInfo *topo_info) +{ +return apicid_core_offset (topo_info) +apicid_core_width (topo_info); +} +/* Bit offset of the Die_ID field */ static inline unsigned apicid_die_offset (X86CPUTopoInfo *topo_info) { -return apicid_core_offset (topo_info) +apicid_core_width (topo_info); +return apicid_module_offset (topo_info) +apicid_module_width (topo_info); }/* Bit offset of the Pkg_ID (socket ID) field */
@@ -127,7 +140,8 @@ static inline void x86_topo_ids_from_idx(X86CPUTopoInfo *topo_info, X86CPUTopoIDs *topo_ids) {unsigned nr_dies = topo_info->dies_per_pkg; -unsigned nr_cores = topo_info->cores_per_die; +unsigned nr_cores = topo_info->cores_per_module * + topo_info->modules_per_die;unsigned nr_threads = topo_info->threads_per_core; topo_ids->pkg_id = cpu_index / (nr_dies * nr_cores * nr_threads);
target/i386/cpu.c +11 -6
@@ -278,10 +278,11 @@ static uint32_t num_cpus_by_topo_level(X86CPUTopoInfo *topo_info,case CPU_TOPO_LEVEL_CORE :return topo_info->threads_per_core;case CPU_TOPO_LEVEL_DIE : -return topo_info->threads_per_core * topo_info->cores_per_die; +return topo_info->threads_per_core * topo_info->cores_per_module * + topo_info->modules_per_die;case CPU_TOPO_LEVEL_PACKAGE : -return topo_info->threads_per_core * topo_info->cores_per_die * - topo_info->dies_per_pkg; +return topo_info->threads_per_core * topo_info->cores_per_module * + topo_info->modules_per_die * topo_info->dies_per_pkg;default :g_assert_not_reached (); }
@@ -450,7 +451,9 @@ static void encode_cache_cpuid8000001d(CPUCacheInfo *cache,/* L3 is shared among multiple cores */ if (cache->level ==3 ) { - l3_threads = topo_info->cores_per_die * topo_info->threads_per_core; + l3_threads = topo_info->modules_per_die * + topo_info->cores_per_module * + topo_info->threads_per_core; *eax |= (l3_threads -1 ) <<14 ; }else { *eax |= ((topo_info->threads_per_core -1 ) <<14 );
@@ -6131,10 +6134,12 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,uint32_t cpus_per_pkg; topo_info.dies_per_pkg = env->nr_dies; - topo_info.cores_per_die = cs->nr_cores / env->nr_dies; + topo_info.modules_per_die = env->nr_modules; + topo_info.cores_per_module = cs->nr_cores / env->nr_dies / env->nr_modules; topo_info.threads_per_core = cs->nr_threads; - cores_per_pkg = topo_info.cores_per_die * topo_info.dies_per_pkg; + cores_per_pkg = topo_info.cores_per_module * topo_info.modules_per_die * + topo_info.dies_per_pkg; cpus_per_pkg = cores_per_pkg * topo_info.threads_per_core;/* Calculate & apply limits for different index ranges */
tests/unit/test-x86-topo.c +24 -21
94 lines not on this page · raw message
Signature · 1 line
2.34.1
Re: [PATCH v7 07/16] i386: Support modules_per_die in X86CPUTopoInfo
Quoted 139 lines · Zhao Liu
From: Zhuocheng Ding <[email protected]>
Support module level in i386 cpu topology structure "X86CPUTopoInfo".
Since x86 does not yet support the "clusters" parameter in "-smp", X86CPUTopoInfo.modules_per_die is currently always 1. Therefore, the module level width in APIC ID, which can be calculated by "
apicid_bitwidth_for_count(topo_info->modules_per_die)", is always 0 for now, so we can directly add APIC ID related helpers to support module level parsing.In addition, update topology structure in
test-x86-topo.c.Signed-off-by Zhuocheng Ding <[email protected]>Co-developed-by Zhao Liu <[email protected]>Signed-off-by Zhao Liu <[email protected]>Tested-by Babu Moger <[email protected]>Tested-by Yongwei Ma <[email protected]>---
Changes since v3:
* Drop the description about not exposing module level in commit message.
* Update topology related calculation in newly added helpers:
num_cpus_by_topo_level()andapicid_offset_by_topo_level().Changes since v1:
* Include module level related helpers (apicid_module_width()andapicid_module_offset()) in this patch. (Yanan)--- hw/i386/x86.c | 3 ++- include/hw/i386/topology.h | 22 +++++++++++++++---- target/i386/cpu.c | 17 +++++++++----- tests/unit/test-x86-topo.c | 45 ++++++++++++++++++++------------------ 4 files changed, 55 insertions(+), 32 deletions(-)hw/i386/x86.c +2
-1@@ -72,7 +72,8 @@ static void init_topo_info(X86CPUTopoInfo *topo_info,MachineState *ms =MACHINE (x86ms); topo_info->dies_per_pkg = ms->smp.dies; - topo_info->cores_per_die = ms->smp.cores; + topo_info->modules_per_die = ms->smp.clusters; + topo_info->cores_per_module = ms->smp.cores; topo_info->threads_per_core = ms->smp.threads; }include/hw/i386/topology.h +18
-4@@ -56,7 +56,8 @@ typedef struct X86CPUTopoIDs {typedef struct X86CPUTopoInfo {unsigned dies_per_pkg; -unsigned cores_per_die; +unsigned modules_per_die; +unsigned cores_per_module;unsigned threads_per_core; }X86CPUTopoInfo ;@@ -77,7 +78,13 @@ static inline unsigned apicid_smt_width(X86CPUTopoInfo *topo_info)/* Bit width of the Core_ID field */ static inline unsigned apicid_core_width (X86CPUTopoInfo *topo_info) { -return apicid_bitwidth_for_count (topo_info->cores_per_die); +return apicid_bitwidth_for_count (topo_info->cores_per_module); +} + +/* Bit width of the Module_ID (cluster ID) field */ +static inline unsigned apicid_module_width (X86CPUTopoInfo *topo_info) +{ +return apicid_bitwidth_for_count (topo_info->modules_per_die); }/* Bit width of the Die_ID field */ @@ -92,10 +99,16 @@ static inline unsigned apicid_core_offset(X86CPUTopoInfo *topo_info)return apicid_smt_width (topo_info); } +/* Bit offset of the Module_ID (cluster ID) field */ +static inline unsigned apicid_module_offset (X86CPUTopoInfo *topo_info) +{ +return apicid_core_offset (topo_info) +apicid_core_width (topo_info); +} +/* Bit offset of the Die_ID field */ static inline unsigned apicid_die_offset (X86CPUTopoInfo *topo_info) { -return apicid_core_offset (topo_info) +apicid_core_width (topo_info); +return apicid_module_offset (topo_info) +apicid_module_width (topo_info); }/* Bit offset of the Pkg_ID (socket ID) field */ @@ -127,7 +140,8 @@ static inline void x86_topo_ids_from_idx(X86CPUTopoInfo *topo_info, X86CPUTopoIDs *topo_ids) {unsigned nr_dies = topo_info->dies_per_pkg; -unsigned nr_cores = topo_info->cores_per_die; +unsigned nr_cores = topo_info->cores_per_module * + topo_info->modules_per_die;unsigned nr_threads = topo_info->threads_per_core; topo_ids->pkg_id = cpu_index / (nr_dies * nr_cores * nr_threads);target/i386/cpu.c +7
-4@@ -278,10 +278,11 @@ static uint32_t num_cpus_by_topo_level(X86CPUTopoInfo *topo_info,case CPU_TOPO_LEVEL_CORE :return topo_info->threads_per_core;case CPU_TOPO_LEVEL_DIE : -return topo_info->threads_per_core * topo_info->cores_per_die; +return topo_info->threads_per_core * topo_info->cores_per_module * + topo_info->modules_per_die;case CPU_TOPO_LEVEL_PACKAGE : -return topo_info->threads_per_core * topo_info->cores_per_die * - topo_info->dies_per_pkg; +return topo_info->threads_per_core * topo_info->cores_per_module * + topo_info->modules_per_die * topo_info->dies_per_pkg;default :g_assert_not_reached (); }@@ -450,7 +451,9 @@ static void encode_cache_cpuid8000001d(CPUCacheInfo *cache,/* L3 is shared among multiple cores */ if (cache->level ==3 ) { - l3_threads = topo_info->cores_per_die * topo_info->threads_per_core; + l3_threads = topo_info->modules_per_die * + topo_info->cores_per_module * + topo_info->threads_per_core; *eax |= (l3_threads -1 ) <<14 ; }else { *eax |= ((topo_info->threads_per_core -1 ) <<14 );@@ -6131,10 +6134,12 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,uint32_t cpus_per_pkg;
target/i386/cpu.c void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count, · line 6142
Quoted 10 lines
target/i386/cpu.c +4
-2topo_info.dies_per_pkg = env->nr_dies; - topo_info.cores_per_die = cs->nr_cores / env->nr_dies; + topo_info.modules_per_die = env->nr_modules; + topo_info.cores_per_module = cs->nr_cores / env->nr_dies / env->nr_modules; topo_info.threads_per_core = cs->nr_threads; - cores_per_pkg = topo_info.cores_per_die * topo_info.dies_per_pkg; + cores_per_pkg = topo_info.cores_per_module * topo_info.modules_per_die * + topo_info.dies_per_pkg;
Nit. maybe we can introduce some helper function like
static inline uint32_t topo_info_cores_per_pkg(X86CPUTopoInfo *topo_info) {
return topo_info.cores_per_module * topo_info.modules_per_die *
topo_info.dies_per_pkg;
}so we don't need to care how it calculates.
Besides,
Quoted 104 lines
target/i386/cpu.c +0
-0cpus_per_pkg = cores_per_pkg * topo_info.threads_per_core;/* Calculate & apply limits for different index ranges */ tests/unit/test-x86-topo.c +24
-21@@ -30,13 +30,16 @@ static void test_topo_bits(void) {X86CPUTopoInfo topo_info = {0 }; -/* simple tests for 1 thread per core, 1 core per die, 1 die per package */ - topo_info = (X86CPUTopoInfo ) {1 ,1 ,1 }; +/* +* simple tests for 1 thread per core, 1 core per module, +* 1 module per die, 1 die per package +*/ + topo_info = (X86CPUTopoInfo ) {1 ,1 ,1 ,1 }; g_assert_cmpuint(apicid_smt_width(&topo_info), ==, 0); g_assert_cmpuint(apicid_core_width(&topo_info), ==, 0); g_assert_cmpuint(apicid_die_width(&topo_info), ==, 0); - topo_info = (X86CPUTopoInfo ) {1 ,1 ,1 }; + topo_info = (X86CPUTopoInfo ) {1 ,1 ,1 ,1 }; g_assert_cmpuint(x86_apicid_from_cpu_idx(&topo_info, 0), ==, 0); g_assert_cmpuint(x86_apicid_from_cpu_idx(&topo_info, 1), ==, 1); g_assert_cmpuint(x86_apicid_from_cpu_idx(&topo_info, 2), ==, 2);@@ -45,39 +48,39 @@ static void test_topo_bits(void)/* Test field width calculation for multiple values */ - topo_info = (X86CPUTopoInfo ) {1 ,1 ,2 }; + topo_info = (X86CPUTopoInfo ) {1 ,1 ,1 ,2 }; g_assert_cmpuint(apicid_smt_width(&topo_info), ==, 1); - topo_info = (X86CPUTopoInfo ) {1 ,1 ,3 }; + topo_info = (X86CPUTopoInfo ) {1 ,1 ,1 ,3 }; g_assert_cmpuint(apicid_smt_width(&topo_info), ==, 2); - topo_info = (X86CPUTopoInfo ) {1 ,1 ,4 }; + topo_info = (X86CPUTopoInfo ) {1 ,1 ,1 ,4 }; g_assert_cmpuint(apicid_smt_width(&topo_info), ==, 2); - topo_info = (X86CPUTopoInfo ) {1 ,1 ,14 }; + topo_info = (X86CPUTopoInfo ) {1 ,1 ,1 ,14 }; g_assert_cmpuint(apicid_smt_width(&topo_info), ==, 4); - topo_info = (X86CPUTopoInfo ) {1 ,1 ,15 }; + topo_info = (X86CPUTopoInfo ) {1 ,1 ,1 ,15 }; g_assert_cmpuint(apicid_smt_width(&topo_info), ==, 4); - topo_info = (X86CPUTopoInfo ) {1 ,1 ,16 }; + topo_info = (X86CPUTopoInfo ) {1 ,1 ,1 ,16 }; g_assert_cmpuint(apicid_smt_width(&topo_info), ==, 4); - topo_info = (X86CPUTopoInfo ) {1 ,1 ,17 }; + topo_info = (X86CPUTopoInfo ) {1 ,1 ,1 ,17 }; g_assert_cmpuint(apicid_smt_width(&topo_info), ==, 5); - topo_info = (X86CPUTopoInfo ) {1 ,30 ,2 }; + topo_info = (X86CPUTopoInfo ) {1 ,1 ,30 ,2 }; g_assert_cmpuint(apicid_core_width(&topo_info), ==, 5); - topo_info = (X86CPUTopoInfo ) {1 ,31 ,2 }; + topo_info = (X86CPUTopoInfo ) {1 ,1 ,31 ,2 }; g_assert_cmpuint(apicid_core_width(&topo_info), ==, 5); - topo_info = (X86CPUTopoInfo ) {1 ,32 ,2 }; + topo_info = (X86CPUTopoInfo ) {1 ,1 ,32 ,2 }; g_assert_cmpuint(apicid_core_width(&topo_info), ==, 5); - topo_info = (X86CPUTopoInfo ) {1 ,33 ,2 }; + topo_info = (X86CPUTopoInfo ) {1 ,1 ,33 ,2 }; g_assert_cmpuint(apicid_core_width(&topo_info), ==, 6); - topo_info = (X86CPUTopoInfo ) {1 ,30 ,2 }; + topo_info = (X86CPUTopoInfo ) {1 ,1 ,30 ,2 }; g_assert_cmpuint(apicid_die_width(&topo_info), ==, 0); - topo_info = (X86CPUTopoInfo ) {2 ,30 ,2 }; + topo_info = (X86CPUTopoInfo ) {2 ,1 ,30 ,2 }; g_assert_cmpuint(apicid_die_width(&topo_info), ==, 1); - topo_info = (X86CPUTopoInfo ) {3 ,30 ,2 }; + topo_info = (X86CPUTopoInfo ) {3 ,1 ,30 ,2 }; g_assert_cmpuint(apicid_die_width(&topo_info), ==, 2); - topo_info = (X86CPUTopoInfo ) {4 ,30 ,2 }; + topo_info = (X86CPUTopoInfo ) {4 ,1 ,30 ,2 }; g_assert_cmpuint(apicid_die_width(&topo_info), ==, 2);/* build a weird topology and see if IDs are calculated correctly @@ -85,18 +88,18 @@ static void test_topo_bits(void)/* This will use 2 bits for thread ID and 3 bits for core ID */ - topo_info = (X86CPUTopoInfo ) {1 ,6 ,3 }; + topo_info = (X86CPUTopoInfo ) {1 ,1 ,6 ,3 }; g_assert_cmpuint(apicid_smt_width(&topo_info), ==, 2); g_assert_cmpuint(apicid_core_offset(&topo_info), ==, 2); g_assert_cmpuint(apicid_die_offset(&topo_info), ==, 5); g_assert_cmpuint(apicid_pkg_offset(&topo_info), ==, 5); - topo_info = (X86CPUTopoInfo ) {1 ,6 ,3 }; + topo_info = (X86CPUTopoInfo ) {1 ,1 ,6 ,3 }; g_assert_cmpuint(x86_apicid_from_cpu_idx(&topo_info, 0), ==, 0); g_assert_cmpuint(x86_apicid_from_cpu_idx(&topo_info, 1), ==, 1); g_assert_cmpuint(x86_apicid_from_cpu_idx(&topo_info, 2), ==, 2); - topo_info = (X86CPUTopoInfo ) {1 ,6 ,3 }; + topo_info = (X86CPUTopoInfo ) {1 ,1 ,6 ,3 }; g_assert_cmpuint(x86_apicid_from_cpu_idx(&topo_info, 1 * 3 + 0), ==, (1 <<2 ) |0 ); g_assert_cmpuint(x86_apicid_from_cpu_idx(&topo_info, 1 * 3 + 1), ==,
Re: [PATCH v7 07/16] i386: Support modules_per_die in X86CPUTopoInfo
Hi Xiaoyao,
Quoted 4 lines · Xiaoyao Li
target/i386/cpu.c void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count, · line 6142
Quoted 3 lines
target/i386/cpu.c +2
-1- cores_per_pkg = topo_info.cores_per_die * topo_info.dies_per_pkg; + cores_per_pkg = topo_info.cores_per_module * topo_info.modules_per_die * + topo_info.dies_per_pkg;
Quoted 8 lines
Nit. maybe we can introduce some helper function like
static inline uint32_t topo_info_cores_per_pkg(X86CPUTopoInfo *topo_info) { return topo_info.cores_per_module * topo_info.modules_per_die * topo_info.dies_per_pkg; }so we don't need to care how it calculates.
Yeah, will add this helper, maybe in another patch.
Quoted 4 lines
Besides,
Reviewed-by Xiaoyao Li <[email protected]>
Thanks!
-Zhao
PATCH v7 8/16 i386: Expose module level in CPUID[0x1F]
In reply to Zhao LiuFrom: Zhao Liu <[email protected]>
Linux kernel (from v6.4, with commit edc0a2b595765 ("x86/topology: Fix erroneous smp_num_siblings on Intel Hybrid platforms") is able to handle platforms with Module level enumerated via CPUID.1F.
Expose the module level in CPUID[0x1F] if the machine has more than 1 modules.
(Tested CPU topology in CPUID[0x1F] leaf with various die/cluster configurations in "-smp".)
---
Changes since v3:
* New patch to expose module level in 0x1F.
* Add Tested-by tag from Yongwei.
--- target/i386/cpu.c | 12 +++++++++++- target/i386/cpu.h | 2 ++ target/i386/kvm/kvm.c | 2 +- 3 files changed, 14 insertions(+), 2 deletions(-)
target/i386/cpu.c +11 -1
@@ -277,6 +277,8 @@ static uint32_t num_cpus_by_topo_level(X86CPUTopoInfo *topo_info,return 1 ;case CPU_TOPO_LEVEL_CORE :return topo_info->threads_per_core; +case CPU_TOPO_LEVEL_MODULE : +return topo_info->threads_per_core * topo_info->cores_per_module;case CPU_TOPO_LEVEL_DIE :return topo_info->threads_per_core * topo_info->cores_per_module * topo_info->modules_per_die;
@@ -297,6 +299,8 @@ static uint32_t apicid_offset_by_topo_level(X86CPUTopoInfo *topo_info,return 0 ;case CPU_TOPO_LEVEL_CORE :return apicid_core_offset (topo_info); +case CPU_TOPO_LEVEL_MODULE : +return apicid_module_offset (topo_info);case CPU_TOPO_LEVEL_DIE :return apicid_die_offset (topo_info);case CPU_TOPO_LEVEL_PACKAGE :
@@ -316,6 +320,8 @@ static uint32_t cpuid1f_topo_type(enum CPUTopoLevel topo_level)return CPUID_1F_ECX_TOPO_LEVEL_SMT ;case CPU_TOPO_LEVEL_CORE :return CPUID_1F_ECX_TOPO_LEVEL_CORE ; +case CPU_TOPO_LEVEL_MODULE : +return CPUID_1F_ECX_TOPO_LEVEL_MODULE ;case CPU_TOPO_LEVEL_DIE :return CPUID_1F_ECX_TOPO_LEVEL_DIE ;default :
@@ -347,6 +353,10 @@ static void encode_topo_cpuid1f(CPUX86State *env, uint32_t count,if (env->nr_dies >1 ) {set_bit (CPU_TOPO_LEVEL_DIE , topo_bitmap); } + +if (env->nr_modules >1 ) { +set_bit (CPU_TOPO_LEVEL_MODULE , topo_bitmap); + } } *ecx = count &0xff ;
@@ -6394,7 +6404,7 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,break ;case 0x1F :/* V2 Extended Topology Enumeration Leaf */ -if (topo_info.dies_per_pkg <2 ) { +if (topo_info.modules_per_die <2 && topo_info.dies_per_pkg <2 ) { *eax = *ebx = *ecx = *edx =0 ;break ; }
target/i386/cpu.h +2 -0
@@ -1018,6 +1018,7 @@ enum CPUTopoLevel {CPU_TOPO_LEVEL_INVALID ,CPU_TOPO_LEVEL_SMT ,CPU_TOPO_LEVEL_CORE , +CPU_TOPO_LEVEL_MODULE ,CPU_TOPO_LEVEL_DIE ,CPU_TOPO_LEVEL_PACKAGE , CPU_TOPO_LEVEL_MAX,
@@ -1032,6 +1033,7 @@ enum CPUTopoLevel {#define CPUID_1F_ECX_TOPO_LEVEL_INVALID CPUID_B_ECX_TOPO_LEVEL_INVALID#define CPUID_1F_ECX_TOPO_LEVEL_SMT CPUID_B_ECX_TOPO_LEVEL_SMT#define CPUID_1F_ECX_TOPO_LEVEL_CORE CPUID_B_ECX_TOPO_LEVEL_CORE +#define CPUID_1F_ECX_TOPO_LEVEL_MODULE 3#define CPUID_1F_ECX_TOPO_LEVEL_DIE 5/* MSR Feature Bits */
target/i386/kvm/kvm.c +1 -1
@@ -1913,7 +1913,7 @@ int kvm_arch_init_vcpu(CPUState *cs)break ; }case 0x1f : -if (env->nr_dies <2 ) { +if (env->nr_modules <2 && env->nr_dies <2 ) {break ; }/* fallthrough */
Signature · 1 line
2.34.1
Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]
Quoted 69 lines · Zhao Liu
From: Zhao Liu <[email protected]>
Linux kernel (from v6.4, with commit
edc0a2b595765("x86/topology: Fix erroneoussmp_num_siblingson Intel Hybrid platforms") is able to handle platforms with Module level enumerated via CPUID.1F.Expose the module level in CPUID[0x1F] if the machine has more than 1 modules.
(Tested CPU topology in CPUID[0x1F] leaf with various die/cluster configurations in "-smp".)
Signed-off-by Zhao Liu <[email protected]>Tested-by Babu Moger <[email protected]>Tested-by Yongwei Ma <[email protected]>---
Changes since v3:
* New patch to expose module level in0x1F.
* Add Tested-by tag from Yongwei.--- target/i386/cpu.c | 12 +++++++++++- target/i386/cpu.h | 2 ++ target/i386/kvm/kvm.c | 2 +- 3 files changed, 14 insertions(+), 2 deletions(-)target/i386/cpu.c +10
-0@@ -277,6 +277,8 @@ static uint32_t num_cpus_by_topo_level(X86CPUTopoInfo *topo_info,return 1 ;case CPU_TOPO_LEVEL_CORE :return topo_info->threads_per_core; +case CPU_TOPO_LEVEL_MODULE : +return topo_info->threads_per_core * topo_info->cores_per_module;case CPU_TOPO_LEVEL_DIE :return topo_info->threads_per_core * topo_info->cores_per_module * topo_info->modules_per_die;@@ -297,6 +299,8 @@ static uint32_t apicid_offset_by_topo_level(X86CPUTopoInfo *topo_info,return 0 ;case CPU_TOPO_LEVEL_CORE :return apicid_core_offset (topo_info); +case CPU_TOPO_LEVEL_MODULE : +return apicid_module_offset (topo_info);case CPU_TOPO_LEVEL_DIE :return apicid_die_offset (topo_info);case CPU_TOPO_LEVEL_PACKAGE :@@ -316,6 +320,8 @@ static uint32_t cpuid1f_topo_type(enum CPUTopoLevel topo_level)return CPUID_1F_ECX_TOPO_LEVEL_SMT ;case CPU_TOPO_LEVEL_CORE :return CPUID_1F_ECX_TOPO_LEVEL_CORE ; +case CPU_TOPO_LEVEL_MODULE : +return CPUID_1F_ECX_TOPO_LEVEL_MODULE ;case CPU_TOPO_LEVEL_DIE :return CPUID_1F_ECX_TOPO_LEVEL_DIE ;default :@@ -347,6 +353,10 @@ static void encode_topo_cpuid1f(CPUX86State *env, uint32_t count,if (env->nr_dies >1 ) {set_bit (CPU_TOPO_LEVEL_DIE , topo_bitmap); } + +if (env->nr_modules >1 ) { +set_bit (CPU_TOPO_LEVEL_MODULE , topo_bitmap); + } } *ecx = count &0xff ;
target/i386/cpu.c void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count, · line 6407
Quoted 5 lines
target/i386/cpu.c +1
-1@@ -6394,7 +6404,7 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,break ;case 0x1F :/* V2 Extended Topology Enumeration Leaf */ -if (topo_info.dies_per_pkg <2 ) { +if (topo_info.modules_per_die <2 && topo_info.dies_per_pkg <2 ) {
maybe we can come up with below function if we have env->valid_cpu_topo[] as I suggested in patch 5.
bool cpu_x86_has_valid_cpuid1f(CPUX86State *env) {
return env->valid_cpu_topo[2] ? true : false;
}...
Quoted 28 lines
*eax = *ebx = *ecx = *edx = 0; break; }target/i386/cpu.h +2
-0@@ -1018,6 +1018,7 @@ enum CPUTopoLevel {CPU_TOPO_LEVEL_INVALID ,CPU_TOPO_LEVEL_SMT ,CPU_TOPO_LEVEL_CORE , +CPU_TOPO_LEVEL_MODULE ,CPU_TOPO_LEVEL_DIE ,CPU_TOPO_LEVEL_PACKAGE , CPU_TOPO_LEVEL_MAX,@@ -1032,6 +1033,7 @@ enum CPUTopoLevel {#define CPUID_1F_ECX_TOPO_LEVEL_INVALID CPUID_B_ECX_TOPO_LEVEL_INVALID#define CPUID_1F_ECX_TOPO_LEVEL_SMT CPUID_B_ECX_TOPO_LEVEL_SMT#define CPUID_1F_ECX_TOPO_LEVEL_CORE CPUID_B_ECX_TOPO_LEVEL_CORE +#define CPUID_1F_ECX_TOPO_LEVEL_MODULE 3#define CPUID_1F_ECX_TOPO_LEVEL_DIE 5/* MSR Feature Bits */ target/i386/kvm/kvm.c +0
-0
target/i386/kvm/kvm.c int kvm_arch_init_vcpu(CPUState *cs) · line 1916
Quoted 5 lines
target/i386/kvm/kvm.c +1
-1@@ -1913,7 +1913,7 @@ int kvm_arch_init_vcpu(CPUState *cs)break ; }case 0x1f : -if (env->nr_dies <2 ) { +if (env->nr_modules <2 && env->nr_dies <2 ) {
then cpu_x86_has_valid_cpuid1f() can be used here.
Quoted 3 lines
target/i386/kvm/kvm.c +0
-0break ; }/* fallthrough */
Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]
Hi Xiaoyao,
Quoted 82 lines · Xiaoyao Li
Date: Thu, 11 Jan 2024 14:04:53 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]Quoted 68 lines · Zhao Liu
From: Zhao Liu <[email protected]>
Linux kernel (from v6.4, with commit
edc0a2b595765("x86/topology: Fix erroneoussmp_num_siblingson Intel Hybrid platforms") is able to handle platforms with Module level enumerated via CPUID.1F.Expose the module level in CPUID[0x1F] if the machine has more than 1 modules.
(Tested CPU topology in CPUID[0x1F] leaf with various die/cluster configurations in "-smp".)
Signed-off-by Zhao Liu <[email protected]>Tested-by Babu Moger <[email protected]>Tested-by Yongwei Ma <[email protected]>---
Changes since v3:
* New patch to expose module level in0x1F.
* Add Tested-by tag from Yongwei.--- target/i386/cpu.c | 12 +++++++++++- target/i386/cpu.h | 2 ++ target/i386/kvm/kvm.c | 2 +- 3 files changed, 14 insertions(+), 2 deletions(-)target/i386/cpu.c +10
-0@@ -277,6 +277,8 @@ static uint32_t num_cpus_by_topo_level(X86CPUTopoInfo *topo_info,return 1 ;case CPU_TOPO_LEVEL_CORE :return topo_info->threads_per_core; +case CPU_TOPO_LEVEL_MODULE : +return topo_info->threads_per_core * topo_info->cores_per_module;case CPU_TOPO_LEVEL_DIE :return topo_info->threads_per_core * topo_info->cores_per_module * topo_info->modules_per_die;@@ -297,6 +299,8 @@ static uint32_t apicid_offset_by_topo_level(X86CPUTopoInfo *topo_info,return 0 ;case CPU_TOPO_LEVEL_CORE :return apicid_core_offset (topo_info); +case CPU_TOPO_LEVEL_MODULE : +return apicid_module_offset (topo_info);case CPU_TOPO_LEVEL_DIE :return apicid_die_offset (topo_info);case CPU_TOPO_LEVEL_PACKAGE :@@ -316,6 +320,8 @@ static uint32_t cpuid1f_topo_type(enum CPUTopoLevel topo_level)return CPUID_1F_ECX_TOPO_LEVEL_SMT ;case CPU_TOPO_LEVEL_CORE :return CPUID_1F_ECX_TOPO_LEVEL_CORE ; +case CPU_TOPO_LEVEL_MODULE : +return CPUID_1F_ECX_TOPO_LEVEL_MODULE ;case CPU_TOPO_LEVEL_DIE :return CPUID_1F_ECX_TOPO_LEVEL_DIE ;default :@@ -347,6 +353,10 @@ static void encode_topo_cpuid1f(CPUX86State *env, uint32_t count,if (env->nr_dies >1 ) {set_bit (CPU_TOPO_LEVEL_DIE , topo_bitmap); } + +if (env->nr_modules >1 ) { +set_bit (CPU_TOPO_LEVEL_MODULE , topo_bitmap); + } } *ecx = count &0xff ;target/i386/cpu.c void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count, · line 6407
Quoted 5 lines
target/i386/cpu.c +1
-1@@ -6394,7 +6404,7 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,break ;case 0x1F :/* V2 Extended Topology Enumeration Leaf */ -if (topo_info.dies_per_pkg <2 ) { +if (topo_info.modules_per_die <2 && topo_info.dies_per_pkg <2 ) {maybe we can come up with below function if we have env->valid_cpu_topo[] as I suggested in patch 5.
Quoted 5 lines
bool cpu_x86_has_valid_cpuid1f(CPUX86State *env) { return env->valid_cpu_topo[2] ? true : false; }...
This makes sense.
Quoted 27 lines
Quoted 27 lines
*eax = *ebx = *ecx = *edx = 0; break; }target/i386/cpu.h +2
-0@@ -1018,6 +1018,7 @@ enum CPUTopoLevel {CPU_TOPO_LEVEL_INVALID ,CPU_TOPO_LEVEL_SMT ,CPU_TOPO_LEVEL_CORE , +CPU_TOPO_LEVEL_MODULE ,CPU_TOPO_LEVEL_DIE ,CPU_TOPO_LEVEL_PACKAGE , CPU_TOPO_LEVEL_MAX,@@ -1032,6 +1033,7 @@ enum CPUTopoLevel {#define CPUID_1F_ECX_TOPO_LEVEL_INVALID CPUID_B_ECX_TOPO_LEVEL_INVALID#define CPUID_1F_ECX_TOPO_LEVEL_SMT CPUID_B_ECX_TOPO_LEVEL_SMT#define CPUID_1F_ECX_TOPO_LEVEL_CORE CPUID_B_ECX_TOPO_LEVEL_CORE +#define CPUID_1F_ECX_TOPO_LEVEL_MODULE 3#define CPUID_1F_ECX_TOPO_LEVEL_DIE 5/* MSR Feature Bits */ target/i386/kvm/kvm.c +0
-0
Quoted 9 lines
target/i386/kvm/kvm.c int kvm_arch_init_vcpu(CPUState *cs) · line 1916
Quoted 6 lines
target/i386/kvm/kvm.c +1
-1@@ -1913,7 +1913,7 @@ int kvm_arch_init_vcpu(CPUState *cs)break ; }case 0x1f : -if (env->nr_dies <2 ) { +if (env->nr_modules <2 && env->nr_dies <2 ) {then
cpu_x86_has_valid_cpuid1f()can be used here.
Good idae, I will also try this.
Thanks,
Zhao
Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]
In reply to Zhao LiuQuoted 69 lines · Zhao Liu
From: Zhao Liu <[email protected]>
Linux kernel (from v6.4, with commit
edc0a2b595765("x86/topology: Fix erroneoussmp_num_siblingson Intel Hybrid platforms") is able to handle platforms with Module level enumerated via CPUID.1F.Expose the module level in CPUID[0x1F] if the machine has more than 1 modules.
(Tested CPU topology in CPUID[0x1F] leaf with various die/cluster configurations in "-smp".)
Signed-off-by Zhao Liu <[email protected]>Tested-by Babu Moger <[email protected]>Tested-by Yongwei Ma <[email protected]>---
Changes since v3:
* New patch to expose module level in0x1F.
* Add Tested-by tag from Yongwei.--- target/i386/cpu.c | 12 +++++++++++- target/i386/cpu.h | 2 ++ target/i386/kvm/kvm.c | 2 +- 3 files changed, 14 insertions(+), 2 deletions(-)target/i386/cpu.c +10
-0@@ -277,6 +277,8 @@ static uint32_t num_cpus_by_topo_level(X86CPUTopoInfo *topo_info,return 1 ;case CPU_TOPO_LEVEL_CORE :return topo_info->threads_per_core; +case CPU_TOPO_LEVEL_MODULE : +return topo_info->threads_per_core * topo_info->cores_per_module;case CPU_TOPO_LEVEL_DIE :return topo_info->threads_per_core * topo_info->cores_per_module * topo_info->modules_per_die;@@ -297,6 +299,8 @@ static uint32_t apicid_offset_by_topo_level(X86CPUTopoInfo *topo_info,return 0 ;case CPU_TOPO_LEVEL_CORE :return apicid_core_offset (topo_info); +case CPU_TOPO_LEVEL_MODULE : +return apicid_module_offset (topo_info);case CPU_TOPO_LEVEL_DIE :return apicid_die_offset (topo_info);case CPU_TOPO_LEVEL_PACKAGE :@@ -316,6 +320,8 @@ static uint32_t cpuid1f_topo_type(enum CPUTopoLevel topo_level)return CPUID_1F_ECX_TOPO_LEVEL_SMT ;case CPU_TOPO_LEVEL_CORE :return CPUID_1F_ECX_TOPO_LEVEL_CORE ; +case CPU_TOPO_LEVEL_MODULE : +return CPUID_1F_ECX_TOPO_LEVEL_MODULE ;case CPU_TOPO_LEVEL_DIE :return CPUID_1F_ECX_TOPO_LEVEL_DIE ;default :@@ -347,6 +353,10 @@ static void encode_topo_cpuid1f(CPUX86State *env, uint32_t count,if (env->nr_dies >1 ) {set_bit (CPU_TOPO_LEVEL_DIE , topo_bitmap); } + +if (env->nr_modules >1 ) { +set_bit (CPU_TOPO_LEVEL_MODULE , topo_bitmap); + } } *ecx = count &0xff ;
target/i386/cpu.c void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count, · line 6407
Quoted 5 lines
target/i386/cpu.c +1
-1@@ -6394,7 +6404,7 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,break ;case 0x1F :/* V2 Extended Topology Enumeration Leaf */ -if (topo_info.dies_per_pkg <2 ) { +if (topo_info.modules_per_die <2 && topo_info.dies_per_pkg <2 ) {
A question:
Is the original checking necessary ?
The 0x1f exists even on cpu w/o modules/dies topology on bare metal, I tried on EMR:
// leaf 00x00000000 0x00: eax=0x00000020 ebx=0x756e6547 ecx=0x6c65746e edx=0x49656e69
// leaf 0x1f0x0000001f 0x00: eax=0x00000001 ebx=0x00000002 ecx=0x00000100 edx=0x00000004 0x0000001f 0x01: eax=0x00000007 ebx=0x00000080 ecx=0x00000201 edx=0x00000004 0x0000001f 0x02: eax=0x00000000 ebx=0x00000000 ecx=0x00000002 edx=0x00000004
// leaf 0xb0x0000000b 0x00: eax=0x00000001 ebx=0x00000002 ecx=0x00000100 edx=0x00000004 0x0000000b 0x01: eax=0x00000007 ebx=0x00000080 ecx=0x00000201 edx=0x00000004 0x0000000b 0x02: eax=0x00000000 ebx=0x00000000 ecx=0x00000002 edx=0x00000004
So here leads to different cpu behavior from bare metal, even in case of "-cpu host".
In SDM Vol2, cpudid instruction section:
" CPUID leaf 1FH is a preferred superset to leaf 0BH. Intel recommends using leaf 1FH when available rather than leaf 0BH and ensuring that any leaf 0BH algorithms are updated to support leaf 1FH. "
My understanding: if 0x1f is existed (leaf 0.eax >= 0x1f) then it should have same values in lp/core level as 0xb.
Quoted 40 lines
target/i386/cpu.c +0
-0*eax = *ebx = *ecx = *edx =0 ;break ; }target/i386/cpu.h +2
-0@@ -1018,6 +1018,7 @@ enum CPUTopoLevel {CPU_TOPO_LEVEL_INVALID ,CPU_TOPO_LEVEL_SMT ,CPU_TOPO_LEVEL_CORE , +CPU_TOPO_LEVEL_MODULE ,CPU_TOPO_LEVEL_DIE ,CPU_TOPO_LEVEL_PACKAGE , CPU_TOPO_LEVEL_MAX,@@ -1032,6 +1033,7 @@ enum CPUTopoLevel {#define CPUID_1F_ECX_TOPO_LEVEL_INVALID CPUID_B_ECX_TOPO_LEVEL_INVALID#define CPUID_1F_ECX_TOPO_LEVEL_SMT CPUID_B_ECX_TOPO_LEVEL_SMT#define CPUID_1F_ECX_TOPO_LEVEL_CORE CPUID_B_ECX_TOPO_LEVEL_CORE +#define CPUID_1F_ECX_TOPO_LEVEL_MODULE 3#define CPUID_1F_ECX_TOPO_LEVEL_DIE 5/* MSR Feature Bits */ target/i386/kvm/kvm.c +1
-2@@ -1913,7 +1913,7 @@ int kvm_arch_init_vcpu(CPUState *cs)break ; }case 0x1f : -if (env->nr_dies <2 ) { +if (env->nr_modules <2 && env->nr_dies <2 ) {break ; }/* fallthrough */ --2.34.1
Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]
Hi Yuan,
Quoted 93 lines · Yuan Yao
Date: Mon, 15 Jan 2024 11:25:24 +0800
From: Yuan Yao <[email protected]>
Subject: Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]Quoted 69 lines · Zhao Liu
From: Zhao Liu <[email protected]>
Linux kernel (from v6.4, with commit
edc0a2b595765("x86/topology: Fix erroneoussmp_num_siblingson Intel Hybrid platforms") is able to handle platforms with Module level enumerated via CPUID.1F.Expose the module level in CPUID[0x1F] if the machine has more than 1 modules.
(Tested CPU topology in CPUID[0x1F] leaf with various die/cluster configurations in "-smp".)
Signed-off-by Zhao Liu <[email protected]>Tested-by Babu Moger <[email protected]>Tested-by Yongwei Ma <[email protected]>---
Changes since v3:
* New patch to expose module level in0x1F.
* Add Tested-by tag from Yongwei.--- target/i386/cpu.c | 12 +++++++++++- target/i386/cpu.h | 2 ++ target/i386/kvm/kvm.c | 2 +- 3 files changed, 14 insertions(+), 2 deletions(-)target/i386/cpu.c +10
-0@@ -277,6 +277,8 @@ static uint32_t num_cpus_by_topo_level(X86CPUTopoInfo *topo_info,return 1 ;case CPU_TOPO_LEVEL_CORE :return topo_info->threads_per_core; +case CPU_TOPO_LEVEL_MODULE : +return topo_info->threads_per_core * topo_info->cores_per_module;case CPU_TOPO_LEVEL_DIE :return topo_info->threads_per_core * topo_info->cores_per_module * topo_info->modules_per_die;@@ -297,6 +299,8 @@ static uint32_t apicid_offset_by_topo_level(X86CPUTopoInfo *topo_info,return 0 ;case CPU_TOPO_LEVEL_CORE :return apicid_core_offset (topo_info); +case CPU_TOPO_LEVEL_MODULE : +return apicid_module_offset (topo_info);case CPU_TOPO_LEVEL_DIE :return apicid_die_offset (topo_info);case CPU_TOPO_LEVEL_PACKAGE :@@ -316,6 +320,8 @@ static uint32_t cpuid1f_topo_type(enum CPUTopoLevel topo_level)return CPUID_1F_ECX_TOPO_LEVEL_SMT ;case CPU_TOPO_LEVEL_CORE :return CPUID_1F_ECX_TOPO_LEVEL_CORE ; +case CPU_TOPO_LEVEL_MODULE : +return CPUID_1F_ECX_TOPO_LEVEL_MODULE ;case CPU_TOPO_LEVEL_DIE :return CPUID_1F_ECX_TOPO_LEVEL_DIE ;default :@@ -347,6 +353,10 @@ static void encode_topo_cpuid1f(CPUX86State *env, uint32_t count,if (env->nr_dies >1 ) {set_bit (CPU_TOPO_LEVEL_DIE , topo_bitmap); } + +if (env->nr_modules >1 ) { +set_bit (CPU_TOPO_LEVEL_MODULE , topo_bitmap); + } } *ecx = count &0xff ;target/i386/cpu.c void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count, · line 6407
Quoted 5 lines
target/i386/cpu.c +1
-1@@ -6394,7 +6404,7 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,break ;case 0x1F :/* V2 Extended Topology Enumeration Leaf */ -if (topo_info.dies_per_pkg <2 ) { +if (topo_info.modules_per_die <2 && topo_info.dies_per_pkg <2 ) {A question:
Is the original checking necessary ?
The0x1fexists even on cpu w/o modules/dies topology on bare metal, I tried on EMR:// leaf 0
0x000000000x00: eax=0x00000020 ebx=0x756e6547 ecx=0x6c65746e edx=0x49656e69// leaf
0x1f0x0000001f0x00: eax=0x00000001 ebx=0x00000002 ecx=0x00000100 edx=0x000000040x0000001f0x01: eax=0x00000007 ebx=0x00000080 ecx=0x00000201 edx=0x000000040x0000001f0x02: eax=0x00000000 ebx=0x00000000 ecx=0x00000002 edx=0x00000004
Quoted 4 lines
// leaf
0xb0x0000000b0x00: eax=0x00000001 ebx=0x00000002 ecx=0x00000100 edx=0x000000040x0000000b0x01: eax=0x00000007 ebx=0x00000080 ecx=0x00000201 edx=0x000000040x0000000b0x02: eax=0x00000000 ebx=0x00000000 ecx=0x00000002 edx=0x00000004
The 0x1f is introduced for CascadeLake-AP with die level. And yes the newer mahcines all have this leaf.
Quoted 6 lines
So here leads to different cpu behavior from bare metal, even in case of "-cpu host".
In SDM Vol2, cpudid instruction section:
Quoted 7 lines
" CPUID leaf 1FH is a preferred superset to leaf 0BH. Intel recommends using leaf 1FH when available rather than leaf 0BH and ensuring that any leaf 0BH algorithms are updated to support leaf 1FH. "
My understanding: if
0x1fis existed (leaf 0.eax >=0x1f) then it should have same values in lp/core level as0xb.
Yes, I think it's time to move to default 0x1f.
The compatibility issue can be solved by a cpuid-0x1f option similar to cpuid-0xb. I'll cook a patch after this patch series.
Thanks,
Zhao
Quoted 41 lines
Quoted 41 lines
target/i386/cpu.c +0
-0*eax = *ebx = *ecx = *edx =0 ;break ; }target/i386/cpu.h +2
-0@@ -1018,6 +1018,7 @@ enum CPUTopoLevel {CPU_TOPO_LEVEL_INVALID ,CPU_TOPO_LEVEL_SMT ,CPU_TOPO_LEVEL_CORE , +CPU_TOPO_LEVEL_MODULE ,CPU_TOPO_LEVEL_DIE ,CPU_TOPO_LEVEL_PACKAGE , CPU_TOPO_LEVEL_MAX,@@ -1032,6 +1033,7 @@ enum CPUTopoLevel {#define CPUID_1F_ECX_TOPO_LEVEL_INVALID CPUID_B_ECX_TOPO_LEVEL_INVALID#define CPUID_1F_ECX_TOPO_LEVEL_SMT CPUID_B_ECX_TOPO_LEVEL_SMT#define CPUID_1F_ECX_TOPO_LEVEL_CORE CPUID_B_ECX_TOPO_LEVEL_CORE +#define CPUID_1F_ECX_TOPO_LEVEL_MODULE 3#define CPUID_1F_ECX_TOPO_LEVEL_DIE 5/* MSR Feature Bits */ target/i386/kvm/kvm.c +1
-2@@ -1913,7 +1913,7 @@ int kvm_arch_init_vcpu(CPUState *cs)break ; }case 0x1f : -if (env->nr_dies <2 ) { +if (env->nr_modules <2 && env->nr_dies <2 ) {break ; }/* fallthrough */ --2.34.1
Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]
Quoted 96 lines · Zhao Liu
Hi Yuan,
Quoted 93 lines · Yuan Yao
Date: Mon, 15 Jan 2024 11:25:24 +0800
From: Yuan Yao <[email protected]>
Subject: Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]Quoted 69 lines · Zhao Liu
From: Zhao Liu <[email protected]>
Linux kernel (from v6.4, with commit
edc0a2b595765("x86/topology: Fix erroneoussmp_num_siblingson Intel Hybrid platforms") is able to handle platforms with Module level enumerated via CPUID.1F.Expose the module level in CPUID[0x1F] if the machine has more than 1 modules.
(Tested CPU topology in CPUID[0x1F] leaf with various die/cluster configurations in "-smp".)
Signed-off-by Zhao Liu <[email protected]>Tested-by Babu Moger <[email protected]>Tested-by Yongwei Ma <[email protected]>---
Changes since v3:
* New patch to expose module level in0x1F.
* Add Tested-by tag from Yongwei.--- target/i386/cpu.c | 12 +++++++++++- target/i386/cpu.h | 2 ++ target/i386/kvm/kvm.c | 2 +- 3 files changed, 14 insertions(+), 2 deletions(-)target/i386/cpu.c +10
-0@@ -277,6 +277,8 @@ static uint32_t num_cpus_by_topo_level(X86CPUTopoInfo *topo_info,return 1 ;case CPU_TOPO_LEVEL_CORE :return topo_info->threads_per_core; +case CPU_TOPO_LEVEL_MODULE : +return topo_info->threads_per_core * topo_info->cores_per_module;case CPU_TOPO_LEVEL_DIE :return topo_info->threads_per_core * topo_info->cores_per_module * topo_info->modules_per_die;@@ -297,6 +299,8 @@ static uint32_t apicid_offset_by_topo_level(X86CPUTopoInfo *topo_info,return 0 ;case CPU_TOPO_LEVEL_CORE :return apicid_core_offset (topo_info); +case CPU_TOPO_LEVEL_MODULE : +return apicid_module_offset (topo_info);case CPU_TOPO_LEVEL_DIE :return apicid_die_offset (topo_info);case CPU_TOPO_LEVEL_PACKAGE :@@ -316,6 +320,8 @@ static uint32_t cpuid1f_topo_type(enum CPUTopoLevel topo_level)return CPUID_1F_ECX_TOPO_LEVEL_SMT ;case CPU_TOPO_LEVEL_CORE :return CPUID_1F_ECX_TOPO_LEVEL_CORE ; +case CPU_TOPO_LEVEL_MODULE : +return CPUID_1F_ECX_TOPO_LEVEL_MODULE ;case CPU_TOPO_LEVEL_DIE :return CPUID_1F_ECX_TOPO_LEVEL_DIE ;default :@@ -347,6 +353,10 @@ static void encode_topo_cpuid1f(CPUX86State *env, uint32_t count,if (env->nr_dies >1 ) {set_bit (CPU_TOPO_LEVEL_DIE , topo_bitmap); } + +if (env->nr_modules >1 ) { +set_bit (CPU_TOPO_LEVEL_MODULE , topo_bitmap); + } } *ecx = count &0xff ;target/i386/cpu.c void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count, · line 6407
Quoted 5 lines
target/i386/cpu.c +1
-1@@ -6394,7 +6404,7 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,break ;case 0x1F :/* V2 Extended Topology Enumeration Leaf */ -if (topo_info.dies_per_pkg <2 ) { +if (topo_info.modules_per_die <2 && topo_info.dies_per_pkg <2 ) {A question:
Is the original checking necessary ?
The0x1fexists even on cpu w/o modules/dies topology on bare metal, I tried on EMR:// leaf 0
0x000000000x00: eax=0x00000020 ebx=0x756e6547 ecx=0x6c65746e edx=0x49656e69// leaf
0x1f0x0000001f0x00: eax=0x00000001 ebx=0x00000002 ecx=0x00000100 edx=0x000000040x0000001f0x01: eax=0x00000007 ebx=0x00000080 ecx=0x00000201 edx=0x000000040x0000001f0x02: eax=0x00000000 ebx=0x00000000 ecx=0x00000002 edx=0x00000004
Quoted 21 lines
Quoted 4 lines
// leaf
0xb0x0000000b0x00: eax=0x00000001 ebx=0x00000002 ecx=0x00000100 edx=0x000000040x0000000b0x01: eax=0x00000007 ebx=0x00000080 ecx=0x00000201 edx=0x000000040x0000000b0x02: eax=0x00000000 ebx=0x00000000 ecx=0x00000002 edx=0x00000004The
0x1fis introduced for CascadeLake-AP with die level. And yes the newer mahcines all have this leaf.Quoted 13 lines
So here leads to different cpu behavior from bare metal, even in case of "-cpu host".
In SDM Vol2, cpudid instruction section:
" CPUID leaf 1FH is a preferred superset to leaf 0BH. Intel recommends using leaf 1FH when available rather than leaf 0BH and ensuring that any leaf 0BH algorithms are updated to support leaf 1FH. "
My understanding: if
0x1fis existed (leaf 0.eax >=0x1f) then it should have same values in lp/core level as0xb.
No. leaf 0x1f reports the same values in lp/core leve as leaf 0xb only when the machine supports these two levels. If the machine supports more levels, they will be different.
e.g., the data on one Alder lake:
0x0000000b 0x00: eax=0x00000001 ebx=0x00000001 ecx=0x00000100 edx=0x00000006
0x0000000b 0x01: eax=0x00000007 ebx=0x00000004 ecx=0x00000201
edx=0x00000006
0x0000000b 0x02: eax=0x00000000 ebx=0x00000000 ecx=0x00000002
edx=0x00000006
0x0000001f 0x00: eax=0x00000001 ebx=0x00000001 ecx=0x00000100 edx=0x00000006
0x0000001f 0x01: eax=0x00000003 ebx=0x00000004 ecx=0x00000201
edx=0x00000006
0x0000001f 0x02: eax=0x00000007 ebx=0x00000004 ecx=0x00000302
edx=0x00000006
0x0000001f 0x03: eax=0x00000000 ebx=0x00000000 ecx=0x00000003
edx=0x00000006
Quoted 1 line
Yes, I think it's time to move to default
0x1f.
we don't need to do so until it's necessary.
Quoted 48 lines
The compatibility issue can be solved by a cpuid-0x1f option similar to cpuid-0xb. I'll cook a patch after this patch series.
Thanks,
ZhaoQuoted 41 lines
Quoted 41 lines
target/i386/cpu.c +0
-0*eax = *ebx = *ecx = *edx =0 ;break ; }target/i386/cpu.h +2
-0@@ -1018,6 +1018,7 @@ enum CPUTopoLevel {CPU_TOPO_LEVEL_INVALID ,CPU_TOPO_LEVEL_SMT ,CPU_TOPO_LEVEL_CORE , +CPU_TOPO_LEVEL_MODULE ,CPU_TOPO_LEVEL_DIE ,CPU_TOPO_LEVEL_PACKAGE , CPU_TOPO_LEVEL_MAX,@@ -1032,6 +1033,7 @@ enum CPUTopoLevel {#define CPUID_1F_ECX_TOPO_LEVEL_INVALID CPUID_B_ECX_TOPO_LEVEL_INVALID#define CPUID_1F_ECX_TOPO_LEVEL_SMT CPUID_B_ECX_TOPO_LEVEL_SMT#define CPUID_1F_ECX_TOPO_LEVEL_CORE CPUID_B_ECX_TOPO_LEVEL_CORE +#define CPUID_1F_ECX_TOPO_LEVEL_MODULE 3#define CPUID_1F_ECX_TOPO_LEVEL_DIE 5/* MSR Feature Bits */ target/i386/kvm/kvm.c +1
-2@@ -1913,7 +1913,7 @@ int kvm_arch_init_vcpu(CPUState *cs)break ; }case 0x1f : -if (env->nr_dies <2 ) { +if (env->nr_modules <2 && env->nr_dies <2 ) {break ; }/* fallthrough */ --2.34.1
Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]
Quoted 125 lines · Xiaoyao Li
Quoted 96 lines · Zhao Liu
Hi Yuan,
Quoted 93 lines · Yuan Yao
Date: Mon, 15 Jan 2024 11:25:24 +0800
From: Yuan Yao <[email protected]>
Subject: Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]Quoted 69 lines · Zhao Liu
From: Zhao Liu <[email protected]>
Linux kernel (from v6.4, with commit
edc0a2b595765("x86/topology: Fix erroneoussmp_num_siblingson Intel Hybrid platforms") is able to handle platforms with Module level enumerated via CPUID.1F.Expose the module level in CPUID[0x1F] if the machine has more than 1 modules.
(Tested CPU topology in CPUID[0x1F] leaf with various die/cluster configurations in "-smp".)
Signed-off-by Zhao Liu <[email protected]>Tested-by Babu Moger <[email protected]>Tested-by Yongwei Ma <[email protected]>---
Changes since v3:
* New patch to expose module level in0x1F.
* Add Tested-by tag from Yongwei.--- target/i386/cpu.c | 12 +++++++++++- target/i386/cpu.h | 2 ++ target/i386/kvm/kvm.c | 2 +- 3 files changed, 14 insertions(+), 2 deletions(-)target/i386/cpu.c +10
-0@@ -277,6 +277,8 @@ static uint32_t num_cpus_by_topo_level(X86CPUTopoInfo *topo_info,return 1 ;case CPU_TOPO_LEVEL_CORE :return topo_info->threads_per_core; +case CPU_TOPO_LEVEL_MODULE : +return topo_info->threads_per_core * topo_info->cores_per_module;case CPU_TOPO_LEVEL_DIE :return topo_info->threads_per_core * topo_info->cores_per_module * topo_info->modules_per_die;@@ -297,6 +299,8 @@ static uint32_t apicid_offset_by_topo_level(X86CPUTopoInfo *topo_info,return 0 ;case CPU_TOPO_LEVEL_CORE :return apicid_core_offset (topo_info); +case CPU_TOPO_LEVEL_MODULE : +return apicid_module_offset (topo_info);case CPU_TOPO_LEVEL_DIE :return apicid_die_offset (topo_info);case CPU_TOPO_LEVEL_PACKAGE :@@ -316,6 +320,8 @@ static uint32_t cpuid1f_topo_type(enum CPUTopoLevel topo_level)return CPUID_1F_ECX_TOPO_LEVEL_SMT ;case CPU_TOPO_LEVEL_CORE :return CPUID_1F_ECX_TOPO_LEVEL_CORE ; +case CPU_TOPO_LEVEL_MODULE : +return CPUID_1F_ECX_TOPO_LEVEL_MODULE ;case CPU_TOPO_LEVEL_DIE :return CPUID_1F_ECX_TOPO_LEVEL_DIE ;default :@@ -347,6 +353,10 @@ static void encode_topo_cpuid1f(CPUX86State *env, uint32_t count,if (env->nr_dies >1 ) {set_bit (CPU_TOPO_LEVEL_DIE , topo_bitmap); } + +if (env->nr_modules >1 ) { +set_bit (CPU_TOPO_LEVEL_MODULE , topo_bitmap); + } } *ecx = count &0xff ;target/i386/cpu.c void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count, · line 6407
Quoted 5 lines
target/i386/cpu.c +1
-1@@ -6394,7 +6404,7 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,break ;case 0x1F :/* V2 Extended Topology Enumeration Leaf */ -if (topo_info.dies_per_pkg <2 ) { +if (topo_info.modules_per_die <2 && topo_info.dies_per_pkg <2 ) {A question:
Is the original checking necessary ?
The0x1fexists even on cpu w/o modules/dies topology on bare metal, I tried on EMR:// leaf 0
0x000000000x00: eax=0x00000020 ebx=0x756e6547 ecx=0x6c65746e edx=0x49656e69// leaf
0x1f0x0000001f0x00: eax=0x00000001 ebx=0x00000002 ecx=0x00000100 edx=0x000000040x0000001f0x01: eax=0x00000007 ebx=0x00000080 ecx=0x00000201 edx=0x000000040x0000001f0x02: eax=0x00000000 ebx=0x00000000 ecx=0x00000002 edx=0x00000004Quoted 21 lines
Quoted 4 lines
// leaf
0xb0x0000000b0x00: eax=0x00000001 ebx=0x00000002 ecx=0x00000100 edx=0x000000040x0000000b0x01: eax=0x00000007 ebx=0x00000080 ecx=0x00000201 edx=0x000000040x0000000b0x02: eax=0x00000000 ebx=0x00000000 ecx=0x00000002 edx=0x00000004The
0x1fis introduced for CascadeLake-AP with die level. And yes the newer mahcines all have this leaf.Quoted 13 lines
So here leads to different cpu behavior from bare metal, even in case of "-cpu host".
In SDM Vol2, cpudid instruction section:
" CPUID leaf 1FH is a preferred superset to leaf 0BH. Intel recommends using leaf 1FH when available rather than leaf 0BH and ensuring that any leaf 0BH algorithms are updated to support leaf 1FH. "
My understanding: if
0x1fis existed (leaf 0.eax >=0x1f) then it should have same values in lp/core level as0xb.No. leaf
0x1freports the same values in lp/core leve as leaf0xbonly when the machine supports these two levels. If the machine supports more levels, they will be different.e.g., the data on one Alder lake:
Quoted 8 lines
0x0000000b0x00: eax=0x00000001 ebx=0x00000001 ecx=0x00000100 edx=0x000000060x0000000b0x01: eax=0x00000007 ebx=0x00000004 ecx=0x00000201 edx=0x000000060x0000000b0x02: eax=0x00000000 ebx=0x00000000 ecx=0x00000002 edx=0x00000006
0x0000001f0x00: eax=0x00000001 ebx=0x00000001 ecx=0x00000100 edx=0x000000060x0000001f0x01: eax=0x00000003 ebx=0x00000004 ecx=0x00000201 edx=0x000000060x0000001f0x02: eax=0x00000007 ebx=0x00000004 ecx=0x00000302 edx=0x000000060x0000001f0x03: eax=0x00000000 ebx=0x00000000 ecx=0x00000003 edx=0x00000006
Ah, so my understanding is incorrect on this.
I tried on one raptor lake i5-i335U, which also hybrid soc but doesn't have module level, in this case 0x1f and 0xb have same values in core/lp level.
Quoted 55 lines
Quoted 3 lines
Yes, I think it's time to move to default
0x1f.we don't need to do so until it's necessary.
Quoted 48 lines
The compatibility issue can be solved by a cpuid-0x1f option similar to cpuid-0xb. I'll cook a patch after this patch series.
Thanks,
ZhaoQuoted 41 lines
Quoted 41 lines
target/i386/cpu.c +0
-0*eax = *ebx = *ecx = *edx =0 ;break ; }target/i386/cpu.h +2
-0@@ -1018,6 +1018,7 @@ enum CPUTopoLevel {CPU_TOPO_LEVEL_INVALID ,CPU_TOPO_LEVEL_SMT ,CPU_TOPO_LEVEL_CORE , +CPU_TOPO_LEVEL_MODULE ,CPU_TOPO_LEVEL_DIE ,CPU_TOPO_LEVEL_PACKAGE , CPU_TOPO_LEVEL_MAX,@@ -1032,6 +1033,7 @@ enum CPUTopoLevel {#define CPUID_1F_ECX_TOPO_LEVEL_INVALID CPUID_B_ECX_TOPO_LEVEL_INVALID#define CPUID_1F_ECX_TOPO_LEVEL_SMT CPUID_B_ECX_TOPO_LEVEL_SMT#define CPUID_1F_ECX_TOPO_LEVEL_CORE CPUID_B_ECX_TOPO_LEVEL_CORE +#define CPUID_1F_ECX_TOPO_LEVEL_MODULE 3#define CPUID_1F_ECX_TOPO_LEVEL_DIE 5/* MSR Feature Bits */ target/i386/kvm/kvm.c +1
-2@@ -1913,7 +1913,7 @@ int kvm_arch_init_vcpu(CPUState *cs)break ; }case 0x1f : -if (env->nr_dies <2 ) { +if (env->nr_modules <2 && env->nr_dies <2 ) {break ; }/* fallthrough */ --2.34.1
Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]
Quoted 8 lines · Yuan Yao
Date: Mon, 15 Jan 2024 13:20:22 +0800
From: Yuan Yao <[email protected]>
Subject: Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]Ah, so my understanding is incorrect on this.
I tried on one raptor lake i5-i335U, which also hybrid soc but doesn't have module level, in this case
0x1fand0xbhave same values in core/lp level.
Some socs have modules/dies but they don't expose them in 0x1f.
If the soc only expose thread/core levels in 0x1f, then its 0x1f is same as 0x0b. Otherwise, it will have more subleaves and different values.
Thanks,
Zhao
Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]
Quoted 11 lines · Zhao Liu
Quoted 8 lines · Yuan Yao
Date: Mon, 15 Jan 2024 13:20:22 +0800
From: Yuan Yao <[email protected]>
Subject: Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]Ah, so my understanding is incorrect on this.
I tried on one raptor lake i5-i335U, which also hybrid soc but doesn't have module level, in this case
0x1fand0xbhave same values in core/lp level.Some socs have modules/dies but they don't expose them in
0x1f.
Here they don't expose because from hardware level they can't or possible software level configuration (i.e. disable some cores in bios) ?
Quoted 8 lines
If the soc only expose thread/core levels in
0x1f, then its0x1fis same as0x0b. Otherwise, it will have more subleaves and different values.Thanks,
Zhao
Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]
Quoted 12 lines · Yuan Yao
Date: Mon, 15 Jan 2024 14:57:30 +0800
From: Yuan Yao <[email protected]>
Subject: Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]Quoted 7 lines · Zhao Liu
Quoted 6 lines · Yuan Yao
Date: Mon, 15 Jan 2024 13:20:22 +0800
From: Yuan Yao <[email protected]>
Subject: Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]Ah, so my understanding is incorrect on this.
Quoted 8 lines
Quoted 4 lines
Quoted 2 lines
I tried on one raptor lake i5-i335U, which also hybrid soc but doesn't have module level, in this case
0x1fand0xbhave same values in core/lp level.Some socs have modules/dies but they don't expose them in
0x1f.Here they don't expose because from hardware level they can't or possible software level configuration (i.e. disable some cores in bios) ?
This leaf is decided at hardware level. Whether or not which levels are exposed sometimes depends if there is the topology-related feature, but there is no clear rule (just as in the ADL family neither ADL-S/P exposes modules, while ADL-N exposes modules).
Regards,
Zhao
Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]
Quoted 18 lines · Zhao Liu
Quoted 17 lines · Yuan Yao
Date: Mon, 15 Jan 2024 14:57:30 +0800
From: Yuan Yao <[email protected]>
Subject: Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]Quoted 11 lines · Zhao Liu
Quoted 8 lines · Yuan Yao
Date: Mon, 15 Jan 2024 13:20:22 +0800
From: Yuan Yao <[email protected]>
Subject: Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]Ah, so my understanding is incorrect on this.
I tried on one raptor lake i5-i335U, which also hybrid soc but doesn't have module level, in this case
0x1fand0xbhave same values in core/lp level.Some socs have modules/dies but they don't expose them in
0x1f.
Quoted 8 lines
Quoted 3 lines
Here they don't expose because from hardware level they can't or possible software level configuration (i.e. disable some cores in bios) ?
This leaf is decided at hardware level. Whether or not which levels are exposed sometimes depends if there is the topology-related feature, but there is no clear rule (just as in the ADL family neither ADL-S/P exposes modules, while ADL-N exposes modules).
I see, thanks for your information!
Quoted 4 lines
Regards,
Zhao
Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]
In reply to Xiaoyao LiHi Xiaoyao,
Quoted 7 lines · Xiaoyao Li
Date: Mon, 15 Jan 2024 12:34:12 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]Quoted 1 line
Yes, I think it's time to move to default
0x1f.we don't need to do so until it's necessary.
Recent and future machines all support 0x1f, and at least SDM has emphasized the preferred use of 0x1f.
Thanks,
Zhao
Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]
Quoted 7 lines · Zhao Liu
Hi Xiaoyao,
Quoted 4 lines · Xiaoyao Li
Date: Mon, 15 Jan 2024 12:34:12 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]
Quoted 6 lines
Quoted 3 lines
Quoted 1 line
Yes, I think it's time to move to default
0x1f.we don't need to do so until it's necessary.
Recent and future machines all support
0x1f, and at least SDM has emphasized the preferred use of0x1f.
The preference is the guideline for software e.g., OS. QEMU doesn't need to emulate cpuid leaf 0x1f to guest if there is only smt and core level. because in this case, they are exactly the same in leaf 0xb and 0x1f. we don't need to bother advertising the duplicate data.
Quoted 3 lines
Thanks,
Zhao
Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]
Quoted 14 lines · Xiaoyao Li
Date: Mon, 15 Jan 2024 14:11:17 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]Quoted 9 lines · Zhao Liu
Hi Xiaoyao,
Quoted 6 lines · Xiaoyao Li
Date: Mon, 15 Jan 2024 12:34:12 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]Quoted 1 line
Yes, I think it's time to move to default
0x1f.
Quoted 7 lines
Quoted 4 lines
Quoted 1 line
we don't need to do so until it's necessary.
Recent and future machines all support
0x1f, and at least SDM has emphasized the preferred use of0x1f.The preference is the guideline for software e.g., OS. QEMU doesn't need to emulate cpuid leaf
0x1fto guest if there is only smt and core level.
Please, QEMU is emulating hardware not writing software. Is there any reason why we shouldn't emulate new and generic hardware behaviors and stick with the old ones?
Quoted 2 lines
because in this case, they are exactly the same in leaf
0xband0x1f. we don't need to bother advertising the duplicate data.
You can't "define" the same 0x0b and 0x1f as duplicates. SDM doesn't have such the definition.
Regards,
Zhao
Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]
Quoted 17 lines · Zhao Liu
Quoted 16 lines · Xiaoyao Li
Date: Mon, 15 Jan 2024 14:11:17 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]Quoted 11 lines · Zhao Liu
Hi Xiaoyao,
Quoted 7 lines · Xiaoyao Li
Date: Mon, 15 Jan 2024 12:34:12 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]Quoted 1 line
Yes, I think it's time to move to default
0x1f.we don't need to do so until it's necessary.
Quoted 7 lines
Quoted 5 lines
Quoted 2 lines
Recent and future machines all support
0x1f, and at least SDM has emphasized the preferred use of0x1f.The preference is the guideline for software e.g., OS. QEMU doesn't need to emulate cpuid leaf
0x1fto guest if there is only smt and core level.Please, QEMU is emulating hardware not writing software.
what I want to conveyed was that, SDM is teaching software how to probe the cpu topology, not suggesting VMM how to advertise cpu topology to guest.
Quoted 3 lines
Is there any
reason why we shouldn't emulate new and generic hardware behaviors and stick with the old ones?
I didn't say we shouldn't, but we don't need to do it if it's unnecessary.
if cpuid 0x1f is advertised to guest by default, it will also introduce the inconsistence. Old product doesn't have cpuid 0x1f, but using QEMU to emualte an old product, it has.
sure we can have code to fix it, that only expose 0x1f to new enough cpu model. But it just make thing complicated.
Quoted 5 lines
Quoted 2 lines
because in this case, they are exactly the same in leaf
0xband0x1f. we don't need to bother advertising the duplicate data.You can't "define" the same
0x0band0x1fas duplicates. SDM doesn't have such the definition.
for QEMU, they are duplicate data that need to be maintained and need to be passed to KVM by KVM_SET_CPUID. For guest, it's also unnecessary, because it doesn't provide any additional information with cpuid leaf 1f.
SDM keeps cpuid 0xb is for backwards compatibility.
Quoted 3 lines
Regards,
Zhao
Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]
Hi Xiaoyao,
Quoted 25 lines · Xiaoyao Li
Date: Mon, 15 Jan 2024 15:16:43 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]Quoted 20 lines · Zhao Liu
Quoted 19 lines · Xiaoyao Li
Date: Mon, 15 Jan 2024 14:11:17 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]Quoted 13 lines · Zhao Liu
Hi Xiaoyao,
Quoted 7 lines · Xiaoyao Li
Date: Mon, 15 Jan 2024 12:34:12 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 08/16] i386: Expose module level in CPUID[0x1F]Quoted 1 line
Yes, I think it's time to move to default
0x1f.we don't need to do so until it's necessary.
Recent and future machines all support
0x1f, and at least SDM has emphasized the preferred use of0x1f.
Quoted 8 lines
Quoted 4 lines
Quoted 2 lines
The preference is the guideline for software e.g., OS. QEMU doesn't need to emulate cpuid leaf
0x1fto guest if there is only smt and core level.Please, QEMU is emulating hardware not writing software.
what I want to conveyed was that, SDM is teaching software how to probe the cpu topology, not suggesting VMM how to advertise cpu topology to guest.
This reflects the hardware's behavioral tendency. Additionally, due to SDM related suggestion about 0x1f preference, lots of new software may rely on 0x1f, making 0x1f as the default enabling leaf helps to enhance Guest compatibility.
Quoted 6 lines
Quoted 4 lines
Is there any
reason why we shouldn't emulate new and generic hardware behaviors and stick with the old ones?I didn't say we shouldn't, but we don't need to do it if it's unnecessary.
Probably never going to deprecate 0x0b, and 0x1f is in fact replacing 0x0b, kind of like a timing issue, when should 0x1f be enabled by default?
Maybe for some new CPU models or -host, we can start making it as default. This eliminates the difference in the CPU topology enumeration interface between Host and Guest. What do you think?
Quoted 4 lines
if cpuid
0x1fis advertised to guest by default, it will also introduce the inconsistence. Old product doesn't have cpuid0x1f, but using QEMU to emualte an old product, it has.
Yes, this is the similar case as 0x0b. Old machine doens't has 0x0b. And QEMU uses cpuid-0xb option to resolve compatibility issue.
Quoted 7 lines
sure we can have code to fix it, that only expose
0x1fto new enough cpu model. But it just make thing complicated.Quoted 3 lines
Quoted 2 lines
because in this case, they are exactly the same in leaf
0xband0x1f. we don't need to bother advertising the duplicate data.
Quoted 6 lines
Quoted 2 lines
You can't "define" the same
0x0band0x1fas duplicates. SDM doesn't have such the definition.for QEMU, they are duplicate data that need to be maintained and need to be passed to KVM by
KVM_SET_CPUID. For guest, it's also unnecessary, because it doesn't provide any additional information with cpuid leaf 1f.
I understand your concerns. The benefit is to follow the behavior of the new hardware and spec recommendations, on new machines people are going to be more accustomed to using 0x1f to get topology, and VMs on new machines that don't have 0x1f will tend to get confused.
I could start by having a look at if we could synchronize Host in -host to enable 0x1f. If there isn't too much block, -host is an acceptable starting point, after all, there are no additional compatibility issues for this case. ;-)
Thanks,
Zhao
Quoted 3 lines
SDM keeps cpuid
0xbis for backwards compatibility.
PATCH v7 9/16 i386: Support module_id in X86CPUTopoIDs
In reply to Zhao LiuFrom: Zhuocheng Ding <[email protected]>
Add module_id member in X86CPUTopoIDs.
module_id can be parsed from APIC ID, so also update APIC ID parsing rule to support module level. With this support, the conversions with module level between X86CPUTopoIDs, X86CPUTopoInfo and APIC ID are completed.
module_id can be also generated from cpu topology, and before i386 supports "clusters" in smp, the default "clusters per die" is only 1, thus the module_id generated in this way is 0, so that it will not conflict with the module_id generated by APIC ID.
---
Changes since v1:
* Merge the patch "i386: Update APIC ID parsing rule to support module level" into this one. (Yanan)
* Move the apicid_module_width() and apicid_module_offset() support into the previous modules_per_die related patch. (Yanan)
--- hw/i386/x86.c | 28 +++++++++++++++++++++------- include/hw/i386/topology.h | 17 +++++++++++++---- 2 files changed, 34 insertions(+), 11 deletions(-)
hw/i386/x86.c +21 -7
@@ -315,11 +315,11 @@ void x86_cpu_pre_plug(HotplugHandler *hotplug_dev,/* * If APIC ID is not set, -* set it based on socket/die/core/thread properties. +* set it based on socket/die/cluster/core/thread properties. */ if (cpu->apic_id ==UNASSIGNED_APIC_ID ) { -int max_socket = (ms->smp.max_cpus -1 ) / - smp_threads / smp_cores / ms->smp.dies; +int max_socket = (ms->smp.max_cpus -1 ) / smp_threads / smp_cores / + ms->smp.clusters / ms->smp.dies;/* * die-id was optional in QEMU 4.0 and older, so keep it optional
@@ -366,17 +366,27 @@ void x86_cpu_pre_plug(HotplugHandler *hotplug_dev, topo_ids.die_id = cpu->die_id; topo_ids.core_id = cpu->core_id; topo_ids.smt_id = cpu->thread_id; + +/* +* TODO: This is the temporary initialization for topo_ids.module_id to +* avoid "maybe-uninitialized" compilation errors. Will remove when +* X86CPU supports cluster_id. +*/ + topo_ids.module_id =0 ; + cpu->apic_id =x86_apicid_from_topo_ids (&topo_info, &topo_ids); } cpu_slot =x86_find_cpu_slot (MACHINE (x86ms), cpu->apic_id, &idx);if (!cpu_slot) {x86_topo_ids_from_apicid (cpu->apic_id, &topo_info, &topo_ids); +error_setg (errp, -"Invalid CPU [socket: %u, die: %u, core: %u, thread: %u] with" -" APIC ID %" PRIu32", valid index range 0:%d" , - topo_ids.pkg_id, topo_ids.die_id, topo_ids.core_id, topo_ids.smt_id, - cpu->apic_id, ms->possible_cpus->len -1 ); +"Invalid CPU [socket: %u, die: %u, module: %u, core: %u, thread: %u]" +" with APIC ID %" PRIu32", valid index range 0:%d" , + topo_ids.pkg_id, topo_ids.die_id, topo_ids.module_id, + topo_ids.core_id, topo_ids.smt_id, cpu->apic_id, + ms->possible_cpus->len -1 );return ; }
@@ -502,6 +512,10 @@ const CPUArchIdList *x86_possible_cpu_arch_ids(MachineState *ms) ms->possible_cpus->cpus[i].props.has_die_id = true; ms->possible_cpus->cpus[i].props.die_id = topo_ids.die_id; } +if (ms->smp.clusters >1 ) { + ms->possible_cpus->cpus[i].props.has_cluster_id = true; + ms->possible_cpus->cpus[i].props.cluster_id = topo_ids.module_id; + } ms->possible_cpus->cpus[i].props.has_core_id = true; ms->possible_cpus->cpus[i].props.core_id = topo_ids.core_id; ms->possible_cpus->cpus[i].props.has_thread_id = true;
include/hw/i386/topology.h +13 -4
@@ -50,6 +50,7 @@ typedef uint32_t apic_id_t;typedef struct X86CPUTopoIDs {unsigned pkg_id;unsigned die_id; +unsigned module_id;unsigned core_id;unsigned smt_id; }X86CPUTopoIDs ;
@@ -127,6 +128,7 @@ static inline apic_id_t x86_apicid_from_topo_ids(X86CPUTopoInfo *topo_info, {return (topo_ids->pkg_id <<apicid_pkg_offset (topo_info)) | (topo_ids->die_id <<apicid_die_offset (topo_info)) | + (topo_ids->module_id <<apicid_module_offset (topo_info)) | (topo_ids->core_id <<apicid_core_offset (topo_info)) | topo_ids->smt_id; }
@@ -140,12 +142,16 @@ static inline void x86_topo_ids_from_idx(X86CPUTopoInfo *topo_info, X86CPUTopoIDs *topo_ids) {unsigned nr_dies = topo_info->dies_per_pkg; -unsigned nr_cores = topo_info->cores_per_module * - topo_info->modules_per_die; +unsigned nr_modules = topo_info->modules_per_die; +unsigned nr_cores = topo_info->cores_per_module;unsigned nr_threads = topo_info->threads_per_core; - topo_ids->pkg_id = cpu_index / (nr_dies * nr_cores * nr_threads); - topo_ids->die_id = cpu_index / (nr_cores * nr_threads) % nr_dies; + topo_ids->pkg_id = cpu_index / (nr_dies * nr_modules * + nr_cores * nr_threads); + topo_ids->die_id = cpu_index / (nr_modules * nr_cores * + nr_threads) % nr_dies; + topo_ids->module_id = cpu_index / (nr_cores * nr_threads) % + nr_modules; topo_ids->core_id = cpu_index / nr_threads % nr_cores; topo_ids->smt_id = cpu_index % nr_threads; }
@@ -163,6 +169,9 @@ static inline void x86_topo_ids_from_apicid(apic_id_t apicid, topo_ids->core_id = (apicid >>apicid_core_offset (topo_info)) & ~(0xFFFFFFFFUL <<apicid_core_width (topo_info)); + topo_ids->module_id = + (apicid >>apicid_module_offset (topo_info)) & + ~(0xFFFFFFFFUL <<apicid_module_width (topo_info)); topo_ids->die_id = (apicid >>apicid_die_offset (topo_info)) & ~(0xFFFFFFFFUL <<apicid_die_width (topo_info));
Signature · 1 line
2.34.1
Re: [PATCH v7 09/16] i386: Support module_id in X86CPUTopoIDs
Quoted 52 lines · Zhao Liu
From: Zhuocheng Ding <[email protected]>
Add
module_idmember in X86CPUTopoIDs.
module_idcan be parsed from APIC ID, so also update APIC ID parsing rule to support module level. With this support, the conversions with module level between X86CPUTopoIDs, X86CPUTopoInfo and APIC ID are completed.
module_idcan be also generated from cpu topology, and before i386 supports "clusters" in smp, the default "clusters per die" is only 1, thus themodule_idgenerated in this way is 0, so that it will not conflict with themodule_idgenerated by APIC ID.Signed-off-by Zhuocheng Ding <[email protected]>Co-developed-by Zhao Liu <[email protected]>Signed-off-by Zhao Liu <[email protected]>Tested-by Babu Moger <[email protected]>Tested-by Yongwei Ma <[email protected]>---
Changes since v1:
* Merge the patch "i386: Update APIC ID parsing rule to support module level" into this one. (Yanan)
* Move theapicid_module_width()andapicid_module_offset()support into the previousmodules_per_dierelated patch. (Yanan)--- hw/i386/x86.c | 28 +++++++++++++++++++++------- include/hw/i386/topology.h | 17 +++++++++++++---- 2 files changed, 34 insertions(+), 11 deletions(-)hw/i386/x86.c +3
-3@@ -315,11 +315,11 @@ void x86_cpu_pre_plug(HotplugHandler *hotplug_dev,/* * If APIC ID is not set, -* set it based on socket/die/core/thread properties. +* set it based on socket/die/cluster/core/thread properties. */ if (cpu->apic_id ==UNASSIGNED_APIC_ID ) { -int max_socket = (ms->smp.max_cpus -1 ) / - smp_threads / smp_cores / ms->smp.dies; +int max_socket = (ms->smp.max_cpus -1 ) / smp_threads / smp_cores / + ms->smp.clusters / ms->smp.dies;/* * die-id was optional in QEMU 4.0 and older, so keep it optional @@ -366,17 +366,27 @@ void x86_cpu_pre_plug(HotplugHandler *hotplug_dev, topo_ids.die_id = cpu->die_id;
hw/i386/x86.c void x86_cpu_pre_plug(HotplugHandler *hotplug_dev, · line 376
Quoted 10 lines
hw/i386/x86.c +7
-0topo_ids.core_id = cpu->core_id; topo_ids.smt_id = cpu->thread_id; + +/* +* TODO: This is the temporary initialization for topo_ids.module_id to +* avoid "maybe-uninitialized" compilation errors. Will remove when +* X86CPU supports cluster_id. +*/ + topo_ids.module_id =0 ;
if you put patch 10 before this patch, then we don't need this trick.
Quoted 82 lines
hw/i386/x86.c +10
-4cpu->apic_id =x86_apicid_from_topo_ids (&topo_info, &topo_ids); } cpu_slot =x86_find_cpu_slot (MACHINE (x86ms), cpu->apic_id, &idx);if (!cpu_slot) {x86_topo_ids_from_apicid (cpu->apic_id, &topo_info, &topo_ids); +error_setg (errp, -"Invalid CPU [socket: %u, die: %u, core: %u, thread: %u] with" -" APIC ID %" PRIu32", valid index range 0:%d" , - topo_ids.pkg_id, topo_ids.die_id, topo_ids.core_id, topo_ids.smt_id, - cpu->apic_id, ms->possible_cpus->len -1 ); +"Invalid CPU [socket: %u, die: %u, module: %u, core: %u, thread: %u]" +" with APIC ID %" PRIu32", valid index range 0:%d" , + topo_ids.pkg_id, topo_ids.die_id, topo_ids.module_id, + topo_ids.core_id, topo_ids.smt_id, cpu->apic_id, + ms->possible_cpus->len -1 );return ; }@@ -502,6 +512,10 @@ const CPUArchIdList *x86_possible_cpu_arch_ids(MachineState *ms) ms->possible_cpus->cpus[i].props.has_die_id = true; ms->possible_cpus->cpus[i].props.die_id = topo_ids.die_id; } +if (ms->smp.clusters >1 ) { + ms->possible_cpus->cpus[i].props.has_cluster_id = true; + ms->possible_cpus->cpus[i].props.cluster_id = topo_ids.module_id; + } ms->possible_cpus->cpus[i].props.has_core_id = true; ms->possible_cpus->cpus[i].props.core_id = topo_ids.core_id; ms->possible_cpus->cpus[i].props.has_thread_id = true;include/hw/i386/topology.h +13
-4@@ -50,6 +50,7 @@ typedef uint32_t apic_id_t;typedef struct X86CPUTopoIDs {unsigned pkg_id;unsigned die_id; +unsigned module_id;unsigned core_id;unsigned smt_id; }X86CPUTopoIDs ;@@ -127,6 +128,7 @@ static inline apic_id_t x86_apicid_from_topo_ids(X86CPUTopoInfo *topo_info, {return (topo_ids->pkg_id <<apicid_pkg_offset (topo_info)) | (topo_ids->die_id <<apicid_die_offset (topo_info)) | + (topo_ids->module_id <<apicid_module_offset (topo_info)) | (topo_ids->core_id <<apicid_core_offset (topo_info)) | topo_ids->smt_id; }@@ -140,12 +142,16 @@ static inline void x86_topo_ids_from_idx(X86CPUTopoInfo *topo_info, X86CPUTopoIDs *topo_ids) {unsigned nr_dies = topo_info->dies_per_pkg; -unsigned nr_cores = topo_info->cores_per_module * - topo_info->modules_per_die; +unsigned nr_modules = topo_info->modules_per_die; +unsigned nr_cores = topo_info->cores_per_module;unsigned nr_threads = topo_info->threads_per_core; - topo_ids->pkg_id = cpu_index / (nr_dies * nr_cores * nr_threads); - topo_ids->die_id = cpu_index / (nr_cores * nr_threads) % nr_dies; + topo_ids->pkg_id = cpu_index / (nr_dies * nr_modules * + nr_cores * nr_threads); + topo_ids->die_id = cpu_index / (nr_modules * nr_cores * + nr_threads) % nr_dies; + topo_ids->module_id = cpu_index / (nr_cores * nr_threads) % + nr_modules; topo_ids->core_id = cpu_index / nr_threads % nr_cores; topo_ids->smt_id = cpu_index % nr_threads; }@@ -163,6 +169,9 @@ static inline void x86_topo_ids_from_apicid(apic_id_t apicid, topo_ids->core_id = (apicid >>apicid_core_offset (topo_info)) & ~(0xFFFFFFFFUL <<apicid_core_width (topo_info)); + topo_ids->module_id = + (apicid >>apicid_module_offset (topo_info)) & + ~(0xFFFFFFFFUL <<apicid_module_width (topo_info)); topo_ids->die_id = (apicid >>apicid_die_offset (topo_info)) & ~(0xFFFFFFFFUL <<apicid_die_width (topo_info));
Re: [PATCH v7 09/16] i386: Support module_id in X86CPUTopoIDs
Hi Xiaoyao,
Quoted 59 lines · Xiaoyao Li
Date: Sun, 14 Jan 2024 20:42:00 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 09/16] i386: Supportmodule_idin X86CPUTopoIDsQuoted 54 lines · Zhao Liu
From: Zhuocheng Ding <[email protected]>
Add
module_idmember in X86CPUTopoIDs.
module_idcan be parsed from APIC ID, so also update APIC ID parsing rule to support module level. With this support, the conversions with module level between X86CPUTopoIDs, X86CPUTopoInfo and APIC ID are completed.
module_idcan be also generated from cpu topology, and before i386 supports "clusters" in smp, the default "clusters per die" is only 1, thus themodule_idgenerated in this way is 0, so that it will not conflict with themodule_idgenerated by APIC ID.Signed-off-by Zhuocheng Ding <[email protected]>Co-developed-by Zhao Liu <[email protected]>Signed-off-by Zhao Liu <[email protected]>Tested-by Babu Moger <[email protected]>Tested-by Yongwei Ma <[email protected]>---
Changes since v1:
* Merge the patch "i386: Update APIC ID parsing rule to support module level" into this one. (Yanan)
* Move theapicid_module_width()andapicid_module_offset()support into the previousmodules_per_dierelated patch. (Yanan)--- hw/i386/x86.c | 28 +++++++++++++++++++++------- include/hw/i386/topology.h | 17 +++++++++++++---- 2 files changed, 34 insertions(+), 11 deletions(-)hw/i386/x86.c +5
-3@@ -315,11 +315,11 @@ void x86_cpu_pre_plug(HotplugHandler *hotplug_dev,/* * If APIC ID is not set, -* set it based on socket/die/core/thread properties. +* set it based on socket/die/cluster/core/thread properties. */ if (cpu->apic_id ==UNASSIGNED_APIC_ID ) { -int max_socket = (ms->smp.max_cpus -1 ) / - smp_threads / smp_cores / ms->smp.dies; +int max_socket = (ms->smp.max_cpus -1 ) / smp_threads / smp_cores / + ms->smp.clusters / ms->smp.dies;/* * die-id was optional in QEMU 4.0 and older, so keep it optional @@ -366,17 +366,27 @@ void x86_cpu_pre_plug(HotplugHandler *hotplug_dev, topo_ids.die_id = cpu->die_id; topo_ids.core_id = cpu->core_id; topo_ids.smt_id = cpu->thread_id; + +/*
Quoted 8 lines
hw/i386/x86.c void x86_cpu_pre_plug(HotplugHandler *hotplug_dev, · line 379
Quoted 6 lines
hw/i386/x86.c +5
-0+* TODO: This is the temporary initialization for topo_ids.module_id to +* avoid "maybe-uninitialized" compilation errors. Will remove when +* X86CPU supports cluster_id. +*/ + topo_ids.module_id =0 ;if you put patch 10 before this patch, then we don't need this trick.
Then, we need another trick to resolve "cpu->cluster_id = topo_ids.module_id;" in patch 10. ;-)
Thanks,
Zhao
Quoted 81 lines
Quoted 80 lines
hw/i386/x86.c +10
-4cpu->apic_id =x86_apicid_from_topo_ids (&topo_info, &topo_ids); } cpu_slot =x86_find_cpu_slot (MACHINE (x86ms), cpu->apic_id, &idx);if (!cpu_slot) {x86_topo_ids_from_apicid (cpu->apic_id, &topo_info, &topo_ids); +error_setg (errp, -"Invalid CPU [socket: %u, die: %u, core: %u, thread: %u] with" -" APIC ID %" PRIu32", valid index range 0:%d" , - topo_ids.pkg_id, topo_ids.die_id, topo_ids.core_id, topo_ids.smt_id, - cpu->apic_id, ms->possible_cpus->len -1 ); +"Invalid CPU [socket: %u, die: %u, module: %u, core: %u, thread: %u]" +" with APIC ID %" PRIu32", valid index range 0:%d" , + topo_ids.pkg_id, topo_ids.die_id, topo_ids.module_id, + topo_ids.core_id, topo_ids.smt_id, cpu->apic_id, + ms->possible_cpus->len -1 );return ; }@@ -502,6 +512,10 @@ const CPUArchIdList *x86_possible_cpu_arch_ids(MachineState *ms) ms->possible_cpus->cpus[i].props.has_die_id = true; ms->possible_cpus->cpus[i].props.die_id = topo_ids.die_id; } +if (ms->smp.clusters >1 ) { + ms->possible_cpus->cpus[i].props.has_cluster_id = true; + ms->possible_cpus->cpus[i].props.cluster_id = topo_ids.module_id; + } ms->possible_cpus->cpus[i].props.has_core_id = true; ms->possible_cpus->cpus[i].props.core_id = topo_ids.core_id; ms->possible_cpus->cpus[i].props.has_thread_id = true;include/hw/i386/topology.h +13
-4@@ -50,6 +50,7 @@ typedef uint32_t apic_id_t;typedef struct X86CPUTopoIDs {unsigned pkg_id;unsigned die_id; +unsigned module_id;unsigned core_id;unsigned smt_id; }X86CPUTopoIDs ;@@ -127,6 +128,7 @@ static inline apic_id_t x86_apicid_from_topo_ids(X86CPUTopoInfo *topo_info, {return (topo_ids->pkg_id <<apicid_pkg_offset (topo_info)) | (topo_ids->die_id <<apicid_die_offset (topo_info)) | + (topo_ids->module_id <<apicid_module_offset (topo_info)) | (topo_ids->core_id <<apicid_core_offset (topo_info)) | topo_ids->smt_id; }@@ -140,12 +142,16 @@ static inline void x86_topo_ids_from_idx(X86CPUTopoInfo *topo_info, X86CPUTopoIDs *topo_ids) {unsigned nr_dies = topo_info->dies_per_pkg; -unsigned nr_cores = topo_info->cores_per_module * - topo_info->modules_per_die; +unsigned nr_modules = topo_info->modules_per_die; +unsigned nr_cores = topo_info->cores_per_module;unsigned nr_threads = topo_info->threads_per_core; - topo_ids->pkg_id = cpu_index / (nr_dies * nr_cores * nr_threads); - topo_ids->die_id = cpu_index / (nr_cores * nr_threads) % nr_dies; + topo_ids->pkg_id = cpu_index / (nr_dies * nr_modules * + nr_cores * nr_threads); + topo_ids->die_id = cpu_index / (nr_modules * nr_cores * + nr_threads) % nr_dies; + topo_ids->module_id = cpu_index / (nr_cores * nr_threads) % + nr_modules; topo_ids->core_id = cpu_index / nr_threads % nr_cores; topo_ids->smt_id = cpu_index % nr_threads; }@@ -163,6 +169,9 @@ static inline void x86_topo_ids_from_apicid(apic_id_t apicid, topo_ids->core_id = (apicid >>apicid_core_offset (topo_info)) & ~(0xFFFFFFFFUL <<apicid_core_width (topo_info)); + topo_ids->module_id = + (apicid >>apicid_module_offset (topo_info)) & + ~(0xFFFFFFFFUL <<apicid_module_width (topo_info)); topo_ids->die_id = (apicid >>apicid_die_offset (topo_info)) & ~(0xFFFFFFFFUL <<apicid_die_width (topo_info));
PATCH v7 10/16 i386/cpu: Introduce cluster-id to X86CPU
In reply to Zhao LiuFrom: Zhuocheng Ding <[email protected]>
Introduce cluster-id other than module-id to be consistent with CpuInstanceProperties.cluster-id, and this avoids the confusion of parameter names when hotplugging.
Following the legacy smp check rules, also add the cluster_id validity into x86_cpu_pre_plug().
---
Changes since v6:
* Update the comment when check cluster-id. Since there's no v8.2, the cluster-id support should at least start from v9.0.
Changes since v5:
* Update the comment when check cluster-id. Since current QEMU is v8.2, the cluster-id support should at least start from v8.3.
Changes since v3:
* Use the imperative in the commit message. (Babu)
--- hw/i386/x86.c | 33 +++++++++++++++++++++++++-------- target/i386/cpu.c | 2 ++ target/i386/cpu.h | 1 + 3 files changed, 28 insertions(+), 8 deletions(-)
hw/i386/x86.c +25 -8
@@ -329,6 +329,14 @@ void x86_cpu_pre_plug(HotplugHandler *hotplug_dev, cpu->die_id =0 ; } +/* +* cluster-id was optional in QEMU 9.0 and older, so keep it optional +* if there's only one cluster per die. +*/ +if (cpu->cluster_id <0 && ms->smp.clusters ==1 ) { + cpu->cluster_id =0 ; + } +if (cpu->socket_id <0 ) {error_setg (errp,"CPU socket-id is not set" );return ;
@@ -345,6 +353,14 @@ void x86_cpu_pre_plug(HotplugHandler *hotplug_dev, cpu->die_id, ms->smp.dies - 1);return ; } +if (cpu->cluster_id <0 ) { +error_setg (errp,"CPU cluster-id is not set" ); +return ; + }else if (cpu->cluster_id > ms->smp.clusters -1 ) { +error_setg (errp,"Invalid CPU cluster-id: %u must be in range 0:%u" , + cpu->cluster_id, ms->smp.clusters -1 ); +return ; + }if (cpu->core_id <0 ) {error_setg (errp,"CPU core-id is not set" );return ;
@@ -364,16 +380,9 @@ void x86_cpu_pre_plug(HotplugHandler *hotplug_dev, topo_ids.pkg_id = cpu->socket_id; topo_ids.die_id = cpu->die_id; + topo_ids.module_id = cpu->cluster_id; topo_ids.core_id = cpu->core_id; topo_ids.smt_id = cpu->thread_id; - -/* -* TODO: This is the temporary initialization for topo_ids.module_id to -* avoid "maybe-uninitialized" compilation errors. Will remove when -* X86CPU supports cluster_id. -*/ - topo_ids.module_id =0 ; - cpu->apic_id =x86_apicid_from_topo_ids (&topo_info, &topo_ids); }
@@ -418,6 +427,14 @@ void x86_cpu_pre_plug(HotplugHandler *hotplug_dev, } cpu->die_id = topo_ids.die_id; +if (cpu->cluster_id !=-1 && cpu->cluster_id != topo_ids.module_id) { +error_setg (errp,"property cluster-id: %u doesn't match set apic-id:" +" 0x%x (cluster-id: %u)" , cpu->cluster_id, cpu->apic_id, + topo_ids.module_id); +return ; + } + cpu->cluster_id = topo_ids.module_id; +if (cpu->core_id !=-1 && cpu->core_id != topo_ids.core_id) { error_setg(errp, "property core-id: %u doesn't match set apic-id:" " 0x%x (core-id: %u)", cpu->core_id, cpu->apic_id,
target/i386/cpu.c +2 -0
@@ -7909,12 +7909,14 @@ static Property x86_cpu_properties[] = {DEFINE_PROP_UINT32 ("apic-id" ,X86CPU , apic_id,0 ),DEFINE_PROP_INT32 ("thread-id" ,X86CPU , thread_id,0 ),DEFINE_PROP_INT32 ("core-id" ,X86CPU , core_id,0 ), +DEFINE_PROP_INT32 ("cluster-id" ,X86CPU , cluster_id,0 ),DEFINE_PROP_INT32 ("die-id" ,X86CPU , die_id,0 ),DEFINE_PROP_INT32 ("socket-id" ,X86CPU , socket_id,0 ), #elseDEFINE_PROP_UINT32 ("apic-id" ,X86CPU , apic_id,UNASSIGNED_APIC_ID ),DEFINE_PROP_INT32 ("thread-id" ,X86CPU , thread_id,-1 ),DEFINE_PROP_INT32 ("core-id" ,X86CPU , core_id,-1 ), +DEFINE_PROP_INT32 ("cluster-id" ,X86CPU , cluster_id,-1 ),DEFINE_PROP_INT32 ("die-id" ,X86CPU , die_id,-1 ), DEFINE_PROP_INT32("socket-id", X86CPU, socket_id, -1), #endif
target/i386/cpu.h +1 -0
@@ -2057,6 +2057,7 @@ struct ArchCPU {int32_t node_id;/* NUMA node this CPU belongs to */ int32_t socket_id;int32_t die_id; +int32_t cluster_id;int32_t core_id;int32_t thread_id;
Signature · 1 line
2.34.1
Re: [PATCH v7 10/16] i386/cpu: Introduce cluster-id to X86CPU
Quoted 5 lines · Zhao Liu
From: Zhuocheng Ding <[email protected]>
Introduce cluster-id other than module-id to be consistent with CpuInstanceProperties.cluster-id, and this avoids the confusion of parameter names when hotplugging.
I don't think reusing 'cluster' from arm for x86's 'module' is a good idea. It introduces confusion around the code.
s390 just added 'drawer' and 'book' in cpu topology[1]. I think we can also add a module level for x86 instead of reusing cluster.
(This is also what I want to reply to the cover letter.)
[1] https://lore.kernel.org/qemu-devel/[email protected]/
Quoted 124 lines
Following the legacy smp check rules, also add the
cluster_idvalidity intox86_cpu_pre_plug().Signed-off-by Zhuocheng Ding <[email protected]>Co-developed-by Zhao Liu <[email protected]>Signed-off-by Zhao Liu <[email protected]>Tested-by Babu Moger <[email protected]>Tested-by Yongwei Ma <[email protected]>---
Changes since v6:
* Update the comment when check cluster-id. Since there's no v8.2, the cluster-id support should at least start from v9.0.Changes since v5:
* Update the comment when check cluster-id. Since current QEMU is v8.2, the cluster-id support should at least start from v8.3.Changes since v3:
* Use the imperative in the commit message. (Babu)--- hw/i386/x86.c | 33 +++++++++++++++++++++++++-------- target/i386/cpu.c | 2 ++ target/i386/cpu.h | 1 + 3 files changed, 28 insertions(+), 8 deletions(-)hw/i386/x86.c +25
-8@@ -329,6 +329,14 @@ void x86_cpu_pre_plug(HotplugHandler *hotplug_dev, cpu->die_id =0 ; } +/* +* cluster-id was optional in QEMU 9.0 and older, so keep it optional +* if there's only one cluster per die. +*/ +if (cpu->cluster_id <0 && ms->smp.clusters ==1 ) { + cpu->cluster_id =0 ; + } +if (cpu->socket_id <0 ) {error_setg (errp,"CPU socket-id is not set" );return ;@@ -345,6 +353,14 @@ void x86_cpu_pre_plug(HotplugHandler *hotplug_dev, cpu->die_id, ms->smp.dies - 1);return ; } +if (cpu->cluster_id <0 ) { +error_setg (errp,"CPU cluster-id is not set" ); +return ; + }else if (cpu->cluster_id > ms->smp.clusters -1 ) { +error_setg (errp,"Invalid CPU cluster-id: %u must be in range 0:%u" , + cpu->cluster_id, ms->smp.clusters -1 ); +return ; + }if (cpu->core_id <0 ) {error_setg (errp,"CPU core-id is not set" );return ;@@ -364,16 +380,9 @@ void x86_cpu_pre_plug(HotplugHandler *hotplug_dev, topo_ids.pkg_id = cpu->socket_id; topo_ids.die_id = cpu->die_id; + topo_ids.module_id = cpu->cluster_id; topo_ids.core_id = cpu->core_id; topo_ids.smt_id = cpu->thread_id; - -/* -* TODO: This is the temporary initialization for topo_ids.module_id to -* avoid "maybe-uninitialized" compilation errors. Will remove when -* X86CPU supports cluster_id. -*/ - topo_ids.module_id =0 ; - cpu->apic_id =x86_apicid_from_topo_ids (&topo_info, &topo_ids); }@@ -418,6 +427,14 @@ void x86_cpu_pre_plug(HotplugHandler *hotplug_dev, } cpu->die_id = topo_ids.die_id; +if (cpu->cluster_id !=-1 && cpu->cluster_id != topo_ids.module_id) { +error_setg (errp,"property cluster-id: %u doesn't match set apic-id:" +" 0x%x (cluster-id: %u)" , cpu->cluster_id, cpu->apic_id, + topo_ids.module_id); +return ; + } + cpu->cluster_id = topo_ids.module_id; +if (cpu->core_id !=-1 && cpu->core_id != topo_ids.core_id) { error_setg(errp, "property core-id: %u doesn't match set apic-id:" " 0x%x (core-id: %u)", cpu->core_id, cpu->apic_id,target/i386/cpu.c +2
-0@@ -7909,12 +7909,14 @@ static Property x86_cpu_properties[] = {DEFINE_PROP_UINT32 ("apic-id" ,X86CPU , apic_id,0 ),DEFINE_PROP_INT32 ("thread-id" ,X86CPU , thread_id,0 ),DEFINE_PROP_INT32 ("core-id" ,X86CPU , core_id,0 ), +DEFINE_PROP_INT32 ("cluster-id" ,X86CPU , cluster_id,0 ),DEFINE_PROP_INT32 ("die-id" ,X86CPU , die_id,0 ),DEFINE_PROP_INT32 ("socket-id" ,X86CPU , socket_id,0 ), #elseDEFINE_PROP_UINT32 ("apic-id" ,X86CPU , apic_id,UNASSIGNED_APIC_ID ),DEFINE_PROP_INT32 ("thread-id" ,X86CPU , thread_id,-1 ),DEFINE_PROP_INT32 ("core-id" ,X86CPU , core_id,-1 ), +DEFINE_PROP_INT32 ("cluster-id" ,X86CPU , cluster_id,-1 ),DEFINE_PROP_INT32 ("die-id" ,X86CPU , die_id,-1 ), DEFINE_PROP_INT32("socket-id", X86CPU, socket_id, -1), #endiftarget/i386/cpu.h +1
-0@@ -2057,6 +2057,7 @@ struct ArchCPU {int32_t node_id;/* NUMA node this CPU belongs to */ int32_t socket_id;int32_t die_id; +int32_t cluster_id;int32_t core_id;int32_t thread_id;
Re: [PATCH v7 10/16] i386/cpu: Introduce cluster-id to X86CPU
Quoted 5 lines · Xiaoyao Li
Date: Sun, 14 Jan 2024 21:49:18 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 10/16] i386/cpu: Introduce cluster-id to X86CPUOn 1/8/2024 4:27 PM, Zhao Liu wrote:
Quoted 8 lines
Quoted 5 lines
From: Zhuocheng Ding <[email protected]>
Introduce cluster-id other than module-id to be consistent with CpuInstanceProperties.cluster-id, and this avoids the confusion of parameter names when hotplugging.
I don't think reusing 'cluster' from arm for x86's 'module' is a good idea. It introduces confusion around the code.
There is a precedent: generic "socket" v.s. i386 "package".
The direct definition of cluster is the level that is above the "core" and shares the hardware resources including L2. In this sense, arm's cluster is the same as x86's module.
Though different arches have different naming styles, but QEMU's generic code still need the uniform topology hierarchy.
Quoted 7 lines
s390 just added 'drawer' and 'book' in cpu topology[1]. I think we can also add a module level for x86 instead of reusing cluster.
(This is also what I want to reply to the cover letter.)
These two new levels have the clear topological hierarchy relationship and don't duplicate existing ones.
"book" or "drawer" may correspond to intel's "cluster".
Maybe, in the future, we could support for arch-specific alias topologies in -smp.
Thanks,
Zhao
Quoted 121 lines
Quoted 120 lines
Following the legacy smp check rules, also add the
cluster_idvalidity intox86_cpu_pre_plug().Signed-off-by Zhuocheng Ding <[email protected]>Co-developed-by Zhao Liu <[email protected]>Signed-off-by Zhao Liu <[email protected]>Tested-by Babu Moger <[email protected]>Tested-by Yongwei Ma <[email protected]>---
Changes since v6:
* Update the comment when check cluster-id. Since there's no v8.2, the cluster-id support should at least start from v9.0.Changes since v5:
* Update the comment when check cluster-id. Since current QEMU is v8.2, the cluster-id support should at least start from v8.3.Changes since v3:
* Use the imperative in the commit message. (Babu)--- hw/i386/x86.c | 33 +++++++++++++++++++++++++-------- target/i386/cpu.c | 2 ++ target/i386/cpu.h | 1 + 3 files changed, 28 insertions(+), 8 deletions(-)hw/i386/x86.c +25
-8@@ -329,6 +329,14 @@ void x86_cpu_pre_plug(HotplugHandler *hotplug_dev, cpu->die_id =0 ; } +/* +* cluster-id was optional in QEMU 9.0 and older, so keep it optional +* if there's only one cluster per die. +*/ +if (cpu->cluster_id <0 && ms->smp.clusters ==1 ) { + cpu->cluster_id =0 ; + } +if (cpu->socket_id <0 ) {error_setg (errp,"CPU socket-id is not set" );return ;@@ -345,6 +353,14 @@ void x86_cpu_pre_plug(HotplugHandler *hotplug_dev, cpu->die_id, ms->smp.dies - 1);return ; } +if (cpu->cluster_id <0 ) { +error_setg (errp,"CPU cluster-id is not set" ); +return ; + }else if (cpu->cluster_id > ms->smp.clusters -1 ) { +error_setg (errp,"Invalid CPU cluster-id: %u must be in range 0:%u" , + cpu->cluster_id, ms->smp.clusters -1 ); +return ; + }if (cpu->core_id <0 ) {error_setg (errp,"CPU core-id is not set" );return ;@@ -364,16 +380,9 @@ void x86_cpu_pre_plug(HotplugHandler *hotplug_dev, topo_ids.pkg_id = cpu->socket_id; topo_ids.die_id = cpu->die_id; + topo_ids.module_id = cpu->cluster_id; topo_ids.core_id = cpu->core_id; topo_ids.smt_id = cpu->thread_id; - -/* -* TODO: This is the temporary initialization for topo_ids.module_id to -* avoid "maybe-uninitialized" compilation errors. Will remove when -* X86CPU supports cluster_id. -*/ - topo_ids.module_id =0 ; - cpu->apic_id =x86_apicid_from_topo_ids (&topo_info, &topo_ids); }@@ -418,6 +427,14 @@ void x86_cpu_pre_plug(HotplugHandler *hotplug_dev, } cpu->die_id = topo_ids.die_id; +if (cpu->cluster_id !=-1 && cpu->cluster_id != topo_ids.module_id) { +error_setg (errp,"property cluster-id: %u doesn't match set apic-id:" +" 0x%x (cluster-id: %u)" , cpu->cluster_id, cpu->apic_id, + topo_ids.module_id); +return ; + } + cpu->cluster_id = topo_ids.module_id; +if (cpu->core_id !=-1 && cpu->core_id != topo_ids.core_id) { error_setg(errp, "property core-id: %u doesn't match set apic-id:" " 0x%x (core-id: %u)", cpu->core_id, cpu->apic_id,target/i386/cpu.c +2
-0@@ -7909,12 +7909,14 @@ static Property x86_cpu_properties[] = {DEFINE_PROP_UINT32 ("apic-id" ,X86CPU , apic_id,0 ),DEFINE_PROP_INT32 ("thread-id" ,X86CPU , thread_id,0 ),DEFINE_PROP_INT32 ("core-id" ,X86CPU , core_id,0 ), +DEFINE_PROP_INT32 ("cluster-id" ,X86CPU , cluster_id,0 ),DEFINE_PROP_INT32 ("die-id" ,X86CPU , die_id,0 ),DEFINE_PROP_INT32 ("socket-id" ,X86CPU , socket_id,0 ), #elseDEFINE_PROP_UINT32 ("apic-id" ,X86CPU , apic_id,UNASSIGNED_APIC_ID ),DEFINE_PROP_INT32 ("thread-id" ,X86CPU , thread_id,-1 ),DEFINE_PROP_INT32 ("core-id" ,X86CPU , core_id,-1 ), +DEFINE_PROP_INT32 ("cluster-id" ,X86CPU , cluster_id,-1 ),DEFINE_PROP_INT32 ("die-id" ,X86CPU , die_id,-1 ), DEFINE_PROP_INT32("socket-id", X86CPU, socket_id, -1), #endiftarget/i386/cpu.h +1
-0@@ -2057,6 +2057,7 @@ struct ArchCPU {int32_t node_id;/* NUMA node this CPU belongs to */ int32_t socket_id;int32_t die_id; +int32_t cluster_id;int32_t core_id;int32_t thread_id;
Re: [PATCH v7 10/16] i386/cpu: Introduce cluster-id to X86CPU
Quoted 8 lines · Zhao Liu
Quoted 7 lines · Xiaoyao Li
Date: Sun, 14 Jan 2024 21:49:18 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 10/16] i386/cpu: Introduce cluster-id to X86CPUQuoted 2 lines · Zhao Liu
From: Zhuocheng Ding <[email protected]>
Quoted 8 lines
Quoted 6 lines
Quoted 3 lines
Introduce cluster-id other than module-id to be consistent with CpuInstanceProperties.cluster-id, and this avoids the confusion of parameter names when hotplugging.
I don't think reusing 'cluster' from arm for x86's 'module' is a good idea. It introduces confusion around the code.
There is a precedent: generic "socket" v.s. i386 "package".
It's not the same thing. "socket" vs "package" is just software people and hardware people chose different name. It's just different naming issue.
however, here it's reusing name issue while 'cluster' has been defined for x86. It does introduce confusion.
Quoted 3 lines
The direct definition of cluster is the level that is above the "core" and shares the hardware resources including L2. In this sense, arm's cluster is the same as x86's module.
then, what about intel implements tile level in the future? why ARM's 'cluster' is mapped to 'module', but not 'tile' ?
reusing 'cluster' for 'module' is just a bad idea.
Quoted 2 lines
Though different arches have different naming styles, but QEMU's generic code still need the uniform topology hierarchy.
generic code can provide as many topology levels as it can. each ARCH can choose to use the ones it supports.
e.g.,
in qapi/machine.json, it says,
# The ordering from highest/coarsest to lowest/finest is:
# @drawers, @books, @sockets, @dies, @clusters, @cores, @threads. #
# Different architectures support different subsets of topology
# containers. # # For example, s390x does not have clusters and dies, and the socket # is the parent container of cores.
we can update it to
# The ordering from highest/coarsest to lowest/finest is:
# @drawers, @books, @sockets, @dies, @clusters, @module, @cores, # @threads.
#
# Different architectures support different subsets of topology
# containers. # # For example, s390x does not have clusters and dies, and the socket # is the parent container of cores. # # For example, x86 does not have drawers and books, and does not support # cluster.
even if cluster of x86 is supported someday in the future, we can remove the ordering requirement from above description.
Quoted 8 lines
Quoted 7 lines
s390 just added 'drawer' and 'book' in cpu topology[1]. I think we can also add a module level for x86 instead of reusing cluster.
(This is also what I want to reply to the cover letter.)
Quoted 7 lines
These two new levels have the clear topological hierarchy relationship and don't duplicate existing ones.
"book" or "drawer" may correspond to intel's "cluster".
Maybe, in the future, we could support for arch-specific alias topologies in -smp.
I don't think we need alias, reusing 'cluster' for 'module' doesn't gain any benefit except avoid adding a new field in SMPconfiguration. All the other cluster code is ARM specific and x86 cannot share.
I don't think it's a problem to add 'module' to SMPconfiguration.
Quoted 124 lines
Thanks,
ZhaoQuoted 121 lines
Quoted 120 lines
Following the legacy smp check rules, also add the
cluster_idvalidity intox86_cpu_pre_plug().Signed-off-by Zhuocheng Ding <[email protected]>Co-developed-by Zhao Liu <[email protected]>Signed-off-by Zhao Liu <[email protected]>Tested-by Babu Moger <[email protected]>Tested-by Yongwei Ma <[email protected]>---
Changes since v6:
* Update the comment when check cluster-id. Since there's no v8.2, the cluster-id support should at least start from v9.0.Changes since v5:
* Update the comment when check cluster-id. Since current QEMU is v8.2, the cluster-id support should at least start from v8.3.Changes since v3:
* Use the imperative in the commit message. (Babu)--- hw/i386/x86.c | 33 +++++++++++++++++++++++++-------- target/i386/cpu.c | 2 ++ target/i386/cpu.h | 1 + 3 files changed, 28 insertions(+), 8 deletions(-)hw/i386/x86.c +25
-8@@ -329,6 +329,14 @@ void x86_cpu_pre_plug(HotplugHandler *hotplug_dev, cpu->die_id =0 ; } +/* +* cluster-id was optional in QEMU 9.0 and older, so keep it optional +* if there's only one cluster per die. +*/ +if (cpu->cluster_id <0 && ms->smp.clusters ==1 ) { + cpu->cluster_id =0 ; + } +if (cpu->socket_id <0 ) {error_setg (errp,"CPU socket-id is not set" );return ;@@ -345,6 +353,14 @@ void x86_cpu_pre_plug(HotplugHandler *hotplug_dev, cpu->die_id, ms->smp.dies - 1);return ; } +if (cpu->cluster_id <0 ) { +error_setg (errp,"CPU cluster-id is not set" ); +return ; + }else if (cpu->cluster_id > ms->smp.clusters -1 ) { +error_setg (errp,"Invalid CPU cluster-id: %u must be in range 0:%u" , + cpu->cluster_id, ms->smp.clusters -1 ); +return ; + }if (cpu->core_id <0 ) {error_setg (errp,"CPU core-id is not set" );return ;@@ -364,16 +380,9 @@ void x86_cpu_pre_plug(HotplugHandler *hotplug_dev, topo_ids.pkg_id = cpu->socket_id; topo_ids.die_id = cpu->die_id; + topo_ids.module_id = cpu->cluster_id; topo_ids.core_id = cpu->core_id; topo_ids.smt_id = cpu->thread_id; - -/* -* TODO: This is the temporary initialization for topo_ids.module_id to -* avoid "maybe-uninitialized" compilation errors. Will remove when -* X86CPU supports cluster_id. -*/ - topo_ids.module_id =0 ; - cpu->apic_id =x86_apicid_from_topo_ids (&topo_info, &topo_ids); }@@ -418,6 +427,14 @@ void x86_cpu_pre_plug(HotplugHandler *hotplug_dev, } cpu->die_id = topo_ids.die_id; +if (cpu->cluster_id !=-1 && cpu->cluster_id != topo_ids.module_id) { +error_setg (errp,"property cluster-id: %u doesn't match set apic-id:" +" 0x%x (cluster-id: %u)" , cpu->cluster_id, cpu->apic_id, + topo_ids.module_id); +return ; + } + cpu->cluster_id = topo_ids.module_id; +if (cpu->core_id !=-1 && cpu->core_id != topo_ids.core_id) { error_setg(errp, "property core-id: %u doesn't match set apic-id:" " 0x%x (core-id: %u)", cpu->core_id, cpu->apic_id,target/i386/cpu.c +2
-0@@ -7909,12 +7909,14 @@ static Property x86_cpu_properties[] = {DEFINE_PROP_UINT32 ("apic-id" ,X86CPU , apic_id,0 ),DEFINE_PROP_INT32 ("thread-id" ,X86CPU , thread_id,0 ),DEFINE_PROP_INT32 ("core-id" ,X86CPU , core_id,0 ), +DEFINE_PROP_INT32 ("cluster-id" ,X86CPU , cluster_id,0 ),DEFINE_PROP_INT32 ("die-id" ,X86CPU , die_id,0 ),DEFINE_PROP_INT32 ("socket-id" ,X86CPU , socket_id,0 ), #elseDEFINE_PROP_UINT32 ("apic-id" ,X86CPU , apic_id,UNASSIGNED_APIC_ID ),DEFINE_PROP_INT32 ("thread-id" ,X86CPU , thread_id,-1 ),DEFINE_PROP_INT32 ("core-id" ,X86CPU , core_id,-1 ), +DEFINE_PROP_INT32 ("cluster-id" ,X86CPU , cluster_id,-1 ),DEFINE_PROP_INT32 ("die-id" ,X86CPU , die_id,-1 ), DEFINE_PROP_INT32("socket-id", X86CPU, socket_id, -1), #endiftarget/i386/cpu.h +1
-0@@ -2057,6 +2057,7 @@ struct ArchCPU {int32_t node_id;/* NUMA node this CPU belongs to */ int32_t socket_id;int32_t die_id; +int32_t cluster_id;int32_t core_id;int32_t thread_id;
Re: [PATCH v7 10/16] i386/cpu: Introduce cluster-id to X86CPU
(Also cc "machine core" maintainers.)
Hi Xiaoyao,
Quoted 17 lines · Xiaoyao Li
Date: Mon, 15 Jan 2024 12:18:17 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 10/16] i386/cpu: Introduce cluster-id to X86CPUQuoted 12 lines · Zhao Liu
Quoted 11 lines · Xiaoyao Li
Date: Sun, 14 Jan 2024 21:49:18 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 10/16] i386/cpu: Introduce cluster-id to X86CPUQuoted 5 lines · Zhao Liu
From: Zhuocheng Ding <[email protected]>
Introduce cluster-id other than module-id to be consistent with CpuInstanceProperties.cluster-id, and this avoids the confusion of parameter names when hotplugging.
Quoted 7 lines
Quoted 4 lines
Quoted 2 lines
I don't think reusing 'cluster' from arm for x86's 'module' is a good idea. It introduces confusion around the code.
There is a precedent: generic "socket" v.s. i386 "package".
It's not the same thing. "socket" vs "package" is just software people and hardware people chose different name. It's just different naming issue.
No, it's a similar issue. Same physical device, different name only.
Furthermore, the topology was introduced for resource layout and silicon fabrication, and similar design ideas and fabrication processes are fairly consistent across common current arches. Therefore, it is possible to abstract similar topological hierarchies for different arches.
Quoted 3 lines
however, here it's reusing name issue while 'cluster' has been defined for x86. It does introduce confusion.
There's nothing fundamentally different between the x86 module and the generic cluster, is there? This is the reason that I don't agree with introducing "modules" in -smp.
Quoted 7 lines
Quoted 4 lines
The direct definition of cluster is the level that is above the "core" and shares the hardware resources including L2. In this sense, arm's cluster is the same as x86's module.
then, what about intel implements tile level in the future? why ARM's 'cluster' is mapped to 'module', but not 'tile' ?
This depends on the actual need.
Module (for x86) and cluster (in general) are similar, and tile (for x86) is used for L3 in practice, so I use module rather than tile to map generic clusters.
And, it should be noted that x86 module is mapped to the generic cluster, not to ARM's. It's just that currently only ARM is using the clusters option in -smp.
I believe QEMU provides the abstract and unified topology hierarchies in -smp, not the arch-specific hierarchies.
Quoted 9 lines
reusing 'cluster' for 'module' is just a bad idea.
Quoted 2 lines
Though different arches have different naming styles, but QEMU's generic code still need the uniform topology hierarchy.
generic code can provide as many topology levels as it can. each ARCH can choose to use the ones it supports.
Quoted 6 lines
e.g.,
in
qapi/machine.json, it says,# The ordering from highest/coarsest to lowest/finest is:
# @drawers, @books, @sockets, @dies, @clusters, @cores, @threads.
This ordering is well-defined...
Quoted 7 lines
#
# Different architectures support different subsets of topology# containers. # # For example, s390x does not have clusters and dies, and the socket # is the parent container of cores.
Quoted 5 lines
we can update it to
# The ordering from highest/coarsest to lowest/finest is:
# @drawers, @books, @sockets, @dies, @clusters, @module, @cores, # @threads.
...but here it's impossible to figure out why cluster is above module, and even I can't come up with the difference between cluster and module.
Quoted 10 lines
#
# Different architectures support different subsets of topology# containers. # # For example, s390x does not have clusters and dies, and the socket # is the parent container of cores. # # For example, x86 does not have drawers and books, and does not support # cluster.
Quoted 2 lines
even if cluster of x86 is supported someday in the future, we can remove the ordering requirement from above description.
x86's cluster is above the package.
To reserve this name for x86, we can't have the well-defined topology ordering.
But topology ordering is necessary in generic code, and many calculations depend on the topology ordering.
Quoted 12 lines
Quoted 12 lines
Quoted 8 lines
s390 just added 'drawer' and 'book' in cpu topology[1]. I think we can also add a module level for x86 instead of reusing cluster.
(This is also what I want to reply to the cover letter.)
These two new levels have the clear topological hierarchy relationship and don't duplicate existing ones.
Quoted 8 lines
Quoted 4 lines
"book" or "drawer" may correspond to intel's "cluster".
Maybe, in the future, we could support for arch-specific alias topologies in -smp.
I don't think we need alias, reusing 'cluster' for 'module' doesn't gain any benefit except avoid adding a new field in SMPconfiguration. All the other cluster code is ARM specific and x86 cannot share.
The point is that there is no difference between intel module and general cluster...Considering only the naming issue, even AMD has the "complex" to correspond to the Intel's "module".
Quoted 2 lines
I don't think it's a problem to add 'module' to SMPconfiguration.
Adding an option is simple, but however, it is not conducive to the topology maintenance of QEMU, reusing the existing generic structure should be the first consideration except when the new level is fundamentally different.
Thanks,
Zhao
Re: [PATCH v7 10/16] i386/cpu: Introduce cluster-id to X86CPU
Quoted 38 lines · Zhao Liu
(Also cc "machine core" maintainers.)
Hi Xiaoyao,
Quoted 17 lines · Xiaoyao Li
Date: Mon, 15 Jan 2024 12:18:17 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 10/16] i386/cpu: Introduce cluster-id to X86CPUQuoted 12 lines · Zhao Liu
Quoted 11 lines · Xiaoyao Li
Date: Sun, 14 Jan 2024 21:49:18 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 10/16] i386/cpu: Introduce cluster-id to X86CPUQuoted 5 lines · Zhao Liu
From: Zhuocheng Ding <[email protected]>
Introduce cluster-id other than module-id to be consistent with CpuInstanceProperties.cluster-id, and this avoids the confusion of parameter names when hotplugging.
Quoted 7 lines
Quoted 4 lines
Quoted 2 lines
I don't think reusing 'cluster' from arm for x86's 'module' is a good idea. It introduces confusion around the code.
There is a precedent: generic "socket" v.s. i386 "package".
It's not the same thing. "socket" vs "package" is just software people and hardware people chose different name. It's just different naming issue.
No, it's a similar issue. Same physical device, different name only.
Furthermore, the topology was introduced for resource layout and silicon fabrication, and similar design ideas and fabrication processes are fairly consistent across common current arches. Therefore, it is possible to abstract similar topological hierarchies for different arches.
Quoted 1 line
Quoted 6 lines
Quoted 2 lines
however, here it's reusing name issue while 'cluster' has been defined for x86. It does introduce confusion.
There's nothing fundamentally different between the x86 module and the generic cluster, is there? This is the reason that I don't agree with introducing "modules" in -smp.
generic cluster just means the cluster of processors, i.e, a group of cpus/lps. It is just a middle level between die and core.
It can be the module level in intel, or tile level. Further, if per die lp number increases in the future, there might be more middle levels in intel between die and core. Then at that time, how to decide what level should cluster be mapped to?
Quoted 68 lines
Quoted 7 lines
Quoted 4 lines
The direct definition of cluster is the level that is above the "core" and shares the hardware resources including L2. In this sense, arm's cluster is the same as x86's module.
then, what about intel implements tile level in the future? why ARM's 'cluster' is mapped to 'module', but not 'tile' ?
This depends on the actual need.
Module (for x86) and cluster (in general) are similar, and tile (for x86) is used for L3 in practice, so I use module rather than tile to map generic cluster.
And, it should be noted that x86 module is mapped to the generic cluster, not to ARM's. It's just that currently only ARM is using the clusters option in -smp.
I believe QEMU provides the abstract and unified topology hierarchies in -smp, not the arch-specific hierarchies.
Quoted 9 lines
reusing 'cluster' for 'module' is just a bad idea.
Quoted 2 lines
Though different arches have different naming styles, but QEMU's generic code still need the uniform topology hierarchy.
generic code can provide as many topology levels as it can. each ARCH can choose to use the ones it supports.
Quoted 6 lines
e.g.,
in
qapi/machine.json, it says,# The ordering from highest/coarsest to lowest/finest is:
# @drawers, @books, @sockets, @dies, @clusters, @cores, @threads.This ordering is well-defined...
Quoted 7 lines
#
# Different architectures support different subsets of topology# containers. # # For example, s390x does not have clusters and dies, and the socket # is the parent container of cores.Quoted 5 lines
we can update it to
# The ordering from highest/coarsest to lowest/finest is:
# @drawers, @books, @sockets, @dies, @clusters, @module, @cores, # @threads....but here it's impossible to figure out why cluster is above module, and even I can't come up with the difference between cluster and module.
Quoted 10 lines
#
# Different architectures support different subsets of topology# containers. # # For example, s390x does not have clusters and dies, and the socket # is the parent container of cores. # # For example, x86 does not have drawers and books, and does not support # cluster.Quoted 2 lines
even if cluster of x86 is supported someday in the future, we can remove the ordering requirement from above description.
Quoted 7 lines
x86's cluster is above the package.
To reserve this name for x86, we can't have the well-defined topology ordering.
But topology ordering is necessary in generic code, and many calculations depend on the topology ordering.
could you point me to the code?
Quoted 17 lines
Quoted 17 lines
Quoted 16 lines
Quoted 8 lines
s390 just added 'drawer' and 'book' in cpu topology[1]. I think we can also add a module level for x86 instead of reusing cluster.
(This is also what I want to reply to the cover letter.)
These two new levels have the clear topological hierarchy relationship and don't duplicate existing ones.
"book" or "drawer" may correspond to intel's "cluster".
Maybe, in the future, we could support for arch-specific alias topologies in -smp.
Quoted 7 lines
Quoted 3 lines
I don't think we need alias, reusing 'cluster' for 'module' doesn't gain any benefit except avoid adding a new field in SMPconfiguration. All the other cluster code is ARM specific and x86 cannot share.
The point is that there is no difference between intel module and general cluster...Considering only the naming issue, even AMD has the "complex" to correspond to the Intel's "module".
does complex of AMD really match with intel module? L3 cache is shared in one complex, while L2 cache is shared in one module for now.
Quoted 11 lines
Quoted 2 lines
I don't think it's a problem to add 'module' to SMPconfiguration.
Adding an option is simple, but however, it is not conducive to the topology maintenance of QEMU, reusing the existing generic structure should be the first consideration except when the new level is fundamentally different.
Thanks,
Zhao
Re: [PATCH v7 10/16] i386/cpu: Introduce cluster-id to X86CPU
Hi Xiaoyao,
Quoted 46 lines · Xiaoyao Li
Date: Mon, 15 Jan 2024 15:45:58 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 10/16] i386/cpu: Introduce cluster-id to X86CPUQuoted 41 lines · Zhao Liu
(Also cc "machine core" maintainers.)
Hi Xiaoyao,
Quoted 17 lines · Xiaoyao Li
Date: Mon, 15 Jan 2024 12:18:17 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 10/16] i386/cpu: Introduce cluster-id to X86CPUQuoted 12 lines · Zhao Liu
Quoted 11 lines · Xiaoyao Li
Date: Sun, 14 Jan 2024 21:49:18 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 10/16] i386/cpu: Introduce cluster-id to X86CPUQuoted 5 lines · Zhao Liu
From: Zhuocheng Ding <[email protected]>
Introduce cluster-id other than module-id to be consistent with CpuInstanceProperties.cluster-id, and this avoids the confusion of parameter names when hotplugging.
Quoted 7 lines
Quoted 4 lines
Quoted 2 lines
I don't think reusing 'cluster' from arm for x86's 'module' is a good idea. It introduces confusion around the code.
There is a precedent: generic "socket" v.s. i386 "package".
It's not the same thing. "socket" vs "package" is just software people and hardware people chose different name. It's just different naming issue.
No, it's a similar issue. Same physical device, different name only.
Furthermore, the topology was introduced for resource layout and silicon fabrication, and similar design ideas and fabrication processes are fairly consistent across common current arches. Therefore, it is possible to abstract similar topological hierarchies for different arches.
Quoted 3 lines
however, here it's reusing name issue while 'cluster' has been defined for x86. It does introduce confusion.
Quoted 6 lines
Quoted 3 lines
There's nothing fundamentally different between the x86 module and the generic cluster, is there? This is the reason that I don't agree with introducing "modules" in -smp.
generic cluster just means the cluster of processors, i.e, a group of cpus/lps. It is just a middle level between die and core.
Not sure if you mean the "cluster" device for TCG GDB? "cluster" device is different with "cluster" option in -smp.
When Yanan introduced the "cluster" option in -smp, he mentioned that it is for sharing L2 and L3 tags, which roughly corresponds to our module.
Quoted 5 lines
It can be the module level in intel, or tile level. Further, if per die lp number increases in the future, there might be more middle levels in intel between die and core. Then at that time, how to decide what level should cluster be mapped to?
Currently, there're 3 levels defined in SDM which are between die and
core: diegrp, tile and module. In our products, L2 is just sharing on the module, so the intel's module and the general cluster are the best match.
There are no commercially available machines for the other levels yet, so there's no way to ensure exactly what the future holds, but we should try to avoid fragmentation of the topology hierarchy and try to maintain the uniform and common topology hierarchies for QEMU.
Unless a new level for -smp is introduced in the future when an unsolvable problem is raised.
Quoted 71 lines
Quoted 71 lines
Quoted 8 lines
Quoted 5 lines
The direct definition of cluster is the level that is above the "core" and shares the hardware resources including L2. In this sense, arm's cluster is the same as x86's module.
then, what about intel implements tile level in the future? why ARM's 'cluster' is mapped to 'module', but not 'tile' ?
This depends on the actual need.
Module (for x86) and cluster (in general) are similar, and tile (for x86) is used for L3 in practice, so I use module rather than tile to map generic cluster.
And, it should be noted that x86 module is mapped to the generic cluster, not to ARM's. It's just that currently only ARM is using the clusters option in -smp.
I believe QEMU provides the abstract and unified topology hierarchies in -smp, not the arch-specific hierarchies.
Quoted 9 lines
reusing 'cluster' for 'module' is just a bad idea.
Quoted 2 lines
Though different arches have different naming styles, but QEMU's generic code still need the uniform topology hierarchy.
generic code can provide as many topology levels as it can. each ARCH can choose to use the ones it supports.
Quoted 6 lines
e.g.,
in
qapi/machine.json, it says,# The ordering from highest/coarsest to lowest/finest is:
# @drawers, @books, @sockets, @dies, @clusters, @cores, @threads.This ordering is well-defined...
Quoted 7 lines
#
# Different architectures support different subsets of topology# containers. # # For example, s390x does not have clusters and dies, and the socket # is the parent container of cores.Quoted 5 lines
we can update it to
# The ordering from highest/coarsest to lowest/finest is:
# @drawers, @books, @sockets, @dies, @clusters, @module, @cores, # @threads....but here it's impossible to figure out why cluster is above module, and even I can't come up with the difference between cluster and module.
Quoted 10 lines
#
# Different architectures support different subsets of topology# containers. # # For example, s390x does not have clusters and dies, and the socket # is the parent container of cores. # # For example, x86 does not have drawers and books, and does not support # cluster.Quoted 2 lines
even if cluster of x86 is supported someday in the future, we can remove the ordering requirement from above description.
x86's cluster is above the package.
Quoted 7 lines
Quoted 5 lines
To reserve this name for x86, we can't have the well-defined topology ordering.
But topology ordering is necessary in generic code, and many calculations depend on the topology ordering.
could you point me to the code?
Yes, e.g., there're 2 helpers: machine_topo_get_cores_per_socket() and machine_topo_get_threads_per_socket().
Quoted 22 lines
Quoted 22 lines
Quoted 21 lines
Quoted 10 lines
Quoted 9 lines
s390 just added 'drawer' and 'book' in cpu topology[1]. I think we can also add a module level for x86 instead of reusing cluster.
(This is also what I want to reply to the cover letter.)
Quoted 7 lines
These two new levels have the clear topological hierarchy relationship and don't duplicate existing ones.
"book" or "drawer" may correspond to intel's "cluster".
Maybe, in the future, we could support for arch-specific alias topologies in -smp.
I don't think we need alias, reusing 'cluster' for 'module' doesn't gain any benefit except avoid adding a new field in SMPconfiguration. All the other cluster code is ARM specific and x86 cannot share.
Quoted 6 lines
Quoted 3 lines
The point is that there is no difference between intel module and general cluster...Considering only the naming issue, even AMD has the "complex" to correspond to the Intel's "module".
does complex of AMD really match with intel module? L3 cache is shared in one complex, while L2 cache is shared in one module for now.
If then it could correspond to intel's tile, which is after all a level below die.
Thanks,
Zhao
Re: [PATCH v7 10/16] i386/cpu: Introduce cluster-id to X86CPU
Quoted 53 lines · Zhao Liu
Hi Xiaoyao,
Quoted 50 lines · Xiaoyao Li
Date: Mon, 15 Jan 2024 15:45:58 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 10/16] i386/cpu: Introduce cluster-id to X86CPUQuoted 44 lines · Zhao Liu
(Also cc "machine core" maintainers.)
u
Hi Xiaoyao,Quoted 17 lines · Xiaoyao Li
Date: Mon, 15 Jan 2024 12:18:17 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 10/16] i386/cpu: Introduce cluster-id to X86CPUQuoted 12 lines · Zhao Liu
Quoted 11 lines · Xiaoyao Li
Date: Sun, 14 Jan 2024 21:49:18 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 10/16] i386/cpu: Introduce cluster-id to X86CPUQuoted 5 lines · Zhao Liu
From: Zhuocheng Ding <[email protected]>
Introduce cluster-id other than module-id to be consistent with CpuInstanceProperties.cluster-id, and this avoids the confusion of parameter names when hotplugging.
Quoted 7 lines
Quoted 4 lines
Quoted 2 lines
I don't think reusing 'cluster' from arm for x86's 'module' is a good idea. It introduces confusion around the code.
There is a precedent: generic "socket" v.s. i386 "package".
It's not the same thing. "socket" vs "package" is just software people and hardware people chose different name. It's just different naming issue.
No, it's a similar issue. Same physical device, different name only.
Furthermore, the topology was introduced for resource layout and silicon fabrication, and similar design ideas and fabrication processes are fairly consistent across common current arches. Therefore, it is possible to abstract similar topological hierarchies for different arches.
Quoted 3 lines
however, here it's reusing name issue while 'cluster' has been defined for x86. It does introduce confusion.
There's nothing fundamentally different between the x86 module and the generic cluster, is there? This is the reason that I don't agree with introducing "modules" in -smp.
Quoted 5 lines
Quoted 2 lines
generic cluster just means the cluster of processors, i.e, a group of cpus/lps. It is just a middle level between die and core.
Not sure if you mean the "cluster" device for TCG GDB? "cluster" device is different with "cluster" option in -smp.
No, I just mean the word 'cluster'. And I thought what you called "generic cluster" means "a cluster of logical processors"
Below I quote the description of Yanan's commit 864c3b5c32f0:
A cluster generally means a group of CPU cores which share L2 cache or other mid-level resources, and it is the shared resources that is used to improve scheduler's behavior. From the point of view of the size range, it's between CPU die and CPU core. For example, on some ARM64 Kunpeng servers, we have 6 clusters in each NUMA node, and 4 CPU cores in each cluster. The 4 CPU cores share a separate L2 cache and a L3 cache tag, which brings cache affinity advantage.
What I get from it, is, cluster is just a middle level between CPU die and CPU core. The cpu cores inside one cluster shares some mid-level resource. L2 cache is just one example of the shared mid-level resource. So it can be either module level, or tile level in x86, or even the diegrp level you mentioned below.
Quoted 4 lines
When Yanan introduced the "cluster" option in -smp, he mentioned that it is for sharing L2 and L3 tags, which roughly corresponds to our module.
Quoted 1 line
Quoted 8 lines
Quoted 4 lines
It can be the module level in intel, or tile level. Further, if per die lp number increases in the future, there might be more middle levels in intel between die and core. Then at that time, how to decide what level should cluster be mapped to?
Currently, there're 3 levels defined in SDM which are between die and
core: diegrp, tile and module. In our products, L2 is just sharing on the module, so the intel's module and the general cluster are the best match.
you said 'generic cluster' a lot of times. But from my point of view, you are referring to current ARM's cluster instead of *generic* cluster.
Anyway, cluster is just a mid-level between die and core. We should not associate it any specific resource. A resource is shared in what level can change, e.g., initially L3 cache is shared in a physical package. When multi-die got supported, L3 cache is shared in one die. Now, on AMD product, L3 cache is shared in one complex, and one die can have 2 complexs thus 2 separate L3 cache in one die.
It doesn't matter calling it cluster, or module, or xyz. It is just a name to represent a cpu topology level between die and core. What matters is, once it gets accepted, it becomes formal ABI for users that 'cluster' means 'module' for x86 users. This is definitely a big confusion for people. Maybe people try to figure out why, and find the reason is that 'cluster' means the level at which L2 cache is shared and that's just the module level in x86 shares L2 cache. Maybe in the future, "L2 is shared in module" get changed just like the example I give for L3 above. Then, that's really the big confusion, and all this become the "historic reason" that cluster is chosen to represent module in x86.
Quoted 82 lines
There are no commercially available machines for the other levels yet, so there's no way to ensure exactly what the future holds, but we should try to avoid fragmentation of the topology hierarchy and try to maintain the uniform and common topology hierarchies for QEMU.
Unless a new level for -smp is introduced in the future when an unsolvable problem is raised.
Quoted 74 lines
Quoted 74 lines
Quoted 8 lines
Quoted 5 lines
The direct definition of cluster is the level that is above the "core" and shares the hardware resources including L2. In this sense, arm's cluster is the same as x86's module.
then, what about intel implements tile level in the future? why ARM's 'cluster' is mapped to 'module', but not 'tile' ?
This depends on the actual need.
Module (for x86) and cluster (in general) are similar, and tile (for x86) is used for L3 in practice, so I use module rather than tile to map generic cluster.
And, it should be noted that x86 module is mapped to the generic cluster, not to ARM's. It's just that currently only ARM is using the clusters option in -smp.
I believe QEMU provides the abstract and unified topology hierarchies in -smp, not the arch-specific hierarchies.
Quoted 9 lines
reusing 'cluster' for 'module' is just a bad idea.
Quoted 2 lines
Though different arches have different naming styles, but QEMU's generic code still need the uniform topology hierarchy.
generic code can provide as many topology levels as it can. each ARCH can choose to use the ones it supports.
Quoted 6 lines
e.g.,
in
qapi/machine.json, it says,# The ordering from highest/coarsest to lowest/finest is:
# @drawers, @books, @sockets, @dies, @clusters, @cores, @threads.This ordering is well-defined...
Quoted 7 lines
#
# Different architectures support different subsets of topology# containers. # # For example, s390x does not have clusters and dies, and the socket # is the parent container of cores.Quoted 5 lines
we can update it to
# The ordering from highest/coarsest to lowest/finest is:
# @drawers, @books, @sockets, @dies, @clusters, @module, @cores, # @threads....but here it's impossible to figure out why cluster is above module, and even I can't come up with the difference between cluster and module.
Quoted 10 lines
#
# Different architectures support different subsets of topology# containers. # # For example, s390x does not have clusters and dies, and the socket # is the parent container of cores. # # For example, x86 does not have drawers and books, and does not support # cluster.Quoted 2 lines
even if cluster of x86 is supported someday in the future, we can remove the ordering requirement from above description.
x86's cluster is above the package.
To reserve this name for x86, we can't have the well-defined topology ordering.
Quoted 7 lines
Quoted 4 lines
Quoted 2 lines
But topology ordering is necessary in generic code, and many calculations depend on the topology ordering.
could you point me to the code?
Yes, e.g., there're 2 helpers:
machine_topo_get_cores_per_socket()andmachine_topo_get_threads_per_socket().
I see. these two helpers are fragile, that they need to be updated every time new level between core and socket is introduced.
Anyway, we can ensure the order for each ARCH, that the valid levels for any ARCH are ordered. e.g., we have
@drawers, @books, @sockets, @dies, @clusters, @module, @cores, @threads
defined,
for s390, the valid levels are
@drawers, @books, @sockets, @cores, @threads
for arm, the valid levels are
@sockets, @dies, @clusters, @cores, @threads
(I'm not sure if die is supported for ARM?)
for x86, the valid levels are
@sockets, @dies, @module, @cores, @threads
All of them are ordered. those unsupported level in each ARCH just get value 1. It won't have any issue in the calculation for the default value, but you provided two functions may not be lucky. anyway, they can be fixed at the time when we really go this approach.
Quoted 26 lines
Quoted 26 lines
Quoted 25 lines
Quoted 13 lines
Quoted 13 lines
Quoted 9 lines
s390 just added 'drawer' and 'book' in cpu topology[1]. I think we can also add a module level for x86 instead of reusing cluster.
(This is also what I want to reply to the cover letter.)
These two new levels have the clear topological hierarchy relationship and don't duplicate existing ones.
Quoted 8 lines
Quoted 4 lines
"book" or "drawer" may correspond to intel's "cluster".
Maybe, in the future, we could support for arch-specific alias topologies in -smp.
I don't think we need alias, reusing 'cluster' for 'module' doesn't gain any benefit except avoid adding a new field in SMPconfiguration. All the other cluster code is ARM specific and x86 cannot share.
The point is that there is no difference between intel module and general cluster...Considering only the naming issue, even AMD has the "complex" to correspond to the Intel's "module".
Quoted 5 lines
Quoted 2 lines
does complex of AMD really match with intel module? L3 cache is shared in one complex, while L2 cache is shared in one module for now.
If then it could correspond to intel's tile, which is after all a level below die.
So if AMD wants to add complex in smp topology, where should complex level get put? between die and cluster?
Quoted 3 lines
Thanks,
Zhao
Re: [PATCH v7 10/16] i386/cpu: Introduce cluster-id to X86CPU
Quoted 77 lines · Xiaoyao Li
Date: Wed, 17 Jan 2024 00:40:12 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 10/16] i386/cpu: Introduce cluster-id to X86CPUQuoted 53 lines · Zhao Liu
Hi Xiaoyao,
Quoted 50 lines · Xiaoyao Li
Date: Mon, 15 Jan 2024 15:45:58 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 10/16] i386/cpu: Introduce cluster-id to X86CPUQuoted 44 lines · Zhao Liu
(Also cc "machine core" maintainers.)
u
Hi Xiaoyao,Quoted 17 lines · Xiaoyao Li
Date: Mon, 15 Jan 2024 12:18:17 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 10/16] i386/cpu: Introduce cluster-id to X86CPUQuoted 12 lines · Zhao Liu
Quoted 11 lines · Xiaoyao Li
Date: Sun, 14 Jan 2024 21:49:18 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 10/16] i386/cpu: Introduce cluster-id to X86CPUQuoted 5 lines · Zhao Liu
From: Zhuocheng Ding <[email protected]>
Introduce cluster-id other than module-id to be consistent with CpuInstanceProperties.cluster-id, and this avoids the confusion of parameter names when hotplugging.
Quoted 7 lines
Quoted 4 lines
Quoted 2 lines
I don't think reusing 'cluster' from arm for x86's 'module' is a good idea. It introduces confusion around the code.
There is a precedent: generic "socket" v.s. i386 "package".
It's not the same thing. "socket" vs "package" is just software people and hardware people chose different name. It's just different naming issue.
No, it's a similar issue. Same physical device, different name only.
Furthermore, the topology was introduced for resource layout and silicon fabrication, and similar design ideas and fabrication processes are fairly consistent across common current arches. Therefore, it is possible to abstract similar topological hierarchies for different arches.
Quoted 3 lines
however, here it's reusing name issue while 'cluster' has been defined for x86. It does introduce confusion.
There's nothing fundamentally different between the x86 module and the generic cluster, is there? This is the reason that I don't agree with introducing "modules" in -smp.
Quoted 5 lines
Quoted 2 lines
generic cluster just means the cluster of processors, i.e, a group of cpus/lps. It is just a middle level between die and core.
Not sure if you mean the "cluster" device for TCG GDB? "cluster" device is different with "cluster" option in -smp.
No, I just mean the word 'cluster'. And I thought what you called "generic cluster" means "a cluster of logical processors"
Below I quote the description of Yanan's commit
864c3b5c32f0:A cluster generally means a group of CPU cores which share L2 cache or other mid-level resources, and it is the shared resources that is used to improve scheduler's behavior. From the point of view of the size range, it's between CPU die and CPU core. For example, on some ARM64 Kunpeng servers, we have 6 clusters in each NUMA node, and 4 CPU cores in each cluster. The 4 CPU cores share a separate L2 cache and a L3 cache tag, which brings cache affinity advantage.
Quoted 2 lines
What I get from it, is, cluster is just a middle level between CPU die and CPU core.
Here the words "a group of CPU" is not the software concept, but a hardware topology.
Quoted 4 lines
The cpu cores inside one cluster shares some mid-level resource. L2 cache is just one example of the shared mid-level resource. So it can be either module level, or tile level in x86, or even the diegrp level you mentioned below.
In actual hardware design, ARM's cluster is close to intel's module.
I'm not seeing any examples of clusters being similar to tile or diegrp. Even within intel, the hardware architects have agreed definitions for each level.
Quoted 10 lines
Quoted 10 lines
When Yanan introduced the "cluster" option in -smp, he mentioned that it is for sharing L2 and L3 tags, which roughly corresponds to our module.
Quoted 5 lines
It can be the module level in intel, or tile level. Further, if per die lp number increases in the future, there might be more middle levels in intel between die and core. Then at that time, how to decide what level should cluster be mapped to?
Quoted 6 lines
Quoted 3 lines
Currently, there're 3 levels defined in SDM which are between die and
core: diegrp, tile and module. In our products, L2 is just sharing on the module, so the intel's module and the general cluster are the best match.you said 'generic cluster' a lot of times. But from my point of view, you are referring to current ARM's cluster instead of *generic* cluster.
No, I'm always talking about the "cluster" in -smp, not ARM specific cluster. ARM just maps its arch-specific cluster to the general one.
Quoted 9 lines
Anyway, cluster is just a mid-level between die and core. We should not associate it any specific resource. A resource is shared in what level can change, e.g., initially L3 cache is shared in a physical package. When multi-die got supported, L3 cache is shared in one die. Now, on AMD product, L3 cache is shared in one complex, and one die can have 2 complexs thus 2 separate L3 cache in one die.
It doesn't matter calling it cluster, or module, or xyz. It is just a name to represent a cpu topology level between die and core.
In the case of more complex topologies, QEMU may consider supporting aliasing.
Vender can support topology aliases for its own arch, but there is no possibility of discarding QEMU's unified topology hierarchy in favor of building arch-specific hierarchies.
Quoted 7 lines
What matters is,
once it gets accepted, it becomes formal ABI for users that 'cluster' means 'module' for x86 users. This is definitely a big confusion for people. Maybe people try to figure out why, and find the reason is that 'cluster' means the level at which L2 cache is shared and that's just the module level in x86 shares L2 cache. Maybe in the future, "L2 is shared in module" get changed just like the example I give for L3 above.
My decision to map modules to clusters is based on existing and future product topology characteristics, which are supported by hardware practices. Of course, in the upcoming our cache topology patch series, users can customize the cache topology hierarchy.
Quoted 1 line
Then, that's really the big confusion,
I don't think this will confuse the user. All details can be explained clearly by document.
Quoted 93 lines
and all this become the "historic reason" that cluster is chosen to represent module in x86.
Quoted 82 lines
There are no commercially available machines for the other levels yet, so there's no way to ensure exactly what the future holds, but we should try to avoid fragmentation of the topology hierarchy and try to maintain the uniform and common topology hierarchies for QEMU.
Unless a new level for -smp is introduced in the future when an unsolvable problem is raised.
Quoted 74 lines
Quoted 74 lines
Quoted 8 lines
Quoted 5 lines
The direct definition of cluster is the level that is above the "core" and shares the hardware resources including L2. In this sense, arm's cluster is the same as x86's module.
then, what about intel implements tile level in the future? why ARM's 'cluster' is mapped to 'module', but not 'tile' ?
This depends on the actual need.
Module (for x86) and cluster (in general) are similar, and tile (for x86) is used for L3 in practice, so I use module rather than tile to map generic cluster.
And, it should be noted that x86 module is mapped to the generic cluster, not to ARM's. It's just that currently only ARM is using the clusters option in -smp.
I believe QEMU provides the abstract and unified topology hierarchies in -smp, not the arch-specific hierarchies.
Quoted 9 lines
reusing 'cluster' for 'module' is just a bad idea.
Quoted 2 lines
Though different arches have different naming styles, but QEMU's generic code still need the uniform topology hierarchy.
generic code can provide as many topology levels as it can. each ARCH can choose to use the ones it supports.
Quoted 6 lines
e.g.,
in
qapi/machine.json, it says,# The ordering from highest/coarsest to lowest/finest is:
# @drawers, @books, @sockets, @dies, @clusters, @cores, @threads.This ordering is well-defined...
Quoted 7 lines
#
# Different architectures support different subsets of topology# containers. # # For example, s390x does not have clusters and dies, and the socket # is the parent container of cores.Quoted 5 lines
we can update it to
# The ordering from highest/coarsest to lowest/finest is:
# @drawers, @books, @sockets, @dies, @clusters, @module, @cores, # @threads....but here it's impossible to figure out why cluster is above module, and even I can't come up with the difference between cluster and module.
Quoted 10 lines
#
# Different architectures support different subsets of topology# containers. # # For example, s390x does not have clusters and dies, and the socket # is the parent container of cores. # # For example, x86 does not have drawers and books, and does not support # cluster.Quoted 2 lines
even if cluster of x86 is supported someday in the future, we can remove the ordering requirement from above description.
x86's cluster is above the package.
To reserve this name for x86, we can't have the well-defined topology ordering.
Quoted 7 lines
Quoted 4 lines
Quoted 2 lines
But topology ordering is necessary in generic code, and many calculations depend on the topology ordering.
could you point me to the code?
Yes, e.g., there're 2 helpers:
machine_topo_get_cores_per_socket()andmachine_topo_get_threads_per_socket().
Quoted 7 lines
I see. these two helpers are fragile, that they need to be updated every time new level between core and socket is introduced.
Anyway, we can ensure the order for each ARCH, that the valid levels for any ARCH are ordered. e.g., we have
@drawers, @books, @sockets, @dies, @clusters, @module, @cores, @threads
Sorry to repeat my previous objection: the order why cluster is above the module can't be well-defined. This is more confusing. It is not possible to add a new level without clearly defining the hierarchical relationship.
Different venders have their different names, there is no reason or possibility to cram all the names of all the vender into this arrangement, and it is not maintainable.
Quoted 47 lines
defined,
for s390, the valid levels are
@drawers, @books, @sockets, @cores, @threads
for arm, the valid levels are
@sockets, @dies, @clusters, @cores, @threads
(I'm not sure if die is supported for ARM?)for x86, the valid levels are
@sockets, @dies, @module, @cores, @threads
All of them are ordered. those unsupported level in each ARCH just get value
1. It won't have any issue in the calculation for the default value, but you provided two functions may not be lucky. anyway, they can be fixed at the time when we really go this approach.Quoted 26 lines
Quoted 26 lines
Quoted 25 lines
Quoted 13 lines
Quoted 13 lines
Quoted 9 lines
s390 just added 'drawer' and 'book' in cpu topology[1]. I think we can also add a module level for x86 instead of reusing cluster.
(This is also what I want to reply to the cover letter.)
These two new levels have the clear topological hierarchy relationship and don't duplicate existing ones.
Quoted 8 lines
Quoted 4 lines
"book" or "drawer" may correspond to intel's "cluster".
Maybe, in the future, we could support for arch-specific alias topologies in -smp.
I don't think we need alias, reusing 'cluster' for 'module' doesn't gain any benefit except avoid adding a new field in SMPconfiguration. All the other cluster code is ARM specific and x86 cannot share.
The point is that there is no difference between intel module and general cluster...Considering only the naming issue, even AMD has the "complex" to correspond to the Intel's "module".
Quoted 8 lines
Quoted 5 lines
Quoted 2 lines
does complex of AMD really match with intel module? L3 cache is shared in one complex, while L2 cache is shared in one module for now.
If then it could correspond to intel's tile, which is after all a level below die.
So if AMD wants to add complex in smp topology, where should complex level get put? between die and cluster?
That's just an example, and just showed that AMD and intel have naming differences, even if they are both x86. We can't be happy with everyone.
Thanks,
Zhao
Re: [PATCH v7 10/16] i386/cpu: Introduce cluster-id to X86CPU
Hi Xiaoyao,
Quoted 19 lines
Quoted 19 lines
Quoted 5 lines
Quoted 2 lines
generic cluster just means the cluster of processors, i.e, a group of cpus/lps. It is just a middle level between die and core.
Not sure if you mean the "cluster" device for TCG GDB? "cluster" device is different with "cluster" option in -smp.
No, I just mean the word 'cluster'. And I thought what you called "generic cluster" means "a cluster of logical processors"
Below I quote the description of Yanan's commit
864c3b5c32f0:A cluster generally means a group of CPU cores which share L2 cache or other mid-level resources, and it is the shared resources that is used to improve scheduler's behavior. From the point of view of the size range, it's between CPU die and CPU core. For example, on some ARM64 Kunpeng servers, we have 6 clusters in each NUMA node, and 4 CPU cores in each cluster. The 4 CPU cores share a separate L2 cache and a L3 cache tag, which brings cache affinity advantage.
Quoted 5 lines
Quoted 2 lines
What I get from it, is, cluster is just a middle level between CPU die and CPU core.
Here the words "a group of CPU" is not the software concept, but a hardware topology.
When I found this material:
https://www.kernel.org/doc/Documentation/devicetree/bindings/cpu/cpu-topology.txt
I realized the most essential difference between cluster and module is that cluster supports nesting, i.e. it can have nesting clusters as a layer of CPU topology.
Even though QEMU's description of cluster looked similar to module when it was introduced, it is impossible to envision whether ARM/RISCV and other device tree-based arches will continue to introduce nesting clusters in the future.
To avoid potential conflicts, it would be better to introduce modules for x86 to differentiate them from clusters.
Thanks,
Zhao
PATCH v7 11/16 tests: Add test case of APIC ID for module level parsing
In reply to Zhao LiuFrom: Zhuocheng Ding <[email protected]>
After i386 supports module level, it's time to add the test for module level's parsing.
--- tests/unit/test-x86-topo.c | 19 +++++++++++++++---- 1 file changed, 15 insertions(+), 4 deletions(-)
tests/unit/test-x86-topo.c +15 -4
@@ -37,6 +37,7 @@ static void test_topo_bits(void) topo_info = (X86CPUTopoInfo ) {1 ,1 ,1 ,1 }; g_assert_cmpuint(apicid_smt_width(&topo_info), ==, 0); g_assert_cmpuint(apicid_core_width(&topo_info), ==, 0); + g_assert_cmpuint(apicid_module_width(&topo_info), ==, 0); g_assert_cmpuint(apicid_die_width(&topo_info), ==, 0); topo_info = (X86CPUTopoInfo ) {1 ,1 ,1 ,1 };
@@ -74,13 +75,22 @@ static void test_topo_bits(void) topo_info = (X86CPUTopoInfo ) {1 ,1 ,33 ,2 }; g_assert_cmpuint(apicid_core_width(&topo_info), ==, 6); - topo_info = (X86CPUTopoInfo ) {1 ,1 ,30 ,2 }; + topo_info = (X86CPUTopoInfo ) {1 ,6 ,30 ,2 }; + g_assert_cmpuint(apicid_module_width(&topo_info), ==, 3); + topo_info = (X86CPUTopoInfo ) {1 ,7 ,30 ,2 }; + g_assert_cmpuint(apicid_module_width(&topo_info), ==, 3); + topo_info = (X86CPUTopoInfo ) {1 ,8 ,30 ,2 }; + g_assert_cmpuint(apicid_module_width(&topo_info), ==, 3); + topo_info = (X86CPUTopoInfo ) {1 ,9 ,30 ,2 }; + g_assert_cmpuint(apicid_module_width(&topo_info), ==, 4); + + topo_info = (X86CPUTopoInfo ) {1 ,6 ,30 ,2 }; g_assert_cmpuint(apicid_die_width(&topo_info), ==, 0); - topo_info = (X86CPUTopoInfo ) {2 ,1 ,30 ,2 }; + topo_info = (X86CPUTopoInfo ) {2 ,6 ,30 ,2 }; g_assert_cmpuint(apicid_die_width(&topo_info), ==, 1); - topo_info = (X86CPUTopoInfo ) {3 ,1 ,30 ,2 }; + topo_info = (X86CPUTopoInfo ) {3 ,6 ,30 ,2 }; g_assert_cmpuint(apicid_die_width(&topo_info), ==, 2); - topo_info = (X86CPUTopoInfo ) {4 ,1 ,30 ,2 }; + topo_info = (X86CPUTopoInfo ) {4 ,6 ,30 ,2 }; g_assert_cmpuint(apicid_die_width(&topo_info), ==, 2);/* build a weird topology and see if IDs are calculated correctly
@@ -91,6 +101,7 @@ static void test_topo_bits(void) topo_info = (X86CPUTopoInfo ) {1 ,1 ,6 ,3 }; g_assert_cmpuint(apicid_smt_width(&topo_info), ==, 2); g_assert_cmpuint(apicid_core_offset(&topo_info), ==, 2); + g_assert_cmpuint(apicid_module_offset(&topo_info), ==, 5); g_assert_cmpuint(apicid_die_offset(&topo_info), ==, 5); g_assert_cmpuint(apicid_pkg_offset(&topo_info), ==, 5);
Signature · 1 line
2.34.1
PATCH v7 12/16 hw/i386/pc: Support smp.clusters for x86 PC machine
In reply to Zhao LiuFrom: Zhuocheng Ding <[email protected]>
As module-level topology support is added to X86CPU, now we can enable the support for the cluster parameter on PC machines. With this support, we can define a 5-level x86 CPU topology with "-smp":
-smp cpus=*,maxcpus=*,sockets=*,dies=*,clusters=*,cores=*,threads=*.
Additionally, add the 5-level topology example in description of "-smp".
--- hw/i386/pc.c | 1 + qemu-options.hx | 10 +++++----- 2 files changed, 6 insertions(+), 5 deletions(-)
hw/i386/pc.c +1 -0
@@ -1849,6 +1849,7 @@ static void pc_machine_class_init(ObjectClass *oc, void *data) mc->default_cpu_type =TARGET_DEFAULT_CPU_TYPE ; mc->nvdimm_supported = true; mc->smp_props.dies_supported = true; + mc->smp_props.clusters_supported = true; mc->default_ram_id ="pc.ram" ; pcmc->default_smbios_ep_type =SMBIOS_ENTRY_POINT_TYPE_64 ;
qemu-options.hx +5 -5
@@ -337,14 +337,14 @@ SRST -smp 8,sockets=2,cores=2,threads=2,maxcpus=8 The following sub-option defines a CPU topology hierarchy (2 sockets - totally on the machine, 2 dies per socket, 2 cores per die, 2 threads - per core) for PC machines which support sockets/dies/cores/threads. - Some members of the option can be omitted but their values will be - automatically computed: + totally on the machine, 2 dies per socket, 2 clusters per die, 2 cores per + cluster, 2 threads per core) for PC machines which support sockets/dies + /clusters/cores/threads. Some members of the option can be omitted but + their values will be automatically computed: :: - -smp 16,sockets=2,dies=2,cores=2,threads=2,maxcpus=16 + -smp 32,sockets=2,dies=2,clusters=2,cores=2,threads=2,maxcpus=32 The following sub-option defines a CPU topology hierarchy (2 sockets totally on the machine, 2 clusters per socket, 2 cores per cluster,
Signature · 1 line
2.34.1
PATCH v7 13/16 i386: Add cache topology info in CPUCacheInfo
In reply to Zhao LiuFrom: Zhao Liu <[email protected]>
Currently, by default, the cache topology is encoded as:
1. i/d cache is shared in one core.
2. L2 cache is shared in one core.
3. L3 cache is shared in one die.
This default general setting has caused a misunderstanding, that is, the cache topology is completely equated with a specific cpu topology, such as the connection between L2 cache and core level, and the connection between L3 cache and die level.
In fact, the settings of these topologies depend on the specific platform and are not static. For example, on Alder Lake-P, every four Atom cores share the same L2 cache.
Thus, we should explicitly define the corresponding cache topology for different cache models to increase scalability.
Except legacy_l2_cache_cpuid2 (its default topo level is CPU_TOPO_LEVEL_UNKNOW), explicitly set the corresponding topology level for all other cache models. In order to be compatible with the existing cache topology, set the CPU_TOPO_LEVEL_CORE level for the i/d cache, set the CPU_TOPO_LEVEL_CORE level for L2 cache, and set the CPU_TOPO_LEVEL_DIE level for L3 cache.
The field for CPUID[4].EAX[bits 25:14] or CPUID[0x8000001D].EAX[bits 25:14] will be set based on CPUCacheInfo.share_level.
---
Changes since v3:
* Fix cache topology uninitialization bugs for some AMD CPUs. (Babu)
* Move the CPUTopoLevel enumeration definition to the previous 0x1f rework patch.
Changes since v1:
* Add the prefix "CPU_TOPO_LEVEL_*" for CPU topology level names. (Yanan)
* (Revert, pls refer "i386: Decouple CPUID[0x1F] subleaf with specific topology level") Rename the "INVALID" level to CPU_TOPO_LEVEL_UNKNOW. (Yanan)
--- target/i386/cpu.c | 36 ++++++++++++++++++++++++++++++++++++ target/i386/cpu.h | 7 +++++++ 2 files changed, 43 insertions(+)
target/i386/cpu.c +36 -0
@@ -558,6 +558,7 @@ static CPUCacheInfo legacy_l1d_cache = { .sets = 64, .partitions = 1, .no_invd_sharing = true, + .share_level = CPU_TOPO_LEVEL_CORE, };/*FIXME: CPUID leaf 0x80000005 is inconsistent with leaves 2 & 4 */
@@ -572,6 +573,7 @@ static CPUCacheInfo legacy_l1d_cache_amd = { .partitions = 1, .lines_per_tag = 1, .no_invd_sharing = true, + .share_level = CPU_TOPO_LEVEL_CORE, };/* L1 instruction cache: */
@@ -585,6 +587,7 @@ static CPUCacheInfo legacy_l1i_cache = { .sets = 64, .partitions = 1, .no_invd_sharing = true, + .share_level = CPU_TOPO_LEVEL_CORE, };/*FIXME: CPUID leaf 0x80000005 is inconsistent with leaves 2 & 4 */
@@ -599,6 +602,7 @@ static CPUCacheInfo legacy_l1i_cache_amd = { .partitions = 1, .lines_per_tag = 1, .no_invd_sharing = true, + .share_level = CPU_TOPO_LEVEL_CORE, };/* Level 2 unified cache: */
@@ -612,6 +616,7 @@ static CPUCacheInfo legacy_l2_cache = { .sets = 4096, .partitions = 1, .no_invd_sharing = true, + .share_level = CPU_TOPO_LEVEL_CORE, };/*FIXME: CPUID leaf 2 descriptor is inconsistent with CPUID leaf 4 */
@@ -621,6 +626,7 @@ static CPUCacheInfo legacy_l2_cache_cpuid2 = { .size = 2 * MiB, .line_size = 64, .associativity = 8, + .share_level = CPU_TOPO_LEVEL_INVALID, };
@@ -634,6 +640,7 @@ static CPUCacheInfo legacy_l2_cache_amd = { .associativity = 16, .sets = 512, .partitions = 1, + .share_level = CPU_TOPO_LEVEL_CORE, };/* Level 3 unified cache: */
@@ -649,6 +656,7 @@ static CPUCacheInfo legacy_l3_cache = { .self_init = true, .inclusive = true, .complex_indexing = true, + .share_level = CPU_TOPO_LEVEL_DIE, };/* TLB definitions: */
@@ -1947,6 +1955,7 @@ static const CPUCaches epyc_cache_info = { .lines_per_tag = 1, .self_init = 1, .no_invd_sharing = true, + .share_level = CPU_TOPO_LEVEL_CORE, }, .l1i_cache = &(CPUCacheInfo ) { .type =INSTRUCTION_CACHE ,
@@ -1959,6 +1968,7 @@ static const CPUCaches epyc_cache_info = { .lines_per_tag = 1, .self_init = 1, .no_invd_sharing = true, + .share_level = CPU_TOPO_LEVEL_CORE, }, .l2_cache = &(CPUCacheInfo ) { .type =UNIFIED_CACHE ,
@@ -1969,6 +1979,7 @@ static const CPUCaches epyc_cache_info = { .partitions = 1, .sets = 1024, .lines_per_tag = 1, + .share_level = CPU_TOPO_LEVEL_CORE, }, .l3_cache = &(CPUCacheInfo ) { .type =UNIFIED_CACHE ,
@@ -1982,6 +1993,7 @@ static const CPUCaches epyc_cache_info = { .self_init = true, .inclusive = true, .complex_indexing = true, + .share_level = CPU_TOPO_LEVEL_DIE, }, };
@@ -1997,6 +2009,7 @@ static CPUCaches epyc_v4_cache_info = { .lines_per_tag = 1, .self_init = 1, .no_invd_sharing = true, + .share_level = CPU_TOPO_LEVEL_CORE, }, .l1i_cache = &(CPUCacheInfo ) { .type =INSTRUCTION_CACHE ,
@@ -2009,6 +2022,7 @@ static CPUCaches epyc_v4_cache_info = { .lines_per_tag = 1, .self_init = 1, .no_invd_sharing = true, + .share_level = CPU_TOPO_LEVEL_CORE, }, .l2_cache = &(CPUCacheInfo ) { .type =UNIFIED_CACHE ,
@@ -2019,6 +2033,7 @@ static CPUCaches epyc_v4_cache_info = { .partitions = 1, .sets = 1024, .lines_per_tag = 1, + .share_level = CPU_TOPO_LEVEL_CORE, }, .l3_cache = &(CPUCacheInfo ) { .type =UNIFIED_CACHE ,
@@ -2032,6 +2047,7 @@ static CPUCaches epyc_v4_cache_info = { .self_init = true, .inclusive = true, .complex_indexing = false, + .share_level = CPU_TOPO_LEVEL_DIE, }, };
@@ -2047,6 +2063,7 @@ static const CPUCaches epyc_rome_cache_info = { .lines_per_tag = 1, .self_init = 1, .no_invd_sharing = true, + .share_level = CPU_TOPO_LEVEL_CORE, }, .l1i_cache = &(CPUCacheInfo ) { .type =INSTRUCTION_CACHE ,
@@ -2059,6 +2076,7 @@ static const CPUCaches epyc_rome_cache_info = { .lines_per_tag = 1, .self_init = 1, .no_invd_sharing = true, + .share_level = CPU_TOPO_LEVEL_CORE, }, .l2_cache = &(CPUCacheInfo ) { .type =UNIFIED_CACHE ,
@@ -2069,6 +2087,7 @@ static const CPUCaches epyc_rome_cache_info = { .partitions = 1, .sets = 1024, .lines_per_tag = 1, + .share_level = CPU_TOPO_LEVEL_CORE, }, .l3_cache = &(CPUCacheInfo ) { .type =UNIFIED_CACHE ,
@@ -2082,6 +2101,7 @@ static const CPUCaches epyc_rome_cache_info = { .self_init = true, .inclusive = true, .complex_indexing = true, + .share_level = CPU_TOPO_LEVEL_DIE, }, };
@@ -2097,6 +2117,7 @@ static const CPUCaches epyc_rome_v3_cache_info = { .lines_per_tag = 1, .self_init = 1, .no_invd_sharing = true, + .share_level = CPU_TOPO_LEVEL_CORE, }, .l1i_cache = &(CPUCacheInfo ) { .type =INSTRUCTION_CACHE ,
@@ -2109,6 +2130,7 @@ static const CPUCaches epyc_rome_v3_cache_info = { .lines_per_tag = 1, .self_init = 1, .no_invd_sharing = true, + .share_level = CPU_TOPO_LEVEL_CORE, }, .l2_cache = &(CPUCacheInfo ) { .type =UNIFIED_CACHE ,
@@ -2119,6 +2141,7 @@ static const CPUCaches epyc_rome_v3_cache_info = { .partitions = 1, .sets = 1024, .lines_per_tag = 1, + .share_level = CPU_TOPO_LEVEL_CORE, }, .l3_cache = &(CPUCacheInfo ) { .type =UNIFIED_CACHE ,
@@ -2132,6 +2155,7 @@ static const CPUCaches epyc_rome_v3_cache_info = { .self_init = true, .inclusive = true, .complex_indexing = false, + .share_level = CPU_TOPO_LEVEL_DIE, }, };
@@ -2147,6 +2171,7 @@ static const CPUCaches epyc_milan_cache_info = { .lines_per_tag = 1, .self_init = 1, .no_invd_sharing = true, + .share_level = CPU_TOPO_LEVEL_CORE, }, .l1i_cache = &(CPUCacheInfo ) { .type =INSTRUCTION_CACHE ,
@@ -2159,6 +2184,7 @@ static const CPUCaches epyc_milan_cache_info = { .lines_per_tag = 1, .self_init = 1, .no_invd_sharing = true, + .share_level = CPU_TOPO_LEVEL_CORE, }, .l2_cache = &(CPUCacheInfo ) { .type =UNIFIED_CACHE ,
@@ -2169,6 +2195,7 @@ static const CPUCaches epyc_milan_cache_info = { .partitions = 1, .sets = 1024, .lines_per_tag = 1, + .share_level = CPU_TOPO_LEVEL_CORE, }, .l3_cache = &(CPUCacheInfo ) { .type =UNIFIED_CACHE ,
@@ -2182,6 +2209,7 @@ static const CPUCaches epyc_milan_cache_info = { .self_init = true, .inclusive = true, .complex_indexing = true, + .share_level = CPU_TOPO_LEVEL_DIE, }, };
@@ -2197,6 +2225,7 @@ static const CPUCaches epyc_milan_v2_cache_info = { .lines_per_tag = 1, .self_init = 1, .no_invd_sharing = true, + .share_level = CPU_TOPO_LEVEL_CORE, }, .l1i_cache = &(CPUCacheInfo ) { .type =INSTRUCTION_CACHE ,
@@ -2209,6 +2238,7 @@ static const CPUCaches epyc_milan_v2_cache_info = { .lines_per_tag = 1, .self_init = 1, .no_invd_sharing = true, + .share_level = CPU_TOPO_LEVEL_CORE, }, .l2_cache = &(CPUCacheInfo ) { .type =UNIFIED_CACHE ,
@@ -2219,6 +2249,7 @@ static const CPUCaches epyc_milan_v2_cache_info = { .partitions = 1, .sets = 1024, .lines_per_tag = 1, + .share_level = CPU_TOPO_LEVEL_CORE, }, .l3_cache = &(CPUCacheInfo ) { .type =UNIFIED_CACHE ,
@@ -2232,6 +2263,7 @@ static const CPUCaches epyc_milan_v2_cache_info = { .self_init = true, .inclusive = true, .complex_indexing = false, + .share_level = CPU_TOPO_LEVEL_DIE, }, };
@@ -2247,6 +2279,7 @@ static const CPUCaches epyc_genoa_cache_info = { .lines_per_tag = 1, .self_init = 1, .no_invd_sharing = true, + .share_level = CPU_TOPO_LEVEL_CORE, }, .l1i_cache = &(CPUCacheInfo ) { .type =INSTRUCTION_CACHE ,
@@ -2259,6 +2292,7 @@ static const CPUCaches epyc_genoa_cache_info = { .lines_per_tag = 1, .self_init = 1, .no_invd_sharing = true, + .share_level = CPU_TOPO_LEVEL_CORE, }, .l2_cache = &(CPUCacheInfo ) { .type =UNIFIED_CACHE ,
@@ -2269,6 +2303,7 @@ static const CPUCaches epyc_genoa_cache_info = { .partitions = 1, .sets = 2048, .lines_per_tag = 1, + .share_level = CPU_TOPO_LEVEL_CORE, }, .l3_cache = &(CPUCacheInfo ) { .type =UNIFIED_CACHE ,
@@ -2282,6 +2317,7 @@ static const CPUCaches epyc_genoa_cache_info = { .self_init = true, .inclusive = true, .complex_indexing = false, + .share_level = CPU_TOPO_LEVEL_DIE, }, };
target/i386/cpu.h +7 -0
@@ -1597,6 +1597,13 @@ typedef struct CPUCacheInfo {* address bits. CPUID[4].EDX[bit 2]. */ bool complex_indexing; + +/* +* Cache Topology. The level that cache is shared in. +* Used to encode CPUID[4].EAX[bits 25:14] or +* CPUID[0x8000001D].EAX[bits 25:14]. +*/ +enum CPUTopoLevel share_level; } CPUCacheInfo;
Signature · 1 line
2.34.1
PATCH v7 14/16 i386: Use CPUCacheInfo.share_level to encode CPUID[4]
In reply to Zhao LiuFrom: Zhao Liu <[email protected]>
CPUID[4].EAX[bits 25:14] is used to represent the cache topology for Intel CPUs.
After cache models have topology information, we can use CPUCacheInfo.share_level to decide which topology level to be encoded into CPUID[4].EAX[bits 25:14].
And since maximum_processor_id (original "num_apic_ids") is parsed based on cpu topology levels, which are verified when parsing smp, it's no need to check this value by "assert(num_apic_ids > 0)" again, so remove this assert.
Additionally, wrap the encoding of CPUID[4].EAX[bits 31:26] into a helper to make the code cleaner.
---
Changes since v1:
* Use "enum CPUTopoLevel share_level" as the parameter in max_processor_ids_for_cache().
* Make cache_into_passthrough case also use max_processor_ids_for_cache() and max_core_ids_in_package() to encode CPUID[4]. (Yanan)
* Rename the title of this patch (the original is "i386: Use CPUCacheInfo.share_level to encode CPUID[4].EAX[bits 25:14]").
--- target/i386/cpu.c | 70 +++++++++++++++++++++++++++++------------------ 1 file changed, 43 insertions(+), 27 deletions(-)
target/i386/cpu.c +43 -27
@@ -235,22 +235,53 @@ static uint8_t cpuid2_cache_descriptor(CPUCacheInfo *cache) ((t) == UNIFIED_CACHE) ? CACHE_TYPE_UNIFIED : \ 0 /* Invalid value */) +static uint32_t max_processor_ids_for_cache (X86CPUTopoInfo *topo_info, +enum CPUTopoLevel share_level) +{ +uint32_t num_ids =0 ; + +switch (share_level) { +case CPU_TOPO_LEVEL_CORE : + num_ids =1 <<apicid_core_offset (topo_info); +break ; +case CPU_TOPO_LEVEL_DIE : + num_ids =1 <<apicid_die_offset (topo_info); +break ; +case CPU_TOPO_LEVEL_PACKAGE : + num_ids =1 <<apicid_pkg_offset (topo_info); +break ; +default : +/* +* Currently there is no use case for SMT and MODULE, so use +* assert directly to facilitate debugging. +*/ +g_assert_not_reached (); + } + +return num_ids -1 ; +} + +static uint32_t max_core_ids_in_package (X86CPUTopoInfo *topo_info) +{ +uint32_t num_cores =1 << (apicid_pkg_offset (topo_info) - +apicid_core_offset (topo_info)); +return num_cores -1 ; +}/* Encode cache info for CPUID[4] */ static void encode_cache_cpuid4 (CPUCacheInfo *cache, -int num_apic_ids,int num_cores, +X86CPUTopoInfo *topo_info,uint32_t *eax,uint32_t *ebx,uint32_t *ecx,uint32_t *edx) {assert (cache->size == cache->line_size * cache->associativity * cache->partitions * cache->sets); -assert (num_apic_ids >0 ); *eax =CACHE_TYPE (cache->type) |CACHE_LEVEL (cache->level) | (cache->self_init ?CACHE_SELF_INIT_LEVEL :0 ) | - ((num_cores -1 ) <<26 ) | - ((num_apic_ids -1 ) <<14 ); + (max_core_ids_in_package (topo_info) <<26 ) | + (max_processor_ids_for_cache (topo_info, cache->share_level) <<14 );assert (cache->line_size >0 );assert (cache->partitions >0 );
@@ -6263,56 +6294,41 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,int host_vcpus_per_cache =1 + ((*eax &0x3FFC000 ) >>14 );if (cores_per_pkg >1 ) { -int addressable_cores_offset = -apicid_pkg_offset (&topo_info) - -apicid_core_offset (&topo_info); - *eax &= ~0xFC000000 ; - *eax |= (1 << (addressable_cores_offset -1 )) <<26 ; + *eax |=max_core_ids_in_package (&topo_info) <<26 ; }if (host_vcpus_per_cache > cpus_per_pkg) { -int pkg_offset =apicid_pkg_offset (&topo_info); - *eax &= ~0x3FFC000 ; - *eax |= (1 << (pkg_offset -1 )) <<14 ; + *eax |= +max_processor_ids_for_cache (&topo_info, +CPU_TOPO_LEVEL_PACKAGE ) <<14 ; } } }else if (cpu->vendor_cpuid_only &&IS_AMD_CPU (env)) { *eax = *ebx = *ecx = *edx =0 ; }else { *eax =0 ; -int addressable_cores_offset =apicid_pkg_offset (&topo_info) - -apicid_core_offset (&topo_info); -int core_offset, die_offset;switch (count) {case 0 :/* L1 dcache info */ - core_offset =apicid_core_offset (&topo_info);encode_cache_cpuid4 (env->cache_info_cpuid4.l1d_cache, - (1 << core_offset), - (1 << addressable_cores_offset), + &topo_info, eax, ebx, ecx, edx);break ;case 1 :/* L1 icache info */ - core_offset =apicid_core_offset (&topo_info);encode_cache_cpuid4 (env->cache_info_cpuid4.l1i_cache, - (1 << core_offset), - (1 << addressable_cores_offset), + &topo_info, eax, ebx, ecx, edx);break ;case 2 :/* L2 cache info */ - core_offset =apicid_core_offset (&topo_info);encode_cache_cpuid4 (env->cache_info_cpuid4.l2_cache, - (1 << core_offset), - (1 << addressable_cores_offset), + &topo_info, eax, ebx, ecx, edx);break ;case 3 :/* L3 cache info */ - die_offset =apicid_die_offset (&topo_info);if (cpu->enable_l3_cache) {encode_cache_cpuid4 (env->cache_info_cpuid4.l3_cache, - (1 << die_offset), - (1 << addressable_cores_offset), + &topo_info, eax, ebx, ecx, edx);break ; }
Signature · 1 line
2.34.1
Re: [PATCH v7 14/16] i386: Use CPUCacheInfo.share_level to encode CPUID[4]
Quoted 39 lines · Zhao Liu
From: Zhao Liu <[email protected]>
CPUID[4].EAX[bits 25:14] is used to represent the cache topology for Intel CPUs.
After cache models have topology information, we can use CPUCacheInfo.share_level to decide which topology level to be encoded into CPUID[4].EAX[bits 25:14].
And since
maximum_processor_id(original "num_apic_ids") is parsed based on cpu topology levels, which are verified when parsing smp, it's no need to check this value by "assert(num_apic_ids > 0)" again, so remove this assert.Additionally, wrap the encoding of CPUID[4].EAX[bits 31:26] into a helper to make the code cleaner.
Signed-off-by Zhao Liu <[email protected]>Tested-by Babu Moger <[email protected]>Tested-by Yongwei Ma <[email protected]>---
Changes since v1:
* Use "enum CPUTopoLevelshare_level" as the parameter inmax_processor_ids_for_cache().
* Makecache_into_passthroughcase also usemax_processor_ids_for_cache()andmax_core_ids_in_package()to encode CPUID[4]. (Yanan)
* Rename the title of this patch (the original is "i386: UseCPUCacheInfo.share_level to encode CPUID[4].EAX[bits 25:14]"). --- target/i386/cpu.c | 70 +++++++++++++++++++++++++++++------------------ 1 file changed, 43 insertions(+), 27 deletions(-)target/i386/cpu.c +0
-0
target/i386/cpu.c static uint8_t cpuid2_cache_descriptor(CPUCacheInfo *cache) · line 239
Quoted 5 lines
target/i386/cpu.c +2
-0@@ -235,22 +235,53 @@ static uint8_t cpuid2_cache_descriptor(CPUCacheInfo *cache) ((t) == UNIFIED_CACHE) ? CACHE_TYPE_UNIFIED : \ 0 /* Invalid value */) +static uint32_t max_processor_ids_for_cache (X86CPUTopoInfo *topo_info, +enum CPUTopoLevel share_level)
I prefer the name to max_lp_ids_share_the_cache()
Quoted 12 lines
target/i386/cpu.c +12
-0+{ +uint32_t num_ids =0 ; + +switch (share_level) { +case CPU_TOPO_LEVEL_CORE : + num_ids =1 <<apicid_core_offset (topo_info); +break ; +case CPU_TOPO_LEVEL_DIE : + num_ids =1 <<apicid_die_offset (topo_info); +break ; +case CPU_TOPO_LEVEL_PACKAGE : + num_ids =1 <<apicid_pkg_offset (topo_info);
target/i386/cpu.c static uint8_t cpuid2_cache_descriptor(CPUCacheInfo *cache) · line 261
Quoted 10 lines
target/i386/cpu.c +10
-0+break ; +default : +/* +* Currently there is no use case for SMT and MODULE, so use +* assert directly to facilitate debugging. +*/ +g_assert_not_reached (); + } + +return num_ids -1 ;
suggest to just return num_ids, and let the caller to do the -1 work.
target/i386/cpu.c static uint8_t cpuid2_cache_descriptor(CPUCacheInfo *cache) · line 268
Quoted 7 lines
target/i386/cpu.c +7
-0+} + +static uint32_t max_core_ids_in_package (X86CPUTopoInfo *topo_info) +{ +uint32_t num_cores =1 << (apicid_pkg_offset (topo_info) - +apicid_core_offset (topo_info)); +return num_cores -1 ;
ditto.
Quoted 10 lines
target/i386/cpu.c +2
-1+}/* Encode cache info for CPUID[4] */ static void encode_cache_cpuid4(CPUCacheInfo *cache, - int num_apic_ids, int num_cores, + X86CPUTopoInfo *topo_info, uint32_t *eax, uint32_t *ebx, uint32_t *ecx, uint32_t *edx) { assert(cache->size == cache->line_size * cache->associativity *
target/i386/cpu.c static uint8_t cpuid2_cache_descriptor(CPUCacheInfo *cache) · line 284
Quoted 10 lines
target/i386/cpu.c +2
-3cache->partitions * cache->sets); -assert (num_apic_ids >0 ); *eax =CACHE_TYPE (cache->type) |CACHE_LEVEL (cache->level) | (cache->self_init ?CACHE_SELF_INIT_LEVEL :0 ) | - ((num_cores -1 ) <<26 ) | - ((num_apic_ids -1 ) <<14 ); + (max_core_ids_in_package (topo_info) <<26 ) | + (max_processor_ids_for_cache (topo_info, cache->share_level) <<14 );
by the way, we can change the order of the two line. :)
Quoted 68 lines
target/i386/cpu.c +8
-23assert (cache->line_size >0 );assert (cache->partitions >0 );@@ -6263,56 +6294,41 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,int host_vcpus_per_cache =1 + ((*eax &0x3FFC000 ) >>14 );if (cores_per_pkg >1 ) { -int addressable_cores_offset = -apicid_pkg_offset (&topo_info) - -apicid_core_offset (&topo_info); - *eax &= ~0xFC000000 ; - *eax |= (1 << (addressable_cores_offset -1 )) <<26 ; + *eax |=max_core_ids_in_package (&topo_info) <<26 ; }if (host_vcpus_per_cache > cpus_per_pkg) { -int pkg_offset =apicid_pkg_offset (&topo_info); - *eax &= ~0x3FFC000 ; - *eax |= (1 << (pkg_offset -1 )) <<14 ; + *eax |= +max_processor_ids_for_cache (&topo_info, +CPU_TOPO_LEVEL_PACKAGE ) <<14 ; } } }else if (cpu->vendor_cpuid_only &&IS_AMD_CPU (env)) { *eax = *ebx = *ecx = *edx =0 ; }else { *eax =0 ; -int addressable_cores_offset =apicid_pkg_offset (&topo_info) - -apicid_core_offset (&topo_info); -int core_offset, die_offset;switch (count) {case 0 :/* L1 dcache info */ - core_offset =apicid_core_offset (&topo_info);encode_cache_cpuid4 (env->cache_info_cpuid4.l1d_cache, - (1 << core_offset), - (1 << addressable_cores_offset), + &topo_info, eax, ebx, ecx, edx);break ;case 1 :/* L1 icache info */ - core_offset =apicid_core_offset (&topo_info);encode_cache_cpuid4 (env->cache_info_cpuid4.l1i_cache, - (1 << core_offset), - (1 << addressable_cores_offset), + &topo_info, eax, ebx, ecx, edx);break ;case 2 :/* L2 cache info */ - core_offset =apicid_core_offset (&topo_info);encode_cache_cpuid4 (env->cache_info_cpuid4.l2_cache, - (1 << core_offset), - (1 << addressable_cores_offset), + &topo_info, eax, ebx, ecx, edx);break ;case 3 :/* L3 cache info */ - die_offset =apicid_die_offset (&topo_info);if (cpu->enable_l3_cache) {encode_cache_cpuid4 (env->cache_info_cpuid4.l3_cache, - (1 << die_offset), - (1 << addressable_cores_offset), + &topo_info, eax, ebx, ecx, edx);break ; }
Re: [PATCH v7 14/16] i386: Use CPUCacheInfo.share_level to encode CPUID[4]
Hi Xiaoyao,
Quoted 40 lines · Xiaoyao Li
Date: Sun, 14 Jan 2024 22:31:50 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 14/16] i386: Use CPUCacheInfo.share_level to encode
CPUID[4]Quoted 34 lines · Zhao Liu
From: Zhao Liu <[email protected]>
CPUID[4].EAX[bits 25:14] is used to represent the cache topology for Intel CPUs.
After cache models have topology information, we can use CPUCacheInfo.share_level to decide which topology level to be encoded into CPUID[4].EAX[bits 25:14].
And since
maximum_processor_id(original "num_apic_ids") is parsed based on cpu topology levels, which are verified when parsing smp, it's no need to check this value by "assert(num_apic_ids > 0)" again, so remove this assert.Additionally, wrap the encoding of CPUID[4].EAX[bits 31:26] into a helper to make the code cleaner.
Signed-off-by Zhao Liu <[email protected]>Tested-by Babu Moger <[email protected]>Tested-by Yongwei Ma <[email protected]>---
Changes since v1:
* Use "enum CPUTopoLevelshare_level" as the parameter inmax_processor_ids_for_cache().
* Makecache_into_passthroughcase also usemax_processor_ids_for_cache()andmax_core_ids_in_package()to encode CPUID[4]. (Yanan)
* Rename the title of this patch (the original is "i386: UseCPUCacheInfo.share_level to encode CPUID[4].EAX[bits 25:14]"). --- target/i386/cpu.c | 70 +++++++++++++++++++++++++++++------------------ 1 file changed, 43 insertions(+), 27 deletions(-)
Quoted 11 lines
target/i386/cpu.c static uint8_t cpuid2_cache_descriptor(CPUCacheInfo *cache) · line 238
Quoted 9 lines
target/i386/cpu.c +2
-0@@ -235,22 +235,53 @@ static uint8_t cpuid2_cache_descriptor(CPUCacheInfo *cache) ((t) == UNIFIED_CACHE) ? CACHE_TYPE_UNIFIED : \ 0 /* Invalid value */) +static uint32_t max_processor_ids_for_cache (X86CPUTopoInfo *topo_info, +enum CPUTopoLevel share_level)I prefer the name to
max_lp_ids_share_the_cache()
Yes, lp is more accurate.
Quoted 17 lines
Quoted 17 lines
target/i386/cpu.c +16
-0+{ +uint32_t num_ids =0 ; + +switch (share_level) { +case CPU_TOPO_LEVEL_CORE : + num_ids =1 <<apicid_core_offset (topo_info); +break ; +case CPU_TOPO_LEVEL_DIE : + num_ids =1 <<apicid_die_offset (topo_info); +break ; +case CPU_TOPO_LEVEL_PACKAGE : + num_ids =1 <<apicid_pkg_offset (topo_info); +break ; +default : +/* +* Currently there is no use case for SMT and MODULE, so use
Quoted 8 lines
target/i386/cpu.c static uint8_t cpuid2_cache_descriptor(CPUCacheInfo *cache) · line 261
Quoted 6 lines
target/i386/cpu.c +6
-0+* assert directly to facilitate debugging. +*/ +g_assert_not_reached (); + } + +return num_ids -1 ;suggest to just return
num_ids, and let the caller to do the -1 work.
Emm, SDM calls the whole "num_ids - 1" (CPUID.0x4.EAX[bits 14-25]) as "maximum number of addressable IDs for logical processors sharing this cache"...
So if this helper just names "num_ids" as max_lp_ids_share_the_cache, I'm not sure there would be ambiguity here?
Quoted 30 lines
target/i386/cpu.c static uint8_t cpuid2_cache_descriptor(CPUCacheInfo *cache) · line 268
Quoted 8 lines
target/i386/cpu.c +7
-0+} + +static uint32_t max_core_ids_in_package (X86CPUTopoInfo *topo_info) +{ +uint32_t num_cores =1 << (apicid_pkg_offset (topo_info) - +apicid_core_offset (topo_info)); +return num_cores -1 ;ditto.
Quoted 8 lines
target/i386/cpu.c +2
-1+}/* Encode cache info for CPUID[4] */ static void encode_cache_cpuid4 (CPUCacheInfo *cache, -int num_apic_ids,int num_cores, +X86CPUTopoInfo *topo_info,uint32_t *eax,uint32_t *ebx,uint32_t *ecx,uint32_t *edx) {target/i386/cpu.c static uint8_t cpuid2_cache_descriptor(CPUCacheInfo *cache) · line 284
Quoted 10 lines
target/i386/cpu.c +2
-3assert (cache->size == cache->line_size * cache->associativity * cache->partitions * cache->sets); -assert (num_apic_ids >0 ); *eax =CACHE_TYPE (cache->type) |CACHE_LEVEL (cache->level) | (cache->self_init ?CACHE_SELF_INIT_LEVEL :0 ) | - ((num_cores -1 ) <<26 ) | - ((num_apic_ids -1 ) <<14 ); + (max_core_ids_in_package (topo_info) <<26 ) | + (max_processor_ids_for_cache (topo_info, cache->share_level) <<14 );
Quoted 1 line
by the way, we can change the order of the two line. :)
Yes!
Thanks,
Zhao
Quoted 67 lines
Quoted 66 lines
target/i386/cpu.c +8
-23assert (cache->line_size >0 );assert (cache->partitions >0 );@@ -6263,56 +6294,41 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,int host_vcpus_per_cache =1 + ((*eax &0x3FFC000 ) >>14 );if (cores_per_pkg >1 ) { -int addressable_cores_offset = -apicid_pkg_offset (&topo_info) - -apicid_core_offset (&topo_info); - *eax &= ~0xFC000000 ; - *eax |= (1 << (addressable_cores_offset -1 )) <<26 ; + *eax |=max_core_ids_in_package (&topo_info) <<26 ; }if (host_vcpus_per_cache > cpus_per_pkg) { -int pkg_offset =apicid_pkg_offset (&topo_info); - *eax &= ~0x3FFC000 ; - *eax |= (1 << (pkg_offset -1 )) <<14 ; + *eax |= +max_processor_ids_for_cache (&topo_info, +CPU_TOPO_LEVEL_PACKAGE ) <<14 ; } } }else if (cpu->vendor_cpuid_only &&IS_AMD_CPU (env)) { *eax = *ebx = *ecx = *edx =0 ; }else { *eax =0 ; -int addressable_cores_offset =apicid_pkg_offset (&topo_info) - -apicid_core_offset (&topo_info); -int core_offset, die_offset;switch (count) {case 0 :/* L1 dcache info */ - core_offset =apicid_core_offset (&topo_info);encode_cache_cpuid4 (env->cache_info_cpuid4.l1d_cache, - (1 << core_offset), - (1 << addressable_cores_offset), + &topo_info, eax, ebx, ecx, edx);break ;case 1 :/* L1 icache info */ - core_offset =apicid_core_offset (&topo_info);encode_cache_cpuid4 (env->cache_info_cpuid4.l1i_cache, - (1 << core_offset), - (1 << addressable_cores_offset), + &topo_info, eax, ebx, ecx, edx);break ;case 2 :/* L2 cache info */ - core_offset =apicid_core_offset (&topo_info);encode_cache_cpuid4 (env->cache_info_cpuid4.l2_cache, - (1 << core_offset), - (1 << addressable_cores_offset), + &topo_info, eax, ebx, ecx, edx);break ;case 3 :/* L3 cache info */ - die_offset =apicid_die_offset (&topo_info);if (cpu->enable_l3_cache) {encode_cache_cpuid4 (env->cache_info_cpuid4.l3_cache, - (1 << die_offset), - (1 << addressable_cores_offset), + &topo_info, eax, ebx, ecx, edx);break ; }
Re: [PATCH v7 14/16] i386: Use CPUCacheInfo.share_level to encode CPUID[4]
Quoted 23 lines · Zhao Liu
Quoted 15 lines
Quoted 15 lines
target/i386/cpu.c +15
-0+{ +uint32_t num_ids =0 ; + +switch (share_level) { +case CPU_TOPO_LEVEL_CORE : + num_ids =1 <<apicid_core_offset (topo_info); +break ; +case CPU_TOPO_LEVEL_DIE : + num_ids =1 <<apicid_die_offset (topo_info); +break ; +case CPU_TOPO_LEVEL_PACKAGE : + num_ids =1 <<apicid_pkg_offset (topo_info); +break ; +default : +/* Quoted 8 lines
target/i386/cpu.c static uint8_t cpuid2_cache_descriptor(CPUCacheInfo *cache) · line 261
Quoted 7 lines
target/i386/cpu.c +7
-0+* Currently there is no use case for SMT and MODULE, so use +* assert directly to facilitate debugging. +*/ +g_assert_not_reached (); + } + +return num_ids -1 ;suggest to just return
num_ids, and let the caller to do the -1 work.
Quoted 6 lines
Emm, SDM calls the whole "
num_ids- 1" (CPUID.0x4.EAX[bits 14-25]) as "maximum number of addressable IDs for logical processors sharing this cache"...So if this helper just names "
num_ids" asmax_lp_ids_share_the_cache, I'm not sure there would be ambiguity here?
I don't think it will.
if this function is going to used anywhere else, people will need to keep in mind to do +1 stuff to get the actual number.
leaving the -1 trick to where CPUID value gets encoded. let's make this function generic.
Re: [PATCH v7 14/16] i386: Use CPUCacheInfo.share_level to encode CPUID[4]
Hi Xiaoyao,
Quoted 36 lines · Xiaoyao Li
Date: Mon, 15 Jan 2024 12:25:19 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 14/16] i386: Use CPUCacheInfo.share_level to encode
CPUID[4]Quoted 23 lines · Zhao Liu
Quoted 15 lines
Quoted 15 lines
target/i386/cpu.c +15
-0+{ +uint32_t num_ids =0 ; + +switch (share_level) { +case CPU_TOPO_LEVEL_CORE : + num_ids =1 <<apicid_core_offset (topo_info); +break ; +case CPU_TOPO_LEVEL_DIE : + num_ids =1 <<apicid_die_offset (topo_info); +break ; +case CPU_TOPO_LEVEL_PACKAGE : + num_ids =1 <<apicid_pkg_offset (topo_info); +break ; +default : +/* Quoted 8 lines
target/i386/cpu.c static uint8_t cpuid2_cache_descriptor(CPUCacheInfo *cache) · line 261
Quoted 7 lines
target/i386/cpu.c +7
-0+* Currently there is no use case for SMT and MODULE, so use +* assert directly to facilitate debugging. +*/ +g_assert_not_reached (); + } + +return num_ids -1 ;suggest to just return
num_ids, and let the caller to do the -1 work.Quoted 6 lines
Emm, SDM calls the whole "
num_ids- 1" (CPUID.0x4.EAX[bits 14-25]) as "maximum number of addressable IDs for logical processors sharing this cache"...So if this helper just names "
num_ids" asmax_lp_ids_share_the_cache, I'm not sure there would be ambiguity here?
Quoted 7 lines
I don't think it will.
if this function is going to used anywhere else, people will need to keep in mind to do +1 stuff to get the actual number.
leaving the -1 trick to where CPUID value gets encoded. let's make this function generic.
This helper is the complete pattern to get addressable IDs, this is to say, the "- 1" is also the part of this calculation.
Its own meaning is self-consistent and generic enough to meet the common definitions of AMD and Intel.
Thanks,
Zhao
Re: [PATCH v7 14/16] i386: Use CPUCacheInfo.share_level to encode CPUID[4]
Quoted 44 lines · Zhao Liu
Hi Xiaoyao,
Quoted 41 lines · Xiaoyao Li
Date: Mon, 15 Jan 2024 12:25:19 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 14/16] i386: Use CPUCacheInfo.share_level to encode
CPUID[4]Quoted 23 lines · Zhao Liu
Quoted 15 lines
Quoted 15 lines
target/i386/cpu.c +15
-0+{ +uint32_t num_ids =0 ; + +switch (share_level) { +case CPU_TOPO_LEVEL_CORE : + num_ids =1 <<apicid_core_offset (topo_info); +break ; +case CPU_TOPO_LEVEL_DIE : + num_ids =1 <<apicid_die_offset (topo_info); +break ; +case CPU_TOPO_LEVEL_PACKAGE : + num_ids =1 <<apicid_pkg_offset (topo_info); +break ; +default : +/* Quoted 8 lines
target/i386/cpu.c static uint8_t cpuid2_cache_descriptor(CPUCacheInfo *cache) · line 261
Quoted 7 lines
target/i386/cpu.c +7
-0+* Currently there is no use case for SMT and MODULE, so use +* assert directly to facilitate debugging. +*/ +g_assert_not_reached (); + } + +return num_ids -1 ;suggest to just return
num_ids, and let the caller to do the -1 work.Quoted 6 lines
Emm, SDM calls the whole "
num_ids- 1" (CPUID.0x4.EAX[bits 14-25]) as "maximum number of addressable IDs for logical processors sharing this cache"...So if this helper just names "
num_ids" asmax_lp_ids_share_the_cache, I'm not sure there would be ambiguity here?I don't think it will.
if this function is going to used anywhere else, people will need to keep in mind to do +1 stuff to get the actual number.
Quoted 8 lines
Quoted 2 lines
leaving the -1 trick to where CPUID value gets encoded. let's make this function generic.
This helper is the complete pattern to get addressable IDs, this is to say, the "- 1" is also the part of this calculation.
Its own meaning is self-consistent and generic enough to meet the common definitions of AMD and Intel.
OK. I stop bikeshedding on it.
Quoted 3 lines
Thanks,
Zhao
Re: [PATCH v7 14/16] i386: Use CPUCacheInfo.share_level to encode CPUID[4]
Hi Xiaoyao,
Quoted 53 lines · Xiaoyao Li
Date: Mon, 15 Jan 2024 15:00:25 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 14/16] i386: Use CPUCacheInfo.share_level to encode
CPUID[4]Quoted 47 lines · Zhao Liu
Hi Xiaoyao,
Quoted 43 lines · Xiaoyao Li
Date: Mon, 15 Jan 2024 12:25:19 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 14/16] i386: Use CPUCacheInfo.share_level to encode
CPUID[4]Quoted 23 lines · Zhao Liu
Quoted 15 lines
Quoted 15 lines
target/i386/cpu.c +15
-0+{ +uint32_t num_ids =0 ; + +switch (share_level) { +case CPU_TOPO_LEVEL_CORE : + num_ids =1 <<apicid_core_offset (topo_info); +break ; +case CPU_TOPO_LEVEL_DIE : + num_ids =1 <<apicid_die_offset (topo_info); +break ; +case CPU_TOPO_LEVEL_PACKAGE : + num_ids =1 <<apicid_pkg_offset (topo_info); +break ; +default : +/* Quoted 8 lines
target/i386/cpu.c static uint8_t cpuid2_cache_descriptor(CPUCacheInfo *cache) · line 261
Quoted 7 lines
target/i386/cpu.c +7
-0+* Currently there is no use case for SMT and MODULE, so use +* assert directly to facilitate debugging. +*/ +g_assert_not_reached (); + } + +return num_ids -1 ;suggest to just return
num_ids, and let the caller to do the -1 work.Quoted 6 lines
Emm, SDM calls the whole "
num_ids- 1" (CPUID.0x4.EAX[bits 14-25]) as "maximum number of addressable IDs for logical processors sharing this cache"...So if this helper just names "
num_ids" asmax_lp_ids_share_the_cache, I'm not sure there would be ambiguity here?I don't think it will.
if this function is going to used anywhere else, people will need to keep in mind to do +1 stuff to get the actual number.
leaving the -1 trick to where CPUID value gets encoded. let's make this function generic.
Quoted 8 lines
Quoted 5 lines
This helper is the complete pattern to get addressable IDs, this is to say, the "- 1" is also the part of this calculation.
Its own meaning is self-consistent and generic enough to meet the common definitions of AMD and Intel.
OK. I stop bikeshedding on it.
Thanks for your review ;-).
Regards,
Zhao
PATCH v7 15/16 i386: Use offsets get NumSharingCache for CPUID[0x8000001D].EAX[bits 25:14]
In reply to Zhao LiuFrom: Zhao Liu <[email protected]>
The commit 8f4202fb1080 ("i386: Populate AMD Processor Cache Information for cpuid 0x8000001D") adds the cache topology for AMD CPU by encoding the number of sharing threads directly.
From AMD's APM, NumSharingCache (CPUID[0x8000001D].EAX[bits 25:14]) means [1]:
The number of logical processors sharing this cache is the value of this field incremented by 1. To determine which logical processors are sharing a cache, determine a Share Id for each processor as follows:
ShareId = LocalApicId >> log2(NumSharingCache+1)
Logical processors with the same ShareId then share a cache. If NumSharingCache+1 is not a power of two, round it up to the next power of two.
From the description above, the calculation of this field should be same as CPUID[4].EAX[bits 25:14] for Intel CPUs. So also use the offsets of APIC ID to calculate this field.
[1]: APM, vol.3, appendix.E.4.15 Function 8000_001Dh--Cache Topology
Information
---
Changes since v3:
* Rewrite the subject. (Babu)
* Delete the original "comment/help" expression, as this behavior is confirmed for AMD CPUs. (Babu)
* Rename "num_apic_ids" (v3) to "num_sharing_cache" to match spec definition. (Babu)
Changes since v1:
* Rename "l3_threads" to "num_apic_ids" in encode_cache_cpuid8000001d(). (Yanan)
* Add the description of the original commit and add Cc.
--- target/i386/cpu.c | 10 ++++------ 1 file changed, 4 insertions(+), 6 deletions(-)
target/i386/cpu.c +4 -6
@@ -483,7 +483,7 @@ static void encode_cache_cpuid8000001d(CPUCacheInfo *cache, uint32_t *eax, uint32_t *ebx, uint32_t *ecx, uint32_t *edx) { - uint32_t l3_threads; + uint32_t num_sharing_cache;assert (cache->size == cache->line_size * cache->associativity * cache->partitions * cache->sets);
@@ -492,13 +492,11 @@ static void encode_cache_cpuid8000001d(CPUCacheInfo *cache,/* L3 is shared among multiple cores */ if (cache->level ==3 ) { - l3_threads = topo_info->modules_per_die * - topo_info->cores_per_module * - topo_info->threads_per_core; - *eax |= (l3_threads -1 ) <<14 ; + num_sharing_cache =1 <<apicid_die_offset (topo_info); }else { - *eax |= ((topo_info->threads_per_core -1 ) <<14 ); + num_sharing_cache =1 <<apicid_core_offset (topo_info); } + *eax |= (num_sharing_cache -1 ) <<14 ;assert (cache->line_size >0 );assert (cache->partitions >0 );
Signature · 1 line
2.34.1
Re: [PATCH v7 15/16] i386: Use offsets get NumSharingCache for CPUID[0x8000001D].EAX[bits 25:14]
Quoted 19 lines · Zhao Liu
From: Zhao Liu <[email protected]>
The commit
8f4202fb1080("i386: Populate AMD Processor Cache Information for cpuid0x8000001D") adds the cache topology for AMD CPU by encoding the number of sharing threads directly.From AMD's APM, NumSharingCache (CPUID[0x8000001D].EAX[bits 25:14])
means [1]:The number of logical processors sharing this cache is the value of this field incremented by 1. To determine which logical processors are sharing a cache, determine a Share Id for each processor as follows:
ShareId = LocalApicId >> log2(NumSharingCache+1)
Logical processors with the same ShareId then share a cache. If NumSharingCache+1 is not a power of two, round it up to the next power of two.
Quoted 6 lines
From the description above, the calculation of this field should be same
as CPUID[4].EAX[bits 25:14] for Intel CPUs. So also use the offsets of APIC ID to calculate this field.[1]: APM, vol.3, appendix.E.4.15 Function 8000_001Dh--Cache Topology
Information
this patch can be dropped because we have next patch.
Quoted 51 lines
Signed-off-by Zhao Liu <[email protected]>Reviewed-by Babu Moger <[email protected]>Tested-by Babu Moger <[email protected]>Tested-by Yongwei Ma <[email protected]>---
Changes since v3:
* Rewrite the subject. (Babu)
* Delete the original "comment/help" expression, as this behavior is confirmed for AMD CPUs. (Babu)
* Rename "num_apic_ids" (v3) to "num_sharing_cache" to match spec definition. (Babu)Changes since v1:
* Rename "l3_threads" to "num_apic_ids" inencode_cache_cpuid8000001d(). (Yanan) * Add the description of the original commit and add Cc. --- target/i386/cpu.c | 10 ++++------ 1 file changed, 4 insertions(+), 6 deletions(-)target/i386/cpu.c +4
-6@@ -483,7 +483,7 @@ static void encode_cache_cpuid8000001d(CPUCacheInfo *cache, uint32_t *eax, uint32_t *ebx, uint32_t *ecx, uint32_t *edx) { - uint32_t l3_threads; + uint32_t num_sharing_cache;assert (cache->size == cache->line_size * cache->associativity * cache->partitions * cache->sets);@@ -492,13 +492,11 @@ static void encode_cache_cpuid8000001d(CPUCacheInfo *cache,/* L3 is shared among multiple cores */ if (cache->level ==3 ) { - l3_threads = topo_info->modules_per_die * - topo_info->cores_per_module * - topo_info->threads_per_core; - *eax |= (l3_threads -1 ) <<14 ; + num_sharing_cache =1 <<apicid_die_offset (topo_info); }else { - *eax |= ((topo_info->threads_per_core -1 ) <<14 ); + num_sharing_cache =1 <<apicid_core_offset (topo_info); } + *eax |= (num_sharing_cache -1 ) <<14 ;assert (cache->line_size >0 );assert (cache->partitions >0 );
Re: [PATCH v7 15/16] i386: Use offsets get NumSharingCache for CPUID[0x8000001D].EAX[bits 25:14]
Hi Xiaoyao,
Quoted 25 lines · Xiaoyao Li
Date: Sun, 14 Jan 2024 22:42:41 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 15/16] i386: Use offsets get NumSharingCache for
CPUID[0x8000001D].EAX[bits 25:14]Quoted 19 lines · Zhao Liu
From: Zhao Liu <[email protected]>
The commit
8f4202fb1080("i386: Populate AMD Processor Cache Information for cpuid0x8000001D") adds the cache topology for AMD CPU by encoding the number of sharing threads directly.From AMD's APM, NumSharingCache (CPUID[0x8000001D].EAX[bits 25:14])
means [1]:The number of logical processors sharing this cache is the value of this field incremented by 1. To determine which logical processors are sharing a cache, determine a Share Id for each processor as follows:
ShareId = LocalApicId >> log2(NumSharingCache+1)
Logical processors with the same ShareId then share a cache. If NumSharingCache+1 is not a power of two, round it up to the next power of two.
Quoted 8 lines
Quoted 6 lines
From the description above, the calculation of this field should be same
as CPUID[4].EAX[bits 25:14] for Intel CPUs. So also use the offsets of APIC ID to calculate this field.[1]: APM, vol.3, appendix.E.4.15 Function 8000_001Dh--Cache Topology
Informationthis patch can be dropped because we have next patch.
This patch is mainly used to explicitly emphasize the change in encoding way and compliance with AMD spec... I didn't tested on AMD machine, so the more granular patch would make it easier for the community to review and test.
Thanks,
Zhao
Quoted 50 lines
Quoted 49 lines
Signed-off-by Zhao Liu <[email protected]>Reviewed-by Babu Moger <[email protected]>Tested-by Babu Moger <[email protected]>Tested-by Yongwei Ma <[email protected]>---
Changes since v3:
* Rewrite the subject. (Babu)
* Delete the original "comment/help" expression, as this behavior is confirmed for AMD CPUs. (Babu)
* Rename "num_apic_ids" (v3) to "num_sharing_cache" to match spec definition. (Babu)Changes since v1:
* Rename "l3_threads" to "num_apic_ids" inencode_cache_cpuid8000001d(). (Yanan) * Add the description of the original commit and add Cc. --- target/i386/cpu.c | 10 ++++------ 1 file changed, 4 insertions(+), 6 deletions(-)target/i386/cpu.c +4
-6@@ -483,7 +483,7 @@ static void encode_cache_cpuid8000001d(CPUCacheInfo *cache, uint32_t *eax, uint32_t *ebx, uint32_t *ecx, uint32_t *edx) { - uint32_t l3_threads; + uint32_t num_sharing_cache;assert (cache->size == cache->line_size * cache->associativity * cache->partitions * cache->sets);@@ -492,13 +492,11 @@ static void encode_cache_cpuid8000001d(CPUCacheInfo *cache,/* L3 is shared among multiple cores */ if (cache->level ==3 ) { - l3_threads = topo_info->modules_per_die * - topo_info->cores_per_module * - topo_info->threads_per_core; - *eax |= (l3_threads -1 ) <<14 ; + num_sharing_cache =1 <<apicid_die_offset (topo_info); }else { - *eax |= ((topo_info->threads_per_core -1 ) <<14 ); + num_sharing_cache =1 <<apicid_core_offset (topo_info); } + *eax |= (num_sharing_cache -1 ) <<14 ;assert (cache->line_size >0 );assert (cache->partitions >0 );
Re: [PATCH v7 15/16] i386: Use offsets get NumSharingCache for CPUID[0x8000001D].EAX[bits 25:14]
Quoted 35 lines · Zhao Liu
Hi Xiaoyao,
Quoted 32 lines · Xiaoyao Li
Date: Sun, 14 Jan 2024 22:42:41 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 15/16] i386: Use offsets get NumSharingCache for
CPUID[0x8000001D].EAX[bits 25:14]Quoted 25 lines · Zhao Liu
From: Zhao Liu <[email protected]>
The commit
8f4202fb1080("i386: Populate AMD Processor Cache Information for cpuid0x8000001D") adds the cache topology for AMD CPU by encoding the number of sharing threads directly.From AMD's APM, NumSharingCache (CPUID[0x8000001D].EAX[bits 25:14])
means [1]:The number of logical processors sharing this cache is the value of this field incremented by 1. To determine which logical processors are sharing a cache, determine a Share Id for each processor as follows:
ShareId = LocalApicId >> log2(NumSharingCache+1)
Logical processors with the same ShareId then share a cache. If NumSharingCache+1 is not a power of two, round it up to the next power of two.
From the description above, the calculation of this field should be same
as CPUID[4].EAX[bits 25:14] for Intel CPUs. So also use the offsets of APIC ID to calculate this field.[1]: APM, vol.3, appendix.E.4.15 Function 8000_001Dh--Cache Topology
Information
Quoted 6 lines
Quoted 1 line
this patch can be dropped because we have next patch.
This patch is mainly used to explicitly emphasize the change in encoding way and compliance with AMD spec... I didn't tested on AMD machine, so the more granular patch would make it easier for the community to review and test.
then please move this patch ahead, e.g., after patch 2.
Quoted 53 lines
Thanks,
ZhaoQuoted 50 lines
Quoted 49 lines
Signed-off-by Zhao Liu <[email protected]>Reviewed-by Babu Moger <[email protected]>Tested-by Babu Moger <[email protected]>Tested-by Yongwei Ma <[email protected]>---
Changes since v3:
* Rewrite the subject. (Babu)
* Delete the original "comment/help" expression, as this behavior is confirmed for AMD CPUs. (Babu)
* Rename "num_apic_ids" (v3) to "num_sharing_cache" to match spec definition. (Babu)Changes since v1:
* Rename "l3_threads" to "num_apic_ids" inencode_cache_cpuid8000001d(). (Yanan) * Add the description of the original commit and add Cc. --- target/i386/cpu.c | 10 ++++------ 1 file changed, 4 insertions(+), 6 deletions(-)target/i386/cpu.c +4
-6@@ -483,7 +483,7 @@ static void encode_cache_cpuid8000001d(CPUCacheInfo *cache, uint32_t *eax, uint32_t *ebx, uint32_t *ecx, uint32_t *edx) { - uint32_t l3_threads; + uint32_t num_sharing_cache;assert (cache->size == cache->line_size * cache->associativity * cache->partitions * cache->sets);@@ -492,13 +492,11 @@ static void encode_cache_cpuid8000001d(CPUCacheInfo *cache,/* L3 is shared among multiple cores */ if (cache->level ==3 ) { - l3_threads = topo_info->modules_per_die * - topo_info->cores_per_module * - topo_info->threads_per_core; - *eax |= (l3_threads -1 ) <<14 ; + num_sharing_cache =1 <<apicid_die_offset (topo_info); }else { - *eax |= ((topo_info->threads_per_core -1 ) <<14 ); + num_sharing_cache =1 <<apicid_core_offset (topo_info); } + *eax |= (num_sharing_cache -1 ) <<14 ;assert (cache->line_size >0 );assert (cache->partitions >0 );
Re: [PATCH v7 15/16] i386: Use offsets get NumSharingCache for CPUID[0x8000001D].EAX[bits 25:14]
Hi Xiaoyao,
Quoted 41 lines · Xiaoyao Li
Date: Mon, 15 Jan 2024 12:27:43 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 15/16] i386: Use offsets get NumSharingCache for
CPUID[0x8000001D].EAX[bits 25:14]Quoted 35 lines · Zhao Liu
Hi Xiaoyao,
Quoted 32 lines · Xiaoyao Li
Date: Sun, 14 Jan 2024 22:42:41 +0800
From: Xiaoyao Li <[email protected]>
Subject: Re: [PATCH v7 15/16] i386: Use offsets get NumSharingCache for
CPUID[0x8000001D].EAX[bits 25:14]Quoted 25 lines · Zhao Liu
From: Zhao Liu <[email protected]>
The commit
8f4202fb1080("i386: Populate AMD Processor Cache Information for cpuid0x8000001D") adds the cache topology for AMD CPU by encoding the number of sharing threads directly.From AMD's APM, NumSharingCache (CPUID[0x8000001D].EAX[bits 25:14])
means [1]:The number of logical processors sharing this cache is the value of this field incremented by 1. To determine which logical processors are sharing a cache, determine a Share Id for each processor as follows:
ShareId = LocalApicId >> log2(NumSharingCache+1)
Logical processors with the same ShareId then share a cache. If NumSharingCache+1 is not a power of two, round it up to the next power of two.
From the description above, the calculation of this field should be same
as CPUID[4].EAX[bits 25:14] for Intel CPUs. So also use the offsets of APIC ID to calculate this field.[1]: APM, vol.3, appendix.E.4.15 Function 8000_001Dh--Cache Topology
Information
Quoted 9 lines
Quoted 6 lines
Quoted 1 line
this patch can be dropped because we have next patch.
This patch is mainly used to explicitly emphasize the change in encoding way and compliance with AMD spec... I didn't tested on AMD machine, so the more granular patch would make it easier for the community to review and test.
then please move this patch ahead, e.g., after patch 2.
OK. Thanks!
-Zhao
PATCH v7 16/16 i386: Use CPUCacheInfo.share_level to encode CPUID[0x8000001D].EAX[bits 25:14]
In reply to Zhao LiuFrom: Zhao Liu <[email protected]>
CPUID[0x8000001D].EAX[bits 25:14] NumSharingCache: number of logical processors sharing cache.
The number of logical processors sharing this cache is NumSharingCache + 1.
After cache models have topology information, we can use CPUCacheInfo.share_level to decide which topology level to be encoded into CPUID[0x8000001D].EAX[bits 25:14].
---
Changes since v3:
* Explain what "CPUID[0x8000001D].EAX[bits 25:14]" means in the commit message. (Babu)
Changes since v1:
* Use cache->share_level as the parameter in max_processor_ids_for_cache().
--- target/i386/cpu.c | 10 +--------- 1 file changed, 1 insertion(+), 9 deletions(-)
target/i386/cpu.c +1 -9
@@ -483,20 +483,12 @@ static void encode_cache_cpuid8000001d(CPUCacheInfo *cache, uint32_t *eax, uint32_t *ebx, uint32_t *ecx, uint32_t *edx) { - uint32_t num_sharing_cache; assert(cache->size == cache->line_size * cache->associativity * cache->partitions * cache->sets); *eax =CACHE_TYPE (cache->type) |CACHE_LEVEL (cache->level) | (cache->self_init ?CACHE_SELF_INIT_LEVEL :0 ); - -/* L3 is shared among multiple cores */ -if (cache->level ==3 ) { - num_sharing_cache =1 <<apicid_die_offset (topo_info); - }else { - num_sharing_cache =1 <<apicid_core_offset (topo_info); - } - *eax |= (num_sharing_cache -1 ) <<14 ; + *eax |=max_processor_ids_for_cache (topo_info, cache->share_level) <<14 ;assert (cache->line_size >0 );assert (cache->partitions >0 );
Signature · 1 line
2.34.1
Hi Zhao,
Ran few basic tests on AMD systems. Changes look good.
Thanks
Babu
Quoted 245 lines · Zhao Liu
From: Zhao Liu <[email protected]>
Hi list,
This is the our v7 patch series, rebased on the master branch at the commit
d328fef93ae7("Merge tag 'pull-20231230' of
https://gitlab.com/rth7680/qemu into staging").No more change since v6 [1] exclude the comment nit update.
Welcome your comments!
PS: Since v5, we have dropped "x-l2-cache-topo" option and now are working on porting the original x-l2-cache-topo option to smp [2]. Just like:
-smp cpus=4,sockets=2,cores=2,threads=1, \ l3-cache=socket,l2-cache=core,l1-i-cache=core,l1-d-cache=coreThe cache topology enhancement in this patch set is the preparation for supporting future user-configurable cache topology (via generic cli interface).
---
# IntroductionThis series adds the cluster support for x86 PC machine, which allows x86 can use smp.clusters to configure the module level CPU topology of x86.
This series also is the preparation to help x86 to define the more flexible cache topology, such as having multiple cores share the same L2 cache at cluster level. (That was what x-l2-cache-topo did, and we will explore a generic way.)
About why we don't share L2 cache at cluster and need a configuration way, pls see section: ## Why not share L2 cache in cluster directly.
# Background
The "clusters" parameter in "smp" is introduced by ARM [3], but x86 hasn't supported it.
At present, x86 defaults L2 cache is shared in one core, but this is not enough. There're some platforms that multiple cores share the same L2 cache, e.g., Alder Lake-P shares L2 cache for one module of Atom cores [4], that is, every four Atom cores shares one L2 cache. Therefore, we need the new CPU topology level (cluster/module).
Another reason is for hybrid architecture. cluster support not only provides another level of topology definition in x86, but would also provide required code change for future our hybrid topology support.
# Overview
## Introduction of module level for x86
"cluster" in smp is the CPU topology level which is between "core" and die.
For x86, the "cluster" in smp is corresponding to the module level [4], which is above the core level. So use the "module" other than "cluster" in x86 code.
And please note that x86 already has a cpu topology level also named "cluster" [5], this level is at the upper level of the package. Here, the cluster in x86 cpu topology is completely different from the "clusters" as the smp parameter. After the module level is introduced, the cluster as the smp parameter will actually refer to the module level of x86.
## Why not share L2 cache in cluster directly
Though "clusters" was introduced to help define L2 cache topology [3], using cluster to define x86's L2 cache topology will cause the compatibility problem:
Currently, x86 defaults that the L2 cache is shared in one core, which actually implies a default setting "cores per L2 cache is 1" and therefore implicitly defaults to having as many L2 caches as cores.
For example (i386 PC machine):
-smp 16,sockets=2,dies=2,cores=2,threads=2,maxcpus=16 (*)Considering the topology of the L2 cache, this (*) implicitly means "1 core per L2 cache" and "2 L2 caches per die".
If we use cluster to configure L2 cache topology with the new default setting "clusters per L2 cache is 1", the above semantics will change to "2 cores per cluster" and "1 cluster per L2 cache", that is, "2 cores per L2 cache".
So the same command (*) will cause changes in the L2 cache topology, further affecting the performance of the virtual machine.
Therefore, x86 should only treat cluster as a cpu topology level and avoid using it to change L2 cache by default for compatibility.
## module level in CPUID
Linux kernel (from v6.4, with commit
edc0a2b595765("x86/topology: Fix erroneoussmp_num_siblingson Intel Hybrid platforms") is able to handle platforms with Module level enumerated via CPUID.1F.Expose the module level in CPUID[0x1F] (for Intel CPUs) if the machine has more than 1 modules since v3.
## New cache topology info in CPUCacheInfo
(This is in preparation for users being able to configure cache topology from the cli later on.)
Currently, by default, the cache topology is encoded as:
1. i/d cache is shared in one core.
2. L2 cache is shared in one core.
3. L3 cache is shared in one die.This default general setting has caused a misunderstanding, that is, the cache topology is completely equated with a specific cpu topology, such as the connection between L2 cache and core level, and the connection between L3 cache and die level.
In fact, the settings of these topologies depend on the specific platform and are not static. For example, on Alder Lake-P, every four Atom cores share the same L2 cache [3].
Thus, in this patch set, we explicitly define the corresponding cache topology for different cpu models and this has two benefits:
1. Easy to expand to new CPU models in the future, which has different cache topology.
2. It can easily support custom cache topology by some command.# Patch description
patch 1 Fixes about x86 topology and Intel l1 cache topology.
patch 2-3 Cleanups about topology related CPUID encoding and QEMU
topology variables.patch 4-5 Refactor CPUID[0x1F] encoding to prepare to introduce module
level.patch 6-12 Add the module as the new CPU topology level in x86, and it
is corresponding to the cluster level in generic code.patch 13,14,16 Add cache topology information in cache models.
patch 15 Update AMD CPUs' cache topology encoding.
[1]: https://lore.kernel.org/qemu-devel/[email protected]/
[2]: https://lists.gnu.org/archive/html/qemu-devel/2023-10/msg01954.html
[3]: https://patchew.org/QEMU/[email protected]/
[4]: https://www.intel.com/content/www/us/en/products/platforms/details/alder-lake-p.html
[5]: SDM, vol.3, ch.9, 9.9.1 Hierarchical Mapping of Shared Resources.Best Regards,
Zhao
---
Changelog:Changes since v6:
* Update the comment when check cluster-id. Since there's no v8.2, the cluster-id support should at least start from v9.0.
* Rebase on commitd328fef93ae7("Merge tag 'pull-20231230' of
https://gitlab.com/rth7680/qemu into staging").Changes since v5:
* The first four patches of v5 [1] have been merged, v6 contains the remaining patches.
* Reabse on the latest master.
* Update the comment when check cluster-id. Since current QEMU is v8.2, the cluster-id support should at least start from v8.3.Changes since v4:
* Drop the "x-l2-cache-topo" option. (Michael)
* Add A/R/T tags.Changes since v3 (main changes):
* Expose module level in CPUID[0x1F].
* Fix compile warnings. (Babu)
* Fixes cache topology uninitialization bugs for some AMD CPUs. (Babu)Changes since v2:
* Add "Tested-by", "Reviewed-by" and "ACKed-by" tags.
* Use newly added wrapped helper to get cores per socket inqemu_init_vcpu().Changes since v1:
* Reordered patches. (Yanan)
* Deprecated the patch to fix comment ofmachine_parse_smp_config(). (Yanan)
* Renametest-x86-cpuid.ctotest-x86-topo.c. (Yanan)
* Split the intel's l1 cache topology fix into a new separate patch. (Yanan)
* Combinedmodule_idand APIC ID for module level support into one patch. (Yanan)
* Makecache_into_passthroughcase of cpuid0x04leaf in
*cpu_x86_cpuid()usemax_processor_ids_for_cache()andmax_core_ids_in_package()to encode CPUID[4]. (Yanan)
* Add the prefix "CPU_TOPO_LEVEL_*" for CPU topology level names. (Yanan)---
Zhao Liu (10):
i386/cpu: Fix i/d-cache topology to core level for Intel CPU i386/cpu: Use APIC ID offset to encode cache topo in CPUID[4] i386/cpu: Consolidate the use oftopo_infoincpu_x86_cpuid()
i386: Split topology types of CPUID[0x1F] from the definitions ofCPUID[0xB] i386: Decouple CPUID[0x1F] subleaf with specific topology level i386: Expose module level in CPUID[0x1F] i386: Add cache topology info in CPUCacheInfo i386: Use CPUCacheInfo.share_level to encode CPUID[4] i386: Use offsets get NumSharingCache for CPUID[0x8000001D].EAX[bits 25:14] i386: Use CPUCacheInfo.share_level to encode CPUID[0x8000001D].EAX[bits 25:14]Zhuocheng Ding (6):
i386: Introduce module-level cpu topology to CPUX86State
i386: Supportmodules_per_diein X86CPUTopoInfo
i386: Supportmodule_idin X86CPUTopoIDs i386/cpu: Introduce cluster-id to X86CPU
tests: Add test case of APIC ID for module level parsing hw/i386/pc: Support smp.clusters for x86 PC machinehw/i386/pc.c | 1 + hw/i386/x86.c | 49 ++++++- include/hw/i386/topology.h | 35 ++++- qemu-options.hx | 10 +- target/i386/cpu.c | 289 +++++++++++++++++++++++++++++-------- target/i386/cpu.h | 43 +++++- target/i386/kvm/kvm.c | 2 +- tests/unit/test-x86-topo.c | 56 ++++--- 8 files changed, 379 insertions(+), 106 deletions(-)
Signature · 2 lines
Thanks Babu Moger
Hi Babu,
Quoted 6 lines · Babu
Date: Mon, 8 Jan 2024 11:46:50 -0600
From: "Moger, Babu" <[email protected]>
Subject: Re: [PATCH v7 00/16] Support smp.clusters for x86 in QEMUHi Zhao,
Quoted 8 lines
Ran few basic tests on AMD systems. Changes look good.
Thanks
BabuTested-by Babu Moger <[email protected]>
Thanks much for your test!
Regards,
Zhao