x86/cpufeatures: Re-arrange a few features and enumerate AVX512 BFLOAT16 intructions
Subsystems (3)
cover x86/cpufeatures: Re-arrange a few features and enumerate AVX512 BFLOAT16 intructions
To enumerate AVX512 BFLOAT16 feature CPUID.7.1:EAX[5] and other future features in CPUID.7.1:EAX, Boris suggests to create a new pure feature bits word.
Boris further suggests to re-define word 11 as scattered features word and move the four X86_FEATURE_CQM_* features in existing word 11 and word 12 into the new word 11. Then use word 12 to hold features in CPUID.7.1:EAX including AVX512 BFLOAT16 instructions.
Also remove x86_cache_max_rmid and x86_cache_occ_scale in cpuinfo_x86 to save memory space because they are only read once by resctrl during boot time. Get the info directly from CPUID in resctrl initialization.
Fenghua Yu (3):
x86/resctrl: Get max rmid and occupancy scale directly from CPUID
instead of cpuinfo_x86
x86/cpufeatures: Combine word 11 and 12 into new scattered features
word 11
x86/cpufeatures: Enumerate new AVX512 BFLOAT16 instructions
arch/x86/include/asm/cpufeature.h | 4 +-- arch/x86/include/asm/cpufeatures.h | 18 +++++++---- arch/x86/include/asm/processor.h | 3 -- arch/x86/kernel/cpu/common.c | 45 ++------------------------ arch/x86/kernel/cpu/cpuid-deps.c | 4 +++ arch/x86/kernel/cpu/resctrl/internal.h | 2 +- arch/x86/kernel/cpu/resctrl/monitor.c | 28 ++++++++++++++-- arch/x86/kernel/cpu/scattered.c | 4 +++ arch/x86/kvm/cpuid.h | 2 -- 9 files changed, 51 insertions(+), 59 deletions(-)
Signature · 1 line
2.19.1
RFC PATCH 1/3 x86/resctrl: Get max rmid and occupancy scale directly from CPUID instead of cpuinfo_x86
Although x86_cache_max_rmid and x86_cache_occ_scale are read only once during resctrl initialization, they are always stored in cpuinfo_x86 on each CPU during run time even if resctrl is not configured.
To save cpuinfo_x86 space and make CPU and resctrl initialization simpler, remove the two fields from cpuinfo_x86 and get max rmid and occupancy scale directly from CPUID during resctrl initialization. And since each known platform that supports resctrl has same max rmid on all CPUs, no need to scan all CPUs to find minimum of max rmid values, i.e. getting max rmid from CPUID on the current CPU is fine.
--- arch/x86/include/asm/processor.h | 3 --- arch/x86/kernel/cpu/common.c | 28 -------------------------- arch/x86/kernel/cpu/resctrl/internal.h | 2 +- arch/x86/kernel/cpu/resctrl/monitor.c | 28 +++++++++++++++++++++++--- 4 files changed, 26 insertions(+), 35 deletions(-)
arch/x86/include/asm/processor.h +0 -3
@@ -99,9 +99,6 @@ struct cpuinfo_x86 {/* in KB - valid for CPUS which support this call: */ unsigned int x86_cache_size;int x86_cache_alignment;/* In bytes */ -/* Cache QoS architectural values: */ -int x86_cache_max_rmid;/* max index */ -int x86_cache_occ_scale;/* scale to bytes */ int x86_power;unsigned long loops_per_jiffy;/* cpuid returned max cores value: */
arch/x86/kernel/cpu/common.c +0 -28
@@ -840,22 +840,9 @@ void get_cpu_cap(struct cpuinfo_x86 *c) c->x86_capability[CPUID_F_0_EDX ] = edx;if (cpu_has (c,X86_FEATURE_CQM_LLC )) { -/* will be overridden if occupancy monitoring exists */ - c->x86_cache_max_rmid = ebx; -/* QoS sub-leaf, EAX=0Fh, ECX=1 */ cpuid_count (0x0000000F ,1 , &eax, &ebx, &ecx, &edx); c->x86_capability[CPUID_F_1_EDX ] = edx; - -if ((cpu_has (c,X86_FEATURE_CQM_OCCUP_LLC )) || - ((cpu_has (c,X86_FEATURE_CQM_MBM_TOTAL )) || - (cpu_has (c,X86_FEATURE_CQM_MBM_LOCAL )))) { - c->x86_cache_max_rmid = ecx; - c->x86_cache_occ_scale = ebx; - } - }else { - c->x86_cache_max_rmid =-1 ; - c->x86_cache_occ_scale =-1 ; } }
@@ -1269,20 +1256,6 @@ static void generic_identify(struct cpuinfo_x86 *c)#endif } -static void x86_init_cache_qos (struct cpuinfo_x86 *c) -{ -/* -* The heavy lifting of max_rmid and cache_occ_scale are handled -* in get_cpu_cap(). Here we just set the max_rmid for the boot_cpu -* in case CQM bits really aren't there in this CPU. -*/ -if (c != &boot_cpu_data) { - boot_cpu_data.x86_cache_max_rmid = -min (boot_cpu_data.x86_cache_max_rmid, - c->x86_cache_max_rmid); - } -} -/* * Validate that ACPI/mptables have the same information about the * effective APIC id and update the package map.
@@ -1391,7 +1364,6 @@ static void identify_cpu(struct cpuinfo_x86 *c)#endif x86_init_rdrand (c); -x86_init_cache_qos (c);setup_pku (c);/*
arch/x86/kernel/cpu/resctrl/internal.h +1 -1
@@ -579,7 +579,7 @@ int closids_supported(void);void closid_free (int closid);int alloc_rmid (void );void free_rmid (u32 rmid); -int rdt_get_mon_l3_config (struct rdt_resource *r); +int __init rdt_get_mon_l3_config (struct rdt_resource *r);void mon_event_count (void *info);int rdtgroup_mondata_show (struct seq_file *m,void *arg); void rmdir_mondata_subdir_allrdtgrp(struct rdt_resource *r,
arch/x86/kernel/cpu/resctrl/monitor.c +25 -3
@@ -617,13 +617,35 @@ static void l3_mon_evt_init(struct rdt_resource *r)list_add_tail (&mbm_local_event.list, &r->evt_list); } -int rdt_get_mon_l3_config (struct rdt_resource *r) +static void __init get_cqm_info (struct rdt_resource *r) +{ +u32 eax, ebx, ecx, edx; + +/* +* At this point, CQM LLC and one of occupancy, MBM total, and +* MBM local monitoring features must be supported. +*/ +cpuid_count (0x0000000F ,0 , &eax, &ebx, &ecx, &edx); +/* will be overridden if occupancy monitoring exists */ + r->num_rmid = ebx +1 ; + +cpuid_count (0x0000000F ,1 , &eax, &ebx, &ecx, &edx); + +if (boot_cpu_has (X86_FEATURE_CQM_OCCUP_LLC )) + r->num_rmid = ecx +1 ; + +if (boot_cpu_has (X86_FEATURE_CQM_MBM_TOTAL ) ||boot_cpu_has (X86_FEATURE_CQM_MBM_LOCAL )) + r->mon_scale = ebx; +else + r->mon_scale =-1 ; +} + +int __init rdt_get_mon_l3_config (struct rdt_resource *r) {unsigned int cl_size = boot_cpu_data.x86_cache_size;int ret; - r->mon_scale = boot_cpu_data.x86_cache_occ_scale; - r->num_rmid = boot_cpu_data.x86_cache_max_rmid +1 ; +get_cqm_info (r);/* * A reasonable upper limit on the max threshold is the number
Signature · 1 line
2.19.1
Borislav Petkov
maintainerRe: [RFC PATCH 1/3] x86/resctrl: Get max rmid and occupancy scale directly from CPUID instead of cpuinfo_x86
Quoted 114 lines · Fenghua Yu
Although
x86_cache_max_rmidandx86_cache_occ_scaleare read only once during resctrl initialization, they are always stored incpuinfo_x86on each CPU during run time even if resctrl is not configured.To save
cpuinfo_x86space and make CPU and resctrl initialization simpler, remove the two fields fromcpuinfo_x86and get max rmid and occupancy scale directly from CPUID during resctrl initialization. And since each known platform that supports resctrl has same max rmid on all CPUs, no need to scan all CPUs to find minimum of max rmid values, i.e. getting max rmid from CPUID on the current CPU is fine.Signed-off-by Fenghua Yu <[email protected]>--- arch/x86/include/asm/processor.h | 3 --- arch/x86/kernel/cpu/common.c | 28 -------------------------- arch/x86/kernel/cpu/resctrl/internal.h | 2 +- arch/x86/kernel/cpu/resctrl/monitor.c | 28 +++++++++++++++++++++++--- 4 files changed, 26 insertions(+), 35 deletions(-)arch/x86/include/asm/processor.h +0
-3@@ -99,9 +99,6 @@ struct cpuinfo_x86 {/* in KB - valid for CPUS which support this call: */ unsigned int x86_cache_size;int x86_cache_alignment;/* In bytes */ -/* Cache QoS architectural values: */ -int x86_cache_max_rmid;/* max index */ -int x86_cache_occ_scale;/* scale to bytes */ int x86_power;unsigned long loops_per_jiffy;/* cpuid returned max cores value: */ arch/x86/kernel/cpu/common.c +0
-28@@ -840,22 +840,9 @@ void get_cpu_cap(struct cpuinfo_x86 *c) c->x86_capability[CPUID_F_0_EDX ] = edx;if (cpu_has (c,X86_FEATURE_CQM_LLC )) { -/* will be overridden if occupancy monitoring exists */ - c->x86_cache_max_rmid = ebx; -/* QoS sub-leaf, EAX=0Fh, ECX=1 */ cpuid_count (0x0000000F ,1 , &eax, &ebx, &ecx, &edx); c->x86_capability[CPUID_F_1_EDX ] = edx; - -if ((cpu_has (c,X86_FEATURE_CQM_OCCUP_LLC )) || - ((cpu_has (c,X86_FEATURE_CQM_MBM_TOTAL )) || - (cpu_has (c,X86_FEATURE_CQM_MBM_LOCAL )))) { - c->x86_cache_max_rmid = ecx; - c->x86_cache_occ_scale = ebx; - } - }else { - c->x86_cache_max_rmid =-1 ; - c->x86_cache_occ_scale =-1 ; } }@@ -1269,20 +1256,6 @@ static void generic_identify(struct cpuinfo_x86 *c)#endif } -static void x86_init_cache_qos (struct cpuinfo_x86 *c) -{ -/* -* The heavy lifting of max_rmid and cache_occ_scale are handled -* in get_cpu_cap(). Here we just set the max_rmid for the boot_cpu -* in case CQM bits really aren't there in this CPU. -*/ -if (c != &boot_cpu_data) { - boot_cpu_data.x86_cache_max_rmid = -min (boot_cpu_data.x86_cache_max_rmid, - c->x86_cache_max_rmid); - } -} -/* * Validate that ACPI/mptables have the same information about the * effective APIC id and update the package map. @@ -1391,7 +1364,6 @@ static void identify_cpu(struct cpuinfo_x86 *c)#endif x86_init_rdrand (c); -x86_init_cache_qos (c);setup_pku (c);/* arch/x86/kernel/cpu/resctrl/internal.h +1
-1@@ -579,7 +579,7 @@ int closids_supported(void);void closid_free (int closid);int alloc_rmid (void );void free_rmid (u32 rmid); -int rdt_get_mon_l3_config (struct rdt_resource *r); +int __init rdt_get_mon_l3_config (struct rdt_resource *r);void mon_event_count (void *info);int rdtgroup_mondata_show (struct seq_file *m,void *arg); void rmdir_mondata_subdir_allrdtgrp(struct rdt_resource *r,arch/x86/kernel/cpu/resctrl/monitor.c +3
-1@@ -617,13 +617,35 @@ static void l3_mon_evt_init(struct rdt_resource *r)list_add_tail (&mbm_local_event.list, &r->evt_list); } -int rdt_get_mon_l3_config (struct rdt_resource *r) +static void __init get_cqm_info (struct rdt_resource *r) +{ +u32 eax, ebx, ecx, edx;
arch/x86/kernel/cpu/resctrl/monitor.c static void l3_mon_evt_init(struct rdt_resource *r) · line 632
Quoted 10 lines
arch/x86/kernel/cpu/resctrl/monitor.c +10
-0+ +/* +* At this point, CQM LLC and one of occupancy, MBM total, and +* MBM local monitoring features must be supported. +*/ +cpuid_count (0x0000000F ,0 , &eax, &ebx, &ecx, &edx); +/* will be overridden if occupancy monitoring exists */ + r->num_rmid = ebx +1 ; + +cpuid_count (0x0000000F ,1 , &eax, &ebx, &ecx, &edx);
Those CPUID accesses should be done *after* testing features, not before.
arch/x86/kernel/cpu/resctrl/monitor.c static void l3_mon_evt_init(struct rdt_resource *r) · line 634
Quoted 2 lines
arch/x86/kernel/cpu/resctrl/monitor.c +2
-0+ +if (boot_cpu_has (X86_FEATURE_CQM_OCCUP_LLC ))
That is already done in get_rdt_mon_resources() and rdt_mon_features caches those bits. I think you wanna test QOS_L3_OCCUP_EVENT_ID in there and then read CPUID 0xf and set ->num_rmid.
arch/x86/kernel/cpu/resctrl/monitor.c static void l3_mon_evt_init(struct rdt_resource *r) · line 637
Quoted 3 lines
arch/x86/kernel/cpu/resctrl/monitor.c +3
-0+ r->num_rmid = ecx +1 ; + +if (boot_cpu_has (X86_FEATURE_CQM_MBM_TOTAL ) ||boot_cpu_has (X86_FEATURE_CQM_MBM_LOCAL ))
Ditto.
Other than that, I like where this cleanup is going...
Signature · 4 lines
Regards/Gruss,
Boris.
Good mailing practices for 400: avoid top-posting and trim the reply.Re: [RFC PATCH 1/3] x86/resctrl: Get max rmid and occupancy scale directly from CPUID instead of cpuinfo_x86
Quoted 129 lines · Borislav Petkov
Quoted 114 lines · Fenghua Yu
Although
x86_cache_max_rmidandx86_cache_occ_scaleare read only once during resctrl initialization, they are always stored incpuinfo_x86on each CPU during run time even if resctrl is not configured.To save
cpuinfo_x86space and make CPU and resctrl initialization simpler, remove the two fields fromcpuinfo_x86and get max rmid and occupancy scale directly from CPUID during resctrl initialization. And since each known platform that supports resctrl has same max rmid on all CPUs, no need to scan all CPUs to find minimum of max rmid values, i.e. getting max rmid from CPUID on the current CPU is fine.Signed-off-by Fenghua Yu <[email protected]>--- arch/x86/include/asm/processor.h | 3 --- arch/x86/kernel/cpu/common.c | 28 -------------------------- arch/x86/kernel/cpu/resctrl/internal.h | 2 +- arch/x86/kernel/cpu/resctrl/monitor.c | 28 +++++++++++++++++++++++--- 4 files changed, 26 insertions(+), 35 deletions(-)arch/x86/include/asm/processor.h +0
-3@@ -99,9 +99,6 @@ struct cpuinfo_x86 {/* in KB - valid for CPUS which support this call: */ unsigned int x86_cache_size;int x86_cache_alignment;/* In bytes */ -/* Cache QoS architectural values: */ -int x86_cache_max_rmid;/* max index */ -int x86_cache_occ_scale;/* scale to bytes */ int x86_power;unsigned long loops_per_jiffy;/* cpuid returned max cores value: */ arch/x86/kernel/cpu/common.c +0
-28@@ -840,22 +840,9 @@ void get_cpu_cap(struct cpuinfo_x86 *c) c->x86_capability[CPUID_F_0_EDX ] = edx;if (cpu_has (c,X86_FEATURE_CQM_LLC )) { -/* will be overridden if occupancy monitoring exists */ - c->x86_cache_max_rmid = ebx; -/* QoS sub-leaf, EAX=0Fh, ECX=1 */ cpuid_count (0x0000000F ,1 , &eax, &ebx, &ecx, &edx); c->x86_capability[CPUID_F_1_EDX ] = edx; - -if ((cpu_has (c,X86_FEATURE_CQM_OCCUP_LLC )) || - ((cpu_has (c,X86_FEATURE_CQM_MBM_TOTAL )) || - (cpu_has (c,X86_FEATURE_CQM_MBM_LOCAL )))) { - c->x86_cache_max_rmid = ecx; - c->x86_cache_occ_scale = ebx; - } - }else { - c->x86_cache_max_rmid =-1 ; - c->x86_cache_occ_scale =-1 ; } }@@ -1269,20 +1256,6 @@ static void generic_identify(struct cpuinfo_x86 *c)#endif } -static void x86_init_cache_qos (struct cpuinfo_x86 *c) -{ -/* -* The heavy lifting of max_rmid and cache_occ_scale are handled -* in get_cpu_cap(). Here we just set the max_rmid for the boot_cpu -* in case CQM bits really aren't there in this CPU. -*/ -if (c != &boot_cpu_data) { - boot_cpu_data.x86_cache_max_rmid = -min (boot_cpu_data.x86_cache_max_rmid, - c->x86_cache_max_rmid); - } -} -/* * Validate that ACPI/mptables have the same information about the * effective APIC id and update the package map. @@ -1391,7 +1364,6 @@ static void identify_cpu(struct cpuinfo_x86 *c)#endif x86_init_rdrand (c); -x86_init_cache_qos (c);setup_pku (c);/* arch/x86/kernel/cpu/resctrl/internal.h +1
-1@@ -579,7 +579,7 @@ int closids_supported(void);void closid_free (int closid);int alloc_rmid (void );void free_rmid (u32 rmid); -int rdt_get_mon_l3_config (struct rdt_resource *r); +int __init rdt_get_mon_l3_config (struct rdt_resource *r);void mon_event_count (void *info);int rdtgroup_mondata_show (struct seq_file *m,void *arg); void rmdir_mondata_subdir_allrdtgrp(struct rdt_resource *r,arch/x86/kernel/cpu/resctrl/monitor.c +3
-1@@ -617,13 +617,35 @@ static void l3_mon_evt_init(struct rdt_resource *r)list_add_tail (&mbm_local_event.list, &r->evt_list); } -int rdt_get_mon_l3_config (struct rdt_resource *r) +static void __init get_cqm_info (struct rdt_resource *r) +{ +u32 eax, ebx, ecx, edx;arch/x86/kernel/cpu/resctrl/monitor.c static void l3_mon_evt_init(struct rdt_resource *r) · line 632
Quoted 10 lines
arch/x86/kernel/cpu/resctrl/monitor.c +10
-0+ +/* +* At this point, CQM LLC and one of occupancy, MBM total, and +* MBM local monitoring features must be supported. +*/ +cpuid_count (0x0000000F ,0 , &eax, &ebx, &ecx, &edx); +/* will be overridden if occupancy monitoring exists */ + r->num_rmid = ebx +1 ; + +cpuid_count (0x0000000F ,1 , &eax, &ebx, &ecx, &edx);Those CPUID accesses should be done *after* testing features, not before.
Quoted 5 lines
arch/x86/kernel/cpu/resctrl/monitor.c static void l3_mon_evt_init(struct rdt_resource *r) · line 634
Quoted 2 lines
arch/x86/kernel/cpu/resctrl/monitor.c +2
-0+ +if (boot_cpu_has (X86_FEATURE_CQM_OCCUP_LLC ))That is already done in
get_rdt_mon_resources()andrdt_mon_featurescaches those bits. I think you wanna testQOS_L3_OCCUP_EVENT_IDin there
Sure. I can test QOS_L3_OCCUP_EVENT_ID instead of cpuinfo_x86.
Quoted 1 line
and then read CPUID
0xfand set ->num_rmid.
When this function is called, X86_FEATURE_CQM_LLC must be supported and one of X86_FEATURE_CQM_OCCUP_LLC, X86_FEATURE_CQM_MBM_LOCAL, and X86_FEATURE_CQM_MBM_TOTAL must be supported. Otherwise, get_rdt_mon_resource() is returned before calling rdt_get_mon_l3_config().
So CPUID.f.0 and CPUID.f.1 must be readable and return meaningful data without testing the features.
But the problem is it's unknown which ones of X86_FEATURE_CQM_OCCUP_LLC, X86_FEATURE_CQM_MBM_LOCAL, and X86_FEATURE_CQM_MBM_TOTAL are supported.
So after reading CPUID.f.1, test X86_FEATURE_CQM_OCCUP_LLC in order to override r->num_rmid from ecx. And test X86_FEATURE_CQM_MBM_TOTAL or X86_FEATURE_CQM_MBM_LOCAL in order to get r>mon_scale from ebx.
Is this sequence OK? If the current sequence is OK, I only need to change code to test the QOS_L3_*_EVENT_ID, right?
Quoted 6 lines
arch/x86/kernel/cpu/resctrl/monitor.c static void l3_mon_evt_init(struct rdt_resource *r) · line 637
Quoted 4 lines
arch/x86/kernel/cpu/resctrl/monitor.c +3
-0+ r->num_rmid = ecx +1 ; + +if (boot_cpu_has (X86_FEATURE_CQM_MBM_TOTAL ) ||boot_cpu_has (X86_FEATURE_CQM_MBM_LOCAL ))Ditto.
Thanks.
-Fenghua
Borislav Petkov
maintainerRe: [RFC PATCH 1/3] x86/resctrl: Get max rmid and occupancy scale directly from CPUID instead of cpuinfo_x86
Quoted 7 lines · Fenghua Yu
When this function is called,
X86_FEATURE_CQM_LLCmust be supported and one ofX86_FEATURE_CQM_OCCUP_LLC,X86_FEATURE_CQM_MBM_LOCAL, andX86_FEATURE_CQM_MBM_TOTALmust be supported. Otherwise,get_rdt_mon_resource()is returned before callingrdt_get_mon_l3_config().So CPUID.f.0 and CPUID.f.1 must be readable and return meaningful data without testing the features.
How's that?
static void __init get_cqm_info(struct rdt_resource *r)
{
u32 eax, ebx, ecx, edx;
/*
* At this point, CQM LLC and one of occupancy, MBM total, and
* MBM local monitoring features must be supported.
*/
cpuid_count(0xf, 1, &eax, &ebx, &ecx, &edx);
if (rdt_mon_features & QOS_L3_OCCUP_EVENT_ID)
r->num_rmid = ecx + 1;
else
/*
* Fallback maximum range (zero-based) of RMID within
* this physical processor of all types, in subleaf 0,
* EBX.
*/
r->num_rmid = cpuid_ebx(0xf) + 1;
if (rdt_mon_features & (QOS_L3_MBM_TOTAL_EVENT_ID |
QOS_L3_MBM_LOCAL_EVENT_ID))
r->mon_scale = ebx;
else
r->mon_scale = -1;
}Signature · 4 lines
Regards/Gruss,
Boris.
Good mailing practices for 400: avoid top-posting and trim the reply.Re: [RFC PATCH 1/3] x86/resctrl: Get max rmid and occupancy scale directly from CPUID instead of cpuinfo_x86
Quoted 29 lines · Borislav Petkov
Quoted 7 lines · Fenghua Yu
When this function is called,
X86_FEATURE_CQM_LLCmust be supported and one ofX86_FEATURE_CQM_OCCUP_LLC,X86_FEATURE_CQM_MBM_LOCAL, andX86_FEATURE_CQM_MBM_TOTALmust be supported. Otherwise,get_rdt_mon_resource()is returned before callingrdt_get_mon_l3_config().So CPUID.f.0 and CPUID.f.1 must be readable and return meaningful data without testing the features.
How's that?
static void
__initget_cqm_info(struct rdt_resource *r){ u32 eax, ebx, ecx, edx; /* * At this point, CQM LLC and one of occupancy, MBM total, and * MBM local monitoring features must be supported. */ cpuid_count(0xf, 1, &eax, &ebx, &ecx, &edx); if (rdt_mon_features & QOS_L3_OCCUP_EVENT_ID) r->num_rmid = ecx + 1; else /* * Fallback maximum range (zero-based) of RMID within * this physical processor of all types, in subleaf 0, * EBX. */
Quoted 8 lines
r->num_rmid = cpuid_ebx(0xf) + 1;
if (rdt_mon_features & (QOS_L3_MBM_TOTAL_EVENT_ID |
QOS_L3_MBM_LOCAL_EVENT_ID))
r->mon_scale = ebx;
else
r->mon_scale = -1;
}This is much better and cleaner! I will copy your code to patch 0001 in the next version.
Thanks.
-Fenghua
RFC PATCH 2/3 x86/cpufeatures: Combine word 11 and 12 into new scattered features word 11
In reply to Fenghua YuIt's a waste for the four X86_FEATURE_CQM_* features to occupy two pure feature bits words. To better utilize feature words, re-define word 11 to host scattered features and move the four X86_FEATURE_CQM_* features into word 11. More scattered features can be added in word 11 in the future.
KVM doesn't support resctrl now. So it's safe to move the X86_FEATURE_CQM_* features to scattered features word 11 for KVM.
--- arch/x86/include/asm/cpufeature.h | 3 +-- arch/x86/include/asm/cpufeatures.h | 17 ++++++++++------- arch/x86/kernel/cpu/common.c | 14 -------------- arch/x86/kernel/cpu/cpuid-deps.c | 3 +++ arch/x86/kernel/cpu/scattered.c | 4 ++++ arch/x86/kvm/cpuid.h | 2 -- 6 files changed, 18 insertions(+), 25 deletions(-)
arch/x86/include/asm/cpufeature.h +1 -2
@@ -22,8 +22,7 @@ enum cpuid_leafsCPUID_LNX_3 ,CPUID_7_0_EBX ,CPUID_D_1_EAX , -CPUID_F_0_EDX , -CPUID_F_1_EDX , +CPUID_LNX_4 ,CPUID_8000_0008_EBX ,CPUID_6_EAX , CPUID_8000_000A_EDX,
arch/x86/include/asm/cpufeatures.h +10 -7
@@ -269,13 +269,16 @@#define X86_FEATURE_XGETBV1 (10*32+ 2)/* XGETBV with ECX = 1 instruction */ #define X86_FEATURE_XSAVES (10*32+ 3)/* XSAVES/XRSTORS instructions */ -/* Intel-defined CPU QoS Sub-leaf, CPUID level 0x0000000F:0 (EDX), word 11 */ -#define X86_FEATURE_CQM_LLC (11*32+ 1)/* LLC QoS if 1 */ - -/* Intel-defined CPU QoS Sub-leaf, CPUID level 0x0000000F:1 (EDX), word 12 */ -#define X86_FEATURE_CQM_OCCUP_LLC (12*32+ 0)/* LLC occupancy monitoring */ -#define X86_FEATURE_CQM_MBM_TOTAL (12*32+ 1)/* LLC Total MBM monitoring */ -#define X86_FEATURE_CQM_MBM_LOCAL (12*32+ 2)/* LLC Local MBM monitoring */ +/* +* Extended auxiliary flags: Linux defined - For features scattered in various +* CPUID levels like 0xf, word 11. +* +* Reuse free bits when adding new feature flags! +*/ +#define X86_FEATURE_CQM_LLC (11*32+ 0)/* LLC QoS if 1 */ +#define X86_FEATURE_CQM_OCCUP_LLC (11*32+ 1)/* LLC occupancy monitoring */ +#define X86_FEATURE_CQM_MBM_TOTAL (11*32+ 2)/* LLC Total MBM monitoring */ +#define X86_FEATURE_CQM_MBM_LOCAL (11*32+ 3)/* LLC Local MBM monitoring */ /* AMD-defined CPU features, CPUID level 0x80000008 (EBX), word 13 */ #define X86_FEATURE_CLZERO (13*32+ 0)/* CLZERO instruction */
arch/x86/kernel/cpu/common.c +0 -14
@@ -832,20 +832,6 @@ void get_cpu_cap(struct cpuinfo_x86 *c) c->x86_capability[CPUID_D_1_EAX ] = eax; } -/* Additional Intel-defined flags: level 0x0000000F */ -if (c->cpuid_level >=0x0000000F ) { - -/* QoS sub-leaf, EAX=0Fh, ECX=0 */ -cpuid_count (0x0000000F ,0 , &eax, &ebx, &ecx, &edx); - c->x86_capability[CPUID_F_0_EDX ] = edx; - -if (cpu_has (c,X86_FEATURE_CQM_LLC )) { -/* QoS sub-leaf, EAX=0Fh, ECX=1 */ -cpuid_count (0x0000000F ,1 , &eax, &ebx, &ecx, &edx); - c->x86_capability[CPUID_F_1_EDX ] = edx; - } - } -/* AMD-defined flags: level 0x80000001 */ eax =cpuid_eax (0x80000000 ); c->extended_cpuid_level = eax;
arch/x86/kernel/cpu/cpuid-deps.c +3 -0
@@ -59,6 +59,9 @@ static const struct cpuid_dep cpuid_deps[] = { { X86_FEATURE_AVX512_4VNNIW, X86_FEATURE_AVX512F }, { X86_FEATURE_AVX512_4FMAPS, X86_FEATURE_AVX512F }, { X86_FEATURE_AVX512_VPOPCNTDQ, X86_FEATURE_AVX512F }, + { X86_FEATURE_CQM_OCCUP_LLC, X86_FEATURE_CQM_LLC }, + { X86_FEATURE_CQM_MBM_TOTAL, X86_FEATURE_CQM_LLC }, + { X86_FEATURE_CQM_MBM_LOCAL, X86_FEATURE_CQM_LLC }, {} };
arch/x86/kernel/cpu/scattered.c +4 -0
@@ -26,6 +26,10 @@ struct cpuid_bit {static const struct cpuid_bit cpuid_bits[] = { {X86_FEATURE_APERFMPERF ,CPUID_ECX ,0 ,0x00000006 ,0 }, {X86_FEATURE_EPB ,CPUID_ECX ,3 ,0x00000006 ,0 }, + {X86_FEATURE_CQM_LLC ,CPUID_EDX ,1 ,0x0000000f ,0 }, + {X86_FEATURE_CQM_OCCUP_LLC ,CPUID_EDX ,0 ,0x0000000f ,1 }, + {X86_FEATURE_CQM_MBM_TOTAL ,CPUID_EDX ,1 ,0x0000000f ,1 }, + {X86_FEATURE_CQM_MBM_LOCAL ,CPUID_EDX ,2 ,0x0000000f ,1 }, {X86_FEATURE_CAT_L3 ,CPUID_EBX ,1 ,0x00000010 ,0 }, {X86_FEATURE_CAT_L2 ,CPUID_EBX ,2 ,0x00000010 ,0 }, {X86_FEATURE_CDP_L3 ,CPUID_ECX ,2 ,0x00000010 ,1 },
arch/x86/kvm/cpuid.h +0 -2
@@ -47,8 +47,6 @@ static const struct cpuid_reg reverse_cpuid[] = { [CPUID_8000_0001_ECX] = {0x80000001, 0, CPUID_ECX}, [CPUID_7_0_EBX] = { 7, 0, CPUID_EBX}, [CPUID_D_1_EAX] = { 0xd, 1, CPUID_EAX}, - [CPUID_F_0_EDX] = { 0xf, 0, CPUID_EDX}, - [CPUID_F_1_EDX] = { 0xf, 1, CPUID_EDX}, [CPUID_8000_0008_EBX] = {0x80000008, 0, CPUID_EBX}, [CPUID_6_EAX] = { 6, 0, CPUID_EAX}, [CPUID_8000_000A_EDX] = {0x8000000a, 0, CPUID_EDX},
Signature · 1 line
2.19.1
Borislav Petkov
maintainerRe: [RFC PATCH 2/3] x86/cpufeatures: Combine word 11 and 12 into new scattered features word 11
Quoted 10 lines · Fenghua Yu
It's a waste for the four X86_FEATURE_CQM_* features to occupy two pure feature bits words. To better utilize feature words, re-define word 11 to host scattered features and move the four X86_FEATURE_CQM_* features into word 11. More scattered features can be added in word 11 in the future.
KVM doesn't support resctrl now. So it's safe to move the X86_FEATURE_CQM_* features to scattered features word 11 for KVM.
Signed-off-by Fenghua Yu <[email protected]>
...
arch/x86/kvm/cpuid.h static const struct cpuid_reg reverse_cpuid[] = { · line 51
Quoted 10 lines
arch/x86/kvm/cpuid.h +0
-2@@ -47,8 +47,6 @@ static const struct cpuid_reg reverse_cpuid[] = { [CPUID_8000_0001_ECX] = {0x80000001, 0, CPUID_ECX}, [CPUID_7_0_EBX] = { 7, 0, CPUID_EBX}, [CPUID_D_1_EAX] = { 0xd, 1, CPUID_EAX}, - [CPUID_F_0_EDX] = { 0xf, 0, CPUID_EDX}, - [CPUID_F_1_EDX] = { 0xf, 1, CPUID_EDX},
I think you're going to have to change those to:
[CPUID_LNX_4] = { 0, 0, 0},
[CPUID_7_1_EAX] = { 7, 1, CPUID_EAX },instead of removing them because kvm is basically hardcoding the feature words and then bitches when array elements in the middle get removed:
In file included from ./include/linux/export.h:45,
from ./include/linux/linkage.h:7,
from ./include/linux/preempt.h:10,
from ./include/linux/hardirq.h:5,
from ./include/linux/kvm_host.h:10,
from arch/x86/kvm/x86.c:22:In function ‘x86_feature_cpuid’,
inlined from ‘guest_cpuid_get_register’ at arch/x86/kvm/cpuid.h:71:33,
inlined from ‘guest_cpuid_has’ at arch/x86/kvm/cpuid.h:100:8,
inlined from ‘kvm_get_msr_common’ at arch/x86/kvm/x86.c:2804:8:./include/linux/compiler.h:345:38: error: call to ‘__compiletime_assert_62’ declared with attribute error: BUILD_BUG_ON failed: x86_leaf >= ARRAY_SIZE(reverse_cpuid)
_compiletime_assert(condition, msg, __compiletime_assert_, __LINE__)
^
./include/linux/compiler.h:326:4: note: in definition of macro ‘__compiletime_assert’
prefix ## suffix(); \
^~~~~~./include/linux/compiler.h:345:2: note: in expansion of macro ‘_compiletime_assert’
_compiletime_assert(condition, msg, __compiletime_assert_, __LINE__) ^~~~~~~~~~~~~~~~~~~./include/linux/build_bug.h:39:37: note: in expansion of macro ‘compiletime_assert’
#define BUILD_BUG_ON_MSG(cond, msg) compiletime_assert(!(cond), msg)
^~~~~~~~~~~~~~~~~~./include/linux/build_bug.h:50:2: note: in expansion of macro ‘BUILD_BUG_ON_MSG’
BUILD_BUG_ON_MSG(condition, "BUILD_BUG_ON failed: " #condition)
^~~~~~~~~~~~~~~~arch/x86/kvm/cpuid.h:62:2: note: in expansion of macro ‘BUILD_BUG_ON’
BUILD_BUG_ON(x86_leaf >= ARRAY_SIZE(reverse_cpuid));
^~~~~~~~~~~~./include/linux/compiler.h:345:38: error: call to ‘__compiletime_assert_63’ declared with attribute error: BUILD_BUG_ON failed: reverse_cpuid[x86_leaf].function == 0
_compiletime_assert(condition, msg, __compiletime_assert_, __LINE__)
^
./include/linux/compiler.h:326:4: note: in definition of macro ‘__compiletime_assert’
prefix ## suffix(); \
^~~~~~./include/linux/compiler.h:345:2: note: in expansion of macro ‘_compiletime_assert’
_compiletime_assert(condition, msg, __compiletime_assert_, __LINE__) ^~~~~~~~~~~~~~~~~~~./include/linux/build_bug.h:39:37: note: in expansion of macro ‘compiletime_assert’
#define BUILD_BUG_ON_MSG(cond, msg) compiletime_assert(!(cond), msg)
^~~~~~~~~~~~~~~~~~./include/linux/build_bug.h:50:2: note: in expansion of macro ‘BUILD_BUG_ON_MSG’
BUILD_BUG_ON_MSG(condition, "BUILD_BUG_ON failed: " #condition) ^~~~~~~~~~~~~~~~arch/x86/kvm/cpuid.h:63:2: note: in expansion of macro ‘BUILD_BUG_ON’
BUILD_BUG_ON(reverse_cpuid[x86_leaf].function == 0);
^~~~~~~~~~~~
make[2]: *** [scripts/Makefile.build:278: arch/x86/kvm/x86.o] Error 1 make[2]: *** Waiting for unfinished jobs....
make[1]: *** [scripts/Makefile.build:489: arch/x86/kvm] Error 2 make[1]: *** Waiting for unfinished jobs....
make: *** [Makefile:1071: arch/x86] Error 2
make: *** Waiting for unfinished jobs....
Signature · 4 lines
Regards/Gruss,
Boris.
Good mailing practices for 400: avoid top-posting and trim the reply.Borislav Petkov
maintainerRe: [RFC PATCH 2/3] x86/cpufeatures: Combine word 11 and 12 into new scattered features word 11
Quoted 27 lines · Borislav Petkov
Quoted 10 lines · Fenghua Yu
It's a waste for the four X86_FEATURE_CQM_* features to occupy two pure feature bits words. To better utilize feature words, re-define word 11 to host scattered features and move the four X86_FEATURE_CQM_* features into word 11. More scattered features can be added in word 11 in the future.
KVM doesn't support resctrl now. So it's safe to move the X86_FEATURE_CQM_* features to scattered features word 11 for KVM.
Signed-off-by Fenghua Yu <[email protected]>...
arch/x86/kvm/cpuid.h static const struct cpuid_reg reverse_cpuid[] = { · line 51
Quoted 10 lines
arch/x86/kvm/cpuid.h +0
-2@@ -47,8 +47,6 @@ static const struct cpuid_reg reverse_cpuid[] = { [CPUID_8000_0001_ECX] = {0x80000001, 0, CPUID_ECX}, [CPUID_7_0_EBX] = { 7, 0, CPUID_EBX}, [CPUID_D_1_EAX] = { 0xd, 1, CPUID_EAX}, - [CPUID_F_0_EDX] = { 0xf, 0, CPUID_EDX}, - [CPUID_F_1_EDX] = { 0xf, 1, CPUID_EDX},I think you're going to have to change those to:
Quoted 5 lines
[CPUID_LNX_4] = { 0, 0, 0}, [CPUID_7_1_EAX] = { 7, 1, CPUID_EAX },instead of removing them because kvm is basically hardcoding the feature words and then bitches when array elements in the middle get removed:
Alternatively - and what I think is the better solution - would be to remove those BUILD_BUG_ONs in x86_feature_cpuid and filter out the Linux-defined leafs dynamically. This way the array won't have holes in it.
Signature · 4 lines
Regards/Gruss,
Boris.
Good mailing practices for 400: avoid top-posting and trim the reply.Re: [RFC PATCH 2/3] x86/cpufeatures: Combine word 11 and 12 into new scattered features word 11
Quoted 31 lines · Borislav Petkov
Quoted 30 lines · Borislav Petkov
Quoted 10 lines · Fenghua Yu
It's a waste for the four X86_FEATURE_CQM_* features to occupy two pure feature bits words. To better utilize feature words, re-define word 11 to host scattered features and move the four X86_FEATURE_CQM_* features into word 11. More scattered features can be added in word 11 in the future.
KVM doesn't support resctrl now. So it's safe to move the X86_FEATURE_CQM_* features to scattered features word 11 for KVM.
Signed-off-by Fenghua Yu <[email protected]>...
arch/x86/kvm/cpuid.h static const struct cpuid_reg reverse_cpuid[] = { · line 51
Quoted 10 lines
arch/x86/kvm/cpuid.h +0
-2@@ -47,8 +47,6 @@ static const struct cpuid_reg reverse_cpuid[] = { [CPUID_8000_0001_ECX] = {0x80000001, 0, CPUID_ECX}, [CPUID_7_0_EBX] = { 7, 0, CPUID_EBX}, [CPUID_D_1_EAX] = { 0xd, 1, CPUID_EAX}, - [CPUID_F_0_EDX] = { 0xf, 0, CPUID_EDX}, - [CPUID_F_1_EDX] = { 0xf, 1, CPUID_EDX},I think you're going to have to change those to:
[CPUID_LNX_4] = { 0, 0, 0}, [CPUID_7_1_EAX] = { 7, 1, CPUID_EAX },
Quoted 7 lines
Quoted 2 lines
instead of removing them because kvm is basically hardcoding the feature words and then bitches when array elements in the middle get removed:
Alternatively - and what I think is the better solution - would be to remove those
BUILD_BUG_ONsinx86_feature_cpuidand filter out the Linux-defined leafs dynamically. This way the array won't have holes in it.
Maybe adding a dummy slot in cpuid_leafs in patch 0002 to avoid the compilation errors?
Those KVM BUILD_BUG_ON() want to find out any empty leaf in reverse_cpuid (and also cpuid_leafs).
The patch 0002 actually creates such empty leaf 12 in cpuid_leafs. So CPUID_7_EDX=17 instead of NCAPINTS-1=18 in cpuid_leafs. It's not desired.
After applying patch 0003, the hole is filled in and there is no compilation error from the KVM BUILD_BUG_ON() checks. So the compilation errors only happens in bisect.
arch/x86/include/asm/cpufeature.h +1 -0
@@ -23,6 +23,7 @@ enum cpuid_leafsCPUID_7_0_EBX ,CPUID_D_1_EAX ,CPUID_LNX_4 , +CPUID_DUMMY ,CPUID_8000_0008_EBX ,CPUID_6_EAX , CPUID_8000_000A_EDX,
Is it OK to add the above patch to the patch 0002 to solve the compilation errors? If it's ok, I will explain in commit message why add CPUID_DUMMY.
Thanks.
-Fenghua
Borislav Petkov
maintainerRe: [RFC PATCH 2/3] x86/cpufeatures: Combine word 11 and 12 into new scattered features word 11
+ Radim and Paolo. See upthread for context.
Quoted 7 lines · Fenghua Yu
Quoted 4 lines
Alternatively - and what I think is the better solution - would be to remove those
BUILD_BUG_ONsinx86_feature_cpuidand filter out the Linux-defined leafs dynamically. This way the array won't have holes in it.Maybe adding a dummy slot in
cpuid_leafsin patch 0002 to avoid the compilation errors?
Maybe you didn't read what you're replying to: "This way the array won't have holes in it". Ontop of yours:
arch/x86/kvm/cpuid.h +4 -2
@@ -47,6 +47,7 @@ static const struct cpuid_reg reverse_cpuid[] = { [CPUID_8000_0001_ECX] = {0x80000001, 0, CPUID_ECX}, [CPUID_7_0_EBX] = { 7, 0, CPUID_EBX}, [CPUID_D_1_EAX] = { 0xd, 1, CPUID_EAX}, + [CPUID_7_1_EAX] = { 7, 1, CPUID_EAX}, [CPUID_8000_0008_EBX] = {0x80000008, 0, CPUID_EBX}, [CPUID_6_EAX] = { 6, 0, CPUID_EAX}, [CPUID_8000_000A_EDX] = {0x8000000a, 0, CPUID_EDX},
@@ -59,8 +60,9 @@ static __always_inline struct cpuid_reg x86_feature_cpuid(unsigned x86_feature) {unsigned x86_leaf = x86_feature /32 ; -BUILD_BUG_ON (x86_leaf >=ARRAY_SIZE (reverse_cpuid)); -BUILD_BUG_ON (reverse_cpuid[x86_leaf].function ==0 ); +if (x86_leaf ==CPUID_LNX_1 || + x86_leaf ==CPUID_LNX_4 ) +return NULL ;return reverse_cpuid[x86_leaf]; }
That's what I mean with filter out dynamically.
Signature · 4 lines
Regards/Gruss,
Boris.
Good mailing practices for 400: avoid top-posting and trim the reply.Re: [RFC PATCH 2/3] x86/cpufeatures: Combine word 11 and 12 into new scattered features word 11
Quoted 19 lines · Borislav Petkov
+ Radim and Paolo. See upthread for context.
Quoted 7 lines · Fenghua Yu
Quoted 4 lines
Alternatively - and what I think is the better solution - would be to remove those
BUILD_BUG_ONsinx86_feature_cpuidand filter out the Linux-defined leafs dynamically. This way the array won't have holes in it.Maybe adding a dummy slot in
cpuid_leafsin patch 0002 to avoid the compilation errors?Maybe you didn't read what you're replying to: "This way the array won't have holes in it". Ontop of yours:
arch/x86/kvm/cpuid.h +0
-0
arch/x86/kvm/cpuid.h static const struct cpuid_reg reverse_cpuid[] = { · line 50
Quoted 4 lines
arch/x86/kvm/cpuid.h +1
-0@@ -47,6 +47,7 @@ static const struct cpuid_reg reverse_cpuid[] = { [CPUID_8000_0001_ECX] = {0x80000001, 0, CPUID_ECX}, [CPUID_7_0_EBX] = { 7, 0, CPUID_EBX}, [CPUID_D_1_EAX] = { 0xd, 1, CPUID_EAX}, + [CPUID_7_1_EAX] = { 7, 1, CPUID_EAX},
CPUID_7_1_EAX is defined in patch 0003. Should I combine patch 0002 and 0003 into one patch?
arch/x86/kvm/cpuid.h static __always_inline struct cpuid_reg x86_feature_cpuid(unsigned x86_feature) · line 65
Quoted 12 lines
arch/x86/kvm/cpuid.h +3
-2[CPUID_8000_0008_EBX] = {0x80000008, 0, CPUID_EBX}, [CPUID_6_EAX] = { 6, 0, CPUID_EAX}, [CPUID_8000_000A_EDX] = {0x8000000a, 0, CPUID_EDX},@@ -59,8 +60,9 @@ static __always_inline struct cpuid_reg x86_feature_cpuid(unsigned x86_feature) {unsigned x86_leaf = x86_feature /32 ; -BUILD_BUG_ON (x86_leaf >=ARRAY_SIZE (reverse_cpuid)); -BUILD_BUG_ON (reverse_cpuid[x86_leaf].function ==0 ); +if (x86_leaf ==CPUID_LNX_1 || + x86_leaf ==CPUID_LNX_4 ) +return NULL ;
Need to check CPUID_LNX_2 and CPUID_LNX_3 as well?
Quoted 5 lines
arch/x86/kvm/cpuid.h +0
-0return reverse_cpuid[x86_leaf]; }That's what I mean with filter out dynamically.
Thanks.
-Fenghua
Borislav Petkov
maintainerRe: [RFC PATCH 2/3] x86/cpufeatures: Combine word 11 and 12 into new scattered features word 11
Quoted 2 lines · Fenghua Yu
CPUID_7_1_EAXis defined in patch 0003. Should I combine patch 0002 and 0003 into one patch?
That was just an example diff. Generally, patches should do one logical thing. In this case, I'd suggest you make the kvm change the first and separate patch, from the other three.
Quoted 1 line
Need to check
CPUID_LNX_2andCPUID_LNX_3as well?
Yes, all linux-defined feature words.
Signature · 4 lines
Regards/Gruss,
Boris.
Good mailing practices for 400: avoid top-posting and trim the reply.Re: [RFC PATCH 2/3] x86/cpufeatures: Combine word 11 and 12 into new scattered features word 11
In reply to Borislav PetkovQuoted 39 lines · Borislav Petkov
+ Radim and Paolo. See upthread for context.
Quoted 7 lines · Fenghua Yu
Quoted 4 lines
Alternatively - and what I think is the better solution - would be to remove those
BUILD_BUG_ONsinx86_feature_cpuidand filter out the Linux-defined leafs dynamically. This way the array won't have holes in it.Maybe adding a dummy slot in
cpuid_leafsin patch 0002 to avoid the compilation errors?Maybe you didn't read what you're replying to: "This way the array won't have holes in it". Ontop of yours:
arch/x86/kvm/cpuid.h +4
-2@@ -47,6 +47,7 @@ static const struct cpuid_reg reverse_cpuid[] = { [CPUID_8000_0001_ECX] = {0x80000001, 0, CPUID_ECX}, [CPUID_7_0_EBX] = { 7, 0, CPUID_EBX}, [CPUID_D_1_EAX] = { 0xd, 1, CPUID_EAX}, + [CPUID_7_1_EAX] = { 7, 1, CPUID_EAX}, [CPUID_8000_0008_EBX] = {0x80000008, 0, CPUID_EBX}, [CPUID_6_EAX] = { 6, 0, CPUID_EAX}, [CPUID_8000_000A_EDX] = {0x8000000a, 0, CPUID_EDX},@@ -59,8 +60,9 @@ static __always_inline struct cpuid_reg x86_feature_cpuid(unsigned x86_feature) {unsigned x86_leaf = x86_feature /32 ; -BUILD_BUG_ON (x86_leaf >=ARRAY_SIZE (reverse_cpuid)); -BUILD_BUG_ON (reverse_cpuid[x86_leaf].function ==0 ); +if (x86_leaf ==CPUID_LNX_1 || + x86_leaf ==CPUID_LNX_4 ) +return NULL ;return reverse_cpuid[x86_leaf]; }
Quoted 1 line
That's what I mean with filter out dynamically.
This is wrong. KVM isn't complaining about shuffling the order of feature words, it's complaining that code is trying to do a reverse CPUID lookup
to a feature that isn't in the reverse_cpuid table. Filtering out
checks dynamically is just hiding bugs.
Quoted 4 lines
In function ‘x86_feature_cpuid’,
inlined from ‘guest_cpuid_get_register’ atarch/x86/kvm/cpuid.h:71:33, inlined from ‘guest_cpuid_has’ atarch/x86/kvm/cpuid.h:100:8, inlined from ‘kvm_get_msr_common’ atarch/x86/kvm/x86.c:2804:8:
This corresponds to "guest_cpuid_has(vcpu, X86_FEATURE_ARCH_CAPABILITIES)", i.e. KVM is trying to query X86_FEATURE_ARCH_CAPABILITIES and is yelling that there is no reverse_cpuid entry for CPUID_7_EDX.
The problem is that 'enum cpuid_leafs' no longer matches up with the word numbers defined in cpufeatures.h, e.g. CPUID_7_EDX == 17 or so, but the entries in cpufeatures.h defined CPUID_7_EDX flags using word 18.
This patch also needs to modify NCAPINTS.
Borislav Petkov
maintainerRe: [RFC PATCH 2/3] x86/cpufeatures: Combine word 11 and 12 into new scattered features word 11
Quoted 4 lines · Sean Christopherson
This is wrong. KVM isn't complaining about shuffling the order of feature words, it's complaining that code is trying to do a reverse CPUID lookup
to a feature that isn't in thereverse_cpuidtable. Filtering out
checks dynamically is just hiding bugs.
No no, reverse_cpuid is hardcoding our feature leafs. This is wrong as we want to be able to change those. And reverse_cpuid[] should be able to handle that.
KVM is complaining because he removed one leaf. He adds it later in patch 3 as a Linux-defined leaf.
All that doesn't matter for KVM - if KVM wants to do reverse lookup, then it should handle Linux-defined leafs just fine.
Signature · 4 lines
Regards/Gruss,
Boris.
Good mailing practices for 400: avoid top-posting and trim the reply.Re: [RFC PATCH 2/3] x86/cpufeatures: Combine word 11 and 12 into new scattered features word 11
Quoted 6 lines · Borislav Petkov
Quoted 4 lines · Sean Christopherson
This is wrong. KVM isn't complaining about shuffling the order of feature words, it's complaining that code is trying to do a reverse CPUID lookup
to a feature that isn't in thereverse_cpuidtable. Filtering out
checks dynamically is just hiding bugs.
Quoted 6 lines
No no,
reverse_cpuidis hardcoding our feature leafs. This is wrong as we want to be able to change those. And reverse_cpuid[] should be able to handle that.KVM is complaining because he removed one leaf. He adds it later in patch 3 as a Linux-defined leaf.
Yes, because removing that leaf breaks 'enum cpuid_leafs'. Patch 3/3 "fixes" it by re-inserting a leaf, which causes 'enum cpuid_leafs' to align with the CPU features.
For example, this assertion also fails:
arch/x86/kernel/cpu/common.c +1 -0
@@ -823,6 +823,7 @@ void get_cpu_cap(struct cpuinfo_x86 *c) c->x86_capability[CPUID_7_0_EBX ] = ebx; c->x86_capability[CPUID_7_ECX ] = ecx; c->x86_capability[CPUID_7_EDX ] = edx; +BUILD_BUG_ON (CPUID_7_EDX !=X86_FEATURE_ARCH_CAPABILITIES /32 ); }/* Extended state features: level 0x0000000d */
In function ‘x86_feature_cpuid’,
inlined from ‘guest_cpuid_get_register’ at arch/x86/kvm/cpuid.h:71:25,
inlined from ‘guest_cpuid_has’ at arch/x86/kvm/cpuid.h:100:6,
inlined from ‘kvm_get_msr_common’ at arch/x86/kvm/x86.c:2824:8:include/linux/compiler.h:345:38: error: call to ‘__compiletime_assert_62’ declared with attribute error: BUILD_BUG_ON failed: x86_leaf >= ARRAY_SIZE(reverse_cpuid)
_compiletime_assert(condition, msg, __compiletime_assert_, __LINE__)
But this assertion passes because its word is 10, i.e. below the 11/12 words that are getting mucked with.
arch/x86/kernel/cpu/common.c +1 -0
@@ -830,6 +830,7 @@ void get_cpu_cap(struct cpuinfo_x86 *c)cpuid_count (0x0000000d ,1 , &eax, &ebx, &ecx, &edx); c->x86_capability[CPUID_D_1_EAX ] = eax; +BUILD_BUG_ON (CPUID_D_1_EAX !=X86_FEATURE_XSAVES /32 ); }/* AMD-defined flags: level 0x80000001 */
Quoted 2 lines
All that doesn't matter for KVM - if KVM wants to do reverse lookup, then it should handle Linux-defined leafs just fine.
KVM can't handle Linux-defined leafs without extra tricks, which is why I removed get_scattered_cpuid_leaf() or whatever it was called.
Borislav Petkov
maintainerRe: [RFC PATCH 2/3] x86/cpufeatures: Combine word 11 and 12 into new scattered features word 11
Quoted 1 line · Sean Christopherson
KVM can't handle Linux-defined leafs without extra tricks
and that's what I'm proposing - an extra trick.
Signature · 4 lines
Regards/Gruss,
Boris.
Good mailing practices for 400: avoid top-posting and trim the reply.Re: [RFC PATCH 2/3] x86/cpufeatures: Combine word 11 and 12 into new scattered features word 11
Quoted 4 lines · Borislav Petkov
Quoted 1 line · Sean Christopherson
KVM can't handle Linux-defined leafs without extra tricks
and that's what I'm proposing - an extra trick.
It's not a trick, it's bug suppression.
Try running a kernel built with only patches 1/2 and 2/2 applied, along with KVM's assertions removed. It'll probably boot fine since most of the affected features are option things, but Linux's feature reporting will be all kinds of screwed up.
E.g. this WARN triggers because CPUID_7_EDX is 17, not 18 as expected, and so the word at c->x86_capability[18] is never written. Run the kernel again after applying patch 3/3 and the warning magically disappears...
/* Intel-defined CPU features, CPUID level 0x00000007:0 (EDX), word 18 */
#define X86_FEATURE_SPEC_CTRL_SSBD (18*32+31) /* "" Speculative Store Bypass Disable */
arch/x86/kernel/cpu/common.c +2 -0
@@ -823,6 +823,8 @@ void get_cpu_cap(struct cpuinfo_x86 *c) c->x86_capability[CPUID_7_0_EBX ] = ebx; c->x86_capability[CPUID_7_ECX ] = ecx; c->x86_capability[CPUID_7_EDX ] = edx; + +WARN_ON ((edx &BIT (31 )) && !cpu_has (c,X86_FEATURE_SPEC_CTRL_SSBD )); }/* Extended state features: level 0x0000000d */
[ 0.085583] WARNING: CPU: 0 PID: 0 at /home/sean/go/src/kernel.org/sinux/arch/x86/kernel/cpu/common.c:827 get_cpu_cap+0x1c9/0x1d0 [ 0.085583] Modules linked in: [ 0.085585] CPU: 0 PID: 0 Comm: swapper/0 Tainted: G W 5.2.0-rc2+ #4 [ 0.085585] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 0.0.0 02/06/2015 [ 0.085586] RIP: 0010:get_cpu_cap+0x1c9/0x1d0 [ 0.085587] Code: <0f> 0b e9 7f fe ff ff 48 8b 05 d9 39 fe 00 ba 01 00 00 00 41 54 b9 [ 0.085588] RSP: 0000:ffffffff82003e80 EFLAGS: 00010246 [ 0.085588] RAX: 0000000000000000 RBX: 00000000009c4fbb RCX: 0000000000000004 [ 0.085589] RDX: 0000000084000000 RSI: 0000000000000000 RDI: ffffffff820f71e0 [ 0.085589] RBP: ffffffff820f71e0 R08: 0000000000000001 R09: 0000000000000001 [ 0.085589] R10: 000000000002c4f8 R11: 0000000000000002 R12: ffffffff82486920 [ 0.085590] R13: ffffffff8248d2e0 R14: ffff88827fff73c0 R15: 000000000000010f [ 0.085590] FS: 0000000000000000(0000) GS:ffff888277a00000(0000) knlGS:0000000000000000 [ 0.085592] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 0.085592] CR2: 00000000ffffffff CR3: 0000000005009001 CR4: 00000000000606b0 [ 0.085592] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [ 0.085593] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [ 0.085593] Call Trace: [ 0.085595] identify_cpu+0xbb/0x540 [ 0.085597] identify_boot_cpu+0xc/0x6f [ 0.085598] check_bugs+0x26/0x852 [ 0.085599] start_kernel+0x47a/0x4ab [ 0.085601] secondary_startup_64+0xa4/0xb0 [ 0.085602] ---[ end trace 33fbe952b06dbb68 ]---
Borislav Petkov
maintainerRe: [RFC PATCH 2/3] x86/cpufeatures: Combine word 11 and 12 into new scattered features word 11
Quoted 6 lines · Sean Christopherson
Quoted 4 lines · Borislav Petkov
Quoted 1 line · Sean Christopherson
KVM can't handle Linux-defined leafs without extra tricks
and that's what I'm proposing - an extra trick.
Quoted 8 lines
It's not a trick, it's bug suppression.
Try running a kernel built with only patches 1/2 and 2/2 applied, along with KVM's assertions removed. It'll probably boot fine since most of the affected features are option things, but Linux's feature reporting will be all kinds of screwed up.
E.g. this WARN triggers because
CPUID_7_EDXis 17, not 18 as expected,
We can decrement NCAPINTS and word 18 in the header. The BUILD_BUG_ONs should not fire then too.
But the easier thing is to not remove any defines in the enum cpuid_leafs thing so that the capabilities array has the proper size for after patch 2.
Signature · 4 lines
Regards/Gruss,
Boris.
Good mailing practices for 400: avoid top-posting and trim the reply.Re: [RFC PATCH 2/3] x86/cpufeatures: Combine word 11 and 12 into new scattered features word 11
Quoted 14 lines · Borislav Petkov
Quoted 13 lines · Sean Christopherson
Quoted 4 lines · Borislav Petkov
Quoted 1 line · Sean Christopherson
KVM can't handle Linux-defined leafs without extra tricks
and that's what I'm proposing - an extra trick.
It's not a trick, it's bug suppression.
Try running a kernel built with only patches 1/2 and 2/2 applied, along with KVM's assertions removed. It'll probably boot fine since most of the affected features are option things, but Linux's feature reporting will be all kinds of screwed up.
Quoted 8 lines
Quoted 1 line
E.g. this WARN triggers because
CPUID_7_EDXis 17, not 18 as expected,We can decrement NCAPINTS and word 18 in the header. The
BUILD_BUG_ONsshould not fire then too.But the easier thing is to not remove any defines in the enum
cpuid_leafsthing so that the capabilities array has the proper size for after patch 2.
Agreed, Fenghua's proposed CPUID_DUMMY is way easier.
Re: [RFC PATCH 2/3] x86/cpufeatures: Combine word 11 and 12 into new scattered features word 11
In reply to Sean ChristophersonQuoted 56 lines · Sean Christopherson
Quoted 39 lines · Borislav Petkov
+ Radim and Paolo. See upthread for context.
Quoted 7 lines · Fenghua Yu
Quoted 4 lines
Alternatively - and what I think is the better solution - would be to remove those
BUILD_BUG_ONsinx86_feature_cpuidand filter out the Linux-defined leafs dynamically. This way the array won't have holes in it.Maybe adding a dummy slot in
cpuid_leafsin patch 0002 to avoid the compilation errors?Maybe you didn't read what you're replying to: "This way the array won't have holes in it". Ontop of yours:
arch/x86/kvm/cpuid.h +4
-2@@ -47,6 +47,7 @@ static const struct cpuid_reg reverse_cpuid[] = { [CPUID_8000_0001_ECX] = {0x80000001, 0, CPUID_ECX}, [CPUID_7_0_EBX] = { 7, 0, CPUID_EBX}, [CPUID_D_1_EAX] = { 0xd, 1, CPUID_EAX}, + [CPUID_7_1_EAX] = { 7, 1, CPUID_EAX}, [CPUID_8000_0008_EBX] = {0x80000008, 0, CPUID_EBX}, [CPUID_6_EAX] = { 6, 0, CPUID_EAX}, [CPUID_8000_000A_EDX] = {0x8000000a, 0, CPUID_EDX},@@ -59,8 +60,9 @@ static __always_inline struct cpuid_reg x86_feature_cpuid(unsigned x86_feature) {unsigned x86_leaf = x86_feature /32 ; -BUILD_BUG_ON (x86_leaf >=ARRAY_SIZE (reverse_cpuid)); -BUILD_BUG_ON (reverse_cpuid[x86_leaf].function ==0 ); +if (x86_leaf ==CPUID_LNX_1 || + x86_leaf ==CPUID_LNX_4 ) +return NULL ;return reverse_cpuid[x86_leaf]; }Quoted 1 line
That's what I mean with filter out dynamically.
This is wrong. KVM isn't complaining about shuffling the order of feature words, it's complaining that code is trying to do a reverse CPUID lookup
to a feature that isn't in thereverse_cpuidtable. Filtering out
checks dynamically is just hiding bugs.Quoted 4 lines
In function ‘x86_feature_cpuid’,
inlined from ‘guest_cpuid_get_register’ atarch/x86/kvm/cpuid.h:71:33, inlined from ‘guest_cpuid_has’ atarch/x86/kvm/cpuid.h:100:8, inlined from ‘kvm_get_msr_common’ atarch/x86/kvm/x86.c:2804:8:This corresponds to "
guest_cpuid_has(vcpu, X86_FEATURE_ARCH_CAPABILITIES)", i.e. KVM is trying to queryX86_FEATURE_ARCH_CAPABILITIESand is yelling that there is noreverse_cpuidentry forCPUID_7_EDX.
Quoted 5 lines
The problem is that 'enum
cpuid_leafs' no longer matches up with the word numbers defined incpufeatures.h, e.g.CPUID_7_EDX== 17 or so, but the entries incpufeatures.hdefinedCPUID_7_EDXflags using word 18.This patch also needs to modify NCAPINTS.
Changing NCAPINTS is heavy lifting to only solve this biset issue. After applying patch 0003, the word 12 hole generated in patch 0002 is filled in and no issue reported. If changing NCAPINTS in patch 0002, it needs to be changed back again after applying 0003.
How about add this patch in patch 0002? I posted this patch in this thread just now and post it here again:
arch/x86/include/asm/cpufeature.h +1 -0
@@ -23,6 +23,7 @@ enum cpuid_leafsCPUID_7_0_EBX ,CPUID_D_1_EAX ,CPUID_LNX_4 , +CPUID_DUMMY ,CPUID_8000_0008_EBX ,CPUID_6_EAX , CPUID_8000_000A_EDX,
Adding this small patch into patch 0002 will solve the build errors without changing the build checks.
Thanks.
-Fenghua
Borislav Petkov
maintainerRe: [RFC PATCH 2/3] x86/cpufeatures: Combine word 11 and 12 into new scattered features word 11
Quoted 2 lines · Fenghua Yu
Adding this small patch into patch 0002 will solve the build errors without changing the build checks.
There's no need for that if you remove the BUILD_BUG_ON checks first.
Signature · 4 lines
Regards/Gruss,
Boris.
Good mailing practices for 400: avoid top-posting and trim the reply.Re: [RFC PATCH 2/3] x86/cpufeatures: Combine word 11 and 12 into new scattered features word 11
Quoted 5 lines · Borislav Petkov
Quoted 2 lines · Fenghua Yu
Adding this small patch into patch 0002 will solve the build errors without changing the build checks.
There's no need for that if you remove the
BUILD_BUG_ONchecks first.
But without this small patch, CPUID_7_EDX is 17 instead of NCAPINTS(19)-1=18 in patch 0002. Of course CPUID_7_EDX is 18 correctly evetually after applying patch 0003 which add the word 12 back.
KVM reports this CPUID_7_EDX is not 18 in bisect (as Sena also pointed out). I think even native kernel should have some check to report this error (similar to other checks of NCAPINTS).
Thanks.
-Fenghua
Borislav Petkov
maintainerRe: [RFC PATCH 2/3] x86/cpufeatures: Combine word 11 and 12 into new scattered features word 11
Quoted 3 lines · Fenghua Yu
But without this small patch,
CPUID_7_EDXis 17 instead of NCAPINTS(19)-1=18 in patch 0002. Of courseCPUID_7_EDXis 18 correctly evetually after applying patch 0003 which add the word 12 back.
Ok, then I guess the easiest should be if not remove defines from cpuid_leafs but rename them, no?
Signature · 4 lines
Regards/Gruss,
Boris.
Good mailing practices for 400: avoid top-posting and trim the reply.Re: [RFC PATCH 2/3] x86/cpufeatures: Combine word 11 and 12 into new scattered features word 11
Quoted 7 lines · Borislav Petkov
Quoted 3 lines · Fenghua Yu
But without this small patch,
CPUID_7_EDXis 17 instead of NCAPINTS(19)-1=18 in patch 0002. Of courseCPUID_7_EDXis 18 correctly evetually after applying patch 0003 which add the word 12 back.Ok, then I guess the easiest should be if not remove defines from
cpuid_leafsbut rename them, no?
So I will re-name the defines in cpuid_leafs to CPUID_LNX_4 and CPUID_DUMMY to match status of word 11 and word 12.
Hopefully that's ok?
Thanks.
-Fenghua
RFC PATCH 3/3 x86/cpufeatures: Enumerate new AVX512 BFLOAT16 instructions
In reply to Fenghua YuAVX512 Vector Neural Network Instructions (VNNI) in Intel Deep Learning Boost support BFLOAT16 format (BF16). BF16 is a short version of FP32 and has several advantages over FP16. BF16 offers more than enough range for deep learning training tasks and doesn't need to handle hardware exception as this is a performance optimization. FP32 accumulation after the multiply is essential to achieve sufficient numerical behavior on an application level.
AVX512 BFLOAT16 instructions can be enumerated by:
CPUID.7.1:EAX[bit 5] AVX512_BF16
Add word 12 to hold features in CPUID.7.1:EAX including AVX512_BF16.
Detailed information of the CPUID bit and AVX512 BFLOAT16 instructions can be found in the latest Intel Architecture Instruction Set Extensions and Future Features Programming Reference.
--- arch/x86/include/asm/cpufeature.h | 1 + arch/x86/include/asm/cpufeatures.h | 3 +++ arch/x86/kernel/cpu/common.c | 3 +++ arch/x86/kernel/cpu/cpuid-deps.c | 1 + 4 files changed, 8 insertions(+)
arch/x86/include/asm/cpufeature.h +1 -0
@@ -23,6 +23,7 @@ enum cpuid_leafsCPUID_7_0_EBX ,CPUID_D_1_EAX ,CPUID_LNX_4 , +CPUID_7_1_EAX ,CPUID_8000_0008_EBX ,CPUID_6_EAX , CPUID_8000_000A_EDX,
arch/x86/include/asm/cpufeatures.h +3 -0
@@ -280,6 +280,9 @@#define X86_FEATURE_CQM_MBM_TOTAL (11*32+ 2)/* LLC Total MBM monitoring */ #define X86_FEATURE_CQM_MBM_LOCAL (11*32+ 3)/* LLC Local MBM monitoring */ +/* Intel-defined CPU features, CPUID level 0x00000007:1 (EAX), word 12 */ +#define X86_FEATURE_AVX512_BF16 (12*32+ 5)/* AVX512 BFLOAT16 instructions */ +/* AMD-defined CPU features, CPUID level 0x80000008 (EBX), word 13 */ #define X86_FEATURE_CLZERO (13*32+ 0)/* CLZERO instruction */ #define X86_FEATURE_IRPERF (13*32+ 1)/* Instructions Retired Count */
arch/x86/kernel/cpu/common.c +3 -0
@@ -823,6 +823,9 @@ void get_cpu_cap(struct cpuinfo_x86 *c) c->x86_capability[CPUID_7_0_EBX ] = ebx; c->x86_capability[CPUID_7_ECX ] = ecx; c->x86_capability[CPUID_7_EDX ] = edx; + +cpuid_count (0x00000007 ,1 , &eax, &ebx, &ecx, &edx); + c->x86_capability[CPUID_7_1_EAX ] = eax; }/* Extended state features: level 0x0000000d */
arch/x86/kernel/cpu/cpuid-deps.c +1 -0
@@ -62,6 +62,7 @@ static const struct cpuid_dep cpuid_deps[] = { { X86_FEATURE_CQM_OCCUP_LLC, X86_FEATURE_CQM_LLC }, { X86_FEATURE_CQM_MBM_TOTAL, X86_FEATURE_CQM_LLC }, { X86_FEATURE_CQM_MBM_LOCAL, X86_FEATURE_CQM_LLC }, + { X86_FEATURE_AVX512_BF16, X86_FEATURE_AVX512VL }, {} };
Signature · 1 line
2.19.1