x86/paravirt: useless assignment instructions cause Unixbench full core performance degradation
Subsystems (4)
PATCH 1/1 x86/paravirt: useless assignment instructions cause Unixbench full core performance degradation
The instructions assigned to the parameters of
the vcpu_is_preempted function in the X86 architecture physical machine are redundant instructions, which cause the multi-core performance of Unixbench to drop by about 300 to 500 points. The C function is as follows:
static bool vcpu_is_preempted(long vcpu);
The parameter assignments in the function osq_lock that call the function vcpu_is_preempted are as follows:
mov 0x14(%rax),%edi
sub $0x1,%edi
The above instructions are unnecessary
in the X86 Native operating environment, causing high cache-misses and degrading performance.
This patch implements the replacement of the above instructions with the nop instruction at system startup.
When the assignment C code changes,
the above assignment instructions may also change accordingly. In order to flexibly replace the instructions generated by the C code, this patch defines the macro ALTERNATIVE_C_CODE
from the ALTERNATIVE macro. Different from the first parameter of the ALTERNATIVE macro, the ALTERNATIVE_C_CODE macro is the first parameter is a C code statement and cannot contain a return statement. To use the macro, for example,
replace the instructions generated by the
C code 'cpu = node->cpu - 1;' with the nop instruction:
OSQ_ALTERNATIVE_C_CODE(cpu = node->cpu - 1, "nop", 0);
The patch effect is as follows two machines,
Unixbench runs with full core score:
1. Machine configuration:
Intel(R) Xeon(R) Silver 4210 CPU @ 2.20GHz
CPU core: 40
Memory: 256G
OS Kernel: 5.19-rc3
Before using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 944608003.9 80943.3
Double-Precision Whetstone 55.0 212045.8 38553.8
Execl Throughput 43.0 42772.3 9947.1
File Copy 1024 bufsize 2000 maxblocks 3960.0 462656.6 1168.3 File Copy 256 bufsize 500 maxblocks 1655.0 120043.6 725.3 File Copy 4096 bufsize 8000 maxblocks 5800.0 1544525.5 2663.0
Pipe Throughput 12440.0 47277698.5 38004.6
Pipe-based Context Switching 4000.0 1894556.8 4736.4
Process Creation 126.0 86077.0 6831.5
Shell Scripts (1 concurrent) 42.4 70236.3 16565.2
Shell Scripts (8 concurrent) 6.0 8978.1 14963.4
System Call Overhead 15000.0 4691260.0 3127.5
========
System Benchmarks Index Score 7980.9
After using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 2253984916.9 193143.5
Double-Precision Whetstone 55.0 438940.3 79807.3
Execl Throughput 43.0 10720.3 2493.1
File Copy 1024 bufsize 2000 maxblocks 3960.0 312233.0 788.5 File Copy 256 bufsize 500 maxblocks 1655.0 80050.9 483.7 File Copy 4096 bufsize 8000 maxblocks 5800.0 1036101.7 1786.4
Pipe Throughput 12440.0 117700315.3 94614.4
Pipe-based Context Switching 4000.0 8421909.8 21054.8
Process Creation 126.0 36742.0 2916.0
Shell Scripts (1 concurrent) 42.4 52846.2 12463.7
Shell Scripts (8 concurrent) 6.0 7058.1 11763.6
System Call Overhead 15000.0 6791548.2 4527.7
========
System Benchmarks Index Score 8260.6
2. Machine configuration:
Hygon C86 7185 32-core Processor
CPU core: 128
Memory: 256G
OS Kernel: 5.19-rc3
Before using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 2256283941.6 193340.5
Double-Precision Whetstone 55.0 439577.3 79923.2
Execl Throughput 43.0 10013.6 2328.7
File Copy 1024 bufsize 2000 maxblocks 3960.0 278121.5 702.3 File Copy 256 bufsize 500 maxblocks 1655.0 71835.5 434.1 File Copy 4096 bufsize 8000 maxblocks 5800.0 905654.2 1561.5
Pipe Throughput 12440.0 117715166.2 94626.3
Pipe-based Context Switching 4000.0 7731331.7 19328.3
Process Creation 126.0 30157.8 2393.5
Shell Scripts (1 concurrent) 42.4 48670.8 11479.0
Shell Scripts (8 concurrent) 6.0 6595.6 10992.7
System Call Overhead 15000.0 6766475.9 4511.0
========
System Benchmarks Index Score 7688.7
After using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 2253984916.9 193143.5
Double-Precision Whetstone 55.0 438940.3 79807.3
Execl Throughput 43.0 10720.3 2493.1
File Copy 1024 bufsize 2000 maxblocks 3960.0 312233.0 788.5 File Copy 256 bufsize 500 maxblocks 1655.0 80050.9 483.7 File Copy 4096 bufsize 8000 maxblocks 5800.0 1036101.7 1786.4
Pipe Throughput 12440.0 117700315.3 94614.4
Pipe-based Context Switching 4000.0 8421909.8 21054.8
Process Creation 126.0 36742.0 2916.0
Shell Scripts (1 concurrent) 42.4 52846.2 12463.7
Shell Scripts (8 concurrent) 6.0 7058.1 11763.6
System Call Overhead 15000.0 6791548.2 4527.7
========
System Benchmarks Index Score 8260.6
--- arch/x86/include/asm/alternative.h | 15 +++++++++++++++ arch/x86/kernel/alternative.c | 11 +++++++++++ include/linux/osq_lock.h | 7 +++++++ kernel/locking/osq_lock.c | 6 +++++- 4 files changed, 38 insertions(+), 1 deletion(-)
arch/x86/include/asm/alternative.h +15 -0
@@ -167,6 +167,21 @@ static inline int alternatives_text_reserved(void *start, void *end) ALTINSTR_REPLACEMENT(newinstr, 1) \".popsection\n" +/* alternative c code primitive: */ +#define ALTERNATIVE_C_CODE (oldinstr, newinstr, feature) \ +do { \ + asm volatile("661:\n\t"); \ + oldinstr; \ + asm volatile("\n662:\n" \ + alt_end_marker ":\n" \ + ".pushsection .altinstructions,\"a\"\n" \ + ALTINSTR_ENTRY(feature, 1) \ + ".popsection\n" \ + ".pushsection .altinstr_replacement, \"ax\"\n" \ + ALTINSTR_REPLACEMENT(newinstr, 1) \ + ".popsection\n"); \ +} while (0) +#define ALTERNATIVE_2 (oldinstr, newinstr1, feature1, newinstr2, feature2)\ OLDINSTR_2(oldinstr, 1, 2) \ ".pushsection .altinstructions,\"a\"\n" \
arch/x86/kernel/alternative.c +11 -0
@@ -251,6 +251,8 @@ static void __init_or_module noinline optimize_nops(u8 *instr, size_t len) } } +extern bool pv_is_native_vcpu_is_preempted (void ); +/* * Replace instructions with better alternatives for this CPU type. This runs * before SMP is initialized to avoid SMP problems with self modifying code.
@@ -285,6 +287,15 @@ void __init_or_module noinline apply_alternatives(struct alt_instr *start, instr = (u8 *)&a->instr_offset + a->instr_offset; replacement = (u8 *)&a->repl_offset + a->repl_offset; + +if (*replacement ==0x90 && a->replacementlen ==1 ) { +#if defined(CONFIG_PARAVIRT_SPINLOCKS ) +if (pv_is_native_vcpu_is_preempted ()) +add_nops (instr, a->instrlen); +#endif +continue ; + } +BUG_ON (a->instrlen >sizeof (insn_buff));BUG_ON (feature >= (NCAPINTS +NBUGINTS ) *32 );
include/linux/osq_lock.h +7 -0
@@ -38,4 +38,11 @@ static inline bool osq_is_locked(struct optimistic_spin_queue *lock) return atomic_read(&lock->tail) != OSQ_UNLOCKED_VAL; } +#ifdef ALTERNATIVE_C_CODE +#define OSQ_ALTERNATIVE_C_CODE (oldinstr, newinstr, feature) \ + ALTERNATIVE_C_CODE(oldinstr, newinstr, feature) +#else +#define OSQ_ALTERNATIVE_C_CODE (oldinstr, newinstr, feature) oldinstr +#endif +#endif
kernel/locking/osq_lock.c +5 -1
@@ -24,7 +24,11 @@ static inline int encode_cpu(int cpu_nr)static inline int node_cpu (struct optimistic_spin_node *node) { -return node->cpu -1 ; +int cpu =0 ; + +OSQ_ALTERNATIVE_C_CODE (cpu = node->cpu -1 ,"nop" ,0 ); + +return cpu; }static inline struct optimistic_spin_node *decode_cpu (int encoded_cpu_val)
Signature · 1 line
2.20.1
Waiman Long
reviewerQuoted 168 lines · Guo Hui
The instructions assigned to the parameters of
thevcpu_is_preemptedfunction in the X86 architecture physical machine are redundant instructions, which cause the multi-core performance of Unixbench to drop by about 300 to 500 points. The C function is as follows:
static boolvcpu_is_preempted(long vcpu);The parameter assignments in the function
osq_lockthat call the functionvcpu_is_preemptedare as follows:
mov 0x14(%rax),%edi
sub $0x1,%ediThe above instructions are unnecessary
in the X86 Native operating environment, causing high cache-misses and degrading performance.This patch implements the replacement of the above instructions with the nop instruction at system startup.
When the assignment C code changes,
the above assignment instructions may also change accordingly. In order to flexibly replace the instructions generated by the C code, this patch defines the macroALTERNATIVE_C_CODE
from the ALTERNATIVE macro. Different from the first parameter of the ALTERNATIVE macro, theALTERNATIVE_C_CODEmacro is the first parameter is a C code statement and cannot contain a return statement. To use the macro, for example,
replace the instructions generated by the
C code 'cpu =node->cpu- 1;' with the nop instruction:
OSQ_ALTERNATIVE_C_CODE(cpu = node->cpu - 1, "nop", 0);The patch effect is as follows two machines,
Unixbench runs with full core score:1. Machine configuration:
Intel(R) Xeon(R) Silver 4210 CPU @ 2.20GHz
CPU core: 40
Memory: 256G
OS Kernel: 5.19-rc3Before using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 944608003.9 80943.3
Double-Precision Whetstone 55.0 212045.8 38553.8
Execl Throughput 43.0 42772.3 9947.1File Copy 1024 bufsize 2000 maxblocks 3960.0 462656.6 1168.3 File Copy 256 bufsize 500 maxblocks 1655.0 120043.6 725.3 File Copy 4096 bufsize 8000 maxblocks 5800.0 1544525.5 2663.0Pipe Throughput 12440.0 47277698.5 38004.6
Pipe-based Context Switching 4000.0 1894556.8 4736.4
Process Creation 126.0 86077.0 6831.5
Shell Scripts (1 concurrent) 42.4 70236.3 16565.2
Shell Scripts (8 concurrent) 6.0 8978.1 14963.4
System Call Overhead 15000.0 4691260.0 3127.5========System Benchmarks Index Score 7980.9
After using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 2253984916.9 193143.5
Double-Precision Whetstone 55.0 438940.3 79807.3
Execl Throughput 43.0 10720.3 2493.1File Copy 1024 bufsize 2000 maxblocks 3960.0 312233.0 788.5 File Copy 256 bufsize 500 maxblocks 1655.0 80050.9 483.7 File Copy 4096 bufsize 8000 maxblocks 5800.0 1036101.7 1786.4Pipe Throughput 12440.0 117700315.3 94614.4
Pipe-based Context Switching 4000.0 8421909.8 21054.8
Process Creation 126.0 36742.0 2916.0
Shell Scripts (1 concurrent) 42.4 52846.2 12463.7
Shell Scripts (8 concurrent) 6.0 7058.1 11763.6
System Call Overhead 15000.0 6791548.2 4527.7========System Benchmarks Index Score 8260.6
2. Machine configuration:
Hygon C86 7185 32-core Processor
CPU core: 128
Memory: 256G
OS Kernel: 5.19-rc3Before using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 2256283941.6 193340.5
Double-Precision Whetstone 55.0 439577.3 79923.2
Execl Throughput 43.0 10013.6 2328.7File Copy 1024 bufsize 2000 maxblocks 3960.0 278121.5 702.3 File Copy 256 bufsize 500 maxblocks 1655.0 71835.5 434.1 File Copy 4096 bufsize 8000 maxblocks 5800.0 905654.2 1561.5Pipe Throughput 12440.0 117715166.2 94626.3
Pipe-based Context Switching 4000.0 7731331.7 19328.3
Process Creation 126.0 30157.8 2393.5
Shell Scripts (1 concurrent) 42.4 48670.8 11479.0
Shell Scripts (8 concurrent) 6.0 6595.6 10992.7
System Call Overhead 15000.0 6766475.9 4511.0========System Benchmarks Index Score 7688.7
After using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 2253984916.9 193143.5
Double-Precision Whetstone 55.0 438940.3 79807.3
Execl Throughput 43.0 10720.3 2493.1File Copy 1024 bufsize 2000 maxblocks 3960.0 312233.0 788.5 File Copy 256 bufsize 500 maxblocks 1655.0 80050.9 483.7 File Copy 4096 bufsize 8000 maxblocks 5800.0 1036101.7 1786.4Pipe Throughput 12440.0 117700315.3 94614.4
Pipe-based Context Switching 4000.0 8421909.8 21054.8
Process Creation 126.0 36742.0 2916.0
Shell Scripts (1 concurrent) 42.4 52846.2 12463.7
Shell Scripts (8 concurrent) 6.0 7058.1 11763.6
System Call Overhead 15000.0 6791548.2 4527.7========System Benchmarks Index Score 8260.6
Signed-off-by Guo Hui <[email protected]>--- arch/x86/include/asm/alternative.h | 15 +++++++++++++++ arch/x86/kernel/alternative.c | 11 +++++++++++ include/linux/osq_lock.h | 7 +++++++ kernel/locking/osq_lock.c | 6 +++++- 4 files changed, 38 insertions(+), 1 deletion(-)arch/x86/include/asm/alternative.h +15
-0@@ -167,6 +167,21 @@ static inline int alternatives_text_reserved(void *start, void *end) ALTINSTR_REPLACEMENT(newinstr, 1) \".popsection\n" +/* alternative c code primitive: */ +#define ALTERNATIVE_C_CODE (oldinstr, newinstr, feature) \ +do { \ + asm volatile("661:\n\t"); \ + oldinstr; \ + asm volatile("\n662:\n" \ + alt_end_marker ":\n" \ + ".pushsection .altinstructions,\"a\"\n" \ + ALTINSTR_ENTRY(feature, 1) \ + ".popsection\n" \ + ".pushsection .altinstr_replacement, \"ax\"\n" \ + ALTINSTR_REPLACEMENT(newinstr, 1) \ + ".popsection\n"); \ +} while (0) +#define ALTERNATIVE_2 (oldinstr, newinstr1, feature1, newinstr2, feature2)\ OLDINSTR_2(oldinstr, 1, 2) \ ".pushsection .altinstructions,\"a\"\n" \arch/x86/kernel/alternative.c +2
-0@@ -251,6 +251,8 @@ static void __init_or_module noinline optimize_nops(u8 *instr, size_t len) } } +extern bool pv_is_native_vcpu_is_preempted (void ); +/* * Replace instructions with better alternatives for this CPU type. This runs * before SMP is initialized to avoid SMP problems with self modifying code. @@ -285,6 +287,15 @@ void __init_or_module noinline apply_alternatives(struct alt_instr *start, instr = (u8 *)&a->instr_offset + a->instr_offset;
arch/x86/kernel/alternative.c void __init_or_module noinline apply_alternatives(struct alt_instr *start, · line 298
Quoted 10 lines
arch/x86/kernel/alternative.c +9
-0replacement = (u8 *)&a->repl_offset + a->repl_offset; + +if (*replacement ==0x90 && a->replacementlen ==1 ) { +#if defined(CONFIG_PARAVIRT_SPINLOCKS ) +if (pv_is_native_vcpu_is_preempted ()) +add_nops (instr, a->instrlen); +#endif +continue ; + } +
This is hacky and it may incorrectly affect other alternatives that patches thing to nop.
Quoted 24 lines
arch/x86/kernel/alternative.c +0
-0BUG_ON (a->instrlen >sizeof (insn_buff));BUG_ON (feature >= (NCAPINTS +NBUGINTS ) *32 );include/linux/osq_lock.h +7
-0@@ -38,4 +38,11 @@ static inline bool osq_is_locked(struct optimistic_spin_queue *lock) return atomic_read(&lock->tail) != OSQ_UNLOCKED_VAL; } +#ifdef ALTERNATIVE_C_CODE +#define OSQ_ALTERNATIVE_C_CODE (oldinstr, newinstr, feature) \ + ALTERNATIVE_C_CODE(oldinstr, newinstr, feature) +#else +#define OSQ_ALTERNATIVE_C_CODE (oldinstr, newinstr, feature) oldinstr +#endif +#endif kernel/locking/osq_lock.c +0
-0
kernel/locking/osq_lock.c static inline int encode_cpu(int cpu_nr) · line 32
Quoted 10 lines
kernel/locking/osq_lock.c +5
-1@@ -24,7 +24,11 @@ static inline int encode_cpu(int cpu_nr)static inline int node_cpu (struct optimistic_spin_node *node) { -return node->cpu -1 ; +int cpu =0 ; + +OSQ_ALTERNATIVE_C_CODE (cpu = node->cpu -1 ,"nop" ,0 ); + +return cpu; }
Why don't you use static_key to control the alternative and add a late_initcall() to set it. I think that will be simpler and you can throw away ALTERNATIVE_C_CODE().
Cheers,
Longman
PATCH v2 1/1 x86/paravirt: useless assignment instructions cause Unixbench full core performance degradation
The instructions assigned to the vcpu_is_preempted function parameter in the X86 architecture physical machine are redundant instructions, causing the multi-core performance of Unixbench to drop by about 4% to 5%. The C function is as follows:
static bool vcpu_is_preempted(long vcpu);
The parameter 'vcpu' in the function osq_lock
that calls the function vcpu_is_preempted is assigned as follows:
The C code is in the function node_cpu:
cpu = node->cpu - 1;
The instructions corresponding to the C code are:
mov 0x14(%rax),%edi
sub $0x1,%edi
The above instructions are unnecessary
in the X86 Native operating environment, causing high cache-misses and degrading performance.
This patch uses static_key to not execute this instruction in the Native runtime environment.
The code is as follows:
DEFINE_STATIC_KEY_FALSE(preemted_key);
static inline int node_cpu(struct optimistic_spin_node *node)
{
int cpu = 0;
if (!static_branch_unlikely(&preemted_key))
cpu = node->cpu - 1;
return cpu;
}The patch effect is as follows two machines,
Unixbench runs with full core score:
1. Machine configuration:
Intel(R) Xeon(R) Silver 4210 CPU @ 2.20GHz
CPU core: 40
Memory: 256G
OS Kernel: 5.19-rc3
Before using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 948326591.2 81261.9 Double-Precision Whetstone 55.0 211986.3 38543.0 Execl Throughput 43.0 43453.2 10105.4 File Copy 1024 bufsize 2000 maxblocks 3960.0 438936.2 1108.4 File Copy 256 bufsize 500 maxblocks 1655.0 118197.4 714.2 File Copy 4096 bufsize 8000 maxblocks 5800.0 1534674.7 2646.0
Pipe Throughput 12440.0 46482107.6 37365.0
Pipe-based Context Switching 4000.0 1915094.2 4787.7
Process Creation 126.0 85442.2 6781.1
Shell Scripts (1 concurrent) 42.4 69400.7 16368.1
Shell Scripts (8 concurrent) 6.0 8877.2 14795.3
System Call Overhead 15000.0 4714906.1 3143.3
========
System Benchmarks Index Score 7923.3
After using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 947032915.5 81151.1 Double-Precision Whetstone 55.0 211971.2 38540.2 Execl Throughput 43.0 45054.8 10477.9 File Copy 1024 bufsize 2000 maxblocks 3960.0 515024.9 1300.6 File Copy 256 bufsize 500 maxblocks 1655.0 146354.6 884.3 File Copy 4096 bufsize 8000 maxblocks 5800.0 1679995.9 2896.5
Pipe Throughput 12440.0 46466394.2 37352.4
Pipe-based Context Switching 4000.0 1898221.4 4745.6
Process Creation 126.0 85653.1 6797.9
Shell Scripts (1 concurrent) 42.4 69437.3 16376.7
Shell Scripts (8 concurrent) 6.0 8898.9 14831.4
System Call Overhead 15000.0 4658746.7 3105.8
========
System Benchmarks Index Score 8248.8
2. Machine configuration:
Hygon C86 7185 32-core Processor
CPU core: 128
Memory: 256G
OS Kernel: 5.19-rc3
Before using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 2256644068.3 193371.4
Double-Precision Whetstone 55.0 438969.9 79812.7
Execl Throughput 43.0 10108.6 2350.8
File Copy 1024 bufsize 2000 maxblocks 3960.0 275892.8 696.7 File Copy 256 bufsize 500 maxblocks 1655.0 72082.7 435.5 File Copy 4096 bufsize 8000 maxblocks 5800.0 925043.4 1594.9
Pipe Throughput 12440.0 118905512.5 95583.2
Pipe-based Context Switching 4000.0 7820945.7 19552.4
Process Creation 126.0 31233.3 2478.8
Shell Scripts (1 concurrent) 42.4 49042.8 11566.7
Shell Scripts (8 concurrent) 6.0 6656.0 11093.3
System Call Overhead 15000.0 6816047.5 4544.0
========
System Benchmarks Index Score 7756.6
After using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 2252272929.4 192996.8
Double-Precision Whetstone 55.0 451847.2 82154.0
Execl Throughput 43.0 10595.1 2464.0
File Copy 1024 bufsize 2000 maxblocks 3960.0 301279.3 760.8 File Copy 256 bufsize 500 maxblocks 1655.0 79291.3 479.1 File Copy 4096 bufsize 8000 maxblocks 5800.0 1039755.2 1792.7
Pipe Throughput 12440.0 118701468.1 95419.2
Pipe-based Context Switching 4000.0 8073453.3 20183.6
Process Creation 126.0 33440.9 2654.0
Shell Scripts (1 concurrent) 42.4 52722.6 12434.6
Shell Scripts (8 concurrent) 6.0 7050.4 11750.6
System Call Overhead 15000.0 6834371.5 4556.2
========
System Benchmarks Index Score 8157.8
--- arch/x86/kernel/paravirt-spinlocks.c | 4 ++++ kernel/locking/osq_lock.c | 9 ++++++++- 2 files changed, 12 insertions(+), 1 deletion(-)
arch/x86/kernel/paravirt-spinlocks.c +4 -0
@@ -33,6 +33,8 @@ bool pv_is_native_vcpu_is_preempted(void) __raw_callee_save___native_vcpu_is_preempted; } +DECLARE_STATIC_KEY_FALSE (preemted_key); +void __init paravirt_set_cap (void ) {if (!pv_is_native_spin_unlock ())
@@ -40,4 +42,6 @@ void __init paravirt_set_cap(void)if (!pv_is_native_vcpu_is_preempted ())setup_force_cpu_cap (X86_FEATURE_VCPUPREEMPT ); +else +static_branch_enable (&preemted_key); }
kernel/locking/osq_lock.c +8 -1
@@ -22,9 +22,16 @@ static inline int encode_cpu(int cpu_nr)return cpu_nr +1 ; } +DEFINE_STATIC_KEY_FALSE (preemted_key); +static inline int node_cpu (struct optimistic_spin_node *node) { -return node->cpu -1 ; +int cpu =0 ; + +if (!static_branch_unlikely (&preemted_key)) + cpu = node->cpu -1 ; + +return cpu; }static inline struct optimistic_spin_node *decode_cpu (int encoded_cpu_val)
Signature · 1 line
2.20.1
Waiman Long
reviewerQuoted 158 lines · Guo Hui
The instructions assigned to the
vcpu_is_preemptedfunction parameter in the X86 architecture physical machine are redundant instructions, causing the multi-core performance of Unixbench to drop by about 4% to 5%. The C function is as follows:
static boolvcpu_is_preempted(long vcpu);The parameter 'vcpu' in the function
osq_lock
that calls the functionvcpu_is_preemptedis assigned as follows:The C code is in the function
node_cpu:
cpu =node->cpu- 1;The instructions corresponding to the C code are:
mov 0x14(%rax),%edi
sub $0x1,%ediThe above instructions are unnecessary
in the X86 Native operating environment, causing high cache-misses and degrading performance.This patch uses
static_keyto not execute this instruction in the Native runtime environment.The code is as follows:
DEFINE_STATIC_KEY_FALSE(preemted_key);static inline int
node_cpu(struct optimistic_spin_node *node){ int cpu = 0; if (!static_branch_unlikely(&preemted_key)) cpu = node->cpu - 1; return cpu; }The patch effect is as follows two machines,
Unixbench runs with full core score:1. Machine configuration:
Intel(R) Xeon(R) Silver 4210 CPU @ 2.20GHz
CPU core: 40
Memory: 256G
OS Kernel: 5.19-rc3Before using the patch:
System Benchmarks Index Values BASELINE RESULT INDEXDhrystone 2 using register variables 116700.0 948326591.2 81261.9 Double-Precision Whetstone 55.0 211986.3 38543.0 Execl Throughput 43.0 43453.2 10105.4 File Copy 1024 bufsize 2000 maxblocks 3960.0 438936.2 1108.4 File Copy 256 bufsize 500 maxblocks 1655.0 118197.4 714.2 File Copy 4096 bufsize 8000 maxblocks 5800.0 1534674.7 2646.0Pipe Throughput 12440.0 46482107.6 37365.0
Pipe-based Context Switching 4000.0 1915094.2 4787.7
Process Creation 126.0 85442.2 6781.1
Shell Scripts (1 concurrent) 42.4 69400.7 16368.1
Shell Scripts (8 concurrent) 6.0 8877.2 14795.3
System Call Overhead 15000.0 4714906.1 3143.3========System Benchmarks Index Score 7923.3
After using the patch:
System Benchmarks Index Values BASELINE RESULT INDEXDhrystone 2 using register variables 116700.0 947032915.5 81151.1 Double-Precision Whetstone 55.0 211971.2 38540.2 Execl Throughput 43.0 45054.8 10477.9 File Copy 1024 bufsize 2000 maxblocks 3960.0 515024.9 1300.6 File Copy 256 bufsize 500 maxblocks 1655.0 146354.6 884.3 File Copy 4096 bufsize 8000 maxblocks 5800.0 1679995.9 2896.5Pipe Throughput 12440.0 46466394.2 37352.4
Pipe-based Context Switching 4000.0 1898221.4 4745.6
Process Creation 126.0 85653.1 6797.9
Shell Scripts (1 concurrent) 42.4 69437.3 16376.7
Shell Scripts (8 concurrent) 6.0 8898.9 14831.4
System Call Overhead 15000.0 4658746.7 3105.8========System Benchmarks Index Score 8248.8
2. Machine configuration:
Hygon C86 7185 32-core Processor
CPU core: 128
Memory: 256G
OS Kernel: 5.19-rc3Before using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 2256644068.3 193371.4
Double-Precision Whetstone 55.0 438969.9 79812.7
Execl Throughput 43.0 10108.6 2350.8File Copy 1024 bufsize 2000 maxblocks 3960.0 275892.8 696.7 File Copy 256 bufsize 500 maxblocks 1655.0 72082.7 435.5 File Copy 4096 bufsize 8000 maxblocks 5800.0 925043.4 1594.9Pipe Throughput 12440.0 118905512.5 95583.2
Pipe-based Context Switching 4000.0 7820945.7 19552.4
Process Creation 126.0 31233.3 2478.8
Shell Scripts (1 concurrent) 42.4 49042.8 11566.7
Shell Scripts (8 concurrent) 6.0 6656.0 11093.3
System Call Overhead 15000.0 6816047.5 4544.0========System Benchmarks Index Score 7756.6
After using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 2252272929.4 192996.8
Double-Precision Whetstone 55.0 451847.2 82154.0
Execl Throughput 43.0 10595.1 2464.0File Copy 1024 bufsize 2000 maxblocks 3960.0 301279.3 760.8 File Copy 256 bufsize 500 maxblocks 1655.0 79291.3 479.1 File Copy 4096 bufsize 8000 maxblocks 5800.0 1039755.2 1792.7Pipe Throughput 12440.0 118701468.1 95419.2
Pipe-based Context Switching 4000.0 8073453.3 20183.6
Process Creation 126.0 33440.9 2654.0
Shell Scripts (1 concurrent) 42.4 52722.6 12434.6
Shell Scripts (8 concurrent) 6.0 7050.4 11750.6
System Call Overhead 15000.0 6834371.5 4556.2========System Benchmarks Index Score 8157.8
Signed-off-by Guo Hui <[email protected]>--- arch/x86/kernel/paravirt-spinlocks.c | 4 ++++ kernel/locking/osq_lock.c | 9 ++++++++- 2 files changed, 12 insertions(+), 1 deletion(-)arch/x86/kernel/paravirt-spinlocks.c +4
-0@@ -33,6 +33,8 @@ bool pv_is_native_vcpu_is_preempted(void) __raw_callee_save___native_vcpu_is_preempted; } +DECLARE_STATIC_KEY_FALSE (preemted_key); +void __init paravirt_set_cap (void ) {if (!pv_is_native_spin_unlock ())@@ -40,4 +42,6 @@ void __init paravirt_set_cap(void)if (!pv_is_native_vcpu_is_preempted ())setup_force_cpu_cap (X86_FEATURE_VCPUPREEMPT ); +else +static_branch_enable (&preemted_key); }kernel/locking/osq_lock.c +2
-0@@ -22,9 +22,16 @@ static inline int encode_cpu(int cpu_nr)return cpu_nr +1 ; } +DEFINE_STATIC_KEY_FALSE (preemted_key); +static inline int node_cpu (struct optimistic_spin_node *node) {
kernel/locking/osq_lock.c static inline int encode_cpu(int cpu_nr) · line 37
Quoted 10 lines
kernel/locking/osq_lock.c +6
-1-return node->cpu -1 ; +int cpu =0 ; + +if (!static_branch_unlikely (&preemted_key)) + cpu = node->cpu -1 ; + +return cpu; }static inline struct optimistic_spin_node *decode_cpu (int encoded_cpu_val)
This looks much better now.
However vcpu_is_preempted() is not just used by x86, it is also used in arm64, powerpc and s390. If you want to default disable node_cpu(), you have to either make the necessary changes in all those places or you have to change it a way that won't affect other arches while disabling it for the x86 case.
Cheers,
Longman
_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization
A message that is not in the archive (sent off-list or not yet received): [email protected]
Waiman Long
reviewerQuoted 13 lines · Guo Hui
Thank you very much Longman, my patch is as you said, only disable
node_cpuon X86, enablenode_cpuon arm64, powerpc, s390 architectures; the code is in filearch/x86/kernel/paravirt-spinlocks.c:
DECLARE_STATIC_KEY_FALSE(preemted_key);
static_branch_enable(&preemted_key);the default value of
preemted_keyis false and the if conditional statement is reversed, the code is in filekernel/locking/osq_lock.c:
DEFINE_STATIC_KEY_FALSE(preemted_key);static inline int
node_cpu(struct optimistic_spin_node *node){
int cpu = 0;
Quoted 7 lines
if (!static_branch_unlikely(&preemted_key)) cpu =
node->cpu- 1;return cpu;
}In this way, only one nop instruction is added to architectures arm64, powerpc and s390, including virtual machines, without any other changes.
You are right. I am probably too tired last night to read the patch more carefully.
Cheers,
Longman
Juergen Gross
maintainerQuoted 131 lines · Guo Hui
The instructions assigned to the
vcpu_is_preemptedfunction parameter in the X86 architecture physical machine are redundant instructions, causing the multi-core performance of Unixbench to drop by about 4% to 5%. The C function is as follows:
static boolvcpu_is_preempted(long vcpu);The parameter 'vcpu' in the function
osq_lock
that calls the functionvcpu_is_preemptedis assigned as follows:The C code is in the function
node_cpu:
cpu =node->cpu- 1;The instructions corresponding to the C code are:
mov 0x14(%rax),%edi
sub $0x1,%ediThe above instructions are unnecessary
in the X86 Native operating environment, causing high cache-misses and degrading performance.This patch uses
static_keyto not execute this instruction in the Native runtime environment.The code is as follows:
DEFINE_STATIC_KEY_FALSE(preemted_key);static inline int
node_cpu(struct optimistic_spin_node *node){ int cpu = 0; if (!static_branch_unlikely(&preemted_key)) cpu = node->cpu - 1; return cpu; }The patch effect is as follows two machines,
Unixbench runs with full core score:1. Machine configuration:
Intel(R) Xeon(R) Silver 4210 CPU @ 2.20GHz
CPU core: 40
Memory: 256G
OS Kernel: 5.19-rc3Before using the patch:
System Benchmarks Index Values BASELINE RESULT INDEXDhrystone 2 using register variables 116700.0 948326591.2 81261.9 Double-Precision Whetstone 55.0 211986.3 38543.0 Execl Throughput 43.0 43453.2 10105.4 File Copy 1024 bufsize 2000 maxblocks 3960.0 438936.2 1108.4 File Copy 256 bufsize 500 maxblocks 1655.0 118197.4 714.2 File Copy 4096 bufsize 8000 maxblocks 5800.0 1534674.7 2646.0Pipe Throughput 12440.0 46482107.6 37365.0
Pipe-based Context Switching 4000.0 1915094.2 4787.7
Process Creation 126.0 85442.2 6781.1
Shell Scripts (1 concurrent) 42.4 69400.7 16368.1
Shell Scripts (8 concurrent) 6.0 8877.2 14795.3
System Call Overhead 15000.0 4714906.1 3143.3========System Benchmarks Index Score 7923.3
After using the patch:
System Benchmarks Index Values BASELINE RESULT INDEXDhrystone 2 using register variables 116700.0 947032915.5 81151.1 Double-Precision Whetstone 55.0 211971.2 38540.2 Execl Throughput 43.0 45054.8 10477.9 File Copy 1024 bufsize 2000 maxblocks 3960.0 515024.9 1300.6 File Copy 256 bufsize 500 maxblocks 1655.0 146354.6 884.3 File Copy 4096 bufsize 8000 maxblocks 5800.0 1679995.9 2896.5Pipe Throughput 12440.0 46466394.2 37352.4
Pipe-based Context Switching 4000.0 1898221.4 4745.6
Process Creation 126.0 85653.1 6797.9
Shell Scripts (1 concurrent) 42.4 69437.3 16376.7
Shell Scripts (8 concurrent) 6.0 8898.9 14831.4
System Call Overhead 15000.0 4658746.7 3105.8========System Benchmarks Index Score 8248.8
2. Machine configuration:
Hygon C86 7185 32-core Processor
CPU core: 128
Memory: 256G
OS Kernel: 5.19-rc3Before using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 2256644068.3 193371.4
Double-Precision Whetstone 55.0 438969.9 79812.7
Execl Throughput 43.0 10108.6 2350.8File Copy 1024 bufsize 2000 maxblocks 3960.0 275892.8 696.7 File Copy 256 bufsize 500 maxblocks 1655.0 72082.7 435.5 File Copy 4096 bufsize 8000 maxblocks 5800.0 925043.4 1594.9Pipe Throughput 12440.0 118905512.5 95583.2
Pipe-based Context Switching 4000.0 7820945.7 19552.4
Process Creation 126.0 31233.3 2478.8
Shell Scripts (1 concurrent) 42.4 49042.8 11566.7
Shell Scripts (8 concurrent) 6.0 6656.0 11093.3
System Call Overhead 15000.0 6816047.5 4544.0========System Benchmarks Index Score 7756.6
After using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 2252272929.4 192996.8
Double-Precision Whetstone 55.0 451847.2 82154.0
Execl Throughput 43.0 10595.1 2464.0File Copy 1024 bufsize 2000 maxblocks 3960.0 301279.3 760.8 File Copy 256 bufsize 500 maxblocks 1655.0 79291.3 479.1 File Copy 4096 bufsize 8000 maxblocks 5800.0 1039755.2 1792.7Pipe Throughput 12440.0 118701468.1 95419.2
Pipe-based Context Switching 4000.0 8073453.3 20183.6
Process Creation 126.0 33440.9 2654.0
Shell Scripts (1 concurrent) 42.4 52722.6 12434.6
Shell Scripts (8 concurrent) 6.0 7050.4 11750.6
System Call Overhead 15000.0 6834371.5 4556.2========System Benchmarks Index Score 8157.8
Signed-off-by Guo Hui <[email protected]>--- arch/x86/kernel/paravirt-spinlocks.c | 4 ++++ kernel/locking/osq_lock.c | 9 ++++++++- 2 files changed, 12 insertions(+), 1 deletion(-)arch/x86/kernel/paravirt-spinlocks.c +0
-0
arch/x86/kernel/paravirt-spinlocks.c bool pv_is_native_vcpu_is_preempted(void) · line 36
Quoted 4 lines
arch/x86/kernel/paravirt-spinlocks.c +1
-0@@ -33,6 +33,8 @@ bool pv_is_native_vcpu_is_preempted(void) __raw_callee_save___native_vcpu_is_preempted; } +DECLARE_STATIC_KEY_FALSE (preemted_key);
Please name it "preempted_key".
Juergen
_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization
PATCH v3 1/1 x86/paravirt: useless assignment instructions cause Unixbench full core performance degradation
The instructions assigned to the vcpu_is_preempted function parameter in the X86 architecture physical machine are redundant instructions, causing the multi-core performance of Unixbench to drop by about 4% to 5%. The C function is as follows:
static bool vcpu_is_preempted(long vcpu);
The parameter 'vcpu' in the function osq_lock
that calls the function vcpu_is_preempted is assigned as follows:
The C code is in the function node_cpu:
cpu = node->cpu - 1;
The instructions corresponding to the C code are:
mov 0x14(%rax),%edi
sub $0x1,%edi
The above instructions are unnecessary
in the X86 Native operating environment, causing high cache-misses and degrading performance.
This patch uses static_key to not execute this instruction in the Native runtime environment.
The code is as follows:
DEFINE_STATIC_KEY_FALSE(preempted_key);
static inline int node_cpu(struct optimistic_spin_node *node)
{
int cpu = 0;
if (!static_branch_unlikely(&preempted_key))
cpu = node->cpu - 1;
return cpu;
}The patch effect is as follows two machines,
Unixbench runs with full core score:
1. Machine configuration:
Intel(R) Xeon(R) Silver 4210 CPU @ 2.20GHz
CPU core: 40
Memory: 256G
OS Kernel: 5.19-rc3
Before using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 948326591.2 81261.9 Double-Precision Whetstone 55.0 211986.3 38543.0 Execl Throughput 43.0 43453.2 10105.4 File Copy 1024 bufsize 2000 maxblocks 3960.0 438936.2 1108.4 File Copy 256 bufsize 500 maxblocks 1655.0 118197.4 714.2 File Copy 4096 bufsize 8000 maxblocks 5800.0 1534674.7 2646.0
Pipe Throughput 12440.0 46482107.6 37365.0
Pipe-based Context Switching 4000.0 1915094.2 4787.7
Process Creation 126.0 85442.2 6781.1
Shell Scripts (1 concurrent) 42.4 69400.7 16368.1
Shell Scripts (8 concurrent) 6.0 8877.2 14795.3
System Call Overhead 15000.0 4714906.1 3143.3
========
System Benchmarks Index Score 7923.3
After using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 947032915.5 81151.1 Double-Precision Whetstone 55.0 211971.2 38540.2 Execl Throughput 43.0 45054.8 10477.9 File Copy 1024 bufsize 2000 maxblocks 3960.0 515024.9 1300.6 File Copy 256 bufsize 500 maxblocks 1655.0 146354.6 884.3 File Copy 4096 bufsize 8000 maxblocks 5800.0 1679995.9 2896.5
Pipe Throughput 12440.0 46466394.2 37352.4
Pipe-based Context Switching 4000.0 1898221.4 4745.6
Process Creation 126.0 85653.1 6797.9
Shell Scripts (1 concurrent) 42.4 69437.3 16376.7
Shell Scripts (8 concurrent) 6.0 8898.9 14831.4
System Call Overhead 15000.0 4658746.7 3105.8
========
System Benchmarks Index Score 8248.8
2. Machine configuration:
Hygon C86 7185 32-core Processor
CPU core: 128
Memory: 256G
OS Kernel: 5.19-rc3
Before using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 2256644068.3 193371.4
Double-Precision Whetstone 55.0 438969.9 79812.7
Execl Throughput 43.0 10108.6 2350.8
File Copy 1024 bufsize 2000 maxblocks 3960.0 275892.8 696.7 File Copy 256 bufsize 500 maxblocks 1655.0 72082.7 435.5 File Copy 4096 bufsize 8000 maxblocks 5800.0 925043.4 1594.9
Pipe Throughput 12440.0 118905512.5 95583.2
Pipe-based Context Switching 4000.0 7820945.7 19552.4
Process Creation 126.0 31233.3 2478.8
Shell Scripts (1 concurrent) 42.4 49042.8 11566.7
Shell Scripts (8 concurrent) 6.0 6656.0 11093.3
System Call Overhead 15000.0 6816047.5 4544.0
========
System Benchmarks Index Score 7756.6
After using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 2252272929.4 192996.8
Double-Precision Whetstone 55.0 451847.2 82154.0
Execl Throughput 43.0 10595.1 2464.0
File Copy 1024 bufsize 2000 maxblocks 3960.0 301279.3 760.8 File Copy 256 bufsize 500 maxblocks 1655.0 79291.3 479.1 File Copy 4096 bufsize 8000 maxblocks 5800.0 1039755.2 1792.7
Pipe Throughput 12440.0 118701468.1 95419.2
Pipe-based Context Switching 4000.0 8073453.3 20183.6
Process Creation 126.0 33440.9 2654.0
Shell Scripts (1 concurrent) 42.4 52722.6 12434.6
Shell Scripts (8 concurrent) 6.0 7050.4 11750.6
System Call Overhead 15000.0 6834371.5 4556.2
========
System Benchmarks Index Score 8157.8
--- arch/x86/kernel/paravirt-spinlocks.c | 4 ++++ kernel/locking/osq_lock.c | 9 ++++++++- 2 files changed, 12 insertions(+), 1 deletion(-)
arch/x86/kernel/paravirt-spinlocks.c +4 -0
@@ -33,6 +33,8 @@ bool pv_is_native_vcpu_is_preempted(void) __raw_callee_save___native_vcpu_is_preempted; } +DECLARE_STATIC_KEY_FALSE (preempted_key); +void __init paravirt_set_cap (void ) {if (!pv_is_native_spin_unlock ())
@@ -40,4 +42,6 @@ void __init paravirt_set_cap(void)if (!pv_is_native_vcpu_is_preempted ())setup_force_cpu_cap (X86_FEATURE_VCPUPREEMPT ); +else +static_branch_enable (&preempted_key); }
kernel/locking/osq_lock.c +8 -1
@@ -22,9 +22,16 @@ static inline int encode_cpu(int cpu_nr)return cpu_nr +1 ; } +DEFINE_STATIC_KEY_FALSE (preempted_key); +static inline int node_cpu (struct optimistic_spin_node *node) { -return node->cpu -1 ; +int cpu =0 ; + +if (!static_branch_unlikely (&preempted_key)) + cpu = node->cpu -1 ; + +return cpu; }static inline struct optimistic_spin_node *decode_cpu (int encoded_cpu_val)
Signature · 1 line
2.20.1
Peter Zijlstra
maintainerQuoted 12 lines · Guo Hui
The instructions assigned to the
vcpu_is_preemptedfunction parameter in the X86 architecture physical machine are redundant instructions, causing the multi-core performance of Unixbench to drop by about 4% to 5%. The C function is as follows:
static boolvcpu_is_preempted(long vcpu);The parameter 'vcpu' in the function
osq_lock
that calls the functionvcpu_is_preemptedis assigned as follows:The C code is in the function
node_cpu:
cpu =node->cpu- 1;
Quoted 7 lines
The instructions corresponding to the C code are:
mov 0x14(%rax),%edi
sub $0x1,%ediThe above instructions are unnecessary
in the X86 Native operating environment, causing high cache-misses and degrading performance.
The above basically says that argument setup is not patched out and causes significant pain due to a cache-miss.
Quoted 20 lines
Signed-off-by Guo Hui <[email protected]>--- arch/x86/kernel/paravirt-spinlocks.c | 4 ++++ kernel/locking/osq_lock.c | 9 ++++++++- 2 files changed, 12 insertions(+), 1 deletion(-)arch/x86/kernel/paravirt-spinlocks.c +2
-0@@ -33,6 +33,8 @@ bool pv_is_native_vcpu_is_preempted(void) __raw_callee_save___native_vcpu_is_preempted; } +DECLARE_STATIC_KEY_FALSE (preemted_key); +void __init paravirt_set_cap (void ) {if (!pv_is_native_spin_unlock ())
arch/x86/kernel/paravirt-spinlocks.c void __init paravirt_set_cap(void) · line 47
Quoted 6 lines
arch/x86/kernel/paravirt-spinlocks.c +2
-0@@ -40,4 +42,6 @@ void __init paravirt_set_cap(void)if (!pv_is_native_vcpu_is_preempted ())setup_force_cpu_cap (X86_FEATURE_VCPUPREEMPT ); +else +static_branch_enable (&preemted_key); }
At least for x86 it makes sense to have the static_key default the other way around. That is, enable it along with vcpu_is_preempted().
Quoted 10 lines
kernel/locking/osq_lock.c +2
-0@@ -22,9 +22,16 @@ static inline int encode_cpu(int cpu_nr)return cpu_nr +1 ; } +DEFINE_STATIC_KEY_FALSE (preemted_key); +
kernel/locking/osq_lock.c static inline int encode_cpu(int cpu_nr) · line 35
Quoted 10 lines
kernel/locking/osq_lock.c +6
-1static inline int node_cpu (struct optimistic_spin_node *node) { -return node->cpu -1 ; +int cpu =0 ; + +if (!static_branch_unlikely (&preemted_key)) + cpu = node->cpu -1 ; + +return cpu; }
Would not something like:
static inline boolvcpu_is_preempted_node(struct optimistic_spin_node *node)
{
if (!static_branch_unlikely(&vcpu_has_preemption))
return false;
return vcpu_is_preempted(node_cpu(node->prev));
}And then use that like:
if (smp_cond_load_relaxed(&node->locked, VAL || need_resched() || vcpu_is_preempted_node(node)))
Not generate better code still?
_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization
A message that is not in the archive (sent off-list or not yet received): [email protected]
Peter Zijlstra
maintainerQuoted 2 lines · Guo Hui
Ok thanks Peter, your suggestion is very good, I will modify my patch as you suggested.
Because it messes up the order in which people normally read text. Why is top-posting such a bad thing?
Top-posting.
What is the most annoying thing in e-mail?
_______________________________________________
Virtualization mailing list [email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization
PATCH v4 1/1 x86/paravirt: useless assignment instructions cause Unixbench full core performance degradation
In reply to Peter ZijlstraThe instructions assigned to the vcpu_is_preempted function parameter in the X86 architecture physical machine are redundant instructions, causing the multi-core performance of Unixbench to drop by about 4% to 5%. The C function is as follows:
static bool vcpu_is_preempted(long vcpu);
The parameter 'vcpu' in the function osq_lock
that calls the function vcpu_is_preempted is assigned as follows:
The C code is in the function node_cpu:
cpu = node->cpu - 1;
The instructions corresponding to the C code are:
mov 0x14(%rax),%edi
sub $0x1,%edi
The above instructions are unnecessary
in the X86 Native operating environment, causing high cache-misses and degrading performance.
This patch uses static_key to not execute this instruction in the Native runtime environment.
The patch effect is as follows two machines,
Unixbench runs with full core score:
1. Machine configuration:
Intel(R) Xeon(R) Silver 4210 CPU @ 2.20GHz
CPU core: 40
Memory: 256G
OS Kernel: 5.19-rc3
Before using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 948326591.2 81261.9 Double-Precision Whetstone 55.0 211986.3 38543.0 Execl Throughput 43.0 43453.2 10105.4 File Copy 1024 bufsize 2000 maxblocks 3960.0 438936.2 1108.4 File Copy 256 bufsize 500 maxblocks 1655.0 118197.4 714.2 File Copy 4096 bufsize 8000 maxblocks 5800.0 1534674.7 2646.0
Pipe Throughput 12440.0 46482107.6 37365.0
Pipe-based Context Switching 4000.0 1915094.2 4787.7
Process Creation 126.0 85442.2 6781.1
Shell Scripts (1 concurrent) 42.4 69400.7 16368.1
Shell Scripts (8 concurrent) 6.0 8877.2 14795.3
System Call Overhead 15000.0 4714906.1 3143.3
========
System Benchmarks Index Score 7923.3
After using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 947032915.5 81151.1 Double-Precision Whetstone 55.0 211971.2 38540.2 Execl Throughput 43.0 45054.8 10477.9 File Copy 1024 bufsize 2000 maxblocks 3960.0 515024.9 1300.6 File Copy 256 bufsize 500 maxblocks 1655.0 146354.6 884.3 File Copy 4096 bufsize 8000 maxblocks 5800.0 1679995.9 2896.5
Pipe Throughput 12440.0 46466394.2 37352.4
Pipe-based Context Switching 4000.0 1898221.4 4745.6
Process Creation 126.0 85653.1 6797.9
Shell Scripts (1 concurrent) 42.4 69437.3 16376.7
Shell Scripts (8 concurrent) 6.0 8898.9 14831.4
System Call Overhead 15000.0 4658746.7 3105.8
========
System Benchmarks Index Score 8248.8
2. Machine configuration:
Hygon C86 7185 32-core Processor
CPU core: 128
Memory: 256G
OS Kernel: 5.19-rc3
Before using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 2256644068.3 193371.4
Double-Precision Whetstone 55.0 438969.9 79812.7
Execl Throughput 43.0 10108.6 2350.8
File Copy 1024 bufsize 2000 maxblocks 3960.0 275892.8 696.7 File Copy 256 bufsize 500 maxblocks 1655.0 72082.7 435.5 File Copy 4096 bufsize 8000 maxblocks 5800.0 925043.4 1594.9
Pipe Throughput 12440.0 118905512.5 95583.2
Pipe-based Context Switching 4000.0 7820945.7 19552.4
Process Creation 126.0 31233.3 2478.8
Shell Scripts (1 concurrent) 42.4 49042.8 11566.7
Shell Scripts (8 concurrent) 6.0 6656.0 11093.3
System Call Overhead 15000.0 6816047.5 4544.0
========
System Benchmarks Index Score 7756.6
After using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 2252272929.4 192996.8
Double-Precision Whetstone 55.0 451847.2 82154.0
Execl Throughput 43.0 10595.1 2464.0
File Copy 1024 bufsize 2000 maxblocks 3960.0 301279.3 760.8 File Copy 256 bufsize 500 maxblocks 1655.0 79291.3 479.1 File Copy 4096 bufsize 8000 maxblocks 5800.0 1039755.2 1792.7
Pipe Throughput 12440.0 118701468.1 95419.2
Pipe-based Context Switching 4000.0 8073453.3 20183.6
Process Creation 126.0 33440.9 2654.0
Shell Scripts (1 concurrent) 42.4 52722.6 12434.6
Shell Scripts (8 concurrent) 6.0 7050.4 11750.6
System Call Overhead 15000.0 6834371.5 4556.2
========
System Benchmarks Index Score 8157.8
--- arch/x86/kernel/paravirt-spinlocks.c | 4 ++++ kernel/locking/osq_lock.c | 12 +++++++++++- 2 files changed, 15 insertions(+), 1 deletion(-)
arch/x86/kernel/paravirt-spinlocks.c +4 -0
@@ -33,6 +33,8 @@ bool pv_is_native_vcpu_is_preempted(void) __raw_callee_save___native_vcpu_is_preempted; } +DECLARE_STATIC_KEY_TRUE (vcpu_has_preemption); +void __init paravirt_set_cap (void ) {if (!pv_is_native_spin_unlock ())
@@ -40,4 +42,6 @@ void __init paravirt_set_cap(void)if (!pv_is_native_vcpu_is_preempted ())setup_force_cpu_cap (X86_FEATURE_VCPUPREEMPT ); +else +static_branch_disable (&vcpu_has_preemption); }
kernel/locking/osq_lock.c +11 -1
@@ -27,6 +27,16 @@ static inline int node_cpu(struct optimistic_spin_node *node)return node->cpu -1 ; } +DEFINE_STATIC_KEY_TRUE (vcpu_has_preemption); + +static inline bool vcpu_is_preempted_node (struct optimistic_spin_node *node) +{ +if (!static_branch_unlikely (&vcpu_has_preemption)) +return false; + +return vcpu_is_preempted (node_cpu (node->prev)); +} +static inline struct optimistic_spin_node *decode_cpu (int encoded_cpu_val) {int cpu_nr = encoded_cpu_val -1 ;
@@ -141,7 +151,7 @@ bool osq_lock(struct optimistic_spin_queue *lock)* polling, be careful. */ if (smp_cond_load_relaxed (&node->locked,VAL ||need_resched () || -vcpu_is_preempted (node_cpu (node->prev)))) +vcpu_is_preempted_node (node)))return true;/* unqueue */
Signature · 1 line
2.20.1
Quoted 103 lines · Guo Hui
The instructions assigned to the
vcpu_is_preemptedfunction parameter in the X86 architecture physical machine are redundant instructions, causing the multi-core performance of Unixbench to drop by about 4% to 5%. The C function is as follows:
static boolvcpu_is_preempted(long vcpu);The parameter 'vcpu' in the function
osq_lock
that calls the functionvcpu_is_preemptedis assigned as follows:The C code is in the function
node_cpu:
cpu =node->cpu- 1;The instructions corresponding to the C code are:
mov 0x14(%rax),%edi
sub $0x1,%ediThe above instructions are unnecessary
in the X86 Native operating environment, causing high cache-misses and degrading performance.This patch uses
static_keyto not execute this instruction in the Native runtime environment.The patch effect is as follows two machines,
Unixbench runs with full core score:1. Machine configuration:
Intel(R) Xeon(R) Silver 4210 CPU @ 2.20GHz
CPU core: 40
Memory: 256G
OS Kernel: 5.19-rc3Before using the patch:
System Benchmarks Index Values BASELINE RESULT INDEXDhrystone 2 using register variables 116700.0 948326591.2 81261.9 Double-Precision Whetstone 55.0 211986.3 38543.0 Execl Throughput 43.0 43453.2 10105.4 File Copy 1024 bufsize 2000 maxblocks 3960.0 438936.2 1108.4 File Copy 256 bufsize 500 maxblocks 1655.0 118197.4 714.2 File Copy 4096 bufsize 8000 maxblocks 5800.0 1534674.7 2646.0Pipe Throughput 12440.0 46482107.6 37365.0
Pipe-based Context Switching 4000.0 1915094.2 4787.7
Process Creation 126.0 85442.2 6781.1
Shell Scripts (1 concurrent) 42.4 69400.7 16368.1
Shell Scripts (8 concurrent) 6.0 8877.2 14795.3
System Call Overhead 15000.0 4714906.1 3143.3========System Benchmarks Index Score 7923.3
After using the patch:
System Benchmarks Index Values BASELINE RESULT INDEXDhrystone 2 using register variables 116700.0 947032915.5 81151.1 Double-Precision Whetstone 55.0 211971.2 38540.2 Execl Throughput 43.0 45054.8 10477.9 File Copy 1024 bufsize 2000 maxblocks 3960.0 515024.9 1300.6 File Copy 256 bufsize 500 maxblocks 1655.0 146354.6 884.3 File Copy 4096 bufsize 8000 maxblocks 5800.0 1679995.9 2896.5Pipe Throughput 12440.0 46466394.2 37352.4
Pipe-based Context Switching 4000.0 1898221.4 4745.6
Process Creation 126.0 85653.1 6797.9
Shell Scripts (1 concurrent) 42.4 69437.3 16376.7
Shell Scripts (8 concurrent) 6.0 8898.9 14831.4
System Call Overhead 15000.0 4658746.7 3105.8========System Benchmarks Index Score 8248.8
2. Machine configuration:
Hygon C86 7185 32-core Processor
CPU core: 128
Memory: 256G
OS Kernel: 5.19-rc3Before using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 2256644068.3 193371.4
Double-Precision Whetstone 55.0 438969.9 79812.7
Execl Throughput 43.0 10108.6 2350.8File Copy 1024 bufsize 2000 maxblocks 3960.0 275892.8 696.7 File Copy 256 bufsize 500 maxblocks 1655.0 72082.7 435.5 File Copy 4096 bufsize 8000 maxblocks 5800.0 925043.4 1594.9Pipe Throughput 12440.0 118905512.5 95583.2
Pipe-based Context Switching 4000.0 7820945.7 19552.4
Process Creation 126.0 31233.3 2478.8
Shell Scripts (1 concurrent) 42.4 49042.8 11566.7
Shell Scripts (8 concurrent) 6.0 6656.0 11093.3
System Call Overhead 15000.0 6816047.5 4544.0========System Benchmarks Index Score 7756.6
After using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 2252272929.4 192996.8
Double-Precision Whetstone 55.0 451847.2 82154.0
Execl Throughput 43.0 10595.1 2464.0File Copy 1024 bufsize 2000 maxblocks 3960.0 301279.3 760.8 File Copy 256 bufsize 500 maxblocks 1655.0 79291.3 479.1 File Copy 4096 bufsize 8000 maxblocks 5800.0 1039755.2 1792.7Pipe Throughput 12440.0 118701468.1 95419.2
Pipe-based Context Switching 4000.0 8073453.3 20183.6
Process Creation 126.0 33440.9 2654.0
Shell Scripts (1 concurrent) 42.4 52722.6 12434.6
Shell Scripts (8 concurrent) 6.0 7050.4 11750.6
System Call Overhead 15000.0 6834371.5 4556.2
Quoted 4 lines
========System Benchmarks Index Score 8157.8
Signed-off-by Guo Hui <[email protected]>
Juergen
Waiman Long
reviewerQuoted 154 lines · Guo Hui
The instructions assigned to the
vcpu_is_preemptedfunction parameter in the X86 architecture physical machine are redundant instructions, causing the multi-core performance of Unixbench to drop by about 4% to 5%. The C function is as follows:
static boolvcpu_is_preempted(long vcpu);The parameter 'vcpu' in the function
osq_lock
that calls the functionvcpu_is_preemptedis assigned as follows:The C code is in the function
node_cpu:
cpu =node->cpu- 1;The instructions corresponding to the C code are:
mov 0x14(%rax),%edi
sub $0x1,%ediThe above instructions are unnecessary
in the X86 Native operating environment, causing high cache-misses and degrading performance.This patch uses
static_keyto not execute this instruction in the Native runtime environment.The patch effect is as follows two machines,
Unixbench runs with full core score:1. Machine configuration:
Intel(R) Xeon(R) Silver 4210 CPU @ 2.20GHz
CPU core: 40
Memory: 256G
OS Kernel: 5.19-rc3Before using the patch:
System Benchmarks Index Values BASELINE RESULT INDEXDhrystone 2 using register variables 116700.0 948326591.2 81261.9 Double-Precision Whetstone 55.0 211986.3 38543.0 Execl Throughput 43.0 43453.2 10105.4 File Copy 1024 bufsize 2000 maxblocks 3960.0 438936.2 1108.4 File Copy 256 bufsize 500 maxblocks 1655.0 118197.4 714.2 File Copy 4096 bufsize 8000 maxblocks 5800.0 1534674.7 2646.0Pipe Throughput 12440.0 46482107.6 37365.0
Pipe-based Context Switching 4000.0 1915094.2 4787.7
Process Creation 126.0 85442.2 6781.1
Shell Scripts (1 concurrent) 42.4 69400.7 16368.1
Shell Scripts (8 concurrent) 6.0 8877.2 14795.3
System Call Overhead 15000.0 4714906.1 3143.3========System Benchmarks Index Score 7923.3
After using the patch:
System Benchmarks Index Values BASELINE RESULT INDEXDhrystone 2 using register variables 116700.0 947032915.5 81151.1 Double-Precision Whetstone 55.0 211971.2 38540.2 Execl Throughput 43.0 45054.8 10477.9 File Copy 1024 bufsize 2000 maxblocks 3960.0 515024.9 1300.6 File Copy 256 bufsize 500 maxblocks 1655.0 146354.6 884.3 File Copy 4096 bufsize 8000 maxblocks 5800.0 1679995.9 2896.5Pipe Throughput 12440.0 46466394.2 37352.4
Pipe-based Context Switching 4000.0 1898221.4 4745.6
Process Creation 126.0 85653.1 6797.9
Shell Scripts (1 concurrent) 42.4 69437.3 16376.7
Shell Scripts (8 concurrent) 6.0 8898.9 14831.4
System Call Overhead 15000.0 4658746.7 3105.8========System Benchmarks Index Score 8248.8
2. Machine configuration:
Hygon C86 7185 32-core Processor
CPU core: 128
Memory: 256G
OS Kernel: 5.19-rc3Before using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 2256644068.3 193371.4
Double-Precision Whetstone 55.0 438969.9 79812.7
Execl Throughput 43.0 10108.6 2350.8File Copy 1024 bufsize 2000 maxblocks 3960.0 275892.8 696.7 File Copy 256 bufsize 500 maxblocks 1655.0 72082.7 435.5 File Copy 4096 bufsize 8000 maxblocks 5800.0 925043.4 1594.9Pipe Throughput 12440.0 118905512.5 95583.2
Pipe-based Context Switching 4000.0 7820945.7 19552.4
Process Creation 126.0 31233.3 2478.8
Shell Scripts (1 concurrent) 42.4 49042.8 11566.7
Shell Scripts (8 concurrent) 6.0 6656.0 11093.3
System Call Overhead 15000.0 6816047.5 4544.0========System Benchmarks Index Score 7756.6
After using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 2252272929.4 192996.8
Double-Precision Whetstone 55.0 451847.2 82154.0
Execl Throughput 43.0 10595.1 2464.0File Copy 1024 bufsize 2000 maxblocks 3960.0 301279.3 760.8 File Copy 256 bufsize 500 maxblocks 1655.0 79291.3 479.1 File Copy 4096 bufsize 8000 maxblocks 5800.0 1039755.2 1792.7Pipe Throughput 12440.0 118701468.1 95419.2
Pipe-based Context Switching 4000.0 8073453.3 20183.6
Process Creation 126.0 33440.9 2654.0
Shell Scripts (1 concurrent) 42.4 52722.6 12434.6
Shell Scripts (8 concurrent) 6.0 7050.4 11750.6
System Call Overhead 15000.0 6834371.5 4556.2========System Benchmarks Index Score 8157.8
Signed-off-by Guo Hui <[email protected]>--- arch/x86/kernel/paravirt-spinlocks.c | 4 ++++ kernel/locking/osq_lock.c | 12 +++++++++++- 2 files changed, 15 insertions(+), 1 deletion(-)arch/x86/kernel/paravirt-spinlocks.c +4
-0@@ -33,6 +33,8 @@ bool pv_is_native_vcpu_is_preempted(void) __raw_callee_save___native_vcpu_is_preempted; } +DECLARE_STATIC_KEY_TRUE (vcpu_has_preemption); +void __init paravirt_set_cap (void ) {if (!pv_is_native_spin_unlock ())@@ -40,4 +42,6 @@ void __init paravirt_set_cap(void)if (!pv_is_native_vcpu_is_preempted ())setup_force_cpu_cap (X86_FEATURE_VCPUPREEMPT ); +else +static_branch_disable (&vcpu_has_preemption); }kernel/locking/osq_lock.c +10
-0@@ -27,6 +27,16 @@ static inline int node_cpu(struct optimistic_spin_node *node)return node->cpu -1 ; } +DEFINE_STATIC_KEY_TRUE (vcpu_has_preemption); + +static inline bool vcpu_is_preempted_node (struct optimistic_spin_node *node) +{ +if (!static_branch_unlikely (&vcpu_has_preemption)) +return false; + +return vcpu_is_preempted (node_cpu (node->prev)); +} +static inline struct optimistic_spin_node *decode_cpu (int encoded_cpu_val) {int cpu_nr = encoded_cpu_val -1 ;
kernel/locking/osq_lock.c bool osq_lock(struct optimistic_spin_queue *lock) · line 157
Quoted 8 lines
kernel/locking/osq_lock.c +1
-1@@ -141,7 +151,7 @@ bool osq_lock(struct optimistic_spin_queue *lock)* polling, be careful. */ if (smp_cond_load_relaxed (&node->locked,VAL ||need_resched () || -vcpu_is_preempted (node_cpu (node->prev)))) +vcpu_is_preempted_node (node)))return true;/* unqueue */
The patch looks good. I do have a minor nit though.
Usually, DEFINE_STATIC_KEY_TRUE() is paired with static_branch_likely() and DEFINE_STATIC_KEY_FALSE() is paired with static_branch_unlikely(). I think what Peter meant is to use DEFINE_STATIC_KEY_FALSE() and enable vcpu_has_preemption together with X86_FEATURE_VCPUPREEMPT in the same if block.
Cheers,
Longman
PATCH v5 1/1 x86/paravirt: useless assignment instructions cause Unixbench full core performance degradation
The instructions assigned to the vcpu_is_preempted function parameter in the X86 architecture physical machine are redundant instructions, causing the multi-core performance of Unixbench to drop by about 4% to 5%. The C function is as follows:
static bool vcpu_is_preempted(long vcpu);
The parameter 'vcpu' in the function osq_lock
that calls the function vcpu_is_preempted is assigned as follows:
The C code is in the function node_cpu:
cpu = node->cpu - 1;
The instructions corresponding to the C code are:
mov 0x14(%rax),%edi
sub $0x1,%edi
The above instructions are unnecessary
in the X86 Native operating environment, causing high cache-misses and degrading performance.
This patch uses static_key to not execute this instruction in the Native runtime environment.
The patch effect is as follows two machines,
Unixbench runs with full core score:
1. Machine configuration:
Intel(R) Xeon(R) Silver 4210 CPU @ 2.20GHz
CPU core: 40
Memory: 256G
OS Kernel: 5.19-rc3
Before using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 948326591.2 81261.9 Double-Precision Whetstone 55.0 211986.3 38543.0 Execl Throughput 43.0 43453.2 10105.4 File Copy 1024 bufsize 2000 maxblocks 3960.0 438936.2 1108.4 File Copy 256 bufsize 500 maxblocks 1655.0 118197.4 714.2 File Copy 4096 bufsize 8000 maxblocks 5800.0 1534674.7 2646.0
Pipe Throughput 12440.0 46482107.6 37365.0
Pipe-based Context Switching 4000.0 1915094.2 4787.7
Process Creation 126.0 85442.2 6781.1
Shell Scripts (1 concurrent) 42.4 69400.7 16368.1
Shell Scripts (8 concurrent) 6.0 8877.2 14795.3
System Call Overhead 15000.0 4714906.1 3143.3
========
System Benchmarks Index Score 7923.3
After using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 947032915.5 81151.1 Double-Precision Whetstone 55.0 211971.2 38540.2 Execl Throughput 43.0 45054.8 10477.9 File Copy 1024 bufsize 2000 maxblocks 3960.0 515024.9 1300.6 File Copy 256 bufsize 500 maxblocks 1655.0 146354.6 884.3 File Copy 4096 bufsize 8000 maxblocks 5800.0 1679995.9 2896.5
Pipe Throughput 12440.0 46466394.2 37352.4
Pipe-based Context Switching 4000.0 1898221.4 4745.6
Process Creation 126.0 85653.1 6797.9
Shell Scripts (1 concurrent) 42.4 69437.3 16376.7
Shell Scripts (8 concurrent) 6.0 8898.9 14831.4
System Call Overhead 15000.0 4658746.7 3105.8
========
System Benchmarks Index Score 8248.8
2. Machine configuration:
Hygon C86 7185 32-core Processor
CPU core: 128
Memory: 256G
OS Kernel: 5.19-rc3
Before using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 2256644068.3 193371.4
Double-Precision Whetstone 55.0 438969.9 79812.7
Execl Throughput 43.0 10108.6 2350.8
File Copy 1024 bufsize 2000 maxblocks 3960.0 275892.8 696.7 File Copy 256 bufsize 500 maxblocks 1655.0 72082.7 435.5 File Copy 4096 bufsize 8000 maxblocks 5800.0 925043.4 1594.9
Pipe Throughput 12440.0 118905512.5 95583.2
Pipe-based Context Switching 4000.0 7820945.7 19552.4
Process Creation 126.0 31233.3 2478.8
Shell Scripts (1 concurrent) 42.4 49042.8 11566.7
Shell Scripts (8 concurrent) 6.0 6656.0 11093.3
System Call Overhead 15000.0 6816047.5 4544.0
========
System Benchmarks Index Score 7756.6
After using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 2252272929.4 192996.8
Double-Precision Whetstone 55.0 451847.2 82154.0
Execl Throughput 43.0 10595.1 2464.0
File Copy 1024 bufsize 2000 maxblocks 3960.0 301279.3 760.8 File Copy 256 bufsize 500 maxblocks 1655.0 79291.3 479.1 File Copy 4096 bufsize 8000 maxblocks 5800.0 1039755.2 1792.7
Pipe Throughput 12440.0 118701468.1 95419.2
Pipe-based Context Switching 4000.0 8073453.3 20183.6
Process Creation 126.0 33440.9 2654.0
Shell Scripts (1 concurrent) 42.4 52722.6 12434.6
Shell Scripts (8 concurrent) 6.0 7050.4 11750.6
System Call Overhead 15000.0 6834371.5 4556.2
========
System Benchmarks Index Score 8157.8
--- arch/x86/kernel/paravirt-spinlocks.c | 6 +++++- kernel/locking/osq_lock.c | 12 +++++++++++- 2 files changed, 16 insertions(+), 2 deletions(-)
arch/x86/kernel/paravirt-spinlocks.c +5 -1
@@ -33,11 +33,15 @@ bool pv_is_native_vcpu_is_preempted(void) __raw_callee_save___native_vcpu_is_preempted; } +DECLARE_STATIC_KEY_FALSE (vcpu_has_preemption); +void __init paravirt_set_cap (void ) {if (!pv_is_native_spin_unlock ())setup_force_cpu_cap (X86_FEATURE_PVUNLOCK ); -if (!pv_is_native_vcpu_is_preempted ()) +if (!pv_is_native_vcpu_is_preempted ()) {setup_force_cpu_cap (X86_FEATURE_VCPUPREEMPT ); +static_branch_enable (&vcpu_has_preemption); + } }
kernel/locking/osq_lock.c +11 -1
@@ -27,6 +27,16 @@ static inline int node_cpu(struct optimistic_spin_node *node)return node->cpu -1 ; } +DEFINE_STATIC_KEY_FALSE (vcpu_has_preemption); + +static inline bool vcpu_is_preempted_node (struct optimistic_spin_node *node) +{ +if (!static_branch_unlikely (&vcpu_has_preemption)) +return false; + +return vcpu_is_preempted (node_cpu (node->prev)); +} +static inline struct optimistic_spin_node *decode_cpu (int encoded_cpu_val) {int cpu_nr = encoded_cpu_val -1 ;
@@ -141,7 +151,7 @@ bool osq_lock(struct optimistic_spin_queue *lock)* polling, be careful. */ if (smp_cond_load_relaxed (&node->locked,VAL ||need_resched () || -vcpu_is_preempted (node_cpu (node->prev)))) +vcpu_is_preempted_node (node)))return true;/* unqueue */
Signature · 1 line
2.20.1
PATCH v6 1/1 x86/paravirt: useless assignment instructions cause Unixbench full core performance degradation
In reply to Waiman LongThe instructions assigned to the vcpu_is_preempted function parameter in the X86 architecture physical machine are redundant instructions, causing the multi-core performance of Unixbench to drop by about 4% to 5%. The C function is as follows:
static bool vcpu_is_preempted(long vcpu);
The parameter 'vcpu' in the function osq_lock
that calls the function vcpu_is_preempted is assigned as follows:
The C code is in the function node_cpu:
cpu = node->cpu - 1;
The instructions corresponding to the C code are:
mov 0x14(%rax),%edi
sub $0x1,%edi
The above instructions are unnecessary
in the X86 Native operating environment, causing high cache-misses and degrading performance.
This patch uses static_key to not execute this instruction in the Native runtime environment.
The patch effect is as follows two machines,
Unixbench runs with full core score:
1. Machine configuration:
Intel(R) Xeon(R) Silver 4210 CPU @ 2.20GHz
CPU core: 40
Memory: 256G
OS Kernel: 5.19-rc3
Before using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 948326591.2 81261.9 Double-Precision Whetstone 55.0 211986.3 38543.0 Execl Throughput 43.0 43453.2 10105.4 File Copy 1024 bufsize 2000 maxblocks 3960.0 438936.2 1108.4 File Copy 256 bufsize 500 maxblocks 1655.0 118197.4 714.2 File Copy 4096 bufsize 8000 maxblocks 5800.0 1534674.7 2646.0
Pipe Throughput 12440.0 46482107.6 37365.0
Pipe-based Context Switching 4000.0 1915094.2 4787.7
Process Creation 126.0 85442.2 6781.1
Shell Scripts (1 concurrent) 42.4 69400.7 16368.1
Shell Scripts (8 concurrent) 6.0 8877.2 14795.3
System Call Overhead 15000.0 4714906.1 3143.3
========
System Benchmarks Index Score 7923.3
After using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 947032915.5 81151.1 Double-Precision Whetstone 55.0 211971.2 38540.2 Execl Throughput 43.0 45054.8 10477.9 File Copy 1024 bufsize 2000 maxblocks 3960.0 515024.9 1300.6 File Copy 256 bufsize 500 maxblocks 1655.0 146354.6 884.3 File Copy 4096 bufsize 8000 maxblocks 5800.0 1679995.9 2896.5
Pipe Throughput 12440.0 46466394.2 37352.4
Pipe-based Context Switching 4000.0 1898221.4 4745.6
Process Creation 126.0 85653.1 6797.9
Shell Scripts (1 concurrent) 42.4 69437.3 16376.7
Shell Scripts (8 concurrent) 6.0 8898.9 14831.4
System Call Overhead 15000.0 4658746.7 3105.8
========
System Benchmarks Index Score 8248.8
2. Machine configuration:
Hygon C86 7185 32-core Processor
CPU core: 128
Memory: 256G
OS Kernel: 5.19-rc3
Before using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 2256644068.3 193371.4
Double-Precision Whetstone 55.0 438969.9 79812.7
Execl Throughput 43.0 10108.6 2350.8
File Copy 1024 bufsize 2000 maxblocks 3960.0 275892.8 696.7 File Copy 256 bufsize 500 maxblocks 1655.0 72082.7 435.5 File Copy 4096 bufsize 8000 maxblocks 5800.0 925043.4 1594.9
Pipe Throughput 12440.0 118905512.5 95583.2
Pipe-based Context Switching 4000.0 7820945.7 19552.4
Process Creation 126.0 31233.3 2478.8
Shell Scripts (1 concurrent) 42.4 49042.8 11566.7
Shell Scripts (8 concurrent) 6.0 6656.0 11093.3
System Call Overhead 15000.0 6816047.5 4544.0
========
System Benchmarks Index Score 7756.6
After using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 2252272929.4 192996.8
Double-Precision Whetstone 55.0 451847.2 82154.0
Execl Throughput 43.0 10595.1 2464.0
File Copy 1024 bufsize 2000 maxblocks 3960.0 301279.3 760.8 File Copy 256 bufsize 500 maxblocks 1655.0 79291.3 479.1 File Copy 4096 bufsize 8000 maxblocks 5800.0 1039755.2 1792.7
Pipe Throughput 12440.0 118701468.1 95419.2
Pipe-based Context Switching 4000.0 8073453.3 20183.6
Process Creation 126.0 33440.9 2654.0
Shell Scripts (1 concurrent) 42.4 52722.6 12434.6
Shell Scripts (8 concurrent) 6.0 7050.4 11750.6
System Call Overhead 15000.0 6834371.5 4556.2
========
System Benchmarks Index Score 8157.8
--- arch/x86/kernel/paravirt-spinlocks.c | 4 ++++ kernel/locking/osq_lock.c | 12 +++++++++++- 2 files changed, 15 insertions(+), 1 deletion(-)
arch/x86/kernel/paravirt-spinlocks.c +4 -0
@@ -33,6 +33,8 @@ bool pv_is_native_vcpu_is_preempted(void) __raw_callee_save___native_vcpu_is_preempted; } +DECLARE_STATIC_KEY_TRUE (vcpu_has_preemption); +void __init paravirt_set_cap (void ) {if (!pv_is_native_spin_unlock ())
@@ -40,4 +42,6 @@ void __init paravirt_set_cap(void)if (!pv_is_native_vcpu_is_preempted ())setup_force_cpu_cap (X86_FEATURE_VCPUPREEMPT ); +else +static_branch_disable (&vcpu_has_preemption); }
kernel/locking/osq_lock.c +11 -1
@@ -27,6 +27,16 @@ static inline int node_cpu(struct optimistic_spin_node *node)return node->cpu -1 ; } +DEFINE_STATIC_KEY_TRUE (vcpu_has_preemption); + +static inline bool vcpu_is_preempted_node (struct optimistic_spin_node *node) +{ +if (static_branch_likely (&vcpu_has_preemption)) +return vcpu_is_preempted (node_cpu (node->prev)); + +return false; +} +static inline struct optimistic_spin_node *decode_cpu (int encoded_cpu_val) {int cpu_nr = encoded_cpu_val -1 ;
@@ -141,7 +151,7 @@ bool osq_lock(struct optimistic_spin_queue *lock)* polling, be careful. */ if (smp_cond_load_relaxed (&node->locked,VAL ||need_resched () || -vcpu_is_preempted (node_cpu (node->prev)))) +vcpu_is_preempted_node (node)))return true;/* unqueue */
Signature · 1 line
2.20.1
Waiman Long
reviewerQuoted 154 lines · Guo Hui
The instructions assigned to the
vcpu_is_preemptedfunction parameter in the X86 architecture physical machine are redundant instructions, causing the multi-core performance of Unixbench to drop by about 4% to 5%. The C function is as follows:
static boolvcpu_is_preempted(long vcpu);The parameter 'vcpu' in the function
osq_lock
that calls the functionvcpu_is_preemptedis assigned as follows:The C code is in the function
node_cpu:
cpu =node->cpu- 1;The instructions corresponding to the C code are:
mov 0x14(%rax),%edi
sub $0x1,%ediThe above instructions are unnecessary
in the X86 Native operating environment, causing high cache-misses and degrading performance.This patch uses
static_keyto not execute this instruction in the Native runtime environment.The patch effect is as follows two machines,
Unixbench runs with full core score:1. Machine configuration:
Intel(R) Xeon(R) Silver 4210 CPU @ 2.20GHz
CPU core: 40
Memory: 256G
OS Kernel: 5.19-rc3Before using the patch:
System Benchmarks Index Values BASELINE RESULT INDEXDhrystone 2 using register variables 116700.0 948326591.2 81261.9 Double-Precision Whetstone 55.0 211986.3 38543.0 Execl Throughput 43.0 43453.2 10105.4 File Copy 1024 bufsize 2000 maxblocks 3960.0 438936.2 1108.4 File Copy 256 bufsize 500 maxblocks 1655.0 118197.4 714.2 File Copy 4096 bufsize 8000 maxblocks 5800.0 1534674.7 2646.0Pipe Throughput 12440.0 46482107.6 37365.0
Pipe-based Context Switching 4000.0 1915094.2 4787.7
Process Creation 126.0 85442.2 6781.1
Shell Scripts (1 concurrent) 42.4 69400.7 16368.1
Shell Scripts (8 concurrent) 6.0 8877.2 14795.3
System Call Overhead 15000.0 4714906.1 3143.3========System Benchmarks Index Score 7923.3
After using the patch:
System Benchmarks Index Values BASELINE RESULT INDEXDhrystone 2 using register variables 116700.0 947032915.5 81151.1 Double-Precision Whetstone 55.0 211971.2 38540.2 Execl Throughput 43.0 45054.8 10477.9 File Copy 1024 bufsize 2000 maxblocks 3960.0 515024.9 1300.6 File Copy 256 bufsize 500 maxblocks 1655.0 146354.6 884.3 File Copy 4096 bufsize 8000 maxblocks 5800.0 1679995.9 2896.5Pipe Throughput 12440.0 46466394.2 37352.4
Pipe-based Context Switching 4000.0 1898221.4 4745.6
Process Creation 126.0 85653.1 6797.9
Shell Scripts (1 concurrent) 42.4 69437.3 16376.7
Shell Scripts (8 concurrent) 6.0 8898.9 14831.4
System Call Overhead 15000.0 4658746.7 3105.8========System Benchmarks Index Score 8248.8
2. Machine configuration:
Hygon C86 7185 32-core Processor
CPU core: 128
Memory: 256G
OS Kernel: 5.19-rc3Before using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 2256644068.3 193371.4
Double-Precision Whetstone 55.0 438969.9 79812.7
Execl Throughput 43.0 10108.6 2350.8File Copy 1024 bufsize 2000 maxblocks 3960.0 275892.8 696.7 File Copy 256 bufsize 500 maxblocks 1655.0 72082.7 435.5 File Copy 4096 bufsize 8000 maxblocks 5800.0 925043.4 1594.9Pipe Throughput 12440.0 118905512.5 95583.2
Pipe-based Context Switching 4000.0 7820945.7 19552.4
Process Creation 126.0 31233.3 2478.8
Shell Scripts (1 concurrent) 42.4 49042.8 11566.7
Shell Scripts (8 concurrent) 6.0 6656.0 11093.3
System Call Overhead 15000.0 6816047.5 4544.0========System Benchmarks Index Score 7756.6
After using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 2252272929.4 192996.8
Double-Precision Whetstone 55.0 451847.2 82154.0
Execl Throughput 43.0 10595.1 2464.0File Copy 1024 bufsize 2000 maxblocks 3960.0 301279.3 760.8 File Copy 256 bufsize 500 maxblocks 1655.0 79291.3 479.1 File Copy 4096 bufsize 8000 maxblocks 5800.0 1039755.2 1792.7Pipe Throughput 12440.0 118701468.1 95419.2
Pipe-based Context Switching 4000.0 8073453.3 20183.6
Process Creation 126.0 33440.9 2654.0
Shell Scripts (1 concurrent) 42.4 52722.6 12434.6
Shell Scripts (8 concurrent) 6.0 7050.4 11750.6
System Call Overhead 15000.0 6834371.5 4556.2========System Benchmarks Index Score 8157.8
Signed-off-by Guo Hui <[email protected]>--- arch/x86/kernel/paravirt-spinlocks.c | 4 ++++ kernel/locking/osq_lock.c | 12 +++++++++++- 2 files changed, 15 insertions(+), 1 deletion(-)arch/x86/kernel/paravirt-spinlocks.c +4
-0@@ -33,6 +33,8 @@ bool pv_is_native_vcpu_is_preempted(void) __raw_callee_save___native_vcpu_is_preempted; } +DECLARE_STATIC_KEY_TRUE (vcpu_has_preemption); +void __init paravirt_set_cap (void ) {if (!pv_is_native_spin_unlock ())@@ -40,4 +42,6 @@ void __init paravirt_set_cap(void)if (!pv_is_native_vcpu_is_preempted ())setup_force_cpu_cap (X86_FEATURE_VCPUPREEMPT ); +else +static_branch_disable (&vcpu_has_preemption); }kernel/locking/osq_lock.c +10
-0@@ -27,6 +27,16 @@ static inline int node_cpu(struct optimistic_spin_node *node)return node->cpu -1 ; } +DEFINE_STATIC_KEY_TRUE (vcpu_has_preemption); + +static inline bool vcpu_is_preempted_node (struct optimistic_spin_node *node) +{ +if (static_branch_likely (&vcpu_has_preemption)) +return vcpu_is_preempted (node_cpu (node->prev)); + +return false; +} +static inline struct optimistic_spin_node *decode_cpu (int encoded_cpu_val) {int cpu_nr = encoded_cpu_val -1 ;
kernel/locking/osq_lock.c bool osq_lock(struct optimistic_spin_queue *lock) · line 157
Quoted 8 lines
kernel/locking/osq_lock.c +1
-1@@ -141,7 +151,7 @@ bool osq_lock(struct optimistic_spin_queue *lock)* polling, be careful. */ if (smp_cond_load_relaxed (&node->locked,VAL ||need_resched () || -vcpu_is_preempted (node_cpu (node->prev)))) +vcpu_is_preempted_node (node)))return true;/* unqueue */
How about a further improvement on configurations that don't use vcpu_is_preempted() at all?
+#ifdef vcpu_is_preempted +DEFINE_STATIC_KEY_TRUE(vcpu_has_preemption);
+
static inline int node_cpu(struct optimistic_spin_node *node) {
return node->cpu - 1;
}
+static inline bool vcpu_is_preempted_node(struct optimistic_spin_node *node) +{
+ if (static_branch_likely(&vcpu_has_preemption))
+ return vcpu_is_preempted(node_cpu(node->prev));
+
+ return false;
+}
+#else
+static inline bool vcpu_is_preempted_node(struct optimistic_spin_node *node) +{
+ return false;
+}
+#endif
+
static inline struct optimistic_spin_node *decode_cpu(int encoded_cpu_val) {
int cpu_nr = encoded_cpu_val - 1;
@@ -141,7 +158,7 @@ bool osq_lock(struct optimistic_spin_queue *lock)
* polling, be careful.
*/
if (smp_cond_load_relaxed(&node->locked, VAL || need_resched() ||
- vcpu_is_preempted(node_cpu(node->prev))))
+ vcpu_is_preempted_node(node))) return true;
/* unqueue */
For those configurations, vcpu_is_preempted_node() will just get compiled out.
Cheers,
Longman
PATCH v7 1/1 x86/paravirt: useless assignment instructions cause Unixbench full core performance degradation
The instructions assigned to the vcpu_is_preempted function parameter in the X86 architecture physical machine are redundant instructions, causing the multi-core performance of Unixbench to drop by about 4% to 5%. The C function is as follows:
static bool vcpu_is_preempted(long vcpu);
The parameter 'vcpu' in the function osq_lock
that calls the function vcpu_is_preempted is assigned as follows:
The C code is in the function node_cpu:
cpu = node->cpu - 1;
The instructions corresponding to the C code are:
mov 0x14(%rax),%edi
sub $0x1,%edi
The above instructions are unnecessary
in the X86 Native operating environment, causing high cache-misses and degrading performance.
This patch uses static_key to not execute this instruction in the Native runtime environment.
The patch effect is as follows two machines,
Unixbench runs with full core score:
1. Machine configuration:
Intel(R) Xeon(R) Silver 4210 CPU @ 2.20GHz
CPU core: 40
Memory: 256G
OS Kernel: 5.19-rc3
Before using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 948326591.2 81261.9 Double-Precision Whetstone 55.0 211986.3 38543.0 Execl Throughput 43.0 43453.2 10105.4 File Copy 1024 bufsize 2000 maxblocks 3960.0 438936.2 1108.4 File Copy 256 bufsize 500 maxblocks 1655.0 118197.4 714.2 File Copy 4096 bufsize 8000 maxblocks 5800.0 1534674.7 2646.0
Pipe Throughput 12440.0 46482107.6 37365.0
Pipe-based Context Switching 4000.0 1915094.2 4787.7
Process Creation 126.0 85442.2 6781.1
Shell Scripts (1 concurrent) 42.4 69400.7 16368.1
Shell Scripts (8 concurrent) 6.0 8877.2 14795.3
System Call Overhead 15000.0 4714906.1 3143.3
========
System Benchmarks Index Score 7923.3
After using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 947032915.5 81151.1 Double-Precision Whetstone 55.0 211971.2 38540.2 Execl Throughput 43.0 45054.8 10477.9 File Copy 1024 bufsize 2000 maxblocks 3960.0 515024.9 1300.6 File Copy 256 bufsize 500 maxblocks 1655.0 146354.6 884.3 File Copy 4096 bufsize 8000 maxblocks 5800.0 1679995.9 2896.5
Pipe Throughput 12440.0 46466394.2 37352.4
Pipe-based Context Switching 4000.0 1898221.4 4745.6
Process Creation 126.0 85653.1 6797.9
Shell Scripts (1 concurrent) 42.4 69437.3 16376.7
Shell Scripts (8 concurrent) 6.0 8898.9 14831.4
System Call Overhead 15000.0 4658746.7 3105.8
========
System Benchmarks Index Score 8248.8
2. Machine configuration:
Hygon C86 7185 32-core Processor
CPU core: 128
Memory: 256G
OS Kernel: 5.19-rc3
Before using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 2256644068.3 193371.4
Double-Precision Whetstone 55.0 438969.9 79812.7
Execl Throughput 43.0 10108.6 2350.8
File Copy 1024 bufsize 2000 maxblocks 3960.0 275892.8 696.7 File Copy 256 bufsize 500 maxblocks 1655.0 72082.7 435.5 File Copy 4096 bufsize 8000 maxblocks 5800.0 925043.4 1594.9
Pipe Throughput 12440.0 118905512.5 95583.2
Pipe-based Context Switching 4000.0 7820945.7 19552.4
Process Creation 126.0 31233.3 2478.8
Shell Scripts (1 concurrent) 42.4 49042.8 11566.7
Shell Scripts (8 concurrent) 6.0 6656.0 11093.3
System Call Overhead 15000.0 6816047.5 4544.0
========
System Benchmarks Index Score 7756.6
After using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 2252272929.4 192996.8
Double-Precision Whetstone 55.0 451847.2 82154.0
Execl Throughput 43.0 10595.1 2464.0
File Copy 1024 bufsize 2000 maxblocks 3960.0 301279.3 760.8 File Copy 256 bufsize 500 maxblocks 1655.0 79291.3 479.1 File Copy 4096 bufsize 8000 maxblocks 5800.0 1039755.2 1792.7
Pipe Throughput 12440.0 118701468.1 95419.2
Pipe-based Context Switching 4000.0 8073453.3 20183.6
Process Creation 126.0 33440.9 2654.0
Shell Scripts (1 concurrent) 42.4 52722.6 12434.6
Shell Scripts (8 concurrent) 6.0 7050.4 11750.6
System Call Overhead 15000.0 6834371.5 4556.2
========
System Benchmarks Index Score 8157.8
--- arch/x86/kernel/paravirt-spinlocks.c | 4 ++++ kernel/locking/osq_lock.c | 19 ++++++++++++++++++- 2 files changed, 22 insertions(+), 1 deletion(-)
arch/x86/kernel/paravirt-spinlocks.c +4 -0
@@ -33,6 +33,8 @@ bool pv_is_native_vcpu_is_preempted(void) __raw_callee_save___native_vcpu_is_preempted; } +DECLARE_STATIC_KEY_TRUE (vcpu_has_preemption); +void __init paravirt_set_cap (void ) {if (!pv_is_native_spin_unlock ())
@@ -40,4 +42,6 @@ void __init paravirt_set_cap(void)if (!pv_is_native_vcpu_is_preempted ())setup_force_cpu_cap (X86_FEATURE_VCPUPREEMPT ); +else +static_branch_disable (&vcpu_has_preemption); }
kernel/locking/osq_lock.c +18 -1
@@ -27,6 +27,23 @@ static inline int node_cpu(struct optimistic_spin_node *node)return node->cpu -1 ; } +#ifdef vcpu_is_preempted +DEFINE_STATIC_KEY_TRUE (vcpu_has_preemption); + +static inline bool vcpu_is_preempted_node (struct optimistic_spin_node *node) +{ +if (static_branch_likely (&vcpu_has_preemption)) +return vcpu_is_preempted (node_cpu (node->prev)); + +return false; +} +#else +static inline bool vcpu_is_preempted_node (struct optimistic_spin_node *node) +{ +return false; +} +#endif +static inline struct optimistic_spin_node *decode_cpu (int encoded_cpu_val) {int cpu_nr = encoded_cpu_val -1 ;
@@ -141,7 +158,7 @@ bool osq_lock(struct optimistic_spin_queue *lock)* polling, be careful. */ if (smp_cond_load_relaxed (&node->locked,VAL ||need_resched () || -vcpu_is_preempted (node_cpu (node->prev)))) +vcpu_is_preempted_node (node)))return true;/* unqueue */
Signature · 1 line
2.20.1
Quoted 161 lines · Guo Hui
The instructions assigned to the
vcpu_is_preemptedfunction parameter in the X86 architecture physical machine are redundant instructions, causing the multi-core performance of Unixbench to drop by about 4% to 5%. The C function is as follows:
static boolvcpu_is_preempted(long vcpu);The parameter 'vcpu' in the function
osq_lock
that calls the functionvcpu_is_preemptedis assigned as follows:The C code is in the function
node_cpu:
cpu =node->cpu- 1;The instructions corresponding to the C code are:
mov 0x14(%rax),%edi
sub $0x1,%ediThe above instructions are unnecessary
in the X86 Native operating environment, causing high cache-misses and degrading performance.This patch uses
static_keyto not execute this instruction in the Native runtime environment.The patch effect is as follows two machines,
Unixbench runs with full core score:1. Machine configuration:
Intel(R) Xeon(R) Silver 4210 CPU @ 2.20GHz
CPU core: 40
Memory: 256G
OS Kernel: 5.19-rc3Before using the patch:
System Benchmarks Index Values BASELINE RESULT INDEXDhrystone 2 using register variables 116700.0 948326591.2 81261.9 Double-Precision Whetstone 55.0 211986.3 38543.0 Execl Throughput 43.0 43453.2 10105.4 File Copy 1024 bufsize 2000 maxblocks 3960.0 438936.2 1108.4 File Copy 256 bufsize 500 maxblocks 1655.0 118197.4 714.2 File Copy 4096 bufsize 8000 maxblocks 5800.0 1534674.7 2646.0Pipe Throughput 12440.0 46482107.6 37365.0
Pipe-based Context Switching 4000.0 1915094.2 4787.7
Process Creation 126.0 85442.2 6781.1
Shell Scripts (1 concurrent) 42.4 69400.7 16368.1
Shell Scripts (8 concurrent) 6.0 8877.2 14795.3
System Call Overhead 15000.0 4714906.1 3143.3========System Benchmarks Index Score 7923.3
After using the patch:
System Benchmarks Index Values BASELINE RESULT INDEXDhrystone 2 using register variables 116700.0 947032915.5 81151.1 Double-Precision Whetstone 55.0 211971.2 38540.2 Execl Throughput 43.0 45054.8 10477.9 File Copy 1024 bufsize 2000 maxblocks 3960.0 515024.9 1300.6 File Copy 256 bufsize 500 maxblocks 1655.0 146354.6 884.3 File Copy 4096 bufsize 8000 maxblocks 5800.0 1679995.9 2896.5Pipe Throughput 12440.0 46466394.2 37352.4
Pipe-based Context Switching 4000.0 1898221.4 4745.6
Process Creation 126.0 85653.1 6797.9
Shell Scripts (1 concurrent) 42.4 69437.3 16376.7
Shell Scripts (8 concurrent) 6.0 8898.9 14831.4
System Call Overhead 15000.0 4658746.7 3105.8========System Benchmarks Index Score 8248.8
2. Machine configuration:
Hygon C86 7185 32-core Processor
CPU core: 128
Memory: 256G
OS Kernel: 5.19-rc3Before using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 2256644068.3 193371.4
Double-Precision Whetstone 55.0 438969.9 79812.7
Execl Throughput 43.0 10108.6 2350.8File Copy 1024 bufsize 2000 maxblocks 3960.0 275892.8 696.7 File Copy 256 bufsize 500 maxblocks 1655.0 72082.7 435.5 File Copy 4096 bufsize 8000 maxblocks 5800.0 925043.4 1594.9Pipe Throughput 12440.0 118905512.5 95583.2
Pipe-based Context Switching 4000.0 7820945.7 19552.4
Process Creation 126.0 31233.3 2478.8
Shell Scripts (1 concurrent) 42.4 49042.8 11566.7
Shell Scripts (8 concurrent) 6.0 6656.0 11093.3
System Call Overhead 15000.0 6816047.5 4544.0========System Benchmarks Index Score 7756.6
After using the patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 2252272929.4 192996.8
Double-Precision Whetstone 55.0 451847.2 82154.0
Execl Throughput 43.0 10595.1 2464.0File Copy 1024 bufsize 2000 maxblocks 3960.0 301279.3 760.8 File Copy 256 bufsize 500 maxblocks 1655.0 79291.3 479.1 File Copy 4096 bufsize 8000 maxblocks 5800.0 1039755.2 1792.7Pipe Throughput 12440.0 118701468.1 95419.2
Pipe-based Context Switching 4000.0 8073453.3 20183.6
Process Creation 126.0 33440.9 2654.0
Shell Scripts (1 concurrent) 42.4 52722.6 12434.6
Shell Scripts (8 concurrent) 6.0 7050.4 11750.6
System Call Overhead 15000.0 6834371.5 4556.2========System Benchmarks Index Score 8157.8
Signed-off-by Guo Hui <[email protected]>--- arch/x86/kernel/paravirt-spinlocks.c | 4 ++++ kernel/locking/osq_lock.c | 19 ++++++++++++++++++- 2 files changed, 22 insertions(+), 1 deletion(-)arch/x86/kernel/paravirt-spinlocks.c +4
-0@@ -33,6 +33,8 @@ bool pv_is_native_vcpu_is_preempted(void) __raw_callee_save___native_vcpu_is_preempted; } +DECLARE_STATIC_KEY_TRUE (vcpu_has_preemption); +void __init paravirt_set_cap (void ) {if (!pv_is_native_spin_unlock ())@@ -40,4 +42,6 @@ void __init paravirt_set_cap(void)if (!pv_is_native_vcpu_is_preempted ())setup_force_cpu_cap (X86_FEATURE_VCPUPREEMPT ); +else +static_branch_disable (&vcpu_has_preemption); }kernel/locking/osq_lock.c +17
-0@@ -27,6 +27,23 @@ static inline int node_cpu(struct optimistic_spin_node *node)return node->cpu -1 ; } +#ifdef vcpu_is_preempted +DEFINE_STATIC_KEY_TRUE (vcpu_has_preemption); + +static inline bool vcpu_is_preempted_node (struct optimistic_spin_node *node) +{ +if (static_branch_likely (&vcpu_has_preemption)) +return vcpu_is_preempted (node_cpu (node->prev)); + +return false; +} +#else +static inline bool vcpu_is_preempted_node (struct optimistic_spin_node *node) +{ +return false; +} +#endif +static inline struct optimistic_spin_node *decode_cpu (int encoded_cpu_val) {int cpu_nr = encoded_cpu_val -1 ;
kernel/locking/osq_lock.c bool osq_lock(struct optimistic_spin_queue *lock) · line 164
Quoted 8 lines
kernel/locking/osq_lock.c +1
-1@@ -141,7 +158,7 @@ bool osq_lock(struct optimistic_spin_queue *lock)* polling, be careful. */ if (smp_cond_load_relaxed (&node->locked,VAL ||need_resched () || -vcpu_is_preempted (node_cpu (node->prev)))) +vcpu_is_preempted_node (node)))return true;/* unqueue */