Re: [PATCH] x86/paravirt: useless assignment instructions cause Unixbench full core performance degradation
From: Waiman Long
Date: Thu Jun 23 2022 - 17:54:18 EST
On 6/23/22 11:50, Guo Hui wrote:
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
Signed-off-by: Guo Hui <guohui@xxxxxxxxxxxxx>
---
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(-)
diff --git a/arch/x86/include/asm/alternative.h b/arch/x86/include/asm/alternative.h
index 9b10c8c76..5979ebe89 100644
--- a/arch/x86/include/asm/alternative.h
+++ b/arch/x86/include/asm/alternative.h
@@ -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" \
diff --git a/arch/x86/kernel/alternative.c b/arch/x86/kernel/alternative.c
index e257f6c80..cf77be884 100644
--- a/arch/x86/kernel/alternative.c
+++ b/arch/x86/kernel/alternative.c
@@ -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;
+ }
+
This is hacky and it may incorrectly affect other alternatives that
patches thing to nop.
BUG_ON(a->instrlen > sizeof(insn_buff));
BUG_ON(feature >= (NCAPINTS + NBUGINTS) * 32);
diff --git a/include/linux/osq_lock.h b/include/linux/osq_lock.h
index 5581dbd3b..ee960e3aa 100644
--- a/include/linux/osq_lock.h
+++ b/include/linux/osq_lock.h
@@ -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
diff --git a/kernel/locking/osq_lock.c b/kernel/locking/osq_lock.c
index d5610ad52..bbe7d640c 100644
--- a/kernel/locking/osq_lock.c
+++ b/kernel/locking/osq_lock.c
@@ -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