Re: [PATCH v6] x86/paravirt: useless assignment instructions cause Unixbench full core performance degradation
From: Waiman Long
Date: Tue Jun 28 2022 - 10:15:32 EST
On 6/28/22 08:54, Guo Hui wrote:
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
Signed-off-by: Guo Hui <guohui@xxxxxxxxxxxxx>
---
arch/x86/kernel/paravirt-spinlocks.c | 4 ++++
kernel/locking/osq_lock.c | 12 +++++++++++-
2 files changed, 15 insertions(+), 1 deletion(-)
diff --git a/arch/x86/kernel/paravirt-spinlocks.c b/arch/x86/kernel/paravirt-spinlocks.c
index 9e1ea99..a2eb375 100644
--- a/arch/x86/kernel/paravirt-spinlocks.c
+++ b/arch/x86/kernel/paravirt-spinlocks.c
@@ -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);
}
diff --git a/kernel/locking/osq_lock.c b/kernel/locking/osq_lock.c
index d5610ad..883e815 100644
--- a/kernel/locking/osq_lock.c
+++ b/kernel/locking/osq_lock.c
@@ -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 */
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