Fix performance issue brought by TSC-sync logic
PATCH 1/1 Fix performance issue brought by TSC-sync logic
Recently we found one performance bug when doing network test with VTd assigned devices - in some extreme case, the network performance in HVM using new Linux kernel could be 1/20 of native. Root cause is one of our sync-tsc-under-deep-C-state patches brings extra kilo-TSC drift between pCPUs and let check-tsc-sync logic in HVM failed. The result is the kernel fails to use platform timer (HPET, PMtimer) for gettimeofday instead of TSC and brings very frequent costly IOport access VMExit - triple per one call.
We provides below 2 patches to address the issue:
tsc1.patch: Minimize the TSC drift between pCPUs by letting BSP/AP set
TSC at the same time in time_calibration_rendezvous(). Looping a few times before writing tsc sounds better, but it may be too costly.
tsc2.patch: only do TSC-sync if really necessary, which narrows its effect a lot.
Thanks,
Xiaowei
xen/arch/x86/time.c +8 -9
@@ -1095,22 +1095,21 @@ static void time_calibration_rendezvous(while (atomic_read (&r->nr_cpus) != (total_cpus -1 ) )cpu_relax (); r->master_stime =read_platform_stime (); -rdtscll (r->master_tsc_stamp); +if (boot_cpu_has (X86_FEATURE_CONSTANT_TSC ) ) +rdtscll (r->master_tsc_stamp);mb ();/* write r->master_* /then/ signal */ atomic_inc (&r->nr_cpus); - c->local_tsc_stamp = r->master_tsc_stamp; }else {atomic_inc (&r->nr_cpus);while (atomic_read (&r->nr_cpus) != total_cpus ) -cpu_relax (); -mb ();/* receive signal /then/ read r->master_* */ -if (boot_cpu_has (X86_FEATURE_CONSTANT_TSC ) ) -wrmsrl (MSR_IA32_TSC , r->master_tsc_stamp); -rdtscll (c->local_tsc_stamp); - } - +mb ();/* receive signal /then/ read r->master_* */ + } + +if (boot_cpu_has (X86_FEATURE_CONSTANT_TSC ) ) +wrmsrl (MSR_IA32_TSC , r->master_tsc_stamp); +rdtscll (c->local_tsc_stamp); c->stime_local_stamp =get_s_time (); c->stime_master_stamp = r->master_stime;
xen/arch/x86/acpi/cpu_idle.c +9 -0
@@ -737,6 +737,15 @@ long set_cx_pminfo(uint32_t cpu, structif ( cpu_id ==0 && pm_idle_save ==NULL ) { +int deepest_cx = acpi_power->states[acpi_power->count -1 ].type; +if ( max_cstate >=3 && deepest_cx >=ACPI_STATE_C3 ) + tsc_may_stop =1 ; +else if ( max_cstate >=2 && deepest_cx >=ACPI_STATE_C2 + && !local_apic_timer_c2_ok ) + tsc_may_stop =1 ; +else + tsc_may_stop =0 ; + pm_idle_save = pm_idle; pm_idle = acpi_processor_idle; }
xen/arch/x86/time.c +8 -3
@@ -1091,6 +1091,8 @@ struct calibration_rendezvous {u64 master_tsc_stamp; }; +int tsc_may_stop__read_mostly =0 ; +static void time_calibration_rendezvous (void *_r) {struct cpu_calibration *c = &this_cpu (cpu_calibration);
@@ -1102,7 +1104,9 @@ static void time_calibration_rendezvous(while (atomic_read (&r->nr_cpus) != (total_cpus -1 ) )cpu_relax (); r->master_stime =read_platform_stime (); -if (boot_cpu_has (X86_FEATURE_CONSTANT_TSC ) ) +if ( !boot_cpu_has (X86_FEATURE_NOSTOP_TSC ) + &&boot_cpu_has (X86_FEATURE_CONSTANT_TSC ) + && tsc_may_stop )rdtscll (r->master_tsc_stamp);mb ();/* write r->master_* /then/ signal */ atomic_inc (&r->nr_cpus);
@@ -1114,7 +1118,7 @@ static void time_calibration_rendezvous(mb ();/* receive signal /then/ read r->master_* */ } -if (boot_cpu_has (X86_FEATURE_CONSTANT_TSC ) ) +if ( r->master_tsc_stamp )wrmsrl (MSR_IA32_TSC , r->master_tsc_stamp);rdtscll (c->local_tsc_stamp); c->stime_local_stamp =get_s_time ();
@@ -1127,7 +1131,8 @@ static void time_calibration(void *unuse {struct calibration_rendezvous r = { .cpu_calibration_map = cpu_online_map, - .nr_cpus =ATOMIC_INIT (0 ) + .nr_cpus =ATOMIC_INIT (0 ), + .master_tsc_stamp =0 };/* @wait=1 because we must wait for all cpus before freeing @r. */
xen/include/asm-x86/time.h +2 -0
@@ -41,4 +41,6 @@ uint64_t acpi_pm_tick_to_ns(uint64_t ticuint64_t acpi_pm_tick_to_ns (uint64_t ticks);uint64_t ns_to_acpi_pm_tick (uint64_t ns); +extern int tsc_may_stop; +#endif /* __X86_TIME_H__ */
_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel
Quoted 10 lines · Xiaowei
Recently we found one performance bug when doing network test with VTd assigned devices - in some extreme case, the network performance in HVM using new Linux kernel could be 1/20 of native. Root cause is one of our sync-tsc-under-deep-C-state patches brings extra kilo-TSC drift between pCPUs and let check-tsc-sync logic in HVM failed. The result is the kernel fails to use platform timer (HPET, PMtimer) for gettimeofday instead of TSC and brings very frequent costly IOport access VMExit - triple per one call.
We provides below 2 patches to address the issue:
Patch 1 looks reasonable. Patch number 2 I'm less keen on, since patch 1 should suffice? Also I think regular re-sync across CPUs is a good idea anyway. And that also reminds me -- isn't the CONSTANT_TSC logic in time.c broken by host S3, and also by CPU hotplug? There's nothing to force sync of AP TSC to BP TSC when an AP comes online after boot. Doesn't init_percpu_time() need to handle that?
-- Keir
Quoted 19 lines
From: Keir Fraser
Sent: Monday, February 23, 2009 8:52 PMQuoted 1 line · Xiaowei
Recently we found one performance bug when doing network
test with VTd
Quoted 1 line
assigned devices - in some extreme case, the network
performance in HVM
Quoted 1 line
using new Linux kernel could be 1/20 of native. Root cause
is one of our
Quoted 1 line
sync-tsc-under-deep-C-state patches brings extra kilo-TSC
drift between
Quoted 6 lines
pCPUs and let check-tsc-sync logic in HVM failed. The result is the kernel fails to use platform timer (HPET, PMtimer) for gettimeofday instead of TSC and brings very frequent costly IOport access VMExit - triple per one call.
We provides below 2 patches to address the issue:
Quoted 10 lines
Patch 1 looks reasonable. Patch number 2 I'm less keen on, since patch 1
should suffice? Also I think regular re-sync across CPUs is a good idea anyway. And that also reminds me -- isn't theCONSTANT_TSClogic intime.c
broken by host S3, and also by CPU hotplug? There's nothing to force sync of
AP TSC to BP TSC when an AP comes online after boot. Doesn'tinit_percpu_time()need to handle that?
Ah, yes, it's broken regarding to S3. We'll work out a patch to handle it.
Thanks,
Kevin
Quoted 11 lines · Keir Fraser
Quoted 9 lines · Xiaowei
Recently we found one performance bug when doing network test with VTd assigned devices - in some extreme case, the network performance in HVM using new Linux kernel could be 1/20 of native. Root cause is one of our sync-tsc-under-deep-C-state patches brings extra kilo-TSC drift between pCPUs and let check-tsc-sync logic in HVM failed. The result is the kernel fails to use platform timer (HPET, PMtimer) for gettimeofday instead of TSC and brings very frequent costly IOport access VMExit - triple per one call.
Quoted 5 lines
Quoted 1 line
We provides below 2 patches to address the issue:
Patch 1 looks reasonable. Patch number 2 I'm less keen on, since patch 1 should suffice? Also I think regular re-sync across CPUs is a good idea anyway.
Here is average of 100 cycles skew results on one core 2 quad machine:
1) TSC-sync: 1300
2) TSC-sync+tsc1.patch: 400
3) without TSC-sync: 200 (a.k.a sync at boot time only)
We can see from 1) to 2), cycles skew improves a lot. However Linux kernel's logic to check TSC sync (check_tsc_warp) is very strict, so even with tsc1.patch, there are still chances to observe checking failed inside VM.
For further improvement to reach the effect of 3), e.g. by taking care of cache consistance amongs CPUs, there will be more overhead. And considering the function is called per second, we are hesitating to do this. What's your idea?:)
Thanks,
xiaowei
Quoted 4 lines · Xiaowei
For further improvement to reach the effect of 3), e.g. by taking care of cache consistance amongs CPUs, there will be more overhead. And considering the function is called per second, we are hesitating to do this. What's your idea?:)
Maybe we should see what that overhead is... Also we may not really need to run that function every second.
-- Keir
Quoted 7 lines · Keir Fraser
Quoted 4 lines · Xiaowei
For further improvement to reach the effect of 3), e.g. by taking care of cache consistance amongs CPUs, there will be more overhead. And considering the function is called per second, we are hesitating to do this. What's your idea?:)
Quoted 6 lines
Maybe we should see what that overhead is... Also we may not really need to run that function every second.
-- Keir
I measured time_calibration_rendezvous()'s overhead on my machine. It's around 5-6k TSC. And I made anther patch to add loop. The cycle skew introduced by TSC-sync is gone with it. And a bit surprisingly, the expected extra overhead is not even noticeable:)
The new patch is attached.
Thanks,
Xiaowei
xen/arch/x86/time.c +56 -23
@@ -1079,38 +1079,69 @@ static void local_time_calibration(void)*/ struct calibration_rendezvous {cpumask_t cpu_calibration_map; -atomic_t nr_cpus; +atomic_t count_start; +atomic_t count_end;s_time_t master_stime;u64 master_tsc_stamp; }; +#define NR_LOOPS 5 +static void time_calibration_rendezvous (void *_r) { +int i;struct cpu_calibration *c = &this_cpu (cpu_calibration);struct calibration_rendezvous *r = _r;unsigned int total_cpus =cpus_weight (r->cpu_calibration_map); -if (smp_processor_id () ==0 ) - { -while (atomic_read (&r->nr_cpus) != (total_cpus -1 ) ) -cpu_relax (); - r->master_stime =read_platform_stime (); -rdtscll (r->master_tsc_stamp); -mb ();/* write r->master_* /then/ signal */ -atomic_inc (&r->nr_cpus); - c->local_tsc_stamp = r->master_tsc_stamp; - } -else - { -atomic_inc (&r->nr_cpus); -while (atomic_read (&r->nr_cpus) != total_cpus ) -cpu_relax (); -mb ();/* receive signal /then/ read r->master_* */ -if (boot_cpu_has (X86_FEATURE_CONSTANT_TSC ) ) -wrmsrl (MSR_IA32_TSC , r->master_tsc_stamp); -rdtscll (c->local_tsc_stamp); - } - +/* +* Loop is used here to get rid of the cache's side effect to enlarge +* the TSC difference among CPUs. +*/ +for ( i =0 ; i <NR_LOOPS ; i++ ) + { +if (smp_processor_id () ==0 ) + { +while (atomic_read (&r->count_start) != (total_cpus -1 ) ) +mb (); + +if ( r->master_stime ==0 ) + { + r->master_stime =read_platform_stime (); +if (boot_cpu_has (X86_FEATURE_CONSTANT_TSC ) ) +rdtscll (r->master_tsc_stamp); + } +atomic_set (&r->count_end,0 ); +wmb (); +atomic_inc (&r->count_start); + +if (boot_cpu_has (X86_FEATURE_CONSTANT_TSC ) && + i ==NR_LOOPS -1 ) +write_tsc ((u32 )r->master_tsc_stamp, (u32)(r->master_tsc_stamp >>32 )); + +while (atomic_read (&r->count_end) != total_cpus -1 ) +mb (); +atomic_set (&r->count_start,0 ); +wmb (); +atomic_inc (&r->count_end); + } +else + { +atomic_inc (&r->count_start); +while (atomic_read (&r->count_start) != total_cpus ) +mb (); + +if (boot_cpu_has (X86_FEATURE_CONSTANT_TSC ) && + i ==NR_LOOPS -1 ) +write_tsc ((u32 )r->master_tsc_stamp, (u32)(r->master_tsc_stamp >>32 )); + +atomic_inc (&r->count_end); +while (atomic_read (&r->count_end) != total_cpus) +mb (); + } + } + +rdtscll (c->local_tsc_stamp); c->stime_local_stamp =get_s_time (); c->stime_master_stamp = r->master_stime;
@@ -1121,7 +1152,9 @@ static void time_calibration(void *unuse {struct calibration_rendezvous r = { .cpu_calibration_map = cpu_online_map, - .nr_cpus =ATOMIC_INIT (0 ) + .count_start =ATOMIC_INIT (0 ), + .count_end =ATOMIC_INIT (0 ), + .master_stime =0 };/* @wait=1 because we must wait for all cpus before freeing @r. */
_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel