During Live Partition Migration (LPM), it is observed that perf
counter values reports zero post migration completion. However
'perf stat' with workload continues to show counts post migration
since PMU gets disabled/enabled during sched switches. But incase
of system/cpu wide monitoring, zero counts were reported with 'perf
stat' after migration completion.
Example:
./perf stat -e r1001e -I 1000
time counts unit events
1.001010437 22,137,414 r1001e
2.002495447 15,455,821 r1001e
<<>> As seen in next below logs, the counter values shows zero
after migration is completed.
<<>>
86.142535370 129,392,333,440 r1001e
87.144714617 0 r1001e
88.146526636 0 r1001e
89.148085029 0 r1001e
Here PMU is enabled during start of perf session and counter
values are read at intervals. Counters are only disabled at the
end of session. The powerpc mobility code presently does not handle
disabling and enabling back of PMU counters during partition
migration. Also since the PMU register values are not saved/restored
during migration, PMU registers like Monitor Mode Control Register 0
(MMCR0), Monitor Mode Control Register 1 (MMCR1) will not contain
the value it was programmed with. Hence PMU counters will not be
enabled correctly post migration.
Fix this in mobility code by handling disabling and enabling of
PMU in all cpu's before and after migration. Patch introduces two
functions 'mobility_pmu_disable' and 'mobility_pmu_enable'.
mobility_pmu_disable() is called before the processor threads goes
to suspend state so as to disable the PMU counters. And disable is
done only if there are any active events running on that cpu.
mobility_pmu_enable() is called after the migrate is done to enable
back the PMU counters.
Since the performance Monitor counters ( PMCs) are not
saved/restored during LPM, results in PMC value being zero and the
'event->hw.prev_count' being non-zero value. This causes problem
during updation of event->count since we always accumulate
(event->hw.prev_count - PMC value) in event->count. If
event->hw.prev_count is greater PMC value, event->count becomes
negative. To fix this, 'prev_count' also needs to be re-initialised
for all events while enabling back the events. Hence read the
existing events and clear the PMC index (stored in event->hw.idx)
for all events im mobility_pmu_disable. By this way, event count
settings will get re-initialised correctly in power_pmu_enable.
Signed-off-by: Athira Rajeev <redacted>
[ Fixed compilation error reported by kernel test robot ]
Reported-by: kernel test robot <redacted>
---
Changelog:
Change from v2 -> v3:
Addressed review comments from Nicholas Piggin.
- Removed the "migrate" field which was added in initial
patch to address updation of event count settings correctly
in power_pmu_enable. Instead read off existing events in
mobility_pmu_disable before power_pmu_enable.
- Moved the mobility_pmu_disable/enable declaration from
rtas.h to perf event header file.
Addressed review comments from Nathan.
- Moved the mobility function calls from stop_machine
context out to pseries_migrate_partition. Also now this
is a per cpu invocation.
Change from v1 -> v2:
- Moved the mobility_pmu_enable and mobility_pmu_disable
declarations under CONFIG_PPC_PERF_CTRS in rtas.h.
Also included 'asm/rtas.h' in core-book3s to fix the
compilation warning reported by kernel test robot.
arch/powerpc/include/asm/perf_event.h | 8 +++++
arch/powerpc/perf/core-book3s.c | 39 +++++++++++++++++++++++
arch/powerpc/platforms/pseries/mobility.c | 7 ++++
3 files changed, 54 insertions(+)
From: Nicholas Piggin <npiggin@gmail.com> Date: 2021-10-29 06:17:14
Excerpts from Athira Rajeev's message of October 29, 2021 1:05 pm:
During Live Partition Migration (LPM), it is observed that perf
counter values reports zero post migration completion. However
'perf stat' with workload continues to show counts post migration
since PMU gets disabled/enabled during sched switches. But incase
of system/cpu wide monitoring, zero counts were reported with 'perf
stat' after migration completion.
Example:
./perf stat -e r1001e -I 1000
time counts unit events
1.001010437 22,137,414 r1001e
2.002495447 15,455,821 r1001e
<<>> As seen in next below logs, the counter values shows zero
after migration is completed.
<<>>
86.142535370 129,392,333,440 r1001e
87.144714617 0 r1001e
88.146526636 0 r1001e
89.148085029 0 r1001e
This is the output without the patch? After the patch it keeps counting
I suppose? And does the very large count go away too?
quoted hunk
Here PMU is enabled during start of perf session and counter
values are read at intervals. Counters are only disabled at the
end of session. The powerpc mobility code presently does not handle
disabling and enabling back of PMU counters during partition
migration. Also since the PMU register values are not saved/restored
during migration, PMU registers like Monitor Mode Control Register 0
(MMCR0), Monitor Mode Control Register 1 (MMCR1) will not contain
the value it was programmed with. Hence PMU counters will not be
enabled correctly post migration.
Fix this in mobility code by handling disabling and enabling of
PMU in all cpu's before and after migration. Patch introduces two
functions 'mobility_pmu_disable' and 'mobility_pmu_enable'.
mobility_pmu_disable() is called before the processor threads goes
to suspend state so as to disable the PMU counters. And disable is
done only if there are any active events running on that cpu.
mobility_pmu_enable() is called after the migrate is done to enable
back the PMU counters.
Since the performance Monitor counters ( PMCs) are not
saved/restored during LPM, results in PMC value being zero and the
'event->hw.prev_count' being non-zero value. This causes problem
during updation of event->count since we always accumulate
(event->hw.prev_count - PMC value) in event->count. If
event->hw.prev_count is greater PMC value, event->count becomes
negative. To fix this, 'prev_count' also needs to be re-initialised
for all events while enabling back the events. Hence read the
existing events and clear the PMC index (stored in event->hw.idx)
for all events im mobility_pmu_disable. By this way, event count
settings will get re-initialised correctly in power_pmu_enable.
Signed-off-by: Athira Rajeev <redacted>
[ Fixed compilation error reported by kernel test robot ]
Reported-by: kernel test robot <redacted>
---
Changelog:
Change from v2 -> v3:
Addressed review comments from Nicholas Piggin.
- Removed the "migrate" field which was added in initial
patch to address updation of event count settings correctly
in power_pmu_enable. Instead read off existing events in
mobility_pmu_disable before power_pmu_enable.
- Moved the mobility_pmu_disable/enable declaration from
rtas.h to perf event header file.
Addressed review comments from Nathan.
- Moved the mobility function calls from stop_machine
context out to pseries_migrate_partition. Also now this
is a per cpu invocation.
Change from v1 -> v2:
- Moved the mobility_pmu_enable and mobility_pmu_disable
declarations under CONFIG_PPC_PERF_CTRS in rtas.h.
Also included 'asm/rtas.h' in core-book3s to fix the
compilation warning reported by kernel test robot.
arch/powerpc/include/asm/perf_event.h | 8 +++++
arch/powerpc/perf/core-book3s.c | 39 +++++++++++++++++++++++
arch/powerpc/platforms/pseries/mobility.c | 7 ++++
3 files changed, 54 insertions(+)
@@ -631,12 +632,18 @@ static int pseries_migrate_partition(u64 handle)if(ret)returnret;+/* Disable PMU before suspend */+on_each_cpu(&mobility_pmu_disable,NULL,0);
Why was this moved out of stop machine and to an IPI?
My concern would be, what are the other CPUs doing at this time? Is it
possible they could take interrupts and schedule? Could that mess up the
perf state here?
Thanks,
Nick
arch/powerpc/platforms/pseries/mobility.c:670:22: error: 'mobility_pmu_disable' undeclared (first use in this function)
670 | on_each_cpu(&mobility_pmu_disable, NULL, 0);
| ^~~~~~~~~~~~~~~~~~~~
arch/powerpc/platforms/pseries/mobility.c:670:22: note: each undeclared identifier is reported only once for each function it appears in
quoted
arch/powerpc/platforms/pseries/mobility.c:679:22: error: 'mobility_pmu_enable' undeclared (first use in this function)
679 | on_each_cpu(&mobility_pmu_enable, NULL, 0);
| ^~~~~~~~~~~~~~~~~~~
vim +/mobility_pmu_disable +670 arch/powerpc/platforms/pseries/mobility.c
660
661 static int pseries_migrate_partition(u64 handle)
662 {
663 int ret;
664
665 ret = wait_for_vasi_session_suspending(handle);
666 if (ret)
667 return ret;
668
669 /* Disable PMU before suspend */
> 670 on_each_cpu(&mobility_pmu_disable, NULL, 0);
671
672 ret = pseries_suspend(handle);
673 if (ret == 0)
674 post_mobility_fixup();
675 else
676 pseries_cancel_migration(handle, ret);
677
678 /* Enable PMU after resuming */
> 679 on_each_cpu(&mobility_pmu_enable, NULL, 0);
680
681 return ret;
682 }
683
---
0-DAY CI Kernel Test Service, Intel Corporation
https://lists.01.org/hyperkitty/list/kbuild-all@lists.01.org
From: Michael Ellerman <mpe@ellerman.id.au> Date: 2021-10-29 13:16:23
Nicholas Piggin [off-list ref] writes:
Excerpts from Athira Rajeev's message of October 29, 2021 1:05 pm:
quoted
During Live Partition Migration (LPM), it is observed that perf
counter values reports zero post migration completion. However
'perf stat' with workload continues to show counts post migration
since PMU gets disabled/enabled during sched switches. But incase
of system/cpu wide monitoring, zero counts were reported with 'perf
stat' after migration completion.
Example:
./perf stat -e r1001e -I 1000
time counts unit events
1.001010437 22,137,414 r1001e
2.002495447 15,455,821 r1001e
<<>> As seen in next below logs, the counter values shows zero
after migration is completed.
<<>>
86.142535370 129,392,333,440 r1001e
87.144714617 0 r1001e
88.146526636 0 r1001e
89.148085029 0 r1001e
This is the output without the patch? After the patch it keeps counting
I suppose? And does the very large count go away too?
quoted
Here PMU is enabled during start of perf session and counter
values are read at intervals. Counters are only disabled at the
end of session. The powerpc mobility code presently does not handle
disabling and enabling back of PMU counters during partition
migration. Also since the PMU register values are not saved/restored
during migration, PMU registers like Monitor Mode Control Register 0
(MMCR0), Monitor Mode Control Register 1 (MMCR1) will not contain
the value it was programmed with. Hence PMU counters will not be
enabled correctly post migration.
Fix this in mobility code by handling disabling and enabling of
PMU in all cpu's before and after migration. Patch introduces two
functions 'mobility_pmu_disable' and 'mobility_pmu_enable'.
mobility_pmu_disable() is called before the processor threads goes
to suspend state so as to disable the PMU counters. And disable is
done only if there are any active events running on that cpu.
mobility_pmu_enable() is called after the migrate is done to enable
back the PMU counters.
Since the performance Monitor counters ( PMCs) are not
saved/restored during LPM, results in PMC value being zero and the
'event->hw.prev_count' being non-zero value. This causes problem
during updation of event->count since we always accumulate
(event->hw.prev_count - PMC value) in event->count. If
event->hw.prev_count is greater PMC value, event->count becomes
negative. To fix this, 'prev_count' also needs to be re-initialised
for all events while enabling back the events. Hence read the
existing events and clear the PMC index (stored in event->hw.idx)
for all events im mobility_pmu_disable. By this way, event count
settings will get re-initialised correctly in power_pmu_enable.
Signed-off-by: Athira Rajeev <redacted>
[ Fixed compilation error reported by kernel test robot ]
Reported-by: kernel test robot <redacted>
---
Changelog:
Change from v2 -> v3:
Addressed review comments from Nicholas Piggin.
- Removed the "migrate" field which was added in initial
patch to address updation of event count settings correctly
in power_pmu_enable. Instead read off existing events in
mobility_pmu_disable before power_pmu_enable.
- Moved the mobility_pmu_disable/enable declaration from
rtas.h to perf event header file.
Addressed review comments from Nathan.
- Moved the mobility function calls from stop_machine
context out to pseries_migrate_partition. Also now this
is a per cpu invocation.
Change from v1 -> v2:
- Moved the mobility_pmu_enable and mobility_pmu_disable
declarations under CONFIG_PPC_PERF_CTRS in rtas.h.
Also included 'asm/rtas.h' in core-book3s to fix the
compilation warning reported by kernel test robot.
arch/powerpc/include/asm/perf_event.h | 8 +++++
arch/powerpc/perf/core-book3s.c | 39 +++++++++++++++++++++++
arch/powerpc/platforms/pseries/mobility.c | 7 ++++
3 files changed, 54 insertions(+)
These are only needed if pseries is built, so should be inside a PSERIES
ifdef.
This function should handle iterating over CPUs, that shouldn't be left
up to the mobility.c code.
And the names should be something like pmu_start_migration(),
pmu_finish_migration().
quoted
+void mobility_pmu_disable(void *unused)
+{
+ struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
+ struct perf_event *event;
+
+ if (cpuhw->n_events != 0) {
+ int i;
+
+ power_pmu_disable(NULL);
+ /*
+ * Read off any pre-existing events because the register
+ * values may not be migrated.
+ */
+ for (i = 0; i < cpuhw->n_events; ++i) {
+ event = cpuhw->event[i];
+ if (event->hw.idx) {
+ power_pmu_read(event);
+ event->hw.idx = 0;
+ }
+ }
+ }
+}
+
/*
* Re-enable all events if disable == 0.
* If we were previously disabled and events were added, then
@@ -1515,6 +1542,18 @@ static void power_pmu_enable(struct pmu *pmu) local_irq_restore(flags); }+/*+ * Called from pseries_migrate_partition() function+ * after migration, from powerpc/mobility code.+ */+void mobility_pmu_enable(void *unused)+{+ struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);++ cpuhw->n_added = cpuhw->n_events;+ power_pmu_enable(NULL);+}+ static int collect_events(struct perf_event *group, int max_count, struct perf_event *ctrs[], u64 *events, unsigned int *flags)
@@ -631,12 +632,18 @@ static int pseries_migrate_partition(u64 handle)if(ret)returnret;+/* Disable PMU before suspend */+on_each_cpu(&mobility_pmu_disable,NULL,0);
Why was this moved out of stop machine and to an IPI?
My concern would be, what are the other CPUs doing at this time? Is it
possible they could take interrupts and schedule? Could that mess up the
perf state here?
pseries_migrate_partition() is called directly from migration_store(),
which is the sysfs store function, which can be called concurrently by
different CPUs.
It's also potentially called from rtas_syscall_dispatch_ibm_suspend_me(),
from sys_rtas(), again with no locking.
So we could have two CPUs calling into here at the same time, which
might not crash, but is unlikely to work well.
I think the lack of locking might have been OK in the past because only
one CPU will successfully get the other CPUs to call do_join() in
pseries_suspend(). But I could be wrong.
Anyway, now that we're mutating the PMU state before suspending we need
to be more careful. So I think we need a lock around the whole sequence.
cheers
On 29-Oct-2021, at 11:46 AM, Nicholas Piggin [off-list ref] wrote:
Excerpts from Athira Rajeev's message of October 29, 2021 1:05 pm:
quoted
During Live Partition Migration (LPM), it is observed that perf
counter values reports zero post migration completion. However
'perf stat' with workload continues to show counts post migration
since PMU gets disabled/enabled during sched switches. But incase
of system/cpu wide monitoring, zero counts were reported with 'perf
stat' after migration completion.
Example:
./perf stat -e r1001e -I 1000
time counts unit events
1.001010437 22,137,414 r1001e
2.002495447 15,455,821 r1001e
<<>> As seen in next below logs, the counter values shows zero
after migration is completed.
<<>>
86.142535370 129,392,333,440 r1001e
87.144714617 0 r1001e
88.146526636 0 r1001e
89.148085029 0 r1001e
This is the output without the patch? After the patch it keeps counting
I suppose? And does the very large count go away too?
Hi Nick,
Thanks for the review comments.
The output is without the patch. After the patch it keeps counting and negative count goes away.
My bad, missed to keep both results in present version. I will add the before and after patch results in next version.
quoted
Here PMU is enabled during start of perf session and counter
values are read at intervals. Counters are only disabled at the
end of session. The powerpc mobility code presently does not handle
disabling and enabling back of PMU counters during partition
migration. Also since the PMU register values are not saved/restored
during migration, PMU registers like Monitor Mode Control Register 0
(MMCR0), Monitor Mode Control Register 1 (MMCR1) will not contain
the value it was programmed with. Hence PMU counters will not be
enabled correctly post migration.
Fix this in mobility code by handling disabling and enabling of
PMU in all cpu's before and after migration. Patch introduces two
functions 'mobility_pmu_disable' and 'mobility_pmu_enable'.
mobility_pmu_disable() is called before the processor threads goes
to suspend state so as to disable the PMU counters. And disable is
done only if there are any active events running on that cpu.
mobility_pmu_enable() is called after the migrate is done to enable
back the PMU counters.
Since the performance Monitor counters ( PMCs) are not
saved/restored during LPM, results in PMC value being zero and the
'event->hw.prev_count' being non-zero value. This causes problem
during updation of event->count since we always accumulate
(event->hw.prev_count - PMC value) in event->count. If
event->hw.prev_count is greater PMC value, event->count becomes
negative. To fix this, 'prev_count' also needs to be re-initialised
for all events while enabling back the events. Hence read the
existing events and clear the PMC index (stored in event->hw.idx)
for all events im mobility_pmu_disable. By this way, event count
settings will get re-initialised correctly in power_pmu_enable.
Signed-off-by: Athira Rajeev <redacted>
[ Fixed compilation error reported by kernel test robot ]
Reported-by: kernel test robot <redacted>
---
Changelog:
Change from v2 -> v3:
Addressed review comments from Nicholas Piggin.
- Removed the "migrate" field which was added in initial
patch to address updation of event count settings correctly
in power_pmu_enable. Instead read off existing events in
mobility_pmu_disable before power_pmu_enable.
- Moved the mobility_pmu_disable/enable declaration from
rtas.h to perf event header file.
Addressed review comments from Nathan.
- Moved the mobility function calls from stop_machine
context out to pseries_migrate_partition. Also now this
is a per cpu invocation.
Change from v1 -> v2:
- Moved the mobility_pmu_enable and mobility_pmu_disable
declarations under CONFIG_PPC_PERF_CTRS in rtas.h.
Also included 'asm/rtas.h' in core-book3s to fix the
compilation warning reported by kernel test robot.
arch/powerpc/include/asm/perf_event.h | 8 +++++
arch/powerpc/perf/core-book3s.c | 39 +++++++++++++++++++++++
arch/powerpc/platforms/pseries/mobility.c | 7 ++++
3 files changed, 54 insertions(+)
local_irq_restore(flags);
}
+/*
+ * Called from pseries_migrate_partition() function
+ * before migration, from powerpc/mobility code.
+ */
+void mobility_pmu_disable(void *unused)
+{
+ struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
+ struct perf_event *event;
+
+ if (cpuhw->n_events != 0) {
+ int i;
+
+ power_pmu_disable(NULL);
+ /*
+ * Read off any pre-existing events because the register
+ * values may not be migrated.
+ */
+ for (i = 0; i < cpuhw->n_events; ++i) {
+ event = cpuhw->event[i];
+ if (event->hw.idx) {
+ power_pmu_read(event);
+ event->hw.idx = 0;
+ }
+ }
+ }
+}
+
/*
* Re-enable all events if disable == 0.
* If we were previously disabled and events were added, then
@@ -631,12 +632,18 @@ static int pseries_migrate_partition(u64 handle)
if (ret)
return ret;
+ /* Disable PMU before suspend */
+ on_each_cpu(&mobility_pmu_disable, NULL, 0);
Why was this moved out of stop machine and to an IPI?
My concern would be, what are the other CPUs doing at this time? Is it
possible they could take interrupts and schedule? Could that mess up the
perf state here?
Thanks,
Nick
From: Nicholas Piggin <npiggin@gmail.com> Date: 2021-11-02 11:36:42
Excerpts from Michael Ellerman's message of October 29, 2021 11:15 pm:
Nicholas Piggin [off-list ref] writes:
quoted
Excerpts from Athira Rajeev's message of October 29, 2021 1:05 pm:
quoted
@@ -631,12 +632,18 @@ static int pseries_migrate_partition(u64 handle) if (ret) return ret;+ /* Disable PMU before suspend */+ on_each_cpu(&mobility_pmu_disable, NULL, 0);
Why was this moved out of stop machine and to an IPI?
My concern would be, what are the other CPUs doing at this time? Is it
possible they could take interrupts and schedule? Could that mess up the
perf state here?
pseries_migrate_partition() is called directly from migration_store(),
which is the sysfs store function, which can be called concurrently by
different CPUs.
It's also potentially called from rtas_syscall_dispatch_ibm_suspend_me(),
from sys_rtas(), again with no locking.
So we could have two CPUs calling into here at the same time, which
might not crash, but is unlikely to work well.
I think the lack of locking might have been OK in the past because only
one CPU will successfully get the other CPUs to call do_join() in
pseries_suspend(). But I could be wrong.
Anyway, now that we're mutating the PMU state before suspending we need
to be more careful. So I think we need a lock around the whole sequence.
My concern is still that we wouldn't necessarily have the other CPUs
under control at that point even if we serialize the migrate path.
They could take interrupts, possibly call into perf subsystem after
the mobility_pmu_disable (e.g., via syscall or context switch) which
might mess things up.
I think the stop machine is a reasonable place for the code in this
case. It's a low level disabling of hardware facility and saving off
registers.
Thanks,
Nick
On 29-Oct-2021, at 6:45 PM, Michael Ellerman [off-list ref] wrote:
Nicholas Piggin <npiggin@gmail.com <mailto:npiggin@gmail.com>> writes:
quoted
Excerpts from Athira Rajeev's message of October 29, 2021 1:05 pm:
quoted
During Live Partition Migration (LPM), it is observed that perf
counter values reports zero post migration completion. However
'perf stat' with workload continues to show counts post migration
since PMU gets disabled/enabled during sched switches. But incase
of system/cpu wide monitoring, zero counts were reported with 'perf
stat' after migration completion.
Example:
./perf stat -e r1001e -I 1000
time counts unit events
1.001010437 22,137,414 r1001e
2.002495447 15,455,821 r1001e
<<>> As seen in next below logs, the counter values shows zero
after migration is completed.
<<>>
86.142535370 129,392,333,440 r1001e
87.144714617 0 r1001e
88.146526636 0 r1001e
89.148085029 0 r1001e
This is the output without the patch? After the patch it keeps counting
I suppose? And does the very large count go away too?
quoted
Here PMU is enabled during start of perf session and counter
values are read at intervals. Counters are only disabled at the
end of session. The powerpc mobility code presently does not handle
disabling and enabling back of PMU counters during partition
migration. Also since the PMU register values are not saved/restored
during migration, PMU registers like Monitor Mode Control Register 0
(MMCR0), Monitor Mode Control Register 1 (MMCR1) will not contain
the value it was programmed with. Hence PMU counters will not be
enabled correctly post migration.
Fix this in mobility code by handling disabling and enabling of
PMU in all cpu's before and after migration. Patch introduces two
functions 'mobility_pmu_disable' and 'mobility_pmu_enable'.
mobility_pmu_disable() is called before the processor threads goes
to suspend state so as to disable the PMU counters. And disable is
done only if there are any active events running on that cpu.
mobility_pmu_enable() is called after the migrate is done to enable
back the PMU counters.
Since the performance Monitor counters ( PMCs) are not
saved/restored during LPM, results in PMC value being zero and the
'event->hw.prev_count' being non-zero value. This causes problem
during updation of event->count since we always accumulate
(event->hw.prev_count - PMC value) in event->count. If
event->hw.prev_count is greater PMC value, event->count becomes
negative. To fix this, 'prev_count' also needs to be re-initialised
for all events while enabling back the events. Hence read the
existing events and clear the PMC index (stored in event->hw.idx)
for all events im mobility_pmu_disable. By this way, event count
settings will get re-initialised correctly in power_pmu_enable.
Signed-off-by: Athira Rajeev <redacted>
[ Fixed compilation error reported by kernel test robot ]
Reported-by: kernel test robot <redacted>
---
Changelog:
Change from v2 -> v3:
Addressed review comments from Nicholas Piggin.
- Removed the "migrate" field which was added in initial
patch to address updation of event count settings correctly
in power_pmu_enable. Instead read off existing events in
mobility_pmu_disable before power_pmu_enable.
- Moved the mobility_pmu_disable/enable declaration from
rtas.h to perf event header file.
Addressed review comments from Nathan.
- Moved the mobility function calls from stop_machine
context out to pseries_migrate_partition. Also now this
is a per cpu invocation.
Change from v1 -> v2:
- Moved the mobility_pmu_enable and mobility_pmu_disable
declarations under CONFIG_PPC_PERF_CTRS in rtas.h.
Also included 'asm/rtas.h' in core-book3s to fix the
compilation warning reported by kernel test robot.
arch/powerpc/include/asm/perf_event.h | 8 +++++
arch/powerpc/perf/core-book3s.c | 39 +++++++++++++++++++++++
arch/powerpc/platforms/pseries/mobility.c | 7 ++++
3 files changed, 54 insertions(+)
local_irq_restore(flags);
}
+/*
+ * Called from pseries_migrate_partition() function
+ * before migration, from powerpc/mobility code.
+ */
These are only needed if pseries is built, so should be inside a PSERIES
ifdef.
Sure mpe, will address this change in next version
This function should handle iterating over CPUs, that shouldn't be left
up to the mobility.c code.
And the names should be something like pmu_start_migration(),
pmu_finish_migration().
Ok, will change
quoted
quoted
+void mobility_pmu_disable(void *unused)
+{
+ struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
+ struct perf_event *event;
+
+ if (cpuhw->n_events != 0) {
+ int i;
+
+ power_pmu_disable(NULL);
+ /*
+ * Read off any pre-existing events because the register
+ * values may not be migrated.
+ */
+ for (i = 0; i < cpuhw->n_events; ++i) {
+ event = cpuhw->event[i];
+ if (event->hw.idx) {
+ power_pmu_read(event);
+ event->hw.idx = 0;
+ }
+ }
+ }
+}
+
/*
* Re-enable all events if disable == 0.
* If we were previously disabled and events were added, then
@@ -631,12 +632,18 @@ static int pseries_migrate_partition(u64 handle)
if (ret)
return ret;
+ /* Disable PMU before suspend */
+ on_each_cpu(&mobility_pmu_disable, NULL, 0);
Why was this moved out of stop machine and to an IPI?
My concern would be, what are the other CPUs doing at this time? Is it
possible they could take interrupts and schedule? Could that mess up the
perf state here?
pseries_migrate_partition() is called directly from migration_store(),
which is the sysfs store function, which can be called concurrently by
different CPUs.
It's also potentially called from rtas_syscall_dispatch_ibm_suspend_me(),
from sys_rtas(), again with no locking.
So we could have two CPUs calling into here at the same time, which
might not crash, but is unlikely to work well.
I think the lack of locking might have been OK in the past because only
one CPU will successfully get the other CPUs to call do_join() in
pseries_suspend(). But I could be wrong.
Anyway, now that we're mutating the PMU state before suspending we need
to be more careful. So I think we need a lock around the whole sequence.
Thanks for the review comments.
The PMU callback invocations were from inside stop machine ( inside “do_join” ) in initial version.
Moved these out to pseries_migrate_partition as an IPI so as to minimise calls to other sub-system from “do_join” context, as pointed by Nathan.
But if it is not safe to have per-cpu invocation from pseries_migrate_partition, should we handle this under the interrupt disable path in do_join itself ( as in the initial version ) ?
Please share your thoughts/suggestions.
Thanks
Athira
Excerpts from Michael Ellerman's message of October 29, 2021 11:15 pm:
quoted
Nicholas Piggin [off-list ref] writes:
quoted
Excerpts from Athira Rajeev's message of October 29, 2021 1:05 pm:
quoted
@@ -631,12 +632,18 @@ static int pseries_migrate_partition(u64 handle) if (ret) return ret;+ /* Disable PMU before suspend */+ on_each_cpu(&mobility_pmu_disable, NULL, 0);
Why was this moved out of stop machine and to an IPI?
My concern would be, what are the other CPUs doing at this time? Is it
possible they could take interrupts and schedule? Could that mess up the
perf state here?
pseries_migrate_partition() is called directly from migration_store(),
which is the sysfs store function, which can be called concurrently by
different CPUs.
It's also potentially called from rtas_syscall_dispatch_ibm_suspend_me(),
from sys_rtas(), again with no locking.
So we could have two CPUs calling into here at the same time, which
might not crash, but is unlikely to work well.
I think the lack of locking might have been OK in the past because only
one CPU will successfully get the other CPUs to call do_join() in
pseries_suspend(). But I could be wrong.
Anyway, now that we're mutating the PMU state before suspending we need
to be more careful. So I think we need a lock around the whole
sequence.
Regardless of the outcome here, generally agreed that some serialization
should be imposed in this path. The way the platform works (and some
extra measures by the drmgr utility) make it so that this code isn't
entered concurrently in usual operation, but it's possible to make it
happen if you are root.
A file-static mutex should be OK.
My concern is still that we wouldn't necessarily have the other CPUs
under control at that point even if we serialize the migrate path.
They could take interrupts, possibly call into perf subsystem after
the mobility_pmu_disable (e.g., via syscall or context switch) which
might mess things up.
I think the stop machine is a reasonable place for the code in this
case. It's a low level disabling of hardware facility and saving off
registers.
That makes sense, but I can't help feeling concerned still. For this to
be safe, power_pmu_enable() and power_pmu_disable() must never sleep or
re-enable interrupts or send IPIs. I don't see anything obviously unsafe
right now, but is that already part of their contract? Is there much
risk they could change in the future to violate those constraints?
That aside, the proposed change seems like we would be hacking around a
more generic perf/pmu limitation in a powerpc-specific way. I see the
same behavior on x86 across suspend/resume.
# perf stat -a -e cache-misses -I 1000 & sleep 2 ; systemctl suspend ; sleep 20 ; kill $(jobs -p)
[1] 189806
# time counts unit events
1.000501710 9,983,649 cache-misses
2.002620321 14,131,072 cache-misses
3.004579071 23,010,971 cache-misses
9.971854783 140,737,491,680,853 cache-misses
10.982669250 0 cache-misses
11.984660498 0 cache-misses
12.986648392 0 cache-misses
13.988561766 0 cache-misses
14.992670615 0 cache-misses
15.994938111 0 cache-misses
16.996703952 0 cache-misses
17.999092812 0 cache-misses
19.000602677 0 cache-misses
20.003272216 0 cache-misses
21.004770295 0 cache-misses
# uname -r
5.13.19-100.fc33.x86_64
From: Michael Ellerman <mpe@ellerman.id.au> Date: 2021-11-04 05:55:51
Nathan Lynch [off-list ref] writes:
Nicholas Piggin [off-list ref] writes:
quoted
Excerpts from Michael Ellerman's message of October 29, 2021 11:15 pm:
quoted
Nicholas Piggin [off-list ref] writes:
quoted
Excerpts from Athira Rajeev's message of October 29, 2021 1:05 pm:
quoted
@@ -631,12 +632,18 @@ static int pseries_migrate_partition(u64 handle) if (ret) return ret;+ /* Disable PMU before suspend */+ on_each_cpu(&mobility_pmu_disable, NULL, 0);
Why was this moved out of stop machine and to an IPI?
My concern would be, what are the other CPUs doing at this time? Is it
possible they could take interrupts and schedule? Could that mess up the
perf state here?
pseries_migrate_partition() is called directly from migration_store(),
which is the sysfs store function, which can be called concurrently by
different CPUs.
It's also potentially called from rtas_syscall_dispatch_ibm_suspend_me(),
from sys_rtas(), again with no locking.
So we could have two CPUs calling into here at the same time, which
might not crash, but is unlikely to work well.
I think the lack of locking might have been OK in the past because only
one CPU will successfully get the other CPUs to call do_join() in
pseries_suspend(). But I could be wrong.
Anyway, now that we're mutating the PMU state before suspending we need
to be more careful. So I think we need a lock around the whole
sequence.
Regardless of the outcome here, generally agreed that some serialization
should be imposed in this path. The way the platform works (and some
extra measures by the drmgr utility) make it so that this code isn't
entered concurrently in usual operation, but it's possible to make it
happen if you are root.
Yeah I agree it's unlikely to be a problem in practice.
A file-static mutex should be OK.
Ack.
quoted
My concern is still that we wouldn't necessarily have the other CPUs
under control at that point even if we serialize the migrate path.
They could take interrupts, possibly call into perf subsystem after
the mobility_pmu_disable (e.g., via syscall or context switch) which
might mess things up.
I think the stop machine is a reasonable place for the code in this
case. It's a low level disabling of hardware facility and saving off
registers.
That makes sense, but I can't help feeling concerned still. For this to
be safe, power_pmu_enable() and power_pmu_disable() must never sleep or
re-enable interrupts or send IPIs. I don't see anything obviously unsafe
right now, but is that already part of their contract? Is there much
risk they could change in the future to violate those constraints?
That aside, the proposed change seems like we would be hacking around a
more generic perf/pmu limitation in a powerpc-specific way. I see the
same behavior on x86 across suspend/resume.
# perf stat -a -e cache-misses -I 1000 & sleep 2 ; systemctl suspend ; sleep 20 ; kill $(jobs -p)
[1] 189806
# time counts unit events
1.000501710 9,983,649 cache-misses
2.002620321 14,131,072 cache-misses
3.004579071 23,010,971 cache-misses
9.971854783 140,737,491,680,853 cache-misses
10.982669250 0 cache-misses
11.984660498 0 cache-misses
12.986648392 0 cache-misses
13.988561766 0 cache-misses
14.992670615 0 cache-misses
15.994938111 0 cache-misses
16.996703952 0 cache-misses
17.999092812 0 cache-misses
19.000602677 0 cache-misses
20.003272216 0 cache-misses
21.004770295 0 cache-misses
# uname -r
5.13.19-100.fc33.x86_64
That is interesting.
Athira, I guess we should bring that to the perf maintainers and see if
there's any interest in solving the issue in a generic fashion.
cheers
From: Nicholas Piggin <npiggin@gmail.com> Date: 2021-11-05 01:46:56
Excerpts from Nathan Lynch's message of November 4, 2021 7:11 am:
Nicholas Piggin [off-list ref] writes:
quoted
Excerpts from Michael Ellerman's message of October 29, 2021 11:15 pm:
quoted
Nicholas Piggin [off-list ref] writes:
quoted
Excerpts from Athira Rajeev's message of October 29, 2021 1:05 pm:
quoted
@@ -631,12 +632,18 @@ static int pseries_migrate_partition(u64 handle) if (ret) return ret;+ /* Disable PMU before suspend */+ on_each_cpu(&mobility_pmu_disable, NULL, 0);
Why was this moved out of stop machine and to an IPI?
My concern would be, what are the other CPUs doing at this time? Is it
possible they could take interrupts and schedule? Could that mess up the
perf state here?
pseries_migrate_partition() is called directly from migration_store(),
which is the sysfs store function, which can be called concurrently by
different CPUs.
It's also potentially called from rtas_syscall_dispatch_ibm_suspend_me(),
from sys_rtas(), again with no locking.
So we could have two CPUs calling into here at the same time, which
might not crash, but is unlikely to work well.
I think the lack of locking might have been OK in the past because only
one CPU will successfully get the other CPUs to call do_join() in
pseries_suspend(). But I could be wrong.
Anyway, now that we're mutating the PMU state before suspending we need
to be more careful. So I think we need a lock around the whole
sequence.
Regardless of the outcome here, generally agreed that some serialization
should be imposed in this path. The way the platform works (and some
extra measures by the drmgr utility) make it so that this code isn't
entered concurrently in usual operation, but it's possible to make it
happen if you are root.
A file-static mutex should be OK.
quoted
My concern is still that we wouldn't necessarily have the other CPUs
under control at that point even if we serialize the migrate path.
They could take interrupts, possibly call into perf subsystem after
the mobility_pmu_disable (e.g., via syscall or context switch) which
might mess things up.
I think the stop machine is a reasonable place for the code in this
case. It's a low level disabling of hardware facility and saving off
registers.
That makes sense, but I can't help feeling concerned still. For this to
be safe, power_pmu_enable() and power_pmu_disable() must never sleep or
re-enable interrupts or send IPIs. I don't see anything obviously unsafe
right now, but is that already part of their contract?
Yes that would have to be. That's much the same as an IPI handler. Maybe
stop machine has a few other things to worry about but I'm not sure that
would get in the way. Just have a note in the implementation or name to
remind it is stop machine context.
Is there much
risk they could change in the future to violate those constraints?
My guess is if it needed to do anything much more complicated then you
would want to shut down perf more completely in a way that other APIs
know about.
That aside, the proposed change seems like we would be hacking around a
more generic perf/pmu limitation in a powerpc-specific way. I see the
same behavior on x86 across suspend/resume.
Arguably it's fixing up a x86 suspend/resume bug that might be fixed
in the same way as this -- presumably it's not restoring PMU registers
at resume.
But if they did something generic in perf core, this code could be
changed over to use it quite easily I would hope. But pinging the
other perf devs to check would be a good idea.
Thanks,
Nick