[ Upstream commit a3c3c66670cee11eb13aa43905904bf29cb92d32 ]
The perf events code fails to account for total_time_enabled of
inactive events.
Here is a failure case for accounting total_time_enabled for
CPU PMU events:
sudo ./perf stat -vvv -e armv8_pmuv3_0/event=0x08/ -e armv8_pmuv3_1/event=0x08/ -- stress-ng --pthread=2 -t 2s
...
armv8_pmuv3_0/event=0x08/: 1138698008 2289429840 2174835740
armv8_pmuv3_1/event=0x08/: 1826791390 1950025700 847648440
` ` `
` ` > total_time_running with child
` > total_time_enabled with child
> count with child
Performance counter stats for 'stress-ng --pthread=2 -t 2s':
1,138,698,008 armv8_pmuv3_0/event=0x08/ (94.99%)
1,826,791,390 armv8_pmuv3_1/event=0x08/ (43.47%)
The two events above are opened on two different CPU PMUs, for example,
each event is opened for a cluster in an Arm big.LITTLE system, they
will never run on the same CPU. In theory, the total enabled time should
be same for both events, as two events are opened and closed together.
As the result show, the two events' total enabled time including
child event is different (2289429840 vs 1950025700).
This is because child events are not accounted properly
if a event is INACTIVE state when the task exits:
perf_event_exit_event()
`> perf_remove_from_context()
`> __perf_remove_from_context()
`> perf_child_detach() -> Accumulate child_total_time_enabled
`> list_del_event() -> Update child event's time
The problem is the time accumulation happens prior to child event's
time updating. Thus, it misses to account the last period's time when
the event exits.
The perf core layer follows the rule that timekeeping is tied to state
change. To address the issue, make __perf_remove_from_context()
handle the task exit case by passing 'DETACH_EXIT' to it and
invoke perf_event_state() for state alongside with accounting the time.
Then, perf_child_detach() populates the time into the parent's time metrics.
After this patch, the bug is fixed:
sudo ./perf stat -vvv -e armv8_pmuv3_0/event=0x08/ -e armv8_pmuv3_1/event=0x08/ -- stress-ng --pthread=2 -t 10s
...
armv8_pmuv3_0/event=0x08/: 15396770398 32157963940 21898169000
armv8_pmuv3_1/event=0x08/: 22428964974 32157963940 10259794940
Performance counter stats for 'stress-ng --pthread=2 -t 10s':
15,396,770,398 armv8_pmuv3_0/event=0x08/ (68.10%)
22,428,964,974 armv8_pmuv3_1/event=0x08/ (31.90%)
[ mingo: Clarified the changelog. ]
Fixes: ef54c1a476aef ("perf: Rework perf_event_exit_event()")
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Yeoreum Yun <yeoreum.yun@arm.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Tested-by: Leo Yan <leo.yan@arm.com>
Link: https://lore.kernel.org/r/20250326082003.1630986-1-yeoreum.yun@arm.com
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit d57e94f5b891925e4f2796266eba31edd5a01903 ]
To save/restore LBR call stack data in system-wide mode, the task_struct
information is required.
Extend the parameters of sched_task() to supply task_struct information.
When schedule in, the LBR call stack data for new task will be restored.
When schedule out, the LBR call stack data for old task will be saved.
Only need to pass the required task_struct information.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20250314172700.438923-4-kan.liang@linux.intel.com
Stable-dep-of: 3cec9fd03543 ("perf/x86/lbr: Fix shorter LBRs call stacks for the system-wide mode")
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit cb4369129339060218baca718a578bb0b826e734 ]
Some PMU specific data has to be saved/restored during context switch,
e.g. LBR call stack data. Currently, the data is saved in event context
structure, but only for per-process event. For system-wide event,
because of missing the LBR call stack data after context switch, LBR
callstacks are always shorter in comparison to per-process mode.
For example,
Per-process mode:
$perf record --call-graph lbr -- taskset -c 0 ./tchain_edit
- 99.90% 99.86% tchain_edit tchain_edit [.] f3
99.86% _start
__libc_start_main
generic_start_main
main
f1
- f2
f3
System-wide mode:
$perf record --call-graph lbr -a -- taskset -c 0 ./tchain_edit
- 99.88% 99.82% tchain_edit tchain_edit [.] f3
- 62.02% main
f1
f2
f3
- 28.83% f1
- f2
f3
- 28.83% f1
- f2
f3
- 8.88% generic_start_main
main
f1
f2
f3
It isn't practical to simply allocate the data for system-wide event in
CPU context structure for all tasks. We have no idea which CPU a task
will be scheduled to. The duplicated LBR data has to be maintained on
every CPU context structure. That's a huge waste. Otherwise, the LBR
data still lost if the task is scheduled to another CPU.
Save the pmu specific data in task_struct. The size of pmu specific data
is 788 bytes for LBR call stack. Usually, the overall amount of threads
doesn't exceed a few thousands. For 10K threads, keeping LBR data would
consume additional ~8MB. The additional space will only be allocated
during LBR call stack monitoring. It will be released when the
monitoring is finished.
Furthermore, moving task_ctx_data from perf_event_context to task_struct
can reduce complexity and make things clearer. E.g. perf doesn't need to
swap task_ctx_data on optimized context switch path.
This patch set is just the first step. There could be other
optimization/extension on top of this patch set. E.g. for cgroup
profiling, perf just needs to save/store the LBR call stack information
for tasks in specific cgroup. That could reduce the additional space.
Also, the LBR call stack can be available for software events, or allow
even debugging use cases, like LBRs on crash later.
Because of the alignment requirement of Intel Arch LBR, the Kmem cache
is used to allocate the PMU specific data. It's required when child task
allocates the space. Save it in struct perf_ctx_data.
The refcount in struct perf_ctx_data is used to track the users of pmu
specific data.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Alexey Budankov <alexey.budankov@linux.intel.com>
Link: https://lore.kernel.org/r/20250314172700.438923-1-kan.liang@linux.intel.com
Stable-dep-of: 3cec9fd03543 ("perf/x86/lbr: Fix shorter LBRs call stacks for the system-wide mode")
Signed-off-by: Sasha Levin <sashal@kernel.org>
There is a fairly obvious race between perf_init_event() doing idr_find() and perf_pmu_register() doing idr_alloc() with an incompletely initialized PMU pointer. Avoid by doing idr_alloc() on a NULL pointer to register the id, and swizzling the real struct pmu pointer at the end using idr_replace(). Also making sure to not set struct pmu members after publishing the struct pmu, duh. [ introduce idr_cmpxchg() in order to better handle the idr_replace() error case -- if it were to return an unexpected pointer, it will already have replaced the value and there is no going back. ] Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Link: https://lore.kernel.org/r/20241104135517.858805880@infradead.org
Commit a63fbed776c7 ("perf/tracing/cpuhotplug: Fix locking order")
placed pmus_lock inside pmus_srcu, this makes perf_pmu_unregister()
trip lockdep.
Move the locking about such that only pmu_idr and pmus (list) are
modified while holding pmus_lock. This avoids doing synchronize_srcu()
while holding pmus_lock and all is well again.
Fixes: a63fbed776c7 ("perf/tracing/cpuhotplug: Fix locking order")
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lore.kernel.org/r/20241104135517.679556858@infradead.org
A low attr::freq value cannot be set via IOC_PERIOD on some platforms.
The perf_event_check_period() introduced in:
81ec3f3c4c4d ("perf/x86: Add check_period PMU callback")
was intended to check the period, rather than the frequency.
A low frequency may be mistakenly rejected by limit_period().
Fix it.
Fixes: 81ec3f3c4c4d ("perf/x86: Add check_period PMU callback")
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Ravi Bangoria <ravi.bangoria@amd.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: stable@vger.kernel.org
Link: https://lore.kernel.org/r/20250117151913.3043942-2-kan.liang@linux.intel.com
Closes: https://lore.kernel.org/lkml/20250115154949.3147-1-ravi.bangoria@amd.com/
Syskaller triggers a warning due to prev_epc->pmu != next_epc->pmu in
perf_event_swap_task_ctx_data(). vmcore shows that two lists have the same
perf_event_pmu_context, but not in the same order.
The problem is that the order of pmu_ctx_list for the parent is impacted by
the time when an event/PMU is added. While the order for a child is
impacted by the event order in the pinned_groups and flexible_groups. So
the order of pmu_ctx_list in the parent and child may be different.
To fix this problem, insert the perf_event_pmu_context to its proper place
after iteration of the pmu_ctx_list.
The follow testcase can trigger above warning:
# perf record -e cycles --call-graph lbr -- taskset -c 3 ./a.out &
# perf stat -e cpu-clock,cs -p xxx // xxx is the pid of a.out
test.c
void main() {
int count = 0;
pid_t pid;
printf("%d running\n", getpid());
sleep(30);
printf("running\n");
pid = fork();
if (pid == -1) {
printf("fork error\n");
return;
}
if (pid == 0) {
while (1) {
count++;
}
} else {
while (1) {
count++;
}
}
}
The testcase first opens an LBR event, so it will allocate task_ctx_data,
and then open tracepoint and software events, so the parent context will
have 3 different perf_event_pmu_contexts. On inheritance, child ctx will
insert the perf_event_pmu_context in another order and the warning will
trigger.
[ mingo: Tidied up the changelog. ]
Fixes: bd2756811766 ("perf: Rewrite core context handling")
Signed-off-by: Luo Gengkun <luogengkun@huaweicloud.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Kan Liang <kan.liang@linux.intel.com>
Link: https://lore.kernel.org/r/20250122073356.1824736-1-luogengkun@huaweicloud.com
The perf_iterate_ctx() function performs RCU list traversal but
currently lacks RCU read lock protection. This causes lockdep warnings
when running perf probe with unshare(1) under CONFIG_PROVE_RCU_LIST=y:
WARNING: suspicious RCU usage
kernel/events/core.c:8168 RCU-list traversed in non-reader section!!
Call Trace:
lockdep_rcu_suspicious
? perf_event_addr_filters_apply
perf_iterate_ctx
perf_event_exec
begin_new_exec
? load_elf_phdrs
load_elf_binary
? lock_acquire
? find_held_lock
? bprm_execve
bprm_execve
do_execveat_common.isra.0
__x64_sys_execve
do_syscall_64
entry_SYSCALL_64_after_hwframe
This protection was previously present but was removed in commit
bd2756811766 ("perf: Rewrite core context handling"). Add back the
necessary rcu_read_lock()/rcu_read_unlock() pair around
perf_iterate_ctx() call in perf_event_exec().
[ mingo: Use scoped_guard() as suggested by Peter ]
Fixes: bd2756811766 ("perf: Rewrite core context handling")
Signed-off-by: Breno Leitao <leitao@debian.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: stable@vger.kernel.org
Link: https://lore.kernel.org/r/20250117-fix_perf_rcu-v1-1-13cb9210fc6a@debian.org
We are adjusting struct page to make it smaller, removing unneeded fields which correctly belong to struct folio. Two of those fields are page->index and page->mapping. Perf is currently making use of both of these. This is unnecessary. This patch eliminates this. Perf establishes its own internally controlled memory-mapped pages using vm_ops hooks. The first page in the mapping is the read/write user control page, and the rest of the mapping consists of read-only pages. The VMA is backed by kernel memory either from the buddy allocator or vmalloc depending on configuration. It is intended to be mapped read/write, but because it has a page_mkwrite() hook, vma_wants_writenotify() indicates that it should be mapped read-only. When a write fault occurs, the provided page_mkwrite() hook, perf_mmap_fault() (doing double duty handing faults as well) uses the vmf->pgoff field to determine if this is the first page, allowing for the desired read/write first page, read-only rest mapping. For this to work the implementation has to carefully work around faulting logic. When a page is write-faulted, the fault() hook is called first, then its page_mkwrite() hook is called (to allow for dirty tracking in file systems). On fault we set the folio's mapping in perf_mmap_fault(), this is because when do_page_mkwrite() is subsequently invoked, it treats a missing mapping as an indicator that the fault should be retried. We also set the folio's index so, given the folio is being treated as faux user memory, it correctly references its offset within the VMA. This explains why the mapping and index fields are used - but it's not necessary. We preallocate pages when perf_mmap() is called for the first time via rb_alloc(), and further allocate auxiliary pages via rb_aux_alloc() as needed if the mapping requires it. This allocation is done in the f_ops->mmap() hook provided in perf_mmap(), and so we can instead simply map all the memory right away here - there's no point in handling (read) page faults when we don't demand page nor need to be notified about them (perf does not). This patch therefore changes this logic to map everything when the mmap() hook is called, establishing a PFN map. It implements vm_ops->pfn_mkwrite() to provide the required read/write vs. read-only behaviour, which does not require the previously implemented workarounds. While it is not ideal to use a VM_PFNMAP here, doing anything else will result in the page_mkwrite() hook need to be provided, which requires the same page->mapping hack this patch seeks to undo. It will also result in the pages being treated as folios and placed on the rmap, which really does not make sense for these mappings. Semantically it makes sense to establish this as some kind of special mapping, as the pages are managed by perf and are not strictly user pages, but currently the only means by which we can do so functionally while maintaining the required R/W and R/O behaviour is a PFN map. There should be no change to actual functionality as a result of this change. Signed-off-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/20250103153151.124163-1-lorenzo.stoakes@oracle.com
While at it, rename the same function in s390 cpum_sf PMU. Signed-off-by: Namhyung Kim <namhyung@kernel.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Tested-by: Ravi Bangoria <ravi.bangoria@amd.com> Reviewed-by: Ravi Bangoria <ravi.bangoria@amd.com> Acked-by: Thomas Richter <tmricht@linux.ibm.com> Link: https://lore.kernel.org/r/20241203180441.1634709-2-namhyung@kernel.org