data->sample_flags may be modified in perf_prepare_sample(),
in perf_tp_event(), different swevents use the same on-stack
perf_sample_data, the previous swevent may change sample_flags in
perf_prepare_sample(), as a result, some members of perf_sample_data are
not correctly initialized when next swevent_event preparing sample
(for example data->id, the value varies according to swevent).
A simple scenario triggers this problem is as follows:
# perf record -e sched:sched_switch --switch-output-event sched:sched_switch -a sleep 1
[ perf record: dump data: Woken up 0 times ]
[ perf record: Dump perf.data.2023041209014396 ]
[ perf record: dump data: Woken up 0 times ]
[ perf record: Dump perf.data.2023041209014662 ]
[ perf record: dump data: Woken up 0 times ]
[ perf record: Dump perf.data.2023041209014910 ]
[ perf record: Woken up 0 times to write data ]
[ perf record: Dump perf.data.2023041209015164 ]
[ perf record: Captured and wrote 0.069 MB perf.data.<timestamp> ]
# ls -l
total 860
-rw------- 1 root root 95694 Apr 12 09:01 perf.data.2023041209014396
-rw------- 1 root root 606430 Apr 12 09:01 perf.data.2023041209014662
-rw------- 1 root root 82246 Apr 12 09:01 perf.data.2023041209014910
-rw------- 1 root root 82342 Apr 12 09:01 perf.data.2023041209015164
# perf script -i perf.data.2023041209014396
0x11d58 [0x80]: failed to process type: 9 [Bad address]
Solution: Add perf_sample_data_flags_{save, restore} helpers to save and
restore sample_flags when processing different swevents
After fix:
# perf record -e sched:sched_switch --switch-output-event sched:sched_switch -a sleep 1
[ perf record: dump data: Woken up 0 times ]
[ perf record: Dump perf.data.2023041209442259 ]
[ perf record: dump data: Woken up 0 times ]
[ perf record: Dump perf.data.2023041209442514 ]
[ perf record: dump data: Woken up 0 times ]
[ perf record: Dump perf.data.2023041209442760 ]
[ perf record: Woken up 0 times to write data ]
[ perf record: Dump perf.data.2023041209443003 ]
[ perf record: Captured and wrote 0.069 MB perf.data.<timestamp> ]
# ls -l
total 864
-rw------- 1 root root 100166 Apr 12 09:44 perf.data.2023041209442259
-rw------- 1 root root 606438 Apr 12 09:44 perf.data.2023041209442514
-rw------- 1 root root 82246 Apr 12 09:44 perf.data.2023041209442760
-rw------- 1 root root 82342 Apr 12 09:44 perf.data.2023041209443003
# perf script -i perf.data.2023041209442259 | head -n 5
perf 232 [000] 66.846217: sched:sched_switch: prev_comm=perf prev_pid=232 prev_prio=120 prev_state=D ==> next_comm=perf next_pid=234 next_prio=120
perf 234 [000] 66.846449: sched:sched_switch: prev_comm=perf prev_pid=234 prev_prio=120 prev_state=S ==> next_comm=perf next_pid=232 next_prio=120
perf 232 [000] 66.846546: sched:sched_switch: prev_comm=perf prev_pid=232 prev_prio=120 prev_state=R ==> next_comm=perf next_pid=234 next_prio=120
perf 234 [000] 66.846606: sched:sched_switch: prev_comm=perf prev_pid=234 prev_prio=120 prev_state=S ==> next_comm=perf next_pid=232 next_prio=120
perf 232 [000] 66.846646: sched:sched_switch: prev_comm=perf prev_pid=232 prev_prio=120 prev_state=R ==> next_comm=perf next_pid=234 next_prio=120
Fixes: bb447c27a467 ("perf/core: Set data->sample_flags in perf_prepare_sample()")
Signed-off-by: Yang Jihong <[email protected]>
---
include/linux/perf_event.h | 11 +++++++++++
kernel/events/core.c | 13 ++++++++++++-
2 files changed, 23 insertions(+), 1 deletion(-)
diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
index d5628a7b5eaa..7499bd3a2ed5 100644
--- a/include/linux/perf_event.h
+++ b/include/linux/perf_event.h
@@ -1185,6 +1185,17 @@ struct perf_sample_data {
PERF_MEM_S(LOCK, NA) |\
PERF_MEM_S(TLB, NA))
+/* Note: caller must ensure flags is not NULL */
+static inline void perf_sample_data_flags_save(struct perf_sample_data *data, u64 *flags)
+{
+ *flags = data->sample_flags;
+}
+
+static inline void perf_sample_data_flags_restore(struct perf_sample_data *data, u64 flags)
+{
+ data->sample_flags = flags;
+}
+
static inline void perf_sample_data_init(struct perf_sample_data *data,
u64 addr, u64 period)
{
diff --git a/kernel/events/core.c b/kernel/events/core.c
index 435815d3be3f..ea1fb63a6037 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -10136,6 +10136,7 @@ void perf_tp_event(u16 event_type, u64 count, void *record, int entry_size,
{
struct perf_sample_data data;
struct perf_event *event;
+ u64 sample_flags;
struct perf_raw_record raw = {
.frag = {
@@ -10150,8 +10151,18 @@ void perf_tp_event(u16 event_type, u64 count, void *record, int entry_size,
perf_trace_buf_update(record, event_type);
hlist_for_each_entry_rcu(event, head, hlist_entry) {
- if (perf_tp_event_match(event, &data, regs))
+ if (perf_tp_event_match(event, &data, regs)) {
+ /*
+ * sample_flags may be modified in perf_event_output_*
+ * (see perf_prepare_sample). For different swevents,
+ * use the same on-stack perf_sample_data, therefore,
+ * need to save samle_flags, and restore it
+ * after perf_swevent_event.
+ */
+ perf_sample_data_flags_save(&data, &sample_flags);
perf_swevent_event(event, count, &data, regs);
+ perf_sample_data_flags_restore(&data, sample_flags);
+ }
}
/*
--
2.30.GIT
On Wed, Apr 12, 2023 at 09:52:40AM +0000, Yang Jihong wrote:
> data->sample_flags may be modified in perf_prepare_sample(),
> in perf_tp_event(), different swevents use the same on-stack
> perf_sample_data, the previous swevent may change sample_flags in
> perf_prepare_sample(), as a result, some members of perf_sample_data are
> not correctly initialized when next swevent_event preparing sample
> (for example data->id, the value varies according to swevent).
>
> A simple scenario triggers this problem is as follows:
>
> # perf record -e sched:sched_switch --switch-output-event sched:sched_switch -a sleep 1
> [ perf record: dump data: Woken up 0 times ]
> [ perf record: Dump perf.data.2023041209014396 ]
> [ perf record: dump data: Woken up 0 times ]
> [ perf record: Dump perf.data.2023041209014662 ]
> [ perf record: dump data: Woken up 0 times ]
> [ perf record: Dump perf.data.2023041209014910 ]
> [ perf record: Woken up 0 times to write data ]
> [ perf record: Dump perf.data.2023041209015164 ]
> [ perf record: Captured and wrote 0.069 MB perf.data.<timestamp> ]
> # ls -l
> total 860
> -rw------- 1 root root 95694 Apr 12 09:01 perf.data.2023041209014396
> -rw------- 1 root root 606430 Apr 12 09:01 perf.data.2023041209014662
> -rw------- 1 root root 82246 Apr 12 09:01 perf.data.2023041209014910
> -rw------- 1 root root 82342 Apr 12 09:01 perf.data.2023041209015164
> # perf script -i perf.data.2023041209014396
> 0x11d58 [0x80]: failed to process type: 9 [Bad address]
>
> Solution: Add perf_sample_data_flags_{save, restore} helpers to save and
> restore sample_flags when processing different swevents
>
> After fix:
>
> # perf record -e sched:sched_switch --switch-output-event sched:sched_switch -a sleep 1
> [ perf record: dump data: Woken up 0 times ]
> [ perf record: Dump perf.data.2023041209442259 ]
> [ perf record: dump data: Woken up 0 times ]
> [ perf record: Dump perf.data.2023041209442514 ]
> [ perf record: dump data: Woken up 0 times ]
> [ perf record: Dump perf.data.2023041209442760 ]
> [ perf record: Woken up 0 times to write data ]
> [ perf record: Dump perf.data.2023041209443003 ]
> [ perf record: Captured and wrote 0.069 MB perf.data.<timestamp> ]
> # ls -l
> total 864
> -rw------- 1 root root 100166 Apr 12 09:44 perf.data.2023041209442259
> -rw------- 1 root root 606438 Apr 12 09:44 perf.data.2023041209442514
> -rw------- 1 root root 82246 Apr 12 09:44 perf.data.2023041209442760
> -rw------- 1 root root 82342 Apr 12 09:44 perf.data.2023041209443003
> # perf script -i perf.data.2023041209442259 | head -n 5
> perf 232 [000] 66.846217: sched:sched_switch: prev_comm=perf prev_pid=232 prev_prio=120 prev_state=D ==> next_comm=perf next_pid=234 next_prio=120
> perf 234 [000] 66.846449: sched:sched_switch: prev_comm=perf prev_pid=234 prev_prio=120 prev_state=S ==> next_comm=perf next_pid=232 next_prio=120
> perf 232 [000] 66.846546: sched:sched_switch: prev_comm=perf prev_pid=232 prev_prio=120 prev_state=R ==> next_comm=perf next_pid=234 next_prio=120
> perf 234 [000] 66.846606: sched:sched_switch: prev_comm=perf prev_pid=234 prev_prio=120 prev_state=S ==> next_comm=perf next_pid=232 next_prio=120
> perf 232 [000] 66.846646: sched:sched_switch: prev_comm=perf prev_pid=232 prev_prio=120 prev_state=R ==> next_comm=perf next_pid=234 next_prio=120
This seems a little bit short on analysis; what actual flags are the
problem? Much of the data will in fact be identical between these
invocations and endlessly re-computing that is wasteful.
I'm thinking perhaps those flags that update ->dyn_size are the problem?
At the same time, Should you not also then clear dyn_size?
Hello,
On 2023/4/17 19:45, Peter Zijlstra wrote:
> On Wed, Apr 12, 2023 at 09:52:40AM +0000, Yang Jihong wrote:
>> data->sample_flags may be modified in perf_prepare_sample(),
>> in perf_tp_event(), different swevents use the same on-stack
>> perf_sample_data, the previous swevent may change sample_flags in
>> perf_prepare_sample(), as a result, some members of perf_sample_data are
>> not correctly initialized when next swevent_event preparing sample
>> (for example data->id, the value varies according to swevent).
>>
>> A simple scenario triggers this problem is as follows:
>>
>> # perf record -e sched:sched_switch --switch-output-event sched:sched_switch -a sleep 1
>> [ perf record: dump data: Woken up 0 times ]
>> [ perf record: Dump perf.data.2023041209014396 ]
>> [ perf record: dump data: Woken up 0 times ]
>> [ perf record: Dump perf.data.2023041209014662 ]
>> [ perf record: dump data: Woken up 0 times ]
>> [ perf record: Dump perf.data.2023041209014910 ]
>> [ perf record: Woken up 0 times to write data ]
>> [ perf record: Dump perf.data.2023041209015164 ]
>> [ perf record: Captured and wrote 0.069 MB perf.data.<timestamp> ]
>> # ls -l
>> total 860
>> -rw------- 1 root root 95694 Apr 12 09:01 perf.data.2023041209014396
>> -rw------- 1 root root 606430 Apr 12 09:01 perf.data.2023041209014662
>> -rw------- 1 root root 82246 Apr 12 09:01 perf.data.2023041209014910
>> -rw------- 1 root root 82342 Apr 12 09:01 perf.data.2023041209015164
>> # perf script -i perf.data.2023041209014396
>> 0x11d58 [0x80]: failed to process type: 9 [Bad address]
>>
>> Solution: Add perf_sample_data_flags_{save, restore} helpers to save and
>> restore sample_flags when processing different swevents
>>
>> After fix:
>>
>> # perf record -e sched:sched_switch --switch-output-event sched:sched_switch -a sleep 1
>> [ perf record: dump data: Woken up 0 times ]
>> [ perf record: Dump perf.data.2023041209442259 ]
>> [ perf record: dump data: Woken up 0 times ]
>> [ perf record: Dump perf.data.2023041209442514 ]
>> [ perf record: dump data: Woken up 0 times ]
>> [ perf record: Dump perf.data.2023041209442760 ]
>> [ perf record: Woken up 0 times to write data ]
>> [ perf record: Dump perf.data.2023041209443003 ]
>> [ perf record: Captured and wrote 0.069 MB perf.data.<timestamp> ]
>> # ls -l
>> total 864
>> -rw------- 1 root root 100166 Apr 12 09:44 perf.data.2023041209442259
>> -rw------- 1 root root 606438 Apr 12 09:44 perf.data.2023041209442514
>> -rw------- 1 root root 82246 Apr 12 09:44 perf.data.2023041209442760
>> -rw------- 1 root root 82342 Apr 12 09:44 perf.data.2023041209443003
>> # perf script -i perf.data.2023041209442259 | head -n 5
>> perf 232 [000] 66.846217: sched:sched_switch: prev_comm=perf prev_pid=232 prev_prio=120 prev_state=D ==> next_comm=perf next_pid=234 next_prio=120
>> perf 234 [000] 66.846449: sched:sched_switch: prev_comm=perf prev_pid=234 prev_prio=120 prev_state=S ==> next_comm=perf next_pid=232 next_prio=120
>> perf 232 [000] 66.846546: sched:sched_switch: prev_comm=perf prev_pid=232 prev_prio=120 prev_state=R ==> next_comm=perf next_pid=234 next_prio=120
>> perf 234 [000] 66.846606: sched:sched_switch: prev_comm=perf prev_pid=234 prev_prio=120 prev_state=S ==> next_comm=perf next_pid=232 next_prio=120
>> perf 232 [000] 66.846646: sched:sched_switch: prev_comm=perf prev_pid=232 prev_prio=120 prev_state=R ==> next_comm=perf next_pid=234 next_prio=120
>
> This seems a little bit short on analysis; what actual flags are the
> problem? Much of the data will in fact be identical between these
> invocations and endlessly re-computing that is wasteful.
>
In the preceding example, data->id is incorrectly initialized,
that is, PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER
The detailed process is as follows:
1. In the example of perf record, create two swevents evt1 and evt2,
and attach to sched:sched_switch, assume that the IDs are 10 and 11
2. In perf_tp_event():
struct perf_sample_data data // on-stack value
hlist_for_each_entry_rcu
-> perf_swevent_event(evt1, &data)
-> perf_swevent_overflow(evt1, &data)
-> ...
-> perf_prepare_sample(evt1, &data)
-> __perf_event_header__init_id(evt1, &data)
-> data->sample_flags |= data->type & PERF_SAMPLE_ID_ALL
// data->sample_flags set PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER
-> if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER))
data->id = primary_event_id(event);
// for evt1, the if condition is met,
// and data->id can be correctly initialized.
...
-> perf_swevent_event(evt2, &data)
-> perf_swevent_overflow(evt2, &data)
-> ...
-> perf_prepare_sample(evt2, &data)
-> __perf_event_header__init_id(evt2, &data)
-> if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER))
data->id = primary_event_id(event);
// data->sample_flags set PERF_SAMPLE_ID |
PERF_SAMPLE_IDENTIFIER in evt1,
// therefore, the condition is not met for evt2.
// As a result, the ID of evt2 is not assigned to data-id.
> I'm thinking perhaps those flags that update ->dyn_size are the problem?
> At the same time, Should you not also then clear dyn_size?
Yes, according to the code, dyn_size should also be cleared.
Maybe we need to change it to the following, which would be more
appropriate?
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -10144,14 +10144,14 @@ void perf_tp_event(u16 event_type, u64 count,
void *record, int entry_size,
},
};
- perf_sample_data_init(&data, 0, 0);
- perf_sample_save_raw_data(&data, &raw);
-
perf_trace_buf_update(record, event_type);
hlist_for_each_entry_rcu(event, head, hlist_entry) {
- if (perf_tp_event_match(event, &data, regs))
+ if (perf_tp_event_match(event, &data, regs)) {
+ perf_sample_data_init(&data, 0, 0);
+ perf_sample_save_raw_data(&data, &raw);
perf_swevent_event(event, count, &data, regs);
+ }
}
Thanks,
Yang.
>
> .
>
On Tue, Apr 18, 2023 at 09:35:23AM +0800, Yang Jihong wrote:
> > I'm thinking perhaps those flags that update ->dyn_size are the problem?
> > At the same time, Should you not also then clear dyn_size?
> Yes, according to the code, dyn_size should also be cleared.
> Maybe we need to change it to the following, which would be more
> appropriate?
>
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -10144,14 +10144,14 @@ void perf_tp_event(u16 event_type, u64 count, void
> *record, int entry_size,
> },
> };
>
> - perf_sample_data_init(&data, 0, 0);
> - perf_sample_save_raw_data(&data, &raw);
> -
> perf_trace_buf_update(record, event_type);
>
> hlist_for_each_entry_rcu(event, head, hlist_entry) {
> - if (perf_tp_event_match(event, &data, regs))
> + if (perf_tp_event_match(event, &data, regs)) {
> + perf_sample_data_init(&data, 0, 0);
> + perf_sample_save_raw_data(&data, &raw);
> perf_swevent_event(event, count, &data, regs);
> + }
> }
That is certainly the safe option. I just went through the list and
while there's certainly a number of options we'll recompute for naught,
most of them are indeed either dyn_size or event specific :/
So yeah, please send the above as v2.
Hello,
On 2023/4/18 18:25, Peter Zijlstra wrote:
> On Tue, Apr 18, 2023 at 09:35:23AM +0800, Yang Jihong wrote:
>
>>> I'm thinking perhaps those flags that update ->dyn_size are the problem?
>>> At the same time, Should you not also then clear dyn_size?
>
>> Yes, according to the code, dyn_size should also be cleared.
>> Maybe we need to change it to the following, which would be more
>> appropriate?
>>
>> --- a/kernel/events/core.c
>> +++ b/kernel/events/core.c
>> @@ -10144,14 +10144,14 @@ void perf_tp_event(u16 event_type, u64 count, void
>> *record, int entry_size,
>> },
>> };
>>
>> - perf_sample_data_init(&data, 0, 0);
>> - perf_sample_save_raw_data(&data, &raw);
>> -
>> perf_trace_buf_update(record, event_type);
>>
>> hlist_for_each_entry_rcu(event, head, hlist_entry) {
>> - if (perf_tp_event_match(event, &data, regs))
>> + if (perf_tp_event_match(event, &data, regs)) {
>> + perf_sample_data_init(&data, 0, 0);
>> + perf_sample_save_raw_data(&data, &raw);
>> perf_swevent_event(event, count, &data, regs);
>> + }
>> }
>
> That is certainly the safe option. I just went through the list and
> while there's certainly a number of options we'll recompute for naught,
> most of them are indeed either dyn_size or event specific :/
>
> So yeah, please send the above as v2.
>
OK, will send v2 according to above fix solution.
Thanks,
Yang.