Skip to contentSkip to conversation

perf cs-etm: Track exception level fixups

PATCH In review James Clark 3 participants 11 messages Jun 26, 2023 – Jul 3, 2023

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

cover perf cs-etm: Track exception level fixups

A couple of changes related to edge cases since commit 8d3031d39fe8 ("perf cs-etm: Track exception level").

I think the second one is low risk seeing as any path requiring a thread leading up to adding to the histogram would already have been crashing. Maybe the thread check could also be added to hist_entry_iter__add() although other users of it don't seem to have the same issue, and there is another use of al.thread above in builtin-report.c so it's probably ok where I've added it.

Applies to perf-tools-next/perf-tools-next (929ff679b69)

James Clark (2):
perf cs-etm: Handle per-thread mode on EL1 host kernel case perf report: Don't add to histogram when there is no thread found

 tools/perf/builtin-report.c |  3 +++
 tools/perf/util/cs-etm.c    | 11 +++++++++--
 2 files changed, 12 insertions(+), 2 deletions(-)

base-commit: 929ff679b694f0f9656aec38b3a7d5c440c5ca24

Signature · 1 line
2.34.1
#1 lore raw

PATCH 1/2 perf cs-etm: Handle per-thread mode on EL1 host kernel case

In per-thread mode there are no context packets so no way to determine which type of context packets exist. But because it's only possible to trace host processes in per-thread mode without context packets then assume host in this case.

This fixes the per-thread test case failures when running on nVHE:

98: Check Arm CoreSight trace data recording and synthesized samples:
--- start ---
...
Recording trace with '-e cs_etm/timestamp=0/ --per-thread' Looking at perf.data file for dumping branch samples:
CoreSight basic testing with '-e cs_etm/timestamp=0/ --per-thread': FAIL Recording trace with '-e cs_etm/timestamp=1/ --per-thread' Looking at perf.data file for dumping branch samples:
CoreSight basic testing with '-e cs_etm/timestamp=1/ --per-thread': FAIL
...

Fixes 8d3031d39fe8 ("perf cs-etm: Track exception level")
---
 tools/perf/util/cs-etm.c | 11 +++++++++--
 1 file changed, 9 insertions(+), 2 deletions(-)
tools/perf/util/cs-etm.c +9 -2
index 1419b40dfbe8..85821cc5650e 100644
@@ -900,10 +900,17 @@ static struct machine *cs_etm__get_machine(struct cs_etm_queue *etmq,
 
 	/*
 	 * For any virtualisation based on nVHE (e.g. pKVM), or host kernels
-	 * running at EL1 assume everything is the host.
+	 * running at EL1, or no context IDs (per-thread mode) assume everything
+	 * is the host.
 	 */
-	if (pid_fmt == CS_ETM_PIDFMT_CTXTID)
+	switch (pid_fmt) {
+	case CS_ETM_PIDFMT_CTXTID:
+	case CS_ETM_PIDFMT_NONE:
 		return &etmq->etm->session->machines.host;
+	case CS_ETM_PIDFMT_CTXTID2:
+	default:
+		break;
+	}
 
 	/*
 	 * Not perfect, but otherwise assume anything in EL1 is the default
Signature · 1 line
2.34.1
#2 lore raw

PATCH 2/2 perf report: Don't add to histogram when there is no thread found

In reply to James Clark

thread__find_map() chooses to exit without assigning a thread to the addr_location in some scenarios, for example when there are samples from a guest and perf_guest == false. This results in a segfault when adding to the histogram because it uses unguarded accesses to the thread member of the addr_location.

Fix it by exiting early if no thread is set. This fixes the referenced commit when using perf report with Coresight but probably isn't exclusive to that case.

Fixes 8d3031d39fe8 ("perf cs-etm: Track exception level")
---
 tools/perf/builtin-report.c | 3 +++
 1 file changed, 3 insertions(+)
tools/perf/builtin-report.c +3 -0
index dcedfe00f04d..1a2caa4ce5c3 100644
@@ -293,6 +293,9 @@ static int process_sample_event(struct perf_tool *tool,
 		goto out_put;
 	}
 
+	if (!al.thread)
+		goto out_put;
+
 	if (rep->stitch_lbr)
 		thread__set_lbr_stitch_enable(al.thread, true);
 
Signature · 1 line
2.34.1
maintainer
#3 lore raw

Re: [PATCH 2/2] perf report: Don't add to histogram when there is no thread found

Quoted 5 lines · James Clark

thread__find_map() chooses to exit without assigning a thread to the addr_location in some scenarios, for example when there are samples from a guest and perf_guest == false. This results in a segfault when adding to the histogram because it uses unguarded accesses to the thread member of the addr_location.

Looking at the commit 0dd5041c9a0ea ("perf addr_location: Add init/exit/copy functions") that introduced the change, I'm not sure if it's the intend behavior.

It might change maps and map, but not thread. Then I think no reason to not set the al->thread at the beginning.

How about this? Ian?
(I guess we can get rid of the duplicate 'al->map = NULL' part)

Thanks,
Namhyung

---8<---

tools/perf/util/event.c +2 -5
index 3860b0c74829..4cbb092e0684 100644
@@ -581,15 +581,14 @@ struct map *thread__find_map(struct thread *thread, u8 cpumode, u64 addr,
 	maps__zput(al->maps);
 	map__zput(al->map);
 	thread__zput(al->thread);
+	al->thread = thread__get(thread);
 
 	al->addr = addr;
 	al->cpumode = cpumode;
 	al->filtered = 0;
 
-	if (machine == NULL) {
-		al->map = NULL;
+	if (machine == NULL)
 		return NULL;
-	}
 
 	if (cpumode == PERF_RECORD_MISC_KERNEL && perf_host) {
 		al->level = 'k';
@@ -605,7 +604,6 @@ struct map *thread__find_map(struct thread *thread, u8 cpumode, u64 addr,
 		al->level = 'u';
 	} else {
 		al->level = 'H';
-		al->map = NULL;
 
 		if ((cpumode == PERF_RECORD_MISC_GUEST_USER ||
 			cpumode == PERF_RECORD_MISC_GUEST_KERNEL) &&
@@ -619,7 +617,6 @@ struct map *thread__find_map(struct thread *thread, u8 cpumode, u64 addr,
 		return NULL;
 	}
 	al->maps = maps__get(maps);
-	al->thread = thread__get(thread);
 	al->map = map__get(maps__find(maps, al->addr));
 	if (al->map != NULL) {
 		/*
reviewer
#4 lore raw

Re: [PATCH 2/2] perf report: Don't add to histogram when there is no thread found

Quoted 12 lines · Namhyung Kim
Quoted 5 lines · James Clark

thread__find_map() chooses to exit without assigning a thread to the addr_location in some scenarios, for example when there are samples from a guest and perf_guest == false. This results in a segfault when adding to the histogram because it uses unguarded accesses to the thread member of the addr_location.

Looking at the commit 0dd5041c9a0ea ("perf addr_location: Add init/exit/copy functions") that introduced the change, I'm not sure if it's the intend behavior.

Quoted 5 lines

It might change maps and map, but not thread. Then I think no reason to not set the al->thread at the beginning.

How about this? Ian?
(I guess we can get rid of the duplicate 'al->map = NULL' part)

It seemed strange that we were failing to find a map (the function's purpose) but then populating the address_location. The change below brings back that somewhat odd behavior. I'm okay with reverting to the old behavior, clearly there were users relying on it. We should probably also copy maps and not just thread, as that was the previous behavior.

Thanks,
Ian

Quoted 44 lines

Thanks,
Namhyung

---8<---

tools/perf/util/event.c +2 -5
index 3860b0c74829..4cbb092e0684 100644
@@ -581,15 +581,14 @@ struct map *thread__find_map(struct thread *thread, u8 cpumode, u64 addr,
        maps__zput(al->maps);
        map__zput(al->map);
        thread__zput(al->thread);
+       al->thread = thread__get(thread);
 
        al->addr = addr;
        al->cpumode = cpumode;
        al->filtered = 0;
 
-       if (machine == NULL) {
-               al->map = NULL;
+       if (machine == NULL)
                return NULL;
-       }
 
        if (cpumode == PERF_RECORD_MISC_KERNEL && perf_host) {
                al->level = 'k';
@@ -605,7 +604,6 @@ struct map *thread__find_map(struct thread *thread, u8 cpumode, u64 addr,
                al->level = 'u';
        } else {
                al->level = 'H';
-               al->map = NULL;
 
                if ((cpumode == PERF_RECORD_MISC_GUEST_USER ||
                        cpumode == PERF_RECORD_MISC_GUEST_KERNEL) &&
@@ -619,7 +617,6 @@ struct map *thread__find_map(struct thread *thread, u8 cpumode, u64 addr,
                return NULL;
        }
        al->maps = maps__get(maps);
-       al->thread = thread__get(thread);
        al->map = map__get(maps__find(maps, al->addr));
        if (al->map != NULL) {
                /*
maintainer
#5 lore raw

Re: [PATCH 2/2] perf report: Don't add to histogram when there is no thread found

Quoted 20 lines · Ian Rogers
Quoted 12 lines · Namhyung Kim
Quoted 5 lines · James Clark

thread__find_map() chooses to exit without assigning a thread to the addr_location in some scenarios, for example when there are samples from a guest and perf_guest == false. This results in a segfault when adding to the histogram because it uses unguarded accesses to the thread member of the addr_location.

Looking at the commit 0dd5041c9a0ea ("perf addr_location: Add init/exit/copy functions") that introduced the change, I'm not sure if it's the intend behavior.

Quoted 5 lines

It might change maps and map, but not thread. Then I think no reason to not set the al->thread at the beginning.

How about this? Ian?
(I guess we can get rid of the duplicate 'al->map = NULL' part)

Quoted 6 lines

It seemed strange that we were failing to find a map (the function's purpose) but then populating the address_location. The change below brings back that somewhat odd behavior. I'm okay with reverting to the old behavior, clearly there were users relying on it. We should probably also copy maps and not just thread, as that was the previous behavior.

Probably. But it used to support samples without maps and I think that's why it ignores the return value of thread__find_map(). So we can expect al.map is NULL and maybe fine to leave it for now.

As machine__resolve() returns -1 if it gets no thread, we should set al.thread when it returns 0.

Can I get your Acked-by?

Thanks,
Namhyung

reviewer Acked-by
#6 lore raw

Re: [PATCH 2/2] perf report: Don't add to histogram when there is no thread found

Quoted 29 lines · Namhyung Kim
Quoted 20 lines · Ian Rogers
Quoted 12 lines · Namhyung Kim
Quoted 5 lines · James Clark

thread__find_map() chooses to exit without assigning a thread to the addr_location in some scenarios, for example when there are samples from a guest and perf_guest == false. This results in a segfault when adding to the histogram because it uses unguarded accesses to the thread member of the addr_location.

Looking at the commit 0dd5041c9a0ea ("perf addr_location: Add init/exit/copy functions") that introduced the change, I'm not sure if it's the intend behavior.

Quoted 5 lines

It might change maps and map, but not thread. Then I think no reason to not set the al->thread at the beginning.

How about this? Ian?
(I guess we can get rid of the duplicate 'al->map = NULL' part)

Quoted 6 lines

It seemed strange that we were failing to find a map (the function's purpose) but then populating the address_location. The change below brings back that somewhat odd behavior. I'm okay with reverting to the old behavior, clearly there were users relying on it. We should probably also copy maps and not just thread, as that was the previous behavior.

Quoted 8 lines

Probably. But it used to support samples without maps and I think that's why it ignores the return value of thread__find_map(). So we can expect al.map is NULL and maybe fine to leave it for now.

As machine__resolve() returns -1 if it gets no thread, we should set al.thread when it returns 0.

Can I get your Acked-by?

Yep:

Thanks,
Ian

Quoted 2 lines

Thanks,
Namhyung

#7 lore raw

Re: [PATCH 2/2] perf report: Don't add to histogram when there is no thread found

Quoted 34 lines · Ian Rogers
Quoted 33 lines · Namhyung Kim
Quoted 20 lines · Ian Rogers
Quoted 12 lines · Namhyung Kim
Quoted 5 lines · James Clark

thread__find_map() chooses to exit without assigning a thread to the addr_location in some scenarios, for example when there are samples from a guest and perf_guest == false. This results in a segfault when adding to the histogram because it uses unguarded accesses to the thread member of the addr_location.

Looking at the commit 0dd5041c9a0ea ("perf addr_location: Add init/exit/copy functions") that introduced the change, I'm not sure if it's the intend behavior.

Quoted 5 lines

It might change maps and map, but not thread. Then I think no reason to not set the al->thread at the beginning.

How about this? Ian?
(I guess we can get rid of the duplicate 'al->map = NULL' part)

Quoted 6 lines

It seemed strange that we were failing to find a map (the function's purpose) but then populating the address_location. The change below brings back that somewhat odd behavior. I'm okay with reverting to the old behavior, clearly there were users relying on it. We should probably also copy maps and not just thread, as that was the previous behavior.

Probably. But it used to support samples without maps and I think that's why it ignores the return value of thread__find_map(). So we can expect al.map is NULL and maybe fine to leave it for now.

Quoted 7 lines
Quoted 4 lines

As machine__resolve() returns -1 if it gets no thread, we should set al.thread when it returns 0.

Can I get your Acked-by?

Yep:

Looks good to me too. Should I resend the set with this change instead of my one?

Quoted 6 lines

Thanks,
Ian

Quoted 2 lines

Thanks,
Namhyung

_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

maintainer
#8 lore raw

Re: [PATCH 2/2] perf report: Don't add to histogram when there is no thread found

Quoted 41 lines · James Clark
Quoted 37 lines · Ian Rogers
Quoted 36 lines · Namhyung Kim
Quoted 20 lines · Ian Rogers
Quoted 12 lines · Namhyung Kim
Quoted 5 lines · James Clark

thread__find_map() chooses to exit without assigning a thread to the addr_location in some scenarios, for example when there are samples from a guest and perf_guest == false. This results in a segfault when adding to the histogram because it uses unguarded accesses to the thread member of the addr_location.

Looking at the commit 0dd5041c9a0ea ("perf addr_location: Add init/exit/copy functions") that introduced the change, I'm not sure if it's the intend behavior.

Quoted 5 lines

It might change maps and map, but not thread. Then I think no reason to not set the al->thread at the beginning.

How about this? Ian?
(I guess we can get rid of the duplicate 'al->map = NULL' part)

Quoted 6 lines

It seemed strange that we were failing to find a map (the function's purpose) but then populating the address_location. The change below brings back that somewhat odd behavior. I'm okay with reverting to the old behavior, clearly there were users relying on it. We should probably also copy maps and not just thread, as that was the previous behavior.

Probably. But it used to support samples without maps and I think that's why it ignores the return value of thread__find_map(). So we can expect al.map is NULL and maybe fine to leave it for now.

As machine__resolve() returns -1 if it gets no thread, we should set al.thread when it returns 0.

Quoted 7 lines
Quoted 4 lines
Quoted 1 line

Can I get your Acked-by?

Yep:

Looks good to me too. Should I resend the set with this change instead of my one?

No, I can take care of that. I'll take this as your Acked-by. :)

Thanks,
Namhyung

maintainer
#9 lore raw

Re: [PATCH 2/2] perf report: Don't add to histogram when there is no thread found

Quoted 45 lines · Namhyung Kim
Quoted 43 lines · James Clark
Quoted 39 lines · Ian Rogers
Quoted 37 lines · Namhyung Kim
Quoted 20 lines · Ian Rogers
Quoted 12 lines · Namhyung Kim
Quoted 5 lines · James Clark

thread__find_map() chooses to exit without assigning a thread to the addr_location in some scenarios, for example when there are samples from a guest and perf_guest == false. This results in a segfault when adding to the histogram because it uses unguarded accesses to the thread member of the addr_location.

Looking at the commit 0dd5041c9a0ea ("perf addr_location: Add init/exit/copy functions") that introduced the change, I'm not sure if it's the intend behavior.

Quoted 5 lines

It might change maps and map, but not thread. Then I think no reason to not set the al->thread at the beginning.

How about this? Ian?
(I guess we can get rid of the duplicate 'al->map = NULL' part)

Quoted 6 lines

It seemed strange that we were failing to find a map (the function's purpose) but then populating the address_location. The change below brings back that somewhat odd behavior. I'm okay with reverting to the old behavior, clearly there were users relying on it. We should probably also copy maps and not just thread, as that was the previous behavior.

Probably. But it used to support samples without maps and I think that's why it ignores the return value of thread__find_map(). So we can expect al.map is NULL and maybe fine to leave it for now.

As machine__resolve() returns -1 if it gets no thread, we should set al.thread when it returns 0.

Can I get your Acked-by?

Quoted 7 lines
Quoted 5 lines
Quoted 2 lines

Yep:

Looks good to me too. Should I resend the set with this change instead of my one?

No, I can take care of that. I'll take this as your Acked-by. :)

This part is applied to perf-tools-next, thanks!

#10 lore raw

Re: [PATCH 2/2] perf report: Don't add to histogram when there is no thread found

Quoted 49 lines · Namhyung Kim
Quoted 48 lines · Namhyung Kim
Quoted 46 lines · James Clark
Quoted 41 lines · Ian Rogers
Quoted 29 lines · Namhyung Kim
Quoted 20 lines · Ian Rogers
Quoted 12 lines · Namhyung Kim
Quoted 5 lines · James Clark

thread__find_map() chooses to exit without assigning a thread to the addr_location in some scenarios, for example when there are samples from a guest and perf_guest == false. This results in a segfault when adding to the histogram because it uses unguarded accesses to the thread member of the addr_location.

Looking at the commit 0dd5041c9a0ea ("perf addr_location: Add init/exit/copy functions") that introduced the change, I'm not sure if it's the intend behavior.

Quoted 5 lines

It might change maps and map, but not thread. Then I think no reason to not set the al->thread at the beginning.

How about this? Ian?
(I guess we can get rid of the duplicate 'al->map = NULL' part)

Quoted 6 lines

It seemed strange that we were failing to find a map (the function's purpose) but then populating the address_location. The change below brings back that somewhat odd behavior. I'm okay with reverting to the old behavior, clearly there were users relying on it. We should probably also copy maps and not just thread, as that was the previous behavior.

Quoted 8 lines

Probably. But it used to support samples without maps and I think that's why it ignores the return value of thread__find_map(). So we can expect al.map is NULL and maybe fine to leave it for now.

As machine__resolve() returns -1 if it gets no thread, we should set al.thread when it returns 0.

Can I get your Acked-by?

Yep:

Quoted 6 lines
Quoted 4 lines
Quoted 2 lines

Looks good to me too. Should I resend the set with this change instead of my one?

No, I can take care of that. I'll take this as your Acked-by. :)

This part is applied to perf-tools-next, thanks!

Thanks Namhyung