perf: fix several memory leaks reported by ASan on perf-test
cover perf: fix several memory leaks reported by ASan on perf-test
This patchset fixes several memory leaks found using ASan on perf-test. After this patch, all memory leaks in perf-test that were found on my machine are fixed.
This does not mean that there are no longer memory leaks in perf (just running perf report or perf top raises a few), but that now it is possible to run the perf tests with ASan to catch future memory errors.
While many of these leaks have little to no effect on the daily use of perf (some were inside the test itself, other were just a missing cleanup on exit), other were (slightly) more serious, namely:
- nsinfo refcounting was broken since it was not decreased when a reference was replaced, causing many nsinfos to never being deallocated.
- dso__new_map refcounting was wrong, missing a __put
- lzma stream was never closed
Below are the results of perf-test before and after this patchset. Only failed and skipped tests are reported for brevity.
The perf binary has been compiled using
make DEBUG=1 EXTRA_CFLAGS='-fno-omit-frame-pointer -fsanitize=address' compilation using clang shows same results.
Tests are run on my laptop:
Intel(R) Core(TM) i7-5500U CPU @ 2.40GHz (2 cores + hyperthreading) Fedora 34 w/ kernel 5.12.14-300.fc34.x86_64
Before this patchset:
1: vmlinux symtab matches kallsyms : FAILED! 19: 'import perf' in python : FAILED! 23.1: Read Only Watchpoint : Skip (missing hardware support) 31: Lookup mmap thread : FAILED! 41: Session topology : FAILED! 42.1: Basic BPF filtering : FAILED! 42.2: BPF pinning : Skip 42.3: BPF prologue generation : Skip 49: Synthesize attr update : FAILED! 58: builtin clang support : Skip (not compiled in) 63: Test libpfm4 support : Skip (not compiled in) 65: maps__merge_in : FAILED! 73: DWARF unwind : FAILED! 78: build id cache operations : FAILED! 82: Use vfs_getname probe to get syscall args filenames : FAILED! 83: Zstd perf.data compression/decompression : FAILED! 86: perf stat --bpf-counters test : Skip
87: Check Arm CoreSight trace data recording and synthesized samples: Skip
88: Check open filename arg using perf trace + vfs_getname : FAILED!
After this patchset:
19: 'import perf' in python : FAILED! 23.1: Read Only Watchpoint : Skip (missing hardware support) 58: builtin clang support : Skip (not compiled in) 63: Test libpfm4 support : Skip (not compiled in) 73: DWARF unwind : FAILED! 86: perf stat --bpf-counters test : Skip
87: Check Arm CoreSight trace data recording and synthesized samples: Skip
After this patchset, without ASan:
23.1: Read Only Watchpoint : Skip (missing hardware support) 58: builtin clang support : Skip (not compiled in) 63: Test libpfm4 support : Skip (not compiled in) 86: perf stat --bpf-counters test : Skip
87: Check Arm CoreSight trace data recording and synthesized samples: Skip
The two failing tests are false positives and can be skipped:
- "19: 'import perf' in python" fails since there are memory leaks inside the python library.
- "73: DWARF unwind" fails due to stack buffer underflow, since it's unwinding the stack.
Riccardo
Riccardo Mancini (20):
perf nsinfo: fix refcounting
perf env: fix sibling_dies memory leak
perf test: session_topology: delete session->evlist perf test: event_update: fix memory leak of evlist perf test: event_update: fix memory leak of unit perf dso: fix memory leak in dso__new_map
perf test: maps__merge_in: fix memory leak of maps perf env: fix memory leak of cpu_pmu_caps
perf report: free generated help strings for sort option perf inject: close inject.output
perf session: cleanup trace_event
perf script: release zstd data
perf script: fix memory leaks in perf_script
perf util/lzma: close lzma stream
perf trace: free malloc'd trace fields on exit perf trace: free syscall->arg_fmt
perf trace: free syscall tp fields in evsel->priv perf trace: free strings in trace__parse_events_option perf test: bpf: free obj_buf
perf util/probe-file: delete namelist on error in del_events
tools/perf/builtin-inject.c | 13 ++++++--- tools/perf/builtin-report.c | 33 ++++++++++++++------- tools/perf/builtin-script.c | 8 ++++++ tools/perf/builtin-trace.c | 51 +++++++++++++++++++++++++++++++-- tools/perf/tests/bpf.c | 1 + tools/perf/tests/event_update.c | 6 ++-- tools/perf/tests/maps.c | 2 ++ tools/perf/tests/topology.c | 1 + tools/perf/util/dso.c | 4 ++- tools/perf/util/env.c | 2 ++ tools/perf/util/lzma.c | 8 ++++-- tools/perf/util/map.c | 2 ++ tools/perf/util/probe-event.c | 4 ++- tools/perf/util/probe-file.c | 4 +-- tools/perf/util/session.c | 1 + tools/perf/util/sort.c | 2 +- tools/perf/util/sort.h | 2 +- 17 files changed, 116 insertions(+), 28 deletions(-)
Signature · 1 line
2.31.1
PATCH 1/20 perf nsinfo: fix refcounting
ASan reports a memory leak of nsinfo during the execution of the perf test "31: Lookup mmap thread".
The leak is caused by a refcounted variable being replaced without dropping the refcount.
This patch makes sure that the refcnt of nsinfo is decreased whenever a refcounted variable is replaced with a new value.
--- tools/perf/builtin-inject.c | 5 +++-- tools/perf/util/map.c | 2 ++ tools/perf/util/probe-event.c | 4 +++- 3 files changed, 8 insertions(+), 3 deletions(-)
tools/perf/builtin-inject.c +3 -2
@@ -361,9 +361,10 @@ static struct dso *findnew_dso(int pid, int tid, const char *filename, dso =machine__findnew_dso_id (machine, filename, id); } -if (dso) +if (dso) { +nsinfo__put (dso->nsinfo); dso->nsinfo = nsi; -else + }else nsinfo__put (nsi);thread__put (thread);
tools/perf/util/map.c +2 -0
@@ -192,6 +192,8 @@ struct map *map__new(struct machine *machine, u64 start, u64 len,if (!(prot &PROT_EXEC ))dso__set_loaded (dso); } + +nsinfo__put (dso->nsinfo); dso->nsinfo = nsi;if (build_id__is_defined (bid))
tools/perf/util/probe-event.c +3 -1
@@ -179,8 +179,10 @@ struct map *get_target_map(const char *target, struct nsinfo *nsi, bool user)struct map *map; map =dso__new_map (target); -if (map && map->dso) +if (map && map->dso) { +nsinfo__put (map->dso->nsinfo); map->dso->nsinfo =nsinfo__get (nsi); + }return map; }else {return kernel_get_module_map (target);
Signature · 1 line
2.31.1
Arnaldo Carvalho de Melo
maintainerRe: [PATCH 01/20] perf nsinfo: fix refcounting
Em Thu, Jul 15, 2021 at 06:07:06PM +0200, Riccardo Mancini escreveu:
Quoted 7 lines
ASan reports a memory leak of nsinfo during the execution of the perf test "31: Lookup mmap thread".
The leak is caused by a refcounted variable being replaced without dropping the refcount.This patch makes sure that the refcnt of nsinfo is decreased whenever a refcounted variable is replaced with a new value.
So, there are multiple fixes in just one patch, I'll split it into three, no need to resend.
I'll try and check if finding Fixes: for the three is easy, that way [email protected] will figure out which of the supported releases need each of them.
- Arnaldo
Quoted 56 lines
Signed-off-by Riccardo Mancini <[email protected]>--- tools/perf/builtin-inject.c | 5 +++-- tools/perf/util/map.c | 2 ++ tools/perf/util/probe-event.c | 4 +++- 3 files changed, 8 insertions(+), 3 deletions(-)tools/perf/builtin-inject.c +3
-2@@ -361,9 +361,10 @@ static struct dso *findnew_dso(int pid, int tid, const char *filename, dso =machine__findnew_dso_id (machine, filename, id); } -if (dso) +if (dso) { +nsinfo__put (dso->nsinfo); dso->nsinfo = nsi; -else + }else nsinfo__put (nsi);thread__put (thread);tools/perf/util/map.c +2
-0@@ -192,6 +192,8 @@ struct map *map__new(struct machine *machine, u64 start, u64 len,if (!(prot &PROT_EXEC ))dso__set_loaded (dso); } + +nsinfo__put (dso->nsinfo); dso->nsinfo = nsi;if (build_id__is_defined (bid))tools/perf/util/probe-event.c +3
-1@@ -179,8 +179,10 @@ struct map *get_target_map(const char *target, struct nsinfo *nsi, bool user)struct map *map; map =dso__new_map (target); -if (map && map->dso) +if (map && map->dso) { +nsinfo__put (map->dso->nsinfo); map->dso->nsinfo =nsinfo__get (nsi); + }return map; }else {return kernel_get_module_map (target);Signature · 1 line
2.31.1
Signature · 1 line
- Arnaldo
Arnaldo Carvalho de Melo
maintainerRe: [PATCH 01/20] perf nsinfo: fix refcounting
Em Thu, Jul 15, 2021 at 04:02:16PM -0300, Arnaldo Carvalho de Melo escreveu:
Quoted 9 lines
Em Thu, Jul 15, 2021 at 06:07:06PM +0200, Riccardo Mancini escreveu:
Quoted 7 lines
ASan reports a memory leak of nsinfo during the execution of the perf test "31: Lookup mmap thread".
The leak is caused by a refcounted variable being replaced without dropping the refcount.This patch makes sure that the refcnt of nsinfo is decreased whenever a refcounted variable is replaced with a new value.
Quoted 6 lines
So, there are multiple fixes in just one patch, I'll split it into three, no need to resend.
I'll try and check if finding Fixes: for the three is easy, that way [email protected] will figure out which of the supported releases need each of them.
The first patch becomes:
commit b36bbdefec297ee5b6e278904360ad14fa3b09de
Author: Riccardo Mancini <[email protected]> Date: Thu Jul 15 18:07:06 2021 +0200
perf inject: Fix dso->nsinfo refcounting
ASan reports a memory leak of nsinfo during the execution of the perf test "31: Lookup mmap thread".
The leak is caused by a refcounted variable being replaced without dropping the refcount.
This patch makes sure that the refcnt of nsinfo is decreased when a refcounted variable is replaced with a new value.
tools/perf/builtin-inject.c +3 -2
@@ -361,9 +361,10 @@ static struct dso *findnew_dso(int pid, int tid, const char *filename, dso =machine__findnew_dso_id (machine, filename, id); } -if (dso) +if (dso) { +nsinfo__put (dso->nsinfo); dso->nsinfo = nsi; -else + }else nsinfo__put (nsi);thread__put (thread);
Arnaldo Carvalho de Melo
maintainerRe: [PATCH 01/20] perf nsinfo: fix refcounting
In reply to Arnaldo Carvalho de MeloEm Thu, Jul 15, 2021 at 04:02:16PM -0300, Arnaldo Carvalho de Melo escreveu:
Quoted 9 lines
Em Thu, Jul 15, 2021 at 06:07:06PM +0200, Riccardo Mancini escreveu:
Quoted 7 lines
ASan reports a memory leak of nsinfo during the execution of the perf test "31: Lookup mmap thread".
The leak is caused by a refcounted variable being replaced without dropping the refcount.This patch makes sure that the refcnt of nsinfo is decreased whenever a refcounted variable is replaced with a new value.
Quoted 6 lines
So, there are multiple fixes in just one patch, I'll split it into three, no need to resend.
I'll try and check if finding Fixes: for the three is easy, that way [email protected] will figure out which of the supported releases need each of them.
The second is at the end of this message, and looking at it I think there is a problem with where this nsinfo reference is stored, it should be in the 'struct map', that is per-thread, not in the 'struct dso', that may be shared by multiple 'struct map's in different cgroups, so we can't set dso->nsinfo in map__new(), right? Kristen?
- Arnaldo
commit 280f3933661f9ca9563ed04ce82313fdbcbb8f75
Author: Riccardo Mancini <[email protected]> Date: Thu Jul 15 18:07:06 2021 +0200
perf map: Fix dso->nsinfo refcounting
ASan reports a memory leak of nsinfo during the execution of
# perf test "31: Lookup mmap thread"
The leak is caused by a refcounted variable being replaced without dropping the refcount.
This patch makes sure that the refcnt of nsinfo is decreased whenever a refcounted variable is replaced with a new value.
tools/perf/util/map.c +2 -0
@@ -192,6 +192,8 @@ struct map *map__new(struct machine *machine, u64 start, u64 len,if (!(prot &PROT_EXEC ))dso__set_loaded (dso); } + +nsinfo__put (dso->nsinfo); dso->nsinfo = nsi;if (build_id__is_defined (bid))
Re: [PATCH 01/20] perf nsinfo: fix refcounting
Quoted 17 lines · Arnaldo Carvalho de Melo
Em Thu, Jul 15, 2021 at 04:02:16PM -0300, Arnaldo Carvalho de Melo escreveu:
Quoted 9 lines
Em Thu, Jul 15, 2021 at 06:07:06PM +0200, Riccardo Mancini escreveu:
Quoted 7 lines
ASan reports a memory leak of nsinfo during the execution of the perf test "31: Lookup mmap thread".
The leak is caused by a refcounted variable being replaced without dropping the refcount.This patch makes sure that the refcnt of nsinfo is decreased whenever a refcounted variable is replaced with a new value.
Quoted 6 lines
So, there are multiple fixes in just one patch, I'll split it into three, no need to resend.
I'll try and check if finding Fixes: for the three is easy, that way [email protected] will figure out which of the supported releases need each of them.
Quoted 5 lines
The second is at the end of this message, and looking at it I think there is a problem with where this nsinfo reference is stored, it should be in the 'struct map', that is per-thread, not in the 'struct dso', that may be shared by multiple 'struct map's in different cgroups, so we can't set
dso->nsinfoinmap__new(), right? Kristen?
Apologies for the long delayed reply, and for the hassle that this has caused.
You're right, Arnaldo, that namespace mapping could differ on a per thread basis. If the dso objects are intended to be read by multiple threads with different views of the target process's namespaces. With containers, though, there might be many different versions of the same object with the same filename, and the only ways we can disambiguate are through the build-ids or by ensuring that the dso refers to a single pid and mount namespace. I may have erred in assuming that all of perf's threads would be able to access the target namespace via the same identifiers. Is that what's going on here?
-K
Quoted 41 lines
commit
280f3933661f9ca9563ed04ce82313fdbcbb8f75Author: Riccardo Mancini <[email protected]> Date: Thu Jul 15 18:07:06 2021 +0200perf map: Fix
dso->nsinforefcountingASan reports a memory leak of nsinfo during the execution of
# perf test "31: Lookup mmap thread"
The leak is caused by a refcounted variable being replaced without dropping the refcount.
This patch makes sure that the refcnt of nsinfo is decreased whenever a refcounted variable is replaced with a new value.
Signed-off-by Riccardo Mancini <[email protected]>Fixes bf2e710b3cb8445c ("perf maps: Lookup maps in both intitial mountns and inner mountns.")Cc Ian Rogers <[email protected]>Cc Jiri Olsa <[email protected]>Cc Krister Johansen <[email protected]>Cc Mark Rutland <[email protected]>Cc Namhyung Kim <[email protected]>Cc Peter Zijlstra <[email protected]>Link http://lore.kernel.org/lkml/55223bc8821b34ccb01f92ef1401c02b6a32e61f.1626343282.git.rickyman7@gmail.com open hereSigned-off-by Arnaldo Carvalho de Melo <[email protected]>tools/perf/util/map.c +2
-0@@ -192,6 +192,8 @@ struct map *map__new(struct machine *machine, u64 start, u64 len,if (!(prot &PROT_EXEC ))dso__set_loaded (dso); } + +nsinfo__put (dso->nsinfo); dso->nsinfo = nsi;if (build_id__is_defined (bid))
Arnaldo Carvalho de Melo
maintainerRe: [PATCH 01/20] perf nsinfo: fix refcounting
Em Tue, Jan 04, 2022 at 10:19:42PM -0800, Krister Johansen escreveu:
Quoted 27 lines · Arnaldo Carvalho de Melo
Quoted 17 lines
Em Thu, Jul 15, 2021 at 04:02:16PM -0300, Arnaldo Carvalho de Melo escreveu:
Quoted 9 lines
Em Thu, Jul 15, 2021 at 06:07:06PM +0200, Riccardo Mancini escreveu:
Quoted 7 lines
ASan reports a memory leak of nsinfo during the execution of the perf test "31: Lookup mmap thread".
The leak is caused by a refcounted variable being replaced without dropping the refcount.This patch makes sure that the refcnt of nsinfo is decreased whenever a refcounted variable is replaced with a new value.
Quoted 6 lines
So, there are multiple fixes in just one patch, I'll split it into three, no need to resend.
I'll try and check if finding Fixes: for the three is easy, that way [email protected] will figure out which of the supported releases need each of them.
Quoted 5 lines
The second is at the end of this message, and looking at it I think there is a problem with where this nsinfo reference is stored, it should be in the 'struct map', that is per-thread, not in the 'struct dso', that may be shared by multiple 'struct map's in different cgroups, so we can't set
dso->nsinfoinmap__new(), right? Kristen?Apologies for the long delayed reply, and for the hassle that this has caused.
Quoted 9 lines
You're right, Arnaldo, that namespace mapping could differ on a per thread basis. If the dso objects are intended to be read by multiple threads with different views of the target process's namespaces. With containers, though, there might be many different versions of the same object with the same filename, and the only ways we can disambiguate are through the build-ids or by ensuring that the dso refers to a single pid and mount namespace. I may have erred in assuming that all of perf's threads would be able to access the target namespace via the same identifiers. Is that what's going on here?
yeah, I think using build-ids, when available is the way to go, and since we can almost take it for granted (haven't checked) that it will be present, as I think gcc/clang adds them by default, seems the way to go.
That or somehow have the DSO list to be perf namespace.
- Arnaldo
Quoted 44 lines
-K
Quoted 41 lines
commit
280f3933661f9ca9563ed04ce82313fdbcbb8f75Author: Riccardo Mancini <[email protected]> Date: Thu Jul 15 18:07:06 2021 +0200perf map: Fix
dso->nsinforefcountingASan reports a memory leak of nsinfo during the execution of
# perf test "31: Lookup mmap thread"
The leak is caused by a refcounted variable being replaced without dropping the refcount.
This patch makes sure that the refcnt of nsinfo is decreased whenever a refcounted variable is replaced with a new value.
Signed-off-by Riccardo Mancini <[email protected]>Fixes bf2e710b3cb8445c ("perf maps: Lookup maps in both intitial mountns and inner mountns.")Cc Ian Rogers <[email protected]>Cc Jiri Olsa <[email protected]>Cc Krister Johansen <[email protected]>Cc Mark Rutland <[email protected]>Cc Namhyung Kim <[email protected]>Cc Peter Zijlstra <[email protected]>Link http://lore.kernel.org/lkml/55223bc8821b34ccb01f92ef1401c02b6a32e61f.1626343282.git.rickyman7@gmail.com open hereSigned-off-by Arnaldo Carvalho de Melo <[email protected]>tools/perf/util/map.c +2
-0@@ -192,6 +192,8 @@ struct map *map__new(struct machine *machine, u64 start, u64 len,if (!(prot &PROT_EXEC ))dso__set_loaded (dso); } + +nsinfo__put (dso->nsinfo); dso->nsinfo = nsi;if (build_id__is_defined (bid))
Signature · 1 line
- Arnaldo
Arnaldo Carvalho de Melo
maintainerRe: [PATCH 01/20] perf nsinfo: fix refcounting
Em Thu, Jan 06, 2022 at 05:27:43PM -0300, Arnaldo Carvalho de Melo escreveu:
Quoted 38 lines
Em Tue, Jan 04, 2022 at 10:19:42PM -0800, Krister Johansen escreveu:
Quoted 27 lines · Arnaldo Carvalho de Melo
Quoted 17 lines
Em Thu, Jul 15, 2021 at 04:02:16PM -0300, Arnaldo Carvalho de Melo escreveu:
Quoted 9 lines
Em Thu, Jul 15, 2021 at 06:07:06PM +0200, Riccardo Mancini escreveu:
Quoted 7 lines
ASan reports a memory leak of nsinfo during the execution of the perf test "31: Lookup mmap thread".
The leak is caused by a refcounted variable being replaced without dropping the refcount.This patch makes sure that the refcnt of nsinfo is decreased whenever a refcounted variable is replaced with a new value.
Quoted 6 lines
So, there are multiple fixes in just one patch, I'll split it into three, no need to resend.
I'll try and check if finding Fixes: for the three is easy, that way [email protected] will figure out which of the supported releases need each of them.
Quoted 5 lines
The second is at the end of this message, and looking at it I think there is a problem with where this nsinfo reference is stored, it should be in the 'struct map', that is per-thread, not in the 'struct dso', that may be shared by multiple 'struct map's in different cgroups, so we can't set
dso->nsinfoinmap__new(), right? Kristen?Apologies for the long delayed reply, and for the hassle that this has caused.
Quoted 9 lines
You're right, Arnaldo, that namespace mapping could differ on a per thread basis. If the dso objects are intended to be read by multiple threads with different views of the target process's namespaces. With containers, though, there might be many different versions of the same object with the same filename, and the only ways we can disambiguate are through the build-ids or by ensuring that the dso refers to a single pid and mount namespace. I may have erred in assuming that all of perf's threads would be able to access the target namespace via the same identifiers. Is that what's going on here?
Quoted 6 lines
yeah, I think using build-ids, when available is the way to go, and since we can almost take it for granted (haven't checked) that it will be present, as I think gcc/clang adds them by default, seems the way to go.
That or somehow have the DSO list to be perf namespace.
To be by namespace, sorry.
But having it by build-id seems the more sensible way to go, so that we can avoid having multiple instances of the same DSO in multiple namespaces, which should be common.
I.e. continue to have a perf 'struct machine' DSO list instead of moving to having the namespace hierarchy represented under 'struct machine' since we have a way to uniquify DSOs, the build id.
One way or the other, nsinfo shouldn't be stored in DSOs, its a map property, which in turn is a thread property.
I.e. please consider thinking about moving it to 'struct map', would you try it, please?
- Arnaldo
Quoted 50 lines
- Arnaldo
Quoted 44 lines
-K
Quoted 41 lines
commit
280f3933661f9ca9563ed04ce82313fdbcbb8f75Author: Riccardo Mancini <[email protected]> Date: Thu Jul 15 18:07:06 2021 +0200perf map: Fix
dso->nsinforefcountingASan reports a memory leak of nsinfo during the execution of
# perf test "31: Lookup mmap thread"
The leak is caused by a refcounted variable being replaced without dropping the refcount.
This patch makes sure that the refcnt of nsinfo is decreased whenever a refcounted variable is replaced with a new value.
Signed-off-by Riccardo Mancini <[email protected]>Fixes bf2e710b3cb8445c ("perf maps: Lookup maps in both intitial mountns and inner mountns.")Cc Ian Rogers <[email protected]>Cc Jiri Olsa <[email protected]>Cc Krister Johansen <[email protected]>Cc Mark Rutland <[email protected]>Cc Namhyung Kim <[email protected]>Cc Peter Zijlstra <[email protected]>Link http://lore.kernel.org/lkml/55223bc8821b34ccb01f92ef1401c02b6a32e61f.1626343282.git.rickyman7@gmail.com open hereSigned-off-by Arnaldo Carvalho de Melo <[email protected]>tools/perf/util/map.c +2
-0@@ -192,6 +192,8 @@ struct map *map__new(struct machine *machine, u64 start, u64 len,if (!(prot &PROT_EXEC ))dso__set_loaded (dso); } + +nsinfo__put (dso->nsinfo); dso->nsinfo = nsi;if (build_id__is_defined (bid))Signature · 1 line
- Arnaldo
Signature · 1 line
- Arnaldo
Arnaldo Carvalho de Melo
maintainerRe: [PATCH 01/20] perf nsinfo: fix refcounting
In reply to Arnaldo Carvalho de MeloEm Thu, Jul 15, 2021 at 04:02:16PM -0300, Arnaldo Carvalho de Melo escreveu:
Quoted 9 lines
Em Thu, Jul 15, 2021 at 06:07:06PM +0200, Riccardo Mancini escreveu:
Quoted 7 lines
ASan reports a memory leak of nsinfo during the execution of the perf test "31: Lookup mmap thread".
The leak is caused by a refcounted variable being replaced without dropping the refcount.This patch makes sure that the refcnt of nsinfo is decreased whenever a refcounted variable is replaced with a new value.
Quoted 6 lines
So, there are multiple fixes in just one patch, I'll split it into three, no need to resend.
I'll try and check if finding Fixes: for the three is easy, that way [email protected] will figure out which of the supported releases need each of them.
3rd:
commit 7416120498f83963f4c5662847a1a0556e771b89
Author: Riccardo Mancini <[email protected]> Date: Thu Jul 15 18:07:06 2021 +0200
perf probe: Fix dso->nsinfo refcounting
ASan reports a memory leak of nsinfo during the execution of:
# perf test "31: Lookup mmap thread".
The leak is caused by a refcounted variable being replaced without dropping the refcount.
This patch makes sure that the refcnt of nsinfo is decreased whenever a refcounted variable is replaced with a new value.
tools/perf/util/probe-event.c +3 -1
@@ -179,8 +179,10 @@ struct map *get_target_map(const char *target, struct nsinfo *nsi, bool user)struct map *map; map =dso__new_map (target); -if (map && map->dso) +if (map && map->dso) { +nsinfo__put (map->dso->nsinfo); map->dso->nsinfo =nsinfo__get (nsi); + }return map; }else {return kernel_get_module_map (target);
PATCH 2/20 perf env: fix sibling_dies memory leak
In reply to Riccardo ManciniASan reports a memory leak in perf_env while running the perf test "41: Session topology", caused by sibling_dies not being freed.
This patch adds the required free.
--- tools/perf/util/env.c | 1 + 1 file changed, 1 insertion(+)
tools/perf/util/env.c +1 -0
@@ -186,6 +186,7 @@ void perf_env__exit(struct perf_env *env)zfree (&env->cpuid);zfree (&env->cmdline);zfree (&env->cmdline_argv); +zfree (&env->sibling_dies);zfree (&env->sibling_cores);zfree (&env->sibling_threads);zfree (&env->pmu_mappings);
Signature · 1 line
2.31.1
Re: [PATCH 02/20] perf env: fix sibling_dies memory leak
Em Thu, Jul 15, 2021 at 06:07:07PM +0200, Riccardo Mancini escreveu:
Quoted 6 lines
ASan reports a memory leak in
perf_envwhile running the perf test "41: Session topology", caused bysibling_diesnot being freed.This patch adds the required free.
Fixes acae8b36cded0ee6 ("perf header: Add die information in CPU topology")
So, I CC the author of the patch and its committer, using this shell script:
$ cat ~/bin/fixes #!/bin/bash
cset=$1
if [ $# -eq 0 ] ; then
read cset
fi
git show --pretty=fuller $cset | grep '^\(Author\|Commit\): ' | sed -r 's/.*: +/Cc: /g' | grep -v "Arnaldo Carvalho de Melo" | sort -u echo Fixes: git one $cset
$
$ grep one -B1 ~/.gitconfig [alias] one = show -s --pretty='format:%h (\"%s\")' $ $ fixes acae8b36cded0ee6 Cc: Kan Liang <[email protected]> Fixes: acae8b36cded0ee6 ("perf header: Add die information in CPU topology") $
Normally I copy the cset SHA and, in vim, select it and run !fixes, i.e.
:'<,'>!fixes
Thanks, applied.
- Arnaldo
Quoted 20 lines
Signed-off-by Riccardo Mancini <[email protected]>--- tools/perf/util/env.c | 1 + 1 file changed, 1 insertion(+)tools/perf/util/env.c +1
-0@@ -186,6 +186,7 @@ void perf_env__exit(struct perf_env *env)zfree (&env->cpuid);zfree (&env->cmdline);zfree (&env->cmdline_argv); +zfree (&env->sibling_dies);zfree (&env->sibling_cores);zfree (&env->sibling_threads);zfree (&env->pmu_mappings);Signature · 1 line
2.31.1
Signature · 1 line
- Arnaldo
Re: [PATCH 02/20] perf env: fix sibling_dies memory leak
Hi Arnaldo,
thanks for reviewing this patchset and, more importantly, for manually adding all Fixes: tags. I'm so sorry for that.
Quoted 31 lines · Arnaldo Carvalho de Melo
Em Thu, Jul 15, 2021 at 06:07:07PM +0200, Riccardo Mancini escreveu:
Quoted 4 lines
ASan reports a memory leak in
perf_envwhile running the perf test "41: Session topology", caused bysibling_diesnot being freed.This patch adds the required free.
Quoted 1 line
Fixes acae8b36cded0ee6 ("perf header: Add die information in CPU topology")So, I CC the author of the patch and its committer, using this shell script:
$ cat ~/bin/fixes #!/bin/bash cset=$1 if [ $# -eq 0 ] ; then read cset fi git show --pretty=fuller $cset | grep '^\(Author\|Commit\): ' | sed -r 's/.*: +/Cc: /g' | grep -v "Arnaldo Carvalho de Melo" | sort -u echo Fixes: `git one $cset` $ $ grep one -B1 ~/.gitconfig [alias] one = show -s --pretty='format:%h (\"%s\")' $ $ fixes acae8b36cded0ee6 Cc: Kan Liang <[email protected]> Fixes: acae8b36cded0ee6 ("perf header: Add die information in CPU topology") $
Quoted 3 lines
Normally I copy the cset SHA and, in vim, select it and run !fixes, i.e.
:'<,'>!fixes
Thanks! I've added this nice script to my PATH. :)
Thanks,
Riccardo
Quoted 26 lines
Thanks, applied.
- Arnaldo
Quoted 20 lines
Signed-off-by Riccardo Mancini <[email protected]>---
tools/perf/util/env.c | 1 +1 file changed, 1 insertion(+)tools/perf/util/env.c +1
-0@@ -186,6 +186,7 @@ void perf_env__exit(struct perf_env *env) zfree(&env->cpuid); zfree(&env->cmdline); zfree(&env->cmdline_argv); + zfree(&env->sibling_dies); zfree(&env->sibling_cores); zfree(&env->sibling_threads); zfree(&env->pmu_mappings);Signature · 1 line
2.31.1
PATCH 3/20 perf test: session_topology: delete session->evlist
In reply to Riccardo ManciniASan reports a memory leak related to session->evlist while running the perf test "41: Session topology".
When perf_data is in write mode, session->evlist is owned by the caller, which should also take care of deleting it.
This patch adds the missing evlist__delete.
--- tools/perf/tests/topology.c | 1 + 1 file changed, 1 insertion(+)
tools/perf/tests/topology.c +1 -0
@@ -61,6 +61,7 @@ static int session_write_header(char *path)TEST_ASSERT_VAL ("failed to write header" , !perf_session__write_header (session, session->evlist, data.file.fd, true)); +evlist__delete (session->evlist);perf_session__delete (session);return 0 ;
Signature · 1 line
2.31.1
Re: [PATCH 03/20] perf test: session_topology: delete session->evlist
Em Thu, Jul 15, 2021 at 06:07:08PM +0200, Riccardo Mancini escreveu:
Quoted 7 lines
ASan reports a memory leak related to
session->evlistwhile running the perf test "41: Session topology".When
perf_datais in write mode,session->evlistis owned by the caller, which should also take care of deleting it.This patch adds the missing
evlist__delete.
Thanks, applied.
- Arnaldo
Quoted 20 lines
Signed-off-by Riccardo Mancini <[email protected]>--- tools/perf/tests/topology.c | 1 + 1 file changed, 1 insertion(+)tools/perf/tests/topology.c +1
-0@@ -61,6 +61,7 @@ static int session_write_header(char *path)TEST_ASSERT_VAL ("failed to write header" , !perf_session__write_header (session, session->evlist, data.file.fd, true)); +evlist__delete (session->evlist);perf_session__delete (session);return 0 ;Signature · 1 line
2.31.1
Signature · 1 line
- Arnaldo
PATCH 4/20 perf test: event_update: fix memory leak of evlist
In reply to Riccardo ManciniASan reports a memory leak when running the perf test "49: Synthesize attr update" caused by evlist not being deleted.
This patch adds the missing evlist__delete and removes the perf_cpu_map__put since it's already being deleted by evlist__delete.
--- tools/perf/tests/event_update.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
tools/perf/tests/event_update.c +1 -1
@@ -118,6 +118,6 @@ int test__event_update(struct test *test __maybe_unused, int subtest __maybe_unuTEST_ASSERT_VAL ("failed to synthesize attr update cpus" , !perf_event__synthesize_event_update_cpus (&tmp.tool, evsel, process_event_cpus)); -perf_cpu_map__put (evsel->core.own_cpus); +evlist__delete (evlist);return 0 ; }
Signature · 1 line
2.31.1
Re: [PATCH 04/20] perf test: event_update: fix memory leak of evlist
Em Thu, Jul 15, 2021 at 06:07:09PM +0200, Riccardo Mancini escreveu:
Quoted 5 lines
ASan reports a memory leak when running the perf test "49: Synthesize attr update" caused by evlist not being deleted.
This patch adds the missing
evlist__deleteand removes theperf_cpu_map__putsince it's already being deleted byevlist__delete.
Thanks, applied.
- Arnaldo
Quoted 20 lines
Signed-off-by Riccardo Mancini <[email protected]>--- tools/perf/tests/event_update.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)tools/perf/tests/event_update.c +1
-1@@ -118,6 +118,6 @@ int test__event_update(struct test *test __maybe_unused, int subtest __maybe_unuTEST_ASSERT_VAL ("failed to synthesize attr update cpus" , !perf_event__synthesize_event_update_cpus (&tmp.tool, evsel, process_event_cpus)); -perf_cpu_map__put (evsel->core.own_cpus); +evlist__delete (evlist);return 0 ; }Signature · 1 line
2.31.1
Signature · 1 line
- Arnaldo
PATCH 5/20 perf test: event_update: fix memory leak of unit
In reply to Riccardo ManciniASan reports a memory leak while running the perf test "49: Synthesize attr update" caused by a string being duplicated but never freed.
This patch adds the missing free().
Note that evsel->unit is not deallocated together with evsel since it is supposed to be a constant string.
--- tools/perf/tests/event_update.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-)
tools/perf/tests/event_update.c +3 -1
@@ -88,6 +88,7 @@ int test__event_update(struct test *test __maybe_unused, int subtest __maybe_unustruct evsel *evsel;struct event_name tmp;struct evlist *evlist =evlist__new_default (); +char *unit =strdup ("KRAVA" );TEST_ASSERT_VAL ("failed to get evlist" , evlist);
@@ -98,7 +99,7 @@ int test__event_update(struct test *test __maybe_unused, int subtest __maybe_unuperf_evlist__id_add (&evlist->core, &evsel->core,0 ,0 ,123 ); - evsel->unit =strdup ("KRAVA" ); + evsel->unit = unit;TEST_ASSERT_VAL ("failed to synthesize attr update unit" , !perf_event__synthesize_event_update_unit (NULL , evsel, process_event_unit));
@@ -118,6 +119,7 @@ int test__event_update(struct test *test __maybe_unused, int subtest __maybe_unuTEST_ASSERT_VAL ("failed to synthesize attr update cpus" , !perf_event__synthesize_event_update_cpus (&tmp.tool, evsel, process_event_cpus)); +free (unit);evlist__delete (evlist);return 0 ; }
Signature · 1 line
2.31.1
Re: [PATCH 05/20] perf test: event_update: fix memory leak of unit
Em Thu, Jul 15, 2021 at 06:07:10PM +0200, Riccardo Mancini escreveu:
Quoted 8 lines
ASan reports a memory leak while running the perf test "49: Synthesize attr update" caused by a string being duplicated but never freed.
This patch adds the missing
free().Note that
evsel->unitis not deallocated together with evsel since it is supposed to be a constant string.
Thanks, applied and added the Fixes tag, which I'm doing to all the patches in the series.
- Arnaldo
Quoted 37 lines
Signed-off-by Riccardo Mancini <[email protected]>--- tools/perf/tests/event_update.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-)tools/perf/tests/event_update.c +3
-1@@ -88,6 +88,7 @@ int test__event_update(struct test *test __maybe_unused, int subtest __maybe_unustruct evsel *evsel;struct event_name tmp;struct evlist *evlist =evlist__new_default (); +char *unit =strdup ("KRAVA" );TEST_ASSERT_VAL ("failed to get evlist" , evlist);@@ -98,7 +99,7 @@ int test__event_update(struct test *test __maybe_unused, int subtest __maybe_unuperf_evlist__id_add (&evlist->core, &evsel->core,0 ,0 ,123 ); - evsel->unit =strdup ("KRAVA" ); + evsel->unit = unit;TEST_ASSERT_VAL ("failed to synthesize attr update unit" , !perf_event__synthesize_event_update_unit (NULL , evsel, process_event_unit));@@ -118,6 +119,7 @@ int test__event_update(struct test *test __maybe_unused, int subtest __maybe_unuTEST_ASSERT_VAL ("failed to synthesize attr update cpus" , !perf_event__synthesize_event_update_cpus (&tmp.tool, evsel, process_event_cpus)); +free (unit);evlist__delete (evlist);return 0 ; }Signature · 1 line
2.31.1
Signature · 1 line
- Arnaldo
PATCH 6/20 perf dso: fix memory leak in dso__new_map
In reply to Riccardo ManciniASan reports a memory leak when running the perf test "65: maps__merge_in". The causes of the leaks are two, this patch addresses only the first one, which is related to dso__new_map. The bug is that dso__new_map creates a new dso but never decreases the refcount it gets from creating it.
This patch adds the missing dso__put.
--- tools/perf/util/dso.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-)
tools/perf/util/dso.c +3 -1
@@ -1154,8 +1154,10 @@ struct map *dso__new_map(const char *name)struct map *map =NULL ;struct dso *dso =dso__new (name); -if (dso) +if (dso) { map =map__new2 (0 , dso); +dso__put (dso); + }return map; }
Signature · 1 line
2.31.1
Re: [PATCH 06/20] perf dso: fix memory leak in dso__new_map
Em Thu, Jul 15, 2021 at 06:07:11PM +0200, Riccardo Mancini escreveu:
Quoted 7 lines
ASan reports a memory leak when running the perf test "65:
maps__merge_in". The causes of the leaks are two, this patch addresses only the first one, which is related todso__new_map. The bug is thatdso__new_mapcreates a new dso but never decreases the refcount it gets from creating it.This patch adds the missing
dso__put.
Thanks, applied.
This needs some extra thinking for some unrelated issue, namely that this dso isn't linked in the machine dso list, I think this is only used by 'perf probe' :-\
I guess this won't be a problem since using the machine dso list is important just for other tools, like when traversing to get the build-ids, etc.
- Arnaldo
Quoted 24 lines
Signed-off-by Riccardo Mancini <[email protected]>--- tools/perf/util/dso.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-)tools/perf/util/dso.c +3
-1@@ -1154,8 +1154,10 @@ struct map *dso__new_map(const char *name)struct map *map =NULL ;struct dso *dso =dso__new (name); -if (dso) +if (dso) { map =map__new2 (0 , dso); +dso__put (dso); + }return map; }Signature · 1 line
2.31.1
Signature · 1 line
- Arnaldo
Re: [PATCH 06/20] perf dso: fix memory leak in dso__new_map
In reply to Riccardo ManciniEm Thu, Jul 15, 2021 at 06:07:11PM +0200, Riccardo Mancini escreveu:
Quoted 7 lines
ASan reports a memory leak when running the perf test "65:
maps__merge_in". The causes of the leaks are two, this patch addresses only the first one, which is related todso__new_map. The bug is thatdso__new_mapcreates a new dso but never decreases the refcount it gets from creating it.This patch adds the missing
dso__put.
Thanks, applied.
- Arnaldo
Quoted 24 lines
Signed-off-by Riccardo Mancini <[email protected]>--- tools/perf/util/dso.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-)tools/perf/util/dso.c +3
-1@@ -1154,8 +1154,10 @@ struct map *dso__new_map(const char *name)struct map *map =NULL ;struct dso *dso =dso__new (name); -if (dso) +if (dso) { map =map__new2 (0 , dso); +dso__put (dso); + }return map; }Signature · 1 line
2.31.1
Signature · 1 line
- Arnaldo
PATCH 7/20 perf test: maps__merge_in: fix memory leak of maps
In reply to Riccardo ManciniASan reports a memory leak when running the perf test "65: maps__merge_in". This is the second and final patch addressing these memory leaks.
This time, the problem is simply that the maps object is never destructed.
This patch adds the missing maps__exit call.
--- tools/perf/tests/maps.c | 2 ++ 1 file changed, 2 insertions(+)
tools/perf/tests/maps.c +2 -0
@@ -116,5 +116,7 @@ int test__maps__merge_in(struct test *t __maybe_unused, int subtest __maybe_unus ret =check_maps (merged3,ARRAY_SIZE (merged3), &maps);TEST_ASSERT_VAL ("merge check failed" , !ret); + +maps__exit (&maps);return TEST_OK ; }
Signature · 1 line
2.31.1
Re: [PATCH 07/20] perf test: maps__merge_in: fix memory leak of maps
Em Thu, Jul 15, 2021 at 06:07:12PM +0200, Riccardo Mancini escreveu:
Quoted 7 lines
ASan reports a memory leak when running the perf test "65:
maps__merge_in". This is the second and final patch addressing these memory leaks.
This time, the problem is simply that the maps object is never destructed.This patch adds the missing
maps__exitcall.
Not really this one, but at least the patch should apply more easily:
Thanks, applied.
- Arnaldo
Quoted 20 lines
Signed-off-by Riccardo Mancini <[email protected]>--- tools/perf/tests/maps.c | 2 ++ 1 file changed, 2 insertions(+)tools/perf/tests/maps.c +2
-0@@ -116,5 +116,7 @@ int test__maps__merge_in(struct test *t __maybe_unused, int subtest __maybe_unus ret =check_maps (merged3,ARRAY_SIZE (merged3), &maps);TEST_ASSERT_VAL ("merge check failed" , !ret); + +maps__exit (&maps);return TEST_OK ; }Signature · 1 line
2.31.1
Signature · 1 line
- Arnaldo
PATCH 8/20 perf env: fix memory leak of cpu_pmu_caps
In reply to Riccardo ManciniASan reports memory leaks while running the perf test
"83: Zstd perf.data compression/decompression".
The first of the leaks is caused by env->cpu_pmu_caps not being freed.
This patch adds the missing (z)free inside perf_env__exit.
--- tools/perf/util/env.c | 1 + 1 file changed, 1 insertion(+)
tools/perf/util/env.c +1 -0
@@ -191,6 +191,7 @@ void perf_env__exit(struct perf_env *env)zfree (&env->sibling_threads);zfree (&env->pmu_mappings);zfree (&env->cpu); +zfree (&env->cpu_pmu_caps);zfree (&env->numa_map);for (i =0 ; i < env->nr_numa_nodes; i++)
Signature · 1 line
2.31.1
Re: [PATCH 08/20] perf env: fix memory leak of cpu_pmu_caps
Em Thu, Jul 15, 2021 at 06:07:13PM +0200, Riccardo Mancini escreveu:
Quoted 5 lines
ASan reports memory leaks while running the perf test
"83: Zstd perf.data compression/decompression".
The first of the leaks is caused byenv->cpu_pmu_capsnot being freed.This patch adds the missing (z)free inside
perf_env__exit.
Thanks, applied.
- Arnaldo
Quoted 20 lines
Signed-off-by Riccardo Mancini <[email protected]>--- tools/perf/util/env.c | 1 + 1 file changed, 1 insertion(+)tools/perf/util/env.c +1
-0@@ -191,6 +191,7 @@ void perf_env__exit(struct perf_env *env)zfree (&env->sibling_threads);zfree (&env->pmu_mappings);zfree (&env->cpu); +zfree (&env->cpu_pmu_caps);zfree (&env->numa_map);for (i =0 ; i < env->nr_numa_nodes; i++)Signature · 1 line
2.31.1
Signature · 1 line
- Arnaldo
PATCH 9/20 perf report: free generated help strings for sort option
In reply to Riccardo ManciniASan reports the memory leak of the strings allocated by sort_help when running perf report.
This patch changes the returned pointer to char* (instead of const char*), saves it in a temporary variable, and finally deallocates it at function exit.
--- tools/perf/builtin-report.c | 33 ++++++++++++++++++++++----------- tools/perf/util/sort.c | 2 +- tools/perf/util/sort.h | 2 +- 3 files changed, 24 insertions(+), 13 deletions(-)
tools/perf/builtin-report.c +22 -11
@@ -1175,6 +1175,8 @@ int cmd_report(int argc, const char **argv) .annotation_opts = annotation__default_options, .skip_empty = true, }; +char *sort_order_help =sort_help ("sort by key(s):" ); +char *field_order_help =sort_help ("output field(s): overhead period sample " );const struct option options[] = {OPT_STRING ('i' ,"input" , &input_name,"file" ,"input file name" ),
@@ -1209,9 +1211,9 @@ int cmd_report(int argc, const char **argv)OPT_BOOLEAN (0 ,"header-only" , &report.header_only,"Show only data header." ),OPT_STRING ('s' ,"sort" , &sort_order,"key[,key2...]" , -sort_help ("sort by key(s):" )), + sort_order_help),OPT_STRING ('F' ,"fields" , &field_order,"key[,keys...]" , -sort_help ("output field(s): overhead period sample " )), + field_order_help),OPT_BOOLEAN (0 ,"show-cpu-utilization" , &symbol_conf.show_cpu_utilization,"Show sample percentage for different cpu modes" ), OPT_BOOLEAN_FLAG(0, "showcpuutilization", &symbol_conf.show_cpu_utilization,
@@ -1344,11 +1346,11 @@ int cmd_report(int argc, const char **argv)char sort_tmp[128 ];if (ret <0 ) -return ret; + goto exit; ret =perf_config (report__config, &report);if (ret) -return ret; + goto exit; argc =parse_options (argc, argv, options, report_usage,0 );if (argc) {
@@ -1362,8 +1364,10 @@ int cmd_report(int argc, const char **argv) report.symbol_filter_str = argv[0 ]; } -if (annotate_check_args (&report.annotation_opts) <0 ) -return -EINVAL ; +if (annotate_check_args (&report.annotation_opts) <0 ) { + ret = -EINVAL ; + goto exit; + }if (report.mmaps_mode) report.tasks_mode = true;
@@ -1377,12 +1381,14 @@ int cmd_report(int argc, const char **argv)if (symbol_conf.vmlinux_name &&access (symbol_conf.vmlinux_name,R_OK )) {pr_err ("Invalid file: %s\n" , symbol_conf.vmlinux_name); -return -EINVAL ; + ret = -EINVAL ; + goto exit; }if (symbol_conf.kallsyms_name &&access (symbol_conf.kallsyms_name,R_OK )) {pr_err ("Invalid file: %s\n" , symbol_conf.kallsyms_name); -return -EINVAL ; + ret = -EINVAL ; + goto exit; }if (report.inverted_callchain)
@@ -1406,12 +1412,14 @@ int cmd_report(int argc, const char **argv) repeat: session =perf_session__new (&data, false, &report.tool); -if (IS_ERR (session)) -return PTR_ERR (session); +if (IS_ERR (session)) { + ret =PTR_ERR (session); + goto exit; + } ret =evswitch__init (&report.evswitch, session->evlist, stderr);if (ret) -return ret; + goto exit;if (zstd_init (&(session->zstd_data),0 ) <0 )pr_warning ("Decompression initialization failed. Reported data may be incomplete.\n" );
@@ -1646,5 +1654,8 @@ int cmd_report(int argc, const char **argv)zstd_fini (&(session->zstd_data));perf_session__delete (session); +exit: +free (sort_order_help); +free (field_order_help);return ret; }
tools/perf/util/sort.c +1 -1
@@ -3370,7 +3370,7 @@ static void add_hpp_sort_string(struct strbuf *sb, struct hpp_dimension *s, intadd_key (sb, s[i].name, llen); } -const char *sort_help (const char *prefix) +char *sort_help (const char *prefix) {struct strbuf sb;char *s;
tools/perf/util/sort.h +1 -1
@@ -302,7 +302,7 @@ void reset_output_field(void);void sort__setup_elide (FILE *fp);void perf_hpp__set_elide (int idx,bool elide); -const char *sort_help (const char *prefix); +char *sort_help (const char *prefix);int report_parse_ignore_callees_opt (const struct option *opt,const char *arg,int unset);
Signature · 1 line
2.31.1
Re: [PATCH 09/20] perf report: free generated help strings for sort option
Em Thu, Jul 15, 2021 at 06:07:14PM +0200, Riccardo Mancini escreveu:
Quoted 6 lines
ASan reports the memory leak of the strings allocated by
sort_helpwhen running perf report.This patch changes the returned pointer to char* (instead of const char*), saves it in a temporary variable, and finally deallocates it at function exit.
Thanks, applied.
- Arnaldo
Quoted 132 lines
Signed-off-by Riccardo Mancini <[email protected]>--- tools/perf/builtin-report.c | 33 ++++++++++++++++++++++----------- tools/perf/util/sort.c | 2 +- tools/perf/util/sort.h | 2 +- 3 files changed, 24 insertions(+), 13 deletions(-)tools/perf/builtin-report.c +22
-11@@ -1175,6 +1175,8 @@ int cmd_report(int argc, const char **argv) .annotation_opts = annotation__default_options, .skip_empty = true, }; +char *sort_order_help =sort_help ("sort by key(s):" ); +char *field_order_help =sort_help ("output field(s): overhead period sample " );const struct option options[] = {OPT_STRING ('i' ,"input" , &input_name,"file" ,"input file name" ),@@ -1209,9 +1211,9 @@ int cmd_report(int argc, const char **argv)OPT_BOOLEAN (0 ,"header-only" , &report.header_only,"Show only data header." ),OPT_STRING ('s' ,"sort" , &sort_order,"key[,key2...]" , -sort_help ("sort by key(s):" )), + sort_order_help),OPT_STRING ('F' ,"fields" , &field_order,"key[,keys...]" , -sort_help ("output field(s): overhead period sample " )), + field_order_help),OPT_BOOLEAN (0 ,"show-cpu-utilization" , &symbol_conf.show_cpu_utilization,"Show sample percentage for different cpu modes" ), OPT_BOOLEAN_FLAG(0, "showcpuutilization", &symbol_conf.show_cpu_utilization,@@ -1344,11 +1346,11 @@ int cmd_report(int argc, const char **argv)char sort_tmp[128 ];if (ret <0 ) -return ret; + goto exit; ret =perf_config (report__config, &report);if (ret) -return ret; + goto exit; argc =parse_options (argc, argv, options, report_usage,0 );if (argc) {@@ -1362,8 +1364,10 @@ int cmd_report(int argc, const char **argv) report.symbol_filter_str = argv[0 ]; } -if (annotate_check_args (&report.annotation_opts) <0 ) -return -EINVAL ; +if (annotate_check_args (&report.annotation_opts) <0 ) { + ret = -EINVAL ; + goto exit; + }if (report.mmaps_mode) report.tasks_mode = true;@@ -1377,12 +1381,14 @@ int cmd_report(int argc, const char **argv)if (symbol_conf.vmlinux_name &&access (symbol_conf.vmlinux_name,R_OK )) {pr_err ("Invalid file: %s\n" , symbol_conf.vmlinux_name); -return -EINVAL ; + ret = -EINVAL ; + goto exit; }if (symbol_conf.kallsyms_name &&access (symbol_conf.kallsyms_name,R_OK )) {pr_err ("Invalid file: %s\n" , symbol_conf.kallsyms_name); -return -EINVAL ; + ret = -EINVAL ; + goto exit; }if (report.inverted_callchain)@@ -1406,12 +1412,14 @@ int cmd_report(int argc, const char **argv) repeat: session =perf_session__new (&data, false, &report.tool); -if (IS_ERR (session)) -return PTR_ERR (session); +if (IS_ERR (session)) { + ret =PTR_ERR (session); + goto exit; + } ret =evswitch__init (&report.evswitch, session->evlist, stderr);if (ret) -return ret; + goto exit;if (zstd_init (&(session->zstd_data),0 ) <0 )pr_warning ("Decompression initialization failed. Reported data may be incomplete.\n" );@@ -1646,5 +1654,8 @@ int cmd_report(int argc, const char **argv)zstd_fini (&(session->zstd_data));perf_session__delete (session); +exit: +free (sort_order_help); +free (field_order_help);return ret; }tools/perf/util/sort.c +1
-1@@ -3370,7 +3370,7 @@ static void add_hpp_sort_string(struct strbuf *sb, struct hpp_dimension *s, intadd_key (sb, s[i].name, llen); } -const char *sort_help (const char *prefix) +char *sort_help (const char *prefix) {struct strbuf sb;char *s;tools/perf/util/sort.h +1
-1@@ -302,7 +302,7 @@ void reset_output_field(void);void sort__setup_elide (FILE *fp);void perf_hpp__set_elide (int idx,bool elide); -const char *sort_help (const char *prefix); +char *sort_help (const char *prefix);int report_parse_ignore_callees_opt (const struct option *opt,const char *arg,int unset);Signature · 1 line
2.31.1
Signature · 1 line
- Arnaldo
PATCH 10/20 perf inject: close inject.output
In reply to Riccardo ManciniASan reports a memory leak when running the perf test "83: Zstd perf.data compression/decompression", which happens inside perf-inject.
The bug is caused by inject.output never being closed.
This patch adds the missing perf_data__close.
--- tools/perf/builtin-inject.c | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-)
tools/perf/builtin-inject.c +6 -2
@@ -993,8 +993,10 @@ int cmd_inject(int argc, const char **argv) data.path = inject.input_name; inject.session =perf_session__new (&data, inject.output.is_pipe, &inject.tool); -if (IS_ERR (inject.session)) -return PTR_ERR (inject.session); +if (IS_ERR (inject.session)) { + ret =PTR_ERR (inject.session); + goto out_close_output; + }if (zstd_init (&(inject.session->zstd_data),0 ) <0 )pr_warning ("Decompression initialization failed.\n" );
@@ -1036,6 +1038,8 @@ int cmd_inject(int argc, const char **argv) out_delete:zstd_fini (&(inject.session->zstd_data));perf_session__delete (inject.session); +out_close_output: +perf_data__close (&inject.output);free (inject.itrace_synth_opts.vm_tm_corr_args);return ret; }
Signature · 1 line
2.31.1
Re: [PATCH 10/20] perf inject: close inject.output
Em Thu, Jul 15, 2021 at 06:07:15PM +0200, Riccardo Mancini escreveu:
Quoted 6 lines
ASan reports a memory leak when running the perf test "83: Zstd perf.data compression/decompression", which happens inside perf-inject.
The bug is caused by inject.output never being closed.This patch adds the missing
perf_data__close.
Thanks, applied.
- Arnaldo
Quoted 34 lines
Signed-off-by Riccardo Mancini <[email protected]>--- tools/perf/builtin-inject.c | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-)tools/perf/builtin-inject.c +6
-2@@ -993,8 +993,10 @@ int cmd_inject(int argc, const char **argv) data.path = inject.input_name; inject.session =perf_session__new (&data, inject.output.is_pipe, &inject.tool); -if (IS_ERR (inject.session)) -return PTR_ERR (inject.session); +if (IS_ERR (inject.session)) { + ret =PTR_ERR (inject.session); + goto out_close_output; + }if (zstd_init (&(inject.session->zstd_data),0 ) <0 )pr_warning ("Decompression initialization failed.\n" );@@ -1036,6 +1038,8 @@ int cmd_inject(int argc, const char **argv) out_delete:zstd_fini (&(inject.session->zstd_data));perf_session__delete (inject.session); +out_close_output: +perf_data__close (&inject.output);free (inject.itrace_synth_opts.vm_tm_corr_args);return ret; }Signature · 1 line
2.31.1
Signature · 1 line
- Arnaldo
PATCH 11/20 perf session: cleanup trace_event
In reply to Riccardo ManciniASan reports several memory leaks when running the perf test "82: Use vfs_getname probe to get syscall args filenames", many of which are related to session->tevent.
This patch will solve this problem, then next patch will fix the remaining memory leaks in perf-script.
This bug is due to a missing deallocation of the trace_event data strutures.
This patch adds the missing trace_event__cleanup in perf_session__delete.
--- tools/perf/util/session.c | 1 + 1 file changed, 1 insertion(+)
tools/perf/util/session.c +1 -0
@@ -306,6 +306,7 @@ void perf_session__delete(struct perf_session *session)evlist__delete (session->evlist);perf_data__close (session->data); } +trace_event__cleanup (&session->tevent);free (session); }
Signature · 1 line
2.31.1
Re: [PATCH 11/20] perf session: cleanup trace_event
Em Thu, Jul 15, 2021 at 06:07:16PM +0200, Riccardo Mancini escreveu:
Quoted 10 lines
ASan reports several memory leaks when running the perf test "82: Use
vfs_getnameprobe to get syscall args filenames", many of which are related tosession->tevent.
This patch will solve this problem, then next patch will fix the remaining memory leaks in perf-script.
This bug is due to a missing deallocation of thetrace_eventdata strutures.This patch adds the missing
trace_event__cleanupinperf_session__delete.
Thanks, applied.
- Arnaldo
Quoted 20 lines
Signed-off-by Riccardo Mancini <[email protected]>--- tools/perf/util/session.c | 1 + 1 file changed, 1 insertion(+)tools/perf/util/session.c +1
-0@@ -306,6 +306,7 @@ void perf_session__delete(struct perf_session *session)evlist__delete (session->evlist);perf_data__close (session->data); } +trace_event__cleanup (&session->tevent);free (session); }Signature · 1 line
2.31.1
Signature · 1 line
- Arnaldo
PATCH 12/20 perf script: release zstd data
In reply to Riccardo ManciniASan reports several memory leak while running the perf test "82: Use vfs_getname probe to get syscall args filenames". One of the leaks is caused by zstd data not being released on exit in perf-script.
This patch adds the missing zstd_fini.
--- tools/perf/builtin-script.c | 1 + 1 file changed, 1 insertion(+)
tools/perf/builtin-script.c +1 -0
@@ -4143,6 +4143,7 @@ int cmd_script(int argc, const char **argv)zfree (&script.ptime_range); } +zstd_fini (&(session->zstd_data));evlist__free_stats (session->evlist);perf_session__delete (session);
Signature · 1 line
2.31.1
Re: [PATCH 12/20] perf script: release zstd data
Em Thu, Jul 15, 2021 at 06:07:17PM +0200, Riccardo Mancini escreveu:
Quoted 6 lines
ASan reports several memory leak while running the perf test "82: Use
vfs_getnameprobe to get syscall args filenames". One of the leaks is caused by zstd data not being released on exit in perf-script.This patch adds the missing
zstd_fini.
Thanks, applied.
- Arnaldo
Quoted 20 lines
Signed-off-by Riccardo Mancini <[email protected]>--- tools/perf/builtin-script.c | 1 + 1 file changed, 1 insertion(+)tools/perf/builtin-script.c +1
-0@@ -4143,6 +4143,7 @@ int cmd_script(int argc, const char **argv)zfree (&script.ptime_range); } +zstd_fini (&(session->zstd_data));evlist__free_stats (session->evlist);perf_session__delete (session);Signature · 1 line
2.31.1
Signature · 1 line
- Arnaldo
PATCH 13/20 perf script: fix memory leaks in perf_script
In reply to Riccardo ManciniASan reports several memory leaks while running the perf test "82: Use vfs_getname probe to get syscall args filenames". Two of these are caused by some refcounts not being decreased on perf-script exit, namely script.threads and script.cpus.
This patch adds the missing __put calls in a new perf_script__exit function, which is called at the end of cmd_script.
This patch concludes the fixes of all remaining memory leaks in perf test "82: Use vfs_getname probe to get syscall args filenames".
--- tools/perf/builtin-script.c | 7 +++++++ 1 file changed, 7 insertions(+)
tools/perf/builtin-script.c +7 -0
@@ -2601,6 +2601,12 @@ static void perf_script__exit_per_event_dump_stats(struct perf_script *script) } } +static void perf_script__exit (struct perf_script *script) +{ +perf_thread_map__put (script->threads); +perf_cpu_map__put (script->cpus); +} +static int __cmd_script (struct perf_script *script) {int ret;
@@ -4146,6 +4152,7 @@ int cmd_script(int argc, const char **argv)zstd_fini (&(session->zstd_data));evlist__free_stats (session->evlist);perf_session__delete (session); +perf_script__exit (&script);if (script_started)cleanup_scripting ();
Signature · 1 line
2.31.1
Re: [PATCH 13/20] perf script: fix memory leaks in perf_script
Em Thu, Jul 15, 2021 at 06:07:18PM +0200, Riccardo Mancini escreveu:
Quoted 10 lines
ASan reports several memory leaks while running the perf test "82: Use
vfs_getnameprobe to get syscall args filenames". Two of these are caused by some refcounts not being decreased on perf-script exit, namely script.threads and script.cpus.This patch adds the missing
__putcalls in a newperf_script__exitfunction, which is called at the end ofcmd_script.This patch concludes the fixes of all remaining memory leaks in perf test "82: Use
vfs_getnameprobe to get syscall args filenames".
Thanks, applied.
- Arnaldo
Quoted 33 lines
Signed-off-by Riccardo Mancini <[email protected]>--- tools/perf/builtin-script.c | 7 +++++++ 1 file changed, 7 insertions(+)tools/perf/builtin-script.c +7
-0@@ -2601,6 +2601,12 @@ static void perf_script__exit_per_event_dump_stats(struct perf_script *script) } } +static void perf_script__exit (struct perf_script *script) +{ +perf_thread_map__put (script->threads); +perf_cpu_map__put (script->cpus); +} +static int __cmd_script (struct perf_script *script) {int ret;@@ -4146,6 +4152,7 @@ int cmd_script(int argc, const char **argv)zstd_fini (&(session->zstd_data));evlist__free_stats (session->evlist);perf_session__delete (session); +perf_script__exit (&script);if (script_started)cleanup_scripting ();Signature · 1 line
2.31.1
Signature · 1 line
- Arnaldo
PATCH 14/20 perf util/lzma: close lzma stream
In reply to Riccardo ManciniASan reports memory leaks when running the perf test "88: Check open filename arg using perf trace + vfs_getname". One of these is caused by the lzma stream never being closed inside lzma_decompress_to_file.
This patch adds the missing lzma_end.
--- tools/perf/util/lzma.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)
tools/perf/util/lzma.c +5 -3
@@ -69,7 +69,7 @@ int lzma_decompress_to_file(const char *input, int output_fd)if (ferror (infile)) {pr_err ("lzma: read error: %s\n" ,strerror (errno)); - goto err_fclose; + goto err_lzma_end; }if (feof (infile))
@@ -83,7 +83,7 @@ int lzma_decompress_to_file(const char *input, int output_fd)if (writen (output_fd, buf_out, write_size) != write_size) {pr_err ("lzma: write error: %s\n" ,strerror (errno)); - goto err_fclose; + goto err_lzma_end; } strm.next_out = buf_out;
@@ -95,11 +95,13 @@ int lzma_decompress_to_file(const char *input, int output_fd)break ;pr_err ("lzma: failed %s\n" ,lzma_strerror (ret)); - goto err_fclose; + goto err_lzma_end; } } err =0 ; +err_lzma_end: +lzma_end (&strm); err_fclose:fclose (infile);return err;
Signature · 1 line
2.31.1
Re: [PATCH 14/20] perf util/lzma: close lzma stream
Em Thu, Jul 15, 2021 at 06:07:19PM +0200, Riccardo Mancini escreveu:
Quoted 6 lines
ASan reports memory leaks when running the perf test "88: Check open filename arg using perf trace +
vfs_getname". One of these is caused by the lzma stream never being closed insidelzma_decompress_to_file.This patch adds the missing
lzma_end.
Thanks, applied.
- Arnaldo
Quoted 45 lines
Signed-off-by Riccardo Mancini <[email protected]>--- tools/perf/util/lzma.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)tools/perf/util/lzma.c +5
-3@@ -69,7 +69,7 @@ int lzma_decompress_to_file(const char *input, int output_fd)if (ferror (infile)) {pr_err ("lzma: read error: %s\n" ,strerror (errno)); - goto err_fclose; + goto err_lzma_end; }if (feof (infile))@@ -83,7 +83,7 @@ int lzma_decompress_to_file(const char *input, int output_fd)if (writen (output_fd, buf_out, write_size) != write_size) {pr_err ("lzma: write error: %s\n" ,strerror (errno)); - goto err_fclose; + goto err_lzma_end; } strm.next_out = buf_out;@@ -95,11 +95,13 @@ int lzma_decompress_to_file(const char *input, int output_fd)break ;pr_err ("lzma: failed %s\n" ,lzma_strerror (ret)); - goto err_fclose; + goto err_lzma_end; } } err =0 ; +err_lzma_end: +lzma_end (&strm); err_fclose:fclose (infile);return err;Signature · 1 line
2.31.1
Signature · 1 line
- Arnaldo
PATCH 15/20 perf trace: free malloc'd trace fields on exit
In reply to Riccardo ManciniASan reports several memory leaks running the perf test "88: Check open filename arg using perf trace + vfs_getname". The first of these leaks is related to struct trace fields never being deallocated.
This patch adds the function trace__exit, which is called at the end of cmd_trace, replacing the existing deallocation, which is now moved inside the new function.
This function deallocates:
- ev_qualifier
- ev_qualifier_ids.entries
- syscalls.table
- sctbl
- perfconfig_events
In order to prevent errors in case any of this field is never initialized, this patch also adds initialization to NULL to these fields.
--- tools/perf/builtin-trace.c | 17 ++++++++++++++++- 1 file changed, 16 insertions(+), 1 deletion(-)
tools/perf/builtin-trace.c +16 -1
@@ -4701,6 +4701,15 @@ static int trace__config(const char *var, const char *value, void *arg)return err; } +static void trace__exit (struct trace *trace) +{ +strlist__delete (trace->ev_qualifier); +free (trace->ev_qualifier_ids.entries); +free (trace->syscalls.table); +syscalltbl__delete (trace->sctbl); +zfree (&trace->perfconfig_events); +} +int cmd_trace (int argc,const char **argv) {const char *trace_usage[] = {
@@ -4731,6 +4740,12 @@ int cmd_trace(int argc, const char **argv) .kernel_syscallchains = false, .max_stack = UINT_MAX, .max_events = ULONG_MAX, + .ev_qualifier = NULL, + .sctbl = NULL, + .ev_qualifier_ids = { + .entries = NULL, + .nr = 0, + }, };const char *map_dump_str =NULL ;const char *output_name =NULL ;
@@ -5135,6 +5150,6 @@ int cmd_trace(int argc, const char **argv)if (output_name !=NULL )fclose (trace.output); out: -zfree (&trace.perfconfig_events); +trace__exit (&trace);return err; }
Signature · 1 line
2.31.1
Re: [PATCH 15/20] perf trace: free malloc'd trace fields on exit
Em Thu, Jul 15, 2021 at 06:07:20PM +0200, Riccardo Mancini escreveu:
Quoted 47 lines
ASan reports several memory leaks running the perf test "88: Check open filename arg using perf trace +
vfs_getname". The first of these leaks is related to struct trace fields never being deallocated.This patch adds the function
trace__exit, which is called at the end ofcmd_trace, replacing the existing deallocation, which is now moved inside the new function.
This function deallocates:
-ev_qualifier
- ev_qualifier_ids.entries
- syscalls.table
- sctbl
-perfconfig_eventsIn order to prevent errors in case any of this field is never initialized, this patch also adds initialization to NULL to these fields.
Signed-off-by Riccardo Mancini <[email protected]>--- tools/perf/builtin-trace.c | 17 ++++++++++++++++- 1 file changed, 16 insertions(+), 1 deletion(-)tools/perf/builtin-trace.c +9
-0@@ -4701,6 +4701,15 @@ static int trace__config(const char *var, const char *value, void *arg)return err; } +static void trace__exit (struct trace *trace) +{ +strlist__delete (trace->ev_qualifier); +free (trace->ev_qualifier_ids.entries); +free (trace->syscalls.table); +syscalltbl__delete (trace->sctbl); +zfree (&trace->perfconfig_events); +} +int cmd_trace (int argc,const char **argv) {const char *trace_usage[] = {@@ -4731,6 +4740,12 @@ int cmd_trace(int argc, const char **argv) .kernel_syscallchains = false, .max_stack = UINT_MAX,
tools/perf/builtin-trace.c int cmd_trace(int argc, const char **argv) · line 4751
Quoted 10 lines
tools/perf/builtin-trace.c +6
-0.max_events = ULONG_MAX, + .ev_qualifier = NULL, + .sctbl = NULL, + .ev_qualifier_ids = { + .entries = NULL, + .nr = 0, + }, };const char *map_dump_str =NULL ;const char *output_name =NULL ;
The above hunk is not needed, as all non explicitely initialized members will be zeroed, so to reduce patch size I'm dropping it.
Thanks, applied.
- Arnaldo
Quoted 11 lines
tools/perf/builtin-trace.c +1
-1@@ -5135,6 +5150,6 @@ int cmd_trace(int argc, const char **argv)if (output_name !=NULL )fclose (trace.output); out: -zfree (&trace.perfconfig_events); +trace__exit (&trace);return err; }Signature · 1 line
2.31.1
Signature · 1 line
- Arnaldo
PATCH 16/20 perf trace: free syscall->arg_fmt
In reply to Riccardo ManciniASan reports several memory leaks running the perf test
"88: Check open filename arg using perf trace + vfs_getname". The second of these leaks is caused by the arg_fmt field of syscall not being deallocated.
This patch adds a new function syscall__exit which is called on all syscalls.table entries in trace__exit, which will free the arg_fmt field.
--- tools/perf/builtin-trace.c | 16 +++++++++++++++- 1 file changed, 15 insertions(+), 1 deletion(-)
tools/perf/builtin-trace.c +15 -1
@@ -2266,6 +2266,14 @@ static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *samreturn augmented_args; } +static void syscall__exit (struct syscall *sc) +{ +if (!sc) +return ; + +free (sc->arg_fmt); +} +static int trace__sys_enter (struct trace *trace,struct evsel *evsel,union perf_event *event__maybe_unused ,struct perf_sample *sample)
@@ -4703,9 +4711,15 @@ static int trace__config(const char *var, const char *value, void *arg)static void trace__exit (struct trace *trace) { +int i; +strlist__delete (trace->ev_qualifier);free (trace->ev_qualifier_ids.entries); -free (trace->syscalls.table); +if (trace->syscalls.table) { +for (i =0 ; i <= trace->sctbl->syscalls.max_id; i++) +syscall__exit (&trace->syscalls.table[i]); +free (trace->syscalls.table); + }syscalltbl__delete (trace->sctbl);zfree (&trace->perfconfig_events); }
Signature · 1 line
2.31.1
Re: [PATCH 16/20] perf trace: free syscall->arg_fmt
Em Thu, Jul 15, 2021 at 06:07:21PM +0200, Riccardo Mancini escreveu:
Quoted 8 lines
ASan reports several memory leaks running the perf test
"88: Check open filename arg using perf trace +vfs_getname". The second of these leaks is caused by thearg_fmtfield of syscall not being deallocated.This patch adds a new function
syscall__exitwhich is called on all syscalls.table entries intrace__exit, which will free thearg_fmtfield.
Thanks, applied.
- Arnaldo
Quoted 44 lines
Signed-off-by Riccardo Mancini <[email protected]>--- tools/perf/builtin-trace.c | 16 +++++++++++++++- 1 file changed, 15 insertions(+), 1 deletion(-)tools/perf/builtin-trace.c +15
-1@@ -2266,6 +2266,14 @@ static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *samreturn augmented_args; } +static void syscall__exit (struct syscall *sc) +{ +if (!sc) +return ; + +free (sc->arg_fmt); +} +static int trace__sys_enter (struct trace *trace,struct evsel *evsel,union perf_event *event__maybe_unused ,struct perf_sample *sample)@@ -4703,9 +4711,15 @@ static int trace__config(const char *var, const char *value, void *arg)static void trace__exit (struct trace *trace) { +int i; +strlist__delete (trace->ev_qualifier);free (trace->ev_qualifier_ids.entries); -free (trace->syscalls.table); +if (trace->syscalls.table) { +for (i =0 ; i <= trace->sctbl->syscalls.max_id; i++) +syscall__exit (&trace->syscalls.table[i]); +free (trace->syscalls.table); + }syscalltbl__delete (trace->sctbl);zfree (&trace->perfconfig_events); }Signature · 1 line
2.31.1
Signature · 1 line
- Arnaldo
PATCH 17/20 perf trace: free syscall tp fields in evsel->priv
In reply to Riccardo ManciniASan reports several memory leaks running the perf test
"88: Check open filename arg using perf trace + vfs_getname". The third of these leaks is related to evsel->priv fields of sycalls never being deallocated.
This patch adds the function evlist__free_syscall_tp_fields which iterates over all evsels in evlist, matching syscalls, and calling the missing frees.
This new function is called at the end of trace__run, right before calling evlist__delete.
--- tools/perf/builtin-trace.c | 17 ++++++++++++++++- 1 file changed, 16 insertions(+), 1 deletion(-)
tools/perf/builtin-trace.c +16 -1
@@ -3103,6 +3103,21 @@ static struct evsel *evsel__new_pgfault(u64 config)return evsel; } +static void evlist__free_syscall_tp_fields (struct evlist *evlist) +{ +struct evsel *evsel; + +evlist__for_each_entry (evlist, evsel) { +struct evsel_trace *et = evsel->priv; + +if (!et || !evsel->tp_format ||strcmp (evsel->tp_format->system,"syscalls" )) +continue ; + +free (et->fmt); +free (et); + } +} +static void trace__handle_event (struct trace *trace,union perf_event *event,struct perf_sample *sample) {const u32 type = event->header.type;
@@ -4138,7 +4153,7 @@ static int trace__run(struct trace *trace, int argc, const char **argv) out_delete_evlist:trace__symbols__exit (trace); - +evlist__free_syscall_tp_fields (evlist);evlist__delete (evlist);cgroup__put (trace->cgroup); trace->evlist =NULL ;
Signature · 1 line
2.31.1
Re: [PATCH 17/20] perf trace: free syscall tp fields in evsel->priv
Em Thu, Jul 15, 2021 at 06:07:22PM +0200, Riccardo Mancini escreveu:
Quoted 10 lines
ASan reports several memory leaks running the perf test
"88: Check open filename arg using perf trace +vfs_getname". The third of these leaks is related toevsel->privfields of sycalls never being deallocated.This patch adds the function
evlist__free_syscall_tp_fieldswhich iterates over all evsels in evlist, matching syscalls, and calling the missing frees.
This new function is called at the end oftrace__run, right before callingevlist__delete.
Thanks, applied.
- Arnaldo
Quoted 43 lines
Signed-off-by Riccardo Mancini <[email protected]>--- tools/perf/builtin-trace.c | 17 ++++++++++++++++- 1 file changed, 16 insertions(+), 1 deletion(-)tools/perf/builtin-trace.c +16
-1@@ -3103,6 +3103,21 @@ static struct evsel *evsel__new_pgfault(u64 config)return evsel; } +static void evlist__free_syscall_tp_fields (struct evlist *evlist) +{ +struct evsel *evsel; + +evlist__for_each_entry (evlist, evsel) { +struct evsel_trace *et = evsel->priv; + +if (!et || !evsel->tp_format ||strcmp (evsel->tp_format->system,"syscalls" )) +continue ; + +free (et->fmt); +free (et); + } +} +static void trace__handle_event (struct trace *trace,union perf_event *event,struct perf_sample *sample) {const u32 type = event->header.type;@@ -4138,7 +4153,7 @@ static int trace__run(struct trace *trace, int argc, const char **argv) out_delete_evlist:trace__symbols__exit (trace); - +evlist__free_syscall_tp_fields (evlist);evlist__delete (evlist);cgroup__put (trace->cgroup); trace->evlist =NULL ;Signature · 1 line
2.31.1
Signature · 1 line
- Arnaldo
PATCH 18/20 perf trace: free strings in trace__parse_events_option
In reply to Riccardo ManciniASan reports several memory leaks running the perf test "88: Check open filename arg using perf trace + vfs_getname". The fourth of these leaks is related to some strings never being freed in trace__parse_events_option.
This patch adds the missing frees.
--- tools/perf/builtin-trace.c | 3 +++ 1 file changed, 3 insertions(+)
tools/perf/builtin-trace.c +3 -0
@@ -4659,6 +4659,9 @@ static int trace__parse_events_option(const struct option *opt, const char *str, err =parse_events_option (&o, lists[0 ],0 ); } out: +free (strace_groups_dir); +free (lists[0 ]); +free (lists[1 ]);if (sep) *sep =',' ;
Signature · 1 line
2.31.1
Re: [PATCH 18/20] perf trace: free strings in trace__parse_events_option
Em Thu, Jul 15, 2021 at 06:07:23PM +0200, Riccardo Mancini escreveu:
Quoted 6 lines
ASan reports several memory leaks running the perf test "88: Check open filename arg using perf trace +
vfs_getname". The fourth of these leaks is related to some strings never being freed intrace__parse_events_option.This patch adds the missing frees.
Thanks, applied.
- Arnaldo
Quoted 22 lines
Signed-off-by Riccardo Mancini <[email protected]>--- tools/perf/builtin-trace.c | 3 +++ 1 file changed, 3 insertions(+)tools/perf/builtin-trace.c +3
-0@@ -4659,6 +4659,9 @@ static int trace__parse_events_option(const struct option *opt, const char *str, err =parse_events_option (&o, lists[0 ],0 ); } out: +free (strace_groups_dir); +free (lists[0 ]); +free (lists[1 ]);if (sep) *sep =',' ;Signature · 1 line
2.31.1
Signature · 1 line
- Arnaldo
PATCH 19/20 perf test: bpf: free obj_buf
In reply to Riccardo ManciniASan reports some memory leaks when running the perf test "42: BPF filter".
The first of these leaks is caused by obj_buf never being deallocated in __test__bpf.
This patch adds the missing free.
--- tools/perf/tests/bpf.c | 1 + 1 file changed, 1 insertion(+)
tools/perf/tests/bpf.c +1 -0
@@ -276,6 +276,7 @@ static int __test__bpf(int idx) } out: +free (obj_buf);bpf__clear ();return ret; }
Signature · 1 line
2.31.1
Re: [PATCH 19/20] perf test: bpf: free obj_buf
Em Thu, Jul 15, 2021 at 06:07:24PM +0200, Riccardo Mancini escreveu:
Quoted 6 lines
ASan reports some memory leaks when running the perf test "42: BPF filter".
The first of these leaks is caused byobj_bufnever being deallocated in__test__bpf.This patch adds the missing free.
Thanks, applied.
- Arnaldo
Quoted 20 lines
Signed-off-by Riccardo Mancini <[email protected]>--- tools/perf/tests/bpf.c | 1 + 1 file changed, 1 insertion(+)tools/perf/tests/bpf.c +1
-0@@ -276,6 +276,7 @@ static int __test__bpf(int idx) } out: +free (obj_buf);bpf__clear ();return ret; }Signature · 1 line
2.31.1
Signature · 1 line
- Arnaldo
Arnaldo Carvalho de Melo
maintainerRe: [PATCH 19/20] perf test: bpf: free obj_buf
In reply to Riccardo ManciniEm Thu, Jul 15, 2021 at 06:07:24PM +0200, Riccardo Mancini escreveu:
Quoted 18 lines
ASan reports some memory leaks when running the perf test "42: BPF filter".
The first of these leaks is caused byobj_bufnever being deallocated in__test__bpf.This patch adds the missing free.
Signed-off-by Riccardo Mancini <[email protected]>--- tools/perf/tests/bpf.c | 1 + 1 file changed, 1 insertion(+)tools/perf/tests/bpf.c +0
-0
tools/perf/tests/bpf.c static int __test__bpf(int idx) · line 282
Quoted 7 lines
tools/perf/tests/bpf.c +1
-0@@ -276,6 +276,7 @@ static int __test__bpf(int idx) } out: +free (obj_buf);bpf__clear ();return ret; }
I followed the advice and added the stdlib.h include, elsewhere we're getting included via some other include by luck.
- Arnaldo
66 6.97 ubuntu:18.04-x-m68k : FAIL gcc version 7.5.0 (Ubuntu 7.5.0-3ubuntu1~18.04)
tests/bpf.c: In function '__test__bpf':
tests/bpf.c:279:2: error: implicit declaration of function 'free' [-Werror=implicit-function-declaration]
free(obj_buf);
^~~~
tests/bpf.c:279:2: error: incompatible implicit declaration of built-in function 'free' [-Werror]
tests/bpf.c:279:2: note: include '<stdlib.h>' or provide a declaration of 'free'
cc1: all warnings being treated as errors
/git/perf-5.14.0-rc1/tools/build/Makefile.build:139: recipe for target 'tests' failed
make[3]: *** [tests] Error 2PATCH 20/20 perf util/probe-file: delete namelist on error in del_events
In reply to Riccardo ManciniASan reports some memory leaks when running the perf test "42: BPF filter".
This second leak is caused by a strlist not being dellocated on error inside probe_file__del_events.
This patch adds a goto label before the deallocation and makes the error path jump to it.
--- tools/perf/util/probe-file.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
tools/perf/util/probe-file.c +2 -2
@@ -377,11 +377,11 @@ int probe_file__del_events(int fd, struct strfilter *filter) ret =probe_file__get_events (fd, filter, namelist);if (ret <0 ) -return ret; + goto out; ret =probe_file__del_strlist (fd, namelist); +out:strlist__delete (namelist); -return ret; }
Signature · 1 line
2.31.1
Re: [PATCH 20/20] perf util/probe-file: delete namelist on error in del_events
Em Thu, Jul 15, 2021 at 06:07:25PM +0200, Riccardo Mancini escreveu:
Quoted 7 lines
ASan reports some memory leaks when running the perf test "42: BPF filter".
This second leak is caused by a strlist not being dellocated on error insideprobe_file__del_events.This patch adds a goto label before the deallocation and makes the error path jump to it.
Thanks, applied.
- Arnaldo
Quoted 26 lines
Signed-off-by Riccardo Mancini <[email protected]>--- tools/perf/util/probe-file.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)tools/perf/util/probe-file.c +2
-2@@ -377,11 +377,11 @@ int probe_file__del_events(int fd, struct strfilter *filter) ret =probe_file__get_events (fd, filter, namelist);if (ret <0 ) -return ret; + goto out; ret =probe_file__del_strlist (fd, namelist); +out:strlist__delete (namelist); -return ret; }Signature · 1 line
2.31.1
Signature · 1 line
- Arnaldo