Re: [PATCH 5.3 111/344] perf unwind: Fix libunwind when tid != pid

From: Andres Freund
Date: Sat Nov 09 2019 - 20:53:41 EST


Hi,

On 2019-10-03 17:51:16 +0200, Greg Kroah-Hartman wrote:
> From: John Keeping <john@xxxxxxxxxxxx>
>
> [ Upstream commit e8ba2906f6b9054102ad035ac9cafad9d4168589 ]
>
> Commit e5adfc3e7e77 ("perf map: Synthesize maps only for thread group
> leader") changed the recording side so that we no longer get mmap events
> for threads other than the thread group leader (when synthesising these
> events for threads which exist before perf is started).
>
> When a file recorded after this change is loaded, the lack of mmap
> records mean that unwinding is not set up for any other threads.
>
> This can be seen in a simple record/report scenario:
>
> perf record --call-graph=dwarf -t $TID
> perf report
>
> If $TID is a process ID then the report will show call graphs, but if
> $TID is a secondary thread the output is as if --call-graph=none was
> specified.
>
> Following the rationale in that commit, move the libunwind fields into
> struct map_groups and update the libunwind functions to take this
> instead of the struct thread. This is only required for
> unwind__finish_access which must now be called from map_groups__delete
> and the others are changed for symmetry.
>
> Note that unwind__get_entries keeps the thread argument since it is
> required for symbol lookup and the libdw unwind provider uses the thread
> ID.
>
> Signed-off-by: John Keeping <john@xxxxxxxxxxxx>
> Reviewed-by: Jiri Olsa <jolsa@xxxxxxxxxx>
> Cc: Alexander Shishkin <alexander.shishkin@xxxxxxxxxxxxxxx>
> Cc: Konstantin Khlebnikov <khlebnikov@xxxxxxxxxxxxxx>
> Cc: Namhyung Kim <namhyung@xxxxxxxxxx>
> Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> Fixes: e5adfc3e7e77 ("perf map: Synthesize maps only for thread group leader")
> Link: http://lkml.kernel.org/r/20190815100146.28842-2-john@xxxxxxxxxxxx
> Signed-off-by: Arnaldo Carvalho de Melo <acme@xxxxxxxxxx>
> Signed-off-by: Sasha Levin <sashal@xxxxxxxxxx>

This unfortunately broke --call-graph dwarf on 5.3 (and presumably older
branches), because while this commit has been included in stable, the
prerequisite

commit ab6cd0e5276e24403751e0b3b8ed807738a8571f
Author: John Keeping <john@xxxxxxxxxxxx>
AuthorDate: 2019-08-15 11:01:44 +0100
Commit: Arnaldo Carvalho de Melo <acme@xxxxxxxxxx>
CommitDate: 2019-08-16 12:25:23 -0300

perf map: Use zalloc for map_groups

In the next commit we will add new fields to map_groups and we need
these to be null if no value is assigned. The simplest way to achieve
this is to request zeroed memory from the allocator.

Signed-off-by: John Keeping <john@xxxxxxxxxxxx>
Reviewed-by: Jiri Olsa <jolsa@xxxxxxxxxx>
Cc: Alexander Shishkin <alexander.shishkin@xxxxxxxxxxxxxxx>
Cc: Konstantin Khlebnikov <khlebnikov@xxxxxxxxxxxxxx>
Cc: Namhyung Kim <namhyung@xxxxxxxxxx>
Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
Cc: john keeping <john@xxxxxxxxxxxx>
Link: http://lkml.kernel.org/r/20190815100146.28842-1-john@xxxxxxxxxxxx
Signed-off-by: Arnaldo Carvalho de Melo <acme@xxxxxxxxxx>

has not.


The crash I get is:

Thread 1 "perf" received signal SIGSEGV, Segmentation fault.
0x0000555555872238 in unwind__flush_access (mg=0x555555c53b50) at util/unwind-libunwind.c:76
76 mg->unwind_libunwind_ops->flush_access(mg);
(gdb) bt
#0 0x0000555555872238 in unwind__flush_access (mg=0x555555c53b50) at util/unwind-libunwind.c:76
#1 0x0000555555800ac4 in ____thread__set_comm (exec=true, timestamp=325096707055731, str=0x7ffff7f96ed8 "sleep", thread=0x555555c53bc0) at util/thread.c:254
#2 __thread__set_comm (thread=thread@entry=0x555555c53bc0, str=str@entry=0x7ffff7f96ed8 "sleep", timestamp=325096707055731, exec=exec@entry=true)
at util/thread.c:268
#3 0x00005555557f132a in machine__process_comm_event (machine=0x555555c4bc68, event=0x7ffff7f96ec8, sample=0x7fffffff8f70) at util/machine.c:600
#4 0x00005555557fa93b in perf_session__deliver_event (session=0x555555c4baf0, event=0x7ffff7f96ec8, tool=0x555555acb9a0 <record>, file_offset=73416)
at util/session.c:1473
#5 0x00005555557feae8 in do_flush (show_progress=true, oe=0x555555c52610) at util/ordered-events.c:243
#6 __ordered_events__flush (oe=oe@entry=0x555555c52610, how=how@entry=OE_FLUSH__FINAL, timestamp=timestamp@entry=0) at util/ordered-events.c:322
#7 0x00005555557fef45 in __ordered_events__flush (timestamp=<optimized out>, how=<optimized out>, oe=<optimized out>) at util/ordered-events.c:338
#8 ordered_events__flush (how=OE_FLUSH__FINAL, oe=0x555555c52610) at util/ordered-events.c:340
#9 ordered_events__flush (oe=oe@entry=0x555555c52610, how=how@entry=OE_FLUSH__FINAL) at util/ordered-events.c:338
#10 0x00005555557fd17c in __perf_session__process_events (session=0x555555c4baf0) at util/session.c:2152
#11 perf_session__process_events (session=session@entry=0x555555c4baf0) at util/session.c:2181
#12 0x0000555555729379 in process_buildids (rec=0x555555acb9a0 <record>) at builtin-record.c:829
#13 record__finish_output (rec=0x555555acb9a0 <record>) at builtin-record.c:1037
#14 0x000055555572c000 in __cmd_record (rec=0x555555acb9a0 <record>, argv=<optimized out>, argc=2) at builtin-record.c:1661
#15 cmd_record (argc=2, argv=<optimized out>) at builtin-record.c:2450
#16 0x000055555579cd9d in run_builtin (p=0x555555ad4958 <commands+216>, argc=5, argv=0x7fffffffdcc0) at perf.c:304
#17 0x0000555555714baa in handle_internal_command (argv=0x7fffffffdcc0, argc=5) at perf.c:356
#18 run_argv (argcp=<synthetic pointer>, argv=<synthetic pointer>) at perf.c:400
#19 main (argc=5, argv=0x7fffffffdcc0) at perf.c:525

(gdb) p *mg
$7 = {maps = {entries = {rb_node = 0x0}, names = {rb_node = 0x0}, lock = {lock = pthread_rwlock_t = {Status = Not acquired, Shared = No,
Prefers = Readers}}}, machine = 0x555555c4bc68, refcnt = {refs = {counter = 1}}, addr_space = 0x693f6967632e6775,
unwind_libunwind_ops = 0xa32313438313d64}

(gdb) p mg->unwind_libunwind_ops
$8 = (struct unwind_libunwind_ops *) 0xa32313438313d64

(gdb) p *mg->unwind_libunwind_ops
Cannot access memory at address 0xa32313438313d64

which makes sense, because map_groups__new() allocates the group with
malloc, and map_groups__init() only initializes map_groups->{maps,machine,refcnt}


A bit surprised that nobody complained about this so far...


Greetings,

Andres Freund