lkml.org 
[lkml]   [2015]   [Nov]   [18]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRE: [PATCH perf/core 00/13] perf memory/refcnt leak fixes
Date
From: Arnaldo Carvalho de Melo [mailto:acme@kernel.org]
>
>Em Wed, Nov 18, 2015 at 03:40:09PM +0900, Masami Hiramatsu escreveu:
>> Hi,
>>
>> Here is a series to fix some memory leaks and refcount
>> leaks on map and dso. This also includes the refcnt APIs
>> with backtrace debugging feature.
>
>Cool, I wonder if this could be usable in the kernel proper... Is there
>such a facility there? I'll check.

As far as I know, there is no same feature, but I guess expanding
kmemleak is possible to provide similar feature.

>But thanks for doing this work, I'll go thru the fixes first, then look
>at the debugging feature.

Thanks!

>
>- Arnaldo
>
>> The story has started from the posible memory leak report
>> reported by Wnag Nan.
>> I've tried to use valgrind to ensure the perf probe doesn't
>> have other memory leaks. The result is here:
>>
>> ----
>> # valgrind ./perf probe vfs_read
>> ==17521== Memcheck, a memory error detector
>> ==17521== Copyright (C) 2002-2013, and GNU GPL'd, by Julian Seward et al.
>> ==17521== Using Valgrind-3.10.0 and LibVEX; rerun with -h for copyright info
>> ==17521== Command: ./perf probe vfs_read
>> ==17521==
>> Added new event:
>> probe:vfs_read (on vfs_read)
>>
>> You can now use it in all perf tools, such as:
>>
>> perf record -e probe:vfs_read -aR sleep 1
>>
>> ==17521==
>> ==17521== HEAP SUMMARY:
>> ==17521== in use at exit: 3,512,761 bytes in 38,012 blocks
>> ==17521== total heap usage: 74,723 allocs, 36,711 frees, 24,014,927
>> bytes allocated
>> ==17521==
>> ==17521== LEAK SUMMARY:
>> ==17521== definitely lost: 6,857 bytes in 49 blocks
>> ==17521== indirectly lost: 3,501,287 bytes in 37,891 blocks
>> ==17521== possibly lost: 0 bytes in 0 blocks
>> ==17521== still reachable: 4,617 bytes in 72 blocks
>> ==17521== suppressed: 0 bytes in 0 blocks
>> ==17521== Rerun with --leak-check=full to see details of leaked memory
>> ==17521==
>> ==17521== For counts of detected and suppressed errors, rerun with: -v
>> ==17521== ERROR SUMMARY: 0 errors from 0 contexts (suppressed: 2 from 2)
>> ----
>>
>> Oops! It leaked almost 4 MB memories. I've tried to find
>> the root causes, and what I've found is there are many
>> leaks in not only perf-probe specific code, but also maps
>> and dsos (and some other pieces).
>>
>> The first 3 patches are just fixing 'easy' memory leaks. However,
>> most of the leaks are caused by refcnt. Since valgrind seems not
>> able to debug this kind of issues, I introduced a hand-made refcnt
>> backtrace APIs for debugging.
>> The rest of patches are for fixing refcnt leak bugs and replcing
>> refcnt apis.
>>
>> After all, most of the issues are gone, except for just a few issues
>> in elfutils. I'll continue to investigate that.
>>
>> ----
>> valgrind ./perf probe vfs_read
>> ==29521== Memcheck, a memory error detector
>> ==29521== Copyright (C) 2002-2013, and GNU GPL'd, by Julian Seward et al.
>> ==29521== Using Valgrind-3.10.0 and LibVEX; rerun with -h for copyright info
>> ==29521== Command: ./perf probe vfs_read
>> ==29521==
>> Added new event:
>> probe:vfs_read (on vfs_read)
>>
>> You can now use it in all perf tools, such as:
>>
>> perf record -e probe:vfs_read -aR sleep 1
>>
>> ==29521==
>> ==29521== HEAP SUMMARY:
>> ==29521== in use at exit: 5,137 bytes in 75 blocks
>> ==29521== total heap usage: 74,723 allocs, 74,648 frees, 24,014,927
>> bytes allocated
>> ==29521==
>> ==29521== LEAK SUMMARY:
>> ==29521== definitely lost: 520 bytes in 3 blocks
>> ==29521== indirectly lost: 0 bytes in 0 blocks
>> ==29521== possibly lost: 0 bytes in 0 blocks
>> ==29521== still reachable: 4,617 bytes in 72 blocks
>> ==29521== suppressed: 0 bytes in 0 blocks
>> ==29521== Rerun with --leak-check=full to see details of leaked memory
>> ==29521==
>> ==29521== For counts of detected and suppressed errors, rerun with: -v
>> ==29521== ERROR SUMMARY: 0 errors from 0 contexts (suppressed: 2 from 2)
>> ----
>>
>> Anyway, I decided to release these fixes and the debugging feature
>> because at least this will improve perf quality.
>>
>> Thank you,
>>
>> ---
>>
>> Masami Hiramatsu (13):
>> perf probe: Fix to free temporal Dwarf_Frame
>> perf: Make perf_exec_path always returns malloc'd string
>> perf: Introduce generic refcount APIs with debug feature
>> perf: make map to use refcnt
>> perf: Fix machine__findnew_module_map to put registered map
>> perf: Fix machine__destroy_kernel_maps to put vmlinux_maps
>> perf: Fix to destroy kernel maps when machine exits
>> perf: Fix to put new map after inserting to map_groups in dso__load_sym
>> perf: Make dso to use refcnt for debug
>> perf: Fix __dsos__addnew to put dso after adding it to the list
>> perf: Fix machine__create_kernel_maps to put kernel dso
>> perf: Fix machine__findnew_module_map to put dso
>> perf: Fix dso__load_sym to put dso
>>
>>
>> tools/perf/config/Makefile | 5 ++
>> tools/perf/util/Build | 1
>> tools/perf/util/dso.c | 9 ++-
>> tools/perf/util/exec_cmd.c | 20 ++++--
>> tools/perf/util/exec_cmd.h | 5 +-
>> tools/perf/util/help.c | 6 +-
>> tools/perf/util/machine.c | 17 ++++-
>> tools/perf/util/map.c | 7 +-
>> tools/perf/util/map.h | 3 +
>> tools/perf/util/probe-finder.c | 9 ++-
>> tools/perf/util/refcnt.c | 125 ++++++++++++++++++++++++++++++++++++++++
>> tools/perf/util/refcnt.h | 65 +++++++++++++++++++++
>> tools/perf/util/symbol-elf.c | 4 +
>> 13 files changed, 250 insertions(+), 26 deletions(-)
>> create mode 100644 tools/perf/util/refcnt.c
>> create mode 100644 tools/perf/util/refcnt.h
>>
>>
>> --
>> Masami HIRAMATSU
>> Linux Technology Research Center, System Productivity Research Dept.
>> Center for Technology Innovation - Systems Engineering
>> Hitachi, Ltd., Research & Development Group
>> E-mail: masami.hiramatsu.pt@hitachi.com


\
 
 \ /
  Last update: 2015-11-19 04:21    [W:0.191 / U:1.236 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site