lkml.org 
[lkml]   [2013]   [Jan]   [23]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: [RFC/PATCH] scripts/tracing: Add trace_analyze.py tool
From
Hi Minchan,

On Wed, Jan 23, 2013 at 1:27 AM, Minchan Kim <minchan@kernel.org> wrote:
> Hi Ezequiel,
>
> On Tue, Jan 22, 2013 at 06:46:58AM -0300, Ezequiel Garcia wrote:
>> From: Ezequiel Garcia <elezegarcia@gmail.com>
>>
>> The purpose of trace_analyze.py tool is to perform static
>> and dynamic memory analysis using a kmem ftrace
>> log file and a built kernel tree.
>>
>> This script and related work has been done on the CEWG/2012 project:
>> "Kernel dynamic memory allocation tracking and reduction"
>> (More info here [1])
>>
>> It produces mainly two kinds of outputs:
>> * an account-like output, similar to the one given by Perf, example below.
>> * a ring-char output, examples here [2].
>>
>> $ ./scripts/tracing/trace_analyze.py -k linux -f kmem.log --account-file account.txt
>> $ ./scripts/tracing/trace_analyze.py -k linux -f kmem.log -c account.txt
>>
>> This will produce an account file like this:
>>
>> current bytes allocated: 669696
>> current bytes requested: 618823
>> current wasted bytes: 50873
>> number of allocs: 7649
>> number of frees: 2563
>> number of callers: 115
>>
>> total waste net alloc/free caller
>> ---------------------------------------------
>> 299200 0 298928 1100/1 alloc_inode+0x4fL
>> 189824 0 140544 1483/385 __d_alloc+0x22L
>> 51904 0 47552 811/68 sysfs_new_dirent+0x4eL
>> [...]
>>
>> [1] http://elinux.org/Kernel_dynamic_memory_analysis
>> [2] http://elinux.org/Kernel_dynamic_memory_analysis#Current_dynamic_footprint
>
> First of all, Thanks for nice work! It could be very useful for
> embedded side.
>
> Questions.
>
> 1. Can we detect different call path but same function?
> I mean
>
> A C
> \ /
> B D
> \ /
> E
> |
> kmalloc
>
> In this case, E could be called by A or C. I would like to know the call path.
> It could point out exact culprit of memory hogger.
>

I'm sorry, I'm not following you:
How can I know which caller in the call path is the 'real' responsible
for the allocation?

The only way I can think of achieving something like this is by using
kmalloc_track_caller() instead of kmalloc().
This is done in cases where an allocer is known to alloc memory on
behalf of its caller.

> 2. Does it support alloc_pages family?
> kmem event trace already supports it. If it supports, maybe we can replace
> CONFIG_PAGE_OWNER hack.
>

Mmm.. no, it doesn't support alloc_pages and friends, for we found
no reason to do it.
However, it sounds like a nice idea, on a first thought.

I'll review CONFIG_PAGE_OWNER patches and see if I can come up with something.

Meantime, and given this is just a script submission, is there anything
preventing to merge this? We can move it to perf, and/or add it
features, etc. later,
on top of this. Does this make sense?

--
Ezequiel


\
 
 \ /
  Last update: 2013-01-23 23:21    [W:0.098 / U:0.040 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site