Re: [PATCH 0/3] mm: cma: /proc/cmainfo

[Minchan Kim]

On Mon, Dec 29, 2014 at 11:52:58AM -0800, Laura Abbott wrote:
> On 12/28/2014 6:36 PM, Minchan Kim wrote:
> >Hello,
> >
> >On Fri, Dec 26, 2014 at 05:39:01PM +0300, Stefan I. Strogin wrote:
> >>Hello all,
> >>
> >>Here is a patch set that adds /proc/cmainfo.
> >>
> >>When compiled with CONFIG_CMA_DEBUG /proc/cmainfo will contain information
> >>about about total, used, maximum free contiguous chunk and all currently
> >>allocated contiguous buffers in CMA regions. The information about allocated
> >>CMA buffers includes pid, comm, allocation latency and stacktrace at the
> >>moment of allocation.
> >
> >It just says what you are doing but you didn't say why we need it.
> >I can guess but clear description(ie, the problem what you want to
> >solve with this patchset) would help others to review, for instance,
> >why we need latency, why we need callstack, why we need new wheel
> >rather than ftrace and so on.
> >
> >Thanks.
> >
> 
> 
> I've been meaning to write something like this for a while so I'm
> happy to see an attempt made to fix this. I can't speak for the
> author's reasons for wanting this information but there are
> several reasons why I was thinking of something similar.
> 
> The most common bug reports seen internally on CMA are 1) CMA is
> too slow and 2) CMA failed to allocate memory. For #1, not all
> allocations may be slow so it's useful to be able to keep track
> of which allocations are taking too long. For #2, migration

Then, I don't think we could keep all of allocations. What we need
is only slow allocations. I hope we can do that with ftrace.

ex)

# cd /sys/kernel/debug/tracing
# echo 1 > options/stacktrace
# echo cam_alloc > set_ftrace_filter
# echo your_threshold > tracing_thresh

I know it doesn't work now but I think it's more flexible
and general way to handle such issues(ie, latency of some functions).
So, I hope we could enhance ftrace rather than new wheel.
Ccing ftrace people.

Futhermore, if we really need to have such information, we need more data
(ex, how many of pages were migrated out, how many pages were dropped
without migrated, how many pages were written back, how many pages were
retried with the page lock and so on).
In this case, event trace would be better.


> failure is fairly common but it's still important to rule out
> a memory leak from a dma client. Seeing all the allocations is
> also very useful for memory tuning (e.g. how big does the CMA
> region need to be, which clients are actually allocating memory).

Memory leak is really general problem and could we handle it with
page_owner?

> 
> ftrace is certainly usable for tracing CMA allocation callers and
> latency. ftrace is still only a fixed size buffer though so it's
> possible for information to be lost if other logging is enabled.

Sorry, I don't get with only above reasons why we need this. :(

> For most of the CMA use cases, there is a very high cost if the
> proper debugging information is not available so the more that
> can be guaranteed the better.
> 
> It's also worth noting that the SLUB allocator has a sysfs
> interface for showing allocation callers when CONFIG_SLUB_DEBUG
> is enabled.
> 
> Thanks,
> Laura
> 
> -- 
> Qualcomm Innovation Center, Inc.
> Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
> a Linux Foundation Collaborative Project
> 
> --
> To unsubscribe, send a message with 'unsubscribe linux-mm' in
> the body to majordomo@xxxxxxxxxx  For more info on Linux MM,
> see: http://www.linux-mm.org/ .
> Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>

-- 
Kind regards,
Minchan Kim
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/