Re: [PATCH v3 0/4] enhance shmem process and swap accounting

[Jerome Marchand <jmarchan-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>]

[-- Attachment #1: Type: text/plain, Size: 5917 bytes --]

On 08/05/2015 03:01 PM, Vlastimil Babka wrote:
> Reposting due to lack of feedback in May. I hope at least patches 1 and 2
> could be merged as they are IMHO bugfixes. 3 and 4 is optional but IMHO useful.
> 
> Changes since v2:
> o Rebase on next-20150805.
> o This means that /proc/pid/maps has the proportional swap share (SwapPss:)
>   field as per https://lkml.org/lkml/2015/6/15/274
>   It's not clear what to do with shmem here so it's 0 for now.
>   - swapped out shmem doesn't have swap entries, so we would have to look at who
>     else has the shmem object (partially) mapped
>   - to be more precise we should also check if his range actually includes 
>     the offset in question, which could get rather involved
>   - or is there some easy way I don't see?

Hmm... This is much more difficult than I envision when commenting on
Minchan patch. One possibility could be to have the pte of paged out
shmem pages set in a similar way than regular swap entry are. But that
would need to use some very precious estate on the pte.
As it is, a zero value, while obviously wrong, has the advantage of not
being misleading like a bad approximation would be (like the kind which
doesn't properly accounts for partial mapping).

Jerome

> o Konstantin suggested for patch 3/4 that I drop the CONFIG_SHMEM #ifdefs
>   I didn't see the point in going against tinyfication when the work is
>   already done, but I can do that if more people think it's better and it
>   would block the series.
> 
> Changes since v1:
> o In Patch 2, rely on SHMEM_I(inode)->swapped if possible, and fallback to
>   radix tree iterator on partially mapped shmem objects, i.e. decouple shmem
>   swap usage determination from the page walk, for performance reasons.
>   Thanks to Jerome and Konstantin for the tips.
>   The downside is that mm/shmem.c had to be touched.
> 
> This series is based on Jerome Marchand's [1] so let me quote the first
> paragraph from there:
> 
> There are several shortcomings with the accounting of shared memory
> (sysV shm, shared anonymous mapping, mapping to a tmpfs file). The
> values in /proc/<pid>/status and statm don't allow to distinguish
> between shmem memory and a shared mapping to a regular file, even
> though theirs implication on memory usage are quite different: at
> reclaim, file mapping can be dropped or write back on disk while shmem
> needs a place in swap. As for shmem pages that are swapped-out or in
> swap cache, they aren't accounted at all.
> 
> The original motivation for myself is that a customer found (IMHO rightfully)
> confusing that e.g. top output for process swap usage is unreliable with
> respect to swapped out shmem pages, which are not accounted for.
> 
> The fundamental difference between private anonymous and shmem pages is that
> the latter has PTE's converted to pte_none, and not swapents. As such, they are
> not accounted to the number of swapents visible e.g. in /proc/pid/status VmSwap
> row. It might be theoretically possible to use swapents when swapping out shmem
> (without extra cost, as one has to change all mappers anyway), and on swap in
> only convert the swapent for the faulting process, leaving swapents in other
> processes until they also fault (so again no extra cost). But I don't know how
> many assumptions this would break, and it would be too disruptive change for a
> relatively small benefit.
> 
> Instead, my approach is to document the limitation of VmSwap, and provide means
> to determine the swap usage for shmem areas for those who are interested and
> willing to pay the price, using /proc/pid/smaps. Because outside of ipcs, I
> don't think it's possible to currently to determine the usage at all.  The
> previous patchset [1] did introduce new shmem-specific fields into smaps
> output, and functions to determine the values. I take a simpler approach,
> noting that smaps output already has a "Swap: X kB" line, where currently X ==
> 0 always for shmem areas. I think we can just consider this a bug and provide
> the proper value by consulting the radix tree, as e.g. mincore_page() does. In the
> patch changelog I explain why this is also not perfect (and cannot be without
> swapents), but still arguably much better than showing a 0.
> 
> The last two patches are adapted from Jerome's patchset and provide a VmRSS
> breakdown to VmAnon, VmFile and VmShm in /proc/pid/status. Hugh noted that
> this is a welcome addition, and I agree that it might help e.g. debugging
> process memory usage at albeit non-zero, but still rather low cost of extra
> per-mm counter and some page flag checks. I updated these patches to 4.0-rc1,
> made them respect !CONFIG_SHMEM so that tiny systems don't pay the cost, and
> optimized the page flag checking somewhat.
> 
> [1] http://lwn.net/Articles/611966/
> 
> Jerome Marchand (2):
>   mm, shmem: Add shmem resident memory accounting
>   mm, procfs: Display VmAnon, VmFile and VmShm in /proc/pid/status
> 
> Vlastimil Babka (2):
>   mm, documentation: clarify /proc/pid/status VmSwap limitations
>   mm, proc: account for shmem swap in /proc/pid/smaps
> 
>  Documentation/filesystems/proc.txt | 18 ++++++++++---
>  arch/s390/mm/pgtable.c             |  5 +---
>  fs/proc/task_mmu.c                 | 52 ++++++++++++++++++++++++++++++++++--
>  include/linux/mm.h                 | 28 ++++++++++++++++++++
>  include/linux/mm_types.h           |  9 ++++---
>  include/linux/shmem_fs.h           |  6 +++++
>  kernel/events/uprobes.c            |  2 +-
>  mm/memory.c                        | 30 +++++++--------------
>  mm/oom_kill.c                      |  5 ++--
>  mm/rmap.c                          | 15 +++--------
>  mm/shmem.c                         | 54 ++++++++++++++++++++++++++++++++++++++
>  11 files changed, 178 insertions(+), 46 deletions(-)
> 



[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 473 bytes --]