Re: [PATCH v2 1/2] procfs: Add 'size' to /proc/<pid>/fdinfo/

[Kalesh Singh <kaleshsingh@google.com>]

On Thu, Jun 30, 2022 at 5:03 AM Brian Foster <bfoster@redhat.com> wrote:
>
> On Thu, Jun 30, 2022 at 07:48:46AM -0400, Brian Foster wrote:
> > On Wed, Jun 29, 2022 at 01:43:11PM -0700, Kalesh Singh wrote:
> > > On Wed, Jun 29, 2022 at 5:23 AM Brian Foster <bfoster@redhat.com> wrote:
> > > >
> > > > On Tue, Jun 28, 2022 at 03:38:02PM -0700, Kalesh Singh wrote:
> > > > > On Tue, Jun 28, 2022 at 4:54 AM Brian Foster <bfoster@redhat.com> wrote:
> > > > > >
> > > > > > On Thu, Jun 23, 2022 at 03:06:06PM -0700, Kalesh Singh wrote:
> > > > > > > To be able to account the amount of memory a process is keeping pinned
> > > > > > > by open file descriptors add a 'size' field to fdinfo output.
> > > > > > >
> > > > > > > dmabufs fds already expose a 'size' field for this reason, remove this
> > > > > > > and make it a common field for all fds. This allows tracking of
> > > > > > > other types of memory (e.g. memfd and ashmem in Android).
> > > > > > >
> > > > > > > Signed-off-by: Kalesh Singh <kaleshsingh@google.com>
> > > > > > > Reviewed-by: Christian König <christian.koenig@amd.com>
> > > > > > > ---
> > > > > > >
> > > > > > > Changes in v2:
> > > > > > >   - Add Christian's Reviewed-by
> > > > > > >
> > > > > > > Changes from rfc:
> > > > > > >   - Split adding 'size' and 'path' into a separate patches, per Christian
> > > > > > >   - Split fdinfo seq_printf into separate lines, per Christian
> > > > > > >   - Fix indentation (use tabs) in documentaion, per Randy
> > > > > > >
> > > > > > >  Documentation/filesystems/proc.rst | 12 ++++++++++--
> > > > > > >  drivers/dma-buf/dma-buf.c          |  1 -
> > > > > > >  fs/proc/fd.c                       |  9 +++++----
> > > > > > >  3 files changed, 15 insertions(+), 7 deletions(-)
> > > > > > >
> > > > ...
> > > > > >
> > > > > > Also not sure if it matters that much for your use case, but something
> > > > > > worth noting at least with shmem is that one can do something like:
> > > > > >
> > > > > > # cat /proc/meminfo | grep Shmem:
> > > > > > Shmem:               764 kB
> > > > > > # xfs_io -fc "falloc -k 0 10m" ./file
> > > > > > # ls -alh file
> > > > > > -rw-------. 1 root root 0 Jun 28 07:22 file
> > > > > > # stat file
> > > > > >   File: file
> > > > > >   Size: 0               Blocks: 20480      IO Block: 4096   regular empty file
> > > > > > # cat /proc/meminfo | grep Shmem:
> > > > > > Shmem:             11004 kB
> > > > > >
> > > > > > ... where the resulting memory usage isn't reflected in i_size (but is
> > > > > > is in i_blocks/bytes).
> > > > >
> > > > > I tried a similar experiment a few times, but I don't see the same
> > > > > results. In my case, there is not any change in shmem. IIUC the
> > > > > fallocate is allocating the disk space not shared memory.
> > > > >
> > > >
> > > > Sorry, it was implied in my previous test was that I was running against
> > > > tmpfs. So regardless of fs, the fallocate keep_size semantics shown in
> > > > both cases is as expected: the underlying blocks are allocated and the
> > > > inode size is unchanged.
> > > >
> > > > What wasn't totally clear to me when I read this patch was 1. whether
> > > > tmpfs refers to Shmem and 2. whether tmpfs allowed this sort of
> > > > operation. The test above seems to confirm both, however, right? E.g., a
> > > > more detailed example:
> > > >
> > > > # mount | grep /tmp
> > > > tmpfs on /tmp type tmpfs (rw,nosuid,nodev,seclabel,nr_inodes=1048576,inode64)
> > > > # cat /proc/meminfo | grep Shmem:
> > > > Shmem:              5300 kB
> > > > # xfs_io -fc "falloc -k 0 1g" /tmp/file
> > > > # stat /tmp/file
> > > >   File: /tmp/file
> > > >   Size: 0               Blocks: 2097152    IO Block: 4096   regular empty file
> > > > Device: 22h/34d Inode: 45          Links: 1
> > > > Access: (0600/-rw-------)  Uid: (    0/    root)   Gid: (    0/    root)
> > > > Context: unconfined_u:object_r:user_tmp_t:s0
> > > > Access: 2022-06-29 08:04:01.301307154 -0400
> > > > Modify: 2022-06-29 08:04:01.301307154 -0400
> > > > Change: 2022-06-29 08:04:01.451312834 -0400
> > > >  Birth: 2022-06-29 08:04:01.301307154 -0400
> > > > # cat /proc/meminfo | grep Shmem:
> > > > Shmem:           1053876 kB
> > > > # rm -f /tmp/file
> > > > # cat /proc/meminfo | grep Shmem:
> > > > Shmem:              5300 kB
> > > >
> > > > So clearly this impacts Shmem.. was your test run against tmpfs or some
> > > > other (disk based) fs?
> > >
> > > Hi Brian,
> > >
> > > Thanks for clarifying. My issue was tmpfs not mounted at /tmp in my system:
> > >
> > > ==> meminfo.start <==
> > > Shmem:               572 kB
> > > ==> meminfo.stop <==
> > > Shmem:             51688 kB
> > >
> >
> > Ok, makes sense.
> >
> > > >
> > > > FWIW, I don't have any objection to exposing inode size if it's commonly
> > > > useful information. My feedback was more just an fyi that i_size doesn't
> > > > necessarily reflect underlying space consumption (whether it's memory or
> > > > disk space) in more generic cases, because it sounds like that is really
> > > > what you're after here. The opposite example to the above would be
> > > > something like an 'xfs_io -fc "truncate 1t" /tmp/file', which shows a
> > > > 1TB inode size with zero additional shmem usage.
> > >
> > > From these cases, it seems the more generic way to do this is by
> > > calculating the actual size consumed using the blocks. (i_blocks *
> > > 512). So in the latter example  'xfs_io -fc "truncate 1t" /tmp/file'
> > > the size consumed would be zero. Let me know if it sounds ok to you
> > > and I can repost the updated version.
> > >
> >
> > That sounds a bit more useful to me if you're interested in space usage,
> > or at least I don't have a better idea for you. ;)
> >
> > One thing to note is that I'm not sure whether all fs' use i_blocks
> > reliably. E.g., XFS populates stat->blocks via a separate block counter
> > in the XFS specific inode structure (see xfs_vn_getattr()). A bunch of
> > other fs' seem to touch it so perhaps that is just an outlier. You could
> > consider fixing that up, perhaps make a ->getattr() call to avoid it, or
> > just use the field directly if it's useful enough as is and there are no
> > other objections. Something to think about anyways..
> >

Hi Brian,

Thanks for pointing it out. Let me take a look into the xfs case.

>
> Oh, I wonder if you're looking for similar "file rss" information this
> series wants to collect/expose..?
>
> https://lore.kernel.org/linux-fsdevel/20220624080444.7619-1-christian.koenig@amd.com/#r

Christian's series seems to have some overlap with what we want to
achieve here. IIUC it exposes the information on the per process
granularity. Perhaps if that approach is agreed on, I think we can use
the file_rss() f_op to expose the  per file size in the fdinfo for the
cases where the i_blocks are unreliable.

Thanks,
Kalesh

>
> Brian
>
> > Brian
> >
> > > Thanks,
> > > Kalesh
> > >
> > > >
> > > > Brian
> > > >
> > > > > cat /proc/meminfo > meminfo.start
> > > > > xfs_io -fc "falloc -k 0 50m" ./xfs_file
> > > > > cat /proc/meminfo > meminfo.stop
> > > > > tail -n +1 meminfo.st* | grep -i '==\|Shmem:'
> > > > >
> > > > > ==> meminfo.start <==
> > > > > Shmem:               484 kB
> > > > > ==> meminfo.stop <==
> > > > > Shmem:               484 kB
> > > > >
> > > > > ls -lh xfs_file
> > > > > -rw------- 1 root root 0 Jun 28 15:12 xfs_file
> > > > >
> > > > > stat xfs_file
> > > > >   File: xfs_file
> > > > >   Size: 0               Blocks: 102400     IO Block: 4096   regular empty file
> > > > >
> > > > > Thanks,
> > > > > Kalesh
> > > > >
> > > > > >
> > > > > > Brian
> > > > > >
> > > > > > >
> > > > > > >       /* show_fd_locks() never deferences files so a stale value is safe */
> > > > > > >       show_fd_locks(m, file, files);
> > > > > > > --
> > > > > > > 2.37.0.rc0.161.g10f37bed90-goog
> > > > > > >
> > > > > >
> > > > >
> > > >
> > > > --
> > > > To unsubscribe from this group and stop receiving emails from it, send an email to kernel-team+unsubscribe@android.com.
> > > >
> > >
>