Re: [PATCH v2] bdev: Reduce time holding bd_mutex in sync in blkdev_close()

[Doug Anderson]

Hi Alexander,

On Tue, Mar 24, 2020 at 2:48 PM Douglas Anderson <dianders@xxxxxxxxxxxx> wrote:
>
> While trying to "dd" to the block device for a USB stick, I
> encountered a hung task warning (blocked for > 120 seconds).  I
> managed to come up with an easy way to reproduce this on my system
> (where /dev/sdb is the block device for my USB stick) with:
>
>   while true; do dd if=/dev/zero of=/dev/sdb bs=4M; done
>
> With my reproduction here are the relevant bits from the hung task
> detector:
>
>  INFO: task udevd:294 blocked for more than 122 seconds.
>  ...
>  udevd           D    0   294      1 0x00400008
>  Call trace:
>   ...
>   mutex_lock_nested+0x40/0x50
>   __blkdev_get+0x7c/0x3d4
>   blkdev_get+0x118/0x138
>   blkdev_open+0x94/0xa8
>   do_dentry_open+0x268/0x3a0
>   vfs_open+0x34/0x40
>   path_openat+0x39c/0xdf4
>   do_filp_open+0x90/0x10c
>   do_sys_open+0x150/0x3c8
>   ...
>
>  ...
>  Showing all locks held in the system:
>  ...
>  1 lock held by dd/2798:
>   #0: ffffff814ac1a3b8 (&bdev->bd_mutex){+.+.}, at: __blkdev_put+0x50/0x204
>  ...
>  dd              D    0  2798   2764 0x00400208
>  Call trace:
>   ...
>   schedule+0x8c/0xbc
>   io_schedule+0x1c/0x40
>   wait_on_page_bit_common+0x238/0x338
>   __lock_page+0x5c/0x68
>   write_cache_pages+0x194/0x500
>   generic_writepages+0x64/0xa4
>   blkdev_writepages+0x24/0x30
>   do_writepages+0x48/0xa8
>   __filemap_fdatawrite_range+0xac/0xd8
>   filemap_write_and_wait+0x30/0x84
>   __blkdev_put+0x88/0x204
>   blkdev_put+0xc4/0xe4
>   blkdev_close+0x28/0x38
>   __fput+0xe0/0x238
>   ____fput+0x1c/0x28
>   task_work_run+0xb0/0xe4
>   do_notify_resume+0xfc0/0x14bc
>   work_pending+0x8/0x14
>
> The problem appears related to the fact that my USB disk is terribly
> slow and that I have a lot of RAM in my system to cache things.
> Specifically my writes seem to be happening at ~15 MB/s and I've got
> ~4 GB of RAM in my system that can be used for buffering.  To write 4
> GB of buffer to disk thus takes ~4000 MB / ~15 MB/s = ~267 seconds.
>
> The 267 second number is a problem because in __blkdev_put() we call
> sync_blockdev() while holding the bd_mutex.  Any other callers who
> want the bd_mutex will be blocked for the whole time.
>
> The problem is made worse because I believe blkdev_put() specifically
> tells other tasks (namely udev) to go try to access the device at right
> around the same time we're going to hold the mutex for a long time.
>
> Putting some traces around this (after disabling the hung task detector),
> I could confirm:
>  dd:    437.608600: __blkdev_put() right before sync_blockdev() for sdb
>  udevd: 437.623901: blkdev_open() right before blkdev_get() for sdb
>  dd:    661.468451: __blkdev_put() right after sync_blockdev() for sdb
>  udevd: 663.820426: blkdev_open() right after blkdev_get() for sdb
>
> A simple fix for this is to realize that sync_blockdev() works fine if
> you're not holding the mutex.  Also, it's not the end of the world if
> you sync a little early (though it can have performance impacts).
> Thus we can make a guess that we're going to need to do the sync and
> then do it without holding the mutex.  We still do one last sync with
> the mutex but it should be much, much faster.
>
> With this, my hung task warnings for my test case are gone.
>
> Signed-off-by: Douglas Anderson <dianders@xxxxxxxxxxxx>
> ---
> I didn't put a "Fixes" annotation here because, as far as I can tell,
> this issue has been here "forever" unless someone knows of something
> else that changed that made this possible to hit.  This could probably
> get picked back to any stable tree that anyone is still maintaining.
>
> Changes in v2:
> - Don't bother holding the mutex when checking "bd_openers".
>
>  fs/block_dev.c | 10 ++++++++++
>  1 file changed, 10 insertions(+)
>
> diff --git a/fs/block_dev.c b/fs/block_dev.c
> index 9501880dff5e..40c57a9cc91a 100644
> --- a/fs/block_dev.c
> +++ b/fs/block_dev.c
> @@ -1892,6 +1892,16 @@ static void __blkdev_put(struct block_device *bdev, fmode_t mode, int for_part)
>         struct gendisk *disk = bdev->bd_disk;
>         struct block_device *victim = NULL;
>
> +       /*
> +        * Sync early if it looks like we're the last one.  If someone else
> +        * opens the block device between now and the decrement of bd_openers
> +        * then we did a sync that we didn't need to, but that's not the end
> +        * of the world and we want to avoid long (could be several minute)
> +        * syncs while holding the mutex.
> +        */
> +       if (bdev->bd_openers == 1)
> +               sync_blockdev(bdev);
> +
>         mutex_lock_nested(&bdev->bd_mutex, for_part);
>         if (for_part)
>                 bdev->bd_part_count--;
> --
> 2.25.1.696.g5e7596f4ac-goog

Are you the right person to land this patch?  If so, is there anything
else that needs to be done?  Jens: if you should be the person to land
(as suggested by "git log" but not by "get_maintainer") I'm happy to
repost with collected tags.  Originally I trusted "get_maintainer" to
help point me to the right person.

Thanks!

-Doug