On Thu, 27 Feb 2014 13:51:25 -0800 Andrew Morton [off-list ref]
wrote:

On Fri, 28 Feb 2014 08:34:43 +1100 NeilBrown [off-list ref] wrote:

quoted

On Thu, 27 Feb 2014 12:58:07 -0800 Andrew Morton [off-list ref]
wrote:

quoted

On Thu, 27 Feb 2014 17:24:45 +1100 NeilBrown [off-list ref] wrote:

quoted

If poll or select is waiting on /proc/mdstat when md-mod is unloaded
an oops will ensure when the poll/select completes.

This is because the wait_queue_head which is registered with poll_wait()
is local to the module and no longer exists when the poll completes and
detaches that wait_queue_head (in poll_free_wait -> remove_wait_queue).

To fix this we need the wait_queue_head to have (at least) the same life
time as the proc_dir_entry.  So this patch places it in that structure.

We:
  - add pde_poll_wait to struct proc_dir_entry
  - call poll_wait() passing this when poll() is called on the proc file
  - export a function proc_wake_up which will call wake_up() on pde_poll_wait

and make use of all that in md.c

This sounds wrong.  If a userspace process is waiting on
md_event_waiters then the md module is "busy" and the rmmod attempt
should fail?

Al Viro says "no" quite firmly.

I think the core argument is that

  rmmod md-mod < /proc/mdstat

would deadlock.

Well, only if the rmmod hangs around waiting for the module to go idle.
I'm thinking rmmod should fail.  EBUSY.

quoted

http://marc.info/?l=linux-fsdevel&m=133024267507384

Why don't a billion blocking-procfs-read sites have this problem?

This issue of a file in /proc (actually /proc/fs/nfsd which is a separate
filesystem) blocking came up on the NFS list recently.
Experiments suggested that no other files in /proc block reads
(though /proc/vmcore takes somewhat longer to 'cat' than most).
So the offending file was changed to never block.

i.e. I think your 'billion' estimate is a little high, by one billion.
:-)

NeilBrown