On Sun, 26 Feb 2012 16:33:25 +0800 "majianpeng" [off-list ref] wrote:
I wrote a letter to maintainer of procfs. He's answer is follow:
quoted
No. Read remove_proc_entry(), especially the part under ->pde_unload_lock.
The whole damn point of that stuff is that opened file on procfs does *not*
pin the module down; IO in progress does, but that's it.
You can't deadlock rmmod foobar </proc/crap/foobar; with your patch we'll
be back to a pile of deadlocks in there.
IOW, NAK.
Hmmm... that sounds like Al Viro.
I see the point of that code now. And your patch would be open to the same
problem wouldn't it. i.e.
rmmod md_mod < /proc/mdstat
would deadlock??
So we still need to find a fix that actually works correctly.
That probably means moving the call to 'poll_wait' into procfs code, and using
a wait_queue_head which is managed by procfs.
So maybe:
- put a wait_queue_head in 'struct proc_dir_entry'
- have proc_reg_poll call poll_wait passing that wait_queue_head
- write a function e.g. proc_poll_wake which calls wake_up on that
wait_queue_head
- have md.c save the return value from proc_create, and call the above
function instead of calling "wake_up(&md_event_waiters)"
That should work.
NeilBrown