[lkml]   [2012]   [Feb]   [27]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCH] hugetlbfs: Add new rw_semaphore to fix truncate/read race
On Mon, Feb 27, 2012 at 03:11:35PM -0800, Andrew Morton wrote:

> This patch comes somewhat out of the blue and I'm unsure what's going on.
> You say there's some (potential?) deadlock with mmap, but it is
> undescribed. Have people observed this deadlock? Has it caused
> lockdep warnings? Please update the changelog to fully describe the
> bug.

There's one simple rule: never, ever take ->i_mutex under ->mmap_sem.
E.g. in any ->mmap() (obvious - mmap(2) calls that under ->mmap_sem) or
any ->release() of mappable file (munmap(2) does fput() under ->mmap_sem
and that will call ->release() if no other references are still around).

Hugetlbfs is slightly unusual since it takes ->i_mutex in read() - usually
that's done in write(), while read() doesn't bother with that. In either
case you do copying to/from userland buffer while holding ->i_mutex, which
nests ->mmap_sem within it.

> Also, the new truncate_sem is undoumented. This leaves readers to work
> out for themselves what it might be for. Please let's add code
> comments which completely describe the race, and how this lock prevents
> it.
> We should also document our locking rules.

Hell, yes. I've spent the last couple of weeks crawling through VM-related
code and locking in there is _scary_. "Convoluted" doesn't even begin to
cover it, especially when it gets to "what locks are required when accessing
this field" ;-/ Got quite a catch out of that trawl by now...

> When should code take this
> lock? What are its ranking rules with respect to i_mutex, i_mmap_mutex
> and possibly others?

 \ /
  Last update: 2012-02-28 01:05    [W:0.054 / U:2.828 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site