Re: a suspected data race bug in Ext4 file system

[Mingming Cao <cmm@us.ibm.com>]

On Thu, 2008-11-06 at 18:51 -0800, Andrew Morton wrote:
> On Fri, 7 Nov 2008 11:09:49 +0900 "______ shin hong" <hongshin@gmail.com> wrote:
> 
> > Dear Ext4 maintainer,
> > 
> > I found a suspected data race bug at ext4_release_file() in ext4/file.c .
> > 
> > Is it necessary to re-check whether the if-statement's
> > condition(atomic_read(&inode->i_writecount) == 1) is satisifed after
> > semaphore down operation?
> > 
> > It seems to use "test and test and set" pattern(double-checking)
> >  but the second "test" seems to be missed.
> > 
> > Please examine this function. Thank you.
> > 
> > p.s. I sent the same report to sct@redhat.com. But I did not received
> > any reply.
> >        I think my mail was discarded so I am sending the same again.
> > 
> 
> An appropriate way to report this possible bug is to Cc: the ext4 mailing
> list, which I have cc'ed here.

I guess we could add the check after grab the i_data_sem, but not sure
how useful is that, there is always a window that during the process of
freeing preallocation, another new writer to the same inode. The
i_data_sem semaphore is hold to prevent race between discard
preallocation and use of preallocation, not to protect i_writecount.

The check of inode's i_writecount is to avoid doing discard for every
file close, the idea is only try to discard the preallocation for the
last writer of the file. But the i_writecount could be >1 at the time of
discard preallocation. If a new writer comes in after i_data_sem is hold
then the preallocation is freed, then by the time the new writer comes
in and need block allocation, the preallocate gets build up
atomatically.

 
Mingming