Re: cifs: bugfix for unreclaimed writeback pages incifs_writev_requeue()

[Jeff Layton]

On Mon, 4 Mar 2013 15:04:49 +0000
Anton Altaparmakov <aia21@xxxxxxxxx> wrote:

> Hi,
> 
> The below commit that is present in 3.9-rc1 is buggy.  It releases the page at which point it may no longer exist and then it unlocks it afterwards.  Even if you are somehow getting away with it I think it is an explosion/memory corruption waiting to happen...
> 
> Best regards,
> 
> 	Anton
> 
> On 2 Mar 2013, at 19:55, Linux Kernel Mailing List <linux-kernel@xxxxxxxxxxxxxxx> wrote:
> 
> > Gitweb:     http://git.kernel.org/linus/;a=commit;h=c51bb0ea40ca038da26b1fa7d450f4078124af03
> > Commit:     c51bb0ea40ca038da26b1fa7d450f4078124af03
> > Parent:     0b7bc84000d71f3647ca33ab1bf5bd928535c846
> > Author:     Ouyang Maochun <ouyang.maochun@xxxxxxxxxx>
> > AuthorDate: Mon Feb 18 09:54:52 2013 -0600
> > Committer:  Steve French <sfrench@xxxxxxxxxx>
> > CommitDate: Thu Feb 28 09:01:47 2013 -0600
> > 
> >    cifs: bugfix for unreclaimed writeback pages in cifs_writev_requeue()
> > 
> >    Pages get the PG_writeback flag set before cifs sends its
> >    request to SMB server in cifs_writepages(), if the SMB service
> >    goes down, cifs may try to recommit the writing requests in
> >    cifs_writev_requeue(). However, it does not clean its PG_writeback
> >    flag and relaimed the pages even if it fails again in
> >    cifs_writev_requeue(), which may lead to the hanging of the
> >    processes accessing the cifs directory. This patch just cleans
> >    the PG_writeback flags and reclaims the pages under that circumstances.
> > 
> >        Steps to reproduce the bug(trying serveral times may trigger the issue):
> >        1.Write from cifs client continuously.(e.g dd if=/dev/zero of=<cifs file>)
> >        2.Stop SMB service from server.(e.g service smb stop)
> >        3.Wait for two minutesï and then start SMB service from
> >    server.(e.g service smb start)
> >        4.The processes which are accessing cifs directory may hang up.
> > 
> >    Signed-off-by: Ouyang Maochun <ouyang.maochun@xxxxxxxxxx>
> >    Signed-off-by: Jiang Yong <jian.yong5@xxxxxxxxxx>
> >    Tested-by: Zhang Xianwei <zhang.xianwei8@xxxxxxxxxx>
> >    Reviewed-by: Wang Liang <wang.liang82@xxxxxxxxxx>
> >    Reviewed-by: Cai Qu <cai.qu@xxxxxxxxxx>
> >    Reviewed-by: Jiang Biao <jiang.biao2@xxxxxxxxxx>
> >    Reviewed-by: Jeff Layton <jlayton@xxxxxxxxxx>
> >    Reviewed-by: Pavel Shilovsky <piastry@xxxxxxxxxxx>
> >    Signed-off-by: Steve French <sfrench@xxxxxxxxxx>
> > ---
> > fs/cifs/cifssmb.c |    5 ++++-
> > 1 files changed, 4 insertions(+), 1 deletions(-)
> > 
> > diff --git a/fs/cifs/cifssmb.c b/fs/cifs/cifssmb.c
> > index 00e12f2..7353bc5 100644
> > --- a/fs/cifs/cifssmb.c
> > +++ b/fs/cifs/cifssmb.c
> > @@ -1909,8 +1909,11 @@ cifs_writev_requeue(struct cifs_writedata *wdata)
> > 	} while (rc == -EAGAIN);
> > 
> > 	for (i = 0; i < wdata->nr_pages; i++) {
> > -		if (rc != 0)
> > +		if (rc != 0) {
> > 			SetPageError(wdata->pages[i]);
> > +			end_page_writeback(wdata->pages[i]);
> > +			page_cache_release(wdata->pages[i]);
> > +		}
> > 		unlock_page(wdata->pages[i]);
> > 	}
> > 
> 

Well spotted...

We definitely should be unlocking the page before releasing it. I
think it's sufficient to simply move the unlock call before the
check of "rc". I'll send out a patch to do just that once I've
smoke-tested it.

Thanks,
-- 
Jeff Layton <jlayton@xxxxxxxxxx>
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/