Re: [PATCH RFC 07/10] fs/ext4: Fail truncate if pages are GUP pinned
From: Ira Weiny
Date: Thu Jun 06 2019 - 12:20:00 EST
On Thu, Jun 06, 2019 at 12:58:55PM +0200, Jan Kara wrote:
> On Wed 05-06-19 18:45:40, ira.weiny@xxxxxxxxx wrote:
> > From: Ira Weiny <ira.weiny@xxxxxxxxx>
> > If pages are actively gup pinned fail the truncate operation.
> > Signed-off-by: Ira Weiny <ira.weiny@xxxxxxxxx>
> > ---
> > fs/ext4/inode.c | 3 +++
> > 1 file changed, 3 insertions(+)
> > diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
> > index 75f543f384e4..1ded83ec08c0 100644
> > --- a/fs/ext4/inode.c
> > +++ b/fs/ext4/inode.c
> > @@ -4250,6 +4250,9 @@ int ext4_break_layouts(struct inode *inode, loff_t offset, loff_t len)
> > if (!page)
> > return 0;
> > + if (page_gup_pinned(page))
> > + return -ETXTBSY;
> > +
> > error = ___wait_var_event(&page->_refcount,
> > atomic_read(&page->_refcount) == 1,
> > TASK_INTERRUPTIBLE, 0, 0,
> This caught my eye. Does this mean that now truncate for a file which has
> temporary gup users (such buffers for DIO) can fail with ETXTBUSY?
I thought about that before and I _thought_ I had accounted for it. But I
think you are right...
> doesn't look desirable.
No not desirable at all... Ah it just dawned on my why I thought it was ok...
I was wrong. :-/
> If we would mandate layout lease while pages are
> pinned as I suggested, this could be dealt with by checking for leases with
> pins (breaking such lease would return error and not break it) and if
> breaking leases succeeds (i.e., there are no long-term pinned pages), we'd
> just wait for the remaining references as we do now.
But I'm going to respond with some of the challenges of this (and ideas I had)
when replying to your other email.
> Jan Kara <jack@xxxxxxxx>
> SUSE Labs, CR