On Thu, May 20, 2021 at 3:30 PM Jan Kara <jack@xxxxxxx> wrote:
> On Thu 20-05-21 14:25:36, Andreas Gruenbacher wrote:
> > Now that we handle self-recursion on the inode glock in gfs2_fault and
> > gfs2_page_mkwrite, we need to take care of more complex deadlock
> > scenarios like the following (example by Jan Kara):
> >
> > Two independent processes P1, P2. Two files F1, F2, and two mappings M1,
> > M2 where M1 is a mapping of F1, M2 is a mapping of F2. Now P1 does DIO
> > to F1 with M2 as a buffer, P2 does DIO to F2 with M1 as a buffer. They
> > can race like:
> >
> > P1 P2
> > read() read()
> > gfs2_file_read_iter() gfs2_file_read_iter()
> > gfs2_file_direct_read() gfs2_file_direct_read()
> > locks glock of F1 locks glock of F2
> > iomap_dio_rw() iomap_dio_rw()
> > bio_iov_iter_get_pages() bio_iov_iter_get_pages()
> > <fault in M2> <fault in M1>
> > gfs2_fault() gfs2_fault()
> > tries to grab glock of F2 tries to grab glock of F1
> >
> > Those kinds of scenarios are much harder to reproduce than
> > self-recursion.
> >
> > We deal with such situations by using the LM_FLAG_OUTER flag to mark
> > "outer" glock taking. Then, when taking an "inner" glock, we use the
> > LM_FLAG_TRY flag so that locking attempts that don't immediately succeed
> > will be aborted. In case of a failed locking attempt, we "unroll" to
> > where the "outer" glock was taken, drop the "outer" glock, and fault in
> > the first offending user page. This will re-trigger the "inner" locking
> > attempt but without the LM_FLAG_TRY flag. Once that has happened, we
> > re-acquire the "outer" glock and retry the original operation.
> >
> > Reported-by: Jan Kara <jack@xxxxxxx>
> > Signed-off-by: Andreas Gruenbacher <agruenba@xxxxxxxxxx>
>
> ...
>
> > diff --git a/fs/gfs2/file.c b/fs/gfs2/file.c
> > index 7d88abb4629b..8b26893f8dc6 100644
> > --- a/fs/gfs2/file.c
> > +++ b/fs/gfs2/file.c
> > @@ -431,21 +431,30 @@ static vm_fault_t gfs2_page_mkwrite(struct vm_fault *vmf)
> > vm_fault_t ret = VM_FAULT_LOCKED;
> > struct gfs2_holder gh;
> > unsigned int length;
> > + u16 flags = 0;
> > loff_t size;
> > int err;
> >
> > sb_start_pagefault(inode->i_sb);
> >
> > - gfs2_holder_init(ip->i_gl, LM_ST_EXCLUSIVE, 0, &gh);
> > + if (current_holds_glock())
> > + flags |= LM_FLAG_TRY;
> > +
> > + gfs2_holder_init(ip->i_gl, LM_ST_EXCLUSIVE, flags, &gh);
> > if (likely(!outer_gh)) {
> > err = gfs2_glock_nq(&gh);
> > if (err) {
> > ret = block_page_mkwrite_return(err);
> > + if (err == GLR_TRYFAILED) {
> > + set_current_needs_retry(true);
> > + ret = VM_FAULT_SIGBUS;
> > + }
>
> I've checked to make sure but do_user_addr_fault() indeed calls do_sigbus()
> which raises the SIGBUS signal. So if the application does not ignore
> SIGBUS, your retry will be visible to the application and can cause all
> sorts of interesting results...
I would have noticed that, but no SIGBUS signals were actually
delivered. So we probably end up in kernelmode_fixup_or_oops() when in
kernel mode, which just does nothing in that case.
Andy Lutomirski, you've been involved with this, could you please shed
some light?
> So you probably need to add a new VM_FAULT_
> return code that will behave like VM_FAULT_SIGBUS except it will not raise
> the signal.
A new VM_FAULT_* flag might make the code easier to read, but I don't
know if we can have one.
> Otherwise it seems to me your approach should work.
Thanks a lot,
Andreas
