On Wed 17-05-17 16:56:27, Jan Kara wrote:
> On Wed 17-05-17 16:13:57, Michal Hocko wrote:
> > On Wed 17-05-17 05:33:01, Tahsin Erdogan wrote:
> > > When a transaction starts, start_this_handle() saves current
> > > PF_MEMALLOC_NOFS value so that it can be restored at journal stop time.
> > > Journal restart is a special case that calls start_this_handle() without
> > > stopping the transaction. start_this_handle() isn't aware that the
> > > original value is already stored so it overwrites it with current value.
> > >
> > > For instance, a call sequence like below leaves PF_MEMALLOC_NOFS flag set
> > > at the end:
> > >
> > > jbd2_journal_start()
> > > jbd2__journal_restart()
> > > jbd2_journal_stop()
> > >
> > > Make jbd2__journal_restart() restore the original value before calling
> > > start_this_handle().
> > >
> > > Fixes: 81378da64de6 ("jbd2: mark the transaction context with the scope GFP_NOFS context")
> > > Signed-off-by: Tahsin Erdogan <tahsin@xxxxxxxxxx>
> > > ---
> > > fs/jbd2/transaction.c | 1 +
> > > 1 file changed, 1 insertion(+)
> > >
> > > diff --git a/fs/jbd2/transaction.c b/fs/jbd2/transaction.c
> > > index 9ee4832b6f8b..dfd6afebdfeb 100644
> > > --- a/fs/jbd2/transaction.c
> > > +++ b/fs/jbd2/transaction.c
> > > @@ -680,6 +680,7 @@ int jbd2__journal_restart(handle_t *handle, int nblocks, gfp_t gfp_mask)
> > >
> > > rwsem_release(&journal->j_trans_commit_map, 1, _THIS_IP_);
> > > handle->h_buffer_credits = nblocks;
> > > + memalloc_nofs_restore(handle->saved_alloc_context);
> > > ret = start_this_handle(journal, handle, gfp_mask);
> > > return ret;
> > > }
> >
> > I remember Jack has mentioned something about nested transaction back
> > then when reviewing the patch. But I cannot remember or find a pointer
> > to that email. I have a vague recollection that there is a reference
> > counting for those transactions.
> >
> > Anyway, Is this patch really correct? So let's say we are in
> > the transaction context already and then you disable the scope
> > NOFS protection, start_this_handle will allocate before it calls
> > memalloc_nofs_save and that would recurse to the filesystem. If
> > anything wouldn't it be better to simply call memalloc_nofs_save only if
> > we start a new transaction? I thought we were doing that already but the
> > code is so convoluted I have hard time to wrap my head around it.
>
> I was thinking about his as well but the fact is jbd2__journal_restart()
> actually does equivalent of jbd2_journal_stop() for the handle above the
> place where memalloc_nofs_restore() was added so in this sense we really
> miss memalloc_nofs_restore() there... So what Tahsin did makes sense.
OK, I will trust you ;). This deserves a comment then, I believe.
Thanks!
--
Michal Hocko
SUSE Labs
