On Fri, 20 Oct 2023 at 08:12, Andy Shevchenko
<andriy.shevchenko@xxxxxxxxx> wrote:
>
> > > --- a/fs/quota/dquot.c
> > > +++ b/fs/quota/dquot.c
> > > @@ -632,8 +632,10 @@ static inline int dquot_write_dquot(struct dquot *dquot)
> > > {
> > > int ret = dquot->dq_sb->dq_op->write_dquot(dquot);
> > > if (ret < 0) {
> > > +#if 0
> > > quota_error(dquot->dq_sb, "Can't write quota structure "
> > > "(error %d). Quota may get out of sync!", ret);
> > > +#endif
> > > /* Clear dirty bit anyway to avoid infinite loop. */
> > > clear_dquot_dirty(dquot);
> > > }
>
> Doing the same on the my branch based on top of v6.6-rc6 does not help.
> So looks like a race condition somewhere happening related to that dirty bit
> (as comment states it needs to be cleaned to avoid infinite loop, that's
> probably what happens).
Hmm. Normally, dirty bits should always be cleared *before* the
write-back, not after it. Otherwise you might lose a dirty event that
happened *during* writeback.
But I don't know the quota code.
... the fact that the #if 0 doesn't help in another case does say that
it's not the quota_error() call itself. Which it really couldn't have
been (apart from timing and compiler bugs), but it's still a data
point, I guess.
Linus
