On Tue, 2020-12-22 at 12:55 -0500, Vivek Goyal wrote:
> On Tue, Dec 22, 2020 at 05:46:37PM +0000, Matthew Wilcox wrote:
> > On Tue, Dec 22, 2020 at 11:29:25AM -0500, Vivek Goyal wrote:
> > > On Tue, Dec 22, 2020 at 04:20:27PM +0000, Matthew Wilcox wrote:
> > > > On Mon, Dec 21, 2020 at 02:50:55PM -0500, Vivek Goyal wrote:
> > > > > +static int ovl_errseq_check_advance(struct super_block *sb, struct
> > > > > file *file)
> > > > > +{
> > > > > + struct ovl_fs *ofs = sb->s_fs_info;
> > > > > + struct super_block *upper_sb;
> > > > > + int ret;
> > > > > +
> > > > > + if (!ovl_upper_mnt(ofs))
> > > > > + return 0;
> > > > > +
> > > > > + upper_sb = ovl_upper_mnt(ofs)->mnt_sb;
> > > > > +
> > > > > + if (!errseq_check(&upper_sb->s_wb_err, file->f_sb_err))
> > > > > + return 0;
> > > > > +
> > > > > + /* Something changed, must use slow path */
> > > > > + spin_lock(&file->f_lock);
> > > > > + ret = errseq_check_and_advance(&upper_sb->s_wb_err,
> > > > > &file->f_sb_err);
> > > > > + spin_unlock(&file->f_lock);
> > > >
> > > > Why are you microoptimising syncfs()? Are there really applications
> > > > which
> > > > call syncfs() in a massively parallel manner on the same file
> > > > descriptor?
> > >
> > > This is atleast theoritical race. I am not aware which application can
> > > trigger this race. So to me it makes sense to fix the race.
> > >
> > > Jeff Layton also posted a fix for syncfs().
> > >
> > > https://lore.kernel.org/linux-fsdevel/20201219134804.20034-1-jlay...@kernel.org/
> > >
> > > To me it makes sense to fix the race irrespective of the fact if somebody
> > > hit it or not. People end up copying code in other parts of kernel and
> > > and they will atleast copy race free code.
> >
> > Let me try again. "Why are you trying to avoid taking the spinlock?"
>
> Aha.., sorry, I misunderstood your question. I don't have a good answer.
> I just copied the code from Jeff Layton's patch.
>
> Agreed that cost of taking spin lock will not be significant until
> syncfs() is called at high frequency. Having said that, most of the
> time taking spin lock will not be needed, so avoiding it with
> a simple call to errseq_check() sounds reasonable too.
>
> I don't have any strong opinions here. I am fine with any of the
> implementation people like.
>
It is a micro-optimization, but we'll almost always be able to avoid
taking the lock altogether. Errors here should be very, very infrequent.
That said I don't have strong feelings on this either.
--
Jeff Layton <jlay...@kernel.org>
