Re: [RFC] fuse: Avoid invalidating attrs if writeback_cache enabled

[Yongji Xie <xieyongji@bytedance.com>]

On Wed, Oct 13, 2021 at 9:52 PM Miklos Szeredi <miklos@szeredi.hu> wrote:
>
> On Mon, 11 Oct 2021 at 16:45, Yongji Xie <xieyongji@bytedance.com> wrote:
> >
> > On Mon, Oct 11, 2021 at 9:21 PM Miklos Szeredi <miklos@szeredi.hu> wrote:
> > >
> > > On Mon, 11 Oct 2021 at 11:07, Xie Yongji <xieyongji@bytedance.com> wrote:
> > > >
> > > > Recently we found the performance of small direct writes is bad
> > > > when writeback_cache enabled. This is because we need to get
> > > > attrs from userspace in fuse_update_get_attr() on each write.
> > > > The timeout for the attributes doesn't work since every direct write
> > > > will invalidate the attrs in fuse_direct_IO().
> > > >
> > > > To fix it, this patch tries to avoid invalidating attrs if writeback_cache
> > > > is enabled since we should trust local size/ctime/mtime in this case.
> > >
> > > Hi,
> > >
> > > Thanks for the patch.
> > >
> > > Just pushed an update to
> > > git://git.kernel.org/pub/scm/linux/kernel/git/mszeredi/fuse.gitt#for-next
> > > (9ca3f8697158 ("fuse: selective attribute invalidation")) that should
> > > fix this behavior.
> > >
> >
> > Looks like fuse_update_get_attr() will still get attrs from userspace
> > each time with this commit applied.
> >
> > > Could you please test?
> > >
> >
> > I applied the commit 9ca3f8697158 ("fuse: selective attribute
> > invalidation")  and tested it. But the issue still exists.
>
> Yeah, my bad.  Pushed a more complete set of fixes to #for-next ending with
>
> e15a9a5fca6c ("fuse: take cache_mask into account in getattr")
>
> You should pull or cherry pick the complete branch.
>

I tested this branch, but it still doesn't fix this issue. The
inval_mask = 0x6C0 and cache_mask = 0x2C0, so we still need to get
attrs from userspace. Should we add STATX_BLOCKS to cache_mask?

Thanks,
Yongji