On Wed, Mar 20, 2024 at 11:37:28AM +0100, Andrey Albershteyn wrote:
> On 2024-03-19 16:30:10, Darrick J. Wong wrote:
> > On Wed, Mar 13, 2024 at 10:54:39AM -0700, Darrick J. Wong wrote:
> > > From: Andrey Albershteyn <aalbersh@xxxxxxxxxx>
> > >
> > > For XFS, fsverity's global workqueue is not really suitable due to:
> > >
> > > 1. High priority workqueues are used within XFS to ensure that data
> > > IO completion cannot stall processing of journal IO completions.
> > > Hence using a WQ_HIGHPRI workqueue directly in the user data IO
> > > path is a potential filesystem livelock/deadlock vector.
> > >
> > > 2. The fsverity workqueue is global - it creates a cross-filesystem
> > > contention point.
> > >
> > > This patch adds per-filesystem, per-cpu workqueue for fsverity
> > > work. This allows iomap to add verification work in the read path on
> > > BIO completion.
> > >
> > > Signed-off-by: Andrey Albershteyn <aalbersh@xxxxxxxxxx>
> > > Reviewed-by: Darrick J. Wong <djwong@xxxxxxxxxx>
> > > Signed-off-by: Darrick J. Wong <djwong@xxxxxxxxxx>
> > > ---
> > > fs/super.c | 7 +++++++
> > > include/linux/fs.h | 2 ++
> > > include/linux/fsverity.h | 22 ++++++++++++++++++++++
> > > 3 files changed, 31 insertions(+)
> > >
> > >
> > > diff --git a/fs/super.c b/fs/super.c
> > > index d35e85295489..338d86864200 100644
> > > --- a/fs/super.c
> > > +++ b/fs/super.c
> > > @@ -642,6 +642,13 @@ void generic_shutdown_super(struct super_block *sb)
> > > sb->s_dio_done_wq = NULL;
> > > }
> > >
> > > +#ifdef CONFIG_FS_VERITY
> > > + if (sb->s_read_done_wq) {
> > > + destroy_workqueue(sb->s_read_done_wq);
> > > + sb->s_read_done_wq = NULL;
> > > + }
> > > +#endif
> > > +
> > > if (sop->put_super)
> > > sop->put_super(sb);
> > >
> > > diff --git a/include/linux/fs.h b/include/linux/fs.h
> > > index ed5966a70495..9db24a825d94 100644
> > > --- a/include/linux/fs.h
> > > +++ b/include/linux/fs.h
> > > @@ -1221,6 +1221,8 @@ struct super_block {
> > > #endif
> > > #ifdef CONFIG_FS_VERITY
> > > const struct fsverity_operations *s_vop;
> > > + /* Completion queue for post read verification */
> > > + struct workqueue_struct *s_read_done_wq;
> > > #endif
> > > #if IS_ENABLED(CONFIG_UNICODE)
> > > struct unicode_map *s_encoding;
> > > diff --git a/include/linux/fsverity.h b/include/linux/fsverity.h
> > > index 0973b521ac5a..45b7c613148a 100644
> > > --- a/include/linux/fsverity.h
> > > +++ b/include/linux/fsverity.h
> > > @@ -241,6 +241,22 @@ void fsverity_enqueue_verify_work(struct work_struct *work);
> > > void fsverity_invalidate_block(struct inode *inode,
> > > struct fsverity_blockbuf *block);
> > >
> > > +static inline int fsverity_set_ops(struct super_block *sb,
> > > + const struct fsverity_operations *ops)
> > > +{
> > > + sb->s_vop = ops;
> > > +
> > > + /* Create per-sb workqueue for post read bio verification */
> > > + struct workqueue_struct *wq = alloc_workqueue(
> > > + "pread/%s", (WQ_FREEZABLE | WQ_MEM_RECLAIM), 0, sb->s_id);
> >
> > Looking at this more closely, why is it that the fsverity_read_queue
> > is unbound and tagged WQ_HIGHPRI, whereas this one is instead FREEZEABLE
> > and MEM_RECLAIM and bound?
> >
> > If it's really feasible to use /one/ workqueue for all the read
> > post-processing then this ought to be a fs/super.c helper ala
> > sb_init_dio_done_wq. That said, from Eric's comments on the v5 thread
> > about fsverity and fscrypt locking horns over workqueue stalls I'm not
> > convinced that's true.
>
> There's good explanation by Dave why WQ_HIGHPRI is not a good fit
> for XFS (potential livelock/deadlock):
>
> https://lore.kernel.org/linux-xfs/20221214054357.GI3600936@xxxxxxxxxxxxxxxxxxx/
>
> Based on his feedback I changed it to per-filesystem.
Ah, ok. Why is the workqueue tagged with MEM_RECLAIM though? Does
letting it run actually help out with reclaim? I guess it does by
allowing pages involved in readahead to get to unlocked state where they
can be ripped out. :)
--D
> --
> - Andrey
>
>
