On Fri, 2022-09-23 at 12:02 -0700, Darrick J. Wong wrote:
> On Wed, Sep 21, 2022 at 10:44:34PM -0700,
> allison.henderson@xxxxxxxxxx wrote:
> > From: Allison Henderson <allison.henderson@xxxxxxxxxx>
> >
> > Renames that generate parent pointer updates can join up to 5
> > inodes locked in sorted order. So we need to increase the
> > number of defer ops inodes and relock them in the same way.
> >
> > Signed-off-by: Allison Henderson <allison.henderson@xxxxxxxxxx>
> > ---
> > fs/xfs/libxfs/xfs_defer.c | 28 ++++++++++++++++++++++++++--
> > fs/xfs/libxfs/xfs_defer.h | 8 +++++++-
> > fs/xfs/xfs_inode.c | 2 +-
> > fs/xfs/xfs_inode.h | 1 +
> > 4 files changed, 35 insertions(+), 4 deletions(-)
> >
> > diff --git a/fs/xfs/libxfs/xfs_defer.c b/fs/xfs/libxfs/xfs_defer.c
> > index 5a321b783398..c0279b57e51d 100644
> > --- a/fs/xfs/libxfs/xfs_defer.c
> > +++ b/fs/xfs/libxfs/xfs_defer.c
> > @@ -820,13 +820,37 @@ xfs_defer_ops_continue(
> > struct xfs_trans *tp,
> > struct xfs_defer_resources *dres)
> > {
> > - unsigned int i;
> > + unsigned int i, j;
> > + struct
> > xfs_inode *sips[XFS_DEFER_OPS_NR_INODES];
> > + struct xfs_inode *temp;
> >
> > ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
> > ASSERT(!(tp->t_flags & XFS_TRANS_DIRTY));
> >
> > /* Lock the captured resources to the new transaction. */
> > - if (dfc->dfc_held.dr_inos == 2)
> > + if (dfc->dfc_held.dr_inos > 2) {
> > + /*
> > + * Renames with parent pointer updates can lock up
> > to 5 inodes,
> > + * sorted by their inode number. So we need to
> > make sure they
> > + * are relocked in the same way.
> > + */
> > + memset(sips, 0, sizeof(sips));
> > + for (i = 0; i < dfc->dfc_held.dr_inos; i++)
> > + sips[i] = dfc->dfc_held.dr_ip[i];
> > +
> > + /* Bubble sort of at most 5 inodes */
> > + for (i = 0; i < dfc->dfc_held.dr_inos; i++) {
> > + for (j = 1; j < dfc->dfc_held.dr_inos; j++)
> > {
> > + if (sips[j]->i_ino < sips[j-1]-
> > >i_ino) {
> > + temp = sips[j];
> > + sips[j] = sips[j-1];
> > + sips[j-1] = temp;
> > + }
> > + }
> > + }
> > +
> > + xfs_lock_inodes(sips, dfc->dfc_held.dr_inos,
> > XFS_ILOCK_EXCL);
> > + } else if (dfc->dfc_held.dr_inos == 2)
> > xfs_lock_two_inodes(dfc->dfc_held.dr_ip[0],
> > XFS_ILOCK_EXCL,
> > dfc->dfc_held.dr_ip[1],
> > XFS_ILOCK_EXCL);
> > else if (dfc->dfc_held.dr_inos == 1)
> > diff --git a/fs/xfs/libxfs/xfs_defer.h b/fs/xfs/libxfs/xfs_defer.h
> > index 114a3a4930a3..3e4029d2ce41 100644
> > --- a/fs/xfs/libxfs/xfs_defer.h
> > +++ b/fs/xfs/libxfs/xfs_defer.h
> > @@ -70,7 +70,13 @@ extern const struct xfs_defer_op_type
> > xfs_attr_defer_type;
> > /*
> > * Deferred operation item relogging limits.
> > */
> > -#define XFS_DEFER_OPS_NR_INODES 2 /* join up to two
> > inodes */
> > +
> > +/*
> > + * Rename w/ parent pointers can require up to 5 inodes with
> > defered ops to
>
> s/defered/deferred/
>
> > + * be joined to the transaction: src_dp, target_dp, src_ip,
> > target_ip, and wip.
> > + * These inodes are locked in sorted order by their inode numbers
> > + */
>
> When would we be processing *five* different inodes?
>
> Does this happen when src_ip is a child of src_dp, target_ip is a
> child
> of target_dp, all four inodes are distinct, and the VFS asks us to
> move
> src_ip from src_dp to target_dp, unlink target_ip from target_dp,
> *and*
> install a whiteout in the dirent in src_dp? In which case src_ip,
> target_ip, and wip all need parent pointer adjustments?
Right, the src_ip is overwriting the target_ip in another directory.
>
> So that's three inodes that need to stay locked for deferred
> operations;
> what about src_dp and target_dp?
Well they need their directory entries updated, also we need the
parents to reconstruct what the old parent pointer for their children
were so that we can remove the correct one from the child ips.
> I don't think they need to stay
> locked, but OTOH it's probably easier (for now) to lock everything
> until
> the end of the entire rename operation instead of making everyone
> reason about when they fall off the transaction chain, right?
>
> If the answer to all that is yes and the typo gets fixed, then
It makes sense to me that they should stay locked. Will update the
typo.
Thanks for the reviews!
Allison
> Reviewed-by: Darrick J. Wong <djwong@xxxxxxxxxx>
>
> --D
>
> > +#define XFS_DEFER_OPS_NR_INODES 5
> > #define XFS_DEFER_OPS_NR_BUFS 2 /* join up to two buffers
> > */
> >
> > /* Resources that must be held across a transaction roll. */
> > diff --git a/fs/xfs/xfs_inode.c b/fs/xfs/xfs_inode.c
> > index c000b74dd203..5ebbfceb1ada 100644
> > --- a/fs/xfs/xfs_inode.c
> > +++ b/fs/xfs/xfs_inode.c
> > @@ -447,7 +447,7 @@ xfs_lock_inumorder(
> > * lock more than one at a time, lockdep will report false
> > positives saying we
> > * have violated locking orders.
> > */
> > -static void
> > +void
> > xfs_lock_inodes(
> > struct xfs_inode **ips,
> > int inodes,
> > diff --git a/fs/xfs/xfs_inode.h b/fs/xfs/xfs_inode.h
> > index fa780f08dc89..2eaed98af814 100644
> > --- a/fs/xfs/xfs_inode.h
> > +++ b/fs/xfs/xfs_inode.h
> > @@ -574,5 +574,6 @@ void xfs_end_io(struct work_struct *work);
> >
> > int xfs_ilock2_io_mmap(struct xfs_inode *ip1, struct xfs_inode
> > *ip2);
> > void xfs_iunlock2_io_mmap(struct xfs_inode *ip1, struct xfs_inode
> > *ip2);
> > +void xfs_lock_inodes(struct xfs_inode **ips, int inodes, uint
> > lock_mode);
> >
> > #endif /* __XFS_INODE_H__ */
> > --
> > 2.25.1
> >
