platform/kernel/linux-rpi.git
3 years agoxfs: move the XFS_IFEXTENTS check into xfs_iread_extents
Christoph Hellwig [Tue, 13 Apr 2021 18:15:09 +0000 (11:15 -0700)]
xfs: move the XFS_IFEXTENTS check into xfs_iread_extents

Move the XFS_IFEXTENTS check from the callers into xfs_iread_extents to
simplify the code.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: drop unnecessary setfilesize helper
Brian Foster [Fri, 9 Apr 2021 17:27:56 +0000 (10:27 -0700)]
xfs: drop unnecessary setfilesize helper

xfs_setfilesize() is the only remaining caller of the internal
__xfs_setfilesize() helper. Fold them into a single function.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: drop unused ioend private merge and setfilesize code
Brian Foster [Fri, 9 Apr 2021 17:27:55 +0000 (10:27 -0700)]
xfs: drop unused ioend private merge and setfilesize code

XFS no longer attaches anthing to ioend->io_private. Remove the
unnecessary ->io_private merging code. This removes the only remaining
user of xfs_setfilesize_ioend() so remove that function as well.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: open code ioend needs workqueue helper
Brian Foster [Fri, 9 Apr 2021 17:27:55 +0000 (10:27 -0700)]
xfs: open code ioend needs workqueue helper

Open code xfs_ioend_needs_workqueue() into the only remaining
caller.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: drop submit side trans alloc for append ioends
Brian Foster [Fri, 9 Apr 2021 17:27:43 +0000 (10:27 -0700)]
xfs: drop submit side trans alloc for append ioends

Per-inode ioend completion batching has a log reservation deadlock
vector between preallocated append transactions and transactions
that are acquired at completion time for other purposes (i.e.,
unwritten extent conversion or COW fork remaps). For example, if the
ioend completion workqueue task executes on a batch of ioends that
are sorted such that an append ioend sits at the tail, it's possible
for the outstanding append transaction reservation to block
allocation of transactions required to process preceding ioends in
the list.

Append ioend completion is historically the common path for on-disk
inode size updates. While file extending writes may have completed
sometime earlier, the on-disk inode size is only updated after
successful writeback completion. These transactions are preallocated
serially from writeback context to mitigate concurrency and
associated log reservation pressure across completions processed by
multi-threaded workqueue tasks.

However, now that delalloc blocks unconditionally map to unwritten
extents at physical block allocation time, size updates via append
ioends are relatively rare. This means that inode size updates most
commonly occur as part of the preexisting completion time
transaction to convert unwritten extents. As a result, there is no
longer a strong need to preallocate size update transactions.

Remove the preallocation of inode size update transactions to avoid
the ioend completion processing log reservation deadlock. Instead,
continue to send all potential size extending ioends to workqueue
context for completion and allocate the transaction from that
context. This ensures that no outstanding log reservation is owned
by the ioend completion worker task when it begins to process
ioends.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: fix return of uninitialized value in variable error
Colin Ian King [Fri, 9 Apr 2021 17:27:34 +0000 (10:27 -0700)]
xfs: fix return of uninitialized value in variable error

A previous commit removed a call to xfs_attr3_leaf_read that
assigned an error return code to variable error. We now have
a few early error return paths to label 'out' that return
error if error is set; however error now is uninitialized
so potentially garbage is being returned.  Fix this by setting
error to zero to restore the original behaviour where error
was zero at the label 'restart'.

Addresses-Coverity: ("Uninitialized scalar variable")
Fixes: 07120f1abdff ("xfs: Add xfs_has_attr and subroutines")
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: get rid of the ip parameter to xchk_setup_*
Darrick J. Wong [Thu, 8 Apr 2021 00:59:39 +0000 (17:59 -0700)]
xfs: get rid of the ip parameter to xchk_setup_*

Now that the scrub context stores a pointer to the file that was used to
invoke the scrub call, the struct xfs_inode pointer that we passed to
all the setup functions is no longer necessary.  This is only ever used
if the caller wants us to scrub the metadata of the open file.

Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Chandan Babu R <chandanrlinux@gmail.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
3 years agoxfs: fix scrub and remount-ro protection when running scrub
Darrick J. Wong [Mon, 22 Mar 2021 16:51:51 +0000 (09:51 -0700)]
xfs: fix scrub and remount-ro protection when running scrub

While running a new fstest that races a readonly remount with scrub
running in repair mode, I observed the kernel tripping over debugging
assertions in the log quiesce code that were checking that the CIL was
empty.  When the sysadmin runs scrub in repair mode, the scrub code
allocates real transactions (with reservations) to change things, but
doesn't increment the superblock writers count to block a readonly
remount attempt while it is running.

We don't require the userspace caller to have a writable file descriptor
to run repairs, so we have to call mnt_want_write_file to obtain freeze
protection and increment the writers count.  It's ok to remove the call
to sb_start_write for the dry-run case because commit 8321ddb2fa29
removed the behavior where scrub and fsfreeze fight over the buffer LRU.

Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Chandan Babu R <chandanrlinux@gmail.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
3 years agoxfs: move the check for post-EOF mappings into xfs_can_free_eofblocks
Darrick J. Wong [Tue, 23 Mar 2021 23:59:31 +0000 (16:59 -0700)]
xfs: move the check for post-EOF mappings into xfs_can_free_eofblocks

Fix the weird split of responsibilities between xfs_can_free_eofblocks
and xfs_free_eofblocks by moving the chunk of code that looks for any
actual post-EOF space mappings from the second function into the first.

This clears the way for deferred inode inactivation to be able to decide
if an inode needs inactivation work before committing the released inode
to the inactivation code paths (vs. marking it for reclaim).

Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
3 years agoxfs: move the xfs_can_free_eofblocks call under the IOLOCK
Darrick J. Wong [Tue, 23 Mar 2021 23:59:31 +0000 (16:59 -0700)]
xfs: move the xfs_can_free_eofblocks call under the IOLOCK

In xfs_inode_free_eofblocks, move the xfs_can_free_eofblocks call
further down in the function to the point where we have taken the
IOLOCK.  This is preparation for the next patch, where we will need that
lock (or equivalent) so that we can check if there are any post-eof
blocks to clean out.

Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
3 years agoxfs: precalculate default inode attribute offset
Dave Chinner [Tue, 6 Apr 2021 14:03:24 +0000 (07:03 -0700)]
xfs: precalculate default inode attribute offset

Default attr fork offset is based on inode size, so is a fixed
geometry parameter of the inode. Move it to the xfs_ino_geometry
structure and stop calculating it on every call to
xfs_default_attroffset().

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Tested-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
3 years agoxfs: default attr fork size does not handle device inodes
Dave Chinner [Tue, 6 Apr 2021 14:02:04 +0000 (07:02 -0700)]
xfs: default attr fork size does not handle device inodes

Device inodes have a non-default data fork size of 8 bytes
as checked/enforced by xfs_repair. xfs_default_attroffset() doesn't
handle this, so lets do a minor refactor so it does.

Fixes: e6a688c33238 ("xfs: initialise attr fork on inode create")
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Tested-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
3 years agoxfs: inode fork allocation depends on XFS_IFEXTENT flag
Dave Chinner [Tue, 6 Apr 2021 14:01:41 +0000 (07:01 -0700)]
xfs: inode fork allocation depends on XFS_IFEXTENT flag

Due to confusion on when the XFS_IFEXTENT needs to be set, the
changes in e6a688c33238 ("xfs: initialise attr fork on inode
create") failed to set the flag when initialising the empty
attribute fork at inode creation. Set this flag the same way
xfs_bmap_add_attrfork() does after attry fork allocation.

Fixes: e6a688c33238 ("xfs: initialise attr fork on inode create")
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Tested-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
3 years agoxfs: eager inode attr fork init needs attr feature awareness
Dave Chinner [Tue, 6 Apr 2021 14:01:00 +0000 (07:01 -0700)]
xfs: eager inode attr fork init needs attr feature awareness

The pitfalls of regression testing on a machine without realising
that selinux was disabled. Only set the attr fork during inode
allocation if the attr feature bits are already set on the
superblock.

Fixes: e6a688c33238 ("xfs: initialise attr fork on inode create")
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Tested-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
3 years agoxfs: scrub: Disable check for unoptimized data fork bmbt node
Chandan Babu R [Tue, 6 Apr 2021 13:59:18 +0000 (06:59 -0700)]
xfs: scrub: Disable check for unoptimized data fork bmbt node

xchk_btree_check_minrecs() checks if the contents of the immediate child of a
bmbt root block can fit within the root block. This check could fail on inodes
with an attr fork since xfs_bmap_add_attrfork_btree() used to demote the
current root node of the data fork as the child of a newly allocated root node
if it found that the size of "struct xfs_btree_block" along with the space
required for records exceeded that of space available in the data fork.

xfs_bmap_add_attrfork_btree() should have used "struct xfs_bmdr_block" instead
of "struct xfs_btree_block" for the above mentioned space requirement
calculation. This commit disables the check for unoptimized (in terms of
disk space usage) data fork bmbt trees since there could be filesystems
in use that already have such a layout.

Suggested-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Chandan Babu R <chandanrlinux@gmail.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
3 years agoxfs: Use struct xfs_bmdr_block instead of struct xfs_btree_block to calculate root...
Chandan Babu R [Fri, 2 Apr 2021 22:07:33 +0000 (15:07 -0700)]
xfs: Use struct xfs_bmdr_block instead of struct xfs_btree_block to calculate root node size

The incore data fork of an inode stores the bmap btree root node as 'struct
xfs_btree_block'. However, the ondisk version of the inode stores the bmap
btree root node as a 'struct xfs_bmdr_block'.

xfs_bmap_add_attrfork_btree() checks if the btree root node fits inside the
data fork of the inode. However, it incorrectly uses 'struct xfs_btree_block'
to compute the size of the bmap btree root node. Since size of 'struct
xfs_btree_block' is larger than that of 'struct xfs_bmdr_block',
xfs_bmap_add_attrfork_btree() could end up unnecessarily demoting the current
root node as the child of newly allocated root node.

This commit optimizes space usage by modifying xfs_bmap_add_attrfork_btree()
to use 'struct xfs_bmdr_block' to check if the bmap btree root node fits
inside the data fork of the inode.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Chandan Babu R <chandanrlinux@gmail.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: deprecate BMV_IF_NO_DMAPI_READ flag
Anthony Iliopoulos [Wed, 31 Mar 2021 20:56:47 +0000 (13:56 -0700)]
xfs: deprecate BMV_IF_NO_DMAPI_READ flag

Use of the flag has had no effect since kernel commit 288699fecaff
("xfs: drop dmapi hooks"), which removed all dmapi related code, so
deprecate it.

Signed-off-by: Anthony Iliopoulos <ailiop@suse.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: merge _xfs_dic2xflags into xfs_ip2xflags
Christoph Hellwig [Mon, 29 Mar 2021 18:11:46 +0000 (11:11 -0700)]
xfs: merge _xfs_dic2xflags into xfs_ip2xflags

Merge _xfs_dic2xflags into its only caller.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: move the di_crtime field to struct xfs_inode
Christoph Hellwig [Mon, 29 Mar 2021 18:11:45 +0000 (11:11 -0700)]
xfs: move the di_crtime field to struct xfs_inode

Move the crtime field from struct xfs_icdinode into stuct xfs_inode and
remove the now entirely unused struct xfs_icdinode.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: move the di_flags2 field to struct xfs_inode
Christoph Hellwig [Mon, 29 Mar 2021 18:11:45 +0000 (11:11 -0700)]
xfs: move the di_flags2 field to struct xfs_inode

In preparation of removing the historic icinode struct, move the flags2
field into the containing xfs_inode structure.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: move the di_flags field to struct xfs_inode
Christoph Hellwig [Mon, 29 Mar 2021 18:11:44 +0000 (11:11 -0700)]
xfs: move the di_flags field to struct xfs_inode

In preparation of removing the historic icinode struct, move the flags
field into the containing xfs_inode structure.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: move the di_forkoff field to struct xfs_inode
Christoph Hellwig [Mon, 29 Mar 2021 18:11:44 +0000 (11:11 -0700)]
xfs: move the di_forkoff field to struct xfs_inode

In preparation of removing the historic icinode struct, move the
forkoff field into the containing xfs_inode structure.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: use a union for i_cowextsize and i_flushiter
Christoph Hellwig [Mon, 29 Mar 2021 18:11:43 +0000 (11:11 -0700)]
xfs: use a union for i_cowextsize and i_flushiter

The i_cowextsize field is only used for v3 inodes, and the i_flushiter
field is only used for v1/v2 inodes.  Use a union to pack the inode a
littler better after adding a few missing guards around their usage.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: use XFS_B_TO_FSB in xfs_ioctl_setattr
Christoph Hellwig [Mon, 29 Mar 2021 18:11:43 +0000 (11:11 -0700)]
xfs: use XFS_B_TO_FSB in xfs_ioctl_setattr

Clean up xfs_ioctl_setattr a bit by using XFS_B_TO_FSB.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: cleanup xfs_fill_fsxattr
Christoph Hellwig [Mon, 29 Mar 2021 18:11:42 +0000 (11:11 -0700)]
xfs: cleanup xfs_fill_fsxattr

Add a local xfs_mount variable, and use the XFS_FSB_TO_B helper.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: move the di_flushiter field to struct xfs_inode
Christoph Hellwig [Mon, 29 Mar 2021 18:11:42 +0000 (11:11 -0700)]
xfs: move the di_flushiter field to struct xfs_inode

In preparation of removing the historic icinode struct, move the
flushiter field into the containing xfs_inode structure.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: move the di_cowextsize field to struct xfs_inode
Christoph Hellwig [Mon, 29 Mar 2021 18:11:42 +0000 (11:11 -0700)]
xfs: move the di_cowextsize field to struct xfs_inode

In preparation of removing the historic icinode struct, move the
cowextsize field into the containing xfs_inode structure.  Also
switch to use the xfs_extlen_t instead of a uint32_t.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: move the di_extsize field to struct xfs_inode
Christoph Hellwig [Mon, 29 Mar 2021 18:11:41 +0000 (11:11 -0700)]
xfs: move the di_extsize field to struct xfs_inode

In preparation of removing the historic icinode struct, move the extsize
field into the containing xfs_inode structure.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: move the di_nblocks field to struct xfs_inode
Christoph Hellwig [Mon, 29 Mar 2021 18:11:40 +0000 (11:11 -0700)]
xfs: move the di_nblocks field to struct xfs_inode

In preparation of removing the historic icinode struct, move the nblocks
field into the containing xfs_inode structure.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: move the di_size field to struct xfs_inode
Christoph Hellwig [Mon, 29 Mar 2021 18:11:40 +0000 (11:11 -0700)]
xfs: move the di_size field to struct xfs_inode

In preparation of removing the historic icinode struct, move the on-disk
size field into the containing xfs_inode structure.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: move the di_projid field to struct xfs_inode
Christoph Hellwig [Mon, 29 Mar 2021 18:11:39 +0000 (11:11 -0700)]
xfs: move the di_projid field to struct xfs_inode

In preparation of removing the historic icinode struct, move the projid
field into the containing xfs_inode structure.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: don't clear the "dinode core" in xfs_inode_alloc
Christoph Hellwig [Mon, 29 Mar 2021 18:11:39 +0000 (11:11 -0700)]
xfs: don't clear the "dinode core" in xfs_inode_alloc

The xfs_icdinode structure just contains a random mix of inode field,
which are all read from the on-disk inode and mostly not looked at
before reading the inode or initializing a new inode cluster.  The
only exceptions are the forkoff and blocks field, which are used
in sanity checks for freshly allocated inodes.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: remove the di_dmevmask and di_dmstate fields from struct xfs_icdinode
Christoph Hellwig [Mon, 29 Mar 2021 18:11:38 +0000 (11:11 -0700)]
xfs: remove the di_dmevmask and di_dmstate fields from struct xfs_icdinode

The legacy DMAPI fields were never set by upstream Linux XFS, and have no
way to be read using the kernel APIs.  So instead of bloating the in-core
inode for them just copy them from the on-disk inode into the log when
logging the inode.  The only caveat is that we need to make sure to zero
the fields for newly read or deleted inodes, which is solved using a new
flag in the inode.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: remove the unused xfs_icdinode_has_bigtime helper
Christoph Hellwig [Mon, 29 Mar 2021 18:11:38 +0000 (11:11 -0700)]
xfs: remove the unused xfs_icdinode_has_bigtime helper

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: handle crtime more carefully in xfs_bulkstat_one_int
Christoph Hellwig [Mon, 29 Mar 2021 18:11:37 +0000 (11:11 -0700)]
xfs: handle crtime more carefully in xfs_bulkstat_one_int

The crtime only exists for v5 inodes, so only copy it for those.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: consistently initialize di_flags2
Christoph Hellwig [Mon, 29 Mar 2021 18:11:37 +0000 (11:11 -0700)]
xfs: consistently initialize di_flags2

Make sure di_flags2 is always initialized.  We currently get this implicitly
by clearing the dinode core on allocating the in-core inode, but that is
about to go away.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: split xfs_imap_to_bp
Christoph Hellwig [Mon, 29 Mar 2021 18:11:37 +0000 (11:11 -0700)]
xfs: split xfs_imap_to_bp

Split looking up the dinode from xfs_imap_to_bp, which can be
significantly simplified as a result.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: scrub: Remove incorrect check executed on block format directories
Chandan Babu R [Fri, 26 Mar 2021 17:52:56 +0000 (10:52 -0700)]
xfs: scrub: Remove incorrect check executed on block format directories

A directory with one directory block which in turns consists of two or more fs
blocks is incorrectly flagged as corrupt by scrub since it assumes that
"Block" format directories have a data fork single extent spanning the file
offset range of [0, Dir block size - 1].

This commit fixes the bug by removing the incorrect check.

Signed-off-by: Chandan Babu R <chandanrlinux@gmail.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: Initialize xfs_alloc_arg->total correctly when allocating minlen extents
Chandan Babu R [Thu, 25 Mar 2021 18:55:10 +0000 (11:55 -0700)]
xfs: Initialize xfs_alloc_arg->total correctly when allocating minlen extents

xfs/538 can cause the following call trace to be printed when executing on a
multi-block directory configuration,

 WARNING: CPU: 1 PID: 2578 at fs/xfs/libxfs/xfs_bmap.c:717 xfs_bmap_extents_to_btree+0x520/0x5d0
 Call Trace:
  ? xfs_buf_rele+0x4f/0x450
  xfs_bmap_add_extent_hole_real+0x747/0x960
  xfs_bmapi_allocate+0x39a/0x440
  xfs_bmapi_write+0x507/0x9e0
  xfs_da_grow_inode_int+0x1cd/0x330
  ? up+0x12/0x60
  xfs_dir2_grow_inode+0x62/0x110
  ? xfs_trans_log_inode+0x234/0x2d0
  xfs_dir2_sf_to_block+0x103/0x940
  ? xfs_dir2_sf_check+0x8c/0x210
  ? xfs_da_compname+0x19/0x30
  ? xfs_dir2_sf_lookup+0xd0/0x3d0
  xfs_dir2_sf_addname+0x10d/0x910
  xfs_dir_createname+0x1ad/0x210
  xfs_create+0x404/0x620
  xfs_generic_create+0x24c/0x320
  path_openat+0xda6/0x1030
  do_filp_open+0x88/0x130
  ? kmem_cache_alloc+0x50/0x210
  ? __cond_resched+0x16/0x40
  ? kmem_cache_alloc+0x50/0x210
  do_sys_openat2+0x97/0x150
  __x64_sys_creat+0x49/0x70
  do_syscall_64+0x33/0x40
  entry_SYSCALL_64_after_hwframe+0x44/0xae

This occurs because xfs_bmap_exact_minlen_extent_alloc() initializes
xfs_alloc_arg->total to xfs_bmalloca->minlen. In the context of
xfs_bmap_exact_minlen_extent_alloc(), xfs_bmalloca->minlen has a value of 1
and hence the space allocator could choose an AG which has less than
xfs_bmalloca->total number of free blocks available. As the transaction
proceeds, one of the future space allocation requests could fail due to
non-availability of free blocks in the AG that was originally chosen.

This commit fixes the bug by assigning xfs_alloc_arg->total to the value of
xfs_bmalloca->total.

Fixes: 301519674699 ("xfs: Introduce error injection to allocate only minlen size extents for files")
Signed-off-by: Chandan Babu R <chandanrlinux@gmail.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
3 years agoxfs: Fix dax inode extent calculation when direct write is performed on an unwritten...
Chandan Babu R [Thu, 25 Mar 2021 18:48:18 +0000 (11:48 -0700)]
xfs: Fix dax inode extent calculation when direct write is performed on an unwritten extent

With dax enabled filesystems, a direct write operation into an existing
unwritten extent results in xfs_iomap_write_direct() zero-ing and converting
the extent into a normal extent before the actual data is copied from the
userspace buffer.

The inode extent count can increase by 2 if the extent range being written to
maps to the middle of the existing unwritten extent range. Hence this commit
uses XFS_IEXT_WRITE_UNWRITTEN_CNT as the extent count delta when such a write
operation is being performed.

Fixes: 727e1acd297c ("xfs: Check for extent overflow when trivally adding a new extent")
Reported-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Chandan Babu R <chandanrlinux@gmail.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
3 years agoxfs: fix xfs_trans slab cache name
Anthony Iliopoulos [Thu, 25 Mar 2021 18:47:05 +0000 (11:47 -0700)]
xfs: fix xfs_trans slab cache name

Removal of kmem_zone_init wrappers accidentally changed a slab cache
name from "xfs_trans" to "xf_trans". Fix this so that userspace
consumers of /proc/slabinfo and /sys/kernel/slab can find it again.

Fixes: b1231760e443 ("xfs: Remove slab init wrappers")
Signed-off-by: Anthony Iliopoulos <ailiop@suse.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: add error injection for per-AG resv failure
Gao Xiang [Wed, 24 Mar 2021 02:05:39 +0000 (19:05 -0700)]
xfs: add error injection for per-AG resv failure

per-AG resv failure after fixing up freespace is hard to test in an
effective way, so directly add an error injection path to observe
such error handling path works as expected.

Signed-off-by: Gao Xiang <hsiangkao@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: support shrinking unused space in the last AG
Gao Xiang [Wed, 24 Mar 2021 02:05:39 +0000 (19:05 -0700)]
xfs: support shrinking unused space in the last AG

As the first step of shrinking, this attempts to enable shrinking
unused space in the last allocation group by fixing up freespace
btree, agi, agf and adjusting super block and use a helper
xfs_ag_shrink_space() to fixup the last AG.

This can be all done in one transaction for now, so I think no
additional protection is needed.

Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Gao Xiang <hsiangkao@redhat.com>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: introduce xfs_ag_shrink_space()
Gao Xiang [Wed, 24 Mar 2021 02:05:38 +0000 (19:05 -0700)]
xfs: introduce xfs_ag_shrink_space()

This patch introduces a helper to shrink unused space in the last AG
by fixing up the freespace btree.

Also make sure that the per-AG reservation works under the new AG
size. If such per-AG reservation or extent allocation fails, roll
the transaction so the new transaction could cancel without any side
effects.

Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Gao Xiang <hsiangkao@redhat.com>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: hoist out xfs_resizefs_init_new_ags()
Gao Xiang [Wed, 24 Mar 2021 02:05:38 +0000 (19:05 -0700)]
xfs: hoist out xfs_resizefs_init_new_ags()

Move out related logic for initializing new added AGs to a new helper
in preparation for shrinking. No logic changes.

Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Gao Xiang <hsiangkao@redhat.com>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: update lazy sb counters immediately for resizefs
Gao Xiang [Wed, 24 Mar 2021 02:05:37 +0000 (19:05 -0700)]
xfs: update lazy sb counters immediately for resizefs

sb_fdblocks will be updated lazily if lazysbcount is enabled,
therefore when shrinking the filesystem sb_fdblocks could be
larger than sb_dblocks and xfs_validate_sb_write() would fail.

Even for growfs case, it'd be better to update lazy sb counters
immediately to reflect the real sb counters.

Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Gao Xiang <hsiangkao@redhat.com>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: Fix a typo
Bhaskar Chowdhury [Tue, 23 Mar 2021 23:59:30 +0000 (16:59 -0700)]
xfs: Fix a typo

s/strutures/structures/

Signed-off-by: Bhaskar Chowdhury <unixbhaskar@gmail.com>
Reviewed-by: Pavel Reichl <preichl@redhat.com>
Acked-by: Randy Dunlap <rdunlap@infradead.org>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: Rudimentary spelling fix
Bhaskar Chowdhury [Tue, 23 Mar 2021 23:59:30 +0000 (16:59 -0700)]
xfs: Rudimentary spelling fix

s/sytemcall/syscall/

Signed-off-by: Bhaskar Chowdhury <unixbhaskar@gmail.com>
Acked-by: Randy Dunlap <rdunlap@infradead.org>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: Rudimentary typo fixes
Bhaskar Chowdhury [Tue, 23 Mar 2021 23:59:30 +0000 (16:59 -0700)]
xfs: Rudimentary typo fixes

s/filesytem/filesystem/
s/instrumention/instrumentation/

Signed-off-by: Bhaskar Chowdhury <unixbhaskar@gmail.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: __percpu_counter_compare() inode count debug too expensive
Dave Chinner [Mon, 22 Mar 2021 16:52:06 +0000 (09:52 -0700)]
xfs: __percpu_counter_compare() inode count debug too expensive

- 21.92% __xfs_trans_commit
     - 21.62% xfs_log_commit_cil
- 11.69% xfs_trans_unreserve_and_mod_sb
   - 11.58% __percpu_counter_compare
      - 11.45% __percpu_counter_sum
 - 10.29% _raw_spin_lock_irqsave
    - 10.28% do_raw_spin_lock
 __pv_queued_spin_lock_slowpath

We debated just getting rid of it last time this came up and
there was no real objection to removing it. Now it's the biggest
scalability limitation for debug kernels even on smallish machines,
so let's just get rid of it.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: reduce debug overhead of dir leaf/node checks
Dave Chinner [Mon, 22 Mar 2021 16:52:05 +0000 (09:52 -0700)]
xfs: reduce debug overhead of dir leaf/node checks

On debug kernels, we call xfs_dir3_leaf_check_int() multiple times
on every directory modification. The robust hash ordering checks it
does on every entry in the leaf on every call results in a massive
CPU overhead which slows down debug kernels by a large amount.

We use xfs_dir3_leaf_check_int() for the verifiers as well, so we
can't just gut the function to reduce overhead. What we can do,
however, is reduce the work it does when it is called from the
debug interfaces, just leaving the high level checks in place and
leaving the robust validation to the verifiers. This means the debug
checks will catch gross errors, but subtle bugs might not be caught
until a verifier is run.

It is easy enough to restore the existing debug behaviour if the
developer needs it (just change a call parameter in the debug code),
but overwise the overhead makes testing large directory block sizes
on debug kernels very slow.

Profile at an unlink rate of ~80k file/s on a 64k block size
filesystem before the patch:

  40.30%  [kernel]  [k] xfs_dir3_leaf_check_int
  10.98%  [kernel]  [k] __xfs_dir3_data_check
   8.10%  [kernel]  [k] xfs_verify_dir_ino
   4.42%  [kernel]  [k] memcpy
   2.22%  [kernel]  [k] xfs_dir2_data_get_ftype
   1.52%  [kernel]  [k] do_raw_spin_lock

Profile after, at an unlink rate of ~125k files/s (+50% improvement)
has largely dropped the leaf verification debug overhead out of the
profile.

  16.53%  [kernel]  [k] __xfs_dir3_data_check
  12.53%  [kernel]  [k] xfs_verify_dir_ino
   7.97%  [kernel]  [k] memcpy
   3.36%  [kernel]  [k] xfs_dir2_data_get_ftype
   2.86%  [kernel]  [k] __pv_queued_spin_lock_slowpath

Create shows a similar change in profile and a +25% improvement in
performance.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: No need for inode number error injection in __xfs_dir3_data_check
Dave Chinner [Mon, 22 Mar 2021 16:52:05 +0000 (09:52 -0700)]
xfs: No need for inode number error injection in __xfs_dir3_data_check

We call xfs_dir_ino_validate() for every dir entry in a directory
when doing validity checking of the directory. It calls
xfs_verify_dir_ino() then emits a corruption report if bad or does
error injection if good. It is extremely costly:

  43.27%  [kernel]  [k] xfs_dir3_leaf_check_int
  10.28%  [kernel]  [k] __xfs_dir3_data_check
   6.61%  [kernel]  [k] xfs_verify_dir_ino
   4.16%  [kernel]  [k] xfs_errortag_test
   4.00%  [kernel]  [k] memcpy
   3.48%  [kernel]  [k] xfs_dir_ino_validate

7% of the cpu usage in this directory traversal workload is
xfs_dir_ino_validate() doing absolutely nothing.

We don't need error injection to simulate a bad inode numbers in the
directory structure because we can do that by fuzzing the structure
on disk.

And we don't need a corruption report, because the
__xfs_dir3_data_check() will emit one if the inode number is bad.

So just call xfs_verify_dir_ino() directly here, and get rid of all
this unnecessary overhead:

  40.30%  [kernel]  [k] xfs_dir3_leaf_check_int
  10.98%  [kernel]  [k] __xfs_dir3_data_check
   8.10%  [kernel]  [k] xfs_verify_dir_ino
   4.42%  [kernel]  [k] memcpy
   2.22%  [kernel]  [k] xfs_dir2_data_get_ftype
   1.52%  [kernel]  [k] do_raw_spin_lock

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: type verification is expensive
Dave Chinner [Mon, 22 Mar 2021 16:52:05 +0000 (09:52 -0700)]
xfs: type verification is expensive

From a concurrent rm -rf workload:

  41.04%  [kernel]  [k] xfs_dir3_leaf_check_int
   9.85%  [kernel]  [k] __xfs_dir3_data_check
   5.60%  [kernel]  [k] xfs_verify_ino
   5.32%  [kernel]  [k] xfs_agino_range
   4.21%  [kernel]  [k] memcpy
   3.06%  [kernel]  [k] xfs_errortag_test
   2.57%  [kernel]  [k] xfs_dir_ino_validate
   1.66%  [kernel]  [k] xfs_dir2_data_get_ftype
   1.17%  [kernel]  [k] do_raw_spin_lock
   1.11%  [kernel]  [k] xfs_verify_dir_ino
   0.84%  [kernel]  [k] __raw_callee_save___pv_queued_spin_unlock
   0.83%  [kernel]  [k] xfs_buf_find
   0.64%  [kernel]  [k] xfs_log_commit_cil

THere's an awful lot of overhead in just range checking inode
numbers in that, but each inode number check is not a lot of code.
The total is a bit over 14.5% of the CPU time is spent validating
inode numbers.

The problem is that they deeply nested global scope functions so the
overhead here is all in function call marshalling.

   text    data     bss     dec     hex filename
   2077       0       0    2077     81d fs/xfs/libxfs/xfs_types.o.orig
   2197       0       0    2197     895 fs/xfs/libxfs/xfs_types.o

There's a small increase in binary size by inlining all the local
nested calls in the verifier functions, but the same workload now
profiles as:

  40.69%  [kernel]  [k] xfs_dir3_leaf_check_int
  10.52%  [kernel]  [k] __xfs_dir3_data_check
   6.68%  [kernel]  [k] xfs_verify_dir_ino
   4.22%  [kernel]  [k] xfs_errortag_test
   4.15%  [kernel]  [k] memcpy
   3.53%  [kernel]  [k] xfs_dir_ino_validate
   1.87%  [kernel]  [k] xfs_dir2_data_get_ftype
   1.37%  [kernel]  [k] do_raw_spin_lock
   0.98%  [kernel]  [k] xfs_buf_find
   0.94%  [kernel]  [k] __raw_callee_save___pv_queued_spin_unlock
   0.73%  [kernel]  [k] xfs_log_commit_cil

Now we only spend just over 10% of the time validing inode numbers
for the same workload. Hence a few "inline" keyworks is good enough
to reduce the validation overhead by 30%...

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: optimise xfs_buf_item_size/format for contiguous regions
Dave Chinner [Mon, 22 Mar 2021 16:52:04 +0000 (09:52 -0700)]
xfs: optimise xfs_buf_item_size/format for contiguous regions

We process the buf_log_item bitmap one set bit at a time with
xfs_next_bit() so we can detect if a region crosses a memcpy
discontinuity in the buffer data address. This has massive overhead
on large buffers (e.g. 64k directory blocks) because we do a lot of
unnecessary checks and xfs_buf_offset() calls.

For example, 16-way concurrent create workload on debug kernel
running CPU bound has this at the top of the profile at ~120k
create/s on 64kb directory block size:

  20.66%  [kernel]  [k] xfs_dir3_leaf_check_int
   7.10%  [kernel]  [k] memcpy
   6.22%  [kernel]  [k] xfs_next_bit
   3.55%  [kernel]  [k] xfs_buf_offset
   3.53%  [kernel]  [k] xfs_buf_item_format
   3.34%  [kernel]  [k] __pv_queued_spin_lock_slowpath
   3.04%  [kernel]  [k] do_raw_spin_lock
   2.84%  [kernel]  [k] xfs_buf_item_size_segment.isra.0
   2.31%  [kernel]  [k] __raw_callee_save___pv_queued_spin_unlock
   1.36%  [kernel]  [k] xfs_log_commit_cil

(debug checks hurt large blocks)

The only buffers with discontinuities in the data address are
unmapped buffers, and they are only used for inode cluster buffers
and only for logging unlinked pointers. IOWs, it is -rare- that we
even need to detect a discontinuity in the buffer item formatting
code.

Optimise all this by using xfs_contig_bits() to find the size of
the contiguous regions, then test for a discontiunity inside it. If
we find one, do the slow "bit at a time" method we do now. If we
don't, then just copy the entire contiguous range in one go.

Profile now looks like:

  25.26%  [kernel]  [k] xfs_dir3_leaf_check_int
   9.25%  [kernel]  [k] memcpy
   5.01%  [kernel]  [k] __pv_queued_spin_lock_slowpath
   2.84%  [kernel]  [k] do_raw_spin_lock
   2.22%  [kernel]  [k] __raw_callee_save___pv_queued_spin_unlock
   1.88%  [kernel]  [k] xfs_buf_find
   1.53%  [kernel]  [k] memmove
   1.47%  [kernel]  [k] xfs_log_commit_cil
....
   0.34%  [kernel]  [k] xfs_buf_item_format
....
   0.21%  [kernel]  [k] xfs_buf_offset
....
   0.16%  [kernel]  [k] xfs_contig_bits
....
   0.13%  [kernel]  [k] xfs_buf_item_size_segment.isra.0

So the bit scanning over for the dirty region tracking for the
buffer log items is basically gone. Debug overhead hurts even more
now...

Perf comparison

dir block  creates unlink
size (kb) time rate time

Original  4 4m08s 220k  5m13s
Original 64 7m21s 115k 13m25s
Patched  4 3m59s 230k  5m03s
Patched 64 6m23s 143k 12m33s

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: xfs_buf_item_size_segment() needs to pass segment offset
Dave Chinner [Mon, 22 Mar 2021 16:52:04 +0000 (09:52 -0700)]
xfs: xfs_buf_item_size_segment() needs to pass segment offset

Otherwise it doesn't correctly calculate the number of vectors
in a logged buffer that has a contiguous map that gets split into
multiple regions because the range spans discontigous memory.

Probably never been hit in practice - we don't log contiguous ranges
on unmapped buffers (inode clusters).

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: reduce buffer log item shadow allocations
Dave Chinner [Mon, 22 Mar 2021 16:52:03 +0000 (09:52 -0700)]
xfs: reduce buffer log item shadow allocations

When we modify btrees repeatedly, we regularly increase the size of
the logged region by a single chunk at a time (per transaction
commit). This results in the CIL formatting code having to
reallocate the log vector buffer every time the buffer dirty region
grows. Hence over a typical 4kB btree buffer, we might grow the log
vector 4096/128 = 32x over a short period where we repeatedly add
or remove records to/from the buffer over a series of running
transaction. This means we are doing 32 memory allocations and frees
over this time during a performance critical path in the journal.

The amount of space tracked in the CIL for the object is calculated
during the ->iop_format() call for the buffer log item, but the
buffer memory allocated for it is calculated by the ->iop_size()
call. The size callout determines the size of the buffer, the format
call determines the space used in the buffer.

Hence we can oversize the buffer space required in the size
calculation without impacting the amount of space used and accounted
to the CIL for the changes being logged. This allows us to reduce
the number of allocations by rounding up the buffer size to allow
for future growth. This can safe a substantial amount of CPU time in
this path:

-   46.52%     2.02%  [kernel]                  [k] xfs_log_commit_cil
   - 44.49% xfs_log_commit_cil
      - 30.78% _raw_spin_lock
         - 30.75% do_raw_spin_lock
              30.27% __pv_queued_spin_lock_slowpath

(oh, ouch!)
....
      - 1.05% kmem_alloc_large
         - 1.02% kmem_alloc
              0.94% __kmalloc

This overhead here us what this patch is aimed at. After:

      - 0.76% kmem_alloc_large
         - 0.75% kmem_alloc
              0.70% __kmalloc

The size of 512 bytes is based on the bitmap chunk size being 128
bytes and that random directory entry updates almost never require
more than 3-4 128 byte regions to be logged in the directory block.

The other observation is for per-ag btrees. When we are inserting
into a new btree block, we'll pack it from the front. Hence the
first few records land in the first 128 bytes so we log only 128
bytes, the next 8-16 records land in the second region so now we log
256 bytes. And so on.  If we are doing random updates, it will only
allocate every 4 random 128 byte regions that are dirtied instead of
every single one.

Any larger than 512 bytes and I noticed an increase in memory
footprint in my scalability workloads. Any less than this and I
didn't really see any significant benefit to CPU usage.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Chandan Babu R <chandanrlinux@gmail.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Gao Xiang <hsiangkao@redhat.com>
3 years agoxfs: initialise attr fork on inode create
Dave Chinner [Mon, 22 Mar 2021 16:52:03 +0000 (09:52 -0700)]
xfs: initialise attr fork on inode create

When we allocate a new inode, we often need to add an attribute to
the inode as part of the create. This can happen as a result of
needing to add default ACLs or security labels before the inode is
made visible to userspace.

This is highly inefficient right now. We do the create transaction
to allocate the inode, then we do an "add attr fork" transaction to
modify the just created empty inode to set the inode fork offset to
allow attributes to be stored, then we go and do the attribute
creation.

This means 3 transactions instead of 1 to allocate an inode, and
this greatly increases the load on the CIL commit code, resulting in
excessive contention on the CIL spin locks and performance
degradation:

 18.99%  [kernel]                [k] __pv_queued_spin_lock_slowpath
  3.57%  [kernel]                [k] do_raw_spin_lock
  2.51%  [kernel]                [k] __raw_callee_save___pv_queued_spin_unlock
  2.48%  [kernel]                [k] memcpy
  2.34%  [kernel]                [k] xfs_log_commit_cil

The typical profile resulting from running fsmark on a selinux enabled
filesytem is adds this overhead to the create path:

  - 15.30% xfs_init_security
     - 15.23% security_inode_init_security
- 13.05% xfs_initxattrs
   - 12.94% xfs_attr_set
      - 6.75% xfs_bmap_add_attrfork
 - 5.51% xfs_trans_commit
    - 5.48% __xfs_trans_commit
       - 5.35% xfs_log_commit_cil
  - 3.86% _raw_spin_lock
     - do_raw_spin_lock
  __pv_queued_spin_lock_slowpath
 - 0.70% xfs_trans_alloc
      0.52% xfs_trans_reserve
      - 5.41% xfs_attr_set_args
 - 5.39% xfs_attr_set_shortform.constprop.0
    - 4.46% xfs_trans_commit
       - 4.46% __xfs_trans_commit
  - 4.33% xfs_log_commit_cil
     - 2.74% _raw_spin_lock
- do_raw_spin_lock
     __pv_queued_spin_lock_slowpath
       0.60% xfs_inode_item_format
      0.90% xfs_attr_try_sf_addname
- 1.99% selinux_inode_init_security
   - 1.02% security_sid_to_context_force
      - 1.00% security_sid_to_context_core
 - 0.92% sidtab_entry_to_string
    - 0.90% sidtab_sid2str_get
 0.59% sidtab_sid2str_put.part.0
   - 0.82% selinux_determine_inode_label
      - 0.77% security_transition_sid
   0.70% security_compute_sid.part.0

And fsmark creation rate performance drops by ~25%. The key point to
note here is that half the additional overhead comes from adding the
attribute fork to the newly created inode. That's crazy, considering
we can do this same thing at inode create time with a couple of
lines of code and no extra overhead.

So, if we know we are going to add an attribute immediately after
creating the inode, let's just initialise the attribute fork inside
the create transaction and chop that whole chunk of code out of
the create fast path. This completely removes the performance
drop caused by enabling SELinux, and the profile looks like:

     - 8.99% xfs_init_security
         - 9.00% security_inode_init_security
            - 6.43% xfs_initxattrs
               - 6.37% xfs_attr_set
                  - 5.45% xfs_attr_set_args
                     - 5.42% xfs_attr_set_shortform.constprop.0
                        - 4.51% xfs_trans_commit
                           - 4.54% __xfs_trans_commit
                              - 4.59% xfs_log_commit_cil
                                 - 2.67% _raw_spin_lock
                                    - 3.28% do_raw_spin_lock
                                         3.08% __pv_queued_spin_lock_slowpath
                                   0.66% xfs_inode_item_format
                        - 0.90% xfs_attr_try_sf_addname
                  - 0.60% xfs_trans_alloc
            - 2.35% selinux_inode_init_security
               - 1.25% security_sid_to_context_force
                  - 1.21% security_sid_to_context_core
                     - 1.19% sidtab_entry_to_string
                        - 1.20% sidtab_sid2str_get
                           - 0.86% sidtab_sid2str_put.part.0
                              - 0.62% _raw_spin_lock_irqsave
                                 - 0.77% do_raw_spin_lock
                                      __pv_queued_spin_lock_slowpath
               - 0.84% selinux_determine_inode_label
                  - 0.83% security_transition_sid
                       0.86% security_compute_sid.part.0

Which indicates the XFS overhead of creating the selinux xattr has
been halved. This doesn't fix the CIL lock contention problem, just
means it's not a limiting factor for this workload. Lock contention
in the security subsystems is going to be an issue soon, though...

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
[djwong: fix compilation error when CONFIG_SECURITY=n]
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Gao Xiang <hsiangkao@redhat.com>
3 years agoxfs: ensure xfs_errortag_random_default matches XFS_ERRTAG_MAX
Gao Xiang [Mon, 22 Mar 2021 16:52:02 +0000 (09:52 -0700)]
xfs: ensure xfs_errortag_random_default matches XFS_ERRTAG_MAX

Add the BUILD_BUG_ON to xfs_errortag_add() in order to make sure that
the length of xfs_errortag_random_default matches XFS_ERRTAG_MAX when
building.

Signed-off-by: Gao Xiang <hsiangkao@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: Skip repetitive warnings about mount options
Pavel Reichl [Mon, 22 Mar 2021 16:52:02 +0000 (09:52 -0700)]
xfs: Skip repetitive warnings about mount options

Skip the warnings about mount option being deprecated if we are
remounting and deprecated option state is not changing.

Bug: https://bugzilla.kernel.org/show_bug.cgi?id=211605
Fix-suggested-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: Pavel Reichl <preichl@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Carlos Maiolino <cmaiolino@redhat.com>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: rename variable mp to parsing_mp
Pavel Reichl [Mon, 22 Mar 2021 16:52:01 +0000 (09:52 -0700)]
xfs: rename variable mp to parsing_mp

Rename mp variable to parsisng_mp so it is easy to distinguish
between current mount point handle and handle for mount point
which mount options are being parsed.

Suggested-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: Pavel Reichl <preichl@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Carlos Maiolino <cmaiolino@redhat.com>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
3 years agoxfs: rename the blockgc workqueue
Darrick J. Wong [Mon, 22 Mar 2021 16:51:55 +0000 (09:51 -0700)]
xfs: rename the blockgc workqueue

Since we're about to start using the blockgc workqueue to dispose of
inactivated inodes, strip the "block" prefix from the name; now it's
merely the general garbage collection (gc) workqueue.

Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
3 years agoxfs: prevent metadata files from being inactivated
Darrick J. Wong [Mon, 22 Mar 2021 16:51:54 +0000 (09:51 -0700)]
xfs: prevent metadata files from being inactivated

Files containing metadata (quota records, rt bitmap and summary info)
are fully managed by the filesystem, which means that all resource
cleanup must be explicit, not automatic.  This means that they should
never be subjected automatic to post-eof truncation, nor should they be
freed automatically even if the link count drops to zero.

In other words, xfs_inactive() should leave these files alone.  Add the
necessary predicate functions to make this happen.  This adds a second
layer of prevention for the kinds of fs corruption that was fixed by
commit f4c32e87de7d.  If we ever decide to support removing metadata
files, we should make all those metadata updates explicit.

Rearrange the order of #includes to fix compiler errors, since
xfs_mount.h is supposed to be included before xfs_inode.h

Followup-to: f4c32e87de7d ("xfs: fix realtime bitmap/summary file truncation when growing rt volume")
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
3 years agoxfs: validate ag btree levels using the precomputed values
Darrick J. Wong [Mon, 22 Mar 2021 16:51:54 +0000 (09:51 -0700)]
xfs: validate ag btree levels using the precomputed values

Use the AG btree height limits that we precomputed into the xfs_mount to
validate the AG headers instead of using XFS_BTREE_MAXLEVELS.

Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
3 years agoxfs: remove return value from xchk_ag_btcur_init
Darrick J. Wong [Mon, 22 Mar 2021 16:51:53 +0000 (09:51 -0700)]
xfs: remove return value from xchk_ag_btcur_init

Functions called by this function cannot fail, so get rid of the return
and error checking.

Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
3 years agoxfs: set the scrub AG number in xchk_ag_read_headers
Darrick J. Wong [Mon, 22 Mar 2021 16:51:53 +0000 (09:51 -0700)]
xfs: set the scrub AG number in xchk_ag_read_headers

Since xchk_ag_read_headers initializes fields in struct xchk_ag, we
might as well set the AG number and save the callers the trouble.

Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
3 years agoxfs: mark a data structure sick if there are cross-referencing errors
Darrick J. Wong [Mon, 22 Mar 2021 16:51:52 +0000 (09:51 -0700)]
xfs: mark a data structure sick if there are cross-referencing errors

If scrub observes cross-referencing errors while scanning a data
structure, mark the data structure sick.  There's /something/
inconsistent, even if we can't really tell what it is.

Fixes: 4860a05d2475 ("xfs: scrub/repair should update filesystem metadata health")
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
3 years agoxfs: bail out of scrub immediately if scan incomplete
Darrick J. Wong [Mon, 22 Mar 2021 16:51:52 +0000 (09:51 -0700)]
xfs: bail out of scrub immediately if scan incomplete

If a scrubber cannot complete its check and signals an incomplete check,
we must bail out immediately without updating health status, trying a
repair, etc. because our scan is incomplete and we therefore do not know
much more.

Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
3 years agoxfs: fix dquot scrub loop cancellation
Darrick J. Wong [Mon, 22 Mar 2021 16:51:52 +0000 (09:51 -0700)]
xfs: fix dquot scrub loop cancellation

When xchk_quota_item figures out that it needs to terminate the scrub
operation, it needs to return some error code to abort the loop, but
instead it returns zero and the loop keeps running.  Fix this by making
it use ECANCELED, and fix the other loop bailout condition check at the
bottom too.

Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
3 years agoxfs: fix uninitialized variables in xrep_calc_ag_resblks
Darrick J. Wong [Mon, 22 Mar 2021 16:51:51 +0000 (09:51 -0700)]
xfs: fix uninitialized variables in xrep_calc_ag_resblks

If we can't read the AGF header, we never actually set a value for
freelen and usedlen.  These two variables are used to make the worst
case estimate of btree size, so it's safe to set them to the AG size as
a fallback.

Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
3 years agoxfs: drop freeze protection when running GETFSMAP
Darrick J. Wong [Mon, 22 Mar 2021 16:51:50 +0000 (09:51 -0700)]
xfs: drop freeze protection when running GETFSMAP

A recent log refactoring patchset from Brian Foster relaxed fsfreeze
behavior with regards to the buffer cache -- now freeze only waits for
pending buffer IO to finish, and does not try to drain the buffer cache
LRU.  As a result, fsfreeze should no longer stall indefinitely while
fsmap runs.  Drop the sb_start_write calls around fsmap invocations.

While we're cleaning things, add a comment to the xfs_trans_alloc_empty
call explaining why we're running around with empty transactions.

Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
3 years agoLinux 5.12-rc4
Linus Torvalds [Sun, 21 Mar 2021 21:56:43 +0000 (14:56 -0700)]
Linux 5.12-rc4

3 years agoMerge tag 'ext4_for_linus_stable' of git://git.kernel.org/pub/scm/linux/kernel/git...
Linus Torvalds [Sun, 21 Mar 2021 21:06:10 +0000 (14:06 -0700)]
Merge tag 'ext4_for_linus_stable' of git://git./linux/kernel/git/tytso/ext4

Pull ext4 fixes from Ted Ts'o:
 "Miscellaneous ext4 bug fixes for v5.12"

* tag 'ext4_for_linus_stable' of git://git.kernel.org/pub/scm/linux/kernel/git/tytso/ext4:
  ext4: initialize ret to suppress smatch warning
  ext4: stop inode update before return
  ext4: fix rename whiteout with fast commit
  ext4: fix timer use-after-free on failed mount
  ext4: fix potential error in ext4_do_update_inode
  ext4: do not try to set xattr into ea_inode if value is empty
  ext4: do not iput inode under running transaction in ext4_rename()
  ext4: find old entry again if failed to rename whiteout
  ext4: fix error handling in ext4_end_enable_verity()
  ext4: fix bh ref count on error paths
  fs/ext4: fix integer overflow in s_log_groups_per_flex
  ext4: add reclaim checks to xattr code
  ext4: shrink race window in ext4_should_retry_alloc()

3 years agoMerge tag 'io_uring-5.12-2021-03-21' of git://git.kernel.dk/linux-block
Linus Torvalds [Sun, 21 Mar 2021 19:25:54 +0000 (12:25 -0700)]
Merge tag 'io_uring-5.12-2021-03-21' of git://git.kernel.dk/linux-block

Pull io_uring followup fixes from Jens Axboe:

 - The SIGSTOP change from Eric, so we properly ignore that for
   PF_IO_WORKER threads.

 - Disallow sending signals to PF_IO_WORKER threads in general, we're
   not interested in having them funnel back to the io_uring owning
   task.

 - Stable fix from Stefan, ensuring we properly break links for short
   send/sendmsg recv/recvmsg if MSG_WAITALL is set.

 - Catch and loop when needing to run task_work before a PF_IO_WORKER
   threads goes to sleep.

* tag 'io_uring-5.12-2021-03-21' of git://git.kernel.dk/linux-block:
  io_uring: call req_set_fail_links() on short send[msg]()/recv[msg]() with MSG_WAITALL
  io-wq: ensure task is running before processing task_work
  signal: don't allow STOP on PF_IO_WORKER threads
  signal: don't allow sending any signals to PF_IO_WORKER threads

3 years agoMerge tag 'staging-5.12-rc4' of git://git.kernel.org/pub/scm/linux/kernel/git/gregkh...
Linus Torvalds [Sun, 21 Mar 2021 18:54:04 +0000 (11:54 -0700)]
Merge tag 'staging-5.12-rc4' of git://git./linux/kernel/git/gregkh/staging

Pull staging and IIO driver fixes from Greg KH:
 "Some small staging and IIO driver fixes:

   - MAINTAINERS changes for the move of the staging mailing list

   - comedi driver fixes to get request_irq() to work correctly

   - counter driver fixes for reported issues with iio devices

   - tiny iio driver fixes for reported issues.

  All of these have been in linux-next with no reported problems"

* tag 'staging-5.12-rc4' of git://git.kernel.org/pub/scm/linux/kernel/git/gregkh/staging:
  staging: vt665x: fix alignment constraints
  staging: comedi: cb_pcidas64: fix request_irq() warn
  staging: comedi: cb_pcidas: fix request_irq() warn
  MAINTAINERS: move the staging subsystem to lists.linux.dev
  MAINTAINERS: move some real subsystems off of the staging mailing list
  iio: gyro: mpu3050: Fix error handling in mpu3050_trigger_handler
  iio: hid-sensor-temperature: Fix issues of timestamp channel
  iio: hid-sensor-humidity: Fix alignment issue of timestamp channel
  counter: stm32-timer-cnt: fix ceiling miss-alignment with reload register
  counter: stm32-timer-cnt: fix ceiling write max value
  counter: stm32-timer-cnt: Report count function when SLAVE_MODE_DISABLED
  iio: adc: ab8500-gpadc: Fix off by 10 to 3
  iio:adc:stm32-adc: Add HAS_IOMEM dependency
  iio: adis16400: Fix an error code in adis16400_initial_setup()
  iio: adc: adi-axi-adc: add proper Kconfig dependencies
  iio: adc: ad7949: fix wrong ADC result due to incorrect bit mask
  iio: hid-sensor-prox: Fix scale not correct issue
  iio:adc:qcom-spmi-vadc: add default scale to LR_MUX2_BAT_ID channel

3 years agoMerge tag 'usb-5.12-rc4' of git://git.kernel.org/pub/scm/linux/kernel/git/gregkh/usb
Linus Torvalds [Sun, 21 Mar 2021 18:49:16 +0000 (11:49 -0700)]
Merge tag 'usb-5.12-rc4' of git://git./linux/kernel/git/gregkh/usb

Pull USB and Thunderbolt driver fixes from Greg KH:
 "Here are some small Thunderbolt and USB driver fixes for some reported
  issues:

   - thunderbolt fixes for minor problems

   - typec fixes for power issues

   - usb-storage quirk addition

   - usbip bugfix

   - dwc3 bugfix when stopping transfers

   - cdnsp bugfix for isoc transfers

   - gadget use-after-free fix

  All have been in linux-next this week with no reported issues"

* tag 'usb-5.12-rc4' of git://git.kernel.org/pub/scm/linux/kernel/git/gregkh/usb:
  usb: typec: tcpm: Skip sink_cap query only when VDM sm is busy
  usb: dwc3: gadget: Prevent EP queuing while stopping transfers
  usb: typec: tcpm: Invoke power_supply_changed for tcpm-source-psy-
  usb: typec: Remove vdo[3] part of tps6598x_rx_identity_reg struct
  usb-storage: Add quirk to defeat Kindle's automatic unload
  usb: gadget: configfs: Fix KASAN use-after-free
  usbip: Fix incorrect double assignment to udc->ud.tcp_rx
  usb: cdnsp: Fixes incorrect value in ISOC TRB
  thunderbolt: Increase runtime PM reference count on DP tunnel discovery
  thunderbolt: Initialize HopID IDAs in tb_switch_alloc()

3 years agoMerge tag 'irq-urgent-2021-03-21' of git://git.kernel.org/pub/scm/linux/kernel/git...
Linus Torvalds [Sun, 21 Mar 2021 18:34:24 +0000 (11:34 -0700)]
Merge tag 'irq-urgent-2021-03-21' of git://git./linux/kernel/git/tip/tip

Pull irq fix from Ingo Molnar:
 "A change to robustify force-threaded IRQ handlers to always disable
  interrupts, plus a DocBook fix.

  The force-threaded IRQ handler change has been accelerated from the
  normal schedule of such a change to keep the bad pattern/workaround of
  spin_lock_irqsave() in handlers or IRQF_NOTHREAD as a kludge from
  spreading"

* tag 'irq-urgent-2021-03-21' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
  genirq: Disable interrupts for force threaded handlers
  genirq/irq_sim: Fix typos in kernel doc (fnode -> fwnode)

3 years agoMerge tag 'perf-urgent-2021-03-21' of git://git.kernel.org/pub/scm/linux/kernel/git...
Linus Torvalds [Sun, 21 Mar 2021 18:26:21 +0000 (11:26 -0700)]
Merge tag 'perf-urgent-2021-03-21' of git://git./linux/kernel/git/tip/tip

Pull perf fixes from Ingo Molnar:
 "Boundary condition fixes for bugs unearthed by the perf fuzzer"

* tag 'perf-urgent-2021-03-21' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
  perf/x86/intel: Fix unchecked MSR access error caused by VLBR_EVENT
  perf/x86/intel: Fix a crash caused by zero PEBS status

3 years agoMerge tag 'locking-urgent-2021-03-21' of git://git.kernel.org/pub/scm/linux/kernel...
Linus Torvalds [Sun, 21 Mar 2021 18:19:29 +0000 (11:19 -0700)]
Merge tag 'locking-urgent-2021-03-21' of git://git./linux/kernel/git/tip/tip

Pull locking fixes from Ingo Molnar:

 - Get static calls & modules right. Hopefully.

 - WW mutex fixes

* tag 'locking-urgent-2021-03-21' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
  static_call: Fix static_call_update() sanity check
  static_call: Align static_call_is_init() patching condition
  static_call: Fix static_call_set_init()
  locking/ww_mutex: Fix acquire/release imbalance in ww_acquire_init()/ww_acquire_fini()
  locking/ww_mutex: Simplify use_ww_ctx & ww_ctx handling

3 years agoMerge tag 'efi-urgent-2021-03-21' of git://git.kernel.org/pub/scm/linux/kernel/git...
Linus Torvalds [Sun, 21 Mar 2021 18:11:22 +0000 (11:11 -0700)]
Merge tag 'efi-urgent-2021-03-21' of git://git./linux/kernel/git/tip/tip

Pull EFI fixes from Ingo Molnar:

 - another missing RT_PROP table related fix, to ensure that the
   efivarfs pseudo filesystem fails gracefully if variable services
   are unsupported

 - use the correct alignment for literal EFI GUIDs

 - fix a use after unmap issue in the memreserve code

* tag 'efi-urgent-2021-03-21' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
  efi: use 32-bit alignment for efi_guid_t literals
  firmware/efi: Fix a use after bug in efi_mem_reserve_persistent
  efivars: respect EFI_UNSUPPORTED return from firmware

3 years agoMerge tag 'x86_urgent_for_v5.12-rc4' of git://git.kernel.org/pub/scm/linux/kernel...
Linus Torvalds [Sun, 21 Mar 2021 18:04:20 +0000 (11:04 -0700)]
Merge tag 'x86_urgent_for_v5.12-rc4' of git://git./linux/kernel/git/tip/tip

Pull x86 fixes from Borislav Petkov:
 "The freshest pile of shiny x86 fixes for 5.12:

   - Add the arch-specific mapping between physical and logical CPUs to
     fix devicetree-node lookups

   - Restore the IRQ2 ignore logic

   - Fix get_nr_restart_syscall() to return the correct restart syscall
     number. Split in a 4-patches set to avoid kABI breakage when
     backporting to dead kernels"

* tag 'x86_urgent_for_v5.12-rc4' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
  x86/apic/of: Fix CPU devicetree-node lookups
  x86/ioapic: Ignore IRQ2 again
  x86: Introduce restart_block->arch_data to remove TS_COMPAT_RESTART
  x86: Introduce TS_COMPAT_RESTART to fix get_nr_restart_syscall()
  x86: Move TS_COMPAT back to asm/thread_info.h
  kernel, fs: Introduce and use set_restart_fn() and arch_set_restart_data()

3 years agoMerge tag 'powerpc-5.12-4' of git://git.kernel.org/pub/scm/linux/kernel/git/powerpc...
Linus Torvalds [Sun, 21 Mar 2021 17:57:35 +0000 (10:57 -0700)]
Merge tag 'powerpc-5.12-4' of git://git./linux/kernel/git/powerpc/linux

Pull powerpc fixes from Michael Ellerman:

 - Fix a possible stack corruption and subsequent DLPAR failure in the
   rpadlpar_io PCI hotplug driver

 - Two build fixes for uncommon configurations

Thanks to Christophe Leroy and Tyrel Datwyler.

* tag 'powerpc-5.12-4' of git://git.kernel.org/pub/scm/linux/kernel/git/powerpc/linux:
  PCI: rpadlpar: Fix potential drc_name corruption in store functions
  powerpc: Force inlining of cpu_has_feature() to avoid build failure
  powerpc/vdso32: Add missing _restgpr_31_x to fix build failure

3 years agoio_uring: call req_set_fail_links() on short send[msg]()/recv[msg]() with MSG_WAITALL
Stefan Metzmacher [Sat, 20 Mar 2021 19:33:36 +0000 (20:33 +0100)]
io_uring: call req_set_fail_links() on short send[msg]()/recv[msg]() with MSG_WAITALL

Without that it's not safe to use them in a linked combination with
others.

Now combinations like IORING_OP_SENDMSG followed by IORING_OP_SPLICE
should be possible.

We already handle short reads and writes for the following opcodes:

- IORING_OP_READV
- IORING_OP_READ_FIXED
- IORING_OP_READ
- IORING_OP_WRITEV
- IORING_OP_WRITE_FIXED
- IORING_OP_WRITE
- IORING_OP_SPLICE
- IORING_OP_TEE

Now we have it for these as well:

- IORING_OP_SENDMSG
- IORING_OP_SEND
- IORING_OP_RECVMSG
- IORING_OP_RECV

For IORING_OP_RECVMSG we also check for the MSG_TRUNC and MSG_CTRUNC
flags in order to call req_set_fail_links().

There might be applications arround depending on the behavior
that even short send[msg]()/recv[msg]() retuns continue an
IOSQE_IO_LINK chain.

It's very unlikely that such applications pass in MSG_WAITALL,
which is only defined in 'man 2 recvmsg', but not in 'man 2 sendmsg'.

It's expected that the low level sock_sendmsg() call just ignores
MSG_WAITALL, as MSG_ZEROCOPY is also ignored without explicitly set
SO_ZEROCOPY.

We also expect the caller to know about the implicit truncation to
MAX_RW_COUNT, which we don't detect.

cc: netdev@vger.kernel.org
Link: https://lore.kernel.org/r/c4e1a4cc0d905314f4d5dc567e65a7b09621aab3.1615908477.git.metze@samba.org
Signed-off-by: Stefan Metzmacher <metze@samba.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
3 years agoio-wq: ensure task is running before processing task_work
Jens Axboe [Sun, 21 Mar 2021 13:06:56 +0000 (07:06 -0600)]
io-wq: ensure task is running before processing task_work

Mark the current task as running if we need to run task_work from the
io-wq threads as part of work handling. If that is the case, then return
as such so that the caller can appropriately loop back and reset if it
was part of a going-to-sleep flush.

Fixes: 3bfe6106693b ("io-wq: fork worker threads from original task")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
3 years agosignal: don't allow STOP on PF_IO_WORKER threads
Eric W. Biederman [Sun, 21 Mar 2021 15:37:48 +0000 (09:37 -0600)]
signal: don't allow STOP on PF_IO_WORKER threads

Just like we don't allow normal signals to IO threads, don't deliver a
STOP to a task that has PF_IO_WORKER set. The IO threads don't take
signals in general, and have no means of flushing out a stop either.

Longer term, we may want to look into allowing stop of these threads,
as it relates to eg process freezing. For now, this prevents a spin
issue if a SIGSTOP is delivered to the parent task.

Reported-by: Stefan Metzmacher <metze@samba.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
3 years agosignal: don't allow sending any signals to PF_IO_WORKER threads
Jens Axboe [Sat, 20 Mar 2021 01:25:13 +0000 (19:25 -0600)]
signal: don't allow sending any signals to PF_IO_WORKER threads

They don't take signals individually, and even if they share signals with
the parent task, don't allow them to be delivered through the worker
thread. Linux does allow this kind of behavior for regular threads, but
it's really a compatability thing that we need not care about for the IO
threads.

Reported-by: Stefan Metzmacher <metze@samba.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
3 years agoext4: initialize ret to suppress smatch warning
Theodore Ts'o [Sun, 21 Mar 2021 04:45:37 +0000 (00:45 -0400)]
ext4: initialize ret to suppress smatch warning

Signed-off-by: Theodore Ts'o <tytso@mit.edu>
3 years agoext4: stop inode update before return
Pan Bian [Sun, 17 Jan 2021 08:57:32 +0000 (00:57 -0800)]
ext4: stop inode update before return

The inode update should be stopped before returing the error code.

Signed-off-by: Pan Bian <bianpan2016@163.com>
Link: https://lore.kernel.org/r/20210117085732.93788-1-bianpan2016@163.com
Fixes: 8016e29f4362 ("ext4: fast commit recovery path")
Cc: stable@kernel.org
Reviewed-by: Harshad Shirwadkar <harshadshirwadkar@gmail.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
3 years agoext4: fix rename whiteout with fast commit
Harshad Shirwadkar [Tue, 16 Mar 2021 22:19:21 +0000 (15:19 -0700)]
ext4: fix rename whiteout with fast commit

This patch adds rename whiteout support in fast commits. Note that the
whiteout object that gets created is actually char device. Which
imples, the function ext4_inode_journal_mode(struct inode *inode)
would return "JOURNAL_DATA" for this inode. This has a consequence in
fast commit code that it will make creation of the whiteout object a
fast-commit ineligible behavior and thus will fall back to full
commits. With this patch, this can be observed by running fast commits
with rename whiteout and seeing the stats generated by ext4_fc_stats
tracepoint as follows:

ext4_fc_stats: dev 254:32 fc ineligible reasons:
XATTR:0, CROSS_RENAME:0, JOURNAL_FLAG_CHANGE:0, NO_MEM:0, SWAP_BOOT:0,
RESIZE:0, RENAME_DIR:0, FALLOC_RANGE:0, INODE_JOURNAL_DATA:16;
num_commits:6, ineligible: 6, numblks: 3

So in short, this patch guarantees that in case of rename whiteout, we
fall back to full commits.

Amir mentioned that instead of creating a new whiteout object for
every rename, we can create a static whiteout object with irrelevant
nlink. That will make fast commits to not fall back to full
commit. But until this happens, this patch will ensure correctness by
falling back to full commits.

Fixes: 8016e29f4362 ("ext4: fast commit recovery path")
Cc: stable@kernel.org
Signed-off-by: Harshad Shirwadkar <harshadshirwadkar@gmail.com>
Link: https://lore.kernel.org/r/20210316221921.1124955-1-harshadshirwadkar@gmail.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
3 years agoext4: fix timer use-after-free on failed mount
Jan Kara [Mon, 15 Mar 2021 16:59:06 +0000 (17:59 +0100)]
ext4: fix timer use-after-free on failed mount

When filesystem mount fails because of corrupted filesystem we first
cancel the s_err_report timer reminding fs errors every day and only
then we flush s_error_work. However s_error_work may report another fs
error and re-arm timer thus resulting in timer use-after-free. Fix the
problem by first flushing the work and only after that canceling the
s_err_report timer.

Reported-by: syzbot+628472a2aac693ab0fcd@syzkaller.appspotmail.com
Fixes: 2d01ddc86606 ("ext4: save error info to sb through journal if available")
CC: stable@vger.kernel.org
Signed-off-by: Jan Kara <jack@suse.cz>
Link: https://lore.kernel.org/r/20210315165906.2175-1-jack@suse.cz
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
3 years agoext4: fix potential error in ext4_do_update_inode
Shijie Luo [Fri, 12 Mar 2021 06:50:51 +0000 (01:50 -0500)]
ext4: fix potential error in ext4_do_update_inode

If set_large_file = 1 and errors occur in ext4_handle_dirty_metadata(),
the error code will be overridden, go to out_brelse to avoid this
situation.

Signed-off-by: Shijie Luo <luoshijie1@huawei.com>
Link: https://lore.kernel.org/r/20210312065051.36314-1-luoshijie1@huawei.com
Cc: stable@kernel.org
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
3 years agoext4: do not try to set xattr into ea_inode if value is empty
zhangyi (F) [Fri, 5 Mar 2021 12:05:08 +0000 (20:05 +0800)]
ext4: do not try to set xattr into ea_inode if value is empty

Syzbot report a warning that ext4 may create an empty ea_inode if set
an empty extent attribute to a file on the file system which is no free
blocks left.

  WARNING: CPU: 6 PID: 10667 at fs/ext4/xattr.c:1640 ext4_xattr_set_entry+0x10f8/0x1114 fs/ext4/xattr.c:1640
  ...
  Call trace:
   ext4_xattr_set_entry+0x10f8/0x1114 fs/ext4/xattr.c:1640
   ext4_xattr_block_set+0x1d0/0x1b1c fs/ext4/xattr.c:1942
   ext4_xattr_set_handle+0x8a0/0xf1c fs/ext4/xattr.c:2390
   ext4_xattr_set+0x120/0x1f0 fs/ext4/xattr.c:2491
   ext4_xattr_trusted_set+0x48/0x5c fs/ext4/xattr_trusted.c:37
   __vfs_setxattr+0x208/0x23c fs/xattr.c:177
  ...

Now, ext4 try to store extent attribute into an external inode if
ext4_xattr_block_set() return -ENOSPC, but for the case of store an
empty extent attribute, store the extent entry into the extent
attribute block is enough. A simple reproduce below.

  fallocate test.img -l 1M
  mkfs.ext4 -F -b 2048 -O ea_inode test.img
  mount test.img /mnt
  dd if=/dev/zero of=/mnt/foo bs=2048 count=500
  setfattr -n "user.test" /mnt/foo

Reported-by: syzbot+98b881fdd8ebf45ab4ae@syzkaller.appspotmail.com
Fixes: 9c6e7853c531 ("ext4: reserve space for xattr entries/names")
Cc: stable@kernel.org
Signed-off-by: zhangyi (F) <yi.zhang@huawei.com>
Link: https://lore.kernel.org/r/20210305120508.298465-1-yi.zhang@huawei.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
3 years agoext4: do not iput inode under running transaction in ext4_rename()
zhangyi (F) [Wed, 3 Mar 2021 13:17:03 +0000 (21:17 +0800)]
ext4: do not iput inode under running transaction in ext4_rename()

In ext4_rename(), when RENAME_WHITEOUT failed to add new entry into
directory, it ends up dropping new created whiteout inode under the
running transaction. After commit <9b88f9fb0d2> ("ext4: Do not iput inode
under running transaction"), we follow the assumptions that evict() does
not get called from a transaction context but in ext4_rename() it breaks
this suggestion. Although it's not a real problem, better to obey it, so
this patch add inode to orphan list and stop transaction before final
iput().

Signed-off-by: zhangyi (F) <yi.zhang@huawei.com>
Link: https://lore.kernel.org/r/20210303131703.330415-2-yi.zhang@huawei.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
3 years agoext4: find old entry again if failed to rename whiteout
zhangyi (F) [Wed, 3 Mar 2021 13:17:02 +0000 (21:17 +0800)]
ext4: find old entry again if failed to rename whiteout

If we failed to add new entry on rename whiteout, we cannot reset the
old->de entry directly, because the old->de could have moved from under
us during make indexed dir. So find the old entry again before reset is
needed, otherwise it may corrupt the filesystem as below.

  /dev/sda: Entry '00000001' in ??? (12) has deleted/unused inode 15. CLEARED.
  /dev/sda: Unattached inode 75
  /dev/sda: UNEXPECTED INCONSISTENCY; RUN fsck MANUALLY.

Fixes: 6b4b8e6b4ad ("ext4: fix bug for rename with RENAME_WHITEOUT")
Cc: stable@vger.kernel.org
Signed-off-by: zhangyi (F) <yi.zhang@huawei.com>
Link: https://lore.kernel.org/r/20210303131703.330415-1-yi.zhang@huawei.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
3 years agogenirq: Disable interrupts for force threaded handlers
Thomas Gleixner [Wed, 17 Mar 2021 14:38:52 +0000 (15:38 +0100)]
genirq: Disable interrupts for force threaded handlers

With interrupt force threading all device interrupt handlers are invoked
from kernel threads. Contrary to hard interrupt context the invocation only
disables bottom halfs, but not interrupts. This was an oversight back then
because any code like this will have an issue:

thread(irq_A)
  irq_handler(A)
    spin_lock(&foo->lock);

interrupt(irq_B)
  irq_handler(B)
    spin_lock(&foo->lock);

This has been triggered with networking (NAPI vs. hrtimers) and console
drivers where printk() happens from an interrupt which interrupted the
force threaded handler.

Now people noticed and started to change the spin_lock() in the handler to
spin_lock_irqsave() which affects performance or add IRQF_NOTHREAD to the
interrupt request which in turn breaks RT.

Fix the root cause and not the symptom and disable interrupts before
invoking the force threaded handler which preserves the regular semantics
and the usefulness of the interrupt force threading as a general debugging
tool.

For not RT this is not changing much, except that during the execution of
the threaded handler interrupts are delayed until the handler
returns. Vs. scheduling and softirq processing there is no difference.

For RT kernels there is no issue.

Fixes: 8d32a307e4fa ("genirq: Provide forced interrupt threading")
Reported-by: Johan Hovold <johan@kernel.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Johan Hovold <johan@kernel.org>
Acked-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lore.kernel.org/r/20210317143859.513307808@linutronix.de
3 years agoMerge tag 'riscv-for-linus-5.12-rc4' of git://git.kernel.org/pub/scm/linux/kernel...
Linus Torvalds [Sat, 20 Mar 2021 18:01:54 +0000 (11:01 -0700)]
Merge tag 'riscv-for-linus-5.12-rc4' of git://git./linux/kernel/git/riscv/linux

Pull RISC-V fixes from Palmer Dabbelt:
 "A handful of fixes for 5.12:

   - fix the SBI remote fence numbers for hypervisor fences, which had
     been transcribed in the wrong order in Linux. These fences are only
     used with the KVM patches applied.

   - fix a whole host of build warnings, these should have no functional
     change.

   - fix init_resources() to prevent an off-by-one error from causing an
     out-of-bounds array reference. This was manifesting during boot on
     vexriscv.

   - ensure the KASAN mappings are visible before proceeding to use
     them"

* tag 'riscv-for-linus-5.12-rc4' of git://git.kernel.org/pub/scm/linux/kernel/git/riscv/linux:
  riscv: Correct SPARSEMEM configuration
  RISC-V: kasan: Declare kasan_shallow_populate() static
  riscv: Ensure page table writes are flushed when initializing KASAN vmalloc
  RISC-V: Fix out-of-bounds accesses in init_resources()
  riscv: Fix compilation error with Canaan SoC
  ftrace: Fix spelling mistake "disabed" -> "disabled"
  riscv: fix bugon.cocci warnings
  riscv: process: Fix no prototype for arch_dup_task_struct
  riscv: ftrace: Use ftrace_get_regs helper
  riscv: process: Fix no prototype for show_regs
  riscv: syscall_table: Reduce W=1 compilation warnings noise
  riscv: time: Fix no prototype for time_init
  riscv: ptrace: Fix no prototype warnings
  riscv: sbi: Fix comment of __sbi_set_timer_v01
  riscv: irq: Fix no prototype warning
  riscv: traps: Fix no prototype warnings
  RISC-V: correct enum sbi_ext_rfence_fid

3 years agoMerge tag '5.12-rc3-smb3' of git://git.samba.org/sfrench/cifs-2.6
Linus Torvalds [Sat, 20 Mar 2021 18:00:25 +0000 (11:00 -0700)]
Merge tag '5.12-rc3-smb3' of git://git.samba.org/sfrench/cifs-2.6

Pull cifs fixes from Steve French:
 "Five cifs/smb3 fixes - three for stable, including an important ACL
  fix and security signature fix"

* tag '5.12-rc3-smb3' of git://git.samba.org/sfrench/cifs-2.6:
  cifs: fix allocation size on newly created files
  cifs: warn and fail if trying to use rootfs without the config option
  fs/cifs/: fix misspellings using codespell tool
  cifs: Fix preauth hash corruption
  cifs: update new ACE pointer after populate_new_aces.

3 years agoMerge tag 'scsi-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/jejb/scsi
Linus Torvalds [Sat, 20 Mar 2021 17:57:10 +0000 (10:57 -0700)]
Merge tag 'scsi-fixes' of git://git./linux/kernel/git/jejb/scsi

Pull SCSI fixes from James Bottomley:
 "Eight fixes, all in drivers, all fairly minor either being fixes in
  error legs, memory leaks on teardown, context errors or semantic
  problems"

* tag 'scsi-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/jejb/scsi:
  scsi: mpt3sas: Do not use GFP_KERNEL in atomic context
  scsi: ufs: ufs-mediatek: Correct operator & -> &&
  scsi: sd_zbc: Update write pointer offset cache
  scsi: lpfc: Fix some error codes in debugfs
  scsi: qla2xxx: Fix broken #endif placement
  scsi: st: Fix a use after free in st_open()
  scsi: myrs: Fix a double free in myrs_cleanup()
  scsi: ibmvfc: Free channel_setup_buf during device tear down

3 years agoMerge tag 'zonefs-5.12-rc4' of git://git.kernel.org/pub/scm/linux/kernel/git/dlemoal...
Linus Torvalds [Sat, 20 Mar 2021 00:32:30 +0000 (17:32 -0700)]
Merge tag 'zonefs-5.12-rc4' of git://git./linux/kernel/git/dlemoal/zonefs

Pull zonefs fixes from Damien Le Moal:

 - fix inode write open reference count (Chao)

 - Fix wrong write offset for asynchronous O_APPEND writes (me)

 - Prevent use of sequential zone file as swap files (me)

* tag 'zonefs-5.12-rc4' of git://git.kernel.org/pub/scm/linux/kernel/git/dlemoal/zonefs:
  zonefs: fix to update .i_wr_refcnt correctly in zonefs_open_zone()
  zonefs: Fix O_APPEND async write handling
  zonefs: prevent use of seq files as swap file

3 years agoMerge tag 'block-5.12-2021-03-19' of git://git.kernel.dk/linux-block
Linus Torvalds [Sat, 20 Mar 2021 00:07:10 +0000 (17:07 -0700)]
Merge tag 'block-5.12-2021-03-19' of git://git.kernel.dk/linux-block

Pull block fixes from Jens Axboe:
 "Just an NVMe pull request this week:

   - fix tag allocation for keep alive

   - fix a unit mismatch for the Write Zeroes limits

   - various TCP transport fixes (Sagi Grimberg, Elad Grupi)

   - fix iosqes and iocqes validation for discovery controllers (Sagi Grimberg)"

* tag 'block-5.12-2021-03-19' of git://git.kernel.dk/linux-block:
  nvmet-tcp: fix kmap leak when data digest in use
  nvmet: don't check iosqes,iocqes for discovery controllers
  nvme-rdma: fix possible hang when failing to set io queues
  nvme-tcp: fix possible hang when failing to set io queues
  nvme-tcp: fix misuse of __smp_processor_id with preemption enabled
  nvme-tcp: fix a NULL deref when receiving a 0-length r2t PDU
  nvme: fix Write Zeroes limitations
  nvme: allocate the keep alive request using BLK_MQ_REQ_NOWAIT
  nvme: merge nvme_keep_alive into nvme_keep_alive_work
  nvme-fabrics: only reserve a single tag

3 years agoMerge tag 'io_uring-5.12-2021-03-19' of git://git.kernel.dk/linux-block
Linus Torvalds [Sat, 20 Mar 2021 00:01:09 +0000 (17:01 -0700)]
Merge tag 'io_uring-5.12-2021-03-19' of git://git.kernel.dk/linux-block

Pull io_uring fixes from Jens Axboe:
 "Quieter week this time, which was both expected and desired. About
  half of the below is fixes for this release, the other half are just
  fixes in general. In detail:

   - Fix the freezing of IO threads, by making the freezer not send them
     fake signals. Make them freezable by default.

   - Like we did for personalities, move the buffer IDR to xarray. Kills
     some code and avoids a use-after-free on teardown.

   - SQPOLL cleanups and fixes (Pavel)

   - Fix linked timeout race (Pavel)

   - Fix potential completion post use-after-free (Pavel)

   - Cleanup and move internal structures outside of general kernel view
     (Stefan)

   - Use MSG_SIGNAL for send/recv from io_uring (Stefan)"

* tag 'io_uring-5.12-2021-03-19' of git://git.kernel.dk/linux-block:
  io_uring: don't leak creds on SQO attach error
  io_uring: use typesafe pointers in io_uring_task
  io_uring: remove structures from include/linux/io_uring.h
  io_uring: imply MSG_NOSIGNAL for send[msg]()/recv[msg]() calls
  io_uring: fix sqpoll cancellation via task_work
  io_uring: add generic callback_head helpers
  io_uring: fix concurrent parking
  io_uring: halt SQO submission on ctx exit
  io_uring: replace sqd rw_semaphore with mutex
  io_uring: fix complete_post use ctx after free
  io_uring: fix ->flags races by linked timeouts
  io_uring: convert io_buffer_idr to XArray
  io_uring: allow IO worker threads to be frozen
  kernel: freezer should treat PF_IO_WORKER like PF_KTHREAD for freezing