While investigating the extent corruption bug I ran into this bug in debug
only code. xfs_bmap_check_leaf_extents() loops through the leaf blocks of
the extent btree checking that every extent is entirely before the next
extent. It also compares the last extent in the previous block to the
first extent in the current block when the previous block has been
released and potentially unmapped. So take a copy of the last extent
instead of a pointer. Also move the last extent check out of the loop
because we only need to do it once.

SGI-PV: 976035
SGI-Modid: xfs-linux-melb:xfs-kern:30718a
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>
Signed-off-by: NChristoph Hellwig <hch@infradead.org>

Most VN_RELE calls either directly contain a XFS_ITOV or have the
corresponding xfs_inode already in scope. Use the IRELE helper instead of
VN_RELE to clarify the code. With a little more work we can kill VN_RELE
altogether and define IRELE in terms of iput directly.

SGI-PV: 976035
SGI-Modid: xfs-linux-melb:xfs-kern:30710a
Signed-off-by: NChristoph Hellwig <hch@infradead.org>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

The three subcases of xfs_ioc_xattr don't share any semantics and almost
no code, so split it into three separate helpers.

SGI-PV: 976035
SGI-Modid: xfs-linux-melb:xfs-kern:30709a
Signed-off-by: NChristoph Hellwig <hch@infradead.org>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

- rename rootvp to root for clarify
- remove useless vn_to_inode call
- check is_bad_inode before calling d_alloc_root
- use iput instead of VN_RELE in the error case

SGI-PV: 976035
SGI-Modid: xfs-linux-melb:xfs-kern:30708a
Signed-off-by: NChristoph Hellwig <hch@infradead.org>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

When writing into preallocated regions there is a case where XFS can oops
or hang doing the unwritten extent conversion on I/O completion. It turns
out that the problem is related to the btree cursor being invalid.

When we do an insert into the tree, we may need to split blocks in the
tree. When we only split at the leaf level (i.e. level 0), everything
works just fine. However, if we have a multi-level split in the btreee,
the cursor passed to the insert function is no longer valid once the
insert is complete.

The leaf level split is handled correctly because all the operations at
level 0 are done using the original cursor, hence it is updated correctly.
However, when we need to update the next level up the tree, we don't use
that cursor - we use a cloned cursor that points to the index in the next
level up where we need to do the insert.

Hence if we need to split a second level, the changes to the tree are
reflected in the cloned cursor and not the original cursor. This
clone-and-move-up-a-level-on-split behaviour recurses all the way to the
top of the tree.

The complexity here is that these cloned cursors do not point to the
original index that was inserted - they point to the newly allocated block
(the right block) and the original cursor pointer to that level may still
point to the left block. Hence, without deep examination of the cloned
cursor and buffers, we cannot update the original cursor with the new path
from the cloned cursor.

In these cases the original cursor could be pointing to the wrong block(s)
and hence a subsequent modification to the tree using that cursor will
lead to corruption of the tree.

The crash case occurs when the tree changes height - we insert a new level
in the tree, and the cursor does not have a buffer in it's path for that
level. Hence any attempt to walk back up the cursor to the root block will
result in a null pointer dereference.

To make matters even more complex, the BMAP BT is rooted in an inode, so
we can have a change of height in the btree *without a root split*. That
is, if the root block in the inode is full when we split a leaf node, we
cannot fit the pointer to the new block in the root, so we allocate a new
block, migrate all the ptrs out of the inode into the new block and point
the inode root block at the newly allocated block. This changes the height
of the tree without a root split having occurred and hence invalidates the
path in the original cursor.

The patch below prevents xfs_bmbt_insert() from returning with an invalid
cursor by detecting the cases that invalidate the original cursor and
refresh it by do a lookup into the btree for the original index we were
inserting at.

Note that the INOBT, AGFBNO and AGFCNT btree implementations also have
this bug, but the cursor is currently always destroyed or revalidated
after an insert for those trees. Hence this patch only address the problem
in the BMBT code.

SGI-PV: 979339
SGI-Modid: xfs-linux-melb:xfs-kern:30701a
Signed-off-by: NDavid Chinner <dgc@sgi.com>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

At ENOSPC, we can get a filesystem shutdown due to a cancelling a dirty
transaction in xfs_mkdir or xfs_create. This is due to the initial
allocation attempt not taking into account inode alignment and hence we
can prepare the AGF freelist for allocation when it's not actually
possible to do an allocation. This results in inode allocation returning
ENOSPC with a dirty transaction, and hence we shut down the filesystem.

Because the first allocation is an exact allocation attempt, we must tell
the allocator that the alignment does not affect the allocation attempt.
i.e. we will accept any extent alignment as long as the extent starts at
the block we want. Unfortunately, this means that if the longest free
extent is less than the length + alignment necessary for fallback
allocation attempts but is long enough to attempt a non-aligned
allocation, we will modify the free list.

If we then have the exact allocation fail, all other allocation attempts
will also fail due to the alignment constraint being taken into account.
Hence the initial attempt needs to set the "alignment slop" field so that
alignment, while not required, must be taken into account when determining
if there is enough space left in the AG to do the allocation.

That means if the exact allocation fails, we will not dirty the freelist
if there is not enough space available fo a subsequent allocation to
succeed. Hence we get an ENOSPC error back to userspace without shutting
down the filesystem.

SGI-PV: 978886
SGI-Modid: xfs-linux-melb:xfs-kern:30699a
Signed-off-by: NDavid Chinner <dgc@sgi.com>
Signed-off-by: NChristoph Hellwig <hch@infradead.org>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

Replace the xfs_ail_entry_t with a struct list_head and clean the
surrounding code up. Also fixes a livelock in xfs_trans_first_push_ail()
by terminating the loop at the head of the list correctly.

SGI-PV: 978682
SGI-Modid: xfs-linux-melb:xfs-kern:30636a
Signed-off-by: NJosef 'Jeff' Sipek <jeffpc@josefsipek.net>
Signed-off-by: NDavid Chinner <dgc@sgi.com>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

When xfs_mountfs is called by xfs_mount xfs_readsb was called 35 lines
above unconditionally, so there is no need to try to read the superblock
if it's not present. If any other port doesn't have the superblock read at
this point it should just call it directly from it's xfs_mount equivalent.

SGI-PV: 976035
SGI-Modid: xfs-linux-melb:xfs-kern:30603a
Signed-off-by: NChristoph Hellwig <hch@infradead.org>
Signed-off-by: NDonald Douwsma <donaldd@sgi.com>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

It's completely unused so we might aswell kill it. Note that there is
another t_sema in struct xlog_ticket, which is used and actually an sv_t
despite the name. That one is left untouched by this patch.

SGI-PV: 971186
SGI-Modid: xfs-linux-melb:xfs-kern:30591a
Signed-off-by: NNiv Sardi <xaiki@sgi.com>
Signed-off-by: NChristoph Hellwig <hch@infradead.org>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

SGI-PV: 976035
SGI-Modid: xfs-linux-melb:xfs-kern:30553a
Signed-off-by: NChristoph Hellwig <hch@infradead.org>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

SGI-PV: 976035
SGI-Modid: xfs-linux-melb:xfs-kern:30552a
Signed-off-by: NChristoph Hellwig <hch@infradead.org>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

SGI-PV: 976035
SGI-Modid: xfs-linux-melb:xfs-kern:30551a
Signed-off-by: NChristoph Hellwig <hch@infradead.org>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

SGI-PV: 976035
SGI-Modid: xfs-linux-melb:xfs-kern:30550a
Signed-off-by: NChristoph Hellwig <hch@infradead.org>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

SGI-PV: 976035
SGI-Modid: xfs-linux-melb:xfs-kern:30548a
Signed-off-by: NChristoph Hellwig <hch@infradead.org>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

SGI-PV: 976035
SGI-Modid: xfs-linux-melb:xfs-kern:30547a
Signed-off-by: NChristoph Hellwig <hch@infradead.org>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

SGI-PV: 976035
SGI-Modid: xfs-linux-melb:xfs-kern:30546a
Signed-off-by: NChristoph Hellwig <hch@infradead.org>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

SGI-PV: 976035
SGI-Modid: xfs-linux-melb:xfs-kern:30545a
Signed-off-by: NChristoph Hellwig <hch@infradead.org>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

Now that the ktrace_enter() code is using atomics, the non-power-of-2
buffer sizes - which require modulus operations to get the index - are
showing up as using substantial CPU in the profiles.

Force the buffer sizes to be rounded up to the nearest power of two and
use masking rather than modulus operations to convert the index counter to
the buffer index. This reduces ktrace_enter overhead to 8% of a CPU time,
and again almost halves the trace intensive test runtime.

SGI-PV: 977546
SGI-Modid: xfs-linux-melb:xfs-kern:30538a
Signed-off-by: NDavid Chinner <dgc@sgi.com>
Signed-off-by: NChristoph Hellwig <hch@infradead.org>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

ktrace_enter() is consuming vast amounts of CPU time due to the use of a
single global lock for protecting buffer index increments. Change it to
use per-buffer atomic counters - this reduces ktrace_enter() overhead
during a trace intensive test on a 4p machine from 58% of all CPU time to
12% and halves test runtime.

SGI-PV: 977546
SGI-Modid: xfs-linux-melb:xfs-kern:30537a
Signed-off-by: NDavid Chinner <dgc@sgi.com>
Signed-off-by: NChristoph Hellwig <hch@infradead.org>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

XFS changes the c/mtime of an inode when truncating it to the same size.
The c/mtime is only supposed to change if the size is changed. Not to be
confused with ftruncate, where the c/mtime is supposed to be changed even
if the size is not changed.

The Linux VFS encodes this semantic difference in the flags it sends down
to ->setattr, which XFS currently ignores. We need to make XFS pay
attention to the VFS flags and hence Do The Right Thing.

SGI-PV: 977547
SGI-Modid: xfs-linux-melb:xfs-kern:30536a
Signed-off-by: NDavid Chinner <dgc@sgi.com>
Signed-off-by: NChristoph Hellwig <hch@infradead.org>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

As Dave pointed out after the export ops changes we now always encode the
parent into the filehandle for regular files, but it's not actually needed
when the filesystem is export with no_subtree_check. This one-liner fixes
xfs_fs_encode_fh to skip encoding the parent unless nessecary.

SGI-PV: 976035
SGI-Modid: xfs-linux-melb:xfs-kern:30535a
Signed-off-by: NChristoph Hellwig <hch@infradead.org>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

We can just use xfs_ilock/xfs_iunlock instead and get rid of the ugly
bhv_vrwlock_t.

SGI-PV: 976035
SGI-Modid: xfs-linux-melb:xfs-kern:30533a
Signed-off-by: NChristoph Hellwig <hch@infradead.org>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

Instead of of xfs_get_dir_entry use a macro to get the xfs_inode from the
dentry in the callers and grab the reference manually.

Only grab the reference once as it's fine to keep it over the dmapi calls.
(And even that reference is actually superflous in Linux but I'll leave
that for another patch)

SGI-PV: 976035
SGI-Modid: xfs-linux-melb:xfs-kern:30531a
Signed-off-by: NChristoph Hellwig <hch@infradead.org>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

Cleanup the unneeded intermediate vnode step in the flushing helpers and
go directly from the xfs_inode to the struct address_space.

SGI-PV: 976035
SGI-Modid: xfs-linux-melb:xfs-kern:30530a
Signed-off-by: NChristoph Hellwig <hch@infradead.org>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

- use proper goto based unwinding instead of the current mess of
  multiple conditionals
- rename ip to inode because that's the normal convention for Linux
  inodes while ip is the convention for xfs_inodes
- remove unlikely checks for the default_acl - branches marked unlikely
  might lead to extreme branch bredictor slowdons if taken and for some
  workloads a default acl is quite common
- properly indent the switch statements
- remove xfs_has_fs_struct as nfsd has a fs_struct in any semi-recent
  kernel

SGI-PV: 976035
SGI-Modid: xfs-linux-melb:xfs-kern:30529a
Signed-off-by: NChristoph Hellwig <hch@infradead.org>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

Now that we update the log tail LSN less frequently on transaction
completion, we pass the contention straight to the global log state lock
(l_iclog_lock) during transaction completion.

We currently have to take this lock to decrement the iclog reference
count. there is a reference count on each iclog, so we need to take þhe
global lock for all refcount changes.

When large numbers of processes are all doing small trnasctions, the iclog
reference counts will be quite high, and the state change that absolutely
requires the l_iclog_lock is the except rather than the norm.

Change the reference counting on the iclogs to use atomic_inc/dec so that
we can use atomic_dec_and_lock during transaction completion and avoid the
need for grabbing the l_iclog_lock for every reference count decrement
except the one that matters - the last.

SGI-PV: 975671
SGI-Modid: xfs-linux-melb:xfs-kern:30505a
Signed-off-by: NDavid Chinner <dgc@sgi.com>
Signed-off-by: NTim Shimmin <tes@sgi.com>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

When hundreds of processors attempt to commit transactions at the same
time, they can contend on the AIL lock when updating the tail LSN held in
the in-core log structure.

At the moment, the tail LSN is only needed when actually writing out an
iclog, so it really does not need to be updated on every single
transaction completion - only those that result in switching iclogs and
flushing them to disk.

The result is that we reduce the number of times we need to grab the AIL
lock and the log grant lock by up to two orders of magnitude on large
processor count machines. The problem has previously been hidden by AIL
lock contention walking the AIL list which was recently solved and
uncovered this issue.

SGI-PV: 975671
SGI-Modid: xfs-linux-melb:xfs-kern:30504a
Signed-off-by: NDavid Chinner <dgc@sgi.com>
Signed-off-by: NTim Shimmin <tes@sgi.com>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

Remove open coded checks for the whether the inode is clean and replace
them with an inlined function.

SGI-PV: 977461
SGI-Modid: xfs-linux-melb:xfs-kern:30503a
Signed-off-by: NDavid Chinner <dgc@sgi.com>
Signed-off-by: NChristoph Hellwig <hch@infradead.org>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

Remove the xfs_icluster structure and replace with a radix tree lookup.

We don't need to keep a list of inodes in each cluster around anymore as
we can look them up quickly when we need to. The only time we need to do
this now is during inode writeback.

Factor the inode cluster writeback code out of xfs_iflush and convert it
to use radix_tree_gang_lookup() instead of walking a list of inodes built
when we first read in the inodes.

This remove 3 pointers from each xfs_inode structure and the xfs_icluster
structure per inode cluster. Hence we reduce the cache footprint of the
xfs_inodes by between 5-10% depending on cluster sparseness.

To be truly efficient we need a radix_tree_gang_lookup_range() call to
stop searching once we are past the end of the cluster instead of trying
to find a full cluster's worth of inodes.

Before (ia64):

$ cat /sys/slab/xfs_inode/object_size 536

After:

$ cat /sys/slab/xfs_inode/object_size 512

SGI-PV: 977460
SGI-Modid: xfs-linux-melb:xfs-kern:30502a
Signed-off-by: NDavid Chinner <dgc@sgi.com>
Signed-off-by: NChristoph Hellwig <hch@infradead.org>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

When pdflush is writing back inodes, it can get stuck on inode cluster
buffers that are currently under I/O. This occurs when we write data to
multiple inodes in the same inode cluster at the same time.

Effectively, delayed allocation marks the inode dirty during the data
writeback. Hence if the inode cluster was flushed during the writeback of
the first inode, the writeback of the second inode will block waiting for
the inode cluster write to complete before writing it again for the newly
dirtied inode.

Basically, we want to avoid this from happening so we don't block pdflush
and slow down all of writeback. Hence we introduce a non-blocking async
inode flush flag that pdflush uses. If this flag is set, we use
non-blocking operations (e.g. try locks) whereever we can to avoid
blocking or extra I/O being issued.

SGI-PV: 970925
SGI-Modid: xfs-linux-melb:xfs-kern:30501a
Signed-off-by: NDavid Chinner <dgc@sgi.com>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

The only difference between the functions is one passes an inode for the
lookup, the other passes an inode number. However, they don't do the same
validity checking or set all the same state on the buffer that is returned
yet they should.

Factor the functions into a common implementation.

SGI-PV: 970925
SGI-Modid: xfs-linux-melb:xfs-kern:30500a
Signed-off-by: NDavid Chinner <dgc@sgi.com>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

SGI-PV: 971186
SGI-Modid: xfs-linux-melb:xfs-kern:30490a
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>
Signed-off-by: NDonald Douwsma <donaldd@sgi.com>

Remove the xfs_refcache, it was only needed while we were still
building for 2.4 kernels.

SGI-PV: 971186
SGI-Modid: xfs-linux-melb:xfs-kern:30472a
Signed-off-by: NDonald Douwsma <donaldd@sgi.com>
Signed-off-by: NChristoph Hellwig <hch@infradead.org>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

On a forced shutdown, xfs_finish_reclaim() will skip flushing the inode.
If the inode flush lock is not already held and there is an outstanding
xfs_iflush_done() then we might free the inode prematurely. By acquiring
and releasing the flush lock we will synchronise with xfs_iflush_done().

SGI-PV: 909874
SGI-Modid: xfs-linux-melb:xfs-kern:30468a
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>
Signed-off-by: NDavid Chinner <dgc@sgi.com>

bailout if fails.

SGI-PV: 973041
SGI-Modid: xfs-linux-melb:xfs-kern:30462a
Signed-off-by: NNiv Sardi <xaiki@sgi.com>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

Since older kernels may look in the sb_bad_features2 slot for flags,
rather than zeroing it out on fixup, we should make it equal to the
sb_features2 value.

Also, if the ATTR2 flag was not found prior to features2 fixup, it was not
set in the mount flags, so re-check after the fixup so that the current
session will use the feature.

Also fix up the comments to reflect these changes.

SGI-PV: 980085
SGI-Modid: xfs-linux-melb:xfs-kern:30778a
Signed-off-by: NEric Sandeen <sandeen@sandeen.net>
Signed-off-by: NDavid Chinner <dgc@sgi.com>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

Due to the xfs_dsb_t structure not being 64 bit aligned, the last field of
the on-disk superblock can vary in location This causes problems when the
filesystem gets moved to a different platform, or there is a 32 bit
userspace and 64 bit kernel.

This patch detects the defect at mount time, logs a warning such as:

XFS: correcting sb_features alignment problem

in dmesg and corrects the problem so that everything is OK. it also
blacklists the bad field in the superblock so it does not get used for
something else later on.

SGI-PV: 977636
SGI-Modid: xfs-linux-melb:xfs-kern:30539a
Signed-off-by: NDavid Chinner <dgc@sgi.com>
Signed-off-by: NChristoph Hellwig <hch@infradead.org>
Signed-off-by: NEric Sandeen <sandeen@sandeen.net>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

Remove macro-to-small-function indirection from xfs_sb.h, and remove some
which are completely unused.

SGI-PV: 976035
SGI-Modid: xfs-linux-melb:xfs-kern:30528a
Signed-off-by: NEric Sandeen <sandeen@sandeen.net>
Signed-off-by: NDonald Douwsma <donaldd@sgi.com>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>

If the radix_tree_preload() fails, we need to destroy the inode we just
read in before trying again. This could leak xfs_vnode structures when
there is memory pressure. Noticed by Christoph Hellwig.

SGI-PV: 977823
SGI-Modid: xfs-linux-melb:xfs-kern:30606a
Signed-off-by: NDavid Chinner <dgc@sgi.com>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>
Signed-off-by: NChristoph Hellwig <hch@infradead.org>

wakeups

Idle state is not being detected properly by the xfsaild push code. The
current idle state is detected by an empty list which may never happen
with mostly idle filesystem or one using lazy superblock counters. A
single dirty item in the list that exists beyond the push target can
result repeated looping attempting to push up to the target because it
fails to check if the push target has been acheived or not.

Fix by considering a dirty list with everything past the target as an idle
state and set the timeout appropriately.

SGI-PV: 977545
SGI-Modid: xfs-linux-melb:xfs-kern:30532a
Signed-off-by: NDavid Chinner <dgc@sgi.com>
Signed-off-by: NChristoph Hellwig <hch@infradead.org>
Signed-off-by: NLachlan McIlroy <lachlan@sgi.com>