The addition of direct formatting of log items into the CIL
linear buffer added alignment restrictions that the start of each
vector needed to be 64 bit aligned. Hence padding was added in
xlog_finish_iovec() to round up the vector length to ensure the next
vector started with the correct alignment.

This adds a small number of bytes to the size of
the linear buffer that is otherwise unused. The issue is that we
then use the linear buffer size to determine the log space used by
the log item, and this includes the unused space. Hence when we
account for space used by the log item, it's more than is actually
written into the iclogs, and hence we slowly leak this space.

This results on log hangs when reserving space, with threads getting
stuck with these stack traces:

Call Trace:
[<ffffffff81d15989>] schedule+0x29/0x70
[<ffffffff8150d3a2>] xlog_grant_head_wait+0xa2/0x1a0
[<ffffffff8150d55d>] xlog_grant_head_check+0xbd/0x140
[<ffffffff8150ee33>] xfs_log_reserve+0x103/0x220
[<ffffffff814b7f05>] xfs_trans_reserve+0x2f5/0x310
.....

The 4 bytes is significant. Brain Foster did all the hard work in
tracking down a reproducable leak to inode chunk allocation (it went
away with the ikeep mount option). His rough numbers were that
creating 50,000 inodes leaked 11 log blocks. This turns out to be
roughly 800 inode chunks or 1600 inode cluster buffers. That
works out at roughly 4 bytes per cluster buffer logged, and at that
I started looking for a 4 byte leak in the buffer logging code.

What I found was that a struct xfs_buf_log_format structure for an
inode cluster buffer is 28 bytes in length. This gets rounded up to
32 bytes, but the vector length remains 28 bytes. Hence the CIL
ticket reservation is decremented by 32 bytes (via lv->lv_buf_len)
for that vector rather than 28 bytes which are written into the log.

The fix for this problem is to separately track the bytes used by
the log vectors in the item and use that instead of the buffer
length when accounting for the log space that will be used by the
formatted log item.

Again, thanks to Brian Foster for doing all the hard work and long
hours to isolate this leak and make finding the bug relatively
simple.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Log forces can occur deep in the call chain when we have relatively
little stack free. Log forces can also happen at close to the call
chain leaves (e.g. xfs_buf_lock()) and hence we can trigger IO from
places where we really don't want to add more stack overhead.

This stack overhead occurs because log forces do foreground CIL
pushes (xlog_cil_push_foreground()) rather than waking the
background push wq and waiting for the for the push to complete.
This foreground push was done to avoid confusing the CFQ Io
scheduler when fsync()s were issued, as it has trouble dealing with
dependent IOs being issued from different process contexts.

Avoiding blowing the stack is much more critical than performance
optimisations for CFQ, especially as we've been recommending against
the use of CFQ for XFS since 3.2 kernels were release because of
it's problems with multi-threaded IO workloads.

Hence convert xlog_cil_push_foreground() to move the push work
to the CIL workqueue. We already do the waiting for the push to
complete in xlog_cil_force_lsn(), so there's nothing else we need to
modify to make this work.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

On 32 bit platforms, the log item vector headers are not 64 bit
aligned or sized. hence if we don't take care to align them
correctly or pad the buffer appropriately for 8 byte alignment, we
can end up with alignment issues when accessing the user buffer
directly as a structure.

To solve this, simply pad the buffer headers to 64 bit offset so
that the data section is always 8 byte aligned.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reported-by: NMichael L. Semon <mlsemon35@gmail.com>
Tested-by: NMichael L. Semon <mlsemon35@gmail.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Convert xfs_log_commit_cil() to a void function since it return nothing
but 0 in any case, after that we can simplify the relative code logic
in xfs_trans_commit() accordingly.
Signed-off-by: NJie Liu <jeff.liu@oracle.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Instead of setting up pointers to memory locations in iop_format which then
get copied into the CIL linear buffer after return move the copy into
the individual inode items.  This avoids the need to always have a memory
block in the exact same layout that gets written into the log around, and
allow the log items to be much more flexible in their in-memory layouts.

The only caveat is that we need to properly align the data for each
iovec so that don't have structures misaligned in subsequent iovecs.

Note that all log item format routines now need to be careful to modify
the copy of the item that was placed into the CIL after calls to
xlog_copy_iovec instead of the in-memory copy.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Share code that was previously duplicated in two branches.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

xfs_trans.h has a dependency on xfs_log.h for a couple of
structures. Most code that does transactions doesn't need to know
anything about the log, but this dependency means that they have to
include xfs_log.h. Decouple the xfs_trans.h and xfs_log.h header
files and clean up the includes to be in dependency order.

In doing this, remove the direct include of xfs_trans_reserve.h from
xfs_trans.h so that we remove the dependency between xfs_trans.h and
xfs_mount.h. Hence the xfs_trans.h include can be moved to the
indicate the actual dependencies other header files have on it.

Note that these are kernel only header files, so this does not
translate to any userspace changes at all.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBen Myers <bpm@sgi.com>
Signed-off-by: NBen Myers <bpm@sgi.com>

All of the buffer operations structures are needed to be exported
for xfs_db, so move them all to a common location rather than
spreading them all over the place. They are verifying the on-disk
format, so while xfs_format.h might be a good place, it is not part
of the on disk format.

Hence we need to create a new header file that we centralise these
related definitions. Start by moving the bffer operations
structures, and then also move all the other definitions that have
crept into xfs_log_format.h and xfs_format.h as there was no other
shared header file to put them in.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NBen Myers <bpm@sgi.com>

Recent analysis of a deadlocked XFS filesystem from a kernel
crash dump indicated that the filesystem was stuck waiting for log
space. The short story of the hang on the RHEL6 kernel is this:

	- the tail of the log is pinned by an inode
	- the inode has been pushed by the xfsaild
	- the inode has been flushed to it's backing buffer and is
	  currently flush locked and hence waiting for backing
	  buffer IO to complete and remove it from the AIL
	- the backing buffer is marked for write - it is on the
	  delayed write queue
	- the inode buffer has been modified directly and logged
	  recently due to unlinked inode list modification
	- the backing buffer is pinned in memory as it is in the
	  active CIL context.
	- the xfsbufd won't start buffer writeback because it is
	  pinned
	- xfssyncd won't force the log because it sees the log as
	  needing to be covered and hence wants to issue a dummy
	  transaction to move the log covering state machine along.

Hence there is no trigger to force the CIL to the log and hence
unpin the inode buffer and therefore complete the inode IO, remove
it from the AIL and hence move the tail of the log along, allowing
transactions to start again.

Mainline kernels also have the same deadlock, though the signature
is slightly different - the inode buffer never reaches the delayed
write lists because xfs_buf_item_push() sees that it is pinned and
hence never adds it to the delayed write list that the xfsaild
flushes.

There are two possible solutions here. The first is to simply force
the log before trying to cover the log and so ensure that the CIL is
emptied before we try to reserve space for the dummy transaction in
the xfs_log_worker(). While this might work most of the time, it is
still racy and is no guarantee that we don't get stuck in
xfs_trans_reserve waiting for log space to come free. Hence it's not
the best way to solve the problem.

The second solution is to modify xfs_log_need_covered() to be aware
of the CIL. We only should be attempting to cover the log if there
is no current activity in the log - covering the log is the process
of ensuring that the head and tail in the log on disk are identical
(i.e. the log is clean and at idle). Hence, by definition, if there
are items in the CIL then the log is not at idle and so we don't
need to attempt to cover it.

When we don't need to cover the log because it is active or idle, we
issue a log force from xfs_log_worker() - if the log is idle, then
this does nothing.  However, if the log is active due to there being
items in the CIL, it will force the items in the CIL to the log and
unpin them.

In the case of the above deadlock scenario, instead of
xfs_log_worker() getting stuck in xfs_trans_reserve() attempting to
cover the log, it will instead force the log, thereby unpinning the
inode buffer, allowing IO to be issued and complete and hence
removing the inode that was pinning the tail of the log from the
AIL. At that point, everything will start moving along again. i.e.
the xfs_log_worker turns back into a watchdog that can alleviate
deadlocks based around pinned items that prevent the tail of the log
from being moved...
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NEric Sandeen <sandeen@redhat.com>
Signed-off-by: NBen Myers <bpm@sgi.com>

The xc_cil_lock is used for two purposes - to protect the CIL
itself, and to protect the push/commit state and lists. These are
two logically separate structures and operations, so can have their
own locks. This means that pushing on the CIL and the commit wait
ordering won't contend for a lock with other transactions that are
completing concurrently. As the CIL insertion is the hottest path
throught eh CIL, this is a big win.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NMark Tinguely <tinguely@sgi.com>
Signed-off-by: NBen Myers <bpm@sgi.com>

Now that all the log item preparation and formatting is done under
the CIL lock, we can get rid of the intermediate log vector chain
used to track items to be inserted into the CIL.

We can already find all the items to be committed from the
transaction handle, so as long as we attach the log vectors to the
item before we insert the items into the CIL, we don't need to
create a log vector chain to pass around.

This means we can move all the item insertion code into and optimise
it into a pair of simple passes across all the items in the
transaction. The first pass does the formatting and accounting, the
second inserts them all into the CIL.

We keep this two pass split so that we can separate the CIL
insertion - which must be done under the CIL spinlock - from the
formatting. We could insert each item into the CIL with a single
pass, but that massively increases the number of times we have to
grab the CIL spinlock. It is much more efficient (and hence
scalable) to do a batch operation and insert all objects in a single
lock grab.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NMark Tinguely <tinguely@sgi.com>
Signed-off-by: NBen Myers <bpm@sgi.com>

Now that we have the size of the log vector that has been allocated,
we can determine if we need to allocate a new log vector for
formatting and insertion. We only need to allocate a new vector if
it won't fit into the existing buffer.

However, we need to hold the CIL context lock while we do this so
that we can't race with a push draining the currently queued log
vectors. It is safe to do this as long as we do GFP_NOFS allocation
to avoid avoid memory allocation recursing into the filesystem.
Hence we can safely overwrite the existing log vector on the CIL if
it is large enough to hold all the dirty regions of the current
item.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NMark Tinguely <tinguely@sgi.com>
Signed-off-by: NBen Myers <bpm@sgi.com>

Now that we have the size of the object before the formatting pass
is called, we can allocation the log vector and it's buffer in a
single allocation rather than two separate allocations.

Store the size of the allocated buffer in the log vector so that
we potentially avoid allocation for future modifications of the
object.

While touching this code, remove the IOP_FORMAT definition.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NMark Tinguely <tinguely@sgi.com>
Signed-off-by: NBen Myers <bpm@sgi.com>

To begin optimising the CIL commit process, we need to have IOP_SIZE
return both the number of vectors and the size of the data pointed
to by the vectors. This enables us to calculate the size ofthe
memory allocation needed before the formatting step and reduces the
number of memory allocations per item by one.

While there, kill the IOP_SIZE macro.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NMark Tinguely <tinguely@sgi.com>
Signed-off-by: NBen Myers <bpm@sgi.com>

And "ordered log vector" is a log vector that is used for
tracking a log item through the CIL and into the AIL as part of the
log checkpointing. These ordered log vectors are special in that
they are not written to to journal in any way, and are not accounted
to the checkpoint being written.

The reason for this behaviour is to allow operations to attach items
to transactions and have them follow the normal transactional
lifecycle without actually having to write them to the journal. This
allows logging of items that track high level logical changes and
writing them to the log, while the physical items being modified
pass through into the AIL and pin the tail of the log (and therefore
the logical item in the log) until all the modified items are
physically written to disk.

IOWs, it allows us to write metadata without physically logging
every individual change but still maintain the full transactional
integrity guarantees we currently have w.r.t. crash recovery.

This change modifies some of the CIL item insertion loops, as
ordered log vectors introduce some new constraints as they don't
track any data. One advantage of this change is that it combines
two log vector chain walks into a single pass, so there is less
overhead in the transaction commit pass as well. It also kills some
unused code in the log vector walk loop when committing the CIL.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NMark Tinguely <tinguely@sgi.com>
Signed-off-by: NBen Myers <bpm@sgi.com>

There are several places where we use KM_SLEEP allocation contexts
and use the fact that they are called from transaction context to
add KM_NOFS where appropriate. Unfortunately, there are several
places where the code makes this assumption but can be called from
outside transaction context but with filesystem locks held. These
places need explicit KM_NOFS annotations to avoid lockdep
complaining about reclaim contexts.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBen Myers <bpm@sgi.com>
Signed-off-by: NBen Myers <bpm@sgi.com>

(cherry picked from commit ac14876c)

There are several places where we use KM_SLEEP allocation contexts
and use the fact that they are called from transaction context to
add KM_NOFS where appropriate. Unfortunately, there are several
places where the code makes this assumption but can be called from
outside transaction context but with filesystem locks held. These
places need explicit KM_NOFS annotations to avoid lockdep
complaining about reclaim contexts.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBen Myers <bpm@sgi.com>
Signed-off-by: NBen Myers <bpm@sgi.com>

xfs_log_commit_iclog() function has been removed by commits 93b8a585:
	xfs: remove the deprecated nodelaylog option

Beginning from Linux 3.3, only delayed logging is supported so that
we call xfs_log_commit_cil() at xfs_trans_commit() only, remove the
useless comments so.
Signed-off-by: NJie Liu <jeff.liu@oracle.com>
Reviewed-by: NMark Tinguely <tinguely@sgi.com>
Signed-off-by: NBen Myers <bpm@sgi.com>

Rename the XFS log structure to xlog to help crash distinquish it from the
other logs in Linux.
Signed-off-by: NMark Tinguely <tinguely@sgi.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NBen Myers <bpm@sgi.com>

Rename the XFS log structure to xlog to help crash distinquish it from the
other logs in Linux.
Signed-off-by: NMark Tinguely <tinguely@sgi.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NBen Myers <bpm@sgi.com>

With the removal of xfs_rw.h and other changes over time, xfs_bit.h
is being included in many files that don't actually need it. Clean
up the includes as necessary.

Also move the only-used-once xfs_ialloc_find_free() static inline
function out of a header file that is widely included to reduce
the number of needless dependencies on xfs_bit.h.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NMark Tinguely <tinguely@sgi.com>
Signed-off-by: NBen Myers <bpm@sgi.com>

Now that the busy extent tracking has been moved out of the
allocation files, clean up the namespace it uses to
"xfs_extent_busy" rather than a mix of "xfs_busy" and
"xfs_alloc_busy".

Signed-off-by: Dave Chinner<dchinner@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NMark Tinguely <tinguely@sgi.com>
Signed-off-by: NBen Myers <bpm@sgi.com>

To make it easier to handle userspace code merges, move all the busy
extent handling out of the allocation code and into it's own file.
The userspace code does not need the busy extent code, so this
simplifies the merging of the kernel code into the userspace
xfsprogs library.

Because the busy extent code has been almost completely rewritten
over the past couple of years, also update the copyright on this new
file to include the authors that made all those changes.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NMark Tinguely <tinguely@sgi.com>
Signed-off-by: NBen Myers <bpm@sgi.com>

Untangle the header file includes a bit by moving the definition of
xfs_agino_t to xfs_types.h. This removes the dependency that xfs_ag.h has on
xfs_inum.h, meaning we don't need to include xfs_inum.h everywhere we include
xfs_ag.h.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NMark Tinguely <tinguely@sgi.com>
Signed-off-by: NBen Myers <bpm@sgi.com>

Doing background CIL flushes adds significant latency to whatever
async transaction that triggers it. To avoid blocking async
transactions on things like waiting for log buffer IO to complete,
move the CIL push off into a workqueue.  By moving the push work
into a workqueue, we remove all the latency that the commit adds
from the foreground transaction commit path. This also means that
single threaded workloads won't do the CIL push procssing, leaving
them more CPU to do more async transactions.

To do this, we need to keep track of the sequence number we have
pushed work for. This avoids having many transaction commits
attempting to schedule work for the same sequence, and ensures that
we only ever have one push (background or forced) in progress at a
time. It also means that we don't need to take the CIL lock in write
mode to check for potential background push races, which reduces
lock contention.

To avoid potential issues with "smart" IO schedulers, don't use the
workqueue for log force triggered flushes. Instead, do them directly
so that the log IO is done directly by the process issuing the log
force and so doesn't get stuck on IO elevator queue idling
incorrectly delaying the log IO from the workqueue.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NMark Tinguely <tinguely@sgi.com>
Signed-off-by: NBen Myers <bpm@sgi.com>

Outside the now removed nodelaylog code this field is only used for
asserts and can be safely removed now.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NBen Myers <bpm@sgi.com>

Now that the nodelaylog mode is gone we can simplify the transaction commit
path a bit by removing the xfs_trans_commit_cil routine.  Restoring the
process flags is merged into xfs_trans_commit which already does it for
the error path, and allocating the log vectors is merged into
xlog_cil_format_items, which already fills them with data, thus avoiding
one loop over all log items.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NBen Myers <bpm@sgi.com>

The delaylog mode has been the default for a long time, and the nodelaylog
option has been scheduled for removal in Linux 3.3.  Remove it and code
only used by it now that we have opened the 3.3 window.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NBen Myers <bpm@sgi.com>

The below patch fixes some typos in various parts of the kernel, as well as fixes some comments.
Please let me know if I missed anything, and I will try to get it changed and resent.
Signed-off-by: NJustin P. Mattock <justinmattock@gmail.com>
Acked-by: NRandy Dunlap <rdunlap@xenotime.net>
Signed-off-by: NJiri Kosina <jkosina@suse.cz>

Now that we have reliably tracking of deleted extents in a
transaction we can easily implement "online" discard support
which calls blkdev_issue_discard once a transaction commits.

The actual discard is a two stage operation as we first have
to mark the busy extent as not available for reuse before we
can start the actual discard.  Note that we don't bother
supporting discard for the non-delaylog mode.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NAlex Elder <aelder@sgi.com>

Instead of finding the per-ag and then taking and releasing the pagb_lock
for every single busy extent completed sort the list of busy extents and
only switch betweens AGs where nessecary.  This becomes especially important
with the online discard support which will hit this lock more often.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NAlex Elder <aelder@sgi.com>

Failure to commit a transaction into the CIL is not handled
correctly. This currently can only happen when racing with a
shutdown and requires an explicit shutdown check, so it rare and can
be avoided. Remove the shutdown check and make the CIL commit a void
function to indicate it will always succeed, thereby removing the
incorrectly handled failure case.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NAlex Elder <aelder@sgi.com>

The kmemleak detector shows this after test 139:

unreferenced object 0xffff880079b88bb0 (size 264):
  comm "xfs_io", pid 4904, jiffies 4294909382 (age 276.824s)
  hex dump (first 32 bytes):
    00 00 00 00 ad 4e ad de ff ff ff ff 00 00 00 00  .....N..........
    ff ff ff ff ff ff ff ff 48 7b c9 82 ff ff ff ff  ........H{......
  backtrace:
    [<ffffffff81afb04d>] kmemleak_alloc+0x2d/0x60
    [<ffffffff8115c6cf>] kmem_cache_alloc+0x13f/0x2b0
    [<ffffffff814aaa97>] kmem_zone_alloc+0x77/0xf0
    [<ffffffff814aab2e>] kmem_zone_zalloc+0x1e/0x50
    [<ffffffff8148f394>] xlog_ticket_alloc+0x34/0x170
    [<ffffffff81494444>] xlog_cil_push+0xa4/0x3f0
    [<ffffffff81494eca>] xlog_cil_force_lsn+0x15a/0x160
    [<ffffffff814933a5>] _xfs_log_force_lsn+0x75/0x2d0
    [<ffffffff814a264d>] _xfs_trans_commit+0x2bd/0x2f0
    [<ffffffff8148bfdd>] xfs_iomap_write_allocate+0x1ad/0x350
    [<ffffffff814ac17f>] xfs_map_blocks+0x21f/0x370
    [<ffffffff814ad1b7>] xfs_vm_writepage+0x1c7/0x550
    [<ffffffff8112200a>] __writepage+0x1a/0x50
    [<ffffffff81122df2>] write_cache_pages+0x1c2/0x4c0
    [<ffffffff81123117>] generic_writepages+0x27/0x30
    [<ffffffff814aba5d>] xfs_vm_writepages+0x5d/0x80

By inspection, the leak occurs when xlog_write() returns and error
and we jump to the abort path without dropping the reference on the
active ticket.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NAlex Elder <aelder@sgi.com>

The log grant queues are one of the few places left using sv_t
constructs for waiting. Given we are touching this code, we should
convert them to plain wait queues. While there, convert all the
other sv_t users in the log code as well.

Seeing as this removes the last users of the sv_t type, remove the
header file defining the wrapper and the fragments that still
reference it.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>

When inserting items into the AIL from the transaction committed
callbacks, we take the AIL lock for every single item that is to be
inserted. For a CIL checkpoint commit, this can be tens of thousands
of individual inserts, yet almost all of the items will be inserted
at the same point in the AIL because they have the same index.

To reduce the overhead and contention on the AIL lock for such
operations, introduce a "bulk insert" operation which allows a list
of log items with the same LSN to be inserted in a single operation
via a list splice. To do this, we need to pre-sort the log items
being committed into a temporary list for insertion.

The complexity is that not every log item will end up with the same
LSN, and not every item is actually inserted into the AIL. Items
that don't match the commit LSN will be inserted and unpinned as per
the current one-at-a-time method (relatively rare), while items that
are not to be inserted will be unpinned and freed immediately. Items
that are to be inserted at the given commit lsn are placed in a
temporary array and inserted into the AIL in bulk each time the
array fills up.

As a result of this, we trade off AIL hold time for a significant
reduction in traffic. lock_stat output shows that the worst case
hold time is unchanged, but contention from AIL inserts drops by an
order of magnitude and the number of lock traversal decreases
significantly.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>

When commiting a transaction, we do a lock CIL state lock round trip
on every single log vector we insert into the CIL. This is resulting
in the lock being as hot as the inode and dcache locks on 8-way
create workloads. Rework the insertion loops to bring the number
of lock round trips to one per transaction for log vectors, and one
more do the busy extents.

Also change the allocation of the log vector buffer not to zero it
as we copy over the entire allocated buffer anyway.

This patch also includes a structural cleanup to the CIL item
insertion provided by Christoph Hellwig.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NAlex Elder <aelder@sgi.com>

I have been seeing occasional pauses in transaction throughput up to
30s long under heavy parallel workloads. The only notable thing was
that the xfsaild was trying to be active during the pauses, but
making no progress. It was running exactly 20 times a second (on the
50ms no-progress backoff), and the number of pushbuf events was
constant across this time as well.  IOWs, the xfsaild appeared to be
stuck on buffers that it could not push out.

Further investigation indicated that it was trying to push out inode
buffers that were pinned and/or locked. The xfsbufd was also getting
woken at the same frequency (by the xfsaild, no doubt) to push out
delayed write buffers. The xfsbufd was not making any progress
because all the buffers in the delwri queue were pinned. This scan-
and-make-no-progress dance went one in the trace for some seconds,
before the xfssyncd came along an issued a log force, and then
things started going again.

However, I noticed something strange about the log force - there
were way too many IO's issued. 516 log buffers were written, to be
exact. That added up to 129MB of log IO, which got me very
interested because it's almost exactly 25% of the size of the log.
He delayed logging code is suppose to aggregate the minimum of 25%
of the log or 8MB worth of changes before flushing. That's what
really puzzled me - why did a log force write 129MB instead of only
8MB?

Essentially what has happened is that no CIL pushes had occurred
since the previous tail push which cleared out 25% of the log space.
That caused all the new transactions to block because there wasn't
log space for them, but they kick the xfsaild to push the tail.
However, the xfsaild was not making progress because there were
buffers it could not lock and flush, and the xfsbufd could not flush
them because they were pinned. As a result, both the xfsaild and the
xfsbufd could not move the tail of the log forward without the CIL
first committing.

The cause of the problem was that the background CIL push, which
should happen when 8MB of aggregated changes have been committed, is
being held off by the concurrent transaction commit load. The
background push does a down_write_trylock() which will fail if there
is a concurrent transaction commit holding the push lock in read
mode. With 8 CPUs all doing transactions as fast as they can, there
was enough concurrent transaction commits to hold off the background
push until tail-pushing could no longer free log space, and the halt
would occur.

It should be noted that there is no reason why it would halt at 25%
of log space used by a single CIL checkpoint. This bug could
definitely violate the "no transaction should be larger than half
the log" requirement and hence result in corruption if the system
crashed under heavy load. This sort of bug is exactly the reason why
delayed logging was tagged as experimental....

The fix is to start blocking background pushes once the threshold
has been exceeded. Rework the threshold calculations to keep the
amount of log space a CIL checkpoint can use to below that of the
AIL push threshold to avoid the problem completely.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NAlex Elder <aelder@sgi.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>

Formatting items requires memory allocation when using delayed
logging. Currently that memory allocation is done while holding the
CIL context lock in read mode. This means that if memory allocation
takes some time (e.g. enters reclaim), we cannot push on the CIL
until the allocation(s) required by formatting complete. This can
stall CIL pushes for some time, and once a push is stalled so are
all new transaction commits.

Fix this splitting the item formatting into two steps. The first
step which does the allocation and memcpy() into the allocated
buffer is now done outside the CIL context lock, and only the CIL
insert is done inside the CIL context lock. This avoids the stall
issue.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>

Delayed logging adds some serialisation to the log force process to
ensure that it does not deference a bad commit context structure
when determining if a CIL push is necessary or not. It does this by
grabing the CIL context lock exclusively, then dropping it before
pushing the CIL if necessary. This causes serialisation of all log
forces and pushes regardless of whether a force is necessary or not.
As a result fsync heavy workloads (like dbench) can be significantly
slower with delayed logging than without.

To avoid this penalty, copy the current sequence from the context to
the CIL structure when they are swapped. This allows us to do
unlocked checks on the current sequence without having to worry
about dereferencing context structures that may have already been
freed. Hence we can remove the CIL context locking in the forcing
code and only call into the push code if the current context matches
the sequence we need to force.

By passing the sequence into the push code, we can check the
sequence again once we have the CIL lock held exclusive and abort if
the sequence has already been pushed. This avoids a lock round-trip
and unnecessary CIL pushes when we have racing push calls.

The result is that the regression in dbench performance goes away -
this change improves dbench performance on a ramdisk from ~2100MB/s
to ~2500MB/s. This compares favourably to not using delayed logging
which retuns ~2500MB/s for the same workload.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>

When we commit a transaction using delayed logging, we need to
unlock the items in the transaciton before we unlock the CIL context
and allow it to be checkpointed. If we unlock them after we release
the CIl context lock, the CIL can checkpoint and complete before
we free the log items. This breaks stale buffer item unlock and
unpin processing as there is an implicit assumption that the unlock
will occur before the unpin.

Also, some log items need to store the LSN of the transaction commit
in the item (inodes and EFIs) and so can race with other transaction
completions if we don't prevent the CIL from checkpointing before
the unlock occurs.

Cc: <stable@kernel.org>
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>