If we get an IO error on a synchronous superblock write, we attach an
error release function to it so that when the last reference goes away
the release function is called and the buffer is invalidated and
unlocked. The buffer is left locked until the release function is
called so that other concurrent users of the buffer will be locked out
until the buffer error is fully processed.

Unfortunately, for the superblock buffer the filesyetm itself holds a
reference to the buffer which prevents the reference count from
dropping to zero and the release function being called. As a result,
once an IO error occurs on a sync write, the buffer will never be
unlocked and all future attempts to lock the buffer will hang.

To make matters worse, this problems is not unique to such buffers;
if there is a concurrent _xfs_buf_find() running, the lookup will grab
a reference to the buffer and then wait on the buffer lock, preventing
the reference count from ever falling to zero and hence unlocking the
buffer.

As such, the whole b_relse function implementation is broken because it
cannot rely on the buffer reference count falling to zero to unlock the
errored buffer. The synchronous write error path is the only path that
uses this callback - it is used to ensure that the synchronous waiter
gets the buffer error before the error state is cleared from the buffer
by the release function.

Given that the only sychronous buffer writes now go through xfs_bwrite
and the error path in question can only occur for a write of a dirty,
logged buffer, we can move most of the b_relse processing to happen
inline in xfs_buf_iodone_callbacks, just like a normal I/O completion.
In addition to that we make sure the error is not cleared in
xfs_buf_iodone_callbacks, so that xfs_bwrite can reliably check it.
Given that xfs_bwrite keeps the buffer locked until it has waited for
it and checked the error this allows to reliably propagate the error
to the caller, and make sure that the buffer is reliably unlocked.

Given that xfs_buf_iodone_callbacks was the only instance of the
b_relse callback we can remove it entirely.

Based on earlier patches by Dave Chinner and Ajeet Yadav.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reported-by: Ajeet Yadav <ajeet.yadav.77@gmail.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Alex Elder <aelder@sgi.com>

Allow manual discards from userspace using the FITRIM ioctl.  This is not
intended to be run during normal workloads, as the freepsace btree walks
can cause large performance degradation.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Alex Elder <aelder@sgi.com>

To ensure the log is covered and the filesystem idles correctly, we
need to ensure that dummy transactions hit the disk and do not stay
pinned in memory.  If the superblock is pinned in memory, it can't
be flushed so the log covering cannot make progress. The result is
dependent on timing - more oftent han not we continue to issues a
log covering transaction every 36s rather than idling after ~90s.

Fix this by making the log covering transaction synchronous. To
avoid additional log force from xfssyncd, make the log covering
transaction take the place of the existing log force in the xfssyncd
background sync process.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>

When two concurrent unaligned, non-overlapping direct IOs are issued
to the same block, the direct Io layer will race to zero the block.
The result is that one of the concurrent IOs will overwrite data
written by the other IO with zeros. This is demonstrated by the
xfsqa test 240.

To avoid this problem, serialise all unaligned direct IOs to an
inode with a big hammer. We need a big hammer approach as we need to
serialise AIO as well, so we can't just block writes on locks.
Hence, the big hammer is calling xfs_ioend_wait() while holding out
other unaligned direct IOs from starting.

We don't bother trying to serialised aligned vs unaligned IOs as
they are overlapping IO and the result of concurrent overlapping IOs
is undefined - the result of either IO is a valid result so we let
them race. Hence we only penalise unaligned IO, which already has a
major overhead compared to aligned IO so this isn't a major problem.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Alex Elder <aelder@sgi.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

The buffered IO and direct IO write paths share a common set of
checks and limiting code prior to issuing the write. Factor that
into a common helper function.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Alex Elder <aelder@sgi.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

Complete the split of the different write IO paths by splitting the
buffered IO write path out of xfs_file_aio_write(). This makes the
different mechanisms of the write patchs easier to follow.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Alex Elder <aelder@sgi.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

The current xfs_file_aio_write code is a mess of locking shenanigans
to handle the different locking requirements of buffered and direct
IO. Start to clean this up by disentangling the direct IO path from
the mess.

This also removes the failed direct IO fallback path to buffered IO.
XFS handles all direct IO cases without needing to fall back to
buffered IO, so we can safely remove this unused path. This greatly
simplifies the logic and locking needed in the write path.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

We need to obtain the i_mutex, i_iolock and i_ilock during the read
and write paths. Add a set of wrapper functions to neatly
encapsulate the lock ordering and shared/exclusive semantics to make
the locking easier to follow and get right.

Note that this changes some of the exclusive locking serialisation in
that serialisation will occur against the i_mutex instead of the
XFS_IOLOCK_EXCL. This does not change any behaviour, and it is
arguably more efficient to use the mutex for such serialisation than
the rw_sem.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Alex Elder <aelder@sgi.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Alex Elder <aelder@sgi.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

xfs_file_aio_write() only returns the error from synchronous
flushing of the data and inode if error == 0. At the point where
error is being checked, it is guaranteed to be > 0. Therefore any
errors returned by the data or fsync flush will never be returned.
Fix the checks so we overwrite the current error once and only if an
error really occurred.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Alex Elder <aelder@sgi.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

The only thing that the grant lock remains to protect is the grant head
manipulations when adding or removing space from the log. These calculations
are already based on atomic variables, so we can already update them safely
without locks. However, the grant head manpulations require atomic multi-step
calculations to be executed, which the algorithms currently don't allow.

To make these multi-step calculations atomic, convert the algorithms to
compare-and-exchange loops on the atomic variables. That is, we sample the old
value, perform the calculation and use atomic64_cmpxchg() to attempt to update
the head with the new value. If the head has not changed since we sampled it,
it will succeed and we are done. Otherwise, we rerun the calculation again from
a new sample of the head.

This allows us to remove the grant lock from around all the grant head space
manipulations, and that effectively removes the grant lock from the log
completely. Hence we can remove the grant lock completely from the log at this
point.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

The log grant ticket wait queues are currently protected by the log
grant lock.  However, the queues are functionally independent from
each other, and operations on them only require serialisation
against other queue operations now that all of the other log
variables they use are atomic values.

Hence, we can make them independent of the grant lock by introducing
new locks just to protect the lists operations. because the lists
are independent, we can use a lock per list and ensure that reserve
and write head queuing do not contend.

To ensure forced shutdowns work correctly in conjunction with the
new fast paths, ensure that we check whether the log has been shut
down in the grant functions once we hold the relevant spin locks but
before we go to sleep. This is needed to co-ordinate correctly with
the wakeups that are issued on the ticket queues so we don't leave
any processes sleeping on the queues during a shutdown.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

Convert the log grant heads to atomic64_t types in preparation for
converting the accounting algorithms to atomic operations. his patch
just converts the variables; the algorithmic changes are in a
separate patch for clarity.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

log->l_tail_lsn is currently protected by the log grant lock. The
lock is only needed for serialising readers against writers, so we
don't really need the lock if we make the l_tail_lsn variable an
atomic. Converting the l_tail_lsn variable to an atomic64_t means we
can start to peel back the grant lock from various operations.

Also, provide functions to safely crack an atomic LSN variable into
it's component pieces and to recombined the components into an
atomic variable. Use them where appropriate.

This also removes the need for explicitly holding a spinlock to read
the l_tail_lsn on 32 bit platforms.
Signed-off-by: Dave Chinner <dchinner@redhat.com>

log->l_last_sync_lsn is updated in only one critical spot - log
buffer Io completion - and is protected by the grant lock here. This
requires the grant lock to be taken for every log buffer IO
completion. Converting the l_last_sync_lsn variable to an atomic64_t
means that we do not need to take the grant lock in log buffer IO
completion to update it.

This also removes the need for explicitly holding a spinlock to read
the l_last_sync_lsn on 32 bit platforms.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

The xlog_grant_push_ail() currently takes the grant lock internally to sample
the tail lsn, last sync lsn and the reserve grant head. Most of the callers
already hold the grant lock but have to drop it before calling
xlog_grant_push_ail(). This is a left over from when the AIL tail pushing was
done in line and hence xlog_grant_push_ail had to drop the grant lock. AIL push
is now done in another thread and hence we can safely hold the grant lock over
the entire xlog_grant_push_ail call.

Push the grant lock outside of xlog_grant_push_ail() to simplify the locking
and synchronisation needed for tail pushing.  This will reduce traffic on the
grant lock by itself, but this is only one step in preparing for the complete
removal of the grant lock.

While there, clean up the formatting of xlog_grant_push_ail() to match the
rest of the XFS code.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

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: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

Prepare for switching the grant heads to atomic variables by
combining the two 32 bit values that make up the grant head into a
single 64 bit variable.  Provide wrapper functions to combine and
split the grant heads appropriately for calculations and use them as
necessary.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

The log grant space calculations are repeated for both write and
reserve grant heads. To make it simpler to convert the calculations
toa different algorithm, factor them so both the gratn heads use the
same calculation functions. Once this is done we can drop the
wrappers that are used in only a couple of place to update both
grant heads at once as they don't provide any particular value.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

Factor repeated debug code out of grant head manipulation functions into a
separate function. This removes ifdef DEBUG spagetti from the code and makes
the code easier to follow.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

The grant write and reserve queues use a roll-your-own double linked
list, so convert it to a standard list_head structure and convert
all the list traversals to use list_for_each_entry(). We can also
get rid of the XLOG_TIC_IN_Q flag as we can use the list_empty()
check to tell if the ticket is in a list or not.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

We now have two copies of AIL delete operations that are mostly
duplicate functionality. The single log item deletes can be
implemented via the bulk updates by turning xfs_trans_ail_delete()
into a simple wrapper. This removes all the duplicate delete
functionality and associated helpers.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

We now have two copies of AIL insert operations that are mostly
duplicate functionality. The single log item updates can be
implemented via the bulk updates by turning xfs_trans_ail_update()
into a simple wrapper. This removes all the duplicate insert
functionality and associated helpers.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

When inode buffer IO completes, usually all of the inodes are removed from the
AIL. This involves processing them one at a time and taking the AIL lock once
for every inode. When all CPUs are processing inode IO completions, this causes
excessive amount sof contention on the AIL lock.

Instead, change the way we process inode IO completion in the buffer
IO done callback. Allow the inode IO done callback to walk the list
of IO done callbacks and pull all the inodes off the buffer in one
go and then process them as a batch.

Once all the inodes for removal are collected, take the AIL lock
once and do a bulk removal operation to minimise traffic on the AIL
lock.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

To allow buffer iodone callbacks to consume multiple items off the
callback list, first we need to convert the xfs_buf_do_callbacks()
to consume items and always pull the next item from the head of the
list.

The means the item list walk is never dependent on knowing the
next item on the list and hence allows callbacks to remove items
from the list as well. This allows callbacks to do bulk operations
by scanning the list for identical callbacks, consuming them all
and then processing them in bulk, negating the need for multiple
callbacks of that type.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

The xfaild often tries to rest to wait for congestion to pass of for
IO to complete, but is regularly woken in tail-pushing situations.
In severe cases, the xfsaild is getting woken tens of thousands of
times a second. Reduce the number needless wakeups by only waking
the xfsaild if the new target is larger than the old one. Further
make short sleeps uninterruptible as they occur when the xfsaild has
decided it needs to back off to allow some IO to complete and being
woken early is counter-productive.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph 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: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

xfs_ail_delete() has a needlessly complex interface. It returns the log item
that was passed in for deletion (which the callers then assert is identical to
the one passed in), and callers of xfs_ail_delete() still need to invalidate
current traversal cursors.

Make xfs_ail_delete() return void, move the cursor invalidation inside it, and
clean up the callers just to use the log item pointer they passed in.

While cleaning up, remove the messy and unnecessary "/* ARGUSED */" comments
around all these functions.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

EFI/EFD interactions are protected from races by the AIL lock. They
are the only type of log items that require the the AIL lock to
serialise internal state, so they need to be separated from the AIL
lock before we can do bulk insert operations on the AIL.

To acheive this, convert the counter of the number of extents in the
EFI to an atomic so it can be safely manipulated by EFD processing
without locks. Also, convert the EFI state flag manipulations to use
atomic bit operations so no locks are needed to record state
changes. Finally, use the state bits to determine when it is safe to
free the EFI and clean up the code to do this neatly.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

XFS_EFI_CANCELED has not been set in the code base since
xfs_efi_cancel() was removed back in 2006 by commit
065d312e ("[XFS] Remove unused
iop_abort log item operation), and even then xfs_efi_cancel() was
never called. I haven't tracked it back further than that (beyond
git history), but it indicates that the handling of EFIs in
cancelled transactions has been broken for a long time.

Basically, when we get an IOP_UNPIN(lip, 1); call from
xfs_trans_uncommit() (i.e. remove == 1), if we don't free the log
item descriptor we leak it. Fix the behviour to be correct and kill
the XFS_EFI_CANCELED flag.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

Now that the buffer reclaim infrastructure can handle different reclaim
priorities for different types of buffers, reconnect the hooks in the
XFS code that has been sitting dormant since it was ported to Linux. This
should finally give use reclaim prioritisation that is on a par with the
functionality that Irix provided XFS 15 years ago.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

Introduce a per-buftarg LRU for memory reclaim to operate on. This
is the last piece we need to put in place so that we can fully
control the buffer lifecycle. This allows XFS to be responsibile for
maintaining the working set of buffers under memory pressure instead
of relying on the VM reclaim not to take pages we need out from
underneath us.

The implementation introduces a b_lru_ref counter into the buffer.
This is currently set to 1 whenever the buffer is referenced and so is used to
determine if the buffer should be added to the LRU or not when freed.
Effectively it allows lazy LRU initialisation of the buffer so we do not need
to touch the LRU list and locks in xfs_buf_find().

Instead, when the buffer is being released and we drop the last
reference to it, we check the b_lru_ref count and if it is none zero
we re-add the buffer reference and add the inode to the LRU. The
b_lru_ref counter is decremented by the shrinker, and whenever the
shrinker comes across a buffer with a zero b_lru_ref counter, if
released the LRU reference on the buffer. In the absence of a lookup
race, this will result in the buffer being freed.

This counting mechanism is used instead of a reference flag so that
it is simple to re-introduce buffer-type specific reclaim reference
counts to prioritise reclaim more effectively. We still have all
those hooks in the XFS code, so this will provide the infrastructure
to re-implement that functionality.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

Before we introduce per-buftarg LRU lists, split the shrinker
implementation into per-buftarg shrinker callbacks. At the moment
we wake all the xfsbufds to run the delayed write queues to free
the dirty buffers and make their pages available for reclaim.
However, with an LRU, we want to be able to free clean, unused
buffers as well, so we need to separate the xfsbufd from the
shrinker callbacks.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Alex Elder <aelder@sgi.com>

now that we are using RCU protection for the inode cache lookups,
the lock is only needed on the modification side. Hence it is not
necessary for the lock to be a rwlock as there are no read side
holders anymore. Convert it to a spin lock to reflect it's exclusive
nature.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Alex Elder <aelder@sgi.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

With delayed logging greatly increasing the sustained parallelism of inode
operations, the inode cache locking is showing significant read vs write
contention when inode reclaim runs at the same time as lookups. There is
also a lot more write lock acquistions than there are read locks (4:1 ratio)
so the read locking is not really buying us much in the way of parallelism.

To avoid the read vs write contention, change the cache to use RCU locking on
the read side. To avoid needing to RCU free every single inode, use the built
in slab RCU freeing mechanism. This requires us to be able to detect lookups of
freed inodes, so enѕure that ever freed inode has an inode number of zero and
the XFS_IRECLAIM flag set. We already check the XFS_IRECLAIM flag in cache hit
lookup path, but also add a check for a zero inode number as well.

We canthen convert all the read locking lockups to use RCU read side locking
and hence remove all read side locking.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Alex Elder <aelder@sgi.com>

Introduce RCU freeing of XFS inodes so that we can convert lookup
traversals to use rcu_read_lock() protection. This patch only
introduces the RCU freeing to minimise the potential conflicts with
mainline if this is merged into mainline via a VFS patchset. It
abuses the i_dentry list for the RCU callback structure because the
VFS patches make this a union so it is safe to use like this and
simplifies and merge issues.

This patch uses basic RCU freeing rather than SLAB_DESTROY_BY_RCU.
The later lookup patches need the same "found free inode" protection
regardless of the RCU freeing method used, so once again the RCU
freeing method can be dealt with apprpriately at merge time without
affecting any other code.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>

A long standing problem for streaming writeѕ through the NFS server
has been that the NFS server opens and closes file descriptors on an
inode for every write. The result of this behaviour is that the
->release() function is called on every close and that results in
XFS truncating speculative preallocation beyond the EOF.  This has
an adverse effect on file layout when multiple files are being
written at the same time - they interleave their extents and can
result in severe fragmentation.

To avoid this problem, keep track of ->release calls made on a dirty
inode. For most cases, an inode is only going to be opened once for
writing and then closed again during it's lifetime in cache. Hence
if there are multiple ->release calls when the inode is dirty, there
is a good chance that the inode is being accessed by the NFS server.
Hence set a flag the first time ->release is called while there are
delalloc blocks still outstanding on the inode.

If this flag is set when ->release is next called, then do no
truncate away the speculative preallocation - leave it there so that
subsequent writes do not need to reallocate the delalloc space. This
will prevent interleaving of extents of different inodes written
concurrently to the same AG.

If we get this wrong, it is not a big deal as we truncate
speculative allocation beyond EOF anyway in xfs_inactive() when the
inode is thrown out of the cache.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

Currently the size of the speculative preallocation during delayed
allocation is fixed by either the allocsize mount option of a
default size. We are seeing a lot of cases where we need to
recommend using the allocsize mount option to prevent fragmentation
when buffered writes land in the same AG.

Rather than using a fixed preallocation size by default (up to 64k),
make it dynamic by basing it on the current inode size. That way the
EOF preallocation will increase as the file size increases.  Hence
for streaming writes we are much more likely to get large
preallocations exactly when we need it to reduce fragementation.

For default settings, the size of the initial extents is determined
by the number of parallel writers and the amount of memory in the
machine. For 4GB RAM and 4 concurrent 32GB file writes:

EXT: FILE-OFFSET           BLOCK-RANGE          AG AG-OFFSET                 TOTAL
   0: [0..1048575]:         1048672..2097247      0 (1048672..2097247)      1048576
   1: [1048576..2097151]:   5242976..6291551      0 (5242976..6291551)      1048576
   2: [2097152..4194303]:   12583008..14680159    0 (12583008..14680159)    2097152
   3: [4194304..8388607]:   25165920..29360223    0 (25165920..29360223)    4194304
   4: [8388608..16777215]:  58720352..67108959    0 (58720352..67108959)    8388608
   5: [16777216..33554423]: 117440584..134217791  0 (117440584..134217791) 16777208
   6: [33554424..50331511]: 184549056..201326143  0 (184549056..201326143) 16777088
   7: [50331512..67108599]: 251657408..268434495  0 (251657408..268434495) 16777088

and for 16 concurrent 16GB file writes:

 EXT: FILE-OFFSET           BLOCK-RANGE          AG AG-OFFSET                 TOTAL
   0: [0..262143]:          2490472..2752615      0 (2490472..2752615)       262144
   1: [262144..524287]:     6291560..6553703      0 (6291560..6553703)       262144
   2: [524288..1048575]:    13631592..14155879    0 (13631592..14155879)     524288
   3: [1048576..2097151]:   30408808..31457383    0 (30408808..31457383)    1048576
   4: [2097152..4194303]:   52428904..54526055    0 (52428904..54526055)    2097152
   5: [4194304..8388607]:   104857704..109052007  0 (104857704..109052007)  4194304
   6: [8388608..16777215]:  209715304..218103911  0 (209715304..218103911)  8388608
   7: [16777216..33554423]: 452984848..469762055  0 (452984848..469762055) 16777208

Because it is hard to take back specualtive preallocation, cases
where there are large slow growing log files on a nearly full
filesystem may cause premature ENOSPC. Hence as the filesystem nears
full, the maximum dynamic prealloc size іs reduced according to this
table (based on 4k block size):

freespace       max prealloc size
  >5%             full extent (8GB)
  4-5%             2GB (8GB >> 2)
  3-4%             1GB (8GB >> 3)
  2-3%           512MB (8GB >> 4)
  1-2%           256MB (8GB >> 5)
  <1%            128MB (8GB >> 6)

This should reduce the amount of space held in speculative
preallocation for such cases.

The allocsize mount option turns off the dynamic behaviour and fixes
the prealloc size to whatever the mount option specifies. i.e. the
behaviour is unchanged.
Signed-off-by: Dave Chinner <dchinner@redhat.com>

When listing attributes, we are doiing memory allocations under the
inode ilock using only KM_SLEEP. This allows memory allocation to
recurse back into the filesystem and do writeback, which may the
ilock we already hold on the current inode. THis will deadlock.
Hence use KM_NOFS for such allocations outside of transaction
context to ensure that reclaim recursion does not occur.
Reported-by: Nick Piggin <npiggin@gmail.com>
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>