With multi-threaded writes we were getting ENOSPC early because somebody would
come in, start flushing delalloc because they couldn't make their reservation,
and in the meantime other threads would come in and use the space that was
getting freed up, so when the original thread went to check to see if they had
space they didn't and they'd return ENOSPC.  So instead if we have some free
space but not enough for our reservation, take the reservation and then start
doing the flushing.  The only time we don't take reservations is when we've
already overcommitted our space, that way we don't have people who come late to
the party way overcommitting ourselves.  This also moves all of the retrying and
flushing code into reserve_metdata_bytes so it's all uniform.  This keeps my
fs_mark test from returning -ENOSPC as soon as it starts and actually lets me
fill up the disk.  Thanks,
Signed-off-by: NJosef Bacik <josef@redhat.com>

Yan Zheng noticed two places we were doing a lot of work
without task->state set to TASK_RUNNING.  This sets the state
properly after we get ready to sleep but decide not to.
Signed-off-by: NChris Mason <chris.mason@oracle.com>

This patch adds metadata ENOSPC handling for the balance code.
It is consisted by following major changes:

1. Avoid COW tree leave in the phrase of merging tree.

2. Handle interaction with snapshot creation.

3. make the backref cache can live across transactions.
Signed-off-by: NYan Zheng <zheng.yan@oracle.com>
Signed-off-by: NChris Mason <chris.mason@oracle.com>

reserve metadata space for handling orphan inodes
Signed-off-by: NYan Zheng <zheng.yan@oracle.com>
Signed-off-by: NChris Mason <chris.mason@oracle.com>

Reserve metadata space for extent tree, checksum tree and root tree
Signed-off-by: NYan Zheng <zheng.yan@oracle.com>
Signed-off-by: NChris Mason <chris.mason@oracle.com>

Besides simplify the code, this change makes sure all metadata
reservation for normal metadata operations are released after
committing transaction.

Changes since V1:

Add code that check if unlink and rmdir will free space.

Add ENOSPC handling for clone ioctl.
Signed-off-by: NYan Zheng <zheng.yan@oracle.com>
Signed-off-by: NChris Mason <chris.mason@oracle.com>

Introducing metadata reseravtion contexts has two major advantages.
First, it makes metadata reseravtion more traceable. Second, it can
reclaim freed space and re-add them to the itself after transaction
committed.

Besides add btrfs_block_rsv structure and related helper functions,
This patch contains following changes:

Move code that decides if freed tree block should be pinned into
btrfs_free_tree_block().

Make space accounting more accurate, mainly for handling read only
block groups.
Signed-off-by: NChris Mason <chris.mason@oracle.com>

This creates the reference to a new snapshot in the same commit as the
snapshot itself.  This avoids the need for a second commit in order for a
snapshot to be persistent, and also avoids the problem of "leaking" a
new snapshot tree root if the host crashes before the second commit takes
place.

It is not at all clear to me why it wasn't always done this way.  If there
is still a reason for the two-stage {create,finish}_pending_snapshots()
approach I'm missing something!  :)

I've been running this for a couple weeks under pretty heavy usage (a few
snapshots per minute) without obvious problems.
Signed-off-by: NSage Weil <sage@newdream.net>
Signed-off-by: NChris Mason <chris.mason@oracle.com>

We only need to call finish_wait() after wait loop.

By the way, this patch makes code of waiting loop similar to
example in wait.h(no functional change)
Signed-off-by: NZhao Lei <zhaolei@cn.fujitsu.com>
Signed-off-by: NMiao Xie <miaox@cn.fujitsu.com>
Signed-off-by: NChris Mason <chris.mason@oracle.com>

include cleanup: Update gfp.h and slab.h includes to prepare for breaking implicit slab.h inclusion from percpu.h

percpu.h is included by sched.h and module.h and thus ends up being
included when building most .c files.  percpu.h includes slab.h which
in turn includes gfp.h making everything defined by the two files
universally available and complicating inclusion dependencies.

percpu.h -> slab.h dependency is about to be removed.  Prepare for
this change by updating users of gfp and slab facilities include those
headers directly instead of assuming availability.  As this conversion
needs to touch large number of source files, the following script is
used as the basis of conversion.

  http://userweb.kernel.org/~tj/misc/slabh-sweep.py

The script does the followings.

* Scan files for gfp and slab usages and update includes such that
  only the necessary includes are there.  ie. if only gfp is used,
  gfp.h, if slab is used, slab.h.

* When the script inserts a new include, it looks at the include
  blocks and try to put the new include such that its order conforms
  to its surrounding.  It's put in the include block which contains
  core kernel includes, in the same order that the rest are ordered -
  alphabetical, Christmas tree, rev-Xmas-tree or at the end if there
  doesn't seem to be any matching order.

* If the script can't find a place to put a new include (mostly
  because the file doesn't have fitting include block), it prints out
  an error message indicating which .h file needs to be added to the
  file.

The conversion was done in the following steps.

1. The initial automatic conversion of all .c files updated slightly
   over 4000 files, deleting around 700 includes and adding ~480 gfp.h
   and ~3000 slab.h inclusions.  The script emitted errors for ~400
   files.

2. Each error was manually checked.  Some didn't need the inclusion,
   some needed manual addition while adding it to implementation .h or
   embedding .c file was more appropriate for others.  This step added
   inclusions to around 150 files.

3. The script was run again and the output was compared to the edits
   from #2 to make sure no file was left behind.

4. Several build tests were done and a couple of problems were fixed.
   e.g. lib/decompress_*.c used malloc/free() wrappers around slab
   APIs requiring slab.h to be added manually.

5. The script was run on all .h files but without automatically
   editing them as sprinkling gfp.h and slab.h inclusions around .h
   files could easily lead to inclusion dependency hell.  Most gfp.h
   inclusion directives were ignored as stuff from gfp.h was usually
   wildly available and often used in preprocessor macros.  Each
   slab.h inclusion directive was examined and added manually as
   necessary.

6. percpu.h was updated not to include slab.h.

7. Build test were done on the following configurations and failures
   were fixed.  CONFIG_GCOV_KERNEL was turned off for all tests (as my
   distributed build env didn't work with gcov compiles) and a few
   more options had to be turned off depending on archs to make things
   build (like ipr on powerpc/64 which failed due to missing writeq).

   * x86 and x86_64 UP and SMP allmodconfig and a custom test config.
   * powerpc and powerpc64 SMP allmodconfig
   * sparc and sparc64 SMP allmodconfig
   * ia64 SMP allmodconfig
   * s390 SMP allmodconfig
   * alpha SMP allmodconfig
   * um on x86_64 SMP allmodconfig

8. percpu.h modifications were reverted so that it could be applied as
   a separate patch and serve as bisection point.

Given the fact that I had only a couple of failures from tests on step
6, I'm fairly confident about the coverage of this conversion patch.
If there is a breakage, it's likely to be something in one of the arch
headers which should be easily discoverable easily on most builds of
the specific arch.
Signed-off-by: NTejun Heo <tj@kernel.org>
Guess-its-ok-by: NChristoph Lameter <cl@linux-foundation.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>

Flush any delalloc extents when we create a snapshot, so that recently
written file data is always included in the snapshot.

A later commit will add the ability to snapshot without the flush, but
most people expect flushing.
Signed-off-by: NSage Weil <sage@newdream.net>
Signed-off-by: NChris Mason <chris.mason@oracle.com>

btrfs inialize rb trees in quite a number of places by settin rb_node =
NULL;  The problem with this is that 17d9ddc7 in the
linux-next tree adds a new field to that struct which needs to be NULL for
the new rbtree library code to work properly.  This patch uses RB_ROOT as
the intializer so all of the relevant fields will be NULL'd.  Without the
patch I get a panic.
Signed-off-by: NEric Paris <eparis@redhat.com>
Acked-by: NVenkatesh Pallipadi <venkatesh.pallipadi@intel.com>
Signed-off-by: NChris Mason <chris.mason@oracle.com>

The bytes_used field in root item was originally planned to
trace the amount of used data and tree blocks. But it never
worked right since we can't trace freeing of data accurately.
This patch changes it to only trace the amount of tree blocks.
Signed-off-by: NYan Zheng <zheng.yan@oracle.com>
Signed-off-by: NChris Mason <chris.mason@oracle.com>

iput() can trigger new transactions if we are dropping the
final reference, so calling it in btrfs_commit_transaction
may end up deadlock. This patch adds delayed iput to avoid
the issue.
Signed-off-by: NYan Zheng <zheng.yan@oracle.com>
Signed-off-by: NChris Mason <chris.mason@oracle.com>

btrfs_lookup_dentry may trigger orphan cleanup, so it's not good
to call it while committing a transaction.
Signed-off-by: NYan Zheng <zheng.yan@oracle.com>
Signed-off-by: NChris Mason <chris.mason@oracle.com>

We allow two log transactions at a time, but use same flag
to mark dirty tree-log btree blocks. So we may flush dirty
blocks belonging to newer log transaction when committing a
log transaction. This patch fixes the issue by using two
flags to mark dirty tree-log btree blocks.
Signed-off-by: NYan Zheng <zheng.yan@oracle.com>
Signed-off-by: NChris Mason <chris.mason@oracle.com>

We use journal_info to tell if we're in a nested transaction to make sure we
don't commit the transaction within a nested transaction.  We use another
method to see if there are any outstanding ioctl trans handles, so if we're
starting one do not set current->journal_info, since it will screw with other
filesystems.  This patch also cleans up the starting stuff so there aren't any
magic numbers.
Signed-off-by: NJosef Bacik <josef@redhat.com>
Signed-off-by: NChris Mason <chris.mason@oracle.com>

Syncing the tree log is a 3 phase operation.

1) write and wait for all the tree log blocks for a given root.

2) write and wait for all the tree log blocks for the
tree of tree log roots.

3) write and wait for the super blocks (barriers here)

This isn't as efficient as it could be because there is
no requirement to wait for the blocks from step one to hit the disk
before we start writing the blocks from step two.  This commit
changes the sequence so that we don't start waiting until
all the tree blocks from both steps one and two have been sent
to disk.

We do this by breaking up btrfs_write_wait_marked_extents into
two functions, which is trivial because it was already broken
up into two parts.
Signed-off-by: NChris Mason <chris.mason@oracle.com>

At the start of a transaction we do a btrfs_reserve_metadata_space() and
specify how many items we plan on modifying.  Then once we've done our
modifications and such, just call btrfs_unreserve_metadata_space() for
the same number of items we reserved.

For keeping track of metadata needed for data I've had to add an extent_io op
for when we merge extents.  This lets us track space properly when we are doing
sequential writes, so we don't end up reserving way more metadata space than
what we need.

The only place where the metadata space accounting is not done is in the
relocation code.  This is because Yan is going to be reworking that code in the
near future, so running btrfs-vol -b could still possibly result in a ENOSPC
related panic.  This patch also turns off the metadata_ratio stuff in order to
allow users to more efficiently use their disk space.

This patch makes it so we track how much metadata we need for an inode's
delayed allocation extents by tracking how many extents are currently
waiting for allocation.  It introduces two new callbacks for the
extent_io tree's, merge_extent_hook and split_extent_hook.  These help
us keep track of when we merge delalloc extents together and split them
up.  Reservations are handled prior to any actually dirty'ing occurs,
and then we unreserve after we dirty.

btrfs_unreserve_metadata_for_delalloc() will make the appropriate
unreservations as needed based on the number of reservations we
currently have and the number of extents we currently have.  Doing the
reservation outside of doing any of the actual dirty'ing lets us do
things like filemap_flush() the inode to try and force delalloc to
happen, or as a last resort actually start allocation on all delalloc
inodes in the fs.  This has survived dbench, fs_mark and an fsx torture
test.
Signed-off-by: NJosef Bacik <jbacik@redhat.com>
Signed-off-by: NChris Mason <chris.mason@oracle.com>

This patch adds snapshot/subvolume destroy ioctl.  A subvolume that isn't being
used and doesn't contains links to other subvolumes can be destroyed.
Signed-off-by: NYan Zheng <zheng.yan@oracle.com>
Signed-off-by: NChris Mason <chris.mason@oracle.com>

btrfs allows subvolumes and snapshots anywhere in the directory tree.
If we snapshot a subvolume that contains a link to other subvolume
called subvolA, subvolA can be accessed through both the original
subvolume and the snapshot. This is similar to creating hard link to
directory, and has the very similar problems.

The aim of this patch is enforcing there is only one access point to
each subvolume. Only the first directory entry (the one added when
the subvolume/snapshot was created) is treated as valid access point.
The first directory entry is distinguished by checking root forward
reference. If the corresponding root forward reference is missing,
we know the entry is not the first one.

This patch also adds snapshot/subvolume rename support, the code
allows rename subvolume link across subvolumes.
Signed-off-by: NYan Zheng <zheng.yan@oracle.com>
Signed-off-by: NChris Mason <chris.mason@oracle.com>

This patch contains two changes to avoid unnecessary tree block reads during
snapshot dropping.

First, check tree block's reference count and flags before reading the tree
block. if reference count > 1 and there is no need to update backrefs, we can
avoid reading the tree block.

Second, save when snapshot was created in root_key.offset. we can compare block
pointer's generation with snapshot's creation generation during updating
backrefs. If a given block was created before snapshot was created, the
snapshot can't be the tree block's owner. So we can avoid reading the block.
Signed-off-by: NYan Zheng <zheng.yan@oracle.com>
Signed-off-by: NChris Mason <chris.mason@oracle.com>

This patch gets rid of two limitations of async block group caching.
The old code delays handling pinned extents when block group is in
caching. To allocate logged file extents, the old code need wait
until block group is fully cached. To get rid of the limitations,
This patch introduces a data structure to track the progress of
caching. Base on the caching progress, we know which extents should
be added to the free space cache when handling the pinned extents.
The logged file extents are also handled in a similar way.

This patch also changes how pinned extents are tracked. The old
code uses one tree to track pinned extents, and copy the pinned
extents tree at transaction commit time. This patch makes it use
two trees to track pinned extents. One tree for extents that are
pinned in the running transaction, one tree for extents that can
be unpinned. At transaction commit time, we swap the two trees.
Signed-off-by: NYan Zheng <zheng.yan@oracle.com>
Signed-off-by: NChris Mason <chris.mason@oracle.com>

The semaphore used by the async caching threads can prevent a
transaction commit, which can make the FS appear to stall.  This
releases the semaphore more often when a transaction commit is
in progress.
Signed-off-by: NChris Mason <chris.mason@oracle.com>

The async block group caching code uses the commit_root pointer
to get a stable version of the extent allocation tree for scanning.
This copy of the tree root isn't going to change and it significantly
reduces the complexity of the scanning code.

During a commit, we have a loop where we update the extent allocation
tree root.  We need to loop because updating the root pointer in
the tree of tree roots may allocate blocks which may change the
extent allocation tree.

Right now the commit_root pointer is changed inside this loop.  It
is more correct to change the commit_root pointer only after all the
looping is done.
Signed-off-by: NYan Zheng <zheng.yan@oracle.com>
Signed-off-by: NChris Mason <chris.mason@oracle.com>

The commit_transaction call to wait_ordered_extents when snap_pending
passes nocow_only=1 to process only NOCOW or PREALLOC extents.  This isn't
correct for the 'flushoncommit' mode, as it skips extents we just started
IO on in start_delalloc_inodes.

So, in the flushoncommit case, wait on all ordered extents.  Otherwise,
only pass the nocow_only flag to wait_ordered_extents if snap_pending.
Signed-off-by: NSage Weil <sage@newdream.net>
Signed-off-by: NChris Mason <chris.mason@oracle.com>

This patch moves the caching of the block group off to a kthread in order to
allow people to allocate sooner.  Instead of blocking up behind the caching
mutex, we instead kick of the caching kthread, and then attempt to make an
allocation.  If we cannot, we wait on the block groups caching waitqueue, which
the caching kthread will wake the waiting threads up everytime it finds 2 meg
worth of space, and then again when its finished caching.  This is how I tested
the speedup from this

mkfs the disk
mount the disk
fill the disk up with fs_mark
unmount the disk
mount the disk
time touch /mnt/foo

Without my changes this took 11 seconds on my box, with these changes it now
takes 1 second.

Another change thats been put in place is we lock the super mirror's in the
pinned extent map in order to keep us from adding that stuff as free space when
caching the block group.  This doesn't really change anything else as far as the
pinned extent map is concerned, since for actual pinned extents we use
EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock
those extents to keep from leaking memory.

I've also added a check where when we are reading block groups from disk, if the
amount of space used == the size of the block group, we go ahead and mark the
block group as cached.  This drastically reduces the amount of time it takes to
cache the block groups.  Using the same test as above, except doing a dd to a
file and then unmounting, it used to take 33 seconds to umount, now it takes 3
seconds.

This version uses the commit_root in the caching kthread, and then keeps track
of how many async caching threads are running at any given time so if one of the
async threads is still running as we cross transactions we can wait until its
finished before handling the pinned extents.  Thank you,
Signed-off-by: NJosef Bacik <jbacik@redhat.com>
Signed-off-by: NChris Mason <chris.mason@oracle.com>

Write dirty block groups may allocate new block, and so may add new delayed
back ref. btrfs_run_delayed_refs may make some block groups dirty.

commit_cowonly_roots does not handle the recursion properly, and some dirty
blocks can be left unwritten at commit time. This patch moves
btrfs_run_delayed_refs into the loop that writes dirty block groups, and makes
the code not break out of the loop until there are no dirty block groups or
delayed back refs.
Signed-off-by: NYan Zheng <zheng.yan@oracle.com>
Signed-off-by: NChris Mason <chris.mason@oracle.com>

The new backref format has restriction on type of backref item.  If a tree
block isn't referenced by its owner tree, full backrefs must be used for the
pointers in it. When a tree block loses its owner tree's reference, backrefs
for the pointers in it should be updated to full backrefs. Current
btrfs_drop_snapshot misses the code that updates backrefs, so it's unsafe for
general use.

This patch adds backrefs update code to btrfs_drop_snapshot.  It isn't a
problem in the restricted form btrfs_drop_snapshot is used today, but for
general snapshot deletion this update is required.
Signed-off-by: NYan Zheng <zheng.yan@oracle.com>
Signed-off-by: NChris Mason <chris.mason@oracle.com>

commit_fs_roots skips updating root items for fs trees that aren't modified.
This is unsafe now that relocation code modifies root item's last_snapshot
field without modifying corresponding fs tree.
Signed-off-by: NYan Zheng <zheng.yan@oracle.com>
Signed-off-by: NChris Mason <chris.mason@oracle.com>

This commit introduces a new kind of back reference for btrfs metadata.
Once a filesystem has been mounted with this commit, IT WILL NO LONGER
BE MOUNTABLE BY OLDER KERNELS.

When a tree block in subvolume tree is cow'd, the reference counts of all
extents it points to are increased by one.  At transaction commit time,
the old root of the subvolume is recorded in a "dead root" data structure,
and the btree it points to is later walked, dropping reference counts
and freeing any blocks where the reference count goes to 0.

The increments done during cow and decrements done after commit cancel out,
and the walk is a very expensive way to go about freeing the blocks that
are no longer referenced by the new btree root.  This commit reduces the
transaction overhead by avoiding the need for dead root records.

When a non-shared tree block is cow'd, we free the old block at once, and the
new block inherits old block's references. When a tree block with reference
count > 1 is cow'd, we increase the reference counts of all extents
the new block points to by one, and decrease the old block's reference count by
one.

This dead tree avoidance code removes the need to modify the reference
counts of lower level extents when a non-shared tree block is cow'd.
But we still need to update back ref for all pointers in the block.
This is because the location of the block is recorded in the back ref
item.

We can solve this by introducing a new type of back ref. The new
back ref provides information about pointer's key, level and in which
tree the pointer lives. This information allow us to find the pointer
by searching the tree. The shortcoming of the new back ref is that it
only works for pointers in tree blocks referenced by their owner trees.

This is mostly a problem for snapshots, where resolving one of these
fuzzy back references would be O(number_of_snapshots) and quite slow.
The solution used here is to use the fuzzy back references in the common
case where a given tree block is only referenced by one root,
and use the full back references when multiple roots have a reference
on a given block.

This commit adds per subvolume red-black tree to keep trace of cached
inodes. The red-black tree helps the balancing code to find cached
inodes whose inode numbers within a given range.

This commit improves the balancing code by introducing several data
structures to keep the state of balancing. The most important one
is the back ref cache. It caches how the upper level tree blocks are
referenced. This greatly reduce the overhead of checking back ref.

The improved balancing code scales significantly better with a large
number of snapshots.

This is a very large commit and was written in a number of
pieces.  But, they depend heavily on the disk format change and were
squashed together to make sure git bisect didn't end up in a
bad state wrt space balancing or the format change.
Signed-off-by: NYan Zheng <zheng.yan@oracle.com>
Signed-off-by: NChris Mason <chris.mason@oracle.com>

After a transaction commit, the old root of the subvol btrees are sent through
snapshot removal.  This is what actually frees up any blocks replaced by
COW, and anything the old blocks pointed to.

Snapshot deletion will pause when a transaction commit has started, which
helps to avoid a huge amount of delayed reference count updates piling up
as the transaction is trying to close.

But, this pause happens after the snapshot deletion process has asked other
procs on the system to throttle back a bit so that it can make progress.

We don't want to throttle everyone while we're waiting for the transaction
commit, it leads to deadlocks in the user transaction ioctls used by Ceph
and makes things slower in general.

This patch changes things to avoid the throttling while we sleep.
Signed-off-by: NChris Mason <chris.mason@oracle.com>

The 'flushoncommit' mount option forces any data dirtied by a write in a
prior transaction to commit as part of the current commit.  This makes
the committed state a fully consistent view of the file system from the
application's perspective (i.e., it includes all completed file system
operations).  This was previously the behavior only when a snapshot is
created.

This is used by Ceph to ensure that completed writes make it to the
platter along with the metadata operations they are bound to (by
BTRFS_IOC_TRANS_{START,END}).
Signed-off-by: NSage Weil <sage@newdream.net>
Signed-off-by: NChris Mason <chris.mason@oracle.com>

Because btrfs is copy-on-write, we end up picking new locations for
blocks very often.  This makes it fairly difficult to maintain perfect
read patterns over time, but we can at least do some optimizations
for writes.

This is done today by remembering the last place we allocated and
trying to find a free space hole big enough to hold more than just one
allocation.  The end result is that we tend to write sequentially to
the drive.

This happens all the time for metadata and it happens for data
when mounted -o ssd.  But, the way we record it is fairly racey
and it tends to fragment the free space over time because we are trying
to allocate fairly large areas at once.

This commit gets rid of the races by adding a free space cluster object
with dedicated locking to make sure that only one process at a time
is out replacing the cluster.

The free space fragmentation is somewhat solved by allowing a cluster
to be comprised of smaller free space extents.  This part definitely
adds some CPU time to the cluster allocations, but it allows the allocator
to consume the small holes left behind by cow.
Signed-off-by: NChris Mason <chris.mason@oracle.com>

Renames and truncates are both common ways to replace old data with new
data.  The filesystem can make an effort to make sure the new data is
on disk before actually replacing the old data.

This is especially important for rename, which many application use as
though it were atomic for both the data and the metadata involved.  The
current btrfs code will happily replace a file that is fully on disk
with one that was just created and still has pending IO.

If we crash after transaction commit but before the IO is done, we'll end
up replacing a good file with a zero length file.  The solution used
here is to create a list of inodes that need special ordering and force
them to disk before the commit is done.  This is similar to the
ext3 style data=ordering, except it is only done on selected files.

Btrfs is able to get away with this because it does not wait on commits
very often, even for fsync (which use a sub-commit).

For renames, we order the file when it wasn't already
on disk and when it is replacing an existing file.  Larger files
are sent to filemap_flush right away (before the transaction handle is
opened).

For truncates, we order if the file goes from non-zero size down to
zero size.  This is a little different, because at the time of the
truncate the file has no dirty bytes to order.  But, we flag the inode
so that it is added to the ordered list on close (via release method).  We
also immediately add it to the ordered list of the current transaction
so that we can try to flush down any writes the application sneaks in
before commit.
Signed-off-by: NChris Mason <chris.mason@oracle.com>

Commits are fairly expensive, and so btrfs has code to sit around for a while
during the commit and let new writers come in.

But, while we're sitting there, new delayed refs might be added, and those
can be expensive to process as well.  Unless the transaction is very very
young, it makes sense to go ahead and let the commit finish without hanging
around.

The commit grow loop isn't as important as it used to be, the fsync logging
code handles most performance critical syncs now.
Signed-off-by: NChris Mason <chris.mason@oracle.com>

To avoid deadlocks and reduce latencies during some critical operations, some
transaction writers are allowed to jump into the running transaction and make
it run a little longer, while others sit around and wait for the commit to
finish.

This is a bit unfair, especially when the callers that jump in do a bunch
of IO that makes all the others procs on the box wait.  This commit
reduces the stalls this produces by pre-reading file extent pointers
during btrfs_finish_ordered_io before the transaction is joined.

It also tunes the drop_snapshot code to politely wait for transactions
that have started writing out their delayed refs to finish.  This avoids
new delayed refs being flooded into the queue while we're trying to
close off the transaction.
Signed-off-by: NChris Mason <chris.mason@oracle.com>

The delayed reference queue maintains pending operations that need to
be done to the extent allocation tree.  These are processed by
finding records in the tree that are not currently being processed one at
a time.

This is slow because it uses lots of time searching through the rbtree
and because it creates lock contention on the extent allocation tree
when lots of different procs are running delayed refs at the same time.

This commit changes things to grab a cluster of refs for processing,
using a cursor into the rbtree as the starting point of the next search.
This way we walk smoothly through the rbtree.
Signed-off-by: NChris Mason <chris.mason@oracle.com>

The extent allocation tree maintains a reference count and full
back reference information for every extent allocated in the
filesystem.  For subvolume and snapshot trees, every time
a block goes through COW, the new copy of the block adds a reference
on every block it points to.

If a btree node points to 150 leaves, then the COW code needs to go
and add backrefs on 150 different extents, which might be spread all
over the extent allocation tree.

These updates currently happen during btrfs_cow_block, and most COWs
happen during btrfs_search_slot.  btrfs_search_slot has locks held
on both the parent and the node we are COWing, and so we really want
to avoid IO during the COW if we can.

This commit adds an rbtree of pending reference count updates and extent
allocations.  The tree is ordered by byte number of the extent and byte number
of the parent for the back reference.  The tree allows us to:

1) Modify back references in something close to disk order, reducing seeks
2) Significantly reduce the number of modifications made as block pointers
are balanced around
3) Do all of the extent insertion and back reference modifications outside
of the performance critical btrfs_search_slot code.

#3 has the added benefit of greatly reducing the btrfs stack footprint.
The extent allocation tree modifications are done without the deep
(and somewhat recursive) call chains used in the past.

These delayed back reference updates must be done before the transaction
commits, and so the rbtree is tied to the transaction.  Throttling is
implemented to help keep the queue of backrefs at a reasonable size.

Since there was a similar mechanism in place for the extent tree
extents, that is removed and replaced by the delayed reference tree.

Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code.
Signed-off-by: NChris Mason <chris.mason@oracle.com>

In order to avoid doing expensive extent management with tree locks held,
btrfs_search_slot will preallocate tree blocks for use by COW without
any tree locks held.

A later commit moves all of the extent allocation work for COW into
a delayed update mechanism, and this preallocation will no longer be
required.
Signed-off-by: NChris Mason <chris.mason@oracle.com>