Btrfs uses async helper threads to submit write bios so the checksumming
helper threads don't block on the disk.
The submit bio threads may process bios for more than one block device,
so when they find one device congested they try to move on to other
devices instead of blocking in get_request_wait for one device.
This does a pretty good job of keeping multiple devices busy, but the
congested flag has a number of problems. A congested device may still
give you a request, and other procs that aren't backing off the congested
device may starve you out.
This commit uses the io_context stored in current to decide if our process
has been made a batching process by the block layer. If so, it keeps
sending IO down for at least one batch. This helps make sure we do
a good amount of work each time we visit a bdev, and avoids large IO
stalls in multi-device workloads.
It's also very ugly. A better solution is in the works with Jens Axboe.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
COW means we cycle though blocks fairly quickly, and once we
free an extent on disk, it doesn't make much sense to keep the pages around.
This commit tries to immediately free the page when we free the extent,
which lowers our memory footprint significantly.
Signed-off-by: Chris 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: Chris Mason <chris.mason@oracle.com>
btrfs_update_delayed_ref is optimized to add and remove different
references in one pass through the delayed ref tree. It is a zero
sum on the total number of refs on a given extent.
But, the code was recording an extra ref in the head node. This
never made it down to the disk but was used when deciding if it was
safe to free the extent while dropping snapshots.
The fix used here is to make sure the ref_mod count is unchanged
on the head ref when btrfs_update_delayed_ref is called.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
The fsync log has code to make sure all of the parents of a file are in the
log along with the file. It uses a minimal log of the parent directory
inodes, just enough to get the parent directory on disk.
If the transaction that originally created a file is fully on disk,
and the file hasn't been renamed or linked into other directories, we
can safely skip the parent directory walk. We know the file is on disk
somewhere and we can go ahead and just log that single file.
This is more important now because unrelated unlinks in the parent directory
might make us force a commit if we try to log the parent.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
The tree logging code allows individual files or directories to be logged
without including operations on other files and directories in the FS.
It tries to commit the minimal set of changes to disk in order to
fsync the single file or directory that was sent to fsync or O_SYNC.
The tree logging code was allowing files and directories to be unlinked
if they were part of a rename operation where only one directory
in the rename was in the fsync log. This patch adds a few new rules
to the tree logging.
1) on rename or unlink, if the inode being unlinked isn't in the fsync
log, we must force a full commit before doing an fsync of the directory
where the unlink was done. The commit isn't done during the unlink,
but it is forced the next time we try to log the parent directory.
Solution: record transid of last unlink/rename per directory when the
directory wasn't already logged. For renames this is only done when
renaming to a different directory.
mkdir foo/some_dir
normal commit
rename foo/some_dir foo2/some_dir
mkdir foo/some_dir
fsync foo/some_dir/some_file
The fsync above will unlink the original some_dir without recording
it in its new location (foo2). After a crash, some_dir will be gone
unless the fsync of some_file forces a full commit
2) we must log any new names for any file or dir that is in the fsync
log. This way we make sure not to lose files that are unlinked during
the same transaction.
2a) we must log any new names for any file or dir during rename
when the directory they are being removed from was logged.
2a is actually the more important variant. Without the extra logging
a crash might unlink the old name without recreating the new one
3) after a crash, we must go through any directories with a link count
of zero and redo the rm -rf
mkdir f1/foo
normal commit
rm -rf f1/foo
fsync(f1)
The directory f1 was fully removed from the FS, but fsync was never
called on f1, only its parent dir. After a crash the rm -rf must
be replayed. This must be able to recurse down the entire
directory tree. The inode link count fixup code takes care of the
ugly details.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
During log replay, inodes are copied from the log to the main filesystem
btrees. Sometimes they have a zero link count in the log but they actually
gain links during the replay or have some in the main btree.
This patch updates the link count to be at least one after copying the
inode out of the log. This makes sure the inode is deleted during an
iput while the rest of the replay code is still working on it.
The log replay has fixup code to make sure that link counts are correct
at the end of the replay, so we could use any non-zero number here and
it would work fine.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
The delayed reference mechanism is responsible for all updates to the
extent allocation trees, including those updates created while processing
the delayed references.
This commit tries to limit the amount of work that gets created during
the final run of delayed refs before a commit. It avoids cowing new blocks
unless it is required to finish the commit, and so it avoids new allocations
that were not really required.
The goal is to avoid infinite loops where we are always making more work
on the final run of delayed refs. Over the long term we'll make a
special log for the last delayed ref updates as well.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
This reads in blocks in the checksum btree before starting the
transaction in btrfs_finish_ordered_io. It makes it much more likely
we'll be able to do operations inside the transaction without
needing any btree reads, which limits transaction latencies overall.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
btrfs_mark_buffer dirty would set dirty bits in the extent_io tree
for the buffers it was dirtying. This may require a kmalloc and it
was not atomic. So, anyone who called btrfs_mark_buffer_dirty had to
set any btree locks they were holding to blocking first.
This commit changes dirty tracking for extent buffers to just use a flag
in the extent buffer. Now that we have one and only one extent buffer
per page, this can be safely done without losing dirty bits along the way.
This also introduces a path->leave_spinning flag that callers of
btrfs_search_slot can use to indicate they will properly deal with a
path returned where all the locks are spinning instead of blocking.
Many of the btree search callers now expect spinning paths,
resulting in better btree concurrency overall.
Signed-off-by: Chris 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: Chris Mason <chris.mason@oracle.com>
This reduces contention on the extent buffer spin locks by testing for a
blocking lock before trying to take the spinlock.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
The fs/btrfs/inode.c code to run delayed allocation during writout
needed some stack usage optimization. This is the first pass, it does
the check for compression earlier on, which allows us to do the common
(no compression) case higher up in the call chain.
Signed-off-by: Chris 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: Chris 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: Chris Mason <chris.mason@oracle.com>
When extents are freed, it is likely that we've removed the last
delayed reference update for the extent. This checks the delayed
ref tree when things are freed, and if no ref updates area left it
immediately processes the delayed ref.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
Many of the tree balancing functions follow the same pattern.
1) cow a block
2) do something to the result
This commit breaks them up into two functions so the variables and
code required for part two don't suck down stack during part one.
Signed-off-by: Chris 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: Chris 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: Chris Mason <chris.mason@oracle.com>
The full flag on the space info structs tells the allocator not to try
and allocate more chunks because the devices in the FS are fully allocated.
When more devices are added, we need to clear the full flag so the allocator
knows it has more space available.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
Storage allocated to different raid levels in btrfs is tracked by
a btrfs_space_info structure, and all of the current space_infos are
collected into a list_head.
Most filesystems have 3 or 4 of these structs total, and the list is
only changed when new raid levels are added or at unmount time.
This commit adds rcu locking on the list head, and properly frees
things at unmount time. It also clears the space_info->full flag
whenever new space is added to the FS.
The locking for the space info list goes like this:
reads: protected by rcu_read_lock()
writes: protected by the chunk_mutex
At unmount time we don't need special locking because all the readers
are gone.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
btrfs_tree_locked was being used to make sure a given extent_buffer was
properly locked in a few places. But, it wasn't correct for UP compiled
kernels.
This switches it to using assert_spin_locked instead, and renames it to
btrfs_assert_tree_locked to better reflect how it was really being used.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
This fixes a problem where we could return -ENOSPC when we may actually have
plenty of space, the space is just pinned. Instead of returning -ENOSPC
immediately, commit the transaction first and then try and do the allocation
again.
This patch also does chunk allocation for metadata if we pass the 80%
threshold for metadata space. This will help with stack usage since the chunk
allocation will happen early on, instead of when the allocation is happening.
Signed-off-by: Josef Bacik <jbacik@redhat.com>
This is a step in the direction of better -ENOSPC handling. Instead of
checking the global bytes counter we check the space_info bytes counters to
make sure we have enough space.
If we don't we go ahead and try to allocate a new chunk, and then if that fails
we return -ENOSPC. This patch adds two counters to btrfs_space_info,
bytes_delalloc and bytes_may_use.
bytes_delalloc account for extents we've actually setup for delalloc and will
be allocated at some point down the line.
bytes_may_use is to keep track of how many bytes we may use for delalloc at
some point. When we actually set the extent_bit for the delalloc bytes we
subtract the reserved bytes from the bytes_may_use counter. This keeps us from
not actually being able to allocate space for any delalloc bytes.
Signed-off-by: Josef Bacik <jbacik@redhat.com>
btrfs_record_root_in_trans needs the trans_mutex held to make sure two
callers don't race to setup the root in a given transaction. This adds
it to all the places that were missing it.
Signed-off-by: Yan Zheng <zheng.yan@oracle.com>
Btrfs is currently using spin_lock_nested with a nested value based
on the tree depth of the block. But, this doesn't quite work because
the max tree depth is bigger than what spin_lock_nested can deal with,
and because locks are sometimes taken before the level field is filled in.
The solution here is to use lockdep_set_class_and_name instead, and to
set the class before unlocking the pages when the block is read from the
disk and just after init of a freshly allocated tree block.
btrfs_clear_path_blocking is also changed to take the locks in the proper
order, and it also makes sure all the locks currently held are properly
set to blocking before it tries to retake the spinlocks. Otherwise, lockdep
gets upset about bad lock orderin.
The lockdep magic cam from Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Chris Mason <chris.mason@oracle.com>
The call to kzalloc is followed by a kmalloc whose result is stored in the
same variable.
The semantic match that finds the problem is as follows:
(http://www.emn.fr/x-info/coccinelle/)
// <smpl>
@r exists@
local idexpression x;
statement S;
expression E;
identifier f,l;
position p1,p2;
expression *ptr != NULL;
@@
(
if ((x@p1 = \(kmalloc\|kzalloc\|kcalloc\)(...)) == NULL) S
|
x@p1 = \(kmalloc\|kzalloc\|kcalloc\)(...);
...
if (x == NULL) S
)
<... when != x
when != if (...) { <+...x...+> }
x->f = E
...>
(
return \(0\|<+...x...+>\|ptr\);
|
return@p2 ...;
)
@script:python@
p1 << r.p1;
p2 << r.p2;
@@
print "* file: %s kmalloc %s return %s" % (p1[0].file,p1[0].line,p2[0].line)
// </smpl>
Signed-off-by: Julia Lawall <julia@diku.dk>
Signed-off-by: Chris Mason <chris.mason@oracle.com>
btrfs_init_path was initially used when the path objects were on the
stack. Now all the work is done by btrfs_alloc_path and btrfs_init_path
isn't required.
This patch removes it, and just uses kmem_cache_zalloc to zero out the object.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
btrfs_releasepage may call kmem_cache_alloc indirectly,
and provide same GFP flags it gets to kmem_cache_alloc.
So it's possible to use __GFP_HIGHMEM with the slab
allocator.
Signed-off-by: Yan Zheng <zheng.yan@oracle.com>
Cleaning old snapshots can make sync(1) somewhat slow, and some users
and applications still use it in a global fsync kind of workload.
This patch changes btrfs not to clean old snapshots during sync, which is
safe from a FS consistency point of view. The major downside is that it
makes it difficult to tell when old snapshots have been reaped and
the space they were using has been reclaimed. A new ioctl will be added
for this purpose instead.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
Larger metadata clusters can significantly improve writeback performance
on ssd drives with large erasure blocks. The larger clusters make it
more likely a given IO will completely overwrite the ssd block, so it
doesn't have to do an internal rwm cycle.
On spinning media, lager metadata clusters end up spreading out the
metadata more over time, which makes fsck slower, so we don't want this
to be the default.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
Btrfs wasn't parsing any new mount options during remount, making it
difficult to set mount options on a root drive.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
Theres a slight problem with finish_current_insert, if we set all to 1 and then
go through and don't actually skip any of the extents on the pending list, we
could exit right after we've added new extents.
This is a problem because by inserting the new extents we could have gotten new
COW's to happen and such, so we may have some pending updates to do or even
more inserts to do after that.
So this patch will only exit if we have never skipped any of the extents in the
pending list, and we have no extents to insert, this will make sure that all of
the pending work is truly done before we return. I've been running with this
patch for a few days with all of my other testing and have not seen issues.
Thanks,
Signed-off-by: Josef Bacik <jbacik@redhat.com>
Btrfs was using spin_is_contended to see if it should drop locks before
doing extent allocations during btrfs_search_slot. The idea was to avoid
expensive searches in the tree unless the lock was actually contended.
But, spin_is_contended is specific to the ticket spinlocks on x86, so this
is causing compile errors everywhere else.
In practice, the contention could easily appear some time after we started
doing the extent allocation, and it makes more sense to always drop the lock
instead.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
The S_ISGID check in btrfs_new_inode caused an oops during subvol creation
because sometimes the dir is null.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
On fast devices that go from congested to uncongested very quickly, pdflush
is waiting too often in congestion_wait, and the FS is backing off to
easily in write_cache_pages.
For now, fix this on the btrfs side by only checking congestion after
some bios have already gone down. Longer term a real fix is needed
for pdflush, but that is a larger project.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
Whenever an item deletion is done, we need to balance all the nodes
in the tree to make sure we don't end up with an empty node if a pointer
is deleted. This balance prep happens from the root of the tree down
so we can drop our locks as we go.
reada_for_balance was triggering read-ahead on neighboring nodes even
when no balancing was required. This adds an extra check to avoid
calling balance_level() and avoid reada_for_balance() when a balance
won't be required.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
btrfs_unlock_up_safe would break out at the first NULL node entry or
unlocked node it found in the path.
Some of the callers have missing nodes at the lower levels of the path, so this
commit fixes things to check all the nodes in the path before returning.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
btrfs_del_leaf does two things. First it removes the pointer in the
parent, and then it frees the block that has the leaf. It has the
parent node locked for both operations.
But, it only needs the parent locked while it is deleting the pointer.
After that it can safely free the block without the parent locked.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
btrfs_truncate_inode_items is setup to stop doing btree searches when
it has finished removing the items for the inode. It used to detect the
end of the inode by looking for an objectid that didn't match the
one we were searching for.
But, this would result in an extra search through the btree, which
adds extra balancing and cow costs to the operation.
This commit adds a check to see if we found the inode item, which means
we can stop searching early.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
The compression code had some checks to make sure we were only
compressing bytes inside of i_size, but it wasn't catching every
case. To make things worse, some incorrect math about the number
of bytes remaining would make it try to compress more pages than the
file really had.
The fix used here is to fall back to the non-compression code in this
case, which does all the proper cleanup of delalloc and other accounting.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
With selinux on we end up calling __btrfs_setxattr when we create an inode,
which calls btrfs_start_transaction(). The problem is we've already called
that in btrfs_new_inode, and in btrfs_start_transaction we end up doing a
wait_current_trans(). If btrfs-transaction has started committing it will wait
for all handles to finish, while the other process is waiting for the
transaction to commit. This is fixed by using btrfs_join_transaction, which
won't wait for the transaction to commit. Thanks,
Signed-off-by: Josef Bacik <jbacik@redhat.com>
Before this patch, new files/dirs would ignore the SGID bit on their
parent directory and always be owned by the creating user's uid/gid.
Signed-off-by: Chris Ball <cjb@laptop.org>
Signed-off-by: Chris Mason <chris.mason@oracle.com>
Every transaction in btrfs creates a new snapshot, and then schedules the
snapshot from the last transaction for deletion. Snapshot deletion
works by walking down the btree and dropping the reference counts
on each btree block during the walk.
If if a given leaf or node has a reference count greater than one,
the reference count is decremented and the subtree pointed to by that
node is ignored.
If the reference count is one, walking continues down into that node
or leaf, and the references of everything it points to are decremented.
The old code would try to work in small pieces, walking down the tree
until it found the lowest leaf or node to free and then returning. This
was very friendly to the rest of the FS because it didn't have a huge
impact on other operations.
But it wouldn't always keep up with the rate that new commits added new
snapshots for deletion, and it wasn't very optimal for the extent
allocation tree because it wasn't finding leaves that were close together
on disk and processing them at the same time.
This changes things to walk down to a level 1 node and then process it
in bulk. All the leaf pointers are sorted and the leaves are dropped
in order based on their extent number.
The extent allocation tree and commit code are now fast enough for
this kind of bulk processing to work without slowing the rest of the FS
down. Overall it does less IO and is better able to keep up with
snapshot deletions under high load.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
Most of the btrfs metadata operations can be protected by a spinlock,
but some operations still need to schedule.
So far, btrfs has been using a mutex along with a trylock loop,
most of the time it is able to avoid going for the full mutex, so
the trylock loop is a big performance gain.
This commit is step one for getting rid of the blocking locks entirely.
btrfs_tree_lock takes a spinlock, and the code explicitly switches
to a blocking lock when it starts an operation that can schedule.
We'll be able get rid of the blocking locks in smaller pieces over time.
Tracing allows us to find the most common cause of blocking, so we
can start with the hot spots first.
The basic idea is:
btrfs_tree_lock() returns with the spin lock held
btrfs_set_lock_blocking() sets the EXTENT_BUFFER_BLOCKING bit in
the extent buffer flags, and then drops the spin lock. The buffer is
still considered locked by all of the btrfs code.
If btrfs_tree_lock gets the spinlock but finds the blocking bit set, it drops
the spin lock and waits on a wait queue for the blocking bit to go away.
Much of the code that needs to set the blocking bit finishes without actually
blocking a good percentage of the time. So, an adaptive spin is still
used against the blocking bit to avoid very high context switch rates.
btrfs_clear_lock_blocking() clears the blocking bit and returns
with the spinlock held again.
btrfs_tree_unlock() can be called on either blocking or spinning locks,
it does the right thing based on the blocking bit.
ctree.c has a helper function to set/clear all the locked buffers in a
path as blocking.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
Before metadata is written to disk, it is updated to reflect that writeout
has begun. Once this update is done, the block must be cow'd before it
can be modified again.
This update was originally synchronized by using a per-fs spinlock. Today
the buffers for the metadata blocks are locked before writeout begins,
and everyone that tests the flag has the buffer locked as well.
So, the per-fs spinlock (called hash_lock for no good reason) is no
longer required.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
