Miao made the ordered operations stuff run async, which introduced a
deadlock where we could get somebody (sync) racing in and committing the
transaction while a commit was already happening. The new committer would
try and flush ordered operations which would hang waiting for the commit to
finish because it is done asynchronously and no longer inherits the callers
trans handle. To fix this we need to make the ordered operations list a per
transaction list. We can get new inodes added to the ordered operation list
by truncating them and then having another process writing to them, so this
makes it so that anybody trying to add an ordered operation _must_ start a
transaction in order to add itself to the list, which will keep new inodes
from getting added to the ordered operations list after we start committing.
This should fix the deadlock and also keeps us from doing a lot more work
than we need to during commit. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Since we don't actually copy the extent information from the source tree in
the fast case we don't need to wait for ordered io to be completed in order
to fsync, we just need to wait for the io to be completed. So when we're
logging our file just attach all of the ordered extents to the log, and then
when the log syncs just wait for IO_DONE on the ordered extents and then
write the super. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Pull btrfs update from Chris Mason:
"A big set of fixes and features.
In terms of line count, most of the code comes from Stefan, who added
the ability to replace a single drive in place. This is different
from how btrfs normally replaces drives, and is much much much faster.
Josef is plowing through our synchronous write performance. This pull
request does not include the DIO_OWN_WAITING patch that was discussed
on the list, but it has a number of other improvements to cut down our
latencies and CPU time during fsync/O_DIRECT writes.
Miao Xie has a big series of fixes and is spreading out ordered
operations over more CPUs. This improves performance and reduces
contention.
I've put in fixes for error handling around hash collisions. These
are going back to individual stable kernels as I test against them.
Otherwise we have a lot of fixes and cleanups, thanks everyone!
raid5/6 is being rebased against the device replacement code. I'll
have it posted this Friday along with a nice series of benchmarks."
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mason/linux-btrfs: (115 commits)
Btrfs: fix a bug of per-file nocow
Btrfs: fix hash overflow handling
Btrfs: don't take inode delalloc mutex if we're a free space inode
Btrfs: fix autodefrag and umount lockup
Btrfs: fix permissions of empty files not affected by umask
Btrfs: put raid properties into global table
Btrfs: fix BUG() in scrub when first superblock reading gives EIO
Btrfs: do not call file_update_time in aio_write
Btrfs: only unlock and relock if we have to
Btrfs: use tokens where we can in the tree log
Btrfs: optimize leaf_space_used
Btrfs: don't memset new tokens
Btrfs: only clear dirty on the buffer if it is marked as dirty
Btrfs: move checks in set_page_dirty under DEBUG
Btrfs: log changed inodes based on the extent map tree
Btrfs: add path->really_keep_locks
Btrfs: do not mark ems as prealloc if we are writing to them
Btrfs: keep track of the extents original block length
Btrfs: inline csums if we're fsyncing
Btrfs: don't bother copying if we're only logging the inode
...
Though the process of the ordered extents is a bit different with the delalloc inode
flush, but we can see it as a subset of the delalloc inode flush, so we also handle
them by flush workers.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
The process of the ordered operations is similar to the delalloc inode flush, so
we handle them by flush workers.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
"Whether" is misspelled in various comments across the tree; this
fixes them. No code changes.
Signed-off-by: Adam Buchbinder <adam.buchbinder@gmail.com>
Signed-off-by: Jiri Kosina <jkosina@suse.cz>
The ordered extent allocation is in the fast path of the IO, so use a slab
to improve the speed of the allocation.
"Size of the struct is 280, so this will fall into the size-512 bucket,
giving 8 objects per page, while own slab will pack 14 objects into a page.
Another benefit I see is to check for leaked objects when the module is
removed (and the cache destroy takes place)."
-- David Sterba
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
If a snapshot is created while we are writing some data into the file,
the i_size of the corresponding file in the snapshot will be wrong, it will
be beyond the end of the last file extent. And btrfsck will report:
root 256 inode 257 errors 100
Steps to reproduce:
# mkfs.btrfs <partition>
# mount <partition> <mnt>
# cd <mnt>
# dd if=/dev/zero of=tmpfile bs=4M count=1024 &
# for ((i=0; i<4; i++))
> do
> btrfs sub snap . $i
> done
This because the algorithm of disk_i_size update is wrong. Though there are
some ordered extents behind the current one which we use to update disk_i_size,
it doesn't mean those extents will be dealt with in the same transaction. So
We shouldn't use the offset of those extents to update disk_i_size. Or we will
get the wrong i_size in the snapshot.
We fix this problem by recording the max real i_size. If we find there is a
ordered extent which is in front of the current one and doesn't complete, we
will record the end of the current one into that ordered extent. Surely, if
the current extent holds the end of other extent(it must be greater than
the current one because it is behind the current one), we will record the
number that the current extent holds. In this way, we can exclude the ordered
extents that may not be dealth with in the same transaction, and be easy to
know the real disk_i_size.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
We noticed that the ordered extent completion doesn't really rely on having
a page and that it could be done independantly of ending the writeback on a
page. This patch makes us not do the threaded endio stuff for normal
buffered writes and direct writes so we can end page writeback as soon as
possible (in irq context) and only start threads to do the ordered work when
it is actually done. Compression needs to be reworked some to take
advantage of this as well, but atm it has to do a find_get_page in its endio
handler so it must be done in its own thread. This makes direct writes
quite a bit faster. Thanks,
Signed-off-by: Josef Bacik <josef@redhat.com>
Make the code aware of compression type, instead of always assuming
zlib compression.
Also make the zlib workspace function as common code for all
compression types.
Signed-off-by: Li Zefan <lizf@cn.fujitsu.com>
The new DIO bio splitting code has problems when the bio
spans more than one ordered extent. This will happen as the
generic DIO code merges our get_blocks calls together into
a bigger single bio.
This fixes things by walking forward in the ordered extent
code finding all the overlapping ordered extents and completing them
all at once.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
This provides basic DIO support for reading and writing. It does not do the
work to recover from mismatching checksums, that will come later. A few design
changes have been made from Jim's code (sorry Jim!)
1) Use the generic direct-io code. Jim originally re-wrote all the generic DIO
code in order to account for all of BTRFS's oddities, but thanks to that work it
seems like the best bet is to just ignore compression and such and just opt to
fallback on buffered IO.
2) Fallback on buffered IO for compressed or inline extents. Jim's code did
it's own buffering to make dio with compressed extents work. Now we just
fallback onto normal buffered IO.
3) Use ordered extents for the writes so that all of the
lock_extent()
lookup_ordered()
type checks continue to work.
4) Do the lock_extent() lookup_ordered() loop in readpage so we don't race with
DIO writes.
I've tested this with fsx and everything works great. This patch depends on my
dio and filemap.c patches to work. Thanks,
Signed-off-by: Josef Bacik <josef@redhat.com>
Signed-off-by: Chris Mason <chris.mason@oracle.com>
When finishing io we run btrfs_dec_test_ordered_pending, and then immediately
run btrfs_lookup_ordered_extent, but btrfs_dec_test_ordered_pending does that
already, so we're searching twice when we don't have to. This patch lets us
pass a btrfs_ordered_extent in to btrfs_dec_test_ordered_pending so if we do
complete io on that ordered extent we can just use the one we found then instead
of having to do another btrfs_lookup_ordered_extent. This made my fio job with
the other patch go from 24 mb/s to 29 mb/s.
Signed-off-by: Josef Bacik <josef@redhat.com>
Signed-off-by: Chris Mason <chris.mason@oracle.com>
The ordered tree used to need a mutex, but currently all we use it for is to
protect the rb_tree, and a spin_lock is just fine for that. Using a spin_lock
instead makes dbench run a little faster, 58 mb/s instead of 51 mb/s, and have
less latency, 3445.138 ms instead of 3820.633 ms.
Signed-off-by: Josef Bacik <josef@redhat.com>
Signed-off-by: Chris 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 17d9ddc72f 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: Eric Paris <eparis@redhat.com>
Acked-by: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
Signed-off-by: Chris 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: Yan Zheng <zheng.yan@oracle.com>
Signed-off-by: Chris Mason <chris.mason@oracle.com>
There are some cases file extents are inserted without involving
ordered struct. In these cases, we update disk_i_size directly,
without checking pending ordered extent and DELALLOC bit. This
patch extends btrfs_ordered_update_i_size() to handle these cases.
Signed-off-by: Yan Zheng <zheng.yan@oracle.com>
Signed-off-by: Chris Mason <chris.mason@oracle.com>
Use filemap_fdatawrite_range and filemap_fdatawait_range instead of
local copies of the functions. For filemap_fdatawait_range that
also means replacing the awkward old wait_on_page_writeback_range
calling convention with the regular filemap byte offsets.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Chris Mason <chris.mason@oracle.com>
Btrfs writes go through delalloc to the data=ordered code. This
makes sure that all of the data is on disk before the metadata
that references it. The tracking means that we have to make sure
each page in an extent is fully written before we add that extent into
the on-disk btree.
This was done in the past by setting the EXTENT_ORDERED bit for the
range of an extent when it was added to the data=ordered code, and then
clearing the EXTENT_ORDERED bit in the extent state tree as each page
finished IO.
One of the reasons we had to do this was because sometimes pages are
magically dirtied without page_mkwrite being called. The EXTENT_ORDERED
bit is checked at writepage time, and if it isn't there, our page become
dirty without going through the proper path.
These bit operations make for a number of rbtree searches for each page,
and can cause considerable lock contention.
This commit switches from the EXTENT_ORDERED bit to use PagePrivate2.
As pages go into the ordered code, PagePrivate2 is set on each one.
This is a cheap operation because we already have all the pages locked
and ready to go.
As IO finishes, the PagePrivate2 bit is cleared and the ordered
accoutning is updated for each page.
At writepage time, if the PagePrivate2 bit is missing, we go into the
writepage fixup code to handle improperly dirtied pages.
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 stores checksums for each data block. Until now, they have
been stored in the subvolume trees, indexed by the inode that is
referencing the data block. This means that when we read the inode,
we've probably read in at least some checksums as well.
But, this has a few problems:
* The checksums are indexed by logical offset in the file. When
compression is on, this means we have to do the expensive checksumming
on the uncompressed data. It would be faster if we could checksum
the compressed data instead.
* If we implement encryption, we'll be checksumming the plain text and
storing that on disk. This is significantly less secure.
* For either compression or encryption, we have to get the plain text
back before we can verify the checksum as correct. This makes the raid
layer balancing and extent moving much more expensive.
* It makes the front end caching code more complex, as we have touch
the subvolume and inodes as we cache extents.
* There is potentitally one copy of the checksum in each subvolume
referencing an extent.
The solution used here is to store the extent checksums in a dedicated
tree. This allows us to index the checksums by phyiscal extent
start and length. It means:
* The checksum is against the data stored on disk, after any compression
or encryption is done.
* The checksum is stored in a central location, and can be verified without
following back references, or reading inodes.
This makes compression significantly faster by reducing the amount of
data that needs to be checksummed. It will also allow much faster
raid management code in general.
The checksums are indexed by a key with a fixed objectid (a magic value
in ctree.h) and offset set to the starting byte of the extent. This
allows us to copy the checksum items into the fsync log tree directly (or
any other tree), without having to invent a second format for them.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
This patch updates btrfs-progs for fallocate support.
fallocate is a little different in Btrfs because we need to tell the
COW system that a given preallocated extent doesn't need to be
cow'd as long as there are no snapshots of it. This leverages the
-o nodatacow checks.
Signed-off-by: Yan Zheng <zheng.yan@oracle.com>
This patch simplifies the nodatacow checker. If all references
were created after the latest snapshot, then we can avoid COW
safely. This patch also updates run_delalloc_nocow to do more
fine-grained checking.
Signed-off-by: Yan Zheng <zheng.yan@oracle.com>
This is a large change for adding compression on reading and writing,
both for inline and regular extents. It does some fairly large
surgery to the writeback paths.
Compression is off by default and enabled by mount -o compress. Even
when the -o compress mount option is not used, it is possible to read
compressed extents off the disk.
If compression for a given set of pages fails to make them smaller, the
file is flagged to avoid future compression attempts later.
* While finding delalloc extents, the pages are locked before being sent down
to the delalloc handler. This allows the delalloc handler to do complex things
such as cleaning the pages, marking them writeback and starting IO on their
behalf.
* Inline extents are inserted at delalloc time now. This allows us to compress
the data before inserting the inline extent, and it allows us to insert
an inline extent that spans multiple pages.
* All of the in-memory extent representations (extent_map.c, ordered-data.c etc)
are changed to record both an in-memory size and an on disk size, as well
as a flag for compression.
From a disk format point of view, the extent pointers in the file are changed
to record the on disk size of a given extent and some encoding flags.
Space in the disk format is allocated for compression encoding, as well
as encryption and a generic 'other' field. Neither the encryption or the
'other' field are currently used.
In order to limit the amount of data read for a single random read in the
file, the size of a compressed extent is limited to 128k. This is a
software only limit, the disk format supports u64 sized compressed extents.
In order to limit the ram consumed while processing extents, the uncompressed
size of a compressed extent is limited to 256k. This is a software only limit
and will be subject to tuning later.
Checksumming is still done on compressed extents, and it is done on the
uncompressed version of the data. This way additional encodings can be
layered on without having to figure out which encoding to checksum.
Compression happens at delalloc time, which is basically singled threaded because
it is usually done by a single pdflush thread. This makes it tricky to
spread the compression load across all the cpus on the box. We'll have to
look at parallel pdflush walks of dirty inodes at a later time.
Decompression is hooked into readpages and it does spread across CPUs nicely.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
This reworks the btrfs O_DIRECT write code a bit. It had always fallen
back to buffered IO and done an invalidate, but needed to be updated
for the data=ordered code. The invalidate wasn't actually removing pages
because they were still inside an ordered extent.
This also combines the O_DIRECT/O_SYNC paths where possible, and kicks
off IO in the main btrfs_file_write loop to keep the pipe down the the
disk full as we process long writes.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
Before setting an extent to delalloc, the code needs to wait for
pending ordered extents.
Also, the relocation code needs to wait for ordered IO before scanning
the block group again. This is because the extents are not removed
until the IO for the new extents is finished
Signed-off-by: Chris Mason <chris.mason@oracle.com>
Stress testing was showing data checksum errors, most of which were caused
by a lookup bug in the extent_map tree. The tree was caching the last
pointer returned, and searches would check the last pointer first.
But, search callers also expect the search to return the very first
matching extent in the range, which wasn't always true with the last
pointer usage.
For now, the code to cache the last return value is just removed. It is
easy to fix, but I think lookups are rare enough that it isn't required anymore.
This commit also replaces do_sync_mapping_range with a local copy of the
related functions.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
Data checksumming is done right before the bio is sent down the IO stack,
which means a single bio might span more than one ordered extent. In
this case, the checksumming data is split between two ordered extents.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
Checksum items are not inserted until the entire ordered extent is on disk,
but individual pages might be clean and available for reclaim long before
the whole extent is on disk.
In order to allow those pages to be freed, we need to be able to search
the list of ordered extents to find the checksum that is going to be inserted
in the tree. This way if the page needs to be read back in before
the checksums are in the btree, we'll be able to verify the checksum on
the page.
This commit adds the ability to search the pending ordered extents for
a given offset in the file, and changes btrfs_releasepage to allow
ordered pages to be freed.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
This changes the ordered data code to update i_size after the extent
is on disk. An on disk i_size is maintained in the in-memory btrfs inode
structures, and this is updated as extents finish.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
The old data=ordered code would force commit to wait until
all the data extents from the transaction were fully on disk. This
introduced large latencies into the commit and stalled new writers
in the transaction for a long time.
The new code changes the way data allocations and extents work:
* When delayed allocation is filled, data extents are reserved, and
the extent bit EXTENT_ORDERED is set on the entire range of the extent.
A struct btrfs_ordered_extent is allocated an inserted into a per-inode
rbtree to track the pending extents.
* As each page is written EXTENT_ORDERED is cleared on the bytes corresponding
to that page.
* When all of the bytes corresponding to a single struct btrfs_ordered_extent
are written, The previously reserved extent is inserted into the FS
btree and into the extent allocation trees. The checksums for the file
data are also updated.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
This allows us to delete an unlinked inode with dirty pages from the list
instead of forcing commit to write these out before deleting the inode.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
Using ilookup5 during data=ordered writeback could deadlock on I_LOCK. This
saves a pointer to the inode instead.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
This forces file data extents down the disk along with the metadata that
references them. The current implementation is fairly simple, and just
writes out all of the dirty pages in an inode before the commit.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
