When running ocfs2 test suite multiple nodes reflink stress test, for a
4 nodes cluster, every unlink() for refcounted file needs about 700s.
The slow unlink is caused by the contention of refcount tree lock since
all nodes are unlink files using the same refcount tree. When the
unlinking file have many extents(over 1600 in our test), most of the
extents has refcounted flag set. In ocfs2_commit_truncate(), it will
execute the following call trace for every extents. This means it needs
get and released refcount tree lock about 1600 times. And when several
nodes are do this at the same time, the performance will be very low.
ocfs2_remove_btree_range()
-- ocfs2_lock_refcount_tree()
---- ocfs2_refcount_lock()
------ __ocfs2_cluster_lock()
ocfs2_refcount_lock() is costly, move it to ocfs2_commit_truncate() to
do lock/unlock once can improve a lot performance.
Signed-off-by: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Wengang <wen.gang.wang@oracle.com>
Reviewed-by: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Add ocfs2_trim_fs to support trimming freed clusters in the
volume. A range will be given and all the freed clusters greater
than minlen will be discarded to the block layer.
Signed-off-by: Tao Ma <boyu.mt@taobao.com>
Signed-off-by: Joel Becker <jlbec@evilplan.org>
Tristan Ye has done some refactoring against our truncate
process, so some functions like ocfs2_prepare_truncate and
ocfs2_free_truncate_context are no use and we'd better
remove them.
Signed-off-by: Tao Ma <tao.ma@oracle.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
The original idea to pull ocfs2_find_cpos_for_left_leaf() out of
alloc.c is to benefit punching-holes optimization patch, it however,
can also be referred by other funcs in the future who want to do the
same job.
Signed-off-by: Tristan Ye <tristan.ye@oracle.com>
Acked-by: Mark Fasheh <mfasheh@suse.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Truncate is just a special case of punching holes(from new i_size to
end), we therefore could take advantage of the existing
ocfs2_remove_btree_range() to reduce the comlexity and redundancy in
alloc.c. The goal here is to make truncate more generic and
straightforward.
Several functions only used by ocfs2_commit_truncate() will smiply be
removed.
ocfs2_remove_btree_range() was originally used by the hole punching
code, which didn't take refcount trees into account (definitely a bug).
We therefore need to change that func a bit to handle refcount trees.
It must take the refcount lock, calculate and reserve blocks for
refcount tree changes, and decrease refcounts at the end. We replace
ocfs2_lock_allocators() here by adding a new func
ocfs2_reserve_blocks_for_rec_trunc() which accepts some extra blocks to
reserve. This will not hurt any other code using
ocfs2_remove_btree_range() (such as dir truncate and hole punching).
I merged the following steps into one patch since they may be
logically doing one thing, though I know it looks a little bit fat
to review.
1). Remove redundant code used by ocfs2_commit_truncate(), since we're
moving to ocfs2_remove_btree_range anyway.
2). Add a new func ocfs2_reserve_blocks_for_rec_trunc() for purpose of
accepting some extra blocks to reserve.
3). Change ocfs2_prepare_refcount_change_for_del() a bit to fit our
needs. It's safe to do this since it's only being called by
truncate.
4). Change ocfs2_remove_btree_range() a bit to take refcount case into
account.
5). Finally, we change ocfs2_commit_truncate() to call
ocfs2_remove_btree_range() in a proper way.
The patch has been tested normally for sanity check, stress tests
with heavier workload will be expected.
Based on this patch, fixing the punching holes bug will be fairly easy.
Signed-off-by: Tristan Ye <tristan.ye@oracle.com>
Acked-by: Mark Fasheh <mfasheh@suse.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
In case the block we are going to free is allocated from
a discontiguous block group, we have to use suballoc_loc
to be the right group.
Signed-off-by: Tao Ma <tao.ma@oracle.com>
ocfs2 refcount tree is stored as an extent tree while
the leaf ocfs2_refcount_rec points to a refcount block.
The following step can trip a kernel panic.
mkfs.ocfs2 -b 512 -C 1M --fs-features=refcount $DEVICE
mount -t ocfs2 $DEVICE $MNT_DIR
FILE_NAME=$RANDOM
FILE_NAME_1=$RANDOM
FILE_REF="${FILE_NAME}_ref"
FILE_REF_1="${FILE_NAME}_ref_1"
for((i=0;i<305;i++))
do
# /mnt/1048576 is a file with 1048576 sizes.
cat /mnt/1048576 >> $MNT_DIR/$FILE_NAME
cat /mnt/1048576 >> $MNT_DIR/$FILE_NAME_1
done
for((i=0;i<3;i++))
do
cat /mnt/1048576 >> $MNT_DIR/$FILE_NAME
done
for((i=0;i<2;i++))
do
cat /mnt/1048576 >> $MNT_DIR/$FILE_NAME
cat /mnt/1048576 >> $MNT_DIR/$FILE_NAME_1
done
cat /mnt/1048576 >> $MNT_DIR/$FILE_NAME
for((i=0;i<11;i++))
do
cat /mnt/1048576 >> $MNT_DIR/$FILE_NAME
cat /mnt/1048576 >> $MNT_DIR/$FILE_NAME_1
done
reflink $MNT_DIR/$FILE_NAME $MNT_DIR/$FILE_REF
# write_f is a program which will write some bytes to a file at offset.
# write_f -f file_name -l offset -w write_bytes.
./write_f -f $MNT_DIR/$FILE_REF -l $[310*1048576] -w 4096
./write_f -f $MNT_DIR/$FILE_REF -l $[306*1048576] -w 4096
./write_f -f $MNT_DIR/$FILE_REF -l $[311*1048576] -w 4096
./write_f -f $MNT_DIR/$FILE_NAME -l $[310*1048576] -w 4096
./write_f -f $MNT_DIR/$FILE_NAME -l $[311*1048576] -w 4096
reflink $MNT_DIR/$FILE_NAME $MNT_DIR/$FILE_REF_1
./write_f -f $MNT_DIR/$FILE_NAME -l $[311*1048576] -w 4096
#kernel panic here.
The reason is that if the ocfs2_extent_rec is the last record
in a leaf extent block, the old solution fails to find the
suitable end cpos. So this patch try to walk through the b-tree,
find the next sub root and get the c_pos the next sub-tree starts
from.
btw, I have runned tristan's test case against the patched kernel
for several days and this type of kernel panic never happens again.
Signed-off-by: Tao Ma <tao.ma@oracle.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
This patch try CoW support for a refcounted record.
the whole process will be:
1. Calculate how many clusters we need to CoW and where we start.
Extents that are not completely encompassed by the write will
be broken on 1MB boundaries.
2. Do CoW for the clusters with the help of page cache.
3. Change the b-tree structure with the new allocated clusters.
Signed-off-by: Tao Ma <tao.ma@oracle.com>
Given a physical cpos and length, decrement the refcount
in the tree. If the refcount for any portion of the extent goes
to zero, that portion is queued for freeing.
Signed-off-by: Tao Ma <tao.ma@oracle.com>
Now fs/ocfs2/alloc.c has more than 7000 lines. It contains our
basic b-tree operation. Although we have already make our b-tree
operation generic, the basic structrue ocfs2_path which is used
to iterate one b-tree branch is still static and limited to only
used in alloc.c. As refcount tree need them and I don't want to
add any more b-tree unrelated code to alloc.c, export them out.
Signed-off-by: Tao Ma <tao.ma@oracle.com>
Add refcount b-tree as a new extent tree so that it can
use the b-tree to store and maniuplate ocfs2_refcount_rec.
Signed-off-by: Tao Ma <tao.ma@oracle.com>
With this commit, extent tree operations are divorced from inodes and
rely on ocfs2_caching_info. Phew!
Signed-off-by: Joel Becker <joel.becker@oracle.com>
ocfs2_find_path and ocfs2_find_leaf() walk our btrees, reading extent
blocks. They need struct ocfs2_caching_info for that, but not struct
inode.
Signed-off-by: Joel Becker <joel.becker@oracle.com>
extent blocks belong to btrees on more than just inodes, so we want to
pass the ocfs2_caching_info structure directly to
ocfs2_read_extent_block(). A number of places in alloc.c can now drop
struct inode from their argument list.
Signed-off-by: Joel Becker <joel.becker@oracle.com>
What do we cache? Metadata blocks. What are most of our non-inode metadata
blocks? Extent blocks for our btrees. struct ocfs2_extent_tree is the
main structure for managing those. So let's store the associated
ocfs2_caching_info there.
This means that ocfs2_et_root_journal_access() doesn't need struct inode
anymore, and any place that has an et can refer to et->et_ci instead of
INODE_CACHE(inode).
Signed-off-by: Joel Becker <joel.becker@oracle.com>
This patch makes use of Ocfs2's flexible btree code to add an additional
tree to directory inodes. The new tree stores an array of small,
fixed-length records in each leaf block. Each record stores a hash value,
and pointer to a block in the traditional (unindexed) directory tree where a
dirent with the given name hash resides. Lookup exclusively uses this tree
to find dirents, thus providing us with constant time name lookups.
Some of the hashing code was copied from ext3. Unfortunately, it has lots of
unfixed checkpatch errors. I left that as-is so that tracking changes would
be easier.
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
Acked-by: Joel Becker <joel.becker@oracle.com>
When an ocfs2 extended attribute is large enough to require its own
allocation tree, we root it with an ocfs2_xattr_value_root. However,
these roots can be a part of inodes, xattr blocks, or xattr buckets.
Thus, they need a different journal access function for each container.
We wrap the bh, its journal access function, and the value root (xv) in
a structure called ocfs2_xattr_valu_buf. This is a package that can
be passed around. In this first pass, we simply pass it to the
extent tree code.
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
The per-metadata-type ocfs2_journal_access_*() functions hook up jbd2
commit triggers and allow us to compute metadata ecc right before the
buffers are written out. This commit provides ecc for inodes, extent
blocks, group descriptors, and quota blocks. It is not safe to use
extened attributes and metaecc at the same time yet.
The ocfs2_extent_tree and ocfs2_path abstractions in alloc.c both hide
the type of block at their root. Before, it didn't matter, but now the
root block must use the appropriate ocfs2_journal_access_*() function.
To keep this abstract, the structures now have a pointer to the matching
journal_access function and a wrapper call to call it.
A few places use naked ocfs2_write_block() calls instead of adding the
blocks to the journal. We make sure to calculate their checksum and ecc
before the write.
Since we pass around the journal_access functions. Let's typedef them
in ocfs2.h.
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
We weren't consistently checking extent blocks after we read them.
Most places checked the signature, but none checked h_blkno or
h_fs_signature. Create a toplevel ocfs2_read_extent_block() that does
the read and the validation.
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
This patch genericizes the high level handling of extent removal.
ocfs2_remove_btree_range() is nearly identical to
__ocfs2_remove_inode_range(), except that extent tree operations have been
used where necessary. We update ocfs2_remove_inode_range() to use the
generic helper. Now extent tree based structures have an easy way to
truncate ranges.
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
Acked-by: Joel Becker <joel.becker@oracle.com>
In ocfs2 xattr set, we reserve metadata and clusters in any place
they are needed. It is time-consuming and ineffective, so this
patch try to reserve metadata and clusters at the beginning of
ocfs2_xattr_set.
Signed-off-by: Tao Ma <tao.ma@oracle.com>
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
Now in ocfs2 xattr set, the whole process are divided into many small
parts and they are wrapped into diffrent transactions and it make the
set doesn't look like a real transaction. So we want to integrate it
into a real one.
In some cases we will allocate some clusters and free some in just one
transaction. e.g, one xattr is larger than inline size, so it and its
value root is stored within the inode while the value is outside in a
cluster. Then we try to update it with a smaller value(larger than the
size of root but smaller than inline size), we may need to free the
outside cluster while allocate a new bucket(one cluster) since now the
inode may be full. The old solution will lock the global_bitmap(if the
local alloc failed in stress test) and then the truncate log. This will
cause a ABBA lock with truncate log flush.
This patch add the clusters free in dealloc_ctxt, so that we can record
the free clusters during the transaction and then free it after we
release the global_bitmap in xattr set.
Signed-off-by: Tao Ma <tao.ma@oracle.com>
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
The original get/put_extent_tree() functions held a reference on
et_root_bh. However, every single caller already has a safe reference,
making the get/put cycle irrelevant.
We change ocfs2_get_*_extent_tree() to ocfs2_init_*_extent_tree(). It
no longer gets a reference on et_root_bh. ocfs2_put_extent_tree() is
removed. Callers now have a simpler init+use pattern.
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
We now have three different kinds of extent trees in ocfs2: inode data
(dinode), extended attributes (xattr_tree), and extended attribute
values (xattr_value). There is a nice abstraction for them,
ocfs2_extent_tree, but it is hidden in alloc.c. All the calling
functions have to pick amongst a varied API and pass in type bits and
often extraneous pointers.
A better way is to make ocfs2_extent_tree a first-class object.
Everyone converts their object to an ocfs2_extent_tree() via the
ocfs2_get_*_extent_tree() calls, then uses the ocfs2_extent_tree for all
tree calls to alloc.c.
This simplifies a lot of callers, making for readability. It also
provides an easy way to add additional extent tree types, as they only
need to be defined in alloc.c with a ocfs2_get_<new>_extent_tree()
function.
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
A caller knows what kind of extent tree they have. There's no reason
they have to call ocfs2_get_extent_tree() with a NULL when they could
just as easily call a specific function to their type of extent tree.
Introduce ocfs2_dinode_get_extent_tree(),
ocfs2_xattr_tree_get_extent_tree(), and
ocfs2_xattr_value_get_extent_tree(). They only take the necessary
arguments, calling into the underlying __ocfs2_get_extent_tree() to do
the real work.
__ocfs2_get_extent_tree() is the old ocfs2_get_extent_tree(), but
without needing any switch-by-type logic.
ocfs2_get_extent_tree() is now a wrapper around the specific calls. It
exists because a couple alloc.c functions can take et_type. This will
go later.
Another benefit is that ocfs2_xattr_value_get_extent_tree() can take a
struct ocfs2_xattr_value_root* instead of void*. This gives us
typechecking where we didn't have it before.
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
In xattr bucket, we want to limit the maximum size of a btree leaf,
otherwise we'll lose the benefits of hashing because we'll have to search
large leaves.
So add a new field in ocfs2_extent_tree which indicates the maximum leaf cluster
size we want so that we can prevent ocfs2_insert_extent() from merging the leaf
record even if it is contiguous with an adjacent record.
Other btree types are not affected by this change.
Signed-off-by: Tao Ma <tao.ma@oracle.com>
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
When necessary, an ocfs2_xattr_block will embed an ocfs2_extent_list to
store large numbers of EAs. This patch adds a new type in
ocfs2_extent_tree_type and adds the implementation so that we can re-use the
b-tree code to handle the storage of many EAs.
Signed-off-by: Tao Ma <tao.ma@oracle.com>
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
Add some thin wrappers around ocfs2_insert_extent() for each of the 3
different btree types, ocfs2_inode_insert_extent(),
ocfs2_xattr_value_insert_extent() and ocfs2_xattr_tree_insert_extent(). The
last is for the xattr index btree, which will be used in a followup patch.
All the old callers in file.c etc will call ocfs2_dinode_insert_extent(),
while the other two handle the xattr issue. And the init of extent tree are
handled by these functions.
When storing xattr value which is too large, we will allocate some clusters
for it and here ocfs2_extent_list and ocfs2_extent_rec will also be used. In
order to re-use the b-tree operation code, a new parameter named "private"
is added into ocfs2_extent_tree and it is used to indicate the root of
ocfs2_exent_list. The reason is that we can't deduce the root from the
buffer_head now. It may be in an inode, an ocfs2_xattr_block or even worse,
in any place in an ocfs2_xattr_bucket.
Signed-off-by: Tao Ma <tao.ma@oracle.com>
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
Factor out the non-inode specifics of ocfs2_do_extend_allocation() into a more generic
function, ocfs2_do_cluster_allocation(). ocfs2_do_extend_allocation calls
ocfs2_do_cluster_allocation() now, but the latter can be used for other
btree types as well.
Signed-off-by: Tao Ma <tao.ma@oracle.com>
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
In the old extent tree operation, we take the hypothesis that we
are using the ocfs2_extent_list in ocfs2_dinode as the tree root.
As xattr will also use ocfs2_extent_list to store large value
for a xattr entry, we refactor the tree operation so that xattr
can use it directly.
The refactoring includes 4 steps:
1. Abstract set/get of last_eb_blk and update_clusters since they may
be stored in different location for dinode and xattr.
2. Add a new structure named ocfs2_extent_tree to indicate the
extent tree the operation will work on.
3. Remove all the use of fe_bh and di, use root_bh and root_el in
extent tree instead. So now all the fe_bh is replaced with
et->root_bh, el with root_el accordingly.
4. Make ocfs2_lock_allocators generic. Now it is limited to be only used
in file extend allocation. But the whole function is useful when we want
to store large EAs.
Note: This patch doesn't touch ocfs2_commit_truncate() since it is not used
for anything other than truncate inode data btrees.
Signed-off-by: Tao Ma <tao.ma@oracle.com>
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
ocfs2_extend_meta_needed(), ocfs2_calc_extend_credits() and
ocfs2_reserve_new_metadata() are all useful for extent tree operations. But
they are all limited to an inode btree because they use a struct
ocfs2_dinode parameter. Change their parameter to struct ocfs2_extent_list
(the part of an ocfs2_dinode they actually use) so that the xattr btree code
can use these functions.
Signed-off-by: Tao Ma <tao.ma@oracle.com>
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
ocfs2_num_free_extents() is used to find the number of free extent records
in an inode btree. Hence, it takes an "ocfs2_dinode" parameter. We want to
use this for extended attribute trees in the future, so genericize the
interface the take a buffer head. A future patch will allow that buffer_head
to contain any structure rooting an ocfs2 btree.
Signed-off-by: Tao Ma <tao.ma@oracle.com>
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
Plug ocfs2 into ->fiemap. Some portions of ocfs2_get_clusters() had to be
refactored so that the extent cache can be skipped in favor of going
directly to the on-disk records. This makes it easier for us to determine
which extent is the last one in the btree. Also, I'm not sure we want to be
caching fiemap lookups anyway as they're not directly related to data
read/write.
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Cc: ocfs2-devel@oss.oracle.com
Cc: linux-fsdevel@vger.kernel.org
Create all new directories with OCFS2_INLINE_DATA_FL and the inline data
bytes formatted as an empty directory. Inode size field reflects the actual
amount of inline data available, which makes searching for dirent space
very similar to the regular directory search.
Inline-data directories are automatically pushed out to extents on any
insert request which is too large for the available space.
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
Reviewed-by: Joel Becker <joel.becker@oracle.com>
This fixes up write, truncate, mmap, and RESVSP/UNRESVP to understand inline
inode data.
For the most part, the changes to the core write code can be relied on to do
the heavy lifting. Any code calling ocfs2_write_begin (including shared
writeable mmap) can count on it doing the right thing with respect to
growing inline data to an extent tree.
Size reducing truncates, including UNRESVP can simply zero that portion of
the inode block being removed. Size increasing truncatesm, including RESVP
have to be a little bit smarter and grow the inode to an extent tree if
necessary.
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
Reviewed-by: Joel Becker <joel.becker@oracle.com>
Provide an internal interface for the removal of arbitrary file regions.
ocfs2_remove_inode_range() takes a byte range within a file and will remove
existing extents within that range. Partial clusters will be zeroed so that
any read from within the region will return zeros.
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
The partial cluster zeroing code used during truncate usually assumes that
the rightmost byte in the range to be zeroed lies on a cluster boundary.
This makes sense for truncate, but punching holes might require zeroing on
non-aligned rightmost boundaries.
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
This can now be trivially supported with re-use of our existing extend code.
ocfs2_allocate_unwritten_extents() takes a start offset and a byte length
and iterates over the inode, adding extents (marked as unwritten) until len
is reached. Existing extents are skipped over.
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
Writes to a region marked as unwritten might result in a record split or
merge. We can support splits by making minor changes to the existing insert
code. Merges require left rotations which mostly re-use right rotation
support functions.
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
Deallocation of suballocator blocks, most notably extent blocks, might
involve multiple suballocator inodes.
The locking for this can get extremely complicated, especially when the
suballocator inodes to delete from aren't known until deep within an
unrelated codepath.
Implement a simple scheme for recording the blocks to be unlinked so that
the actual deallocation can be done in a context which won't deadlock.
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
Due to the size of our group bitmaps, we'll never have a leaf node extent
record with more than 16 bits worth of clusters. Split e_clusters up so that
leaf nodes can get a flags field where we can mark unwritten extents.
Interior nodes whose length references all the child nodes beneath it can't
split their e_clusters field, so we use a union to preserve sizing there.
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
Since we don't zero on extend anymore, truncate needs to be fixed up to zero
the part of a file between i_size and and end of it's cluster. Otherwise a
subsequent extend could expose bad data.
This introduced a new helper, which can be used in ocfs2_write().
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
The code in extent_map.c is not prepared to deal with a subtree being
rotated between lookups. This can happen when filling holes in sparse files.
Instead of a lengthy patch to update the code (which would likely lose the
benefit of caching subtree roots), we remove most of the algorithms and
implement a simple path based lookup. A less ambitious extent caching scheme
will be added in a later patch.
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
Introduce tree rotations into the b-tree code. This will allow ocfs2 to
support sparse files. Much of the added code is designed to be generic (in
the ocfs2 sense) so that it can later be re-used to implement large
extended attributes.
This patch only adds the rotation code and does minimal updates to callers
of the extent api.
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
