In normal xattr set, the set sequence is inode, xattr block
and finally xattr bucket if we meet with a ENOSPC. But there
is a corner case.
So consider we will set a xattr whose value will be stored in
a cluster, and there is no xattr block by now. So we will
reserve 1 xattr block and 1 cluster for setting it. Now if we
fail in value extension(in case the volume is almost full and
we can't allocate the cluster because the check in
ocfs2_test_bg_bit_allocatable), ENOSPC will be returned. So
we will try to create a bucket(this time there is a chance that
the reserved cluster will be used), and when we try value extension
again, kernel bug happens. We did meet with it. Check the bug below.
http://oss.oracle.com/bugzilla/show_bug.cgi?id=1251
This patch just try to avoid this by adding a set_abort in
ocfs2_xattr_set_ctxt, so in case ENOSPC happens in value extension,
we will check whether it is caused by the real ENOSPC or just the
full of inode or xattr block. If it is the first case, we set set_abort
so that we don't try any further. we are safe to exit directly here
ince it is really ENOSPC.
Signed-off-by: Tao Ma <tao.ma@oracle.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Currently we process a dirty lockres with the lockres->spinlock taken. While
during the process, we may need to lock on dlm->ast_lock. This breaks the
dependency of dlm->ast_lock(lock first) and lockres->spinlock(lock second).
This patch fixes the problem.
Since we can't release lockres->spinlock, we have to take dlm->ast_lock
just before taking the lockres->spinlock and release it after lockres->spinlock
is released. And use __dlm_queue_bast()/__dlm_queue_ast(), the nolock version,
in dlm_shuffle_lists(). There are no too many locks on a lockres, so there is no
performance harm.
Signed-off-by: Wengang Wang <wen.gang.wang@oracle.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
In ocfs2_prepare_xattr_entry, if we fail to grow an existing value,
xa_cleanup_value_truncate() will leave the old entry in place. Thus, we
reset its value size. However, if we were allocating a new value, we
must not reset the value size or we will BUG(). This resolves
oss.oracle.com bug 1247.
Signed-off-by: Tao Ma <tao.ma@oracle.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Use kstrdup when the goal of an allocation is copy a string into the
allocated region.
The semantic patch that makes this change is as follows:
(http://coccinelle.lip6.fr/)
// <smpl>
@@
expression from,to;
expression flag,E1,E2;
statement S;
@@
- to = kmalloc(strlen(from) + 1,flag);
+ to = kstrdup(from, flag);
... when != \(from = E1 \| to = E1 \)
if (to==NULL || ...) S
... when != \(from = E2 \| to = E2 \)
- strcpy(to, from);
// </smpl>
Signed-off-by: Julia Lawall <julia@diku.dk>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Drop cast on the result of kmalloc and similar functions.
The semantic patch that makes this change is as follows:
(http://coccinelle.lip6.fr/)
// <smpl>
@@
type T;
@@
- (T *)
(\(kmalloc\|kzalloc\|kcalloc\|kmem_cache_alloc\|kmem_cache_zalloc\|
kmem_cache_alloc_node\|kmalloc_node\|kzalloc_node\)(...))
// </smpl>
Signed-off-by: Julia Lawall <julia@diku.dk>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
This patch simplifies the logic of handling existing holes and
skipping extent blocks and removes some confusing comments.
The patch survived the fill_verify_holes testcase in ocfs2-test.
It also passed my manual sanity check and stress tests with enormous
extent records.
Currently punching a hole on a file with 3+ extent tree depth was
really a performance disaster. It can even take several hours,
though we may not hit this in real life with such a huge extent
number.
One simple way to improve the performance is quite straightforward.
From the logic of truncate, we can punch the hole from hole_end to
hole_start, which reduces the overhead of btree operations in a
significant way, such as tree rotation and moving.
Following is the testing result when punching hole from 0 to file end
in bytes, on a 1G file, 1G file consists of 256k extent records, each record
cover 4k data(just one cluster, clustersize is 4k):
===========================================================================
* Original punching-hole mechanism:
===========================================================================
I waited 1 hour for its completion, unfortunately it's still ongoing.
===========================================================================
* Patched punching-hode mechanism:
===========================================================================
real 0m2.518s
user 0m0.000s
sys 0m2.445s
That means we've gained up to 1000 times improvement on performance in this
case, whee! It's fairly cool. and it looks like that performance gain will
be raising when extent records grow.
The patch was based on my former 2 patches, which were about truncating
codes optimization and fixup to handle CoW on punching hole.
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>
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>
Based on the previous patch of optimizing truncate, the bugfix for
refcount trees when punching holes can be fairly easy
and straightforward since most of work we should take into account for
refcounting have been completed already in ocfs2_remove_btree_range().
This patch performs CoW for refcounted extents when a hole being punched
whose start or end offset were in the middle of a cluster, which means
partial zeroing of the cluster will be performed soon.
The patch has been tested fixing the following bug:
http://oss.oracle.com/bugzilla/show_bug.cgi?id=1216
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>
Once file or link creation gets going, it can't be interrupted by a
signal. They're not idempotent.
This blocks signals in ocfs2_mknod(), ocfs2_link(), and ocfs2_symlink()
once we start actually changing things. ocfs2_mknod() covers mknod(),
creat(), mkdir(), and open(O_CREAT).
Signed-off-by: Joel Becker <joel.becker@oracle.com>
ocfs2 sometimes needs to block signals around dlm operations, but it
currently does it with sigprocmask(). Even worse, it's checking the
error code of sigprocmask(). The in-kernel sigprocmask() can only error
if you get the SIG_* argument wrong. We don't.
Wrap the sigprocmask() calls with ocfs2_[un]block_signals(). These
functions are void, but they will BUG() if somehow sigprocmask() returns
an error.
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Lockres hash size of 16KB is far too small for large filesystems (where we
have hundreds of thousands of lock resources stored in the table).
This patch increases it to 128KB.
Signed-off-by: Sunil Mushran <sunil.mushran@oracle.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
In ocfs2, we use ocfs2_extend_trans() to extend a journal handle's
blocks. But if jbd2_journal_extend() fails, it will only restart
with the the new number of blocks. This tends to be awkward since
in most cases we want additional reserved blocks. It makes our code
harder to mantain since the caller can't be sure all the original
blocks will not be accessed and dirtied again. There are 15 callers
of ocfs2_extend_trans() in fs/ocfs2, and 12 of them have to add
h_buffer_credits before they call ocfs2_extend_trans(). This makes
ocfs2_extend_trans() really extend atop the original block count.
Signed-off-by: Tao Ma <tao.ma@oracle.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Two tiny cleanup for allocation reservation.
1. Remove some extra codes in ocfs2_local_alloc_find_clear_bits.
2. Remove an unuseful variables in ocfs2_find_resv_lhs.
Signed-off-by: Tao Ma <tao.ma@oracle.com>
Acked-by: Mark Fasheh <mfasheh@suse.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
When we allocate some bits from the reservation, we always
allocate from the r_start(see ocfs2_resmap_resv_bits).
So there should be no reason to check between r_start
and start. And I don't think we will change this behaviour
later by allocating from some bits after r_start. Why not make
ocfs2_adjust_resv_from_alloc simple for now?
The only chance we have to adjust the reservation is when we haven't
reached the end. With this patch, the function is more readable.
Note:
btw, this patch also fixes an original bug in the function
which I haven't found before.
if (end < ocfs2_resv_end(resv))
rhs = end - ocfs2_resv_end(resv);
This code is of course buggy. ;)
Signed-off-by: Tao Ma <tao.ma@oracle.com>
Acked-by: Mark Fasheh <mfasheh@suse.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
OCFS2 has never really supported intr. This patch acknowledges this reality
and makes nointr the default mount option. In a later patch, we intend to
support intr.
Signed-off-by: Sunil Mushran <sunil.mushran@oracle.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
o2dlm join and leave messages are more than informational as they are
required for debugging locking issues. This patch changes them from
KERN_INFO to KERN_NOTICE.
Signed-off-by: Sunil Mushran <sunil.mushran@oracle.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
This patch logs socket state changes that lead to socket shutdown.
Signed-off-by: Srinivas Eeda <srinivas.eeda@oracle.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Print the node number of a peer node if sending it a message failed.
Signed-off-by: Wengang Wang <wen.gang.wang@oracle.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
The default behavior for directory reservations stays the same, but we add a
mount option so people can tweak the size of directory reservations
according to their workloads.
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
The default reservation size of 4 (32-bit windows) is a bit too ambitious.
Scale it back to 16 bits (resv_level=2). I have been testing various sizes
on a 4-node cluster which runs a mixed workload that is heavily threaded.
With a 256MB local alloc, I get *roughly* the following levels of average file
fragmentation:
resv_level=0 70%
resv_level=1 21%
resv_level=2 23%
resv_level=3 24%
resv_level=4 60%
resv_level=5 did not test
resv_level=6 60%
resv_level=2 seemed like a good compromise between not letting windows be
too small, but not so big that heavier workloads will immediately suffer
without tuning.
This patch also change the behavior of directory reservations - they now
track file reservations. The previous compromise of giving directory
windows only 8 bits wound up fragmenting more at some window sizes because
file allocations had smaller unused windows to poach from.
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
I have observed that the current size of 8M gives us pretty poor
fragmentation on multi-threaded workloads which do lots of writes.
Generally, I can increase the size of local alloc windows and observe a
marked decrease in fragmentation, even up and beyond window sizes of 512
megabytes. This makes sense for a couple reasons - larger local alloc means
more room for reservation windows. On multi-node workloads the larger local
alloc helps as well because we don't have to do window slides as often.
Also, I removed the OCFS2_DEFAULT_LOCAL_ALLOC_SIZE constant as it is no
longer used and the comment above it was out of date.
To test fragmentation, I used a workload which launched 4 threads that did
4k writes into a series of about 140 alternating files.
With resv_level=2, and a 4k/4k file system I observed the following average
fragmentation for various localalloc= parameters:
localalloc= avg. fragmentation
8 48
32 16
64 10
120 7
On larger cluster sizes, the difference is more dramatic.
The new default size top out at 256M, which we'll only get for cluster
sizes of 32K and above.
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
This patch pulls the local alloc sizing code into localalloc.c and provides
a callout to it from ocfs2_fill_super(). Behavior is essentially unchanged
except that I correctly calculate the maximum local alloc size. The old code
in ocfs2_parse_options() calculated the max size as:
ocfs2_local_alloc_size(sb) * 8
which is correct, in bits. Unfortunately though the option passed in is in
megabytes. Ultimately, this bug made no real difference - the shrink code
would catch a too-large size and bring it down to something reasonable.
Still, it's less than efficient as-is.
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Inodes are always allocated from the global bitmap now so we don't need this
any more. Also, the existing implementation bounces reservations around
needlessly.
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
Otherwise, the need for a very large contiguous allocation tends to
wreak havoc on many inode allocation reservations on the local alloc, thus
ruining any chances for contiguousness.
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
Use the reservations system for unindexed dir tree allocations. We don't
bother with the indexed tree as reads from it are mostly random anyway.
Directory reservations are marked seperately, to allow the reservations code
a chance to optimize their window sizes. This patch allocates only 8 bits
for directory windows as they generally are not expected to grow as quickly
as file data. Future improvements to dir window sizing can trivially be
made.
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
This patch improves Ocfs2 allocation policy by allowing an inode to
reserve a portion of the local alloc bitmap for itself. The reserved
portion (allocation window) is advisory in that other allocation
windows might steal it if the local alloc bitmap becomes
full. Otherwise, the reservations are honored and guaranteed to be
free. When the local alloc window is moved to a different portion of
the bitmap, existing reservations are discarded.
Reservation windows are represented internally by a red-black
tree. Within that tree, each node represents the reservation window of
one inode. An LRU of active reservations is also maintained. When new
data is written, we allocate it from the inodes window. When all bits
in a window are exhausted, we allocate a new one as close to the
previous one as possible. Should we not find free space, an existing
reservation is pulled off the LRU and cannibalized.
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
jbd[2]_journal_dirty_metadata() only returns 0. It's been returning 0
since before the kernel moved to git. There is no point in checking
this error.
ocfs2_journal_dirty() has been faithfully returning the status since the
beginning. All over ocfs2, we have blocks of code checking this can't
fail status. In the past few years, we've tried to avoid adding these
checks, because they are pointless. But anyone who looks at our code
assumes they are needed.
Finally, ocfs2_journal_dirty() is made a void function. All error
checking is removed from other files. We'll BUG_ON() the status of
jbd2_journal_dirty_metadata() just in case they change it someday. They
won't.
Signed-off-by: Joel Becker <joel.becker@oracle.com>
When the local alloc file changes windows, unused bits are freed back to the
global bitmap. By defnition, those bits can not be in use by any file. Also,
the local alloc will never have been able to allocate those bits if they
were part of a previous truncate. Therefore it makes sense that we should
clear unused local alloc bits in the undo buffer so that they can be used
immediatly.
[ Modified to call it ocfs2_release_clusters() -- Joel ]
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
You can't store a pointer that you haven't filled in yet and expect it
to work.
Signed-off-by: Tao Ma <tao.ma@oracle.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
When replacing a xattr's value, in some case we wipe its name/value
first and then re-add it. The wipe is done by
ocfs2_xa_block_wipe_namevalue() when the xattr is in the inode or
block. We currently adjust name_offset for all the entries which have
(offset < name_offset). This does not adjust the entrie we're replacing.
Since we are replacing the entry, we don't adjust the total entry count.
When we calculate a new namevalue location, we trust the entries
now-wrong offset in ocfs2_xa_get_free_start(). The solution is to
also adjust the name_offset for the replaced entry, allowing
ocfs2_xa_get_free_start() to calculate the new namevalue location
correctly.
The following script can trigger a kernel panic easily.
echo 'y'|mkfs.ocfs2 --fs-features=local,xattr -b 4K $DEVICE
mount -t ocfs2 $DEVICE $MNT_DIR
FILE=$MNT_DIR/$RANDOM
for((i=0;i<76;i++))
do
string_76="a$string_76"
done
string_78="aa$string_76"
string_82="aaaa$string_78"
touch $FILE
setfattr -n 'user.test1234567890' -v $string_76 $FILE
setfattr -n 'user.test1234567890' -v $string_78 $FILE
setfattr -n 'user.test1234567890' -v $string_82 $FILE
Signed-off-by: Tao Ma <tao.ma@oracle.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
What we were doing before was to ask for the current window size as the
maximum allocation. This had the effect of limiting the amount of allocation
we could get for the local alloc during times when the window size was
shrunk due to fragmentation. In some cases, that could actually *increase*
fragmentation by artificially limiting the number of bits we can accept. So
while we still want to ask for a minimum number of bits equal to window
size, there is no reason why we should limit the number of bits the local
alloc should accept. Hence always allow the maximum number of local alloc
bits.
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
ac_last_group is used to record the last block group we
used during allocation. But the initialization process
only calls ocfs2_which_suballoc_group and fails to
use suballoc_loc properly. So let us do it.
Another function ocfs2_test_suballoc_bit also needs fix.
I have searched all the callers of ocfs2_which_suballoc_group,
and all the callers notices suballoc_loc now.
Signed-off-by: Tao Ma <tao.ma@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_group_bitmap_size has to handle the case when the
volume don't have discontiguous block group support. So
pass the feature_incompat in and check it.
Signed-off-by: Tao Ma <tao.ma@oracle.com>
The fixes include:
1. some endian problems.
2. we should use bit/bpc in ocfs2_block_group_grow_discontig to
allocate clusters.
3. set num_clusters properly in __ocfs2_claim_clusters.
4. change name from ocfs2_supports_discontig_bh to
ocfs2_supports_discontig_bg.
Signed-off-by: Tao Ma <tao.ma@oracle.com>
Rather than extending the transaction every time we add an extent to a
discontiguous block group, we grab enough credits to fill the extent
list up front. This means we can free the bits in the same transaction
if we end up not getting enough space.
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Get the suballoc_loc from ocfs2_claim_new_inode() or
ocfs2_claim_metadata(). Store it on the appropriate field of the block
we just allocated.
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Rather than calculating the resulting block number, return it on the
ocfs2_suballoc_result structure. This way we can calculate block
numbers for discontiguous block groups.
Cluster groups keep doing it the old way.
Signed-off-by: Joel Becker <joel.becker@oracle.com>
They all take an ocfs2_alloc_context, which has the allocation inode.
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Signed-off-by: Tao Ma <tao.ma@oracle.com>
A discontiguous block group can find a range of free bits that straddle
more than one region of its space. Callers can't handle that, so we
trim the returned bits until they fit within one region.
Only cluster allocations ask for min_bits>1. Discontiguous block groups
are only for block allocations. So min_bits doesn't matter here.
Signed-off-by: Joel Becker <joel.becker@oracle.com>
We're going to be adding more info to a suballocator allocation. Rather
than growing every function in the chain, let's pass a result structure
around.
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Signed-off-by: Tao Ma <tao.ma@oracle.com>
If we cannot get a contiguous region for a block group, allocate a
discontiguous one when the filesystem supports it.
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Signed-off-by: Tao Ma <tao.ma@oracle.com>
