This idea is from ext4. By this patch, we can make the dio write parallel,
and improve the performance. But because we can not update isize without
i_mutex, the unlocked dio write just can be done in front of the EOF.
We needn't worry about the race between dio write and truncate, because the
truncate need wait untill all the dio write end.
And we also needn't worry about the race between dio write and punch hole,
because we have extent lock to protect our operation.
I ran fio to test the performance of this feature.
== Hardware ==
CPU: Intel(R) Core(TM)2 Duo CPU E7500 @ 2.93GHz
Mem: 2GB
SSD: Intel X25-M 120GB (Test Partition: 60GB)
== config file ==
[global]
ioengine=psync
direct=1
bs=4k
size=32G
runtime=60
directory=/mnt/btrfs/
filename=testfile
group_reporting
thread
[file1]
numjobs=1 # 2 4
rw=randwrite
== result (KBps) ==
write 1 2 4
lock 24936 24738 24726
nolock 24962 30866 32101
== result (iops) ==
write 1 2 4
lock 6234 6184 6181
nolock 6240 7716 8025
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Currently, we can do unlocked dio reads, but the following race
is possible:
dio_read_task truncate_task
->btrfs_setattr()
->btrfs_direct_IO
->__blockdev_direct_IO
->btrfs_get_block
->btrfs_truncate()
#alloc truncated blocks
#to other inode
->submit_io()
#INFORMATION LEAK
In order to avoid this problem, we must serialize unlocked dio reads with
truncate. There are two approaches:
- use extent lock to protect the extent that we truncate
- use inode_dio_wait() to make sure the truncating task will wait for
the read DIO.
If we use the 1st one, we will meet the endless truncation problem due to
the nonlocked read DIO after we implement the nonlocked write DIO. It is
because we still need invoke inode_dio_wait() avoid the race between write
DIO and truncation. By that time, we have to introduce
btrfs_inode_{block, resume}_nolock_dio()
again. That is we have to implement this patch again, so I choose the 2nd
way to fix the problem.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
The deadlock problem happened when running fsstress(a test program in LTP).
Steps to reproduce:
# mkfs.btrfs -b 100M <partition>
# mount <partition> <mnt>
# <Path>/fsstress -p 3 -n 10000000 -d <mnt>
The reason is:
btrfs_direct_IO()
|->do_direct_IO()
|->get_page()
|->get_blocks()
| |->btrfs_delalloc_resereve_space()
| |->btrfs_add_ordered_extent() ------- Add a new ordered extent
|->dio_send_cur_page(page0) -------------- We didn't submit bio here
|->get_page()
|->get_blocks()
|->btrfs_delalloc_resereve_space()
|->flush_space()
|->btrfs_start_ordered_extent()
|->wait_event() ---------- Wait the completion of
the ordered extent that is
mentioned above
But because we didn't submit the bio that is mentioned above, the ordered
extent can not complete, we would wait for its completion forever.
There are two methods which can fix this deadlock problem:
1. submit the bio before we invoke get_blocks()
2. reserve the space before we do dio
Though the 1st is the simplest way, we need modify the code of VFS, and it
is likely to break contiguous requests, and introduce performance regression
for the other filesystems.
So we have to choose the 2nd way.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
I noticed we were getting lots of warnings with xfstest 83 because we have
reservations outstanding. This is because we moved the orphan add outside
of the truncate, but we don't actually cleanup our reservation if something
fails. This fixes the problem and I no longer see warnings. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Sometimes xfstest 83 will fail to remount the scratch device because we've
gotten ourselves so full that we cannot cleanup the orphan items. In this
case check to see if we're doing the orphan cleanup and if we are allow us
to steal our reservation from the global block rsv. With this patch I've
not been able to reproduce the failed mount problem. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
The argument "inherit" of btrfs_ioctl_snap_create_transid() was assigned
to NULL during we created the snapshots, so we didn't free it though we
called kfree() in the caller.
But since we are sure the snapshot creation is done after the function -
btrfs_ioctl_snap_create_transid() - completes, it is safe that we don't
assign the pointer "inherit" to NULL, and just free it in the caller of
btrfs_ioctl_snap_create_transid(). In this way, the code can become more
readable.
Reported-by: Alex Lyakas <alex.btrfs@zadarastorage.com>
Cc: Arne Jansen <sensille@gmx.net>
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
open_ctree() need read the metadata to initialize the global information
of btrfs. But it may fail after it submit some bio, and then it will jump
to the error path. Unfortunately, it doesn't check if there are some bios
in flight, and just stop all the worker threads. As a result, when the
submitted bios end, they can not find any worker thread which can deal with
subsequent work, then oops happen.
kernel BUG at fs/btrfs/async-thread.c:605!
Fix this problem by invoking invalidate_inode_pages2() before we stop the
worker threads. This function will wait until the bio end because it need
lock the pages which are going to be invalidated, and if a page is under
disk read IO, it must be locked. invalidate_inode_pages2() need wait until
end bio handler to unlocked it.
Reported-and-Tested-by: Tsutomu Itoh <t-itoh@jp.fujitsu.com>
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
This patch adds the flag, BTRFS_SEND_FLAG_NO_FILE_DATA to the btrfs send
ioctl code. When this flag is set, the btrfs send code will never write file
data into the stream (thus also avoiding expensive reads of that data in the
first place). BTRFS_SEND_C_UPDATE_EXTENT commands will be sent (instead of
BTRFS_SEND_C_WRITE) with an offset, length pair indicating the extent in
question.
This patch does not affect the operation of BTRFS_SEND_C_CLONE commands -
they will continue to be sent when a search finds an appropriate extent to
clone from.
Signed-off-by: Mark Fasheh <mfasheh@suse.de>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
For write, we also reserve some space for COW blocks during updating
the checksum tree, and we calculate the number of blocks by checking
if the number of bytes outstanding that are going to need csums needs
one more block for csum.
When we add these checksum into the checksum tree, we use ordered sums
list.
Every ordered sum contains csums for each sector, and we'll first try
to look up an existing csum item,
a) if we don't yet have a proper csum item, then we need to insert one,
b) or if we find one but the csum item is not big enough, then we need
to extend it.
The point is we'll unlock the whole path and then insert or extend.
So others can hack in and update the tree.
Each insert or extend needs update the tree with COW on, and we may need
to insert/extend for many times.
That means what we've reserved for updating checksum tree is NOT enough
indeed.
The case is even more serious with having several write threads at the
same time, it can end up eating our reserved space quickly and starting
eating globle reserve pool instead.
I don't yet come up with a way to calculate the worse case for updating
csum, but extending the checksum item as much as possible can be helpful
in my test.
The idea behind is that it can reduce the times we insert/extend so that
it saves us precious reserved space.
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
The entry point at the defrag ioctl always sets "cache only" to 0;
the codepaths haven't run for a long time as far as I can
tell. Chris says they're dead code, so remove them.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
I hit a deadlock where transaction commit was waiting on num_writers to be
0. This happened because somebody came into btrfs_commit_transaction and
noticed we had aborted and it went to cleanup_transaction. This shouldn't
happen because cleanup_transaction is really to fixup a bad commit, it
doesn't do the normal trans handle cleanup things. So if we have an error
just do the normal btrfs_end_transaction dance and return. Once we are in
the actual commit path we can use cleanup_transaction and be good to go.
Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
I noticed we would deadlock if we aborted a transaction while doing
compressed io. This is because we don't unlock our pages if something goes
horribly wrong. To fix this we need to make sure that we call
extent_clear_unlock_delalloc in order to unlock all the pages. If we have
to cow in the async submission thread we need to make sure to unlock our
locked_page as the cow error path will not unlock the locked page as it
depends on the caller to unlock that page. With this patch we no longer
deadlock on the page lock when we have an aborted transaction. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
People have been complaining about random ENOSPC errors that will clear up
after a umount or just a given amount of time. Chris was able to reproduce
this with stress.sh and lots of processes and so was I. Basically the
overcommit stuff would really let us get out of hand, in my tests I saw up
to 30 gigs of outstanding reservations with only 2 gigs total of metadata
space. This usually worked out fine but with so much outstanding
reservation the flushing stuff short circuits to make sure we don't hang
forever flushing when we really need ENOSPC. Plus we allocate chunks in
order to alleviate the pressure, but this doesn't actually help us since we
only use the non-allocated area in our over commit logic.
So instead of basing overcommit on the amount of non-allocated space,
instead just do it based on how much total space we have, and then limit it
to the non-allocated space in case we are short on space to spill over into.
This allows us to have the same performance as well as no longer giving
random ENOSPC. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Dave sent me a panic where we were doing the orphan cleanup and panic'ed
trying to release our reservation from the orphan block rsv. The reason for
this is because our orphan block rsv had been free'd out from underneath us
because the transaction commit found that there were no orphan inodes
according to its count and decided to free it. This is incorrect so make
sure we inc the orphan inodes count so the accounting is all done properly.
This would also cause the warning in the orphan commit code normally if you
had any orphans to cleanup as they would only decrement the orphan count so
you'd get a negative orphan count which could cause problems during runtime.
Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
When a transaction aborts or there's an EIO on an ordered extent or any
error really we will not free up the space we reserved for this ordered
extent. This results in warnings from the block group cache cleanup in the
case of a transaction abort, or leaking space in the case of EIO on an
ordered extent. Fix this up by free'ing the reserved space if we have an
error at all trying to complete an ordered extent. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
When we abort we've been just free'ing up all the ordered extents and
hoping for the best. This results in lots of warnings from various places,
warnings from btrfs_destroy_inode() because it's ENOSPC accounting isn't
fixed. It will also screw up lots of pages who have been set private but
never get cleared because the ordered extents are never allowed to be
submitted. This patch fixes those warnings. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
I hit this error when reproducing a bug that would end in a transaction
abort. We take the delayed ref head's mutex to keep anybody from processing
it while we're destroying it, but we fail to drop the mutex before we carry
on and free the damned thing. Fix this by doing the remove logic for the
head ourselves and unlock the mutex, that way we can avoid use after free's
or hung tasks waiting on that mutex to come back so they know the delayed
ref completed. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
WARN_ON isn't enough, we need to stop the loop if for any reason
we would overrun the devices_info array.
I tried to track down the connection between the length of
the alloc_devices list and the rw_devices counter but
it wasn't immediately obvious, so be defensive about it.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
__btrfs_std_error didn't always properly call va_end,
and might call va_start even if fmt was NULL.
Move all the varargs handling into the block where we
have fmt.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
I don't think that BTRFS_DEV_EXTENT_KEY is supposed
to fall through to BTRFS_DEV_STATS_KEY ...
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
At least backref_tree_panic() can apparently pass
in a null fs_info, so handle that in __btrfs_panic
to get the message out on the console.
The btrfs_panic macro also uses fs_info, but that's
largely pointless; it's testing to see if
BTRFS_MOUNT_PANIC_ON_FATAL_ERROR is not set.
But if it *were* set, __btrfs_panic() would have,
well, paniced and we wouldn't be here, testing it!
So just BUG() at this point.
And since we only use fs_info once now, just use it
directly.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
No need to test the result, we can't get a
null pointer from list_entry()
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Because of how little we allocate chunks now we can get really tight on
metadata space before we will allocate a new chunk. This resulted in being
unable to add device extents when allocating a new metadata chunk as we did
not have enough space. This is because we were allowed to overcommit too
much metadata without actually making sure we had enough space to make
allocations. The idea behind overcommit is that we are allowed to say "sure
you can have that reservation" when most of the free space is occupied by
reservations, not actual allocations. But in this case where a majority of
the total space is in use by actual allocations we can screw ourselves by
not being able to make real allocations when it matters. So make sure we
have enough real space for our global reserve, and if not then don't allow
overcommitting. Thanks,
Reported-and-tested-by: Jim Schutt <jaschut@sandia.gov>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
I got a double free error when unmounting a file system that failed to add a
chunk during its operation. This is because we will kfree the mapping that
we created but leave the extent_map in the em_tree for chunks. So to fix
this just remove the extent_map when we error out so we don't run into this
problem. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
If we error out allocating a dev extent we will have already created the
block group and such which will cause problems since the allocator may have
tried to allocate out of the block group that no longer exists. This will
cause BUG_ON()'s in the bio submission path. This also makes a failure to
allocate a dev extent a non-abort error, we will just clean up the dev
extents we did allocate and exit. Now if we fail to delete the dev extents
we will abort since we can't have half of the dev extents hanging around,
but this will make us much less likely to abort. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
There is no lock to protect fs_info->fs_state, it will introduce
some problems, such as the value may be covered by the other task
when several tasks modify it. For example:
Task0 - CPU0 Task1 - CPU1
mov %fs_state rax
or $0x1 rax
mov %fs_state rax
or $0x2 rax
mov rax %fs_state
mov rax %fs_state
The expected value is 3, but in fact, it is 2.
Though this problem doesn't happen now (because there is only one
flag currently), the code is error prone, if we add other flags,
the above problem will happen to a certainty.
Now we use bit operation for it to fix the above problem.
In this way, we can make the code more robust and be easy to
add new flags.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
There is no lock to protect
fs_info->avail_{data, metadata, system}_alloc_bits,
it may introduce some problem, such as the wrong profile
information, so we add a seqlock to protect them.
Signed-off-by: Zhao Lei <zhaolei@cn.fujitsu.com>
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
We need not use a global lock to protect the delalloc_bytes of the
inode, just use its own lock. In this way, we can reduce the lock
contention and ->delalloc_lock will just protect delalloc inode
list.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
fs_info->delalloc_bytes is accessed very frequently, so use percpu
counter instead of the u64 variant for it to reduce the lock
contention.
This patch also fixed the problem that we access the variant
without the lock protection.At worst, we would not flush the
delalloc inodes, and just return ENOSPC error when we still have
some free space in the fs.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
->dirty_metadata_bytes is accessed very frequently, so use percpu
counter instead of the u64 variant to reduce the contention of
the lock.
This patch also fixed the problem that we access it without
lock protection in __btrfs_btree_balance_dirty(), which may
cause we skip the dirty pages flush.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
fs_info->alloc_start is a 64bits variant, can be accessed by
multi-task, but it is not protected strictly, it can be changed
while we are accessing it. On 32bit machine, we will get wrong
value because we access it by two instructions.(In fact, it is
also possible that the same problem happens on the 64bit machine,
because the compiler may split the 64bit operation into two 32bit
operation.)
For example:
Assuming -> alloc_start is 0x0000 0000 0001 0000 at the beginning,
then we remount and set ->alloc_start to 0x0000 0100 0000 0000.
Task0 Task1
load high 32 bits
set high 32 bits
set low 32 bits
load low 32 bits
Task1 will get 0.
This patch fixes this problem by using two locks to protect it
fs_info->chunk_mutex
sb->s_umount
On the read side, we just need get one of these two locks, and on
the write side, we must lock all of them.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Though ->max_inline is a 64bit variant, and may be accessed by
multi-task, but it is just suggestive number, so we needn't add
anything to protect fs_info->max_inline, just add a comment to
explain wny we don't use a lock to protect it.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
The header file will then be installed under /usr/include/linux so that
userspace applications can refer to Btrfs ioctls by name and use the same
structs used internally in the kernel.
Signed-off-by: Filipe Brandenburger <filbranden@google.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
CAP_DAC_READ_SEARCH overrides read and search permission check on
file and directory. It seems fit for BTRFS_IOC_INO_PATHS.
Signed-off-by: Kusanagi Kouichi <slash@ac.auone-net.jp>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
This reverts commit 2794ed013b.
Wasn't supposed to get used in btrfs_mknod, it was supposed to be in
btrfs_create, which was done in commit
9185aa587b.
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
btrfs_run_ordered_operations() needn't traverse the ordered operation list
repeatedly, it is because the transaction commiter will invoke it again when
there is no other writer in this transaction, it can ensure that no one can
add new objects into the ordered operation list.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
btrfs_start_delalloc_inodes() needn't traverse and flush the delalloc inodes
repeatedly. It is because we can regard the data that the users write after
we start delalloc inodes flush as the one which is after the delalloc inodes
flush is done, and we can flush it next time.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
We forget to check the return value of btrfs_run_ordered_operations() when
flushing all the pending stuffs, fix it.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
We forget to check the return value of btrfs_start_delalloc_inodes(), fix it.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
The current code of raid attr arry is hard to understand and it is easy to
introduce some problem if we modify the array. So I changed it and made it
more readable.
Cc: Liu Bo <bo.li.liu@oracle.com>
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
This'd save us a rbtree search which may become expensive in large filesystem.
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
This does not change the logic of code, but can save us a read_lock.
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
The API in tree log code has done sort of changes, and it proves that
we can benifit from using token, so do the same thing here.
function_graph tracer's timer shows that it costs nearly half time
of before(39.788us -> 22.391us).
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
