Ben points out that:
When writing files out using O_SYNC, jbd's 1 jiffy delay results in a
significant drop in throughput as the disk sits idle. The patch below
results in a 4-5x performance improvement (from 6.5MB/s to ~24-30MB/s on my
IDE test box) when writing out files using O_SYNC.
So optimise the batching code by omitting it entirely if the process which is
doing a sync write is the same as the one which did the most recent sync
write. If that's true, we're unlikely to get any other processes joining the
transaction.
(Has been in -mm for ages - it took me a long time to get on to performance
testing it)
Numbers, on write-cache-disabled IDE:
/usr/bin/time -p synctest -n 10 -uf -t 1 -p 1 dir-name
Unpatched:
40 seconds
Patched:
35 seconds
Batching disabled:
35 seconds
This is the problematic single-process-doing-fsync case. With multiple
fsyncing processes the numbers are AFACIT unaltered by the patch.
Aside: performance testing and instrumentation shows that the transaction
batching almost doesn't help (testing with synctest -n 1 -uf -t 100 -p 10
dir-name on non-writeback-caching IDE). This is because by the time one
process is running a synchronous commit, a bunch of other processes already
have a transaction handle open, so they're all going to batch into the same
transaction anyway.
The batching seems to offer maybe 5-10% speedup with this workload, but I'm
pretty sure it was more important than that when it was first developed 4-odd
years ago...
Cc: "Stephen C. Tweedie" <sct@redhat.com>
Cc: Benjamin LaHaise <bcrl@kvack.org>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This is the fs/ part of the big kfree cleanup patch.
Remove pointless checks for NULL prior to calling kfree() in fs/.
Signed-off-by: Jesper Juhl <jesper.juhl@gmail.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
- ->releasepage() annotated (s/int/gfp_t), instances updated
- missing gfp_t in fs/* added
- fixed misannotation from the original sweep caught by bitwise checks:
XFS used __nocast both for gfp_t and for flags used by XFS allocator.
The latter left with unsigned int __nocast; we might want to add a
different type for those but for now let's leave them alone. That,
BTW, is a case when __nocast use had been actively confusing - it had
been used in the same code for two different and similar types, with
no way to catch misuses. Switch of gfp_t to bitwise had caught that
immediately...
One tricky bit is left alone to be dealt with later - mapping->flags is
a mix of gfp_t and error indications. Left alone for now.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Use schedule_timeout_{,un}interruptible() instead of
set_current_state()/schedule_timeout() to reduce kernel size. Also use helper
functions to convert between human time units and jiffies rather than constant
HZ division to avoid rounding errors.
Signed-off-by: Nishanth Aravamudan <nacc@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
attached patch should fix the following race:
Proc 1 Proc 2
__flush_batch()
ll_rw_block()
do_get_write_access()
lock_buffer
jh is only waiting for checkpoint
-> b_transaction == NULL ->
do nothing
unlock_buffer
test_set_buffer_locked()
test_clear_buffer_dirty()
__journal_file_buffer()
change the data
submit_bh()
and we have sent wrong data to disk... We now clean the dirty buffer flag
under buffer lock in all cases and hence we know that whenever a buffer is
starting to be journaled we either finish the pending write-out before
attaching a buffer to a transaction or we won't write the buffer until the
transaction is going to be committed.
The test in jbd_unexpected_dirty_buffer() is redundant - remove it.
Furthermore we have to clear the buffer dirty bit under the buffer lock to
prevent races with buffer write-out (and hence prevent returning a buffer with
IO happening).
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This fixes the lots-of-fsx-linux-instances-cause-a-slow-leak bug.
It's been there since 2.6.6, caused by:
ftp://ftp.kernel.org/pub/linux/kernel/people/akpm/patches/2.6/2.6.5/2.6.5-mm4/broken-out/jbd-move-locked-buffers.patch
That patch moves under-writeout ordered-data buffers onto a separate journal
list during commit. It took out the old code which was based on a single
list.
The old code (necessarily) had logic which would restart I/O against buffers
which had been redirtied while they were on the committing transaction's
t_sync_datalist list. The new code only writes buffers once, ignoring
redirtyings by a later transaction, which is good.
But over on the truncate side of things, in journal_unmap_buffer(), we're
treating buffers on the t_locked_list as inviolable things which belong to the
committing transaction, and we just leave them alone during concurrent
truncate-vs-commit.
The net effect is that when truncate tries to invalidate a page whose buffers
are on t_locked_list and have been redirtied, journal_unmap_buffer() just
leaves those buffers alone. truncate will remove the page from its mapping
and we end up with an anonymous clean page with dirty buffers, which is an
illegal state for a page. The JBD commit will not clean those buffers as they
are removed from t_locked_list. The VM (try_to_free_buffers) cannot reclaim
these pages.
The patch teaches journal_unmap_buffer() about buffers which are on the
committing transaction's t_locked_list. These buffers have been written and
I/O has completed. We can take them off the transaction and undirty them
within the context of journal_invalidatepage()->journal_unmap_buffer().
Acked-by: "Stephen C. Tweedie" <sct@redhat.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Initial git repository build. I'm not bothering with the full history,
even though we have it. We can create a separate "historical" git
archive of that later if we want to, and in the meantime it's about
3.2GB when imported into git - space that would just make the early
git days unnecessarily complicated, when we don't have a lot of good
infrastructure for it.
Let it rip!
