Jon Nelson has found a test case which causes postgresql to fail with
the error:
psql:t.sql:4: ERROR: invalid page header in block 38269 of relation base/16384/16581
Under memory pressure, it looks like part of a file can end up getting
replaced by zero's. Until we can figure out the cause, we'll roll
back the change and use block_write_full_page() instead of
ext4_bio_write_page(). The new, more efficient writing function can
be used via the mount option mblk_io_submit, so we can test and fix
the new page I/O code.
To reproduce the problem, install postgres 8.4 or 9.0, and pin enough
memory such that the system just at the end of triggering writeback
before running the following sql script:
begin;
create temporary table foo as select x as a, ARRAY[x] as b FROM
generate_series(1, 10000000 ) AS x;
create index foo_a_idx on foo (a);
create index foo_b_idx on foo USING GIN (b);
rollback;
If the temporary table is created on a hard drive partition which is
encrypted using dm_crypt, then under memory pressure, approximately
30-40% of the time, pgsql will issue the above failure.
This patch should fix this problem, and the problem will come back if
the file system is mounted with the mblk_io_submit mount option.
Reported-by: Jon Nelson <jnelson@jamponi.net>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
* 'for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tytso/ext4:
ext4: Add new ext4 inode tracepoints
ext4: Don't call sb_issue_discard() in ext4_free_blocks()
ext4: do not try to grab the s_umount semaphore in ext4_quota_off
ext4: fix potential race when freeing ext4_io_page structures
ext4: handle writeback of inodes which are being freed
ext4: initialize the percpu counters before replaying the journal
ext4: "ret" may be used uninitialized in ext4_lazyinit_thread()
ext4: fix lazyinit hang after removing request
Linus noted, and complained to me, that doing while lots of "git diff"'s
of kernel sources, these spinlocks were responsible for 27% of the
spinlock cost on his two-processor system as reported by perf.
Git was doing lots of parallel stats, and this was putting a lot of
pressure on ext4_getattr(). A spinlock to protect a single
memory-to-memory copy is pointless, so remove it.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
We need to make check if a page does not have buffes by checking
page_has_buffers(page) before calling page_buffers(page) in
ext4_writepage(). Otherwise page_buffers() could throw a BUG_ON.
Thanks also to Markus Trippelsdorf and Avinash Kurup who also reported
the problem.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Reported-by: Sedat Dilek <sedat.dilek@googlemail.com>
Tested-by: Sedat Dilek <sedat.dilek@googlemail.com>
Surprisingly chown() on ext4 is not SMP scalable operation.
Due to unconditional orphan_del(NULL, inode) in ext4_setattr()
result in significant performance overhead because of global orphan
mutex, especially in no-journal mode (where orphan_add() is noop).
It is possible to skip explicit orphan_del if possible.
Results of fchown() micro-benchmark in no-journal mode
while (1) {
iteration++;
fchown(fd, uid, gid);
fchown(fd, uid + 1, gid + 1)
}
measured: iterations per millisecond
| nr_tasks | w/o patch | with patch |
| 1 | 142 | 185 |
| 4 | 109 | 642 |
Signed-off-by: Dmitry Monakhov <dmonakhov@openvz.org>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
These functions have no need to be exported beyond file context.
No functions needed to be moved for this commit; just some function
declarations changed to be static and removed from header files.
(A similar patch was submitted by Eric Sandeen, but I wanted to handle
code movement in separate patches to make sure code changes didn't
accidentally get dropped.)
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
As pointed out in a prior patch, updating the mapping's
writeback_index based on pages written isn't quite right;
what the writeback index is really supposed to reflect is
the next page which should be scanned for writeback during
periodic flush.
As in write_cache_pages(), write_cache_pages_da() does
this scanning for us as we assemble the mpd for later
writeout. If we keep track of the next page after the
current scan, we can easily update writeback_index without
worrying about pages written vs. pages skipped, etc.
Without this, an fsync will reset writeback_index to
0 (its starting index) + however many pages it wrote, which
can mess up the progress of periodic flush.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
This is analogous to Jan Kara's commit,
f446daaea9
mm: implement writeback livelock avoidance using page tagging
but since we forked write_cache_pages, we need to reimplement
it there (and in ext4_da_writepages, since range_cyclic handling
was moved to there)
If you start a large buffered IO to a file, and then set
fsync after it, you'll find that fsync does not complete
until the other IO stops.
If you continue re-dirtying the file (say, putting dd
with conv=notrunc in a loop), when fsync finally completes
(after all IO is done), it reports via tracing that
it has written many more pages than the file contains;
in other words it has synced and re-synced pages in
the file multiple times.
This then leads to problems with our writeback_index
update, since it advances it by pages written, and
essentially sets writeback_index off the end of the
file...
With the following patch, we only sync as much as was
dirty at the time of the sync.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
This doesn't fix anything at all, it just removes a vestige
of prior use from __mpage_da_writepage()
__mpage_da_writepage() had a *void argument leftover from
its previous life as a callback; make it reflect the actual type.
Fixing this up makes it slightly more obvious to read, and
enables proper typechecking.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Fail block allocation if sb_getblk() returns NULL. In that case,
sb_find_get_block() also likely to fail so that it should skip
calling ext4_forget().
Signed-off-by: Namhyung Kim <namhyung@gmail.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Call the block I/O layer directly instad of going through the buffer
layer. This should give us much better performance and scalability,
as well as lowering our CPU utilization when doing buffered writeback.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
This massively simplifies the ext4_da_writepages() code path by
completely removing mpage_put_bnr_bhs(), which is almost 100 lines of
code iterating over a set of pages using pagevec_lookup(), and folds
that functionality into mpage_da_submit_io()'s existing
pagevec_lookup() loop.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Expand the call:
if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
ext4_bh_delay_or_unwritten))
goto redirty_page
into mpage_da_submit_io().
This will allow us to merge in mpage_put_bnr_to_bhs() in the next
patch.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
As a prepratory step to switching to bio_submit, inline
ext4_writepage() into mpage_da_submit() and then simplify things a
bit. This makes it clearer what mpage_da_submit needs to do.
Also, move the ClearPageChecked(page) call into
__ext4_journalled_writepage(), as a minor bit of cleanup refactoring.
This also allows us to pull i_size_read() and
ext4_should_journal_data() out of the loop, which should be a very
minor CPU savings.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
The actual code in ext4_writepage() is unnecessarily convoluted.
Simplify it so it is easier to understand, but otherwise logically
equivalent.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Eventually we need to completely reorganize the ext4 writepage
callpath, but for now, we simplify things a little by calling
mpage_da_submit_io() from mpage_da_map_blocks(), since all of the
places where we call mpage_da_map_blocks() it is followed up by a call
to mpage_da_submit_io().
We're also a wee bit better with respect to error handling, but there
are still a number of issues where it's not clear what the right thing
is to do with ext4 functions deep in the writeback codepath fails.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
By queuing the io end on the unwritten workqueue before adding it
to our inode's list of completed IOs, I think we run the risk
of the work getting completed, and the IO freed, before we try
to add it to the inode's i_completed_io_list.
It should be safe to add it to the inode's list of completed
IOs, and -then- queue it for completion, I think.
Thanks to Dave Chinner for pointing out the race.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Jiaying Zhang <jiayingz@google.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
On linux-2.6.36-rc2, if we execute the following script, we can hang
the system when the /bin/sync command is executed:
========================================================================
#!/bin/sh
echo -n "HANG UP TEST: "
/bin/dd if=/dev/zero of=/tmp/img bs=1k count=1 seek=1M 2> /dev/null
/sbin/mkfs.ext4 -Fq /tmp/img
/bin/mount -o loop -t ext4 /tmp/img /mnt
/bin/dd if=/dev/zero of=/mnt/file bs=1 count=1 \
seek=$((16*1024*1024*1024*1024-4096)) 2> /dev/null
/bin/sync
/bin/umount /mnt
echo "DONE"
exit 0
========================================================================
We can see the following backtrace if we get the kdump when this
hangup occurs:
======================================================================
kthread()
=> bdi_writeback_thread()
=> wb_do_writeback()
=> wb_writeback()
=> writeback_inodes_wb()
=> writeback_sb_inodes()
=> writeback_single_inode()
=> ext4_da_writepages() ---+
^ infinite |
| loop |
+-------------+
======================================================================
The reason why this hangup happens is described as follows:
1) We write the last extent block of the file whose size is the filesystem
maximum size.
2) "BH_Delay" flag is set on the buffer_head of its block.
3) - the member, "m_lblk" of struct mpage_da_data is 4294967295 (UINT_MAX)
- the member, "m_len" of struct mpage_da_data is 1
mpage_put_bnr_to_bhs() which is called via ext4_da_writepages()
cannot clear "BH_Delay" flag of the buffer_head because the type of
m_lblk is ext4_lblk_t and then m_lblk + m_len is overflow.
Therefore an infinite loop occurs because ext4_da_writepages()
cannot write the page (which corresponds to the block) since
"BH_Delay" flag isn't cleared.
----------------------------------------------------------------------
static void mpage_put_bnr_to_bhs(struct mpage_da_data *mpd,
struct ext4_map_blocks *map)
{
...
int blocks = map->m_len;
...
do {
// cur_logical = 4294967295
// map->m_lblk = 4294967295
// blocks = 1
// *** map->m_lblk + blocks == 0 (OVERFLOW!) ***
// (cur_logical >= map->m_lblk + blocks) => true
if (cur_logical >= map->m_lblk + blocks)
break;
----------------------------------------------------------------------
NOTE: Mounting with the nodelalloc option will avoid this codepath,
and thus, avoid this hang
Signed-off-by: Toshiyuki Okajima <toshi.okajima@jp.fujitsu.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
I'm uneasy with lots of stuff going on in ext4_da_writepages(),
but bumping nr_to_write from LLONG_MAX to -8 clearly isn't
making anything better, so avoid the multiplier in that case.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Today we simply break out of the inner loop when we have accumulated
max_pages; this keeps scanning forwad and doing pagevec_lookup_tag()
in the while (!done) loop, this does potentially a lot of work
with no net effect.
When we have accumulated max_pages, just clean up and return.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
__block_write_begin and block_prepare_write are identical except for slightly
different calling conventions. Convert all callers to the __block_write_begin
calling conventions and drop block_prepare_write.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs-2.6: (96 commits)
no need for list_for_each_entry_safe()/resetting with superblock list
Fix sget() race with failing mount
vfs: don't hold s_umount over close_bdev_exclusive() call
sysv: do not mark superblock dirty on remount
sysv: do not mark superblock dirty on mount
btrfs: remove junk sb_dirt change
BFS: clean up the superblock usage
AFFS: wait for sb synchronization when needed
AFFS: clean up dirty flag usage
cifs: truncate fallout
mbcache: fix shrinker function return value
mbcache: Remove unused features
add f_flags to struct statfs(64)
pass a struct path to vfs_statfs
update VFS documentation for method changes.
All filesystems that need invalidate_inode_buffers() are doing that explicitly
convert remaining ->clear_inode() to ->evict_inode()
Make ->drop_inode() just return whether inode needs to be dropped
fs/inode.c:clear_inode() is gone
fs/inode.c:evict() doesn't care about delete vs. non-delete paths now
...
Fix up trivial conflicts in fs/nilfs2/super.c
Replace inode_setattr with opencoded variants of it in all callers. This
moves the remaining call to vmtruncate into the filesystem methods where it
can be replaced with the proper truncate sequence.
In a few cases it was obvious that we would never end up calling vmtruncate
so it was left out in the opencoded variant:
spufs: explicitly checks for ATTR_SIZE earlier
btrfs,hugetlbfs,logfs,dlmfs: explicitly clears ATTR_SIZE earlier
ufs: contains an opencoded simple_seattr + truncate that sets the filesize just above
In addition to that ncpfs called inode_setattr with handcrafted iattrs,
which allowed to trim down the opencoded variant.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Split up the block_write_begin implementation - __block_write_begin is a new
trivial wrapper for block_prepare_write that always takes an already
allocated page and can be either called from block_write_begin or filesystem
code that already has a page allocated. Remove the handling of already
allocated pages from block_write_begin after switching all callers that
do it to __block_write_begin.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Move the call to vmtruncate to get rid of accessive blocks to the callers
in prepearation of the new truncate calling sequence. This was only done
for DIO_LOCKING filesystems, so the __blockdev_direct_IO_newtrunc variant
was not needed anyway. Get rid of blockdev_direct_IO_no_locking and
its _newtrunc variant while at it as just opencoding the two additional
paramters is shorted than the name suffix.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
In data=journal mode, we still use block_write_begin() to prepare
page for writing. This function can occasionally mark buffer dirty
which violates journalling assumptions - when a buffer is part of
a transaction, it should be dirty and a buffer can be already part
of a forget list of some transaction when block_write_begin()
gets called. This violation of journalling assumptions then results
in "JBD: Spotted dirty metadata buffer..." warnings.
In fact, temporary dirtying the buffer while the page is still locked
does not really cause problems to the journalling because we won't write
the buffer until the page gets unlocked. So we just have to make sure
to clear dirty bits before unlocking the page.
Signed-off-by: Jan Kara <jack@suse.cz>
Lockstat reports have shown that j_state_lock is a major source of
lock contention, especially on systems with more than 4 CPU cores. So
change it to be a read/write spinlock.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
There were some error paths in ext4_delete_inode() which was not
dropping the inode from the orphan list. This could lead to a BUG_ON
on umount when the orphan list is discovered to be non-empty.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
I often get emails containing the "This should not happen!!" message,
conveniently trimmed to remove things like:
sd 0:0:0:0: [sda] Unhandled error code
sd 0:0:0:0: [sda] Result: hostbyte=DID_OK driverbyte=DRIVER_TIMEOUT
sd 0:0:0:0: [sda] CDB: Write(10): 2a 00 03 13 c9 70 00 00 28 00
end_request: I/O error, dev sda, sector 51628400
Aborting journal on device dm-0-8.
EXT4-fs error (device dm-0): ext4_journal_start_sb: Detected aborted journal
EXT4-fs (dm-0): Remounting filesystem read-only
I don't think there is any value to the verbosity if the reason is
due to a filesystem abort; it just obfuscates the root cause.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
ext4_get_blocks got renamed to ext4_map_blocks, but left stale
comments and a prototype littered around.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
If the user attempts to make a non-extent-mapped file to be too large,
return EFBIG, but don't call ext4_std_err() which will end up marking
the file system as containing an error.
Thanks to Toshiyuki Okajima-san at Fujitsu for pointing this out.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
This patch is to be applied upon Christoph's "direct-io: move aio_complete
into ->end_io" patch. It adds iocb and result fields to struct ext4_io_end_t,
so that we can call aio_complete from ext4_end_io_nolock() after the extent
conversion has finished.
I have verified with Christoph's aio-dio test that used to fail after a few
runs on an original kernel but now succeeds on the patched kernel.
See http://thread.gmane.org/gmane.comp.file-systems.ext4/19659 for details.
Signed-off-by: Jiaying Zhang <jiayingz@google.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Filesystems with unwritten extent support must not complete an AIO request
until the transaction to convert the extent has been commited. That means
the aio_complete calls needs to be moved into the ->end_io callback so
that the filesystem can control when to call it exactly.
This makes a bit of a mess out of dio_complete and the ->end_io callback
prototype even more complicated.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
We have experienced bitmap inconsistencies after crash during file
delete under heavy load. The crash is not file system related and I
the following patch in ext4_free_branches() fixes the recovery
problem.
If the transaction is restarted and there is a crash before the new
transaction is committed, then after recovery, the blocks that this
indirect block points to have been freed, but the indirect block
itself has not been freed and may still point to some of the free
blocks (because of the ext4_forget()).
So ext4_forget() should be called inside ext4_free_blocks() to avoid
this problem.
Signed-off-by: Amir Goldstein <amir73il@users.sf.net>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Save number of file system errors, and the time function name, line
number, block number, and inode number of the first and most recent
errors reported on the file system in the superblock.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Filesystems with unwritten extent support must not complete an AIO request
until the transaction to convert the extent has been commited. That means
the aio_complete calls needs to be moved into the ->end_io callback so
that the filesystem can control when to call it exactly.
This makes a bit of a mess out of dio_complete and the ->end_io callback
prototype even more complicated.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Alex Elder <aelder@sgi.com>
The nobh option was only supported for writeback mode, but given that all
write paths actually create buffer heads it effectively was a no-op already.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
No real bugs found, just removed some dead code.
Found by gcc 4.6's new warnings.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
A few functions were still modifying i_flags in a racy manner.
Signed-off-by: Dmitry Monakhov <dmonakhov@openvz.org>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
* 'for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tytso/ext4: (40 commits)
ext4: Make fsync sync new parent directories in no-journal mode
ext4: Drop whitespace at end of lines
ext4: Fix compat EXT4_IOC_ADD_GROUP
ext4: Conditionally define compat ioctl numbers
tracing: Convert more ext4 events to DEFINE_EVENT
ext4: Add new tracepoints to track mballoc's buddy bitmap loads
ext4: Add a missing trace hook
ext4: restart ext4_ext_remove_space() after transaction restart
ext4: Clear the EXT4_EOFBLOCKS_FL flag only when warranted
ext4: Avoid crashing on NULL ptr dereference on a filesystem error
ext4: Use bitops to read/modify i_flags in struct ext4_inode_info
ext4: Convert calls of ext4_error() to EXT4_ERROR_INODE()
ext4: Convert callers of ext4_get_blocks() to use ext4_map_blocks()
ext4: Add new abstraction ext4_map_blocks() underneath ext4_get_blocks()
ext4: Use our own write_cache_pages()
ext4: Show journal_checksum option
ext4: Fix for ext4_mb_collect_stats()
ext4: check for a good block group before loading buddy pages
ext4: Prevent creation of files larger than RLIMIT_FSIZE using fallocate
ext4: Remove extraneous newlines in ext4_msg() calls
...
Fixed up trivial conflict in fs/ext4/fsync.c
Quota must being initialized if size or uid/git changes requested.
But initialization performed in two different places:
in case of i_size file system is responsible for dquot init
, but in case of uid/gid init will be called internally in
dquot_transfer().
This ambiguity makes code harder to understand.
Let's move this logic to one common helper function.
Signed-off-by: Dmitry Monakhov <dmonakhov@openvz.org>
Signed-off-by: Jan Kara <jack@suse.cz>