Akira-san has been reporting rare deadlocks of his machine when running
xfstests test 269 on ext4 filesystem. The problem turned out to be in
ext4_da_reserve_metadata() and ext4_da_reserve_space() which called
ext4_should_retry_alloc() while holding i_data_sem. Since
ext4_should_retry_alloc() can force a transaction commit, this is a
lock ordering violation and leads to deadlocks.

Fix the problem by just removing the retry loops. These functions should
just report ENOSPC to the caller (e.g. ext4_da_write_begin()) and that
function must take care of retrying after dropping all necessary locks.
Reported-and-tested-by: NAkira Fujita <a-fujita@rs.jp.nec.com>
Reviewed-by: NZheng Liu <wenqing.lz@taobao.com>
Signed-off-by: NJan Kara <jack@suse.cz>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>
Cc: stable@vger.kernel.org

Return a non-zero st_blocks to userspace for statfs() and friends.
Some versions of tar will assume that files with st_blocks == 0
do not contain any data and will skip reading them entirely.
Signed-off-by: NAndreas Dilger <andreas.dilger@intel.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

Pair the two trace events to make troubeshooting writepages
easier, and it should be more convinient to write a simple script
to parse the traces.

Cc: linux-ext4@vger.kernel.org
Cc: Jan Kara <jack@suse.cz>
Signed-off-by: NMing Lei <ming.lei@canonical.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

Commit 4e7ea81d(ext4: restructure writeback path) introduces another
performance regression on random write:

- one more page may be added to ext4 extent in
  mpage_prepare_extent_to_map, and will be submitted for I/O so
  nr_to_write will become -1 before 'done' is set

- the worse thing is that dirty pages may still be retrieved from page
  cache after nr_to_write becomes negative, so lots of small chunks
  can be submitted to block device when page writeback is catching up
  with write path, and performance is hurted.

On one arm A15 board with sata 3.0 SSD(CPU: 1.5GHz dura core, RAM:
2GB, SATA controller: 3.0Gbps), this patch can improve below test's
result from 157MB/sec to 174MB/sec(>10%):

	dd if=/dev/zero of=./z.img bs=8K count=512K

The above test is actually prototype of block write in bonnie++
utility.

This patch makes sure no more pages than nr_to_write can be added to
extent for mapping, so that nr_to_write won't become negative.

Cc: linux-ext4@vger.kernel.org
Acked-by: NJan Kara <jack@suse.cz>
Signed-off-by: NMing Lei <ming.lei@canonical.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

Document give_up_on_write argument of mpage_map_and_submit_extent().
Signed-off-by: NJan Kara <jack@suse.cz>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

The Linux Kernel Performance project guys have reported that commit
4e7ea81d introduces a performance regression for the following fio
workload:

[global]
direct=0
ioengine=mmap
size=1500M
bs=4k
pre_read=1
numjobs=1
overwrite=1
loops=5
runtime=300
group_reporting
invalidate=0
directory=/mnt/
file_service_type=random:36
file_service_type=random:36

[job0]
startdelay=0
rw=randrw
filename=data0/f1:data0/f2

[job1]
startdelay=0
rw=randrw
filename=data0/f2:data0/f1
...

[job7]
startdelay=0
rw=randrw
filename=data0/f2:data0/f1

The culprit of the problem is that after the commit ext4_writepages()
are more aggressive in writing back pages. Thus we have less consecutive
dirty pages resulting in more seeking.

This increased aggressivity is caused by a bug in the condition
terminating ext4_writepages(). We start writing from the beginning of
the file even if we should have terminated ext4_writepages() because
wbc->nr_to_write <= 0.

After fixing the condition the throughput of the fio workload is about 20%
better than before writeback reorganization.
Reported-by: N"Yan, Zheng" <zheng.z.yan@intel.com>
Signed-off-by: NJan Kara <jack@suse.cz>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

truncate_pagecache() doesn't care about old size since commit
cedabed4 ("vfs: Fix vmtruncate() regression").  Let's drop it.
Signed-off-by: NKirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Add support to the core direct-io code to defer AIO completions to user
context using a workqueue.  This replaces opencoded and less efficient
code in XFS and ext4 (we save a memory allocation for each direct IO)
and will be needed to properly support O_(D)SYNC for AIO.

The communication between the filesystem and the direct I/O code requires
a new buffer head flag, which is a bit ugly but not avoidable until the
direct I/O code stops abusing the buffer_head structure for communicating
with the filesystems.

Currently this creates a per-superblock unbound workqueue for these
completions, which is taken from an earlier patch by Jan Kara.  I'm
not really convinced about this use and would prefer a "normal" global
workqueue with a high concurrency limit, but this needs further discussion.

JK: Fixed ext4 part, dynamic allocation of the workqueue.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJan Kara <jack@suse.cz>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

Signed-off-by: NAnatol Pomozov <anatol.pomozov@gmail.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

Use wait_for_stable_page() instead of wait_on_page_writeback()
Signed-off-by: NDmitry Monakhov <dmonakhov@openvz.org>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>
Reviewed-by: NJan Kara <jack@suse.cz>

The following race can lead to a loss of i_disksize update from truncate
thus resulting in a wrong inode size if the inode size isn't updated
again before inode is reclaimed:

ext4_setattr()				mpage_map_and_submit_extent()
  EXT4_I(inode)->i_disksize = attr->ia_size;
  ...					  ...
					  disksize = ((loff_t)mpd->first_page) << PAGE_CACHE_SHIFT
					  /* False because i_size isn't
					   * updated yet */
					  if (disksize > i_size_read(inode))
					  /* True, because i_disksize is
					   * already truncated */
					  if (disksize > EXT4_I(inode)->i_disksize)
					    /* Overwrite i_disksize
					     * update from truncate */
					    ext4_update_i_disksize()
  i_size_write(inode, attr->ia_size);

For other places updating i_disksize such race cannot happen because
i_mutex prevents these races. Writeback is the only place where we do
not hold i_mutex and we cannot grab it there because of lock ordering.

We fix the race by doing both i_disksize and i_size update in truncate
atomically under i_data_sem and in mpage_map_and_submit_extent() we move
the check against i_size under i_data_sem as well.
Signed-off-by: NJan Kara <jack@suse.cz>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>
Cc: stable@vger.kernel.org

Merge conditions in ext4_setattr() handling inode size changes, also
move ext4_begin_ordered_truncate() call somewhat earlier because it
simplifies error recovery in case of failure. Also add error handling in
case i_disksize update fails.
Signed-off-by: NJan Kara <jack@suse.cz>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>
Cc: stable@vger.kernel.org

Inode size can arbitrarily change while writeback is in progress. When
ext4_writepages() has prepared a long extent for mapping and truncate
then reduces i_size, mpage_map_and_submit_buffers() will always map just
one buffer in a page instead of all of them due to lblk < blocks check.
So we end up not using all blocks we've allocated (thus leaking them)
and also delalloc accounting goes wrong manifesting as a warning like:

ext4_da_release_space:1333: ext4_da_release_space: ino 12, to_free 1
with only 0 reserved data blocks

Note that the problem can happen only when blocksize < pagesize because
otherwise we have only a single buffer in the page.

Fix the problem by removing the size check from the mapping loop. We
have an extent allocated so we have to use it all before checking for
i_size. We also rename add_page_bufs_to_extent() to
mpage_process_page_bufs() and make that function submit the page for IO
if all buffers (upto EOF) in it are mapped.
Reported-by: NDave Jones <davej@redhat.com>
Reported-by: NZheng Liu <gnehzuil.liu@gmail.com>
Signed-off-by: NJan Kara <jack@suse.cz>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>
Cc: stable@vger.kernel.org

Currently the logic whether the current buffer can be added to an extent
of buffers to map is split between mpage_add_bh_to_extent() and
add_page_bufs_to_extent(). Move the whole logic to
mpage_add_bh_to_extent() which makes things a bit more straightforward
and make following i_size fixes easier.
Signed-off-by: NJan Kara <jack@suse.cz>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>
Cc: stable@vger.kernel.org

Don't use an unsigned long long for the es_status flags; this requires
that we pass 64-bit values around which is painful on 32-bit systems.
Instead pass the extent status flags around using the low 4 bits of an
unsigned int, and shift them into place when we are reading or writing
es_pblk.
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>
Reviewed-by: NZheng Liu <wenqing.lz@taobao.com>

Commit 0713ed0c added
jbd2_journal_file_inode() call into ext4_block_zero_page_range().
However that function gets called from truncate path and thus inode
needn't have jinode attached - that happens in ext4_file_open() but
the file needn't be ever open since mount. Calling
jbd2_journal_file_inode() without jinode attached results in the oops.

We fix the problem by attaching jinode to inode also in ext4_truncate()
and ext4_punch_hole() when we are going to zero out partial blocks.
Reported-by: Nmajianpeng <majianpeng@gmail.com>
Signed-off-by: NJan Kara <jack@suse.cz>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

In commit 921f266b: ext4: add self-testing infrastructure to do a
sanity check, some sanity checks were added in map_blocks to make sure
'retval == map->m_len'.

Enable these checks by default and report any assertion failures using
ext4_warning() and WARN_ON() since they can help us to figure out some
bugs that are otherwise hard to hit.
Signed-off-by: NZheng Liu <wenqing.lz@taobao.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

If there are no items in the extent status tree, ext4_es_lru_add() is
a no-op.  So it is not sufficient to call ext4_es_lru_add() before we
try to lookup an entry in the extent status tree.  We also need to
call it at the end of ext4_ext_map_blocks(), after items have been
added to the extent status tree.

This could lead to inodes with that have extent status trees but which
are not in the LRU list, which means they won't get considered for
eviction by the es_shrinker.
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>
Cc: Zheng Liu <wenqing.lz@taobao.com>
Cc: stable@vger.kernel.org

Replace "assertation" with "assertion" in lots and lots of debugging
messages.

Correct the comment stating when ext4_es_insert_extent() is used.  It
was no doubt tree at one point, but it is no longer true...
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>
Cc: Zheng Liu <gnehzuil.liu@gmail.com>

The loop in mpage_map_and_submit_extent() is guaranteed to always run
at least once since the caller of mpage_map_and_submit_extent() makes
sure map->m_len > 0. So make that explicit using do-while instead of
pure while which also silences the compiler warning about
uninitialized 'err' variable.
Signed-off-by: NJan Kara <jack@suse.cz>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>
Reviewed-by: NLukas Czerner <lczerner@redhat.com>

The function mpage_released_unused_page() must only be called once;
otherwise the kernel will BUG() when the second call to
mpage_released_unused_page() tries to unlock the pages which had been
unlocked by the first call.

Also restructure the error handling so that we only give up on writing
the dirty pages in the case of ENOSPC where retrying the allocation
won't help.  Otherwise, a transient failure, such as a kmalloc()
failure in calling ext4_map_blocks() might cause us to give up on
those pages, leading to a scary message in /var/log/messages plus data
loss.
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>
Reviewed-by: NJan Kara <jack@suse.cz>

Currently if we pass range into ext4_zero_partial_blocks() which covers
entire block we would attempt to zero it even though we should only zero
unaligned part of the block.

Fix this by checking whether the range covers the whole block skip
zeroing if so.
Signed-off-by: NLukas Czerner <lczerner@redhat.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

The function ext4_write_inline_data_end() can return an error.  So we
need to assign it to a signed integer variable to check for an error
return (since copied is an unsigned int).
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>
Cc: Zheng Liu <wenqing.lz@taobao.com>
Cc: stable@vger.kernel.org

Comparing unsigned variable with 0 always returns false.
err = 0 is duplicated and unnecessary.

[ tytso: Also cleaned up error handling in ext4_block_zero_page_range() ]
Signed-off-by: N"Jon Ernst" <jonernst07@gmx.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

No need to pass file pointer when we can directly pass inode pointer.
Signed-off-by: NAshish Sangwan <a.sangwan@samsung.com>
Signed-off-by: NNamjae Jeon <namjae.jeon@samsung.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

Now we maintain an proper in-order LRU list in ext4 to reclaim entries
from extent status tree when we are under heavy memory pressure.  For
keeping this order, a spin lock is used to protect this list.  But this
lock burns a lot of CPU time.  We can use the following steps to trigger
it.

  % cd /dev/shm
  % dd if=/dev/zero of=ext4-img bs=1M count=2k
  % mkfs.ext4 ext4-img
  % mount -t ext4 -o loop ext4-img /mnt
  % cd /mnt
  % for ((i=0;i<160;i++)); do truncate -s 64g $i; done
  % for ((i=0;i<160;i++)); do cp $i /dev/null &; done
  % perf record -a -g
  % perf report

This commit tries to fix this problem.  Now a new member called
i_touch_when is added into ext4_inode_info to record the last access
time for an inode.  Meanwhile we never need to keep a proper in-order
LRU list.  So this can avoid to burns some CPU time.  When we try to
reclaim some entries from extent status tree, we use list_sort() to get
a proper in-order list.  Then we traverse this list to discard some
entries.  In ext4_sb_info, we use s_es_last_sorted to record the last
time of sorting this list.  When we traverse the list, we skip the inode
that is newer than this time, and move this inode to the tail of LRU
list.  When the head of the list is newer than s_es_last_sorted, we will
sort the LRU list again.

In this commit, we break the loop if s_extent_cache_cnt == 0 because
that means that all extents in extent status tree have been reclaimed.

Meanwhile in this commit, ext4_es_{un}register_shrinker()'s prototype is
changed to save a local variable in these functions.
Reported-by: NDave Hansen <dave.hansen@intel.com>
Signed-off-by: NZheng Liu <wenqing.lz@taobao.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

Rename ext4_da_writepages() to ext4_writepages() and use it for all
modes.  We still need to iterate over all the pages in the case of
data=journalling, but in the case of nodelalloc/data=ordered (which is
what file systems mounted using ext3 backwards compatibility will use)
this will allow us to use a much more efficient I/O submission path.
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

Now that we clear PageWriteback after extent conversion, there's no
need to wait for io_end processing in ext4_evict_inode().  Running
AIO/DIO keeps file reference until aio_complete() is called so
ext4_evict_inode() cannot be called.  For io_end structures resulting
from buffered IO waiting is happening because we wait for
PageWriteback in truncate_inode_pages().
Reviewed-by: NZheng Liu <wenqing.lz@taobao.com>
Signed-off-by: NJan Kara <jack@suse.cz>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

We don't have to wait for extent conversion in ext4_punch_hole() as
buffered IO for the punched range has been flushed and waited upon
(thus all extent conversions for that range have completed).  Also we
wait for all DIO to finish using inode_dio_wait() so there cannot be
any extent conversions pending due to direct IO.

Also remove ext4_flush_unwritten_io() since it's unused now.
Reviewed-by: NZheng Liu <wenqing.lz@taobao.com>
Signed-off-by: NJan Kara <jack@suse.cz>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

Since PageWriteback bit is now cleared after extents are converted
from unwritten to written ones, we have full exclusion of writeback
path from truncate (truncate_inode_pages() waits for PageWriteback
bits to get cleared on all invalidated pages).  Exclusion from DIO
path is achieved by inode_dio_wait() call in ext4_setattr().  So
there's no need to wait for extent convertion in ext4_truncate()
anymore.
Reviewed-by: NZheng Liu <wenqing.lz@taobao.com>
Signed-off-by: NJan Kara <jack@suse.cz>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

Make sure extent conversion after DIO happens while i_dio_count is
still elevated so that inode_dio_wait() waits until extent conversion
is done.  This removes the need for explicit waiting for extent
conversion in some cases.
Reviewed-by: NZheng Liu <wenqing.lz@taobao.com>
Signed-off-by: NJan Kara <jack@suse.cz>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

Later we would like to clear PageWriteback bit only after extent
conversion from unwritten to written extents is performed.  However it
is not possible to start a transaction after PageWriteback is set
because that violates lock ordering (and is easy to deadlock).  So we
have to reserve a transaction before locking pages and sending them
for IO and later we use the transaction for extent conversion from
ext4_end_io().
Reviewed-by: NZheng Liu <wenqing.lz@taobao.com>
Signed-off-by: NJan Kara <jack@suse.cz>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

There isn't any need for setting BH_Uninit on buffers anymore.  It was
only used to signal we need to mark io_end as needing extent
conversion in add_bh_to_extent() but now we can mark the io_end
directly when mapping extent.
Reviewed-by: NZheng Liu <wenqing.lz@taobao.com>
Signed-off-by: NJan Kara <jack@suse.cz>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

There are two issues with current writeback path in ext4.  For one we
don't necessarily map complete pages when blocksize < pagesize and
thus needn't do any writeback in one iteration.  We always map some
blocks though so we will eventually finish mapping the page.  Just if
writeback races with other operations on the file, forward progress is
not really guaranteed. The second problem is that current code
structure makes it hard to associate all the bios to some range of
pages with one io_end structure so that unwritten extents can be
converted after all the bios are finished.  This will be especially
difficult later when io_end will be associated with reserved
transaction handle.

We restructure the writeback path to a relatively simple loop which
first prepares extent of pages, then maps one or more extents so that
no page is partially mapped, and once page is fully mapped it is
submitted for IO. We keep all the mapping and IO submission
information in mpage_da_data structure to somewhat reduce stack usage.
Resulting code is somewhat shorter than the old one and hopefully also
easier to read.
Reviewed-by: NZheng Liu <wenqing.lz@taobao.com>
Signed-off-by: NJan Kara <jack@suse.cz>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

We limit the number of blocks written in a single loop of
ext4_da_writepages() to 64 when inode uses indirect blocks.  That is
unnecessary as credit estimates for mapping logically continguous run
of blocks is rather low even for inode with indirect blocks.  So just
lift this limitation and properly calculate the number of necessary
credits.

This better credit estimate will also later allow us to always write
at least a single page in one iteration.
Reviewed-by: NZheng Liu <wenqing.lz@taobao.com>
Signed-off-by: NJan Kara <jack@suse.cz>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

ext4_ind_trans_blocks() wrongly used 'chunk' argument to decide whether
blocks mapped are logically contiguous. That is wrong since the argument
informs whether the blocks are physically contiguous. As the blocks
mapped are always logically contiguous and that's all
ext4_ind_trans_blocks() cares about, just remove the 'chunk' argument.
Reviewed-by: NZheng Liu <wenqing.lz@taobao.com>
Signed-off-by: NJan Kara <jack@suse.cz>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

Writeback code got better in how it submits IO and now the number of
pages requested to be written is usually higher than original 1024.
The number is now dynamically computed based on observed throughput
and is set to be about 0.5 s worth of writeback.  E.g. on ordinary
SATA drive this ends up somewhere around 10000 as my testing shows.
So remove the unnecessary smarts from ext4_da_writepages().
Reviewed-by: NZheng Liu <wenqing.lz@taobao.com>
Signed-off-by: NJan Kara <jack@suse.cz>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

Change writeback path to create just one io_end structure for the
extent to which we submit IO and share it among bios writing that
extent. This prevents needless splitting and joining of unwritten
extents when they cannot be submitted as a single bio.

Bugs in ENOMEM handling found by Linux File System Verification project
(linuxtesting.org) and fixed by Alexey Khoroshilov
<khoroshilov@ispras.ru>.

CC: Alexey Khoroshilov <khoroshilov@ispras.ru>
Signed-off-by: NJan Kara <jack@suse.cz>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

The arithmetics adding delalloc blocks to the number of used blocks in
ext4_getattr() can easily overflow on 32-bit archs as we first multiply
number of blocks by blocksize and then divide back by 512. Make the
arithmetics more clever and also use proper type (unsigned long long
instead of unsigned long).
Signed-off-by: NJan Kara <jack@suse.cz>
Signed-off-by: NTheodore Ts'o <tytso@mit.edu>

The discard_partial_page_buffers is no longer used anywhere so we can
simply remove it including the *_no_lock variant and
EXT4_DISCARD_PARTIAL_PG_ZERO_UNMAPPED define.
Signed-off-by: NLukas Czerner <lczerner@redhat.com>
Reviewed-by: NJan Kara <jack@suse.cz>
Signed-off-by: NTheodore Ts'o <tytso@mit.edu>