We (Linux Kernel Performance project) found a regression introduced
by commit:

  f7fec032 ext4: track all extent status in extent status tree

The commit causes about 20% performance decrease in fio random write
test. Profiler shows that rb_next() uses a lot of CPU time. The call
stack is:

  rb_next
  ext4_es_find_delayed_extent
  ext4_map_blocks
  _ext4_get_block
  ext4_get_block_write
  __blockdev_direct_IO
  ext4_direct_IO
  generic_file_direct_write
  __generic_file_aio_write
  ext4_file_write
  aio_rw_vect_retry
  aio_run_iocb
  do_io_submit
  sys_io_submit
  system_call_fastpath
  io_submit
  td_io_getevents
  io_u_queued_complete
  thread_main
  main
  __libc_start_main

The cause is that ext4_es_find_delayed_extent() doesn't have an
upper bound, it keeps searching until a delayed extent is found.
When there are a lots of non-delayed entries in the extent state
tree, ext4_es_find_delayed_extent() may uses a lot of CPU time.
Reported-by: NLKP project <lkp@linux.intel.com>
Signed-off-by: NYan, Zheng <zheng.z.yan@intel.com>
Signed-off-by: NZheng Liu <wenqing.lz@taobao.com>
Cc: "Theodore Ts'o" <tytso@mit.edu>

Signed-off-by: NDarrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

Move ext4_ind_migrate() into migrate.c file since it makes much more
sense and ext4_ext_migrate() is there as well.

Also fix tiny style problem - add spaces around "=" in "i=0".
Signed-off-by: NLukas Czerner <lczerner@redhat.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

Signed-off-by: NDmitry Monakhov <dmonakhov@openvz.org>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>
Cc: stable@vger.kernel.org

Currently in ENOSPC condition when writing into unwritten space, or
punching a hole, we might need to split the extent and grow extent tree.
However since we can not allocate any new metadata blocks we'll have to
zero out unwritten part of extent or punched out part of extent, or in
the worst case return ENOSPC even though use actually does not allocate
any space.

Also in delalloc path we do reserve metadata and data blocks for the
time we're going to write out, however metadata block reservation is
very tricky especially since we expect that logical connectivity implies
physical connectivity, however that might not be the case and hence we
might end up allocating more metadata blocks than previously reserved.
So in future, metadata reservation checks should be removed since we can
not assure that we do not under reserve.

And this is where reserved space comes into the picture. When mounting
the file system we slice off a little bit of the file system space (2%
or 4096 clusters, whichever is smaller) which can be then used for the
cases mentioned above to prevent costly zeroout, or unexpected ENOSPC.

The number of reserved clusters can be set via sysfs, however it can
never be bigger than number of free clusters in the file system.

Note that this patch fixes the failure of xfstest 274 as expected.
Signed-off-by: NLukas Czerner <lczerner@redhat.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>
Reviewed-by: NCarlos Maiolino <cmaiolino@redhat.com>

existing locking ordering: journal-> i_data_sem, but
ext4_ind_migrate() grab locks in opposite order which may result in
deadlock.
Signed-off-by: NDmitry Monakhov <dmonakhov@openvz.org>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

Currently when inserting extent in ext4_ext_insert_extent() we would
only try to to see if we can append new extent to the found extent. If
we can not, then we proceed with adding new extent into the extent tree,
but then possibly merging it back again.

We can avoid this situation by trying to append and prepend new extent
to the existing ones. However since the new extent can be on either
sides of the existing extent, we have to pick the right extent to try to
append/prepend to.

This patch adds the conditions to pick the right extent to
append/prepend to and adds the actual prepending condition as well. This
will also eliminate the need to use "reserved" block for possibly
growing extent tree.
Signed-off-by: NLukas Czerner <lczerner@redhat.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

Currently when converting extent to initialized we attempt to transfer
initialized block to the left neighbour if possible when certain
criteria are met. However we do not attempt to do the same for the
right neighbor.

This commit adds the possibility to transfer initialized block to the
right neighbour if:

1. We're not converting the whole extent
2. Both extents are stored in the same extent tree node
3. Right neighbor is initialized
4. Right neighbor is logically abutting the current one
5. Right neighbor is physically abutting the current one
6. Right neighbor would not overflow the length limit

This is basically the same logic as with transferring to the left. This
will gain us some performance benefits since it is faster than inserting
extent and then merging it.

It would also prevent some situation in delalloc patch when we might run
out of metadata reservation. This is due to the fact that we would
attempt to split the extent first (possibly allocating new metadata
block) even though we did not counted for that because it can (and will)
be merged again. This commit fix that scenario, because we no longer
need to split the extent in such case.
Signed-off-by: NLukas Czerner <lczerner@redhat.com>

In order to make it simpler to test the code which support
i_blocks/indirect-mapped inodes, support the conversion of inodes
which are less than 12 blocks and which are contained in no more than
a single extent.

The primary intended use of this code is to converting freshly created
zero-length files and empty directories.

Note that the version of chattr in e2fsprogs 1.42.7 and earlier has a
check that prevents the clearing of the extent flag.  A simple patch
which allows "chattr -e <file>" to work will be checked into the
e2fsprogs git repository.
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

Move common code in ext4_ind_truncate() and ext4_ext_truncate() into
ext4_truncate().  This saves over 60 lines of code.
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

Move common code in ext4_ind_punch_hole() and ext4_ext_punch_hole()
into ext4_punch_hole().  This saves over 150 lines of code.

This also fixes a potential bug when the punch_hole() code is racing
against indirect-to-extents or extents-to-indirect migation.  We are
currently using i_mutex to protect against changes to the inode flag;
specifically, the append-only, immutable, and extents inode flags.  So
we need to take i_mutex before deciding whether to use the
extents-specific or indirect-specific punch_hole code.

Also, there was a missing call to ext4_inode_block_unlocked_dio() in
the indirect punch codepath.  This was added in commit 02d262df
to block DIO readers racing against the punch operation in the
codepath for extent-mapped inodes, but it was missing for
indirect-block mapped inodes.  One of the advantages of refactoring
the code is that it makes such oversights much less likely.
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

When an extent was zeroed out, we forgot to do convert from cpu to le16.
It could make us hit a BUG_ON when we try to write dirty pages out.  So
fix it.

[ Also fix a bug found by Dmitry Monakhov where we were missing
  le32_to_cpu() calls in the new indirect punch hole code.

  There are a number of other big endian warnings found by static code
  analyzers, but we'll wait for the next merge window to fix them all
  up.  These fixes are designed to be Obviously Correct by code
  inspection, and easy to demonstrate that it won't make any
  difference (and hence, won't introduce any bugs) on little endian
  architectures such as x86.  --tytso ]
Signed-off-by: NZheng Liu <wenqing.lz@taobao.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>
Reported-by: NCAI Qian <caiqian@redhat.com>
Reported-by: NChristian Kujau <lists@nerdbynature.de>
Cc: Dmitry Monakhov <dmonakhov@openvz.org>

Currently when converting extent to initialized, we have to decide
whether to zeroout part/all of the uninitialized extent in order to
avoid extent tree growing rapidly.

The decision is made by comparing the size of the extent with the
configurable value s_extent_max_zeroout_kb which is in kibibytes units.

However when converting it to number of blocks we currently use it as it
was in bytes. This is obviously bug and it will result in ext4 _never_
zeroout extents, but rather always split and convert parts to
initialized while leaving the rest uninitialized in default setting.

Fix this by using s_extent_max_zeroout_kb as kibibytes.
Signed-off-by: NLukas Czerner <lczerner@redhat.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>
Cc: stable@vger.kernel.org

Currently in ext4_ext_map_blocks() in delayed allocation writeback
we would update the reservation and after that check whether we claimed
cluster outside of the range of the allocation and if so, we'll give the
block back to the reservation pool.

However this also means that if the number of reserved data block
dropped to zero before the correction, we would release all the metadata
reservation as well, however we might still need it because the we're
not done with the delayed allocation and there might be more blocks to
come. This will result in error messages such as:

EXT4-fs warning (device sdb): ext4_da_update_reserve_space:361: ino 12,
allocated 1 with only 0 reserved metadata blocks (releasing 1 blocks
with reserved 1 data blocks)

This will only happen on bigalloc file system and it can be easily
reproduced using fiemap-tester from xfstests like this:

./src/fiemap-tester -m DHDHDHDHD -S -p0 /mnt/test/file

Or using xfstests such as 225.

Fix this by doing the correction first and updating the reservation
after that so that we do not accidentally decrease
i_reserved_data_blocks to zero.
Signed-off-by: NLukas Czerner <lczerner@redhat.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

This commit fixes a wrong return value of the number of the allocated
blocks in ext4_split_extent.  When the length of blocks we want to
allocate is greater than the length of the current extent, we return a
wrong number.  Let's see what happens in the following case when we
call ext4_split_extent().

  map: [48, 72]
  ex:  [32, 64, u]

'ex' will be split into two parts:
  ex1: [32, 47, u]
  ex2: [48, 64, w]

'map->m_len' is returned from this function, and the value is 24.  But
the real length is 16.  So it should be fixed.

Meanwhile in this commit we use right length of the allocated blocks
when get_reserved_cluster_alloc in ext4_ext_handle_uninitialized_extents
is called.
Signed-off-by: NZheng Liu <wenqing.lz@taobao.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>
Cc: Dmitry Monakhov <dmonakhov@openvz.org>
Cc: stable@vger.kernel.org

When we try to split an extent, this extent could be zeroed out and mark
as initialized.  But we don't know this in ext4_map_blocks because it
only returns a length of allocated extent.  Meanwhile we will mark this
extent as uninitialized because we only check m_flags.

This commit update extent status tree when we try to split an unwritten
extent.  We don't need to worry about the status of this extent because
we always mark it as initialized.
Signed-off-by: NZheng Liu <wenqing.lz@taobao.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>
Cc: Dmitry Monakhov <dmonakhov@openvz.org>

The ext4_ext_handle_uninitialized_extents() function was assuming the
return value of ext4_ext_map_blocks() is equal to map->m_len.  This
incorrect assumption was harmless until we started use status tree as
a extent cache because we need to update status tree according to
'm_len' value.

Meanwhile this commit marks EXT4_MAP_MAPPED flag after unwritten extent
conversion.  It shouldn't cause a bug because we update status tree
according to checking EXT4_MAP_UNWRITTEN flag.  But it should be fixed.

After applied this commit, the following error message from self-testing
infrastructure disappears.

    ...
    kernel: ES len assertation failed for inode: 230 retval 1 !=
    map->m_len 3 in ext4_map_blocks (allocation)
    ...
Signed-off-by: NZheng Liu <wenqing.lz@taobao.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>
Cc: Dmitry Monakhov <dmonakhov@openvz.org>

Now that we don't merge uninitialized extents anymore,
ext4_fallocate() is free to operate on the inode while there are still
some extent conversions pending - it won't disturb them in any way.
Reviewed-by: NZheng Liu <wenqing.lz@taobao.com>
Reviewed-by: NDmitry Monakhov <dmonakhov@openvz.org>
Signed-off-by: NJan Kara <jack@suse.cz>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

Splitting extents inside endio is a bad thing, but unfortunately it is
still possible.  In fact we are pretty close to the moment when all
related issues will be fixed.  Let's warn developer if it still the
case.
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>

Derived from Jan's patch:http://permalink.gmane.org/gmane.comp.file-systems.ext4/36470

Merging of uninitialized extents creates all sorts of interesting race
possibilities when writeback / DIO races with fallocate. Thus
ext4_convert_unwritten_extents_endio() has to deal with a case where
extent to be converted needs to be split out first. That isn't nice
for two reasons:

1) It may need allocation of extent tree block so ENOSPC is possible.
2) It complicates end_io handling code

So we disable merging of uninitialized extents which allows us to simplify
the code. Extents will get merged after they are converted to initialized
ones.
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>

When ext4_split_extent_at() ends up doing zeroout & conversion to
initialized instead of split & conversion, ext4_split_extent() gets
confused and can wrongly mark the extent back as uninitialized
resulting in end IO code getting confused from large unwritten extents
and may result in data loss.

The example of problematic behavior is:
			    lblk len              lblk len
  ext4_split_extent() (ex=[1000,30,uninit], map=[1010,10])
    ext4_split_extent_at() (split [1000,30,uninit] at 1020)
      ext4_ext_insert_extent() -> ENOSPC
      ext4_ext_zeroout()
	 -> extent [1000,30] is now initialized
    ext4_split_extent_at() (split [1000,30,init] at 1010,
			     MARK_UNINIT1 | MARK_UNINIT2)
      -> extent is split and parts marked as uninitialized

Fix the problem by rechecking extent type after the first
ext4_split_extent_at() returns. None of split_flags can not be applied
to initialized extent so this patch also add BUG_ON to prevent similar
issues in future.

TESTCASE: https://github.com/dmonakhov/xfstests/commit/b8a55eb5ce28c6ff29e620ab090902fcd5833597Signed-off-by: NDmitry Monakhov <dmonakhov@openvz.org>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>
Reviewed-by: NJan Kara <jack@suse.cz>

Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

Single extent cache could be removed because we have extent status tree
as a extent cache, and it would be better.
Signed-off-by: NZheng Liu <wenqing.lz@taobao.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>
Cc: Jan kara <jack@suse.cz>

After tracking all extent status, we already have a extent cache in
memory.  Every time we want to lookup a block mapping, we can first
try to lookup it in extent status tree to avoid a potential disk I/O.

A new function called ext4_es_lookup_extent is defined to finish this
work.  When we try to lookup a block mapping, we always call
ext4_map_blocks and/or ext4_da_map_blocks.  So in these functions we
first try to lookup a block mapping in extent status tree.

A new flag EXT4_GET_BLOCKS_NO_PUT_HOLE is used in ext4_da_map_blocks
in order not to put a hole into extent status tree because this hole
will be converted to delayed extent in the tree immediately.
Signed-off-by: NZheng Liu <wenqing.lz@taobao.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>
Cc: Jan kara <jack@suse.cz>

By recording the phycisal block and status, extent status tree is able
to track the status of every extents.  When we call _map_blocks
functions to lookup an extent or create a new written/unwritten/delayed
extent, this extent will be inserted into extent status tree.

We don't load all extents from disk in alloc_inode() because it costs
too much memory, and if a file is opened and closed frequently it will
takes too much time to load all extent information.  So currently when
we create/lookup an extent, this extent will be inserted into extent
status tree.  Hence, the extent status tree may not comprehensively
contain all of the extents found in the file.

Here a condition we need to take care is that an extent might contains
unwritten and delayed status simultaneously because an extent is delayed
allocated and could be allocated by fallocate.  At this time we need to
keep delayed status because later we need to update delayed reservation
space using it.
Signed-off-by: NZheng Liu <wenqing.lz@taobao.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>
Cc: Jan kara <jack@suse.cz>

This commit lets ext4_ext_map_blocks return EXT4_MAP_UNWRITTEN flag
because in later commit ext4_map_blocks needs to use this flag to
determine the extent status.
Signed-off-by: NZheng Liu <wenqing.lz@taobao.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>
Reviewed-by: NJan Kara <jack@suse.cz>

This commit renames ext4_es_find_extent with ext4_es_find_delayed_extent
and improve this function.  First, we split input and output parameter.
Second, this function never return the first block of the next delayed
extent after 'es'.
Signed-off-by: NZheng Liu <wenqing.lz@taobao.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>
Cc: Jan kara <jack@suse.cz>

This commit refines the extent status tree code.

1) A prefix 'es_' is added to to the extent status tree structure
members.

2) Refactored es_remove_extent() so that __es_remove_extent() can be
used by es_insert_extent() to remove the old extent entry(-ies) before
inserting a new one.

3) Rename extent_status_end() to ext4_es_end()

4) ext4_es_can_be_merged() is define to check whether two extents can
be merged or not.

5) Update and clarified comments.
Signed-off-by: NZheng Liu <wenqing.lz@taobao.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>
Reviewed-by: NJan Kara <jack@suse.cz>

So we can better understand what bits of ext4 are responsible for
long-running jbd2 handles, use jbd2__journal_start() so we can pass
context information for logging purposes.

The recommended way for finding the longer-running handles is:

   T=/sys/kernel/debug/tracing
   EVENT=$T/events/jbd2/jbd2_handle_stats
   echo "interval > 5" > $EVENT/filter
   echo 1 > $EVENT/enable

   ./run-my-fs-benchmark

   cat $T/trace > /tmp/problem-handles

This will list handles that were active for longer than 20ms.  Having
longer-running handles is bad, because a commit started at the wrong
time could stall for those 20+ milliseconds, which could delay an
fsync() or an O_SYNC operation.  Here is an example line from the
trace file describing a handle which lived on for 311 jiffies, or over
1.2 seconds:

postmark-2917  [000] ....   196.435786: jbd2_handle_stats: dev 254,32 
   tid 570 type 2 line_no 2541 interval 311 sync 0 requested_blocks 1
   dirtied_blocks 0
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

In two places we call WARN_ON() before we print out the debug message,
however we agreed that the WARN_ON() is unnecessary at those places so
remove them.

Also use ext4_warning() instead of ext4_msg() and printk().
Signed-off-by: NLukas Czerner <lczerner@redhat.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

This patch add supports for indirect file support punching hole.  It
is almost the same as ext4_ext_punch_hole.  First, we invalidate all
pages between this hole, and then we try to deallocate all blocks of
this hole.

A recursive function is used to handle deallocation of blocks.  In
this function, it iterates over the entries in inode's i_blocks or
indirect blocks, and try to free the block for each one of them.

After applying this patch, xfstest #255 will not pass w/o extent because
indirect-based file doesn't support unwritten extents.
Signed-off-by: NZheng Liu <wenqing.lz@taobao.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

Because the function 'sb_getblk' seldomly fails to return NULL
value,it will be better to use 'unlikely' to optimize it.
Signed-off-by: NWang Shilong <wangsl-fnst@cn.fujitsu.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

The only reason for sb_getblk() failing is if it can't allocate the
buffer_head.  So ENOMEM is more appropriate than EIO.  In addition,
make sure that the file system is marked as being inconsistent if
sb_getblk() fails.
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>
Cc: stable@vger.kernel.org

When depth of extent tree is greater than 1, logical start value of
interior node is not correctly updated in ext4_ext_rm_idx.
Signed-off-by: NForrest Liu <forrestl@synology.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>
Reviewed-by: NAshish Sangwan <ashishsangwan2@gmail.com>
Cc: stable@vger.kernel.org

Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>
Signed-off-by: NZhi Yong Wu <wuzhy@linux.vnet.ibm.com>
Reviewed-by: NZheng Liu <gnehzuil.liu@gmail.com>

If we are punching hole in a file, we will return ENOTSUPP.
As for the fallocation of some extents, we will convert the
inline data to a normal extent based file first.
Signed-off-by: NTao Ma <boyu.mt@taobao.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

fiemap is used to find the disk layout of a file, as for inline data,
let us just pretend like a file with just one extent.
Signed-off-by: NTao Ma <boyu.mt@taobao.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

For a normal write case (not journalled write, not delayed
allocation), we write to the inline if the file is small and convert
it to an extent based file when the write is larger than the max
inline size.
Signed-off-by: NTao Ma <boyu.mt@taobao.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

Previously, ext4_extents.h was being included at the end of ext4.h,
which was bad for a number of reasons: (a) it was not being included
in the expected place, and (b) it caused the header to be included
multiple times.  There were #ifdef's to prevent this from causing any
problems, but it still was unnecessary.

By moving the function declarations that were in ext4_extents.h to
ext4.h, which is standard practice for where the function declarations
for the rest of ext4.h can be found, we can remove ext4_extents.h from
being included in ext4.h at all, and then we can only include
ext4_extents.h where it is needed in ext4's source files.

It should be possible to move a few more things into ext4.h, and
further reduce the number of source files that need to #include
ext4_extents.h, but that's a cleanup for another day.
Reported-by: NSachin Kamat <sachin.kamat@linaro.org>
Reported-by: NWei Yongjun <weiyj.lk@gmail.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

This commit is simple cleanup of fiemap codepath which has not been
included in previous commit to make the changes clearer. In this commit
we rename cbex variable to newex in ext4_fill_fiemap_extents() because
callback is no longer present
Signed-off-by: NLukas Czerner <lczerner@redhat.com>
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>