Move the code that splits the ordered extents and records the physical
location for them to the storage layer so that the higher level consumers
don't have to care about physical block numbers at all.  This will also
allow to eventually remove accounting for the zone append write sizes in
the upper layer with a little bit more block layer work.
Reviewed-by: NNaohiro Aota <naohiro.aota@wdc.com>
Reviewed-by: NJosef Bacik <josef@toxicpanda.com>
Reviewed-by: NJohannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

Instead of letting the callers of btrfs_submit_bio deal with checksumming
the (meta)data in the bio and making decisions on when to offload the
checksumming to the bio, leave that to btrfs_submit_bio.  Do do so the
existing btrfs_submit_bio function is split into an upper and a lower
half, so that the lower half can be offloaded to a workqueue.

Note that this changes the behavior for direct writes to raid56 volumes so
that async checksum offloading is not skipped when more I/O is expected.
This runs counter to the argument explaining why it was done, although I
can't measure any affects of the change.  Commits later in this series
will make sure the entire direct writes is offloaded to the workqueue
at once and thus make sure it is sent to the raid56 code from a single
thread.
Reviewed-by: NJosef Bacik <josef@toxicpanda.com>
Reviewed-by: NJohannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

To prepare for further bio submission changes btrfs_csum_one_bio
should be able to take all it's arguments from the btrfs_bio structure.
It can always use the bbio->inode already, and once the compression code
is updated to set ->file_offset that one can be used unconditionally
as well instead of looking at the page mapping now that btrfs doesn't
allow ordered extents to span discontiguous data ranges.

The only slightly tricky bit is the one_ordered flag set by the
compressed writes.  Replace that one with the driver private bio
flag, which gets cleared before the bio is handed off to the block layer
so that we don't get in the way of driver use.

Note: this leaves an argument and a flag to btrfs_wq_submit_bio unused.
But that whole mechanism will be removed in its current form in the
next patch.
Reviewed-by: NJosef Bacik <josef@toxicpanda.com>
Reviewed-by: NJohannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

The submit helpers are now trivial and can be called directly.  Note
that btree_csum_one_bio has to be moved up in the file a bit to avoid a
forward declaration.
Reviewed-by: NJosef Bacik <josef@toxicpanda.com>
Reviewed-by: NJohannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

This flag is unused now, so remove it.  Re-expand the mirror_num field
to 8 bits, and move it to the I/O completion internal section of the
structure.
Reviewed-by: NJohannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

Rename iter to saved_iter and move it next to the repair internals
and nothing outside of bio.c should be touching it.
Reviewed-by: NJohannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

struct io_failure_record and the io_failure_tree tree are unused now,
so remove them. This in turn makes struct btrfs_inode smaller by 16
bytes.
Reviewed-by: NJohannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

The device field is only used by the simple end I/O handler, and for
that it can simply be stored in the bi_private field of the bio,
which is currently used for the fs_info that can be retrieved through
bbio->inode as well.
Reviewed-by: NJohannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

Remove the unused btrfs_verify_data_csum helper, and fold
btrfs_check_data_csum into its only caller.
Reviewed-by: NJohannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

btrfs_bio_for_each_sector is unused now, so remove it.
Reviewed-by: NJohannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

btrfs_bio_free_csum has only one caller left, and that caller is always
for an data inode and doesn't need zeroing of the csum pointer as that
pointer will never be touched again.  Just open code the conditional
kfree there.
Reviewed-by: NJohannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

Currently btrfs handles checksum validation and repair in the end I/O
handler for the btrfs_bio.  This leads to a lot of duplicate code
plus issues with varying semantics or bugs, e.g.

- the until recently broken repair for compressed extents
- the fact that encoded reads validate the checksums but do not kick
  of read repair
- the inconsistent checking of the BTRFS_FS_STATE_NO_CSUMS flag

This commit revamps the checksum validation and repair code to instead
work below the btrfs_submit_bio interfaces.

In case of a checksum failure (or a plain old I/O error), the repair
is now kicked off before the upper level ->end_io handler is invoked.

Progress of an in-progress repair is tracked by a small structure
that is allocated using a mempool for each original bio with failed
sectors, which holds a reference to the original bio.   This new
structure is allocated using a mempool to guarantee forward progress
even under memory pressure.  The mempool will be replenished when
the repair completes, just as the mempools backing the bios.

There is one significant behavior change here:  If repair fails or
is impossible to start with, the whole bio will be failed to the
upper layer.  This is the behavior that all I/O submitters except
for buffered I/O already emulated in their end_io handler.  For
buffered I/O this now means that a large readahead request can
fail due to a single bad sector, but as readahead errors are ignored
the following readpage if the sector is actually accessed will
still be able to read.  This also matches the I/O failure handling
in other file systems.
Reviewed-by: NJosef Bacik <josef@toxicpanda.com>
Reviewed-by: NJohannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

Add a new checksumming helper that wraps btrfs_check_data_csum and
does all the checks to if we're dealing with some form of nodatacsum
I/O.  This helper will be used by the new storage layer checksum
validation and repair code.
Reviewed-by: NJohannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

Instead of calling btrfs_lookup_bio_sums in every caller of
btrfs_submit_bio that reads data, do the call once in btrfs_submit_bio.
Reviewed-by: NJohannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

All callers of btrfs_submit_bio that want to validate checksums
currently have to store a copy of the iter in the btrfs_bio.  Move
the assignment into common code.
Reviewed-by: NJohannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

Add a bbio local variable and to prepare for calling functions that
return a blk_status_t, rename the existing int used for error handling
so that ret can be reused for the blk_status_t, and a label that can be
reused for failing the passed in bio.
Reviewed-by: NAnand Jain <anand.jain@oracle.com>
Reviewed-by: NJohannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

The csums argument is always NULL now, so remove it and always allocate
the csums array in the btrfs_bio.  Also pass the btrfs_bio instead of
inode + bio to document that this function requires a btrfs_bio and
not just any bio.
Reviewed-by: NJohannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: NAnand Jain <anand.jain@oracle.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

To prepare for pending changes drop the optimization to only look up
csums once per bio that is submitted from the iomap layer.  In the
short run this does cause additional lookups for fragmented direct
reads, but later in the series, the bio based lookup will be used on
the entire bio submitted from iomap, restoring the old behavior
in common code.
Reviewed-by: NAnand Jain <anand.jain@oracle.com>
Reviewed-by: NJohannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

All btrfs_bio I/Os are associated with an inode.  Add a pointer to that
inode, which will allow to simplify a lot of calling conventions, and
which will be needed in the I/O completion path in the future.

This grow the btrfs_bio structure by a pointer, but that grows will
be offset by the removal of the device pointer soon.
Reviewed-by: NAnand Jain <anand.jain@oracle.com>
Reviewed-by: NJohannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

Update the comments on btrfs_bio to better describe the structure.
Reviewed-by: NAnand Jain <anand.jain@oracle.com>
Reviewed-by: NJohannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

bio_split_rw can be used by file systems to split and incoming write
bio into multiple bios fitting the hardware limit for use as ZONE_APPEND
bios.  Export it for initial use in btrfs.
Reviewed-by: NJosef Bacik <josef@toxicpanda.com>
Reviewed-by: NChaitanya Kulkarni <kch@nvidia.com>
Reviewed-by: NJohannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

In rbio_update_error_bitmap(), we need to calculate the length of the
rbio.  As since it's called in the endio function, we can not directly
grab the length from bi_iter.

Currently we call bio_for_each_segment_all(), which will always return a
range inside a page.  But that's not necessary as we don't really care
about anything inside the page.

So use bio_for_each_bvec_all(), which can return a bvec across multiple
continuous pages thus reduce the loops.
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NQu Wenruo <wqu@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

There quite a few spelling mistakes as found using codespell. Fix them.
Signed-off-by: NColin Ian King <colin.i.king@gmail.com>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

During fiemap, when checking if a data extent is shared we are doing the
backref walking even if we already know the leaf is shared, which is a
waste of time since if the leaf shared then the data extent is also
shared. So skip the backref walking when we know we are in a shared leaf.

The following test was measures the gains for a case where all leaves
are shared due to a snapshot:

   $ cat test.sh
   #!/bin/bash

   DEV=/dev/sdj
   MNT=/mnt/sdj

   umount $DEV &> /dev/null
   mkfs.btrfs -f $DEV
   # Use compression to quickly create files with a lot of extents
   # (each with a size of 128K).
   mount -o compress=lzo $DEV $MNT

   # 40G gives 327680 extents, each with a size of 128K.
   xfs_io -f -c "pwrite -S 0xab -b 1M 0 40G" $MNT/foobar

   # Add some more files to increase the size of the fs and extent
   # trees (in the real world there's a lot of files and extents
   # from other files).
   xfs_io -f -c "pwrite -S 0xcd -b 1M 0 20G" $MNT/file1
   xfs_io -f -c "pwrite -S 0xef -b 1M 0 20G" $MNT/file2
   xfs_io -f -c "pwrite -S 0x73 -b 1M 0 20G" $MNT/file3

   # Create a snapshot so all the extents become indirectly shared
   # through subtrees, with a generation less than or equals to the
   # generation used to create the snapshot.
   btrfs subvolume snapshot -r $MNT $MNT/snap1

   # Unmount and mount again to clear cached metadata.
   umount $MNT
   mount -o compress=lzo $DEV $MNT

   start=$(date +%s%N)
   # The filefrag tool  uses the fiemap ioctl.
   filefrag $MNT/foobar
   end=$(date +%s%N)
   dur=$(( (end - start) / 1000000 ))
   echo "fiemap took $dur milliseconds (metadata not cached)"
   echo

   start=$(date +%s%N)
   filefrag $MNT/foobar
   end=$(date +%s%N)
   dur=$(( (end - start) / 1000000 ))
   echo "fiemap took $dur milliseconds (metadata cached)"

   umount $MNT

The results were the following on a non-debug kernel (Debian's default
kernel config).

Before this patch:

   (...)
   /mnt/sdi/foobar: 327680 extents found
   fiemap took 1821 milliseconds (metadata not cached)

   /mnt/sdi/foobar: 327680 extents found
   fiemap took 399 milliseconds (metadata cached)

After this patch:

   (...)
   /mnt/sdi/foobar: 327680 extents found
   fiemap took 591 milliseconds (metadata not cached)

   /mnt/sdi/foobar: 327680 extents found
   fiemap took 123 milliseconds (metadata cached)

That's a speedup of 3.1x and 3.2x.
Signed-off-by: NFilipe Manana <fdmanana@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

During fiemap, when accessing the cache that stores the sharedness of an
extent, we need to either be holding a transaction handle or the commit
root semaphore. I left comments about this in the comment that precedes
store_backref_shared_cache() and lookup_backref_shared_cache(), but have
actually not enforced it through assertions. So assert that the commit
root semaphore is held if we are not holding a transaction handle.
Reviewed-by: NJohannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: NFilipe Manana <fdmanana@suse.com>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

Async discard does not acquire the block group reference count while it
holds a reference on the discard list. This is generally OK, as the
paths which destroy block groups tend to try to synchronize on
cancelling async discard work. However, relying on cancelling work
requires careful analysis to be sure it is safe from races with
unpinning scheduling more work.

While I am unable to find a race with unpinning in the current code for
either the unused bgs or relocation paths, I believe we have one in an
older version of auto relocation in a Meta internal build. This suggests
that this is in fact an error prone model, and could be fragile to
future changes to these bg deletion paths.

To make this ownership more clear, add a refcount for async discard. If
work is queued for a block group, its refcount should be incremented,
and when work is completed or canceled, it should be decremented.

CC: stable@vger.kernel.org # 5.15+
Signed-off-by: NBoris Burkov <boris@bur.io>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

Whenever we add or remove an entry to a directory, we issue an utimes
command for the directory. If we add 1000 entries to a directory (create
1000 files under it or move 1000 files to it), then we issue the same
utimes command 1000 times, which increases the send stream size, results
in more pipe IO, one search in the send b+tree, allocating one path for
the search, etc, as well as making the receiver do a system call for each
duplicated utimes command.

We also issue an utimes command when we create a new directory, but later
we might add entries to it corresponding to inodes with an higher inode
number, so it's pointless to issue the utimes command before we create
the last inode under the directory.

So use a lru cache to track directories for which we must send a utimes
command. When we need to remove an entry from the cache, we issue the
utimes command for the respective directory. When finishing the send
operation, we go over each cache element and issue the respective utimes
command. Finally the caching is entirely optional, just a performance
optimization, meaning that if we fail to cache (due to memory allocation
failure), we issue the utimes command right away, that is, we fallback
to the previous, unoptimized, behaviour.

This patch belongs to a patchset comprised of the following patches:

  btrfs: send: directly return from did_overwrite_ref() and simplify it
  btrfs: send: avoid unnecessary generation search at did_overwrite_ref()
  btrfs: send: directly return from will_overwrite_ref() and simplify it
  btrfs: send: avoid extra b+tree searches when checking reference overrides
  btrfs: send: remove send_progress argument from can_rmdir()
  btrfs: send: avoid duplicated orphan dir allocation and initialization
  btrfs: send: avoid unnecessary orphan dir rbtree search at can_rmdir()
  btrfs: send: reduce searches on parent root when checking if dir can be removed
  btrfs: send: iterate waiting dir move rbtree only once when processing refs
  btrfs: send: initialize all the red black trees earlier
  btrfs: send: genericize the backref cache to allow it to be reused
  btrfs: adapt lru cache to allow for 64 bits keys on 32 bits systems
  btrfs: send: cache information about created directories
  btrfs: allow a generation number to be associated with lru cache entries
  btrfs: add an api to delete a specific entry from the lru cache
  btrfs: send: use the lru cache to implement the name cache
  btrfs: send: update size of roots array for backref cache entries
  btrfs: send: cache utimes operations for directories if possible

The following test was run before and after applying the whole patchset,
and on a non-debug kernel (Debian's default kernel config):

   #!/bin/bash

   MNT=/mnt/sdi
   DEV=/dev/sdi

   mkfs.btrfs -f $DEV > /dev/null
   mount $DEV $MNT

   mkdir $MNT/A
   for ((i = 1; i <= 20000; i++)); do
       echo -n > $MNT/A/file_$i
   done

   btrfs subvolume snapshot -r $MNT $MNT/snap1

   mkdir $MNT/B
   for ((i = 20000; i <= 40000; i++)); do
       echo -n > $MNT/B/file_$i
   done

   mv $MNT/A/file_* $MNT/B/

   btrfs subvolume snapshot -r $MNT $MNT/snap2

   start=$(date +%s%N)
   btrfs send -p $MNT/snap1 $MNT/snap2 > /dev/null
   end=$(date +%s%N)

   dur=$(( (end - start) / 1000000 ))
   echo "Incremental send took $dur milliseconds"

   umount $MNT

Before the whole patchset: 18408 milliseconds
After the whole patchset:   1942 milliseconds  (9.5x speedup)

Using 60000 files instead of 40000:

Before the whole patchset: 39764 milliseconds
After the whole patchset:   3076 milliseconds  (12.9x speedup)

Using 20000 files instead of 40000:

Before the whole patchset:  5072 milliseconds
After the whole patchset:    916 milliseconds  (5.5x speedup)
Signed-off-by: NFilipe Manana <fdmanana@suse.com>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

Currently we limit the size of the roots array, for backref cache entries,
to 12 elements. This is because that number is enough for most cases and
to make the backref cache entry size to be exactly 128 bytes, so that
memory is allocated from the kmalloc-128 slab and no space is wasted.

However recent changes in the series refactored the backref cache to be
more generic and allow it to be reused for other purposes, which resulted
in increasing the size of the embedded structure btrfs_lru_cache_entry in
order to allow for supporting inode numbers as keys on 32 bits system and
allow multiple generations per key. This resulted in increasing the size
of struct backref_cache_entry from 128 bytes to 152 bytes. Since the cache
entries are allocated with kmalloc(), it means we end up using the slab
kmalloc-192, so we end up wasting 40 bytes of memory. So bump the size of
the roots array from 12 elements to 17 elements, so we end up using 192
bytes for each backref cache entry.

This patch is part of a larger patchset and the changelog of the last
patch in the series contains a sample performance test and results.
The patches that comprise the patchset are the following:

  btrfs: send: directly return from did_overwrite_ref() and simplify it
  btrfs: send: avoid unnecessary generation search at did_overwrite_ref()
  btrfs: send: directly return from will_overwrite_ref() and simplify it
  btrfs: send: avoid extra b+tree searches when checking reference overrides
  btrfs: send: remove send_progress argument from can_rmdir()
  btrfs: send: avoid duplicated orphan dir allocation and initialization
  btrfs: send: avoid unnecessary orphan dir rbtree search at can_rmdir()
  btrfs: send: reduce searches on parent root when checking if dir can be removed
  btrfs: send: iterate waiting dir move rbtree only once when processing refs
  btrfs: send: initialize all the red black trees earlier
  btrfs: send: genericize the backref cache to allow it to be reused
  btrfs: adapt lru cache to allow for 64 bits keys on 32 bits systems
  btrfs: send: cache information about created directories
  btrfs: allow a generation number to be associated with lru cache entries
  btrfs: add an api to delete a specific entry from the lru cache
  btrfs: send: use the lru cache to implement the name cache
  btrfs: send: update size of roots array for backref cache entries
  btrfs: send: cache utimes operations for directories if possible
Signed-off-by: NFilipe Manana <fdmanana@suse.com>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

The name cache in send is basically a lru cache implemented with a radix
tree and linked lists, very similar to the lru cache module which is used
for the send backref cache and the cache of previously created directories
during a send operation. So remove all the custom caching code for the
name cache and make it use the lru cache instead.

One particular detail to note is that the current cache behaves a bit
differently when it comes to eviction of entries. Namely when after
inserting a new name in the cache, if the cache now has 256 entries, we
evict the last 128 LRU entries. The lru_cache.{c,h} module behaves a bit
differently in that once we reach the cache limit, we evict a single LRU
entry. In practice this doesn't make much difference, but it's actually
better to evict just one entry instead of half of the entries, as there's
always a chance we will need a name stored in one of that last 128 removed
entries.

This patch is part of a larger patchset and the changelog of the last
patch in the series contains a sample performance test and results.
The patches that comprise the patchset are the following:

  btrfs: send: directly return from did_overwrite_ref() and simplify it
  btrfs: send: avoid unnecessary generation search at did_overwrite_ref()
  btrfs: send: directly return from will_overwrite_ref() and simplify it
  btrfs: send: avoid extra b+tree searches when checking reference overrides
  btrfs: send: remove send_progress argument from can_rmdir()
  btrfs: send: avoid duplicated orphan dir allocation and initialization
  btrfs: send: avoid unnecessary orphan dir rbtree search at can_rmdir()
  btrfs: send: reduce searches on parent root when checking if dir can be removed
  btrfs: send: iterate waiting dir move rbtree only once when processing refs
  btrfs: send: initialize all the red black trees earlier
  btrfs: send: genericize the backref cache to allow it to be reused
  btrfs: adapt lru cache to allow for 64 bits keys on 32 bits systems
  btrfs: send: cache information about created directories
  btrfs: allow a generation number to be associated with lru cache entries
  btrfs: add an api to delete a specific entry from the lru cache
  btrfs: send: use the lru cache to implement the name cache
  btrfs: send: update size of roots array for backref cache entries
  btrfs: send: cache utimes operations for directories if possible
Signed-off-by: NFilipe Manana <fdmanana@suse.com>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

In order to replace the open coded name cache in send with the lru cache,
we need an API for the lru cache to delete a specific entry for which we
did a previous lookup. This adds the API for it, and a next patch in the
series will use it.

This patch is part of a larger patchset and the changelog of the last
patch in the series contains a sample performance test and results.
The patches that comprise the patchset are the following:

  btrfs: send: directly return from did_overwrite_ref() and simplify it
  btrfs: send: avoid unnecessary generation search at did_overwrite_ref()
  btrfs: send: directly return from will_overwrite_ref() and simplify it
  btrfs: send: avoid extra b+tree searches when checking reference overrides
  btrfs: send: remove send_progress argument from can_rmdir()
  btrfs: send: avoid duplicated orphan dir allocation and initialization
  btrfs: send: avoid unnecessary orphan dir rbtree search at can_rmdir()
  btrfs: send: reduce searches on parent root when checking if dir can be removed
  btrfs: send: iterate waiting dir move rbtree only once when processing refs
  btrfs: send: initialize all the red black trees earlier
  btrfs: send: genericize the backref cache to allow it to be reused
  btrfs: adapt lru cache to allow for 64 bits keys on 32 bits systems
  btrfs: send: cache information about created directories
  btrfs: allow a generation number to be associated with lru cache entries
  btrfs: add an api to delete a specific entry from the lru cache
  btrfs: send: use the lru cache to implement the name cache
  btrfs: send: update size of roots array for backref cache entries
  btrfs: send: cache utimes operations for directories if possible
Signed-off-by: NFilipe Manana <fdmanana@suse.com>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

This allows an optional generation number to be associated to each entry
of the lru cache. Entries with the same key but different generations, are
stored in the linked list to which the maple tree points to. This is meant
to be used when there's a small number of different generations, so the
impact of searching a linked list is negligible. The goal is to get rid of
the open coded name cache in the send code (which uses a radix tree and
a similar linked list of values/entries) and use instead the lru cache
module. For that particular use case we have at most 2 generations that
are associated to each key (inode number): one generation for the send
root and another generation for the parent root. The actual migration of
the send name cache is done in the next patch in the series.

This patch is part of a larger patchset and the changelog of the last
patch in the series contains a sample performance test and results.
The patches that comprise the patchset are the following:

  btrfs: send: directly return from did_overwrite_ref() and simplify it
  btrfs: send: avoid unnecessary generation search at did_overwrite_ref()
  btrfs: send: directly return from will_overwrite_ref() and simplify it
  btrfs: send: avoid extra b+tree searches when checking reference overrides
  btrfs: send: remove send_progress argument from can_rmdir()
  btrfs: send: avoid duplicated orphan dir allocation and initialization
  btrfs: send: avoid unnecessary orphan dir rbtree search at can_rmdir()
  btrfs: send: reduce searches on parent root when checking if dir can be removed
  btrfs: send: iterate waiting dir move rbtree only once when processing refs
  btrfs: send: initialize all the red black trees earlier
  btrfs: send: genericize the backref cache to allow it to be reused
  btrfs: adapt lru cache to allow for 64 bits keys on 32 bits systems
  btrfs: send: cache information about created directories
  btrfs: allow a generation number to be associated with lru cache entries
  btrfs: add an api to delete a specific entry from the lru cache
  btrfs: send: use the lru cache to implement the name cache
  btrfs: send: update size of roots array for backref cache entries
  btrfs: send: cache utimes operations for directories if possible
Signed-off-by: NFilipe Manana <fdmanana@suse.com>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

During an incremental send, when processing the reference for an inode
we need to check if the directory where the new reference is located was
already created before creating the new reference. This check, which is
done by the helper did_create_dir(), can be expensive if the directory
has many entries, since it consists in searching the send root's b+tree
and visiting every single dir index key until we either find one which
points to an inode with a number smaller than the current inode's number
or until we visited all index keys. So it doesn't scale well for very
large directories.

So improve on this by caching created directories using a lru cache, and
limiting its size to 64 entries, which results in using at most 4096
bytes of memory. The caching is optional, if we fail to allocate memory,
we just proceed as before and use the existing slower path.

This patch is part of a larger patchset and the changelog of the last
patch in the series contains a sample performance test and results.
The patches that comprise the patchset are the following:

  btrfs: send: directly return from did_overwrite_ref() and simplify it
  btrfs: send: avoid unnecessary generation search at did_overwrite_ref()
  btrfs: send: directly return from will_overwrite_ref() and simplify it
  btrfs: send: avoid extra b+tree searches when checking reference overrides
  btrfs: send: remove send_progress argument from can_rmdir()
  btrfs: send: avoid duplicated orphan dir allocation and initialization
  btrfs: send: avoid unnecessary orphan dir rbtree search at can_rmdir()
  btrfs: send: reduce searches on parent root when checking if dir can be removed
  btrfs: send: iterate waiting dir move rbtree only once when processing refs
  btrfs: send: initialize all the red black trees earlier
  btrfs: send: genericize the backref cache to allow it to be reused
  btrfs: adapt lru cache to allow for 64 bits keys on 32 bits systems
  btrfs: send: cache information about created directories
  btrfs: allow a generation number to be associated with lru cache entries
  btrfs: add an api to delete a specific entry from the lru cache
  btrfs: send: use the lru cache to implement the name cache
  btrfs: send: update size of roots array for backref cache entries
  btrfs: send: cache utimes operations for directories if possible
Signed-off-by: NFilipe Manana <fdmanana@suse.com>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

The lru cache is backed by a maple tree, which uses the unsigned long
type for keys, and that type has a width of 32 bits on 32 bits systems
and a width of 64 bits on 64 bits systems.

Currently there is only one user of the lru cache, the send backref cache,
which uses a sector number as a key, a logical address right shifted by
fs_info->sectorsize_bits, so a 32 bits width is not yet a problem (the
same happens with the radix tree we use to track extent buffers,
fs_info->buffer_radix).

However the next patches in the series will start using the lru cache for
cases where inode numbers are the keys, and the inode numbers are always
64 bits, even if we are running on a 32 bits system.

So adapt the lru cache to allow multiple values under the same key, by
having the maple tree store a head entry that points to a list of entries
instead of pointing to a single entry. This is a similar approach to what
we currently do for the name cache in send (which uses a radix tree that
has indexes with an unsigned long type as well), and will allow later to
use the lru cache for the send name cache as well.

This patch is part of a larger patchset and the changelog of the last
patch in the series contains a sample performance test and results.
The patches that comprise the patchset are the following:

  btrfs: send: directly return from did_overwrite_ref() and simplify it
  btrfs: send: avoid unnecessary generation search at did_overwrite_ref()
  btrfs: send: directly return from will_overwrite_ref() and simplify it
  btrfs: send: avoid extra b+tree searches when checking reference overrides
  btrfs: send: remove send_progress argument from can_rmdir()
  btrfs: send: avoid duplicated orphan dir allocation and initialization
  btrfs: send: avoid unnecessary orphan dir rbtree search at can_rmdir()
  btrfs: send: reduce searches on parent root when checking if dir can be removed
  btrfs: send: iterate waiting dir move rbtree only once when processing refs
  btrfs: send: initialize all the red black trees earlier
  btrfs: send: genericize the backref cache to allow it to be reused
  btrfs: adapt lru cache to allow for 64 bits keys on 32 bits systems
  btrfs: send: cache information about created directories
  btrfs: allow a generation number to be associated with lru cache entries
  btrfs: add an api to delete a specific entry from the lru cache
  btrfs: send: use the lru cache to implement the name cache
  btrfs: send: update size of roots array for backref cache entries
  btrfs: send: cache utimes operations for directories if possible
Signed-off-by: NFilipe Manana <fdmanana@suse.com>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

The backref cache is a cache backed by a maple tree and a linked list to
keep track of temporal access to cached entries (the LRU entry always at
the head of the list). This type of caching method is going to be useful
in other scenarios, so make the cache implementation more generic and
move it into its own header and source files.

This patch is part of a larger patchset and the changelog of the last
patch in the series contains a sample performance test and results.
The patches that comprise the patchset are the following:

  btrfs: send: directly return from did_overwrite_ref() and simplify it
  btrfs: send: avoid unnecessary generation search at did_overwrite_ref()
  btrfs: send: directly return from will_overwrite_ref() and simplify it
  btrfs: send: avoid extra b+tree searches when checking reference overrides
  btrfs: send: remove send_progress argument from can_rmdir()
  btrfs: send: avoid duplicated orphan dir allocation and initialization
  btrfs: send: avoid unnecessary orphan dir rbtree search at can_rmdir()
  btrfs: send: reduce searches on parent root when checking if dir can be removed
  btrfs: send: iterate waiting dir move rbtree only once when processing refs
  btrfs: send: initialize all the red black trees earlier
  btrfs: send: genericize the backref cache to allow it to be reused
  btrfs: adapt lru cache to allow for 64 bits keys on 32 bits systems
  btrfs: send: cache information about created directories
  btrfs: allow a generation number to be associated with lru cache entries
  btrfs: add an api to delete a specific entry from the lru cache
  btrfs: send: use the lru cache to implement the name cache
  btrfs: send: update size of roots array for backref cache entries
  btrfs: send: cache utimes operations for directories if possible
Signed-off-by: NFilipe Manana <fdmanana@suse.com>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

After we allocate the send context object and before we initialize all
the red black trees, we can jump to the 'out' label if some errors happen,
and then under the 'out' label we use RB_EMPTY_ROOT() against some of the
those trees, which we have not yet initialized. This happens to work out
ok because the send context object was initialized to zeroes with kzalloc
and the RB_ROOT initializer just happens to have the following definition:

    #define RB_ROOT (struct rb_root) { NULL, }

But it's really neither clean nor a good practice as RB_ROOT is supposed
to be opaque and in case it changes or we change those red black trees to
some other data structure, it leaves us in a precarious situation.

So initialize all the red black trees immediately after allocating the
send context and before any jump into the 'out' label.

This patch is part of a larger patchset and the changelog of the last
patch in the series contains a sample performance test and results.
The patches that comprise the patchset are the following:

  btrfs: send: directly return from did_overwrite_ref() and simplify it
  btrfs: send: avoid unnecessary generation search at did_overwrite_ref()
  btrfs: send: directly return from will_overwrite_ref() and simplify it
  btrfs: send: avoid extra b+tree searches when checking reference overrides
  btrfs: send: remove send_progress argument from can_rmdir()
  btrfs: send: avoid duplicated orphan dir allocation and initialization
  btrfs: send: avoid unnecessary orphan dir rbtree search at can_rmdir()
  btrfs: send: reduce searches on parent root when checking if dir can be removed
  btrfs: send: iterate waiting dir move rbtree only once when processing refs
  btrfs: send: initialize all the red black trees earlier
  btrfs: send: genericize the backref cache to allow it to be reused
  btrfs: adapt lru cache to allow for 64 bits keys on 32 bits systems
  btrfs: send: cache information about created directories
  btrfs: allow a generation number to be associated with lru cache entries
  btrfs: add an api to delete a specific entry from the lru cache
  btrfs: send: use the lru cache to implement the name cache
  btrfs: send: update size of roots array for backref cache entries
  btrfs: send: cache utimes operations for directories if possible
Signed-off-by: NFilipe Manana <fdmanana@suse.com>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

When processing the new references for an inode, we unnecessarily iterate
twice the waiting dir moves rbtree, once with is_waiting_for_move() and
if we found an entry in the rbtree, we iterate it again with a call to
get_waiting_dir_move(). This is pointless, we can make this simpler and
more efficient by calling only get_waiting_dir_move(), so just do that.

This patch is part of a larger patchset and the changelog of the last
patch in the series contains a sample performance test and results.
The patches that comprise the patchset are the following:

  btrfs: send: directly return from did_overwrite_ref() and simplify it
  btrfs: send: avoid unnecessary generation search at did_overwrite_ref()
  btrfs: send: directly return from will_overwrite_ref() and simplify it
  btrfs: send: avoid extra b+tree searches when checking reference overrides
  btrfs: send: remove send_progress argument from can_rmdir()
  btrfs: send: avoid duplicated orphan dir allocation and initialization
  btrfs: send: avoid unnecessary orphan dir rbtree search at can_rmdir()
  btrfs: send: reduce searches on parent root when checking if dir can be removed
  btrfs: send: iterate waiting dir move rbtree only once when processing refs
  btrfs: send: initialize all the red black trees earlier
  btrfs: send: genericize the backref cache to allow it to be reused
  btrfs: adapt lru cache to allow for 64 bits keys on 32 bits systems
  btrfs: send: cache information about created directories
  btrfs: allow a generation number to be associated with lru cache entries
  btrfs: add an api to delete a specific entry from the lru cache
  btrfs: send: use the lru cache to implement the name cache
  btrfs: send: update size of roots array for backref cache entries
  btrfs: send: cache utimes operations for directories if possible
Signed-off-by: NFilipe Manana <fdmanana@suse.com>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

During an incremental send, every time we remove a reference (dentry) for
an inode and the parent directory does not exists anymore in the send
root, we go check if we can remove the directory by making a call to
can_rmdir(). This helper can only return true (value 1) if all dentries
were already removed, and for that it always does a search on the parent
root for dir index keys - if it finds any dentry referring to an inode
with a number higher then the inode currently being processed, then the
directory can not be removed and it must return false (value 0).

However that means if a directory that was deleted had 1000 dentries, and
each one pointed to an inode with a number higher then the number of the
directory's inode, we end up doing 1000 searches on the parent root.
Typically files are created in a directory after the directory was created
and therefore they get an higher inode number than the directory. It's
also common to have the each dentry pointing to an inode with a higher
number then the inodes the previous dentries point to, for example when
creating a series of files inside a directory, a very common pattern.

So improve on that by having the first call to can_rmdir() for a directory
to check the number of the inode that the last dentry points to and cache
that inode number in the orphan dir structure. Then every subsequent call
to can_rmdir() can avoid doing a search on the parent root if the number
of the inode currently being processed is smaller than cached inode number
at the directory's orphan dir structure.

This patch is part of a larger patchset and the changelog of the last
patch in the series contains a sample performance test and results.
The patches that comprise the patchset are the following:

  btrfs: send: directly return from did_overwrite_ref() and simplify it
  btrfs: send: avoid unnecessary generation search at did_overwrite_ref()
  btrfs: send: directly return from will_overwrite_ref() and simplify it
  btrfs: send: avoid extra b+tree searches when checking reference overrides
  btrfs: send: remove send_progress argument from can_rmdir()
  btrfs: send: avoid duplicated orphan dir allocation and initialization
  btrfs: send: avoid unnecessary orphan dir rbtree search at can_rmdir()
  btrfs: send: reduce searches on parent root when checking if dir can be removed
  btrfs: send: iterate waiting dir move rbtree only once when processing refs
  btrfs: send: initialize all the red black trees earlier
  btrfs: send: genericize the backref cache to allow it to be reused
  btrfs: adapt lru cache to allow for 64 bits keys on 32 bits systems
  btrfs: send: cache information about created directories
  btrfs: allow a generation number to be associated with lru cache entries
  btrfs: add an api to delete a specific entry from the lru cache
  btrfs: send: use the lru cache to implement the name cache
  btrfs: send: update size of roots array for backref cache entries
  btrfs: send: cache utimes operations for directories if possible
Signed-off-by: NFilipe Manana <fdmanana@suse.com>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

At can_rmdir() we start by searching the orphan dirs rbtree for an orphan
dir object for the target directory. Later when iterating over the dir
index keys, if we find that any dir entry points to inode for which there
is a pending dir move or the inode was not yet processed, we exit because
we can't remove the directory yet. However we end up always calling
add_orphan_dir_info(), which will iterate again the rbtree and if there is
already an orphan dir object (created by the first call to can_rmdir()),
it returns the existing object. This is unnecessary work because in case
there is already an existing orphan dir object, we got a reference to it
at the start of can_rmdir(). So skip the call to add_orphan_dir_info()
if we already have a reference for an orphan dir object.

This patch is part of a larger patchset and the changelog of the last
patch in the series contains a sample performance test and results.
The patches that comprise the patchset are the following:

  btrfs: send: directly return from did_overwrite_ref() and simplify it
  btrfs: send: avoid unnecessary generation search at did_overwrite_ref()
  btrfs: send: directly return from will_overwrite_ref() and simplify it
  btrfs: send: avoid extra b+tree searches when checking reference overrides
  btrfs: send: remove send_progress argument from can_rmdir()
  btrfs: send: avoid duplicated orphan dir allocation and initialization
  btrfs: send: avoid unnecessary orphan dir rbtree search at can_rmdir()
  btrfs: send: reduce searches on parent root when checking if dir can be removed
  btrfs: send: iterate waiting dir move rbtree only once when processing refs
  btrfs: send: initialize all the red black trees earlier
  btrfs: send: genericize the backref cache to allow it to be reused
  btrfs: adapt lru cache to allow for 64 bits keys on 32 bits systems
  btrfs: send: cache information about created directories
  btrfs: allow a generation number to be associated with lru cache entries
  btrfs: add an api to delete a specific entry from the lru cache
  btrfs: send: use the lru cache to implement the name cache
  btrfs: send: update size of roots array for backref cache entries
  btrfs: send: cache utimes operations for directories if possible
Signed-off-by: NFilipe Manana <fdmanana@suse.com>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

At can_rmdir() we are allocating and initializing an orphan dir object
twice. This can be deduplicated outside of the loop that iterates over
the dir index keys. So deduplicate that code, even because other patch
in the series will need to add more initialization code and another one
will add one more condition.

This patch is part of a larger patchset and the changelog of the last
patch in the series contains a sample performance test and results.
The patches that comprise the patchset are the following:

  btrfs: send: directly return from did_overwrite_ref() and simplify it
  btrfs: send: avoid unnecessary generation search at did_overwrite_ref()
  btrfs: send: directly return from will_overwrite_ref() and simplify it
  btrfs: send: avoid extra b+tree searches when checking reference overrides
  btrfs: send: remove send_progress argument from can_rmdir()
  btrfs: send: avoid duplicated orphan dir allocation and initialization
  btrfs: send: avoid unnecessary orphan dir rbtree search at can_rmdir()
  btrfs: send: reduce searches on parent root when checking if dir can be removed
  btrfs: send: iterate waiting dir move rbtree only once when processing refs
  btrfs: send: initialize all the red black trees earlier
  btrfs: send: genericize the backref cache to allow it to be reused
  btrfs: adapt lru cache to allow for 64 bits keys on 32 bits systems
  btrfs: send: cache information about created directories
  btrfs: allow a generation number to be associated with lru cache entries
  btrfs: add an api to delete a specific entry from the lru cache
  btrfs: send: use the lru cache to implement the name cache
  btrfs: send: update size of roots array for backref cache entries
  btrfs: send: cache utimes operations for directories if possible
Signed-off-by: NFilipe Manana <fdmanana@suse.com>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

All callers of can_rmdir() pass sctx->cur_ino as the value for the
send_progress argument, so remove the argument and directly use
sctx->cur_ino.

This patch is part of a larger patchset and the changelog of the last
patch in the series contains a sample performance test and results.
The patches that comprise the patchset are the following:

  btrfs: send: directly return from did_overwrite_ref() and simplify it
  btrfs: send: avoid unnecessary generation search at did_overwrite_ref()
  btrfs: send: directly return from will_overwrite_ref() and simplify it
  btrfs: send: avoid extra b+tree searches when checking reference overrides
  btrfs: send: remove send_progress argument from can_rmdir()
  btrfs: send: avoid duplicated orphan dir allocation and initialization
  btrfs: send: avoid unnecessary orphan dir rbtree search at can_rmdir()
  btrfs: send: reduce searches on parent root when checking if dir can be removed
  btrfs: send: iterate waiting dir move rbtree only once when processing refs
  btrfs: send: initialize all the red black trees earlier
  btrfs: send: genericize the backref cache to allow it to be reused
  btrfs: adapt lru cache to allow for 64 bits keys on 32 bits systems
  btrfs: send: cache information about created directories
  btrfs: allow a generation number to be associated with lru cache entries
  btrfs: add an api to delete a specific entry from the lru cache
  btrfs: send: use the lru cache to implement the name cache
  btrfs: send: update size of roots array for backref cache entries
  btrfs: send: cache utimes operations for directories if possible
Signed-off-by: NFilipe Manana <fdmanana@suse.com>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>