When doing a direct IO write against a file range that either has
preallocated extents in that range or has regular extents and the file
has the NOCOW attribute set, the write fails with -ENOSPC when all of
the following conditions are met:

1) There are no data blocks groups with enough free space matching
   the size of the write;

2) There's not enough unallocated space for allocating a new data block
   group;

3) The extents in the target file range are not shared, neither through
   snapshots nor through reflinks.

This is wrong because a NOCOW write can be done in such case, and in fact
it's possible to do it using a buffered IO write, since when failing to
allocate data space, the buffered IO path checks if a NOCOW write is
possible.

The failure in direct IO write path comes from the fact that early on,
at btrfs_dio_iomap_begin(), we try to allocate data space for the write
and if it that fails we return the error and stop - we never check if we
can do NOCOW. But later, at btrfs_get_blocks_direct_write(), we check
if we can do a NOCOW write into the range, or a subset of the range, and
then release the previously reserved data space.

Fix this by doing the data reservation only if needed, when we must COW,
at btrfs_get_blocks_direct_write() instead of doing it at
btrfs_dio_iomap_begin(). This also simplifies a bit the logic and removes
the inneficiency of doing unnecessary data reservations.

The following example test script reproduces the problem:

  $ cat dio-nocow-enospc.sh
  #!/bin/bash

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

  # Use a small fixed size (1G) filesystem so that it's quick to fill
  # it up.
  # Make sure the mixed block groups feature is not enabled because we
  # later want to not have more space available for allocating data
  # extents but still have enough metadata space free for the file writes.
  mkfs.btrfs -f -b $((1024 * 1024 * 1024)) -O ^mixed-bg $DEV
  mount $DEV $MNT

  # Create our test file with the NOCOW attribute set.
  touch $MNT/foobar
  chattr +C $MNT/foobar

  # Now fill in all unallocated space with data for our test file.
  # This will allocate a data block group that will be full and leave
  # no (or a very small amount of) unallocated space in the device, so
  # that it will not be possible to allocate a new block group later.
  echo
  echo "Creating test file with initial data..."
  xfs_io -c "pwrite -S 0xab -b 1M 0 900M" $MNT/foobar

  # Now try a direct IO write against file range [0, 10M[.
  # This should succeed since this is a NOCOW file and an extent for the
  # range was previously allocated.
  echo
  echo "Trying direct IO write over allocated space..."
  xfs_io -d -c "pwrite -S 0xcd -b 10M 0 10M" $MNT/foobar

  umount $MNT

When running the test:

  $ ./dio-nocow-enospc.sh
  (...)

  Creating test file with initial data...
  wrote 943718400/943718400 bytes at offset 0
  900 MiB, 900 ops; 0:00:01.43 (625.526 MiB/sec and 625.5265 ops/sec)

  Trying direct IO write over allocated space...
  pwrite: No space left on device

A test case for fstests will follow, testing both this direct IO write
scenario as well as the buffered IO write scenario to make it less likely
to get future regressions on the buffered IO case.
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>