For append write workloads, extending the file requires a certain
amount of exclusive locking to be done up front to ensure sanity in
things like ensuring that we've zeroed any allocated regions
between the old EOF and the start of the new IO.

For single threads, this typically isn't a problem, and for large
IOs we don't serialise enough for it to be a problem for two
threads on really fast block devices. However for smaller IO and
larger thread counts we have a problem.

Take 4 concurrent sequential, single block sized and aligned IOs.
After the first IO is submitted but before it completes, we end up
with this state:

        IO 1    IO 2    IO 3    IO 4
      +-------+-------+-------+-------+
      ^       ^
      |       |
      |       |
      |       |
      |       \- ip->i_new_size
      \- ip->i_size

And the IO is done without exclusive locking because offset <=
ip->i_size. When we submit IO 2, we see offset > ip->i_size, and
grab the IO lock exclusive, because there is a chance we need to do
EOF zeroing. However, there is already an IO in progress that avoids
the need for IO zeroing because offset <= ip->i_new_size. hence we
could avoid holding the IO lock exlcusive for this. Hence after
submission of the second IO, we'd end up this state:

        IO 1    IO 2    IO 3    IO 4
      +-------+-------+-------+-------+
      ^               ^
      |               |
      |               |
      |               |
      |               \- ip->i_new_size
      \- ip->i_size

There is no need to grab the i_mutex of the IO lock in exclusive
mode if we don't need to invalidate the page cache. Taking these
locks on every direct IO effective serialises them as taking the IO
lock in exclusive mode has to wait for all shared holders to drop
the lock. That only happens when IO is complete, so effective it
prevents dispatch of concurrent direct IO writes to the same inode.

And so you can see that for the third concurrent IO, we'd avoid
exclusive locking for the same reason we avoided the exclusive lock
for the second IO.

Fixing this is a bit more complex than that, because we need to hold
a write-submission local value of ip->i_new_size to that clearing
the value is only done if no other thread has updated it before our
IO completes.....
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Alex Elder <aelder@sgi.com>