fsnotify: Rearrange fast path to minimise overhead when there is no watcher

fsnotify: Rearrange fast path to minimise overhead when there is no watcher

The fsnotify paths are trivial to hit even when there are no watchers and
they are surprisingly expensive. For example, every successful vfs_write()
hits fsnotify_modify which calls both fsnotify_parent and fsnotify unless
FMODE_NONOTIFY is set which is an internal flag invisible to userspace.
As it stands, fsnotify_parent is a guaranteed functional call even if there
are no watchers and fsnotify() does a substantial amount of unnecessary
work before it checks if there are any watchers. A perf profile showed
that applying mnt->mnt_fsnotify_mask in fnotify() was almost half of the
total samples taken in that function during a test. This patch rearranges
the fast paths to reduce the amount of work done when there are no
watchers.

The test motivating this was "perf bench sched messaging --pipe". Despite
the fact the pipes are anonymous, fsnotify is still called a lot and
the overhead is noticeable even though it's completely pointless. It's
likely the overhead is negligible for real IO so this is an extreme
example. This is a comparison of hackbench using processes and pipes on
a 1-socket machine with 8 CPU threads without fanotify watchers.

                              5.7.0                  5.7.0
                            vanilla      fastfsnotify-v1r1
Amean     1       0.4837 (   0.00%)      0.4630 *   4.27%*
Amean     3       1.5447 (   0.00%)      1.4557 (   5.76%)
Amean     5       2.6037 (   0.00%)      2.4363 (   6.43%)
Amean     7       3.5987 (   0.00%)      3.4757 (   3.42%)
Amean     12      5.8267 (   0.00%)      5.6983 (   2.20%)
Amean     18      8.4400 (   0.00%)      8.1327 (   3.64%)
Amean     24     11.0187 (   0.00%)     10.0290 *   8.98%*
Amean     30     13.1013 (   0.00%)     12.8510 (   1.91%)
Amean     32     13.9190 (   0.00%)     13.2410 (   4.87%)

                       5.7.0       5.7.0
                     vanilla fastfsnotify-v1r1
Duration User         157.05      152.79
Duration System      1279.98     1219.32
Duration Elapsed      182.81      174.52

This is showing that the latencies are improved by roughly 2-9%. The
variability is not shown but some of these results are within the noise
as this workload heavily overloads the machine. That said, the system CPU
usage is reduced by quite a bit so it makes sense to avoid the overhead
even if it is a bit tricky to detect at times. A perf profile of just 1
group of tasks showed that 5.14% of samples taken were in either fsnotify()
or fsnotify_parent(). With the patch, 2.8% of samples were in fsnotify,
mostly function entry and the initial check for watchers.  The check for
watchers is complicated enough that inlining it may be controversial.

[Amir] Slightly simplify with mnt_or_sb_mask => marks_mask

Link: https://lore.kernel.org/r/20200708111156.24659-1-amir73il@gmail.com
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Amir Goldstein <amir73il@gmail.com>
Signed-off-by: Jan Kara <jack@suse.cz>