This is a very slow operation.  Right now POSIX_FADV_DONTNEED is the top
user because it has to freeze page references when removing it from the
cache.  invalidate_bdev() calls it for the same reason.  Both are
triggered from userspace, so it's easy to generate a storm.

mlock/mlockall no longer calls lru_add_drain_all - I've seen here
serious slowdown on older kernels.

There are some less obvious paths in memory migration/CMA/offlining which
shouldn't call frequently.

The worst case requires a non-trivial workload because lru_add_drain_all()
skips cpus where vectors are empty.  Something must constantly generate a
flow of pages for each cpu.  Also cpus must be busy to make scheduling
per-cpu works slower.  And the machine must be big enough (64+ cpus in our
case).

In our case that was a massive series of mlock calls in map-reduce while
other tasks write logs (and generates flows of new pages in per-cpu
vectors).  Mlock calls were serialized by mutex and accumulated latency up
to 10 seconds or more.

The kernel does not call lru_add_drain_all on mlock paths since 4.15, but
the same scenario could be triggered by fadvise(POSIX_FADV_DONTNEED) or
any other remaining user.

There is no reason to do the drain again if somebody else already drained
all the per-cpu vectors while we waited for the lock.

Piggyback on a drain starting and finishing while we wait for the lock:
all pages pending at the time of our entry were drained from the vectors.

Callers like POSIX_FADV_DONTNEED retry their operations once after
draining per-cpu vectors when pages have unexpected references.

Link: http://lkml.kernel.org/r/157019456205.3142.3369423180908482020.stgit@buzz


Signed-off-by: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Matthew Wilcox <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au>