Now, the scalability of swap code will drop much when the swap device
becomes fragmented, because the swap slots allocation batching stops
working.  To solve the problem, in this patch, we will try to scan a
little more swap slots with restricted effort to batch the swap slots
allocation even if the swap device is fragmented.  Test shows that the
benchmark score can increase up to 37.1% with the patch.  Details are as
follows.

The swap code has a per-cpu cache of swap slots.  These batch swap space
allocations to improve swap subsystem scaling.  In the following code
path,

  add_to_swap()
    get_swap_page()
      refill_swap_slots_cache()
        get_swap_pages()
	  scan_swap_map_slots()

scan_swap_map_slots() and get_swap_pages() can return multiple swap
slots for each call.  These slots will be cached in the per-CPU swap
slots cache, so that several following swap slot requests will be
fulfilled there to avoid the lock contention in the lower level swap
space allocation/freeing code path.

But this only works when there are free swap clusters.  If a swap device
becomes so fragmented that there's no free swap clusters,
scan_swap_map_slots() and get_swap_pages() will return only one swap
slot for each call in the above code path.  Effectively, this falls back
to the situation before the swap slots cache was introduced, the heavy
lock contention on the swap related locks kills the scalability.

Why does it work in this way? Because the swap device could be large,
and the free swap slot scanning could be quite time consuming, to avoid
taking too much time to scanning free swap slots, the conservative
method was used.

In fact, this can be improved via scanning a little more free slots with
strictly restricted effort.  Which is implemented in this patch.  In
scan_swap_map_slots(), after the first free swap slot is gotten, we will
try to scan a little more, but only if we haven't scanned too many slots
(< LATENCY_LIMIT).  That is, the added scanning latency is strictly
restricted.

To test the patch, we have run 16-process pmbench memory benchmark on a
2-socket server machine with 48 cores.  Multiple ram disks are
configured as the swap devices.  The pmbench working-set size is much
larger than the available memory so that swapping is triggered.  The
memory read/write ratio is 80/20 and the accessing pattern is random, so
the swap space becomes highly fragmented during the test.  In the
original implementation, the lock contention on swap related locks is
very heavy.  The perf profiling data of the lock contention code path is
as following,

 _raw_spin_lock.get_swap_pages.get_swap_page.add_to_swap:             21.03
 _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node:    1.92
 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node:      1.72
 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages:       0.69

While after applying this patch, it becomes,

 _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node:    4.89
 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node:      3.85
 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages:       1.1
 _raw_spin_lock_irqsave.pagevec_lru_move_fn.__lru_cache_add.do_swap_page: 0.88

That is, the lock contention on the swap locks is eliminated.

And the pmbench score increases 37.1%.  The swapin throughput increases
45.7% from 2.02 GB/s to 2.94 GB/s.  While the swapout throughput increases
45.3% from 2.04 GB/s to 2.97 GB/s.
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Acked-by: Tim Chen <tim.c.chen@linux.intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Link: http://lkml.kernel.org/r/20200427030023.264780-1-ying.huang@intel.comSigned-off-by: Linus Torvalds <torvalds@linux-foundation.org>