mm, swap: use rbtree for swap_extent

swap_extent is used to map swap page offset to backing device's block
offset.  For a continuous block range, one swap_extent is used and all
these swap_extents are managed in a linked list.

These swap_extents are used by map_swap_entry() during swap's read and
write path.  To find out the backing device's block offset for a page
offset, the swap_extent list will be traversed linearly, with
curr_swap_extent being used as a cache to speed up the search.

This works well as long as swap_extents are not huge or when the number
of processes that access swap device are few, but when the swap device
has many extents and there are a number of processes accessing the swap
device concurrently, it can be a problem.  On one of our servers, the
disk's remaining size is tight:

  $df -h
  Filesystem      Size  Used Avail Use% Mounted on
  ... ...
  /dev/nvme0n1p1  1.8T  1.3T  504G  72% /home/t4

When creating a 80G swapfile there, there are as many as 84656 swap
extents.  The end result is, kernel spends abou 30% time in
map_swap_entry() and swap throughput is only 70MB/s.

As a comparison, when I used smaller sized swapfile, like 4G whose
swap_extent dropped to 2000, swap throughput is back to 400-500MB/s and
map_swap_entry() is about 3%.

One downside of using rbtree for swap_extent is, 'struct rbtree' takes
24 bytes while 'struct list_head' takes 16 bytes, that's 8 bytes more
for each swap_extent.  For a swapfile that has 80k swap_extents, that
means 625KiB more memory consumed.

Test:

Since it's not possible to reboot that server, I can not test this patch
diretly there.  Instead, I tested it on another server with NVMe disk.

I created a 20G swapfile on an NVMe backed XFS fs.  By default, the
filesystem is quite clean and the created swapfile has only 2 extents.
Testing vanilla and this patch shows no obvious performance difference
when swapfile is not fragmented.

To see the patch's effects, I used some tweaks to manually fragment the
swapfile by breaking the extent at 1M boundary.  This made the swapfile
have 20K extents.

  nr_task=4
  kernel   swapout(KB/s) map_swap_entry(perf)  swapin(KB/s) map_swap_entry(perf)
  vanilla  165191           90.77%             171798          90.21%
  patched  858993 +420%      2.16%             715827 +317%     0.77%

  nr_task=8
  kernel   swapout(KB/s) map_swap_entry(perf)  swapin(KB/s) map_swap_entry(perf)
  vanilla  306783           92.19%             318145          87.76%
  patched  954437 +211%      2.35%            1073741 +237%     1.57%

swapout: the throughput of swap out, in KB/s, higher is better 1st
map_swap_entry: cpu cycles percent sampled by perf swapin: the
throughput of swap in, in KB/s, higher is better.  2nd map_swap_entry:
cpu cycles percent sampled by perf

nr_task=1 doesn't show any difference, this is due to the curr_swap_extent
can be effectively used to cache the correct swap extent for single task
workload.

[akpm@linux-foundation.org: s/BUG_ON(1)/BUG()/]
Link: http://lkml.kernel.org/r/20190523142404.GA181@aaronlu
Signed-off-by: Aaron Lu <ziqian.lzq@antfin.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

1 files changed, 1 insertions, 1 deletions

diff --git a/mm/page_io.c b/mm/page_io.c
index a39aac2f8c8d..24ee600f9131 100644
--- a/mm/page_io.c
+++ b/mm/page_io.c
@@ -163,7 +163,7 @@ int generic_swapfile_activate(struct swap_info_struct *sis,
 	blocks_per_page = PAGE_SIZE >> blkbits;
 
 	/*
-	 * Map all the blocks into the extent list.  This code doesn't try
+	 * Map all the blocks into the extent tree.  This code doesn't try
 	 * to be very smart.
 	 */
 	probe_block = 0;