zram: drop max_zpage_size and use zs_huge_class_size()
authorSergey Senozhatsky <sergey.senozhatsky.work@gmail.com>
Thu, 5 Apr 2018 23:24:47 +0000 (16:24 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 6 Apr 2018 04:36:26 +0000 (21:36 -0700)
Remove ZRAM's enforced "huge object" value and use zsmalloc huge-class
watermark instead, which makes more sense.

TEST
- I used a 1G zram device, LZO compression back-end, original
  data set size was 444MB. Looking at zsmalloc classes stats the
  test ended up to be pretty fair.

BASE ZRAM/ZSMALLOC
=====================
zram mm_stat

498978816 191482495 199831552        0 199831552    15634        0

zsmalloc classes

 class  size almost_full almost_empty obj_allocated   obj_used pages_used pages_per_zspage freeable
...
   151  2448           0            0          1240       1240        744                3        0
   168  2720           0            0          4200       4200       2800                2        0
   190  3072           0            0         10100      10100       7575                3        0
   202  3264           0            0           380        380        304                4        0
   254  4096           0            0         10620      10620      10620                1        0

 Total                 7           46        106982     106187      48787                         0

PATCHED ZRAM/ZSMALLOC
=====================

zram mm_stat

498978816 182579184 194248704        0 194248704    15628        0

zsmalloc classes

 class  size almost_full almost_empty obj_allocated   obj_used pages_used pages_per_zspage freeable
...
   151  2448           0            0          1240       1240        744                3        0
   168  2720           0            0          4200       4200       2800                2        0
   190  3072           0            0         10100      10100       7575                3        0
   202  3264           0            0          7180       7180       5744                4        0
   254  4096           0            0          3820       3820       3820                1        0

 Total                 8           45        106959     106193      47424                         0

As we can see, we reduced the number of objects stored in class-4096,
because a huge number of objects which we previously forcibly stored in
class-4096 now stored in non-huge class-3264.  This results in lower
memory consumption:

- zsmalloc now uses 47424 physical pages, which is less than 48787 pages
  zsmalloc used before.

- objects that we store in class-3264 share zspages.  That's why overall
  the number of pages that both class-4096 and class-3264 consumed went
  down from 10924 to 9564.

[sergey.senozhatsky.work@gmail.com: add pool param to zs_huge_class_size()]
Link: http://lkml.kernel.org/r/20180314081833.1096-3-sergey.senozhatsky@gmail.com
Link: http://lkml.kernel.org/r/20180306070639.7389-3-sergey.senozhatsky@gmail.com
Signed-off-by: Sergey Senozhatsky <sergey.senozhatsky@gmail.com>
Acked-by: Minchan Kim <minchan@kernel.org>
Cc: Mike Rapoport <rppt@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
drivers/block/zram/zram_drv.c
drivers/block/zram/zram_drv.h

index 0afa6c8..bd38f46 100644 (file)
@@ -44,6 +44,11 @@ static const char *default_compressor = "lzo";
 
 /* Module params (documentation at end) */
 static unsigned int num_devices = 1;
+/*
+ * Pages that compress to sizes equals or greater than this are stored
+ * uncompressed in memory.
+ */
+static size_t huge_class_size;
 
 static void zram_free_page(struct zram *zram, size_t index);
 
@@ -786,6 +791,8 @@ static bool zram_meta_alloc(struct zram *zram, u64 disksize)
                return false;
        }
 
+       if (!huge_class_size)
+               huge_class_size = zs_huge_class_size(zram->mem_pool);
        return true;
 }
 
@@ -965,7 +972,7 @@ compress_again:
                return ret;
        }
 
-       if (unlikely(comp_len > max_zpage_size)) {
+       if (unlikely(comp_len >= huge_class_size)) {
                if (zram_wb_enabled(zram) && allow_wb) {
                        zcomp_stream_put(zram->comp);
                        ret = write_to_bdev(zram, bvec, index, bio, &element);
index 31762db..d71c800 100644 (file)
 
 #include "zcomp.h"
 
-/*-- Configurable parameters */
-
-/*
- * Pages that compress to size greater than this are stored
- * uncompressed in memory.
- */
-static const size_t max_zpage_size = PAGE_SIZE / 4 * 3;
-
-/*
- * NOTE: max_zpage_size must be less than or equal to:
- *   ZS_MAX_ALLOC_SIZE. Otherwise, zs_malloc() would
- * always return failure.
- */
-
-/*-- End of configurable params */
-
 #define SECTOR_SHIFT           9
 #define SECTORS_PER_PAGE_SHIFT (PAGE_SHIFT - SECTOR_SHIFT)
 #define SECTORS_PER_PAGE       (1 << SECTORS_PER_PAGE_SHIFT)