slub: optimize bulk slowpath free by detached freelist

This change focus on improving the speed of object freeing in the
"slowpath" of kmem_cache_free_bulk.

The calls slab_free (fastpath) and __slab_free (slowpath) have been
extended with support for bulk free, which amortize the overhead of
the (locked) cmpxchg_double.

To use the new bulking feature, we build what I call a detached
freelist.  The detached freelist takes advantage of three properties:

 1) the free function call owns the object that is about to be freed,
    thus writing into this memory is synchronization-free.

 2) many freelist's can co-exist side-by-side in the same slab-page
    each with a separate head pointer.

 3) it is the visibility of the head pointer that needs synchronization.

Given these properties, the brilliant part is that the detached
freelist can be constructed without any need for synchronization.  The
freelist is constructed directly in the page objects, without any
synchronization needed.  The detached freelist is allocated on the
stack of the function call kmem_cache_free_bulk.  Thus, the freelist
head pointer is not visible to other CPUs.

All objects in a SLUB freelist must belong to the same slab-page.
Thus, constructing the detached freelist is about matching objects
that belong to the same slab-page.  The bulk free array is scanned is
a progressive manor with a limited look-ahead facility.

Kmem debug support is handled in call of slab_free().

Notice kmem_cache_free_bulk no longer need to disable IRQs. This
only slowed down single free bulk with approx 3 cycles.

Performance data:
 Benchmarked[1] obj size 256 bytes on CPU i7-4790K @ 4.00GHz

SLUB fastpath single object quick reuse: 47 cycles(tsc) 11.931 ns

To get stable and comparable numbers, the kernel have been booted with
"slab_merge" (this also improve performance for larger bulk sizes).

Performance data, compared against fallback bulking:

bulk -  fallback bulk            - improvement with this patch
   1 -  62 cycles(tsc) 15.662 ns - 49 cycles(tsc) 12.407 ns- improved 21.0%
   2 -  55 cycles(tsc) 13.935 ns - 30 cycles(tsc) 7.506 ns - improved 45.5%
   3 -  53 cycles(tsc) 13.341 ns - 23 cycles(tsc) 5.865 ns - improved 56.6%
   4 -  52 cycles(tsc) 13.081 ns - 20 cycles(tsc) 5.048 ns - improved 61.5%
   8 -  50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.659 ns - improved 64.0%
  16 -  49 cycles(tsc) 12.412 ns - 17 cycles(tsc) 4.495 ns - improved 65.3%
  30 -  49 cycles(tsc) 12.484 ns - 18 cycles(tsc) 4.533 ns - improved 63.3%
  32 -  50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.707 ns - improved 64.0%
  34 -  96 cycles(tsc) 24.243 ns - 23 cycles(tsc) 5.976 ns - improved 76.0%
  48 -  83 cycles(tsc) 20.818 ns - 21 cycles(tsc) 5.329 ns - improved 74.7%
  64 -  74 cycles(tsc) 18.700 ns - 20 cycles(tsc) 5.127 ns - improved 73.0%
 128 -  90 cycles(tsc) 22.734 ns - 27 cycles(tsc) 6.833 ns - improved 70.0%
 158 -  99 cycles(tsc) 24.776 ns - 30 cycles(tsc) 7.583 ns - improved 69.7%
 250 - 104 cycles(tsc) 26.089 ns - 37 cycles(tsc) 9.280 ns - improved 64.4%

Performance data, compared current in-kernel bulking:

bulk - curr in-kernel  - improvement with this patch
   1 -  46 cycles(tsc) - 49 cycles(tsc) - improved (cycles:-3) -6.5%
   2 -  27 cycles(tsc) - 30 cycles(tsc) - improved (cycles:-3) -11.1%
   3 -  21 cycles(tsc) - 23 cycles(tsc) - improved (cycles:-2) -9.5%
   4 -  18 cycles(tsc) - 20 cycles(tsc) - improved (cycles:-2) -11.1%
   8 -  17 cycles(tsc) - 18 cycles(tsc) - improved (cycles:-1) -5.9%
  16 -  18 cycles(tsc) - 17 cycles(tsc) - improved (cycles: 1)  5.6%
  30 -  18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0)  0.0%
  32 -  18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0)  0.0%
  34 -  78 cycles(tsc) - 23 cycles(tsc) - improved (cycles:55) 70.5%
  48 -  60 cycles(tsc) - 21 cycles(tsc) - improved (cycles:39) 65.0%
  64 -  49 cycles(tsc) - 20 cycles(tsc) - improved (cycles:29) 59.2%
 128 -  69 cycles(tsc) - 27 cycles(tsc) - improved (cycles:42) 60.9%
 158 -  79 cycles(tsc) - 30 cycles(tsc) - improved (cycles:49) 62.0%
 250 -  86 cycles(tsc) - 37 cycles(tsc) - improved (cycles:49) 57.0%

Performance with normal SLUB merging is significantly slower for
larger bulking.  This is believed to (primarily) be an effect of not
having to share the per-CPU data-structures, as tuning per-CPU size
can achieve similar performance.

bulk - slab_nomerge   -  normal SLUB merge
   1 -  49 cycles(tsc) - 49 cycles(tsc) - merge slower with cycles:0
   2 -  30 cycles(tsc) - 30 cycles(tsc) - merge slower with cycles:0
   3 -  23 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:0
   4 -  20 cycles(tsc) - 20 cycles(tsc) - merge slower with cycles:0
   8 -  18 cycles(tsc) - 18 cycles(tsc) - merge slower with cycles:0
  16 -  17 cycles(tsc) - 17 cycles(tsc) - merge slower with cycles:0
  30 -  18 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:5
  32 -  18 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:4
  34 -  23 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:-1
  48 -  21 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:1
  64 -  20 cycles(tsc) - 48 cycles(tsc) - merge slower with cycles:28
 128 -  27 cycles(tsc) - 57 cycles(tsc) - merge slower with cycles:30
 158 -  30 cycles(tsc) - 59 cycles(tsc) - merge slower with cycles:29
 250 -  37 cycles(tsc) - 56 cycles(tsc) - merge slower with cycles:19

Joint work with Alexander Duyck.

[1] https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/mm/slab_bulk_test01.c



[akpm@linux-foundation.org: BUG_ON -> WARN_ON;return]
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: Alexander Duyck <alexander.h.duyck@redhat.com>
Acked-by: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>