Re: [RFC v2 00/34] SLUB: reduce irq disabled scope and make it RT compatible

[Mike Galbraith <efault@gmx.de>]

On Sat, 2021-07-03 at 09:24 +0200, Mike Galbraith wrote:
>
> It also appears to be saying that there's something RT specific to
> stare at in addition to the list_lock business.

The what is ___slab_alloc() consuming 3.9% CPU in tip-rt-slub whereas
it consumes < 1% in both tip-rt (sans slub patches) and tip-slub.

The why remains to ponder.

            5.13.0.g60ab3ed-tip-rt           5.13.0.g60ab3ed-tip-rt-slub      5.13.0.g60ab3ed-tip-slub
    25.18%  copy_user_enhanced_fast_string   copy_user_enhanced_fast_string   copy_user_enhanced_fast_string
     5.08%  unix_stream_read_generic         unix_stream_read_generic         unix_stream_read_generic
     3.39%  rt_spin_lock                 *** ___slab_alloc ***                __skb_datagram_iter
     2.80%  __skb_datagram_iter              rt_spin_lock                     _raw_spin_lock
     2.11%  get_page_from_freelist           __skb_datagram_iter              __alloc_skb
     2.01%  skb_release_data                 rt_spin_unlock                   skb_release_data
     1.94%  rt_spin_unlock                   get_page_from_freelist           __alloc_pages
     1.85%  __alloc_skb                      migrate_enable                   unix_stream_sendmsg
     1.68%  __schedule                       skb_release_data                 _raw_spin_lock_irqsave
     1.67%  unix_stream_sendmsg              __schedule                       free_pcppages_bulk
     1.50%  free_pcppages_bulk               unix_stream_sendmsg              __slab_free
     1.38%  migrate_enable                   free_pcppages_bulk               __fget_light
     1.24%  __fget_light                     __alloc_pages                    vfs_write
     1.16%  __slab_free                      migrate_disable                  __schedule
     1.14%  __alloc_pages                    __fget_light                     get_page_from_freelist
     1.10%  fsnotify                         __slab_free                      new_sync_write
     1.07%  kfree                            fsnotify                         fsnotify

5.13.0.g60ab3ed-tip-rt-slub ___slab_alloc() consumes 3.90%
  0.40 │       mov    0x28(%r13),%edx
  0.42 │       add    %r15,%rdx
       │     __swab():
       │     #endif
       │
       │     static __always_inline unsigned long __swab(const unsigned long y)
       │     {
       │     #if __BITS_PER_LONG == 64
       │     return __swab64(y);
  0.05 │       mov    %rdx,%rax
  1.14 │       bswap  %rax
       │     freelist_ptr():
       │     return (void *)((unsigned long)ptr ^ s->random ^  <== CONFIG_SLAB_FREELIST_HARDENED
  0.72 │       xor    0xb0(%r13),%rax
 65.41 │       xor    (%rdx),%rax                              <== huh? miss = 65% of that 3.9% kernel util?
       │     next_tid():
       │     return tid + TID_STEP;
  0.09 │       addq   $0x200,0x48(%r12)
       │     ___slab_alloc():
       │     * freelist is pointing to the list of objects to be used.
       │     * page is pointing to the page from which the objects are obtained.
       │     * That page must be frozen for per cpu allocations to work.
       │     */
       │     VM_BUG_ON(!c->page->frozen);
       │     c->freelist = get_freepointer(s, freelist);
  0.05 │       mov    %rax,0x40(%r12)
       │     c->tid = next_tid(c->tid);
       │     local_unlock_irqrestore(&s->cpu_slab->lock, flags);

5.13.0.g60ab3ed-tip-rt ___slab_alloc() consumes < 1%
Percent│     }
       │
       │     /* must check again c->freelist in case of cpu migration or IRQ */
       │     freelist = c->freelist;
  0.02 │ a1:   mov    (%r14),%r13
       │     if (freelist)
       │       test   %r13,%r13
  0.02 │     ↓ je     460
       │     get_freepointer():
       │     return freelist_dereference(s, object + s->offset);
  0.23 │ ad:   mov    0x28(%r12),%edx
  0.18 │       add    %r13,%rdx
       │     __swab():
       │     #endif
       │
       │     static __always_inline unsigned long __swab(const unsigned long y)
       │     {
       │     #if __BITS_PER_LONG == 64
       │     return __swab64(y);
  0.06 │       mov    %rdx,%rax
  1.16 │       bswap  %rax
       │     freelist_ptr():
       │     return (void *)((unsigned long)ptr ^ s->random ^
  0.23 │       xor    0xb0(%r12),%rax
 35.25 │       xor    (%rdx),%rax              <== 35% of < 1% kernel util
       │     next_tid():
       │     return tid + TID_STEP;
  0.28 │       addq   $0x200,0x8(%r14)
       │     ___slab_alloc():
       │     * freelist is pointing to the list of objects to be used.
       │     * page is pointing to the page from which the objects are obtained.
       │     * That page must be frozen for per cpu allocations to work.
       │     */
       │     VM_BUG_ON(!c->page->frozen);
       │     c->freelist = get_freepointer(s, freelist);

5.13.0.g60ab3ed-tip-slub ___slab_alloc() also consumes < 1%
Percent│     load_freelist:
       │
       │     lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock));
  0.28 │ 84:   add    this_cpu_off,%rax
       │     get_freepointer():
       │     return freelist_dereference(s, object + s->offset);
  0.14 │       mov    0x28(%r14),%eax
       │     ___slab_alloc():
       │     * freelist is pointing to the list of objects to be used.
       │     * page is pointing to the page from which the objects are obtained.
       │     * That page must be frozen for per cpu allocations to work.
       │     */
       │     VM_BUG_ON(!c->page->frozen);
       │     c->freelist = get_freepointer(s, freelist);
 34.36 │       mov    0x0(%r13,%rax,1),%rax
       │     next_tid():
       │     return tid + TID_STEP;
  0.10 │       addq   $0x1,0x8(%r12)
       │     ___slab_alloc():
       │     c->freelist = get_freepointer(s, freelist);
  0.04 │       mov    %rax,(%r12)
       │     c->tid = next_tid(c->tid);
       │     local_unlock_irqrestore(&s->cpu_slab->lock, flags);
  0.12 │       mov    (%r14),%rax
  0.03 │       add    this_cpu_off,%rax
       │     arch_local_irq_restore():
       │     return arch_irqs_disabled_flags(flags);
       │     }