git.openwrt.org Git - openwrt/staging/linusw.git/blob

   1 From e604c3ccb4dfbdde2467fccef9bb36170a392695 Mon Sep 17 00:00:00 2001
   2 From: "T.J. Alumbaugh" <talumbau@google.com>
   3 Date: Wed, 18 Jan 2023 00:18:27 +0000
   4 Subject: [PATCH 18/19] UPSTREAM: mm: multi-gen LRU: simplify
   5  lru_gen_look_around()
   6
   7 Update the folio generation in place with or without
   8 current->reclaim_state->mm_walk.  The LRU lock is held for longer, if
   9 mm_walk is NULL and the number of folios to update is more than
  10 PAGEVEC_SIZE.
  11
  12 This causes a measurable regression from the LRU lock contention during a
  13 microbencmark.  But a tiny regression is not worth the complexity.
  14
  15 Link: https://lkml.kernel.org/r/20230118001827.1040870-8-talumbau@google.com
  16 Change-Id: I9ce18b4f4062e6c1c13c98ece9422478eb8e1846
  17 Signed-off-by: T.J. Alumbaugh <talumbau@google.com>
  18 Cc: Yu Zhao <yuzhao@google.com>
  19 Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
  20 (cherry picked from commit abf086721a2f1e6897c57796f7268df1b194c750)
  21 Bug: 274865848
  22 Signed-off-by: T.J. Mercier <tjmercier@google.com>
  23 ---
  24  mm/vmscan.c | 73 +++++++++++++++++------------------------------------
  25  1 file changed, 23 insertions(+), 50 deletions(-)
  26
  27 diff --git a/mm/vmscan.c b/mm/vmscan.c
  28 index 8f496c2e670a9..f6ce7a1fd78a3 100644
  29 --- a/mm/vmscan.c
  30 +++ b/mm/vmscan.c
  31 @@ -4571,13 +4571,12 @@ static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
  32  void lru_gen_look_around(struct page_vma_mapped_walk *pvmw)
  33  {
  34         int i;
  35 -       pte_t *pte;
  36         unsigned long start;
  37         unsigned long end;
  38 -       unsigned long addr;
  39         struct lru_gen_mm_walk *walk;
  40         int young = 0;
  41 -       unsigned long bitmap[BITS_TO_LONGS(MIN_LRU_BATCH)] = {};
  42 +       pte_t *pte = pvmw->pte;
  43 +       unsigned long addr = pvmw->address;
  44         struct folio *folio = pfn_folio(pvmw->pfn);
  45         struct mem_cgroup *memcg = folio_memcg(folio);
  46         struct pglist_data *pgdat = folio_pgdat(folio);
  47 @@ -4594,25 +4593,28 @@ void lru_gen_look_around(struct page_vma_mapped_walk *pvmw)
  48         /* avoid taking the LRU lock under the PTL when possible */
  49         walk = current->reclaim_state ? current->reclaim_state->mm_walk : NULL;
  50
  51 -       start = max(pvmw->address & PMD_MASK, pvmw->vma->vm_start);
  52 -       end = min(pvmw->address | ~PMD_MASK, pvmw->vma->vm_end - 1) + 1;
  53 +       start = max(addr & PMD_MASK, pvmw->vma->vm_start);
  54 +       end = min(addr | ~PMD_MASK, pvmw->vma->vm_end - 1) + 1;
  55
  56         if (end - start > MIN_LRU_BATCH * PAGE_SIZE) {
  57 -               if (pvmw->address - start < MIN_LRU_BATCH * PAGE_SIZE / 2)
  58 +               if (addr - start < MIN_LRU_BATCH * PAGE_SIZE / 2)
  59                         end = start + MIN_LRU_BATCH * PAGE_SIZE;
  60 -               else if (end - pvmw->address < MIN_LRU_BATCH * PAGE_SIZE / 2)
  61 +               else if (end - addr < MIN_LRU_BATCH * PAGE_SIZE / 2)
  62                         start = end - MIN_LRU_BATCH * PAGE_SIZE;
  63                 else {
  64 -                       start = pvmw->address - MIN_LRU_BATCH * PAGE_SIZE / 2;
  65 -                       end = pvmw->address + MIN_LRU_BATCH * PAGE_SIZE / 2;
  66 +                       start = addr - MIN_LRU_BATCH * PAGE_SIZE / 2;
  67 +                       end = addr + MIN_LRU_BATCH * PAGE_SIZE / 2;
  68                 }
  69         }
  70
  71 -       pte = pvmw->pte - (pvmw->address - start) / PAGE_SIZE;
  72 +       /* folio_update_gen() requires stable folio_memcg() */
  73 +       if (!mem_cgroup_trylock_pages(memcg))
  74 +               return;
  75
  76 -       rcu_read_lock();
  77         arch_enter_lazy_mmu_mode();
  78
  79 +       pte -= (addr - start) / PAGE_SIZE;
  80 +
  81         for (i = 0, addr = start; addr != end; i++, addr += PAGE_SIZE) {
  82                 unsigned long pfn;
  83
  84 @@ -4637,56 +4639,27 @@ void lru_gen_look_around(struct page_vma_mapped_walk *pvmw)
  85                       !folio_test_swapcache(folio)))
  86                         folio_mark_dirty(folio);
  87
  88 +               if (walk) {
  89 +                       old_gen = folio_update_gen(folio, new_gen);
  90 +                       if (old_gen >= 0 && old_gen != new_gen)
  91 +                               update_batch_size(walk, folio, old_gen, new_gen);
  92 +
  93 +                       continue;
  94 +               }
  95 +
  96                 old_gen = folio_lru_gen(folio);
  97                 if (old_gen < 0)
  98                         folio_set_referenced(folio);
  99                 else if (old_gen != new_gen)
 100 -                       __set_bit(i, bitmap);
 101 +                       folio_activate(folio);
 102         }
 103
 104         arch_leave_lazy_mmu_mode();
 105 -       rcu_read_unlock();
 106 +       mem_cgroup_unlock_pages();
 107
 108         /* feedback from rmap walkers to page table walkers */
 109         if (suitable_to_scan(i, young))
 110                 update_bloom_filter(lruvec, max_seq, pvmw->pmd);
 111 -
 112 -       if (!walk && bitmap_weight(bitmap, MIN_LRU_BATCH) < PAGEVEC_SIZE) {
 113 -               for_each_set_bit(i, bitmap, MIN_LRU_BATCH) {
 114 -                       folio = pfn_folio(pte_pfn(pte[i]));
 115 -                       folio_activate(folio);
 116 -               }
 117 -               return;
 118 -       }
 119 -
 120 -       /* folio_update_gen() requires stable folio_memcg() */
 121 -       if (!mem_cgroup_trylock_pages(memcg))
 122 -               return;
 123 -
 124 -       if (!walk) {
 125 -               spin_lock_irq(&lruvec->lru_lock);
 126 -               new_gen = lru_gen_from_seq(lruvec->lrugen.max_seq);
 127 -       }
 128 -
 129 -       for_each_set_bit(i, bitmap, MIN_LRU_BATCH) {
 130 -               folio = pfn_folio(pte_pfn(pte[i]));
 131 -               if (folio_memcg_rcu(folio) != memcg)
 132 -                       continue;
 133 -
 134 -               old_gen = folio_update_gen(folio, new_gen);
 135 -               if (old_gen < 0 || old_gen == new_gen)
 136 -                       continue;
 137 -
 138 -               if (walk)
 139 -                       update_batch_size(walk, folio, old_gen, new_gen);
 140 -               else
 141 -                       lru_gen_update_size(lruvec, folio, old_gen, new_gen);
 142 -       }
 143 -
 144 -       if (!walk)
 145 -               spin_unlock_irq(&lruvec->lru_lock);
 146 -
 147 -       mem_cgroup_unlock_pages();
 148  }
 149
 150  /******************************************************************************
 151 --
 152 2.40.1
 153