git.openwrt.org Git - openwrt/staging/stintel.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 --- a/mm/vmscan.c
  28 +++ b/mm/vmscan.c
  29 @@ -4571,13 +4571,12 @@ static void lru_gen_age_node(struct pgli
  30  void lru_gen_look_around(struct page_vma_mapped_walk *pvmw)
  31  {
  32         int i;
  33 -       pte_t *pte;
  34         unsigned long start;
  35         unsigned long end;
  36 -       unsigned long addr;
  37         struct lru_gen_mm_walk *walk;
  38         int young = 0;
  39 -       unsigned long bitmap[BITS_TO_LONGS(MIN_LRU_BATCH)] = {};
  40 +       pte_t *pte = pvmw->pte;
  41 +       unsigned long addr = pvmw->address;
  42         struct folio *folio = pfn_folio(pvmw->pfn);
  43         struct mem_cgroup *memcg = folio_memcg(folio);
  44         struct pglist_data *pgdat = folio_pgdat(folio);
  45 @@ -4594,25 +4593,28 @@ void lru_gen_look_around(struct page_vma
  46         /* avoid taking the LRU lock under the PTL when possible */
  47         walk = current->reclaim_state ? current->reclaim_state->mm_walk : NULL;
  48
  49 -       start = max(pvmw->address & PMD_MASK, pvmw->vma->vm_start);
  50 -       end = min(pvmw->address | ~PMD_MASK, pvmw->vma->vm_end - 1) + 1;
  51 +       start = max(addr & PMD_MASK, pvmw->vma->vm_start);
  52 +       end = min(addr | ~PMD_MASK, pvmw->vma->vm_end - 1) + 1;
  53
  54         if (end - start > MIN_LRU_BATCH * PAGE_SIZE) {
  55 -               if (pvmw->address - start < MIN_LRU_BATCH * PAGE_SIZE / 2)
  56 +               if (addr - start < MIN_LRU_BATCH * PAGE_SIZE / 2)
  57                         end = start + MIN_LRU_BATCH * PAGE_SIZE;
  58 -               else if (end - pvmw->address < MIN_LRU_BATCH * PAGE_SIZE / 2)
  59 +               else if (end - addr < MIN_LRU_BATCH * PAGE_SIZE / 2)
  60                         start = end - MIN_LRU_BATCH * PAGE_SIZE;
  61                 else {
  62 -                       start = pvmw->address - MIN_LRU_BATCH * PAGE_SIZE / 2;
  63 -                       end = pvmw->address + MIN_LRU_BATCH * PAGE_SIZE / 2;
  64 +                       start = addr - MIN_LRU_BATCH * PAGE_SIZE / 2;
  65 +                       end = addr + MIN_LRU_BATCH * PAGE_SIZE / 2;
  66                 }
  67         }
  68
  69 -       pte = pvmw->pte - (pvmw->address - start) / PAGE_SIZE;
  70 +       /* folio_update_gen() requires stable folio_memcg() */
  71 +       if (!mem_cgroup_trylock_pages(memcg))
  72 +               return;
  73
  74 -       rcu_read_lock();
  75         arch_enter_lazy_mmu_mode();
  76
  77 +       pte -= (addr - start) / PAGE_SIZE;
  78 +
  79         for (i = 0, addr = start; addr != end; i++, addr += PAGE_SIZE) {
  80                 unsigned long pfn;
  81
  82 @@ -4637,56 +4639,27 @@ void lru_gen_look_around(struct page_vma
  83                       !folio_test_swapcache(folio)))
  84                         folio_mark_dirty(folio);
  85
  86 +               if (walk) {
  87 +                       old_gen = folio_update_gen(folio, new_gen);
  88 +                       if (old_gen >= 0 && old_gen != new_gen)
  89 +                               update_batch_size(walk, folio, old_gen, new_gen);
  90 +
  91 +                       continue;
  92 +               }
  93 +
  94                 old_gen = folio_lru_gen(folio);
  95                 if (old_gen < 0)
  96                         folio_set_referenced(folio);
  97                 else if (old_gen != new_gen)
  98 -                       __set_bit(i, bitmap);
  99 +                       folio_activate(folio);
 100         }
 101
 102         arch_leave_lazy_mmu_mode();
 103 -       rcu_read_unlock();
 104 +       mem_cgroup_unlock_pages();
 105
 106         /* feedback from rmap walkers to page table walkers */
 107         if (suitable_to_scan(i, young))
 108                 update_bloom_filter(lruvec, max_seq, pvmw->pmd);
 109 -
 110 -       if (!walk && bitmap_weight(bitmap, MIN_LRU_BATCH) < PAGEVEC_SIZE) {
 111 -               for_each_set_bit(i, bitmap, MIN_LRU_BATCH) {
 112 -                       folio = pfn_folio(pte_pfn(pte[i]));
 113 -                       folio_activate(folio);
 114 -               }
 115 -               return;
 116 -       }
 117 -
 118 -       /* folio_update_gen() requires stable folio_memcg() */
 119 -       if (!mem_cgroup_trylock_pages(memcg))
 120 -               return;
 121 -
 122 -       if (!walk) {
 123 -               spin_lock_irq(&lruvec->lru_lock);
 124 -               new_gen = lru_gen_from_seq(lruvec->lrugen.max_seq);
 125 -       }
 126 -
 127 -       for_each_set_bit(i, bitmap, MIN_LRU_BATCH) {
 128 -               folio = pfn_folio(pte_pfn(pte[i]));
 129 -               if (folio_memcg_rcu(folio) != memcg)
 130 -                       continue;
 131 -
 132 -               old_gen = folio_update_gen(folio, new_gen);
 133 -               if (old_gen < 0 || old_gen == new_gen)
 134 -                       continue;
 135 -
 136 -               if (walk)
 137 -                       update_batch_size(walk, folio, old_gen, new_gen);
 138 -               else
 139 -                       lru_gen_update_size(lruvec, folio, old_gen, new_gen);
 140 -       }
 141 -
 142 -       if (!walk)
 143 -               spin_unlock_irq(&lruvec->lru_lock);
 144 -
 145 -       mem_cgroup_unlock_pages();
 146  }
 147
 148  /******************************************************************************