target/linux/generic/backport-6.1/020-v6.3-19-mm-add-vma_has_recency.patch

   1 From 6c7f552a48b49a8612786a28a2239fbc24fac289 Mon Sep 17 00:00:00 2001
   2 From: Yu Zhao <yuzhao@google.com>
   3 Date: Fri, 30 Dec 2022 14:52:51 -0700
   4 Subject: [PATCH 19/29] mm: add vma_has_recency()
   5
   6 Add vma_has_recency() to indicate whether a VMA may exhibit temporal
   7 locality that the LRU algorithm relies on.
   8
   9 This function returns false for VMAs marked by VM_SEQ_READ or
  10 VM_RAND_READ.  While the former flag indicates linear access, i.e., a
  11 special case of spatial locality, both flags indicate a lack of temporal
  12 locality, i.e., the reuse of an area within a relatively small duration.
  13
  14 "Recency" is chosen over "locality" to avoid confusion between temporal
  15 and spatial localities.
  16
  17 Before this patch, the active/inactive LRU only ignored the accessed bit
  18 from VMAs marked by VM_SEQ_READ.  After this patch, the active/inactive
  19 LRU and MGLRU share the same logic: they both ignore the accessed bit if
  20 vma_has_recency() returns false.
  21
  22 For the active/inactive LRU, the following fio test showed a [6, 8]%
  23 increase in IOPS when randomly accessing mapped files under memory
  24 pressure.
  25
  26   kb=$(awk '/MemTotal/ { print $2 }' /proc/meminfo)
  27   kb=$((kb - 8*1024*1024))
  28
  29   modprobe brd rd_nr=1 rd_size=$kb
  30   dd if=/dev/zero of=/dev/ram0 bs=1M
  31
  32   mkfs.ext4 /dev/ram0
  33   mount /dev/ram0 /mnt/
  34   swapoff -a
  35
  36   fio --name=test --directory=/mnt/ --ioengine=mmap --numjobs=8 \
  37       --size=8G --rw=randrw --time_based --runtime=10m \
  38       --group_reporting
  39
  40 The discussion that led to this patch is here [1].  Additional test
  41 results are available in that thread.
  42
  43 [1] https://lore.kernel.org/r/Y31s%2FK8T85jh05wH@google.com/
  44
  45 Link: https://lkml.kernel.org/r/20221230215252.2628425-1-yuzhao@google.com
  46 Signed-off-by: Yu Zhao <yuzhao@google.com>
  47 Cc: Alexander Viro <viro@zeniv.linux.org.uk>
  48 Cc: Andrea Righi <andrea.righi@canonical.com>
  49 Cc: Johannes Weiner <hannes@cmpxchg.org>
  50 Cc: Michael Larabel <Michael@MichaelLarabel.com>
  51 Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
  52 ---
  53  include/linux/mm_inline.h |  9 +++++++++
  54  mm/memory.c               |  8 ++++----
  55  mm/rmap.c                 | 42 +++++++++++++++++----------------------
  56  mm/vmscan.c               |  5 ++++-
  57  4 files changed, 35 insertions(+), 29 deletions(-)
  58
  59 --- a/include/linux/mm_inline.h
  60 +++ b/include/linux/mm_inline.h
  61 @@ -578,4 +578,12 @@ pte_install_uffd_wp_if_needed(struct vm_
  62  #endif
  63  }
  64
  65 +static inline bool vma_has_recency(struct vm_area_struct *vma)
  66 +{
  67 +       if (vma->vm_flags & (VM_SEQ_READ | VM_RAND_READ))
  68 +               return false;
  69 +
  70 +       return true;
  71 +}
  72 +
  73  #endif
  74 --- a/mm/memory.c
  75 +++ b/mm/memory.c
  76 @@ -1435,8 +1435,7 @@ again:
  77                                         force_flush = 1;
  78                                         set_page_dirty(page);
  79                                 }
  80 -                               if (pte_young(ptent) &&
  81 -                                   likely(!(vma->vm_flags & VM_SEQ_READ)))
  82 +                               if (pte_young(ptent) && likely(vma_has_recency(vma)))
  83                                         mark_page_accessed(page);
  84                         }
  85                         rss[mm_counter(page)]--;
  86 @@ -5170,8 +5169,8 @@ static inline void mm_account_fault(stru
  87  #ifdef CONFIG_LRU_GEN
  88  static void lru_gen_enter_fault(struct vm_area_struct *vma)
  89  {
  90 -       /* the LRU algorithm doesn't apply to sequential or random reads */
  91 -       current->in_lru_fault = !(vma->vm_flags & (VM_SEQ_READ | VM_RAND_READ));
  92 +       /* the LRU algorithm only applies to accesses with recency */
  93 +       current->in_lru_fault = vma_has_recency(vma);
  94  }
  95
  96  static void lru_gen_exit_fault(void)
  97 --- a/mm/rmap.c
  98 +++ b/mm/rmap.c
  99 @@ -823,25 +823,14 @@ static bool folio_referenced_one(struct
 100                 }
 101
 102                 if (pvmw.pte) {
 103 -                       if (lru_gen_enabled() && pte_young(*pvmw.pte) &&
 104 -                           !(vma->vm_flags & (VM_SEQ_READ | VM_RAND_READ))) {
 105 +                       if (lru_gen_enabled() && pte_young(*pvmw.pte)) {
 106                                 lru_gen_look_around(&pvmw);
 107                                 referenced++;
 108                         }
 109
 110                         if (ptep_clear_flush_young_notify(vma, address,
 111 -                                               pvmw.pte)) {
 112 -                               /*
 113 -                                * Don't treat a reference through
 114 -                                * a sequentially read mapping as such.
 115 -                                * If the folio has been used in another mapping,
 116 -                                * we will catch it; if this other mapping is
 117 -                                * already gone, the unmap path will have set
 118 -                                * the referenced flag or activated the folio.
 119 -                                */
 120 -                               if (likely(!(vma->vm_flags & VM_SEQ_READ)))
 121 -                                       referenced++;
 122 -                       }
 123 +                                               pvmw.pte))
 124 +                               referenced++;
 125                 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
 126                         if (pmdp_clear_flush_young_notify(vma, address,
 127                                                 pvmw.pmd))
 128 @@ -875,7 +864,20 @@ static bool invalid_folio_referenced_vma
 129         struct folio_referenced_arg *pra = arg;
 130         struct mem_cgroup *memcg = pra->memcg;
 131
 132 -       if (!mm_match_cgroup(vma->vm_mm, memcg))
 133 +       /*
 134 +        * Ignore references from this mapping if it has no recency. If the
 135 +        * page has been used in another mapping, we will catch it; if this
 136 +        * other mapping is already gone, the unmap path will have set the
 137 +        * referenced flag or activated the page in zap_pte_range().
 138 +        */
 139 +       if (!vma_has_recency(vma))
 140 +               return true;
 141 +
 142 +       /*
 143 +        * If we are reclaiming on behalf of a cgroup, skip counting on behalf
 144 +        * of references from different cgroups.
 145 +        */
 146 +       if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
 147                 return true;
 148
 149         return false;
 150 @@ -906,6 +908,7 @@ int folio_referenced(struct folio *folio
 151                 .arg = (void *)&pra,
 152                 .anon_lock = folio_lock_anon_vma_read,
 153                 .try_lock = true,
 154 +               .invalid_vma = invalid_folio_referenced_vma,
 155         };
 156
 157         *vm_flags = 0;
 158 @@ -921,15 +924,6 @@ int folio_referenced(struct folio *folio
 159                         return 1;
 160         }
 161
 162 -       /*
 163 -        * If we are reclaiming on behalf of a cgroup, skip
 164 -        * counting on behalf of references from different
 165 -        * cgroups
 166 -        */
 167 -       if (memcg) {
 168 -               rwc.invalid_vma = invalid_folio_referenced_vma;
 169 -       }
 170 -
 171         rmap_walk(folio, &rwc);
 172         *vm_flags = pra.vm_flags;
 173
 174 --- a/mm/vmscan.c
 175 +++ b/mm/vmscan.c
 176 @@ -3766,7 +3766,10 @@ static int should_skip_vma(unsigned long
 177         if (is_vm_hugetlb_page(vma))
 178                 return true;
 179
 180 -       if (vma->vm_flags & (VM_LOCKED | VM_SPECIAL | VM_SEQ_READ | VM_RAND_READ))
 181 +       if (!vma_has_recency(vma))
 182 +               return true;
 183 +
 184 +       if (vma->vm_flags & (VM_LOCKED | VM_SPECIAL))
 185                 return true;
 186
 187         if (vma == get_gate_vma(vma->vm_mm))