target/linux/generic/backport-6.1/020-v6.3-03-UPSTREAM-mm-multi-gen-LRU-remove-eviction-fairness-s.patch

   1 From 14f9a7a15f3d1af351f30e0438fd747b7ac253b0 Mon Sep 17 00:00:00 2001
   2 From: Yu Zhao <yuzhao@google.com>
   3 Date: Wed, 21 Dec 2022 21:19:01 -0700
   4 Subject: [PATCH 03/19] UPSTREAM: mm: multi-gen LRU: remove eviction fairness
   5  safeguard
   6
   7 Recall that the eviction consumes the oldest generation: first it
   8 bucket-sorts folios whose gen counters were updated by the aging and
   9 reclaims the rest; then it increments lrugen->min_seq.
  10
  11 The current eviction fairness safeguard for global reclaim has a
  12 dilemma: when there are multiple eligible memcgs, should it continue
  13 or stop upon meeting the reclaim goal? If it continues, it overshoots
  14 and increases direct reclaim latency; if it stops, it loses fairness
  15 between memcgs it has taken memory away from and those it has yet to.
  16
  17 With memcg LRU, the eviction, while ensuring eventual fairness, will
  18 stop upon meeting its goal. Therefore the current eviction fairness
  19 safeguard for global reclaim will not be needed.
  20
  21 Note that memcg LRU only applies to global reclaim. For memcg reclaim,
  22 the eviction will continue, even if it is overshooting. This becomes
  23 unconditional due to code simplification.
  24
  25 Link: https://lkml.kernel.org/r/20221222041905.2431096-4-yuzhao@google.com
  26 Signed-off-by: Yu Zhao <yuzhao@google.com>
  27 Cc: Johannes Weiner <hannes@cmpxchg.org>
  28 Cc: Jonathan Corbet <corbet@lwn.net>
  29 Cc: Michael Larabel <Michael@MichaelLarabel.com>
  30 Cc: Michal Hocko <mhocko@kernel.org>
  31 Cc: Mike Rapoport <rppt@kernel.org>
  32 Cc: Roman Gushchin <roman.gushchin@linux.dev>
  33 Cc: Suren Baghdasaryan <surenb@google.com>
  34 Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
  35 Bug: 274865848
  36 (cherry picked from commit a579086c99ed70cc4bfc104348dbe3dd8f2787e6)
  37 Change-Id: I08ac1b3c90e29cafd0566785aaa4bcdb5db7d22c
  38 Signed-off-by: T.J. Mercier <tjmercier@google.com>
  39 ---
  40  mm/vmscan.c | 81 +++++++++++++++--------------------------------------
  41  1 file changed, 23 insertions(+), 58 deletions(-)
  42
  43 --- a/mm/vmscan.c
  44 +++ b/mm/vmscan.c
  45 @@ -448,6 +448,11 @@ static bool cgroup_reclaim(struct scan_c
  46         return sc->target_mem_cgroup;
  47  }
  48
  49 +static bool global_reclaim(struct scan_control *sc)
  50 +{
  51 +       return !sc->target_mem_cgroup || mem_cgroup_is_root(sc->target_mem_cgroup);
  52 +}
  53 +
  54  /**
  55   * writeback_throttling_sane - is the usual dirty throttling mechanism available?
  56   * @sc: scan_control in question
  57 @@ -498,6 +503,11 @@ static bool cgroup_reclaim(struct scan_c
  58         return false;
  59  }
  60
  61 +static bool global_reclaim(struct scan_control *sc)
  62 +{
  63 +       return true;
  64 +}
  65 +
  66  static bool writeback_throttling_sane(struct scan_control *sc)
  67  {
  68         return true;
  69 @@ -4993,8 +5003,7 @@ static int isolate_folios(struct lruvec
  70         return scanned;
  71  }
  72
  73 -static int evict_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness,
  74 -                       bool *need_swapping)
  75 +static int evict_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness)
  76  {
  77         int type;
  78         int scanned;
  79 @@ -5083,9 +5092,6 @@ retry:
  80                 goto retry;
  81         }
  82
  83 -       if (need_swapping && type == LRU_GEN_ANON)
  84 -               *need_swapping = true;
  85 -
  86         return scanned;
  87  }
  88
  89 @@ -5124,67 +5130,26 @@ done:
  90         return min_seq[!can_swap] + MIN_NR_GENS <= max_seq ? nr_to_scan : 0;
  91  }
  92
  93 -static bool should_abort_scan(struct lruvec *lruvec, unsigned long seq,
  94 -                             struct scan_control *sc, bool need_swapping)
  95 +static unsigned long get_nr_to_reclaim(struct scan_control *sc)
  96  {
  97 -       int i;
  98 -       DEFINE_MAX_SEQ(lruvec);
  99 -
 100 -       if (!current_is_kswapd()) {
 101 -               /* age each memcg at most once to ensure fairness */
 102 -               if (max_seq - seq > 1)
 103 -                       return true;
 104 -
 105 -               /* over-swapping can increase allocation latency */
 106 -               if (sc->nr_reclaimed >= sc->nr_to_reclaim && need_swapping)
 107 -                       return true;
 108 -
 109 -               /* give this thread a chance to exit and free its memory */
 110 -               if (fatal_signal_pending(current)) {
 111 -                       sc->nr_reclaimed += MIN_LRU_BATCH;
 112 -                       return true;
 113 -               }
 114 -
 115 -               if (cgroup_reclaim(sc))
 116 -                       return false;
 117 -       } else if (sc->nr_reclaimed - sc->last_reclaimed < sc->nr_to_reclaim)
 118 -               return false;
 119 -
 120 -       /* keep scanning at low priorities to ensure fairness */
 121 -       if (sc->priority > DEF_PRIORITY - 2)
 122 -               return false;
 123 -
 124 -       /*
 125 -        * A minimum amount of work was done under global memory pressure. For
 126 -        * kswapd, it may be overshooting. For direct reclaim, the allocation
 127 -        * may succeed if all suitable zones are somewhat safe. In either case,
 128 -        * it's better to stop now, and restart later if necessary.
 129 -        */
 130 -       for (i = 0; i <= sc->reclaim_idx; i++) {
 131 -               unsigned long wmark;
 132 -               struct zone *zone = lruvec_pgdat(lruvec)->node_zones + i;
 133 -
 134 -               if (!managed_zone(zone))
 135 -                       continue;
 136 -
 137 -               wmark = current_is_kswapd() ? high_wmark_pages(zone) : low_wmark_pages(zone);
 138 -               if (wmark > zone_page_state(zone, NR_FREE_PAGES))
 139 -                       return false;
 140 -       }
 141 +       /* don't abort memcg reclaim to ensure fairness */
 142 +       if (!global_reclaim(sc))
 143 +               return -1;
 144
 145 -       sc->nr_reclaimed += MIN_LRU_BATCH;
 146 +       /* discount the previous progress for kswapd */
 147 +       if (current_is_kswapd())
 148 +               return sc->nr_to_reclaim + sc->last_reclaimed;
 149
 150 -       return true;
 151 +       return max(sc->nr_to_reclaim, compact_gap(sc->order));
 152  }
 153
 154  static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
 155  {
 156         struct blk_plug plug;
 157         bool need_aging = false;
 158 -       bool need_swapping = false;
 159         unsigned long scanned = 0;
 160         unsigned long reclaimed = sc->nr_reclaimed;
 161 -       DEFINE_MAX_SEQ(lruvec);
 162 +       unsigned long nr_to_reclaim = get_nr_to_reclaim(sc);
 163
 164         lru_add_drain();
 165
 166 @@ -5208,7 +5173,7 @@ static void lru_gen_shrink_lruvec(struct
 167                 if (!nr_to_scan)
 168                         goto done;
 169
 170 -               delta = evict_folios(lruvec, sc, swappiness, &need_swapping);
 171 +               delta = evict_folios(lruvec, sc, swappiness);
 172                 if (!delta)
 173                         goto done;
 174
 175 @@ -5216,7 +5181,7 @@ static void lru_gen_shrink_lruvec(struct
 176                 if (scanned >= nr_to_scan)
 177                         break;
 178
 179 -               if (should_abort_scan(lruvec, max_seq, sc, need_swapping))
 180 +               if (sc->nr_reclaimed >= nr_to_reclaim)
 181                         break;
 182
 183                 cond_resched();
 184 @@ -5666,7 +5631,7 @@ static int run_eviction(struct lruvec *l
 185                 if (sc->nr_reclaimed >= nr_to_reclaim)
 186                         return 0;
 187
 188 -               if (!evict_folios(lruvec, sc, swappiness, NULL))
 189 +               if (!evict_folios(lruvec, sc, swappiness))
 190                         return 0;
 191
 192                 cond_resched();