target/linux/generic/backport-5.15/020-v6.1-11-mm-multi-gen-LRU-thrashing-prevention.patch

   1 From 73d1ff551760f0c79c47ab70faa4c2ca91413f5c Mon Sep 17 00:00:00 2001
   2 From: Yu Zhao <yuzhao@google.com>
   3 Date: Sun, 18 Sep 2022 02:00:08 -0600
   4 Subject: [PATCH 11/29] mm: multi-gen LRU: thrashing prevention
   5 MIME-Version: 1.0
   6 Content-Type: text/plain; charset=UTF-8
   7 Content-Transfer-Encoding: 8bit
   8
   9 Add /sys/kernel/mm/lru_gen/min_ttl_ms for thrashing prevention, as
  10 requested by many desktop users [1].
  11
  12 When set to value N, it prevents the working set of N milliseconds from
  13 getting evicted.  The OOM killer is triggered if this working set cannot
  14 be kept in memory.  Based on the average human detectable lag (~100ms),
  15 N=1000 usually eliminates intolerable lags due to thrashing.  Larger
  16 values like N=3000 make lags less noticeable at the risk of premature OOM
  17 kills.
  18
  19 Compared with the size-based approach [2], this time-based approach
  20 has the following advantages:
  21
  22 1. It is easier to configure because it is agnostic to applications
  23    and memory sizes.
  24 2. It is more reliable because it is directly wired to the OOM killer.
  25
  26 [1] https://lore.kernel.org/r/Ydza%2FzXKY9ATRoh6@google.com/
  27 [2] https://lore.kernel.org/r/20101028191523.GA14972@google.com/
  28
  29 Link: https://lkml.kernel.org/r/20220918080010.2920238-12-yuzhao@google.com
  30 Signed-off-by: Yu Zhao <yuzhao@google.com>
  31 Acked-by: Brian Geffon <bgeffon@google.com>
  32 Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
  33 Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
  34 Acked-by: Steven Barrett <steven@liquorix.net>
  35 Acked-by: Suleiman Souhlal <suleiman@google.com>
  36 Tested-by: Daniel Byrne <djbyrne@mtu.edu>
  37 Tested-by: Donald Carr <d@chaos-reins.com>
  38 Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
  39 Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
  40 Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
  41 Tested-by: Sofia Trinh <sofia.trinh@edi.works>
  42 Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
  43 Cc: Andi Kleen <ak@linux.intel.com>
  44 Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
  45 Cc: Barry Song <baohua@kernel.org>
  46 Cc: Catalin Marinas <catalin.marinas@arm.com>
  47 Cc: Dave Hansen <dave.hansen@linux.intel.com>
  48 Cc: Hillf Danton <hdanton@sina.com>
  49 Cc: Jens Axboe <axboe@kernel.dk>
  50 Cc: Johannes Weiner <hannes@cmpxchg.org>
  51 Cc: Jonathan Corbet <corbet@lwn.net>
  52 Cc: Linus Torvalds <torvalds@linux-foundation.org>
  53 Cc: Matthew Wilcox <willy@infradead.org>
  54 Cc: Mel Gorman <mgorman@suse.de>
  55 Cc: Miaohe Lin <linmiaohe@huawei.com>
  56 Cc: Michael Larabel <Michael@MichaelLarabel.com>
  57 Cc: Michal Hocko <mhocko@kernel.org>
  58 Cc: Mike Rapoport <rppt@kernel.org>
  59 Cc: Mike Rapoport <rppt@linux.ibm.com>
  60 Cc: Peter Zijlstra <peterz@infradead.org>
  61 Cc: Qi Zheng <zhengqi.arch@bytedance.com>
  62 Cc: Tejun Heo <tj@kernel.org>
  63 Cc: Vlastimil Babka <vbabka@suse.cz>
  64 Cc: Will Deacon <will@kernel.org>
  65 Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
  66 ---
  67  include/linux/mmzone.h |  2 ++
  68  mm/vmscan.c            | 74 ++++++++++++++++++++++++++++++++++++++++--
  69  2 files changed, 73 insertions(+), 3 deletions(-)
  70
  71 diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
  72 index edaf035503ed..6b85ba1f4e18 100644
  73 --- a/include/linux/mmzone.h
  74 +++ b/include/linux/mmzone.h
  75 @@ -399,6 +399,8 @@ struct lru_gen_struct {
  76         unsigned long max_seq;
  77         /* the eviction increments the oldest generation numbers */
  78         unsigned long min_seq[ANON_AND_FILE];
  79 +       /* the birth time of each generation in jiffies */
  80 +       unsigned long timestamps[MAX_NR_GENS];
  81         /* the multi-gen LRU lists, lazily sorted on eviction */
  82         struct list_head lists[MAX_NR_GENS][ANON_AND_FILE][MAX_NR_ZONES];
  83         /* the multi-gen LRU sizes, eventually consistent */
  84 diff --git a/mm/vmscan.c b/mm/vmscan.c
  85 index be37d996bc92..642ee7bef61d 100644
  86 --- a/mm/vmscan.c
  87 +++ b/mm/vmscan.c
  88 @@ -4064,6 +4064,7 @@ static void inc_max_seq(struct lruvec *lruvec, bool can_swap)
  89         for (type = 0; type < ANON_AND_FILE; type++)
  90                 reset_ctrl_pos(lruvec, type, false);
  91
  92 +       WRITE_ONCE(lrugen->timestamps[next], jiffies);
  93         /* make sure preceding modifications appear */
  94         smp_store_release(&lrugen->max_seq, lrugen->max_seq + 1);
  95
  96 @@ -4193,7 +4194,7 @@ static bool should_run_aging(struct lruvec *lruvec, unsigned long max_seq, unsig
  97         return false;
  98  }
  99
 100 -static void age_lruvec(struct lruvec *lruvec, struct scan_control *sc)
 101 +static bool age_lruvec(struct lruvec *lruvec, struct scan_control *sc, unsigned long min_ttl)
 102  {
 103         bool need_aging;
 104         unsigned long nr_to_scan;
 105 @@ -4207,16 +4208,36 @@ static void age_lruvec(struct lruvec *lruvec, struct scan_control *sc)
 106         mem_cgroup_calculate_protection(NULL, memcg);
 107
 108         if (mem_cgroup_below_min(memcg))
 109 -               return;
 110 +               return false;
 111
 112         need_aging = should_run_aging(lruvec, max_seq, min_seq, sc, swappiness, &nr_to_scan);
 113 +
 114 +       if (min_ttl) {
 115 +               int gen = lru_gen_from_seq(min_seq[LRU_GEN_FILE]);
 116 +               unsigned long birth = READ_ONCE(lruvec->lrugen.timestamps[gen]);
 117 +
 118 +               if (time_is_after_jiffies(birth + min_ttl))
 119 +                       return false;
 120 +
 121 +               /* the size is likely too small to be helpful */
 122 +               if (!nr_to_scan && sc->priority != DEF_PRIORITY)
 123 +                       return false;
 124 +       }
 125 +
 126         if (need_aging)
 127                 try_to_inc_max_seq(lruvec, max_seq, sc, swappiness);
 128 +
 129 +       return true;
 130  }
 131
 132 +/* to protect the working set of the last N jiffies */
 133 +static unsigned long lru_gen_min_ttl __read_mostly;
 134 +
 135  static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
 136  {
 137         struct mem_cgroup *memcg;
 138 +       bool success = false;
 139 +       unsigned long min_ttl = READ_ONCE(lru_gen_min_ttl);
 140
 141         VM_WARN_ON_ONCE(!current_is_kswapd());
 142
 143 @@ -4239,12 +4260,32 @@ static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
 144         do {
 145                 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
 146
 147 -               age_lruvec(lruvec, sc);
 148 +               if (age_lruvec(lruvec, sc, min_ttl))
 149 +                       success = true;
 150
 151                 cond_resched();
 152         } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
 153
 154         clear_mm_walk();
 155 +
 156 +       /* check the order to exclude compaction-induced reclaim */
 157 +       if (success || !min_ttl || sc->order)
 158 +               return;
 159 +
 160 +       /*
 161 +        * The main goal is to OOM kill if every generation from all memcgs is
 162 +        * younger than min_ttl. However, another possibility is all memcgs are
 163 +        * either below min or empty.
 164 +        */
 165 +       if (mutex_trylock(&oom_lock)) {
 166 +               struct oom_control oc = {
 167 +                       .gfp_mask = sc->gfp_mask,
 168 +               };
 169 +
 170 +               out_of_memory(&oc);
 171 +
 172 +               mutex_unlock(&oom_lock);
 173 +       }
 174  }
 175
 176  /*
 177 @@ -5002,6 +5043,28 @@ static void lru_gen_change_state(bool enabled)
 178   *                          sysfs interface
 179   ******************************************************************************/
 180
 181 +static ssize_t show_min_ttl(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
 182 +{
 183 +       return sprintf(buf, "%u\n", jiffies_to_msecs(READ_ONCE(lru_gen_min_ttl)));
 184 +}
 185 +
 186 +static ssize_t store_min_ttl(struct kobject *kobj, struct kobj_attribute *attr,
 187 +                            const char *buf, size_t len)
 188 +{
 189 +       unsigned int msecs;
 190 +
 191 +       if (kstrtouint(buf, 0, &msecs))
 192 +               return -EINVAL;
 193 +
 194 +       WRITE_ONCE(lru_gen_min_ttl, msecs_to_jiffies(msecs));
 195 +
 196 +       return len;
 197 +}
 198 +
 199 +static struct kobj_attribute lru_gen_min_ttl_attr = __ATTR(
 200 +       min_ttl_ms, 0644, show_min_ttl, store_min_ttl
 201 +);
 202 +
 203  static ssize_t show_enabled(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
 204  {
 205         unsigned int caps = 0;
 206 @@ -5050,6 +5113,7 @@ static struct kobj_attribute lru_gen_enabled_attr = __ATTR(
 207  );
 208
 209  static struct attribute *lru_gen_attrs[] = {
 210 +       &lru_gen_min_ttl_attr.attr,
 211         &lru_gen_enabled_attr.attr,
 212         NULL
 213  };
 214 @@ -5065,12 +5129,16 @@ static struct attribute_group lru_gen_attr_group = {
 215
 216  void lru_gen_init_lruvec(struct lruvec *lruvec)
 217  {
 218 +       int i;
 219         int gen, type, zone;
 220         struct lru_gen_struct *lrugen = &lruvec->lrugen;
 221
 222         lrugen->max_seq = MIN_NR_GENS + 1;
 223         lrugen->enabled = lru_gen_enabled();
 224
 225 +       for (i = 0; i <= MIN_NR_GENS + 1; i++)
 226 +               lrugen->timestamps[i] = jiffies;
 227 +
 228         for_each_gen_type_zone(gen, type, zone)
 229                 INIT_LIST_HEAD(&lrugen->lists[gen][type][zone]);
 230
 231 --
 232 2.40.0
 233