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

   1 From 348fdbada9fb3f0bf1a53651be46319105af187f Mon Sep 17 00:00:00 2001
   2 From: Yu Zhao <yuzhao@google.com>
   3 Date: Wed, 21 Dec 2022 21:18:59 -0700
   4 Subject: [PATCH 21/29] mm: multi-gen LRU: rename lru_gen_struct to
   5  lru_gen_page
   6
   7 Patch series "mm: multi-gen LRU: memcg LRU", v3.
   8
   9 Overview
  10 ========
  11
  12 An memcg LRU is a per-node LRU of memcgs.  It is also an LRU of LRUs,
  13 since each node and memcg combination has an LRU of pages (see
  14 mem_cgroup_lruvec()).
  15
  16 Its goal is to improve the scalability of global reclaim, which is
  17 critical to system-wide memory overcommit in data centers.  Note that
  18 memcg reclaim is currently out of scope.
  19
  20 Its memory bloat is a pointer to each lruvec and negligible to each
  21 pglist_data.  In terms of traversing memcgs during global reclaim, it
  22 improves the best-case complexity from O(n) to O(1) and does not affect
  23 the worst-case complexity O(n).  Therefore, on average, it has a sublinear
  24 complexity in contrast to the current linear complexity.
  25
  26 The basic structure of an memcg LRU can be understood by an analogy to
  27 the active/inactive LRU (of pages):
  28 1. It has the young and the old (generations), i.e., the counterparts
  29    to the active and the inactive;
  30 2. The increment of max_seq triggers promotion, i.e., the counterpart
  31    to activation;
  32 3. Other events trigger similar operations, e.g., offlining an memcg
  33    triggers demotion, i.e., the counterpart to deactivation.
  34
  35 In terms of global reclaim, it has two distinct features:
  36 1. Sharding, which allows each thread to start at a random memcg (in
  37    the old generation) and improves parallelism;
  38 2. Eventual fairness, which allows direct reclaim to bail out at will
  39    and reduces latency without affecting fairness over some time.
  40
  41 The commit message in patch 6 details the workflow:
  42 https://lore.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com/
  43
  44 The following is a simple test to quickly verify its effectiveness.
  45
  46   Test design:
  47   1. Create multiple memcgs.
  48   2. Each memcg contains a job (fio).
  49   3. All jobs access the same amount of memory randomly.
  50   4. The system does not experience global memory pressure.
  51   5. Periodically write to the root memory.reclaim.
  52
  53   Desired outcome:
  54   1. All memcgs have similar pgsteal counts, i.e., stddev(pgsteal)
  55      over mean(pgsteal) is close to 0%.
  56   2. The total pgsteal is close to the total requested through
  57      memory.reclaim, i.e., sum(pgsteal) over sum(requested) is close
  58      to 100%.
  59
  60   Actual outcome [1]:
  61                                      MGLRU off    MGLRU on
  62   stddev(pgsteal) / mean(pgsteal)    75%          20%
  63   sum(pgsteal) / sum(requested)      425%         95%
  64
  65   ####################################################################
  66   MEMCGS=128
  67
  68   for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
  69       mkdir /sys/fs/cgroup/memcg$memcg
  70   done
  71
  72   start() {
  73       echo $BASHPID > /sys/fs/cgroup/memcg$memcg/cgroup.procs
  74
  75       fio -name=memcg$memcg --numjobs=1 --ioengine=mmap \
  76           --filename=/dev/zero --size=1920M --rw=randrw \
  77           --rate=64m,64m --random_distribution=random \
  78           --fadvise_hint=0 --time_based --runtime=10h \
  79           --group_reporting --minimal
  80   }
  81
  82   for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
  83       start &
  84   done
  85
  86   sleep 600
  87
  88   for ((i = 0; i < 600; i++)); do
  89       echo 256m >/sys/fs/cgroup/memory.reclaim
  90       sleep 6
  91   done
  92
  93   for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
  94       grep "pgsteal " /sys/fs/cgroup/memcg$memcg/memory.stat
  95   done
  96   ####################################################################
  97
  98 [1]: This was obtained from running the above script (touches less
  99      than 256GB memory) on an EPYC 7B13 with 512GB DRAM for over an
 100      hour.
 101
 102 This patch (of 8):
 103
 104 The new name lru_gen_page will be more distinct from the coming
 105 lru_gen_memcg.
 106
 107 Link: https://lkml.kernel.org/r/20221222041905.2431096-1-yuzhao@google.com
 108 Link: https://lkml.kernel.org/r/20221222041905.2431096-2-yuzhao@google.com
 109 Signed-off-by: Yu Zhao <yuzhao@google.com>
 110 Cc: Johannes Weiner <hannes@cmpxchg.org>
 111 Cc: Jonathan Corbet <corbet@lwn.net>
 112 Cc: Michael Larabel <Michael@MichaelLarabel.com>
 113 Cc: Michal Hocko <mhocko@kernel.org>
 114 Cc: Mike Rapoport <rppt@kernel.org>
 115 Cc: Roman Gushchin <roman.gushchin@linux.dev>
 116 Cc: Suren Baghdasaryan <surenb@google.com>
 117 Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
 118 ---
 119  include/linux/mm_inline.h |  4 ++--
 120  include/linux/mmzone.h    |  6 +++---
 121  mm/vmscan.c               | 34 +++++++++++++++++-----------------
 122  mm/workingset.c           |  4 ++--
 123  4 files changed, 24 insertions(+), 24 deletions(-)
 124
 125 diff --git a/include/linux/mm_inline.h b/include/linux/mm_inline.h
 126 index 8a6a2a23f9b6..27c4890503c5 100644
 127 --- a/include/linux/mm_inline.h
 128 +++ b/include/linux/mm_inline.h
 129 @@ -168,7 +168,7 @@ static inline void lru_gen_update_size(struct lruvec *lruvec, struct page *page,
 130         int zone = page_zonenum(page);
 131         int delta = thp_nr_pages(page);
 132         enum lru_list lru = type * LRU_INACTIVE_FILE;
 133 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 134 +       struct lru_gen_page *lrugen = &lruvec->lrugen;
 135
 136         VM_WARN_ON_ONCE(old_gen != -1 && old_gen >= MAX_NR_GENS);
 137         VM_WARN_ON_ONCE(new_gen != -1 && new_gen >= MAX_NR_GENS);
 138 @@ -214,7 +214,7 @@ static inline bool lru_gen_add_page(struct lruvec *lruvec, struct page *page, bo
 139         int gen = page_lru_gen(page);
 140         int type = page_is_file_lru(page);
 141         int zone = page_zonenum(page);
 142 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 143 +       struct lru_gen_page *lrugen = &lruvec->lrugen;
 144
 145         VM_WARN_ON_ONCE_PAGE(gen != -1, page);
 146
 147 diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
 148 index 6b85ba1f4e18..5856b026c089 100644
 149 --- a/include/linux/mmzone.h
 150 +++ b/include/linux/mmzone.h
 151 @@ -394,7 +394,7 @@ enum {
 152   * The number of pages in each generation is eventually consistent and therefore
 153   * can be transiently negative when reset_batch_size() is pending.
 154   */
 155 -struct lru_gen_struct {
 156 +struct lru_gen_page {
 157         /* the aging increments the youngest generation number */
 158         unsigned long max_seq;
 159         /* the eviction increments the oldest generation numbers */
 160 @@ -451,7 +451,7 @@ struct lru_gen_mm_state {
 161  struct lru_gen_mm_walk {
 162         /* the lruvec under reclaim */
 163         struct lruvec *lruvec;
 164 -       /* unstable max_seq from lru_gen_struct */
 165 +       /* unstable max_seq from lru_gen_page */
 166         unsigned long max_seq;
 167         /* the next address within an mm to scan */
 168         unsigned long next_addr;
 169 @@ -514,7 +514,7 @@ struct lruvec {
 170         unsigned long                   flags;
 171  #ifdef CONFIG_LRU_GEN
 172         /* evictable pages divided into generations */
 173 -       struct lru_gen_struct           lrugen;
 174 +       struct lru_gen_page             lrugen;
 175         /* to concurrently iterate lru_gen_mm_list */
 176         struct lru_gen_mm_state         mm_state;
 177  #endif
 178 diff --git a/mm/vmscan.c b/mm/vmscan.c
 179 index 4ab376abeaae..3b1b5bd9736a 100644
 180 --- a/mm/vmscan.c
 181 +++ b/mm/vmscan.c
 182 @@ -2910,7 +2910,7 @@ static int get_nr_gens(struct lruvec *lruvec, int type)
 183
 184  static bool __maybe_unused seq_is_valid(struct lruvec *lruvec)
 185  {
 186 -       /* see the comment on lru_gen_struct */
 187 +       /* see the comment on lru_gen_page */
 188         return get_nr_gens(lruvec, LRU_GEN_FILE) >= MIN_NR_GENS &&
 189                get_nr_gens(lruvec, LRU_GEN_FILE) <= get_nr_gens(lruvec, LRU_GEN_ANON) &&
 190                get_nr_gens(lruvec, LRU_GEN_ANON) <= MAX_NR_GENS;
 191 @@ -3316,7 +3316,7 @@ struct ctrl_pos {
 192  static void read_ctrl_pos(struct lruvec *lruvec, int type, int tier, int gain,
 193                           struct ctrl_pos *pos)
 194  {
 195 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 196 +       struct lru_gen_page *lrugen = &lruvec->lrugen;
 197         int hist = lru_hist_from_seq(lrugen->min_seq[type]);
 198
 199         pos->refaulted = lrugen->avg_refaulted[type][tier] +
 200 @@ -3331,7 +3331,7 @@ static void read_ctrl_pos(struct lruvec *lruvec, int type, int tier, int gain,
 201  static void reset_ctrl_pos(struct lruvec *lruvec, int type, bool carryover)
 202  {
 203         int hist, tier;
 204 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 205 +       struct lru_gen_page *lrugen = &lruvec->lrugen;
 206         bool clear = carryover ? NR_HIST_GENS == 1 : NR_HIST_GENS > 1;
 207         unsigned long seq = carryover ? lrugen->min_seq[type] : lrugen->max_seq + 1;
 208
 209 @@ -3408,7 +3408,7 @@ static int page_update_gen(struct page *page, int gen)
 210  static int page_inc_gen(struct lruvec *lruvec, struct page *page, bool reclaiming)
 211  {
 212         int type = page_is_file_lru(page);
 213 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 214 +       struct lru_gen_page *lrugen = &lruvec->lrugen;
 215         int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]);
 216         unsigned long new_flags, old_flags = READ_ONCE(page->flags);
 217
 218 @@ -3453,7 +3453,7 @@ static void update_batch_size(struct lru_gen_mm_walk *walk, struct page *page,
 219  static void reset_batch_size(struct lruvec *lruvec, struct lru_gen_mm_walk *walk)
 220  {
 221         int gen, type, zone;
 222 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 223 +       struct lru_gen_page *lrugen = &lruvec->lrugen;
 224
 225         walk->batched = 0;
 226
 227 @@ -3979,7 +3979,7 @@ static bool inc_min_seq(struct lruvec *lruvec, int type, bool can_swap)
 228  {
 229         int zone;
 230         int remaining = MAX_LRU_BATCH;
 231 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 232 +       struct lru_gen_page *lrugen = &lruvec->lrugen;
 233         int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]);
 234
 235         if (type == LRU_GEN_ANON && !can_swap)
 236 @@ -4015,7 +4015,7 @@ static bool try_to_inc_min_seq(struct lruvec *lruvec, bool can_swap)
 237  {
 238         int gen, type, zone;
 239         bool success = false;
 240 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 241 +       struct lru_gen_page *lrugen = &lruvec->lrugen;
 242         DEFINE_MIN_SEQ(lruvec);
 243
 244         VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
 245 @@ -4036,7 +4036,7 @@ static bool try_to_inc_min_seq(struct lruvec *lruvec, bool can_swap)
 246                 ;
 247         }
 248
 249 -       /* see the comment on lru_gen_struct */
 250 +       /* see the comment on lru_gen_page */
 251         if (can_swap) {
 252                 min_seq[LRU_GEN_ANON] = min(min_seq[LRU_GEN_ANON], min_seq[LRU_GEN_FILE]);
 253                 min_seq[LRU_GEN_FILE] = max(min_seq[LRU_GEN_ANON], lrugen->min_seq[LRU_GEN_FILE]);
 254 @@ -4058,7 +4058,7 @@ static void inc_max_seq(struct lruvec *lruvec, bool can_swap, bool force_scan)
 255  {
 256         int prev, next;
 257         int type, zone;
 258 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 259 +       struct lru_gen_page *lrugen = &lruvec->lrugen;
 260
 261         spin_lock_irq(&lruvec->lru_lock);
 262
 263 @@ -4116,7 +4116,7 @@ static bool try_to_inc_max_seq(struct lruvec *lruvec, unsigned long max_seq,
 264         bool success;
 265         struct lru_gen_mm_walk *walk;
 266         struct mm_struct *mm = NULL;
 267 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 268 +       struct lru_gen_page *lrugen = &lruvec->lrugen;
 269
 270         VM_WARN_ON_ONCE(max_seq > READ_ONCE(lrugen->max_seq));
 271
 272 @@ -4181,7 +4181,7 @@ static bool should_run_aging(struct lruvec *lruvec, unsigned long max_seq, unsig
 273         unsigned long old = 0;
 274         unsigned long young = 0;
 275         unsigned long total = 0;
 276 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 277 +       struct lru_gen_page *lrugen = &lruvec->lrugen;
 278         struct mem_cgroup *memcg = lruvec_memcg(lruvec);
 279
 280         for (type = !can_swap; type < ANON_AND_FILE; type++) {
 281 @@ -4466,7 +4466,7 @@ static bool sort_page(struct lruvec *lruvec, struct page *page, int tier_idx)
 282         int delta = thp_nr_pages(page);
 283         int refs = page_lru_refs(page);
 284         int tier = lru_tier_from_refs(refs);
 285 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 286 +       struct lru_gen_page *lrugen = &lruvec->lrugen;
 287
 288         VM_WARN_ON_ONCE_PAGE(gen >= MAX_NR_GENS, page);
 289
 290 @@ -4566,7 +4566,7 @@ static int scan_pages(struct lruvec *lruvec, struct scan_control *sc,
 291         int scanned = 0;
 292         int isolated = 0;
 293         int remaining = MAX_LRU_BATCH;
 294 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 295 +       struct lru_gen_page *lrugen = &lruvec->lrugen;
 296         struct mem_cgroup *memcg = lruvec_memcg(lruvec);
 297
 298         VM_WARN_ON_ONCE(!list_empty(list));
 299 @@ -4967,7 +4967,7 @@ static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc
 300
 301  static bool __maybe_unused state_is_valid(struct lruvec *lruvec)
 302  {
 303 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 304 +       struct lru_gen_page *lrugen = &lruvec->lrugen;
 305
 306         if (lrugen->enabled) {
 307                 enum lru_list lru;
 308 @@ -5247,7 +5247,7 @@ static void lru_gen_seq_show_full(struct seq_file *m, struct lruvec *lruvec,
 309         int i;
 310         int type, tier;
 311         int hist = lru_hist_from_seq(seq);
 312 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 313 +       struct lru_gen_page *lrugen = &lruvec->lrugen;
 314
 315         for (tier = 0; tier < MAX_NR_TIERS; tier++) {
 316                 seq_printf(m, "            %10d", tier);
 317 @@ -5296,7 +5296,7 @@ static int lru_gen_seq_show(struct seq_file *m, void *v)
 318         unsigned long seq;
 319         bool full = !debugfs_real_fops(m->file)->write;
 320         struct lruvec *lruvec = v;
 321 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 322 +       struct lru_gen_page *lrugen = &lruvec->lrugen;
 323         int nid = lruvec_pgdat(lruvec)->node_id;
 324         struct mem_cgroup *memcg = lruvec_memcg(lruvec);
 325         DEFINE_MAX_SEQ(lruvec);
 326 @@ -5549,7 +5549,7 @@ void lru_gen_init_lruvec(struct lruvec *lruvec)
 327  {
 328         int i;
 329         int gen, type, zone;
 330 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 331 +       struct lru_gen_page *lrugen = &lruvec->lrugen;
 332
 333         lrugen->max_seq = MIN_NR_GENS + 1;
 334         lrugen->enabled = lru_gen_enabled();
 335 diff --git a/mm/workingset.c b/mm/workingset.c
 336 index aeba62cebf8c..a5e1798c6d60 100644
 337 --- a/mm/workingset.c
 338 +++ b/mm/workingset.c
 339 @@ -223,7 +223,7 @@ static void *lru_gen_eviction(struct page *page)
 340         unsigned long token;
 341         unsigned long min_seq;
 342         struct lruvec *lruvec;
 343 -       struct lru_gen_struct *lrugen;
 344 +       struct lru_gen_page *lrugen;
 345         int type = page_is_file_lru(page);
 346         int delta = thp_nr_pages(page);
 347         int refs = page_lru_refs(page);
 348 @@ -252,7 +252,7 @@ static void lru_gen_refault(struct page *page, void *shadow)
 349         unsigned long token;
 350         unsigned long min_seq;
 351         struct lruvec *lruvec;
 352 -       struct lru_gen_struct *lrugen;
 353 +       struct lru_gen_page *lrugen;
 354         struct mem_cgroup *memcg;
 355         struct pglist_data *pgdat;
 356         int type = page_is_file_lru(page);
 357 --
 358 2.40.0
 359