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

   1 From 8c20e2eb5f2a0175b774134685e4d7bd93e85ff8 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 01/19] UPSTREAM: mm: multi-gen LRU: rename lru_gen_struct to
   5  lru_gen_folio
   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 folios (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 folios):
  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_folio 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 Bug: 274865848
 119 (cherry picked from commit 391655fe08d1f942359a11148aa9aaf3f99d6d6f)
 120 Change-Id: I7df67e0e2435ba28f10eaa57d28d98b61a9210a6
 121 Signed-off-by: T.J. Mercier <tjmercier@google.com>
 122 ---
 123  include/linux/mm_inline.h |  4 ++--
 124  include/linux/mmzone.h    |  6 +++---
 125  mm/vmscan.c               | 34 +++++++++++++++++-----------------
 126  mm/workingset.c           |  4 ++--
 127  4 files changed, 24 insertions(+), 24 deletions(-)
 128
 129 diff --git a/include/linux/mm_inline.h b/include/linux/mm_inline.h
 130 index e8ed225d8f7ca..f63968bd7de59 100644
 131 --- a/include/linux/mm_inline.h
 132 +++ b/include/linux/mm_inline.h
 133 @@ -178,7 +178,7 @@ static inline void lru_gen_update_size(struct lruvec *lruvec, struct folio *foli
 134         int zone = folio_zonenum(folio);
 135         int delta = folio_nr_pages(folio);
 136         enum lru_list lru = type * LRU_INACTIVE_FILE;
 137 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 138 +       struct lru_gen_folio *lrugen = &lruvec->lrugen;
 139
 140         VM_WARN_ON_ONCE(old_gen != -1 && old_gen >= MAX_NR_GENS);
 141         VM_WARN_ON_ONCE(new_gen != -1 && new_gen >= MAX_NR_GENS);
 142 @@ -224,7 +224,7 @@ static inline bool lru_gen_add_folio(struct lruvec *lruvec, struct folio *folio,
 143         int gen = folio_lru_gen(folio);
 144         int type = folio_is_file_lru(folio);
 145         int zone = folio_zonenum(folio);
 146 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 147 +       struct lru_gen_folio *lrugen = &lruvec->lrugen;
 148
 149         VM_WARN_ON_ONCE_FOLIO(gen != -1, folio);
 150
 151 diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
 152 index 5f74891556f33..bd3e4689f72dc 100644
 153 --- a/include/linux/mmzone.h
 154 +++ b/include/linux/mmzone.h
 155 @@ -404,7 +404,7 @@ enum {
 156   * The number of pages in each generation is eventually consistent and therefore
 157   * can be transiently negative when reset_batch_size() is pending.
 158   */
 159 -struct lru_gen_struct {
 160 +struct lru_gen_folio {
 161         /* the aging increments the youngest generation number */
 162         unsigned long max_seq;
 163         /* the eviction increments the oldest generation numbers */
 164 @@ -461,7 +461,7 @@ struct lru_gen_mm_state {
 165  struct lru_gen_mm_walk {
 166         /* the lruvec under reclaim */
 167         struct lruvec *lruvec;
 168 -       /* unstable max_seq from lru_gen_struct */
 169 +       /* unstable max_seq from lru_gen_folio */
 170         unsigned long max_seq;
 171         /* the next address within an mm to scan */
 172         unsigned long next_addr;
 173 @@ -524,7 +524,7 @@ struct lruvec {
 174         unsigned long                   flags;
 175  #ifdef CONFIG_LRU_GEN
 176         /* evictable pages divided into generations */
 177 -       struct lru_gen_struct           lrugen;
 178 +       struct lru_gen_folio            lrugen;
 179         /* to concurrently iterate lru_gen_mm_list */
 180         struct lru_gen_mm_state         mm_state;
 181  #endif
 182 diff --git a/mm/vmscan.c b/mm/vmscan.c
 183 index d18296109aa7e..27142caf284c1 100644
 184 --- a/mm/vmscan.c
 185 +++ b/mm/vmscan.c
 186 @@ -3190,7 +3190,7 @@ static int get_nr_gens(struct lruvec *lruvec, int type)
 187
 188  static bool __maybe_unused seq_is_valid(struct lruvec *lruvec)
 189  {
 190 -       /* see the comment on lru_gen_struct */
 191 +       /* see the comment on lru_gen_folio */
 192         return get_nr_gens(lruvec, LRU_GEN_FILE) >= MIN_NR_GENS &&
 193                get_nr_gens(lruvec, LRU_GEN_FILE) <= get_nr_gens(lruvec, LRU_GEN_ANON) &&
 194                get_nr_gens(lruvec, LRU_GEN_ANON) <= MAX_NR_GENS;
 195 @@ -3596,7 +3596,7 @@ struct ctrl_pos {
 196  static void read_ctrl_pos(struct lruvec *lruvec, int type, int tier, int gain,
 197                           struct ctrl_pos *pos)
 198  {
 199 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 200 +       struct lru_gen_folio *lrugen = &lruvec->lrugen;
 201         int hist = lru_hist_from_seq(lrugen->min_seq[type]);
 202
 203         pos->refaulted = lrugen->avg_refaulted[type][tier] +
 204 @@ -3611,7 +3611,7 @@ static void read_ctrl_pos(struct lruvec *lruvec, int type, int tier, int gain,
 205  static void reset_ctrl_pos(struct lruvec *lruvec, int type, bool carryover)
 206  {
 207         int hist, tier;
 208 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 209 +       struct lru_gen_folio *lrugen = &lruvec->lrugen;
 210         bool clear = carryover ? NR_HIST_GENS == 1 : NR_HIST_GENS > 1;
 211         unsigned long seq = carryover ? lrugen->min_seq[type] : lrugen->max_seq + 1;
 212
 213 @@ -3688,7 +3688,7 @@ static int folio_update_gen(struct folio *folio, int gen)
 214  static int folio_inc_gen(struct lruvec *lruvec, struct folio *folio, bool reclaiming)
 215  {
 216         int type = folio_is_file_lru(folio);
 217 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 218 +       struct lru_gen_folio *lrugen = &lruvec->lrugen;
 219         int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]);
 220         unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
 221
 222 @@ -3733,7 +3733,7 @@ static void update_batch_size(struct lru_gen_mm_walk *walk, struct folio *folio,
 223  static void reset_batch_size(struct lruvec *lruvec, struct lru_gen_mm_walk *walk)
 224  {
 225         int gen, type, zone;
 226 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 227 +       struct lru_gen_folio *lrugen = &lruvec->lrugen;
 228
 229         walk->batched = 0;
 230
 231 @@ -4250,7 +4250,7 @@ static bool inc_min_seq(struct lruvec *lruvec, int type, bool can_swap)
 232  {
 233         int zone;
 234         int remaining = MAX_LRU_BATCH;
 235 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 236 +       struct lru_gen_folio *lrugen = &lruvec->lrugen;
 237         int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]);
 238
 239         if (type == LRU_GEN_ANON && !can_swap)
 240 @@ -4286,7 +4286,7 @@ static bool try_to_inc_min_seq(struct lruvec *lruvec, bool can_swap)
 241  {
 242         int gen, type, zone;
 243         bool success = false;
 244 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 245 +       struct lru_gen_folio *lrugen = &lruvec->lrugen;
 246         DEFINE_MIN_SEQ(lruvec);
 247
 248         VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
 249 @@ -4307,7 +4307,7 @@ static bool try_to_inc_min_seq(struct lruvec *lruvec, bool can_swap)
 250                 ;
 251         }
 252
 253 -       /* see the comment on lru_gen_struct */
 254 +       /* see the comment on lru_gen_folio */
 255         if (can_swap) {
 256                 min_seq[LRU_GEN_ANON] = min(min_seq[LRU_GEN_ANON], min_seq[LRU_GEN_FILE]);
 257                 min_seq[LRU_GEN_FILE] = max(min_seq[LRU_GEN_ANON], lrugen->min_seq[LRU_GEN_FILE]);
 258 @@ -4329,7 +4329,7 @@ static void inc_max_seq(struct lruvec *lruvec, bool can_swap, bool force_scan)
 259  {
 260         int prev, next;
 261         int type, zone;
 262 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 263 +       struct lru_gen_folio *lrugen = &lruvec->lrugen;
 264
 265         spin_lock_irq(&lruvec->lru_lock);
 266
 267 @@ -4387,7 +4387,7 @@ static bool try_to_inc_max_seq(struct lruvec *lruvec, unsigned long max_seq,
 268         bool success;
 269         struct lru_gen_mm_walk *walk;
 270         struct mm_struct *mm = NULL;
 271 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 272 +       struct lru_gen_folio *lrugen = &lruvec->lrugen;
 273
 274         VM_WARN_ON_ONCE(max_seq > READ_ONCE(lrugen->max_seq));
 275
 276 @@ -4452,7 +4452,7 @@ static bool should_run_aging(struct lruvec *lruvec, unsigned long max_seq, unsig
 277         unsigned long old = 0;
 278         unsigned long young = 0;
 279         unsigned long total = 0;
 280 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 281 +       struct lru_gen_folio *lrugen = &lruvec->lrugen;
 282         struct mem_cgroup *memcg = lruvec_memcg(lruvec);
 283
 284         for (type = !can_swap; type < ANON_AND_FILE; type++) {
 285 @@ -4737,7 +4737,7 @@ static bool sort_folio(struct lruvec *lruvec, struct folio *folio, int tier_idx)
 286         int delta = folio_nr_pages(folio);
 287         int refs = folio_lru_refs(folio);
 288         int tier = lru_tier_from_refs(refs);
 289 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 290 +       struct lru_gen_folio *lrugen = &lruvec->lrugen;
 291
 292         VM_WARN_ON_ONCE_FOLIO(gen >= MAX_NR_GENS, folio);
 293
 294 @@ -4837,7 +4837,7 @@ static int scan_folios(struct lruvec *lruvec, struct scan_control *sc,
 295         int scanned = 0;
 296         int isolated = 0;
 297         int remaining = MAX_LRU_BATCH;
 298 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 299 +       struct lru_gen_folio *lrugen = &lruvec->lrugen;
 300         struct mem_cgroup *memcg = lruvec_memcg(lruvec);
 301
 302         VM_WARN_ON_ONCE(!list_empty(list));
 303 @@ -5237,7 +5237,7 @@ static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc
 304
 305  static bool __maybe_unused state_is_valid(struct lruvec *lruvec)
 306  {
 307 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 308 +       struct lru_gen_folio *lrugen = &lruvec->lrugen;
 309
 310         if (lrugen->enabled) {
 311                 enum lru_list lru;
 312 @@ -5519,7 +5519,7 @@ static void lru_gen_seq_show_full(struct seq_file *m, struct lruvec *lruvec,
 313         int i;
 314         int type, tier;
 315         int hist = lru_hist_from_seq(seq);
 316 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 317 +       struct lru_gen_folio *lrugen = &lruvec->lrugen;
 318
 319         for (tier = 0; tier < MAX_NR_TIERS; tier++) {
 320                 seq_printf(m, "            %10d", tier);
 321 @@ -5569,7 +5569,7 @@ static int lru_gen_seq_show(struct seq_file *m, void *v)
 322         unsigned long seq;
 323         bool full = !debugfs_real_fops(m->file)->write;
 324         struct lruvec *lruvec = v;
 325 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 326 +       struct lru_gen_folio *lrugen = &lruvec->lrugen;
 327         int nid = lruvec_pgdat(lruvec)->node_id;
 328         struct mem_cgroup *memcg = lruvec_memcg(lruvec);
 329         DEFINE_MAX_SEQ(lruvec);
 330 @@ -5823,7 +5823,7 @@ void lru_gen_init_lruvec(struct lruvec *lruvec)
 331  {
 332         int i;
 333         int gen, type, zone;
 334 -       struct lru_gen_struct *lrugen = &lruvec->lrugen;
 335 +       struct lru_gen_folio *lrugen = &lruvec->lrugen;
 336
 337         lrugen->max_seq = MIN_NR_GENS + 1;
 338         lrugen->enabled = lru_gen_enabled();
 339 diff --git a/mm/workingset.c b/mm/workingset.c
 340 index ae7e984b23c6b..688aaa73f64e8 100644
 341 --- a/mm/workingset.c
 342 +++ b/mm/workingset.c
 343 @@ -223,7 +223,7 @@ static void *lru_gen_eviction(struct folio *folio)
 344         unsigned long token;
 345         unsigned long min_seq;
 346         struct lruvec *lruvec;
 347 -       struct lru_gen_struct *lrugen;
 348 +       struct lru_gen_folio *lrugen;
 349         int type = folio_is_file_lru(folio);
 350         int delta = folio_nr_pages(folio);
 351         int refs = folio_lru_refs(folio);
 352 @@ -252,7 +252,7 @@ static void lru_gen_refault(struct folio *folio, void *shadow)
 353         unsigned long token;
 354         unsigned long min_seq;
 355         struct lruvec *lruvec;
 356 -       struct lru_gen_struct *lrugen;
 357 +       struct lru_gen_folio *lrugen;
 358         struct mem_cgroup *memcg;
 359         struct pglist_data *pgdat;
 360         int type = folio_is_file_lru(folio);
 361 --
 362 2.40.1
 363