git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/movinggc.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Moving/copying garbage collector
   4  *
   5  * Copyright 2012 Google, Inc.
   6  */
   7
   8 #include "bcachefs.h"
   9 #include "alloc_foreground.h"
  10 #include "btree_iter.h"
  11 #include "btree_update.h"
  12 #include "buckets.h"
  13 #include "clock.h"
  14 #include "disk_groups.h"
  15 #include "error.h"
  16 #include "extents.h"
  17 #include "eytzinger.h"
  18 #include "io.h"
  19 #include "keylist.h"
  20 #include "move.h"
  21 #include "movinggc.h"
  22 #include "super-io.h"
  23
  24 #include <trace/events/bcachefs.h>
  25 #include <linux/freezer.h>
  26 #include <linux/kthread.h>
  27 #include <linux/math64.h>
  28 #include <linux/sched/task.h>
  29 #include <linux/sort.h>
  30 #include <linux/wait.h>
  31
  32 /*
  33  * We can't use the entire copygc reserve in one iteration of copygc: we may
  34  * need the buckets we're freeing up to go back into the copygc reserve to make
  35  * forward progress, but if the copygc reserve is full they'll be available for
  36  * any allocation - and it's possible that in a given iteration, we free up most
  37  * of the buckets we're going to free before we allocate most of the buckets
  38  * we're going to allocate.
  39  *
  40  * If we only use half of the reserve per iteration, then in steady state we'll
  41  * always have room in the reserve for the buckets we're going to need in the
  42  * next iteration:
  43  */
  44 #define COPYGC_BUCKETS_PER_ITER(ca)                                     \
  45         ((ca)->free[RESERVE_MOVINGGC].size / 2)
  46
  47 static int bucket_offset_cmp(const void *_l, const void *_r, size_t size)
  48 {
  49         const struct copygc_heap_entry *l = _l;
  50         const struct copygc_heap_entry *r = _r;
  51
  52         return  cmp_int(l->dev,    r->dev) ?:
  53                 cmp_int(l->offset, r->offset);
  54 }
  55
  56 static enum data_cmd copygc_pred(struct bch_fs *c, void *arg,
  57                                  struct bkey_s_c k,
  58                                  struct bch_io_opts *io_opts,
  59                                  struct data_opts *data_opts)
  60 {
  61         copygc_heap *h = &c->copygc_heap;
  62         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
  63         const union bch_extent_entry *entry;
  64         struct extent_ptr_decoded p = { 0 };
  65
  66         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
  67                 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
  68                 struct copygc_heap_entry search = {
  69                         .dev    = p.ptr.dev,
  70                         .offset = p.ptr.offset,
  71                 };
  72
  73                 ssize_t i = eytzinger0_find_le(h->data, h->used,
  74                                                sizeof(h->data[0]),
  75                                                bucket_offset_cmp, &search);
  76 #if 0
  77                 /* eytzinger search verify code: */
  78                 ssize_t j = -1, k;
  79
  80                 for (k = 0; k < h->used; k++)
  81                         if (h->data[k].offset <= ptr->offset &&
  82                             (j < 0 || h->data[k].offset > h->data[j].offset))
  83                                 j = k;
  84
  85                 BUG_ON(i != j);
  86 #endif
  87                 if (i >= 0 &&
  88                     p.ptr.offset < h->data[i].offset + ca->mi.bucket_size &&
  89                     p.ptr.gen == h->data[i].gen) {
  90                         /*
  91                          * We need to use the journal reserve here, because
  92                          *  - journal reclaim depends on btree key cache
  93                          *    flushing to make forward progress,
  94                          *  - which has to make forward progress when the
  95                          *    journal is pre-reservation full,
  96                          *  - and depends on allocation - meaning allocator and
  97                          *    copygc
  98                          */
  99
 100                         data_opts->target               = io_opts->background_target;
 101                         data_opts->nr_replicas          = 1;
 102                         data_opts->btree_insert_flags   = BTREE_INSERT_USE_RESERVE|
 103                                 BTREE_INSERT_JOURNAL_RESERVED;
 104                         data_opts->rewrite_dev          = p.ptr.dev;
 105
 106                         if (p.has_ec)
 107                                 data_opts->nr_replicas += p.ec.redundancy;
 108
 109                         return DATA_REWRITE;
 110                 }
 111         }
 112
 113         return DATA_SKIP;
 114 }
 115
 116 static bool have_copygc_reserve(struct bch_dev *ca)
 117 {
 118         bool ret;
 119
 120         spin_lock(&ca->fs->freelist_lock);
 121         ret = fifo_full(&ca->free[RESERVE_MOVINGGC]) ||
 122                 ca->allocator_state != ALLOCATOR_running;
 123         spin_unlock(&ca->fs->freelist_lock);
 124
 125         return ret;
 126 }
 127
 128 static inline int fragmentation_cmp(copygc_heap *heap,
 129                                    struct copygc_heap_entry l,
 130                                    struct copygc_heap_entry r)
 131 {
 132         return cmp_int(l.fragmentation, r.fragmentation);
 133 }
 134
 135 static int bch2_copygc(struct bch_fs *c)
 136 {
 137         copygc_heap *h = &c->copygc_heap;
 138         struct copygc_heap_entry e, *i;
 139         struct bucket_array *buckets;
 140         struct bch_move_stats move_stats;
 141         u64 sectors_to_move = 0, sectors_not_moved = 0;
 142         u64 sectors_reserved = 0;
 143         u64 buckets_to_move, buckets_not_moved = 0;
 144         struct bch_dev *ca;
 145         unsigned dev_idx;
 146         size_t b, heap_size = 0;
 147         int ret;
 148
 149         memset(&move_stats, 0, sizeof(move_stats));
 150         /*
 151          * Find buckets with lowest sector counts, skipping completely
 152          * empty buckets, by building a maxheap sorted by sector count,
 153          * and repeatedly replacing the maximum element until all
 154          * buckets have been visited.
 155          */
 156         h->used = 0;
 157
 158         for_each_rw_member(ca, c, dev_idx)
 159                 heap_size += ca->mi.nbuckets >> 7;
 160
 161         if (h->size < heap_size) {
 162                 free_heap(&c->copygc_heap);
 163                 if (!init_heap(&c->copygc_heap, heap_size, GFP_KERNEL)) {
 164                         bch_err(c, "error allocating copygc heap");
 165                         return 0;
 166                 }
 167         }
 168
 169         for_each_rw_member(ca, c, dev_idx) {
 170                 closure_wait_event(&c->freelist_wait, have_copygc_reserve(ca));
 171
 172                 spin_lock(&ca->fs->freelist_lock);
 173                 sectors_reserved += fifo_used(&ca->free[RESERVE_MOVINGGC]) * ca->mi.bucket_size;
 174                 spin_unlock(&ca->fs->freelist_lock);
 175
 176                 down_read(&ca->bucket_lock);
 177                 buckets = bucket_array(ca);
 178
 179                 for (b = buckets->first_bucket; b < buckets->nbuckets; b++) {
 180                         struct bucket *g = buckets->b + b;
 181                         struct bucket_mark m = READ_ONCE(g->mark);
 182                         struct copygc_heap_entry e;
 183
 184                         if (m.owned_by_allocator ||
 185                             m.data_type != BCH_DATA_user ||
 186                             !bucket_sectors_used(m) ||
 187                             bucket_sectors_used(m) >= ca->mi.bucket_size)
 188                                 continue;
 189
 190                         WARN_ON(m.stripe && !g->stripe_redundancy);
 191
 192                         e = (struct copygc_heap_entry) {
 193                                 .dev            = dev_idx,
 194                                 .gen            = m.gen,
 195                                 .replicas       = 1 + g->stripe_redundancy,
 196                                 .fragmentation  = bucket_sectors_used(m) * (1U << 15)
 197                                         / ca->mi.bucket_size,
 198                                 .sectors        = bucket_sectors_used(m),
 199                                 .offset         = bucket_to_sector(ca, b),
 200                         };
 201                         heap_add_or_replace(h, e, -fragmentation_cmp, NULL);
 202                 }
 203                 up_read(&ca->bucket_lock);
 204         }
 205
 206         if (!sectors_reserved) {
 207                 bch2_fs_fatal_error(c, "stuck, ran out of copygc reserve!");
 208                 return -1;
 209         }
 210
 211         /*
 212          * Our btree node allocations also come out of RESERVE_MOVINGGC:
 213          */
 214         sectors_to_move = (sectors_to_move * 3) / 4;
 215
 216         for (i = h->data; i < h->data + h->used; i++)
 217                 sectors_to_move += i->sectors * i->replicas;
 218
 219         while (sectors_to_move > sectors_reserved) {
 220                 BUG_ON(!heap_pop(h, e, -fragmentation_cmp, NULL));
 221                 sectors_to_move -= e.sectors * e.replicas;
 222         }
 223
 224         buckets_to_move = h->used;
 225
 226         if (!buckets_to_move)
 227                 return 0;
 228
 229         eytzinger0_sort(h->data, h->used,
 230                         sizeof(h->data[0]),
 231                         bucket_offset_cmp, NULL);
 232
 233         ret = bch2_move_data(c,
 234                              0,                 POS_MIN,
 235                              BTREE_ID_NR,       POS_MAX,
 236                              NULL,
 237                              writepoint_ptr(&c->copygc_write_point),
 238                              copygc_pred, NULL,
 239                              &move_stats);
 240
 241         for_each_rw_member(ca, c, dev_idx) {
 242                 down_read(&ca->bucket_lock);
 243                 buckets = bucket_array(ca);
 244                 for (i = h->data; i < h->data + h->used; i++) {
 245                         struct bucket_mark m;
 246                         size_t b;
 247
 248                         if (i->dev != dev_idx)
 249                                 continue;
 250
 251                         b = sector_to_bucket(ca, i->offset);
 252                         m = READ_ONCE(buckets->b[b].mark);
 253
 254                         if (i->gen == m.gen &&
 255                             bucket_sectors_used(m)) {
 256                                 sectors_not_moved += bucket_sectors_used(m);
 257                                 buckets_not_moved++;
 258                         }
 259                 }
 260                 up_read(&ca->bucket_lock);
 261         }
 262
 263         if (sectors_not_moved && !ret)
 264                 bch_warn_ratelimited(c,
 265                         "copygc finished but %llu/%llu sectors, %llu/%llu buckets not moved (move stats: moved %llu sectors, raced %llu keys, %llu sectors)",
 266                          sectors_not_moved, sectors_to_move,
 267                          buckets_not_moved, buckets_to_move,
 268                          atomic64_read(&move_stats.sectors_moved),
 269                          atomic64_read(&move_stats.keys_raced),
 270                          atomic64_read(&move_stats.sectors_raced));
 271
 272         trace_copygc(c,
 273                      atomic64_read(&move_stats.sectors_moved), sectors_not_moved,
 274                      buckets_to_move, buckets_not_moved);
 275         return 0;
 276 }
 277
 278 /*
 279  * Copygc runs when the amount of fragmented data is above some arbitrary
 280  * threshold:
 281  *
 282  * The threshold at the limit - when the device is full - is the amount of space
 283  * we reserved in bch2_recalc_capacity; we can't have more than that amount of
 284  * disk space stranded due to fragmentation and store everything we have
 285  * promised to store.
 286  *
 287  * But we don't want to be running copygc unnecessarily when the device still
 288  * has plenty of free space - rather, we want copygc to smoothly run every so
 289  * often and continually reduce the amount of fragmented space as the device
 290  * fills up. So, we increase the threshold by half the current free space.
 291  */
 292 unsigned long bch2_copygc_wait_amount(struct bch_fs *c)
 293 {
 294         struct bch_dev *ca;
 295         unsigned dev_idx;
 296         s64 wait = S64_MAX, fragmented_allowed, fragmented;
 297
 298         for_each_rw_member(ca, c, dev_idx) {
 299                 struct bch_dev_usage usage = bch2_dev_usage_read(ca);
 300
 301                 fragmented_allowed = ((__dev_buckets_reclaimable(ca, usage) *
 302                                         ca->mi.bucket_size) >> 1);
 303                 fragmented = usage.d[BCH_DATA_user].fragmented;
 304
 305                 wait = min(wait, max(0LL, fragmented_allowed - fragmented));
 306         }
 307
 308         return wait;
 309 }
 310
 311 static int bch2_copygc_thread(void *arg)
 312 {
 313         struct bch_fs *c = arg;
 314         struct io_clock *clock = &c->io_clock[WRITE];
 315         u64 last, wait;
 316
 317         set_freezable();
 318
 319         while (!kthread_should_stop()) {
 320                 cond_resched();
 321
 322                 if (kthread_wait_freezable(c->copy_gc_enabled))
 323                         break;
 324
 325                 last = atomic64_read(&clock->now);
 326                 wait = bch2_copygc_wait_amount(c);
 327
 328                 if (wait > clock->max_slop) {
 329                         trace_copygc_wait(c, wait, last + wait);
 330                         c->copygc_wait = last + wait;
 331                         bch2_kthread_io_clock_wait(clock, last + wait,
 332                                         MAX_SCHEDULE_TIMEOUT);
 333                         continue;
 334                 }
 335
 336                 c->copygc_wait = 0;
 337
 338                 if (bch2_copygc(c))
 339                         break;
 340         }
 341
 342         return 0;
 343 }
 344
 345 void bch2_copygc_stop(struct bch_fs *c)
 346 {
 347         if (c->copygc_thread) {
 348                 kthread_stop(c->copygc_thread);
 349                 put_task_struct(c->copygc_thread);
 350         }
 351         c->copygc_thread = NULL;
 352 }
 353
 354 int bch2_copygc_start(struct bch_fs *c)
 355 {
 356         struct task_struct *t;
 357
 358         if (c->copygc_thread)
 359                 return 0;
 360
 361         if (c->opts.nochanges)
 362                 return 0;
 363
 364         if (bch2_fs_init_fault("copygc_start"))
 365                 return -ENOMEM;
 366
 367         t = kthread_create(bch2_copygc_thread, c, "bch-copygc/%s", c->name);
 368         if (IS_ERR(t)) {
 369                 bch_err(c, "error creating copygc thread: %li", PTR_ERR(t));
 370                 return PTR_ERR(t);
 371         }
 372
 373         get_task_struct(t);
 374
 375         c->copygc_thread = t;
 376         wake_up_process(c->copygc_thread);
 377
 378         return 0;
 379 }
 380
 381 void bch2_fs_copygc_init(struct bch_fs *c)
 382 {
 383 }