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;
  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                         data_opts->target               = io_opts->background_target;
  91                         data_opts->nr_replicas          = 1;
  92                         data_opts->btree_insert_flags   = BTREE_INSERT_USE_RESERVE;
  93                         data_opts->rewrite_dev          = p.ptr.dev;
  94
  95                         if (p.has_ec) {
  96                                 struct stripe *m = genradix_ptr(&c->stripes[0], p.ec.idx);
  97
  98                                 data_opts->nr_replicas += m->nr_redundant;
  99                         }
 100
 101                         return DATA_REWRITE;
 102                 }
 103         }
 104
 105         return DATA_SKIP;
 106 }
 107
 108 static bool have_copygc_reserve(struct bch_dev *ca)
 109 {
 110         bool ret;
 111
 112         spin_lock(&ca->fs->freelist_lock);
 113         ret = fifo_full(&ca->free[RESERVE_MOVINGGC]) ||
 114                 ca->allocator_state != ALLOCATOR_RUNNING;
 115         spin_unlock(&ca->fs->freelist_lock);
 116
 117         return ret;
 118 }
 119
 120 static inline int fragmentation_cmp(copygc_heap *heap,
 121                                    struct copygc_heap_entry l,
 122                                    struct copygc_heap_entry r)
 123 {
 124         return cmp_int(l.fragmentation, r.fragmentation);
 125 }
 126
 127 static int bch2_copygc(struct bch_fs *c)
 128 {
 129         copygc_heap *h = &c->copygc_heap;
 130         struct copygc_heap_entry e, *i;
 131         struct bucket_array *buckets;
 132         struct bch_move_stats move_stats;
 133         u64 sectors_to_move = 0, sectors_not_moved = 0;
 134         u64 sectors_reserved = 0;
 135         u64 buckets_to_move, buckets_not_moved = 0;
 136         struct bch_dev *ca;
 137         unsigned dev_idx;
 138         size_t b, heap_size = 0;
 139         int ret;
 140
 141         memset(&move_stats, 0, sizeof(move_stats));
 142         /*
 143          * Find buckets with lowest sector counts, skipping completely
 144          * empty buckets, by building a maxheap sorted by sector count,
 145          * and repeatedly replacing the maximum element until all
 146          * buckets have been visited.
 147          */
 148         h->used = 0;
 149
 150         for_each_rw_member(ca, c, dev_idx)
 151                 heap_size += ca->mi.nbuckets >> 7;
 152
 153         if (h->size < heap_size) {
 154                 free_heap(&c->copygc_heap);
 155                 if (!init_heap(&c->copygc_heap, heap_size, GFP_KERNEL)) {
 156                         bch_err(c, "error allocating copygc heap");
 157                         return 0;
 158                 }
 159         }
 160
 161         for_each_rw_member(ca, c, dev_idx) {
 162                 closure_wait_event(&c->freelist_wait, have_copygc_reserve(ca));
 163
 164                 spin_lock(&ca->fs->freelist_lock);
 165                 sectors_reserved += fifo_used(&ca->free[RESERVE_MOVINGGC]) * ca->mi.bucket_size;
 166                 spin_unlock(&ca->fs->freelist_lock);
 167
 168                 down_read(&ca->bucket_lock);
 169                 buckets = bucket_array(ca);
 170
 171                 for (b = buckets->first_bucket; b < buckets->nbuckets; b++) {
 172                         struct bucket *g = buckets->b + b;
 173                         struct bucket_mark m = READ_ONCE(g->mark);
 174                         struct copygc_heap_entry e;
 175
 176                         if (m.owned_by_allocator ||
 177                             m.data_type != BCH_DATA_user ||
 178                             !bucket_sectors_used(m) ||
 179                             bucket_sectors_used(m) >= ca->mi.bucket_size)
 180                                 continue;
 181
 182                         WARN_ON(m.stripe && !g->ec_redundancy);
 183
 184                         e = (struct copygc_heap_entry) {
 185                                 .dev            = dev_idx,
 186                                 .gen            = m.gen,
 187                                 .replicas       = 1 + g->ec_redundancy,
 188                                 .fragmentation  = bucket_sectors_used(m) * (1U << 15)
 189                                         / ca->mi.bucket_size,
 190                                 .sectors        = bucket_sectors_used(m),
 191                                 .offset         = bucket_to_sector(ca, b),
 192                         };
 193                         heap_add_or_replace(h, e, -fragmentation_cmp, NULL);
 194                 }
 195                 up_read(&ca->bucket_lock);
 196         }
 197
 198         if (!sectors_reserved) {
 199                 bch2_fs_fatal_error(c, "stuck, ran out of copygc reserve!");
 200                 return -1;
 201         }
 202
 203         for (i = h->data; i < h->data + h->used; i++)
 204                 sectors_to_move += i->sectors * i->replicas;
 205
 206         while (sectors_to_move > sectors_reserved) {
 207                 BUG_ON(!heap_pop(h, e, -fragmentation_cmp, NULL));
 208                 sectors_to_move -= e.sectors * e.replicas;
 209         }
 210
 211         buckets_to_move = h->used;
 212
 213         if (!buckets_to_move)
 214                 return 0;
 215
 216         eytzinger0_sort(h->data, h->used,
 217                         sizeof(h->data[0]),
 218                         bucket_offset_cmp, NULL);
 219
 220         ret = bch2_move_data(c, &c->copygc_pd.rate,
 221                              writepoint_ptr(&c->copygc_write_point),
 222                              POS_MIN, POS_MAX,
 223                              copygc_pred, NULL,
 224                              &move_stats);
 225
 226         for_each_rw_member(ca, c, dev_idx) {
 227                 down_read(&ca->bucket_lock);
 228                 buckets = bucket_array(ca);
 229                 for (i = h->data; i < h->data + h->used; i++) {
 230                         struct bucket_mark m;
 231                         size_t b;
 232
 233                         if (i->dev != dev_idx)
 234                                 continue;
 235
 236                         b = sector_to_bucket(ca, i->offset);
 237                         m = READ_ONCE(buckets->b[b].mark);
 238
 239                         if (i->gen == m.gen &&
 240                             bucket_sectors_used(m)) {
 241                                 sectors_not_moved += bucket_sectors_used(m);
 242                                 buckets_not_moved++;
 243                         }
 244                 }
 245                 up_read(&ca->bucket_lock);
 246         }
 247
 248         if (sectors_not_moved && !ret)
 249                 bch_warn_ratelimited(c,
 250                         "copygc finished but %llu/%llu sectors, %llu/%llu buckets not moved (move stats: moved %llu sectors, raced %llu keys, %llu sectors)",
 251                          sectors_not_moved, sectors_to_move,
 252                          buckets_not_moved, buckets_to_move,
 253                          atomic64_read(&move_stats.sectors_moved),
 254                          atomic64_read(&move_stats.keys_raced),
 255                          atomic64_read(&move_stats.sectors_raced));
 256
 257         trace_copygc(c,
 258                      atomic64_read(&move_stats.sectors_moved), sectors_not_moved,
 259                      buckets_to_move, buckets_not_moved);
 260         return 0;
 261 }
 262
 263 /*
 264  * Copygc runs when the amount of fragmented data is above some arbitrary
 265  * threshold:
 266  *
 267  * The threshold at the limit - when the device is full - is the amount of space
 268  * we reserved in bch2_recalc_capacity; we can't have more than that amount of
 269  * disk space stranded due to fragmentation and store everything we have
 270  * promised to store.
 271  *
 272  * But we don't want to be running copygc unnecessarily when the device still
 273  * has plenty of free space - rather, we want copygc to smoothly run every so
 274  * often and continually reduce the amount of fragmented space as the device
 275  * fills up. So, we increase the threshold by half the current free space.
 276  */
 277 unsigned long bch2_copygc_wait_amount(struct bch_fs *c)
 278 {
 279         struct bch_dev *ca;
 280         unsigned dev_idx;
 281         u64 fragmented_allowed = c->copygc_threshold;
 282         u64 fragmented = 0;
 283
 284         for_each_rw_member(ca, c, dev_idx) {
 285                 struct bch_dev_usage usage = bch2_dev_usage_read(ca);
 286
 287                 fragmented_allowed += ((__dev_buckets_available(ca, usage) *
 288                                         ca->mi.bucket_size) >> 1);
 289                 fragmented += usage.sectors_fragmented;
 290         }
 291
 292         return max_t(s64, 0, fragmented_allowed - fragmented);
 293 }
 294
 295 static int bch2_copygc_thread(void *arg)
 296 {
 297         struct bch_fs *c = arg;
 298         struct io_clock *clock = &c->io_clock[WRITE];
 299         unsigned long last, wait;
 300
 301         set_freezable();
 302
 303         while (!kthread_should_stop()) {
 304                 if (kthread_wait_freezable(c->copy_gc_enabled))
 305                         break;
 306
 307                 last = atomic_long_read(&clock->now);
 308                 wait = bch2_copygc_wait_amount(c);
 309
 310                 if (wait > clock->max_slop) {
 311                         bch2_kthread_io_clock_wait(clock, last + wait,
 312                                         MAX_SCHEDULE_TIMEOUT);
 313                         continue;
 314                 }
 315
 316                 if (bch2_copygc(c))
 317                         break;
 318         }
 319
 320         return 0;
 321 }
 322
 323 void bch2_copygc_stop(struct bch_fs *c)
 324 {
 325         c->copygc_pd.rate.rate = UINT_MAX;
 326         bch2_ratelimit_reset(&c->copygc_pd.rate);
 327
 328         if (c->copygc_thread) {
 329                 kthread_stop(c->copygc_thread);
 330                 put_task_struct(c->copygc_thread);
 331         }
 332         c->copygc_thread = NULL;
 333 }
 334
 335 int bch2_copygc_start(struct bch_fs *c)
 336 {
 337         struct task_struct *t;
 338
 339         if (c->copygc_thread)
 340                 return 0;
 341
 342         if (c->opts.nochanges)
 343                 return 0;
 344
 345         if (bch2_fs_init_fault("copygc_start"))
 346                 return -ENOMEM;
 347
 348         t = kthread_create(bch2_copygc_thread, c, "bch-copygc/%s", c->name);
 349         if (IS_ERR(t))
 350                 return PTR_ERR(t);
 351
 352         get_task_struct(t);
 353
 354         c->copygc_thread = t;
 355         wake_up_process(c->copygc_thread);
 356
 357         return 0;
 358 }
 359
 360 void bch2_fs_copygc_init(struct bch_fs *c)
 361 {
 362         bch2_pd_controller_init(&c->copygc_pd);
 363         c->copygc_pd.d_term = 0;
 364 }