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                         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                                 data_opts->nr_replicas += p.ec.redundancy;
  97
  98                         return DATA_REWRITE;
  99                 }
 100         }
 101
 102         return DATA_SKIP;
 103 }
 104
 105 static bool have_copygc_reserve(struct bch_dev *ca)
 106 {
 107         bool ret;
 108
 109         spin_lock(&ca->fs->freelist_lock);
 110         ret = fifo_full(&ca->free[RESERVE_MOVINGGC]) ||
 111                 ca->allocator_state != ALLOCATOR_RUNNING;
 112         spin_unlock(&ca->fs->freelist_lock);
 113
 114         return ret;
 115 }
 116
 117 static inline int fragmentation_cmp(copygc_heap *heap,
 118                                    struct copygc_heap_entry l,
 119                                    struct copygc_heap_entry r)
 120 {
 121         return cmp_int(l.fragmentation, r.fragmentation);
 122 }
 123
 124 static int bch2_copygc(struct bch_fs *c)
 125 {
 126         copygc_heap *h = &c->copygc_heap;
 127         struct copygc_heap_entry e, *i;
 128         struct bucket_array *buckets;
 129         struct bch_move_stats move_stats;
 130         u64 sectors_to_move = 0, sectors_not_moved = 0;
 131         u64 sectors_reserved = 0;
 132         u64 buckets_to_move, buckets_not_moved = 0;
 133         struct bch_dev *ca;
 134         unsigned dev_idx;
 135         size_t b, heap_size = 0;
 136         int ret;
 137
 138         memset(&move_stats, 0, sizeof(move_stats));
 139         /*
 140          * Find buckets with lowest sector counts, skipping completely
 141          * empty buckets, by building a maxheap sorted by sector count,
 142          * and repeatedly replacing the maximum element until all
 143          * buckets have been visited.
 144          */
 145         h->used = 0;
 146
 147         for_each_rw_member(ca, c, dev_idx)
 148                 heap_size += ca->mi.nbuckets >> 7;
 149
 150         if (h->size < heap_size) {
 151                 free_heap(&c->copygc_heap);
 152                 if (!init_heap(&c->copygc_heap, heap_size, GFP_KERNEL)) {
 153                         bch_err(c, "error allocating copygc heap");
 154                         return 0;
 155                 }
 156         }
 157
 158         for_each_rw_member(ca, c, dev_idx) {
 159                 closure_wait_event(&c->freelist_wait, have_copygc_reserve(ca));
 160
 161                 spin_lock(&ca->fs->freelist_lock);
 162                 sectors_reserved += fifo_used(&ca->free[RESERVE_MOVINGGC]) * ca->mi.bucket_size;
 163                 spin_unlock(&ca->fs->freelist_lock);
 164
 165                 down_read(&ca->bucket_lock);
 166                 buckets = bucket_array(ca);
 167
 168                 for (b = buckets->first_bucket; b < buckets->nbuckets; b++) {
 169                         struct bucket *g = buckets->b + b;
 170                         struct bucket_mark m = READ_ONCE(g->mark);
 171                         struct copygc_heap_entry e;
 172
 173                         if (m.owned_by_allocator ||
 174                             m.data_type != BCH_DATA_user ||
 175                             !bucket_sectors_used(m) ||
 176                             bucket_sectors_used(m) >= ca->mi.bucket_size)
 177                                 continue;
 178
 179                         WARN_ON(m.stripe && !g->stripe_redundancy);
 180
 181                         e = (struct copygc_heap_entry) {
 182                                 .dev            = dev_idx,
 183                                 .gen            = m.gen,
 184                                 .replicas       = 1 + g->stripe_redundancy,
 185                                 .fragmentation  = bucket_sectors_used(m) * (1U << 15)
 186                                         / ca->mi.bucket_size,
 187                                 .sectors        = bucket_sectors_used(m),
 188                                 .offset         = bucket_to_sector(ca, b),
 189                         };
 190                         heap_add_or_replace(h, e, -fragmentation_cmp, NULL);
 191                 }
 192                 up_read(&ca->bucket_lock);
 193         }
 194
 195         if (!sectors_reserved) {
 196                 bch2_fs_fatal_error(c, "stuck, ran out of copygc reserve!");
 197                 return -1;
 198         }
 199
 200         /*
 201          * Our btree node allocations also come out of RESERVE_MOVINGGC:
 202          */
 203         sectors_to_move = (sectors_to_move * 3) / 4;
 204
 205         for (i = h->data; i < h->data + h->used; i++)
 206                 sectors_to_move += i->sectors * i->replicas;
 207
 208         while (sectors_to_move > sectors_reserved) {
 209                 BUG_ON(!heap_pop(h, e, -fragmentation_cmp, NULL));
 210                 sectors_to_move -= e.sectors * e.replicas;
 211         }
 212
 213         buckets_to_move = h->used;
 214
 215         if (!buckets_to_move)
 216                 return 0;
 217
 218         eytzinger0_sort(h->data, h->used,
 219                         sizeof(h->data[0]),
 220                         bucket_offset_cmp, NULL);
 221
 222         ret = bch2_move_data(c, &c->copygc_pd.rate,
 223                              writepoint_ptr(&c->copygc_write_point),
 224                              POS_MIN, POS_MAX,
 225                              copygc_pred, NULL,
 226                              &move_stats);
 227
 228         for_each_rw_member(ca, c, dev_idx) {
 229                 down_read(&ca->bucket_lock);
 230                 buckets = bucket_array(ca);
 231                 for (i = h->data; i < h->data + h->used; i++) {
 232                         struct bucket_mark m;
 233                         size_t b;
 234
 235                         if (i->dev != dev_idx)
 236                                 continue;
 237
 238                         b = sector_to_bucket(ca, i->offset);
 239                         m = READ_ONCE(buckets->b[b].mark);
 240
 241                         if (i->gen == m.gen &&
 242                             bucket_sectors_used(m)) {
 243                                 sectors_not_moved += bucket_sectors_used(m);
 244                                 buckets_not_moved++;
 245                         }
 246                 }
 247                 up_read(&ca->bucket_lock);
 248         }
 249
 250         if (sectors_not_moved && !ret)
 251                 bch_warn_ratelimited(c,
 252                         "copygc finished but %llu/%llu sectors, %llu/%llu buckets not moved (move stats: moved %llu sectors, raced %llu keys, %llu sectors)",
 253                          sectors_not_moved, sectors_to_move,
 254                          buckets_not_moved, buckets_to_move,
 255                          atomic64_read(&move_stats.sectors_moved),
 256                          atomic64_read(&move_stats.keys_raced),
 257                          atomic64_read(&move_stats.sectors_raced));
 258
 259         trace_copygc(c,
 260                      atomic64_read(&move_stats.sectors_moved), sectors_not_moved,
 261                      buckets_to_move, buckets_not_moved);
 262         return 0;
 263 }
 264
 265 /*
 266  * Copygc runs when the amount of fragmented data is above some arbitrary
 267  * threshold:
 268  *
 269  * The threshold at the limit - when the device is full - is the amount of space
 270  * we reserved in bch2_recalc_capacity; we can't have more than that amount of
 271  * disk space stranded due to fragmentation and store everything we have
 272  * promised to store.
 273  *
 274  * But we don't want to be running copygc unnecessarily when the device still
 275  * has plenty of free space - rather, we want copygc to smoothly run every so
 276  * often and continually reduce the amount of fragmented space as the device
 277  * fills up. So, we increase the threshold by half the current free space.
 278  */
 279 unsigned long bch2_copygc_wait_amount(struct bch_fs *c)
 280 {
 281         struct bch_dev *ca;
 282         unsigned dev_idx;
 283         u64 fragmented_allowed = c->copygc_threshold;
 284         u64 fragmented = 0;
 285
 286         for_each_rw_member(ca, c, dev_idx) {
 287                 struct bch_dev_usage usage = bch2_dev_usage_read(ca);
 288
 289                 fragmented_allowed += ((__dev_buckets_available(ca, usage) *
 290                                         ca->mi.bucket_size) >> 1);
 291                 fragmented += usage.d[BCH_DATA_user].fragmented;
 292         }
 293
 294         return max_t(s64, 0, fragmented_allowed - fragmented);
 295 }
 296
 297 static int bch2_copygc_thread(void *arg)
 298 {
 299         struct bch_fs *c = arg;
 300         struct io_clock *clock = &c->io_clock[WRITE];
 301         u64 last, wait;
 302
 303         set_freezable();
 304
 305         while (!kthread_should_stop()) {
 306                 if (kthread_wait_freezable(c->copy_gc_enabled))
 307                         break;
 308
 309                 last = atomic64_read(&clock->now);
 310                 wait = bch2_copygc_wait_amount(c);
 311
 312                 if (wait > clock->max_slop) {
 313                         bch2_kthread_io_clock_wait(clock, last + wait,
 314                                         MAX_SCHEDULE_TIMEOUT);
 315                         continue;
 316                 }
 317
 318                 if (bch2_copygc(c))
 319                         break;
 320         }
 321
 322         return 0;
 323 }
 324
 325 void bch2_copygc_stop(struct bch_fs *c)
 326 {
 327         c->copygc_pd.rate.rate = UINT_MAX;
 328         bch2_ratelimit_reset(&c->copygc_pd.rate);
 329
 330         if (c->copygc_thread) {
 331                 kthread_stop(c->copygc_thread);
 332                 put_task_struct(c->copygc_thread);
 333         }
 334         c->copygc_thread = NULL;
 335 }
 336
 337 int bch2_copygc_start(struct bch_fs *c)
 338 {
 339         struct task_struct *t;
 340
 341         if (c->copygc_thread)
 342                 return 0;
 343
 344         if (c->opts.nochanges)
 345                 return 0;
 346
 347         if (bch2_fs_init_fault("copygc_start"))
 348                 return -ENOMEM;
 349
 350         t = kthread_create(bch2_copygc_thread, c, "bch-copygc/%s", c->name);
 351         if (IS_ERR(t)) {
 352                 bch_err(c, "error creating copygc thread: %li", PTR_ERR(t));
 353                 return PTR_ERR(t);
 354         }
 355
 356         get_task_struct(t);
 357
 358         c->copygc_thread = t;
 359         wake_up_process(c->copygc_thread);
 360
 361         return 0;
 362 }
 363
 364 void bch2_fs_copygc_init(struct bch_fs *c)
 365 {
 366         bch2_pd_controller_init(&c->copygc_pd);
 367         c->copygc_pd.d_term = 0;
 368 }