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_background.h"
  10 #include "alloc_foreground.h"
  11 #include "btree_iter.h"
  12 #include "btree_update.h"
  13 #include "buckets.h"
  14 #include "clock.h"
  15 #include "disk_groups.h"
  16 #include "error.h"
  17 #include "extents.h"
  18 #include "eytzinger.h"
  19 #include "io.h"
  20 #include "keylist.h"
  21 #include "move.h"
  22 #include "movinggc.h"
  23 #include "super-io.h"
  24
  25 #include <trace/events/bcachefs.h>
  26 #include <linux/freezer.h>
  27 #include <linux/kthread.h>
  28 #include <linux/math64.h>
  29 #include <linux/sched/task.h>
  30 #include <linux/sort.h>
  31 #include <linux/wait.h>
  32
  33 static int bucket_offset_cmp(const void *_l, const void *_r, size_t size)
  34 {
  35         const struct copygc_heap_entry *l = _l;
  36         const struct copygc_heap_entry *r = _r;
  37
  38         return  cmp_int(l->dev,    r->dev) ?:
  39                 cmp_int(l->offset, r->offset);
  40 }
  41
  42 static enum data_cmd copygc_pred(struct bch_fs *c, void *arg,
  43                                  struct bkey_s_c k,
  44                                  struct bch_io_opts *io_opts,
  45                                  struct data_opts *data_opts)
  46 {
  47         copygc_heap *h = &c->copygc_heap;
  48         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
  49         const union bch_extent_entry *entry;
  50         struct extent_ptr_decoded p = { 0 };
  51
  52         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
  53                 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
  54                 struct copygc_heap_entry search = {
  55                         .dev    = p.ptr.dev,
  56                         .offset = p.ptr.offset,
  57                 };
  58                 ssize_t i;
  59
  60                 if (p.ptr.cached)
  61                         continue;
  62
  63                 i = eytzinger0_find_le(h->data, h->used,
  64                                        sizeof(h->data[0]),
  65                                        bucket_offset_cmp, &search);
  66 #if 0
  67                 /* eytzinger search verify code: */
  68                 ssize_t j = -1, k;
  69
  70                 for (k = 0; k < h->used; k++)
  71                         if (h->data[k].offset <= ptr->offset &&
  72                             (j < 0 || h->data[k].offset > h->data[j].offset))
  73                                 j = k;
  74
  75                 BUG_ON(i != j);
  76 #endif
  77                 if (i >= 0 &&
  78                     p.ptr.dev == h->data[i].dev &&
  79                     p.ptr.offset < h->data[i].offset + ca->mi.bucket_size &&
  80                     p.ptr.gen == h->data[i].gen) {
  81                         /*
  82                          * We need to use the journal reserve here, because
  83                          *  - journal reclaim depends on btree key cache
  84                          *    flushing to make forward progress,
  85                          *  - which has to make forward progress when the
  86                          *    journal is pre-reservation full,
  87                          *  - and depends on allocation - meaning allocator and
  88                          *    copygc
  89                          */
  90
  91                         data_opts->target               = io_opts->background_target;
  92                         data_opts->nr_replicas          = 1;
  93                         data_opts->btree_insert_flags   = BTREE_INSERT_USE_RESERVE|
  94                                 JOURNAL_WATERMARK_copygc;
  95                         data_opts->rewrite_dev          = p.ptr.dev;
  96
  97                         if (p.has_ec)
  98                                 data_opts->nr_replicas += p.ec.redundancy;
  99
 100                         return DATA_REWRITE;
 101                 }
 102         }
 103
 104         return DATA_SKIP;
 105 }
 106
 107 static inline int fragmentation_cmp(copygc_heap *heap,
 108                                    struct copygc_heap_entry l,
 109                                    struct copygc_heap_entry r)
 110 {
 111         return cmp_int(l.fragmentation, r.fragmentation);
 112 }
 113
 114 static int walk_buckets_to_copygc(struct bch_fs *c)
 115 {
 116         copygc_heap *h = &c->copygc_heap;
 117         struct btree_trans trans;
 118         struct btree_iter iter;
 119         struct bkey_s_c k;
 120         struct bch_alloc_v4 a;
 121         int ret;
 122
 123         bch2_trans_init(&trans, c, 0, 0);
 124
 125         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 126                            BTREE_ITER_PREFETCH, k, ret) {
 127                 struct bch_dev *ca = bch_dev_bkey_exists(c, iter.pos.inode);
 128                 struct copygc_heap_entry e;
 129
 130                 bch2_alloc_to_v4(k, &a);
 131
 132                 if (a.data_type != BCH_DATA_user ||
 133                     a.dirty_sectors >= ca->mi.bucket_size ||
 134                     bch2_bucket_is_open(c, iter.pos.inode, iter.pos.offset))
 135                         continue;
 136
 137                 e = (struct copygc_heap_entry) {
 138                         .dev            = iter.pos.inode,
 139                         .gen            = a.gen,
 140                         .replicas       = 1 + a.stripe_redundancy,
 141                         .fragmentation  = (u64) a.dirty_sectors * (1ULL << 31)
 142                                 / ca->mi.bucket_size,
 143                         .sectors        = a.dirty_sectors,
 144                         .offset         = bucket_to_sector(ca, iter.pos.offset),
 145                 };
 146                 heap_add_or_replace(h, e, -fragmentation_cmp, NULL);
 147
 148         }
 149         bch2_trans_iter_exit(&trans, &iter);
 150
 151         bch2_trans_exit(&trans);
 152         return ret;
 153 }
 154
 155 static int bucket_inorder_cmp(const void *_l, const void *_r)
 156 {
 157         const struct copygc_heap_entry *l = _l;
 158         const struct copygc_heap_entry *r = _r;
 159
 160         return cmp_int(l->dev, r->dev) ?: cmp_int(l->offset, r->offset);
 161 }
 162
 163 static int check_copygc_was_done(struct bch_fs *c,
 164                                  u64 *sectors_not_moved,
 165                                  u64 *buckets_not_moved)
 166 {
 167         copygc_heap *h = &c->copygc_heap;
 168         struct btree_trans trans;
 169         struct btree_iter iter;
 170         struct bkey_s_c k;
 171         struct bch_alloc_v4 a;
 172         struct copygc_heap_entry *i;
 173         int ret = 0;
 174
 175         sort(h->data, h->used, sizeof(h->data[0]), bucket_inorder_cmp, NULL);
 176
 177         bch2_trans_init(&trans, c, 0, 0);
 178         bch2_trans_iter_init(&trans, &iter, BTREE_ID_alloc, POS_MIN, 0);
 179
 180         for (i = h->data; i < h->data + h->used; i++) {
 181                 struct bch_dev *ca = bch_dev_bkey_exists(c, i->dev);
 182
 183                 bch2_btree_iter_set_pos(&iter, POS(i->dev, sector_to_bucket(ca, i->offset)));
 184
 185                 ret = lockrestart_do(&trans,
 186                                 bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
 187                 if (ret)
 188                         break;
 189
 190                 bch2_alloc_to_v4(k, &a);
 191
 192                 if (a.gen == i->gen && a.dirty_sectors) {
 193                         *sectors_not_moved += a.dirty_sectors;
 194                         *buckets_not_moved += 1;
 195                 }
 196         }
 197         bch2_trans_iter_exit(&trans, &iter);
 198
 199         bch2_trans_exit(&trans);
 200         return ret;
 201 }
 202
 203 static int bch2_copygc(struct bch_fs *c)
 204 {
 205         copygc_heap *h = &c->copygc_heap;
 206         struct copygc_heap_entry e, *i;
 207         struct bch_move_stats move_stats;
 208         u64 sectors_to_move = 0, sectors_to_write = 0, sectors_not_moved = 0;
 209         u64 sectors_reserved = 0;
 210         u64 buckets_to_move, buckets_not_moved = 0;
 211         struct bch_dev *ca;
 212         unsigned dev_idx;
 213         size_t heap_size = 0;
 214         int ret;
 215
 216         bch_move_stats_init(&move_stats, "copygc");
 217
 218         /*
 219          * Find buckets with lowest sector counts, skipping completely
 220          * empty buckets, by building a maxheap sorted by sector count,
 221          * and repeatedly replacing the maximum element until all
 222          * buckets have been visited.
 223          */
 224         h->used = 0;
 225
 226         for_each_rw_member(ca, c, dev_idx)
 227                 heap_size += ca->mi.nbuckets >> 7;
 228
 229         if (h->size < heap_size) {
 230                 free_heap(&c->copygc_heap);
 231                 if (!init_heap(&c->copygc_heap, heap_size, GFP_KERNEL)) {
 232                         bch_err(c, "error allocating copygc heap");
 233                         return 0;
 234                 }
 235         }
 236
 237         for_each_rw_member(ca, c, dev_idx) {
 238                 s64 avail = min(dev_buckets_available(ca, RESERVE_movinggc),
 239                                 ca->mi.nbuckets >> 6);
 240
 241                 sectors_reserved += avail * ca->mi.bucket_size;
 242         }
 243
 244         ret = walk_buckets_to_copygc(c);
 245         if (ret) {
 246                 bch2_fs_fatal_error(c, "error walking buckets to copygc!");
 247                 return ret;
 248         }
 249
 250         if (!h->used) {
 251                 bch_err_ratelimited(c, "copygc requested to run but found no buckets to move!");
 252                 return 0;
 253         }
 254
 255         /*
 256          * Our btree node allocations also come out of RESERVE_movingc:
 257          */
 258         sectors_reserved = (sectors_reserved * 3) / 4;
 259         if (!sectors_reserved) {
 260                 bch2_fs_fatal_error(c, "stuck, ran out of copygc reserve!");
 261                 return -1;
 262         }
 263
 264         for (i = h->data; i < h->data + h->used; i++) {
 265                 sectors_to_move += i->sectors;
 266                 sectors_to_write += i->sectors * i->replicas;
 267         }
 268
 269         while (sectors_to_write > sectors_reserved) {
 270                 BUG_ON(!heap_pop(h, e, -fragmentation_cmp, NULL));
 271                 sectors_to_write -= e.sectors * e.replicas;
 272         }
 273
 274         buckets_to_move = h->used;
 275
 276         if (!buckets_to_move) {
 277                 bch_err_ratelimited(c, "copygc cannot run - sectors_reserved %llu!",
 278                                     sectors_reserved);
 279                 return 0;
 280         }
 281
 282         eytzinger0_sort(h->data, h->used,
 283                         sizeof(h->data[0]),
 284                         bucket_offset_cmp, NULL);
 285
 286         ret = bch2_move_data(c,
 287                              0,                 POS_MIN,
 288                              BTREE_ID_NR,       POS_MAX,
 289                              NULL,
 290                              writepoint_ptr(&c->copygc_write_point),
 291                              copygc_pred, NULL,
 292                              &move_stats);
 293         if (ret) {
 294                 bch_err(c, "error %i from bch2_move_data() in copygc", ret);
 295                 return ret;
 296         }
 297
 298         ret = check_copygc_was_done(c, &sectors_not_moved, &buckets_not_moved);
 299         if (ret) {
 300                 bch_err(c, "error %i from check_copygc_was_done()", ret);
 301                 return ret;
 302         }
 303
 304         if (sectors_not_moved)
 305                 bch_warn_ratelimited(c,
 306                         "copygc finished but %llu/%llu sectors, %llu/%llu buckets not moved (move stats: moved %llu sectors, raced %llu keys, %llu sectors)",
 307                          sectors_not_moved, sectors_to_move,
 308                          buckets_not_moved, buckets_to_move,
 309                          atomic64_read(&move_stats.sectors_moved),
 310                          atomic64_read(&move_stats.keys_raced),
 311                          atomic64_read(&move_stats.sectors_raced));
 312
 313         trace_copygc(c,
 314                      atomic64_read(&move_stats.sectors_moved), sectors_not_moved,
 315                      buckets_to_move, buckets_not_moved);
 316         return 0;
 317 }
 318
 319 /*
 320  * Copygc runs when the amount of fragmented data is above some arbitrary
 321  * threshold:
 322  *
 323  * The threshold at the limit - when the device is full - is the amount of space
 324  * we reserved in bch2_recalc_capacity; we can't have more than that amount of
 325  * disk space stranded due to fragmentation and store everything we have
 326  * promised to store.
 327  *
 328  * But we don't want to be running copygc unnecessarily when the device still
 329  * has plenty of free space - rather, we want copygc to smoothly run every so
 330  * often and continually reduce the amount of fragmented space as the device
 331  * fills up. So, we increase the threshold by half the current free space.
 332  */
 333 unsigned long bch2_copygc_wait_amount(struct bch_fs *c)
 334 {
 335         struct bch_dev *ca;
 336         unsigned dev_idx;
 337         s64 wait = S64_MAX, fragmented_allowed, fragmented;
 338
 339         for_each_rw_member(ca, c, dev_idx) {
 340                 struct bch_dev_usage usage = bch2_dev_usage_read(ca);
 341
 342                 fragmented_allowed = ((__dev_buckets_available(ca, usage, RESERVE_none) *
 343                                        ca->mi.bucket_size) >> 1);
 344                 fragmented = usage.d[BCH_DATA_user].fragmented;
 345
 346                 wait = min(wait, max(0LL, fragmented_allowed - fragmented));
 347         }
 348
 349         return wait;
 350 }
 351
 352 static int bch2_copygc_thread(void *arg)
 353 {
 354         struct bch_fs *c = arg;
 355         struct io_clock *clock = &c->io_clock[WRITE];
 356         u64 last, wait;
 357
 358         set_freezable();
 359
 360         while (!kthread_should_stop()) {
 361                 cond_resched();
 362
 363                 if (kthread_wait_freezable(c->copy_gc_enabled))
 364                         break;
 365
 366                 last = atomic64_read(&clock->now);
 367                 wait = bch2_copygc_wait_amount(c);
 368
 369                 if (wait > clock->max_slop) {
 370                         trace_copygc_wait(c, wait, last + wait);
 371                         c->copygc_wait = last + wait;
 372                         bch2_kthread_io_clock_wait(clock, last + wait,
 373                                         MAX_SCHEDULE_TIMEOUT);
 374                         continue;
 375                 }
 376
 377                 c->copygc_wait = 0;
 378
 379                 if (bch2_copygc(c))
 380                         break;
 381         }
 382
 383         return 0;
 384 }
 385
 386 void bch2_copygc_stop(struct bch_fs *c)
 387 {
 388         if (c->copygc_thread) {
 389                 kthread_stop(c->copygc_thread);
 390                 put_task_struct(c->copygc_thread);
 391         }
 392         c->copygc_thread = NULL;
 393 }
 394
 395 int bch2_copygc_start(struct bch_fs *c)
 396 {
 397         struct task_struct *t;
 398
 399         if (c->copygc_thread)
 400                 return 0;
 401
 402         if (c->opts.nochanges)
 403                 return 0;
 404
 405         if (bch2_fs_init_fault("copygc_start"))
 406                 return -ENOMEM;
 407
 408         t = kthread_create(bch2_copygc_thread, c, "bch-copygc/%s", c->name);
 409         if (IS_ERR(t)) {
 410                 bch_err(c, "error creating copygc thread: %li", PTR_ERR(t));
 411                 return PTR_ERR(t);
 412         }
 413
 414         get_task_struct(t);
 415
 416         c->copygc_thread = t;
 417         wake_up_process(c->copygc_thread);
 418
 419         return 0;
 420 }
 421
 422 void bch2_fs_copygc_init(struct bch_fs *c)
 423 {
 424 }