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 "btree_write_buffer.h"
  14 #include "buckets.h"
  15 #include "clock.h"
  16 #include "disk_groups.h"
  17 #include "errcode.h"
  18 #include "error.h"
  19 #include "extents.h"
  20 #include "eytzinger.h"
  21 #include "io.h"
  22 #include "keylist.h"
  23 #include "lru.h"
  24 #include "move.h"
  25 #include "movinggc.h"
  26 #include "super-io.h"
  27 #include "trace.h"
  28
  29 #include <linux/bsearch.h>
  30 #include <linux/freezer.h>
  31 #include <linux/kthread.h>
  32 #include <linux/math64.h>
  33 #include <linux/sched/task.h>
  34 #include <linux/sort.h>
  35 #include <linux/wait.h>
  36
  37 struct buckets_in_flight {
  38         struct rhashtable               table;
  39         struct move_bucket_in_flight    *first;
  40         struct move_bucket_in_flight    *last;
  41         size_t                          nr;
  42         size_t                          sectors;
  43 };
  44
  45 static const struct rhashtable_params bch_move_bucket_params = {
  46         .head_offset    = offsetof(struct move_bucket_in_flight, hash),
  47         .key_offset     = offsetof(struct move_bucket_in_flight, bucket.k),
  48         .key_len        = sizeof(struct move_bucket_key),
  49 };
  50
  51 static struct move_bucket_in_flight *
  52 move_bucket_in_flight_add(struct buckets_in_flight *list, struct move_bucket b)
  53 {
  54         struct move_bucket_in_flight *new = kzalloc(sizeof(*new), GFP_KERNEL);
  55         int ret;
  56
  57         if (!new)
  58                 return ERR_PTR(-ENOMEM);
  59
  60         new->bucket = b;
  61
  62         ret = rhashtable_lookup_insert_fast(&list->table, &new->hash,
  63                                             bch_move_bucket_params);
  64         if (ret) {
  65                 kfree(new);
  66                 return ERR_PTR(ret);
  67         }
  68
  69         if (!list->first)
  70                 list->first = new;
  71         else
  72                 list->last->next = new;
  73
  74         list->last = new;
  75         list->nr++;
  76         list->sectors += b.sectors;
  77         return new;
  78 }
  79
  80 static int bch2_bucket_is_movable(struct btree_trans *trans,
  81                                   struct move_bucket *b, u64 time)
  82 {
  83         struct btree_iter iter;
  84         struct bkey_s_c k;
  85         struct bch_alloc_v4 _a;
  86         const struct bch_alloc_v4 *a;
  87         int ret;
  88
  89         if (bch2_bucket_is_open(trans->c,
  90                                 b->k.bucket.inode,
  91                                 b->k.bucket.offset))
  92                 return 0;
  93
  94         k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_alloc,
  95                                b->k.bucket, BTREE_ITER_CACHED);
  96         ret = bkey_err(k);
  97         if (ret)
  98                 return ret;
  99
 100         a = bch2_alloc_to_v4(k, &_a);
 101         b->k.gen        = a->gen;
 102         b->sectors      = a->dirty_sectors;
 103
 104         ret = data_type_movable(a->data_type) &&
 105                 a->fragmentation_lru &&
 106                 a->fragmentation_lru <= time;
 107
 108         bch2_trans_iter_exit(trans, &iter);
 109         return ret;
 110 }
 111
 112 static void move_buckets_wait(struct btree_trans *trans,
 113                               struct moving_context *ctxt,
 114                               struct buckets_in_flight *list,
 115                               bool flush)
 116 {
 117         struct move_bucket_in_flight *i;
 118         int ret;
 119
 120         while ((i = list->first)) {
 121                 if (flush)
 122                         move_ctxt_wait_event(ctxt, trans, !atomic_read(&i->count));
 123
 124                 if (atomic_read(&i->count))
 125                         break;
 126
 127                 list->first = i->next;
 128                 if (!list->first)
 129                         list->last = NULL;
 130
 131                 list->nr--;
 132                 list->sectors -= i->bucket.sectors;
 133
 134                 ret = rhashtable_remove_fast(&list->table, &i->hash,
 135                                              bch_move_bucket_params);
 136                 BUG_ON(ret);
 137                 kfree(i);
 138         }
 139
 140         bch2_trans_unlock(trans);
 141 }
 142
 143 static bool bucket_in_flight(struct buckets_in_flight *list,
 144                              struct move_bucket_key k)
 145 {
 146         return rhashtable_lookup_fast(&list->table, &k, bch_move_bucket_params);
 147 }
 148
 149 typedef DARRAY(struct move_bucket) move_buckets;
 150
 151 static int bch2_copygc_get_buckets(struct btree_trans *trans,
 152                         struct moving_context *ctxt,
 153                         struct buckets_in_flight *buckets_in_flight,
 154                         move_buckets *buckets)
 155 {
 156         struct bch_fs *c = trans->c;
 157         struct btree_iter iter;
 158         struct bkey_s_c k;
 159         size_t nr_to_get = max(16UL, buckets_in_flight->nr / 4);
 160         size_t saw = 0, in_flight = 0, not_movable = 0, sectors = 0;
 161         int ret;
 162
 163         move_buckets_wait(trans, ctxt, buckets_in_flight, false);
 164
 165         ret = bch2_btree_write_buffer_flush(trans);
 166         if (bch2_fs_fatal_err_on(ret, c, "%s: error %s from bch2_btree_write_buffer_flush()",
 167                                  __func__, bch2_err_str(ret)))
 168                 return ret;
 169
 170         ret = for_each_btree_key2_upto(trans, iter, BTREE_ID_lru,
 171                                   lru_pos(BCH_LRU_FRAGMENTATION_START, 0, 0),
 172                                   lru_pos(BCH_LRU_FRAGMENTATION_START, U64_MAX, LRU_TIME_MAX),
 173                                   0, k, ({
 174                 struct move_bucket b = { .k.bucket = u64_to_bucket(k.k->p.offset) };
 175                 int ret = 0;
 176
 177                 saw++;
 178
 179                 if (!bch2_bucket_is_movable(trans, &b, lru_pos_time(k.k->p)))
 180                         not_movable++;
 181                 else if (bucket_in_flight(buckets_in_flight, b.k))
 182                         in_flight++;
 183                 else {
 184                         ret = darray_push(buckets, b) ?: buckets->nr >= nr_to_get;
 185                         if (ret >= 0)
 186                                 sectors += b.sectors;
 187                 }
 188                 ret;
 189         }));
 190
 191         pr_debug("have: %zu (%zu) saw %zu in flight %zu not movable %zu got %zu (%zu)/%zu buckets ret %i",
 192                  buckets_in_flight->nr, buckets_in_flight->sectors,
 193                  saw, in_flight, not_movable, buckets->nr, sectors, nr_to_get, ret);
 194
 195         return ret < 0 ? ret : 0;
 196 }
 197
 198 noinline
 199 static int bch2_copygc(struct btree_trans *trans,
 200                        struct moving_context *ctxt,
 201                        struct buckets_in_flight *buckets_in_flight)
 202 {
 203         struct bch_fs *c = trans->c;
 204         struct data_update_opts data_opts = {
 205                 .btree_insert_flags = BTREE_INSERT_USE_RESERVE|JOURNAL_WATERMARK_copygc,
 206         };
 207         move_buckets buckets = { 0 };
 208         struct move_bucket_in_flight *f;
 209         struct move_bucket *i;
 210         u64 moved = atomic64_read(&ctxt->stats->sectors_moved);
 211         int ret = 0;
 212
 213         ret = bch2_copygc_get_buckets(trans, ctxt, buckets_in_flight, &buckets);
 214         if (ret)
 215                 goto err;
 216
 217         darray_for_each(buckets, i) {
 218                 if (unlikely(freezing(current)))
 219                         break;
 220
 221                 f = move_bucket_in_flight_add(buckets_in_flight, *i);
 222                 ret = PTR_ERR_OR_ZERO(f);
 223                 if (ret == -EEXIST) /* rare race: copygc_get_buckets returned same bucket more than once */
 224                         continue;
 225                 if (ret == -ENOMEM) { /* flush IO, continue later */
 226                         ret = 0;
 227                         break;
 228                 }
 229
 230                 ret = __bch2_evacuate_bucket(trans, ctxt, f, f->bucket.k.bucket,
 231                                              f->bucket.k.gen, data_opts);
 232                 if (ret)
 233                         goto err;
 234         }
 235 err:
 236         darray_exit(&buckets);
 237
 238         /* no entries in LRU btree found, or got to end: */
 239         if (bch2_err_matches(ret, ENOENT))
 240                 ret = 0;
 241
 242         if (ret < 0 && !bch2_err_matches(ret, EROFS))
 243                 bch_err(c, "error from bch2_move_data() in copygc: %s", bch2_err_str(ret));
 244
 245         moved = atomic64_read(&ctxt->stats->sectors_moved) - moved;
 246         trace_and_count(c, copygc, c, moved, 0, 0, 0);
 247         return ret;
 248 }
 249
 250 /*
 251  * Copygc runs when the amount of fragmented data is above some arbitrary
 252  * threshold:
 253  *
 254  * The threshold at the limit - when the device is full - is the amount of space
 255  * we reserved in bch2_recalc_capacity; we can't have more than that amount of
 256  * disk space stranded due to fragmentation and store everything we have
 257  * promised to store.
 258  *
 259  * But we don't want to be running copygc unnecessarily when the device still
 260  * has plenty of free space - rather, we want copygc to smoothly run every so
 261  * often and continually reduce the amount of fragmented space as the device
 262  * fills up. So, we increase the threshold by half the current free space.
 263  */
 264 unsigned long bch2_copygc_wait_amount(struct bch_fs *c)
 265 {
 266         struct bch_dev *ca;
 267         unsigned dev_idx;
 268         s64 wait = S64_MAX, fragmented_allowed, fragmented;
 269         unsigned i;
 270
 271         for_each_rw_member(ca, c, dev_idx) {
 272                 struct bch_dev_usage usage = bch2_dev_usage_read(ca);
 273
 274                 fragmented_allowed = ((__dev_buckets_available(ca, usage, BCH_WATERMARK_stripe) *
 275                                        ca->mi.bucket_size) >> 1);
 276                 fragmented = 0;
 277
 278                 for (i = 0; i < BCH_DATA_NR; i++)
 279                         if (data_type_movable(i))
 280                                 fragmented += usage.d[i].fragmented;
 281
 282                 wait = min(wait, max(0LL, fragmented_allowed - fragmented));
 283         }
 284
 285         return wait;
 286 }
 287
 288 void bch2_copygc_wait_to_text(struct printbuf *out, struct bch_fs *c)
 289 {
 290         prt_printf(out, "Currently waiting for:     ");
 291         prt_human_readable_u64(out, max(0LL, c->copygc_wait -
 292                                         atomic64_read(&c->io_clock[WRITE].now)) << 9);
 293         prt_newline(out);
 294
 295         prt_printf(out, "Currently waiting since:   ");
 296         prt_human_readable_u64(out, max(0LL,
 297                                         atomic64_read(&c->io_clock[WRITE].now) -
 298                                         c->copygc_wait_at) << 9);
 299         prt_newline(out);
 300
 301         prt_printf(out, "Currently calculated wait: ");
 302         prt_human_readable_u64(out, bch2_copygc_wait_amount(c));
 303         prt_newline(out);
 304 }
 305
 306 static int bch2_copygc_thread(void *arg)
 307 {
 308         struct bch_fs *c = arg;
 309         struct btree_trans trans;
 310         struct moving_context ctxt;
 311         struct bch_move_stats move_stats;
 312         struct io_clock *clock = &c->io_clock[WRITE];
 313         struct buckets_in_flight move_buckets;
 314         u64 last, wait;
 315         int ret = 0;
 316
 317         memset(&move_buckets, 0, sizeof(move_buckets));
 318
 319         ret = rhashtable_init(&move_buckets.table, &bch_move_bucket_params);
 320         if (ret) {
 321                 bch_err(c, "error allocating copygc buckets in flight: %s",
 322                         bch2_err_str(ret));
 323                 return ret;
 324         }
 325
 326         set_freezable();
 327         bch2_trans_init(&trans, c, 0, 0);
 328
 329         bch2_move_stats_init(&move_stats, "copygc");
 330         bch2_moving_ctxt_init(&ctxt, c, NULL, &move_stats,
 331                               writepoint_ptr(&c->copygc_write_point),
 332                               false);
 333
 334         while (!ret && !kthread_should_stop()) {
 335                 bch2_trans_unlock(&trans);
 336                 cond_resched();
 337
 338                 if (!c->copy_gc_enabled) {
 339                         move_buckets_wait(&trans, &ctxt, &move_buckets, true);
 340                         kthread_wait_freezable(c->copy_gc_enabled);
 341                 }
 342
 343                 if (unlikely(freezing(current))) {
 344                         move_buckets_wait(&trans, &ctxt, &move_buckets, true);
 345                         __refrigerator(false);
 346                         continue;
 347                 }
 348
 349                 last = atomic64_read(&clock->now);
 350                 wait = bch2_copygc_wait_amount(c);
 351
 352                 if (wait > clock->max_slop) {
 353                         c->copygc_wait_at = last;
 354                         c->copygc_wait = last + wait;
 355                         move_buckets_wait(&trans, &ctxt, &move_buckets, true);
 356                         trace_and_count(c, copygc_wait, c, wait, last + wait);
 357                         bch2_kthread_io_clock_wait(clock, last + wait,
 358                                         MAX_SCHEDULE_TIMEOUT);
 359                         continue;
 360                 }
 361
 362                 c->copygc_wait = 0;
 363
 364                 c->copygc_running = true;
 365                 ret = bch2_copygc(&trans, &ctxt, &move_buckets);
 366                 c->copygc_running = false;
 367
 368                 wake_up(&c->copygc_running_wq);
 369         }
 370
 371         move_buckets_wait(&trans, &ctxt, &move_buckets, true);
 372         rhashtable_destroy(&move_buckets.table);
 373         bch2_trans_exit(&trans);
 374         bch2_moving_ctxt_exit(&ctxt);
 375
 376         return 0;
 377 }
 378
 379 void bch2_copygc_stop(struct bch_fs *c)
 380 {
 381         if (c->copygc_thread) {
 382                 kthread_stop(c->copygc_thread);
 383                 put_task_struct(c->copygc_thread);
 384         }
 385         c->copygc_thread = NULL;
 386 }
 387
 388 int bch2_copygc_start(struct bch_fs *c)
 389 {
 390         struct task_struct *t;
 391         int ret;
 392
 393         if (c->copygc_thread)
 394                 return 0;
 395
 396         if (c->opts.nochanges)
 397                 return 0;
 398
 399         if (bch2_fs_init_fault("copygc_start"))
 400                 return -ENOMEM;
 401
 402         t = kthread_create(bch2_copygc_thread, c, "bch-copygc/%s", c->name);
 403         ret = PTR_ERR_OR_ZERO(t);
 404         if (ret) {
 405                 bch_err(c, "error creating copygc thread: %s", bch2_err_str(ret));
 406                 return ret;
 407         }
 408
 409         get_task_struct(t);
 410
 411         c->copygc_thread = t;
 412         wake_up_process(c->copygc_thread);
 413
 414         return 0;
 415 }
 416
 417 void bch2_fs_copygc_init(struct bch_fs *c)
 418 {
 419         init_waitqueue_head(&c->copygc_running_wq);
 420         c->copygc_running = false;
 421 }