fs/bcachefs/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 "errcode.h"
  17 #include "error.h"
  18 #include "lru.h"
  19 #include "move.h"
  20 #include "movinggc.h"
  21 #include "trace.h"
  22
  23 #include <linux/freezer.h>
  24 #include <linux/kthread.h>
  25 #include <linux/math64.h>
  26 #include <linux/sched/task.h>
  27 #include <linux/wait.h>
  28
  29 struct buckets_in_flight {
  30         struct rhashtable               table;
  31         struct move_bucket_in_flight    *first;
  32         struct move_bucket_in_flight    *last;
  33         size_t                          nr;
  34         size_t                          sectors;
  35 };
  36
  37 static const struct rhashtable_params bch_move_bucket_params = {
  38         .head_offset    = offsetof(struct move_bucket_in_flight, hash),
  39         .key_offset     = offsetof(struct move_bucket_in_flight, bucket.k),
  40         .key_len        = sizeof(struct move_bucket_key),
  41 };
  42
  43 static struct move_bucket_in_flight *
  44 move_bucket_in_flight_add(struct buckets_in_flight *list, struct move_bucket b)
  45 {
  46         struct move_bucket_in_flight *new = kzalloc(sizeof(*new), GFP_KERNEL);
  47         int ret;
  48
  49         if (!new)
  50                 return ERR_PTR(-ENOMEM);
  51
  52         new->bucket = b;
  53
  54         ret = rhashtable_lookup_insert_fast(&list->table, &new->hash,
  55                                             bch_move_bucket_params);
  56         if (ret) {
  57                 kfree(new);
  58                 return ERR_PTR(ret);
  59         }
  60
  61         if (!list->first)
  62                 list->first = new;
  63         else
  64                 list->last->next = new;
  65
  66         list->last = new;
  67         list->nr++;
  68         list->sectors += b.sectors;
  69         return new;
  70 }
  71
  72 static int bch2_bucket_is_movable(struct btree_trans *trans,
  73                                   struct move_bucket *b, u64 time)
  74 {
  75         struct btree_iter iter;
  76         struct bkey_s_c k;
  77         struct bch_alloc_v4 _a;
  78         const struct bch_alloc_v4 *a;
  79         int ret;
  80
  81         if (bch2_bucket_is_open(trans->c,
  82                                 b->k.bucket.inode,
  83                                 b->k.bucket.offset))
  84                 return 0;
  85
  86         k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_alloc,
  87                                b->k.bucket, BTREE_ITER_CACHED);
  88         ret = bkey_err(k);
  89         if (ret)
  90                 return ret;
  91
  92         a = bch2_alloc_to_v4(k, &_a);
  93         b->k.gen        = a->gen;
  94         b->sectors      = a->dirty_sectors;
  95
  96         ret = data_type_movable(a->data_type) &&
  97                 a->fragmentation_lru &&
  98                 a->fragmentation_lru <= time;
  99
 100         bch2_trans_iter_exit(trans, &iter);
 101         return ret;
 102 }
 103
 104 static void move_buckets_wait(struct moving_context *ctxt,
 105                               struct buckets_in_flight *list,
 106                               bool flush)
 107 {
 108         struct move_bucket_in_flight *i;
 109         int ret;
 110
 111         while ((i = list->first)) {
 112                 if (flush)
 113                         move_ctxt_wait_event(ctxt, !atomic_read(&i->count));
 114
 115                 if (atomic_read(&i->count))
 116                         break;
 117
 118                 list->first = i->next;
 119                 if (!list->first)
 120                         list->last = NULL;
 121
 122                 list->nr--;
 123                 list->sectors -= i->bucket.sectors;
 124
 125                 ret = rhashtable_remove_fast(&list->table, &i->hash,
 126                                              bch_move_bucket_params);
 127                 BUG_ON(ret);
 128                 kfree(i);
 129         }
 130
 131         bch2_trans_unlock_long(ctxt->trans);
 132 }
 133
 134 static bool bucket_in_flight(struct buckets_in_flight *list,
 135                              struct move_bucket_key k)
 136 {
 137         return rhashtable_lookup_fast(&list->table, &k, bch_move_bucket_params);
 138 }
 139
 140 typedef DARRAY(struct move_bucket) move_buckets;
 141
 142 static int bch2_copygc_get_buckets(struct moving_context *ctxt,
 143                         struct buckets_in_flight *buckets_in_flight,
 144                         move_buckets *buckets)
 145 {
 146         struct btree_trans *trans = ctxt->trans;
 147         struct bch_fs *c = trans->c;
 148         struct btree_iter iter;
 149         struct bkey_s_c k;
 150         size_t nr_to_get = max_t(size_t, 16U, buckets_in_flight->nr / 4);
 151         size_t saw = 0, in_flight = 0, not_movable = 0, sectors = 0;
 152         int ret;
 153
 154         move_buckets_wait(ctxt, buckets_in_flight, false);
 155
 156         ret = bch2_btree_write_buffer_flush(trans);
 157         if (bch2_fs_fatal_err_on(ret, c, "%s: error %s from bch2_btree_write_buffer_flush()",
 158                                  __func__, bch2_err_str(ret)))
 159                 return ret;
 160
 161         ret = for_each_btree_key2_upto(trans, iter, BTREE_ID_lru,
 162                                   lru_pos(BCH_LRU_FRAGMENTATION_START, 0, 0),
 163                                   lru_pos(BCH_LRU_FRAGMENTATION_START, U64_MAX, LRU_TIME_MAX),
 164                                   0, k, ({
 165                 struct move_bucket b = { .k.bucket = u64_to_bucket(k.k->p.offset) };
 166                 int ret2 = 0;
 167
 168                 saw++;
 169
 170                 if (!bch2_bucket_is_movable(trans, &b, lru_pos_time(k.k->p)))
 171                         not_movable++;
 172                 else if (bucket_in_flight(buckets_in_flight, b.k))
 173                         in_flight++;
 174                 else {
 175                         ret2 = darray_push(buckets, b) ?: buckets->nr >= nr_to_get;
 176                         if (ret2 >= 0)
 177                                 sectors += b.sectors;
 178                 }
 179                 ret2;
 180         }));
 181
 182         pr_debug("have: %zu (%zu) saw %zu in flight %zu not movable %zu got %zu (%zu)/%zu buckets ret %i",
 183                  buckets_in_flight->nr, buckets_in_flight->sectors,
 184                  saw, in_flight, not_movable, buckets->nr, sectors, nr_to_get, ret);
 185
 186         return ret < 0 ? ret : 0;
 187 }
 188
 189 noinline
 190 static int bch2_copygc(struct moving_context *ctxt,
 191                        struct buckets_in_flight *buckets_in_flight,
 192                        bool *did_work)
 193 {
 194         struct btree_trans *trans = ctxt->trans;
 195         struct bch_fs *c = trans->c;
 196         struct data_update_opts data_opts = {
 197                 .btree_insert_flags = BCH_WATERMARK_copygc,
 198         };
 199         move_buckets buckets = { 0 };
 200         struct move_bucket_in_flight *f;
 201         struct move_bucket *i;
 202         u64 moved = atomic64_read(&ctxt->stats->sectors_moved);
 203         int ret = 0;
 204
 205         ret = bch2_copygc_get_buckets(ctxt, buckets_in_flight, &buckets);
 206         if (ret)
 207                 goto err;
 208
 209         darray_for_each(buckets, i) {
 210                 if (kthread_should_stop() || freezing(current))
 211                         break;
 212
 213                 f = move_bucket_in_flight_add(buckets_in_flight, *i);
 214                 ret = PTR_ERR_OR_ZERO(f);
 215                 if (ret == -EEXIST) { /* rare race: copygc_get_buckets returned same bucket more than once */
 216                         ret = 0;
 217                         continue;
 218                 }
 219                 if (ret == -ENOMEM) { /* flush IO, continue later */
 220                         ret = 0;
 221                         break;
 222                 }
 223
 224                 ret = __bch2_evacuate_bucket(ctxt, f, f->bucket.k.bucket,
 225                                              f->bucket.k.gen, data_opts);
 226                 if (ret)
 227                         goto err;
 228
 229                 *did_work = true;
 230         }
 231 err:
 232         darray_exit(&buckets);
 233
 234         /* no entries in LRU btree found, or got to end: */
 235         if (bch2_err_matches(ret, ENOENT))
 236                 ret = 0;
 237
 238         if (ret < 0 && !bch2_err_matches(ret, EROFS))
 239                 bch_err_msg(c, ret, "from bch2_move_data()");
 240
 241         moved = atomic64_read(&ctxt->stats->sectors_moved) - moved;
 242         trace_and_count(c, copygc, c, moved, 0, 0, 0);
 243         return ret;
 244 }
 245
 246 /*
 247  * Copygc runs when the amount of fragmented data is above some arbitrary
 248  * threshold:
 249  *
 250  * The threshold at the limit - when the device is full - is the amount of space
 251  * we reserved in bch2_recalc_capacity; we can't have more than that amount of
 252  * disk space stranded due to fragmentation and store everything we have
 253  * promised to store.
 254  *
 255  * But we don't want to be running copygc unnecessarily when the device still
 256  * has plenty of free space - rather, we want copygc to smoothly run every so
 257  * often and continually reduce the amount of fragmented space as the device
 258  * fills up. So, we increase the threshold by half the current free space.
 259  */
 260 unsigned long bch2_copygc_wait_amount(struct bch_fs *c)
 261 {
 262         struct bch_dev *ca;
 263         unsigned dev_idx;
 264         s64 wait = S64_MAX, fragmented_allowed, fragmented;
 265         unsigned i;
 266
 267         for_each_rw_member(ca, c, dev_idx) {
 268                 struct bch_dev_usage usage = bch2_dev_usage_read(ca);
 269
 270                 fragmented_allowed = ((__dev_buckets_available(ca, usage, BCH_WATERMARK_stripe) *
 271                                        ca->mi.bucket_size) >> 1);
 272                 fragmented = 0;
 273
 274                 for (i = 0; i < BCH_DATA_NR; i++)
 275                         if (data_type_movable(i))
 276                                 fragmented += usage.d[i].fragmented;
 277
 278                 wait = min(wait, max(0LL, fragmented_allowed - fragmented));
 279         }
 280
 281         return wait;
 282 }
 283
 284 void bch2_copygc_wait_to_text(struct printbuf *out, struct bch_fs *c)
 285 {
 286         prt_printf(out, "Currently waiting for:     ");
 287         prt_human_readable_u64(out, max(0LL, c->copygc_wait -
 288                                         atomic64_read(&c->io_clock[WRITE].now)) << 9);
 289         prt_newline(out);
 290
 291         prt_printf(out, "Currently waiting since:   ");
 292         prt_human_readable_u64(out, max(0LL,
 293                                         atomic64_read(&c->io_clock[WRITE].now) -
 294                                         c->copygc_wait_at) << 9);
 295         prt_newline(out);
 296
 297         prt_printf(out, "Currently calculated wait: ");
 298         prt_human_readable_u64(out, bch2_copygc_wait_amount(c));
 299         prt_newline(out);
 300 }
 301
 302 static int bch2_copygc_thread(void *arg)
 303 {
 304         struct bch_fs *c = arg;
 305         struct moving_context ctxt;
 306         struct bch_move_stats move_stats;
 307         struct io_clock *clock = &c->io_clock[WRITE];
 308         struct buckets_in_flight *buckets;
 309         u64 last, wait;
 310         int ret = 0;
 311
 312         buckets = kzalloc(sizeof(struct buckets_in_flight), GFP_KERNEL);
 313         if (!buckets)
 314                 return -ENOMEM;
 315         ret = rhashtable_init(&buckets->table, &bch_move_bucket_params);
 316         if (ret) {
 317                 kfree(buckets);
 318                 bch_err_msg(c, ret, "allocating copygc buckets in flight");
 319                 return ret;
 320         }
 321
 322         set_freezable();
 323
 324         bch2_move_stats_init(&move_stats, "copygc");
 325         bch2_moving_ctxt_init(&ctxt, c, NULL, &move_stats,
 326                               writepoint_ptr(&c->copygc_write_point),
 327                               false);
 328
 329         while (!ret && !kthread_should_stop()) {
 330                 bool did_work = false;
 331
 332                 bch2_trans_unlock_long(ctxt.trans);
 333                 cond_resched();
 334
 335                 if (!c->copy_gc_enabled) {
 336                         move_buckets_wait(&ctxt, buckets, true);
 337                         kthread_wait_freezable(c->copy_gc_enabled);
 338                 }
 339
 340                 if (unlikely(freezing(current))) {
 341                         move_buckets_wait(&ctxt, buckets, true);
 342                         __refrigerator(false);
 343                         continue;
 344                 }
 345
 346                 last = atomic64_read(&clock->now);
 347                 wait = bch2_copygc_wait_amount(c);
 348
 349                 if (wait > clock->max_slop) {
 350                         c->copygc_wait_at = last;
 351                         c->copygc_wait = last + wait;
 352                         move_buckets_wait(&ctxt, buckets, true);
 353                         trace_and_count(c, copygc_wait, c, wait, last + wait);
 354                         bch2_kthread_io_clock_wait(clock, last + wait,
 355                                         MAX_SCHEDULE_TIMEOUT);
 356                         continue;
 357                 }
 358
 359                 c->copygc_wait = 0;
 360
 361                 c->copygc_running = true;
 362                 ret = bch2_copygc(&ctxt, buckets, &did_work);
 363                 c->copygc_running = false;
 364
 365                 wake_up(&c->copygc_running_wq);
 366
 367                 if (!wait && !did_work) {
 368                         u64 min_member_capacity = bch2_min_rw_member_capacity(c);
 369
 370                         if (min_member_capacity == U64_MAX)
 371                                 min_member_capacity = 128 * 2048;
 372
 373                         bch2_trans_unlock_long(ctxt.trans);
 374                         bch2_kthread_io_clock_wait(clock, last + (min_member_capacity >> 6),
 375                                         MAX_SCHEDULE_TIMEOUT);
 376                 }
 377         }
 378
 379         move_buckets_wait(&ctxt, buckets, true);
 380
 381         rhashtable_destroy(&buckets->table);
 382         kfree(buckets);
 383         bch2_moving_ctxt_exit(&ctxt);
 384         bch2_move_stats_exit(&move_stats, c);
 385
 386         return 0;
 387 }
 388
 389 void bch2_copygc_stop(struct bch_fs *c)
 390 {
 391         if (c->copygc_thread) {
 392                 kthread_stop(c->copygc_thread);
 393                 put_task_struct(c->copygc_thread);
 394         }
 395         c->copygc_thread = NULL;
 396 }
 397
 398 int bch2_copygc_start(struct bch_fs *c)
 399 {
 400         struct task_struct *t;
 401         int ret;
 402
 403         if (c->copygc_thread)
 404                 return 0;
 405
 406         if (c->opts.nochanges)
 407                 return 0;
 408
 409         if (bch2_fs_init_fault("copygc_start"))
 410                 return -ENOMEM;
 411
 412         t = kthread_create(bch2_copygc_thread, c, "bch-copygc/%s", c->name);
 413         ret = PTR_ERR_OR_ZERO(t);
 414         if (ret) {
 415                 bch_err_msg(c, ret, "creating copygc thread");
 416                 return ret;
 417         }
 418
 419         get_task_struct(t);
 420
 421         c->copygc_thread = t;
 422         wake_up_process(c->copygc_thread);
 423
 424         return 0;
 425 }
 426
 427 void bch2_fs_copygc_init(struct bch_fs *c)
 428 {
 429         init_waitqueue_head(&c->copygc_running_wq);
 430         c->copygc_running = false;
 431 }