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 "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      = bch2_bucket_sectors_dirty(*a);
  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_tryflush(trans);
 157         if (bch2_err_matches(ret, EROFS))
 158                 return ret;
 159
 160         if (bch2_fs_fatal_err_on(ret, c, "%s: error %s from bch2_btree_write_buffer_tryflush()",
 161                                  __func__, bch2_err_str(ret)))
 162                 return ret;
 163
 164         ret = for_each_btree_key2_upto(trans, iter, BTREE_ID_lru,
 165                                   lru_pos(BCH_LRU_FRAGMENTATION_START, 0, 0),
 166                                   lru_pos(BCH_LRU_FRAGMENTATION_START, U64_MAX, LRU_TIME_MAX),
 167                                   0, k, ({
 168                 struct move_bucket b = { .k.bucket = u64_to_bucket(k.k->p.offset) };
 169                 int ret2 = 0;
 170
 171                 saw++;
 172
 173                 ret2 = bch2_bucket_is_movable(trans, &b, lru_pos_time(k.k->p));
 174                 if (ret2 < 0)
 175                         goto err;
 176
 177                 if (!ret2)
 178                         not_movable++;
 179                 else if (bucket_in_flight(buckets_in_flight, b.k))
 180                         in_flight++;
 181                 else {
 182                         ret2 = darray_push(buckets, b);
 183                         if (ret2)
 184                                 goto err;
 185                         sectors += b.sectors;
 186                 }
 187
 188                 ret2 = buckets->nr >= nr_to_get;
 189 err:
 190                 ret2;
 191         }));
 192
 193         pr_debug("have: %zu (%zu) saw %zu in flight %zu not movable %zu got %zu (%zu)/%zu buckets ret %i",
 194                  buckets_in_flight->nr, buckets_in_flight->sectors,
 195                  saw, in_flight, not_movable, buckets->nr, sectors, nr_to_get, ret);
 196
 197         return ret < 0 ? ret : 0;
 198 }
 199
 200 noinline
 201 static int bch2_copygc(struct moving_context *ctxt,
 202                        struct buckets_in_flight *buckets_in_flight,
 203                        bool *did_work)
 204 {
 205         struct btree_trans *trans = ctxt->trans;
 206         struct bch_fs *c = trans->c;
 207         struct data_update_opts data_opts = {
 208                 .btree_insert_flags = BCH_WATERMARK_copygc,
 209         };
 210         move_buckets buckets = { 0 };
 211         struct move_bucket_in_flight *f;
 212         struct move_bucket *i;
 213         u64 moved = atomic64_read(&ctxt->stats->sectors_moved);
 214         int ret = 0;
 215
 216         ret = bch2_copygc_get_buckets(ctxt, buckets_in_flight, &buckets);
 217         if (ret)
 218                 goto err;
 219
 220         darray_for_each(buckets, i) {
 221                 if (kthread_should_stop() || freezing(current))
 222                         break;
 223
 224                 f = move_bucket_in_flight_add(buckets_in_flight, *i);
 225                 ret = PTR_ERR_OR_ZERO(f);
 226                 if (ret == -EEXIST) { /* rare race: copygc_get_buckets returned same bucket more than once */
 227                         ret = 0;
 228                         continue;
 229                 }
 230                 if (ret == -ENOMEM) { /* flush IO, continue later */
 231                         ret = 0;
 232                         break;
 233                 }
 234
 235                 ret = bch2_evacuate_bucket(ctxt, f, f->bucket.k.bucket,
 236                                              f->bucket.k.gen, data_opts);
 237                 if (ret)
 238                         goto err;
 239
 240                 *did_work = true;
 241         }
 242 err:
 243         darray_exit(&buckets);
 244
 245         /* no entries in LRU btree found, or got to end: */
 246         if (bch2_err_matches(ret, ENOENT))
 247                 ret = 0;
 248
 249         if (ret < 0 && !bch2_err_matches(ret, EROFS))
 250                 bch_err_msg(c, ret, "from bch2_move_data()");
 251
 252         moved = atomic64_read(&ctxt->stats->sectors_moved) - moved;
 253         trace_and_count(c, copygc, c, moved, 0, 0, 0);
 254         return ret;
 255 }
 256
 257 /*
 258  * Copygc runs when the amount of fragmented data is above some arbitrary
 259  * threshold:
 260  *
 261  * The threshold at the limit - when the device is full - is the amount of space
 262  * we reserved in bch2_recalc_capacity; we can't have more than that amount of
 263  * disk space stranded due to fragmentation and store everything we have
 264  * promised to store.
 265  *
 266  * But we don't want to be running copygc unnecessarily when the device still
 267  * has plenty of free space - rather, we want copygc to smoothly run every so
 268  * often and continually reduce the amount of fragmented space as the device
 269  * fills up. So, we increase the threshold by half the current free space.
 270  */
 271 unsigned long bch2_copygc_wait_amount(struct bch_fs *c)
 272 {
 273         struct bch_dev *ca;
 274         unsigned dev_idx;
 275         s64 wait = S64_MAX, fragmented_allowed, fragmented;
 276         unsigned i;
 277
 278         for_each_rw_member(ca, c, dev_idx) {
 279                 struct bch_dev_usage usage = bch2_dev_usage_read(ca);
 280
 281                 fragmented_allowed = ((__dev_buckets_available(ca, usage, BCH_WATERMARK_stripe) *
 282                                        ca->mi.bucket_size) >> 1);
 283                 fragmented = 0;
 284
 285                 for (i = 0; i < BCH_DATA_NR; i++)
 286                         if (data_type_movable(i))
 287                                 fragmented += usage.d[i].fragmented;
 288
 289                 wait = min(wait, max(0LL, fragmented_allowed - fragmented));
 290         }
 291
 292         return wait;
 293 }
 294
 295 void bch2_copygc_wait_to_text(struct printbuf *out, struct bch_fs *c)
 296 {
 297         prt_printf(out, "Currently waiting for:     ");
 298         prt_human_readable_u64(out, max(0LL, c->copygc_wait -
 299                                         atomic64_read(&c->io_clock[WRITE].now)) << 9);
 300         prt_newline(out);
 301
 302         prt_printf(out, "Currently waiting since:   ");
 303         prt_human_readable_u64(out, max(0LL,
 304                                         atomic64_read(&c->io_clock[WRITE].now) -
 305                                         c->copygc_wait_at) << 9);
 306         prt_newline(out);
 307
 308         prt_printf(out, "Currently calculated wait: ");
 309         prt_human_readable_u64(out, bch2_copygc_wait_amount(c));
 310         prt_newline(out);
 311 }
 312
 313 static int bch2_copygc_thread(void *arg)
 314 {
 315         struct bch_fs *c = arg;
 316         struct moving_context ctxt;
 317         struct bch_move_stats move_stats;
 318         struct io_clock *clock = &c->io_clock[WRITE];
 319         struct buckets_in_flight *buckets;
 320         u64 last, wait;
 321         int ret = 0;
 322
 323         buckets = kzalloc(sizeof(struct buckets_in_flight), GFP_KERNEL);
 324         if (!buckets)
 325                 return -ENOMEM;
 326         ret = rhashtable_init(&buckets->table, &bch_move_bucket_params);
 327         if (ret) {
 328                 kfree(buckets);
 329                 bch_err_msg(c, ret, "allocating copygc buckets in flight");
 330                 return ret;
 331         }
 332
 333         set_freezable();
 334
 335         bch2_move_stats_init(&move_stats, "copygc");
 336         bch2_moving_ctxt_init(&ctxt, c, NULL, &move_stats,
 337                               writepoint_ptr(&c->copygc_write_point),
 338                               false);
 339
 340         while (!ret && !kthread_should_stop()) {
 341                 bool did_work = false;
 342
 343                 bch2_trans_unlock_long(ctxt.trans);
 344                 cond_resched();
 345
 346                 if (!c->copy_gc_enabled) {
 347                         move_buckets_wait(&ctxt, buckets, true);
 348                         kthread_wait_freezable(c->copy_gc_enabled ||
 349                                                kthread_should_stop());
 350                 }
 351
 352                 if (unlikely(freezing(current))) {
 353                         move_buckets_wait(&ctxt, buckets, true);
 354                         __refrigerator(false);
 355                         continue;
 356                 }
 357
 358                 last = atomic64_read(&clock->now);
 359                 wait = bch2_copygc_wait_amount(c);
 360
 361                 if (wait > clock->max_slop) {
 362                         c->copygc_wait_at = last;
 363                         c->copygc_wait = last + wait;
 364                         move_buckets_wait(&ctxt, buckets, true);
 365                         trace_and_count(c, copygc_wait, c, wait, last + wait);
 366                         bch2_kthread_io_clock_wait(clock, last + wait,
 367                                         MAX_SCHEDULE_TIMEOUT);
 368                         continue;
 369                 }
 370
 371                 c->copygc_wait = 0;
 372
 373                 c->copygc_running = true;
 374                 ret = bch2_copygc(&ctxt, buckets, &did_work);
 375                 c->copygc_running = false;
 376
 377                 wake_up(&c->copygc_running_wq);
 378
 379                 if (!wait && !did_work) {
 380                         u64 min_member_capacity = bch2_min_rw_member_capacity(c);
 381
 382                         if (min_member_capacity == U64_MAX)
 383                                 min_member_capacity = 128 * 2048;
 384
 385                         bch2_trans_unlock_long(ctxt.trans);
 386                         bch2_kthread_io_clock_wait(clock, last + (min_member_capacity >> 6),
 387                                         MAX_SCHEDULE_TIMEOUT);
 388                 }
 389         }
 390
 391         move_buckets_wait(&ctxt, buckets, true);
 392
 393         rhashtable_destroy(&buckets->table);
 394         kfree(buckets);
 395         bch2_moving_ctxt_exit(&ctxt);
 396         bch2_move_stats_exit(&move_stats, c);
 397
 398         return 0;
 399 }
 400
 401 void bch2_copygc_stop(struct bch_fs *c)
 402 {
 403         if (c->copygc_thread) {
 404                 kthread_stop(c->copygc_thread);
 405                 put_task_struct(c->copygc_thread);
 406         }
 407         c->copygc_thread = NULL;
 408 }
 409
 410 int bch2_copygc_start(struct bch_fs *c)
 411 {
 412         struct task_struct *t;
 413         int ret;
 414
 415         if (c->copygc_thread)
 416                 return 0;
 417
 418         if (c->opts.nochanges)
 419                 return 0;
 420
 421         if (bch2_fs_init_fault("copygc_start"))
 422                 return -ENOMEM;
 423
 424         t = kthread_create(bch2_copygc_thread, c, "bch-copygc/%s", c->name);
 425         ret = PTR_ERR_OR_ZERO(t);
 426         if (ret) {
 427                 bch_err_msg(c, ret, "creating copygc thread");
 428                 return ret;
 429         }
 430
 431         get_task_struct(t);
 432
 433         c->copygc_thread = t;
 434         wake_up_process(c->copygc_thread);
 435
 436         return 0;
 437 }
 438
 439 void bch2_fs_copygc_init(struct bch_fs *c)
 440 {
 441         init_waitqueue_head(&c->copygc_running_wq);
 442         c->copygc_running = false;
 443 }