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      = 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 btree_trans *trans,
 105                               struct moving_context *ctxt,
 106                               struct buckets_in_flight *list,
 107                               bool flush)
 108 {
 109         struct move_bucket_in_flight *i;
 110         int ret;
 111
 112         while ((i = list->first)) {
 113                 if (flush)
 114                         move_ctxt_wait_event(ctxt, trans, !atomic_read(&i->count));
 115
 116                 if (atomic_read(&i->count))
 117                         break;
 118
 119                 list->first = i->next;
 120                 if (!list->first)
 121                         list->last = NULL;
 122
 123                 list->nr--;
 124                 list->sectors -= i->bucket.sectors;
 125
 126                 ret = rhashtable_remove_fast(&list->table, &i->hash,
 127                                              bch_move_bucket_params);
 128                 BUG_ON(ret);
 129                 kfree(i);
 130         }
 131
 132         bch2_trans_unlock(trans);
 133 }
 134
 135 static bool bucket_in_flight(struct buckets_in_flight *list,
 136                              struct move_bucket_key k)
 137 {
 138         return rhashtable_lookup_fast(&list->table, &k, bch_move_bucket_params);
 139 }
 140
 141 typedef DARRAY(struct move_bucket) move_buckets;
 142
 143 static int bch2_copygc_get_buckets(struct btree_trans *trans,
 144                         struct moving_context *ctxt,
 145                         struct buckets_in_flight *buckets_in_flight,
 146                         move_buckets *buckets)
 147 {
 148         struct bch_fs *c = trans->c;
 149         struct btree_iter iter;
 150         struct bkey_s_c k;
 151         size_t nr_to_get = max_t(size_t, 16U, buckets_in_flight->nr / 4);
 152         size_t saw = 0, in_flight = 0, not_movable = 0, sectors = 0;
 153         int ret;
 154
 155         move_buckets_wait(trans, ctxt, buckets_in_flight, false);
 156
 157         ret = bch2_btree_write_buffer_flush(trans);
 158         if (bch2_fs_fatal_err_on(ret, c, "%s: error %s from bch2_btree_write_buffer_flush()",
 159                                  __func__, bch2_err_str(ret)))
 160                 return ret;
 161
 162         ret = for_each_btree_key2_upto(trans, iter, BTREE_ID_lru,
 163                                   lru_pos(BCH_LRU_FRAGMENTATION_START, 0, 0),
 164                                   lru_pos(BCH_LRU_FRAGMENTATION_START, U64_MAX, LRU_TIME_MAX),
 165                                   0, k, ({
 166                 struct move_bucket b = { .k.bucket = u64_to_bucket(k.k->p.offset) };
 167                 int ret2 = 0;
 168
 169                 saw++;
 170
 171                 if (!bch2_bucket_is_movable(trans, &b, lru_pos_time(k.k->p)))
 172                         not_movable++;
 173                 else if (bucket_in_flight(buckets_in_flight, b.k))
 174                         in_flight++;
 175                 else {
 176                         ret2 = darray_push(buckets, b) ?: buckets->nr >= nr_to_get;
 177                         if (ret2 >= 0)
 178                                 sectors += b.sectors;
 179                 }
 180                 ret2;
 181         }));
 182
 183         pr_debug("have: %zu (%zu) saw %zu in flight %zu not movable %zu got %zu (%zu)/%zu buckets ret %i",
 184                  buckets_in_flight->nr, buckets_in_flight->sectors,
 185                  saw, in_flight, not_movable, buckets->nr, sectors, nr_to_get, ret);
 186
 187         return ret < 0 ? ret : 0;
 188 }
 189
 190 noinline
 191 static int bch2_copygc(struct btree_trans *trans,
 192                        struct moving_context *ctxt,
 193                        struct buckets_in_flight *buckets_in_flight)
 194 {
 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(trans, ctxt, buckets_in_flight, &buckets);
 206         if (ret)
 207                 goto err;
 208
 209         darray_for_each(buckets, i) {
 210                 if (unlikely(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(trans, ctxt, f, f->bucket.k.bucket,
 225                                              f->bucket.k.gen, data_opts);
 226                 if (ret)
 227                         goto err;
 228         }
 229 err:
 230         darray_exit(&buckets);
 231
 232         /* no entries in LRU btree found, or got to end: */
 233         if (bch2_err_matches(ret, ENOENT))
 234                 ret = 0;
 235
 236         if (ret < 0 && !bch2_err_matches(ret, EROFS))
 237                 bch_err_msg(c, ret, "from bch2_move_data()");
 238
 239         moved = atomic64_read(&ctxt->stats->sectors_moved) - moved;
 240         trace_and_count(c, copygc, c, moved, 0, 0, 0);
 241         return ret;
 242 }
 243
 244 /*
 245  * Copygc runs when the amount of fragmented data is above some arbitrary
 246  * threshold:
 247  *
 248  * The threshold at the limit - when the device is full - is the amount of space
 249  * we reserved in bch2_recalc_capacity; we can't have more than that amount of
 250  * disk space stranded due to fragmentation and store everything we have
 251  * promised to store.
 252  *
 253  * But we don't want to be running copygc unnecessarily when the device still
 254  * has plenty of free space - rather, we want copygc to smoothly run every so
 255  * often and continually reduce the amount of fragmented space as the device
 256  * fills up. So, we increase the threshold by half the current free space.
 257  */
 258 unsigned long bch2_copygc_wait_amount(struct bch_fs *c)
 259 {
 260         struct bch_dev *ca;
 261         unsigned dev_idx;
 262         s64 wait = S64_MAX, fragmented_allowed, fragmented;
 263         unsigned i;
 264
 265         for_each_rw_member(ca, c, dev_idx) {
 266                 struct bch_dev_usage usage = bch2_dev_usage_read(ca);
 267
 268                 fragmented_allowed = ((__dev_buckets_available(ca, usage, BCH_WATERMARK_stripe) *
 269                                        ca->mi.bucket_size) >> 1);
 270                 fragmented = 0;
 271
 272                 for (i = 0; i < BCH_DATA_NR; i++)
 273                         if (data_type_movable(i))
 274                                 fragmented += usage.d[i].fragmented;
 275
 276                 wait = min(wait, max(0LL, fragmented_allowed - fragmented));
 277         }
 278
 279         return wait;
 280 }
 281
 282 void bch2_copygc_wait_to_text(struct printbuf *out, struct bch_fs *c)
 283 {
 284         prt_printf(out, "Currently waiting for:     ");
 285         prt_human_readable_u64(out, max(0LL, c->copygc_wait -
 286                                         atomic64_read(&c->io_clock[WRITE].now)) << 9);
 287         prt_newline(out);
 288
 289         prt_printf(out, "Currently waiting since:   ");
 290         prt_human_readable_u64(out, max(0LL,
 291                                         atomic64_read(&c->io_clock[WRITE].now) -
 292                                         c->copygc_wait_at) << 9);
 293         prt_newline(out);
 294
 295         prt_printf(out, "Currently calculated wait: ");
 296         prt_human_readable_u64(out, bch2_copygc_wait_amount(c));
 297         prt_newline(out);
 298 }
 299
 300 static int bch2_copygc_thread(void *arg)
 301 {
 302         struct bch_fs *c = arg;
 303         struct btree_trans *trans;
 304         struct moving_context ctxt;
 305         struct bch_move_stats move_stats;
 306         struct io_clock *clock = &c->io_clock[WRITE];
 307         struct buckets_in_flight buckets;
 308         u64 last, wait;
 309         int ret = 0;
 310
 311         memset(&buckets, 0, sizeof(buckets));
 312
 313         ret = rhashtable_init(&buckets.table, &bch_move_bucket_params);
 314         if (ret) {
 315                 bch_err_msg(c, ret, "allocating copygc buckets in flight");
 316                 return ret;
 317         }
 318
 319         set_freezable();
 320         trans = bch2_trans_get(c);
 321
 322         bch2_move_stats_init(&move_stats, "copygc");
 323         bch2_moving_ctxt_init(&ctxt, c, NULL, &move_stats,
 324                               writepoint_ptr(&c->copygc_write_point),
 325                               false);
 326
 327         while (!ret && !kthread_should_stop()) {
 328                 bch2_trans_unlock(trans);
 329                 cond_resched();
 330
 331                 if (!c->copy_gc_enabled) {
 332                         move_buckets_wait(trans, &ctxt, &buckets, true);
 333                         kthread_wait_freezable(c->copy_gc_enabled);
 334                 }
 335
 336                 if (unlikely(freezing(current))) {
 337                         move_buckets_wait(trans, &ctxt, &buckets, true);
 338                         __refrigerator(false);
 339                         continue;
 340                 }
 341
 342                 last = atomic64_read(&clock->now);
 343                 wait = bch2_copygc_wait_amount(c);
 344
 345                 if (wait > clock->max_slop) {
 346                         c->copygc_wait_at = last;
 347                         c->copygc_wait = last + wait;
 348                         move_buckets_wait(trans, &ctxt, &buckets, true);
 349                         trace_and_count(c, copygc_wait, c, wait, last + wait);
 350                         bch2_kthread_io_clock_wait(clock, last + wait,
 351                                         MAX_SCHEDULE_TIMEOUT);
 352                         continue;
 353                 }
 354
 355                 c->copygc_wait = 0;
 356
 357                 c->copygc_running = true;
 358                 ret = bch2_copygc(trans, &ctxt, &buckets);
 359                 c->copygc_running = false;
 360
 361                 wake_up(&c->copygc_running_wq);
 362         }
 363
 364         move_buckets_wait(trans, &ctxt, &buckets, true);
 365         rhashtable_destroy(&buckets.table);
 366         bch2_trans_put(trans);
 367         bch2_moving_ctxt_exit(&ctxt);
 368
 369         return 0;
 370 }
 371
 372 void bch2_copygc_stop(struct bch_fs *c)
 373 {
 374         if (c->copygc_thread) {
 375                 kthread_stop(c->copygc_thread);
 376                 put_task_struct(c->copygc_thread);
 377         }
 378         c->copygc_thread = NULL;
 379 }
 380
 381 int bch2_copygc_start(struct bch_fs *c)
 382 {
 383         struct task_struct *t;
 384         int ret;
 385
 386         if (c->copygc_thread)
 387                 return 0;
 388
 389         if (c->opts.nochanges)
 390                 return 0;
 391
 392         if (bch2_fs_init_fault("copygc_start"))
 393                 return -ENOMEM;
 394
 395         t = kthread_create(bch2_copygc_thread, c, "bch-copygc/%s", c->name);
 396         ret = PTR_ERR_OR_ZERO(t);
 397         if (ret) {
 398                 bch_err_msg(c, ret, "creating copygc thread");
 399                 return ret;
 400         }
 401
 402         get_task_struct(t);
 403
 404         c->copygc_thread = t;
 405         wake_up_process(c->copygc_thread);
 406
 407         return 0;
 408 }
 409
 410 void bch2_fs_copygc_init(struct bch_fs *c)
 411 {
 412         init_waitqueue_head(&c->copygc_running_wq);
 413         c->copygc_running = false;
 414 }