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
  28 #include <trace/events/bcachefs.h>
  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         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
  95                              b->k.bucket, BTREE_ITER_CACHED);
  96         k = bch2_btree_iter_peek_slot(&iter);
  97         ret = bkey_err(k);
  98         bch2_trans_iter_exit(trans, &iter);
  99
 100         if (ret)
 101                 return ret;
 102
 103         a = bch2_alloc_to_v4(k, &_a);
 104         b->k.gen        = a->gen;
 105         b->sectors      = a->dirty_sectors;
 106
 107         ret = data_type_movable(a->data_type) &&
 108                 a->fragmentation_lru &&
 109                 a->fragmentation_lru <= time;
 110
 111         if (!ret) {
 112                 struct printbuf buf = PRINTBUF;
 113
 114                 bch2_bkey_val_to_text(&buf, trans->c, k);
 115                 pr_debug("%s", buf.buf);
 116                 printbuf_exit(&buf);
 117         }
 118
 119         return ret;
 120 }
 121
 122 static void move_buckets_wait(struct btree_trans *trans,
 123                               struct moving_context *ctxt,
 124                               struct buckets_in_flight *list,
 125                               bool flush)
 126 {
 127         struct move_bucket_in_flight *i;
 128         int ret;
 129
 130         while ((i = list->first)) {
 131                 if (flush)
 132                         move_ctxt_wait_event(ctxt, trans, !atomic_read(&i->count));
 133
 134                 if (atomic_read(&i->count))
 135                         break;
 136
 137                 /*
 138                  * moving_ctxt_exit calls bch2_write as it flushes pending
 139                  * reads, which inits another btree_trans; this one must be
 140                  * unlocked:
 141                  */
 142                 bch2_verify_bucket_evacuated(trans, i->bucket.k.bucket, i->bucket.k.gen);
 143
 144                 list->first = i->next;
 145                 if (!list->first)
 146                         list->last = NULL;
 147
 148                 list->nr--;
 149                 list->sectors -= i->bucket.sectors;
 150
 151                 ret = rhashtable_remove_fast(&list->table, &i->hash,
 152                                              bch_move_bucket_params);
 153                 BUG_ON(ret);
 154                 kfree(i);
 155         }
 156
 157         bch2_trans_unlock(trans);
 158 }
 159
 160 static bool bucket_in_flight(struct buckets_in_flight *list,
 161                              struct move_bucket_key k)
 162 {
 163         return rhashtable_lookup_fast(&list->table, &k, bch_move_bucket_params);
 164 }
 165
 166 typedef DARRAY(struct move_bucket) move_buckets;
 167
 168 static int bch2_copygc_get_buckets(struct btree_trans *trans,
 169                         struct moving_context *ctxt,
 170                         struct buckets_in_flight *buckets_in_flight,
 171                         move_buckets *buckets)
 172 {
 173         struct bch_fs *c = trans->c;
 174         struct btree_iter iter;
 175         struct bkey_s_c k;
 176         size_t nr_to_get = max(16UL, buckets_in_flight->nr / 4);
 177         size_t saw = 0, in_flight = 0, not_movable = 0, sectors = 0;
 178         int ret;
 179
 180         move_buckets_wait(trans, ctxt, buckets_in_flight, false);
 181
 182         ret = bch2_btree_write_buffer_flush(trans);
 183         if (bch2_fs_fatal_err_on(ret, c, "%s: error %s from bch2_btree_write_buffer_flush()",
 184                                  __func__, bch2_err_str(ret)))
 185                 return ret;
 186
 187         ret = for_each_btree_key2_upto(trans, iter, BTREE_ID_lru,
 188                                   lru_pos(BCH_LRU_FRAGMENTATION_START, 0, 0),
 189                                   lru_pos(BCH_LRU_FRAGMENTATION_START, U64_MAX, LRU_TIME_MAX),
 190                                   0, k, ({
 191                 struct move_bucket b = { .k.bucket = u64_to_bucket(k.k->p.offset) };
 192                 int ret = 0;
 193
 194                 saw++;
 195
 196                 if (!bch2_bucket_is_movable(trans, &b, lru_pos_time(k.k->p)))
 197                         not_movable++;
 198                 else if (bucket_in_flight(buckets_in_flight, b.k))
 199                         in_flight++;
 200                 else {
 201                         ret = darray_push(buckets, b) ?: buckets->nr >= nr_to_get;
 202                         if (ret >= 0)
 203                                 sectors += b.sectors;
 204                 }
 205                 ret;
 206         }));
 207
 208         pr_debug("have: %zu (%zu) saw %zu in flight %zu not movable %zu got %zu (%zu)/%zu buckets ret %i",
 209                  buckets_in_flight->nr, buckets_in_flight->sectors,
 210                  saw, in_flight, not_movable, buckets->nr, sectors, nr_to_get, ret);
 211
 212         return ret < 0 ? ret : 0;
 213 }
 214
 215 static int bch2_copygc(struct btree_trans *trans,
 216                        struct moving_context *ctxt,
 217                        struct buckets_in_flight *buckets_in_flight)
 218 {
 219         struct bch_fs *c = trans->c;
 220         struct data_update_opts data_opts = {
 221                 .btree_insert_flags = BTREE_INSERT_USE_RESERVE|JOURNAL_WATERMARK_copygc,
 222         };
 223         move_buckets buckets = { 0 };
 224         struct move_bucket_in_flight *f;
 225         struct move_bucket *i;
 226         u64 moved = atomic64_read(&ctxt->stats->sectors_moved);
 227         int ret = 0;
 228
 229         ret = bch2_copygc_get_buckets(trans, ctxt, buckets_in_flight, &buckets);
 230         if (ret)
 231                 goto err;
 232
 233         darray_for_each(buckets, i) {
 234                 if (unlikely(freezing(current)))
 235                         break;
 236
 237                 f = move_bucket_in_flight_add(buckets_in_flight, *i);
 238                 ret = PTR_ERR_OR_ZERO(f);
 239                 if (ret == -EEXIST) /* rare race: copygc_get_buckets returned same bucket more than once */
 240                         continue;
 241                 if (ret == -ENOMEM) { /* flush IO, continue later */
 242                         ret = 0;
 243                         break;
 244                 }
 245
 246                 ret = __bch2_evacuate_bucket(trans, ctxt, f, f->bucket.k.bucket,
 247                                              f->bucket.k.gen, data_opts);
 248                 if (ret)
 249                         goto err;
 250         }
 251 err:
 252         darray_exit(&buckets);
 253
 254         /* no entries in LRU btree found, or got to end: */
 255         if (ret == -ENOENT)
 256                 ret = 0;
 257
 258         if (ret < 0 && !bch2_err_matches(ret, EROFS))
 259                 bch_err(c, "error from bch2_move_data() in copygc: %s", bch2_err_str(ret));
 260
 261         moved = atomic64_read(&ctxt->stats->sectors_moved) - moved;
 262         trace_and_count(c, copygc, c, moved, 0, 0, 0);
 263         return ret;
 264 }
 265
 266 /*
 267  * Copygc runs when the amount of fragmented data is above some arbitrary
 268  * threshold:
 269  *
 270  * The threshold at the limit - when the device is full - is the amount of space
 271  * we reserved in bch2_recalc_capacity; we can't have more than that amount of
 272  * disk space stranded due to fragmentation and store everything we have
 273  * promised to store.
 274  *
 275  * But we don't want to be running copygc unnecessarily when the device still
 276  * has plenty of free space - rather, we want copygc to smoothly run every so
 277  * often and continually reduce the amount of fragmented space as the device
 278  * fills up. So, we increase the threshold by half the current free space.
 279  */
 280 unsigned long bch2_copygc_wait_amount(struct bch_fs *c)
 281 {
 282         struct bch_dev *ca;
 283         unsigned dev_idx;
 284         s64 wait = S64_MAX, fragmented_allowed, fragmented;
 285         unsigned i;
 286
 287         for_each_rw_member(ca, c, dev_idx) {
 288                 struct bch_dev_usage usage = bch2_dev_usage_read(ca);
 289
 290                 fragmented_allowed = ((__dev_buckets_available(ca, usage, RESERVE_stripe) *
 291                                        ca->mi.bucket_size) >> 1);
 292                 fragmented = 0;
 293
 294                 for (i = 0; i < BCH_DATA_NR; i++)
 295                         if (data_type_movable(i))
 296                                 fragmented += usage.d[i].fragmented;
 297
 298                 wait = min(wait, max(0LL, fragmented_allowed - fragmented));
 299         }
 300
 301         return wait;
 302 }
 303
 304 void bch2_copygc_wait_to_text(struct printbuf *out, struct bch_fs *c)
 305 {
 306         prt_printf(out, "Currently waiting for:     ");
 307         prt_human_readable_u64(out, max(0LL, c->copygc_wait -
 308                                         atomic64_read(&c->io_clock[WRITE].now)) << 9);
 309         prt_newline(out);
 310
 311         prt_printf(out, "Currently waiting since:   ");
 312         prt_human_readable_u64(out, max(0LL,
 313                                         atomic64_read(&c->io_clock[WRITE].now) -
 314                                         c->copygc_wait_at) << 9);
 315         prt_newline(out);
 316
 317         prt_printf(out, "Currently calculated wait: ");
 318         prt_human_readable_u64(out, bch2_copygc_wait_amount(c));
 319         prt_newline(out);
 320 }
 321
 322 static int bch2_copygc_thread(void *arg)
 323 {
 324         struct bch_fs *c = arg;
 325         struct btree_trans trans;
 326         struct moving_context ctxt;
 327         struct bch_move_stats move_stats;
 328         struct io_clock *clock = &c->io_clock[WRITE];
 329         struct buckets_in_flight move_buckets;
 330         u64 last, wait;
 331         int ret = 0;
 332
 333         memset(&move_buckets, 0, sizeof(move_buckets));
 334
 335         ret = rhashtable_init(&move_buckets.table, &bch_move_bucket_params);
 336         if (ret) {
 337                 bch_err(c, "error allocating copygc buckets in flight: %s",
 338                         bch2_err_str(ret));
 339                 return ret;
 340         }
 341
 342         set_freezable();
 343         bch2_trans_init(&trans, c, 0, 0);
 344
 345         bch2_move_stats_init(&move_stats, "copygc");
 346         bch2_moving_ctxt_init(&ctxt, c, NULL, &move_stats,
 347                               writepoint_ptr(&c->copygc_write_point),
 348                               false);
 349
 350         while (!ret && !kthread_should_stop()) {
 351                 bch2_trans_unlock(&trans);
 352                 cond_resched();
 353
 354                 if (!c->copy_gc_enabled) {
 355                         move_buckets_wait(&trans, &ctxt, &move_buckets, true);
 356                         kthread_wait_freezable(c->copy_gc_enabled);
 357                 }
 358
 359                 if (unlikely(freezing(current))) {
 360                         move_buckets_wait(&trans, &ctxt, &move_buckets, true);
 361                         __refrigerator(false);
 362                         continue;
 363                 }
 364
 365                 last = atomic64_read(&clock->now);
 366                 wait = bch2_copygc_wait_amount(c);
 367
 368                 if (wait > clock->max_slop) {
 369                         c->copygc_wait_at = last;
 370                         c->copygc_wait = last + wait;
 371                         move_buckets_wait(&trans, &ctxt, &move_buckets, true);
 372                         trace_and_count(c, copygc_wait, c, wait, last + wait);
 373                         bch2_kthread_io_clock_wait(clock, last + wait,
 374                                         MAX_SCHEDULE_TIMEOUT);
 375                         continue;
 376                 }
 377
 378                 c->copygc_wait = 0;
 379
 380                 c->copygc_running = true;
 381                 ret = bch2_copygc(&trans, &ctxt, &move_buckets);
 382                 c->copygc_running = false;
 383
 384                 wake_up(&c->copygc_running_wq);
 385         }
 386
 387         move_buckets_wait(&trans, &ctxt, &move_buckets, true);
 388         bch2_trans_exit(&trans);
 389         bch2_moving_ctxt_exit(&ctxt);
 390
 391         return 0;
 392 }
 393
 394 void bch2_copygc_stop(struct bch_fs *c)
 395 {
 396         if (c->copygc_thread) {
 397                 kthread_stop(c->copygc_thread);
 398                 put_task_struct(c->copygc_thread);
 399         }
 400         c->copygc_thread = NULL;
 401 }
 402
 403 int bch2_copygc_start(struct bch_fs *c)
 404 {
 405         struct task_struct *t;
 406         int ret;
 407
 408         if (c->copygc_thread)
 409                 return 0;
 410
 411         if (c->opts.nochanges)
 412                 return 0;
 413
 414         if (bch2_fs_init_fault("copygc_start"))
 415                 return -ENOMEM;
 416
 417         t = kthread_create(bch2_copygc_thread, c, "bch-copygc/%s", c->name);
 418         ret = PTR_ERR_OR_ZERO(t);
 419         if (ret) {
 420                 bch_err(c, "error creating copygc thread: %s", bch2_err_str(ret));
 421                 return ret;
 422         }
 423
 424         get_task_struct(t);
 425
 426         c->copygc_thread = t;
 427         wake_up_process(c->copygc_thread);
 428
 429         return 0;
 430 }
 431
 432 void bch2_fs_copygc_init(struct bch_fs *c)
 433 {
 434         init_waitqueue_head(&c->copygc_running_wq);
 435         c->copygc_running = false;
 436 }