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         .automatic_shrinking    = true,
  42 };
  43
  44 static struct move_bucket_in_flight *
  45 move_bucket_in_flight_add(struct buckets_in_flight *list, struct move_bucket b)
  46 {
  47         struct move_bucket_in_flight *new = kzalloc(sizeof(*new), GFP_KERNEL);
  48         int ret;
  49
  50         if (!new)
  51                 return ERR_PTR(-ENOMEM);
  52
  53         new->bucket = b;
  54
  55         ret = rhashtable_lookup_insert_fast(&list->table, &new->hash,
  56                                             bch_move_bucket_params);
  57         if (ret) {
  58                 kfree(new);
  59                 return ERR_PTR(ret);
  60         }
  61
  62         if (!list->first)
  63                 list->first = new;
  64         else
  65                 list->last->next = new;
  66
  67         list->last = new;
  68         list->nr++;
  69         list->sectors += b.sectors;
  70         return new;
  71 }
  72
  73 static int bch2_bucket_is_movable(struct btree_trans *trans,
  74                                   struct move_bucket *b, u64 time)
  75 {
  76         struct bch_fs *c = trans->c;
  77         struct btree_iter iter;
  78         struct bkey_s_c k;
  79         struct bch_alloc_v4 _a;
  80         const struct bch_alloc_v4 *a;
  81         int ret;
  82
  83         if (bch2_bucket_is_open(trans->c,
  84                                 b->k.bucket.inode,
  85                                 b->k.bucket.offset))
  86                 return 0;
  87
  88         k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_alloc,
  89                                b->k.bucket, BTREE_ITER_cached);
  90         ret = bkey_err(k);
  91         if (ret)
  92                 return ret;
  93
  94         struct bch_dev *ca = bch2_dev_tryget(c, k.k->p.inode);
  95         if (!ca)
  96                 goto out;
  97
  98         a = bch2_alloc_to_v4(k, &_a);
  99         b->k.gen        = a->gen;
 100         b->sectors      = bch2_bucket_sectors_dirty(*a);
 101         u64 lru_idx     = alloc_lru_idx_fragmentation(*a, ca);
 102
 103         ret = lru_idx && lru_idx <= time;
 104
 105         bch2_dev_put(ca);
 106 out:
 107         bch2_trans_iter_exit(trans, &iter);
 108         return ret;
 109 }
 110
 111 static void move_buckets_wait(struct moving_context *ctxt,
 112                               struct buckets_in_flight *list,
 113                               bool flush)
 114 {
 115         struct move_bucket_in_flight *i;
 116         int ret;
 117
 118         while ((i = list->first)) {
 119                 if (flush)
 120                         move_ctxt_wait_event(ctxt, !atomic_read(&i->count));
 121
 122                 if (atomic_read(&i->count))
 123                         break;
 124
 125                 list->first = i->next;
 126                 if (!list->first)
 127                         list->last = NULL;
 128
 129                 list->nr--;
 130                 list->sectors -= i->bucket.sectors;
 131
 132                 ret = rhashtable_remove_fast(&list->table, &i->hash,
 133                                              bch_move_bucket_params);
 134                 BUG_ON(ret);
 135                 kfree(i);
 136         }
 137
 138         bch2_trans_unlock_long(ctxt->trans);
 139 }
 140
 141 static bool bucket_in_flight(struct buckets_in_flight *list,
 142                              struct move_bucket_key k)
 143 {
 144         return rhashtable_lookup_fast(&list->table, &k, bch_move_bucket_params);
 145 }
 146
 147 typedef DARRAY(struct move_bucket) move_buckets;
 148
 149 static int bch2_copygc_get_buckets(struct moving_context *ctxt,
 150                         struct buckets_in_flight *buckets_in_flight,
 151                         move_buckets *buckets)
 152 {
 153         struct btree_trans *trans = ctxt->trans;
 154         struct bch_fs *c = trans->c;
 155         size_t nr_to_get = max_t(size_t, 16U, buckets_in_flight->nr / 4);
 156         size_t saw = 0, in_flight = 0, not_movable = 0, sectors = 0;
 157         int ret;
 158
 159         move_buckets_wait(ctxt, buckets_in_flight, false);
 160
 161         ret = bch2_btree_write_buffer_tryflush(trans);
 162         if (bch2_err_matches(ret, EROFS))
 163                 return ret;
 164
 165         if (bch2_fs_fatal_err_on(ret, c, "%s: from bch2_btree_write_buffer_tryflush()", bch2_err_str(ret)))
 166                 return ret;
 167
 168         bch2_trans_begin(trans);
 169
 170         ret = for_each_btree_key_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 ret2 = 0;
 176
 177                 saw++;
 178
 179                 ret2 = bch2_bucket_is_movable(trans, &b, lru_pos_time(k.k->p));
 180                 if (ret2 < 0)
 181                         goto err;
 182
 183                 if (!ret2)
 184                         not_movable++;
 185                 else if (bucket_in_flight(buckets_in_flight, b.k))
 186                         in_flight++;
 187                 else {
 188                         ret2 = darray_push(buckets, b);
 189                         if (ret2)
 190                                 goto err;
 191                         sectors += b.sectors;
 192                 }
 193
 194                 ret2 = buckets->nr >= nr_to_get;
 195 err:
 196                 ret2;
 197         }));
 198
 199         pr_debug("have: %zu (%zu) saw %zu in flight %zu not movable %zu got %zu (%zu)/%zu buckets ret %i",
 200                  buckets_in_flight->nr, buckets_in_flight->sectors,
 201                  saw, in_flight, not_movable, buckets->nr, sectors, nr_to_get, ret);
 202
 203         return ret < 0 ? ret : 0;
 204 }
 205
 206 noinline
 207 static int bch2_copygc(struct moving_context *ctxt,
 208                        struct buckets_in_flight *buckets_in_flight,
 209                        bool *did_work)
 210 {
 211         struct btree_trans *trans = ctxt->trans;
 212         struct bch_fs *c = trans->c;
 213         struct data_update_opts data_opts = {
 214                 .btree_insert_flags = BCH_WATERMARK_copygc,
 215         };
 216         move_buckets buckets = { 0 };
 217         struct move_bucket_in_flight *f;
 218         u64 moved = atomic64_read(&ctxt->stats->sectors_moved);
 219         int ret = 0;
 220
 221         ret = bch2_copygc_get_buckets(ctxt, buckets_in_flight, &buckets);
 222         if (ret)
 223                 goto err;
 224
 225         darray_for_each(buckets, i) {
 226                 if (kthread_should_stop() || freezing(current))
 227                         break;
 228
 229                 f = move_bucket_in_flight_add(buckets_in_flight, *i);
 230                 ret = PTR_ERR_OR_ZERO(f);
 231                 if (ret == -EEXIST) { /* rare race: copygc_get_buckets returned same bucket more than once */
 232                         ret = 0;
 233                         continue;
 234                 }
 235                 if (ret == -ENOMEM) { /* flush IO, continue later */
 236                         ret = 0;
 237                         break;
 238                 }
 239
 240                 ret = bch2_evacuate_bucket(ctxt, f, f->bucket.k.bucket,
 241                                              f->bucket.k.gen, data_opts);
 242                 if (ret)
 243                         goto err;
 244
 245                 *did_work = true;
 246         }
 247 err:
 248         darray_exit(&buckets);
 249
 250         /* no entries in LRU btree found, or got to end: */
 251         if (bch2_err_matches(ret, ENOENT))
 252                 ret = 0;
 253
 254         if (ret < 0 && !bch2_err_matches(ret, EROFS))
 255                 bch_err_msg(c, ret, "from bch2_move_data()");
 256
 257         moved = atomic64_read(&ctxt->stats->sectors_moved) - moved;
 258         trace_and_count(c, copygc, c, moved, 0, 0, 0);
 259         return ret;
 260 }
 261
 262 /*
 263  * Copygc runs when the amount of fragmented data is above some arbitrary
 264  * threshold:
 265  *
 266  * The threshold at the limit - when the device is full - is the amount of space
 267  * we reserved in bch2_recalc_capacity; we can't have more than that amount of
 268  * disk space stranded due to fragmentation and store everything we have
 269  * promised to store.
 270  *
 271  * But we don't want to be running copygc unnecessarily when the device still
 272  * has plenty of free space - rather, we want copygc to smoothly run every so
 273  * often and continually reduce the amount of fragmented space as the device
 274  * fills up. So, we increase the threshold by half the current free space.
 275  */
 276 unsigned long bch2_copygc_wait_amount(struct bch_fs *c)
 277 {
 278         s64 wait = S64_MAX, fragmented_allowed, fragmented;
 279
 280         for_each_rw_member(c, ca) {
 281                 struct bch_dev_usage usage = bch2_dev_usage_read(ca);
 282
 283                 fragmented_allowed = ((__dev_buckets_available(ca, usage, BCH_WATERMARK_stripe) *
 284                                        ca->mi.bucket_size) >> 1);
 285                 fragmented = 0;
 286
 287                 for (unsigned i = 0; i < BCH_DATA_NR; i++)
 288                         if (data_type_movable(i))
 289                                 fragmented += usage.d[i].fragmented;
 290
 291                 wait = min(wait, max(0LL, fragmented_allowed - fragmented));
 292         }
 293
 294         return wait;
 295 }
 296
 297 void bch2_copygc_wait_to_text(struct printbuf *out, struct bch_fs *c)
 298 {
 299         printbuf_tabstop_push(out, 32);
 300         prt_printf(out, "running:\t%u\n",               c->copygc_running);
 301         prt_printf(out, "copygc_wait:\t%llu\n",         c->copygc_wait);
 302         prt_printf(out, "copygc_wait_at:\t%llu\n",      c->copygc_wait_at);
 303
 304         prt_printf(out, "Currently waiting for:\t");
 305         prt_human_readable_u64(out, max(0LL, c->copygc_wait -
 306                                         atomic64_read(&c->io_clock[WRITE].now)) << 9);
 307         prt_newline(out);
 308
 309         prt_printf(out, "Currently waiting since:\t");
 310         prt_human_readable_u64(out, max(0LL,
 311                                         atomic64_read(&c->io_clock[WRITE].now) -
 312                                         c->copygc_wait_at) << 9);
 313         prt_newline(out);
 314
 315         prt_printf(out, "Currently calculated wait:\t");
 316         prt_human_readable_u64(out, bch2_copygc_wait_amount(c));
 317         prt_newline(out);
 318 }
 319
 320 static int bch2_copygc_thread(void *arg)
 321 {
 322         struct bch_fs *c = arg;
 323         struct moving_context ctxt;
 324         struct bch_move_stats move_stats;
 325         struct io_clock *clock = &c->io_clock[WRITE];
 326         struct buckets_in_flight *buckets;
 327         u64 last, wait;
 328         int ret = 0;
 329
 330         buckets = kzalloc(sizeof(struct buckets_in_flight), GFP_KERNEL);
 331         if (!buckets)
 332                 return -ENOMEM;
 333         ret = rhashtable_init(&buckets->table, &bch_move_bucket_params);
 334         bch_err_msg(c, ret, "allocating copygc buckets in flight");
 335         if (ret) {
 336                 kfree(buckets);
 337                 return ret;
 338         }
 339
 340         set_freezable();
 341
 342         bch2_move_stats_init(&move_stats, "copygc");
 343         bch2_moving_ctxt_init(&ctxt, c, NULL, &move_stats,
 344                               writepoint_ptr(&c->copygc_write_point),
 345                               false);
 346
 347         while (!ret && !kthread_should_stop()) {
 348                 bool did_work = false;
 349
 350                 bch2_trans_unlock_long(ctxt.trans);
 351                 cond_resched();
 352
 353                 if (!c->copy_gc_enabled) {
 354                         move_buckets_wait(&ctxt, buckets, true);
 355                         kthread_wait_freezable(c->copy_gc_enabled ||
 356                                                kthread_should_stop());
 357                 }
 358
 359                 if (unlikely(freezing(current))) {
 360                         move_buckets_wait(&ctxt, 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(&ctxt, 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(&ctxt, buckets, &did_work);
 382                 c->copygc_running = false;
 383
 384                 wake_up(&c->copygc_running_wq);
 385
 386                 if (!wait && !did_work) {
 387                         u64 min_member_capacity = bch2_min_rw_member_capacity(c);
 388
 389                         if (min_member_capacity == U64_MAX)
 390                                 min_member_capacity = 128 * 2048;
 391
 392                         move_buckets_wait(&ctxt, buckets, true);
 393                         bch2_kthread_io_clock_wait(clock, last + (min_member_capacity >> 6),
 394                                         MAX_SCHEDULE_TIMEOUT);
 395                 }
 396         }
 397
 398         move_buckets_wait(&ctxt, buckets, true);
 399
 400         rhashtable_destroy(&buckets->table);
 401         kfree(buckets);
 402         bch2_moving_ctxt_exit(&ctxt);
 403         bch2_move_stats_exit(&move_stats, c);
 404
 405         return 0;
 406 }
 407
 408 void bch2_copygc_stop(struct bch_fs *c)
 409 {
 410         if (c->copygc_thread) {
 411                 kthread_stop(c->copygc_thread);
 412                 put_task_struct(c->copygc_thread);
 413         }
 414         c->copygc_thread = NULL;
 415 }
 416
 417 int bch2_copygc_start(struct bch_fs *c)
 418 {
 419         struct task_struct *t;
 420         int ret;
 421
 422         if (c->copygc_thread)
 423                 return 0;
 424
 425         if (c->opts.nochanges)
 426                 return 0;
 427
 428         if (bch2_fs_init_fault("copygc_start"))
 429                 return -ENOMEM;
 430
 431         t = kthread_create(bch2_copygc_thread, c, "bch-copygc/%s", c->name);
 432         ret = PTR_ERR_OR_ZERO(t);
 433         bch_err_msg(c, ret, "creating copygc thread");
 434         if (ret)
 435                 return ret;
 436
 437         get_task_struct(t);
 438
 439         c->copygc_thread = t;
 440         wake_up_process(c->copygc_thread);
 441
 442         return 0;
 443 }
 444
 445 void bch2_fs_copygc_init(struct bch_fs *c)
 446 {
 447         init_waitqueue_head(&c->copygc_running_wq);
 448         c->copygc_running = false;
 449 }