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