git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/ec.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 /* erasure coding */
   4
   5 #include "bcachefs.h"
   6 #include "alloc_foreground.h"
   7 #include "backpointers.h"
   8 #include "bkey_buf.h"
   9 #include "bset.h"
  10 #include "btree_gc.h"
  11 #include "btree_update.h"
  12 #include "btree_write_buffer.h"
  13 #include "buckets.h"
  14 #include "checksum.h"
  15 #include "disk_groups.h"
  16 #include "ec.h"
  17 #include "error.h"
  18 #include "io_read.h"
  19 #include "keylist.h"
  20 #include "recovery.h"
  21 #include "replicas.h"
  22 #include "super-io.h"
  23 #include "util.h"
  24
  25 #include <linux/sort.h>
  26
  27 #ifdef __KERNEL__
  28
  29 #include <linux/raid/pq.h>
  30 #include <linux/raid/xor.h>
  31
  32 static void raid5_recov(unsigned disks, unsigned failed_idx,
  33                         size_t size, void **data)
  34 {
  35         unsigned i = 2, nr;
  36
  37         BUG_ON(failed_idx >= disks);
  38
  39         swap(data[0], data[failed_idx]);
  40         memcpy(data[0], data[1], size);
  41
  42         while (i < disks) {
  43                 nr = min_t(unsigned, disks - i, MAX_XOR_BLOCKS);
  44                 xor_blocks(nr, size, data[0], data + i);
  45                 i += nr;
  46         }
  47
  48         swap(data[0], data[failed_idx]);
  49 }
  50
  51 static void raid_gen(int nd, int np, size_t size, void **v)
  52 {
  53         if (np >= 1)
  54                 raid5_recov(nd + np, nd, size, v);
  55         if (np >= 2)
  56                 raid6_call.gen_syndrome(nd + np, size, v);
  57         BUG_ON(np > 2);
  58 }
  59
  60 static void raid_rec(int nr, int *ir, int nd, int np, size_t size, void **v)
  61 {
  62         switch (nr) {
  63         case 0:
  64                 break;
  65         case 1:
  66                 if (ir[0] < nd + 1)
  67                         raid5_recov(nd + 1, ir[0], size, v);
  68                 else
  69                         raid6_call.gen_syndrome(nd + np, size, v);
  70                 break;
  71         case 2:
  72                 if (ir[1] < nd) {
  73                         /* data+data failure. */
  74                         raid6_2data_recov(nd + np, size, ir[0], ir[1], v);
  75                 } else if (ir[0] < nd) {
  76                         /* data + p/q failure */
  77
  78                         if (ir[1] == nd) /* data + p failure */
  79                                 raid6_datap_recov(nd + np, size, ir[0], v);
  80                         else { /* data + q failure */
  81                                 raid5_recov(nd + 1, ir[0], size, v);
  82                                 raid6_call.gen_syndrome(nd + np, size, v);
  83                         }
  84                 } else {
  85                         raid_gen(nd, np, size, v);
  86                 }
  87                 break;
  88         default:
  89                 BUG();
  90         }
  91 }
  92
  93 #else
  94
  95 #include <raid/raid.h>
  96
  97 #endif
  98
  99 struct ec_bio {
 100         struct bch_dev          *ca;
 101         struct ec_stripe_buf    *buf;
 102         size_t                  idx;
 103         struct bio              bio;
 104 };
 105
 106 /* Stripes btree keys: */
 107
 108 int bch2_stripe_invalid(struct bch_fs *c, struct bkey_s_c k,
 109                         enum bkey_invalid_flags flags,
 110                         struct printbuf *err)
 111 {
 112         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 113         int ret = 0;
 114
 115         bkey_fsck_err_on(bkey_eq(k.k->p, POS_MIN) ||
 116                          bpos_gt(k.k->p, POS(0, U32_MAX)), c, err,
 117                          stripe_pos_bad,
 118                          "stripe at bad pos");
 119
 120         bkey_fsck_err_on(bkey_val_u64s(k.k) < stripe_val_u64s(s), c, err,
 121                          stripe_val_size_bad,
 122                          "incorrect value size (%zu < %u)",
 123                          bkey_val_u64s(k.k), stripe_val_u64s(s));
 124
 125         ret = bch2_bkey_ptrs_invalid(c, k, flags, err);
 126 fsck_err:
 127         return ret;
 128 }
 129
 130 void bch2_stripe_to_text(struct printbuf *out, struct bch_fs *c,
 131                          struct bkey_s_c k)
 132 {
 133         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 134         unsigned i, nr_data = s->nr_blocks - s->nr_redundant;
 135
 136         prt_printf(out, "algo %u sectors %u blocks %u:%u csum %u gran %u",
 137                s->algorithm,
 138                le16_to_cpu(s->sectors),
 139                nr_data,
 140                s->nr_redundant,
 141                s->csum_type,
 142                1U << s->csum_granularity_bits);
 143
 144         for (i = 0; i < s->nr_blocks; i++) {
 145                 const struct bch_extent_ptr *ptr = s->ptrs + i;
 146                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 147                 u32 offset;
 148                 u64 b = sector_to_bucket_and_offset(ca, ptr->offset, &offset);
 149
 150                 prt_printf(out, " %u:%llu:%u", ptr->dev, b, offset);
 151                 if (i < nr_data)
 152                         prt_printf(out, "#%u", stripe_blockcount_get(s, i));
 153                 prt_printf(out, " gen %u", ptr->gen);
 154                 if (ptr_stale(ca, ptr))
 155                         prt_printf(out, " stale");
 156         }
 157 }
 158
 159 /* returns blocknr in stripe that we matched: */
 160 static const struct bch_extent_ptr *bkey_matches_stripe(struct bch_stripe *s,
 161                                                 struct bkey_s_c k, unsigned *block)
 162 {
 163         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 164         const struct bch_extent_ptr *ptr;
 165         unsigned i, nr_data = s->nr_blocks - s->nr_redundant;
 166
 167         bkey_for_each_ptr(ptrs, ptr)
 168                 for (i = 0; i < nr_data; i++)
 169                         if (__bch2_ptr_matches_stripe(&s->ptrs[i], ptr,
 170                                                       le16_to_cpu(s->sectors))) {
 171                                 *block = i;
 172                                 return ptr;
 173                         }
 174
 175         return NULL;
 176 }
 177
 178 static bool extent_has_stripe_ptr(struct bkey_s_c k, u64 idx)
 179 {
 180         switch (k.k->type) {
 181         case KEY_TYPE_extent: {
 182                 struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
 183                 const union bch_extent_entry *entry;
 184
 185                 extent_for_each_entry(e, entry)
 186                         if (extent_entry_type(entry) ==
 187                             BCH_EXTENT_ENTRY_stripe_ptr &&
 188                             entry->stripe_ptr.idx == idx)
 189                                 return true;
 190
 191                 break;
 192         }
 193         }
 194
 195         return false;
 196 }
 197
 198 /* Stripe bufs: */
 199
 200 static void ec_stripe_buf_exit(struct ec_stripe_buf *buf)
 201 {
 202         if (buf->key.k.type == KEY_TYPE_stripe) {
 203                 struct bkey_i_stripe *s = bkey_i_to_stripe(&buf->key);
 204                 unsigned i;
 205
 206                 for (i = 0; i < s->v.nr_blocks; i++) {
 207                         kvpfree(buf->data[i], buf->size << 9);
 208                         buf->data[i] = NULL;
 209                 }
 210         }
 211 }
 212
 213 /* XXX: this is a non-mempoolified memory allocation: */
 214 static int ec_stripe_buf_init(struct ec_stripe_buf *buf,
 215                               unsigned offset, unsigned size)
 216 {
 217         struct bch_stripe *v = &bkey_i_to_stripe(&buf->key)->v;
 218         unsigned csum_granularity = 1U << v->csum_granularity_bits;
 219         unsigned end = offset + size;
 220         unsigned i;
 221
 222         BUG_ON(end > le16_to_cpu(v->sectors));
 223
 224         offset  = round_down(offset, csum_granularity);
 225         end     = min_t(unsigned, le16_to_cpu(v->sectors),
 226                         round_up(end, csum_granularity));
 227
 228         buf->offset     = offset;
 229         buf->size       = end - offset;
 230
 231         memset(buf->valid, 0xFF, sizeof(buf->valid));
 232
 233         for (i = 0; i < v->nr_blocks; i++) {
 234                 buf->data[i] = kvpmalloc(buf->size << 9, GFP_KERNEL);
 235                 if (!buf->data[i])
 236                         goto err;
 237         }
 238
 239         return 0;
 240 err:
 241         ec_stripe_buf_exit(buf);
 242         return -BCH_ERR_ENOMEM_stripe_buf;
 243 }
 244
 245 /* Checksumming: */
 246
 247 static struct bch_csum ec_block_checksum(struct ec_stripe_buf *buf,
 248                                          unsigned block, unsigned offset)
 249 {
 250         struct bch_stripe *v = &bkey_i_to_stripe(&buf->key)->v;
 251         unsigned csum_granularity = 1 << v->csum_granularity_bits;
 252         unsigned end = buf->offset + buf->size;
 253         unsigned len = min(csum_granularity, end - offset);
 254
 255         BUG_ON(offset >= end);
 256         BUG_ON(offset <  buf->offset);
 257         BUG_ON(offset & (csum_granularity - 1));
 258         BUG_ON(offset + len != le16_to_cpu(v->sectors) &&
 259                (len & (csum_granularity - 1)));
 260
 261         return bch2_checksum(NULL, v->csum_type,
 262                              null_nonce(),
 263                              buf->data[block] + ((offset - buf->offset) << 9),
 264                              len << 9);
 265 }
 266
 267 static void ec_generate_checksums(struct ec_stripe_buf *buf)
 268 {
 269         struct bch_stripe *v = &bkey_i_to_stripe(&buf->key)->v;
 270         unsigned i, j, csums_per_device = stripe_csums_per_device(v);
 271
 272         if (!v->csum_type)
 273                 return;
 274
 275         BUG_ON(buf->offset);
 276         BUG_ON(buf->size != le16_to_cpu(v->sectors));
 277
 278         for (i = 0; i < v->nr_blocks; i++)
 279                 for (j = 0; j < csums_per_device; j++)
 280                         stripe_csum_set(v, i, j,
 281                                 ec_block_checksum(buf, i, j << v->csum_granularity_bits));
 282 }
 283
 284 static void ec_validate_checksums(struct bch_fs *c, struct ec_stripe_buf *buf)
 285 {
 286         struct bch_stripe *v = &bkey_i_to_stripe(&buf->key)->v;
 287         unsigned csum_granularity = 1 << v->csum_granularity_bits;
 288         unsigned i;
 289
 290         if (!v->csum_type)
 291                 return;
 292
 293         for (i = 0; i < v->nr_blocks; i++) {
 294                 unsigned offset = buf->offset;
 295                 unsigned end = buf->offset + buf->size;
 296
 297                 if (!test_bit(i, buf->valid))
 298                         continue;
 299
 300                 while (offset < end) {
 301                         unsigned j = offset >> v->csum_granularity_bits;
 302                         unsigned len = min(csum_granularity, end - offset);
 303                         struct bch_csum want = stripe_csum_get(v, i, j);
 304                         struct bch_csum got = ec_block_checksum(buf, i, offset);
 305
 306                         if (bch2_crc_cmp(want, got)) {
 307                                 struct printbuf err = PRINTBUF;
 308                                 struct bch_dev *ca = bch_dev_bkey_exists(c, v->ptrs[i].dev);
 309
 310                                 prt_printf(&err, "stripe checksum error: expected %0llx:%0llx got %0llx:%0llx (type %s)\n",
 311                                            want.hi, want.lo,
 312                                            got.hi, got.lo,
 313                                            bch2_csum_types[v->csum_type]);
 314                                 prt_printf(&err, "  for %ps at %u of\n  ", (void *) _RET_IP_, i);
 315                                 bch2_bkey_val_to_text(&err, c, bkey_i_to_s_c(&buf->key));
 316                                 bch_err_ratelimited(ca, "%s", err.buf);
 317                                 printbuf_exit(&err);
 318
 319                                 clear_bit(i, buf->valid);
 320
 321                                 bch2_io_error(ca, BCH_MEMBER_ERROR_checksum);
 322                                 break;
 323                         }
 324
 325                         offset += len;
 326                 }
 327         }
 328 }
 329
 330 /* Erasure coding: */
 331
 332 static void ec_generate_ec(struct ec_stripe_buf *buf)
 333 {
 334         struct bch_stripe *v = &bkey_i_to_stripe(&buf->key)->v;
 335         unsigned nr_data = v->nr_blocks - v->nr_redundant;
 336         unsigned bytes = le16_to_cpu(v->sectors) << 9;
 337
 338         raid_gen(nr_data, v->nr_redundant, bytes, buf->data);
 339 }
 340
 341 static unsigned ec_nr_failed(struct ec_stripe_buf *buf)
 342 {
 343         struct bch_stripe *v = &bkey_i_to_stripe(&buf->key)->v;
 344
 345         return v->nr_blocks - bitmap_weight(buf->valid, v->nr_blocks);
 346 }
 347
 348 static int ec_do_recov(struct bch_fs *c, struct ec_stripe_buf *buf)
 349 {
 350         struct bch_stripe *v = &bkey_i_to_stripe(&buf->key)->v;
 351         unsigned i, failed[BCH_BKEY_PTRS_MAX], nr_failed = 0;
 352         unsigned nr_data = v->nr_blocks - v->nr_redundant;
 353         unsigned bytes = buf->size << 9;
 354
 355         if (ec_nr_failed(buf) > v->nr_redundant) {
 356                 bch_err_ratelimited(c,
 357                         "error doing reconstruct read: unable to read enough blocks");
 358                 return -1;
 359         }
 360
 361         for (i = 0; i < nr_data; i++)
 362                 if (!test_bit(i, buf->valid))
 363                         failed[nr_failed++] = i;
 364
 365         raid_rec(nr_failed, failed, nr_data, v->nr_redundant, bytes, buf->data);
 366         return 0;
 367 }
 368
 369 /* IO: */
 370
 371 static void ec_block_endio(struct bio *bio)
 372 {
 373         struct ec_bio *ec_bio = container_of(bio, struct ec_bio, bio);
 374         struct bch_stripe *v = &bkey_i_to_stripe(&ec_bio->buf->key)->v;
 375         struct bch_extent_ptr *ptr = &v->ptrs[ec_bio->idx];
 376         struct bch_dev *ca = ec_bio->ca;
 377         struct closure *cl = bio->bi_private;
 378
 379         if (bch2_dev_io_err_on(bio->bi_status, ca,
 380                                bio_data_dir(bio)
 381                                ? BCH_MEMBER_ERROR_write
 382                                : BCH_MEMBER_ERROR_read,
 383                                "erasure coding %s error: %s",
 384                                bio_data_dir(bio) ? "write" : "read",
 385                                bch2_blk_status_to_str(bio->bi_status)))
 386                 clear_bit(ec_bio->idx, ec_bio->buf->valid);
 387
 388         if (ptr_stale(ca, ptr)) {
 389                 bch_err_ratelimited(ca->fs,
 390                                     "error %s stripe: stale pointer after io",
 391                                     bio_data_dir(bio) == READ ? "reading from" : "writing to");
 392                 clear_bit(ec_bio->idx, ec_bio->buf->valid);
 393         }
 394
 395         bio_put(&ec_bio->bio);
 396         percpu_ref_put(&ca->io_ref);
 397         closure_put(cl);
 398 }
 399
 400 static void ec_block_io(struct bch_fs *c, struct ec_stripe_buf *buf,
 401                         blk_opf_t opf, unsigned idx, struct closure *cl)
 402 {
 403         struct bch_stripe *v = &bkey_i_to_stripe(&buf->key)->v;
 404         unsigned offset = 0, bytes = buf->size << 9;
 405         struct bch_extent_ptr *ptr = &v->ptrs[idx];
 406         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 407         enum bch_data_type data_type = idx < v->nr_blocks - v->nr_redundant
 408                 ? BCH_DATA_user
 409                 : BCH_DATA_parity;
 410         int rw = op_is_write(opf);
 411
 412         if (ptr_stale(ca, ptr)) {
 413                 bch_err_ratelimited(c,
 414                                     "error %s stripe: stale pointer",
 415                                     rw == READ ? "reading from" : "writing to");
 416                 clear_bit(idx, buf->valid);
 417                 return;
 418         }
 419
 420         if (!bch2_dev_get_ioref(ca, rw)) {
 421                 clear_bit(idx, buf->valid);
 422                 return;
 423         }
 424
 425         this_cpu_add(ca->io_done->sectors[rw][data_type], buf->size);
 426
 427         while (offset < bytes) {
 428                 unsigned nr_iovecs = min_t(size_t, BIO_MAX_VECS,
 429                                            DIV_ROUND_UP(bytes, PAGE_SIZE));
 430                 unsigned b = min_t(size_t, bytes - offset,
 431                                    nr_iovecs << PAGE_SHIFT);
 432                 struct ec_bio *ec_bio;
 433
 434                 ec_bio = container_of(bio_alloc_bioset(ca->disk_sb.bdev,
 435                                                        nr_iovecs,
 436                                                        opf,
 437                                                        GFP_KERNEL,
 438                                                        &c->ec_bioset),
 439                                       struct ec_bio, bio);
 440
 441                 ec_bio->ca                      = ca;
 442                 ec_bio->buf                     = buf;
 443                 ec_bio->idx                     = idx;
 444
 445                 ec_bio->bio.bi_iter.bi_sector   = ptr->offset + buf->offset + (offset >> 9);
 446                 ec_bio->bio.bi_end_io           = ec_block_endio;
 447                 ec_bio->bio.bi_private          = cl;
 448
 449                 bch2_bio_map(&ec_bio->bio, buf->data[idx] + offset, b);
 450
 451                 closure_get(cl);
 452                 percpu_ref_get(&ca->io_ref);
 453
 454                 submit_bio(&ec_bio->bio);
 455
 456                 offset += b;
 457         }
 458
 459         percpu_ref_put(&ca->io_ref);
 460 }
 461
 462 static int get_stripe_key_trans(struct btree_trans *trans, u64 idx,
 463                                 struct ec_stripe_buf *stripe)
 464 {
 465         struct btree_iter iter;
 466         struct bkey_s_c k;
 467         int ret;
 468
 469         k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_stripes,
 470                                POS(0, idx), BTREE_ITER_SLOTS);
 471         ret = bkey_err(k);
 472         if (ret)
 473                 goto err;
 474         if (k.k->type != KEY_TYPE_stripe) {
 475                 ret = -ENOENT;
 476                 goto err;
 477         }
 478         bkey_reassemble(&stripe->key, k);
 479 err:
 480         bch2_trans_iter_exit(trans, &iter);
 481         return ret;
 482 }
 483
 484 /* recovery read path: */
 485 int bch2_ec_read_extent(struct btree_trans *trans, struct bch_read_bio *rbio)
 486 {
 487         struct bch_fs *c = trans->c;
 488         struct ec_stripe_buf *buf;
 489         struct closure cl;
 490         struct bch_stripe *v;
 491         unsigned i, offset;
 492         int ret = 0;
 493
 494         closure_init_stack(&cl);
 495
 496         BUG_ON(!rbio->pick.has_ec);
 497
 498         buf = kzalloc(sizeof(*buf), GFP_NOFS);
 499         if (!buf)
 500                 return -BCH_ERR_ENOMEM_ec_read_extent;
 501
 502         ret = lockrestart_do(trans, get_stripe_key_trans(trans, rbio->pick.ec.idx, buf));
 503         if (ret) {
 504                 bch_err_ratelimited(c,
 505                         "error doing reconstruct read: error %i looking up stripe", ret);
 506                 kfree(buf);
 507                 return -EIO;
 508         }
 509
 510         v = &bkey_i_to_stripe(&buf->key)->v;
 511
 512         if (!bch2_ptr_matches_stripe(v, rbio->pick)) {
 513                 bch_err_ratelimited(c,
 514                         "error doing reconstruct read: pointer doesn't match stripe");
 515                 ret = -EIO;
 516                 goto err;
 517         }
 518
 519         offset = rbio->bio.bi_iter.bi_sector - v->ptrs[rbio->pick.ec.block].offset;
 520         if (offset + bio_sectors(&rbio->bio) > le16_to_cpu(v->sectors)) {
 521                 bch_err_ratelimited(c,
 522                         "error doing reconstruct read: read is bigger than stripe");
 523                 ret = -EIO;
 524                 goto err;
 525         }
 526
 527         ret = ec_stripe_buf_init(buf, offset, bio_sectors(&rbio->bio));
 528         if (ret)
 529                 goto err;
 530
 531         for (i = 0; i < v->nr_blocks; i++)
 532                 ec_block_io(c, buf, REQ_OP_READ, i, &cl);
 533
 534         closure_sync(&cl);
 535
 536         if (ec_nr_failed(buf) > v->nr_redundant) {
 537                 bch_err_ratelimited(c,
 538                         "error doing reconstruct read: unable to read enough blocks");
 539                 ret = -EIO;
 540                 goto err;
 541         }
 542
 543         ec_validate_checksums(c, buf);
 544
 545         ret = ec_do_recov(c, buf);
 546         if (ret)
 547                 goto err;
 548
 549         memcpy_to_bio(&rbio->bio, rbio->bio.bi_iter,
 550                       buf->data[rbio->pick.ec.block] + ((offset - buf->offset) << 9));
 551 err:
 552         ec_stripe_buf_exit(buf);
 553         kfree(buf);
 554         return ret;
 555 }
 556
 557 /* stripe bucket accounting: */
 558
 559 static int __ec_stripe_mem_alloc(struct bch_fs *c, size_t idx, gfp_t gfp)
 560 {
 561         ec_stripes_heap n, *h = &c->ec_stripes_heap;
 562
 563         if (idx >= h->size) {
 564                 if (!init_heap(&n, max(1024UL, roundup_pow_of_two(idx + 1)), gfp))
 565                         return -BCH_ERR_ENOMEM_ec_stripe_mem_alloc;
 566
 567                 mutex_lock(&c->ec_stripes_heap_lock);
 568                 if (n.size > h->size) {
 569                         memcpy(n.data, h->data, h->used * sizeof(h->data[0]));
 570                         n.used = h->used;
 571                         swap(*h, n);
 572                 }
 573                 mutex_unlock(&c->ec_stripes_heap_lock);
 574
 575                 free_heap(&n);
 576         }
 577
 578         if (!genradix_ptr_alloc(&c->stripes, idx, gfp))
 579                 return -BCH_ERR_ENOMEM_ec_stripe_mem_alloc;
 580
 581         if (c->gc_pos.phase != GC_PHASE_NOT_RUNNING &&
 582             !genradix_ptr_alloc(&c->gc_stripes, idx, gfp))
 583                 return -BCH_ERR_ENOMEM_ec_stripe_mem_alloc;
 584
 585         return 0;
 586 }
 587
 588 static int ec_stripe_mem_alloc(struct btree_trans *trans,
 589                                struct btree_iter *iter)
 590 {
 591         return allocate_dropping_locks_errcode(trans,
 592                         __ec_stripe_mem_alloc(trans->c, iter->pos.offset, _gfp));
 593 }
 594
 595 /*
 596  * Hash table of open stripes:
 597  * Stripes that are being created or modified are kept in a hash table, so that
 598  * stripe deletion can skip them.
 599  */
 600
 601 static bool __bch2_stripe_is_open(struct bch_fs *c, u64 idx)
 602 {
 603         unsigned hash = hash_64(idx, ilog2(ARRAY_SIZE(c->ec_stripes_new)));
 604         struct ec_stripe_new *s;
 605
 606         hlist_for_each_entry(s, &c->ec_stripes_new[hash], hash)
 607                 if (s->idx == idx)
 608                         return true;
 609         return false;
 610 }
 611
 612 static bool bch2_stripe_is_open(struct bch_fs *c, u64 idx)
 613 {
 614         bool ret = false;
 615
 616         spin_lock(&c->ec_stripes_new_lock);
 617         ret = __bch2_stripe_is_open(c, idx);
 618         spin_unlock(&c->ec_stripes_new_lock);
 619
 620         return ret;
 621 }
 622
 623 static bool bch2_try_open_stripe(struct bch_fs *c,
 624                                  struct ec_stripe_new *s,
 625                                  u64 idx)
 626 {
 627         bool ret;
 628
 629         spin_lock(&c->ec_stripes_new_lock);
 630         ret = !__bch2_stripe_is_open(c, idx);
 631         if (ret) {
 632                 unsigned hash = hash_64(idx, ilog2(ARRAY_SIZE(c->ec_stripes_new)));
 633
 634                 s->idx = idx;
 635                 hlist_add_head(&s->hash, &c->ec_stripes_new[hash]);
 636         }
 637         spin_unlock(&c->ec_stripes_new_lock);
 638
 639         return ret;
 640 }
 641
 642 static void bch2_stripe_close(struct bch_fs *c, struct ec_stripe_new *s)
 643 {
 644         BUG_ON(!s->idx);
 645
 646         spin_lock(&c->ec_stripes_new_lock);
 647         hlist_del_init(&s->hash);
 648         spin_unlock(&c->ec_stripes_new_lock);
 649
 650         s->idx = 0;
 651 }
 652
 653 /* Heap of all existing stripes, ordered by blocks_nonempty */
 654
 655 static u64 stripe_idx_to_delete(struct bch_fs *c)
 656 {
 657         ec_stripes_heap *h = &c->ec_stripes_heap;
 658
 659         lockdep_assert_held(&c->ec_stripes_heap_lock);
 660
 661         if (h->used &&
 662             h->data[0].blocks_nonempty == 0 &&
 663             !bch2_stripe_is_open(c, h->data[0].idx))
 664                 return h->data[0].idx;
 665
 666         return 0;
 667 }
 668
 669 static inline int ec_stripes_heap_cmp(ec_stripes_heap *h,
 670                                       struct ec_stripe_heap_entry l,
 671                                       struct ec_stripe_heap_entry r)
 672 {
 673         return ((l.blocks_nonempty > r.blocks_nonempty) -
 674                 (l.blocks_nonempty < r.blocks_nonempty));
 675 }
 676
 677 static inline void ec_stripes_heap_set_backpointer(ec_stripes_heap *h,
 678                                                    size_t i)
 679 {
 680         struct bch_fs *c = container_of(h, struct bch_fs, ec_stripes_heap);
 681
 682         genradix_ptr(&c->stripes, h->data[i].idx)->heap_idx = i;
 683 }
 684
 685 static void heap_verify_backpointer(struct bch_fs *c, size_t idx)
 686 {
 687         ec_stripes_heap *h = &c->ec_stripes_heap;
 688         struct stripe *m = genradix_ptr(&c->stripes, idx);
 689
 690         BUG_ON(m->heap_idx >= h->used);
 691         BUG_ON(h->data[m->heap_idx].idx != idx);
 692 }
 693
 694 void bch2_stripes_heap_del(struct bch_fs *c,
 695                            struct stripe *m, size_t idx)
 696 {
 697         mutex_lock(&c->ec_stripes_heap_lock);
 698         heap_verify_backpointer(c, idx);
 699
 700         heap_del(&c->ec_stripes_heap, m->heap_idx,
 701                  ec_stripes_heap_cmp,
 702                  ec_stripes_heap_set_backpointer);
 703         mutex_unlock(&c->ec_stripes_heap_lock);
 704 }
 705
 706 void bch2_stripes_heap_insert(struct bch_fs *c,
 707                               struct stripe *m, size_t idx)
 708 {
 709         mutex_lock(&c->ec_stripes_heap_lock);
 710         BUG_ON(heap_full(&c->ec_stripes_heap));
 711
 712         heap_add(&c->ec_stripes_heap, ((struct ec_stripe_heap_entry) {
 713                         .idx = idx,
 714                         .blocks_nonempty = m->blocks_nonempty,
 715                 }),
 716                  ec_stripes_heap_cmp,
 717                  ec_stripes_heap_set_backpointer);
 718
 719         heap_verify_backpointer(c, idx);
 720         mutex_unlock(&c->ec_stripes_heap_lock);
 721 }
 722
 723 void bch2_stripes_heap_update(struct bch_fs *c,
 724                               struct stripe *m, size_t idx)
 725 {
 726         ec_stripes_heap *h = &c->ec_stripes_heap;
 727         bool do_deletes;
 728         size_t i;
 729
 730         mutex_lock(&c->ec_stripes_heap_lock);
 731         heap_verify_backpointer(c, idx);
 732
 733         h->data[m->heap_idx].blocks_nonempty = m->blocks_nonempty;
 734
 735         i = m->heap_idx;
 736         heap_sift_up(h,   i, ec_stripes_heap_cmp,
 737                      ec_stripes_heap_set_backpointer);
 738         heap_sift_down(h, i, ec_stripes_heap_cmp,
 739                        ec_stripes_heap_set_backpointer);
 740
 741         heap_verify_backpointer(c, idx);
 742
 743         do_deletes = stripe_idx_to_delete(c) != 0;
 744         mutex_unlock(&c->ec_stripes_heap_lock);
 745
 746         if (do_deletes)
 747                 bch2_do_stripe_deletes(c);
 748 }
 749
 750 /* stripe deletion */
 751
 752 static int ec_stripe_delete(struct btree_trans *trans, u64 idx)
 753 {
 754         struct bch_fs *c = trans->c;
 755         struct btree_iter iter;
 756         struct bkey_s_c k;
 757         struct bkey_s_c_stripe s;
 758         int ret;
 759
 760         k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_stripes, POS(0, idx),
 761                                BTREE_ITER_INTENT);
 762         ret = bkey_err(k);
 763         if (ret)
 764                 goto err;
 765
 766         if (k.k->type != KEY_TYPE_stripe) {
 767                 bch2_fs_inconsistent(c, "attempting to delete nonexistent stripe %llu", idx);
 768                 ret = -EINVAL;
 769                 goto err;
 770         }
 771
 772         s = bkey_s_c_to_stripe(k);
 773         for (unsigned i = 0; i < s.v->nr_blocks; i++)
 774                 if (stripe_blockcount_get(s.v, i)) {
 775                         struct printbuf buf = PRINTBUF;
 776
 777                         bch2_bkey_val_to_text(&buf, c, k);
 778                         bch2_fs_inconsistent(c, "attempting to delete nonempty stripe %s", buf.buf);
 779                         printbuf_exit(&buf);
 780                         ret = -EINVAL;
 781                         goto err;
 782                 }
 783
 784         ret = bch2_btree_delete_at(trans, &iter, 0);
 785 err:
 786         bch2_trans_iter_exit(trans, &iter);
 787         return ret;
 788 }
 789
 790 static void ec_stripe_delete_work(struct work_struct *work)
 791 {
 792         struct bch_fs *c =
 793                 container_of(work, struct bch_fs, ec_stripe_delete_work);
 794
 795         while (1) {
 796                 mutex_lock(&c->ec_stripes_heap_lock);
 797                 u64 idx = stripe_idx_to_delete(c);
 798                 mutex_unlock(&c->ec_stripes_heap_lock);
 799
 800                 if (!idx)
 801                         break;
 802
 803                 int ret = bch2_trans_do(c, NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
 804                                         ec_stripe_delete(trans, idx));
 805                 bch_err_fn(c, ret);
 806                 if (ret)
 807                         break;
 808         }
 809
 810         bch2_write_ref_put(c, BCH_WRITE_REF_stripe_delete);
 811 }
 812
 813 void bch2_do_stripe_deletes(struct bch_fs *c)
 814 {
 815         if (bch2_write_ref_tryget(c, BCH_WRITE_REF_stripe_delete) &&
 816             !queue_work(c->write_ref_wq, &c->ec_stripe_delete_work))
 817                 bch2_write_ref_put(c, BCH_WRITE_REF_stripe_delete);
 818 }
 819
 820 /* stripe creation: */
 821
 822 static int ec_stripe_key_update(struct btree_trans *trans,
 823                                 struct bkey_i_stripe *new,
 824                                 bool create)
 825 {
 826         struct bch_fs *c = trans->c;
 827         struct btree_iter iter;
 828         struct bkey_s_c k;
 829         int ret;
 830
 831         k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_stripes,
 832                                new->k.p, BTREE_ITER_INTENT);
 833         ret = bkey_err(k);
 834         if (ret)
 835                 goto err;
 836
 837         if (k.k->type != (create ? KEY_TYPE_deleted : KEY_TYPE_stripe)) {
 838                 bch2_fs_inconsistent(c, "error %s stripe: got existing key type %s",
 839                                      create ? "creating" : "updating",
 840                                      bch2_bkey_types[k.k->type]);
 841                 ret = -EINVAL;
 842                 goto err;
 843         }
 844
 845         if (k.k->type == KEY_TYPE_stripe) {
 846                 const struct bch_stripe *old = bkey_s_c_to_stripe(k).v;
 847                 unsigned i;
 848
 849                 if (old->nr_blocks != new->v.nr_blocks) {
 850                         bch_err(c, "error updating stripe: nr_blocks does not match");
 851                         ret = -EINVAL;
 852                         goto err;
 853                 }
 854
 855                 for (i = 0; i < new->v.nr_blocks; i++) {
 856                         unsigned v = stripe_blockcount_get(old, i);
 857
 858                         BUG_ON(v &&
 859                                (old->ptrs[i].dev != new->v.ptrs[i].dev ||
 860                                 old->ptrs[i].gen != new->v.ptrs[i].gen ||
 861                                 old->ptrs[i].offset != new->v.ptrs[i].offset));
 862
 863                         stripe_blockcount_set(&new->v, i, v);
 864                 }
 865         }
 866
 867         ret = bch2_trans_update(trans, &iter, &new->k_i, 0);
 868 err:
 869         bch2_trans_iter_exit(trans, &iter);
 870         return ret;
 871 }
 872
 873 static int ec_stripe_update_extent(struct btree_trans *trans,
 874                                    struct bpos bucket, u8 gen,
 875                                    struct ec_stripe_buf *s,
 876                                    struct bpos *bp_pos)
 877 {
 878         struct bch_stripe *v = &bkey_i_to_stripe(&s->key)->v;
 879         struct bch_fs *c = trans->c;
 880         struct bch_backpointer bp;
 881         struct btree_iter iter;
 882         struct bkey_s_c k;
 883         const struct bch_extent_ptr *ptr_c;
 884         struct bch_extent_ptr *ptr, *ec_ptr = NULL;
 885         struct bch_extent_stripe_ptr stripe_ptr;
 886         struct bkey_i *n;
 887         int ret, dev, block;
 888
 889         ret = bch2_get_next_backpointer(trans, bucket, gen,
 890                                 bp_pos, &bp, BTREE_ITER_CACHED);
 891         if (ret)
 892                 return ret;
 893         if (bpos_eq(*bp_pos, SPOS_MAX))
 894                 return 0;
 895
 896         if (bp.level) {
 897                 struct printbuf buf = PRINTBUF;
 898                 struct btree_iter node_iter;
 899                 struct btree *b;
 900
 901                 b = bch2_backpointer_get_node(trans, &node_iter, *bp_pos, bp);
 902                 bch2_trans_iter_exit(trans, &node_iter);
 903
 904                 if (!b)
 905                         return 0;
 906
 907                 prt_printf(&buf, "found btree node in erasure coded bucket: b=%px\n", b);
 908                 bch2_backpointer_to_text(&buf, &bp);
 909
 910                 bch2_fs_inconsistent(c, "%s", buf.buf);
 911                 printbuf_exit(&buf);
 912                 return -EIO;
 913         }
 914
 915         k = bch2_backpointer_get_key(trans, &iter, *bp_pos, bp, BTREE_ITER_INTENT);
 916         ret = bkey_err(k);
 917         if (ret)
 918                 return ret;
 919         if (!k.k) {
 920                 /*
 921                  * extent no longer exists - we could flush the btree
 922                  * write buffer and retry to verify, but no need:
 923                  */
 924                 return 0;
 925         }
 926
 927         if (extent_has_stripe_ptr(k, s->key.k.p.offset))
 928                 goto out;
 929
 930         ptr_c = bkey_matches_stripe(v, k, &block);
 931         /*
 932          * It doesn't generally make sense to erasure code cached ptrs:
 933          * XXX: should we be incrementing a counter?
 934          */
 935         if (!ptr_c || ptr_c->cached)
 936                 goto out;
 937
 938         dev = v->ptrs[block].dev;
 939
 940         n = bch2_trans_kmalloc(trans, bkey_bytes(k.k) + sizeof(stripe_ptr));
 941         ret = PTR_ERR_OR_ZERO(n);
 942         if (ret)
 943                 goto out;
 944
 945         bkey_reassemble(n, k);
 946
 947         bch2_bkey_drop_ptrs(bkey_i_to_s(n), ptr, ptr->dev != dev);
 948         ec_ptr = bch2_bkey_has_device(bkey_i_to_s(n), dev);
 949         BUG_ON(!ec_ptr);
 950
 951         stripe_ptr = (struct bch_extent_stripe_ptr) {
 952                 .type = 1 << BCH_EXTENT_ENTRY_stripe_ptr,
 953                 .block          = block,
 954                 .redundancy     = v->nr_redundant,
 955                 .idx            = s->key.k.p.offset,
 956         };
 957
 958         __extent_entry_insert(n,
 959                         (union bch_extent_entry *) ec_ptr,
 960                         (union bch_extent_entry *) &stripe_ptr);
 961
 962         ret = bch2_trans_update(trans, &iter, n, 0);
 963 out:
 964         bch2_trans_iter_exit(trans, &iter);
 965         return ret;
 966 }
 967
 968 static int ec_stripe_update_bucket(struct btree_trans *trans, struct ec_stripe_buf *s,
 969                                    unsigned block)
 970 {
 971         struct bch_fs *c = trans->c;
 972         struct bch_stripe *v = &bkey_i_to_stripe(&s->key)->v;
 973         struct bch_extent_ptr bucket = v->ptrs[block];
 974         struct bpos bucket_pos = PTR_BUCKET_POS(c, &bucket);
 975         struct bpos bp_pos = POS_MIN;
 976         int ret = 0;
 977
 978         while (1) {
 979                 ret = commit_do(trans, NULL, NULL,
 980                                 BCH_TRANS_COMMIT_no_check_rw|
 981                                 BCH_TRANS_COMMIT_no_enospc,
 982                         ec_stripe_update_extent(trans, bucket_pos, bucket.gen,
 983                                                 s, &bp_pos));
 984                 if (ret)
 985                         break;
 986                 if (bkey_eq(bp_pos, POS_MAX))
 987                         break;
 988
 989                 bp_pos = bpos_nosnap_successor(bp_pos);
 990         }
 991
 992         return ret;
 993 }
 994
 995 static int ec_stripe_update_extents(struct bch_fs *c, struct ec_stripe_buf *s)
 996 {
 997         struct btree_trans *trans = bch2_trans_get(c);
 998         struct bch_stripe *v = &bkey_i_to_stripe(&s->key)->v;
 999         unsigned i, nr_data = v->nr_blocks - v->nr_redundant;
1000         int ret = 0;
1001
1002         ret = bch2_btree_write_buffer_flush_sync(trans);
1003         if (ret)
1004                 goto err;
1005
1006         for (i = 0; i < nr_data; i++) {
1007                 ret = ec_stripe_update_bucket(trans, s, i);
1008                 if (ret)
1009                         break;
1010         }
1011 err:
1012         bch2_trans_put(trans);
1013
1014         return ret;
1015 }
1016
1017 static void zero_out_rest_of_ec_bucket(struct bch_fs *c,
1018                                        struct ec_stripe_new *s,
1019                                        unsigned block,
1020                                        struct open_bucket *ob)
1021 {
1022         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1023         unsigned offset = ca->mi.bucket_size - ob->sectors_free;
1024         int ret;
1025
1026         if (!bch2_dev_get_ioref(ca, WRITE)) {
1027                 s->err = -BCH_ERR_erofs_no_writes;
1028                 return;
1029         }
1030
1031         memset(s->new_stripe.data[block] + (offset << 9),
1032                0,
1033                ob->sectors_free << 9);
1034
1035         ret = blkdev_issue_zeroout(ca->disk_sb.bdev,
1036                         ob->bucket * ca->mi.bucket_size + offset,
1037                         ob->sectors_free,
1038                         GFP_KERNEL, 0);
1039
1040         percpu_ref_put(&ca->io_ref);
1041
1042         if (ret)
1043                 s->err = ret;
1044 }
1045
1046 void bch2_ec_stripe_new_free(struct bch_fs *c, struct ec_stripe_new *s)
1047 {
1048         if (s->idx)
1049                 bch2_stripe_close(c, s);
1050         kfree(s);
1051 }
1052
1053 /*
1054  * data buckets of new stripe all written: create the stripe
1055  */
1056 static void ec_stripe_create(struct ec_stripe_new *s)
1057 {
1058         struct bch_fs *c = s->c;
1059         struct open_bucket *ob;
1060         struct bch_stripe *v = &bkey_i_to_stripe(&s->new_stripe.key)->v;
1061         unsigned i, nr_data = v->nr_blocks - v->nr_redundant;
1062         int ret;
1063
1064         BUG_ON(s->h->s == s);
1065
1066         closure_sync(&s->iodone);
1067
1068         if (!s->err) {
1069                 for (i = 0; i < nr_data; i++)
1070                         if (s->blocks[i]) {
1071                                 ob = c->open_buckets + s->blocks[i];
1072
1073                                 if (ob->sectors_free)
1074                                         zero_out_rest_of_ec_bucket(c, s, i, ob);
1075                         }
1076         }
1077
1078         if (s->err) {
1079                 if (!bch2_err_matches(s->err, EROFS))
1080                         bch_err(c, "error creating stripe: error writing data buckets");
1081                 goto err;
1082         }
1083
1084         if (s->have_existing_stripe) {
1085                 ec_validate_checksums(c, &s->existing_stripe);
1086
1087                 if (ec_do_recov(c, &s->existing_stripe)) {
1088                         bch_err(c, "error creating stripe: error reading existing stripe");
1089                         goto err;
1090                 }
1091
1092                 for (i = 0; i < nr_data; i++)
1093                         if (stripe_blockcount_get(&bkey_i_to_stripe(&s->existing_stripe.key)->v, i))
1094                                 swap(s->new_stripe.data[i],
1095                                      s->existing_stripe.data[i]);
1096
1097                 ec_stripe_buf_exit(&s->existing_stripe);
1098         }
1099
1100         BUG_ON(!s->allocated);
1101         BUG_ON(!s->idx);
1102
1103         ec_generate_ec(&s->new_stripe);
1104
1105         ec_generate_checksums(&s->new_stripe);
1106
1107         /* write p/q: */
1108         for (i = nr_data; i < v->nr_blocks; i++)
1109                 ec_block_io(c, &s->new_stripe, REQ_OP_WRITE, i, &s->iodone);
1110         closure_sync(&s->iodone);
1111
1112         if (ec_nr_failed(&s->new_stripe)) {
1113                 bch_err(c, "error creating stripe: error writing redundancy buckets");
1114                 goto err;
1115         }
1116
1117         ret = bch2_trans_do(c, &s->res, NULL,
1118                             BCH_TRANS_COMMIT_no_check_rw|
1119                             BCH_TRANS_COMMIT_no_enospc,
1120                             ec_stripe_key_update(trans,
1121                                         bkey_i_to_stripe(&s->new_stripe.key),
1122                                         !s->have_existing_stripe));
1123         bch_err_msg(c, ret, "creating stripe key");
1124         if (ret) {
1125                 goto err;
1126         }
1127
1128         ret = ec_stripe_update_extents(c, &s->new_stripe);
1129         bch_err_msg(c, ret, "error updating extents");
1130         if (ret)
1131                 goto err;
1132 err:
1133         bch2_disk_reservation_put(c, &s->res);
1134
1135         for (i = 0; i < v->nr_blocks; i++)
1136                 if (s->blocks[i]) {
1137                         ob = c->open_buckets + s->blocks[i];
1138
1139                         if (i < nr_data) {
1140                                 ob->ec = NULL;
1141                                 __bch2_open_bucket_put(c, ob);
1142                         } else {
1143                                 bch2_open_bucket_put(c, ob);
1144                         }
1145                 }
1146
1147         mutex_lock(&c->ec_stripe_new_lock);
1148         list_del(&s->list);
1149         mutex_unlock(&c->ec_stripe_new_lock);
1150         wake_up(&c->ec_stripe_new_wait);
1151
1152         ec_stripe_buf_exit(&s->existing_stripe);
1153         ec_stripe_buf_exit(&s->new_stripe);
1154         closure_debug_destroy(&s->iodone);
1155
1156         ec_stripe_new_put(c, s, STRIPE_REF_stripe);
1157 }
1158
1159 static struct ec_stripe_new *get_pending_stripe(struct bch_fs *c)
1160 {
1161         struct ec_stripe_new *s;
1162
1163         mutex_lock(&c->ec_stripe_new_lock);
1164         list_for_each_entry(s, &c->ec_stripe_new_list, list)
1165                 if (!atomic_read(&s->ref[STRIPE_REF_io]))
1166                         goto out;
1167         s = NULL;
1168 out:
1169         mutex_unlock(&c->ec_stripe_new_lock);
1170
1171         return s;
1172 }
1173
1174 static void ec_stripe_create_work(struct work_struct *work)
1175 {
1176         struct bch_fs *c = container_of(work,
1177                 struct bch_fs, ec_stripe_create_work);
1178         struct ec_stripe_new *s;
1179
1180         while ((s = get_pending_stripe(c)))
1181                 ec_stripe_create(s);
1182
1183         bch2_write_ref_put(c, BCH_WRITE_REF_stripe_create);
1184 }
1185
1186 void bch2_ec_do_stripe_creates(struct bch_fs *c)
1187 {
1188         bch2_write_ref_get(c, BCH_WRITE_REF_stripe_create);
1189
1190         if (!queue_work(system_long_wq, &c->ec_stripe_create_work))
1191                 bch2_write_ref_put(c, BCH_WRITE_REF_stripe_create);
1192 }
1193
1194 static void ec_stripe_set_pending(struct bch_fs *c, struct ec_stripe_head *h)
1195 {
1196         struct ec_stripe_new *s = h->s;
1197
1198         BUG_ON(!s->allocated && !s->err);
1199
1200         h->s            = NULL;
1201         s->pending      = true;
1202
1203         mutex_lock(&c->ec_stripe_new_lock);
1204         list_add(&s->list, &c->ec_stripe_new_list);
1205         mutex_unlock(&c->ec_stripe_new_lock);
1206
1207         ec_stripe_new_put(c, s, STRIPE_REF_io);
1208 }
1209
1210 void bch2_ec_bucket_cancel(struct bch_fs *c, struct open_bucket *ob)
1211 {
1212         struct ec_stripe_new *s = ob->ec;
1213
1214         s->err = -EIO;
1215 }
1216
1217 void *bch2_writepoint_ec_buf(struct bch_fs *c, struct write_point *wp)
1218 {
1219         struct open_bucket *ob = ec_open_bucket(c, &wp->ptrs);
1220         struct bch_dev *ca;
1221         unsigned offset;
1222
1223         if (!ob)
1224                 return NULL;
1225
1226         BUG_ON(!ob->ec->new_stripe.data[ob->ec_idx]);
1227
1228         ca      = bch_dev_bkey_exists(c, ob->dev);
1229         offset  = ca->mi.bucket_size - ob->sectors_free;
1230
1231         return ob->ec->new_stripe.data[ob->ec_idx] + (offset << 9);
1232 }
1233
1234 static int unsigned_cmp(const void *_l, const void *_r)
1235 {
1236         unsigned l = *((const unsigned *) _l);
1237         unsigned r = *((const unsigned *) _r);
1238
1239         return cmp_int(l, r);
1240 }
1241
1242 /* pick most common bucket size: */
1243 static unsigned pick_blocksize(struct bch_fs *c,
1244                                struct bch_devs_mask *devs)
1245 {
1246         unsigned nr = 0, sizes[BCH_SB_MEMBERS_MAX];
1247         struct {
1248                 unsigned nr, size;
1249         } cur = { 0, 0 }, best = { 0, 0 };
1250
1251         for_each_member_device_rcu(c, ca, devs)
1252                 sizes[nr++] = ca->mi.bucket_size;
1253
1254         sort(sizes, nr, sizeof(unsigned), unsigned_cmp, NULL);
1255
1256         for (unsigned i = 0; i < nr; i++) {
1257                 if (sizes[i] != cur.size) {
1258                         if (cur.nr > best.nr)
1259                                 best = cur;
1260
1261                         cur.nr = 0;
1262                         cur.size = sizes[i];
1263                 }
1264
1265                 cur.nr++;
1266         }
1267
1268         if (cur.nr > best.nr)
1269                 best = cur;
1270
1271         return best.size;
1272 }
1273
1274 static bool may_create_new_stripe(struct bch_fs *c)
1275 {
1276         return false;
1277 }
1278
1279 static void ec_stripe_key_init(struct bch_fs *c,
1280                                struct bkey_i *k,
1281                                unsigned nr_data,
1282                                unsigned nr_parity,
1283                                unsigned stripe_size)
1284 {
1285         struct bkey_i_stripe *s = bkey_stripe_init(k);
1286         unsigned u64s;
1287
1288         s->v.sectors                    = cpu_to_le16(stripe_size);
1289         s->v.algorithm                  = 0;
1290         s->v.nr_blocks                  = nr_data + nr_parity;
1291         s->v.nr_redundant               = nr_parity;
1292         s->v.csum_granularity_bits      = ilog2(c->opts.encoded_extent_max >> 9);
1293         s->v.csum_type                  = BCH_CSUM_crc32c;
1294         s->v.pad                        = 0;
1295
1296         while ((u64s = stripe_val_u64s(&s->v)) > BKEY_VAL_U64s_MAX) {
1297                 BUG_ON(1 << s->v.csum_granularity_bits >=
1298                        le16_to_cpu(s->v.sectors) ||
1299                        s->v.csum_granularity_bits == U8_MAX);
1300                 s->v.csum_granularity_bits++;
1301         }
1302
1303         set_bkey_val_u64s(&s->k, u64s);
1304 }
1305
1306 static int ec_new_stripe_alloc(struct bch_fs *c, struct ec_stripe_head *h)
1307 {
1308         struct ec_stripe_new *s;
1309
1310         lockdep_assert_held(&h->lock);
1311
1312         s = kzalloc(sizeof(*s), GFP_KERNEL);
1313         if (!s)
1314                 return -BCH_ERR_ENOMEM_ec_new_stripe_alloc;
1315
1316         mutex_init(&s->lock);
1317         closure_init(&s->iodone, NULL);
1318         atomic_set(&s->ref[STRIPE_REF_stripe], 1);
1319         atomic_set(&s->ref[STRIPE_REF_io], 1);
1320         s->c            = c;
1321         s->h            = h;
1322         s->nr_data      = min_t(unsigned, h->nr_active_devs,
1323                                 BCH_BKEY_PTRS_MAX) - h->redundancy;
1324         s->nr_parity    = h->redundancy;
1325
1326         ec_stripe_key_init(c, &s->new_stripe.key,
1327                            s->nr_data, s->nr_parity, h->blocksize);
1328
1329         h->s = s;
1330         return 0;
1331 }
1332
1333 static struct ec_stripe_head *
1334 ec_new_stripe_head_alloc(struct bch_fs *c, unsigned target,
1335                          unsigned algo, unsigned redundancy,
1336                          enum bch_watermark watermark)
1337 {
1338         struct ec_stripe_head *h;
1339
1340         h = kzalloc(sizeof(*h), GFP_KERNEL);
1341         if (!h)
1342                 return NULL;
1343
1344         mutex_init(&h->lock);
1345         BUG_ON(!mutex_trylock(&h->lock));
1346
1347         h->target       = target;
1348         h->algo         = algo;
1349         h->redundancy   = redundancy;
1350         h->watermark    = watermark;
1351
1352         rcu_read_lock();
1353         h->devs = target_rw_devs(c, BCH_DATA_user, target);
1354
1355         for_each_member_device_rcu(c, ca, &h->devs)
1356                 if (!ca->mi.durability)
1357                         __clear_bit(ca->dev_idx, h->devs.d);
1358
1359         h->blocksize = pick_blocksize(c, &h->devs);
1360
1361         for_each_member_device_rcu(c, ca, &h->devs)
1362                 if (ca->mi.bucket_size == h->blocksize)
1363                         h->nr_active_devs++;
1364
1365         rcu_read_unlock();
1366
1367         /*
1368          * If we only have redundancy + 1 devices, we're better off with just
1369          * replication:
1370          */
1371         if (h->nr_active_devs < h->redundancy + 2)
1372                 bch_err(c, "insufficient devices available to create stripe (have %u, need %u) - mismatched bucket sizes?",
1373                         h->nr_active_devs, h->redundancy + 2);
1374
1375         list_add(&h->list, &c->ec_stripe_head_list);
1376         return h;
1377 }
1378
1379 void bch2_ec_stripe_head_put(struct bch_fs *c, struct ec_stripe_head *h)
1380 {
1381         if (h->s &&
1382             h->s->allocated &&
1383             bitmap_weight(h->s->blocks_allocated,
1384                           h->s->nr_data) == h->s->nr_data)
1385                 ec_stripe_set_pending(c, h);
1386
1387         mutex_unlock(&h->lock);
1388 }
1389
1390 static struct ec_stripe_head *
1391 __bch2_ec_stripe_head_get(struct btree_trans *trans,
1392                           unsigned target,
1393                           unsigned algo,
1394                           unsigned redundancy,
1395                           enum bch_watermark watermark)
1396 {
1397         struct bch_fs *c = trans->c;
1398         struct ec_stripe_head *h;
1399         int ret;
1400
1401         if (!redundancy)
1402                 return NULL;
1403
1404         ret = bch2_trans_mutex_lock(trans, &c->ec_stripe_head_lock);
1405         if (ret)
1406                 return ERR_PTR(ret);
1407
1408         if (test_bit(BCH_FS_going_ro, &c->flags)) {
1409                 h = ERR_PTR(-BCH_ERR_erofs_no_writes);
1410                 goto found;
1411         }
1412
1413         list_for_each_entry(h, &c->ec_stripe_head_list, list)
1414                 if (h->target           == target &&
1415                     h->algo             == algo &&
1416                     h->redundancy       == redundancy &&
1417                     h->watermark        == watermark) {
1418                         ret = bch2_trans_mutex_lock(trans, &h->lock);
1419                         if (ret)
1420                                 h = ERR_PTR(ret);
1421                         goto found;
1422                 }
1423
1424         h = ec_new_stripe_head_alloc(c, target, algo, redundancy, watermark);
1425 found:
1426         if (!IS_ERR_OR_NULL(h) &&
1427             h->nr_active_devs < h->redundancy + 2) {
1428                 mutex_unlock(&h->lock);
1429                 h = NULL;
1430         }
1431         mutex_unlock(&c->ec_stripe_head_lock);
1432         return h;
1433 }
1434
1435 static int new_stripe_alloc_buckets(struct btree_trans *trans, struct ec_stripe_head *h,
1436                                     enum bch_watermark watermark, struct closure *cl)
1437 {
1438         struct bch_fs *c = trans->c;
1439         struct bch_devs_mask devs = h->devs;
1440         struct open_bucket *ob;
1441         struct open_buckets buckets;
1442         struct bch_stripe *v = &bkey_i_to_stripe(&h->s->new_stripe.key)->v;
1443         unsigned i, j, nr_have_parity = 0, nr_have_data = 0;
1444         bool have_cache = true;
1445         int ret = 0;
1446
1447         BUG_ON(v->nr_blocks     != h->s->nr_data + h->s->nr_parity);
1448         BUG_ON(v->nr_redundant  != h->s->nr_parity);
1449
1450         for_each_set_bit(i, h->s->blocks_gotten, v->nr_blocks) {
1451                 __clear_bit(v->ptrs[i].dev, devs.d);
1452                 if (i < h->s->nr_data)
1453                         nr_have_data++;
1454                 else
1455                         nr_have_parity++;
1456         }
1457
1458         BUG_ON(nr_have_data     > h->s->nr_data);
1459         BUG_ON(nr_have_parity   > h->s->nr_parity);
1460
1461         buckets.nr = 0;
1462         if (nr_have_parity < h->s->nr_parity) {
1463                 ret = bch2_bucket_alloc_set_trans(trans, &buckets,
1464                                             &h->parity_stripe,
1465                                             &devs,
1466                                             h->s->nr_parity,
1467                                             &nr_have_parity,
1468                                             &have_cache, 0,
1469                                             BCH_DATA_parity,
1470                                             watermark,
1471                                             cl);
1472
1473                 open_bucket_for_each(c, &buckets, ob, i) {
1474                         j = find_next_zero_bit(h->s->blocks_gotten,
1475                                                h->s->nr_data + h->s->nr_parity,
1476                                                h->s->nr_data);
1477                         BUG_ON(j >= h->s->nr_data + h->s->nr_parity);
1478
1479                         h->s->blocks[j] = buckets.v[i];
1480                         v->ptrs[j] = bch2_ob_ptr(c, ob);
1481                         __set_bit(j, h->s->blocks_gotten);
1482                 }
1483
1484                 if (ret)
1485                         return ret;
1486         }
1487
1488         buckets.nr = 0;
1489         if (nr_have_data < h->s->nr_data) {
1490                 ret = bch2_bucket_alloc_set_trans(trans, &buckets,
1491                                             &h->block_stripe,
1492                                             &devs,
1493                                             h->s->nr_data,
1494                                             &nr_have_data,
1495                                             &have_cache, 0,
1496                                             BCH_DATA_user,
1497                                             watermark,
1498                                             cl);
1499
1500                 open_bucket_for_each(c, &buckets, ob, i) {
1501                         j = find_next_zero_bit(h->s->blocks_gotten,
1502                                                h->s->nr_data, 0);
1503                         BUG_ON(j >= h->s->nr_data);
1504
1505                         h->s->blocks[j] = buckets.v[i];
1506                         v->ptrs[j] = bch2_ob_ptr(c, ob);
1507                         __set_bit(j, h->s->blocks_gotten);
1508                 }
1509
1510                 if (ret)
1511                         return ret;
1512         }
1513
1514         return 0;
1515 }
1516
1517 /* XXX: doesn't obey target: */
1518 static s64 get_existing_stripe(struct bch_fs *c,
1519                                struct ec_stripe_head *head)
1520 {
1521         ec_stripes_heap *h = &c->ec_stripes_heap;
1522         struct stripe *m;
1523         size_t heap_idx;
1524         u64 stripe_idx;
1525         s64 ret = -1;
1526
1527         if (may_create_new_stripe(c))
1528                 return -1;
1529
1530         mutex_lock(&c->ec_stripes_heap_lock);
1531         for (heap_idx = 0; heap_idx < h->used; heap_idx++) {
1532                 /* No blocks worth reusing, stripe will just be deleted: */
1533                 if (!h->data[heap_idx].blocks_nonempty)
1534                         continue;
1535
1536                 stripe_idx = h->data[heap_idx].idx;
1537
1538                 m = genradix_ptr(&c->stripes, stripe_idx);
1539
1540                 if (m->algorithm        == head->algo &&
1541                     m->nr_redundant     == head->redundancy &&
1542                     m->sectors          == head->blocksize &&
1543                     m->blocks_nonempty  < m->nr_blocks - m->nr_redundant &&
1544                     bch2_try_open_stripe(c, head->s, stripe_idx)) {
1545                         ret = stripe_idx;
1546                         break;
1547                 }
1548         }
1549         mutex_unlock(&c->ec_stripes_heap_lock);
1550         return ret;
1551 }
1552
1553 static int __bch2_ec_stripe_head_reuse(struct btree_trans *trans, struct ec_stripe_head *h)
1554 {
1555         struct bch_fs *c = trans->c;
1556         struct bch_stripe *new_v = &bkey_i_to_stripe(&h->s->new_stripe.key)->v;
1557         struct bch_stripe *existing_v;
1558         unsigned i;
1559         s64 idx;
1560         int ret;
1561
1562         /*
1563          * If we can't allocate a new stripe, and there's no stripes with empty
1564          * blocks for us to reuse, that means we have to wait on copygc:
1565          */
1566         idx = get_existing_stripe(c, h);
1567         if (idx < 0)
1568                 return -BCH_ERR_stripe_alloc_blocked;
1569
1570         ret = get_stripe_key_trans(trans, idx, &h->s->existing_stripe);
1571         if (ret) {
1572                 bch2_stripe_close(c, h->s);
1573                 if (!bch2_err_matches(ret, BCH_ERR_transaction_restart))
1574                         bch2_fs_fatal_error(c, "error reading stripe key: %s", bch2_err_str(ret));
1575                 return ret;
1576         }
1577
1578         existing_v = &bkey_i_to_stripe(&h->s->existing_stripe.key)->v;
1579
1580         BUG_ON(existing_v->nr_redundant != h->s->nr_parity);
1581         h->s->nr_data = existing_v->nr_blocks -
1582                 existing_v->nr_redundant;
1583
1584         ret = ec_stripe_buf_init(&h->s->existing_stripe, 0, h->blocksize);
1585         if (ret) {
1586                 bch2_stripe_close(c, h->s);
1587                 return ret;
1588         }
1589
1590         BUG_ON(h->s->existing_stripe.size != h->blocksize);
1591         BUG_ON(h->s->existing_stripe.size != le16_to_cpu(existing_v->sectors));
1592
1593         /*
1594          * Free buckets we initially allocated - they might conflict with
1595          * blocks from the stripe we're reusing:
1596          */
1597         for_each_set_bit(i, h->s->blocks_gotten, new_v->nr_blocks) {
1598                 bch2_open_bucket_put(c, c->open_buckets + h->s->blocks[i]);
1599                 h->s->blocks[i] = 0;
1600         }
1601         memset(h->s->blocks_gotten, 0, sizeof(h->s->blocks_gotten));
1602         memset(h->s->blocks_allocated, 0, sizeof(h->s->blocks_allocated));
1603
1604         for (i = 0; i < existing_v->nr_blocks; i++) {
1605                 if (stripe_blockcount_get(existing_v, i)) {
1606                         __set_bit(i, h->s->blocks_gotten);
1607                         __set_bit(i, h->s->blocks_allocated);
1608                 }
1609
1610                 ec_block_io(c, &h->s->existing_stripe, READ, i, &h->s->iodone);
1611         }
1612
1613         bkey_copy(&h->s->new_stripe.key, &h->s->existing_stripe.key);
1614         h->s->have_existing_stripe = true;
1615
1616         return 0;
1617 }
1618
1619 static int __bch2_ec_stripe_head_reserve(struct btree_trans *trans, struct ec_stripe_head *h)
1620 {
1621         struct bch_fs *c = trans->c;
1622         struct btree_iter iter;
1623         struct bkey_s_c k;
1624         struct bpos min_pos = POS(0, 1);
1625         struct bpos start_pos = bpos_max(min_pos, POS(0, c->ec_stripe_hint));
1626         int ret;
1627
1628         if (!h->s->res.sectors) {
1629                 ret = bch2_disk_reservation_get(c, &h->s->res,
1630                                         h->blocksize,
1631                                         h->s->nr_parity,
1632                                         BCH_DISK_RESERVATION_NOFAIL);
1633                 if (ret)
1634                         return ret;
1635         }
1636
1637         for_each_btree_key_norestart(trans, iter, BTREE_ID_stripes, start_pos,
1638                            BTREE_ITER_SLOTS|BTREE_ITER_INTENT, k, ret) {
1639                 if (bkey_gt(k.k->p, POS(0, U32_MAX))) {
1640                         if (start_pos.offset) {
1641                                 start_pos = min_pos;
1642                                 bch2_btree_iter_set_pos(&iter, start_pos);
1643                                 continue;
1644                         }
1645
1646                         ret = -BCH_ERR_ENOSPC_stripe_create;
1647                         break;
1648                 }
1649
1650                 if (bkey_deleted(k.k) &&
1651                     bch2_try_open_stripe(c, h->s, k.k->p.offset))
1652                         break;
1653         }
1654
1655         c->ec_stripe_hint = iter.pos.offset;
1656
1657         if (ret)
1658                 goto err;
1659
1660         ret = ec_stripe_mem_alloc(trans, &iter);
1661         if (ret) {
1662                 bch2_stripe_close(c, h->s);
1663                 goto err;
1664         }
1665
1666         h->s->new_stripe.key.k.p = iter.pos;
1667 out:
1668         bch2_trans_iter_exit(trans, &iter);
1669         return ret;
1670 err:
1671         bch2_disk_reservation_put(c, &h->s->res);
1672         goto out;
1673 }
1674
1675 struct ec_stripe_head *bch2_ec_stripe_head_get(struct btree_trans *trans,
1676                                                unsigned target,
1677                                                unsigned algo,
1678                                                unsigned redundancy,
1679                                                enum bch_watermark watermark,
1680                                                struct closure *cl)
1681 {
1682         struct bch_fs *c = trans->c;
1683         struct ec_stripe_head *h;
1684         bool waiting = false;
1685         int ret;
1686
1687         h = __bch2_ec_stripe_head_get(trans, target, algo, redundancy, watermark);
1688         if (IS_ERR_OR_NULL(h))
1689                 return h;
1690
1691         if (!h->s) {
1692                 ret = ec_new_stripe_alloc(c, h);
1693                 if (ret) {
1694                         bch_err(c, "failed to allocate new stripe");
1695                         goto err;
1696                 }
1697         }
1698
1699         if (h->s->allocated)
1700                 goto allocated;
1701
1702         if (h->s->have_existing_stripe)
1703                 goto alloc_existing;
1704
1705         /* First, try to allocate a full stripe: */
1706         ret =   new_stripe_alloc_buckets(trans, h, BCH_WATERMARK_stripe, NULL) ?:
1707                 __bch2_ec_stripe_head_reserve(trans, h);
1708         if (!ret)
1709                 goto allocate_buf;
1710         if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
1711             bch2_err_matches(ret, ENOMEM))
1712                 goto err;
1713
1714         /*
1715          * Not enough buckets available for a full stripe: we must reuse an
1716          * existing stripe:
1717          */
1718         while (1) {
1719                 ret = __bch2_ec_stripe_head_reuse(trans, h);
1720                 if (!ret)
1721                         break;
1722                 if (waiting || !cl || ret != -BCH_ERR_stripe_alloc_blocked)
1723                         goto err;
1724
1725                 if (watermark == BCH_WATERMARK_copygc) {
1726                         ret =   new_stripe_alloc_buckets(trans, h, watermark, NULL) ?:
1727                                 __bch2_ec_stripe_head_reserve(trans, h);
1728                         if (ret)
1729                                 goto err;
1730                         goto allocate_buf;
1731                 }
1732
1733                 /* XXX freelist_wait? */
1734                 closure_wait(&c->freelist_wait, cl);
1735                 waiting = true;
1736         }
1737
1738         if (waiting)
1739                 closure_wake_up(&c->freelist_wait);
1740 alloc_existing:
1741         /*
1742          * Retry allocating buckets, with the watermark for this
1743          * particular write:
1744          */
1745         ret = new_stripe_alloc_buckets(trans, h, watermark, cl);
1746         if (ret)
1747                 goto err;
1748
1749 allocate_buf:
1750         ret = ec_stripe_buf_init(&h->s->new_stripe, 0, h->blocksize);
1751         if (ret)
1752                 goto err;
1753
1754         h->s->allocated = true;
1755 allocated:
1756         BUG_ON(!h->s->idx);
1757         BUG_ON(!h->s->new_stripe.data[0]);
1758         BUG_ON(trans->restarted);
1759         return h;
1760 err:
1761         bch2_ec_stripe_head_put(c, h);
1762         return ERR_PTR(ret);
1763 }
1764
1765 static void __bch2_ec_stop(struct bch_fs *c, struct bch_dev *ca)
1766 {
1767         struct ec_stripe_head *h;
1768         struct open_bucket *ob;
1769         unsigned i;
1770
1771         mutex_lock(&c->ec_stripe_head_lock);
1772         list_for_each_entry(h, &c->ec_stripe_head_list, list) {
1773                 mutex_lock(&h->lock);
1774                 if (!h->s)
1775                         goto unlock;
1776
1777                 if (!ca)
1778                         goto found;
1779
1780                 for (i = 0; i < bkey_i_to_stripe(&h->s->new_stripe.key)->v.nr_blocks; i++) {
1781                         if (!h->s->blocks[i])
1782                                 continue;
1783
1784                         ob = c->open_buckets + h->s->blocks[i];
1785                         if (ob->dev == ca->dev_idx)
1786                                 goto found;
1787                 }
1788                 goto unlock;
1789 found:
1790                 h->s->err = -BCH_ERR_erofs_no_writes;
1791                 ec_stripe_set_pending(c, h);
1792 unlock:
1793                 mutex_unlock(&h->lock);
1794         }
1795         mutex_unlock(&c->ec_stripe_head_lock);
1796 }
1797
1798 void bch2_ec_stop_dev(struct bch_fs *c, struct bch_dev *ca)
1799 {
1800         __bch2_ec_stop(c, ca);
1801 }
1802
1803 void bch2_fs_ec_stop(struct bch_fs *c)
1804 {
1805         __bch2_ec_stop(c, NULL);
1806 }
1807
1808 static bool bch2_fs_ec_flush_done(struct bch_fs *c)
1809 {
1810         bool ret;
1811
1812         mutex_lock(&c->ec_stripe_new_lock);
1813         ret = list_empty(&c->ec_stripe_new_list);
1814         mutex_unlock(&c->ec_stripe_new_lock);
1815
1816         return ret;
1817 }
1818
1819 void bch2_fs_ec_flush(struct bch_fs *c)
1820 {
1821         wait_event(c->ec_stripe_new_wait, bch2_fs_ec_flush_done(c));
1822 }
1823
1824 int bch2_stripes_read(struct bch_fs *c)
1825 {
1826         int ret = bch2_trans_run(c,
1827                 for_each_btree_key(trans, iter, BTREE_ID_stripes, POS_MIN,
1828                                    BTREE_ITER_PREFETCH, k, ({
1829                         if (k.k->type != KEY_TYPE_stripe)
1830                                 continue;
1831
1832                         ret = __ec_stripe_mem_alloc(c, k.k->p.offset, GFP_KERNEL);
1833                         if (ret)
1834                                 break;
1835
1836                         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
1837
1838                         struct stripe *m = genradix_ptr(&c->stripes, k.k->p.offset);
1839                         m->sectors      = le16_to_cpu(s->sectors);
1840                         m->algorithm    = s->algorithm;
1841                         m->nr_blocks    = s->nr_blocks;
1842                         m->nr_redundant = s->nr_redundant;
1843                         m->blocks_nonempty = 0;
1844
1845                         for (unsigned i = 0; i < s->nr_blocks; i++)
1846                                 m->blocks_nonempty += !!stripe_blockcount_get(s, i);
1847
1848                         bch2_stripes_heap_insert(c, m, k.k->p.offset);
1849                         0;
1850                 })));
1851         bch_err_fn(c, ret);
1852         return ret;
1853 }
1854
1855 void bch2_stripes_heap_to_text(struct printbuf *out, struct bch_fs *c)
1856 {
1857         ec_stripes_heap *h = &c->ec_stripes_heap;
1858         struct stripe *m;
1859         size_t i;
1860
1861         mutex_lock(&c->ec_stripes_heap_lock);
1862         for (i = 0; i < min_t(size_t, h->used, 50); i++) {
1863                 m = genradix_ptr(&c->stripes, h->data[i].idx);
1864
1865                 prt_printf(out, "%zu %u/%u+%u", h->data[i].idx,
1866                        h->data[i].blocks_nonempty,
1867                        m->nr_blocks - m->nr_redundant,
1868                        m->nr_redundant);
1869                 if (bch2_stripe_is_open(c, h->data[i].idx))
1870                         prt_str(out, " open");
1871                 prt_newline(out);
1872         }
1873         mutex_unlock(&c->ec_stripes_heap_lock);
1874 }
1875
1876 void bch2_new_stripes_to_text(struct printbuf *out, struct bch_fs *c)
1877 {
1878         struct ec_stripe_head *h;
1879         struct ec_stripe_new *s;
1880
1881         mutex_lock(&c->ec_stripe_head_lock);
1882         list_for_each_entry(h, &c->ec_stripe_head_list, list) {
1883                 prt_printf(out, "target %u algo %u redundancy %u %s:\n",
1884                        h->target, h->algo, h->redundancy,
1885                        bch2_watermarks[h->watermark]);
1886
1887                 if (h->s)
1888                         prt_printf(out, "\tidx %llu blocks %u+%u allocated %u\n",
1889                                h->s->idx, h->s->nr_data, h->s->nr_parity,
1890                                bitmap_weight(h->s->blocks_allocated,
1891                                              h->s->nr_data));
1892         }
1893         mutex_unlock(&c->ec_stripe_head_lock);
1894
1895         prt_printf(out, "in flight:\n");
1896
1897         mutex_lock(&c->ec_stripe_new_lock);
1898         list_for_each_entry(s, &c->ec_stripe_new_list, list) {
1899                 prt_printf(out, "\tidx %llu blocks %u+%u ref %u %u %s\n",
1900                            s->idx, s->nr_data, s->nr_parity,
1901                            atomic_read(&s->ref[STRIPE_REF_io]),
1902                            atomic_read(&s->ref[STRIPE_REF_stripe]),
1903                            bch2_watermarks[s->h->watermark]);
1904         }
1905         mutex_unlock(&c->ec_stripe_new_lock);
1906 }
1907
1908 void bch2_fs_ec_exit(struct bch_fs *c)
1909 {
1910         struct ec_stripe_head *h;
1911         unsigned i;
1912
1913         while (1) {
1914                 mutex_lock(&c->ec_stripe_head_lock);
1915                 h = list_first_entry_or_null(&c->ec_stripe_head_list,
1916                                              struct ec_stripe_head, list);
1917                 if (h)
1918                         list_del(&h->list);
1919                 mutex_unlock(&c->ec_stripe_head_lock);
1920                 if (!h)
1921                         break;
1922
1923                 if (h->s) {
1924                         for (i = 0; i < bkey_i_to_stripe(&h->s->new_stripe.key)->v.nr_blocks; i++)
1925                                 BUG_ON(h->s->blocks[i]);
1926
1927                         kfree(h->s);
1928                 }
1929                 kfree(h);
1930         }
1931
1932         BUG_ON(!list_empty(&c->ec_stripe_new_list));
1933
1934         free_heap(&c->ec_stripes_heap);
1935         genradix_free(&c->stripes);
1936         bioset_exit(&c->ec_bioset);
1937 }
1938
1939 void bch2_fs_ec_init_early(struct bch_fs *c)
1940 {
1941         spin_lock_init(&c->ec_stripes_new_lock);
1942         mutex_init(&c->ec_stripes_heap_lock);
1943
1944         INIT_LIST_HEAD(&c->ec_stripe_head_list);
1945         mutex_init(&c->ec_stripe_head_lock);
1946
1947         INIT_LIST_HEAD(&c->ec_stripe_new_list);
1948         mutex_init(&c->ec_stripe_new_lock);
1949         init_waitqueue_head(&c->ec_stripe_new_wait);
1950
1951         INIT_WORK(&c->ec_stripe_create_work, ec_stripe_create_work);
1952         INIT_WORK(&c->ec_stripe_delete_work, ec_stripe_delete_work);
1953 }
1954
1955 int bch2_fs_ec_init(struct bch_fs *c)
1956 {
1957         return bioset_init(&c->ec_bioset, 1, offsetof(struct ec_bio, bio),
1958                            BIOSET_NEED_BVECS);
1959 }