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 "bset.h"
   8 #include "btree_gc.h"
   9 #include "btree_update.h"
  10 #include "buckets.h"
  11 #include "disk_groups.h"
  12 #include "ec.h"
  13 #include "error.h"
  14 #include "io.h"
  15 #include "keylist.h"
  16 #include "recovery.h"
  17 #include "super-io.h"
  18 #include "util.h"
  19
  20 #include <linux/sort.h>
  21
  22 #ifdef __KERNEL__
  23
  24 #include <linux/raid/pq.h>
  25 #include <linux/raid/xor.h>
  26
  27 static void raid5_recov(unsigned disks, unsigned failed_idx,
  28                         size_t size, void **data)
  29 {
  30         unsigned i = 2, nr;
  31
  32         BUG_ON(failed_idx >= disks);
  33
  34         swap(data[0], data[failed_idx]);
  35         memcpy(data[0], data[1], size);
  36
  37         while (i < disks) {
  38                 nr = min_t(unsigned, disks - i, MAX_XOR_BLOCKS);
  39                 xor_blocks(nr, size, data[0], data + i);
  40                 i += nr;
  41         }
  42
  43         swap(data[0], data[failed_idx]);
  44 }
  45
  46 static void raid_gen(int nd, int np, size_t size, void **v)
  47 {
  48         if (np >= 1)
  49                 raid5_recov(nd + np, nd, size, v);
  50         if (np >= 2)
  51                 raid6_call.gen_syndrome(nd + np, size, v);
  52         BUG_ON(np > 2);
  53 }
  54
  55 static void raid_rec(int nr, int *ir, int nd, int np, size_t size, void **v)
  56 {
  57         switch (nr) {
  58         case 0:
  59                 break;
  60         case 1:
  61                 if (ir[0] < nd + 1)
  62                         raid5_recov(nd + 1, ir[0], size, v);
  63                 else
  64                         raid6_call.gen_syndrome(nd + np, size, v);
  65                 break;
  66         case 2:
  67                 if (ir[1] < nd) {
  68                         /* data+data failure. */
  69                         raid6_2data_recov(nd + np, size, ir[0], ir[1], v);
  70                 } else if (ir[0] < nd) {
  71                         /* data + p/q failure */
  72
  73                         if (ir[1] == nd) /* data + p failure */
  74                                 raid6_datap_recov(nd + np, size, ir[0], v);
  75                         else { /* data + q failure */
  76                                 raid5_recov(nd + 1, ir[0], size, v);
  77                                 raid6_call.gen_syndrome(nd + np, size, v);
  78                         }
  79                 } else {
  80                         raid_gen(nd, np, size, v);
  81                 }
  82                 break;
  83         default:
  84                 BUG();
  85         }
  86 }
  87
  88 #else
  89
  90 #include <raid/raid.h>
  91
  92 #endif
  93
  94 struct ec_bio {
  95         struct bch_dev          *ca;
  96         struct ec_stripe_buf    *buf;
  97         size_t                  idx;
  98         struct bio              bio;
  99 };
 100
 101 /* Stripes btree keys: */
 102
 103 const char *bch2_stripe_invalid(const struct bch_fs *c, struct bkey_s_c k)
 104 {
 105         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 106
 107         if (k.k->p.inode)
 108                 return "invalid stripe key";
 109
 110         if (bkey_val_bytes(k.k) < sizeof(*s))
 111                 return "incorrect value size";
 112
 113         if (bkey_val_bytes(k.k) < sizeof(*s) ||
 114             bkey_val_u64s(k.k) < stripe_val_u64s(s))
 115                 return "incorrect value size";
 116
 117         return bch2_bkey_ptrs_invalid(c, k);
 118 }
 119
 120 void bch2_stripe_to_text(struct printbuf *out, struct bch_fs *c,
 121                          struct bkey_s_c k)
 122 {
 123         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 124         unsigned i;
 125
 126         pr_buf(out, "algo %u sectors %u blocks %u:%u csum %u gran %u",
 127                s->algorithm,
 128                le16_to_cpu(s->sectors),
 129                s->nr_blocks - s->nr_redundant,
 130                s->nr_redundant,
 131                s->csum_type,
 132                1U << s->csum_granularity_bits);
 133
 134         for (i = 0; i < s->nr_blocks; i++)
 135                 pr_buf(out, " %u:%llu:%u", s->ptrs[i].dev,
 136                        (u64) s->ptrs[i].offset,
 137                        stripe_blockcount_get(s, i));
 138
 139         bch2_bkey_ptrs_to_text(out, c, k);
 140 }
 141
 142 static int ptr_matches_stripe(struct bch_fs *c,
 143                               struct bch_stripe *v,
 144                               const struct bch_extent_ptr *ptr)
 145 {
 146         unsigned i;
 147
 148         for (i = 0; i < v->nr_blocks - v->nr_redundant; i++) {
 149                 const struct bch_extent_ptr *ptr2 = v->ptrs + i;
 150
 151                 if (ptr->dev == ptr2->dev &&
 152                     ptr->gen == ptr2->gen &&
 153                     ptr->offset >= ptr2->offset &&
 154                     ptr->offset <  ptr2->offset + le16_to_cpu(v->sectors))
 155                         return i;
 156         }
 157
 158         return -1;
 159 }
 160
 161 static int extent_matches_stripe(struct bch_fs *c,
 162                                  struct bch_stripe *v,
 163                                  struct bkey_s_c k)
 164 {
 165
 166         switch (k.k->type) {
 167         case KEY_TYPE_extent: {
 168                 struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
 169                 const struct bch_extent_ptr *ptr;
 170                 int idx;
 171
 172                 extent_for_each_ptr(e, ptr) {
 173                         idx = ptr_matches_stripe(c, v, ptr);
 174                         if (idx >= 0)
 175                                 return idx;
 176                 }
 177                 break;
 178         }
 179         }
 180
 181         return -1;
 182 }
 183
 184 static bool extent_has_stripe_ptr(struct bkey_s_c k, u64 idx)
 185 {
 186         switch (k.k->type) {
 187         case KEY_TYPE_extent: {
 188                 struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
 189                 const union bch_extent_entry *entry;
 190
 191                 extent_for_each_entry(e, entry)
 192                         if (extent_entry_type(entry) ==
 193                             BCH_EXTENT_ENTRY_stripe_ptr &&
 194                             entry->stripe_ptr.idx == idx)
 195                                 return true;
 196
 197                 break;
 198         }
 199         }
 200
 201         return false;
 202 }
 203
 204 static void ec_stripe_key_init(struct bch_fs *c,
 205                                struct bkey_i_stripe *s,
 206                                struct open_buckets *blocks,
 207                                struct open_buckets *parity,
 208                                unsigned stripe_size)
 209 {
 210         struct open_bucket *ob;
 211         unsigned i, u64s;
 212
 213         bkey_stripe_init(&s->k_i);
 214         s->v.sectors                    = cpu_to_le16(stripe_size);
 215         s->v.algorithm                  = 0;
 216         s->v.nr_blocks                  = parity->nr + blocks->nr;
 217         s->v.nr_redundant               = parity->nr;
 218         s->v.csum_granularity_bits      = ilog2(c->sb.encoded_extent_max);
 219         s->v.csum_type                  = BCH_CSUM_CRC32C;
 220         s->v.pad                        = 0;
 221
 222         open_bucket_for_each(c, blocks, ob, i)
 223                 s->v.ptrs[i]                    = ob->ptr;
 224
 225         open_bucket_for_each(c, parity, ob, i)
 226                 s->v.ptrs[blocks->nr + i]       = ob->ptr;
 227
 228         while ((u64s = stripe_val_u64s(&s->v)) > BKEY_VAL_U64s_MAX) {
 229                 BUG_ON(1 << s->v.csum_granularity_bits >=
 230                        le16_to_cpu(s->v.sectors) ||
 231                        s->v.csum_granularity_bits == U8_MAX);
 232                 s->v.csum_granularity_bits++;
 233         }
 234
 235         set_bkey_val_u64s(&s->k, u64s);
 236 }
 237
 238 /* Checksumming: */
 239
 240 static void ec_generate_checksums(struct ec_stripe_buf *buf)
 241 {
 242         struct bch_stripe *v = &buf->key.v;
 243         unsigned csum_granularity = 1 << v->csum_granularity_bits;
 244         unsigned csums_per_device = stripe_csums_per_device(v);
 245         unsigned csum_bytes = bch_crc_bytes[v->csum_type];
 246         unsigned i, j;
 247
 248         if (!csum_bytes)
 249                 return;
 250
 251         BUG_ON(buf->offset);
 252         BUG_ON(buf->size != le16_to_cpu(v->sectors));
 253
 254         for (i = 0; i < v->nr_blocks; i++) {
 255                 for (j = 0; j < csums_per_device; j++) {
 256                         unsigned offset = j << v->csum_granularity_bits;
 257                         unsigned len = min(csum_granularity, buf->size - offset);
 258
 259                         struct bch_csum csum =
 260                                 bch2_checksum(NULL, v->csum_type,
 261                                               null_nonce(),
 262                                               buf->data[i] + (offset << 9),
 263                                               len << 9);
 264
 265                         memcpy(stripe_csum(v, i, j), &csum, csum_bytes);
 266                 }
 267         }
 268 }
 269
 270 static void ec_validate_checksums(struct bch_fs *c, struct ec_stripe_buf *buf)
 271 {
 272         struct bch_stripe *v = &buf->key.v;
 273         unsigned csum_granularity = 1 << v->csum_granularity_bits;
 274         unsigned csum_bytes = bch_crc_bytes[v->csum_type];
 275         unsigned i;
 276
 277         if (!csum_bytes)
 278                 return;
 279
 280         for (i = 0; i < v->nr_blocks; i++) {
 281                 unsigned offset = buf->offset;
 282                 unsigned end = buf->offset + buf->size;
 283
 284                 if (!test_bit(i, buf->valid))
 285                         continue;
 286
 287                 while (offset < end) {
 288                         unsigned j = offset >> v->csum_granularity_bits;
 289                         unsigned len = min(csum_granularity, end - offset);
 290                         struct bch_csum csum;
 291
 292                         BUG_ON(offset & (csum_granularity - 1));
 293                         BUG_ON(offset + len != le16_to_cpu(v->sectors) &&
 294                                ((offset + len) & (csum_granularity - 1)));
 295
 296                         csum = bch2_checksum(NULL, v->csum_type,
 297                                              null_nonce(),
 298                                              buf->data[i] + ((offset - buf->offset) << 9),
 299                                              len << 9);
 300
 301                         if (memcmp(stripe_csum(v, i, j), &csum, csum_bytes)) {
 302                                 __bcache_io_error(c,
 303                                         "checksum error while doing reconstruct read (%u:%u)",
 304                                         i, j);
 305                                 clear_bit(i, buf->valid);
 306                                 break;
 307                         }
 308
 309                         offset += len;
 310                 }
 311         }
 312 }
 313
 314 /* Erasure coding: */
 315
 316 static void ec_generate_ec(struct ec_stripe_buf *buf)
 317 {
 318         struct bch_stripe *v = &buf->key.v;
 319         unsigned nr_data = v->nr_blocks - v->nr_redundant;
 320         unsigned bytes = le16_to_cpu(v->sectors) << 9;
 321
 322         raid_gen(nr_data, v->nr_redundant, bytes, buf->data);
 323 }
 324
 325 static unsigned __ec_nr_failed(struct ec_stripe_buf *buf, unsigned nr)
 326 {
 327         return nr - bitmap_weight(buf->valid, nr);
 328 }
 329
 330 static unsigned ec_nr_failed(struct ec_stripe_buf *buf)
 331 {
 332         return __ec_nr_failed(buf, buf->key.v.nr_blocks);
 333 }
 334
 335 static int ec_do_recov(struct bch_fs *c, struct ec_stripe_buf *buf)
 336 {
 337         struct bch_stripe *v = &buf->key.v;
 338         unsigned i, failed[EC_STRIPE_MAX], nr_failed = 0;
 339         unsigned nr_data = v->nr_blocks - v->nr_redundant;
 340         unsigned bytes = buf->size << 9;
 341
 342         if (ec_nr_failed(buf) > v->nr_redundant) {
 343                 __bcache_io_error(c,
 344                         "error doing reconstruct read: unable to read enough blocks");
 345                 return -1;
 346         }
 347
 348         for (i = 0; i < nr_data; i++)
 349                 if (!test_bit(i, buf->valid))
 350                         failed[nr_failed++] = i;
 351
 352         raid_rec(nr_failed, failed, nr_data, v->nr_redundant, bytes, buf->data);
 353         return 0;
 354 }
 355
 356 /* IO: */
 357
 358 static void ec_block_endio(struct bio *bio)
 359 {
 360         struct ec_bio *ec_bio = container_of(bio, struct ec_bio, bio);
 361         struct bch_dev *ca = ec_bio->ca;
 362         struct closure *cl = bio->bi_private;
 363
 364         if (bch2_dev_io_err_on(bio->bi_status, ca, "erasure coding"))
 365                 clear_bit(ec_bio->idx, ec_bio->buf->valid);
 366
 367         bio_put(&ec_bio->bio);
 368         percpu_ref_put(&ca->io_ref);
 369         closure_put(cl);
 370 }
 371
 372 static void ec_block_io(struct bch_fs *c, struct ec_stripe_buf *buf,
 373                         unsigned rw, unsigned idx, struct closure *cl)
 374 {
 375         struct bch_stripe *v = &buf->key.v;
 376         unsigned offset = 0, bytes = buf->size << 9;
 377         struct bch_extent_ptr *ptr = &v->ptrs[idx];
 378         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 379
 380         if (!bch2_dev_get_ioref(ca, rw)) {
 381                 clear_bit(idx, buf->valid);
 382                 return;
 383         }
 384
 385         while (offset < bytes) {
 386                 unsigned nr_iovecs = min_t(size_t, BIO_MAX_PAGES,
 387                                            DIV_ROUND_UP(bytes, PAGE_SIZE));
 388                 unsigned b = min_t(size_t, bytes - offset,
 389                                    nr_iovecs << PAGE_SHIFT);
 390                 struct ec_bio *ec_bio;
 391
 392                 ec_bio = container_of(bio_alloc_bioset(GFP_KERNEL, nr_iovecs,
 393                                                        &c->ec_bioset),
 394                                       struct ec_bio, bio);
 395
 396                 ec_bio->ca                      = ca;
 397                 ec_bio->buf                     = buf;
 398                 ec_bio->idx                     = idx;
 399
 400                 bio_set_dev(&ec_bio->bio, ca->disk_sb.bdev);
 401                 bio_set_op_attrs(&ec_bio->bio, rw, 0);
 402
 403                 ec_bio->bio.bi_iter.bi_sector   = ptr->offset + buf->offset + (offset >> 9);
 404                 ec_bio->bio.bi_end_io           = ec_block_endio;
 405                 ec_bio->bio.bi_private          = cl;
 406
 407                 bch2_bio_map(&ec_bio->bio, buf->data[idx] + offset, b);
 408
 409                 closure_get(cl);
 410                 percpu_ref_get(&ca->io_ref);
 411
 412                 submit_bio(&ec_bio->bio);
 413
 414                 offset += b;
 415         }
 416
 417         percpu_ref_put(&ca->io_ref);
 418 }
 419
 420 /* recovery read path: */
 421 int bch2_ec_read_extent(struct bch_fs *c, struct bch_read_bio *rbio)
 422 {
 423         struct btree_trans trans;
 424         struct btree_iter *iter;
 425         struct ec_stripe_buf *buf;
 426         struct closure cl;
 427         struct bkey_s_c k;
 428         struct bch_stripe *v;
 429         unsigned stripe_idx;
 430         unsigned offset, end;
 431         unsigned i, nr_data, csum_granularity;
 432         int ret = 0, idx;
 433
 434         closure_init_stack(&cl);
 435
 436         BUG_ON(!rbio->pick.idx ||
 437                rbio->pick.idx - 1 >= rbio->pick.ec_nr);
 438
 439         stripe_idx = rbio->pick.ec[rbio->pick.idx - 1].idx;
 440
 441         buf = kzalloc(sizeof(*buf), GFP_NOIO);
 442         if (!buf)
 443                 return -ENOMEM;
 444
 445         bch2_trans_init(&trans, c, 0, 0);
 446
 447         iter = bch2_trans_get_iter(&trans, BTREE_ID_EC,
 448                                    POS(0, stripe_idx),
 449                                    BTREE_ITER_SLOTS);
 450         k = bch2_btree_iter_peek_slot(iter);
 451         if (bkey_err(k) || k.k->type != KEY_TYPE_stripe) {
 452                 __bcache_io_error(c,
 453                         "error doing reconstruct read: stripe not found");
 454                 kfree(buf);
 455                 return bch2_trans_exit(&trans) ?: -EIO;
 456         }
 457
 458         bkey_reassemble(&buf->key.k_i, k);
 459         bch2_trans_exit(&trans);
 460
 461         v = &buf->key.v;
 462
 463         nr_data = v->nr_blocks - v->nr_redundant;
 464
 465         idx = ptr_matches_stripe(c, v, &rbio->pick.ptr);
 466         BUG_ON(idx < 0);
 467
 468         csum_granularity = 1U << v->csum_granularity_bits;
 469
 470         offset  = rbio->bio.bi_iter.bi_sector - v->ptrs[idx].offset;
 471         end     = offset + bio_sectors(&rbio->bio);
 472
 473         BUG_ON(end > le16_to_cpu(v->sectors));
 474
 475         buf->offset     = round_down(offset, csum_granularity);
 476         buf->size       = min_t(unsigned, le16_to_cpu(v->sectors),
 477                                 round_up(end, csum_granularity)) - buf->offset;
 478
 479         for (i = 0; i < v->nr_blocks; i++) {
 480                 buf->data[i] = kmalloc(buf->size << 9, GFP_NOIO);
 481                 if (!buf->data[i]) {
 482                         ret = -ENOMEM;
 483                         goto err;
 484                 }
 485         }
 486
 487         memset(buf->valid, 0xFF, sizeof(buf->valid));
 488
 489         for (i = 0; i < v->nr_blocks; i++) {
 490                 struct bch_extent_ptr *ptr = v->ptrs + i;
 491                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 492
 493                 if (ptr_stale(ca, ptr)) {
 494                         __bcache_io_error(c,
 495                                           "error doing reconstruct read: stale pointer");
 496                         clear_bit(i, buf->valid);
 497                         continue;
 498                 }
 499
 500                 ec_block_io(c, buf, REQ_OP_READ, i, &cl);
 501         }
 502
 503         closure_sync(&cl);
 504
 505         if (ec_nr_failed(buf) > v->nr_redundant) {
 506                 __bcache_io_error(c,
 507                         "error doing reconstruct read: unable to read enough blocks");
 508                 ret = -EIO;
 509                 goto err;
 510         }
 511
 512         ec_validate_checksums(c, buf);
 513
 514         ret = ec_do_recov(c, buf);
 515         if (ret)
 516                 goto err;
 517
 518         memcpy_to_bio(&rbio->bio, rbio->bio.bi_iter,
 519                       buf->data[idx] + ((offset - buf->offset) << 9));
 520 err:
 521         for (i = 0; i < v->nr_blocks; i++)
 522                 kfree(buf->data[i]);
 523         kfree(buf);
 524         return ret;
 525 }
 526
 527 /* stripe bucket accounting: */
 528
 529 static int __ec_stripe_mem_alloc(struct bch_fs *c, size_t idx, gfp_t gfp)
 530 {
 531         ec_stripes_heap n, *h = &c->ec_stripes_heap;
 532
 533         if (idx >= h->size) {
 534                 if (!init_heap(&n, max(1024UL, roundup_pow_of_two(idx + 1)), gfp))
 535                         return -ENOMEM;
 536
 537                 spin_lock(&c->ec_stripes_heap_lock);
 538                 if (n.size > h->size) {
 539                         memcpy(n.data, h->data, h->used * sizeof(h->data[0]));
 540                         n.used = h->used;
 541                         swap(*h, n);
 542                 }
 543                 spin_unlock(&c->ec_stripes_heap_lock);
 544
 545                 free_heap(&n);
 546         }
 547
 548         if (!genradix_ptr_alloc(&c->stripes[0], idx, gfp))
 549                 return -ENOMEM;
 550
 551         if (c->gc_pos.phase != GC_PHASE_NOT_RUNNING &&
 552             !genradix_ptr_alloc(&c->stripes[1], idx, gfp))
 553                 return -ENOMEM;
 554
 555         return 0;
 556 }
 557
 558 static int ec_stripe_mem_alloc(struct bch_fs *c,
 559                                struct btree_iter *iter)
 560 {
 561         size_t idx = iter->pos.offset;
 562         int ret = 0;
 563
 564         if (!__ec_stripe_mem_alloc(c, idx, GFP_NOWAIT))
 565                 return ret;
 566
 567         bch2_trans_unlock(iter->trans);
 568         ret = -EINTR;
 569
 570         if (!__ec_stripe_mem_alloc(c, idx, GFP_KERNEL))
 571                 return ret;
 572
 573         return -ENOMEM;
 574 }
 575
 576 static ssize_t stripe_idx_to_delete(struct bch_fs *c)
 577 {
 578         ec_stripes_heap *h = &c->ec_stripes_heap;
 579
 580         return h->used && h->data[0].blocks_nonempty == 0
 581                 ? h->data[0].idx : -1;
 582 }
 583
 584 static inline int ec_stripes_heap_cmp(ec_stripes_heap *h,
 585                                       struct ec_stripe_heap_entry l,
 586                                       struct ec_stripe_heap_entry r)
 587 {
 588         return ((l.blocks_nonempty > r.blocks_nonempty) -
 589                 (l.blocks_nonempty < r.blocks_nonempty));
 590 }
 591
 592 static inline void ec_stripes_heap_set_backpointer(ec_stripes_heap *h,
 593                                                    size_t i)
 594 {
 595         struct bch_fs *c = container_of(h, struct bch_fs, ec_stripes_heap);
 596
 597         genradix_ptr(&c->stripes[0], h->data[i].idx)->heap_idx = i;
 598 }
 599
 600 static void heap_verify_backpointer(struct bch_fs *c, size_t idx)
 601 {
 602         ec_stripes_heap *h = &c->ec_stripes_heap;
 603         struct stripe *m = genradix_ptr(&c->stripes[0], idx);
 604
 605         BUG_ON(!m->alive);
 606         BUG_ON(m->heap_idx >= h->used);
 607         BUG_ON(h->data[m->heap_idx].idx != idx);
 608 }
 609
 610 void bch2_stripes_heap_update(struct bch_fs *c,
 611                               struct stripe *m, size_t idx)
 612 {
 613         ec_stripes_heap *h = &c->ec_stripes_heap;
 614         size_t i;
 615
 616         if (m->alive) {
 617                 heap_verify_backpointer(c, idx);
 618
 619                 h->data[m->heap_idx].blocks_nonempty = m->blocks_nonempty;
 620
 621                 i = m->heap_idx;
 622                 heap_sift_up(h,   i, ec_stripes_heap_cmp,
 623                              ec_stripes_heap_set_backpointer);
 624                 heap_sift_down(h, i, ec_stripes_heap_cmp,
 625                                ec_stripes_heap_set_backpointer);
 626
 627                 heap_verify_backpointer(c, idx);
 628         } else {
 629                 bch2_stripes_heap_insert(c, m, idx);
 630         }
 631
 632         if (stripe_idx_to_delete(c) >= 0 &&
 633             !percpu_ref_is_dying(&c->writes))
 634                 schedule_work(&c->ec_stripe_delete_work);
 635 }
 636
 637 void bch2_stripes_heap_del(struct bch_fs *c,
 638                            struct stripe *m, size_t idx)
 639 {
 640         heap_verify_backpointer(c, idx);
 641
 642         m->alive = false;
 643         heap_del(&c->ec_stripes_heap, m->heap_idx,
 644                  ec_stripes_heap_cmp,
 645                  ec_stripes_heap_set_backpointer);
 646 }
 647
 648 void bch2_stripes_heap_insert(struct bch_fs *c,
 649                               struct stripe *m, size_t idx)
 650 {
 651         BUG_ON(heap_full(&c->ec_stripes_heap));
 652
 653         heap_add(&c->ec_stripes_heap, ((struct ec_stripe_heap_entry) {
 654                         .idx = idx,
 655                         .blocks_nonempty = m->blocks_nonempty,
 656                 }),
 657                  ec_stripes_heap_cmp,
 658                  ec_stripes_heap_set_backpointer);
 659         m->alive = true;
 660
 661         heap_verify_backpointer(c, idx);
 662 }
 663
 664 /* stripe deletion */
 665
 666 static int ec_stripe_delete(struct bch_fs *c, size_t idx)
 667 {
 668         return bch2_btree_delete_range(c, BTREE_ID_EC,
 669                                        POS(0, idx),
 670                                        POS(0, idx + 1),
 671                                        NULL);
 672 }
 673
 674 static void ec_stripe_delete_work(struct work_struct *work)
 675 {
 676         struct bch_fs *c =
 677                 container_of(work, struct bch_fs, ec_stripe_delete_work);
 678         ssize_t idx;
 679
 680         down_read(&c->gc_lock);
 681         mutex_lock(&c->ec_stripe_create_lock);
 682
 683         while (1) {
 684                 spin_lock(&c->ec_stripes_heap_lock);
 685                 idx = stripe_idx_to_delete(c);
 686                 spin_unlock(&c->ec_stripes_heap_lock);
 687
 688                 if (idx < 0)
 689                         break;
 690
 691                 if (ec_stripe_delete(c, idx))
 692                         break;
 693         }
 694
 695         mutex_unlock(&c->ec_stripe_create_lock);
 696         up_read(&c->gc_lock);
 697 }
 698
 699 /* stripe creation: */
 700
 701 static int ec_stripe_bkey_insert(struct bch_fs *c,
 702                                  struct bkey_i_stripe *stripe)
 703 {
 704         struct btree_trans trans;
 705         struct btree_iter *iter;
 706         struct bkey_s_c k;
 707         int ret;
 708
 709         bch2_trans_init(&trans, c, 0, 0);
 710 retry:
 711         bch2_trans_begin(&trans);
 712
 713         /* XXX: start pos hint */
 714         for_each_btree_key(&trans, iter, BTREE_ID_EC, POS_MIN,
 715                            BTREE_ITER_SLOTS|BTREE_ITER_INTENT, k, ret) {
 716                 if (bkey_cmp(k.k->p, POS(0, U32_MAX)) > 0)
 717                         break;
 718
 719                 if (bkey_deleted(k.k))
 720                         goto found_slot;
 721         }
 722
 723         if (!ret)
 724                 ret = -ENOSPC;
 725         goto err;
 726 found_slot:
 727         ret = ec_stripe_mem_alloc(c, iter);
 728         if (ret)
 729                 goto err;
 730
 731         stripe->k.p = iter->pos;
 732
 733         bch2_trans_update(&trans, BTREE_INSERT_ENTRY(iter, &stripe->k_i));
 734
 735         ret = bch2_trans_commit(&trans, NULL, NULL,
 736                                 BTREE_INSERT_ATOMIC|
 737                                 BTREE_INSERT_NOFAIL);
 738 err:
 739         if (ret == -EINTR)
 740                 goto retry;
 741         bch2_trans_exit(&trans);
 742
 743         return ret;
 744 }
 745
 746 static void extent_stripe_ptr_add(struct bkey_s_extent e,
 747                                   struct ec_stripe_buf *s,
 748                                   struct bch_extent_ptr *ptr,
 749                                   unsigned block)
 750 {
 751         struct bch_extent_stripe_ptr *dst = (void *) ptr;
 752         union bch_extent_entry *end = extent_entry_last(e);
 753
 754         memmove_u64s_up(dst + 1, dst, (u64 *) end - (u64 *) dst);
 755         e.k->u64s += sizeof(*dst) / sizeof(u64);
 756
 757         *dst = (struct bch_extent_stripe_ptr) {
 758                 .type = 1 << BCH_EXTENT_ENTRY_stripe_ptr,
 759                 .block          = block,
 760                 .idx            = s->key.k.p.offset,
 761         };
 762 }
 763
 764 static int ec_stripe_update_ptrs(struct bch_fs *c,
 765                                  struct ec_stripe_buf *s,
 766                                  struct bkey *pos)
 767 {
 768         struct btree_trans trans;
 769         struct btree_iter *iter;
 770         struct bkey_s_c k;
 771         struct bkey_s_extent e;
 772         struct bch_extent_ptr *ptr;
 773         BKEY_PADDED(k) tmp;
 774         int ret = 0, dev, idx;
 775
 776         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
 777
 778         iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
 779                                    bkey_start_pos(pos),
 780                                    BTREE_ITER_INTENT);
 781
 782         while ((k = bch2_btree_iter_peek(iter)).k &&
 783                !(ret = bkey_err(k)) &&
 784                bkey_cmp(bkey_start_pos(k.k), pos->p) < 0) {
 785                 if (extent_has_stripe_ptr(k, s->key.k.p.offset)) {
 786                         bch2_btree_iter_next(iter);
 787                         continue;
 788                 }
 789
 790                 idx = extent_matches_stripe(c, &s->key.v, k);
 791                 if (idx < 0) {
 792                         bch2_btree_iter_next(iter);
 793                         continue;
 794                 }
 795
 796                 bch2_btree_iter_set_pos(iter, bkey_start_pos(k.k));
 797
 798                 dev = s->key.v.ptrs[idx].dev;
 799
 800                 bkey_reassemble(&tmp.k, k);
 801                 e = bkey_i_to_s_extent(&tmp.k);
 802
 803                 extent_for_each_ptr(e, ptr)
 804                         if (ptr->dev != dev)
 805                                 ptr->cached = true;
 806
 807                 ptr = (void *) bch2_extent_has_device(e.c, dev);
 808                 BUG_ON(!ptr);
 809
 810                 extent_stripe_ptr_add(e, s, ptr, idx);
 811
 812                 bch2_trans_update(&trans, BTREE_INSERT_ENTRY(iter, &tmp.k));
 813
 814                 ret = bch2_trans_commit(&trans, NULL, NULL,
 815                                         BTREE_INSERT_ATOMIC|
 816                                         BTREE_INSERT_NOFAIL|
 817                                         BTREE_INSERT_USE_RESERVE);
 818                 if (ret == -EINTR)
 819                         ret = 0;
 820                 if (ret)
 821                         break;
 822         }
 823
 824         bch2_trans_exit(&trans);
 825
 826         return ret;
 827 }
 828
 829 /*
 830  * data buckets of new stripe all written: create the stripe
 831  */
 832 static void ec_stripe_create(struct ec_stripe_new *s)
 833 {
 834         struct bch_fs *c = s->c;
 835         struct open_bucket *ob;
 836         struct bkey_i *k;
 837         struct bch_stripe *v = &s->stripe.key.v;
 838         unsigned i, nr_data = v->nr_blocks - v->nr_redundant;
 839         struct closure cl;
 840         int ret;
 841
 842         BUG_ON(s->h->s == s);
 843
 844         closure_init_stack(&cl);
 845
 846         if (s->err) {
 847                 bch_err(c, "error creating stripe: error writing data buckets");
 848                 goto err;
 849         }
 850
 851         if (!percpu_ref_tryget(&c->writes))
 852                 goto err;
 853
 854         BUG_ON(bitmap_weight(s->blocks_allocated,
 855                              s->blocks.nr) != s->blocks.nr);
 856
 857         ec_generate_ec(&s->stripe);
 858
 859         ec_generate_checksums(&s->stripe);
 860
 861         /* write p/q: */
 862         for (i = nr_data; i < v->nr_blocks; i++)
 863                 ec_block_io(c, &s->stripe, REQ_OP_WRITE, i, &cl);
 864
 865         closure_sync(&cl);
 866
 867         for (i = nr_data; i < v->nr_blocks; i++)
 868                 if (!test_bit(i, s->stripe.valid)) {
 869                         bch_err(c, "error creating stripe: error writing redundancy buckets");
 870                         goto err_put_writes;
 871                 }
 872
 873         mutex_lock(&c->ec_stripe_create_lock);
 874
 875         ret = ec_stripe_bkey_insert(c, &s->stripe.key);
 876         if (ret) {
 877                 bch_err(c, "error creating stripe: error creating stripe key");
 878                 goto err_unlock;
 879         }
 880
 881         for_each_keylist_key(&s->keys, k) {
 882                 ret = ec_stripe_update_ptrs(c, &s->stripe, &k->k);
 883                 if (ret)
 884                         break;
 885         }
 886
 887 err_unlock:
 888         mutex_unlock(&c->ec_stripe_create_lock);
 889 err_put_writes:
 890         percpu_ref_put(&c->writes);
 891 err:
 892         open_bucket_for_each(c, &s->blocks, ob, i) {
 893                 ob->ec = NULL;
 894                 __bch2_open_bucket_put(c, ob);
 895         }
 896
 897         bch2_open_buckets_put(c, &s->parity);
 898
 899         bch2_keylist_free(&s->keys, s->inline_keys);
 900
 901         mutex_lock(&s->h->lock);
 902         list_del(&s->list);
 903         mutex_unlock(&s->h->lock);
 904
 905         for (i = 0; i < s->stripe.key.v.nr_blocks; i++)
 906                 kvpfree(s->stripe.data[i], s->stripe.size << 9);
 907         kfree(s);
 908 }
 909
 910 static struct ec_stripe_new *ec_stripe_set_pending(struct ec_stripe_head *h)
 911 {
 912         struct ec_stripe_new *s = h->s;
 913
 914         list_add(&s->list, &h->stripes);
 915         h->s = NULL;
 916
 917         return s;
 918 }
 919
 920 static void ec_stripe_new_put(struct ec_stripe_new *s)
 921 {
 922         BUG_ON(atomic_read(&s->pin) <= 0);
 923         if (atomic_dec_and_test(&s->pin))
 924                 ec_stripe_create(s);
 925 }
 926
 927 /* have a full bucket - hand it off to be erasure coded: */
 928 void bch2_ec_bucket_written(struct bch_fs *c, struct open_bucket *ob)
 929 {
 930         struct ec_stripe_new *s = ob->ec;
 931
 932         if (ob->sectors_free)
 933                 s->err = -1;
 934
 935         ec_stripe_new_put(s);
 936 }
 937
 938 void bch2_ec_bucket_cancel(struct bch_fs *c, struct open_bucket *ob)
 939 {
 940         struct ec_stripe_new *s = ob->ec;
 941
 942         s->err = -EIO;
 943 }
 944
 945 void *bch2_writepoint_ec_buf(struct bch_fs *c, struct write_point *wp)
 946 {
 947         struct open_bucket *ob = ec_open_bucket(c, &wp->ptrs);
 948         struct bch_dev *ca;
 949         unsigned offset;
 950
 951         if (!ob)
 952                 return NULL;
 953
 954         ca      = bch_dev_bkey_exists(c, ob->ptr.dev);
 955         offset  = ca->mi.bucket_size - ob->sectors_free;
 956
 957         return ob->ec->stripe.data[ob->ec_idx] + (offset << 9);
 958 }
 959
 960 void bch2_ec_add_backpointer(struct bch_fs *c, struct write_point *wp,
 961                              struct bpos pos, unsigned sectors)
 962 {
 963         struct open_bucket *ob = ec_open_bucket(c, &wp->ptrs);
 964         struct ec_stripe_new *ec;
 965
 966         if (!ob)
 967                 return;
 968
 969         ec = ob->ec;
 970         mutex_lock(&ec->lock);
 971
 972         if (bch2_keylist_realloc(&ec->keys, ec->inline_keys,
 973                                  ARRAY_SIZE(ec->inline_keys),
 974                                  BKEY_U64s)) {
 975                 BUG();
 976         }
 977
 978         bkey_init(&ec->keys.top->k);
 979         ec->keys.top->k.p       = pos;
 980         bch2_key_resize(&ec->keys.top->k, sectors);
 981         bch2_keylist_push(&ec->keys);
 982
 983         mutex_unlock(&ec->lock);
 984 }
 985
 986 static int unsigned_cmp(const void *_l, const void *_r)
 987 {
 988         unsigned l = *((const unsigned *) _l);
 989         unsigned r = *((const unsigned *) _r);
 990
 991         return cmp_int(l, r);
 992 }
 993
 994 /* pick most common bucket size: */
 995 static unsigned pick_blocksize(struct bch_fs *c,
 996                                struct bch_devs_mask *devs)
 997 {
 998         struct bch_dev *ca;
 999         unsigned i, nr = 0, sizes[BCH_SB_MEMBERS_MAX];
1000         struct {
1001                 unsigned nr, size;
1002         } cur = { 0, 0 }, best = { 0, 0 };
1003
1004         for_each_member_device_rcu(ca, c, i, devs)
1005                 sizes[nr++] = ca->mi.bucket_size;
1006
1007         sort(sizes, nr, sizeof(unsigned), unsigned_cmp, NULL);
1008
1009         for (i = 0; i < nr; i++) {
1010                 if (sizes[i] != cur.size) {
1011                         if (cur.nr > best.nr)
1012                                 best = cur;
1013
1014                         cur.nr = 0;
1015                         cur.size = sizes[i];
1016                 }
1017
1018                 cur.nr++;
1019         }
1020
1021         if (cur.nr > best.nr)
1022                 best = cur;
1023
1024         return best.size;
1025 }
1026
1027 int bch2_ec_stripe_new_alloc(struct bch_fs *c, struct ec_stripe_head *h)
1028 {
1029         struct ec_stripe_new *s;
1030         unsigned i;
1031
1032         BUG_ON(h->parity.nr != h->redundancy);
1033         BUG_ON(!h->blocks.nr);
1034         BUG_ON(h->parity.nr + h->blocks.nr > EC_STRIPE_MAX);
1035         lockdep_assert_held(&h->lock);
1036
1037         s = kzalloc(sizeof(*s), GFP_KERNEL);
1038         if (!s)
1039                 return -ENOMEM;
1040
1041         mutex_init(&s->lock);
1042         atomic_set(&s->pin, 1);
1043         s->c            = c;
1044         s->h            = h;
1045         s->blocks       = h->blocks;
1046         s->parity       = h->parity;
1047
1048         memset(&h->blocks, 0, sizeof(h->blocks));
1049         memset(&h->parity, 0, sizeof(h->parity));
1050
1051         bch2_keylist_init(&s->keys, s->inline_keys);
1052
1053         s->stripe.offset        = 0;
1054         s->stripe.size          = h->blocksize;
1055         memset(s->stripe.valid, 0xFF, sizeof(s->stripe.valid));
1056
1057         ec_stripe_key_init(c, &s->stripe.key,
1058                            &s->blocks, &s->parity,
1059                            h->blocksize);
1060
1061         for (i = 0; i < s->stripe.key.v.nr_blocks; i++) {
1062                 s->stripe.data[i] = kvpmalloc(s->stripe.size << 9, GFP_KERNEL);
1063                 if (!s->stripe.data[i])
1064                         goto err;
1065         }
1066
1067         h->s = s;
1068
1069         return 0;
1070 err:
1071         for (i = 0; i < s->stripe.key.v.nr_blocks; i++)
1072                 kvpfree(s->stripe.data[i], s->stripe.size << 9);
1073         kfree(s);
1074         return -ENOMEM;
1075 }
1076
1077 static struct ec_stripe_head *
1078 ec_new_stripe_head_alloc(struct bch_fs *c, unsigned target,
1079                          unsigned algo, unsigned redundancy)
1080 {
1081         struct ec_stripe_head *h;
1082         struct bch_dev *ca;
1083         unsigned i;
1084
1085         h = kzalloc(sizeof(*h), GFP_KERNEL);
1086         if (!h)
1087                 return NULL;
1088
1089         mutex_init(&h->lock);
1090         mutex_lock(&h->lock);
1091         INIT_LIST_HEAD(&h->stripes);
1092
1093         h->target       = target;
1094         h->algo         = algo;
1095         h->redundancy   = redundancy;
1096
1097         rcu_read_lock();
1098         h->devs = target_rw_devs(c, BCH_DATA_USER, target);
1099
1100         for_each_member_device_rcu(ca, c, i, &h->devs)
1101                 if (!ca->mi.durability)
1102                         __clear_bit(i, h->devs.d);
1103
1104         h->blocksize = pick_blocksize(c, &h->devs);
1105
1106         for_each_member_device_rcu(ca, c, i, &h->devs)
1107                 if (ca->mi.bucket_size == h->blocksize)
1108                         h->nr_active_devs++;
1109
1110         rcu_read_unlock();
1111         list_add(&h->list, &c->ec_new_stripe_list);
1112         return h;
1113 }
1114
1115 void bch2_ec_stripe_head_put(struct ec_stripe_head *h)
1116 {
1117         struct ec_stripe_new *s = NULL;
1118
1119         if (h->s &&
1120             bitmap_weight(h->s->blocks_allocated,
1121                           h->s->blocks.nr) == h->s->blocks.nr)
1122                 s = ec_stripe_set_pending(h);
1123
1124         mutex_unlock(&h->lock);
1125
1126         if (s)
1127                 ec_stripe_new_put(s);
1128 }
1129
1130 struct ec_stripe_head *bch2_ec_stripe_head_get(struct bch_fs *c,
1131                                                unsigned target,
1132                                                unsigned algo,
1133                                                unsigned redundancy)
1134 {
1135         struct ec_stripe_head *h;
1136
1137         if (!redundancy)
1138                 return NULL;
1139
1140         mutex_lock(&c->ec_new_stripe_lock);
1141         list_for_each_entry(h, &c->ec_new_stripe_list, list)
1142                 if (h->target           == target &&
1143                     h->algo             == algo &&
1144                     h->redundancy       == redundancy) {
1145                         mutex_lock(&h->lock);
1146                         goto found;
1147                 }
1148
1149         h = ec_new_stripe_head_alloc(c, target, algo, redundancy);
1150 found:
1151         mutex_unlock(&c->ec_new_stripe_lock);
1152         return h;
1153 }
1154
1155 void bch2_ec_stop_dev(struct bch_fs *c, struct bch_dev *ca)
1156 {
1157         struct ec_stripe_head *h;
1158         struct open_bucket *ob;
1159         unsigned i;
1160
1161         mutex_lock(&c->ec_new_stripe_lock);
1162         list_for_each_entry(h, &c->ec_new_stripe_list, list) {
1163                 struct ec_stripe_new *s = NULL;
1164
1165                 mutex_lock(&h->lock);
1166                 bch2_open_buckets_stop_dev(c, ca,
1167                                            &h->blocks,
1168                                            BCH_DATA_USER);
1169                 bch2_open_buckets_stop_dev(c, ca,
1170                                            &h->parity,
1171                                            BCH_DATA_USER);
1172
1173                 if (!h->s)
1174                         goto unlock;
1175
1176                 open_bucket_for_each(c, &h->s->blocks, ob, i)
1177                         if (ob->ptr.dev == ca->dev_idx)
1178                                 goto found;
1179                 open_bucket_for_each(c, &h->s->parity, ob, i)
1180                         if (ob->ptr.dev == ca->dev_idx)
1181                                 goto found;
1182                 goto unlock;
1183 found:
1184                 h->s->err = -1;
1185                 s = ec_stripe_set_pending(h);
1186 unlock:
1187                 mutex_unlock(&h->lock);
1188
1189                 if (s)
1190                         ec_stripe_new_put(s);
1191         }
1192         mutex_unlock(&c->ec_new_stripe_lock);
1193 }
1194
1195 static int __bch2_stripe_write_key(struct btree_trans *trans,
1196                                    struct btree_iter *iter,
1197                                    struct stripe *m,
1198                                    size_t idx,
1199                                    struct bkey_i_stripe *new_key,
1200                                    unsigned flags)
1201 {
1202         struct bch_fs *c = trans->c;
1203         struct bkey_s_c k;
1204         unsigned i;
1205         int ret;
1206
1207         bch2_btree_iter_set_pos(iter, POS(0, idx));
1208
1209         k = bch2_btree_iter_peek_slot(iter);
1210         ret = bkey_err(k);
1211         if (ret)
1212                 return ret;
1213
1214         if (k.k->type != KEY_TYPE_stripe)
1215                 return -EIO;
1216
1217         bkey_reassemble(&new_key->k_i, k);
1218
1219         spin_lock(&c->ec_stripes_heap_lock);
1220
1221         for (i = 0; i < new_key->v.nr_blocks; i++)
1222                 stripe_blockcount_set(&new_key->v, i,
1223                                       m->block_sectors[i]);
1224         m->dirty = false;
1225
1226         spin_unlock(&c->ec_stripes_heap_lock);
1227
1228         bch2_trans_update(trans, BTREE_INSERT_ENTRY(iter, &new_key->k_i));
1229
1230         return bch2_trans_commit(trans, NULL, NULL,
1231                                  BTREE_INSERT_NOFAIL|flags);
1232 }
1233
1234 int bch2_stripes_write(struct bch_fs *c, unsigned flags, bool *wrote)
1235 {
1236         struct btree_trans trans;
1237         struct btree_iter *iter;
1238         struct genradix_iter giter;
1239         struct bkey_i_stripe *new_key;
1240         struct stripe *m;
1241         int ret = 0;
1242
1243         new_key = kmalloc(255 * sizeof(u64), GFP_KERNEL);
1244         BUG_ON(!new_key);
1245
1246         bch2_trans_init(&trans, c, 0, 0);
1247
1248         iter = bch2_trans_get_iter(&trans, BTREE_ID_EC, POS_MIN,
1249                                    BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
1250
1251         genradix_for_each(&c->stripes[0], giter, m) {
1252                 if (!m->dirty)
1253                         continue;
1254
1255                 ret = __bch2_stripe_write_key(&trans, iter, m, giter.pos,
1256                                               new_key, flags);
1257                 if (ret)
1258                         break;
1259
1260                 *wrote = true;
1261         }
1262
1263         bch2_trans_exit(&trans);
1264
1265         kfree(new_key);
1266
1267         return ret;
1268 }
1269
1270 int bch2_stripes_read(struct bch_fs *c, struct journal_keys *journal_keys)
1271 {
1272         struct btree_trans trans;
1273         struct btree_iter *btree_iter;
1274         struct journal_iter journal_iter;
1275         struct bkey_s_c btree_k, journal_k, k;
1276         int ret;
1277
1278         ret = bch2_fs_ec_start(c);
1279         if (ret)
1280                 return ret;
1281
1282         bch2_trans_init(&trans, c, 0, 0);
1283
1284         btree_iter      = bch2_trans_get_iter(&trans, BTREE_ID_EC, POS_MIN, 0);
1285         journal_iter    = bch2_journal_iter_init(journal_keys, BTREE_ID_EC);
1286
1287         btree_k         = bch2_btree_iter_peek(btree_iter);
1288         journal_k       = bch2_journal_iter_peek(&journal_iter);
1289
1290         while (1) {
1291                 if (btree_k.k && journal_k.k) {
1292                         int cmp = bkey_cmp(btree_k.k->p, journal_k.k->p);
1293
1294                         if (cmp < 0) {
1295                                 k = btree_k;
1296                                 btree_k = bch2_btree_iter_next(btree_iter);
1297                         } else if (cmp == 0) {
1298                                 btree_k = bch2_btree_iter_next(btree_iter);
1299                                 k = journal_k;
1300                                 journal_k = bch2_journal_iter_next(&journal_iter);
1301                         } else {
1302                                 k = journal_k;
1303                                 journal_k = bch2_journal_iter_next(&journal_iter);
1304                         }
1305                 } else if (btree_k.k) {
1306                         k = btree_k;
1307                         btree_k = bch2_btree_iter_next(btree_iter);
1308                 } else if (journal_k.k) {
1309                         k = journal_k;
1310                         journal_k = bch2_journal_iter_next(&journal_iter);
1311                 } else {
1312                         break;
1313                 }
1314
1315                 bch2_mark_key(c, k, 0, 0, NULL, 0,
1316                               BCH_BUCKET_MARK_ALLOC_READ|
1317                               BCH_BUCKET_MARK_NOATOMIC);
1318         }
1319
1320         ret = bch2_trans_exit(&trans) ?: ret;
1321         if (ret) {
1322                 bch_err(c, "error reading stripes: %i", ret);
1323                 return ret;
1324         }
1325
1326         return 0;
1327 }
1328
1329 int bch2_ec_mem_alloc(struct bch_fs *c, bool gc)
1330 {
1331         struct btree_trans trans;
1332         struct btree_iter *iter;
1333         struct bkey_s_c k;
1334         size_t i, idx = 0;
1335         int ret = 0;
1336
1337         bch2_trans_init(&trans, c, 0, 0);
1338
1339         iter = bch2_trans_get_iter(&trans, BTREE_ID_EC, POS(0, U64_MAX), 0);
1340
1341         k = bch2_btree_iter_prev(iter);
1342         if (!IS_ERR_OR_NULL(k.k))
1343                 idx = k.k->p.offset + 1;
1344         ret = bch2_trans_exit(&trans);
1345         if (ret)
1346                 return ret;
1347
1348         if (!gc &&
1349             !init_heap(&c->ec_stripes_heap, roundup_pow_of_two(idx),
1350                        GFP_KERNEL))
1351                 return -ENOMEM;
1352 #if 0
1353         ret = genradix_prealloc(&c->stripes[gc], idx, GFP_KERNEL);
1354 #else
1355         for (i = 0; i < idx; i++)
1356                 if (!genradix_ptr_alloc(&c->stripes[gc], i, GFP_KERNEL))
1357                         return -ENOMEM;
1358 #endif
1359         return 0;
1360 }
1361
1362 int bch2_fs_ec_start(struct bch_fs *c)
1363 {
1364         return bch2_ec_mem_alloc(c, false);
1365 }
1366
1367 void bch2_fs_ec_exit(struct bch_fs *c)
1368 {
1369         struct ec_stripe_head *h;
1370
1371         while (1) {
1372                 mutex_lock(&c->ec_new_stripe_lock);
1373                 h = list_first_entry_or_null(&c->ec_new_stripe_list,
1374                                              struct ec_stripe_head, list);
1375                 if (h)
1376                         list_del(&h->list);
1377                 mutex_unlock(&c->ec_new_stripe_lock);
1378                 if (!h)
1379                         break;
1380
1381                 BUG_ON(h->s);
1382                 BUG_ON(!list_empty(&h->stripes));
1383                 kfree(h);
1384         }
1385
1386         free_heap(&c->ec_stripes_heap);
1387         genradix_free(&c->stripes[0]);
1388         bioset_exit(&c->ec_bioset);
1389 }
1390
1391 int bch2_fs_ec_init(struct bch_fs *c)
1392 {
1393         INIT_WORK(&c->ec_stripe_delete_work, ec_stripe_delete_work);
1394
1395         return bioset_init(&c->ec_bioset, 1, offsetof(struct ec_bio, bio),
1396                            BIOSET_NEED_BVECS);
1397 }