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 "bkey_buf.h"
   8 #include "bset.h"
   9 #include "btree_gc.h"
  10 #include "btree_update.h"
  11 #include "buckets.h"
  12 #include "disk_groups.h"
  13 #include "ec.h"
  14 #include "error.h"
  15 #include "io.h"
  16 #include "keylist.h"
  17 #include "recovery.h"
  18 #include "replicas.h"
  19 #include "super-io.h"
  20 #include "util.h"
  21
  22 #include <linux/sort.h>
  23
  24 #ifdef __KERNEL__
  25
  26 #include <linux/raid/pq.h>
  27 #include <linux/raid/xor.h>
  28
  29 static void raid5_recov(unsigned disks, unsigned failed_idx,
  30                         size_t size, void **data)
  31 {
  32         unsigned i = 2, nr;
  33
  34         BUG_ON(failed_idx >= disks);
  35
  36         swap(data[0], data[failed_idx]);
  37         memcpy(data[0], data[1], size);
  38
  39         while (i < disks) {
  40                 nr = min_t(unsigned, disks - i, MAX_XOR_BLOCKS);
  41                 xor_blocks(nr, size, data[0], data + i);
  42                 i += nr;
  43         }
  44
  45         swap(data[0], data[failed_idx]);
  46 }
  47
  48 static void raid_gen(int nd, int np, size_t size, void **v)
  49 {
  50         if (np >= 1)
  51                 raid5_recov(nd + np, nd, size, v);
  52         if (np >= 2)
  53                 raid6_call.gen_syndrome(nd + np, size, v);
  54         BUG_ON(np > 2);
  55 }
  56
  57 static void raid_rec(int nr, int *ir, int nd, int np, size_t size, void **v)
  58 {
  59         switch (nr) {
  60         case 0:
  61                 break;
  62         case 1:
  63                 if (ir[0] < nd + 1)
  64                         raid5_recov(nd + 1, ir[0], size, v);
  65                 else
  66                         raid6_call.gen_syndrome(nd + np, size, v);
  67                 break;
  68         case 2:
  69                 if (ir[1] < nd) {
  70                         /* data+data failure. */
  71                         raid6_2data_recov(nd + np, size, ir[0], ir[1], v);
  72                 } else if (ir[0] < nd) {
  73                         /* data + p/q failure */
  74
  75                         if (ir[1] == nd) /* data + p failure */
  76                                 raid6_datap_recov(nd + np, size, ir[0], v);
  77                         else { /* data + q failure */
  78                                 raid5_recov(nd + 1, ir[0], size, v);
  79                                 raid6_call.gen_syndrome(nd + np, size, v);
  80                         }
  81                 } else {
  82                         raid_gen(nd, np, size, v);
  83                 }
  84                 break;
  85         default:
  86                 BUG();
  87         }
  88 }
  89
  90 #else
  91
  92 #include <raid/raid.h>
  93
  94 #endif
  95
  96 struct ec_bio {
  97         struct bch_dev          *ca;
  98         struct ec_stripe_buf    *buf;
  99         size_t                  idx;
 100         struct bio              bio;
 101 };
 102
 103 /* Stripes btree keys: */
 104
 105 const char *bch2_stripe_invalid(const struct bch_fs *c, struct bkey_s_c k)
 106 {
 107         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 108
 109         if (!bkey_cmp(k.k->p, POS_MIN))
 110                 return "stripe at pos 0";
 111
 112         if (k.k->p.inode)
 113                 return "invalid stripe key";
 114
 115         if (bkey_val_bytes(k.k) < sizeof(*s))
 116                 return "incorrect value size";
 117
 118         if (bkey_val_bytes(k.k) < sizeof(*s) ||
 119             bkey_val_u64s(k.k) < stripe_val_u64s(s))
 120                 return "incorrect value size";
 121
 122         return bch2_bkey_ptrs_invalid(c, k);
 123 }
 124
 125 void bch2_stripe_to_text(struct printbuf *out, struct bch_fs *c,
 126                          struct bkey_s_c k)
 127 {
 128         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 129         unsigned i;
 130
 131         pr_buf(out, "algo %u sectors %u blocks %u:%u csum %u gran %u",
 132                s->algorithm,
 133                le16_to_cpu(s->sectors),
 134                s->nr_blocks - s->nr_redundant,
 135                s->nr_redundant,
 136                s->csum_type,
 137                1U << s->csum_granularity_bits);
 138
 139         for (i = 0; i < s->nr_blocks; i++)
 140                 pr_buf(out, " %u:%llu:%u", s->ptrs[i].dev,
 141                        (u64) s->ptrs[i].offset,
 142                        stripe_blockcount_get(s, i));
 143 }
 144
 145 /* returns blocknr in stripe that we matched: */
 146 static const struct bch_extent_ptr *bkey_matches_stripe(struct bch_stripe *s,
 147                                                 struct bkey_s_c k, unsigned *block)
 148 {
 149         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 150         const struct bch_extent_ptr *ptr;
 151         unsigned i, nr_data = s->nr_blocks - s->nr_redundant;
 152
 153         bkey_for_each_ptr(ptrs, ptr)
 154                 for (i = 0; i < nr_data; i++)
 155                         if (__bch2_ptr_matches_stripe(&s->ptrs[i], ptr,
 156                                                       le16_to_cpu(s->sectors))) {
 157                                 *block = i;
 158                                 return ptr;
 159                         }
 160
 161         return NULL;
 162 }
 163
 164 static bool extent_has_stripe_ptr(struct bkey_s_c k, u64 idx)
 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 union bch_extent_entry *entry;
 170
 171                 extent_for_each_entry(e, entry)
 172                         if (extent_entry_type(entry) ==
 173                             BCH_EXTENT_ENTRY_stripe_ptr &&
 174                             entry->stripe_ptr.idx == idx)
 175                                 return true;
 176
 177                 break;
 178         }
 179         }
 180
 181         return false;
 182 }
 183
 184 /* Stripe bufs: */
 185
 186 static void ec_stripe_buf_exit(struct ec_stripe_buf *buf)
 187 {
 188         unsigned i;
 189
 190         for (i = 0; i < buf->key.v.nr_blocks; i++) {
 191                 kvpfree(buf->data[i], buf->size << 9);
 192                 buf->data[i] = NULL;
 193         }
 194 }
 195
 196 static int ec_stripe_buf_init(struct ec_stripe_buf *buf,
 197                                unsigned offset, unsigned size)
 198 {
 199         struct bch_stripe *v = &buf->key.v;
 200         unsigned csum_granularity = 1U << v->csum_granularity_bits;
 201         unsigned end = offset + size;
 202         unsigned i;
 203
 204         BUG_ON(end > le16_to_cpu(v->sectors));
 205
 206         offset  = round_down(offset, csum_granularity);
 207         end     = min_t(unsigned, le16_to_cpu(v->sectors),
 208                         round_up(end, csum_granularity));
 209
 210         buf->offset     = offset;
 211         buf->size       = end - offset;
 212
 213         memset(buf->valid, 0xFF, sizeof(buf->valid));
 214
 215         for (i = 0; i < buf->key.v.nr_blocks; i++) {
 216                 buf->data[i] = kvpmalloc(buf->size << 9, GFP_KERNEL);
 217                 if (!buf->data[i])
 218                         goto err;
 219         }
 220
 221         return 0;
 222 err:
 223         ec_stripe_buf_exit(buf);
 224         return -ENOMEM;
 225 }
 226
 227 /* Checksumming: */
 228
 229 static struct bch_csum ec_block_checksum(struct ec_stripe_buf *buf,
 230                                          unsigned block, unsigned offset)
 231 {
 232         struct bch_stripe *v = &buf->key.v;
 233         unsigned csum_granularity = 1 << v->csum_granularity_bits;
 234         unsigned end = buf->offset + buf->size;
 235         unsigned len = min(csum_granularity, end - offset);
 236
 237         BUG_ON(offset >= end);
 238         BUG_ON(offset <  buf->offset);
 239         BUG_ON(offset & (csum_granularity - 1));
 240         BUG_ON(offset + len != le16_to_cpu(v->sectors) &&
 241                (len & (csum_granularity - 1)));
 242
 243         return bch2_checksum(NULL, v->csum_type,
 244                              null_nonce(),
 245                              buf->data[block] + ((offset - buf->offset) << 9),
 246                              len << 9);
 247 }
 248
 249 static void ec_generate_checksums(struct ec_stripe_buf *buf)
 250 {
 251         struct bch_stripe *v = &buf->key.v;
 252         unsigned i, j, csums_per_device = stripe_csums_per_device(v);
 253
 254         if (!v->csum_type)
 255                 return;
 256
 257         BUG_ON(buf->offset);
 258         BUG_ON(buf->size != le16_to_cpu(v->sectors));
 259
 260         for (i = 0; i < v->nr_blocks; i++)
 261                 for (j = 0; j < csums_per_device; j++)
 262                         stripe_csum_set(v, i, j,
 263                                 ec_block_checksum(buf, i, j << v->csum_granularity_bits));
 264 }
 265
 266 static void ec_validate_checksums(struct bch_fs *c, struct ec_stripe_buf *buf)
 267 {
 268         struct bch_stripe *v = &buf->key.v;
 269         unsigned csum_granularity = 1 << v->csum_granularity_bits;
 270         unsigned i;
 271
 272         if (!v->csum_type)
 273                 return;
 274
 275         for (i = 0; i < v->nr_blocks; i++) {
 276                 unsigned offset = buf->offset;
 277                 unsigned end = buf->offset + buf->size;
 278
 279                 if (!test_bit(i, buf->valid))
 280                         continue;
 281
 282                 while (offset < end) {
 283                         unsigned j = offset >> v->csum_granularity_bits;
 284                         unsigned len = min(csum_granularity, end - offset);
 285                         struct bch_csum want = stripe_csum_get(v, i, j);
 286                         struct bch_csum got = ec_block_checksum(buf, i, offset);
 287
 288                         if (bch2_crc_cmp(want, got)) {
 289                                 struct printbuf buf2 = PRINTBUF;
 290
 291                                 bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(&buf->key.k_i));
 292
 293                                 bch_err_ratelimited(c,
 294                                         "stripe checksum error for %ps at %u:%u: csum type %u, expected %llx got %llx\n%s",
 295                                         (void *) _RET_IP_, i, j, v->csum_type,
 296                                         want.lo, got.lo, buf2.buf);
 297                                 printbuf_exit(&buf2);
 298                                 clear_bit(i, buf->valid);
 299                                 break;
 300                         }
 301
 302                         offset += len;
 303                 }
 304         }
 305 }
 306
 307 /* Erasure coding: */
 308
 309 static void ec_generate_ec(struct ec_stripe_buf *buf)
 310 {
 311         struct bch_stripe *v = &buf->key.v;
 312         unsigned nr_data = v->nr_blocks - v->nr_redundant;
 313         unsigned bytes = le16_to_cpu(v->sectors) << 9;
 314
 315         raid_gen(nr_data, v->nr_redundant, bytes, buf->data);
 316 }
 317
 318 static unsigned ec_nr_failed(struct ec_stripe_buf *buf)
 319 {
 320         return buf->key.v.nr_blocks -
 321                 bitmap_weight(buf->valid, buf->key.v.nr_blocks);
 322 }
 323
 324 static int ec_do_recov(struct bch_fs *c, struct ec_stripe_buf *buf)
 325 {
 326         struct bch_stripe *v = &buf->key.v;
 327         unsigned i, failed[BCH_BKEY_PTRS_MAX], nr_failed = 0;
 328         unsigned nr_data = v->nr_blocks - v->nr_redundant;
 329         unsigned bytes = buf->size << 9;
 330
 331         if (ec_nr_failed(buf) > v->nr_redundant) {
 332                 bch_err_ratelimited(c,
 333                         "error doing reconstruct read: unable to read enough blocks");
 334                 return -1;
 335         }
 336
 337         for (i = 0; i < nr_data; i++)
 338                 if (!test_bit(i, buf->valid))
 339                         failed[nr_failed++] = i;
 340
 341         raid_rec(nr_failed, failed, nr_data, v->nr_redundant, bytes, buf->data);
 342         return 0;
 343 }
 344
 345 /* IO: */
 346
 347 static void ec_block_endio(struct bio *bio)
 348 {
 349         struct ec_bio *ec_bio = container_of(bio, struct ec_bio, bio);
 350         struct bch_stripe *v = &ec_bio->buf->key.v;
 351         struct bch_extent_ptr *ptr = &v->ptrs[ec_bio->idx];
 352         struct bch_dev *ca = ec_bio->ca;
 353         struct closure *cl = bio->bi_private;
 354
 355         if (bch2_dev_io_err_on(bio->bi_status, ca, "erasure coding %s error: %s",
 356                                bio_data_dir(bio) ? "write" : "read",
 357                                bch2_blk_status_to_str(bio->bi_status)))
 358                 clear_bit(ec_bio->idx, ec_bio->buf->valid);
 359
 360         if (ptr_stale(ca, ptr)) {
 361                 bch_err_ratelimited(ca->fs,
 362                                     "error %s stripe: stale pointer after io",
 363                                     bio_data_dir(bio) == READ ? "reading from" : "writing to");
 364                 clear_bit(ec_bio->idx, ec_bio->buf->valid);
 365         }
 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         enum bch_data_type data_type = idx < buf->key.v.nr_blocks - buf->key.v.nr_redundant
 380                 ? BCH_DATA_user
 381                 : BCH_DATA_parity;
 382
 383         if (ptr_stale(ca, ptr)) {
 384                 bch_err_ratelimited(c,
 385                                     "error %s stripe: stale pointer",
 386                                     rw == READ ? "reading from" : "writing to");
 387                 clear_bit(idx, buf->valid);
 388                 return;
 389         }
 390
 391         if (!bch2_dev_get_ioref(ca, rw)) {
 392                 clear_bit(idx, buf->valid);
 393                 return;
 394         }
 395
 396         this_cpu_add(ca->io_done->sectors[rw][data_type], buf->size);
 397
 398         while (offset < bytes) {
 399                 unsigned nr_iovecs = min_t(size_t, BIO_MAX_VECS,
 400                                            DIV_ROUND_UP(bytes, PAGE_SIZE));
 401                 unsigned b = min_t(size_t, bytes - offset,
 402                                    nr_iovecs << PAGE_SHIFT);
 403                 struct ec_bio *ec_bio;
 404
 405                 ec_bio = container_of(bio_alloc_bioset(GFP_KERNEL, nr_iovecs,
 406                                                        &c->ec_bioset),
 407                                       struct ec_bio, bio);
 408
 409                 ec_bio->ca                      = ca;
 410                 ec_bio->buf                     = buf;
 411                 ec_bio->idx                     = idx;
 412
 413                 bio_set_dev(&ec_bio->bio, ca->disk_sb.bdev);
 414                 bio_set_op_attrs(&ec_bio->bio, rw, 0);
 415
 416                 ec_bio->bio.bi_iter.bi_sector   = ptr->offset + buf->offset + (offset >> 9);
 417                 ec_bio->bio.bi_end_io           = ec_block_endio;
 418                 ec_bio->bio.bi_private          = cl;
 419
 420                 bch2_bio_map(&ec_bio->bio, buf->data[idx] + offset, b);
 421
 422                 closure_get(cl);
 423                 percpu_ref_get(&ca->io_ref);
 424
 425                 submit_bio(&ec_bio->bio);
 426
 427                 offset += b;
 428         }
 429
 430         percpu_ref_put(&ca->io_ref);
 431 }
 432
 433 static int get_stripe_key(struct bch_fs *c, u64 idx, struct ec_stripe_buf *stripe)
 434 {
 435         struct btree_trans trans;
 436         struct btree_iter iter;
 437         struct bkey_s_c k;
 438         int ret;
 439
 440         bch2_trans_init(&trans, c, 0, 0);
 441         bch2_trans_iter_init(&trans, &iter, BTREE_ID_stripes,
 442                              POS(0, idx), BTREE_ITER_SLOTS);
 443         k = bch2_btree_iter_peek_slot(&iter);
 444         ret = bkey_err(k);
 445         if (ret)
 446                 goto err;
 447         if (k.k->type != KEY_TYPE_stripe) {
 448                 ret = -ENOENT;
 449                 goto err;
 450         }
 451         bkey_reassemble(&stripe->key.k_i, k);
 452 err:
 453         bch2_trans_iter_exit(&trans, &iter);
 454         bch2_trans_exit(&trans);
 455         return ret;
 456 }
 457
 458 /* recovery read path: */
 459 int bch2_ec_read_extent(struct bch_fs *c, struct bch_read_bio *rbio)
 460 {
 461         struct ec_stripe_buf *buf;
 462         struct closure cl;
 463         struct bch_stripe *v;
 464         unsigned i, offset;
 465         int ret = 0;
 466
 467         closure_init_stack(&cl);
 468
 469         BUG_ON(!rbio->pick.has_ec);
 470
 471         buf = kzalloc(sizeof(*buf), GFP_NOIO);
 472         if (!buf)
 473                 return -ENOMEM;
 474
 475         ret = get_stripe_key(c, rbio->pick.ec.idx, buf);
 476         if (ret) {
 477                 bch_err_ratelimited(c,
 478                         "error doing reconstruct read: error %i looking up stripe", ret);
 479                 kfree(buf);
 480                 return -EIO;
 481         }
 482
 483         v = &buf->key.v;
 484
 485         if (!bch2_ptr_matches_stripe(v, rbio->pick)) {
 486                 bch_err_ratelimited(c,
 487                         "error doing reconstruct read: pointer doesn't match stripe");
 488                 ret = -EIO;
 489                 goto err;
 490         }
 491
 492         offset = rbio->bio.bi_iter.bi_sector - v->ptrs[rbio->pick.ec.block].offset;
 493         if (offset + bio_sectors(&rbio->bio) > le16_to_cpu(v->sectors)) {
 494                 bch_err_ratelimited(c,
 495                         "error doing reconstruct read: read is bigger than stripe");
 496                 ret = -EIO;
 497                 goto err;
 498         }
 499
 500         ret = ec_stripe_buf_init(buf, offset, bio_sectors(&rbio->bio));
 501         if (ret)
 502                 goto err;
 503
 504         for (i = 0; i < v->nr_blocks; i++)
 505                 ec_block_io(c, buf, REQ_OP_READ, i, &cl);
 506
 507         closure_sync(&cl);
 508
 509         if (ec_nr_failed(buf) > v->nr_redundant) {
 510                 bch_err_ratelimited(c,
 511                         "error doing reconstruct read: unable to read enough blocks");
 512                 ret = -EIO;
 513                 goto err;
 514         }
 515
 516         ec_validate_checksums(c, buf);
 517
 518         ret = ec_do_recov(c, buf);
 519         if (ret)
 520                 goto err;
 521
 522         memcpy_to_bio(&rbio->bio, rbio->bio.bi_iter,
 523                       buf->data[rbio->pick.ec.block] + ((offset - buf->offset) << 9));
 524 err:
 525         ec_stripe_buf_exit(buf);
 526         kfree(buf);
 527         return ret;
 528 }
 529
 530 /* stripe bucket accounting: */
 531
 532 static int __ec_stripe_mem_alloc(struct bch_fs *c, size_t idx, gfp_t gfp)
 533 {
 534         ec_stripes_heap n, *h = &c->ec_stripes_heap;
 535
 536         if (idx >= h->size) {
 537                 if (!init_heap(&n, max(1024UL, roundup_pow_of_two(idx + 1)), gfp))
 538                         return -ENOMEM;
 539
 540                 spin_lock(&c->ec_stripes_heap_lock);
 541                 if (n.size > h->size) {
 542                         memcpy(n.data, h->data, h->used * sizeof(h->data[0]));
 543                         n.used = h->used;
 544                         swap(*h, n);
 545                 }
 546                 spin_unlock(&c->ec_stripes_heap_lock);
 547
 548                 free_heap(&n);
 549         }
 550
 551         if (!genradix_ptr_alloc(&c->stripes, idx, gfp))
 552                 return -ENOMEM;
 553
 554         if (c->gc_pos.phase != GC_PHASE_NOT_RUNNING &&
 555             !genradix_ptr_alloc(&c->gc_stripes, idx, gfp))
 556                 return -ENOMEM;
 557
 558         return 0;
 559 }
 560
 561 static int ec_stripe_mem_alloc(struct btree_trans *trans,
 562                                struct btree_iter *iter)
 563 {
 564         size_t idx = iter->pos.offset;
 565         int ret = 0;
 566
 567         if (!__ec_stripe_mem_alloc(trans->c, idx, GFP_NOWAIT|__GFP_NOWARN))
 568                 return ret;
 569
 570         bch2_trans_unlock(trans);
 571         ret = -EINTR;
 572
 573         if (!__ec_stripe_mem_alloc(trans->c, idx, GFP_KERNEL))
 574                 return ret;
 575
 576         return -ENOMEM;
 577 }
 578
 579 static ssize_t stripe_idx_to_delete(struct bch_fs *c)
 580 {
 581         ec_stripes_heap *h = &c->ec_stripes_heap;
 582
 583         return h->used && h->data[0].blocks_nonempty == 0
 584                 ? h->data[0].idx : -1;
 585 }
 586
 587 static inline int ec_stripes_heap_cmp(ec_stripes_heap *h,
 588                                       struct ec_stripe_heap_entry l,
 589                                       struct ec_stripe_heap_entry r)
 590 {
 591         return ((l.blocks_nonempty > r.blocks_nonempty) -
 592                 (l.blocks_nonempty < r.blocks_nonempty));
 593 }
 594
 595 static inline void ec_stripes_heap_set_backpointer(ec_stripes_heap *h,
 596                                                    size_t i)
 597 {
 598         struct bch_fs *c = container_of(h, struct bch_fs, ec_stripes_heap);
 599
 600         genradix_ptr(&c->stripes, h->data[i].idx)->heap_idx = i;
 601 }
 602
 603 static void heap_verify_backpointer(struct bch_fs *c, size_t idx)
 604 {
 605         ec_stripes_heap *h = &c->ec_stripes_heap;
 606         struct stripe *m = genradix_ptr(&c->stripes, idx);
 607
 608         BUG_ON(!m->alive);
 609         BUG_ON(m->heap_idx >= h->used);
 610         BUG_ON(h->data[m->heap_idx].idx != idx);
 611 }
 612
 613 void bch2_stripes_heap_del(struct bch_fs *c,
 614                            struct stripe *m, size_t idx)
 615 {
 616         if (!m->on_heap)
 617                 return;
 618
 619         m->on_heap = false;
 620
 621         heap_verify_backpointer(c, idx);
 622
 623         heap_del(&c->ec_stripes_heap, m->heap_idx,
 624                  ec_stripes_heap_cmp,
 625                  ec_stripes_heap_set_backpointer);
 626 }
 627
 628 void bch2_stripes_heap_insert(struct bch_fs *c,
 629                               struct stripe *m, size_t idx)
 630 {
 631         if (m->on_heap)
 632                 return;
 633
 634         BUG_ON(heap_full(&c->ec_stripes_heap));
 635
 636         m->on_heap = true;
 637
 638         heap_add(&c->ec_stripes_heap, ((struct ec_stripe_heap_entry) {
 639                         .idx = idx,
 640                         .blocks_nonempty = m->blocks_nonempty,
 641                 }),
 642                  ec_stripes_heap_cmp,
 643                  ec_stripes_heap_set_backpointer);
 644
 645         heap_verify_backpointer(c, idx);
 646 }
 647
 648 void bch2_stripes_heap_update(struct bch_fs *c,
 649                               struct stripe *m, size_t idx)
 650 {
 651         ec_stripes_heap *h = &c->ec_stripes_heap;
 652         size_t i;
 653
 654         if (!m->on_heap)
 655                 return;
 656
 657         heap_verify_backpointer(c, idx);
 658
 659         h->data[m->heap_idx].blocks_nonempty = m->blocks_nonempty;
 660
 661         i = m->heap_idx;
 662         heap_sift_up(h,   i, ec_stripes_heap_cmp,
 663                      ec_stripes_heap_set_backpointer);
 664         heap_sift_down(h, i, ec_stripes_heap_cmp,
 665                        ec_stripes_heap_set_backpointer);
 666
 667         heap_verify_backpointer(c, idx);
 668
 669         if (stripe_idx_to_delete(c) >= 0 &&
 670             !percpu_ref_is_dying(&c->writes))
 671                 schedule_work(&c->ec_stripe_delete_work);
 672 }
 673
 674 /* stripe deletion */
 675
 676 static int ec_stripe_delete(struct bch_fs *c, size_t idx)
 677 {
 678         return bch2_btree_delete_range(c, BTREE_ID_stripes,
 679                                        POS(0, idx),
 680                                        POS(0, idx + 1),
 681                                        0, NULL);
 682 }
 683
 684 static void ec_stripe_delete_work(struct work_struct *work)
 685 {
 686         struct bch_fs *c =
 687                 container_of(work, struct bch_fs, ec_stripe_delete_work);
 688         ssize_t idx;
 689
 690         while (1) {
 691                 spin_lock(&c->ec_stripes_heap_lock);
 692                 idx = stripe_idx_to_delete(c);
 693                 if (idx < 0) {
 694                         spin_unlock(&c->ec_stripes_heap_lock);
 695                         break;
 696                 }
 697
 698                 bch2_stripes_heap_del(c, genradix_ptr(&c->stripes, idx), idx);
 699                 spin_unlock(&c->ec_stripes_heap_lock);
 700
 701                 if (ec_stripe_delete(c, idx))
 702                         break;
 703         }
 704 }
 705
 706 /* stripe creation: */
 707
 708 static int ec_stripe_bkey_insert(struct btree_trans *trans,
 709                                  struct bkey_i_stripe *stripe,
 710                                  struct disk_reservation *res)
 711 {
 712         struct bch_fs *c = trans->c;
 713         struct btree_iter iter;
 714         struct bkey_s_c k;
 715         struct bpos min_pos = POS(0, 1);
 716         struct bpos start_pos = bpos_max(min_pos, POS(0, c->ec_stripe_hint));
 717         int ret;
 718
 719         for_each_btree_key(trans, iter, BTREE_ID_stripes, start_pos,
 720                            BTREE_ITER_SLOTS|BTREE_ITER_INTENT, k, ret) {
 721                 if (bkey_cmp(k.k->p, POS(0, U32_MAX)) > 0) {
 722                         if (start_pos.offset) {
 723                                 start_pos = min_pos;
 724                                 bch2_btree_iter_set_pos(&iter, start_pos);
 725                                 continue;
 726                         }
 727
 728                         ret = -ENOSPC;
 729                         break;
 730                 }
 731
 732                 if (bkey_deleted(k.k))
 733                         goto found_slot;
 734         }
 735
 736         goto err;
 737 found_slot:
 738         start_pos = iter.pos;
 739
 740         ret = ec_stripe_mem_alloc(trans, &iter);
 741         if (ret)
 742                 goto err;
 743
 744         stripe->k.p = iter.pos;
 745
 746         ret = bch2_trans_update(trans, &iter, &stripe->k_i, 0);
 747
 748         c->ec_stripe_hint = start_pos.offset;
 749 err:
 750         bch2_trans_iter_exit(trans, &iter);
 751
 752         return ret;
 753 }
 754
 755 static int ec_stripe_bkey_update(struct btree_trans *trans,
 756                                  struct bkey_i_stripe *new,
 757                                  struct disk_reservation *res)
 758 {
 759         struct btree_iter iter;
 760         struct bkey_s_c k;
 761         const struct bch_stripe *existing;
 762         unsigned i;
 763         int ret;
 764
 765         bch2_trans_iter_init(trans, &iter, BTREE_ID_stripes,
 766                              new->k.p, BTREE_ITER_INTENT);
 767         k = bch2_btree_iter_peek_slot(&iter);
 768         ret = bkey_err(k);
 769         if (ret)
 770                 goto err;
 771
 772         if (!k.k || k.k->type != KEY_TYPE_stripe) {
 773                 bch_err(trans->c, "error updating stripe: not found");
 774                 ret = -ENOENT;
 775                 goto err;
 776         }
 777
 778         existing = bkey_s_c_to_stripe(k).v;
 779
 780         if (existing->nr_blocks != new->v.nr_blocks) {
 781                 bch_err(trans->c, "error updating stripe: nr_blocks does not match");
 782                 ret = -EINVAL;
 783                 goto err;
 784         }
 785
 786         for (i = 0; i < new->v.nr_blocks; i++)
 787                 stripe_blockcount_set(&new->v, i,
 788                         stripe_blockcount_get(existing, i));
 789
 790         ret = bch2_trans_update(trans, &iter, &new->k_i, 0);
 791 err:
 792         bch2_trans_iter_exit(trans, &iter);
 793         return ret;
 794 }
 795
 796 static void extent_stripe_ptr_add(struct bkey_s_extent e,
 797                                   struct ec_stripe_buf *s,
 798                                   struct bch_extent_ptr *ptr,
 799                                   unsigned block)
 800 {
 801         struct bch_extent_stripe_ptr *dst = (void *) ptr;
 802         union bch_extent_entry *end = extent_entry_last(e);
 803
 804         memmove_u64s_up(dst + 1, dst, (u64 *) end - (u64 *) dst);
 805         e.k->u64s += sizeof(*dst) / sizeof(u64);
 806
 807         *dst = (struct bch_extent_stripe_ptr) {
 808                 .type = 1 << BCH_EXTENT_ENTRY_stripe_ptr,
 809                 .block          = block,
 810                 .redundancy     = s->key.v.nr_redundant,
 811                 .idx            = s->key.k.p.offset,
 812         };
 813 }
 814
 815 static int ec_stripe_update_ptrs(struct bch_fs *c,
 816                                  struct ec_stripe_buf *s,
 817                                  struct bkey *pos)
 818 {
 819         struct btree_trans trans;
 820         struct btree_iter iter;
 821         struct bkey_s_c k;
 822         struct bkey_s_extent e;
 823         struct bkey_buf sk;
 824         struct bpos next_pos;
 825         int ret = 0, dev, block;
 826
 827         bch2_bkey_buf_init(&sk);
 828         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
 829
 830         /* XXX this doesn't support the reflink btree */
 831
 832         bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
 833                              bkey_start_pos(pos),
 834                              BTREE_ITER_INTENT);
 835 retry:
 836         while (bch2_trans_begin(&trans),
 837                (k = bch2_btree_iter_peek(&iter)).k &&
 838                !(ret = bkey_err(k)) &&
 839                bkey_cmp(bkey_start_pos(k.k), pos->p) < 0) {
 840                 const struct bch_extent_ptr *ptr_c;
 841                 struct bch_extent_ptr *ptr, *ec_ptr = NULL;
 842
 843                 if (extent_has_stripe_ptr(k, s->key.k.p.offset)) {
 844                         bch2_btree_iter_advance(&iter);
 845                         continue;
 846                 }
 847
 848                 ptr_c = bkey_matches_stripe(&s->key.v, k, &block);
 849                 /*
 850                  * It doesn't generally make sense to erasure code cached ptrs:
 851                  * XXX: should we be incrementing a counter?
 852                  */
 853                 if (!ptr_c || ptr_c->cached) {
 854                         bch2_btree_iter_advance(&iter);
 855                         continue;
 856                 }
 857
 858                 dev = s->key.v.ptrs[block].dev;
 859
 860                 bch2_bkey_buf_reassemble(&sk, c, k);
 861                 e = bkey_i_to_s_extent(sk.k);
 862
 863                 bch2_bkey_drop_ptrs(e.s, ptr, ptr->dev != dev);
 864                 ec_ptr = (void *) bch2_bkey_has_device(e.s_c, dev);
 865                 BUG_ON(!ec_ptr);
 866
 867                 extent_stripe_ptr_add(e, s, ec_ptr, block);
 868
 869                 bch2_btree_iter_set_pos(&iter, bkey_start_pos(&sk.k->k));
 870                 next_pos = sk.k->k.p;
 871
 872                 ret   = bch2_btree_iter_traverse(&iter) ?:
 873                         bch2_trans_update(&trans, &iter, sk.k, 0) ?:
 874                         bch2_trans_commit(&trans, NULL, NULL,
 875                                         BTREE_INSERT_NOFAIL);
 876                 if (!ret)
 877                         bch2_btree_iter_set_pos(&iter, next_pos);
 878                 if (ret)
 879                         break;
 880         }
 881         if (ret == -EINTR)
 882                 goto retry;
 883         bch2_trans_iter_exit(&trans, &iter);
 884
 885         bch2_trans_exit(&trans);
 886         bch2_bkey_buf_exit(&sk, c);
 887
 888         return ret;
 889 }
 890
 891 /*
 892  * data buckets of new stripe all written: create the stripe
 893  */
 894 static void ec_stripe_create(struct ec_stripe_new *s)
 895 {
 896         struct bch_fs *c = s->c;
 897         struct open_bucket *ob;
 898         struct bkey_i *k;
 899         struct stripe *m;
 900         struct bch_stripe *v = &s->new_stripe.key.v;
 901         unsigned i, nr_data = v->nr_blocks - v->nr_redundant;
 902         int ret;
 903
 904         BUG_ON(s->h->s == s);
 905
 906         closure_sync(&s->iodone);
 907
 908         if (s->err) {
 909                 if (s->err != -EROFS)
 910                         bch_err(c, "error creating stripe: error writing data buckets");
 911                 goto err;
 912         }
 913
 914         if (s->have_existing_stripe) {
 915                 ec_validate_checksums(c, &s->existing_stripe);
 916
 917                 if (ec_do_recov(c, &s->existing_stripe)) {
 918                         bch_err(c, "error creating stripe: error reading existing stripe");
 919                         goto err;
 920                 }
 921
 922                 for (i = 0; i < nr_data; i++)
 923                         if (stripe_blockcount_get(&s->existing_stripe.key.v, i))
 924                                 swap(s->new_stripe.data[i],
 925                                      s->existing_stripe.data[i]);
 926
 927                 ec_stripe_buf_exit(&s->existing_stripe);
 928         }
 929
 930         BUG_ON(!s->allocated);
 931
 932         if (!percpu_ref_tryget(&c->writes))
 933                 goto err;
 934
 935         ec_generate_ec(&s->new_stripe);
 936
 937         ec_generate_checksums(&s->new_stripe);
 938
 939         /* write p/q: */
 940         for (i = nr_data; i < v->nr_blocks; i++)
 941                 ec_block_io(c, &s->new_stripe, REQ_OP_WRITE, i, &s->iodone);
 942         closure_sync(&s->iodone);
 943
 944         if (ec_nr_failed(&s->new_stripe)) {
 945                 bch_err(c, "error creating stripe: error writing redundancy buckets");
 946                 goto err_put_writes;
 947         }
 948
 949         ret = bch2_trans_do(c, &s->res, NULL, BTREE_INSERT_NOFAIL,
 950                             s->have_existing_stripe
 951                             ? ec_stripe_bkey_update(&trans, &s->new_stripe.key, &s->res)
 952                             : ec_stripe_bkey_insert(&trans, &s->new_stripe.key, &s->res));
 953         if (ret) {
 954                 bch_err(c, "error creating stripe: error creating stripe key");
 955                 goto err_put_writes;
 956         }
 957
 958         for_each_keylist_key(&s->keys, k) {
 959                 ret = ec_stripe_update_ptrs(c, &s->new_stripe, &k->k);
 960                 if (ret) {
 961                         bch_err(c, "error creating stripe: error %i updating pointers", ret);
 962                         break;
 963                 }
 964         }
 965
 966         spin_lock(&c->ec_stripes_heap_lock);
 967         m = genradix_ptr(&c->stripes, s->new_stripe.key.k.p.offset);
 968
 969         BUG_ON(m->on_heap);
 970         bch2_stripes_heap_insert(c, m, s->new_stripe.key.k.p.offset);
 971         spin_unlock(&c->ec_stripes_heap_lock);
 972 err_put_writes:
 973         percpu_ref_put(&c->writes);
 974 err:
 975         bch2_disk_reservation_put(c, &s->res);
 976
 977         for (i = 0; i < v->nr_blocks; i++)
 978                 if (s->blocks[i]) {
 979                         ob = c->open_buckets + s->blocks[i];
 980
 981                         if (i < nr_data) {
 982                                 ob->ec = NULL;
 983                                 __bch2_open_bucket_put(c, ob);
 984                         } else {
 985                                 bch2_open_bucket_put(c, ob);
 986                         }
 987                 }
 988
 989         bch2_keylist_free(&s->keys, s->inline_keys);
 990
 991         ec_stripe_buf_exit(&s->existing_stripe);
 992         ec_stripe_buf_exit(&s->new_stripe);
 993         closure_debug_destroy(&s->iodone);
 994         kfree(s);
 995 }
 996
 997 static void ec_stripe_create_work(struct work_struct *work)
 998 {
 999         struct bch_fs *c = container_of(work,
1000                 struct bch_fs, ec_stripe_create_work);
1001         struct ec_stripe_new *s, *n;
1002 restart:
1003         mutex_lock(&c->ec_stripe_new_lock);
1004         list_for_each_entry_safe(s, n, &c->ec_stripe_new_list, list)
1005                 if (!atomic_read(&s->pin)) {
1006                         list_del(&s->list);
1007                         mutex_unlock(&c->ec_stripe_new_lock);
1008                         ec_stripe_create(s);
1009                         goto restart;
1010                 }
1011         mutex_unlock(&c->ec_stripe_new_lock);
1012 }
1013
1014 static void ec_stripe_new_put(struct bch_fs *c, struct ec_stripe_new *s)
1015 {
1016         BUG_ON(atomic_read(&s->pin) <= 0);
1017
1018         if (atomic_dec_and_test(&s->pin)) {
1019                 BUG_ON(!s->pending);
1020                 queue_work(system_long_wq, &c->ec_stripe_create_work);
1021         }
1022 }
1023
1024 static void ec_stripe_set_pending(struct bch_fs *c, struct ec_stripe_head *h)
1025 {
1026         struct ec_stripe_new *s = h->s;
1027
1028         BUG_ON(!s->allocated && !s->err);
1029
1030         h->s            = NULL;
1031         s->pending      = true;
1032
1033         mutex_lock(&c->ec_stripe_new_lock);
1034         list_add(&s->list, &c->ec_stripe_new_list);
1035         mutex_unlock(&c->ec_stripe_new_lock);
1036
1037         ec_stripe_new_put(c, s);
1038 }
1039
1040 /* have a full bucket - hand it off to be erasure coded: */
1041 void bch2_ec_bucket_written(struct bch_fs *c, struct open_bucket *ob)
1042 {
1043         struct ec_stripe_new *s = ob->ec;
1044
1045         if (ob->sectors_free)
1046                 s->err = -1;
1047
1048         ec_stripe_new_put(c, s);
1049 }
1050
1051 void bch2_ec_bucket_cancel(struct bch_fs *c, struct open_bucket *ob)
1052 {
1053         struct ec_stripe_new *s = ob->ec;
1054
1055         s->err = -EIO;
1056 }
1057
1058 void *bch2_writepoint_ec_buf(struct bch_fs *c, struct write_point *wp)
1059 {
1060         struct open_bucket *ob = ec_open_bucket(c, &wp->ptrs);
1061         struct bch_dev *ca;
1062         unsigned offset;
1063
1064         if (!ob)
1065                 return NULL;
1066
1067         ca      = bch_dev_bkey_exists(c, ob->dev);
1068         offset  = ca->mi.bucket_size - ob->sectors_free;
1069
1070         return ob->ec->new_stripe.data[ob->ec_idx] + (offset << 9);
1071 }
1072
1073 void bch2_ob_add_backpointer(struct bch_fs *c, struct open_bucket *ob,
1074                              struct bkey *k)
1075 {
1076         struct ec_stripe_new *ec = ob->ec;
1077
1078         if (!ec)
1079                 return;
1080
1081         mutex_lock(&ec->lock);
1082
1083         if (bch2_keylist_realloc(&ec->keys, ec->inline_keys,
1084                                  ARRAY_SIZE(ec->inline_keys),
1085                                  BKEY_U64s)) {
1086                 BUG();
1087         }
1088
1089         bkey_init(&ec->keys.top->k);
1090         ec->keys.top->k.p       = k->p;
1091         ec->keys.top->k.size    = k->size;
1092         bch2_keylist_push(&ec->keys);
1093
1094         mutex_unlock(&ec->lock);
1095 }
1096
1097 static int unsigned_cmp(const void *_l, const void *_r)
1098 {
1099         unsigned l = *((const unsigned *) _l);
1100         unsigned r = *((const unsigned *) _r);
1101
1102         return cmp_int(l, r);
1103 }
1104
1105 /* pick most common bucket size: */
1106 static unsigned pick_blocksize(struct bch_fs *c,
1107                                struct bch_devs_mask *devs)
1108 {
1109         struct bch_dev *ca;
1110         unsigned i, nr = 0, sizes[BCH_SB_MEMBERS_MAX];
1111         struct {
1112                 unsigned nr, size;
1113         } cur = { 0, 0 }, best = { 0, 0 };
1114
1115         for_each_member_device_rcu(ca, c, i, devs)
1116                 sizes[nr++] = ca->mi.bucket_size;
1117
1118         sort(sizes, nr, sizeof(unsigned), unsigned_cmp, NULL);
1119
1120         for (i = 0; i < nr; i++) {
1121                 if (sizes[i] != cur.size) {
1122                         if (cur.nr > best.nr)
1123                                 best = cur;
1124
1125                         cur.nr = 0;
1126                         cur.size = sizes[i];
1127                 }
1128
1129                 cur.nr++;
1130         }
1131
1132         if (cur.nr > best.nr)
1133                 best = cur;
1134
1135         return best.size;
1136 }
1137
1138 static bool may_create_new_stripe(struct bch_fs *c)
1139 {
1140         return false;
1141 }
1142
1143 static void ec_stripe_key_init(struct bch_fs *c,
1144                                struct bkey_i_stripe *s,
1145                                unsigned nr_data,
1146                                unsigned nr_parity,
1147                                unsigned stripe_size)
1148 {
1149         unsigned u64s;
1150
1151         bkey_stripe_init(&s->k_i);
1152         s->v.sectors                    = cpu_to_le16(stripe_size);
1153         s->v.algorithm                  = 0;
1154         s->v.nr_blocks                  = nr_data + nr_parity;
1155         s->v.nr_redundant               = nr_parity;
1156         s->v.csum_granularity_bits      = ilog2(c->opts.encoded_extent_max >> 9);
1157         s->v.csum_type                  = BCH_CSUM_crc32c;
1158         s->v.pad                        = 0;
1159
1160         while ((u64s = stripe_val_u64s(&s->v)) > BKEY_VAL_U64s_MAX) {
1161                 BUG_ON(1 << s->v.csum_granularity_bits >=
1162                        le16_to_cpu(s->v.sectors) ||
1163                        s->v.csum_granularity_bits == U8_MAX);
1164                 s->v.csum_granularity_bits++;
1165         }
1166
1167         set_bkey_val_u64s(&s->k, u64s);
1168 }
1169
1170 static int ec_new_stripe_alloc(struct bch_fs *c, struct ec_stripe_head *h)
1171 {
1172         struct ec_stripe_new *s;
1173
1174         lockdep_assert_held(&h->lock);
1175
1176         s = kzalloc(sizeof(*s), GFP_KERNEL);
1177         if (!s)
1178                 return -ENOMEM;
1179
1180         mutex_init(&s->lock);
1181         closure_init(&s->iodone, NULL);
1182         atomic_set(&s->pin, 1);
1183         s->c            = c;
1184         s->h            = h;
1185         s->nr_data      = min_t(unsigned, h->nr_active_devs,
1186                                 BCH_BKEY_PTRS_MAX) - h->redundancy;
1187         s->nr_parity    = h->redundancy;
1188
1189         bch2_keylist_init(&s->keys, s->inline_keys);
1190
1191         ec_stripe_key_init(c, &s->new_stripe.key, s->nr_data,
1192                            s->nr_parity, h->blocksize);
1193
1194         h->s = s;
1195         return 0;
1196 }
1197
1198 static struct ec_stripe_head *
1199 ec_new_stripe_head_alloc(struct bch_fs *c, unsigned target,
1200                          unsigned algo, unsigned redundancy,
1201                          bool copygc)
1202 {
1203         struct ec_stripe_head *h;
1204         struct bch_dev *ca;
1205         unsigned i;
1206
1207         h = kzalloc(sizeof(*h), GFP_KERNEL);
1208         if (!h)
1209                 return NULL;
1210
1211         mutex_init(&h->lock);
1212         mutex_lock(&h->lock);
1213
1214         h->target       = target;
1215         h->algo         = algo;
1216         h->redundancy   = redundancy;
1217         h->copygc       = copygc;
1218
1219         rcu_read_lock();
1220         h->devs = target_rw_devs(c, BCH_DATA_user, target);
1221
1222         for_each_member_device_rcu(ca, c, i, &h->devs)
1223                 if (!ca->mi.durability)
1224                         __clear_bit(i, h->devs.d);
1225
1226         h->blocksize = pick_blocksize(c, &h->devs);
1227
1228         for_each_member_device_rcu(ca, c, i, &h->devs)
1229                 if (ca->mi.bucket_size == h->blocksize)
1230                         h->nr_active_devs++;
1231
1232         rcu_read_unlock();
1233         list_add(&h->list, &c->ec_stripe_head_list);
1234         return h;
1235 }
1236
1237 void bch2_ec_stripe_head_put(struct bch_fs *c, struct ec_stripe_head *h)
1238 {
1239         if (h->s &&
1240             h->s->allocated &&
1241             bitmap_weight(h->s->blocks_allocated,
1242                           h->s->nr_data) == h->s->nr_data)
1243                 ec_stripe_set_pending(c, h);
1244
1245         mutex_unlock(&h->lock);
1246 }
1247
1248 struct ec_stripe_head *__bch2_ec_stripe_head_get(struct bch_fs *c,
1249                                                  unsigned target,
1250                                                  unsigned algo,
1251                                                  unsigned redundancy,
1252                                                  bool copygc)
1253 {
1254         struct ec_stripe_head *h;
1255
1256         if (!redundancy)
1257                 return NULL;
1258
1259         mutex_lock(&c->ec_stripe_head_lock);
1260         list_for_each_entry(h, &c->ec_stripe_head_list, list)
1261                 if (h->target           == target &&
1262                     h->algo             == algo &&
1263                     h->redundancy       == redundancy &&
1264                     h->copygc           == copygc) {
1265                         mutex_lock(&h->lock);
1266                         goto found;
1267                 }
1268
1269         h = ec_new_stripe_head_alloc(c, target, algo, redundancy, copygc);
1270 found:
1271         mutex_unlock(&c->ec_stripe_head_lock);
1272         return h;
1273 }
1274
1275 static int new_stripe_alloc_buckets(struct bch_fs *c, struct ec_stripe_head *h,
1276                                     struct closure *cl)
1277 {
1278         struct bch_devs_mask devs = h->devs;
1279         struct open_bucket *ob;
1280         struct open_buckets buckets;
1281         unsigned i, j, nr_have_parity = 0, nr_have_data = 0;
1282         bool have_cache = true;
1283         int ret = 0;
1284
1285         for (i = 0; i < h->s->new_stripe.key.v.nr_blocks; i++) {
1286                 if (test_bit(i, h->s->blocks_gotten)) {
1287                         __clear_bit(h->s->new_stripe.key.v.ptrs[i].dev, devs.d);
1288                         if (i < h->s->nr_data)
1289                                 nr_have_data++;
1290                         else
1291                                 nr_have_parity++;
1292                 }
1293         }
1294
1295         BUG_ON(nr_have_data     > h->s->nr_data);
1296         BUG_ON(nr_have_parity   > h->s->nr_parity);
1297
1298         buckets.nr = 0;
1299         if (nr_have_parity < h->s->nr_parity) {
1300                 ret = bch2_bucket_alloc_set(c, &buckets,
1301                                             &h->parity_stripe,
1302                                             &devs,
1303                                             h->s->nr_parity,
1304                                             &nr_have_parity,
1305                                             &have_cache,
1306                                             h->copygc
1307                                             ? RESERVE_movinggc
1308                                             : RESERVE_none,
1309                                             0,
1310                                             cl);
1311
1312                 open_bucket_for_each(c, &buckets, ob, i) {
1313                         j = find_next_zero_bit(h->s->blocks_gotten,
1314                                                h->s->nr_data + h->s->nr_parity,
1315                                                h->s->nr_data);
1316                         BUG_ON(j >= h->s->nr_data + h->s->nr_parity);
1317
1318                         h->s->blocks[j] = buckets.v[i];
1319                         h->s->new_stripe.key.v.ptrs[j] = bch2_ob_ptr(c, ob);
1320                         __set_bit(j, h->s->blocks_gotten);
1321                 }
1322
1323                 if (ret)
1324                         return ret;
1325         }
1326
1327         buckets.nr = 0;
1328         if (nr_have_data < h->s->nr_data) {
1329                 ret = bch2_bucket_alloc_set(c, &buckets,
1330                                             &h->block_stripe,
1331                                             &devs,
1332                                             h->s->nr_data,
1333                                             &nr_have_data,
1334                                             &have_cache,
1335                                             h->copygc
1336                                             ? RESERVE_movinggc
1337                                             : RESERVE_none,
1338                                             0,
1339                                             cl);
1340
1341                 open_bucket_for_each(c, &buckets, ob, i) {
1342                         j = find_next_zero_bit(h->s->blocks_gotten,
1343                                                h->s->nr_data, 0);
1344                         BUG_ON(j >= h->s->nr_data);
1345
1346                         h->s->blocks[j] = buckets.v[i];
1347                         h->s->new_stripe.key.v.ptrs[j] = bch2_ob_ptr(c, ob);
1348                         __set_bit(j, h->s->blocks_gotten);
1349                 }
1350
1351                 if (ret)
1352                         return ret;
1353         }
1354
1355         return 0;
1356 }
1357
1358 /* XXX: doesn't obey target: */
1359 static s64 get_existing_stripe(struct bch_fs *c,
1360                                struct ec_stripe_head *head)
1361 {
1362         ec_stripes_heap *h = &c->ec_stripes_heap;
1363         struct stripe *m;
1364         size_t heap_idx;
1365         u64 stripe_idx;
1366         s64 ret = -1;
1367
1368         if (may_create_new_stripe(c))
1369                 return -1;
1370
1371         spin_lock(&c->ec_stripes_heap_lock);
1372         for (heap_idx = 0; heap_idx < h->used; heap_idx++) {
1373                 /* No blocks worth reusing, stripe will just be deleted: */
1374                 if (!h->data[heap_idx].blocks_nonempty)
1375                         continue;
1376
1377                 stripe_idx = h->data[heap_idx].idx;
1378                 m = genradix_ptr(&c->stripes, stripe_idx);
1379
1380                 if (m->algorithm        == head->algo &&
1381                     m->nr_redundant     == head->redundancy &&
1382                     m->sectors          == head->blocksize &&
1383                     m->blocks_nonempty  < m->nr_blocks - m->nr_redundant) {
1384                         bch2_stripes_heap_del(c, m, stripe_idx);
1385                         ret = stripe_idx;
1386                         break;
1387                 }
1388         }
1389         spin_unlock(&c->ec_stripes_heap_lock);
1390         return ret;
1391 }
1392
1393 static int __bch2_ec_stripe_head_reuse(struct bch_fs *c,
1394                                                    struct ec_stripe_head *h)
1395 {
1396         unsigned i;
1397         s64 idx;
1398         int ret;
1399
1400         idx = get_existing_stripe(c, h);
1401         if (idx < 0) {
1402                 bch_err(c, "failed to find an existing stripe");
1403                 return -ENOSPC;
1404         }
1405
1406         h->s->have_existing_stripe = true;
1407         ret = get_stripe_key(c, idx, &h->s->existing_stripe);
1408         if (ret) {
1409                 bch2_fs_fatal_error(c, "error reading stripe key: %i", ret);
1410                 return ret;
1411         }
1412
1413         if (ec_stripe_buf_init(&h->s->existing_stripe, 0, h->blocksize)) {
1414                 /*
1415                  * this is a problem: we have deleted from the
1416                  * stripes heap already
1417                  */
1418                 BUG();
1419         }
1420
1421         BUG_ON(h->s->existing_stripe.size != h->blocksize);
1422         BUG_ON(h->s->existing_stripe.size != h->s->existing_stripe.key.v.sectors);
1423
1424         for (i = 0; i < h->s->existing_stripe.key.v.nr_blocks; i++) {
1425                 if (stripe_blockcount_get(&h->s->existing_stripe.key.v, i)) {
1426                         __set_bit(i, h->s->blocks_gotten);
1427                         __set_bit(i, h->s->blocks_allocated);
1428                 }
1429
1430                 ec_block_io(c, &h->s->existing_stripe, READ, i, &h->s->iodone);
1431         }
1432
1433         bkey_copy(&h->s->new_stripe.key.k_i,
1434                         &h->s->existing_stripe.key.k_i);
1435
1436         return 0;
1437 }
1438
1439 static int __bch2_ec_stripe_head_reserve(struct bch_fs *c,
1440                                                         struct ec_stripe_head *h)
1441 {
1442         int ret;
1443
1444         ret = bch2_disk_reservation_get(c, &h->s->res,
1445                         h->blocksize,
1446                         h->s->nr_parity, 0);
1447
1448         if (ret) {
1449                 /*
1450                  * This means we need to wait for copygc to
1451                  * empty out buckets from existing stripes:
1452                  */
1453                 bch_err(c, "failed to reserve stripe");
1454         }
1455
1456         return ret;
1457 }
1458
1459 struct ec_stripe_head *bch2_ec_stripe_head_get(struct bch_fs *c,
1460                                                unsigned target,
1461                                                unsigned algo,
1462                                                unsigned redundancy,
1463                                                bool copygc,
1464                                                struct closure *cl)
1465 {
1466         struct ec_stripe_head *h;
1467         int ret;
1468         bool needs_stripe_new;
1469
1470         h = __bch2_ec_stripe_head_get(c, target, algo, redundancy, copygc);
1471         if (!h) {
1472                 bch_err(c, "no stripe head");
1473                 return NULL;
1474         }
1475
1476         needs_stripe_new = !h->s;
1477         if (needs_stripe_new) {
1478                 if (ec_new_stripe_alloc(c, h)) {
1479                         ret = -ENOMEM;
1480                         bch_err(c, "failed to allocate new stripe");
1481                         goto err;
1482                 }
1483
1484                 if (ec_stripe_buf_init(&h->s->new_stripe, 0, h->blocksize))
1485                         BUG();
1486         }
1487
1488         /*
1489          * Try reserve a new stripe before reusing an
1490          * existing stripe. This will prevent unnecessary
1491          * read amplification during write oriented workloads.
1492          */
1493         ret = 0;
1494         if (!h->s->allocated && !h->s->res.sectors && !h->s->have_existing_stripe)
1495                 ret = __bch2_ec_stripe_head_reserve(c, h);
1496         if (ret && needs_stripe_new)
1497                 ret = __bch2_ec_stripe_head_reuse(c, h);
1498         if (ret)
1499                 goto err;
1500
1501         if (!h->s->allocated) {
1502                 ret = new_stripe_alloc_buckets(c, h, cl);
1503                 if (ret)
1504                         goto err;
1505
1506                 h->s->allocated = true;
1507         }
1508
1509         return h;
1510
1511 err:
1512         bch2_ec_stripe_head_put(c, h);
1513         return ERR_PTR(ret);
1514 }
1515
1516 void bch2_ec_stop_dev(struct bch_fs *c, struct bch_dev *ca)
1517 {
1518         struct ec_stripe_head *h;
1519         struct open_bucket *ob;
1520         unsigned i;
1521
1522         mutex_lock(&c->ec_stripe_head_lock);
1523         list_for_each_entry(h, &c->ec_stripe_head_list, list) {
1524
1525                 mutex_lock(&h->lock);
1526                 if (!h->s)
1527                         goto unlock;
1528
1529                 for (i = 0; i < h->s->new_stripe.key.v.nr_blocks; i++) {
1530                         if (!h->s->blocks[i])
1531                                 continue;
1532
1533                         ob = c->open_buckets + h->s->blocks[i];
1534                         if (ob->dev == ca->dev_idx)
1535                                 goto found;
1536                 }
1537                 goto unlock;
1538 found:
1539                 h->s->err = -EROFS;
1540                 ec_stripe_set_pending(c, h);
1541 unlock:
1542                 mutex_unlock(&h->lock);
1543         }
1544         mutex_unlock(&c->ec_stripe_head_lock);
1545 }
1546
1547 void bch2_stripes_heap_start(struct bch_fs *c)
1548 {
1549         struct genradix_iter iter;
1550         struct stripe *m;
1551
1552         genradix_for_each(&c->stripes, iter, m)
1553                 if (m->alive)
1554                         bch2_stripes_heap_insert(c, m, iter.pos);
1555 }
1556
1557 int bch2_stripes_read(struct bch_fs *c)
1558 {
1559         struct btree_trans trans;
1560         struct btree_iter iter;
1561         struct bkey_s_c k;
1562         const struct bch_stripe *s;
1563         struct stripe *m;
1564         unsigned i;
1565         int ret;
1566
1567         bch2_trans_init(&trans, c, 0, 0);
1568
1569         for_each_btree_key(&trans, iter, BTREE_ID_stripes, POS_MIN,
1570                            BTREE_ITER_PREFETCH, k, ret) {
1571                 if (k.k->type != KEY_TYPE_stripe)
1572                         continue;
1573
1574                 ret = __ec_stripe_mem_alloc(c, k.k->p.offset, GFP_KERNEL);
1575                 if (ret)
1576                         break;
1577
1578                 s = bkey_s_c_to_stripe(k).v;
1579
1580                 m = genradix_ptr(&c->stripes, k.k->p.offset);
1581                 m->alive        = true;
1582                 m->sectors      = le16_to_cpu(s->sectors);
1583                 m->algorithm    = s->algorithm;
1584                 m->nr_blocks    = s->nr_blocks;
1585                 m->nr_redundant = s->nr_redundant;
1586                 m->blocks_nonempty = 0;
1587
1588                 for (i = 0; i < s->nr_blocks; i++)
1589                         m->blocks_nonempty += !!stripe_blockcount_get(s, i);
1590
1591                 spin_lock(&c->ec_stripes_heap_lock);
1592                 bch2_stripes_heap_update(c, m, k.k->p.offset);
1593                 spin_unlock(&c->ec_stripes_heap_lock);
1594         }
1595         bch2_trans_iter_exit(&trans, &iter);
1596
1597         bch2_trans_exit(&trans);
1598
1599         if (ret)
1600                 bch_err(c, "error reading stripes: %i", ret);
1601
1602         return ret;
1603 }
1604
1605 void bch2_stripes_heap_to_text(struct printbuf *out, struct bch_fs *c)
1606 {
1607         ec_stripes_heap *h = &c->ec_stripes_heap;
1608         struct stripe *m;
1609         size_t i;
1610
1611         spin_lock(&c->ec_stripes_heap_lock);
1612         for (i = 0; i < min_t(size_t, h->used, 20); i++) {
1613                 m = genradix_ptr(&c->stripes, h->data[i].idx);
1614
1615                 pr_buf(out, "%zu %u/%u+%u\n", h->data[i].idx,
1616                        h->data[i].blocks_nonempty,
1617                        m->nr_blocks - m->nr_redundant,
1618                        m->nr_redundant);
1619         }
1620         spin_unlock(&c->ec_stripes_heap_lock);
1621 }
1622
1623 void bch2_new_stripes_to_text(struct printbuf *out, struct bch_fs *c)
1624 {
1625         struct ec_stripe_head *h;
1626         struct ec_stripe_new *s;
1627
1628         mutex_lock(&c->ec_stripe_head_lock);
1629         list_for_each_entry(h, &c->ec_stripe_head_list, list) {
1630                 pr_buf(out, "target %u algo %u redundancy %u:\n",
1631                        h->target, h->algo, h->redundancy);
1632
1633                 if (h->s)
1634                         pr_buf(out, "\tpending: blocks %u+%u allocated %u\n",
1635                                h->s->nr_data, h->s->nr_parity,
1636                                bitmap_weight(h->s->blocks_allocated,
1637                                              h->s->nr_data));
1638         }
1639         mutex_unlock(&c->ec_stripe_head_lock);
1640
1641         mutex_lock(&c->ec_stripe_new_lock);
1642         list_for_each_entry(s, &c->ec_stripe_new_list, list) {
1643                 pr_buf(out, "\tin flight: blocks %u+%u pin %u\n",
1644                        s->nr_data, s->nr_parity,
1645                        atomic_read(&s->pin));
1646         }
1647         mutex_unlock(&c->ec_stripe_new_lock);
1648 }
1649
1650 void bch2_fs_ec_exit(struct bch_fs *c)
1651 {
1652         struct ec_stripe_head *h;
1653
1654         while (1) {
1655                 mutex_lock(&c->ec_stripe_head_lock);
1656                 h = list_first_entry_or_null(&c->ec_stripe_head_list,
1657                                              struct ec_stripe_head, list);
1658                 if (h)
1659                         list_del(&h->list);
1660                 mutex_unlock(&c->ec_stripe_head_lock);
1661                 if (!h)
1662                         break;
1663
1664                 BUG_ON(h->s);
1665                 kfree(h);
1666         }
1667
1668         BUG_ON(!list_empty(&c->ec_stripe_new_list));
1669
1670         free_heap(&c->ec_stripes_heap);
1671         genradix_free(&c->stripes);
1672         bioset_exit(&c->ec_bioset);
1673 }
1674
1675 int bch2_fs_ec_init(struct bch_fs *c)
1676 {
1677         INIT_WORK(&c->ec_stripe_create_work, ec_stripe_create_work);
1678         INIT_WORK(&c->ec_stripe_delete_work, ec_stripe_delete_work);
1679
1680         return bioset_init(&c->ec_bioset, 1, offsetof(struct ec_bio, bio),
1681                            BIOSET_NEED_BVECS);
1682 }