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 int bch2_stripe_invalid(const struct bch_fs *c, struct bkey_s_c k,
 106                         int rw, struct printbuf *err)
 107 {
 108         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 109
 110         if (!bkey_cmp(k.k->p, POS_MIN)) {
 111                 pr_buf(err, "stripe at POS_MIN");
 112                 return -EINVAL;
 113         }
 114
 115         if (k.k->p.inode) {
 116                 pr_buf(err, "nonzero inode field");
 117                 return -EINVAL;
 118         }
 119
 120         if (bkey_val_bytes(k.k) < sizeof(*s)) {
 121                 pr_buf(err, "incorrect value size (%zu < %zu)",
 122                        bkey_val_bytes(k.k), sizeof(*s));
 123                 return -EINVAL;
 124         }
 125
 126         if (bkey_val_u64s(k.k) < stripe_val_u64s(s)) {
 127                 pr_buf(err, "incorrect value size (%zu < %u)",
 128                        bkey_val_u64s(k.k), stripe_val_u64s(s));
 129                 return -EINVAL;
 130         }
 131
 132         return bch2_bkey_ptrs_invalid(c, k, rw, err);
 133 }
 134
 135 void bch2_stripe_to_text(struct printbuf *out, struct bch_fs *c,
 136                          struct bkey_s_c k)
 137 {
 138         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 139         unsigned i;
 140
 141         pr_buf(out, "algo %u sectors %u blocks %u:%u csum %u gran %u",
 142                s->algorithm,
 143                le16_to_cpu(s->sectors),
 144                s->nr_blocks - s->nr_redundant,
 145                s->nr_redundant,
 146                s->csum_type,
 147                1U << s->csum_granularity_bits);
 148
 149         for (i = 0; i < s->nr_blocks; i++)
 150                 pr_buf(out, " %u:%llu:%u", s->ptrs[i].dev,
 151                        (u64) s->ptrs[i].offset,
 152                        stripe_blockcount_get(s, i));
 153 }
 154
 155 /* returns blocknr in stripe that we matched: */
 156 static const struct bch_extent_ptr *bkey_matches_stripe(struct bch_stripe *s,
 157                                                 struct bkey_s_c k, unsigned *block)
 158 {
 159         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 160         const struct bch_extent_ptr *ptr;
 161         unsigned i, nr_data = s->nr_blocks - s->nr_redundant;
 162
 163         bkey_for_each_ptr(ptrs, ptr)
 164                 for (i = 0; i < nr_data; i++)
 165                         if (__bch2_ptr_matches_stripe(&s->ptrs[i], ptr,
 166                                                       le16_to_cpu(s->sectors))) {
 167                                 *block = i;
 168                                 return ptr;
 169                         }
 170
 171         return NULL;
 172 }
 173
 174 static bool extent_has_stripe_ptr(struct bkey_s_c k, u64 idx)
 175 {
 176         switch (k.k->type) {
 177         case KEY_TYPE_extent: {
 178                 struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
 179                 const union bch_extent_entry *entry;
 180
 181                 extent_for_each_entry(e, entry)
 182                         if (extent_entry_type(entry) ==
 183                             BCH_EXTENT_ENTRY_stripe_ptr &&
 184                             entry->stripe_ptr.idx == idx)
 185                                 return true;
 186
 187                 break;
 188         }
 189         }
 190
 191         return false;
 192 }
 193
 194 /* Stripe bufs: */
 195
 196 static void ec_stripe_buf_exit(struct ec_stripe_buf *buf)
 197 {
 198         unsigned i;
 199
 200         for (i = 0; i < buf->key.v.nr_blocks; i++) {
 201                 kvpfree(buf->data[i], buf->size << 9);
 202                 buf->data[i] = NULL;
 203         }
 204 }
 205
 206 static int ec_stripe_buf_init(struct ec_stripe_buf *buf,
 207                                unsigned offset, unsigned size)
 208 {
 209         struct bch_stripe *v = &buf->key.v;
 210         unsigned csum_granularity = 1U << v->csum_granularity_bits;
 211         unsigned end = offset + size;
 212         unsigned i;
 213
 214         BUG_ON(end > le16_to_cpu(v->sectors));
 215
 216         offset  = round_down(offset, csum_granularity);
 217         end     = min_t(unsigned, le16_to_cpu(v->sectors),
 218                         round_up(end, csum_granularity));
 219
 220         buf->offset     = offset;
 221         buf->size       = end - offset;
 222
 223         memset(buf->valid, 0xFF, sizeof(buf->valid));
 224
 225         for (i = 0; i < buf->key.v.nr_blocks; i++) {
 226                 buf->data[i] = kvpmalloc(buf->size << 9, GFP_KERNEL);
 227                 if (!buf->data[i])
 228                         goto err;
 229         }
 230
 231         return 0;
 232 err:
 233         ec_stripe_buf_exit(buf);
 234         return -ENOMEM;
 235 }
 236
 237 /* Checksumming: */
 238
 239 static struct bch_csum ec_block_checksum(struct ec_stripe_buf *buf,
 240                                          unsigned block, unsigned offset)
 241 {
 242         struct bch_stripe *v = &buf->key.v;
 243         unsigned csum_granularity = 1 << v->csum_granularity_bits;
 244         unsigned end = buf->offset + buf->size;
 245         unsigned len = min(csum_granularity, end - offset);
 246
 247         BUG_ON(offset >= end);
 248         BUG_ON(offset <  buf->offset);
 249         BUG_ON(offset & (csum_granularity - 1));
 250         BUG_ON(offset + len != le16_to_cpu(v->sectors) &&
 251                (len & (csum_granularity - 1)));
 252
 253         return bch2_checksum(NULL, v->csum_type,
 254                              null_nonce(),
 255                              buf->data[block] + ((offset - buf->offset) << 9),
 256                              len << 9);
 257 }
 258
 259 static void ec_generate_checksums(struct ec_stripe_buf *buf)
 260 {
 261         struct bch_stripe *v = &buf->key.v;
 262         unsigned i, j, csums_per_device = stripe_csums_per_device(v);
 263
 264         if (!v->csum_type)
 265                 return;
 266
 267         BUG_ON(buf->offset);
 268         BUG_ON(buf->size != le16_to_cpu(v->sectors));
 269
 270         for (i = 0; i < v->nr_blocks; i++)
 271                 for (j = 0; j < csums_per_device; j++)
 272                         stripe_csum_set(v, i, j,
 273                                 ec_block_checksum(buf, i, j << v->csum_granularity_bits));
 274 }
 275
 276 static void ec_validate_checksums(struct bch_fs *c, struct ec_stripe_buf *buf)
 277 {
 278         struct bch_stripe *v = &buf->key.v;
 279         unsigned csum_granularity = 1 << v->csum_granularity_bits;
 280         unsigned i;
 281
 282         if (!v->csum_type)
 283                 return;
 284
 285         for (i = 0; i < v->nr_blocks; i++) {
 286                 unsigned offset = buf->offset;
 287                 unsigned end = buf->offset + buf->size;
 288
 289                 if (!test_bit(i, buf->valid))
 290                         continue;
 291
 292                 while (offset < end) {
 293                         unsigned j = offset >> v->csum_granularity_bits;
 294                         unsigned len = min(csum_granularity, end - offset);
 295                         struct bch_csum want = stripe_csum_get(v, i, j);
 296                         struct bch_csum got = ec_block_checksum(buf, i, offset);
 297
 298                         if (bch2_crc_cmp(want, got)) {
 299                                 struct printbuf buf2 = PRINTBUF;
 300
 301                                 bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(&buf->key.k_i));
 302
 303                                 bch_err_ratelimited(c,
 304                                         "stripe checksum error for %ps at %u:%u: csum type %u, expected %llx got %llx\n%s",
 305                                         (void *) _RET_IP_, i, j, v->csum_type,
 306                                         want.lo, got.lo, buf2.buf);
 307                                 printbuf_exit(&buf2);
 308                                 clear_bit(i, buf->valid);
 309                                 break;
 310                         }
 311
 312                         offset += len;
 313                 }
 314         }
 315 }
 316
 317 /* Erasure coding: */
 318
 319 static void ec_generate_ec(struct ec_stripe_buf *buf)
 320 {
 321         struct bch_stripe *v = &buf->key.v;
 322         unsigned nr_data = v->nr_blocks - v->nr_redundant;
 323         unsigned bytes = le16_to_cpu(v->sectors) << 9;
 324
 325         raid_gen(nr_data, v->nr_redundant, bytes, buf->data);
 326 }
 327
 328 static unsigned ec_nr_failed(struct ec_stripe_buf *buf)
 329 {
 330         return buf->key.v.nr_blocks -
 331                 bitmap_weight(buf->valid, buf->key.v.nr_blocks);
 332 }
 333
 334 static int ec_do_recov(struct bch_fs *c, struct ec_stripe_buf *buf)
 335 {
 336         struct bch_stripe *v = &buf->key.v;
 337         unsigned i, failed[BCH_BKEY_PTRS_MAX], nr_failed = 0;
 338         unsigned nr_data = v->nr_blocks - v->nr_redundant;
 339         unsigned bytes = buf->size << 9;
 340
 341         if (ec_nr_failed(buf) > v->nr_redundant) {
 342                 bch_err_ratelimited(c,
 343                         "error doing reconstruct read: unable to read enough blocks");
 344                 return -1;
 345         }
 346
 347         for (i = 0; i < nr_data; i++)
 348                 if (!test_bit(i, buf->valid))
 349                         failed[nr_failed++] = i;
 350
 351         raid_rec(nr_failed, failed, nr_data, v->nr_redundant, bytes, buf->data);
 352         return 0;
 353 }
 354
 355 /* IO: */
 356
 357 static void ec_block_endio(struct bio *bio)
 358 {
 359         struct ec_bio *ec_bio = container_of(bio, struct ec_bio, bio);
 360         struct bch_stripe *v = &ec_bio->buf->key.v;
 361         struct bch_extent_ptr *ptr = &v->ptrs[ec_bio->idx];
 362         struct bch_dev *ca = ec_bio->ca;
 363         struct closure *cl = bio->bi_private;
 364
 365         if (bch2_dev_io_err_on(bio->bi_status, ca, "erasure coding %s error: %s",
 366                                bio_data_dir(bio) ? "write" : "read",
 367                                bch2_blk_status_to_str(bio->bi_status)))
 368                 clear_bit(ec_bio->idx, ec_bio->buf->valid);
 369
 370         if (ptr_stale(ca, ptr)) {
 371                 bch_err_ratelimited(ca->fs,
 372                                     "error %s stripe: stale pointer after io",
 373                                     bio_data_dir(bio) == READ ? "reading from" : "writing to");
 374                 clear_bit(ec_bio->idx, ec_bio->buf->valid);
 375         }
 376
 377         bio_put(&ec_bio->bio);
 378         percpu_ref_put(&ca->io_ref);
 379         closure_put(cl);
 380 }
 381
 382 static void ec_block_io(struct bch_fs *c, struct ec_stripe_buf *buf,
 383                         unsigned rw, unsigned idx, struct closure *cl)
 384 {
 385         struct bch_stripe *v = &buf->key.v;
 386         unsigned offset = 0, bytes = buf->size << 9;
 387         struct bch_extent_ptr *ptr = &v->ptrs[idx];
 388         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 389         enum bch_data_type data_type = idx < buf->key.v.nr_blocks - buf->key.v.nr_redundant
 390                 ? BCH_DATA_user
 391                 : BCH_DATA_parity;
 392
 393         if (ptr_stale(ca, ptr)) {
 394                 bch_err_ratelimited(c,
 395                                     "error %s stripe: stale pointer",
 396                                     rw == READ ? "reading from" : "writing to");
 397                 clear_bit(idx, buf->valid);
 398                 return;
 399         }
 400
 401         if (!bch2_dev_get_ioref(ca, rw)) {
 402                 clear_bit(idx, buf->valid);
 403                 return;
 404         }
 405
 406         this_cpu_add(ca->io_done->sectors[rw][data_type], buf->size);
 407
 408         while (offset < bytes) {
 409                 unsigned nr_iovecs = min_t(size_t, BIO_MAX_VECS,
 410                                            DIV_ROUND_UP(bytes, PAGE_SIZE));
 411                 unsigned b = min_t(size_t, bytes - offset,
 412                                    nr_iovecs << PAGE_SHIFT);
 413                 struct ec_bio *ec_bio;
 414
 415                 ec_bio = container_of(bio_alloc_bioset(ca->disk_sb.bdev,
 416                                                        nr_iovecs,
 417                                                        rw,
 418                                                        GFP_KERNEL,
 419                                                        &c->ec_bioset),
 420                                       struct ec_bio, bio);
 421
 422                 ec_bio->ca                      = ca;
 423                 ec_bio->buf                     = buf;
 424                 ec_bio->idx                     = idx;
 425
 426                 ec_bio->bio.bi_iter.bi_sector   = ptr->offset + buf->offset + (offset >> 9);
 427                 ec_bio->bio.bi_end_io           = ec_block_endio;
 428                 ec_bio->bio.bi_private          = cl;
 429
 430                 bch2_bio_map(&ec_bio->bio, buf->data[idx] + offset, b);
 431
 432                 closure_get(cl);
 433                 percpu_ref_get(&ca->io_ref);
 434
 435                 submit_bio(&ec_bio->bio);
 436
 437                 offset += b;
 438         }
 439
 440         percpu_ref_put(&ca->io_ref);
 441 }
 442
 443 static int get_stripe_key(struct bch_fs *c, u64 idx, struct ec_stripe_buf *stripe)
 444 {
 445         struct btree_trans trans;
 446         struct btree_iter iter;
 447         struct bkey_s_c k;
 448         int ret;
 449
 450         bch2_trans_init(&trans, c, 0, 0);
 451         bch2_trans_iter_init(&trans, &iter, BTREE_ID_stripes,
 452                              POS(0, idx), BTREE_ITER_SLOTS);
 453         k = bch2_btree_iter_peek_slot(&iter);
 454         ret = bkey_err(k);
 455         if (ret)
 456                 goto err;
 457         if (k.k->type != KEY_TYPE_stripe) {
 458                 ret = -ENOENT;
 459                 goto err;
 460         }
 461         bkey_reassemble(&stripe->key.k_i, k);
 462 err:
 463         bch2_trans_iter_exit(&trans, &iter);
 464         bch2_trans_exit(&trans);
 465         return ret;
 466 }
 467
 468 /* recovery read path: */
 469 int bch2_ec_read_extent(struct bch_fs *c, struct bch_read_bio *rbio)
 470 {
 471         struct ec_stripe_buf *buf;
 472         struct closure cl;
 473         struct bch_stripe *v;
 474         unsigned i, offset;
 475         int ret = 0;
 476
 477         closure_init_stack(&cl);
 478
 479         BUG_ON(!rbio->pick.has_ec);
 480
 481         buf = kzalloc(sizeof(*buf), GFP_NOIO);
 482         if (!buf)
 483                 return -ENOMEM;
 484
 485         ret = get_stripe_key(c, rbio->pick.ec.idx, buf);
 486         if (ret) {
 487                 bch_err_ratelimited(c,
 488                         "error doing reconstruct read: error %i looking up stripe", ret);
 489                 kfree(buf);
 490                 return -EIO;
 491         }
 492
 493         v = &buf->key.v;
 494
 495         if (!bch2_ptr_matches_stripe(v, rbio->pick)) {
 496                 bch_err_ratelimited(c,
 497                         "error doing reconstruct read: pointer doesn't match stripe");
 498                 ret = -EIO;
 499                 goto err;
 500         }
 501
 502         offset = rbio->bio.bi_iter.bi_sector - v->ptrs[rbio->pick.ec.block].offset;
 503         if (offset + bio_sectors(&rbio->bio) > le16_to_cpu(v->sectors)) {
 504                 bch_err_ratelimited(c,
 505                         "error doing reconstruct read: read is bigger than stripe");
 506                 ret = -EIO;
 507                 goto err;
 508         }
 509
 510         ret = ec_stripe_buf_init(buf, offset, bio_sectors(&rbio->bio));
 511         if (ret)
 512                 goto err;
 513
 514         for (i = 0; i < v->nr_blocks; i++)
 515                 ec_block_io(c, buf, REQ_OP_READ, i, &cl);
 516
 517         closure_sync(&cl);
 518
 519         if (ec_nr_failed(buf) > v->nr_redundant) {
 520                 bch_err_ratelimited(c,
 521                         "error doing reconstruct read: unable to read enough blocks");
 522                 ret = -EIO;
 523                 goto err;
 524         }
 525
 526         ec_validate_checksums(c, buf);
 527
 528         ret = ec_do_recov(c, buf);
 529         if (ret)
 530                 goto err;
 531
 532         memcpy_to_bio(&rbio->bio, rbio->bio.bi_iter,
 533                       buf->data[rbio->pick.ec.block] + ((offset - buf->offset) << 9));
 534 err:
 535         ec_stripe_buf_exit(buf);
 536         kfree(buf);
 537         return ret;
 538 }
 539
 540 /* stripe bucket accounting: */
 541
 542 static int __ec_stripe_mem_alloc(struct bch_fs *c, size_t idx, gfp_t gfp)
 543 {
 544         ec_stripes_heap n, *h = &c->ec_stripes_heap;
 545
 546         if (idx >= h->size) {
 547                 if (!init_heap(&n, max(1024UL, roundup_pow_of_two(idx + 1)), gfp))
 548                         return -ENOMEM;
 549
 550                 spin_lock(&c->ec_stripes_heap_lock);
 551                 if (n.size > h->size) {
 552                         memcpy(n.data, h->data, h->used * sizeof(h->data[0]));
 553                         n.used = h->used;
 554                         swap(*h, n);
 555                 }
 556                 spin_unlock(&c->ec_stripes_heap_lock);
 557
 558                 free_heap(&n);
 559         }
 560
 561         if (!genradix_ptr_alloc(&c->stripes, idx, gfp))
 562                 return -ENOMEM;
 563
 564         if (c->gc_pos.phase != GC_PHASE_NOT_RUNNING &&
 565             !genradix_ptr_alloc(&c->gc_stripes, idx, gfp))
 566                 return -ENOMEM;
 567
 568         return 0;
 569 }
 570
 571 static int ec_stripe_mem_alloc(struct btree_trans *trans,
 572                                struct btree_iter *iter)
 573 {
 574         size_t idx = iter->pos.offset;
 575         int ret = 0;
 576
 577         if (!__ec_stripe_mem_alloc(trans->c, idx, GFP_NOWAIT|__GFP_NOWARN))
 578                 return ret;
 579
 580         bch2_trans_unlock(trans);
 581         ret = -EINTR;
 582
 583         if (!__ec_stripe_mem_alloc(trans->c, idx, GFP_KERNEL))
 584                 return ret;
 585
 586         return -ENOMEM;
 587 }
 588
 589 static ssize_t stripe_idx_to_delete(struct bch_fs *c)
 590 {
 591         ec_stripes_heap *h = &c->ec_stripes_heap;
 592
 593         return h->used && h->data[0].blocks_nonempty == 0
 594                 ? h->data[0].idx : -1;
 595 }
 596
 597 static inline int ec_stripes_heap_cmp(ec_stripes_heap *h,
 598                                       struct ec_stripe_heap_entry l,
 599                                       struct ec_stripe_heap_entry r)
 600 {
 601         return ((l.blocks_nonempty > r.blocks_nonempty) -
 602                 (l.blocks_nonempty < r.blocks_nonempty));
 603 }
 604
 605 static inline void ec_stripes_heap_set_backpointer(ec_stripes_heap *h,
 606                                                    size_t i)
 607 {
 608         struct bch_fs *c = container_of(h, struct bch_fs, ec_stripes_heap);
 609
 610         genradix_ptr(&c->stripes, h->data[i].idx)->heap_idx = i;
 611 }
 612
 613 static void heap_verify_backpointer(struct bch_fs *c, size_t idx)
 614 {
 615         ec_stripes_heap *h = &c->ec_stripes_heap;
 616         struct stripe *m = genradix_ptr(&c->stripes, idx);
 617
 618         BUG_ON(!m->alive);
 619         BUG_ON(m->heap_idx >= h->used);
 620         BUG_ON(h->data[m->heap_idx].idx != idx);
 621 }
 622
 623 void bch2_stripes_heap_del(struct bch_fs *c,
 624                            struct stripe *m, size_t idx)
 625 {
 626         if (!m->on_heap)
 627                 return;
 628
 629         m->on_heap = false;
 630
 631         heap_verify_backpointer(c, idx);
 632
 633         heap_del(&c->ec_stripes_heap, m->heap_idx,
 634                  ec_stripes_heap_cmp,
 635                  ec_stripes_heap_set_backpointer);
 636 }
 637
 638 void bch2_stripes_heap_insert(struct bch_fs *c,
 639                               struct stripe *m, size_t idx)
 640 {
 641         if (m->on_heap)
 642                 return;
 643
 644         BUG_ON(heap_full(&c->ec_stripes_heap));
 645
 646         m->on_heap = true;
 647
 648         heap_add(&c->ec_stripes_heap, ((struct ec_stripe_heap_entry) {
 649                         .idx = idx,
 650                         .blocks_nonempty = m->blocks_nonempty,
 651                 }),
 652                  ec_stripes_heap_cmp,
 653                  ec_stripes_heap_set_backpointer);
 654
 655         heap_verify_backpointer(c, idx);
 656 }
 657
 658 void bch2_stripes_heap_update(struct bch_fs *c,
 659                               struct stripe *m, size_t idx)
 660 {
 661         ec_stripes_heap *h = &c->ec_stripes_heap;
 662         size_t i;
 663
 664         if (!m->on_heap)
 665                 return;
 666
 667         heap_verify_backpointer(c, idx);
 668
 669         h->data[m->heap_idx].blocks_nonempty = m->blocks_nonempty;
 670
 671         i = m->heap_idx;
 672         heap_sift_up(h,   i, ec_stripes_heap_cmp,
 673                      ec_stripes_heap_set_backpointer);
 674         heap_sift_down(h, i, ec_stripes_heap_cmp,
 675                        ec_stripes_heap_set_backpointer);
 676
 677         heap_verify_backpointer(c, idx);
 678
 679         if (stripe_idx_to_delete(c) >= 0 &&
 680             !percpu_ref_is_dying(&c->writes))
 681                 schedule_work(&c->ec_stripe_delete_work);
 682 }
 683
 684 /* stripe deletion */
 685
 686 static int ec_stripe_delete(struct bch_fs *c, size_t idx)
 687 {
 688         return bch2_btree_delete_range(c, BTREE_ID_stripes,
 689                                        POS(0, idx),
 690                                        POS(0, idx + 1),
 691                                        0, NULL);
 692 }
 693
 694 static void ec_stripe_delete_work(struct work_struct *work)
 695 {
 696         struct bch_fs *c =
 697                 container_of(work, struct bch_fs, ec_stripe_delete_work);
 698         ssize_t idx;
 699
 700         while (1) {
 701                 spin_lock(&c->ec_stripes_heap_lock);
 702                 idx = stripe_idx_to_delete(c);
 703                 if (idx < 0) {
 704                         spin_unlock(&c->ec_stripes_heap_lock);
 705                         break;
 706                 }
 707
 708                 bch2_stripes_heap_del(c, genradix_ptr(&c->stripes, idx), idx);
 709                 spin_unlock(&c->ec_stripes_heap_lock);
 710
 711                 if (ec_stripe_delete(c, idx))
 712                         break;
 713         }
 714 }
 715
 716 /* stripe creation: */
 717
 718 static int ec_stripe_bkey_insert(struct btree_trans *trans,
 719                                  struct bkey_i_stripe *stripe,
 720                                  struct disk_reservation *res)
 721 {
 722         struct bch_fs *c = trans->c;
 723         struct btree_iter iter;
 724         struct bkey_s_c k;
 725         struct bpos min_pos = POS(0, 1);
 726         struct bpos start_pos = bpos_max(min_pos, POS(0, c->ec_stripe_hint));
 727         int ret;
 728
 729         for_each_btree_key(trans, iter, BTREE_ID_stripes, start_pos,
 730                            BTREE_ITER_SLOTS|BTREE_ITER_INTENT, k, ret) {
 731                 if (bkey_cmp(k.k->p, POS(0, U32_MAX)) > 0) {
 732                         if (start_pos.offset) {
 733                                 start_pos = min_pos;
 734                                 bch2_btree_iter_set_pos(&iter, start_pos);
 735                                 continue;
 736                         }
 737
 738                         ret = -ENOSPC;
 739                         break;
 740                 }
 741
 742                 if (bkey_deleted(k.k))
 743                         goto found_slot;
 744         }
 745
 746         goto err;
 747 found_slot:
 748         start_pos = iter.pos;
 749
 750         ret = ec_stripe_mem_alloc(trans, &iter);
 751         if (ret)
 752                 goto err;
 753
 754         stripe->k.p = iter.pos;
 755
 756         ret = bch2_trans_update(trans, &iter, &stripe->k_i, 0);
 757
 758         c->ec_stripe_hint = start_pos.offset;
 759 err:
 760         bch2_trans_iter_exit(trans, &iter);
 761
 762         return ret;
 763 }
 764
 765 static int ec_stripe_bkey_update(struct btree_trans *trans,
 766                                  struct bkey_i_stripe *new,
 767                                  struct disk_reservation *res)
 768 {
 769         struct btree_iter iter;
 770         struct bkey_s_c k;
 771         const struct bch_stripe *existing;
 772         unsigned i;
 773         int ret;
 774
 775         bch2_trans_iter_init(trans, &iter, BTREE_ID_stripes,
 776                              new->k.p, BTREE_ITER_INTENT);
 777         k = bch2_btree_iter_peek_slot(&iter);
 778         ret = bkey_err(k);
 779         if (ret)
 780                 goto err;
 781
 782         if (!k.k || k.k->type != KEY_TYPE_stripe) {
 783                 bch_err(trans->c, "error updating stripe: not found");
 784                 ret = -ENOENT;
 785                 goto err;
 786         }
 787
 788         existing = bkey_s_c_to_stripe(k).v;
 789
 790         if (existing->nr_blocks != new->v.nr_blocks) {
 791                 bch_err(trans->c, "error updating stripe: nr_blocks does not match");
 792                 ret = -EINVAL;
 793                 goto err;
 794         }
 795
 796         for (i = 0; i < new->v.nr_blocks; i++)
 797                 stripe_blockcount_set(&new->v, i,
 798                         stripe_blockcount_get(existing, i));
 799
 800         ret = bch2_trans_update(trans, &iter, &new->k_i, 0);
 801 err:
 802         bch2_trans_iter_exit(trans, &iter);
 803         return ret;
 804 }
 805
 806 static void extent_stripe_ptr_add(struct bkey_s_extent e,
 807                                   struct ec_stripe_buf *s,
 808                                   struct bch_extent_ptr *ptr,
 809                                   unsigned block)
 810 {
 811         struct bch_extent_stripe_ptr *dst = (void *) ptr;
 812         union bch_extent_entry *end = extent_entry_last(e);
 813
 814         memmove_u64s_up(dst + 1, dst, (u64 *) end - (u64 *) dst);
 815         e.k->u64s += sizeof(*dst) / sizeof(u64);
 816
 817         *dst = (struct bch_extent_stripe_ptr) {
 818                 .type = 1 << BCH_EXTENT_ENTRY_stripe_ptr,
 819                 .block          = block,
 820                 .redundancy     = s->key.v.nr_redundant,
 821                 .idx            = s->key.k.p.offset,
 822         };
 823 }
 824
 825 static int ec_stripe_update_ptrs(struct bch_fs *c,
 826                                  struct ec_stripe_buf *s,
 827                                  struct bkey *pos)
 828 {
 829         struct btree_trans trans;
 830         struct btree_iter iter;
 831         struct bkey_s_c k;
 832         struct bkey_s_extent e;
 833         struct bkey_buf sk;
 834         struct bpos next_pos;
 835         int ret = 0, dev, block;
 836
 837         bch2_bkey_buf_init(&sk);
 838         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
 839
 840         /* XXX this doesn't support the reflink btree */
 841
 842         bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
 843                              bkey_start_pos(pos),
 844                              BTREE_ITER_INTENT);
 845 retry:
 846         while (bch2_trans_begin(&trans),
 847                (k = bch2_btree_iter_peek(&iter)).k &&
 848                !(ret = bkey_err(k)) &&
 849                bkey_cmp(bkey_start_pos(k.k), pos->p) < 0) {
 850                 const struct bch_extent_ptr *ptr_c;
 851                 struct bch_extent_ptr *ptr, *ec_ptr = NULL;
 852
 853                 if (extent_has_stripe_ptr(k, s->key.k.p.offset)) {
 854                         bch2_btree_iter_advance(&iter);
 855                         continue;
 856                 }
 857
 858                 ptr_c = bkey_matches_stripe(&s->key.v, k, &block);
 859                 /*
 860                  * It doesn't generally make sense to erasure code cached ptrs:
 861                  * XXX: should we be incrementing a counter?
 862                  */
 863                 if (!ptr_c || ptr_c->cached) {
 864                         bch2_btree_iter_advance(&iter);
 865                         continue;
 866                 }
 867
 868                 dev = s->key.v.ptrs[block].dev;
 869
 870                 bch2_bkey_buf_reassemble(&sk, c, k);
 871                 e = bkey_i_to_s_extent(sk.k);
 872
 873                 bch2_bkey_drop_ptrs(e.s, ptr, ptr->dev != dev);
 874                 ec_ptr = (void *) bch2_bkey_has_device(e.s_c, dev);
 875                 BUG_ON(!ec_ptr);
 876
 877                 extent_stripe_ptr_add(e, s, ec_ptr, block);
 878
 879                 bch2_btree_iter_set_pos(&iter, bkey_start_pos(&sk.k->k));
 880                 next_pos = sk.k->k.p;
 881
 882                 ret   = bch2_btree_iter_traverse(&iter) ?:
 883                         bch2_trans_update(&trans, &iter, sk.k, 0) ?:
 884                         bch2_trans_commit(&trans, NULL, NULL,
 885                                         BTREE_INSERT_NOFAIL);
 886                 if (!ret)
 887                         bch2_btree_iter_set_pos(&iter, next_pos);
 888                 if (ret)
 889                         break;
 890         }
 891         if (ret == -EINTR)
 892                 goto retry;
 893         bch2_trans_iter_exit(&trans, &iter);
 894
 895         bch2_trans_exit(&trans);
 896         bch2_bkey_buf_exit(&sk, c);
 897
 898         return ret;
 899 }
 900
 901 /*
 902  * data buckets of new stripe all written: create the stripe
 903  */
 904 static void ec_stripe_create(struct ec_stripe_new *s)
 905 {
 906         struct bch_fs *c = s->c;
 907         struct open_bucket *ob;
 908         struct bkey_i *k;
 909         struct stripe *m;
 910         struct bch_stripe *v = &s->new_stripe.key.v;
 911         unsigned i, nr_data = v->nr_blocks - v->nr_redundant;
 912         int ret;
 913
 914         BUG_ON(s->h->s == s);
 915
 916         closure_sync(&s->iodone);
 917
 918         if (s->err) {
 919                 if (s->err != -EROFS)
 920                         bch_err(c, "error creating stripe: error writing data buckets");
 921                 goto err;
 922         }
 923
 924         if (s->have_existing_stripe) {
 925                 ec_validate_checksums(c, &s->existing_stripe);
 926
 927                 if (ec_do_recov(c, &s->existing_stripe)) {
 928                         bch_err(c, "error creating stripe: error reading existing stripe");
 929                         goto err;
 930                 }
 931
 932                 for (i = 0; i < nr_data; i++)
 933                         if (stripe_blockcount_get(&s->existing_stripe.key.v, i))
 934                                 swap(s->new_stripe.data[i],
 935                                      s->existing_stripe.data[i]);
 936
 937                 ec_stripe_buf_exit(&s->existing_stripe);
 938         }
 939
 940         BUG_ON(!s->allocated);
 941
 942         if (!percpu_ref_tryget(&c->writes))
 943                 goto err;
 944
 945         ec_generate_ec(&s->new_stripe);
 946
 947         ec_generate_checksums(&s->new_stripe);
 948
 949         /* write p/q: */
 950         for (i = nr_data; i < v->nr_blocks; i++)
 951                 ec_block_io(c, &s->new_stripe, REQ_OP_WRITE, i, &s->iodone);
 952         closure_sync(&s->iodone);
 953
 954         if (ec_nr_failed(&s->new_stripe)) {
 955                 bch_err(c, "error creating stripe: error writing redundancy buckets");
 956                 goto err_put_writes;
 957         }
 958
 959         ret = bch2_trans_do(c, &s->res, NULL, BTREE_INSERT_NOFAIL,
 960                             s->have_existing_stripe
 961                             ? ec_stripe_bkey_update(&trans, &s->new_stripe.key, &s->res)
 962                             : ec_stripe_bkey_insert(&trans, &s->new_stripe.key, &s->res));
 963         if (ret) {
 964                 bch_err(c, "error creating stripe: error creating stripe key");
 965                 goto err_put_writes;
 966         }
 967
 968         for_each_keylist_key(&s->keys, k) {
 969                 ret = ec_stripe_update_ptrs(c, &s->new_stripe, &k->k);
 970                 if (ret) {
 971                         bch_err(c, "error creating stripe: error %i updating pointers", ret);
 972                         break;
 973                 }
 974         }
 975
 976         spin_lock(&c->ec_stripes_heap_lock);
 977         m = genradix_ptr(&c->stripes, s->new_stripe.key.k.p.offset);
 978
 979         BUG_ON(m->on_heap);
 980         bch2_stripes_heap_insert(c, m, s->new_stripe.key.k.p.offset);
 981         spin_unlock(&c->ec_stripes_heap_lock);
 982 err_put_writes:
 983         percpu_ref_put(&c->writes);
 984 err:
 985         bch2_disk_reservation_put(c, &s->res);
 986
 987         for (i = 0; i < v->nr_blocks; i++)
 988                 if (s->blocks[i]) {
 989                         ob = c->open_buckets + s->blocks[i];
 990
 991                         if (i < nr_data) {
 992                                 ob->ec = NULL;
 993                                 __bch2_open_bucket_put(c, ob);
 994                         } else {
 995                                 bch2_open_bucket_put(c, ob);
 996                         }
 997                 }
 998
 999         bch2_keylist_free(&s->keys, s->inline_keys);
1000
1001         ec_stripe_buf_exit(&s->existing_stripe);
1002         ec_stripe_buf_exit(&s->new_stripe);
1003         closure_debug_destroy(&s->iodone);
1004         kfree(s);
1005 }
1006
1007 static void ec_stripe_create_work(struct work_struct *work)
1008 {
1009         struct bch_fs *c = container_of(work,
1010                 struct bch_fs, ec_stripe_create_work);
1011         struct ec_stripe_new *s, *n;
1012 restart:
1013         mutex_lock(&c->ec_stripe_new_lock);
1014         list_for_each_entry_safe(s, n, &c->ec_stripe_new_list, list)
1015                 if (!atomic_read(&s->pin)) {
1016                         list_del(&s->list);
1017                         mutex_unlock(&c->ec_stripe_new_lock);
1018                         ec_stripe_create(s);
1019                         goto restart;
1020                 }
1021         mutex_unlock(&c->ec_stripe_new_lock);
1022 }
1023
1024 static void ec_stripe_new_put(struct bch_fs *c, struct ec_stripe_new *s)
1025 {
1026         BUG_ON(atomic_read(&s->pin) <= 0);
1027
1028         if (atomic_dec_and_test(&s->pin)) {
1029                 BUG_ON(!s->pending);
1030                 queue_work(system_long_wq, &c->ec_stripe_create_work);
1031         }
1032 }
1033
1034 static void ec_stripe_set_pending(struct bch_fs *c, struct ec_stripe_head *h)
1035 {
1036         struct ec_stripe_new *s = h->s;
1037
1038         BUG_ON(!s->allocated && !s->err);
1039
1040         h->s            = NULL;
1041         s->pending      = true;
1042
1043         mutex_lock(&c->ec_stripe_new_lock);
1044         list_add(&s->list, &c->ec_stripe_new_list);
1045         mutex_unlock(&c->ec_stripe_new_lock);
1046
1047         ec_stripe_new_put(c, s);
1048 }
1049
1050 /* have a full bucket - hand it off to be erasure coded: */
1051 void bch2_ec_bucket_written(struct bch_fs *c, struct open_bucket *ob)
1052 {
1053         struct ec_stripe_new *s = ob->ec;
1054
1055         if (ob->sectors_free)
1056                 s->err = -1;
1057
1058         ec_stripe_new_put(c, s);
1059 }
1060
1061 void bch2_ec_bucket_cancel(struct bch_fs *c, struct open_bucket *ob)
1062 {
1063         struct ec_stripe_new *s = ob->ec;
1064
1065         s->err = -EIO;
1066 }
1067
1068 void *bch2_writepoint_ec_buf(struct bch_fs *c, struct write_point *wp)
1069 {
1070         struct open_bucket *ob = ec_open_bucket(c, &wp->ptrs);
1071         struct bch_dev *ca;
1072         unsigned offset;
1073
1074         if (!ob)
1075                 return NULL;
1076
1077         ca      = bch_dev_bkey_exists(c, ob->dev);
1078         offset  = ca->mi.bucket_size - ob->sectors_free;
1079
1080         return ob->ec->new_stripe.data[ob->ec_idx] + (offset << 9);
1081 }
1082
1083 void bch2_ob_add_backpointer(struct bch_fs *c, struct open_bucket *ob,
1084                              struct bkey *k)
1085 {
1086         struct ec_stripe_new *ec = ob->ec;
1087
1088         if (!ec)
1089                 return;
1090
1091         mutex_lock(&ec->lock);
1092
1093         if (bch2_keylist_realloc(&ec->keys, ec->inline_keys,
1094                                  ARRAY_SIZE(ec->inline_keys),
1095                                  BKEY_U64s)) {
1096                 BUG();
1097         }
1098
1099         bkey_init(&ec->keys.top->k);
1100         ec->keys.top->k.p       = k->p;
1101         ec->keys.top->k.size    = k->size;
1102         bch2_keylist_push(&ec->keys);
1103
1104         mutex_unlock(&ec->lock);
1105 }
1106
1107 static int unsigned_cmp(const void *_l, const void *_r)
1108 {
1109         unsigned l = *((const unsigned *) _l);
1110         unsigned r = *((const unsigned *) _r);
1111
1112         return cmp_int(l, r);
1113 }
1114
1115 /* pick most common bucket size: */
1116 static unsigned pick_blocksize(struct bch_fs *c,
1117                                struct bch_devs_mask *devs)
1118 {
1119         struct bch_dev *ca;
1120         unsigned i, nr = 0, sizes[BCH_SB_MEMBERS_MAX];
1121         struct {
1122                 unsigned nr, size;
1123         } cur = { 0, 0 }, best = { 0, 0 };
1124
1125         for_each_member_device_rcu(ca, c, i, devs)
1126                 sizes[nr++] = ca->mi.bucket_size;
1127
1128         sort(sizes, nr, sizeof(unsigned), unsigned_cmp, NULL);
1129
1130         for (i = 0; i < nr; i++) {
1131                 if (sizes[i] != cur.size) {
1132                         if (cur.nr > best.nr)
1133                                 best = cur;
1134
1135                         cur.nr = 0;
1136                         cur.size = sizes[i];
1137                 }
1138
1139                 cur.nr++;
1140         }
1141
1142         if (cur.nr > best.nr)
1143                 best = cur;
1144
1145         return best.size;
1146 }
1147
1148 static bool may_create_new_stripe(struct bch_fs *c)
1149 {
1150         return false;
1151 }
1152
1153 static void ec_stripe_key_init(struct bch_fs *c,
1154                                struct bkey_i_stripe *s,
1155                                unsigned nr_data,
1156                                unsigned nr_parity,
1157                                unsigned stripe_size)
1158 {
1159         unsigned u64s;
1160
1161         bkey_stripe_init(&s->k_i);
1162         s->v.sectors                    = cpu_to_le16(stripe_size);
1163         s->v.algorithm                  = 0;
1164         s->v.nr_blocks                  = nr_data + nr_parity;
1165         s->v.nr_redundant               = nr_parity;
1166         s->v.csum_granularity_bits      = ilog2(c->opts.encoded_extent_max >> 9);
1167         s->v.csum_type                  = BCH_CSUM_crc32c;
1168         s->v.pad                        = 0;
1169
1170         while ((u64s = stripe_val_u64s(&s->v)) > BKEY_VAL_U64s_MAX) {
1171                 BUG_ON(1 << s->v.csum_granularity_bits >=
1172                        le16_to_cpu(s->v.sectors) ||
1173                        s->v.csum_granularity_bits == U8_MAX);
1174                 s->v.csum_granularity_bits++;
1175         }
1176
1177         set_bkey_val_u64s(&s->k, u64s);
1178 }
1179
1180 static int ec_new_stripe_alloc(struct bch_fs *c, struct ec_stripe_head *h)
1181 {
1182         struct ec_stripe_new *s;
1183
1184         lockdep_assert_held(&h->lock);
1185
1186         s = kzalloc(sizeof(*s), GFP_KERNEL);
1187         if (!s)
1188                 return -ENOMEM;
1189
1190         mutex_init(&s->lock);
1191         closure_init(&s->iodone, NULL);
1192         atomic_set(&s->pin, 1);
1193         s->c            = c;
1194         s->h            = h;
1195         s->nr_data      = min_t(unsigned, h->nr_active_devs,
1196                                 BCH_BKEY_PTRS_MAX) - h->redundancy;
1197         s->nr_parity    = h->redundancy;
1198
1199         bch2_keylist_init(&s->keys, s->inline_keys);
1200
1201         ec_stripe_key_init(c, &s->new_stripe.key, s->nr_data,
1202                            s->nr_parity, h->blocksize);
1203
1204         h->s = s;
1205         return 0;
1206 }
1207
1208 static struct ec_stripe_head *
1209 ec_new_stripe_head_alloc(struct bch_fs *c, unsigned target,
1210                          unsigned algo, unsigned redundancy,
1211                          bool copygc)
1212 {
1213         struct ec_stripe_head *h;
1214         struct bch_dev *ca;
1215         unsigned i;
1216
1217         h = kzalloc(sizeof(*h), GFP_KERNEL);
1218         if (!h)
1219                 return NULL;
1220
1221         mutex_init(&h->lock);
1222         mutex_lock(&h->lock);
1223
1224         h->target       = target;
1225         h->algo         = algo;
1226         h->redundancy   = redundancy;
1227         h->copygc       = copygc;
1228
1229         rcu_read_lock();
1230         h->devs = target_rw_devs(c, BCH_DATA_user, target);
1231
1232         for_each_member_device_rcu(ca, c, i, &h->devs)
1233                 if (!ca->mi.durability)
1234                         __clear_bit(i, h->devs.d);
1235
1236         h->blocksize = pick_blocksize(c, &h->devs);
1237
1238         for_each_member_device_rcu(ca, c, i, &h->devs)
1239                 if (ca->mi.bucket_size == h->blocksize)
1240                         h->nr_active_devs++;
1241
1242         rcu_read_unlock();
1243         list_add(&h->list, &c->ec_stripe_head_list);
1244         return h;
1245 }
1246
1247 void bch2_ec_stripe_head_put(struct bch_fs *c, struct ec_stripe_head *h)
1248 {
1249         if (h->s &&
1250             h->s->allocated &&
1251             bitmap_weight(h->s->blocks_allocated,
1252                           h->s->nr_data) == h->s->nr_data)
1253                 ec_stripe_set_pending(c, h);
1254
1255         mutex_unlock(&h->lock);
1256 }
1257
1258 struct ec_stripe_head *__bch2_ec_stripe_head_get(struct bch_fs *c,
1259                                                  unsigned target,
1260                                                  unsigned algo,
1261                                                  unsigned redundancy,
1262                                                  bool copygc)
1263 {
1264         struct ec_stripe_head *h;
1265
1266         if (!redundancy)
1267                 return NULL;
1268
1269         mutex_lock(&c->ec_stripe_head_lock);
1270         list_for_each_entry(h, &c->ec_stripe_head_list, list)
1271                 if (h->target           == target &&
1272                     h->algo             == algo &&
1273                     h->redundancy       == redundancy &&
1274                     h->copygc           == copygc) {
1275                         mutex_lock(&h->lock);
1276                         goto found;
1277                 }
1278
1279         h = ec_new_stripe_head_alloc(c, target, algo, redundancy, copygc);
1280 found:
1281         mutex_unlock(&c->ec_stripe_head_lock);
1282         return h;
1283 }
1284
1285 static int new_stripe_alloc_buckets(struct bch_fs *c, struct ec_stripe_head *h,
1286                                     struct closure *cl)
1287 {
1288         struct bch_devs_mask devs = h->devs;
1289         struct open_bucket *ob;
1290         struct open_buckets buckets;
1291         unsigned i, j, nr_have_parity = 0, nr_have_data = 0;
1292         bool have_cache = true;
1293         int ret = 0;
1294
1295         for (i = 0; i < h->s->new_stripe.key.v.nr_blocks; i++) {
1296                 if (test_bit(i, h->s->blocks_gotten)) {
1297                         __clear_bit(h->s->new_stripe.key.v.ptrs[i].dev, devs.d);
1298                         if (i < h->s->nr_data)
1299                                 nr_have_data++;
1300                         else
1301                                 nr_have_parity++;
1302                 }
1303         }
1304
1305         BUG_ON(nr_have_data     > h->s->nr_data);
1306         BUG_ON(nr_have_parity   > h->s->nr_parity);
1307
1308         buckets.nr = 0;
1309         if (nr_have_parity < h->s->nr_parity) {
1310                 ret = bch2_bucket_alloc_set(c, &buckets,
1311                                             &h->parity_stripe,
1312                                             &devs,
1313                                             h->s->nr_parity,
1314                                             &nr_have_parity,
1315                                             &have_cache,
1316                                             h->copygc
1317                                             ? RESERVE_movinggc
1318                                             : RESERVE_none,
1319                                             0,
1320                                             cl);
1321
1322                 open_bucket_for_each(c, &buckets, ob, i) {
1323                         j = find_next_zero_bit(h->s->blocks_gotten,
1324                                                h->s->nr_data + h->s->nr_parity,
1325                                                h->s->nr_data);
1326                         BUG_ON(j >= h->s->nr_data + h->s->nr_parity);
1327
1328                         h->s->blocks[j] = buckets.v[i];
1329                         h->s->new_stripe.key.v.ptrs[j] = bch2_ob_ptr(c, ob);
1330                         __set_bit(j, h->s->blocks_gotten);
1331                 }
1332
1333                 if (ret)
1334                         return ret;
1335         }
1336
1337         buckets.nr = 0;
1338         if (nr_have_data < h->s->nr_data) {
1339                 ret = bch2_bucket_alloc_set(c, &buckets,
1340                                             &h->block_stripe,
1341                                             &devs,
1342                                             h->s->nr_data,
1343                                             &nr_have_data,
1344                                             &have_cache,
1345                                             h->copygc
1346                                             ? RESERVE_movinggc
1347                                             : RESERVE_none,
1348                                             0,
1349                                             cl);
1350
1351                 open_bucket_for_each(c, &buckets, ob, i) {
1352                         j = find_next_zero_bit(h->s->blocks_gotten,
1353                                                h->s->nr_data, 0);
1354                         BUG_ON(j >= h->s->nr_data);
1355
1356                         h->s->blocks[j] = buckets.v[i];
1357                         h->s->new_stripe.key.v.ptrs[j] = bch2_ob_ptr(c, ob);
1358                         __set_bit(j, h->s->blocks_gotten);
1359                 }
1360
1361                 if (ret)
1362                         return ret;
1363         }
1364
1365         return 0;
1366 }
1367
1368 /* XXX: doesn't obey target: */
1369 static s64 get_existing_stripe(struct bch_fs *c,
1370                                struct ec_stripe_head *head)
1371 {
1372         ec_stripes_heap *h = &c->ec_stripes_heap;
1373         struct stripe *m;
1374         size_t heap_idx;
1375         u64 stripe_idx;
1376         s64 ret = -1;
1377
1378         if (may_create_new_stripe(c))
1379                 return -1;
1380
1381         spin_lock(&c->ec_stripes_heap_lock);
1382         for (heap_idx = 0; heap_idx < h->used; heap_idx++) {
1383                 /* No blocks worth reusing, stripe will just be deleted: */
1384                 if (!h->data[heap_idx].blocks_nonempty)
1385                         continue;
1386
1387                 stripe_idx = h->data[heap_idx].idx;
1388                 m = genradix_ptr(&c->stripes, stripe_idx);
1389
1390                 if (m->algorithm        == head->algo &&
1391                     m->nr_redundant     == head->redundancy &&
1392                     m->sectors          == head->blocksize &&
1393                     m->blocks_nonempty  < m->nr_blocks - m->nr_redundant) {
1394                         bch2_stripes_heap_del(c, m, stripe_idx);
1395                         ret = stripe_idx;
1396                         break;
1397                 }
1398         }
1399         spin_unlock(&c->ec_stripes_heap_lock);
1400         return ret;
1401 }
1402
1403 static int __bch2_ec_stripe_head_reuse(struct bch_fs *c,
1404                                                    struct ec_stripe_head *h)
1405 {
1406         unsigned i;
1407         s64 idx;
1408         int ret;
1409
1410         idx = get_existing_stripe(c, h);
1411         if (idx < 0) {
1412                 bch_err(c, "failed to find an existing stripe");
1413                 return -ENOSPC;
1414         }
1415
1416         h->s->have_existing_stripe = true;
1417         ret = get_stripe_key(c, idx, &h->s->existing_stripe);
1418         if (ret) {
1419                 bch2_fs_fatal_error(c, "error reading stripe key: %i", ret);
1420                 return ret;
1421         }
1422
1423         if (ec_stripe_buf_init(&h->s->existing_stripe, 0, h->blocksize)) {
1424                 /*
1425                  * this is a problem: we have deleted from the
1426                  * stripes heap already
1427                  */
1428                 BUG();
1429         }
1430
1431         BUG_ON(h->s->existing_stripe.size != h->blocksize);
1432         BUG_ON(h->s->existing_stripe.size != h->s->existing_stripe.key.v.sectors);
1433
1434         for (i = 0; i < h->s->existing_stripe.key.v.nr_blocks; i++) {
1435                 if (stripe_blockcount_get(&h->s->existing_stripe.key.v, i)) {
1436                         __set_bit(i, h->s->blocks_gotten);
1437                         __set_bit(i, h->s->blocks_allocated);
1438                 }
1439
1440                 ec_block_io(c, &h->s->existing_stripe, READ, i, &h->s->iodone);
1441         }
1442
1443         bkey_copy(&h->s->new_stripe.key.k_i,
1444                         &h->s->existing_stripe.key.k_i);
1445
1446         return 0;
1447 }
1448
1449 static int __bch2_ec_stripe_head_reserve(struct bch_fs *c,
1450                                                         struct ec_stripe_head *h)
1451 {
1452         int ret;
1453
1454         ret = bch2_disk_reservation_get(c, &h->s->res,
1455                         h->blocksize,
1456                         h->s->nr_parity, 0);
1457
1458         if (ret) {
1459                 /*
1460                  * This means we need to wait for copygc to
1461                  * empty out buckets from existing stripes:
1462                  */
1463                 bch_err(c, "failed to reserve stripe");
1464         }
1465
1466         return ret;
1467 }
1468
1469 struct ec_stripe_head *bch2_ec_stripe_head_get(struct bch_fs *c,
1470                                                unsigned target,
1471                                                unsigned algo,
1472                                                unsigned redundancy,
1473                                                bool copygc,
1474                                                struct closure *cl)
1475 {
1476         struct ec_stripe_head *h;
1477         int ret;
1478         bool needs_stripe_new;
1479
1480         h = __bch2_ec_stripe_head_get(c, target, algo, redundancy, copygc);
1481         if (!h) {
1482                 bch_err(c, "no stripe head");
1483                 return NULL;
1484         }
1485
1486         needs_stripe_new = !h->s;
1487         if (needs_stripe_new) {
1488                 if (ec_new_stripe_alloc(c, h)) {
1489                         ret = -ENOMEM;
1490                         bch_err(c, "failed to allocate new stripe");
1491                         goto err;
1492                 }
1493
1494                 if (ec_stripe_buf_init(&h->s->new_stripe, 0, h->blocksize))
1495                         BUG();
1496         }
1497
1498         /*
1499          * Try reserve a new stripe before reusing an
1500          * existing stripe. This will prevent unnecessary
1501          * read amplification during write oriented workloads.
1502          */
1503         ret = 0;
1504         if (!h->s->allocated && !h->s->res.sectors && !h->s->have_existing_stripe)
1505                 ret = __bch2_ec_stripe_head_reserve(c, h);
1506         if (ret && needs_stripe_new)
1507                 ret = __bch2_ec_stripe_head_reuse(c, h);
1508         if (ret)
1509                 goto err;
1510
1511         if (!h->s->allocated) {
1512                 ret = new_stripe_alloc_buckets(c, h, cl);
1513                 if (ret)
1514                         goto err;
1515
1516                 h->s->allocated = true;
1517         }
1518
1519         return h;
1520
1521 err:
1522         bch2_ec_stripe_head_put(c, h);
1523         return ERR_PTR(ret);
1524 }
1525
1526 void bch2_ec_stop_dev(struct bch_fs *c, struct bch_dev *ca)
1527 {
1528         struct ec_stripe_head *h;
1529         struct open_bucket *ob;
1530         unsigned i;
1531
1532         mutex_lock(&c->ec_stripe_head_lock);
1533         list_for_each_entry(h, &c->ec_stripe_head_list, list) {
1534
1535                 mutex_lock(&h->lock);
1536                 if (!h->s)
1537                         goto unlock;
1538
1539                 for (i = 0; i < h->s->new_stripe.key.v.nr_blocks; i++) {
1540                         if (!h->s->blocks[i])
1541                                 continue;
1542
1543                         ob = c->open_buckets + h->s->blocks[i];
1544                         if (ob->dev == ca->dev_idx)
1545                                 goto found;
1546                 }
1547                 goto unlock;
1548 found:
1549                 h->s->err = -EROFS;
1550                 ec_stripe_set_pending(c, h);
1551 unlock:
1552                 mutex_unlock(&h->lock);
1553         }
1554         mutex_unlock(&c->ec_stripe_head_lock);
1555 }
1556
1557 void bch2_stripes_heap_start(struct bch_fs *c)
1558 {
1559         struct genradix_iter iter;
1560         struct stripe *m;
1561
1562         genradix_for_each(&c->stripes, iter, m)
1563                 if (m->alive)
1564                         bch2_stripes_heap_insert(c, m, iter.pos);
1565 }
1566
1567 int bch2_stripes_read(struct bch_fs *c)
1568 {
1569         struct btree_trans trans;
1570         struct btree_iter iter;
1571         struct bkey_s_c k;
1572         const struct bch_stripe *s;
1573         struct stripe *m;
1574         unsigned i;
1575         int ret;
1576
1577         bch2_trans_init(&trans, c, 0, 0);
1578
1579         for_each_btree_key(&trans, iter, BTREE_ID_stripes, POS_MIN,
1580                            BTREE_ITER_PREFETCH, k, ret) {
1581                 if (k.k->type != KEY_TYPE_stripe)
1582                         continue;
1583
1584                 ret = __ec_stripe_mem_alloc(c, k.k->p.offset, GFP_KERNEL);
1585                 if (ret)
1586                         break;
1587
1588                 s = bkey_s_c_to_stripe(k).v;
1589
1590                 m = genradix_ptr(&c->stripes, k.k->p.offset);
1591                 m->alive        = true;
1592                 m->sectors      = le16_to_cpu(s->sectors);
1593                 m->algorithm    = s->algorithm;
1594                 m->nr_blocks    = s->nr_blocks;
1595                 m->nr_redundant = s->nr_redundant;
1596                 m->blocks_nonempty = 0;
1597
1598                 for (i = 0; i < s->nr_blocks; i++)
1599                         m->blocks_nonempty += !!stripe_blockcount_get(s, i);
1600
1601                 spin_lock(&c->ec_stripes_heap_lock);
1602                 bch2_stripes_heap_update(c, m, k.k->p.offset);
1603                 spin_unlock(&c->ec_stripes_heap_lock);
1604         }
1605         bch2_trans_iter_exit(&trans, &iter);
1606
1607         bch2_trans_exit(&trans);
1608
1609         if (ret)
1610                 bch_err(c, "error reading stripes: %i", ret);
1611
1612         return ret;
1613 }
1614
1615 void bch2_stripes_heap_to_text(struct printbuf *out, struct bch_fs *c)
1616 {
1617         ec_stripes_heap *h = &c->ec_stripes_heap;
1618         struct stripe *m;
1619         size_t i;
1620
1621         spin_lock(&c->ec_stripes_heap_lock);
1622         for (i = 0; i < min_t(size_t, h->used, 20); i++) {
1623                 m = genradix_ptr(&c->stripes, h->data[i].idx);
1624
1625                 pr_buf(out, "%zu %u/%u+%u\n", h->data[i].idx,
1626                        h->data[i].blocks_nonempty,
1627                        m->nr_blocks - m->nr_redundant,
1628                        m->nr_redundant);
1629         }
1630         spin_unlock(&c->ec_stripes_heap_lock);
1631 }
1632
1633 void bch2_new_stripes_to_text(struct printbuf *out, struct bch_fs *c)
1634 {
1635         struct ec_stripe_head *h;
1636         struct ec_stripe_new *s;
1637
1638         mutex_lock(&c->ec_stripe_head_lock);
1639         list_for_each_entry(h, &c->ec_stripe_head_list, list) {
1640                 pr_buf(out, "target %u algo %u redundancy %u:\n",
1641                        h->target, h->algo, h->redundancy);
1642
1643                 if (h->s)
1644                         pr_buf(out, "\tpending: blocks %u+%u allocated %u\n",
1645                                h->s->nr_data, h->s->nr_parity,
1646                                bitmap_weight(h->s->blocks_allocated,
1647                                              h->s->nr_data));
1648         }
1649         mutex_unlock(&c->ec_stripe_head_lock);
1650
1651         mutex_lock(&c->ec_stripe_new_lock);
1652         list_for_each_entry(s, &c->ec_stripe_new_list, list) {
1653                 pr_buf(out, "\tin flight: blocks %u+%u pin %u\n",
1654                        s->nr_data, s->nr_parity,
1655                        atomic_read(&s->pin));
1656         }
1657         mutex_unlock(&c->ec_stripe_new_lock);
1658 }
1659
1660 void bch2_fs_ec_exit(struct bch_fs *c)
1661 {
1662         struct ec_stripe_head *h;
1663
1664         while (1) {
1665                 mutex_lock(&c->ec_stripe_head_lock);
1666                 h = list_first_entry_or_null(&c->ec_stripe_head_list,
1667                                              struct ec_stripe_head, list);
1668                 if (h)
1669                         list_del(&h->list);
1670                 mutex_unlock(&c->ec_stripe_head_lock);
1671                 if (!h)
1672                         break;
1673
1674                 BUG_ON(h->s);
1675                 kfree(h);
1676         }
1677
1678         BUG_ON(!list_empty(&c->ec_stripe_new_list));
1679
1680         free_heap(&c->ec_stripes_heap);
1681         genradix_free(&c->stripes);
1682         bioset_exit(&c->ec_bioset);
1683 }
1684
1685 void bch2_fs_ec_init_early(struct bch_fs *c)
1686 {
1687         INIT_WORK(&c->ec_stripe_create_work, ec_stripe_create_work);
1688         INIT_WORK(&c->ec_stripe_delete_work, ec_stripe_delete_work);
1689 }
1690
1691 int bch2_fs_ec_init(struct bch_fs *c)
1692 {
1693         return bioset_init(&c->ec_bioset, 1, offsetof(struct ec_bio, bio),
1694                            BIOSET_NEED_BVECS);
1695 }