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                 prt_printf(err, "stripe at POS_MIN");
 112                 return -EINVAL;
 113         }
 114
 115         if (k.k->p.inode) {
 116                 prt_printf(err, "nonzero inode field");
 117                 return -EINVAL;
 118         }
 119
 120         if (bkey_val_bytes(k.k) < sizeof(*s)) {
 121                 prt_printf(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                 prt_printf(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         prt_printf(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                 prt_printf(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
 576         if (!__ec_stripe_mem_alloc(trans->c, idx, GFP_NOWAIT|__GFP_NOWARN))
 577                 return 0;
 578
 579         bch2_trans_unlock(trans);
 580
 581         return   __ec_stripe_mem_alloc(trans->c, idx, GFP_KERNEL) ?:
 582                 bch2_trans_relock(trans);
 583 }
 584
 585 static ssize_t stripe_idx_to_delete(struct bch_fs *c)
 586 {
 587         ec_stripes_heap *h = &c->ec_stripes_heap;
 588
 589         return h->used && h->data[0].blocks_nonempty == 0
 590                 ? h->data[0].idx : -1;
 591 }
 592
 593 static inline int ec_stripes_heap_cmp(ec_stripes_heap *h,
 594                                       struct ec_stripe_heap_entry l,
 595                                       struct ec_stripe_heap_entry r)
 596 {
 597         return ((l.blocks_nonempty > r.blocks_nonempty) -
 598                 (l.blocks_nonempty < r.blocks_nonempty));
 599 }
 600
 601 static inline void ec_stripes_heap_set_backpointer(ec_stripes_heap *h,
 602                                                    size_t i)
 603 {
 604         struct bch_fs *c = container_of(h, struct bch_fs, ec_stripes_heap);
 605
 606         genradix_ptr(&c->stripes, h->data[i].idx)->heap_idx = i;
 607 }
 608
 609 static void heap_verify_backpointer(struct bch_fs *c, size_t idx)
 610 {
 611         ec_stripes_heap *h = &c->ec_stripes_heap;
 612         struct stripe *m = genradix_ptr(&c->stripes, idx);
 613
 614         BUG_ON(!m->alive);
 615         BUG_ON(m->heap_idx >= h->used);
 616         BUG_ON(h->data[m->heap_idx].idx != idx);
 617 }
 618
 619 void bch2_stripes_heap_del(struct bch_fs *c,
 620                            struct stripe *m, size_t idx)
 621 {
 622         if (!m->on_heap)
 623                 return;
 624
 625         m->on_heap = false;
 626
 627         heap_verify_backpointer(c, idx);
 628
 629         heap_del(&c->ec_stripes_heap, m->heap_idx,
 630                  ec_stripes_heap_cmp,
 631                  ec_stripes_heap_set_backpointer);
 632 }
 633
 634 void bch2_stripes_heap_insert(struct bch_fs *c,
 635                               struct stripe *m, size_t idx)
 636 {
 637         if (m->on_heap)
 638                 return;
 639
 640         BUG_ON(heap_full(&c->ec_stripes_heap));
 641
 642         m->on_heap = true;
 643
 644         heap_add(&c->ec_stripes_heap, ((struct ec_stripe_heap_entry) {
 645                         .idx = idx,
 646                         .blocks_nonempty = m->blocks_nonempty,
 647                 }),
 648                  ec_stripes_heap_cmp,
 649                  ec_stripes_heap_set_backpointer);
 650
 651         heap_verify_backpointer(c, idx);
 652 }
 653
 654 void bch2_stripes_heap_update(struct bch_fs *c,
 655                               struct stripe *m, size_t idx)
 656 {
 657         ec_stripes_heap *h = &c->ec_stripes_heap;
 658         size_t i;
 659
 660         if (!m->on_heap)
 661                 return;
 662
 663         heap_verify_backpointer(c, idx);
 664
 665         h->data[m->heap_idx].blocks_nonempty = m->blocks_nonempty;
 666
 667         i = m->heap_idx;
 668         heap_sift_up(h,   i, ec_stripes_heap_cmp,
 669                      ec_stripes_heap_set_backpointer);
 670         heap_sift_down(h, i, ec_stripes_heap_cmp,
 671                        ec_stripes_heap_set_backpointer);
 672
 673         heap_verify_backpointer(c, idx);
 674
 675         if (stripe_idx_to_delete(c) >= 0 &&
 676             !percpu_ref_is_dying(&c->writes))
 677                 schedule_work(&c->ec_stripe_delete_work);
 678 }
 679
 680 /* stripe deletion */
 681
 682 static int ec_stripe_delete(struct bch_fs *c, size_t idx)
 683 {
 684         return bch2_btree_delete_range(c, BTREE_ID_stripes,
 685                                        POS(0, idx),
 686                                        POS(0, idx + 1),
 687                                        0, NULL);
 688 }
 689
 690 static void ec_stripe_delete_work(struct work_struct *work)
 691 {
 692         struct bch_fs *c =
 693                 container_of(work, struct bch_fs, ec_stripe_delete_work);
 694         ssize_t idx;
 695
 696         while (1) {
 697                 spin_lock(&c->ec_stripes_heap_lock);
 698                 idx = stripe_idx_to_delete(c);
 699                 if (idx < 0) {
 700                         spin_unlock(&c->ec_stripes_heap_lock);
 701                         break;
 702                 }
 703
 704                 bch2_stripes_heap_del(c, genradix_ptr(&c->stripes, idx), idx);
 705                 spin_unlock(&c->ec_stripes_heap_lock);
 706
 707                 if (ec_stripe_delete(c, idx))
 708                         break;
 709         }
 710 }
 711
 712 /* stripe creation: */
 713
 714 static int ec_stripe_bkey_insert(struct btree_trans *trans,
 715                                  struct bkey_i_stripe *stripe,
 716                                  struct disk_reservation *res)
 717 {
 718         struct bch_fs *c = trans->c;
 719         struct btree_iter iter;
 720         struct bkey_s_c k;
 721         struct bpos min_pos = POS(0, 1);
 722         struct bpos start_pos = bpos_max(min_pos, POS(0, c->ec_stripe_hint));
 723         int ret;
 724
 725         for_each_btree_key_norestart(trans, iter, BTREE_ID_stripes, start_pos,
 726                            BTREE_ITER_SLOTS|BTREE_ITER_INTENT, k, ret) {
 727                 if (bkey_cmp(k.k->p, POS(0, U32_MAX)) > 0) {
 728                         if (start_pos.offset) {
 729                                 start_pos = min_pos;
 730                                 bch2_btree_iter_set_pos(&iter, start_pos);
 731                                 continue;
 732                         }
 733
 734                         ret = -ENOSPC;
 735                         break;
 736                 }
 737
 738                 if (bkey_deleted(k.k))
 739                         break;
 740         }
 741
 742         c->ec_stripe_hint = iter.pos.offset;
 743
 744         if (ret)
 745                 goto err;
 746
 747         ret = ec_stripe_mem_alloc(trans, &iter);
 748         if (ret)
 749                 goto err;
 750
 751         stripe->k.p = iter.pos;
 752
 753         ret = bch2_trans_update(trans, &iter, &stripe->k_i, 0);
 754 err:
 755         bch2_trans_iter_exit(trans, &iter);
 756
 757         return ret;
 758 }
 759
 760 static int ec_stripe_bkey_update(struct btree_trans *trans,
 761                                  struct bkey_i_stripe *new,
 762                                  struct disk_reservation *res)
 763 {
 764         struct btree_iter iter;
 765         struct bkey_s_c k;
 766         const struct bch_stripe *existing;
 767         unsigned i;
 768         int ret;
 769
 770         bch2_trans_iter_init(trans, &iter, BTREE_ID_stripes,
 771                              new->k.p, BTREE_ITER_INTENT);
 772         k = bch2_btree_iter_peek_slot(&iter);
 773         ret = bkey_err(k);
 774         if (ret)
 775                 goto err;
 776
 777         if (!k.k || k.k->type != KEY_TYPE_stripe) {
 778                 bch_err(trans->c, "error updating stripe: not found");
 779                 ret = -ENOENT;
 780                 goto err;
 781         }
 782
 783         existing = bkey_s_c_to_stripe(k).v;
 784
 785         if (existing->nr_blocks != new->v.nr_blocks) {
 786                 bch_err(trans->c, "error updating stripe: nr_blocks does not match");
 787                 ret = -EINVAL;
 788                 goto err;
 789         }
 790
 791         for (i = 0; i < new->v.nr_blocks; i++)
 792                 stripe_blockcount_set(&new->v, i,
 793                         stripe_blockcount_get(existing, i));
 794
 795         ret = bch2_trans_update(trans, &iter, &new->k_i, 0);
 796 err:
 797         bch2_trans_iter_exit(trans, &iter);
 798         return ret;
 799 }
 800
 801 static void extent_stripe_ptr_add(struct bkey_s_extent e,
 802                                   struct ec_stripe_buf *s,
 803                                   struct bch_extent_ptr *ptr,
 804                                   unsigned block)
 805 {
 806         struct bch_extent_stripe_ptr *dst = (void *) ptr;
 807         union bch_extent_entry *end = extent_entry_last(e);
 808
 809         memmove_u64s_up(dst + 1, dst, (u64 *) end - (u64 *) dst);
 810         e.k->u64s += sizeof(*dst) / sizeof(u64);
 811
 812         *dst = (struct bch_extent_stripe_ptr) {
 813                 .type = 1 << BCH_EXTENT_ENTRY_stripe_ptr,
 814                 .block          = block,
 815                 .redundancy     = s->key.v.nr_redundant,
 816                 .idx            = s->key.k.p.offset,
 817         };
 818 }
 819
 820 static int ec_stripe_update_extent(struct btree_trans *trans,
 821                                    struct btree_iter *iter,
 822                                    struct bkey_s_c k,
 823                                    struct ec_stripe_buf *s,
 824                                    struct bpos end)
 825 {
 826         const struct bch_extent_ptr *ptr_c;
 827         struct bch_extent_ptr *ptr, *ec_ptr = NULL;
 828         struct bkey_i *n;
 829         int ret, dev, block;
 830
 831         if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
 832                 return 1;
 833
 834         if (extent_has_stripe_ptr(k, s->key.k.p.offset))
 835                 return 0;
 836
 837         ptr_c = bkey_matches_stripe(&s->key.v, k, &block);
 838         /*
 839          * It doesn't generally make sense to erasure code cached ptrs:
 840          * XXX: should we be incrementing a counter?
 841          */
 842         if (!ptr_c || ptr_c->cached)
 843                 return 0;
 844
 845         dev = s->key.v.ptrs[block].dev;
 846
 847         n = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 848         ret = PTR_ERR_OR_ZERO(n);
 849         if (ret)
 850                 return ret;
 851
 852         bkey_reassemble(n, k);
 853
 854         bch2_bkey_drop_ptrs(bkey_i_to_s(n), ptr, ptr->dev != dev);
 855         ec_ptr = (void *) bch2_bkey_has_device(bkey_i_to_s_c(n), dev);
 856         BUG_ON(!ec_ptr);
 857
 858         extent_stripe_ptr_add(bkey_i_to_s_extent(n), s, ec_ptr, block);
 859
 860         return bch2_trans_update(trans, iter, n, 0);
 861 }
 862
 863 static int ec_stripe_update_extents(struct bch_fs *c,
 864                                  struct ec_stripe_buf *s,
 865                                  struct bkey *pos)
 866 {
 867         struct btree_iter iter;
 868         struct bkey_s_c k;
 869
 870         return bch2_trans_run(c,
 871                 for_each_btree_key_commit(&trans, iter,
 872                         BTREE_ID_extents, bkey_start_pos(pos),
 873                         BTREE_ITER_NOT_EXTENTS|BTREE_ITER_INTENT, k,
 874                         NULL, NULL, BTREE_INSERT_NOFAIL,
 875                 ec_stripe_update_extent(&trans, &iter, k, s, pos->p)));
 876 }
 877
 878 /*
 879  * data buckets of new stripe all written: create the stripe
 880  */
 881 static void ec_stripe_create(struct ec_stripe_new *s)
 882 {
 883         struct bch_fs *c = s->c;
 884         struct open_bucket *ob;
 885         struct bkey_i *k;
 886         struct stripe *m;
 887         struct bch_stripe *v = &s->new_stripe.key.v;
 888         unsigned i, nr_data = v->nr_blocks - v->nr_redundant;
 889         int ret;
 890
 891         BUG_ON(s->h->s == s);
 892
 893         closure_sync(&s->iodone);
 894
 895         if (s->err) {
 896                 if (s->err != -EROFS)
 897                         bch_err(c, "error creating stripe: error writing data buckets");
 898                 goto err;
 899         }
 900
 901         if (s->have_existing_stripe) {
 902                 ec_validate_checksums(c, &s->existing_stripe);
 903
 904                 if (ec_do_recov(c, &s->existing_stripe)) {
 905                         bch_err(c, "error creating stripe: error reading existing stripe");
 906                         goto err;
 907                 }
 908
 909                 for (i = 0; i < nr_data; i++)
 910                         if (stripe_blockcount_get(&s->existing_stripe.key.v, i))
 911                                 swap(s->new_stripe.data[i],
 912                                      s->existing_stripe.data[i]);
 913
 914                 ec_stripe_buf_exit(&s->existing_stripe);
 915         }
 916
 917         BUG_ON(!s->allocated);
 918
 919         if (!percpu_ref_tryget_live(&c->writes))
 920                 goto err;
 921
 922         ec_generate_ec(&s->new_stripe);
 923
 924         ec_generate_checksums(&s->new_stripe);
 925
 926         /* write p/q: */
 927         for (i = nr_data; i < v->nr_blocks; i++)
 928                 ec_block_io(c, &s->new_stripe, REQ_OP_WRITE, i, &s->iodone);
 929         closure_sync(&s->iodone);
 930
 931         if (ec_nr_failed(&s->new_stripe)) {
 932                 bch_err(c, "error creating stripe: error writing redundancy buckets");
 933                 goto err_put_writes;
 934         }
 935
 936         ret = bch2_trans_do(c, &s->res, NULL, BTREE_INSERT_NOFAIL,
 937                             s->have_existing_stripe
 938                             ? ec_stripe_bkey_update(&trans, &s->new_stripe.key, &s->res)
 939                             : ec_stripe_bkey_insert(&trans, &s->new_stripe.key, &s->res));
 940         if (ret) {
 941                 bch_err(c, "error creating stripe: error creating stripe key");
 942                 goto err_put_writes;
 943         }
 944
 945         for_each_keylist_key(&s->keys, k) {
 946                 ret = ec_stripe_update_extents(c, &s->new_stripe, &k->k);
 947                 if (ret) {
 948                         bch_err(c, "error creating stripe: error updating pointers: %s",
 949                                 bch2_err_str(ret));
 950                         break;
 951                 }
 952         }
 953
 954         spin_lock(&c->ec_stripes_heap_lock);
 955         m = genradix_ptr(&c->stripes, s->new_stripe.key.k.p.offset);
 956
 957         BUG_ON(m->on_heap);
 958         bch2_stripes_heap_insert(c, m, s->new_stripe.key.k.p.offset);
 959         spin_unlock(&c->ec_stripes_heap_lock);
 960 err_put_writes:
 961         percpu_ref_put(&c->writes);
 962 err:
 963         bch2_disk_reservation_put(c, &s->res);
 964
 965         for (i = 0; i < v->nr_blocks; i++)
 966                 if (s->blocks[i]) {
 967                         ob = c->open_buckets + s->blocks[i];
 968
 969                         if (i < nr_data) {
 970                                 ob->ec = NULL;
 971                                 __bch2_open_bucket_put(c, ob);
 972                         } else {
 973                                 bch2_open_bucket_put(c, ob);
 974                         }
 975                 }
 976
 977         bch2_keylist_free(&s->keys, s->inline_keys);
 978
 979         ec_stripe_buf_exit(&s->existing_stripe);
 980         ec_stripe_buf_exit(&s->new_stripe);
 981         closure_debug_destroy(&s->iodone);
 982         kfree(s);
 983 }
 984
 985 static void ec_stripe_create_work(struct work_struct *work)
 986 {
 987         struct bch_fs *c = container_of(work,
 988                 struct bch_fs, ec_stripe_create_work);
 989         struct ec_stripe_new *s, *n;
 990 restart:
 991         mutex_lock(&c->ec_stripe_new_lock);
 992         list_for_each_entry_safe(s, n, &c->ec_stripe_new_list, list)
 993                 if (!atomic_read(&s->pin)) {
 994                         list_del(&s->list);
 995                         mutex_unlock(&c->ec_stripe_new_lock);
 996                         ec_stripe_create(s);
 997                         goto restart;
 998                 }
 999         mutex_unlock(&c->ec_stripe_new_lock);
1000 }
1001
1002 static void ec_stripe_new_put(struct bch_fs *c, struct ec_stripe_new *s)
1003 {
1004         BUG_ON(atomic_read(&s->pin) <= 0);
1005
1006         if (atomic_dec_and_test(&s->pin)) {
1007                 BUG_ON(!s->pending);
1008                 queue_work(system_long_wq, &c->ec_stripe_create_work);
1009         }
1010 }
1011
1012 static void ec_stripe_set_pending(struct bch_fs *c, struct ec_stripe_head *h)
1013 {
1014         struct ec_stripe_new *s = h->s;
1015
1016         BUG_ON(!s->allocated && !s->err);
1017
1018         h->s            = NULL;
1019         s->pending      = true;
1020
1021         mutex_lock(&c->ec_stripe_new_lock);
1022         list_add(&s->list, &c->ec_stripe_new_list);
1023         mutex_unlock(&c->ec_stripe_new_lock);
1024
1025         ec_stripe_new_put(c, s);
1026 }
1027
1028 /* have a full bucket - hand it off to be erasure coded: */
1029 void bch2_ec_bucket_written(struct bch_fs *c, struct open_bucket *ob)
1030 {
1031         struct ec_stripe_new *s = ob->ec;
1032
1033         if (ob->sectors_free)
1034                 s->err = -1;
1035
1036         ec_stripe_new_put(c, s);
1037 }
1038
1039 void bch2_ec_bucket_cancel(struct bch_fs *c, struct open_bucket *ob)
1040 {
1041         struct ec_stripe_new *s = ob->ec;
1042
1043         s->err = -EIO;
1044 }
1045
1046 void *bch2_writepoint_ec_buf(struct bch_fs *c, struct write_point *wp)
1047 {
1048         struct open_bucket *ob = ec_open_bucket(c, &wp->ptrs);
1049         struct bch_dev *ca;
1050         unsigned offset;
1051
1052         if (!ob)
1053                 return NULL;
1054
1055         ca      = bch_dev_bkey_exists(c, ob->dev);
1056         offset  = ca->mi.bucket_size - ob->sectors_free;
1057
1058         return ob->ec->new_stripe.data[ob->ec_idx] + (offset << 9);
1059 }
1060
1061 void bch2_ob_add_backpointer(struct bch_fs *c, struct open_bucket *ob,
1062                              struct bkey *k)
1063 {
1064         struct ec_stripe_new *ec = ob->ec;
1065
1066         if (!ec)
1067                 return;
1068
1069         mutex_lock(&ec->lock);
1070
1071         if (bch2_keylist_realloc(&ec->keys, ec->inline_keys,
1072                                  ARRAY_SIZE(ec->inline_keys),
1073                                  BKEY_U64s)) {
1074                 BUG();
1075         }
1076
1077         bkey_init(&ec->keys.top->k);
1078         ec->keys.top->k.p       = k->p;
1079         ec->keys.top->k.size    = k->size;
1080         bch2_keylist_push(&ec->keys);
1081
1082         mutex_unlock(&ec->lock);
1083 }
1084
1085 static int unsigned_cmp(const void *_l, const void *_r)
1086 {
1087         unsigned l = *((const unsigned *) _l);
1088         unsigned r = *((const unsigned *) _r);
1089
1090         return cmp_int(l, r);
1091 }
1092
1093 /* pick most common bucket size: */
1094 static unsigned pick_blocksize(struct bch_fs *c,
1095                                struct bch_devs_mask *devs)
1096 {
1097         struct bch_dev *ca;
1098         unsigned i, nr = 0, sizes[BCH_SB_MEMBERS_MAX];
1099         struct {
1100                 unsigned nr, size;
1101         } cur = { 0, 0 }, best = { 0, 0 };
1102
1103         for_each_member_device_rcu(ca, c, i, devs)
1104                 sizes[nr++] = ca->mi.bucket_size;
1105
1106         sort(sizes, nr, sizeof(unsigned), unsigned_cmp, NULL);
1107
1108         for (i = 0; i < nr; i++) {
1109                 if (sizes[i] != cur.size) {
1110                         if (cur.nr > best.nr)
1111                                 best = cur;
1112
1113                         cur.nr = 0;
1114                         cur.size = sizes[i];
1115                 }
1116
1117                 cur.nr++;
1118         }
1119
1120         if (cur.nr > best.nr)
1121                 best = cur;
1122
1123         return best.size;
1124 }
1125
1126 static bool may_create_new_stripe(struct bch_fs *c)
1127 {
1128         return false;
1129 }
1130
1131 static void ec_stripe_key_init(struct bch_fs *c,
1132                                struct bkey_i_stripe *s,
1133                                unsigned nr_data,
1134                                unsigned nr_parity,
1135                                unsigned stripe_size)
1136 {
1137         unsigned u64s;
1138
1139         bkey_stripe_init(&s->k_i);
1140         s->v.sectors                    = cpu_to_le16(stripe_size);
1141         s->v.algorithm                  = 0;
1142         s->v.nr_blocks                  = nr_data + nr_parity;
1143         s->v.nr_redundant               = nr_parity;
1144         s->v.csum_granularity_bits      = ilog2(c->opts.encoded_extent_max >> 9);
1145         s->v.csum_type                  = BCH_CSUM_crc32c;
1146         s->v.pad                        = 0;
1147
1148         while ((u64s = stripe_val_u64s(&s->v)) > BKEY_VAL_U64s_MAX) {
1149                 BUG_ON(1 << s->v.csum_granularity_bits >=
1150                        le16_to_cpu(s->v.sectors) ||
1151                        s->v.csum_granularity_bits == U8_MAX);
1152                 s->v.csum_granularity_bits++;
1153         }
1154
1155         set_bkey_val_u64s(&s->k, u64s);
1156 }
1157
1158 static int ec_new_stripe_alloc(struct bch_fs *c, struct ec_stripe_head *h)
1159 {
1160         struct ec_stripe_new *s;
1161
1162         lockdep_assert_held(&h->lock);
1163
1164         s = kzalloc(sizeof(*s), GFP_KERNEL);
1165         if (!s)
1166                 return -ENOMEM;
1167
1168         mutex_init(&s->lock);
1169         closure_init(&s->iodone, NULL);
1170         atomic_set(&s->pin, 1);
1171         s->c            = c;
1172         s->h            = h;
1173         s->nr_data      = min_t(unsigned, h->nr_active_devs,
1174                                 BCH_BKEY_PTRS_MAX) - h->redundancy;
1175         s->nr_parity    = h->redundancy;
1176
1177         bch2_keylist_init(&s->keys, s->inline_keys);
1178
1179         ec_stripe_key_init(c, &s->new_stripe.key, s->nr_data,
1180                            s->nr_parity, h->blocksize);
1181
1182         h->s = s;
1183         return 0;
1184 }
1185
1186 static struct ec_stripe_head *
1187 ec_new_stripe_head_alloc(struct bch_fs *c, unsigned target,
1188                          unsigned algo, unsigned redundancy,
1189                          bool copygc)
1190 {
1191         struct ec_stripe_head *h;
1192         struct bch_dev *ca;
1193         unsigned i;
1194
1195         h = kzalloc(sizeof(*h), GFP_KERNEL);
1196         if (!h)
1197                 return NULL;
1198
1199         mutex_init(&h->lock);
1200         mutex_lock(&h->lock);
1201
1202         h->target       = target;
1203         h->algo         = algo;
1204         h->redundancy   = redundancy;
1205         h->copygc       = copygc;
1206
1207         rcu_read_lock();
1208         h->devs = target_rw_devs(c, BCH_DATA_user, target);
1209
1210         for_each_member_device_rcu(ca, c, i, &h->devs)
1211                 if (!ca->mi.durability)
1212                         __clear_bit(i, h->devs.d);
1213
1214         h->blocksize = pick_blocksize(c, &h->devs);
1215
1216         for_each_member_device_rcu(ca, c, i, &h->devs)
1217                 if (ca->mi.bucket_size == h->blocksize)
1218                         h->nr_active_devs++;
1219
1220         rcu_read_unlock();
1221         list_add(&h->list, &c->ec_stripe_head_list);
1222         return h;
1223 }
1224
1225 void bch2_ec_stripe_head_put(struct bch_fs *c, struct ec_stripe_head *h)
1226 {
1227         if (h->s &&
1228             h->s->allocated &&
1229             bitmap_weight(h->s->blocks_allocated,
1230                           h->s->nr_data) == h->s->nr_data)
1231                 ec_stripe_set_pending(c, h);
1232
1233         mutex_unlock(&h->lock);
1234 }
1235
1236 struct ec_stripe_head *__bch2_ec_stripe_head_get(struct bch_fs *c,
1237                                                  unsigned target,
1238                                                  unsigned algo,
1239                                                  unsigned redundancy,
1240                                                  bool copygc)
1241 {
1242         struct ec_stripe_head *h;
1243
1244         if (!redundancy)
1245                 return NULL;
1246
1247         mutex_lock(&c->ec_stripe_head_lock);
1248         list_for_each_entry(h, &c->ec_stripe_head_list, list)
1249                 if (h->target           == target &&
1250                     h->algo             == algo &&
1251                     h->redundancy       == redundancy &&
1252                     h->copygc           == copygc) {
1253                         mutex_lock(&h->lock);
1254                         goto found;
1255                 }
1256
1257         h = ec_new_stripe_head_alloc(c, target, algo, redundancy, copygc);
1258 found:
1259         mutex_unlock(&c->ec_stripe_head_lock);
1260         return h;
1261 }
1262
1263 static int new_stripe_alloc_buckets(struct bch_fs *c, struct ec_stripe_head *h,
1264                                     struct closure *cl)
1265 {
1266         struct bch_devs_mask devs = h->devs;
1267         struct open_bucket *ob;
1268         struct open_buckets buckets;
1269         unsigned i, j, nr_have_parity = 0, nr_have_data = 0;
1270         bool have_cache = true;
1271         int ret = 0;
1272
1273         for (i = 0; i < h->s->new_stripe.key.v.nr_blocks; i++) {
1274                 if (test_bit(i, h->s->blocks_gotten)) {
1275                         __clear_bit(h->s->new_stripe.key.v.ptrs[i].dev, devs.d);
1276                         if (i < h->s->nr_data)
1277                                 nr_have_data++;
1278                         else
1279                                 nr_have_parity++;
1280                 }
1281         }
1282
1283         BUG_ON(nr_have_data     > h->s->nr_data);
1284         BUG_ON(nr_have_parity   > h->s->nr_parity);
1285
1286         buckets.nr = 0;
1287         if (nr_have_parity < h->s->nr_parity) {
1288                 ret = bch2_bucket_alloc_set(c, &buckets,
1289                                             &h->parity_stripe,
1290                                             &devs,
1291                                             h->s->nr_parity,
1292                                             &nr_have_parity,
1293                                             &have_cache,
1294                                             h->copygc
1295                                             ? RESERVE_movinggc
1296                                             : RESERVE_none,
1297                                             0,
1298                                             cl);
1299
1300                 open_bucket_for_each(c, &buckets, ob, i) {
1301                         j = find_next_zero_bit(h->s->blocks_gotten,
1302                                                h->s->nr_data + h->s->nr_parity,
1303                                                h->s->nr_data);
1304                         BUG_ON(j >= h->s->nr_data + h->s->nr_parity);
1305
1306                         h->s->blocks[j] = buckets.v[i];
1307                         h->s->new_stripe.key.v.ptrs[j] = bch2_ob_ptr(c, ob);
1308                         __set_bit(j, h->s->blocks_gotten);
1309                 }
1310
1311                 if (ret)
1312                         return ret;
1313         }
1314
1315         buckets.nr = 0;
1316         if (nr_have_data < h->s->nr_data) {
1317                 ret = bch2_bucket_alloc_set(c, &buckets,
1318                                             &h->block_stripe,
1319                                             &devs,
1320                                             h->s->nr_data,
1321                                             &nr_have_data,
1322                                             &have_cache,
1323                                             h->copygc
1324                                             ? RESERVE_movinggc
1325                                             : RESERVE_none,
1326                                             0,
1327                                             cl);
1328
1329                 open_bucket_for_each(c, &buckets, ob, i) {
1330                         j = find_next_zero_bit(h->s->blocks_gotten,
1331                                                h->s->nr_data, 0);
1332                         BUG_ON(j >= h->s->nr_data);
1333
1334                         h->s->blocks[j] = buckets.v[i];
1335                         h->s->new_stripe.key.v.ptrs[j] = bch2_ob_ptr(c, ob);
1336                         __set_bit(j, h->s->blocks_gotten);
1337                 }
1338
1339                 if (ret)
1340                         return ret;
1341         }
1342
1343         return 0;
1344 }
1345
1346 /* XXX: doesn't obey target: */
1347 static s64 get_existing_stripe(struct bch_fs *c,
1348                                struct ec_stripe_head *head)
1349 {
1350         ec_stripes_heap *h = &c->ec_stripes_heap;
1351         struct stripe *m;
1352         size_t heap_idx;
1353         u64 stripe_idx;
1354         s64 ret = -1;
1355
1356         if (may_create_new_stripe(c))
1357                 return -1;
1358
1359         spin_lock(&c->ec_stripes_heap_lock);
1360         for (heap_idx = 0; heap_idx < h->used; heap_idx++) {
1361                 /* No blocks worth reusing, stripe will just be deleted: */
1362                 if (!h->data[heap_idx].blocks_nonempty)
1363                         continue;
1364
1365                 stripe_idx = h->data[heap_idx].idx;
1366                 m = genradix_ptr(&c->stripes, stripe_idx);
1367
1368                 if (m->algorithm        == head->algo &&
1369                     m->nr_redundant     == head->redundancy &&
1370                     m->sectors          == head->blocksize &&
1371                     m->blocks_nonempty  < m->nr_blocks - m->nr_redundant) {
1372                         bch2_stripes_heap_del(c, m, stripe_idx);
1373                         ret = stripe_idx;
1374                         break;
1375                 }
1376         }
1377         spin_unlock(&c->ec_stripes_heap_lock);
1378         return ret;
1379 }
1380
1381 static int __bch2_ec_stripe_head_reuse(struct bch_fs *c,
1382                                                    struct ec_stripe_head *h)
1383 {
1384         unsigned i;
1385         s64 idx;
1386         int ret;
1387
1388         idx = get_existing_stripe(c, h);
1389         if (idx < 0) {
1390                 bch_err(c, "failed to find an existing stripe");
1391                 return -ENOSPC;
1392         }
1393
1394         h->s->have_existing_stripe = true;
1395         ret = get_stripe_key(c, idx, &h->s->existing_stripe);
1396         if (ret) {
1397                 bch2_fs_fatal_error(c, "error reading stripe key: %i", ret);
1398                 return ret;
1399         }
1400
1401         if (ec_stripe_buf_init(&h->s->existing_stripe, 0, h->blocksize)) {
1402                 /*
1403                  * this is a problem: we have deleted from the
1404                  * stripes heap already
1405                  */
1406                 BUG();
1407         }
1408
1409         BUG_ON(h->s->existing_stripe.size != h->blocksize);
1410         BUG_ON(h->s->existing_stripe.size != h->s->existing_stripe.key.v.sectors);
1411
1412         for (i = 0; i < h->s->existing_stripe.key.v.nr_blocks; i++) {
1413                 if (stripe_blockcount_get(&h->s->existing_stripe.key.v, i)) {
1414                         __set_bit(i, h->s->blocks_gotten);
1415                         __set_bit(i, h->s->blocks_allocated);
1416                 }
1417
1418                 ec_block_io(c, &h->s->existing_stripe, READ, i, &h->s->iodone);
1419         }
1420
1421         bkey_copy(&h->s->new_stripe.key.k_i,
1422                         &h->s->existing_stripe.key.k_i);
1423
1424         return 0;
1425 }
1426
1427 static int __bch2_ec_stripe_head_reserve(struct bch_fs *c,
1428                                                         struct ec_stripe_head *h)
1429 {
1430         int ret;
1431
1432         ret = bch2_disk_reservation_get(c, &h->s->res,
1433                         h->blocksize,
1434                         h->s->nr_parity, 0);
1435
1436         if (ret) {
1437                 /*
1438                  * This means we need to wait for copygc to
1439                  * empty out buckets from existing stripes:
1440                  */
1441                 bch_err(c, "failed to reserve stripe");
1442         }
1443
1444         return ret;
1445 }
1446
1447 struct ec_stripe_head *bch2_ec_stripe_head_get(struct bch_fs *c,
1448                                                unsigned target,
1449                                                unsigned algo,
1450                                                unsigned redundancy,
1451                                                bool copygc,
1452                                                struct closure *cl)
1453 {
1454         struct ec_stripe_head *h;
1455         int ret;
1456         bool needs_stripe_new;
1457
1458         h = __bch2_ec_stripe_head_get(c, target, algo, redundancy, copygc);
1459         if (!h) {
1460                 bch_err(c, "no stripe head");
1461                 return NULL;
1462         }
1463
1464         needs_stripe_new = !h->s;
1465         if (needs_stripe_new) {
1466                 if (ec_new_stripe_alloc(c, h)) {
1467                         ret = -ENOMEM;
1468                         bch_err(c, "failed to allocate new stripe");
1469                         goto err;
1470                 }
1471
1472                 if (ec_stripe_buf_init(&h->s->new_stripe, 0, h->blocksize))
1473                         BUG();
1474         }
1475
1476         /*
1477          * Try reserve a new stripe before reusing an
1478          * existing stripe. This will prevent unnecessary
1479          * read amplification during write oriented workloads.
1480          */
1481         ret = 0;
1482         if (!h->s->allocated && !h->s->res.sectors && !h->s->have_existing_stripe)
1483                 ret = __bch2_ec_stripe_head_reserve(c, h);
1484         if (ret && needs_stripe_new)
1485                 ret = __bch2_ec_stripe_head_reuse(c, h);
1486         if (ret)
1487                 goto err;
1488
1489         if (!h->s->allocated) {
1490                 ret = new_stripe_alloc_buckets(c, h, cl);
1491                 if (ret)
1492                         goto err;
1493
1494                 h->s->allocated = true;
1495         }
1496
1497         return h;
1498
1499 err:
1500         bch2_ec_stripe_head_put(c, h);
1501         return ERR_PTR(ret);
1502 }
1503
1504 void bch2_ec_stop_dev(struct bch_fs *c, struct bch_dev *ca)
1505 {
1506         struct ec_stripe_head *h;
1507         struct open_bucket *ob;
1508         unsigned i;
1509
1510         mutex_lock(&c->ec_stripe_head_lock);
1511         list_for_each_entry(h, &c->ec_stripe_head_list, list) {
1512
1513                 mutex_lock(&h->lock);
1514                 if (!h->s)
1515                         goto unlock;
1516
1517                 for (i = 0; i < h->s->new_stripe.key.v.nr_blocks; i++) {
1518                         if (!h->s->blocks[i])
1519                                 continue;
1520
1521                         ob = c->open_buckets + h->s->blocks[i];
1522                         if (ob->dev == ca->dev_idx)
1523                                 goto found;
1524                 }
1525                 goto unlock;
1526 found:
1527                 h->s->err = -EROFS;
1528                 ec_stripe_set_pending(c, h);
1529 unlock:
1530                 mutex_unlock(&h->lock);
1531         }
1532         mutex_unlock(&c->ec_stripe_head_lock);
1533 }
1534
1535 void bch2_stripes_heap_start(struct bch_fs *c)
1536 {
1537         struct genradix_iter iter;
1538         struct stripe *m;
1539
1540         genradix_for_each(&c->stripes, iter, m)
1541                 if (m->alive)
1542                         bch2_stripes_heap_insert(c, m, iter.pos);
1543 }
1544
1545 int bch2_stripes_read(struct bch_fs *c)
1546 {
1547         struct btree_trans trans;
1548         struct btree_iter iter;
1549         struct bkey_s_c k;
1550         const struct bch_stripe *s;
1551         struct stripe *m;
1552         unsigned i;
1553         int ret;
1554
1555         bch2_trans_init(&trans, c, 0, 0);
1556
1557         for_each_btree_key(&trans, iter, BTREE_ID_stripes, POS_MIN,
1558                            BTREE_ITER_PREFETCH, k, ret) {
1559                 if (k.k->type != KEY_TYPE_stripe)
1560                         continue;
1561
1562                 ret = __ec_stripe_mem_alloc(c, k.k->p.offset, GFP_KERNEL);
1563                 if (ret)
1564                         break;
1565
1566                 s = bkey_s_c_to_stripe(k).v;
1567
1568                 m = genradix_ptr(&c->stripes, k.k->p.offset);
1569                 m->alive        = true;
1570                 m->sectors      = le16_to_cpu(s->sectors);
1571                 m->algorithm    = s->algorithm;
1572                 m->nr_blocks    = s->nr_blocks;
1573                 m->nr_redundant = s->nr_redundant;
1574                 m->blocks_nonempty = 0;
1575
1576                 for (i = 0; i < s->nr_blocks; i++)
1577                         m->blocks_nonempty += !!stripe_blockcount_get(s, i);
1578
1579                 spin_lock(&c->ec_stripes_heap_lock);
1580                 bch2_stripes_heap_update(c, m, k.k->p.offset);
1581                 spin_unlock(&c->ec_stripes_heap_lock);
1582         }
1583         bch2_trans_iter_exit(&trans, &iter);
1584
1585         bch2_trans_exit(&trans);
1586
1587         if (ret)
1588                 bch_err(c, "error reading stripes: %i", ret);
1589
1590         return ret;
1591 }
1592
1593 void bch2_stripes_heap_to_text(struct printbuf *out, struct bch_fs *c)
1594 {
1595         ec_stripes_heap *h = &c->ec_stripes_heap;
1596         struct stripe *m;
1597         size_t i;
1598
1599         spin_lock(&c->ec_stripes_heap_lock);
1600         for (i = 0; i < min_t(size_t, h->used, 20); i++) {
1601                 m = genradix_ptr(&c->stripes, h->data[i].idx);
1602
1603                 prt_printf(out, "%zu %u/%u+%u\n", h->data[i].idx,
1604                        h->data[i].blocks_nonempty,
1605                        m->nr_blocks - m->nr_redundant,
1606                        m->nr_redundant);
1607         }
1608         spin_unlock(&c->ec_stripes_heap_lock);
1609 }
1610
1611 void bch2_new_stripes_to_text(struct printbuf *out, struct bch_fs *c)
1612 {
1613         struct ec_stripe_head *h;
1614         struct ec_stripe_new *s;
1615
1616         mutex_lock(&c->ec_stripe_head_lock);
1617         list_for_each_entry(h, &c->ec_stripe_head_list, list) {
1618                 prt_printf(out, "target %u algo %u redundancy %u:\n",
1619                        h->target, h->algo, h->redundancy);
1620
1621                 if (h->s)
1622                         prt_printf(out, "\tpending: blocks %u+%u allocated %u\n",
1623                                h->s->nr_data, h->s->nr_parity,
1624                                bitmap_weight(h->s->blocks_allocated,
1625                                              h->s->nr_data));
1626         }
1627         mutex_unlock(&c->ec_stripe_head_lock);
1628
1629         mutex_lock(&c->ec_stripe_new_lock);
1630         list_for_each_entry(s, &c->ec_stripe_new_list, list) {
1631                 prt_printf(out, "\tin flight: blocks %u+%u pin %u\n",
1632                        s->nr_data, s->nr_parity,
1633                        atomic_read(&s->pin));
1634         }
1635         mutex_unlock(&c->ec_stripe_new_lock);
1636 }
1637
1638 void bch2_fs_ec_exit(struct bch_fs *c)
1639 {
1640         struct ec_stripe_head *h;
1641
1642         while (1) {
1643                 mutex_lock(&c->ec_stripe_head_lock);
1644                 h = list_first_entry_or_null(&c->ec_stripe_head_list,
1645                                              struct ec_stripe_head, list);
1646                 if (h)
1647                         list_del(&h->list);
1648                 mutex_unlock(&c->ec_stripe_head_lock);
1649                 if (!h)
1650                         break;
1651
1652                 BUG_ON(h->s);
1653                 kfree(h);
1654         }
1655
1656         BUG_ON(!list_empty(&c->ec_stripe_new_list));
1657
1658         free_heap(&c->ec_stripes_heap);
1659         genradix_free(&c->stripes);
1660         bioset_exit(&c->ec_bioset);
1661 }
1662
1663 void bch2_fs_ec_init_early(struct bch_fs *c)
1664 {
1665         INIT_WORK(&c->ec_stripe_create_work, ec_stripe_create_work);
1666         INIT_WORK(&c->ec_stripe_delete_work, ec_stripe_delete_work);
1667 }
1668
1669 int bch2_fs_ec_init(struct bch_fs *c)
1670 {
1671         return bioset_init(&c->ec_bioset, 1, offsetof(struct ec_bio, bio),
1672                            BIOSET_NEED_BVECS);
1673 }