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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 /* erasure coding */
   4
   5 #include "bcachefs.h"
   6 #include "alloc_foreground.h"
   7 #include "backpointers.h"
   8 #include "bkey_buf.h"
   9 #include "bset.h"
  10 #include "btree_gc.h"
  11 #include "btree_update.h"
  12 #include "btree_write_buffer.h"
  13 #include "buckets.h"
  14 #include "disk_groups.h"
  15 #include "ec.h"
  16 #include "error.h"
  17 #include "io.h"
  18 #include "keylist.h"
  19 #include "recovery.h"
  20 #include "replicas.h"
  21 #include "super-io.h"
  22 #include "util.h"
  23
  24 #include <linux/sort.h>
  25
  26 #ifdef __KERNEL__
  27
  28 #include <linux/raid/pq.h>
  29 #include <linux/raid/xor.h>
  30
  31 static void raid5_recov(unsigned disks, unsigned failed_idx,
  32                         size_t size, void **data)
  33 {
  34         unsigned i = 2, nr;
  35
  36         BUG_ON(failed_idx >= disks);
  37
  38         swap(data[0], data[failed_idx]);
  39         memcpy(data[0], data[1], size);
  40
  41         while (i < disks) {
  42                 nr = min_t(unsigned, disks - i, MAX_XOR_BLOCKS);
  43                 xor_blocks(nr, size, data[0], data + i);
  44                 i += nr;
  45         }
  46
  47         swap(data[0], data[failed_idx]);
  48 }
  49
  50 static void raid_gen(int nd, int np, size_t size, void **v)
  51 {
  52         if (np >= 1)
  53                 raid5_recov(nd + np, nd, size, v);
  54         if (np >= 2)
  55                 raid6_call.gen_syndrome(nd + np, size, v);
  56         BUG_ON(np > 2);
  57 }
  58
  59 static void raid_rec(int nr, int *ir, int nd, int np, size_t size, void **v)
  60 {
  61         switch (nr) {
  62         case 0:
  63                 break;
  64         case 1:
  65                 if (ir[0] < nd + 1)
  66                         raid5_recov(nd + 1, ir[0], size, v);
  67                 else
  68                         raid6_call.gen_syndrome(nd + np, size, v);
  69                 break;
  70         case 2:
  71                 if (ir[1] < nd) {
  72                         /* data+data failure. */
  73                         raid6_2data_recov(nd + np, size, ir[0], ir[1], v);
  74                 } else if (ir[0] < nd) {
  75                         /* data + p/q failure */
  76
  77                         if (ir[1] == nd) /* data + p failure */
  78                                 raid6_datap_recov(nd + np, size, ir[0], v);
  79                         else { /* data + q failure */
  80                                 raid5_recov(nd + 1, ir[0], size, v);
  81                                 raid6_call.gen_syndrome(nd + np, size, v);
  82                         }
  83                 } else {
  84                         raid_gen(nd, np, size, v);
  85                 }
  86                 break;
  87         default:
  88                 BUG();
  89         }
  90 }
  91
  92 #else
  93
  94 #include <raid/raid.h>
  95
  96 #endif
  97
  98 struct ec_bio {
  99         struct bch_dev          *ca;
 100         struct ec_stripe_buf    *buf;
 101         size_t                  idx;
 102         struct bio              bio;
 103 };
 104
 105 /* Stripes btree keys: */
 106
 107 int bch2_stripe_invalid(const struct bch_fs *c, struct bkey_s_c k,
 108                         unsigned flags, struct printbuf *err)
 109 {
 110         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 111
 112         if (bkey_eq(k.k->p, POS_MIN)) {
 113                 prt_printf(err, "stripe at POS_MIN");
 114                 return -BCH_ERR_invalid_bkey;
 115         }
 116
 117         if (k.k->p.inode) {
 118                 prt_printf(err, "nonzero inode field");
 119                 return -BCH_ERR_invalid_bkey;
 120         }
 121
 122         if (bkey_val_bytes(k.k) < sizeof(*s)) {
 123                 prt_printf(err, "incorrect value size (%zu < %zu)",
 124                        bkey_val_bytes(k.k), sizeof(*s));
 125                 return -BCH_ERR_invalid_bkey;
 126         }
 127
 128         if (bkey_val_u64s(k.k) < stripe_val_u64s(s)) {
 129                 prt_printf(err, "incorrect value size (%zu < %u)",
 130                        bkey_val_u64s(k.k), stripe_val_u64s(s));
 131                 return -BCH_ERR_invalid_bkey;
 132         }
 133
 134         return bch2_bkey_ptrs_invalid(c, k, flags, err);
 135 }
 136
 137 void bch2_stripe_to_text(struct printbuf *out, struct bch_fs *c,
 138                          struct bkey_s_c k)
 139 {
 140         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 141         unsigned i;
 142
 143         prt_printf(out, "algo %u sectors %u blocks %u:%u csum %u gran %u",
 144                s->algorithm,
 145                le16_to_cpu(s->sectors),
 146                s->nr_blocks - s->nr_redundant,
 147                s->nr_redundant,
 148                s->csum_type,
 149                1U << s->csum_granularity_bits);
 150
 151         for (i = 0; i < s->nr_blocks; i++)
 152                 prt_printf(out, " %u:%llu:%u", s->ptrs[i].dev,
 153                        (u64) s->ptrs[i].offset,
 154                        stripe_blockcount_get(s, i));
 155 }
 156
 157 /* returns blocknr in stripe that we matched: */
 158 static const struct bch_extent_ptr *bkey_matches_stripe(struct bch_stripe *s,
 159                                                 struct bkey_s_c k, unsigned *block)
 160 {
 161         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 162         const struct bch_extent_ptr *ptr;
 163         unsigned i, nr_data = s->nr_blocks - s->nr_redundant;
 164
 165         bkey_for_each_ptr(ptrs, ptr)
 166                 for (i = 0; i < nr_data; i++)
 167                         if (__bch2_ptr_matches_stripe(&s->ptrs[i], ptr,
 168                                                       le16_to_cpu(s->sectors))) {
 169                                 *block = i;
 170                                 return ptr;
 171                         }
 172
 173         return NULL;
 174 }
 175
 176 static bool extent_has_stripe_ptr(struct bkey_s_c k, u64 idx)
 177 {
 178         switch (k.k->type) {
 179         case KEY_TYPE_extent: {
 180                 struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
 181                 const union bch_extent_entry *entry;
 182
 183                 extent_for_each_entry(e, entry)
 184                         if (extent_entry_type(entry) ==
 185                             BCH_EXTENT_ENTRY_stripe_ptr &&
 186                             entry->stripe_ptr.idx == idx)
 187                                 return true;
 188
 189                 break;
 190         }
 191         }
 192
 193         return false;
 194 }
 195
 196 /* Stripe bufs: */
 197
 198 static void ec_stripe_buf_exit(struct ec_stripe_buf *buf)
 199 {
 200         unsigned i;
 201
 202         for (i = 0; i < buf->key.v.nr_blocks; i++) {
 203                 kvpfree(buf->data[i], buf->size << 9);
 204                 buf->data[i] = NULL;
 205         }
 206 }
 207
 208 static int ec_stripe_buf_init(struct ec_stripe_buf *buf,
 209                                unsigned offset, unsigned size)
 210 {
 211         struct bch_stripe *v = &buf->key.v;
 212         unsigned csum_granularity = 1U << v->csum_granularity_bits;
 213         unsigned end = offset + size;
 214         unsigned i;
 215
 216         BUG_ON(end > le16_to_cpu(v->sectors));
 217
 218         offset  = round_down(offset, csum_granularity);
 219         end     = min_t(unsigned, le16_to_cpu(v->sectors),
 220                         round_up(end, csum_granularity));
 221
 222         buf->offset     = offset;
 223         buf->size       = end - offset;
 224
 225         memset(buf->valid, 0xFF, sizeof(buf->valid));
 226
 227         for (i = 0; i < buf->key.v.nr_blocks; i++) {
 228                 buf->data[i] = kvpmalloc(buf->size << 9, GFP_KERNEL);
 229                 if (!buf->data[i])
 230                         goto err;
 231         }
 232
 233         return 0;
 234 err:
 235         ec_stripe_buf_exit(buf);
 236         return -ENOMEM;
 237 }
 238
 239 /* Checksumming: */
 240
 241 static struct bch_csum ec_block_checksum(struct ec_stripe_buf *buf,
 242                                          unsigned block, unsigned offset)
 243 {
 244         struct bch_stripe *v = &buf->key.v;
 245         unsigned csum_granularity = 1 << v->csum_granularity_bits;
 246         unsigned end = buf->offset + buf->size;
 247         unsigned len = min(csum_granularity, end - offset);
 248
 249         BUG_ON(offset >= end);
 250         BUG_ON(offset <  buf->offset);
 251         BUG_ON(offset & (csum_granularity - 1));
 252         BUG_ON(offset + len != le16_to_cpu(v->sectors) &&
 253                (len & (csum_granularity - 1)));
 254
 255         return bch2_checksum(NULL, v->csum_type,
 256                              null_nonce(),
 257                              buf->data[block] + ((offset - buf->offset) << 9),
 258                              len << 9);
 259 }
 260
 261 static void ec_generate_checksums(struct ec_stripe_buf *buf)
 262 {
 263         struct bch_stripe *v = &buf->key.v;
 264         unsigned i, j, csums_per_device = stripe_csums_per_device(v);
 265
 266         if (!v->csum_type)
 267                 return;
 268
 269         BUG_ON(buf->offset);
 270         BUG_ON(buf->size != le16_to_cpu(v->sectors));
 271
 272         for (i = 0; i < v->nr_blocks; i++)
 273                 for (j = 0; j < csums_per_device; j++)
 274                         stripe_csum_set(v, i, j,
 275                                 ec_block_checksum(buf, i, j << v->csum_granularity_bits));
 276 }
 277
 278 static void ec_validate_checksums(struct bch_fs *c, struct ec_stripe_buf *buf)
 279 {
 280         struct bch_stripe *v = &buf->key.v;
 281         unsigned csum_granularity = 1 << v->csum_granularity_bits;
 282         unsigned i;
 283
 284         if (!v->csum_type)
 285                 return;
 286
 287         for (i = 0; i < v->nr_blocks; i++) {
 288                 unsigned offset = buf->offset;
 289                 unsigned end = buf->offset + buf->size;
 290
 291                 if (!test_bit(i, buf->valid))
 292                         continue;
 293
 294                 while (offset < end) {
 295                         unsigned j = offset >> v->csum_granularity_bits;
 296                         unsigned len = min(csum_granularity, end - offset);
 297                         struct bch_csum want = stripe_csum_get(v, i, j);
 298                         struct bch_csum got = ec_block_checksum(buf, i, offset);
 299
 300                         if (bch2_crc_cmp(want, got)) {
 301                                 struct printbuf buf2 = PRINTBUF;
 302
 303                                 bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(&buf->key.k_i));
 304
 305                                 bch_err_ratelimited(c,
 306                                         "stripe checksum error for %ps at %u:%u: csum type %u, expected %llx got %llx\n%s",
 307                                         (void *) _RET_IP_, i, j, v->csum_type,
 308                                         want.lo, got.lo, buf2.buf);
 309                                 printbuf_exit(&buf2);
 310                                 clear_bit(i, buf->valid);
 311                                 break;
 312                         }
 313
 314                         offset += len;
 315                 }
 316         }
 317 }
 318
 319 /* Erasure coding: */
 320
 321 static void ec_generate_ec(struct ec_stripe_buf *buf)
 322 {
 323         struct bch_stripe *v = &buf->key.v;
 324         unsigned nr_data = v->nr_blocks - v->nr_redundant;
 325         unsigned bytes = le16_to_cpu(v->sectors) << 9;
 326
 327         raid_gen(nr_data, v->nr_redundant, bytes, buf->data);
 328 }
 329
 330 static unsigned ec_nr_failed(struct ec_stripe_buf *buf)
 331 {
 332         return buf->key.v.nr_blocks -
 333                 bitmap_weight(buf->valid, buf->key.v.nr_blocks);
 334 }
 335
 336 static int ec_do_recov(struct bch_fs *c, struct ec_stripe_buf *buf)
 337 {
 338         struct bch_stripe *v = &buf->key.v;
 339         unsigned i, failed[BCH_BKEY_PTRS_MAX], nr_failed = 0;
 340         unsigned nr_data = v->nr_blocks - v->nr_redundant;
 341         unsigned bytes = buf->size << 9;
 342
 343         if (ec_nr_failed(buf) > v->nr_redundant) {
 344                 bch_err_ratelimited(c,
 345                         "error doing reconstruct read: unable to read enough blocks");
 346                 return -1;
 347         }
 348
 349         for (i = 0; i < nr_data; i++)
 350                 if (!test_bit(i, buf->valid))
 351                         failed[nr_failed++] = i;
 352
 353         raid_rec(nr_failed, failed, nr_data, v->nr_redundant, bytes, buf->data);
 354         return 0;
 355 }
 356
 357 /* IO: */
 358
 359 static void ec_block_endio(struct bio *bio)
 360 {
 361         struct ec_bio *ec_bio = container_of(bio, struct ec_bio, bio);
 362         struct bch_stripe *v = &ec_bio->buf->key.v;
 363         struct bch_extent_ptr *ptr = &v->ptrs[ec_bio->idx];
 364         struct bch_dev *ca = ec_bio->ca;
 365         struct closure *cl = bio->bi_private;
 366
 367         if (bch2_dev_io_err_on(bio->bi_status, ca, "erasure coding %s error: %s",
 368                                bio_data_dir(bio) ? "write" : "read",
 369                                bch2_blk_status_to_str(bio->bi_status)))
 370                 clear_bit(ec_bio->idx, ec_bio->buf->valid);
 371
 372         if (ptr_stale(ca, ptr)) {
 373                 bch_err_ratelimited(ca->fs,
 374                                     "error %s stripe: stale pointer after io",
 375                                     bio_data_dir(bio) == READ ? "reading from" : "writing to");
 376                 clear_bit(ec_bio->idx, ec_bio->buf->valid);
 377         }
 378
 379         bio_put(&ec_bio->bio);
 380         percpu_ref_put(&ca->io_ref);
 381         closure_put(cl);
 382 }
 383
 384 static void ec_block_io(struct bch_fs *c, struct ec_stripe_buf *buf,
 385                         unsigned rw, unsigned idx, struct closure *cl)
 386 {
 387         struct bch_stripe *v = &buf->key.v;
 388         unsigned offset = 0, bytes = buf->size << 9;
 389         struct bch_extent_ptr *ptr = &v->ptrs[idx];
 390         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 391         enum bch_data_type data_type = idx < buf->key.v.nr_blocks - buf->key.v.nr_redundant
 392                 ? BCH_DATA_user
 393                 : BCH_DATA_parity;
 394
 395         if (ptr_stale(ca, ptr)) {
 396                 bch_err_ratelimited(c,
 397                                     "error %s stripe: stale pointer",
 398                                     rw == READ ? "reading from" : "writing to");
 399                 clear_bit(idx, buf->valid);
 400                 return;
 401         }
 402
 403         if (!bch2_dev_get_ioref(ca, rw)) {
 404                 clear_bit(idx, buf->valid);
 405                 return;
 406         }
 407
 408         this_cpu_add(ca->io_done->sectors[rw][data_type], buf->size);
 409
 410         while (offset < bytes) {
 411                 unsigned nr_iovecs = min_t(size_t, BIO_MAX_VECS,
 412                                            DIV_ROUND_UP(bytes, PAGE_SIZE));
 413                 unsigned b = min_t(size_t, bytes - offset,
 414                                    nr_iovecs << PAGE_SHIFT);
 415                 struct ec_bio *ec_bio;
 416
 417                 ec_bio = container_of(bio_alloc_bioset(ca->disk_sb.bdev,
 418                                                        nr_iovecs,
 419                                                        rw,
 420                                                        GFP_KERNEL,
 421                                                        &c->ec_bioset),
 422                                       struct ec_bio, bio);
 423
 424                 ec_bio->ca                      = ca;
 425                 ec_bio->buf                     = buf;
 426                 ec_bio->idx                     = idx;
 427
 428                 ec_bio->bio.bi_iter.bi_sector   = ptr->offset + buf->offset + (offset >> 9);
 429                 ec_bio->bio.bi_end_io           = ec_block_endio;
 430                 ec_bio->bio.bi_private          = cl;
 431
 432                 bch2_bio_map(&ec_bio->bio, buf->data[idx] + offset, b);
 433
 434                 closure_get(cl);
 435                 percpu_ref_get(&ca->io_ref);
 436
 437                 submit_bio(&ec_bio->bio);
 438
 439                 offset += b;
 440         }
 441
 442         percpu_ref_put(&ca->io_ref);
 443 }
 444
 445 static int get_stripe_key(struct bch_fs *c, u64 idx, struct ec_stripe_buf *stripe)
 446 {
 447         struct btree_trans trans;
 448         struct btree_iter iter;
 449         struct bkey_s_c k;
 450         int ret;
 451
 452         bch2_trans_init(&trans, c, 0, 0);
 453         bch2_trans_iter_init(&trans, &iter, BTREE_ID_stripes,
 454                              POS(0, idx), BTREE_ITER_SLOTS);
 455         k = bch2_btree_iter_peek_slot(&iter);
 456         ret = bkey_err(k);
 457         if (ret)
 458                 goto err;
 459         if (k.k->type != KEY_TYPE_stripe) {
 460                 ret = -ENOENT;
 461                 goto err;
 462         }
 463         bkey_reassemble(&stripe->key.k_i, k);
 464 err:
 465         bch2_trans_iter_exit(&trans, &iter);
 466         bch2_trans_exit(&trans);
 467         return ret;
 468 }
 469
 470 /* recovery read path: */
 471 int bch2_ec_read_extent(struct bch_fs *c, struct bch_read_bio *rbio)
 472 {
 473         struct ec_stripe_buf *buf;
 474         struct closure cl;
 475         struct bch_stripe *v;
 476         unsigned i, offset;
 477         int ret = 0;
 478
 479         closure_init_stack(&cl);
 480
 481         BUG_ON(!rbio->pick.has_ec);
 482
 483         buf = kzalloc(sizeof(*buf), GFP_NOIO);
 484         if (!buf)
 485                 return -ENOMEM;
 486
 487         ret = get_stripe_key(c, rbio->pick.ec.idx, buf);
 488         if (ret) {
 489                 bch_err_ratelimited(c,
 490                         "error doing reconstruct read: error %i looking up stripe", ret);
 491                 kfree(buf);
 492                 return -EIO;
 493         }
 494
 495         v = &buf->key.v;
 496
 497         if (!bch2_ptr_matches_stripe(v, rbio->pick)) {
 498                 bch_err_ratelimited(c,
 499                         "error doing reconstruct read: pointer doesn't match stripe");
 500                 ret = -EIO;
 501                 goto err;
 502         }
 503
 504         offset = rbio->bio.bi_iter.bi_sector - v->ptrs[rbio->pick.ec.block].offset;
 505         if (offset + bio_sectors(&rbio->bio) > le16_to_cpu(v->sectors)) {
 506                 bch_err_ratelimited(c,
 507                         "error doing reconstruct read: read is bigger than stripe");
 508                 ret = -EIO;
 509                 goto err;
 510         }
 511
 512         ret = ec_stripe_buf_init(buf, offset, bio_sectors(&rbio->bio));
 513         if (ret)
 514                 goto err;
 515
 516         for (i = 0; i < v->nr_blocks; i++)
 517                 ec_block_io(c, buf, REQ_OP_READ, i, &cl);
 518
 519         closure_sync(&cl);
 520
 521         if (ec_nr_failed(buf) > v->nr_redundant) {
 522                 bch_err_ratelimited(c,
 523                         "error doing reconstruct read: unable to read enough blocks");
 524                 ret = -EIO;
 525                 goto err;
 526         }
 527
 528         ec_validate_checksums(c, buf);
 529
 530         ret = ec_do_recov(c, buf);
 531         if (ret)
 532                 goto err;
 533
 534         memcpy_to_bio(&rbio->bio, rbio->bio.bi_iter,
 535                       buf->data[rbio->pick.ec.block] + ((offset - buf->offset) << 9));
 536 err:
 537         ec_stripe_buf_exit(buf);
 538         kfree(buf);
 539         return ret;
 540 }
 541
 542 /* stripe bucket accounting: */
 543
 544 static int __ec_stripe_mem_alloc(struct bch_fs *c, size_t idx, gfp_t gfp)
 545 {
 546         ec_stripes_heap n, *h = &c->ec_stripes_heap;
 547
 548         if (idx >= h->size) {
 549                 if (!init_heap(&n, max(1024UL, roundup_pow_of_two(idx + 1)), gfp))
 550                         return -ENOMEM;
 551
 552                 spin_lock(&c->ec_stripes_heap_lock);
 553                 if (n.size > h->size) {
 554                         memcpy(n.data, h->data, h->used * sizeof(h->data[0]));
 555                         n.used = h->used;
 556                         swap(*h, n);
 557                 }
 558                 spin_unlock(&c->ec_stripes_heap_lock);
 559
 560                 free_heap(&n);
 561         }
 562
 563         if (!genradix_ptr_alloc(&c->stripes, idx, gfp))
 564                 return -ENOMEM;
 565
 566         if (c->gc_pos.phase != GC_PHASE_NOT_RUNNING &&
 567             !genradix_ptr_alloc(&c->gc_stripes, idx, gfp))
 568                 return -ENOMEM;
 569
 570         return 0;
 571 }
 572
 573 static int ec_stripe_mem_alloc(struct btree_trans *trans,
 574                                struct btree_iter *iter)
 575 {
 576         size_t idx = iter->pos.offset;
 577
 578         if (!__ec_stripe_mem_alloc(trans->c, idx, GFP_NOWAIT|__GFP_NOWARN))
 579                 return 0;
 580
 581         bch2_trans_unlock(trans);
 582
 583         return   __ec_stripe_mem_alloc(trans->c, idx, GFP_KERNEL) ?:
 584                 bch2_trans_relock(trans);
 585 }
 586
 587 static ssize_t stripe_idx_to_delete(struct bch_fs *c)
 588 {
 589         ec_stripes_heap *h = &c->ec_stripes_heap;
 590
 591         return h->used && h->data[0].blocks_nonempty == 0
 592                 ? h->data[0].idx : -1;
 593 }
 594
 595 static inline int ec_stripes_heap_cmp(ec_stripes_heap *h,
 596                                       struct ec_stripe_heap_entry l,
 597                                       struct ec_stripe_heap_entry r)
 598 {
 599         return ((l.blocks_nonempty > r.blocks_nonempty) -
 600                 (l.blocks_nonempty < r.blocks_nonempty));
 601 }
 602
 603 static inline void ec_stripes_heap_set_backpointer(ec_stripes_heap *h,
 604                                                    size_t i)
 605 {
 606         struct bch_fs *c = container_of(h, struct bch_fs, ec_stripes_heap);
 607
 608         genradix_ptr(&c->stripes, h->data[i].idx)->heap_idx = i;
 609 }
 610
 611 static void heap_verify_backpointer(struct bch_fs *c, size_t idx)
 612 {
 613         ec_stripes_heap *h = &c->ec_stripes_heap;
 614         struct stripe *m = genradix_ptr(&c->stripes, idx);
 615
 616         BUG_ON(!m->alive);
 617         BUG_ON(m->heap_idx >= h->used);
 618         BUG_ON(h->data[m->heap_idx].idx != idx);
 619 }
 620
 621 void bch2_stripes_heap_del(struct bch_fs *c,
 622                            struct stripe *m, size_t idx)
 623 {
 624         if (!m->on_heap)
 625                 return;
 626
 627         m->on_heap = false;
 628
 629         heap_verify_backpointer(c, idx);
 630
 631         heap_del(&c->ec_stripes_heap, m->heap_idx,
 632                  ec_stripes_heap_cmp,
 633                  ec_stripes_heap_set_backpointer);
 634 }
 635
 636 void bch2_stripes_heap_insert(struct bch_fs *c,
 637                               struct stripe *m, size_t idx)
 638 {
 639         if (m->on_heap)
 640                 return;
 641
 642         BUG_ON(heap_full(&c->ec_stripes_heap));
 643
 644         m->on_heap = true;
 645
 646         heap_add(&c->ec_stripes_heap, ((struct ec_stripe_heap_entry) {
 647                         .idx = idx,
 648                         .blocks_nonempty = m->blocks_nonempty,
 649                 }),
 650                  ec_stripes_heap_cmp,
 651                  ec_stripes_heap_set_backpointer);
 652
 653         heap_verify_backpointer(c, idx);
 654 }
 655
 656 void bch2_stripes_heap_update(struct bch_fs *c,
 657                               struct stripe *m, size_t idx)
 658 {
 659         ec_stripes_heap *h = &c->ec_stripes_heap;
 660         size_t i;
 661
 662         if (!m->on_heap)
 663                 return;
 664
 665         heap_verify_backpointer(c, idx);
 666
 667         h->data[m->heap_idx].blocks_nonempty = m->blocks_nonempty;
 668
 669         i = m->heap_idx;
 670         heap_sift_up(h,   i, ec_stripes_heap_cmp,
 671                      ec_stripes_heap_set_backpointer);
 672         heap_sift_down(h, i, ec_stripes_heap_cmp,
 673                        ec_stripes_heap_set_backpointer);
 674
 675         heap_verify_backpointer(c, idx);
 676
 677         if (stripe_idx_to_delete(c) >= 0)
 678                 bch2_do_stripe_deletes(c);
 679 }
 680
 681 /* stripe deletion */
 682
 683 static int ec_stripe_delete(struct bch_fs *c, size_t idx)
 684 {
 685         return bch2_btree_delete_range(c, BTREE_ID_stripes,
 686                                        POS(0, idx),
 687                                        POS(0, idx),
 688                                        0, NULL);
 689 }
 690
 691 static void ec_stripe_delete_work(struct work_struct *work)
 692 {
 693         struct bch_fs *c =
 694                 container_of(work, struct bch_fs, ec_stripe_delete_work);
 695         ssize_t idx;
 696
 697         while (1) {
 698                 spin_lock(&c->ec_stripes_heap_lock);
 699                 idx = stripe_idx_to_delete(c);
 700                 if (idx < 0) {
 701                         spin_unlock(&c->ec_stripes_heap_lock);
 702                         break;
 703                 }
 704
 705                 bch2_stripes_heap_del(c, genradix_ptr(&c->stripes, idx), idx);
 706                 spin_unlock(&c->ec_stripes_heap_lock);
 707
 708                 if (ec_stripe_delete(c, idx))
 709                         break;
 710         }
 711
 712         bch2_write_ref_put(c, BCH_WRITE_REF_stripe_delete);
 713 }
 714
 715 void bch2_do_stripe_deletes(struct bch_fs *c)
 716 {
 717         if (bch2_write_ref_tryget(c, BCH_WRITE_REF_stripe_delete) &&
 718             !schedule_work(&c->ec_stripe_delete_work))
 719                 bch2_write_ref_put(c, BCH_WRITE_REF_stripe_delete);
 720 }
 721
 722 /* stripe creation: */
 723
 724 static int ec_stripe_bkey_insert(struct btree_trans *trans,
 725                                  struct bkey_i_stripe *stripe,
 726                                  struct disk_reservation *res)
 727 {
 728         struct bch_fs *c = trans->c;
 729         struct btree_iter iter;
 730         struct bkey_s_c k;
 731         struct bpos min_pos = POS(0, 1);
 732         struct bpos start_pos = bpos_max(min_pos, POS(0, c->ec_stripe_hint));
 733         int ret;
 734
 735         for_each_btree_key_norestart(trans, iter, BTREE_ID_stripes, start_pos,
 736                            BTREE_ITER_SLOTS|BTREE_ITER_INTENT, k, ret) {
 737                 if (bkey_gt(k.k->p, POS(0, U32_MAX))) {
 738                         if (start_pos.offset) {
 739                                 start_pos = min_pos;
 740                                 bch2_btree_iter_set_pos(&iter, start_pos);
 741                                 continue;
 742                         }
 743
 744                         ret = -BCH_ERR_ENOSPC_stripe_create;
 745                         break;
 746                 }
 747
 748                 if (bkey_deleted(k.k))
 749                         break;
 750         }
 751
 752         c->ec_stripe_hint = iter.pos.offset;
 753
 754         if (ret)
 755                 goto err;
 756
 757         ret = ec_stripe_mem_alloc(trans, &iter);
 758         if (ret)
 759                 goto err;
 760
 761         stripe->k.p = iter.pos;
 762
 763         ret = bch2_trans_update(trans, &iter, &stripe->k_i, 0);
 764 err:
 765         bch2_trans_iter_exit(trans, &iter);
 766
 767         return ret;
 768 }
 769
 770 static int ec_stripe_bkey_update(struct btree_trans *trans,
 771                                  struct bkey_i_stripe *new,
 772                                  struct disk_reservation *res)
 773 {
 774         struct btree_iter iter;
 775         struct bkey_s_c k;
 776         const struct bch_stripe *existing;
 777         unsigned i;
 778         int ret;
 779
 780         bch2_trans_iter_init(trans, &iter, BTREE_ID_stripes,
 781                              new->k.p, BTREE_ITER_INTENT);
 782         k = bch2_btree_iter_peek_slot(&iter);
 783         ret = bkey_err(k);
 784         if (ret)
 785                 goto err;
 786
 787         if (!k.k || k.k->type != KEY_TYPE_stripe) {
 788                 bch_err(trans->c, "error updating stripe: not found");
 789                 ret = -ENOENT;
 790                 goto err;
 791         }
 792
 793         existing = bkey_s_c_to_stripe(k).v;
 794
 795         if (existing->nr_blocks != new->v.nr_blocks) {
 796                 bch_err(trans->c, "error updating stripe: nr_blocks does not match");
 797                 ret = -EINVAL;
 798                 goto err;
 799         }
 800
 801         for (i = 0; i < new->v.nr_blocks; i++)
 802                 stripe_blockcount_set(&new->v, i,
 803                         stripe_blockcount_get(existing, i));
 804
 805         ret = bch2_trans_update(trans, &iter, &new->k_i, 0);
 806 err:
 807         bch2_trans_iter_exit(trans, &iter);
 808         return ret;
 809 }
 810
 811 static void extent_stripe_ptr_add(struct bkey_s_extent e,
 812                                   struct ec_stripe_buf *s,
 813                                   struct bch_extent_ptr *ptr,
 814                                   unsigned block)
 815 {
 816         struct bch_extent_stripe_ptr *dst = (void *) ptr;
 817         union bch_extent_entry *end = extent_entry_last(e);
 818
 819         memmove_u64s_up(dst + 1, dst, (u64 *) end - (u64 *) dst);
 820         e.k->u64s += sizeof(*dst) / sizeof(u64);
 821
 822         *dst = (struct bch_extent_stripe_ptr) {
 823                 .type = 1 << BCH_EXTENT_ENTRY_stripe_ptr,
 824                 .block          = block,
 825                 .redundancy     = s->key.v.nr_redundant,
 826                 .idx            = s->key.k.p.offset,
 827         };
 828 }
 829
 830 static int ec_stripe_update_extent(struct btree_trans *trans,
 831                                    struct btree_iter *iter,
 832                                    struct bkey_s_c k,
 833                                    struct ec_stripe_buf *s)
 834 {
 835         const struct bch_extent_ptr *ptr_c;
 836         struct bch_extent_ptr *ptr, *ec_ptr = NULL;
 837         struct bkey_i *n;
 838         int ret, dev, block;
 839
 840         if (extent_has_stripe_ptr(k, s->key.k.p.offset))
 841                 return 0;
 842
 843         ptr_c = bkey_matches_stripe(&s->key.v, k, &block);
 844         /*
 845          * It doesn't generally make sense to erasure code cached ptrs:
 846          * XXX: should we be incrementing a counter?
 847          */
 848         if (!ptr_c || ptr_c->cached)
 849                 return 0;
 850
 851         dev = s->key.v.ptrs[block].dev;
 852
 853         n = bch2_bkey_make_mut(trans, k);
 854         ret = PTR_ERR_OR_ZERO(n);
 855         if (ret)
 856                 return ret;
 857
 858         bch2_bkey_drop_ptrs(bkey_i_to_s(n), ptr, ptr->dev != dev);
 859         ec_ptr = (void *) bch2_bkey_has_device(bkey_i_to_s_c(n), dev);
 860         BUG_ON(!ec_ptr);
 861
 862         extent_stripe_ptr_add(bkey_i_to_s_extent(n), s, ec_ptr, block);
 863
 864         return bch2_trans_update(trans, iter, n, 0);
 865 }
 866
 867 static int ec_stripe_update_bucket(struct btree_trans *trans, struct ec_stripe_buf *s,
 868                                    unsigned block)
 869 {
 870         struct bch_fs *c = trans->c;
 871         struct bch_extent_ptr bucket = s->key.v.ptrs[block];
 872         struct bpos bucket_pos = PTR_BUCKET_POS(c, &bucket);
 873         struct bch_backpointer bp;
 874         struct btree_iter iter;
 875         struct bkey_s_c k;
 876         u64 bp_offset = 0;
 877         int ret = 0;
 878 retry:
 879         while (1) {
 880                 bch2_trans_begin(trans);
 881
 882                 ret = bch2_get_next_backpointer(trans, bucket_pos, bucket.gen,
 883                                                 &bp_offset, &bp,
 884                                                 BTREE_ITER_CACHED);
 885                 if (ret)
 886                         break;
 887                 if (bp_offset == U64_MAX)
 888                         break;
 889
 890                 if (bch2_fs_inconsistent_on(bp.level, c, "found btree node in erasure coded bucket!?")) {
 891                         ret = -EIO;
 892                         break;
 893                 }
 894
 895                 k = bch2_backpointer_get_key(trans, &iter, bucket_pos, bp_offset, bp);
 896                 ret = bkey_err(k);
 897                 if (ret)
 898                         break;
 899                 if (!k.k)
 900                         continue;
 901
 902                 ret = ec_stripe_update_extent(trans, &iter, k, s);
 903                 bch2_trans_iter_exit(trans, &iter);
 904                 if (ret)
 905                         break;
 906
 907                 bp_offset++;
 908         }
 909
 910         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 911                 goto retry;
 912
 913         return ret;
 914 }
 915
 916 static int ec_stripe_update_extents(struct bch_fs *c, struct ec_stripe_buf *s)
 917 {
 918         struct btree_trans trans;
 919         struct bch_stripe *v = &s->key.v;
 920         unsigned i, nr_data = v->nr_blocks - v->nr_redundant;
 921         int ret = 0;
 922
 923         bch2_trans_init(&trans, c, 0, 0);
 924
 925         ret = bch2_btree_write_buffer_flush(&trans);
 926         if (ret)
 927                 goto err;
 928
 929         for (i = 0; i < nr_data; i++) {
 930                 ret = ec_stripe_update_bucket(&trans, s, i);
 931                 if (ret)
 932                         break;
 933         }
 934 err:
 935         bch2_trans_exit(&trans);
 936
 937         return ret;
 938 }
 939
 940 /*
 941  * data buckets of new stripe all written: create the stripe
 942  */
 943 static void ec_stripe_create(struct ec_stripe_new *s)
 944 {
 945         struct bch_fs *c = s->c;
 946         struct open_bucket *ob;
 947         struct stripe *m;
 948         struct bch_stripe *v = &s->new_stripe.key.v;
 949         unsigned i, nr_data = v->nr_blocks - v->nr_redundant;
 950         int ret;
 951
 952         BUG_ON(s->h->s == s);
 953
 954         closure_sync(&s->iodone);
 955
 956         if (s->err) {
 957                 if (!bch2_err_matches(s->err, EROFS))
 958                         bch_err(c, "error creating stripe: error writing data buckets");
 959                 goto err;
 960         }
 961
 962         if (s->have_existing_stripe) {
 963                 ec_validate_checksums(c, &s->existing_stripe);
 964
 965                 if (ec_do_recov(c, &s->existing_stripe)) {
 966                         bch_err(c, "error creating stripe: error reading existing stripe");
 967                         goto err;
 968                 }
 969
 970                 for (i = 0; i < nr_data; i++)
 971                         if (stripe_blockcount_get(&s->existing_stripe.key.v, i))
 972                                 swap(s->new_stripe.data[i],
 973                                      s->existing_stripe.data[i]);
 974
 975                 ec_stripe_buf_exit(&s->existing_stripe);
 976         }
 977
 978         BUG_ON(!s->allocated);
 979
 980         if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_stripe_create))
 981                 goto err;
 982
 983         ec_generate_ec(&s->new_stripe);
 984
 985         ec_generate_checksums(&s->new_stripe);
 986
 987         /* write p/q: */
 988         for (i = nr_data; i < v->nr_blocks; i++)
 989                 ec_block_io(c, &s->new_stripe, REQ_OP_WRITE, i, &s->iodone);
 990         closure_sync(&s->iodone);
 991
 992         if (ec_nr_failed(&s->new_stripe)) {
 993                 bch_err(c, "error creating stripe: error writing redundancy buckets");
 994                 goto err_put_writes;
 995         }
 996
 997         ret = bch2_trans_do(c, &s->res, NULL, BTREE_INSERT_NOFAIL,
 998                             s->have_existing_stripe
 999                             ? ec_stripe_bkey_update(&trans, &s->new_stripe.key, &s->res)
1000                             : ec_stripe_bkey_insert(&trans, &s->new_stripe.key, &s->res));
1001         if (ret) {
1002                 bch_err(c, "error creating stripe: error creating stripe key");
1003                 goto err_put_writes;
1004         }
1005
1006         ret = ec_stripe_update_extents(c, &s->new_stripe);
1007         if (ret)
1008                 bch_err(c, "error creating stripe: error updating pointers: %s",
1009                         bch2_err_str(ret));
1010
1011         spin_lock(&c->ec_stripes_heap_lock);
1012         m = genradix_ptr(&c->stripes, s->new_stripe.key.k.p.offset);
1013
1014         BUG_ON(m->on_heap);
1015         bch2_stripes_heap_insert(c, m, s->new_stripe.key.k.p.offset);
1016         spin_unlock(&c->ec_stripes_heap_lock);
1017 err_put_writes:
1018         bch2_write_ref_put(c, BCH_WRITE_REF_stripe_create);
1019 err:
1020         bch2_disk_reservation_put(c, &s->res);
1021
1022         for (i = 0; i < v->nr_blocks; i++)
1023                 if (s->blocks[i]) {
1024                         ob = c->open_buckets + s->blocks[i];
1025
1026                         if (i < nr_data) {
1027                                 ob->ec = NULL;
1028                                 __bch2_open_bucket_put(c, ob);
1029                         } else {
1030                                 bch2_open_bucket_put(c, ob);
1031                         }
1032                 }
1033
1034         ec_stripe_buf_exit(&s->existing_stripe);
1035         ec_stripe_buf_exit(&s->new_stripe);
1036         closure_debug_destroy(&s->iodone);
1037         kfree(s);
1038 }
1039
1040 static void ec_stripe_create_work(struct work_struct *work)
1041 {
1042         struct bch_fs *c = container_of(work,
1043                 struct bch_fs, ec_stripe_create_work);
1044         struct ec_stripe_new *s, *n;
1045 restart:
1046         mutex_lock(&c->ec_stripe_new_lock);
1047         list_for_each_entry_safe(s, n, &c->ec_stripe_new_list, list)
1048                 if (!atomic_read(&s->pin)) {
1049                         list_del(&s->list);
1050                         mutex_unlock(&c->ec_stripe_new_lock);
1051                         ec_stripe_create(s);
1052                         goto restart;
1053                 }
1054         mutex_unlock(&c->ec_stripe_new_lock);
1055 }
1056
1057 static void ec_stripe_new_put(struct bch_fs *c, struct ec_stripe_new *s)
1058 {
1059         BUG_ON(atomic_read(&s->pin) <= 0);
1060
1061         if (atomic_dec_and_test(&s->pin)) {
1062                 BUG_ON(!s->pending);
1063                 queue_work(system_long_wq, &c->ec_stripe_create_work);
1064         }
1065 }
1066
1067 static void ec_stripe_set_pending(struct bch_fs *c, struct ec_stripe_head *h)
1068 {
1069         struct ec_stripe_new *s = h->s;
1070
1071         BUG_ON(!s->allocated && !s->err);
1072
1073         h->s            = NULL;
1074         s->pending      = true;
1075
1076         mutex_lock(&c->ec_stripe_new_lock);
1077         list_add(&s->list, &c->ec_stripe_new_list);
1078         mutex_unlock(&c->ec_stripe_new_lock);
1079
1080         ec_stripe_new_put(c, s);
1081 }
1082
1083 /* have a full bucket - hand it off to be erasure coded: */
1084 void bch2_ec_bucket_written(struct bch_fs *c, struct open_bucket *ob)
1085 {
1086         struct ec_stripe_new *s = ob->ec;
1087
1088         if (ob->sectors_free)
1089                 s->err = -1;
1090
1091         ec_stripe_new_put(c, s);
1092 }
1093
1094 void bch2_ec_bucket_cancel(struct bch_fs *c, struct open_bucket *ob)
1095 {
1096         struct ec_stripe_new *s = ob->ec;
1097
1098         s->err = -EIO;
1099 }
1100
1101 void *bch2_writepoint_ec_buf(struct bch_fs *c, struct write_point *wp)
1102 {
1103         struct open_bucket *ob = ec_open_bucket(c, &wp->ptrs);
1104         struct bch_dev *ca;
1105         unsigned offset;
1106
1107         if (!ob)
1108                 return NULL;
1109
1110         ca      = bch_dev_bkey_exists(c, ob->dev);
1111         offset  = ca->mi.bucket_size - ob->sectors_free;
1112
1113         return ob->ec->new_stripe.data[ob->ec_idx] + (offset << 9);
1114 }
1115
1116 static int unsigned_cmp(const void *_l, const void *_r)
1117 {
1118         unsigned l = *((const unsigned *) _l);
1119         unsigned r = *((const unsigned *) _r);
1120
1121         return cmp_int(l, r);
1122 }
1123
1124 /* pick most common bucket size: */
1125 static unsigned pick_blocksize(struct bch_fs *c,
1126                                struct bch_devs_mask *devs)
1127 {
1128         struct bch_dev *ca;
1129         unsigned i, nr = 0, sizes[BCH_SB_MEMBERS_MAX];
1130         struct {
1131                 unsigned nr, size;
1132         } cur = { 0, 0 }, best = { 0, 0 };
1133
1134         for_each_member_device_rcu(ca, c, i, devs)
1135                 sizes[nr++] = ca->mi.bucket_size;
1136
1137         sort(sizes, nr, sizeof(unsigned), unsigned_cmp, NULL);
1138
1139         for (i = 0; i < nr; i++) {
1140                 if (sizes[i] != cur.size) {
1141                         if (cur.nr > best.nr)
1142                                 best = cur;
1143
1144                         cur.nr = 0;
1145                         cur.size = sizes[i];
1146                 }
1147
1148                 cur.nr++;
1149         }
1150
1151         if (cur.nr > best.nr)
1152                 best = cur;
1153
1154         return best.size;
1155 }
1156
1157 static bool may_create_new_stripe(struct bch_fs *c)
1158 {
1159         return false;
1160 }
1161
1162 static void ec_stripe_key_init(struct bch_fs *c,
1163                                struct bkey_i_stripe *s,
1164                                unsigned nr_data,
1165                                unsigned nr_parity,
1166                                unsigned stripe_size)
1167 {
1168         unsigned u64s;
1169
1170         bkey_stripe_init(&s->k_i);
1171         s->v.sectors                    = cpu_to_le16(stripe_size);
1172         s->v.algorithm                  = 0;
1173         s->v.nr_blocks                  = nr_data + nr_parity;
1174         s->v.nr_redundant               = nr_parity;
1175         s->v.csum_granularity_bits      = ilog2(c->opts.encoded_extent_max >> 9);
1176         s->v.csum_type                  = BCH_CSUM_crc32c;
1177         s->v.pad                        = 0;
1178
1179         while ((u64s = stripe_val_u64s(&s->v)) > BKEY_VAL_U64s_MAX) {
1180                 BUG_ON(1 << s->v.csum_granularity_bits >=
1181                        le16_to_cpu(s->v.sectors) ||
1182                        s->v.csum_granularity_bits == U8_MAX);
1183                 s->v.csum_granularity_bits++;
1184         }
1185
1186         set_bkey_val_u64s(&s->k, u64s);
1187 }
1188
1189 static int ec_new_stripe_alloc(struct bch_fs *c, struct ec_stripe_head *h)
1190 {
1191         struct ec_stripe_new *s;
1192
1193         lockdep_assert_held(&h->lock);
1194
1195         s = kzalloc(sizeof(*s), GFP_KERNEL);
1196         if (!s)
1197                 return -ENOMEM;
1198
1199         mutex_init(&s->lock);
1200         closure_init(&s->iodone, NULL);
1201         atomic_set(&s->pin, 1);
1202         s->c            = c;
1203         s->h            = h;
1204         s->nr_data      = min_t(unsigned, h->nr_active_devs,
1205                                 BCH_BKEY_PTRS_MAX) - h->redundancy;
1206         s->nr_parity    = h->redundancy;
1207
1208         ec_stripe_key_init(c, &s->new_stripe.key, s->nr_data,
1209                            s->nr_parity, h->blocksize);
1210
1211         h->s = s;
1212         return 0;
1213 }
1214
1215 static struct ec_stripe_head *
1216 ec_new_stripe_head_alloc(struct bch_fs *c, unsigned target,
1217                          unsigned algo, unsigned redundancy,
1218                          bool copygc)
1219 {
1220         struct ec_stripe_head *h;
1221         struct bch_dev *ca;
1222         unsigned i;
1223
1224         h = kzalloc(sizeof(*h), GFP_KERNEL);
1225         if (!h)
1226                 return NULL;
1227
1228         mutex_init(&h->lock);
1229         mutex_lock(&h->lock);
1230
1231         h->target       = target;
1232         h->algo         = algo;
1233         h->redundancy   = redundancy;
1234         h->copygc       = copygc;
1235
1236         rcu_read_lock();
1237         h->devs = target_rw_devs(c, BCH_DATA_user, target);
1238
1239         for_each_member_device_rcu(ca, c, i, &h->devs)
1240                 if (!ca->mi.durability)
1241                         __clear_bit(i, h->devs.d);
1242
1243         h->blocksize = pick_blocksize(c, &h->devs);
1244
1245         for_each_member_device_rcu(ca, c, i, &h->devs)
1246                 if (ca->mi.bucket_size == h->blocksize)
1247                         h->nr_active_devs++;
1248
1249         rcu_read_unlock();
1250         list_add(&h->list, &c->ec_stripe_head_list);
1251         return h;
1252 }
1253
1254 void bch2_ec_stripe_head_put(struct bch_fs *c, struct ec_stripe_head *h)
1255 {
1256         if (h->s &&
1257             h->s->allocated &&
1258             bitmap_weight(h->s->blocks_allocated,
1259                           h->s->nr_data) == h->s->nr_data)
1260                 ec_stripe_set_pending(c, h);
1261
1262         mutex_unlock(&h->lock);
1263 }
1264
1265 struct ec_stripe_head *__bch2_ec_stripe_head_get(struct bch_fs *c,
1266                                                  unsigned target,
1267                                                  unsigned algo,
1268                                                  unsigned redundancy,
1269                                                  bool copygc)
1270 {
1271         struct ec_stripe_head *h;
1272
1273         if (!redundancy)
1274                 return NULL;
1275
1276         mutex_lock(&c->ec_stripe_head_lock);
1277         list_for_each_entry(h, &c->ec_stripe_head_list, list)
1278                 if (h->target           == target &&
1279                     h->algo             == algo &&
1280                     h->redundancy       == redundancy &&
1281                     h->copygc           == copygc) {
1282                         mutex_lock(&h->lock);
1283                         goto found;
1284                 }
1285
1286         h = ec_new_stripe_head_alloc(c, target, algo, redundancy, copygc);
1287 found:
1288         mutex_unlock(&c->ec_stripe_head_lock);
1289         return h;
1290 }
1291
1292 static int new_stripe_alloc_buckets(struct bch_fs *c, struct ec_stripe_head *h,
1293                                     struct closure *cl)
1294 {
1295         struct bch_devs_mask devs = h->devs;
1296         struct open_bucket *ob;
1297         struct open_buckets buckets;
1298         unsigned i, j, nr_have_parity = 0, nr_have_data = 0;
1299         bool have_cache = true;
1300         int ret = 0;
1301
1302         for (i = 0; i < h->s->new_stripe.key.v.nr_blocks; i++) {
1303                 if (test_bit(i, h->s->blocks_gotten)) {
1304                         __clear_bit(h->s->new_stripe.key.v.ptrs[i].dev, devs.d);
1305                         if (i < h->s->nr_data)
1306                                 nr_have_data++;
1307                         else
1308                                 nr_have_parity++;
1309                 }
1310         }
1311
1312         BUG_ON(nr_have_data     > h->s->nr_data);
1313         BUG_ON(nr_have_parity   > h->s->nr_parity);
1314
1315         buckets.nr = 0;
1316         if (nr_have_parity < h->s->nr_parity) {
1317                 ret = bch2_bucket_alloc_set(c, &buckets,
1318                                             &h->parity_stripe,
1319                                             &devs,
1320                                             h->s->nr_parity,
1321                                             &nr_have_parity,
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 + h->s->nr_parity,
1332                                                h->s->nr_data);
1333                         BUG_ON(j >= h->s->nr_data + h->s->nr_parity);
1334
1335                         h->s->blocks[j] = buckets.v[i];
1336                         h->s->new_stripe.key.v.ptrs[j] = bch2_ob_ptr(c, ob);
1337                         __set_bit(j, h->s->blocks_gotten);
1338                 }
1339
1340                 if (ret)
1341                         return ret;
1342         }
1343
1344         buckets.nr = 0;
1345         if (nr_have_data < h->s->nr_data) {
1346                 ret = bch2_bucket_alloc_set(c, &buckets,
1347                                             &h->block_stripe,
1348                                             &devs,
1349                                             h->s->nr_data,
1350                                             &nr_have_data,
1351                                             &have_cache,
1352                                             h->copygc
1353                                             ? RESERVE_movinggc
1354                                             : RESERVE_none,
1355                                             0,
1356                                             cl);
1357
1358                 open_bucket_for_each(c, &buckets, ob, i) {
1359                         j = find_next_zero_bit(h->s->blocks_gotten,
1360                                                h->s->nr_data, 0);
1361                         BUG_ON(j >= h->s->nr_data);
1362
1363                         h->s->blocks[j] = buckets.v[i];
1364                         h->s->new_stripe.key.v.ptrs[j] = bch2_ob_ptr(c, ob);
1365                         __set_bit(j, h->s->blocks_gotten);
1366                 }
1367
1368                 if (ret)
1369                         return ret;
1370         }
1371
1372         return 0;
1373 }
1374
1375 /* XXX: doesn't obey target: */
1376 static s64 get_existing_stripe(struct bch_fs *c,
1377                                struct ec_stripe_head *head)
1378 {
1379         ec_stripes_heap *h = &c->ec_stripes_heap;
1380         struct stripe *m;
1381         size_t heap_idx;
1382         u64 stripe_idx;
1383         s64 ret = -1;
1384
1385         if (may_create_new_stripe(c))
1386                 return -1;
1387
1388         spin_lock(&c->ec_stripes_heap_lock);
1389         for (heap_idx = 0; heap_idx < h->used; heap_idx++) {
1390                 /* No blocks worth reusing, stripe will just be deleted: */
1391                 if (!h->data[heap_idx].blocks_nonempty)
1392                         continue;
1393
1394                 stripe_idx = h->data[heap_idx].idx;
1395                 m = genradix_ptr(&c->stripes, stripe_idx);
1396
1397                 if (m->algorithm        == head->algo &&
1398                     m->nr_redundant     == head->redundancy &&
1399                     m->sectors          == head->blocksize &&
1400                     m->blocks_nonempty  < m->nr_blocks - m->nr_redundant) {
1401                         bch2_stripes_heap_del(c, m, stripe_idx);
1402                         ret = stripe_idx;
1403                         break;
1404                 }
1405         }
1406         spin_unlock(&c->ec_stripes_heap_lock);
1407         return ret;
1408 }
1409
1410 static int __bch2_ec_stripe_head_reuse(struct bch_fs *c,
1411                                                    struct ec_stripe_head *h)
1412 {
1413         unsigned i;
1414         s64 idx;
1415         int ret;
1416
1417         idx = get_existing_stripe(c, h);
1418         if (idx < 0)
1419                 return -BCH_ERR_ENOSPC_stripe_reuse;
1420
1421         h->s->have_existing_stripe = true;
1422         ret = get_stripe_key(c, idx, &h->s->existing_stripe);
1423         if (ret) {
1424                 bch2_fs_fatal_error(c, "error reading stripe key: %i", ret);
1425                 return ret;
1426         }
1427
1428         if (ec_stripe_buf_init(&h->s->existing_stripe, 0, h->blocksize)) {
1429                 /*
1430                  * this is a problem: we have deleted from the
1431                  * stripes heap already
1432                  */
1433                 BUG();
1434         }
1435
1436         BUG_ON(h->s->existing_stripe.size != h->blocksize);
1437         BUG_ON(h->s->existing_stripe.size != h->s->existing_stripe.key.v.sectors);
1438
1439         for (i = 0; i < h->s->existing_stripe.key.v.nr_blocks; i++) {
1440                 if (stripe_blockcount_get(&h->s->existing_stripe.key.v, i)) {
1441                         __set_bit(i, h->s->blocks_gotten);
1442                         __set_bit(i, h->s->blocks_allocated);
1443                 }
1444
1445                 ec_block_io(c, &h->s->existing_stripe, READ, i, &h->s->iodone);
1446         }
1447
1448         bkey_copy(&h->s->new_stripe.key.k_i,
1449                         &h->s->existing_stripe.key.k_i);
1450
1451         return 0;
1452 }
1453
1454 static int __bch2_ec_stripe_head_reserve(struct bch_fs *c,
1455                                                         struct ec_stripe_head *h)
1456 {
1457         return bch2_disk_reservation_get(c, &h->s->res,
1458                                          h->blocksize,
1459                                          h->s->nr_parity, 0);
1460 }
1461
1462 struct ec_stripe_head *bch2_ec_stripe_head_get(struct bch_fs *c,
1463                                                unsigned target,
1464                                                unsigned algo,
1465                                                unsigned redundancy,
1466                                                bool copygc,
1467                                                struct closure *cl)
1468 {
1469         struct ec_stripe_head *h;
1470         int ret;
1471         bool needs_stripe_new;
1472
1473         h = __bch2_ec_stripe_head_get(c, target, algo, redundancy, copygc);
1474         if (!h) {
1475                 bch_err(c, "no stripe head");
1476                 return NULL;
1477         }
1478
1479         needs_stripe_new = !h->s;
1480         if (needs_stripe_new) {
1481                 if (ec_new_stripe_alloc(c, h)) {
1482                         ret = -ENOMEM;
1483                         bch_err(c, "failed to allocate new stripe");
1484                         goto err;
1485                 }
1486
1487                 if (ec_stripe_buf_init(&h->s->new_stripe, 0, h->blocksize))
1488                         BUG();
1489         }
1490
1491         /*
1492          * Try reserve a new stripe before reusing an
1493          * existing stripe. This will prevent unnecessary
1494          * read amplification during write oriented workloads.
1495          */
1496         ret = 0;
1497         if (!h->s->allocated && !h->s->res.sectors && !h->s->have_existing_stripe)
1498                 ret = __bch2_ec_stripe_head_reserve(c, h);
1499         if (ret && needs_stripe_new)
1500                 ret = __bch2_ec_stripe_head_reuse(c, h);
1501         if (ret) {
1502                 bch_err_ratelimited(c, "failed to get stripe: %s", bch2_err_str(ret));
1503                 goto err;
1504         }
1505
1506         if (!h->s->allocated) {
1507                 ret = new_stripe_alloc_buckets(c, h, cl);
1508                 if (ret)
1509                         goto err;
1510
1511                 h->s->allocated = true;
1512         }
1513
1514         return h;
1515
1516 err:
1517         bch2_ec_stripe_head_put(c, h);
1518         return ERR_PTR(ret);
1519 }
1520
1521 void bch2_ec_stop_dev(struct bch_fs *c, struct bch_dev *ca)
1522 {
1523         struct ec_stripe_head *h;
1524         struct open_bucket *ob;
1525         unsigned i;
1526
1527         mutex_lock(&c->ec_stripe_head_lock);
1528         list_for_each_entry(h, &c->ec_stripe_head_list, list) {
1529
1530                 mutex_lock(&h->lock);
1531                 if (!h->s)
1532                         goto unlock;
1533
1534                 for (i = 0; i < h->s->new_stripe.key.v.nr_blocks; i++) {
1535                         if (!h->s->blocks[i])
1536                                 continue;
1537
1538                         ob = c->open_buckets + h->s->blocks[i];
1539                         if (ob->dev == ca->dev_idx)
1540                                 goto found;
1541                 }
1542                 goto unlock;
1543 found:
1544                 h->s->err = -EROFS;
1545                 ec_stripe_set_pending(c, h);
1546 unlock:
1547                 mutex_unlock(&h->lock);
1548         }
1549         mutex_unlock(&c->ec_stripe_head_lock);
1550 }
1551
1552 void bch2_stripes_heap_start(struct bch_fs *c)
1553 {
1554         struct genradix_iter iter;
1555         struct stripe *m;
1556
1557         genradix_for_each(&c->stripes, iter, m)
1558                 if (m->alive)
1559                         bch2_stripes_heap_insert(c, m, iter.pos);
1560 }
1561
1562 int bch2_stripes_read(struct bch_fs *c)
1563 {
1564         struct btree_trans trans;
1565         struct btree_iter iter;
1566         struct bkey_s_c k;
1567         const struct bch_stripe *s;
1568         struct stripe *m;
1569         unsigned i;
1570         int ret;
1571
1572         bch2_trans_init(&trans, c, 0, 0);
1573
1574         for_each_btree_key(&trans, iter, BTREE_ID_stripes, POS_MIN,
1575                            BTREE_ITER_PREFETCH, k, ret) {
1576                 if (k.k->type != KEY_TYPE_stripe)
1577                         continue;
1578
1579                 ret = __ec_stripe_mem_alloc(c, k.k->p.offset, GFP_KERNEL);
1580                 if (ret)
1581                         break;
1582
1583                 s = bkey_s_c_to_stripe(k).v;
1584
1585                 m = genradix_ptr(&c->stripes, k.k->p.offset);
1586                 m->alive        = true;
1587                 m->sectors      = le16_to_cpu(s->sectors);
1588                 m->algorithm    = s->algorithm;
1589                 m->nr_blocks    = s->nr_blocks;
1590                 m->nr_redundant = s->nr_redundant;
1591                 m->blocks_nonempty = 0;
1592
1593                 for (i = 0; i < s->nr_blocks; i++)
1594                         m->blocks_nonempty += !!stripe_blockcount_get(s, i);
1595
1596                 spin_lock(&c->ec_stripes_heap_lock);
1597                 bch2_stripes_heap_update(c, m, k.k->p.offset);
1598                 spin_unlock(&c->ec_stripes_heap_lock);
1599         }
1600         bch2_trans_iter_exit(&trans, &iter);
1601
1602         bch2_trans_exit(&trans);
1603
1604         if (ret)
1605                 bch_err(c, "error reading stripes: %i", ret);
1606
1607         return ret;
1608 }
1609
1610 void bch2_stripes_heap_to_text(struct printbuf *out, struct bch_fs *c)
1611 {
1612         ec_stripes_heap *h = &c->ec_stripes_heap;
1613         struct stripe *m;
1614         size_t i;
1615
1616         spin_lock(&c->ec_stripes_heap_lock);
1617         for (i = 0; i < min_t(size_t, h->used, 20); i++) {
1618                 m = genradix_ptr(&c->stripes, h->data[i].idx);
1619
1620                 prt_printf(out, "%zu %u/%u+%u\n", h->data[i].idx,
1621                        h->data[i].blocks_nonempty,
1622                        m->nr_blocks - m->nr_redundant,
1623                        m->nr_redundant);
1624         }
1625         spin_unlock(&c->ec_stripes_heap_lock);
1626 }
1627
1628 void bch2_new_stripes_to_text(struct printbuf *out, struct bch_fs *c)
1629 {
1630         struct ec_stripe_head *h;
1631         struct ec_stripe_new *s;
1632
1633         mutex_lock(&c->ec_stripe_head_lock);
1634         list_for_each_entry(h, &c->ec_stripe_head_list, list) {
1635                 prt_printf(out, "target %u algo %u redundancy %u:\n",
1636                        h->target, h->algo, h->redundancy);
1637
1638                 if (h->s)
1639                         prt_printf(out, "\tpending: blocks %u+%u allocated %u\n",
1640                                h->s->nr_data, h->s->nr_parity,
1641                                bitmap_weight(h->s->blocks_allocated,
1642                                              h->s->nr_data));
1643         }
1644         mutex_unlock(&c->ec_stripe_head_lock);
1645
1646         mutex_lock(&c->ec_stripe_new_lock);
1647         list_for_each_entry(s, &c->ec_stripe_new_list, list) {
1648                 prt_printf(out, "\tin flight: blocks %u+%u pin %u\n",
1649                        s->nr_data, s->nr_parity,
1650                        atomic_read(&s->pin));
1651         }
1652         mutex_unlock(&c->ec_stripe_new_lock);
1653 }
1654
1655 void bch2_fs_ec_exit(struct bch_fs *c)
1656 {
1657         struct ec_stripe_head *h;
1658
1659         while (1) {
1660                 mutex_lock(&c->ec_stripe_head_lock);
1661                 h = list_first_entry_or_null(&c->ec_stripe_head_list,
1662                                              struct ec_stripe_head, list);
1663                 if (h)
1664                         list_del(&h->list);
1665                 mutex_unlock(&c->ec_stripe_head_lock);
1666                 if (!h)
1667                         break;
1668
1669                 BUG_ON(h->s);
1670                 kfree(h);
1671         }
1672
1673         BUG_ON(!list_empty(&c->ec_stripe_new_list));
1674
1675         free_heap(&c->ec_stripes_heap);
1676         genradix_free(&c->stripes);
1677         bioset_exit(&c->ec_bioset);
1678 }
1679
1680 void bch2_fs_ec_init_early(struct bch_fs *c)
1681 {
1682         INIT_WORK(&c->ec_stripe_create_work, ec_stripe_create_work);
1683         INIT_WORK(&c->ec_stripe_delete_work, ec_stripe_delete_work);
1684 }
1685
1686 int bch2_fs_ec_init(struct bch_fs *c)
1687 {
1688         return bioset_init(&c->ec_bioset, 1, offsetof(struct ec_bio, bio),
1689                            BIOSET_NEED_BVECS);
1690 }