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, nr_data = s->nr_blocks - s->nr_redundant;
 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                nr_data,
 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                 const struct bch_extent_ptr *ptr = s->ptrs + i;
 153                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 154                 u32 offset;
 155                 u64 b = sector_to_bucket_and_offset(ca, ptr->offset, &offset);
 156
 157                 prt_printf(out, " %u:%llu:%u", ptr->dev, b, offset);
 158                 if (i < nr_data)
 159                         prt_printf(out, "#%u", stripe_blockcount_get(s, i));
 160                 if (ptr_stale(ca, ptr))
 161                         prt_printf(out, " stale");
 162         }
 163 }
 164
 165 /* returns blocknr in stripe that we matched: */
 166 static const struct bch_extent_ptr *bkey_matches_stripe(struct bch_stripe *s,
 167                                                 struct bkey_s_c k, unsigned *block)
 168 {
 169         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 170         const struct bch_extent_ptr *ptr;
 171         unsigned i, nr_data = s->nr_blocks - s->nr_redundant;
 172
 173         bkey_for_each_ptr(ptrs, ptr)
 174                 for (i = 0; i < nr_data; i++)
 175                         if (__bch2_ptr_matches_stripe(&s->ptrs[i], ptr,
 176                                                       le16_to_cpu(s->sectors))) {
 177                                 *block = i;
 178                                 return ptr;
 179                         }
 180
 181         return NULL;
 182 }
 183
 184 static bool extent_has_stripe_ptr(struct bkey_s_c k, u64 idx)
 185 {
 186         switch (k.k->type) {
 187         case KEY_TYPE_extent: {
 188                 struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
 189                 const union bch_extent_entry *entry;
 190
 191                 extent_for_each_entry(e, entry)
 192                         if (extent_entry_type(entry) ==
 193                             BCH_EXTENT_ENTRY_stripe_ptr &&
 194                             entry->stripe_ptr.idx == idx)
 195                                 return true;
 196
 197                 break;
 198         }
 199         }
 200
 201         return false;
 202 }
 203
 204 /* Stripe bufs: */
 205
 206 static void ec_stripe_buf_exit(struct ec_stripe_buf *buf)
 207 {
 208         unsigned i;
 209
 210         for (i = 0; i < buf->key.v.nr_blocks; i++) {
 211                 kvpfree(buf->data[i], buf->size << 9);
 212                 buf->data[i] = NULL;
 213         }
 214 }
 215
 216 /* XXX: this is a non-mempoolified memory allocation: */
 217 static int ec_stripe_buf_init(struct ec_stripe_buf *buf,
 218                               unsigned offset, unsigned size)
 219 {
 220         struct bch_stripe *v = &buf->key.v;
 221         unsigned csum_granularity = 1U << v->csum_granularity_bits;
 222         unsigned end = offset + size;
 223         unsigned i;
 224
 225         BUG_ON(end > le16_to_cpu(v->sectors));
 226
 227         offset  = round_down(offset, csum_granularity);
 228         end     = min_t(unsigned, le16_to_cpu(v->sectors),
 229                         round_up(end, csum_granularity));
 230
 231         buf->offset     = offset;
 232         buf->size       = end - offset;
 233
 234         memset(buf->valid, 0xFF, sizeof(buf->valid));
 235
 236         for (i = 0; i < buf->key.v.nr_blocks; i++) {
 237                 buf->data[i] = kvpmalloc(buf->size << 9, GFP_KERNEL);
 238                 if (!buf->data[i])
 239                         goto err;
 240         }
 241
 242         return 0;
 243 err:
 244         ec_stripe_buf_exit(buf);
 245         return -BCH_ERR_ENOMEM_stripe_buf;
 246 }
 247
 248 /* Checksumming: */
 249
 250 static struct bch_csum ec_block_checksum(struct ec_stripe_buf *buf,
 251                                          unsigned block, unsigned offset)
 252 {
 253         struct bch_stripe *v = &buf->key.v;
 254         unsigned csum_granularity = 1 << v->csum_granularity_bits;
 255         unsigned end = buf->offset + buf->size;
 256         unsigned len = min(csum_granularity, end - offset);
 257
 258         BUG_ON(offset >= end);
 259         BUG_ON(offset <  buf->offset);
 260         BUG_ON(offset & (csum_granularity - 1));
 261         BUG_ON(offset + len != le16_to_cpu(v->sectors) &&
 262                (len & (csum_granularity - 1)));
 263
 264         return bch2_checksum(NULL, v->csum_type,
 265                              null_nonce(),
 266                              buf->data[block] + ((offset - buf->offset) << 9),
 267                              len << 9);
 268 }
 269
 270 static void ec_generate_checksums(struct ec_stripe_buf *buf)
 271 {
 272         struct bch_stripe *v = &buf->key.v;
 273         unsigned i, j, csums_per_device = stripe_csums_per_device(v);
 274
 275         if (!v->csum_type)
 276                 return;
 277
 278         BUG_ON(buf->offset);
 279         BUG_ON(buf->size != le16_to_cpu(v->sectors));
 280
 281         for (i = 0; i < v->nr_blocks; i++)
 282                 for (j = 0; j < csums_per_device; j++)
 283                         stripe_csum_set(v, i, j,
 284                                 ec_block_checksum(buf, i, j << v->csum_granularity_bits));
 285 }
 286
 287 static void ec_validate_checksums(struct bch_fs *c, struct ec_stripe_buf *buf)
 288 {
 289         struct bch_stripe *v = &buf->key.v;
 290         unsigned csum_granularity = 1 << v->csum_granularity_bits;
 291         unsigned i;
 292
 293         if (!v->csum_type)
 294                 return;
 295
 296         for (i = 0; i < v->nr_blocks; i++) {
 297                 unsigned offset = buf->offset;
 298                 unsigned end = buf->offset + buf->size;
 299
 300                 if (!test_bit(i, buf->valid))
 301                         continue;
 302
 303                 while (offset < end) {
 304                         unsigned j = offset >> v->csum_granularity_bits;
 305                         unsigned len = min(csum_granularity, end - offset);
 306                         struct bch_csum want = stripe_csum_get(v, i, j);
 307                         struct bch_csum got = ec_block_checksum(buf, i, offset);
 308
 309                         if (bch2_crc_cmp(want, got)) {
 310                                 struct printbuf buf2 = PRINTBUF;
 311
 312                                 bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(&buf->key.k_i));
 313
 314                                 bch_err_ratelimited(c,
 315                                         "stripe checksum error for %ps at %u:%u: csum type %u, expected %llx got %llx\n%s",
 316                                         (void *) _RET_IP_, i, j, v->csum_type,
 317                                         want.lo, got.lo, buf2.buf);
 318                                 printbuf_exit(&buf2);
 319                                 clear_bit(i, buf->valid);
 320                                 break;
 321                         }
 322
 323                         offset += len;
 324                 }
 325         }
 326 }
 327
 328 /* Erasure coding: */
 329
 330 static void ec_generate_ec(struct ec_stripe_buf *buf)
 331 {
 332         struct bch_stripe *v = &buf->key.v;
 333         unsigned nr_data = v->nr_blocks - v->nr_redundant;
 334         unsigned bytes = le16_to_cpu(v->sectors) << 9;
 335
 336         raid_gen(nr_data, v->nr_redundant, bytes, buf->data);
 337 }
 338
 339 static unsigned ec_nr_failed(struct ec_stripe_buf *buf)
 340 {
 341         return buf->key.v.nr_blocks -
 342                 bitmap_weight(buf->valid, buf->key.v.nr_blocks);
 343 }
 344
 345 static int ec_do_recov(struct bch_fs *c, struct ec_stripe_buf *buf)
 346 {
 347         struct bch_stripe *v = &buf->key.v;
 348         unsigned i, failed[BCH_BKEY_PTRS_MAX], nr_failed = 0;
 349         unsigned nr_data = v->nr_blocks - v->nr_redundant;
 350         unsigned bytes = buf->size << 9;
 351
 352         if (ec_nr_failed(buf) > v->nr_redundant) {
 353                 bch_err_ratelimited(c,
 354                         "error doing reconstruct read: unable to read enough blocks");
 355                 return -1;
 356         }
 357
 358         for (i = 0; i < nr_data; i++)
 359                 if (!test_bit(i, buf->valid))
 360                         failed[nr_failed++] = i;
 361
 362         raid_rec(nr_failed, failed, nr_data, v->nr_redundant, bytes, buf->data);
 363         return 0;
 364 }
 365
 366 /* IO: */
 367
 368 static void ec_block_endio(struct bio *bio)
 369 {
 370         struct ec_bio *ec_bio = container_of(bio, struct ec_bio, bio);
 371         struct bch_stripe *v = &ec_bio->buf->key.v;
 372         struct bch_extent_ptr *ptr = &v->ptrs[ec_bio->idx];
 373         struct bch_dev *ca = ec_bio->ca;
 374         struct closure *cl = bio->bi_private;
 375
 376         if (bch2_dev_io_err_on(bio->bi_status, ca, "erasure coding %s error: %s",
 377                                bio_data_dir(bio) ? "write" : "read",
 378                                bch2_blk_status_to_str(bio->bi_status)))
 379                 clear_bit(ec_bio->idx, ec_bio->buf->valid);
 380
 381         if (ptr_stale(ca, ptr)) {
 382                 bch_err_ratelimited(ca->fs,
 383                                     "error %s stripe: stale pointer after io",
 384                                     bio_data_dir(bio) == READ ? "reading from" : "writing to");
 385                 clear_bit(ec_bio->idx, ec_bio->buf->valid);
 386         }
 387
 388         bio_put(&ec_bio->bio);
 389         percpu_ref_put(&ca->io_ref);
 390         closure_put(cl);
 391 }
 392
 393 static void ec_block_io(struct bch_fs *c, struct ec_stripe_buf *buf,
 394                         unsigned rw, unsigned idx, struct closure *cl)
 395 {
 396         struct bch_stripe *v = &buf->key.v;
 397         unsigned offset = 0, bytes = buf->size << 9;
 398         struct bch_extent_ptr *ptr = &v->ptrs[idx];
 399         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 400         enum bch_data_type data_type = idx < buf->key.v.nr_blocks - buf->key.v.nr_redundant
 401                 ? BCH_DATA_user
 402                 : BCH_DATA_parity;
 403
 404         if (ptr_stale(ca, ptr)) {
 405                 bch_err_ratelimited(c,
 406                                     "error %s stripe: stale pointer",
 407                                     rw == READ ? "reading from" : "writing to");
 408                 clear_bit(idx, buf->valid);
 409                 return;
 410         }
 411
 412         if (!bch2_dev_get_ioref(ca, rw)) {
 413                 clear_bit(idx, buf->valid);
 414                 return;
 415         }
 416
 417         this_cpu_add(ca->io_done->sectors[rw][data_type], buf->size);
 418
 419         while (offset < bytes) {
 420                 unsigned nr_iovecs = min_t(size_t, BIO_MAX_VECS,
 421                                            DIV_ROUND_UP(bytes, PAGE_SIZE));
 422                 unsigned b = min_t(size_t, bytes - offset,
 423                                    nr_iovecs << PAGE_SHIFT);
 424                 struct ec_bio *ec_bio;
 425
 426                 ec_bio = container_of(bio_alloc_bioset(ca->disk_sb.bdev,
 427                                                        nr_iovecs,
 428                                                        rw,
 429                                                        GFP_KERNEL,
 430                                                        &c->ec_bioset),
 431                                       struct ec_bio, bio);
 432
 433                 ec_bio->ca                      = ca;
 434                 ec_bio->buf                     = buf;
 435                 ec_bio->idx                     = idx;
 436
 437                 ec_bio->bio.bi_iter.bi_sector   = ptr->offset + buf->offset + (offset >> 9);
 438                 ec_bio->bio.bi_end_io           = ec_block_endio;
 439                 ec_bio->bio.bi_private          = cl;
 440
 441                 bch2_bio_map(&ec_bio->bio, buf->data[idx] + offset, b);
 442
 443                 closure_get(cl);
 444                 percpu_ref_get(&ca->io_ref);
 445
 446                 submit_bio(&ec_bio->bio);
 447
 448                 offset += b;
 449         }
 450
 451         percpu_ref_put(&ca->io_ref);
 452 }
 453
 454 static int get_stripe_key_trans(struct btree_trans *trans, u64 idx,
 455                                 struct ec_stripe_buf *stripe)
 456 {
 457         struct btree_iter iter;
 458         struct bkey_s_c k;
 459         int ret;
 460
 461         bch2_trans_iter_init(trans, &iter, BTREE_ID_stripes,
 462                              POS(0, idx), BTREE_ITER_SLOTS);
 463         k = bch2_btree_iter_peek_slot(&iter);
 464         ret = bkey_err(k);
 465         if (ret)
 466                 goto err;
 467         if (k.k->type != KEY_TYPE_stripe) {
 468                 ret = -ENOENT;
 469                 goto err;
 470         }
 471         bkey_reassemble(&stripe->key.k_i, k);
 472 err:
 473         bch2_trans_iter_exit(trans, &iter);
 474         return ret;
 475 }
 476
 477 static int get_stripe_key(struct bch_fs *c, u64 idx, struct ec_stripe_buf *stripe)
 478 {
 479         return bch2_trans_run(c, get_stripe_key_trans(&trans, idx, stripe));
 480 }
 481
 482 /* recovery read path: */
 483 int bch2_ec_read_extent(struct bch_fs *c, struct bch_read_bio *rbio)
 484 {
 485         struct ec_stripe_buf *buf;
 486         struct closure cl;
 487         struct bch_stripe *v;
 488         unsigned i, offset;
 489         int ret = 0;
 490
 491         closure_init_stack(&cl);
 492
 493         BUG_ON(!rbio->pick.has_ec);
 494
 495         buf = kzalloc(sizeof(*buf), GFP_NOIO);
 496         if (!buf)
 497                 return -ENOMEM;
 498
 499         ret = get_stripe_key(c, rbio->pick.ec.idx, buf);
 500         if (ret) {
 501                 bch_err_ratelimited(c,
 502                         "error doing reconstruct read: error %i looking up stripe", ret);
 503                 kfree(buf);
 504                 return -EIO;
 505         }
 506
 507         v = &buf->key.v;
 508
 509         if (!bch2_ptr_matches_stripe(v, rbio->pick)) {
 510                 bch_err_ratelimited(c,
 511                         "error doing reconstruct read: pointer doesn't match stripe");
 512                 ret = -EIO;
 513                 goto err;
 514         }
 515
 516         offset = rbio->bio.bi_iter.bi_sector - v->ptrs[rbio->pick.ec.block].offset;
 517         if (offset + bio_sectors(&rbio->bio) > le16_to_cpu(v->sectors)) {
 518                 bch_err_ratelimited(c,
 519                         "error doing reconstruct read: read is bigger than stripe");
 520                 ret = -EIO;
 521                 goto err;
 522         }
 523
 524         ret = ec_stripe_buf_init(buf, offset, bio_sectors(&rbio->bio));
 525         if (ret)
 526                 goto err;
 527
 528         for (i = 0; i < v->nr_blocks; i++)
 529                 ec_block_io(c, buf, REQ_OP_READ, i, &cl);
 530
 531         closure_sync(&cl);
 532
 533         if (ec_nr_failed(buf) > v->nr_redundant) {
 534                 bch_err_ratelimited(c,
 535                         "error doing reconstruct read: unable to read enough blocks");
 536                 ret = -EIO;
 537                 goto err;
 538         }
 539
 540         ec_validate_checksums(c, buf);
 541
 542         ret = ec_do_recov(c, buf);
 543         if (ret)
 544                 goto err;
 545
 546         memcpy_to_bio(&rbio->bio, rbio->bio.bi_iter,
 547                       buf->data[rbio->pick.ec.block] + ((offset - buf->offset) << 9));
 548 err:
 549         ec_stripe_buf_exit(buf);
 550         kfree(buf);
 551         return ret;
 552 }
 553
 554 /* stripe bucket accounting: */
 555
 556 static int __ec_stripe_mem_alloc(struct bch_fs *c, size_t idx, gfp_t gfp)
 557 {
 558         ec_stripes_heap n, *h = &c->ec_stripes_heap;
 559
 560         if (idx >= h->size) {
 561                 if (!init_heap(&n, max(1024UL, roundup_pow_of_two(idx + 1)), gfp))
 562                         return -ENOMEM;
 563
 564                 mutex_lock(&c->ec_stripes_heap_lock);
 565                 if (n.size > h->size) {
 566                         memcpy(n.data, h->data, h->used * sizeof(h->data[0]));
 567                         n.used = h->used;
 568                         swap(*h, n);
 569                 }
 570                 mutex_unlock(&c->ec_stripes_heap_lock);
 571
 572                 free_heap(&n);
 573         }
 574
 575         if (!genradix_ptr_alloc(&c->stripes, idx, gfp))
 576                 return -ENOMEM;
 577
 578         if (c->gc_pos.phase != GC_PHASE_NOT_RUNNING &&
 579             !genradix_ptr_alloc(&c->gc_stripes, idx, gfp))
 580                 return -ENOMEM;
 581
 582         return 0;
 583 }
 584
 585 static int ec_stripe_mem_alloc(struct btree_trans *trans,
 586                                struct btree_iter *iter)
 587 {
 588         size_t idx = iter->pos.offset;
 589
 590         if (!__ec_stripe_mem_alloc(trans->c, idx, GFP_NOWAIT|__GFP_NOWARN))
 591                 return 0;
 592
 593         bch2_trans_unlock(trans);
 594
 595         return   __ec_stripe_mem_alloc(trans->c, idx, GFP_KERNEL) ?:
 596                 bch2_trans_relock(trans);
 597 }
 598
 599 /*
 600  * Hash table of open stripes:
 601  * Stripes that are being created or modified are kept in a hash table, so that
 602  * stripe deletion can skip them.
 603  */
 604
 605 static bool __bch2_stripe_is_open(struct bch_fs *c, u64 idx)
 606 {
 607         unsigned hash = hash_64(idx, ilog2(ARRAY_SIZE(c->ec_stripes_new)));
 608         struct ec_stripe_new *s;
 609
 610         hlist_for_each_entry(s, &c->ec_stripes_new[hash], hash)
 611                 if (s->idx == idx)
 612                         return true;
 613         return false;
 614 }
 615
 616 static bool bch2_stripe_is_open(struct bch_fs *c, u64 idx)
 617 {
 618         bool ret = false;
 619
 620         spin_lock(&c->ec_stripes_new_lock);
 621         ret = __bch2_stripe_is_open(c, idx);
 622         spin_unlock(&c->ec_stripes_new_lock);
 623
 624         return ret;
 625 }
 626
 627 static bool bch2_try_open_stripe(struct bch_fs *c,
 628                                  struct ec_stripe_new *s,
 629                                  u64 idx)
 630 {
 631         bool ret;
 632
 633         spin_lock(&c->ec_stripes_new_lock);
 634         ret = !__bch2_stripe_is_open(c, idx);
 635         if (ret) {
 636                 unsigned hash = hash_64(idx, ilog2(ARRAY_SIZE(c->ec_stripes_new)));
 637
 638                 s->idx = idx;
 639                 hlist_add_head(&s->hash, &c->ec_stripes_new[hash]);
 640         }
 641         spin_unlock(&c->ec_stripes_new_lock);
 642
 643         return ret;
 644 }
 645
 646 static void bch2_stripe_close(struct bch_fs *c, struct ec_stripe_new *s)
 647 {
 648         BUG_ON(!s->idx);
 649
 650         spin_lock(&c->ec_stripes_new_lock);
 651         hlist_del_init(&s->hash);
 652         spin_unlock(&c->ec_stripes_new_lock);
 653
 654         s->idx = 0;
 655 }
 656
 657 /* Heap of all existing stripes, ordered by blocks_nonempty */
 658
 659 static u64 stripe_idx_to_delete(struct bch_fs *c)
 660 {
 661         ec_stripes_heap *h = &c->ec_stripes_heap;
 662         size_t heap_idx;
 663
 664         lockdep_assert_held(&c->ec_stripes_heap_lock);
 665
 666         for (heap_idx = 0; heap_idx < h->used; heap_idx++)
 667                 if (h->data[heap_idx].blocks_nonempty == 0 &&
 668                     !bch2_stripe_is_open(c, h->data[heap_idx].idx))
 669                         return h->data[heap_idx].idx;
 670
 671         return 0;
 672 }
 673
 674 static inline int ec_stripes_heap_cmp(ec_stripes_heap *h,
 675                                       struct ec_stripe_heap_entry l,
 676                                       struct ec_stripe_heap_entry r)
 677 {
 678         return ((l.blocks_nonempty > r.blocks_nonempty) -
 679                 (l.blocks_nonempty < r.blocks_nonempty));
 680 }
 681
 682 static inline void ec_stripes_heap_set_backpointer(ec_stripes_heap *h,
 683                                                    size_t i)
 684 {
 685         struct bch_fs *c = container_of(h, struct bch_fs, ec_stripes_heap);
 686
 687         genradix_ptr(&c->stripes, h->data[i].idx)->heap_idx = i;
 688 }
 689
 690 static void heap_verify_backpointer(struct bch_fs *c, size_t idx)
 691 {
 692         ec_stripes_heap *h = &c->ec_stripes_heap;
 693         struct stripe *m = genradix_ptr(&c->stripes, idx);
 694
 695         BUG_ON(m->heap_idx >= h->used);
 696         BUG_ON(h->data[m->heap_idx].idx != idx);
 697 }
 698
 699 void bch2_stripes_heap_del(struct bch_fs *c,
 700                            struct stripe *m, size_t idx)
 701 {
 702         mutex_lock(&c->ec_stripes_heap_lock);
 703         heap_verify_backpointer(c, idx);
 704
 705         heap_del(&c->ec_stripes_heap, m->heap_idx,
 706                  ec_stripes_heap_cmp,
 707                  ec_stripes_heap_set_backpointer);
 708         mutex_unlock(&c->ec_stripes_heap_lock);
 709 }
 710
 711 void bch2_stripes_heap_insert(struct bch_fs *c,
 712                               struct stripe *m, size_t idx)
 713 {
 714         mutex_lock(&c->ec_stripes_heap_lock);
 715         BUG_ON(heap_full(&c->ec_stripes_heap));
 716
 717         heap_add(&c->ec_stripes_heap, ((struct ec_stripe_heap_entry) {
 718                         .idx = idx,
 719                         .blocks_nonempty = m->blocks_nonempty,
 720                 }),
 721                  ec_stripes_heap_cmp,
 722                  ec_stripes_heap_set_backpointer);
 723
 724         heap_verify_backpointer(c, idx);
 725         mutex_unlock(&c->ec_stripes_heap_lock);
 726 }
 727
 728 void bch2_stripes_heap_update(struct bch_fs *c,
 729                               struct stripe *m, size_t idx)
 730 {
 731         ec_stripes_heap *h = &c->ec_stripes_heap;
 732         bool do_deletes;
 733         size_t i;
 734
 735         mutex_lock(&c->ec_stripes_heap_lock);
 736         heap_verify_backpointer(c, idx);
 737
 738         h->data[m->heap_idx].blocks_nonempty = m->blocks_nonempty;
 739
 740         i = m->heap_idx;
 741         heap_sift_up(h,   i, ec_stripes_heap_cmp,
 742                      ec_stripes_heap_set_backpointer);
 743         heap_sift_down(h, i, ec_stripes_heap_cmp,
 744                        ec_stripes_heap_set_backpointer);
 745
 746         heap_verify_backpointer(c, idx);
 747
 748         do_deletes = stripe_idx_to_delete(c) != 0;
 749         mutex_unlock(&c->ec_stripes_heap_lock);
 750
 751         if (do_deletes)
 752                 bch2_do_stripe_deletes(c);
 753 }
 754
 755 /* stripe deletion */
 756
 757 static int ec_stripe_delete(struct btree_trans *trans, u64 idx)
 758 {
 759         struct bch_fs *c = trans->c;
 760         struct btree_iter iter;
 761         struct bkey_s_c k;
 762         struct bkey_s_c_stripe s;
 763         int ret;
 764
 765         bch2_trans_iter_init(trans, &iter, BTREE_ID_stripes, POS(0, idx),
 766                              BTREE_ITER_INTENT);
 767         k = bch2_btree_iter_peek_slot(&iter);
 768         ret = bkey_err(k);
 769         if (ret)
 770                 goto err;
 771
 772         if (k.k->type != KEY_TYPE_stripe) {
 773                 bch2_fs_inconsistent(c, "attempting to delete nonexistent stripe %llu", idx);
 774                 ret = -EINVAL;
 775                 goto err;
 776         }
 777
 778         s = bkey_s_c_to_stripe(k);
 779         for (unsigned i = 0; i < s.v->nr_blocks; i++)
 780                 if (stripe_blockcount_get(s.v, i)) {
 781                         struct printbuf buf = PRINTBUF;
 782
 783                         bch2_bkey_val_to_text(&buf, c, k);
 784                         bch2_fs_inconsistent(c, "attempting to delete nonempty stripe %s", buf.buf);
 785                         printbuf_exit(&buf);
 786                         ret = -EINVAL;
 787                         goto err;
 788                 }
 789
 790         ret = bch2_btree_delete_at(trans, &iter, 0);
 791 err:
 792         bch2_trans_iter_exit(trans, &iter);
 793         return ret;
 794 }
 795
 796 static void ec_stripe_delete_work(struct work_struct *work)
 797 {
 798         struct bch_fs *c =
 799                 container_of(work, struct bch_fs, ec_stripe_delete_work);
 800         struct btree_trans trans;
 801         int ret;
 802         u64 idx;
 803
 804         bch2_trans_init(&trans, c, 0, 0);
 805
 806         while (1) {
 807                 mutex_lock(&c->ec_stripes_heap_lock);
 808                 idx = stripe_idx_to_delete(c);
 809                 mutex_unlock(&c->ec_stripes_heap_lock);
 810
 811                 if (!idx)
 812                         break;
 813
 814                 ret = commit_do(&trans, NULL, NULL, BTREE_INSERT_NOFAIL,
 815                                 ec_stripe_delete(&trans, idx));
 816                 if (ret) {
 817                         bch_err(c, "%s: err %s", __func__, bch2_err_str(ret));
 818                         break;
 819                 }
 820         }
 821
 822         bch2_trans_exit(&trans);
 823
 824         bch2_write_ref_put(c, BCH_WRITE_REF_stripe_delete);
 825 }
 826
 827 void bch2_do_stripe_deletes(struct bch_fs *c)
 828 {
 829         if (bch2_write_ref_tryget(c, BCH_WRITE_REF_stripe_delete) &&
 830             !schedule_work(&c->ec_stripe_delete_work))
 831                 bch2_write_ref_put(c, BCH_WRITE_REF_stripe_delete);
 832 }
 833
 834 /* stripe creation: */
 835
 836 static int ec_stripe_key_update(struct btree_trans *trans,
 837                                 struct bkey_i_stripe *new,
 838                                 bool create)
 839 {
 840         struct bch_fs *c = trans->c;
 841         struct btree_iter iter;
 842         struct bkey_s_c k;
 843         int ret;
 844
 845         bch2_trans_iter_init(trans, &iter, BTREE_ID_stripes,
 846                              new->k.p, BTREE_ITER_INTENT);
 847         k = bch2_btree_iter_peek_slot(&iter);
 848         ret = bkey_err(k);
 849         if (ret)
 850                 goto err;
 851
 852         if (k.k->type != (create ? KEY_TYPE_deleted : KEY_TYPE_stripe)) {
 853                 bch2_fs_inconsistent(c, "error %s stripe: got existing key type %s",
 854                                      create ? "creating" : "updating",
 855                                      bch2_bkey_types[k.k->type]);
 856                 ret = -EINVAL;
 857                 goto err;
 858         }
 859
 860         if (k.k->type == KEY_TYPE_stripe) {
 861                 const struct bch_stripe *old = bkey_s_c_to_stripe(k).v;
 862                 unsigned i;
 863
 864                 if (old->nr_blocks != new->v.nr_blocks) {
 865                         bch_err(c, "error updating stripe: nr_blocks does not match");
 866                         ret = -EINVAL;
 867                         goto err;
 868                 }
 869
 870                 for (i = 0; i < new->v.nr_blocks; i++) {
 871                         unsigned v = stripe_blockcount_get(old, i);
 872
 873                         BUG_ON(v &&
 874                                (old->ptrs[i].dev != new->v.ptrs[i].dev ||
 875                                 old->ptrs[i].gen != new->v.ptrs[i].gen ||
 876                                 old->ptrs[i].offset != new->v.ptrs[i].offset));
 877
 878                         stripe_blockcount_set(&new->v, i, v);
 879                 }
 880         }
 881
 882         ret = bch2_trans_update(trans, &iter, &new->k_i, 0);
 883 err:
 884         bch2_trans_iter_exit(trans, &iter);
 885         return ret;
 886 }
 887
 888 static int ec_stripe_update_extent(struct btree_trans *trans,
 889                                    struct bpos bucket, u8 gen,
 890                                    struct ec_stripe_buf *s,
 891                                    u64 *bp_offset)
 892 {
 893         struct bch_fs *c = trans->c;
 894         struct bch_backpointer bp;
 895         struct btree_iter iter;
 896         struct bkey_s_c k;
 897         const struct bch_extent_ptr *ptr_c;
 898         struct bch_extent_ptr *ptr, *ec_ptr = NULL;
 899         struct bch_extent_stripe_ptr stripe_ptr;
 900         struct bkey_i *n;
 901         int ret, dev, block;
 902
 903         ret = bch2_get_next_backpointer(trans, bucket, gen,
 904                                 bp_offset, &bp, BTREE_ITER_CACHED);
 905         if (ret)
 906                 return ret;
 907         if (*bp_offset == U64_MAX)
 908                 return 0;
 909
 910         if (bp.level) {
 911                 struct printbuf buf = PRINTBUF;
 912                 struct btree_iter node_iter;
 913                 struct btree *b;
 914
 915                 b = bch2_backpointer_get_node(trans, &node_iter, bucket, *bp_offset, bp);
 916                 bch2_trans_iter_exit(trans, &node_iter);
 917
 918                 if (!b)
 919                         return 0;
 920
 921                 prt_printf(&buf, "found btree node in erasure coded bucket: b=%px\n", b);
 922                 bch2_backpointer_to_text(&buf, &bp);
 923
 924                 bch2_fs_inconsistent(c, "%s", buf.buf);
 925                 printbuf_exit(&buf);
 926                 return -EIO;
 927         }
 928
 929         k = bch2_backpointer_get_key(trans, &iter, bucket, *bp_offset, bp);
 930         ret = bkey_err(k);
 931         if (ret)
 932                 return ret;
 933         if (!k.k) {
 934                 /*
 935                  * extent no longer exists - we could flush the btree
 936                  * write buffer and retry to verify, but no need:
 937                  */
 938                 return 0;
 939         }
 940
 941         if (extent_has_stripe_ptr(k, s->key.k.p.offset))
 942                 goto out;
 943
 944         ptr_c = bkey_matches_stripe(&s->key.v, k, &block);
 945         /*
 946          * It doesn't generally make sense to erasure code cached ptrs:
 947          * XXX: should we be incrementing a counter?
 948          */
 949         if (!ptr_c || ptr_c->cached)
 950                 goto out;
 951
 952         dev = s->key.v.ptrs[block].dev;
 953
 954         n = bch2_trans_kmalloc(trans, bkey_bytes(k.k) + sizeof(stripe_ptr));
 955         ret = PTR_ERR_OR_ZERO(n);
 956         if (ret)
 957                 goto out;
 958
 959         bkey_reassemble(n, k);
 960
 961         bch2_bkey_drop_ptrs(bkey_i_to_s(n), ptr, ptr->dev != dev);
 962         ec_ptr = (void *) bch2_bkey_has_device(bkey_i_to_s_c(n), dev);
 963         BUG_ON(!ec_ptr);
 964
 965         stripe_ptr = (struct bch_extent_stripe_ptr) {
 966                 .type = 1 << BCH_EXTENT_ENTRY_stripe_ptr,
 967                 .block          = block,
 968                 .redundancy     = s->key.v.nr_redundant,
 969                 .idx            = s->key.k.p.offset,
 970         };
 971
 972         __extent_entry_insert(n,
 973                         (union bch_extent_entry *) ec_ptr,
 974                         (union bch_extent_entry *) &stripe_ptr);
 975
 976         ret = bch2_trans_update(trans, &iter, n, 0);
 977 out:
 978         bch2_trans_iter_exit(trans, &iter);
 979         return ret;
 980 }
 981
 982 static int ec_stripe_update_bucket(struct btree_trans *trans, struct ec_stripe_buf *s,
 983                                    unsigned block)
 984 {
 985         struct bch_fs *c = trans->c;
 986         struct bch_extent_ptr bucket = s->key.v.ptrs[block];
 987         struct bpos bucket_pos = PTR_BUCKET_POS(c, &bucket);
 988         u64 bp_offset = 0;
 989         int ret = 0;
 990
 991         while (1) {
 992                 ret = commit_do(trans, NULL, NULL,
 993                                 BTREE_INSERT_NOFAIL,
 994                         ec_stripe_update_extent(trans, bucket_pos, bucket.gen,
 995                                                 s, &bp_offset));
 996                 if (ret)
 997                         break;
 998                 if (bp_offset == U64_MAX)
 999                         break;
1000
1001                 bp_offset++;
1002         }
1003
1004         return ret;
1005 }
1006
1007 static int ec_stripe_update_extents(struct bch_fs *c, struct ec_stripe_buf *s)
1008 {
1009         struct btree_trans trans;
1010         struct bch_stripe *v = &s->key.v;
1011         unsigned i, nr_data = v->nr_blocks - v->nr_redundant;
1012         int ret = 0;
1013
1014         bch2_trans_init(&trans, c, 0, 0);
1015
1016         ret = bch2_btree_write_buffer_flush(&trans);
1017         if (ret)
1018                 goto err;
1019
1020         for (i = 0; i < nr_data; i++) {
1021                 ret = ec_stripe_update_bucket(&trans, s, i);
1022                 if (ret)
1023                         break;
1024         }
1025 err:
1026         bch2_trans_exit(&trans);
1027
1028         return ret;
1029 }
1030
1031 static void zero_out_rest_of_ec_bucket(struct bch_fs *c,
1032                                        struct ec_stripe_new *s,
1033                                        unsigned block,
1034                                        struct open_bucket *ob)
1035 {
1036         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1037         unsigned offset = ca->mi.bucket_size - ob->sectors_free;
1038         int ret;
1039
1040         if (!bch2_dev_get_ioref(ca, WRITE)) {
1041                 s->err = -EROFS;
1042                 return;
1043         }
1044
1045         memset(s->new_stripe.data[block] + (offset << 9),
1046                0,
1047                ob->sectors_free << 9);
1048
1049         ret = blkdev_issue_zeroout(ca->disk_sb.bdev,
1050                         ob->bucket * ca->mi.bucket_size + offset,
1051                         ob->sectors_free,
1052                         GFP_KERNEL, 0);
1053
1054         percpu_ref_put(&ca->io_ref);
1055
1056         if (ret)
1057                 s->err = ret;
1058 }
1059
1060 /*
1061  * data buckets of new stripe all written: create the stripe
1062  */
1063 static void ec_stripe_create(struct ec_stripe_new *s)
1064 {
1065         struct bch_fs *c = s->c;
1066         struct open_bucket *ob;
1067         struct bch_stripe *v = &s->new_stripe.key.v;
1068         unsigned i, nr_data = v->nr_blocks - v->nr_redundant;
1069         int ret;
1070
1071         BUG_ON(s->h->s == s);
1072
1073         closure_sync(&s->iodone);
1074
1075         for (i = 0; i < nr_data; i++)
1076                 if (s->blocks[i]) {
1077                         ob = c->open_buckets + s->blocks[i];
1078
1079                         if (ob->sectors_free)
1080                                 zero_out_rest_of_ec_bucket(c, s, i, ob);
1081                 }
1082
1083         if (s->err) {
1084                 if (!bch2_err_matches(s->err, EROFS))
1085                         bch_err(c, "error creating stripe: error writing data buckets");
1086                 goto err;
1087         }
1088
1089         if (s->have_existing_stripe) {
1090                 ec_validate_checksums(c, &s->existing_stripe);
1091
1092                 if (ec_do_recov(c, &s->existing_stripe)) {
1093                         bch_err(c, "error creating stripe: error reading existing stripe");
1094                         goto err;
1095                 }
1096
1097                 for (i = 0; i < nr_data; i++)
1098                         if (stripe_blockcount_get(&s->existing_stripe.key.v, i))
1099                                 swap(s->new_stripe.data[i],
1100                                      s->existing_stripe.data[i]);
1101
1102                 ec_stripe_buf_exit(&s->existing_stripe);
1103         }
1104
1105         BUG_ON(!s->allocated);
1106         BUG_ON(!s->idx);
1107
1108         ec_generate_ec(&s->new_stripe);
1109
1110         ec_generate_checksums(&s->new_stripe);
1111
1112         /* write p/q: */
1113         for (i = nr_data; i < v->nr_blocks; i++)
1114                 ec_block_io(c, &s->new_stripe, REQ_OP_WRITE, i, &s->iodone);
1115         closure_sync(&s->iodone);
1116
1117         if (ec_nr_failed(&s->new_stripe)) {
1118                 bch_err(c, "error creating stripe: error writing redundancy buckets");
1119                 goto err;
1120         }
1121
1122         ret = bch2_trans_do(c, &s->res, NULL, BTREE_INSERT_NOFAIL,
1123                             ec_stripe_key_update(&trans, &s->new_stripe.key,
1124                                                  !s->have_existing_stripe));
1125         if (ret) {
1126                 bch_err(c, "error creating stripe: error creating stripe key");
1127                 goto err;
1128         }
1129
1130         ret = ec_stripe_update_extents(c, &s->new_stripe);
1131         if (ret) {
1132                 bch_err(c, "error creating stripe: error updating pointers: %s",
1133                         bch2_err_str(ret));
1134                 goto err;
1135         }
1136 err:
1137         bch2_disk_reservation_put(c, &s->res);
1138
1139         for (i = 0; i < v->nr_blocks; i++)
1140                 if (s->blocks[i]) {
1141                         ob = c->open_buckets + s->blocks[i];
1142
1143                         if (i < nr_data) {
1144                                 ob->ec = NULL;
1145                                 __bch2_open_bucket_put(c, ob);
1146                         } else {
1147                                 bch2_open_bucket_put(c, ob);
1148                         }
1149                 }
1150
1151         mutex_lock(&c->ec_stripe_new_lock);
1152         list_del(&s->list);
1153         mutex_unlock(&c->ec_stripe_new_lock);
1154
1155         if (s->idx)
1156                 bch2_stripe_close(c, s);
1157
1158         ec_stripe_buf_exit(&s->existing_stripe);
1159         ec_stripe_buf_exit(&s->new_stripe);
1160         closure_debug_destroy(&s->iodone);
1161         kfree(s);
1162 }
1163
1164 static struct ec_stripe_new *get_pending_stripe(struct bch_fs *c)
1165 {
1166         struct ec_stripe_new *s;
1167
1168         mutex_lock(&c->ec_stripe_new_lock);
1169         list_for_each_entry(s, &c->ec_stripe_new_list, list)
1170                 if (!atomic_read(&s->pin))
1171                         goto out;
1172         s = NULL;
1173 out:
1174         mutex_unlock(&c->ec_stripe_new_lock);
1175
1176         return s;
1177 }
1178
1179 static void ec_stripe_create_work(struct work_struct *work)
1180 {
1181         struct bch_fs *c = container_of(work,
1182                 struct bch_fs, ec_stripe_create_work);
1183         struct ec_stripe_new *s;
1184
1185         while ((s = get_pending_stripe(c)))
1186                 ec_stripe_create(s);
1187
1188         bch2_write_ref_put(c, BCH_WRITE_REF_stripe_create);
1189 }
1190
1191 void bch2_ec_do_stripe_creates(struct bch_fs *c)
1192 {
1193         bch2_write_ref_get(c, BCH_WRITE_REF_stripe_create);
1194
1195         if (!queue_work(system_long_wq, &c->ec_stripe_create_work))
1196                 bch2_write_ref_put(c, BCH_WRITE_REF_stripe_create);
1197 }
1198
1199 static void ec_stripe_set_pending(struct bch_fs *c, struct ec_stripe_head *h)
1200 {
1201         struct ec_stripe_new *s = h->s;
1202
1203         BUG_ON(!s->allocated && !s->err);
1204
1205         h->s            = NULL;
1206         s->pending      = true;
1207
1208         mutex_lock(&c->ec_stripe_new_lock);
1209         list_add(&s->list, &c->ec_stripe_new_list);
1210         mutex_unlock(&c->ec_stripe_new_lock);
1211
1212         ec_stripe_new_put(c, s);
1213 }
1214
1215 void bch2_ec_bucket_cancel(struct bch_fs *c, struct open_bucket *ob)
1216 {
1217         struct ec_stripe_new *s = ob->ec;
1218
1219         s->err = -EIO;
1220 }
1221
1222 void *bch2_writepoint_ec_buf(struct bch_fs *c, struct write_point *wp)
1223 {
1224         struct open_bucket *ob = ec_open_bucket(c, &wp->ptrs);
1225         struct bch_dev *ca;
1226         unsigned offset;
1227
1228         if (!ob)
1229                 return NULL;
1230
1231         BUG_ON(!ob->ec->new_stripe.data[ob->ec_idx]);
1232
1233         ca      = bch_dev_bkey_exists(c, ob->dev);
1234         offset  = ca->mi.bucket_size - ob->sectors_free;
1235
1236         return ob->ec->new_stripe.data[ob->ec_idx] + (offset << 9);
1237 }
1238
1239 static int unsigned_cmp(const void *_l, const void *_r)
1240 {
1241         unsigned l = *((const unsigned *) _l);
1242         unsigned r = *((const unsigned *) _r);
1243
1244         return cmp_int(l, r);
1245 }
1246
1247 /* pick most common bucket size: */
1248 static unsigned pick_blocksize(struct bch_fs *c,
1249                                struct bch_devs_mask *devs)
1250 {
1251         struct bch_dev *ca;
1252         unsigned i, nr = 0, sizes[BCH_SB_MEMBERS_MAX];
1253         struct {
1254                 unsigned nr, size;
1255         } cur = { 0, 0 }, best = { 0, 0 };
1256
1257         for_each_member_device_rcu(ca, c, i, devs)
1258                 sizes[nr++] = ca->mi.bucket_size;
1259
1260         sort(sizes, nr, sizeof(unsigned), unsigned_cmp, NULL);
1261
1262         for (i = 0; i < nr; i++) {
1263                 if (sizes[i] != cur.size) {
1264                         if (cur.nr > best.nr)
1265                                 best = cur;
1266
1267                         cur.nr = 0;
1268                         cur.size = sizes[i];
1269                 }
1270
1271                 cur.nr++;
1272         }
1273
1274         if (cur.nr > best.nr)
1275                 best = cur;
1276
1277         return best.size;
1278 }
1279
1280 static bool may_create_new_stripe(struct bch_fs *c)
1281 {
1282         return false;
1283 }
1284
1285 static void ec_stripe_key_init(struct bch_fs *c,
1286                                struct bkey_i_stripe *s,
1287                                unsigned nr_data,
1288                                unsigned nr_parity,
1289                                unsigned stripe_size)
1290 {
1291         unsigned u64s;
1292
1293         bkey_stripe_init(&s->k_i);
1294         s->v.sectors                    = cpu_to_le16(stripe_size);
1295         s->v.algorithm                  = 0;
1296         s->v.nr_blocks                  = nr_data + nr_parity;
1297         s->v.nr_redundant               = nr_parity;
1298         s->v.csum_granularity_bits      = ilog2(c->opts.encoded_extent_max >> 9);
1299         s->v.csum_type                  = BCH_CSUM_crc32c;
1300         s->v.pad                        = 0;
1301
1302         while ((u64s = stripe_val_u64s(&s->v)) > BKEY_VAL_U64s_MAX) {
1303                 BUG_ON(1 << s->v.csum_granularity_bits >=
1304                        le16_to_cpu(s->v.sectors) ||
1305                        s->v.csum_granularity_bits == U8_MAX);
1306                 s->v.csum_granularity_bits++;
1307         }
1308
1309         set_bkey_val_u64s(&s->k, u64s);
1310 }
1311
1312 static int ec_new_stripe_alloc(struct bch_fs *c, struct ec_stripe_head *h)
1313 {
1314         struct ec_stripe_new *s;
1315
1316         lockdep_assert_held(&h->lock);
1317
1318         s = kzalloc(sizeof(*s), GFP_KERNEL);
1319         if (!s)
1320                 return -ENOMEM;
1321
1322         mutex_init(&s->lock);
1323         closure_init(&s->iodone, NULL);
1324         atomic_set(&s->pin, 1);
1325         s->c            = c;
1326         s->h            = h;
1327         s->nr_data      = min_t(unsigned, h->nr_active_devs,
1328                                 BCH_BKEY_PTRS_MAX) - h->redundancy;
1329         s->nr_parity    = h->redundancy;
1330
1331         ec_stripe_key_init(c, &s->new_stripe.key, s->nr_data,
1332                            s->nr_parity, h->blocksize);
1333
1334         h->s = s;
1335         return 0;
1336 }
1337
1338 static struct ec_stripe_head *
1339 ec_new_stripe_head_alloc(struct bch_fs *c, unsigned target,
1340                          unsigned algo, unsigned redundancy,
1341                          enum alloc_reserve reserve)
1342 {
1343         struct ec_stripe_head *h;
1344         struct bch_dev *ca;
1345         unsigned i;
1346
1347         h = kzalloc(sizeof(*h), GFP_KERNEL);
1348         if (!h)
1349                 return NULL;
1350
1351         mutex_init(&h->lock);
1352         BUG_ON(!mutex_trylock(&h->lock));
1353
1354         h->target       = target;
1355         h->algo         = algo;
1356         h->redundancy   = redundancy;
1357         h->reserve      = reserve;
1358
1359         rcu_read_lock();
1360         h->devs = target_rw_devs(c, BCH_DATA_user, target);
1361
1362         for_each_member_device_rcu(ca, c, i, &h->devs)
1363                 if (!ca->mi.durability)
1364                         __clear_bit(i, h->devs.d);
1365
1366         h->blocksize = pick_blocksize(c, &h->devs);
1367
1368         for_each_member_device_rcu(ca, c, i, &h->devs)
1369                 if (ca->mi.bucket_size == h->blocksize)
1370                         h->nr_active_devs++;
1371
1372         rcu_read_unlock();
1373         list_add(&h->list, &c->ec_stripe_head_list);
1374         return h;
1375 }
1376
1377 void bch2_ec_stripe_head_put(struct bch_fs *c, struct ec_stripe_head *h)
1378 {
1379         if (h->s &&
1380             h->s->allocated &&
1381             bitmap_weight(h->s->blocks_allocated,
1382                           h->s->nr_data) == h->s->nr_data)
1383                 ec_stripe_set_pending(c, h);
1384
1385         mutex_unlock(&h->lock);
1386 }
1387
1388 struct ec_stripe_head *__bch2_ec_stripe_head_get(struct btree_trans *trans,
1389                                                  unsigned target,
1390                                                  unsigned algo,
1391                                                  unsigned redundancy,
1392                                                  enum alloc_reserve reserve)
1393 {
1394         struct bch_fs *c = trans->c;
1395         struct ec_stripe_head *h;
1396         int ret;
1397
1398         if (!redundancy)
1399                 return NULL;
1400
1401         ret = bch2_trans_mutex_lock(trans, &c->ec_stripe_head_lock);
1402         if (ret)
1403                 return ERR_PTR(ret);
1404
1405         list_for_each_entry(h, &c->ec_stripe_head_list, list)
1406                 if (h->target           == target &&
1407                     h->algo             == algo &&
1408                     h->redundancy       == redundancy &&
1409                     h->reserve          == reserve) {
1410                         ret = bch2_trans_mutex_lock(trans, &h->lock);
1411                         if (ret)
1412                                 h = ERR_PTR(ret);
1413                         goto found;
1414                 }
1415
1416         h = ec_new_stripe_head_alloc(c, target, algo, redundancy, reserve);
1417 found:
1418         mutex_unlock(&c->ec_stripe_head_lock);
1419         return h;
1420 }
1421
1422 static int new_stripe_alloc_buckets(struct btree_trans *trans, struct ec_stripe_head *h,
1423                                     enum alloc_reserve reserve, struct closure *cl)
1424 {
1425         struct bch_fs *c = trans->c;
1426         struct bch_devs_mask devs = h->devs;
1427         struct open_bucket *ob;
1428         struct open_buckets buckets;
1429         unsigned i, j, nr_have_parity = 0, nr_have_data = 0;
1430         bool have_cache = true;
1431         int ret = 0;
1432
1433         BUG_ON(h->s->new_stripe.key.v.nr_blocks         != h->s->nr_data + h->s->nr_parity);
1434         BUG_ON(h->s->new_stripe.key.v.nr_redundant      != h->s->nr_parity);
1435
1436         for_each_set_bit(i, h->s->blocks_gotten, h->s->new_stripe.key.v.nr_blocks) {
1437                 __clear_bit(h->s->new_stripe.key.v.ptrs[i].dev, devs.d);
1438                 if (i < h->s->nr_data)
1439                         nr_have_data++;
1440                 else
1441                         nr_have_parity++;
1442         }
1443
1444         BUG_ON(nr_have_data     > h->s->nr_data);
1445         BUG_ON(nr_have_parity   > h->s->nr_parity);
1446
1447         buckets.nr = 0;
1448         if (nr_have_parity < h->s->nr_parity) {
1449                 ret = bch2_bucket_alloc_set_trans(trans, &buckets,
1450                                             &h->parity_stripe,
1451                                             &devs,
1452                                             h->s->nr_parity,
1453                                             &nr_have_parity,
1454                                             &have_cache,
1455                                             BCH_DATA_parity,
1456                                             reserve,
1457                                             cl);
1458
1459                 open_bucket_for_each(c, &buckets, ob, i) {
1460                         j = find_next_zero_bit(h->s->blocks_gotten,
1461                                                h->s->nr_data + h->s->nr_parity,
1462                                                h->s->nr_data);
1463                         BUG_ON(j >= h->s->nr_data + h->s->nr_parity);
1464
1465                         h->s->blocks[j] = buckets.v[i];
1466                         h->s->new_stripe.key.v.ptrs[j] = bch2_ob_ptr(c, ob);
1467                         __set_bit(j, h->s->blocks_gotten);
1468                 }
1469
1470                 if (ret)
1471                         return ret;
1472         }
1473
1474         buckets.nr = 0;
1475         if (nr_have_data < h->s->nr_data) {
1476                 ret = bch2_bucket_alloc_set_trans(trans, &buckets,
1477                                             &h->block_stripe,
1478                                             &devs,
1479                                             h->s->nr_data,
1480                                             &nr_have_data,
1481                                             &have_cache,
1482                                             BCH_DATA_user,
1483                                             reserve,
1484                                             cl);
1485
1486                 open_bucket_for_each(c, &buckets, ob, i) {
1487                         j = find_next_zero_bit(h->s->blocks_gotten,
1488                                                h->s->nr_data, 0);
1489                         BUG_ON(j >= h->s->nr_data);
1490
1491                         h->s->blocks[j] = buckets.v[i];
1492                         h->s->new_stripe.key.v.ptrs[j] = bch2_ob_ptr(c, ob);
1493                         __set_bit(j, h->s->blocks_gotten);
1494                 }
1495
1496                 if (ret)
1497                         return ret;
1498         }
1499
1500         return 0;
1501 }
1502
1503 /* XXX: doesn't obey target: */
1504 static s64 get_existing_stripe(struct bch_fs *c,
1505                                struct ec_stripe_head *head)
1506 {
1507         ec_stripes_heap *h = &c->ec_stripes_heap;
1508         struct stripe *m;
1509         size_t heap_idx;
1510         u64 stripe_idx;
1511         s64 ret = -1;
1512
1513         if (may_create_new_stripe(c))
1514                 return -1;
1515
1516         mutex_lock(&c->ec_stripes_heap_lock);
1517         for (heap_idx = 0; heap_idx < h->used; heap_idx++) {
1518                 /* No blocks worth reusing, stripe will just be deleted: */
1519                 if (!h->data[heap_idx].blocks_nonempty)
1520                         continue;
1521
1522                 stripe_idx = h->data[heap_idx].idx;
1523
1524                 m = genradix_ptr(&c->stripes, stripe_idx);
1525
1526                 if (m->algorithm        == head->algo &&
1527                     m->nr_redundant     == head->redundancy &&
1528                     m->sectors          == head->blocksize &&
1529                     m->blocks_nonempty  < m->nr_blocks - m->nr_redundant &&
1530                     bch2_try_open_stripe(c, head->s, stripe_idx)) {
1531                         ret = stripe_idx;
1532                         break;
1533                 }
1534         }
1535         mutex_unlock(&c->ec_stripes_heap_lock);
1536         return ret;
1537 }
1538
1539 static int __bch2_ec_stripe_head_reuse(struct btree_trans *trans, struct ec_stripe_head *h)
1540 {
1541         struct bch_fs *c = trans->c;
1542         unsigned i;
1543         s64 idx;
1544         int ret;
1545
1546         /*
1547          * If we can't allocate a new stripe, and there's no stripes with empty
1548          * blocks for us to reuse, that means we have to wait on copygc:
1549          */
1550         idx = get_existing_stripe(c, h);
1551         if (idx < 0)
1552                 return -BCH_ERR_stripe_alloc_blocked;
1553
1554         ret = get_stripe_key_trans(trans, idx, &h->s->existing_stripe);
1555         if (ret) {
1556                 bch2_stripe_close(c, h->s);
1557                 if (!bch2_err_matches(ret, BCH_ERR_transaction_restart))
1558                         bch2_fs_fatal_error(c, "error reading stripe key: %s", bch2_err_str(ret));
1559                 return ret;
1560         }
1561
1562         BUG_ON(h->s->existing_stripe.key.v.nr_redundant != h->s->nr_parity);
1563         h->s->nr_data = h->s->existing_stripe.key.v.nr_blocks -
1564                 h->s->existing_stripe.key.v.nr_redundant;
1565
1566         ret = ec_stripe_buf_init(&h->s->existing_stripe, 0, h->blocksize);
1567         if (ret) {
1568                 bch2_stripe_close(c, h->s);
1569                 return ret;
1570         }
1571
1572         BUG_ON(h->s->existing_stripe.size != h->blocksize);
1573         BUG_ON(h->s->existing_stripe.size != h->s->existing_stripe.key.v.sectors);
1574
1575         /*
1576          * Free buckets we initially allocated - they might conflict with
1577          * blocks from the stripe we're reusing:
1578          */
1579         for_each_set_bit(i, h->s->blocks_gotten, h->s->new_stripe.key.v.nr_blocks) {
1580                 bch2_open_bucket_put(c, c->open_buckets + h->s->blocks[i]);
1581                 h->s->blocks[i] = 0;
1582         }
1583         memset(h->s->blocks_gotten, 0, sizeof(h->s->blocks_gotten));
1584         memset(h->s->blocks_allocated, 0, sizeof(h->s->blocks_allocated));
1585
1586         for (i = 0; i < h->s->existing_stripe.key.v.nr_blocks; i++) {
1587                 if (stripe_blockcount_get(&h->s->existing_stripe.key.v, i)) {
1588                         __set_bit(i, h->s->blocks_gotten);
1589                         __set_bit(i, h->s->blocks_allocated);
1590                 }
1591
1592                 ec_block_io(c, &h->s->existing_stripe, READ, i, &h->s->iodone);
1593         }
1594
1595         bkey_copy(&h->s->new_stripe.key.k_i, &h->s->existing_stripe.key.k_i);
1596         h->s->have_existing_stripe = true;
1597
1598         return 0;
1599 }
1600
1601 static int __bch2_ec_stripe_head_reserve(struct btree_trans *trans, struct ec_stripe_head *h)
1602 {
1603         struct bch_fs *c = trans->c;
1604         struct btree_iter iter;
1605         struct bkey_s_c k;
1606         struct bpos min_pos = POS(0, 1);
1607         struct bpos start_pos = bpos_max(min_pos, POS(0, c->ec_stripe_hint));
1608         int ret;
1609
1610         if (!h->s->res.sectors) {
1611                 ret = bch2_disk_reservation_get(c, &h->s->res,
1612                                         h->blocksize,
1613                                         h->s->nr_parity,
1614                                         BCH_DISK_RESERVATION_NOFAIL);
1615                 if (ret)
1616                         return ret;
1617         }
1618
1619         for_each_btree_key_norestart(trans, iter, BTREE_ID_stripes, start_pos,
1620                            BTREE_ITER_SLOTS|BTREE_ITER_INTENT, k, ret) {
1621                 if (bkey_gt(k.k->p, POS(0, U32_MAX))) {
1622                         if (start_pos.offset) {
1623                                 start_pos = min_pos;
1624                                 bch2_btree_iter_set_pos(&iter, start_pos);
1625                                 continue;
1626                         }
1627
1628                         ret = -BCH_ERR_ENOSPC_stripe_create;
1629                         break;
1630                 }
1631
1632                 if (bkey_deleted(k.k) &&
1633                     bch2_try_open_stripe(c, h->s, k.k->p.offset))
1634                         break;
1635         }
1636
1637         c->ec_stripe_hint = iter.pos.offset;
1638
1639         if (ret)
1640                 goto err;
1641
1642         ret = ec_stripe_mem_alloc(trans, &iter);
1643         if (ret) {
1644                 bch2_stripe_close(c, h->s);
1645                 goto err;
1646         }
1647
1648         h->s->new_stripe.key.k.p = iter.pos;
1649 out:
1650         bch2_trans_iter_exit(trans, &iter);
1651         return ret;
1652 err:
1653         bch2_disk_reservation_put(c, &h->s->res);
1654         goto out;
1655 }
1656
1657 struct ec_stripe_head *bch2_ec_stripe_head_get(struct btree_trans *trans,
1658                                                unsigned target,
1659                                                unsigned algo,
1660                                                unsigned redundancy,
1661                                                enum alloc_reserve reserve,
1662                                                struct closure *cl)
1663 {
1664         struct bch_fs *c = trans->c;
1665         struct ec_stripe_head *h;
1666         bool waiting = false;
1667         int ret;
1668
1669         h = __bch2_ec_stripe_head_get(trans, target, algo, redundancy, reserve);
1670         if (!h)
1671                 bch_err(c, "no stripe head");
1672         if (IS_ERR_OR_NULL(h))
1673                 return h;
1674
1675         if (!h->s) {
1676                 if (ec_new_stripe_alloc(c, h)) {
1677                         ret = -ENOMEM;
1678                         bch_err(c, "failed to allocate new stripe");
1679                         goto err;
1680                 }
1681         }
1682
1683         if (h->s->allocated)
1684                 goto allocated;
1685
1686         if (h->s->have_existing_stripe)
1687                 goto alloc_existing;
1688
1689         /* First, try to allocate a full stripe: */
1690         ret =   new_stripe_alloc_buckets(trans, h, RESERVE_stripe, NULL) ?:
1691                 __bch2_ec_stripe_head_reserve(trans, h);
1692         if (!ret)
1693                 goto allocate_buf;
1694         if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
1695             bch2_err_matches(ret, ENOMEM))
1696                 goto err;
1697
1698         /*
1699          * Not enough buckets available for a full stripe: we must reuse an
1700          * existing stripe:
1701          */
1702         while (1) {
1703                 ret = __bch2_ec_stripe_head_reuse(trans, h);
1704                 if (!ret)
1705                         break;
1706                 if (waiting || !cl || ret != -BCH_ERR_stripe_alloc_blocked)
1707                         goto err;
1708
1709                 /* XXX freelist_wait? */
1710                 closure_wait(&c->freelist_wait, cl);
1711                 waiting = true;
1712         }
1713
1714         if (waiting)
1715                 closure_wake_up(&c->freelist_wait);
1716 alloc_existing:
1717         /*
1718          * Retry allocating buckets, with the reserve watermark for this
1719          * particular write:
1720          */
1721         ret = new_stripe_alloc_buckets(trans, h, reserve, cl);
1722         if (ret)
1723                 goto err;
1724
1725 allocate_buf:
1726         ret = ec_stripe_buf_init(&h->s->new_stripe, 0, h->blocksize);
1727         if (ret)
1728                 goto err;
1729
1730         h->s->allocated = true;
1731 allocated:
1732         BUG_ON(!h->s->idx);
1733         BUG_ON(!h->s->new_stripe.data[0]);
1734         BUG_ON(trans->restarted);
1735         return h;
1736 err:
1737         bch2_ec_stripe_head_put(c, h);
1738         return ERR_PTR(ret);
1739 }
1740
1741 void bch2_ec_stop_dev(struct bch_fs *c, struct bch_dev *ca)
1742 {
1743         struct ec_stripe_head *h;
1744         struct open_bucket *ob;
1745         unsigned i;
1746
1747         mutex_lock(&c->ec_stripe_head_lock);
1748         list_for_each_entry(h, &c->ec_stripe_head_list, list) {
1749
1750                 mutex_lock(&h->lock);
1751                 if (!h->s)
1752                         goto unlock;
1753
1754                 for (i = 0; i < h->s->new_stripe.key.v.nr_blocks; i++) {
1755                         if (!h->s->blocks[i])
1756                                 continue;
1757
1758                         ob = c->open_buckets + h->s->blocks[i];
1759                         if (ob->dev == ca->dev_idx)
1760                                 goto found;
1761                 }
1762                 goto unlock;
1763 found:
1764                 h->s->err = -EROFS;
1765                 ec_stripe_set_pending(c, h);
1766 unlock:
1767                 mutex_unlock(&h->lock);
1768         }
1769         mutex_unlock(&c->ec_stripe_head_lock);
1770 }
1771
1772 int bch2_stripes_read(struct bch_fs *c)
1773 {
1774         struct btree_trans trans;
1775         struct btree_iter iter;
1776         struct bkey_s_c k;
1777         const struct bch_stripe *s;
1778         struct stripe *m;
1779         unsigned i;
1780         int ret;
1781
1782         bch2_trans_init(&trans, c, 0, 0);
1783
1784         for_each_btree_key(&trans, iter, BTREE_ID_stripes, POS_MIN,
1785                            BTREE_ITER_PREFETCH, k, ret) {
1786                 if (k.k->type != KEY_TYPE_stripe)
1787                         continue;
1788
1789                 ret = __ec_stripe_mem_alloc(c, k.k->p.offset, GFP_KERNEL);
1790                 if (ret)
1791                         break;
1792
1793                 s = bkey_s_c_to_stripe(k).v;
1794
1795                 m = genradix_ptr(&c->stripes, k.k->p.offset);
1796                 m->sectors      = le16_to_cpu(s->sectors);
1797                 m->algorithm    = s->algorithm;
1798                 m->nr_blocks    = s->nr_blocks;
1799                 m->nr_redundant = s->nr_redundant;
1800                 m->blocks_nonempty = 0;
1801
1802                 for (i = 0; i < s->nr_blocks; i++)
1803                         m->blocks_nonempty += !!stripe_blockcount_get(s, i);
1804
1805                 bch2_stripes_heap_insert(c, m, k.k->p.offset);
1806         }
1807         bch2_trans_iter_exit(&trans, &iter);
1808
1809         bch2_trans_exit(&trans);
1810
1811         if (ret)
1812                 bch_err(c, "error reading stripes: %i", ret);
1813
1814         return ret;
1815 }
1816
1817 void bch2_stripes_heap_to_text(struct printbuf *out, struct bch_fs *c)
1818 {
1819         ec_stripes_heap *h = &c->ec_stripes_heap;
1820         struct stripe *m;
1821         size_t i;
1822
1823         mutex_lock(&c->ec_stripes_heap_lock);
1824         for (i = 0; i < min_t(size_t, h->used, 20); i++) {
1825                 m = genradix_ptr(&c->stripes, h->data[i].idx);
1826
1827                 prt_printf(out, "%zu %u/%u+%u\n", h->data[i].idx,
1828                        h->data[i].blocks_nonempty,
1829                        m->nr_blocks - m->nr_redundant,
1830                        m->nr_redundant);
1831         }
1832         mutex_unlock(&c->ec_stripes_heap_lock);
1833 }
1834
1835 void bch2_new_stripes_to_text(struct printbuf *out, struct bch_fs *c)
1836 {
1837         struct ec_stripe_head *h;
1838         struct ec_stripe_new *s;
1839
1840         mutex_lock(&c->ec_stripe_head_lock);
1841         list_for_each_entry(h, &c->ec_stripe_head_list, list) {
1842                 prt_printf(out, "target %u algo %u redundancy %u:\n",
1843                        h->target, h->algo, h->redundancy);
1844
1845                 if (h->s)
1846                         prt_printf(out, "\tpending: idx %llu blocks %u+%u allocated %u\n",
1847                                h->s->idx, h->s->nr_data, h->s->nr_parity,
1848                                bitmap_weight(h->s->blocks_allocated,
1849                                              h->s->nr_data));
1850         }
1851         mutex_unlock(&c->ec_stripe_head_lock);
1852
1853         mutex_lock(&c->ec_stripe_new_lock);
1854         list_for_each_entry(s, &c->ec_stripe_new_list, list) {
1855                 prt_printf(out, "\tin flight: idx %llu blocks %u+%u pin %u\n",
1856                            s->idx, s->nr_data, s->nr_parity,
1857                            atomic_read(&s->pin));
1858         }
1859         mutex_unlock(&c->ec_stripe_new_lock);
1860 }
1861
1862 void bch2_fs_ec_exit(struct bch_fs *c)
1863 {
1864         struct ec_stripe_head *h;
1865         unsigned i;
1866
1867         while (1) {
1868                 mutex_lock(&c->ec_stripe_head_lock);
1869                 h = list_first_entry_or_null(&c->ec_stripe_head_list,
1870                                              struct ec_stripe_head, list);
1871                 if (h)
1872                         list_del(&h->list);
1873                 mutex_unlock(&c->ec_stripe_head_lock);
1874                 if (!h)
1875                         break;
1876
1877                 if (h->s) {
1878                         for (i = 0; i < h->s->new_stripe.key.v.nr_blocks; i++)
1879                                 BUG_ON(h->s->blocks[i]);
1880
1881                         kfree(h->s);
1882                 }
1883                 kfree(h);
1884         }
1885
1886         BUG_ON(!list_empty(&c->ec_stripe_new_list));
1887
1888         free_heap(&c->ec_stripes_heap);
1889         genradix_free(&c->stripes);
1890         bioset_exit(&c->ec_bioset);
1891 }
1892
1893 void bch2_fs_ec_init_early(struct bch_fs *c)
1894 {
1895         INIT_WORK(&c->ec_stripe_create_work, ec_stripe_create_work);
1896         INIT_WORK(&c->ec_stripe_delete_work, ec_stripe_delete_work);
1897 }
1898
1899 int bch2_fs_ec_init(struct bch_fs *c)
1900 {
1901         spin_lock_init(&c->ec_stripes_new_lock);
1902
1903         return bioset_init(&c->ec_bioset, 1, offsetof(struct ec_bio, bio),
1904                            BIOSET_NEED_BVECS);
1905 }