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 static int ec_stripe_buf_init(struct ec_stripe_buf *buf,
 217                                unsigned offset, unsigned size)
 218 {
 219         struct bch_stripe *v = &buf->key.v;
 220         unsigned csum_granularity = 1U << v->csum_granularity_bits;
 221         unsigned end = offset + size;
 222         unsigned i;
 223
 224         BUG_ON(end > le16_to_cpu(v->sectors));
 225
 226         offset  = round_down(offset, csum_granularity);
 227         end     = min_t(unsigned, le16_to_cpu(v->sectors),
 228                         round_up(end, csum_granularity));
 229
 230         buf->offset     = offset;
 231         buf->size       = end - offset;
 232
 233         memset(buf->valid, 0xFF, sizeof(buf->valid));
 234
 235         for (i = 0; i < buf->key.v.nr_blocks; i++) {
 236                 buf->data[i] = kvpmalloc(buf->size << 9, GFP_KERNEL);
 237                 if (!buf->data[i])
 238                         goto err;
 239         }
 240
 241         return 0;
 242 err:
 243         ec_stripe_buf_exit(buf);
 244         return -ENOMEM;
 245 }
 246
 247 /* Checksumming: */
 248
 249 static struct bch_csum ec_block_checksum(struct ec_stripe_buf *buf,
 250                                          unsigned block, unsigned offset)
 251 {
 252         struct bch_stripe *v = &buf->key.v;
 253         unsigned csum_granularity = 1 << v->csum_granularity_bits;
 254         unsigned end = buf->offset + buf->size;
 255         unsigned len = min(csum_granularity, end - offset);
 256
 257         BUG_ON(offset >= end);
 258         BUG_ON(offset <  buf->offset);
 259         BUG_ON(offset & (csum_granularity - 1));
 260         BUG_ON(offset + len != le16_to_cpu(v->sectors) &&
 261                (len & (csum_granularity - 1)));
 262
 263         return bch2_checksum(NULL, v->csum_type,
 264                              null_nonce(),
 265                              buf->data[block] + ((offset - buf->offset) << 9),
 266                              len << 9);
 267 }
 268
 269 static void ec_generate_checksums(struct ec_stripe_buf *buf)
 270 {
 271         struct bch_stripe *v = &buf->key.v;
 272         unsigned i, j, csums_per_device = stripe_csums_per_device(v);
 273
 274         if (!v->csum_type)
 275                 return;
 276
 277         BUG_ON(buf->offset);
 278         BUG_ON(buf->size != le16_to_cpu(v->sectors));
 279
 280         for (i = 0; i < v->nr_blocks; i++)
 281                 for (j = 0; j < csums_per_device; j++)
 282                         stripe_csum_set(v, i, j,
 283                                 ec_block_checksum(buf, i, j << v->csum_granularity_bits));
 284 }
 285
 286 static void ec_validate_checksums(struct bch_fs *c, struct ec_stripe_buf *buf)
 287 {
 288         struct bch_stripe *v = &buf->key.v;
 289         unsigned csum_granularity = 1 << v->csum_granularity_bits;
 290         unsigned i;
 291
 292         if (!v->csum_type)
 293                 return;
 294
 295         for (i = 0; i < v->nr_blocks; i++) {
 296                 unsigned offset = buf->offset;
 297                 unsigned end = buf->offset + buf->size;
 298
 299                 if (!test_bit(i, buf->valid))
 300                         continue;
 301
 302                 while (offset < end) {
 303                         unsigned j = offset >> v->csum_granularity_bits;
 304                         unsigned len = min(csum_granularity, end - offset);
 305                         struct bch_csum want = stripe_csum_get(v, i, j);
 306                         struct bch_csum got = ec_block_checksum(buf, i, offset);
 307
 308                         if (bch2_crc_cmp(want, got)) {
 309                                 struct printbuf buf2 = PRINTBUF;
 310
 311                                 bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(&buf->key.k_i));
 312
 313                                 bch_err_ratelimited(c,
 314                                         "stripe checksum error for %ps at %u:%u: csum type %u, expected %llx got %llx\n%s",
 315                                         (void *) _RET_IP_, i, j, v->csum_type,
 316                                         want.lo, got.lo, buf2.buf);
 317                                 printbuf_exit(&buf2);
 318                                 clear_bit(i, buf->valid);
 319                                 break;
 320                         }
 321
 322                         offset += len;
 323                 }
 324         }
 325 }
 326
 327 /* Erasure coding: */
 328
 329 static void ec_generate_ec(struct ec_stripe_buf *buf)
 330 {
 331         struct bch_stripe *v = &buf->key.v;
 332         unsigned nr_data = v->nr_blocks - v->nr_redundant;
 333         unsigned bytes = le16_to_cpu(v->sectors) << 9;
 334
 335         raid_gen(nr_data, v->nr_redundant, bytes, buf->data);
 336 }
 337
 338 static unsigned ec_nr_failed(struct ec_stripe_buf *buf)
 339 {
 340         return buf->key.v.nr_blocks -
 341                 bitmap_weight(buf->valid, buf->key.v.nr_blocks);
 342 }
 343
 344 static int ec_do_recov(struct bch_fs *c, struct ec_stripe_buf *buf)
 345 {
 346         struct bch_stripe *v = &buf->key.v;
 347         unsigned i, failed[BCH_BKEY_PTRS_MAX], nr_failed = 0;
 348         unsigned nr_data = v->nr_blocks - v->nr_redundant;
 349         unsigned bytes = buf->size << 9;
 350
 351         if (ec_nr_failed(buf) > v->nr_redundant) {
 352                 bch_err_ratelimited(c,
 353                         "error doing reconstruct read: unable to read enough blocks");
 354                 return -1;
 355         }
 356
 357         for (i = 0; i < nr_data; i++)
 358                 if (!test_bit(i, buf->valid))
 359                         failed[nr_failed++] = i;
 360
 361         raid_rec(nr_failed, failed, nr_data, v->nr_redundant, bytes, buf->data);
 362         return 0;
 363 }
 364
 365 /* IO: */
 366
 367 static void ec_block_endio(struct bio *bio)
 368 {
 369         struct ec_bio *ec_bio = container_of(bio, struct ec_bio, bio);
 370         struct bch_stripe *v = &ec_bio->buf->key.v;
 371         struct bch_extent_ptr *ptr = &v->ptrs[ec_bio->idx];
 372         struct bch_dev *ca = ec_bio->ca;
 373         struct closure *cl = bio->bi_private;
 374
 375         if (bch2_dev_io_err_on(bio->bi_status, ca, "erasure coding %s error: %s",
 376                                bio_data_dir(bio) ? "write" : "read",
 377                                bch2_blk_status_to_str(bio->bi_status)))
 378                 clear_bit(ec_bio->idx, ec_bio->buf->valid);
 379
 380         if (ptr_stale(ca, ptr)) {
 381                 bch_err_ratelimited(ca->fs,
 382                                     "error %s stripe: stale pointer after io",
 383                                     bio_data_dir(bio) == READ ? "reading from" : "writing to");
 384                 clear_bit(ec_bio->idx, ec_bio->buf->valid);
 385         }
 386
 387         bio_put(&ec_bio->bio);
 388         percpu_ref_put(&ca->io_ref);
 389         closure_put(cl);
 390 }
 391
 392 static void ec_block_io(struct bch_fs *c, struct ec_stripe_buf *buf,
 393                         unsigned rw, unsigned idx, struct closure *cl)
 394 {
 395         struct bch_stripe *v = &buf->key.v;
 396         unsigned offset = 0, bytes = buf->size << 9;
 397         struct bch_extent_ptr *ptr = &v->ptrs[idx];
 398         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 399         enum bch_data_type data_type = idx < buf->key.v.nr_blocks - buf->key.v.nr_redundant
 400                 ? BCH_DATA_user
 401                 : BCH_DATA_parity;
 402
 403         if (ptr_stale(ca, ptr)) {
 404                 bch_err_ratelimited(c,
 405                                     "error %s stripe: stale pointer",
 406                                     rw == READ ? "reading from" : "writing to");
 407                 clear_bit(idx, buf->valid);
 408                 return;
 409         }
 410
 411         if (!bch2_dev_get_ioref(ca, rw)) {
 412                 clear_bit(idx, buf->valid);
 413                 return;
 414         }
 415
 416         this_cpu_add(ca->io_done->sectors[rw][data_type], buf->size);
 417
 418         while (offset < bytes) {
 419                 unsigned nr_iovecs = min_t(size_t, BIO_MAX_VECS,
 420                                            DIV_ROUND_UP(bytes, PAGE_SIZE));
 421                 unsigned b = min_t(size_t, bytes - offset,
 422                                    nr_iovecs << PAGE_SHIFT);
 423                 struct ec_bio *ec_bio;
 424
 425                 ec_bio = container_of(bio_alloc_bioset(ca->disk_sb.bdev,
 426                                                        nr_iovecs,
 427                                                        rw,
 428                                                        GFP_KERNEL,
 429                                                        &c->ec_bioset),
 430                                       struct ec_bio, bio);
 431
 432                 ec_bio->ca                      = ca;
 433                 ec_bio->buf                     = buf;
 434                 ec_bio->idx                     = idx;
 435
 436                 ec_bio->bio.bi_iter.bi_sector   = ptr->offset + buf->offset + (offset >> 9);
 437                 ec_bio->bio.bi_end_io           = ec_block_endio;
 438                 ec_bio->bio.bi_private          = cl;
 439
 440                 bch2_bio_map(&ec_bio->bio, buf->data[idx] + offset, b);
 441
 442                 closure_get(cl);
 443                 percpu_ref_get(&ca->io_ref);
 444
 445                 submit_bio(&ec_bio->bio);
 446
 447                 offset += b;
 448         }
 449
 450         percpu_ref_put(&ca->io_ref);
 451 }
 452
 453 static int get_stripe_key_trans(struct btree_trans *trans, u64 idx,
 454                                 struct ec_stripe_buf *stripe)
 455 {
 456         struct btree_iter iter;
 457         struct bkey_s_c k;
 458         int ret;
 459
 460         bch2_trans_iter_init(trans, &iter, BTREE_ID_stripes,
 461                              POS(0, idx), BTREE_ITER_SLOTS);
 462         k = bch2_btree_iter_peek_slot(&iter);
 463         ret = bkey_err(k);
 464         if (ret)
 465                 goto err;
 466         if (k.k->type != KEY_TYPE_stripe) {
 467                 ret = -ENOENT;
 468                 goto err;
 469         }
 470         bkey_reassemble(&stripe->key.k_i, k);
 471 err:
 472         bch2_trans_iter_exit(trans, &iter);
 473         return ret;
 474 }
 475
 476 static int get_stripe_key(struct bch_fs *c, u64 idx, struct ec_stripe_buf *stripe)
 477 {
 478         return bch2_trans_run(c, get_stripe_key_trans(&trans, idx, stripe));
 479 }
 480
 481 /* recovery read path: */
 482 int bch2_ec_read_extent(struct bch_fs *c, struct bch_read_bio *rbio)
 483 {
 484         struct ec_stripe_buf *buf;
 485         struct closure cl;
 486         struct bch_stripe *v;
 487         unsigned i, offset;
 488         int ret = 0;
 489
 490         closure_init_stack(&cl);
 491
 492         BUG_ON(!rbio->pick.has_ec);
 493
 494         buf = kzalloc(sizeof(*buf), GFP_NOIO);
 495         if (!buf)
 496                 return -ENOMEM;
 497
 498         ret = get_stripe_key(c, rbio->pick.ec.idx, buf);
 499         if (ret) {
 500                 bch_err_ratelimited(c,
 501                         "error doing reconstruct read: error %i looking up stripe", ret);
 502                 kfree(buf);
 503                 return -EIO;
 504         }
 505
 506         v = &buf->key.v;
 507
 508         if (!bch2_ptr_matches_stripe(v, rbio->pick)) {
 509                 bch_err_ratelimited(c,
 510                         "error doing reconstruct read: pointer doesn't match stripe");
 511                 ret = -EIO;
 512                 goto err;
 513         }
 514
 515         offset = rbio->bio.bi_iter.bi_sector - v->ptrs[rbio->pick.ec.block].offset;
 516         if (offset + bio_sectors(&rbio->bio) > le16_to_cpu(v->sectors)) {
 517                 bch_err_ratelimited(c,
 518                         "error doing reconstruct read: read is bigger than stripe");
 519                 ret = -EIO;
 520                 goto err;
 521         }
 522
 523         ret = ec_stripe_buf_init(buf, offset, bio_sectors(&rbio->bio));
 524         if (ret)
 525                 goto err;
 526
 527         for (i = 0; i < v->nr_blocks; i++)
 528                 ec_block_io(c, buf, REQ_OP_READ, i, &cl);
 529
 530         closure_sync(&cl);
 531
 532         if (ec_nr_failed(buf) > v->nr_redundant) {
 533                 bch_err_ratelimited(c,
 534                         "error doing reconstruct read: unable to read enough blocks");
 535                 ret = -EIO;
 536                 goto err;
 537         }
 538
 539         ec_validate_checksums(c, buf);
 540
 541         ret = ec_do_recov(c, buf);
 542         if (ret)
 543                 goto err;
 544
 545         memcpy_to_bio(&rbio->bio, rbio->bio.bi_iter,
 546                       buf->data[rbio->pick.ec.block] + ((offset - buf->offset) << 9));
 547 err:
 548         ec_stripe_buf_exit(buf);
 549         kfree(buf);
 550         return ret;
 551 }
 552
 553 /* stripe bucket accounting: */
 554
 555 static int __ec_stripe_mem_alloc(struct bch_fs *c, size_t idx, gfp_t gfp)
 556 {
 557         ec_stripes_heap n, *h = &c->ec_stripes_heap;
 558
 559         if (idx >= h->size) {
 560                 if (!init_heap(&n, max(1024UL, roundup_pow_of_two(idx + 1)), gfp))
 561                         return -ENOMEM;
 562
 563                 mutex_lock(&c->ec_stripes_heap_lock);
 564                 if (n.size > h->size) {
 565                         memcpy(n.data, h->data, h->used * sizeof(h->data[0]));
 566                         n.used = h->used;
 567                         swap(*h, n);
 568                 }
 569                 mutex_unlock(&c->ec_stripes_heap_lock);
 570
 571                 free_heap(&n);
 572         }
 573
 574         if (!genradix_ptr_alloc(&c->stripes, idx, gfp))
 575                 return -ENOMEM;
 576
 577         if (c->gc_pos.phase != GC_PHASE_NOT_RUNNING &&
 578             !genradix_ptr_alloc(&c->gc_stripes, idx, gfp))
 579                 return -ENOMEM;
 580
 581         return 0;
 582 }
 583
 584 static int ec_stripe_mem_alloc(struct btree_trans *trans,
 585                                struct btree_iter *iter)
 586 {
 587         size_t idx = iter->pos.offset;
 588
 589         if (!__ec_stripe_mem_alloc(trans->c, idx, GFP_NOWAIT|__GFP_NOWARN))
 590                 return 0;
 591
 592         bch2_trans_unlock(trans);
 593
 594         return   __ec_stripe_mem_alloc(trans->c, idx, GFP_KERNEL) ?:
 595                 bch2_trans_relock(trans);
 596 }
 597
 598 /*
 599  * Hash table of open stripes:
 600  * Stripes that are being created or modified are kept in a hash table, so that
 601  * stripe deletion can skip them.
 602  */
 603
 604 static bool __bch2_stripe_is_open(struct bch_fs *c, u64 idx)
 605 {
 606         unsigned hash = hash_64(idx, ilog2(ARRAY_SIZE(c->ec_stripes_new)));
 607         struct ec_stripe_new *s;
 608
 609         hlist_for_each_entry(s, &c->ec_stripes_new[hash], hash)
 610                 if (s->idx == idx)
 611                         return true;
 612         return false;
 613 }
 614
 615 static bool bch2_stripe_is_open(struct bch_fs *c, u64 idx)
 616 {
 617         bool ret = false;
 618
 619         spin_lock(&c->ec_stripes_new_lock);
 620         ret = __bch2_stripe_is_open(c, idx);
 621         spin_unlock(&c->ec_stripes_new_lock);
 622
 623         return ret;
 624 }
 625
 626 static bool bch2_try_open_stripe(struct bch_fs *c,
 627                                  struct ec_stripe_new *s,
 628                                  u64 idx)
 629 {
 630         bool ret;
 631
 632         spin_lock(&c->ec_stripes_new_lock);
 633         ret = !__bch2_stripe_is_open(c, idx);
 634         if (ret) {
 635                 unsigned hash = hash_64(idx, ilog2(ARRAY_SIZE(c->ec_stripes_new)));
 636
 637                 s->idx = idx;
 638                 hlist_add_head(&s->hash, &c->ec_stripes_new[hash]);
 639         }
 640         spin_unlock(&c->ec_stripes_new_lock);
 641
 642         return ret;
 643 }
 644
 645 static void bch2_stripe_close(struct bch_fs *c, struct ec_stripe_new *s)
 646 {
 647         BUG_ON(!s->idx);
 648
 649         spin_lock(&c->ec_stripes_new_lock);
 650         hlist_del(&s->hash);
 651         spin_unlock(&c->ec_stripes_new_lock);
 652
 653         s->idx = 0;
 654 }
 655
 656 /* Heap of all existing stripes, ordered by blocks_nonempty */
 657
 658 static u64 stripe_idx_to_delete(struct bch_fs *c)
 659 {
 660         ec_stripes_heap *h = &c->ec_stripes_heap;
 661         size_t heap_idx;
 662
 663         lockdep_assert_held(&c->ec_stripes_heap_lock);
 664
 665         for (heap_idx = 0; heap_idx < h->used; heap_idx++)
 666                 if (h->data[heap_idx].blocks_nonempty == 0 &&
 667                     !bch2_stripe_is_open(c, h->data[heap_idx].idx))
 668                         return h->data[heap_idx].idx;
 669
 670         return 0;
 671 }
 672
 673 static inline int ec_stripes_heap_cmp(ec_stripes_heap *h,
 674                                       struct ec_stripe_heap_entry l,
 675                                       struct ec_stripe_heap_entry r)
 676 {
 677         return ((l.blocks_nonempty > r.blocks_nonempty) -
 678                 (l.blocks_nonempty < r.blocks_nonempty));
 679 }
 680
 681 static inline void ec_stripes_heap_set_backpointer(ec_stripes_heap *h,
 682                                                    size_t i)
 683 {
 684         struct bch_fs *c = container_of(h, struct bch_fs, ec_stripes_heap);
 685
 686         genradix_ptr(&c->stripes, h->data[i].idx)->heap_idx = i;
 687 }
 688
 689 static void heap_verify_backpointer(struct bch_fs *c, size_t idx)
 690 {
 691         ec_stripes_heap *h = &c->ec_stripes_heap;
 692         struct stripe *m = genradix_ptr(&c->stripes, idx);
 693
 694         BUG_ON(m->heap_idx >= h->used);
 695         BUG_ON(h->data[m->heap_idx].idx != idx);
 696 }
 697
 698 void bch2_stripes_heap_del(struct bch_fs *c,
 699                            struct stripe *m, size_t idx)
 700 {
 701         mutex_lock(&c->ec_stripes_heap_lock);
 702         heap_verify_backpointer(c, idx);
 703
 704         heap_del(&c->ec_stripes_heap, m->heap_idx,
 705                  ec_stripes_heap_cmp,
 706                  ec_stripes_heap_set_backpointer);
 707         mutex_unlock(&c->ec_stripes_heap_lock);
 708 }
 709
 710 void bch2_stripes_heap_insert(struct bch_fs *c,
 711                               struct stripe *m, size_t idx)
 712 {
 713         mutex_lock(&c->ec_stripes_heap_lock);
 714         BUG_ON(heap_full(&c->ec_stripes_heap));
 715
 716         heap_add(&c->ec_stripes_heap, ((struct ec_stripe_heap_entry) {
 717                         .idx = idx,
 718                         .blocks_nonempty = m->blocks_nonempty,
 719                 }),
 720                  ec_stripes_heap_cmp,
 721                  ec_stripes_heap_set_backpointer);
 722
 723         heap_verify_backpointer(c, idx);
 724         mutex_unlock(&c->ec_stripes_heap_lock);
 725 }
 726
 727 void bch2_stripes_heap_update(struct bch_fs *c,
 728                               struct stripe *m, size_t idx)
 729 {
 730         ec_stripes_heap *h = &c->ec_stripes_heap;
 731         bool do_deletes;
 732         size_t i;
 733
 734         mutex_lock(&c->ec_stripes_heap_lock);
 735         heap_verify_backpointer(c, idx);
 736
 737         h->data[m->heap_idx].blocks_nonempty = m->blocks_nonempty;
 738
 739         i = m->heap_idx;
 740         heap_sift_up(h,   i, ec_stripes_heap_cmp,
 741                      ec_stripes_heap_set_backpointer);
 742         heap_sift_down(h, i, ec_stripes_heap_cmp,
 743                        ec_stripes_heap_set_backpointer);
 744
 745         heap_verify_backpointer(c, idx);
 746
 747         do_deletes = stripe_idx_to_delete(c) != 0;
 748         mutex_unlock(&c->ec_stripes_heap_lock);
 749
 750         if (do_deletes)
 751                 bch2_do_stripe_deletes(c);
 752 }
 753
 754 /* stripe deletion */
 755
 756 static int ec_stripe_delete(struct btree_trans *trans, u64 idx)
 757 {
 758         struct bch_fs *c = trans->c;
 759         struct btree_iter iter;
 760         struct bkey_s_c k;
 761         struct bkey_s_c_stripe s;
 762         int ret;
 763
 764         bch2_trans_iter_init(trans, &iter, BTREE_ID_stripes, POS(0, idx),
 765                              BTREE_ITER_INTENT);
 766         k = bch2_btree_iter_peek_slot(&iter);
 767         ret = bkey_err(k);
 768         if (ret)
 769                 goto err;
 770
 771         if (k.k->type != KEY_TYPE_stripe) {
 772                 bch2_fs_inconsistent(c, "attempting to delete nonexistent stripe %llu", idx);
 773                 ret = -EINVAL;
 774                 goto err;
 775         }
 776
 777         s = bkey_s_c_to_stripe(k);
 778         for (unsigned i = 0; i < s.v->nr_blocks; i++)
 779                 if (stripe_blockcount_get(s.v, i)) {
 780                         struct printbuf buf = PRINTBUF;
 781
 782                         bch2_bkey_val_to_text(&buf, c, k);
 783                         bch2_fs_inconsistent(c, "attempting to delete nonempty stripe %s", buf.buf);
 784                         printbuf_exit(&buf);
 785                         ret = -EINVAL;
 786                         goto err;
 787                 }
 788
 789         ret = bch2_btree_delete_at(trans, &iter, 0);
 790 err:
 791         bch2_trans_iter_exit(trans, &iter);
 792         return ret;
 793 }
 794
 795 static void ec_stripe_delete_work(struct work_struct *work)
 796 {
 797         struct bch_fs *c =
 798                 container_of(work, struct bch_fs, ec_stripe_delete_work);
 799         struct btree_trans trans;
 800         int ret;
 801         u64 idx;
 802
 803         bch2_trans_init(&trans, c, 0, 0);
 804
 805         while (1) {
 806                 mutex_lock(&c->ec_stripes_heap_lock);
 807                 idx = stripe_idx_to_delete(c);
 808                 mutex_unlock(&c->ec_stripes_heap_lock);
 809
 810                 if (!idx)
 811                         break;
 812
 813                 ret = commit_do(&trans, NULL, NULL, BTREE_INSERT_NOFAIL,
 814                                 ec_stripe_delete(&trans, idx));
 815                 if (ret) {
 816                         bch_err(c, "%s: err %s", __func__, bch2_err_str(ret));
 817                         break;
 818                 }
 819         }
 820
 821         bch2_trans_exit(&trans);
 822
 823         bch2_write_ref_put(c, BCH_WRITE_REF_stripe_delete);
 824 }
 825
 826 void bch2_do_stripe_deletes(struct bch_fs *c)
 827 {
 828         if (bch2_write_ref_tryget(c, BCH_WRITE_REF_stripe_delete) &&
 829             !schedule_work(&c->ec_stripe_delete_work))
 830                 bch2_write_ref_put(c, BCH_WRITE_REF_stripe_delete);
 831 }
 832
 833 /* stripe creation: */
 834
 835 static int ec_stripe_key_update(struct btree_trans *trans,
 836                                 struct bkey_i_stripe *new,
 837                                 bool create)
 838 {
 839         struct bch_fs *c = trans->c;
 840         struct btree_iter iter;
 841         struct bkey_s_c k;
 842         int ret;
 843
 844         bch2_trans_iter_init(trans, &iter, BTREE_ID_stripes,
 845                              new->k.p, BTREE_ITER_INTENT);
 846         k = bch2_btree_iter_peek_slot(&iter);
 847         ret = bkey_err(k);
 848         if (ret)
 849                 goto err;
 850
 851         if (k.k->type != (create ? KEY_TYPE_deleted : KEY_TYPE_stripe)) {
 852                 bch2_fs_inconsistent(c, "error %s stripe: got existing key type %s",
 853                                      create ? "creating" : "updating",
 854                                      bch2_bkey_types[k.k->type]);
 855                 ret = -EINVAL;
 856                 goto err;
 857         }
 858
 859         if (k.k->type == KEY_TYPE_stripe) {
 860                 const struct bch_stripe *old = bkey_s_c_to_stripe(k).v;
 861                 unsigned i;
 862
 863                 if (old->nr_blocks != new->v.nr_blocks) {
 864                         bch_err(c, "error updating stripe: nr_blocks does not match");
 865                         ret = -EINVAL;
 866                         goto err;
 867                 }
 868
 869                 for (i = 0; i < new->v.nr_blocks; i++)
 870                         stripe_blockcount_set(&new->v, i, stripe_blockcount_get(old, i));
 871         }
 872
 873         ret = bch2_trans_update(trans, &iter, &new->k_i, 0);
 874 err:
 875         bch2_trans_iter_exit(trans, &iter);
 876         return ret;
 877 }
 878
 879 static int ec_stripe_update_extent(struct btree_trans *trans,
 880                                    struct bpos bucket, u8 gen,
 881                                    struct ec_stripe_buf *s,
 882                                    u64 *bp_offset)
 883 {
 884         struct bch_fs *c = trans->c;
 885         struct bch_backpointer bp;
 886         struct btree_iter iter;
 887         struct bkey_s_c k;
 888         const struct bch_extent_ptr *ptr_c;
 889         struct bch_extent_ptr *ptr, *ec_ptr = NULL;
 890         struct bch_extent_stripe_ptr stripe_ptr;
 891         struct bkey_i *n;
 892         int ret, dev, block;
 893
 894         ret = bch2_get_next_backpointer(trans, bucket, gen,
 895                                 bp_offset, &bp, BTREE_ITER_CACHED);
 896         if (ret)
 897                 return ret;
 898         if (*bp_offset == U64_MAX)
 899                 return 0;
 900
 901         if (bch2_fs_inconsistent_on(bp.level, c, "found btree node in erasure coded bucket!?"))
 902                 return -EIO;
 903
 904         k = bch2_backpointer_get_key(trans, &iter, bucket, *bp_offset, bp);
 905         ret = bkey_err(k);
 906         if (ret)
 907                 return ret;
 908         if (!k.k) {
 909                 /*
 910                  * extent no longer exists - we could flush the btree
 911                  * write buffer and retry to verify, but no need:
 912                  */
 913                 return 0;
 914         }
 915
 916         if (extent_has_stripe_ptr(k, s->key.k.p.offset))
 917                 goto out;
 918
 919         ptr_c = bkey_matches_stripe(&s->key.v, k, &block);
 920         /*
 921          * It doesn't generally make sense to erasure code cached ptrs:
 922          * XXX: should we be incrementing a counter?
 923          */
 924         if (!ptr_c || ptr_c->cached)
 925                 goto out;
 926
 927         dev = s->key.v.ptrs[block].dev;
 928
 929         n = bch2_trans_kmalloc(trans, bkey_bytes(k.k) + sizeof(stripe_ptr));
 930         ret = PTR_ERR_OR_ZERO(n);
 931         if (ret)
 932                 goto out;
 933
 934         bkey_reassemble(n, k);
 935
 936         bch2_bkey_drop_ptrs(bkey_i_to_s(n), ptr, ptr->dev != dev);
 937         ec_ptr = (void *) bch2_bkey_has_device(bkey_i_to_s_c(n), dev);
 938         BUG_ON(!ec_ptr);
 939
 940         stripe_ptr = (struct bch_extent_stripe_ptr) {
 941                 .type = 1 << BCH_EXTENT_ENTRY_stripe_ptr,
 942                 .block          = block,
 943                 .redundancy     = s->key.v.nr_redundant,
 944                 .idx            = s->key.k.p.offset,
 945         };
 946
 947         __extent_entry_insert(n,
 948                         (union bch_extent_entry *) ec_ptr,
 949                         (union bch_extent_entry *) &stripe_ptr);
 950
 951         ret = bch2_trans_update(trans, &iter, n, 0);
 952 out:
 953         bch2_trans_iter_exit(trans, &iter);
 954         return ret;
 955 }
 956
 957 static int ec_stripe_update_bucket(struct btree_trans *trans, struct ec_stripe_buf *s,
 958                                    unsigned block)
 959 {
 960         struct bch_fs *c = trans->c;
 961         struct bch_extent_ptr bucket = s->key.v.ptrs[block];
 962         struct bpos bucket_pos = PTR_BUCKET_POS(c, &bucket);
 963         u64 bp_offset = 0;
 964         int ret = 0;
 965
 966         while (1) {
 967                 ret = commit_do(trans, NULL, NULL,
 968                                 BTREE_INSERT_NOFAIL,
 969                         ec_stripe_update_extent(trans, bucket_pos, bucket.gen,
 970                                                 s, &bp_offset));
 971                 if (ret)
 972                         break;
 973                 if (bp_offset == U64_MAX)
 974                         break;
 975
 976                 bp_offset++;
 977         }
 978
 979         return ret;
 980 }
 981
 982 static int ec_stripe_update_extents(struct bch_fs *c, struct ec_stripe_buf *s)
 983 {
 984         struct btree_trans trans;
 985         struct bch_stripe *v = &s->key.v;
 986         unsigned i, nr_data = v->nr_blocks - v->nr_redundant;
 987         int ret = 0;
 988
 989         bch2_trans_init(&trans, c, 0, 0);
 990
 991         ret = bch2_btree_write_buffer_flush(&trans);
 992         if (ret)
 993                 goto err;
 994
 995         for (i = 0; i < nr_data; i++) {
 996                 ret = ec_stripe_update_bucket(&trans, s, i);
 997                 if (ret)
 998                         break;
 999         }
1000 err:
1001         bch2_trans_exit(&trans);
1002
1003         return ret;
1004 }
1005
1006 static void zero_out_rest_of_ec_bucket(struct bch_fs *c,
1007                                        struct ec_stripe_new *s,
1008                                        unsigned block,
1009                                        struct open_bucket *ob)
1010 {
1011         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1012         unsigned offset = ca->mi.bucket_size - ob->sectors_free;
1013         int ret;
1014
1015         if (!bch2_dev_get_ioref(ca, WRITE)) {
1016                 s->err = -EROFS;
1017                 return;
1018         }
1019
1020         memset(s->new_stripe.data[block] + (offset << 9),
1021                0,
1022                ob->sectors_free << 9);
1023
1024         ret = blkdev_issue_zeroout(ca->disk_sb.bdev,
1025                         ob->bucket * ca->mi.bucket_size + offset,
1026                         ob->sectors_free,
1027                         GFP_KERNEL, 0);
1028
1029         percpu_ref_put(&ca->io_ref);
1030
1031         if (ret)
1032                 s->err = ret;
1033 }
1034
1035 /*
1036  * data buckets of new stripe all written: create the stripe
1037  */
1038 static void ec_stripe_create(struct ec_stripe_new *s)
1039 {
1040         struct bch_fs *c = s->c;
1041         struct open_bucket *ob;
1042         struct bch_stripe *v = &s->new_stripe.key.v;
1043         unsigned i, nr_data = v->nr_blocks - v->nr_redundant;
1044         int ret;
1045
1046         BUG_ON(s->h->s == s);
1047
1048         closure_sync(&s->iodone);
1049
1050         for (i = 0; i < nr_data; i++)
1051                 if (s->blocks[i]) {
1052                         ob = c->open_buckets + s->blocks[i];
1053
1054                         if (ob->sectors_free)
1055                                 zero_out_rest_of_ec_bucket(c, s, i, ob);
1056                 }
1057
1058         if (s->err) {
1059                 if (!bch2_err_matches(s->err, EROFS))
1060                         bch_err(c, "error creating stripe: error writing data buckets");
1061                 goto err;
1062         }
1063
1064         if (s->have_existing_stripe) {
1065                 ec_validate_checksums(c, &s->existing_stripe);
1066
1067                 if (ec_do_recov(c, &s->existing_stripe)) {
1068                         bch_err(c, "error creating stripe: error reading existing stripe");
1069                         goto err;
1070                 }
1071
1072                 for (i = 0; i < nr_data; i++)
1073                         if (stripe_blockcount_get(&s->existing_stripe.key.v, i))
1074                                 swap(s->new_stripe.data[i],
1075                                      s->existing_stripe.data[i]);
1076
1077                 ec_stripe_buf_exit(&s->existing_stripe);
1078         }
1079
1080         BUG_ON(!s->allocated);
1081
1082         ec_generate_ec(&s->new_stripe);
1083
1084         ec_generate_checksums(&s->new_stripe);
1085
1086         /* write p/q: */
1087         for (i = nr_data; i < v->nr_blocks; i++)
1088                 ec_block_io(c, &s->new_stripe, REQ_OP_WRITE, i, &s->iodone);
1089         closure_sync(&s->iodone);
1090
1091         if (ec_nr_failed(&s->new_stripe)) {
1092                 bch_err(c, "error creating stripe: error writing redundancy buckets");
1093                 goto err;
1094         }
1095
1096         ret = bch2_trans_do(c, &s->res, NULL, BTREE_INSERT_NOFAIL,
1097                             ec_stripe_key_update(&trans, &s->new_stripe.key,
1098                                                  !s->have_existing_stripe));
1099         if (ret) {
1100                 bch_err(c, "error creating stripe: error creating stripe key");
1101                 goto err;
1102         }
1103
1104         ret = ec_stripe_update_extents(c, &s->new_stripe);
1105         if (ret) {
1106                 bch_err(c, "error creating stripe: error updating pointers: %s",
1107                         bch2_err_str(ret));
1108                 goto err;
1109         }
1110
1111         bch2_stripe_close(c, s);
1112 err:
1113         bch2_disk_reservation_put(c, &s->res);
1114
1115         for (i = 0; i < v->nr_blocks; i++)
1116                 if (s->blocks[i]) {
1117                         ob = c->open_buckets + s->blocks[i];
1118
1119                         if (i < nr_data) {
1120                                 ob->ec = NULL;
1121                                 __bch2_open_bucket_put(c, ob);
1122                         } else {
1123                                 bch2_open_bucket_put(c, ob);
1124                         }
1125                 }
1126
1127         ec_stripe_buf_exit(&s->existing_stripe);
1128         ec_stripe_buf_exit(&s->new_stripe);
1129         closure_debug_destroy(&s->iodone);
1130         kfree(s);
1131 }
1132
1133 static struct ec_stripe_new *get_pending_stripe(struct bch_fs *c)
1134 {
1135         struct ec_stripe_new *s;
1136
1137         mutex_lock(&c->ec_stripe_new_lock);
1138         list_for_each_entry(s, &c->ec_stripe_new_list, list)
1139                 if (!atomic_read(&s->pin)) {
1140                         list_del(&s->list);
1141                         goto out;
1142                 }
1143         s = NULL;
1144 out:
1145         mutex_unlock(&c->ec_stripe_new_lock);
1146
1147         return s;
1148 }
1149
1150 static void ec_stripe_create_work(struct work_struct *work)
1151 {
1152         struct bch_fs *c = container_of(work,
1153                 struct bch_fs, ec_stripe_create_work);
1154         struct ec_stripe_new *s;
1155
1156         while ((s = get_pending_stripe(c)))
1157                 ec_stripe_create(s);
1158
1159         bch2_write_ref_put(c, BCH_WRITE_REF_stripe_create);
1160 }
1161
1162 void bch2_ec_do_stripe_creates(struct bch_fs *c)
1163 {
1164         bch2_write_ref_get(c, BCH_WRITE_REF_stripe_create);
1165
1166         if (!queue_work(system_long_wq, &c->ec_stripe_create_work))
1167                 bch2_write_ref_put(c, BCH_WRITE_REF_stripe_create);
1168 }
1169
1170 static void ec_stripe_new_put(struct bch_fs *c, struct ec_stripe_new *s)
1171 {
1172         BUG_ON(atomic_read(&s->pin) <= 0);
1173
1174         if (atomic_dec_and_test(&s->pin))
1175                 bch2_ec_do_stripe_creates(c);
1176 }
1177
1178 static void ec_stripe_set_pending(struct bch_fs *c, struct ec_stripe_head *h)
1179 {
1180         struct ec_stripe_new *s = h->s;
1181
1182         BUG_ON(!s->allocated && !s->err);
1183
1184         h->s            = NULL;
1185         s->pending      = true;
1186
1187         mutex_lock(&c->ec_stripe_new_lock);
1188         list_add(&s->list, &c->ec_stripe_new_list);
1189         mutex_unlock(&c->ec_stripe_new_lock);
1190
1191         ec_stripe_new_put(c, s);
1192 }
1193
1194 /* have a full bucket - hand it off to be erasure coded: */
1195 void bch2_ec_bucket_written(struct bch_fs *c, struct open_bucket *ob)
1196 {
1197         struct ec_stripe_new *s = ob->ec;
1198
1199         ec_stripe_new_put(c, s);
1200 }
1201
1202 void bch2_ec_bucket_cancel(struct bch_fs *c, struct open_bucket *ob)
1203 {
1204         struct ec_stripe_new *s = ob->ec;
1205
1206         s->err = -EIO;
1207 }
1208
1209 void *bch2_writepoint_ec_buf(struct bch_fs *c, struct write_point *wp)
1210 {
1211         struct open_bucket *ob = ec_open_bucket(c, &wp->ptrs);
1212         struct bch_dev *ca;
1213         unsigned offset;
1214
1215         if (!ob)
1216                 return NULL;
1217
1218         ca      = bch_dev_bkey_exists(c, ob->dev);
1219         offset  = ca->mi.bucket_size - ob->sectors_free;
1220
1221         return ob->ec->new_stripe.data[ob->ec_idx] + (offset << 9);
1222 }
1223
1224 static int unsigned_cmp(const void *_l, const void *_r)
1225 {
1226         unsigned l = *((const unsigned *) _l);
1227         unsigned r = *((const unsigned *) _r);
1228
1229         return cmp_int(l, r);
1230 }
1231
1232 /* pick most common bucket size: */
1233 static unsigned pick_blocksize(struct bch_fs *c,
1234                                struct bch_devs_mask *devs)
1235 {
1236         struct bch_dev *ca;
1237         unsigned i, nr = 0, sizes[BCH_SB_MEMBERS_MAX];
1238         struct {
1239                 unsigned nr, size;
1240         } cur = { 0, 0 }, best = { 0, 0 };
1241
1242         for_each_member_device_rcu(ca, c, i, devs)
1243                 sizes[nr++] = ca->mi.bucket_size;
1244
1245         sort(sizes, nr, sizeof(unsigned), unsigned_cmp, NULL);
1246
1247         for (i = 0; i < nr; i++) {
1248                 if (sizes[i] != cur.size) {
1249                         if (cur.nr > best.nr)
1250                                 best = cur;
1251
1252                         cur.nr = 0;
1253                         cur.size = sizes[i];
1254                 }
1255
1256                 cur.nr++;
1257         }
1258
1259         if (cur.nr > best.nr)
1260                 best = cur;
1261
1262         return best.size;
1263 }
1264
1265 static bool may_create_new_stripe(struct bch_fs *c)
1266 {
1267         return false;
1268 }
1269
1270 static void ec_stripe_key_init(struct bch_fs *c,
1271                                struct bkey_i_stripe *s,
1272                                unsigned nr_data,
1273                                unsigned nr_parity,
1274                                unsigned stripe_size)
1275 {
1276         unsigned u64s;
1277
1278         bkey_stripe_init(&s->k_i);
1279         s->v.sectors                    = cpu_to_le16(stripe_size);
1280         s->v.algorithm                  = 0;
1281         s->v.nr_blocks                  = nr_data + nr_parity;
1282         s->v.nr_redundant               = nr_parity;
1283         s->v.csum_granularity_bits      = ilog2(c->opts.encoded_extent_max >> 9);
1284         s->v.csum_type                  = BCH_CSUM_crc32c;
1285         s->v.pad                        = 0;
1286
1287         while ((u64s = stripe_val_u64s(&s->v)) > BKEY_VAL_U64s_MAX) {
1288                 BUG_ON(1 << s->v.csum_granularity_bits >=
1289                        le16_to_cpu(s->v.sectors) ||
1290                        s->v.csum_granularity_bits == U8_MAX);
1291                 s->v.csum_granularity_bits++;
1292         }
1293
1294         set_bkey_val_u64s(&s->k, u64s);
1295 }
1296
1297 static int ec_new_stripe_alloc(struct bch_fs *c, struct ec_stripe_head *h)
1298 {
1299         struct ec_stripe_new *s;
1300
1301         lockdep_assert_held(&h->lock);
1302
1303         s = kzalloc(sizeof(*s), GFP_KERNEL);
1304         if (!s)
1305                 return -ENOMEM;
1306
1307         mutex_init(&s->lock);
1308         closure_init(&s->iodone, NULL);
1309         atomic_set(&s->pin, 1);
1310         s->c            = c;
1311         s->h            = h;
1312         s->nr_data      = min_t(unsigned, h->nr_active_devs,
1313                                 BCH_BKEY_PTRS_MAX) - h->redundancy;
1314         s->nr_parity    = h->redundancy;
1315
1316         ec_stripe_key_init(c, &s->new_stripe.key, s->nr_data,
1317                            s->nr_parity, h->blocksize);
1318
1319         h->s = s;
1320         return 0;
1321 }
1322
1323 static struct ec_stripe_head *
1324 ec_new_stripe_head_alloc(struct bch_fs *c, unsigned target,
1325                          unsigned algo, unsigned redundancy,
1326                          bool copygc)
1327 {
1328         struct ec_stripe_head *h;
1329         struct bch_dev *ca;
1330         unsigned i;
1331
1332         h = kzalloc(sizeof(*h), GFP_KERNEL);
1333         if (!h)
1334                 return NULL;
1335
1336         mutex_init(&h->lock);
1337         BUG_ON(!mutex_trylock(&h->lock));
1338
1339         h->target       = target;
1340         h->algo         = algo;
1341         h->redundancy   = redundancy;
1342         h->copygc       = copygc;
1343
1344         rcu_read_lock();
1345         h->devs = target_rw_devs(c, BCH_DATA_user, target);
1346
1347         for_each_member_device_rcu(ca, c, i, &h->devs)
1348                 if (!ca->mi.durability)
1349                         __clear_bit(i, h->devs.d);
1350
1351         h->blocksize = pick_blocksize(c, &h->devs);
1352
1353         for_each_member_device_rcu(ca, c, i, &h->devs)
1354                 if (ca->mi.bucket_size == h->blocksize)
1355                         h->nr_active_devs++;
1356
1357         rcu_read_unlock();
1358         list_add(&h->list, &c->ec_stripe_head_list);
1359         return h;
1360 }
1361
1362 void bch2_ec_stripe_head_put(struct bch_fs *c, struct ec_stripe_head *h)
1363 {
1364         if (h->s &&
1365             h->s->allocated &&
1366             bitmap_weight(h->s->blocks_allocated,
1367                           h->s->nr_data) == h->s->nr_data)
1368                 ec_stripe_set_pending(c, h);
1369
1370         mutex_unlock(&h->lock);
1371 }
1372
1373 struct ec_stripe_head *__bch2_ec_stripe_head_get(struct btree_trans *trans,
1374                                                  unsigned target,
1375                                                  unsigned algo,
1376                                                  unsigned redundancy,
1377                                                  bool copygc)
1378 {
1379         struct bch_fs *c = trans->c;
1380         struct ec_stripe_head *h;
1381         int ret;
1382
1383         if (!redundancy)
1384                 return NULL;
1385
1386         ret = bch2_trans_mutex_lock(trans, &c->ec_stripe_head_lock);
1387         if (ret)
1388                 return ERR_PTR(ret);
1389
1390         list_for_each_entry(h, &c->ec_stripe_head_list, list)
1391                 if (h->target           == target &&
1392                     h->algo             == algo &&
1393                     h->redundancy       == redundancy &&
1394                     h->copygc           == copygc) {
1395                         ret = bch2_trans_mutex_lock(trans, &h->lock);
1396                         if (ret)
1397                                 h = ERR_PTR(ret);
1398                         goto found;
1399                 }
1400
1401         h = ec_new_stripe_head_alloc(c, target, algo, redundancy, copygc);
1402 found:
1403         mutex_unlock(&c->ec_stripe_head_lock);
1404         return h;
1405 }
1406
1407 static int new_stripe_alloc_buckets(struct btree_trans *trans, struct ec_stripe_head *h,
1408                                     struct closure *cl)
1409 {
1410         struct bch_fs *c = trans->c;
1411         struct bch_devs_mask devs = h->devs;
1412         struct open_bucket *ob;
1413         struct open_buckets buckets;
1414         unsigned i, j, nr_have_parity = 0, nr_have_data = 0;
1415         bool have_cache = true;
1416         int ret = 0;
1417
1418         for (i = 0; i < h->s->new_stripe.key.v.nr_blocks; i++) {
1419                 if (test_bit(i, h->s->blocks_gotten)) {
1420                         __clear_bit(h->s->new_stripe.key.v.ptrs[i].dev, devs.d);
1421                         if (i < h->s->nr_data)
1422                                 nr_have_data++;
1423                         else
1424                                 nr_have_parity++;
1425                 }
1426         }
1427
1428         BUG_ON(nr_have_data     > h->s->nr_data);
1429         BUG_ON(nr_have_parity   > h->s->nr_parity);
1430
1431         buckets.nr = 0;
1432         if (nr_have_parity < h->s->nr_parity) {
1433                 ret = bch2_bucket_alloc_set_trans(trans, &buckets,
1434                                             &h->parity_stripe,
1435                                             &devs,
1436                                             h->s->nr_parity,
1437                                             &nr_have_parity,
1438                                             &have_cache,
1439                                             BCH_DATA_parity,
1440                                             h->copygc
1441                                             ? RESERVE_movinggc
1442                                             : RESERVE_none,
1443                                             cl);
1444
1445                 open_bucket_for_each(c, &buckets, ob, i) {
1446                         j = find_next_zero_bit(h->s->blocks_gotten,
1447                                                h->s->nr_data + h->s->nr_parity,
1448                                                h->s->nr_data);
1449                         BUG_ON(j >= h->s->nr_data + h->s->nr_parity);
1450
1451                         h->s->blocks[j] = buckets.v[i];
1452                         h->s->new_stripe.key.v.ptrs[j] = bch2_ob_ptr(c, ob);
1453                         __set_bit(j, h->s->blocks_gotten);
1454                 }
1455
1456                 if (ret)
1457                         return ret;
1458         }
1459
1460         buckets.nr = 0;
1461         if (nr_have_data < h->s->nr_data) {
1462                 ret = bch2_bucket_alloc_set_trans(trans, &buckets,
1463                                             &h->block_stripe,
1464                                             &devs,
1465                                             h->s->nr_data,
1466                                             &nr_have_data,
1467                                             &have_cache,
1468                                             BCH_DATA_user,
1469                                             h->copygc
1470                                             ? RESERVE_movinggc
1471                                             : RESERVE_none,
1472                                             cl);
1473
1474                 open_bucket_for_each(c, &buckets, ob, i) {
1475                         j = find_next_zero_bit(h->s->blocks_gotten,
1476                                                h->s->nr_data, 0);
1477                         BUG_ON(j >= h->s->nr_data);
1478
1479                         h->s->blocks[j] = buckets.v[i];
1480                         h->s->new_stripe.key.v.ptrs[j] = bch2_ob_ptr(c, ob);
1481                         __set_bit(j, h->s->blocks_gotten);
1482                 }
1483
1484                 if (ret)
1485                         return ret;
1486         }
1487
1488         return 0;
1489 }
1490
1491 /* XXX: doesn't obey target: */
1492 static s64 get_existing_stripe(struct bch_fs *c,
1493                                struct ec_stripe_head *head)
1494 {
1495         ec_stripes_heap *h = &c->ec_stripes_heap;
1496         struct stripe *m;
1497         size_t heap_idx;
1498         u64 stripe_idx;
1499         s64 ret = -1;
1500
1501         if (may_create_new_stripe(c))
1502                 return -1;
1503
1504         mutex_lock(&c->ec_stripes_heap_lock);
1505         for (heap_idx = 0; heap_idx < h->used; heap_idx++) {
1506                 /* No blocks worth reusing, stripe will just be deleted: */
1507                 if (!h->data[heap_idx].blocks_nonempty)
1508                         continue;
1509
1510                 stripe_idx = h->data[heap_idx].idx;
1511
1512                 m = genradix_ptr(&c->stripes, stripe_idx);
1513
1514                 if (m->algorithm        == head->algo &&
1515                     m->nr_redundant     == head->redundancy &&
1516                     m->sectors          == head->blocksize &&
1517                     m->blocks_nonempty  < m->nr_blocks - m->nr_redundant &&
1518                     bch2_try_open_stripe(c, head->s, stripe_idx)) {
1519                         ret = stripe_idx;
1520                         break;
1521                 }
1522         }
1523         mutex_unlock(&c->ec_stripes_heap_lock);
1524         return ret;
1525 }
1526
1527 static int __bch2_ec_stripe_head_reuse(struct btree_trans *trans, struct ec_stripe_head *h)
1528 {
1529         struct bch_fs *c = trans->c;
1530         unsigned i;
1531         s64 idx;
1532         int ret;
1533
1534         idx = get_existing_stripe(c, h);
1535         if (idx < 0)
1536                 return -BCH_ERR_ENOSPC_stripe_reuse;
1537
1538         h->s->have_existing_stripe = true;
1539         ret = get_stripe_key_trans(trans, idx, &h->s->existing_stripe);
1540         if (ret) {
1541                 bch2_fs_fatal_error(c, "error reading stripe key: %i", ret);
1542                 return ret;
1543         }
1544
1545         if (ec_stripe_buf_init(&h->s->existing_stripe, 0, h->blocksize)) {
1546                 /*
1547                  * this is a problem: we have deleted from the
1548                  * stripes heap already
1549                  */
1550                 BUG();
1551         }
1552
1553         BUG_ON(h->s->existing_stripe.size != h->blocksize);
1554         BUG_ON(h->s->existing_stripe.size != h->s->existing_stripe.key.v.sectors);
1555
1556         for (i = 0; i < h->s->existing_stripe.key.v.nr_blocks; i++) {
1557                 if (stripe_blockcount_get(&h->s->existing_stripe.key.v, i)) {
1558                         __set_bit(i, h->s->blocks_gotten);
1559                         __set_bit(i, h->s->blocks_allocated);
1560                 }
1561
1562                 ec_block_io(c, &h->s->existing_stripe, READ, i, &h->s->iodone);
1563         }
1564
1565         bkey_copy(&h->s->new_stripe.key.k_i,
1566                   &h->s->existing_stripe.key.k_i);
1567
1568         return 0;
1569 }
1570
1571 static int __bch2_ec_stripe_head_reserve(struct btree_trans *trans, struct ec_stripe_head *h)
1572 {
1573         struct bch_fs *c = trans->c;
1574         struct btree_iter iter;
1575         struct bkey_s_c k;
1576         struct bpos min_pos = POS(0, 1);
1577         struct bpos start_pos = bpos_max(min_pos, POS(0, c->ec_stripe_hint));
1578         int ret;
1579
1580         BUG_ON(h->s->res.sectors);
1581
1582         ret = bch2_disk_reservation_get(c, &h->s->res,
1583                                         h->blocksize,
1584                                         h->s->nr_parity, 0);
1585         if (ret)
1586                 return ret;
1587
1588         for_each_btree_key_norestart(trans, iter, BTREE_ID_stripes, start_pos,
1589                            BTREE_ITER_SLOTS|BTREE_ITER_INTENT, k, ret) {
1590                 if (bkey_gt(k.k->p, POS(0, U32_MAX))) {
1591                         if (start_pos.offset) {
1592                                 start_pos = min_pos;
1593                                 bch2_btree_iter_set_pos(&iter, start_pos);
1594                                 continue;
1595                         }
1596
1597                         ret = -BCH_ERR_ENOSPC_stripe_create;
1598                         break;
1599                 }
1600
1601                 if (bkey_deleted(k.k) &&
1602                     bch2_try_open_stripe(c, h->s, k.k->p.offset))
1603                         break;
1604         }
1605
1606         c->ec_stripe_hint = iter.pos.offset;
1607
1608         if (ret)
1609                 goto err;
1610
1611         ret = ec_stripe_mem_alloc(trans, &iter);
1612         if (ret) {
1613                 bch2_stripe_close(c, h->s);
1614                 goto err;
1615         }
1616
1617         h->s->new_stripe.key.k.p = iter.pos;
1618 out:
1619         bch2_trans_iter_exit(trans, &iter);
1620         return ret;
1621 err:
1622         bch2_disk_reservation_put(c, &h->s->res);
1623         goto out;
1624 }
1625
1626 struct ec_stripe_head *bch2_ec_stripe_head_get(struct btree_trans *trans,
1627                                                unsigned target,
1628                                                unsigned algo,
1629                                                unsigned redundancy,
1630                                                bool copygc,
1631                                                struct closure *cl)
1632 {
1633         struct bch_fs *c = trans->c;
1634         struct ec_stripe_head *h;
1635         int ret;
1636         bool needs_stripe_new;
1637
1638         h = __bch2_ec_stripe_head_get(trans, target, algo, redundancy, copygc);
1639         if (!h)
1640                 bch_err(c, "no stripe head");
1641         if (IS_ERR_OR_NULL(h))
1642                 return h;
1643
1644         needs_stripe_new = !h->s;
1645         if (needs_stripe_new) {
1646                 if (ec_new_stripe_alloc(c, h)) {
1647                         ret = -ENOMEM;
1648                         bch_err(c, "failed to allocate new stripe");
1649                         goto err;
1650                 }
1651
1652                 if (ec_stripe_buf_init(&h->s->new_stripe, 0, h->blocksize))
1653                         BUG();
1654         }
1655
1656         /*
1657          * Try reserve a new stripe before reusing an
1658          * existing stripe. This will prevent unnecessary
1659          * read amplification during write oriented workloads.
1660          */
1661         ret = 0;
1662         if (!h->s->allocated && !h->s->res.sectors && !h->s->have_existing_stripe)
1663                 ret = __bch2_ec_stripe_head_reserve(trans, h);
1664         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1665                 goto err;
1666
1667         if (ret && needs_stripe_new)
1668                 ret = __bch2_ec_stripe_head_reuse(trans, h);
1669         if (ret) {
1670                 bch_err_ratelimited(c, "failed to get stripe: %s", bch2_err_str(ret));
1671                 goto err;
1672         }
1673
1674         if (!h->s->allocated) {
1675                 ret = new_stripe_alloc_buckets(trans, h, cl);
1676                 if (ret)
1677                         goto err;
1678
1679                 h->s->allocated = true;
1680         }
1681
1682         BUG_ON(trans->restarted);
1683         return h;
1684 err:
1685         bch2_ec_stripe_head_put(c, h);
1686         return ERR_PTR(ret);
1687 }
1688
1689 void bch2_ec_stop_dev(struct bch_fs *c, struct bch_dev *ca)
1690 {
1691         struct ec_stripe_head *h;
1692         struct open_bucket *ob;
1693         unsigned i;
1694
1695         mutex_lock(&c->ec_stripe_head_lock);
1696         list_for_each_entry(h, &c->ec_stripe_head_list, list) {
1697
1698                 mutex_lock(&h->lock);
1699                 if (!h->s)
1700                         goto unlock;
1701
1702                 for (i = 0; i < h->s->new_stripe.key.v.nr_blocks; i++) {
1703                         if (!h->s->blocks[i])
1704                                 continue;
1705
1706                         ob = c->open_buckets + h->s->blocks[i];
1707                         if (ob->dev == ca->dev_idx)
1708                                 goto found;
1709                 }
1710                 goto unlock;
1711 found:
1712                 h->s->err = -EROFS;
1713                 ec_stripe_set_pending(c, h);
1714 unlock:
1715                 mutex_unlock(&h->lock);
1716         }
1717         mutex_unlock(&c->ec_stripe_head_lock);
1718 }
1719
1720 int bch2_stripes_read(struct bch_fs *c)
1721 {
1722         struct btree_trans trans;
1723         struct btree_iter iter;
1724         struct bkey_s_c k;
1725         const struct bch_stripe *s;
1726         struct stripe *m;
1727         unsigned i;
1728         int ret;
1729
1730         bch2_trans_init(&trans, c, 0, 0);
1731
1732         for_each_btree_key(&trans, iter, BTREE_ID_stripes, POS_MIN,
1733                            BTREE_ITER_PREFETCH, k, ret) {
1734                 if (k.k->type != KEY_TYPE_stripe)
1735                         continue;
1736
1737                 ret = __ec_stripe_mem_alloc(c, k.k->p.offset, GFP_KERNEL);
1738                 if (ret)
1739                         break;
1740
1741                 s = bkey_s_c_to_stripe(k).v;
1742
1743                 m = genradix_ptr(&c->stripes, k.k->p.offset);
1744                 m->sectors      = le16_to_cpu(s->sectors);
1745                 m->algorithm    = s->algorithm;
1746                 m->nr_blocks    = s->nr_blocks;
1747                 m->nr_redundant = s->nr_redundant;
1748                 m->blocks_nonempty = 0;
1749
1750                 for (i = 0; i < s->nr_blocks; i++)
1751                         m->blocks_nonempty += !!stripe_blockcount_get(s, i);
1752
1753                 bch2_stripes_heap_insert(c, m, k.k->p.offset);
1754         }
1755         bch2_trans_iter_exit(&trans, &iter);
1756
1757         bch2_trans_exit(&trans);
1758
1759         if (ret)
1760                 bch_err(c, "error reading stripes: %i", ret);
1761
1762         return ret;
1763 }
1764
1765 void bch2_stripes_heap_to_text(struct printbuf *out, struct bch_fs *c)
1766 {
1767         ec_stripes_heap *h = &c->ec_stripes_heap;
1768         struct stripe *m;
1769         size_t i;
1770
1771         mutex_lock(&c->ec_stripes_heap_lock);
1772         for (i = 0; i < min_t(size_t, h->used, 20); i++) {
1773                 m = genradix_ptr(&c->stripes, h->data[i].idx);
1774
1775                 prt_printf(out, "%zu %u/%u+%u\n", h->data[i].idx,
1776                        h->data[i].blocks_nonempty,
1777                        m->nr_blocks - m->nr_redundant,
1778                        m->nr_redundant);
1779         }
1780         mutex_unlock(&c->ec_stripes_heap_lock);
1781 }
1782
1783 void bch2_new_stripes_to_text(struct printbuf *out, struct bch_fs *c)
1784 {
1785         struct ec_stripe_head *h;
1786         struct ec_stripe_new *s;
1787
1788         mutex_lock(&c->ec_stripe_head_lock);
1789         list_for_each_entry(h, &c->ec_stripe_head_list, list) {
1790                 prt_printf(out, "target %u algo %u redundancy %u:\n",
1791                        h->target, h->algo, h->redundancy);
1792
1793                 if (h->s)
1794                         prt_printf(out, "\tpending: blocks %u+%u allocated %u\n",
1795                                h->s->nr_data, h->s->nr_parity,
1796                                bitmap_weight(h->s->blocks_allocated,
1797                                              h->s->nr_data));
1798         }
1799         mutex_unlock(&c->ec_stripe_head_lock);
1800
1801         mutex_lock(&c->ec_stripe_new_lock);
1802         list_for_each_entry(s, &c->ec_stripe_new_list, list) {
1803                 prt_printf(out, "\tin flight: blocks %u+%u pin %u\n",
1804                        s->nr_data, s->nr_parity,
1805                        atomic_read(&s->pin));
1806         }
1807         mutex_unlock(&c->ec_stripe_new_lock);
1808 }
1809
1810 void bch2_fs_ec_exit(struct bch_fs *c)
1811 {
1812         struct ec_stripe_head *h;
1813         unsigned i;
1814
1815         while (1) {
1816                 mutex_lock(&c->ec_stripe_head_lock);
1817                 h = list_first_entry_or_null(&c->ec_stripe_head_list,
1818                                              struct ec_stripe_head, list);
1819                 if (h)
1820                         list_del(&h->list);
1821                 mutex_unlock(&c->ec_stripe_head_lock);
1822                 if (!h)
1823                         break;
1824
1825                 if (h->s) {
1826                         for (i = 0; i < h->s->new_stripe.key.v.nr_blocks; i++)
1827                                 BUG_ON(h->s->blocks[i]);
1828
1829                         kfree(h->s);
1830                 }
1831                 kfree(h);
1832         }
1833
1834         BUG_ON(!list_empty(&c->ec_stripe_new_list));
1835
1836         free_heap(&c->ec_stripes_heap);
1837         genradix_free(&c->stripes);
1838         bioset_exit(&c->ec_bioset);
1839 }
1840
1841 void bch2_fs_ec_init_early(struct bch_fs *c)
1842 {
1843         INIT_WORK(&c->ec_stripe_create_work, ec_stripe_create_work);
1844         INIT_WORK(&c->ec_stripe_delete_work, ec_stripe_delete_work);
1845 }
1846
1847 int bch2_fs_ec_init(struct bch_fs *c)
1848 {
1849         spin_lock_init(&c->ec_stripes_new_lock);
1850
1851         return bioset_init(&c->ec_bioset, 1, offsetof(struct ec_bio, bio),
1852                            BIOSET_NEED_BVECS);
1853 }