1 // SPDX-License-Identifier: GPL-2.0 2 3 /* erasure coding */ 4 5 #include "bcachefs.h" 6 #include "alloc_background.h" 7 #include "alloc_foreground.h" 8 #include "backpointers.h" 9 #include "bkey_buf.h" 10 #include "bset.h" 11 #include "btree_gc.h" 12 #include "btree_update.h" 13 #include "btree_write_buffer.h" 14 #include "buckets.h" 15 #include "checksum.h" 16 #include "disk_groups.h" 17 #include "ec.h" 18 #include "error.h" 19 #include "io_read.h" 20 #include "keylist.h" 21 #include "recovery.h" 22 #include "replicas.h" 23 #include "super-io.h" 24 #include "util.h" 25 26 #include <linux/sort.h> 27 28 #ifdef __KERNEL__ 29 30 #include <linux/raid/pq.h> 31 #include <linux/raid/xor.h> 32 33 static void raid5_recov(unsigned disks, unsigned failed_idx, 34 size_t size, void **data) 35 { 36 unsigned i = 2, nr; 37 38 BUG_ON(failed_idx >= disks); 39 40 swap(data[0], data[failed_idx]); 41 memcpy(data[0], data[1], size); 42 43 while (i < disks) { 44 nr = min_t(unsigned, disks - i, MAX_XOR_BLOCKS); 45 xor_blocks(nr, size, data[0], data + i); 46 i += nr; 47 } 48 49 swap(data[0], data[failed_idx]); 50 } 51 52 static void raid_gen(int nd, int np, size_t size, void **v) 53 { 54 if (np >= 1) 55 raid5_recov(nd + np, nd, size, v); 56 if (np >= 2) 57 raid6_call.gen_syndrome(nd + np, size, v); 58 BUG_ON(np > 2); 59 } 60 61 static void raid_rec(int nr, int *ir, int nd, int np, size_t size, void **v) 62 { 63 switch (nr) { 64 case 0: 65 break; 66 case 1: 67 if (ir[0] < nd + 1) 68 raid5_recov(nd + 1, ir[0], size, v); 69 else 70 raid6_call.gen_syndrome(nd + np, size, v); 71 break; 72 case 2: 73 if (ir[1] < nd) { 74 /* data+data failure. */ 75 raid6_2data_recov(nd + np, size, ir[0], ir[1], v); 76 } else if (ir[0] < nd) { 77 /* data + p/q failure */ 78 79 if (ir[1] == nd) /* data + p failure */ 80 raid6_datap_recov(nd + np, size, ir[0], v); 81 else { /* data + q failure */ 82 raid5_recov(nd + 1, ir[0], size, v); 83 raid6_call.gen_syndrome(nd + np, size, v); 84 } 85 } else { 86 raid_gen(nd, np, size, v); 87 } 88 break; 89 default: 90 BUG(); 91 } 92 } 93 94 #else 95 96 #include <raid/raid.h> 97 98 #endif 99 100 struct ec_bio { 101 struct bch_dev *ca; 102 struct ec_stripe_buf *buf; 103 size_t idx; 104 struct bio bio; 105 }; 106 107 /* Stripes btree keys: */ 108 109 int bch2_stripe_invalid(struct bch_fs *c, struct bkey_s_c k, 110 enum bkey_invalid_flags flags, 111 struct printbuf *err) 112 { 113 const struct bch_stripe *s = bkey_s_c_to_stripe(k).v; 114 int ret = 0; 115 116 bkey_fsck_err_on(bkey_eq(k.k->p, POS_MIN) || 117 bpos_gt(k.k->p, POS(0, U32_MAX)), c, err, 118 stripe_pos_bad, 119 "stripe at bad pos"); 120 121 bkey_fsck_err_on(bkey_val_u64s(k.k) < stripe_val_u64s(s), c, err, 122 stripe_val_size_bad, 123 "incorrect value size (%zu < %u)", 124 bkey_val_u64s(k.k), stripe_val_u64s(s)); 125 126 ret = bch2_bkey_ptrs_invalid(c, k, flags, err); 127 fsck_err: 128 return ret; 129 } 130 131 void bch2_stripe_to_text(struct printbuf *out, struct bch_fs *c, 132 struct bkey_s_c k) 133 { 134 const struct bch_stripe *s = bkey_s_c_to_stripe(k).v; 135 unsigned i, nr_data = s->nr_blocks - s->nr_redundant; 136 137 prt_printf(out, "algo %u sectors %u blocks %u:%u csum %u gran %u", 138 s->algorithm, 139 le16_to_cpu(s->sectors), 140 nr_data, 141 s->nr_redundant, 142 s->csum_type, 143 1U << s->csum_granularity_bits); 144 145 for (i = 0; i < s->nr_blocks; i++) { 146 const struct bch_extent_ptr *ptr = s->ptrs + i; 147 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev); 148 u32 offset; 149 u64 b = sector_to_bucket_and_offset(ca, ptr->offset, &offset); 150 151 prt_printf(out, " %u:%llu:%u", ptr->dev, b, offset); 152 if (i < nr_data) 153 prt_printf(out, "#%u", stripe_blockcount_get(s, i)); 154 prt_printf(out, " gen %u", ptr->gen); 155 if (ptr_stale(ca, ptr)) 156 prt_printf(out, " stale"); 157 } 158 } 159 160 /* Triggers: */ 161 162 static int bch2_trans_mark_stripe_bucket(struct btree_trans *trans, 163 struct bkey_s_c_stripe s, 164 unsigned idx, bool deleting) 165 { 166 struct bch_fs *c = trans->c; 167 const struct bch_extent_ptr *ptr = &s.v->ptrs[idx]; 168 struct btree_iter iter; 169 struct bkey_i_alloc_v4 *a; 170 enum bch_data_type data_type = idx >= s.v->nr_blocks - s.v->nr_redundant 171 ? BCH_DATA_parity : 0; 172 s64 sectors = data_type ? le16_to_cpu(s.v->sectors) : 0; 173 int ret = 0; 174 175 if (deleting) 176 sectors = -sectors; 177 178 a = bch2_trans_start_alloc_update(trans, &iter, PTR_BUCKET_POS(c, ptr)); 179 if (IS_ERR(a)) 180 return PTR_ERR(a); 181 182 ret = bch2_check_bucket_ref(trans, s.s_c, ptr, sectors, data_type, 183 a->v.gen, a->v.data_type, 184 a->v.dirty_sectors); 185 if (ret) 186 goto err; 187 188 if (!deleting) { 189 if (bch2_trans_inconsistent_on(a->v.stripe || 190 a->v.stripe_redundancy, trans, 191 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: multiple stripes using same bucket (%u, %llu)", 192 iter.pos.inode, iter.pos.offset, a->v.gen, 193 bch2_data_type_str(a->v.data_type), 194 a->v.dirty_sectors, 195 a->v.stripe, s.k->p.offset)) { 196 ret = -EIO; 197 goto err; 198 } 199 200 if (bch2_trans_inconsistent_on(data_type && a->v.dirty_sectors, trans, 201 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: data already in stripe bucket %llu", 202 iter.pos.inode, iter.pos.offset, a->v.gen, 203 bch2_data_type_str(a->v.data_type), 204 a->v.dirty_sectors, 205 s.k->p.offset)) { 206 ret = -EIO; 207 goto err; 208 } 209 210 a->v.stripe = s.k->p.offset; 211 a->v.stripe_redundancy = s.v->nr_redundant; 212 a->v.data_type = BCH_DATA_stripe; 213 } else { 214 if (bch2_trans_inconsistent_on(a->v.stripe != s.k->p.offset || 215 a->v.stripe_redundancy != s.v->nr_redundant, trans, 216 "bucket %llu:%llu gen %u: not marked as stripe when deleting stripe %llu (got %u)", 217 iter.pos.inode, iter.pos.offset, a->v.gen, 218 s.k->p.offset, a->v.stripe)) { 219 ret = -EIO; 220 goto err; 221 } 222 223 a->v.stripe = 0; 224 a->v.stripe_redundancy = 0; 225 a->v.data_type = alloc_data_type(a->v, BCH_DATA_user); 226 } 227 228 a->v.dirty_sectors += sectors; 229 if (data_type) 230 a->v.data_type = !deleting ? data_type : 0; 231 232 ret = bch2_trans_update(trans, &iter, &a->k_i, 0); 233 if (ret) 234 goto err; 235 err: 236 bch2_trans_iter_exit(trans, &iter); 237 return ret; 238 } 239 240 static int mark_stripe_bucket(struct btree_trans *trans, 241 struct bkey_s_c k, 242 unsigned ptr_idx, 243 unsigned flags) 244 { 245 struct bch_fs *c = trans->c; 246 const struct bch_stripe *s = bkey_s_c_to_stripe(k).v; 247 unsigned nr_data = s->nr_blocks - s->nr_redundant; 248 bool parity = ptr_idx >= nr_data; 249 enum bch_data_type data_type = parity ? BCH_DATA_parity : BCH_DATA_stripe; 250 s64 sectors = parity ? le16_to_cpu(s->sectors) : 0; 251 const struct bch_extent_ptr *ptr = s->ptrs + ptr_idx; 252 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev); 253 struct bucket old, new, *g; 254 struct printbuf buf = PRINTBUF; 255 int ret = 0; 256 257 BUG_ON(!(flags & BTREE_TRIGGER_GC)); 258 259 /* * XXX doesn't handle deletion */ 260 261 percpu_down_read(&c->mark_lock); 262 g = PTR_GC_BUCKET(ca, ptr); 263 264 if (g->dirty_sectors || 265 (g->stripe && g->stripe != k.k->p.offset)) { 266 bch2_fs_inconsistent(c, 267 "bucket %u:%zu gen %u: multiple stripes using same bucket\n%s", 268 ptr->dev, PTR_BUCKET_NR(ca, ptr), g->gen, 269 (bch2_bkey_val_to_text(&buf, c, k), buf.buf)); 270 ret = -EINVAL; 271 goto err; 272 } 273 274 bucket_lock(g); 275 old = *g; 276 277 ret = bch2_check_bucket_ref(trans, k, ptr, sectors, data_type, 278 g->gen, g->data_type, 279 g->dirty_sectors); 280 if (ret) 281 goto err; 282 283 g->data_type = data_type; 284 g->dirty_sectors += sectors; 285 286 g->stripe = k.k->p.offset; 287 g->stripe_redundancy = s->nr_redundant; 288 new = *g; 289 err: 290 bucket_unlock(g); 291 if (!ret) 292 bch2_dev_usage_update_m(c, ca, &old, &new); 293 percpu_up_read(&c->mark_lock); 294 printbuf_exit(&buf); 295 return ret; 296 } 297 298 int bch2_trigger_stripe(struct btree_trans *trans, 299 enum btree_id btree_id, unsigned level, 300 struct bkey_s_c old, struct bkey_s _new, 301 unsigned flags) 302 { 303 struct bkey_s_c new = _new.s_c; 304 struct bch_fs *c = trans->c; 305 u64 idx = new.k->p.offset; 306 const struct bch_stripe *old_s = old.k->type == KEY_TYPE_stripe 307 ? bkey_s_c_to_stripe(old).v : NULL; 308 const struct bch_stripe *new_s = new.k->type == KEY_TYPE_stripe 309 ? bkey_s_c_to_stripe(new).v : NULL; 310 311 if (flags & BTREE_TRIGGER_TRANSACTIONAL) { 312 /* 313 * If the pointers aren't changing, we don't need to do anything: 314 */ 315 if (new_s && old_s && 316 new_s->nr_blocks == old_s->nr_blocks && 317 new_s->nr_redundant == old_s->nr_redundant && 318 !memcmp(old_s->ptrs, new_s->ptrs, 319 new_s->nr_blocks * sizeof(struct bch_extent_ptr))) 320 return 0; 321 322 BUG_ON(new_s && old_s && 323 (new_s->nr_blocks != old_s->nr_blocks || 324 new_s->nr_redundant != old_s->nr_redundant)); 325 326 if (new_s) { 327 s64 sectors = le16_to_cpu(new_s->sectors); 328 329 struct bch_replicas_padded r; 330 bch2_bkey_to_replicas(&r.e, new); 331 int ret = bch2_update_replicas_list(trans, &r.e, sectors * new_s->nr_redundant); 332 if (ret) 333 return ret; 334 } 335 336 if (old_s) { 337 s64 sectors = -((s64) le16_to_cpu(old_s->sectors)); 338 339 struct bch_replicas_padded r; 340 bch2_bkey_to_replicas(&r.e, old); 341 int ret = bch2_update_replicas_list(trans, &r.e, sectors * old_s->nr_redundant); 342 if (ret) 343 return ret; 344 } 345 346 unsigned nr_blocks = new_s ? new_s->nr_blocks : old_s->nr_blocks; 347 for (unsigned i = 0; i < nr_blocks; i++) { 348 if (new_s && old_s && 349 !memcmp(&new_s->ptrs[i], 350 &old_s->ptrs[i], 351 sizeof(new_s->ptrs[i]))) 352 continue; 353 354 if (new_s) { 355 int ret = bch2_trans_mark_stripe_bucket(trans, 356 bkey_s_c_to_stripe(new), i, false); 357 if (ret) 358 return ret; 359 } 360 361 if (old_s) { 362 int ret = bch2_trans_mark_stripe_bucket(trans, 363 bkey_s_c_to_stripe(old), i, true); 364 if (ret) 365 return ret; 366 } 367 } 368 } 369 370 if (flags & BTREE_TRIGGER_ATOMIC) { 371 struct stripe *m = genradix_ptr(&c->stripes, idx); 372 373 if (!m) { 374 struct printbuf buf1 = PRINTBUF; 375 struct printbuf buf2 = PRINTBUF; 376 377 bch2_bkey_val_to_text(&buf1, c, old); 378 bch2_bkey_val_to_text(&buf2, c, new); 379 bch_err_ratelimited(c, "error marking nonexistent stripe %llu while marking\n" 380 "old %s\n" 381 "new %s", idx, buf1.buf, buf2.buf); 382 printbuf_exit(&buf2); 383 printbuf_exit(&buf1); 384 bch2_inconsistent_error(c); 385 return -1; 386 } 387 388 if (!new_s) { 389 bch2_stripes_heap_del(c, m, idx); 390 391 memset(m, 0, sizeof(*m)); 392 } else { 393 m->sectors = le16_to_cpu(new_s->sectors); 394 m->algorithm = new_s->algorithm; 395 m->nr_blocks = new_s->nr_blocks; 396 m->nr_redundant = new_s->nr_redundant; 397 m->blocks_nonempty = 0; 398 399 for (unsigned i = 0; i < new_s->nr_blocks; i++) 400 m->blocks_nonempty += !!stripe_blockcount_get(new_s, i); 401 402 if (!old_s) 403 bch2_stripes_heap_insert(c, m, idx); 404 else 405 bch2_stripes_heap_update(c, m, idx); 406 } 407 } 408 409 if (flags & BTREE_TRIGGER_GC) { 410 struct gc_stripe *m = 411 genradix_ptr_alloc(&c->gc_stripes, idx, GFP_KERNEL); 412 413 if (!m) { 414 bch_err(c, "error allocating memory for gc_stripes, idx %llu", 415 idx); 416 return -BCH_ERR_ENOMEM_mark_stripe; 417 } 418 /* 419 * This will be wrong when we bring back runtime gc: we should 420 * be unmarking the old key and then marking the new key 421 */ 422 m->alive = true; 423 m->sectors = le16_to_cpu(new_s->sectors); 424 m->nr_blocks = new_s->nr_blocks; 425 m->nr_redundant = new_s->nr_redundant; 426 427 for (unsigned i = 0; i < new_s->nr_blocks; i++) 428 m->ptrs[i] = new_s->ptrs[i]; 429 430 bch2_bkey_to_replicas(&m->r.e, new); 431 432 /* 433 * gc recalculates this field from stripe ptr 434 * references: 435 */ 436 memset(m->block_sectors, 0, sizeof(m->block_sectors)); 437 438 for (unsigned i = 0; i < new_s->nr_blocks; i++) { 439 int ret = mark_stripe_bucket(trans, new, i, flags); 440 if (ret) 441 return ret; 442 } 443 444 int ret = bch2_update_replicas(c, new, &m->r.e, 445 ((s64) m->sectors * m->nr_redundant), 446 0, true); 447 if (ret) { 448 struct printbuf buf = PRINTBUF; 449 450 bch2_bkey_val_to_text(&buf, c, new); 451 bch2_fs_fatal_error(c, ": no replicas entry for %s", buf.buf); 452 printbuf_exit(&buf); 453 return ret; 454 } 455 } 456 457 return 0; 458 } 459 460 /* returns blocknr in stripe that we matched: */ 461 static const struct bch_extent_ptr *bkey_matches_stripe(struct bch_stripe *s, 462 struct bkey_s_c k, unsigned *block) 463 { 464 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k); 465 unsigned i, nr_data = s->nr_blocks - s->nr_redundant; 466 467 bkey_for_each_ptr(ptrs, ptr) 468 for (i = 0; i < nr_data; i++) 469 if (__bch2_ptr_matches_stripe(&s->ptrs[i], ptr, 470 le16_to_cpu(s->sectors))) { 471 *block = i; 472 return ptr; 473 } 474 475 return NULL; 476 } 477 478 static bool extent_has_stripe_ptr(struct bkey_s_c k, u64 idx) 479 { 480 switch (k.k->type) { 481 case KEY_TYPE_extent: { 482 struct bkey_s_c_extent e = bkey_s_c_to_extent(k); 483 const union bch_extent_entry *entry; 484 485 extent_for_each_entry(e, entry) 486 if (extent_entry_type(entry) == 487 BCH_EXTENT_ENTRY_stripe_ptr && 488 entry->stripe_ptr.idx == idx) 489 return true; 490 491 break; 492 } 493 } 494 495 return false; 496 } 497 498 /* Stripe bufs: */ 499 500 static void ec_stripe_buf_exit(struct ec_stripe_buf *buf) 501 { 502 if (buf->key.k.type == KEY_TYPE_stripe) { 503 struct bkey_i_stripe *s = bkey_i_to_stripe(&buf->key); 504 unsigned i; 505 506 for (i = 0; i < s->v.nr_blocks; i++) { 507 kvfree(buf->data[i]); 508 buf->data[i] = NULL; 509 } 510 } 511 } 512 513 /* XXX: this is a non-mempoolified memory allocation: */ 514 static int ec_stripe_buf_init(struct ec_stripe_buf *buf, 515 unsigned offset, unsigned size) 516 { 517 struct bch_stripe *v = &bkey_i_to_stripe(&buf->key)->v; 518 unsigned csum_granularity = 1U << v->csum_granularity_bits; 519 unsigned end = offset + size; 520 unsigned i; 521 522 BUG_ON(end > le16_to_cpu(v->sectors)); 523 524 offset = round_down(offset, csum_granularity); 525 end = min_t(unsigned, le16_to_cpu(v->sectors), 526 round_up(end, csum_granularity)); 527 528 buf->offset = offset; 529 buf->size = end - offset; 530 531 memset(buf->valid, 0xFF, sizeof(buf->valid)); 532 533 for (i = 0; i < v->nr_blocks; i++) { 534 buf->data[i] = kvmalloc(buf->size << 9, GFP_KERNEL); 535 if (!buf->data[i]) 536 goto err; 537 } 538 539 return 0; 540 err: 541 ec_stripe_buf_exit(buf); 542 return -BCH_ERR_ENOMEM_stripe_buf; 543 } 544 545 /* Checksumming: */ 546 547 static struct bch_csum ec_block_checksum(struct ec_stripe_buf *buf, 548 unsigned block, unsigned offset) 549 { 550 struct bch_stripe *v = &bkey_i_to_stripe(&buf->key)->v; 551 unsigned csum_granularity = 1 << v->csum_granularity_bits; 552 unsigned end = buf->offset + buf->size; 553 unsigned len = min(csum_granularity, end - offset); 554 555 BUG_ON(offset >= end); 556 BUG_ON(offset < buf->offset); 557 BUG_ON(offset & (csum_granularity - 1)); 558 BUG_ON(offset + len != le16_to_cpu(v->sectors) && 559 (len & (csum_granularity - 1))); 560 561 return bch2_checksum(NULL, v->csum_type, 562 null_nonce(), 563 buf->data[block] + ((offset - buf->offset) << 9), 564 len << 9); 565 } 566 567 static void ec_generate_checksums(struct ec_stripe_buf *buf) 568 { 569 struct bch_stripe *v = &bkey_i_to_stripe(&buf->key)->v; 570 unsigned i, j, csums_per_device = stripe_csums_per_device(v); 571 572 if (!v->csum_type) 573 return; 574 575 BUG_ON(buf->offset); 576 BUG_ON(buf->size != le16_to_cpu(v->sectors)); 577 578 for (i = 0; i < v->nr_blocks; i++) 579 for (j = 0; j < csums_per_device; j++) 580 stripe_csum_set(v, i, j, 581 ec_block_checksum(buf, i, j << v->csum_granularity_bits)); 582 } 583 584 static void ec_validate_checksums(struct bch_fs *c, struct ec_stripe_buf *buf) 585 { 586 struct bch_stripe *v = &bkey_i_to_stripe(&buf->key)->v; 587 unsigned csum_granularity = 1 << v->csum_granularity_bits; 588 unsigned i; 589 590 if (!v->csum_type) 591 return; 592 593 for (i = 0; i < v->nr_blocks; i++) { 594 unsigned offset = buf->offset; 595 unsigned end = buf->offset + buf->size; 596 597 if (!test_bit(i, buf->valid)) 598 continue; 599 600 while (offset < end) { 601 unsigned j = offset >> v->csum_granularity_bits; 602 unsigned len = min(csum_granularity, end - offset); 603 struct bch_csum want = stripe_csum_get(v, i, j); 604 struct bch_csum got = ec_block_checksum(buf, i, offset); 605 606 if (bch2_crc_cmp(want, got)) { 607 struct printbuf err = PRINTBUF; 608 struct bch_dev *ca = bch_dev_bkey_exists(c, v->ptrs[i].dev); 609 610 prt_printf(&err, "stripe checksum error: expected %0llx:%0llx got %0llx:%0llx (type %s)\n", 611 want.hi, want.lo, 612 got.hi, got.lo, 613 bch2_csum_types[v->csum_type]); 614 prt_printf(&err, " for %ps at %u of\n ", (void *) _RET_IP_, i); 615 bch2_bkey_val_to_text(&err, c, bkey_i_to_s_c(&buf->key)); 616 bch_err_ratelimited(ca, "%s", err.buf); 617 printbuf_exit(&err); 618 619 clear_bit(i, buf->valid); 620 621 bch2_io_error(ca, BCH_MEMBER_ERROR_checksum); 622 break; 623 } 624 625 offset += len; 626 } 627 } 628 } 629 630 /* Erasure coding: */ 631 632 static void ec_generate_ec(struct ec_stripe_buf *buf) 633 { 634 struct bch_stripe *v = &bkey_i_to_stripe(&buf->key)->v; 635 unsigned nr_data = v->nr_blocks - v->nr_redundant; 636 unsigned bytes = le16_to_cpu(v->sectors) << 9; 637 638 raid_gen(nr_data, v->nr_redundant, bytes, buf->data); 639 } 640 641 static unsigned ec_nr_failed(struct ec_stripe_buf *buf) 642 { 643 struct bch_stripe *v = &bkey_i_to_stripe(&buf->key)->v; 644 645 return v->nr_blocks - bitmap_weight(buf->valid, v->nr_blocks); 646 } 647 648 static int ec_do_recov(struct bch_fs *c, struct ec_stripe_buf *buf) 649 { 650 struct bch_stripe *v = &bkey_i_to_stripe(&buf->key)->v; 651 unsigned i, failed[BCH_BKEY_PTRS_MAX], nr_failed = 0; 652 unsigned nr_data = v->nr_blocks - v->nr_redundant; 653 unsigned bytes = buf->size << 9; 654 655 if (ec_nr_failed(buf) > v->nr_redundant) { 656 bch_err_ratelimited(c, 657 "error doing reconstruct read: unable to read enough blocks"); 658 return -1; 659 } 660 661 for (i = 0; i < nr_data; i++) 662 if (!test_bit(i, buf->valid)) 663 failed[nr_failed++] = i; 664 665 raid_rec(nr_failed, failed, nr_data, v->nr_redundant, bytes, buf->data); 666 return 0; 667 } 668 669 /* IO: */ 670 671 static void ec_block_endio(struct bio *bio) 672 { 673 struct ec_bio *ec_bio = container_of(bio, struct ec_bio, bio); 674 struct bch_stripe *v = &bkey_i_to_stripe(&ec_bio->buf->key)->v; 675 struct bch_extent_ptr *ptr = &v->ptrs[ec_bio->idx]; 676 struct bch_dev *ca = ec_bio->ca; 677 struct closure *cl = bio->bi_private; 678 679 if (bch2_dev_io_err_on(bio->bi_status, ca, 680 bio_data_dir(bio) 681 ? BCH_MEMBER_ERROR_write 682 : BCH_MEMBER_ERROR_read, 683 "erasure coding %s error: %s", 684 bio_data_dir(bio) ? "write" : "read", 685 bch2_blk_status_to_str(bio->bi_status))) 686 clear_bit(ec_bio->idx, ec_bio->buf->valid); 687 688 if (ptr_stale(ca, ptr)) { 689 bch_err_ratelimited(ca->fs, 690 "error %s stripe: stale pointer after io", 691 bio_data_dir(bio) == READ ? "reading from" : "writing to"); 692 clear_bit(ec_bio->idx, ec_bio->buf->valid); 693 } 694 695 bio_put(&ec_bio->bio); 696 percpu_ref_put(&ca->io_ref); 697 closure_put(cl); 698 } 699 700 static void ec_block_io(struct bch_fs *c, struct ec_stripe_buf *buf, 701 blk_opf_t opf, unsigned idx, struct closure *cl) 702 { 703 struct bch_stripe *v = &bkey_i_to_stripe(&buf->key)->v; 704 unsigned offset = 0, bytes = buf->size << 9; 705 struct bch_extent_ptr *ptr = &v->ptrs[idx]; 706 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev); 707 enum bch_data_type data_type = idx < v->nr_blocks - v->nr_redundant 708 ? BCH_DATA_user 709 : BCH_DATA_parity; 710 int rw = op_is_write(opf); 711 712 if (ptr_stale(ca, ptr)) { 713 bch_err_ratelimited(c, 714 "error %s stripe: stale pointer", 715 rw == READ ? "reading from" : "writing to"); 716 clear_bit(idx, buf->valid); 717 return; 718 } 719 720 if (!bch2_dev_get_ioref(ca, rw)) { 721 clear_bit(idx, buf->valid); 722 return; 723 } 724 725 this_cpu_add(ca->io_done->sectors[rw][data_type], buf->size); 726 727 while (offset < bytes) { 728 unsigned nr_iovecs = min_t(size_t, BIO_MAX_VECS, 729 DIV_ROUND_UP(bytes, PAGE_SIZE)); 730 unsigned b = min_t(size_t, bytes - offset, 731 nr_iovecs << PAGE_SHIFT); 732 struct ec_bio *ec_bio; 733 734 ec_bio = container_of(bio_alloc_bioset(ca->disk_sb.bdev, 735 nr_iovecs, 736 opf, 737 GFP_KERNEL, 738 &c->ec_bioset), 739 struct ec_bio, bio); 740 741 ec_bio->ca = ca; 742 ec_bio->buf = buf; 743 ec_bio->idx = idx; 744 745 ec_bio->bio.bi_iter.bi_sector = ptr->offset + buf->offset + (offset >> 9); 746 ec_bio->bio.bi_end_io = ec_block_endio; 747 ec_bio->bio.bi_private = cl; 748 749 bch2_bio_map(&ec_bio->bio, buf->data[idx] + offset, b); 750 751 closure_get(cl); 752 percpu_ref_get(&ca->io_ref); 753 754 submit_bio(&ec_bio->bio); 755 756 offset += b; 757 } 758 759 percpu_ref_put(&ca->io_ref); 760 } 761 762 static int get_stripe_key_trans(struct btree_trans *trans, u64 idx, 763 struct ec_stripe_buf *stripe) 764 { 765 struct btree_iter iter; 766 struct bkey_s_c k; 767 int ret; 768 769 k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_stripes, 770 POS(0, idx), BTREE_ITER_SLOTS); 771 ret = bkey_err(k); 772 if (ret) 773 goto err; 774 if (k.k->type != KEY_TYPE_stripe) { 775 ret = -ENOENT; 776 goto err; 777 } 778 bkey_reassemble(&stripe->key, k); 779 err: 780 bch2_trans_iter_exit(trans, &iter); 781 return ret; 782 } 783 784 /* recovery read path: */ 785 int bch2_ec_read_extent(struct btree_trans *trans, struct bch_read_bio *rbio) 786 { 787 struct bch_fs *c = trans->c; 788 struct ec_stripe_buf *buf; 789 struct closure cl; 790 struct bch_stripe *v; 791 unsigned i, offset; 792 int ret = 0; 793 794 closure_init_stack(&cl); 795 796 BUG_ON(!rbio->pick.has_ec); 797 798 buf = kzalloc(sizeof(*buf), GFP_NOFS); 799 if (!buf) 800 return -BCH_ERR_ENOMEM_ec_read_extent; 801 802 ret = lockrestart_do(trans, get_stripe_key_trans(trans, rbio->pick.ec.idx, buf)); 803 if (ret) { 804 bch_err_ratelimited(c, 805 "error doing reconstruct read: error %i looking up stripe", ret); 806 kfree(buf); 807 return -EIO; 808 } 809 810 v = &bkey_i_to_stripe(&buf->key)->v; 811 812 if (!bch2_ptr_matches_stripe(v, rbio->pick)) { 813 bch_err_ratelimited(c, 814 "error doing reconstruct read: pointer doesn't match stripe"); 815 ret = -EIO; 816 goto err; 817 } 818 819 offset = rbio->bio.bi_iter.bi_sector - v->ptrs[rbio->pick.ec.block].offset; 820 if (offset + bio_sectors(&rbio->bio) > le16_to_cpu(v->sectors)) { 821 bch_err_ratelimited(c, 822 "error doing reconstruct read: read is bigger than stripe"); 823 ret = -EIO; 824 goto err; 825 } 826 827 ret = ec_stripe_buf_init(buf, offset, bio_sectors(&rbio->bio)); 828 if (ret) 829 goto err; 830 831 for (i = 0; i < v->nr_blocks; i++) 832 ec_block_io(c, buf, REQ_OP_READ, i, &cl); 833 834 closure_sync(&cl); 835 836 if (ec_nr_failed(buf) > v->nr_redundant) { 837 bch_err_ratelimited(c, 838 "error doing reconstruct read: unable to read enough blocks"); 839 ret = -EIO; 840 goto err; 841 } 842 843 ec_validate_checksums(c, buf); 844 845 ret = ec_do_recov(c, buf); 846 if (ret) 847 goto err; 848 849 memcpy_to_bio(&rbio->bio, rbio->bio.bi_iter, 850 buf->data[rbio->pick.ec.block] + ((offset - buf->offset) << 9)); 851 err: 852 ec_stripe_buf_exit(buf); 853 kfree(buf); 854 return ret; 855 } 856 857 /* stripe bucket accounting: */ 858 859 static int __ec_stripe_mem_alloc(struct bch_fs *c, size_t idx, gfp_t gfp) 860 { 861 ec_stripes_heap n, *h = &c->ec_stripes_heap; 862 863 if (idx >= h->size) { 864 if (!init_heap(&n, max(1024UL, roundup_pow_of_two(idx + 1)), gfp)) 865 return -BCH_ERR_ENOMEM_ec_stripe_mem_alloc; 866 867 mutex_lock(&c->ec_stripes_heap_lock); 868 if (n.size > h->size) { 869 memcpy(n.data, h->data, h->used * sizeof(h->data[0])); 870 n.used = h->used; 871 swap(*h, n); 872 } 873 mutex_unlock(&c->ec_stripes_heap_lock); 874 875 free_heap(&n); 876 } 877 878 if (!genradix_ptr_alloc(&c->stripes, idx, gfp)) 879 return -BCH_ERR_ENOMEM_ec_stripe_mem_alloc; 880 881 if (c->gc_pos.phase != GC_PHASE_NOT_RUNNING && 882 !genradix_ptr_alloc(&c->gc_stripes, idx, gfp)) 883 return -BCH_ERR_ENOMEM_ec_stripe_mem_alloc; 884 885 return 0; 886 } 887 888 static int ec_stripe_mem_alloc(struct btree_trans *trans, 889 struct btree_iter *iter) 890 { 891 return allocate_dropping_locks_errcode(trans, 892 __ec_stripe_mem_alloc(trans->c, iter->pos.offset, _gfp)); 893 } 894 895 /* 896 * Hash table of open stripes: 897 * Stripes that are being created or modified are kept in a hash table, so that 898 * stripe deletion can skip them. 899 */ 900 901 static bool __bch2_stripe_is_open(struct bch_fs *c, u64 idx) 902 { 903 unsigned hash = hash_64(idx, ilog2(ARRAY_SIZE(c->ec_stripes_new))); 904 struct ec_stripe_new *s; 905 906 hlist_for_each_entry(s, &c->ec_stripes_new[hash], hash) 907 if (s->idx == idx) 908 return true; 909 return false; 910 } 911 912 static bool bch2_stripe_is_open(struct bch_fs *c, u64 idx) 913 { 914 bool ret = false; 915 916 spin_lock(&c->ec_stripes_new_lock); 917 ret = __bch2_stripe_is_open(c, idx); 918 spin_unlock(&c->ec_stripes_new_lock); 919 920 return ret; 921 } 922 923 static bool bch2_try_open_stripe(struct bch_fs *c, 924 struct ec_stripe_new *s, 925 u64 idx) 926 { 927 bool ret; 928 929 spin_lock(&c->ec_stripes_new_lock); 930 ret = !__bch2_stripe_is_open(c, idx); 931 if (ret) { 932 unsigned hash = hash_64(idx, ilog2(ARRAY_SIZE(c->ec_stripes_new))); 933 934 s->idx = idx; 935 hlist_add_head(&s->hash, &c->ec_stripes_new[hash]); 936 } 937 spin_unlock(&c->ec_stripes_new_lock); 938 939 return ret; 940 } 941 942 static void bch2_stripe_close(struct bch_fs *c, struct ec_stripe_new *s) 943 { 944 BUG_ON(!s->idx); 945 946 spin_lock(&c->ec_stripes_new_lock); 947 hlist_del_init(&s->hash); 948 spin_unlock(&c->ec_stripes_new_lock); 949 950 s->idx = 0; 951 } 952 953 /* Heap of all existing stripes, ordered by blocks_nonempty */ 954 955 static u64 stripe_idx_to_delete(struct bch_fs *c) 956 { 957 ec_stripes_heap *h = &c->ec_stripes_heap; 958 959 lockdep_assert_held(&c->ec_stripes_heap_lock); 960 961 if (h->used && 962 h->data[0].blocks_nonempty == 0 && 963 !bch2_stripe_is_open(c, h->data[0].idx)) 964 return h->data[0].idx; 965 966 return 0; 967 } 968 969 static inline int ec_stripes_heap_cmp(ec_stripes_heap *h, 970 struct ec_stripe_heap_entry l, 971 struct ec_stripe_heap_entry r) 972 { 973 return ((l.blocks_nonempty > r.blocks_nonempty) - 974 (l.blocks_nonempty < r.blocks_nonempty)); 975 } 976 977 static inline void ec_stripes_heap_set_backpointer(ec_stripes_heap *h, 978 size_t i) 979 { 980 struct bch_fs *c = container_of(h, struct bch_fs, ec_stripes_heap); 981 982 genradix_ptr(&c->stripes, h->data[i].idx)->heap_idx = i; 983 } 984 985 static void heap_verify_backpointer(struct bch_fs *c, size_t idx) 986 { 987 ec_stripes_heap *h = &c->ec_stripes_heap; 988 struct stripe *m = genradix_ptr(&c->stripes, idx); 989 990 BUG_ON(m->heap_idx >= h->used); 991 BUG_ON(h->data[m->heap_idx].idx != idx); 992 } 993 994 void bch2_stripes_heap_del(struct bch_fs *c, 995 struct stripe *m, size_t idx) 996 { 997 mutex_lock(&c->ec_stripes_heap_lock); 998 heap_verify_backpointer(c, idx); 999 1000 heap_del(&c->ec_stripes_heap, m->heap_idx, 1001 ec_stripes_heap_cmp, 1002 ec_stripes_heap_set_backpointer); 1003 mutex_unlock(&c->ec_stripes_heap_lock); 1004 } 1005 1006 void bch2_stripes_heap_insert(struct bch_fs *c, 1007 struct stripe *m, size_t idx) 1008 { 1009 mutex_lock(&c->ec_stripes_heap_lock); 1010 BUG_ON(heap_full(&c->ec_stripes_heap)); 1011 1012 heap_add(&c->ec_stripes_heap, ((struct ec_stripe_heap_entry) { 1013 .idx = idx, 1014 .blocks_nonempty = m->blocks_nonempty, 1015 }), 1016 ec_stripes_heap_cmp, 1017 ec_stripes_heap_set_backpointer); 1018 1019 heap_verify_backpointer(c, idx); 1020 mutex_unlock(&c->ec_stripes_heap_lock); 1021 } 1022 1023 void bch2_stripes_heap_update(struct bch_fs *c, 1024 struct stripe *m, size_t idx) 1025 { 1026 ec_stripes_heap *h = &c->ec_stripes_heap; 1027 bool do_deletes; 1028 size_t i; 1029 1030 mutex_lock(&c->ec_stripes_heap_lock); 1031 heap_verify_backpointer(c, idx); 1032 1033 h->data[m->heap_idx].blocks_nonempty = m->blocks_nonempty; 1034 1035 i = m->heap_idx; 1036 heap_sift_up(h, i, ec_stripes_heap_cmp, 1037 ec_stripes_heap_set_backpointer); 1038 heap_sift_down(h, i, ec_stripes_heap_cmp, 1039 ec_stripes_heap_set_backpointer); 1040 1041 heap_verify_backpointer(c, idx); 1042 1043 do_deletes = stripe_idx_to_delete(c) != 0; 1044 mutex_unlock(&c->ec_stripes_heap_lock); 1045 1046 if (do_deletes) 1047 bch2_do_stripe_deletes(c); 1048 } 1049 1050 /* stripe deletion */ 1051 1052 static int ec_stripe_delete(struct btree_trans *trans, u64 idx) 1053 { 1054 struct bch_fs *c = trans->c; 1055 struct btree_iter iter; 1056 struct bkey_s_c k; 1057 struct bkey_s_c_stripe s; 1058 int ret; 1059 1060 k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_stripes, POS(0, idx), 1061 BTREE_ITER_INTENT); 1062 ret = bkey_err(k); 1063 if (ret) 1064 goto err; 1065 1066 if (k.k->type != KEY_TYPE_stripe) { 1067 bch2_fs_inconsistent(c, "attempting to delete nonexistent stripe %llu", idx); 1068 ret = -EINVAL; 1069 goto err; 1070 } 1071 1072 s = bkey_s_c_to_stripe(k); 1073 for (unsigned i = 0; i < s.v->nr_blocks; i++) 1074 if (stripe_blockcount_get(s.v, i)) { 1075 struct printbuf buf = PRINTBUF; 1076 1077 bch2_bkey_val_to_text(&buf, c, k); 1078 bch2_fs_inconsistent(c, "attempting to delete nonempty stripe %s", buf.buf); 1079 printbuf_exit(&buf); 1080 ret = -EINVAL; 1081 goto err; 1082 } 1083 1084 ret = bch2_btree_delete_at(trans, &iter, 0); 1085 err: 1086 bch2_trans_iter_exit(trans, &iter); 1087 return ret; 1088 } 1089 1090 static void ec_stripe_delete_work(struct work_struct *work) 1091 { 1092 struct bch_fs *c = 1093 container_of(work, struct bch_fs, ec_stripe_delete_work); 1094 1095 while (1) { 1096 mutex_lock(&c->ec_stripes_heap_lock); 1097 u64 idx = stripe_idx_to_delete(c); 1098 mutex_unlock(&c->ec_stripes_heap_lock); 1099 1100 if (!idx) 1101 break; 1102 1103 int ret = bch2_trans_do(c, NULL, NULL, BCH_TRANS_COMMIT_no_enospc, 1104 ec_stripe_delete(trans, idx)); 1105 bch_err_fn(c, ret); 1106 if (ret) 1107 break; 1108 } 1109 1110 bch2_write_ref_put(c, BCH_WRITE_REF_stripe_delete); 1111 } 1112 1113 void bch2_do_stripe_deletes(struct bch_fs *c) 1114 { 1115 if (bch2_write_ref_tryget(c, BCH_WRITE_REF_stripe_delete) && 1116 !queue_work(c->write_ref_wq, &c->ec_stripe_delete_work)) 1117 bch2_write_ref_put(c, BCH_WRITE_REF_stripe_delete); 1118 } 1119 1120 /* stripe creation: */ 1121 1122 static int ec_stripe_key_update(struct btree_trans *trans, 1123 struct bkey_i_stripe *new, 1124 bool create) 1125 { 1126 struct bch_fs *c = trans->c; 1127 struct btree_iter iter; 1128 struct bkey_s_c k; 1129 int ret; 1130 1131 k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_stripes, 1132 new->k.p, BTREE_ITER_INTENT); 1133 ret = bkey_err(k); 1134 if (ret) 1135 goto err; 1136 1137 if (k.k->type != (create ? KEY_TYPE_deleted : KEY_TYPE_stripe)) { 1138 bch2_fs_inconsistent(c, "error %s stripe: got existing key type %s", 1139 create ? "creating" : "updating", 1140 bch2_bkey_types[k.k->type]); 1141 ret = -EINVAL; 1142 goto err; 1143 } 1144 1145 if (k.k->type == KEY_TYPE_stripe) { 1146 const struct bch_stripe *old = bkey_s_c_to_stripe(k).v; 1147 unsigned i; 1148 1149 if (old->nr_blocks != new->v.nr_blocks) { 1150 bch_err(c, "error updating stripe: nr_blocks does not match"); 1151 ret = -EINVAL; 1152 goto err; 1153 } 1154 1155 for (i = 0; i < new->v.nr_blocks; i++) { 1156 unsigned v = stripe_blockcount_get(old, i); 1157 1158 BUG_ON(v && 1159 (old->ptrs[i].dev != new->v.ptrs[i].dev || 1160 old->ptrs[i].gen != new->v.ptrs[i].gen || 1161 old->ptrs[i].offset != new->v.ptrs[i].offset)); 1162 1163 stripe_blockcount_set(&new->v, i, v); 1164 } 1165 } 1166 1167 ret = bch2_trans_update(trans, &iter, &new->k_i, 0); 1168 err: 1169 bch2_trans_iter_exit(trans, &iter); 1170 return ret; 1171 } 1172 1173 static int ec_stripe_update_extent(struct btree_trans *trans, 1174 struct bpos bucket, u8 gen, 1175 struct ec_stripe_buf *s, 1176 struct bpos *bp_pos) 1177 { 1178 struct bch_stripe *v = &bkey_i_to_stripe(&s->key)->v; 1179 struct bch_fs *c = trans->c; 1180 struct bch_backpointer bp; 1181 struct btree_iter iter; 1182 struct bkey_s_c k; 1183 const struct bch_extent_ptr *ptr_c; 1184 struct bch_extent_ptr *ptr, *ec_ptr = NULL; 1185 struct bch_extent_stripe_ptr stripe_ptr; 1186 struct bkey_i *n; 1187 int ret, dev, block; 1188 1189 ret = bch2_get_next_backpointer(trans, bucket, gen, 1190 bp_pos, &bp, BTREE_ITER_CACHED); 1191 if (ret) 1192 return ret; 1193 if (bpos_eq(*bp_pos, SPOS_MAX)) 1194 return 0; 1195 1196 if (bp.level) { 1197 struct printbuf buf = PRINTBUF; 1198 struct btree_iter node_iter; 1199 struct btree *b; 1200 1201 b = bch2_backpointer_get_node(trans, &node_iter, *bp_pos, bp); 1202 bch2_trans_iter_exit(trans, &node_iter); 1203 1204 if (!b) 1205 return 0; 1206 1207 prt_printf(&buf, "found btree node in erasure coded bucket: b=%px\n", b); 1208 bch2_backpointer_to_text(&buf, &bp); 1209 1210 bch2_fs_inconsistent(c, "%s", buf.buf); 1211 printbuf_exit(&buf); 1212 return -EIO; 1213 } 1214 1215 k = bch2_backpointer_get_key(trans, &iter, *bp_pos, bp, BTREE_ITER_INTENT); 1216 ret = bkey_err(k); 1217 if (ret) 1218 return ret; 1219 if (!k.k) { 1220 /* 1221 * extent no longer exists - we could flush the btree 1222 * write buffer and retry to verify, but no need: 1223 */ 1224 return 0; 1225 } 1226 1227 if (extent_has_stripe_ptr(k, s->key.k.p.offset)) 1228 goto out; 1229 1230 ptr_c = bkey_matches_stripe(v, k, &block); 1231 /* 1232 * It doesn't generally make sense to erasure code cached ptrs: 1233 * XXX: should we be incrementing a counter? 1234 */ 1235 if (!ptr_c || ptr_c->cached) 1236 goto out; 1237 1238 dev = v->ptrs[block].dev; 1239 1240 n = bch2_trans_kmalloc(trans, bkey_bytes(k.k) + sizeof(stripe_ptr)); 1241 ret = PTR_ERR_OR_ZERO(n); 1242 if (ret) 1243 goto out; 1244 1245 bkey_reassemble(n, k); 1246 1247 bch2_bkey_drop_ptrs(bkey_i_to_s(n), ptr, ptr->dev != dev); 1248 ec_ptr = bch2_bkey_has_device(bkey_i_to_s(n), dev); 1249 BUG_ON(!ec_ptr); 1250 1251 stripe_ptr = (struct bch_extent_stripe_ptr) { 1252 .type = 1 << BCH_EXTENT_ENTRY_stripe_ptr, 1253 .block = block, 1254 .redundancy = v->nr_redundant, 1255 .idx = s->key.k.p.offset, 1256 }; 1257 1258 __extent_entry_insert(n, 1259 (union bch_extent_entry *) ec_ptr, 1260 (union bch_extent_entry *) &stripe_ptr); 1261 1262 ret = bch2_trans_update(trans, &iter, n, 0); 1263 out: 1264 bch2_trans_iter_exit(trans, &iter); 1265 return ret; 1266 } 1267 1268 static int ec_stripe_update_bucket(struct btree_trans *trans, struct ec_stripe_buf *s, 1269 unsigned block) 1270 { 1271 struct bch_fs *c = trans->c; 1272 struct bch_stripe *v = &bkey_i_to_stripe(&s->key)->v; 1273 struct bch_extent_ptr bucket = v->ptrs[block]; 1274 struct bpos bucket_pos = PTR_BUCKET_POS(c, &bucket); 1275 struct bpos bp_pos = POS_MIN; 1276 int ret = 0; 1277 1278 while (1) { 1279 ret = commit_do(trans, NULL, NULL, 1280 BCH_TRANS_COMMIT_no_check_rw| 1281 BCH_TRANS_COMMIT_no_enospc, 1282 ec_stripe_update_extent(trans, bucket_pos, bucket.gen, 1283 s, &bp_pos)); 1284 if (ret) 1285 break; 1286 if (bkey_eq(bp_pos, POS_MAX)) 1287 break; 1288 1289 bp_pos = bpos_nosnap_successor(bp_pos); 1290 } 1291 1292 return ret; 1293 } 1294 1295 static int ec_stripe_update_extents(struct bch_fs *c, struct ec_stripe_buf *s) 1296 { 1297 struct btree_trans *trans = bch2_trans_get(c); 1298 struct bch_stripe *v = &bkey_i_to_stripe(&s->key)->v; 1299 unsigned i, nr_data = v->nr_blocks - v->nr_redundant; 1300 int ret = 0; 1301 1302 ret = bch2_btree_write_buffer_flush_sync(trans); 1303 if (ret) 1304 goto err; 1305 1306 for (i = 0; i < nr_data; i++) { 1307 ret = ec_stripe_update_bucket(trans, s, i); 1308 if (ret) 1309 break; 1310 } 1311 err: 1312 bch2_trans_put(trans); 1313 1314 return ret; 1315 } 1316 1317 static void zero_out_rest_of_ec_bucket(struct bch_fs *c, 1318 struct ec_stripe_new *s, 1319 unsigned block, 1320 struct open_bucket *ob) 1321 { 1322 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev); 1323 unsigned offset = ca->mi.bucket_size - ob->sectors_free; 1324 int ret; 1325 1326 if (!bch2_dev_get_ioref(ca, WRITE)) { 1327 s->err = -BCH_ERR_erofs_no_writes; 1328 return; 1329 } 1330 1331 memset(s->new_stripe.data[block] + (offset << 9), 1332 0, 1333 ob->sectors_free << 9); 1334 1335 ret = blkdev_issue_zeroout(ca->disk_sb.bdev, 1336 ob->bucket * ca->mi.bucket_size + offset, 1337 ob->sectors_free, 1338 GFP_KERNEL, 0); 1339 1340 percpu_ref_put(&ca->io_ref); 1341 1342 if (ret) 1343 s->err = ret; 1344 } 1345 1346 void bch2_ec_stripe_new_free(struct bch_fs *c, struct ec_stripe_new *s) 1347 { 1348 if (s->idx) 1349 bch2_stripe_close(c, s); 1350 kfree(s); 1351 } 1352 1353 /* 1354 * data buckets of new stripe all written: create the stripe 1355 */ 1356 static void ec_stripe_create(struct ec_stripe_new *s) 1357 { 1358 struct bch_fs *c = s->c; 1359 struct open_bucket *ob; 1360 struct bch_stripe *v = &bkey_i_to_stripe(&s->new_stripe.key)->v; 1361 unsigned i, nr_data = v->nr_blocks - v->nr_redundant; 1362 int ret; 1363 1364 BUG_ON(s->h->s == s); 1365 1366 closure_sync(&s->iodone); 1367 1368 if (!s->err) { 1369 for (i = 0; i < nr_data; i++) 1370 if (s->blocks[i]) { 1371 ob = c->open_buckets + s->blocks[i]; 1372 1373 if (ob->sectors_free) 1374 zero_out_rest_of_ec_bucket(c, s, i, ob); 1375 } 1376 } 1377 1378 if (s->err) { 1379 if (!bch2_err_matches(s->err, EROFS)) 1380 bch_err(c, "error creating stripe: error writing data buckets"); 1381 goto err; 1382 } 1383 1384 if (s->have_existing_stripe) { 1385 ec_validate_checksums(c, &s->existing_stripe); 1386 1387 if (ec_do_recov(c, &s->existing_stripe)) { 1388 bch_err(c, "error creating stripe: error reading existing stripe"); 1389 goto err; 1390 } 1391 1392 for (i = 0; i < nr_data; i++) 1393 if (stripe_blockcount_get(&bkey_i_to_stripe(&s->existing_stripe.key)->v, i)) 1394 swap(s->new_stripe.data[i], 1395 s->existing_stripe.data[i]); 1396 1397 ec_stripe_buf_exit(&s->existing_stripe); 1398 } 1399 1400 BUG_ON(!s->allocated); 1401 BUG_ON(!s->idx); 1402 1403 ec_generate_ec(&s->new_stripe); 1404 1405 ec_generate_checksums(&s->new_stripe); 1406 1407 /* write p/q: */ 1408 for (i = nr_data; i < v->nr_blocks; i++) 1409 ec_block_io(c, &s->new_stripe, REQ_OP_WRITE, i, &s->iodone); 1410 closure_sync(&s->iodone); 1411 1412 if (ec_nr_failed(&s->new_stripe)) { 1413 bch_err(c, "error creating stripe: error writing redundancy buckets"); 1414 goto err; 1415 } 1416 1417 ret = bch2_trans_do(c, &s->res, NULL, 1418 BCH_TRANS_COMMIT_no_check_rw| 1419 BCH_TRANS_COMMIT_no_enospc, 1420 ec_stripe_key_update(trans, 1421 bkey_i_to_stripe(&s->new_stripe.key), 1422 !s->have_existing_stripe)); 1423 bch_err_msg(c, ret, "creating stripe key"); 1424 if (ret) { 1425 goto err; 1426 } 1427 1428 ret = ec_stripe_update_extents(c, &s->new_stripe); 1429 bch_err_msg(c, ret, "error updating extents"); 1430 if (ret) 1431 goto err; 1432 err: 1433 bch2_disk_reservation_put(c, &s->res); 1434 1435 for (i = 0; i < v->nr_blocks; i++) 1436 if (s->blocks[i]) { 1437 ob = c->open_buckets + s->blocks[i]; 1438 1439 if (i < nr_data) { 1440 ob->ec = NULL; 1441 __bch2_open_bucket_put(c, ob); 1442 } else { 1443 bch2_open_bucket_put(c, ob); 1444 } 1445 } 1446 1447 mutex_lock(&c->ec_stripe_new_lock); 1448 list_del(&s->list); 1449 mutex_unlock(&c->ec_stripe_new_lock); 1450 wake_up(&c->ec_stripe_new_wait); 1451 1452 ec_stripe_buf_exit(&s->existing_stripe); 1453 ec_stripe_buf_exit(&s->new_stripe); 1454 closure_debug_destroy(&s->iodone); 1455 1456 ec_stripe_new_put(c, s, STRIPE_REF_stripe); 1457 } 1458 1459 static struct ec_stripe_new *get_pending_stripe(struct bch_fs *c) 1460 { 1461 struct ec_stripe_new *s; 1462 1463 mutex_lock(&c->ec_stripe_new_lock); 1464 list_for_each_entry(s, &c->ec_stripe_new_list, list) 1465 if (!atomic_read(&s->ref[STRIPE_REF_io])) 1466 goto out; 1467 s = NULL; 1468 out: 1469 mutex_unlock(&c->ec_stripe_new_lock); 1470 1471 return s; 1472 } 1473 1474 static void ec_stripe_create_work(struct work_struct *work) 1475 { 1476 struct bch_fs *c = container_of(work, 1477 struct bch_fs, ec_stripe_create_work); 1478 struct ec_stripe_new *s; 1479 1480 while ((s = get_pending_stripe(c))) 1481 ec_stripe_create(s); 1482 1483 bch2_write_ref_put(c, BCH_WRITE_REF_stripe_create); 1484 } 1485 1486 void bch2_ec_do_stripe_creates(struct bch_fs *c) 1487 { 1488 bch2_write_ref_get(c, BCH_WRITE_REF_stripe_create); 1489 1490 if (!queue_work(system_long_wq, &c->ec_stripe_create_work)) 1491 bch2_write_ref_put(c, BCH_WRITE_REF_stripe_create); 1492 } 1493 1494 static void ec_stripe_set_pending(struct bch_fs *c, struct ec_stripe_head *h) 1495 { 1496 struct ec_stripe_new *s = h->s; 1497 1498 BUG_ON(!s->allocated && !s->err); 1499 1500 h->s = NULL; 1501 s->pending = true; 1502 1503 mutex_lock(&c->ec_stripe_new_lock); 1504 list_add(&s->list, &c->ec_stripe_new_list); 1505 mutex_unlock(&c->ec_stripe_new_lock); 1506 1507 ec_stripe_new_put(c, s, STRIPE_REF_io); 1508 } 1509 1510 void bch2_ec_bucket_cancel(struct bch_fs *c, struct open_bucket *ob) 1511 { 1512 struct ec_stripe_new *s = ob->ec; 1513 1514 s->err = -EIO; 1515 } 1516 1517 void *bch2_writepoint_ec_buf(struct bch_fs *c, struct write_point *wp) 1518 { 1519 struct open_bucket *ob = ec_open_bucket(c, &wp->ptrs); 1520 struct bch_dev *ca; 1521 unsigned offset; 1522 1523 if (!ob) 1524 return NULL; 1525 1526 BUG_ON(!ob->ec->new_stripe.data[ob->ec_idx]); 1527 1528 ca = bch_dev_bkey_exists(c, ob->dev); 1529 offset = ca->mi.bucket_size - ob->sectors_free; 1530 1531 return ob->ec->new_stripe.data[ob->ec_idx] + (offset << 9); 1532 } 1533 1534 static int unsigned_cmp(const void *_l, const void *_r) 1535 { 1536 unsigned l = *((const unsigned *) _l); 1537 unsigned r = *((const unsigned *) _r); 1538 1539 return cmp_int(l, r); 1540 } 1541 1542 /* pick most common bucket size: */ 1543 static unsigned pick_blocksize(struct bch_fs *c, 1544 struct bch_devs_mask *devs) 1545 { 1546 unsigned nr = 0, sizes[BCH_SB_MEMBERS_MAX]; 1547 struct { 1548 unsigned nr, size; 1549 } cur = { 0, 0 }, best = { 0, 0 }; 1550 1551 for_each_member_device_rcu(c, ca, devs) 1552 sizes[nr++] = ca->mi.bucket_size; 1553 1554 sort(sizes, nr, sizeof(unsigned), unsigned_cmp, NULL); 1555 1556 for (unsigned i = 0; i < nr; i++) { 1557 if (sizes[i] != cur.size) { 1558 if (cur.nr > best.nr) 1559 best = cur; 1560 1561 cur.nr = 0; 1562 cur.size = sizes[i]; 1563 } 1564 1565 cur.nr++; 1566 } 1567 1568 if (cur.nr > best.nr) 1569 best = cur; 1570 1571 return best.size; 1572 } 1573 1574 static bool may_create_new_stripe(struct bch_fs *c) 1575 { 1576 return false; 1577 } 1578 1579 static void ec_stripe_key_init(struct bch_fs *c, 1580 struct bkey_i *k, 1581 unsigned nr_data, 1582 unsigned nr_parity, 1583 unsigned stripe_size) 1584 { 1585 struct bkey_i_stripe *s = bkey_stripe_init(k); 1586 unsigned u64s; 1587 1588 s->v.sectors = cpu_to_le16(stripe_size); 1589 s->v.algorithm = 0; 1590 s->v.nr_blocks = nr_data + nr_parity; 1591 s->v.nr_redundant = nr_parity; 1592 s->v.csum_granularity_bits = ilog2(c->opts.encoded_extent_max >> 9); 1593 s->v.csum_type = BCH_CSUM_crc32c; 1594 s->v.pad = 0; 1595 1596 while ((u64s = stripe_val_u64s(&s->v)) > BKEY_VAL_U64s_MAX) { 1597 BUG_ON(1 << s->v.csum_granularity_bits >= 1598 le16_to_cpu(s->v.sectors) || 1599 s->v.csum_granularity_bits == U8_MAX); 1600 s->v.csum_granularity_bits++; 1601 } 1602 1603 set_bkey_val_u64s(&s->k, u64s); 1604 } 1605 1606 static int ec_new_stripe_alloc(struct bch_fs *c, struct ec_stripe_head *h) 1607 { 1608 struct ec_stripe_new *s; 1609 1610 lockdep_assert_held(&h->lock); 1611 1612 s = kzalloc(sizeof(*s), GFP_KERNEL); 1613 if (!s) 1614 return -BCH_ERR_ENOMEM_ec_new_stripe_alloc; 1615 1616 mutex_init(&s->lock); 1617 closure_init(&s->iodone, NULL); 1618 atomic_set(&s->ref[STRIPE_REF_stripe], 1); 1619 atomic_set(&s->ref[STRIPE_REF_io], 1); 1620 s->c = c; 1621 s->h = h; 1622 s->nr_data = min_t(unsigned, h->nr_active_devs, 1623 BCH_BKEY_PTRS_MAX) - h->redundancy; 1624 s->nr_parity = h->redundancy; 1625 1626 ec_stripe_key_init(c, &s->new_stripe.key, 1627 s->nr_data, s->nr_parity, h->blocksize); 1628 1629 h->s = s; 1630 return 0; 1631 } 1632 1633 static struct ec_stripe_head * 1634 ec_new_stripe_head_alloc(struct bch_fs *c, unsigned target, 1635 unsigned algo, unsigned redundancy, 1636 enum bch_watermark watermark) 1637 { 1638 struct ec_stripe_head *h; 1639 1640 h = kzalloc(sizeof(*h), GFP_KERNEL); 1641 if (!h) 1642 return NULL; 1643 1644 mutex_init(&h->lock); 1645 BUG_ON(!mutex_trylock(&h->lock)); 1646 1647 h->target = target; 1648 h->algo = algo; 1649 h->redundancy = redundancy; 1650 h->watermark = watermark; 1651 1652 rcu_read_lock(); 1653 h->devs = target_rw_devs(c, BCH_DATA_user, target); 1654 1655 for_each_member_device_rcu(c, ca, &h->devs) 1656 if (!ca->mi.durability) 1657 __clear_bit(ca->dev_idx, h->devs.d); 1658 1659 h->blocksize = pick_blocksize(c, &h->devs); 1660 1661 for_each_member_device_rcu(c, ca, &h->devs) 1662 if (ca->mi.bucket_size == h->blocksize) 1663 h->nr_active_devs++; 1664 1665 rcu_read_unlock(); 1666 1667 /* 1668 * If we only have redundancy + 1 devices, we're better off with just 1669 * replication: 1670 */ 1671 if (h->nr_active_devs < h->redundancy + 2) 1672 bch_err(c, "insufficient devices available to create stripe (have %u, need %u) - mismatched bucket sizes?", 1673 h->nr_active_devs, h->redundancy + 2); 1674 1675 list_add(&h->list, &c->ec_stripe_head_list); 1676 return h; 1677 } 1678 1679 void bch2_ec_stripe_head_put(struct bch_fs *c, struct ec_stripe_head *h) 1680 { 1681 if (h->s && 1682 h->s->allocated && 1683 bitmap_weight(h->s->blocks_allocated, 1684 h->s->nr_data) == h->s->nr_data) 1685 ec_stripe_set_pending(c, h); 1686 1687 mutex_unlock(&h->lock); 1688 } 1689 1690 static struct ec_stripe_head * 1691 __bch2_ec_stripe_head_get(struct btree_trans *trans, 1692 unsigned target, 1693 unsigned algo, 1694 unsigned redundancy, 1695 enum bch_watermark watermark) 1696 { 1697 struct bch_fs *c = trans->c; 1698 struct ec_stripe_head *h; 1699 int ret; 1700 1701 if (!redundancy) 1702 return NULL; 1703 1704 ret = bch2_trans_mutex_lock(trans, &c->ec_stripe_head_lock); 1705 if (ret) 1706 return ERR_PTR(ret); 1707 1708 if (test_bit(BCH_FS_going_ro, &c->flags)) { 1709 h = ERR_PTR(-BCH_ERR_erofs_no_writes); 1710 goto found; 1711 } 1712 1713 list_for_each_entry(h, &c->ec_stripe_head_list, list) 1714 if (h->target == target && 1715 h->algo == algo && 1716 h->redundancy == redundancy && 1717 h->watermark == watermark) { 1718 ret = bch2_trans_mutex_lock(trans, &h->lock); 1719 if (ret) 1720 h = ERR_PTR(ret); 1721 goto found; 1722 } 1723 1724 h = ec_new_stripe_head_alloc(c, target, algo, redundancy, watermark); 1725 found: 1726 if (!IS_ERR_OR_NULL(h) && 1727 h->nr_active_devs < h->redundancy + 2) { 1728 mutex_unlock(&h->lock); 1729 h = NULL; 1730 } 1731 mutex_unlock(&c->ec_stripe_head_lock); 1732 return h; 1733 } 1734 1735 static int new_stripe_alloc_buckets(struct btree_trans *trans, struct ec_stripe_head *h, 1736 enum bch_watermark watermark, struct closure *cl) 1737 { 1738 struct bch_fs *c = trans->c; 1739 struct bch_devs_mask devs = h->devs; 1740 struct open_bucket *ob; 1741 struct open_buckets buckets; 1742 struct bch_stripe *v = &bkey_i_to_stripe(&h->s->new_stripe.key)->v; 1743 unsigned i, j, nr_have_parity = 0, nr_have_data = 0; 1744 bool have_cache = true; 1745 int ret = 0; 1746 1747 BUG_ON(v->nr_blocks != h->s->nr_data + h->s->nr_parity); 1748 BUG_ON(v->nr_redundant != h->s->nr_parity); 1749 1750 for_each_set_bit(i, h->s->blocks_gotten, v->nr_blocks) { 1751 __clear_bit(v->ptrs[i].dev, devs.d); 1752 if (i < h->s->nr_data) 1753 nr_have_data++; 1754 else 1755 nr_have_parity++; 1756 } 1757 1758 BUG_ON(nr_have_data > h->s->nr_data); 1759 BUG_ON(nr_have_parity > h->s->nr_parity); 1760 1761 buckets.nr = 0; 1762 if (nr_have_parity < h->s->nr_parity) { 1763 ret = bch2_bucket_alloc_set_trans(trans, &buckets, 1764 &h->parity_stripe, 1765 &devs, 1766 h->s->nr_parity, 1767 &nr_have_parity, 1768 &have_cache, 0, 1769 BCH_DATA_parity, 1770 watermark, 1771 cl); 1772 1773 open_bucket_for_each(c, &buckets, ob, i) { 1774 j = find_next_zero_bit(h->s->blocks_gotten, 1775 h->s->nr_data + h->s->nr_parity, 1776 h->s->nr_data); 1777 BUG_ON(j >= h->s->nr_data + h->s->nr_parity); 1778 1779 h->s->blocks[j] = buckets.v[i]; 1780 v->ptrs[j] = bch2_ob_ptr(c, ob); 1781 __set_bit(j, h->s->blocks_gotten); 1782 } 1783 1784 if (ret) 1785 return ret; 1786 } 1787 1788 buckets.nr = 0; 1789 if (nr_have_data < h->s->nr_data) { 1790 ret = bch2_bucket_alloc_set_trans(trans, &buckets, 1791 &h->block_stripe, 1792 &devs, 1793 h->s->nr_data, 1794 &nr_have_data, 1795 &have_cache, 0, 1796 BCH_DATA_user, 1797 watermark, 1798 cl); 1799 1800 open_bucket_for_each(c, &buckets, ob, i) { 1801 j = find_next_zero_bit(h->s->blocks_gotten, 1802 h->s->nr_data, 0); 1803 BUG_ON(j >= h->s->nr_data); 1804 1805 h->s->blocks[j] = buckets.v[i]; 1806 v->ptrs[j] = bch2_ob_ptr(c, ob); 1807 __set_bit(j, h->s->blocks_gotten); 1808 } 1809 1810 if (ret) 1811 return ret; 1812 } 1813 1814 return 0; 1815 } 1816 1817 /* XXX: doesn't obey target: */ 1818 static s64 get_existing_stripe(struct bch_fs *c, 1819 struct ec_stripe_head *head) 1820 { 1821 ec_stripes_heap *h = &c->ec_stripes_heap; 1822 struct stripe *m; 1823 size_t heap_idx; 1824 u64 stripe_idx; 1825 s64 ret = -1; 1826 1827 if (may_create_new_stripe(c)) 1828 return -1; 1829 1830 mutex_lock(&c->ec_stripes_heap_lock); 1831 for (heap_idx = 0; heap_idx < h->used; heap_idx++) { 1832 /* No blocks worth reusing, stripe will just be deleted: */ 1833 if (!h->data[heap_idx].blocks_nonempty) 1834 continue; 1835 1836 stripe_idx = h->data[heap_idx].idx; 1837 1838 m = genradix_ptr(&c->stripes, stripe_idx); 1839 1840 if (m->algorithm == head->algo && 1841 m->nr_redundant == head->redundancy && 1842 m->sectors == head->blocksize && 1843 m->blocks_nonempty < m->nr_blocks - m->nr_redundant && 1844 bch2_try_open_stripe(c, head->s, stripe_idx)) { 1845 ret = stripe_idx; 1846 break; 1847 } 1848 } 1849 mutex_unlock(&c->ec_stripes_heap_lock); 1850 return ret; 1851 } 1852 1853 static int __bch2_ec_stripe_head_reuse(struct btree_trans *trans, struct ec_stripe_head *h) 1854 { 1855 struct bch_fs *c = trans->c; 1856 struct bch_stripe *new_v = &bkey_i_to_stripe(&h->s->new_stripe.key)->v; 1857 struct bch_stripe *existing_v; 1858 unsigned i; 1859 s64 idx; 1860 int ret; 1861 1862 /* 1863 * If we can't allocate a new stripe, and there's no stripes with empty 1864 * blocks for us to reuse, that means we have to wait on copygc: 1865 */ 1866 idx = get_existing_stripe(c, h); 1867 if (idx < 0) 1868 return -BCH_ERR_stripe_alloc_blocked; 1869 1870 ret = get_stripe_key_trans(trans, idx, &h->s->existing_stripe); 1871 bch2_fs_fatal_err_on(ret && !bch2_err_matches(ret, BCH_ERR_transaction_restart), c, 1872 "reading stripe key: %s", bch2_err_str(ret)); 1873 if (ret) { 1874 bch2_stripe_close(c, h->s); 1875 return ret; 1876 } 1877 1878 existing_v = &bkey_i_to_stripe(&h->s->existing_stripe.key)->v; 1879 1880 BUG_ON(existing_v->nr_redundant != h->s->nr_parity); 1881 h->s->nr_data = existing_v->nr_blocks - 1882 existing_v->nr_redundant; 1883 1884 ret = ec_stripe_buf_init(&h->s->existing_stripe, 0, h->blocksize); 1885 if (ret) { 1886 bch2_stripe_close(c, h->s); 1887 return ret; 1888 } 1889 1890 BUG_ON(h->s->existing_stripe.size != h->blocksize); 1891 BUG_ON(h->s->existing_stripe.size != le16_to_cpu(existing_v->sectors)); 1892 1893 /* 1894 * Free buckets we initially allocated - they might conflict with 1895 * blocks from the stripe we're reusing: 1896 */ 1897 for_each_set_bit(i, h->s->blocks_gotten, new_v->nr_blocks) { 1898 bch2_open_bucket_put(c, c->open_buckets + h->s->blocks[i]); 1899 h->s->blocks[i] = 0; 1900 } 1901 memset(h->s->blocks_gotten, 0, sizeof(h->s->blocks_gotten)); 1902 memset(h->s->blocks_allocated, 0, sizeof(h->s->blocks_allocated)); 1903 1904 for (i = 0; i < existing_v->nr_blocks; i++) { 1905 if (stripe_blockcount_get(existing_v, i)) { 1906 __set_bit(i, h->s->blocks_gotten); 1907 __set_bit(i, h->s->blocks_allocated); 1908 } 1909 1910 ec_block_io(c, &h->s->existing_stripe, READ, i, &h->s->iodone); 1911 } 1912 1913 bkey_copy(&h->s->new_stripe.key, &h->s->existing_stripe.key); 1914 h->s->have_existing_stripe = true; 1915 1916 return 0; 1917 } 1918 1919 static int __bch2_ec_stripe_head_reserve(struct btree_trans *trans, struct ec_stripe_head *h) 1920 { 1921 struct bch_fs *c = trans->c; 1922 struct btree_iter iter; 1923 struct bkey_s_c k; 1924 struct bpos min_pos = POS(0, 1); 1925 struct bpos start_pos = bpos_max(min_pos, POS(0, c->ec_stripe_hint)); 1926 int ret; 1927 1928 if (!h->s->res.sectors) { 1929 ret = bch2_disk_reservation_get(c, &h->s->res, 1930 h->blocksize, 1931 h->s->nr_parity, 1932 BCH_DISK_RESERVATION_NOFAIL); 1933 if (ret) 1934 return ret; 1935 } 1936 1937 for_each_btree_key_norestart(trans, iter, BTREE_ID_stripes, start_pos, 1938 BTREE_ITER_SLOTS|BTREE_ITER_INTENT, k, ret) { 1939 if (bkey_gt(k.k->p, POS(0, U32_MAX))) { 1940 if (start_pos.offset) { 1941 start_pos = min_pos; 1942 bch2_btree_iter_set_pos(&iter, start_pos); 1943 continue; 1944 } 1945 1946 ret = -BCH_ERR_ENOSPC_stripe_create; 1947 break; 1948 } 1949 1950 if (bkey_deleted(k.k) && 1951 bch2_try_open_stripe(c, h->s, k.k->p.offset)) 1952 break; 1953 } 1954 1955 c->ec_stripe_hint = iter.pos.offset; 1956 1957 if (ret) 1958 goto err; 1959 1960 ret = ec_stripe_mem_alloc(trans, &iter); 1961 if (ret) { 1962 bch2_stripe_close(c, h->s); 1963 goto err; 1964 } 1965 1966 h->s->new_stripe.key.k.p = iter.pos; 1967 out: 1968 bch2_trans_iter_exit(trans, &iter); 1969 return ret; 1970 err: 1971 bch2_disk_reservation_put(c, &h->s->res); 1972 goto out; 1973 } 1974 1975 struct ec_stripe_head *bch2_ec_stripe_head_get(struct btree_trans *trans, 1976 unsigned target, 1977 unsigned algo, 1978 unsigned redundancy, 1979 enum bch_watermark watermark, 1980 struct closure *cl) 1981 { 1982 struct bch_fs *c = trans->c; 1983 struct ec_stripe_head *h; 1984 bool waiting = false; 1985 int ret; 1986 1987 h = __bch2_ec_stripe_head_get(trans, target, algo, redundancy, watermark); 1988 if (IS_ERR_OR_NULL(h)) 1989 return h; 1990 1991 if (!h->s) { 1992 ret = ec_new_stripe_alloc(c, h); 1993 if (ret) { 1994 bch_err(c, "failed to allocate new stripe"); 1995 goto err; 1996 } 1997 } 1998 1999 if (h->s->allocated) 2000 goto allocated; 2001 2002 if (h->s->have_existing_stripe) 2003 goto alloc_existing; 2004 2005 /* First, try to allocate a full stripe: */ 2006 ret = new_stripe_alloc_buckets(trans, h, BCH_WATERMARK_stripe, NULL) ?: 2007 __bch2_ec_stripe_head_reserve(trans, h); 2008 if (!ret) 2009 goto allocate_buf; 2010 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) || 2011 bch2_err_matches(ret, ENOMEM)) 2012 goto err; 2013 2014 /* 2015 * Not enough buckets available for a full stripe: we must reuse an 2016 * existing stripe: 2017 */ 2018 while (1) { 2019 ret = __bch2_ec_stripe_head_reuse(trans, h); 2020 if (!ret) 2021 break; 2022 if (waiting || !cl || ret != -BCH_ERR_stripe_alloc_blocked) 2023 goto err; 2024 2025 if (watermark == BCH_WATERMARK_copygc) { 2026 ret = new_stripe_alloc_buckets(trans, h, watermark, NULL) ?: 2027 __bch2_ec_stripe_head_reserve(trans, h); 2028 if (ret) 2029 goto err; 2030 goto allocate_buf; 2031 } 2032 2033 /* XXX freelist_wait? */ 2034 closure_wait(&c->freelist_wait, cl); 2035 waiting = true; 2036 } 2037 2038 if (waiting) 2039 closure_wake_up(&c->freelist_wait); 2040 alloc_existing: 2041 /* 2042 * Retry allocating buckets, with the watermark for this 2043 * particular write: 2044 */ 2045 ret = new_stripe_alloc_buckets(trans, h, watermark, cl); 2046 if (ret) 2047 goto err; 2048 2049 allocate_buf: 2050 ret = ec_stripe_buf_init(&h->s->new_stripe, 0, h->blocksize); 2051 if (ret) 2052 goto err; 2053 2054 h->s->allocated = true; 2055 allocated: 2056 BUG_ON(!h->s->idx); 2057 BUG_ON(!h->s->new_stripe.data[0]); 2058 BUG_ON(trans->restarted); 2059 return h; 2060 err: 2061 bch2_ec_stripe_head_put(c, h); 2062 return ERR_PTR(ret); 2063 } 2064 2065 static void __bch2_ec_stop(struct bch_fs *c, struct bch_dev *ca) 2066 { 2067 struct ec_stripe_head *h; 2068 struct open_bucket *ob; 2069 unsigned i; 2070 2071 mutex_lock(&c->ec_stripe_head_lock); 2072 list_for_each_entry(h, &c->ec_stripe_head_list, list) { 2073 mutex_lock(&h->lock); 2074 if (!h->s) 2075 goto unlock; 2076 2077 if (!ca) 2078 goto found; 2079 2080 for (i = 0; i < bkey_i_to_stripe(&h->s->new_stripe.key)->v.nr_blocks; i++) { 2081 if (!h->s->blocks[i]) 2082 continue; 2083 2084 ob = c->open_buckets + h->s->blocks[i]; 2085 if (ob->dev == ca->dev_idx) 2086 goto found; 2087 } 2088 goto unlock; 2089 found: 2090 h->s->err = -BCH_ERR_erofs_no_writes; 2091 ec_stripe_set_pending(c, h); 2092 unlock: 2093 mutex_unlock(&h->lock); 2094 } 2095 mutex_unlock(&c->ec_stripe_head_lock); 2096 } 2097 2098 void bch2_ec_stop_dev(struct bch_fs *c, struct bch_dev *ca) 2099 { 2100 __bch2_ec_stop(c, ca); 2101 } 2102 2103 void bch2_fs_ec_stop(struct bch_fs *c) 2104 { 2105 __bch2_ec_stop(c, NULL); 2106 } 2107 2108 static bool bch2_fs_ec_flush_done(struct bch_fs *c) 2109 { 2110 bool ret; 2111 2112 mutex_lock(&c->ec_stripe_new_lock); 2113 ret = list_empty(&c->ec_stripe_new_list); 2114 mutex_unlock(&c->ec_stripe_new_lock); 2115 2116 return ret; 2117 } 2118 2119 void bch2_fs_ec_flush(struct bch_fs *c) 2120 { 2121 wait_event(c->ec_stripe_new_wait, bch2_fs_ec_flush_done(c)); 2122 } 2123 2124 int bch2_stripes_read(struct bch_fs *c) 2125 { 2126 int ret = bch2_trans_run(c, 2127 for_each_btree_key(trans, iter, BTREE_ID_stripes, POS_MIN, 2128 BTREE_ITER_PREFETCH, k, ({ 2129 if (k.k->type != KEY_TYPE_stripe) 2130 continue; 2131 2132 ret = __ec_stripe_mem_alloc(c, k.k->p.offset, GFP_KERNEL); 2133 if (ret) 2134 break; 2135 2136 const struct bch_stripe *s = bkey_s_c_to_stripe(k).v; 2137 2138 struct stripe *m = genradix_ptr(&c->stripes, k.k->p.offset); 2139 m->sectors = le16_to_cpu(s->sectors); 2140 m->algorithm = s->algorithm; 2141 m->nr_blocks = s->nr_blocks; 2142 m->nr_redundant = s->nr_redundant; 2143 m->blocks_nonempty = 0; 2144 2145 for (unsigned i = 0; i < s->nr_blocks; i++) 2146 m->blocks_nonempty += !!stripe_blockcount_get(s, i); 2147 2148 bch2_stripes_heap_insert(c, m, k.k->p.offset); 2149 0; 2150 }))); 2151 bch_err_fn(c, ret); 2152 return ret; 2153 } 2154 2155 void bch2_stripes_heap_to_text(struct printbuf *out, struct bch_fs *c) 2156 { 2157 ec_stripes_heap *h = &c->ec_stripes_heap; 2158 struct stripe *m; 2159 size_t i; 2160 2161 mutex_lock(&c->ec_stripes_heap_lock); 2162 for (i = 0; i < min_t(size_t, h->used, 50); i++) { 2163 m = genradix_ptr(&c->stripes, h->data[i].idx); 2164 2165 prt_printf(out, "%zu %u/%u+%u", h->data[i].idx, 2166 h->data[i].blocks_nonempty, 2167 m->nr_blocks - m->nr_redundant, 2168 m->nr_redundant); 2169 if (bch2_stripe_is_open(c, h->data[i].idx)) 2170 prt_str(out, " open"); 2171 prt_newline(out); 2172 } 2173 mutex_unlock(&c->ec_stripes_heap_lock); 2174 } 2175 2176 void bch2_new_stripes_to_text(struct printbuf *out, struct bch_fs *c) 2177 { 2178 struct ec_stripe_head *h; 2179 struct ec_stripe_new *s; 2180 2181 mutex_lock(&c->ec_stripe_head_lock); 2182 list_for_each_entry(h, &c->ec_stripe_head_list, list) { 2183 prt_printf(out, "target %u algo %u redundancy %u %s:\n", 2184 h->target, h->algo, h->redundancy, 2185 bch2_watermarks[h->watermark]); 2186 2187 if (h->s) 2188 prt_printf(out, "\tidx %llu blocks %u+%u allocated %u\n", 2189 h->s->idx, h->s->nr_data, h->s->nr_parity, 2190 bitmap_weight(h->s->blocks_allocated, 2191 h->s->nr_data)); 2192 } 2193 mutex_unlock(&c->ec_stripe_head_lock); 2194 2195 prt_printf(out, "in flight:\n"); 2196 2197 mutex_lock(&c->ec_stripe_new_lock); 2198 list_for_each_entry(s, &c->ec_stripe_new_list, list) { 2199 prt_printf(out, "\tidx %llu blocks %u+%u ref %u %u %s\n", 2200 s->idx, s->nr_data, s->nr_parity, 2201 atomic_read(&s->ref[STRIPE_REF_io]), 2202 atomic_read(&s->ref[STRIPE_REF_stripe]), 2203 bch2_watermarks[s->h->watermark]); 2204 } 2205 mutex_unlock(&c->ec_stripe_new_lock); 2206 } 2207 2208 void bch2_fs_ec_exit(struct bch_fs *c) 2209 { 2210 struct ec_stripe_head *h; 2211 unsigned i; 2212 2213 while (1) { 2214 mutex_lock(&c->ec_stripe_head_lock); 2215 h = list_first_entry_or_null(&c->ec_stripe_head_list, 2216 struct ec_stripe_head, list); 2217 if (h) 2218 list_del(&h->list); 2219 mutex_unlock(&c->ec_stripe_head_lock); 2220 if (!h) 2221 break; 2222 2223 if (h->s) { 2224 for (i = 0; i < bkey_i_to_stripe(&h->s->new_stripe.key)->v.nr_blocks; i++) 2225 BUG_ON(h->s->blocks[i]); 2226 2227 kfree(h->s); 2228 } 2229 kfree(h); 2230 } 2231 2232 BUG_ON(!list_empty(&c->ec_stripe_new_list)); 2233 2234 free_heap(&c->ec_stripes_heap); 2235 genradix_free(&c->stripes); 2236 bioset_exit(&c->ec_bioset); 2237 } 2238 2239 void bch2_fs_ec_init_early(struct bch_fs *c) 2240 { 2241 spin_lock_init(&c->ec_stripes_new_lock); 2242 mutex_init(&c->ec_stripes_heap_lock); 2243 2244 INIT_LIST_HEAD(&c->ec_stripe_head_list); 2245 mutex_init(&c->ec_stripe_head_lock); 2246 2247 INIT_LIST_HEAD(&c->ec_stripe_new_list); 2248 mutex_init(&c->ec_stripe_new_lock); 2249 init_waitqueue_head(&c->ec_stripe_new_wait); 2250 2251 INIT_WORK(&c->ec_stripe_create_work, ec_stripe_create_work); 2252 INIT_WORK(&c->ec_stripe_delete_work, ec_stripe_delete_work); 2253 } 2254 2255 int bch2_fs_ec_init(struct bch_fs *c) 2256 { 2257 return bioset_init(&c->ec_bioset, 1, offsetof(struct ec_bio, bio), 2258 BIOSET_NEED_BVECS); 2259 } 2260