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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 /* erasure coding */
   4
   5 #include "bcachefs.h"
   6 #include "alloc_foreground.h"
   7 #include "bkey_buf.h"
   8 #include "bset.h"
   9 #include "btree_gc.h"
  10 #include "btree_update.h"
  11 #include "buckets.h"
  12 #include "disk_groups.h"
  13 #include "ec.h"
  14 #include "error.h"
  15 #include "io.h"
  16 #include "keylist.h"
  17 #include "recovery.h"
  18 #include "super-io.h"
  19 #include "util.h"
  20
  21 #include <linux/sort.h>
  22
  23 #ifdef __KERNEL__
  24
  25 #include <linux/raid/pq.h>
  26 #include <linux/raid/xor.h>
  27
  28 static void raid5_recov(unsigned disks, unsigned failed_idx,
  29                         size_t size, void **data)
  30 {
  31         unsigned i = 2, nr;
  32
  33         BUG_ON(failed_idx >= disks);
  34
  35         swap(data[0], data[failed_idx]);
  36         memcpy(data[0], data[1], size);
  37
  38         while (i < disks) {
  39                 nr = min_t(unsigned, disks - i, MAX_XOR_BLOCKS);
  40                 xor_blocks(nr, size, data[0], data + i);
  41                 i += nr;
  42         }
  43
  44         swap(data[0], data[failed_idx]);
  45 }
  46
  47 static void raid_gen(int nd, int np, size_t size, void **v)
  48 {
  49         if (np >= 1)
  50                 raid5_recov(nd + np, nd, size, v);
  51         if (np >= 2)
  52                 raid6_call.gen_syndrome(nd + np, size, v);
  53         BUG_ON(np > 2);
  54 }
  55
  56 static void raid_rec(int nr, int *ir, int nd, int np, size_t size, void **v)
  57 {
  58         switch (nr) {
  59         case 0:
  60                 break;
  61         case 1:
  62                 if (ir[0] < nd + 1)
  63                         raid5_recov(nd + 1, ir[0], size, v);
  64                 else
  65                         raid6_call.gen_syndrome(nd + np, size, v);
  66                 break;
  67         case 2:
  68                 if (ir[1] < nd) {
  69                         /* data+data failure. */
  70                         raid6_2data_recov(nd + np, size, ir[0], ir[1], v);
  71                 } else if (ir[0] < nd) {
  72                         /* data + p/q failure */
  73
  74                         if (ir[1] == nd) /* data + p failure */
  75                                 raid6_datap_recov(nd + np, size, ir[0], v);
  76                         else { /* data + q failure */
  77                                 raid5_recov(nd + 1, ir[0], size, v);
  78                                 raid6_call.gen_syndrome(nd + np, size, v);
  79                         }
  80                 } else {
  81                         raid_gen(nd, np, size, v);
  82                 }
  83                 break;
  84         default:
  85                 BUG();
  86         }
  87 }
  88
  89 #else
  90
  91 #include <raid/raid.h>
  92
  93 #endif
  94
  95 struct ec_bio {
  96         struct bch_dev          *ca;
  97         struct ec_stripe_buf    *buf;
  98         size_t                  idx;
  99         struct bio              bio;
 100 };
 101
 102 /* Stripes btree keys: */
 103
 104 const char *bch2_stripe_invalid(const struct bch_fs *c, struct bkey_s_c k)
 105 {
 106         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 107
 108         if (k.k->p.inode)
 109                 return "invalid stripe key";
 110
 111         if (bkey_val_bytes(k.k) < sizeof(*s))
 112                 return "incorrect value size";
 113
 114         if (bkey_val_bytes(k.k) < sizeof(*s) ||
 115             bkey_val_u64s(k.k) < stripe_val_u64s(s))
 116                 return "incorrect value size";
 117
 118         return bch2_bkey_ptrs_invalid(c, k);
 119 }
 120
 121 void bch2_stripe_to_text(struct printbuf *out, struct bch_fs *c,
 122                          struct bkey_s_c k)
 123 {
 124         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 125         unsigned i;
 126
 127         pr_buf(out, "algo %u sectors %u blocks %u:%u csum %u gran %u",
 128                s->algorithm,
 129                le16_to_cpu(s->sectors),
 130                s->nr_blocks - s->nr_redundant,
 131                s->nr_redundant,
 132                s->csum_type,
 133                1U << s->csum_granularity_bits);
 134
 135         for (i = 0; i < s->nr_blocks; i++)
 136                 pr_buf(out, " %u:%llu:%u", s->ptrs[i].dev,
 137                        (u64) s->ptrs[i].offset,
 138                        stripe_blockcount_get(s, i));
 139 }
 140
 141 /* returns blocknr in stripe that we matched: */
 142 static int bkey_matches_stripe(struct bch_stripe *s,
 143                                struct bkey_s_c k)
 144 {
 145         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 146         const struct bch_extent_ptr *ptr;
 147         unsigned i, nr_data = s->nr_blocks - s->nr_redundant;
 148
 149         bkey_for_each_ptr(ptrs, ptr)
 150                 for (i = 0; i < nr_data; i++)
 151                         if (__bch2_ptr_matches_stripe(s, ptr, i))
 152                                 return i;
 153
 154         return -1;
 155 }
 156
 157 static bool extent_has_stripe_ptr(struct bkey_s_c k, u64 idx)
 158 {
 159         switch (k.k->type) {
 160         case KEY_TYPE_extent: {
 161                 struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
 162                 const union bch_extent_entry *entry;
 163
 164                 extent_for_each_entry(e, entry)
 165                         if (extent_entry_type(entry) ==
 166                             BCH_EXTENT_ENTRY_stripe_ptr &&
 167                             entry->stripe_ptr.idx == idx)
 168                                 return true;
 169
 170                 break;
 171         }
 172         }
 173
 174         return false;
 175 }
 176
 177 /* Stripe bufs: */
 178
 179 static void ec_stripe_buf_exit(struct ec_stripe_buf *buf)
 180 {
 181         unsigned i;
 182
 183         for (i = 0; i < buf->key.v.nr_blocks; i++) {
 184                 kvpfree(buf->data[i], buf->size << 9);
 185                 buf->data[i] = NULL;
 186         }
 187 }
 188
 189 static int ec_stripe_buf_init(struct ec_stripe_buf *buf,
 190                                unsigned offset, unsigned size)
 191 {
 192         struct bch_stripe *v = &buf->key.v;
 193         unsigned csum_granularity = 1U << v->csum_granularity_bits;
 194         unsigned end = offset + size;
 195         unsigned i;
 196
 197         BUG_ON(end > le16_to_cpu(v->sectors));
 198
 199         offset  = round_down(offset, csum_granularity);
 200         end     = min_t(unsigned, le16_to_cpu(v->sectors),
 201                         round_up(end, csum_granularity));
 202
 203         buf->offset     = offset;
 204         buf->size       = end - offset;
 205
 206         memset(buf->valid, 0xFF, sizeof(buf->valid));
 207
 208         for (i = 0; i < buf->key.v.nr_blocks; i++) {
 209                 buf->data[i] = kvpmalloc(buf->size << 9, GFP_KERNEL);
 210                 if (!buf->data[i])
 211                         goto err;
 212         }
 213
 214         return 0;
 215 err:
 216         ec_stripe_buf_exit(buf);
 217         return -ENOMEM;
 218 }
 219
 220 /* Checksumming: */
 221
 222 static struct bch_csum ec_block_checksum(struct ec_stripe_buf *buf,
 223                                          unsigned block, unsigned offset)
 224 {
 225         struct bch_stripe *v = &buf->key.v;
 226         unsigned csum_granularity = 1 << v->csum_granularity_bits;
 227         unsigned end = buf->offset + buf->size;
 228         unsigned len = min(csum_granularity, end - offset);
 229
 230         BUG_ON(offset >= end);
 231         BUG_ON(offset <  buf->offset);
 232         BUG_ON(offset & (csum_granularity - 1));
 233         BUG_ON(offset + len != le16_to_cpu(v->sectors) &&
 234                (len & (csum_granularity - 1)));
 235
 236         return bch2_checksum(NULL, v->csum_type,
 237                              null_nonce(),
 238                              buf->data[block] + ((offset - buf->offset) << 9),
 239                              len << 9);
 240 }
 241
 242 static void ec_generate_checksums(struct ec_stripe_buf *buf)
 243 {
 244         struct bch_stripe *v = &buf->key.v;
 245         unsigned i, j, csums_per_device = stripe_csums_per_device(v);
 246
 247         if (!v->csum_type)
 248                 return;
 249
 250         BUG_ON(buf->offset);
 251         BUG_ON(buf->size != le16_to_cpu(v->sectors));
 252
 253         for (i = 0; i < v->nr_blocks; i++)
 254                 for (j = 0; j < csums_per_device; j++)
 255                         stripe_csum_set(v, i, j,
 256                                 ec_block_checksum(buf, i, j << v->csum_granularity_bits));
 257 }
 258
 259 static void ec_validate_checksums(struct bch_fs *c, struct ec_stripe_buf *buf)
 260 {
 261         struct bch_stripe *v = &buf->key.v;
 262         unsigned csum_granularity = 1 << v->csum_granularity_bits;
 263         unsigned i;
 264
 265         if (!v->csum_type)
 266                 return;
 267
 268         for (i = 0; i < v->nr_blocks; i++) {
 269                 unsigned offset = buf->offset;
 270                 unsigned end = buf->offset + buf->size;
 271
 272                 if (!test_bit(i, buf->valid))
 273                         continue;
 274
 275                 while (offset < end) {
 276                         unsigned j = offset >> v->csum_granularity_bits;
 277                         unsigned len = min(csum_granularity, end - offset);
 278                         struct bch_csum want = stripe_csum_get(v, i, j);
 279                         struct bch_csum got = ec_block_checksum(buf, i, offset);
 280
 281                         if (bch2_crc_cmp(want, got)) {
 282                                 bch_err_ratelimited(c,
 283                                         "stripe checksum error at %u:%u: csum type %u, expected %llx got %llx",
 284                                         i, j, v->csum_type,
 285                                         want.lo, got.lo);
 286                                 clear_bit(i, buf->valid);
 287                                 break;
 288                         }
 289
 290                         offset += len;
 291                 }
 292         }
 293 }
 294
 295 /* Erasure coding: */
 296
 297 static void ec_generate_ec(struct ec_stripe_buf *buf)
 298 {
 299         struct bch_stripe *v = &buf->key.v;
 300         unsigned nr_data = v->nr_blocks - v->nr_redundant;
 301         unsigned bytes = le16_to_cpu(v->sectors) << 9;
 302
 303         raid_gen(nr_data, v->nr_redundant, bytes, buf->data);
 304 }
 305
 306 static unsigned ec_nr_failed(struct ec_stripe_buf *buf)
 307 {
 308         return buf->key.v.nr_blocks -
 309                 bitmap_weight(buf->valid, buf->key.v.nr_blocks);
 310 }
 311
 312 static int ec_do_recov(struct bch_fs *c, struct ec_stripe_buf *buf)
 313 {
 314         struct bch_stripe *v = &buf->key.v;
 315         unsigned i, failed[BCH_BKEY_PTRS_MAX], nr_failed = 0;
 316         unsigned nr_data = v->nr_blocks - v->nr_redundant;
 317         unsigned bytes = buf->size << 9;
 318
 319         if (ec_nr_failed(buf) > v->nr_redundant) {
 320                 bch_err_ratelimited(c,
 321                         "error doing reconstruct read: unable to read enough blocks");
 322                 return -1;
 323         }
 324
 325         for (i = 0; i < nr_data; i++)
 326                 if (!test_bit(i, buf->valid))
 327                         failed[nr_failed++] = i;
 328
 329         raid_rec(nr_failed, failed, nr_data, v->nr_redundant, bytes, buf->data);
 330         return 0;
 331 }
 332
 333 /* IO: */
 334
 335 static void ec_block_endio(struct bio *bio)
 336 {
 337         struct ec_bio *ec_bio = container_of(bio, struct ec_bio, bio);
 338         struct bch_dev *ca = ec_bio->ca;
 339         struct closure *cl = bio->bi_private;
 340
 341         if (bch2_dev_io_err_on(bio->bi_status, ca, "erasure coding %s error: %s",
 342                                bio_data_dir(bio) ? "write" : "read",
 343                                bch2_blk_status_to_str(bio->bi_status)))
 344                 clear_bit(ec_bio->idx, ec_bio->buf->valid);
 345
 346         bio_put(&ec_bio->bio);
 347         percpu_ref_put(&ca->io_ref);
 348         closure_put(cl);
 349 }
 350
 351 static void ec_block_io(struct bch_fs *c, struct ec_stripe_buf *buf,
 352                         unsigned rw, unsigned idx, struct closure *cl)
 353 {
 354         struct bch_stripe *v = &buf->key.v;
 355         unsigned offset = 0, bytes = buf->size << 9;
 356         struct bch_extent_ptr *ptr = &v->ptrs[idx];
 357         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 358         enum bch_data_type data_type = idx < buf->key.v.nr_blocks - buf->key.v.nr_redundant
 359                 ? BCH_DATA_user
 360                 : BCH_DATA_parity;
 361
 362         if (ptr_stale(ca, ptr)) {
 363                 bch_err_ratelimited(c,
 364                                     "error %s stripe: stale pointer",
 365                                     rw == READ ? "reading from" : "writing to");
 366                 clear_bit(idx, buf->valid);
 367                 return;
 368         }
 369
 370         if (!bch2_dev_get_ioref(ca, rw)) {
 371                 clear_bit(idx, buf->valid);
 372                 return;
 373         }
 374
 375         this_cpu_add(ca->io_done->sectors[rw][data_type], buf->size);
 376
 377         while (offset < bytes) {
 378                 unsigned nr_iovecs = min_t(size_t, BIO_MAX_PAGES,
 379                                            DIV_ROUND_UP(bytes, PAGE_SIZE));
 380                 unsigned b = min_t(size_t, bytes - offset,
 381                                    nr_iovecs << PAGE_SHIFT);
 382                 struct ec_bio *ec_bio;
 383
 384                 ec_bio = container_of(bio_alloc_bioset(GFP_KERNEL, nr_iovecs,
 385                                                        &c->ec_bioset),
 386                                       struct ec_bio, bio);
 387
 388                 ec_bio->ca                      = ca;
 389                 ec_bio->buf                     = buf;
 390                 ec_bio->idx                     = idx;
 391
 392                 bio_set_dev(&ec_bio->bio, ca->disk_sb.bdev);
 393                 bio_set_op_attrs(&ec_bio->bio, rw, 0);
 394
 395                 ec_bio->bio.bi_iter.bi_sector   = ptr->offset + buf->offset + (offset >> 9);
 396                 ec_bio->bio.bi_end_io           = ec_block_endio;
 397                 ec_bio->bio.bi_private          = cl;
 398
 399                 bch2_bio_map(&ec_bio->bio, buf->data[idx] + offset, b);
 400
 401                 closure_get(cl);
 402                 percpu_ref_get(&ca->io_ref);
 403
 404                 submit_bio(&ec_bio->bio);
 405
 406                 offset += b;
 407         }
 408
 409         percpu_ref_put(&ca->io_ref);
 410 }
 411
 412 static int get_stripe_key(struct bch_fs *c, u64 idx, struct ec_stripe_buf *stripe)
 413 {
 414         struct btree_trans trans;
 415         struct btree_iter *iter;
 416         struct bkey_s_c k;
 417         int ret;
 418
 419         bch2_trans_init(&trans, c, 0, 0);
 420         iter = bch2_trans_get_iter(&trans, BTREE_ID_EC, POS(0, idx), BTREE_ITER_SLOTS);
 421         k = bch2_btree_iter_peek_slot(iter);
 422         ret = bkey_err(k);
 423         if (ret)
 424                 goto err;
 425         if (k.k->type != KEY_TYPE_stripe) {
 426                 ret = -ENOENT;
 427                 goto err;
 428         }
 429         bkey_reassemble(&stripe->key.k_i, k);
 430 err:
 431         bch2_trans_exit(&trans);
 432         return ret;
 433 }
 434
 435 /* recovery read path: */
 436 int bch2_ec_read_extent(struct bch_fs *c, struct bch_read_bio *rbio)
 437 {
 438         struct ec_stripe_buf *buf;
 439         struct closure cl;
 440         struct bch_stripe *v;
 441         unsigned i, offset;
 442         int ret = 0;
 443
 444         closure_init_stack(&cl);
 445
 446         BUG_ON(!rbio->pick.has_ec);
 447
 448         buf = kzalloc(sizeof(*buf), GFP_NOIO);
 449         if (!buf)
 450                 return -ENOMEM;
 451
 452         ret = get_stripe_key(c, rbio->pick.ec.idx, buf);
 453         if (ret) {
 454                 bch_err_ratelimited(c,
 455                         "error doing reconstruct read: error %i looking up stripe", ret);
 456                 kfree(buf);
 457                 return -EIO;
 458         }
 459
 460         v = &buf->key.v;
 461
 462         if (!bch2_ptr_matches_stripe(v, rbio->pick)) {
 463                 bch_err_ratelimited(c,
 464                         "error doing reconstruct read: pointer doesn't match stripe");
 465                 ret = -EIO;
 466                 goto err;
 467         }
 468
 469         offset = rbio->bio.bi_iter.bi_sector - v->ptrs[rbio->pick.ec.block].offset;
 470         if (offset + bio_sectors(&rbio->bio) > le16_to_cpu(v->sectors)) {
 471                 bch_err_ratelimited(c,
 472                         "error doing reconstruct read: read is bigger than stripe");
 473                 ret = -EIO;
 474                 goto err;
 475         }
 476
 477         ret = ec_stripe_buf_init(buf, offset, bio_sectors(&rbio->bio));
 478         if (ret)
 479                 goto err;
 480
 481         for (i = 0; i < v->nr_blocks; i++)
 482                 ec_block_io(c, buf, REQ_OP_READ, i, &cl);
 483
 484         closure_sync(&cl);
 485
 486         if (ec_nr_failed(buf) > v->nr_redundant) {
 487                 bch_err_ratelimited(c,
 488                         "error doing reconstruct read: unable to read enough blocks");
 489                 ret = -EIO;
 490                 goto err;
 491         }
 492
 493         ec_validate_checksums(c, buf);
 494
 495         ret = ec_do_recov(c, buf);
 496         if (ret)
 497                 goto err;
 498
 499         memcpy_to_bio(&rbio->bio, rbio->bio.bi_iter,
 500                       buf->data[rbio->pick.ec.block] + ((offset - buf->offset) << 9));
 501 err:
 502         ec_stripe_buf_exit(buf);
 503         kfree(buf);
 504         return ret;
 505 }
 506
 507 /* stripe bucket accounting: */
 508
 509 static int __ec_stripe_mem_alloc(struct bch_fs *c, size_t idx, gfp_t gfp)
 510 {
 511         ec_stripes_heap n, *h = &c->ec_stripes_heap;
 512
 513         if (idx >= h->size) {
 514                 if (!init_heap(&n, max(1024UL, roundup_pow_of_two(idx + 1)), gfp))
 515                         return -ENOMEM;
 516
 517                 spin_lock(&c->ec_stripes_heap_lock);
 518                 if (n.size > h->size) {
 519                         memcpy(n.data, h->data, h->used * sizeof(h->data[0]));
 520                         n.used = h->used;
 521                         swap(*h, n);
 522                 }
 523                 spin_unlock(&c->ec_stripes_heap_lock);
 524
 525                 free_heap(&n);
 526         }
 527
 528         if (!genradix_ptr_alloc(&c->stripes[0], idx, gfp))
 529                 return -ENOMEM;
 530
 531         if (c->gc_pos.phase != GC_PHASE_NOT_RUNNING &&
 532             !genradix_ptr_alloc(&c->stripes[1], idx, gfp))
 533                 return -ENOMEM;
 534
 535         return 0;
 536 }
 537
 538 static int ec_stripe_mem_alloc(struct bch_fs *c,
 539                                struct btree_iter *iter)
 540 {
 541         size_t idx = iter->pos.offset;
 542         int ret = 0;
 543
 544         if (!__ec_stripe_mem_alloc(c, idx, GFP_NOWAIT|__GFP_NOWARN))
 545                 return ret;
 546
 547         bch2_trans_unlock(iter->trans);
 548         ret = -EINTR;
 549
 550         if (!__ec_stripe_mem_alloc(c, idx, GFP_KERNEL))
 551                 return ret;
 552
 553         return -ENOMEM;
 554 }
 555
 556 static ssize_t stripe_idx_to_delete(struct bch_fs *c)
 557 {
 558         ec_stripes_heap *h = &c->ec_stripes_heap;
 559
 560         return h->used && h->data[0].blocks_nonempty == 0
 561                 ? h->data[0].idx : -1;
 562 }
 563
 564 static inline int ec_stripes_heap_cmp(ec_stripes_heap *h,
 565                                       struct ec_stripe_heap_entry l,
 566                                       struct ec_stripe_heap_entry r)
 567 {
 568         return ((l.blocks_nonempty > r.blocks_nonempty) -
 569                 (l.blocks_nonempty < r.blocks_nonempty));
 570 }
 571
 572 static inline void ec_stripes_heap_set_backpointer(ec_stripes_heap *h,
 573                                                    size_t i)
 574 {
 575         struct bch_fs *c = container_of(h, struct bch_fs, ec_stripes_heap);
 576
 577         genradix_ptr(&c->stripes[0], h->data[i].idx)->heap_idx = i;
 578 }
 579
 580 static void heap_verify_backpointer(struct bch_fs *c, size_t idx)
 581 {
 582         ec_stripes_heap *h = &c->ec_stripes_heap;
 583         struct stripe *m = genradix_ptr(&c->stripes[0], idx);
 584
 585         BUG_ON(!m->alive);
 586         BUG_ON(m->heap_idx >= h->used);
 587         BUG_ON(h->data[m->heap_idx].idx != idx);
 588 }
 589
 590 void bch2_stripes_heap_del(struct bch_fs *c,
 591                            struct stripe *m, size_t idx)
 592 {
 593         if (!m->on_heap)
 594                 return;
 595
 596         m->on_heap = false;
 597
 598         heap_verify_backpointer(c, idx);
 599
 600         heap_del(&c->ec_stripes_heap, m->heap_idx,
 601                  ec_stripes_heap_cmp,
 602                  ec_stripes_heap_set_backpointer);
 603 }
 604
 605 void bch2_stripes_heap_insert(struct bch_fs *c,
 606                               struct stripe *m, size_t idx)
 607 {
 608         if (m->on_heap)
 609                 return;
 610
 611         BUG_ON(heap_full(&c->ec_stripes_heap));
 612
 613         m->on_heap = true;
 614
 615         heap_add(&c->ec_stripes_heap, ((struct ec_stripe_heap_entry) {
 616                         .idx = idx,
 617                         .blocks_nonempty = m->blocks_nonempty,
 618                 }),
 619                  ec_stripes_heap_cmp,
 620                  ec_stripes_heap_set_backpointer);
 621
 622         heap_verify_backpointer(c, idx);
 623 }
 624
 625 void bch2_stripes_heap_update(struct bch_fs *c,
 626                               struct stripe *m, size_t idx)
 627 {
 628         ec_stripes_heap *h = &c->ec_stripes_heap;
 629         size_t i;
 630
 631         if (!m->on_heap)
 632                 return;
 633
 634         heap_verify_backpointer(c, idx);
 635
 636         h->data[m->heap_idx].blocks_nonempty = m->blocks_nonempty;
 637
 638         i = m->heap_idx;
 639         heap_sift_up(h,   i, ec_stripes_heap_cmp,
 640                      ec_stripes_heap_set_backpointer);
 641         heap_sift_down(h, i, ec_stripes_heap_cmp,
 642                        ec_stripes_heap_set_backpointer);
 643
 644         heap_verify_backpointer(c, idx);
 645
 646         if (stripe_idx_to_delete(c) >= 0 &&
 647             !percpu_ref_is_dying(&c->writes))
 648                 schedule_work(&c->ec_stripe_delete_work);
 649 }
 650
 651 /* stripe deletion */
 652
 653 static int ec_stripe_delete(struct bch_fs *c, size_t idx)
 654 {
 655         //pr_info("deleting stripe %zu", idx);
 656         return bch2_btree_delete_range(c, BTREE_ID_EC,
 657                                        POS(0, idx),
 658                                        POS(0, idx + 1),
 659                                        NULL);
 660 }
 661
 662 static void ec_stripe_delete_work(struct work_struct *work)
 663 {
 664         struct bch_fs *c =
 665                 container_of(work, struct bch_fs, ec_stripe_delete_work);
 666         ssize_t idx;
 667
 668         while (1) {
 669                 spin_lock(&c->ec_stripes_heap_lock);
 670                 idx = stripe_idx_to_delete(c);
 671                 if (idx < 0) {
 672                         spin_unlock(&c->ec_stripes_heap_lock);
 673                         break;
 674                 }
 675
 676                 bch2_stripes_heap_del(c, genradix_ptr(&c->stripes[0], idx), idx);
 677                 spin_unlock(&c->ec_stripes_heap_lock);
 678
 679                 if (ec_stripe_delete(c, idx))
 680                         break;
 681         }
 682 }
 683
 684 /* stripe creation: */
 685
 686 static int ec_stripe_bkey_insert(struct bch_fs *c,
 687                                  struct bkey_i_stripe *stripe,
 688                                  struct disk_reservation *res)
 689 {
 690         struct btree_trans trans;
 691         struct btree_iter *iter;
 692         struct bkey_s_c k;
 693         struct bpos min_pos = POS(0, 1);
 694         struct bpos start_pos = bpos_max(min_pos, POS(0, c->ec_stripe_hint));
 695         int ret;
 696
 697         bch2_trans_init(&trans, c, 0, 0);
 698 retry:
 699         bch2_trans_begin(&trans);
 700
 701         for_each_btree_key(&trans, iter, BTREE_ID_EC, start_pos,
 702                            BTREE_ITER_SLOTS|BTREE_ITER_INTENT, k, ret) {
 703                 if (bkey_cmp(k.k->p, POS(0, U32_MAX)) > 0) {
 704                         if (start_pos.offset) {
 705                                 start_pos = min_pos;
 706                                 bch2_btree_iter_set_pos(iter, start_pos);
 707                                 continue;
 708                         }
 709
 710                         ret = -ENOSPC;
 711                         break;
 712                 }
 713
 714                 if (bkey_deleted(k.k))
 715                         goto found_slot;
 716         }
 717
 718         goto err;
 719 found_slot:
 720         start_pos = iter->pos;
 721
 722         ret = ec_stripe_mem_alloc(c, iter);
 723         if (ret)
 724                 goto err;
 725
 726         stripe->k.p = iter->pos;
 727
 728         bch2_trans_update(&trans, iter, &stripe->k_i, 0);
 729
 730         ret = bch2_trans_commit(&trans, res, NULL,
 731                                 BTREE_INSERT_NOFAIL);
 732 err:
 733         bch2_trans_iter_put(&trans, iter);
 734
 735         if (ret == -EINTR)
 736                 goto retry;
 737
 738         c->ec_stripe_hint = ret ? start_pos.offset : start_pos.offset + 1;
 739         bch2_trans_exit(&trans);
 740
 741         return ret;
 742 }
 743
 744 static int ec_stripe_bkey_update(struct btree_trans *trans,
 745                                  struct bkey_i_stripe *new)
 746 {
 747         struct bch_fs *c = trans->c;
 748         struct btree_iter *iter;
 749         struct bkey_s_c k;
 750         const struct bch_stripe *existing;
 751         unsigned i;
 752         int ret;
 753
 754         iter = bch2_trans_get_iter(trans, BTREE_ID_EC,
 755                                    new->k.p, BTREE_ITER_INTENT);
 756         k = bch2_btree_iter_peek_slot(iter);
 757         ret = bkey_err(k);
 758         if (ret)
 759                 goto err;
 760
 761         if (!k.k || k.k->type != KEY_TYPE_stripe) {
 762                 bch_err(c, "error updating stripe: not found");
 763                 ret = -ENOENT;
 764                 goto err;
 765         }
 766
 767         existing = bkey_s_c_to_stripe(k).v;
 768
 769         if (existing->nr_blocks != new->v.nr_blocks) {
 770                 bch_err(c, "error updating stripe: nr_blocks does not match");
 771                 ret = -EINVAL;
 772                 goto err;
 773         }
 774
 775         for (i = 0; i < new->v.nr_blocks; i++)
 776                 stripe_blockcount_set(&new->v, i,
 777                         stripe_blockcount_get(existing, i));
 778
 779         bch2_trans_update(trans, iter, &new->k_i, 0);
 780 err:
 781         bch2_trans_iter_put(trans, iter);
 782         return ret;
 783 }
 784
 785 static void extent_stripe_ptr_add(struct bkey_s_extent e,
 786                                   struct ec_stripe_buf *s,
 787                                   struct bch_extent_ptr *ptr,
 788                                   unsigned block)
 789 {
 790         struct bch_extent_stripe_ptr *dst = (void *) ptr;
 791         union bch_extent_entry *end = extent_entry_last(e);
 792
 793         memmove_u64s_up(dst + 1, dst, (u64 *) end - (u64 *) dst);
 794         e.k->u64s += sizeof(*dst) / sizeof(u64);
 795
 796         *dst = (struct bch_extent_stripe_ptr) {
 797                 .type = 1 << BCH_EXTENT_ENTRY_stripe_ptr,
 798                 .block          = block,
 799                 .idx            = s->key.k.p.offset,
 800         };
 801 }
 802
 803 static int ec_stripe_update_ptrs(struct bch_fs *c,
 804                                  struct ec_stripe_buf *s,
 805                                  struct bkey *pos)
 806 {
 807         struct btree_trans trans;
 808         struct btree_iter *iter;
 809         struct bkey_s_c k;
 810         struct bkey_s_extent e;
 811         struct bkey_buf sk;
 812         int ret = 0, dev, block;
 813
 814         bch2_bkey_buf_init(&sk);
 815         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
 816
 817         /* XXX this doesn't support the reflink btree */
 818
 819         iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
 820                                    bkey_start_pos(pos),
 821                                    BTREE_ITER_INTENT);
 822
 823         while ((k = bch2_btree_iter_peek(iter)).k &&
 824                !(ret = bkey_err(k)) &&
 825                bkey_cmp(bkey_start_pos(k.k), pos->p) < 0) {
 826                 struct bch_extent_ptr *ptr, *ec_ptr = NULL;
 827
 828                 if (extent_has_stripe_ptr(k, s->key.k.p.offset)) {
 829                         bch2_btree_iter_next(iter);
 830                         continue;
 831                 }
 832
 833                 block = bkey_matches_stripe(&s->key.v, k);
 834                 if (block < 0) {
 835                         bch2_btree_iter_next(iter);
 836                         continue;
 837                 }
 838
 839                 dev = s->key.v.ptrs[block].dev;
 840
 841                 bch2_bkey_buf_reassemble(&sk, c, k);
 842                 e = bkey_i_to_s_extent(sk.k);
 843
 844                 bch2_bkey_drop_ptrs(e.s, ptr, ptr->dev != dev);
 845                 ec_ptr = (void *) bch2_bkey_has_device(e.s_c, dev);
 846                 BUG_ON(!ec_ptr);
 847
 848                 extent_stripe_ptr_add(e, s, ec_ptr, block);
 849
 850                 bch2_btree_iter_set_pos(iter, bkey_start_pos(&sk.k->k));
 851                 bch2_trans_update(&trans, iter, sk.k, 0);
 852
 853                 ret = bch2_trans_commit(&trans, NULL, NULL,
 854                                         BTREE_INSERT_NOFAIL);
 855                 if (ret == -EINTR)
 856                         ret = 0;
 857                 if (ret)
 858                         break;
 859         }
 860
 861         bch2_trans_exit(&trans);
 862         bch2_bkey_buf_exit(&sk, c);
 863
 864         return ret;
 865 }
 866
 867 /*
 868  * data buckets of new stripe all written: create the stripe
 869  */
 870 static void ec_stripe_create(struct ec_stripe_new *s)
 871 {
 872         struct bch_fs *c = s->c;
 873         struct open_bucket *ob;
 874         struct bkey_i *k;
 875         struct stripe *m;
 876         struct bch_stripe *v = &s->new_stripe.key.v;
 877         unsigned i, nr_data = v->nr_blocks - v->nr_redundant;
 878         int ret;
 879
 880         BUG_ON(s->h->s == s);
 881
 882         closure_sync(&s->iodone);
 883
 884         if (s->err) {
 885                 if (s->err != -EROFS)
 886                         bch_err(c, "error creating stripe: error writing data buckets");
 887                 goto err;
 888         }
 889
 890         if (s->have_existing_stripe) {
 891                 ec_validate_checksums(c, &s->existing_stripe);
 892
 893                 if (ec_do_recov(c, &s->existing_stripe)) {
 894                         bch_err(c, "error creating stripe: error reading existing stripe");
 895                         goto err;
 896                 }
 897
 898                 for (i = 0; i < nr_data; i++)
 899                         if (stripe_blockcount_get(&s->existing_stripe.key.v, i))
 900                                 swap(s->new_stripe.data[i],
 901                                      s->existing_stripe.data[i]);
 902
 903                 ec_stripe_buf_exit(&s->existing_stripe);
 904         }
 905
 906         BUG_ON(!s->allocated);
 907
 908         if (!percpu_ref_tryget(&c->writes))
 909                 goto err;
 910
 911         ec_generate_ec(&s->new_stripe);
 912
 913         ec_generate_checksums(&s->new_stripe);
 914
 915         /* write p/q: */
 916         for (i = nr_data; i < v->nr_blocks; i++)
 917                 ec_block_io(c, &s->new_stripe, REQ_OP_WRITE, i, &s->iodone);
 918         closure_sync(&s->iodone);
 919
 920         if (ec_nr_failed(&s->new_stripe)) {
 921                 bch_err(c, "error creating stripe: error writing redundancy buckets");
 922                 goto err_put_writes;
 923         }
 924
 925         ret = s->have_existing_stripe
 926                 ? bch2_trans_do(c, &s->res, NULL, BTREE_INSERT_NOFAIL,
 927                                 ec_stripe_bkey_update(&trans, &s->new_stripe.key))
 928                 : ec_stripe_bkey_insert(c, &s->new_stripe.key, &s->res);
 929         if (ret) {
 930                 bch_err(c, "error creating stripe: error creating stripe key");
 931                 goto err_put_writes;
 932         }
 933
 934         for_each_keylist_key(&s->keys, k) {
 935                 ret = ec_stripe_update_ptrs(c, &s->new_stripe, &k->k);
 936                 if (ret) {
 937                         bch_err(c, "error creating stripe: error %i updating pointers", ret);
 938                         break;
 939                 }
 940         }
 941
 942         spin_lock(&c->ec_stripes_heap_lock);
 943         m = genradix_ptr(&c->stripes[0], s->new_stripe.key.k.p.offset);
 944
 945         BUG_ON(m->on_heap);
 946         bch2_stripes_heap_insert(c, m, s->new_stripe.key.k.p.offset);
 947         spin_unlock(&c->ec_stripes_heap_lock);
 948 err_put_writes:
 949         percpu_ref_put(&c->writes);
 950 err:
 951         bch2_disk_reservation_put(c, &s->res);
 952
 953         for (i = 0; i < v->nr_blocks; i++)
 954                 if (s->blocks[i]) {
 955                         ob = c->open_buckets + s->blocks[i];
 956
 957                         if (i < nr_data) {
 958                                 ob->ec = NULL;
 959                                 __bch2_open_bucket_put(c, ob);
 960                         } else {
 961                                 bch2_open_bucket_put(c, ob);
 962                         }
 963                 }
 964
 965         bch2_keylist_free(&s->keys, s->inline_keys);
 966
 967         ec_stripe_buf_exit(&s->existing_stripe);
 968         ec_stripe_buf_exit(&s->new_stripe);
 969         closure_debug_destroy(&s->iodone);
 970         kfree(s);
 971 }
 972
 973 static void ec_stripe_create_work(struct work_struct *work)
 974 {
 975         struct bch_fs *c = container_of(work,
 976                 struct bch_fs, ec_stripe_create_work);
 977         struct ec_stripe_new *s, *n;
 978 restart:
 979         mutex_lock(&c->ec_stripe_new_lock);
 980         list_for_each_entry_safe(s, n, &c->ec_stripe_new_list, list)
 981                 if (!atomic_read(&s->pin)) {
 982                         list_del(&s->list);
 983                         mutex_unlock(&c->ec_stripe_new_lock);
 984                         ec_stripe_create(s);
 985                         goto restart;
 986                 }
 987         mutex_unlock(&c->ec_stripe_new_lock);
 988 }
 989
 990 static void ec_stripe_new_put(struct bch_fs *c, struct ec_stripe_new *s)
 991 {
 992         BUG_ON(atomic_read(&s->pin) <= 0);
 993
 994         if (atomic_dec_and_test(&s->pin)) {
 995                 BUG_ON(!s->pending);
 996                 queue_work(system_long_wq, &c->ec_stripe_create_work);
 997         }
 998 }
 999
1000 static void ec_stripe_set_pending(struct bch_fs *c, struct ec_stripe_head *h)
1001 {
1002         struct ec_stripe_new *s = h->s;
1003
1004         BUG_ON(!s->allocated && !s->err);
1005
1006         h->s            = NULL;
1007         s->pending      = true;
1008
1009         mutex_lock(&c->ec_stripe_new_lock);
1010         list_add(&s->list, &c->ec_stripe_new_list);
1011         mutex_unlock(&c->ec_stripe_new_lock);
1012
1013         ec_stripe_new_put(c, s);
1014 }
1015
1016 /* have a full bucket - hand it off to be erasure coded: */
1017 void bch2_ec_bucket_written(struct bch_fs *c, struct open_bucket *ob)
1018 {
1019         struct ec_stripe_new *s = ob->ec;
1020
1021         if (ob->sectors_free)
1022                 s->err = -1;
1023
1024         ec_stripe_new_put(c, s);
1025 }
1026
1027 void bch2_ec_bucket_cancel(struct bch_fs *c, struct open_bucket *ob)
1028 {
1029         struct ec_stripe_new *s = ob->ec;
1030
1031         s->err = -EIO;
1032 }
1033
1034 void *bch2_writepoint_ec_buf(struct bch_fs *c, struct write_point *wp)
1035 {
1036         struct open_bucket *ob = ec_open_bucket(c, &wp->ptrs);
1037         struct bch_dev *ca;
1038         unsigned offset;
1039
1040         if (!ob)
1041                 return NULL;
1042
1043         ca      = bch_dev_bkey_exists(c, ob->ptr.dev);
1044         offset  = ca->mi.bucket_size - ob->sectors_free;
1045
1046         return ob->ec->new_stripe.data[ob->ec_idx] + (offset << 9);
1047 }
1048
1049 void bch2_ec_add_backpointer(struct bch_fs *c, struct write_point *wp,
1050                              struct bpos pos, unsigned sectors)
1051 {
1052         struct open_bucket *ob = ec_open_bucket(c, &wp->ptrs);
1053         struct ec_stripe_new *ec;
1054
1055         if (!ob)
1056                 return;
1057
1058         //pr_info("adding backpointer at %llu:%llu", pos.inode, pos.offset);
1059
1060         ec = ob->ec;
1061         mutex_lock(&ec->lock);
1062
1063         if (bch2_keylist_realloc(&ec->keys, ec->inline_keys,
1064                                  ARRAY_SIZE(ec->inline_keys),
1065                                  BKEY_U64s)) {
1066                 BUG();
1067         }
1068
1069         bkey_init(&ec->keys.top->k);
1070         ec->keys.top->k.p       = pos;
1071         bch2_key_resize(&ec->keys.top->k, sectors);
1072         bch2_keylist_push(&ec->keys);
1073
1074         mutex_unlock(&ec->lock);
1075 }
1076
1077 static int unsigned_cmp(const void *_l, const void *_r)
1078 {
1079         unsigned l = *((const unsigned *) _l);
1080         unsigned r = *((const unsigned *) _r);
1081
1082         return cmp_int(l, r);
1083 }
1084
1085 /* pick most common bucket size: */
1086 static unsigned pick_blocksize(struct bch_fs *c,
1087                                struct bch_devs_mask *devs)
1088 {
1089         struct bch_dev *ca;
1090         unsigned i, nr = 0, sizes[BCH_SB_MEMBERS_MAX];
1091         struct {
1092                 unsigned nr, size;
1093         } cur = { 0, 0 }, best = { 0, 0 };
1094
1095         for_each_member_device_rcu(ca, c, i, devs)
1096                 sizes[nr++] = ca->mi.bucket_size;
1097
1098         sort(sizes, nr, sizeof(unsigned), unsigned_cmp, NULL);
1099
1100         for (i = 0; i < nr; i++) {
1101                 if (sizes[i] != cur.size) {
1102                         if (cur.nr > best.nr)
1103                                 best = cur;
1104
1105                         cur.nr = 0;
1106                         cur.size = sizes[i];
1107                 }
1108
1109                 cur.nr++;
1110         }
1111
1112         if (cur.nr > best.nr)
1113                 best = cur;
1114
1115         return best.size;
1116 }
1117
1118 static bool may_create_new_stripe(struct bch_fs *c)
1119 {
1120         return false;
1121 }
1122
1123 static void ec_stripe_key_init(struct bch_fs *c,
1124                                struct bkey_i_stripe *s,
1125                                unsigned nr_data,
1126                                unsigned nr_parity,
1127                                unsigned stripe_size)
1128 {
1129         unsigned u64s;
1130
1131         bkey_stripe_init(&s->k_i);
1132         s->v.sectors                    = cpu_to_le16(stripe_size);
1133         s->v.algorithm                  = 0;
1134         s->v.nr_blocks                  = nr_data + nr_parity;
1135         s->v.nr_redundant               = nr_parity;
1136         s->v.csum_granularity_bits      = ilog2(c->sb.encoded_extent_max);
1137         s->v.csum_type                  = BCH_CSUM_CRC32C;
1138         s->v.pad                        = 0;
1139
1140         while ((u64s = stripe_val_u64s(&s->v)) > BKEY_VAL_U64s_MAX) {
1141                 BUG_ON(1 << s->v.csum_granularity_bits >=
1142                        le16_to_cpu(s->v.sectors) ||
1143                        s->v.csum_granularity_bits == U8_MAX);
1144                 s->v.csum_granularity_bits++;
1145         }
1146
1147         set_bkey_val_u64s(&s->k, u64s);
1148 }
1149
1150 static int ec_new_stripe_alloc(struct bch_fs *c, struct ec_stripe_head *h)
1151 {
1152         struct ec_stripe_new *s;
1153
1154         lockdep_assert_held(&h->lock);
1155
1156         s = kzalloc(sizeof(*s), GFP_KERNEL);
1157         if (!s)
1158                 return -ENOMEM;
1159
1160         mutex_init(&s->lock);
1161         closure_init(&s->iodone, NULL);
1162         atomic_set(&s->pin, 1);
1163         s->c            = c;
1164         s->h            = h;
1165         s->nr_data      = min_t(unsigned, h->nr_active_devs,
1166                                 BCH_BKEY_PTRS_MAX) - h->redundancy;
1167         s->nr_parity    = h->redundancy;
1168
1169         bch2_keylist_init(&s->keys, s->inline_keys);
1170
1171         ec_stripe_key_init(c, &s->new_stripe.key, s->nr_data,
1172                            s->nr_parity, h->blocksize);
1173
1174         h->s = s;
1175         return 0;
1176 }
1177
1178 static struct ec_stripe_head *
1179 ec_new_stripe_head_alloc(struct bch_fs *c, unsigned target,
1180                          unsigned algo, unsigned redundancy,
1181                          bool copygc)
1182 {
1183         struct ec_stripe_head *h;
1184         struct bch_dev *ca;
1185         unsigned i;
1186
1187         h = kzalloc(sizeof(*h), GFP_KERNEL);
1188         if (!h)
1189                 return NULL;
1190
1191         mutex_init(&h->lock);
1192         mutex_lock(&h->lock);
1193
1194         h->target       = target;
1195         h->algo         = algo;
1196         h->redundancy   = redundancy;
1197         h->copygc       = copygc;
1198
1199         rcu_read_lock();
1200         h->devs = target_rw_devs(c, BCH_DATA_user, target);
1201
1202         for_each_member_device_rcu(ca, c, i, &h->devs)
1203                 if (!ca->mi.durability)
1204                         __clear_bit(i, h->devs.d);
1205
1206         h->blocksize = pick_blocksize(c, &h->devs);
1207
1208         for_each_member_device_rcu(ca, c, i, &h->devs)
1209                 if (ca->mi.bucket_size == h->blocksize)
1210                         h->nr_active_devs++;
1211
1212         rcu_read_unlock();
1213         list_add(&h->list, &c->ec_stripe_head_list);
1214         return h;
1215 }
1216
1217 void bch2_ec_stripe_head_put(struct bch_fs *c, struct ec_stripe_head *h)
1218 {
1219         if (h->s &&
1220             h->s->allocated &&
1221             bitmap_weight(h->s->blocks_allocated,
1222                           h->s->nr_data) == h->s->nr_data)
1223                 ec_stripe_set_pending(c, h);
1224
1225         mutex_unlock(&h->lock);
1226 }
1227
1228 struct ec_stripe_head *__bch2_ec_stripe_head_get(struct bch_fs *c,
1229                                                  unsigned target,
1230                                                  unsigned algo,
1231                                                  unsigned redundancy,
1232                                                  bool copygc)
1233 {
1234         struct ec_stripe_head *h;
1235
1236         if (!redundancy)
1237                 return NULL;
1238
1239         mutex_lock(&c->ec_stripe_head_lock);
1240         list_for_each_entry(h, &c->ec_stripe_head_list, list)
1241                 if (h->target           == target &&
1242                     h->algo             == algo &&
1243                     h->redundancy       == redundancy &&
1244                     h->copygc           == copygc) {
1245                         mutex_lock(&h->lock);
1246                         goto found;
1247                 }
1248
1249         h = ec_new_stripe_head_alloc(c, target, algo, redundancy, copygc);
1250 found:
1251         mutex_unlock(&c->ec_stripe_head_lock);
1252         return h;
1253 }
1254
1255 static enum bucket_alloc_ret
1256 new_stripe_alloc_buckets(struct bch_fs *c, struct ec_stripe_head *h,
1257                          struct closure *cl)
1258 {
1259         struct bch_devs_mask devs = h->devs;
1260         struct open_bucket *ob;
1261         struct open_buckets buckets;
1262         unsigned i, j, nr_have_parity = 0, nr_have_data = 0;
1263         bool have_cache = true;
1264         enum bucket_alloc_ret ret = ALLOC_SUCCESS;
1265
1266         for (i = 0; i < h->s->new_stripe.key.v.nr_blocks; i++) {
1267                 if (test_bit(i, h->s->blocks_gotten)) {
1268                         __clear_bit(h->s->new_stripe.key.v.ptrs[i].dev, devs.d);
1269                         if (i < h->s->nr_data)
1270                                 nr_have_data++;
1271                         else
1272                                 nr_have_parity++;
1273                 }
1274         }
1275
1276         BUG_ON(nr_have_data     > h->s->nr_data);
1277         BUG_ON(nr_have_parity   > h->s->nr_parity);
1278
1279         percpu_down_read(&c->mark_lock);
1280         rcu_read_lock();
1281
1282         buckets.nr = 0;
1283         if (nr_have_parity < h->s->nr_parity) {
1284                 ret = bch2_bucket_alloc_set(c, &buckets,
1285                                             &h->parity_stripe,
1286                                             &devs,
1287                                             h->s->nr_parity,
1288                                             &nr_have_parity,
1289                                             &have_cache,
1290                                             h->copygc
1291                                             ? RESERVE_MOVINGGC
1292                                             : RESERVE_NONE,
1293                                             0,
1294                                             cl);
1295
1296                 open_bucket_for_each(c, &buckets, ob, i) {
1297                         j = find_next_zero_bit(h->s->blocks_gotten,
1298                                                h->s->nr_data + h->s->nr_parity,
1299                                                h->s->nr_data);
1300                         BUG_ON(j >= h->s->nr_data + h->s->nr_parity);
1301
1302                         h->s->blocks[j] = buckets.v[i];
1303                         h->s->new_stripe.key.v.ptrs[j] = ob->ptr;
1304                         __set_bit(j, h->s->blocks_gotten);
1305                 }
1306
1307                 if (ret)
1308                         goto err;
1309         }
1310
1311         buckets.nr = 0;
1312         if (nr_have_data < h->s->nr_data) {
1313                 ret = bch2_bucket_alloc_set(c, &buckets,
1314                                             &h->block_stripe,
1315                                             &devs,
1316                                             h->s->nr_data,
1317                                             &nr_have_data,
1318                                             &have_cache,
1319                                             h->copygc
1320                                             ? RESERVE_MOVINGGC
1321                                             : RESERVE_NONE,
1322                                             0,
1323                                             cl);
1324
1325                 open_bucket_for_each(c, &buckets, ob, i) {
1326                         j = find_next_zero_bit(h->s->blocks_gotten,
1327                                                h->s->nr_data, 0);
1328                         BUG_ON(j >= h->s->nr_data);
1329
1330                         h->s->blocks[j] = buckets.v[i];
1331                         h->s->new_stripe.key.v.ptrs[j] = ob->ptr;
1332                         __set_bit(j, h->s->blocks_gotten);
1333                 }
1334
1335                 if (ret)
1336                         goto err;
1337         }
1338 err:
1339         rcu_read_unlock();
1340         percpu_up_read(&c->mark_lock);
1341         return ret;
1342 }
1343
1344 /* XXX: doesn't obey target: */
1345 static s64 get_existing_stripe(struct bch_fs *c,
1346                                struct ec_stripe_head *head)
1347 {
1348         ec_stripes_heap *h = &c->ec_stripes_heap;
1349         struct stripe *m;
1350         size_t heap_idx;
1351         u64 stripe_idx;
1352
1353         if (may_create_new_stripe(c))
1354                 return -1;
1355
1356         spin_lock(&c->ec_stripes_heap_lock);
1357         for (heap_idx = 0; heap_idx < h->used; heap_idx++) {
1358                 if (!h->data[heap_idx].blocks_nonempty)
1359                         continue;
1360
1361                 stripe_idx = h->data[heap_idx].idx;
1362                 m = genradix_ptr(&c->stripes[0], stripe_idx);
1363
1364                 if (m->algorithm        == head->algo &&
1365                     m->nr_redundant     == head->redundancy &&
1366                     m->sectors          == head->blocksize &&
1367                     m->blocks_nonempty  < m->nr_blocks - m->nr_redundant) {
1368                         bch2_stripes_heap_del(c, m, stripe_idx);
1369                         spin_unlock(&c->ec_stripes_heap_lock);
1370                         return stripe_idx;
1371                 }
1372         }
1373
1374         spin_unlock(&c->ec_stripes_heap_lock);
1375         return -1;
1376 }
1377
1378 struct ec_stripe_head *bch2_ec_stripe_head_get(struct bch_fs *c,
1379                                                unsigned target,
1380                                                unsigned algo,
1381                                                unsigned redundancy,
1382                                                bool copygc,
1383                                                struct closure *cl)
1384 {
1385         struct ec_stripe_head *h;
1386         unsigned i;
1387         s64 idx;
1388         int ret;
1389
1390         h = __bch2_ec_stripe_head_get(c, target, algo, redundancy, copygc);
1391         if (!h) {
1392                 bch_err(c, "no stripe head");
1393                 return NULL;
1394         }
1395
1396         if (!h->s) {
1397                 if (ec_new_stripe_alloc(c, h)) {
1398                         bch2_ec_stripe_head_put(c, h);
1399                         bch_err(c, "failed to allocate new stripe");
1400                         return NULL;
1401                 }
1402
1403                 idx = get_existing_stripe(c, h);
1404                 if (idx >= 0) {
1405                         h->s->have_existing_stripe = true;
1406                         ret = get_stripe_key(c, idx, &h->s->existing_stripe);
1407                         if (ret) {
1408                                 bch2_fs_fatal_error(c, "error reading stripe key: %i", ret);
1409                                 bch2_ec_stripe_head_put(c, h);
1410                                 return NULL;
1411                         }
1412
1413                         if (ec_stripe_buf_init(&h->s->existing_stripe, 0, h->blocksize)) {
1414                                 /*
1415                                  * this is a problem: we have deleted from the
1416                                  * stripes heap already
1417                                  */
1418                                 BUG();
1419                         }
1420
1421                         BUG_ON(h->s->existing_stripe.size != h->blocksize);
1422                         BUG_ON(h->s->existing_stripe.size != h->s->existing_stripe.key.v.sectors);
1423
1424                         for (i = 0; i < h->s->existing_stripe.key.v.nr_blocks; i++) {
1425                                 if (stripe_blockcount_get(&h->s->existing_stripe.key.v, i)) {
1426                                         __set_bit(i, h->s->blocks_gotten);
1427                                         __set_bit(i, h->s->blocks_allocated);
1428                                 }
1429
1430                                 ec_block_io(c, &h->s->existing_stripe, READ, i, &h->s->iodone);
1431                         }
1432
1433                         bkey_copy(&h->s->new_stripe.key.k_i,
1434                                   &h->s->existing_stripe.key.k_i);
1435                 }
1436
1437                 if (ec_stripe_buf_init(&h->s->new_stripe, 0, h->blocksize)) {
1438                         BUG();
1439                 }
1440         }
1441
1442         if (!h->s->allocated) {
1443                 if (!h->s->have_existing_stripe &&
1444                     !h->s->res.sectors) {
1445                         ret = bch2_disk_reservation_get(c, &h->s->res,
1446                                         h->blocksize,
1447                                         h->s->nr_parity, 0);
1448                         if (ret) {
1449                                 /*
1450                                  * This means we need to wait for copygc to
1451                                  * empty out buckets from existing stripes:
1452                                  */
1453                                 bch2_ec_stripe_head_put(c, h);
1454                                 h = NULL;
1455                                 goto out;
1456                         }
1457                 }
1458
1459                 ret = new_stripe_alloc_buckets(c, h, cl);
1460                 if (ret) {
1461                         bch2_ec_stripe_head_put(c, h);
1462                         h = ERR_PTR(-ret);
1463                         goto out;
1464                 }
1465
1466                 h->s->allocated = true;
1467         }
1468 out:
1469         return h;
1470 }
1471
1472 void bch2_ec_stop_dev(struct bch_fs *c, struct bch_dev *ca)
1473 {
1474         struct ec_stripe_head *h;
1475         struct open_bucket *ob;
1476         unsigned i;
1477
1478         mutex_lock(&c->ec_stripe_head_lock);
1479         list_for_each_entry(h, &c->ec_stripe_head_list, list) {
1480
1481                 mutex_lock(&h->lock);
1482                 if (!h->s)
1483                         goto unlock;
1484
1485                 for (i = 0; i < h->s->new_stripe.key.v.nr_blocks; i++) {
1486                         if (!h->s->blocks[i])
1487                                 continue;
1488
1489                         ob = c->open_buckets + h->s->blocks[i];
1490                         if (ob->ptr.dev == ca->dev_idx)
1491                                 goto found;
1492                 }
1493                 goto unlock;
1494 found:
1495                 h->s->err = -EROFS;
1496                 ec_stripe_set_pending(c, h);
1497 unlock:
1498                 mutex_unlock(&h->lock);
1499         }
1500         mutex_unlock(&c->ec_stripe_head_lock);
1501 }
1502
1503 void bch2_stripes_heap_start(struct bch_fs *c)
1504 {
1505         struct genradix_iter iter;
1506         struct stripe *m;
1507
1508         genradix_for_each(&c->stripes[0], iter, m)
1509                 if (m->alive)
1510                         bch2_stripes_heap_insert(c, m, iter.pos);
1511 }
1512
1513 static int __bch2_stripe_write_key(struct btree_trans *trans,
1514                                    struct btree_iter *iter,
1515                                    struct stripe *m,
1516                                    size_t idx,
1517                                    struct bkey_i_stripe *new_key)
1518 {
1519         const struct bch_stripe *v;
1520         struct bkey_s_c k;
1521         unsigned i;
1522         int ret;
1523
1524         bch2_btree_iter_set_pos(iter, POS(0, idx));
1525
1526         k = bch2_btree_iter_peek_slot(iter);
1527         ret = bkey_err(k);
1528         if (ret)
1529                 return ret;
1530
1531         if (k.k->type != KEY_TYPE_stripe)
1532                 return -EIO;
1533
1534         v = bkey_s_c_to_stripe(k).v;
1535         for (i = 0; i < v->nr_blocks; i++)
1536                 if (m->block_sectors[i] != stripe_blockcount_get(v, i))
1537                         goto write;
1538         return 0;
1539 write:
1540         bkey_reassemble(&new_key->k_i, k);
1541
1542         for (i = 0; i < new_key->v.nr_blocks; i++)
1543                 stripe_blockcount_set(&new_key->v, i,
1544                                       m->block_sectors[i]);
1545
1546         bch2_trans_update(trans, iter, &new_key->k_i, 0);
1547         return 0;
1548 }
1549
1550 int bch2_stripes_write(struct bch_fs *c, unsigned flags)
1551 {
1552         struct btree_trans trans;
1553         struct btree_iter *iter;
1554         struct genradix_iter giter;
1555         struct bkey_i_stripe *new_key;
1556         struct stripe *m;
1557         int ret = 0;
1558
1559         new_key = kmalloc(255 * sizeof(u64), GFP_KERNEL);
1560         BUG_ON(!new_key);
1561
1562         bch2_trans_init(&trans, c, 0, 0);
1563
1564         iter = bch2_trans_get_iter(&trans, BTREE_ID_EC, POS_MIN,
1565                                    BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
1566
1567         genradix_for_each(&c->stripes[0], giter, m) {
1568                 if (!m->alive)
1569                         continue;
1570
1571                 ret = __bch2_trans_do(&trans, NULL, NULL,
1572                                       BTREE_INSERT_NOFAIL|flags,
1573                         __bch2_stripe_write_key(&trans, iter, m,
1574                                         giter.pos, new_key));
1575
1576                 if (ret)
1577                         break;
1578         }
1579
1580         bch2_trans_exit(&trans);
1581
1582         kfree(new_key);
1583
1584         return ret;
1585 }
1586
1587 static int bch2_stripes_read_fn(struct bch_fs *c, enum btree_id id,
1588                               unsigned level, struct bkey_s_c k)
1589 {
1590         int ret = 0;
1591
1592         if (k.k->type == KEY_TYPE_stripe) {
1593                 ret = __ec_stripe_mem_alloc(c, k.k->p.offset, GFP_KERNEL) ?:
1594                         bch2_mark_key(c, k, 0, 0, NULL, 0,
1595                                       BTREE_TRIGGER_NOATOMIC);
1596                 if (ret)
1597                         return ret;
1598         }
1599
1600         return ret;
1601 }
1602
1603 int bch2_stripes_read(struct bch_fs *c, struct journal_keys *journal_keys)
1604 {
1605         int ret = bch2_btree_and_journal_walk(c, journal_keys, BTREE_ID_EC,
1606                                           NULL, bch2_stripes_read_fn);
1607         if (ret)
1608                 bch_err(c, "error reading stripes: %i", ret);
1609
1610         return ret;
1611 }
1612
1613 int bch2_ec_mem_alloc(struct bch_fs *c, bool gc)
1614 {
1615         struct btree_trans trans;
1616         struct btree_iter *iter;
1617         struct bkey_s_c k;
1618         size_t i, idx = 0;
1619         int ret = 0;
1620
1621         bch2_trans_init(&trans, c, 0, 0);
1622
1623         iter = bch2_trans_get_iter(&trans, BTREE_ID_EC, POS(0, U64_MAX), 0);
1624
1625         k = bch2_btree_iter_prev(iter);
1626         if (!IS_ERR_OR_NULL(k.k))
1627                 idx = k.k->p.offset + 1;
1628         ret = bch2_trans_exit(&trans);
1629         if (ret)
1630                 return ret;
1631
1632         if (!idx)
1633                 return 0;
1634
1635         if (!gc &&
1636             !init_heap(&c->ec_stripes_heap, roundup_pow_of_two(idx),
1637                        GFP_KERNEL))
1638                 return -ENOMEM;
1639 #if 0
1640         ret = genradix_prealloc(&c->stripes[gc], idx, GFP_KERNEL);
1641 #else
1642         for (i = 0; i < idx; i++)
1643                 if (!genradix_ptr_alloc(&c->stripes[gc], i, GFP_KERNEL))
1644                         return -ENOMEM;
1645 #endif
1646         return 0;
1647 }
1648
1649 void bch2_stripes_heap_to_text(struct printbuf *out, struct bch_fs *c)
1650 {
1651         ec_stripes_heap *h = &c->ec_stripes_heap;
1652         struct stripe *m;
1653         size_t i;
1654
1655         spin_lock(&c->ec_stripes_heap_lock);
1656         for (i = 0; i < min_t(size_t, h->used, 20); i++) {
1657                 m = genradix_ptr(&c->stripes[0], h->data[i].idx);
1658
1659                 pr_buf(out, "%zu %u/%u+%u\n", h->data[i].idx,
1660                        h->data[i].blocks_nonempty,
1661                        m->nr_blocks - m->nr_redundant,
1662                        m->nr_redundant);
1663         }
1664         spin_unlock(&c->ec_stripes_heap_lock);
1665 }
1666
1667 void bch2_new_stripes_to_text(struct printbuf *out, struct bch_fs *c)
1668 {
1669         struct ec_stripe_head *h;
1670         struct ec_stripe_new *s;
1671
1672         mutex_lock(&c->ec_stripe_head_lock);
1673         list_for_each_entry(h, &c->ec_stripe_head_list, list) {
1674                 pr_buf(out, "target %u algo %u redundancy %u:\n",
1675                        h->target, h->algo, h->redundancy);
1676
1677                 if (h->s)
1678                         pr_buf(out, "\tpending: blocks %u+%u allocated %u\n",
1679                                h->s->nr_data, h->s->nr_parity,
1680                                bitmap_weight(h->s->blocks_allocated,
1681                                              h->s->nr_data));
1682         }
1683         mutex_unlock(&c->ec_stripe_head_lock);
1684
1685         mutex_lock(&c->ec_stripe_new_lock);
1686         list_for_each_entry(s, &c->ec_stripe_new_list, list) {
1687                 pr_buf(out, "\tin flight: blocks %u+%u pin %u\n",
1688                        s->nr_data, s->nr_parity,
1689                        atomic_read(&s->pin));
1690         }
1691         mutex_unlock(&c->ec_stripe_new_lock);
1692 }
1693
1694 void bch2_fs_ec_exit(struct bch_fs *c)
1695 {
1696         struct ec_stripe_head *h;
1697
1698         while (1) {
1699                 mutex_lock(&c->ec_stripe_head_lock);
1700                 h = list_first_entry_or_null(&c->ec_stripe_head_list,
1701                                              struct ec_stripe_head, list);
1702                 if (h)
1703                         list_del(&h->list);
1704                 mutex_unlock(&c->ec_stripe_head_lock);
1705                 if (!h)
1706                         break;
1707
1708                 BUG_ON(h->s);
1709                 kfree(h);
1710         }
1711
1712         BUG_ON(!list_empty(&c->ec_stripe_new_list));
1713
1714         free_heap(&c->ec_stripes_heap);
1715         genradix_free(&c->stripes[0]);
1716         bioset_exit(&c->ec_bioset);
1717 }
1718
1719 int bch2_fs_ec_init(struct bch_fs *c)
1720 {
1721         INIT_WORK(&c->ec_stripe_create_work, ec_stripe_create_work);
1722         INIT_WORK(&c->ec_stripe_delete_work, ec_stripe_delete_work);
1723
1724         return bioset_init(&c->ec_bioset, 1, offsetof(struct ec_bio, bio),
1725                            BIOSET_NEED_BVECS);
1726 }