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 (!bkey_cmp(k.k->p, POS_MIN))
 109                 return "stripe at pos 0";
 110
 111         if (k.k->p.inode)
 112                 return "invalid stripe key";
 113
 114         if (bkey_val_bytes(k.k) < sizeof(*s))
 115                 return "incorrect value size";
 116
 117         if (bkey_val_bytes(k.k) < sizeof(*s) ||
 118             bkey_val_u64s(k.k) < stripe_val_u64s(s))
 119                 return "incorrect value size";
 120
 121         return bch2_bkey_ptrs_invalid(c, k);
 122 }
 123
 124 void bch2_stripe_to_text(struct printbuf *out, struct bch_fs *c,
 125                          struct bkey_s_c k)
 126 {
 127         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 128         unsigned i;
 129
 130         pr_buf(out, "algo %u sectors %u blocks %u:%u csum %u gran %u",
 131                s->algorithm,
 132                le16_to_cpu(s->sectors),
 133                s->nr_blocks - s->nr_redundant,
 134                s->nr_redundant,
 135                s->csum_type,
 136                1U << s->csum_granularity_bits);
 137
 138         for (i = 0; i < s->nr_blocks; i++)
 139                 pr_buf(out, " %u:%llu:%u", s->ptrs[i].dev,
 140                        (u64) s->ptrs[i].offset,
 141                        stripe_blockcount_get(s, i));
 142 }
 143
 144 /* returns blocknr in stripe that we matched: */
 145 static int bkey_matches_stripe(struct bch_stripe *s,
 146                                struct bkey_s_c k)
 147 {
 148         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 149         const struct bch_extent_ptr *ptr;
 150         unsigned i, nr_data = s->nr_blocks - s->nr_redundant;
 151
 152         bkey_for_each_ptr(ptrs, ptr)
 153                 for (i = 0; i < nr_data; i++)
 154                         if (__bch2_ptr_matches_stripe(&s->ptrs[i], ptr,
 155                                                       le16_to_cpu(s->sectors)))
 156                                 return i;
 157
 158         return -1;
 159 }
 160
 161 static bool extent_has_stripe_ptr(struct bkey_s_c k, u64 idx)
 162 {
 163         switch (k.k->type) {
 164         case KEY_TYPE_extent: {
 165                 struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
 166                 const union bch_extent_entry *entry;
 167
 168                 extent_for_each_entry(e, entry)
 169                         if (extent_entry_type(entry) ==
 170                             BCH_EXTENT_ENTRY_stripe_ptr &&
 171                             entry->stripe_ptr.idx == idx)
 172                                 return true;
 173
 174                 break;
 175         }
 176         }
 177
 178         return false;
 179 }
 180
 181 /* Stripe bufs: */
 182
 183 static void ec_stripe_buf_exit(struct ec_stripe_buf *buf)
 184 {
 185         unsigned i;
 186
 187         for (i = 0; i < buf->key.v.nr_blocks; i++) {
 188                 kvpfree(buf->data[i], buf->size << 9);
 189                 buf->data[i] = NULL;
 190         }
 191 }
 192
 193 static int ec_stripe_buf_init(struct ec_stripe_buf *buf,
 194                                unsigned offset, unsigned size)
 195 {
 196         struct bch_stripe *v = &buf->key.v;
 197         unsigned csum_granularity = 1U << v->csum_granularity_bits;
 198         unsigned end = offset + size;
 199         unsigned i;
 200
 201         BUG_ON(end > le16_to_cpu(v->sectors));
 202
 203         offset  = round_down(offset, csum_granularity);
 204         end     = min_t(unsigned, le16_to_cpu(v->sectors),
 205                         round_up(end, csum_granularity));
 206
 207         buf->offset     = offset;
 208         buf->size       = end - offset;
 209
 210         memset(buf->valid, 0xFF, sizeof(buf->valid));
 211
 212         for (i = 0; i < buf->key.v.nr_blocks; i++) {
 213                 buf->data[i] = kvpmalloc(buf->size << 9, GFP_KERNEL);
 214                 if (!buf->data[i])
 215                         goto err;
 216         }
 217
 218         return 0;
 219 err:
 220         ec_stripe_buf_exit(buf);
 221         return -ENOMEM;
 222 }
 223
 224 /* Checksumming: */
 225
 226 static struct bch_csum ec_block_checksum(struct ec_stripe_buf *buf,
 227                                          unsigned block, unsigned offset)
 228 {
 229         struct bch_stripe *v = &buf->key.v;
 230         unsigned csum_granularity = 1 << v->csum_granularity_bits;
 231         unsigned end = buf->offset + buf->size;
 232         unsigned len = min(csum_granularity, end - offset);
 233
 234         BUG_ON(offset >= end);
 235         BUG_ON(offset <  buf->offset);
 236         BUG_ON(offset & (csum_granularity - 1));
 237         BUG_ON(offset + len != le16_to_cpu(v->sectors) &&
 238                (len & (csum_granularity - 1)));
 239
 240         return bch2_checksum(NULL, v->csum_type,
 241                              null_nonce(),
 242                              buf->data[block] + ((offset - buf->offset) << 9),
 243                              len << 9);
 244 }
 245
 246 static void ec_generate_checksums(struct ec_stripe_buf *buf)
 247 {
 248         struct bch_stripe *v = &buf->key.v;
 249         unsigned i, j, csums_per_device = stripe_csums_per_device(v);
 250
 251         if (!v->csum_type)
 252                 return;
 253
 254         BUG_ON(buf->offset);
 255         BUG_ON(buf->size != le16_to_cpu(v->sectors));
 256
 257         for (i = 0; i < v->nr_blocks; i++)
 258                 for (j = 0; j < csums_per_device; j++)
 259                         stripe_csum_set(v, i, j,
 260                                 ec_block_checksum(buf, i, j << v->csum_granularity_bits));
 261 }
 262
 263 static void ec_validate_checksums(struct bch_fs *c, struct ec_stripe_buf *buf)
 264 {
 265         struct bch_stripe *v = &buf->key.v;
 266         unsigned csum_granularity = 1 << v->csum_granularity_bits;
 267         unsigned i;
 268
 269         if (!v->csum_type)
 270                 return;
 271
 272         for (i = 0; i < v->nr_blocks; i++) {
 273                 unsigned offset = buf->offset;
 274                 unsigned end = buf->offset + buf->size;
 275
 276                 if (!test_bit(i, buf->valid))
 277                         continue;
 278
 279                 while (offset < end) {
 280                         unsigned j = offset >> v->csum_granularity_bits;
 281                         unsigned len = min(csum_granularity, end - offset);
 282                         struct bch_csum want = stripe_csum_get(v, i, j);
 283                         struct bch_csum got = ec_block_checksum(buf, i, offset);
 284
 285                         if (bch2_crc_cmp(want, got)) {
 286                                 char buf2[200];
 287
 288                                 bch2_bkey_val_to_text(&PBUF(buf2), c, bkey_i_to_s_c(&buf->key.k_i));
 289
 290                                 bch_err_ratelimited(c,
 291                                         "stripe checksum error for %ps at %u:%u: csum type %u, expected %llx got %llx\n%s",
 292                                         (void *) _RET_IP_, i, j, v->csum_type,
 293                                         want.lo, got.lo, buf2);
 294                                 clear_bit(i, buf->valid);
 295                                 break;
 296                         }
 297
 298                         offset += len;
 299                 }
 300         }
 301 }
 302
 303 /* Erasure coding: */
 304
 305 static void ec_generate_ec(struct ec_stripe_buf *buf)
 306 {
 307         struct bch_stripe *v = &buf->key.v;
 308         unsigned nr_data = v->nr_blocks - v->nr_redundant;
 309         unsigned bytes = le16_to_cpu(v->sectors) << 9;
 310
 311         raid_gen(nr_data, v->nr_redundant, bytes, buf->data);
 312 }
 313
 314 static unsigned ec_nr_failed(struct ec_stripe_buf *buf)
 315 {
 316         return buf->key.v.nr_blocks -
 317                 bitmap_weight(buf->valid, buf->key.v.nr_blocks);
 318 }
 319
 320 static int ec_do_recov(struct bch_fs *c, struct ec_stripe_buf *buf)
 321 {
 322         struct bch_stripe *v = &buf->key.v;
 323         unsigned i, failed[BCH_BKEY_PTRS_MAX], nr_failed = 0;
 324         unsigned nr_data = v->nr_blocks - v->nr_redundant;
 325         unsigned bytes = buf->size << 9;
 326
 327         if (ec_nr_failed(buf) > v->nr_redundant) {
 328                 bch_err_ratelimited(c,
 329                         "error doing reconstruct read: unable to read enough blocks");
 330                 return -1;
 331         }
 332
 333         for (i = 0; i < nr_data; i++)
 334                 if (!test_bit(i, buf->valid))
 335                         failed[nr_failed++] = i;
 336
 337         raid_rec(nr_failed, failed, nr_data, v->nr_redundant, bytes, buf->data);
 338         return 0;
 339 }
 340
 341 /* IO: */
 342
 343 static void ec_block_endio(struct bio *bio)
 344 {
 345         struct ec_bio *ec_bio = container_of(bio, struct ec_bio, bio);
 346         struct bch_stripe *v = &ec_bio->buf->key.v;
 347         struct bch_extent_ptr *ptr = &v->ptrs[ec_bio->idx];
 348         struct bch_dev *ca = ec_bio->ca;
 349         struct closure *cl = bio->bi_private;
 350
 351         if (bch2_dev_io_err_on(bio->bi_status, ca, "erasure coding %s error: %s",
 352                                bio_data_dir(bio) ? "write" : "read",
 353                                bch2_blk_status_to_str(bio->bi_status)))
 354                 clear_bit(ec_bio->idx, ec_bio->buf->valid);
 355
 356         if (ptr_stale(ca, ptr)) {
 357                 bch_err_ratelimited(ca->fs,
 358                                     "error %s stripe: stale pointer after io",
 359                                     bio_data_dir(bio) == READ ? "reading from" : "writing to");
 360                 clear_bit(ec_bio->idx, ec_bio->buf->valid);
 361         }
 362
 363         bio_put(&ec_bio->bio);
 364         percpu_ref_put(&ca->io_ref);
 365         closure_put(cl);
 366 }
 367
 368 static void ec_block_io(struct bch_fs *c, struct ec_stripe_buf *buf,
 369                         unsigned rw, unsigned idx, struct closure *cl)
 370 {
 371         struct bch_stripe *v = &buf->key.v;
 372         unsigned offset = 0, bytes = buf->size << 9;
 373         struct bch_extent_ptr *ptr = &v->ptrs[idx];
 374         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 375         enum bch_data_type data_type = idx < buf->key.v.nr_blocks - buf->key.v.nr_redundant
 376                 ? BCH_DATA_user
 377                 : BCH_DATA_parity;
 378
 379         if (ptr_stale(ca, ptr)) {
 380                 bch_err_ratelimited(c,
 381                                     "error %s stripe: stale pointer",
 382                                     rw == READ ? "reading from" : "writing to");
 383                 clear_bit(idx, buf->valid);
 384                 return;
 385         }
 386
 387         if (!bch2_dev_get_ioref(ca, rw)) {
 388                 clear_bit(idx, buf->valid);
 389                 return;
 390         }
 391
 392         this_cpu_add(ca->io_done->sectors[rw][data_type], buf->size);
 393
 394         while (offset < bytes) {
 395                 unsigned nr_iovecs = min_t(size_t, BIO_MAX_VECS,
 396                                            DIV_ROUND_UP(bytes, PAGE_SIZE));
 397                 unsigned b = min_t(size_t, bytes - offset,
 398                                    nr_iovecs << PAGE_SHIFT);
 399                 struct ec_bio *ec_bio;
 400
 401                 ec_bio = container_of(bio_alloc_bioset(GFP_KERNEL, nr_iovecs,
 402                                                        &c->ec_bioset),
 403                                       struct ec_bio, bio);
 404
 405                 ec_bio->ca                      = ca;
 406                 ec_bio->buf                     = buf;
 407                 ec_bio->idx                     = idx;
 408
 409                 bio_set_dev(&ec_bio->bio, ca->disk_sb.bdev);
 410                 bio_set_op_attrs(&ec_bio->bio, rw, 0);
 411
 412                 ec_bio->bio.bi_iter.bi_sector   = ptr->offset + buf->offset + (offset >> 9);
 413                 ec_bio->bio.bi_end_io           = ec_block_endio;
 414                 ec_bio->bio.bi_private          = cl;
 415
 416                 bch2_bio_map(&ec_bio->bio, buf->data[idx] + offset, b);
 417
 418                 closure_get(cl);
 419                 percpu_ref_get(&ca->io_ref);
 420
 421                 submit_bio(&ec_bio->bio);
 422
 423                 offset += b;
 424         }
 425
 426         percpu_ref_put(&ca->io_ref);
 427 }
 428
 429 static int get_stripe_key(struct bch_fs *c, u64 idx, struct ec_stripe_buf *stripe)
 430 {
 431         struct btree_trans trans;
 432         struct btree_iter *iter;
 433         struct bkey_s_c k;
 434         int ret;
 435
 436         bch2_trans_init(&trans, c, 0, 0);
 437         iter = bch2_trans_get_iter(&trans, BTREE_ID_stripes, POS(0, idx), BTREE_ITER_SLOTS);
 438         k = bch2_btree_iter_peek_slot(iter);
 439         ret = bkey_err(k);
 440         if (ret)
 441                 goto err;
 442         if (k.k->type != KEY_TYPE_stripe) {
 443                 ret = -ENOENT;
 444                 goto err;
 445         }
 446         bkey_reassemble(&stripe->key.k_i, k);
 447 err:
 448         bch2_trans_exit(&trans);
 449         return ret;
 450 }
 451
 452 /* recovery read path: */
 453 int bch2_ec_read_extent(struct bch_fs *c, struct bch_read_bio *rbio)
 454 {
 455         struct ec_stripe_buf *buf;
 456         struct closure cl;
 457         struct bch_stripe *v;
 458         unsigned i, offset;
 459         int ret = 0;
 460
 461         closure_init_stack(&cl);
 462
 463         BUG_ON(!rbio->pick.has_ec);
 464
 465         buf = kzalloc(sizeof(*buf), GFP_NOIO);
 466         if (!buf)
 467                 return -ENOMEM;
 468
 469         ret = get_stripe_key(c, rbio->pick.ec.idx, buf);
 470         if (ret) {
 471                 bch_err_ratelimited(c,
 472                         "error doing reconstruct read: error %i looking up stripe", ret);
 473                 kfree(buf);
 474                 return -EIO;
 475         }
 476
 477         v = &buf->key.v;
 478
 479         if (!bch2_ptr_matches_stripe(v, rbio->pick)) {
 480                 bch_err_ratelimited(c,
 481                         "error doing reconstruct read: pointer doesn't match stripe");
 482                 ret = -EIO;
 483                 goto err;
 484         }
 485
 486         offset = rbio->bio.bi_iter.bi_sector - v->ptrs[rbio->pick.ec.block].offset;
 487         if (offset + bio_sectors(&rbio->bio) > le16_to_cpu(v->sectors)) {
 488                 bch_err_ratelimited(c,
 489                         "error doing reconstruct read: read is bigger than stripe");
 490                 ret = -EIO;
 491                 goto err;
 492         }
 493
 494         ret = ec_stripe_buf_init(buf, offset, bio_sectors(&rbio->bio));
 495         if (ret)
 496                 goto err;
 497
 498         for (i = 0; i < v->nr_blocks; i++)
 499                 ec_block_io(c, buf, REQ_OP_READ, i, &cl);
 500
 501         closure_sync(&cl);
 502
 503         if (ec_nr_failed(buf) > v->nr_redundant) {
 504                 bch_err_ratelimited(c,
 505                         "error doing reconstruct read: unable to read enough blocks");
 506                 ret = -EIO;
 507                 goto err;
 508         }
 509
 510         ec_validate_checksums(c, buf);
 511
 512         ret = ec_do_recov(c, buf);
 513         if (ret)
 514                 goto err;
 515
 516         memcpy_to_bio(&rbio->bio, rbio->bio.bi_iter,
 517                       buf->data[rbio->pick.ec.block] + ((offset - buf->offset) << 9));
 518 err:
 519         ec_stripe_buf_exit(buf);
 520         kfree(buf);
 521         return ret;
 522 }
 523
 524 /* stripe bucket accounting: */
 525
 526 static int __ec_stripe_mem_alloc(struct bch_fs *c, size_t idx, gfp_t gfp)
 527 {
 528         ec_stripes_heap n, *h = &c->ec_stripes_heap;
 529
 530         if (idx >= h->size) {
 531                 if (!init_heap(&n, max(1024UL, roundup_pow_of_two(idx + 1)), gfp))
 532                         return -ENOMEM;
 533
 534                 spin_lock(&c->ec_stripes_heap_lock);
 535                 if (n.size > h->size) {
 536                         memcpy(n.data, h->data, h->used * sizeof(h->data[0]));
 537                         n.used = h->used;
 538                         swap(*h, n);
 539                 }
 540                 spin_unlock(&c->ec_stripes_heap_lock);
 541
 542                 free_heap(&n);
 543         }
 544
 545         if (!genradix_ptr_alloc(&c->stripes[0], idx, gfp))
 546                 return -ENOMEM;
 547
 548         if (c->gc_pos.phase != GC_PHASE_NOT_RUNNING &&
 549             !genradix_ptr_alloc(&c->stripes[1], idx, gfp))
 550                 return -ENOMEM;
 551
 552         return 0;
 553 }
 554
 555 static int ec_stripe_mem_alloc(struct bch_fs *c,
 556                                struct btree_iter *iter)
 557 {
 558         size_t idx = iter->pos.offset;
 559         int ret = 0;
 560
 561         if (!__ec_stripe_mem_alloc(c, idx, GFP_NOWAIT|__GFP_NOWARN))
 562                 return ret;
 563
 564         bch2_trans_unlock(iter->trans);
 565         ret = -EINTR;
 566
 567         if (!__ec_stripe_mem_alloc(c, idx, GFP_KERNEL))
 568                 return ret;
 569
 570         return -ENOMEM;
 571 }
 572
 573 static ssize_t stripe_idx_to_delete(struct bch_fs *c)
 574 {
 575         ec_stripes_heap *h = &c->ec_stripes_heap;
 576
 577         return h->used && h->data[0].blocks_nonempty == 0
 578                 ? h->data[0].idx : -1;
 579 }
 580
 581 static inline int ec_stripes_heap_cmp(ec_stripes_heap *h,
 582                                       struct ec_stripe_heap_entry l,
 583                                       struct ec_stripe_heap_entry r)
 584 {
 585         return ((l.blocks_nonempty > r.blocks_nonempty) -
 586                 (l.blocks_nonempty < r.blocks_nonempty));
 587 }
 588
 589 static inline void ec_stripes_heap_set_backpointer(ec_stripes_heap *h,
 590                                                    size_t i)
 591 {
 592         struct bch_fs *c = container_of(h, struct bch_fs, ec_stripes_heap);
 593
 594         genradix_ptr(&c->stripes[0], h->data[i].idx)->heap_idx = i;
 595 }
 596
 597 static void heap_verify_backpointer(struct bch_fs *c, size_t idx)
 598 {
 599         ec_stripes_heap *h = &c->ec_stripes_heap;
 600         struct stripe *m = genradix_ptr(&c->stripes[0], idx);
 601
 602         BUG_ON(!m->alive);
 603         BUG_ON(m->heap_idx >= h->used);
 604         BUG_ON(h->data[m->heap_idx].idx != idx);
 605 }
 606
 607 void bch2_stripes_heap_del(struct bch_fs *c,
 608                            struct stripe *m, size_t idx)
 609 {
 610         if (!m->on_heap)
 611                 return;
 612
 613         m->on_heap = false;
 614
 615         heap_verify_backpointer(c, idx);
 616
 617         heap_del(&c->ec_stripes_heap, m->heap_idx,
 618                  ec_stripes_heap_cmp,
 619                  ec_stripes_heap_set_backpointer);
 620 }
 621
 622 void bch2_stripes_heap_insert(struct bch_fs *c,
 623                               struct stripe *m, size_t idx)
 624 {
 625         if (m->on_heap)
 626                 return;
 627
 628         BUG_ON(heap_full(&c->ec_stripes_heap));
 629
 630         m->on_heap = true;
 631
 632         heap_add(&c->ec_stripes_heap, ((struct ec_stripe_heap_entry) {
 633                         .idx = idx,
 634                         .blocks_nonempty = m->blocks_nonempty,
 635                 }),
 636                  ec_stripes_heap_cmp,
 637                  ec_stripes_heap_set_backpointer);
 638
 639         heap_verify_backpointer(c, idx);
 640 }
 641
 642 void bch2_stripes_heap_update(struct bch_fs *c,
 643                               struct stripe *m, size_t idx)
 644 {
 645         ec_stripes_heap *h = &c->ec_stripes_heap;
 646         size_t i;
 647
 648         if (!m->on_heap)
 649                 return;
 650
 651         heap_verify_backpointer(c, idx);
 652
 653         h->data[m->heap_idx].blocks_nonempty = m->blocks_nonempty;
 654
 655         i = m->heap_idx;
 656         heap_sift_up(h,   i, ec_stripes_heap_cmp,
 657                      ec_stripes_heap_set_backpointer);
 658         heap_sift_down(h, i, ec_stripes_heap_cmp,
 659                        ec_stripes_heap_set_backpointer);
 660
 661         heap_verify_backpointer(c, idx);
 662
 663         if (stripe_idx_to_delete(c) >= 0 &&
 664             !percpu_ref_is_dying(&c->writes))
 665                 schedule_work(&c->ec_stripe_delete_work);
 666 }
 667
 668 /* stripe deletion */
 669
 670 static int ec_stripe_delete(struct bch_fs *c, size_t idx)
 671 {
 672         return bch2_btree_delete_range(c, BTREE_ID_stripes,
 673                                        POS(0, idx),
 674                                        POS(0, idx + 1),
 675                                        NULL);
 676 }
 677
 678 static void ec_stripe_delete_work(struct work_struct *work)
 679 {
 680         struct bch_fs *c =
 681                 container_of(work, struct bch_fs, ec_stripe_delete_work);
 682         ssize_t idx;
 683
 684         while (1) {
 685                 spin_lock(&c->ec_stripes_heap_lock);
 686                 idx = stripe_idx_to_delete(c);
 687                 if (idx < 0) {
 688                         spin_unlock(&c->ec_stripes_heap_lock);
 689                         break;
 690                 }
 691
 692                 bch2_stripes_heap_del(c, genradix_ptr(&c->stripes[0], idx), idx);
 693                 spin_unlock(&c->ec_stripes_heap_lock);
 694
 695                 if (ec_stripe_delete(c, idx))
 696                         break;
 697         }
 698 }
 699
 700 /* stripe creation: */
 701
 702 static int ec_stripe_bkey_insert(struct bch_fs *c,
 703                                  struct bkey_i_stripe *stripe,
 704                                  struct disk_reservation *res)
 705 {
 706         struct btree_trans trans;
 707         struct btree_iter *iter;
 708         struct bkey_s_c k;
 709         struct bpos min_pos = POS(0, 1);
 710         struct bpos start_pos = bpos_max(min_pos, POS(0, c->ec_stripe_hint));
 711         int ret;
 712
 713         bch2_trans_init(&trans, c, 0, 0);
 714 retry:
 715         bch2_trans_begin(&trans);
 716
 717         for_each_btree_key(&trans, iter, BTREE_ID_stripes, start_pos,
 718                            BTREE_ITER_SLOTS|BTREE_ITER_INTENT, k, ret) {
 719                 if (bkey_cmp(k.k->p, POS(0, U32_MAX)) > 0) {
 720                         if (start_pos.offset) {
 721                                 start_pos = min_pos;
 722                                 bch2_btree_iter_set_pos(iter, start_pos);
 723                                 continue;
 724                         }
 725
 726                         ret = -ENOSPC;
 727                         break;
 728                 }
 729
 730                 if (bkey_deleted(k.k))
 731                         goto found_slot;
 732         }
 733
 734         goto err;
 735 found_slot:
 736         start_pos = iter->pos;
 737
 738         ret = ec_stripe_mem_alloc(c, iter);
 739         if (ret)
 740                 goto err;
 741
 742         stripe->k.p = iter->pos;
 743
 744         ret   = bch2_trans_update(&trans, iter, &stripe->k_i, 0) ?:
 745                 bch2_trans_commit(&trans, res, NULL,
 746                                 BTREE_INSERT_NOFAIL);
 747 err:
 748         bch2_trans_iter_put(&trans, iter);
 749
 750         if (ret == -EINTR)
 751                 goto retry;
 752
 753         c->ec_stripe_hint = ret ? start_pos.offset : start_pos.offset + 1;
 754         bch2_trans_exit(&trans);
 755
 756         return ret;
 757 }
 758
 759 static int ec_stripe_bkey_update(struct btree_trans *trans,
 760                                  struct bkey_i_stripe *new)
 761 {
 762         struct btree_iter *iter;
 763         struct bkey_s_c k;
 764         const struct bch_stripe *existing;
 765         unsigned i;
 766         int ret;
 767
 768         iter = bch2_trans_get_iter(trans, BTREE_ID_stripes,
 769                                    new->k.p, BTREE_ITER_INTENT);
 770         k = bch2_btree_iter_peek_slot(iter);
 771         ret = bkey_err(k);
 772         if (ret)
 773                 goto err;
 774
 775         if (!k.k || k.k->type != KEY_TYPE_stripe) {
 776                 bch_err(trans->c, "error updating stripe: not found");
 777                 ret = -ENOENT;
 778                 goto err;
 779         }
 780
 781         existing = bkey_s_c_to_stripe(k).v;
 782
 783         if (existing->nr_blocks != new->v.nr_blocks) {
 784                 bch_err(trans->c, "error updating stripe: nr_blocks does not match");
 785                 ret = -EINVAL;
 786                 goto err;
 787         }
 788
 789         for (i = 0; i < new->v.nr_blocks; i++)
 790                 stripe_blockcount_set(&new->v, i,
 791                         stripe_blockcount_get(existing, i));
 792
 793         ret = bch2_trans_update(trans, iter, &new->k_i, 0);
 794 err:
 795         bch2_trans_iter_put(trans, iter);
 796         return ret;
 797 }
 798
 799 static void extent_stripe_ptr_add(struct bkey_s_extent e,
 800                                   struct ec_stripe_buf *s,
 801                                   struct bch_extent_ptr *ptr,
 802                                   unsigned block)
 803 {
 804         struct bch_extent_stripe_ptr *dst = (void *) ptr;
 805         union bch_extent_entry *end = extent_entry_last(e);
 806
 807         memmove_u64s_up(dst + 1, dst, (u64 *) end - (u64 *) dst);
 808         e.k->u64s += sizeof(*dst) / sizeof(u64);
 809
 810         *dst = (struct bch_extent_stripe_ptr) {
 811                 .type = 1 << BCH_EXTENT_ENTRY_stripe_ptr,
 812                 .block          = block,
 813                 .redundancy     = s->key.v.nr_redundant,
 814                 .idx            = s->key.k.p.offset,
 815         };
 816 }
 817
 818 static int ec_stripe_update_ptrs(struct bch_fs *c,
 819                                  struct ec_stripe_buf *s,
 820                                  struct bkey *pos)
 821 {
 822         struct btree_trans trans;
 823         struct btree_iter *iter;
 824         struct bkey_s_c k;
 825         struct bkey_s_extent e;
 826         struct bkey_buf sk;
 827         int ret = 0, dev, block;
 828
 829         bch2_bkey_buf_init(&sk);
 830         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
 831
 832         /* XXX this doesn't support the reflink btree */
 833
 834         iter = bch2_trans_get_iter(&trans, BTREE_ID_extents,
 835                                    bkey_start_pos(pos),
 836                                    BTREE_ITER_INTENT);
 837
 838         while ((k = bch2_btree_iter_peek(iter)).k &&
 839                !(ret = bkey_err(k)) &&
 840                bkey_cmp(bkey_start_pos(k.k), pos->p) < 0) {
 841                 struct bch_extent_ptr *ptr, *ec_ptr = NULL;
 842
 843                 if (extent_has_stripe_ptr(k, s->key.k.p.offset)) {
 844                         bch2_btree_iter_advance(iter);
 845                         continue;
 846                 }
 847
 848                 block = bkey_matches_stripe(&s->key.v, k);
 849                 if (block < 0) {
 850                         bch2_btree_iter_advance(iter);
 851                         continue;
 852                 }
 853
 854                 dev = s->key.v.ptrs[block].dev;
 855
 856                 bch2_bkey_buf_reassemble(&sk, c, k);
 857                 e = bkey_i_to_s_extent(sk.k);
 858
 859                 bch2_bkey_drop_ptrs(e.s, ptr, ptr->dev != dev);
 860                 ec_ptr = (void *) bch2_bkey_has_device(e.s_c, dev);
 861                 BUG_ON(!ec_ptr);
 862
 863                 extent_stripe_ptr_add(e, s, ec_ptr, block);
 864
 865                 bch2_btree_iter_set_pos(iter, bkey_start_pos(&sk.k->k));
 866                 ret   = bch2_trans_update(&trans, iter, sk.k, 0) ?:
 867                         bch2_trans_commit(&trans, NULL, NULL,
 868                                         BTREE_INSERT_NOFAIL);
 869                 if (ret == -EINTR)
 870                         ret = 0;
 871                 if (ret)
 872                         break;
 873         }
 874         bch2_trans_iter_put(&trans, iter);
 875
 876         bch2_trans_exit(&trans);
 877         bch2_bkey_buf_exit(&sk, c);
 878
 879         return ret;
 880 }
 881
 882 /*
 883  * data buckets of new stripe all written: create the stripe
 884  */
 885 static void ec_stripe_create(struct ec_stripe_new *s)
 886 {
 887         struct bch_fs *c = s->c;
 888         struct open_bucket *ob;
 889         struct bkey_i *k;
 890         struct stripe *m;
 891         struct bch_stripe *v = &s->new_stripe.key.v;
 892         unsigned i, nr_data = v->nr_blocks - v->nr_redundant;
 893         int ret;
 894
 895         BUG_ON(s->h->s == s);
 896
 897         closure_sync(&s->iodone);
 898
 899         if (s->err) {
 900                 if (s->err != -EROFS)
 901                         bch_err(c, "error creating stripe: error writing data buckets");
 902                 goto err;
 903         }
 904
 905         if (s->have_existing_stripe) {
 906                 ec_validate_checksums(c, &s->existing_stripe);
 907
 908                 if (ec_do_recov(c, &s->existing_stripe)) {
 909                         bch_err(c, "error creating stripe: error reading existing stripe");
 910                         goto err;
 911                 }
 912
 913                 for (i = 0; i < nr_data; i++)
 914                         if (stripe_blockcount_get(&s->existing_stripe.key.v, i))
 915                                 swap(s->new_stripe.data[i],
 916                                      s->existing_stripe.data[i]);
 917
 918                 ec_stripe_buf_exit(&s->existing_stripe);
 919         }
 920
 921         BUG_ON(!s->allocated);
 922
 923         if (!percpu_ref_tryget(&c->writes))
 924                 goto err;
 925
 926         ec_generate_ec(&s->new_stripe);
 927
 928         ec_generate_checksums(&s->new_stripe);
 929
 930         /* write p/q: */
 931         for (i = nr_data; i < v->nr_blocks; i++)
 932                 ec_block_io(c, &s->new_stripe, REQ_OP_WRITE, i, &s->iodone);
 933         closure_sync(&s->iodone);
 934
 935         if (ec_nr_failed(&s->new_stripe)) {
 936                 bch_err(c, "error creating stripe: error writing redundancy buckets");
 937                 goto err_put_writes;
 938         }
 939
 940         ret = s->have_existing_stripe
 941                 ? bch2_trans_do(c, &s->res, NULL, BTREE_INSERT_NOFAIL,
 942                                 ec_stripe_bkey_update(&trans, &s->new_stripe.key))
 943                 : ec_stripe_bkey_insert(c, &s->new_stripe.key, &s->res);
 944         if (ret) {
 945                 bch_err(c, "error creating stripe: error creating stripe key");
 946                 goto err_put_writes;
 947         }
 948
 949         for_each_keylist_key(&s->keys, k) {
 950                 ret = ec_stripe_update_ptrs(c, &s->new_stripe, &k->k);
 951                 if (ret) {
 952                         bch_err(c, "error creating stripe: error %i updating pointers", ret);
 953                         break;
 954                 }
 955         }
 956
 957         spin_lock(&c->ec_stripes_heap_lock);
 958         m = genradix_ptr(&c->stripes[0], s->new_stripe.key.k.p.offset);
 959
 960         BUG_ON(m->on_heap);
 961         bch2_stripes_heap_insert(c, m, s->new_stripe.key.k.p.offset);
 962         spin_unlock(&c->ec_stripes_heap_lock);
 963 err_put_writes:
 964         percpu_ref_put(&c->writes);
 965 err:
 966         bch2_disk_reservation_put(c, &s->res);
 967
 968         for (i = 0; i < v->nr_blocks; i++)
 969                 if (s->blocks[i]) {
 970                         ob = c->open_buckets + s->blocks[i];
 971
 972                         if (i < nr_data) {
 973                                 ob->ec = NULL;
 974                                 __bch2_open_bucket_put(c, ob);
 975                         } else {
 976                                 bch2_open_bucket_put(c, ob);
 977                         }
 978                 }
 979
 980         bch2_keylist_free(&s->keys, s->inline_keys);
 981
 982         ec_stripe_buf_exit(&s->existing_stripe);
 983         ec_stripe_buf_exit(&s->new_stripe);
 984         closure_debug_destroy(&s->iodone);
 985         kfree(s);
 986 }
 987
 988 static void ec_stripe_create_work(struct work_struct *work)
 989 {
 990         struct bch_fs *c = container_of(work,
 991                 struct bch_fs, ec_stripe_create_work);
 992         struct ec_stripe_new *s, *n;
 993 restart:
 994         mutex_lock(&c->ec_stripe_new_lock);
 995         list_for_each_entry_safe(s, n, &c->ec_stripe_new_list, list)
 996                 if (!atomic_read(&s->pin)) {
 997                         list_del(&s->list);
 998                         mutex_unlock(&c->ec_stripe_new_lock);
 999                         ec_stripe_create(s);
1000                         goto restart;
1001                 }
1002         mutex_unlock(&c->ec_stripe_new_lock);
1003 }
1004
1005 static void ec_stripe_new_put(struct bch_fs *c, struct ec_stripe_new *s)
1006 {
1007         BUG_ON(atomic_read(&s->pin) <= 0);
1008
1009         if (atomic_dec_and_test(&s->pin)) {
1010                 BUG_ON(!s->pending);
1011                 queue_work(system_long_wq, &c->ec_stripe_create_work);
1012         }
1013 }
1014
1015 static void ec_stripe_set_pending(struct bch_fs *c, struct ec_stripe_head *h)
1016 {
1017         struct ec_stripe_new *s = h->s;
1018
1019         BUG_ON(!s->allocated && !s->err);
1020
1021         h->s            = NULL;
1022         s->pending      = true;
1023
1024         mutex_lock(&c->ec_stripe_new_lock);
1025         list_add(&s->list, &c->ec_stripe_new_list);
1026         mutex_unlock(&c->ec_stripe_new_lock);
1027
1028         ec_stripe_new_put(c, s);
1029 }
1030
1031 /* have a full bucket - hand it off to be erasure coded: */
1032 void bch2_ec_bucket_written(struct bch_fs *c, struct open_bucket *ob)
1033 {
1034         struct ec_stripe_new *s = ob->ec;
1035
1036         if (ob->sectors_free)
1037                 s->err = -1;
1038
1039         ec_stripe_new_put(c, s);
1040 }
1041
1042 void bch2_ec_bucket_cancel(struct bch_fs *c, struct open_bucket *ob)
1043 {
1044         struct ec_stripe_new *s = ob->ec;
1045
1046         s->err = -EIO;
1047 }
1048
1049 void *bch2_writepoint_ec_buf(struct bch_fs *c, struct write_point *wp)
1050 {
1051         struct open_bucket *ob = ec_open_bucket(c, &wp->ptrs);
1052         struct bch_dev *ca;
1053         unsigned offset;
1054
1055         if (!ob)
1056                 return NULL;
1057
1058         ca      = bch_dev_bkey_exists(c, ob->ptr.dev);
1059         offset  = ca->mi.bucket_size - ob->sectors_free;
1060
1061         return ob->ec->new_stripe.data[ob->ec_idx] + (offset << 9);
1062 }
1063
1064 void bch2_ec_add_backpointer(struct bch_fs *c, struct write_point *wp,
1065                              struct bpos pos, unsigned sectors)
1066 {
1067         struct open_bucket *ob = ec_open_bucket(c, &wp->ptrs);
1068         struct ec_stripe_new *ec;
1069
1070         if (!ob)
1071                 return;
1072
1073         ec = ob->ec;
1074         mutex_lock(&ec->lock);
1075
1076         if (bch2_keylist_realloc(&ec->keys, ec->inline_keys,
1077                                  ARRAY_SIZE(ec->inline_keys),
1078                                  BKEY_U64s)) {
1079                 BUG();
1080         }
1081
1082         bkey_init(&ec->keys.top->k);
1083         ec->keys.top->k.p       = pos;
1084         bch2_key_resize(&ec->keys.top->k, sectors);
1085         bch2_keylist_push(&ec->keys);
1086
1087         mutex_unlock(&ec->lock);
1088 }
1089
1090 static int unsigned_cmp(const void *_l, const void *_r)
1091 {
1092         unsigned l = *((const unsigned *) _l);
1093         unsigned r = *((const unsigned *) _r);
1094
1095         return cmp_int(l, r);
1096 }
1097
1098 /* pick most common bucket size: */
1099 static unsigned pick_blocksize(struct bch_fs *c,
1100                                struct bch_devs_mask *devs)
1101 {
1102         struct bch_dev *ca;
1103         unsigned i, nr = 0, sizes[BCH_SB_MEMBERS_MAX];
1104         struct {
1105                 unsigned nr, size;
1106         } cur = { 0, 0 }, best = { 0, 0 };
1107
1108         for_each_member_device_rcu(ca, c, i, devs)
1109                 sizes[nr++] = ca->mi.bucket_size;
1110
1111         sort(sizes, nr, sizeof(unsigned), unsigned_cmp, NULL);
1112
1113         for (i = 0; i < nr; i++) {
1114                 if (sizes[i] != cur.size) {
1115                         if (cur.nr > best.nr)
1116                                 best = cur;
1117
1118                         cur.nr = 0;
1119                         cur.size = sizes[i];
1120                 }
1121
1122                 cur.nr++;
1123         }
1124
1125         if (cur.nr > best.nr)
1126                 best = cur;
1127
1128         return best.size;
1129 }
1130
1131 static bool may_create_new_stripe(struct bch_fs *c)
1132 {
1133         return false;
1134 }
1135
1136 static void ec_stripe_key_init(struct bch_fs *c,
1137                                struct bkey_i_stripe *s,
1138                                unsigned nr_data,
1139                                unsigned nr_parity,
1140                                unsigned stripe_size)
1141 {
1142         unsigned u64s;
1143
1144         bkey_stripe_init(&s->k_i);
1145         s->v.sectors                    = cpu_to_le16(stripe_size);
1146         s->v.algorithm                  = 0;
1147         s->v.nr_blocks                  = nr_data + nr_parity;
1148         s->v.nr_redundant               = nr_parity;
1149         s->v.csum_granularity_bits      = ilog2(c->sb.encoded_extent_max);
1150         s->v.csum_type                  = BCH_CSUM_CRC32C;
1151         s->v.pad                        = 0;
1152
1153         while ((u64s = stripe_val_u64s(&s->v)) > BKEY_VAL_U64s_MAX) {
1154                 BUG_ON(1 << s->v.csum_granularity_bits >=
1155                        le16_to_cpu(s->v.sectors) ||
1156                        s->v.csum_granularity_bits == U8_MAX);
1157                 s->v.csum_granularity_bits++;
1158         }
1159
1160         set_bkey_val_u64s(&s->k, u64s);
1161 }
1162
1163 static int ec_new_stripe_alloc(struct bch_fs *c, struct ec_stripe_head *h)
1164 {
1165         struct ec_stripe_new *s;
1166
1167         lockdep_assert_held(&h->lock);
1168
1169         s = kzalloc(sizeof(*s), GFP_KERNEL);
1170         if (!s)
1171                 return -ENOMEM;
1172
1173         mutex_init(&s->lock);
1174         closure_init(&s->iodone, NULL);
1175         atomic_set(&s->pin, 1);
1176         s->c            = c;
1177         s->h            = h;
1178         s->nr_data      = min_t(unsigned, h->nr_active_devs,
1179                                 BCH_BKEY_PTRS_MAX) - h->redundancy;
1180         s->nr_parity    = h->redundancy;
1181
1182         bch2_keylist_init(&s->keys, s->inline_keys);
1183
1184         ec_stripe_key_init(c, &s->new_stripe.key, s->nr_data,
1185                            s->nr_parity, h->blocksize);
1186
1187         h->s = s;
1188         return 0;
1189 }
1190
1191 static struct ec_stripe_head *
1192 ec_new_stripe_head_alloc(struct bch_fs *c, unsigned target,
1193                          unsigned algo, unsigned redundancy,
1194                          bool copygc)
1195 {
1196         struct ec_stripe_head *h;
1197         struct bch_dev *ca;
1198         unsigned i;
1199
1200         h = kzalloc(sizeof(*h), GFP_KERNEL);
1201         if (!h)
1202                 return NULL;
1203
1204         mutex_init(&h->lock);
1205         mutex_lock(&h->lock);
1206
1207         h->target       = target;
1208         h->algo         = algo;
1209         h->redundancy   = redundancy;
1210         h->copygc       = copygc;
1211
1212         rcu_read_lock();
1213         h->devs = target_rw_devs(c, BCH_DATA_user, target);
1214
1215         for_each_member_device_rcu(ca, c, i, &h->devs)
1216                 if (!ca->mi.durability)
1217                         __clear_bit(i, h->devs.d);
1218
1219         h->blocksize = pick_blocksize(c, &h->devs);
1220
1221         for_each_member_device_rcu(ca, c, i, &h->devs)
1222                 if (ca->mi.bucket_size == h->blocksize)
1223                         h->nr_active_devs++;
1224
1225         rcu_read_unlock();
1226         list_add(&h->list, &c->ec_stripe_head_list);
1227         return h;
1228 }
1229
1230 void bch2_ec_stripe_head_put(struct bch_fs *c, struct ec_stripe_head *h)
1231 {
1232         if (h->s &&
1233             h->s->allocated &&
1234             bitmap_weight(h->s->blocks_allocated,
1235                           h->s->nr_data) == h->s->nr_data)
1236                 ec_stripe_set_pending(c, h);
1237
1238         mutex_unlock(&h->lock);
1239 }
1240
1241 struct ec_stripe_head *__bch2_ec_stripe_head_get(struct bch_fs *c,
1242                                                  unsigned target,
1243                                                  unsigned algo,
1244                                                  unsigned redundancy,
1245                                                  bool copygc)
1246 {
1247         struct ec_stripe_head *h;
1248
1249         if (!redundancy)
1250                 return NULL;
1251
1252         mutex_lock(&c->ec_stripe_head_lock);
1253         list_for_each_entry(h, &c->ec_stripe_head_list, list)
1254                 if (h->target           == target &&
1255                     h->algo             == algo &&
1256                     h->redundancy       == redundancy &&
1257                     h->copygc           == copygc) {
1258                         mutex_lock(&h->lock);
1259                         goto found;
1260                 }
1261
1262         h = ec_new_stripe_head_alloc(c, target, algo, redundancy, copygc);
1263 found:
1264         mutex_unlock(&c->ec_stripe_head_lock);
1265         return h;
1266 }
1267
1268 static enum bucket_alloc_ret
1269 new_stripe_alloc_buckets(struct bch_fs *c, struct ec_stripe_head *h,
1270                          struct closure *cl)
1271 {
1272         struct bch_devs_mask devs = h->devs;
1273         struct open_bucket *ob;
1274         struct open_buckets buckets;
1275         unsigned i, j, nr_have_parity = 0, nr_have_data = 0;
1276         bool have_cache = true;
1277         enum bucket_alloc_ret ret = ALLOC_SUCCESS;
1278
1279         for (i = 0; i < h->s->new_stripe.key.v.nr_blocks; i++) {
1280                 if (test_bit(i, h->s->blocks_gotten)) {
1281                         __clear_bit(h->s->new_stripe.key.v.ptrs[i].dev, devs.d);
1282                         if (i < h->s->nr_data)
1283                                 nr_have_data++;
1284                         else
1285                                 nr_have_parity++;
1286                 }
1287         }
1288
1289         BUG_ON(nr_have_data     > h->s->nr_data);
1290         BUG_ON(nr_have_parity   > h->s->nr_parity);
1291
1292         percpu_down_read(&c->mark_lock);
1293         rcu_read_lock();
1294
1295         buckets.nr = 0;
1296         if (nr_have_parity < h->s->nr_parity) {
1297                 ret = bch2_bucket_alloc_set(c, &buckets,
1298                                             &h->parity_stripe,
1299                                             &devs,
1300                                             h->s->nr_parity,
1301                                             &nr_have_parity,
1302                                             &have_cache,
1303                                             h->copygc
1304                                             ? RESERVE_MOVINGGC
1305                                             : RESERVE_NONE,
1306                                             0,
1307                                             cl);
1308
1309                 open_bucket_for_each(c, &buckets, ob, i) {
1310                         j = find_next_zero_bit(h->s->blocks_gotten,
1311                                                h->s->nr_data + h->s->nr_parity,
1312                                                h->s->nr_data);
1313                         BUG_ON(j >= h->s->nr_data + h->s->nr_parity);
1314
1315                         h->s->blocks[j] = buckets.v[i];
1316                         h->s->new_stripe.key.v.ptrs[j] = ob->ptr;
1317                         __set_bit(j, h->s->blocks_gotten);
1318                 }
1319
1320                 if (ret)
1321                         goto err;
1322         }
1323
1324         buckets.nr = 0;
1325         if (nr_have_data < h->s->nr_data) {
1326                 ret = bch2_bucket_alloc_set(c, &buckets,
1327                                             &h->block_stripe,
1328                                             &devs,
1329                                             h->s->nr_data,
1330                                             &nr_have_data,
1331                                             &have_cache,
1332                                             h->copygc
1333                                             ? RESERVE_MOVINGGC
1334                                             : RESERVE_NONE,
1335                                             0,
1336                                             cl);
1337
1338                 open_bucket_for_each(c, &buckets, ob, i) {
1339                         j = find_next_zero_bit(h->s->blocks_gotten,
1340                                                h->s->nr_data, 0);
1341                         BUG_ON(j >= h->s->nr_data);
1342
1343                         h->s->blocks[j] = buckets.v[i];
1344                         h->s->new_stripe.key.v.ptrs[j] = ob->ptr;
1345                         __set_bit(j, h->s->blocks_gotten);
1346                 }
1347
1348                 if (ret)
1349                         goto err;
1350         }
1351 err:
1352         rcu_read_unlock();
1353         percpu_up_read(&c->mark_lock);
1354         return ret;
1355 }
1356
1357 /* XXX: doesn't obey target: */
1358 static s64 get_existing_stripe(struct bch_fs *c,
1359                                struct ec_stripe_head *head)
1360 {
1361         ec_stripes_heap *h = &c->ec_stripes_heap;
1362         struct stripe *m;
1363         size_t heap_idx;
1364         u64 stripe_idx;
1365         s64 ret = -1;
1366
1367         if (may_create_new_stripe(c))
1368                 return -1;
1369
1370         spin_lock(&c->ec_stripes_heap_lock);
1371         for (heap_idx = 0; heap_idx < h->used; heap_idx++) {
1372                 /* No blocks worth reusing, stripe will just be deleted: */
1373                 if (!h->data[heap_idx].blocks_nonempty)
1374                         continue;
1375
1376                 stripe_idx = h->data[heap_idx].idx;
1377                 m = genradix_ptr(&c->stripes[0], stripe_idx);
1378
1379                 if (m->algorithm        == head->algo &&
1380                     m->nr_redundant     == head->redundancy &&
1381                     m->sectors          == head->blocksize &&
1382                     m->blocks_nonempty  < m->nr_blocks - m->nr_redundant) {
1383                         bch2_stripes_heap_del(c, m, stripe_idx);
1384                         ret = stripe_idx;
1385                         break;
1386                 }
1387         }
1388         spin_unlock(&c->ec_stripes_heap_lock);
1389         return ret;
1390 }
1391
1392 static int __bch2_ec_stripe_head_reuse(struct bch_fs *c,
1393                                                    struct ec_stripe_head *h)
1394 {
1395         unsigned i;
1396         s64 idx;
1397         int ret;
1398
1399         idx = get_existing_stripe(c, h);
1400         if (idx < 0) {
1401                 bch_err(c, "failed to find an existing stripe");
1402                 return -ENOSPC;
1403         }
1404
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                 return ret;
1410         }
1411
1412         if (ec_stripe_buf_init(&h->s->existing_stripe, 0, h->blocksize)) {
1413                 /*
1414                  * this is a problem: we have deleted from the
1415                  * stripes heap already
1416                  */
1417                 BUG();
1418         }
1419
1420         BUG_ON(h->s->existing_stripe.size != h->blocksize);
1421         BUG_ON(h->s->existing_stripe.size != h->s->existing_stripe.key.v.sectors);
1422
1423         for (i = 0; i < h->s->existing_stripe.key.v.nr_blocks; i++) {
1424                 if (stripe_blockcount_get(&h->s->existing_stripe.key.v, i)) {
1425                         __set_bit(i, h->s->blocks_gotten);
1426                         __set_bit(i, h->s->blocks_allocated);
1427                 }
1428
1429                 ec_block_io(c, &h->s->existing_stripe, READ, i, &h->s->iodone);
1430         }
1431
1432         bkey_copy(&h->s->new_stripe.key.k_i,
1433                         &h->s->existing_stripe.key.k_i);
1434
1435         return 0;
1436 }
1437
1438 static int __bch2_ec_stripe_head_reserve(struct bch_fs *c,
1439                                                         struct ec_stripe_head *h)
1440 {
1441         int ret;
1442
1443         ret = bch2_disk_reservation_get(c, &h->s->res,
1444                         h->blocksize,
1445                         h->s->nr_parity, 0);
1446
1447         if (ret) {
1448                 /*
1449                  * This means we need to wait for copygc to
1450                  * empty out buckets from existing stripes:
1451                  */
1452                 bch_err(c, "failed to reserve stripe");
1453         }
1454
1455         return ret;
1456 }
1457
1458 struct ec_stripe_head *bch2_ec_stripe_head_get(struct bch_fs *c,
1459                                                unsigned target,
1460                                                unsigned algo,
1461                                                unsigned redundancy,
1462                                                bool copygc,
1463                                                struct closure *cl)
1464 {
1465         struct ec_stripe_head *h;
1466         int ret;
1467         bool needs_stripe_new;
1468
1469         h = __bch2_ec_stripe_head_get(c, target, algo, redundancy, copygc);
1470         if (!h) {
1471                 bch_err(c, "no stripe head");
1472                 return NULL;
1473         }
1474
1475         needs_stripe_new = !h->s;
1476         if (needs_stripe_new) {
1477                 if (ec_new_stripe_alloc(c, h)) {
1478                         ret = -ENOMEM;
1479                         bch_err(c, "failed to allocate new stripe");
1480                         goto err;
1481                 }
1482
1483                 if (ec_stripe_buf_init(&h->s->new_stripe, 0, h->blocksize))
1484                         BUG();
1485         }
1486
1487         /*
1488          * Try reserve a new stripe before reusing an
1489          * existing stripe. This will prevent unnecessary
1490          * read amplification during write oriented workloads.
1491          */
1492         ret = 0;
1493         if (!h->s->allocated && !h->s->res.sectors && !h->s->have_existing_stripe)
1494                 ret = __bch2_ec_stripe_head_reserve(c, h);
1495         if (ret && needs_stripe_new)
1496                 ret = __bch2_ec_stripe_head_reuse(c, h);
1497         if (ret)
1498                 goto err;
1499
1500         if (!h->s->allocated) {
1501                 ret = new_stripe_alloc_buckets(c, h, cl);
1502                 if (ret)
1503                         goto err;
1504
1505                 h->s->allocated = true;
1506         }
1507
1508         return h;
1509
1510 err:
1511         bch2_ec_stripe_head_put(c, h);
1512         return ERR_PTR(-ret);
1513 }
1514
1515 void bch2_ec_stop_dev(struct bch_fs *c, struct bch_dev *ca)
1516 {
1517         struct ec_stripe_head *h;
1518         struct open_bucket *ob;
1519         unsigned i;
1520
1521         mutex_lock(&c->ec_stripe_head_lock);
1522         list_for_each_entry(h, &c->ec_stripe_head_list, list) {
1523
1524                 mutex_lock(&h->lock);
1525                 if (!h->s)
1526                         goto unlock;
1527
1528                 for (i = 0; i < h->s->new_stripe.key.v.nr_blocks; i++) {
1529                         if (!h->s->blocks[i])
1530                                 continue;
1531
1532                         ob = c->open_buckets + h->s->blocks[i];
1533                         if (ob->ptr.dev == ca->dev_idx)
1534                                 goto found;
1535                 }
1536                 goto unlock;
1537 found:
1538                 h->s->err = -EROFS;
1539                 ec_stripe_set_pending(c, h);
1540 unlock:
1541                 mutex_unlock(&h->lock);
1542         }
1543         mutex_unlock(&c->ec_stripe_head_lock);
1544 }
1545
1546 void bch2_stripes_heap_start(struct bch_fs *c)
1547 {
1548         struct genradix_iter iter;
1549         struct stripe *m;
1550
1551         genradix_for_each(&c->stripes[0], iter, m)
1552                 if (m->alive)
1553                         bch2_stripes_heap_insert(c, m, iter.pos);
1554 }
1555
1556 static int __bch2_stripe_write_key(struct btree_trans *trans,
1557                                    struct btree_iter *iter,
1558                                    struct stripe *m,
1559                                    size_t idx,
1560                                    struct bkey_i_stripe *new_key)
1561 {
1562         const struct bch_stripe *v;
1563         struct bkey_s_c k;
1564         unsigned i;
1565         int ret;
1566
1567         bch2_btree_iter_set_pos(iter, POS(0, idx));
1568
1569         k = bch2_btree_iter_peek_slot(iter);
1570         ret = bkey_err(k);
1571         if (ret)
1572                 return ret;
1573
1574         if (k.k->type != KEY_TYPE_stripe)
1575                 return -EIO;
1576
1577         v = bkey_s_c_to_stripe(k).v;
1578         for (i = 0; i < v->nr_blocks; i++)
1579                 if (m->block_sectors[i] != stripe_blockcount_get(v, i))
1580                         goto write;
1581         return 0;
1582 write:
1583         bkey_reassemble(&new_key->k_i, k);
1584
1585         for (i = 0; i < new_key->v.nr_blocks; i++)
1586                 stripe_blockcount_set(&new_key->v, i,
1587                                       m->block_sectors[i]);
1588
1589         return bch2_trans_update(trans, iter, &new_key->k_i, 0);
1590 }
1591
1592 int bch2_stripes_write(struct bch_fs *c, unsigned flags)
1593 {
1594         struct btree_trans trans;
1595         struct btree_iter *iter;
1596         struct genradix_iter giter;
1597         struct bkey_i_stripe *new_key;
1598         struct stripe *m;
1599         int ret = 0;
1600
1601         new_key = kmalloc(255 * sizeof(u64), GFP_KERNEL);
1602         BUG_ON(!new_key);
1603
1604         bch2_trans_init(&trans, c, 0, 0);
1605
1606         iter = bch2_trans_get_iter(&trans, BTREE_ID_stripes, POS_MIN,
1607                                    BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
1608
1609         genradix_for_each(&c->stripes[0], giter, m) {
1610                 if (!m->alive)
1611                         continue;
1612
1613                 ret = __bch2_trans_do(&trans, NULL, NULL,
1614                                       BTREE_INSERT_NOFAIL|flags,
1615                         __bch2_stripe_write_key(&trans, iter, m,
1616                                         giter.pos, new_key));
1617
1618                 if (ret)
1619                         break;
1620         }
1621         bch2_trans_iter_put(&trans, iter);
1622
1623         bch2_trans_exit(&trans);
1624
1625         kfree(new_key);
1626
1627         return ret;
1628 }
1629
1630 static int bch2_stripes_read_fn(struct bch_fs *c, struct bkey_s_c k)
1631 {
1632         int ret = 0;
1633
1634         if (k.k->type == KEY_TYPE_stripe)
1635                 ret = __ec_stripe_mem_alloc(c, k.k->p.offset, GFP_KERNEL) ?:
1636                         bch2_mark_key(c, k, 0, 0, NULL, 0,
1637                                       BTREE_TRIGGER_NOATOMIC);
1638
1639         return ret;
1640 }
1641
1642 int bch2_stripes_read(struct bch_fs *c)
1643 {
1644         int ret = bch2_btree_and_journal_walk(c, BTREE_ID_stripes,
1645                                               bch2_stripes_read_fn);
1646         if (ret)
1647                 bch_err(c, "error reading stripes: %i", ret);
1648
1649         return ret;
1650 }
1651
1652 int bch2_ec_mem_alloc(struct bch_fs *c, bool gc)
1653 {
1654         struct btree_trans trans;
1655         struct btree_iter *iter;
1656         struct bkey_s_c k;
1657         size_t i, idx = 0;
1658         int ret = 0;
1659
1660         bch2_trans_init(&trans, c, 0, 0);
1661         iter = bch2_trans_get_iter(&trans, BTREE_ID_stripes, POS(0, U64_MAX), 0);
1662
1663         k = bch2_btree_iter_prev(iter);
1664         if (!IS_ERR_OR_NULL(k.k))
1665                 idx = k.k->p.offset + 1;
1666
1667         bch2_trans_iter_put(&trans, iter);
1668         ret = bch2_trans_exit(&trans);
1669         if (ret)
1670                 return ret;
1671
1672         if (!idx)
1673                 return 0;
1674
1675         if (!gc &&
1676             !init_heap(&c->ec_stripes_heap, roundup_pow_of_two(idx),
1677                        GFP_KERNEL))
1678                 return -ENOMEM;
1679 #if 0
1680         ret = genradix_prealloc(&c->stripes[gc], idx, GFP_KERNEL);
1681 #else
1682         for (i = 0; i < idx; i++)
1683                 if (!genradix_ptr_alloc(&c->stripes[gc], i, GFP_KERNEL))
1684                         return -ENOMEM;
1685 #endif
1686         return 0;
1687 }
1688
1689 void bch2_stripes_heap_to_text(struct printbuf *out, struct bch_fs *c)
1690 {
1691         ec_stripes_heap *h = &c->ec_stripes_heap;
1692         struct stripe *m;
1693         size_t i;
1694
1695         spin_lock(&c->ec_stripes_heap_lock);
1696         for (i = 0; i < min_t(size_t, h->used, 20); i++) {
1697                 m = genradix_ptr(&c->stripes[0], h->data[i].idx);
1698
1699                 pr_buf(out, "%zu %u/%u+%u\n", h->data[i].idx,
1700                        h->data[i].blocks_nonempty,
1701                        m->nr_blocks - m->nr_redundant,
1702                        m->nr_redundant);
1703         }
1704         spin_unlock(&c->ec_stripes_heap_lock);
1705 }
1706
1707 void bch2_new_stripes_to_text(struct printbuf *out, struct bch_fs *c)
1708 {
1709         struct ec_stripe_head *h;
1710         struct ec_stripe_new *s;
1711
1712         mutex_lock(&c->ec_stripe_head_lock);
1713         list_for_each_entry(h, &c->ec_stripe_head_list, list) {
1714                 pr_buf(out, "target %u algo %u redundancy %u:\n",
1715                        h->target, h->algo, h->redundancy);
1716
1717                 if (h->s)
1718                         pr_buf(out, "\tpending: blocks %u+%u allocated %u\n",
1719                                h->s->nr_data, h->s->nr_parity,
1720                                bitmap_weight(h->s->blocks_allocated,
1721                                              h->s->nr_data));
1722         }
1723         mutex_unlock(&c->ec_stripe_head_lock);
1724
1725         mutex_lock(&c->ec_stripe_new_lock);
1726         list_for_each_entry(s, &c->ec_stripe_new_list, list) {
1727                 pr_buf(out, "\tin flight: blocks %u+%u pin %u\n",
1728                        s->nr_data, s->nr_parity,
1729                        atomic_read(&s->pin));
1730         }
1731         mutex_unlock(&c->ec_stripe_new_lock);
1732 }
1733
1734 void bch2_fs_ec_exit(struct bch_fs *c)
1735 {
1736         struct ec_stripe_head *h;
1737
1738         while (1) {
1739                 mutex_lock(&c->ec_stripe_head_lock);
1740                 h = list_first_entry_or_null(&c->ec_stripe_head_list,
1741                                              struct ec_stripe_head, list);
1742                 if (h)
1743                         list_del(&h->list);
1744                 mutex_unlock(&c->ec_stripe_head_lock);
1745                 if (!h)
1746                         break;
1747
1748                 BUG_ON(h->s);
1749                 kfree(h);
1750         }
1751
1752         BUG_ON(!list_empty(&c->ec_stripe_new_list));
1753
1754         free_heap(&c->ec_stripes_heap);
1755         genradix_free(&c->stripes[0]);
1756         bioset_exit(&c->ec_bioset);
1757 }
1758
1759 int bch2_fs_ec_init(struct bch_fs *c)
1760 {
1761         INIT_WORK(&c->ec_stripe_create_work, ec_stripe_create_work);
1762         INIT_WORK(&c->ec_stripe_delete_work, ec_stripe_delete_work);
1763
1764         return bioset_init(&c->ec_bioset, 1, offsetof(struct ec_bio, bio),
1765                            BIOSET_NEED_BVECS);
1766 }