1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Some low level IO code, and hacks for various block layer limitations
4 *
5 * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
6 * Copyright 2012 Google, Inc.
7 */
8
9 #include "bcachefs.h"
10 #include "alloc_background.h"
11 #include "alloc_foreground.h"
12 #include "btree_update.h"
13 #include "buckets.h"
14 #include "checksum.h"
15 #include "clock.h"
16 #include "compress.h"
17 #include "data_update.h"
18 #include "disk_groups.h"
19 #include "ec.h"
20 #include "error.h"
21 #include "io_read.h"
22 #include "io_misc.h"
23 #include "io_write.h"
24 #include "subvolume.h"
25 #include "trace.h"
26
27 #include <linux/sched/mm.h>
28
29 #ifndef CONFIG_BCACHEFS_NO_LATENCY_ACCT
30
31 static bool bch2_target_congested(struct bch_fs *c, u16 target)
32 {
33 const struct bch_devs_mask *devs;
34 unsigned d, nr = 0, total = 0;
35 u64 now = local_clock(), last;
36 s64 congested;
37 struct bch_dev *ca;
38
39 if (!target)
40 return false;
41
42 rcu_read_lock();
43 devs = bch2_target_to_mask(c, target) ?:
44 &c->rw_devs[BCH_DATA_user];
45
46 for_each_set_bit(d, devs->d, BCH_SB_MEMBERS_MAX) {
47 ca = rcu_dereference(c->devs[d]);
48 if (!ca)
49 continue;
50
51 congested = atomic_read(&ca->congested);
52 last = READ_ONCE(ca->congested_last);
53 if (time_after64(now, last))
54 congested -= (now - last) >> 12;
55
56 total += max(congested, 0LL);
57 nr++;
58 }
59 rcu_read_unlock();
60
61 return bch2_rand_range(nr * CONGESTED_MAX) < total;
62 }
63
64 #else
65
66 static bool bch2_target_congested(struct bch_fs *c, u16 target)
67 {
68 return false;
69 }
70
71 #endif
72
73 /* Cache promotion on read */
74
75 struct promote_op {
76 struct rcu_head rcu;
77 u64 start_time;
78
79 struct rhash_head hash;
80 struct bpos pos;
81
82 struct data_update write;
83 struct bio_vec bi_inline_vecs[]; /* must be last */
84 };
85
86 static const struct rhashtable_params bch_promote_params = {
87 .head_offset = offsetof(struct promote_op, hash),
88 .key_offset = offsetof(struct promote_op, pos),
89 .key_len = sizeof(struct bpos),
90 .automatic_shrinking = true,
91 };
92
93 static inline int should_promote(struct bch_fs *c, struct bkey_s_c k,
94 struct bpos pos,
95 struct bch_io_opts opts,
96 unsigned flags,
97 struct bch_io_failures *failed)
98 {
99 if (!failed) {
100 BUG_ON(!opts.promote_target);
101
102 if (!(flags & BCH_READ_MAY_PROMOTE))
103 return -BCH_ERR_nopromote_may_not;
104
105 if (bch2_bkey_has_target(c, k, opts.promote_target))
106 return -BCH_ERR_nopromote_already_promoted;
107
108 if (bkey_extent_is_unwritten(k))
109 return -BCH_ERR_nopromote_unwritten;
110
111 if (bch2_target_congested(c, opts.promote_target))
112 return -BCH_ERR_nopromote_congested;
113 }
114
115 if (rhashtable_lookup_fast(&c->promote_table, &pos,
116 bch_promote_params))
117 return -BCH_ERR_nopromote_in_flight;
118
119 return 0;
120 }
121
122 static void promote_free(struct bch_fs *c, struct promote_op *op)
123 {
124 int ret;
125
126 bch2_data_update_exit(&op->write);
127
128 ret = rhashtable_remove_fast(&c->promote_table, &op->hash,
129 bch_promote_params);
130 BUG_ON(ret);
131 bch2_write_ref_put(c, BCH_WRITE_REF_promote);
132 kfree_rcu(op, rcu);
133 }
134
135 static void promote_done(struct bch_write_op *wop)
136 {
137 struct promote_op *op =
138 container_of(wop, struct promote_op, write.op);
139 struct bch_fs *c = op->write.op.c;
140
141 bch2_time_stats_update(&c->times[BCH_TIME_data_promote],
142 op->start_time);
143 promote_free(c, op);
144 }
145
146 static void promote_start(struct promote_op *op, struct bch_read_bio *rbio)
147 {
148 struct bio *bio = &op->write.op.wbio.bio;
149
150 trace_and_count(op->write.op.c, read_promote, &rbio->bio);
151
152 /* we now own pages: */
153 BUG_ON(!rbio->bounce);
154 BUG_ON(rbio->bio.bi_vcnt > bio->bi_max_vecs);
155
156 memcpy(bio->bi_io_vec, rbio->bio.bi_io_vec,
157 sizeof(struct bio_vec) * rbio->bio.bi_vcnt);
158 swap(bio->bi_vcnt, rbio->bio.bi_vcnt);
159
160 bch2_data_update_read_done(&op->write, rbio->pick.crc);
161 }
162
163 static struct promote_op *__promote_alloc(struct btree_trans *trans,
164 enum btree_id btree_id,
165 struct bkey_s_c k,
166 struct bpos pos,
167 struct extent_ptr_decoded *pick,
168 struct bch_io_opts opts,
169 unsigned sectors,
170 struct bch_read_bio **rbio,
171 struct bch_io_failures *failed)
172 {
173 struct bch_fs *c = trans->c;
174 struct promote_op *op = NULL;
175 struct bio *bio;
176 unsigned pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
177 int ret;
178
179 if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_promote))
180 return ERR_PTR(-BCH_ERR_nopromote_no_writes);
181
182 op = kzalloc(struct_size(op, bi_inline_vecs, pages), GFP_KERNEL);
183 if (!op) {
184 ret = -BCH_ERR_nopromote_enomem;
185 goto err;
186 }
187
188 op->start_time = local_clock();
189 op->pos = pos;
190
191 /*
192 * We don't use the mempool here because extents that aren't
193 * checksummed or compressed can be too big for the mempool:
194 */
195 *rbio = kzalloc(sizeof(struct bch_read_bio) +
196 sizeof(struct bio_vec) * pages,
197 GFP_KERNEL);
198 if (!*rbio) {
199 ret = -BCH_ERR_nopromote_enomem;
200 goto err;
201 }
202
203 rbio_init(&(*rbio)->bio, opts);
204 bio_init(&(*rbio)->bio, NULL, (*rbio)->bio.bi_inline_vecs, pages, 0);
205
206 if (bch2_bio_alloc_pages(&(*rbio)->bio, sectors << 9, GFP_KERNEL)) {
207 ret = -BCH_ERR_nopromote_enomem;
208 goto err;
209 }
210
211 (*rbio)->bounce = true;
212 (*rbio)->split = true;
213 (*rbio)->kmalloc = true;
214
215 if (rhashtable_lookup_insert_fast(&c->promote_table, &op->hash,
216 bch_promote_params)) {
217 ret = -BCH_ERR_nopromote_in_flight;
218 goto err;
219 }
220
221 bio = &op->write.op.wbio.bio;
222 bio_init(bio, NULL, bio->bi_inline_vecs, pages, 0);
223
224 struct data_update_opts update_opts = {};
225
226 if (!failed) {
227 update_opts.target = opts.promote_target;
228 update_opts.extra_replicas = 1;
229 update_opts.write_flags = BCH_WRITE_ALLOC_NOWAIT|BCH_WRITE_CACHED;
230 } else {
231 update_opts.target = opts.foreground_target;
232
233 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
234 unsigned i = 0;
235 bkey_for_each_ptr(ptrs, ptr) {
236 if (bch2_dev_io_failures(failed, ptr->dev))
237 update_opts.rewrite_ptrs |= BIT(i);
238 i++;
239 }
240 }
241
242 ret = bch2_data_update_init(trans, NULL, NULL, &op->write,
243 writepoint_hashed((unsigned long) current),
244 opts,
245 update_opts,
246 btree_id, k);
247 /*
248 * possible errors: -BCH_ERR_nocow_lock_blocked,
249 * -BCH_ERR_ENOSPC_disk_reservation:
250 */
251 if (ret) {
252 BUG_ON(rhashtable_remove_fast(&c->promote_table, &op->hash,
253 bch_promote_params));
254 goto err;
255 }
256
257 op->write.op.end_io = promote_done;
258
259 return op;
260 err:
261 if (*rbio)
262 bio_free_pages(&(*rbio)->bio);
263 kfree(*rbio);
264 *rbio = NULL;
265 kfree(op);
266 bch2_write_ref_put(c, BCH_WRITE_REF_promote);
267 return ERR_PTR(ret);
268 }
269
270 noinline
271 static struct promote_op *promote_alloc(struct btree_trans *trans,
272 struct bvec_iter iter,
273 struct bkey_s_c k,
274 struct extent_ptr_decoded *pick,
275 struct bch_io_opts opts,
276 unsigned flags,
277 struct bch_read_bio **rbio,
278 bool *bounce,
279 bool *read_full,
280 struct bch_io_failures *failed)
281 {
282 struct bch_fs *c = trans->c;
283 /*
284 * if failed != NULL we're not actually doing a promote, we're
285 * recovering from an io/checksum error
286 */
287 bool promote_full = (failed ||
288 *read_full ||
289 READ_ONCE(c->promote_whole_extents));
290 /* data might have to be decompressed in the write path: */
291 unsigned sectors = promote_full
292 ? max(pick->crc.compressed_size, pick->crc.live_size)
293 : bvec_iter_sectors(iter);
294 struct bpos pos = promote_full
295 ? bkey_start_pos(k.k)
296 : POS(k.k->p.inode, iter.bi_sector);
297 struct promote_op *promote;
298 int ret;
299
300 ret = should_promote(c, k, pos, opts, flags, failed);
301 if (ret)
302 goto nopromote;
303
304 promote = __promote_alloc(trans,
305 k.k->type == KEY_TYPE_reflink_v
306 ? BTREE_ID_reflink
307 : BTREE_ID_extents,
308 k, pos, pick, opts, sectors, rbio, failed);
309 ret = PTR_ERR_OR_ZERO(promote);
310 if (ret)
311 goto nopromote;
312
313 *bounce = true;
314 *read_full = promote_full;
315 return promote;
316 nopromote:
317 trace_read_nopromote(c, ret);
318 return NULL;
319 }
320
321 /* Read */
322
323 #define READ_RETRY_AVOID 1
324 #define READ_RETRY 2
325 #define READ_ERR 3
326
327 enum rbio_context {
328 RBIO_CONTEXT_NULL,
329 RBIO_CONTEXT_HIGHPRI,
330 RBIO_CONTEXT_UNBOUND,
331 };
332
333 static inline struct bch_read_bio *
334 bch2_rbio_parent(struct bch_read_bio *rbio)
335 {
336 return rbio->split ? rbio->parent : rbio;
337 }
338
339 __always_inline
340 static void bch2_rbio_punt(struct bch_read_bio *rbio, work_func_t fn,
341 enum rbio_context context,
342 struct workqueue_struct *wq)
343 {
344 if (context <= rbio->context) {
345 fn(&rbio->work);
346 } else {
347 rbio->work.func = fn;
348 rbio->context = context;
349 queue_work(wq, &rbio->work);
350 }
351 }
352
353 static inline struct bch_read_bio *bch2_rbio_free(struct bch_read_bio *rbio)
354 {
355 BUG_ON(rbio->bounce && !rbio->split);
356
357 if (rbio->promote)
358 promote_free(rbio->c, rbio->promote);
359 rbio->promote = NULL;
360
361 if (rbio->bounce)
362 bch2_bio_free_pages_pool(rbio->c, &rbio->bio);
363
364 if (rbio->split) {
365 struct bch_read_bio *parent = rbio->parent;
366
367 if (rbio->kmalloc)
368 kfree(rbio);
369 else
370 bio_put(&rbio->bio);
371
372 rbio = parent;
373 }
374
375 return rbio;
376 }
377
378 /*
379 * Only called on a top level bch_read_bio to complete an entire read request,
380 * not a split:
381 */
382 static void bch2_rbio_done(struct bch_read_bio *rbio)
383 {
384 if (rbio->start_time)
385 bch2_time_stats_update(&rbio->c->times[BCH_TIME_data_read],
386 rbio->start_time);
387 bio_endio(&rbio->bio);
388 }
389
390 static void bch2_read_retry_nodecode(struct bch_fs *c, struct bch_read_bio *rbio,
391 struct bvec_iter bvec_iter,
392 struct bch_io_failures *failed,
393 unsigned flags)
394 {
395 struct btree_trans *trans = bch2_trans_get(c);
396 struct btree_iter iter;
397 struct bkey_buf sk;
398 struct bkey_s_c k;
399 int ret;
400
401 flags &= ~BCH_READ_LAST_FRAGMENT;
402 flags |= BCH_READ_MUST_CLONE;
403
404 bch2_bkey_buf_init(&sk);
405
406 bch2_trans_iter_init(trans, &iter, rbio->data_btree,
407 rbio->read_pos, BTREE_ITER_slots);
408 retry:
409 bch2_trans_begin(trans);
410 rbio->bio.bi_status = 0;
411
412 k = bch2_btree_iter_peek_slot(&iter);
413 if (bkey_err(k))
414 goto err;
415
416 bch2_bkey_buf_reassemble(&sk, c, k);
417 k = bkey_i_to_s_c(sk.k);
418
419 if (!bch2_bkey_matches_ptr(c, k,
420 rbio->pick.ptr,
421 rbio->data_pos.offset -
422 rbio->pick.crc.offset)) {
423 /* extent we wanted to read no longer exists: */
424 rbio->hole = true;
425 goto out;
426 }
427
428 ret = __bch2_read_extent(trans, rbio, bvec_iter,
429 rbio->read_pos,
430 rbio->data_btree,
431 k, 0, failed, flags);
432 if (ret == READ_RETRY)
433 goto retry;
434 if (ret)
435 goto err;
436 out:
437 bch2_rbio_done(rbio);
438 bch2_trans_iter_exit(trans, &iter);
439 bch2_trans_put(trans);
440 bch2_bkey_buf_exit(&sk, c);
441 return;
442 err:
443 rbio->bio.bi_status = BLK_STS_IOERR;
444 goto out;
445 }
446
447 static void bch2_rbio_retry(struct work_struct *work)
448 {
449 struct bch_read_bio *rbio =
450 container_of(work, struct bch_read_bio, work);
451 struct bch_fs *c = rbio->c;
452 struct bvec_iter iter = rbio->bvec_iter;
453 unsigned flags = rbio->flags;
454 subvol_inum inum = {
455 .subvol = rbio->subvol,
456 .inum = rbio->read_pos.inode,
457 };
458 struct bch_io_failures failed = { .nr = 0 };
459
460 trace_and_count(c, read_retry, &rbio->bio);
461
462 if (rbio->retry == READ_RETRY_AVOID)
463 bch2_mark_io_failure(&failed, &rbio->pick);
464
465 rbio->bio.bi_status = 0;
466
467 rbio = bch2_rbio_free(rbio);
468
469 flags |= BCH_READ_IN_RETRY;
470 flags &= ~BCH_READ_MAY_PROMOTE;
471
472 if (flags & BCH_READ_NODECODE) {
473 bch2_read_retry_nodecode(c, rbio, iter, &failed, flags);
474 } else {
475 flags &= ~BCH_READ_LAST_FRAGMENT;
476 flags |= BCH_READ_MUST_CLONE;
477
478 __bch2_read(c, rbio, iter, inum, &failed, flags);
479 }
480 }
481
482 static void bch2_rbio_error(struct bch_read_bio *rbio, int retry,
483 blk_status_t error)
484 {
485 rbio->retry = retry;
486
487 if (rbio->flags & BCH_READ_IN_RETRY)
488 return;
489
490 if (retry == READ_ERR) {
491 rbio = bch2_rbio_free(rbio);
492
493 rbio->bio.bi_status = error;
494 bch2_rbio_done(rbio);
495 } else {
496 bch2_rbio_punt(rbio, bch2_rbio_retry,
497 RBIO_CONTEXT_UNBOUND, system_unbound_wq);
498 }
499 }
500
501 static int __bch2_rbio_narrow_crcs(struct btree_trans *trans,
502 struct bch_read_bio *rbio)
503 {
504 struct bch_fs *c = rbio->c;
505 u64 data_offset = rbio->data_pos.offset - rbio->pick.crc.offset;
506 struct bch_extent_crc_unpacked new_crc;
507 struct btree_iter iter;
508 struct bkey_i *new;
509 struct bkey_s_c k;
510 int ret = 0;
511
512 if (crc_is_compressed(rbio->pick.crc))
513 return 0;
514
515 k = bch2_bkey_get_iter(trans, &iter, rbio->data_btree, rbio->data_pos,
516 BTREE_ITER_slots|BTREE_ITER_intent);
517 if ((ret = bkey_err(k)))
518 goto out;
519
520 if (bversion_cmp(k.k->version, rbio->version) ||
521 !bch2_bkey_matches_ptr(c, k, rbio->pick.ptr, data_offset))
522 goto out;
523
524 /* Extent was merged? */
525 if (bkey_start_offset(k.k) < data_offset ||
526 k.k->p.offset > data_offset + rbio->pick.crc.uncompressed_size)
527 goto out;
528
529 if (bch2_rechecksum_bio(c, &rbio->bio, rbio->version,
530 rbio->pick.crc, NULL, &new_crc,
531 bkey_start_offset(k.k) - data_offset, k.k->size,
532 rbio->pick.crc.csum_type)) {
533 bch_err(c, "error verifying existing checksum while narrowing checksum (memory corruption?)");
534 ret = 0;
535 goto out;
536 }
537
538 /*
539 * going to be temporarily appending another checksum entry:
540 */
541 new = bch2_trans_kmalloc(trans, bkey_bytes(k.k) +
542 sizeof(struct bch_extent_crc128));
543 if ((ret = PTR_ERR_OR_ZERO(new)))
544 goto out;
545
546 bkey_reassemble(new, k);
547
548 if (!bch2_bkey_narrow_crcs(new, new_crc))
549 goto out;
550
551 ret = bch2_trans_update(trans, &iter, new,
552 BTREE_UPDATE_internal_snapshot_node);
553 out:
554 bch2_trans_iter_exit(trans, &iter);
555 return ret;
556 }
557
558 static noinline void bch2_rbio_narrow_crcs(struct bch_read_bio *rbio)
559 {
560 bch2_trans_do(rbio->c, NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
561 __bch2_rbio_narrow_crcs(trans, rbio));
562 }
563
564 /* Inner part that may run in process context */
565 static void __bch2_read_endio(struct work_struct *work)
566 {
567 struct bch_read_bio *rbio =
568 container_of(work, struct bch_read_bio, work);
569 struct bch_fs *c = rbio->c;
570 struct bio *src = &rbio->bio;
571 struct bio *dst = &bch2_rbio_parent(rbio)->bio;
572 struct bvec_iter dst_iter = rbio->bvec_iter;
573 struct bch_extent_crc_unpacked crc = rbio->pick.crc;
574 struct nonce nonce = extent_nonce(rbio->version, crc);
575 unsigned nofs_flags;
576 struct bch_csum csum;
577 int ret;
578
579 nofs_flags = memalloc_nofs_save();
580
581 /* Reset iterator for checksumming and copying bounced data: */
582 if (rbio->bounce) {
583 src->bi_iter.bi_size = crc.compressed_size << 9;
584 src->bi_iter.bi_idx = 0;
585 src->bi_iter.bi_bvec_done = 0;
586 } else {
587 src->bi_iter = rbio->bvec_iter;
588 }
589
590 csum = bch2_checksum_bio(c, crc.csum_type, nonce, src);
591 if (bch2_crc_cmp(csum, rbio->pick.crc.csum) && !c->opts.no_data_io)
592 goto csum_err;
593
594 /*
595 * XXX
596 * We need to rework the narrow_crcs path to deliver the read completion
597 * first, and then punt to a different workqueue, otherwise we're
598 * holding up reads while doing btree updates which is bad for memory
599 * reclaim.
600 */
601 if (unlikely(rbio->narrow_crcs))
602 bch2_rbio_narrow_crcs(rbio);
603
604 if (rbio->flags & BCH_READ_NODECODE)
605 goto nodecode;
606
607 /* Adjust crc to point to subset of data we want: */
608 crc.offset += rbio->offset_into_extent;
609 crc.live_size = bvec_iter_sectors(rbio->bvec_iter);
610
611 if (crc_is_compressed(crc)) {
612 ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
613 if (ret)
614 goto decrypt_err;
615
616 if (bch2_bio_uncompress(c, src, dst, dst_iter, crc) &&
617 !c->opts.no_data_io)
618 goto decompression_err;
619 } else {
620 /* don't need to decrypt the entire bio: */
621 nonce = nonce_add(nonce, crc.offset << 9);
622 bio_advance(src, crc.offset << 9);
623
624 BUG_ON(src->bi_iter.bi_size < dst_iter.bi_size);
625 src->bi_iter.bi_size = dst_iter.bi_size;
626
627 ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
628 if (ret)
629 goto decrypt_err;
630
631 if (rbio->bounce) {
632 struct bvec_iter src_iter = src->bi_iter;
633
634 bio_copy_data_iter(dst, &dst_iter, src, &src_iter);
635 }
636 }
637
638 if (rbio->promote) {
639 /*
640 * Re encrypt data we decrypted, so it's consistent with
641 * rbio->crc:
642 */
643 ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
644 if (ret)
645 goto decrypt_err;
646
647 promote_start(rbio->promote, rbio);
648 rbio->promote = NULL;
649 }
650 nodecode:
651 if (likely(!(rbio->flags & BCH_READ_IN_RETRY))) {
652 rbio = bch2_rbio_free(rbio);
653 bch2_rbio_done(rbio);
654 }
655 out:
656 memalloc_nofs_restore(nofs_flags);
657 return;
658 csum_err:
659 /*
660 * Checksum error: if the bio wasn't bounced, we may have been
661 * reading into buffers owned by userspace (that userspace can
662 * scribble over) - retry the read, bouncing it this time:
663 */
664 if (!rbio->bounce && (rbio->flags & BCH_READ_USER_MAPPED)) {
665 rbio->flags |= BCH_READ_MUST_BOUNCE;
666 bch2_rbio_error(rbio, READ_RETRY, BLK_STS_IOERR);
667 goto out;
668 }
669
670 struct printbuf buf = PRINTBUF;
671 buf.atomic++;
672 prt_str(&buf, "data ");
673 bch2_csum_err_msg(&buf, crc.csum_type, rbio->pick.crc.csum, csum);
674
675 struct bch_dev *ca = rbio->have_ioref ? bch2_dev_have_ref(c, rbio->pick.ptr.dev) : NULL;
676 if (ca) {
677 bch_err_inum_offset_ratelimited(ca,
678 rbio->read_pos.inode,
679 rbio->read_pos.offset << 9,
680 "data %s", buf.buf);
681 bch2_io_error(ca, BCH_MEMBER_ERROR_checksum);
682 }
683 printbuf_exit(&buf);
684 bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
685 goto out;
686 decompression_err:
687 bch_err_inum_offset_ratelimited(c, rbio->read_pos.inode,
688 rbio->read_pos.offset << 9,
689 "decompression error");
690 bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
691 goto out;
692 decrypt_err:
693 bch_err_inum_offset_ratelimited(c, rbio->read_pos.inode,
694 rbio->read_pos.offset << 9,
695 "decrypt error");
696 bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
697 goto out;
698 }
699
700 static void bch2_read_endio(struct bio *bio)
701 {
702 struct bch_read_bio *rbio =
703 container_of(bio, struct bch_read_bio, bio);
704 struct bch_fs *c = rbio->c;
705 struct bch_dev *ca = rbio->have_ioref ? bch2_dev_have_ref(c, rbio->pick.ptr.dev) : NULL;
706 struct workqueue_struct *wq = NULL;
707 enum rbio_context context = RBIO_CONTEXT_NULL;
708
709 if (rbio->have_ioref) {
710 bch2_latency_acct(ca, rbio->submit_time, READ);
711 percpu_ref_put(&ca->io_ref);
712 }
713
714 if (!rbio->split)
715 rbio->bio.bi_end_io = rbio->end_io;
716
717 if (bio->bi_status) {
718 if (ca) {
719 bch_err_inum_offset_ratelimited(ca,
720 rbio->read_pos.inode,
721 rbio->read_pos.offset,
722 "data read error: %s",
723 bch2_blk_status_to_str(bio->bi_status));
724 bch2_io_error(ca, BCH_MEMBER_ERROR_read);
725 }
726 bch2_rbio_error(rbio, READ_RETRY_AVOID, bio->bi_status);
727 return;
728 }
729
730 if (((rbio->flags & BCH_READ_RETRY_IF_STALE) && race_fault()) ||
731 (ca && dev_ptr_stale(ca, &rbio->pick.ptr))) {
732 trace_and_count(c, read_reuse_race, &rbio->bio);
733
734 if (rbio->flags & BCH_READ_RETRY_IF_STALE)
735 bch2_rbio_error(rbio, READ_RETRY, BLK_STS_AGAIN);
736 else
737 bch2_rbio_error(rbio, READ_ERR, BLK_STS_AGAIN);
738 return;
739 }
740
741 if (rbio->narrow_crcs ||
742 rbio->promote ||
743 crc_is_compressed(rbio->pick.crc) ||
744 bch2_csum_type_is_encryption(rbio->pick.crc.csum_type))
745 context = RBIO_CONTEXT_UNBOUND, wq = system_unbound_wq;
746 else if (rbio->pick.crc.csum_type)
747 context = RBIO_CONTEXT_HIGHPRI, wq = system_highpri_wq;
748
749 bch2_rbio_punt(rbio, __bch2_read_endio, context, wq);
750 }
751
752 int __bch2_read_indirect_extent(struct btree_trans *trans,
753 unsigned *offset_into_extent,
754 struct bkey_buf *orig_k)
755 {
756 struct btree_iter iter;
757 struct bkey_s_c k;
758 u64 reflink_offset;
759 int ret;
760
761 reflink_offset = le64_to_cpu(bkey_i_to_reflink_p(orig_k->k)->v.idx) +
762 *offset_into_extent;
763
764 k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_reflink,
765 POS(0, reflink_offset), 0);
766 ret = bkey_err(k);
767 if (ret)
768 goto err;
769
770 if (k.k->type != KEY_TYPE_reflink_v &&
771 k.k->type != KEY_TYPE_indirect_inline_data) {
772 bch_err_inum_offset_ratelimited(trans->c,
773 orig_k->k->k.p.inode,
774 orig_k->k->k.p.offset << 9,
775 "%llu len %u points to nonexistent indirect extent %llu",
776 orig_k->k->k.p.offset,
777 orig_k->k->k.size,
778 reflink_offset);
779 bch2_inconsistent_error(trans->c);
780 ret = -EIO;
781 goto err;
782 }
783
784 *offset_into_extent = iter.pos.offset - bkey_start_offset(k.k);
785 bch2_bkey_buf_reassemble(orig_k, trans->c, k);
786 err:
787 bch2_trans_iter_exit(trans, &iter);
788 return ret;
789 }
790
791 static noinline void read_from_stale_dirty_pointer(struct btree_trans *trans,
792 struct bch_dev *ca,
793 struct bkey_s_c k,
794 struct bch_extent_ptr ptr)
795 {
796 struct bch_fs *c = trans->c;
797 struct btree_iter iter;
798 struct printbuf buf = PRINTBUF;
799 int ret;
800
801 bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
802 PTR_BUCKET_POS(ca, &ptr),
803 BTREE_ITER_cached);
804
805 u8 *gen = bucket_gen(ca, iter.pos.offset);
806 if (gen) {
807
808 prt_printf(&buf, "Attempting to read from stale dirty pointer:\n");
809 printbuf_indent_add(&buf, 2);
810
811 bch2_bkey_val_to_text(&buf, c, k);
812 prt_newline(&buf);
813
814 prt_printf(&buf, "memory gen: %u", *gen);
815
816 ret = lockrestart_do(trans, bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
817 if (!ret) {
818 prt_newline(&buf);
819 bch2_bkey_val_to_text(&buf, c, k);
820 }
821 } else {
822 prt_printf(&buf, "Attempting to read from invalid bucket %llu:%llu:\n",
823 iter.pos.inode, iter.pos.offset);
824 printbuf_indent_add(&buf, 2);
825
826 prt_printf(&buf, "first bucket %u nbuckets %llu\n",
827 ca->mi.first_bucket, ca->mi.nbuckets);
828
829 bch2_bkey_val_to_text(&buf, c, k);
830 prt_newline(&buf);
831 }
832
833 bch2_fs_inconsistent(c, "%s", buf.buf);
834
835 bch2_trans_iter_exit(trans, &iter);
836 printbuf_exit(&buf);
837 }
838
839 int __bch2_read_extent(struct btree_trans *trans, struct bch_read_bio *orig,
840 struct bvec_iter iter, struct bpos read_pos,
841 enum btree_id data_btree, struct bkey_s_c k,
842 unsigned offset_into_extent,
843 struct bch_io_failures *failed, unsigned flags)
844 {
845 struct bch_fs *c = trans->c;
846 struct extent_ptr_decoded pick;
847 struct bch_read_bio *rbio = NULL;
848 struct promote_op *promote = NULL;
849 bool bounce = false, read_full = false, narrow_crcs = false;
850 struct bpos data_pos = bkey_start_pos(k.k);
851 int pick_ret;
852
853 if (bkey_extent_is_inline_data(k.k)) {
854 unsigned bytes = min_t(unsigned, iter.bi_size,
855 bkey_inline_data_bytes(k.k));
856
857 swap(iter.bi_size, bytes);
858 memcpy_to_bio(&orig->bio, iter, bkey_inline_data_p(k));
859 swap(iter.bi_size, bytes);
860 bio_advance_iter(&orig->bio, &iter, bytes);
861 zero_fill_bio_iter(&orig->bio, iter);
862 goto out_read_done;
863 }
864 retry_pick:
865 pick_ret = bch2_bkey_pick_read_device(c, k, failed, &pick);
866
867 /* hole or reservation - just zero fill: */
868 if (!pick_ret)
869 goto hole;
870
871 if (pick_ret < 0) {
872 bch_err_inum_offset_ratelimited(c,
873 read_pos.inode, read_pos.offset << 9,
874 "no device to read from");
875 goto err;
876 }
877
878 struct bch_dev *ca = bch2_dev_get_ioref(c, pick.ptr.dev, READ);
879
880 /*
881 * Stale dirty pointers are treated as IO errors, but @failed isn't
882 * allocated unless we're in the retry path - so if we're not in the
883 * retry path, don't check here, it'll be caught in bch2_read_endio()
884 * and we'll end up in the retry path:
885 */
886 if ((flags & BCH_READ_IN_RETRY) &&
887 !pick.ptr.cached &&
888 ca &&
889 unlikely(dev_ptr_stale(ca, &pick.ptr))) {
890 read_from_stale_dirty_pointer(trans, ca, k, pick.ptr);
891 bch2_mark_io_failure(failed, &pick);
892 percpu_ref_put(&ca->io_ref);
893 goto retry_pick;
894 }
895
896 /*
897 * Unlock the iterator while the btree node's lock is still in
898 * cache, before doing the IO:
899 */
900 bch2_trans_unlock(trans);
901
902 if (flags & BCH_READ_NODECODE) {
903 /*
904 * can happen if we retry, and the extent we were going to read
905 * has been merged in the meantime:
906 */
907 if (pick.crc.compressed_size > orig->bio.bi_vcnt * PAGE_SECTORS) {
908 if (ca)
909 percpu_ref_put(&ca->io_ref);
910 goto hole;
911 }
912
913 iter.bi_size = pick.crc.compressed_size << 9;
914 goto get_bio;
915 }
916
917 if (!(flags & BCH_READ_LAST_FRAGMENT) ||
918 bio_flagged(&orig->bio, BIO_CHAIN))
919 flags |= BCH_READ_MUST_CLONE;
920
921 narrow_crcs = !(flags & BCH_READ_IN_RETRY) &&
922 bch2_can_narrow_extent_crcs(k, pick.crc);
923
924 if (narrow_crcs && (flags & BCH_READ_USER_MAPPED))
925 flags |= BCH_READ_MUST_BOUNCE;
926
927 EBUG_ON(offset_into_extent + bvec_iter_sectors(iter) > k.k->size);
928
929 if (crc_is_compressed(pick.crc) ||
930 (pick.crc.csum_type != BCH_CSUM_none &&
931 (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
932 (bch2_csum_type_is_encryption(pick.crc.csum_type) &&
933 (flags & BCH_READ_USER_MAPPED)) ||
934 (flags & BCH_READ_MUST_BOUNCE)))) {
935 read_full = true;
936 bounce = true;
937 }
938
939 if (orig->opts.promote_target)// || failed)
940 promote = promote_alloc(trans, iter, k, &pick, orig->opts, flags,
941 &rbio, &bounce, &read_full, failed);
942
943 if (!read_full) {
944 EBUG_ON(crc_is_compressed(pick.crc));
945 EBUG_ON(pick.crc.csum_type &&
946 (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
947 bvec_iter_sectors(iter) != pick.crc.live_size ||
948 pick.crc.offset ||
949 offset_into_extent));
950
951 data_pos.offset += offset_into_extent;
952 pick.ptr.offset += pick.crc.offset +
953 offset_into_extent;
954 offset_into_extent = 0;
955 pick.crc.compressed_size = bvec_iter_sectors(iter);
956 pick.crc.uncompressed_size = bvec_iter_sectors(iter);
957 pick.crc.offset = 0;
958 pick.crc.live_size = bvec_iter_sectors(iter);
959 }
960 get_bio:
961 if (rbio) {
962 /*
963 * promote already allocated bounce rbio:
964 * promote needs to allocate a bio big enough for uncompressing
965 * data in the write path, but we're not going to use it all
966 * here:
967 */
968 EBUG_ON(rbio->bio.bi_iter.bi_size <
969 pick.crc.compressed_size << 9);
970 rbio->bio.bi_iter.bi_size =
971 pick.crc.compressed_size << 9;
972 } else if (bounce) {
973 unsigned sectors = pick.crc.compressed_size;
974
975 rbio = rbio_init(bio_alloc_bioset(NULL,
976 DIV_ROUND_UP(sectors, PAGE_SECTORS),
977 0,
978 GFP_NOFS,
979 &c->bio_read_split),
980 orig->opts);
981
982 bch2_bio_alloc_pages_pool(c, &rbio->bio, sectors << 9);
983 rbio->bounce = true;
984 rbio->split = true;
985 } else if (flags & BCH_READ_MUST_CLONE) {
986 /*
987 * Have to clone if there were any splits, due to error
988 * reporting issues (if a split errored, and retrying didn't
989 * work, when it reports the error to its parent (us) we don't
990 * know if the error was from our bio, and we should retry, or
991 * from the whole bio, in which case we don't want to retry and
992 * lose the error)
993 */
994 rbio = rbio_init(bio_alloc_clone(NULL, &orig->bio, GFP_NOFS,
995 &c->bio_read_split),
996 orig->opts);
997 rbio->bio.bi_iter = iter;
998 rbio->split = true;
999 } else {
1000 rbio = orig;
1001 rbio->bio.bi_iter = iter;
1002 EBUG_ON(bio_flagged(&rbio->bio, BIO_CHAIN));
1003 }
1004
1005 EBUG_ON(bio_sectors(&rbio->bio) != pick.crc.compressed_size);
1006
1007 rbio->c = c;
1008 rbio->submit_time = local_clock();
1009 if (rbio->split)
1010 rbio->parent = orig;
1011 else
1012 rbio->end_io = orig->bio.bi_end_io;
1013 rbio->bvec_iter = iter;
1014 rbio->offset_into_extent= offset_into_extent;
1015 rbio->flags = flags;
1016 rbio->have_ioref = ca != NULL;
1017 rbio->narrow_crcs = narrow_crcs;
1018 rbio->hole = 0;
1019 rbio->retry = 0;
1020 rbio->context = 0;
1021 /* XXX: only initialize this if needed */
1022 rbio->devs_have = bch2_bkey_devs(k);
1023 rbio->pick = pick;
1024 rbio->subvol = orig->subvol;
1025 rbio->read_pos = read_pos;
1026 rbio->data_btree = data_btree;
1027 rbio->data_pos = data_pos;
1028 rbio->version = k.k->version;
1029 rbio->promote = promote;
1030 INIT_WORK(&rbio->work, NULL);
1031
1032 if (flags & BCH_READ_NODECODE)
1033 orig->pick = pick;
1034
1035 rbio->bio.bi_opf = orig->bio.bi_opf;
1036 rbio->bio.bi_iter.bi_sector = pick.ptr.offset;
1037 rbio->bio.bi_end_io = bch2_read_endio;
1038
1039 if (rbio->bounce)
1040 trace_and_count(c, read_bounce, &rbio->bio);
1041
1042 this_cpu_add(c->counters[BCH_COUNTER_io_read], bio_sectors(&rbio->bio));
1043 bch2_increment_clock(c, bio_sectors(&rbio->bio), READ);
1044
1045 /*
1046 * If it's being moved internally, we don't want to flag it as a cache
1047 * hit:
1048 */
1049 if (ca && pick.ptr.cached && !(flags & BCH_READ_NODECODE))
1050 bch2_bucket_io_time_reset(trans, pick.ptr.dev,
1051 PTR_BUCKET_NR(ca, &pick.ptr), READ);
1052
1053 if (!(flags & (BCH_READ_IN_RETRY|BCH_READ_LAST_FRAGMENT))) {
1054 bio_inc_remaining(&orig->bio);
1055 trace_and_count(c, read_split, &orig->bio);
1056 }
1057
1058 if (!rbio->pick.idx) {
1059 if (!rbio->have_ioref) {
1060 bch_err_inum_offset_ratelimited(c,
1061 read_pos.inode,
1062 read_pos.offset << 9,
1063 "no device to read from");
1064 bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
1065 goto out;
1066 }
1067
1068 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_user],
1069 bio_sectors(&rbio->bio));
1070 bio_set_dev(&rbio->bio, ca->disk_sb.bdev);
1071
1072 if (unlikely(c->opts.no_data_io)) {
1073 if (likely(!(flags & BCH_READ_IN_RETRY)))
1074 bio_endio(&rbio->bio);
1075 } else {
1076 if (likely(!(flags & BCH_READ_IN_RETRY)))
1077 submit_bio(&rbio->bio);
1078 else
1079 submit_bio_wait(&rbio->bio);
1080 }
1081
1082 /*
1083 * We just submitted IO which may block, we expect relock fail
1084 * events and shouldn't count them:
1085 */
1086 trans->notrace_relock_fail = true;
1087 } else {
1088 /* Attempting reconstruct read: */
1089 if (bch2_ec_read_extent(trans, rbio)) {
1090 bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
1091 goto out;
1092 }
1093
1094 if (likely(!(flags & BCH_READ_IN_RETRY)))
1095 bio_endio(&rbio->bio);
1096 }
1097 out:
1098 if (likely(!(flags & BCH_READ_IN_RETRY))) {
1099 return 0;
1100 } else {
1101 int ret;
1102
1103 rbio->context = RBIO_CONTEXT_UNBOUND;
1104 bch2_read_endio(&rbio->bio);
1105
1106 ret = rbio->retry;
1107 rbio = bch2_rbio_free(rbio);
1108
1109 if (ret == READ_RETRY_AVOID) {
1110 bch2_mark_io_failure(failed, &pick);
1111 ret = READ_RETRY;
1112 }
1113
1114 if (!ret)
1115 goto out_read_done;
1116
1117 return ret;
1118 }
1119
1120 err:
1121 if (flags & BCH_READ_IN_RETRY)
1122 return READ_ERR;
1123
1124 orig->bio.bi_status = BLK_STS_IOERR;
1125 goto out_read_done;
1126
1127 hole:
1128 /*
1129 * won't normally happen in the BCH_READ_NODECODE
1130 * (bch2_move_extent()) path, but if we retry and the extent we wanted
1131 * to read no longer exists we have to signal that:
1132 */
1133 if (flags & BCH_READ_NODECODE)
1134 orig->hole = true;
1135
1136 zero_fill_bio_iter(&orig->bio, iter);
1137 out_read_done:
1138 if (flags & BCH_READ_LAST_FRAGMENT)
1139 bch2_rbio_done(orig);
1140 return 0;
1141 }
1142
1143 void __bch2_read(struct bch_fs *c, struct bch_read_bio *rbio,
1144 struct bvec_iter bvec_iter, subvol_inum inum,
1145 struct bch_io_failures *failed, unsigned flags)
1146 {
1147 struct btree_trans *trans = bch2_trans_get(c);
1148 struct btree_iter iter;
1149 struct bkey_buf sk;
1150 struct bkey_s_c k;
1151 int ret;
1152
1153 BUG_ON(flags & BCH_READ_NODECODE);
1154
1155 bch2_bkey_buf_init(&sk);
1156 bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
1157 POS(inum.inum, bvec_iter.bi_sector),
1158 BTREE_ITER_slots);
1159
1160 while (1) {
1161 unsigned bytes, sectors, offset_into_extent;
1162 enum btree_id data_btree = BTREE_ID_extents;
1163
1164 bch2_trans_begin(trans);
1165
1166 u32 snapshot;
1167 ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
1168 if (ret)
1169 goto err;
1170
1171 bch2_btree_iter_set_snapshot(&iter, snapshot);
1172
1173 bch2_btree_iter_set_pos(&iter,
1174 POS(inum.inum, bvec_iter.bi_sector));
1175
1176 k = bch2_btree_iter_peek_slot(&iter);
1177 ret = bkey_err(k);
1178 if (ret)
1179 goto err;
1180
1181 offset_into_extent = iter.pos.offset -
1182 bkey_start_offset(k.k);
1183 sectors = k.k->size - offset_into_extent;
1184
1185 bch2_bkey_buf_reassemble(&sk, c, k);
1186
1187 ret = bch2_read_indirect_extent(trans, &data_btree,
1188 &offset_into_extent, &sk);
1189 if (ret)
1190 goto err;
1191
1192 k = bkey_i_to_s_c(sk.k);
1193
1194 /*
1195 * With indirect extents, the amount of data to read is the min
1196 * of the original extent and the indirect extent:
1197 */
1198 sectors = min(sectors, k.k->size - offset_into_extent);
1199
1200 bytes = min(sectors, bvec_iter_sectors(bvec_iter)) << 9;
1201 swap(bvec_iter.bi_size, bytes);
1202
1203 if (bvec_iter.bi_size == bytes)
1204 flags |= BCH_READ_LAST_FRAGMENT;
1205
1206 ret = __bch2_read_extent(trans, rbio, bvec_iter, iter.pos,
1207 data_btree, k,
1208 offset_into_extent, failed, flags);
1209 if (ret)
1210 goto err;
1211
1212 if (flags & BCH_READ_LAST_FRAGMENT)
1213 break;
1214
1215 swap(bvec_iter.bi_size, bytes);
1216 bio_advance_iter(&rbio->bio, &bvec_iter, bytes);
1217
1218 ret = btree_trans_too_many_iters(trans);
1219 if (ret)
1220 goto err;
1221 err:
1222 if (ret &&
1223 !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
1224 ret != READ_RETRY &&
1225 ret != READ_RETRY_AVOID)
1226 break;
1227 }
1228
1229 bch2_trans_iter_exit(trans, &iter);
1230 bch2_trans_put(trans);
1231 bch2_bkey_buf_exit(&sk, c);
1232
1233 if (ret) {
1234 bch_err_inum_offset_ratelimited(c, inum.inum,
1235 bvec_iter.bi_sector << 9,
1236 "read error %i from btree lookup", ret);
1237 rbio->bio.bi_status = BLK_STS_IOERR;
1238 bch2_rbio_done(rbio);
1239 }
1240 }
1241
1242 void bch2_fs_io_read_exit(struct bch_fs *c)
1243 {
1244 if (c->promote_table.tbl)
1245 rhashtable_destroy(&c->promote_table);
1246 bioset_exit(&c->bio_read_split);
1247 bioset_exit(&c->bio_read);
1248 }
1249
1250 int bch2_fs_io_read_init(struct bch_fs *c)
1251 {
1252 if (bioset_init(&c->bio_read, 1, offsetof(struct bch_read_bio, bio),
1253 BIOSET_NEED_BVECS))
1254 return -BCH_ERR_ENOMEM_bio_read_init;
1255
1256 if (bioset_init(&c->bio_read_split, 1, offsetof(struct bch_read_bio, bio),
1257 BIOSET_NEED_BVECS))
1258 return -BCH_ERR_ENOMEM_bio_read_split_init;
1259
1260 if (rhashtable_init(&c->promote_table, &bch_promote_params))
1261 return -BCH_ERR_ENOMEM_promote_table_init;
1262
1263 return 0;
1264 }
1265
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.