xref: /linux/fs/bcachefs/io_read.c (revision b4db9f840283caca0d904436f187ef56a9126eaa)
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 };
91 
92 static inline int should_promote(struct bch_fs *c, struct bkey_s_c k,
93 				  struct bpos pos,
94 				  struct bch_io_opts opts,
95 				  unsigned flags)
96 {
97 	BUG_ON(!opts.promote_target);
98 
99 	if (!(flags & BCH_READ_MAY_PROMOTE))
100 		return -BCH_ERR_nopromote_may_not;
101 
102 	if (bch2_bkey_has_target(c, k, opts.promote_target))
103 		return -BCH_ERR_nopromote_already_promoted;
104 
105 	if (bkey_extent_is_unwritten(k))
106 		return -BCH_ERR_nopromote_unwritten;
107 
108 	if (bch2_target_congested(c, opts.promote_target))
109 		return -BCH_ERR_nopromote_congested;
110 
111 	if (rhashtable_lookup_fast(&c->promote_table, &pos,
112 				   bch_promote_params))
113 		return -BCH_ERR_nopromote_in_flight;
114 
115 	return 0;
116 }
117 
118 static void promote_free(struct bch_fs *c, struct promote_op *op)
119 {
120 	int ret;
121 
122 	bch2_data_update_exit(&op->write);
123 
124 	ret = rhashtable_remove_fast(&c->promote_table, &op->hash,
125 				     bch_promote_params);
126 	BUG_ON(ret);
127 	bch2_write_ref_put(c, BCH_WRITE_REF_promote);
128 	kfree_rcu(op, rcu);
129 }
130 
131 static void promote_done(struct bch_write_op *wop)
132 {
133 	struct promote_op *op =
134 		container_of(wop, struct promote_op, write.op);
135 	struct bch_fs *c = op->write.op.c;
136 
137 	bch2_time_stats_update(&c->times[BCH_TIME_data_promote],
138 			       op->start_time);
139 	promote_free(c, op);
140 }
141 
142 static void promote_start(struct promote_op *op, struct bch_read_bio *rbio)
143 {
144 	struct bio *bio = &op->write.op.wbio.bio;
145 
146 	trace_and_count(op->write.op.c, read_promote, &rbio->bio);
147 
148 	/* we now own pages: */
149 	BUG_ON(!rbio->bounce);
150 	BUG_ON(rbio->bio.bi_vcnt > bio->bi_max_vecs);
151 
152 	memcpy(bio->bi_io_vec, rbio->bio.bi_io_vec,
153 	       sizeof(struct bio_vec) * rbio->bio.bi_vcnt);
154 	swap(bio->bi_vcnt, rbio->bio.bi_vcnt);
155 
156 	bch2_data_update_read_done(&op->write, rbio->pick.crc);
157 }
158 
159 static struct promote_op *__promote_alloc(struct btree_trans *trans,
160 					  enum btree_id btree_id,
161 					  struct bkey_s_c k,
162 					  struct bpos pos,
163 					  struct extent_ptr_decoded *pick,
164 					  struct bch_io_opts opts,
165 					  unsigned sectors,
166 					  struct bch_read_bio **rbio)
167 {
168 	struct bch_fs *c = trans->c;
169 	struct promote_op *op = NULL;
170 	struct bio *bio;
171 	unsigned pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
172 	int ret;
173 
174 	if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_promote))
175 		return ERR_PTR(-BCH_ERR_nopromote_no_writes);
176 
177 	op = kzalloc(struct_size(op, bi_inline_vecs, pages), GFP_KERNEL);
178 	if (!op) {
179 		ret = -BCH_ERR_nopromote_enomem;
180 		goto err;
181 	}
182 
183 	op->start_time = local_clock();
184 	op->pos = pos;
185 
186 	/*
187 	 * We don't use the mempool here because extents that aren't
188 	 * checksummed or compressed can be too big for the mempool:
189 	 */
190 	*rbio = kzalloc(sizeof(struct bch_read_bio) +
191 			sizeof(struct bio_vec) * pages,
192 			GFP_KERNEL);
193 	if (!*rbio) {
194 		ret = -BCH_ERR_nopromote_enomem;
195 		goto err;
196 	}
197 
198 	rbio_init(&(*rbio)->bio, opts);
199 	bio_init(&(*rbio)->bio, NULL, (*rbio)->bio.bi_inline_vecs, pages, 0);
200 
201 	if (bch2_bio_alloc_pages(&(*rbio)->bio, sectors << 9, GFP_KERNEL)) {
202 		ret = -BCH_ERR_nopromote_enomem;
203 		goto err;
204 	}
205 
206 	(*rbio)->bounce		= true;
207 	(*rbio)->split		= true;
208 	(*rbio)->kmalloc	= true;
209 
210 	if (rhashtable_lookup_insert_fast(&c->promote_table, &op->hash,
211 					  bch_promote_params)) {
212 		ret = -BCH_ERR_nopromote_in_flight;
213 		goto err;
214 	}
215 
216 	bio = &op->write.op.wbio.bio;
217 	bio_init(bio, NULL, bio->bi_inline_vecs, pages, 0);
218 
219 	ret = bch2_data_update_init(trans, NULL, NULL, &op->write,
220 			writepoint_hashed((unsigned long) current),
221 			opts,
222 			(struct data_update_opts) {
223 				.target		= opts.promote_target,
224 				.extra_replicas	= 1,
225 				.write_flags	= BCH_WRITE_ALLOC_NOWAIT|BCH_WRITE_CACHED,
226 			},
227 			btree_id, k);
228 	/*
229 	 * possible errors: -BCH_ERR_nocow_lock_blocked,
230 	 * -BCH_ERR_ENOSPC_disk_reservation:
231 	 */
232 	if (ret) {
233 		BUG_ON(rhashtable_remove_fast(&c->promote_table, &op->hash,
234 					      bch_promote_params));
235 		goto err;
236 	}
237 
238 	op->write.op.end_io = promote_done;
239 
240 	return op;
241 err:
242 	if (*rbio)
243 		bio_free_pages(&(*rbio)->bio);
244 	kfree(*rbio);
245 	*rbio = NULL;
246 	kfree(op);
247 	bch2_write_ref_put(c, BCH_WRITE_REF_promote);
248 	return ERR_PTR(ret);
249 }
250 
251 noinline
252 static struct promote_op *promote_alloc(struct btree_trans *trans,
253 					struct bvec_iter iter,
254 					struct bkey_s_c k,
255 					struct extent_ptr_decoded *pick,
256 					struct bch_io_opts opts,
257 					unsigned flags,
258 					struct bch_read_bio **rbio,
259 					bool *bounce,
260 					bool *read_full)
261 {
262 	struct bch_fs *c = trans->c;
263 	bool promote_full = *read_full || READ_ONCE(c->promote_whole_extents);
264 	/* data might have to be decompressed in the write path: */
265 	unsigned sectors = promote_full
266 		? max(pick->crc.compressed_size, pick->crc.live_size)
267 		: bvec_iter_sectors(iter);
268 	struct bpos pos = promote_full
269 		? bkey_start_pos(k.k)
270 		: POS(k.k->p.inode, iter.bi_sector);
271 	struct promote_op *promote;
272 	int ret;
273 
274 	ret = should_promote(c, k, pos, opts, flags);
275 	if (ret)
276 		goto nopromote;
277 
278 	promote = __promote_alloc(trans,
279 				  k.k->type == KEY_TYPE_reflink_v
280 				  ? BTREE_ID_reflink
281 				  : BTREE_ID_extents,
282 				  k, pos, pick, opts, sectors, rbio);
283 	ret = PTR_ERR_OR_ZERO(promote);
284 	if (ret)
285 		goto nopromote;
286 
287 	*bounce		= true;
288 	*read_full	= promote_full;
289 	return promote;
290 nopromote:
291 	trace_read_nopromote(c, ret);
292 	return NULL;
293 }
294 
295 /* Read */
296 
297 #define READ_RETRY_AVOID	1
298 #define READ_RETRY		2
299 #define READ_ERR		3
300 
301 enum rbio_context {
302 	RBIO_CONTEXT_NULL,
303 	RBIO_CONTEXT_HIGHPRI,
304 	RBIO_CONTEXT_UNBOUND,
305 };
306 
307 static inline struct bch_read_bio *
308 bch2_rbio_parent(struct bch_read_bio *rbio)
309 {
310 	return rbio->split ? rbio->parent : rbio;
311 }
312 
313 __always_inline
314 static void bch2_rbio_punt(struct bch_read_bio *rbio, work_func_t fn,
315 			   enum rbio_context context,
316 			   struct workqueue_struct *wq)
317 {
318 	if (context <= rbio->context) {
319 		fn(&rbio->work);
320 	} else {
321 		rbio->work.func		= fn;
322 		rbio->context		= context;
323 		queue_work(wq, &rbio->work);
324 	}
325 }
326 
327 static inline struct bch_read_bio *bch2_rbio_free(struct bch_read_bio *rbio)
328 {
329 	BUG_ON(rbio->bounce && !rbio->split);
330 
331 	if (rbio->promote)
332 		promote_free(rbio->c, rbio->promote);
333 	rbio->promote = NULL;
334 
335 	if (rbio->bounce)
336 		bch2_bio_free_pages_pool(rbio->c, &rbio->bio);
337 
338 	if (rbio->split) {
339 		struct bch_read_bio *parent = rbio->parent;
340 
341 		if (rbio->kmalloc)
342 			kfree(rbio);
343 		else
344 			bio_put(&rbio->bio);
345 
346 		rbio = parent;
347 	}
348 
349 	return rbio;
350 }
351 
352 /*
353  * Only called on a top level bch_read_bio to complete an entire read request,
354  * not a split:
355  */
356 static void bch2_rbio_done(struct bch_read_bio *rbio)
357 {
358 	if (rbio->start_time)
359 		bch2_time_stats_update(&rbio->c->times[BCH_TIME_data_read],
360 				       rbio->start_time);
361 	bio_endio(&rbio->bio);
362 }
363 
364 static void bch2_read_retry_nodecode(struct bch_fs *c, struct bch_read_bio *rbio,
365 				     struct bvec_iter bvec_iter,
366 				     struct bch_io_failures *failed,
367 				     unsigned flags)
368 {
369 	struct btree_trans *trans = bch2_trans_get(c);
370 	struct btree_iter iter;
371 	struct bkey_buf sk;
372 	struct bkey_s_c k;
373 	int ret;
374 
375 	flags &= ~BCH_READ_LAST_FRAGMENT;
376 	flags |= BCH_READ_MUST_CLONE;
377 
378 	bch2_bkey_buf_init(&sk);
379 
380 	bch2_trans_iter_init(trans, &iter, rbio->data_btree,
381 			     rbio->read_pos, BTREE_ITER_SLOTS);
382 retry:
383 	rbio->bio.bi_status = 0;
384 
385 	k = bch2_btree_iter_peek_slot(&iter);
386 	if (bkey_err(k))
387 		goto err;
388 
389 	bch2_bkey_buf_reassemble(&sk, c, k);
390 	k = bkey_i_to_s_c(sk.k);
391 	bch2_trans_unlock(trans);
392 
393 	if (!bch2_bkey_matches_ptr(c, k,
394 				   rbio->pick.ptr,
395 				   rbio->data_pos.offset -
396 				   rbio->pick.crc.offset)) {
397 		/* extent we wanted to read no longer exists: */
398 		rbio->hole = true;
399 		goto out;
400 	}
401 
402 	ret = __bch2_read_extent(trans, rbio, bvec_iter,
403 				 rbio->read_pos,
404 				 rbio->data_btree,
405 				 k, 0, failed, flags);
406 	if (ret == READ_RETRY)
407 		goto retry;
408 	if (ret)
409 		goto err;
410 out:
411 	bch2_rbio_done(rbio);
412 	bch2_trans_iter_exit(trans, &iter);
413 	bch2_trans_put(trans);
414 	bch2_bkey_buf_exit(&sk, c);
415 	return;
416 err:
417 	rbio->bio.bi_status = BLK_STS_IOERR;
418 	goto out;
419 }
420 
421 static void bch2_rbio_retry(struct work_struct *work)
422 {
423 	struct bch_read_bio *rbio =
424 		container_of(work, struct bch_read_bio, work);
425 	struct bch_fs *c	= rbio->c;
426 	struct bvec_iter iter	= rbio->bvec_iter;
427 	unsigned flags		= rbio->flags;
428 	subvol_inum inum = {
429 		.subvol = rbio->subvol,
430 		.inum	= rbio->read_pos.inode,
431 	};
432 	struct bch_io_failures failed = { .nr = 0 };
433 
434 	trace_and_count(c, read_retry, &rbio->bio);
435 
436 	if (rbio->retry == READ_RETRY_AVOID)
437 		bch2_mark_io_failure(&failed, &rbio->pick);
438 
439 	rbio->bio.bi_status = 0;
440 
441 	rbio = bch2_rbio_free(rbio);
442 
443 	flags |= BCH_READ_IN_RETRY;
444 	flags &= ~BCH_READ_MAY_PROMOTE;
445 
446 	if (flags & BCH_READ_NODECODE) {
447 		bch2_read_retry_nodecode(c, rbio, iter, &failed, flags);
448 	} else {
449 		flags &= ~BCH_READ_LAST_FRAGMENT;
450 		flags |= BCH_READ_MUST_CLONE;
451 
452 		__bch2_read(c, rbio, iter, inum, &failed, flags);
453 	}
454 }
455 
456 static void bch2_rbio_error(struct bch_read_bio *rbio, int retry,
457 			    blk_status_t error)
458 {
459 	rbio->retry = retry;
460 
461 	if (rbio->flags & BCH_READ_IN_RETRY)
462 		return;
463 
464 	if (retry == READ_ERR) {
465 		rbio = bch2_rbio_free(rbio);
466 
467 		rbio->bio.bi_status = error;
468 		bch2_rbio_done(rbio);
469 	} else {
470 		bch2_rbio_punt(rbio, bch2_rbio_retry,
471 			       RBIO_CONTEXT_UNBOUND, system_unbound_wq);
472 	}
473 }
474 
475 static int __bch2_rbio_narrow_crcs(struct btree_trans *trans,
476 				   struct bch_read_bio *rbio)
477 {
478 	struct bch_fs *c = rbio->c;
479 	u64 data_offset = rbio->data_pos.offset - rbio->pick.crc.offset;
480 	struct bch_extent_crc_unpacked new_crc;
481 	struct btree_iter iter;
482 	struct bkey_i *new;
483 	struct bkey_s_c k;
484 	int ret = 0;
485 
486 	if (crc_is_compressed(rbio->pick.crc))
487 		return 0;
488 
489 	k = bch2_bkey_get_iter(trans, &iter, rbio->data_btree, rbio->data_pos,
490 			       BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
491 	if ((ret = bkey_err(k)))
492 		goto out;
493 
494 	if (bversion_cmp(k.k->version, rbio->version) ||
495 	    !bch2_bkey_matches_ptr(c, k, rbio->pick.ptr, data_offset))
496 		goto out;
497 
498 	/* Extent was merged? */
499 	if (bkey_start_offset(k.k) < data_offset ||
500 	    k.k->p.offset > data_offset + rbio->pick.crc.uncompressed_size)
501 		goto out;
502 
503 	if (bch2_rechecksum_bio(c, &rbio->bio, rbio->version,
504 			rbio->pick.crc, NULL, &new_crc,
505 			bkey_start_offset(k.k) - data_offset, k.k->size,
506 			rbio->pick.crc.csum_type)) {
507 		bch_err(c, "error verifying existing checksum while narrowing checksum (memory corruption?)");
508 		ret = 0;
509 		goto out;
510 	}
511 
512 	/*
513 	 * going to be temporarily appending another checksum entry:
514 	 */
515 	new = bch2_trans_kmalloc(trans, bkey_bytes(k.k) +
516 				 sizeof(struct bch_extent_crc128));
517 	if ((ret = PTR_ERR_OR_ZERO(new)))
518 		goto out;
519 
520 	bkey_reassemble(new, k);
521 
522 	if (!bch2_bkey_narrow_crcs(new, new_crc))
523 		goto out;
524 
525 	ret = bch2_trans_update(trans, &iter, new,
526 				BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
527 out:
528 	bch2_trans_iter_exit(trans, &iter);
529 	return ret;
530 }
531 
532 static noinline void bch2_rbio_narrow_crcs(struct bch_read_bio *rbio)
533 {
534 	bch2_trans_do(rbio->c, NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
535 		      __bch2_rbio_narrow_crcs(trans, rbio));
536 }
537 
538 /* Inner part that may run in process context */
539 static void __bch2_read_endio(struct work_struct *work)
540 {
541 	struct bch_read_bio *rbio =
542 		container_of(work, struct bch_read_bio, work);
543 	struct bch_fs *c	= rbio->c;
544 	struct bch_dev *ca	= bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
545 	struct bio *src		= &rbio->bio;
546 	struct bio *dst		= &bch2_rbio_parent(rbio)->bio;
547 	struct bvec_iter dst_iter = rbio->bvec_iter;
548 	struct bch_extent_crc_unpacked crc = rbio->pick.crc;
549 	struct nonce nonce = extent_nonce(rbio->version, crc);
550 	unsigned nofs_flags;
551 	struct bch_csum csum;
552 	int ret;
553 
554 	nofs_flags = memalloc_nofs_save();
555 
556 	/* Reset iterator for checksumming and copying bounced data: */
557 	if (rbio->bounce) {
558 		src->bi_iter.bi_size		= crc.compressed_size << 9;
559 		src->bi_iter.bi_idx		= 0;
560 		src->bi_iter.bi_bvec_done	= 0;
561 	} else {
562 		src->bi_iter			= rbio->bvec_iter;
563 	}
564 
565 	csum = bch2_checksum_bio(c, crc.csum_type, nonce, src);
566 	if (bch2_crc_cmp(csum, rbio->pick.crc.csum) && !c->opts.no_data_io)
567 		goto csum_err;
568 
569 	/*
570 	 * XXX
571 	 * We need to rework the narrow_crcs path to deliver the read completion
572 	 * first, and then punt to a different workqueue, otherwise we're
573 	 * holding up reads while doing btree updates which is bad for memory
574 	 * reclaim.
575 	 */
576 	if (unlikely(rbio->narrow_crcs))
577 		bch2_rbio_narrow_crcs(rbio);
578 
579 	if (rbio->flags & BCH_READ_NODECODE)
580 		goto nodecode;
581 
582 	/* Adjust crc to point to subset of data we want: */
583 	crc.offset     += rbio->offset_into_extent;
584 	crc.live_size	= bvec_iter_sectors(rbio->bvec_iter);
585 
586 	if (crc_is_compressed(crc)) {
587 		ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
588 		if (ret)
589 			goto decrypt_err;
590 
591 		if (bch2_bio_uncompress(c, src, dst, dst_iter, crc) &&
592 		    !c->opts.no_data_io)
593 			goto decompression_err;
594 	} else {
595 		/* don't need to decrypt the entire bio: */
596 		nonce = nonce_add(nonce, crc.offset << 9);
597 		bio_advance(src, crc.offset << 9);
598 
599 		BUG_ON(src->bi_iter.bi_size < dst_iter.bi_size);
600 		src->bi_iter.bi_size = dst_iter.bi_size;
601 
602 		ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
603 		if (ret)
604 			goto decrypt_err;
605 
606 		if (rbio->bounce) {
607 			struct bvec_iter src_iter = src->bi_iter;
608 
609 			bio_copy_data_iter(dst, &dst_iter, src, &src_iter);
610 		}
611 	}
612 
613 	if (rbio->promote) {
614 		/*
615 		 * Re encrypt data we decrypted, so it's consistent with
616 		 * rbio->crc:
617 		 */
618 		ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
619 		if (ret)
620 			goto decrypt_err;
621 
622 		promote_start(rbio->promote, rbio);
623 		rbio->promote = NULL;
624 	}
625 nodecode:
626 	if (likely(!(rbio->flags & BCH_READ_IN_RETRY))) {
627 		rbio = bch2_rbio_free(rbio);
628 		bch2_rbio_done(rbio);
629 	}
630 out:
631 	memalloc_nofs_restore(nofs_flags);
632 	return;
633 csum_err:
634 	/*
635 	 * Checksum error: if the bio wasn't bounced, we may have been
636 	 * reading into buffers owned by userspace (that userspace can
637 	 * scribble over) - retry the read, bouncing it this time:
638 	 */
639 	if (!rbio->bounce && (rbio->flags & BCH_READ_USER_MAPPED)) {
640 		rbio->flags |= BCH_READ_MUST_BOUNCE;
641 		bch2_rbio_error(rbio, READ_RETRY, BLK_STS_IOERR);
642 		goto out;
643 	}
644 
645 	struct printbuf buf = PRINTBUF;
646 	buf.atomic++;
647 	prt_str(&buf, "data ");
648 	bch2_csum_err_msg(&buf, crc.csum_type, rbio->pick.crc.csum, csum);
649 
650 	bch_err_inum_offset_ratelimited(ca,
651 		rbio->read_pos.inode,
652 		rbio->read_pos.offset << 9,
653 		"data %s", buf.buf);
654 	printbuf_exit(&buf);
655 
656 	bch2_io_error(ca, BCH_MEMBER_ERROR_checksum);
657 	bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
658 	goto out;
659 decompression_err:
660 	bch_err_inum_offset_ratelimited(c, rbio->read_pos.inode,
661 					rbio->read_pos.offset << 9,
662 					"decompression error");
663 	bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
664 	goto out;
665 decrypt_err:
666 	bch_err_inum_offset_ratelimited(c, rbio->read_pos.inode,
667 					rbio->read_pos.offset << 9,
668 					"decrypt error");
669 	bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
670 	goto out;
671 }
672 
673 static void bch2_read_endio(struct bio *bio)
674 {
675 	struct bch_read_bio *rbio =
676 		container_of(bio, struct bch_read_bio, bio);
677 	struct bch_fs *c	= rbio->c;
678 	struct bch_dev *ca	= bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
679 	struct workqueue_struct *wq = NULL;
680 	enum rbio_context context = RBIO_CONTEXT_NULL;
681 
682 	if (rbio->have_ioref) {
683 		bch2_latency_acct(ca, rbio->submit_time, READ);
684 		percpu_ref_put(&ca->io_ref);
685 	}
686 
687 	if (!rbio->split)
688 		rbio->bio.bi_end_io = rbio->end_io;
689 
690 	if (bch2_dev_inum_io_err_on(bio->bi_status, ca, BCH_MEMBER_ERROR_read,
691 				    rbio->read_pos.inode,
692 				    rbio->read_pos.offset,
693 				    "data read error: %s",
694 			       bch2_blk_status_to_str(bio->bi_status))) {
695 		bch2_rbio_error(rbio, READ_RETRY_AVOID, bio->bi_status);
696 		return;
697 	}
698 
699 	if (((rbio->flags & BCH_READ_RETRY_IF_STALE) && race_fault()) ||
700 	    ptr_stale(ca, &rbio->pick.ptr)) {
701 		trace_and_count(c, read_reuse_race, &rbio->bio);
702 
703 		if (rbio->flags & BCH_READ_RETRY_IF_STALE)
704 			bch2_rbio_error(rbio, READ_RETRY, BLK_STS_AGAIN);
705 		else
706 			bch2_rbio_error(rbio, READ_ERR, BLK_STS_AGAIN);
707 		return;
708 	}
709 
710 	if (rbio->narrow_crcs ||
711 	    rbio->promote ||
712 	    crc_is_compressed(rbio->pick.crc) ||
713 	    bch2_csum_type_is_encryption(rbio->pick.crc.csum_type))
714 		context = RBIO_CONTEXT_UNBOUND,	wq = system_unbound_wq;
715 	else if (rbio->pick.crc.csum_type)
716 		context = RBIO_CONTEXT_HIGHPRI,	wq = system_highpri_wq;
717 
718 	bch2_rbio_punt(rbio, __bch2_read_endio, context, wq);
719 }
720 
721 int __bch2_read_indirect_extent(struct btree_trans *trans,
722 				unsigned *offset_into_extent,
723 				struct bkey_buf *orig_k)
724 {
725 	struct btree_iter iter;
726 	struct bkey_s_c k;
727 	u64 reflink_offset;
728 	int ret;
729 
730 	reflink_offset = le64_to_cpu(bkey_i_to_reflink_p(orig_k->k)->v.idx) +
731 		*offset_into_extent;
732 
733 	k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_reflink,
734 			       POS(0, reflink_offset), 0);
735 	ret = bkey_err(k);
736 	if (ret)
737 		goto err;
738 
739 	if (k.k->type != KEY_TYPE_reflink_v &&
740 	    k.k->type != KEY_TYPE_indirect_inline_data) {
741 		bch_err_inum_offset_ratelimited(trans->c,
742 			orig_k->k->k.p.inode,
743 			orig_k->k->k.p.offset << 9,
744 			"%llu len %u points to nonexistent indirect extent %llu",
745 			orig_k->k->k.p.offset,
746 			orig_k->k->k.size,
747 			reflink_offset);
748 		bch2_inconsistent_error(trans->c);
749 		ret = -EIO;
750 		goto err;
751 	}
752 
753 	*offset_into_extent = iter.pos.offset - bkey_start_offset(k.k);
754 	bch2_bkey_buf_reassemble(orig_k, trans->c, k);
755 err:
756 	bch2_trans_iter_exit(trans, &iter);
757 	return ret;
758 }
759 
760 static noinline void read_from_stale_dirty_pointer(struct btree_trans *trans,
761 						   struct bkey_s_c k,
762 						   struct bch_extent_ptr ptr)
763 {
764 	struct bch_fs *c = trans->c;
765 	struct bch_dev *ca = bch_dev_bkey_exists(c, ptr.dev);
766 	struct btree_iter iter;
767 	struct printbuf buf = PRINTBUF;
768 	int ret;
769 
770 	bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
771 			     PTR_BUCKET_POS(c, &ptr),
772 			     BTREE_ITER_CACHED);
773 
774 	prt_printf(&buf, "Attempting to read from stale dirty pointer:");
775 	printbuf_indent_add(&buf, 2);
776 	prt_newline(&buf);
777 
778 	bch2_bkey_val_to_text(&buf, c, k);
779 	prt_newline(&buf);
780 
781 	prt_printf(&buf, "memory gen: %u", *bucket_gen(ca, iter.pos.offset));
782 
783 	ret = lockrestart_do(trans, bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
784 	if (!ret) {
785 		prt_newline(&buf);
786 		bch2_bkey_val_to_text(&buf, c, k);
787 	}
788 
789 	bch2_fs_inconsistent(c, "%s", buf.buf);
790 
791 	bch2_trans_iter_exit(trans, &iter);
792 	printbuf_exit(&buf);
793 }
794 
795 int __bch2_read_extent(struct btree_trans *trans, struct bch_read_bio *orig,
796 		       struct bvec_iter iter, struct bpos read_pos,
797 		       enum btree_id data_btree, struct bkey_s_c k,
798 		       unsigned offset_into_extent,
799 		       struct bch_io_failures *failed, unsigned flags)
800 {
801 	struct bch_fs *c = trans->c;
802 	struct extent_ptr_decoded pick;
803 	struct bch_read_bio *rbio = NULL;
804 	struct bch_dev *ca = NULL;
805 	struct promote_op *promote = NULL;
806 	bool bounce = false, read_full = false, narrow_crcs = false;
807 	struct bpos data_pos = bkey_start_pos(k.k);
808 	int pick_ret;
809 
810 	if (bkey_extent_is_inline_data(k.k)) {
811 		unsigned bytes = min_t(unsigned, iter.bi_size,
812 				       bkey_inline_data_bytes(k.k));
813 
814 		swap(iter.bi_size, bytes);
815 		memcpy_to_bio(&orig->bio, iter, bkey_inline_data_p(k));
816 		swap(iter.bi_size, bytes);
817 		bio_advance_iter(&orig->bio, &iter, bytes);
818 		zero_fill_bio_iter(&orig->bio, iter);
819 		goto out_read_done;
820 	}
821 retry_pick:
822 	pick_ret = bch2_bkey_pick_read_device(c, k, failed, &pick);
823 
824 	/* hole or reservation - just zero fill: */
825 	if (!pick_ret)
826 		goto hole;
827 
828 	if (pick_ret < 0) {
829 		bch_err_inum_offset_ratelimited(c,
830 				read_pos.inode, read_pos.offset << 9,
831 				"no device to read from");
832 		goto err;
833 	}
834 
835 	ca = bch_dev_bkey_exists(c, pick.ptr.dev);
836 
837 	/*
838 	 * Stale dirty pointers are treated as IO errors, but @failed isn't
839 	 * allocated unless we're in the retry path - so if we're not in the
840 	 * retry path, don't check here, it'll be caught in bch2_read_endio()
841 	 * and we'll end up in the retry path:
842 	 */
843 	if ((flags & BCH_READ_IN_RETRY) &&
844 	    !pick.ptr.cached &&
845 	    unlikely(ptr_stale(ca, &pick.ptr))) {
846 		read_from_stale_dirty_pointer(trans, k, pick.ptr);
847 		bch2_mark_io_failure(failed, &pick);
848 		goto retry_pick;
849 	}
850 
851 	/*
852 	 * Unlock the iterator while the btree node's lock is still in
853 	 * cache, before doing the IO:
854 	 */
855 	bch2_trans_unlock(trans);
856 
857 	if (flags & BCH_READ_NODECODE) {
858 		/*
859 		 * can happen if we retry, and the extent we were going to read
860 		 * has been merged in the meantime:
861 		 */
862 		if (pick.crc.compressed_size > orig->bio.bi_vcnt * PAGE_SECTORS)
863 			goto hole;
864 
865 		iter.bi_size	= pick.crc.compressed_size << 9;
866 		goto get_bio;
867 	}
868 
869 	if (!(flags & BCH_READ_LAST_FRAGMENT) ||
870 	    bio_flagged(&orig->bio, BIO_CHAIN))
871 		flags |= BCH_READ_MUST_CLONE;
872 
873 	narrow_crcs = !(flags & BCH_READ_IN_RETRY) &&
874 		bch2_can_narrow_extent_crcs(k, pick.crc);
875 
876 	if (narrow_crcs && (flags & BCH_READ_USER_MAPPED))
877 		flags |= BCH_READ_MUST_BOUNCE;
878 
879 	EBUG_ON(offset_into_extent + bvec_iter_sectors(iter) > k.k->size);
880 
881 	if (crc_is_compressed(pick.crc) ||
882 	    (pick.crc.csum_type != BCH_CSUM_none &&
883 	     (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
884 	      (bch2_csum_type_is_encryption(pick.crc.csum_type) &&
885 	       (flags & BCH_READ_USER_MAPPED)) ||
886 	      (flags & BCH_READ_MUST_BOUNCE)))) {
887 		read_full = true;
888 		bounce = true;
889 	}
890 
891 	if (orig->opts.promote_target)
892 		promote = promote_alloc(trans, iter, k, &pick, orig->opts, flags,
893 					&rbio, &bounce, &read_full);
894 
895 	if (!read_full) {
896 		EBUG_ON(crc_is_compressed(pick.crc));
897 		EBUG_ON(pick.crc.csum_type &&
898 			(bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
899 			 bvec_iter_sectors(iter) != pick.crc.live_size ||
900 			 pick.crc.offset ||
901 			 offset_into_extent));
902 
903 		data_pos.offset += offset_into_extent;
904 		pick.ptr.offset += pick.crc.offset +
905 			offset_into_extent;
906 		offset_into_extent		= 0;
907 		pick.crc.compressed_size	= bvec_iter_sectors(iter);
908 		pick.crc.uncompressed_size	= bvec_iter_sectors(iter);
909 		pick.crc.offset			= 0;
910 		pick.crc.live_size		= bvec_iter_sectors(iter);
911 	}
912 get_bio:
913 	if (rbio) {
914 		/*
915 		 * promote already allocated bounce rbio:
916 		 * promote needs to allocate a bio big enough for uncompressing
917 		 * data in the write path, but we're not going to use it all
918 		 * here:
919 		 */
920 		EBUG_ON(rbio->bio.bi_iter.bi_size <
921 		       pick.crc.compressed_size << 9);
922 		rbio->bio.bi_iter.bi_size =
923 			pick.crc.compressed_size << 9;
924 	} else if (bounce) {
925 		unsigned sectors = pick.crc.compressed_size;
926 
927 		rbio = rbio_init(bio_alloc_bioset(NULL,
928 						  DIV_ROUND_UP(sectors, PAGE_SECTORS),
929 						  0,
930 						  GFP_NOFS,
931 						  &c->bio_read_split),
932 				 orig->opts);
933 
934 		bch2_bio_alloc_pages_pool(c, &rbio->bio, sectors << 9);
935 		rbio->bounce	= true;
936 		rbio->split	= true;
937 	} else if (flags & BCH_READ_MUST_CLONE) {
938 		/*
939 		 * Have to clone if there were any splits, due to error
940 		 * reporting issues (if a split errored, and retrying didn't
941 		 * work, when it reports the error to its parent (us) we don't
942 		 * know if the error was from our bio, and we should retry, or
943 		 * from the whole bio, in which case we don't want to retry and
944 		 * lose the error)
945 		 */
946 		rbio = rbio_init(bio_alloc_clone(NULL, &orig->bio, GFP_NOFS,
947 						 &c->bio_read_split),
948 				 orig->opts);
949 		rbio->bio.bi_iter = iter;
950 		rbio->split	= true;
951 	} else {
952 		rbio = orig;
953 		rbio->bio.bi_iter = iter;
954 		EBUG_ON(bio_flagged(&rbio->bio, BIO_CHAIN));
955 	}
956 
957 	EBUG_ON(bio_sectors(&rbio->bio) != pick.crc.compressed_size);
958 
959 	rbio->c			= c;
960 	rbio->submit_time	= local_clock();
961 	if (rbio->split)
962 		rbio->parent	= orig;
963 	else
964 		rbio->end_io	= orig->bio.bi_end_io;
965 	rbio->bvec_iter		= iter;
966 	rbio->offset_into_extent= offset_into_extent;
967 	rbio->flags		= flags;
968 	rbio->have_ioref	= pick_ret > 0 && bch2_dev_get_ioref(ca, READ);
969 	rbio->narrow_crcs	= narrow_crcs;
970 	rbio->hole		= 0;
971 	rbio->retry		= 0;
972 	rbio->context		= 0;
973 	/* XXX: only initialize this if needed */
974 	rbio->devs_have		= bch2_bkey_devs(k);
975 	rbio->pick		= pick;
976 	rbio->subvol		= orig->subvol;
977 	rbio->read_pos		= read_pos;
978 	rbio->data_btree	= data_btree;
979 	rbio->data_pos		= data_pos;
980 	rbio->version		= k.k->version;
981 	rbio->promote		= promote;
982 	INIT_WORK(&rbio->work, NULL);
983 
984 	rbio->bio.bi_opf	= orig->bio.bi_opf;
985 	rbio->bio.bi_iter.bi_sector = pick.ptr.offset;
986 	rbio->bio.bi_end_io	= bch2_read_endio;
987 
988 	if (rbio->bounce)
989 		trace_and_count(c, read_bounce, &rbio->bio);
990 
991 	this_cpu_add(c->counters[BCH_COUNTER_io_read], bio_sectors(&rbio->bio));
992 	bch2_increment_clock(c, bio_sectors(&rbio->bio), READ);
993 
994 	/*
995 	 * If it's being moved internally, we don't want to flag it as a cache
996 	 * hit:
997 	 */
998 	if (pick.ptr.cached && !(flags & BCH_READ_NODECODE))
999 		bch2_bucket_io_time_reset(trans, pick.ptr.dev,
1000 			PTR_BUCKET_NR(ca, &pick.ptr), READ);
1001 
1002 	if (!(flags & (BCH_READ_IN_RETRY|BCH_READ_LAST_FRAGMENT))) {
1003 		bio_inc_remaining(&orig->bio);
1004 		trace_and_count(c, read_split, &orig->bio);
1005 	}
1006 
1007 	if (!rbio->pick.idx) {
1008 		if (!rbio->have_ioref) {
1009 			bch_err_inum_offset_ratelimited(c,
1010 					read_pos.inode,
1011 					read_pos.offset << 9,
1012 					"no device to read from");
1013 			bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
1014 			goto out;
1015 		}
1016 
1017 		this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_user],
1018 			     bio_sectors(&rbio->bio));
1019 		bio_set_dev(&rbio->bio, ca->disk_sb.bdev);
1020 
1021 		if (unlikely(c->opts.no_data_io)) {
1022 			if (likely(!(flags & BCH_READ_IN_RETRY)))
1023 				bio_endio(&rbio->bio);
1024 		} else {
1025 			if (likely(!(flags & BCH_READ_IN_RETRY)))
1026 				submit_bio(&rbio->bio);
1027 			else
1028 				submit_bio_wait(&rbio->bio);
1029 		}
1030 
1031 		/*
1032 		 * We just submitted IO which may block, we expect relock fail
1033 		 * events and shouldn't count them:
1034 		 */
1035 		trans->notrace_relock_fail = true;
1036 	} else {
1037 		/* Attempting reconstruct read: */
1038 		if (bch2_ec_read_extent(trans, rbio)) {
1039 			bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
1040 			goto out;
1041 		}
1042 
1043 		if (likely(!(flags & BCH_READ_IN_RETRY)))
1044 			bio_endio(&rbio->bio);
1045 	}
1046 out:
1047 	if (likely(!(flags & BCH_READ_IN_RETRY))) {
1048 		return 0;
1049 	} else {
1050 		int ret;
1051 
1052 		rbio->context = RBIO_CONTEXT_UNBOUND;
1053 		bch2_read_endio(&rbio->bio);
1054 
1055 		ret = rbio->retry;
1056 		rbio = bch2_rbio_free(rbio);
1057 
1058 		if (ret == READ_RETRY_AVOID) {
1059 			bch2_mark_io_failure(failed, &pick);
1060 			ret = READ_RETRY;
1061 		}
1062 
1063 		if (!ret)
1064 			goto out_read_done;
1065 
1066 		return ret;
1067 	}
1068 
1069 err:
1070 	if (flags & BCH_READ_IN_RETRY)
1071 		return READ_ERR;
1072 
1073 	orig->bio.bi_status = BLK_STS_IOERR;
1074 	goto out_read_done;
1075 
1076 hole:
1077 	/*
1078 	 * won't normally happen in the BCH_READ_NODECODE
1079 	 * (bch2_move_extent()) path, but if we retry and the extent we wanted
1080 	 * to read no longer exists we have to signal that:
1081 	 */
1082 	if (flags & BCH_READ_NODECODE)
1083 		orig->hole = true;
1084 
1085 	zero_fill_bio_iter(&orig->bio, iter);
1086 out_read_done:
1087 	if (flags & BCH_READ_LAST_FRAGMENT)
1088 		bch2_rbio_done(orig);
1089 	return 0;
1090 }
1091 
1092 void __bch2_read(struct bch_fs *c, struct bch_read_bio *rbio,
1093 		 struct bvec_iter bvec_iter, subvol_inum inum,
1094 		 struct bch_io_failures *failed, unsigned flags)
1095 {
1096 	struct btree_trans *trans = bch2_trans_get(c);
1097 	struct btree_iter iter;
1098 	struct bkey_buf sk;
1099 	struct bkey_s_c k;
1100 	u32 snapshot;
1101 	int ret;
1102 
1103 	BUG_ON(flags & BCH_READ_NODECODE);
1104 
1105 	bch2_bkey_buf_init(&sk);
1106 retry:
1107 	bch2_trans_begin(trans);
1108 	iter = (struct btree_iter) { NULL };
1109 
1110 	ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
1111 	if (ret)
1112 		goto err;
1113 
1114 	bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
1115 			     SPOS(inum.inum, bvec_iter.bi_sector, snapshot),
1116 			     BTREE_ITER_SLOTS);
1117 	while (1) {
1118 		unsigned bytes, sectors, offset_into_extent;
1119 		enum btree_id data_btree = BTREE_ID_extents;
1120 
1121 		/*
1122 		 * read_extent -> io_time_reset may cause a transaction restart
1123 		 * without returning an error, we need to check for that here:
1124 		 */
1125 		ret = bch2_trans_relock(trans);
1126 		if (ret)
1127 			break;
1128 
1129 		bch2_btree_iter_set_pos(&iter,
1130 				POS(inum.inum, bvec_iter.bi_sector));
1131 
1132 		k = bch2_btree_iter_peek_slot(&iter);
1133 		ret = bkey_err(k);
1134 		if (ret)
1135 			break;
1136 
1137 		offset_into_extent = iter.pos.offset -
1138 			bkey_start_offset(k.k);
1139 		sectors = k.k->size - offset_into_extent;
1140 
1141 		bch2_bkey_buf_reassemble(&sk, c, k);
1142 
1143 		ret = bch2_read_indirect_extent(trans, &data_btree,
1144 					&offset_into_extent, &sk);
1145 		if (ret)
1146 			break;
1147 
1148 		k = bkey_i_to_s_c(sk.k);
1149 
1150 		/*
1151 		 * With indirect extents, the amount of data to read is the min
1152 		 * of the original extent and the indirect extent:
1153 		 */
1154 		sectors = min(sectors, k.k->size - offset_into_extent);
1155 
1156 		bytes = min(sectors, bvec_iter_sectors(bvec_iter)) << 9;
1157 		swap(bvec_iter.bi_size, bytes);
1158 
1159 		if (bvec_iter.bi_size == bytes)
1160 			flags |= BCH_READ_LAST_FRAGMENT;
1161 
1162 		ret = __bch2_read_extent(trans, rbio, bvec_iter, iter.pos,
1163 					 data_btree, k,
1164 					 offset_into_extent, failed, flags);
1165 		if (ret)
1166 			break;
1167 
1168 		if (flags & BCH_READ_LAST_FRAGMENT)
1169 			break;
1170 
1171 		swap(bvec_iter.bi_size, bytes);
1172 		bio_advance_iter(&rbio->bio, &bvec_iter, bytes);
1173 
1174 		ret = btree_trans_too_many_iters(trans);
1175 		if (ret)
1176 			break;
1177 	}
1178 err:
1179 	bch2_trans_iter_exit(trans, &iter);
1180 
1181 	if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
1182 	    ret == READ_RETRY ||
1183 	    ret == READ_RETRY_AVOID)
1184 		goto retry;
1185 
1186 	bch2_trans_put(trans);
1187 	bch2_bkey_buf_exit(&sk, c);
1188 
1189 	if (ret) {
1190 		bch_err_inum_offset_ratelimited(c, inum.inum,
1191 						bvec_iter.bi_sector << 9,
1192 						"read error %i from btree lookup", ret);
1193 		rbio->bio.bi_status = BLK_STS_IOERR;
1194 		bch2_rbio_done(rbio);
1195 	}
1196 }
1197 
1198 void bch2_fs_io_read_exit(struct bch_fs *c)
1199 {
1200 	if (c->promote_table.tbl)
1201 		rhashtable_destroy(&c->promote_table);
1202 	bioset_exit(&c->bio_read_split);
1203 	bioset_exit(&c->bio_read);
1204 }
1205 
1206 int bch2_fs_io_read_init(struct bch_fs *c)
1207 {
1208 	if (bioset_init(&c->bio_read, 1, offsetof(struct bch_read_bio, bio),
1209 			BIOSET_NEED_BVECS))
1210 		return -BCH_ERR_ENOMEM_bio_read_init;
1211 
1212 	if (bioset_init(&c->bio_read_split, 1, offsetof(struct bch_read_bio, bio),
1213 			BIOSET_NEED_BVECS))
1214 		return -BCH_ERR_ENOMEM_bio_read_split_init;
1215 
1216 	if (rhashtable_init(&c->promote_table, &bch_promote_params))
1217 		return -BCH_ERR_ENOMEM_promote_table_init;
1218 
1219 	return 0;
1220 }
1221