xref: /linux/fs/bcachefs/extents.c (revision 289088d54623a1a50bb3ff79f7331bbe501ea591)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com>
4  *
5  * Code for managing the extent btree and dynamically updating the writeback
6  * dirty sector count.
7  */
8 
9 #include "bcachefs.h"
10 #include "bkey_methods.h"
11 #include "btree_cache.h"
12 #include "btree_gc.h"
13 #include "btree_io.h"
14 #include "btree_iter.h"
15 #include "buckets.h"
16 #include "checksum.h"
17 #include "compress.h"
18 #include "debug.h"
19 #include "disk_groups.h"
20 #include "error.h"
21 #include "extents.h"
22 #include "inode.h"
23 #include "journal.h"
24 #include "replicas.h"
25 #include "super.h"
26 #include "super-io.h"
27 #include "trace.h"
28 #include "util.h"
29 
30 static unsigned bch2_crc_field_size_max[] = {
31 	[BCH_EXTENT_ENTRY_crc32] = CRC32_SIZE_MAX,
32 	[BCH_EXTENT_ENTRY_crc64] = CRC64_SIZE_MAX,
33 	[BCH_EXTENT_ENTRY_crc128] = CRC128_SIZE_MAX,
34 };
35 
36 static void bch2_extent_crc_pack(union bch_extent_crc *,
37 				 struct bch_extent_crc_unpacked,
38 				 enum bch_extent_entry_type);
39 
40 struct bch_dev_io_failures *bch2_dev_io_failures(struct bch_io_failures *f,
41 						 unsigned dev)
42 {
43 	struct bch_dev_io_failures *i;
44 
45 	for (i = f->devs; i < f->devs + f->nr; i++)
46 		if (i->dev == dev)
47 			return i;
48 
49 	return NULL;
50 }
51 
52 void bch2_mark_io_failure(struct bch_io_failures *failed,
53 			  struct extent_ptr_decoded *p)
54 {
55 	struct bch_dev_io_failures *f = bch2_dev_io_failures(failed, p->ptr.dev);
56 
57 	if (!f) {
58 		BUG_ON(failed->nr >= ARRAY_SIZE(failed->devs));
59 
60 		f = &failed->devs[failed->nr++];
61 		f->dev		= p->ptr.dev;
62 		f->idx		= p->idx;
63 		f->nr_failed	= 1;
64 		f->nr_retries	= 0;
65 	} else if (p->idx != f->idx) {
66 		f->idx		= p->idx;
67 		f->nr_failed	= 1;
68 		f->nr_retries	= 0;
69 	} else {
70 		f->nr_failed++;
71 	}
72 }
73 
74 static inline u64 dev_latency(struct bch_fs *c, unsigned dev)
75 {
76 	struct bch_dev *ca = bch2_dev_rcu(c, dev);
77 	return ca ? atomic64_read(&ca->cur_latency[READ]) : S64_MAX;
78 }
79 
80 /*
81  * returns true if p1 is better than p2:
82  */
83 static inline bool ptr_better(struct bch_fs *c,
84 			      const struct extent_ptr_decoded p1,
85 			      const struct extent_ptr_decoded p2)
86 {
87 	if (likely(!p1.idx && !p2.idx)) {
88 		u64 l1 = dev_latency(c, p1.ptr.dev);
89 		u64 l2 = dev_latency(c, p2.ptr.dev);
90 
91 		/* Pick at random, biased in favor of the faster device: */
92 
93 		return bch2_rand_range(l1 + l2) > l1;
94 	}
95 
96 	if (bch2_force_reconstruct_read)
97 		return p1.idx > p2.idx;
98 
99 	return p1.idx < p2.idx;
100 }
101 
102 /*
103  * This picks a non-stale pointer, preferably from a device other than @avoid.
104  * Avoid can be NULL, meaning pick any. If there are no non-stale pointers to
105  * other devices, it will still pick a pointer from avoid.
106  */
107 int bch2_bkey_pick_read_device(struct bch_fs *c, struct bkey_s_c k,
108 			       struct bch_io_failures *failed,
109 			       struct extent_ptr_decoded *pick)
110 {
111 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
112 	const union bch_extent_entry *entry;
113 	struct extent_ptr_decoded p;
114 	struct bch_dev_io_failures *f;
115 	int ret = 0;
116 
117 	if (k.k->type == KEY_TYPE_error)
118 		return -EIO;
119 
120 	rcu_read_lock();
121 	bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
122 		/*
123 		 * Unwritten extent: no need to actually read, treat it as a
124 		 * hole and return 0s:
125 		 */
126 		if (p.ptr.unwritten) {
127 			ret = 0;
128 			break;
129 		}
130 
131 		/*
132 		 * If there are any dirty pointers it's an error if we can't
133 		 * read:
134 		 */
135 		if (!ret && !p.ptr.cached)
136 			ret = -EIO;
137 
138 		struct bch_dev *ca = bch2_dev_rcu(c, p.ptr.dev);
139 
140 		if (p.ptr.cached && (!ca || dev_ptr_stale_rcu(ca, &p.ptr)))
141 			continue;
142 
143 		f = failed ? bch2_dev_io_failures(failed, p.ptr.dev) : NULL;
144 		if (f)
145 			p.idx = f->nr_failed < f->nr_retries
146 				? f->idx
147 				: f->idx + 1;
148 
149 		if (!p.idx && !ca)
150 			p.idx++;
151 
152 		if (!p.idx && p.has_ec && bch2_force_reconstruct_read)
153 			p.idx++;
154 
155 		if (!p.idx && !bch2_dev_is_readable(ca))
156 			p.idx++;
157 
158 		if (p.idx >= (unsigned) p.has_ec + 1)
159 			continue;
160 
161 		if (ret > 0 && !ptr_better(c, p, *pick))
162 			continue;
163 
164 		*pick = p;
165 		ret = 1;
166 	}
167 	rcu_read_unlock();
168 
169 	return ret;
170 }
171 
172 /* KEY_TYPE_btree_ptr: */
173 
174 int bch2_btree_ptr_validate(struct bch_fs *c, struct bkey_s_c k,
175 			    enum bch_validate_flags flags)
176 {
177 	int ret = 0;
178 
179 	bkey_fsck_err_on(bkey_val_u64s(k.k) > BCH_REPLICAS_MAX,
180 			 c, btree_ptr_val_too_big,
181 			 "value too big (%zu > %u)", bkey_val_u64s(k.k), BCH_REPLICAS_MAX);
182 
183 	ret = bch2_bkey_ptrs_validate(c, k, flags);
184 fsck_err:
185 	return ret;
186 }
187 
188 void bch2_btree_ptr_to_text(struct printbuf *out, struct bch_fs *c,
189 			    struct bkey_s_c k)
190 {
191 	bch2_bkey_ptrs_to_text(out, c, k);
192 }
193 
194 int bch2_btree_ptr_v2_validate(struct bch_fs *c, struct bkey_s_c k,
195 			       enum bch_validate_flags flags)
196 {
197 	struct bkey_s_c_btree_ptr_v2 bp = bkey_s_c_to_btree_ptr_v2(k);
198 	int ret = 0;
199 
200 	bkey_fsck_err_on(bkey_val_u64s(k.k) > BKEY_BTREE_PTR_VAL_U64s_MAX,
201 			 c, btree_ptr_v2_val_too_big,
202 			 "value too big (%zu > %zu)",
203 			 bkey_val_u64s(k.k), BKEY_BTREE_PTR_VAL_U64s_MAX);
204 
205 	bkey_fsck_err_on(bpos_ge(bp.v->min_key, bp.k->p),
206 			 c, btree_ptr_v2_min_key_bad,
207 			 "min_key > key");
208 
209 	if (flags & BCH_VALIDATE_write)
210 		bkey_fsck_err_on(!bp.v->sectors_written,
211 				 c, btree_ptr_v2_written_0,
212 				 "sectors_written == 0");
213 
214 	ret = bch2_bkey_ptrs_validate(c, k, flags);
215 fsck_err:
216 	return ret;
217 }
218 
219 void bch2_btree_ptr_v2_to_text(struct printbuf *out, struct bch_fs *c,
220 			       struct bkey_s_c k)
221 {
222 	struct bkey_s_c_btree_ptr_v2 bp = bkey_s_c_to_btree_ptr_v2(k);
223 
224 	prt_printf(out, "seq %llx written %u min_key %s",
225 	       le64_to_cpu(bp.v->seq),
226 	       le16_to_cpu(bp.v->sectors_written),
227 	       BTREE_PTR_RANGE_UPDATED(bp.v) ? "R " : "");
228 
229 	bch2_bpos_to_text(out, bp.v->min_key);
230 	prt_printf(out, " ");
231 	bch2_bkey_ptrs_to_text(out, c, k);
232 }
233 
234 void bch2_btree_ptr_v2_compat(enum btree_id btree_id, unsigned version,
235 			      unsigned big_endian, int write,
236 			      struct bkey_s k)
237 {
238 	struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(k);
239 
240 	compat_bpos(0, btree_id, version, big_endian, write, &bp.v->min_key);
241 
242 	if (version < bcachefs_metadata_version_inode_btree_change &&
243 	    btree_id_is_extents(btree_id) &&
244 	    !bkey_eq(bp.v->min_key, POS_MIN))
245 		bp.v->min_key = write
246 			? bpos_nosnap_predecessor(bp.v->min_key)
247 			: bpos_nosnap_successor(bp.v->min_key);
248 }
249 
250 /* KEY_TYPE_extent: */
251 
252 bool bch2_extent_merge(struct bch_fs *c, struct bkey_s l, struct bkey_s_c r)
253 {
254 	struct bkey_ptrs   l_ptrs = bch2_bkey_ptrs(l);
255 	struct bkey_ptrs_c r_ptrs = bch2_bkey_ptrs_c(r);
256 	union bch_extent_entry *en_l;
257 	const union bch_extent_entry *en_r;
258 	struct extent_ptr_decoded lp, rp;
259 	bool use_right_ptr;
260 
261 	en_l = l_ptrs.start;
262 	en_r = r_ptrs.start;
263 	while (en_l < l_ptrs.end && en_r < r_ptrs.end) {
264 		if (extent_entry_type(en_l) != extent_entry_type(en_r))
265 			return false;
266 
267 		en_l = extent_entry_next(en_l);
268 		en_r = extent_entry_next(en_r);
269 	}
270 
271 	if (en_l < l_ptrs.end || en_r < r_ptrs.end)
272 		return false;
273 
274 	en_l = l_ptrs.start;
275 	en_r = r_ptrs.start;
276 	lp.crc = bch2_extent_crc_unpack(l.k, NULL);
277 	rp.crc = bch2_extent_crc_unpack(r.k, NULL);
278 
279 	while (__bkey_ptr_next_decode(l.k, l_ptrs.end, lp, en_l) &&
280 	       __bkey_ptr_next_decode(r.k, r_ptrs.end, rp, en_r)) {
281 		if (lp.ptr.offset + lp.crc.offset + lp.crc.live_size !=
282 		    rp.ptr.offset + rp.crc.offset ||
283 		    lp.ptr.dev			!= rp.ptr.dev ||
284 		    lp.ptr.gen			!= rp.ptr.gen ||
285 		    lp.ptr.unwritten		!= rp.ptr.unwritten ||
286 		    lp.has_ec			!= rp.has_ec)
287 			return false;
288 
289 		/* Extents may not straddle buckets: */
290 		rcu_read_lock();
291 		struct bch_dev *ca = bch2_dev_rcu(c, lp.ptr.dev);
292 		bool same_bucket = ca && PTR_BUCKET_NR(ca, &lp.ptr) == PTR_BUCKET_NR(ca, &rp.ptr);
293 		rcu_read_unlock();
294 
295 		if (!same_bucket)
296 			return false;
297 
298 		if (lp.has_ec			!= rp.has_ec ||
299 		    (lp.has_ec &&
300 		     (lp.ec.block		!= rp.ec.block ||
301 		      lp.ec.redundancy		!= rp.ec.redundancy ||
302 		      lp.ec.idx			!= rp.ec.idx)))
303 			return false;
304 
305 		if (lp.crc.compression_type	!= rp.crc.compression_type ||
306 		    lp.crc.nonce		!= rp.crc.nonce)
307 			return false;
308 
309 		if (lp.crc.offset + lp.crc.live_size + rp.crc.live_size <=
310 		    lp.crc.uncompressed_size) {
311 			/* can use left extent's crc entry */
312 		} else if (lp.crc.live_size <= rp.crc.offset) {
313 			/* can use right extent's crc entry */
314 		} else {
315 			/* check if checksums can be merged: */
316 			if (lp.crc.csum_type		!= rp.crc.csum_type ||
317 			    lp.crc.nonce		!= rp.crc.nonce ||
318 			    crc_is_compressed(lp.crc) ||
319 			    !bch2_checksum_mergeable(lp.crc.csum_type))
320 				return false;
321 
322 			if (lp.crc.offset + lp.crc.live_size != lp.crc.compressed_size ||
323 			    rp.crc.offset)
324 				return false;
325 
326 			if (lp.crc.csum_type &&
327 			    lp.crc.uncompressed_size +
328 			    rp.crc.uncompressed_size > (c->opts.encoded_extent_max >> 9))
329 				return false;
330 		}
331 
332 		en_l = extent_entry_next(en_l);
333 		en_r = extent_entry_next(en_r);
334 	}
335 
336 	en_l = l_ptrs.start;
337 	en_r = r_ptrs.start;
338 	while (en_l < l_ptrs.end && en_r < r_ptrs.end) {
339 		if (extent_entry_is_crc(en_l)) {
340 			struct bch_extent_crc_unpacked crc_l = bch2_extent_crc_unpack(l.k, entry_to_crc(en_l));
341 			struct bch_extent_crc_unpacked crc_r = bch2_extent_crc_unpack(r.k, entry_to_crc(en_r));
342 
343 			if (crc_l.uncompressed_size + crc_r.uncompressed_size >
344 			    bch2_crc_field_size_max[extent_entry_type(en_l)])
345 				return false;
346 		}
347 
348 		en_l = extent_entry_next(en_l);
349 		en_r = extent_entry_next(en_r);
350 	}
351 
352 	use_right_ptr = false;
353 	en_l = l_ptrs.start;
354 	en_r = r_ptrs.start;
355 	while (en_l < l_ptrs.end) {
356 		if (extent_entry_type(en_l) == BCH_EXTENT_ENTRY_ptr &&
357 		    use_right_ptr)
358 			en_l->ptr = en_r->ptr;
359 
360 		if (extent_entry_is_crc(en_l)) {
361 			struct bch_extent_crc_unpacked crc_l =
362 				bch2_extent_crc_unpack(l.k, entry_to_crc(en_l));
363 			struct bch_extent_crc_unpacked crc_r =
364 				bch2_extent_crc_unpack(r.k, entry_to_crc(en_r));
365 
366 			use_right_ptr = false;
367 
368 			if (crc_l.offset + crc_l.live_size + crc_r.live_size <=
369 			    crc_l.uncompressed_size) {
370 				/* can use left extent's crc entry */
371 			} else if (crc_l.live_size <= crc_r.offset) {
372 				/* can use right extent's crc entry */
373 				crc_r.offset -= crc_l.live_size;
374 				bch2_extent_crc_pack(entry_to_crc(en_l), crc_r,
375 						     extent_entry_type(en_l));
376 				use_right_ptr = true;
377 			} else {
378 				crc_l.csum = bch2_checksum_merge(crc_l.csum_type,
379 								 crc_l.csum,
380 								 crc_r.csum,
381 								 crc_r.uncompressed_size << 9);
382 
383 				crc_l.uncompressed_size	+= crc_r.uncompressed_size;
384 				crc_l.compressed_size	+= crc_r.compressed_size;
385 				bch2_extent_crc_pack(entry_to_crc(en_l), crc_l,
386 						     extent_entry_type(en_l));
387 			}
388 		}
389 
390 		en_l = extent_entry_next(en_l);
391 		en_r = extent_entry_next(en_r);
392 	}
393 
394 	bch2_key_resize(l.k, l.k->size + r.k->size);
395 	return true;
396 }
397 
398 /* KEY_TYPE_reservation: */
399 
400 int bch2_reservation_validate(struct bch_fs *c, struct bkey_s_c k,
401 			      enum bch_validate_flags flags)
402 {
403 	struct bkey_s_c_reservation r = bkey_s_c_to_reservation(k);
404 	int ret = 0;
405 
406 	bkey_fsck_err_on(!r.v->nr_replicas || r.v->nr_replicas > BCH_REPLICAS_MAX,
407 			 c, reservation_key_nr_replicas_invalid,
408 			 "invalid nr_replicas (%u)", r.v->nr_replicas);
409 fsck_err:
410 	return ret;
411 }
412 
413 void bch2_reservation_to_text(struct printbuf *out, struct bch_fs *c,
414 			      struct bkey_s_c k)
415 {
416 	struct bkey_s_c_reservation r = bkey_s_c_to_reservation(k);
417 
418 	prt_printf(out, "generation %u replicas %u",
419 	       le32_to_cpu(r.v->generation),
420 	       r.v->nr_replicas);
421 }
422 
423 bool bch2_reservation_merge(struct bch_fs *c, struct bkey_s _l, struct bkey_s_c _r)
424 {
425 	struct bkey_s_reservation l = bkey_s_to_reservation(_l);
426 	struct bkey_s_c_reservation r = bkey_s_c_to_reservation(_r);
427 
428 	if (l.v->generation != r.v->generation ||
429 	    l.v->nr_replicas != r.v->nr_replicas)
430 		return false;
431 
432 	bch2_key_resize(l.k, l.k->size + r.k->size);
433 	return true;
434 }
435 
436 /* Extent checksum entries: */
437 
438 /* returns true if not equal */
439 static inline bool bch2_crc_unpacked_cmp(struct bch_extent_crc_unpacked l,
440 					 struct bch_extent_crc_unpacked r)
441 {
442 	return (l.csum_type		!= r.csum_type ||
443 		l.compression_type	!= r.compression_type ||
444 		l.compressed_size	!= r.compressed_size ||
445 		l.uncompressed_size	!= r.uncompressed_size ||
446 		l.offset		!= r.offset ||
447 		l.live_size		!= r.live_size ||
448 		l.nonce			!= r.nonce ||
449 		bch2_crc_cmp(l.csum, r.csum));
450 }
451 
452 static inline bool can_narrow_crc(struct bch_extent_crc_unpacked u,
453 				  struct bch_extent_crc_unpacked n)
454 {
455 	return !crc_is_compressed(u) &&
456 		u.csum_type &&
457 		u.uncompressed_size > u.live_size &&
458 		bch2_csum_type_is_encryption(u.csum_type) ==
459 		bch2_csum_type_is_encryption(n.csum_type);
460 }
461 
462 bool bch2_can_narrow_extent_crcs(struct bkey_s_c k,
463 				 struct bch_extent_crc_unpacked n)
464 {
465 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
466 	struct bch_extent_crc_unpacked crc;
467 	const union bch_extent_entry *i;
468 
469 	if (!n.csum_type)
470 		return false;
471 
472 	bkey_for_each_crc(k.k, ptrs, crc, i)
473 		if (can_narrow_crc(crc, n))
474 			return true;
475 
476 	return false;
477 }
478 
479 /*
480  * We're writing another replica for this extent, so while we've got the data in
481  * memory we'll be computing a new checksum for the currently live data.
482  *
483  * If there are other replicas we aren't moving, and they are checksummed but
484  * not compressed, we can modify them to point to only the data that is
485  * currently live (so that readers won't have to bounce) while we've got the
486  * checksum we need:
487  */
488 bool bch2_bkey_narrow_crcs(struct bkey_i *k, struct bch_extent_crc_unpacked n)
489 {
490 	struct bkey_ptrs ptrs = bch2_bkey_ptrs(bkey_i_to_s(k));
491 	struct bch_extent_crc_unpacked u;
492 	struct extent_ptr_decoded p;
493 	union bch_extent_entry *i;
494 	bool ret = false;
495 
496 	/* Find a checksum entry that covers only live data: */
497 	if (!n.csum_type) {
498 		bkey_for_each_crc(&k->k, ptrs, u, i)
499 			if (!crc_is_compressed(u) &&
500 			    u.csum_type &&
501 			    u.live_size == u.uncompressed_size) {
502 				n = u;
503 				goto found;
504 			}
505 		return false;
506 	}
507 found:
508 	BUG_ON(crc_is_compressed(n));
509 	BUG_ON(n.offset);
510 	BUG_ON(n.live_size != k->k.size);
511 
512 restart_narrow_pointers:
513 	ptrs = bch2_bkey_ptrs(bkey_i_to_s(k));
514 
515 	bkey_for_each_ptr_decode(&k->k, ptrs, p, i)
516 		if (can_narrow_crc(p.crc, n)) {
517 			bch2_bkey_drop_ptr_noerror(bkey_i_to_s(k), &i->ptr);
518 			p.ptr.offset += p.crc.offset;
519 			p.crc = n;
520 			bch2_extent_ptr_decoded_append(k, &p);
521 			ret = true;
522 			goto restart_narrow_pointers;
523 		}
524 
525 	return ret;
526 }
527 
528 static void bch2_extent_crc_pack(union bch_extent_crc *dst,
529 				 struct bch_extent_crc_unpacked src,
530 				 enum bch_extent_entry_type type)
531 {
532 #define set_common_fields(_dst, _src)					\
533 		_dst.type		= 1 << type;			\
534 		_dst.csum_type		= _src.csum_type,		\
535 		_dst.compression_type	= _src.compression_type,	\
536 		_dst._compressed_size	= _src.compressed_size - 1,	\
537 		_dst._uncompressed_size	= _src.uncompressed_size - 1,	\
538 		_dst.offset		= _src.offset
539 
540 	switch (type) {
541 	case BCH_EXTENT_ENTRY_crc32:
542 		set_common_fields(dst->crc32, src);
543 		dst->crc32.csum		= (u32 __force) *((__le32 *) &src.csum.lo);
544 		break;
545 	case BCH_EXTENT_ENTRY_crc64:
546 		set_common_fields(dst->crc64, src);
547 		dst->crc64.nonce	= src.nonce;
548 		dst->crc64.csum_lo	= (u64 __force) src.csum.lo;
549 		dst->crc64.csum_hi	= (u64 __force) *((__le16 *) &src.csum.hi);
550 		break;
551 	case BCH_EXTENT_ENTRY_crc128:
552 		set_common_fields(dst->crc128, src);
553 		dst->crc128.nonce	= src.nonce;
554 		dst->crc128.csum	= src.csum;
555 		break;
556 	default:
557 		BUG();
558 	}
559 #undef set_common_fields
560 }
561 
562 void bch2_extent_crc_append(struct bkey_i *k,
563 			    struct bch_extent_crc_unpacked new)
564 {
565 	struct bkey_ptrs ptrs = bch2_bkey_ptrs(bkey_i_to_s(k));
566 	union bch_extent_crc *crc = (void *) ptrs.end;
567 	enum bch_extent_entry_type type;
568 
569 	if (bch_crc_bytes[new.csum_type]	<= 4 &&
570 	    new.uncompressed_size		<= CRC32_SIZE_MAX &&
571 	    new.nonce				<= CRC32_NONCE_MAX)
572 		type = BCH_EXTENT_ENTRY_crc32;
573 	else if (bch_crc_bytes[new.csum_type]	<= 10 &&
574 		   new.uncompressed_size	<= CRC64_SIZE_MAX &&
575 		   new.nonce			<= CRC64_NONCE_MAX)
576 		type = BCH_EXTENT_ENTRY_crc64;
577 	else if (bch_crc_bytes[new.csum_type]	<= 16 &&
578 		   new.uncompressed_size	<= CRC128_SIZE_MAX &&
579 		   new.nonce			<= CRC128_NONCE_MAX)
580 		type = BCH_EXTENT_ENTRY_crc128;
581 	else
582 		BUG();
583 
584 	bch2_extent_crc_pack(crc, new, type);
585 
586 	k->k.u64s += extent_entry_u64s(ptrs.end);
587 
588 	EBUG_ON(bkey_val_u64s(&k->k) > BKEY_EXTENT_VAL_U64s_MAX);
589 }
590 
591 /* Generic code for keys with pointers: */
592 
593 unsigned bch2_bkey_nr_ptrs(struct bkey_s_c k)
594 {
595 	return bch2_bkey_devs(k).nr;
596 }
597 
598 unsigned bch2_bkey_nr_ptrs_allocated(struct bkey_s_c k)
599 {
600 	return k.k->type == KEY_TYPE_reservation
601 		? bkey_s_c_to_reservation(k).v->nr_replicas
602 		: bch2_bkey_dirty_devs(k).nr;
603 }
604 
605 unsigned bch2_bkey_nr_ptrs_fully_allocated(struct bkey_s_c k)
606 {
607 	unsigned ret = 0;
608 
609 	if (k.k->type == KEY_TYPE_reservation) {
610 		ret = bkey_s_c_to_reservation(k).v->nr_replicas;
611 	} else {
612 		struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
613 		const union bch_extent_entry *entry;
614 		struct extent_ptr_decoded p;
615 
616 		bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
617 			ret += !p.ptr.cached && !crc_is_compressed(p.crc);
618 	}
619 
620 	return ret;
621 }
622 
623 unsigned bch2_bkey_sectors_compressed(struct bkey_s_c k)
624 {
625 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
626 	const union bch_extent_entry *entry;
627 	struct extent_ptr_decoded p;
628 	unsigned ret = 0;
629 
630 	bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
631 		if (!p.ptr.cached && crc_is_compressed(p.crc))
632 			ret += p.crc.compressed_size;
633 
634 	return ret;
635 }
636 
637 bool bch2_bkey_is_incompressible(struct bkey_s_c k)
638 {
639 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
640 	const union bch_extent_entry *entry;
641 	struct bch_extent_crc_unpacked crc;
642 
643 	bkey_for_each_crc(k.k, ptrs, crc, entry)
644 		if (crc.compression_type == BCH_COMPRESSION_TYPE_incompressible)
645 			return true;
646 	return false;
647 }
648 
649 unsigned bch2_bkey_replicas(struct bch_fs *c, struct bkey_s_c k)
650 {
651 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
652 	const union bch_extent_entry *entry;
653 	struct extent_ptr_decoded p = { 0 };
654 	unsigned replicas = 0;
655 
656 	bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
657 		if (p.ptr.cached)
658 			continue;
659 
660 		if (p.has_ec)
661 			replicas += p.ec.redundancy;
662 
663 		replicas++;
664 
665 	}
666 
667 	return replicas;
668 }
669 
670 static inline unsigned __extent_ptr_durability(struct bch_dev *ca, struct extent_ptr_decoded *p)
671 {
672 	if (p->ptr.cached)
673 		return 0;
674 
675 	return p->has_ec
676 		? p->ec.redundancy + 1
677 		: ca->mi.durability;
678 }
679 
680 unsigned bch2_extent_ptr_desired_durability(struct bch_fs *c, struct extent_ptr_decoded *p)
681 {
682 	struct bch_dev *ca = bch2_dev_rcu(c, p->ptr.dev);
683 
684 	return ca ? __extent_ptr_durability(ca, p) : 0;
685 }
686 
687 unsigned bch2_extent_ptr_durability(struct bch_fs *c, struct extent_ptr_decoded *p)
688 {
689 	struct bch_dev *ca = bch2_dev_rcu(c, p->ptr.dev);
690 
691 	if (!ca || ca->mi.state == BCH_MEMBER_STATE_failed)
692 		return 0;
693 
694 	return __extent_ptr_durability(ca, p);
695 }
696 
697 unsigned bch2_bkey_durability(struct bch_fs *c, struct bkey_s_c k)
698 {
699 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
700 	const union bch_extent_entry *entry;
701 	struct extent_ptr_decoded p;
702 	unsigned durability = 0;
703 
704 	rcu_read_lock();
705 	bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
706 		durability += bch2_extent_ptr_durability(c, &p);
707 	rcu_read_unlock();
708 
709 	return durability;
710 }
711 
712 static unsigned bch2_bkey_durability_safe(struct bch_fs *c, struct bkey_s_c k)
713 {
714 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
715 	const union bch_extent_entry *entry;
716 	struct extent_ptr_decoded p;
717 	unsigned durability = 0;
718 
719 	rcu_read_lock();
720 	bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
721 		if (p.ptr.dev < c->sb.nr_devices && c->devs[p.ptr.dev])
722 			durability += bch2_extent_ptr_durability(c, &p);
723 	rcu_read_unlock();
724 
725 	return durability;
726 }
727 
728 void bch2_bkey_extent_entry_drop(struct bkey_i *k, union bch_extent_entry *entry)
729 {
730 	union bch_extent_entry *end = bkey_val_end(bkey_i_to_s(k));
731 	union bch_extent_entry *next = extent_entry_next(entry);
732 
733 	memmove_u64s(entry, next, (u64 *) end - (u64 *) next);
734 	k->k.u64s -= extent_entry_u64s(entry);
735 }
736 
737 void bch2_extent_ptr_decoded_append(struct bkey_i *k,
738 				    struct extent_ptr_decoded *p)
739 {
740 	struct bkey_ptrs ptrs = bch2_bkey_ptrs(bkey_i_to_s(k));
741 	struct bch_extent_crc_unpacked crc =
742 		bch2_extent_crc_unpack(&k->k, NULL);
743 	union bch_extent_entry *pos;
744 
745 	if (!bch2_crc_unpacked_cmp(crc, p->crc)) {
746 		pos = ptrs.start;
747 		goto found;
748 	}
749 
750 	bkey_for_each_crc(&k->k, ptrs, crc, pos)
751 		if (!bch2_crc_unpacked_cmp(crc, p->crc)) {
752 			pos = extent_entry_next(pos);
753 			goto found;
754 		}
755 
756 	bch2_extent_crc_append(k, p->crc);
757 	pos = bkey_val_end(bkey_i_to_s(k));
758 found:
759 	p->ptr.type = 1 << BCH_EXTENT_ENTRY_ptr;
760 	__extent_entry_insert(k, pos, to_entry(&p->ptr));
761 
762 	if (p->has_ec) {
763 		p->ec.type = 1 << BCH_EXTENT_ENTRY_stripe_ptr;
764 		__extent_entry_insert(k, pos, to_entry(&p->ec));
765 	}
766 }
767 
768 static union bch_extent_entry *extent_entry_prev(struct bkey_ptrs ptrs,
769 					  union bch_extent_entry *entry)
770 {
771 	union bch_extent_entry *i = ptrs.start;
772 
773 	if (i == entry)
774 		return NULL;
775 
776 	while (extent_entry_next(i) != entry)
777 		i = extent_entry_next(i);
778 	return i;
779 }
780 
781 /*
782  * Returns pointer to the next entry after the one being dropped:
783  */
784 union bch_extent_entry *bch2_bkey_drop_ptr_noerror(struct bkey_s k,
785 						   struct bch_extent_ptr *ptr)
786 {
787 	struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
788 	union bch_extent_entry *entry = to_entry(ptr), *next;
789 	union bch_extent_entry *ret = entry;
790 	bool drop_crc = true;
791 
792 	EBUG_ON(ptr < &ptrs.start->ptr ||
793 		ptr >= &ptrs.end->ptr);
794 	EBUG_ON(ptr->type != 1 << BCH_EXTENT_ENTRY_ptr);
795 
796 	for (next = extent_entry_next(entry);
797 	     next != ptrs.end;
798 	     next = extent_entry_next(next)) {
799 		if (extent_entry_is_crc(next)) {
800 			break;
801 		} else if (extent_entry_is_ptr(next)) {
802 			drop_crc = false;
803 			break;
804 		}
805 	}
806 
807 	extent_entry_drop(k, entry);
808 
809 	while ((entry = extent_entry_prev(ptrs, entry))) {
810 		if (extent_entry_is_ptr(entry))
811 			break;
812 
813 		if ((extent_entry_is_crc(entry) && drop_crc) ||
814 		    extent_entry_is_stripe_ptr(entry)) {
815 			ret = (void *) ret - extent_entry_bytes(entry);
816 			extent_entry_drop(k, entry);
817 		}
818 	}
819 
820 	return ret;
821 }
822 
823 union bch_extent_entry *bch2_bkey_drop_ptr(struct bkey_s k,
824 					   struct bch_extent_ptr *ptr)
825 {
826 	bool have_dirty = bch2_bkey_dirty_devs(k.s_c).nr;
827 	union bch_extent_entry *ret =
828 		bch2_bkey_drop_ptr_noerror(k, ptr);
829 
830 	/*
831 	 * If we deleted all the dirty pointers and there's still cached
832 	 * pointers, we could set the cached pointers to dirty if they're not
833 	 * stale - but to do that correctly we'd need to grab an open_bucket
834 	 * reference so that we don't race with bucket reuse:
835 	 */
836 	if (have_dirty &&
837 	    !bch2_bkey_dirty_devs(k.s_c).nr) {
838 		k.k->type = KEY_TYPE_error;
839 		set_bkey_val_u64s(k.k, 0);
840 		ret = NULL;
841 	} else if (!bch2_bkey_nr_ptrs(k.s_c)) {
842 		k.k->type = KEY_TYPE_deleted;
843 		set_bkey_val_u64s(k.k, 0);
844 		ret = NULL;
845 	}
846 
847 	return ret;
848 }
849 
850 void bch2_bkey_drop_device(struct bkey_s k, unsigned dev)
851 {
852 	bch2_bkey_drop_ptrs(k, ptr, ptr->dev == dev);
853 }
854 
855 void bch2_bkey_drop_device_noerror(struct bkey_s k, unsigned dev)
856 {
857 	struct bch_extent_ptr *ptr = bch2_bkey_has_device(k, dev);
858 
859 	if (ptr)
860 		bch2_bkey_drop_ptr_noerror(k, ptr);
861 }
862 
863 const struct bch_extent_ptr *bch2_bkey_has_device_c(struct bkey_s_c k, unsigned dev)
864 {
865 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
866 
867 	bkey_for_each_ptr(ptrs, ptr)
868 		if (ptr->dev == dev)
869 			return ptr;
870 
871 	return NULL;
872 }
873 
874 bool bch2_bkey_has_target(struct bch_fs *c, struct bkey_s_c k, unsigned target)
875 {
876 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
877 	struct bch_dev *ca;
878 	bool ret = false;
879 
880 	rcu_read_lock();
881 	bkey_for_each_ptr(ptrs, ptr)
882 		if (bch2_dev_in_target(c, ptr->dev, target) &&
883 		    (ca = bch2_dev_rcu(c, ptr->dev)) &&
884 		    (!ptr->cached ||
885 		     !dev_ptr_stale_rcu(ca, ptr))) {
886 			ret = true;
887 			break;
888 		}
889 	rcu_read_unlock();
890 
891 	return ret;
892 }
893 
894 bool bch2_bkey_matches_ptr(struct bch_fs *c, struct bkey_s_c k,
895 			   struct bch_extent_ptr m, u64 offset)
896 {
897 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
898 	const union bch_extent_entry *entry;
899 	struct extent_ptr_decoded p;
900 
901 	bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
902 		if (p.ptr.dev	== m.dev &&
903 		    p.ptr.gen	== m.gen &&
904 		    (s64) p.ptr.offset + p.crc.offset - bkey_start_offset(k.k) ==
905 		    (s64) m.offset  - offset)
906 			return true;
907 
908 	return false;
909 }
910 
911 /*
912  * Returns true if two extents refer to the same data:
913  */
914 bool bch2_extents_match(struct bkey_s_c k1, struct bkey_s_c k2)
915 {
916 	if (k1.k->type != k2.k->type)
917 		return false;
918 
919 	if (bkey_extent_is_direct_data(k1.k)) {
920 		struct bkey_ptrs_c ptrs1 = bch2_bkey_ptrs_c(k1);
921 		struct bkey_ptrs_c ptrs2 = bch2_bkey_ptrs_c(k2);
922 		const union bch_extent_entry *entry1, *entry2;
923 		struct extent_ptr_decoded p1, p2;
924 
925 		if (bkey_extent_is_unwritten(k1) != bkey_extent_is_unwritten(k2))
926 			return false;
927 
928 		bkey_for_each_ptr_decode(k1.k, ptrs1, p1, entry1)
929 			bkey_for_each_ptr_decode(k2.k, ptrs2, p2, entry2)
930 				if (p1.ptr.dev		== p2.ptr.dev &&
931 				    p1.ptr.gen		== p2.ptr.gen &&
932 				    (s64) p1.ptr.offset + p1.crc.offset - bkey_start_offset(k1.k) ==
933 				    (s64) p2.ptr.offset + p2.crc.offset - bkey_start_offset(k2.k))
934 					return true;
935 
936 		return false;
937 	} else {
938 		/* KEY_TYPE_deleted, etc. */
939 		return true;
940 	}
941 }
942 
943 struct bch_extent_ptr *
944 bch2_extent_has_ptr(struct bkey_s_c k1, struct extent_ptr_decoded p1, struct bkey_s k2)
945 {
946 	struct bkey_ptrs ptrs2 = bch2_bkey_ptrs(k2);
947 	union bch_extent_entry *entry2;
948 	struct extent_ptr_decoded p2;
949 
950 	bkey_for_each_ptr_decode(k2.k, ptrs2, p2, entry2)
951 		if (p1.ptr.dev		== p2.ptr.dev &&
952 		    p1.ptr.gen		== p2.ptr.gen &&
953 		    (s64) p1.ptr.offset + p1.crc.offset - bkey_start_offset(k1.k) ==
954 		    (s64) p2.ptr.offset + p2.crc.offset - bkey_start_offset(k2.k))
955 			return &entry2->ptr;
956 
957 	return NULL;
958 }
959 
960 void bch2_extent_ptr_set_cached(struct bkey_s k, struct bch_extent_ptr *ptr)
961 {
962 	struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
963 	union bch_extent_entry *entry;
964 	union bch_extent_entry *ec = NULL;
965 
966 	bkey_extent_entry_for_each(ptrs, entry) {
967 		if (&entry->ptr == ptr) {
968 			ptr->cached = true;
969 			if (ec)
970 				extent_entry_drop(k, ec);
971 			return;
972 		}
973 
974 		if (extent_entry_is_stripe_ptr(entry))
975 			ec = entry;
976 		else if (extent_entry_is_ptr(entry))
977 			ec = NULL;
978 	}
979 
980 	BUG();
981 }
982 
983 /*
984  * bch_extent_normalize - clean up an extent, dropping stale pointers etc.
985  *
986  * Returns true if @k should be dropped entirely
987  *
988  * For existing keys, only called when btree nodes are being rewritten, not when
989  * they're merely being compacted/resorted in memory.
990  */
991 bool bch2_extent_normalize(struct bch_fs *c, struct bkey_s k)
992 {
993 	struct bch_dev *ca;
994 
995 	rcu_read_lock();
996 	bch2_bkey_drop_ptrs(k, ptr,
997 		ptr->cached &&
998 		(ca = bch2_dev_rcu(c, ptr->dev)) &&
999 		dev_ptr_stale_rcu(ca, ptr) > 0);
1000 	rcu_read_unlock();
1001 
1002 	return bkey_deleted(k.k);
1003 }
1004 
1005 void bch2_extent_ptr_to_text(struct printbuf *out, struct bch_fs *c, const struct bch_extent_ptr *ptr)
1006 {
1007 	out->atomic++;
1008 	rcu_read_lock();
1009 	struct bch_dev *ca = bch2_dev_rcu(c, ptr->dev);
1010 	if (!ca) {
1011 		prt_printf(out, "ptr: %u:%llu gen %u%s", ptr->dev,
1012 			   (u64) ptr->offset, ptr->gen,
1013 			   ptr->cached ? " cached" : "");
1014 	} else {
1015 		u32 offset;
1016 		u64 b = sector_to_bucket_and_offset(ca, ptr->offset, &offset);
1017 
1018 		prt_printf(out, "ptr: %u:%llu:%u gen %u",
1019 			   ptr->dev, b, offset, ptr->gen);
1020 		if (ptr->cached)
1021 			prt_str(out, " cached");
1022 		if (ptr->unwritten)
1023 			prt_str(out, " unwritten");
1024 		int stale = dev_ptr_stale_rcu(ca, ptr);
1025 		if (stale > 0)
1026 			prt_printf(out, " stale");
1027 		else if (stale)
1028 			prt_printf(out, " invalid");
1029 	}
1030 	rcu_read_unlock();
1031 	--out->atomic;
1032 }
1033 
1034 void bch2_extent_crc_unpacked_to_text(struct printbuf *out, struct bch_extent_crc_unpacked *crc)
1035 {
1036 	prt_printf(out, "crc: c_size %u size %u offset %u nonce %u csum ",
1037 		   crc->compressed_size,
1038 		   crc->uncompressed_size,
1039 		   crc->offset, crc->nonce);
1040 	bch2_prt_csum_type(out, crc->csum_type);
1041 	prt_printf(out, " %0llx:%0llx ", crc->csum.hi, crc->csum.lo);
1042 	prt_str(out, " compress ");
1043 	bch2_prt_compression_type(out, crc->compression_type);
1044 }
1045 
1046 void bch2_bkey_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
1047 			    struct bkey_s_c k)
1048 {
1049 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1050 	const union bch_extent_entry *entry;
1051 	bool first = true;
1052 
1053 	if (c)
1054 		prt_printf(out, "durability: %u ", bch2_bkey_durability_safe(c, k));
1055 
1056 	bkey_extent_entry_for_each(ptrs, entry) {
1057 		if (!first)
1058 			prt_printf(out, " ");
1059 
1060 		switch (__extent_entry_type(entry)) {
1061 		case BCH_EXTENT_ENTRY_ptr:
1062 			bch2_extent_ptr_to_text(out, c, entry_to_ptr(entry));
1063 			break;
1064 
1065 		case BCH_EXTENT_ENTRY_crc32:
1066 		case BCH_EXTENT_ENTRY_crc64:
1067 		case BCH_EXTENT_ENTRY_crc128: {
1068 			struct bch_extent_crc_unpacked crc =
1069 				bch2_extent_crc_unpack(k.k, entry_to_crc(entry));
1070 
1071 			bch2_extent_crc_unpacked_to_text(out, &crc);
1072 			break;
1073 		}
1074 		case BCH_EXTENT_ENTRY_stripe_ptr: {
1075 			const struct bch_extent_stripe_ptr *ec = &entry->stripe_ptr;
1076 
1077 			prt_printf(out, "ec: idx %llu block %u",
1078 			       (u64) ec->idx, ec->block);
1079 			break;
1080 		}
1081 		case BCH_EXTENT_ENTRY_rebalance: {
1082 			const struct bch_extent_rebalance *r = &entry->rebalance;
1083 
1084 			prt_str(out, "rebalance: target ");
1085 			if (c)
1086 				bch2_target_to_text(out, c, r->target);
1087 			else
1088 				prt_printf(out, "%u", r->target);
1089 			prt_str(out, " compression ");
1090 			bch2_compression_opt_to_text(out, r->compression);
1091 			break;
1092 		}
1093 		default:
1094 			prt_printf(out, "(invalid extent entry %.16llx)", *((u64 *) entry));
1095 			return;
1096 		}
1097 
1098 		first = false;
1099 	}
1100 }
1101 
1102 static int extent_ptr_validate(struct bch_fs *c,
1103 			       struct bkey_s_c k,
1104 			       enum bch_validate_flags flags,
1105 			       const struct bch_extent_ptr *ptr,
1106 			       unsigned size_ondisk,
1107 			       bool metadata)
1108 {
1109 	int ret = 0;
1110 
1111 	rcu_read_lock();
1112 	struct bch_dev *ca = bch2_dev_rcu(c, ptr->dev);
1113 	if (!ca) {
1114 		rcu_read_unlock();
1115 		return 0;
1116 	}
1117 	u32 bucket_offset;
1118 	u64 bucket = sector_to_bucket_and_offset(ca, ptr->offset, &bucket_offset);
1119 	unsigned first_bucket	= ca->mi.first_bucket;
1120 	u64 nbuckets		= ca->mi.nbuckets;
1121 	unsigned bucket_size	= ca->mi.bucket_size;
1122 	rcu_read_unlock();
1123 
1124 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1125 	bkey_for_each_ptr(ptrs, ptr2)
1126 		bkey_fsck_err_on(ptr != ptr2 && ptr->dev == ptr2->dev,
1127 				 c, ptr_to_duplicate_device,
1128 				 "multiple pointers to same device (%u)", ptr->dev);
1129 
1130 
1131 	bkey_fsck_err_on(bucket >= nbuckets,
1132 			 c, ptr_after_last_bucket,
1133 			 "pointer past last bucket (%llu > %llu)", bucket, nbuckets);
1134 	bkey_fsck_err_on(bucket < first_bucket,
1135 			 c, ptr_before_first_bucket,
1136 			 "pointer before first bucket (%llu < %u)", bucket, first_bucket);
1137 	bkey_fsck_err_on(bucket_offset + size_ondisk > bucket_size,
1138 			 c, ptr_spans_multiple_buckets,
1139 			 "pointer spans multiple buckets (%u + %u > %u)",
1140 		       bucket_offset, size_ondisk, bucket_size);
1141 fsck_err:
1142 	return ret;
1143 }
1144 
1145 int bch2_bkey_ptrs_validate(struct bch_fs *c, struct bkey_s_c k,
1146 			    enum bch_validate_flags flags)
1147 {
1148 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1149 	const union bch_extent_entry *entry;
1150 	struct bch_extent_crc_unpacked crc;
1151 	unsigned size_ondisk = k.k->size;
1152 	unsigned nonce = UINT_MAX;
1153 	unsigned nr_ptrs = 0;
1154 	bool have_written = false, have_unwritten = false, have_ec = false, crc_since_last_ptr = false;
1155 	int ret = 0;
1156 
1157 	if (bkey_is_btree_ptr(k.k))
1158 		size_ondisk = btree_sectors(c);
1159 
1160 	bkey_extent_entry_for_each(ptrs, entry) {
1161 		bkey_fsck_err_on(__extent_entry_type(entry) >= BCH_EXTENT_ENTRY_MAX,
1162 				 c, extent_ptrs_invalid_entry,
1163 				 "invalid extent entry type (got %u, max %u)",
1164 				 __extent_entry_type(entry), BCH_EXTENT_ENTRY_MAX);
1165 
1166 		bkey_fsck_err_on(bkey_is_btree_ptr(k.k) &&
1167 				 !extent_entry_is_ptr(entry),
1168 				 c, btree_ptr_has_non_ptr,
1169 				 "has non ptr field");
1170 
1171 		switch (extent_entry_type(entry)) {
1172 		case BCH_EXTENT_ENTRY_ptr:
1173 			ret = extent_ptr_validate(c, k, flags, &entry->ptr, size_ondisk, false);
1174 			if (ret)
1175 				return ret;
1176 
1177 			bkey_fsck_err_on(entry->ptr.cached && have_ec,
1178 					 c, ptr_cached_and_erasure_coded,
1179 					 "cached, erasure coded ptr");
1180 
1181 			if (!entry->ptr.unwritten)
1182 				have_written = true;
1183 			else
1184 				have_unwritten = true;
1185 
1186 			have_ec = false;
1187 			crc_since_last_ptr = false;
1188 			nr_ptrs++;
1189 			break;
1190 		case BCH_EXTENT_ENTRY_crc32:
1191 		case BCH_EXTENT_ENTRY_crc64:
1192 		case BCH_EXTENT_ENTRY_crc128:
1193 			crc = bch2_extent_crc_unpack(k.k, entry_to_crc(entry));
1194 
1195 			bkey_fsck_err_on(crc.offset + crc.live_size > crc.uncompressed_size,
1196 					 c, ptr_crc_uncompressed_size_too_small,
1197 					 "checksum offset + key size > uncompressed size");
1198 			bkey_fsck_err_on(!bch2_checksum_type_valid(c, crc.csum_type),
1199 					 c, ptr_crc_csum_type_unknown,
1200 					 "invalid checksum type");
1201 			bkey_fsck_err_on(crc.compression_type >= BCH_COMPRESSION_TYPE_NR,
1202 					 c, ptr_crc_compression_type_unknown,
1203 					 "invalid compression type");
1204 
1205 			if (bch2_csum_type_is_encryption(crc.csum_type)) {
1206 				if (nonce == UINT_MAX)
1207 					nonce = crc.offset + crc.nonce;
1208 				else if (nonce != crc.offset + crc.nonce)
1209 					bkey_fsck_err(c, ptr_crc_nonce_mismatch,
1210 						      "incorrect nonce");
1211 			}
1212 
1213 			bkey_fsck_err_on(crc_since_last_ptr,
1214 					 c, ptr_crc_redundant,
1215 					 "redundant crc entry");
1216 			crc_since_last_ptr = true;
1217 
1218 			bkey_fsck_err_on(crc_is_encoded(crc) &&
1219 					 (crc.uncompressed_size > c->opts.encoded_extent_max >> 9) &&
1220 					 (flags & (BCH_VALIDATE_write|BCH_VALIDATE_commit)),
1221 					 c, ptr_crc_uncompressed_size_too_big,
1222 					 "too large encoded extent");
1223 
1224 			size_ondisk = crc.compressed_size;
1225 			break;
1226 		case BCH_EXTENT_ENTRY_stripe_ptr:
1227 			bkey_fsck_err_on(have_ec,
1228 					 c, ptr_stripe_redundant,
1229 					 "redundant stripe entry");
1230 			have_ec = true;
1231 			break;
1232 		case BCH_EXTENT_ENTRY_rebalance: {
1233 			/*
1234 			 * this shouldn't be a fsck error, for forward
1235 			 * compatibility; the rebalance code should just refetch
1236 			 * the compression opt if it's unknown
1237 			 */
1238 #if 0
1239 			const struct bch_extent_rebalance *r = &entry->rebalance;
1240 
1241 			if (!bch2_compression_opt_valid(r->compression)) {
1242 				struct bch_compression_opt opt = __bch2_compression_decode(r->compression);
1243 				prt_printf(err, "invalid compression opt %u:%u",
1244 					   opt.type, opt.level);
1245 				return -BCH_ERR_invalid_bkey;
1246 			}
1247 #endif
1248 			break;
1249 		}
1250 		}
1251 	}
1252 
1253 	bkey_fsck_err_on(!nr_ptrs,
1254 			 c, extent_ptrs_no_ptrs,
1255 			 "no ptrs");
1256 	bkey_fsck_err_on(nr_ptrs > BCH_BKEY_PTRS_MAX,
1257 			 c, extent_ptrs_too_many_ptrs,
1258 			 "too many ptrs: %u > %u", nr_ptrs, BCH_BKEY_PTRS_MAX);
1259 	bkey_fsck_err_on(have_written && have_unwritten,
1260 			 c, extent_ptrs_written_and_unwritten,
1261 			 "extent with unwritten and written ptrs");
1262 	bkey_fsck_err_on(k.k->type != KEY_TYPE_extent && have_unwritten,
1263 			 c, extent_ptrs_unwritten,
1264 			 "has unwritten ptrs");
1265 	bkey_fsck_err_on(crc_since_last_ptr,
1266 			 c, extent_ptrs_redundant_crc,
1267 			 "redundant crc entry");
1268 	bkey_fsck_err_on(have_ec,
1269 			 c, extent_ptrs_redundant_stripe,
1270 			 "redundant stripe entry");
1271 fsck_err:
1272 	return ret;
1273 }
1274 
1275 void bch2_ptr_swab(struct bkey_s k)
1276 {
1277 	struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
1278 	union bch_extent_entry *entry;
1279 	u64 *d;
1280 
1281 	for (d =  (u64 *) ptrs.start;
1282 	     d != (u64 *) ptrs.end;
1283 	     d++)
1284 		*d = swab64(*d);
1285 
1286 	for (entry = ptrs.start;
1287 	     entry < ptrs.end;
1288 	     entry = extent_entry_next(entry)) {
1289 		switch (extent_entry_type(entry)) {
1290 		case BCH_EXTENT_ENTRY_ptr:
1291 			break;
1292 		case BCH_EXTENT_ENTRY_crc32:
1293 			entry->crc32.csum = swab32(entry->crc32.csum);
1294 			break;
1295 		case BCH_EXTENT_ENTRY_crc64:
1296 			entry->crc64.csum_hi = swab16(entry->crc64.csum_hi);
1297 			entry->crc64.csum_lo = swab64(entry->crc64.csum_lo);
1298 			break;
1299 		case BCH_EXTENT_ENTRY_crc128:
1300 			entry->crc128.csum.hi = (__force __le64)
1301 				swab64((__force u64) entry->crc128.csum.hi);
1302 			entry->crc128.csum.lo = (__force __le64)
1303 				swab64((__force u64) entry->crc128.csum.lo);
1304 			break;
1305 		case BCH_EXTENT_ENTRY_stripe_ptr:
1306 			break;
1307 		case BCH_EXTENT_ENTRY_rebalance:
1308 			break;
1309 		}
1310 	}
1311 }
1312 
1313 const struct bch_extent_rebalance *bch2_bkey_rebalance_opts(struct bkey_s_c k)
1314 {
1315 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1316 	const union bch_extent_entry *entry;
1317 
1318 	bkey_extent_entry_for_each(ptrs, entry)
1319 		if (__extent_entry_type(entry) == BCH_EXTENT_ENTRY_rebalance)
1320 			return &entry->rebalance;
1321 
1322 	return NULL;
1323 }
1324 
1325 unsigned bch2_bkey_ptrs_need_rebalance(struct bch_fs *c, struct bkey_s_c k,
1326 				       unsigned target, unsigned compression)
1327 {
1328 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1329 	unsigned rewrite_ptrs = 0;
1330 
1331 	if (compression) {
1332 		unsigned compression_type = bch2_compression_opt_to_type(compression);
1333 		const union bch_extent_entry *entry;
1334 		struct extent_ptr_decoded p;
1335 		unsigned i = 0;
1336 
1337 		bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1338 			if (p.crc.compression_type == BCH_COMPRESSION_TYPE_incompressible ||
1339 			    p.ptr.unwritten) {
1340 				rewrite_ptrs = 0;
1341 				goto incompressible;
1342 			}
1343 
1344 			if (!p.ptr.cached && p.crc.compression_type != compression_type)
1345 				rewrite_ptrs |= 1U << i;
1346 			i++;
1347 		}
1348 	}
1349 incompressible:
1350 	if (target && bch2_target_accepts_data(c, BCH_DATA_user, target)) {
1351 		unsigned i = 0;
1352 
1353 		bkey_for_each_ptr(ptrs, ptr) {
1354 			if (!ptr->cached && !bch2_dev_in_target(c, ptr->dev, target))
1355 				rewrite_ptrs |= 1U << i;
1356 			i++;
1357 		}
1358 	}
1359 
1360 	return rewrite_ptrs;
1361 }
1362 
1363 bool bch2_bkey_needs_rebalance(struct bch_fs *c, struct bkey_s_c k)
1364 {
1365 	const struct bch_extent_rebalance *r = bch2_bkey_rebalance_opts(k);
1366 
1367 	/*
1368 	 * If it's an indirect extent, we don't delete the rebalance entry when
1369 	 * done so that we know what options were applied - check if it still
1370 	 * needs work done:
1371 	 */
1372 	if (r &&
1373 	    k.k->type == KEY_TYPE_reflink_v &&
1374 	    !bch2_bkey_ptrs_need_rebalance(c, k, r->target, r->compression))
1375 		r = NULL;
1376 
1377 	return r != NULL;
1378 }
1379 
1380 int bch2_bkey_set_needs_rebalance(struct bch_fs *c, struct bkey_i *_k,
1381 				  struct bch_io_opts *opts)
1382 {
1383 	struct bkey_s k = bkey_i_to_s(_k);
1384 	struct bch_extent_rebalance *r;
1385 	unsigned target = opts->background_target;
1386 	unsigned compression = background_compression(*opts);
1387 	bool needs_rebalance;
1388 
1389 	if (!bkey_extent_is_direct_data(k.k))
1390 		return 0;
1391 
1392 	/* get existing rebalance entry: */
1393 	r = (struct bch_extent_rebalance *) bch2_bkey_rebalance_opts(k.s_c);
1394 	if (r) {
1395 		if (k.k->type == KEY_TYPE_reflink_v) {
1396 			/*
1397 			 * indirect extents: existing options take precedence,
1398 			 * so that we don't move extents back and forth if
1399 			 * they're referenced by different inodes with different
1400 			 * options:
1401 			 */
1402 			if (r->target)
1403 				target = r->target;
1404 			if (r->compression)
1405 				compression = r->compression;
1406 		}
1407 
1408 		r->target	= target;
1409 		r->compression	= compression;
1410 	}
1411 
1412 	needs_rebalance = bch2_bkey_ptrs_need_rebalance(c, k.s_c, target, compression);
1413 
1414 	if (needs_rebalance && !r) {
1415 		union bch_extent_entry *new = bkey_val_end(k);
1416 
1417 		new->rebalance.type		= 1U << BCH_EXTENT_ENTRY_rebalance;
1418 		new->rebalance.compression	= compression;
1419 		new->rebalance.target		= target;
1420 		new->rebalance.unused		= 0;
1421 		k.k->u64s += extent_entry_u64s(new);
1422 	} else if (!needs_rebalance && r && k.k->type != KEY_TYPE_reflink_v) {
1423 		/*
1424 		 * For indirect extents, don't delete the rebalance entry when
1425 		 * we're finished so that we know we specifically moved it or
1426 		 * compressed it to its current location/compression type
1427 		 */
1428 		extent_entry_drop(k, (union bch_extent_entry *) r);
1429 	}
1430 
1431 	return 0;
1432 }
1433 
1434 /* Generic extent code: */
1435 
1436 int bch2_cut_front_s(struct bpos where, struct bkey_s k)
1437 {
1438 	unsigned new_val_u64s = bkey_val_u64s(k.k);
1439 	int val_u64s_delta;
1440 	u64 sub;
1441 
1442 	if (bkey_le(where, bkey_start_pos(k.k)))
1443 		return 0;
1444 
1445 	EBUG_ON(bkey_gt(where, k.k->p));
1446 
1447 	sub = where.offset - bkey_start_offset(k.k);
1448 
1449 	k.k->size -= sub;
1450 
1451 	if (!k.k->size) {
1452 		k.k->type = KEY_TYPE_deleted;
1453 		new_val_u64s = 0;
1454 	}
1455 
1456 	switch (k.k->type) {
1457 	case KEY_TYPE_extent:
1458 	case KEY_TYPE_reflink_v: {
1459 		struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
1460 		union bch_extent_entry *entry;
1461 		bool seen_crc = false;
1462 
1463 		bkey_extent_entry_for_each(ptrs, entry) {
1464 			switch (extent_entry_type(entry)) {
1465 			case BCH_EXTENT_ENTRY_ptr:
1466 				if (!seen_crc)
1467 					entry->ptr.offset += sub;
1468 				break;
1469 			case BCH_EXTENT_ENTRY_crc32:
1470 				entry->crc32.offset += sub;
1471 				break;
1472 			case BCH_EXTENT_ENTRY_crc64:
1473 				entry->crc64.offset += sub;
1474 				break;
1475 			case BCH_EXTENT_ENTRY_crc128:
1476 				entry->crc128.offset += sub;
1477 				break;
1478 			case BCH_EXTENT_ENTRY_stripe_ptr:
1479 				break;
1480 			case BCH_EXTENT_ENTRY_rebalance:
1481 				break;
1482 			}
1483 
1484 			if (extent_entry_is_crc(entry))
1485 				seen_crc = true;
1486 		}
1487 
1488 		break;
1489 	}
1490 	case KEY_TYPE_reflink_p: {
1491 		struct bkey_s_reflink_p p = bkey_s_to_reflink_p(k);
1492 
1493 		le64_add_cpu(&p.v->idx, sub);
1494 		break;
1495 	}
1496 	case KEY_TYPE_inline_data:
1497 	case KEY_TYPE_indirect_inline_data: {
1498 		void *p = bkey_inline_data_p(k);
1499 		unsigned bytes = bkey_inline_data_bytes(k.k);
1500 
1501 		sub = min_t(u64, sub << 9, bytes);
1502 
1503 		memmove(p, p + sub, bytes - sub);
1504 
1505 		new_val_u64s -= sub >> 3;
1506 		break;
1507 	}
1508 	}
1509 
1510 	val_u64s_delta = bkey_val_u64s(k.k) - new_val_u64s;
1511 	BUG_ON(val_u64s_delta < 0);
1512 
1513 	set_bkey_val_u64s(k.k, new_val_u64s);
1514 	memset(bkey_val_end(k), 0, val_u64s_delta * sizeof(u64));
1515 	return -val_u64s_delta;
1516 }
1517 
1518 int bch2_cut_back_s(struct bpos where, struct bkey_s k)
1519 {
1520 	unsigned new_val_u64s = bkey_val_u64s(k.k);
1521 	int val_u64s_delta;
1522 	u64 len = 0;
1523 
1524 	if (bkey_ge(where, k.k->p))
1525 		return 0;
1526 
1527 	EBUG_ON(bkey_lt(where, bkey_start_pos(k.k)));
1528 
1529 	len = where.offset - bkey_start_offset(k.k);
1530 
1531 	k.k->p.offset = where.offset;
1532 	k.k->size = len;
1533 
1534 	if (!len) {
1535 		k.k->type = KEY_TYPE_deleted;
1536 		new_val_u64s = 0;
1537 	}
1538 
1539 	switch (k.k->type) {
1540 	case KEY_TYPE_inline_data:
1541 	case KEY_TYPE_indirect_inline_data:
1542 		new_val_u64s = (bkey_inline_data_offset(k.k) +
1543 				min(bkey_inline_data_bytes(k.k), k.k->size << 9)) >> 3;
1544 		break;
1545 	}
1546 
1547 	val_u64s_delta = bkey_val_u64s(k.k) - new_val_u64s;
1548 	BUG_ON(val_u64s_delta < 0);
1549 
1550 	set_bkey_val_u64s(k.k, new_val_u64s);
1551 	memset(bkey_val_end(k), 0, val_u64s_delta * sizeof(u64));
1552 	return -val_u64s_delta;
1553 }
1554