1 // SPDX-License-Identifier: GPL-2.0
2 #include "bcachefs.h"
3 #include "bkey_buf.h"
4 #include "btree_update.h"
5 #include "buckets.h"
6 #include "error.h"
7 #include "extents.h"
8 #include "inode.h"
9 #include "io_misc.h"
10 #include "io_write.h"
11 #include "rebalance.h"
12 #include "reflink.h"
13 #include "subvolume.h"
14 #include "super-io.h"
15
16 #include <linux/sched/signal.h>
17
18 static inline unsigned bkey_type_to_indirect(const struct bkey *k)
19 {
20 switch (k->type) {
21 case KEY_TYPE_extent:
22 return KEY_TYPE_reflink_v;
23 case KEY_TYPE_inline_data:
24 return KEY_TYPE_indirect_inline_data;
25 default:
26 return 0;
27 }
28 }
29
30 /* reflink pointers */
31
32 int bch2_reflink_p_validate(struct bch_fs *c, struct bkey_s_c k,
33 enum bch_validate_flags flags)
34 {
35 struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
36 int ret = 0;
37
38 bkey_fsck_err_on(le64_to_cpu(p.v->idx) < le32_to_cpu(p.v->front_pad),
39 c, reflink_p_front_pad_bad,
40 "idx < front_pad (%llu < %u)",
41 le64_to_cpu(p.v->idx), le32_to_cpu(p.v->front_pad));
42 fsck_err:
43 return ret;
44 }
45
46 void bch2_reflink_p_to_text(struct printbuf *out, struct bch_fs *c,
47 struct bkey_s_c k)
48 {
49 struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
50
51 prt_printf(out, "idx %llu front_pad %u back_pad %u",
52 le64_to_cpu(p.v->idx),
53 le32_to_cpu(p.v->front_pad),
54 le32_to_cpu(p.v->back_pad));
55 }
56
57 bool bch2_reflink_p_merge(struct bch_fs *c, struct bkey_s _l, struct bkey_s_c _r)
58 {
59 struct bkey_s_reflink_p l = bkey_s_to_reflink_p(_l);
60 struct bkey_s_c_reflink_p r = bkey_s_c_to_reflink_p(_r);
61
62 /*
63 * Disabled for now, the triggers code needs to be reworked for merging
64 * of reflink pointers to work:
65 */
66 return false;
67
68 if (le64_to_cpu(l.v->idx) + l.k->size != le64_to_cpu(r.v->idx))
69 return false;
70
71 bch2_key_resize(l.k, l.k->size + r.k->size);
72 return true;
73 }
74
75 static int trans_trigger_reflink_p_segment(struct btree_trans *trans,
76 struct bkey_s_c_reflink_p p, u64 *idx,
77 enum btree_iter_update_trigger_flags flags)
78 {
79 struct bch_fs *c = trans->c;
80 struct btree_iter iter;
81 struct bkey_i *k;
82 __le64 *refcount;
83 int add = !(flags & BTREE_TRIGGER_overwrite) ? 1 : -1;
84 struct printbuf buf = PRINTBUF;
85 int ret;
86
87 k = bch2_bkey_get_mut_noupdate(trans, &iter,
88 BTREE_ID_reflink, POS(0, *idx),
89 BTREE_ITER_with_updates);
90 ret = PTR_ERR_OR_ZERO(k);
91 if (ret)
92 goto err;
93
94 refcount = bkey_refcount(bkey_i_to_s(k));
95 if (!refcount) {
96 bch2_bkey_val_to_text(&buf, c, p.s_c);
97 bch2_trans_inconsistent(trans,
98 "nonexistent indirect extent at %llu while marking\n %s",
99 *idx, buf.buf);
100 ret = -EIO;
101 goto err;
102 }
103
104 if (!*refcount && (flags & BTREE_TRIGGER_overwrite)) {
105 bch2_bkey_val_to_text(&buf, c, p.s_c);
106 bch2_trans_inconsistent(trans,
107 "indirect extent refcount underflow at %llu while marking\n %s",
108 *idx, buf.buf);
109 ret = -EIO;
110 goto err;
111 }
112
113 if (flags & BTREE_TRIGGER_insert) {
114 struct bch_reflink_p *v = (struct bch_reflink_p *) p.v;
115 u64 pad;
116
117 pad = max_t(s64, le32_to_cpu(v->front_pad),
118 le64_to_cpu(v->idx) - bkey_start_offset(&k->k));
119 BUG_ON(pad > U32_MAX);
120 v->front_pad = cpu_to_le32(pad);
121
122 pad = max_t(s64, le32_to_cpu(v->back_pad),
123 k->k.p.offset - p.k->size - le64_to_cpu(v->idx));
124 BUG_ON(pad > U32_MAX);
125 v->back_pad = cpu_to_le32(pad);
126 }
127
128 le64_add_cpu(refcount, add);
129
130 bch2_btree_iter_set_pos_to_extent_start(&iter);
131 ret = bch2_trans_update(trans, &iter, k, 0);
132 if (ret)
133 goto err;
134
135 *idx = k->k.p.offset;
136 err:
137 bch2_trans_iter_exit(trans, &iter);
138 printbuf_exit(&buf);
139 return ret;
140 }
141
142 static s64 gc_trigger_reflink_p_segment(struct btree_trans *trans,
143 struct bkey_s_c_reflink_p p, u64 *idx,
144 enum btree_iter_update_trigger_flags flags,
145 size_t r_idx)
146 {
147 struct bch_fs *c = trans->c;
148 struct reflink_gc *r;
149 int add = !(flags & BTREE_TRIGGER_overwrite) ? 1 : -1;
150 u64 start = le64_to_cpu(p.v->idx);
151 u64 end = le64_to_cpu(p.v->idx) + p.k->size;
152 u64 next_idx = end + le32_to_cpu(p.v->back_pad);
153 s64 ret = 0;
154 struct printbuf buf = PRINTBUF;
155
156 if (r_idx >= c->reflink_gc_nr)
157 goto not_found;
158
159 r = genradix_ptr(&c->reflink_gc_table, r_idx);
160 next_idx = min(next_idx, r->offset - r->size);
161 if (*idx < next_idx)
162 goto not_found;
163
164 BUG_ON((s64) r->refcount + add < 0);
165
166 if (flags & BTREE_TRIGGER_gc)
167 r->refcount += add;
168 *idx = r->offset;
169 return 0;
170 not_found:
171 BUG_ON(!(flags & BTREE_TRIGGER_check_repair));
172
173 if (fsck_err(trans, reflink_p_to_missing_reflink_v,
174 "pointer to missing indirect extent\n"
175 " %s\n"
176 " missing range %llu-%llu",
177 (bch2_bkey_val_to_text(&buf, c, p.s_c), buf.buf),
178 *idx, next_idx)) {
179 struct bkey_i *update = bch2_bkey_make_mut_noupdate(trans, p.s_c);
180 ret = PTR_ERR_OR_ZERO(update);
181 if (ret)
182 goto err;
183
184 if (next_idx <= start) {
185 bkey_i_to_reflink_p(update)->v.front_pad = cpu_to_le32(start - next_idx);
186 } else if (*idx >= end) {
187 bkey_i_to_reflink_p(update)->v.back_pad = cpu_to_le32(*idx - end);
188 } else {
189 bkey_error_init(update);
190 update->k.p = p.k->p;
191 update->k.size = p.k->size;
192 set_bkey_val_u64s(&update->k, 0);
193 }
194
195 ret = bch2_btree_insert_trans(trans, BTREE_ID_extents, update, BTREE_TRIGGER_norun);
196 }
197
198 *idx = next_idx;
199 err:
200 fsck_err:
201 printbuf_exit(&buf);
202 return ret;
203 }
204
205 static int __trigger_reflink_p(struct btree_trans *trans,
206 enum btree_id btree_id, unsigned level, struct bkey_s_c k,
207 enum btree_iter_update_trigger_flags flags)
208 {
209 struct bch_fs *c = trans->c;
210 struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
211 int ret = 0;
212
213 u64 idx = le64_to_cpu(p.v->idx) - le32_to_cpu(p.v->front_pad);
214 u64 end = le64_to_cpu(p.v->idx) + p.k->size + le32_to_cpu(p.v->back_pad);
215
216 if (flags & BTREE_TRIGGER_transactional) {
217 while (idx < end && !ret)
218 ret = trans_trigger_reflink_p_segment(trans, p, &idx, flags);
219 }
220
221 if (flags & (BTREE_TRIGGER_check_repair|BTREE_TRIGGER_gc)) {
222 size_t l = 0, r = c->reflink_gc_nr;
223
224 while (l < r) {
225 size_t m = l + (r - l) / 2;
226 struct reflink_gc *ref = genradix_ptr(&c->reflink_gc_table, m);
227 if (ref->offset <= idx)
228 l = m + 1;
229 else
230 r = m;
231 }
232
233 while (idx < end && !ret)
234 ret = gc_trigger_reflink_p_segment(trans, p, &idx, flags, l++);
235 }
236
237 return ret;
238 }
239
240 int bch2_trigger_reflink_p(struct btree_trans *trans,
241 enum btree_id btree_id, unsigned level,
242 struct bkey_s_c old,
243 struct bkey_s new,
244 enum btree_iter_update_trigger_flags flags)
245 {
246 if ((flags & BTREE_TRIGGER_transactional) &&
247 (flags & BTREE_TRIGGER_insert)) {
248 struct bch_reflink_p *v = bkey_s_to_reflink_p(new).v;
249
250 v->front_pad = v->back_pad = 0;
251 }
252
253 return trigger_run_overwrite_then_insert(__trigger_reflink_p, trans, btree_id, level, old, new, flags);
254 }
255
256 /* indirect extents */
257
258 int bch2_reflink_v_validate(struct bch_fs *c, struct bkey_s_c k,
259 enum bch_validate_flags flags)
260 {
261 return bch2_bkey_ptrs_validate(c, k, flags);
262 }
263
264 void bch2_reflink_v_to_text(struct printbuf *out, struct bch_fs *c,
265 struct bkey_s_c k)
266 {
267 struct bkey_s_c_reflink_v r = bkey_s_c_to_reflink_v(k);
268
269 prt_printf(out, "refcount: %llu ", le64_to_cpu(r.v->refcount));
270
271 bch2_bkey_ptrs_to_text(out, c, k);
272 }
273
274 #if 0
275 Currently disabled, needs to be debugged:
276
277 bool bch2_reflink_v_merge(struct bch_fs *c, struct bkey_s _l, struct bkey_s_c _r)
278 {
279 struct bkey_s_reflink_v l = bkey_s_to_reflink_v(_l);
280 struct bkey_s_c_reflink_v r = bkey_s_c_to_reflink_v(_r);
281
282 return l.v->refcount == r.v->refcount && bch2_extent_merge(c, _l, _r);
283 }
284 #endif
285
286 static inline void
287 check_indirect_extent_deleting(struct bkey_s new,
288 enum btree_iter_update_trigger_flags *flags)
289 {
290 if ((*flags & BTREE_TRIGGER_insert) && !*bkey_refcount(new)) {
291 new.k->type = KEY_TYPE_deleted;
292 new.k->size = 0;
293 set_bkey_val_u64s(new.k, 0);
294 *flags &= ~BTREE_TRIGGER_insert;
295 }
296 }
297
298 int bch2_trigger_reflink_v(struct btree_trans *trans,
299 enum btree_id btree_id, unsigned level,
300 struct bkey_s_c old, struct bkey_s new,
301 enum btree_iter_update_trigger_flags flags)
302 {
303 if ((flags & BTREE_TRIGGER_transactional) &&
304 (flags & BTREE_TRIGGER_insert))
305 check_indirect_extent_deleting(new, &flags);
306
307 return bch2_trigger_extent(trans, btree_id, level, old, new, flags);
308 }
309
310 /* indirect inline data */
311
312 int bch2_indirect_inline_data_validate(struct bch_fs *c, struct bkey_s_c k,
313 enum bch_validate_flags flags)
314 {
315 return 0;
316 }
317
318 void bch2_indirect_inline_data_to_text(struct printbuf *out,
319 struct bch_fs *c, struct bkey_s_c k)
320 {
321 struct bkey_s_c_indirect_inline_data d = bkey_s_c_to_indirect_inline_data(k);
322 unsigned datalen = bkey_inline_data_bytes(k.k);
323
324 prt_printf(out, "refcount %llu datalen %u: %*phN",
325 le64_to_cpu(d.v->refcount), datalen,
326 min(datalen, 32U), d.v->data);
327 }
328
329 int bch2_trigger_indirect_inline_data(struct btree_trans *trans,
330 enum btree_id btree_id, unsigned level,
331 struct bkey_s_c old, struct bkey_s new,
332 enum btree_iter_update_trigger_flags flags)
333 {
334 check_indirect_extent_deleting(new, &flags);
335
336 return 0;
337 }
338
339 static int bch2_make_extent_indirect(struct btree_trans *trans,
340 struct btree_iter *extent_iter,
341 struct bkey_i *orig)
342 {
343 struct bch_fs *c = trans->c;
344 struct btree_iter reflink_iter = { NULL };
345 struct bkey_s_c k;
346 struct bkey_i *r_v;
347 struct bkey_i_reflink_p *r_p;
348 __le64 *refcount;
349 int ret;
350
351 if (orig->k.type == KEY_TYPE_inline_data)
352 bch2_check_set_feature(c, BCH_FEATURE_reflink_inline_data);
353
354 bch2_trans_iter_init(trans, &reflink_iter, BTREE_ID_reflink, POS_MAX,
355 BTREE_ITER_intent);
356 k = bch2_btree_iter_peek_prev(&reflink_iter);
357 ret = bkey_err(k);
358 if (ret)
359 goto err;
360
361 r_v = bch2_trans_kmalloc(trans, sizeof(__le64) + bkey_bytes(&orig->k));
362 ret = PTR_ERR_OR_ZERO(r_v);
363 if (ret)
364 goto err;
365
366 bkey_init(&r_v->k);
367 r_v->k.type = bkey_type_to_indirect(&orig->k);
368 r_v->k.p = reflink_iter.pos;
369 bch2_key_resize(&r_v->k, orig->k.size);
370 r_v->k.version = orig->k.version;
371
372 set_bkey_val_bytes(&r_v->k, sizeof(__le64) + bkey_val_bytes(&orig->k));
373
374 refcount = bkey_refcount(bkey_i_to_s(r_v));
375 *refcount = 0;
376 memcpy(refcount + 1, &orig->v, bkey_val_bytes(&orig->k));
377
378 ret = bch2_trans_update(trans, &reflink_iter, r_v, 0);
379 if (ret)
380 goto err;
381
382 /*
383 * orig is in a bkey_buf which statically allocates 5 64s for the val,
384 * so we know it will be big enough:
385 */
386 orig->k.type = KEY_TYPE_reflink_p;
387 r_p = bkey_i_to_reflink_p(orig);
388 set_bkey_val_bytes(&r_p->k, sizeof(r_p->v));
389
390 /* FORTIFY_SOURCE is broken here, and doesn't provide unsafe_memset() */
391 #if !defined(__NO_FORTIFY) && defined(__OPTIMIZE__) && defined(CONFIG_FORTIFY_SOURCE)
392 __underlying_memset(&r_p->v, 0, sizeof(r_p->v));
393 #else
394 memset(&r_p->v, 0, sizeof(r_p->v));
395 #endif
396
397 r_p->v.idx = cpu_to_le64(bkey_start_offset(&r_v->k));
398
399 ret = bch2_trans_update(trans, extent_iter, &r_p->k_i,
400 BTREE_UPDATE_internal_snapshot_node);
401 err:
402 bch2_trans_iter_exit(trans, &reflink_iter);
403
404 return ret;
405 }
406
407 static struct bkey_s_c get_next_src(struct btree_iter *iter, struct bpos end)
408 {
409 struct bkey_s_c k;
410 int ret;
411
412 for_each_btree_key_upto_continue_norestart(*iter, end, 0, k, ret) {
413 if (bkey_extent_is_unwritten(k))
414 continue;
415
416 if (bkey_extent_is_data(k.k))
417 return k;
418 }
419
420 if (bkey_ge(iter->pos, end))
421 bch2_btree_iter_set_pos(iter, end);
422 return ret ? bkey_s_c_err(ret) : bkey_s_c_null;
423 }
424
425 s64 bch2_remap_range(struct bch_fs *c,
426 subvol_inum dst_inum, u64 dst_offset,
427 subvol_inum src_inum, u64 src_offset,
428 u64 remap_sectors,
429 u64 new_i_size, s64 *i_sectors_delta)
430 {
431 struct btree_trans *trans;
432 struct btree_iter dst_iter, src_iter;
433 struct bkey_s_c src_k;
434 struct bkey_buf new_dst, new_src;
435 struct bpos dst_start = POS(dst_inum.inum, dst_offset);
436 struct bpos src_start = POS(src_inum.inum, src_offset);
437 struct bpos dst_end = dst_start, src_end = src_start;
438 struct bch_io_opts opts;
439 struct bpos src_want;
440 u64 dst_done = 0;
441 u32 dst_snapshot, src_snapshot;
442 int ret = 0, ret2 = 0;
443
444 if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_reflink))
445 return -BCH_ERR_erofs_no_writes;
446
447 bch2_check_set_feature(c, BCH_FEATURE_reflink);
448
449 dst_end.offset += remap_sectors;
450 src_end.offset += remap_sectors;
451
452 bch2_bkey_buf_init(&new_dst);
453 bch2_bkey_buf_init(&new_src);
454 trans = bch2_trans_get(c);
455
456 ret = bch2_inum_opts_get(trans, src_inum, &opts);
457 if (ret)
458 goto err;
459
460 bch2_trans_iter_init(trans, &src_iter, BTREE_ID_extents, src_start,
461 BTREE_ITER_intent);
462 bch2_trans_iter_init(trans, &dst_iter, BTREE_ID_extents, dst_start,
463 BTREE_ITER_intent);
464
465 while ((ret == 0 ||
466 bch2_err_matches(ret, BCH_ERR_transaction_restart)) &&
467 bkey_lt(dst_iter.pos, dst_end)) {
468 struct disk_reservation disk_res = { 0 };
469
470 bch2_trans_begin(trans);
471
472 if (fatal_signal_pending(current)) {
473 ret = -EINTR;
474 break;
475 }
476
477 ret = bch2_subvolume_get_snapshot(trans, src_inum.subvol,
478 &src_snapshot);
479 if (ret)
480 continue;
481
482 bch2_btree_iter_set_snapshot(&src_iter, src_snapshot);
483
484 ret = bch2_subvolume_get_snapshot(trans, dst_inum.subvol,
485 &dst_snapshot);
486 if (ret)
487 continue;
488
489 bch2_btree_iter_set_snapshot(&dst_iter, dst_snapshot);
490
491 if (dst_inum.inum < src_inum.inum) {
492 /* Avoid some lock cycle transaction restarts */
493 ret = bch2_btree_iter_traverse(&dst_iter);
494 if (ret)
495 continue;
496 }
497
498 dst_done = dst_iter.pos.offset - dst_start.offset;
499 src_want = POS(src_start.inode, src_start.offset + dst_done);
500 bch2_btree_iter_set_pos(&src_iter, src_want);
501
502 src_k = get_next_src(&src_iter, src_end);
503 ret = bkey_err(src_k);
504 if (ret)
505 continue;
506
507 if (bkey_lt(src_want, src_iter.pos)) {
508 ret = bch2_fpunch_at(trans, &dst_iter, dst_inum,
509 min(dst_end.offset,
510 dst_iter.pos.offset +
511 src_iter.pos.offset - src_want.offset),
512 i_sectors_delta);
513 continue;
514 }
515
516 if (src_k.k->type != KEY_TYPE_reflink_p) {
517 bch2_btree_iter_set_pos_to_extent_start(&src_iter);
518
519 bch2_bkey_buf_reassemble(&new_src, c, src_k);
520 src_k = bkey_i_to_s_c(new_src.k);
521
522 ret = bch2_make_extent_indirect(trans, &src_iter,
523 new_src.k);
524 if (ret)
525 continue;
526
527 BUG_ON(src_k.k->type != KEY_TYPE_reflink_p);
528 }
529
530 if (src_k.k->type == KEY_TYPE_reflink_p) {
531 struct bkey_s_c_reflink_p src_p =
532 bkey_s_c_to_reflink_p(src_k);
533 struct bkey_i_reflink_p *dst_p =
534 bkey_reflink_p_init(new_dst.k);
535
536 u64 offset = le64_to_cpu(src_p.v->idx) +
537 (src_want.offset -
538 bkey_start_offset(src_k.k));
539
540 dst_p->v.idx = cpu_to_le64(offset);
541 } else {
542 BUG();
543 }
544
545 new_dst.k->k.p = dst_iter.pos;
546 bch2_key_resize(&new_dst.k->k,
547 min(src_k.k->p.offset - src_want.offset,
548 dst_end.offset - dst_iter.pos.offset));
549
550 ret = bch2_bkey_set_needs_rebalance(c, new_dst.k, &opts) ?:
551 bch2_extent_update(trans, dst_inum, &dst_iter,
552 new_dst.k, &disk_res,
553 new_i_size, i_sectors_delta,
554 true);
555 bch2_disk_reservation_put(c, &disk_res);
556 }
557 bch2_trans_iter_exit(trans, &dst_iter);
558 bch2_trans_iter_exit(trans, &src_iter);
559
560 BUG_ON(!ret && !bkey_eq(dst_iter.pos, dst_end));
561 BUG_ON(bkey_gt(dst_iter.pos, dst_end));
562
563 dst_done = dst_iter.pos.offset - dst_start.offset;
564 new_i_size = min(dst_iter.pos.offset << 9, new_i_size);
565
566 do {
567 struct bch_inode_unpacked inode_u;
568 struct btree_iter inode_iter = { NULL };
569
570 bch2_trans_begin(trans);
571
572 ret2 = bch2_inode_peek(trans, &inode_iter, &inode_u,
573 dst_inum, BTREE_ITER_intent);
574
575 if (!ret2 &&
576 inode_u.bi_size < new_i_size) {
577 inode_u.bi_size = new_i_size;
578 ret2 = bch2_inode_write(trans, &inode_iter, &inode_u) ?:
579 bch2_trans_commit(trans, NULL, NULL,
580 BCH_TRANS_COMMIT_no_enospc);
581 }
582
583 bch2_trans_iter_exit(trans, &inode_iter);
584 } while (bch2_err_matches(ret2, BCH_ERR_transaction_restart));
585 err:
586 bch2_trans_put(trans);
587 bch2_bkey_buf_exit(&new_src, c);
588 bch2_bkey_buf_exit(&new_dst, c);
589
590 bch2_write_ref_put(c, BCH_WRITE_REF_reflink);
591
592 return dst_done ?: ret ?: ret2;
593 }
594
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.