Loading...
Note: File does not exist in v4.6.
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
18static 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
32int bch2_reflink_p_invalid(struct bch_fs *c, struct bkey_s_c k,
33 enum bkey_invalid_flags flags,
34 struct printbuf *err)
35{
36 struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
37 int ret = 0;
38
39 bkey_fsck_err_on(le64_to_cpu(p.v->idx) < le32_to_cpu(p.v->front_pad),
40 c, err, reflink_p_front_pad_bad,
41 "idx < front_pad (%llu < %u)",
42 le64_to_cpu(p.v->idx), le32_to_cpu(p.v->front_pad));
43fsck_err:
44 return ret;
45}
46
47void bch2_reflink_p_to_text(struct printbuf *out, struct bch_fs *c,
48 struct bkey_s_c k)
49{
50 struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
51
52 prt_printf(out, "idx %llu front_pad %u back_pad %u",
53 le64_to_cpu(p.v->idx),
54 le32_to_cpu(p.v->front_pad),
55 le32_to_cpu(p.v->back_pad));
56}
57
58bool bch2_reflink_p_merge(struct bch_fs *c, struct bkey_s _l, struct bkey_s_c _r)
59{
60 struct bkey_s_reflink_p l = bkey_s_to_reflink_p(_l);
61 struct bkey_s_c_reflink_p r = bkey_s_c_to_reflink_p(_r);
62
63 /*
64 * Disabled for now, the triggers code needs to be reworked for merging
65 * of reflink pointers to work:
66 */
67 return false;
68
69 if (le64_to_cpu(l.v->idx) + l.k->size != le64_to_cpu(r.v->idx))
70 return false;
71
72 bch2_key_resize(l.k, l.k->size + r.k->size);
73 return true;
74}
75
76static int trans_trigger_reflink_p_segment(struct btree_trans *trans,
77 struct bkey_s_c_reflink_p p,
78 u64 *idx, unsigned flags)
79{
80 struct bch_fs *c = trans->c;
81 struct btree_iter iter;
82 struct bkey_i *k;
83 __le64 *refcount;
84 int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
85 struct printbuf buf = PRINTBUF;
86 int ret;
87
88 k = bch2_bkey_get_mut_noupdate(trans, &iter,
89 BTREE_ID_reflink, POS(0, *idx),
90 BTREE_ITER_WITH_UPDATES);
91 ret = PTR_ERR_OR_ZERO(k);
92 if (ret)
93 goto err;
94
95 refcount = bkey_refcount(bkey_i_to_s(k));
96 if (!refcount) {
97 bch2_bkey_val_to_text(&buf, c, p.s_c);
98 bch2_trans_inconsistent(trans,
99 "nonexistent indirect extent at %llu while marking\n %s",
100 *idx, buf.buf);
101 ret = -EIO;
102 goto err;
103 }
104
105 if (!*refcount && (flags & BTREE_TRIGGER_OVERWRITE)) {
106 bch2_bkey_val_to_text(&buf, c, p.s_c);
107 bch2_trans_inconsistent(trans,
108 "indirect extent refcount underflow at %llu while marking\n %s",
109 *idx, buf.buf);
110 ret = -EIO;
111 goto err;
112 }
113
114 if (flags & BTREE_TRIGGER_INSERT) {
115 struct bch_reflink_p *v = (struct bch_reflink_p *) p.v;
116 u64 pad;
117
118 pad = max_t(s64, le32_to_cpu(v->front_pad),
119 le64_to_cpu(v->idx) - bkey_start_offset(&k->k));
120 BUG_ON(pad > U32_MAX);
121 v->front_pad = cpu_to_le32(pad);
122
123 pad = max_t(s64, le32_to_cpu(v->back_pad),
124 k->k.p.offset - p.k->size - le64_to_cpu(v->idx));
125 BUG_ON(pad > U32_MAX);
126 v->back_pad = cpu_to_le32(pad);
127 }
128
129 le64_add_cpu(refcount, add);
130
131 bch2_btree_iter_set_pos_to_extent_start(&iter);
132 ret = bch2_trans_update(trans, &iter, k, 0);
133 if (ret)
134 goto err;
135
136 *idx = k->k.p.offset;
137err:
138 bch2_trans_iter_exit(trans, &iter);
139 printbuf_exit(&buf);
140 return ret;
141}
142
143static s64 gc_trigger_reflink_p_segment(struct btree_trans *trans,
144 struct bkey_s_c_reflink_p p,
145 u64 *idx, unsigned flags, 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 r->refcount += add;
167 *idx = r->offset;
168 return 0;
169not_found:
170 if (fsck_err(c, reflink_p_to_missing_reflink_v,
171 "pointer to missing indirect extent\n"
172 " %s\n"
173 " missing range %llu-%llu",
174 (bch2_bkey_val_to_text(&buf, c, p.s_c), buf.buf),
175 *idx, next_idx)) {
176 struct bkey_i *update = bch2_bkey_make_mut_noupdate(trans, p.s_c);
177 ret = PTR_ERR_OR_ZERO(update);
178 if (ret)
179 goto err;
180
181 if (next_idx <= start) {
182 bkey_i_to_reflink_p(update)->v.front_pad = cpu_to_le32(start - next_idx);
183 } else if (*idx >= end) {
184 bkey_i_to_reflink_p(update)->v.back_pad = cpu_to_le32(*idx - end);
185 } else {
186 bkey_error_init(update);
187 update->k.p = p.k->p;
188 update->k.p.offset = next_idx;
189 update->k.size = next_idx - *idx;
190 set_bkey_val_u64s(&update->k, 0);
191 }
192
193 ret = bch2_btree_insert_trans(trans, BTREE_ID_extents, update, BTREE_TRIGGER_NORUN);
194 }
195
196 *idx = next_idx;
197err:
198fsck_err:
199 printbuf_exit(&buf);
200 return ret;
201}
202
203static int __trigger_reflink_p(struct btree_trans *trans,
204 enum btree_id btree_id, unsigned level,
205 struct bkey_s_c k, unsigned flags)
206{
207 struct bch_fs *c = trans->c;
208 struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
209 int ret = 0;
210
211 u64 idx = le64_to_cpu(p.v->idx) - le32_to_cpu(p.v->front_pad);
212 u64 end = le64_to_cpu(p.v->idx) + p.k->size + le32_to_cpu(p.v->back_pad);
213
214 if (flags & BTREE_TRIGGER_TRANSACTIONAL) {
215 while (idx < end && !ret)
216 ret = trans_trigger_reflink_p_segment(trans, p, &idx, flags);
217 }
218
219 if (flags & BTREE_TRIGGER_GC) {
220 size_t l = 0, r = c->reflink_gc_nr;
221
222 while (l < r) {
223 size_t m = l + (r - l) / 2;
224 struct reflink_gc *ref = genradix_ptr(&c->reflink_gc_table, m);
225 if (ref->offset <= idx)
226 l = m + 1;
227 else
228 r = m;
229 }
230
231 while (idx < end && !ret)
232 ret = gc_trigger_reflink_p_segment(trans, p, &idx, flags, l++);
233 }
234
235 return ret;
236}
237
238int bch2_trigger_reflink_p(struct btree_trans *trans,
239 enum btree_id btree_id, unsigned level,
240 struct bkey_s_c old,
241 struct bkey_s new,
242 unsigned flags)
243{
244 if ((flags & BTREE_TRIGGER_TRANSACTIONAL) &&
245 (flags & BTREE_TRIGGER_INSERT)) {
246 struct bch_reflink_p *v = bkey_s_to_reflink_p(new).v;
247
248 v->front_pad = v->back_pad = 0;
249 }
250
251 return trigger_run_overwrite_then_insert(__trigger_reflink_p, trans, btree_id, level, old, new, flags);
252}
253
254/* indirect extents */
255
256int bch2_reflink_v_invalid(struct bch_fs *c, struct bkey_s_c k,
257 enum bkey_invalid_flags flags,
258 struct printbuf *err)
259{
260 return bch2_bkey_ptrs_invalid(c, k, flags, err);
261}
262
263void bch2_reflink_v_to_text(struct printbuf *out, struct bch_fs *c,
264 struct bkey_s_c k)
265{
266 struct bkey_s_c_reflink_v r = bkey_s_c_to_reflink_v(k);
267
268 prt_printf(out, "refcount: %llu ", le64_to_cpu(r.v->refcount));
269
270 bch2_bkey_ptrs_to_text(out, c, k);
271}
272
273#if 0
274Currently disabled, needs to be debugged:
275
276bool bch2_reflink_v_merge(struct bch_fs *c, struct bkey_s _l, struct bkey_s_c _r)
277{
278 struct bkey_s_reflink_v l = bkey_s_to_reflink_v(_l);
279 struct bkey_s_c_reflink_v r = bkey_s_c_to_reflink_v(_r);
280
281 return l.v->refcount == r.v->refcount && bch2_extent_merge(c, _l, _r);
282}
283#endif
284
285static inline void check_indirect_extent_deleting(struct bkey_s new, unsigned *flags)
286{
287 if ((*flags & BTREE_TRIGGER_INSERT) && !*bkey_refcount(new)) {
288 new.k->type = KEY_TYPE_deleted;
289 new.k->size = 0;
290 set_bkey_val_u64s(new.k, 0);
291 *flags &= ~BTREE_TRIGGER_INSERT;
292 }
293}
294
295int bch2_trigger_reflink_v(struct btree_trans *trans,
296 enum btree_id btree_id, unsigned level,
297 struct bkey_s_c old, struct bkey_s new,
298 unsigned flags)
299{
300 if ((flags & BTREE_TRIGGER_TRANSACTIONAL) &&
301 (flags & BTREE_TRIGGER_INSERT))
302 check_indirect_extent_deleting(new, &flags);
303
304 return bch2_trigger_extent(trans, btree_id, level, old, new, flags);
305}
306
307/* indirect inline data */
308
309int bch2_indirect_inline_data_invalid(struct bch_fs *c, struct bkey_s_c k,
310 enum bkey_invalid_flags flags,
311 struct printbuf *err)
312{
313 return 0;
314}
315
316void bch2_indirect_inline_data_to_text(struct printbuf *out,
317 struct bch_fs *c, struct bkey_s_c k)
318{
319 struct bkey_s_c_indirect_inline_data d = bkey_s_c_to_indirect_inline_data(k);
320 unsigned datalen = bkey_inline_data_bytes(k.k);
321
322 prt_printf(out, "refcount %llu datalen %u: %*phN",
323 le64_to_cpu(d.v->refcount), datalen,
324 min(datalen, 32U), d.v->data);
325}
326
327int bch2_trigger_indirect_inline_data(struct btree_trans *trans,
328 enum btree_id btree_id, unsigned level,
329 struct bkey_s_c old, struct bkey_s new,
330 unsigned flags)
331{
332 check_indirect_extent_deleting(new, &flags);
333
334 return 0;
335}
336
337static int bch2_make_extent_indirect(struct btree_trans *trans,
338 struct btree_iter *extent_iter,
339 struct bkey_i *orig)
340{
341 struct bch_fs *c = trans->c;
342 struct btree_iter reflink_iter = { NULL };
343 struct bkey_s_c k;
344 struct bkey_i *r_v;
345 struct bkey_i_reflink_p *r_p;
346 __le64 *refcount;
347 int ret;
348
349 if (orig->k.type == KEY_TYPE_inline_data)
350 bch2_check_set_feature(c, BCH_FEATURE_reflink_inline_data);
351
352 bch2_trans_iter_init(trans, &reflink_iter, BTREE_ID_reflink, POS_MAX,
353 BTREE_ITER_INTENT);
354 k = bch2_btree_iter_peek_prev(&reflink_iter);
355 ret = bkey_err(k);
356 if (ret)
357 goto err;
358
359 r_v = bch2_trans_kmalloc(trans, sizeof(__le64) + bkey_bytes(&orig->k));
360 ret = PTR_ERR_OR_ZERO(r_v);
361 if (ret)
362 goto err;
363
364 bkey_init(&r_v->k);
365 r_v->k.type = bkey_type_to_indirect(&orig->k);
366 r_v->k.p = reflink_iter.pos;
367 bch2_key_resize(&r_v->k, orig->k.size);
368 r_v->k.version = orig->k.version;
369
370 set_bkey_val_bytes(&r_v->k, sizeof(__le64) + bkey_val_bytes(&orig->k));
371
372 refcount = bkey_refcount(bkey_i_to_s(r_v));
373 *refcount = 0;
374 memcpy(refcount + 1, &orig->v, bkey_val_bytes(&orig->k));
375
376 ret = bch2_trans_update(trans, &reflink_iter, r_v, 0);
377 if (ret)
378 goto err;
379
380 /*
381 * orig is in a bkey_buf which statically allocates 5 64s for the val,
382 * so we know it will be big enough:
383 */
384 orig->k.type = KEY_TYPE_reflink_p;
385 r_p = bkey_i_to_reflink_p(orig);
386 set_bkey_val_bytes(&r_p->k, sizeof(r_p->v));
387
388 /* FORTIFY_SOURCE is broken here, and doesn't provide unsafe_memset() */
389#if !defined(__NO_FORTIFY) && defined(__OPTIMIZE__) && defined(CONFIG_FORTIFY_SOURCE)
390 __underlying_memset(&r_p->v, 0, sizeof(r_p->v));
391#else
392 memset(&r_p->v, 0, sizeof(r_p->v));
393#endif
394
395 r_p->v.idx = cpu_to_le64(bkey_start_offset(&r_v->k));
396
397 ret = bch2_trans_update(trans, extent_iter, &r_p->k_i,
398 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
399err:
400 bch2_trans_iter_exit(trans, &reflink_iter);
401
402 return ret;
403}
404
405static struct bkey_s_c get_next_src(struct btree_iter *iter, struct bpos end)
406{
407 struct bkey_s_c k;
408 int ret;
409
410 for_each_btree_key_upto_continue_norestart(*iter, end, 0, k, ret) {
411 if (bkey_extent_is_unwritten(k))
412 continue;
413
414 if (bkey_extent_is_data(k.k))
415 return k;
416 }
417
418 if (bkey_ge(iter->pos, end))
419 bch2_btree_iter_set_pos(iter, end);
420 return ret ? bkey_s_c_err(ret) : bkey_s_c_null;
421}
422
423s64 bch2_remap_range(struct bch_fs *c,
424 subvol_inum dst_inum, u64 dst_offset,
425 subvol_inum src_inum, u64 src_offset,
426 u64 remap_sectors,
427 u64 new_i_size, s64 *i_sectors_delta)
428{
429 struct btree_trans *trans;
430 struct btree_iter dst_iter, src_iter;
431 struct bkey_s_c src_k;
432 struct bkey_buf new_dst, new_src;
433 struct bpos dst_start = POS(dst_inum.inum, dst_offset);
434 struct bpos src_start = POS(src_inum.inum, src_offset);
435 struct bpos dst_end = dst_start, src_end = src_start;
436 struct bch_io_opts opts;
437 struct bpos src_want;
438 u64 dst_done = 0;
439 u32 dst_snapshot, src_snapshot;
440 int ret = 0, ret2 = 0;
441
442 if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_reflink))
443 return -BCH_ERR_erofs_no_writes;
444
445 bch2_check_set_feature(c, BCH_FEATURE_reflink);
446
447 dst_end.offset += remap_sectors;
448 src_end.offset += remap_sectors;
449
450 bch2_bkey_buf_init(&new_dst);
451 bch2_bkey_buf_init(&new_src);
452 trans = bch2_trans_get(c);
453
454 ret = bch2_inum_opts_get(trans, src_inum, &opts);
455 if (ret)
456 goto err;
457
458 bch2_trans_iter_init(trans, &src_iter, BTREE_ID_extents, src_start,
459 BTREE_ITER_INTENT);
460 bch2_trans_iter_init(trans, &dst_iter, BTREE_ID_extents, dst_start,
461 BTREE_ITER_INTENT);
462
463 while ((ret == 0 ||
464 bch2_err_matches(ret, BCH_ERR_transaction_restart)) &&
465 bkey_lt(dst_iter.pos, dst_end)) {
466 struct disk_reservation disk_res = { 0 };
467
468 bch2_trans_begin(trans);
469
470 if (fatal_signal_pending(current)) {
471 ret = -EINTR;
472 break;
473 }
474
475 ret = bch2_subvolume_get_snapshot(trans, src_inum.subvol,
476 &src_snapshot);
477 if (ret)
478 continue;
479
480 bch2_btree_iter_set_snapshot(&src_iter, src_snapshot);
481
482 ret = bch2_subvolume_get_snapshot(trans, dst_inum.subvol,
483 &dst_snapshot);
484 if (ret)
485 continue;
486
487 bch2_btree_iter_set_snapshot(&dst_iter, dst_snapshot);
488
489 if (dst_inum.inum < src_inum.inum) {
490 /* Avoid some lock cycle transaction restarts */
491 ret = bch2_btree_iter_traverse(&dst_iter);
492 if (ret)
493 continue;
494 }
495
496 dst_done = dst_iter.pos.offset - dst_start.offset;
497 src_want = POS(src_start.inode, src_start.offset + dst_done);
498 bch2_btree_iter_set_pos(&src_iter, src_want);
499
500 src_k = get_next_src(&src_iter, src_end);
501 ret = bkey_err(src_k);
502 if (ret)
503 continue;
504
505 if (bkey_lt(src_want, src_iter.pos)) {
506 ret = bch2_fpunch_at(trans, &dst_iter, dst_inum,
507 min(dst_end.offset,
508 dst_iter.pos.offset +
509 src_iter.pos.offset - src_want.offset),
510 i_sectors_delta);
511 continue;
512 }
513
514 if (src_k.k->type != KEY_TYPE_reflink_p) {
515 bch2_btree_iter_set_pos_to_extent_start(&src_iter);
516
517 bch2_bkey_buf_reassemble(&new_src, c, src_k);
518 src_k = bkey_i_to_s_c(new_src.k);
519
520 ret = bch2_make_extent_indirect(trans, &src_iter,
521 new_src.k);
522 if (ret)
523 continue;
524
525 BUG_ON(src_k.k->type != KEY_TYPE_reflink_p);
526 }
527
528 if (src_k.k->type == KEY_TYPE_reflink_p) {
529 struct bkey_s_c_reflink_p src_p =
530 bkey_s_c_to_reflink_p(src_k);
531 struct bkey_i_reflink_p *dst_p =
532 bkey_reflink_p_init(new_dst.k);
533
534 u64 offset = le64_to_cpu(src_p.v->idx) +
535 (src_want.offset -
536 bkey_start_offset(src_k.k));
537
538 dst_p->v.idx = cpu_to_le64(offset);
539 } else {
540 BUG();
541 }
542
543 new_dst.k->k.p = dst_iter.pos;
544 bch2_key_resize(&new_dst.k->k,
545 min(src_k.k->p.offset - src_want.offset,
546 dst_end.offset - dst_iter.pos.offset));
547
548 ret = bch2_bkey_set_needs_rebalance(c, new_dst.k, &opts) ?:
549 bch2_extent_update(trans, dst_inum, &dst_iter,
550 new_dst.k, &disk_res,
551 new_i_size, i_sectors_delta,
552 true);
553 bch2_disk_reservation_put(c, &disk_res);
554 }
555 bch2_trans_iter_exit(trans, &dst_iter);
556 bch2_trans_iter_exit(trans, &src_iter);
557
558 BUG_ON(!ret && !bkey_eq(dst_iter.pos, dst_end));
559 BUG_ON(bkey_gt(dst_iter.pos, dst_end));
560
561 dst_done = dst_iter.pos.offset - dst_start.offset;
562 new_i_size = min(dst_iter.pos.offset << 9, new_i_size);
563
564 do {
565 struct bch_inode_unpacked inode_u;
566 struct btree_iter inode_iter = { NULL };
567
568 bch2_trans_begin(trans);
569
570 ret2 = bch2_inode_peek(trans, &inode_iter, &inode_u,
571 dst_inum, BTREE_ITER_INTENT);
572
573 if (!ret2 &&
574 inode_u.bi_size < new_i_size) {
575 inode_u.bi_size = new_i_size;
576 ret2 = bch2_inode_write(trans, &inode_iter, &inode_u) ?:
577 bch2_trans_commit(trans, NULL, NULL,
578 BCH_TRANS_COMMIT_no_enospc);
579 }
580
581 bch2_trans_iter_exit(trans, &inode_iter);
582 } while (bch2_err_matches(ret2, BCH_ERR_transaction_restart));
583err:
584 bch2_trans_put(trans);
585 bch2_bkey_buf_exit(&new_src, c);
586 bch2_bkey_buf_exit(&new_dst, c);
587
588 bch2_write_ref_put(c, BCH_WRITE_REF_reflink);
589
590 return dst_done ?: ret ?: ret2;
591}