projects / linux-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
Merge tag 'soc-drivers-6.9' of git://git.kernel.org/pub/scm/linux/kernel/git/soc/soc
[linux-block.git] / fs / bcachefs / btree_iter.c
CommitLineData
1c6fdbd8
KO		 1// SPDX-License-Identifier: GPL-2.0
2
3#include "bcachefs.h"
4#include "bkey_methods.h"
07a1006a 5#include "bkey_buf.h"
1c6fdbd8
KO		 6#include "btree_cache.h"
7#include "btree_iter.h"
401585fe 8#include "btree_journal_iter.h"
2ca88e5a 9#include "btree_key_cache.h"
1c6fdbd8 10#include "btree_locking.h"
57b0b3db 11#include "btree_update.h"
1c6fdbd8 12#include "debug.h"
50dc0f69 13#include "error.h"
1c6fdbd8 14#include "extents.h"
2ca88e5a 15#include "journal.h"
67997234 16#include "journal_io.h"
35d5aff2 17#include "replicas.h"
8e877caa 18#include "snapshot.h"
1c6fdbd8
KO		 19#include "trace.h"
20
a0cb8d78 21#include <linux/random.h>
1c6fdbd8
KO		 22#include <linux/prefetch.h>
23
67e0dd8f 24static inline void btree_path_list_remove(struct btree_trans *, struct btree_path *);
07f383c7
KO		 25static inline void btree_path_list_add(struct btree_trans *,
26 btree_path_idx_t, btree_path_idx_t);
67e0dd8f 27
ee2c6ea7
KO		 28static inline unsigned long btree_iter_ip_allocated(struct btree_iter *iter)
29{
94c69faf 30#ifdef TRACK_PATH_ALLOCATED
ee2c6ea7
KO		 31 return iter->ip_allocated;
32#else
33 return 0;
34#endif
35}
36
788cc25d 37static btree_path_idx_t btree_path_alloc(struct btree_trans *, btree_path_idx_t);
79904fa2 38static void bch2_trans_srcu_lock(struct btree_trans *);
67e0dd8f
KO		 39
40static inline int __btree_path_cmp(const struct btree_path *l,
41 enum btree_id r_btree_id,
42 bool r_cached,
43 struct bpos r_pos,
44 unsigned r_level)
0423fb71 45{
7abda8c1
KO		 46 /*
47 * Must match lock ordering as defined by __bch2_btree_node_lock:
48 */
67e0dd8f 49 return cmp_int(l->btree_id, r_btree_id) ?:
32b26e8c 50 cmp_int((int) l->cached, (int) r_cached) ?:
67e0dd8f
KO		 51 bpos_cmp(l->pos, r_pos) ?:
52 -cmp_int(l->level, r_level);
53}
54
55static inline int btree_path_cmp(const struct btree_path *l,
56 const struct btree_path *r)
57{
58 return __btree_path_cmp(l, r->btree_id, r->cached, r->pos, r->level);
0423fb71
KO		 59}
60
e751c01a
KO		 61static inline struct bpos bkey_successor(struct btree_iter *iter, struct bpos p)
62{
e751c01a
KO		 63 /* Are we iterating over keys in all snapshots? */
64 if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
65 p = bpos_successor(p);
66 } else {
67 p = bpos_nosnap_successor(p);
68 p.snapshot = iter->snapshot;
69 }
70
71 return p;
72}
73
74static inline struct bpos bkey_predecessor(struct btree_iter *iter, struct bpos p)
75{
e751c01a
KO		 76 /* Are we iterating over keys in all snapshots? */
77 if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
78 p = bpos_predecessor(p);
79 } else {
80 p = bpos_nosnap_predecessor(p);
81 p.snapshot = iter->snapshot;
82 }
83
84 return p;
85}
86
9626aeb1 87static inline struct bpos btree_iter_search_key(struct btree_iter *iter)
a00fd8c5 88{
9626aeb1 89 struct bpos pos = iter->pos;
cbdf24ce 90
9626aeb1 91 if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
e88a75eb 92 !bkey_eq(pos, POS_MAX))
e751c01a 93 pos = bkey_successor(iter, pos);
9626aeb1 94 return pos;
a00fd8c5
KO		 95}
96
67e0dd8f 97static inline bool btree_path_pos_before_node(struct btree_path *path,
e3ecf4f5
KO		 98 struct btree *b)
99{
e88a75eb 100 return bpos_lt(path->pos, b->data->min_key);
e3ecf4f5
KO		 101}
102
67e0dd8f 103static inline bool btree_path_pos_after_node(struct btree_path *path,
e3ecf4f5
KO		 104 struct btree *b)
105{
e88a75eb 106 return bpos_gt(path->pos, b->key.k.p);
e3ecf4f5
KO		 107}
108
67e0dd8f 109static inline bool btree_path_pos_in_node(struct btree_path *path,
e3ecf4f5
KO		 110 struct btree *b)
111{
67e0dd8f
KO		 112 return path->btree_id == b->c.btree_id &&
113 !btree_path_pos_before_node(path, b) &&
114 !btree_path_pos_after_node(path, b);
e3ecf4f5
KO		 115}
116
1c6fdbd8
KO		 117/* Btree iterator: */
118
119#ifdef CONFIG_BCACHEFS_DEBUG
120
67e0dd8f
KO		 121static void bch2_btree_path_verify_cached(struct btree_trans *trans,
122 struct btree_path *path)
d211b408
KO		 123{
124 struct bkey_cached *ck;
67e0dd8f 125 bool locked = btree_node_locked(path, 0);
d211b408 126
67e0dd8f 127 if (!bch2_btree_node_relock(trans, path, 0))
d211b408
KO		 128 return;
129
67e0dd8f
KO		 130 ck = (void *) path->l[0].b;
131 BUG_ON(ck->key.btree_id != path->btree_id ||
e88a75eb 132 !bkey_eq(ck->key.pos, path->pos));
d211b408
KO		 133
134 if (!locked)
8bfe14e8 135 btree_node_unlock(trans, path, 0);
d211b408
KO		 136}
137
67e0dd8f
KO		 138static void bch2_btree_path_verify_level(struct btree_trans *trans,
139 struct btree_path *path, unsigned level)
1c6fdbd8 140{
67e0dd8f 141 struct btree_path_level *l;
4ce41957
KO		 142 struct btree_node_iter tmp;
143 bool locked;
2dac0eae 144 struct bkey_packed *p, *k;
fa8e94fa
KO		 145 struct printbuf buf1 = PRINTBUF;
146 struct printbuf buf2 = PRINTBUF;
147 struct printbuf buf3 = PRINTBUF;
2dac0eae 148 const char *msg;
1c6fdbd8 149
29364f34 150 if (!bch2_debug_check_iterators)
f13f5a8c
KO		 151 return;
152
67e0dd8f 153 l = &path->l[level];
4ce41957 154 tmp = l->iter;
67e0dd8f 155 locked = btree_node_locked(path, level);
4ce41957 156
67e0dd8f 157 if (path->cached) {
d211b408 158 if (!level)
67e0dd8f 159 bch2_btree_path_verify_cached(trans, path);
d211b408
KO		 160 return;
161 }
162
67e0dd8f 163 if (!btree_path_node(path, level))
2dac0eae
KO		 164 return;
165
e9174370 166 if (!bch2_btree_node_relock_notrace(trans, path, level))
271a3d3a
KO		 167 return;
168
67e0dd8f 169 BUG_ON(!btree_path_pos_in_node(path, l->b));
2dac0eae 170
2dac0eae 171 bch2_btree_node_iter_verify(&l->iter, l->b);
1c6fdbd8
KO		 172
173 /*
1ae29c1f 174 * For interior nodes, the iterator will have skipped past deleted keys:
1c6fdbd8 175 */
1ae29c1f 176 p = level
c052cf82 177 ? bch2_btree_node_iter_prev(&tmp, l->b)
2dac0eae
KO		 178 : bch2_btree_node_iter_prev_all(&tmp, l->b);
179 k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
1c6fdbd8 180
67e0dd8f 181 if (p && bkey_iter_pos_cmp(l->b, p, &path->pos) >= 0) {
2dac0eae
KO		 182 msg = "before";
183 goto err;
1c6fdbd8
KO		 184 }
185
67e0dd8f 186 if (k && bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
2dac0eae
KO		 187 msg = "after";
188 goto err;
189 }
f21566f1 190
2dac0eae 191 if (!locked)
8bfe14e8 192 btree_node_unlock(trans, path, level);
2dac0eae
KO		 193 return;
194err:
fa8e94fa 195 bch2_bpos_to_text(&buf1, path->pos);
2dac0eae
KO		 196
197 if (p) {
198 struct bkey uk = bkey_unpack_key(l->b, p);
a1019576 199
fa8e94fa
KO		 200 bch2_bkey_to_text(&buf2, &uk);
201 } else {
401ec4db 202 prt_printf(&buf2, "(none)");
2dac0eae 203 }
1c6fdbd8 204
2dac0eae
KO		 205 if (k) {
206 struct bkey uk = bkey_unpack_key(l->b, k);
a1019576 207
fa8e94fa
KO		 208 bch2_bkey_to_text(&buf3, &uk);
209 } else {
401ec4db 210 prt_printf(&buf3, "(none)");
1c6fdbd8 211 }
2dac0eae 212
67e0dd8f
KO		 213 panic("path should be %s key at level %u:\n"
214 "path pos %s\n"
2dac0eae
KO		 215 "prev key %s\n"
216 "cur key %s\n",
fa8e94fa 217 msg, level, buf1.buf, buf2.buf, buf3.buf);
1c6fdbd8
KO		 218}
219
67e0dd8f
KO		 220static void bch2_btree_path_verify(struct btree_trans *trans,
221 struct btree_path *path)
1c6fdbd8 222{
5aab6635 223 struct bch_fs *c = trans->c;
2dac0eae 224 unsigned i;
1c6fdbd8 225
67e0dd8f
KO		 226 EBUG_ON(path->btree_id >= BTREE_ID_NR);
227
228 for (i = 0; i < (!path->cached ? BTREE_MAX_DEPTH : 1); i++) {
229 if (!path->l[i].b) {
d7407292 230 BUG_ON(!path->cached &&
faa6cb6c 231 bch2_btree_id_root(c, path->btree_id)->b->c.level > i);
67e0dd8f
KO		 232 break;
233 }
234
235 bch2_btree_path_verify_level(trans, path, i);
236 }
237
238 bch2_btree_path_verify_locks(path);
239}
240
241void bch2_trans_verify_paths(struct btree_trans *trans)
242{
243 struct btree_path *path;
ccb7b08f 244 unsigned iter;
67e0dd8f 245
ccb7b08f 246 trans_for_each_path(trans, path, iter)
67e0dd8f
KO		 247 bch2_btree_path_verify(trans, path);
248}
249
250static void bch2_btree_iter_verify(struct btree_iter *iter)
251{
252 struct btree_trans *trans = iter->trans;
253
254 BUG_ON(iter->btree_id >= BTREE_ID_NR);
255
07f383c7 256 BUG_ON(!!(iter->flags & BTREE_ITER_CACHED) != btree_iter_path(trans, iter)->cached);
7e1a3aa9 257
e751c01a
KO		 258 BUG_ON((iter->flags & BTREE_ITER_IS_EXTENTS) &&
259 (iter->flags & BTREE_ITER_ALL_SNAPSHOTS));
260
f21566f1 261 BUG_ON(!(iter->flags & __BTREE_ITER_ALL_SNAPSHOTS) &&
e751c01a 262 (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
9fcdd23b 263 !btree_type_has_snapshot_field(iter->btree_id));
e751c01a 264
1f2d9192 265 if (iter->update_path)
07f383c7
KO		 266 bch2_btree_path_verify(trans, &trans->paths[iter->update_path]);
267 bch2_btree_path_verify(trans, btree_iter_path(trans, iter));
2dac0eae
KO		 268}
269
7e1a3aa9
KO		 270static void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter)
271{
a861c722
KO		 272 BUG_ON((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
273 !iter->pos.snapshot);
274
e751c01a
KO		 275 BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
276 iter->pos.snapshot != iter->snapshot);
277
e88a75eb
KO		 278 BUG_ON(bkey_lt(iter->pos, bkey_start_pos(&iter->k)) ||
279 bkey_gt(iter->pos, iter->k.p));
7e1a3aa9
KO		 280}
281
c075ff70
KO		 282static int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k)
283{
284 struct btree_trans *trans = iter->trans;
285 struct btree_iter copy;
286 struct bkey_s_c prev;
287 int ret = 0;
288
289 if (!bch2_debug_check_iterators)
290 return 0;
291
292 if (!(iter->flags & BTREE_ITER_FILTER_SNAPSHOTS))
293 return 0;
294
295 if (bkey_err(k) || !k.k)
296 return 0;
297
298 BUG_ON(!bch2_snapshot_is_ancestor(trans->c,
299 iter->snapshot,
300 k.k->p.snapshot));
301
302 bch2_trans_iter_init(trans, &copy, iter->btree_id, iter->pos,
ffa7d262 303 BTREE_ITER_NOPRESERVE|
c075ff70
KO		 304 BTREE_ITER_ALL_SNAPSHOTS);
305 prev = bch2_btree_iter_prev(&copy);
306 if (!prev.k)
307 goto out;
308
309 ret = bkey_err(prev);
310 if (ret)
311 goto out;
312
e88a75eb 313 if (bkey_eq(prev.k->p, k.k->p) &&
c075ff70
KO		 314 bch2_snapshot_is_ancestor(trans->c, iter->snapshot,
315 prev.k->p.snapshot) > 0) {
fa8e94fa 316 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
c075ff70 317
fa8e94fa
KO		 318 bch2_bkey_to_text(&buf1, k.k);
319 bch2_bkey_to_text(&buf2, prev.k);
c075ff70
KO		 320
321 panic("iter snap %u\n"
322 "k %s\n"
323 "prev %s\n",
324 iter->snapshot,
fa8e94fa 325 buf1.buf, buf2.buf);
c075ff70
KO		 326 }
327out:
328 bch2_trans_iter_exit(trans, &copy);
329 return ret;
330}
331
32b26e8c
KO		 332void bch2_assert_pos_locked(struct btree_trans *trans, enum btree_id id,
333 struct bpos pos, bool key_cache)
334{
335 struct btree_path *path;
1f75ba4e 336 struct trans_for_each_path_inorder_iter iter;
fa8e94fa 337 struct printbuf buf = PRINTBUF;
32b26e8c 338
5c0bb66a
KO		 339 btree_trans_sort_paths(trans);
340
1f75ba4e 341 trans_for_each_path_inorder(trans, path, iter) {
32b26e8c
KO		 342 int cmp = cmp_int(path->btree_id, id) ?:
343 cmp_int(path->cached, key_cache);
344
345 if (cmp > 0)
346 break;
347 if (cmp < 0)
348 continue;
349
d4263e56 350 if (!btree_node_locked(path, 0) ||
32b26e8c
KO		 351 !path->should_be_locked)
352 continue;
353
354 if (!key_cache) {
e88a75eb
KO		 355 if (bkey_ge(pos, path->l[0].b->data->min_key) &&
356 bkey_le(pos, path->l[0].b->key.k.p))
32b26e8c
KO		 357 return;
358 } else {
e88a75eb 359 if (bkey_eq(pos, path->pos))
32b26e8c
KO		 360 return;
361 }
362 }
363
364 bch2_dump_trans_paths_updates(trans);
fa8e94fa
KO		 365 bch2_bpos_to_text(&buf, pos);
366
32b26e8c 367 panic("not locked: %s %s%s\n",
88dfe193 368 bch2_btree_id_str(id), buf.buf,
32b26e8c
KO		 369 key_cache ? " cached" : "");
370}
371
271a3d3a
KO		 372#else
373
67e0dd8f
KO		 374static inline void bch2_btree_path_verify_level(struct btree_trans *trans,
375 struct btree_path *path, unsigned l) {}
376static inline void bch2_btree_path_verify(struct btree_trans *trans,
377 struct btree_path *path) {}
7d6f9b64 378static inline void bch2_btree_iter_verify(struct btree_iter *iter) {}
7e1a3aa9 379static inline void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter) {}
c075ff70 380static inline int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k) { return 0; }
271a3d3a 381
1c6fdbd8
KO		 382#endif
383
67e0dd8f
KO		 384/* Btree path: fixups after btree updates */
385
23bbd2bb
KO		 386static void btree_node_iter_set_set_pos(struct btree_node_iter *iter,
387 struct btree *b,
388 struct bset_tree *t,
389 struct bkey_packed *k)
390{
391 struct btree_node_iter_set *set;
392
393 btree_node_iter_for_each(iter, set)
394 if (set->end == t->end_offset) {
395 set->k = __btree_node_key_to_offset(b, k);
396 bch2_btree_node_iter_sort(iter, b);
397 return;
398 }
399
400 bch2_btree_node_iter_push(iter, b, k, btree_bkey_last(b, t));
401}
402
67e0dd8f 403static void __bch2_btree_path_fix_key_modified(struct btree_path *path,
9626aeb1
KO		 404 struct btree *b,
405 struct bkey_packed *where)
887c2a4e 406{
67e0dd8f 407 struct btree_path_level *l = &path->l[b->c.level];
887c2a4e 408
9626aeb1
KO		 409 if (where != bch2_btree_node_iter_peek_all(&l->iter, l->b))
410 return;
411
67e0dd8f 412 if (bkey_iter_pos_cmp(l->b, where, &path->pos) < 0)
9626aeb1 413 bch2_btree_node_iter_advance(&l->iter, l->b);
887c2a4e
KO		 414}
415
67e0dd8f 416void bch2_btree_path_fix_key_modified(struct btree_trans *trans,
887c2a4e
KO		 417 struct btree *b,
418 struct bkey_packed *where)
419{
67e0dd8f 420 struct btree_path *path;
b0b67378 421 unsigned i;
887c2a4e 422
b0b67378 423 trans_for_each_path_with_node(trans, b, path, i) {
67e0dd8f
KO		 424 __bch2_btree_path_fix_key_modified(path, b, where);
425 bch2_btree_path_verify_level(trans, path, b->c.level);
887c2a4e
KO		 426 }
427}
428
67e0dd8f
KO		 429static void __bch2_btree_node_iter_fix(struct btree_path *path,
430 struct btree *b,
431 struct btree_node_iter *node_iter,
432 struct bset_tree *t,
433 struct bkey_packed *where,
434 unsigned clobber_u64s,
435 unsigned new_u64s)
1c6fdbd8
KO		 436{
437 const struct bkey_packed *end = btree_bkey_last(b, t);
438 struct btree_node_iter_set *set;
439 unsigned offset = __btree_node_key_to_offset(b, where);
440 int shift = new_u64s - clobber_u64s;
271a3d3a 441 unsigned old_end = t->end_offset - shift;
c0fc30da
KO		 442 unsigned orig_iter_pos = node_iter->data[0].k;
443 bool iter_current_key_modified =
444 orig_iter_pos >= offset &&
445 orig_iter_pos <= offset + clobber_u64s;
1c6fdbd8
KO		 446
447 btree_node_iter_for_each(node_iter, set)
448 if (set->end == old_end)
449 goto found;
450
451 /* didn't find the bset in the iterator - might have to readd it: */
452 if (new_u64s &&
67e0dd8f 453 bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
1c6fdbd8 454 bch2_btree_node_iter_push(node_iter, b, where, end);
c0fc30da
KO		 455 goto fixup_done;
456 } else {
457 /* Iterator is after key that changed */
059e4134 458 return;
1c6fdbd8 459 }
1c6fdbd8 460found:
271a3d3a 461 set->end = t->end_offset;
1c6fdbd8
KO		 462
463 /* Iterator hasn't gotten to the key that changed yet: */
464 if (set->k < offset)
059e4134 465 return;
1c6fdbd8
KO		 466
467 if (new_u64s &&
67e0dd8f 468 bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
1c6fdbd8
KO		 469 set->k = offset;
470 } else if (set->k < offset + clobber_u64s) {
471 set->k = offset + new_u64s;
472 if (set->k == set->end)
473 bch2_btree_node_iter_set_drop(node_iter, set);
474 } else {
c0fc30da 475 /* Iterator is after key that changed */
1c6fdbd8 476 set->k = (int) set->k + shift;
059e4134 477 return;
1c6fdbd8
KO		 478 }
479
1c6fdbd8 480 bch2_btree_node_iter_sort(node_iter, b);
c0fc30da
KO		 481fixup_done:
482 if (node_iter->data[0].k != orig_iter_pos)
483 iter_current_key_modified = true;
56e0e7c7 484
1c6fdbd8 485 /*
23bbd2bb
KO		 486 * When a new key is added, and the node iterator now points to that
487 * key, the iterator might have skipped past deleted keys that should
488 * come after the key the iterator now points to. We have to rewind to
c0fc30da
KO		 489 * before those deleted keys - otherwise
490 * bch2_btree_node_iter_prev_all() breaks:
1c6fdbd8 491 */
23bbd2bb 492 if (!bch2_btree_node_iter_end(node_iter) &&
c0fc30da 493 iter_current_key_modified &&
1ae29c1f 494 b->c.level) {
23bbd2bb
KO		 495 struct bkey_packed *k, *k2, *p;
496
497 k = bch2_btree_node_iter_peek_all(node_iter, b);
1c6fdbd8
KO		 498
499 for_each_bset(b, t) {
23bbd2bb
KO		 500 bool set_pos = false;
501
502 if (node_iter->data[0].end == t->end_offset)
1c6fdbd8
KO		 503 continue;
504
23bbd2bb
KO		 505 k2 = bch2_btree_node_iter_bset_pos(node_iter, b, t);
506
507 while ((p = bch2_bkey_prev_all(b, t, k2)) &&
508 bkey_iter_cmp(b, k, p) < 0) {
509 k2 = p;
510 set_pos = true;
1c6fdbd8 511 }
23bbd2bb
KO		 512
513 if (set_pos)
514 btree_node_iter_set_set_pos(node_iter,
515 b, t, k2);
1c6fdbd8
KO		 516 }
517 }
518}
519
9f6bd307 520void bch2_btree_node_iter_fix(struct btree_trans *trans,
67e0dd8f 521 struct btree_path *path,
216c9fac
KO		 522 struct btree *b,
523 struct btree_node_iter *node_iter,
524 struct bkey_packed *where,
525 unsigned clobber_u64s,
526 unsigned new_u64s)
1c6fdbd8 527{
216c9fac 528 struct bset_tree *t = bch2_bkey_to_bset_inlined(b, where);
67e0dd8f 529 struct btree_path *linked;
b0b67378 530 unsigned i;
1c6fdbd8 531
67e0dd8f
KO		 532 if (node_iter != &path->l[b->c.level].iter) {
533 __bch2_btree_node_iter_fix(path, b, node_iter, t,
059e4134 534 where, clobber_u64s, new_u64s);
2dac0eae 535
29364f34 536 if (bch2_debug_check_iterators)
2dac0eae 537 bch2_btree_node_iter_verify(node_iter, b);
059e4134 538 }
1c6fdbd8 539
b0b67378 540 trans_for_each_path_with_node(trans, b, linked, i) {
1c6fdbd8 541 __bch2_btree_node_iter_fix(linked, b,
059e4134
KO		 542 &linked->l[b->c.level].iter, t,
543 where, clobber_u64s, new_u64s);
67e0dd8f 544 bch2_btree_path_verify_level(trans, linked, b->c.level);
059e4134 545 }
1c6fdbd8
KO		 546}
547
67e0dd8f
KO		 548/* Btree path level: pointer to a particular btree node and node iter */
549
550static inline struct bkey_s_c __btree_iter_unpack(struct bch_fs *c,
551 struct btree_path_level *l,
1c6fdbd8
KO		 552 struct bkey *u,
553 struct bkey_packed *k)
554{
1c6fdbd8
KO		 555 if (unlikely(!k)) {
556 /*
557 * signal to bch2_btree_iter_peek_slot() that we're currently at
558 * a hole
559 */
26609b61 560 u->type = KEY_TYPE_deleted;
1c6fdbd8
KO		 561 return bkey_s_c_null;
562 }
563
2ce8fbd9 564 return bkey_disassemble(l->b, k, u);
1c6fdbd8
KO		 565}
566
67e0dd8f
KO		 567static inline struct bkey_s_c btree_path_level_peek_all(struct bch_fs *c,
568 struct btree_path_level *l,
569 struct bkey *u)
1c6fdbd8 570{
67e0dd8f 571 return __btree_iter_unpack(c, l, u,
1c6fdbd8
KO		 572 bch2_btree_node_iter_peek_all(&l->iter, l->b));
573}
574
67e0dd8f
KO		 575static inline struct bkey_s_c btree_path_level_peek(struct btree_trans *trans,
576 struct btree_path *path,
577 struct btree_path_level *l,
578 struct bkey *u)
1c6fdbd8 579{
67e0dd8f 580 struct bkey_s_c k = __btree_iter_unpack(trans->c, l, u,
1c6fdbd8 581 bch2_btree_node_iter_peek(&l->iter, l->b));
ca58cbd4 582
67e0dd8f
KO		 583 path->pos = k.k ? k.k->p : l->b->key.k.p;
584 trans->paths_sorted = false;
01eed771 585 bch2_btree_path_verify_level(trans, path, l - path->l);
ca58cbd4 586 return k;
1c6fdbd8
KO		 587}
588
67e0dd8f
KO		 589static inline struct bkey_s_c btree_path_level_prev(struct btree_trans *trans,
590 struct btree_path *path,
591 struct btree_path_level *l,
592 struct bkey *u)
ccf5a109 593{
67e0dd8f 594 struct bkey_s_c k = __btree_iter_unpack(trans->c, l, u,
ccf5a109 595 bch2_btree_node_iter_prev(&l->iter, l->b));
ca58cbd4 596
67e0dd8f
KO		 597 path->pos = k.k ? k.k->p : l->b->data->min_key;
598 trans->paths_sorted = false;
01eed771 599 bch2_btree_path_verify_level(trans, path, l - path->l);
ca58cbd4 600 return k;
ccf5a109
KO		 601}
602
67e0dd8f
KO		 603static inline bool btree_path_advance_to_pos(struct btree_path *path,
604 struct btree_path_level *l,
a00fd8c5 605 int max_advance)
1c6fdbd8 606{
a00fd8c5
KO		 607 struct bkey_packed *k;
608 int nr_advanced = 0;
609
610 while ((k = bch2_btree_node_iter_peek_all(&l->iter, l->b)) &&
67e0dd8f 611 bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
a00fd8c5
KO		 612 if (max_advance > 0 && nr_advanced >= max_advance)
613 return false;
614
615 bch2_btree_node_iter_advance(&l->iter, l->b);
616 nr_advanced++;
617 }
618
619 return true;
1c6fdbd8
KO		 620}
621
67e0dd8f 622static inline void __btree_path_level_init(struct btree_path *path,
78cf784e 623 unsigned level)
1c6fdbd8 624{
67e0dd8f 625 struct btree_path_level *l = &path->l[level];
a00fd8c5 626
67e0dd8f 627 bch2_btree_node_iter_init(&l->iter, l->b, &path->pos);
1c6fdbd8 628
537c49d6
KO		 629 /*
630 * Iterators to interior nodes should always be pointed at the first non
631 * whiteout:
632 */
633 if (level)
634 bch2_btree_node_iter_peek(&l->iter, l->b);
1c6fdbd8
KO		 635}
636
b2d1d56b
KO		 637void bch2_btree_path_level_init(struct btree_trans *trans,
638 struct btree_path *path,
639 struct btree *b)
1c6fdbd8 640{
67e0dd8f 641 BUG_ON(path->cached);
2ca88e5a 642
67e0dd8f 643 EBUG_ON(!btree_path_pos_in_node(path, b));
1c6fdbd8 644
1fb4fe63 645 path->l[b->c.level].lock_seq = six_lock_seq(&b->c.lock);
67e0dd8f
KO		 646 path->l[b->c.level].b = b;
647 __btree_path_level_init(path, b->c.level);
1c6fdbd8
KO		 648}
649
67e0dd8f
KO		 650/* Btree path: fixups after btree node updates: */
651
08f78031
KO		 652static void bch2_trans_revalidate_updates_in_node(struct btree_trans *trans, struct btree *b)
653{
654 struct bch_fs *c = trans->c;
08f78031
KO		 655
656 trans_for_each_update(trans, i)
657 if (!i->cached &&
658 i->level == b->c.level &&
659 i->btree_id == b->c.btree_id &&
660 bpos_cmp(i->k->k.p, b->data->min_key) >= 0 &&
661 bpos_cmp(i->k->k.p, b->data->max_key) <= 0) {
7f9821a7 662 i->old_v = bch2_btree_path_peek_slot(trans->paths + i->path, &i->old_k).v;
08f78031
KO		 663
664 if (unlikely(trans->journal_replay_not_finished)) {
665 struct bkey_i *j_k =
666 bch2_journal_keys_peek_slot(c, i->btree_id, i->level,
667 i->k->k.p);
668
669 if (j_k) {
670 i->old_k = j_k->k;
671 i->old_v = &j_k->v;
672 }
673 }
674 }
675}
676
1c6fdbd8
KO		 677/*
678 * A btree node is being replaced - update the iterator to point to the new
679 * node:
680 */
f3360005
KO		 681void bch2_trans_node_add(struct btree_trans *trans,
682 struct btree_path *path,
683 struct btree *b)
1c6fdbd8 684{
f3360005 685 struct btree_path *prev;
1c6fdbd8 686
f3360005
KO		 687 BUG_ON(!btree_path_pos_in_node(path, b));
688
689 while ((prev = prev_btree_path(trans, path)) &&
690 btree_path_pos_in_node(prev, b))
691 path = prev;
692
693 for (;
694 path && btree_path_pos_in_node(path, b);
695 path = next_btree_path(trans, path))
696 if (path->uptodate == BTREE_ITER_UPTODATE && !path->cached) {
1d3ecd7e
KO		 697 enum btree_node_locked_type t =
698 btree_lock_want(path, b->c.level);
1c6fdbd8 699
38474c26 700 if (t != BTREE_NODE_UNLOCKED) {
8bfe14e8 701 btree_node_unlock(trans, path, b->c.level);
c43a6ef9 702 six_lock_increment(&b->c.lock, (enum six_lock_type) t);
5b7fbdcd 703 mark_btree_node_locked(trans, path, b->c.level, t);
1c6fdbd8
KO		 704 }
705
e4215d0f 706 bch2_btree_path_level_init(trans, path, b);
1c6fdbd8 707 }
08f78031
KO		 708
709 bch2_trans_revalidate_updates_in_node(trans, b);
1c6fdbd8
KO		 710}
711
1c6fdbd8
KO		 712/*
713 * A btree node has been modified in such a way as to invalidate iterators - fix
714 * them:
715 */
f7a966a3 716void bch2_trans_node_reinit_iter(struct btree_trans *trans, struct btree *b)
1c6fdbd8 717{
67e0dd8f 718 struct btree_path *path;
b0b67378 719 unsigned i;
1c6fdbd8 720
b0b67378 721 trans_for_each_path_with_node(trans, b, path, i)
67e0dd8f 722 __btree_path_level_init(path, b->c.level);
08f78031
KO		 723
724 bch2_trans_revalidate_updates_in_node(trans, b);
1c6fdbd8
KO		 725}
726
67e0dd8f
KO		 727/* Btree path: traverse, set_pos: */
728
67e0dd8f
KO		 729static inline int btree_path_lock_root(struct btree_trans *trans,
730 struct btree_path *path,
a301dc38
KO		 731 unsigned depth_want,
732 unsigned long trace_ip)
1c6fdbd8 733{
e5af273f 734 struct bch_fs *c = trans->c;
faa6cb6c 735 struct btree *b, **rootp = &bch2_btree_id_root(c, path->btree_id)->b;
1c6fdbd8
KO		 736 enum six_lock_type lock_type;
737 unsigned i;
549d173c 738 int ret;
1c6fdbd8 739
67e0dd8f 740 EBUG_ON(path->nodes_locked);
1c6fdbd8
KO		 741
742 while (1) {
bd2bb273 743 b = READ_ONCE(*rootp);
67e0dd8f 744 path->level = READ_ONCE(b->c.level);
1c6fdbd8 745
67e0dd8f 746 if (unlikely(path->level < depth_want)) {
1c6fdbd8
KO		 747 /*
748 * the root is at a lower depth than the depth we want:
749 * got to the end of the btree, or we're walking nodes
750 * greater than some depth and there are no nodes >=
751 * that depth
752 */
67e0dd8f
KO		 753 path->level = depth_want;
754 for (i = path->level; i < BTREE_MAX_DEPTH; i++)
755 path->l[i].b = NULL;
8812600c 756 return 1;
1c6fdbd8
KO		 757 }
758
67e0dd8f 759 lock_type = __btree_lock_want(path, path->level);
0d7009d7
KO		 760 ret = btree_node_lock(trans, path, &b->c,
761 path->level, lock_type, trace_ip);
549d173c
KO		 762 if (unlikely(ret)) {
763 if (bch2_err_matches(ret, BCH_ERR_lock_fail_root_changed))
764 continue;
765 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
766 return ret;
767 BUG();
e5af273f 768 }
1c6fdbd8 769
bd2bb273 770 if (likely(b == READ_ONCE(*rootp) &&
67e0dd8f 771 b->c.level == path->level &&
1c6fdbd8 772 !race_fault())) {
67e0dd8f 773 for (i = 0; i < path->level; i++)
315c9ba6 774 path->l[i].b = ERR_PTR(-BCH_ERR_no_btree_node_lock_root);
67e0dd8f
KO		 775 path->l[path->level].b = b;
776 for (i = path->level + 1; i < BTREE_MAX_DEPTH; i++)
777 path->l[i].b = NULL;
778
5b7fbdcd
KO		 779 mark_btree_node_locked(trans, path, path->level,
780 (enum btree_node_locked_type) lock_type);
e4215d0f 781 bch2_btree_path_level_init(trans, path, b);
1c6fdbd8 782 return 0;
1c6fdbd8
KO		 783 }
784
c43a6ef9 785 six_unlock_type(&b->c.lock, lock_type);
1c6fdbd8
KO		 786 }
787}
788
789noinline
67e0dd8f 790static int btree_path_prefetch(struct btree_trans *trans, struct btree_path *path)
1c6fdbd8 791{
9f6bd307 792 struct bch_fs *c = trans->c;
67e0dd8f 793 struct btree_path_level *l = path_l(path);
1c6fdbd8
KO		 794 struct btree_node_iter node_iter = l->iter;
795 struct bkey_packed *k;
07a1006a 796 struct bkey_buf tmp;
3c471b65 797 unsigned nr = test_bit(BCH_FS_started, &c->flags)
67e0dd8f
KO		 798 ? (path->level > 1 ? 0 : 2)
799 : (path->level > 1 ? 1 : 16);
800 bool was_locked = btree_node_locked(path, path->level);
8b3e9bd6 801 int ret = 0;
1c6fdbd8 802
07a1006a
KO		 803 bch2_bkey_buf_init(&tmp);
804
dccedaaa 805 while (nr-- && !ret) {
67e0dd8f 806 if (!bch2_btree_node_relock(trans, path, path->level))
07a1006a 807 break;
1c6fdbd8
KO		 808
809 bch2_btree_node_iter_advance(&node_iter, l->b);
810 k = bch2_btree_node_iter_peek(&node_iter, l->b);
811 if (!k)
812 break;
813
07a1006a 814 bch2_bkey_buf_unpack(&tmp, c, l->b, k);
1306f87d 815 ret = bch2_btree_node_prefetch(trans, path, tmp.k, path->btree_id,
67e0dd8f 816 path->level - 1);
1c6fdbd8
KO		 817 }
818
819 if (!was_locked)
8bfe14e8 820 btree_node_unlock(trans, path, path->level);
07a1006a
KO		 821
822 bch2_bkey_buf_exit(&tmp, c);
8b3e9bd6 823 return ret;
1c6fdbd8
KO		 824}
825
5222a460
KO		 826static int btree_path_prefetch_j(struct btree_trans *trans, struct btree_path *path,
827 struct btree_and_journal_iter *jiter)
828{
829 struct bch_fs *c = trans->c;
830 struct bkey_s_c k;
831 struct bkey_buf tmp;
3c471b65 832 unsigned nr = test_bit(BCH_FS_started, &c->flags)
5222a460
KO		 833 ? (path->level > 1 ? 0 : 2)
834 : (path->level > 1 ? 1 : 16);
835 bool was_locked = btree_node_locked(path, path->level);
836 int ret = 0;
837
838 bch2_bkey_buf_init(&tmp);
839
dccedaaa 840 while (nr-- && !ret) {
5222a460
KO		 841 if (!bch2_btree_node_relock(trans, path, path->level))
842 break;
843
844 bch2_btree_and_journal_iter_advance(jiter);
845 k = bch2_btree_and_journal_iter_peek(jiter);
846 if (!k.k)
847 break;
848
849 bch2_bkey_buf_reassemble(&tmp, c, k);
1306f87d 850 ret = bch2_btree_node_prefetch(trans, path, tmp.k, path->btree_id,
5222a460
KO		 851 path->level - 1);
852 }
853
854 if (!was_locked)
8bfe14e8 855 btree_node_unlock(trans, path, path->level);
5222a460
KO		 856
857 bch2_bkey_buf_exit(&tmp, c);
858 return ret;
859}
860
78cf784e 861static noinline void btree_node_mem_ptr_set(struct btree_trans *trans,
67e0dd8f 862 struct btree_path *path,
72141e1f
KO		 863 unsigned plevel, struct btree *b)
864{
67e0dd8f
KO		 865 struct btree_path_level *l = &path->l[plevel];
866 bool locked = btree_node_locked(path, plevel);
72141e1f
KO		 867 struct bkey_packed *k;
868 struct bch_btree_ptr_v2 *bp;
869
67e0dd8f 870 if (!bch2_btree_node_relock(trans, path, plevel))
72141e1f
KO		 871 return;
872
873 k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
874 BUG_ON(k->type != KEY_TYPE_btree_ptr_v2);
875
876 bp = (void *) bkeyp_val(&l->b->format, k);
877 bp->mem_ptr = (unsigned long)b;
878
879 if (!locked)
8bfe14e8 880 btree_node_unlock(trans, path, plevel);
72141e1f
KO		 881}
882
5222a460
KO		 883static noinline int btree_node_iter_and_journal_peek(struct btree_trans *trans,
884 struct btree_path *path,
885 unsigned flags,
886 struct bkey_buf *out)
887{
888 struct bch_fs *c = trans->c;
889 struct btree_path_level *l = path_l(path);
890 struct btree_and_journal_iter jiter;
891 struct bkey_s_c k;
892 int ret = 0;
893
894 __bch2_btree_and_journal_iter_init_node_iter(&jiter, c, l->b, l->iter, path->pos);
895
896 k = bch2_btree_and_journal_iter_peek(&jiter);
897
898 bch2_bkey_buf_reassemble(out, c, k);
899
49a5192c
KO		 900 if ((flags & BTREE_ITER_PREFETCH) &&
901 c->opts.btree_node_prefetch)
5222a460
KO		 902 ret = btree_path_prefetch_j(trans, path, &jiter);
903
904 bch2_btree_and_journal_iter_exit(&jiter);
905 return ret;
906}
907
67e0dd8f
KO		 908static __always_inline int btree_path_down(struct btree_trans *trans,
909 struct btree_path *path,
910 unsigned flags,
a301dc38 911 unsigned long trace_ip)
1c6fdbd8 912{
6e075b54 913 struct bch_fs *c = trans->c;
67e0dd8f 914 struct btree_path_level *l = path_l(path);
1c6fdbd8 915 struct btree *b;
67e0dd8f
KO		 916 unsigned level = path->level - 1;
917 enum six_lock_type lock_type = __btree_lock_want(path, level);
07a1006a
KO		 918 struct bkey_buf tmp;
919 int ret;
1c6fdbd8 920
67e0dd8f 921 EBUG_ON(!btree_node_locked(path, path->level));
1c6fdbd8 922
07a1006a 923 bch2_bkey_buf_init(&tmp);
5222a460
KO		 924
925 if (unlikely(trans->journal_replay_not_finished)) {
926 ret = btree_node_iter_and_journal_peek(trans, path, flags, &tmp);
927 if (ret)
928 goto err;
929 } else {
930 bch2_bkey_buf_unpack(&tmp, c, l->b,
931 bch2_btree_node_iter_peek(&l->iter, l->b));
932
49a5192c
KO		 933 if ((flags & BTREE_ITER_PREFETCH) &&
934 c->opts.btree_node_prefetch) {
5222a460
KO		 935 ret = btree_path_prefetch(trans, path);
936 if (ret)
937 goto err;
938 }
939 }
1c6fdbd8 940
67e0dd8f 941 b = bch2_btree_node_get(trans, path, tmp.k, level, lock_type, trace_ip);
07a1006a
KO		 942 ret = PTR_ERR_OR_ZERO(b);
943 if (unlikely(ret))
944 goto err;
1c6fdbd8 945
5222a460
KO		 946 if (likely(!trans->journal_replay_not_finished &&
947 tmp.k->k.type == KEY_TYPE_btree_ptr_v2) &&
07a1006a 948 unlikely(b != btree_node_mem_ptr(tmp.k)))
67e0dd8f 949 btree_node_mem_ptr_set(trans, path, level + 1, b);
72141e1f 950
67e0dd8f 951 if (btree_node_read_locked(path, level + 1))
8bfe14e8 952 btree_node_unlock(trans, path, level + 1);
223b560e 953
5b7fbdcd
KO		 954 mark_btree_node_locked(trans, path, level,
955 (enum btree_node_locked_type) lock_type);
67e0dd8f 956 path->level = level;
e4215d0f 957 bch2_btree_path_level_init(trans, path, b);
c205321b 958
67e0dd8f 959 bch2_btree_path_verify_locks(path);
07a1006a
KO		 960err:
961 bch2_bkey_buf_exit(&tmp, c);
962 return ret;
1c6fdbd8
KO		 963}
964
1c6fdbd8 965
8f9ad91a 966static int bch2_btree_path_traverse_all(struct btree_trans *trans)
1c6fdbd8 967{
e542029e 968 struct bch_fs *c = trans->c;
0d7009d7 969 struct btree_path *path;
8f9ad91a 970 unsigned long trace_ip = _RET_IP_;
ccb7b08f
KO		 971 unsigned i;
972 int ret = 0;
e542029e 973
2ca88e5a 974 if (trans->in_traverse_all)
549d173c 975 return -BCH_ERR_transaction_restart_in_traverse_all;
2ca88e5a
KO		 976
977 trans->in_traverse_all = true;
978retry_all:
549d173c 979 trans->restarted = 0;
5e2d8be8 980 trans->last_restarted_ip = 0;
e5af273f 981
ccb7b08f 982 trans_for_each_path(trans, path, i)
67e0dd8f 983 path->should_be_locked = false;
e542029e 984
67e0dd8f 985 btree_trans_sort_paths(trans);
2527dd91 986
58fbf808 987 bch2_trans_unlock(trans);
a301dc38 988 cond_resched();
1c6fdbd8 989
8f9ad91a 990 if (unlikely(trans->memory_allocation_failure)) {
1c6fdbd8
KO		 991 struct closure cl;
992
993 closure_init_stack(&cl);
994
995 do {
a564c9fa 996 ret = bch2_btree_cache_cannibalize_lock(trans, &cl);
1c6fdbd8
KO		 997 closure_sync(&cl);
998 } while (ret);
999 }
1000
1c6fdbd8 1001 /* Now, redo traversals in correct order: */
0423fb71
KO		 1002 i = 0;
1003 while (i < trans->nr_sorted) {
96ed47d1 1004 btree_path_idx_t idx = trans->sorted[i];
2ca88e5a 1005
33aa419d
KO		 1006 /*
1007 * Traversing a path can cause another path to be added at about
1008 * the same position:
1009 */
96ed47d1
KO		 1010 if (trans->paths[idx].uptodate) {
1011 __btree_path_get(&trans->paths[idx], false);
1012 ret = bch2_btree_path_traverse_one(trans, idx, 0, _THIS_IP_);
1013 __btree_path_put(&trans->paths[idx], false);
c82ed304 1014
549d173c 1015 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
65d48e35 1016 bch2_err_matches(ret, ENOMEM))
33aa419d 1017 goto retry_all;
a897ef68
KO		 1018 if (ret)
1019 goto err;
33aa419d 1020 } else {
0423fb71 1021 i++;
33aa419d 1022 }
e542029e 1023 }
0423fb71
KO		 1024
1025 /*
ee94c413
KO		 1026 * We used to assert that all paths had been traversed here
1027 * (path->uptodate < BTREE_ITER_NEED_TRAVERSE); however, since
1e81f89b 1028 * path->should_be_locked is not set yet, we might have unlocked and
ee94c413 1029 * then failed to relock a path - that's fine.
0423fb71 1030 */
a897ef68 1031err:
a564c9fa 1032 bch2_btree_cache_cannibalize_unlock(trans);
2ca88e5a 1033
669f87a5
KO		 1034 trans->in_traverse_all = false;
1035
674cfc26 1036 trace_and_count(c, trans_traverse_all, trans, trace_ip);
1c6fdbd8 1037 return ret;
bf7b87a4 1038}
1c6fdbd8 1039
29aa78f1
KO		 1040static inline bool btree_path_check_pos_in_node(struct btree_path *path,
1041 unsigned l, int check_pos)
f4b61341 1042{
67e0dd8f 1043 if (check_pos < 0 && btree_path_pos_before_node(path, path->l[l].b))
f4b61341 1044 return false;
67e0dd8f 1045 if (check_pos > 0 && btree_path_pos_after_node(path, path->l[l].b))
f4b61341
KO		 1046 return false;
1047 return true;
1048}
1049
29aa78f1
KO		 1050static inline bool btree_path_good_node(struct btree_trans *trans,
1051 struct btree_path *path,
1052 unsigned l, int check_pos)
1053{
1054 return is_btree_node(path, l) &&
1055 bch2_btree_node_relock(trans, path, l) &&
1056 btree_path_check_pos_in_node(path, l, check_pos);
1057}
1058
c4bce586
KO		 1059static void btree_path_set_level_down(struct btree_trans *trans,
1060 struct btree_path *path,
1061 unsigned new_level)
1062{
1063 unsigned l;
1064
1065 path->level = new_level;
1066
1067 for (l = path->level + 1; l < BTREE_MAX_DEPTH; l++)
1068 if (btree_lock_want(path, l) == BTREE_NODE_UNLOCKED)
8bfe14e8 1069 btree_node_unlock(trans, path, l);
c4bce586
KO		 1070
1071 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1072 bch2_btree_path_verify(trans, path);
1073}
1074
29aa78f1
KO		 1075static noinline unsigned __btree_path_up_until_good_node(struct btree_trans *trans,
1076 struct btree_path *path,
1077 int check_pos)
1c6fdbd8 1078{
8ee0134e 1079 unsigned i, l = path->level;
29cea6f4 1080again:
67e0dd8f 1081 while (btree_path_node(path, l) &&
315c9ba6
KO		 1082 !btree_path_good_node(trans, path, l, check_pos))
1083 __btree_path_set_level_up(trans, path, l++);
1c6fdbd8 1084
8ee0134e
KO		 1085 /* If we need intent locks, take them too: */
1086 for (i = l + 1;
1087 i < path->locks_want && btree_path_node(path, i);
1088 i++)
29cea6f4 1089 if (!bch2_btree_node_relock(trans, path, i)) {
315c9ba6
KO		 1090 while (l <= i)
1091 __btree_path_set_level_up(trans, path, l++);
29cea6f4
KO		 1092 goto again;
1093 }
8ee0134e 1094
1c6fdbd8
KO		 1095 return l;
1096}
1097
29aa78f1
KO		 1098static inline unsigned btree_path_up_until_good_node(struct btree_trans *trans,
1099 struct btree_path *path,
1100 int check_pos)
1101{
1102 return likely(btree_node_locked(path, path->level) &&
1103 btree_path_check_pos_in_node(path, path->level, check_pos))
1104 ? path->level
1105 : __btree_path_up_until_good_node(trans, path, check_pos);
1106}
1107
1c6fdbd8
KO		 1108/*
1109 * This is the main state machine for walking down the btree - walks down to a
1110 * specified depth
1111 *
1112 * Returns 0 on success, -EIO on error (error reading in a btree node).
1113 *
1114 * On error, caller (peek_node()/peek_key()) must return NULL; the error is
b7607ce9 1115 * stashed in the iterator and returned from bch2_trans_exit().
1c6fdbd8 1116 */
4e3d1899 1117int bch2_btree_path_traverse_one(struct btree_trans *trans,
96ed47d1 1118 btree_path_idx_t path_idx,
4e3d1899
KO		 1119 unsigned flags,
1120 unsigned long trace_ip)
1c6fdbd8 1121{
96ed47d1 1122 struct btree_path *path = &trans->paths[path_idx];
8ee0134e 1123 unsigned depth_want = path->level;
0f35e086 1124 int ret = -((int) trans->restarted);
1c6fdbd8 1125
549d173c 1126 if (unlikely(ret))
979735df 1127 goto out;
979735df 1128
c4accde4
KO		 1129 if (unlikely(!trans->srcu_held))
1130 bch2_trans_srcu_lock(trans);
1131
e829b717 1132 /*
67e0dd8f
KO		 1133 * Ensure we obey path->should_be_locked: if it's set, we can't unlock
1134 * and re-traverse the path without a transaction restart:
e829b717 1135 */
67e0dd8f 1136 if (path->should_be_locked) {
549d173c 1137 ret = bch2_btree_path_relock(trans, path, trace_ip);
e829b717
KO		 1138 goto out;
1139 }
1140
67e0dd8f
KO		 1141 if (path->cached) {
1142 ret = bch2_btree_path_traverse_cached(trans, path, flags);
531a0095
KO		 1143 goto out;
1144 }
2ca88e5a 1145
96ed47d1
KO		 1146 path = &trans->paths[path_idx];
1147
67e0dd8f 1148 if (unlikely(path->level >= BTREE_MAX_DEPTH))
531a0095 1149 goto out;
2ca88e5a 1150
67e0dd8f 1151 path->level = btree_path_up_until_good_node(trans, path, 0);
1c6fdbd8 1152
845cffed
KO		 1153 EBUG_ON(btree_path_node(path, path->level) &&
1154 !btree_node_locked(path, path->level));
1155
1c6fdbd8 1156 /*
67e0dd8f 1157 * Note: path->nodes[path->level] may be temporarily NULL here - that
1c6fdbd8
KO		 1158 * would indicate to other code that we got to the end of the btree,
1159 * here it indicates that relocking the root failed - it's critical that
67e0dd8f 1160 * btree_path_lock_root() comes next and that it can't fail
1c6fdbd8 1161 */
67e0dd8f
KO		 1162 while (path->level > depth_want) {
1163 ret = btree_path_node(path, path->level)
1164 ? btree_path_down(trans, path, flags, trace_ip)
1165 : btree_path_lock_root(trans, path, depth_want, trace_ip);
1c6fdbd8 1166 if (unlikely(ret)) {
531a0095
KO		 1167 if (ret == 1) {
1168 /*
f21566f1
KO		 1169 * No nodes at this level - got to the end of
1170 * the btree:
531a0095
KO		 1171 */
1172 ret = 0;
1173 goto out;
1174 }
8812600c 1175
8bfe14e8 1176 __bch2_btree_path_unlock(trans, path);
67e0dd8f 1177 path->level = depth_want;
315c9ba6 1178 path->l[path->level].b = ERR_PTR(ret);
531a0095 1179 goto out;
1c6fdbd8
KO		 1180 }
1181 }
1182
67e0dd8f 1183 path->uptodate = BTREE_ITER_UPTODATE;
531a0095 1184out:
434b1c75
KO		 1185 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted)
1186 panic("ret %s (%i) trans->restarted %s (%i)\n",
1187 bch2_err_str(ret), ret,
1188 bch2_err_str(trans->restarted), trans->restarted);
67e0dd8f 1189 bch2_btree_path_verify(trans, path);
531a0095 1190 return ret;
1c6fdbd8
KO		 1191}
1192
0cc455b3 1193static inline void btree_path_copy(struct btree_trans *trans, struct btree_path *dst,
67e0dd8f
KO		 1194 struct btree_path *src)
1195{
1196 unsigned i, offset = offsetof(struct btree_path, pos);
1197
1198 memcpy((void *) dst + offset,
1199 (void *) src + offset,
1200 sizeof(struct btree_path) - offset);
1201
c81f5836
KO		 1202 for (i = 0; i < BTREE_MAX_DEPTH; i++) {
1203 unsigned t = btree_node_locked_type(dst, i);
1204
1205 if (t != BTREE_NODE_UNLOCKED)
1206 six_lock_increment(&dst->l[i].b->c.lock, t);
1207 }
67e0dd8f
KO		 1208}
1209
788cc25d
KO		 1210static btree_path_idx_t btree_path_clone(struct btree_trans *trans, btree_path_idx_t src,
1211 bool intent)
67e0dd8f 1212{
788cc25d 1213 btree_path_idx_t new = btree_path_alloc(trans, src);
788cc25d
KO		 1214 btree_path_copy(trans, trans->paths + new, trans->paths + src);
1215 __btree_path_get(trans->paths + new, intent);
c404f203
KO		 1216 return new;
1217}
1218
0cc455b3 1219__flatten
f6363aca
KO		 1220btree_path_idx_t __bch2_btree_path_make_mut(struct btree_trans *trans,
1221 btree_path_idx_t path, bool intent, unsigned long ip)
c404f203 1222{
f6363aca 1223 __btree_path_put(trans->paths + path, intent);
788cc25d 1224 path = btree_path_clone(trans, path, intent);
f6363aca 1225 trans->paths[path].preserve = false;
67e0dd8f
KO		 1226 return path;
1227}
1228
4617d946 1229btree_path_idx_t __must_check
67e0dd8f 1230__bch2_btree_path_set_pos(struct btree_trans *trans,
4617d946
KO		 1231 btree_path_idx_t path_idx, struct bpos new_pos,
1232 bool intent, unsigned long ip)
67e0dd8f 1233{
4617d946 1234 int cmp = bpos_cmp(new_pos, trans->paths[path_idx].pos);
5ce8b92d 1235
12344c7c 1236 bch2_trans_verify_not_in_restart(trans);
4617d946 1237 EBUG_ON(!trans->paths[path_idx].ref);
67e0dd8f 1238
f6363aca 1239 path_idx = bch2_btree_path_make_mut(trans, path_idx, intent, ip);
67e0dd8f 1240
4617d946 1241 struct btree_path *path = trans->paths + path_idx;
67e0dd8f 1242 path->pos = new_pos;
67e0dd8f
KO		 1243 trans->paths_sorted = false;
1244
1245 if (unlikely(path->cached)) {
8bfe14e8 1246 btree_node_unlock(trans, path, 0);
315c9ba6 1247 path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_up);
67e0dd8f
KO		 1248 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1249 goto out;
1250 }
1251
cd5bd162 1252 unsigned level = btree_path_up_until_good_node(trans, path, cmp);
67e0dd8f 1253
5c792e1b
KO		 1254 if (btree_path_node(path, level)) {
1255 struct btree_path_level *l = &path->l[level];
1256
1257 BUG_ON(!btree_node_locked(path, level));
67e0dd8f
KO		 1258 /*
1259 * We might have to skip over many keys, or just a few: try
1260 * advancing the node iterator, and if we have to skip over too
1261 * many keys just reinit it (or if we're rewinding, since that
1262 * is expensive).
1263 */
1264 if (cmp < 0 ||
5c792e1b
KO		 1265 !btree_path_advance_to_pos(path, l, 8))
1266 bch2_btree_node_iter_init(&l->iter, l->b, &path->pos);
1267
1268 /*
1269 * Iterators to interior nodes should always be pointed at the first non
1270 * whiteout:
1271 */
1272 if (unlikely(level))
1273 bch2_btree_node_iter_peek(&l->iter, l->b);
67e0dd8f
KO		 1274 }
1275
5c792e1b 1276 if (unlikely(level != path->level)) {
67e0dd8f 1277 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
8bfe14e8 1278 __bch2_btree_path_unlock(trans, path);
1d3ecd7e 1279 }
67e0dd8f
KO		 1280out:
1281 bch2_btree_path_verify(trans, path);
4617d946 1282 return path_idx;
67e0dd8f
KO		 1283}
1284
67e0dd8f
KO		 1285/* Btree path: main interface: */
1286
1287static struct btree_path *have_path_at_pos(struct btree_trans *trans, struct btree_path *path)
1288{
a300261a 1289 struct btree_path *sib;
67e0dd8f 1290
a300261a
KO		 1291 sib = prev_btree_path(trans, path);
1292 if (sib && !btree_path_cmp(sib, path))
1293 return sib;
67e0dd8f 1294
a300261a
KO		 1295 sib = next_btree_path(trans, path);
1296 if (sib && !btree_path_cmp(sib, path))
1297 return sib;
67e0dd8f
KO		 1298
1299 return NULL;
1300}
1301
9a74f63c 1302static struct btree_path *have_node_at_pos(struct btree_trans *trans, struct btree_path *path)
67e0dd8f 1303{
a300261a 1304 struct btree_path *sib;
67e0dd8f 1305
a300261a
KO		 1306 sib = prev_btree_path(trans, path);
1307 if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
1308 return sib;
67e0dd8f 1309
a300261a
KO		 1310 sib = next_btree_path(trans, path);
1311 if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
1312 return sib;
67e0dd8f 1313
9a74f63c 1314 return NULL;
67e0dd8f
KO		 1315}
1316
74e600c1 1317static inline void __bch2_path_free(struct btree_trans *trans, btree_path_idx_t path)
08070cba 1318{
74e600c1
KO		 1319 __bch2_btree_path_unlock(trans, trans->paths + path);
1320 btree_path_list_remove(trans, trans->paths + path);
1321 __clear_bit(path, trans->paths_allocated);
08070cba
KO		 1322}
1323
74e600c1 1324void bch2_path_put(struct btree_trans *trans, btree_path_idx_t path_idx, bool intent)
67e0dd8f 1325{
74e600c1 1326 struct btree_path *path = trans->paths + path_idx, *dup;
67e0dd8f 1327
67e0dd8f
KO		 1328 if (!__btree_path_put(path, intent))
1329 return;
1330
a300261a
KO		 1331 dup = path->preserve
1332 ? have_path_at_pos(trans, path)
1333 : have_node_at_pos(trans, path);
1334
1335 if (!dup && !(!path->preserve && !is_btree_node(path, path->level)))
1336 return;
67e0dd8f 1337
9a74f63c 1338 if (path->should_be_locked &&
a300261a 1339 !trans->restarted &&
b97de453 1340 (!dup || !bch2_btree_path_relock_norestart(trans, dup)))
9a74f63c
KO		 1341 return;
1342
a300261a
KO		 1343 if (dup) {
1344 dup->preserve |= path->preserve;
1345 dup->should_be_locked |= path->should_be_locked;
1346 }
1347
74e600c1 1348 __bch2_path_free(trans, path_idx);
67e0dd8f
KO		 1349}
1350
74e600c1 1351static void bch2_path_put_nokeep(struct btree_trans *trans, btree_path_idx_t path,
7dcbdbd8
KO		 1352 bool intent)
1353{
74e600c1 1354 if (!__btree_path_put(trans->paths + path, intent))
7dcbdbd8
KO		 1355 return;
1356
1357 __bch2_path_free(trans, path);
1358}
1359
12344c7c
KO		 1360void __noreturn bch2_trans_restart_error(struct btree_trans *trans, u32 restart_count)
1361{
1362 panic("trans->restart_count %u, should be %u, last restarted by %pS\n",
1363 trans->restart_count, restart_count,
5e2d8be8 1364 (void *) trans->last_begin_ip);
12344c7c
KO		 1365}
1366
1367void __noreturn bch2_trans_in_restart_error(struct btree_trans *trans)
1368{
1369 panic("in transaction restart: %s, last restarted by %pS\n",
1370 bch2_err_str(trans->restarted),
1371 (void *) trans->last_restarted_ip);
1372}
1373
5c0bb66a 1374noinline __cold
8570d775 1375void bch2_trans_updates_to_text(struct printbuf *buf, struct btree_trans *trans)
67e0dd8f 1376{
401ec4db 1377 prt_printf(buf, "transaction updates for %s journal seq %llu",
8570d775 1378 trans->fn, trans->journal_res.seq);
401ec4db
KO		 1379 prt_newline(buf);
1380 printbuf_indent_add(buf, 2);
2158fe46
KO		 1381
1382 trans_for_each_update(trans, i) {
1383 struct bkey_s_c old = { &i->old_k, i->old_v };
1384
401ec4db 1385 prt_printf(buf, "update: btree=%s cached=%u %pS",
88dfe193 1386 bch2_btree_id_str(i->btree_id),
8570d775
KO		 1387 i->cached,
1388 (void *) i->ip_allocated);
401ec4db 1389 prt_newline(buf);
8570d775 1390
401ec4db 1391 prt_printf(buf, " old ");
8570d775 1392 bch2_bkey_val_to_text(buf, trans->c, old);
401ec4db 1393 prt_newline(buf);
8570d775 1394
401ec4db 1395 prt_printf(buf, " new ");
8570d775 1396 bch2_bkey_val_to_text(buf, trans->c, bkey_i_to_s_c(i->k));
401ec4db 1397 prt_newline(buf);
2158fe46
KO		 1398 }
1399
67997234
KO		 1400 for (struct jset_entry *e = trans->journal_entries;
1401 e != btree_trans_journal_entries_top(trans);
1402 e = vstruct_next(e))
1403 bch2_journal_entry_to_text(buf, trans->c, e);
920e69bc 1404
401ec4db 1405 printbuf_indent_sub(buf, 2);
8570d775
KO		 1406}
1407
1408noinline __cold
1409void bch2_dump_trans_updates(struct btree_trans *trans)
1410{
1411 struct printbuf buf = PRINTBUF;
1412
1413 bch2_trans_updates_to_text(&buf, trans);
a8f35428 1414 bch2_print_string_as_lines(KERN_ERR, buf.buf);
8570d775 1415 printbuf_exit(&buf);
2158fe46
KO		 1416}
1417
311e446a 1418static void bch2_btree_path_to_text(struct printbuf *out, struct btree_trans *trans, btree_path_idx_t path_idx)
5c0bb66a 1419{
311e446a
KO		 1420 struct btree_path *path = trans->paths + path_idx;
1421
5c0bb66a 1422 prt_printf(out, "path: idx %2u ref %u:%u %c %c btree=%s l=%u pos ",
311e446a 1423 path_idx, path->ref, path->intent_ref,
5c0bb66a
KO		 1424 path->preserve ? 'P' : ' ',
1425 path->should_be_locked ? 'S' : ' ',
88dfe193 1426 bch2_btree_id_str(path->btree_id),
5c0bb66a
KO		 1427 path->level);
1428 bch2_bpos_to_text(out, path->pos);
1429
1430 prt_printf(out, " locks %u", path->nodes_locked);
94c69faf 1431#ifdef TRACK_PATH_ALLOCATED
5c0bb66a
KO		 1432 prt_printf(out, " %pS", (void *) path->ip_allocated);
1433#endif
1434 prt_newline(out);
1435}
1436
73bd774d 1437static noinline __cold
5c0bb66a
KO		 1438void __bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans,
1439 bool nosort)
2158fe46 1440{
1f75ba4e 1441 struct trans_for_each_path_inorder_iter iter;
67e0dd8f 1442
5c0bb66a
KO		 1443 if (!nosort)
1444 btree_trans_sort_paths(trans);
67e0dd8f 1445
311e446a
KO		 1446 trans_for_each_path_idx_inorder(trans, iter)
1447 bch2_btree_path_to_text(out, trans, iter.path_idx);
5c0bb66a 1448}
fa8e94fa 1449
5c0bb66a
KO		 1450noinline __cold
1451void bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans)
1452{
1453 __bch2_trans_paths_to_text(out, trans, false);
1454}
fa8e94fa 1455
73bd774d 1456static noinline __cold
5c0bb66a
KO		 1457void __bch2_dump_trans_paths_updates(struct btree_trans *trans, bool nosort)
1458{
1459 struct printbuf buf = PRINTBUF;
1460
1461 __bch2_trans_paths_to_text(&buf, trans, nosort);
a8f35428 1462 bch2_trans_updates_to_text(&buf, trans);
67e0dd8f 1463
a8f35428 1464 bch2_print_string_as_lines(KERN_ERR, buf.buf);
2158fe46 1465 printbuf_exit(&buf);
67e0dd8f
KO		 1466}
1467
5c0bb66a
KO		 1468noinline __cold
1469void bch2_dump_trans_paths_updates(struct btree_trans *trans)
1470{
1471 __bch2_dump_trans_paths_updates(trans, false);
1472}
1473
1474noinline __cold
1475static void bch2_trans_update_max_paths(struct btree_trans *trans)
1476{
1477 struct btree_transaction_stats *s = btree_trans_stats(trans);
1478 struct printbuf buf = PRINTBUF;
2c3b0fc3 1479 size_t nr = bitmap_weight(trans->paths_allocated, trans->nr_paths);
5c0bb66a 1480
5c0bb66a
KO		 1481 bch2_trans_paths_to_text(&buf, trans);
1482
1483 if (!buf.allocation_failure) {
1484 mutex_lock(&s->lock);
6e92d155
KO		 1485 if (nr > s->nr_max_paths) {
1486 s->nr_max_paths = nr;
5c0bb66a
KO		 1487 swap(s->max_paths_text, buf.buf);
1488 }
1489 mutex_unlock(&s->lock);
1490 }
1491
1492 printbuf_exit(&buf);
1493
6474b706 1494 trans->nr_paths_max = nr;
5c0bb66a
KO		 1495}
1496
e153a0d7
KO		 1497noinline __cold
1498int __bch2_btree_trans_too_many_iters(struct btree_trans *trans)
1499{
1500 if (trace_trans_restart_too_many_iters_enabled()) {
1501 struct printbuf buf = PRINTBUF;
1502
1503 bch2_trans_paths_to_text(&buf, trans);
1504 trace_trans_restart_too_many_iters(trans, _THIS_IP_, buf.buf);
1505 printbuf_exit(&buf);
1506 }
1507
1508 count_event(trans->c, trans_restart_too_many_iters);
1509
1510 return btree_trans_restart(trans, BCH_ERR_transaction_restart_too_many_iters);
1511}
1512
3f3bc66e
KO		 1513static noinline void btree_path_overflow(struct btree_trans *trans)
1514{
1515 bch2_dump_trans_paths_updates(trans);
0c99e17d
KO		 1516 bch_err(trans->c, "trans path overflow");
1517}
1518
1519static noinline void btree_paths_realloc(struct btree_trans *trans)
1520{
1521 unsigned nr = trans->nr_paths * 2;
1522
1523 void *p = kzalloc(BITS_TO_LONGS(nr) * sizeof(unsigned long) +
1524 sizeof(struct btree_trans_paths) +
1525 nr * sizeof(struct btree_path) +
1526 nr * sizeof(btree_path_idx_t) + 8 +
1527 nr * sizeof(struct btree_insert_entry), GFP_KERNEL|__GFP_NOFAIL);
1528
1529 unsigned long *paths_allocated = p;
1530 memcpy(paths_allocated, trans->paths_allocated, BITS_TO_LONGS(trans->nr_paths) * sizeof(unsigned long));
1531 p += BITS_TO_LONGS(nr) * sizeof(unsigned long);
1532
1533 p += sizeof(struct btree_trans_paths);
1534 struct btree_path *paths = p;
1535 *trans_paths_nr(paths) = nr;
1536 memcpy(paths, trans->paths, trans->nr_paths * sizeof(struct btree_path));
1537 p += nr * sizeof(struct btree_path);
1538
1539 btree_path_idx_t *sorted = p;
1540 memcpy(sorted, trans->sorted, trans->nr_sorted * sizeof(btree_path_idx_t));
1541 p += nr * sizeof(btree_path_idx_t) + 8;
1542
1543 struct btree_insert_entry *updates = p;
1544 memcpy(updates, trans->updates, trans->nr_paths * sizeof(struct btree_insert_entry));
1545
1546 unsigned long *old = trans->paths_allocated;
1547
1548 rcu_assign_pointer(trans->paths_allocated, paths_allocated);
1549 rcu_assign_pointer(trans->paths, paths);
1550 rcu_assign_pointer(trans->sorted, sorted);
1551 rcu_assign_pointer(trans->updates, updates);
1552
1553 trans->nr_paths = nr;
1554
1555 if (old != trans->_paths_allocated)
1556 kfree_rcu_mightsleep(old);
3f3bc66e
KO		 1557}
1558
788cc25d
KO		 1559static inline btree_path_idx_t btree_path_alloc(struct btree_trans *trans,
1560 btree_path_idx_t pos)
67e0dd8f 1561{
2c3b0fc3 1562 btree_path_idx_t idx = find_first_zero_bit(trans->paths_allocated, trans->nr_paths);
67e0dd8f 1563
0c99e17d
KO		 1564 if (unlikely(idx == trans->nr_paths)) {
1565 if (trans->nr_paths == BTREE_ITER_MAX) {
1566 btree_path_overflow(trans);
1567 return 0;
1568 }
1569
1570 btree_paths_realloc(trans);
1571 }
67e0dd8f 1572
992fa4e6
KO		 1573 /*
1574 * Do this before marking the new path as allocated, since it won't be
1575 * initialized yet:
1576 */
6474b706 1577 if (unlikely(idx > trans->nr_paths_max))
992fa4e6
KO		 1578 bch2_trans_update_max_paths(trans);
1579
6e92d155 1580 __set_bit(idx, trans->paths_allocated);
67e0dd8f 1581
788cc25d 1582 struct btree_path *path = &trans->paths[idx];
67e0dd8f
KO		 1583 path->ref = 0;
1584 path->intent_ref = 0;
1585 path->nodes_locked = 0;
67e0dd8f 1586
07f383c7 1587 btree_path_list_add(trans, pos, idx);
5c0bb66a 1588 trans->paths_sorted = false;
788cc25d 1589 return idx;
67e0dd8f
KO		 1590}
1591
255ebbbf
KO		 1592btree_path_idx_t bch2_path_get(struct btree_trans *trans,
1593 enum btree_id btree_id, struct bpos pos,
1594 unsigned locks_want, unsigned level,
1595 unsigned flags, unsigned long ip)
67e0dd8f 1596{
788cc25d 1597 struct btree_path *path;
f3e1f444
KO		 1598 bool cached = flags & BTREE_ITER_CACHED;
1599 bool intent = flags & BTREE_ITER_INTENT;
1f75ba4e 1600 struct trans_for_each_path_inorder_iter iter;
566eabd3 1601 btree_path_idx_t path_pos = 0, path_idx;
67e0dd8f 1602
12344c7c 1603 bch2_trans_verify_not_in_restart(trans);
7071878b
KO		 1604 bch2_trans_verify_locks(trans);
1605
eb7bd15f 1606 btree_trans_sort_paths(trans);
67e0dd8f 1607
1f75ba4e 1608 trans_for_each_path_inorder(trans, path, iter) {
807dda8c
KO		 1609 if (__btree_path_cmp(path,
1610 btree_id,
1611 cached,
1612 pos,
1613 level) > 0)
1614 break;
67e0dd8f 1615
1f75ba4e 1616 path_pos = iter.path_idx;
67e0dd8f
KO		 1617 }
1618
807dda8c 1619 if (path_pos &&
788cc25d
KO		 1620 trans->paths[path_pos].cached == cached &&
1621 trans->paths[path_pos].btree_id == btree_id &&
1622 trans->paths[path_pos].level == level) {
1623 __btree_path_get(trans->paths + path_pos, intent);
566eabd3
KO		 1624 path_idx = bch2_btree_path_set_pos(trans, path_pos, pos, intent, ip);
1625 path = trans->paths + path_idx;
67e0dd8f 1626 } else {
566eabd3
KO		 1627 path_idx = btree_path_alloc(trans, path_pos);
1628 path = trans->paths + path_idx;
67e0dd8f
KO		 1629
1630 __btree_path_get(path, intent);
1631 path->pos = pos;
1632 path->btree_id = btree_id;
1633 path->cached = cached;
67e0dd8f
KO		 1634 path->uptodate = BTREE_ITER_NEED_TRAVERSE;
1635 path->should_be_locked = false;
1636 path->level = level;
1637 path->locks_want = locks_want;
1638 path->nodes_locked = 0;
1f75ba4e 1639 for (unsigned i = 0; i < ARRAY_SIZE(path->l); i++)
315c9ba6 1640 path->l[i].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
94c69faf 1641#ifdef TRACK_PATH_ALLOCATED
ee2c6ea7 1642 path->ip_allocated = ip;
67e0dd8f
KO		 1643#endif
1644 trans->paths_sorted = false;
1645 }
1646
f3e1f444
KO		 1647 if (!(flags & BTREE_ITER_NOPRESERVE))
1648 path->preserve = true;
1649
67e0dd8f
KO		 1650 if (path->intent_ref)
1651 locks_want = max(locks_want, level + 1);
1652
1653 /*
1654 * If the path has locks_want greater than requested, we don't downgrade
1655 * it here - on transaction restart because btree node split needs to
1656 * upgrade locks, we might be putting/getting the iterator again.
1657 * Downgrading iterators only happens via bch2_trans_downgrade(), after
1658 * a successful transaction commit.
1659 */
1660
1661 locks_want = min(locks_want, BTREE_MAX_DEPTH);
cd5afabe 1662 if (locks_want > path->locks_want)
be9e782d 1663 bch2_btree_path_upgrade_noupgrade_sibs(trans, path, locks_want, NULL);
67e0dd8f 1664
566eabd3 1665 return path_idx;
67e0dd8f
KO		 1666}
1667
b2d1d56b 1668struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *path, struct bkey *u)
67e0dd8f
KO		 1669{
1670
f42238b5
KO		 1671 struct btree_path_level *l = path_l(path);
1672 struct bkey_packed *_k;
67e0dd8f
KO		 1673 struct bkey_s_c k;
1674
f42238b5
KO		 1675 if (unlikely(!l->b))
1676 return bkey_s_c_null;
f7b6ca23 1677
f42238b5
KO		 1678 EBUG_ON(path->uptodate != BTREE_ITER_UPTODATE);
1679 EBUG_ON(!btree_node_locked(path, path->level));
67e0dd8f 1680
f42238b5 1681 if (!path->cached) {
f7b6ca23 1682 _k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
67e0dd8f
KO		 1683 k = _k ? bkey_disassemble(l->b, _k, u) : bkey_s_c_null;
1684
e88a75eb 1685 EBUG_ON(k.k && bkey_deleted(k.k) && bpos_eq(k.k->p, path->pos));
67e0dd8f 1686
e88a75eb 1687 if (!k.k || !bpos_eq(path->pos, k.k->p))
67e0dd8f
KO		 1688 goto hole;
1689 } else {
1690 struct bkey_cached *ck = (void *) path->l[0].b;
1691
f7b6ca23
KO		 1692 EBUG_ON(ck &&
1693 (path->btree_id != ck->key.btree_id ||
e88a75eb 1694 !bkey_eq(path->pos, ck->key.pos)));
087e53c2
KO		 1695 if (!ck || !ck->valid)
1696 return bkey_s_c_null;
67e0dd8f 1697
2e63e180 1698 *u = ck->k->k;
67e0dd8f
KO		 1699 k = bkey_i_to_s_c(ck->k);
1700 }
1701
1702 return k;
1703hole:
1704 bkey_init(u);
1705 u->p = path->pos;
1706 return (struct bkey_s_c) { u, NULL };
1707}
1708
1709/* Btree iterators: */
1710
db92f2ea
KO		 1711int __must_check
1712__bch2_btree_iter_traverse(struct btree_iter *iter)
1713{
07f383c7 1714 return bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
db92f2ea
KO		 1715}
1716
08070cba
KO		 1717int __must_check
1718bch2_btree_iter_traverse(struct btree_iter *iter)
1719{
4617d946 1720 struct btree_trans *trans = iter->trans;
66a0a497
KO		 1721 int ret;
1722
07f383c7 1723 iter->path = bch2_btree_path_set_pos(trans, iter->path,
67e0dd8f 1724 btree_iter_search_key(iter),
ee2c6ea7
KO		 1725 iter->flags & BTREE_ITER_INTENT,
1726 btree_iter_ip_allocated(iter));
08070cba 1727
07f383c7 1728 ret = bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
66a0a497
KO		 1729 if (ret)
1730 return ret;
1731
07f383c7 1732 btree_path_set_should_be_locked(trans->paths + iter->path);
66a0a497 1733 return 0;
08070cba
KO		 1734}
1735
1c6fdbd8
KO		 1736/* Iterate across nodes (leaf and interior nodes) */
1737
1738struct btree *bch2_btree_iter_peek_node(struct btree_iter *iter)
1739{
502027a8 1740 struct btree_trans *trans = iter->trans;
f21566f1 1741 struct btree *b = NULL;
1c6fdbd8
KO		 1742 int ret;
1743
07f383c7 1744 EBUG_ON(trans->paths[iter->path].cached);
7e1a3aa9 1745 bch2_btree_iter_verify(iter);
1c6fdbd8 1746
07f383c7 1747 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
1c6fdbd8 1748 if (ret)
d355c6f4 1749 goto err;
1c6fdbd8 1750
07f383c7
KO		 1751 struct btree_path *path = btree_iter_path(trans, iter);
1752 b = btree_path_node(path, path->level);
1c6fdbd8 1753 if (!b)
f21566f1 1754 goto out;
1c6fdbd8 1755
e88a75eb 1756 BUG_ON(bpos_lt(b->key.k.p, iter->pos));
1c6fdbd8 1757
f21566f1 1758 bkey_init(&iter->k);
67e0dd8f 1759 iter->k.p = iter->pos = b->key.k.p;
502027a8 1760
07f383c7 1761 iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
ee2c6ea7
KO		 1762 iter->flags & BTREE_ITER_INTENT,
1763 btree_iter_ip_allocated(iter));
07f383c7 1764 btree_path_set_should_be_locked(btree_iter_path(trans, iter));
f21566f1
KO		 1765out:
1766 bch2_btree_iter_verify_entry_exit(iter);
1767 bch2_btree_iter_verify(iter);
2dac0eae 1768
1c6fdbd8 1769 return b;
d355c6f4
KO		 1770err:
1771 b = ERR_PTR(ret);
1772 goto out;
1c6fdbd8
KO		 1773}
1774
511b629a
KO		 1775struct btree *bch2_btree_iter_peek_node_and_restart(struct btree_iter *iter)
1776{
1777 struct btree *b;
1778
1779 while (b = bch2_btree_iter_peek_node(iter),
1780 bch2_err_matches(PTR_ERR_OR_ZERO(b), BCH_ERR_transaction_restart))
1781 bch2_trans_begin(iter->trans);
1782
1783 return b;
1784}
1785
2dac0eae 1786struct btree *bch2_btree_iter_next_node(struct btree_iter *iter)
1c6fdbd8 1787{
67e0dd8f 1788 struct btree_trans *trans = iter->trans;
f21566f1 1789 struct btree *b = NULL;
1c6fdbd8
KO		 1790 int ret;
1791
07f383c7 1792 EBUG_ON(trans->paths[iter->path].cached);
12344c7c 1793 bch2_trans_verify_not_in_restart(trans);
7e1a3aa9 1794 bch2_btree_iter_verify(iter);
1c6fdbd8 1795
07f383c7
KO		 1796 struct btree_path *path = btree_iter_path(trans, iter);
1797
979735df 1798 /* already at end? */
67e0dd8f 1799 if (!btree_path_node(path, path->level))
979735df 1800 return NULL;
1c6fdbd8 1801
979735df
KO		 1802 /* got to end? */
1803 if (!btree_path_node(path, path->level + 1)) {
8bfe14e8 1804 btree_path_set_level_up(trans, path);
979735df
KO		 1805 return NULL;
1806 }
1c6fdbd8 1807
979735df 1808 if (!bch2_btree_node_relock(trans, path, path->level + 1)) {
8bfe14e8 1809 __bch2_btree_path_unlock(trans, path);
315c9ba6
KO		 1810 path->l[path->level].b = ERR_PTR(-BCH_ERR_no_btree_node_relock);
1811 path->l[path->level + 1].b = ERR_PTR(-BCH_ERR_no_btree_node_relock);
7071878b 1812 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
674cfc26 1813 trace_and_count(trans->c, trans_restart_relock_next_node, trans, _THIS_IP_, path);
549d173c 1814 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_relock);
d355c6f4 1815 goto err;
979735df 1816 }
1c6fdbd8 1817
979735df 1818 b = btree_path_node(path, path->level + 1);
1c6fdbd8 1819
e88a75eb 1820 if (bpos_eq(iter->pos, b->key.k.p)) {
315c9ba6 1821 __btree_path_set_level_up(trans, path, path->level++);
979735df 1822 } else {
1c6fdbd8
KO		 1823 /*
1824 * Haven't gotten to the end of the parent node: go back down to
1825 * the next child node
1826 */
07f383c7
KO		 1827 iter->path = bch2_btree_path_set_pos(trans, iter->path,
1828 bpos_successor(iter->pos),
1829 iter->flags & BTREE_ITER_INTENT,
1830 btree_iter_ip_allocated(iter));
1c6fdbd8 1831
07f383c7 1832 path = btree_iter_path(trans, iter);
c4bce586 1833 btree_path_set_level_down(trans, path, iter->min_depth);
345ca825 1834
07f383c7 1835 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
d355c6f4
KO		 1836 if (ret)
1837 goto err;
1c6fdbd8 1838
07f383c7 1839 path = btree_iter_path(trans, iter);
67e0dd8f 1840 b = path->l[path->level].b;
1c6fdbd8
KO		 1841 }
1842
f21566f1 1843 bkey_init(&iter->k);
67e0dd8f 1844 iter->k.p = iter->pos = b->key.k.p;
502027a8 1845
07f383c7 1846 iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
ee2c6ea7
KO		 1847 iter->flags & BTREE_ITER_INTENT,
1848 btree_iter_ip_allocated(iter));
07f383c7
KO		 1849 btree_path_set_should_be_locked(btree_iter_path(trans, iter));
1850 EBUG_ON(btree_iter_path(trans, iter)->uptodate);
f21566f1
KO		 1851out:
1852 bch2_btree_iter_verify_entry_exit(iter);
1853 bch2_btree_iter_verify(iter);
2dac0eae 1854
1c6fdbd8 1855 return b;
d355c6f4
KO		 1856err:
1857 b = ERR_PTR(ret);
1858 goto out;
1c6fdbd8
KO		 1859}
1860
1861/* Iterate across keys (in leaf nodes only) */
1862
e0ba3b64 1863inline bool bch2_btree_iter_advance(struct btree_iter *iter)
434094be 1864{
e56978c8
KO		 1865 struct bpos pos = iter->k.p;
1866 bool ret = !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1867 ? bpos_eq(pos, SPOS_MAX)
1868 : bkey_eq(pos, SPOS_MAX));
3d495595 1869
e56978c8
KO		 1870 if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1871 pos = bkey_successor(iter, pos);
1872 bch2_btree_iter_set_pos(iter, pos);
1873 return ret;
3d495595
KO		 1874}
1875
e0ba3b64 1876inline bool bch2_btree_iter_rewind(struct btree_iter *iter)
3d495595
KO		 1877{
1878 struct bpos pos = bkey_start_pos(&iter->k);
e88a75eb
KO		 1879 bool ret = !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1880 ? bpos_eq(pos, POS_MIN)
1881 : bkey_eq(pos, POS_MIN));
3d495595 1882
7e1a3aa9 1883 if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
e751c01a 1884 pos = bkey_predecessor(iter, pos);
3d495595 1885 bch2_btree_iter_set_pos(iter, pos);
7e1a3aa9 1886 return ret;
434094be
KO		 1887}
1888
359e89ad
KO		 1889static noinline
1890void bch2_btree_trans_peek_prev_updates(struct btree_trans *trans, struct btree_iter *iter,
1891 struct bkey_s_c *k)
1892{
1893 struct bpos end = path_l(btree_iter_path(trans, iter))->b->data->min_key;
1894
1895 trans_for_each_update(trans, i)
1896 if (!i->key_cache_already_flushed &&
1897 i->btree_id == iter->btree_id &&
1898 bpos_le(i->k->k.p, iter->pos) &&
1899 bpos_ge(i->k->k.p, k->k ? k->k->p : end)) {
1900 iter->k = i->k->k;
1901 *k = bkey_i_to_s_c(i->k);
1902 }
1903}
1904
eb686359
KO		 1905static noinline
1906void bch2_btree_trans_peek_updates(struct btree_trans *trans, struct btree_iter *iter,
1907 struct bkey_s_c *k)
1908{
1909 struct btree_path *path = btree_iter_path(trans, iter);
1910 struct bpos end = path_l(path)->b->key.k.p;
1911
1912 trans_for_each_update(trans, i)
1913 if (!i->key_cache_already_flushed &&
1914 i->btree_id == iter->btree_id &&
1915 bpos_ge(i->k->k.p, path->pos) &&
1916 bpos_le(i->k->k.p, k->k ? k->k->p : end)) {
1917 iter->k = i->k->k;
1918 *k = bkey_i_to_s_c(i->k);
1919 }
1920}
1921
3763cb95 1922static noinline
c558c577
KO		 1923void bch2_btree_trans_peek_slot_updates(struct btree_trans *trans, struct btree_iter *iter,
1924 struct bkey_s_c *k)
3763cb95 1925{
c558c577
KO		 1926 trans_for_each_update(trans, i)
1927 if (!i->key_cache_already_flushed &&
1928 i->btree_id == iter->btree_id &&
1929 bpos_eq(i->k->k.p, iter->pos)) {
1930 iter->k = i->k->k;
1931 *k = bkey_i_to_s_c(i->k);
1932 }
3763cb95
KO		 1933}
1934
73bd774d
KO		 1935static struct bkey_i *bch2_btree_journal_peek(struct btree_trans *trans,
1936 struct btree_iter *iter,
1937 struct bpos end_pos)
30525f68 1938{
07f383c7
KO		 1939 struct btree_path *path = btree_iter_path(trans, iter);
1940
ad9c7992 1941 return bch2_journal_keys_peek_upto(trans->c, iter->btree_id,
07f383c7
KO		 1942 path->level,
1943 path->pos,
ad9c7992
KO		 1944 end_pos,
1945 &iter->journal_idx);
30525f68
KO		 1946}
1947
5222a460
KO		 1948static noinline
1949struct bkey_s_c btree_trans_peek_slot_journal(struct btree_trans *trans,
1950 struct btree_iter *iter)
1951{
07f383c7
KO		 1952 struct btree_path *path = btree_iter_path(trans, iter);
1953 struct bkey_i *k = bch2_btree_journal_peek(trans, iter, path->pos);
5222a460 1954
5650bb46 1955 if (k) {
5222a460
KO		 1956 iter->k = k->k;
1957 return bkey_i_to_s_c(k);
1958 } else {
1959 return bkey_s_c_null;
1960 }
1961}
1962
1963static noinline
1964struct bkey_s_c btree_trans_peek_journal(struct btree_trans *trans,
1965 struct btree_iter *iter,
1966 struct bkey_s_c k)
1967{
07f383c7 1968 struct btree_path *path = btree_iter_path(trans, iter);
5222a460 1969 struct bkey_i *next_journal =
30525f68 1970 bch2_btree_journal_peek(trans, iter,
07f383c7 1971 k.k ? k.k->p : path_l(path)->b->key.k.p);
5222a460 1972
5650bb46 1973 if (next_journal) {
5222a460
KO		 1974 iter->k = next_journal->k;
1975 k = bkey_i_to_s_c(next_journal);
1976 }
1977
1978 return k;
1979}
1980
f7b6ca23
KO		 1981/*
1982 * Checks btree key cache for key at iter->pos and returns it if present, or
1983 * bkey_s_c_null:
1984 */
1985static noinline
dcced069 1986struct bkey_s_c btree_trans_peek_key_cache(struct btree_iter *iter, struct bpos pos)
f7b6ca23
KO		 1987{
1988 struct btree_trans *trans = iter->trans;
1989 struct bch_fs *c = trans->c;
1990 struct bkey u;
52bf51b9 1991 struct bkey_s_c k;
f7b6ca23
KO		 1992 int ret;
1993
1617d56d
KO		 1994 if ((iter->flags & BTREE_ITER_KEY_CACHE_FILL) &&
1995 bpos_eq(iter->pos, pos))
1996 return bkey_s_c_null;
1997
f7b6ca23
KO		 1998 if (!bch2_btree_key_cache_find(c, iter->btree_id, pos))
1999 return bkey_s_c_null;
2000
2001 if (!iter->key_cache_path)
07f383c7 2002 iter->key_cache_path = bch2_path_get(trans, iter->btree_id, pos,
f7b6ca23 2003 iter->flags & BTREE_ITER_INTENT, 0,
087e53c2 2004 iter->flags|BTREE_ITER_CACHED|
ee2c6ea7
KO		 2005 BTREE_ITER_CACHED_NOFILL,
2006 _THIS_IP_);
f7b6ca23 2007
07f383c7 2008 iter->key_cache_path = bch2_btree_path_set_pos(trans, iter->key_cache_path, pos,
ee2c6ea7
KO		 2009 iter->flags & BTREE_ITER_INTENT,
2010 btree_iter_ip_allocated(iter));
f7b6ca23 2011
07f383c7 2012 ret = bch2_btree_path_traverse(trans, iter->key_cache_path,
dcced069 2013 iter->flags|BTREE_ITER_CACHED) ?:
07f383c7 2014 bch2_btree_path_relock(trans, btree_iter_path(trans, iter), _THIS_IP_);
f7b6ca23
KO		 2015 if (unlikely(ret))
2016 return bkey_s_c_err(ret);
2017
07f383c7 2018 btree_path_set_should_be_locked(trans->paths + iter->key_cache_path);
f7b6ca23 2019
07f383c7 2020 k = bch2_btree_path_peek_slot(trans->paths + iter->key_cache_path, &u);
52bf51b9
KO		 2021 if (k.k && !bkey_err(k)) {
2022 iter->k = u;
2023 k.k = &iter->k;
2024 }
2025 return k;
f7b6ca23
KO		 2026}
2027
a1e82d35 2028static struct bkey_s_c __bch2_btree_iter_peek(struct btree_iter *iter, struct bpos search_key)
1c6fdbd8 2029{
67e0dd8f 2030 struct btree_trans *trans = iter->trans;
f7b6ca23 2031 struct bkey_s_c k, k2;
8e432d98 2032 int ret;
1c6fdbd8 2033
07f383c7 2034 EBUG_ON(btree_iter_path(trans, iter)->cached);
7e1a3aa9 2035 bch2_btree_iter_verify(iter);
1c6fdbd8 2036
1c6fdbd8 2037 while (1) {
969576ec
KO		 2038 struct btree_path_level *l;
2039
07f383c7 2040 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
ee2c6ea7
KO		 2041 iter->flags & BTREE_ITER_INTENT,
2042 btree_iter_ip_allocated(iter));
c8476a4e 2043
07f383c7 2044 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
e6e024e9
KO		 2045 if (unlikely(ret)) {
2046 /* ensure that iter->k is consistent with iter->pos: */
2047 bch2_btree_iter_set_pos(iter, iter->pos);
2048 k = bkey_s_c_err(ret);
2049 goto out;
2050 }
1c6fdbd8 2051
07f383c7
KO		 2052 struct btree_path *path = btree_iter_path(trans, iter);
2053 l = path_l(path);
969576ec
KO		 2054
2055 if (unlikely(!l->b)) {
2056 /* No btree nodes at requested level: */
2057 bch2_btree_iter_set_pos(iter, SPOS_MAX);
2058 k = bkey_s_c_null;
2059 goto out;
2060 }
2061
07f383c7 2062 btree_path_set_should_be_locked(path);
f7b6ca23 2063
969576ec 2064 k = btree_path_level_peek_all(trans->c, l, &iter->k);
e6e024e9 2065
f7b6ca23
KO		 2066 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
2067 k.k &&
2068 (k2 = btree_trans_peek_key_cache(iter, k.k->p)).k) {
8192f8a5
KO		 2069 k = k2;
2070 ret = bkey_err(k);
f7b6ca23 2071 if (ret) {
f7b6ca23
KO		 2072 bch2_btree_iter_set_pos(iter, iter->pos);
2073 goto out;
2074 }
f7b6ca23
KO		 2075 }
2076
5222a460
KO		 2077 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL))
2078 k = btree_trans_peek_journal(trans, iter, k);
2079
eb686359
KO		 2080 if (unlikely((iter->flags & BTREE_ITER_WITH_UPDATES) &&
2081 trans->nr_updates))
2082 bch2_btree_trans_peek_updates(trans, iter, &k);
818664f5 2083
5222a460
KO		 2084 if (k.k && bkey_deleted(k.k)) {
2085 /*
2086 * If we've got a whiteout, and it's after the search
2087 * key, advance the search key to the whiteout instead
2088 * of just after the whiteout - it might be a btree
2089 * whiteout, with a real key at the same position, since
2090 * in the btree deleted keys sort before non deleted.
2091 */
e88a75eb 2092 search_key = !bpos_eq(search_key, k.k->p)
5222a460
KO		 2093 ? k.k->p
2094 : bpos_successor(k.k->p);
2095 continue;
2096 }
2097
818664f5 2098 if (likely(k.k)) {
c075ff70 2099 break;
e88a75eb 2100 } else if (likely(!bpos_eq(l->b->key.k.p, SPOS_MAX))) {
e6e024e9 2101 /* Advance to next leaf node: */
969576ec 2102 search_key = bpos_successor(l->b->key.k.p);
e6e024e9
KO		 2103 } else {
2104 /* End of btree: */
c8476a4e
KO		 2105 bch2_btree_iter_set_pos(iter, SPOS_MAX);
2106 k = bkey_s_c_null;
2107 goto out;
2108 }
1c6fdbd8 2109 }
a1e82d35 2110out:
a1e82d35
KO		 2111 bch2_btree_iter_verify(iter);
2112
2113 return k;
2114}
2115
2116/**
96dea3d5
KO		 2117 * bch2_btree_iter_peek_upto() - returns first key greater than or equal to
2118 * iterator's current position
2119 * @iter: iterator to peek from
2120 * @end: search limit: returns keys less than or equal to @end
2121 *
2122 * Returns: key if found, or an error extractable with bkey_err().
a1e82d35 2123 */
85d8cf16 2124struct bkey_s_c bch2_btree_iter_peek_upto(struct btree_iter *iter, struct bpos end)
a1e82d35
KO		 2125{
2126 struct btree_trans *trans = iter->trans;
2127 struct bpos search_key = btree_iter_search_key(iter);
2128 struct bkey_s_c k;
85d8cf16 2129 struct bpos iter_pos;
a1e82d35
KO		 2130 int ret;
2131
c72f687a 2132 EBUG_ON((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) && bkey_eq(end, POS_MAX));
b0babf2a 2133
1f2d9192 2134 if (iter->update_path) {
07f383c7 2135 bch2_path_put_nokeep(trans, iter->update_path,
7dcbdbd8 2136 iter->flags & BTREE_ITER_INTENT);
07f383c7 2137 iter->update_path = 0;
1f2d9192
KO		 2138 }
2139
a1e82d35
KO		 2140 bch2_btree_iter_verify_entry_exit(iter);
2141
2142 while (1) {
2143 k = __bch2_btree_iter_peek(iter, search_key);
c72f687a
KO		 2144 if (unlikely(!k.k))
2145 goto end;
2146 if (unlikely(bkey_err(k)))
fd211bc7 2147 goto out_no_locked;
a1e82d35 2148
85d8cf16 2149 /*
fa014953
KO		 2150 * We need to check against @end before FILTER_SNAPSHOTS because
2151 * if we get to a different inode that requested we might be
2152 * seeing keys for a different snapshot tree that will all be
2153 * filtered out.
2154 *
2155 * But we can't do the full check here, because bkey_start_pos()
2156 * isn't monotonically increasing before FILTER_SNAPSHOTS, and
2157 * that's what we check against in extents mode:
85d8cf16 2158 */
204f4514
KO		 2159 if (unlikely(!(iter->flags & BTREE_ITER_IS_EXTENTS)
2160 ? bkey_gt(k.k->p, end)
2161 : k.k->p.inode > end.inode))
c72f687a 2162 goto end;
85d8cf16 2163
1f2d9192 2164 if (iter->update_path &&
07f383c7
KO		 2165 !bkey_eq(trans->paths[iter->update_path].pos, k.k->p)) {
2166 bch2_path_put_nokeep(trans, iter->update_path,
7dcbdbd8 2167 iter->flags & BTREE_ITER_INTENT);
07f383c7 2168 iter->update_path = 0;
1f2d9192
KO		 2169 }
2170
2171 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2172 (iter->flags & BTREE_ITER_INTENT) &&
2173 !(iter->flags & BTREE_ITER_IS_EXTENTS) &&
2174 !iter->update_path) {
2175 struct bpos pos = k.k->p;
2176
2177 if (pos.snapshot < iter->snapshot) {
2178 search_key = bpos_successor(k.k->p);
2179 continue;
2180 }
2181
2182 pos.snapshot = iter->snapshot;
2183
2184 /*
2185 * advance, same as on exit for iter->path, but only up
2186 * to snapshot
2187 */
07f383c7 2188 __btree_path_get(trans->paths + iter->path, iter->flags & BTREE_ITER_INTENT);
1f2d9192
KO		 2189 iter->update_path = iter->path;
2190
07f383c7
KO		 2191 iter->update_path = bch2_btree_path_set_pos(trans,
2192 iter->update_path, pos,
ee2c6ea7
KO		 2193 iter->flags & BTREE_ITER_INTENT,
2194 _THIS_IP_);
07f383c7 2195 ret = bch2_btree_path_traverse(trans, iter->update_path, iter->flags);
ef073286
KO		 2196 if (unlikely(ret)) {
2197 k = bkey_s_c_err(ret);
2198 goto out_no_locked;
2199 }
1f2d9192
KO		 2200 }
2201
a1e82d35
KO		 2202 /*
2203 * We can never have a key in a leaf node at POS_MAX, so
2204 * we don't have to check these successor() calls:
2205 */
2206 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2207 !bch2_snapshot_is_ancestor(trans->c,
2208 iter->snapshot,
2209 k.k->p.snapshot)) {
2210 search_key = bpos_successor(k.k->p);
2211 continue;
2212 }
2213
2214 if (bkey_whiteout(k.k) &&
2215 !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2216 search_key = bkey_successor(iter, k.k->p);
2217 continue;
2218 }
2219
fa014953
KO		 2220 /*
2221 * iter->pos should be mononotically increasing, and always be
2222 * equal to the key we just returned - except extents can
2223 * straddle iter->pos:
2224 */
2225 if (!(iter->flags & BTREE_ITER_IS_EXTENTS))
2226 iter_pos = k.k->p;
2227 else
2228 iter_pos = bkey_max(iter->pos, bkey_start_pos(k.k));
2229
2230 if (unlikely(!(iter->flags & BTREE_ITER_IS_EXTENTS)
2231 ? bkey_gt(iter_pos, end)
2232 : bkey_ge(iter_pos, end)))
2233 goto end;
2234
a1e82d35
KO		 2235 break;
2236 }
2237
85d8cf16 2238 iter->pos = iter_pos;
1f2d9192 2239
07f383c7 2240 iter->path = bch2_btree_path_set_pos(trans, iter->path, k.k->p,
ee2c6ea7
KO		 2241 iter->flags & BTREE_ITER_INTENT,
2242 btree_iter_ip_allocated(iter));
fd211bc7 2243
07f383c7 2244 btree_path_set_should_be_locked(btree_iter_path(trans, iter));
fd211bc7 2245out_no_locked:
1f2d9192 2246 if (iter->update_path) {
07f383c7 2247 ret = bch2_btree_path_relock(trans, trans->paths + iter->update_path, _THIS_IP_);
fd211bc7 2248 if (unlikely(ret))
549d173c 2249 k = bkey_s_c_err(ret);
fd211bc7 2250 else
07f383c7 2251 btree_path_set_should_be_locked(trans->paths + iter->update_path);
1f2d9192 2252 }
1f2d9192 2253
a1e82d35 2254 if (!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS))
c075ff70
KO		 2255 iter->pos.snapshot = iter->snapshot;
2256
a1e82d35
KO		 2257 ret = bch2_btree_iter_verify_ret(iter, k);
2258 if (unlikely(ret)) {
2259 bch2_btree_iter_set_pos(iter, iter->pos);
2260 k = bkey_s_c_err(ret);
2261 }
67e0dd8f 2262
7e1a3aa9 2263 bch2_btree_iter_verify_entry_exit(iter);
c075ff70 2264
1c6fdbd8 2265 return k;
c72f687a
KO		 2266end:
2267 bch2_btree_iter_set_pos(iter, end);
2268 k = bkey_s_c_null;
2269 goto out_no_locked;
1c6fdbd8
KO		 2270}
2271
f4b61341 2272/**
96dea3d5 2273 * bch2_btree_iter_next() - returns first key greater than iterator's current
f4b61341 2274 * position
96dea3d5
KO		 2275 * @iter: iterator to peek from
2276 *
2277 * Returns: key if found, or an error extractable with bkey_err().
f4b61341 2278 */
1c6fdbd8
KO		 2279struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
2280{
e0ba3b64 2281 if (!bch2_btree_iter_advance(iter))
2e70ce56 2282 return bkey_s_c_null;
f4b61341 2283
2dac0eae 2284 return bch2_btree_iter_peek(iter);
1c6fdbd8
KO		 2285}
2286
ccf5a109 2287/**
96dea3d5 2288 * bch2_btree_iter_peek_prev() - returns first key less than or equal to
ccf5a109 2289 * iterator's current position
96dea3d5
KO		 2290 * @iter: iterator to peek from
2291 *
2292 * Returns: key if found, or an error extractable with bkey_err().
ccf5a109
KO		 2293 */
2294struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
1c6fdbd8 2295{
67e0dd8f 2296 struct btree_trans *trans = iter->trans;
c8476a4e 2297 struct bpos search_key = iter->pos;
1c6fdbd8 2298 struct bkey_s_c k;
c075ff70
KO		 2299 struct bkey saved_k;
2300 const struct bch_val *saved_v;
542e6396 2301 btree_path_idx_t saved_path = 0;
1c6fdbd8
KO		 2302 int ret;
2303
07f383c7
KO		 2304 EBUG_ON(btree_iter_path(trans, iter)->cached ||
2305 btree_iter_path(trans, iter)->level);
5222a460
KO		 2306
2307 if (iter->flags & BTREE_ITER_WITH_JOURNAL)
2308 return bkey_s_c_err(-EIO);
2309
7e1a3aa9
KO		 2310 bch2_btree_iter_verify(iter);
2311 bch2_btree_iter_verify_entry_exit(iter);
2312
c075ff70
KO		 2313 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2314 search_key.snapshot = U32_MAX;
2315
1c6fdbd8 2316 while (1) {
07f383c7 2317 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
ee2c6ea7
KO		 2318 iter->flags & BTREE_ITER_INTENT,
2319 btree_iter_ip_allocated(iter));
c8476a4e 2320
07f383c7 2321 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
7e1a3aa9 2322 if (unlikely(ret)) {
e6e024e9
KO		 2323 /* ensure that iter->k is consistent with iter->pos: */
2324 bch2_btree_iter_set_pos(iter, iter->pos);
7e1a3aa9 2325 k = bkey_s_c_err(ret);
fd211bc7 2326 goto out_no_locked;
7e1a3aa9 2327 }
1c6fdbd8 2328
07f383c7
KO		 2329 struct btree_path *path = btree_iter_path(trans, iter);
2330
2331 k = btree_path_level_peek(trans, path, &path->l[0], &iter->k);
3d495595
KO		 2332 if (!k.k ||
2333 ((iter->flags & BTREE_ITER_IS_EXTENTS)
e88a75eb
KO		 2334 ? bpos_ge(bkey_start_pos(k.k), search_key)
2335 : bpos_gt(k.k->p, search_key)))
07f383c7 2336 k = btree_path_level_prev(trans, path, &path->l[0], &iter->k);
ccf5a109 2337
359e89ad
KO		 2338 if (unlikely((iter->flags & BTREE_ITER_WITH_UPDATES) &&
2339 trans->nr_updates))
2340 bch2_btree_trans_peek_prev_updates(trans, iter, &k);
2341
e6e024e9 2342 if (likely(k.k)) {
c075ff70
KO		 2343 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) {
2344 if (k.k->p.snapshot == iter->snapshot)
2345 goto got_key;
2346
2347 /*
2348 * If we have a saved candidate, and we're no
2349 * longer at the same _key_ (not pos), return
2350 * that candidate
2351 */
e88a75eb 2352 if (saved_path && !bkey_eq(k.k->p, saved_k.p)) {
07f383c7 2353 bch2_path_put_nokeep(trans, iter->path,
c075ff70 2354 iter->flags & BTREE_ITER_INTENT);
542e6396
KO		 2355 iter->path = saved_path;
2356 saved_path = 0;
c075ff70
KO		 2357 iter->k = saved_k;
2358 k.v = saved_v;
2359 goto got_key;
2360 }
2361
0bc64d7e 2362 if (bch2_snapshot_is_ancestor(trans->c,
c075ff70
KO		 2363 iter->snapshot,
2364 k.k->p.snapshot)) {
2365 if (saved_path)
542e6396 2366 bch2_path_put_nokeep(trans, saved_path,
c075ff70 2367 iter->flags & BTREE_ITER_INTENT);
542e6396 2368 saved_path = btree_path_clone(trans, iter->path,
c075ff70 2369 iter->flags & BTREE_ITER_INTENT);
07f383c7 2370 path = btree_iter_path(trans, iter);
c075ff70
KO		 2371 saved_k = *k.k;
2372 saved_v = k.v;
2373 }
2374
2375 search_key = bpos_predecessor(k.k->p);
2376 continue;
2377 }
2378got_key:
2379 if (bkey_whiteout(k.k) &&
2380 !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2381 search_key = bkey_predecessor(iter, k.k->p);
2382 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2383 search_key.snapshot = U32_MAX;
2384 continue;
2385 }
2386
07f383c7 2387 btree_path_set_should_be_locked(path);
1c6fdbd8 2388 break;
07f383c7 2389 } else if (likely(!bpos_eq(path->l[0].b->data->min_key, POS_MIN))) {
e6e024e9 2390 /* Advance to previous leaf node: */
07f383c7 2391 search_key = bpos_predecessor(path->l[0].b->data->min_key);
e6e024e9
KO		 2392 } else {
2393 /* Start of btree: */
c8476a4e 2394 bch2_btree_iter_set_pos(iter, POS_MIN);
7e1a3aa9 2395 k = bkey_s_c_null;
fd211bc7 2396 goto out_no_locked;
7e1a3aa9 2397 }
ccf5a109 2398 }
1c6fdbd8 2399
e88a75eb 2400 EBUG_ON(bkey_gt(bkey_start_pos(k.k), iter->pos));
3d495595
KO		 2401
2402 /* Extents can straddle iter->pos: */
e88a75eb 2403 if (bkey_lt(k.k->p, iter->pos))
3d495595 2404 iter->pos = k.k->p;
c075ff70
KO		 2405
2406 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2407 iter->pos.snapshot = iter->snapshot;
fd211bc7 2408out_no_locked:
c075ff70 2409 if (saved_path)
542e6396 2410 bch2_path_put_nokeep(trans, saved_path, iter->flags & BTREE_ITER_INTENT);
67e0dd8f 2411
7e1a3aa9
KO		 2412 bch2_btree_iter_verify_entry_exit(iter);
2413 bch2_btree_iter_verify(iter);
67e0dd8f 2414
1c6fdbd8
KO		 2415 return k;
2416}
2417
ccf5a109 2418/**
96dea3d5 2419 * bch2_btree_iter_prev() - returns first key less than iterator's current
ccf5a109 2420 * position
96dea3d5
KO		 2421 * @iter: iterator to peek from
2422 *
2423 * Returns: key if found, or an error extractable with bkey_err().
ccf5a109
KO		 2424 */
2425struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter)
2426{
e0ba3b64 2427 if (!bch2_btree_iter_rewind(iter))
2e70ce56 2428 return bkey_s_c_null;
ccf5a109 2429
2e70ce56 2430 return bch2_btree_iter_peek_prev(iter);
ccf5a109
KO		 2431}
2432
c3801239 2433struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
271a3d3a 2434{
9f6bd307 2435 struct btree_trans *trans = iter->trans;
953ee28a 2436 struct bpos search_key;
271a3d3a 2437 struct bkey_s_c k;
c3801239
KO		 2438 int ret;
2439
7e1a3aa9
KO		 2440 bch2_btree_iter_verify(iter);
2441 bch2_btree_iter_verify_entry_exit(iter);
07f383c7 2442 EBUG_ON(btree_iter_path(trans, iter)->level && (iter->flags & BTREE_ITER_WITH_KEY_CACHE));
7e1a3aa9 2443
1d214eb1
KO		 2444 /* extents can't span inode numbers: */
2445 if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
953ee28a 2446 unlikely(iter->pos.offset == KEY_OFFSET_MAX)) {
1d214eb1
KO		 2447 if (iter->pos.inode == KEY_INODE_MAX)
2448 return bkey_s_c_null;
c3801239 2449
1d214eb1
KO		 2450 bch2_btree_iter_set_pos(iter, bpos_nosnap_successor(iter->pos));
2451 }
8042b5b7 2452
953ee28a 2453 search_key = btree_iter_search_key(iter);
07f383c7 2454 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
ee2c6ea7
KO		 2455 iter->flags & BTREE_ITER_INTENT,
2456 btree_iter_ip_allocated(iter));
953ee28a 2457
07f383c7 2458 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
23dfb3a2
KO		 2459 if (unlikely(ret)) {
2460 k = bkey_s_c_err(ret);
2461 goto out_no_locked;
2462 }
1c6fdbd8 2463
c075ff70
KO		 2464 if ((iter->flags & BTREE_ITER_CACHED) ||
2465 !(iter->flags & (BTREE_ITER_IS_EXTENTS|BTREE_ITER_FILTER_SNAPSHOTS))) {
c558c577 2466 k = bkey_s_c_null;
1c6fdbd8 2467
c558c577
KO		 2468 if (unlikely((iter->flags & BTREE_ITER_WITH_UPDATES) &&
2469 trans->nr_updates)) {
2470 bch2_btree_trans_peek_slot_updates(trans, iter, &k);
2471 if (k.k)
2472 goto out;
1d214eb1
KO		 2473 }
2474
5222a460
KO		 2475 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL) &&
2476 (k = btree_trans_peek_slot_journal(trans, iter)).k)
2477 goto out;
2478
f7b6ca23 2479 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
dcced069 2480 (k = btree_trans_peek_key_cache(iter, iter->pos)).k) {
8192f8a5 2481 if (!bkey_err(k))
f7b6ca23 2482 iter->k = *k.k;
8192f8a5
KO		 2483 /* We're not returning a key from iter->path: */
2484 goto out_no_locked;
f7b6ca23
KO		 2485 }
2486
07f383c7 2487 k = bch2_btree_path_peek_slot(trans->paths + iter->path, &iter->k);
f42238b5
KO		 2488 if (unlikely(!k.k))
2489 goto out_no_locked;
1d214eb1
KO		 2490 } else {
2491 struct bpos next;
c72f687a
KO		 2492 struct bpos end = iter->pos;
2493
2494 if (iter->flags & BTREE_ITER_IS_EXTENTS)
2495 end.offset = U64_MAX;
1d214eb1 2496
07f383c7 2497 EBUG_ON(btree_iter_path(trans, iter)->level);
2ae4573e 2498
b1d87f52 2499 if (iter->flags & BTREE_ITER_INTENT) {
67e0dd8f 2500 struct btree_iter iter2;
b1d87f52 2501
67e0dd8f 2502 bch2_trans_copy_iter(&iter2, iter);
85d8cf16 2503 k = bch2_btree_iter_peek_upto(&iter2, end);
b1d87f52 2504
67e0dd8f
KO		 2505 if (k.k && !bkey_err(k)) {
2506 iter->k = iter2.k;
2507 k.k = &iter->k;
2508 }
2509 bch2_trans_iter_exit(trans, &iter2);
b1d87f52
KO		 2510 } else {
2511 struct bpos pos = iter->pos;
2512
c72f687a 2513 k = bch2_btree_iter_peek_upto(iter, end);
23dfb3a2
KO		 2514 if (unlikely(bkey_err(k)))
2515 bch2_btree_iter_set_pos(iter, pos);
2516 else
2517 iter->pos = pos;
b1d87f52 2518 }
1d214eb1
KO		 2519
2520 if (unlikely(bkey_err(k)))
f42238b5 2521 goto out_no_locked;
1d214eb1
KO		 2522
2523 next = k.k ? bkey_start_pos(k.k) : POS_MAX;
2524
e88a75eb 2525 if (bkey_lt(iter->pos, next)) {
1d214eb1
KO		 2526 bkey_init(&iter->k);
2527 iter->k.p = iter->pos;
c075ff70
KO		 2528
2529 if (iter->flags & BTREE_ITER_IS_EXTENTS) {
2530 bch2_key_resize(&iter->k,
2531 min_t(u64, KEY_SIZE_MAX,
2532 (next.inode == iter->pos.inode
2533 ? next.offset
2534 : KEY_OFFSET_MAX) -
2535 iter->pos.offset));
2536 EBUG_ON(!iter->k.size);
2537 }
1d214eb1
KO		 2538
2539 k = (struct bkey_s_c) { &iter->k, NULL };
1d214eb1 2540 }
271a3d3a 2541 }
5222a460 2542out:
07f383c7 2543 btree_path_set_should_be_locked(btree_iter_path(trans, iter));
fd211bc7 2544out_no_locked:
7e1a3aa9
KO		 2545 bch2_btree_iter_verify_entry_exit(iter);
2546 bch2_btree_iter_verify(iter);
c075ff70
KO		 2547 ret = bch2_btree_iter_verify_ret(iter, k);
2548 if (unlikely(ret))
2549 return bkey_s_c_err(ret);
66a0a497 2550
63f1a598 2551 return k;
1c6fdbd8
KO		 2552}
2553
1c6fdbd8
KO		 2554struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *iter)
2555{
e0ba3b64 2556 if (!bch2_btree_iter_advance(iter))
2e70ce56 2557 return bkey_s_c_null;
1c6fdbd8 2558
2dac0eae 2559 return bch2_btree_iter_peek_slot(iter);
1c6fdbd8
KO		 2560}
2561
18fc6ae5
KO		 2562struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_iter *iter)
2563{
e0ba3b64 2564 if (!bch2_btree_iter_rewind(iter))
18fc6ae5
KO		 2565 return bkey_s_c_null;
2566
2567 return bch2_btree_iter_peek_slot(iter);
2568}
2569
0f2ea655
KO		 2570struct bkey_s_c bch2_btree_iter_peek_and_restart_outlined(struct btree_iter *iter)
2571{
2572 struct bkey_s_c k;
2573
2574 while (btree_trans_too_many_iters(iter->trans) ||
2575 (k = bch2_btree_iter_peek_type(iter, iter->flags),
2576 bch2_err_matches(bkey_err(k), BCH_ERR_transaction_restart)))
2577 bch2_trans_begin(iter->trans);
2578
2579 return k;
2580}
2581
1c6fdbd8
KO		 2582/* new transactional stuff: */
2583
0423fb71
KO		 2584#ifdef CONFIG_BCACHEFS_DEBUG
2585static void btree_trans_verify_sorted_refs(struct btree_trans *trans)
2586{
67e0dd8f 2587 struct btree_path *path;
0423fb71
KO		 2588 unsigned i;
2589
2c3b0fc3 2590 BUG_ON(trans->nr_sorted != bitmap_weight(trans->paths_allocated, trans->nr_paths) - 1);
0423fb71 2591
ccb7b08f 2592 trans_for_each_path(trans, path, i) {
67e0dd8f 2593 BUG_ON(path->sorted_idx >= trans->nr_sorted);
ccb7b08f 2594 BUG_ON(trans->sorted[path->sorted_idx] != i);
0423fb71
KO		 2595 }
2596
2597 for (i = 0; i < trans->nr_sorted; i++) {
2598 unsigned idx = trans->sorted[i];
2599
6e92d155 2600 BUG_ON(!test_bit(idx, trans->paths_allocated));
67e0dd8f 2601 BUG_ON(trans->paths[idx].sorted_idx != i);
0423fb71
KO		 2602 }
2603}
0423fb71
KO		 2604
2605static void btree_trans_verify_sorted(struct btree_trans *trans)
2606{
67e0dd8f 2607 struct btree_path *path, *prev = NULL;
1f75ba4e 2608 struct trans_for_each_path_inorder_iter iter;
0423fb71 2609
ee4d17d0
KO		 2610 if (!bch2_debug_check_iterators)
2611 return;
2612
1f75ba4e 2613 trans_for_each_path_inorder(trans, path, iter) {
eb7bd15f 2614 if (prev && btree_path_cmp(prev, path) > 0) {
5c0bb66a 2615 __bch2_dump_trans_paths_updates(trans, true);
eb7bd15f
KO		 2616 panic("trans paths out of order!\n");
2617 }
67e0dd8f 2618 prev = path;
0423fb71 2619 }
0423fb71 2620}
068bcaa5
KO		 2621#else
2622static inline void btree_trans_verify_sorted_refs(struct btree_trans *trans) {}
2623static inline void btree_trans_verify_sorted(struct btree_trans *trans) {}
2624#endif
0423fb71 2625
068bcaa5 2626void __bch2_btree_trans_sort_paths(struct btree_trans *trans)
0423fb71
KO		 2627{
2628 int i, l = 0, r = trans->nr_sorted, inc = 1;
2629 bool swapped;
2630
068bcaa5
KO		 2631 btree_trans_verify_sorted_refs(trans);
2632
2633 if (trans->paths_sorted)
2634 goto out;
2635
0423fb71
KO		 2636 /*
2637 * Cocktail shaker sort: this is efficient because iterators will be
5c0bb66a 2638 * mostly sorted.
0423fb71
KO		 2639 */
2640 do {
2641 swapped = false;
2642
2643 for (i = inc > 0 ? l : r - 2;
2644 i + 1 < r && i >= l;
2645 i += inc) {
67e0dd8f
KO		 2646 if (btree_path_cmp(trans->paths + trans->sorted[i],
2647 trans->paths + trans->sorted[i + 1]) > 0) {
0423fb71 2648 swap(trans->sorted[i], trans->sorted[i + 1]);
67e0dd8f
KO		 2649 trans->paths[trans->sorted[i]].sorted_idx = i;
2650 trans->paths[trans->sorted[i + 1]].sorted_idx = i + 1;
0423fb71
KO		 2651 swapped = true;
2652 }
2653 }
2654
2655 if (inc > 0)
2656 --r;
2657 else
2658 l++;
2659 inc = -inc;
2660 } while (swapped);
2661
67e0dd8f 2662 trans->paths_sorted = true;
068bcaa5 2663out:
0423fb71
KO		 2664 btree_trans_verify_sorted(trans);
2665}
2666
67e0dd8f
KO		 2667static inline void btree_path_list_remove(struct btree_trans *trans,
2668 struct btree_path *path)
0423fb71 2669{
67e0dd8f 2670 EBUG_ON(path->sorted_idx >= trans->nr_sorted);
0423fb71
KO		 2671#ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
2672 trans->nr_sorted--;
67e0dd8f
KO		 2673 memmove_u64s_down_small(trans->sorted + path->sorted_idx,
2674 trans->sorted + path->sorted_idx + 1,
0c99e17d
KO		 2675 DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx,
2676 sizeof(u64) / sizeof(btree_path_idx_t)));
0423fb71 2677#else
67e0dd8f 2678 array_remove_item(trans->sorted, trans->nr_sorted, path->sorted_idx);
0423fb71 2679#endif
0c99e17d 2680 for (unsigned i = path->sorted_idx; i < trans->nr_sorted; i++)
67e0dd8f 2681 trans->paths[trans->sorted[i]].sorted_idx = i;
0423fb71
KO		 2682}
2683
67e0dd8f 2684static inline void btree_path_list_add(struct btree_trans *trans,
07f383c7
KO		 2685 btree_path_idx_t pos,
2686 btree_path_idx_t path_idx)
0423fb71 2687{
07f383c7 2688 struct btree_path *path = trans->paths + path_idx;
0423fb71 2689
07f383c7 2690 path->sorted_idx = pos ? trans->paths[pos].sorted_idx + 1 : trans->nr_sorted;
0423fb71
KO		 2691
2692#ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
67e0dd8f
KO		 2693 memmove_u64s_up_small(trans->sorted + path->sorted_idx + 1,
2694 trans->sorted + path->sorted_idx,
0c99e17d
KO		 2695 DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx,
2696 sizeof(u64) / sizeof(btree_path_idx_t)));
0423fb71 2697 trans->nr_sorted++;
07f383c7 2698 trans->sorted[path->sorted_idx] = path_idx;
0423fb71 2699#else
07f383c7 2700 array_insert_item(trans->sorted, trans->nr_sorted, path->sorted_idx, path_idx);
0423fb71
KO		 2701#endif
2702
0c99e17d 2703 for (unsigned i = path->sorted_idx; i < trans->nr_sorted; i++)
67e0dd8f 2704 trans->paths[trans->sorted[i]].sorted_idx = i;
64bc0011 2705
0423fb71 2706 btree_trans_verify_sorted_refs(trans);
7c26ecae
KO		 2707}
2708
67e0dd8f 2709void bch2_trans_iter_exit(struct btree_trans *trans, struct btree_iter *iter)
64bc0011 2710{
1f2d9192 2711 if (iter->update_path)
07f383c7 2712 bch2_path_put_nokeep(trans, iter->update_path,
1f2d9192 2713 iter->flags & BTREE_ITER_INTENT);
11f11737 2714 if (iter->path)
07f383c7 2715 bch2_path_put(trans, iter->path,
11f11737 2716 iter->flags & BTREE_ITER_INTENT);
f7b6ca23 2717 if (iter->key_cache_path)
07f383c7 2718 bch2_path_put(trans, iter->key_cache_path,
f7b6ca23 2719 iter->flags & BTREE_ITER_INTENT);
07f383c7
KO		 2720 iter->path = 0;
2721 iter->update_path = 0;
2722 iter->key_cache_path = 0;
2723 iter->trans = NULL;
64bc0011
KO		 2724}
2725
c96f108b
KO		 2726void bch2_trans_iter_init_outlined(struct btree_trans *trans,
2727 struct btree_iter *iter,
2728 enum btree_id btree_id, struct bpos pos,
2729 unsigned flags)
2730{
2731 bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0,
ee2c6ea7
KO		 2732 bch2_btree_iter_flags(trans, btree_id, flags),
2733 _RET_IP_);
424eb881
KO		 2734}
2735
67e0dd8f
KO		 2736void bch2_trans_node_iter_init(struct btree_trans *trans,
2737 struct btree_iter *iter,
2738 enum btree_id btree_id,
2739 struct bpos pos,
2740 unsigned locks_want,
2741 unsigned depth,
2742 unsigned flags)
1c6fdbd8 2743{
1e81f89b
KO		 2744 flags |= BTREE_ITER_NOT_EXTENTS;
2745 flags |= __BTREE_ITER_ALL_SNAPSHOTS;
2746 flags |= BTREE_ITER_ALL_SNAPSHOTS;
c96f108b
KO		 2747
2748 bch2_trans_iter_init_common(trans, iter, btree_id, pos, locks_want, depth,
ee2c6ea7
KO		 2749 __bch2_btree_iter_flags(trans, btree_id, flags),
2750 _RET_IP_);
c96f108b
KO		 2751
2752 iter->min_depth = depth;
2753
07f383c7
KO		 2754 struct btree_path *path = btree_iter_path(trans, iter);
2755 BUG_ON(path->locks_want < min(locks_want, BTREE_MAX_DEPTH));
2756 BUG_ON(path->level != depth);
2757 BUG_ON(iter->min_depth != depth);
67e0dd8f 2758}
7c26ecae 2759
67e0dd8f
KO		 2760void bch2_trans_copy_iter(struct btree_iter *dst, struct btree_iter *src)
2761{
07f383c7
KO		 2762 struct btree_trans *trans = src->trans;
2763
67e0dd8f
KO		 2764 *dst = *src;
2765 if (src->path)
07f383c7 2766 __btree_path_get(trans->paths + src->path, src->flags & BTREE_ITER_INTENT);
1f2d9192 2767 if (src->update_path)
07f383c7
KO		 2768 __btree_path_get(trans->paths + src->update_path, src->flags & BTREE_ITER_INTENT);
2769 dst->key_cache_path = 0;
1c6fdbd8
KO		 2770}
2771
14d8f26a 2772void *__bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
1c6fdbd8 2773{
31403dca 2774 struct bch_fs *c = trans->c;
616928c3 2775 unsigned new_top = trans->mem_top + size;
e06af207
KO		 2776 unsigned old_bytes = trans->mem_bytes;
2777 unsigned new_bytes = roundup_pow_of_two(new_top);
78367aaa 2778 int ret;
14d8f26a 2779 void *new_mem;
73a117d2
KO		 2780 void *p;
2781
14d8f26a 2782 WARN_ON_ONCE(new_bytes > BTREE_TRANS_MEM_MAX);
e131b6aa 2783
31403dca 2784 struct btree_transaction_stats *s = btree_trans_stats(trans);
83322e8c 2785 s->max_mem = max(s->max_mem, new_bytes);
31403dca 2786
78367aaa
KO		 2787 new_mem = krealloc(trans->mem, new_bytes, GFP_NOWAIT|__GFP_NOWARN);
2788 if (unlikely(!new_mem)) {
2789 bch2_trans_unlock(trans);
2790
2791 new_mem = krealloc(trans->mem, new_bytes, GFP_KERNEL);
2792 if (!new_mem && new_bytes <= BTREE_TRANS_MEM_MAX) {
31403dca 2793 new_mem = mempool_alloc(&c->btree_trans_mem_pool, GFP_KERNEL);
78367aaa
KO		 2794 new_bytes = BTREE_TRANS_MEM_MAX;
2795 kfree(trans->mem);
2796 }
1c6fdbd8 2797
78367aaa
KO		 2798 if (!new_mem)
2799 return ERR_PTR(-BCH_ERR_ENOMEM_trans_kmalloc);
2800
2801 trans->mem = new_mem;
2802 trans->mem_bytes = new_bytes;
2803
2804 ret = bch2_trans_relock(trans);
2805 if (ret)
2806 return ERR_PTR(ret);
2807 }
1c6fdbd8 2808
14d8f26a
KO		 2809 trans->mem = new_mem;
2810 trans->mem_bytes = new_bytes;
1c6fdbd8 2811
14d8f26a 2812 if (old_bytes) {
31403dca 2813 trace_and_count(c, trans_restart_mem_realloced, trans, _RET_IP_, new_bytes);
14d8f26a 2814 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_mem_realloced));
1c6fdbd8
KO		 2815 }
2816
20bceecb 2817 p = trans->mem + trans->mem_top;
1c6fdbd8 2818 trans->mem_top += size;
7ed158f2 2819 memset(p, 0, size);
20bceecb 2820 return p;
1c6fdbd8
KO		 2821}
2822
f82755e4
KO		 2823static inline void check_srcu_held_too_long(struct btree_trans *trans)
2824{
2825 WARN(trans->srcu_held && time_after(jiffies, trans->srcu_lock_time + HZ * 10),
2826 "btree trans held srcu lock (delaying memory reclaim) for %lu seconds",
2827 (jiffies - trans->srcu_lock_time) / HZ);
2828}
2829
c4accde4 2830void bch2_trans_srcu_unlock(struct btree_trans *trans)
e242b92a 2831{
c4accde4
KO		 2832 if (trans->srcu_held) {
2833 struct bch_fs *c = trans->c;
2834 struct btree_path *path;
ccb7b08f 2835 unsigned i;
e242b92a 2836
ccb7b08f 2837 trans_for_each_path(trans, path, i)
c4accde4
KO		 2838 if (path->cached && !btree_node_locked(path, 0))
2839 path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_srcu_reset);
e242b92a 2840
f82755e4 2841 check_srcu_held_too_long(trans);
c4accde4
KO		 2842 srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
2843 trans->srcu_held = false;
2844 }
2845}
2846
79904fa2 2847static void bch2_trans_srcu_lock(struct btree_trans *trans)
c4accde4
KO		 2848{
2849 if (!trans->srcu_held) {
2850 trans->srcu_idx = srcu_read_lock(&trans->c->btree_trans_barrier);
2851 trans->srcu_lock_time = jiffies;
2852 trans->srcu_held = true;
2853 }
e242b92a
KO		 2854}
2855
d38494c4 2856/**
955af634 2857 * bch2_trans_begin() - reset a transaction after a interrupted attempt
d38494c4 2858 * @trans: transaction to reset
d38494c4 2859 *
96dea3d5
KO		 2860 * Returns: current restart counter, to be used with trans_was_restarted()
2861 *
549d173c
KO		 2862 * While iterating over nodes or updating nodes a attempt to lock a btree node
2863 * may return BCH_ERR_transaction_restart when the trylock fails. When this
2864 * occurs bch2_trans_begin() should be called and the transaction retried.
d38494c4 2865 */
e941ae7d 2866u32 bch2_trans_begin(struct btree_trans *trans)
1c6fdbd8 2867{
67e0dd8f 2868 struct btree_path *path;
ccb7b08f 2869 unsigned i;
f375d6ca 2870 u64 now;
a8e00bd4 2871
0fbf71f8 2872 bch2_trans_reset_updates(trans);
1c6fdbd8 2873
c497df8b 2874 trans->restart_count++;
163e885a 2875 trans->mem_top = 0;
24de63da 2876 trans->journal_entries = NULL;
bf7b87a4 2877
ccb7b08f 2878 trans_for_each_path(trans, path, i) {
25a77231
KO		 2879 path->should_be_locked = false;
2880
e1f7fa06
KO		 2881 /*
2882 * If the transaction wasn't restarted, we're presuming to be
2883 * doing something new: dont keep iterators excpt the ones that
2884 * are in use - except for the subvolumes btree:
2885 */
2886 if (!trans->restarted && path->btree_id != BTREE_ID_subvolumes)
2887 path->preserve = false;
2888
67e0dd8f
KO		 2889 /*
2890 * XXX: we probably shouldn't be doing this if the transaction
2891 * was restarted, but currently we still overflow transaction
2892 * iterators if we do that
2893 */
2894 if (!path->ref && !path->preserve)
ccb7b08f 2895 __bch2_path_free(trans, i);
67e0dd8f 2896 else
25a77231 2897 path->preserve = false;
67e0dd8f
KO		 2898 }
2899
f375d6ca 2900 now = local_clock();
89056f24
KO		 2901
2902 if (!IS_ENABLED(CONFIG_BCACHEFS_NO_LATENCY_ACCT) &&
2903 time_after64(now, trans->last_begin_time + 10))
2904 __bch2_time_stats_update(&btree_trans_stats(trans)->duration,
2905 trans->last_begin_time, now);
2906
43de721a
KO		 2907 if (!trans->restarted &&
2908 (need_resched() ||
89056f24 2909 time_after64(now, trans->last_begin_time + BTREE_TRANS_MAX_LOCK_HOLD_TIME_NS))) {
49c7cd9d 2910 drop_locks_do(trans, (cond_resched(), 0));
f375d6ca 2911 now = local_clock();
43de721a 2912 }
f375d6ca 2913 trans->last_begin_time = now;
50dc0f69 2914
c4accde4
KO		 2915 if (unlikely(trans->srcu_held &&
2916 time_after(jiffies, trans->srcu_lock_time + msecs_to_jiffies(10))))
2917 bch2_trans_srcu_unlock(trans);
e242b92a 2918
5e2d8be8 2919 trans->last_begin_ip = _RET_IP_;
60b55388 2920 if (trans->restarted) {
67e0dd8f 2921 bch2_btree_path_traverse_all(trans);
60b55388
KO		 2922 trans->notrace_relock_fail = false;
2923 }
e5af273f 2924
e941ae7d
KO		 2925 return trans->restart_count;
2926}
2927
83322e8c 2928const char *bch2_btree_transaction_fns[BCH_TRANSACTIONS_NR] = { "(unknown)" };
307e3c13
KO		 2929
2930unsigned bch2_trans_get_fn_idx(const char *fn)
4aba7d45 2931{
83322e8c 2932 for (unsigned i = 0; i < ARRAY_SIZE(bch2_btree_transaction_fns); i++)
307e3c13
KO		 2933 if (!bch2_btree_transaction_fns[i] ||
2934 bch2_btree_transaction_fns[i] == fn) {
2935 bch2_btree_transaction_fns[i] = fn;
4aba7d45
KO		 2936 return i;
2937 }
2938
2939 pr_warn_once("BCH_TRANSACTIONS_NR not big enough!");
83322e8c 2940 return 0;
4aba7d45
KO		 2941}
2942
6bd68ec2 2943struct btree_trans *__bch2_trans_get(struct bch_fs *c, unsigned fn_idx)
c0ebe3e4 2944 __acquires(&c->btree_trans_barrier)
1c6fdbd8 2945{
6bd68ec2 2946 struct btree_trans *trans;
5c0bb66a 2947
31403dca
KO		 2948 if (IS_ENABLED(__KERNEL__)) {
2949 trans = this_cpu_xchg(c->btree_trans_bufs->trans, NULL);
2950 if (trans) {
5cc6daf7 2951 memset(trans, 0, offsetof(struct btree_trans, list));
31403dca
KO		 2952 goto got_trans;
2953 }
2954 }
6bd68ec2 2955
31403dca 2956 trans = mempool_alloc(&c->btree_trans_pool, GFP_NOFS);
ae1ede58 2957 memset(trans, 0, sizeof(*trans));
31403dca
KO		 2958 closure_init_stack(&trans->ref);
2959
2960 seqmutex_lock(&c->btree_trans_lock);
2961 if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG)) {
2962 struct btree_trans *pos;
2963 pid_t pid = current->pid;
2964
2965 trans->locking_wait.task = current;
2966
2967 list_for_each_entry(pos, &c->btree_trans_list, list) {
2968 struct task_struct *pos_task = READ_ONCE(pos->locking_wait.task);
2969 /*
2970 * We'd much prefer to be stricter here and completely
2971 * disallow multiple btree_trans in the same thread -
2972 * but the data move path calls bch2_write when we
2973 * already have a btree_trans initialized.
2974 */
2975 BUG_ON(pos_task &&
2976 pid == pos_task->pid &&
2977 bch2_trans_locked(pos));
2978
2979 if (pos_task && pid < pos_task->pid) {
2980 list_add_tail(&trans->list, &pos->list);
2981 goto list_add_done;
2982 }
2983 }
2984 }
2985 list_add_tail(&trans->list, &c->btree_trans_list);
2986list_add_done:
2987 seqmutex_unlock(&c->btree_trans_lock);
2988got_trans:
1c6fdbd8 2989 trans->c = c;
d7e4e513 2990 trans->last_begin_time = local_clock();
307e3c13 2991 trans->fn_idx = fn_idx;
33bd5d06 2992 trans->locking_wait.task = current;
5222a460 2993 trans->journal_replay_not_finished =
8a443d3e
KO		 2994 unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)) &&
2995 atomic_inc_not_zero(&c->journal_keys.ref);
2c3b0fc3 2996 trans->nr_paths = ARRAY_SIZE(trans->_paths);
6474b706
KO		 2997 trans->paths_allocated = trans->_paths_allocated;
2998 trans->sorted = trans->_sorted;
2999 trans->paths = trans->_paths;
5cc6daf7 3000 trans->updates = trans->_updates;
6474b706 3001
0c99e17d 3002 *trans_paths_nr(trans->paths) = BTREE_ITER_INITIAL;
fea153a8 3003
f8fd5871
KO		 3004 trans->paths_allocated[0] = 1;
3005
31403dca
KO		 3006 if (fn_idx < BCH_TRANSACTIONS_NR) {
3007 trans->fn = bch2_btree_transaction_fns[fn_idx];
616928c3 3008
31403dca 3009 struct btree_transaction_stats *s = &c->btree_transaction_stats[fn_idx];
e131b6aa 3010
31403dca
KO		 3011 if (s->max_mem) {
3012 unsigned expected_mem_bytes = roundup_pow_of_two(s->max_mem);
3013
3014 trans->mem = kmalloc(expected_mem_bytes, GFP_KERNEL);
3015 if (likely(trans->mem))
3016 trans->mem_bytes = expected_mem_bytes;
e131b6aa 3017 }
920e69bc 3018
6474b706 3019 trans->nr_paths_max = s->nr_max_paths;
24de63da 3020 trans->journal_entries_size = s->journal_entries_size;
920e69bc 3021 }
5c0bb66a 3022
c4accde4 3023 trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
e242b92a 3024 trans->srcu_lock_time = jiffies;
c4accde4 3025 trans->srcu_held = true;
6bd68ec2 3026 return trans;
1c6fdbd8
KO		 3027}
3028
67e0dd8f
KO		 3029static void check_btree_paths_leaked(struct btree_trans *trans)
3030{
3031#ifdef CONFIG_BCACHEFS_DEBUG
3032 struct bch_fs *c = trans->c;
3033 struct btree_path *path;
ccb7b08f 3034 unsigned i;
67e0dd8f 3035
ccb7b08f 3036 trans_for_each_path(trans, path, i)
67e0dd8f
KO		 3037 if (path->ref)
3038 goto leaked;
3039 return;
3040leaked:
669f87a5 3041 bch_err(c, "btree paths leaked from %s!", trans->fn);
ccb7b08f 3042 trans_for_each_path(trans, path, i)
67e0dd8f
KO		 3043 if (path->ref)
3044 printk(KERN_ERR " btree %s %pS\n",
88dfe193 3045 bch2_btree_id_str(path->btree_id),
67e0dd8f
KO		 3046 (void *) path->ip_allocated);
3047 /* Be noisy about this: */
3048 bch2_fatal_error(c);
3049#endif
3050}
3051
6bd68ec2 3052void bch2_trans_put(struct btree_trans *trans)
c0ebe3e4 3053 __releases(&c->btree_trans_barrier)
1c6fdbd8 3054{
1a21bf98
KO		 3055 struct bch_fs *c = trans->c;
3056
ece254b2 3057 bch2_trans_unlock(trans);
1c6fdbd8 3058
67e0dd8f 3059 trans_for_each_update(trans, i)
7f9821a7 3060 __btree_path_put(trans->paths + i->path, true);
31403dca
KO		 3061 trans->nr_updates = 0;
3062 trans->locking_wait.task = NULL;
509d3e0a 3063
67e0dd8f 3064 check_btree_paths_leaked(trans);
50dc0f69 3065
f82755e4
KO		 3066 if (trans->srcu_held) {
3067 check_srcu_held_too_long(trans);
c4accde4 3068 srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
f82755e4 3069 }
876c7af3 3070
9620c3ec
KO		 3071 if (trans->fs_usage_deltas) {
3072 if (trans->fs_usage_deltas->size + sizeof(trans->fs_usage_deltas) ==
3073 REPLICAS_DELTA_LIST_MAX)
3074 mempool_free(trans->fs_usage_deltas,
67e0dd8f 3075 &c->replicas_delta_pool);
9620c3ec
KO		 3076 else
3077 kfree(trans->fs_usage_deltas);
3078 }
e131b6aa 3079
8a443d3e
KO		 3080 if (unlikely(trans->journal_replay_not_finished))
3081 bch2_journal_keys_put(c);
3082
0c99e17d
KO		 3083 unsigned long *paths_allocated = trans->paths_allocated;
3084 trans->paths_allocated = NULL;
3085 trans->paths = NULL;
3086
3087 if (paths_allocated != trans->_paths_allocated)
3088 kfree_rcu_mightsleep(paths_allocated);
3089
e131b6aa 3090 if (trans->mem_bytes == BTREE_TRANS_MEM_MAX)
67e0dd8f 3091 mempool_free(trans->mem, &c->btree_trans_mem_pool);
e131b6aa
KO		 3092 else
3093 kfree(trans->mem);
1a21bf98 3094
6bd68ec2
KO		 3095 /* Userspace doesn't have a real percpu implementation: */
3096 if (IS_ENABLED(__KERNEL__))
3097 trans = this_cpu_xchg(c->btree_trans_bufs->trans, trans);
31403dca
KO		 3098
3099 if (trans) {
3100 closure_sync(&trans->ref);
3101
3102 seqmutex_lock(&c->btree_trans_lock);
3103 list_del(&trans->list);
3104 seqmutex_unlock(&c->btree_trans_lock);
3105
6bd68ec2 3106 mempool_free(trans, &c->btree_trans_pool);
31403dca 3107 }
1c6fdbd8 3108}
36e9d698 3109
b3d1e6ca 3110static void __maybe_unused
afbc7194
KO		 3111bch2_btree_bkey_cached_common_to_text(struct printbuf *out,
3112 struct btree_bkey_cached_common *b)
d211b408 3113{
c240c3a9
KO		 3114 struct six_lock_count c = six_lock_counts(&b->lock);
3115 struct task_struct *owner;
3116 pid_t pid;
3117
3118 rcu_read_lock();
3119 owner = READ_ONCE(b->lock.owner);
3e3e02e6 3120 pid = owner ? owner->pid : 0;
c240c3a9
KO		 3121 rcu_read_unlock();
3122
afbc7194
KO		 3123 prt_tab(out);
3124 prt_printf(out, "%px %c l=%u %s:", b, b->cached ? 'c' : 'b',
88dfe193 3125 b->level, bch2_btree_id_str(b->btree_id));
4e6defd1 3126 bch2_bpos_to_text(out, btree_node_pos(b));
c240c3a9 3127
afbc7194
KO		 3128 prt_tab(out);
3129 prt_printf(out, " locks %u:%u:%u held by pid %u",
c240c3a9 3130 c.n[0], c.n[1], c.n[2], pid);
d211b408
KO		 3131}
3132
50b13bee 3133void bch2_btree_trans_to_text(struct printbuf *out, struct btree_trans *trans)
5c1d808a 3134{
86b74451 3135 struct btree_bkey_cached_common *b;
c7ce2732 3136 static char lock_types[] = { 'r', 'i', 'w' };
33981244 3137 struct task_struct *task = READ_ONCE(trans->locking_wait.task);
f154c3eb 3138 unsigned l, idx;
495aabed 3139
fea153a8
KO		 3140 /* before rcu_read_lock(): */
3141 bch2_printbuf_make_room(out, 4096);
3142
afbc7194
KO		 3143 if (!out->nr_tabstops) {
3144 printbuf_tabstop_push(out, 16);
3145 printbuf_tabstop_push(out, 32);
3146 }
3147
33981244 3148 prt_printf(out, "%i %s\n", task ? task->pid : 0, trans->fn);
495aabed 3149
fea153a8
KO		 3150 /* trans->paths is rcu protected vs. freeing */
3151 rcu_read_lock();
3152 out->atomic++;
3153
3154 struct btree_path *paths = rcu_dereference(trans->paths);
3155 if (!paths)
3156 goto out;
3157
3158 unsigned long *paths_allocated = trans_paths_allocated(paths);
3159
3160 trans_for_each_path_idx_from(paths_allocated, *trans_paths_nr(paths), idx, 1) {
3161 struct btree_path *path = paths + idx;
50b13bee
KO		 3162 if (!path->nodes_locked)
3163 continue;
495aabed 3164
50b13bee 3165 prt_printf(out, " path %u %c l=%u %s:",
1f75ba4e 3166 idx,
50b13bee
KO		 3167 path->cached ? 'c' : 'b',
3168 path->level,
88dfe193 3169 bch2_btree_id_str(path->btree_id));
50b13bee 3170 bch2_bpos_to_text(out, path->pos);
afbc7194 3171 prt_newline(out);
50b13bee
KO		 3172
3173 for (l = 0; l < BTREE_MAX_DEPTH; l++) {
86b74451 3174 if (btree_node_locked(path, l) &&
315c9ba6 3175 !IS_ERR_OR_NULL(b = (void *) READ_ONCE(path->l[l].b))) {
d4263e56
KO		 3176 prt_printf(out, " %c l=%u ",
3177 lock_types[btree_node_locked_type(path, l)], l);
afbc7194
KO		 3178 bch2_btree_bkey_cached_common_to_text(out, b);
3179 prt_newline(out);
495aabed
KO		 3180 }
3181 }
50b13bee 3182 }
495aabed 3183
50b13bee
KO		 3184 b = READ_ONCE(trans->locking);
3185 if (b) {
9d7f2a41
KO		 3186 prt_printf(out, " blocked for %lluus on",
3187 div_u64(local_clock() - trans->locking_wait.start_time,
3188 1000));
afbc7194
KO		 3189 prt_newline(out);
3190 prt_printf(out, " %c", lock_types[trans->locking_wait.lock_want]);
3191 bch2_btree_bkey_cached_common_to_text(out, b);
3192 prt_newline(out);
495aabed 3193 }
fea153a8
KO		 3194out:
3195 --out->atomic;
3196 rcu_read_unlock();
495aabed
KO		 3197}
3198
36e9d698
KO		 3199void bch2_fs_btree_iter_exit(struct bch_fs *c)
3200{
4aba7d45 3201 struct btree_transaction_stats *s;
6bd68ec2
KO		 3202 struct btree_trans *trans;
3203 int cpu;
3204
31403dca
KO		 3205 if (c->btree_trans_bufs)
3206 for_each_possible_cpu(cpu) {
3207 struct btree_trans *trans =
3208 per_cpu_ptr(c->btree_trans_bufs, cpu)->trans;
3209
3210 if (trans) {
3211 closure_sync(&trans->ref);
3212
3213 seqmutex_lock(&c->btree_trans_lock);
3214 list_del(&trans->list);
3215 seqmutex_unlock(&c->btree_trans_lock);
3216 }
3217 kfree(trans);
3218 }
3219 free_percpu(c->btree_trans_bufs);
3220
6bd68ec2
KO		 3221 trans = list_first_entry_or_null(&c->btree_trans_list, struct btree_trans, list);
3222 if (trans)
3223 panic("%s leaked btree_trans\n", trans->fn);
3224
4aba7d45
KO		 3225 for (s = c->btree_transaction_stats;
3226 s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
5c0bb66a
KO		 3227 s++) {
3228 kfree(s->max_paths_text);
4aba7d45 3229 bch2_time_stats_exit(&s->lock_hold_times);
5c0bb66a 3230 }
4aba7d45 3231
99fafb04
KO		 3232 if (c->btree_trans_barrier_initialized)
3233 cleanup_srcu_struct(&c->btree_trans_barrier);
e131b6aa 3234 mempool_exit(&c->btree_trans_mem_pool);
6bd68ec2 3235 mempool_exit(&c->btree_trans_pool);
36e9d698
KO		 3236}
3237
50a8a732 3238void bch2_fs_btree_iter_init_early(struct bch_fs *c)
36e9d698 3239{
4aba7d45 3240 struct btree_transaction_stats *s;
36e9d698 3241
4aba7d45
KO		 3242 for (s = c->btree_transaction_stats;
3243 s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
5c0bb66a 3244 s++) {
89056f24 3245 bch2_time_stats_init(&s->duration);
4aba7d45 3246 bch2_time_stats_init(&s->lock_hold_times);
5c0bb66a
KO		 3247 mutex_init(&s->lock);
3248 }
4aba7d45 3249
495aabed 3250 INIT_LIST_HEAD(&c->btree_trans_list);
a5b696ee 3251 seqmutex_init(&c->btree_trans_lock);
50a8a732
TB		 3252}
3253
3254int bch2_fs_btree_iter_init(struct bch_fs *c)
3255{
3256 int ret;
495aabed 3257
6bd68ec2
KO		 3258 c->btree_trans_bufs = alloc_percpu(struct btree_trans_buf);
3259 if (!c->btree_trans_bufs)
3260 return -ENOMEM;
3261
3262 ret = mempool_init_kmalloc_pool(&c->btree_trans_pool, 1,
3263 sizeof(struct btree_trans)) ?:
e131b6aa 3264 mempool_init_kmalloc_pool(&c->btree_trans_mem_pool, 1,
99fafb04
KO		 3265 BTREE_TRANS_MEM_MAX) ?:
3266 init_srcu_struct(&c->btree_trans_barrier);
3267 if (!ret)
3268 c->btree_trans_barrier_initialized = true;
3269 return ret;
36e9d698 3270}
Linux 6.x block layer and io_uring tree(s)