227094efd05072eda0bef61a6c18c03237b5184d
[sfrench/cifs-2.6.git] / fs / btrfs / delayed-ref.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2009 Oracle.  All rights reserved.
4  */
5
6 #include <linux/sched.h>
7 #include <linux/slab.h>
8 #include <linux/sort.h>
9 #include "ctree.h"
10 #include "delayed-ref.h"
11 #include "transaction.h"
12 #include "qgroup.h"
13
14 struct kmem_cache *btrfs_delayed_ref_head_cachep;
15 struct kmem_cache *btrfs_delayed_tree_ref_cachep;
16 struct kmem_cache *btrfs_delayed_data_ref_cachep;
17 struct kmem_cache *btrfs_delayed_extent_op_cachep;
18 /*
19  * delayed back reference update tracking.  For subvolume trees
20  * we queue up extent allocations and backref maintenance for
21  * delayed processing.   This avoids deep call chains where we
22  * add extents in the middle of btrfs_search_slot, and it allows
23  * us to buffer up frequently modified backrefs in an rb tree instead
24  * of hammering updates on the extent allocation tree.
25  */
26
27 /*
28  * compare two delayed tree backrefs with same bytenr and type
29  */
30 static int comp_tree_refs(struct btrfs_delayed_tree_ref *ref1,
31                           struct btrfs_delayed_tree_ref *ref2)
32 {
33         if (ref1->node.type == BTRFS_TREE_BLOCK_REF_KEY) {
34                 if (ref1->root < ref2->root)
35                         return -1;
36                 if (ref1->root > ref2->root)
37                         return 1;
38         } else {
39                 if (ref1->parent < ref2->parent)
40                         return -1;
41                 if (ref1->parent > ref2->parent)
42                         return 1;
43         }
44         return 0;
45 }
46
47 /*
48  * compare two delayed data backrefs with same bytenr and type
49  */
50 static int comp_data_refs(struct btrfs_delayed_data_ref *ref1,
51                           struct btrfs_delayed_data_ref *ref2)
52 {
53         if (ref1->node.type == BTRFS_EXTENT_DATA_REF_KEY) {
54                 if (ref1->root < ref2->root)
55                         return -1;
56                 if (ref1->root > ref2->root)
57                         return 1;
58                 if (ref1->objectid < ref2->objectid)
59                         return -1;
60                 if (ref1->objectid > ref2->objectid)
61                         return 1;
62                 if (ref1->offset < ref2->offset)
63                         return -1;
64                 if (ref1->offset > ref2->offset)
65                         return 1;
66         } else {
67                 if (ref1->parent < ref2->parent)
68                         return -1;
69                 if (ref1->parent > ref2->parent)
70                         return 1;
71         }
72         return 0;
73 }
74
75 static int comp_refs(struct btrfs_delayed_ref_node *ref1,
76                      struct btrfs_delayed_ref_node *ref2,
77                      bool check_seq)
78 {
79         int ret = 0;
80
81         if (ref1->type < ref2->type)
82                 return -1;
83         if (ref1->type > ref2->type)
84                 return 1;
85         if (ref1->type == BTRFS_TREE_BLOCK_REF_KEY ||
86             ref1->type == BTRFS_SHARED_BLOCK_REF_KEY)
87                 ret = comp_tree_refs(btrfs_delayed_node_to_tree_ref(ref1),
88                                      btrfs_delayed_node_to_tree_ref(ref2));
89         else
90                 ret = comp_data_refs(btrfs_delayed_node_to_data_ref(ref1),
91                                      btrfs_delayed_node_to_data_ref(ref2));
92         if (ret)
93                 return ret;
94         if (check_seq) {
95                 if (ref1->seq < ref2->seq)
96                         return -1;
97                 if (ref1->seq > ref2->seq)
98                         return 1;
99         }
100         return 0;
101 }
102
103 /* insert a new ref to head ref rbtree */
104 static struct btrfs_delayed_ref_head *htree_insert(struct rb_root *root,
105                                                    struct rb_node *node)
106 {
107         struct rb_node **p = &root->rb_node;
108         struct rb_node *parent_node = NULL;
109         struct btrfs_delayed_ref_head *entry;
110         struct btrfs_delayed_ref_head *ins;
111         u64 bytenr;
112
113         ins = rb_entry(node, struct btrfs_delayed_ref_head, href_node);
114         bytenr = ins->bytenr;
115         while (*p) {
116                 parent_node = *p;
117                 entry = rb_entry(parent_node, struct btrfs_delayed_ref_head,
118                                  href_node);
119
120                 if (bytenr < entry->bytenr)
121                         p = &(*p)->rb_left;
122                 else if (bytenr > entry->bytenr)
123                         p = &(*p)->rb_right;
124                 else
125                         return entry;
126         }
127
128         rb_link_node(node, parent_node, p);
129         rb_insert_color(node, root);
130         return NULL;
131 }
132
133 static struct btrfs_delayed_ref_node* tree_insert(struct rb_root *root,
134                 struct btrfs_delayed_ref_node *ins)
135 {
136         struct rb_node **p = &root->rb_node;
137         struct rb_node *node = &ins->ref_node;
138         struct rb_node *parent_node = NULL;
139         struct btrfs_delayed_ref_node *entry;
140
141         while (*p) {
142                 int comp;
143
144                 parent_node = *p;
145                 entry = rb_entry(parent_node, struct btrfs_delayed_ref_node,
146                                  ref_node);
147                 comp = comp_refs(ins, entry, true);
148                 if (comp < 0)
149                         p = &(*p)->rb_left;
150                 else if (comp > 0)
151                         p = &(*p)->rb_right;
152                 else
153                         return entry;
154         }
155
156         rb_link_node(node, parent_node, p);
157         rb_insert_color(node, root);
158         return NULL;
159 }
160
161 /*
162  * find an head entry based on bytenr. This returns the delayed ref
163  * head if it was able to find one, or NULL if nothing was in that spot.
164  * If return_bigger is given, the next bigger entry is returned if no exact
165  * match is found.
166  */
167 static struct btrfs_delayed_ref_head *
168 find_ref_head(struct rb_root *root, u64 bytenr,
169               int return_bigger)
170 {
171         struct rb_node *n;
172         struct btrfs_delayed_ref_head *entry;
173
174         n = root->rb_node;
175         entry = NULL;
176         while (n) {
177                 entry = rb_entry(n, struct btrfs_delayed_ref_head, href_node);
178
179                 if (bytenr < entry->bytenr)
180                         n = n->rb_left;
181                 else if (bytenr > entry->bytenr)
182                         n = n->rb_right;
183                 else
184                         return entry;
185         }
186         if (entry && return_bigger) {
187                 if (bytenr > entry->bytenr) {
188                         n = rb_next(&entry->href_node);
189                         if (!n)
190                                 n = rb_first(root);
191                         entry = rb_entry(n, struct btrfs_delayed_ref_head,
192                                          href_node);
193                         return entry;
194                 }
195                 return entry;
196         }
197         return NULL;
198 }
199
200 int btrfs_delayed_ref_lock(struct btrfs_trans_handle *trans,
201                            struct btrfs_delayed_ref_head *head)
202 {
203         struct btrfs_delayed_ref_root *delayed_refs;
204
205         delayed_refs = &trans->transaction->delayed_refs;
206         lockdep_assert_held(&delayed_refs->lock);
207         if (mutex_trylock(&head->mutex))
208                 return 0;
209
210         refcount_inc(&head->refs);
211         spin_unlock(&delayed_refs->lock);
212
213         mutex_lock(&head->mutex);
214         spin_lock(&delayed_refs->lock);
215         if (RB_EMPTY_NODE(&head->href_node)) {
216                 mutex_unlock(&head->mutex);
217                 btrfs_put_delayed_ref_head(head);
218                 return -EAGAIN;
219         }
220         btrfs_put_delayed_ref_head(head);
221         return 0;
222 }
223
224 static inline void drop_delayed_ref(struct btrfs_trans_handle *trans,
225                                     struct btrfs_delayed_ref_root *delayed_refs,
226                                     struct btrfs_delayed_ref_head *head,
227                                     struct btrfs_delayed_ref_node *ref)
228 {
229         lockdep_assert_held(&head->lock);
230         rb_erase(&ref->ref_node, &head->ref_tree);
231         RB_CLEAR_NODE(&ref->ref_node);
232         if (!list_empty(&ref->add_list))
233                 list_del(&ref->add_list);
234         ref->in_tree = 0;
235         btrfs_put_delayed_ref(ref);
236         atomic_dec(&delayed_refs->num_entries);
237         if (trans->delayed_ref_updates)
238                 trans->delayed_ref_updates--;
239 }
240
241 static bool merge_ref(struct btrfs_trans_handle *trans,
242                       struct btrfs_delayed_ref_root *delayed_refs,
243                       struct btrfs_delayed_ref_head *head,
244                       struct btrfs_delayed_ref_node *ref,
245                       u64 seq)
246 {
247         struct btrfs_delayed_ref_node *next;
248         struct rb_node *node = rb_next(&ref->ref_node);
249         bool done = false;
250
251         while (!done && node) {
252                 int mod;
253
254                 next = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
255                 node = rb_next(node);
256                 if (seq && next->seq >= seq)
257                         break;
258                 if (comp_refs(ref, next, false))
259                         break;
260
261                 if (ref->action == next->action) {
262                         mod = next->ref_mod;
263                 } else {
264                         if (ref->ref_mod < next->ref_mod) {
265                                 swap(ref, next);
266                                 done = true;
267                         }
268                         mod = -next->ref_mod;
269                 }
270
271                 drop_delayed_ref(trans, delayed_refs, head, next);
272                 ref->ref_mod += mod;
273                 if (ref->ref_mod == 0) {
274                         drop_delayed_ref(trans, delayed_refs, head, ref);
275                         done = true;
276                 } else {
277                         /*
278                          * Can't have multiples of the same ref on a tree block.
279                          */
280                         WARN_ON(ref->type == BTRFS_TREE_BLOCK_REF_KEY ||
281                                 ref->type == BTRFS_SHARED_BLOCK_REF_KEY);
282                 }
283         }
284
285         return done;
286 }
287
288 void btrfs_merge_delayed_refs(struct btrfs_trans_handle *trans,
289                               struct btrfs_delayed_ref_root *delayed_refs,
290                               struct btrfs_delayed_ref_head *head)
291 {
292         struct btrfs_fs_info *fs_info = trans->fs_info;
293         struct btrfs_delayed_ref_node *ref;
294         struct rb_node *node;
295         u64 seq = 0;
296
297         lockdep_assert_held(&head->lock);
298
299         if (RB_EMPTY_ROOT(&head->ref_tree))
300                 return;
301
302         /* We don't have too many refs to merge for data. */
303         if (head->is_data)
304                 return;
305
306         spin_lock(&fs_info->tree_mod_seq_lock);
307         if (!list_empty(&fs_info->tree_mod_seq_list)) {
308                 struct seq_list *elem;
309
310                 elem = list_first_entry(&fs_info->tree_mod_seq_list,
311                                         struct seq_list, list);
312                 seq = elem->seq;
313         }
314         spin_unlock(&fs_info->tree_mod_seq_lock);
315
316 again:
317         for (node = rb_first(&head->ref_tree); node; node = rb_next(node)) {
318                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
319                 if (seq && ref->seq >= seq)
320                         continue;
321                 if (merge_ref(trans, delayed_refs, head, ref, seq))
322                         goto again;
323         }
324 }
325
326 int btrfs_check_delayed_seq(struct btrfs_fs_info *fs_info, u64 seq)
327 {
328         struct seq_list *elem;
329         int ret = 0;
330
331         spin_lock(&fs_info->tree_mod_seq_lock);
332         if (!list_empty(&fs_info->tree_mod_seq_list)) {
333                 elem = list_first_entry(&fs_info->tree_mod_seq_list,
334                                         struct seq_list, list);
335                 if (seq >= elem->seq) {
336                         btrfs_debug(fs_info,
337                                 "holding back delayed_ref %#x.%x, lowest is %#x.%x",
338                                 (u32)(seq >> 32), (u32)seq,
339                                 (u32)(elem->seq >> 32), (u32)elem->seq);
340                         ret = 1;
341                 }
342         }
343
344         spin_unlock(&fs_info->tree_mod_seq_lock);
345         return ret;
346 }
347
348 struct btrfs_delayed_ref_head *
349 btrfs_select_ref_head(struct btrfs_trans_handle *trans)
350 {
351         struct btrfs_delayed_ref_root *delayed_refs;
352         struct btrfs_delayed_ref_head *head;
353         u64 start;
354         bool loop = false;
355
356         delayed_refs = &trans->transaction->delayed_refs;
357
358 again:
359         start = delayed_refs->run_delayed_start;
360         head = find_ref_head(&delayed_refs->href_root, start, 1);
361         if (!head && !loop) {
362                 delayed_refs->run_delayed_start = 0;
363                 start = 0;
364                 loop = true;
365                 head = find_ref_head(&delayed_refs->href_root, start, 1);
366                 if (!head)
367                         return NULL;
368         } else if (!head && loop) {
369                 return NULL;
370         }
371
372         while (head->processing) {
373                 struct rb_node *node;
374
375                 node = rb_next(&head->href_node);
376                 if (!node) {
377                         if (loop)
378                                 return NULL;
379                         delayed_refs->run_delayed_start = 0;
380                         start = 0;
381                         loop = true;
382                         goto again;
383                 }
384                 head = rb_entry(node, struct btrfs_delayed_ref_head,
385                                 href_node);
386         }
387
388         head->processing = 1;
389         WARN_ON(delayed_refs->num_heads_ready == 0);
390         delayed_refs->num_heads_ready--;
391         delayed_refs->run_delayed_start = head->bytenr +
392                 head->num_bytes;
393         return head;
394 }
395
396 /*
397  * Helper to insert the ref_node to the tail or merge with tail.
398  *
399  * Return 0 for insert.
400  * Return >0 for merge.
401  */
402 static int insert_delayed_ref(struct btrfs_trans_handle *trans,
403                               struct btrfs_delayed_ref_root *root,
404                               struct btrfs_delayed_ref_head *href,
405                               struct btrfs_delayed_ref_node *ref)
406 {
407         struct btrfs_delayed_ref_node *exist;
408         int mod;
409         int ret = 0;
410
411         spin_lock(&href->lock);
412         exist = tree_insert(&href->ref_tree, ref);
413         if (!exist)
414                 goto inserted;
415
416         /* Now we are sure we can merge */
417         ret = 1;
418         if (exist->action == ref->action) {
419                 mod = ref->ref_mod;
420         } else {
421                 /* Need to change action */
422                 if (exist->ref_mod < ref->ref_mod) {
423                         exist->action = ref->action;
424                         mod = -exist->ref_mod;
425                         exist->ref_mod = ref->ref_mod;
426                         if (ref->action == BTRFS_ADD_DELAYED_REF)
427                                 list_add_tail(&exist->add_list,
428                                               &href->ref_add_list);
429                         else if (ref->action == BTRFS_DROP_DELAYED_REF) {
430                                 ASSERT(!list_empty(&exist->add_list));
431                                 list_del(&exist->add_list);
432                         } else {
433                                 ASSERT(0);
434                         }
435                 } else
436                         mod = -ref->ref_mod;
437         }
438         exist->ref_mod += mod;
439
440         /* remove existing tail if its ref_mod is zero */
441         if (exist->ref_mod == 0)
442                 drop_delayed_ref(trans, root, href, exist);
443         spin_unlock(&href->lock);
444         return ret;
445 inserted:
446         if (ref->action == BTRFS_ADD_DELAYED_REF)
447                 list_add_tail(&ref->add_list, &href->ref_add_list);
448         atomic_inc(&root->num_entries);
449         trans->delayed_ref_updates++;
450         spin_unlock(&href->lock);
451         return ret;
452 }
453
454 /*
455  * helper function to update the accounting in the head ref
456  * existing and update must have the same bytenr
457  */
458 static noinline void
459 update_existing_head_ref(struct btrfs_delayed_ref_root *delayed_refs,
460                          struct btrfs_delayed_ref_head *existing,
461                          struct btrfs_delayed_ref_head *update,
462                          int *old_ref_mod_ret)
463 {
464         int old_ref_mod;
465
466         BUG_ON(existing->is_data != update->is_data);
467
468         spin_lock(&existing->lock);
469         if (update->must_insert_reserved) {
470                 /* if the extent was freed and then
471                  * reallocated before the delayed ref
472                  * entries were processed, we can end up
473                  * with an existing head ref without
474                  * the must_insert_reserved flag set.
475                  * Set it again here
476                  */
477                 existing->must_insert_reserved = update->must_insert_reserved;
478
479                 /*
480                  * update the num_bytes so we make sure the accounting
481                  * is done correctly
482                  */
483                 existing->num_bytes = update->num_bytes;
484
485         }
486
487         if (update->extent_op) {
488                 if (!existing->extent_op) {
489                         existing->extent_op = update->extent_op;
490                 } else {
491                         if (update->extent_op->update_key) {
492                                 memcpy(&existing->extent_op->key,
493                                        &update->extent_op->key,
494                                        sizeof(update->extent_op->key));
495                                 existing->extent_op->update_key = true;
496                         }
497                         if (update->extent_op->update_flags) {
498                                 existing->extent_op->flags_to_set |=
499                                         update->extent_op->flags_to_set;
500                                 existing->extent_op->update_flags = true;
501                         }
502                         btrfs_free_delayed_extent_op(update->extent_op);
503                 }
504         }
505         /*
506          * update the reference mod on the head to reflect this new operation,
507          * only need the lock for this case cause we could be processing it
508          * currently, for refs we just added we know we're a-ok.
509          */
510         old_ref_mod = existing->total_ref_mod;
511         if (old_ref_mod_ret)
512                 *old_ref_mod_ret = old_ref_mod;
513         existing->ref_mod += update->ref_mod;
514         existing->total_ref_mod += update->ref_mod;
515
516         /*
517          * If we are going to from a positive ref mod to a negative or vice
518          * versa we need to make sure to adjust pending_csums accordingly.
519          */
520         if (existing->is_data) {
521                 if (existing->total_ref_mod >= 0 && old_ref_mod < 0)
522                         delayed_refs->pending_csums -= existing->num_bytes;
523                 if (existing->total_ref_mod < 0 && old_ref_mod >= 0)
524                         delayed_refs->pending_csums += existing->num_bytes;
525         }
526         spin_unlock(&existing->lock);
527 }
528
529 static void init_delayed_ref_head(struct btrfs_delayed_ref_head *head_ref,
530                                   struct btrfs_qgroup_extent_record *qrecord,
531                                   u64 bytenr, u64 num_bytes, u64 ref_root,
532                                   u64 reserved, int action, bool is_data,
533                                   bool is_system)
534 {
535         int count_mod = 1;
536         int must_insert_reserved = 0;
537
538         /* If reserved is provided, it must be a data extent. */
539         BUG_ON(!is_data && reserved);
540
541         /*
542          * The head node stores the sum of all the mods, so dropping a ref
543          * should drop the sum in the head node by one.
544          */
545         if (action == BTRFS_UPDATE_DELAYED_HEAD)
546                 count_mod = 0;
547         else if (action == BTRFS_DROP_DELAYED_REF)
548                 count_mod = -1;
549
550         /*
551          * BTRFS_ADD_DELAYED_EXTENT means that we need to update the reserved
552          * accounting when the extent is finally added, or if a later
553          * modification deletes the delayed ref without ever inserting the
554          * extent into the extent allocation tree.  ref->must_insert_reserved
555          * is the flag used to record that accounting mods are required.
556          *
557          * Once we record must_insert_reserved, switch the action to
558          * BTRFS_ADD_DELAYED_REF because other special casing is not required.
559          */
560         if (action == BTRFS_ADD_DELAYED_EXTENT)
561                 must_insert_reserved = 1;
562         else
563                 must_insert_reserved = 0;
564
565         refcount_set(&head_ref->refs, 1);
566         head_ref->bytenr = bytenr;
567         head_ref->num_bytes = num_bytes;
568         head_ref->ref_mod = count_mod;
569         head_ref->must_insert_reserved = must_insert_reserved;
570         head_ref->is_data = is_data;
571         head_ref->is_system = is_system;
572         head_ref->ref_tree = RB_ROOT;
573         INIT_LIST_HEAD(&head_ref->ref_add_list);
574         RB_CLEAR_NODE(&head_ref->href_node);
575         head_ref->processing = 0;
576         head_ref->total_ref_mod = count_mod;
577         head_ref->qgroup_reserved = 0;
578         head_ref->qgroup_ref_root = 0;
579         spin_lock_init(&head_ref->lock);
580         mutex_init(&head_ref->mutex);
581
582         if (qrecord) {
583                 if (ref_root && reserved) {
584                         head_ref->qgroup_ref_root = ref_root;
585                         head_ref->qgroup_reserved = reserved;
586                 }
587
588                 qrecord->bytenr = bytenr;
589                 qrecord->num_bytes = num_bytes;
590                 qrecord->old_roots = NULL;
591         }
592 }
593
594 /*
595  * helper function to actually insert a head node into the rbtree.
596  * this does all the dirty work in terms of maintaining the correct
597  * overall modification count.
598  */
599 static noinline struct btrfs_delayed_ref_head *
600 add_delayed_ref_head(struct btrfs_trans_handle *trans,
601                      struct btrfs_delayed_ref_head *head_ref,
602                      struct btrfs_qgroup_extent_record *qrecord,
603                      u64 bytenr, u64 num_bytes, u64 ref_root, u64 reserved,
604                      int action, int is_data, int is_system,
605                      int *qrecord_inserted_ret,
606                      int *old_ref_mod, int *new_ref_mod)
607
608 {
609         struct btrfs_delayed_ref_head *existing;
610         struct btrfs_delayed_ref_root *delayed_refs;
611         int count_mod = 1;
612         int must_insert_reserved = 0;
613         int qrecord_inserted = 0;
614
615         /* If reserved is provided, it must be a data extent. */
616         BUG_ON(!is_data && reserved);
617
618         /*
619          * the head node stores the sum of all the mods, so dropping a ref
620          * should drop the sum in the head node by one.
621          */
622         if (action == BTRFS_UPDATE_DELAYED_HEAD)
623                 count_mod = 0;
624         else if (action == BTRFS_DROP_DELAYED_REF)
625                 count_mod = -1;
626
627         /*
628          * BTRFS_ADD_DELAYED_EXTENT means that we need to update
629          * the reserved accounting when the extent is finally added, or
630          * if a later modification deletes the delayed ref without ever
631          * inserting the extent into the extent allocation tree.
632          * ref->must_insert_reserved is the flag used to record
633          * that accounting mods are required.
634          *
635          * Once we record must_insert_reserved, switch the action to
636          * BTRFS_ADD_DELAYED_REF because other special casing is not required.
637          */
638         if (action == BTRFS_ADD_DELAYED_EXTENT)
639                 must_insert_reserved = 1;
640         else
641                 must_insert_reserved = 0;
642
643         delayed_refs = &trans->transaction->delayed_refs;
644
645         refcount_set(&head_ref->refs, 1);
646         head_ref->bytenr = bytenr;
647         head_ref->num_bytes = num_bytes;
648         head_ref->ref_mod = count_mod;
649         head_ref->must_insert_reserved = must_insert_reserved;
650         head_ref->is_data = is_data;
651         head_ref->is_system = is_system;
652         head_ref->ref_tree = RB_ROOT;
653         INIT_LIST_HEAD(&head_ref->ref_add_list);
654         RB_CLEAR_NODE(&head_ref->href_node);
655         head_ref->processing = 0;
656         head_ref->total_ref_mod = count_mod;
657         head_ref->qgroup_reserved = 0;
658         head_ref->qgroup_ref_root = 0;
659         spin_lock_init(&head_ref->lock);
660         mutex_init(&head_ref->mutex);
661
662         /* Record qgroup extent info if provided */
663         if (qrecord) {
664                 if (ref_root && reserved) {
665                         head_ref->qgroup_ref_root = ref_root;
666                         head_ref->qgroup_reserved = reserved;
667                 }
668
669                 qrecord->bytenr = bytenr;
670                 qrecord->num_bytes = num_bytes;
671                 qrecord->old_roots = NULL;
672
673                 if(btrfs_qgroup_trace_extent_nolock(trans->fs_info,
674                                         delayed_refs, qrecord))
675                         kfree(qrecord);
676                 else
677                         qrecord_inserted = 1;
678         }
679
680         trace_add_delayed_ref_head(trans->fs_info, head_ref, action);
681
682         existing = htree_insert(&delayed_refs->href_root,
683                                 &head_ref->href_node);
684         if (existing) {
685                 WARN_ON(ref_root && reserved && existing->qgroup_ref_root
686                         && existing->qgroup_reserved);
687                 update_existing_head_ref(delayed_refs, existing, head_ref,
688                                          old_ref_mod);
689                 /*
690                  * we've updated the existing ref, free the newly
691                  * allocated ref
692                  */
693                 kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
694                 head_ref = existing;
695         } else {
696                 if (old_ref_mod)
697                         *old_ref_mod = 0;
698                 if (is_data && count_mod < 0)
699                         delayed_refs->pending_csums += num_bytes;
700                 delayed_refs->num_heads++;
701                 delayed_refs->num_heads_ready++;
702                 atomic_inc(&delayed_refs->num_entries);
703                 trans->delayed_ref_updates++;
704         }
705         if (qrecord_inserted_ret)
706                 *qrecord_inserted_ret = qrecord_inserted;
707         if (new_ref_mod)
708                 *new_ref_mod = head_ref->total_ref_mod;
709         return head_ref;
710 }
711
712 /*
713  * init_delayed_ref_common - Initialize the structure which represents a
714  *                           modification to a an extent.
715  *
716  * @fs_info:    Internal to the mounted filesystem mount structure.
717  *
718  * @ref:        The structure which is going to be initialized.
719  *
720  * @bytenr:     The logical address of the extent for which a modification is
721  *              going to be recorded.
722  *
723  * @num_bytes:  Size of the extent whose modification is being recorded.
724  *
725  * @ref_root:   The id of the root where this modification has originated, this
726  *              can be either one of the well-known metadata trees or the
727  *              subvolume id which references this extent.
728  *
729  * @action:     Can be one of BTRFS_ADD_DELAYED_REF/BTRFS_DROP_DELAYED_REF or
730  *              BTRFS_ADD_DELAYED_EXTENT
731  *
732  * @ref_type:   Holds the type of the extent which is being recorded, can be
733  *              one of BTRFS_SHARED_BLOCK_REF_KEY/BTRFS_TREE_BLOCK_REF_KEY
734  *              when recording a metadata extent or BTRFS_SHARED_DATA_REF_KEY/
735  *              BTRFS_EXTENT_DATA_REF_KEY when recording data extent
736  */
737 static void init_delayed_ref_common(struct btrfs_fs_info *fs_info,
738                                     struct btrfs_delayed_ref_node *ref,
739                                     u64 bytenr, u64 num_bytes, u64 ref_root,
740                                     int action, u8 ref_type)
741 {
742         u64 seq = 0;
743
744         if (action == BTRFS_ADD_DELAYED_EXTENT)
745                 action = BTRFS_ADD_DELAYED_REF;
746
747         if (is_fstree(ref_root))
748                 seq = atomic64_read(&fs_info->tree_mod_seq);
749
750         refcount_set(&ref->refs, 1);
751         ref->bytenr = bytenr;
752         ref->num_bytes = num_bytes;
753         ref->ref_mod = 1;
754         ref->action = action;
755         ref->is_head = 0;
756         ref->in_tree = 1;
757         ref->seq = seq;
758         ref->type = ref_type;
759         RB_CLEAR_NODE(&ref->ref_node);
760         INIT_LIST_HEAD(&ref->add_list);
761 }
762
763 /*
764  * add a delayed tree ref.  This does all of the accounting required
765  * to make sure the delayed ref is eventually processed before this
766  * transaction commits.
767  */
768 int btrfs_add_delayed_tree_ref(struct btrfs_fs_info *fs_info,
769                                struct btrfs_trans_handle *trans,
770                                u64 bytenr, u64 num_bytes, u64 parent,
771                                u64 ref_root,  int level, int action,
772                                struct btrfs_delayed_extent_op *extent_op,
773                                int *old_ref_mod, int *new_ref_mod)
774 {
775         struct btrfs_delayed_tree_ref *ref;
776         struct btrfs_delayed_ref_head *head_ref;
777         struct btrfs_delayed_ref_root *delayed_refs;
778         struct btrfs_qgroup_extent_record *record = NULL;
779         int qrecord_inserted;
780         int is_system = (ref_root == BTRFS_CHUNK_TREE_OBJECTID);
781         int ret;
782         u8 ref_type;
783
784         BUG_ON(extent_op && extent_op->is_data);
785         ref = kmem_cache_alloc(btrfs_delayed_tree_ref_cachep, GFP_NOFS);
786         if (!ref)
787                 return -ENOMEM;
788
789         if (parent)
790                 ref_type = BTRFS_SHARED_BLOCK_REF_KEY;
791         else
792                 ref_type = BTRFS_TREE_BLOCK_REF_KEY;
793         init_delayed_ref_common(fs_info, &ref->node, bytenr, num_bytes,
794                                 ref_root, action, ref_type);
795         ref->root = ref_root;
796         ref->parent = parent;
797         ref->level = level;
798
799         head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
800         if (!head_ref)
801                 goto free_ref;
802
803         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) &&
804             is_fstree(ref_root)) {
805                 record = kmalloc(sizeof(*record), GFP_NOFS);
806                 if (!record)
807                         goto free_head_ref;
808         }
809
810         head_ref->extent_op = extent_op;
811
812         delayed_refs = &trans->transaction->delayed_refs;
813         spin_lock(&delayed_refs->lock);
814
815         /*
816          * insert both the head node and the new ref without dropping
817          * the spin lock
818          */
819         head_ref = add_delayed_ref_head(trans, head_ref, record, bytenr,
820                                         num_bytes, 0, 0, action, 0,
821                                         is_system, &qrecord_inserted,
822                                         old_ref_mod, new_ref_mod);
823
824
825         ret = insert_delayed_ref(trans, delayed_refs, head_ref, &ref->node);
826         spin_unlock(&delayed_refs->lock);
827
828         trace_add_delayed_tree_ref(fs_info, &ref->node, ref,
829                                    action == BTRFS_ADD_DELAYED_EXTENT ?
830                                    BTRFS_ADD_DELAYED_REF : action);
831         if (ret > 0)
832                 kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
833
834         if (qrecord_inserted)
835                 btrfs_qgroup_trace_extent_post(fs_info, record);
836
837         return 0;
838
839 free_head_ref:
840         kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
841 free_ref:
842         kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
843
844         return -ENOMEM;
845 }
846
847 /*
848  * add a delayed data ref. it's similar to btrfs_add_delayed_tree_ref.
849  */
850 int btrfs_add_delayed_data_ref(struct btrfs_fs_info *fs_info,
851                                struct btrfs_trans_handle *trans,
852                                u64 bytenr, u64 num_bytes,
853                                u64 parent, u64 ref_root,
854                                u64 owner, u64 offset, u64 reserved, int action,
855                                int *old_ref_mod, int *new_ref_mod)
856 {
857         struct btrfs_delayed_data_ref *ref;
858         struct btrfs_delayed_ref_head *head_ref;
859         struct btrfs_delayed_ref_root *delayed_refs;
860         struct btrfs_qgroup_extent_record *record = NULL;
861         int qrecord_inserted;
862         int ret;
863         u8 ref_type;
864
865         ref = kmem_cache_alloc(btrfs_delayed_data_ref_cachep, GFP_NOFS);
866         if (!ref)
867                 return -ENOMEM;
868
869         if (parent)
870                 ref_type = BTRFS_SHARED_DATA_REF_KEY;
871         else
872                 ref_type = BTRFS_EXTENT_DATA_REF_KEY;
873         init_delayed_ref_common(fs_info, &ref->node, bytenr, num_bytes,
874                                 ref_root, action, ref_type);
875         ref->root = ref_root;
876         ref->parent = parent;
877         ref->objectid = owner;
878         ref->offset = offset;
879
880
881         head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
882         if (!head_ref) {
883                 kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
884                 return -ENOMEM;
885         }
886
887         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) &&
888             is_fstree(ref_root)) {
889                 record = kmalloc(sizeof(*record), GFP_NOFS);
890                 if (!record) {
891                         kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
892                         kmem_cache_free(btrfs_delayed_ref_head_cachep,
893                                         head_ref);
894                         return -ENOMEM;
895                 }
896         }
897
898         head_ref->extent_op = NULL;
899
900         delayed_refs = &trans->transaction->delayed_refs;
901         spin_lock(&delayed_refs->lock);
902
903         /*
904          * insert both the head node and the new ref without dropping
905          * the spin lock
906          */
907         head_ref = add_delayed_ref_head(trans, head_ref, record, bytenr,
908                                         num_bytes, ref_root, reserved,
909                                         action, 1, 0, &qrecord_inserted,
910                                         old_ref_mod, new_ref_mod);
911
912         ret = insert_delayed_ref(trans, delayed_refs, head_ref, &ref->node);
913         spin_unlock(&delayed_refs->lock);
914
915         trace_add_delayed_data_ref(trans->fs_info, &ref->node, ref,
916                                    action == BTRFS_ADD_DELAYED_EXTENT ?
917                                    BTRFS_ADD_DELAYED_REF : action);
918         if (ret > 0)
919                 kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
920
921
922         if (qrecord_inserted)
923                 return btrfs_qgroup_trace_extent_post(fs_info, record);
924         return 0;
925 }
926
927 int btrfs_add_delayed_extent_op(struct btrfs_fs_info *fs_info,
928                                 struct btrfs_trans_handle *trans,
929                                 u64 bytenr, u64 num_bytes,
930                                 struct btrfs_delayed_extent_op *extent_op)
931 {
932         struct btrfs_delayed_ref_head *head_ref;
933         struct btrfs_delayed_ref_root *delayed_refs;
934
935         head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
936         if (!head_ref)
937                 return -ENOMEM;
938
939         head_ref->extent_op = extent_op;
940
941         delayed_refs = &trans->transaction->delayed_refs;
942         spin_lock(&delayed_refs->lock);
943
944         /*
945          * extent_ops just modify the flags of an extent and they don't result
946          * in ref count changes, hence it's safe to pass false/0 for is_system
947          * argument
948          */
949         add_delayed_ref_head(trans, head_ref, NULL, bytenr, num_bytes, 0, 0,
950                              BTRFS_UPDATE_DELAYED_HEAD, extent_op->is_data,
951                              0, NULL, NULL, NULL);
952
953         spin_unlock(&delayed_refs->lock);
954         return 0;
955 }
956
957 /*
958  * this does a simple search for the head node for a given extent.
959  * It must be called with the delayed ref spinlock held, and it returns
960  * the head node if any where found, or NULL if not.
961  */
962 struct btrfs_delayed_ref_head *
963 btrfs_find_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs, u64 bytenr)
964 {
965         return find_ref_head(&delayed_refs->href_root, bytenr, 0);
966 }
967
968 void __cold btrfs_delayed_ref_exit(void)
969 {
970         kmem_cache_destroy(btrfs_delayed_ref_head_cachep);
971         kmem_cache_destroy(btrfs_delayed_tree_ref_cachep);
972         kmem_cache_destroy(btrfs_delayed_data_ref_cachep);
973         kmem_cache_destroy(btrfs_delayed_extent_op_cachep);
974 }
975
976 int __init btrfs_delayed_ref_init(void)
977 {
978         btrfs_delayed_ref_head_cachep = kmem_cache_create(
979                                 "btrfs_delayed_ref_head",
980                                 sizeof(struct btrfs_delayed_ref_head), 0,
981                                 SLAB_MEM_SPREAD, NULL);
982         if (!btrfs_delayed_ref_head_cachep)
983                 goto fail;
984
985         btrfs_delayed_tree_ref_cachep = kmem_cache_create(
986                                 "btrfs_delayed_tree_ref",
987                                 sizeof(struct btrfs_delayed_tree_ref), 0,
988                                 SLAB_MEM_SPREAD, NULL);
989         if (!btrfs_delayed_tree_ref_cachep)
990                 goto fail;
991
992         btrfs_delayed_data_ref_cachep = kmem_cache_create(
993                                 "btrfs_delayed_data_ref",
994                                 sizeof(struct btrfs_delayed_data_ref), 0,
995                                 SLAB_MEM_SPREAD, NULL);
996         if (!btrfs_delayed_data_ref_cachep)
997                 goto fail;
998
999         btrfs_delayed_extent_op_cachep = kmem_cache_create(
1000                                 "btrfs_delayed_extent_op",
1001                                 sizeof(struct btrfs_delayed_extent_op), 0,
1002                                 SLAB_MEM_SPREAD, NULL);
1003         if (!btrfs_delayed_extent_op_cachep)
1004                 goto fail;
1005
1006         return 0;
1007 fail:
1008         btrfs_delayed_ref_exit();
1009         return -ENOMEM;
1010 }