2 * Copyright (c) 2007-2008 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
37 static int hammer_mem_lookup(hammer_cursor_t cursor);
38 static int hammer_mem_first(hammer_cursor_t cursor);
39 static int hammer_frontend_trunc_callback(hammer_record_t record,
41 static int hammer_bulk_scan_callback(hammer_record_t record, void *data);
42 static int hammer_record_needs_overwrite_delete(hammer_record_t record);
43 static int hammer_delete_general(hammer_cursor_t cursor, hammer_inode_t ip,
44 hammer_btree_leaf_elm_t leaf);
45 static int hammer_cursor_localize_data(hammer_data_ondisk_t data,
46 hammer_btree_leaf_elm_t leaf);
48 struct rec_trunc_info {
53 struct hammer_bulk_info {
54 hammer_record_t record;
55 hammer_record_t conflict;
59 * Red-black tree support. Comparison code for insertion.
62 hammer_rec_rb_compare(hammer_record_t rec1, hammer_record_t rec2)
64 if (rec1->leaf.base.rec_type < rec2->leaf.base.rec_type)
66 if (rec1->leaf.base.rec_type > rec2->leaf.base.rec_type)
69 if (rec1->leaf.base.key < rec2->leaf.base.key)
71 if (rec1->leaf.base.key > rec2->leaf.base.key)
75 * For search & insertion purposes records deleted by the
76 * frontend or deleted/committed by the backend are silently
77 * ignored. Otherwise pipelined insertions will get messed
80 * rec1 is greater then rec2 if rec1 is marked deleted.
81 * rec1 is less then rec2 if rec2 is marked deleted.
83 * Multiple deleted records may be present, do not return 0
84 * if both are marked deleted.
86 if (rec1->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
87 HAMMER_RECF_COMMITTED)) {
90 if (rec2->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
91 HAMMER_RECF_COMMITTED)) {
99 * Basic record comparison code similar to hammer_btree_cmp().
101 * obj_id is not compared and may not yet be assigned in the record.
104 hammer_rec_cmp(hammer_base_elm_t elm, hammer_record_t rec)
106 if (elm->rec_type < rec->leaf.base.rec_type)
108 if (elm->rec_type > rec->leaf.base.rec_type)
111 if (elm->key < rec->leaf.base.key)
113 if (elm->key > rec->leaf.base.key)
117 * Never match against an item deleted by the frontend
118 * or backend, or committed by the backend.
120 * elm is less then rec if rec is marked deleted.
122 if (rec->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
123 HAMMER_RECF_COMMITTED)) {
130 * Ranged scan to locate overlapping record(s). This is used by
131 * hammer_ip_get_bulk() to locate an overlapping record. We have
132 * to use a ranged scan because the keys for data records with the
133 * same file base offset can be different due to differing data_len's.
135 * NOTE: The base file offset of a data record is (key - data_len), not (key).
138 hammer_rec_overlap_cmp(hammer_record_t rec, void *data)
140 struct hammer_bulk_info *info = data;
141 hammer_btree_leaf_elm_t leaf = &info->record->leaf;
143 if (rec->leaf.base.rec_type < leaf->base.rec_type)
145 if (rec->leaf.base.rec_type > leaf->base.rec_type)
151 if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) {
152 /* rec_beg >= leaf_end */
153 if (rec->leaf.base.key - rec->leaf.data_len >= leaf->base.key)
155 /* rec_end <= leaf_beg */
156 if (rec->leaf.base.key <= leaf->base.key - leaf->data_len)
159 if (rec->leaf.base.key < leaf->base.key)
161 if (rec->leaf.base.key > leaf->base.key)
166 * We have to return 0 at this point, even if DELETED_FE is set,
167 * because returning anything else will cause the scan to ignore
168 * one of the branches when we really want it to check both.
174 * RB_SCAN comparison code for hammer_mem_first(). The argument order
175 * is reversed so the comparison result has to be negated. key_beg and
176 * key_end are both range-inclusive.
178 * Localized deletions are not cached in-memory.
182 hammer_rec_scan_cmp(hammer_record_t rec, void *data)
184 hammer_cursor_t cursor = data;
187 r = hammer_rec_cmp(&cursor->key_beg, rec);
190 r = hammer_rec_cmp(&cursor->key_end, rec);
197 * This compare function is used when simply looking up key_beg.
201 hammer_rec_find_cmp(hammer_record_t rec, void *data)
203 hammer_cursor_t cursor = data;
206 r = hammer_rec_cmp(&cursor->key_beg, rec);
215 * Locate blocks within the truncation range. Partial blocks do not count.
219 hammer_rec_trunc_cmp(hammer_record_t rec, void *data)
221 struct rec_trunc_info *info = data;
223 if (rec->leaf.base.rec_type < info->rec_type)
225 if (rec->leaf.base.rec_type > info->rec_type)
228 switch(rec->leaf.base.rec_type) {
229 case HAMMER_RECTYPE_DB:
231 * DB record key is not beyond the truncation point, retain.
233 if (rec->leaf.base.key < info->trunc_off)
236 case HAMMER_RECTYPE_DATA:
238 * DATA record offset start is not beyond the truncation point,
241 if (rec->leaf.base.key - rec->leaf.data_len < info->trunc_off)
245 hpanic("unexpected record type");
249 * The record start is >= the truncation point, return match,
250 * the record should be destroyed.
255 RB_GENERATE(hammer_rec_rb_tree, hammer_record, rb_node, hammer_rec_rb_compare);
258 * Allocate a record for the caller to finish filling in. The record is
259 * returned referenced. In order to manually set data call this function
260 * with data_len=0 and then manually set record->leaf.data_len and
261 * record->data later.
264 hammer_alloc_mem_record(hammer_inode_t ip, int data_len)
266 hammer_record_t record;
270 ++hammer_count_records;
271 record = kmalloc(sizeof(*record), hmp->m_misc,
272 M_WAITOK | M_ZERO | M_USE_RESERVE);
273 record->flush_state = HAMMER_FST_IDLE;
275 record->leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
276 record->leaf.data_len = data_len;
277 hammer_ref(&record->lock);
280 record->data = kmalloc(data_len, hmp->m_misc, M_WAITOK | M_ZERO);
281 record->flags |= HAMMER_RECF_ALLOCDATA;
282 ++hammer_count_record_datas;
289 hammer_wait_mem_record_ident(hammer_record_t record, const char *ident)
291 while (record->flush_state == HAMMER_FST_FLUSH) {
292 record->flags |= HAMMER_RECF_WANTED;
293 tsleep(record, 0, ident, 0);
298 * Called from the backend, hammer_inode.c, after a record has been
299 * flushed to disk. The record has been exclusively locked by the
300 * caller and interlocked with BE.
302 * We clean up the state, unlock, and release the record (the record
303 * was referenced by the fact that it was in the HAMMER_FST_FLUSH state).
306 hammer_flush_record_done(hammer_record_t record, int error)
308 hammer_inode_t target_ip;
310 KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
311 KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE);
314 * If an error occured, the backend was unable to sync the
315 * record to its media. Leave the record intact.
318 hammer_critical_error(record->ip->hmp, record->ip, error,
319 "while flushing record");
322 --record->flush_group->refs;
323 record->flush_group = NULL;
326 * Adjust the flush state and dependancy based on success or
329 if (record->flags & (HAMMER_RECF_DELETED_BE | HAMMER_RECF_COMMITTED)) {
330 if ((target_ip = record->target_ip) != NULL) {
331 TAILQ_REMOVE(&target_ip->target_list, record,
333 record->target_ip = NULL;
334 hammer_test_inode(target_ip);
336 record->flush_state = HAMMER_FST_IDLE;
338 if (record->target_ip) {
339 record->flush_state = HAMMER_FST_SETUP;
340 hammer_test_inode(record->ip);
341 hammer_test_inode(record->target_ip);
343 record->flush_state = HAMMER_FST_IDLE;
346 record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
351 if (record->flags & HAMMER_RECF_WANTED) {
352 record->flags &= ~HAMMER_RECF_WANTED;
355 hammer_rel_mem_record(record);
359 * Release a memory record. Records marked for deletion are immediately
360 * removed from the RB-Tree but otherwise left intact until the last ref
364 hammer_rel_mem_record(struct hammer_record *record)
367 hammer_reserve_t resv;
369 hammer_inode_t target_ip;
372 hammer_rel(&record->lock);
374 if (hammer_norefs(&record->lock)) {
376 * Upon release of the last reference wakeup any waiters.
377 * The record structure may get destroyed so callers will
378 * loop up and do a relookup.
380 * WARNING! Record must be removed from RB-TREE before we
381 * might possibly block. hammer_test_inode() can block!
387 * Upon release of the last reference a record marked deleted
388 * by the front or backend, or committed by the backend,
391 if (record->flags & (HAMMER_RECF_DELETED_FE |
392 HAMMER_RECF_DELETED_BE |
393 HAMMER_RECF_COMMITTED)) {
394 KKASSERT(hammer_isactive(&ip->lock) > 0);
395 KKASSERT(record->flush_state != HAMMER_FST_FLUSH);
398 * target_ip may have zero refs, we have to ref it
399 * to prevent it from being ripped out from under
402 if ((target_ip = record->target_ip) != NULL) {
403 TAILQ_REMOVE(&target_ip->target_list,
404 record, target_entry);
405 record->target_ip = NULL;
406 hammer_ref(&target_ip->lock);
410 * Remove the record from the RB-Tree
412 if (record->flags & HAMMER_RECF_ONRBTREE) {
413 RB_REMOVE(hammer_rec_rb_tree,
416 record->flags &= ~HAMMER_RECF_ONRBTREE;
417 KKASSERT(ip->rsv_recs > 0);
418 if (RB_EMPTY(&ip->rec_tree)) {
419 ip->flags &= ~HAMMER_INODE_XDIRTY;
420 ip->sync_flags &= ~HAMMER_INODE_XDIRTY;
428 * We must wait for any direct-IO to complete before
429 * we can destroy the record because the bio may
430 * have a reference to it.
433 (HAMMER_RECG_DIRECT_IO | HAMMER_RECG_DIRECT_INVAL)) {
434 hammer_io_direct_wait(record);
438 * Account for the completion after the direct IO
444 hmp->rsv_databytes -= record->leaf.data_len;
446 if (RB_EMPTY(&ip->rec_tree))
447 hammer_test_inode(ip);
448 if ((ip->flags & HAMMER_INODE_RECSW) &&
449 ip->rsv_recs <= hammer_limit_inode_recs/2) {
450 ip->flags &= ~HAMMER_INODE_RECSW;
451 wakeup(&ip->rsv_recs);
456 * Do this test after removing record from the RB-Tree.
459 hammer_test_inode(target_ip);
460 hammer_rel_inode(target_ip, 0);
463 if (record->flags & HAMMER_RECF_ALLOCDATA) {
464 --hammer_count_record_datas;
465 kfree(record->data, hmp->m_misc);
466 record->flags &= ~HAMMER_RECF_ALLOCDATA;
470 * Release the reservation.
472 * If the record was not committed we can theoretically
473 * undo the reservation. However, doing so might
474 * create weird edge cases with the ordering of
475 * direct writes because the related buffer cache
476 * elements are per-vnode. So we don't try.
478 if ((resv = record->resv) != NULL) {
479 /* XXX undo leaf.data_offset,leaf.data_len */
480 hammer_blockmap_reserve_complete(hmp, resv);
484 --hammer_count_records;
485 kfree(record, hmp->m_misc);
491 * Record visibility depends on whether the record is being accessed by
492 * the backend or the frontend. Backend tests ignore the frontend delete
493 * flag. Frontend tests do NOT ignore the backend delete/commit flags and
494 * must also check for commit races.
496 * Return non-zero if the record is visible, zero if it isn't or if it is
497 * deleted. Returns 0 if the record has been comitted (unless the special
498 * delete-visibility flag is set). A committed record must be located
499 * via the media B-Tree. Returns non-zero if the record is good.
501 * If HAMMER_CURSOR_DELETE_VISIBILITY is set we allow deleted memory
502 * records to be returned. This is so pending deletions are detected
503 * when using an iterator to locate an unused hash key, or when we need
504 * to locate historical records on-disk to destroy.
508 hammer_ip_iterate_mem_good(hammer_cursor_t cursor, hammer_record_t record)
510 if (cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY)
512 if (cursor->flags & HAMMER_CURSOR_BACKEND) {
513 if (record->flags & (HAMMER_RECF_DELETED_BE |
514 HAMMER_RECF_COMMITTED)) {
518 if (record->flags & (HAMMER_RECF_DELETED_FE |
519 HAMMER_RECF_DELETED_BE |
520 HAMMER_RECF_COMMITTED)) {
528 * This callback is used as part of the RB_SCAN function for in-memory
529 * records. We terminate it (return -1) as soon as we get a match.
531 * This routine is used by frontend code.
533 * The primary compare code does not account for ASOF lookups. This
534 * code handles that case as well as a few others.
538 hammer_rec_scan_callback(hammer_record_t rec, void *data)
540 hammer_cursor_t cursor = data;
543 * We terminate on success, so this should be NULL on entry.
545 KKASSERT(cursor->iprec == NULL);
548 * Skip if the record was marked deleted or committed.
550 if (hammer_ip_iterate_mem_good(cursor, rec) == 0)
554 * Skip if not visible due to our as-of TID
556 if (cursor->flags & HAMMER_CURSOR_ASOF) {
557 if (cursor->asof < rec->leaf.base.create_tid)
559 if (rec->leaf.base.delete_tid &&
560 cursor->asof >= rec->leaf.base.delete_tid) {
566 * ref the record. The record is protected from backend B-Tree
567 * interactions by virtue of the cursor's IP lock.
569 hammer_ref(&rec->lock);
572 * The record may have been deleted or committed while we
573 * were blocked. XXX remove?
575 if (hammer_ip_iterate_mem_good(cursor, rec) == 0) {
576 hammer_rel_mem_record(rec);
581 * Set the matching record and stop the scan.
589 * Lookup an in-memory record given the key specified in the cursor. Works
590 * just like hammer_btree_lookup() but operates on an inode's in-memory
593 * The lookup must fail if the record is marked for deferred deletion.
595 * The API for mem/btree_lookup() does not mess with the ATE/EOF bits.
599 hammer_mem_lookup(hammer_cursor_t cursor)
601 KKASSERT(cursor->ip != NULL);
603 hammer_rel_mem_record(cursor->iprec);
604 cursor->iprec = NULL;
606 hammer_rec_rb_tree_RB_SCAN(&cursor->ip->rec_tree, hammer_rec_find_cmp,
607 hammer_rec_scan_callback, cursor);
609 return (cursor->iprec ? 0 : ENOENT);
613 * hammer_mem_first() - locate the first in-memory record matching the
614 * cursor within the bounds of the key range.
616 * WARNING! API is slightly different from btree_first(). hammer_mem_first()
617 * will set ATEMEM the same as MEMEOF, and does not return any error.
621 hammer_mem_first(hammer_cursor_t cursor)
623 KKASSERT(cursor->ip != NULL);
625 hammer_rel_mem_record(cursor->iprec);
626 cursor->iprec = NULL;
628 hammer_rec_rb_tree_RB_SCAN(&cursor->ip->rec_tree, hammer_rec_scan_cmp,
629 hammer_rec_scan_callback, cursor);
632 cursor->flags &= ~(HAMMER_CURSOR_MEMEOF | HAMMER_CURSOR_ATEMEM);
634 cursor->flags |= HAMMER_CURSOR_MEMEOF | HAMMER_CURSOR_ATEMEM;
636 return (cursor->iprec ? 0 : ENOENT);
639 /************************************************************************
640 * HAMMER IN-MEMORY RECORD FUNCTIONS *
641 ************************************************************************
643 * These functions manipulate in-memory records. Such records typically
644 * exist prior to being committed to disk or indexed via the on-disk B-Tree.
648 * Add a directory entry (dip,ncp) which references inode (ip).
650 * Note that the low 32 bits of the namekey are set temporarily to create
651 * a unique in-memory record, and may be modified a second time when the
652 * record is synchronized to disk. In particular, the low 32 bits cannot be
653 * all 0's when synching to disk, which is not handled here.
655 * NOTE: bytes does not include any terminating \0 on name, and name might
659 hammer_ip_add_direntry(struct hammer_transaction *trans,
660 struct hammer_inode *dip, const char *name, int bytes,
661 struct hammer_inode *ip)
663 struct hammer_cursor cursor;
664 hammer_record_t record;
666 uint32_t max_iterations;
668 KKASSERT(dip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY);
670 record = hammer_alloc_mem_record(dip, HAMMER_ENTRY_SIZE(bytes));
672 record->type = HAMMER_MEM_RECORD_ADD;
673 record->leaf.base.localization = dip->obj_localization |
674 hammer_dir_localization(dip);
675 record->leaf.base.obj_id = dip->obj_id;
676 record->leaf.base.key = hammer_direntry_namekey(dip, name, bytes,
678 record->leaf.base.rec_type = HAMMER_RECTYPE_DIRENTRY;
679 record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
680 record->data->entry.obj_id = ip->obj_id;
681 record->data->entry.localization = ip->obj_localization;
682 bcopy(name, record->data->entry.name, bytes);
684 ++ip->ino_data.nlinks;
685 ip->ino_data.ctime = trans->time;
686 hammer_modify_inode(trans, ip, HAMMER_INODE_DDIRTY);
689 * Find an unused namekey. Both the in-memory record tree and
690 * the B-Tree are checked. We do not want historically deleted
691 * names to create a collision as our iteration space may be limited,
692 * and since create_tid wouldn't match anyway an ASOF search
693 * must be used to locate collisions.
695 * delete-visibility is set so pending deletions do not give us
696 * a false-negative on our ability to use an iterator.
698 * The iterator must not rollover the key. Directory keys only
699 * use the positive key space.
701 hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
702 cursor.key_beg = record->leaf.base;
703 cursor.flags |= HAMMER_CURSOR_ASOF;
704 cursor.flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
705 cursor.asof = ip->obj_asof;
707 while (hammer_ip_lookup(&cursor) == 0) {
708 ++record->leaf.base.key;
709 KKASSERT(record->leaf.base.key > 0);
710 cursor.key_beg.key = record->leaf.base.key;
711 if (--max_iterations == 0) {
712 hammer_rel_mem_record(record);
719 * The target inode and the directory entry are bound together.
721 record->target_ip = ip;
722 record->flush_state = HAMMER_FST_SETUP;
723 TAILQ_INSERT_TAIL(&ip->target_list, record, target_entry);
726 * The inode now has a dependancy and must be taken out of the idle
727 * state. An inode not in an idle state is given an extra reference.
729 * When transitioning to a SETUP state flag for an automatic reflush
730 * when the dependancies are disposed of if someone is waiting on
733 if (ip->flush_state == HAMMER_FST_IDLE) {
734 hammer_ref(&ip->lock);
735 ip->flush_state = HAMMER_FST_SETUP;
736 if (ip->flags & HAMMER_INODE_FLUSHW)
737 ip->flags |= HAMMER_INODE_REFLUSH;
739 error = hammer_mem_add(record);
741 dip->ino_data.mtime = trans->time;
742 hammer_modify_inode(trans, dip, HAMMER_INODE_MTIME);
745 hammer_done_cursor(&cursor);
750 * Delete the directory entry and update the inode link count. The
751 * cursor must be seeked to the directory entry record being deleted.
753 * The related inode should be share-locked by the caller. The caller is
754 * on the frontend. It could also be NULL indicating that the directory
755 * entry being removed has no related inode.
757 * This function can return EDEADLK requiring the caller to terminate
758 * the cursor, any locks, wait on the returned record, and retry.
761 hammer_ip_del_direntry(struct hammer_transaction *trans,
762 hammer_cursor_t cursor, struct hammer_inode *dip,
763 struct hammer_inode *ip)
765 hammer_record_t record;
768 if (hammer_cursor_inmem(cursor)) {
770 * In-memory (unsynchronized) records can simply be freed.
772 * Even though the HAMMER_RECF_DELETED_FE flag is ignored
773 * by the backend, we must still avoid races against the
774 * backend potentially syncing the record to the media.
776 * We cannot call hammer_ip_delete_record(), that routine may
777 * only be called from the backend.
779 record = cursor->iprec;
780 if (record->flags & (HAMMER_RECF_INTERLOCK_BE |
781 HAMMER_RECF_DELETED_BE |
782 HAMMER_RECF_COMMITTED)) {
783 KKASSERT(cursor->deadlk_rec == NULL);
784 hammer_ref(&record->lock);
785 cursor->deadlk_rec = record;
788 KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
789 record->flags |= HAMMER_RECF_DELETED_FE;
794 * If the record is on-disk we have to queue the deletion by
795 * the record's key. This also causes lookups to skip the
796 * record (lookups for the purposes of finding an unused
797 * directory key do not skip the record).
799 KKASSERT(dip->flags &
800 (HAMMER_INODE_ONDISK | HAMMER_INODE_DONDISK));
801 record = hammer_alloc_mem_record(dip, 0);
802 record->type = HAMMER_MEM_RECORD_DEL;
803 record->leaf.base = cursor->leaf->base;
804 KKASSERT(dip->obj_id == record->leaf.base.obj_id);
807 * ip may be NULL, indicating the deletion of a directory
808 * entry which has no related inode.
810 record->target_ip = ip;
812 record->flush_state = HAMMER_FST_SETUP;
813 TAILQ_INSERT_TAIL(&ip->target_list, record,
816 record->flush_state = HAMMER_FST_IDLE;
820 * The inode now has a dependancy and must be taken out of
821 * the idle state. An inode not in an idle state is given
822 * an extra reference.
824 * When transitioning to a SETUP state flag for an automatic
825 * reflush when the dependancies are disposed of if someone
826 * is waiting on the inode.
828 if (ip && ip->flush_state == HAMMER_FST_IDLE) {
829 hammer_ref(&ip->lock);
830 ip->flush_state = HAMMER_FST_SETUP;
831 if (ip->flags & HAMMER_INODE_FLUSHW)
832 ip->flags |= HAMMER_INODE_REFLUSH;
835 error = hammer_mem_add(record);
839 * One less link. The file may still be open in the OS even after
840 * all links have gone away.
842 * We have to terminate the cursor before syncing the inode to
843 * avoid deadlocking against ourselves. XXX this may no longer
846 * If nlinks drops to zero and the vnode is inactive (or there is
847 * no vnode), call hammer_inode_unloadable_check() to zonk the
848 * inode. If we don't do this here the inode will not be destroyed
849 * on-media until we unmount.
853 --ip->ino_data.nlinks; /* do before we might block */
854 ip->ino_data.ctime = trans->time;
856 dip->ino_data.mtime = trans->time;
857 hammer_modify_inode(trans, dip, HAMMER_INODE_MTIME);
859 hammer_modify_inode(trans, ip, HAMMER_INODE_DDIRTY);
860 if (ip->ino_data.nlinks == 0 &&
861 (ip->vp == NULL || (ip->vp->v_flag & VINACTIVE))) {
862 hammer_done_cursor(cursor);
863 hammer_inode_unloadable_check(ip, 1);
864 hammer_flush_inode(ip, 0);
873 * Add a record to an inode.
875 * The caller must allocate the record with hammer_alloc_mem_record(ip,len) and
876 * initialize the following additional fields that are not initialized by these
879 * The related inode should be share-locked by the caller. The caller is
882 * record->leaf.base.key
883 * record->leaf.base.rec_type
884 * record->leaf.base.localization
887 hammer_ip_add_record(struct hammer_transaction *trans, hammer_record_t record)
889 hammer_inode_t ip = record->ip;
892 KKASSERT(record->leaf.base.localization != 0);
893 record->leaf.base.obj_id = ip->obj_id;
894 record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
895 error = hammer_mem_add(record);
900 * Locate a pre-existing bulk record in memory. The caller wishes to
901 * replace the record with a new one. The existing record may have a
902 * different length (and thus a different key) so we have to use an
903 * overlap check function.
905 static hammer_record_t
906 hammer_ip_get_bulk(hammer_record_t record)
908 struct hammer_bulk_info info;
909 hammer_inode_t ip = record->ip;
911 info.record = record;
912 info.conflict = NULL;
913 hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_overlap_cmp,
914 hammer_bulk_scan_callback, &info);
916 return(info.conflict); /* may be NULL */
920 * Take records vetted by overlap_cmp. The first non-deleted record
921 * (if any) stops the scan.
924 hammer_bulk_scan_callback(hammer_record_t record, void *data)
926 struct hammer_bulk_info *info = data;
928 if (record->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
929 HAMMER_RECF_COMMITTED)) {
932 hammer_ref(&record->lock);
933 info->conflict = record;
934 return(-1); /* stop scan */
938 * Reserve blockmap space placemarked with an in-memory record.
940 * This routine is called by the frontend in order to be able to directly
941 * flush a buffer cache buffer. The frontend has locked the related buffer
942 * cache buffers and we should be able to manipulate any overlapping
945 * The caller is responsible for adding the returned record and deleting
946 * the returned conflicting record (if any), typically by calling
947 * hammer_ip_replace_bulk() (via hammer_io_direct_write()).
950 hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset, void *data, int bytes,
953 hammer_record_t record;
954 hammer_dedup_cache_t dcp;
959 * Create a record to cover the direct write. The record cannot
960 * be added to the in-memory RB tree here as it might conflict
961 * with an existing memory record. See hammer_io_direct_write().
963 * The backend is responsible for finalizing the space reserved in
966 * XXX bytes not aligned, depend on the reservation code to
967 * align the reservation.
969 record = hammer_alloc_mem_record(ip, 0);
970 zone = hammer_data_zone_index(bytes);
974 crc = crc32(data, bytes);
976 if (hammer_live_dedup == 0)
978 if ((dcp = hammer_dedup_cache_lookup(ip->hmp, crc)) != NULL) {
979 struct hammer_dedup_cache tmp = *dcp;
981 record->resv = hammer_blockmap_reserve_dedup(ip->hmp, zone,
982 bytes, tmp.data_offset, errorp);
983 if (record->resv == NULL)
986 if (!hammer_dedup_validate(&tmp, zone, bytes, data)) {
987 hammer_blockmap_reserve_complete(ip->hmp, record->resv);
991 record->leaf.data_offset = tmp.data_offset;
992 record->flags |= HAMMER_RECF_DEDUPED;
995 record->resv = hammer_blockmap_reserve(ip->hmp, zone, bytes,
996 &record->leaf.data_offset, errorp);
997 if (record->resv == NULL) {
998 hdkprintf("reservation failed\n");
999 hammer_rel_mem_record(record);
1004 record->type = HAMMER_MEM_RECORD_DATA;
1005 record->leaf.base.rec_type = HAMMER_RECTYPE_DATA;
1006 record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
1007 record->leaf.base.obj_id = ip->obj_id;
1008 record->leaf.base.key = file_offset + bytes;
1009 record->leaf.base.localization = ip->obj_localization |
1010 HAMMER_LOCALIZE_MISC;
1011 record->leaf.data_len = bytes;
1012 record->leaf.data_crc = crc;
1013 KKASSERT(*errorp == 0);
1019 * Called by hammer_io_direct_write() prior to any possible completion
1020 * of the BIO to emplace the memory record associated with the I/O and
1021 * to replace any prior memory record which might still be active.
1023 * Setting the FE deleted flag on the old record (if any) avoids any RB
1024 * tree insertion conflict, amoung other things.
1026 * This has to be done prior to the caller completing any related buffer
1027 * cache I/O or a reinstantiation of the buffer may load data from the
1028 * old media location instead of the new media location. The holding
1029 * of the locked buffer cache buffer serves to interlock the record
1030 * replacement operation.
1033 hammer_ip_replace_bulk(hammer_mount_t hmp, hammer_record_t record)
1035 hammer_record_t conflict;
1036 int error __debugvar;
1038 while ((conflict = hammer_ip_get_bulk(record)) != NULL) {
1039 if ((conflict->flags & HAMMER_RECF_INTERLOCK_BE) == 0) {
1040 conflict->flags |= HAMMER_RECF_DELETED_FE;
1043 conflict->flags |= HAMMER_RECF_WANTED;
1044 tsleep(conflict, 0, "hmrrc3", 0);
1045 hammer_rel_mem_record(conflict);
1047 error = hammer_mem_add(record);
1049 hammer_rel_mem_record(conflict);
1050 KKASSERT(error == 0);
1054 * Frontend truncation code. Scan in-memory records only. On-disk records
1055 * and records in a flushing state are handled by the backend. The vnops
1056 * setattr code will handle the block containing the truncation point.
1058 * Partial blocks are not deleted.
1060 * This code is only called on regular files.
1063 hammer_ip_frontend_trunc(struct hammer_inode *ip, off_t file_size)
1065 struct rec_trunc_info info;
1067 switch(ip->ino_data.obj_type) {
1068 case HAMMER_OBJTYPE_REGFILE:
1069 info.rec_type = HAMMER_RECTYPE_DATA;
1071 case HAMMER_OBJTYPE_DBFILE:
1072 info.rec_type = HAMMER_RECTYPE_DB;
1077 info.trunc_off = file_size;
1078 hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_trunc_cmp,
1079 hammer_frontend_trunc_callback, &info);
1084 * Scan callback for frontend records to destroy during a truncation.
1085 * We must ensure that DELETED_FE is set on the record or the frontend
1086 * will get confused in future read() calls.
1088 * NOTE: DELETED_FE cannot be set while the record interlock (BE) is held.
1089 * In this rare case we must wait for the interlock to be cleared.
1091 * NOTE: This function is only called on regular files. There are further
1092 * restrictions to the setting of DELETED_FE on directory records
1093 * undergoing a flush due to sensitive inode link count calculations.
1096 hammer_frontend_trunc_callback(hammer_record_t record, void *data __unused)
1098 if (record->flags & HAMMER_RECF_DELETED_FE)
1101 if (record->flush_state == HAMMER_FST_FLUSH)
1104 hammer_ref(&record->lock);
1105 while (record->flags & HAMMER_RECF_INTERLOCK_BE)
1106 hammer_wait_mem_record_ident(record, "hmmtrr");
1107 record->flags |= HAMMER_RECF_DELETED_FE;
1108 hammer_rel_mem_record(record);
1113 * Return 1 if the caller must check for and delete existing records
1114 * before writing out a new data record.
1116 * Return 0 if the caller can just insert the record into the B-Tree without
1120 hammer_record_needs_overwrite_delete(hammer_record_t record)
1122 hammer_inode_t ip = record->ip;
1123 int64_t file_offset;
1126 if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE)
1127 file_offset = record->leaf.base.key;
1129 file_offset = record->leaf.base.key - record->leaf.data_len;
1130 r = (file_offset < ip->save_trunc_off);
1131 if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1132 if (ip->save_trunc_off <= record->leaf.base.key)
1133 ip->save_trunc_off = record->leaf.base.key + 1;
1135 if (ip->save_trunc_off < record->leaf.base.key)
1136 ip->save_trunc_off = record->leaf.base.key;
1142 * Backend code. Sync a record to the media.
1145 hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t record)
1147 hammer_transaction_t trans = cursor->trans;
1148 int64_t file_offset;
1154 KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
1155 KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE);
1156 KKASSERT(record->leaf.base.localization != 0);
1159 * Any direct-write related to the record must complete before we
1160 * can sync the record to the on-disk media.
1162 if (record->gflags & (HAMMER_RECG_DIRECT_IO | HAMMER_RECG_DIRECT_INVAL))
1163 hammer_io_direct_wait(record);
1166 * If this is a bulk-data record placemarker there may be an existing
1167 * record on-disk, indicating a data overwrite. If there is the
1168 * on-disk record must be deleted before we can insert our new record.
1170 * We've synthesized this record and do not know what the create_tid
1171 * on-disk is, nor how much data it represents.
1173 * Keep in mind that (key) for data records is (base_offset + len),
1174 * not (base_offset). Also, we only want to get rid of on-disk
1175 * records since we are trying to sync our in-memory record, call
1176 * hammer_ip_delete_range() with truncating set to 1 to make sure
1177 * it skips in-memory records.
1179 * It is ok for the lookup to return ENOENT.
1181 * NOTE OPTIMIZATION: sync_trunc_off is used to determine if we have
1182 * to call hammer_ip_delete_range() or not. This also means we must
1183 * update sync_trunc_off() as we write.
1185 if (record->type == HAMMER_MEM_RECORD_DATA &&
1186 hammer_record_needs_overwrite_delete(record)) {
1187 file_offset = record->leaf.base.key - record->leaf.data_len;
1188 bytes = (record->leaf.data_len + HAMMER_BUFMASK) &
1190 KKASSERT((file_offset & HAMMER_BUFMASK) == 0);
1191 error = hammer_ip_delete_range(
1193 file_offset, file_offset + bytes - 1,
1195 if (error && error != ENOENT)
1200 * If this is a general record there may be an on-disk version
1201 * that must be deleted before we can insert the new record.
1203 if (record->type == HAMMER_MEM_RECORD_GENERAL) {
1204 error = hammer_delete_general(cursor, record->ip, &record->leaf);
1205 if (error && error != ENOENT)
1212 hammer_normalize_cursor(cursor);
1213 cursor->key_beg = record->leaf.base;
1214 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1215 cursor->flags |= HAMMER_CURSOR_BACKEND;
1216 cursor->flags &= ~HAMMER_CURSOR_INSERT;
1219 * Records can wind up on-media before the inode itself is on-media.
1222 record->ip->flags |= HAMMER_INODE_DONDISK;
1225 * If we are deleting a directory entry an exact match must be
1228 if (record->type == HAMMER_MEM_RECORD_DEL) {
1229 error = hammer_btree_lookup(cursor);
1231 KKASSERT(cursor->iprec == NULL);
1232 error = hammer_ip_delete_record(cursor, record->ip,
1235 record->flags |= HAMMER_RECF_DELETED_BE |
1236 HAMMER_RECF_COMMITTED;
1237 ++record->ip->rec_generation;
1246 * Issue a lookup to position the cursor and locate the insertion
1247 * point. The target key should not exist. If we are creating a
1248 * directory entry we may have to iterate the low 32 bits of the
1249 * key to find an unused key.
1251 hammer_sync_lock_sh(trans);
1252 cursor->flags |= HAMMER_CURSOR_INSERT;
1253 error = hammer_btree_lookup(cursor);
1254 if (hammer_debug_inode)
1255 hdkprintf("DOINSERT LOOKUP %d\n", error);
1257 hdkprintf("duplicate rec at (%016llx)\n",
1258 (long long)record->leaf.base.key);
1259 if (hammer_debug_critical)
1260 Debugger("duplicate record1");
1264 if (error != ENOENT)
1268 * Allocate the record and data. The result buffers will be
1269 * marked as being modified and further calls to
1270 * hammer_modify_buffer() will result in unneeded UNDO records.
1272 * Support zero-fill records (data == NULL and data_len != 0)
1274 if (record->type == HAMMER_MEM_RECORD_DATA) {
1276 * The data portion of a bulk-data record has already been
1277 * committed to disk, we need only adjust the layer2
1278 * statistics in the same transaction as our B-Tree insert.
1280 KKASSERT(record->leaf.data_offset != 0);
1281 error = hammer_blockmap_finalize(trans,
1283 record->leaf.data_offset,
1284 record->leaf.data_len);
1286 if (hammer_live_dedup == 2 &&
1287 (record->flags & HAMMER_RECF_DEDUPED) == 0) {
1288 hammer_dedup_cache_add(record->ip, &record->leaf);
1290 } else if (record->data && record->leaf.data_len) {
1292 * Wholely cached record, with data. Allocate the data.
1294 bdata = hammer_alloc_data(trans, record->leaf.data_len,
1295 record->leaf.base.rec_type,
1296 &record->leaf.data_offset,
1297 &cursor->data_buffer,
1301 hammer_crc_set_leaf(record->data, &record->leaf);
1302 hammer_modify_buffer_noundo(trans, cursor->data_buffer);
1303 bcopy(record->data, bdata, record->leaf.data_len);
1304 hammer_modify_buffer_done(cursor->data_buffer);
1307 * Wholely cached record, without data.
1309 record->leaf.data_offset = 0;
1310 record->leaf.data_crc = 0;
1313 error = hammer_btree_insert(cursor, &record->leaf, &doprop);
1314 if (hammer_debug_inode && error) {
1315 hdkprintf("BTREE INSERT error %d @ %016llx:%d key %016llx\n",
1317 (long long)cursor->node->node_offset,
1319 (long long)record->leaf.base.key);
1323 * Our record is on-disk and we normally mark the in-memory version
1324 * as having been committed (and not BE-deleted).
1326 * If the record represented a directory deletion but we had to
1327 * sync a valid directory entry to disk due to dependancies,
1328 * we must convert the record to a covering delete so the
1329 * frontend does not have visibility on the synced entry.
1331 * WARNING: cursor's leaf pointer may have changed after do_propagation
1336 hammer_btree_do_propagation(cursor,
1340 if (record->flags & HAMMER_RECF_CONVERT_DELETE) {
1342 * Must convert deleted directory entry add
1343 * to a directory entry delete.
1345 KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
1346 record->flags &= ~HAMMER_RECF_DELETED_FE;
1347 record->type = HAMMER_MEM_RECORD_DEL;
1348 KKASSERT(record->ip->obj_id == record->leaf.base.obj_id);
1349 KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
1350 record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
1351 KKASSERT((record->flags & (HAMMER_RECF_COMMITTED |
1352 HAMMER_RECF_DELETED_BE)) == 0);
1353 /* converted record is not yet committed */
1354 /* hammer_flush_record_done takes care of the rest */
1357 * Everything went fine and we are now done with
1360 record->flags |= HAMMER_RECF_COMMITTED;
1361 ++record->ip->rec_generation;
1364 if (record->leaf.data_offset) {
1365 hammer_blockmap_free(trans, record->leaf.data_offset,
1366 record->leaf.data_len);
1370 hammer_sync_unlock(trans);
1376 * Add the record to the inode's rec_tree. The low 32 bits of a directory
1377 * entry's key is used to deal with hash collisions in the upper 32 bits.
1378 * A unique 64 bit key is generated in-memory and may be regenerated a
1379 * second time when the directory record is flushed to the on-disk B-Tree.
1381 * A referenced record is passed to this function. This function
1382 * eats the reference. If an error occurs the record will be deleted.
1384 * A copy of the temporary record->data pointer provided by the caller
1388 hammer_mem_add(hammer_record_t record)
1390 hammer_mount_t hmp = record->ip->hmp;
1393 * Make a private copy of record->data
1396 KKASSERT(record->flags & HAMMER_RECF_ALLOCDATA);
1399 * Insert into the RB tree. A unique key should have already
1400 * been selected if this is a directory entry.
1402 if (RB_INSERT(hammer_rec_rb_tree, &record->ip->rec_tree, record)) {
1403 record->flags |= HAMMER_RECF_DELETED_FE;
1404 hammer_rel_mem_record(record);
1408 ++record->ip->rsv_recs;
1409 record->ip->hmp->rsv_databytes += record->leaf.data_len;
1410 record->flags |= HAMMER_RECF_ONRBTREE;
1411 hammer_modify_inode(NULL, record->ip, HAMMER_INODE_XDIRTY);
1412 hammer_rel_mem_record(record);
1416 /************************************************************************
1417 * HAMMER INODE MERGED-RECORD FUNCTIONS *
1418 ************************************************************************
1420 * These functions augment the B-Tree scanning functions in hammer_btree.c
1421 * by merging in-memory records with on-disk records.
1425 * Locate a particular record either in-memory or on-disk.
1427 * NOTE: This is basically a standalone routine, hammer_ip_next() may
1428 * NOT be called to iterate results.
1431 hammer_ip_lookup(hammer_cursor_t cursor)
1436 * If the element is in-memory return it without searching the
1439 KKASSERT(cursor->ip);
1440 error = hammer_mem_lookup(cursor);
1442 cursor->leaf = &cursor->iprec->leaf;
1445 if (error != ENOENT)
1449 * If the inode has on-disk components search the on-disk B-Tree.
1451 if ((cursor->ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) == 0)
1453 error = hammer_btree_lookup(cursor);
1455 error = hammer_btree_extract_leaf(cursor);
1460 * Helper for hammer_ip_first()/hammer_ip_next()
1462 * NOTE: Both ATEDISK and DISKEOF will be set the same. This sets up
1463 * hammer_ip_first() for calling hammer_ip_next(), and sets up the re-seek
1464 * state if hammer_ip_next() needs to re-seek.
1468 _hammer_ip_seek_btree(hammer_cursor_t cursor)
1470 hammer_inode_t ip = cursor->ip;
1473 if (ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) {
1474 error = hammer_btree_lookup(cursor);
1475 if (error == ENOENT || error == EDEADLK) {
1476 if (hammer_debug_general & 0x2000) {
1477 hdkprintf("error %d node %p %016llx index %d\n",
1478 error, cursor->node,
1479 (long long)cursor->node->node_offset,
1482 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1483 error = hammer_btree_iterate(cursor);
1486 cursor->flags &= ~(HAMMER_CURSOR_DISKEOF |
1487 HAMMER_CURSOR_ATEDISK);
1489 cursor->flags |= HAMMER_CURSOR_DISKEOF |
1490 HAMMER_CURSOR_ATEDISK;
1491 if (error == ENOENT)
1495 cursor->flags |= HAMMER_CURSOR_DISKEOF | HAMMER_CURSOR_ATEDISK;
1502 * Helper for hammer_ip_next()
1504 * The caller has determined that the media cursor is further along than the
1505 * memory cursor and must be reseeked after a generation number change.
1509 _hammer_ip_reseek(hammer_cursor_t cursor)
1511 struct hammer_base_elm save;
1512 hammer_btree_elm_t elm;
1513 int error __debugvar;
1520 hkprintf("Debug: re-seeked during scan @ino=%016llx\n",
1521 (long long)cursor->ip->obj_id);
1522 save = cursor->key_beg;
1523 cursor->key_beg = cursor->iprec->leaf.base;
1524 error = _hammer_ip_seek_btree(cursor);
1525 KKASSERT(error == 0);
1526 cursor->key_beg = save;
1529 * If the memory record was previous returned to
1530 * the caller and the media record matches
1531 * (-1/+1: only create_tid differs), then iterate
1532 * the media record to avoid a double result.
1534 if ((cursor->flags & HAMMER_CURSOR_ATEDISK) == 0 &&
1535 (cursor->flags & HAMMER_CURSOR_LASTWASMEM)) {
1536 elm = &cursor->node->ondisk->elms[cursor->index];
1537 r = hammer_btree_cmp(&elm->base, &cursor->iprec->leaf.base);
1538 if (cursor->flags & HAMMER_CURSOR_ASOF) {
1539 if (r >= -1 && r <= 1) {
1540 hkprintf("Debug: iterated after "
1541 "re-seek (asof r=%d)\n", r);
1542 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1547 hkprintf("Debug: iterated after "
1549 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1558 * Locate the first record within the cursor's key_beg/key_end range,
1559 * restricted to a particular inode. 0 is returned on success, ENOENT
1560 * if no records matched the requested range, or some other error.
1562 * When 0 is returned hammer_ip_next() may be used to iterate additional
1563 * records within the requested range.
1565 * This function can return EDEADLK, requiring the caller to terminate
1566 * the cursor and try again.
1570 hammer_ip_first(hammer_cursor_t cursor)
1572 hammer_inode_t ip __debugvar = cursor->ip;
1575 KKASSERT(ip != NULL);
1578 * Clean up fields and setup for merged scan
1580 cursor->flags &= ~HAMMER_CURSOR_RETEST;
1583 * Search the in-memory record list (Red-Black tree). Unlike the
1584 * B-Tree search, mem_first checks for records in the range.
1586 * This function will setup both ATEMEM and MEMEOF properly for
1587 * the ip iteration. ATEMEM will be set if MEMEOF is set.
1589 hammer_mem_first(cursor);
1592 * Detect generation changes during blockages, including
1593 * blockages which occur on the initial btree search.
1595 cursor->rec_generation = cursor->ip->rec_generation;
1598 * Initial search and result
1600 error = _hammer_ip_seek_btree(cursor);
1602 error = hammer_ip_next(cursor);
1608 * Retrieve the next record in a merged iteration within the bounds of the
1609 * cursor. This call may be made multiple times after the cursor has been
1610 * initially searched with hammer_ip_first().
1612 * There are numerous special cases in this code to deal with races between
1613 * in-memory records and on-media records.
1615 * 0 is returned on success, ENOENT if no further records match the
1616 * requested range, or some other error code is returned.
1619 hammer_ip_next(hammer_cursor_t cursor)
1621 hammer_btree_elm_t elm;
1622 hammer_record_t rec;
1623 hammer_record_t tmprec;
1629 * Get the next on-disk record
1631 * NOTE: If we deleted the last on-disk record we had scanned
1632 * ATEDISK will be clear and RETEST will be set, forcing
1633 * a call to iterate. The fact that ATEDISK is clear causes
1634 * iterate to re-test the 'current' element. If ATEDISK is
1635 * set, iterate will skip the 'current' element.
1638 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
1639 if (cursor->flags & (HAMMER_CURSOR_ATEDISK |
1640 HAMMER_CURSOR_RETEST)) {
1641 error = hammer_btree_iterate(cursor);
1642 cursor->flags &= ~HAMMER_CURSOR_RETEST;
1644 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1645 hammer_cache_node(&cursor->ip->cache[1],
1647 } else if (error == ENOENT) {
1648 cursor->flags |= HAMMER_CURSOR_DISKEOF |
1649 HAMMER_CURSOR_ATEDISK;
1656 * If the generation changed the backend has deleted or committed
1657 * one or more memory records since our last check.
1659 * When this case occurs if the disk cursor is > current memory record
1660 * or the disk cursor is at EOF, we must re-seek the disk-cursor.
1661 * Since the cursor is ahead it must have not yet been eaten (if
1662 * not at eof anyway). (XXX data offset case?)
1664 * NOTE: we are not doing a full check here. That will be handled
1667 * If we have exhausted all memory records we do not have to do any
1670 while (cursor->rec_generation != cursor->ip->rec_generation &&
1672 hkprintf("Debug: generation changed during scan @ino=%016llx\n",
1673 (long long)cursor->ip->obj_id);
1674 cursor->rec_generation = cursor->ip->rec_generation;
1675 if (cursor->flags & HAMMER_CURSOR_MEMEOF)
1677 if (cursor->flags & HAMMER_CURSOR_DISKEOF) {
1680 KKASSERT((cursor->flags & HAMMER_CURSOR_ATEDISK) == 0);
1681 elm = &cursor->node->ondisk->elms[cursor->index];
1682 r = hammer_btree_cmp(&elm->base,
1683 &cursor->iprec->leaf.base);
1687 * Do we re-seek the media cursor?
1690 if (_hammer_ip_reseek(cursor))
1696 * We can now safely get the next in-memory record. We cannot
1699 * hammer_rec_scan_cmp: Is the record still in our general range,
1700 * (non-inclusive of snapshot exclusions)?
1701 * hammer_rec_scan_callback: Is the record in our snapshot?
1704 if ((cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) {
1706 * If the current memory record was eaten then get the next
1707 * one. Stale records are skipped.
1709 if (cursor->flags & HAMMER_CURSOR_ATEMEM) {
1710 tmprec = cursor->iprec;
1711 cursor->iprec = NULL;
1712 rec = hammer_rec_rb_tree_RB_NEXT(tmprec);
1714 if (hammer_rec_scan_cmp(rec, cursor) != 0)
1716 if (hammer_rec_scan_callback(rec, cursor) != 0)
1718 rec = hammer_rec_rb_tree_RB_NEXT(rec);
1720 if (cursor->iprec) {
1721 KKASSERT(cursor->iprec == rec);
1722 cursor->flags &= ~HAMMER_CURSOR_ATEMEM;
1724 cursor->flags |= HAMMER_CURSOR_MEMEOF;
1726 cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1731 * MEMORY RECORD VALIDITY TEST
1733 * (We still can't block, which is why tmprec is being held so
1736 * If the memory record is no longer valid we skip it. It may
1737 * have been deleted by the frontend. If it was deleted or
1738 * committed by the backend the generation change re-seeked the
1739 * disk cursor and the record will be present there.
1741 if (error == 0 && (cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) {
1742 KKASSERT(cursor->iprec);
1743 KKASSERT((cursor->flags & HAMMER_CURSOR_ATEMEM) == 0);
1744 if (!hammer_ip_iterate_mem_good(cursor, cursor->iprec)) {
1745 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1747 hammer_rel_mem_record(tmprec);
1752 hammer_rel_mem_record(tmprec);
1755 * Extract either the disk or memory record depending on their
1756 * relative position.
1759 switch(cursor->flags & (HAMMER_CURSOR_ATEDISK | HAMMER_CURSOR_ATEMEM)) {
1762 * Both entries valid. Compare the entries and nominally
1763 * return the first one in the sort order. Numerous cases
1764 * require special attention, however.
1766 elm = &cursor->node->ondisk->elms[cursor->index];
1767 r = hammer_btree_cmp(&elm->base, &cursor->iprec->leaf.base);
1770 * If the two entries differ only by their key (-2/2) or
1771 * create_tid (-1/1), and are DATA records, we may have a
1772 * nominal match. We have to calculate the base file
1773 * offset of the data.
1775 if (r <= 2 && r >= -2 && r != 0 &&
1776 cursor->ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE &&
1777 cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1778 int64_t base1 = elm->leaf.base.key - elm->leaf.data_len;
1779 int64_t base2 = cursor->iprec->leaf.base.key -
1780 cursor->iprec->leaf.data_len;
1786 error = hammer_btree_extract_leaf(cursor);
1787 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1788 cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1793 * If the entries match exactly the memory entry is either
1794 * an on-disk directory entry deletion or a bulk data
1795 * overwrite. If it is a directory entry deletion we eat
1798 * For the bulk-data overwrite case it is possible to have
1799 * visibility into both, which simply means the syncer
1800 * hasn't gotten around to doing the delete+insert sequence
1801 * on the B-Tree. Use the memory entry and throw away the
1804 * If the in-memory record is not either of these we
1805 * probably caught the syncer while it was syncing it to
1806 * the media. Since we hold a shared lock on the cursor,
1807 * the in-memory record had better be marked deleted at
1811 if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL) {
1812 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1813 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1814 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1817 } else if (cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1818 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1819 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1821 /* fall through to memory entry */
1823 hpanic("duplicate mem/B-Tree entry %p %d %08x",
1825 cursor->iprec->type,
1826 cursor->iprec->flags);
1827 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1831 /* fall through to the memory entry */
1832 case HAMMER_CURSOR_ATEDISK:
1834 * Only the memory entry is valid.
1836 cursor->leaf = &cursor->iprec->leaf;
1837 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1838 cursor->flags |= HAMMER_CURSOR_LASTWASMEM;
1841 * If the memory entry is an on-disk deletion we should have
1842 * also had found a B-Tree record. If the backend beat us
1843 * to it it would have interlocked the cursor and we should
1844 * have seen the in-memory record marked DELETED_FE.
1846 if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL &&
1847 (cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1848 hpanic("del-on-disk with no B-Tree entry iprec %p flags %08x",
1850 cursor->iprec->flags);
1853 case HAMMER_CURSOR_ATEMEM:
1855 * Only the disk entry is valid
1857 error = hammer_btree_extract_leaf(cursor);
1858 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1859 cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1863 * Neither entry is valid
1865 * XXX error not set properly
1867 cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1868 cursor->leaf = NULL;
1876 * Resolve the cursor->data pointer for the current cursor position in
1877 * a merged iteration.
1880 hammer_ip_resolve_data(hammer_cursor_t cursor)
1882 hammer_record_t record;
1885 if (hammer_cursor_inmem(cursor)) {
1887 * The data associated with an in-memory record is usually
1888 * kmalloced, but reserve-ahead data records will have an
1889 * on-disk reference.
1891 * NOTE: Reserve-ahead data records must be handled in the
1892 * context of the related high level buffer cache buffer
1893 * to interlock against async writes.
1895 record = cursor->iprec;
1896 cursor->data = record->data;
1898 if (cursor->data == NULL) {
1899 KKASSERT(record->leaf.base.rec_type ==
1900 HAMMER_RECTYPE_DATA);
1901 cursor->data = hammer_bread_ext(cursor->trans->hmp,
1902 record->leaf.data_offset,
1903 record->leaf.data_len,
1905 &cursor->data_buffer);
1909 * Loading leaf here isn't necessary if it's guaranteed that
1910 * the cursor is at a leaf node (which basically should be)
1911 * because hammer_btree_extract_data() does that.
1913 cursor->leaf = &cursor->node->ondisk->elms[cursor->index].leaf;
1914 error = hammer_btree_extract_data(cursor);
1920 * Backend truncation / record replacement - delete records in range.
1922 * Delete all records within the specified range for inode ip. In-memory
1923 * records still associated with the frontend are ignored.
1925 * If truncating is non-zero in-memory records associated with the back-end
1926 * are ignored. If truncating is > 1 we can return EWOULDBLOCK.
1930 * * An unaligned range will cause new records to be added to cover
1931 * the edge cases. (XXX not implemented yet).
1933 * * Replacement via reservations (see hammer_ip_sync_record_cursor())
1934 * also do not deal with unaligned ranges.
1936 * * ran_end is inclusive (e.g. 0,1023 instead of 0,1024).
1938 * * Record keys for regular file data have to be special-cased since
1939 * they indicate the end of the range (key = base + bytes).
1941 * * This function may be asked to delete ridiculously huge ranges, for
1942 * example if someone truncates or removes a 1TB regular file. We
1943 * must be very careful on restarts and we may have to stop w/
1944 * EWOULDBLOCK to avoid blowing out the buffer cache.
1947 hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1948 int64_t ran_beg, int64_t ran_end, int truncating)
1950 hammer_transaction_t trans = cursor->trans;
1951 hammer_btree_leaf_elm_t leaf;
1956 KKASSERT(trans->type == HAMMER_TRANS_FLS);
1958 hammer_normalize_cursor(cursor);
1959 cursor->key_beg.localization = ip->obj_localization |
1960 HAMMER_LOCALIZE_MISC;
1961 cursor->key_beg.obj_id = ip->obj_id;
1962 cursor->key_beg.create_tid = 0;
1963 cursor->key_beg.delete_tid = 0;
1964 cursor->key_beg.obj_type = 0;
1966 if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1967 cursor->key_beg.key = ran_beg;
1968 cursor->key_beg.rec_type = HAMMER_RECTYPE_DB;
1971 * The key in the B-Tree is (base+bytes), so the first possible
1972 * matching key is ran_beg + 1.
1974 cursor->key_beg.key = ran_beg + 1;
1975 cursor->key_beg.rec_type = HAMMER_RECTYPE_DATA;
1978 cursor->key_end = cursor->key_beg;
1979 if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1980 cursor->key_end.key = ran_end;
1982 tmp64 = ran_end + MAXPHYS + 1; /* work around GCC-4 bug */
1983 if (tmp64 < ran_end)
1984 cursor->key_end.key = 0x7FFFFFFFFFFFFFFFLL;
1986 cursor->key_end.key = ran_end + MAXPHYS + 1;
1989 cursor->asof = ip->obj_asof;
1990 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1991 cursor->flags |= HAMMER_CURSOR_ASOF;
1992 cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
1993 cursor->flags |= HAMMER_CURSOR_BACKEND;
1994 cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE;
1996 error = hammer_ip_first(cursor);
1999 * Iterate through matching records and mark them as deleted.
2001 while (error == 0) {
2002 leaf = cursor->leaf;
2004 KKASSERT(leaf->base.delete_tid == 0);
2005 KKASSERT(leaf->base.obj_id == ip->obj_id);
2008 * There may be overlap cases for regular file data. Also
2009 * remember the key for a regular file record is (base + len),
2012 * Note that due to duplicates (mem & media) allowed by
2013 * DELETE_VISIBILITY, off can wind up less then ran_beg.
2015 if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) {
2016 off = leaf->base.key - leaf->data_len;
2018 * Check the left edge case. We currently do not
2019 * split existing records.
2021 if (off < ran_beg && leaf->base.key > ran_beg) {
2022 hpanic("hammer left edge case %016llx %d",
2023 (long long)leaf->base.key,
2028 * Check the right edge case. Note that the
2029 * record can be completely out of bounds, which
2030 * terminates the search.
2032 * base->key is exclusive of the right edge while
2033 * ran_end is inclusive of the right edge. The
2034 * (key - data_len) left boundary is inclusive.
2036 * XXX theory-check this test at some point, are
2037 * we missing a + 1 somewhere? Note that ran_end
2040 if (leaf->base.key - 1 > ran_end) {
2041 if (leaf->base.key - leaf->data_len > ran_end)
2043 hpanic("hammer right edge case");
2046 off = leaf->base.key;
2050 * Delete the record. When truncating we do not delete
2051 * in-memory (data) records because they represent data
2052 * written after the truncation.
2054 * This will also physically destroy the B-Tree entry and
2055 * data if the retention policy dictates. The function
2056 * will set HAMMER_CURSOR_RETEST to cause hammer_ip_next()
2057 * to retest the new 'current' element.
2059 if (truncating == 0 || hammer_cursor_ondisk(cursor)) {
2060 error = hammer_ip_delete_record(cursor, ip, trans->tid);
2062 * If we have built up too many meta-buffers we risk
2063 * deadlocking the kernel and must stop. This can
2064 * occur when deleting ridiculously huge files.
2065 * sync_trunc_off is updated so the next cycle does
2066 * not re-iterate records we have already deleted.
2068 * This is only done with formal truncations.
2070 if (truncating > 1 && error == 0 &&
2071 hammer_flusher_meta_limit(ip->hmp)) {
2072 ip->sync_trunc_off = off;
2073 error = EWOULDBLOCK;
2078 ran_beg = off; /* for restart */
2079 error = hammer_ip_next(cursor);
2082 hammer_cache_node(&ip->cache[1], cursor->node);
2084 if (error == EDEADLK) {
2085 hammer_done_cursor(cursor);
2086 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
2090 if (error == ENOENT)
2096 * This backend function deletes the specified record on-disk, similar to
2097 * delete_range but for a specific record. Unlike the exact deletions
2098 * used when deleting a directory entry this function uses an ASOF search
2099 * like delete_range.
2101 * This function may be called with ip->obj_asof set for a slave snapshot,
2102 * so don't use it. We always delete non-historical records only.
2105 hammer_delete_general(hammer_cursor_t cursor, hammer_inode_t ip,
2106 hammer_btree_leaf_elm_t leaf)
2108 hammer_transaction_t trans = cursor->trans;
2111 KKASSERT(trans->type == HAMMER_TRANS_FLS);
2113 hammer_normalize_cursor(cursor);
2114 cursor->key_beg = leaf->base;
2115 cursor->asof = HAMMER_MAX_TID;
2116 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
2117 cursor->flags |= HAMMER_CURSOR_ASOF;
2118 cursor->flags |= HAMMER_CURSOR_BACKEND;
2119 cursor->flags &= ~HAMMER_CURSOR_INSERT;
2121 error = hammer_btree_lookup(cursor);
2123 error = hammer_ip_delete_record(cursor, ip, trans->tid);
2125 if (error == EDEADLK) {
2126 hammer_done_cursor(cursor);
2127 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
2135 * This function deletes remaining auxillary records when an inode is
2136 * being deleted. This function explicitly does not delete the
2137 * inode record, directory entry, data, or db records. Those must be
2138 * properly disposed of prior to this call.
2141 hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip, int *countp)
2143 hammer_transaction_t trans = cursor->trans;
2144 hammer_btree_leaf_elm_t leaf __debugvar;
2147 KKASSERT(trans->type == HAMMER_TRANS_FLS);
2149 hammer_normalize_cursor(cursor);
2150 cursor->key_beg.localization = ip->obj_localization |
2151 HAMMER_LOCALIZE_MISC;
2152 cursor->key_beg.obj_id = ip->obj_id;
2153 cursor->key_beg.create_tid = 0;
2154 cursor->key_beg.delete_tid = 0;
2155 cursor->key_beg.obj_type = 0;
2156 cursor->key_beg.rec_type = HAMMER_RECTYPE_CLEAN_START;
2157 cursor->key_beg.key = HAMMER_MIN_KEY;
2159 cursor->key_end = cursor->key_beg;
2160 cursor->key_end.rec_type = HAMMER_RECTYPE_MAX;
2161 cursor->key_end.key = HAMMER_MAX_KEY;
2163 cursor->asof = ip->obj_asof;
2164 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
2165 cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2166 cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
2167 cursor->flags |= HAMMER_CURSOR_BACKEND;
2169 error = hammer_ip_first(cursor);
2172 * Iterate through matching records and mark them as deleted.
2174 while (error == 0) {
2175 leaf = cursor->leaf;
2177 KKASSERT(leaf->base.delete_tid == 0);
2180 * Mark the record and B-Tree entry as deleted. This will
2181 * also physically delete the B-Tree entry, record, and
2182 * data if the retention policy dictates. The function
2183 * will set HAMMER_CURSOR_RETEST to cause hammer_ip_next()
2184 * to retest the new 'current' element.
2186 * Directory entries (and delete-on-disk directory entries)
2187 * must be synced and cannot be deleted.
2189 error = hammer_ip_delete_record(cursor, ip, trans->tid);
2193 error = hammer_ip_next(cursor);
2196 hammer_cache_node(&ip->cache[1], cursor->node);
2197 if (error == EDEADLK) {
2198 hammer_done_cursor(cursor);
2199 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
2203 if (error == ENOENT)
2209 * Delete the record at the current cursor. On success the cursor will
2210 * be positioned appropriately for an iteration but may no longer be at
2213 * This routine is only called from the backend.
2215 * NOTE: This can return EDEADLK, requiring the caller to terminate the
2219 hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
2222 hammer_record_t iprec;
2225 KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
2229 * In-memory (unsynchronized) records can simply be freed. This
2230 * only occurs in range iterations since all other records are
2231 * individually synchronized. Thus there should be no confusion with
2234 * An in-memory record may be deleted before being committed to disk,
2235 * but could have been accessed in the mean time. The reservation
2236 * code will deal with the case.
2238 if (hammer_cursor_inmem(cursor)) {
2239 iprec = cursor->iprec;
2240 KKASSERT((iprec->flags & HAMMER_RECF_INTERLOCK_BE) ==0);
2241 iprec->flags |= HAMMER_RECF_DELETED_FE;
2242 iprec->flags |= HAMMER_RECF_DELETED_BE;
2243 KKASSERT(iprec->ip == ip);
2244 ++ip->rec_generation;
2249 * On-disk records are marked as deleted by updating their delete_tid.
2250 * This does not effect their position in the B-Tree (which is based
2251 * on their create_tid).
2253 * Frontend B-Tree operations track inodes so we tell
2254 * hammer_delete_at_cursor() not to.
2256 error = hammer_btree_extract_leaf(cursor);
2258 error = hammer_delete_at_cursor(
2260 HAMMER_DELETE_ADJUST | hammer_nohistory(ip),
2262 cursor->trans->time32,
2269 * Used to write a generic record w/optional data to the media b-tree
2270 * when no inode context is available. Used by the mirroring and
2273 * Caller must set cursor->key_beg to leaf->base. The cursor must be
2274 * flagged for backend operation and not flagged ASOF (since we are
2275 * doing an insertion).
2277 * This function will acquire the appropriate sync lock and will set
2278 * the cursor insertion flag for the operation, do the btree lookup,
2279 * and the insertion, and clear the insertion flag and sync lock before
2280 * returning. The cursor state will be such that the caller can continue
2281 * scanning (used by the mirroring code).
2283 * mode: HAMMER_CREATE_MODE_UMIRROR copyin data, check crc
2284 * HAMMER_CREATE_MODE_SYS bcopy data, generate crc
2286 * NOTE: EDEADLK can be returned. The caller must do deadlock handling and
2289 * EALREADY can be returned if the record already exists (WARNING,
2290 * because ASOF cannot be used no check is made for illegal
2293 * NOTE: Do not use the function for normal inode-related records as this
2294 * functions goes directly to the media and is not integrated with
2295 * in-memory records.
2298 hammer_create_at_cursor(hammer_cursor_t cursor, hammer_btree_leaf_elm_t leaf,
2299 void *udata, int mode)
2301 hammer_transaction_t trans;
2302 hammer_buffer_t data_buffer;
2303 hammer_off_t ndata_offset;
2304 hammer_tid_t high_tid;
2309 trans = cursor->trans;
2314 KKASSERT((cursor->flags &
2315 (HAMMER_CURSOR_BACKEND | HAMMER_CURSOR_ASOF)) ==
2316 (HAMMER_CURSOR_BACKEND));
2318 hammer_sync_lock_sh(trans);
2320 if (leaf->data_len) {
2321 ndata = hammer_alloc_data(trans, leaf->data_len,
2322 leaf->base.rec_type,
2323 &ndata_offset, &data_buffer,
2325 if (ndata == NULL) {
2326 hammer_sync_unlock(trans);
2329 leaf->data_offset = ndata_offset;
2330 hammer_modify_buffer_noundo(trans, data_buffer);
2333 case HAMMER_CREATE_MODE_UMIRROR:
2334 error = copyin(udata, ndata, leaf->data_len);
2336 if (hammer_crc_test_leaf(ndata, leaf) == 0) {
2337 hdkprintf("CRC DATA @ %016llx/%d MISMATCH ON PIPE\n",
2338 (long long)ndata_offset,
2342 error = hammer_cursor_localize_data(
2347 case HAMMER_CREATE_MODE_SYS:
2348 bcopy(udata, ndata, leaf->data_len);
2350 hammer_crc_set_leaf(ndata, leaf);
2353 hpanic("bad mode %d", mode);
2354 break; /* NOT REACHED */
2356 hammer_modify_buffer_done(data_buffer);
2358 leaf->data_offset = 0;
2366 * Do the insertion. This can fail with a EDEADLK or EALREADY
2368 cursor->flags |= HAMMER_CURSOR_INSERT;
2369 error = hammer_btree_lookup(cursor);
2370 if (error != ENOENT) {
2375 error = hammer_btree_insert(cursor, leaf, &doprop);
2378 * Cursor is left on current element, we want to skip it now.
2379 * (in case the caller is scanning)
2381 cursor->flags |= HAMMER_CURSOR_ATEDISK;
2382 cursor->flags &= ~HAMMER_CURSOR_INSERT;
2385 * If the insertion happens to be creating (and not just replacing)
2386 * an inode we have to track it.
2389 leaf->base.rec_type == HAMMER_RECTYPE_INODE &&
2390 leaf->base.delete_tid == 0) {
2391 hammer_modify_volume_field(trans, trans->rootvol,
2393 ++trans->hmp->rootvol->ondisk->vol0_stat_inodes;
2394 hammer_modify_volume_done(trans->rootvol);
2398 * vol0_next_tid must track the highest TID stored in the filesystem.
2399 * We do not need to generate undo for this update.
2401 high_tid = leaf->base.create_tid;
2402 if (high_tid < leaf->base.delete_tid)
2403 high_tid = leaf->base.delete_tid;
2404 if (trans->rootvol->ondisk->vol0_next_tid < high_tid) {
2405 hammer_modify_volume_noundo(trans, trans->rootvol);
2406 trans->rootvol->ondisk->vol0_next_tid = high_tid;
2407 hammer_modify_volume_done(trans->rootvol);
2411 * WARNING! cursor's leaf pointer may have changed after
2412 * do_propagation returns.
2414 if (error == 0 && doprop)
2415 hammer_btree_do_propagation(cursor, NULL, leaf);
2421 if (error && leaf->data_offset) {
2422 hammer_blockmap_free(trans, leaf->data_offset, leaf->data_len);
2425 hammer_sync_unlock(trans);
2427 hammer_rel_buffer(data_buffer, 0);
2432 * Delete the B-Tree element at the current cursor and do any necessary
2433 * mirror propagation.
2435 * The cursor must be properly positioned for an iteration on return but
2436 * may be pointing at an internal element.
2438 * An element can be un-deleted by passing a delete_tid of 0 with
2439 * HAMMER_DELETE_ADJUST.
2441 * This function will store the number of bytes deleted in *stat_bytes
2442 * if stat_bytes is not NULL.
2445 hammer_delete_at_cursor(hammer_cursor_t cursor, int delete_flags,
2446 hammer_tid_t delete_tid, uint32_t delete_ts,
2447 int track, int64_t *stat_bytes)
2449 struct hammer_btree_leaf_elm save_leaf;
2450 hammer_transaction_t trans;
2451 hammer_btree_leaf_elm_t leaf;
2453 hammer_btree_elm_t elm;
2454 hammer_off_t data_offset;
2462 error = hammer_cursor_upgrade(cursor);
2466 trans = cursor->trans;
2467 node = cursor->node;
2468 elm = &node->ondisk->elms[cursor->index];
2470 KKASSERT(elm->base.btype == HAMMER_BTREE_TYPE_RECORD);
2472 hammer_sync_lock_sh(trans);
2478 * Adjust the delete_tid. Update the mirror_tid propagation field
2479 * as well. delete_tid can be 0 (undelete -- used by mirroring).
2481 if (delete_flags & HAMMER_DELETE_ADJUST) {
2482 if (elm->base.rec_type == HAMMER_RECTYPE_INODE) {
2483 if (elm->leaf.base.delete_tid == 0 && delete_tid)
2485 if (elm->leaf.base.delete_tid && delete_tid == 0)
2489 hammer_modify_node(trans, node, elm, sizeof(*elm));
2490 elm->leaf.base.delete_tid = delete_tid;
2491 elm->leaf.delete_ts = delete_ts;
2492 hammer_modify_node_done(node);
2494 if (elm->leaf.base.delete_tid > node->ondisk->mirror_tid) {
2495 hammer_modify_node_field(trans, node, mirror_tid);
2496 node->ondisk->mirror_tid = elm->leaf.base.delete_tid;
2497 hammer_modify_node_done(node);
2499 if (hammer_debug_general & 0x0002) {
2500 hdkprintf("propagate %016llx @%016llx\n",
2501 (long long)elm->leaf.base.delete_tid,
2502 (long long)node->node_offset);
2507 * Adjust for the iteration. We have deleted the current
2508 * element and want to clear ATEDISK so the iteration does
2509 * not skip the element after, which now becomes the current
2510 * element. This element must be re-tested if doing an
2511 * iteration, which is handled by the RETEST flag.
2513 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
2514 cursor->flags |= HAMMER_CURSOR_RETEST;
2515 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
2519 * An on-disk record cannot have the same delete_tid
2520 * as its create_tid. In a chain of record updates
2521 * this could result in a duplicate record.
2523 KKASSERT(elm->leaf.base.delete_tid !=
2524 elm->leaf.base.create_tid);
2528 * Destroy the B-Tree element if asked (typically if a nohistory
2529 * file or mount, or when called by the pruning code).
2531 * Adjust the ATEDISK flag to properly support iterations.
2533 if (delete_flags & HAMMER_DELETE_DESTROY) {
2534 data_offset = elm->leaf.data_offset;
2535 data_len = elm->leaf.data_len;
2537 save_leaf = elm->leaf;
2540 if (elm->base.rec_type == HAMMER_RECTYPE_INODE &&
2541 elm->leaf.base.delete_tid == 0) {
2545 error = hammer_btree_delete(cursor, &ndelete);
2548 * The deletion moves the next element (if any) to
2549 * the current element position. We must clear
2550 * ATEDISK so this element is not skipped and we
2551 * must set RETEST to force any iteration to re-test
2554 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
2555 cursor->flags |= HAMMER_CURSOR_RETEST;
2556 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
2558 bytes += (ndelete * sizeof(struct hammer_node_ondisk));
2560 switch(HAMMER_ZONE(data_offset)) {
2561 case HAMMER_ZONE_LARGE_DATA:
2562 case HAMMER_ZONE_SMALL_DATA:
2563 case HAMMER_ZONE_META:
2564 hammer_blockmap_free(trans,
2565 data_offset, data_len);
2575 * Track inode count and next_tid. This is used by the mirroring
2576 * and PFS code. icount can be negative, zero, or positive.
2578 if (error == 0 && track) {
2580 hammer_modify_volume_field(trans, trans->rootvol,
2582 trans->rootvol->ondisk->vol0_stat_inodes += icount;
2583 hammer_modify_volume_done(trans->rootvol);
2585 if (trans->rootvol->ondisk->vol0_next_tid < delete_tid) {
2586 hammer_modify_volume_noundo(trans, trans->rootvol);
2587 trans->rootvol->ondisk->vol0_next_tid = delete_tid;
2588 hammer_modify_volume_done(trans->rootvol);
2593 * mirror_tid propagation occurs if the node's mirror_tid had to be
2594 * updated while adjusting the delete_tid.
2596 * This occurs when deleting even in nohistory mode, but does not
2597 * occur when pruning an already-deleted node.
2599 * cursor->ip is NULL when called from the pruning, mirroring,
2600 * and pfs code. If non-NULL propagation will be conditionalized
2601 * on whether the PFS is in no-history mode or not.
2603 * WARNING: cursor's leaf pointer may have changed after do_propagation
2608 hammer_btree_do_propagation(cursor, cursor->ip->pfsm, leaf);
2610 hammer_btree_do_propagation(cursor, NULL, leaf);
2613 *stat_bytes = bytes;
2614 hammer_sync_unlock(trans);
2619 * Determine whether we can remove a directory. This routine checks whether
2620 * a directory is empty or not and enforces flush connectivity.
2622 * Flush connectivity requires that we block if the target directory is
2623 * currently flushing, otherwise it may not end up in the same flush group.
2625 * Returns 0 on success, ENOTEMPTY or EDEADLK (or other errors) on failure.
2628 hammer_ip_check_directory_empty(hammer_transaction_t trans, hammer_inode_t ip)
2630 struct hammer_cursor cursor;
2634 * Check directory empty
2636 hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2638 cursor.key_beg.localization = ip->obj_localization |
2639 hammer_dir_localization(ip);
2640 cursor.key_beg.obj_id = ip->obj_id;
2641 cursor.key_beg.create_tid = 0;
2642 cursor.key_beg.delete_tid = 0;
2643 cursor.key_beg.obj_type = 0;
2644 cursor.key_beg.rec_type = HAMMER_RECTYPE_ENTRY_START;
2645 cursor.key_beg.key = HAMMER_MIN_KEY;
2647 cursor.key_end = cursor.key_beg;
2648 cursor.key_end.rec_type = HAMMER_RECTYPE_MAX;
2649 cursor.key_end.key = HAMMER_MAX_KEY;
2651 cursor.asof = ip->obj_asof;
2652 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2654 error = hammer_ip_first(&cursor);
2655 if (error == ENOENT)
2657 else if (error == 0)
2659 hammer_done_cursor(&cursor);
2664 * Localize the data payload. Directory entries may need their
2665 * localization adjusted.
2669 hammer_cursor_localize_data(hammer_data_ondisk_t data,
2670 hammer_btree_leaf_elm_t leaf)
2672 uint32_t localization;
2674 if (leaf->base.rec_type == HAMMER_RECTYPE_DIRENTRY) {
2675 localization = leaf->base.localization &
2676 HAMMER_LOCALIZE_PSEUDOFS_MASK;
2677 if (data->entry.localization != localization) {
2678 data->entry.localization = localization;
2679 hammer_crc_set_leaf(data, leaf);