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
34 * $DragonFly: src/sys/vfs/hammer/hammer_object.c,v 1.81 2008/07/02 21:57:54 dillon Exp $
39 static int hammer_mem_add(hammer_record_t record
);
40 static int hammer_mem_lookup(hammer_cursor_t cursor
);
41 static int hammer_mem_first(hammer_cursor_t cursor
);
42 static int hammer_frontend_trunc_callback(hammer_record_t record
,
44 static int hammer_record_needs_overwrite_delete(hammer_record_t record
);
45 static int hammer_delete_general(hammer_cursor_t cursor
, hammer_inode_t ip
,
46 hammer_btree_leaf_elm_t leaf
);
48 struct rec_trunc_info
{
54 * Red-black tree support. Comparison code for insertion.
57 hammer_rec_rb_compare(hammer_record_t rec1
, hammer_record_t rec2
)
59 if (rec1
->leaf
.base
.rec_type
< rec2
->leaf
.base
.rec_type
)
61 if (rec1
->leaf
.base
.rec_type
> rec2
->leaf
.base
.rec_type
)
64 if (rec1
->leaf
.base
.key
< rec2
->leaf
.base
.key
)
66 if (rec1
->leaf
.base
.key
> rec2
->leaf
.base
.key
)
70 * Never match against an item deleted by the front-end.
72 * rec1 is greater then rec2 if rec1 is marked deleted.
73 * rec1 is less then rec2 if rec2 is marked deleted.
75 * Multiple deleted records may be present, do not return 0
76 * if both are marked deleted.
78 if (rec1
->flags
& HAMMER_RECF_DELETED_FE
)
80 if (rec2
->flags
& HAMMER_RECF_DELETED_FE
)
87 * Basic record comparison code similar to hammer_btree_cmp().
90 hammer_rec_cmp(hammer_base_elm_t elm
, hammer_record_t rec
)
92 if (elm
->rec_type
< rec
->leaf
.base
.rec_type
)
94 if (elm
->rec_type
> rec
->leaf
.base
.rec_type
)
97 if (elm
->key
< rec
->leaf
.base
.key
)
99 if (elm
->key
> rec
->leaf
.base
.key
)
103 * Never match against an item deleted by the front-end.
104 * elm is less then rec if rec is marked deleted.
106 if (rec
->flags
& HAMMER_RECF_DELETED_FE
)
112 * Special LOOKUP_INFO to locate an overlapping record. This used by
113 * the reservation code to implement small-block records (whos keys will
114 * be different depending on data_len, when representing the same base
117 * NOTE: The base file offset of a data record is (key - data_len), not (key).
120 hammer_rec_overlap_compare(hammer_btree_leaf_elm_t leaf
, hammer_record_t rec
)
122 if (leaf
->base
.rec_type
< rec
->leaf
.base
.rec_type
)
124 if (leaf
->base
.rec_type
> rec
->leaf
.base
.rec_type
)
130 if (leaf
->base
.rec_type
== HAMMER_RECTYPE_DATA
) {
131 /* leaf_end <= rec_beg */
132 if (leaf
->base
.key
<= rec
->leaf
.base
.key
- rec
->leaf
.data_len
)
134 /* leaf_beg >= rec_end */
135 if (leaf
->base
.key
- leaf
->data_len
>= rec
->leaf
.base
.key
)
138 if (leaf
->base
.key
< rec
->leaf
.base
.key
)
140 if (leaf
->base
.key
> rec
->leaf
.base
.key
)
145 * Never match against an item deleted by the front-end.
146 * leaf is less then rec if rec is marked deleted.
148 * We must still return the proper code for the scan to continue
149 * along the correct branches.
151 if (rec
->flags
& HAMMER_RECF_DELETED_FE
) {
152 if (leaf
->base
.key
< rec
->leaf
.base
.key
)
154 if (leaf
->base
.key
> rec
->leaf
.base
.key
)
162 * RB_SCAN comparison code for hammer_mem_first(). The argument order
163 * is reversed so the comparison result has to be negated. key_beg and
164 * key_end are both range-inclusive.
166 * Localized deletions are not cached in-memory.
170 hammer_rec_scan_cmp(hammer_record_t rec
, void *data
)
172 hammer_cursor_t cursor
= data
;
175 r
= hammer_rec_cmp(&cursor
->key_beg
, rec
);
178 r
= hammer_rec_cmp(&cursor
->key_end
, rec
);
185 * This compare function is used when simply looking up key_beg.
189 hammer_rec_find_cmp(hammer_record_t rec
, void *data
)
191 hammer_cursor_t cursor
= data
;
194 r
= hammer_rec_cmp(&cursor
->key_beg
, rec
);
203 * Locate blocks within the truncation range. Partial blocks do not count.
207 hammer_rec_trunc_cmp(hammer_record_t rec
, void *data
)
209 struct rec_trunc_info
*info
= data
;
211 if (rec
->leaf
.base
.rec_type
< info
->rec_type
)
213 if (rec
->leaf
.base
.rec_type
> info
->rec_type
)
216 switch(rec
->leaf
.base
.rec_type
) {
217 case HAMMER_RECTYPE_DB
:
219 * DB record key is not beyond the truncation point, retain.
221 if (rec
->leaf
.base
.key
< info
->trunc_off
)
224 case HAMMER_RECTYPE_DATA
:
226 * DATA record offset start is not beyond the truncation point,
229 if (rec
->leaf
.base
.key
- rec
->leaf
.data_len
< info
->trunc_off
)
233 panic("hammer_rec_trunc_cmp: unexpected record type");
237 * The record start is >= the truncation point, return match,
238 * the record should be destroyed.
243 RB_GENERATE(hammer_rec_rb_tree
, hammer_record
, rb_node
, hammer_rec_rb_compare
);
244 RB_GENERATE_XLOOKUP(hammer_rec_rb_tree
, INFO
, hammer_record
, rb_node
,
245 hammer_rec_overlap_compare
, hammer_btree_leaf_elm_t
);
248 * Allocate a record for the caller to finish filling in. The record is
249 * returned referenced.
252 hammer_alloc_mem_record(hammer_inode_t ip
, int data_len
)
254 hammer_record_t record
;
256 ++hammer_count_records
;
257 record
= kmalloc(sizeof(*record
), M_HAMMER
,
258 M_WAITOK
| M_ZERO
| M_USE_RESERVE
);
259 record
->flush_state
= HAMMER_FST_IDLE
;
261 record
->leaf
.base
.btype
= HAMMER_BTREE_TYPE_RECORD
;
262 record
->leaf
.data_len
= data_len
;
263 hammer_ref(&record
->lock
);
266 record
->data
= kmalloc(data_len
, M_HAMMER
, M_WAITOK
| M_ZERO
);
267 record
->flags
|= HAMMER_RECF_ALLOCDATA
;
268 ++hammer_count_record_datas
;
275 hammer_wait_mem_record_ident(hammer_record_t record
, const char *ident
)
277 while (record
->flush_state
== HAMMER_FST_FLUSH
) {
278 record
->flags
|= HAMMER_RECF_WANTED
;
279 tsleep(record
, 0, ident
, 0);
284 * Called from the backend, hammer_inode.c, after a record has been
285 * flushed to disk. The record has been exclusively locked by the
286 * caller and interlocked with BE.
288 * We clean up the state, unlock, and release the record (the record
289 * was referenced by the fact that it was in the HAMMER_FST_FLUSH state).
292 hammer_flush_record_done(hammer_record_t record
, int error
)
294 hammer_inode_t target_ip
;
296 KKASSERT(record
->flush_state
== HAMMER_FST_FLUSH
);
297 KKASSERT(record
->flags
& HAMMER_RECF_INTERLOCK_BE
);
301 * An error occured, the backend was unable to sync the
302 * record to its media. Leave the record intact.
304 Debugger("flush_record_done error");
307 if (record
->flags
& HAMMER_RECF_DELETED_BE
) {
308 if ((target_ip
= record
->target_ip
) != NULL
) {
309 TAILQ_REMOVE(&target_ip
->target_list
, record
,
311 record
->target_ip
= NULL
;
312 hammer_test_inode(target_ip
);
314 record
->flush_state
= HAMMER_FST_IDLE
;
316 if (record
->target_ip
) {
317 record
->flush_state
= HAMMER_FST_SETUP
;
318 hammer_test_inode(record
->ip
);
319 hammer_test_inode(record
->target_ip
);
321 record
->flush_state
= HAMMER_FST_IDLE
;
324 record
->flags
&= ~HAMMER_RECF_INTERLOCK_BE
;
325 if (record
->flags
& HAMMER_RECF_WANTED
) {
326 record
->flags
&= ~HAMMER_RECF_WANTED
;
329 hammer_rel_mem_record(record
);
333 * Release a memory record. Records marked for deletion are immediately
334 * removed from the RB-Tree but otherwise left intact until the last ref
338 hammer_rel_mem_record(struct hammer_record
*record
)
340 hammer_inode_t ip
, target_ip
;
342 hammer_unref(&record
->lock
);
344 if (record
->lock
.refs
== 0) {
346 * Upon release of the last reference wakeup any waiters.
347 * The record structure may get destroyed so callers will
348 * loop up and do a relookup.
350 * WARNING! Record must be removed from RB-TREE before we
351 * might possibly block. hammer_test_inode() can block!
356 * Upon release of the last reference a record marked deleted
359 if (record
->flags
& HAMMER_RECF_DELETED_FE
) {
360 KKASSERT(ip
->lock
.refs
> 0);
361 KKASSERT(record
->flush_state
!= HAMMER_FST_FLUSH
);
364 * target_ip may have zero refs, we have to ref it
365 * to prevent it from being ripped out from under
368 if ((target_ip
= record
->target_ip
) != NULL
) {
369 TAILQ_REMOVE(&target_ip
->target_list
,
370 record
, target_entry
);
371 record
->target_ip
= NULL
;
372 hammer_ref(&target_ip
->lock
);
375 if (record
->flags
& HAMMER_RECF_ONRBTREE
) {
376 RB_REMOVE(hammer_rec_rb_tree
,
377 &record
->ip
->rec_tree
,
379 KKASSERT(ip
->rsv_recs
> 0);
382 ip
->hmp
->rsv_databytes
-= record
->leaf
.data_len
;
383 record
->flags
&= ~HAMMER_RECF_ONRBTREE
;
385 if (RB_EMPTY(&record
->ip
->rec_tree
)) {
386 record
->ip
->flags
&= ~HAMMER_INODE_XDIRTY
;
387 record
->ip
->sync_flags
&= ~HAMMER_INODE_XDIRTY
;
388 hammer_test_inode(record
->ip
);
393 * Do this test after removing record from the B-Tree.
396 hammer_test_inode(target_ip
);
397 hammer_rel_inode(target_ip
, 0);
400 if (record
->flags
& HAMMER_RECF_ALLOCDATA
) {
401 --hammer_count_record_datas
;
402 kfree(record
->data
, M_HAMMER
);
403 record
->flags
&= ~HAMMER_RECF_ALLOCDATA
;
406 hammer_blockmap_reserve_complete(ip
->hmp
,
411 --hammer_count_records
;
412 kfree(record
, M_HAMMER
);
418 * Record visibility depends on whether the record is being accessed by
419 * the backend or the frontend.
421 * Return non-zero if the record is visible, zero if it isn't or if it is
426 hammer_ip_iterate_mem_good(hammer_cursor_t cursor
, hammer_record_t record
)
428 if (cursor
->flags
& HAMMER_CURSOR_BACKEND
) {
429 if (record
->flags
& HAMMER_RECF_DELETED_BE
)
432 if (record
->flags
& HAMMER_RECF_DELETED_FE
)
439 * This callback is used as part of the RB_SCAN function for in-memory
440 * records. We terminate it (return -1) as soon as we get a match.
442 * This routine is used by frontend code.
444 * The primary compare code does not account for ASOF lookups. This
445 * code handles that case as well as a few others.
449 hammer_rec_scan_callback(hammer_record_t rec
, void *data
)
451 hammer_cursor_t cursor
= data
;
454 * We terminate on success, so this should be NULL on entry.
456 KKASSERT(cursor
->iprec
== NULL
);
459 * Skip if the record was marked deleted.
461 if (hammer_ip_iterate_mem_good(cursor
, rec
) == 0)
465 * Skip if not visible due to our as-of TID
467 if (cursor
->flags
& HAMMER_CURSOR_ASOF
) {
468 if (cursor
->asof
< rec
->leaf
.base
.create_tid
)
470 if (rec
->leaf
.base
.delete_tid
&&
471 cursor
->asof
>= rec
->leaf
.base
.delete_tid
) {
477 * If the record is queued to the flusher we have to block until
478 * it isn't. Otherwise we may see duplication between our memory
479 * cache and the media.
481 hammer_ref(&rec
->lock
);
483 #warning "This deadlocks"
485 if (rec
->flush_state
== HAMMER_FST_FLUSH
)
486 hammer_wait_mem_record(rec
);
490 * The record may have been deleted while we were blocked.
492 if (hammer_ip_iterate_mem_good(cursor
, rec
) == 0) {
493 hammer_rel_mem_record(rec
);
498 * Set the matching record and stop the scan.
506 * Lookup an in-memory record given the key specified in the cursor. Works
507 * just like hammer_btree_lookup() but operates on an inode's in-memory
510 * The lookup must fail if the record is marked for deferred deletion.
514 hammer_mem_lookup(hammer_cursor_t cursor
)
518 KKASSERT(cursor
->ip
);
520 hammer_rel_mem_record(cursor
->iprec
);
521 cursor
->iprec
= NULL
;
523 hammer_rec_rb_tree_RB_SCAN(&cursor
->ip
->rec_tree
, hammer_rec_find_cmp
,
524 hammer_rec_scan_callback
, cursor
);
526 if (cursor
->iprec
== NULL
)
534 * hammer_mem_first() - locate the first in-memory record matching the
535 * cursor within the bounds of the key range.
539 hammer_mem_first(hammer_cursor_t cursor
)
544 KKASSERT(ip
!= NULL
);
547 hammer_rel_mem_record(cursor
->iprec
);
548 cursor
->iprec
= NULL
;
551 hammer_rec_rb_tree_RB_SCAN(&ip
->rec_tree
, hammer_rec_scan_cmp
,
552 hammer_rec_scan_callback
, cursor
);
555 * Adjust scan.node and keep it linked into the RB-tree so we can
556 * hold the cursor through third party modifications of the RB-tree.
564 hammer_mem_done(hammer_cursor_t cursor
)
567 hammer_rel_mem_record(cursor
->iprec
);
568 cursor
->iprec
= NULL
;
572 /************************************************************************
573 * HAMMER IN-MEMORY RECORD FUNCTIONS *
574 ************************************************************************
576 * These functions manipulate in-memory records. Such records typically
577 * exist prior to being committed to disk or indexed via the on-disk B-Tree.
581 * Add a directory entry (dip,ncp) which references inode (ip).
583 * Note that the low 32 bits of the namekey are set temporarily to create
584 * a unique in-memory record, and may be modified a second time when the
585 * record is synchronized to disk. In particular, the low 32 bits cannot be
586 * all 0's when synching to disk, which is not handled here.
588 * NOTE: bytes does not include any terminating \0 on name, and name might
592 hammer_ip_add_directory(struct hammer_transaction
*trans
,
593 struct hammer_inode
*dip
, const char *name
, int bytes
,
594 struct hammer_inode
*ip
)
596 struct hammer_cursor cursor
;
597 hammer_record_t record
;
602 record
= hammer_alloc_mem_record(dip
, HAMMER_ENTRY_SIZE(bytes
));
603 if (++trans
->hmp
->namekey_iterator
== 0)
604 ++trans
->hmp
->namekey_iterator
;
606 record
->type
= HAMMER_MEM_RECORD_ADD
;
607 record
->leaf
.base
.localization
= dip
->obj_localization
+
608 HAMMER_LOCALIZE_MISC
;
609 record
->leaf
.base
.obj_id
= dip
->obj_id
;
610 record
->leaf
.base
.key
= hammer_directory_namekey(name
, bytes
);
611 record
->leaf
.base
.key
+= trans
->hmp
->namekey_iterator
;
612 record
->leaf
.base
.rec_type
= HAMMER_RECTYPE_DIRENTRY
;
613 record
->leaf
.base
.obj_type
= ip
->ino_leaf
.base
.obj_type
;
614 record
->data
->entry
.obj_id
= ip
->obj_id
;
615 record
->data
->entry
.localization
= ip
->obj_localization
;
616 bcopy(name
, record
->data
->entry
.name
, bytes
);
618 ++ip
->ino_data
.nlinks
;
619 hammer_modify_inode(ip
, HAMMER_INODE_DDIRTY
);
622 * Find an unused namekey. Both the in-memory record tree and
623 * the B-Tree are checked. Exact matches also match create_tid
624 * so use an ASOF search to (mostly) ignore it.
626 * delete-visibility is set so pending deletions do not give us
627 * a false-negative on our ability to use an iterator.
629 hammer_init_cursor(trans
, &cursor
, &dip
->cache
[1], dip
);
630 cursor
.key_beg
= record
->leaf
.base
;
631 cursor
.flags
|= HAMMER_CURSOR_ASOF
;
632 cursor
.flags
|= HAMMER_CURSOR_DELETE_VISIBILITY
;
633 cursor
.asof
= ip
->obj_asof
;
636 while (hammer_ip_lookup(&cursor
) == 0) {
637 iterator
= (u_int32_t
)record
->leaf
.base
.key
+ 1;
640 record
->leaf
.base
.key
&= ~0xFFFFFFFFLL
;
641 record
->leaf
.base
.key
|= iterator
;
642 cursor
.key_beg
.key
= record
->leaf
.base
.key
;
643 if (++count
== 1000000000) {
644 hammer_rel_mem_record(record
);
651 * The target inode and the directory entry are bound together.
653 record
->target_ip
= ip
;
654 record
->flush_state
= HAMMER_FST_SETUP
;
655 TAILQ_INSERT_TAIL(&ip
->target_list
, record
, target_entry
);
658 * The inode now has a dependancy and must be taken out of the idle
659 * state. An inode not in an idle state is given an extra reference.
661 if (ip
->flush_state
== HAMMER_FST_IDLE
) {
662 hammer_ref(&ip
->lock
);
663 ip
->flush_state
= HAMMER_FST_SETUP
;
665 error
= hammer_mem_add(record
);
667 hammer_done_cursor(&cursor
);
672 * Delete the directory entry and update the inode link count. The
673 * cursor must be seeked to the directory entry record being deleted.
675 * The related inode should be share-locked by the caller. The caller is
678 * This function can return EDEADLK requiring the caller to terminate
679 * the cursor, any locks, wait on the returned record, and retry.
682 hammer_ip_del_directory(struct hammer_transaction
*trans
,
683 hammer_cursor_t cursor
, struct hammer_inode
*dip
,
684 struct hammer_inode
*ip
)
686 hammer_record_t record
;
689 if (hammer_cursor_inmem(cursor
)) {
691 * In-memory (unsynchronized) records can simply be freed.
692 * Even though the HAMMER_RECF_DELETED_FE flag is ignored
693 * by the backend, we must still avoid races against the
694 * backend potentially syncing the record to the media.
696 * We cannot call hammer_ip_delete_record(), that routine may
697 * only be called from the backend.
699 record
= cursor
->iprec
;
700 if (record
->flags
& HAMMER_RECF_INTERLOCK_BE
) {
701 KKASSERT(cursor
->deadlk_rec
== NULL
);
702 hammer_ref(&record
->lock
);
703 cursor
->deadlk_rec
= record
;
706 KKASSERT(record
->type
== HAMMER_MEM_RECORD_ADD
);
707 record
->flags
|= HAMMER_RECF_DELETED_FE
;
712 * If the record is on-disk we have to queue the deletion by
713 * the record's key. This also causes lookups to skip the
716 KKASSERT(dip
->flags
&
717 (HAMMER_INODE_ONDISK
| HAMMER_INODE_DONDISK
));
718 record
= hammer_alloc_mem_record(dip
, 0);
719 record
->type
= HAMMER_MEM_RECORD_DEL
;
720 record
->leaf
.base
= cursor
->leaf
->base
;
722 record
->target_ip
= ip
;
723 record
->flush_state
= HAMMER_FST_SETUP
;
724 TAILQ_INSERT_TAIL(&ip
->target_list
, record
, target_entry
);
727 * The inode now has a dependancy and must be taken out of
728 * the idle state. An inode not in an idle state is given
729 * an extra reference.
731 if (ip
->flush_state
== HAMMER_FST_IDLE
) {
732 hammer_ref(&ip
->lock
);
733 ip
->flush_state
= HAMMER_FST_SETUP
;
736 error
= hammer_mem_add(record
);
740 * One less link. The file may still be open in the OS even after
741 * all links have gone away.
743 * We have to terminate the cursor before syncing the inode to
744 * avoid deadlocking against ourselves. XXX this may no longer
747 * If nlinks drops to zero and the vnode is inactive (or there is
748 * no vnode), call hammer_inode_unloadable_check() to zonk the
749 * inode. If we don't do this here the inode will not be destroyed
750 * on-media until we unmount.
753 --ip
->ino_data
.nlinks
;
754 hammer_modify_inode(ip
, HAMMER_INODE_DDIRTY
);
755 if (ip
->ino_data
.nlinks
== 0 &&
756 (ip
->vp
== NULL
|| (ip
->vp
->v_flag
& VINACTIVE
))) {
757 hammer_done_cursor(cursor
);
758 hammer_inode_unloadable_check(ip
, 1);
759 hammer_flush_inode(ip
, 0);
767 * Add a record to an inode.
769 * The caller must allocate the record with hammer_alloc_mem_record(ip) and
770 * initialize the following additional fields:
772 * The related inode should be share-locked by the caller. The caller is
775 * record->rec.entry.base.base.key
776 * record->rec.entry.base.base.rec_type
777 * record->rec.entry.base.base.data_len
778 * record->data (a copy will be kmalloc'd if it cannot be embedded)
781 hammer_ip_add_record(struct hammer_transaction
*trans
, hammer_record_t record
)
783 hammer_inode_t ip
= record
->ip
;
786 KKASSERT(record
->leaf
.base
.localization
!= 0);
787 record
->leaf
.base
.obj_id
= ip
->obj_id
;
788 record
->leaf
.base
.obj_type
= ip
->ino_leaf
.base
.obj_type
;
789 error
= hammer_mem_add(record
);
794 * Locate a bulk record in-memory. Bulk records allow disk space to be
795 * reserved so the front-end can flush large data writes without having
796 * to queue the BIO to the flusher. Only the related record gets queued
799 static hammer_record_t
800 hammer_ip_get_bulk(hammer_inode_t ip
, off_t file_offset
, int bytes
)
802 hammer_record_t record
;
803 struct hammer_btree_leaf_elm leaf
;
805 bzero(&leaf
, sizeof(leaf
));
806 leaf
.base
.obj_id
= ip
->obj_id
;
807 leaf
.base
.key
= file_offset
+ bytes
;
808 leaf
.base
.create_tid
= 0;
809 leaf
.base
.delete_tid
= 0;
810 leaf
.base
.rec_type
= HAMMER_RECTYPE_DATA
;
811 leaf
.base
.obj_type
= 0; /* unused */
812 leaf
.base
.btype
= HAMMER_BTREE_TYPE_RECORD
; /* unused */
813 leaf
.base
.localization
= ip
->obj_localization
+ HAMMER_LOCALIZE_MISC
;
814 leaf
.data_len
= bytes
;
816 record
= hammer_rec_rb_tree_RB_LOOKUP_INFO(&ip
->rec_tree
, &leaf
);
818 hammer_ref(&record
->lock
);
823 * Reserve blockmap space placemarked with an in-memory record.
825 * This routine is called by the frontend in order to be able to directly
826 * flush a buffer cache buffer. The frontend has locked the related buffer
827 * cache buffers and we should be able to manipulate any overlapping
831 hammer_ip_add_bulk(hammer_inode_t ip
, off_t file_offset
, void *data
, int bytes
,
834 hammer_record_t record
;
835 hammer_record_t conflict
;
840 * Deal with conflicting in-memory records. We cannot have multiple
841 * in-memory records for the same offset without seriously confusing
842 * the backend, including but not limited to the backend issuing
843 * delete-create-delete sequences and asserting on the delete_tid
844 * being the same as the create_tid.
846 * If we encounter a record with the backend interlock set we cannot
847 * immediately delete it without confusing the backend.
849 while ((conflict
= hammer_ip_get_bulk(ip
, file_offset
, bytes
)) !=NULL
) {
850 if (conflict
->flags
& HAMMER_RECF_INTERLOCK_BE
) {
851 conflict
->flags
|= HAMMER_RECF_WANTED
;
852 tsleep(conflict
, 0, "hmrrc3", 0);
854 conflict
->flags
|= HAMMER_RECF_DELETED_FE
;
856 hammer_rel_mem_record(conflict
);
860 * Create a record to cover the direct write. This is called with
861 * the related BIO locked so there should be no possible conflict.
863 * The backend is responsible for finalizing the space reserved in
866 * XXX bytes not aligned, depend on the reservation code to
867 * align the reservation.
869 record
= hammer_alloc_mem_record(ip
, 0);
870 zone
= (bytes
>= HAMMER_BUFSIZE
) ? HAMMER_ZONE_LARGE_DATA_INDEX
:
871 HAMMER_ZONE_SMALL_DATA_INDEX
;
872 record
->resv
= hammer_blockmap_reserve(ip
->hmp
, zone
, bytes
,
873 &record
->leaf
.data_offset
,
875 if (record
->resv
== NULL
) {
876 kprintf("hammer_ip_add_bulk: reservation failed\n");
877 hammer_rel_mem_record(record
);
880 record
->type
= HAMMER_MEM_RECORD_DATA
;
881 record
->leaf
.base
.rec_type
= HAMMER_RECTYPE_DATA
;
882 record
->leaf
.base
.obj_type
= ip
->ino_leaf
.base
.obj_type
;
883 record
->leaf
.base
.obj_id
= ip
->obj_id
;
884 record
->leaf
.base
.key
= file_offset
+ bytes
;
885 record
->leaf
.base
.localization
= ip
->obj_localization
+
886 HAMMER_LOCALIZE_MISC
;
887 record
->leaf
.data_len
= bytes
;
888 hammer_crc_set_leaf(data
, &record
->leaf
);
889 flags
= record
->flags
;
891 hammer_ref(&record
->lock
); /* mem_add eats a reference */
892 *errorp
= hammer_mem_add(record
);
894 conflict
= hammer_ip_get_bulk(ip
, file_offset
, bytes
);
895 kprintf("hammer_ip_add_bulk: error %d conflict %p file_offset %lld bytes %d\n",
896 *errorp
, conflict
, file_offset
, bytes
);
898 kprintf("conflict %lld %d\n", conflict
->leaf
.base
.key
, conflict
->leaf
.data_len
);
900 hammer_rel_mem_record(conflict
);
902 KKASSERT(*errorp
== 0);
903 conflict
= hammer_ip_get_bulk(ip
, file_offset
, bytes
);
904 if (conflict
!= record
) {
905 kprintf("conflict mismatch %p %p %08x\n", conflict
, record
, record
->flags
);
907 kprintf("conflict mismatch %lld/%d %lld/%d\n", conflict
->leaf
.base
.key
, conflict
->leaf
.data_len
, record
->leaf
.base
.key
, record
->leaf
.data_len
);
909 KKASSERT(conflict
== record
);
910 hammer_rel_mem_record(conflict
);
916 * Frontend truncation code. Scan in-memory records only. On-disk records
917 * and records in a flushing state are handled by the backend. The vnops
918 * setattr code will handle the block containing the truncation point.
920 * Partial blocks are not deleted.
923 hammer_ip_frontend_trunc(struct hammer_inode
*ip
, off_t file_size
)
925 struct rec_trunc_info info
;
927 switch(ip
->ino_data
.obj_type
) {
928 case HAMMER_OBJTYPE_REGFILE
:
929 info
.rec_type
= HAMMER_RECTYPE_DATA
;
931 case HAMMER_OBJTYPE_DBFILE
:
932 info
.rec_type
= HAMMER_RECTYPE_DB
;
937 info
.trunc_off
= file_size
;
938 hammer_rec_rb_tree_RB_SCAN(&ip
->rec_tree
, hammer_rec_trunc_cmp
,
939 hammer_frontend_trunc_callback
, &info
);
944 hammer_frontend_trunc_callback(hammer_record_t record
, void *data __unused
)
946 if (record
->flags
& HAMMER_RECF_DELETED_FE
)
948 if (record
->flush_state
== HAMMER_FST_FLUSH
)
950 KKASSERT((record
->flags
& HAMMER_RECF_INTERLOCK_BE
) == 0);
951 hammer_ref(&record
->lock
);
952 record
->flags
|= HAMMER_RECF_DELETED_FE
;
953 hammer_rel_mem_record(record
);
958 * Return 1 if the caller must check for and delete existing records
959 * before writing out a new data record.
961 * Return 0 if the caller can just insert the record into the B-Tree without
965 hammer_record_needs_overwrite_delete(hammer_record_t record
)
967 hammer_inode_t ip
= record
->ip
;
971 if (ip
->ino_data
.obj_type
== HAMMER_OBJTYPE_DBFILE
)
972 file_offset
= record
->leaf
.base
.key
;
974 file_offset
= record
->leaf
.base
.key
- record
->leaf
.data_len
;
975 r
= (file_offset
< ip
->save_trunc_off
);
976 if (ip
->ino_data
.obj_type
== HAMMER_OBJTYPE_DBFILE
) {
977 if (ip
->save_trunc_off
<= record
->leaf
.base
.key
)
978 ip
->save_trunc_off
= record
->leaf
.base
.key
+ 1;
980 if (ip
->save_trunc_off
< record
->leaf
.base
.key
)
981 ip
->save_trunc_off
= record
->leaf
.base
.key
;
987 * Backend code. Sync a record to the media.
990 hammer_ip_sync_record_cursor(hammer_cursor_t cursor
, hammer_record_t record
)
992 hammer_transaction_t trans
= cursor
->trans
;
998 KKASSERT(record
->flush_state
== HAMMER_FST_FLUSH
);
999 KKASSERT(record
->flags
& HAMMER_RECF_INTERLOCK_BE
);
1000 KKASSERT(record
->leaf
.base
.localization
!= 0);
1003 * If this is a bulk-data record placemarker there may be an existing
1004 * record on-disk, indicating a data overwrite. If there is the
1005 * on-disk record must be deleted before we can insert our new record.
1007 * We've synthesized this record and do not know what the create_tid
1008 * on-disk is, nor how much data it represents.
1010 * Keep in mind that (key) for data records is (base_offset + len),
1011 * not (base_offset). Also, we only want to get rid of on-disk
1012 * records since we are trying to sync our in-memory record, call
1013 * hammer_ip_delete_range() with truncating set to 1 to make sure
1014 * it skips in-memory records.
1016 * It is ok for the lookup to return ENOENT.
1018 * NOTE OPTIMIZATION: sync_trunc_off is used to determine if we have
1019 * to call hammer_ip_delete_range() or not. This also means we must
1020 * update sync_trunc_off() as we write.
1022 if (record
->type
== HAMMER_MEM_RECORD_DATA
&&
1023 hammer_record_needs_overwrite_delete(record
)) {
1024 file_offset
= record
->leaf
.base
.key
- record
->leaf
.data_len
;
1025 bytes
= (record
->leaf
.data_len
+ HAMMER_BUFMASK
) &
1027 KKASSERT((file_offset
& HAMMER_BUFMASK
) == 0);
1028 error
= hammer_ip_delete_range(
1030 file_offset
, file_offset
+ bytes
- 1,
1032 if (error
&& error
!= ENOENT
)
1037 * If this is a general record there may be an on-disk version
1038 * that must be deleted before we can insert the new record.
1040 if (record
->type
== HAMMER_MEM_RECORD_GENERAL
) {
1041 error
= hammer_delete_general(cursor
, record
->ip
,
1043 if (error
&& error
!= ENOENT
)
1050 hammer_normalize_cursor(cursor
);
1051 cursor
->key_beg
= record
->leaf
.base
;
1052 cursor
->flags
&= ~HAMMER_CURSOR_INITMASK
;
1053 cursor
->flags
|= HAMMER_CURSOR_BACKEND
;
1054 cursor
->flags
&= ~HAMMER_CURSOR_INSERT
;
1057 * Records can wind up on-media before the inode itself is on-media.
1060 record
->ip
->flags
|= HAMMER_INODE_DONDISK
;
1063 * If we are deleting a directory entry an exact match must be
1066 if (record
->type
== HAMMER_MEM_RECORD_DEL
) {
1067 error
= hammer_btree_lookup(cursor
);
1069 error
= hammer_ip_delete_record(cursor
, record
->ip
,
1072 record
->flags
|= HAMMER_RECF_DELETED_FE
;
1073 record
->flags
|= HAMMER_RECF_DELETED_BE
;
1082 * Issue a lookup to position the cursor and locate the cluster. The
1083 * target key should not exist. If we are creating a directory entry
1084 * we may have to iterate the low 32 bits of the key to find an unused
1087 cursor
->flags
|= HAMMER_CURSOR_INSERT
;
1089 error
= hammer_btree_lookup(cursor
);
1090 if (hammer_debug_inode
)
1091 kprintf("DOINSERT LOOKUP %d\n", error
);
1093 kprintf("hammer_ip_sync_record: duplicate rec "
1094 "at (%016llx)\n", record
->leaf
.base
.key
);
1095 Debugger("duplicate record1");
1099 if (record
->type
== HAMMER_MEM_RECORD_DATA
)
1100 kprintf("sync_record %016llx ---------------- %016llx %d\n",
1101 record
->leaf
.base
.key
- record
->leaf
.data_len
,
1102 record
->leaf
.data_offset
, error
);
1105 if (error
!= ENOENT
)
1109 * Allocate the record and data. The result buffers will be
1110 * marked as being modified and further calls to
1111 * hammer_modify_buffer() will result in unneeded UNDO records.
1113 * Support zero-fill records (data == NULL and data_len != 0)
1115 if (record
->type
== HAMMER_MEM_RECORD_DATA
) {
1117 * The data portion of a bulk-data record has already been
1118 * committed to disk, we need only adjust the layer2
1119 * statistics in the same transaction as our B-Tree insert.
1121 KKASSERT(record
->leaf
.data_offset
!= 0);
1122 hammer_blockmap_finalize(trans
, record
->leaf
.data_offset
,
1123 record
->leaf
.data_len
);
1125 } else if (record
->data
&& record
->leaf
.data_len
) {
1127 * Wholely cached record, with data. Allocate the data.
1129 bdata
= hammer_alloc_data(trans
, record
->leaf
.data_len
,
1130 record
->leaf
.base
.rec_type
,
1131 &record
->leaf
.data_offset
,
1132 &cursor
->data_buffer
, &error
);
1135 hammer_crc_set_leaf(record
->data
, &record
->leaf
);
1136 hammer_modify_buffer(trans
, cursor
->data_buffer
, NULL
, 0);
1137 bcopy(record
->data
, bdata
, record
->leaf
.data_len
);
1138 hammer_modify_buffer_done(cursor
->data_buffer
);
1141 * Wholely cached record, without data.
1143 record
->leaf
.data_offset
= 0;
1144 record
->leaf
.data_crc
= 0;
1147 error
= hammer_btree_insert(cursor
, &record
->leaf
);
1148 if (hammer_debug_inode
&& error
)
1149 kprintf("BTREE INSERT error %d @ %016llx:%d key %016llx\n", error
, cursor
->node
->node_offset
, cursor
->index
, record
->leaf
.base
.key
);
1152 * Our record is on-disk, normally mark the in-memory version as
1153 * deleted. If the record represented a directory deletion but
1154 * we had to sync a valid directory entry to disk we must convert
1155 * the record to a covering delete so the frontend does not have
1156 * visibility on the synced entry.
1159 if (record
->flags
& HAMMER_RECF_CONVERT_DELETE
) {
1160 KKASSERT(record
->type
== HAMMER_MEM_RECORD_ADD
);
1161 record
->flags
&= ~HAMMER_RECF_DELETED_FE
;
1162 record
->type
= HAMMER_MEM_RECORD_DEL
;
1163 KKASSERT(record
->flush_state
== HAMMER_FST_FLUSH
);
1164 record
->flags
&= ~HAMMER_RECF_CONVERT_DELETE
;
1165 /* hammer_flush_record_done takes care of the rest */
1167 record
->flags
|= HAMMER_RECF_DELETED_FE
;
1168 record
->flags
|= HAMMER_RECF_DELETED_BE
;
1171 if (record
->leaf
.data_offset
) {
1172 hammer_blockmap_free(trans
, record
->leaf
.data_offset
,
1173 record
->leaf
.data_len
);
1182 * Add the record to the inode's rec_tree. The low 32 bits of a directory
1183 * entry's key is used to deal with hash collisions in the upper 32 bits.
1184 * A unique 64 bit key is generated in-memory and may be regenerated a
1185 * second time when the directory record is flushed to the on-disk B-Tree.
1187 * A referenced record is passed to this function. This function
1188 * eats the reference. If an error occurs the record will be deleted.
1190 * A copy of the temporary record->data pointer provided by the caller
1195 hammer_mem_add(hammer_record_t record
)
1197 hammer_mount_t hmp
= record
->ip
->hmp
;
1200 * Make a private copy of record->data
1203 KKASSERT(record
->flags
& HAMMER_RECF_ALLOCDATA
);
1206 * Insert into the RB tree. A unique key should have already
1207 * been selected if this is a directory entry.
1209 if (RB_INSERT(hammer_rec_rb_tree
, &record
->ip
->rec_tree
, record
)) {
1210 record
->flags
|= HAMMER_RECF_DELETED_FE
;
1211 hammer_rel_mem_record(record
);
1214 ++hmp
->count_newrecords
;
1216 ++record
->ip
->rsv_recs
;
1217 record
->ip
->hmp
->rsv_databytes
+= record
->leaf
.data_len
;
1218 record
->flags
|= HAMMER_RECF_ONRBTREE
;
1219 hammer_modify_inode(record
->ip
, HAMMER_INODE_XDIRTY
);
1220 hammer_rel_mem_record(record
);
1224 /************************************************************************
1225 * HAMMER INODE MERGED-RECORD FUNCTIONS *
1226 ************************************************************************
1228 * These functions augment the B-Tree scanning functions in hammer_btree.c
1229 * by merging in-memory records with on-disk records.
1233 * Locate a particular record either in-memory or on-disk.
1235 * NOTE: This is basically a standalone routine, hammer_ip_next() may
1236 * NOT be called to iterate results.
1239 hammer_ip_lookup(hammer_cursor_t cursor
)
1244 * If the element is in-memory return it without searching the
1247 KKASSERT(cursor
->ip
);
1248 error
= hammer_mem_lookup(cursor
);
1250 cursor
->leaf
= &cursor
->iprec
->leaf
;
1253 if (error
!= ENOENT
)
1257 * If the inode has on-disk components search the on-disk B-Tree.
1259 if ((cursor
->ip
->flags
& (HAMMER_INODE_ONDISK
|HAMMER_INODE_DONDISK
)) == 0)
1261 error
= hammer_btree_lookup(cursor
);
1263 error
= hammer_btree_extract(cursor
, HAMMER_CURSOR_GET_LEAF
);
1268 * Locate the first record within the cursor's key_beg/key_end range,
1269 * restricted to a particular inode. 0 is returned on success, ENOENT
1270 * if no records matched the requested range, or some other error.
1272 * When 0 is returned hammer_ip_next() may be used to iterate additional
1273 * records within the requested range.
1275 * This function can return EDEADLK, requiring the caller to terminate
1276 * the cursor and try again.
1279 hammer_ip_first(hammer_cursor_t cursor
)
1281 hammer_inode_t ip
= cursor
->ip
;
1284 KKASSERT(ip
!= NULL
);
1287 * Clean up fields and setup for merged scan
1289 cursor
->flags
&= ~HAMMER_CURSOR_DELBTREE
;
1290 cursor
->flags
|= HAMMER_CURSOR_ATEDISK
| HAMMER_CURSOR_ATEMEM
;
1291 cursor
->flags
|= HAMMER_CURSOR_DISKEOF
| HAMMER_CURSOR_MEMEOF
;
1292 if (cursor
->iprec
) {
1293 hammer_rel_mem_record(cursor
->iprec
);
1294 cursor
->iprec
= NULL
;
1298 * Search the on-disk B-Tree. hammer_btree_lookup() only does an
1299 * exact lookup so if we get ENOENT we have to call the iterate
1300 * function to validate the first record after the begin key.
1302 * The ATEDISK flag is used by hammer_btree_iterate to determine
1303 * whether it must index forwards or not. It is also used here
1304 * to select the next record from in-memory or on-disk.
1306 * EDEADLK can only occur if the lookup hit an empty internal
1307 * element and couldn't delete it. Since this could only occur
1308 * in-range, we can just iterate from the failure point.
1310 if (ip
->flags
& (HAMMER_INODE_ONDISK
|HAMMER_INODE_DONDISK
)) {
1311 error
= hammer_btree_lookup(cursor
);
1312 if (error
== ENOENT
|| error
== EDEADLK
) {
1313 cursor
->flags
&= ~HAMMER_CURSOR_ATEDISK
;
1314 if (hammer_debug_general
& 0x2000)
1315 kprintf("error %d node %p %016llx index %d\n", error
, cursor
->node
, cursor
->node
->node_offset
, cursor
->index
);
1316 error
= hammer_btree_iterate(cursor
);
1318 if (error
&& error
!= ENOENT
)
1321 cursor
->flags
&= ~HAMMER_CURSOR_DISKEOF
;
1322 cursor
->flags
&= ~HAMMER_CURSOR_ATEDISK
;
1324 cursor
->flags
|= HAMMER_CURSOR_ATEDISK
;
1329 * Search the in-memory record list (Red-Black tree). Unlike the
1330 * B-Tree search, mem_first checks for records in the range.
1332 error
= hammer_mem_first(cursor
);
1333 if (error
&& error
!= ENOENT
)
1336 cursor
->flags
&= ~HAMMER_CURSOR_MEMEOF
;
1337 cursor
->flags
&= ~HAMMER_CURSOR_ATEMEM
;
1338 if (hammer_ip_iterate_mem_good(cursor
, cursor
->iprec
) == 0)
1339 cursor
->flags
|= HAMMER_CURSOR_ATEMEM
;
1343 * This will return the first matching record.
1345 return(hammer_ip_next(cursor
));
1349 * Retrieve the next record in a merged iteration within the bounds of the
1350 * cursor. This call may be made multiple times after the cursor has been
1351 * initially searched with hammer_ip_first().
1353 * 0 is returned on success, ENOENT if no further records match the
1354 * requested range, or some other error code is returned.
1357 hammer_ip_next(hammer_cursor_t cursor
)
1359 hammer_btree_elm_t elm
;
1360 hammer_record_t rec
, save
;
1366 * Load the current on-disk and in-memory record. If we ate any
1367 * records we have to get the next one.
1369 * If we deleted the last on-disk record we had scanned ATEDISK will
1370 * be clear and DELBTREE will be set, forcing a call to iterate. The
1371 * fact that ATEDISK is clear causes iterate to re-test the 'current'
1372 * element. If ATEDISK is set, iterate will skip the 'current'
1375 * Get the next on-disk record
1377 if (cursor
->flags
& (HAMMER_CURSOR_ATEDISK
|HAMMER_CURSOR_DELBTREE
)) {
1378 if ((cursor
->flags
& HAMMER_CURSOR_DISKEOF
) == 0) {
1379 error
= hammer_btree_iterate(cursor
);
1380 cursor
->flags
&= ~HAMMER_CURSOR_DELBTREE
;
1382 cursor
->flags
&= ~HAMMER_CURSOR_ATEDISK
;
1383 hammer_cache_node(&cursor
->ip
->cache
[1],
1386 cursor
->flags
|= HAMMER_CURSOR_DISKEOF
|
1387 HAMMER_CURSOR_ATEDISK
;
1394 * Get the next in-memory record. The record can be ripped out
1395 * of the RB tree so we maintain a scan_info structure to track
1398 * hammer_rec_scan_cmp: Is the record still in our general range,
1399 * (non-inclusive of snapshot exclusions)?
1400 * hammer_rec_scan_callback: Is the record in our snapshot?
1402 if (cursor
->flags
& HAMMER_CURSOR_ATEMEM
) {
1403 if ((cursor
->flags
& HAMMER_CURSOR_MEMEOF
) == 0) {
1404 save
= cursor
->iprec
;
1405 cursor
->iprec
= NULL
;
1406 rec
= save
? hammer_rec_rb_tree_RB_NEXT(save
) : NULL
;
1408 if (hammer_rec_scan_cmp(rec
, cursor
) != 0)
1410 if (hammer_rec_scan_callback(rec
, cursor
) != 0)
1412 rec
= hammer_rec_rb_tree_RB_NEXT(rec
);
1415 hammer_rel_mem_record(save
);
1416 if (cursor
->iprec
) {
1417 KKASSERT(cursor
->iprec
== rec
);
1418 cursor
->flags
&= ~HAMMER_CURSOR_ATEMEM
;
1420 cursor
->flags
|= HAMMER_CURSOR_MEMEOF
;
1426 * The memory record may have become stale while being held in
1427 * cursor->iprec. We are interlocked against the backend on
1428 * with regards to B-Tree entries.
1430 if ((cursor
->flags
& HAMMER_CURSOR_ATEMEM
) == 0) {
1431 if (hammer_ip_iterate_mem_good(cursor
, cursor
->iprec
) == 0) {
1432 cursor
->flags
|= HAMMER_CURSOR_ATEMEM
;
1438 * Extract either the disk or memory record depending on their
1439 * relative position.
1442 switch(cursor
->flags
& (HAMMER_CURSOR_ATEDISK
| HAMMER_CURSOR_ATEMEM
)) {
1445 * Both entries valid. Compare the entries and nominally
1446 * return the first one in the sort order. Numerous cases
1447 * require special attention, however.
1449 elm
= &cursor
->node
->ondisk
->elms
[cursor
->index
];
1450 r
= hammer_btree_cmp(&elm
->base
, &cursor
->iprec
->leaf
.base
);
1453 * If the two entries differ only by their key (-2/2) or
1454 * create_tid (-1/1), and are DATA records, we may have a
1455 * nominal match. We have to calculate the base file
1456 * offset of the data.
1458 if (r
<= 2 && r
>= -2 && r
!= 0 &&
1459 cursor
->ip
->ino_data
.obj_type
== HAMMER_OBJTYPE_REGFILE
&&
1460 cursor
->iprec
->type
== HAMMER_MEM_RECORD_DATA
) {
1461 int64_t base1
= elm
->leaf
.base
.key
- elm
->leaf
.data_len
;
1462 int64_t base2
= cursor
->iprec
->leaf
.base
.key
-
1463 cursor
->iprec
->leaf
.data_len
;
1469 error
= hammer_btree_extract(cursor
,
1470 HAMMER_CURSOR_GET_LEAF
);
1471 cursor
->flags
|= HAMMER_CURSOR_ATEDISK
;
1476 * If the entries match exactly the memory entry is either
1477 * an on-disk directory entry deletion or a bulk data
1478 * overwrite. If it is a directory entry deletion we eat
1481 * For the bulk-data overwrite case it is possible to have
1482 * visibility into both, which simply means the syncer
1483 * hasn't gotten around to doing the delete+insert sequence
1484 * on the B-Tree. Use the memory entry and throw away the
1487 * If the in-memory record is not either of these we
1488 * probably caught the syncer while it was syncing it to
1489 * the media. Since we hold a shared lock on the cursor,
1490 * the in-memory record had better be marked deleted at
1494 if (cursor
->iprec
->type
== HAMMER_MEM_RECORD_DEL
) {
1495 if ((cursor
->flags
& HAMMER_CURSOR_DELETE_VISIBILITY
) == 0) {
1496 cursor
->flags
|= HAMMER_CURSOR_ATEDISK
;
1497 cursor
->flags
|= HAMMER_CURSOR_ATEMEM
;
1500 } else if (cursor
->iprec
->type
== HAMMER_MEM_RECORD_DATA
) {
1501 if ((cursor
->flags
& HAMMER_CURSOR_DELETE_VISIBILITY
) == 0) {
1502 cursor
->flags
|= HAMMER_CURSOR_ATEDISK
;
1504 /* fall through to memory entry */
1506 panic("hammer_ip_next: duplicate mem/b-tree entry");
1507 cursor
->flags
|= HAMMER_CURSOR_ATEMEM
;
1511 /* fall through to the memory entry */
1512 case HAMMER_CURSOR_ATEDISK
:
1514 * Only the memory entry is valid.
1516 cursor
->leaf
= &cursor
->iprec
->leaf
;
1517 cursor
->flags
|= HAMMER_CURSOR_ATEMEM
;
1520 * If the memory entry is an on-disk deletion we should have
1521 * also had found a B-Tree record. If the backend beat us
1522 * to it it would have interlocked the cursor and we should
1523 * have seen the in-memory record marked DELETED_FE.
1525 if (cursor
->iprec
->type
== HAMMER_MEM_RECORD_DEL
&&
1526 (cursor
->flags
& HAMMER_CURSOR_DELETE_VISIBILITY
) == 0) {
1527 panic("hammer_ip_next: del-on-disk with no b-tree entry");
1530 case HAMMER_CURSOR_ATEMEM
:
1532 * Only the disk entry is valid
1534 error
= hammer_btree_extract(cursor
, HAMMER_CURSOR_GET_LEAF
);
1535 cursor
->flags
|= HAMMER_CURSOR_ATEDISK
;
1539 * Neither entry is valid
1541 * XXX error not set properly
1543 cursor
->leaf
= NULL
;
1551 * Resolve the cursor->data pointer for the current cursor position in
1552 * a merged iteration.
1555 hammer_ip_resolve_data(hammer_cursor_t cursor
)
1557 hammer_record_t record
;
1560 if (hammer_cursor_inmem(cursor
)) {
1562 * The data associated with an in-memory record is usually
1563 * kmalloced, but reserve-ahead data records will have an
1564 * on-disk reference.
1566 * NOTE: Reserve-ahead data records must be handled in the
1567 * context of the related high level buffer cache buffer
1568 * to interlock against async writes.
1570 record
= cursor
->iprec
;
1571 cursor
->data
= record
->data
;
1573 if (cursor
->data
== NULL
) {
1574 KKASSERT(record
->leaf
.base
.rec_type
==
1575 HAMMER_RECTYPE_DATA
);
1576 cursor
->data
= hammer_bread_ext(cursor
->trans
->hmp
,
1577 record
->leaf
.data_offset
,
1578 record
->leaf
.data_len
,
1580 &cursor
->data_buffer
);
1583 cursor
->leaf
= &cursor
->node
->ondisk
->elms
[cursor
->index
].leaf
;
1584 error
= hammer_btree_extract(cursor
, HAMMER_CURSOR_GET_DATA
);
1590 * Backend truncation / record replacement - delete records in range.
1592 * Delete all records within the specified range for inode ip. In-memory
1593 * records still associated with the frontend are ignored.
1595 * If truncating is non-zero in-memory records associated with the back-end
1596 * are ignored. If truncating is > 1 we can return EWOULDBLOCK.
1600 * * An unaligned range will cause new records to be added to cover
1601 * the edge cases. (XXX not implemented yet).
1603 * * Replacement via reservations (see hammer_ip_sync_record_cursor())
1604 * also do not deal with unaligned ranges.
1606 * * ran_end is inclusive (e.g. 0,1023 instead of 0,1024).
1608 * * Record keys for regular file data have to be special-cased since
1609 * they indicate the end of the range (key = base + bytes).
1611 * * This function may be asked to delete ridiculously huge ranges, for
1612 * example if someone truncates or removes a 1TB regular file. We
1613 * must be very careful on restarts and we may have to stop w/
1614 * EWOULDBLOCK to avoid blowing out the buffer cache.
1617 hammer_ip_delete_range(hammer_cursor_t cursor
, hammer_inode_t ip
,
1618 int64_t ran_beg
, int64_t ran_end
, int truncating
)
1620 hammer_transaction_t trans
= cursor
->trans
;
1621 hammer_btree_leaf_elm_t leaf
;
1627 kprintf("delete_range %p %016llx-%016llx\n", ip
, ran_beg
, ran_end
);
1630 KKASSERT(trans
->type
== HAMMER_TRANS_FLS
);
1632 hammer_normalize_cursor(cursor
);
1633 cursor
->key_beg
.localization
= ip
->obj_localization
+
1634 HAMMER_LOCALIZE_MISC
;
1635 cursor
->key_beg
.obj_id
= ip
->obj_id
;
1636 cursor
->key_beg
.create_tid
= 0;
1637 cursor
->key_beg
.delete_tid
= 0;
1638 cursor
->key_beg
.obj_type
= 0;
1640 if (ip
->ino_data
.obj_type
== HAMMER_OBJTYPE_DBFILE
) {
1641 cursor
->key_beg
.key
= ran_beg
;
1642 cursor
->key_beg
.rec_type
= HAMMER_RECTYPE_DB
;
1645 * The key in the B-Tree is (base+bytes), so the first possible
1646 * matching key is ran_beg + 1.
1648 cursor
->key_beg
.key
= ran_beg
+ 1;
1649 cursor
->key_beg
.rec_type
= HAMMER_RECTYPE_DATA
;
1652 cursor
->key_end
= cursor
->key_beg
;
1653 if (ip
->ino_data
.obj_type
== HAMMER_OBJTYPE_DBFILE
) {
1654 cursor
->key_end
.key
= ran_end
;
1656 tmp64
= ran_end
+ MAXPHYS
+ 1; /* work around GCC-4 bug */
1657 if (tmp64
< ran_end
)
1658 cursor
->key_end
.key
= 0x7FFFFFFFFFFFFFFFLL
;
1660 cursor
->key_end
.key
= ran_end
+ MAXPHYS
+ 1;
1663 cursor
->asof
= ip
->obj_asof
;
1664 cursor
->flags
&= ~HAMMER_CURSOR_INITMASK
;
1665 cursor
->flags
|= HAMMER_CURSOR_ASOF
;
1666 cursor
->flags
|= HAMMER_CURSOR_DELETE_VISIBILITY
;
1667 cursor
->flags
|= HAMMER_CURSOR_BACKEND
;
1668 cursor
->flags
|= HAMMER_CURSOR_END_INCLUSIVE
;
1670 error
= hammer_ip_first(cursor
);
1673 * Iterate through matching records and mark them as deleted.
1675 while (error
== 0) {
1676 leaf
= cursor
->leaf
;
1678 KKASSERT(leaf
->base
.delete_tid
== 0);
1681 * There may be overlap cases for regular file data. Also
1682 * remember the key for a regular file record is (base + len),
1685 if (leaf
->base
.rec_type
== HAMMER_RECTYPE_DATA
) {
1686 off
= leaf
->base
.key
- leaf
->data_len
;
1688 * Check the left edge case. We currently do not
1689 * split existing records.
1691 if (off
< ran_beg
) {
1692 panic("hammer left edge case %016llx %d\n",
1693 leaf
->base
.key
, leaf
->data_len
);
1697 * Check the right edge case. Note that the
1698 * record can be completely out of bounds, which
1699 * terminates the search.
1701 * base->key is exclusive of the right edge while
1702 * ran_end is inclusive of the right edge. The
1703 * (key - data_len) left boundary is inclusive.
1705 * XXX theory-check this test at some point, are
1706 * we missing a + 1 somewhere? Note that ran_end
1709 if (leaf
->base
.key
- 1 > ran_end
) {
1710 if (leaf
->base
.key
- leaf
->data_len
> ran_end
)
1712 panic("hammer right edge case\n");
1715 off
= leaf
->base
.key
;
1719 * Delete the record. When truncating we do not delete
1720 * in-memory (data) records because they represent data
1721 * written after the truncation.
1723 * This will also physically destroy the B-Tree entry and
1724 * data if the retention policy dictates. The function
1725 * will set HAMMER_CURSOR_DELBTREE which hammer_ip_next()
1726 * uses to perform a fixup.
1728 if (truncating
== 0 || hammer_cursor_ondisk(cursor
)) {
1729 error
= hammer_ip_delete_record(cursor
, ip
, trans
->tid
);
1731 * If we have built up too many meta-buffers we risk
1732 * deadlocking the kernel and must stop. This can
1733 * occur when deleting ridiculously huge files.
1734 * sync_trunc_off is updated so the next cycle does
1735 * not re-iterate records we have already deleted.
1737 * This is only done with formal truncations.
1739 if (truncating
> 1 && error
== 0 &&
1740 hammer_flusher_meta_limit(ip
->hmp
)) {
1741 ip
->sync_trunc_off
= off
;
1742 error
= EWOULDBLOCK
;
1747 ran_beg
= off
; /* for restart */
1748 error
= hammer_ip_next(cursor
);
1751 hammer_cache_node(&ip
->cache
[1], cursor
->node
);
1753 if (error
== EDEADLK
) {
1754 hammer_done_cursor(cursor
);
1755 error
= hammer_init_cursor(trans
, cursor
, &ip
->cache
[1], ip
);
1759 if (error
== ENOENT
)
1765 * This backend function deletes the specified record on-disk, similar to
1766 * delete_range but for a specific record. Unlike the exact deletions
1767 * used when deleting a directory entry this function uses an ASOF search
1768 * like delete_range.
1770 * This function may be called with ip->obj_asof set for a slave snapshot,
1771 * so don't use it. We always delete non-historical records only.
1774 hammer_delete_general(hammer_cursor_t cursor
, hammer_inode_t ip
,
1775 hammer_btree_leaf_elm_t leaf
)
1777 hammer_transaction_t trans
= cursor
->trans
;
1780 KKASSERT(trans
->type
== HAMMER_TRANS_FLS
);
1782 hammer_normalize_cursor(cursor
);
1783 cursor
->key_beg
= leaf
->base
;
1784 cursor
->asof
= HAMMER_MAX_TID
;
1785 cursor
->flags
&= ~HAMMER_CURSOR_INITMASK
;
1786 cursor
->flags
|= HAMMER_CURSOR_ASOF
;
1787 cursor
->flags
|= HAMMER_CURSOR_BACKEND
;
1788 cursor
->flags
&= ~HAMMER_CURSOR_INSERT
;
1790 error
= hammer_btree_lookup(cursor
);
1792 error
= hammer_ip_delete_record(cursor
, ip
, trans
->tid
);
1794 if (error
== EDEADLK
) {
1795 hammer_done_cursor(cursor
);
1796 error
= hammer_init_cursor(trans
, cursor
, &ip
->cache
[1], ip
);
1804 * This function deletes remaining auxillary records when an inode is
1805 * being deleted. This function explicitly does not delete the
1806 * inode record, directory entry, data, or db records. Those must be
1807 * properly disposed of prior to this call.
1810 hammer_ip_delete_clean(hammer_cursor_t cursor
, hammer_inode_t ip
, int *countp
)
1812 hammer_transaction_t trans
= cursor
->trans
;
1813 hammer_btree_leaf_elm_t leaf
;
1816 KKASSERT(trans
->type
== HAMMER_TRANS_FLS
);
1818 hammer_normalize_cursor(cursor
);
1819 cursor
->key_beg
.localization
= ip
->obj_localization
+
1820 HAMMER_LOCALIZE_MISC
;
1821 cursor
->key_beg
.obj_id
= ip
->obj_id
;
1822 cursor
->key_beg
.create_tid
= 0;
1823 cursor
->key_beg
.delete_tid
= 0;
1824 cursor
->key_beg
.obj_type
= 0;
1825 cursor
->key_beg
.rec_type
= HAMMER_RECTYPE_CLEAN_START
;
1826 cursor
->key_beg
.key
= HAMMER_MIN_KEY
;
1828 cursor
->key_end
= cursor
->key_beg
;
1829 cursor
->key_end
.rec_type
= HAMMER_RECTYPE_MAX
;
1830 cursor
->key_end
.key
= HAMMER_MAX_KEY
;
1832 cursor
->asof
= ip
->obj_asof
;
1833 cursor
->flags
&= ~HAMMER_CURSOR_INITMASK
;
1834 cursor
->flags
|= HAMMER_CURSOR_END_INCLUSIVE
| HAMMER_CURSOR_ASOF
;
1835 cursor
->flags
|= HAMMER_CURSOR_DELETE_VISIBILITY
;
1836 cursor
->flags
|= HAMMER_CURSOR_BACKEND
;
1838 error
= hammer_ip_first(cursor
);
1841 * Iterate through matching records and mark them as deleted.
1843 while (error
== 0) {
1844 leaf
= cursor
->leaf
;
1846 KKASSERT(leaf
->base
.delete_tid
== 0);
1849 * Mark the record and B-Tree entry as deleted. This will
1850 * also physically delete the B-Tree entry, record, and
1851 * data if the retention policy dictates. The function
1852 * will set HAMMER_CURSOR_DELBTREE which hammer_ip_next()
1853 * uses to perform a fixup.
1855 * Directory entries (and delete-on-disk directory entries)
1856 * must be synced and cannot be deleted.
1858 error
= hammer_ip_delete_record(cursor
, ip
, trans
->tid
);
1862 error
= hammer_ip_next(cursor
);
1865 hammer_cache_node(&ip
->cache
[1], cursor
->node
);
1866 if (error
== EDEADLK
) {
1867 hammer_done_cursor(cursor
);
1868 error
= hammer_init_cursor(trans
, cursor
, &ip
->cache
[1], ip
);
1872 if (error
== ENOENT
)
1878 * Delete the record at the current cursor. On success the cursor will
1879 * be positioned appropriately for an iteration but may no longer be at
1882 * This routine is only called from the backend.
1884 * NOTE: This can return EDEADLK, requiring the caller to terminate the
1888 hammer_ip_delete_record(hammer_cursor_t cursor
, hammer_inode_t ip
,
1891 hammer_off_t zone2_offset
;
1892 hammer_record_t iprec
;
1893 hammer_btree_elm_t elm
;
1898 KKASSERT(cursor
->flags
& HAMMER_CURSOR_BACKEND
);
1900 hmp
= cursor
->node
->hmp
;
1903 * In-memory (unsynchronized) records can simply be freed. This
1904 * only occurs in range iterations since all other records are
1905 * individually synchronized. Thus there should be no confusion with
1908 * An in-memory record may be deleted before being committed to disk,
1909 * but could have been accessed in the mean time. The backing store
1910 * may never been marked allocated and so hammer_blockmap_free() may
1911 * never get called on it. Because of this we have to make sure that
1912 * we've gotten rid of any related hammer_buffer or buffer cache
1915 if (hammer_cursor_inmem(cursor
)) {
1916 iprec
= cursor
->iprec
;
1917 KKASSERT((iprec
->flags
& HAMMER_RECF_INTERLOCK_BE
) ==0);
1918 iprec
->flags
|= HAMMER_RECF_DELETED_FE
;
1919 iprec
->flags
|= HAMMER_RECF_DELETED_BE
;
1921 if (iprec
->leaf
.data_offset
&& iprec
->leaf
.data_len
) {
1922 zone2_offset
= hammer_blockmap_lookup(hmp
, iprec
->leaf
.data_offset
, &error
);
1923 KKASSERT(error
== 0);
1924 hammer_del_buffers(hmp
,
1925 iprec
->leaf
.data_offset
,
1927 iprec
->leaf
.data_len
);
1933 * On-disk records are marked as deleted by updating their delete_tid.
1934 * This does not effect their position in the B-Tree (which is based
1935 * on their create_tid).
1937 error
= hammer_btree_extract(cursor
, HAMMER_CURSOR_GET_LEAF
);
1941 * If we were mounted with the nohistory option, we physically
1942 * delete the record.
1944 dodelete
= hammer_nohistory(ip
);
1947 error
= hammer_cursor_upgrade(cursor
);
1949 elm
= &cursor
->node
->ondisk
->elms
[cursor
->index
];
1950 hammer_modify_node(cursor
->trans
, cursor
->node
,
1952 elm
->leaf
.base
.delete_tid
= tid
;
1953 elm
->leaf
.delete_ts
= cursor
->trans
->time32
;
1954 hammer_modify_node_done(cursor
->node
);
1957 * An on-disk record cannot have the same delete_tid
1958 * as its create_tid. In a chain of record updates
1959 * this could result in a duplicate record.
1961 KKASSERT(elm
->leaf
.base
.delete_tid
!= elm
->leaf
.base
.create_tid
);
1965 if (error
== 0 && dodelete
) {
1966 error
= hammer_delete_at_cursor(cursor
, NULL
);
1968 panic("hammer_ip_delete_record: unable to physically delete the record!\n");
1976 hammer_delete_at_cursor(hammer_cursor_t cursor
, int64_t *stat_bytes
)
1978 hammer_btree_elm_t elm
;
1979 hammer_off_t data_offset
;
1984 elm
= &cursor
->node
->ondisk
->elms
[cursor
->index
];
1985 KKASSERT(elm
->base
.btype
== HAMMER_BTREE_TYPE_RECORD
);
1987 data_offset
= elm
->leaf
.data_offset
;
1988 data_len
= elm
->leaf
.data_len
;
1989 rec_type
= elm
->leaf
.base
.rec_type
;
1991 error
= hammer_btree_delete(cursor
);
1994 * This forces a fixup for the iteration because
1995 * the cursor is now either sitting at the 'next'
1996 * element or sitting at the end of a leaf.
1998 if ((cursor
->flags
& HAMMER_CURSOR_DISKEOF
) == 0) {
1999 cursor
->flags
|= HAMMER_CURSOR_DELBTREE
;
2000 cursor
->flags
&= ~HAMMER_CURSOR_ATEDISK
;
2004 switch(data_offset
& HAMMER_OFF_ZONE_MASK
) {
2005 case HAMMER_ZONE_LARGE_DATA
:
2006 case HAMMER_ZONE_SMALL_DATA
:
2007 case HAMMER_ZONE_META
:
2008 hammer_blockmap_free(cursor
->trans
,
2009 data_offset
, data_len
);
2019 * Determine whether we can remove a directory. This routine checks whether
2020 * a directory is empty or not and enforces flush connectivity.
2022 * Flush connectivity requires that we block if the target directory is
2023 * currently flushing, otherwise it may not end up in the same flush group.
2025 * Returns 0 on success, ENOTEMPTY or EDEADLK (or other errors) on failure.
2028 hammer_ip_check_directory_empty(hammer_transaction_t trans
, hammer_inode_t ip
)
2030 struct hammer_cursor cursor
;
2034 * Check directory empty
2036 hammer_init_cursor(trans
, &cursor
, &ip
->cache
[1], ip
);
2038 cursor
.key_beg
.localization
= ip
->obj_localization
+
2039 HAMMER_LOCALIZE_MISC
;
2040 cursor
.key_beg
.obj_id
= ip
->obj_id
;
2041 cursor
.key_beg
.create_tid
= 0;
2042 cursor
.key_beg
.delete_tid
= 0;
2043 cursor
.key_beg
.obj_type
= 0;
2044 cursor
.key_beg
.rec_type
= HAMMER_RECTYPE_INODE
+ 1;
2045 cursor
.key_beg
.key
= HAMMER_MIN_KEY
;
2047 cursor
.key_end
= cursor
.key_beg
;
2048 cursor
.key_end
.rec_type
= 0xFFFF;
2049 cursor
.key_end
.key
= HAMMER_MAX_KEY
;
2051 cursor
.asof
= ip
->obj_asof
;
2052 cursor
.flags
|= HAMMER_CURSOR_END_INCLUSIVE
| HAMMER_CURSOR_ASOF
;
2054 error
= hammer_ip_first(&cursor
);
2055 if (error
== ENOENT
)
2057 else if (error
== 0)
2059 hammer_done_cursor(&cursor
);