2 * Copyright (c) 2011-2015 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@dragonflybsd.org>
6 * and Venkatesh Srinivas <vsrinivas@dragonflybsd.org>
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in
16 * the documentation and/or other materials provided with the
18 * 3. Neither the name of The DragonFly Project nor the names of its
19 * contributors may be used to endorse or promote products derived
20 * from this software without specific, prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
23 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
24 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
25 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
26 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
27 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
28 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
29 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
30 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
31 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
32 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * This subsystem implements most of the core support functions for
37 * the hammer2_chain structure.
39 * Chains are the in-memory version on media objects (volume header, inodes,
40 * indirect blocks, data blocks, etc). Chains represent a portion of the
43 * Chains are no-longer delete-duplicated. Instead, the original in-memory
44 * chain will be moved along with its block reference (e.g. for things like
45 * renames, hardlink operations, modifications, etc), and will be indexed
46 * on a secondary list for flush handling instead of propagating a flag
49 * Concurrent front-end operations can still run against backend flushes
50 * as long as they do not cross the current flush boundary. An operation
51 * running above the current flush (in areas not yet flushed) can become
52 * part of the current flush while ano peration running below the current
53 * flush can become part of the next flush.
55 #include <sys/cdefs.h>
56 #include <sys/param.h>
57 #include <sys/systm.h>
58 #include <sys/types.h>
60 #include <sys/kern_syscall.h>
63 #include <crypto/sha2/sha2.h>
67 static hammer2_chain_t
*hammer2_chain_create_indirect(
68 hammer2_chain_t
*parent
,
69 hammer2_key_t key
, int keybits
,
70 hammer2_tid_t mtid
, int for_type
, int *errorp
);
71 static hammer2_io_t
*hammer2_chain_drop_data(hammer2_chain_t
*chain
);
72 static hammer2_chain_t
*hammer2_combined_find(
73 hammer2_chain_t
*parent
,
74 hammer2_blockref_t
*base
, int count
,
75 hammer2_key_t
*key_nextp
,
76 hammer2_key_t key_beg
, hammer2_key_t key_end
,
77 hammer2_blockref_t
**bresp
);
80 * Basic RBTree for chains (core->rbtree and core->dbtree). Chains cannot
81 * overlap in the RB trees. Deleted chains are moved from rbtree to either
84 * Chains in delete-duplicate sequences can always iterate through core_entry
85 * to locate the live version of the chain.
87 RB_GENERATE(hammer2_chain_tree
, hammer2_chain
, rbnode
, hammer2_chain_cmp
);
90 hammer2_chain_cmp(hammer2_chain_t
*chain1
, hammer2_chain_t
*chain2
)
98 * Compare chains. Overlaps are not supposed to happen and catch
99 * any software issues early we count overlaps as a match.
101 c1_beg
= chain1
->bref
.key
;
102 c1_end
= c1_beg
+ ((hammer2_key_t
)1 << chain1
->bref
.keybits
) - 1;
103 c2_beg
= chain2
->bref
.key
;
104 c2_end
= c2_beg
+ ((hammer2_key_t
)1 << chain2
->bref
.keybits
) - 1;
106 if (c1_end
< c2_beg
) /* fully to the left */
108 if (c1_beg
> c2_end
) /* fully to the right */
110 return(0); /* overlap (must not cross edge boundary) */
114 * Assert that a chain has no media data associated with it.
117 hammer2_chain_assert_no_data(hammer2_chain_t
*chain
)
119 KKASSERT(chain
->dio
== NULL
);
120 if (chain
->bref
.type
!= HAMMER2_BREF_TYPE_VOLUME
&&
121 chain
->bref
.type
!= HAMMER2_BREF_TYPE_FREEMAP
&&
123 panic("hammer2_assert_no_data: chain %p still has data", chain
);
128 * Make a chain visible to the flusher. The flusher needs to be able to
129 * do flushes of subdirectory chains or single files so it does a top-down
130 * recursion using the ONFLUSH flag for the recursion. It locates MODIFIED
131 * or UPDATE chains and flushes back up the chain to the volume root.
133 * This routine sets ONFLUSH upward until it hits the volume root. For
134 * simplicity we ignore PFSROOT boundaries whos rules can be complex.
135 * Extra ONFLUSH flagging doesn't hurt the filesystem.
138 hammer2_chain_setflush(hammer2_chain_t
*chain
)
140 hammer2_chain_t
*parent
;
142 if ((chain
->flags
& HAMMER2_CHAIN_ONFLUSH
) == 0) {
143 hammer2_spin_sh(&chain
->core
.spin
);
144 while ((chain
->flags
& HAMMER2_CHAIN_ONFLUSH
) == 0) {
145 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_ONFLUSH
);
146 if ((parent
= chain
->parent
) == NULL
)
148 hammer2_spin_sh(&parent
->core
.spin
);
149 hammer2_spin_unsh(&chain
->core
.spin
);
152 hammer2_spin_unsh(&chain
->core
.spin
);
157 * Allocate a new disconnected chain element representing the specified
158 * bref. chain->refs is set to 1 and the passed bref is copied to
159 * chain->bref. chain->bytes is derived from the bref.
161 * chain->pmp inherits pmp unless the chain is an inode (other than the
164 * NOTE: Returns a referenced but unlocked (because there is no core) chain.
167 hammer2_chain_alloc(hammer2_dev_t
*hmp
, hammer2_pfs_t
*pmp
,
168 hammer2_blockref_t
*bref
)
170 hammer2_chain_t
*chain
;
174 * Special case - radix of 0 indicates a chain that does not
175 * need a data reference (context is completely embedded in the
178 if ((int)(bref
->data_off
& HAMMER2_OFF_MASK_RADIX
))
179 bytes
= 1U << (int)(bref
->data_off
& HAMMER2_OFF_MASK_RADIX
);
183 atomic_add_long(&hammer2_chain_allocs
, 1);
186 * Construct the appropriate system structure.
189 case HAMMER2_BREF_TYPE_DIRENT
:
190 case HAMMER2_BREF_TYPE_INODE
:
191 case HAMMER2_BREF_TYPE_INDIRECT
:
192 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
193 case HAMMER2_BREF_TYPE_DATA
:
194 case HAMMER2_BREF_TYPE_FREEMAP_LEAF
:
196 * Chain's are really only associated with the hmp but we
197 * maintain a pmp association for per-mount memory tracking
198 * purposes. The pmp can be NULL.
200 chain
= kmalloc(sizeof(*chain
), hmp
->mchain
, M_WAITOK
| M_ZERO
);
202 case HAMMER2_BREF_TYPE_VOLUME
:
203 case HAMMER2_BREF_TYPE_FREEMAP
:
205 * Only hammer2_chain_bulksnap() calls this function with these
208 chain
= kmalloc(sizeof(*chain
), hmp
->mchain
, M_WAITOK
| M_ZERO
);
212 panic("hammer2_chain_alloc: unrecognized blockref type: %d",
217 * Initialize the new chain structure. pmp must be set to NULL for
218 * chains belonging to the super-root topology of a device mount.
220 if (pmp
== hmp
->spmp
)
226 chain
->bytes
= bytes
;
228 chain
->flags
= HAMMER2_CHAIN_ALLOCATED
;
231 * Set the PFS boundary flag if this chain represents a PFS root.
233 if (bref
->flags
& HAMMER2_BREF_FLAG_PFSROOT
)
234 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_PFSBOUNDARY
);
235 hammer2_chain_core_init(chain
);
241 * Initialize a chain's core structure. This structure used to be allocated
242 * but is now embedded.
244 * The core is not locked. No additional refs on the chain are made.
245 * (trans) must not be NULL if (core) is not NULL.
248 hammer2_chain_core_init(hammer2_chain_t
*chain
)
251 * Fresh core under nchain (no multi-homing of ochain's
254 RB_INIT(&chain
->core
.rbtree
); /* live chains */
255 hammer2_mtx_init(&chain
->lock
, "h2chain");
259 * Add a reference to a chain element, preventing its destruction.
261 * (can be called with spinlock held)
264 hammer2_chain_ref(hammer2_chain_t
*chain
)
266 if (atomic_fetchadd_int(&chain
->refs
, 1) == 0) {
268 * 0->non-zero transition must ensure that chain is removed
271 * NOTE: Already holding lru_spin here so we cannot call
272 * hammer2_chain_ref() to get it off lru_list, do
275 if (chain
->flags
& HAMMER2_CHAIN_ONLRU
) {
276 hammer2_pfs_t
*pmp
= chain
->pmp
;
277 hammer2_spin_ex(&pmp
->lru_spin
);
278 if (chain
->flags
& HAMMER2_CHAIN_ONLRU
) {
279 atomic_add_int(&pmp
->lru_count
, -1);
280 atomic_clear_int(&chain
->flags
,
281 HAMMER2_CHAIN_ONLRU
);
282 TAILQ_REMOVE(&pmp
->lru_list
, chain
, lru_node
);
284 hammer2_spin_unex(&pmp
->lru_spin
);
288 kprintf("REFC %p %d %08x\n", chain
, chain
->refs
- 1, chain
->flags
);
294 * Ref a locked chain and force the data to be held across an unlock.
295 * Chain must be currently locked. The user of the chain who desires
296 * to release the hold must call hammer2_chain_lock_unhold() to relock
297 * and unhold the chain, then unlock normally, or may simply call
298 * hammer2_chain_drop_unhold() (which is safer against deadlocks).
301 hammer2_chain_ref_hold(hammer2_chain_t
*chain
)
303 atomic_add_int(&chain
->lockcnt
, 1);
304 hammer2_chain_ref(chain
);
308 * Insert the chain in the core rbtree.
310 * Normal insertions are placed in the live rbtree. Insertion of a deleted
311 * chain is a special case used by the flush code that is placed on the
312 * unstaged deleted list to avoid confusing the live view.
314 #define HAMMER2_CHAIN_INSERT_SPIN 0x0001
315 #define HAMMER2_CHAIN_INSERT_LIVE 0x0002
316 #define HAMMER2_CHAIN_INSERT_RACE 0x0004
320 hammer2_chain_insert(hammer2_chain_t
*parent
, hammer2_chain_t
*chain
,
321 int flags
, int generation
)
323 hammer2_chain_t
*xchain
;
326 if (flags
& HAMMER2_CHAIN_INSERT_SPIN
)
327 hammer2_spin_ex(&parent
->core
.spin
);
330 * Interlocked by spinlock, check for race
332 if ((flags
& HAMMER2_CHAIN_INSERT_RACE
) &&
333 parent
->core
.generation
!= generation
) {
341 xchain
= RB_INSERT(hammer2_chain_tree
, &parent
->core
.rbtree
, chain
);
342 KASSERT(xchain
== NULL
,
343 ("hammer2_chain_insert: collision %p %p (key=%016jx)",
344 chain
, xchain
, chain
->bref
.key
));
345 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_ONRBTREE
);
346 chain
->parent
= parent
;
347 ++parent
->core
.chain_count
;
348 ++parent
->core
.generation
; /* XXX incs for _get() too, XXX */
351 * We have to keep track of the effective live-view blockref count
352 * so the create code knows when to push an indirect block.
354 if (flags
& HAMMER2_CHAIN_INSERT_LIVE
)
355 atomic_add_int(&parent
->core
.live_count
, 1);
357 if (flags
& HAMMER2_CHAIN_INSERT_SPIN
)
358 hammer2_spin_unex(&parent
->core
.spin
);
363 * Drop the caller's reference to the chain. When the ref count drops to
364 * zero this function will try to disassociate the chain from its parent and
365 * deallocate it, then recursely drop the parent using the implied ref
366 * from the chain's chain->parent.
368 * Nobody should own chain's mutex on the 1->0 transition, unless this drop
369 * races an acquisition by another cpu. Therefore we can loop if we are
370 * unable to acquire the mutex, and refs is unlikely to be 1 unless we again
371 * race against another drop.
373 static hammer2_chain_t
*hammer2_chain_lastdrop(hammer2_chain_t
*chain
);
376 hammer2_chain_drop(hammer2_chain_t
*chain
)
380 if (hammer2_debug
& 0x200000)
383 kprintf("DROP %p %d %08x\n", chain
, chain
->refs
- 1, chain
->flags
);
387 KKASSERT(chain
->refs
> 0);
395 if (mtx_lock_ex_try(&chain
->lock
) == 0)
396 chain
= hammer2_chain_lastdrop(chain
);
397 /* retry the same chain */
399 if (atomic_cmpset_int(&chain
->refs
, refs
, refs
- 1))
401 /* retry the same chain */
408 * Unhold a held and probably not-locked chain, ensure that the data is
409 * dropped on the 1->0 transition of lockcnt by obtaining an exclusive
410 * lock and then simply unlocking the chain.
413 hammer2_chain_drop_unhold(hammer2_chain_t
*chain
)
419 lockcnt
= chain
->lockcnt
;
422 if (atomic_cmpset_int(&chain
->lockcnt
,
423 lockcnt
, lockcnt
- 1)) {
426 } else if (mtx_lock_ex_try(&chain
->lock
) == 0) {
427 hammer2_chain_unlock(chain
);
431 * This situation can easily occur on SMP due to
432 * the gap inbetween the 1->0 transition and the
433 * final unlock. We cannot safely block on the
434 * mutex because lockcnt might go above 1.
436 * XXX Sleep for one tick if it takes too long.
439 if (iter
> 1000 + hz
) {
440 kprintf("hammer2: h2race1 %p\n", chain
);
443 tsleep(&iter
, 0, "h2race1", 1);
448 hammer2_chain_drop(chain
);
452 * Handles the (potential) last drop of chain->refs from 1->0. Called with
453 * the mutex exclusively locked, refs == 1, and lockcnt 0. SMP races are
454 * possible against refs and lockcnt. We must dispose of the mutex on chain.
456 * This function returns an unlocked chain for recursive drop or NULL. It
457 * can return the same chain if it determines it has raced another ref.
461 * When two chains need to be recursively dropped we use the chain we
462 * would otherwise free to placehold the additional chain. It's a bit
463 * convoluted but we can't just recurse without potentially blowing out
466 * The chain cannot be freed if it has any children.
467 * The chain cannot be freed if flagged MODIFIED unless we can dispose of it.
468 * The chain cannot be freed if flagged UPDATE unless we can dispose of it.
469 * Any dedup registration can remain intact.
471 * The core spinlock is allowed to nest child-to-parent (not parent-to-child).
475 hammer2_chain_lastdrop(hammer2_chain_t
*chain
)
479 hammer2_chain_t
*parent
;
480 hammer2_chain_t
*rdrop
;
487 * On last drop if there is no parent and data_off is good (at
488 * least does not represent the volume root), the modified chain
489 * is probably going to be destroyed. We have to make sure that
490 * the data area is not registered for dedup.
492 * XXX removed. In fact, we do not have to make sure that the
493 * data area is not registered for dedup. The data area
494 * can, in fact, still be used for dedup because it is
495 * still allocated in the freemap and the underlying I/O
496 * will still be flushed.
498 if (chain
->parent
== NULL
&&
499 (chain
->flags
& HAMMER2_CHAIN_MODIFIED
) &&
500 (chain
->bref
.data_off
& ~HAMMER2_OFF_MASK_RADIX
)) {
502 hammer2_io_dedup_delete(hmp
, chain
->bref
.type
,
503 chain
->bref
.data_off
, chain
->bytes
);
507 * We need chain's spinlock to interlock the sub-tree test.
508 * We already have chain's mutex, protecting chain->parent.
510 * Remember that chain->refs can be in flux.
512 hammer2_spin_ex(&chain
->core
.spin
);
514 if ((parent
= chain
->parent
) != NULL
) {
516 * If the chain has a parent the UPDATE bit prevents scrapping
517 * as the chain is needed to properly flush the parent. Try
518 * to complete the 1->0 transition and return NULL. Retry
519 * (return chain) if we are unable to complete the 1->0
520 * transition, else return NULL (nothing more to do).
522 * If the chain has a parent the MODIFIED bit prevents
525 * Chains with UPDATE/MODIFIED are *not* put on the LRU list!
527 if (chain
->flags
& (HAMMER2_CHAIN_UPDATE
|
528 HAMMER2_CHAIN_MODIFIED
)) {
529 if (atomic_cmpset_int(&chain
->refs
, 1, 0)) {
530 hammer2_spin_unex(&chain
->core
.spin
);
532 dio
= hammer2_chain_drop_data(chain
, 0);
534 hammer2_io_bqrelse(&dio
);
536 hammer2_chain_assert_no_data(chain
);
537 hammer2_mtx_unlock(&chain
->lock
);
540 hammer2_spin_unex(&chain
->core
.spin
);
541 hammer2_mtx_unlock(&chain
->lock
);
545 /* spinlock still held */
548 * The chain has no parent and can be flagged for destruction.
549 * Since it has no parent, UPDATE can also be cleared.
551 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_DESTROY
);
552 if (chain
->flags
& HAMMER2_CHAIN_UPDATE
)
553 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_UPDATE
);
556 * If the chain has children we must still flush the chain.
557 * Any dedup is already handled by the underlying DIO, so
558 * we do not have to specifically flush it here.
560 * In the case where it has children, the DESTROY flag test
561 * in the flush code will prevent unnecessary flushes of
562 * MODIFIED chains that are not flagged DEDUP so don't worry
565 if (chain
->core
.chain_count
) {
567 * Put on flushq (should ensure refs > 1), retry
570 hammer2_spin_unex(&chain
->core
.spin
);
571 hammer2_delayed_flush(chain
);
572 hammer2_mtx_unlock(&chain
->lock
);
574 return(chain
); /* retry drop */
578 * Otherwise we can scrap the MODIFIED bit if it is set,
579 * and continue along the freeing path.
581 * Be sure to clean-out any dedup bits. Without a parent
582 * this chain will no longer be visible to the flush code.
583 * Easy check data_off to avoid the volume root.
585 if (chain
->flags
& HAMMER2_CHAIN_MODIFIED
) {
586 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_MODIFIED
);
587 atomic_add_long(&hammer2_count_modified_chains
, -1);
589 hammer2_pfs_memory_wakeup(chain
->pmp
);
591 /* spinlock still held */
594 /* spinlock still held */
600 * If any children exist we must leave the chain intact with refs == 0.
601 * They exist because chains are retained below us which have refs or
602 * may require flushing.
604 * Retry (return chain) if we fail to transition the refs to 0, else
605 * return NULL indication nothing more to do.
607 * Chains with children are NOT put on the LRU list.
609 if (chain
->core
.chain_count
) {
610 if (atomic_cmpset_int(&chain
->refs
, 1, 0)) {
611 hammer2_spin_unex(&chain
->core
.spin
);
612 hammer2_chain_assert_no_data(chain
);
613 hammer2_mtx_unlock(&chain
->lock
);
616 hammer2_spin_unex(&chain
->core
.spin
);
617 hammer2_mtx_unlock(&chain
->lock
);
621 /* spinlock still held */
622 /* no chains left under us */
625 * chain->core has no children left so no accessors can get to our
626 * chain from there. Now we have to lock the parent core to interlock
627 * remaining possible accessors that might bump chain's refs before
628 * we can safely drop chain's refs with intent to free the chain.
631 pmp
= chain
->pmp
; /* can be NULL */
634 parent
= chain
->parent
;
637 * WARNING! chain's spin lock is still held here, and other spinlocks
638 * will be acquired and released in the code below. We
639 * cannot be making fancy procedure calls!
643 * We can cache the chain if it is associated with a pmp
644 * and not flagged as being destroyed or requesting a full
645 * release. In this situation the chain is not removed
646 * from its parent, i.e. it can still be looked up.
648 * We intentionally do not cache DATA chains because these
649 * were likely used to load data into the logical buffer cache
650 * and will not be accessed again for some time.
653 (HAMMER2_CHAIN_DESTROY
| HAMMER2_CHAIN_RELEASE
)) == 0 &&
655 chain
->bref
.type
!= HAMMER2_BREF_TYPE_DATA
) {
657 hammer2_spin_ex(&parent
->core
.spin
);
658 if (atomic_cmpset_int(&chain
->refs
, 1, 0) == 0) {
660 * 1->0 transition failed, retry. Do not drop
661 * the chain's data yet!
664 hammer2_spin_unex(&parent
->core
.spin
);
665 hammer2_spin_unex(&chain
->core
.spin
);
666 hammer2_mtx_unlock(&chain
->lock
);
675 dio
= hammer2_chain_drop_data(chain
, 1);
677 hammer2_chain_assert_no_data(chain
);
679 KKASSERT((chain
->flags
& HAMMER2_CHAIN_ONLRU
) == 0);
680 hammer2_spin_ex(&pmp
->lru_spin
);
681 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_ONLRU
);
682 TAILQ_INSERT_TAIL(&pmp
->lru_list
, chain
, lru_node
);
685 * If we are over the LRU limit we need to drop something.
687 if (pmp
->lru_count
> HAMMER2_LRU_LIMIT
) {
688 rdrop
= TAILQ_FIRST(&pmp
->lru_list
);
689 atomic_clear_int(&rdrop
->flags
, HAMMER2_CHAIN_ONLRU
);
690 TAILQ_REMOVE(&pmp
->lru_list
, rdrop
, lru_node
);
691 atomic_add_int(&rdrop
->refs
, 1);
692 atomic_set_int(&rdrop
->flags
, HAMMER2_CHAIN_RELEASE
);
694 atomic_add_int(&pmp
->lru_count
, 1);
696 hammer2_spin_unex(&pmp
->lru_spin
);
698 hammer2_spin_unex(&parent
->core
.spin
);
699 parent
= NULL
; /* safety */
701 hammer2_spin_unex(&chain
->core
.spin
);
702 hammer2_mtx_unlock(&chain
->lock
);
705 hammer2_io_bqrelse(&dio
);
713 * Spinlock the parent and try to drop the last ref on chain.
714 * On success determine if we should dispose of the chain
715 * (remove the chain from its parent, etc).
717 * (normal core locks are top-down recursive but we define
718 * core spinlocks as bottom-up recursive, so this is safe).
721 hammer2_spin_ex(&parent
->core
.spin
);
722 if (atomic_cmpset_int(&chain
->refs
, 1, 0) == 0) {
724 /* XXX remove, don't try to drop data on fail */
725 hammer2_spin_unex(&parent
->core
.spin
);
726 dio
= hammer2_chain_drop_data(chain
, 0);
727 hammer2_spin_unex(&chain
->core
.spin
);
729 hammer2_io_bqrelse(&dio
);
732 * 1->0 transition failed, retry.
734 hammer2_spin_unex(&parent
->core
.spin
);
735 hammer2_spin_unex(&chain
->core
.spin
);
736 hammer2_mtx_unlock(&chain
->lock
);
742 * 1->0 transition successful, parent spin held to prevent
743 * new lookups, chain spinlock held to protect parent field.
744 * Remove chain from the parent.
746 if (chain
->flags
& HAMMER2_CHAIN_ONRBTREE
) {
747 RB_REMOVE(hammer2_chain_tree
,
748 &parent
->core
.rbtree
, chain
);
749 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_ONRBTREE
);
750 --parent
->core
.chain_count
;
751 chain
->parent
= NULL
;
755 * If our chain was the last chain in the parent's core the
756 * core is now empty and its parent might have to be
757 * re-dropped if it has 0 refs.
759 if (parent
->core
.chain_count
== 0) {
761 atomic_add_int(&rdrop
->refs
, 1);
763 if (atomic_cmpset_int(&rdrop->refs, 0, 1) == 0)
767 hammer2_spin_unex(&parent
->core
.spin
);
768 parent
= NULL
; /* safety */
774 if (atomic_cmpset_int(&chain
->refs
, 1, 0) == 0) {
776 * 1->0 transition failed, retry.
778 hammer2_spin_unex(&parent
->core
.spin
);
779 hammer2_spin_unex(&chain
->core
.spin
);
780 hammer2_mtx_unlock(&chain
->lock
);
787 * Successful 1->0 transition, no parent, no children... no way for
788 * anyone to ref this chain any more. We can clean-up and free it.
790 * We still have the core spinlock, and core's chain_count is 0.
791 * Any parent spinlock is gone.
793 hammer2_spin_unex(&chain
->core
.spin
);
794 hammer2_chain_assert_no_data(chain
);
795 hammer2_mtx_unlock(&chain
->lock
);
796 KKASSERT(RB_EMPTY(&chain
->core
.rbtree
) &&
797 chain
->core
.chain_count
== 0);
800 * All locks are gone, no pointers remain to the chain, finish
803 KKASSERT((chain
->flags
& (HAMMER2_CHAIN_UPDATE
|
804 HAMMER2_CHAIN_MODIFIED
)) == 0);
806 dio
= hammer2_chain_drop_data(chain
, 1);
808 hammer2_io_bqrelse(&dio
);
812 * Once chain resources are gone we can use the now dead chain
813 * structure to placehold what might otherwise require a recursive
814 * drop, because we have potentially two things to drop and can only
815 * return one directly.
817 if (chain
->flags
& HAMMER2_CHAIN_ALLOCATED
) {
818 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_ALLOCATED
);
820 kfree(chain
, hmp
->mchain
);
824 * Possible chaining loop when parent re-drop needed.
830 * On last lock release.
832 static hammer2_io_t
*
833 hammer2_chain_drop_data(hammer2_chain_t
*chain
)
837 if ((dio
= chain
->dio
) != NULL
) {
841 switch(chain
->bref
.type
) {
842 case HAMMER2_BREF_TYPE_VOLUME
:
843 case HAMMER2_BREF_TYPE_FREEMAP
:
846 if (chain
->data
!= NULL
) {
847 hammer2_spin_unex(&chain
->core
.spin
);
848 panic("chain data not null: "
849 "chain %p bref %016jx.%02x "
850 "refs %d parent %p dio %p data %p",
851 chain
, chain
->bref
.data_off
,
852 chain
->bref
.type
, chain
->refs
,
854 chain
->dio
, chain
->data
);
856 KKASSERT(chain
->data
== NULL
);
864 * Lock a referenced chain element, acquiring its data with I/O if necessary,
865 * and specify how you would like the data to be resolved.
867 * If an I/O or other fatal error occurs, chain->error will be set to non-zero.
869 * The lock is allowed to recurse, multiple locking ops will aggregate
870 * the requested resolve types. Once data is assigned it will not be
871 * removed until the last unlock.
873 * HAMMER2_RESOLVE_NEVER - Do not resolve the data element.
874 * (typically used to avoid device/logical buffer
877 * HAMMER2_RESOLVE_MAYBE - Do not resolve data elements for chains in
878 * the INITIAL-create state (indirect blocks only).
880 * Do not resolve data elements for DATA chains.
881 * (typically used to avoid device/logical buffer
884 * HAMMER2_RESOLVE_ALWAYS- Always resolve the data element.
886 * HAMMER2_RESOLVE_SHARED- (flag) The chain is locked shared, otherwise
887 * it will be locked exclusive.
889 * NOTE: Embedded elements (volume header, inodes) are always resolved
892 * NOTE: Specifying HAMMER2_RESOLVE_ALWAYS on a newly-created non-embedded
893 * element will instantiate and zero its buffer, and flush it on
896 * NOTE: (data) elements are normally locked RESOLVE_NEVER or RESOLVE_MAYBE
897 * so as not to instantiate a device buffer, which could alias against
898 * a logical file buffer. However, if ALWAYS is specified the
899 * device buffer will be instantiated anyway.
901 * WARNING! This function blocks on I/O if data needs to be fetched. This
902 * blocking can run concurrent with other compatible lock holders
903 * who do not need data returning. The lock is not upgraded to
904 * exclusive during a data fetch, a separate bit is used to
905 * interlock I/O. However, an exclusive lock holder can still count
906 * on being interlocked against an I/O fetch managed by a shared
910 hammer2_chain_lock(hammer2_chain_t
*chain
, int how
)
913 * Ref and lock the element. Recursive locks are allowed.
915 KKASSERT(chain
->refs
> 0);
916 atomic_add_int(&chain
->lockcnt
, 1);
919 * Get the appropriate lock. If LOCKAGAIN is flagged with SHARED
920 * the caller expects a shared lock to already be present and we
921 * are giving it another ref. This case must importantly not block
922 * if there is a pending exclusive lock request.
924 if (how
& HAMMER2_RESOLVE_SHARED
) {
925 if (how
& HAMMER2_RESOLVE_LOCKAGAIN
) {
926 hammer2_mtx_sh_again(&chain
->lock
);
928 hammer2_mtx_sh(&chain
->lock
);
931 hammer2_mtx_ex(&chain
->lock
);
933 ++curthread
->td_tracker
;
936 * If we already have a valid data pointer no further action is
943 * Do we have to resolve the data? This is generally only
944 * applicable to HAMMER2_BREF_TYPE_DATA which is special-cased.
945 * Other BREF types expects the data to be there.
947 switch(how
& HAMMER2_RESOLVE_MASK
) {
948 case HAMMER2_RESOLVE_NEVER
:
950 case HAMMER2_RESOLVE_MAYBE
:
951 if (chain
->flags
& HAMMER2_CHAIN_INITIAL
)
953 if (chain
->bref
.type
== HAMMER2_BREF_TYPE_DATA
)
956 if (chain
->bref
.type
== HAMMER2_BREF_TYPE_FREEMAP_NODE
)
958 if (chain
->bref
.type
== HAMMER2_BREF_TYPE_FREEMAP_LEAF
)
962 case HAMMER2_RESOLVE_ALWAYS
:
968 * Caller requires data
970 hammer2_chain_load_data(chain
);
974 * Lock the chain, retain the hold, and drop the data persistence count.
975 * The data should remain valid because we never transitioned lockcnt
979 hammer2_chain_lock_unhold(hammer2_chain_t
*chain
, int how
)
981 hammer2_chain_lock(chain
, how
);
982 atomic_add_int(&chain
->lockcnt
, -1);
987 * Downgrade an exclusive chain lock to a shared chain lock.
989 * NOTE: There is no upgrade equivalent due to the ease of
990 * deadlocks in that direction.
993 hammer2_chain_lock_downgrade(hammer2_chain_t
*chain
)
995 hammer2_mtx_downgrade(&chain
->lock
);
1001 * Obtains a second shared lock on the chain, does not account the second
1002 * shared lock as being owned by the current thread.
1004 * Caller must already own a shared lock on this chain.
1006 * The lock function is required to obtain the second shared lock without
1007 * blocking on pending exclusive requests.
1010 hammer2_chain_push_shared_lock(hammer2_chain_t
*chain
)
1012 hammer2_mtx_sh_again(&chain
->lock
);
1013 atomic_add_int(&chain
->lockcnt
, 1);
1014 /* do not count in td_tracker for this thread */
1018 * Accounts for a shared lock that was pushed to us as being owned by our
1022 hammer2_chain_pull_shared_lock(hammer2_chain_t
*chain
)
1024 ++curthread
->td_tracker
;
1029 * Issue I/O and install chain->data. Caller must hold a chain lock, lock
1030 * may be of any type.
1032 * Once chain->data is set it cannot be disposed of until all locks are
1036 hammer2_chain_load_data(hammer2_chain_t
*chain
)
1038 hammer2_blockref_t
*bref
;
1045 * Degenerate case, data already present, or chain is not expected
1050 if ((chain
->bref
.data_off
& HAMMER2_OFF_MASK_RADIX
) == 0)
1054 KKASSERT(hmp
!= NULL
);
1057 * Gain the IOINPROG bit, interlocked block.
1063 oflags
= chain
->flags
;
1065 if (oflags
& HAMMER2_CHAIN_IOINPROG
) {
1066 nflags
= oflags
| HAMMER2_CHAIN_IOSIGNAL
;
1067 tsleep_interlock(&chain
->flags
, 0);
1068 if (atomic_cmpset_int(&chain
->flags
, oflags
, nflags
)) {
1069 tsleep(&chain
->flags
, PINTERLOCKED
,
1074 nflags
= oflags
| HAMMER2_CHAIN_IOINPROG
;
1075 if (atomic_cmpset_int(&chain
->flags
, oflags
, nflags
)) {
1083 * We own CHAIN_IOINPROG
1085 * Degenerate case if we raced another load.
1091 * We must resolve to a device buffer, either by issuing I/O or
1092 * by creating a zero-fill element. We do not mark the buffer
1093 * dirty when creating a zero-fill element (the hammer2_chain_modify()
1094 * API must still be used to do that).
1096 * The device buffer is variable-sized in powers of 2 down
1097 * to HAMMER2_MIN_ALLOC (typically 1K). A 64K physical storage
1098 * chunk always contains buffers of the same size. (XXX)
1100 * The minimum physical IO size may be larger than the variable
1103 bref
= &chain
->bref
;
1106 * The getblk() optimization can only be used on newly created
1107 * elements if the physical block size matches the request.
1109 if (chain
->flags
& HAMMER2_CHAIN_INITIAL
) {
1110 error
= hammer2_io_new(hmp
, bref
->type
,
1111 bref
->data_off
, chain
->bytes
,
1114 error
= hammer2_io_bread(hmp
, bref
->type
,
1115 bref
->data_off
, chain
->bytes
,
1117 hammer2_adjreadcounter(&chain
->bref
, chain
->bytes
);
1120 chain
->error
= HAMMER2_ERROR_IO
;
1121 kprintf("hammer2_chain_lock: I/O error %016jx: %d\n",
1122 (intmax_t)bref
->data_off
, error
);
1123 hammer2_io_bqrelse(&chain
->dio
);
1129 * This isn't perfect and can be ignored on OSs which do not have
1130 * an indication as to whether a buffer is coming from cache or
1131 * if I/O was actually issued for the read. TESTEDGOOD will work
1132 * pretty well without the B_IOISSUED logic because chains are
1133 * cached, but in that situation (without B_IOISSUED) it will not
1134 * detect whether a re-read via I/O is corrupted verses the original
1137 * We can't re-run the CRC on every fresh lock. That would be
1138 * insanely expensive.
1140 * If the underlying kernel buffer covers the entire chain we can
1141 * use the B_IOISSUED indication to determine if we have to re-run
1142 * the CRC on chain data for chains that managed to stay cached
1143 * across the kernel disposal of the original buffer.
1145 if ((dio
= chain
->dio
) != NULL
&& dio
->bp
) {
1146 struct buf
*bp
= dio
->bp
;
1148 if (dio
->psize
== chain
->bytes
&&
1149 (bp
->b_flags
& B_IOISSUED
)) {
1150 atomic_clear_int(&chain
->flags
,
1151 HAMMER2_CHAIN_TESTEDGOOD
);
1152 bp
->b_flags
&= ~B_IOISSUED
;
1157 * NOTE: A locked chain's data cannot be modified without first
1158 * calling hammer2_chain_modify().
1162 * Clear INITIAL. In this case we used io_new() and the buffer has
1163 * been zero'd and marked dirty.
1165 bdata
= hammer2_io_data(chain
->dio
, chain
->bref
.data_off
);
1167 if (chain
->flags
& HAMMER2_CHAIN_INITIAL
) {
1168 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_INITIAL
);
1169 chain
->bref
.flags
|= HAMMER2_BREF_FLAG_ZERO
;
1170 } else if (chain
->flags
& HAMMER2_CHAIN_MODIFIED
) {
1172 * check data not currently synchronized due to
1173 * modification. XXX assumes data stays in the buffer
1174 * cache, which might not be true (need biodep on flush
1175 * to calculate crc? or simple crc?).
1177 } else if ((chain
->flags
& HAMMER2_CHAIN_TESTEDGOOD
) == 0) {
1178 if (hammer2_chain_testcheck(chain
, bdata
) == 0) {
1179 chain
->error
= HAMMER2_ERROR_CHECK
;
1181 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_TESTEDGOOD
);
1186 * Setup the data pointer, either pointing it to an embedded data
1187 * structure and copying the data from the buffer, or pointing it
1190 * The buffer is not retained when copying to an embedded data
1191 * structure in order to avoid potential deadlocks or recursions
1192 * on the same physical buffer.
1194 * WARNING! Other threads can start using the data the instant we
1195 * set chain->data non-NULL.
1197 switch (bref
->type
) {
1198 case HAMMER2_BREF_TYPE_VOLUME
:
1199 case HAMMER2_BREF_TYPE_FREEMAP
:
1201 * Copy data from bp to embedded buffer
1203 panic("hammer2_chain_load_data: unresolved volume header");
1205 case HAMMER2_BREF_TYPE_DIRENT
:
1206 KKASSERT(chain
->bytes
!= 0);
1208 case HAMMER2_BREF_TYPE_INODE
:
1209 case HAMMER2_BREF_TYPE_FREEMAP_LEAF
:
1210 case HAMMER2_BREF_TYPE_INDIRECT
:
1211 case HAMMER2_BREF_TYPE_DATA
:
1212 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
1215 * Point data at the device buffer and leave dio intact.
1217 chain
->data
= (void *)bdata
;
1222 * Release HAMMER2_CHAIN_IOINPROG and signal waiters if requested.
1229 oflags
= chain
->flags
;
1230 nflags
= oflags
& ~(HAMMER2_CHAIN_IOINPROG
|
1231 HAMMER2_CHAIN_IOSIGNAL
);
1232 KKASSERT(oflags
& HAMMER2_CHAIN_IOINPROG
);
1233 if (atomic_cmpset_int(&chain
->flags
, oflags
, nflags
)) {
1234 if (oflags
& HAMMER2_CHAIN_IOSIGNAL
)
1235 wakeup(&chain
->flags
);
1242 * Unlock and deref a chain element.
1244 * Remember that the presence of children under chain prevent the chain's
1245 * destruction but do not add additional references, so the dio will still
1249 hammer2_chain_unlock(hammer2_chain_t
*chain
)
1255 --curthread
->td_tracker
;
1258 * If multiple locks are present (or being attempted) on this
1259 * particular chain we can just unlock, drop refs, and return.
1261 * Otherwise fall-through on the 1->0 transition.
1264 lockcnt
= chain
->lockcnt
;
1265 KKASSERT(lockcnt
> 0);
1268 if (atomic_cmpset_int(&chain
->lockcnt
,
1269 lockcnt
, lockcnt
- 1)) {
1270 hammer2_mtx_unlock(&chain
->lock
);
1273 } else if (hammer2_mtx_upgrade_try(&chain
->lock
) == 0) {
1274 /* while holding the mutex exclusively */
1275 if (atomic_cmpset_int(&chain
->lockcnt
, 1, 0))
1279 * This situation can easily occur on SMP due to
1280 * the gap inbetween the 1->0 transition and the
1281 * final unlock. We cannot safely block on the
1282 * mutex because lockcnt might go above 1.
1284 * XXX Sleep for one tick if it takes too long.
1286 if (++iter
> 1000) {
1287 if (iter
> 1000 + hz
) {
1288 kprintf("hammer2: h2race2 %p\n", chain
);
1291 tsleep(&iter
, 0, "h2race2", 1);
1299 * Last unlock / mutex upgraded to exclusive. Drop the data
1302 dio
= hammer2_chain_drop_data(chain
);
1304 hammer2_io_bqrelse(&dio
);
1305 hammer2_mtx_unlock(&chain
->lock
);
1309 * Unlock and hold chain data intact
1312 hammer2_chain_unlock_hold(hammer2_chain_t
*chain
)
1314 atomic_add_int(&chain
->lockcnt
, 1);
1315 hammer2_chain_unlock(chain
);
1319 * Helper to obtain the blockref[] array base and count for a chain.
1321 * XXX Not widely used yet, various use cases need to be validated and
1322 * converted to use this function.
1325 hammer2_blockref_t
*
1326 hammer2_chain_base_and_count(hammer2_chain_t
*parent
, int *countp
)
1328 hammer2_blockref_t
*base
;
1331 if (parent
->flags
& HAMMER2_CHAIN_INITIAL
) {
1334 switch(parent
->bref
.type
) {
1335 case HAMMER2_BREF_TYPE_INODE
:
1336 count
= HAMMER2_SET_COUNT
;
1338 case HAMMER2_BREF_TYPE_INDIRECT
:
1339 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
1340 count
= parent
->bytes
/ sizeof(hammer2_blockref_t
);
1342 case HAMMER2_BREF_TYPE_VOLUME
:
1343 count
= HAMMER2_SET_COUNT
;
1345 case HAMMER2_BREF_TYPE_FREEMAP
:
1346 count
= HAMMER2_SET_COUNT
;
1349 panic("hammer2_chain_create_indirect: "
1350 "unrecognized blockref type: %d",
1356 switch(parent
->bref
.type
) {
1357 case HAMMER2_BREF_TYPE_INODE
:
1358 base
= &parent
->data
->ipdata
.u
.blockset
.blockref
[0];
1359 count
= HAMMER2_SET_COUNT
;
1361 case HAMMER2_BREF_TYPE_INDIRECT
:
1362 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
1363 base
= &parent
->data
->npdata
[0];
1364 count
= parent
->bytes
/ sizeof(hammer2_blockref_t
);
1366 case HAMMER2_BREF_TYPE_VOLUME
:
1367 base
= &parent
->data
->voldata
.
1368 sroot_blockset
.blockref
[0];
1369 count
= HAMMER2_SET_COUNT
;
1371 case HAMMER2_BREF_TYPE_FREEMAP
:
1372 base
= &parent
->data
->blkset
.blockref
[0];
1373 count
= HAMMER2_SET_COUNT
;
1376 panic("hammer2_chain_create_indirect: "
1377 "unrecognized blockref type: %d",
1389 * This counts the number of live blockrefs in a block array and
1390 * also calculates the point at which all remaining blockrefs are empty.
1391 * This routine can only be called on a live chain.
1393 * NOTE: Flag is not set until after the count is complete, allowing
1394 * callers to test the flag without holding the spinlock.
1396 * NOTE: If base is NULL the related chain is still in the INITIAL
1397 * state and there are no blockrefs to count.
1399 * NOTE: live_count may already have some counts accumulated due to
1400 * creation and deletion and could even be initially negative.
1403 hammer2_chain_countbrefs(hammer2_chain_t
*chain
,
1404 hammer2_blockref_t
*base
, int count
)
1406 hammer2_spin_ex(&chain
->core
.spin
);
1407 if ((chain
->flags
& HAMMER2_CHAIN_COUNTEDBREFS
) == 0) {
1409 while (--count
>= 0) {
1410 if (base
[count
].type
)
1413 chain
->core
.live_zero
= count
+ 1;
1414 while (count
>= 0) {
1415 if (base
[count
].type
)
1416 atomic_add_int(&chain
->core
.live_count
,
1421 chain
->core
.live_zero
= 0;
1423 /* else do not modify live_count */
1424 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_COUNTEDBREFS
);
1426 hammer2_spin_unex(&chain
->core
.spin
);
1430 * Resize the chain's physical storage allocation in-place. This function does
1431 * not usually adjust the data pointer and must be followed by (typically) a
1432 * hammer2_chain_modify() call to copy any old data over and adjust the
1435 * Chains can be resized smaller without reallocating the storage. Resizing
1436 * larger will reallocate the storage. Excess or prior storage is reclaimed
1437 * asynchronously at a later time.
1439 * An nradix value of 0 is special-cased to mean that the storage should
1440 * be disassociated, that is the chain is being resized to 0 bytes (not 1
1443 * Must be passed an exclusively locked parent and chain.
1445 * This function is mostly used with DATA blocks locked RESOLVE_NEVER in order
1446 * to avoid instantiating a device buffer that conflicts with the vnode data
1447 * buffer. However, because H2 can compress or encrypt data, the chain may
1448 * have a dio assigned to it in those situations, and they do not conflict.
1450 * XXX return error if cannot resize.
1453 hammer2_chain_resize(hammer2_chain_t
*chain
,
1454 hammer2_tid_t mtid
, hammer2_off_t dedup_off
,
1455 int nradix
, int flags
)
1464 * Only data and indirect blocks can be resized for now.
1465 * (The volu root, inodes, and freemap elements use a fixed size).
1467 KKASSERT(chain
!= &hmp
->vchain
);
1468 KKASSERT(chain
->bref
.type
== HAMMER2_BREF_TYPE_DATA
||
1469 chain
->bref
.type
== HAMMER2_BREF_TYPE_INDIRECT
||
1470 chain
->bref
.type
== HAMMER2_BREF_TYPE_DIRENT
);
1473 * Nothing to do if the element is already the proper size
1475 obytes
= chain
->bytes
;
1476 nbytes
= (nradix
) ? (1U << nradix
) : 0;
1477 if (obytes
== nbytes
)
1481 * Make sure the old data is instantiated so we can copy it. If this
1482 * is a data block, the device data may be superfluous since the data
1483 * might be in a logical block, but compressed or encrypted data is
1486 * NOTE: The modify will set BMAPUPD for us if BMAPPED is set.
1488 hammer2_chain_modify(chain
, mtid
, dedup_off
, 0);
1491 * Relocate the block, even if making it smaller (because different
1492 * block sizes may be in different regions).
1494 * NOTE: Operation does not copy the data and may only be used
1495 * to resize data blocks in-place, or directory entry blocks
1496 * which are about to be modified in some manner.
1498 hammer2_freemap_alloc(chain
, nbytes
);
1499 chain
->bytes
= nbytes
;
1502 * We don't want the followup chain_modify() to try to copy data
1503 * from the old (wrong-sized) buffer. It won't know how much to
1504 * copy. This case should only occur during writes when the
1505 * originator already has the data to write in-hand.
1508 KKASSERT(chain
->bref
.type
== HAMMER2_BREF_TYPE_DATA
||
1509 chain
->bref
.type
== HAMMER2_BREF_TYPE_DIRENT
);
1510 hammer2_io_brelse(&chain
->dio
);
1516 * Set the chain modified so its data can be changed by the caller, or
1517 * install deduplicated data. The caller must call this routine for each
1518 * set of modifications it makes, even if the chain is already flagged
1521 * Sets bref.modify_tid to mtid only if mtid != 0. Note that bref.modify_tid
1522 * is a CLC (cluster level change) field and is not updated by parent
1523 * propagation during a flush.
1527 * If the DEDUPABLE flag is set in the chain the storage must be reallocated
1528 * even if the chain is still flagged MODIFIED. In this case the chain's
1529 * DEDUPABLE flag will be cleared once the new storage has been assigned.
1531 * If the caller passes a non-zero dedup_off we will use it to assign the
1532 * new storage. The MODIFIED flag will be *CLEARED* in this case, and
1533 * DEDUPABLE will be set (NOTE: the UPDATE flag is always set). The caller
1534 * must not modify the data content upon return.
1537 hammer2_chain_modify(hammer2_chain_t
*chain
, hammer2_tid_t mtid
,
1538 hammer2_off_t dedup_off
, int flags
)
1540 hammer2_blockref_t obref
;
1549 obref
= chain
->bref
;
1550 KKASSERT((chain
->flags
& HAMMER2_CHAIN_FICTITIOUS
) == 0);
1553 * Data is not optional for freemap chains (we must always be sure
1554 * to copy the data on COW storage allocations).
1556 if (chain
->bref
.type
== HAMMER2_BREF_TYPE_FREEMAP_NODE
||
1557 chain
->bref
.type
== HAMMER2_BREF_TYPE_FREEMAP_LEAF
) {
1558 KKASSERT((chain
->flags
& HAMMER2_CHAIN_INITIAL
) ||
1559 (flags
& HAMMER2_MODIFY_OPTDATA
) == 0);
1563 * Data must be resolved if already assigned, unless explicitly
1564 * flagged otherwise.
1566 if (chain
->data
== NULL
&& chain
->bytes
!= 0 &&
1567 (flags
& HAMMER2_MODIFY_OPTDATA
) == 0 &&
1568 (chain
->bref
.data_off
& ~HAMMER2_OFF_MASK_RADIX
)) {
1569 hammer2_chain_load_data(chain
);
1573 * Set MODIFIED to indicate that the chain has been modified. A new
1574 * allocation is required when modifying a chain.
1576 * Set UPDATE to ensure that the blockref is updated in the parent.
1579 * If MODIFIED is already set determine if we can reuse the assigned
1580 * data block or if we need a new data block.
1582 if ((chain
->flags
& HAMMER2_CHAIN_MODIFIED
) == 0) {
1584 * Must set modified bit.
1586 atomic_add_long(&hammer2_count_modified_chains
, 1);
1587 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_MODIFIED
);
1588 hammer2_pfs_memory_inc(chain
->pmp
); /* can be NULL */
1591 * We may be able to avoid a copy-on-write if the chain's
1592 * check mode is set to NONE and the chain's current
1593 * modify_tid is beyond the last explicit snapshot tid.
1595 * This implements HAMMER2's overwrite-in-place feature.
1597 * NOTE! This data-block cannot be used as a de-duplication
1598 * source when the check mode is set to NONE.
1600 if ((chain
->bref
.type
== HAMMER2_BREF_TYPE_DATA
||
1601 chain
->bref
.type
== HAMMER2_BREF_TYPE_DIRENT
) &&
1602 (chain
->flags
& HAMMER2_CHAIN_INITIAL
) == 0 &&
1603 (chain
->flags
& HAMMER2_CHAIN_DEDUPABLE
) == 0 &&
1604 HAMMER2_DEC_CHECK(chain
->bref
.methods
) ==
1605 HAMMER2_CHECK_NONE
&&
1607 chain
->bref
.modify_tid
>
1608 chain
->pmp
->iroot
->meta
.pfs_lsnap_tid
) {
1610 * Sector overwrite allowed.
1615 * Sector overwrite not allowed, must copy-on-write.
1619 } else if (chain
->flags
& HAMMER2_CHAIN_DEDUPABLE
) {
1621 * If the modified chain was registered for dedup we need
1622 * a new allocation. This only happens for delayed-flush
1623 * chains (i.e. which run through the front-end buffer
1629 * Already flagged modified, no new allocation is needed.
1635 * Flag parent update required.
1637 if ((chain
->flags
& HAMMER2_CHAIN_UPDATE
) == 0)
1638 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_UPDATE
);
1641 * The modification or re-modification requires an allocation and
1644 * If dedup_off is non-zero, caller already has a data offset
1645 * containing the caller's desired data. The dedup offset is
1646 * allowed to be in a partially free state and we must be sure
1647 * to reset it to a fully allocated state to force two bulkfree
1648 * passes to free it again. The chain will not be marked MODIFIED
1649 * in the dedup case, as the dedup data cannot be changed without
1652 * NOTE: Only applicable when chain->bytes != 0.
1654 * XXX can a chain already be marked MODIFIED without a data
1655 * assignment? If not, assert here instead of testing the case.
1657 if (chain
!= &hmp
->vchain
&& chain
!= &hmp
->fchain
&&
1659 if ((chain
->bref
.data_off
& ~HAMMER2_OFF_MASK_RADIX
) == 0 ||
1663 * NOTE: We do not have to remove the dedup
1664 * registration because the area is still
1665 * allocated and the underlying DIO will
1669 chain
->bref
.data_off
= dedup_off
;
1670 chain
->bytes
= 1 << (dedup_off
&
1671 HAMMER2_OFF_MASK_RADIX
);
1672 atomic_clear_int(&chain
->flags
,
1673 HAMMER2_CHAIN_MODIFIED
);
1674 atomic_add_long(&hammer2_count_modified_chains
,
1677 hammer2_pfs_memory_wakeup(chain
->pmp
);
1678 hammer2_freemap_adjust(hmp
, &chain
->bref
,
1679 HAMMER2_FREEMAP_DORECOVER
);
1680 atomic_set_int(&chain
->flags
,
1681 HAMMER2_CHAIN_DEDUPABLE
);
1683 hammer2_freemap_alloc(chain
, chain
->bytes
);
1684 atomic_clear_int(&chain
->flags
,
1685 HAMMER2_CHAIN_DEDUPABLE
);
1687 /* XXX failed allocation */
1692 * Update mirror_tid and modify_tid. modify_tid is only updated
1693 * if not passed as zero (during flushes, parent propagation passes
1696 * NOTE: chain->pmp could be the device spmp.
1698 chain
->bref
.mirror_tid
= hmp
->voldata
.mirror_tid
+ 1;
1700 chain
->bref
.modify_tid
= mtid
;
1703 * Set BMAPUPD to tell the flush code that an existing blockmap entry
1704 * requires updating as well as to tell the delete code that the
1705 * chain's blockref might not exactly match (in terms of physical size
1706 * or block offset) the one in the parent's blocktable. The base key
1707 * of course will still match.
1709 if (chain
->flags
& HAMMER2_CHAIN_BMAPPED
)
1710 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_BMAPUPD
);
1713 * Short-cut data blocks which the caller does not need an actual
1714 * data reference to (aka OPTDATA), as long as the chain does not
1715 * already have a data pointer to the data. This generally means
1716 * that the modifications are being done via the logical buffer cache.
1717 * The INITIAL flag relates only to the device data buffer and thus
1718 * remains unchange in this situation.
1720 * This code also handles bytes == 0 (most dirents).
1722 if (chain
->bref
.type
== HAMMER2_BREF_TYPE_DATA
&&
1723 (flags
& HAMMER2_MODIFY_OPTDATA
) &&
1724 chain
->data
== NULL
) {
1725 KKASSERT(chain
->dio
== NULL
);
1730 * Clearing the INITIAL flag (for indirect blocks) indicates that
1731 * we've processed the uninitialized storage allocation.
1733 * If this flag is already clear we are likely in a copy-on-write
1734 * situation but we have to be sure NOT to bzero the storage if
1735 * no data is present.
1737 if (chain
->flags
& HAMMER2_CHAIN_INITIAL
) {
1738 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_INITIAL
);
1745 * Instantiate data buffer and possibly execute COW operation
1747 switch(chain
->bref
.type
) {
1748 case HAMMER2_BREF_TYPE_VOLUME
:
1749 case HAMMER2_BREF_TYPE_FREEMAP
:
1751 * The data is embedded, no copy-on-write operation is
1754 KKASSERT(chain
->dio
== NULL
);
1756 case HAMMER2_BREF_TYPE_DIRENT
:
1758 * The data might be fully embedded.
1760 if (chain
->bytes
== 0) {
1761 KKASSERT(chain
->dio
== NULL
);
1765 case HAMMER2_BREF_TYPE_INODE
:
1766 case HAMMER2_BREF_TYPE_FREEMAP_LEAF
:
1767 case HAMMER2_BREF_TYPE_DATA
:
1768 case HAMMER2_BREF_TYPE_INDIRECT
:
1769 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
1771 * Perform the copy-on-write operation
1773 * zero-fill or copy-on-write depending on whether
1774 * chain->data exists or not and set the dirty state for
1775 * the new buffer. hammer2_io_new() will handle the
1778 * If a dedup_off was supplied this is an existing block
1779 * and no COW, copy, or further modification is required.
1781 KKASSERT(chain
!= &hmp
->vchain
&& chain
!= &hmp
->fchain
);
1783 if (wasinitial
&& dedup_off
== 0) {
1784 error
= hammer2_io_new(hmp
, chain
->bref
.type
,
1785 chain
->bref
.data_off
,
1786 chain
->bytes
, &dio
);
1788 error
= hammer2_io_bread(hmp
, chain
->bref
.type
,
1789 chain
->bref
.data_off
,
1790 chain
->bytes
, &dio
);
1792 hammer2_adjreadcounter(&chain
->bref
, chain
->bytes
);
1795 * If an I/O error occurs make sure callers cannot accidently
1796 * modify the old buffer's contents and corrupt the filesystem.
1799 kprintf("hammer2_chain_modify: hmp=%p I/O error\n",
1801 chain
->error
= HAMMER2_ERROR_IO
;
1802 hammer2_io_brelse(&dio
);
1803 hammer2_io_brelse(&chain
->dio
);
1808 bdata
= hammer2_io_data(dio
, chain
->bref
.data_off
);
1812 * COW (unless a dedup).
1814 KKASSERT(chain
->dio
!= NULL
);
1815 if (chain
->data
!= (void *)bdata
&& dedup_off
== 0) {
1816 bcopy(chain
->data
, bdata
, chain
->bytes
);
1818 } else if (wasinitial
== 0) {
1820 * We have a problem. We were asked to COW but
1821 * we don't have any data to COW with!
1823 panic("hammer2_chain_modify: having a COW %p\n",
1828 * Retire the old buffer, replace with the new. Dirty or
1829 * redirty the new buffer.
1831 * WARNING! The system buffer cache may have already flushed
1832 * the buffer, so we must be sure to [re]dirty it
1833 * for further modification.
1835 * If dedup_off was supplied, the caller is not
1836 * expected to make any further modification to the
1840 hammer2_io_bqrelse(&chain
->dio
);
1841 chain
->data
= (void *)bdata
;
1844 hammer2_io_setdirty(dio
);
1847 panic("hammer2_chain_modify: illegal non-embedded type %d",
1854 * setflush on parent indicating that the parent must recurse down
1855 * to us. Do not call on chain itself which might already have it
1859 hammer2_chain_setflush(chain
->parent
);
1863 * Modify the chain associated with an inode.
1866 hammer2_chain_modify_ip(hammer2_inode_t
*ip
, hammer2_chain_t
*chain
,
1867 hammer2_tid_t mtid
, int flags
)
1869 hammer2_inode_modify(ip
);
1870 hammer2_chain_modify(chain
, mtid
, 0, flags
);
1874 * Volume header data locks
1877 hammer2_voldata_lock(hammer2_dev_t
*hmp
)
1879 lockmgr(&hmp
->vollk
, LK_EXCLUSIVE
);
1883 hammer2_voldata_unlock(hammer2_dev_t
*hmp
)
1885 lockmgr(&hmp
->vollk
, LK_RELEASE
);
1889 hammer2_voldata_modify(hammer2_dev_t
*hmp
)
1891 if ((hmp
->vchain
.flags
& HAMMER2_CHAIN_MODIFIED
) == 0) {
1892 atomic_add_long(&hammer2_count_modified_chains
, 1);
1893 atomic_set_int(&hmp
->vchain
.flags
, HAMMER2_CHAIN_MODIFIED
);
1894 hammer2_pfs_memory_inc(hmp
->vchain
.pmp
);
1899 * This function returns the chain at the nearest key within the specified
1900 * range. The returned chain will be referenced but not locked.
1902 * This function will recurse through chain->rbtree as necessary and will
1903 * return a *key_nextp suitable for iteration. *key_nextp is only set if
1904 * the iteration value is less than the current value of *key_nextp.
1906 * The caller should use (*key_nextp) to calculate the actual range of
1907 * the returned element, which will be (key_beg to *key_nextp - 1), because
1908 * there might be another element which is superior to the returned element
1911 * (*key_nextp) can be passed as key_beg in an iteration only while non-NULL
1912 * chains continue to be returned. On EOF (*key_nextp) may overflow since
1913 * it will wind up being (key_end + 1).
1915 * WARNING! Must be called with child's spinlock held. Spinlock remains
1916 * held through the operation.
1918 struct hammer2_chain_find_info
{
1919 hammer2_chain_t
*best
;
1920 hammer2_key_t key_beg
;
1921 hammer2_key_t key_end
;
1922 hammer2_key_t key_next
;
1925 static int hammer2_chain_find_cmp(hammer2_chain_t
*child
, void *data
);
1926 static int hammer2_chain_find_callback(hammer2_chain_t
*child
, void *data
);
1930 hammer2_chain_find(hammer2_chain_t
*parent
, hammer2_key_t
*key_nextp
,
1931 hammer2_key_t key_beg
, hammer2_key_t key_end
)
1933 struct hammer2_chain_find_info info
;
1936 info
.key_beg
= key_beg
;
1937 info
.key_end
= key_end
;
1938 info
.key_next
= *key_nextp
;
1940 RB_SCAN(hammer2_chain_tree
, &parent
->core
.rbtree
,
1941 hammer2_chain_find_cmp
, hammer2_chain_find_callback
,
1943 *key_nextp
= info
.key_next
;
1945 kprintf("chain_find %p %016jx:%016jx next=%016jx\n",
1946 parent
, key_beg
, key_end
, *key_nextp
);
1954 hammer2_chain_find_cmp(hammer2_chain_t
*child
, void *data
)
1956 struct hammer2_chain_find_info
*info
= data
;
1957 hammer2_key_t child_beg
;
1958 hammer2_key_t child_end
;
1960 child_beg
= child
->bref
.key
;
1961 child_end
= child_beg
+ ((hammer2_key_t
)1 << child
->bref
.keybits
) - 1;
1963 if (child_end
< info
->key_beg
)
1965 if (child_beg
> info
->key_end
)
1972 hammer2_chain_find_callback(hammer2_chain_t
*child
, void *data
)
1974 struct hammer2_chain_find_info
*info
= data
;
1975 hammer2_chain_t
*best
;
1976 hammer2_key_t child_end
;
1979 * WARNING! Layerq is scanned forwards, exact matches should keep
1980 * the existing info->best.
1982 if ((best
= info
->best
) == NULL
) {
1984 * No previous best. Assign best
1987 } else if (best
->bref
.key
<= info
->key_beg
&&
1988 child
->bref
.key
<= info
->key_beg
) {
1993 /*info->best = child;*/
1994 } else if (child
->bref
.key
< best
->bref
.key
) {
1996 * Child has a nearer key and best is not flush with key_beg.
1997 * Set best to child. Truncate key_next to the old best key.
2000 if (info
->key_next
> best
->bref
.key
|| info
->key_next
== 0)
2001 info
->key_next
= best
->bref
.key
;
2002 } else if (child
->bref
.key
== best
->bref
.key
) {
2004 * If our current best is flush with the child then this
2005 * is an illegal overlap.
2007 * key_next will automatically be limited to the smaller of
2008 * the two end-points.
2014 * Keep the current best but truncate key_next to the child's
2017 * key_next will also automatically be limited to the smaller
2018 * of the two end-points (probably not necessary for this case
2019 * but we do it anyway).
2021 if (info
->key_next
> child
->bref
.key
|| info
->key_next
== 0)
2022 info
->key_next
= child
->bref
.key
;
2026 * Always truncate key_next based on child's end-of-range.
2028 child_end
= child
->bref
.key
+ ((hammer2_key_t
)1 << child
->bref
.keybits
);
2029 if (child_end
&& (info
->key_next
> child_end
|| info
->key_next
== 0))
2030 info
->key_next
= child_end
;
2036 * Retrieve the specified chain from a media blockref, creating the
2037 * in-memory chain structure which reflects it. The returned chain is
2038 * held but not locked. The caller must lock it to crc-check and
2039 * dereference its data, and should check chain->error after locking
2040 * before assuming that the data is good.
2042 * To handle insertion races pass the INSERT_RACE flag along with the
2043 * generation number of the core. NULL will be returned if the generation
2044 * number changes before we have a chance to insert the chain. Insert
2045 * races can occur because the parent might be held shared.
2047 * Caller must hold the parent locked shared or exclusive since we may
2048 * need the parent's bref array to find our block.
2050 * WARNING! chain->pmp is always set to NULL for any chain representing
2051 * part of the super-root topology.
2054 hammer2_chain_get(hammer2_chain_t
*parent
, int generation
,
2055 hammer2_blockref_t
*bref
)
2057 hammer2_dev_t
*hmp
= parent
->hmp
;
2058 hammer2_chain_t
*chain
;
2062 * Allocate a chain structure representing the existing media
2063 * entry. Resulting chain has one ref and is not locked.
2065 if (bref
->flags
& HAMMER2_BREF_FLAG_PFSROOT
)
2066 chain
= hammer2_chain_alloc(hmp
, NULL
, bref
);
2068 chain
= hammer2_chain_alloc(hmp
, parent
->pmp
, bref
);
2069 /* ref'd chain returned */
2072 * Flag that the chain is in the parent's blockmap so delete/flush
2073 * knows what to do with it.
2075 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_BMAPPED
);
2078 * Link the chain into its parent. A spinlock is required to safely
2079 * access the RBTREE, and it is possible to collide with another
2080 * hammer2_chain_get() operation because the caller might only hold
2081 * a shared lock on the parent.
2083 * NOTE: Get races can occur quite often when we distribute
2084 * asynchronous read-aheads across multiple threads.
2086 KKASSERT(parent
->refs
> 0);
2087 error
= hammer2_chain_insert(parent
, chain
,
2088 HAMMER2_CHAIN_INSERT_SPIN
|
2089 HAMMER2_CHAIN_INSERT_RACE
,
2092 KKASSERT((chain
->flags
& HAMMER2_CHAIN_ONRBTREE
) == 0);
2093 /*kprintf("chain %p get race\n", chain);*/
2094 hammer2_chain_drop(chain
);
2097 KKASSERT(chain
->flags
& HAMMER2_CHAIN_ONRBTREE
);
2101 * Return our new chain referenced but not locked, or NULL if
2108 * Lookup initialization/completion API
2111 hammer2_chain_lookup_init(hammer2_chain_t
*parent
, int flags
)
2113 hammer2_chain_ref(parent
);
2114 if (flags
& HAMMER2_LOOKUP_SHARED
) {
2115 hammer2_chain_lock(parent
, HAMMER2_RESOLVE_ALWAYS
|
2116 HAMMER2_RESOLVE_SHARED
);
2118 hammer2_chain_lock(parent
, HAMMER2_RESOLVE_ALWAYS
);
2124 hammer2_chain_lookup_done(hammer2_chain_t
*parent
)
2127 hammer2_chain_unlock(parent
);
2128 hammer2_chain_drop(parent
);
2133 * Take the locked chain and return a locked parent. The chain remains
2136 * This will work even if the chain is errored, and the caller can check
2137 * parent->error on return if desired since the parent will be locked.
2139 * This function handles the lock order reversal.
2142 hammer2_chain_getparent(hammer2_chain_t
*chain
, int how
)
2144 hammer2_chain_t
*parent
;
2147 * Be careful of order, chain must be unlocked before parent
2148 * is locked below to avoid a deadlock.
2150 * Safe access to fu->parent requires fu's core spinlock.
2153 hammer2_spin_ex(&chain
->core
.spin
);
2154 parent
= chain
->parent
;
2155 if (parent
== NULL
) {
2156 hammer2_spin_unex(&chain
->core
.spin
);
2157 panic("hammer2_chain_getparent: no parent");
2159 hammer2_chain_ref(parent
);
2160 hammer2_spin_unex(&chain
->core
.spin
);
2162 hammer2_chain_unlock(chain
);
2163 hammer2_chain_lock(parent
, how
);
2164 hammer2_chain_lock(chain
, how
);
2167 * Parent relinking races are quite common. We have to get it right
2168 * or we will blow up the block table.
2170 if (chain
->parent
!= parent
) {
2171 hammer2_chain_unlock(parent
);
2172 hammer2_chain_drop(parent
);
2179 * Take the locked chain and return a locked parent. The chain is unlocked
2180 * and dropped. *chainp is set to the returned parent as a convenience.
2182 * This will work even if the chain is errored, and the caller can check
2183 * parent->error on return if desired since the parent will be locked.
2185 * This function handles the lock order reversal.
2188 hammer2_chain_repparent(hammer2_chain_t
**chainp
, int how
)
2190 hammer2_chain_t
*chain
;
2191 hammer2_chain_t
*parent
;
2194 * Be careful of order, chain must be unlocked before parent
2195 * is locked below to avoid a deadlock.
2197 * Safe access to fu->parent requires fu's core spinlock.
2201 hammer2_spin_ex(&chain
->core
.spin
);
2202 parent
= chain
->parent
;
2203 if (parent
== NULL
) {
2204 hammer2_spin_unex(&chain
->core
.spin
);
2205 panic("hammer2_chain_getparent: no parent");
2207 hammer2_chain_ref(parent
);
2208 hammer2_spin_unex(&chain
->core
.spin
);
2210 hammer2_chain_unlock(chain
);
2211 hammer2_chain_lock(parent
, how
);
2214 * Parent relinking races are quite common. We have to get it right
2215 * or we will blow up the block table.
2217 if (chain
->parent
!= parent
) {
2218 hammer2_chain_lock(chain
, how
);
2219 hammer2_chain_unlock(parent
);
2220 hammer2_chain_drop(parent
);
2223 hammer2_chain_drop(chain
);
2230 * Locate the first chain whos key range overlaps (key_beg, key_end) inclusive.
2231 * (*parentp) typically points to an inode but can also point to a related
2232 * indirect block and this function will recurse upwards and find the inode
2235 * This function unconditionally sets *errorp, replacing any previous value.
2237 * (*parentp) must be exclusively locked and referenced and can be an inode
2238 * or an existing indirect block within the inode. If (*parent) is errored
2239 * out, this function will not attempt to recurse the radix tree and
2240 * will return NULL along with an appropriate *errorp. If NULL is returned
2241 * and *errorp is 0, the requested lookup could not be located.
2243 * On return (*parentp) will be modified to point at the deepest parent chain
2244 * element encountered during the search, as a helper for an insertion or
2245 * deletion. The new (*parentp) will be locked and referenced and the old
2246 * will be unlocked and dereferenced (no change if they are both the same).
2247 * This is particularly important if the caller wishes to insert a new chain,
2248 * (*parentp) will be set properly even if NULL is returned, as long as no
2251 * The matching chain will be returned exclusively locked. If NOLOCK is
2252 * requested the chain will be returned only referenced. Note that the
2253 * parent chain must always be locked shared or exclusive, matching the
2254 * HAMMER2_LOOKUP_SHARED flag. We can conceivably lock it SHARED temporarily
2255 * when NOLOCK is specified but that complicates matters if *parentp must
2256 * inherit the chain.
2258 * NOLOCK also implies NODATA, since an unlocked chain usually has a NULL
2259 * data pointer or can otherwise be in flux.
2261 * NULL is returned if no match was found, but (*parentp) will still
2262 * potentially be adjusted.
2264 * On return (*key_nextp) will point to an iterative value for key_beg.
2265 * (If NULL is returned (*key_nextp) is set to (key_end + 1)).
2267 * This function will also recurse up the chain if the key is not within the
2268 * current parent's range. (*parentp) can never be set to NULL. An iteration
2269 * can simply allow (*parentp) to float inside the loop.
2271 * NOTE! chain->data is not always resolved. By default it will not be
2272 * resolved for BREF_TYPE_DATA, FREEMAP_NODE, or FREEMAP_LEAF. Use
2273 * HAMMER2_LOOKUP_ALWAYS to force resolution (but be careful w/
2274 * BREF_TYPE_DATA as the device buffer can alias the logical file
2279 hammer2_chain_lookup(hammer2_chain_t
**parentp
, hammer2_key_t
*key_nextp
,
2280 hammer2_key_t key_beg
, hammer2_key_t key_end
,
2281 int *errorp
, int flags
)
2284 hammer2_chain_t
*parent
;
2285 hammer2_chain_t
*chain
;
2286 hammer2_blockref_t
*base
;
2287 hammer2_blockref_t
*bref
;
2288 hammer2_blockref_t bcopy
;
2289 hammer2_key_t scan_beg
;
2290 hammer2_key_t scan_end
;
2292 int how_always
= HAMMER2_RESOLVE_ALWAYS
;
2293 int how_maybe
= HAMMER2_RESOLVE_MAYBE
;
2296 int maxloops
= 300000;
2298 if (flags
& HAMMER2_LOOKUP_ALWAYS
) {
2299 how_maybe
= how_always
;
2300 how
= HAMMER2_RESOLVE_ALWAYS
;
2301 } else if (flags
& (HAMMER2_LOOKUP_NODATA
| HAMMER2_LOOKUP_NOLOCK
)) {
2302 how
= HAMMER2_RESOLVE_NEVER
;
2304 how
= HAMMER2_RESOLVE_MAYBE
;
2306 if (flags
& HAMMER2_LOOKUP_SHARED
) {
2307 how_maybe
|= HAMMER2_RESOLVE_SHARED
;
2308 how_always
|= HAMMER2_RESOLVE_SHARED
;
2309 how
|= HAMMER2_RESOLVE_SHARED
;
2313 * Recurse (*parentp) upward if necessary until the parent completely
2314 * encloses the key range or we hit the inode.
2316 * Handle races against the flusher deleting indirect nodes on its
2317 * way back up by continuing to recurse upward past the deletion.
2323 while (parent
->bref
.type
== HAMMER2_BREF_TYPE_INDIRECT
||
2324 parent
->bref
.type
== HAMMER2_BREF_TYPE_FREEMAP_NODE
) {
2325 scan_beg
= parent
->bref
.key
;
2326 scan_end
= scan_beg
+
2327 ((hammer2_key_t
)1 << parent
->bref
.keybits
) - 1;
2328 if ((parent
->flags
& HAMMER2_CHAIN_DELETED
) == 0) {
2329 if (key_beg
>= scan_beg
&& key_end
<= scan_end
)
2332 parent
= hammer2_chain_repparent(parentp
, how_maybe
);
2336 if (--maxloops
== 0)
2337 panic("hammer2_chain_lookup: maxloops");
2339 * Locate the blockref array. Currently we do a fully associative
2340 * search through the array.
2342 switch(parent
->bref
.type
) {
2343 case HAMMER2_BREF_TYPE_INODE
:
2345 * Special shortcut for embedded data returns the inode
2346 * itself. Callers must detect this condition and access
2347 * the embedded data (the strategy code does this for us).
2349 * This is only applicable to regular files and softlinks.
2351 * We need a second lock on parent. Since we already have
2352 * a lock we must pass LOCKAGAIN to prevent unexpected
2353 * blocking (we don't want to block on a second shared
2354 * ref if an exclusive lock is pending)
2356 if (parent
->data
->ipdata
.meta
.op_flags
&
2357 HAMMER2_OPFLAG_DIRECTDATA
) {
2358 if (flags
& HAMMER2_LOOKUP_NODIRECT
) {
2360 *key_nextp
= key_end
+ 1;
2363 hammer2_chain_ref(parent
);
2364 if ((flags
& HAMMER2_LOOKUP_NOLOCK
) == 0)
2365 hammer2_chain_lock(parent
,
2367 HAMMER2_RESOLVE_LOCKAGAIN
);
2368 *key_nextp
= key_end
+ 1;
2371 base
= &parent
->data
->ipdata
.u
.blockset
.blockref
[0];
2372 count
= HAMMER2_SET_COUNT
;
2374 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
2375 case HAMMER2_BREF_TYPE_INDIRECT
:
2377 * Handle MATCHIND on the parent
2379 if (flags
& HAMMER2_LOOKUP_MATCHIND
) {
2380 scan_beg
= parent
->bref
.key
;
2381 scan_end
= scan_beg
+
2382 ((hammer2_key_t
)1 << parent
->bref
.keybits
) - 1;
2383 if (key_beg
== scan_beg
&& key_end
== scan_end
) {
2385 hammer2_chain_ref(chain
);
2386 hammer2_chain_lock(chain
, how_maybe
);
2387 *key_nextp
= scan_end
+ 1;
2393 * Optimize indirect blocks in the INITIAL state to avoid
2396 if (parent
->flags
& HAMMER2_CHAIN_INITIAL
) {
2399 if (parent
->data
== NULL
) {
2400 kprintf("parent->data is NULL %p\n", parent
);
2402 tsleep(parent
, 0, "xxx", 0);
2404 base
= &parent
->data
->npdata
[0];
2406 count
= parent
->bytes
/ sizeof(hammer2_blockref_t
);
2408 case HAMMER2_BREF_TYPE_VOLUME
:
2409 base
= &parent
->data
->voldata
.sroot_blockset
.blockref
[0];
2410 count
= HAMMER2_SET_COUNT
;
2412 case HAMMER2_BREF_TYPE_FREEMAP
:
2413 base
= &parent
->data
->blkset
.blockref
[0];
2414 count
= HAMMER2_SET_COUNT
;
2417 kprintf("hammer2_chain_lookup: unrecognized "
2418 "blockref(B) type: %d",
2421 tsleep(&base
, 0, "dead", 0);
2422 panic("hammer2_chain_lookup: unrecognized "
2423 "blockref(B) type: %d",
2425 base
= NULL
; /* safety */
2426 count
= 0; /* safety */
2430 * No lookup is possible if the parent is errored. We delayed
2431 * this check as long as we could to ensure that the parent backup,
2432 * embedded data, and MATCHIND code could still execute.
2434 if (parent
->error
) {
2435 *errorp
= parent
->error
;
2440 * Merged scan to find next candidate.
2442 * hammer2_base_*() functions require the parent->core.live_* fields
2443 * to be synchronized.
2445 * We need to hold the spinlock to access the block array and RB tree
2446 * and to interlock chain creation.
2448 if ((parent
->flags
& HAMMER2_CHAIN_COUNTEDBREFS
) == 0)
2449 hammer2_chain_countbrefs(parent
, base
, count
);
2454 hammer2_spin_ex(&parent
->core
.spin
);
2455 chain
= hammer2_combined_find(parent
, base
, count
,
2459 generation
= parent
->core
.generation
;
2462 * Exhausted parent chain, iterate.
2465 hammer2_spin_unex(&parent
->core
.spin
);
2466 if (key_beg
== key_end
) /* short cut single-key case */
2470 * Stop if we reached the end of the iteration.
2472 if (parent
->bref
.type
!= HAMMER2_BREF_TYPE_INDIRECT
&&
2473 parent
->bref
.type
!= HAMMER2_BREF_TYPE_FREEMAP_NODE
) {
2478 * Calculate next key, stop if we reached the end of the
2479 * iteration, otherwise go up one level and loop.
2481 key_beg
= parent
->bref
.key
+
2482 ((hammer2_key_t
)1 << parent
->bref
.keybits
);
2483 if (key_beg
== 0 || key_beg
> key_end
)
2485 parent
= hammer2_chain_repparent(parentp
, how_maybe
);
2490 * Selected from blockref or in-memory chain.
2492 if (chain
== NULL
) {
2494 hammer2_spin_unex(&parent
->core
.spin
);
2495 chain
= hammer2_chain_get(parent
, generation
,
2497 if (chain
== NULL
) {
2499 kprintf("retry lookup parent %p keys %016jx:%016jx\n",
2500 parent, key_beg, key_end);
2504 if (bcmp(&bcopy
, bref
, sizeof(bcopy
))) {
2505 hammer2_chain_drop(chain
);
2509 hammer2_chain_ref(chain
);
2510 hammer2_spin_unex(&parent
->core
.spin
);
2514 * chain is referenced but not locked. We must lock the chain
2515 * to obtain definitive state.
2517 if (chain
->bref
.type
== HAMMER2_BREF_TYPE_INDIRECT
||
2518 chain
->bref
.type
== HAMMER2_BREF_TYPE_FREEMAP_NODE
) {
2519 hammer2_chain_lock(chain
, how_maybe
);
2521 hammer2_chain_lock(chain
, how
);
2523 KKASSERT(chain
->parent
== parent
);
2526 * Skip deleted chains (XXX cache 'i' end-of-block-array? XXX)
2528 * NOTE: Chain's key range is not relevant as there might be
2529 * one-offs within the range that are not deleted.
2531 * NOTE: Lookups can race delete-duplicate because
2532 * delete-duplicate does not lock the parent's core
2533 * (they just use the spinlock on the core).
2535 if (chain
->flags
& HAMMER2_CHAIN_DELETED
) {
2536 kprintf("skip deleted chain %016jx.%02x key=%016jx\n",
2537 chain
->bref
.data_off
, chain
->bref
.type
,
2539 hammer2_chain_unlock(chain
);
2540 hammer2_chain_drop(chain
);
2541 key_beg
= *key_nextp
;
2542 if (key_beg
== 0 || key_beg
> key_end
)
2548 * If the chain element is an indirect block it becomes the new
2549 * parent and we loop on it. We must maintain our top-down locks
2550 * to prevent the flusher from interfering (i.e. doing a
2551 * delete-duplicate and leaving us recursing down a deleted chain).
2553 * The parent always has to be locked with at least RESOLVE_MAYBE
2554 * so we can access its data. It might need a fixup if the caller
2555 * passed incompatible flags. Be careful not to cause a deadlock
2556 * as a data-load requires an exclusive lock.
2558 * If HAMMER2_LOOKUP_MATCHIND is set and the indirect block's key
2559 * range is within the requested key range we return the indirect
2560 * block and do NOT loop. This is usually only used to acquire
2563 if (chain
->bref
.type
== HAMMER2_BREF_TYPE_INDIRECT
||
2564 chain
->bref
.type
== HAMMER2_BREF_TYPE_FREEMAP_NODE
) {
2565 hammer2_chain_unlock(parent
);
2566 hammer2_chain_drop(parent
);
2567 *parentp
= parent
= chain
;
2572 * All done, return the chain.
2574 * If the caller does not want a locked chain, replace the lock with
2575 * a ref. Perhaps this can eventually be optimized to not obtain the
2576 * lock in the first place for situations where the data does not
2577 * need to be resolved.
2579 * NOTE! A chain->error must be tested by the caller upon return.
2580 * *errorp is only set based on issues which occur while
2581 * trying to reach the chain.
2584 if (flags
& HAMMER2_LOOKUP_NOLOCK
)
2585 hammer2_chain_unlock(chain
);
2591 * After having issued a lookup we can iterate all matching keys.
2593 * If chain is non-NULL we continue the iteration from just after it's index.
2595 * If chain is NULL we assume the parent was exhausted and continue the
2596 * iteration at the next parent.
2598 * If a fatal error occurs (typically an I/O error), a dummy chain is
2599 * returned with chain->error and error-identifying information set. This
2600 * chain will assert if you try to do anything fancy with it.
2602 * XXX Depending on where the error occurs we should allow continued iteration.
2604 * parent must be locked on entry and remains locked throughout. chain's
2605 * lock status must match flags. Chain is always at least referenced.
2607 * WARNING! The MATCHIND flag does not apply to this function.
2610 hammer2_chain_next(hammer2_chain_t
**parentp
, hammer2_chain_t
*chain
,
2611 hammer2_key_t
*key_nextp
,
2612 hammer2_key_t key_beg
, hammer2_key_t key_end
,
2613 int *errorp
, int flags
)
2615 hammer2_chain_t
*parent
;
2619 * Calculate locking flags for upward recursion.
2621 how_maybe
= HAMMER2_RESOLVE_MAYBE
;
2622 if (flags
& HAMMER2_LOOKUP_SHARED
)
2623 how_maybe
|= HAMMER2_RESOLVE_SHARED
;
2629 * Calculate the next index and recalculate the parent if necessary.
2632 key_beg
= chain
->bref
.key
+
2633 ((hammer2_key_t
)1 << chain
->bref
.keybits
);
2634 if ((flags
& (HAMMER2_LOOKUP_NOLOCK
|
2635 HAMMER2_LOOKUP_NOUNLOCK
)) == 0) {
2636 hammer2_chain_unlock(chain
);
2638 hammer2_chain_drop(chain
);
2641 * chain invalid past this point, but we can still do a
2642 * pointer comparison w/parent.
2644 * Any scan where the lookup returned degenerate data embedded
2645 * in the inode has an invalid index and must terminate.
2647 if (chain
== parent
)
2649 if (key_beg
== 0 || key_beg
> key_end
)
2652 } else if (parent
->bref
.type
!= HAMMER2_BREF_TYPE_INDIRECT
&&
2653 parent
->bref
.type
!= HAMMER2_BREF_TYPE_FREEMAP_NODE
) {
2655 * We reached the end of the iteration.
2660 * Continue iteration with next parent unless the current
2661 * parent covers the range.
2663 * (This also handles the case of a deleted, empty indirect
2666 key_beg
= parent
->bref
.key
+
2667 ((hammer2_key_t
)1 << parent
->bref
.keybits
);
2668 if (key_beg
== 0 || key_beg
> key_end
)
2670 parent
= hammer2_chain_repparent(parentp
, how_maybe
);
2676 return (hammer2_chain_lookup(parentp
, key_nextp
,
2682 * Caller wishes to iterate chains under parent, loading new chains into
2683 * chainp. Caller must initialize *chainp to NULL and *firstp to 1, and
2684 * then call hammer2_chain_scan() repeatedly until a non-zero return.
2685 * During the scan, *firstp will be set to 0 and (*chainp) will be replaced
2686 * with the returned chain for the scan. The returned *chainp will be
2687 * locked and referenced. Any prior contents will be unlocked and dropped.
2689 * Caller should check the return value. A normal scan EOF will return
2690 * exactly HAMMER2_ERRORF_EOF. Any other non-zero value indicates an
2691 * error trying to access parent data. Any error in the returned chain
2692 * must be tested separately by the caller.
2694 * (*chainp) is dropped on each scan, but will only be set if the returned
2695 * element itself can recurse. Leaf elements are NOT resolved, loaded, or
2696 * returned via *chainp. The caller will get their bref only.
2698 * The raw scan function is similar to lookup/next but does not seek to a key.
2699 * Blockrefs are iterated via first_bref = (parent, NULL) and
2700 * next_chain = (parent, bref).
2702 * The passed-in parent must be locked and its data resolved. The function
2703 * nominally returns a locked and referenced *chainp != NULL for chains
2704 * the caller might need to recurse on (and will dipose of any *chainp passed
2705 * in). The caller must check the chain->bref.type either way.
2708 hammer2_chain_scan(hammer2_chain_t
*parent
, hammer2_chain_t
**chainp
,
2709 hammer2_blockref_t
*bref
, int *firstp
,
2713 hammer2_blockref_t
*base
;
2714 hammer2_blockref_t
*bref_ptr
;
2716 hammer2_key_t next_key
;
2717 hammer2_chain_t
*chain
= NULL
;
2719 int how_always
= HAMMER2_RESOLVE_ALWAYS
;
2720 int how_maybe
= HAMMER2_RESOLVE_MAYBE
;
2723 int maxloops
= 300000;
2730 * Scan flags borrowed from lookup.
2732 if (flags
& HAMMER2_LOOKUP_ALWAYS
) {
2733 how_maybe
= how_always
;
2734 how
= HAMMER2_RESOLVE_ALWAYS
;
2735 } else if (flags
& (HAMMER2_LOOKUP_NODATA
| HAMMER2_LOOKUP_NOLOCK
)) {
2736 how
= HAMMER2_RESOLVE_NEVER
;
2738 how
= HAMMER2_RESOLVE_MAYBE
;
2740 if (flags
& HAMMER2_LOOKUP_SHARED
) {
2741 how_maybe
|= HAMMER2_RESOLVE_SHARED
;
2742 how_always
|= HAMMER2_RESOLVE_SHARED
;
2743 how
|= HAMMER2_RESOLVE_SHARED
;
2747 * Calculate key to locate first/next element, unlocking the previous
2748 * element as we go. Be careful, the key calculation can overflow.
2750 * (also reset bref to NULL)
2756 key
= bref
->key
+ ((hammer2_key_t
)1 << bref
->keybits
);
2757 if ((chain
= *chainp
) != NULL
) {
2759 hammer2_chain_unlock(chain
);
2760 hammer2_chain_drop(chain
);
2764 error
|= HAMMER2_ERROR_EOF
;
2770 if (parent
->error
) {
2771 error
= parent
->error
;
2774 if (--maxloops
== 0)
2775 panic("hammer2_chain_scan: maxloops");
2778 * Locate the blockref array. Currently we do a fully associative
2779 * search through the array.
2781 switch(parent
->bref
.type
) {
2782 case HAMMER2_BREF_TYPE_INODE
:
2784 * An inode with embedded data has no sub-chains.
2786 * WARNING! Bulk scan code may pass a static chain marked
2787 * as BREF_TYPE_INODE with a copy of the volume
2788 * root blockset to snapshot the volume.
2790 if (parent
->data
->ipdata
.meta
.op_flags
&
2791 HAMMER2_OPFLAG_DIRECTDATA
) {
2792 error
|= HAMMER2_ERROR_EOF
;
2795 base
= &parent
->data
->ipdata
.u
.blockset
.blockref
[0];
2796 count
= HAMMER2_SET_COUNT
;
2798 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
2799 case HAMMER2_BREF_TYPE_INDIRECT
:
2801 * Optimize indirect blocks in the INITIAL state to avoid
2804 if (parent
->flags
& HAMMER2_CHAIN_INITIAL
) {
2807 if (parent
->data
== NULL
)
2808 panic("parent->data is NULL");
2809 base
= &parent
->data
->npdata
[0];
2811 count
= parent
->bytes
/ sizeof(hammer2_blockref_t
);
2813 case HAMMER2_BREF_TYPE_VOLUME
:
2814 base
= &parent
->data
->voldata
.sroot_blockset
.blockref
[0];
2815 count
= HAMMER2_SET_COUNT
;
2817 case HAMMER2_BREF_TYPE_FREEMAP
:
2818 base
= &parent
->data
->blkset
.blockref
[0];
2819 count
= HAMMER2_SET_COUNT
;
2822 panic("hammer2_chain_lookup: unrecognized blockref type: %d",
2824 base
= NULL
; /* safety */
2825 count
= 0; /* safety */
2829 * Merged scan to find next candidate.
2831 * hammer2_base_*() functions require the parent->core.live_* fields
2832 * to be synchronized.
2834 * We need to hold the spinlock to access the block array and RB tree
2835 * and to interlock chain creation.
2837 if ((parent
->flags
& HAMMER2_CHAIN_COUNTEDBREFS
) == 0)
2838 hammer2_chain_countbrefs(parent
, base
, count
);
2842 hammer2_spin_ex(&parent
->core
.spin
);
2843 chain
= hammer2_combined_find(parent
, base
, count
,
2845 key
, HAMMER2_KEY_MAX
,
2847 generation
= parent
->core
.generation
;
2850 * Exhausted parent chain, we're done.
2852 if (bref_ptr
== NULL
) {
2853 hammer2_spin_unex(&parent
->core
.spin
);
2854 KKASSERT(chain
== NULL
);
2855 error
|= HAMMER2_ERROR_EOF
;
2860 * Copy into the supplied stack-based blockref.
2865 * Selected from blockref or in-memory chain.
2867 if (chain
== NULL
) {
2868 switch(bref
->type
) {
2869 case HAMMER2_BREF_TYPE_INODE
:
2870 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
2871 case HAMMER2_BREF_TYPE_INDIRECT
:
2872 case HAMMER2_BREF_TYPE_VOLUME
:
2873 case HAMMER2_BREF_TYPE_FREEMAP
:
2875 * Recursion, always get the chain
2877 hammer2_spin_unex(&parent
->core
.spin
);
2878 chain
= hammer2_chain_get(parent
, generation
, bref
);
2879 if (chain
== NULL
) {
2880 kprintf("retry scan parent %p keys %016jx\n",
2884 if (bcmp(bref
, bref_ptr
, sizeof(*bref
))) {
2885 hammer2_chain_drop(chain
);
2892 * No recursion, do not waste time instantiating
2893 * a chain, just iterate using the bref.
2895 hammer2_spin_unex(&parent
->core
.spin
);
2900 * Recursion or not we need the chain in order to supply
2903 hammer2_chain_ref(chain
);
2904 hammer2_spin_unex(&parent
->core
.spin
);
2908 * chain is referenced but not locked. We must lock the chain
2909 * to obtain definitive state.
2912 hammer2_chain_lock(chain
, how
);
2915 * Skip deleted chains (XXX cache 'i' end-of-block-array? XXX)
2917 * NOTE: chain's key range is not relevant as there might be
2918 * one-offs within the range that are not deleted.
2920 * NOTE: XXX this could create problems with scans used in
2921 * situations other than mount-time recovery.
2923 * NOTE: Lookups can race delete-duplicate because
2924 * delete-duplicate does not lock the parent's core
2925 * (they just use the spinlock on the core).
2927 if (chain
&& (chain
->flags
& HAMMER2_CHAIN_DELETED
)) {
2928 hammer2_chain_unlock(chain
);
2929 hammer2_chain_drop(chain
);
2934 error
|= HAMMER2_ERROR_EOF
;
2942 * All done, return the bref or NULL, supply chain if necessary.
2950 * Create and return a new hammer2 system memory structure of the specified
2951 * key, type and size and insert it under (*parentp). This is a full
2952 * insertion, based on the supplied key/keybits, and may involve creating
2953 * indirect blocks and moving other chains around via delete/duplicate.
2955 * THE CALLER MUST HAVE ALREADY PROPERLY SEEKED (*parentp) TO THE INSERTION
2956 * POINT SANS ANY REQUIRED INDIRECT BLOCK CREATIONS DUE TO THE ARRAY BEING
2957 * FULL. This typically means that the caller is creating the chain after
2958 * doing a hammer2_chain_lookup().
2960 * (*parentp) must be exclusive locked and may be replaced on return
2961 * depending on how much work the function had to do.
2963 * (*parentp) must not be errored or this function will assert.
2965 * (*chainp) usually starts out NULL and returns the newly created chain,
2966 * but if the caller desires the caller may allocate a disconnected chain
2967 * and pass it in instead.
2969 * This function should NOT be used to insert INDIRECT blocks. It is
2970 * typically used to create/insert inodes and data blocks.
2972 * Caller must pass-in an exclusively locked parent the new chain is to
2973 * be inserted under, and optionally pass-in a disconnected, exclusively
2974 * locked chain to insert (else we create a new chain). The function will
2975 * adjust (*parentp) as necessary, create or connect the chain, and
2976 * return an exclusively locked chain in *chainp.
2978 * When creating a PFSROOT inode under the super-root, pmp is typically NULL
2979 * and will be reassigned.
2981 * NOTE: returns HAMMER_ERROR_* flags
2984 hammer2_chain_create(hammer2_chain_t
**parentp
, hammer2_chain_t
**chainp
,
2985 hammer2_pfs_t
*pmp
, int methods
,
2986 hammer2_key_t key
, int keybits
, int type
, size_t bytes
,
2987 hammer2_tid_t mtid
, hammer2_off_t dedup_off
, int flags
)
2990 hammer2_chain_t
*chain
;
2991 hammer2_chain_t
*parent
;
2992 hammer2_blockref_t
*base
;
2993 hammer2_blockref_t dummy
;
2997 int maxloops
= 300000;
3000 * Topology may be crossing a PFS boundary.
3003 KKASSERT(hammer2_mtx_owned(&parent
->lock
));
3004 KKASSERT(parent
->error
== 0);
3008 if (chain
== NULL
) {
3010 * First allocate media space and construct the dummy bref,
3011 * then allocate the in-memory chain structure. Set the
3012 * INITIAL flag for fresh chains which do not have embedded
3015 * XXX for now set the check mode of the child based on
3016 * the parent or, if the parent is an inode, the
3017 * specification in the inode.
3019 bzero(&dummy
, sizeof(dummy
));
3022 dummy
.keybits
= keybits
;
3023 dummy
.data_off
= hammer2_getradix(bytes
);
3026 * Inherit methods from parent by default. Primarily used
3027 * for BREF_TYPE_DATA. Non-data types *must* be set to
3028 * a non-NONE check algorithm.
3031 dummy
.methods
= parent
->bref
.methods
;
3033 dummy
.methods
= (uint8_t)methods
;
3035 if (type
!= HAMMER2_BREF_TYPE_DATA
&&
3036 HAMMER2_DEC_CHECK(dummy
.methods
) == HAMMER2_CHECK_NONE
) {
3038 HAMMER2_ENC_CHECK(HAMMER2_CHECK_DEFAULT
);
3041 chain
= hammer2_chain_alloc(hmp
, pmp
, &dummy
);
3044 * Lock the chain manually, chain_lock will load the chain
3045 * which we do NOT want to do. (note: chain->refs is set
3046 * to 1 by chain_alloc() for us, but lockcnt is not).
3049 hammer2_mtx_ex(&chain
->lock
);
3051 ++curthread
->td_tracker
;
3054 * Set INITIAL to optimize I/O. The flag will generally be
3055 * processed when we call hammer2_chain_modify().
3057 * Recalculate bytes to reflect the actual media block
3058 * allocation. Handle special case radix 0 == 0 bytes.
3060 bytes
= (size_t)(chain
->bref
.data_off
& HAMMER2_OFF_MASK_RADIX
);
3062 bytes
= (hammer2_off_t
)1 << bytes
;
3063 chain
->bytes
= bytes
;
3066 case HAMMER2_BREF_TYPE_VOLUME
:
3067 case HAMMER2_BREF_TYPE_FREEMAP
:
3068 panic("hammer2_chain_create: called with volume type");
3070 case HAMMER2_BREF_TYPE_INDIRECT
:
3071 panic("hammer2_chain_create: cannot be used to"
3072 "create indirect block");
3074 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
3075 panic("hammer2_chain_create: cannot be used to"
3076 "create freemap root or node");
3078 case HAMMER2_BREF_TYPE_FREEMAP_LEAF
:
3079 KKASSERT(bytes
== sizeof(chain
->data
->bmdata
));
3081 case HAMMER2_BREF_TYPE_DIRENT
:
3082 case HAMMER2_BREF_TYPE_INODE
:
3083 case HAMMER2_BREF_TYPE_DATA
:
3086 * leave chain->data NULL, set INITIAL
3088 KKASSERT(chain
->data
== NULL
);
3089 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_INITIAL
);
3094 * We are reattaching a previously deleted chain, possibly
3095 * under a new parent and possibly with a new key/keybits.
3096 * The chain does not have to be in a modified state. The
3097 * UPDATE flag will be set later on in this routine.
3099 * Do NOT mess with the current state of the INITIAL flag.
3101 chain
->bref
.key
= key
;
3102 chain
->bref
.keybits
= keybits
;
3103 if (chain
->flags
& HAMMER2_CHAIN_DELETED
)
3104 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_DELETED
);
3105 KKASSERT(chain
->parent
== NULL
);
3107 if (flags
& HAMMER2_INSERT_PFSROOT
)
3108 chain
->bref
.flags
|= HAMMER2_BREF_FLAG_PFSROOT
;
3110 chain
->bref
.flags
&= ~HAMMER2_BREF_FLAG_PFSROOT
;
3113 * Calculate how many entries we have in the blockref array and
3114 * determine if an indirect block is required.
3117 if (--maxloops
== 0)
3118 panic("hammer2_chain_create: maxloops");
3120 switch(parent
->bref
.type
) {
3121 case HAMMER2_BREF_TYPE_INODE
:
3122 if ((parent
->data
->ipdata
.meta
.op_flags
&
3123 HAMMER2_OPFLAG_DIRECTDATA
) != 0) {
3124 kprintf("hammer2: parent set for direct-data! "
3125 "pkey=%016jx ckey=%016jx\n",
3129 KKASSERT((parent
->data
->ipdata
.meta
.op_flags
&
3130 HAMMER2_OPFLAG_DIRECTDATA
) == 0);
3131 KKASSERT(parent
->data
!= NULL
);
3132 base
= &parent
->data
->ipdata
.u
.blockset
.blockref
[0];
3133 count
= HAMMER2_SET_COUNT
;
3135 case HAMMER2_BREF_TYPE_INDIRECT
:
3136 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
3137 if (parent
->flags
& HAMMER2_CHAIN_INITIAL
)
3140 base
= &parent
->data
->npdata
[0];
3141 count
= parent
->bytes
/ sizeof(hammer2_blockref_t
);
3143 case HAMMER2_BREF_TYPE_VOLUME
:
3144 KKASSERT(parent
->data
!= NULL
);
3145 base
= &parent
->data
->voldata
.sroot_blockset
.blockref
[0];
3146 count
= HAMMER2_SET_COUNT
;
3148 case HAMMER2_BREF_TYPE_FREEMAP
:
3149 KKASSERT(parent
->data
!= NULL
);
3150 base
= &parent
->data
->blkset
.blockref
[0];
3151 count
= HAMMER2_SET_COUNT
;
3154 panic("hammer2_chain_create: unrecognized blockref type: %d",
3162 * Make sure we've counted the brefs
3164 if ((parent
->flags
& HAMMER2_CHAIN_COUNTEDBREFS
) == 0)
3165 hammer2_chain_countbrefs(parent
, base
, count
);
3167 KASSERT(parent
->core
.live_count
>= 0 &&
3168 parent
->core
.live_count
<= count
,
3169 ("bad live_count %d/%d (%02x, %d)",
3170 parent
->core
.live_count
, count
,
3171 parent
->bref
.type
, parent
->bytes
));
3174 * If no free blockref could be found we must create an indirect
3175 * block and move a number of blockrefs into it. With the parent
3176 * locked we can safely lock each child in order to delete+duplicate
3177 * it without causing a deadlock.
3179 * This may return the new indirect block or the old parent depending
3180 * on where the key falls. NULL is returned on error.
3182 if (parent
->core
.live_count
== count
) {
3183 hammer2_chain_t
*nparent
;
3185 nparent
= hammer2_chain_create_indirect(parent
, key
, keybits
,
3186 mtid
, type
, &error
);
3187 if (nparent
== NULL
) {
3189 hammer2_chain_drop(chain
);
3193 if (parent
!= nparent
) {
3194 hammer2_chain_unlock(parent
);
3195 hammer2_chain_drop(parent
);
3196 parent
= *parentp
= nparent
;
3201 if (chain
->flags
& HAMMER2_CHAIN_DELETED
)
3202 kprintf("Inserting deleted chain @%016jx\n",
3206 * Link the chain into its parent.
3208 if (chain
->parent
!= NULL
)
3209 panic("hammer2: hammer2_chain_create: chain already connected");
3210 KKASSERT(chain
->parent
== NULL
);
3211 hammer2_chain_insert(parent
, chain
,
3212 HAMMER2_CHAIN_INSERT_SPIN
|
3213 HAMMER2_CHAIN_INSERT_LIVE
,
3218 * Mark the newly created chain modified. This will cause
3219 * UPDATE to be set and process the INITIAL flag.
3221 * Device buffers are not instantiated for DATA elements
3222 * as these are handled by logical buffers.
3224 * Indirect and freemap node indirect blocks are handled
3225 * by hammer2_chain_create_indirect() and not by this
3228 * Data for all other bref types is expected to be
3229 * instantiated (INODE, LEAF).
3231 switch(chain
->bref
.type
) {
3232 case HAMMER2_BREF_TYPE_DATA
:
3233 case HAMMER2_BREF_TYPE_FREEMAP_LEAF
:
3234 case HAMMER2_BREF_TYPE_DIRENT
:
3235 case HAMMER2_BREF_TYPE_INODE
:
3236 hammer2_chain_modify(chain
, mtid
, dedup_off
,
3237 HAMMER2_MODIFY_OPTDATA
);
3241 * Remaining types are not supported by this function.
3242 * In particular, INDIRECT and LEAF_NODE types are
3243 * handled by create_indirect().
3245 panic("hammer2_chain_create: bad type: %d",
3252 * When reconnecting a chain we must set UPDATE and
3253 * setflush so the flush recognizes that it must update
3254 * the bref in the parent.
3256 if ((chain
->flags
& HAMMER2_CHAIN_UPDATE
) == 0)
3257 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_UPDATE
);
3261 * We must setflush(parent) to ensure that it recurses through to
3262 * chain. setflush(chain) might not work because ONFLUSH is possibly
3263 * already set in the chain (so it won't recurse up to set it in the
3266 hammer2_chain_setflush(parent
);
3275 * Move the chain from its old parent to a new parent. The chain must have
3276 * already been deleted or already disconnected (or never associated) with
3277 * a parent. The chain is reassociated with the new parent and the deleted
3278 * flag will be cleared (no longer deleted). The chain's modification state
3281 * THE CALLER MUST HAVE ALREADY PROPERLY SEEKED (parent) TO THE INSERTION
3282 * POINT SANS ANY REQUIRED INDIRECT BLOCK CREATIONS DUE TO THE ARRAY BEING
3283 * FULL. This typically means that the caller is creating the chain after
3284 * doing a hammer2_chain_lookup().
3286 * A non-NULL bref is typically passed when key and keybits must be overridden.
3287 * Note that hammer2_cluster_duplicate() *ONLY* uses the key and keybits fields
3288 * from a passed-in bref and uses the old chain's bref for everything else.
3290 * Neither (parent) or (chain) can be errored.
3292 * If (parent) is non-NULL then the chain is inserted under the parent.
3294 * If (parent) is NULL then the newly duplicated chain is not inserted
3295 * anywhere, similar to if it had just been chain_alloc()'d (suitable for
3296 * passing into hammer2_chain_create() after this function returns).
3298 * WARNING! This function calls create which means it can insert indirect
3299 * blocks. This can cause other unrelated chains in the parent to
3300 * be moved to a newly inserted indirect block in addition to the
3304 hammer2_chain_rename(hammer2_blockref_t
*bref
,
3305 hammer2_chain_t
**parentp
, hammer2_chain_t
*chain
,
3306 hammer2_tid_t mtid
, int flags
)
3309 hammer2_chain_t
*parent
;
3313 * WARNING! We should never resolve DATA to device buffers
3314 * (XXX allow it if the caller did?), and since
3315 * we currently do not have the logical buffer cache
3316 * buffer in-hand to fix its cached physical offset
3317 * we also force the modify code to not COW it. XXX
3320 KKASSERT(chain
->parent
== NULL
);
3321 KKASSERT(chain
->error
== 0);
3324 * Now create a duplicate of the chain structure, associating
3325 * it with the same core, making it the same size, pointing it
3326 * to the same bref (the same media block).
3328 * NOTE: Handle special radix == 0 case (means 0 bytes).
3331 bref
= &chain
->bref
;
3332 bytes
= (size_t)(bref
->data_off
& HAMMER2_OFF_MASK_RADIX
);
3334 bytes
= (hammer2_off_t
)1 << bytes
;
3337 * If parent is not NULL the duplicated chain will be entered under
3338 * the parent and the UPDATE bit set to tell flush to update
3341 * We must setflush(parent) to ensure that it recurses through to
3342 * chain. setflush(chain) might not work because ONFLUSH is possibly
3343 * already set in the chain (so it won't recurse up to set it in the
3346 * Having both chains locked is extremely important for atomicy.
3348 if (parentp
&& (parent
= *parentp
) != NULL
) {
3349 KKASSERT(hammer2_mtx_owned(&parent
->lock
));
3350 KKASSERT(parent
->refs
> 0);
3351 KKASSERT(parent
->error
== 0);
3353 hammer2_chain_create(parentp
, &chain
,
3354 chain
->pmp
, HAMMER2_METH_DEFAULT
,
3355 bref
->key
, bref
->keybits
, bref
->type
,
3356 chain
->bytes
, mtid
, 0, flags
);
3357 KKASSERT(chain
->flags
& HAMMER2_CHAIN_UPDATE
);
3358 hammer2_chain_setflush(*parentp
);
3363 * Helper function for deleting chains.
3365 * The chain is removed from the live view (the RBTREE) as well as the parent's
3366 * blockmap. Both chain and its parent must be locked.
3368 * parent may not be errored. chain can be errored.
3371 _hammer2_chain_delete_helper(hammer2_chain_t
*parent
, hammer2_chain_t
*chain
,
3372 hammer2_tid_t mtid
, int flags
)
3376 KKASSERT((chain
->flags
& (HAMMER2_CHAIN_DELETED
|
3377 HAMMER2_CHAIN_FICTITIOUS
)) == 0);
3378 KKASSERT(chain
->parent
== parent
);
3381 if (chain
->flags
& HAMMER2_CHAIN_BMAPPED
) {
3383 * Chain is blockmapped, so there must be a parent.
3384 * Atomically remove the chain from the parent and remove
3385 * the blockmap entry. The parent must be set modified
3386 * to remove the blockmap entry.
3388 hammer2_blockref_t
*base
;
3391 KKASSERT(parent
!= NULL
);
3392 KKASSERT(parent
->error
== 0);
3393 KKASSERT((parent
->flags
& HAMMER2_CHAIN_INITIAL
) == 0);
3394 hammer2_chain_modify(parent
, mtid
, 0, 0);
3397 * Calculate blockmap pointer
3399 KKASSERT(chain
->flags
& HAMMER2_CHAIN_ONRBTREE
);
3400 hammer2_spin_ex(&chain
->core
.spin
);
3401 hammer2_spin_ex(&parent
->core
.spin
);
3403 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_DELETED
);
3404 atomic_add_int(&parent
->core
.live_count
, -1);
3405 ++parent
->core
.generation
;
3406 RB_REMOVE(hammer2_chain_tree
, &parent
->core
.rbtree
, chain
);
3407 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_ONRBTREE
);
3408 --parent
->core
.chain_count
;
3409 chain
->parent
= NULL
;
3411 switch(parent
->bref
.type
) {
3412 case HAMMER2_BREF_TYPE_INODE
:
3414 * Access the inode's block array. However, there
3415 * is no block array if the inode is flagged
3419 (parent
->data
->ipdata
.meta
.op_flags
&
3420 HAMMER2_OPFLAG_DIRECTDATA
) == 0) {
3422 &parent
->data
->ipdata
.u
.blockset
.blockref
[0];
3426 count
= HAMMER2_SET_COUNT
;
3428 case HAMMER2_BREF_TYPE_INDIRECT
:
3429 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
3431 base
= &parent
->data
->npdata
[0];
3434 count
= parent
->bytes
/ sizeof(hammer2_blockref_t
);
3436 case HAMMER2_BREF_TYPE_VOLUME
:
3437 base
= &parent
->data
->voldata
.
3438 sroot_blockset
.blockref
[0];
3439 count
= HAMMER2_SET_COUNT
;
3441 case HAMMER2_BREF_TYPE_FREEMAP
:
3442 base
= &parent
->data
->blkset
.blockref
[0];
3443 count
= HAMMER2_SET_COUNT
;
3448 panic("hammer2_flush_pass2: "
3449 "unrecognized blockref type: %d",
3454 * delete blockmapped chain from its parent.
3456 * The parent is not affected by any statistics in chain
3457 * which are pending synchronization. That is, there is
3458 * nothing to undo in the parent since they have not yet
3459 * been incorporated into the parent.
3461 * The parent is affected by statistics stored in inodes.
3462 * Those have already been synchronized, so they must be
3463 * undone. XXX split update possible w/delete in middle?
3466 hammer2_base_delete(parent
, base
, count
, chain
);
3468 hammer2_spin_unex(&parent
->core
.spin
);
3469 hammer2_spin_unex(&chain
->core
.spin
);
3470 } else if (chain
->flags
& HAMMER2_CHAIN_ONRBTREE
) {
3472 * Chain is not blockmapped but a parent is present.
3473 * Atomically remove the chain from the parent. There is
3474 * no blockmap entry to remove.
3476 * Because chain was associated with a parent but not
3477 * synchronized, the chain's *_count_up fields contain
3478 * inode adjustment statistics which must be undone.
3480 hammer2_spin_ex(&chain
->core
.spin
);
3481 hammer2_spin_ex(&parent
->core
.spin
);
3482 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_DELETED
);
3483 atomic_add_int(&parent
->core
.live_count
, -1);
3484 ++parent
->core
.generation
;
3485 RB_REMOVE(hammer2_chain_tree
, &parent
->core
.rbtree
, chain
);
3486 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_ONRBTREE
);
3487 --parent
->core
.chain_count
;
3488 chain
->parent
= NULL
;
3489 hammer2_spin_unex(&parent
->core
.spin
);
3490 hammer2_spin_unex(&chain
->core
.spin
);
3493 * Chain is not blockmapped and has no parent. This
3494 * is a degenerate case.
3496 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_DELETED
);
3501 * Create an indirect block that covers one or more of the elements in the
3502 * current parent. Either returns the existing parent with no locking or
3503 * ref changes or returns the new indirect block locked and referenced
3504 * and leaving the original parent lock/ref intact as well.
3506 * If an error occurs, NULL is returned and *errorp is set to the error.
3508 * The returned chain depends on where the specified key falls.
3510 * The key/keybits for the indirect mode only needs to follow three rules:
3512 * (1) That all elements underneath it fit within its key space and
3514 * (2) That all elements outside it are outside its key space.
3516 * (3) When creating the new indirect block any elements in the current
3517 * parent that fit within the new indirect block's keyspace must be
3518 * moved into the new indirect block.
3520 * (4) The keyspace chosen for the inserted indirect block CAN cover a wider
3521 * keyspace the the current parent, but lookup/iteration rules will
3522 * ensure (and must ensure) that rule (2) for all parents leading up
3523 * to the nearest inode or the root volume header is adhered to. This
3524 * is accomplished by always recursing through matching keyspaces in
3525 * the hammer2_chain_lookup() and hammer2_chain_next() API.
3527 * The current implementation calculates the current worst-case keyspace by
3528 * iterating the current parent and then divides it into two halves, choosing
3529 * whichever half has the most elements (not necessarily the half containing
3530 * the requested key).
3532 * We can also opt to use the half with the least number of elements. This
3533 * causes lower-numbered keys (aka logical file offsets) to recurse through
3534 * fewer indirect blocks and higher-numbered keys to recurse through more.
3535 * This also has the risk of not moving enough elements to the new indirect
3536 * block and being forced to create several indirect blocks before the element
3539 * Must be called with an exclusively locked parent.
3541 * NOTE: *errorp set to HAMMER_ERROR_* flags
3543 static int hammer2_chain_indkey_freemap(hammer2_chain_t
*parent
,
3544 hammer2_key_t
*keyp
, int keybits
,
3545 hammer2_blockref_t
*base
, int count
);
3546 static int hammer2_chain_indkey_file(hammer2_chain_t
*parent
,
3547 hammer2_key_t
*keyp
, int keybits
,
3548 hammer2_blockref_t
*base
, int count
,
3550 static int hammer2_chain_indkey_dir(hammer2_chain_t
*parent
,
3551 hammer2_key_t
*keyp
, int keybits
,
3552 hammer2_blockref_t
*base
, int count
,
3556 hammer2_chain_create_indirect(hammer2_chain_t
*parent
,
3557 hammer2_key_t create_key
, int create_bits
,
3558 hammer2_tid_t mtid
, int for_type
, int *errorp
)
3561 hammer2_blockref_t
*base
;
3562 hammer2_blockref_t
*bref
;
3563 hammer2_blockref_t bcopy
;
3564 hammer2_chain_t
*chain
;
3565 hammer2_chain_t
*ichain
;
3566 hammer2_chain_t dummy
;
3567 hammer2_key_t key
= create_key
;
3568 hammer2_key_t key_beg
;
3569 hammer2_key_t key_end
;
3570 hammer2_key_t key_next
;
3571 int keybits
= create_bits
;
3578 int maxloops
= 300000;
3581 * Calculate the base blockref pointer or NULL if the chain
3582 * is known to be empty. We need to calculate the array count
3583 * for RB lookups either way.
3587 KKASSERT(hammer2_mtx_owned(&parent
->lock
));
3589 /*hammer2_chain_modify(&parent, HAMMER2_MODIFY_OPTDATA);*/
3590 base
= hammer2_chain_base_and_count(parent
, &count
);
3593 * dummy used in later chain allocation (no longer used for lookups).
3595 bzero(&dummy
, sizeof(dummy
));
3598 * How big should our new indirect block be? It has to be at least
3599 * as large as its parent for splits to work properly.
3601 * The freemap uses a specific indirect block size. The number of
3602 * levels are built dynamically and ultimately depend on the size
3603 * volume. Because freemap blocks are taken from the reserved areas
3604 * of the volume our goal is efficiency (fewer levels) and not so
3605 * much to save disk space.
3607 * The first indirect block level for a directory usually uses
3608 * HAMMER2_IND_BYTES_MIN (4KB = 32 directory entries). Due to
3609 * the hash mechanism, this typically gives us a nominal
3610 * 32 * 4 entries with one level of indirection.
3612 * We use HAMMER2_IND_BYTES_NOM (16KB = 128 blockrefs) for FILE
3613 * indirect blocks. The initial 4 entries in the inode gives us
3614 * 256KB. Up to 4 indirect blocks gives us 32MB. Three levels
3615 * of indirection gives us 137GB, and so forth. H2 can support
3616 * huge file sizes but they are not typical, so we try to stick
3617 * with compactness and do not use a larger indirect block size.
3619 * We could use 64KB (PBUFSIZE), giving us 512 blockrefs, but
3620 * due to the way indirect blocks are created this usually winds
3621 * up being extremely inefficient for small files. Even though
3622 * 16KB requires more levels of indirection for very large files,
3623 * the 16KB records can be ganged together into 64KB DIOs.
3625 if (for_type
== HAMMER2_BREF_TYPE_FREEMAP_NODE
||
3626 for_type
== HAMMER2_BREF_TYPE_FREEMAP_LEAF
) {
3627 nbytes
= HAMMER2_FREEMAP_LEVELN_PSIZE
;
3628 } else if (parent
->bref
.type
== HAMMER2_BREF_TYPE_INODE
) {
3629 if (parent
->data
->ipdata
.meta
.type
==
3630 HAMMER2_OBJTYPE_DIRECTORY
)
3631 nbytes
= HAMMER2_IND_BYTES_MIN
; /* 4KB = 32 entries */
3633 nbytes
= HAMMER2_IND_BYTES_NOM
; /* 16KB = ~8MB file */
3636 nbytes
= HAMMER2_IND_BYTES_NOM
;
3638 if (nbytes
< count
* sizeof(hammer2_blockref_t
)) {
3639 KKASSERT(for_type
!= HAMMER2_BREF_TYPE_FREEMAP_NODE
&&
3640 for_type
!= HAMMER2_BREF_TYPE_FREEMAP_LEAF
);
3641 nbytes
= count
* sizeof(hammer2_blockref_t
);
3643 ncount
= nbytes
/ sizeof(hammer2_blockref_t
);
3646 * When creating an indirect block for a freemap node or leaf
3647 * the key/keybits must be fitted to static radix levels because
3648 * particular radix levels use particular reserved blocks in the
3651 * This routine calculates the key/radix of the indirect block
3652 * we need to create, and whether it is on the high-side or the
3656 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
3657 case HAMMER2_BREF_TYPE_FREEMAP_LEAF
:
3658 keybits
= hammer2_chain_indkey_freemap(parent
, &key
, keybits
,
3661 case HAMMER2_BREF_TYPE_DATA
:
3662 keybits
= hammer2_chain_indkey_file(parent
, &key
, keybits
,
3663 base
, count
, ncount
);
3665 case HAMMER2_BREF_TYPE_DIRENT
:
3666 case HAMMER2_BREF_TYPE_INODE
:
3667 keybits
= hammer2_chain_indkey_dir(parent
, &key
, keybits
,
3668 base
, count
, ncount
);
3671 panic("illegal indirect block for bref type %d", for_type
);
3676 * Normalize the key for the radix being represented, keeping the
3677 * high bits and throwing away the low bits.
3679 key
&= ~(((hammer2_key_t
)1 << keybits
) - 1);
3682 * Ok, create our new indirect block
3684 if (for_type
== HAMMER2_BREF_TYPE_FREEMAP_NODE
||
3685 for_type
== HAMMER2_BREF_TYPE_FREEMAP_LEAF
) {
3686 dummy
.bref
.type
= HAMMER2_BREF_TYPE_FREEMAP_NODE
;
3688 dummy
.bref
.type
= HAMMER2_BREF_TYPE_INDIRECT
;
3690 dummy
.bref
.key
= key
;
3691 dummy
.bref
.keybits
= keybits
;
3692 dummy
.bref
.data_off
= hammer2_getradix(nbytes
);
3693 dummy
.bref
.methods
=
3694 HAMMER2_ENC_CHECK(HAMMER2_DEC_CHECK(parent
->bref
.methods
)) |
3695 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE
);
3697 ichain
= hammer2_chain_alloc(hmp
, parent
->pmp
, &dummy
.bref
);
3698 atomic_set_int(&ichain
->flags
, HAMMER2_CHAIN_INITIAL
);
3699 hammer2_chain_lock(ichain
, HAMMER2_RESOLVE_MAYBE
);
3700 /* ichain has one ref at this point */
3703 * We have to mark it modified to allocate its block, but use
3704 * OPTDATA to allow it to remain in the INITIAL state. Otherwise
3705 * it won't be acted upon by the flush code.
3707 hammer2_chain_modify(ichain
, mtid
, 0, HAMMER2_MODIFY_OPTDATA
);
3710 * Iterate the original parent and move the matching brefs into
3711 * the new indirect block.
3713 * XXX handle flushes.
3716 key_end
= HAMMER2_KEY_MAX
;
3717 key_next
= 0; /* avoid gcc warnings */
3718 hammer2_spin_ex(&parent
->core
.spin
);
3724 * Parent may have been modified, relocating its block array.
3725 * Reload the base pointer.
3727 base
= hammer2_chain_base_and_count(parent
, &count
);
3729 if (++loops
> 100000) {
3730 hammer2_spin_unex(&parent
->core
.spin
);
3731 panic("excessive loops r=%d p=%p base/count %p:%d %016jx\n",
3732 reason
, parent
, base
, count
, key_next
);
3736 * NOTE: spinlock stays intact, returned chain (if not NULL)
3737 * is not referenced or locked which means that we
3738 * cannot safely check its flagged / deletion status
3741 chain
= hammer2_combined_find(parent
, base
, count
,
3745 generation
= parent
->core
.generation
;
3748 key_next
= bref
->key
+ ((hammer2_key_t
)1 << bref
->keybits
);
3751 * Skip keys that are not within the key/radix of the new
3752 * indirect block. They stay in the parent.
3754 if ((~(((hammer2_key_t
)1 << keybits
) - 1) &
3755 (key
^ bref
->key
)) != 0) {
3756 goto next_key_spinlocked
;
3760 * Load the new indirect block by acquiring the related
3761 * chains (potentially from media as it might not be
3762 * in-memory). Then move it to the new parent (ichain).
3764 * chain is referenced but not locked. We must lock the
3765 * chain to obtain definitive state.
3769 * Use chain already present in the RBTREE
3771 hammer2_chain_ref(chain
);
3772 hammer2_spin_unex(&parent
->core
.spin
);
3773 hammer2_chain_lock(chain
, HAMMER2_RESOLVE_NEVER
);
3776 * Get chain for blockref element. _get returns NULL
3777 * on insertion race.
3780 hammer2_spin_unex(&parent
->core
.spin
);
3781 chain
= hammer2_chain_get(parent
, generation
, &bcopy
);
3782 if (chain
== NULL
) {
3784 hammer2_spin_ex(&parent
->core
.spin
);
3787 if (bcmp(&bcopy
, bref
, sizeof(bcopy
))) {
3788 kprintf("REASON 2\n");
3790 hammer2_chain_drop(chain
);
3791 hammer2_spin_ex(&parent
->core
.spin
);
3794 hammer2_chain_lock(chain
, HAMMER2_RESOLVE_NEVER
);
3798 * This is always live so if the chain has been deleted
3799 * we raced someone and we have to retry.
3801 * NOTE: Lookups can race delete-duplicate because
3802 * delete-duplicate does not lock the parent's core
3803 * (they just use the spinlock on the core).
3805 * (note reversed logic for this one)
3807 if (chain
->parent
!= parent
||
3808 (chain
->flags
& HAMMER2_CHAIN_DELETED
)) {
3809 hammer2_chain_unlock(chain
);
3810 hammer2_chain_drop(chain
);
3811 kprintf("hammer2_chain_create_indirect "
3812 "RETRY (%p,%p)->%p %08x\n",
3813 parent
, chain
->parent
, chain
, chain
->flags
);
3814 hammer2_spin_ex(&parent
->core
.spin
);
3819 * Shift the chain to the indirect block.
3821 * WARNING! No reason for us to load chain data, pass NOSTATS
3822 * to prevent delete/insert from trying to access
3823 * inode stats (and thus asserting if there is no
3824 * chain->data loaded).
3826 * WARNING! The (parent, chain) deletion may modify the parent
3827 * and invalidate the base pointer.
3829 hammer2_chain_delete(parent
, chain
, mtid
, 0);
3830 hammer2_chain_rename(NULL
, &ichain
, chain
, mtid
, 0);
3831 hammer2_chain_unlock(chain
);
3832 hammer2_chain_drop(chain
);
3833 KKASSERT(parent
->refs
> 0);
3835 base
= NULL
; /* safety */
3836 hammer2_spin_ex(&parent
->core
.spin
);
3837 next_key_spinlocked
:
3838 if (--maxloops
== 0)
3839 panic("hammer2_chain_create_indirect: maxloops");
3841 if (key_next
== 0 || key_next
> key_end
)
3846 hammer2_spin_unex(&parent
->core
.spin
);
3849 * Insert the new indirect block into the parent now that we've
3850 * cleared out some entries in the parent. We calculated a good
3851 * insertion index in the loop above (ichain->index).
3853 * We don't have to set UPDATE here because we mark ichain
3854 * modified down below (so the normal modified -> flush -> set-moved
3855 * sequence applies).
3857 * The insertion shouldn't race as this is a completely new block
3858 * and the parent is locked.
3860 base
= NULL
; /* safety, parent modify may change address */
3861 KKASSERT((ichain
->flags
& HAMMER2_CHAIN_ONRBTREE
) == 0);
3862 hammer2_chain_insert(parent
, ichain
,
3863 HAMMER2_CHAIN_INSERT_SPIN
|
3864 HAMMER2_CHAIN_INSERT_LIVE
,
3868 * Make sure flushes propogate after our manual insertion.
3870 hammer2_chain_setflush(ichain
);
3871 hammer2_chain_setflush(parent
);
3874 * Figure out what to return.
3876 if (~(((hammer2_key_t
)1 << keybits
) - 1) &
3877 (create_key
^ key
)) {
3879 * Key being created is outside the key range,
3880 * return the original parent.
3882 hammer2_chain_unlock(ichain
);
3883 hammer2_chain_drop(ichain
);
3886 * Otherwise its in the range, return the new parent.
3887 * (leave both the new and old parent locked).
3896 * Freemap indirect blocks
3898 * Calculate the keybits and highside/lowside of the freemap node the
3899 * caller is creating.
3901 * This routine will specify the next higher-level freemap key/radix
3902 * representing the lowest-ordered set. By doing so, eventually all
3903 * low-ordered sets will be moved one level down.
3905 * We have to be careful here because the freemap reserves a limited
3906 * number of blocks for a limited number of levels. So we can't just
3907 * push indiscriminately.
3910 hammer2_chain_indkey_freemap(hammer2_chain_t
*parent
, hammer2_key_t
*keyp
,
3911 int keybits
, hammer2_blockref_t
*base
, int count
)
3913 hammer2_chain_t
*chain
;
3914 hammer2_blockref_t
*bref
;
3916 hammer2_key_t key_beg
;
3917 hammer2_key_t key_end
;
3918 hammer2_key_t key_next
;
3921 int maxloops
= 300000;
3929 * Calculate the range of keys in the array being careful to skip
3930 * slots which are overridden with a deletion.
3933 key_end
= HAMMER2_KEY_MAX
;
3934 hammer2_spin_ex(&parent
->core
.spin
);
3937 if (--maxloops
== 0) {
3938 panic("indkey_freemap shit %p %p:%d\n",
3939 parent
, base
, count
);
3941 chain
= hammer2_combined_find(parent
, base
, count
,
3953 * Skip deleted chains.
3955 if (chain
&& (chain
->flags
& HAMMER2_CHAIN_DELETED
)) {
3956 if (key_next
== 0 || key_next
> key_end
)
3963 * Use the full live (not deleted) element for the scan
3964 * iteration. HAMMER2 does not allow partial replacements.
3966 * XXX should be built into hammer2_combined_find().
3968 key_next
= bref
->key
+ ((hammer2_key_t
)1 << bref
->keybits
);
3970 if (keybits
> bref
->keybits
) {
3972 keybits
= bref
->keybits
;
3973 } else if (keybits
== bref
->keybits
&& bref
->key
< key
) {
3980 hammer2_spin_unex(&parent
->core
.spin
);
3983 * Return the keybits for a higher-level FREEMAP_NODE covering
3987 case HAMMER2_FREEMAP_LEVEL0_RADIX
:
3988 keybits
= HAMMER2_FREEMAP_LEVEL1_RADIX
;
3990 case HAMMER2_FREEMAP_LEVEL1_RADIX
:
3991 keybits
= HAMMER2_FREEMAP_LEVEL2_RADIX
;
3993 case HAMMER2_FREEMAP_LEVEL2_RADIX
:
3994 keybits
= HAMMER2_FREEMAP_LEVEL3_RADIX
;
3996 case HAMMER2_FREEMAP_LEVEL3_RADIX
:
3997 keybits
= HAMMER2_FREEMAP_LEVEL4_RADIX
;
3999 case HAMMER2_FREEMAP_LEVEL4_RADIX
:
4000 keybits
= HAMMER2_FREEMAP_LEVEL5_RADIX
;
4002 case HAMMER2_FREEMAP_LEVEL5_RADIX
:
4003 panic("hammer2_chain_indkey_freemap: level too high");
4006 panic("hammer2_chain_indkey_freemap: bad radix");
4015 * File indirect blocks
4017 * Calculate the key/keybits for the indirect block to create by scanning
4018 * existing keys. The key being created is also passed in *keyp and can be
4019 * inside or outside the indirect block. Regardless, the indirect block
4020 * must hold at least two keys in order to guarantee sufficient space.
4022 * We use a modified version of the freemap's fixed radix tree, but taylored
4023 * for file data. Basically we configure an indirect block encompassing the
4027 hammer2_chain_indkey_file(hammer2_chain_t
*parent
, hammer2_key_t
*keyp
,
4028 int keybits
, hammer2_blockref_t
*base
, int count
,
4031 hammer2_chain_t
*chain
;
4032 hammer2_blockref_t
*bref
;
4034 hammer2_key_t key_beg
;
4035 hammer2_key_t key_end
;
4036 hammer2_key_t key_next
;
4040 int maxloops
= 300000;
4048 * Calculate the range of keys in the array being careful to skip
4049 * slots which are overridden with a deletion.
4051 * Locate the smallest key.
4054 key_end
= HAMMER2_KEY_MAX
;
4055 hammer2_spin_ex(&parent
->core
.spin
);
4058 if (--maxloops
== 0) {
4059 panic("indkey_freemap shit %p %p:%d\n",
4060 parent
, base
, count
);
4062 chain
= hammer2_combined_find(parent
, base
, count
,
4074 * Skip deleted chains.
4076 if (chain
&& (chain
->flags
& HAMMER2_CHAIN_DELETED
)) {
4077 if (key_next
== 0 || key_next
> key_end
)
4084 * Use the full live (not deleted) element for the scan
4085 * iteration. HAMMER2 does not allow partial replacements.
4087 * XXX should be built into hammer2_combined_find().
4089 key_next
= bref
->key
+ ((hammer2_key_t
)1 << bref
->keybits
);
4091 if (keybits
> bref
->keybits
) {
4093 keybits
= bref
->keybits
;
4094 } else if (keybits
== bref
->keybits
&& bref
->key
< key
) {
4101 hammer2_spin_unex(&parent
->core
.spin
);
4104 * Calculate the static keybits for a higher-level indirect block
4105 * that contains the key.
4110 case HAMMER2_IND_BYTES_MIN
/ sizeof(hammer2_blockref_t
):
4111 nradix
= HAMMER2_IND_RADIX_MIN
- HAMMER2_BLOCKREF_RADIX
;
4113 case HAMMER2_IND_BYTES_NOM
/ sizeof(hammer2_blockref_t
):
4114 nradix
= HAMMER2_IND_RADIX_NOM
- HAMMER2_BLOCKREF_RADIX
;
4116 case HAMMER2_IND_BYTES_MAX
/ sizeof(hammer2_blockref_t
):
4117 nradix
= HAMMER2_IND_RADIX_MAX
- HAMMER2_BLOCKREF_RADIX
;
4120 panic("bad ncount %d\n", ncount
);
4126 * The largest radix that can be returned for an indirect block is
4127 * 63 bits. (The largest practical indirect block radix is actually
4128 * 62 bits because the top-level inode or volume root contains four
4129 * entries, but allow 63 to be returned).
4134 return keybits
+ nradix
;
4140 * Directory indirect blocks.
4142 * Covers both the inode index (directory of inodes), and directory contents
4143 * (filenames hardlinked to inodes).
4145 * Because directory keys are hashed we generally try to cut the space in
4146 * half. We accomodate the inode index (which tends to have linearly
4147 * increasing inode numbers) by ensuring that the keyspace is at least large
4148 * enough to fill up the indirect block being created.
4151 hammer2_chain_indkey_dir(hammer2_chain_t
*parent
, hammer2_key_t
*keyp
,
4152 int keybits
, hammer2_blockref_t
*base
, int count
,
4155 hammer2_blockref_t
*bref
;
4156 hammer2_chain_t
*chain
;
4157 hammer2_key_t key_beg
;
4158 hammer2_key_t key_end
;
4159 hammer2_key_t key_next
;
4164 int maxloops
= 300000;
4167 * Shortcut if the parent is the inode. In this situation the
4168 * parent has 4+1 directory entries and we are creating an indirect
4169 * block capable of holding many more.
4171 if (parent
->bref
.type
== HAMMER2_BREF_TYPE_INODE
) {
4180 * Calculate the range of keys in the array being careful to skip
4181 * slots which are overridden with a deletion.
4184 key_end
= HAMMER2_KEY_MAX
;
4185 hammer2_spin_ex(&parent
->core
.spin
);
4188 if (--maxloops
== 0) {
4189 panic("indkey_freemap shit %p %p:%d\n",
4190 parent
, base
, count
);
4192 chain
= hammer2_combined_find(parent
, base
, count
,
4206 if (chain
&& (chain
->flags
& HAMMER2_CHAIN_DELETED
)) {
4207 if (key_next
== 0 || key_next
> key_end
)
4214 * Use the full live (not deleted) element for the scan
4215 * iteration. HAMMER2 does not allow partial replacements.
4217 * XXX should be built into hammer2_combined_find().
4219 key_next
= bref
->key
+ ((hammer2_key_t
)1 << bref
->keybits
);
4222 * Expand our calculated key range (key, keybits) to fit
4223 * the scanned key. nkeybits represents the full range
4224 * that we will later cut in half (two halves @ nkeybits - 1).
4227 if (nkeybits
< bref
->keybits
) {
4228 if (bref
->keybits
> 64) {
4229 kprintf("bad bref chain %p bref %p\n",
4233 nkeybits
= bref
->keybits
;
4235 while (nkeybits
< 64 &&
4236 (~(((hammer2_key_t
)1 << nkeybits
) - 1) &
4237 (key
^ bref
->key
)) != 0) {
4242 * If the new key range is larger we have to determine
4243 * which side of the new key range the existing keys fall
4244 * under by checking the high bit, then collapsing the
4245 * locount into the hicount or vise-versa.
4247 if (keybits
!= nkeybits
) {
4248 if (((hammer2_key_t
)1 << (nkeybits
- 1)) & key
) {
4259 * The newly scanned key will be in the lower half or the
4260 * upper half of the (new) key range.
4262 if (((hammer2_key_t
)1 << (nkeybits
- 1)) & bref
->key
)
4271 hammer2_spin_unex(&parent
->core
.spin
);
4272 bref
= NULL
; /* now invalid (safety) */
4275 * Adjust keybits to represent half of the full range calculated
4276 * above (radix 63 max) for our new indirect block.
4281 * Expand keybits to hold at least ncount elements. ncount will be
4282 * a power of 2. This is to try to completely fill leaf nodes (at
4283 * least for keys which are not hashes).
4285 * We aren't counting 'in' or 'out', we are counting 'high side'
4286 * and 'low side' based on the bit at (1LL << keybits). We want
4287 * everything to be inside in these cases so shift it all to
4288 * the low or high side depending on the new high bit.
4290 while (((hammer2_key_t
)1 << keybits
) < ncount
) {
4292 if (key
& ((hammer2_key_t
)1 << keybits
)) {
4301 if (hicount
> locount
)
4302 key
|= (hammer2_key_t
)1 << keybits
;
4304 key
&= ~(hammer2_key_t
)1 << keybits
;
4314 * Directory indirect blocks.
4316 * Covers both the inode index (directory of inodes), and directory contents
4317 * (filenames hardlinked to inodes).
4319 * Because directory keys are hashed we generally try to cut the space in
4320 * half. We accomodate the inode index (which tends to have linearly
4321 * increasing inode numbers) by ensuring that the keyspace is at least large
4322 * enough to fill up the indirect block being created.
4325 hammer2_chain_indkey_dir(hammer2_chain_t
*parent
, hammer2_key_t
*keyp
,
4326 int keybits
, hammer2_blockref_t
*base
, int count
,
4329 hammer2_blockref_t
*bref
;
4330 hammer2_chain_t
*chain
;
4331 hammer2_key_t key_beg
;
4332 hammer2_key_t key_end
;
4333 hammer2_key_t key_next
;
4338 int maxloops
= 300000;
4341 * Shortcut if the parent is the inode. In this situation the
4342 * parent has 4+1 directory entries and we are creating an indirect
4343 * block capable of holding many more.
4345 if (parent
->bref
.type
== HAMMER2_BREF_TYPE_INODE
) {
4354 * Calculate the range of keys in the array being careful to skip
4355 * slots which are overridden with a deletion.
4358 key_end
= HAMMER2_KEY_MAX
;
4359 hammer2_spin_ex(&parent
->core
.spin
);
4362 if (--maxloops
== 0) {
4363 panic("indkey_freemap shit %p %p:%d\n",
4364 parent
, base
, count
);
4366 chain
= hammer2_combined_find(parent
, base
, count
,
4380 if (chain
&& (chain
->flags
& HAMMER2_CHAIN_DELETED
)) {
4381 if (key_next
== 0 || key_next
> key_end
)
4388 * Use the full live (not deleted) element for the scan
4389 * iteration. HAMMER2 does not allow partial replacements.
4391 * XXX should be built into hammer2_combined_find().
4393 key_next
= bref
->key
+ ((hammer2_key_t
)1 << bref
->keybits
);
4396 * Expand our calculated key range (key, keybits) to fit
4397 * the scanned key. nkeybits represents the full range
4398 * that we will later cut in half (two halves @ nkeybits - 1).
4401 if (nkeybits
< bref
->keybits
) {
4402 if (bref
->keybits
> 64) {
4403 kprintf("bad bref chain %p bref %p\n",
4407 nkeybits
= bref
->keybits
;
4409 while (nkeybits
< 64 &&
4410 (~(((hammer2_key_t
)1 << nkeybits
) - 1) &
4411 (key
^ bref
->key
)) != 0) {
4416 * If the new key range is larger we have to determine
4417 * which side of the new key range the existing keys fall
4418 * under by checking the high bit, then collapsing the
4419 * locount into the hicount or vise-versa.
4421 if (keybits
!= nkeybits
) {
4422 if (((hammer2_key_t
)1 << (nkeybits
- 1)) & key
) {
4433 * The newly scanned key will be in the lower half or the
4434 * upper half of the (new) key range.
4436 if (((hammer2_key_t
)1 << (nkeybits
- 1)) & bref
->key
)
4445 hammer2_spin_unex(&parent
->core
.spin
);
4446 bref
= NULL
; /* now invalid (safety) */
4449 * Adjust keybits to represent half of the full range calculated
4450 * above (radix 63 max) for our new indirect block.
4455 * Expand keybits to hold at least ncount elements. ncount will be
4456 * a power of 2. This is to try to completely fill leaf nodes (at
4457 * least for keys which are not hashes).
4459 * We aren't counting 'in' or 'out', we are counting 'high side'
4460 * and 'low side' based on the bit at (1LL << keybits). We want
4461 * everything to be inside in these cases so shift it all to
4462 * the low or high side depending on the new high bit.
4464 while (((hammer2_key_t
)1 << keybits
) < ncount
) {
4466 if (key
& ((hammer2_key_t
)1 << keybits
)) {
4475 if (hicount
> locount
)
4476 key
|= (hammer2_key_t
)1 << keybits
;
4478 key
&= ~(hammer2_key_t
)1 << keybits
;
4488 * Sets CHAIN_DELETED and remove the chain's blockref from the parent if
4491 * Both parent and chain must be locked exclusively.
4493 * This function will modify the parent if the blockref requires removal
4494 * from the parent's block table.
4496 * This function is NOT recursive. Any entity already pushed into the
4497 * chain (such as an inode) may still need visibility into its contents,
4498 * as well as the ability to read and modify the contents. For example,
4499 * for an unlinked file which is still open.
4501 * Also note that the flusher is responsible for cleaning up empty
4505 hammer2_chain_delete(hammer2_chain_t
*parent
, hammer2_chain_t
*chain
,
4506 hammer2_tid_t mtid
, int flags
)
4508 KKASSERT(hammer2_mtx_owned(&chain
->lock
));
4511 * Nothing to do if already marked.
4513 * We need the spinlock on the core whos RBTREE contains chain
4514 * to protect against races.
4516 if ((chain
->flags
& HAMMER2_CHAIN_DELETED
) == 0) {
4517 KKASSERT((chain
->flags
& HAMMER2_CHAIN_DELETED
) == 0 &&
4518 chain
->parent
== parent
);
4519 _hammer2_chain_delete_helper(parent
, chain
, mtid
, flags
);
4523 * Permanent deletions mark the chain as destroyed.
4525 if (flags
& HAMMER2_DELETE_PERMANENT
)
4526 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_DESTROY
);
4527 hammer2_chain_setflush(chain
);
4531 * Returns the index of the nearest element in the blockref array >= elm.
4532 * Returns (count) if no element could be found.
4534 * Sets *key_nextp to the next key for loop purposes but does not modify
4535 * it if the next key would be higher than the current value of *key_nextp.
4536 * Note that *key_nexp can overflow to 0, which should be tested by the
4539 * WARNING! Must be called with parent's spinlock held. Spinlock remains
4540 * held through the operation.
4543 hammer2_base_find(hammer2_chain_t
*parent
,
4544 hammer2_blockref_t
*base
, int count
,
4545 hammer2_key_t
*key_nextp
,
4546 hammer2_key_t key_beg
, hammer2_key_t key_end
)
4548 hammer2_blockref_t
*scan
;
4549 hammer2_key_t scan_end
;
4554 * Require the live chain's already have their core's counted
4555 * so we can optimize operations.
4557 KKASSERT(parent
->flags
& HAMMER2_CHAIN_COUNTEDBREFS
);
4562 if (count
== 0 || base
== NULL
)
4566 * Sequential optimization using parent->cache_index. This is
4567 * the most likely scenario.
4569 * We can avoid trailing empty entries on live chains, otherwise
4570 * we might have to check the whole block array.
4572 i
= parent
->cache_index
; /* SMP RACE OK */
4574 limit
= parent
->core
.live_zero
;
4579 KKASSERT(i
< count
);
4585 while (i
> 0 && (scan
->type
== 0 || scan
->key
> key_beg
)) {
4589 parent
->cache_index
= i
;
4592 * Search forwards, stop when we find a scan element which
4593 * encloses the key or until we know that there are no further
4597 if (scan
->type
!= 0) {
4598 scan_end
= scan
->key
+
4599 ((hammer2_key_t
)1 << scan
->keybits
) - 1;
4600 if (scan
->key
> key_beg
|| scan_end
>= key_beg
)
4609 parent
->cache_index
= i
;
4613 scan_end
= scan
->key
+
4614 ((hammer2_key_t
)1 << scan
->keybits
);
4615 if (scan_end
&& (*key_nextp
> scan_end
||
4617 *key_nextp
= scan_end
;
4625 * Do a combined search and return the next match either from the blockref
4626 * array or from the in-memory chain. Sets *bresp to the returned bref in
4627 * both cases, or sets it to NULL if the search exhausted. Only returns
4628 * a non-NULL chain if the search matched from the in-memory chain.
4630 * When no in-memory chain has been found and a non-NULL bref is returned
4634 * The returned chain is not locked or referenced. Use the returned bref
4635 * to determine if the search exhausted or not. Iterate if the base find
4636 * is chosen but matches a deleted chain.
4638 * WARNING! Must be called with parent's spinlock held. Spinlock remains
4639 * held through the operation.
4641 static hammer2_chain_t
*
4642 hammer2_combined_find(hammer2_chain_t
*parent
,
4643 hammer2_blockref_t
*base
, int count
,
4644 hammer2_key_t
*key_nextp
,
4645 hammer2_key_t key_beg
, hammer2_key_t key_end
,
4646 hammer2_blockref_t
**bresp
)
4648 hammer2_blockref_t
*bref
;
4649 hammer2_chain_t
*chain
;
4653 * Lookup in block array and in rbtree.
4655 *key_nextp
= key_end
+ 1;
4656 i
= hammer2_base_find(parent
, base
, count
, key_nextp
,
4658 chain
= hammer2_chain_find(parent
, key_nextp
, key_beg
, key_end
);
4663 if (i
== count
&& chain
== NULL
) {
4669 * Only chain matched.
4672 bref
= &chain
->bref
;
4677 * Only blockref matched.
4679 if (chain
== NULL
) {
4685 * Both in-memory and blockref matched, select the nearer element.
4687 * If both are flush with the left-hand side or both are the
4688 * same distance away, select the chain. In this situation the
4689 * chain must have been loaded from the matching blockmap.
4691 if ((chain
->bref
.key
<= key_beg
&& base
[i
].key
<= key_beg
) ||
4692 chain
->bref
.key
== base
[i
].key
) {
4693 KKASSERT(chain
->bref
.key
== base
[i
].key
);
4694 bref
= &chain
->bref
;
4699 * Select the nearer key
4701 if (chain
->bref
.key
< base
[i
].key
) {
4702 bref
= &chain
->bref
;
4709 * If the bref is out of bounds we've exhausted our search.
4712 if (bref
->key
> key_end
) {
4722 * Locate the specified block array element and delete it. The element
4725 * The spin lock on the related chain must be held.
4727 * NOTE: live_count was adjusted when the chain was deleted, so it does not
4728 * need to be adjusted when we commit the media change.
4731 hammer2_base_delete(hammer2_chain_t
*parent
,
4732 hammer2_blockref_t
*base
, int count
,
4733 hammer2_chain_t
*chain
)
4735 hammer2_blockref_t
*elm
= &chain
->bref
;
4736 hammer2_blockref_t
*scan
;
4737 hammer2_key_t key_next
;
4741 * Delete element. Expect the element to exist.
4743 * XXX see caller, flush code not yet sophisticated enough to prevent
4744 * re-flushed in some cases.
4746 key_next
= 0; /* max range */
4747 i
= hammer2_base_find(parent
, base
, count
, &key_next
,
4748 elm
->key
, elm
->key
);
4750 if (i
== count
|| scan
->type
== 0 ||
4751 scan
->key
!= elm
->key
||
4752 ((chain
->flags
& HAMMER2_CHAIN_BMAPUPD
) == 0 &&
4753 scan
->keybits
!= elm
->keybits
)) {
4754 hammer2_spin_unex(&parent
->core
.spin
);
4755 panic("delete base %p element not found at %d/%d elm %p\n",
4756 base
, i
, count
, elm
);
4761 * Update stats and zero the entry.
4763 * NOTE: Handle radix == 0 (0 bytes) case.
4765 if ((int)(scan
->data_off
& HAMMER2_OFF_MASK_RADIX
)) {
4766 parent
->bref
.embed
.stats
.data_count
-= (hammer2_off_t
)1 <<
4767 (int)(scan
->data_off
& HAMMER2_OFF_MASK_RADIX
);
4769 switch(scan
->type
) {
4770 case HAMMER2_BREF_TYPE_INODE
:
4771 parent
->bref
.embed
.stats
.inode_count
-= 1;
4773 case HAMMER2_BREF_TYPE_DATA
:
4774 case HAMMER2_BREF_TYPE_INDIRECT
:
4775 parent
->bref
.embed
.stats
.data_count
-=
4776 scan
->embed
.stats
.data_count
;
4777 parent
->bref
.embed
.stats
.inode_count
-=
4778 scan
->embed
.stats
.inode_count
;
4784 bzero(scan
, sizeof(*scan
));
4787 * We can only optimize parent->core.live_zero for live chains.
4789 if (parent
->core
.live_zero
== i
+ 1) {
4790 while (--i
>= 0 && base
[i
].type
== 0)
4792 parent
->core
.live_zero
= i
+ 1;
4796 * Clear appropriate blockmap flags in chain.
4798 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_BMAPPED
|
4799 HAMMER2_CHAIN_BMAPUPD
);
4803 * Insert the specified element. The block array must not already have the
4804 * element and must have space available for the insertion.
4806 * The spin lock on the related chain must be held.
4808 * NOTE: live_count was adjusted when the chain was deleted, so it does not
4809 * need to be adjusted when we commit the media change.
4812 hammer2_base_insert(hammer2_chain_t
*parent
,
4813 hammer2_blockref_t
*base
, int count
,
4814 hammer2_chain_t
*chain
)
4816 hammer2_blockref_t
*elm
= &chain
->bref
;
4817 hammer2_key_t key_next
;
4826 * Insert new element. Expect the element to not already exist
4827 * unless we are replacing it.
4829 * XXX see caller, flush code not yet sophisticated enough to prevent
4830 * re-flushed in some cases.
4832 key_next
= 0; /* max range */
4833 i
= hammer2_base_find(parent
, base
, count
, &key_next
,
4834 elm
->key
, elm
->key
);
4837 * Shortcut fill optimization, typical ordered insertion(s) may not
4840 KKASSERT(i
>= 0 && i
<= count
);
4843 * Set appropriate blockmap flags in chain.
4845 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_BMAPPED
);
4848 * Update stats and zero the entry
4850 if ((int)(elm
->data_off
& HAMMER2_OFF_MASK_RADIX
)) {
4851 parent
->bref
.embed
.stats
.data_count
+= (hammer2_off_t
)1 <<
4852 (int)(elm
->data_off
& HAMMER2_OFF_MASK_RADIX
);
4855 case HAMMER2_BREF_TYPE_INODE
:
4856 parent
->bref
.embed
.stats
.inode_count
+= 1;
4858 case HAMMER2_BREF_TYPE_DATA
:
4859 case HAMMER2_BREF_TYPE_INDIRECT
:
4860 parent
->bref
.embed
.stats
.data_count
+=
4861 elm
->embed
.stats
.data_count
;
4862 parent
->bref
.embed
.stats
.inode_count
+=
4863 elm
->embed
.stats
.inode_count
;
4871 * We can only optimize parent->core.live_zero for live chains.
4873 if (i
== count
&& parent
->core
.live_zero
< count
) {
4874 i
= parent
->core
.live_zero
++;
4879 xkey
= elm
->key
+ ((hammer2_key_t
)1 << elm
->keybits
) - 1;
4880 if (i
!= count
&& (base
[i
].key
< elm
->key
|| xkey
>= base
[i
].key
)) {
4881 hammer2_spin_unex(&parent
->core
.spin
);
4882 panic("insert base %p overlapping elements at %d elm %p\n",
4887 * Try to find an empty slot before or after.
4891 while (j
> 0 || k
< count
) {
4893 if (j
>= 0 && base
[j
].type
== 0) {
4897 bcopy(&base
[j
+1], &base
[j
],
4898 (i
- j
- 1) * sizeof(*base
));
4904 if (k
< count
&& base
[k
].type
== 0) {
4905 bcopy(&base
[i
], &base
[i
+1],
4906 (k
- i
) * sizeof(hammer2_blockref_t
));
4910 * We can only update parent->core.live_zero for live
4913 if (parent
->core
.live_zero
<= k
)
4914 parent
->core
.live_zero
= k
+ 1;
4919 panic("hammer2_base_insert: no room!");
4926 for (l
= 0; l
< count
; ++l
) {
4928 key_next
= base
[l
].key
+
4929 ((hammer2_key_t
)1 << base
[l
].keybits
) - 1;
4933 while (++l
< count
) {
4935 if (base
[l
].key
<= key_next
)
4936 panic("base_insert %d %d,%d,%d fail %p:%d", u
, i
, j
, k
, base
, l
);
4937 key_next
= base
[l
].key
+
4938 ((hammer2_key_t
)1 << base
[l
].keybits
) - 1;
4948 * Sort the blockref array for the chain. Used by the flush code to
4949 * sort the blockref[] array.
4951 * The chain must be exclusively locked AND spin-locked.
4953 typedef hammer2_blockref_t
*hammer2_blockref_p
;
4957 hammer2_base_sort_callback(const void *v1
, const void *v2
)
4959 hammer2_blockref_p bref1
= *(const hammer2_blockref_p
*)v1
;
4960 hammer2_blockref_p bref2
= *(const hammer2_blockref_p
*)v2
;
4963 * Make sure empty elements are placed at the end of the array
4965 if (bref1
->type
== 0) {
4966 if (bref2
->type
== 0)
4969 } else if (bref2
->type
== 0) {
4976 if (bref1
->key
< bref2
->key
)
4978 if (bref1
->key
> bref2
->key
)
4984 hammer2_base_sort(hammer2_chain_t
*chain
)
4986 hammer2_blockref_t
*base
;
4989 switch(chain
->bref
.type
) {
4990 case HAMMER2_BREF_TYPE_INODE
:
4992 * Special shortcut for embedded data returns the inode
4993 * itself. Callers must detect this condition and access
4994 * the embedded data (the strategy code does this for us).
4996 * This is only applicable to regular files and softlinks.
4998 if (chain
->data
->ipdata
.meta
.op_flags
&
4999 HAMMER2_OPFLAG_DIRECTDATA
) {
5002 base
= &chain
->data
->ipdata
.u
.blockset
.blockref
[0];
5003 count
= HAMMER2_SET_COUNT
;
5005 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
5006 case HAMMER2_BREF_TYPE_INDIRECT
:
5008 * Optimize indirect blocks in the INITIAL state to avoid
5011 KKASSERT((chain
->flags
& HAMMER2_CHAIN_INITIAL
) == 0);
5012 base
= &chain
->data
->npdata
[0];
5013 count
= chain
->bytes
/ sizeof(hammer2_blockref_t
);
5015 case HAMMER2_BREF_TYPE_VOLUME
:
5016 base
= &chain
->data
->voldata
.sroot_blockset
.blockref
[0];
5017 count
= HAMMER2_SET_COUNT
;
5019 case HAMMER2_BREF_TYPE_FREEMAP
:
5020 base
= &chain
->data
->blkset
.blockref
[0];
5021 count
= HAMMER2_SET_COUNT
;
5024 kprintf("hammer2_chain_lookup: unrecognized "
5025 "blockref(A) type: %d",
5028 tsleep(&base
, 0, "dead", 0);
5029 panic("hammer2_chain_lookup: unrecognized "
5030 "blockref(A) type: %d",
5032 base
= NULL
; /* safety */
5033 count
= 0; /* safety */
5035 kqsort(base
, count
, sizeof(*base
), hammer2_base_sort_callback
);
5041 * Chain memory management
5044 hammer2_chain_wait(hammer2_chain_t
*chain
)
5046 tsleep(chain
, 0, "chnflw", 1);
5049 const hammer2_media_data_t
*
5050 hammer2_chain_rdata(hammer2_chain_t
*chain
)
5052 KKASSERT(chain
->data
!= NULL
);
5053 return (chain
->data
);
5056 hammer2_media_data_t
*
5057 hammer2_chain_wdata(hammer2_chain_t
*chain
)
5059 KKASSERT(chain
->data
!= NULL
);
5060 return (chain
->data
);
5064 * Set the check data for a chain. This can be a heavy-weight operation
5065 * and typically only runs on-flush. For file data check data is calculated
5066 * when the logical buffers are flushed.
5069 hammer2_chain_setcheck(hammer2_chain_t
*chain
, void *bdata
)
5071 chain
->bref
.flags
&= ~HAMMER2_BREF_FLAG_ZERO
;
5073 switch(HAMMER2_DEC_CHECK(chain
->bref
.methods
)) {
5074 case HAMMER2_CHECK_NONE
:
5076 case HAMMER2_CHECK_DISABLED
:
5078 case HAMMER2_CHECK_ISCSI32
:
5079 chain
->bref
.check
.iscsi32
.value
=
5080 hammer2_icrc32(bdata
, chain
->bytes
);
5082 case HAMMER2_CHECK_XXHASH64
:
5083 chain
->bref
.check
.xxhash64
.value
=
5084 XXH64(bdata
, chain
->bytes
, XXH_HAMMER2_SEED
);
5086 case HAMMER2_CHECK_SHA192
:
5088 SHA256_CTX hash_ctx
;
5090 uint8_t digest
[SHA256_DIGEST_LENGTH
];
5091 uint64_t digest64
[SHA256_DIGEST_LENGTH
/8];
5094 SHA256_Init(&hash_ctx
);
5095 SHA256_Update(&hash_ctx
, bdata
, chain
->bytes
);
5096 SHA256_Final(u
.digest
, &hash_ctx
);
5097 u
.digest64
[2] ^= u
.digest64
[3];
5099 chain
->bref
.check
.sha192
.data
,
5100 sizeof(chain
->bref
.check
.sha192
.data
));
5103 case HAMMER2_CHECK_FREEMAP
:
5104 chain
->bref
.check
.freemap
.icrc32
=
5105 hammer2_icrc32(bdata
, chain
->bytes
);
5108 kprintf("hammer2_chain_setcheck: unknown check type %02x\n",
5109 chain
->bref
.methods
);
5115 hammer2_chain_testcheck(hammer2_chain_t
*chain
, void *bdata
)
5121 if (chain
->bref
.flags
& HAMMER2_BREF_FLAG_ZERO
)
5124 switch(HAMMER2_DEC_CHECK(chain
->bref
.methods
)) {
5125 case HAMMER2_CHECK_NONE
:
5128 case HAMMER2_CHECK_DISABLED
:
5131 case HAMMER2_CHECK_ISCSI32
:
5132 check32
= hammer2_icrc32(bdata
, chain
->bytes
);
5133 r
= (chain
->bref
.check
.iscsi32
.value
== check32
);
5135 kprintf("chain %016jx.%02x meth=%02x CHECK FAIL "
5136 "(flags=%08x, bref/data %08x/%08x)\n",
5137 chain
->bref
.data_off
,
5139 chain
->bref
.methods
,
5141 chain
->bref
.check
.iscsi32
.value
,
5144 hammer2_check_icrc32
+= chain
->bytes
;
5146 case HAMMER2_CHECK_XXHASH64
:
5147 check64
= XXH64(bdata
, chain
->bytes
, XXH_HAMMER2_SEED
);
5148 r
= (chain
->bref
.check
.xxhash64
.value
== check64
);
5150 kprintf("chain %016jx.%02x key=%016jx "
5151 "meth=%02x CHECK FAIL "
5152 "(flags=%08x, bref/data %016jx/%016jx)\n",
5153 chain
->bref
.data_off
,
5156 chain
->bref
.methods
,
5158 chain
->bref
.check
.xxhash64
.value
,
5161 hammer2_check_xxhash64
+= chain
->bytes
;
5163 case HAMMER2_CHECK_SHA192
:
5165 SHA256_CTX hash_ctx
;
5167 uint8_t digest
[SHA256_DIGEST_LENGTH
];
5168 uint64_t digest64
[SHA256_DIGEST_LENGTH
/8];
5171 SHA256_Init(&hash_ctx
);
5172 SHA256_Update(&hash_ctx
, bdata
, chain
->bytes
);
5173 SHA256_Final(u
.digest
, &hash_ctx
);
5174 u
.digest64
[2] ^= u
.digest64
[3];
5176 chain
->bref
.check
.sha192
.data
,
5177 sizeof(chain
->bref
.check
.sha192
.data
)) == 0) {
5181 kprintf("chain %016jx.%02x meth=%02x "
5183 chain
->bref
.data_off
,
5185 chain
->bref
.methods
);
5189 case HAMMER2_CHECK_FREEMAP
:
5190 r
= (chain
->bref
.check
.freemap
.icrc32
==
5191 hammer2_icrc32(bdata
, chain
->bytes
));
5193 kprintf("chain %016jx.%02x meth=%02x "
5195 chain
->bref
.data_off
,
5197 chain
->bref
.methods
);
5198 kprintf("freemap.icrc %08x icrc32 %08x (%d)\n",
5199 chain
->bref
.check
.freemap
.icrc32
,
5200 hammer2_icrc32(bdata
, chain
->bytes
),
5203 kprintf("dio %p buf %016jx,%d bdata %p/%p\n",
5204 chain
->dio
, chain
->dio
->bp
->b_loffset
,
5205 chain
->dio
->bp
->b_bufsize
, bdata
,
5206 chain
->dio
->bp
->b_data
);
5211 kprintf("hammer2_chain_setcheck: unknown check type %02x\n",
5212 chain
->bref
.methods
);
5220 * Acquire the chain and parent representing the specified inode for the
5221 * device at the specified cluster index.
5223 * The flags passed in are LOOKUP flags, not RESOLVE flags.
5225 * If we are unable to locate the hardlink, INVAL is returned and *chainp
5226 * will be NULL. *parentp may still be set error or not, or NULL if the
5227 * parent itself could not be resolved.
5229 * Caller must pass-in a valid or NULL *parentp or *chainp. The passed-in
5230 * *parentp and *chainp will be unlocked if not NULL.
5233 hammer2_chain_inode_find(hammer2_pfs_t
*pmp
, hammer2_key_t inum
,
5234 int clindex
, int flags
,
5235 hammer2_chain_t
**parentp
, hammer2_chain_t
**chainp
)
5237 hammer2_chain_t
*parent
;
5238 hammer2_chain_t
*rchain
;
5239 hammer2_key_t key_dummy
;
5243 resolve_flags
= (flags
& HAMMER2_LOOKUP_SHARED
) ?
5244 HAMMER2_RESOLVE_SHARED
: 0;
5247 * Caller expects us to replace these.
5250 hammer2_chain_unlock(*chainp
);
5251 hammer2_chain_drop(*chainp
);
5255 hammer2_chain_unlock(*parentp
);
5256 hammer2_chain_drop(*parentp
);
5261 * Inodes hang off of the iroot (bit 63 is clear, differentiating
5262 * inodes from root directory entries in the key lookup).
5264 parent
= hammer2_inode_chain(pmp
->iroot
, clindex
, resolve_flags
);
5267 rchain
= hammer2_chain_lookup(&parent
, &key_dummy
,
5271 error
= HAMMER2_ERROR_IO
;
5280 * Used by the bulkscan code to snapshot the synchronized storage for
5281 * a volume, allowing it to be scanned concurrently against normal
5285 hammer2_chain_bulksnap(hammer2_dev_t
*hmp
)
5287 hammer2_chain_t
*copy
;
5289 copy
= hammer2_chain_alloc(hmp
, hmp
->spmp
, &hmp
->vchain
.bref
);
5290 copy
->data
= kmalloc(sizeof(copy
->data
->voldata
),
5293 hammer2_voldata_lock(hmp
);
5294 copy
->data
->voldata
= hmp
->volsync
;
5295 hammer2_voldata_unlock(hmp
);
5301 hammer2_chain_bulkdrop(hammer2_chain_t
*copy
)
5303 KKASSERT(copy
->bref
.type
== HAMMER2_BREF_TYPE_VOLUME
);
5304 KKASSERT(copy
->data
);
5305 kfree(copy
->data
, copy
->hmp
->mchain
);
5307 atomic_add_long(&hammer2_chain_allocs
, -1);
5308 hammer2_chain_drop(copy
);
5312 * Create a snapshot of the specified (chain) with the specified label.
5313 * The originating hammer2_inode must be exclusively locked for
5314 * safety. The device's bulklk should be held by the caller. The caller
5315 * is responsible for synchronizing the filesystem to storage before
5316 * taking the snapshot.
5319 hammer2_chain_snapshot(hammer2_chain_t
*chain
, hammer2_ioc_pfs_t
*pmp
,
5323 const hammer2_inode_data_t
*ripdata
;
5324 hammer2_inode_data_t
*wipdata
;
5325 hammer2_chain_t
*nchain
;
5326 hammer2_inode_t
*nip
;
5335 kprintf("snapshot %s\n", pmp
->name
);
5337 name_len
= strlen(pmp
->name
);
5338 lhc
= hammer2_dirhash(pmp
->name
, name_len
);
5343 ripdata
= &chain
->data
->ipdata
;
5345 opfs_clid
= ripdata
->meta
.pfs_clid
;
5350 * Create the snapshot directory under the super-root
5352 * Set PFS type, generate a unique filesystem id, and generate
5353 * a cluster id. Use the same clid when snapshotting a PFS root,
5354 * which theoretically allows the snapshot to be used as part of
5355 * the same cluster (perhaps as a cache).
5357 * Copy the (flushed) blockref array. Theoretically we could use
5358 * chain_duplicate() but it becomes difficult to disentangle
5359 * the shared core so for now just brute-force it.
5364 hammer2_chain_unlock(chain
);
5365 nip
= hammer2_inode_create(hmp
->spmp
->iroot
, hmp
->spmp
->iroot
,
5366 &vat
, proc0
.p_ucred
,
5367 pmp
->name
, name_len
, 0,
5369 HAMMER2_INSERT_PFSROOT
, &error
);
5370 hammer2_chain_lock(chain
, HAMMER2_RESOLVE_ALWAYS
);
5373 hammer2_inode_modify(nip
);
5374 nchain
= hammer2_inode_chain(nip
, 0, HAMMER2_RESOLVE_ALWAYS
);
5375 hammer2_chain_modify(nchain
, mtid
, 0, 0);
5376 wipdata
= &nchain
->data
->ipdata
;
5378 nip
->meta
.pfs_type
= HAMMER2_PFSTYPE_MASTER
;
5379 nip
->meta
.pfs_subtype
= HAMMER2_PFSSUBTYPE_SNAPSHOT
;
5380 nip
->meta
.op_flags
|= HAMMER2_OPFLAG_PFSROOT
;
5381 kern_uuidgen(&nip
->meta
.pfs_fsid
, 1);
5384 * Give the snapshot its own private cluster id. As a
5385 * snapshot no further synchronization with the original
5386 * cluster will be done.
5389 if (chain
->flags
& HAMMER2_CHAIN_PFSBOUNDARY
)
5390 nip
->meta
.pfs_clid
= opfs_clid
;
5392 kern_uuidgen(&nip
->meta
.pfs_clid
, 1);
5394 kern_uuidgen(&nip
->meta
.pfs_clid
, 1);
5395 nchain
->bref
.flags
|= HAMMER2_BREF_FLAG_PFSROOT
;
5397 /* XXX hack blockset copy */
5398 /* XXX doesn't work with real cluster */
5399 wipdata
->meta
= nip
->meta
;
5400 wipdata
->u
.blockset
= ripdata
->u
.blockset
;
5402 hammer2_flush(nchain
, 1);
5403 KKASSERT(wipdata
== &nchain
->data
->ipdata
);
5404 hammer2_pfsalloc(nchain
, wipdata
, nchain
->bref
.modify_tid
, 0);
5406 hammer2_chain_unlock(nchain
);
5407 hammer2_chain_drop(nchain
);
5408 hammer2_inode_chain_sync(nip
);
5409 hammer2_inode_unlock(nip
);
5410 hammer2_inode_run_sideq(hmp
->spmp
);
5416 * Returns non-zero if the chain (INODE or DIRENT) matches the
5420 hammer2_chain_dirent_test(hammer2_chain_t
*chain
, const char *name
,
5423 const hammer2_inode_data_t
*ripdata
;
5424 const hammer2_dirent_head_t
*den
;
5426 if (chain
->bref
.type
== HAMMER2_BREF_TYPE_INODE
) {
5427 ripdata
= &chain
->data
->ipdata
;
5428 if (ripdata
->meta
.name_len
== name_len
&&
5429 bcmp(ripdata
->filename
, name
, name_len
) == 0) {
5433 if (chain
->bref
.type
== HAMMER2_BREF_TYPE_DIRENT
&&
5434 chain
->bref
.embed
.dirent
.namlen
== name_len
) {
5435 den
= &chain
->bref
.embed
.dirent
;
5436 if (name_len
> sizeof(chain
->bref
.check
.buf
) &&
5437 bcmp(chain
->data
->buf
, name
, name_len
) == 0) {
5440 if (name_len
<= sizeof(chain
->bref
.check
.buf
) &&
5441 bcmp(chain
->bref
.check
.buf
, name
, name_len
) == 0) {