2 * Copyright (c) 2011-2018 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);
79 static struct krate krate_h2me = { .freq = 1 };
82 * Basic RBTree for chains (core->rbtree and core->dbtree). Chains cannot
83 * overlap in the RB trees. Deleted chains are moved from rbtree to either
86 * Chains in delete-duplicate sequences can always iterate through core_entry
87 * to locate the live version of the chain.
89 RB_GENERATE(hammer2_chain_tree, hammer2_chain, rbnode, hammer2_chain_cmp);
92 hammer2_chain_cmp(hammer2_chain_t *chain1, hammer2_chain_t *chain2)
100 * Compare chains. Overlaps are not supposed to happen and catch
101 * any software issues early we count overlaps as a match.
103 c1_beg = chain1->bref.key;
104 c1_end = c1_beg + ((hammer2_key_t)1 << chain1->bref.keybits) - 1;
105 c2_beg = chain2->bref.key;
106 c2_end = c2_beg + ((hammer2_key_t)1 << chain2->bref.keybits) - 1;
108 if (c1_end < c2_beg) /* fully to the left */
110 if (c1_beg > c2_end) /* fully to the right */
112 return(0); /* overlap (must not cross edge boundary) */
116 * Assert that a chain has no media data associated with it.
119 hammer2_chain_assert_no_data(hammer2_chain_t *chain)
121 KKASSERT(chain->dio == NULL);
122 if (chain->bref.type != HAMMER2_BREF_TYPE_VOLUME &&
123 chain->bref.type != HAMMER2_BREF_TYPE_FREEMAP &&
125 panic("hammer2_assert_no_data: chain %p still has data", chain);
130 * Make a chain visible to the flusher. The flusher needs to be able to
131 * do flushes of subdirectory chains or single files so it does a top-down
132 * recursion using the ONFLUSH flag for the recursion. It locates MODIFIED
133 * or UPDATE chains and flushes back up the chain to the volume root.
135 * This routine sets ONFLUSH upward until it hits the volume root. For
136 * simplicity we ignore PFSROOT boundaries whos rules can be complex.
137 * Extra ONFLUSH flagging doesn't hurt the filesystem.
140 hammer2_chain_setflush(hammer2_chain_t *chain)
142 hammer2_chain_t *parent;
144 if ((chain->flags & HAMMER2_CHAIN_ONFLUSH) == 0) {
145 hammer2_spin_sh(&chain->core.spin);
146 while ((chain->flags & HAMMER2_CHAIN_ONFLUSH) == 0) {
147 atomic_set_int(&chain->flags, HAMMER2_CHAIN_ONFLUSH);
148 if ((parent = chain->parent) == NULL)
150 hammer2_spin_sh(&parent->core.spin);
151 hammer2_spin_unsh(&chain->core.spin);
154 hammer2_spin_unsh(&chain->core.spin);
159 * Allocate a new disconnected chain element representing the specified
160 * bref. chain->refs is set to 1 and the passed bref is copied to
161 * chain->bref. chain->bytes is derived from the bref.
163 * chain->pmp inherits pmp unless the chain is an inode (other than the
166 * NOTE: Returns a referenced but unlocked (because there is no core) chain.
169 hammer2_chain_alloc(hammer2_dev_t *hmp, hammer2_pfs_t *pmp,
170 hammer2_blockref_t *bref)
172 hammer2_chain_t *chain;
176 * Special case - radix of 0 indicates a chain that does not
177 * need a data reference (context is completely embedded in the
180 if ((int)(bref->data_off & HAMMER2_OFF_MASK_RADIX))
181 bytes = 1U << (int)(bref->data_off & HAMMER2_OFF_MASK_RADIX);
185 atomic_add_long(&hammer2_chain_allocs, 1);
188 * Construct the appropriate system structure.
191 case HAMMER2_BREF_TYPE_DIRENT:
192 case HAMMER2_BREF_TYPE_INODE:
193 case HAMMER2_BREF_TYPE_INDIRECT:
194 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
195 case HAMMER2_BREF_TYPE_DATA:
196 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
198 * Chain's are really only associated with the hmp but we
199 * maintain a pmp association for per-mount memory tracking
200 * purposes. The pmp can be NULL.
202 chain = kmalloc(sizeof(*chain), hmp->mchain, M_WAITOK | M_ZERO);
204 case HAMMER2_BREF_TYPE_VOLUME:
205 case HAMMER2_BREF_TYPE_FREEMAP:
207 * Only hammer2_chain_bulksnap() calls this function with these
210 chain = kmalloc(sizeof(*chain), hmp->mchain, M_WAITOK | M_ZERO);
214 panic("hammer2_chain_alloc: unrecognized blockref type: %d",
219 * Initialize the new chain structure. pmp must be set to NULL for
220 * chains belonging to the super-root topology of a device mount.
222 if (pmp == hmp->spmp)
228 chain->bytes = bytes;
230 chain->flags = HAMMER2_CHAIN_ALLOCATED;
233 * Set the PFS boundary flag if this chain represents a PFS root.
235 if (bref->flags & HAMMER2_BREF_FLAG_PFSROOT)
236 atomic_set_int(&chain->flags, HAMMER2_CHAIN_PFSBOUNDARY);
237 hammer2_chain_core_init(chain);
243 * Initialize a chain's core structure. This structure used to be allocated
244 * but is now embedded.
246 * The core is not locked. No additional refs on the chain are made.
247 * (trans) must not be NULL if (core) is not NULL.
250 hammer2_chain_core_init(hammer2_chain_t *chain)
253 * Fresh core under nchain (no multi-homing of ochain's
256 RB_INIT(&chain->core.rbtree); /* live chains */
257 hammer2_mtx_init(&chain->lock, "h2chain");
261 * Add a reference to a chain element, preventing its destruction.
263 * (can be called with spinlock held)
266 hammer2_chain_ref(hammer2_chain_t *chain)
268 if (atomic_fetchadd_int(&chain->refs, 1) == 0) {
270 * Just flag that the chain was used and should be recycled
271 * on the LRU if it encounters it later.
273 if (chain->flags & HAMMER2_CHAIN_ONLRU)
274 atomic_set_int(&chain->flags, HAMMER2_CHAIN_LRUHINT);
278 * REMOVED - reduces contention, lru_list is more heuristical
281 * 0->non-zero transition must ensure that chain is removed
284 * NOTE: Already holding lru_spin here so we cannot call
285 * hammer2_chain_ref() to get it off lru_list, do
288 if (chain->flags & HAMMER2_CHAIN_ONLRU) {
289 hammer2_pfs_t *pmp = chain->pmp;
290 hammer2_spin_ex(&pmp->lru_spin);
291 if (chain->flags & HAMMER2_CHAIN_ONLRU) {
292 atomic_add_int(&pmp->lru_count, -1);
293 atomic_clear_int(&chain->flags,
294 HAMMER2_CHAIN_ONLRU);
295 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node);
297 hammer2_spin_unex(&pmp->lru_spin);
304 * Ref a locked chain and force the data to be held across an unlock.
305 * Chain must be currently locked. The user of the chain who desires
306 * to release the hold must call hammer2_chain_lock_unhold() to relock
307 * and unhold the chain, then unlock normally, or may simply call
308 * hammer2_chain_drop_unhold() (which is safer against deadlocks).
311 hammer2_chain_ref_hold(hammer2_chain_t *chain)
313 atomic_add_int(&chain->lockcnt, 1);
314 hammer2_chain_ref(chain);
318 * Insert the chain in the core rbtree.
320 * Normal insertions are placed in the live rbtree. Insertion of a deleted
321 * chain is a special case used by the flush code that is placed on the
322 * unstaged deleted list to avoid confusing the live view.
324 #define HAMMER2_CHAIN_INSERT_SPIN 0x0001
325 #define HAMMER2_CHAIN_INSERT_LIVE 0x0002
326 #define HAMMER2_CHAIN_INSERT_RACE 0x0004
330 hammer2_chain_insert(hammer2_chain_t *parent, hammer2_chain_t *chain,
331 int flags, int generation)
333 hammer2_chain_t *xchain;
336 if (flags & HAMMER2_CHAIN_INSERT_SPIN)
337 hammer2_spin_ex(&parent->core.spin);
340 * Interlocked by spinlock, check for race
342 if ((flags & HAMMER2_CHAIN_INSERT_RACE) &&
343 parent->core.generation != generation) {
344 error = HAMMER2_ERROR_EAGAIN;
351 xchain = RB_INSERT(hammer2_chain_tree, &parent->core.rbtree, chain);
352 KASSERT(xchain == NULL,
353 ("hammer2_chain_insert: collision %p %p (key=%016jx)",
354 chain, xchain, chain->bref.key));
355 atomic_set_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE);
356 chain->parent = parent;
357 ++parent->core.chain_count;
358 ++parent->core.generation; /* XXX incs for _get() too, XXX */
361 * We have to keep track of the effective live-view blockref count
362 * so the create code knows when to push an indirect block.
364 if (flags & HAMMER2_CHAIN_INSERT_LIVE)
365 atomic_add_int(&parent->core.live_count, 1);
367 if (flags & HAMMER2_CHAIN_INSERT_SPIN)
368 hammer2_spin_unex(&parent->core.spin);
373 * Drop the caller's reference to the chain. When the ref count drops to
374 * zero this function will try to disassociate the chain from its parent and
375 * deallocate it, then recursely drop the parent using the implied ref
376 * from the chain's chain->parent.
378 * Nobody should own chain's mutex on the 1->0 transition, unless this drop
379 * races an acquisition by another cpu. Therefore we can loop if we are
380 * unable to acquire the mutex, and refs is unlikely to be 1 unless we again
381 * race against another drop.
383 static hammer2_chain_t *hammer2_chain_lastdrop(hammer2_chain_t *chain,
385 static void hammer2_chain_lru_flush(hammer2_pfs_t *pmp);
388 hammer2_chain_drop(hammer2_chain_t *chain)
392 if (hammer2_debug & 0x200000)
395 KKASSERT(chain->refs > 0);
403 if (hammer2_mtx_ex_try(&chain->lock) == 0)
404 chain = hammer2_chain_lastdrop(chain, 0);
405 /* retry the same chain, or chain from lastdrop */
407 if (atomic_cmpset_int(&chain->refs, refs, refs - 1))
409 /* retry the same chain */
416 * Unhold a held and probably not-locked chain, ensure that the data is
417 * dropped on the 1->0 transition of lockcnt by obtaining an exclusive
418 * lock and then simply unlocking the chain.
421 hammer2_chain_drop_unhold(hammer2_chain_t *chain)
427 lockcnt = chain->lockcnt;
430 if (atomic_cmpset_int(&chain->lockcnt,
431 lockcnt, lockcnt - 1)) {
434 } else if (hammer2_mtx_ex_try(&chain->lock) == 0) {
435 hammer2_chain_unlock(chain);
439 * This situation can easily occur on SMP due to
440 * the gap inbetween the 1->0 transition and the
441 * final unlock. We cannot safely block on the
442 * mutex because lockcnt might go above 1.
444 * XXX Sleep for one tick if it takes too long.
447 if (iter > 1000 + hz) {
448 kprintf("hammer2: h2race1 %p\n", chain);
451 tsleep(&iter, 0, "h2race1", 1);
456 hammer2_chain_drop(chain);
460 * Handles the (potential) last drop of chain->refs from 1->0. Called with
461 * the mutex exclusively locked, refs == 1, and lockcnt 0. SMP races are
462 * possible against refs and lockcnt. We must dispose of the mutex on chain.
464 * This function returns an unlocked chain for recursive drop or NULL. It
465 * can return the same chain if it determines it has raced another ref.
469 * When two chains need to be recursively dropped we use the chain we
470 * would otherwise free to placehold the additional chain. It's a bit
471 * convoluted but we can't just recurse without potentially blowing out
474 * The chain cannot be freed if it has any children.
475 * The chain cannot be freed if flagged MODIFIED unless we can dispose of it.
476 * The chain cannot be freed if flagged UPDATE unless we can dispose of it.
477 * Any dedup registration can remain intact.
479 * The core spinlock is allowed to nest child-to-parent (not parent-to-child).
483 hammer2_chain_lastdrop(hammer2_chain_t *chain, int depth)
487 hammer2_chain_t *parent;
488 hammer2_chain_t *rdrop;
495 * On last drop if there is no parent and data_off is good (at
496 * least does not represent the volume root), the modified chain
497 * is probably going to be destroyed. We have to make sure that
498 * the data area is not registered for dedup.
500 * XXX removed. In fact, we do not have to make sure that the
501 * data area is not registered for dedup. The data area
502 * can, in fact, still be used for dedup because it is
503 * still allocated in the freemap and the underlying I/O
504 * will still be flushed.
506 if (chain->parent == NULL &&
507 (chain->flags & HAMMER2_CHAIN_MODIFIED) &&
508 (chain->bref.data_off & ~HAMMER2_OFF_MASK_RADIX)) {
510 hammer2_io_dedup_delete(hmp, chain->bref.type,
511 chain->bref.data_off, chain->bytes);
515 * We need chain's spinlock to interlock the sub-tree test.
516 * We already have chain's mutex, protecting chain->parent.
518 * Remember that chain->refs can be in flux.
520 hammer2_spin_ex(&chain->core.spin);
522 if (chain->parent != NULL) {
524 * If the chain has a parent the UPDATE bit prevents scrapping
525 * as the chain is needed to properly flush the parent. Try
526 * to complete the 1->0 transition and return NULL. Retry
527 * (return chain) if we are unable to complete the 1->0
528 * transition, else return NULL (nothing more to do).
530 * If the chain has a parent the MODIFIED bit prevents
533 * Chains with UPDATE/MODIFIED are *not* put on the LRU list!
535 if (chain->flags & (HAMMER2_CHAIN_UPDATE |
536 HAMMER2_CHAIN_MODIFIED)) {
537 if (atomic_cmpset_int(&chain->refs, 1, 0)) {
538 hammer2_spin_unex(&chain->core.spin);
540 dio = hammer2_chain_drop_data(chain, 0);
542 hammer2_io_bqrelse(&dio);
544 hammer2_chain_assert_no_data(chain);
545 hammer2_mtx_unlock(&chain->lock);
548 hammer2_spin_unex(&chain->core.spin);
549 hammer2_mtx_unlock(&chain->lock);
553 /* spinlock still held */
554 } else if (chain->bref.type == HAMMER2_BREF_TYPE_VOLUME ||
555 chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP) {
557 * Retain the static vchain and fchain. Clear bits that
558 * are not relevant. Do not clear the MODIFIED bit,
559 * and certainly do not put it on the delayed-flush queue.
561 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
564 * The chain has no parent and can be flagged for destruction.
565 * Since it has no parent, UPDATE can also be cleared.
567 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DESTROY);
568 if (chain->flags & HAMMER2_CHAIN_UPDATE)
569 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
572 * If the chain has children we must still flush the chain.
573 * Any dedup is already handled by the underlying DIO, so
574 * we do not have to specifically flush it here.
576 * In the case where it has children, the DESTROY flag test
577 * in the flush code will prevent unnecessary flushes of
578 * MODIFIED chains that are not flagged DEDUP so don't worry
581 if (chain->core.chain_count) {
583 * Put on flushq (should ensure refs > 1), retry
586 hammer2_spin_unex(&chain->core.spin);
587 hammer2_delayed_flush(chain);
588 hammer2_mtx_unlock(&chain->lock);
590 return(chain); /* retry drop */
594 * Otherwise we can scrap the MODIFIED bit if it is set,
595 * and continue along the freeing path.
597 * Be sure to clean-out any dedup bits. Without a parent
598 * this chain will no longer be visible to the flush code.
599 * Easy check data_off to avoid the volume root.
601 if (chain->flags & HAMMER2_CHAIN_MODIFIED) {
602 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
603 atomic_add_long(&hammer2_count_modified_chains, -1);
605 hammer2_pfs_memory_wakeup(chain->pmp);
607 /* spinlock still held */
610 /* spinlock still held */
616 * If any children exist we must leave the chain intact with refs == 0.
617 * They exist because chains are retained below us which have refs or
618 * may require flushing.
620 * Retry (return chain) if we fail to transition the refs to 0, else
621 * return NULL indication nothing more to do.
623 * Chains with children are NOT put on the LRU list.
625 if (chain->core.chain_count) {
626 if (atomic_cmpset_int(&chain->refs, 1, 0)) {
627 hammer2_spin_unex(&chain->core.spin);
628 hammer2_chain_assert_no_data(chain);
629 hammer2_mtx_unlock(&chain->lock);
632 hammer2_spin_unex(&chain->core.spin);
633 hammer2_mtx_unlock(&chain->lock);
637 /* spinlock still held */
638 /* no chains left under us */
641 * chain->core has no children left so no accessors can get to our
642 * chain from there. Now we have to lock the parent core to interlock
643 * remaining possible accessors that might bump chain's refs before
644 * we can safely drop chain's refs with intent to free the chain.
647 pmp = chain->pmp; /* can be NULL */
650 parent = chain->parent;
653 * WARNING! chain's spin lock is still held here, and other spinlocks
654 * will be acquired and released in the code below. We
655 * cannot be making fancy procedure calls!
659 * We can cache the chain if it is associated with a pmp
660 * and not flagged as being destroyed or requesting a full
661 * release. In this situation the chain is not removed
662 * from its parent, i.e. it can still be looked up.
664 * We intentionally do not cache DATA chains because these
665 * were likely used to load data into the logical buffer cache
666 * and will not be accessed again for some time.
669 (HAMMER2_CHAIN_DESTROY | HAMMER2_CHAIN_RELEASE)) == 0 &&
671 chain->bref.type != HAMMER2_BREF_TYPE_DATA) {
673 hammer2_spin_ex(&parent->core.spin);
674 if (atomic_cmpset_int(&chain->refs, 1, 0) == 0) {
676 * 1->0 transition failed, retry. Do not drop
677 * the chain's data yet!
680 hammer2_spin_unex(&parent->core.spin);
681 hammer2_spin_unex(&chain->core.spin);
682 hammer2_mtx_unlock(&chain->lock);
691 dio = hammer2_chain_drop_data(chain, 1);
693 hammer2_chain_assert_no_data(chain);
696 * Make sure we are on the LRU list, clean up excessive
697 * LRU entries. We can only really drop one but there might
698 * be other entries that we can remove from the lru_list
701 * NOTE: HAMMER2_CHAIN_ONLRU may only be safely set when
702 * chain->core.spin AND pmp->lru_spin are held, but
703 * can be safely cleared only holding pmp->lru_spin.
705 if ((chain->flags & HAMMER2_CHAIN_ONLRU) == 0) {
706 hammer2_spin_ex(&pmp->lru_spin);
707 if ((chain->flags & HAMMER2_CHAIN_ONLRU) == 0) {
708 atomic_set_int(&chain->flags,
709 HAMMER2_CHAIN_ONLRU);
710 TAILQ_INSERT_TAIL(&pmp->lru_list,
712 atomic_add_int(&pmp->lru_count, 1);
714 if (pmp->lru_count < HAMMER2_LRU_LIMIT)
715 depth = 1; /* disable lru_list flush */
716 hammer2_spin_unex(&pmp->lru_spin);
718 /* disable lru flush */
723 hammer2_spin_unex(&parent->core.spin);
724 parent = NULL; /* safety */
726 hammer2_spin_unex(&chain->core.spin);
727 hammer2_mtx_unlock(&chain->lock);
730 hammer2_io_bqrelse(&dio);
734 * lru_list hysteresis (see above for depth overrides).
735 * Note that depth also prevents excessive lastdrop recursion.
738 hammer2_chain_lru_flush(pmp);
745 * Make sure we are not on the LRU list.
747 if (chain->flags & HAMMER2_CHAIN_ONLRU) {
748 hammer2_spin_ex(&pmp->lru_spin);
749 if (chain->flags & HAMMER2_CHAIN_ONLRU) {
750 atomic_add_int(&pmp->lru_count, -1);
751 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU);
752 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node);
754 hammer2_spin_unex(&pmp->lru_spin);
758 * Spinlock the parent and try to drop the last ref on chain.
759 * On success determine if we should dispose of the chain
760 * (remove the chain from its parent, etc).
762 * (normal core locks are top-down recursive but we define
763 * core spinlocks as bottom-up recursive, so this is safe).
766 hammer2_spin_ex(&parent->core.spin);
767 if (atomic_cmpset_int(&chain->refs, 1, 0) == 0) {
769 /* XXX remove, don't try to drop data on fail */
770 hammer2_spin_unex(&parent->core.spin);
771 dio = hammer2_chain_drop_data(chain, 0);
772 hammer2_spin_unex(&chain->core.spin);
774 hammer2_io_bqrelse(&dio);
777 * 1->0 transition failed, retry.
779 hammer2_spin_unex(&parent->core.spin);
780 hammer2_spin_unex(&chain->core.spin);
781 hammer2_mtx_unlock(&chain->lock);
787 * 1->0 transition successful, parent spin held to prevent
788 * new lookups, chain spinlock held to protect parent field.
789 * Remove chain from the parent.
791 * If the chain is being removed from the parent's btree but
792 * is not bmapped, we have to adjust live_count downward. If
793 * it is bmapped then the blockref is retained in the parent
794 * as is its associated live_count. This case can occur when
795 * a chain added to the topology is unable to flush and is
796 * then later deleted.
798 if (chain->flags & HAMMER2_CHAIN_ONRBTREE) {
799 if ((parent->flags & HAMMER2_CHAIN_COUNTEDBREFS) &&
800 (chain->flags & HAMMER2_CHAIN_BMAPPED) == 0) {
801 atomic_add_int(&parent->core.live_count, -1);
803 RB_REMOVE(hammer2_chain_tree,
804 &parent->core.rbtree, chain);
805 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE);
806 --parent->core.chain_count;
807 chain->parent = NULL;
811 * If our chain was the last chain in the parent's core the
812 * core is now empty and its parent might have to be
813 * re-dropped if it has 0 refs.
815 if (parent->core.chain_count == 0) {
817 atomic_add_int(&rdrop->refs, 1);
819 if (atomic_cmpset_int(&rdrop->refs, 0, 1) == 0)
823 hammer2_spin_unex(&parent->core.spin);
824 parent = NULL; /* safety */
830 if (atomic_cmpset_int(&chain->refs, 1, 0) == 0) {
832 * 1->0 transition failed, retry.
834 hammer2_spin_unex(&parent->core.spin);
835 hammer2_spin_unex(&chain->core.spin);
836 hammer2_mtx_unlock(&chain->lock);
843 * Successful 1->0 transition, no parent, no children... no way for
844 * anyone to ref this chain any more. We can clean-up and free it.
846 * We still have the core spinlock, and core's chain_count is 0.
847 * Any parent spinlock is gone.
849 hammer2_spin_unex(&chain->core.spin);
850 hammer2_chain_assert_no_data(chain);
851 hammer2_mtx_unlock(&chain->lock);
852 KKASSERT(RB_EMPTY(&chain->core.rbtree) &&
853 chain->core.chain_count == 0);
856 * All locks are gone, no pointers remain to the chain, finish
859 KKASSERT((chain->flags & (HAMMER2_CHAIN_UPDATE |
860 HAMMER2_CHAIN_MODIFIED)) == 0);
862 dio = hammer2_chain_drop_data(chain, 1);
864 hammer2_io_bqrelse(&dio);
868 * Once chain resources are gone we can use the now dead chain
869 * structure to placehold what might otherwise require a recursive
870 * drop, because we have potentially two things to drop and can only
871 * return one directly.
873 if (chain->flags & HAMMER2_CHAIN_ALLOCATED) {
874 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ALLOCATED);
876 kfree(chain, hmp->mchain);
880 * Possible chaining loop when parent re-drop needed.
886 * Heuristical flush of the LRU, try to reduce the number of entries
887 * on the LRU to (HAMMER2_LRU_LIMIT * 2 / 3). This procedure is called
888 * only when lru_count exceeds HAMMER2_LRU_LIMIT.
892 hammer2_chain_lru_flush(hammer2_pfs_t *pmp)
894 hammer2_chain_t *chain;
898 hammer2_spin_ex(&pmp->lru_spin);
899 while (pmp->lru_count > HAMMER2_LRU_LIMIT * 2 / 3) {
901 * Pick a chain off the lru_list, just recycle it quickly
902 * if LRUHINT is set (the chain was ref'd but left on
903 * the lru_list, so cycle to the end).
905 chain = TAILQ_FIRST(&pmp->lru_list);
906 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node);
908 if (chain->flags & HAMMER2_CHAIN_LRUHINT) {
909 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_LRUHINT);
910 TAILQ_INSERT_TAIL(&pmp->lru_list, chain, lru_node);
916 * Ok, we are off the LRU. We must adjust refs before we
917 * can safely clear the ONLRU flag.
919 atomic_add_int(&pmp->lru_count, -1);
920 if (atomic_cmpset_int(&chain->refs, 0, 1)) {
921 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU);
922 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
925 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU);
928 hammer2_spin_unex(&pmp->lru_spin);
933 * If we picked a chain off the lru list we may be able to lastdrop
934 * it. Use a depth of 1 to prevent excessive lastdrop recursion.
944 if (hammer2_mtx_ex_try(&chain->lock) == 0)
945 chain = hammer2_chain_lastdrop(chain, 1);
946 /* retry the same chain, or chain from lastdrop */
948 if (atomic_cmpset_int(&chain->refs, refs, refs - 1))
950 /* retry the same chain */
958 * On last lock release.
960 static hammer2_io_t *
961 hammer2_chain_drop_data(hammer2_chain_t *chain)
965 if ((dio = chain->dio) != NULL) {
969 switch(chain->bref.type) {
970 case HAMMER2_BREF_TYPE_VOLUME:
971 case HAMMER2_BREF_TYPE_FREEMAP:
974 if (chain->data != NULL) {
975 hammer2_spin_unex(&chain->core.spin);
976 panic("chain data not null: "
977 "chain %p bref %016jx.%02x "
978 "refs %d parent %p dio %p data %p",
979 chain, chain->bref.data_off,
980 chain->bref.type, chain->refs,
982 chain->dio, chain->data);
984 KKASSERT(chain->data == NULL);
992 * Lock a referenced chain element, acquiring its data with I/O if necessary,
993 * and specify how you would like the data to be resolved.
995 * If an I/O or other fatal error occurs, chain->error will be set to non-zero.
997 * The lock is allowed to recurse, multiple locking ops will aggregate
998 * the requested resolve types. Once data is assigned it will not be
999 * removed until the last unlock.
1001 * HAMMER2_RESOLVE_NEVER - Do not resolve the data element.
1002 * (typically used to avoid device/logical buffer
1003 * aliasing for data)
1005 * HAMMER2_RESOLVE_MAYBE - Do not resolve data elements for chains in
1006 * the INITIAL-create state (indirect blocks only).
1008 * Do not resolve data elements for DATA chains.
1009 * (typically used to avoid device/logical buffer
1010 * aliasing for data)
1012 * HAMMER2_RESOLVE_ALWAYS- Always resolve the data element.
1014 * HAMMER2_RESOLVE_SHARED- (flag) The chain is locked shared, otherwise
1015 * it will be locked exclusive.
1017 * HAMMER2_RESOLVE_NONBLOCK- (flag) The chain is locked non-blocking. If
1018 * the lock fails, EAGAIN is returned.
1020 * NOTE: Embedded elements (volume header, inodes) are always resolved
1023 * NOTE: Specifying HAMMER2_RESOLVE_ALWAYS on a newly-created non-embedded
1024 * element will instantiate and zero its buffer, and flush it on
1027 * NOTE: (data) elements are normally locked RESOLVE_NEVER or RESOLVE_MAYBE
1028 * so as not to instantiate a device buffer, which could alias against
1029 * a logical file buffer. However, if ALWAYS is specified the
1030 * device buffer will be instantiated anyway.
1032 * NOTE: The return value is always 0 unless NONBLOCK is specified, in which
1033 * case it can be either 0 or EAGAIN.
1035 * WARNING! This function blocks on I/O if data needs to be fetched. This
1036 * blocking can run concurrent with other compatible lock holders
1037 * who do not need data returning. The lock is not upgraded to
1038 * exclusive during a data fetch, a separate bit is used to
1039 * interlock I/O. However, an exclusive lock holder can still count
1040 * on being interlocked against an I/O fetch managed by a shared
1044 hammer2_chain_lock(hammer2_chain_t *chain, int how)
1046 KKASSERT(chain->refs > 0);
1048 if (how & HAMMER2_RESOLVE_NONBLOCK) {
1050 * For non-blocking operation attempt to get the lock
1051 * before bumping lockcnt, just so we don't have to deal
1052 * with dropping lockcnt (and dealing with the underlying
1055 * NOTE: LOCKAGAIN must always succeed without blocking.
1057 if (how & HAMMER2_RESOLVE_SHARED) {
1058 if (how & HAMMER2_RESOLVE_LOCKAGAIN) {
1059 hammer2_mtx_sh_again(&chain->lock);
1061 if (hammer2_mtx_sh_try(&chain->lock) != 0)
1065 if (hammer2_mtx_ex_try(&chain->lock) != 0)
1068 atomic_add_int(&chain->lockcnt, 1);
1069 ++curthread->td_tracker;
1072 * Lock the element. Recursive locks are allowed. lockcnt
1073 * ensures that data is left intact.
1075 atomic_add_int(&chain->lockcnt, 1);
1078 * Get the appropriate lock. If LOCKAGAIN is flagged with
1079 * SHARED the caller expects a shared lock to already be
1080 * present and we are giving it another ref. This case must
1081 * importantly not block if there is a pending exclusive lock
1084 if (how & HAMMER2_RESOLVE_SHARED) {
1085 if (how & HAMMER2_RESOLVE_LOCKAGAIN) {
1086 hammer2_mtx_sh_again(&chain->lock);
1088 hammer2_mtx_sh(&chain->lock);
1091 hammer2_mtx_ex(&chain->lock);
1093 ++curthread->td_tracker;
1097 * If we already have a valid data pointer make sure the data is
1098 * synchronized to the current cpu, and then no further action is
1103 hammer2_io_bkvasync(chain->dio);
1108 * Do we have to resolve the data? This is generally only
1109 * applicable to HAMMER2_BREF_TYPE_DATA which is special-cased.
1110 * Other BREF types expects the data to be there.
1112 switch(how & HAMMER2_RESOLVE_MASK) {
1113 case HAMMER2_RESOLVE_NEVER:
1115 case HAMMER2_RESOLVE_MAYBE:
1116 if (chain->flags & HAMMER2_CHAIN_INITIAL)
1118 if (chain->bref.type == HAMMER2_BREF_TYPE_DATA)
1121 if (chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE)
1123 if (chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_LEAF)
1127 case HAMMER2_RESOLVE_ALWAYS:
1133 * Caller requires data
1135 hammer2_chain_load_data(chain);
1141 * Lock the chain, retain the hold, and drop the data persistence count.
1142 * The data should remain valid because we never transitioned lockcnt
1146 hammer2_chain_lock_unhold(hammer2_chain_t *chain, int how)
1148 hammer2_chain_lock(chain, how);
1149 atomic_add_int(&chain->lockcnt, -1);
1154 * Downgrade an exclusive chain lock to a shared chain lock.
1156 * NOTE: There is no upgrade equivalent due to the ease of
1157 * deadlocks in that direction.
1160 hammer2_chain_lock_downgrade(hammer2_chain_t *chain)
1162 hammer2_mtx_downgrade(&chain->lock);
1167 * Issue I/O and install chain->data. Caller must hold a chain lock, lock
1168 * may be of any type.
1170 * Once chain->data is set it cannot be disposed of until all locks are
1173 * Make sure the data is synchronized to the current cpu.
1176 hammer2_chain_load_data(hammer2_chain_t *chain)
1178 hammer2_blockref_t *bref;
1185 * Degenerate case, data already present, or chain has no media
1186 * reference to load.
1190 hammer2_io_bkvasync(chain->dio);
1193 if ((chain->bref.data_off & ~HAMMER2_OFF_MASK_RADIX) == 0)
1197 KKASSERT(hmp != NULL);
1200 * Gain the IOINPROG bit, interlocked block.
1206 oflags = chain->flags;
1208 if (oflags & HAMMER2_CHAIN_IOINPROG) {
1209 nflags = oflags | HAMMER2_CHAIN_IOSIGNAL;
1210 tsleep_interlock(&chain->flags, 0);
1211 if (atomic_cmpset_int(&chain->flags, oflags, nflags)) {
1212 tsleep(&chain->flags, PINTERLOCKED,
1217 nflags = oflags | HAMMER2_CHAIN_IOINPROG;
1218 if (atomic_cmpset_int(&chain->flags, oflags, nflags)) {
1226 * We own CHAIN_IOINPROG
1228 * Degenerate case if we raced another load.
1232 hammer2_io_bkvasync(chain->dio);
1237 * We must resolve to a device buffer, either by issuing I/O or
1238 * by creating a zero-fill element. We do not mark the buffer
1239 * dirty when creating a zero-fill element (the hammer2_chain_modify()
1240 * API must still be used to do that).
1242 * The device buffer is variable-sized in powers of 2 down
1243 * to HAMMER2_MIN_ALLOC (typically 1K). A 64K physical storage
1244 * chunk always contains buffers of the same size. (XXX)
1246 * The minimum physical IO size may be larger than the variable
1249 bref = &chain->bref;
1252 * The getblk() optimization can only be used on newly created
1253 * elements if the physical block size matches the request.
1255 if (chain->flags & HAMMER2_CHAIN_INITIAL) {
1256 error = hammer2_io_new(hmp, bref->type,
1257 bref->data_off, chain->bytes,
1260 error = hammer2_io_bread(hmp, bref->type,
1261 bref->data_off, chain->bytes,
1263 hammer2_adjreadcounter(&chain->bref, chain->bytes);
1266 chain->error = HAMMER2_ERROR_EIO;
1267 kprintf("hammer2_chain_lock: I/O error %016jx: %d\n",
1268 (intmax_t)bref->data_off, error);
1269 hammer2_io_bqrelse(&chain->dio);
1275 * This isn't perfect and can be ignored on OSs which do not have
1276 * an indication as to whether a buffer is coming from cache or
1277 * if I/O was actually issued for the read. TESTEDGOOD will work
1278 * pretty well without the B_IOISSUED logic because chains are
1279 * cached, but in that situation (without B_IOISSUED) it will not
1280 * detect whether a re-read via I/O is corrupted verses the original
1283 * We can't re-run the CRC on every fresh lock. That would be
1284 * insanely expensive.
1286 * If the underlying kernel buffer covers the entire chain we can
1287 * use the B_IOISSUED indication to determine if we have to re-run
1288 * the CRC on chain data for chains that managed to stay cached
1289 * across the kernel disposal of the original buffer.
1291 if ((dio = chain->dio) != NULL && dio->bp) {
1292 struct buf *bp = dio->bp;
1294 if (dio->psize == chain->bytes &&
1295 (bp->b_flags & B_IOISSUED)) {
1296 atomic_clear_int(&chain->flags,
1297 HAMMER2_CHAIN_TESTEDGOOD);
1298 bp->b_flags &= ~B_IOISSUED;
1303 * NOTE: A locked chain's data cannot be modified without first
1304 * calling hammer2_chain_modify().
1308 * Clear INITIAL. In this case we used io_new() and the buffer has
1309 * been zero'd and marked dirty.
1311 * NOTE: hammer2_io_data() call issues bkvasync()
1313 bdata = hammer2_io_data(chain->dio, chain->bref.data_off);
1315 if (chain->flags & HAMMER2_CHAIN_INITIAL) {
1316 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1317 chain->bref.flags |= HAMMER2_BREF_FLAG_ZERO;
1318 } else if (chain->flags & HAMMER2_CHAIN_MODIFIED) {
1320 * check data not currently synchronized due to
1321 * modification. XXX assumes data stays in the buffer
1322 * cache, which might not be true (need biodep on flush
1323 * to calculate crc? or simple crc?).
1325 } else if ((chain->flags & HAMMER2_CHAIN_TESTEDGOOD) == 0) {
1326 if (hammer2_chain_testcheck(chain, bdata) == 0) {
1327 chain->error = HAMMER2_ERROR_CHECK;
1329 atomic_set_int(&chain->flags, HAMMER2_CHAIN_TESTEDGOOD);
1334 * Setup the data pointer, either pointing it to an embedded data
1335 * structure and copying the data from the buffer, or pointing it
1338 * The buffer is not retained when copying to an embedded data
1339 * structure in order to avoid potential deadlocks or recursions
1340 * on the same physical buffer.
1342 * WARNING! Other threads can start using the data the instant we
1343 * set chain->data non-NULL.
1345 switch (bref->type) {
1346 case HAMMER2_BREF_TYPE_VOLUME:
1347 case HAMMER2_BREF_TYPE_FREEMAP:
1349 * Copy data from bp to embedded buffer
1351 panic("hammer2_chain_load_data: unresolved volume header");
1353 case HAMMER2_BREF_TYPE_DIRENT:
1354 KKASSERT(chain->bytes != 0);
1356 case HAMMER2_BREF_TYPE_INODE:
1357 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
1358 case HAMMER2_BREF_TYPE_INDIRECT:
1359 case HAMMER2_BREF_TYPE_DATA:
1360 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1363 * Point data at the device buffer and leave dio intact.
1365 chain->data = (void *)bdata;
1370 * Release HAMMER2_CHAIN_IOINPROG and signal waiters if requested.
1377 oflags = chain->flags;
1378 nflags = oflags & ~(HAMMER2_CHAIN_IOINPROG |
1379 HAMMER2_CHAIN_IOSIGNAL);
1380 KKASSERT(oflags & HAMMER2_CHAIN_IOINPROG);
1381 if (atomic_cmpset_int(&chain->flags, oflags, nflags)) {
1382 if (oflags & HAMMER2_CHAIN_IOSIGNAL)
1383 wakeup(&chain->flags);
1390 * Unlock and deref a chain element.
1392 * Remember that the presence of children under chain prevent the chain's
1393 * destruction but do not add additional references, so the dio will still
1397 hammer2_chain_unlock(hammer2_chain_t *chain)
1403 --curthread->td_tracker;
1406 * If multiple locks are present (or being attempted) on this
1407 * particular chain we can just unlock, drop refs, and return.
1409 * Otherwise fall-through on the 1->0 transition.
1412 lockcnt = chain->lockcnt;
1413 KKASSERT(lockcnt > 0);
1416 if (atomic_cmpset_int(&chain->lockcnt,
1417 lockcnt, lockcnt - 1)) {
1418 hammer2_mtx_unlock(&chain->lock);
1421 } else if (hammer2_mtx_upgrade_try(&chain->lock) == 0) {
1422 /* while holding the mutex exclusively */
1423 if (atomic_cmpset_int(&chain->lockcnt, 1, 0))
1427 * This situation can easily occur on SMP due to
1428 * the gap inbetween the 1->0 transition and the
1429 * final unlock. We cannot safely block on the
1430 * mutex because lockcnt might go above 1.
1432 * XXX Sleep for one tick if it takes too long.
1434 if (++iter > 1000) {
1435 if (iter > 1000 + hz) {
1436 kprintf("hammer2: h2race2 %p\n", chain);
1439 tsleep(&iter, 0, "h2race2", 1);
1447 * Last unlock / mutex upgraded to exclusive. Drop the data
1450 dio = hammer2_chain_drop_data(chain);
1452 hammer2_io_bqrelse(&dio);
1453 hammer2_mtx_unlock(&chain->lock);
1457 * Unlock and hold chain data intact
1460 hammer2_chain_unlock_hold(hammer2_chain_t *chain)
1462 atomic_add_int(&chain->lockcnt, 1);
1463 hammer2_chain_unlock(chain);
1467 * Helper to obtain the blockref[] array base and count for a chain.
1469 * XXX Not widely used yet, various use cases need to be validated and
1470 * converted to use this function.
1473 hammer2_blockref_t *
1474 hammer2_chain_base_and_count(hammer2_chain_t *parent, int *countp)
1476 hammer2_blockref_t *base;
1479 if (parent->flags & HAMMER2_CHAIN_INITIAL) {
1482 switch(parent->bref.type) {
1483 case HAMMER2_BREF_TYPE_INODE:
1484 count = HAMMER2_SET_COUNT;
1486 case HAMMER2_BREF_TYPE_INDIRECT:
1487 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1488 count = parent->bytes / sizeof(hammer2_blockref_t);
1490 case HAMMER2_BREF_TYPE_VOLUME:
1491 count = HAMMER2_SET_COUNT;
1493 case HAMMER2_BREF_TYPE_FREEMAP:
1494 count = HAMMER2_SET_COUNT;
1497 panic("hammer2_chain_create_indirect: "
1498 "unrecognized blockref type: %d",
1504 switch(parent->bref.type) {
1505 case HAMMER2_BREF_TYPE_INODE:
1506 base = &parent->data->ipdata.u.blockset.blockref[0];
1507 count = HAMMER2_SET_COUNT;
1509 case HAMMER2_BREF_TYPE_INDIRECT:
1510 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1511 base = &parent->data->npdata[0];
1512 count = parent->bytes / sizeof(hammer2_blockref_t);
1514 case HAMMER2_BREF_TYPE_VOLUME:
1515 base = &parent->data->voldata.
1516 sroot_blockset.blockref[0];
1517 count = HAMMER2_SET_COUNT;
1519 case HAMMER2_BREF_TYPE_FREEMAP:
1520 base = &parent->data->blkset.blockref[0];
1521 count = HAMMER2_SET_COUNT;
1524 panic("hammer2_chain_create_indirect: "
1525 "unrecognized blockref type: %d",
1537 * This counts the number of live blockrefs in a block array and
1538 * also calculates the point at which all remaining blockrefs are empty.
1539 * This routine can only be called on a live chain.
1541 * Caller holds the chain locked, but possibly with a shared lock. We
1542 * must use an exclusive spinlock to prevent corruption.
1544 * NOTE: Flag is not set until after the count is complete, allowing
1545 * callers to test the flag without holding the spinlock.
1547 * NOTE: If base is NULL the related chain is still in the INITIAL
1548 * state and there are no blockrefs to count.
1550 * NOTE: live_count may already have some counts accumulated due to
1551 * creation and deletion and could even be initially negative.
1554 hammer2_chain_countbrefs(hammer2_chain_t *chain,
1555 hammer2_blockref_t *base, int count)
1557 hammer2_spin_ex(&chain->core.spin);
1558 if ((chain->flags & HAMMER2_CHAIN_COUNTEDBREFS) == 0) {
1560 while (--count >= 0) {
1561 if (base[count].type)
1564 chain->core.live_zero = count + 1;
1565 while (count >= 0) {
1566 if (base[count].type)
1567 atomic_add_int(&chain->core.live_count,
1572 chain->core.live_zero = 0;
1574 /* else do not modify live_count */
1575 atomic_set_int(&chain->flags, HAMMER2_CHAIN_COUNTEDBREFS);
1577 hammer2_spin_unex(&chain->core.spin);
1581 * Resize the chain's physical storage allocation in-place. This function does
1582 * not usually adjust the data pointer and must be followed by (typically) a
1583 * hammer2_chain_modify() call to copy any old data over and adjust the
1586 * Chains can be resized smaller without reallocating the storage. Resizing
1587 * larger will reallocate the storage. Excess or prior storage is reclaimed
1588 * asynchronously at a later time.
1590 * An nradix value of 0 is special-cased to mean that the storage should
1591 * be disassociated, that is the chain is being resized to 0 bytes (not 1
1594 * Must be passed an exclusively locked parent and chain.
1596 * This function is mostly used with DATA blocks locked RESOLVE_NEVER in order
1597 * to avoid instantiating a device buffer that conflicts with the vnode data
1598 * buffer. However, because H2 can compress or encrypt data, the chain may
1599 * have a dio assigned to it in those situations, and they do not conflict.
1601 * XXX return error if cannot resize.
1604 hammer2_chain_resize(hammer2_chain_t *chain,
1605 hammer2_tid_t mtid, hammer2_off_t dedup_off,
1606 int nradix, int flags)
1616 * Only data and indirect blocks can be resized for now.
1617 * (The volu root, inodes, and freemap elements use a fixed size).
1619 KKASSERT(chain != &hmp->vchain);
1620 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_DATA ||
1621 chain->bref.type == HAMMER2_BREF_TYPE_INDIRECT ||
1622 chain->bref.type == HAMMER2_BREF_TYPE_DIRENT);
1625 * Nothing to do if the element is already the proper size
1627 obytes = chain->bytes;
1628 nbytes = (nradix) ? (1U << nradix) : 0;
1629 if (obytes == nbytes)
1630 return (chain->error);
1633 * Make sure the old data is instantiated so we can copy it. If this
1634 * is a data block, the device data may be superfluous since the data
1635 * might be in a logical block, but compressed or encrypted data is
1638 * NOTE: The modify will set BMAPUPD for us if BMAPPED is set.
1640 error = hammer2_chain_modify(chain, mtid, dedup_off, 0);
1645 * Relocate the block, even if making it smaller (because different
1646 * block sizes may be in different regions).
1648 * NOTE: Operation does not copy the data and may only be used
1649 * to resize data blocks in-place, or directory entry blocks
1650 * which are about to be modified in some manner.
1652 error = hammer2_freemap_alloc(chain, nbytes);
1656 chain->bytes = nbytes;
1659 * We don't want the followup chain_modify() to try to copy data
1660 * from the old (wrong-sized) buffer. It won't know how much to
1661 * copy. This case should only occur during writes when the
1662 * originator already has the data to write in-hand.
1665 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_DATA ||
1666 chain->bref.type == HAMMER2_BREF_TYPE_DIRENT);
1667 hammer2_io_brelse(&chain->dio);
1670 return (chain->error);
1674 * Set the chain modified so its data can be changed by the caller, or
1675 * install deduplicated data. The caller must call this routine for each
1676 * set of modifications it makes, even if the chain is already flagged
1679 * Sets bref.modify_tid to mtid only if mtid != 0. Note that bref.modify_tid
1680 * is a CLC (cluster level change) field and is not updated by parent
1681 * propagation during a flush.
1683 * Returns an appropriate HAMMER2_ERROR_* code, which will generally reflect
1684 * chain->error except for HAMMER2_ERROR_ENOSPC. If the allocation fails
1685 * due to no space available, HAMMER2_ERROR_ENOSPC is returned and the chain
1686 * remains unmodified with its old data ref intact and chain->error
1691 * If the DEDUPABLE flag is set in the chain the storage must be reallocated
1692 * even if the chain is still flagged MODIFIED. In this case the chain's
1693 * DEDUPABLE flag will be cleared once the new storage has been assigned.
1695 * If the caller passes a non-zero dedup_off we will use it to assign the
1696 * new storage. The MODIFIED flag will be *CLEARED* in this case, and
1697 * DEDUPABLE will be set (NOTE: the UPDATE flag is always set). The caller
1698 * must not modify the data content upon return.
1701 hammer2_chain_modify(hammer2_chain_t *chain, hammer2_tid_t mtid,
1702 hammer2_off_t dedup_off, int flags)
1704 hammer2_blockref_t obref;
1715 obref = chain->bref;
1716 KKASSERT((chain->flags & HAMMER2_CHAIN_FICTITIOUS) == 0);
1719 * Data is not optional for freemap chains (we must always be sure
1720 * to copy the data on COW storage allocations).
1722 if (chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE ||
1723 chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) {
1724 KKASSERT((chain->flags & HAMMER2_CHAIN_INITIAL) ||
1725 (flags & HAMMER2_MODIFY_OPTDATA) == 0);
1729 * Data must be resolved if already assigned, unless explicitly
1730 * flagged otherwise. If we cannot safety load the data the
1731 * modification fails and we return early.
1733 if (chain->data == NULL && chain->bytes != 0 &&
1734 (flags & HAMMER2_MODIFY_OPTDATA) == 0 &&
1735 (chain->bref.data_off & ~HAMMER2_OFF_MASK_RADIX)) {
1736 hammer2_chain_load_data(chain);
1738 return (chain->error);
1743 * Set MODIFIED to indicate that the chain has been modified. A new
1744 * allocation is required when modifying a chain.
1746 * Set UPDATE to ensure that the blockref is updated in the parent.
1748 * If MODIFIED is already set determine if we can reuse the assigned
1749 * data block or if we need a new data block.
1751 if ((chain->flags & HAMMER2_CHAIN_MODIFIED) == 0) {
1753 * Must set modified bit.
1755 atomic_add_long(&hammer2_count_modified_chains, 1);
1756 atomic_set_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
1757 hammer2_pfs_memory_inc(chain->pmp); /* can be NULL */
1761 * We may be able to avoid a copy-on-write if the chain's
1762 * check mode is set to NONE and the chain's current
1763 * modify_tid is beyond the last explicit snapshot tid.
1765 * This implements HAMMER2's overwrite-in-place feature.
1767 * NOTE! This data-block cannot be used as a de-duplication
1768 * source when the check mode is set to NONE.
1770 if ((chain->bref.type == HAMMER2_BREF_TYPE_DATA ||
1771 chain->bref.type == HAMMER2_BREF_TYPE_DIRENT) &&
1772 (chain->flags & HAMMER2_CHAIN_INITIAL) == 0 &&
1773 (chain->flags & HAMMER2_CHAIN_DEDUPABLE) == 0 &&
1774 HAMMER2_DEC_CHECK(chain->bref.methods) ==
1775 HAMMER2_CHECK_NONE &&
1777 chain->bref.modify_tid >
1778 chain->pmp->iroot->meta.pfs_lsnap_tid) {
1780 * Sector overwrite allowed.
1785 * Sector overwrite not allowed, must copy-on-write.
1789 } else if (chain->flags & HAMMER2_CHAIN_DEDUPABLE) {
1791 * If the modified chain was registered for dedup we need
1792 * a new allocation. This only happens for delayed-flush
1793 * chains (i.e. which run through the front-end buffer
1800 * Already flagged modified, no new allocation is needed.
1807 * Flag parent update required.
1809 if ((chain->flags & HAMMER2_CHAIN_UPDATE) == 0) {
1810 atomic_set_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
1817 * The modification or re-modification requires an allocation and
1818 * possible COW. If an error occurs, the previous content and data
1819 * reference is retained and the modification fails.
1821 * If dedup_off is non-zero, the caller is requesting a deduplication
1822 * rather than a modification. The MODIFIED bit is not set and the
1823 * data offset is set to the deduplication offset. The data cannot
1826 * NOTE: The dedup offset is allowed to be in a partially free state
1827 * and we must be sure to reset it to a fully allocated state
1828 * to force two bulkfree passes to free it again.
1830 * NOTE: Only applicable when chain->bytes != 0.
1832 * XXX can a chain already be marked MODIFIED without a data
1833 * assignment? If not, assert here instead of testing the case.
1835 if (chain != &hmp->vchain && chain != &hmp->fchain &&
1837 if ((chain->bref.data_off & ~HAMMER2_OFF_MASK_RADIX) == 0 ||
1841 * NOTE: We do not have to remove the dedup
1842 * registration because the area is still
1843 * allocated and the underlying DIO will
1847 chain->bref.data_off = dedup_off;
1848 chain->bytes = 1 << (dedup_off &
1849 HAMMER2_OFF_MASK_RADIX);
1851 atomic_clear_int(&chain->flags,
1852 HAMMER2_CHAIN_MODIFIED);
1853 atomic_add_long(&hammer2_count_modified_chains,
1856 hammer2_pfs_memory_wakeup(chain->pmp);
1857 hammer2_freemap_adjust(hmp, &chain->bref,
1858 HAMMER2_FREEMAP_DORECOVER);
1859 atomic_set_int(&chain->flags,
1860 HAMMER2_CHAIN_DEDUPABLE);
1862 error = hammer2_freemap_alloc(chain,
1864 atomic_clear_int(&chain->flags,
1865 HAMMER2_CHAIN_DEDUPABLE);
1871 * Stop here if error. We have to undo any flag bits we might
1876 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
1877 atomic_add_long(&hammer2_count_modified_chains, -1);
1879 hammer2_pfs_memory_wakeup(chain->pmp);
1882 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
1888 * Update mirror_tid and modify_tid. modify_tid is only updated
1889 * if not passed as zero (during flushes, parent propagation passes
1892 * NOTE: chain->pmp could be the device spmp.
1894 chain->bref.mirror_tid = hmp->voldata.mirror_tid + 1;
1896 chain->bref.modify_tid = mtid;
1899 * Set BMAPUPD to tell the flush code that an existing blockmap entry
1900 * requires updating as well as to tell the delete code that the
1901 * chain's blockref might not exactly match (in terms of physical size
1902 * or block offset) the one in the parent's blocktable. The base key
1903 * of course will still match.
1905 if (chain->flags & HAMMER2_CHAIN_BMAPPED)
1906 atomic_set_int(&chain->flags, HAMMER2_CHAIN_BMAPUPD);
1909 * Short-cut data blocks which the caller does not need an actual
1910 * data reference to (aka OPTDATA), as long as the chain does not
1911 * already have a data pointer to the data. This generally means
1912 * that the modifications are being done via the logical buffer cache.
1913 * The INITIAL flag relates only to the device data buffer and thus
1914 * remains unchange in this situation.
1916 * This code also handles bytes == 0 (most dirents).
1918 if (chain->bref.type == HAMMER2_BREF_TYPE_DATA &&
1919 (flags & HAMMER2_MODIFY_OPTDATA) &&
1920 chain->data == NULL) {
1921 KKASSERT(chain->dio == NULL);
1926 * Clearing the INITIAL flag (for indirect blocks) indicates that
1927 * we've processed the uninitialized storage allocation.
1929 * If this flag is already clear we are likely in a copy-on-write
1930 * situation but we have to be sure NOT to bzero the storage if
1931 * no data is present.
1933 if (chain->flags & HAMMER2_CHAIN_INITIAL) {
1934 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1941 * Instantiate data buffer and possibly execute COW operation
1943 switch(chain->bref.type) {
1944 case HAMMER2_BREF_TYPE_VOLUME:
1945 case HAMMER2_BREF_TYPE_FREEMAP:
1947 * The data is embedded, no copy-on-write operation is
1950 KKASSERT(chain->dio == NULL);
1952 case HAMMER2_BREF_TYPE_DIRENT:
1954 * The data might be fully embedded.
1956 if (chain->bytes == 0) {
1957 KKASSERT(chain->dio == NULL);
1961 case HAMMER2_BREF_TYPE_INODE:
1962 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
1963 case HAMMER2_BREF_TYPE_DATA:
1964 case HAMMER2_BREF_TYPE_INDIRECT:
1965 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1967 * Perform the copy-on-write operation
1969 * zero-fill or copy-on-write depending on whether
1970 * chain->data exists or not and set the dirty state for
1971 * the new buffer. hammer2_io_new() will handle the
1974 * If a dedup_off was supplied this is an existing block
1975 * and no COW, copy, or further modification is required.
1977 KKASSERT(chain != &hmp->vchain && chain != &hmp->fchain);
1979 if (wasinitial && dedup_off == 0) {
1980 error = hammer2_io_new(hmp, chain->bref.type,
1981 chain->bref.data_off,
1982 chain->bytes, &dio);
1984 error = hammer2_io_bread(hmp, chain->bref.type,
1985 chain->bref.data_off,
1986 chain->bytes, &dio);
1988 hammer2_adjreadcounter(&chain->bref, chain->bytes);
1991 * If an I/O error occurs make sure callers cannot accidently
1992 * modify the old buffer's contents and corrupt the filesystem.
1994 * NOTE: hammer2_io_data() call issues bkvasync()
1997 kprintf("hammer2_chain_modify: hmp=%p I/O error\n",
1999 chain->error = HAMMER2_ERROR_EIO;
2000 hammer2_io_brelse(&dio);
2001 hammer2_io_brelse(&chain->dio);
2006 bdata = hammer2_io_data(dio, chain->bref.data_off);
2010 * COW (unless a dedup).
2012 KKASSERT(chain->dio != NULL);
2013 if (chain->data != (void *)bdata && dedup_off == 0) {
2014 bcopy(chain->data, bdata, chain->bytes);
2016 } else if (wasinitial == 0) {
2018 * We have a problem. We were asked to COW but
2019 * we don't have any data to COW with!
2021 panic("hammer2_chain_modify: having a COW %p\n",
2026 * Retire the old buffer, replace with the new. Dirty or
2027 * redirty the new buffer.
2029 * WARNING! The system buffer cache may have already flushed
2030 * the buffer, so we must be sure to [re]dirty it
2031 * for further modification.
2033 * If dedup_off was supplied, the caller is not
2034 * expected to make any further modification to the
2038 hammer2_io_bqrelse(&chain->dio);
2039 chain->data = (void *)bdata;
2042 hammer2_io_setdirty(dio);
2045 panic("hammer2_chain_modify: illegal non-embedded type %d",
2052 * setflush on parent indicating that the parent must recurse down
2053 * to us. Do not call on chain itself which might already have it
2057 hammer2_chain_setflush(chain->parent);
2058 return (chain->error);
2062 * Modify the chain associated with an inode.
2065 hammer2_chain_modify_ip(hammer2_inode_t *ip, hammer2_chain_t *chain,
2066 hammer2_tid_t mtid, int flags)
2070 hammer2_inode_modify(ip);
2071 error = hammer2_chain_modify(chain, mtid, 0, flags);
2077 * Volume header data locks
2080 hammer2_voldata_lock(hammer2_dev_t *hmp)
2082 lockmgr(&hmp->vollk, LK_EXCLUSIVE);
2086 hammer2_voldata_unlock(hammer2_dev_t *hmp)
2088 lockmgr(&hmp->vollk, LK_RELEASE);
2092 hammer2_voldata_modify(hammer2_dev_t *hmp)
2094 if ((hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) == 0) {
2095 atomic_add_long(&hammer2_count_modified_chains, 1);
2096 atomic_set_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
2097 hammer2_pfs_memory_inc(hmp->vchain.pmp);
2102 * This function returns the chain at the nearest key within the specified
2103 * range. The returned chain will be referenced but not locked.
2105 * This function will recurse through chain->rbtree as necessary and will
2106 * return a *key_nextp suitable for iteration. *key_nextp is only set if
2107 * the iteration value is less than the current value of *key_nextp.
2109 * The caller should use (*key_nextp) to calculate the actual range of
2110 * the returned element, which will be (key_beg to *key_nextp - 1), because
2111 * there might be another element which is superior to the returned element
2114 * (*key_nextp) can be passed as key_beg in an iteration only while non-NULL
2115 * chains continue to be returned. On EOF (*key_nextp) may overflow since
2116 * it will wind up being (key_end + 1).
2118 * WARNING! Must be called with child's spinlock held. Spinlock remains
2119 * held through the operation.
2121 struct hammer2_chain_find_info {
2122 hammer2_chain_t *best;
2123 hammer2_key_t key_beg;
2124 hammer2_key_t key_end;
2125 hammer2_key_t key_next;
2128 static int hammer2_chain_find_cmp(hammer2_chain_t *child, void *data);
2129 static int hammer2_chain_find_callback(hammer2_chain_t *child, void *data);
2133 hammer2_chain_find(hammer2_chain_t *parent, hammer2_key_t *key_nextp,
2134 hammer2_key_t key_beg, hammer2_key_t key_end)
2136 struct hammer2_chain_find_info info;
2139 info.key_beg = key_beg;
2140 info.key_end = key_end;
2141 info.key_next = *key_nextp;
2143 RB_SCAN(hammer2_chain_tree, &parent->core.rbtree,
2144 hammer2_chain_find_cmp, hammer2_chain_find_callback,
2146 *key_nextp = info.key_next;
2148 kprintf("chain_find %p %016jx:%016jx next=%016jx\n",
2149 parent, key_beg, key_end, *key_nextp);
2157 hammer2_chain_find_cmp(hammer2_chain_t *child, void *data)
2159 struct hammer2_chain_find_info *info = data;
2160 hammer2_key_t child_beg;
2161 hammer2_key_t child_end;
2163 child_beg = child->bref.key;
2164 child_end = child_beg + ((hammer2_key_t)1 << child->bref.keybits) - 1;
2166 if (child_end < info->key_beg)
2168 if (child_beg > info->key_end)
2175 hammer2_chain_find_callback(hammer2_chain_t *child, void *data)
2177 struct hammer2_chain_find_info *info = data;
2178 hammer2_chain_t *best;
2179 hammer2_key_t child_end;
2182 * WARNING! Layerq is scanned forwards, exact matches should keep
2183 * the existing info->best.
2185 if ((best = info->best) == NULL) {
2187 * No previous best. Assign best
2190 } else if (best->bref.key <= info->key_beg &&
2191 child->bref.key <= info->key_beg) {
2196 /*info->best = child;*/
2197 } else if (child->bref.key < best->bref.key) {
2199 * Child has a nearer key and best is not flush with key_beg.
2200 * Set best to child. Truncate key_next to the old best key.
2203 if (info->key_next > best->bref.key || info->key_next == 0)
2204 info->key_next = best->bref.key;
2205 } else if (child->bref.key == best->bref.key) {
2207 * If our current best is flush with the child then this
2208 * is an illegal overlap.
2210 * key_next will automatically be limited to the smaller of
2211 * the two end-points.
2217 * Keep the current best but truncate key_next to the child's
2220 * key_next will also automatically be limited to the smaller
2221 * of the two end-points (probably not necessary for this case
2222 * but we do it anyway).
2224 if (info->key_next > child->bref.key || info->key_next == 0)
2225 info->key_next = child->bref.key;
2229 * Always truncate key_next based on child's end-of-range.
2231 child_end = child->bref.key + ((hammer2_key_t)1 << child->bref.keybits);
2232 if (child_end && (info->key_next > child_end || info->key_next == 0))
2233 info->key_next = child_end;
2239 * Retrieve the specified chain from a media blockref, creating the
2240 * in-memory chain structure which reflects it. The returned chain is
2241 * held and locked according to (how) (HAMMER2_RESOLVE_*). The caller must
2242 * handle crc-checks and so forth, and should check chain->error before
2243 * assuming that the data is good.
2245 * To handle insertion races pass the INSERT_RACE flag along with the
2246 * generation number of the core. NULL will be returned if the generation
2247 * number changes before we have a chance to insert the chain. Insert
2248 * races can occur because the parent might be held shared.
2250 * Caller must hold the parent locked shared or exclusive since we may
2251 * need the parent's bref array to find our block.
2253 * WARNING! chain->pmp is always set to NULL for any chain representing
2254 * part of the super-root topology.
2257 hammer2_chain_get(hammer2_chain_t *parent, int generation,
2258 hammer2_blockref_t *bref, int how)
2260 hammer2_dev_t *hmp = parent->hmp;
2261 hammer2_chain_t *chain;
2265 * Allocate a chain structure representing the existing media
2266 * entry. Resulting chain has one ref and is not locked.
2268 if (bref->flags & HAMMER2_BREF_FLAG_PFSROOT)
2269 chain = hammer2_chain_alloc(hmp, NULL, bref);
2271 chain = hammer2_chain_alloc(hmp, parent->pmp, bref);
2272 /* ref'd chain returned */
2275 * Flag that the chain is in the parent's blockmap so delete/flush
2276 * knows what to do with it.
2278 atomic_set_int(&chain->flags, HAMMER2_CHAIN_BMAPPED);
2281 * chain must be locked to avoid unexpected ripouts
2283 hammer2_chain_lock(chain, how);
2286 * Link the chain into its parent. A spinlock is required to safely
2287 * access the RBTREE, and it is possible to collide with another
2288 * hammer2_chain_get() operation because the caller might only hold
2289 * a shared lock on the parent.
2291 * NOTE: Get races can occur quite often when we distribute
2292 * asynchronous read-aheads across multiple threads.
2294 KKASSERT(parent->refs > 0);
2295 error = hammer2_chain_insert(parent, chain,
2296 HAMMER2_CHAIN_INSERT_SPIN |
2297 HAMMER2_CHAIN_INSERT_RACE,
2300 KKASSERT((chain->flags & HAMMER2_CHAIN_ONRBTREE) == 0);
2301 /*kprintf("chain %p get race\n", chain);*/
2302 hammer2_chain_unlock(chain);
2303 hammer2_chain_drop(chain);
2306 KKASSERT(chain->flags & HAMMER2_CHAIN_ONRBTREE);
2310 * Return our new chain referenced but not locked, or NULL if
2317 * Lookup initialization/completion API
2320 hammer2_chain_lookup_init(hammer2_chain_t *parent, int flags)
2322 hammer2_chain_ref(parent);
2323 if (flags & HAMMER2_LOOKUP_SHARED) {
2324 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS |
2325 HAMMER2_RESOLVE_SHARED);
2327 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2333 hammer2_chain_lookup_done(hammer2_chain_t *parent)
2336 hammer2_chain_unlock(parent);
2337 hammer2_chain_drop(parent);
2342 * Take the locked chain and return a locked parent. The chain remains
2343 * locked on return, but may have to be temporarily unlocked to acquire
2344 * the parent. Because of this, (chain) must be stable and cannot be
2345 * deleted while it was temporarily unlocked (typically means that (chain)
2348 * Pass HAMMER2_RESOLVE_* flags in flags.
2350 * This will work even if the chain is errored, and the caller can check
2351 * parent->error on return if desired since the parent will be locked.
2353 * This function handles the lock order reversal.
2356 hammer2_chain_getparent(hammer2_chain_t *chain, int flags)
2358 hammer2_chain_t *parent;
2361 * Be careful of order, chain must be unlocked before parent
2362 * is locked below to avoid a deadlock. Try it trivially first.
2364 parent = chain->parent;
2366 panic("hammer2_chain_getparent: no parent");
2367 hammer2_chain_ref(parent);
2368 if (hammer2_chain_lock(parent, flags|HAMMER2_RESOLVE_NONBLOCK) == 0)
2372 hammer2_chain_unlock(chain);
2373 hammer2_chain_lock(parent, flags);
2374 hammer2_chain_lock(chain, flags);
2377 * Parent relinking races are quite common. We have to get
2378 * it right or we will blow up the block table.
2380 if (chain->parent == parent)
2382 hammer2_chain_unlock(parent);
2383 hammer2_chain_drop(parent);
2385 parent = chain->parent;
2387 panic("hammer2_chain_getparent: no parent");
2388 hammer2_chain_ref(parent);
2394 * Take the locked chain and return a locked parent. The chain is unlocked
2395 * and dropped. *chainp is set to the returned parent as a convenience.
2396 * Pass HAMMER2_RESOLVE_* flags in flags.
2398 * This will work even if the chain is errored, and the caller can check
2399 * parent->error on return if desired since the parent will be locked.
2401 * The chain does NOT need to be stable. We use a tracking structure
2402 * to track the expected parent if the chain is deleted out from under us.
2404 * This function handles the lock order reversal.
2407 hammer2_chain_repparent(hammer2_chain_t **chainp, int flags)
2409 hammer2_chain_t *chain;
2410 hammer2_chain_t *parent;
2411 struct hammer2_reptrack reptrack;
2412 struct hammer2_reptrack **repp;
2415 * Be careful of order, chain must be unlocked before parent
2416 * is locked below to avoid a deadlock. Try it trivially first.
2419 parent = chain->parent;
2420 if (parent == NULL) {
2421 hammer2_spin_unex(&chain->core.spin);
2422 panic("hammer2_chain_repparent: no parent");
2424 hammer2_chain_ref(parent);
2425 if (hammer2_chain_lock(parent, flags|HAMMER2_RESOLVE_NONBLOCK) == 0) {
2426 hammer2_chain_unlock(chain);
2427 hammer2_chain_drop(chain);
2434 * Ok, now it gets a bit nasty. There are multiple situations where
2435 * the parent might be in the middle of a deletion, or where the child
2436 * (chain) might be deleted the instant we let go of its lock.
2437 * We can potentially end up in a no-win situation!
2439 * In particular, the indirect_maintenance() case can cause these
2442 * To deal with this we install a reptrack structure in the parent
2443 * This reptrack structure 'owns' the parent ref and will automatically
2444 * migrate to the parent's parent if the parent is deleted permanently.
2446 hammer2_spin_init(&reptrack.spin, "h2reptrk");
2447 reptrack.chain = parent;
2448 hammer2_chain_ref(parent); /* for the reptrack */
2450 hammer2_spin_ex(&parent->core.spin);
2451 reptrack.next = parent->core.reptrack;
2452 parent->core.reptrack = &reptrack;
2453 hammer2_spin_unex(&parent->core.spin);
2455 hammer2_chain_unlock(chain);
2456 hammer2_chain_drop(chain);
2457 chain = NULL; /* gone */
2460 * At the top of this loop, chain is gone and parent is refd both
2461 * by us explicitly AND via our reptrack. We are attempting to
2465 hammer2_chain_lock(parent, flags);
2467 if (reptrack.chain == parent)
2469 hammer2_chain_unlock(parent);
2470 hammer2_chain_drop(parent);
2472 kprintf("hammer2: debug REPTRACK %p->%p\n",
2473 parent, reptrack.chain);
2474 hammer2_spin_ex(&reptrack.spin);
2475 parent = reptrack.chain;
2476 hammer2_chain_ref(parent);
2477 hammer2_spin_unex(&reptrack.spin);
2481 * Once parent is locked and matches our reptrack, our reptrack
2482 * will be stable and we have our parent. We can unlink our
2485 * WARNING! Remember that the chain lock might be shared. Chains
2486 * locked shared have stable parent linkages.
2488 hammer2_spin_ex(&parent->core.spin);
2489 repp = &parent->core.reptrack;
2490 while (*repp != &reptrack)
2491 repp = &(*repp)->next;
2492 *repp = reptrack.next;
2493 hammer2_spin_unex(&parent->core.spin);
2495 hammer2_chain_drop(parent); /* reptrack ref */
2496 *chainp = parent; /* return parent lock+ref */
2502 * Dispose of any linked reptrack structures in (chain) by shifting them to
2503 * (parent). Both (chain) and (parent) must be exclusively locked.
2505 * This is interlocked against any children of (chain) on the other side.
2506 * No children so remain as-of when this is called so we can test
2507 * core.reptrack without holding the spin-lock.
2509 * Used whenever the caller intends to permanently delete chains related
2510 * to topological recursions (BREF_TYPE_INDIRECT, BREF_TYPE_FREEMAP_NODE),
2511 * where the chains underneath the node being deleted are given a new parent
2512 * above the node being deleted.
2516 hammer2_chain_repchange(hammer2_chain_t *parent, hammer2_chain_t *chain)
2518 struct hammer2_reptrack *reptrack;
2520 KKASSERT(chain->core.live_count == 0 && RB_EMPTY(&chain->core.rbtree));
2521 while (chain->core.reptrack) {
2522 hammer2_spin_ex(&parent->core.spin);
2523 hammer2_spin_ex(&chain->core.spin);
2524 reptrack = chain->core.reptrack;
2525 if (reptrack == NULL) {
2526 hammer2_spin_unex(&chain->core.spin);
2527 hammer2_spin_unex(&parent->core.spin);
2530 hammer2_spin_ex(&reptrack->spin);
2531 chain->core.reptrack = reptrack->next;
2532 reptrack->chain = parent;
2533 reptrack->next = parent->core.reptrack;
2534 parent->core.reptrack = reptrack;
2535 hammer2_chain_ref(parent); /* reptrack */
2537 hammer2_spin_unex(&chain->core.spin);
2538 hammer2_spin_unex(&parent->core.spin);
2539 kprintf("hammer2: debug repchange %p %p->%p\n",
2540 reptrack, chain, parent);
2541 hammer2_chain_drop(chain); /* reptrack */
2546 * Locate the first chain whos key range overlaps (key_beg, key_end) inclusive.
2547 * (*parentp) typically points to an inode but can also point to a related
2548 * indirect block and this function will recurse upwards and find the inode
2549 * or the nearest undeleted indirect block covering the key range.
2551 * This function unconditionally sets *errorp, replacing any previous value.
2553 * (*parentp) must be exclusive or shared locked (depending on flags) and
2554 * referenced and can be an inode or an existing indirect block within the
2557 * If (*parent) is errored out, this function will not attempt to recurse
2558 * the radix tree and will return NULL along with an appropriate *errorp.
2559 * If NULL is returned and *errorp is 0, the requested lookup could not be
2562 * On return (*parentp) will be modified to point at the deepest parent chain
2563 * element encountered during the search, as a helper for an insertion or
2566 * The new (*parentp) will be locked shared or exclusive (depending on flags),
2567 * and referenced, and the old will be unlocked and dereferenced (no change
2568 * if they are both the same). This is particularly important if the caller
2569 * wishes to insert a new chain, (*parentp) will be set properly even if NULL
2570 * is returned, as long as no error occurred.
2572 * The matching chain will be returned locked according to flags.
2576 * NULL is returned if no match was found, but (*parentp) will still
2577 * potentially be adjusted.
2579 * On return (*key_nextp) will point to an iterative value for key_beg.
2580 * (If NULL is returned (*key_nextp) is set to (key_end + 1)).
2582 * This function will also recurse up the chain if the key is not within the
2583 * current parent's range. (*parentp) can never be set to NULL. An iteration
2584 * can simply allow (*parentp) to float inside the loop.
2586 * NOTE! chain->data is not always resolved. By default it will not be
2587 * resolved for BREF_TYPE_DATA, FREEMAP_NODE, or FREEMAP_LEAF. Use
2588 * HAMMER2_LOOKUP_ALWAYS to force resolution (but be careful w/
2589 * BREF_TYPE_DATA as the device buffer can alias the logical file
2594 hammer2_chain_lookup(hammer2_chain_t **parentp, hammer2_key_t *key_nextp,
2595 hammer2_key_t key_beg, hammer2_key_t key_end,
2596 int *errorp, int flags)
2599 hammer2_chain_t *parent;
2600 hammer2_chain_t *chain;
2601 hammer2_blockref_t *base;
2602 hammer2_blockref_t *bref;
2603 hammer2_blockref_t bcopy;
2604 hammer2_key_t scan_beg;
2605 hammer2_key_t scan_end;
2607 int how_always = HAMMER2_RESOLVE_ALWAYS;
2608 int how_maybe = HAMMER2_RESOLVE_MAYBE;
2611 int maxloops = 300000;
2612 volatile hammer2_mtx_t save_mtx;
2614 if (flags & HAMMER2_LOOKUP_ALWAYS) {
2615 how_maybe = how_always;
2616 how = HAMMER2_RESOLVE_ALWAYS;
2617 } else if (flags & HAMMER2_LOOKUP_NODATA) {
2618 how = HAMMER2_RESOLVE_NEVER;
2620 how = HAMMER2_RESOLVE_MAYBE;
2622 if (flags & HAMMER2_LOOKUP_SHARED) {
2623 how_maybe |= HAMMER2_RESOLVE_SHARED;
2624 how_always |= HAMMER2_RESOLVE_SHARED;
2625 how |= HAMMER2_RESOLVE_SHARED;
2629 * Recurse (*parentp) upward if necessary until the parent completely
2630 * encloses the key range or we hit the inode.
2632 * Handle races against the flusher deleting indirect nodes on its
2633 * way back up by continuing to recurse upward past the deletion.
2639 while (parent->bref.type == HAMMER2_BREF_TYPE_INDIRECT ||
2640 parent->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2641 scan_beg = parent->bref.key;
2642 scan_end = scan_beg +
2643 ((hammer2_key_t)1 << parent->bref.keybits) - 1;
2644 if ((parent->flags & HAMMER2_CHAIN_DELETED) == 0) {
2645 if (key_beg >= scan_beg && key_end <= scan_end)
2648 parent = hammer2_chain_repparent(parentp, how_maybe);
2651 if (--maxloops == 0)
2652 panic("hammer2_chain_lookup: maxloops");
2654 * Locate the blockref array. Currently we do a fully associative
2655 * search through the array.
2657 switch(parent->bref.type) {
2658 case HAMMER2_BREF_TYPE_INODE:
2660 * Special shortcut for embedded data returns the inode
2661 * itself. Callers must detect this condition and access
2662 * the embedded data (the strategy code does this for us).
2664 * This is only applicable to regular files and softlinks.
2666 * We need a second lock on parent. Since we already have
2667 * a lock we must pass LOCKAGAIN to prevent unexpected
2668 * blocking (we don't want to block on a second shared
2669 * ref if an exclusive lock is pending)
2671 if (parent->data->ipdata.meta.op_flags &
2672 HAMMER2_OPFLAG_DIRECTDATA) {
2673 if (flags & HAMMER2_LOOKUP_NODIRECT) {
2675 *key_nextp = key_end + 1;
2678 hammer2_chain_ref(parent);
2679 hammer2_chain_lock(parent, how_always |
2680 HAMMER2_RESOLVE_LOCKAGAIN);
2681 *key_nextp = key_end + 1;
2684 base = &parent->data->ipdata.u.blockset.blockref[0];
2685 count = HAMMER2_SET_COUNT;
2687 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2688 case HAMMER2_BREF_TYPE_INDIRECT:
2690 * Handle MATCHIND on the parent
2692 if (flags & HAMMER2_LOOKUP_MATCHIND) {
2693 scan_beg = parent->bref.key;
2694 scan_end = scan_beg +
2695 ((hammer2_key_t)1 << parent->bref.keybits) - 1;
2696 if (key_beg == scan_beg && key_end == scan_end) {
2698 hammer2_chain_ref(chain);
2699 hammer2_chain_lock(chain, how_maybe);
2700 *key_nextp = scan_end + 1;
2706 * Optimize indirect blocks in the INITIAL state to avoid
2709 if (parent->flags & HAMMER2_CHAIN_INITIAL) {
2712 if (parent->data == NULL) {
2713 kprintf("parent->data is NULL %p\n", parent);
2715 tsleep(parent, 0, "xxx", 0);
2717 base = &parent->data->npdata[0];
2719 count = parent->bytes / sizeof(hammer2_blockref_t);
2721 case HAMMER2_BREF_TYPE_VOLUME:
2722 base = &parent->data->voldata.sroot_blockset.blockref[0];
2723 count = HAMMER2_SET_COUNT;
2725 case HAMMER2_BREF_TYPE_FREEMAP:
2726 base = &parent->data->blkset.blockref[0];
2727 count = HAMMER2_SET_COUNT;
2730 kprintf("hammer2_chain_lookup: unrecognized "
2731 "blockref(B) type: %d",
2734 tsleep(&base, 0, "dead", 0);
2735 panic("hammer2_chain_lookup: unrecognized "
2736 "blockref(B) type: %d",
2738 base = NULL; /* safety */
2739 count = 0; /* safety */
2743 * No lookup is possible if the parent is errored. We delayed
2744 * this check as long as we could to ensure that the parent backup,
2745 * embedded data, and MATCHIND code could still execute.
2747 if (parent->error) {
2748 *errorp = parent->error;
2753 * Merged scan to find next candidate.
2755 * hammer2_base_*() functions require the parent->core.live_* fields
2756 * to be synchronized.
2758 * We need to hold the spinlock to access the block array and RB tree
2759 * and to interlock chain creation.
2761 if ((parent->flags & HAMMER2_CHAIN_COUNTEDBREFS) == 0)
2762 hammer2_chain_countbrefs(parent, base, count);
2767 hammer2_spin_ex(&parent->core.spin);
2768 chain = hammer2_combined_find(parent, base, count,
2772 generation = parent->core.generation;
2775 * Exhausted parent chain, iterate.
2778 KKASSERT(chain == NULL);
2779 hammer2_spin_unex(&parent->core.spin);
2780 if (key_beg == key_end) /* short cut single-key case */
2784 * Stop if we reached the end of the iteration.
2786 if (parent->bref.type != HAMMER2_BREF_TYPE_INDIRECT &&
2787 parent->bref.type != HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2792 * Calculate next key, stop if we reached the end of the
2793 * iteration, otherwise go up one level and loop.
2795 key_beg = parent->bref.key +
2796 ((hammer2_key_t)1 << parent->bref.keybits);
2797 if (key_beg == 0 || key_beg > key_end)
2799 parent = hammer2_chain_repparent(parentp, how_maybe);
2804 * Selected from blockref or in-memory chain.
2807 if (chain == NULL) {
2808 hammer2_spin_unex(&parent->core.spin);
2809 if (bcopy.type == HAMMER2_BREF_TYPE_INDIRECT ||
2810 bcopy.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2811 chain = hammer2_chain_get(parent, generation,
2814 chain = hammer2_chain_get(parent, generation,
2820 hammer2_chain_ref(chain);
2821 hammer2_spin_unex(&parent->core.spin);
2824 * chain is referenced but not locked. We must lock the
2825 * chain to obtain definitive state.
2827 if (bcopy.type == HAMMER2_BREF_TYPE_INDIRECT ||
2828 bcopy.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2829 hammer2_chain_lock(chain, how_maybe);
2831 hammer2_chain_lock(chain, how);
2833 KKASSERT(chain->parent == parent);
2835 if (bcmp(&bcopy, &chain->bref, sizeof(bcopy)) ||
2836 chain->parent != parent) {
2837 hammer2_chain_unlock(chain);
2838 hammer2_chain_drop(chain);
2839 chain = NULL; /* SAFETY */
2845 * Skip deleted chains (XXX cache 'i' end-of-block-array? XXX)
2847 * NOTE: Chain's key range is not relevant as there might be
2848 * one-offs within the range that are not deleted.
2850 * NOTE: Lookups can race delete-duplicate because
2851 * delete-duplicate does not lock the parent's core
2852 * (they just use the spinlock on the core).
2854 if (chain->flags & HAMMER2_CHAIN_DELETED) {
2855 kprintf("skip deleted chain %016jx.%02x key=%016jx\n",
2856 chain->bref.data_off, chain->bref.type,
2858 hammer2_chain_unlock(chain);
2859 hammer2_chain_drop(chain);
2860 chain = NULL; /* SAFETY */
2861 key_beg = *key_nextp;
2862 if (key_beg == 0 || key_beg > key_end)
2868 * If the chain element is an indirect block it becomes the new
2869 * parent and we loop on it. We must maintain our top-down locks
2870 * to prevent the flusher from interfering (i.e. doing a
2871 * delete-duplicate and leaving us recursing down a deleted chain).
2873 * The parent always has to be locked with at least RESOLVE_MAYBE
2874 * so we can access its data. It might need a fixup if the caller
2875 * passed incompatible flags. Be careful not to cause a deadlock
2876 * as a data-load requires an exclusive lock.
2878 * If HAMMER2_LOOKUP_MATCHIND is set and the indirect block's key
2879 * range is within the requested key range we return the indirect
2880 * block and do NOT loop. This is usually only used to acquire
2883 if (chain->bref.type == HAMMER2_BREF_TYPE_INDIRECT ||
2884 chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2885 save_mtx = parent->lock;
2886 hammer2_chain_unlock(parent);
2887 hammer2_chain_drop(parent);
2888 *parentp = parent = chain;
2889 chain = NULL; /* SAFETY */
2894 * All done, return the locked chain.
2896 * If the caller does not want a locked chain, replace the lock with
2897 * a ref. Perhaps this can eventually be optimized to not obtain the
2898 * lock in the first place for situations where the data does not
2899 * need to be resolved.
2901 * NOTE! A chain->error must be tested by the caller upon return.
2902 * *errorp is only set based on issues which occur while
2903 * trying to reach the chain.
2909 * After having issued a lookup we can iterate all matching keys.
2911 * If chain is non-NULL we continue the iteration from just after it's index.
2913 * If chain is NULL we assume the parent was exhausted and continue the
2914 * iteration at the next parent.
2916 * If a fatal error occurs (typically an I/O error), a dummy chain is
2917 * returned with chain->error and error-identifying information set. This
2918 * chain will assert if you try to do anything fancy with it.
2920 * XXX Depending on where the error occurs we should allow continued iteration.
2922 * parent must be locked on entry and remains locked throughout. chain's
2923 * lock status must match flags. Chain is always at least referenced.
2925 * WARNING! The MATCHIND flag does not apply to this function.
2928 hammer2_chain_next(hammer2_chain_t **parentp, hammer2_chain_t *chain,
2929 hammer2_key_t *key_nextp,
2930 hammer2_key_t key_beg, hammer2_key_t key_end,
2931 int *errorp, int flags)
2933 hammer2_chain_t *parent;
2937 * Calculate locking flags for upward recursion.
2939 how_maybe = HAMMER2_RESOLVE_MAYBE;
2940 if (flags & HAMMER2_LOOKUP_SHARED)
2941 how_maybe |= HAMMER2_RESOLVE_SHARED;
2947 * Calculate the next index and recalculate the parent if necessary.
2950 key_beg = chain->bref.key +
2951 ((hammer2_key_t)1 << chain->bref.keybits);
2952 hammer2_chain_unlock(chain);
2953 hammer2_chain_drop(chain);
2956 * chain invalid past this point, but we can still do a
2957 * pointer comparison w/parent.
2959 * Any scan where the lookup returned degenerate data embedded
2960 * in the inode has an invalid index and must terminate.
2962 if (chain == parent)
2964 if (key_beg == 0 || key_beg > key_end)
2967 } else if (parent->bref.type != HAMMER2_BREF_TYPE_INDIRECT &&
2968 parent->bref.type != HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2970 * We reached the end of the iteration.
2975 * Continue iteration with next parent unless the current
2976 * parent covers the range.
2978 * (This also handles the case of a deleted, empty indirect
2981 key_beg = parent->bref.key +
2982 ((hammer2_key_t)1 << parent->bref.keybits);
2983 if (key_beg == 0 || key_beg > key_end)
2985 parent = hammer2_chain_repparent(parentp, how_maybe);
2991 return (hammer2_chain_lookup(parentp, key_nextp,
2997 * Caller wishes to iterate chains under parent, loading new chains into
2998 * chainp. Caller must initialize *chainp to NULL and *firstp to 1, and
2999 * then call hammer2_chain_scan() repeatedly until a non-zero return.
3000 * During the scan, *firstp will be set to 0 and (*chainp) will be replaced
3001 * with the returned chain for the scan. The returned *chainp will be
3002 * locked and referenced. Any prior contents will be unlocked and dropped.
3004 * Caller should check the return value. A normal scan EOF will return
3005 * exactly HAMMER2_ERROR_EOF. Any other non-zero value indicates an
3006 * error trying to access parent data. Any error in the returned chain
3007 * must be tested separately by the caller.
3009 * (*chainp) is dropped on each scan, but will only be set if the returned
3010 * element itself can recurse. Leaf elements are NOT resolved, loaded, or
3011 * returned via *chainp. The caller will get their bref only.
3013 * The raw scan function is similar to lookup/next but does not seek to a key.
3014 * Blockrefs are iterated via first_bref = (parent, NULL) and
3015 * next_chain = (parent, bref).
3017 * The passed-in parent must be locked and its data resolved. The function
3018 * nominally returns a locked and referenced *chainp != NULL for chains
3019 * the caller might need to recurse on (and will dipose of any *chainp passed
3020 * in). The caller must check the chain->bref.type either way.
3023 hammer2_chain_scan(hammer2_chain_t *parent, hammer2_chain_t **chainp,
3024 hammer2_blockref_t *bref, int *firstp,
3028 hammer2_blockref_t *base;
3029 hammer2_blockref_t *bref_ptr;
3031 hammer2_key_t next_key;
3032 hammer2_chain_t *chain = NULL;
3034 int how_always = HAMMER2_RESOLVE_ALWAYS;
3035 int how_maybe = HAMMER2_RESOLVE_MAYBE;
3038 int maxloops = 300000;
3045 * Scan flags borrowed from lookup.
3047 if (flags & HAMMER2_LOOKUP_ALWAYS) {
3048 how_maybe = how_always;
3049 how = HAMMER2_RESOLVE_ALWAYS;
3050 } else if (flags & HAMMER2_LOOKUP_NODATA) {
3051 how = HAMMER2_RESOLVE_NEVER;
3053 how = HAMMER2_RESOLVE_MAYBE;
3055 if (flags & HAMMER2_LOOKUP_SHARED) {
3056 how_maybe |= HAMMER2_RESOLVE_SHARED;
3057 how_always |= HAMMER2_RESOLVE_SHARED;
3058 how |= HAMMER2_RESOLVE_SHARED;
3062 * Calculate key to locate first/next element, unlocking the previous
3063 * element as we go. Be careful, the key calculation can overflow.
3065 * (also reset bref to NULL)
3071 key = bref->key + ((hammer2_key_t)1 << bref->keybits);
3072 if ((chain = *chainp) != NULL) {
3074 hammer2_chain_unlock(chain);
3075 hammer2_chain_drop(chain);
3079 error |= HAMMER2_ERROR_EOF;
3085 if (parent->error) {
3086 error = parent->error;
3089 if (--maxloops == 0)
3090 panic("hammer2_chain_scan: maxloops");
3093 * Locate the blockref array. Currently we do a fully associative
3094 * search through the array.
3096 switch(parent->bref.type) {
3097 case HAMMER2_BREF_TYPE_INODE:
3099 * An inode with embedded data has no sub-chains.
3101 * WARNING! Bulk scan code may pass a static chain marked
3102 * as BREF_TYPE_INODE with a copy of the volume
3103 * root blockset to snapshot the volume.
3105 if (parent->data->ipdata.meta.op_flags &
3106 HAMMER2_OPFLAG_DIRECTDATA) {
3107 error |= HAMMER2_ERROR_EOF;
3110 base = &parent->data->ipdata.u.blockset.blockref[0];
3111 count = HAMMER2_SET_COUNT;
3113 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3114 case HAMMER2_BREF_TYPE_INDIRECT:
3116 * Optimize indirect blocks in the INITIAL state to avoid
3119 if (parent->flags & HAMMER2_CHAIN_INITIAL) {
3122 if (parent->data == NULL)
3123 panic("parent->data is NULL");
3124 base = &parent->data->npdata[0];
3126 count = parent->bytes / sizeof(hammer2_blockref_t);
3128 case HAMMER2_BREF_TYPE_VOLUME:
3129 base = &parent->data->voldata.sroot_blockset.blockref[0];
3130 count = HAMMER2_SET_COUNT;
3132 case HAMMER2_BREF_TYPE_FREEMAP:
3133 base = &parent->data->blkset.blockref[0];
3134 count = HAMMER2_SET_COUNT;
3137 panic("hammer2_chain_lookup: unrecognized blockref type: %d",
3139 base = NULL; /* safety */
3140 count = 0; /* safety */
3144 * Merged scan to find next candidate.
3146 * hammer2_base_*() functions require the parent->core.live_* fields
3147 * to be synchronized.
3149 * We need to hold the spinlock to access the block array and RB tree
3150 * and to interlock chain creation.
3152 if ((parent->flags & HAMMER2_CHAIN_COUNTEDBREFS) == 0)
3153 hammer2_chain_countbrefs(parent, base, count);
3157 hammer2_spin_ex(&parent->core.spin);
3158 chain = hammer2_combined_find(parent, base, count,
3160 key, HAMMER2_KEY_MAX,
3162 generation = parent->core.generation;
3165 * Exhausted parent chain, we're done.
3167 if (bref_ptr == NULL) {
3168 hammer2_spin_unex(&parent->core.spin);
3169 KKASSERT(chain == NULL);
3170 error |= HAMMER2_ERROR_EOF;
3175 * Copy into the supplied stack-based blockref.
3180 * Selected from blockref or in-memory chain.
3182 if (chain == NULL) {
3183 switch(bref->type) {
3184 case HAMMER2_BREF_TYPE_INODE:
3185 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3186 case HAMMER2_BREF_TYPE_INDIRECT:
3187 case HAMMER2_BREF_TYPE_VOLUME:
3188 case HAMMER2_BREF_TYPE_FREEMAP:
3190 * Recursion, always get the chain
3192 hammer2_spin_unex(&parent->core.spin);
3193 chain = hammer2_chain_get(parent, generation,
3200 * No recursion, do not waste time instantiating
3201 * a chain, just iterate using the bref.
3203 hammer2_spin_unex(&parent->core.spin);
3208 * Recursion or not we need the chain in order to supply
3211 hammer2_chain_ref(chain);
3212 hammer2_spin_unex(&parent->core.spin);
3213 hammer2_chain_lock(chain, how);
3216 (bcmp(bref, &chain->bref, sizeof(*bref)) ||
3217 chain->parent != parent)) {
3218 hammer2_chain_unlock(chain);
3219 hammer2_chain_drop(chain);
3225 * Skip deleted chains (XXX cache 'i' end-of-block-array? XXX)
3227 * NOTE: chain's key range is not relevant as there might be
3228 * one-offs within the range that are not deleted.
3230 * NOTE: XXX this could create problems with scans used in
3231 * situations other than mount-time recovery.
3233 * NOTE: Lookups can race delete-duplicate because
3234 * delete-duplicate does not lock the parent's core
3235 * (they just use the spinlock on the core).
3237 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) {
3238 hammer2_chain_unlock(chain);
3239 hammer2_chain_drop(chain);
3244 error |= HAMMER2_ERROR_EOF;
3252 * All done, return the bref or NULL, supply chain if necessary.
3260 * Create and return a new hammer2 system memory structure of the specified
3261 * key, type and size and insert it under (*parentp). This is a full
3262 * insertion, based on the supplied key/keybits, and may involve creating
3263 * indirect blocks and moving other chains around via delete/duplicate.
3265 * THE CALLER MUST HAVE ALREADY PROPERLY SEEKED (*parentp) TO THE INSERTION
3266 * POINT SANS ANY REQUIRED INDIRECT BLOCK CREATIONS DUE TO THE ARRAY BEING
3267 * FULL. This typically means that the caller is creating the chain after
3268 * doing a hammer2_chain_lookup().
3270 * (*parentp) must be exclusive locked and may be replaced on return
3271 * depending on how much work the function had to do.
3273 * (*parentp) must not be errored or this function will assert.
3275 * (*chainp) usually starts out NULL and returns the newly created chain,
3276 * but if the caller desires the caller may allocate a disconnected chain
3277 * and pass it in instead.
3279 * This function should NOT be used to insert INDIRECT blocks. It is
3280 * typically used to create/insert inodes and data blocks.
3282 * Caller must pass-in an exclusively locked parent the new chain is to
3283 * be inserted under, and optionally pass-in a disconnected, exclusively
3284 * locked chain to insert (else we create a new chain). The function will
3285 * adjust (*parentp) as necessary, create or connect the chain, and
3286 * return an exclusively locked chain in *chainp.
3288 * When creating a PFSROOT inode under the super-root, pmp is typically NULL
3289 * and will be reassigned.
3291 * NOTE: returns HAMMER_ERROR_* flags
3294 hammer2_chain_create(hammer2_chain_t **parentp, hammer2_chain_t **chainp,
3295 hammer2_pfs_t *pmp, int methods,
3296 hammer2_key_t key, int keybits, int type, size_t bytes,
3297 hammer2_tid_t mtid, hammer2_off_t dedup_off, int flags)
3300 hammer2_chain_t *chain;
3301 hammer2_chain_t *parent;
3302 hammer2_blockref_t *base;
3303 hammer2_blockref_t dummy;
3307 int maxloops = 300000;
3310 * Topology may be crossing a PFS boundary.
3313 KKASSERT(hammer2_mtx_owned(&parent->lock));
3314 KKASSERT(parent->error == 0);
3318 if (chain == NULL) {
3320 * First allocate media space and construct the dummy bref,
3321 * then allocate the in-memory chain structure. Set the
3322 * INITIAL flag for fresh chains which do not have embedded
3325 * XXX for now set the check mode of the child based on
3326 * the parent or, if the parent is an inode, the
3327 * specification in the inode.
3329 bzero(&dummy, sizeof(dummy));
3332 dummy.keybits = keybits;
3333 dummy.data_off = hammer2_getradix(bytes);
3336 * Inherit methods from parent by default. Primarily used
3337 * for BREF_TYPE_DATA. Non-data types *must* be set to
3338 * a non-NONE check algorithm.
3341 dummy.methods = parent->bref.methods;
3343 dummy.methods = (uint8_t)methods;
3345 if (type != HAMMER2_BREF_TYPE_DATA &&
3346 HAMMER2_DEC_CHECK(dummy.methods) == HAMMER2_CHECK_NONE) {
3348 HAMMER2_ENC_CHECK(HAMMER2_CHECK_DEFAULT);
3351 chain = hammer2_chain_alloc(hmp, pmp, &dummy);
3354 * Lock the chain manually, chain_lock will load the chain
3355 * which we do NOT want to do. (note: chain->refs is set
3356 * to 1 by chain_alloc() for us, but lockcnt is not).
3359 hammer2_mtx_ex(&chain->lock);
3361 ++curthread->td_tracker;
3364 * Set INITIAL to optimize I/O. The flag will generally be
3365 * processed when we call hammer2_chain_modify().
3367 * Recalculate bytes to reflect the actual media block
3368 * allocation. Handle special case radix 0 == 0 bytes.
3370 bytes = (size_t)(chain->bref.data_off & HAMMER2_OFF_MASK_RADIX);
3372 bytes = (hammer2_off_t)1 << bytes;
3373 chain->bytes = bytes;
3376 case HAMMER2_BREF_TYPE_VOLUME:
3377 case HAMMER2_BREF_TYPE_FREEMAP:
3378 panic("hammer2_chain_create: called with volume type");
3380 case HAMMER2_BREF_TYPE_INDIRECT:
3381 panic("hammer2_chain_create: cannot be used to"
3382 "create indirect block");
3384 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3385 panic("hammer2_chain_create: cannot be used to"
3386 "create freemap root or node");
3388 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
3389 KKASSERT(bytes == sizeof(chain->data->bmdata));
3391 case HAMMER2_BREF_TYPE_DIRENT:
3392 case HAMMER2_BREF_TYPE_INODE:
3393 case HAMMER2_BREF_TYPE_DATA:
3396 * leave chain->data NULL, set INITIAL
3398 KKASSERT(chain->data == NULL);
3399 atomic_set_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
3404 * We are reattaching a previously deleted chain, possibly
3405 * under a new parent and possibly with a new key/keybits.
3406 * The chain does not have to be in a modified state. The
3407 * UPDATE flag will be set later on in this routine.
3409 * Do NOT mess with the current state of the INITIAL flag.
3411 chain->bref.key = key;
3412 chain->bref.keybits = keybits;
3413 if (chain->flags & HAMMER2_CHAIN_DELETED)
3414 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_DELETED);
3415 KKASSERT(chain->parent == NULL);
3419 * Set the appropriate bref flag if requested.
3421 * NOTE! Callers can call this function to move chains without
3422 * knowing about special flags, so don't clear bref flags
3425 if (flags & HAMMER2_INSERT_PFSROOT)
3426 chain->bref.flags |= HAMMER2_BREF_FLAG_PFSROOT;
3429 * Calculate how many entries we have in the blockref array and
3430 * determine if an indirect block is required.
3433 if (--maxloops == 0)
3434 panic("hammer2_chain_create: maxloops");
3436 switch(parent->bref.type) {
3437 case HAMMER2_BREF_TYPE_INODE:
3438 if ((parent->data->ipdata.meta.op_flags &
3439 HAMMER2_OPFLAG_DIRECTDATA) != 0) {
3440 kprintf("hammer2: parent set for direct-data! "
3441 "pkey=%016jx ckey=%016jx\n",
3445 KKASSERT((parent->data->ipdata.meta.op_flags &
3446 HAMMER2_OPFLAG_DIRECTDATA) == 0);
3447 KKASSERT(parent->data != NULL);
3448 base = &parent->data->ipdata.u.blockset.blockref[0];
3449 count = HAMMER2_SET_COUNT;
3451 case HAMMER2_BREF_TYPE_INDIRECT:
3452 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3453 if (parent->flags & HAMMER2_CHAIN_INITIAL)
3456 base = &parent->data->npdata[0];
3457 count = parent->bytes / sizeof(hammer2_blockref_t);
3459 case HAMMER2_BREF_TYPE_VOLUME:
3460 KKASSERT(parent->data != NULL);
3461 base = &parent->data->voldata.sroot_blockset.blockref[0];
3462 count = HAMMER2_SET_COUNT;
3464 case HAMMER2_BREF_TYPE_FREEMAP:
3465 KKASSERT(parent->data != NULL);
3466 base = &parent->data->blkset.blockref[0];
3467 count = HAMMER2_SET_COUNT;
3470 panic("hammer2_chain_create: unrecognized blockref type: %d",
3478 * Make sure we've counted the brefs
3480 if ((parent->flags & HAMMER2_CHAIN_COUNTEDBREFS) == 0)
3481 hammer2_chain_countbrefs(parent, base, count);
3483 KASSERT(parent->core.live_count >= 0 &&
3484 parent->core.live_count <= count,
3485 ("bad live_count %d/%d (%02x, %d)",
3486 parent->core.live_count, count,
3487 parent->bref.type, parent->bytes));
3490 * If no free blockref could be found we must create an indirect
3491 * block and move a number of blockrefs into it. With the parent
3492 * locked we can safely lock each child in order to delete+duplicate
3493 * it without causing a deadlock.
3495 * This may return the new indirect block or the old parent depending
3496 * on where the key falls. NULL is returned on error.
3498 if (parent->core.live_count == count) {
3499 hammer2_chain_t *nparent;
3501 KKASSERT((flags & HAMMER2_INSERT_SAMEPARENT) == 0);
3503 nparent = hammer2_chain_create_indirect(parent, key, keybits,
3504 mtid, type, &error);
3505 if (nparent == NULL) {
3507 hammer2_chain_drop(chain);
3511 if (parent != nparent) {
3512 hammer2_chain_unlock(parent);
3513 hammer2_chain_drop(parent);
3514 parent = *parentp = nparent;
3519 if (chain->flags & HAMMER2_CHAIN_DELETED)
3520 kprintf("Inserting deleted chain @%016jx\n",
3524 * Link the chain into its parent.
3526 if (chain->parent != NULL)
3527 panic("hammer2: hammer2_chain_create: chain already connected");
3528 KKASSERT(chain->parent == NULL);
3529 KKASSERT(parent->core.live_count < count);
3530 hammer2_chain_insert(parent, chain,
3531 HAMMER2_CHAIN_INSERT_SPIN |
3532 HAMMER2_CHAIN_INSERT_LIVE,
3537 * Mark the newly created chain modified. This will cause
3538 * UPDATE to be set and process the INITIAL flag.
3540 * Device buffers are not instantiated for DATA elements
3541 * as these are handled by logical buffers.
3543 * Indirect and freemap node indirect blocks are handled
3544 * by hammer2_chain_create_indirect() and not by this
3547 * Data for all other bref types is expected to be
3548 * instantiated (INODE, LEAF).
3550 switch(chain->bref.type) {
3551 case HAMMER2_BREF_TYPE_DATA:
3552 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
3553 case HAMMER2_BREF_TYPE_DIRENT:
3554 case HAMMER2_BREF_TYPE_INODE:
3555 error = hammer2_chain_modify(chain, mtid, dedup_off,
3556 HAMMER2_MODIFY_OPTDATA);
3560 * Remaining types are not supported by this function.
3561 * In particular, INDIRECT and LEAF_NODE types are
3562 * handled by create_indirect().
3564 panic("hammer2_chain_create: bad type: %d",
3571 * When reconnecting a chain we must set UPDATE and
3572 * setflush so the flush recognizes that it must update
3573 * the bref in the parent.
3575 if ((chain->flags & HAMMER2_CHAIN_UPDATE) == 0)
3576 atomic_set_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
3580 * We must setflush(parent) to ensure that it recurses through to
3581 * chain. setflush(chain) might not work because ONFLUSH is possibly
3582 * already set in the chain (so it won't recurse up to set it in the
3585 hammer2_chain_setflush(parent);
3594 * Move the chain from its old parent to a new parent. The chain must have
3595 * already been deleted or already disconnected (or never associated) with
3596 * a parent. The chain is reassociated with the new parent and the deleted
3597 * flag will be cleared (no longer deleted). The chain's modification state
3600 * THE CALLER MUST HAVE ALREADY PROPERLY SEEKED (parent) TO THE INSERTION
3601 * POINT SANS ANY REQUIRED INDIRECT BLOCK CREATIONS DUE TO THE ARRAY BEING
3602 * FULL. This typically means that the caller is creating the chain after
3603 * doing a hammer2_chain_lookup().
3605 * Neither (parent) or (chain) can be errored.
3607 * If (parent) is non-NULL then the chain is inserted under the parent.
3609 * If (parent) is NULL then the newly duplicated chain is not inserted
3610 * anywhere, similar to if it had just been chain_alloc()'d (suitable for
3611 * passing into hammer2_chain_create() after this function returns).
3613 * WARNING! This function calls create which means it can insert indirect
3614 * blocks. This can cause other unrelated chains in the parent to
3615 * be moved to a newly inserted indirect block in addition to the
3619 hammer2_chain_rename(hammer2_chain_t **parentp, hammer2_chain_t *chain,
3620 hammer2_tid_t mtid, int flags)
3622 hammer2_blockref_t *bref;
3624 hammer2_chain_t *parent;
3628 * WARNING! We should never resolve DATA to device buffers
3629 * (XXX allow it if the caller did?), and since
3630 * we currently do not have the logical buffer cache
3631 * buffer in-hand to fix its cached physical offset
3632 * we also force the modify code to not COW it. XXX
3635 KKASSERT(chain->parent == NULL);
3636 KKASSERT(chain->error == 0);
3639 * Now create a duplicate of the chain structure, associating
3640 * it with the same core, making it the same size, pointing it
3641 * to the same bref (the same media block).
3643 * NOTE: Handle special radix == 0 case (means 0 bytes).
3645 bref = &chain->bref;
3646 bytes = (size_t)(bref->data_off & HAMMER2_OFF_MASK_RADIX);
3648 bytes = (hammer2_off_t)1 << bytes;
3651 * If parent is not NULL the duplicated chain will be entered under
3652 * the parent and the UPDATE bit set to tell flush to update
3655 * We must setflush(parent) to ensure that it recurses through to
3656 * chain. setflush(chain) might not work because ONFLUSH is possibly
3657 * already set in the chain (so it won't recurse up to set it in the
3660 * Having both chains locked is extremely important for atomicy.
3662 if (parentp && (parent = *parentp) != NULL) {
3663 KKASSERT(hammer2_mtx_owned(&parent->lock));
3664 KKASSERT(parent->refs > 0);
3665 KKASSERT(parent->error == 0);
3667 hammer2_chain_create(parentp, &chain,
3668 chain->pmp, HAMMER2_METH_DEFAULT,
3669 bref->key, bref->keybits, bref->type,
3670 chain->bytes, mtid, 0, flags);
3671 KKASSERT(chain->flags & HAMMER2_CHAIN_UPDATE);
3672 hammer2_chain_setflush(*parentp);
3677 * Helper function for deleting chains.
3679 * The chain is removed from the live view (the RBTREE) as well as the parent's
3680 * blockmap. Both chain and its parent must be locked.
3682 * parent may not be errored. chain can be errored.
3685 _hammer2_chain_delete_helper(hammer2_chain_t *parent, hammer2_chain_t *chain,
3686 hammer2_tid_t mtid, int flags)
3691 KKASSERT((chain->flags & (HAMMER2_CHAIN_DELETED |
3692 HAMMER2_CHAIN_FICTITIOUS)) == 0);
3693 KKASSERT(chain->parent == parent);
3696 if (chain->flags & HAMMER2_CHAIN_BMAPPED) {
3698 * Chain is blockmapped, so there must be a parent.
3699 * Atomically remove the chain from the parent and remove
3700 * the blockmap entry. The parent must be set modified
3701 * to remove the blockmap entry.
3703 hammer2_blockref_t *base;
3706 KKASSERT(parent != NULL);
3707 KKASSERT(parent->error == 0);
3708 KKASSERT((parent->flags & HAMMER2_CHAIN_INITIAL) == 0);
3709 error = hammer2_chain_modify(parent, mtid, 0, 0);
3714 * Calculate blockmap pointer
3716 KKASSERT(chain->flags & HAMMER2_CHAIN_ONRBTREE);
3717 hammer2_spin_ex(&chain->core.spin);
3718 hammer2_spin_ex(&parent->core.spin);
3720 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DELETED);
3721 atomic_add_int(&parent->core.live_count, -1);
3722 ++parent->core.generation;
3723 RB_REMOVE(hammer2_chain_tree, &parent->core.rbtree, chain);
3724 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE);
3725 --parent->core.chain_count;
3726 chain->parent = NULL;
3728 switch(parent->bref.type) {
3729 case HAMMER2_BREF_TYPE_INODE:
3731 * Access the inode's block array. However, there
3732 * is no block array if the inode is flagged
3736 (parent->data->ipdata.meta.op_flags &
3737 HAMMER2_OPFLAG_DIRECTDATA) == 0) {
3739 &parent->data->ipdata.u.blockset.blockref[0];
3743 count = HAMMER2_SET_COUNT;
3745 case HAMMER2_BREF_TYPE_INDIRECT:
3746 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3748 base = &parent->data->npdata[0];
3751 count = parent->bytes / sizeof(hammer2_blockref_t);
3753 case HAMMER2_BREF_TYPE_VOLUME:
3754 base = &parent->data->voldata.
3755 sroot_blockset.blockref[0];
3756 count = HAMMER2_SET_COUNT;
3758 case HAMMER2_BREF_TYPE_FREEMAP:
3759 base = &parent->data->blkset.blockref[0];
3760 count = HAMMER2_SET_COUNT;
3765 panic("hammer2_flush_pass2: "
3766 "unrecognized blockref type: %d",
3771 * delete blockmapped chain from its parent.
3773 * The parent is not affected by any statistics in chain
3774 * which are pending synchronization. That is, there is
3775 * nothing to undo in the parent since they have not yet
3776 * been incorporated into the parent.
3778 * The parent is affected by statistics stored in inodes.
3779 * Those have already been synchronized, so they must be
3780 * undone. XXX split update possible w/delete in middle?
3783 hammer2_base_delete(parent, base, count, chain);
3785 hammer2_spin_unex(&parent->core.spin);
3786 hammer2_spin_unex(&chain->core.spin);
3787 } else if (chain->flags & HAMMER2_CHAIN_ONRBTREE) {
3789 * Chain is not blockmapped but a parent is present.
3790 * Atomically remove the chain from the parent. There is
3791 * no blockmap entry to remove.
3793 * Because chain was associated with a parent but not
3794 * synchronized, the chain's *_count_up fields contain
3795 * inode adjustment statistics which must be undone.
3797 hammer2_spin_ex(&chain->core.spin);
3798 hammer2_spin_ex(&parent->core.spin);
3799 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DELETED);
3800 atomic_add_int(&parent->core.live_count, -1);
3801 ++parent->core.generation;
3802 RB_REMOVE(hammer2_chain_tree, &parent->core.rbtree, chain);
3803 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE);
3804 --parent->core.chain_count;
3805 chain->parent = NULL;
3806 hammer2_spin_unex(&parent->core.spin);
3807 hammer2_spin_unex(&chain->core.spin);
3810 * Chain is not blockmapped and has no parent. This
3811 * is a degenerate case.
3813 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DELETED);
3820 * Create an indirect block that covers one or more of the elements in the
3821 * current parent. Either returns the existing parent with no locking or
3822 * ref changes or returns the new indirect block locked and referenced
3823 * and leaving the original parent lock/ref intact as well.
3825 * If an error occurs, NULL is returned and *errorp is set to the H2 error.
3827 * The returned chain depends on where the specified key falls.
3829 * The key/keybits for the indirect mode only needs to follow three rules:
3831 * (1) That all elements underneath it fit within its key space and
3833 * (2) That all elements outside it are outside its key space.
3835 * (3) When creating the new indirect block any elements in the current
3836 * parent that fit within the new indirect block's keyspace must be
3837 * moved into the new indirect block.
3839 * (4) The keyspace chosen for the inserted indirect block CAN cover a wider
3840 * keyspace the the current parent, but lookup/iteration rules will
3841 * ensure (and must ensure) that rule (2) for all parents leading up
3842 * to the nearest inode or the root volume header is adhered to. This
3843 * is accomplished by always recursing through matching keyspaces in
3844 * the hammer2_chain_lookup() and hammer2_chain_next() API.
3846 * The current implementation calculates the current worst-case keyspace by
3847 * iterating the current parent and then divides it into two halves, choosing
3848 * whichever half has the most elements (not necessarily the half containing
3849 * the requested key).
3851 * We can also opt to use the half with the least number of elements. This
3852 * causes lower-numbered keys (aka logical file offsets) to recurse through
3853 * fewer indirect blocks and higher-numbered keys to recurse through more.
3854 * This also has the risk of not moving enough elements to the new indirect
3855 * block and being forced to create several indirect blocks before the element
3858 * Must be called with an exclusively locked parent.
3860 * NOTE: *errorp set to HAMMER_ERROR_* flags
3862 static int hammer2_chain_indkey_freemap(hammer2_chain_t *parent,
3863 hammer2_key_t *keyp, int keybits,
3864 hammer2_blockref_t *base, int count);
3865 static int hammer2_chain_indkey_file(hammer2_chain_t *parent,
3866 hammer2_key_t *keyp, int keybits,
3867 hammer2_blockref_t *base, int count,
3869 static int hammer2_chain_indkey_dir(hammer2_chain_t *parent,
3870 hammer2_key_t *keyp, int keybits,
3871 hammer2_blockref_t *base, int count,
3875 hammer2_chain_create_indirect(hammer2_chain_t *parent,
3876 hammer2_key_t create_key, int create_bits,
3877 hammer2_tid_t mtid, int for_type, int *errorp)
3880 hammer2_blockref_t *base;
3881 hammer2_blockref_t *bref;
3882 hammer2_blockref_t bcopy;
3883 hammer2_chain_t *chain;
3884 hammer2_chain_t *ichain;
3885 hammer2_chain_t dummy;
3886 hammer2_key_t key = create_key;
3887 hammer2_key_t key_beg;
3888 hammer2_key_t key_end;
3889 hammer2_key_t key_next;
3890 int keybits = create_bits;
3898 int maxloops = 300000;
3901 * Calculate the base blockref pointer or NULL if the chain
3902 * is known to be empty. We need to calculate the array count
3903 * for RB lookups either way.
3906 KKASSERT(hammer2_mtx_owned(&parent->lock));
3909 * Pre-modify the parent now to avoid having to deal with error
3910 * processing if we tried to later (in the middle of our loop).
3912 *errorp = hammer2_chain_modify(parent, mtid, 0, 0);
3914 kprintf("hammer2_create_indirect: error %08x %s\n",
3915 *errorp, hammer2_error_str(*errorp));
3919 /*hammer2_chain_modify(&parent, HAMMER2_MODIFY_OPTDATA);*/
3920 base = hammer2_chain_base_and_count(parent, &count);
3923 * dummy used in later chain allocation (no longer used for lookups).
3925 bzero(&dummy, sizeof(dummy));
3928 * How big should our new indirect block be? It has to be at least
3929 * as large as its parent for splits to work properly.
3931 * The freemap uses a specific indirect block size. The number of
3932 * levels are built dynamically and ultimately depend on the size
3933 * volume. Because freemap blocks are taken from the reserved areas
3934 * of the volume our goal is efficiency (fewer levels) and not so
3935 * much to save disk space.
3937 * The first indirect block level for a directory usually uses
3938 * HAMMER2_IND_BYTES_MIN (4KB = 32 directory entries). Due to
3939 * the hash mechanism, this typically gives us a nominal
3940 * 32 * 4 entries with one level of indirection.
3942 * We use HAMMER2_IND_BYTES_NOM (16KB = 128 blockrefs) for FILE
3943 * indirect blocks. The initial 4 entries in the inode gives us
3944 * 256KB. Up to 4 indirect blocks gives us 32MB. Three levels
3945 * of indirection gives us 137GB, and so forth. H2 can support
3946 * huge file sizes but they are not typical, so we try to stick
3947 * with compactness and do not use a larger indirect block size.
3949 * We could use 64KB (PBUFSIZE), giving us 512 blockrefs, but
3950 * due to the way indirect blocks are created this usually winds
3951 * up being extremely inefficient for small files. Even though
3952 * 16KB requires more levels of indirection for very large files,
3953 * the 16KB records can be ganged together into 64KB DIOs.
3955 if (for_type == HAMMER2_BREF_TYPE_FREEMAP_NODE ||
3956 for_type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) {
3957 nbytes = HAMMER2_FREEMAP_LEVELN_PSIZE;
3958 } else if (parent->bref.type == HAMMER2_BREF_TYPE_INODE) {
3959 if (parent->data->ipdata.meta.type ==
3960 HAMMER2_OBJTYPE_DIRECTORY)
3961 nbytes = HAMMER2_IND_BYTES_MIN; /* 4KB = 32 entries */
3963 nbytes = HAMMER2_IND_BYTES_NOM; /* 16KB = ~8MB file */
3966 nbytes = HAMMER2_IND_BYTES_NOM;
3968 if (nbytes < count * sizeof(hammer2_blockref_t)) {
3969 KKASSERT(for_type != HAMMER2_BREF_TYPE_FREEMAP_NODE &&
3970 for_type != HAMMER2_BREF_TYPE_FREEMAP_LEAF);
3971 nbytes = count * sizeof(hammer2_blockref_t);
3973 ncount = nbytes / sizeof(hammer2_blockref_t);
3976 * When creating an indirect block for a freemap node or leaf
3977 * the key/keybits must be fitted to static radix levels because
3978 * particular radix levels use particular reserved blocks in the
3981 * This routine calculates the key/radix of the indirect block
3982 * we need to create, and whether it is on the high-side or the
3986 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3987 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
3988 keybits = hammer2_chain_indkey_freemap(parent, &key, keybits,
3991 case HAMMER2_BREF_TYPE_DATA:
3992 keybits = hammer2_chain_indkey_file(parent, &key, keybits,
3993 base, count, ncount);
3995 case HAMMER2_BREF_TYPE_DIRENT:
3996 case HAMMER2_BREF_TYPE_INODE:
3997 keybits = hammer2_chain_indkey_dir(parent, &key, keybits,
3998 base, count, ncount);
4001 panic("illegal indirect block for bref type %d", for_type);
4006 * Normalize the key for the radix being represented, keeping the
4007 * high bits and throwing away the low bits.
4009 key &= ~(((hammer2_key_t)1 << keybits) - 1);
4012 * Ok, create our new indirect block
4014 if (for_type == HAMMER2_BREF_TYPE_FREEMAP_NODE ||
4015 for_type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) {
4016 dummy.bref.type = HAMMER2_BREF_TYPE_FREEMAP_NODE;
4018 dummy.bref.type = HAMMER2_BREF_TYPE_INDIRECT;
4020 dummy.bref.key = key;
4021 dummy.bref.keybits = keybits;
4022 dummy.bref.data_off = hammer2_getradix(nbytes);
4023 dummy.bref.methods =
4024 HAMMER2_ENC_CHECK(HAMMER2_DEC_CHECK(parent->bref.methods)) |
4025 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
4027 ichain = hammer2_chain_alloc(hmp, parent->pmp, &dummy.bref);
4028 atomic_set_int(&ichain->flags, HAMMER2_CHAIN_INITIAL);
4029 hammer2_chain_lock(ichain, HAMMER2_RESOLVE_MAYBE);
4030 /* ichain has one ref at this point */
4033 * We have to mark it modified to allocate its block, but use
4034 * OPTDATA to allow it to remain in the INITIAL state. Otherwise
4035 * it won't be acted upon by the flush code.
4037 *errorp = hammer2_chain_modify(ichain, mtid, 0, HAMMER2_MODIFY_OPTDATA);
4039 kprintf("hammer2_alloc_indirect: error %08x %s\n",
4040 *errorp, hammer2_error_str(*errorp));
4041 hammer2_chain_unlock(ichain);
4042 hammer2_chain_drop(ichain);
4047 * Iterate the original parent and move the matching brefs into
4048 * the new indirect block.
4050 * XXX handle flushes.
4053 key_end = HAMMER2_KEY_MAX;
4054 key_next = 0; /* avoid gcc warnings */
4055 hammer2_spin_ex(&parent->core.spin);
4061 * Parent may have been modified, relocating its block array.
4062 * Reload the base pointer.
4064 base = hammer2_chain_base_and_count(parent, &count);
4066 if (++loops > 100000) {
4067 hammer2_spin_unex(&parent->core.spin);
4068 panic("excessive loops r=%d p=%p base/count %p:%d %016jx\n",
4069 reason, parent, base, count, key_next);
4073 * NOTE: spinlock stays intact, returned chain (if not NULL)
4074 * is not referenced or locked which means that we
4075 * cannot safely check its flagged / deletion status
4078 chain = hammer2_combined_find(parent, base, count,
4082 generation = parent->core.generation;
4085 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
4088 * Skip keys that are not within the key/radix of the new
4089 * indirect block. They stay in the parent.
4091 if ((~(((hammer2_key_t)1 << keybits) - 1) &
4092 (key ^ bref->key)) != 0) {
4093 goto next_key_spinlocked;
4097 * Load the new indirect block by acquiring the related
4098 * chains (potentially from media as it might not be
4099 * in-memory). Then move it to the new parent (ichain).
4101 * chain is referenced but not locked. We must lock the
4102 * chain to obtain definitive state.
4107 * Use chain already present in the RBTREE
4109 hammer2_chain_ref(chain);
4110 hammer2_spin_unex(&parent->core.spin);
4111 hammer2_chain_lock(chain, HAMMER2_RESOLVE_NEVER);
4114 * Get chain for blockref element. _get returns NULL
4115 * on insertion race.
4117 hammer2_spin_unex(&parent->core.spin);
4118 chain = hammer2_chain_get(parent, generation, &bcopy,
4119 HAMMER2_RESOLVE_NEVER);
4120 if (chain == NULL) {
4122 hammer2_spin_ex(&parent->core.spin);
4128 * This is always live so if the chain has been deleted
4129 * we raced someone and we have to retry.
4131 * NOTE: Lookups can race delete-duplicate because
4132 * delete-duplicate does not lock the parent's core
4133 * (they just use the spinlock on the core).
4135 * (note reversed logic for this one)
4137 if (bcmp(&bcopy, &chain->bref, sizeof(bcopy)) ||
4138 chain->parent != parent ||
4139 (chain->flags & HAMMER2_CHAIN_DELETED)) {
4140 hammer2_chain_unlock(chain);
4141 hammer2_chain_drop(chain);
4142 if (hammer2_debug & 0x0040) {
4143 kprintf("LOST PARENT RETRY "
4144 "RETRY (%p,%p)->%p %08x\n",
4145 parent, chain->parent, chain, chain->flags);
4147 hammer2_spin_ex(&parent->core.spin);
4152 * Shift the chain to the indirect block.
4154 * WARNING! No reason for us to load chain data, pass NOSTATS
4155 * to prevent delete/insert from trying to access
4156 * inode stats (and thus asserting if there is no
4157 * chain->data loaded).
4159 * WARNING! The (parent, chain) deletion may modify the parent
4160 * and invalidate the base pointer.
4162 * WARNING! Parent must already be marked modified, so we
4163 * can assume that chain_delete always suceeds.
4165 * WARNING! hammer2_chain_repchange() does not have to be
4166 * called (and doesn't work anyway because we are
4167 * only doing a partial shift). A recursion that is
4168 * in-progress can continue at the current parent
4169 * and will be able to properly find its next key.
4171 error = hammer2_chain_delete(parent, chain, mtid, 0);
4172 KKASSERT(error == 0);
4173 hammer2_chain_rename(&ichain, chain, mtid, 0);
4174 hammer2_chain_unlock(chain);
4175 hammer2_chain_drop(chain);
4176 KKASSERT(parent->refs > 0);
4178 base = NULL; /* safety */
4179 hammer2_spin_ex(&parent->core.spin);
4180 next_key_spinlocked:
4181 if (--maxloops == 0)
4182 panic("hammer2_chain_create_indirect: maxloops");
4184 if (key_next == 0 || key_next > key_end)
4189 hammer2_spin_unex(&parent->core.spin);
4192 * Insert the new indirect block into the parent now that we've
4193 * cleared out some entries in the parent. We calculated a good
4194 * insertion index in the loop above (ichain->index).
4196 * We don't have to set UPDATE here because we mark ichain
4197 * modified down below (so the normal modified -> flush -> set-moved
4198 * sequence applies).
4200 * The insertion shouldn't race as this is a completely new block
4201 * and the parent is locked.
4203 base = NULL; /* safety, parent modify may change address */
4204 KKASSERT((ichain->flags & HAMMER2_CHAIN_ONRBTREE) == 0);
4205 KKASSERT(parent->core.live_count < count);
4206 hammer2_chain_insert(parent, ichain,
4207 HAMMER2_CHAIN_INSERT_SPIN |
4208 HAMMER2_CHAIN_INSERT_LIVE,
4212 * Make sure flushes propogate after our manual insertion.
4214 hammer2_chain_setflush(ichain);
4215 hammer2_chain_setflush(parent);
4218 * Figure out what to return.
4220 if (~(((hammer2_key_t)1 << keybits) - 1) &
4221 (create_key ^ key)) {
4223 * Key being created is outside the key range,
4224 * return the original parent.
4226 hammer2_chain_unlock(ichain);
4227 hammer2_chain_drop(ichain);
4230 * Otherwise its in the range, return the new parent.
4231 * (leave both the new and old parent locked).
4240 * Do maintenance on an indirect chain. Both parent and chain are locked.
4242 * Returns non-zero if (chain) is deleted, either due to being empty or
4243 * because its children were safely moved into the parent.
4246 hammer2_chain_indirect_maintenance(hammer2_chain_t *parent,
4247 hammer2_chain_t *chain)
4249 hammer2_blockref_t *chain_base;
4250 hammer2_blockref_t *base;
4251 hammer2_blockref_t *bref;
4252 hammer2_blockref_t bcopy;
4253 hammer2_key_t key_next;
4254 hammer2_key_t key_beg;
4255 hammer2_key_t key_end;
4256 hammer2_chain_t *sub;
4263 * Make sure we have an accurate live_count
4265 if ((chain->flags & (HAMMER2_CHAIN_INITIAL |
4266 HAMMER2_CHAIN_COUNTEDBREFS)) == 0) {
4267 base = &chain->data->npdata[0];
4268 count = chain->bytes / sizeof(hammer2_blockref_t);
4269 hammer2_chain_countbrefs(chain, base, count);
4273 * If the indirect block is empty we can delete it.
4274 * (ignore deletion error)
4276 if (chain->core.live_count == 0 && RB_EMPTY(&chain->core.rbtree)) {
4277 hammer2_chain_delete(parent, chain,
4278 chain->bref.modify_tid,
4279 HAMMER2_DELETE_PERMANENT);
4280 hammer2_chain_repchange(parent, chain);
4284 base = hammer2_chain_base_and_count(parent, &count);
4286 if ((parent->flags & (HAMMER2_CHAIN_INITIAL |
4287 HAMMER2_CHAIN_COUNTEDBREFS)) == 0) {
4288 hammer2_chain_countbrefs(parent, base, count);
4292 * Determine if we can collapse chain into parent, calculate
4293 * hysteresis for chain emptiness.
4295 if (parent->core.live_count + chain->core.live_count - 1 > count)
4297 chain_count = chain->bytes / sizeof(hammer2_blockref_t);
4298 if (chain->core.live_count > chain_count * 3 / 4)
4302 * Ok, theoretically we can collapse chain's contents into
4303 * parent. chain is locked, but any in-memory children of chain
4304 * are not. For this to work, we must be able to dispose of any
4305 * in-memory children of chain.
4307 * For now require that there are no in-memory children of chain.
4309 * WARNING! Both chain and parent must remain locked across this
4314 * Parent must be marked modified. Don't try to collapse it if we
4315 * can't mark it modified. Once modified, destroy chain to make room
4316 * and to get rid of what will be a conflicting key (this is included
4317 * in the calculation above). Finally, move the children of chain
4318 * into chain's parent.
4320 * This order creates an accounting problem for bref.embed.stats
4321 * because we destroy chain before we remove its children. Any
4322 * elements whos blockref is already synchronized will be counted
4323 * twice. To deal with the problem we clean out chain's stats prior
4326 error = hammer2_chain_modify(parent, 0, 0, 0);
4328 krateprintf(&krate_h2me, "hammer2: indirect_maint: %s\n",
4329 hammer2_error_str(error));
4332 error = hammer2_chain_modify(chain, chain->bref.modify_tid, 0, 0);
4334 krateprintf(&krate_h2me, "hammer2: indirect_maint: %s\n",
4335 hammer2_error_str(error));
4339 chain->bref.embed.stats.inode_count = 0;
4340 chain->bref.embed.stats.data_count = 0;
4341 error = hammer2_chain_delete(parent, chain,
4342 chain->bref.modify_tid,
4343 HAMMER2_DELETE_PERMANENT);
4344 KKASSERT(error == 0);
4347 * The combined_find call requires core.spin to be held. One would
4348 * think there wouldn't be any conflicts since we hold chain
4349 * exclusively locked, but the caching mechanism for 0-ref children
4350 * does not require a chain lock.
4352 hammer2_spin_ex(&chain->core.spin);
4356 key_end = HAMMER2_KEY_MAX;
4358 chain_base = &chain->data->npdata[0];
4359 chain_count = chain->bytes / sizeof(hammer2_blockref_t);
4360 sub = hammer2_combined_find(chain, chain_base, chain_count,
4364 generation = chain->core.generation;
4367 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
4371 hammer2_chain_ref(sub);
4372 hammer2_spin_unex(&chain->core.spin);
4373 hammer2_chain_lock(sub, HAMMER2_RESOLVE_NEVER);
4375 hammer2_spin_unex(&chain->core.spin);
4376 sub = hammer2_chain_get(chain, generation, &bcopy,
4377 HAMMER2_RESOLVE_NEVER);
4379 hammer2_spin_ex(&chain->core.spin);
4383 if (bcmp(&bcopy, &sub->bref, sizeof(bcopy)) ||
4384 sub->parent != chain ||
4385 (sub->flags & HAMMER2_CHAIN_DELETED)) {
4386 hammer2_chain_unlock(sub);
4387 hammer2_chain_drop(sub);
4388 hammer2_spin_ex(&chain->core.spin);
4389 sub = NULL; /* safety */
4392 error = hammer2_chain_delete(chain, sub,
4393 sub->bref.modify_tid, 0);
4394 KKASSERT(error == 0);
4395 hammer2_chain_rename(&parent, sub,
4396 sub->bref.modify_tid,
4397 HAMMER2_INSERT_SAMEPARENT);
4398 hammer2_chain_unlock(sub);
4399 hammer2_chain_drop(sub);
4400 hammer2_spin_ex(&chain->core.spin);
4406 hammer2_spin_unex(&chain->core.spin);
4408 hammer2_chain_repchange(parent, chain);
4414 * Freemap indirect blocks
4416 * Calculate the keybits and highside/lowside of the freemap node the
4417 * caller is creating.
4419 * This routine will specify the next higher-level freemap key/radix
4420 * representing the lowest-ordered set. By doing so, eventually all
4421 * low-ordered sets will be moved one level down.
4423 * We have to be careful here because the freemap reserves a limited
4424 * number of blocks for a limited number of levels. So we can't just
4425 * push indiscriminately.
4428 hammer2_chain_indkey_freemap(hammer2_chain_t *parent, hammer2_key_t *keyp,
4429 int keybits, hammer2_blockref_t *base, int count)
4431 hammer2_chain_t *chain;
4432 hammer2_blockref_t *bref;
4434 hammer2_key_t key_beg;
4435 hammer2_key_t key_end;
4436 hammer2_key_t key_next;
4439 int maxloops = 300000;
4447 * Calculate the range of keys in the array being careful to skip
4448 * slots which are overridden with a deletion.
4451 key_end = HAMMER2_KEY_MAX;
4452 hammer2_spin_ex(&parent->core.spin);
4455 if (--maxloops == 0) {
4456 panic("indkey_freemap shit %p %p:%d\n",
4457 parent, base, count);
4459 chain = hammer2_combined_find(parent, base, count,
4471 * Skip deleted chains.
4473 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) {
4474 if (key_next == 0 || key_next > key_end)
4481 * Use the full live (not deleted) element for the scan
4482 * iteration. HAMMER2 does not allow partial replacements.
4484 * XXX should be built into hammer2_combined_find().
4486 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
4488 if (keybits > bref->keybits) {
4490 keybits = bref->keybits;
4491 } else if (keybits == bref->keybits && bref->key < key) {
4498 hammer2_spin_unex(&parent->core.spin);
4501 * Return the keybits for a higher-level FREEMAP_NODE covering
4505 case HAMMER2_FREEMAP_LEVEL0_RADIX:
4506 keybits = HAMMER2_FREEMAP_LEVEL1_RADIX;
4508 case HAMMER2_FREEMAP_LEVEL1_RADIX:
4509 keybits = HAMMER2_FREEMAP_LEVEL2_RADIX;
4511 case HAMMER2_FREEMAP_LEVEL2_RADIX:
4512 keybits = HAMMER2_FREEMAP_LEVEL3_RADIX;
4514 case HAMMER2_FREEMAP_LEVEL3_RADIX:
4515 keybits = HAMMER2_FREEMAP_LEVEL4_RADIX;
4517 case HAMMER2_FREEMAP_LEVEL4_RADIX:
4518 keybits = HAMMER2_FREEMAP_LEVEL5_RADIX;
4520 case HAMMER2_FREEMAP_LEVEL5_RADIX:
4521 panic("hammer2_chain_indkey_freemap: level too high");
4524 panic("hammer2_chain_indkey_freemap: bad radix");
4533 * File indirect blocks
4535 * Calculate the key/keybits for the indirect block to create by scanning
4536 * existing keys. The key being created is also passed in *keyp and can be
4537 * inside or outside the indirect block. Regardless, the indirect block
4538 * must hold at least two keys in order to guarantee sufficient space.
4540 * We use a modified version of the freemap's fixed radix tree, but taylored
4541 * for file data. Basically we configure an indirect block encompassing the
4545 hammer2_chain_indkey_file(hammer2_chain_t *parent, hammer2_key_t *keyp,
4546 int keybits, hammer2_blockref_t *base, int count,
4549 hammer2_chain_t *chain;
4550 hammer2_blockref_t *bref;
4552 hammer2_key_t key_beg;
4553 hammer2_key_t key_end;
4554 hammer2_key_t key_next;
4558 int maxloops = 300000;
4566 * Calculate the range of keys in the array being careful to skip
4567 * slots which are overridden with a deletion.
4569 * Locate the smallest key.
4572 key_end = HAMMER2_KEY_MAX;
4573 hammer2_spin_ex(&parent->core.spin);
4576 if (--maxloops == 0) {
4577 panic("indkey_freemap shit %p %p:%d\n",
4578 parent, base, count);
4580 chain = hammer2_combined_find(parent, base, count,
4592 * Skip deleted chains.
4594 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) {
4595 if (key_next == 0 || key_next > key_end)
4602 * Use the full live (not deleted) element for the scan
4603 * iteration. HAMMER2 does not allow partial replacements.
4605 * XXX should be built into hammer2_combined_find().
4607 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
4609 if (keybits > bref->keybits) {
4611 keybits = bref->keybits;
4612 } else if (keybits == bref->keybits && bref->key < key) {
4619 hammer2_spin_unex(&parent->core.spin);
4622 * Calculate the static keybits for a higher-level indirect block
4623 * that contains the key.
4628 case HAMMER2_IND_BYTES_MIN / sizeof(hammer2_blockref_t):
4629 nradix = HAMMER2_IND_RADIX_MIN - HAMMER2_BLOCKREF_RADIX;
4631 case HAMMER2_IND_BYTES_NOM / sizeof(hammer2_blockref_t):
4632 nradix = HAMMER2_IND_RADIX_NOM - HAMMER2_BLOCKREF_RADIX;
4634 case HAMMER2_IND_BYTES_MAX / sizeof(hammer2_blockref_t):
4635 nradix = HAMMER2_IND_RADIX_MAX - HAMMER2_BLOCKREF_RADIX;
4638 panic("bad ncount %d\n", ncount);
4644 * The largest radix that can be returned for an indirect block is
4645 * 63 bits. (The largest practical indirect block radix is actually
4646 * 62 bits because the top-level inode or volume root contains four
4647 * entries, but allow 63 to be returned).
4652 return keybits + nradix;
4658 * Directory indirect blocks.
4660 * Covers both the inode index (directory of inodes), and directory contents
4661 * (filenames hardlinked to inodes).
4663 * Because directory keys are hashed we generally try to cut the space in
4664 * half. We accomodate the inode index (which tends to have linearly
4665 * increasing inode numbers) by ensuring that the keyspace is at least large
4666 * enough to fill up the indirect block being created.
4669 hammer2_chain_indkey_dir(hammer2_chain_t *parent, hammer2_key_t *keyp,
4670 int keybits, hammer2_blockref_t *base, int count,
4673 hammer2_blockref_t *bref;
4674 hammer2_chain_t *chain;
4675 hammer2_key_t key_beg;
4676 hammer2_key_t key_end;
4677 hammer2_key_t key_next;
4682 int maxloops = 300000;
4685 * NOTE: We can't take a shortcut here anymore for inodes because
4686 * the root directory can contain a mix of inodes and directory
4687 * entries (we used to just return 63 if parent->bref.type was
4688 * HAMMER2_BREF_TYPE_INODE.
4695 * Calculate the range of keys in the array being careful to skip
4696 * slots which are overridden with a deletion.
4699 key_end = HAMMER2_KEY_MAX;
4700 hammer2_spin_ex(&parent->core.spin);
4703 if (--maxloops == 0) {
4704 panic("indkey_freemap shit %p %p:%d\n",
4705 parent, base, count);
4707 chain = hammer2_combined_find(parent, base, count,
4721 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) {
4722 if (key_next == 0 || key_next > key_end)
4729 * Use the full live (not deleted) element for the scan
4730 * iteration. HAMMER2 does not allow partial replacements.
4732 * XXX should be built into hammer2_combined_find().
4734 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
4737 * Expand our calculated key range (key, keybits) to fit
4738 * the scanned key. nkeybits represents the full range
4739 * that we will later cut in half (two halves @ nkeybits - 1).
4742 if (nkeybits < bref->keybits) {
4743 if (bref->keybits > 64) {
4744 kprintf("bad bref chain %p bref %p\n",
4748 nkeybits = bref->keybits;
4750 while (nkeybits < 64 &&
4751 (~(((hammer2_key_t)1 << nkeybits) - 1) &
4752 (key ^ bref->key)) != 0) {
4757 * If the new key range is larger we have to determine
4758 * which side of the new key range the existing keys fall
4759 * under by checking the high bit, then collapsing the
4760 * locount into the hicount or vise-versa.
4762 if (keybits != nkeybits) {
4763 if (((hammer2_key_t)1 << (nkeybits - 1)) & key) {
4774 * The newly scanned key will be in the lower half or the
4775 * upper half of the (new) key range.
4777 if (((hammer2_key_t)1 << (nkeybits - 1)) & bref->key)
4786 hammer2_spin_unex(&parent->core.spin);
4787 bref = NULL; /* now invalid (safety) */
4790 * Adjust keybits to represent half of the full range calculated
4791 * above (radix 63 max) for our new indirect block.
4796 * Expand keybits to hold at least ncount elements. ncount will be
4797 * a power of 2. This is to try to completely fill leaf nodes (at
4798 * least for keys which are not hashes).
4800 * We aren't counting 'in' or 'out', we are counting 'high side'
4801 * and 'low side' based on the bit at (1LL << keybits). We want
4802 * everything to be inside in these cases so shift it all to
4803 * the low or high side depending on the new high bit.
4805 while (((hammer2_key_t)1 << keybits) < ncount) {
4807 if (key & ((hammer2_key_t)1 << keybits)) {
4816 if (hicount > locount)
4817 key |= (hammer2_key_t)1 << keybits;
4819 key &= ~(hammer2_key_t)1 << keybits;
4829 * Directory indirect blocks.
4831 * Covers both the inode index (directory of inodes), and directory contents
4832 * (filenames hardlinked to inodes).
4834 * Because directory keys are hashed we generally try to cut the space in
4835 * half. We accomodate the inode index (which tends to have linearly
4836 * increasing inode numbers) by ensuring that the keyspace is at least large
4837 * enough to fill up the indirect block being created.
4840 hammer2_chain_indkey_dir(hammer2_chain_t *parent, hammer2_key_t *keyp,
4841 int keybits, hammer2_blockref_t *base, int count,
4844 hammer2_blockref_t *bref;
4845 hammer2_chain_t *chain;
4846 hammer2_key_t key_beg;
4847 hammer2_key_t key_end;
4848 hammer2_key_t key_next;
4853 int maxloops = 300000;
4856 * Shortcut if the parent is the inode. In this situation the
4857 * parent has 4+1 directory entries and we are creating an indirect
4858 * block capable of holding many more.
4860 if (parent->bref.type == HAMMER2_BREF_TYPE_INODE) {
4869 * Calculate the range of keys in the array being careful to skip
4870 * slots which are overridden with a deletion.
4873 key_end = HAMMER2_KEY_MAX;
4874 hammer2_spin_ex(&parent->core.spin);
4877 if (--maxloops == 0) {
4878 panic("indkey_freemap shit %p %p:%d\n",
4879 parent, base, count);
4881 chain = hammer2_combined_find(parent, base, count,
4895 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) {
4896 if (key_next == 0 || key_next > key_end)
4903 * Use the full live (not deleted) element for the scan
4904 * iteration. HAMMER2 does not allow partial replacements.
4906 * XXX should be built into hammer2_combined_find().
4908 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
4911 * Expand our calculated key range (key, keybits) to fit
4912 * the scanned key. nkeybits represents the full range
4913 * that we will later cut in half (two halves @ nkeybits - 1).
4916 if (nkeybits < bref->keybits) {
4917 if (bref->keybits > 64) {
4918 kprintf("bad bref chain %p bref %p\n",
4922 nkeybits = bref->keybits;
4924 while (nkeybits < 64 &&
4925 (~(((hammer2_key_t)1 << nkeybits) - 1) &
4926 (key ^ bref->key)) != 0) {
4931 * If the new key range is larger we have to determine
4932 * which side of the new key range the existing keys fall
4933 * under by checking the high bit, then collapsing the
4934 * locount into the hicount or vise-versa.
4936 if (keybits != nkeybits) {
4937 if (((hammer2_key_t)1 << (nkeybits - 1)) & key) {
4948 * The newly scanned key will be in the lower half or the
4949 * upper half of the (new) key range.
4951 if (((hammer2_key_t)1 << (nkeybits - 1)) & bref->key)
4960 hammer2_spin_unex(&parent->core.spin);
4961 bref = NULL; /* now invalid (safety) */
4964 * Adjust keybits to represent half of the full range calculated
4965 * above (radix 63 max) for our new indirect block.
4970 * Expand keybits to hold at least ncount elements. ncount will be
4971 * a power of 2. This is to try to completely fill leaf nodes (at
4972 * least for keys which are not hashes).
4974 * We aren't counting 'in' or 'out', we are counting 'high side'
4975 * and 'low side' based on the bit at (1LL << keybits). We want
4976 * everything to be inside in these cases so shift it all to
4977 * the low or high side depending on the new high bit.
4979 while (((hammer2_key_t)1 << keybits) < ncount) {
4981 if (key & ((hammer2_key_t)1 << keybits)) {
4990 if (hicount > locount)
4991 key |= (hammer2_key_t)1 << keybits;
4993 key &= ~(hammer2_key_t)1 << keybits;
5003 * Sets CHAIN_DELETED and remove the chain's blockref from the parent if
5006 * Both parent and chain must be locked exclusively.
5008 * This function will modify the parent if the blockref requires removal
5009 * from the parent's block table.
5011 * This function is NOT recursive. Any entity already pushed into the
5012 * chain (such as an inode) may still need visibility into its contents,
5013 * as well as the ability to read and modify the contents. For example,
5014 * for an unlinked file which is still open.
5016 * Also note that the flusher is responsible for cleaning up empty
5020 hammer2_chain_delete(hammer2_chain_t *parent, hammer2_chain_t *chain,
5021 hammer2_tid_t mtid, int flags)
5025 KKASSERT(hammer2_mtx_owned(&chain->lock));
5028 * Nothing to do if already marked.
5030 * We need the spinlock on the core whos RBTREE contains chain
5031 * to protect against races.
5033 if ((chain->flags & HAMMER2_CHAIN_DELETED) == 0) {
5034 KKASSERT((chain->flags & HAMMER2_CHAIN_DELETED) == 0 &&
5035 chain->parent == parent);
5036 error = _hammer2_chain_delete_helper(parent, chain,
5041 * Permanent deletions mark the chain as destroyed.
5043 * NOTE: We do not setflush the chain unless the deletion is
5044 * permanent, since the deletion of a chain does not actually
5045 * require it to be flushed.
5048 if (flags & HAMMER2_DELETE_PERMANENT) {
5049 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DESTROY);
5050 hammer2_chain_setflush(chain);
5058 * Returns the index of the nearest element in the blockref array >= elm.
5059 * Returns (count) if no element could be found.
5061 * Sets *key_nextp to the next key for loop purposes but does not modify
5062 * it if the next key would be higher than the current value of *key_nextp.
5063 * Note that *key_nexp can overflow to 0, which should be tested by the
5066 * WARNING! Must be called with parent's spinlock held. Spinlock remains
5067 * held through the operation.
5070 hammer2_base_find(hammer2_chain_t *parent,
5071 hammer2_blockref_t *base, int count,
5072 hammer2_key_t *key_nextp,
5073 hammer2_key_t key_beg, hammer2_key_t key_end)
5075 hammer2_blockref_t *scan;
5076 hammer2_key_t scan_end;
5081 * Require the live chain's already have their core's counted
5082 * so we can optimize operations.
5084 KKASSERT(parent->flags & HAMMER2_CHAIN_COUNTEDBREFS);
5089 if (count == 0 || base == NULL)
5093 * Sequential optimization using parent->cache_index. This is
5094 * the most likely scenario.
5096 * We can avoid trailing empty entries on live chains, otherwise
5097 * we might have to check the whole block array.
5099 i = parent->cache_index; /* SMP RACE OK */
5101 limit = parent->core.live_zero;
5106 KKASSERT(i < count);
5112 while (i > 0 && (scan->type == 0 || scan->key > key_beg)) {
5116 parent->cache_index = i;
5119 * Search forwards, stop when we find a scan element which
5120 * encloses the key or until we know that there are no further
5124 if (scan->type != 0) {
5125 scan_end = scan->key +
5126 ((hammer2_key_t)1 << scan->keybits) - 1;
5127 if (scan->key > key_beg || scan_end >= key_beg)
5136 parent->cache_index = i;
5140 scan_end = scan->key +
5141 ((hammer2_key_t)1 << scan->keybits);
5142 if (scan_end && (*key_nextp > scan_end ||
5144 *key_nextp = scan_end;
5152 * Do a combined search and return the next match either from the blockref
5153 * array or from the in-memory chain. Sets *bresp to the returned bref in
5154 * both cases, or sets it to NULL if the search exhausted. Only returns
5155 * a non-NULL chain if the search matched from the in-memory chain.
5157 * When no in-memory chain has been found and a non-NULL bref is returned
5161 * The returned chain is not locked or referenced. Use the returned bref
5162 * to determine if the search exhausted or not. Iterate if the base find
5163 * is chosen but matches a deleted chain.
5165 * WARNING! Must be called with parent's spinlock held. Spinlock remains
5166 * held through the operation.
5169 hammer2_combined_find(hammer2_chain_t *parent,
5170 hammer2_blockref_t *base, int count,
5171 hammer2_key_t *key_nextp,
5172 hammer2_key_t key_beg, hammer2_key_t key_end,
5173 hammer2_blockref_t **bresp)
5175 hammer2_blockref_t *bref;
5176 hammer2_chain_t *chain;
5180 * Lookup in block array and in rbtree.
5182 *key_nextp = key_end + 1;
5183 i = hammer2_base_find(parent, base, count, key_nextp,
5185 chain = hammer2_chain_find(parent, key_nextp, key_beg, key_end);
5190 if (i == count && chain == NULL) {
5196 * Only chain matched.
5199 bref = &chain->bref;
5204 * Only blockref matched.
5206 if (chain == NULL) {
5212 * Both in-memory and blockref matched, select the nearer element.
5214 * If both are flush with the left-hand side or both are the
5215 * same distance away, select the chain. In this situation the
5216 * chain must have been loaded from the matching blockmap.
5218 if ((chain->bref.key <= key_beg && base[i].key <= key_beg) ||
5219 chain->bref.key == base[i].key) {
5220 KKASSERT(chain->bref.key == base[i].key);
5221 bref = &chain->bref;
5226 * Select the nearer key
5228 if (chain->bref.key < base[i].key) {
5229 bref = &chain->bref;
5236 * If the bref is out of bounds we've exhausted our search.
5239 if (bref->key > key_end) {
5249 * Locate the specified block array element and delete it. The element
5252 * The spin lock on the related chain must be held.
5254 * NOTE: live_count was adjusted when the chain was deleted, so it does not
5255 * need to be adjusted when we commit the media change.
5258 hammer2_base_delete(hammer2_chain_t *parent,
5259 hammer2_blockref_t *base, int count,
5260 hammer2_chain_t *chain)
5262 hammer2_blockref_t *elm = &chain->bref;
5263 hammer2_blockref_t *scan;
5264 hammer2_key_t key_next;
5268 * Delete element. Expect the element to exist.
5270 * XXX see caller, flush code not yet sophisticated enough to prevent
5271 * re-flushed in some cases.
5273 key_next = 0; /* max range */
5274 i = hammer2_base_find(parent, base, count, &key_next,
5275 elm->key, elm->key);
5277 if (i == count || scan->type == 0 ||
5278 scan->key != elm->key ||
5279 ((chain->flags & HAMMER2_CHAIN_BMAPUPD) == 0 &&
5280 scan->keybits != elm->keybits)) {
5281 hammer2_spin_unex(&parent->core.spin);
5282 panic("delete base %p element not found at %d/%d elm %p\n",
5283 base, i, count, elm);
5288 * Update stats and zero the entry.
5290 * NOTE: Handle radix == 0 (0 bytes) case.
5292 if ((int)(scan->data_off & HAMMER2_OFF_MASK_RADIX)) {
5293 parent->bref.embed.stats.data_count -= (hammer2_off_t)1 <<
5294 (int)(scan->data_off & HAMMER2_OFF_MASK_RADIX);
5296 switch(scan->type) {
5297 case HAMMER2_BREF_TYPE_INODE:
5298 --parent->bref.embed.stats.inode_count;
5300 case HAMMER2_BREF_TYPE_DATA:
5301 if (parent->bref.leaf_count == HAMMER2_BLOCKREF_LEAF_MAX) {
5302 atomic_set_int(&chain->flags,
5303 HAMMER2_CHAIN_HINT_LEAF_COUNT);
5305 if (parent->bref.leaf_count)
5306 --parent->bref.leaf_count;
5309 case HAMMER2_BREF_TYPE_INDIRECT:
5310 if (scan->type != HAMMER2_BREF_TYPE_DATA) {
5311 parent->bref.embed.stats.data_count -=
5312 scan->embed.stats.data_count;
5313 parent->bref.embed.stats.inode_count -=
5314 scan->embed.stats.inode_count;
5316 if (scan->type == HAMMER2_BREF_TYPE_INODE)
5318 if (parent->bref.leaf_count == HAMMER2_BLOCKREF_LEAF_MAX) {
5319 atomic_set_int(&chain->flags,
5320 HAMMER2_CHAIN_HINT_LEAF_COUNT);
5322 if (parent->bref.leaf_count <= scan->leaf_count)
5323 parent->bref.leaf_count = 0;
5325 parent->bref.leaf_count -= scan->leaf_count;
5328 case HAMMER2_BREF_TYPE_DIRENT:
5329 if (parent->bref.leaf_count == HAMMER2_BLOCKREF_LEAF_MAX) {
5330 atomic_set_int(&chain->flags,
5331 HAMMER2_CHAIN_HINT_LEAF_COUNT);
5333 if (parent->bref.leaf_count)
5334 --parent->bref.leaf_count;
5340 bzero(scan, sizeof(*scan));
5343 * We can only optimize parent->core.live_zero for live chains.
5345 if (parent->core.live_zero == i + 1) {
5346 while (--i >= 0 && base[i].type == 0)
5348 parent->core.live_zero = i + 1;
5352 * Clear appropriate blockmap flags in chain.
5354 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_BMAPPED |
5355 HAMMER2_CHAIN_BMAPUPD);
5359 * Insert the specified element. The block array must not already have the
5360 * element and must have space available for the insertion.
5362 * The spin lock on the related chain must be held.
5364 * NOTE: live_count was adjusted when the chain was deleted, so it does not
5365 * need to be adjusted when we commit the media change.
5368 hammer2_base_insert(hammer2_chain_t *parent,
5369 hammer2_blockref_t *base, int count,
5370 hammer2_chain_t *chain, hammer2_blockref_t *elm)
5372 hammer2_key_t key_next;
5381 * Insert new element. Expect the element to not already exist
5382 * unless we are replacing it.
5384 * XXX see caller, flush code not yet sophisticated enough to prevent
5385 * re-flushed in some cases.
5387 key_next = 0; /* max range */
5388 i = hammer2_base_find(parent, base, count, &key_next,
5389 elm->key, elm->key);
5392 * Shortcut fill optimization, typical ordered insertion(s) may not
5395 KKASSERT(i >= 0 && i <= count);
5398 * Set appropriate blockmap flags in chain (if not NULL)
5401 atomic_set_int(&chain->flags, HAMMER2_CHAIN_BMAPPED);
5404 * Update stats and zero the entry
5406 if ((int)(elm->data_off & HAMMER2_OFF_MASK_RADIX)) {
5407 parent->bref.embed.stats.data_count += (hammer2_off_t)1 <<
5408 (int)(elm->data_off & HAMMER2_OFF_MASK_RADIX);
5411 case HAMMER2_BREF_TYPE_INODE:
5412 ++parent->bref.embed.stats.inode_count;
5414 case HAMMER2_BREF_TYPE_DATA:
5415 if (parent->bref.leaf_count != HAMMER2_BLOCKREF_LEAF_MAX)
5416 ++parent->bref.leaf_count;
5418 case HAMMER2_BREF_TYPE_INDIRECT:
5419 if (elm->type != HAMMER2_BREF_TYPE_DATA) {
5420 parent->bref.embed.stats.data_count +=
5421 elm->embed.stats.data_count;
5422 parent->bref.embed.stats.inode_count +=
5423 elm->embed.stats.inode_count;
5425 if (elm->type == HAMMER2_BREF_TYPE_INODE)
5427 if (parent->bref.leaf_count + elm->leaf_count <
5428 HAMMER2_BLOCKREF_LEAF_MAX) {
5429 parent->bref.leaf_count += elm->leaf_count;
5431 parent->bref.leaf_count = HAMMER2_BLOCKREF_LEAF_MAX;
5434 case HAMMER2_BREF_TYPE_DIRENT:
5435 if (parent->bref.leaf_count != HAMMER2_BLOCKREF_LEAF_MAX)
5436 ++parent->bref.leaf_count;
5444 * We can only optimize parent->core.live_zero for live chains.
5446 if (i == count && parent->core.live_zero < count) {
5447 i = parent->core.live_zero++;
5452 xkey = elm->key + ((hammer2_key_t)1 << elm->keybits) - 1;
5453 if (i != count && (base[i].key < elm->key || xkey >= base[i].key)) {
5454 hammer2_spin_unex(&parent->core.spin);
5455 panic("insert base %p overlapping elements at %d elm %p\n",
5460 * Try to find an empty slot before or after.
5464 while (j > 0 || k < count) {
5466 if (j >= 0 && base[j].type == 0) {
5470 bcopy(&base[j+1], &base[j],
5471 (i - j - 1) * sizeof(*base));
5477 if (k < count && base[k].type == 0) {
5478 bcopy(&base[i], &base[i+1],
5479 (k - i) * sizeof(hammer2_blockref_t));
5483 * We can only update parent->core.live_zero for live
5486 if (parent->core.live_zero <= k)
5487 parent->core.live_zero = k + 1;
5492 panic("hammer2_base_insert: no room!");
5499 for (l = 0; l < count; ++l) {
5501 key_next = base[l].key +
5502 ((hammer2_key_t)1 << base[l].keybits) - 1;
5506 while (++l < count) {
5508 if (base[l].key <= key_next)
5509 panic("base_insert %d %d,%d,%d fail %p:%d", u, i, j, k, base, l);
5510 key_next = base[l].key +
5511 ((hammer2_key_t)1 << base[l].keybits) - 1;
5521 * Sort the blockref array for the chain. Used by the flush code to
5522 * sort the blockref[] array.
5524 * The chain must be exclusively locked AND spin-locked.
5526 typedef hammer2_blockref_t *hammer2_blockref_p;
5530 hammer2_base_sort_callback(const void *v1, const void *v2)
5532 hammer2_blockref_p bref1 = *(const hammer2_blockref_p *)v1;
5533 hammer2_blockref_p bref2 = *(const hammer2_blockref_p *)v2;
5536 * Make sure empty elements are placed at the end of the array
5538 if (bref1->type == 0) {
5539 if (bref2->type == 0)
5542 } else if (bref2->type == 0) {
5549 if (bref1->key < bref2->key)
5551 if (bref1->key > bref2->key)
5557 hammer2_base_sort(hammer2_chain_t *chain)
5559 hammer2_blockref_t *base;
5562 switch(chain->bref.type) {
5563 case HAMMER2_BREF_TYPE_INODE:
5565 * Special shortcut for embedded data returns the inode
5566 * itself. Callers must detect this condition and access
5567 * the embedded data (the strategy code does this for us).
5569 * This is only applicable to regular files and softlinks.
5571 if (chain->data->ipdata.meta.op_flags &
5572 HAMMER2_OPFLAG_DIRECTDATA) {
5575 base = &chain->data->ipdata.u.blockset.blockref[0];
5576 count = HAMMER2_SET_COUNT;
5578 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
5579 case HAMMER2_BREF_TYPE_INDIRECT:
5581 * Optimize indirect blocks in the INITIAL state to avoid
5584 KKASSERT((chain->flags & HAMMER2_CHAIN_INITIAL) == 0);
5585 base = &chain->data->npdata[0];
5586 count = chain->bytes / sizeof(hammer2_blockref_t);
5588 case HAMMER2_BREF_TYPE_VOLUME:
5589 base = &chain->data->voldata.sroot_blockset.blockref[0];
5590 count = HAMMER2_SET_COUNT;
5592 case HAMMER2_BREF_TYPE_FREEMAP:
5593 base = &chain->data->blkset.blockref[0];
5594 count = HAMMER2_SET_COUNT;
5597 kprintf("hammer2_chain_lookup: unrecognized "
5598 "blockref(A) type: %d",
5601 tsleep(&base, 0, "dead", 0);
5602 panic("hammer2_chain_lookup: unrecognized "
5603 "blockref(A) type: %d",
5605 base = NULL; /* safety */
5606 count = 0; /* safety */
5608 kqsort(base, count, sizeof(*base), hammer2_base_sort_callback);
5614 * Chain memory management
5617 hammer2_chain_wait(hammer2_chain_t *chain)
5619 tsleep(chain, 0, "chnflw", 1);
5622 const hammer2_media_data_t *
5623 hammer2_chain_rdata(hammer2_chain_t *chain)
5625 KKASSERT(chain->data != NULL);
5626 return (chain->data);
5629 hammer2_media_data_t *
5630 hammer2_chain_wdata(hammer2_chain_t *chain)
5632 KKASSERT(chain->data != NULL);
5633 return (chain->data);
5637 * Set the check data for a chain. This can be a heavy-weight operation
5638 * and typically only runs on-flush. For file data check data is calculated
5639 * when the logical buffers are flushed.
5642 hammer2_chain_setcheck(hammer2_chain_t *chain, void *bdata)
5644 chain->bref.flags &= ~HAMMER2_BREF_FLAG_ZERO;
5646 switch(HAMMER2_DEC_CHECK(chain->bref.methods)) {
5647 case HAMMER2_CHECK_NONE:
5649 case HAMMER2_CHECK_DISABLED:
5651 case HAMMER2_CHECK_ISCSI32:
5652 chain->bref.check.iscsi32.value =
5653 hammer2_icrc32(bdata, chain->bytes);
5655 case HAMMER2_CHECK_XXHASH64:
5656 chain->bref.check.xxhash64.value =
5657 XXH64(bdata, chain->bytes, XXH_HAMMER2_SEED);
5659 case HAMMER2_CHECK_SHA192:
5661 SHA256_CTX hash_ctx;
5663 uint8_t digest[SHA256_DIGEST_LENGTH];
5664 uint64_t digest64[SHA256_DIGEST_LENGTH/8];
5667 SHA256_Init(&hash_ctx);
5668 SHA256_Update(&hash_ctx, bdata, chain->bytes);
5669 SHA256_Final(u.digest, &hash_ctx);
5670 u.digest64[2] ^= u.digest64[3];
5672 chain->bref.check.sha192.data,
5673 sizeof(chain->bref.check.sha192.data));
5676 case HAMMER2_CHECK_FREEMAP:
5677 chain->bref.check.freemap.icrc32 =
5678 hammer2_icrc32(bdata, chain->bytes);
5681 kprintf("hammer2_chain_setcheck: unknown check type %02x\n",
5682 chain->bref.methods);
5688 hammer2_chain_testcheck(hammer2_chain_t *chain, void *bdata)
5694 if (chain->bref.flags & HAMMER2_BREF_FLAG_ZERO)
5697 switch(HAMMER2_DEC_CHECK(chain->bref.methods)) {
5698 case HAMMER2_CHECK_NONE:
5701 case HAMMER2_CHECK_DISABLED:
5704 case HAMMER2_CHECK_ISCSI32:
5705 check32 = hammer2_icrc32(bdata, chain->bytes);
5706 r = (chain->bref.check.iscsi32.value == check32);
5708 kprintf("chain %016jx.%02x meth=%02x CHECK FAIL "
5709 "(flags=%08x, bref/data %08x/%08x)\n",
5710 chain->bref.data_off,
5712 chain->bref.methods,
5714 chain->bref.check.iscsi32.value,
5717 hammer2_process_icrc32 += chain->bytes;
5719 case HAMMER2_CHECK_XXHASH64:
5720 check64 = XXH64(bdata, chain->bytes, XXH_HAMMER2_SEED);
5721 r = (chain->bref.check.xxhash64.value == check64);
5723 kprintf("chain %016jx.%02x key=%016jx "
5724 "meth=%02x CHECK FAIL "
5725 "(flags=%08x, bref/data %016jx/%016jx)\n",
5726 chain->bref.data_off,
5729 chain->bref.methods,
5731 chain->bref.check.xxhash64.value,
5734 hammer2_process_xxhash64 += chain->bytes;
5736 case HAMMER2_CHECK_SHA192:
5738 SHA256_CTX hash_ctx;
5740 uint8_t digest[SHA256_DIGEST_LENGTH];
5741 uint64_t digest64[SHA256_DIGEST_LENGTH/8];
5744 SHA256_Init(&hash_ctx);
5745 SHA256_Update(&hash_ctx, bdata, chain->bytes);
5746 SHA256_Final(u.digest, &hash_ctx);
5747 u.digest64[2] ^= u.digest64[3];
5749 chain->bref.check.sha192.data,
5750 sizeof(chain->bref.check.sha192.data)) == 0) {
5754 kprintf("chain %016jx.%02x meth=%02x "
5756 chain->bref.data_off,
5758 chain->bref.methods);
5762 case HAMMER2_CHECK_FREEMAP:
5763 r = (chain->bref.check.freemap.icrc32 ==
5764 hammer2_icrc32(bdata, chain->bytes));
5766 kprintf("chain %016jx.%02x meth=%02x "
5768 chain->bref.data_off,
5770 chain->bref.methods);
5771 kprintf("freemap.icrc %08x icrc32 %08x (%d)\n",
5772 chain->bref.check.freemap.icrc32,
5773 hammer2_icrc32(bdata, chain->bytes),
5776 kprintf("dio %p buf %016jx,%d bdata %p/%p\n",
5777 chain->dio, chain->dio->bp->b_loffset,
5778 chain->dio->bp->b_bufsize, bdata,
5779 chain->dio->bp->b_data);
5784 kprintf("hammer2_chain_setcheck: unknown check type %02x\n",
5785 chain->bref.methods);
5793 * Acquire the chain and parent representing the specified inode for the
5794 * device at the specified cluster index.
5796 * The flags passed in are LOOKUP flags, not RESOLVE flags.
5798 * If we are unable to locate the hardlink, INVAL is returned and *chainp
5799 * will be NULL. *parentp may still be set error or not, or NULL if the
5800 * parent itself could not be resolved.
5802 * Caller must pass-in a valid (and locked), or NULL *parentp or *chainp.
5803 * This function replaces *parentp and *chainp. Generally speaking, if
5804 * the caller found a directory entry and wants the inode, the caller should
5805 * pass the parent,chain representing the directory entry so this function
5806 * can dispose of it properly to avoid any possible lock order reversals.
5809 hammer2_chain_inode_find(hammer2_pfs_t *pmp, hammer2_key_t inum,
5810 int clindex, int flags,
5811 hammer2_chain_t **parentp, hammer2_chain_t **chainp)
5813 hammer2_chain_t *parent;
5814 hammer2_chain_t *rchain;
5815 hammer2_key_t key_dummy;
5819 resolve_flags = (flags & HAMMER2_LOOKUP_SHARED) ?
5820 HAMMER2_RESOLVE_SHARED : 0;
5823 * Caller expects us to replace these.
5826 hammer2_chain_unlock(*chainp);
5827 hammer2_chain_drop(*chainp);
5831 hammer2_chain_unlock(*parentp);
5832 hammer2_chain_drop(*parentp);
5837 * Inodes hang off of the iroot (bit 63 is clear, differentiating
5838 * inodes from root directory entries in the key lookup).
5840 parent = hammer2_inode_chain(pmp->iroot, clindex, resolve_flags);
5843 rchain = hammer2_chain_lookup(&parent, &key_dummy,
5847 error = HAMMER2_ERROR_EIO;
5856 * Used by the bulkscan code to snapshot the synchronized storage for
5857 * a volume, allowing it to be scanned concurrently against normal
5861 hammer2_chain_bulksnap(hammer2_dev_t *hmp)
5863 hammer2_chain_t *copy;
5865 copy = hammer2_chain_alloc(hmp, hmp->spmp, &hmp->vchain.bref);
5866 copy->data = kmalloc(sizeof(copy->data->voldata),
5869 hammer2_voldata_lock(hmp);
5870 copy->data->voldata = hmp->volsync;
5871 hammer2_voldata_unlock(hmp);
5877 hammer2_chain_bulkdrop(hammer2_chain_t *copy)
5879 KKASSERT(copy->bref.type == HAMMER2_BREF_TYPE_VOLUME);
5880 KKASSERT(copy->data);
5881 kfree(copy->data, copy->hmp->mchain);
5883 atomic_add_long(&hammer2_chain_allocs, -1);
5884 hammer2_chain_drop(copy);
5888 * Returns non-zero if the chain (INODE or DIRENT) matches the
5892 hammer2_chain_dirent_test(hammer2_chain_t *chain, const char *name,
5895 const hammer2_inode_data_t *ripdata;
5896 const hammer2_dirent_head_t *den;
5898 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
5899 ripdata = &chain->data->ipdata;
5900 if (ripdata->meta.name_len == name_len &&
5901 bcmp(ripdata->filename, name, name_len) == 0) {
5905 if (chain->bref.type == HAMMER2_BREF_TYPE_DIRENT &&
5906 chain->bref.embed.dirent.namlen == name_len) {
5907 den = &chain->bref.embed.dirent;
5908 if (name_len > sizeof(chain->bref.check.buf) &&
5909 bcmp(chain->data->buf, name, name_len) == 0) {
5912 if (name_len <= sizeof(chain->bref.check.buf) &&
5913 bcmp(chain->bref.check.buf, name, name_len) == 0) {