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 void hammer2_chain_rename_obref(hammer2_chain_t **parentp,
72 hammer2_chain_t *chain, hammer2_tid_t mtid,
73 int flags, hammer2_blockref_t *obref);
74 static int hammer2_chain_delete_obref(hammer2_chain_t *parent,
75 hammer2_chain_t *chain,
76 hammer2_tid_t mtid, int flags,
77 hammer2_blockref_t *obref);
78 static hammer2_io_t *hammer2_chain_drop_data(hammer2_chain_t *chain);
79 static hammer2_chain_t *hammer2_combined_find(
80 hammer2_chain_t *parent,
81 hammer2_blockref_t *base, int count,
82 hammer2_key_t *key_nextp,
83 hammer2_key_t key_beg, hammer2_key_t key_end,
84 hammer2_blockref_t **bresp);
86 static struct krate krate_h2me = { .freq = 1 };
89 * Basic RBTree for chains (core->rbtree and core->dbtree). Chains cannot
90 * overlap in the RB trees. Deleted chains are moved from rbtree to either
93 * Chains in delete-duplicate sequences can always iterate through core_entry
94 * to locate the live version of the chain.
96 RB_GENERATE(hammer2_chain_tree, hammer2_chain, rbnode, hammer2_chain_cmp);
99 hammer2_chain_cmp(hammer2_chain_t *chain1, hammer2_chain_t *chain2)
101 hammer2_key_t c1_beg;
102 hammer2_key_t c1_end;
103 hammer2_key_t c2_beg;
104 hammer2_key_t c2_end;
107 * Compare chains. Overlaps are not supposed to happen and catch
108 * any software issues early we count overlaps as a match.
110 c1_beg = chain1->bref.key;
111 c1_end = c1_beg + ((hammer2_key_t)1 << chain1->bref.keybits) - 1;
112 c2_beg = chain2->bref.key;
113 c2_end = c2_beg + ((hammer2_key_t)1 << chain2->bref.keybits) - 1;
115 if (c1_end < c2_beg) /* fully to the left */
117 if (c1_beg > c2_end) /* fully to the right */
119 return(0); /* overlap (must not cross edge boundary) */
123 * Assert that a chain has no media data associated with it.
126 hammer2_chain_assert_no_data(hammer2_chain_t *chain)
128 KKASSERT(chain->dio == NULL);
129 if (chain->bref.type != HAMMER2_BREF_TYPE_VOLUME &&
130 chain->bref.type != HAMMER2_BREF_TYPE_FREEMAP &&
132 panic("hammer2_assert_no_data: chain %p still has data", chain);
137 * Make a chain visible to the flusher. The flusher operates using a top-down
138 * recursion based on the ONFLUSH flag. It locates MODIFIED and UPDATE chains,
139 * flushes them, and updates blocks back to the volume root.
141 * This routine sets the ONFLUSH flag upward from the triggering chain until
142 * it hits an inode root or the volume root. Inode chains serve as inflection
143 * points, requiring the flusher to bridge across trees. Inodes include
144 * regular inodes, PFS roots (pmp->iroot), and the media super root
148 hammer2_chain_setflush(hammer2_chain_t *chain)
150 hammer2_chain_t *parent;
152 if ((chain->flags & HAMMER2_CHAIN_ONFLUSH) == 0) {
153 hammer2_spin_sh(&chain->core.spin);
154 while ((chain->flags & HAMMER2_CHAIN_ONFLUSH) == 0) {
155 atomic_set_int(&chain->flags, HAMMER2_CHAIN_ONFLUSH);
156 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE)
158 if ((parent = chain->parent) == NULL)
160 hammer2_spin_sh(&parent->core.spin);
161 hammer2_spin_unsh(&chain->core.spin);
164 hammer2_spin_unsh(&chain->core.spin);
169 * Allocate a new disconnected chain element representing the specified
170 * bref. chain->refs is set to 1 and the passed bref is copied to
171 * chain->bref. chain->bytes is derived from the bref.
173 * chain->pmp inherits pmp unless the chain is an inode (other than the
176 * NOTE: Returns a referenced but unlocked (because there is no core) chain.
179 hammer2_chain_alloc(hammer2_dev_t *hmp, hammer2_pfs_t *pmp,
180 hammer2_blockref_t *bref)
182 hammer2_chain_t *chain;
186 * Special case - radix of 0 indicates a chain that does not
187 * need a data reference (context is completely embedded in the
190 if ((int)(bref->data_off & HAMMER2_OFF_MASK_RADIX))
191 bytes = 1U << (int)(bref->data_off & HAMMER2_OFF_MASK_RADIX);
195 atomic_add_long(&hammer2_chain_allocs, 1);
198 * Construct the appropriate system structure.
201 case HAMMER2_BREF_TYPE_DIRENT:
202 case HAMMER2_BREF_TYPE_INODE:
203 case HAMMER2_BREF_TYPE_INDIRECT:
204 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
205 case HAMMER2_BREF_TYPE_DATA:
206 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
208 * Chain's are really only associated with the hmp but we
209 * maintain a pmp association for per-mount memory tracking
210 * purposes. The pmp can be NULL.
212 chain = kmalloc(sizeof(*chain), hmp->mchain, M_WAITOK | M_ZERO);
214 case HAMMER2_BREF_TYPE_VOLUME:
215 case HAMMER2_BREF_TYPE_FREEMAP:
217 * Only hammer2_chain_bulksnap() calls this function with these
220 chain = kmalloc(sizeof(*chain), hmp->mchain, M_WAITOK | M_ZERO);
224 panic("hammer2_chain_alloc: unrecognized blockref type: %d",
229 * Initialize the new chain structure. pmp must be set to NULL for
230 * chains belonging to the super-root topology of a device mount.
232 if (pmp == hmp->spmp)
239 chain->bytes = bytes;
241 chain->flags = HAMMER2_CHAIN_ALLOCATED;
242 lockinit(&chain->diolk, "chdio", 0, 0);
245 * Set the PFS boundary flag if this chain represents a PFS root.
247 if (bref->flags & HAMMER2_BREF_FLAG_PFSROOT)
248 atomic_set_int(&chain->flags, HAMMER2_CHAIN_PFSBOUNDARY);
249 hammer2_chain_core_init(chain);
255 * Initialize a chain's core structure. This structure used to be allocated
256 * but is now embedded.
258 * The core is not locked. No additional refs on the chain are made.
259 * (trans) must not be NULL if (core) is not NULL.
262 hammer2_chain_core_init(hammer2_chain_t *chain)
265 * Fresh core under nchain (no multi-homing of ochain's
268 RB_INIT(&chain->core.rbtree); /* live chains */
269 hammer2_mtx_init(&chain->lock, "h2chain");
273 * Add a reference to a chain element, preventing its destruction.
275 * (can be called with spinlock held)
278 hammer2_chain_ref(hammer2_chain_t *chain)
280 if (atomic_fetchadd_int(&chain->refs, 1) == 0) {
282 * Just flag that the chain was used and should be recycled
283 * on the LRU if it encounters it later.
285 if (chain->flags & HAMMER2_CHAIN_ONLRU)
286 atomic_set_int(&chain->flags, HAMMER2_CHAIN_LRUHINT);
290 * REMOVED - reduces contention, lru_list is more heuristical
293 * 0->non-zero transition must ensure that chain is removed
296 * NOTE: Already holding lru_spin here so we cannot call
297 * hammer2_chain_ref() to get it off lru_list, do
300 if (chain->flags & HAMMER2_CHAIN_ONLRU) {
301 hammer2_pfs_t *pmp = chain->pmp;
302 hammer2_spin_ex(&pmp->lru_spin);
303 if (chain->flags & HAMMER2_CHAIN_ONLRU) {
304 atomic_add_int(&pmp->lru_count, -1);
305 atomic_clear_int(&chain->flags,
306 HAMMER2_CHAIN_ONLRU);
307 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node);
309 hammer2_spin_unex(&pmp->lru_spin);
316 * Ref a locked chain and force the data to be held across an unlock.
317 * Chain must be currently locked. The user of the chain who desires
318 * to release the hold must call hammer2_chain_lock_unhold() to relock
319 * and unhold the chain, then unlock normally, or may simply call
320 * hammer2_chain_drop_unhold() (which is safer against deadlocks).
323 hammer2_chain_ref_hold(hammer2_chain_t *chain)
325 atomic_add_int(&chain->lockcnt, 1);
326 hammer2_chain_ref(chain);
330 * Insert the chain in the core rbtree.
332 * Normal insertions are placed in the live rbtree. Insertion of a deleted
333 * chain is a special case used by the flush code that is placed on the
334 * unstaged deleted list to avoid confusing the live view.
336 #define HAMMER2_CHAIN_INSERT_SPIN 0x0001
337 #define HAMMER2_CHAIN_INSERT_LIVE 0x0002
338 #define HAMMER2_CHAIN_INSERT_RACE 0x0004
342 hammer2_chain_insert(hammer2_chain_t *parent, hammer2_chain_t *chain,
343 int flags, int generation)
345 hammer2_chain_t *xchain;
348 if (flags & HAMMER2_CHAIN_INSERT_SPIN)
349 hammer2_spin_ex(&parent->core.spin);
352 * Interlocked by spinlock, check for race
354 if ((flags & HAMMER2_CHAIN_INSERT_RACE) &&
355 parent->core.generation != generation) {
356 error = HAMMER2_ERROR_EAGAIN;
363 xchain = RB_INSERT(hammer2_chain_tree, &parent->core.rbtree, chain);
364 KASSERT(xchain == NULL,
365 ("hammer2_chain_insert: collision %p %p (key=%016jx)",
366 chain, xchain, chain->bref.key));
367 atomic_set_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE);
368 chain->parent = parent;
369 ++parent->core.chain_count;
370 ++parent->core.generation; /* XXX incs for _get() too, XXX */
373 * We have to keep track of the effective live-view blockref count
374 * so the create code knows when to push an indirect block.
376 if (flags & HAMMER2_CHAIN_INSERT_LIVE)
377 atomic_add_int(&parent->core.live_count, 1);
379 if (flags & HAMMER2_CHAIN_INSERT_SPIN)
380 hammer2_spin_unex(&parent->core.spin);
385 * Drop the caller's reference to the chain. When the ref count drops to
386 * zero this function will try to disassociate the chain from its parent and
387 * deallocate it, then recursely drop the parent using the implied ref
388 * from the chain's chain->parent.
390 * Nobody should own chain's mutex on the 1->0 transition, unless this drop
391 * races an acquisition by another cpu. Therefore we can loop if we are
392 * unable to acquire the mutex, and refs is unlikely to be 1 unless we again
393 * race against another drop.
395 static hammer2_chain_t *hammer2_chain_lastdrop(hammer2_chain_t *chain,
397 static void hammer2_chain_lru_flush(hammer2_pfs_t *pmp);
400 hammer2_chain_drop(hammer2_chain_t *chain)
404 if (hammer2_debug & 0x200000)
407 KKASSERT(chain->refs > 0);
415 if (hammer2_mtx_ex_try(&chain->lock) == 0)
416 chain = hammer2_chain_lastdrop(chain, 0);
417 /* retry the same chain, or chain from lastdrop */
419 if (atomic_cmpset_int(&chain->refs, refs, refs - 1))
421 /* retry the same chain */
428 * Unhold a held and probably not-locked chain, ensure that the data is
429 * dropped on the 1->0 transition of lockcnt by obtaining an exclusive
430 * lock and then simply unlocking the chain.
433 hammer2_chain_drop_unhold(hammer2_chain_t *chain)
439 lockcnt = chain->lockcnt;
442 if (atomic_cmpset_int(&chain->lockcnt,
443 lockcnt, lockcnt - 1)) {
446 } else if (hammer2_mtx_ex_try(&chain->lock) == 0) {
447 hammer2_chain_unlock(chain);
451 * This situation can easily occur on SMP due to
452 * the gap inbetween the 1->0 transition and the
453 * final unlock. We cannot safely block on the
454 * mutex because lockcnt might go above 1.
456 * XXX Sleep for one tick if it takes too long.
459 if (iter > 1000 + hz) {
460 kprintf("hammer2: h2race1 %p\n", chain);
463 tsleep(&iter, 0, "h2race1", 1);
468 hammer2_chain_drop(chain);
472 * Handles the (potential) last drop of chain->refs from 1->0. Called with
473 * the mutex exclusively locked, refs == 1, and lockcnt 0. SMP races are
474 * possible against refs and lockcnt. We must dispose of the mutex on chain.
476 * This function returns an unlocked chain for recursive drop or NULL. It
477 * can return the same chain if it determines it has raced another ref.
481 * When two chains need to be recursively dropped we use the chain we
482 * would otherwise free to placehold the additional chain. It's a bit
483 * convoluted but we can't just recurse without potentially blowing out
486 * The chain cannot be freed if it has any children.
487 * The chain cannot be freed if flagged MODIFIED unless we can dispose of it.
488 * The chain cannot be freed if flagged UPDATE unless we can dispose of it.
489 * Any dedup registration can remain intact.
491 * The core spinlock is allowed to nest child-to-parent (not parent-to-child).
495 hammer2_chain_lastdrop(hammer2_chain_t *chain, int depth)
499 hammer2_chain_t *parent;
500 hammer2_chain_t *rdrop;
503 * We need chain's spinlock to interlock the sub-tree test.
504 * We already have chain's mutex, protecting chain->parent.
506 * Remember that chain->refs can be in flux.
508 hammer2_spin_ex(&chain->core.spin);
510 if (chain->parent != NULL) {
512 * If the chain has a parent the UPDATE bit prevents scrapping
513 * as the chain is needed to properly flush the parent. Try
514 * to complete the 1->0 transition and return NULL. Retry
515 * (return chain) if we are unable to complete the 1->0
516 * transition, else return NULL (nothing more to do).
518 * If the chain has a parent the MODIFIED bit prevents
521 * Chains with UPDATE/MODIFIED are *not* put on the LRU list!
523 if (chain->flags & (HAMMER2_CHAIN_UPDATE |
524 HAMMER2_CHAIN_MODIFIED)) {
525 if (atomic_cmpset_int(&chain->refs, 1, 0)) {
526 hammer2_spin_unex(&chain->core.spin);
527 hammer2_chain_assert_no_data(chain);
528 hammer2_mtx_unlock(&chain->lock);
531 hammer2_spin_unex(&chain->core.spin);
532 hammer2_mtx_unlock(&chain->lock);
536 /* spinlock still held */
537 } else if (chain->bref.type == HAMMER2_BREF_TYPE_VOLUME ||
538 chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP) {
540 * Retain the static vchain and fchain. Clear bits that
541 * are not relevant. Do not clear the MODIFIED bit,
542 * and certainly do not put it on the delayed-flush queue.
544 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
547 * The chain has no parent and can be flagged for destruction.
548 * Since it has no parent, UPDATE can also be cleared.
550 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DESTROY);
551 if (chain->flags & HAMMER2_CHAIN_UPDATE)
552 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
555 * If the chain has children we must propagate the DESTROY
556 * flag downward and rip the disconnected topology apart.
557 * This is accomplished by calling hammer2_flush() on the
560 * Any dedup is already handled by the underlying DIO, so
561 * we do not have to specifically flush it here.
563 if (chain->core.chain_count) {
564 hammer2_spin_unex(&chain->core.spin);
565 hammer2_flush(chain, HAMMER2_FLUSH_TOP |
567 hammer2_mtx_unlock(&chain->lock);
569 return(chain); /* retry drop */
573 * Otherwise we can scrap the MODIFIED bit if it is set,
574 * and continue along the freeing path.
576 * Be sure to clean-out any dedup bits. Without a parent
577 * this chain will no longer be visible to the flush code.
578 * Easy check data_off to avoid the volume root.
580 if (chain->flags & HAMMER2_CHAIN_MODIFIED) {
581 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
582 atomic_add_long(&hammer2_count_modified_chains, -1);
584 hammer2_pfs_memory_wakeup(chain->pmp);
586 /* spinlock still held */
589 /* spinlock still held */
592 * If any children exist we must leave the chain intact with refs == 0.
593 * They exist because chains are retained below us which have refs or
594 * may require flushing.
596 * Retry (return chain) if we fail to transition the refs to 0, else
597 * return NULL indication nothing more to do.
599 * Chains with children are NOT put on the LRU list.
601 if (chain->core.chain_count) {
602 if (atomic_cmpset_int(&chain->refs, 1, 0)) {
603 hammer2_spin_unex(&chain->core.spin);
604 hammer2_chain_assert_no_data(chain);
605 hammer2_mtx_unlock(&chain->lock);
608 hammer2_spin_unex(&chain->core.spin);
609 hammer2_mtx_unlock(&chain->lock);
613 /* spinlock still held */
614 /* no chains left under us */
617 * chain->core has no children left so no accessors can get to our
618 * chain from there. Now we have to lock the parent core to interlock
619 * remaining possible accessors that might bump chain's refs before
620 * we can safely drop chain's refs with intent to free the chain.
623 pmp = chain->pmp; /* can be NULL */
626 parent = chain->parent;
629 * WARNING! chain's spin lock is still held here, and other spinlocks
630 * will be acquired and released in the code below. We
631 * cannot be making fancy procedure calls!
635 * We can cache the chain if it is associated with a pmp
636 * and not flagged as being destroyed or requesting a full
637 * release. In this situation the chain is not removed
638 * from its parent, i.e. it can still be looked up.
640 * We intentionally do not cache DATA chains because these
641 * were likely used to load data into the logical buffer cache
642 * and will not be accessed again for some time.
645 (HAMMER2_CHAIN_DESTROY | HAMMER2_CHAIN_RELEASE)) == 0 &&
647 chain->bref.type != HAMMER2_BREF_TYPE_DATA) {
649 hammer2_spin_ex(&parent->core.spin);
650 if (atomic_cmpset_int(&chain->refs, 1, 0) == 0) {
652 * 1->0 transition failed, retry. Do not drop
653 * the chain's data yet!
656 hammer2_spin_unex(&parent->core.spin);
657 hammer2_spin_unex(&chain->core.spin);
658 hammer2_mtx_unlock(&chain->lock);
666 hammer2_chain_assert_no_data(chain);
669 * Make sure we are on the LRU list, clean up excessive
670 * LRU entries. We can only really drop one but there might
671 * be other entries that we can remove from the lru_list
674 * NOTE: HAMMER2_CHAIN_ONLRU may only be safely set when
675 * chain->core.spin AND pmp->lru_spin are held, but
676 * can be safely cleared only holding pmp->lru_spin.
678 if ((chain->flags & HAMMER2_CHAIN_ONLRU) == 0) {
679 hammer2_spin_ex(&pmp->lru_spin);
680 if ((chain->flags & HAMMER2_CHAIN_ONLRU) == 0) {
681 atomic_set_int(&chain->flags,
682 HAMMER2_CHAIN_ONLRU);
683 TAILQ_INSERT_TAIL(&pmp->lru_list,
685 atomic_add_int(&pmp->lru_count, 1);
687 if (pmp->lru_count < HAMMER2_LRU_LIMIT)
688 depth = 1; /* disable lru_list flush */
689 hammer2_spin_unex(&pmp->lru_spin);
691 /* disable lru flush */
696 hammer2_spin_unex(&parent->core.spin);
697 parent = NULL; /* safety */
699 hammer2_spin_unex(&chain->core.spin);
700 hammer2_mtx_unlock(&chain->lock);
703 * lru_list hysteresis (see above for depth overrides).
704 * Note that depth also prevents excessive lastdrop recursion.
707 hammer2_chain_lru_flush(pmp);
714 * Make sure we are not on the LRU list.
716 if (chain->flags & HAMMER2_CHAIN_ONLRU) {
717 hammer2_spin_ex(&pmp->lru_spin);
718 if (chain->flags & HAMMER2_CHAIN_ONLRU) {
719 atomic_add_int(&pmp->lru_count, -1);
720 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU);
721 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node);
723 hammer2_spin_unex(&pmp->lru_spin);
727 * Spinlock the parent and try to drop the last ref on chain.
728 * On success determine if we should dispose of the chain
729 * (remove the chain from its parent, etc).
731 * (normal core locks are top-down recursive but we define
732 * core spinlocks as bottom-up recursive, so this is safe).
735 hammer2_spin_ex(&parent->core.spin);
736 if (atomic_cmpset_int(&chain->refs, 1, 0) == 0) {
738 * 1->0 transition failed, retry.
740 hammer2_spin_unex(&parent->core.spin);
741 hammer2_spin_unex(&chain->core.spin);
742 hammer2_mtx_unlock(&chain->lock);
748 * 1->0 transition successful, parent spin held to prevent
749 * new lookups, chain spinlock held to protect parent field.
750 * Remove chain from the parent.
752 * If the chain is being removed from the parent's btree but
753 * is not bmapped, we have to adjust live_count downward. If
754 * it is bmapped then the blockref is retained in the parent
755 * as is its associated live_count. This case can occur when
756 * a chain added to the topology is unable to flush and is
757 * then later deleted.
759 if (chain->flags & HAMMER2_CHAIN_ONRBTREE) {
760 if ((parent->flags & HAMMER2_CHAIN_COUNTEDBREFS) &&
761 (chain->flags & HAMMER2_CHAIN_BMAPPED) == 0) {
762 atomic_add_int(&parent->core.live_count, -1);
764 RB_REMOVE(hammer2_chain_tree,
765 &parent->core.rbtree, chain);
766 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE);
767 --parent->core.chain_count;
768 chain->parent = NULL;
772 * If our chain was the last chain in the parent's core the
773 * core is now empty and its parent might have to be
774 * re-dropped if it has 0 refs.
776 if (parent->core.chain_count == 0) {
778 atomic_add_int(&rdrop->refs, 1);
780 if (atomic_cmpset_int(&rdrop->refs, 0, 1) == 0)
784 hammer2_spin_unex(&parent->core.spin);
785 parent = NULL; /* safety */
791 if (atomic_cmpset_int(&chain->refs, 1, 0) == 0) {
793 * 1->0 transition failed, retry.
795 hammer2_spin_unex(&parent->core.spin);
796 hammer2_spin_unex(&chain->core.spin);
797 hammer2_mtx_unlock(&chain->lock);
804 * Successful 1->0 transition, no parent, no children... no way for
805 * anyone to ref this chain any more. We can clean-up and free it.
807 * We still have the core spinlock, and core's chain_count is 0.
808 * Any parent spinlock is gone.
810 hammer2_spin_unex(&chain->core.spin);
811 hammer2_chain_assert_no_data(chain);
812 hammer2_mtx_unlock(&chain->lock);
813 KKASSERT(RB_EMPTY(&chain->core.rbtree) &&
814 chain->core.chain_count == 0);
817 * All locks are gone, no pointers remain to the chain, finish
820 KKASSERT((chain->flags & (HAMMER2_CHAIN_UPDATE |
821 HAMMER2_CHAIN_MODIFIED)) == 0);
824 * Once chain resources are gone we can use the now dead chain
825 * structure to placehold what might otherwise require a recursive
826 * drop, because we have potentially two things to drop and can only
827 * return one directly.
829 if (chain->flags & HAMMER2_CHAIN_ALLOCATED) {
830 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ALLOCATED);
832 kfree(chain, hmp->mchain);
836 * Possible chaining loop when parent re-drop needed.
842 * Heuristical flush of the LRU, try to reduce the number of entries
843 * on the LRU to (HAMMER2_LRU_LIMIT * 2 / 3). This procedure is called
844 * only when lru_count exceeds HAMMER2_LRU_LIMIT.
848 hammer2_chain_lru_flush(hammer2_pfs_t *pmp)
850 hammer2_chain_t *chain;
854 hammer2_spin_ex(&pmp->lru_spin);
855 while (pmp->lru_count > HAMMER2_LRU_LIMIT * 2 / 3) {
857 * Pick a chain off the lru_list, just recycle it quickly
858 * if LRUHINT is set (the chain was ref'd but left on
859 * the lru_list, so cycle to the end).
861 chain = TAILQ_FIRST(&pmp->lru_list);
862 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node);
864 if (chain->flags & HAMMER2_CHAIN_LRUHINT) {
865 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_LRUHINT);
866 TAILQ_INSERT_TAIL(&pmp->lru_list, chain, lru_node);
872 * Ok, we are off the LRU. We must adjust refs before we
873 * can safely clear the ONLRU flag.
875 atomic_add_int(&pmp->lru_count, -1);
876 if (atomic_cmpset_int(&chain->refs, 0, 1)) {
877 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU);
878 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
881 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU);
884 hammer2_spin_unex(&pmp->lru_spin);
889 * If we picked a chain off the lru list we may be able to lastdrop
890 * it. Use a depth of 1 to prevent excessive lastdrop recursion.
900 if (hammer2_mtx_ex_try(&chain->lock) == 0)
901 chain = hammer2_chain_lastdrop(chain, 1);
902 /* retry the same chain, or chain from lastdrop */
904 if (atomic_cmpset_int(&chain->refs, refs, refs - 1))
906 /* retry the same chain */
914 * On last lock release.
916 static hammer2_io_t *
917 hammer2_chain_drop_data(hammer2_chain_t *chain)
921 if ((dio = chain->dio) != NULL) {
925 switch(chain->bref.type) {
926 case HAMMER2_BREF_TYPE_VOLUME:
927 case HAMMER2_BREF_TYPE_FREEMAP:
930 if (chain->data != NULL) {
931 hammer2_spin_unex(&chain->core.spin);
932 panic("chain data not null: "
933 "chain %p bref %016jx.%02x "
934 "refs %d parent %p dio %p data %p",
935 chain, chain->bref.data_off,
936 chain->bref.type, chain->refs,
938 chain->dio, chain->data);
940 KKASSERT(chain->data == NULL);
948 * Lock a referenced chain element, acquiring its data with I/O if necessary,
949 * and specify how you would like the data to be resolved.
951 * If an I/O or other fatal error occurs, chain->error will be set to non-zero.
953 * The lock is allowed to recurse, multiple locking ops will aggregate
954 * the requested resolve types. Once data is assigned it will not be
955 * removed until the last unlock.
957 * HAMMER2_RESOLVE_NEVER - Do not resolve the data element.
958 * (typically used to avoid device/logical buffer
961 * HAMMER2_RESOLVE_MAYBE - Do not resolve data elements for chains in
962 * the INITIAL-create state (indirect blocks only).
964 * Do not resolve data elements for DATA chains.
965 * (typically used to avoid device/logical buffer
968 * HAMMER2_RESOLVE_ALWAYS- Always resolve the data element.
970 * HAMMER2_RESOLVE_SHARED- (flag) The chain is locked shared, otherwise
971 * it will be locked exclusive.
973 * HAMMER2_RESOLVE_NONBLOCK- (flag) The chain is locked non-blocking. If
974 * the lock fails, EAGAIN is returned.
976 * NOTE: Embedded elements (volume header, inodes) are always resolved
979 * NOTE: Specifying HAMMER2_RESOLVE_ALWAYS on a newly-created non-embedded
980 * element will instantiate and zero its buffer, and flush it on
983 * NOTE: (data) elements are normally locked RESOLVE_NEVER or RESOLVE_MAYBE
984 * so as not to instantiate a device buffer, which could alias against
985 * a logical file buffer. However, if ALWAYS is specified the
986 * device buffer will be instantiated anyway.
988 * NOTE: The return value is always 0 unless NONBLOCK is specified, in which
989 * case it can be either 0 or EAGAIN.
991 * WARNING! This function blocks on I/O if data needs to be fetched. This
992 * blocking can run concurrent with other compatible lock holders
993 * who do not need data returning. The lock is not upgraded to
994 * exclusive during a data fetch, a separate bit is used to
995 * interlock I/O. However, an exclusive lock holder can still count
996 * on being interlocked against an I/O fetch managed by a shared
1000 hammer2_chain_lock(hammer2_chain_t *chain, int how)
1002 KKASSERT(chain->refs > 0);
1004 if (how & HAMMER2_RESOLVE_NONBLOCK) {
1006 * We still have to bump lockcnt before acquiring the lock,
1007 * even for non-blocking operation, because the unlock code
1008 * live-loops on lockcnt == 1 when dropping the last lock.
1010 * If the non-blocking operation fails we have to use a
1011 * ref+drop+unhold sequence to undo the mess (or write a
1012 * hammer2_chain_unhold() function that doesn't drop).
1014 * NOTE: LOCKAGAIN must always succeed without blocking,
1015 * even if NONBLOCK is specified.
1017 atomic_add_int(&chain->lockcnt, 1);
1018 if (how & HAMMER2_RESOLVE_SHARED) {
1019 if (how & HAMMER2_RESOLVE_LOCKAGAIN) {
1020 hammer2_mtx_sh_again(&chain->lock);
1022 if (hammer2_mtx_sh_try(&chain->lock) != 0) {
1023 hammer2_chain_ref(chain);
1024 hammer2_chain_drop_unhold(chain);
1029 if (hammer2_mtx_ex_try(&chain->lock) != 0) {
1030 hammer2_chain_ref(chain);
1031 hammer2_chain_drop_unhold(chain);
1035 ++curthread->td_tracker;
1038 * Get the appropriate lock. If LOCKAGAIN is flagged with
1039 * SHARED the caller expects a shared lock to already be
1040 * present and we are giving it another ref. This case must
1041 * importantly not block if there is a pending exclusive lock
1044 atomic_add_int(&chain->lockcnt, 1);
1045 if (how & HAMMER2_RESOLVE_SHARED) {
1046 if (how & HAMMER2_RESOLVE_LOCKAGAIN) {
1047 hammer2_mtx_sh_again(&chain->lock);
1049 hammer2_mtx_sh(&chain->lock);
1052 hammer2_mtx_ex(&chain->lock);
1054 ++curthread->td_tracker;
1058 * If we already have a valid data pointer make sure the data is
1059 * synchronized to the current cpu, and then no further action is
1064 hammer2_io_bkvasync(chain->dio);
1069 * Do we have to resolve the data? This is generally only
1070 * applicable to HAMMER2_BREF_TYPE_DATA which is special-cased.
1071 * Other BREF types expects the data to be there.
1073 switch(how & HAMMER2_RESOLVE_MASK) {
1074 case HAMMER2_RESOLVE_NEVER:
1076 case HAMMER2_RESOLVE_MAYBE:
1077 if (chain->flags & HAMMER2_CHAIN_INITIAL)
1079 if (chain->bref.type == HAMMER2_BREF_TYPE_DATA)
1082 if (chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE)
1084 if (chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_LEAF)
1088 case HAMMER2_RESOLVE_ALWAYS:
1094 * Caller requires data
1096 hammer2_chain_load_data(chain);
1102 * Lock the chain, retain the hold, and drop the data persistence count.
1103 * The data should remain valid because we never transitioned lockcnt
1107 hammer2_chain_lock_unhold(hammer2_chain_t *chain, int how)
1109 hammer2_chain_lock(chain, how);
1110 atomic_add_int(&chain->lockcnt, -1);
1115 * Downgrade an exclusive chain lock to a shared chain lock.
1117 * NOTE: There is no upgrade equivalent due to the ease of
1118 * deadlocks in that direction.
1121 hammer2_chain_lock_downgrade(hammer2_chain_t *chain)
1123 hammer2_mtx_downgrade(&chain->lock);
1128 * Issue I/O and install chain->data. Caller must hold a chain lock, lock
1129 * may be of any type.
1131 * Once chain->data is set it cannot be disposed of until all locks are
1134 * Make sure the data is synchronized to the current cpu.
1137 hammer2_chain_load_data(hammer2_chain_t *chain)
1139 hammer2_blockref_t *bref;
1146 * Degenerate case, data already present, or chain has no media
1147 * reference to load.
1151 hammer2_io_bkvasync(chain->dio);
1154 if ((chain->bref.data_off & ~HAMMER2_OFF_MASK_RADIX) == 0)
1158 KKASSERT(hmp != NULL);
1161 * Gain the IOINPROG bit, interlocked block.
1167 oflags = chain->flags;
1169 if (oflags & HAMMER2_CHAIN_IOINPROG) {
1170 nflags = oflags | HAMMER2_CHAIN_IOSIGNAL;
1171 tsleep_interlock(&chain->flags, 0);
1172 if (atomic_cmpset_int(&chain->flags, oflags, nflags)) {
1173 tsleep(&chain->flags, PINTERLOCKED,
1178 nflags = oflags | HAMMER2_CHAIN_IOINPROG;
1179 if (atomic_cmpset_int(&chain->flags, oflags, nflags)) {
1187 * We own CHAIN_IOINPROG
1189 * Degenerate case if we raced another load.
1193 hammer2_io_bkvasync(chain->dio);
1198 * We must resolve to a device buffer, either by issuing I/O or
1199 * by creating a zero-fill element. We do not mark the buffer
1200 * dirty when creating a zero-fill element (the hammer2_chain_modify()
1201 * API must still be used to do that).
1203 * The device buffer is variable-sized in powers of 2 down
1204 * to HAMMER2_MIN_ALLOC (typically 1K). A 64K physical storage
1205 * chunk always contains buffers of the same size. (XXX)
1207 * The minimum physical IO size may be larger than the variable
1210 bref = &chain->bref;
1213 * The getblk() optimization can only be used on newly created
1214 * elements if the physical block size matches the request.
1216 if (chain->flags & HAMMER2_CHAIN_INITIAL) {
1217 error = hammer2_io_new(hmp, bref->type,
1218 bref->data_off, chain->bytes,
1221 error = hammer2_io_bread(hmp, bref->type,
1222 bref->data_off, chain->bytes,
1224 hammer2_adjreadcounter(&chain->bref, chain->bytes);
1227 chain->error = HAMMER2_ERROR_EIO;
1228 kprintf("hammer2_chain_lock: I/O error %016jx: %d\n",
1229 (intmax_t)bref->data_off, error);
1230 hammer2_io_bqrelse(&chain->dio);
1236 * This isn't perfect and can be ignored on OSs which do not have
1237 * an indication as to whether a buffer is coming from cache or
1238 * if I/O was actually issued for the read. TESTEDGOOD will work
1239 * pretty well without the B_IOISSUED logic because chains are
1240 * cached, but in that situation (without B_IOISSUED) it will not
1241 * detect whether a re-read via I/O is corrupted verses the original
1244 * We can't re-run the CRC on every fresh lock. That would be
1245 * insanely expensive.
1247 * If the underlying kernel buffer covers the entire chain we can
1248 * use the B_IOISSUED indication to determine if we have to re-run
1249 * the CRC on chain data for chains that managed to stay cached
1250 * across the kernel disposal of the original buffer.
1252 if ((dio = chain->dio) != NULL && dio->bp) {
1253 struct buf *bp = dio->bp;
1255 if (dio->psize == chain->bytes &&
1256 (bp->b_flags & B_IOISSUED)) {
1257 atomic_clear_int(&chain->flags,
1258 HAMMER2_CHAIN_TESTEDGOOD);
1259 bp->b_flags &= ~B_IOISSUED;
1264 * NOTE: A locked chain's data cannot be modified without first
1265 * calling hammer2_chain_modify().
1269 * Clear INITIAL. In this case we used io_new() and the buffer has
1270 * been zero'd and marked dirty.
1272 * NOTE: hammer2_io_data() call issues bkvasync()
1274 bdata = hammer2_io_data(chain->dio, chain->bref.data_off);
1276 if (chain->flags & HAMMER2_CHAIN_INITIAL) {
1277 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1278 chain->bref.flags |= HAMMER2_BREF_FLAG_ZERO;
1279 } else if (chain->flags & HAMMER2_CHAIN_MODIFIED) {
1281 * check data not currently synchronized due to
1282 * modification. XXX assumes data stays in the buffer
1283 * cache, which might not be true (need biodep on flush
1284 * to calculate crc? or simple crc?).
1286 } else if ((chain->flags & HAMMER2_CHAIN_TESTEDGOOD) == 0) {
1287 if (hammer2_chain_testcheck(chain, bdata) == 0) {
1288 chain->error = HAMMER2_ERROR_CHECK;
1290 atomic_set_int(&chain->flags, HAMMER2_CHAIN_TESTEDGOOD);
1295 * Setup the data pointer, either pointing it to an embedded data
1296 * structure and copying the data from the buffer, or pointing it
1299 * The buffer is not retained when copying to an embedded data
1300 * structure in order to avoid potential deadlocks or recursions
1301 * on the same physical buffer.
1303 * WARNING! Other threads can start using the data the instant we
1304 * set chain->data non-NULL.
1306 switch (bref->type) {
1307 case HAMMER2_BREF_TYPE_VOLUME:
1308 case HAMMER2_BREF_TYPE_FREEMAP:
1310 * Copy data from bp to embedded buffer
1312 panic("hammer2_chain_load_data: unresolved volume header");
1314 case HAMMER2_BREF_TYPE_DIRENT:
1315 KKASSERT(chain->bytes != 0);
1317 case HAMMER2_BREF_TYPE_INODE:
1318 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
1319 case HAMMER2_BREF_TYPE_INDIRECT:
1320 case HAMMER2_BREF_TYPE_DATA:
1321 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1324 * Point data at the device buffer and leave dio intact.
1326 chain->data = (void *)bdata;
1331 * Release HAMMER2_CHAIN_IOINPROG and signal waiters if requested.
1338 oflags = chain->flags;
1339 nflags = oflags & ~(HAMMER2_CHAIN_IOINPROG |
1340 HAMMER2_CHAIN_IOSIGNAL);
1341 KKASSERT(oflags & HAMMER2_CHAIN_IOINPROG);
1342 if (atomic_cmpset_int(&chain->flags, oflags, nflags)) {
1343 if (oflags & HAMMER2_CHAIN_IOSIGNAL)
1344 wakeup(&chain->flags);
1351 * Unlock and deref a chain element.
1353 * Remember that the presence of children under chain prevent the chain's
1354 * destruction but do not add additional references, so the dio will still
1358 hammer2_chain_unlock(hammer2_chain_t *chain)
1364 --curthread->td_tracker;
1367 * If multiple locks are present (or being attempted) on this
1368 * particular chain we can just unlock, drop refs, and return.
1370 * Otherwise fall-through on the 1->0 transition.
1373 lockcnt = chain->lockcnt;
1374 KKASSERT(lockcnt > 0);
1377 if (atomic_cmpset_int(&chain->lockcnt,
1378 lockcnt, lockcnt - 1)) {
1379 hammer2_mtx_unlock(&chain->lock);
1382 } else if (hammer2_mtx_upgrade_try(&chain->lock) == 0) {
1383 /* while holding the mutex exclusively */
1384 if (atomic_cmpset_int(&chain->lockcnt, 1, 0))
1388 * This situation can easily occur on SMP due to
1389 * the gap inbetween the 1->0 transition and the
1390 * final unlock. We cannot safely block on the
1391 * mutex because lockcnt might go above 1.
1393 * XXX Sleep for one tick if it takes too long.
1395 if (++iter > 1000) {
1396 if (iter > 1000 + hz) {
1397 kprintf("hammer2: h2race2 %p\n", chain);
1400 tsleep(&iter, 0, "h2race2", 1);
1408 * Last unlock / mutex upgraded to exclusive. Drop the data
1411 dio = hammer2_chain_drop_data(chain);
1413 hammer2_io_bqrelse(&dio);
1414 hammer2_mtx_unlock(&chain->lock);
1418 * Unlock and hold chain data intact
1421 hammer2_chain_unlock_hold(hammer2_chain_t *chain)
1423 atomic_add_int(&chain->lockcnt, 1);
1424 hammer2_chain_unlock(chain);
1428 * Helper to obtain the blockref[] array base and count for a chain.
1430 * XXX Not widely used yet, various use cases need to be validated and
1431 * converted to use this function.
1434 hammer2_blockref_t *
1435 hammer2_chain_base_and_count(hammer2_chain_t *parent, int *countp)
1437 hammer2_blockref_t *base;
1440 if (parent->flags & HAMMER2_CHAIN_INITIAL) {
1443 switch(parent->bref.type) {
1444 case HAMMER2_BREF_TYPE_INODE:
1445 count = HAMMER2_SET_COUNT;
1447 case HAMMER2_BREF_TYPE_INDIRECT:
1448 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1449 count = parent->bytes / sizeof(hammer2_blockref_t);
1451 case HAMMER2_BREF_TYPE_VOLUME:
1452 count = HAMMER2_SET_COUNT;
1454 case HAMMER2_BREF_TYPE_FREEMAP:
1455 count = HAMMER2_SET_COUNT;
1458 panic("hammer2_chain_create_indirect: "
1459 "unrecognized blockref type: %d",
1465 switch(parent->bref.type) {
1466 case HAMMER2_BREF_TYPE_INODE:
1467 base = &parent->data->ipdata.u.blockset.blockref[0];
1468 count = HAMMER2_SET_COUNT;
1470 case HAMMER2_BREF_TYPE_INDIRECT:
1471 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1472 base = &parent->data->npdata[0];
1473 count = parent->bytes / sizeof(hammer2_blockref_t);
1475 case HAMMER2_BREF_TYPE_VOLUME:
1476 base = &parent->data->voldata.
1477 sroot_blockset.blockref[0];
1478 count = HAMMER2_SET_COUNT;
1480 case HAMMER2_BREF_TYPE_FREEMAP:
1481 base = &parent->data->blkset.blockref[0];
1482 count = HAMMER2_SET_COUNT;
1485 panic("hammer2_chain_create_indirect: "
1486 "unrecognized blockref type: %d",
1498 * This counts the number of live blockrefs in a block array and
1499 * also calculates the point at which all remaining blockrefs are empty.
1500 * This routine can only be called on a live chain.
1502 * Caller holds the chain locked, but possibly with a shared lock. We
1503 * must use an exclusive spinlock to prevent corruption.
1505 * NOTE: Flag is not set until after the count is complete, allowing
1506 * callers to test the flag without holding the spinlock.
1508 * NOTE: If base is NULL the related chain is still in the INITIAL
1509 * state and there are no blockrefs to count.
1511 * NOTE: live_count may already have some counts accumulated due to
1512 * creation and deletion and could even be initially negative.
1515 hammer2_chain_countbrefs(hammer2_chain_t *chain,
1516 hammer2_blockref_t *base, int count)
1518 hammer2_spin_ex(&chain->core.spin);
1519 if ((chain->flags & HAMMER2_CHAIN_COUNTEDBREFS) == 0) {
1521 while (--count >= 0) {
1522 if (base[count].type)
1525 chain->core.live_zero = count + 1;
1526 while (count >= 0) {
1527 if (base[count].type)
1528 atomic_add_int(&chain->core.live_count,
1533 chain->core.live_zero = 0;
1535 /* else do not modify live_count */
1536 atomic_set_int(&chain->flags, HAMMER2_CHAIN_COUNTEDBREFS);
1538 hammer2_spin_unex(&chain->core.spin);
1542 * Resize the chain's physical storage allocation in-place. This function does
1543 * not usually adjust the data pointer and must be followed by (typically) a
1544 * hammer2_chain_modify() call to copy any old data over and adjust the
1547 * Chains can be resized smaller without reallocating the storage. Resizing
1548 * larger will reallocate the storage. Excess or prior storage is reclaimed
1549 * asynchronously at a later time.
1551 * An nradix value of 0 is special-cased to mean that the storage should
1552 * be disassociated, that is the chain is being resized to 0 bytes (not 1
1555 * Must be passed an exclusively locked parent and chain.
1557 * This function is mostly used with DATA blocks locked RESOLVE_NEVER in order
1558 * to avoid instantiating a device buffer that conflicts with the vnode data
1559 * buffer. However, because H2 can compress or encrypt data, the chain may
1560 * have a dio assigned to it in those situations, and they do not conflict.
1562 * XXX return error if cannot resize.
1565 hammer2_chain_resize(hammer2_chain_t *chain,
1566 hammer2_tid_t mtid, hammer2_off_t dedup_off,
1567 int nradix, int flags)
1577 * Only data and indirect blocks can be resized for now.
1578 * (The volu root, inodes, and freemap elements use a fixed size).
1580 KKASSERT(chain != &hmp->vchain);
1581 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_DATA ||
1582 chain->bref.type == HAMMER2_BREF_TYPE_INDIRECT ||
1583 chain->bref.type == HAMMER2_BREF_TYPE_DIRENT);
1586 * Nothing to do if the element is already the proper size
1588 obytes = chain->bytes;
1589 nbytes = (nradix) ? (1U << nradix) : 0;
1590 if (obytes == nbytes)
1591 return (chain->error);
1594 * Make sure the old data is instantiated so we can copy it. If this
1595 * is a data block, the device data may be superfluous since the data
1596 * might be in a logical block, but compressed or encrypted data is
1599 * NOTE: The modify will set BMAPUPD for us if BMAPPED is set.
1601 error = hammer2_chain_modify(chain, mtid, dedup_off, 0);
1606 * Relocate the block, even if making it smaller (because different
1607 * block sizes may be in different regions).
1609 * NOTE: Operation does not copy the data and may only be used
1610 * to resize data blocks in-place, or directory entry blocks
1611 * which are about to be modified in some manner.
1613 error = hammer2_freemap_alloc(chain, nbytes);
1617 chain->bytes = nbytes;
1620 * We don't want the followup chain_modify() to try to copy data
1621 * from the old (wrong-sized) buffer. It won't know how much to
1622 * copy. This case should only occur during writes when the
1623 * originator already has the data to write in-hand.
1626 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_DATA ||
1627 chain->bref.type == HAMMER2_BREF_TYPE_DIRENT);
1628 hammer2_io_brelse(&chain->dio);
1631 return (chain->error);
1635 * Set the chain modified so its data can be changed by the caller, or
1636 * install deduplicated data. The caller must call this routine for each
1637 * set of modifications it makes, even if the chain is already flagged
1640 * Sets bref.modify_tid to mtid only if mtid != 0. Note that bref.modify_tid
1641 * is a CLC (cluster level change) field and is not updated by parent
1642 * propagation during a flush.
1644 * Returns an appropriate HAMMER2_ERROR_* code, which will generally reflect
1645 * chain->error except for HAMMER2_ERROR_ENOSPC. If the allocation fails
1646 * due to no space available, HAMMER2_ERROR_ENOSPC is returned and the chain
1647 * remains unmodified with its old data ref intact and chain->error
1652 * If the DEDUPABLE flag is set in the chain the storage must be reallocated
1653 * even if the chain is still flagged MODIFIED. In this case the chain's
1654 * DEDUPABLE flag will be cleared once the new storage has been assigned.
1656 * If the caller passes a non-zero dedup_off we will use it to assign the
1657 * new storage. The MODIFIED flag will be *CLEARED* in this case, and
1658 * DEDUPABLE will be set (NOTE: the UPDATE flag is always set). The caller
1659 * must not modify the data content upon return.
1662 hammer2_chain_modify(hammer2_chain_t *chain, hammer2_tid_t mtid,
1663 hammer2_off_t dedup_off, int flags)
1665 hammer2_blockref_t obref;
1676 obref = chain->bref;
1677 KKASSERT((chain->flags & HAMMER2_CHAIN_FICTITIOUS) == 0);
1680 * Data is not optional for freemap chains (we must always be sure
1681 * to copy the data on COW storage allocations).
1683 if (chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE ||
1684 chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) {
1685 KKASSERT((chain->flags & HAMMER2_CHAIN_INITIAL) ||
1686 (flags & HAMMER2_MODIFY_OPTDATA) == 0);
1690 * Data must be resolved if already assigned, unless explicitly
1691 * flagged otherwise. If we cannot safety load the data the
1692 * modification fails and we return early.
1694 if (chain->data == NULL && chain->bytes != 0 &&
1695 (flags & HAMMER2_MODIFY_OPTDATA) == 0 &&
1696 (chain->bref.data_off & ~HAMMER2_OFF_MASK_RADIX)) {
1697 hammer2_chain_load_data(chain);
1699 return (chain->error);
1704 * Set MODIFIED to indicate that the chain has been modified. A new
1705 * allocation is required when modifying a chain.
1707 * Set UPDATE to ensure that the blockref is updated in the parent.
1709 * If MODIFIED is already set determine if we can reuse the assigned
1710 * data block or if we need a new data block.
1712 if ((chain->flags & HAMMER2_CHAIN_MODIFIED) == 0) {
1714 * Must set modified bit.
1716 atomic_add_long(&hammer2_count_modified_chains, 1);
1717 atomic_set_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
1718 hammer2_pfs_memory_inc(chain->pmp); /* can be NULL */
1722 * We may be able to avoid a copy-on-write if the chain's
1723 * check mode is set to NONE and the chain's current
1724 * modify_tid is beyond the last explicit snapshot tid.
1726 * This implements HAMMER2's overwrite-in-place feature.
1728 * NOTE! This data-block cannot be used as a de-duplication
1729 * source when the check mode is set to NONE.
1731 if ((chain->bref.type == HAMMER2_BREF_TYPE_DATA ||
1732 chain->bref.type == HAMMER2_BREF_TYPE_DIRENT) &&
1733 (chain->flags & HAMMER2_CHAIN_INITIAL) == 0 &&
1734 (chain->flags & HAMMER2_CHAIN_DEDUPABLE) == 0 &&
1735 HAMMER2_DEC_CHECK(chain->bref.methods) ==
1736 HAMMER2_CHECK_NONE &&
1738 chain->bref.modify_tid >
1739 chain->pmp->iroot->meta.pfs_lsnap_tid) {
1741 * Sector overwrite allowed.
1746 * Sector overwrite not allowed, must copy-on-write.
1750 } else if (chain->flags & HAMMER2_CHAIN_DEDUPABLE) {
1752 * If the modified chain was registered for dedup we need
1753 * a new allocation. This only happens for delayed-flush
1754 * chains (i.e. which run through the front-end buffer
1761 * Already flagged modified, no new allocation is needed.
1768 * Flag parent update required.
1770 if ((chain->flags & HAMMER2_CHAIN_UPDATE) == 0) {
1771 atomic_set_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
1778 * The XOP code returns held but unlocked focus chains. This
1779 * prevents the chain from being destroyed but does not prevent
1780 * it from being modified. diolk is used to interlock modifications
1781 * against XOP frontend accesses to the focus.
1783 * This allows us to theoretically avoid deadlocking the frontend
1784 * if one of the backends lock up by not formally locking the
1785 * focused chain in the frontend. In addition, the synchronization
1786 * code relies on this mechanism to avoid deadlocking concurrent
1787 * synchronization threads.
1789 lockmgr(&chain->diolk, LK_EXCLUSIVE);
1792 * The modification or re-modification requires an allocation and
1793 * possible COW. If an error occurs, the previous content and data
1794 * reference is retained and the modification fails.
1796 * If dedup_off is non-zero, the caller is requesting a deduplication
1797 * rather than a modification. The MODIFIED bit is not set and the
1798 * data offset is set to the deduplication offset. The data cannot
1801 * NOTE: The dedup offset is allowed to be in a partially free state
1802 * and we must be sure to reset it to a fully allocated state
1803 * to force two bulkfree passes to free it again.
1805 * NOTE: Only applicable when chain->bytes != 0.
1807 * XXX can a chain already be marked MODIFIED without a data
1808 * assignment? If not, assert here instead of testing the case.
1810 if (chain != &hmp->vchain && chain != &hmp->fchain &&
1812 if ((chain->bref.data_off & ~HAMMER2_OFF_MASK_RADIX) == 0 ||
1816 * NOTE: We do not have to remove the dedup
1817 * registration because the area is still
1818 * allocated and the underlying DIO will
1822 chain->bref.data_off = dedup_off;
1823 chain->bytes = 1 << (dedup_off &
1824 HAMMER2_OFF_MASK_RADIX);
1826 atomic_clear_int(&chain->flags,
1827 HAMMER2_CHAIN_MODIFIED);
1828 atomic_add_long(&hammer2_count_modified_chains,
1831 hammer2_pfs_memory_wakeup(chain->pmp);
1832 hammer2_freemap_adjust(hmp, &chain->bref,
1833 HAMMER2_FREEMAP_DORECOVER);
1834 atomic_set_int(&chain->flags,
1835 HAMMER2_CHAIN_DEDUPABLE);
1837 error = hammer2_freemap_alloc(chain,
1839 atomic_clear_int(&chain->flags,
1840 HAMMER2_CHAIN_DEDUPABLE);
1846 * Stop here if error. We have to undo any flag bits we might
1851 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
1852 atomic_add_long(&hammer2_count_modified_chains, -1);
1854 hammer2_pfs_memory_wakeup(chain->pmp);
1857 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
1859 lockmgr(&chain->diolk, LK_RELEASE);
1865 * Update mirror_tid and modify_tid. modify_tid is only updated
1866 * if not passed as zero (during flushes, parent propagation passes
1869 * NOTE: chain->pmp could be the device spmp.
1871 chain->bref.mirror_tid = hmp->voldata.mirror_tid + 1;
1873 chain->bref.modify_tid = mtid;
1876 * Set BMAPUPD to tell the flush code that an existing blockmap entry
1877 * requires updating as well as to tell the delete code that the
1878 * chain's blockref might not exactly match (in terms of physical size
1879 * or block offset) the one in the parent's blocktable. The base key
1880 * of course will still match.
1882 if (chain->flags & HAMMER2_CHAIN_BMAPPED)
1883 atomic_set_int(&chain->flags, HAMMER2_CHAIN_BMAPUPD);
1886 * Short-cut data blocks which the caller does not need an actual
1887 * data reference to (aka OPTDATA), as long as the chain does not
1888 * already have a data pointer to the data. This generally means
1889 * that the modifications are being done via the logical buffer cache.
1890 * The INITIAL flag relates only to the device data buffer and thus
1891 * remains unchange in this situation.
1893 * This code also handles bytes == 0 (most dirents).
1895 if (chain->bref.type == HAMMER2_BREF_TYPE_DATA &&
1896 (flags & HAMMER2_MODIFY_OPTDATA) &&
1897 chain->data == NULL) {
1898 KKASSERT(chain->dio == NULL);
1903 * Clearing the INITIAL flag (for indirect blocks) indicates that
1904 * we've processed the uninitialized storage allocation.
1906 * If this flag is already clear we are likely in a copy-on-write
1907 * situation but we have to be sure NOT to bzero the storage if
1908 * no data is present.
1910 if (chain->flags & HAMMER2_CHAIN_INITIAL) {
1911 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1918 * Instantiate data buffer and possibly execute COW operation
1920 switch(chain->bref.type) {
1921 case HAMMER2_BREF_TYPE_VOLUME:
1922 case HAMMER2_BREF_TYPE_FREEMAP:
1924 * The data is embedded, no copy-on-write operation is
1927 KKASSERT(chain->dio == NULL);
1929 case HAMMER2_BREF_TYPE_DIRENT:
1931 * The data might be fully embedded.
1933 if (chain->bytes == 0) {
1934 KKASSERT(chain->dio == NULL);
1938 case HAMMER2_BREF_TYPE_INODE:
1939 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
1940 case HAMMER2_BREF_TYPE_DATA:
1941 case HAMMER2_BREF_TYPE_INDIRECT:
1942 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1944 * Perform the copy-on-write operation
1946 * zero-fill or copy-on-write depending on whether
1947 * chain->data exists or not and set the dirty state for
1948 * the new buffer. hammer2_io_new() will handle the
1951 * If a dedup_off was supplied this is an existing block
1952 * and no COW, copy, or further modification is required.
1954 KKASSERT(chain != &hmp->vchain && chain != &hmp->fchain);
1956 if (wasinitial && dedup_off == 0) {
1957 error = hammer2_io_new(hmp, chain->bref.type,
1958 chain->bref.data_off,
1959 chain->bytes, &dio);
1961 error = hammer2_io_bread(hmp, chain->bref.type,
1962 chain->bref.data_off,
1963 chain->bytes, &dio);
1965 hammer2_adjreadcounter(&chain->bref, chain->bytes);
1968 * If an I/O error occurs make sure callers cannot accidently
1969 * modify the old buffer's contents and corrupt the filesystem.
1971 * NOTE: hammer2_io_data() call issues bkvasync()
1974 kprintf("hammer2_chain_modify: hmp=%p I/O error\n",
1976 chain->error = HAMMER2_ERROR_EIO;
1977 hammer2_io_brelse(&dio);
1978 hammer2_io_brelse(&chain->dio);
1983 bdata = hammer2_io_data(dio, chain->bref.data_off);
1987 * COW (unless a dedup).
1989 KKASSERT(chain->dio != NULL);
1990 if (chain->data != (void *)bdata && dedup_off == 0) {
1991 bcopy(chain->data, bdata, chain->bytes);
1993 } else if (wasinitial == 0) {
1995 * We have a problem. We were asked to COW but
1996 * we don't have any data to COW with!
1998 panic("hammer2_chain_modify: having a COW %p\n",
2003 * Retire the old buffer, replace with the new. Dirty or
2004 * redirty the new buffer.
2006 * WARNING! The system buffer cache may have already flushed
2007 * the buffer, so we must be sure to [re]dirty it
2008 * for further modification.
2010 * If dedup_off was supplied, the caller is not
2011 * expected to make any further modification to the
2014 * WARNING! hammer2_get_gdata() assumes dio never transitions
2015 * through NULL in order to optimize away unnecessary
2021 if ((tio = chain->dio) != NULL)
2022 hammer2_io_bqrelse(&tio);
2023 chain->data = (void *)bdata;
2026 hammer2_io_setdirty(dio);
2030 panic("hammer2_chain_modify: illegal non-embedded type %d",
2037 * setflush on parent indicating that the parent must recurse down
2038 * to us. Do not call on chain itself which might already have it
2042 hammer2_chain_setflush(chain->parent);
2043 lockmgr(&chain->diolk, LK_RELEASE);
2045 return (chain->error);
2049 * Modify the chain associated with an inode.
2052 hammer2_chain_modify_ip(hammer2_inode_t *ip, hammer2_chain_t *chain,
2053 hammer2_tid_t mtid, int flags)
2057 hammer2_inode_modify(ip);
2058 error = hammer2_chain_modify(chain, mtid, 0, flags);
2064 * Volume header data locks
2067 hammer2_voldata_lock(hammer2_dev_t *hmp)
2069 lockmgr(&hmp->vollk, LK_EXCLUSIVE);
2073 hammer2_voldata_unlock(hammer2_dev_t *hmp)
2075 lockmgr(&hmp->vollk, LK_RELEASE);
2079 hammer2_voldata_modify(hammer2_dev_t *hmp)
2081 if ((hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) == 0) {
2082 atomic_add_long(&hammer2_count_modified_chains, 1);
2083 atomic_set_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
2084 hammer2_pfs_memory_inc(hmp->vchain.pmp);
2089 * This function returns the chain at the nearest key within the specified
2090 * range. The returned chain will be referenced but not locked.
2092 * This function will recurse through chain->rbtree as necessary and will
2093 * return a *key_nextp suitable for iteration. *key_nextp is only set if
2094 * the iteration value is less than the current value of *key_nextp.
2096 * The caller should use (*key_nextp) to calculate the actual range of
2097 * the returned element, which will be (key_beg to *key_nextp - 1), because
2098 * there might be another element which is superior to the returned element
2101 * (*key_nextp) can be passed as key_beg in an iteration only while non-NULL
2102 * chains continue to be returned. On EOF (*key_nextp) may overflow since
2103 * it will wind up being (key_end + 1).
2105 * WARNING! Must be called with child's spinlock held. Spinlock remains
2106 * held through the operation.
2108 struct hammer2_chain_find_info {
2109 hammer2_chain_t *best;
2110 hammer2_key_t key_beg;
2111 hammer2_key_t key_end;
2112 hammer2_key_t key_next;
2115 static int hammer2_chain_find_cmp(hammer2_chain_t *child, void *data);
2116 static int hammer2_chain_find_callback(hammer2_chain_t *child, void *data);
2120 hammer2_chain_find(hammer2_chain_t *parent, hammer2_key_t *key_nextp,
2121 hammer2_key_t key_beg, hammer2_key_t key_end)
2123 struct hammer2_chain_find_info info;
2126 info.key_beg = key_beg;
2127 info.key_end = key_end;
2128 info.key_next = *key_nextp;
2130 RB_SCAN(hammer2_chain_tree, &parent->core.rbtree,
2131 hammer2_chain_find_cmp, hammer2_chain_find_callback,
2133 *key_nextp = info.key_next;
2135 kprintf("chain_find %p %016jx:%016jx next=%016jx\n",
2136 parent, key_beg, key_end, *key_nextp);
2144 hammer2_chain_find_cmp(hammer2_chain_t *child, void *data)
2146 struct hammer2_chain_find_info *info = data;
2147 hammer2_key_t child_beg;
2148 hammer2_key_t child_end;
2150 child_beg = child->bref.key;
2151 child_end = child_beg + ((hammer2_key_t)1 << child->bref.keybits) - 1;
2153 if (child_end < info->key_beg)
2155 if (child_beg > info->key_end)
2162 hammer2_chain_find_callback(hammer2_chain_t *child, void *data)
2164 struct hammer2_chain_find_info *info = data;
2165 hammer2_chain_t *best;
2166 hammer2_key_t child_end;
2169 * WARNING! Layerq is scanned forwards, exact matches should keep
2170 * the existing info->best.
2172 if ((best = info->best) == NULL) {
2174 * No previous best. Assign best
2177 } else if (best->bref.key <= info->key_beg &&
2178 child->bref.key <= info->key_beg) {
2183 /*info->best = child;*/
2184 } else if (child->bref.key < best->bref.key) {
2186 * Child has a nearer key and best is not flush with key_beg.
2187 * Set best to child. Truncate key_next to the old best key.
2190 if (info->key_next > best->bref.key || info->key_next == 0)
2191 info->key_next = best->bref.key;
2192 } else if (child->bref.key == best->bref.key) {
2194 * If our current best is flush with the child then this
2195 * is an illegal overlap.
2197 * key_next will automatically be limited to the smaller of
2198 * the two end-points.
2204 * Keep the current best but truncate key_next to the child's
2207 * key_next will also automatically be limited to the smaller
2208 * of the two end-points (probably not necessary for this case
2209 * but we do it anyway).
2211 if (info->key_next > child->bref.key || info->key_next == 0)
2212 info->key_next = child->bref.key;
2216 * Always truncate key_next based on child's end-of-range.
2218 child_end = child->bref.key + ((hammer2_key_t)1 << child->bref.keybits);
2219 if (child_end && (info->key_next > child_end || info->key_next == 0))
2220 info->key_next = child_end;
2226 * Retrieve the specified chain from a media blockref, creating the
2227 * in-memory chain structure which reflects it. The returned chain is
2228 * held and locked according to (how) (HAMMER2_RESOLVE_*). The caller must
2229 * handle crc-checks and so forth, and should check chain->error before
2230 * assuming that the data is good.
2232 * To handle insertion races pass the INSERT_RACE flag along with the
2233 * generation number of the core. NULL will be returned if the generation
2234 * number changes before we have a chance to insert the chain. Insert
2235 * races can occur because the parent might be held shared.
2237 * Caller must hold the parent locked shared or exclusive since we may
2238 * need the parent's bref array to find our block.
2240 * WARNING! chain->pmp is always set to NULL for any chain representing
2241 * part of the super-root topology.
2244 hammer2_chain_get(hammer2_chain_t *parent, int generation,
2245 hammer2_blockref_t *bref, int how)
2247 hammer2_dev_t *hmp = parent->hmp;
2248 hammer2_chain_t *chain;
2252 * Allocate a chain structure representing the existing media
2253 * entry. Resulting chain has one ref and is not locked.
2255 if (bref->flags & HAMMER2_BREF_FLAG_PFSROOT)
2256 chain = hammer2_chain_alloc(hmp, NULL, bref);
2258 chain = hammer2_chain_alloc(hmp, parent->pmp, bref);
2259 /* ref'd chain returned */
2262 * Flag that the chain is in the parent's blockmap so delete/flush
2263 * knows what to do with it.
2265 atomic_set_int(&chain->flags, HAMMER2_CHAIN_BMAPPED);
2268 * chain must be locked to avoid unexpected ripouts
2270 hammer2_chain_lock(chain, how);
2273 * Link the chain into its parent. A spinlock is required to safely
2274 * access the RBTREE, and it is possible to collide with another
2275 * hammer2_chain_get() operation because the caller might only hold
2276 * a shared lock on the parent.
2278 * NOTE: Get races can occur quite often when we distribute
2279 * asynchronous read-aheads across multiple threads.
2281 KKASSERT(parent->refs > 0);
2282 error = hammer2_chain_insert(parent, chain,
2283 HAMMER2_CHAIN_INSERT_SPIN |
2284 HAMMER2_CHAIN_INSERT_RACE,
2287 KKASSERT((chain->flags & HAMMER2_CHAIN_ONRBTREE) == 0);
2288 /*kprintf("chain %p get race\n", chain);*/
2289 hammer2_chain_unlock(chain);
2290 hammer2_chain_drop(chain);
2293 KKASSERT(chain->flags & HAMMER2_CHAIN_ONRBTREE);
2297 * Return our new chain referenced but not locked, or NULL if
2304 * Lookup initialization/completion API
2307 hammer2_chain_lookup_init(hammer2_chain_t *parent, int flags)
2309 hammer2_chain_ref(parent);
2310 if (flags & HAMMER2_LOOKUP_SHARED) {
2311 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS |
2312 HAMMER2_RESOLVE_SHARED);
2314 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2320 hammer2_chain_lookup_done(hammer2_chain_t *parent)
2323 hammer2_chain_unlock(parent);
2324 hammer2_chain_drop(parent);
2329 * Take the locked chain and return a locked parent. The chain remains
2330 * locked on return, but may have to be temporarily unlocked to acquire
2331 * the parent. Because of this, (chain) must be stable and cannot be
2332 * deleted while it was temporarily unlocked (typically means that (chain)
2335 * Pass HAMMER2_RESOLVE_* flags in flags.
2337 * This will work even if the chain is errored, and the caller can check
2338 * parent->error on return if desired since the parent will be locked.
2340 * This function handles the lock order reversal.
2343 hammer2_chain_getparent(hammer2_chain_t *chain, int flags)
2345 hammer2_chain_t *parent;
2348 * Be careful of order, chain must be unlocked before parent
2349 * is locked below to avoid a deadlock. Try it trivially first.
2351 parent = chain->parent;
2353 panic("hammer2_chain_getparent: no parent");
2354 hammer2_chain_ref(parent);
2355 if (hammer2_chain_lock(parent, flags|HAMMER2_RESOLVE_NONBLOCK) == 0)
2359 hammer2_chain_unlock(chain);
2360 hammer2_chain_lock(parent, flags);
2361 hammer2_chain_lock(chain, flags);
2364 * Parent relinking races are quite common. We have to get
2365 * it right or we will blow up the block table.
2367 if (chain->parent == parent)
2369 hammer2_chain_unlock(parent);
2370 hammer2_chain_drop(parent);
2372 parent = chain->parent;
2374 panic("hammer2_chain_getparent: no parent");
2375 hammer2_chain_ref(parent);
2381 * Take the locked chain and return a locked parent. The chain is unlocked
2382 * and dropped. *chainp is set to the returned parent as a convenience.
2383 * Pass HAMMER2_RESOLVE_* flags in flags.
2385 * This will work even if the chain is errored, and the caller can check
2386 * parent->error on return if desired since the parent will be locked.
2388 * The chain does NOT need to be stable. We use a tracking structure
2389 * to track the expected parent if the chain is deleted out from under us.
2391 * This function handles the lock order reversal.
2394 hammer2_chain_repparent(hammer2_chain_t **chainp, int flags)
2396 hammer2_chain_t *chain;
2397 hammer2_chain_t *parent;
2398 struct hammer2_reptrack reptrack;
2399 struct hammer2_reptrack **repp;
2402 * Be careful of order, chain must be unlocked before parent
2403 * is locked below to avoid a deadlock. Try it trivially first.
2406 parent = chain->parent;
2407 if (parent == NULL) {
2408 hammer2_spin_unex(&chain->core.spin);
2409 panic("hammer2_chain_repparent: no parent");
2411 hammer2_chain_ref(parent);
2412 if (hammer2_chain_lock(parent, flags|HAMMER2_RESOLVE_NONBLOCK) == 0) {
2413 hammer2_chain_unlock(chain);
2414 hammer2_chain_drop(chain);
2421 * Ok, now it gets a bit nasty. There are multiple situations where
2422 * the parent might be in the middle of a deletion, or where the child
2423 * (chain) might be deleted the instant we let go of its lock.
2424 * We can potentially end up in a no-win situation!
2426 * In particular, the indirect_maintenance() case can cause these
2429 * To deal with this we install a reptrack structure in the parent
2430 * This reptrack structure 'owns' the parent ref and will automatically
2431 * migrate to the parent's parent if the parent is deleted permanently.
2433 hammer2_spin_init(&reptrack.spin, "h2reptrk");
2434 reptrack.chain = parent;
2435 hammer2_chain_ref(parent); /* for the reptrack */
2437 hammer2_spin_ex(&parent->core.spin);
2438 reptrack.next = parent->core.reptrack;
2439 parent->core.reptrack = &reptrack;
2440 hammer2_spin_unex(&parent->core.spin);
2442 hammer2_chain_unlock(chain);
2443 hammer2_chain_drop(chain);
2444 chain = NULL; /* gone */
2447 * At the top of this loop, chain is gone and parent is refd both
2448 * by us explicitly AND via our reptrack. We are attempting to
2452 hammer2_chain_lock(parent, flags);
2454 if (reptrack.chain == parent)
2456 hammer2_chain_unlock(parent);
2457 hammer2_chain_drop(parent);
2459 kprintf("hammer2: debug REPTRACK %p->%p\n",
2460 parent, reptrack.chain);
2461 hammer2_spin_ex(&reptrack.spin);
2462 parent = reptrack.chain;
2463 hammer2_chain_ref(parent);
2464 hammer2_spin_unex(&reptrack.spin);
2468 * Once parent is locked and matches our reptrack, our reptrack
2469 * will be stable and we have our parent. We can unlink our
2472 * WARNING! Remember that the chain lock might be shared. Chains
2473 * locked shared have stable parent linkages.
2475 hammer2_spin_ex(&parent->core.spin);
2476 repp = &parent->core.reptrack;
2477 while (*repp != &reptrack)
2478 repp = &(*repp)->next;
2479 *repp = reptrack.next;
2480 hammer2_spin_unex(&parent->core.spin);
2482 hammer2_chain_drop(parent); /* reptrack ref */
2483 *chainp = parent; /* return parent lock+ref */
2489 * Dispose of any linked reptrack structures in (chain) by shifting them to
2490 * (parent). Both (chain) and (parent) must be exclusively locked.
2492 * This is interlocked against any children of (chain) on the other side.
2493 * No children so remain as-of when this is called so we can test
2494 * core.reptrack without holding the spin-lock.
2496 * Used whenever the caller intends to permanently delete chains related
2497 * to topological recursions (BREF_TYPE_INDIRECT, BREF_TYPE_FREEMAP_NODE),
2498 * where the chains underneath the node being deleted are given a new parent
2499 * above the node being deleted.
2503 hammer2_chain_repchange(hammer2_chain_t *parent, hammer2_chain_t *chain)
2505 struct hammer2_reptrack *reptrack;
2507 KKASSERT(chain->core.live_count == 0 && RB_EMPTY(&chain->core.rbtree));
2508 while (chain->core.reptrack) {
2509 hammer2_spin_ex(&parent->core.spin);
2510 hammer2_spin_ex(&chain->core.spin);
2511 reptrack = chain->core.reptrack;
2512 if (reptrack == NULL) {
2513 hammer2_spin_unex(&chain->core.spin);
2514 hammer2_spin_unex(&parent->core.spin);
2517 hammer2_spin_ex(&reptrack->spin);
2518 chain->core.reptrack = reptrack->next;
2519 reptrack->chain = parent;
2520 reptrack->next = parent->core.reptrack;
2521 parent->core.reptrack = reptrack;
2522 hammer2_chain_ref(parent); /* reptrack */
2524 hammer2_spin_unex(&chain->core.spin);
2525 hammer2_spin_unex(&parent->core.spin);
2526 kprintf("hammer2: debug repchange %p %p->%p\n",
2527 reptrack, chain, parent);
2528 hammer2_chain_drop(chain); /* reptrack */
2533 * Locate the first chain whos key range overlaps (key_beg, key_end) inclusive.
2534 * (*parentp) typically points to an inode but can also point to a related
2535 * indirect block and this function will recurse upwards and find the inode
2536 * or the nearest undeleted indirect block covering the key range.
2538 * This function unconditionally sets *errorp, replacing any previous value.
2540 * (*parentp) must be exclusive or shared locked (depending on flags) and
2541 * referenced and can be an inode or an existing indirect block within the
2544 * If (*parent) is errored out, this function will not attempt to recurse
2545 * the radix tree and will return NULL along with an appropriate *errorp.
2546 * If NULL is returned and *errorp is 0, the requested lookup could not be
2549 * On return (*parentp) will be modified to point at the deepest parent chain
2550 * element encountered during the search, as a helper for an insertion or
2553 * The new (*parentp) will be locked shared or exclusive (depending on flags),
2554 * and referenced, and the old will be unlocked and dereferenced (no change
2555 * if they are both the same). This is particularly important if the caller
2556 * wishes to insert a new chain, (*parentp) will be set properly even if NULL
2557 * is returned, as long as no error occurred.
2559 * The matching chain will be returned locked according to flags.
2563 * NULL is returned if no match was found, but (*parentp) will still
2564 * potentially be adjusted.
2566 * On return (*key_nextp) will point to an iterative value for key_beg.
2567 * (If NULL is returned (*key_nextp) is set to (key_end + 1)).
2569 * This function will also recurse up the chain if the key is not within the
2570 * current parent's range. (*parentp) can never be set to NULL. An iteration
2571 * can simply allow (*parentp) to float inside the loop.
2573 * NOTE! chain->data is not always resolved. By default it will not be
2574 * resolved for BREF_TYPE_DATA, FREEMAP_NODE, or FREEMAP_LEAF. Use
2575 * HAMMER2_LOOKUP_ALWAYS to force resolution (but be careful w/
2576 * BREF_TYPE_DATA as the device buffer can alias the logical file
2581 hammer2_chain_lookup(hammer2_chain_t **parentp, hammer2_key_t *key_nextp,
2582 hammer2_key_t key_beg, hammer2_key_t key_end,
2583 int *errorp, int flags)
2586 hammer2_chain_t *parent;
2587 hammer2_chain_t *chain;
2588 hammer2_blockref_t *base;
2589 hammer2_blockref_t *bref;
2590 hammer2_blockref_t bcopy;
2591 hammer2_key_t scan_beg;
2592 hammer2_key_t scan_end;
2594 int how_always = HAMMER2_RESOLVE_ALWAYS;
2595 int how_maybe = HAMMER2_RESOLVE_MAYBE;
2598 int maxloops = 300000;
2599 volatile hammer2_mtx_t save_mtx;
2601 if (flags & HAMMER2_LOOKUP_ALWAYS) {
2602 how_maybe = how_always;
2603 how = HAMMER2_RESOLVE_ALWAYS;
2604 } else if (flags & HAMMER2_LOOKUP_NODATA) {
2605 how = HAMMER2_RESOLVE_NEVER;
2607 how = HAMMER2_RESOLVE_MAYBE;
2609 if (flags & HAMMER2_LOOKUP_SHARED) {
2610 how_maybe |= HAMMER2_RESOLVE_SHARED;
2611 how_always |= HAMMER2_RESOLVE_SHARED;
2612 how |= HAMMER2_RESOLVE_SHARED;
2616 * Recurse (*parentp) upward if necessary until the parent completely
2617 * encloses the key range or we hit the inode.
2619 * Handle races against the flusher deleting indirect nodes on its
2620 * way back up by continuing to recurse upward past the deletion.
2626 while (parent->bref.type == HAMMER2_BREF_TYPE_INDIRECT ||
2627 parent->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2628 scan_beg = parent->bref.key;
2629 scan_end = scan_beg +
2630 ((hammer2_key_t)1 << parent->bref.keybits) - 1;
2631 if ((parent->flags & HAMMER2_CHAIN_DELETED) == 0) {
2632 if (key_beg >= scan_beg && key_end <= scan_end)
2635 parent = hammer2_chain_repparent(parentp, how_maybe);
2638 if (--maxloops == 0)
2639 panic("hammer2_chain_lookup: maxloops");
2641 * Locate the blockref array. Currently we do a fully associative
2642 * search through the array.
2644 switch(parent->bref.type) {
2645 case HAMMER2_BREF_TYPE_INODE:
2647 * Special shortcut for embedded data returns the inode
2648 * itself. Callers must detect this condition and access
2649 * the embedded data (the strategy code does this for us).
2651 * This is only applicable to regular files and softlinks.
2653 * We need a second lock on parent. Since we already have
2654 * a lock we must pass LOCKAGAIN to prevent unexpected
2655 * blocking (we don't want to block on a second shared
2656 * ref if an exclusive lock is pending)
2658 if (parent->data->ipdata.meta.op_flags &
2659 HAMMER2_OPFLAG_DIRECTDATA) {
2660 if (flags & HAMMER2_LOOKUP_NODIRECT) {
2662 *key_nextp = key_end + 1;
2665 hammer2_chain_ref(parent);
2666 hammer2_chain_lock(parent, how_always |
2667 HAMMER2_RESOLVE_LOCKAGAIN);
2668 *key_nextp = key_end + 1;
2671 base = &parent->data->ipdata.u.blockset.blockref[0];
2672 count = HAMMER2_SET_COUNT;
2674 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2675 case HAMMER2_BREF_TYPE_INDIRECT:
2677 * Handle MATCHIND on the parent
2679 if (flags & HAMMER2_LOOKUP_MATCHIND) {
2680 scan_beg = parent->bref.key;
2681 scan_end = scan_beg +
2682 ((hammer2_key_t)1 << parent->bref.keybits) - 1;
2683 if (key_beg == scan_beg && key_end == scan_end) {
2685 hammer2_chain_ref(chain);
2686 hammer2_chain_lock(chain, how_maybe);
2687 *key_nextp = scan_end + 1;
2693 * Optimize indirect blocks in the INITIAL state to avoid
2696 if (parent->flags & HAMMER2_CHAIN_INITIAL) {
2699 if (parent->data == NULL) {
2700 kprintf("parent->data is NULL %p\n", parent);
2702 tsleep(parent, 0, "xxx", 0);
2704 base = &parent->data->npdata[0];
2706 count = parent->bytes / sizeof(hammer2_blockref_t);
2708 case HAMMER2_BREF_TYPE_VOLUME:
2709 base = &parent->data->voldata.sroot_blockset.blockref[0];
2710 count = HAMMER2_SET_COUNT;
2712 case HAMMER2_BREF_TYPE_FREEMAP:
2713 base = &parent->data->blkset.blockref[0];
2714 count = HAMMER2_SET_COUNT;
2717 kprintf("hammer2_chain_lookup: unrecognized "
2718 "blockref(B) type: %d",
2721 tsleep(&base, 0, "dead", 0);
2722 panic("hammer2_chain_lookup: unrecognized "
2723 "blockref(B) type: %d",
2725 base = NULL; /* safety */
2726 count = 0; /* safety */
2730 * No lookup is possible if the parent is errored. We delayed
2731 * this check as long as we could to ensure that the parent backup,
2732 * embedded data, and MATCHIND code could still execute.
2734 if (parent->error) {
2735 *errorp = parent->error;
2740 * Merged scan to find next candidate.
2742 * hammer2_base_*() functions require the parent->core.live_* fields
2743 * to be synchronized.
2745 * We need to hold the spinlock to access the block array and RB tree
2746 * and to interlock chain creation.
2748 if ((parent->flags & HAMMER2_CHAIN_COUNTEDBREFS) == 0)
2749 hammer2_chain_countbrefs(parent, base, count);
2754 hammer2_spin_ex(&parent->core.spin);
2755 chain = hammer2_combined_find(parent, base, count,
2759 generation = parent->core.generation;
2762 * Exhausted parent chain, iterate.
2765 KKASSERT(chain == NULL);
2766 hammer2_spin_unex(&parent->core.spin);
2767 if (key_beg == key_end) /* short cut single-key case */
2771 * Stop if we reached the end of the iteration.
2773 if (parent->bref.type != HAMMER2_BREF_TYPE_INDIRECT &&
2774 parent->bref.type != HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2779 * Calculate next key, stop if we reached the end of the
2780 * iteration, otherwise go up one level and loop.
2782 key_beg = parent->bref.key +
2783 ((hammer2_key_t)1 << parent->bref.keybits);
2784 if (key_beg == 0 || key_beg > key_end)
2786 parent = hammer2_chain_repparent(parentp, how_maybe);
2791 * Selected from blockref or in-memory chain.
2794 if (chain == NULL) {
2795 hammer2_spin_unex(&parent->core.spin);
2796 if (bcopy.type == HAMMER2_BREF_TYPE_INDIRECT ||
2797 bcopy.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2798 chain = hammer2_chain_get(parent, generation,
2801 chain = hammer2_chain_get(parent, generation,
2807 hammer2_chain_ref(chain);
2808 hammer2_spin_unex(&parent->core.spin);
2811 * chain is referenced but not locked. We must lock the
2812 * chain to obtain definitive state.
2814 if (bcopy.type == HAMMER2_BREF_TYPE_INDIRECT ||
2815 bcopy.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2816 hammer2_chain_lock(chain, how_maybe);
2818 hammer2_chain_lock(chain, how);
2820 KKASSERT(chain->parent == parent);
2822 if (bcmp(&bcopy, &chain->bref, sizeof(bcopy)) ||
2823 chain->parent != parent) {
2824 hammer2_chain_unlock(chain);
2825 hammer2_chain_drop(chain);
2826 chain = NULL; /* SAFETY */
2832 * Skip deleted chains (XXX cache 'i' end-of-block-array? XXX)
2834 * NOTE: Chain's key range is not relevant as there might be
2835 * one-offs within the range that are not deleted.
2837 * NOTE: Lookups can race delete-duplicate because
2838 * delete-duplicate does not lock the parent's core
2839 * (they just use the spinlock on the core).
2841 if (chain->flags & HAMMER2_CHAIN_DELETED) {
2842 kprintf("skip deleted chain %016jx.%02x key=%016jx\n",
2843 chain->bref.data_off, chain->bref.type,
2845 hammer2_chain_unlock(chain);
2846 hammer2_chain_drop(chain);
2847 chain = NULL; /* SAFETY */
2848 key_beg = *key_nextp;
2849 if (key_beg == 0 || key_beg > key_end)
2855 * If the chain element is an indirect block it becomes the new
2856 * parent and we loop on it. We must maintain our top-down locks
2857 * to prevent the flusher from interfering (i.e. doing a
2858 * delete-duplicate and leaving us recursing down a deleted chain).
2860 * The parent always has to be locked with at least RESOLVE_MAYBE
2861 * so we can access its data. It might need a fixup if the caller
2862 * passed incompatible flags. Be careful not to cause a deadlock
2863 * as a data-load requires an exclusive lock.
2865 * If HAMMER2_LOOKUP_MATCHIND is set and the indirect block's key
2866 * range is within the requested key range we return the indirect
2867 * block and do NOT loop. This is usually only used to acquire
2870 if (chain->bref.type == HAMMER2_BREF_TYPE_INDIRECT ||
2871 chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2872 save_mtx = parent->lock;
2873 hammer2_chain_unlock(parent);
2874 hammer2_chain_drop(parent);
2875 *parentp = parent = chain;
2876 chain = NULL; /* SAFETY */
2881 * All done, return the locked chain.
2883 * If the caller does not want a locked chain, replace the lock with
2884 * a ref. Perhaps this can eventually be optimized to not obtain the
2885 * lock in the first place for situations where the data does not
2886 * need to be resolved.
2888 * NOTE! A chain->error must be tested by the caller upon return.
2889 * *errorp is only set based on issues which occur while
2890 * trying to reach the chain.
2896 * After having issued a lookup we can iterate all matching keys.
2898 * If chain is non-NULL we continue the iteration from just after it's index.
2900 * If chain is NULL we assume the parent was exhausted and continue the
2901 * iteration at the next parent.
2903 * If a fatal error occurs (typically an I/O error), a dummy chain is
2904 * returned with chain->error and error-identifying information set. This
2905 * chain will assert if you try to do anything fancy with it.
2907 * XXX Depending on where the error occurs we should allow continued iteration.
2909 * parent must be locked on entry and remains locked throughout. chain's
2910 * lock status must match flags. Chain is always at least referenced.
2912 * WARNING! The MATCHIND flag does not apply to this function.
2915 hammer2_chain_next(hammer2_chain_t **parentp, hammer2_chain_t *chain,
2916 hammer2_key_t *key_nextp,
2917 hammer2_key_t key_beg, hammer2_key_t key_end,
2918 int *errorp, int flags)
2920 hammer2_chain_t *parent;
2924 * Calculate locking flags for upward recursion.
2926 how_maybe = HAMMER2_RESOLVE_MAYBE;
2927 if (flags & HAMMER2_LOOKUP_SHARED)
2928 how_maybe |= HAMMER2_RESOLVE_SHARED;
2934 * Calculate the next index and recalculate the parent if necessary.
2937 key_beg = chain->bref.key +
2938 ((hammer2_key_t)1 << chain->bref.keybits);
2939 hammer2_chain_unlock(chain);
2940 hammer2_chain_drop(chain);
2943 * chain invalid past this point, but we can still do a
2944 * pointer comparison w/parent.
2946 * Any scan where the lookup returned degenerate data embedded
2947 * in the inode has an invalid index and must terminate.
2949 if (chain == parent)
2951 if (key_beg == 0 || key_beg > key_end)
2954 } else if (parent->bref.type != HAMMER2_BREF_TYPE_INDIRECT &&
2955 parent->bref.type != HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2957 * We reached the end of the iteration.
2962 * Continue iteration with next parent unless the current
2963 * parent covers the range.
2965 * (This also handles the case of a deleted, empty indirect
2968 key_beg = parent->bref.key +
2969 ((hammer2_key_t)1 << parent->bref.keybits);
2970 if (key_beg == 0 || key_beg > key_end)
2972 parent = hammer2_chain_repparent(parentp, how_maybe);
2978 return (hammer2_chain_lookup(parentp, key_nextp,
2984 * Caller wishes to iterate chains under parent, loading new chains into
2985 * chainp. Caller must initialize *chainp to NULL and *firstp to 1, and
2986 * then call hammer2_chain_scan() repeatedly until a non-zero return.
2987 * During the scan, *firstp will be set to 0 and (*chainp) will be replaced
2988 * with the returned chain for the scan. The returned *chainp will be
2989 * locked and referenced. Any prior contents will be unlocked and dropped.
2991 * Caller should check the return value. A normal scan EOF will return
2992 * exactly HAMMER2_ERROR_EOF. Any other non-zero value indicates an
2993 * error trying to access parent data. Any error in the returned chain
2994 * must be tested separately by the caller.
2996 * (*chainp) is dropped on each scan, but will only be set if the returned
2997 * element itself can recurse. Leaf elements are NOT resolved, loaded, or
2998 * returned via *chainp. The caller will get their bref only.
3000 * The raw scan function is similar to lookup/next but does not seek to a key.
3001 * Blockrefs are iterated via first_bref = (parent, NULL) and
3002 * next_chain = (parent, bref).
3004 * The passed-in parent must be locked and its data resolved. The function
3005 * nominally returns a locked and referenced *chainp != NULL for chains
3006 * the caller might need to recurse on (and will dipose of any *chainp passed
3007 * in). The caller must check the chain->bref.type either way.
3010 hammer2_chain_scan(hammer2_chain_t *parent, hammer2_chain_t **chainp,
3011 hammer2_blockref_t *bref, int *firstp,
3015 hammer2_blockref_t *base;
3016 hammer2_blockref_t *bref_ptr;
3018 hammer2_key_t next_key;
3019 hammer2_chain_t *chain = NULL;
3021 int how_always = HAMMER2_RESOLVE_ALWAYS;
3022 int how_maybe = HAMMER2_RESOLVE_MAYBE;
3025 int maxloops = 300000;
3032 * Scan flags borrowed from lookup.
3034 if (flags & HAMMER2_LOOKUP_ALWAYS) {
3035 how_maybe = how_always;
3036 how = HAMMER2_RESOLVE_ALWAYS;
3037 } else if (flags & HAMMER2_LOOKUP_NODATA) {
3038 how = HAMMER2_RESOLVE_NEVER;
3040 how = HAMMER2_RESOLVE_MAYBE;
3042 if (flags & HAMMER2_LOOKUP_SHARED) {
3043 how_maybe |= HAMMER2_RESOLVE_SHARED;
3044 how_always |= HAMMER2_RESOLVE_SHARED;
3045 how |= HAMMER2_RESOLVE_SHARED;
3049 * Calculate key to locate first/next element, unlocking the previous
3050 * element as we go. Be careful, the key calculation can overflow.
3052 * (also reset bref to NULL)
3058 key = bref->key + ((hammer2_key_t)1 << bref->keybits);
3059 if ((chain = *chainp) != NULL) {
3061 hammer2_chain_unlock(chain);
3062 hammer2_chain_drop(chain);
3066 error |= HAMMER2_ERROR_EOF;
3072 if (parent->error) {
3073 error = parent->error;
3076 if (--maxloops == 0)
3077 panic("hammer2_chain_scan: maxloops");
3080 * Locate the blockref array. Currently we do a fully associative
3081 * search through the array.
3083 switch(parent->bref.type) {
3084 case HAMMER2_BREF_TYPE_INODE:
3086 * An inode with embedded data has no sub-chains.
3088 * WARNING! Bulk scan code may pass a static chain marked
3089 * as BREF_TYPE_INODE with a copy of the volume
3090 * root blockset to snapshot the volume.
3092 if (parent->data->ipdata.meta.op_flags &
3093 HAMMER2_OPFLAG_DIRECTDATA) {
3094 error |= HAMMER2_ERROR_EOF;
3097 base = &parent->data->ipdata.u.blockset.blockref[0];
3098 count = HAMMER2_SET_COUNT;
3100 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3101 case HAMMER2_BREF_TYPE_INDIRECT:
3103 * Optimize indirect blocks in the INITIAL state to avoid
3106 if (parent->flags & HAMMER2_CHAIN_INITIAL) {
3109 if (parent->data == NULL)
3110 panic("parent->data is NULL");
3111 base = &parent->data->npdata[0];
3113 count = parent->bytes / sizeof(hammer2_blockref_t);
3115 case HAMMER2_BREF_TYPE_VOLUME:
3116 base = &parent->data->voldata.sroot_blockset.blockref[0];
3117 count = HAMMER2_SET_COUNT;
3119 case HAMMER2_BREF_TYPE_FREEMAP:
3120 base = &parent->data->blkset.blockref[0];
3121 count = HAMMER2_SET_COUNT;
3124 panic("hammer2_chain_lookup: unrecognized blockref type: %d",
3126 base = NULL; /* safety */
3127 count = 0; /* safety */
3131 * Merged scan to find next candidate.
3133 * hammer2_base_*() functions require the parent->core.live_* fields
3134 * to be synchronized.
3136 * We need to hold the spinlock to access the block array and RB tree
3137 * and to interlock chain creation.
3139 if ((parent->flags & HAMMER2_CHAIN_COUNTEDBREFS) == 0)
3140 hammer2_chain_countbrefs(parent, base, count);
3144 hammer2_spin_ex(&parent->core.spin);
3145 chain = hammer2_combined_find(parent, base, count,
3147 key, HAMMER2_KEY_MAX,
3149 generation = parent->core.generation;
3152 * Exhausted parent chain, we're done.
3154 if (bref_ptr == NULL) {
3155 hammer2_spin_unex(&parent->core.spin);
3156 KKASSERT(chain == NULL);
3157 error |= HAMMER2_ERROR_EOF;
3162 * Copy into the supplied stack-based blockref.
3167 * Selected from blockref or in-memory chain.
3169 if (chain == NULL) {
3170 switch(bref->type) {
3171 case HAMMER2_BREF_TYPE_INODE:
3172 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3173 case HAMMER2_BREF_TYPE_INDIRECT:
3174 case HAMMER2_BREF_TYPE_VOLUME:
3175 case HAMMER2_BREF_TYPE_FREEMAP:
3177 * Recursion, always get the chain
3179 hammer2_spin_unex(&parent->core.spin);
3180 chain = hammer2_chain_get(parent, generation,
3187 * No recursion, do not waste time instantiating
3188 * a chain, just iterate using the bref.
3190 hammer2_spin_unex(&parent->core.spin);
3195 * Recursion or not we need the chain in order to supply
3198 hammer2_chain_ref(chain);
3199 hammer2_spin_unex(&parent->core.spin);
3200 hammer2_chain_lock(chain, how);
3203 (bcmp(bref, &chain->bref, sizeof(*bref)) ||
3204 chain->parent != parent)) {
3205 hammer2_chain_unlock(chain);
3206 hammer2_chain_drop(chain);
3212 * Skip deleted chains (XXX cache 'i' end-of-block-array? XXX)
3214 * NOTE: chain's key range is not relevant as there might be
3215 * one-offs within the range that are not deleted.
3217 * NOTE: XXX this could create problems with scans used in
3218 * situations other than mount-time recovery.
3220 * NOTE: Lookups can race delete-duplicate because
3221 * delete-duplicate does not lock the parent's core
3222 * (they just use the spinlock on the core).
3224 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) {
3225 hammer2_chain_unlock(chain);
3226 hammer2_chain_drop(chain);
3231 error |= HAMMER2_ERROR_EOF;
3239 * All done, return the bref or NULL, supply chain if necessary.
3247 * Create and return a new hammer2 system memory structure of the specified
3248 * key, type and size and insert it under (*parentp). This is a full
3249 * insertion, based on the supplied key/keybits, and may involve creating
3250 * indirect blocks and moving other chains around via delete/duplicate.
3252 * This call can be made with parent == NULL as long as a non -1 methods
3253 * is supplied. hmp must also be supplied in this situation (otherwise
3254 * hmp is extracted from the supplied parent). The chain will be detached
3255 * from the topology. A later call with both parent and chain can be made
3258 * THE CALLER MUST HAVE ALREADY PROPERLY SEEKED (*parentp) TO THE INSERTION
3259 * POINT SANS ANY REQUIRED INDIRECT BLOCK CREATIONS DUE TO THE ARRAY BEING
3260 * FULL. This typically means that the caller is creating the chain after
3261 * doing a hammer2_chain_lookup().
3263 * (*parentp) must be exclusive locked and may be replaced on return
3264 * depending on how much work the function had to do.
3266 * (*parentp) must not be errored or this function will assert.
3268 * (*chainp) usually starts out NULL and returns the newly created chain,
3269 * but if the caller desires the caller may allocate a disconnected chain
3270 * and pass it in instead.
3272 * This function should NOT be used to insert INDIRECT blocks. It is
3273 * typically used to create/insert inodes and data blocks.
3275 * Caller must pass-in an exclusively locked parent the new chain is to
3276 * be inserted under, and optionally pass-in a disconnected, exclusively
3277 * locked chain to insert (else we create a new chain). The function will
3278 * adjust (*parentp) as necessary, create or connect the chain, and
3279 * return an exclusively locked chain in *chainp.
3281 * When creating a PFSROOT inode under the super-root, pmp is typically NULL
3282 * and will be reassigned.
3284 * NOTE: returns HAMMER_ERROR_* flags
3287 hammer2_chain_create(hammer2_chain_t **parentp, hammer2_chain_t **chainp,
3288 hammer2_dev_t *hmp, hammer2_pfs_t *pmp, int methods,
3289 hammer2_key_t key, int keybits, int type, size_t bytes,
3290 hammer2_tid_t mtid, hammer2_off_t dedup_off, int flags)
3292 hammer2_chain_t *chain;
3293 hammer2_chain_t *parent;
3294 hammer2_blockref_t *base;
3295 hammer2_blockref_t dummy;
3299 int maxloops = 300000;
3302 * Topology may be crossing a PFS boundary.
3306 KKASSERT(hammer2_mtx_owned(&parent->lock));
3307 KKASSERT(parent->error == 0);
3312 if (chain == NULL) {
3314 * First allocate media space and construct the dummy bref,
3315 * then allocate the in-memory chain structure. Set the
3316 * INITIAL flag for fresh chains which do not have embedded
3319 * XXX for now set the check mode of the child based on
3320 * the parent or, if the parent is an inode, the
3321 * specification in the inode.
3323 bzero(&dummy, sizeof(dummy));
3326 dummy.keybits = keybits;
3327 dummy.data_off = hammer2_getradix(bytes);
3330 * Inherit methods from parent by default. Primarily used
3331 * for BREF_TYPE_DATA. Non-data types *must* be set to
3332 * a non-NONE check algorithm.
3335 dummy.methods = parent->bref.methods;
3337 dummy.methods = (uint8_t)methods;
3339 if (type != HAMMER2_BREF_TYPE_DATA &&
3340 HAMMER2_DEC_CHECK(dummy.methods) == HAMMER2_CHECK_NONE) {
3342 HAMMER2_ENC_CHECK(HAMMER2_CHECK_DEFAULT);
3345 chain = hammer2_chain_alloc(hmp, pmp, &dummy);
3348 * Lock the chain manually, chain_lock will load the chain
3349 * which we do NOT want to do. (note: chain->refs is set
3350 * to 1 by chain_alloc() for us, but lockcnt is not).
3353 hammer2_mtx_ex(&chain->lock);
3355 ++curthread->td_tracker;
3358 * Set INITIAL to optimize I/O. The flag will generally be
3359 * processed when we call hammer2_chain_modify().
3361 * Recalculate bytes to reflect the actual media block
3362 * allocation. Handle special case radix 0 == 0 bytes.
3364 bytes = (size_t)(chain->bref.data_off & HAMMER2_OFF_MASK_RADIX);
3366 bytes = (hammer2_off_t)1 << bytes;
3367 chain->bytes = bytes;
3370 case HAMMER2_BREF_TYPE_VOLUME:
3371 case HAMMER2_BREF_TYPE_FREEMAP:
3372 panic("hammer2_chain_create: called with volume type");
3374 case HAMMER2_BREF_TYPE_INDIRECT:
3375 panic("hammer2_chain_create: cannot be used to"
3376 "create indirect block");
3378 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3379 panic("hammer2_chain_create: cannot be used to"
3380 "create freemap root or node");
3382 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
3383 KKASSERT(bytes == sizeof(chain->data->bmdata));
3385 case HAMMER2_BREF_TYPE_DIRENT:
3386 case HAMMER2_BREF_TYPE_INODE:
3387 case HAMMER2_BREF_TYPE_DATA:
3390 * leave chain->data NULL, set INITIAL
3392 KKASSERT(chain->data == NULL);
3393 atomic_set_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
3398 * We are reattaching a previously deleted chain, possibly
3399 * under a new parent and possibly with a new key/keybits.
3400 * The chain does not have to be in a modified state. The
3401 * UPDATE flag will be set later on in this routine.
3403 * Do NOT mess with the current state of the INITIAL flag.
3405 chain->bref.key = key;
3406 chain->bref.keybits = keybits;
3407 if (chain->flags & HAMMER2_CHAIN_DELETED)
3408 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_DELETED);
3409 KKASSERT(chain->parent == NULL);
3413 * Set the appropriate bref flag if requested.
3415 * NOTE! Callers can call this function to move chains without
3416 * knowing about special flags, so don't clear bref flags
3419 if (flags & HAMMER2_INSERT_PFSROOT)
3420 chain->bref.flags |= HAMMER2_BREF_FLAG_PFSROOT;
3426 * Calculate how many entries we have in the blockref array and
3427 * determine if an indirect block is required when inserting into
3431 if (--maxloops == 0)
3432 panic("hammer2_chain_create: maxloops");
3434 switch(parent->bref.type) {
3435 case HAMMER2_BREF_TYPE_INODE:
3436 if ((parent->data->ipdata.meta.op_flags &
3437 HAMMER2_OPFLAG_DIRECTDATA) != 0) {
3438 kprintf("hammer2: parent set for direct-data! "
3439 "pkey=%016jx ckey=%016jx\n",
3443 KKASSERT((parent->data->ipdata.meta.op_flags &
3444 HAMMER2_OPFLAG_DIRECTDATA) == 0);
3445 KKASSERT(parent->data != NULL);
3446 base = &parent->data->ipdata.u.blockset.blockref[0];
3447 count = HAMMER2_SET_COUNT;
3449 case HAMMER2_BREF_TYPE_INDIRECT:
3450 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3451 if (parent->flags & HAMMER2_CHAIN_INITIAL)
3454 base = &parent->data->npdata[0];
3455 count = parent->bytes / sizeof(hammer2_blockref_t);
3457 case HAMMER2_BREF_TYPE_VOLUME:
3458 KKASSERT(parent->data != NULL);
3459 base = &parent->data->voldata.sroot_blockset.blockref[0];
3460 count = HAMMER2_SET_COUNT;
3462 case HAMMER2_BREF_TYPE_FREEMAP:
3463 KKASSERT(parent->data != NULL);
3464 base = &parent->data->blkset.blockref[0];
3465 count = HAMMER2_SET_COUNT;
3468 panic("hammer2_chain_create: unrecognized blockref type: %d",
3476 * Make sure we've counted the brefs
3478 if ((parent->flags & HAMMER2_CHAIN_COUNTEDBREFS) == 0)
3479 hammer2_chain_countbrefs(parent, base, count);
3481 KASSERT(parent->core.live_count >= 0 &&
3482 parent->core.live_count <= count,
3483 ("bad live_count %d/%d (%02x, %d)",
3484 parent->core.live_count, count,
3485 parent->bref.type, parent->bytes));
3488 * If no free blockref could be found we must create an indirect
3489 * block and move a number of blockrefs into it. With the parent
3490 * locked we can safely lock each child in order to delete+duplicate
3491 * it without causing a deadlock.
3493 * This may return the new indirect block or the old parent depending
3494 * on where the key falls. NULL is returned on error.
3496 if (parent->core.live_count == count) {
3497 hammer2_chain_t *nparent;
3499 KKASSERT((flags & HAMMER2_INSERT_SAMEPARENT) == 0);
3501 nparent = hammer2_chain_create_indirect(parent, key, keybits,
3502 mtid, type, &error);
3503 if (nparent == NULL) {
3505 hammer2_chain_drop(chain);
3509 if (parent != nparent) {
3510 hammer2_chain_unlock(parent);
3511 hammer2_chain_drop(parent);
3512 parent = *parentp = nparent;
3518 * fall through if parent, or skip to here if no parent.
3521 if (chain->flags & HAMMER2_CHAIN_DELETED)
3522 kprintf("Inserting deleted chain @%016jx\n",
3526 * Link the chain into its parent.
3528 if (chain->parent != NULL)
3529 panic("hammer2: hammer2_chain_create: chain already connected");
3530 KKASSERT(chain->parent == NULL);
3532 KKASSERT(parent->core.live_count < count);
3533 hammer2_chain_insert(parent, chain,
3534 HAMMER2_CHAIN_INSERT_SPIN |
3535 HAMMER2_CHAIN_INSERT_LIVE,
3541 * Mark the newly created chain modified. This will cause
3542 * UPDATE to be set and process the INITIAL flag.
3544 * Device buffers are not instantiated for DATA elements
3545 * as these are handled by logical buffers.
3547 * Indirect and freemap node indirect blocks are handled
3548 * by hammer2_chain_create_indirect() and not by this
3551 * Data for all other bref types is expected to be
3552 * instantiated (INODE, LEAF).
3554 switch(chain->bref.type) {
3555 case HAMMER2_BREF_TYPE_DATA:
3556 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
3557 case HAMMER2_BREF_TYPE_DIRENT:
3558 case HAMMER2_BREF_TYPE_INODE:
3559 error = hammer2_chain_modify(chain, mtid, dedup_off,
3560 HAMMER2_MODIFY_OPTDATA);
3564 * Remaining types are not supported by this function.
3565 * In particular, INDIRECT and LEAF_NODE types are
3566 * handled by create_indirect().
3568 panic("hammer2_chain_create: bad type: %d",
3575 * When reconnecting a chain we must set UPDATE and
3576 * setflush so the flush recognizes that it must update
3577 * the bref in the parent.
3579 if ((chain->flags & HAMMER2_CHAIN_UPDATE) == 0)
3580 atomic_set_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
3584 * We must setflush(parent) to ensure that it recurses through to
3585 * chain. setflush(chain) might not work because ONFLUSH is possibly
3586 * already set in the chain (so it won't recurse up to set it in the
3590 hammer2_chain_setflush(parent);
3599 * Move the chain from its old parent to a new parent. The chain must have
3600 * already been deleted or already disconnected (or never associated) with
3601 * a parent. The chain is reassociated with the new parent and the deleted
3602 * flag will be cleared (no longer deleted). The chain's modification state
3605 * THE CALLER MUST HAVE ALREADY PROPERLY SEEKED (parent) TO THE INSERTION
3606 * POINT SANS ANY REQUIRED INDIRECT BLOCK CREATIONS DUE TO THE ARRAY BEING
3607 * FULL. This typically means that the caller is creating the chain after
3608 * doing a hammer2_chain_lookup().
3610 * Neither (parent) or (chain) can be errored.
3612 * If (parent) is non-NULL then the chain is inserted under the parent.
3614 * If (parent) is NULL then the newly duplicated chain is not inserted
3615 * anywhere, similar to if it had just been chain_alloc()'d (suitable for
3616 * passing into hammer2_chain_create() after this function returns).
3618 * WARNING! This function calls create which means it can insert indirect
3619 * blocks. This can cause other unrelated chains in the parent to
3620 * be moved to a newly inserted indirect block in addition to the
3624 hammer2_chain_rename(hammer2_chain_t **parentp, hammer2_chain_t *chain,
3625 hammer2_tid_t mtid, int flags)
3627 hammer2_blockref_t *bref;
3629 hammer2_chain_t *parent;
3633 * WARNING! We should never resolve DATA to device buffers
3634 * (XXX allow it if the caller did?), and since
3635 * we currently do not have the logical buffer cache
3636 * buffer in-hand to fix its cached physical offset
3637 * we also force the modify code to not COW it. XXX
3639 * NOTE! We allow error'd chains to be renamed. The bref itself
3640 * is good and can be renamed. The content, however, may
3644 KKASSERT(chain->parent == NULL);
3645 /*KKASSERT(chain->error == 0); allow */
3648 * Now create a duplicate of the chain structure, associating
3649 * it with the same core, making it the same size, pointing it
3650 * to the same bref (the same media block).
3652 * NOTE: Handle special radix == 0 case (means 0 bytes).
3654 bref = &chain->bref;
3655 bytes = (size_t)(bref->data_off & HAMMER2_OFF_MASK_RADIX);
3657 bytes = (hammer2_off_t)1 << bytes;
3660 * If parent is not NULL the duplicated chain will be entered under
3661 * the parent and the UPDATE bit set to tell flush to update
3664 * We must setflush(parent) to ensure that it recurses through to
3665 * chain. setflush(chain) might not work because ONFLUSH is possibly
3666 * already set in the chain (so it won't recurse up to set it in the
3669 * Having both chains locked is extremely important for atomicy.
3671 if (parentp && (parent = *parentp) != NULL) {
3672 KKASSERT(hammer2_mtx_owned(&parent->lock));
3673 KKASSERT(parent->refs > 0);
3674 KKASSERT(parent->error == 0);
3676 hammer2_chain_create(parentp, &chain, NULL, chain->pmp,
3677 HAMMER2_METH_DEFAULT,
3678 bref->key, bref->keybits, bref->type,
3679 chain->bytes, mtid, 0, flags);
3680 KKASSERT(chain->flags & HAMMER2_CHAIN_UPDATE);
3681 hammer2_chain_setflush(*parentp);
3686 * This works in tandem with delete_obref() to install a blockref in
3687 * (typically) an indirect block that is associated with the chain being
3688 * moved to *parentp.
3690 * The reason we need this function is that the caller needs to maintain
3691 * the blockref as it was, and not generate a new blockref for what might
3692 * be a modified chain. Otherwise stuff will leak into the flush that
3693 * the flush code's FLUSH_INODE_STOP flag is unable to catch.
3695 * It is EXTREMELY important that we properly set CHAIN_BMAPUPD and
3696 * CHAIN_UPDATE. We must set BMAPUPD if the bref does not match, and
3697 * we must clear CHAIN_UPDATE (that was likely set by the chain_rename) if
3698 * it does. Otherwise we can end up in a situation where H2 is unable to
3699 * clean up the in-memory chain topology.
3701 * The reason for this is that flushes do not generally flush through
3702 * BREF_TYPE_INODE chains and depend on a hammer2_inode_t queued to syncq
3703 * or sideq to properly flush and dispose of the related inode chain's flags.
3704 * Situations where the inode is not actually modified by the frontend,
3705 * but where we have to move the related chains around as we insert or cleanup
3706 * indirect blocks, can leave us with a 'dirty' (non-disposable) in-memory
3707 * inode chain that does not have a hammer2_inode_t associated with it.
3710 hammer2_chain_rename_obref(hammer2_chain_t **parentp, hammer2_chain_t *chain,
3711 hammer2_tid_t mtid, int flags,
3712 hammer2_blockref_t *obref)
3714 hammer2_chain_rename(parentp, chain, mtid, flags);
3717 hammer2_blockref_t *tbase;
3720 KKASSERT((chain->flags & HAMMER2_CHAIN_BMAPPED) == 0);
3721 hammer2_chain_modify(*parentp, mtid, 0, 0);
3722 tbase = hammer2_chain_base_and_count(*parentp, &tcount);
3723 hammer2_base_insert(*parentp, tbase, tcount, chain, obref);
3724 if (bcmp(obref, &chain->bref, sizeof(chain->bref))) {
3725 atomic_set_int(&chain->flags, HAMMER2_CHAIN_BMAPUPD |
3726 HAMMER2_CHAIN_UPDATE);
3728 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
3734 * Helper function for deleting chains.
3736 * The chain is removed from the live view (the RBTREE) as well as the parent's
3737 * blockmap. Both chain and its parent must be locked.
3739 * parent may not be errored. chain can be errored.
3742 _hammer2_chain_delete_helper(hammer2_chain_t *parent, hammer2_chain_t *chain,
3743 hammer2_tid_t mtid, int flags,
3744 hammer2_blockref_t *obref)
3749 KKASSERT((chain->flags & (HAMMER2_CHAIN_DELETED |
3750 HAMMER2_CHAIN_FICTITIOUS)) == 0);
3751 KKASSERT(chain->parent == parent);
3754 if (chain->flags & HAMMER2_CHAIN_BMAPPED) {
3756 * Chain is blockmapped, so there must be a parent.
3757 * Atomically remove the chain from the parent and remove
3758 * the blockmap entry. The parent must be set modified
3759 * to remove the blockmap entry.
3761 hammer2_blockref_t *base;
3764 KKASSERT(parent != NULL);
3765 KKASSERT(parent->error == 0);
3766 KKASSERT((parent->flags & HAMMER2_CHAIN_INITIAL) == 0);
3767 error = hammer2_chain_modify(parent, mtid, 0, 0);
3772 * Calculate blockmap pointer
3774 KKASSERT(chain->flags & HAMMER2_CHAIN_ONRBTREE);
3775 hammer2_spin_ex(&chain->core.spin);
3776 hammer2_spin_ex(&parent->core.spin);
3778 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DELETED);
3779 atomic_add_int(&parent->core.live_count, -1);
3780 ++parent->core.generation;
3781 RB_REMOVE(hammer2_chain_tree, &parent->core.rbtree, chain);
3782 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE);
3783 --parent->core.chain_count;
3784 chain->parent = NULL;
3786 switch(parent->bref.type) {
3787 case HAMMER2_BREF_TYPE_INODE:
3789 * Access the inode's block array. However, there
3790 * is no block array if the inode is flagged
3794 (parent->data->ipdata.meta.op_flags &
3795 HAMMER2_OPFLAG_DIRECTDATA) == 0) {
3797 &parent->data->ipdata.u.blockset.blockref[0];
3801 count = HAMMER2_SET_COUNT;
3803 case HAMMER2_BREF_TYPE_INDIRECT:
3804 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3806 base = &parent->data->npdata[0];
3809 count = parent->bytes / sizeof(hammer2_blockref_t);
3811 case HAMMER2_BREF_TYPE_VOLUME:
3812 base = &parent->data->voldata.
3813 sroot_blockset.blockref[0];
3814 count = HAMMER2_SET_COUNT;
3816 case HAMMER2_BREF_TYPE_FREEMAP:
3817 base = &parent->data->blkset.blockref[0];
3818 count = HAMMER2_SET_COUNT;
3823 panic("hammer2_flush_pass2: "
3824 "unrecognized blockref type: %d",
3829 * delete blockmapped chain from its parent.
3831 * The parent is not affected by any statistics in chain
3832 * which are pending synchronization. That is, there is
3833 * nothing to undo in the parent since they have not yet
3834 * been incorporated into the parent.
3836 * The parent is affected by statistics stored in inodes.
3837 * Those have already been synchronized, so they must be
3838 * undone. XXX split update possible w/delete in middle?
3841 hammer2_base_delete(parent, base, count, chain, obref);
3843 hammer2_spin_unex(&parent->core.spin);
3844 hammer2_spin_unex(&chain->core.spin);
3845 } else if (chain->flags & HAMMER2_CHAIN_ONRBTREE) {
3847 * Chain is not blockmapped but a parent is present.
3848 * Atomically remove the chain from the parent. There is
3849 * no blockmap entry to remove.
3851 * Because chain was associated with a parent but not
3852 * synchronized, the chain's *_count_up fields contain
3853 * inode adjustment statistics which must be undone.
3855 hammer2_spin_ex(&chain->core.spin);
3856 hammer2_spin_ex(&parent->core.spin);
3857 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DELETED);
3858 atomic_add_int(&parent->core.live_count, -1);
3859 ++parent->core.generation;
3860 RB_REMOVE(hammer2_chain_tree, &parent->core.rbtree, chain);
3861 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE);
3862 --parent->core.chain_count;
3863 chain->parent = NULL;
3864 hammer2_spin_unex(&parent->core.spin);
3865 hammer2_spin_unex(&chain->core.spin);
3868 * Chain is not blockmapped and has no parent. This
3869 * is a degenerate case.
3871 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DELETED);
3878 * Create an indirect block that covers one or more of the elements in the
3879 * current parent. Either returns the existing parent with no locking or
3880 * ref changes or returns the new indirect block locked and referenced
3881 * and leaving the original parent lock/ref intact as well.
3883 * If an error occurs, NULL is returned and *errorp is set to the H2 error.
3885 * The returned chain depends on where the specified key falls.
3887 * The key/keybits for the indirect mode only needs to follow three rules:
3889 * (1) That all elements underneath it fit within its key space and
3891 * (2) That all elements outside it are outside its key space.
3893 * (3) When creating the new indirect block any elements in the current
3894 * parent that fit within the new indirect block's keyspace must be
3895 * moved into the new indirect block.
3897 * (4) The keyspace chosen for the inserted indirect block CAN cover a wider
3898 * keyspace the the current parent, but lookup/iteration rules will
3899 * ensure (and must ensure) that rule (2) for all parents leading up
3900 * to the nearest inode or the root volume header is adhered to. This
3901 * is accomplished by always recursing through matching keyspaces in
3902 * the hammer2_chain_lookup() and hammer2_chain_next() API.
3904 * The current implementation calculates the current worst-case keyspace by
3905 * iterating the current parent and then divides it into two halves, choosing
3906 * whichever half has the most elements (not necessarily the half containing
3907 * the requested key).
3909 * We can also opt to use the half with the least number of elements. This
3910 * causes lower-numbered keys (aka logical file offsets) to recurse through
3911 * fewer indirect blocks and higher-numbered keys to recurse through more.
3912 * This also has the risk of not moving enough elements to the new indirect
3913 * block and being forced to create several indirect blocks before the element
3916 * Must be called with an exclusively locked parent.
3918 * NOTE: *errorp set to HAMMER_ERROR_* flags
3920 static int hammer2_chain_indkey_freemap(hammer2_chain_t *parent,
3921 hammer2_key_t *keyp, int keybits,
3922 hammer2_blockref_t *base, int count);
3923 static int hammer2_chain_indkey_file(hammer2_chain_t *parent,
3924 hammer2_key_t *keyp, int keybits,
3925 hammer2_blockref_t *base, int count,
3927 static int hammer2_chain_indkey_dir(hammer2_chain_t *parent,
3928 hammer2_key_t *keyp, int keybits,
3929 hammer2_blockref_t *base, int count,
3933 hammer2_chain_create_indirect(hammer2_chain_t *parent,
3934 hammer2_key_t create_key, int create_bits,
3935 hammer2_tid_t mtid, int for_type, int *errorp)
3938 hammer2_blockref_t *base;
3939 hammer2_blockref_t *bref;
3940 hammer2_blockref_t bcopy;
3941 hammer2_chain_t *chain;
3942 hammer2_chain_t *ichain;
3943 hammer2_chain_t dummy;
3944 hammer2_key_t key = create_key;
3945 hammer2_key_t key_beg;
3946 hammer2_key_t key_end;
3947 hammer2_key_t key_next;
3948 int keybits = create_bits;
3956 int maxloops = 300000;
3959 * Calculate the base blockref pointer or NULL if the chain
3960 * is known to be empty. We need to calculate the array count
3961 * for RB lookups either way.
3964 KKASSERT(hammer2_mtx_owned(&parent->lock));
3967 * Pre-modify the parent now to avoid having to deal with error
3968 * processing if we tried to later (in the middle of our loop).
3970 * We are going to be moving bref's around, the indirect blocks
3971 * cannot be in an initial state. Do not pass MODIFY_OPTDATA.
3973 *errorp = hammer2_chain_modify(parent, mtid, 0, 0);
3975 kprintf("hammer2_create_indirect: error %08x %s\n",
3976 *errorp, hammer2_error_str(*errorp));
3979 KKASSERT((parent->flags & HAMMER2_CHAIN_INITIAL) == 0);
3981 /*hammer2_chain_modify(&parent, HAMMER2_MODIFY_OPTDATA);*/
3982 base = hammer2_chain_base_and_count(parent, &count);
3985 * dummy used in later chain allocation (no longer used for lookups).
3987 bzero(&dummy, sizeof(dummy));
3990 * How big should our new indirect block be? It has to be at least
3991 * as large as its parent for splits to work properly.
3993 * The freemap uses a specific indirect block size. The number of
3994 * levels are built dynamically and ultimately depend on the size
3995 * volume. Because freemap blocks are taken from the reserved areas
3996 * of the volume our goal is efficiency (fewer levels) and not so
3997 * much to save disk space.
3999 * The first indirect block level for a directory usually uses
4000 * HAMMER2_IND_BYTES_MIN (4KB = 32 directory entries). Due to
4001 * the hash mechanism, this typically gives us a nominal
4002 * 32 * 4 entries with one level of indirection.
4004 * We use HAMMER2_IND_BYTES_NOM (16KB = 128 blockrefs) for FILE
4005 * indirect blocks. The initial 4 entries in the inode gives us
4006 * 256KB. Up to 4 indirect blocks gives us 32MB. Three levels
4007 * of indirection gives us 137GB, and so forth. H2 can support
4008 * huge file sizes but they are not typical, so we try to stick
4009 * with compactness and do not use a larger indirect block size.
4011 * We could use 64KB (PBUFSIZE), giving us 512 blockrefs, but
4012 * due to the way indirect blocks are created this usually winds
4013 * up being extremely inefficient for small files. Even though
4014 * 16KB requires more levels of indirection for very large files,
4015 * the 16KB records can be ganged together into 64KB DIOs.
4017 if (for_type == HAMMER2_BREF_TYPE_FREEMAP_NODE ||
4018 for_type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) {
4019 nbytes = HAMMER2_FREEMAP_LEVELN_PSIZE;
4020 } else if (parent->bref.type == HAMMER2_BREF_TYPE_INODE) {
4021 if (parent->data->ipdata.meta.type ==
4022 HAMMER2_OBJTYPE_DIRECTORY)
4023 nbytes = HAMMER2_IND_BYTES_MIN; /* 4KB = 32 entries */
4025 nbytes = HAMMER2_IND_BYTES_NOM; /* 16KB = ~8MB file */
4028 nbytes = HAMMER2_IND_BYTES_NOM;
4030 if (nbytes < count * sizeof(hammer2_blockref_t)) {
4031 KKASSERT(for_type != HAMMER2_BREF_TYPE_FREEMAP_NODE &&
4032 for_type != HAMMER2_BREF_TYPE_FREEMAP_LEAF);
4033 nbytes = count * sizeof(hammer2_blockref_t);
4035 ncount = nbytes / sizeof(hammer2_blockref_t);
4038 * When creating an indirect block for a freemap node or leaf
4039 * the key/keybits must be fitted to static radix levels because
4040 * particular radix levels use particular reserved blocks in the
4043 * This routine calculates the key/radix of the indirect block
4044 * we need to create, and whether it is on the high-side or the
4048 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
4049 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
4050 keybits = hammer2_chain_indkey_freemap(parent, &key, keybits,
4053 case HAMMER2_BREF_TYPE_DATA:
4054 keybits = hammer2_chain_indkey_file(parent, &key, keybits,
4055 base, count, ncount);
4057 case HAMMER2_BREF_TYPE_DIRENT:
4058 case HAMMER2_BREF_TYPE_INODE:
4059 keybits = hammer2_chain_indkey_dir(parent, &key, keybits,
4060 base, count, ncount);
4063 panic("illegal indirect block for bref type %d", for_type);
4068 * Normalize the key for the radix being represented, keeping the
4069 * high bits and throwing away the low bits.
4071 key &= ~(((hammer2_key_t)1 << keybits) - 1);
4074 * Ok, create our new indirect block
4076 if (for_type == HAMMER2_BREF_TYPE_FREEMAP_NODE ||
4077 for_type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) {
4078 dummy.bref.type = HAMMER2_BREF_TYPE_FREEMAP_NODE;
4080 dummy.bref.type = HAMMER2_BREF_TYPE_INDIRECT;
4082 dummy.bref.key = key;
4083 dummy.bref.keybits = keybits;
4084 dummy.bref.data_off = hammer2_getradix(nbytes);
4085 dummy.bref.methods =
4086 HAMMER2_ENC_CHECK(HAMMER2_DEC_CHECK(parent->bref.methods)) |
4087 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
4089 ichain = hammer2_chain_alloc(hmp, parent->pmp, &dummy.bref);
4090 atomic_set_int(&ichain->flags, HAMMER2_CHAIN_INITIAL);
4091 hammer2_chain_lock(ichain, HAMMER2_RESOLVE_MAYBE);
4092 /* ichain has one ref at this point */
4095 * We have to mark it modified to allocate its block, but use
4096 * OPTDATA to allow it to remain in the INITIAL state. Otherwise
4097 * it won't be acted upon by the flush code.
4099 * XXX remove OPTDATA, we need a fully initialized indirect block to
4100 * be able to move the original blockref.
4102 *errorp = hammer2_chain_modify(ichain, mtid, 0, 0);
4104 kprintf("hammer2_alloc_indirect: error %08x %s\n",
4105 *errorp, hammer2_error_str(*errorp));
4106 hammer2_chain_unlock(ichain);
4107 hammer2_chain_drop(ichain);
4110 KKASSERT((ichain->flags & HAMMER2_CHAIN_INITIAL) == 0);
4113 * Iterate the original parent and move the matching brefs into
4114 * the new indirect block.
4116 * XXX handle flushes.
4119 key_end = HAMMER2_KEY_MAX;
4120 key_next = 0; /* avoid gcc warnings */
4121 hammer2_spin_ex(&parent->core.spin);
4127 * Parent may have been modified, relocating its block array.
4128 * Reload the base pointer.
4130 base = hammer2_chain_base_and_count(parent, &count);
4132 if (++loops > 100000) {
4133 hammer2_spin_unex(&parent->core.spin);
4134 panic("excessive loops r=%d p=%p base/count %p:%d %016jx\n",
4135 reason, parent, base, count, key_next);
4139 * NOTE: spinlock stays intact, returned chain (if not NULL)
4140 * is not referenced or locked which means that we
4141 * cannot safely check its flagged / deletion status
4144 chain = hammer2_combined_find(parent, base, count,
4148 generation = parent->core.generation;
4151 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
4154 * Skip keys that are not within the key/radix of the new
4155 * indirect block. They stay in the parent.
4157 if ((~(((hammer2_key_t)1 << keybits) - 1) &
4158 (key ^ bref->key)) != 0) {
4159 goto next_key_spinlocked;
4163 * Load the new indirect block by acquiring the related
4164 * chains (potentially from media as it might not be
4165 * in-memory). Then move it to the new parent (ichain).
4167 * chain is referenced but not locked. We must lock the
4168 * chain to obtain definitive state.
4173 * Use chain already present in the RBTREE
4175 hammer2_chain_ref(chain);
4176 hammer2_spin_unex(&parent->core.spin);
4177 hammer2_chain_lock(chain, HAMMER2_RESOLVE_NEVER);
4180 * Get chain for blockref element. _get returns NULL
4181 * on insertion race.
4183 hammer2_spin_unex(&parent->core.spin);
4184 chain = hammer2_chain_get(parent, generation, &bcopy,
4185 HAMMER2_RESOLVE_NEVER);
4186 if (chain == NULL) {
4188 hammer2_spin_ex(&parent->core.spin);
4194 * This is always live so if the chain has been deleted
4195 * we raced someone and we have to retry.
4197 * NOTE: Lookups can race delete-duplicate because
4198 * delete-duplicate does not lock the parent's core
4199 * (they just use the spinlock on the core).
4201 * (note reversed logic for this one)
4203 if (bcmp(&bcopy, &chain->bref, sizeof(bcopy)) ||
4204 chain->parent != parent ||
4205 (chain->flags & HAMMER2_CHAIN_DELETED)) {
4206 hammer2_chain_unlock(chain);
4207 hammer2_chain_drop(chain);
4208 if (hammer2_debug & 0x0040) {
4209 kprintf("LOST PARENT RETRY "
4210 "RETRY (%p,%p)->%p %08x\n",
4211 parent, chain->parent, chain, chain->flags);
4213 hammer2_spin_ex(&parent->core.spin);
4218 * Shift the chain to the indirect block.
4220 * WARNING! No reason for us to load chain data, pass NOSTATS
4221 * to prevent delete/insert from trying to access
4222 * inode stats (and thus asserting if there is no
4223 * chain->data loaded).
4225 * WARNING! The (parent, chain) deletion may modify the parent
4226 * and invalidate the base pointer.
4228 * WARNING! Parent must already be marked modified, so we
4229 * can assume that chain_delete always suceeds.
4231 * WARNING! hammer2_chain_repchange() does not have to be
4232 * called (and doesn't work anyway because we are
4233 * only doing a partial shift). A recursion that is
4234 * in-progress can continue at the current parent
4235 * and will be able to properly find its next key.
4237 error = hammer2_chain_delete_obref(parent, chain, mtid, 0,
4239 KKASSERT(error == 0);
4240 hammer2_chain_rename_obref(&ichain, chain, mtid, 0, &bcopy);
4241 hammer2_chain_unlock(chain);
4242 hammer2_chain_drop(chain);
4243 KKASSERT(parent->refs > 0);
4245 base = NULL; /* safety */
4246 hammer2_spin_ex(&parent->core.spin);
4247 next_key_spinlocked:
4248 if (--maxloops == 0)
4249 panic("hammer2_chain_create_indirect: maxloops");
4251 if (key_next == 0 || key_next > key_end)
4256 hammer2_spin_unex(&parent->core.spin);
4259 * Insert the new indirect block into the parent now that we've
4260 * cleared out some entries in the parent. We calculated a good
4261 * insertion index in the loop above (ichain->index).
4263 * We don't have to set UPDATE here because we mark ichain
4264 * modified down below (so the normal modified -> flush -> set-moved
4265 * sequence applies).
4267 * The insertion shouldn't race as this is a completely new block
4268 * and the parent is locked.
4270 base = NULL; /* safety, parent modify may change address */
4271 KKASSERT((ichain->flags & HAMMER2_CHAIN_ONRBTREE) == 0);
4272 KKASSERT(parent->core.live_count < count);
4273 hammer2_chain_insert(parent, ichain,
4274 HAMMER2_CHAIN_INSERT_SPIN |
4275 HAMMER2_CHAIN_INSERT_LIVE,
4279 * Make sure flushes propogate after our manual insertion.
4281 hammer2_chain_setflush(ichain);
4282 hammer2_chain_setflush(parent);
4285 * Figure out what to return.
4287 if (~(((hammer2_key_t)1 << keybits) - 1) &
4288 (create_key ^ key)) {
4290 * Key being created is outside the key range,
4291 * return the original parent.
4293 hammer2_chain_unlock(ichain);
4294 hammer2_chain_drop(ichain);
4297 * Otherwise its in the range, return the new parent.
4298 * (leave both the new and old parent locked).
4307 * Do maintenance on an indirect chain. Both parent and chain are locked.
4309 * Returns non-zero if (chain) is deleted, either due to being empty or
4310 * because its children were safely moved into the parent.
4313 hammer2_chain_indirect_maintenance(hammer2_chain_t *parent,
4314 hammer2_chain_t *chain)
4316 hammer2_blockref_t *chain_base;
4317 hammer2_blockref_t *base;
4318 hammer2_blockref_t *bref;
4319 hammer2_blockref_t bcopy;
4320 hammer2_key_t key_next;
4321 hammer2_key_t key_beg;
4322 hammer2_key_t key_end;
4323 hammer2_chain_t *sub;
4330 * Make sure we have an accurate live_count
4332 if ((chain->flags & (HAMMER2_CHAIN_INITIAL |
4333 HAMMER2_CHAIN_COUNTEDBREFS)) == 0) {
4334 base = &chain->data->npdata[0];
4335 count = chain->bytes / sizeof(hammer2_blockref_t);
4336 hammer2_chain_countbrefs(chain, base, count);
4340 * If the indirect block is empty we can delete it.
4341 * (ignore deletion error)
4343 if (chain->core.live_count == 0 && RB_EMPTY(&chain->core.rbtree)) {
4344 hammer2_chain_delete(parent, chain,
4345 chain->bref.modify_tid,
4346 HAMMER2_DELETE_PERMANENT);
4347 hammer2_chain_repchange(parent, chain);
4351 base = hammer2_chain_base_and_count(parent, &count);
4353 if ((parent->flags & (HAMMER2_CHAIN_INITIAL |
4354 HAMMER2_CHAIN_COUNTEDBREFS)) == 0) {
4355 hammer2_chain_countbrefs(parent, base, count);
4359 * Determine if we can collapse chain into parent, calculate
4360 * hysteresis for chain emptiness.
4362 if (parent->core.live_count + chain->core.live_count - 1 > count)
4364 chain_count = chain->bytes / sizeof(hammer2_blockref_t);
4365 if (chain->core.live_count > chain_count * 3 / 4)
4369 * Ok, theoretically we can collapse chain's contents into
4370 * parent. chain is locked, but any in-memory children of chain
4371 * are not. For this to work, we must be able to dispose of any
4372 * in-memory children of chain.
4374 * For now require that there are no in-memory children of chain.
4376 * WARNING! Both chain and parent must remain locked across this
4381 * Parent must be marked modified. Don't try to collapse it if we
4382 * can't mark it modified. Once modified, destroy chain to make room
4383 * and to get rid of what will be a conflicting key (this is included
4384 * in the calculation above). Finally, move the children of chain
4385 * into chain's parent.
4387 * This order creates an accounting problem for bref.embed.stats
4388 * because we destroy chain before we remove its children. Any
4389 * elements whos blockref is already synchronized will be counted
4390 * twice. To deal with the problem we clean out chain's stats prior
4393 error = hammer2_chain_modify(parent, 0, 0, 0);
4395 krateprintf(&krate_h2me, "hammer2: indirect_maint: %s\n",
4396 hammer2_error_str(error));
4399 error = hammer2_chain_modify(chain, chain->bref.modify_tid, 0, 0);
4401 krateprintf(&krate_h2me, "hammer2: indirect_maint: %s\n",
4402 hammer2_error_str(error));
4406 chain->bref.embed.stats.inode_count = 0;
4407 chain->bref.embed.stats.data_count = 0;
4408 error = hammer2_chain_delete(parent, chain,
4409 chain->bref.modify_tid,
4410 HAMMER2_DELETE_PERMANENT);
4411 KKASSERT(error == 0);
4414 * The combined_find call requires core.spin to be held. One would
4415 * think there wouldn't be any conflicts since we hold chain
4416 * exclusively locked, but the caching mechanism for 0-ref children
4417 * does not require a chain lock.
4419 hammer2_spin_ex(&chain->core.spin);
4423 key_end = HAMMER2_KEY_MAX;
4425 chain_base = &chain->data->npdata[0];
4426 chain_count = chain->bytes / sizeof(hammer2_blockref_t);
4427 sub = hammer2_combined_find(chain, chain_base, chain_count,
4431 generation = chain->core.generation;
4434 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
4438 hammer2_chain_ref(sub);
4439 hammer2_spin_unex(&chain->core.spin);
4440 hammer2_chain_lock(sub, HAMMER2_RESOLVE_NEVER);
4442 hammer2_spin_unex(&chain->core.spin);
4443 sub = hammer2_chain_get(chain, generation, &bcopy,
4444 HAMMER2_RESOLVE_NEVER);
4446 hammer2_spin_ex(&chain->core.spin);
4450 if (bcmp(&bcopy, &sub->bref, sizeof(bcopy)) ||
4451 sub->parent != chain ||
4452 (sub->flags & HAMMER2_CHAIN_DELETED)) {
4453 hammer2_chain_unlock(sub);
4454 hammer2_chain_drop(sub);
4455 hammer2_spin_ex(&chain->core.spin);
4456 sub = NULL; /* safety */
4459 error = hammer2_chain_delete_obref(chain, sub,
4460 sub->bref.modify_tid, 0,
4462 KKASSERT(error == 0);
4463 hammer2_chain_rename_obref(&parent, sub,
4464 sub->bref.modify_tid,
4465 HAMMER2_INSERT_SAMEPARENT, &bcopy);
4466 hammer2_chain_unlock(sub);
4467 hammer2_chain_drop(sub);
4468 hammer2_spin_ex(&chain->core.spin);
4474 hammer2_spin_unex(&chain->core.spin);
4476 hammer2_chain_repchange(parent, chain);
4482 * Freemap indirect blocks
4484 * Calculate the keybits and highside/lowside of the freemap node the
4485 * caller is creating.
4487 * This routine will specify the next higher-level freemap key/radix
4488 * representing the lowest-ordered set. By doing so, eventually all
4489 * low-ordered sets will be moved one level down.
4491 * We have to be careful here because the freemap reserves a limited
4492 * number of blocks for a limited number of levels. So we can't just
4493 * push indiscriminately.
4496 hammer2_chain_indkey_freemap(hammer2_chain_t *parent, hammer2_key_t *keyp,
4497 int keybits, hammer2_blockref_t *base, int count)
4499 hammer2_chain_t *chain;
4500 hammer2_blockref_t *bref;
4502 hammer2_key_t key_beg;
4503 hammer2_key_t key_end;
4504 hammer2_key_t key_next;
4507 int maxloops = 300000;
4515 * Calculate the range of keys in the array being careful to skip
4516 * slots which are overridden with a deletion.
4519 key_end = HAMMER2_KEY_MAX;
4520 hammer2_spin_ex(&parent->core.spin);
4523 if (--maxloops == 0) {
4524 panic("indkey_freemap shit %p %p:%d\n",
4525 parent, base, count);
4527 chain = hammer2_combined_find(parent, base, count,
4539 * Skip deleted chains.
4541 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) {
4542 if (key_next == 0 || key_next > key_end)
4549 * Use the full live (not deleted) element for the scan
4550 * iteration. HAMMER2 does not allow partial replacements.
4552 * XXX should be built into hammer2_combined_find().
4554 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
4556 if (keybits > bref->keybits) {
4558 keybits = bref->keybits;
4559 } else if (keybits == bref->keybits && bref->key < key) {
4566 hammer2_spin_unex(&parent->core.spin);
4569 * Return the keybits for a higher-level FREEMAP_NODE covering
4573 case HAMMER2_FREEMAP_LEVEL0_RADIX:
4574 keybits = HAMMER2_FREEMAP_LEVEL1_RADIX;
4576 case HAMMER2_FREEMAP_LEVEL1_RADIX:
4577 keybits = HAMMER2_FREEMAP_LEVEL2_RADIX;
4579 case HAMMER2_FREEMAP_LEVEL2_RADIX:
4580 keybits = HAMMER2_FREEMAP_LEVEL3_RADIX;
4582 case HAMMER2_FREEMAP_LEVEL3_RADIX:
4583 keybits = HAMMER2_FREEMAP_LEVEL4_RADIX;
4585 case HAMMER2_FREEMAP_LEVEL4_RADIX:
4586 keybits = HAMMER2_FREEMAP_LEVEL5_RADIX;
4588 case HAMMER2_FREEMAP_LEVEL5_RADIX:
4589 panic("hammer2_chain_indkey_freemap: level too high");
4592 panic("hammer2_chain_indkey_freemap: bad radix");
4601 * File indirect blocks
4603 * Calculate the key/keybits for the indirect block to create by scanning
4604 * existing keys. The key being created is also passed in *keyp and can be
4605 * inside or outside the indirect block. Regardless, the indirect block
4606 * must hold at least two keys in order to guarantee sufficient space.
4608 * We use a modified version of the freemap's fixed radix tree, but taylored
4609 * for file data. Basically we configure an indirect block encompassing the
4613 hammer2_chain_indkey_file(hammer2_chain_t *parent, hammer2_key_t *keyp,
4614 int keybits, hammer2_blockref_t *base, int count,
4617 hammer2_chain_t *chain;
4618 hammer2_blockref_t *bref;
4620 hammer2_key_t key_beg;
4621 hammer2_key_t key_end;
4622 hammer2_key_t key_next;
4626 int maxloops = 300000;
4634 * Calculate the range of keys in the array being careful to skip
4635 * slots which are overridden with a deletion.
4637 * Locate the smallest key.
4640 key_end = HAMMER2_KEY_MAX;
4641 hammer2_spin_ex(&parent->core.spin);
4644 if (--maxloops == 0) {
4645 panic("indkey_freemap shit %p %p:%d\n",
4646 parent, base, count);
4648 chain = hammer2_combined_find(parent, base, count,
4660 * Skip deleted chains.
4662 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) {
4663 if (key_next == 0 || key_next > key_end)
4670 * Use the full live (not deleted) element for the scan
4671 * iteration. HAMMER2 does not allow partial replacements.
4673 * XXX should be built into hammer2_combined_find().
4675 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
4677 if (keybits > bref->keybits) {
4679 keybits = bref->keybits;
4680 } else if (keybits == bref->keybits && bref->key < key) {
4687 hammer2_spin_unex(&parent->core.spin);
4690 * Calculate the static keybits for a higher-level indirect block
4691 * that contains the key.
4696 case HAMMER2_IND_BYTES_MIN / sizeof(hammer2_blockref_t):
4697 nradix = HAMMER2_IND_RADIX_MIN - HAMMER2_BLOCKREF_RADIX;
4699 case HAMMER2_IND_BYTES_NOM / sizeof(hammer2_blockref_t):
4700 nradix = HAMMER2_IND_RADIX_NOM - HAMMER2_BLOCKREF_RADIX;
4702 case HAMMER2_IND_BYTES_MAX / sizeof(hammer2_blockref_t):
4703 nradix = HAMMER2_IND_RADIX_MAX - HAMMER2_BLOCKREF_RADIX;
4706 panic("bad ncount %d\n", ncount);
4712 * The largest radix that can be returned for an indirect block is
4713 * 63 bits. (The largest practical indirect block radix is actually
4714 * 62 bits because the top-level inode or volume root contains four
4715 * entries, but allow 63 to be returned).
4720 return keybits + nradix;
4726 * Directory indirect blocks.
4728 * Covers both the inode index (directory of inodes), and directory contents
4729 * (filenames hardlinked to inodes).
4731 * Because directory keys are hashed we generally try to cut the space in
4732 * half. We accomodate the inode index (which tends to have linearly
4733 * increasing inode numbers) by ensuring that the keyspace is at least large
4734 * enough to fill up the indirect block being created.
4737 hammer2_chain_indkey_dir(hammer2_chain_t *parent, hammer2_key_t *keyp,
4738 int keybits, hammer2_blockref_t *base, int count,
4741 hammer2_blockref_t *bref;
4742 hammer2_chain_t *chain;
4743 hammer2_key_t key_beg;
4744 hammer2_key_t key_end;
4745 hammer2_key_t key_next;
4750 int maxloops = 300000;
4753 * NOTE: We can't take a shortcut here anymore for inodes because
4754 * the root directory can contain a mix of inodes and directory
4755 * entries (we used to just return 63 if parent->bref.type was
4756 * HAMMER2_BREF_TYPE_INODE.
4763 * Calculate the range of keys in the array being careful to skip
4764 * slots which are overridden with a deletion.
4767 key_end = HAMMER2_KEY_MAX;
4768 hammer2_spin_ex(&parent->core.spin);
4771 if (--maxloops == 0) {
4772 panic("indkey_freemap shit %p %p:%d\n",
4773 parent, base, count);
4775 chain = hammer2_combined_find(parent, base, count,
4789 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) {
4790 if (key_next == 0 || key_next > key_end)
4797 * Use the full live (not deleted) element for the scan
4798 * iteration. HAMMER2 does not allow partial replacements.
4800 * XXX should be built into hammer2_combined_find().
4802 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
4805 * Expand our calculated key range (key, keybits) to fit
4806 * the scanned key. nkeybits represents the full range
4807 * that we will later cut in half (two halves @ nkeybits - 1).
4810 if (nkeybits < bref->keybits) {
4811 if (bref->keybits > 64) {
4812 kprintf("bad bref chain %p bref %p\n",
4816 nkeybits = bref->keybits;
4818 while (nkeybits < 64 &&
4819 (~(((hammer2_key_t)1 << nkeybits) - 1) &
4820 (key ^ bref->key)) != 0) {
4825 * If the new key range is larger we have to determine
4826 * which side of the new key range the existing keys fall
4827 * under by checking the high bit, then collapsing the
4828 * locount into the hicount or vise-versa.
4830 if (keybits != nkeybits) {
4831 if (((hammer2_key_t)1 << (nkeybits - 1)) & key) {
4842 * The newly scanned key will be in the lower half or the
4843 * upper half of the (new) key range.
4845 if (((hammer2_key_t)1 << (nkeybits - 1)) & bref->key)
4854 hammer2_spin_unex(&parent->core.spin);
4855 bref = NULL; /* now invalid (safety) */
4858 * Adjust keybits to represent half of the full range calculated
4859 * above (radix 63 max) for our new indirect block.
4864 * Expand keybits to hold at least ncount elements. ncount will be
4865 * a power of 2. This is to try to completely fill leaf nodes (at
4866 * least for keys which are not hashes).
4868 * We aren't counting 'in' or 'out', we are counting 'high side'
4869 * and 'low side' based on the bit at (1LL << keybits). We want
4870 * everything to be inside in these cases so shift it all to
4871 * the low or high side depending on the new high bit.
4873 while (((hammer2_key_t)1 << keybits) < ncount) {
4875 if (key & ((hammer2_key_t)1 << keybits)) {
4884 if (hicount > locount)
4885 key |= (hammer2_key_t)1 << keybits;
4887 key &= ~(hammer2_key_t)1 << keybits;
4897 * Directory indirect blocks.
4899 * Covers both the inode index (directory of inodes), and directory contents
4900 * (filenames hardlinked to inodes).
4902 * Because directory keys are hashed we generally try to cut the space in
4903 * half. We accomodate the inode index (which tends to have linearly
4904 * increasing inode numbers) by ensuring that the keyspace is at least large
4905 * enough to fill up the indirect block being created.
4908 hammer2_chain_indkey_dir(hammer2_chain_t *parent, hammer2_key_t *keyp,
4909 int keybits, hammer2_blockref_t *base, int count,
4912 hammer2_blockref_t *bref;
4913 hammer2_chain_t *chain;
4914 hammer2_key_t key_beg;
4915 hammer2_key_t key_end;
4916 hammer2_key_t key_next;
4921 int maxloops = 300000;
4924 * Shortcut if the parent is the inode. In this situation the
4925 * parent has 4+1 directory entries and we are creating an indirect
4926 * block capable of holding many more.
4928 if (parent->bref.type == HAMMER2_BREF_TYPE_INODE) {
4937 * Calculate the range of keys in the array being careful to skip
4938 * slots which are overridden with a deletion.
4941 key_end = HAMMER2_KEY_MAX;
4942 hammer2_spin_ex(&parent->core.spin);
4945 if (--maxloops == 0) {
4946 panic("indkey_freemap shit %p %p:%d\n",
4947 parent, base, count);
4949 chain = hammer2_combined_find(parent, base, count,
4963 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) {
4964 if (key_next == 0 || key_next > key_end)
4971 * Use the full live (not deleted) element for the scan
4972 * iteration. HAMMER2 does not allow partial replacements.
4974 * XXX should be built into hammer2_combined_find().
4976 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
4979 * Expand our calculated key range (key, keybits) to fit
4980 * the scanned key. nkeybits represents the full range
4981 * that we will later cut in half (two halves @ nkeybits - 1).
4984 if (nkeybits < bref->keybits) {
4985 if (bref->keybits > 64) {
4986 kprintf("bad bref chain %p bref %p\n",
4990 nkeybits = bref->keybits;
4992 while (nkeybits < 64 &&
4993 (~(((hammer2_key_t)1 << nkeybits) - 1) &
4994 (key ^ bref->key)) != 0) {
4999 * If the new key range is larger we have to determine
5000 * which side of the new key range the existing keys fall
5001 * under by checking the high bit, then collapsing the
5002 * locount into the hicount or vise-versa.
5004 if (keybits != nkeybits) {
5005 if (((hammer2_key_t)1 << (nkeybits - 1)) & key) {
5016 * The newly scanned key will be in the lower half or the
5017 * upper half of the (new) key range.
5019 if (((hammer2_key_t)1 << (nkeybits - 1)) & bref->key)
5028 hammer2_spin_unex(&parent->core.spin);
5029 bref = NULL; /* now invalid (safety) */
5032 * Adjust keybits to represent half of the full range calculated
5033 * above (radix 63 max) for our new indirect block.
5038 * Expand keybits to hold at least ncount elements. ncount will be
5039 * a power of 2. This is to try to completely fill leaf nodes (at
5040 * least for keys which are not hashes).
5042 * We aren't counting 'in' or 'out', we are counting 'high side'
5043 * and 'low side' based on the bit at (1LL << keybits). We want
5044 * everything to be inside in these cases so shift it all to
5045 * the low or high side depending on the new high bit.
5047 while (((hammer2_key_t)1 << keybits) < ncount) {
5049 if (key & ((hammer2_key_t)1 << keybits)) {
5058 if (hicount > locount)
5059 key |= (hammer2_key_t)1 << keybits;
5061 key &= ~(hammer2_key_t)1 << keybits;
5071 * Sets CHAIN_DELETED and remove the chain's blockref from the parent if
5074 * Both parent and chain must be locked exclusively.
5076 * This function will modify the parent if the blockref requires removal
5077 * from the parent's block table.
5079 * This function is NOT recursive. Any entity already pushed into the
5080 * chain (such as an inode) may still need visibility into its contents,
5081 * as well as the ability to read and modify the contents. For example,
5082 * for an unlinked file which is still open.
5084 * Also note that the flusher is responsible for cleaning up empty
5088 hammer2_chain_delete(hammer2_chain_t *parent, hammer2_chain_t *chain,
5089 hammer2_tid_t mtid, int flags)
5093 KKASSERT(hammer2_mtx_owned(&chain->lock));
5096 * Nothing to do if already marked.
5098 * We need the spinlock on the core whos RBTREE contains chain
5099 * to protect against races.
5101 if ((chain->flags & HAMMER2_CHAIN_DELETED) == 0) {
5102 KKASSERT((chain->flags & HAMMER2_CHAIN_DELETED) == 0 &&
5103 chain->parent == parent);
5104 error = _hammer2_chain_delete_helper(parent, chain,
5109 * Permanent deletions mark the chain as destroyed.
5111 * NOTE: We do not setflush the chain unless the deletion is
5112 * permanent, since the deletion of a chain does not actually
5113 * require it to be flushed.
5116 if (flags & HAMMER2_DELETE_PERMANENT) {
5117 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DESTROY);
5118 hammer2_chain_setflush(chain);
5126 hammer2_chain_delete_obref(hammer2_chain_t *parent, hammer2_chain_t *chain,
5127 hammer2_tid_t mtid, int flags,
5128 hammer2_blockref_t *obref)
5132 KKASSERT(hammer2_mtx_owned(&chain->lock));
5135 * Nothing to do if already marked.
5137 * We need the spinlock on the core whos RBTREE contains chain
5138 * to protect against races.
5141 if ((chain->flags & HAMMER2_CHAIN_DELETED) == 0) {
5142 KKASSERT((chain->flags & HAMMER2_CHAIN_DELETED) == 0 &&
5143 chain->parent == parent);
5144 error = _hammer2_chain_delete_helper(parent, chain,
5145 mtid, flags, obref);
5149 * Permanent deletions mark the chain as destroyed.
5151 * NOTE: We do not setflush the chain unless the deletion is
5152 * permanent, since the deletion of a chain does not actually
5153 * require it to be flushed.
5156 if (flags & HAMMER2_DELETE_PERMANENT) {
5157 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DESTROY);
5158 hammer2_chain_setflush(chain);
5166 * Returns the index of the nearest element in the blockref array >= elm.
5167 * Returns (count) if no element could be found.
5169 * Sets *key_nextp to the next key for loop purposes but does not modify
5170 * it if the next key would be higher than the current value of *key_nextp.
5171 * Note that *key_nexp can overflow to 0, which should be tested by the
5174 * WARNING! Must be called with parent's spinlock held. Spinlock remains
5175 * held through the operation.
5178 hammer2_base_find(hammer2_chain_t *parent,
5179 hammer2_blockref_t *base, int count,
5180 hammer2_key_t *key_nextp,
5181 hammer2_key_t key_beg, hammer2_key_t key_end)
5183 hammer2_blockref_t *scan;
5184 hammer2_key_t scan_end;
5189 * Require the live chain's already have their core's counted
5190 * so we can optimize operations.
5192 KKASSERT(parent->flags & HAMMER2_CHAIN_COUNTEDBREFS);
5197 if (count == 0 || base == NULL)
5201 * Sequential optimization using parent->cache_index. This is
5202 * the most likely scenario.
5204 * We can avoid trailing empty entries on live chains, otherwise
5205 * we might have to check the whole block array.
5207 i = parent->cache_index; /* SMP RACE OK */
5209 limit = parent->core.live_zero;
5214 KKASSERT(i < count);
5220 while (i > 0 && (scan->type == 0 || scan->key > key_beg)) {
5224 parent->cache_index = i;
5227 * Search forwards, stop when we find a scan element which
5228 * encloses the key or until we know that there are no further
5232 if (scan->type != 0) {
5233 scan_end = scan->key +
5234 ((hammer2_key_t)1 << scan->keybits) - 1;
5235 if (scan->key > key_beg || scan_end >= key_beg)
5244 parent->cache_index = i;
5248 scan_end = scan->key +
5249 ((hammer2_key_t)1 << scan->keybits);
5250 if (scan_end && (*key_nextp > scan_end ||
5252 *key_nextp = scan_end;
5260 * Do a combined search and return the next match either from the blockref
5261 * array or from the in-memory chain. Sets *bresp to the returned bref in
5262 * both cases, or sets it to NULL if the search exhausted. Only returns
5263 * a non-NULL chain if the search matched from the in-memory chain.
5265 * When no in-memory chain has been found and a non-NULL bref is returned
5269 * The returned chain is not locked or referenced. Use the returned bref
5270 * to determine if the search exhausted or not. Iterate if the base find
5271 * is chosen but matches a deleted chain.
5273 * WARNING! Must be called with parent's spinlock held. Spinlock remains
5274 * held through the operation.
5277 hammer2_combined_find(hammer2_chain_t *parent,
5278 hammer2_blockref_t *base, int count,
5279 hammer2_key_t *key_nextp,
5280 hammer2_key_t key_beg, hammer2_key_t key_end,
5281 hammer2_blockref_t **bresp)
5283 hammer2_blockref_t *bref;
5284 hammer2_chain_t *chain;
5288 * Lookup in block array and in rbtree.
5290 *key_nextp = key_end + 1;
5291 i = hammer2_base_find(parent, base, count, key_nextp,
5293 chain = hammer2_chain_find(parent, key_nextp, key_beg, key_end);
5298 if (i == count && chain == NULL) {
5304 * Only chain matched.
5307 bref = &chain->bref;
5312 * Only blockref matched.
5314 if (chain == NULL) {
5320 * Both in-memory and blockref matched, select the nearer element.
5322 * If both are flush with the left-hand side or both are the
5323 * same distance away, select the chain. In this situation the
5324 * chain must have been loaded from the matching blockmap.
5326 if ((chain->bref.key <= key_beg && base[i].key <= key_beg) ||
5327 chain->bref.key == base[i].key) {
5328 KKASSERT(chain->bref.key == base[i].key);
5329 bref = &chain->bref;
5334 * Select the nearer key
5336 if (chain->bref.key < base[i].key) {
5337 bref = &chain->bref;
5344 * If the bref is out of bounds we've exhausted our search.
5347 if (bref->key > key_end) {
5357 * Locate the specified block array element and delete it. The element
5360 * The spin lock on the related chain must be held.
5362 * NOTE: live_count was adjusted when the chain was deleted, so it does not
5363 * need to be adjusted when we commit the media change.
5366 hammer2_base_delete(hammer2_chain_t *parent,
5367 hammer2_blockref_t *base, int count,
5368 hammer2_chain_t *chain,
5369 hammer2_blockref_t *obref)
5371 hammer2_blockref_t *elm = &chain->bref;
5372 hammer2_blockref_t *scan;
5373 hammer2_key_t key_next;
5377 * Delete element. Expect the element to exist.
5379 * XXX see caller, flush code not yet sophisticated enough to prevent
5380 * re-flushed in some cases.
5382 key_next = 0; /* max range */
5383 i = hammer2_base_find(parent, base, count, &key_next,
5384 elm->key, elm->key);
5386 if (i == count || scan->type == 0 ||
5387 scan->key != elm->key ||
5388 ((chain->flags & HAMMER2_CHAIN_BMAPUPD) == 0 &&
5389 scan->keybits != elm->keybits)) {
5390 hammer2_spin_unex(&parent->core.spin);
5391 panic("delete base %p element not found at %d/%d elm %p\n",
5392 base, i, count, elm);
5397 * Update stats and zero the entry.
5399 * NOTE: Handle radix == 0 (0 bytes) case.
5401 if ((int)(scan->data_off & HAMMER2_OFF_MASK_RADIX)) {
5402 parent->bref.embed.stats.data_count -= (hammer2_off_t)1 <<
5403 (int)(scan->data_off & HAMMER2_OFF_MASK_RADIX);
5405 switch(scan->type) {
5406 case HAMMER2_BREF_TYPE_INODE:
5407 --parent->bref.embed.stats.inode_count;
5409 case HAMMER2_BREF_TYPE_DATA:
5410 if (parent->bref.leaf_count == HAMMER2_BLOCKREF_LEAF_MAX) {
5411 atomic_set_int(&chain->flags,
5412 HAMMER2_CHAIN_HINT_LEAF_COUNT);
5414 if (parent->bref.leaf_count)
5415 --parent->bref.leaf_count;
5418 case HAMMER2_BREF_TYPE_INDIRECT:
5419 if (scan->type != HAMMER2_BREF_TYPE_DATA) {
5420 parent->bref.embed.stats.data_count -=
5421 scan->embed.stats.data_count;
5422 parent->bref.embed.stats.inode_count -=
5423 scan->embed.stats.inode_count;
5425 if (scan->type == HAMMER2_BREF_TYPE_INODE)
5427 if (parent->bref.leaf_count == HAMMER2_BLOCKREF_LEAF_MAX) {
5428 atomic_set_int(&chain->flags,
5429 HAMMER2_CHAIN_HINT_LEAF_COUNT);
5431 if (parent->bref.leaf_count <= scan->leaf_count)
5432 parent->bref.leaf_count = 0;
5434 parent->bref.leaf_count -= scan->leaf_count;
5437 case HAMMER2_BREF_TYPE_DIRENT:
5438 if (parent->bref.leaf_count == HAMMER2_BLOCKREF_LEAF_MAX) {
5439 atomic_set_int(&chain->flags,
5440 HAMMER2_CHAIN_HINT_LEAF_COUNT);
5442 if (parent->bref.leaf_count)
5443 --parent->bref.leaf_count;
5451 bzero(scan, sizeof(*scan));
5454 * We can only optimize parent->core.live_zero for live chains.
5456 if (parent->core.live_zero == i + 1) {
5457 while (--i >= 0 && base[i].type == 0)
5459 parent->core.live_zero = i + 1;
5463 * Clear appropriate blockmap flags in chain.
5465 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_BMAPPED |
5466 HAMMER2_CHAIN_BMAPUPD);
5470 * Insert the specified element. The block array must not already have the
5471 * element and must have space available for the insertion.
5473 * The spin lock on the related chain must be held.
5475 * NOTE: live_count was adjusted when the chain was deleted, so it does not
5476 * need to be adjusted when we commit the media change.
5479 hammer2_base_insert(hammer2_chain_t *parent,
5480 hammer2_blockref_t *base, int count,
5481 hammer2_chain_t *chain, hammer2_blockref_t *elm)
5483 hammer2_key_t key_next;
5492 * Insert new element. Expect the element to not already exist
5493 * unless we are replacing it.
5495 * XXX see caller, flush code not yet sophisticated enough to prevent
5496 * re-flushed in some cases.
5498 key_next = 0; /* max range */
5499 i = hammer2_base_find(parent, base, count, &key_next,
5500 elm->key, elm->key);
5503 * Shortcut fill optimization, typical ordered insertion(s) may not
5506 KKASSERT(i >= 0 && i <= count);
5509 * Set appropriate blockmap flags in chain (if not NULL)
5512 atomic_set_int(&chain->flags, HAMMER2_CHAIN_BMAPPED);
5515 * Update stats and zero the entry
5517 if ((int)(elm->data_off & HAMMER2_OFF_MASK_RADIX)) {
5518 parent->bref.embed.stats.data_count += (hammer2_off_t)1 <<
5519 (int)(elm->data_off & HAMMER2_OFF_MASK_RADIX);
5522 case HAMMER2_BREF_TYPE_INODE:
5523 ++parent->bref.embed.stats.inode_count;
5525 case HAMMER2_BREF_TYPE_DATA:
5526 if (parent->bref.leaf_count != HAMMER2_BLOCKREF_LEAF_MAX)
5527 ++parent->bref.leaf_count;
5529 case HAMMER2_BREF_TYPE_INDIRECT:
5530 if (elm->type != HAMMER2_BREF_TYPE_DATA) {
5531 parent->bref.embed.stats.data_count +=
5532 elm->embed.stats.data_count;
5533 parent->bref.embed.stats.inode_count +=
5534 elm->embed.stats.inode_count;
5536 if (elm->type == HAMMER2_BREF_TYPE_INODE)
5538 if (parent->bref.leaf_count + elm->leaf_count <
5539 HAMMER2_BLOCKREF_LEAF_MAX) {
5540 parent->bref.leaf_count += elm->leaf_count;
5542 parent->bref.leaf_count = HAMMER2_BLOCKREF_LEAF_MAX;
5545 case HAMMER2_BREF_TYPE_DIRENT:
5546 if (parent->bref.leaf_count != HAMMER2_BLOCKREF_LEAF_MAX)
5547 ++parent->bref.leaf_count;
5555 * We can only optimize parent->core.live_zero for live chains.
5557 if (i == count && parent->core.live_zero < count) {
5558 i = parent->core.live_zero++;
5563 xkey = elm->key + ((hammer2_key_t)1 << elm->keybits) - 1;
5564 if (i != count && (base[i].key < elm->key || xkey >= base[i].key)) {
5565 hammer2_spin_unex(&parent->core.spin);
5566 panic("insert base %p overlapping elements at %d elm %p\n",
5571 * Try to find an empty slot before or after.
5575 while (j > 0 || k < count) {
5577 if (j >= 0 && base[j].type == 0) {
5581 bcopy(&base[j+1], &base[j],
5582 (i - j - 1) * sizeof(*base));
5588 if (k < count && base[k].type == 0) {
5589 bcopy(&base[i], &base[i+1],
5590 (k - i) * sizeof(hammer2_blockref_t));
5594 * We can only update parent->core.live_zero for live
5597 if (parent->core.live_zero <= k)
5598 parent->core.live_zero = k + 1;
5603 panic("hammer2_base_insert: no room!");
5610 for (l = 0; l < count; ++l) {
5612 key_next = base[l].key +
5613 ((hammer2_key_t)1 << base[l].keybits) - 1;
5617 while (++l < count) {
5619 if (base[l].key <= key_next)
5620 panic("base_insert %d %d,%d,%d fail %p:%d", u, i, j, k, base, l);
5621 key_next = base[l].key +
5622 ((hammer2_key_t)1 << base[l].keybits) - 1;
5632 * Sort the blockref array for the chain. Used by the flush code to
5633 * sort the blockref[] array.
5635 * The chain must be exclusively locked AND spin-locked.
5637 typedef hammer2_blockref_t *hammer2_blockref_p;
5641 hammer2_base_sort_callback(const void *v1, const void *v2)
5643 hammer2_blockref_p bref1 = *(const hammer2_blockref_p *)v1;
5644 hammer2_blockref_p bref2 = *(const hammer2_blockref_p *)v2;
5647 * Make sure empty elements are placed at the end of the array
5649 if (bref1->type == 0) {
5650 if (bref2->type == 0)
5653 } else if (bref2->type == 0) {
5660 if (bref1->key < bref2->key)
5662 if (bref1->key > bref2->key)
5668 hammer2_base_sort(hammer2_chain_t *chain)
5670 hammer2_blockref_t *base;
5673 switch(chain->bref.type) {
5674 case HAMMER2_BREF_TYPE_INODE:
5676 * Special shortcut for embedded data returns the inode
5677 * itself. Callers must detect this condition and access
5678 * the embedded data (the strategy code does this for us).
5680 * This is only applicable to regular files and softlinks.
5682 if (chain->data->ipdata.meta.op_flags &
5683 HAMMER2_OPFLAG_DIRECTDATA) {
5686 base = &chain->data->ipdata.u.blockset.blockref[0];
5687 count = HAMMER2_SET_COUNT;
5689 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
5690 case HAMMER2_BREF_TYPE_INDIRECT:
5692 * Optimize indirect blocks in the INITIAL state to avoid
5695 KKASSERT((chain->flags & HAMMER2_CHAIN_INITIAL) == 0);
5696 base = &chain->data->npdata[0];
5697 count = chain->bytes / sizeof(hammer2_blockref_t);
5699 case HAMMER2_BREF_TYPE_VOLUME:
5700 base = &chain->data->voldata.sroot_blockset.blockref[0];
5701 count = HAMMER2_SET_COUNT;
5703 case HAMMER2_BREF_TYPE_FREEMAP:
5704 base = &chain->data->blkset.blockref[0];
5705 count = HAMMER2_SET_COUNT;
5708 kprintf("hammer2_chain_lookup: unrecognized "
5709 "blockref(A) type: %d",
5712 tsleep(&base, 0, "dead", 0);
5713 panic("hammer2_chain_lookup: unrecognized "
5714 "blockref(A) type: %d",
5716 base = NULL; /* safety */
5717 count = 0; /* safety */
5719 kqsort(base, count, sizeof(*base), hammer2_base_sort_callback);
5725 * Chain memory management
5728 hammer2_chain_wait(hammer2_chain_t *chain)
5730 tsleep(chain, 0, "chnflw", 1);
5733 const hammer2_media_data_t *
5734 hammer2_chain_rdata(hammer2_chain_t *chain)
5736 KKASSERT(chain->data != NULL);
5737 return (chain->data);
5740 hammer2_media_data_t *
5741 hammer2_chain_wdata(hammer2_chain_t *chain)
5743 KKASSERT(chain->data != NULL);
5744 return (chain->data);
5748 * Set the check data for a chain. This can be a heavy-weight operation
5749 * and typically only runs on-flush. For file data check data is calculated
5750 * when the logical buffers are flushed.
5753 hammer2_chain_setcheck(hammer2_chain_t *chain, void *bdata)
5755 chain->bref.flags &= ~HAMMER2_BREF_FLAG_ZERO;
5757 switch(HAMMER2_DEC_CHECK(chain->bref.methods)) {
5758 case HAMMER2_CHECK_NONE:
5760 case HAMMER2_CHECK_DISABLED:
5762 case HAMMER2_CHECK_ISCSI32:
5763 chain->bref.check.iscsi32.value =
5764 hammer2_icrc32(bdata, chain->bytes);
5766 case HAMMER2_CHECK_XXHASH64:
5767 chain->bref.check.xxhash64.value =
5768 XXH64(bdata, chain->bytes, XXH_HAMMER2_SEED);
5770 case HAMMER2_CHECK_SHA192:
5772 SHA256_CTX hash_ctx;
5774 uint8_t digest[SHA256_DIGEST_LENGTH];
5775 uint64_t digest64[SHA256_DIGEST_LENGTH/8];
5778 SHA256_Init(&hash_ctx);
5779 SHA256_Update(&hash_ctx, bdata, chain->bytes);
5780 SHA256_Final(u.digest, &hash_ctx);
5781 u.digest64[2] ^= u.digest64[3];
5783 chain->bref.check.sha192.data,
5784 sizeof(chain->bref.check.sha192.data));
5787 case HAMMER2_CHECK_FREEMAP:
5788 chain->bref.check.freemap.icrc32 =
5789 hammer2_icrc32(bdata, chain->bytes);
5792 kprintf("hammer2_chain_setcheck: unknown check type %02x\n",
5793 chain->bref.methods);
5799 hammer2_chain_testcheck(hammer2_chain_t *chain, void *bdata)
5805 if (chain->bref.flags & HAMMER2_BREF_FLAG_ZERO)
5808 switch(HAMMER2_DEC_CHECK(chain->bref.methods)) {
5809 case HAMMER2_CHECK_NONE:
5812 case HAMMER2_CHECK_DISABLED:
5815 case HAMMER2_CHECK_ISCSI32:
5816 check32 = hammer2_icrc32(bdata, chain->bytes);
5817 r = (chain->bref.check.iscsi32.value == check32);
5819 kprintf("chain %016jx.%02x meth=%02x CHECK FAIL "
5820 "(flags=%08x, bref/data %08x/%08x)\n",
5821 chain->bref.data_off,
5823 chain->bref.methods,
5825 chain->bref.check.iscsi32.value,
5828 hammer2_process_icrc32 += chain->bytes;
5830 case HAMMER2_CHECK_XXHASH64:
5831 check64 = XXH64(bdata, chain->bytes, XXH_HAMMER2_SEED);
5832 r = (chain->bref.check.xxhash64.value == check64);
5834 kprintf("chain %016jx.%02x key=%016jx "
5835 "meth=%02x CHECK FAIL "
5836 "(flags=%08x, bref/data %016jx/%016jx)\n",
5837 chain->bref.data_off,
5840 chain->bref.methods,
5842 chain->bref.check.xxhash64.value,
5845 hammer2_process_xxhash64 += chain->bytes;
5847 case HAMMER2_CHECK_SHA192:
5849 SHA256_CTX hash_ctx;
5851 uint8_t digest[SHA256_DIGEST_LENGTH];
5852 uint64_t digest64[SHA256_DIGEST_LENGTH/8];
5855 SHA256_Init(&hash_ctx);
5856 SHA256_Update(&hash_ctx, bdata, chain->bytes);
5857 SHA256_Final(u.digest, &hash_ctx);
5858 u.digest64[2] ^= u.digest64[3];
5860 chain->bref.check.sha192.data,
5861 sizeof(chain->bref.check.sha192.data)) == 0) {
5865 kprintf("chain %016jx.%02x meth=%02x "
5867 chain->bref.data_off,
5869 chain->bref.methods);
5873 case HAMMER2_CHECK_FREEMAP:
5874 r = (chain->bref.check.freemap.icrc32 ==
5875 hammer2_icrc32(bdata, chain->bytes));
5877 kprintf("chain %016jx.%02x meth=%02x "
5879 chain->bref.data_off,
5881 chain->bref.methods);
5882 kprintf("freemap.icrc %08x icrc32 %08x (%d)\n",
5883 chain->bref.check.freemap.icrc32,
5884 hammer2_icrc32(bdata, chain->bytes),
5887 kprintf("dio %p buf %016jx,%d bdata %p/%p\n",
5888 chain->dio, chain->dio->bp->b_loffset,
5889 chain->dio->bp->b_bufsize, bdata,
5890 chain->dio->bp->b_data);
5895 kprintf("hammer2_chain_setcheck: unknown check type %02x\n",
5896 chain->bref.methods);
5904 * Acquire the chain and parent representing the specified inode for the
5905 * device at the specified cluster index.
5907 * The flags passed in are LOOKUP flags, not RESOLVE flags.
5909 * If we are unable to locate the inode, HAMMER2_ERROR_EIO is returned and
5910 * *chainp will be NULL. *parentp may still be set error or not, or NULL
5911 * if the parent itself could not be resolved.
5913 * The caller may pass-in a locked *parentp and/or *chainp, or neither.
5914 * They will be unlocked and released by this function. The *parentp and
5915 * *chainp representing the located inode are returned locked.
5918 hammer2_chain_inode_find(hammer2_pfs_t *pmp, hammer2_key_t inum,
5919 int clindex, int flags,
5920 hammer2_chain_t **parentp, hammer2_chain_t **chainp)
5922 hammer2_chain_t *parent;
5923 hammer2_chain_t *rchain;
5924 hammer2_key_t key_dummy;
5925 hammer2_inode_t *ip;
5929 resolve_flags = (flags & HAMMER2_LOOKUP_SHARED) ?
5930 HAMMER2_RESOLVE_SHARED : 0;
5933 * Caller expects us to replace these.
5936 hammer2_chain_unlock(*chainp);
5937 hammer2_chain_drop(*chainp);
5941 hammer2_chain_unlock(*parentp);
5942 hammer2_chain_drop(*parentp);
5947 * Be very careful, this is a backend function and we CANNOT
5948 * lock any frontend inode structure we find. But we have to
5949 * look the inode up this way first in case it exists but is
5950 * detached from the radix tree.
5952 ip = hammer2_inode_lookup(pmp, inum);
5954 *chainp = hammer2_inode_chain_and_parent(ip, clindex,
5957 hammer2_inode_drop(ip);
5960 hammer2_chain_unlock(*chainp);
5961 hammer2_chain_drop(*chainp);
5964 hammer2_chain_unlock(*parentp);
5965 hammer2_chain_drop(*parentp);
5971 * Inodes hang off of the iroot (bit 63 is clear, differentiating
5972 * inodes from root directory entries in the key lookup).
5974 parent = hammer2_inode_chain(pmp->iroot, clindex, resolve_flags);
5977 rchain = hammer2_chain_lookup(&parent, &key_dummy,
5981 error = HAMMER2_ERROR_EIO;
5990 * Used by the bulkscan code to snapshot the synchronized storage for
5991 * a volume, allowing it to be scanned concurrently against normal
5995 hammer2_chain_bulksnap(hammer2_dev_t *hmp)
5997 hammer2_chain_t *copy;
5999 copy = hammer2_chain_alloc(hmp, hmp->spmp, &hmp->vchain.bref);
6000 copy->data = kmalloc(sizeof(copy->data->voldata),
6003 hammer2_voldata_lock(hmp);
6004 copy->data->voldata = hmp->volsync;
6005 hammer2_voldata_unlock(hmp);
6011 hammer2_chain_bulkdrop(hammer2_chain_t *copy)
6013 KKASSERT(copy->bref.type == HAMMER2_BREF_TYPE_VOLUME);
6014 KKASSERT(copy->data);
6015 kfree(copy->data, copy->hmp->mchain);
6017 atomic_add_long(&hammer2_chain_allocs, -1);
6018 hammer2_chain_drop(copy);
6022 * Returns non-zero if the chain (INODE or DIRENT) matches the
6026 hammer2_chain_dirent_test(hammer2_chain_t *chain, const char *name,
6029 const hammer2_inode_data_t *ripdata;
6030 const hammer2_dirent_head_t *den;
6032 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
6033 ripdata = &chain->data->ipdata;
6034 if (ripdata->meta.name_len == name_len &&
6035 bcmp(ripdata->filename, name, name_len) == 0) {
6039 if (chain->bref.type == HAMMER2_BREF_TYPE_DIRENT &&
6040 chain->bref.embed.dirent.namlen == name_len) {
6041 den = &chain->bref.embed.dirent;
6042 if (name_len > sizeof(chain->bref.check.buf) &&
6043 bcmp(chain->data->buf, name, name_len) == 0) {
6046 if (name_len <= sizeof(chain->bref.check.buf) &&
6047 bcmp(chain->bref.check.buf, name, name_len) == 0) {