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).
91 RB_GENERATE(hammer2_chain_tree, hammer2_chain, rbnode, hammer2_chain_cmp);
94 hammer2_chain_cmp(hammer2_chain_t *chain1, hammer2_chain_t *chain2)
102 * Compare chains. Overlaps are not supposed to happen and catch
103 * any software issues early we count overlaps as a match.
105 c1_beg = chain1->bref.key;
106 c1_end = c1_beg + ((hammer2_key_t)1 << chain1->bref.keybits) - 1;
107 c2_beg = chain2->bref.key;
108 c2_end = c2_beg + ((hammer2_key_t)1 << chain2->bref.keybits) - 1;
110 if (c1_end < c2_beg) /* fully to the left */
112 if (c1_beg > c2_end) /* fully to the right */
114 return(0); /* overlap (must not cross edge boundary) */
118 * Assert that a chain has no media data associated with it.
121 hammer2_chain_assert_no_data(hammer2_chain_t *chain)
123 KKASSERT(chain->dio == NULL);
124 if (chain->bref.type != HAMMER2_BREF_TYPE_VOLUME &&
125 chain->bref.type != HAMMER2_BREF_TYPE_FREEMAP &&
127 panic("hammer2_chain_assert_no_data: chain %p still has data",
133 * Make a chain visible to the flusher. The flusher operates using a top-down
134 * recursion based on the ONFLUSH flag. It locates MODIFIED and UPDATE chains,
135 * flushes them, and updates blocks back to the volume root.
137 * This routine sets the ONFLUSH flag upward from the triggering chain until
138 * it hits an inode root or the volume root. Inode chains serve as inflection
139 * points, requiring the flusher to bridge across trees. Inodes include
140 * regular inodes, PFS roots (pmp->iroot), and the media super root
144 hammer2_chain_setflush(hammer2_chain_t *chain)
146 hammer2_chain_t *parent;
148 if ((chain->flags & HAMMER2_CHAIN_ONFLUSH) == 0) {
149 hammer2_spin_sh(&chain->core.spin);
150 while ((chain->flags & HAMMER2_CHAIN_ONFLUSH) == 0) {
151 atomic_set_int(&chain->flags, HAMMER2_CHAIN_ONFLUSH);
152 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE)
154 if ((parent = chain->parent) == NULL)
156 hammer2_spin_sh(&parent->core.spin);
157 hammer2_spin_unsh(&chain->core.spin);
160 hammer2_spin_unsh(&chain->core.spin);
165 * Allocate a new disconnected chain element representing the specified
166 * bref. chain->refs is set to 1 and the passed bref is copied to
167 * chain->bref. chain->bytes is derived from the bref.
169 * chain->pmp inherits pmp unless the chain is an inode (other than the
172 * NOTE: Returns a referenced but unlocked (because there is no core) chain.
175 hammer2_chain_alloc(hammer2_dev_t *hmp, hammer2_pfs_t *pmp,
176 hammer2_blockref_t *bref)
178 hammer2_chain_t *chain;
182 * Special case - radix of 0 indicates a chain that does not
183 * need a data reference (context is completely embedded in the
186 if ((int)(bref->data_off & HAMMER2_OFF_MASK_RADIX))
187 bytes = 1U << (int)(bref->data_off & HAMMER2_OFF_MASK_RADIX);
191 atomic_add_long(&hammer2_chain_allocs, 1);
194 * Construct the appropriate system structure.
197 case HAMMER2_BREF_TYPE_DIRENT:
198 case HAMMER2_BREF_TYPE_INODE:
199 case HAMMER2_BREF_TYPE_INDIRECT:
200 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
201 case HAMMER2_BREF_TYPE_DATA:
202 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
203 chain = kmalloc(sizeof(*chain), hmp->mchain, M_WAITOK | M_ZERO);
205 case HAMMER2_BREF_TYPE_VOLUME:
206 case HAMMER2_BREF_TYPE_FREEMAP:
208 * Only hammer2_chain_bulksnap() calls this function with these
211 chain = kmalloc(sizeof(*chain), hmp->mchain, M_WAITOK | M_ZERO);
215 panic("hammer2_chain_alloc: unrecognized blockref type: %d",
220 * Initialize the new chain structure. pmp must be set to NULL for
221 * chains belonging to the super-root topology of a device mount.
223 if (pmp == hmp->spmp)
230 chain->bytes = bytes;
232 chain->flags = HAMMER2_CHAIN_ALLOCATED;
233 lockinit(&chain->diolk, "chdio", 0, 0);
236 * Set the PFS boundary flag if this chain represents a PFS root.
238 if (bref->flags & HAMMER2_BREF_FLAG_PFSROOT)
239 atomic_set_int(&chain->flags, HAMMER2_CHAIN_PFSBOUNDARY);
240 hammer2_chain_core_init(chain);
246 * Initialize a chain's core structure. This structure used to be allocated
247 * but is now embedded.
249 * The core is not locked. No additional refs on the chain are made.
250 * (trans) must not be NULL if (core) is not NULL.
253 hammer2_chain_core_init(hammer2_chain_t *chain)
256 * Fresh core under nchain (no multi-homing of ochain's
259 RB_INIT(&chain->core.rbtree); /* live chains */
260 hammer2_mtx_init(&chain->lock, "h2chain");
264 * Add a reference to a chain element, preventing its destruction.
266 * (can be called with spinlock held)
269 hammer2_chain_ref(hammer2_chain_t *chain)
271 if (atomic_fetchadd_int(&chain->refs, 1) == 0) {
273 * Just flag that the chain was used and should be recycled
274 * on the LRU if it encounters it later.
276 if (chain->flags & HAMMER2_CHAIN_ONLRU)
277 atomic_set_int(&chain->flags, HAMMER2_CHAIN_LRUHINT);
281 * REMOVED - reduces contention, lru_list is more heuristical
284 * 0->non-zero transition must ensure that chain is removed
287 * NOTE: Already holding lru_spin here so we cannot call
288 * hammer2_chain_ref() to get it off lru_list, do
291 if (chain->flags & HAMMER2_CHAIN_ONLRU) {
292 hammer2_pfs_t *pmp = chain->pmp;
293 hammer2_spin_ex(&pmp->lru_spin);
294 if (chain->flags & HAMMER2_CHAIN_ONLRU) {
295 atomic_add_int(&pmp->lru_count, -1);
296 atomic_clear_int(&chain->flags,
297 HAMMER2_CHAIN_ONLRU);
298 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node);
300 hammer2_spin_unex(&pmp->lru_spin);
307 * Ref a locked chain and force the data to be held across an unlock.
308 * Chain must be currently locked. The user of the chain who desires
309 * to release the hold must call hammer2_chain_lock_unhold() to relock
310 * and unhold the chain, then unlock normally, or may simply call
311 * hammer2_chain_drop_unhold() (which is safer against deadlocks).
314 hammer2_chain_ref_hold(hammer2_chain_t *chain)
316 atomic_add_int(&chain->lockcnt, 1);
317 hammer2_chain_ref(chain);
321 * Insert the chain in the core rbtree.
323 * Normal insertions are placed in the live rbtree. Insertion of a deleted
324 * chain is a special case used by the flush code that is placed on the
325 * unstaged deleted list to avoid confusing the live view.
327 #define HAMMER2_CHAIN_INSERT_SPIN 0x0001
328 #define HAMMER2_CHAIN_INSERT_LIVE 0x0002
329 #define HAMMER2_CHAIN_INSERT_RACE 0x0004
333 hammer2_chain_insert(hammer2_chain_t *parent, hammer2_chain_t *chain,
334 int flags, int generation)
336 hammer2_chain_t *xchain;
339 if (flags & HAMMER2_CHAIN_INSERT_SPIN)
340 hammer2_spin_ex(&parent->core.spin);
343 * Interlocked by spinlock, check for race
345 if ((flags & HAMMER2_CHAIN_INSERT_RACE) &&
346 parent->core.generation != generation) {
347 error = HAMMER2_ERROR_EAGAIN;
354 xchain = RB_INSERT(hammer2_chain_tree, &parent->core.rbtree, chain);
355 KASSERT(xchain == NULL,
356 ("hammer2_chain_insert: collision %p %p (key=%016jx)",
357 chain, xchain, chain->bref.key));
358 atomic_set_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE);
359 chain->parent = parent;
360 ++parent->core.chain_count;
361 ++parent->core.generation; /* XXX incs for _get() too, XXX */
364 * We have to keep track of the effective live-view blockref count
365 * so the create code knows when to push an indirect block.
367 if (flags & HAMMER2_CHAIN_INSERT_LIVE)
368 atomic_add_int(&parent->core.live_count, 1);
370 if (flags & HAMMER2_CHAIN_INSERT_SPIN)
371 hammer2_spin_unex(&parent->core.spin);
376 * Drop the caller's reference to the chain. When the ref count drops to
377 * zero this function will try to disassociate the chain from its parent and
378 * deallocate it, then recursely drop the parent using the implied ref
379 * from the chain's chain->parent.
381 * Nobody should own chain's mutex on the 1->0 transition, unless this drop
382 * races an acquisition by another cpu. Therefore we can loop if we are
383 * unable to acquire the mutex, and refs is unlikely to be 1 unless we again
384 * race against another drop.
386 static hammer2_chain_t *hammer2_chain_lastdrop(hammer2_chain_t *chain,
388 static void hammer2_chain_lru_flush(hammer2_pfs_t *pmp);
391 hammer2_chain_drop(hammer2_chain_t *chain)
395 if (hammer2_debug & 0x200000)
398 KKASSERT(chain->refs > 0);
406 if (hammer2_mtx_ex_try(&chain->lock) == 0)
407 chain = hammer2_chain_lastdrop(chain, 0);
408 /* retry the same chain, or chain from lastdrop */
410 if (atomic_cmpset_int(&chain->refs, refs, refs - 1))
412 /* retry the same chain */
419 * Unhold a held and probably not-locked chain, ensure that the data is
420 * dropped on the 1->0 transition of lockcnt by obtaining an exclusive
421 * lock and then simply unlocking the chain.
424 hammer2_chain_unhold(hammer2_chain_t *chain)
430 lockcnt = chain->lockcnt;
433 if (atomic_cmpset_int(&chain->lockcnt,
434 lockcnt, lockcnt - 1)) {
437 } else if (hammer2_mtx_ex_try(&chain->lock) == 0) {
438 hammer2_chain_unlock(chain);
442 * This situation can easily occur on SMP due to
443 * the gap inbetween the 1->0 transition and the
444 * final unlock. We cannot safely block on the
445 * mutex because lockcnt might go above 1.
447 * XXX Sleep for one tick if it takes too long.
450 if (iter > 1000 + hz) {
451 kprintf("hammer2: h2race1 %p\n", chain);
454 tsleep(&iter, 0, "h2race1", 1);
462 hammer2_chain_drop_unhold(hammer2_chain_t *chain)
464 hammer2_chain_unhold(chain);
465 hammer2_chain_drop(chain);
469 hammer2_chain_rehold(hammer2_chain_t *chain)
471 hammer2_chain_lock(chain, HAMMER2_RESOLVE_SHARED);
472 atomic_add_int(&chain->lockcnt, 1);
473 hammer2_chain_unlock(chain);
477 * Handles the (potential) last drop of chain->refs from 1->0. Called with
478 * the mutex exclusively locked, refs == 1, and lockcnt 0. SMP races are
479 * possible against refs and lockcnt. We must dispose of the mutex on chain.
481 * This function returns an unlocked chain for recursive drop or NULL. It
482 * can return the same chain if it determines it has raced another ref.
486 * When two chains need to be recursively dropped we use the chain we
487 * would otherwise free to placehold the additional chain. It's a bit
488 * convoluted but we can't just recurse without potentially blowing out
491 * The chain cannot be freed if it has any children.
492 * The chain cannot be freed if flagged MODIFIED unless we can dispose of it.
493 * The chain cannot be freed if flagged UPDATE unless we can dispose of it.
494 * Any dedup registration can remain intact.
496 * The core spinlock is allowed to nest child-to-parent (not parent-to-child).
500 hammer2_chain_lastdrop(hammer2_chain_t *chain, int depth)
504 hammer2_chain_t *parent;
505 hammer2_chain_t *rdrop;
508 * We need chain's spinlock to interlock the sub-tree test.
509 * We already have chain's mutex, protecting chain->parent.
511 * Remember that chain->refs can be in flux.
513 hammer2_spin_ex(&chain->core.spin);
515 if (chain->parent != NULL) {
517 * If the chain has a parent the UPDATE bit prevents scrapping
518 * as the chain is needed to properly flush the parent. Try
519 * to complete the 1->0 transition and return NULL. Retry
520 * (return chain) if we are unable to complete the 1->0
521 * transition, else return NULL (nothing more to do).
523 * If the chain has a parent the MODIFIED bit prevents
526 * Chains with UPDATE/MODIFIED are *not* put on the LRU list!
528 if (chain->flags & (HAMMER2_CHAIN_UPDATE |
529 HAMMER2_CHAIN_MODIFIED)) {
530 if (atomic_cmpset_int(&chain->refs, 1, 0)) {
531 hammer2_spin_unex(&chain->core.spin);
532 hammer2_chain_assert_no_data(chain);
533 hammer2_mtx_unlock(&chain->lock);
536 hammer2_spin_unex(&chain->core.spin);
537 hammer2_mtx_unlock(&chain->lock);
541 /* spinlock still held */
542 } else if (chain->bref.type == HAMMER2_BREF_TYPE_VOLUME ||
543 chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP) {
545 * Retain the static vchain and fchain. Clear bits that
546 * are not relevant. Do not clear the MODIFIED bit,
547 * and certainly do not put it on the delayed-flush queue.
549 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
552 * The chain has no parent and can be flagged for destruction.
553 * Since it has no parent, UPDATE can also be cleared.
555 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DESTROY);
556 if (chain->flags & HAMMER2_CHAIN_UPDATE)
557 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
560 * If the chain has children we must propagate the DESTROY
561 * flag downward and rip the disconnected topology apart.
562 * This is accomplished by calling hammer2_flush() on the
565 * Any dedup is already handled by the underlying DIO, so
566 * we do not have to specifically flush it here.
568 if (chain->core.chain_count) {
569 hammer2_spin_unex(&chain->core.spin);
570 hammer2_flush(chain, HAMMER2_FLUSH_TOP |
572 hammer2_mtx_unlock(&chain->lock);
574 return(chain); /* retry drop */
578 * Otherwise we can scrap the MODIFIED bit if it is set,
579 * and continue along the freeing path.
581 * Be sure to clean-out any dedup bits. Without a parent
582 * this chain will no longer be visible to the flush code.
583 * Easy check data_off to avoid the volume root.
585 if (chain->flags & HAMMER2_CHAIN_MODIFIED) {
586 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
587 atomic_add_long(&hammer2_count_modified_chains, -1);
589 hammer2_pfs_memory_wakeup(chain->pmp);
591 /* spinlock still held */
594 /* spinlock still held */
597 * If any children exist we must leave the chain intact with refs == 0.
598 * They exist because chains are retained below us which have refs or
599 * may require flushing.
601 * Retry (return chain) if we fail to transition the refs to 0, else
602 * return NULL indication nothing more to do.
604 * Chains with children are NOT put on the LRU list.
606 if (chain->core.chain_count) {
607 if (atomic_cmpset_int(&chain->refs, 1, 0)) {
608 hammer2_spin_unex(&chain->core.spin);
609 hammer2_chain_assert_no_data(chain);
610 hammer2_mtx_unlock(&chain->lock);
613 hammer2_spin_unex(&chain->core.spin);
614 hammer2_mtx_unlock(&chain->lock);
618 /* spinlock still held */
619 /* no chains left under us */
622 * chain->core has no children left so no accessors can get to our
623 * chain from there. Now we have to lock the parent core to interlock
624 * remaining possible accessors that might bump chain's refs before
625 * we can safely drop chain's refs with intent to free the chain.
628 pmp = chain->pmp; /* can be NULL */
631 parent = chain->parent;
634 * WARNING! chain's spin lock is still held here, and other spinlocks
635 * will be acquired and released in the code below. We
636 * cannot be making fancy procedure calls!
640 * We can cache the chain if it is associated with a pmp
641 * and not flagged as being destroyed or requesting a full
642 * release. In this situation the chain is not removed
643 * from its parent, i.e. it can still be looked up.
645 * We intentionally do not cache DATA chains because these
646 * were likely used to load data into the logical buffer cache
647 * and will not be accessed again for some time.
650 (HAMMER2_CHAIN_DESTROY | HAMMER2_CHAIN_RELEASE)) == 0 &&
652 chain->bref.type != HAMMER2_BREF_TYPE_DATA) {
654 hammer2_spin_ex(&parent->core.spin);
655 if (atomic_cmpset_int(&chain->refs, 1, 0) == 0) {
657 * 1->0 transition failed, retry. Do not drop
658 * the chain's data yet!
661 hammer2_spin_unex(&parent->core.spin);
662 hammer2_spin_unex(&chain->core.spin);
663 hammer2_mtx_unlock(&chain->lock);
671 hammer2_chain_assert_no_data(chain);
674 * Make sure we are on the LRU list, clean up excessive
675 * LRU entries. We can only really drop one but there might
676 * be other entries that we can remove from the lru_list
679 * NOTE: HAMMER2_CHAIN_ONLRU may only be safely set when
680 * chain->core.spin AND pmp->lru_spin are held, but
681 * can be safely cleared only holding pmp->lru_spin.
683 if ((chain->flags & HAMMER2_CHAIN_ONLRU) == 0) {
684 hammer2_spin_ex(&pmp->lru_spin);
685 if ((chain->flags & HAMMER2_CHAIN_ONLRU) == 0) {
686 atomic_set_int(&chain->flags,
687 HAMMER2_CHAIN_ONLRU);
688 TAILQ_INSERT_TAIL(&pmp->lru_list,
690 atomic_add_int(&pmp->lru_count, 1);
692 if (pmp->lru_count < HAMMER2_LRU_LIMIT)
693 depth = 1; /* disable lru_list flush */
694 hammer2_spin_unex(&pmp->lru_spin);
696 /* disable lru flush */
701 hammer2_spin_unex(&parent->core.spin);
702 parent = NULL; /* safety */
704 hammer2_spin_unex(&chain->core.spin);
705 hammer2_mtx_unlock(&chain->lock);
708 * lru_list hysteresis (see above for depth overrides).
709 * Note that depth also prevents excessive lastdrop recursion.
712 hammer2_chain_lru_flush(pmp);
719 * Make sure we are not on the LRU list.
721 if (chain->flags & HAMMER2_CHAIN_ONLRU) {
722 hammer2_spin_ex(&pmp->lru_spin);
723 if (chain->flags & HAMMER2_CHAIN_ONLRU) {
724 atomic_add_int(&pmp->lru_count, -1);
725 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU);
726 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node);
728 hammer2_spin_unex(&pmp->lru_spin);
732 * Spinlock the parent and try to drop the last ref on chain.
733 * On success determine if we should dispose of the chain
734 * (remove the chain from its parent, etc).
736 * (normal core locks are top-down recursive but we define
737 * core spinlocks as bottom-up recursive, so this is safe).
740 hammer2_spin_ex(&parent->core.spin);
741 if (atomic_cmpset_int(&chain->refs, 1, 0) == 0) {
743 * 1->0 transition failed, retry.
745 hammer2_spin_unex(&parent->core.spin);
746 hammer2_spin_unex(&chain->core.spin);
747 hammer2_mtx_unlock(&chain->lock);
753 * 1->0 transition successful, parent spin held to prevent
754 * new lookups, chain spinlock held to protect parent field.
755 * Remove chain from the parent.
757 * If the chain is being removed from the parent's btree but
758 * is not bmapped, we have to adjust live_count downward. If
759 * it is bmapped then the blockref is retained in the parent
760 * as is its associated live_count. This case can occur when
761 * a chain added to the topology is unable to flush and is
762 * then later deleted.
764 if (chain->flags & HAMMER2_CHAIN_ONRBTREE) {
765 if ((parent->flags & HAMMER2_CHAIN_COUNTEDBREFS) &&
766 (chain->flags & HAMMER2_CHAIN_BMAPPED) == 0) {
767 atomic_add_int(&parent->core.live_count, -1);
769 RB_REMOVE(hammer2_chain_tree,
770 &parent->core.rbtree, chain);
771 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE);
772 --parent->core.chain_count;
773 chain->parent = NULL;
777 * If our chain was the last chain in the parent's core the
778 * core is now empty and its parent might have to be
779 * re-dropped if it has 0 refs.
781 if (parent->core.chain_count == 0) {
783 atomic_add_int(&rdrop->refs, 1);
785 if (atomic_cmpset_int(&rdrop->refs, 0, 1) == 0)
789 hammer2_spin_unex(&parent->core.spin);
790 parent = NULL; /* safety */
796 if (atomic_cmpset_int(&chain->refs, 1, 0) == 0) {
798 * 1->0 transition failed, retry.
800 hammer2_spin_unex(&parent->core.spin);
801 hammer2_spin_unex(&chain->core.spin);
802 hammer2_mtx_unlock(&chain->lock);
809 * Successful 1->0 transition, no parent, no children... no way for
810 * anyone to ref this chain any more. We can clean-up and free it.
812 * We still have the core spinlock, and core's chain_count is 0.
813 * Any parent spinlock is gone.
815 hammer2_spin_unex(&chain->core.spin);
816 hammer2_chain_assert_no_data(chain);
817 hammer2_mtx_unlock(&chain->lock);
818 KKASSERT(RB_EMPTY(&chain->core.rbtree) &&
819 chain->core.chain_count == 0);
822 * All locks are gone, no pointers remain to the chain, finish
825 KKASSERT((chain->flags & (HAMMER2_CHAIN_UPDATE |
826 HAMMER2_CHAIN_MODIFIED)) == 0);
829 * Once chain resources are gone we can use the now dead chain
830 * structure to placehold what might otherwise require a recursive
831 * drop, because we have potentially two things to drop and can only
832 * return one directly.
834 if (chain->flags & HAMMER2_CHAIN_ALLOCATED) {
835 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ALLOCATED);
837 kfree(chain, hmp->mchain);
841 * Possible chaining loop when parent re-drop needed.
847 * Heuristical flush of the LRU, try to reduce the number of entries
848 * on the LRU to (HAMMER2_LRU_LIMIT * 2 / 3). This procedure is called
849 * only when lru_count exceeds HAMMER2_LRU_LIMIT.
853 hammer2_chain_lru_flush(hammer2_pfs_t *pmp)
855 hammer2_chain_t *chain;
859 hammer2_spin_ex(&pmp->lru_spin);
860 while (pmp->lru_count > HAMMER2_LRU_LIMIT * 2 / 3) {
862 * Pick a chain off the lru_list, just recycle it quickly
863 * if LRUHINT is set (the chain was ref'd but left on
864 * the lru_list, so cycle to the end).
866 chain = TAILQ_FIRST(&pmp->lru_list);
867 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node);
869 if (chain->flags & HAMMER2_CHAIN_LRUHINT) {
870 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_LRUHINT);
871 TAILQ_INSERT_TAIL(&pmp->lru_list, chain, lru_node);
877 * Ok, we are off the LRU. We must adjust refs before we
878 * can safely clear the ONLRU flag.
880 atomic_add_int(&pmp->lru_count, -1);
881 if (atomic_cmpset_int(&chain->refs, 0, 1)) {
882 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU);
883 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
886 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU);
889 hammer2_spin_unex(&pmp->lru_spin);
894 * If we picked a chain off the lru list we may be able to lastdrop
895 * it. Use a depth of 1 to prevent excessive lastdrop recursion.
905 if (hammer2_mtx_ex_try(&chain->lock) == 0)
906 chain = hammer2_chain_lastdrop(chain, 1);
907 /* retry the same chain, or chain from lastdrop */
909 if (atomic_cmpset_int(&chain->refs, refs, refs - 1))
911 /* retry the same chain */
919 * On last lock release.
921 static hammer2_io_t *
922 hammer2_chain_drop_data(hammer2_chain_t *chain)
926 if ((dio = chain->dio) != NULL) {
930 switch(chain->bref.type) {
931 case HAMMER2_BREF_TYPE_VOLUME:
932 case HAMMER2_BREF_TYPE_FREEMAP:
935 if (chain->data != NULL) {
936 hammer2_spin_unex(&chain->core.spin);
937 panic("chain data not null: "
938 "chain %p bref %016jx.%02x "
939 "refs %d parent %p dio %p data %p",
940 chain, chain->bref.data_off,
941 chain->bref.type, chain->refs,
943 chain->dio, chain->data);
945 KKASSERT(chain->data == NULL);
953 * Lock a referenced chain element, acquiring its data with I/O if necessary,
954 * and specify how you would like the data to be resolved.
956 * If an I/O or other fatal error occurs, chain->error will be set to non-zero.
958 * The lock is allowed to recurse, multiple locking ops will aggregate
959 * the requested resolve types. Once data is assigned it will not be
960 * removed until the last unlock.
962 * HAMMER2_RESOLVE_NEVER - Do not resolve the data element.
963 * (typically used to avoid device/logical buffer
966 * HAMMER2_RESOLVE_MAYBE - Do not resolve data elements for chains in
967 * the INITIAL-create state (indirect blocks only).
969 * Do not resolve data elements for DATA chains.
970 * (typically used to avoid device/logical buffer
973 * HAMMER2_RESOLVE_ALWAYS- Always resolve the data element.
975 * HAMMER2_RESOLVE_SHARED- (flag) The chain is locked shared, otherwise
976 * it will be locked exclusive.
978 * HAMMER2_RESOLVE_NONBLOCK- (flag) The chain is locked non-blocking. If
979 * the lock fails, EAGAIN is returned.
981 * NOTE: Embedded elements (volume header, inodes) are always resolved
984 * NOTE: Specifying HAMMER2_RESOLVE_ALWAYS on a newly-created non-embedded
985 * element will instantiate and zero its buffer, and flush it on
988 * NOTE: (data) elements are normally locked RESOLVE_NEVER or RESOLVE_MAYBE
989 * so as not to instantiate a device buffer, which could alias against
990 * a logical file buffer. However, if ALWAYS is specified the
991 * device buffer will be instantiated anyway.
993 * NOTE: The return value is always 0 unless NONBLOCK is specified, in which
994 * case it can be either 0 or EAGAIN.
996 * WARNING! This function blocks on I/O if data needs to be fetched. This
997 * blocking can run concurrent with other compatible lock holders
998 * who do not need data returning. The lock is not upgraded to
999 * exclusive during a data fetch, a separate bit is used to
1000 * interlock I/O. However, an exclusive lock holder can still count
1001 * on being interlocked against an I/O fetch managed by a shared
1005 hammer2_chain_lock(hammer2_chain_t *chain, int how)
1007 KKASSERT(chain->refs > 0);
1009 if (how & HAMMER2_RESOLVE_NONBLOCK) {
1011 * We still have to bump lockcnt before acquiring the lock,
1012 * even for non-blocking operation, because the unlock code
1013 * live-loops on lockcnt == 1 when dropping the last lock.
1015 * If the non-blocking operation fails we have to use an
1016 * unhold sequence to undo the mess.
1018 * NOTE: LOCKAGAIN must always succeed without blocking,
1019 * even if NONBLOCK is specified.
1021 atomic_add_int(&chain->lockcnt, 1);
1022 if (how & HAMMER2_RESOLVE_SHARED) {
1023 if (how & HAMMER2_RESOLVE_LOCKAGAIN) {
1024 hammer2_mtx_sh_again(&chain->lock);
1026 if (hammer2_mtx_sh_try(&chain->lock) != 0) {
1027 hammer2_chain_unhold(chain);
1032 if (hammer2_mtx_ex_try(&chain->lock) != 0) {
1033 hammer2_chain_unhold(chain);
1037 ++curthread->td_tracker;
1040 * Get the appropriate lock. If LOCKAGAIN is flagged with
1041 * SHARED the caller expects a shared lock to already be
1042 * present and we are giving it another ref. This case must
1043 * importantly not block if there is a pending exclusive lock
1046 atomic_add_int(&chain->lockcnt, 1);
1047 if (how & HAMMER2_RESOLVE_SHARED) {
1048 if (how & HAMMER2_RESOLVE_LOCKAGAIN) {
1049 hammer2_mtx_sh_again(&chain->lock);
1051 hammer2_mtx_sh(&chain->lock);
1054 hammer2_mtx_ex(&chain->lock);
1056 ++curthread->td_tracker;
1060 * If we already have a valid data pointer make sure the data is
1061 * synchronized to the current cpu, and then no further action is
1066 hammer2_io_bkvasync(chain->dio);
1071 * Do we have to resolve the data? This is generally only
1072 * applicable to HAMMER2_BREF_TYPE_DATA which is special-cased.
1073 * Other BREF types expects the data to be there.
1075 switch(how & HAMMER2_RESOLVE_MASK) {
1076 case HAMMER2_RESOLVE_NEVER:
1078 case HAMMER2_RESOLVE_MAYBE:
1079 if (chain->flags & HAMMER2_CHAIN_INITIAL)
1081 if (chain->bref.type == HAMMER2_BREF_TYPE_DATA)
1084 if (chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE)
1086 if (chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_LEAF)
1090 case HAMMER2_RESOLVE_ALWAYS:
1096 * Caller requires data
1098 hammer2_chain_load_data(chain);
1104 * Lock the chain, retain the hold, and drop the data persistence count.
1105 * The data should remain valid because we never transitioned lockcnt
1109 hammer2_chain_lock_unhold(hammer2_chain_t *chain, int how)
1111 hammer2_chain_lock(chain, how);
1112 atomic_add_int(&chain->lockcnt, -1);
1117 * Downgrade an exclusive chain lock to a shared chain lock.
1119 * NOTE: There is no upgrade equivalent due to the ease of
1120 * deadlocks in that direction.
1123 hammer2_chain_lock_downgrade(hammer2_chain_t *chain)
1125 hammer2_mtx_downgrade(&chain->lock);
1130 * Issue I/O and install chain->data. Caller must hold a chain lock, lock
1131 * may be of any type.
1133 * Once chain->data is set it cannot be disposed of until all locks are
1136 * Make sure the data is synchronized to the current cpu.
1139 hammer2_chain_load_data(hammer2_chain_t *chain)
1141 hammer2_blockref_t *bref;
1148 * Degenerate case, data already present, or chain has no media
1149 * reference to load.
1151 KKASSERT(chain->lock.mtx_lock & MTX_MASK);
1154 hammer2_io_bkvasync(chain->dio);
1157 if ((chain->bref.data_off & ~HAMMER2_OFF_MASK_RADIX) == 0)
1161 KKASSERT(hmp != NULL);
1164 * Gain the IOINPROG bit, interlocked block.
1170 oflags = chain->flags;
1172 if (oflags & HAMMER2_CHAIN_IOINPROG) {
1173 nflags = oflags | HAMMER2_CHAIN_IOSIGNAL;
1174 tsleep_interlock(&chain->flags, 0);
1175 if (atomic_cmpset_int(&chain->flags, oflags, nflags)) {
1176 tsleep(&chain->flags, PINTERLOCKED,
1181 nflags = oflags | HAMMER2_CHAIN_IOINPROG;
1182 if (atomic_cmpset_int(&chain->flags, oflags, nflags)) {
1190 * We own CHAIN_IOINPROG
1192 * Degenerate case if we raced another load.
1196 hammer2_io_bkvasync(chain->dio);
1201 * We must resolve to a device buffer, either by issuing I/O or
1202 * by creating a zero-fill element. We do not mark the buffer
1203 * dirty when creating a zero-fill element (the hammer2_chain_modify()
1204 * API must still be used to do that).
1206 * The device buffer is variable-sized in powers of 2 down
1207 * to HAMMER2_MIN_ALLOC (typically 1K). A 64K physical storage
1208 * chunk always contains buffers of the same size. (XXX)
1210 * The minimum physical IO size may be larger than the variable
1213 bref = &chain->bref;
1216 * The getblk() optimization can only be used on newly created
1217 * elements if the physical block size matches the request.
1219 if (chain->flags & HAMMER2_CHAIN_INITIAL) {
1220 error = hammer2_io_new(hmp, bref->type,
1221 bref->data_off, chain->bytes,
1224 error = hammer2_io_bread(hmp, bref->type,
1225 bref->data_off, chain->bytes,
1227 hammer2_adjreadcounter(&chain->bref, chain->bytes);
1230 chain->error = HAMMER2_ERROR_EIO;
1231 kprintf("hammer2_chain_lock: I/O error %016jx: %d\n",
1232 (intmax_t)bref->data_off, error);
1233 hammer2_io_bqrelse(&chain->dio);
1239 * This isn't perfect and can be ignored on OSs which do not have
1240 * an indication as to whether a buffer is coming from cache or
1241 * if I/O was actually issued for the read. TESTEDGOOD will work
1242 * pretty well without the B_IOISSUED logic because chains are
1243 * cached, but in that situation (without B_IOISSUED) it will not
1244 * detect whether a re-read via I/O is corrupted verses the original
1247 * We can't re-run the CRC on every fresh lock. That would be
1248 * insanely expensive.
1250 * If the underlying kernel buffer covers the entire chain we can
1251 * use the B_IOISSUED indication to determine if we have to re-run
1252 * the CRC on chain data for chains that managed to stay cached
1253 * across the kernel disposal of the original buffer.
1255 if ((dio = chain->dio) != NULL && dio->bp) {
1256 struct buf *bp = dio->bp;
1258 if (dio->psize == chain->bytes &&
1259 (bp->b_flags & B_IOISSUED)) {
1260 atomic_clear_int(&chain->flags,
1261 HAMMER2_CHAIN_TESTEDGOOD);
1262 bp->b_flags &= ~B_IOISSUED;
1267 * NOTE: A locked chain's data cannot be modified without first
1268 * calling hammer2_chain_modify().
1272 * Clear INITIAL. In this case we used io_new() and the buffer has
1273 * been zero'd and marked dirty.
1275 * NOTE: hammer2_io_data() call issues bkvasync()
1277 bdata = hammer2_io_data(chain->dio, chain->bref.data_off);
1279 if (chain->flags & HAMMER2_CHAIN_INITIAL) {
1280 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1281 chain->bref.flags |= HAMMER2_BREF_FLAG_ZERO;
1282 } else if (chain->flags & HAMMER2_CHAIN_MODIFIED) {
1284 * check data not currently synchronized due to
1285 * modification. XXX assumes data stays in the buffer
1286 * cache, which might not be true (need biodep on flush
1287 * to calculate crc? or simple crc?).
1289 } else if ((chain->flags & HAMMER2_CHAIN_TESTEDGOOD) == 0) {
1290 if (hammer2_chain_testcheck(chain, bdata) == 0) {
1291 chain->error = HAMMER2_ERROR_CHECK;
1293 atomic_set_int(&chain->flags, HAMMER2_CHAIN_TESTEDGOOD);
1298 * Setup the data pointer, either pointing it to an embedded data
1299 * structure and copying the data from the buffer, or pointing it
1302 * The buffer is not retained when copying to an embedded data
1303 * structure in order to avoid potential deadlocks or recursions
1304 * on the same physical buffer.
1306 * WARNING! Other threads can start using the data the instant we
1307 * set chain->data non-NULL.
1309 switch (bref->type) {
1310 case HAMMER2_BREF_TYPE_VOLUME:
1311 case HAMMER2_BREF_TYPE_FREEMAP:
1313 * Copy data from bp to embedded buffer
1315 panic("hammer2_chain_load_data: unresolved volume header");
1317 case HAMMER2_BREF_TYPE_DIRENT:
1318 KKASSERT(chain->bytes != 0);
1320 case HAMMER2_BREF_TYPE_INODE:
1321 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
1322 case HAMMER2_BREF_TYPE_INDIRECT:
1323 case HAMMER2_BREF_TYPE_DATA:
1324 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1327 * Point data at the device buffer and leave dio intact.
1329 chain->data = (void *)bdata;
1334 * Release HAMMER2_CHAIN_IOINPROG and signal waiters if requested.
1341 oflags = chain->flags;
1342 nflags = oflags & ~(HAMMER2_CHAIN_IOINPROG |
1343 HAMMER2_CHAIN_IOSIGNAL);
1344 KKASSERT(oflags & HAMMER2_CHAIN_IOINPROG);
1345 if (atomic_cmpset_int(&chain->flags, oflags, nflags)) {
1346 if (oflags & HAMMER2_CHAIN_IOSIGNAL)
1347 wakeup(&chain->flags);
1354 * Unlock and deref a chain element.
1356 * Remember that the presence of children under chain prevent the chain's
1357 * destruction but do not add additional references, so the dio will still
1361 hammer2_chain_unlock(hammer2_chain_t *chain)
1367 --curthread->td_tracker;
1370 * If multiple locks are present (or being attempted) on this
1371 * particular chain we can just unlock, drop refs, and return.
1373 * Otherwise fall-through on the 1->0 transition.
1376 lockcnt = chain->lockcnt;
1377 KKASSERT(lockcnt > 0);
1380 if (atomic_cmpset_int(&chain->lockcnt,
1381 lockcnt, lockcnt - 1)) {
1382 hammer2_mtx_unlock(&chain->lock);
1385 } else if (hammer2_mtx_upgrade_try(&chain->lock) == 0) {
1386 /* while holding the mutex exclusively */
1387 if (atomic_cmpset_int(&chain->lockcnt, 1, 0))
1391 * This situation can easily occur on SMP due to
1392 * the gap inbetween the 1->0 transition and the
1393 * final unlock. We cannot safely block on the
1394 * mutex because lockcnt might go above 1.
1396 * XXX Sleep for one tick if it takes too long.
1398 if (++iter > 1000) {
1399 if (iter > 1000 + hz) {
1400 kprintf("hammer2: h2race2 %p\n", chain);
1403 tsleep(&iter, 0, "h2race2", 1);
1411 * Last unlock / mutex upgraded to exclusive. Drop the data
1414 dio = hammer2_chain_drop_data(chain);
1416 hammer2_io_bqrelse(&dio);
1417 hammer2_mtx_unlock(&chain->lock);
1421 * Unlock and hold chain data intact
1424 hammer2_chain_unlock_hold(hammer2_chain_t *chain)
1426 atomic_add_int(&chain->lockcnt, 1);
1427 hammer2_chain_unlock(chain);
1431 * Helper to obtain the blockref[] array base and count for a chain.
1433 * XXX Not widely used yet, various use cases need to be validated and
1434 * converted to use this function.
1437 hammer2_blockref_t *
1438 hammer2_chain_base_and_count(hammer2_chain_t *parent, int *countp)
1440 hammer2_blockref_t *base;
1443 if (parent->flags & HAMMER2_CHAIN_INITIAL) {
1446 switch(parent->bref.type) {
1447 case HAMMER2_BREF_TYPE_INODE:
1448 count = HAMMER2_SET_COUNT;
1450 case HAMMER2_BREF_TYPE_INDIRECT:
1451 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1452 count = parent->bytes / sizeof(hammer2_blockref_t);
1454 case HAMMER2_BREF_TYPE_VOLUME:
1455 count = HAMMER2_SET_COUNT;
1457 case HAMMER2_BREF_TYPE_FREEMAP:
1458 count = HAMMER2_SET_COUNT;
1461 panic("hammer2_chain_base_and_count: "
1462 "unrecognized blockref type: %d",
1468 switch(parent->bref.type) {
1469 case HAMMER2_BREF_TYPE_INODE:
1470 base = &parent->data->ipdata.u.blockset.blockref[0];
1471 count = HAMMER2_SET_COUNT;
1473 case HAMMER2_BREF_TYPE_INDIRECT:
1474 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1475 base = &parent->data->npdata[0];
1476 count = parent->bytes / sizeof(hammer2_blockref_t);
1478 case HAMMER2_BREF_TYPE_VOLUME:
1479 base = &parent->data->voldata.
1480 sroot_blockset.blockref[0];
1481 count = HAMMER2_SET_COUNT;
1483 case HAMMER2_BREF_TYPE_FREEMAP:
1484 base = &parent->data->blkset.blockref[0];
1485 count = HAMMER2_SET_COUNT;
1488 panic("hammer2_chain_base_and_count: "
1489 "unrecognized blockref type: %d",
1501 * This counts the number of live blockrefs in a block array and
1502 * also calculates the point at which all remaining blockrefs are empty.
1503 * This routine can only be called on a live chain.
1505 * Caller holds the chain locked, but possibly with a shared lock. We
1506 * must use an exclusive spinlock to prevent corruption.
1508 * NOTE: Flag is not set until after the count is complete, allowing
1509 * callers to test the flag without holding the spinlock.
1511 * NOTE: If base is NULL the related chain is still in the INITIAL
1512 * state and there are no blockrefs to count.
1514 * NOTE: live_count may already have some counts accumulated due to
1515 * creation and deletion and could even be initially negative.
1518 hammer2_chain_countbrefs(hammer2_chain_t *chain,
1519 hammer2_blockref_t *base, int count)
1521 hammer2_spin_ex(&chain->core.spin);
1522 if ((chain->flags & HAMMER2_CHAIN_COUNTEDBREFS) == 0) {
1524 while (--count >= 0) {
1525 if (base[count].type != HAMMER2_BREF_TYPE_EMPTY)
1528 chain->core.live_zero = count + 1;
1529 while (count >= 0) {
1530 if (base[count].type != HAMMER2_BREF_TYPE_EMPTY)
1531 atomic_add_int(&chain->core.live_count,
1536 chain->core.live_zero = 0;
1538 /* else do not modify live_count */
1539 atomic_set_int(&chain->flags, HAMMER2_CHAIN_COUNTEDBREFS);
1541 hammer2_spin_unex(&chain->core.spin);
1545 * Resize the chain's physical storage allocation in-place. This function does
1546 * not usually adjust the data pointer and must be followed by (typically) a
1547 * hammer2_chain_modify() call to copy any old data over and adjust the
1550 * Chains can be resized smaller without reallocating the storage. Resizing
1551 * larger will reallocate the storage. Excess or prior storage is reclaimed
1552 * asynchronously at a later time.
1554 * An nradix value of 0 is special-cased to mean that the storage should
1555 * be disassociated, that is the chain is being resized to 0 bytes (not 1
1558 * Must be passed an exclusively locked parent and chain.
1560 * This function is mostly used with DATA blocks locked RESOLVE_NEVER in order
1561 * to avoid instantiating a device buffer that conflicts with the vnode data
1562 * buffer. However, because H2 can compress or encrypt data, the chain may
1563 * have a dio assigned to it in those situations, and they do not conflict.
1565 * XXX return error if cannot resize.
1568 hammer2_chain_resize(hammer2_chain_t *chain,
1569 hammer2_tid_t mtid, hammer2_off_t dedup_off,
1570 int nradix, int flags)
1580 * Only data and indirect blocks can be resized for now.
1581 * (The volu root, inodes, and freemap elements use a fixed size).
1583 KKASSERT(chain != &hmp->vchain);
1584 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_DATA ||
1585 chain->bref.type == HAMMER2_BREF_TYPE_INDIRECT ||
1586 chain->bref.type == HAMMER2_BREF_TYPE_DIRENT);
1589 * Nothing to do if the element is already the proper size
1591 obytes = chain->bytes;
1592 nbytes = (nradix) ? (1U << nradix) : 0;
1593 if (obytes == nbytes)
1594 return (chain->error);
1597 * Make sure the old data is instantiated so we can copy it. If this
1598 * is a data block, the device data may be superfluous since the data
1599 * might be in a logical block, but compressed or encrypted data is
1602 * NOTE: The modify will set BMAPUPD for us if BMAPPED is set.
1604 error = hammer2_chain_modify(chain, mtid, dedup_off, 0);
1609 * Relocate the block, even if making it smaller (because different
1610 * block sizes may be in different regions).
1612 * NOTE: Operation does not copy the data and may only be used
1613 * to resize data blocks in-place, or directory entry blocks
1614 * which are about to be modified in some manner.
1616 error = hammer2_freemap_alloc(chain, nbytes);
1620 chain->bytes = nbytes;
1623 * We don't want the followup chain_modify() to try to copy data
1624 * from the old (wrong-sized) buffer. It won't know how much to
1625 * copy. This case should only occur during writes when the
1626 * originator already has the data to write in-hand.
1629 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_DATA ||
1630 chain->bref.type == HAMMER2_BREF_TYPE_DIRENT);
1631 hammer2_io_brelse(&chain->dio);
1634 return (chain->error);
1638 * Set the chain modified so its data can be changed by the caller, or
1639 * install deduplicated data. The caller must call this routine for each
1640 * set of modifications it makes, even if the chain is already flagged
1643 * Sets bref.modify_tid to mtid only if mtid != 0. Note that bref.modify_tid
1644 * is a CLC (cluster level change) field and is not updated by parent
1645 * propagation during a flush.
1647 * Returns an appropriate HAMMER2_ERROR_* code, which will generally reflect
1648 * chain->error except for HAMMER2_ERROR_ENOSPC. If the allocation fails
1649 * due to no space available, HAMMER2_ERROR_ENOSPC is returned and the chain
1650 * remains unmodified with its old data ref intact and chain->error
1655 * If the DEDUPABLE flag is set in the chain the storage must be reallocated
1656 * even if the chain is still flagged MODIFIED. In this case the chain's
1657 * DEDUPABLE flag will be cleared once the new storage has been assigned.
1659 * If the caller passes a non-zero dedup_off we will use it to assign the
1660 * new storage. The MODIFIED flag will be *CLEARED* in this case, and
1661 * DEDUPABLE will be set (NOTE: the UPDATE flag is always set). The caller
1662 * must not modify the data content upon return.
1665 hammer2_chain_modify(hammer2_chain_t *chain, hammer2_tid_t mtid,
1666 hammer2_off_t dedup_off, int flags)
1668 hammer2_blockref_t obref;
1679 obref = chain->bref;
1680 KKASSERT((chain->flags & HAMMER2_CHAIN_FICTITIOUS) == 0);
1681 KKASSERT(chain->lock.mtx_lock & MTX_EXCLUSIVE);
1684 * Data is not optional for freemap chains (we must always be sure
1685 * to copy the data on COW storage allocations).
1687 if (chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE ||
1688 chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) {
1689 KKASSERT((chain->flags & HAMMER2_CHAIN_INITIAL) ||
1690 (flags & HAMMER2_MODIFY_OPTDATA) == 0);
1694 * Data must be resolved if already assigned, unless explicitly
1695 * flagged otherwise. If we cannot safety load the data the
1696 * modification fails and we return early.
1698 if (chain->data == NULL && chain->bytes != 0 &&
1699 (flags & HAMMER2_MODIFY_OPTDATA) == 0 &&
1700 (chain->bref.data_off & ~HAMMER2_OFF_MASK_RADIX)) {
1701 hammer2_chain_load_data(chain);
1703 return (chain->error);
1708 * Set MODIFIED to indicate that the chain has been modified. A new
1709 * allocation is required when modifying a chain.
1711 * Set UPDATE to ensure that the blockref is updated in the parent.
1713 * If MODIFIED is already set determine if we can reuse the assigned
1714 * data block or if we need a new data block.
1716 if ((chain->flags & HAMMER2_CHAIN_MODIFIED) == 0) {
1718 * Must set modified bit.
1720 atomic_add_long(&hammer2_count_modified_chains, 1);
1721 atomic_set_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
1722 hammer2_pfs_memory_inc(chain->pmp); /* can be NULL */
1726 * We may be able to avoid a copy-on-write if the chain's
1727 * check mode is set to NONE and the chain's current
1728 * modify_tid is beyond the last explicit snapshot tid.
1730 * This implements HAMMER2's overwrite-in-place feature.
1732 * NOTE! This data-block cannot be used as a de-duplication
1733 * source when the check mode is set to NONE.
1735 if ((chain->bref.type == HAMMER2_BREF_TYPE_DATA ||
1736 chain->bref.type == HAMMER2_BREF_TYPE_DIRENT) &&
1737 (chain->flags & HAMMER2_CHAIN_INITIAL) == 0 &&
1738 (chain->flags & HAMMER2_CHAIN_DEDUPABLE) == 0 &&
1739 HAMMER2_DEC_CHECK(chain->bref.methods) ==
1740 HAMMER2_CHECK_NONE &&
1742 chain->bref.modify_tid >
1743 chain->pmp->iroot->meta.pfs_lsnap_tid) {
1745 * Sector overwrite allowed.
1750 * Sector overwrite not allowed, must copy-on-write.
1754 } else if (chain->flags & HAMMER2_CHAIN_DEDUPABLE) {
1756 * If the modified chain was registered for dedup we need
1757 * a new allocation. This only happens for delayed-flush
1758 * chains (i.e. which run through the front-end buffer
1765 * Already flagged modified, no new allocation is needed.
1772 * Flag parent update required.
1774 if ((chain->flags & HAMMER2_CHAIN_UPDATE) == 0) {
1775 atomic_set_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
1782 * The XOP code returns held but unlocked focus chains. This
1783 * prevents the chain from being destroyed but does not prevent
1784 * it from being modified. diolk is used to interlock modifications
1785 * against XOP frontend accesses to the focus.
1787 * This allows us to theoretically avoid deadlocking the frontend
1788 * if one of the backends lock up by not formally locking the
1789 * focused chain in the frontend. In addition, the synchronization
1790 * code relies on this mechanism to avoid deadlocking concurrent
1791 * synchronization threads.
1793 lockmgr(&chain->diolk, LK_EXCLUSIVE);
1796 * The modification or re-modification requires an allocation and
1797 * possible COW. If an error occurs, the previous content and data
1798 * reference is retained and the modification fails.
1800 * If dedup_off is non-zero, the caller is requesting a deduplication
1801 * rather than a modification. The MODIFIED bit is not set and the
1802 * data offset is set to the deduplication offset. The data cannot
1805 * NOTE: The dedup offset is allowed to be in a partially free state
1806 * and we must be sure to reset it to a fully allocated state
1807 * to force two bulkfree passes to free it again.
1809 * NOTE: Only applicable when chain->bytes != 0.
1811 * XXX can a chain already be marked MODIFIED without a data
1812 * assignment? If not, assert here instead of testing the case.
1814 if (chain != &hmp->vchain && chain != &hmp->fchain &&
1816 if ((chain->bref.data_off & ~HAMMER2_OFF_MASK_RADIX) == 0 ||
1820 * NOTE: We do not have to remove the dedup
1821 * registration because the area is still
1822 * allocated and the underlying DIO will
1826 chain->bref.data_off = dedup_off;
1827 chain->bytes = 1 << (dedup_off &
1828 HAMMER2_OFF_MASK_RADIX);
1830 atomic_clear_int(&chain->flags,
1831 HAMMER2_CHAIN_MODIFIED);
1832 atomic_add_long(&hammer2_count_modified_chains,
1835 hammer2_pfs_memory_wakeup(chain->pmp);
1836 hammer2_freemap_adjust(hmp, &chain->bref,
1837 HAMMER2_FREEMAP_DORECOVER);
1838 atomic_set_int(&chain->flags,
1839 HAMMER2_CHAIN_DEDUPABLE);
1841 error = hammer2_freemap_alloc(chain,
1843 atomic_clear_int(&chain->flags,
1844 HAMMER2_CHAIN_DEDUPABLE);
1850 * Stop here if error. We have to undo any flag bits we might
1855 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
1856 atomic_add_long(&hammer2_count_modified_chains, -1);
1858 hammer2_pfs_memory_wakeup(chain->pmp);
1861 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
1863 lockmgr(&chain->diolk, LK_RELEASE);
1869 * Update mirror_tid and modify_tid. modify_tid is only updated
1870 * if not passed as zero (during flushes, parent propagation passes
1873 * NOTE: chain->pmp could be the device spmp.
1875 chain->bref.mirror_tid = hmp->voldata.mirror_tid + 1;
1877 chain->bref.modify_tid = mtid;
1880 * Set BMAPUPD to tell the flush code that an existing blockmap entry
1881 * requires updating as well as to tell the delete code that the
1882 * chain's blockref might not exactly match (in terms of physical size
1883 * or block offset) the one in the parent's blocktable. The base key
1884 * of course will still match.
1886 if (chain->flags & HAMMER2_CHAIN_BMAPPED)
1887 atomic_set_int(&chain->flags, HAMMER2_CHAIN_BMAPUPD);
1890 * Short-cut data blocks which the caller does not need an actual
1891 * data reference to (aka OPTDATA), as long as the chain does not
1892 * already have a data pointer to the data. This generally means
1893 * that the modifications are being done via the logical buffer cache.
1894 * The INITIAL flag relates only to the device data buffer and thus
1895 * remains unchange in this situation.
1897 * This code also handles bytes == 0 (most dirents).
1899 if (chain->bref.type == HAMMER2_BREF_TYPE_DATA &&
1900 (flags & HAMMER2_MODIFY_OPTDATA) &&
1901 chain->data == NULL) {
1902 KKASSERT(chain->dio == NULL);
1907 * Clearing the INITIAL flag (for indirect blocks) indicates that
1908 * we've processed the uninitialized storage allocation.
1910 * If this flag is already clear we are likely in a copy-on-write
1911 * situation but we have to be sure NOT to bzero the storage if
1912 * no data is present.
1914 if (chain->flags & HAMMER2_CHAIN_INITIAL) {
1915 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1922 * Instantiate data buffer and possibly execute COW operation
1924 switch(chain->bref.type) {
1925 case HAMMER2_BREF_TYPE_VOLUME:
1926 case HAMMER2_BREF_TYPE_FREEMAP:
1928 * The data is embedded, no copy-on-write operation is
1931 KKASSERT(chain->dio == NULL);
1933 case HAMMER2_BREF_TYPE_DIRENT:
1935 * The data might be fully embedded.
1937 if (chain->bytes == 0) {
1938 KKASSERT(chain->dio == NULL);
1942 case HAMMER2_BREF_TYPE_INODE:
1943 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
1944 case HAMMER2_BREF_TYPE_DATA:
1945 case HAMMER2_BREF_TYPE_INDIRECT:
1946 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1948 * Perform the copy-on-write operation
1950 * zero-fill or copy-on-write depending on whether
1951 * chain->data exists or not and set the dirty state for
1952 * the new buffer. hammer2_io_new() will handle the
1955 * If a dedup_off was supplied this is an existing block
1956 * and no COW, copy, or further modification is required.
1958 KKASSERT(chain != &hmp->vchain && chain != &hmp->fchain);
1960 if (wasinitial && dedup_off == 0) {
1961 error = hammer2_io_new(hmp, chain->bref.type,
1962 chain->bref.data_off,
1963 chain->bytes, &dio);
1965 error = hammer2_io_bread(hmp, chain->bref.type,
1966 chain->bref.data_off,
1967 chain->bytes, &dio);
1969 hammer2_adjreadcounter(&chain->bref, chain->bytes);
1972 * If an I/O error occurs make sure callers cannot accidently
1973 * modify the old buffer's contents and corrupt the filesystem.
1975 * NOTE: hammer2_io_data() call issues bkvasync()
1978 kprintf("hammer2_chain_modify: hmp=%p I/O error\n",
1980 chain->error = HAMMER2_ERROR_EIO;
1981 hammer2_io_brelse(&dio);
1982 hammer2_io_brelse(&chain->dio);
1987 bdata = hammer2_io_data(dio, chain->bref.data_off);
1991 * COW (unless a dedup).
1993 KKASSERT(chain->dio != NULL);
1994 if (chain->data != (void *)bdata && dedup_off == 0) {
1995 bcopy(chain->data, bdata, chain->bytes);
1997 } else if (wasinitial == 0) {
1999 * We have a problem. We were asked to COW but
2000 * we don't have any data to COW with!
2002 panic("hammer2_chain_modify: having a COW %p\n",
2007 * Retire the old buffer, replace with the new. Dirty or
2008 * redirty the new buffer.
2010 * WARNING! The system buffer cache may have already flushed
2011 * the buffer, so we must be sure to [re]dirty it
2012 * for further modification.
2014 * If dedup_off was supplied, the caller is not
2015 * expected to make any further modification to the
2018 * WARNING! hammer2_get_gdata() assumes dio never transitions
2019 * through NULL in order to optimize away unnecessary
2025 if ((tio = chain->dio) != NULL)
2026 hammer2_io_bqrelse(&tio);
2027 chain->data = (void *)bdata;
2030 hammer2_io_setdirty(dio);
2034 panic("hammer2_chain_modify: illegal non-embedded type %d",
2041 * setflush on parent indicating that the parent must recurse down
2042 * to us. Do not call on chain itself which might already have it
2046 hammer2_chain_setflush(chain->parent);
2047 lockmgr(&chain->diolk, LK_RELEASE);
2049 return (chain->error);
2053 * Modify the chain associated with an inode.
2056 hammer2_chain_modify_ip(hammer2_inode_t *ip, hammer2_chain_t *chain,
2057 hammer2_tid_t mtid, int flags)
2061 hammer2_inode_modify(ip);
2062 error = hammer2_chain_modify(chain, mtid, 0, flags);
2068 * Volume header data locks
2071 hammer2_voldata_lock(hammer2_dev_t *hmp)
2073 lockmgr(&hmp->vollk, LK_EXCLUSIVE);
2077 hammer2_voldata_unlock(hammer2_dev_t *hmp)
2079 lockmgr(&hmp->vollk, LK_RELEASE);
2083 hammer2_voldata_modify(hammer2_dev_t *hmp)
2085 if ((hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) == 0) {
2086 atomic_add_long(&hammer2_count_modified_chains, 1);
2087 atomic_set_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
2088 hammer2_pfs_memory_inc(hmp->vchain.pmp);
2093 * This function returns the chain at the nearest key within the specified
2094 * range. The returned chain will be referenced but not locked.
2096 * This function will recurse through chain->rbtree as necessary and will
2097 * return a *key_nextp suitable for iteration. *key_nextp is only set if
2098 * the iteration value is less than the current value of *key_nextp.
2100 * The caller should use (*key_nextp) to calculate the actual range of
2101 * the returned element, which will be (key_beg to *key_nextp - 1), because
2102 * there might be another element which is superior to the returned element
2105 * (*key_nextp) can be passed as key_beg in an iteration only while non-NULL
2106 * chains continue to be returned. On EOF (*key_nextp) may overflow since
2107 * it will wind up being (key_end + 1).
2109 * WARNING! Must be called with child's spinlock held. Spinlock remains
2110 * held through the operation.
2112 struct hammer2_chain_find_info {
2113 hammer2_chain_t *best;
2114 hammer2_key_t key_beg;
2115 hammer2_key_t key_end;
2116 hammer2_key_t key_next;
2119 static int hammer2_chain_find_cmp(hammer2_chain_t *child, void *data);
2120 static int hammer2_chain_find_callback(hammer2_chain_t *child, void *data);
2124 hammer2_chain_find(hammer2_chain_t *parent, hammer2_key_t *key_nextp,
2125 hammer2_key_t key_beg, hammer2_key_t key_end)
2127 struct hammer2_chain_find_info info;
2130 info.key_beg = key_beg;
2131 info.key_end = key_end;
2132 info.key_next = *key_nextp;
2134 RB_SCAN(hammer2_chain_tree, &parent->core.rbtree,
2135 hammer2_chain_find_cmp, hammer2_chain_find_callback,
2137 *key_nextp = info.key_next;
2139 kprintf("chain_find %p %016jx:%016jx next=%016jx\n",
2140 parent, key_beg, key_end, *key_nextp);
2148 hammer2_chain_find_cmp(hammer2_chain_t *child, void *data)
2150 struct hammer2_chain_find_info *info = data;
2151 hammer2_key_t child_beg;
2152 hammer2_key_t child_end;
2154 child_beg = child->bref.key;
2155 child_end = child_beg + ((hammer2_key_t)1 << child->bref.keybits) - 1;
2157 if (child_end < info->key_beg)
2159 if (child_beg > info->key_end)
2166 hammer2_chain_find_callback(hammer2_chain_t *child, void *data)
2168 struct hammer2_chain_find_info *info = data;
2169 hammer2_chain_t *best;
2170 hammer2_key_t child_end;
2173 * WARNING! Layerq is scanned forwards, exact matches should keep
2174 * the existing info->best.
2176 if ((best = info->best) == NULL) {
2178 * No previous best. Assign best
2181 } else if (best->bref.key <= info->key_beg &&
2182 child->bref.key <= info->key_beg) {
2187 /*info->best = child;*/
2188 } else if (child->bref.key < best->bref.key) {
2190 * Child has a nearer key and best is not flush with key_beg.
2191 * Set best to child. Truncate key_next to the old best key.
2194 if (info->key_next > best->bref.key || info->key_next == 0)
2195 info->key_next = best->bref.key;
2196 } else if (child->bref.key == best->bref.key) {
2198 * If our current best is flush with the child then this
2199 * is an illegal overlap.
2201 * key_next will automatically be limited to the smaller of
2202 * the two end-points.
2208 * Keep the current best but truncate key_next to the child's
2211 * key_next will also automatically be limited to the smaller
2212 * of the two end-points (probably not necessary for this case
2213 * but we do it anyway).
2215 if (info->key_next > child->bref.key || info->key_next == 0)
2216 info->key_next = child->bref.key;
2220 * Always truncate key_next based on child's end-of-range.
2222 child_end = child->bref.key + ((hammer2_key_t)1 << child->bref.keybits);
2223 if (child_end && (info->key_next > child_end || info->key_next == 0))
2224 info->key_next = child_end;
2230 * Retrieve the specified chain from a media blockref, creating the
2231 * in-memory chain structure which reflects it. The returned chain is
2232 * held and locked according to (how) (HAMMER2_RESOLVE_*). The caller must
2233 * handle crc-checks and so forth, and should check chain->error before
2234 * assuming that the data is good.
2236 * To handle insertion races pass the INSERT_RACE flag along with the
2237 * generation number of the core. NULL will be returned if the generation
2238 * number changes before we have a chance to insert the chain. Insert
2239 * races can occur because the parent might be held shared.
2241 * Caller must hold the parent locked shared or exclusive since we may
2242 * need the parent's bref array to find our block.
2244 * WARNING! chain->pmp is always set to NULL for any chain representing
2245 * part of the super-root topology.
2248 hammer2_chain_get(hammer2_chain_t *parent, int generation,
2249 hammer2_blockref_t *bref, int how)
2251 hammer2_dev_t *hmp = parent->hmp;
2252 hammer2_chain_t *chain;
2256 * Allocate a chain structure representing the existing media
2257 * entry. Resulting chain has one ref and is not locked.
2259 if (bref->flags & HAMMER2_BREF_FLAG_PFSROOT)
2260 chain = hammer2_chain_alloc(hmp, NULL, bref);
2262 chain = hammer2_chain_alloc(hmp, parent->pmp, bref);
2263 /* ref'd chain returned */
2266 * Flag that the chain is in the parent's blockmap so delete/flush
2267 * knows what to do with it.
2269 atomic_set_int(&chain->flags, HAMMER2_CHAIN_BMAPPED);
2272 * chain must be locked to avoid unexpected ripouts
2274 hammer2_chain_lock(chain, how);
2277 * Link the chain into its parent. A spinlock is required to safely
2278 * access the RBTREE, and it is possible to collide with another
2279 * hammer2_chain_get() operation because the caller might only hold
2280 * a shared lock on the parent.
2282 * NOTE: Get races can occur quite often when we distribute
2283 * asynchronous read-aheads across multiple threads.
2285 KKASSERT(parent->refs > 0);
2286 error = hammer2_chain_insert(parent, chain,
2287 HAMMER2_CHAIN_INSERT_SPIN |
2288 HAMMER2_CHAIN_INSERT_RACE,
2291 KKASSERT((chain->flags & HAMMER2_CHAIN_ONRBTREE) == 0);
2292 /*kprintf("chain %p get race\n", chain);*/
2293 hammer2_chain_unlock(chain);
2294 hammer2_chain_drop(chain);
2297 KKASSERT(chain->flags & HAMMER2_CHAIN_ONRBTREE);
2301 * Return our new chain referenced but not locked, or NULL if
2308 * Lookup initialization/completion API
2311 hammer2_chain_lookup_init(hammer2_chain_t *parent, int flags)
2313 hammer2_chain_ref(parent);
2314 if (flags & HAMMER2_LOOKUP_SHARED) {
2315 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS |
2316 HAMMER2_RESOLVE_SHARED);
2318 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2324 hammer2_chain_lookup_done(hammer2_chain_t *parent)
2327 hammer2_chain_unlock(parent);
2328 hammer2_chain_drop(parent);
2333 * Take the locked chain and return a locked parent. The chain remains
2334 * locked on return, but may have to be temporarily unlocked to acquire
2335 * the parent. Because of this, (chain) must be stable and cannot be
2336 * deleted while it was temporarily unlocked (typically means that (chain)
2339 * Pass HAMMER2_RESOLVE_* flags in flags.
2341 * This will work even if the chain is errored, and the caller can check
2342 * parent->error on return if desired since the parent will be locked.
2344 * This function handles the lock order reversal.
2347 hammer2_chain_getparent(hammer2_chain_t *chain, int flags)
2349 hammer2_chain_t *parent;
2352 * Be careful of order, chain must be unlocked before parent
2353 * is locked below to avoid a deadlock. Try it trivially first.
2355 parent = chain->parent;
2357 panic("hammer2_chain_getparent: no parent");
2358 hammer2_chain_ref(parent);
2359 if (hammer2_chain_lock(parent, flags|HAMMER2_RESOLVE_NONBLOCK) == 0)
2363 hammer2_chain_unlock(chain);
2364 hammer2_chain_lock(parent, flags);
2365 hammer2_chain_lock(chain, flags);
2368 * Parent relinking races are quite common. We have to get
2369 * it right or we will blow up the block table.
2371 if (chain->parent == parent)
2373 hammer2_chain_unlock(parent);
2374 hammer2_chain_drop(parent);
2376 parent = chain->parent;
2378 panic("hammer2_chain_getparent: no parent");
2379 hammer2_chain_ref(parent);
2385 * Take the locked chain and return a locked parent. The chain is unlocked
2386 * and dropped. *chainp is set to the returned parent as a convenience.
2387 * Pass HAMMER2_RESOLVE_* flags in flags.
2389 * This will work even if the chain is errored, and the caller can check
2390 * parent->error on return if desired since the parent will be locked.
2392 * The chain does NOT need to be stable. We use a tracking structure
2393 * to track the expected parent if the chain is deleted out from under us.
2395 * This function handles the lock order reversal.
2398 hammer2_chain_repparent(hammer2_chain_t **chainp, int flags)
2400 hammer2_chain_t *chain;
2401 hammer2_chain_t *parent;
2402 struct hammer2_reptrack reptrack;
2403 struct hammer2_reptrack **repp;
2406 * Be careful of order, chain must be unlocked before parent
2407 * is locked below to avoid a deadlock. Try it trivially first.
2410 parent = chain->parent;
2411 if (parent == NULL) {
2412 hammer2_spin_unex(&chain->core.spin);
2413 panic("hammer2_chain_repparent: no parent");
2415 hammer2_chain_ref(parent);
2416 if (hammer2_chain_lock(parent, flags|HAMMER2_RESOLVE_NONBLOCK) == 0) {
2417 hammer2_chain_unlock(chain);
2418 hammer2_chain_drop(chain);
2425 * Ok, now it gets a bit nasty. There are multiple situations where
2426 * the parent might be in the middle of a deletion, or where the child
2427 * (chain) might be deleted the instant we let go of its lock.
2428 * We can potentially end up in a no-win situation!
2430 * In particular, the indirect_maintenance() case can cause these
2433 * To deal with this we install a reptrack structure in the parent
2434 * This reptrack structure 'owns' the parent ref and will automatically
2435 * migrate to the parent's parent if the parent is deleted permanently.
2437 hammer2_spin_init(&reptrack.spin, "h2reptrk");
2438 reptrack.chain = parent;
2439 hammer2_chain_ref(parent); /* for the reptrack */
2441 hammer2_spin_ex(&parent->core.spin);
2442 reptrack.next = parent->core.reptrack;
2443 parent->core.reptrack = &reptrack;
2444 hammer2_spin_unex(&parent->core.spin);
2446 hammer2_chain_unlock(chain);
2447 hammer2_chain_drop(chain);
2448 chain = NULL; /* gone */
2451 * At the top of this loop, chain is gone and parent is refd both
2452 * by us explicitly AND via our reptrack. We are attempting to
2456 hammer2_chain_lock(parent, flags);
2458 if (reptrack.chain == parent)
2460 hammer2_chain_unlock(parent);
2461 hammer2_chain_drop(parent);
2463 kprintf("hammer2: debug REPTRACK %p->%p\n",
2464 parent, reptrack.chain);
2465 hammer2_spin_ex(&reptrack.spin);
2466 parent = reptrack.chain;
2467 hammer2_chain_ref(parent);
2468 hammer2_spin_unex(&reptrack.spin);
2472 * Once parent is locked and matches our reptrack, our reptrack
2473 * will be stable and we have our parent. We can unlink our
2476 * WARNING! Remember that the chain lock might be shared. Chains
2477 * locked shared have stable parent linkages.
2479 hammer2_spin_ex(&parent->core.spin);
2480 repp = &parent->core.reptrack;
2481 while (*repp != &reptrack)
2482 repp = &(*repp)->next;
2483 *repp = reptrack.next;
2484 hammer2_spin_unex(&parent->core.spin);
2486 hammer2_chain_drop(parent); /* reptrack ref */
2487 *chainp = parent; /* return parent lock+ref */
2493 * Dispose of any linked reptrack structures in (chain) by shifting them to
2494 * (parent). Both (chain) and (parent) must be exclusively locked.
2496 * This is interlocked against any children of (chain) on the other side.
2497 * No children so remain as-of when this is called so we can test
2498 * core.reptrack without holding the spin-lock.
2500 * Used whenever the caller intends to permanently delete chains related
2501 * to topological recursions (BREF_TYPE_INDIRECT, BREF_TYPE_FREEMAP_NODE),
2502 * where the chains underneath the node being deleted are given a new parent
2503 * above the node being deleted.
2507 hammer2_chain_repchange(hammer2_chain_t *parent, hammer2_chain_t *chain)
2509 struct hammer2_reptrack *reptrack;
2511 KKASSERT(chain->core.live_count == 0 && RB_EMPTY(&chain->core.rbtree));
2512 while (chain->core.reptrack) {
2513 hammer2_spin_ex(&parent->core.spin);
2514 hammer2_spin_ex(&chain->core.spin);
2515 reptrack = chain->core.reptrack;
2516 if (reptrack == NULL) {
2517 hammer2_spin_unex(&chain->core.spin);
2518 hammer2_spin_unex(&parent->core.spin);
2521 hammer2_spin_ex(&reptrack->spin);
2522 chain->core.reptrack = reptrack->next;
2523 reptrack->chain = parent;
2524 reptrack->next = parent->core.reptrack;
2525 parent->core.reptrack = reptrack;
2526 hammer2_chain_ref(parent); /* reptrack */
2528 hammer2_spin_unex(&chain->core.spin);
2529 hammer2_spin_unex(&parent->core.spin);
2530 kprintf("hammer2: debug repchange %p %p->%p\n",
2531 reptrack, chain, parent);
2532 hammer2_chain_drop(chain); /* reptrack */
2537 * Locate the first chain whos key range overlaps (key_beg, key_end) inclusive.
2538 * (*parentp) typically points to an inode but can also point to a related
2539 * indirect block and this function will recurse upwards and find the inode
2540 * or the nearest undeleted indirect block covering the key range.
2542 * This function unconditionally sets *errorp, replacing any previous value.
2544 * (*parentp) must be exclusive or shared locked (depending on flags) and
2545 * referenced and can be an inode or an existing indirect block within the
2548 * If (*parent) is errored out, this function will not attempt to recurse
2549 * the radix tree and will return NULL along with an appropriate *errorp.
2550 * If NULL is returned and *errorp is 0, the requested lookup could not be
2553 * On return (*parentp) will be modified to point at the deepest parent chain
2554 * element encountered during the search, as a helper for an insertion or
2557 * The new (*parentp) will be locked shared or exclusive (depending on flags),
2558 * and referenced, and the old will be unlocked and dereferenced (no change
2559 * if they are both the same). This is particularly important if the caller
2560 * wishes to insert a new chain, (*parentp) will be set properly even if NULL
2561 * is returned, as long as no error occurred.
2563 * The matching chain will be returned locked according to flags.
2567 * NULL is returned if no match was found, but (*parentp) will still
2568 * potentially be adjusted.
2570 * On return (*key_nextp) will point to an iterative value for key_beg.
2571 * (If NULL is returned (*key_nextp) is set to (key_end + 1)).
2573 * This function will also recurse up the chain if the key is not within the
2574 * current parent's range. (*parentp) can never be set to NULL. An iteration
2575 * can simply allow (*parentp) to float inside the loop.
2577 * NOTE! chain->data is not always resolved. By default it will not be
2578 * resolved for BREF_TYPE_DATA, FREEMAP_NODE, or FREEMAP_LEAF. Use
2579 * HAMMER2_LOOKUP_ALWAYS to force resolution (but be careful w/
2580 * BREF_TYPE_DATA as the device buffer can alias the logical file
2585 hammer2_chain_lookup(hammer2_chain_t **parentp, hammer2_key_t *key_nextp,
2586 hammer2_key_t key_beg, hammer2_key_t key_end,
2587 int *errorp, int flags)
2590 hammer2_chain_t *parent;
2591 hammer2_chain_t *chain;
2592 hammer2_blockref_t *base;
2593 hammer2_blockref_t *bref;
2594 hammer2_blockref_t bcopy;
2595 hammer2_key_t scan_beg;
2596 hammer2_key_t scan_end;
2598 int how_always = HAMMER2_RESOLVE_ALWAYS;
2599 int how_maybe = HAMMER2_RESOLVE_MAYBE;
2602 int maxloops = 300000;
2603 volatile hammer2_mtx_t save_mtx;
2605 if (flags & HAMMER2_LOOKUP_ALWAYS) {
2606 how_maybe = how_always;
2607 how = HAMMER2_RESOLVE_ALWAYS;
2608 } else if (flags & HAMMER2_LOOKUP_NODATA) {
2609 how = HAMMER2_RESOLVE_NEVER;
2611 how = HAMMER2_RESOLVE_MAYBE;
2613 if (flags & HAMMER2_LOOKUP_SHARED) {
2614 how_maybe |= HAMMER2_RESOLVE_SHARED;
2615 how_always |= HAMMER2_RESOLVE_SHARED;
2616 how |= HAMMER2_RESOLVE_SHARED;
2620 * Recurse (*parentp) upward if necessary until the parent completely
2621 * encloses the key range or we hit the inode.
2623 * Handle races against the flusher deleting indirect nodes on its
2624 * way back up by continuing to recurse upward past the deletion.
2630 while (parent->bref.type == HAMMER2_BREF_TYPE_INDIRECT ||
2631 parent->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2632 scan_beg = parent->bref.key;
2633 scan_end = scan_beg +
2634 ((hammer2_key_t)1 << parent->bref.keybits) - 1;
2635 if ((parent->flags & HAMMER2_CHAIN_DELETED) == 0) {
2636 if (key_beg >= scan_beg && key_end <= scan_end)
2639 parent = hammer2_chain_repparent(parentp, how_maybe);
2642 if (--maxloops == 0)
2643 panic("hammer2_chain_lookup: maxloops");
2645 * Locate the blockref array. Currently we do a fully associative
2646 * search through the array.
2648 switch(parent->bref.type) {
2649 case HAMMER2_BREF_TYPE_INODE:
2651 * Special shortcut for embedded data returns the inode
2652 * itself. Callers must detect this condition and access
2653 * the embedded data (the strategy code does this for us).
2655 * This is only applicable to regular files and softlinks.
2657 * We need a second lock on parent. Since we already have
2658 * a lock we must pass LOCKAGAIN to prevent unexpected
2659 * blocking (we don't want to block on a second shared
2660 * ref if an exclusive lock is pending)
2662 if (parent->data->ipdata.meta.op_flags &
2663 HAMMER2_OPFLAG_DIRECTDATA) {
2664 if (flags & HAMMER2_LOOKUP_NODIRECT) {
2666 *key_nextp = key_end + 1;
2669 hammer2_chain_ref(parent);
2670 hammer2_chain_lock(parent, how_always |
2671 HAMMER2_RESOLVE_LOCKAGAIN);
2672 *key_nextp = key_end + 1;
2675 base = &parent->data->ipdata.u.blockset.blockref[0];
2676 count = HAMMER2_SET_COUNT;
2678 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2679 case HAMMER2_BREF_TYPE_INDIRECT:
2681 * Handle MATCHIND on the parent
2683 if (flags & HAMMER2_LOOKUP_MATCHIND) {
2684 scan_beg = parent->bref.key;
2685 scan_end = scan_beg +
2686 ((hammer2_key_t)1 << parent->bref.keybits) - 1;
2687 if (key_beg == scan_beg && key_end == scan_end) {
2689 hammer2_chain_ref(chain);
2690 hammer2_chain_lock(chain, how_maybe);
2691 *key_nextp = scan_end + 1;
2697 * Optimize indirect blocks in the INITIAL state to avoid
2700 if (parent->flags & HAMMER2_CHAIN_INITIAL) {
2703 if (parent->data == NULL) {
2704 kprintf("parent->data is NULL %p\n", parent);
2706 tsleep(parent, 0, "xxx", 0);
2708 base = &parent->data->npdata[0];
2710 count = parent->bytes / sizeof(hammer2_blockref_t);
2712 case HAMMER2_BREF_TYPE_VOLUME:
2713 base = &parent->data->voldata.sroot_blockset.blockref[0];
2714 count = HAMMER2_SET_COUNT;
2716 case HAMMER2_BREF_TYPE_FREEMAP:
2717 base = &parent->data->blkset.blockref[0];
2718 count = HAMMER2_SET_COUNT;
2721 kprintf("hammer2_chain_lookup: unrecognized "
2722 "blockref(B) type: %d",
2725 tsleep(&base, 0, "dead", 0);
2726 panic("hammer2_chain_lookup: unrecognized "
2727 "blockref(B) type: %d",
2729 base = NULL; /* safety */
2730 count = 0; /* safety */
2734 * No lookup is possible if the parent is errored. We delayed
2735 * this check as long as we could to ensure that the parent backup,
2736 * embedded data, and MATCHIND code could still execute.
2738 if (parent->error) {
2739 *errorp = parent->error;
2744 * Merged scan to find next candidate.
2746 * hammer2_base_*() functions require the parent->core.live_* fields
2747 * to be synchronized.
2749 * We need to hold the spinlock to access the block array and RB tree
2750 * and to interlock chain creation.
2752 if ((parent->flags & HAMMER2_CHAIN_COUNTEDBREFS) == 0)
2753 hammer2_chain_countbrefs(parent, base, count);
2758 hammer2_spin_ex(&parent->core.spin);
2759 chain = hammer2_combined_find(parent, base, count,
2763 generation = parent->core.generation;
2766 * Exhausted parent chain, iterate.
2769 KKASSERT(chain == NULL);
2770 hammer2_spin_unex(&parent->core.spin);
2771 if (key_beg == key_end) /* short cut single-key case */
2775 * Stop if we reached the end of the iteration.
2777 if (parent->bref.type != HAMMER2_BREF_TYPE_INDIRECT &&
2778 parent->bref.type != HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2783 * Calculate next key, stop if we reached the end of the
2784 * iteration, otherwise go up one level and loop.
2786 key_beg = parent->bref.key +
2787 ((hammer2_key_t)1 << parent->bref.keybits);
2788 if (key_beg == 0 || key_beg > key_end)
2790 parent = hammer2_chain_repparent(parentp, how_maybe);
2795 * Selected from blockref or in-memory chain.
2798 if (chain == NULL) {
2799 hammer2_spin_unex(&parent->core.spin);
2800 if (bcopy.type == HAMMER2_BREF_TYPE_INDIRECT ||
2801 bcopy.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2802 chain = hammer2_chain_get(parent, generation,
2805 chain = hammer2_chain_get(parent, generation,
2811 hammer2_chain_ref(chain);
2812 hammer2_spin_unex(&parent->core.spin);
2815 * chain is referenced but not locked. We must lock the
2816 * chain to obtain definitive state.
2818 if (bcopy.type == HAMMER2_BREF_TYPE_INDIRECT ||
2819 bcopy.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2820 hammer2_chain_lock(chain, how_maybe);
2822 hammer2_chain_lock(chain, how);
2824 KKASSERT(chain->parent == parent);
2826 if (bcmp(&bcopy, &chain->bref, sizeof(bcopy)) ||
2827 chain->parent != parent) {
2828 hammer2_chain_unlock(chain);
2829 hammer2_chain_drop(chain);
2830 chain = NULL; /* SAFETY */
2836 * Skip deleted chains (XXX cache 'i' end-of-block-array? XXX)
2838 * NOTE: Chain's key range is not relevant as there might be
2839 * one-offs within the range that are not deleted.
2841 * NOTE: Lookups can race delete-duplicate because
2842 * delete-duplicate does not lock the parent's core
2843 * (they just use the spinlock on the core).
2845 if (chain->flags & HAMMER2_CHAIN_DELETED) {
2846 kprintf("skip deleted chain %016jx.%02x key=%016jx\n",
2847 chain->bref.data_off, chain->bref.type,
2849 hammer2_chain_unlock(chain);
2850 hammer2_chain_drop(chain);
2851 chain = NULL; /* SAFETY */
2852 key_beg = *key_nextp;
2853 if (key_beg == 0 || key_beg > key_end)
2859 * If the chain element is an indirect block it becomes the new
2860 * parent and we loop on it. We must maintain our top-down locks
2861 * to prevent the flusher from interfering (i.e. doing a
2862 * delete-duplicate and leaving us recursing down a deleted chain).
2864 * The parent always has to be locked with at least RESOLVE_MAYBE
2865 * so we can access its data. It might need a fixup if the caller
2866 * passed incompatible flags. Be careful not to cause a deadlock
2867 * as a data-load requires an exclusive lock.
2869 * If HAMMER2_LOOKUP_MATCHIND is set and the indirect block's key
2870 * range is within the requested key range we return the indirect
2871 * block and do NOT loop. This is usually only used to acquire
2874 if (chain->bref.type == HAMMER2_BREF_TYPE_INDIRECT ||
2875 chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2876 save_mtx = parent->lock;
2877 hammer2_chain_unlock(parent);
2878 hammer2_chain_drop(parent);
2879 *parentp = parent = chain;
2880 chain = NULL; /* SAFETY */
2885 * All done, return the locked chain.
2887 * If the caller does not want a locked chain, replace the lock with
2888 * a ref. Perhaps this can eventually be optimized to not obtain the
2889 * lock in the first place for situations where the data does not
2890 * need to be resolved.
2892 * NOTE! A chain->error must be tested by the caller upon return.
2893 * *errorp is only set based on issues which occur while
2894 * trying to reach the chain.
2900 * After having issued a lookup we can iterate all matching keys.
2902 * If chain is non-NULL we continue the iteration from just after it's index.
2904 * If chain is NULL we assume the parent was exhausted and continue the
2905 * iteration at the next parent.
2907 * If a fatal error occurs (typically an I/O error), a dummy chain is
2908 * returned with chain->error and error-identifying information set. This
2909 * chain will assert if you try to do anything fancy with it.
2911 * XXX Depending on where the error occurs we should allow continued iteration.
2913 * parent must be locked on entry and remains locked throughout. chain's
2914 * lock status must match flags. Chain is always at least referenced.
2916 * WARNING! The MATCHIND flag does not apply to this function.
2919 hammer2_chain_next(hammer2_chain_t **parentp, hammer2_chain_t *chain,
2920 hammer2_key_t *key_nextp,
2921 hammer2_key_t key_beg, hammer2_key_t key_end,
2922 int *errorp, int flags)
2924 hammer2_chain_t *parent;
2928 * Calculate locking flags for upward recursion.
2930 how_maybe = HAMMER2_RESOLVE_MAYBE;
2931 if (flags & HAMMER2_LOOKUP_SHARED)
2932 how_maybe |= HAMMER2_RESOLVE_SHARED;
2938 * Calculate the next index and recalculate the parent if necessary.
2941 key_beg = chain->bref.key +
2942 ((hammer2_key_t)1 << chain->bref.keybits);
2943 hammer2_chain_unlock(chain);
2944 hammer2_chain_drop(chain);
2947 * chain invalid past this point, but we can still do a
2948 * pointer comparison w/parent.
2950 * Any scan where the lookup returned degenerate data embedded
2951 * in the inode has an invalid index and must terminate.
2953 if (chain == parent)
2955 if (key_beg == 0 || key_beg > key_end)
2958 } else if (parent->bref.type != HAMMER2_BREF_TYPE_INDIRECT &&
2959 parent->bref.type != HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2961 * We reached the end of the iteration.
2966 * Continue iteration with next parent unless the current
2967 * parent covers the range.
2969 * (This also handles the case of a deleted, empty indirect
2972 key_beg = parent->bref.key +
2973 ((hammer2_key_t)1 << parent->bref.keybits);
2974 if (key_beg == 0 || key_beg > key_end)
2976 parent = hammer2_chain_repparent(parentp, how_maybe);
2982 return (hammer2_chain_lookup(parentp, key_nextp,
2988 * Caller wishes to iterate chains under parent, loading new chains into
2989 * chainp. Caller must initialize *chainp to NULL and *firstp to 1, and
2990 * then call hammer2_chain_scan() repeatedly until a non-zero return.
2991 * During the scan, *firstp will be set to 0 and (*chainp) will be replaced
2992 * with the returned chain for the scan. The returned *chainp will be
2993 * locked and referenced. Any prior contents will be unlocked and dropped.
2995 * Caller should check the return value. A normal scan EOF will return
2996 * exactly HAMMER2_ERROR_EOF. Any other non-zero value indicates an
2997 * error trying to access parent data. Any error in the returned chain
2998 * must be tested separately by the caller.
3000 * (*chainp) is dropped on each scan, but will only be set if the returned
3001 * element itself can recurse. Leaf elements are NOT resolved, loaded, or
3002 * returned via *chainp. The caller will get their bref only.
3004 * The raw scan function is similar to lookup/next but does not seek to a key.
3005 * Blockrefs are iterated via first_bref = (parent, NULL) and
3006 * next_chain = (parent, bref).
3008 * The passed-in parent must be locked and its data resolved. The function
3009 * nominally returns a locked and referenced *chainp != NULL for chains
3010 * the caller might need to recurse on (and will dipose of any *chainp passed
3011 * in). The caller must check the chain->bref.type either way.
3014 hammer2_chain_scan(hammer2_chain_t *parent, hammer2_chain_t **chainp,
3015 hammer2_blockref_t *bref, int *firstp,
3019 hammer2_blockref_t *base;
3020 hammer2_blockref_t *bref_ptr;
3022 hammer2_key_t next_key;
3023 hammer2_chain_t *chain = NULL;
3025 int how_always = HAMMER2_RESOLVE_ALWAYS;
3026 int how_maybe = HAMMER2_RESOLVE_MAYBE;
3029 int maxloops = 300000;
3036 * Scan flags borrowed from lookup.
3038 if (flags & HAMMER2_LOOKUP_ALWAYS) {
3039 how_maybe = how_always;
3040 how = HAMMER2_RESOLVE_ALWAYS;
3041 } else if (flags & HAMMER2_LOOKUP_NODATA) {
3042 how = HAMMER2_RESOLVE_NEVER;
3044 how = HAMMER2_RESOLVE_MAYBE;
3046 if (flags & HAMMER2_LOOKUP_SHARED) {
3047 how_maybe |= HAMMER2_RESOLVE_SHARED;
3048 how_always |= HAMMER2_RESOLVE_SHARED;
3049 how |= HAMMER2_RESOLVE_SHARED;
3053 * Calculate key to locate first/next element, unlocking the previous
3054 * element as we go. Be careful, the key calculation can overflow.
3056 * (also reset bref to NULL)
3062 key = bref->key + ((hammer2_key_t)1 << bref->keybits);
3063 if ((chain = *chainp) != NULL) {
3065 hammer2_chain_unlock(chain);
3066 hammer2_chain_drop(chain);
3070 error |= HAMMER2_ERROR_EOF;
3076 if (parent->error) {
3077 error = parent->error;
3080 if (--maxloops == 0)
3081 panic("hammer2_chain_scan: maxloops");
3084 * Locate the blockref array. Currently we do a fully associative
3085 * search through the array.
3087 switch(parent->bref.type) {
3088 case HAMMER2_BREF_TYPE_INODE:
3090 * An inode with embedded data has no sub-chains.
3092 * WARNING! Bulk scan code may pass a static chain marked
3093 * as BREF_TYPE_INODE with a copy of the volume
3094 * root blockset to snapshot the volume.
3096 if (parent->data->ipdata.meta.op_flags &
3097 HAMMER2_OPFLAG_DIRECTDATA) {
3098 error |= HAMMER2_ERROR_EOF;
3101 base = &parent->data->ipdata.u.blockset.blockref[0];
3102 count = HAMMER2_SET_COUNT;
3104 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3105 case HAMMER2_BREF_TYPE_INDIRECT:
3107 * Optimize indirect blocks in the INITIAL state to avoid
3110 if (parent->flags & HAMMER2_CHAIN_INITIAL) {
3113 if (parent->data == NULL)
3114 panic("parent->data is NULL");
3115 base = &parent->data->npdata[0];
3117 count = parent->bytes / sizeof(hammer2_blockref_t);
3119 case HAMMER2_BREF_TYPE_VOLUME:
3120 base = &parent->data->voldata.sroot_blockset.blockref[0];
3121 count = HAMMER2_SET_COUNT;
3123 case HAMMER2_BREF_TYPE_FREEMAP:
3124 base = &parent->data->blkset.blockref[0];
3125 count = HAMMER2_SET_COUNT;
3128 panic("hammer2_chain_scan: unrecognized blockref type: %d",
3130 base = NULL; /* safety */
3131 count = 0; /* safety */
3135 * Merged scan to find next candidate.
3137 * hammer2_base_*() functions require the parent->core.live_* fields
3138 * to be synchronized.
3140 * We need to hold the spinlock to access the block array and RB tree
3141 * and to interlock chain creation.
3143 if ((parent->flags & HAMMER2_CHAIN_COUNTEDBREFS) == 0)
3144 hammer2_chain_countbrefs(parent, base, count);
3148 hammer2_spin_ex(&parent->core.spin);
3149 chain = hammer2_combined_find(parent, base, count,
3151 key, HAMMER2_KEY_MAX,
3153 generation = parent->core.generation;
3156 * Exhausted parent chain, we're done.
3158 if (bref_ptr == NULL) {
3159 hammer2_spin_unex(&parent->core.spin);
3160 KKASSERT(chain == NULL);
3161 error |= HAMMER2_ERROR_EOF;
3166 * Copy into the supplied stack-based blockref.
3171 * Selected from blockref or in-memory chain.
3173 if (chain == NULL) {
3174 switch(bref->type) {
3175 case HAMMER2_BREF_TYPE_INODE:
3176 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3177 case HAMMER2_BREF_TYPE_INDIRECT:
3178 case HAMMER2_BREF_TYPE_VOLUME:
3179 case HAMMER2_BREF_TYPE_FREEMAP:
3181 * Recursion, always get the chain
3183 hammer2_spin_unex(&parent->core.spin);
3184 chain = hammer2_chain_get(parent, generation,
3191 * No recursion, do not waste time instantiating
3192 * a chain, just iterate using the bref.
3194 hammer2_spin_unex(&parent->core.spin);
3199 * Recursion or not we need the chain in order to supply
3202 hammer2_chain_ref(chain);
3203 hammer2_spin_unex(&parent->core.spin);
3204 hammer2_chain_lock(chain, how);
3207 (bcmp(bref, &chain->bref, sizeof(*bref)) ||
3208 chain->parent != parent)) {
3209 hammer2_chain_unlock(chain);
3210 hammer2_chain_drop(chain);
3216 * Skip deleted chains (XXX cache 'i' end-of-block-array? XXX)
3218 * NOTE: chain's key range is not relevant as there might be
3219 * one-offs within the range that are not deleted.
3221 * NOTE: XXX this could create problems with scans used in
3222 * situations other than mount-time recovery.
3224 * NOTE: Lookups can race delete-duplicate because
3225 * delete-duplicate does not lock the parent's core
3226 * (they just use the spinlock on the core).
3228 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) {
3229 hammer2_chain_unlock(chain);
3230 hammer2_chain_drop(chain);
3235 error |= HAMMER2_ERROR_EOF;
3243 * All done, return the bref or NULL, supply chain if necessary.
3251 * Create and return a new hammer2 system memory structure of the specified
3252 * key, type and size and insert it under (*parentp). This is a full
3253 * insertion, based on the supplied key/keybits, and may involve creating
3254 * indirect blocks and moving other chains around via delete/duplicate.
3256 * This call can be made with parent == NULL as long as a non -1 methods
3257 * is supplied. hmp must also be supplied in this situation (otherwise
3258 * hmp is extracted from the supplied parent). The chain will be detached
3259 * from the topology. A later call with both parent and chain can be made
3262 * THE CALLER MUST HAVE ALREADY PROPERLY SEEKED (*parentp) TO THE INSERTION
3263 * POINT SANS ANY REQUIRED INDIRECT BLOCK CREATIONS DUE TO THE ARRAY BEING
3264 * FULL. This typically means that the caller is creating the chain after
3265 * doing a hammer2_chain_lookup().
3267 * (*parentp) must be exclusive locked and may be replaced on return
3268 * depending on how much work the function had to do.
3270 * (*parentp) must not be errored or this function will assert.
3272 * (*chainp) usually starts out NULL and returns the newly created chain,
3273 * but if the caller desires the caller may allocate a disconnected chain
3274 * and pass it in instead.
3276 * This function should NOT be used to insert INDIRECT blocks. It is
3277 * typically used to create/insert inodes and data blocks.
3279 * Caller must pass-in an exclusively locked parent the new chain is to
3280 * be inserted under, and optionally pass-in a disconnected, exclusively
3281 * locked chain to insert (else we create a new chain). The function will
3282 * adjust (*parentp) as necessary, create or connect the chain, and
3283 * return an exclusively locked chain in *chainp.
3285 * When creating a PFSROOT inode under the super-root, pmp is typically NULL
3286 * and will be reassigned.
3288 * NOTE: returns HAMMER_ERROR_* flags
3291 hammer2_chain_create(hammer2_chain_t **parentp, hammer2_chain_t **chainp,
3292 hammer2_dev_t *hmp, hammer2_pfs_t *pmp, int methods,
3293 hammer2_key_t key, int keybits, int type, size_t bytes,
3294 hammer2_tid_t mtid, hammer2_off_t dedup_off, int flags)
3296 hammer2_chain_t *chain;
3297 hammer2_chain_t *parent;
3298 hammer2_blockref_t *base;
3299 hammer2_blockref_t dummy;
3303 int maxloops = 300000;
3306 * Topology may be crossing a PFS boundary.
3310 KKASSERT(hammer2_mtx_owned(&parent->lock));
3311 KKASSERT(parent->error == 0);
3316 if (chain == NULL) {
3318 * First allocate media space and construct the dummy bref,
3319 * then allocate the in-memory chain structure. Set the
3320 * INITIAL flag for fresh chains which do not have embedded
3323 * XXX for now set the check mode of the child based on
3324 * the parent or, if the parent is an inode, the
3325 * specification in the inode.
3327 bzero(&dummy, sizeof(dummy));
3330 dummy.keybits = keybits;
3331 dummy.data_off = hammer2_getradix(bytes);
3334 * Inherit methods from parent by default. Primarily used
3335 * for BREF_TYPE_DATA. Non-data types *must* be set to
3336 * a non-NONE check algorithm.
3339 dummy.methods = parent->bref.methods;
3341 dummy.methods = (uint8_t)methods;
3343 if (type != HAMMER2_BREF_TYPE_DATA &&
3344 HAMMER2_DEC_CHECK(dummy.methods) == HAMMER2_CHECK_NONE) {
3346 HAMMER2_ENC_CHECK(HAMMER2_CHECK_DEFAULT);
3349 chain = hammer2_chain_alloc(hmp, pmp, &dummy);
3352 * Lock the chain manually, chain_lock will load the chain
3353 * which we do NOT want to do. (note: chain->refs is set
3354 * to 1 by chain_alloc() for us, but lockcnt is not).
3357 hammer2_mtx_ex(&chain->lock);
3359 ++curthread->td_tracker;
3362 * Set INITIAL to optimize I/O. The flag will generally be
3363 * processed when we call hammer2_chain_modify().
3365 * Recalculate bytes to reflect the actual media block
3366 * allocation. Handle special case radix 0 == 0 bytes.
3368 bytes = (size_t)(chain->bref.data_off & HAMMER2_OFF_MASK_RADIX);
3370 bytes = (hammer2_off_t)1 << bytes;
3371 chain->bytes = bytes;
3374 case HAMMER2_BREF_TYPE_VOLUME:
3375 case HAMMER2_BREF_TYPE_FREEMAP:
3376 panic("hammer2_chain_create: called with volume type");
3378 case HAMMER2_BREF_TYPE_INDIRECT:
3379 panic("hammer2_chain_create: cannot be used to"
3380 "create indirect block");
3382 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3383 panic("hammer2_chain_create: cannot be used to"
3384 "create freemap root or node");
3386 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
3387 KKASSERT(bytes == sizeof(chain->data->bmdata));
3389 case HAMMER2_BREF_TYPE_DIRENT:
3390 case HAMMER2_BREF_TYPE_INODE:
3391 case HAMMER2_BREF_TYPE_DATA:
3394 * leave chain->data NULL, set INITIAL
3396 KKASSERT(chain->data == NULL);
3397 atomic_set_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
3402 * We are reattaching a previously deleted chain, possibly
3403 * under a new parent and possibly with a new key/keybits.
3404 * The chain does not have to be in a modified state. The
3405 * UPDATE flag will be set later on in this routine.
3407 * Do NOT mess with the current state of the INITIAL flag.
3409 chain->bref.key = key;
3410 chain->bref.keybits = keybits;
3411 if (chain->flags & HAMMER2_CHAIN_DELETED)
3412 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_DELETED);
3413 KKASSERT(chain->parent == NULL);
3417 * Set the appropriate bref flag if requested.
3419 * NOTE! Callers can call this function to move chains without
3420 * knowing about special flags, so don't clear bref flags
3423 if (flags & HAMMER2_INSERT_PFSROOT)
3424 chain->bref.flags |= HAMMER2_BREF_FLAG_PFSROOT;
3430 * Calculate how many entries we have in the blockref array and
3431 * determine if an indirect block is required when inserting into
3435 if (--maxloops == 0)
3436 panic("hammer2_chain_create: maxloops");
3438 switch(parent->bref.type) {
3439 case HAMMER2_BREF_TYPE_INODE:
3440 if ((parent->data->ipdata.meta.op_flags &
3441 HAMMER2_OPFLAG_DIRECTDATA) != 0) {
3442 kprintf("hammer2: parent set for direct-data! "
3443 "pkey=%016jx ckey=%016jx\n",
3447 KKASSERT((parent->data->ipdata.meta.op_flags &
3448 HAMMER2_OPFLAG_DIRECTDATA) == 0);
3449 KKASSERT(parent->data != NULL);
3450 base = &parent->data->ipdata.u.blockset.blockref[0];
3451 count = HAMMER2_SET_COUNT;
3453 case HAMMER2_BREF_TYPE_INDIRECT:
3454 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3455 if (parent->flags & HAMMER2_CHAIN_INITIAL)
3458 base = &parent->data->npdata[0];
3459 count = parent->bytes / sizeof(hammer2_blockref_t);
3461 case HAMMER2_BREF_TYPE_VOLUME:
3462 KKASSERT(parent->data != NULL);
3463 base = &parent->data->voldata.sroot_blockset.blockref[0];
3464 count = HAMMER2_SET_COUNT;
3466 case HAMMER2_BREF_TYPE_FREEMAP:
3467 KKASSERT(parent->data != NULL);
3468 base = &parent->data->blkset.blockref[0];
3469 count = HAMMER2_SET_COUNT;
3472 panic("hammer2_chain_create: unrecognized blockref type: %d",
3480 * Make sure we've counted the brefs
3482 if ((parent->flags & HAMMER2_CHAIN_COUNTEDBREFS) == 0)
3483 hammer2_chain_countbrefs(parent, base, count);
3485 KASSERT(parent->core.live_count >= 0 &&
3486 parent->core.live_count <= count,
3487 ("bad live_count %d/%d (%02x, %d)",
3488 parent->core.live_count, count,
3489 parent->bref.type, parent->bytes));
3492 * If no free blockref could be found we must create an indirect
3493 * block and move a number of blockrefs into it. With the parent
3494 * locked we can safely lock each child in order to delete+duplicate
3495 * it without causing a deadlock.
3497 * This may return the new indirect block or the old parent depending
3498 * on where the key falls. NULL is returned on error.
3500 if (parent->core.live_count == count) {
3501 hammer2_chain_t *nparent;
3503 KKASSERT((flags & HAMMER2_INSERT_SAMEPARENT) == 0);
3505 nparent = hammer2_chain_create_indirect(parent, key, keybits,
3506 mtid, type, &error);
3507 if (nparent == NULL) {
3509 hammer2_chain_drop(chain);
3513 if (parent != nparent) {
3514 hammer2_chain_unlock(parent);
3515 hammer2_chain_drop(parent);
3516 parent = *parentp = nparent;
3522 * fall through if parent, or skip to here if no parent.
3525 if (chain->flags & HAMMER2_CHAIN_DELETED)
3526 kprintf("Inserting deleted chain @%016jx\n",
3530 * Link the chain into its parent.
3532 if (chain->parent != NULL)
3533 panic("hammer2: hammer2_chain_create: chain already connected");
3534 KKASSERT(chain->parent == NULL);
3536 KKASSERT(parent->core.live_count < count);
3537 hammer2_chain_insert(parent, chain,
3538 HAMMER2_CHAIN_INSERT_SPIN |
3539 HAMMER2_CHAIN_INSERT_LIVE,
3545 * Mark the newly created chain modified. This will cause
3546 * UPDATE to be set and process the INITIAL flag.
3548 * Device buffers are not instantiated for DATA elements
3549 * as these are handled by logical buffers.
3551 * Indirect and freemap node indirect blocks are handled
3552 * by hammer2_chain_create_indirect() and not by this
3555 * Data for all other bref types is expected to be
3556 * instantiated (INODE, LEAF).
3558 switch(chain->bref.type) {
3559 case HAMMER2_BREF_TYPE_DATA:
3560 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
3561 case HAMMER2_BREF_TYPE_DIRENT:
3562 case HAMMER2_BREF_TYPE_INODE:
3563 error = hammer2_chain_modify(chain, mtid, dedup_off,
3564 HAMMER2_MODIFY_OPTDATA);
3568 * Remaining types are not supported by this function.
3569 * In particular, INDIRECT and LEAF_NODE types are
3570 * handled by create_indirect().
3572 panic("hammer2_chain_create: bad type: %d",
3579 * When reconnecting a chain we must set UPDATE and
3580 * setflush so the flush recognizes that it must update
3581 * the bref in the parent.
3583 if ((chain->flags & HAMMER2_CHAIN_UPDATE) == 0)
3584 atomic_set_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
3588 * We must setflush(parent) to ensure that it recurses through to
3589 * chain. setflush(chain) might not work because ONFLUSH is possibly
3590 * already set in the chain (so it won't recurse up to set it in the
3594 hammer2_chain_setflush(parent);
3603 * Move the chain from its old parent to a new parent. The chain must have
3604 * already been deleted or already disconnected (or never associated) with
3605 * a parent. The chain is reassociated with the new parent and the deleted
3606 * flag will be cleared (no longer deleted). The chain's modification state
3609 * THE CALLER MUST HAVE ALREADY PROPERLY SEEKED (parent) TO THE INSERTION
3610 * POINT SANS ANY REQUIRED INDIRECT BLOCK CREATIONS DUE TO THE ARRAY BEING
3611 * FULL. This typically means that the caller is creating the chain after
3612 * doing a hammer2_chain_lookup().
3614 * Neither (parent) or (chain) can be errored.
3616 * If (parent) is non-NULL then the chain is inserted under the parent.
3618 * If (parent) is NULL then the newly duplicated chain is not inserted
3619 * anywhere, similar to if it had just been chain_alloc()'d (suitable for
3620 * passing into hammer2_chain_create() after this function returns).
3622 * WARNING! This function calls create which means it can insert indirect
3623 * blocks. This can cause other unrelated chains in the parent to
3624 * be moved to a newly inserted indirect block in addition to the
3628 hammer2_chain_rename(hammer2_chain_t **parentp, hammer2_chain_t *chain,
3629 hammer2_tid_t mtid, int flags)
3631 hammer2_blockref_t *bref;
3633 hammer2_chain_t *parent;
3637 * WARNING! We should never resolve DATA to device buffers
3638 * (XXX allow it if the caller did?), and since
3639 * we currently do not have the logical buffer cache
3640 * buffer in-hand to fix its cached physical offset
3641 * we also force the modify code to not COW it. XXX
3643 * NOTE! We allow error'd chains to be renamed. The bref itself
3644 * is good and can be renamed. The content, however, may
3648 KKASSERT(chain->parent == NULL);
3649 /*KKASSERT(chain->error == 0); allow */
3652 * Now create a duplicate of the chain structure, associating
3653 * it with the same core, making it the same size, pointing it
3654 * to the same bref (the same media block).
3656 * NOTE: Handle special radix == 0 case (means 0 bytes).
3658 bref = &chain->bref;
3659 bytes = (size_t)(bref->data_off & HAMMER2_OFF_MASK_RADIX);
3661 bytes = (hammer2_off_t)1 << bytes;
3664 * If parent is not NULL the duplicated chain will be entered under
3665 * the parent and the UPDATE bit set to tell flush to update
3668 * We must setflush(parent) to ensure that it recurses through to
3669 * chain. setflush(chain) might not work because ONFLUSH is possibly
3670 * already set in the chain (so it won't recurse up to set it in the
3673 * Having both chains locked is extremely important for atomicy.
3675 if (parentp && (parent = *parentp) != NULL) {
3676 KKASSERT(hammer2_mtx_owned(&parent->lock));
3677 KKASSERT(parent->refs > 0);
3678 KKASSERT(parent->error == 0);
3680 hammer2_chain_create(parentp, &chain, NULL, chain->pmp,
3681 HAMMER2_METH_DEFAULT,
3682 bref->key, bref->keybits, bref->type,
3683 chain->bytes, mtid, 0, flags);
3684 KKASSERT(chain->flags & HAMMER2_CHAIN_UPDATE);
3685 hammer2_chain_setflush(*parentp);
3690 * This works in tandem with delete_obref() to install a blockref in
3691 * (typically) an indirect block that is associated with the chain being
3692 * moved to *parentp.
3694 * The reason we need this function is that the caller needs to maintain
3695 * the blockref as it was, and not generate a new blockref for what might
3696 * be a modified chain. Otherwise stuff will leak into the flush that
3697 * the flush code's FLUSH_INODE_STOP flag is unable to catch.
3699 * It is EXTREMELY important that we properly set CHAIN_BMAPUPD and
3700 * CHAIN_UPDATE. We must set BMAPUPD if the bref does not match, and
3701 * we must clear CHAIN_UPDATE (that was likely set by the chain_rename) if
3702 * it does. Otherwise we can end up in a situation where H2 is unable to
3703 * clean up the in-memory chain topology.
3705 * The reason for this is that flushes do not generally flush through
3706 * BREF_TYPE_INODE chains and depend on a hammer2_inode_t queued to syncq
3707 * or sideq to properly flush and dispose of the related inode chain's flags.
3708 * Situations where the inode is not actually modified by the frontend,
3709 * but where we have to move the related chains around as we insert or cleanup
3710 * indirect blocks, can leave us with a 'dirty' (non-disposable) in-memory
3711 * inode chain that does not have a hammer2_inode_t associated with it.
3714 hammer2_chain_rename_obref(hammer2_chain_t **parentp, hammer2_chain_t *chain,
3715 hammer2_tid_t mtid, int flags,
3716 hammer2_blockref_t *obref)
3718 hammer2_chain_rename(parentp, chain, mtid, flags);
3720 if (obref->type != HAMMER2_BREF_TYPE_EMPTY) {
3721 hammer2_blockref_t *tbase;
3724 KKASSERT((chain->flags & HAMMER2_CHAIN_BMAPPED) == 0);
3725 hammer2_chain_modify(*parentp, mtid, 0, 0);
3726 tbase = hammer2_chain_base_and_count(*parentp, &tcount);
3727 hammer2_base_insert(*parentp, tbase, tcount, chain, obref);
3728 if (bcmp(obref, &chain->bref, sizeof(chain->bref))) {
3729 atomic_set_int(&chain->flags, HAMMER2_CHAIN_BMAPUPD |
3730 HAMMER2_CHAIN_UPDATE);
3732 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
3738 * Helper function for deleting chains.
3740 * The chain is removed from the live view (the RBTREE) as well as the parent's
3741 * blockmap. Both chain and its parent must be locked.
3743 * parent may not be errored. chain can be errored.
3746 _hammer2_chain_delete_helper(hammer2_chain_t *parent, hammer2_chain_t *chain,
3747 hammer2_tid_t mtid, int flags,
3748 hammer2_blockref_t *obref)
3753 KKASSERT((chain->flags & (HAMMER2_CHAIN_DELETED |
3754 HAMMER2_CHAIN_FICTITIOUS)) == 0);
3755 KKASSERT(chain->parent == parent);
3758 if (chain->flags & HAMMER2_CHAIN_BMAPPED) {
3760 * Chain is blockmapped, so there must be a parent.
3761 * Atomically remove the chain from the parent and remove
3762 * the blockmap entry. The parent must be set modified
3763 * to remove the blockmap entry.
3765 hammer2_blockref_t *base;
3768 KKASSERT(parent != NULL);
3769 KKASSERT(parent->error == 0);
3770 KKASSERT((parent->flags & HAMMER2_CHAIN_INITIAL) == 0);
3771 error = hammer2_chain_modify(parent, mtid, 0, 0);
3776 * Calculate blockmap pointer
3778 KKASSERT(chain->flags & HAMMER2_CHAIN_ONRBTREE);
3779 hammer2_spin_ex(&chain->core.spin);
3780 hammer2_spin_ex(&parent->core.spin);
3782 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DELETED);
3783 atomic_add_int(&parent->core.live_count, -1);
3784 ++parent->core.generation;
3785 RB_REMOVE(hammer2_chain_tree, &parent->core.rbtree, chain);
3786 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE);
3787 --parent->core.chain_count;
3788 chain->parent = NULL;
3790 switch(parent->bref.type) {
3791 case HAMMER2_BREF_TYPE_INODE:
3793 * Access the inode's block array. However, there
3794 * is no block array if the inode is flagged
3798 (parent->data->ipdata.meta.op_flags &
3799 HAMMER2_OPFLAG_DIRECTDATA) == 0) {
3801 &parent->data->ipdata.u.blockset.blockref[0];
3805 count = HAMMER2_SET_COUNT;
3807 case HAMMER2_BREF_TYPE_INDIRECT:
3808 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3810 base = &parent->data->npdata[0];
3813 count = parent->bytes / sizeof(hammer2_blockref_t);
3815 case HAMMER2_BREF_TYPE_VOLUME:
3816 base = &parent->data->voldata.
3817 sroot_blockset.blockref[0];
3818 count = HAMMER2_SET_COUNT;
3820 case HAMMER2_BREF_TYPE_FREEMAP:
3821 base = &parent->data->blkset.blockref[0];
3822 count = HAMMER2_SET_COUNT;
3827 panic("_hammer2_chain_delete_helper: "
3828 "unrecognized blockref type: %d",
3833 * delete blockmapped chain from its parent.
3835 * The parent is not affected by any statistics in chain
3836 * which are pending synchronization. That is, there is
3837 * nothing to undo in the parent since they have not yet
3838 * been incorporated into the parent.
3840 * The parent is affected by statistics stored in inodes.
3841 * Those have already been synchronized, so they must be
3842 * undone. XXX split update possible w/delete in middle?
3845 hammer2_base_delete(parent, base, count, chain, obref);
3847 hammer2_spin_unex(&parent->core.spin);
3848 hammer2_spin_unex(&chain->core.spin);
3849 } else if (chain->flags & HAMMER2_CHAIN_ONRBTREE) {
3851 * Chain is not blockmapped but a parent is present.
3852 * Atomically remove the chain from the parent. There is
3853 * no blockmap entry to remove.
3855 * Because chain was associated with a parent but not
3856 * synchronized, the chain's *_count_up fields contain
3857 * inode adjustment statistics which must be undone.
3859 hammer2_spin_ex(&chain->core.spin);
3860 hammer2_spin_ex(&parent->core.spin);
3861 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DELETED);
3862 atomic_add_int(&parent->core.live_count, -1);
3863 ++parent->core.generation;
3864 RB_REMOVE(hammer2_chain_tree, &parent->core.rbtree, chain);
3865 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE);
3866 --parent->core.chain_count;
3867 chain->parent = NULL;
3868 hammer2_spin_unex(&parent->core.spin);
3869 hammer2_spin_unex(&chain->core.spin);
3872 * Chain is not blockmapped and has no parent. This
3873 * is a degenerate case.
3875 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DELETED);
3882 * Create an indirect block that covers one or more of the elements in the
3883 * current parent. Either returns the existing parent with no locking or
3884 * ref changes or returns the new indirect block locked and referenced
3885 * and leaving the original parent lock/ref intact as well.
3887 * If an error occurs, NULL is returned and *errorp is set to the H2 error.
3889 * The returned chain depends on where the specified key falls.
3891 * The key/keybits for the indirect mode only needs to follow three rules:
3893 * (1) That all elements underneath it fit within its key space and
3895 * (2) That all elements outside it are outside its key space.
3897 * (3) When creating the new indirect block any elements in the current
3898 * parent that fit within the new indirect block's keyspace must be
3899 * moved into the new indirect block.
3901 * (4) The keyspace chosen for the inserted indirect block CAN cover a wider
3902 * keyspace the the current parent, but lookup/iteration rules will
3903 * ensure (and must ensure) that rule (2) for all parents leading up
3904 * to the nearest inode or the root volume header is adhered to. This
3905 * is accomplished by always recursing through matching keyspaces in
3906 * the hammer2_chain_lookup() and hammer2_chain_next() API.
3908 * The current implementation calculates the current worst-case keyspace by
3909 * iterating the current parent and then divides it into two halves, choosing
3910 * whichever half has the most elements (not necessarily the half containing
3911 * the requested key).
3913 * We can also opt to use the half with the least number of elements. This
3914 * causes lower-numbered keys (aka logical file offsets) to recurse through
3915 * fewer indirect blocks and higher-numbered keys to recurse through more.
3916 * This also has the risk of not moving enough elements to the new indirect
3917 * block and being forced to create several indirect blocks before the element
3920 * Must be called with an exclusively locked parent.
3922 * NOTE: *errorp set to HAMMER_ERROR_* flags
3924 static int hammer2_chain_indkey_freemap(hammer2_chain_t *parent,
3925 hammer2_key_t *keyp, int keybits,
3926 hammer2_blockref_t *base, int count);
3927 static int hammer2_chain_indkey_file(hammer2_chain_t *parent,
3928 hammer2_key_t *keyp, int keybits,
3929 hammer2_blockref_t *base, int count,
3931 static int hammer2_chain_indkey_dir(hammer2_chain_t *parent,
3932 hammer2_key_t *keyp, int keybits,
3933 hammer2_blockref_t *base, int count,
3937 hammer2_chain_create_indirect(hammer2_chain_t *parent,
3938 hammer2_key_t create_key, int create_bits,
3939 hammer2_tid_t mtid, int for_type, int *errorp)
3942 hammer2_blockref_t *base;
3943 hammer2_blockref_t *bref;
3944 hammer2_blockref_t bcopy;
3945 hammer2_blockref_t dummy;
3946 hammer2_chain_t *chain;
3947 hammer2_chain_t *ichain;
3948 hammer2_key_t key = create_key;
3949 hammer2_key_t key_beg;
3950 hammer2_key_t key_end;
3951 hammer2_key_t key_next;
3952 int keybits = create_bits;
3960 int maxloops = 300000;
3963 * Calculate the base blockref pointer or NULL if the chain
3964 * is known to be empty. We need to calculate the array count
3965 * for RB lookups either way.
3968 KKASSERT(hammer2_mtx_owned(&parent->lock));
3971 * Pre-modify the parent now to avoid having to deal with error
3972 * processing if we tried to later (in the middle of our loop).
3974 * We are going to be moving bref's around, the indirect blocks
3975 * cannot be in an initial state. Do not pass MODIFY_OPTDATA.
3977 *errorp = hammer2_chain_modify(parent, mtid, 0, 0);
3979 kprintf("hammer2_create_indirect: error %08x %s\n",
3980 *errorp, hammer2_error_str(*errorp));
3983 KKASSERT((parent->flags & HAMMER2_CHAIN_INITIAL) == 0);
3985 /*hammer2_chain_modify(&parent, HAMMER2_MODIFY_OPTDATA);*/
3986 base = hammer2_chain_base_and_count(parent, &count);
3989 * How big should our new indirect block be? It has to be at least
3990 * as large as its parent for splits to work properly.
3992 * The freemap uses a specific indirect block size. The number of
3993 * levels are built dynamically and ultimately depend on the size
3994 * volume. Because freemap blocks are taken from the reserved areas
3995 * of the volume our goal is efficiency (fewer levels) and not so
3996 * much to save disk space.
3998 * The first indirect block level for a directory usually uses
3999 * HAMMER2_IND_BYTES_MIN (4KB = 32 directory entries). Due to
4000 * the hash mechanism, this typically gives us a nominal
4001 * 32 * 4 entries with one level of indirection.
4003 * We use HAMMER2_IND_BYTES_NOM (16KB = 128 blockrefs) for FILE
4004 * indirect blocks. The initial 4 entries in the inode gives us
4005 * 256KB. Up to 4 indirect blocks gives us 32MB. Three levels
4006 * of indirection gives us 137GB, and so forth. H2 can support
4007 * huge file sizes but they are not typical, so we try to stick
4008 * with compactness and do not use a larger indirect block size.
4010 * We could use 64KB (PBUFSIZE), giving us 512 blockrefs, but
4011 * due to the way indirect blocks are created this usually winds
4012 * up being extremely inefficient for small files. Even though
4013 * 16KB requires more levels of indirection for very large files,
4014 * the 16KB records can be ganged together into 64KB DIOs.
4016 if (for_type == HAMMER2_BREF_TYPE_FREEMAP_NODE ||
4017 for_type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) {
4018 nbytes = HAMMER2_FREEMAP_LEVELN_PSIZE;
4019 } else if (parent->bref.type == HAMMER2_BREF_TYPE_INODE) {
4020 if (parent->data->ipdata.meta.type ==
4021 HAMMER2_OBJTYPE_DIRECTORY)
4022 nbytes = HAMMER2_IND_BYTES_MIN; /* 4KB = 32 entries */
4024 nbytes = HAMMER2_IND_BYTES_NOM; /* 16KB = ~8MB file */
4027 nbytes = HAMMER2_IND_BYTES_NOM;
4029 if (nbytes < count * sizeof(hammer2_blockref_t)) {
4030 KKASSERT(for_type != HAMMER2_BREF_TYPE_FREEMAP_NODE &&
4031 for_type != HAMMER2_BREF_TYPE_FREEMAP_LEAF);
4032 nbytes = count * sizeof(hammer2_blockref_t);
4034 ncount = nbytes / sizeof(hammer2_blockref_t);
4037 * When creating an indirect block for a freemap node or leaf
4038 * the key/keybits must be fitted to static radix levels because
4039 * particular radix levels use particular reserved blocks in the
4042 * This routine calculates the key/radix of the indirect block
4043 * we need to create, and whether it is on the high-side or the
4047 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
4048 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
4049 keybits = hammer2_chain_indkey_freemap(parent, &key, keybits,
4052 case HAMMER2_BREF_TYPE_DATA:
4053 keybits = hammer2_chain_indkey_file(parent, &key, keybits,
4054 base, count, ncount);
4056 case HAMMER2_BREF_TYPE_DIRENT:
4057 case HAMMER2_BREF_TYPE_INODE:
4058 keybits = hammer2_chain_indkey_dir(parent, &key, keybits,
4059 base, count, ncount);
4062 panic("illegal indirect block for bref type %d", for_type);
4067 * Normalize the key for the radix being represented, keeping the
4068 * high bits and throwing away the low bits.
4070 key &= ~(((hammer2_key_t)1 << keybits) - 1);
4073 * Ok, create our new indirect block
4075 bzero(&dummy, sizeof(dummy));
4076 if (for_type == HAMMER2_BREF_TYPE_FREEMAP_NODE ||
4077 for_type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) {
4078 dummy.type = HAMMER2_BREF_TYPE_FREEMAP_NODE;
4080 dummy.type = HAMMER2_BREF_TYPE_INDIRECT;
4083 dummy.keybits = keybits;
4084 dummy.data_off = hammer2_getradix(nbytes);
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);
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.
5140 obref->type = HAMMER2_BREF_TYPE_EMPTY;
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 == HAMMER2_BREF_TYPE_EMPTY ||
5221 scan->key > key_beg)) {
5225 parent->cache_index = i;
5228 * Search forwards, stop when we find a scan element which
5229 * encloses the key or until we know that there are no further
5233 if (scan->type != HAMMER2_BREF_TYPE_EMPTY) {
5234 scan_end = scan->key +
5235 ((hammer2_key_t)1 << scan->keybits) - 1;
5236 if (scan->key > key_beg || scan_end >= key_beg)
5245 parent->cache_index = i;
5249 scan_end = scan->key +
5250 ((hammer2_key_t)1 << scan->keybits);
5251 if (scan_end && (*key_nextp > scan_end ||
5253 *key_nextp = scan_end;
5261 * Do a combined search and return the next match either from the blockref
5262 * array or from the in-memory chain. Sets *bresp to the returned bref in
5263 * both cases, or sets it to NULL if the search exhausted. Only returns
5264 * a non-NULL chain if the search matched from the in-memory chain.
5266 * When no in-memory chain has been found and a non-NULL bref is returned
5270 * The returned chain is not locked or referenced. Use the returned bref
5271 * to determine if the search exhausted or not. Iterate if the base find
5272 * is chosen but matches a deleted chain.
5274 * WARNING! Must be called with parent's spinlock held. Spinlock remains
5275 * held through the operation.
5278 hammer2_combined_find(hammer2_chain_t *parent,
5279 hammer2_blockref_t *base, int count,
5280 hammer2_key_t *key_nextp,
5281 hammer2_key_t key_beg, hammer2_key_t key_end,
5282 hammer2_blockref_t **bresp)
5284 hammer2_blockref_t *bref;
5285 hammer2_chain_t *chain;
5289 * Lookup in block array and in rbtree.
5291 *key_nextp = key_end + 1;
5292 i = hammer2_base_find(parent, base, count, key_nextp,
5294 chain = hammer2_chain_find(parent, key_nextp, key_beg, key_end);
5299 if (i == count && chain == NULL) {
5305 * Only chain matched.
5308 bref = &chain->bref;
5313 * Only blockref matched.
5315 if (chain == NULL) {
5321 * Both in-memory and blockref matched, select the nearer element.
5323 * If both are flush with the left-hand side or both are the
5324 * same distance away, select the chain. In this situation the
5325 * chain must have been loaded from the matching blockmap.
5327 if ((chain->bref.key <= key_beg && base[i].key <= key_beg) ||
5328 chain->bref.key == base[i].key) {
5329 KKASSERT(chain->bref.key == base[i].key);
5330 bref = &chain->bref;
5335 * Select the nearer key
5337 if (chain->bref.key < base[i].key) {
5338 bref = &chain->bref;
5345 * If the bref is out of bounds we've exhausted our search.
5348 if (bref->key > key_end) {
5358 * Locate the specified block array element and delete it. The element
5361 * The spin lock on the related chain must be held.
5363 * NOTE: live_count was adjusted when the chain was deleted, so it does not
5364 * need to be adjusted when we commit the media change.
5367 hammer2_base_delete(hammer2_chain_t *parent,
5368 hammer2_blockref_t *base, int count,
5369 hammer2_chain_t *chain,
5370 hammer2_blockref_t *obref)
5372 hammer2_blockref_t *elm = &chain->bref;
5373 hammer2_blockref_t *scan;
5374 hammer2_key_t key_next;
5378 * Delete element. Expect the element to exist.
5380 * XXX see caller, flush code not yet sophisticated enough to prevent
5381 * re-flushed in some cases.
5383 key_next = 0; /* max range */
5384 i = hammer2_base_find(parent, base, count, &key_next,
5385 elm->key, elm->key);
5387 if (i == count || scan->type == HAMMER2_BREF_TYPE_EMPTY ||
5388 scan->key != elm->key ||
5389 ((chain->flags & HAMMER2_CHAIN_BMAPUPD) == 0 &&
5390 scan->keybits != elm->keybits)) {
5391 hammer2_spin_unex(&parent->core.spin);
5392 panic("delete base %p element not found at %d/%d elm %p\n",
5393 base, i, count, elm);
5398 * Update stats and zero the entry.
5400 * NOTE: Handle radix == 0 (0 bytes) case.
5402 if ((int)(scan->data_off & HAMMER2_OFF_MASK_RADIX)) {
5403 parent->bref.embed.stats.data_count -= (hammer2_off_t)1 <<
5404 (int)(scan->data_off & HAMMER2_OFF_MASK_RADIX);
5406 switch(scan->type) {
5407 case HAMMER2_BREF_TYPE_INODE:
5408 --parent->bref.embed.stats.inode_count;
5410 case HAMMER2_BREF_TYPE_DATA:
5411 if (parent->bref.leaf_count == HAMMER2_BLOCKREF_LEAF_MAX) {
5412 atomic_set_int(&chain->flags,
5413 HAMMER2_CHAIN_HINT_LEAF_COUNT);
5415 if (parent->bref.leaf_count)
5416 --parent->bref.leaf_count;
5419 case HAMMER2_BREF_TYPE_INDIRECT:
5420 if (scan->type != HAMMER2_BREF_TYPE_DATA) {
5421 parent->bref.embed.stats.data_count -=
5422 scan->embed.stats.data_count;
5423 parent->bref.embed.stats.inode_count -=
5424 scan->embed.stats.inode_count;
5426 if (scan->type == HAMMER2_BREF_TYPE_INODE)
5428 if (parent->bref.leaf_count == HAMMER2_BLOCKREF_LEAF_MAX) {
5429 atomic_set_int(&chain->flags,
5430 HAMMER2_CHAIN_HINT_LEAF_COUNT);
5432 if (parent->bref.leaf_count <= scan->leaf_count)
5433 parent->bref.leaf_count = 0;
5435 parent->bref.leaf_count -= scan->leaf_count;
5438 case HAMMER2_BREF_TYPE_DIRENT:
5439 if (parent->bref.leaf_count == HAMMER2_BLOCKREF_LEAF_MAX) {
5440 atomic_set_int(&chain->flags,
5441 HAMMER2_CHAIN_HINT_LEAF_COUNT);
5443 if (parent->bref.leaf_count)
5444 --parent->bref.leaf_count;
5452 bzero(scan, sizeof(*scan));
5455 * We can only optimize parent->core.live_zero for live chains.
5457 if (parent->core.live_zero == i + 1) {
5458 while (--i >= 0 && base[i].type == HAMMER2_BREF_TYPE_EMPTY)
5460 parent->core.live_zero = i + 1;
5464 * Clear appropriate blockmap flags in chain.
5466 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_BMAPPED |
5467 HAMMER2_CHAIN_BMAPUPD);
5471 * Insert the specified element. The block array must not already have the
5472 * element and must have space available for the insertion.
5474 * The spin lock on the related chain must be held.
5476 * NOTE: live_count was adjusted when the chain was deleted, so it does not
5477 * need to be adjusted when we commit the media change.
5480 hammer2_base_insert(hammer2_chain_t *parent,
5481 hammer2_blockref_t *base, int count,
5482 hammer2_chain_t *chain, hammer2_blockref_t *elm)
5484 hammer2_key_t key_next;
5493 * Insert new element. Expect the element to not already exist
5494 * unless we are replacing it.
5496 * XXX see caller, flush code not yet sophisticated enough to prevent
5497 * re-flushed in some cases.
5499 key_next = 0; /* max range */
5500 i = hammer2_base_find(parent, base, count, &key_next,
5501 elm->key, elm->key);
5504 * Shortcut fill optimization, typical ordered insertion(s) may not
5507 KKASSERT(i >= 0 && i <= count);
5510 * Set appropriate blockmap flags in chain (if not NULL)
5513 atomic_set_int(&chain->flags, HAMMER2_CHAIN_BMAPPED);
5516 * Update stats and zero the entry
5518 if ((int)(elm->data_off & HAMMER2_OFF_MASK_RADIX)) {
5519 parent->bref.embed.stats.data_count += (hammer2_off_t)1 <<
5520 (int)(elm->data_off & HAMMER2_OFF_MASK_RADIX);
5523 case HAMMER2_BREF_TYPE_INODE:
5524 ++parent->bref.embed.stats.inode_count;
5526 case HAMMER2_BREF_TYPE_DATA:
5527 if (parent->bref.leaf_count != HAMMER2_BLOCKREF_LEAF_MAX)
5528 ++parent->bref.leaf_count;
5530 case HAMMER2_BREF_TYPE_INDIRECT:
5531 if (elm->type != HAMMER2_BREF_TYPE_DATA) {
5532 parent->bref.embed.stats.data_count +=
5533 elm->embed.stats.data_count;
5534 parent->bref.embed.stats.inode_count +=
5535 elm->embed.stats.inode_count;
5537 if (elm->type == HAMMER2_BREF_TYPE_INODE)
5539 if (parent->bref.leaf_count + elm->leaf_count <
5540 HAMMER2_BLOCKREF_LEAF_MAX) {
5541 parent->bref.leaf_count += elm->leaf_count;
5543 parent->bref.leaf_count = HAMMER2_BLOCKREF_LEAF_MAX;
5546 case HAMMER2_BREF_TYPE_DIRENT:
5547 if (parent->bref.leaf_count != HAMMER2_BLOCKREF_LEAF_MAX)
5548 ++parent->bref.leaf_count;
5556 * We can only optimize parent->core.live_zero for live chains.
5558 if (i == count && parent->core.live_zero < count) {
5559 i = parent->core.live_zero++;
5564 xkey = elm->key + ((hammer2_key_t)1 << elm->keybits) - 1;
5565 if (i != count && (base[i].key < elm->key || xkey >= base[i].key)) {
5566 hammer2_spin_unex(&parent->core.spin);
5567 panic("insert base %p overlapping elements at %d elm %p\n",
5572 * Try to find an empty slot before or after.
5576 while (j > 0 || k < count) {
5578 if (j >= 0 && base[j].type == HAMMER2_BREF_TYPE_EMPTY) {
5582 bcopy(&base[j+1], &base[j],
5583 (i - j - 1) * sizeof(*base));
5589 if (k < count && base[k].type == HAMMER2_BREF_TYPE_EMPTY) {
5590 bcopy(&base[i], &base[i+1],
5591 (k - i) * sizeof(hammer2_blockref_t));
5595 * We can only update parent->core.live_zero for live
5598 if (parent->core.live_zero <= k)
5599 parent->core.live_zero = k + 1;
5604 panic("hammer2_base_insert: no room!");
5611 for (l = 0; l < count; ++l) {
5612 if (base[l].type != HAMMER2_BREF_TYPE_EMPTY) {
5613 key_next = base[l].key +
5614 ((hammer2_key_t)1 << base[l].keybits) - 1;
5618 while (++l < count) {
5619 if (base[l].type != HAMMER2_BREF_TYPE_EMPTY) {
5620 if (base[l].key <= key_next)
5621 panic("base_insert %d %d,%d,%d fail %p:%d", u, i, j, k, base, l);
5622 key_next = base[l].key +
5623 ((hammer2_key_t)1 << base[l].keybits) - 1;
5633 * Sort the blockref array for the chain. Used by the flush code to
5634 * sort the blockref[] array.
5636 * The chain must be exclusively locked AND spin-locked.
5638 typedef hammer2_blockref_t *hammer2_blockref_p;
5642 hammer2_base_sort_callback(const void *v1, const void *v2)
5644 hammer2_blockref_p bref1 = *(const hammer2_blockref_p *)v1;
5645 hammer2_blockref_p bref2 = *(const hammer2_blockref_p *)v2;
5648 * Make sure empty elements are placed at the end of the array
5650 if (bref1->type == HAMMER2_BREF_TYPE_EMPTY) {
5651 if (bref2->type == HAMMER2_BREF_TYPE_EMPTY)
5654 } else if (bref2->type == HAMMER2_BREF_TYPE_EMPTY) {
5661 if (bref1->key < bref2->key)
5663 if (bref1->key > bref2->key)
5669 hammer2_base_sort(hammer2_chain_t *chain)
5671 hammer2_blockref_t *base;
5674 switch(chain->bref.type) {
5675 case HAMMER2_BREF_TYPE_INODE:
5677 * Special shortcut for embedded data returns the inode
5678 * itself. Callers must detect this condition and access
5679 * the embedded data (the strategy code does this for us).
5681 * This is only applicable to regular files and softlinks.
5683 if (chain->data->ipdata.meta.op_flags &
5684 HAMMER2_OPFLAG_DIRECTDATA) {
5687 base = &chain->data->ipdata.u.blockset.blockref[0];
5688 count = HAMMER2_SET_COUNT;
5690 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
5691 case HAMMER2_BREF_TYPE_INDIRECT:
5693 * Optimize indirect blocks in the INITIAL state to avoid
5696 KKASSERT((chain->flags & HAMMER2_CHAIN_INITIAL) == 0);
5697 base = &chain->data->npdata[0];
5698 count = chain->bytes / sizeof(hammer2_blockref_t);
5700 case HAMMER2_BREF_TYPE_VOLUME:
5701 base = &chain->data->voldata.sroot_blockset.blockref[0];
5702 count = HAMMER2_SET_COUNT;
5704 case HAMMER2_BREF_TYPE_FREEMAP:
5705 base = &chain->data->blkset.blockref[0];
5706 count = HAMMER2_SET_COUNT;
5709 kprintf("hammer2_chain_lookup: unrecognized "
5710 "blockref(A) type: %d",
5713 tsleep(&base, 0, "dead", 0);
5714 panic("hammer2_base_sort: unrecognized "
5715 "blockref(A) type: %d",
5717 base = NULL; /* safety */
5718 count = 0; /* safety */
5720 kqsort(base, count, sizeof(*base), hammer2_base_sort_callback);
5726 * Chain memory management
5729 hammer2_chain_wait(hammer2_chain_t *chain)
5731 tsleep(chain, 0, "chnflw", 1);
5734 const hammer2_media_data_t *
5735 hammer2_chain_rdata(hammer2_chain_t *chain)
5737 KKASSERT(chain->data != NULL);
5738 return (chain->data);
5741 hammer2_media_data_t *
5742 hammer2_chain_wdata(hammer2_chain_t *chain)
5744 KKASSERT(chain->data != NULL);
5745 return (chain->data);
5749 * Set the check data for a chain. This can be a heavy-weight operation
5750 * and typically only runs on-flush. For file data check data is calculated
5751 * when the logical buffers are flushed.
5754 hammer2_chain_setcheck(hammer2_chain_t *chain, void *bdata)
5756 chain->bref.flags &= ~HAMMER2_BREF_FLAG_ZERO;
5758 switch(HAMMER2_DEC_CHECK(chain->bref.methods)) {
5759 case HAMMER2_CHECK_NONE:
5761 case HAMMER2_CHECK_DISABLED:
5763 case HAMMER2_CHECK_ISCSI32:
5764 chain->bref.check.iscsi32.value =
5765 hammer2_icrc32(bdata, chain->bytes);
5767 case HAMMER2_CHECK_XXHASH64:
5768 chain->bref.check.xxhash64.value =
5769 XXH64(bdata, chain->bytes, XXH_HAMMER2_SEED);
5771 case HAMMER2_CHECK_SHA192:
5773 SHA256_CTX hash_ctx;
5775 uint8_t digest[SHA256_DIGEST_LENGTH];
5776 uint64_t digest64[SHA256_DIGEST_LENGTH/8];
5779 SHA256_Init(&hash_ctx);
5780 SHA256_Update(&hash_ctx, bdata, chain->bytes);
5781 SHA256_Final(u.digest, &hash_ctx);
5782 u.digest64[2] ^= u.digest64[3];
5784 chain->bref.check.sha192.data,
5785 sizeof(chain->bref.check.sha192.data));
5788 case HAMMER2_CHECK_FREEMAP:
5789 chain->bref.check.freemap.icrc32 =
5790 hammer2_icrc32(bdata, chain->bytes);
5793 kprintf("hammer2_chain_setcheck: unknown check type %02x\n",
5794 chain->bref.methods);
5800 * Characterize a failed check code and try to trace back to the inode.
5803 hammer2_characterize_failed_chain(hammer2_chain_t *chain, uint64_t check,
5806 hammer2_chain_t *lchain;
5807 hammer2_chain_t *ochain;
5809 kprintf("chain %016jx.%02x (%s) meth=%02x CHECK FAIL "
5810 "(flags=%08x, bref/data ",
5811 chain->bref.data_off,
5813 hammer2_bref_type_str(&chain->bref),
5814 chain->bref.methods,
5817 kprintf("%08x/%08x)\n",
5818 chain->bref.check.iscsi32.value,
5821 kprintf("%016jx/%016jx)\n",
5822 chain->bref.check.xxhash64.value,
5827 * Run up the chains to try to find the governing inode so we
5830 * XXX This error reporting is not really MPSAFE
5834 while (chain && chain->bref.type != HAMMER2_BREF_TYPE_INODE) {
5836 chain = chain->parent;
5839 if (chain && chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
5840 ((chain->bref.flags & HAMMER2_BREF_FLAG_PFSROOT) == 0 ||
5841 (lchain->bref.key & HAMMER2_DIRHASH_VISIBLE))) {
5842 kprintf(" Resides at/in inode %ld\n",
5844 } else if (chain && chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
5845 kprintf(" Resides in inode index - CRITICAL!!!\n");
5847 kprintf(" Resides in root index - CRITICAL!!!\n");
5850 const char *pfsname = "UNKNOWN";
5854 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
5855 if (ochain->pmp->pfs_hmps[i] == ochain->hmp &&
5856 ochain->pmp->pfs_names[i]) {
5857 pfsname = ochain->pmp->pfs_names[i];
5862 kprintf(" In pfs %s on device %s\n",
5863 pfsname, ochain->hmp->devrepname);
5868 hammer2_chain_testcheck(hammer2_chain_t *chain, void *bdata)
5874 if (chain->bref.flags & HAMMER2_BREF_FLAG_ZERO)
5877 switch(HAMMER2_DEC_CHECK(chain->bref.methods)) {
5878 case HAMMER2_CHECK_NONE:
5881 case HAMMER2_CHECK_DISABLED:
5884 case HAMMER2_CHECK_ISCSI32:
5885 check32 = hammer2_icrc32(bdata, chain->bytes);
5886 r = (chain->bref.check.iscsi32.value == check32);
5888 hammer2_characterize_failed_chain(chain, check32, 32);
5890 hammer2_process_icrc32 += chain->bytes;
5892 case HAMMER2_CHECK_XXHASH64:
5893 check64 = XXH64(bdata, chain->bytes, XXH_HAMMER2_SEED);
5894 r = (chain->bref.check.xxhash64.value == check64);
5896 hammer2_characterize_failed_chain(chain, check64, 64);
5898 hammer2_process_xxhash64 += chain->bytes;
5900 case HAMMER2_CHECK_SHA192:
5902 SHA256_CTX hash_ctx;
5904 uint8_t digest[SHA256_DIGEST_LENGTH];
5905 uint64_t digest64[SHA256_DIGEST_LENGTH/8];
5908 SHA256_Init(&hash_ctx);
5909 SHA256_Update(&hash_ctx, bdata, chain->bytes);
5910 SHA256_Final(u.digest, &hash_ctx);
5911 u.digest64[2] ^= u.digest64[3];
5913 chain->bref.check.sha192.data,
5914 sizeof(chain->bref.check.sha192.data)) == 0) {
5918 kprintf("chain %016jx.%02x meth=%02x "
5920 chain->bref.data_off,
5922 chain->bref.methods);
5926 case HAMMER2_CHECK_FREEMAP:
5927 r = (chain->bref.check.freemap.icrc32 ==
5928 hammer2_icrc32(bdata, chain->bytes));
5930 kprintf("chain %016jx.%02x meth=%02x "
5932 chain->bref.data_off,
5934 chain->bref.methods);
5935 kprintf("freemap.icrc %08x icrc32 %08x (%d)\n",
5936 chain->bref.check.freemap.icrc32,
5937 hammer2_icrc32(bdata, chain->bytes),
5940 kprintf("dio %p buf %016jx,%d bdata %p/%p\n",
5941 chain->dio, chain->dio->bp->b_loffset,
5942 chain->dio->bp->b_bufsize, bdata,
5943 chain->dio->bp->b_data);
5948 kprintf("hammer2_chain_setcheck: unknown check type %02x\n",
5949 chain->bref.methods);
5957 * Acquire the chain and parent representing the specified inode for the
5958 * device at the specified cluster index.
5960 * The flags passed in are LOOKUP flags, not RESOLVE flags.
5962 * If we are unable to locate the inode, HAMMER2_ERROR_EIO is returned and
5963 * *chainp will be NULL. *parentp may still be set error or not, or NULL
5964 * if the parent itself could not be resolved.
5966 * The caller may pass-in a locked *parentp and/or *chainp, or neither.
5967 * They will be unlocked and released by this function. The *parentp and
5968 * *chainp representing the located inode are returned locked.
5971 hammer2_chain_inode_find(hammer2_pfs_t *pmp, hammer2_key_t inum,
5972 int clindex, int flags,
5973 hammer2_chain_t **parentp, hammer2_chain_t **chainp)
5975 hammer2_chain_t *parent;
5976 hammer2_chain_t *rchain;
5977 hammer2_key_t key_dummy;
5978 hammer2_inode_t *ip;
5982 resolve_flags = (flags & HAMMER2_LOOKUP_SHARED) ?
5983 HAMMER2_RESOLVE_SHARED : 0;
5986 * Caller expects us to replace these.
5989 hammer2_chain_unlock(*chainp);
5990 hammer2_chain_drop(*chainp);
5994 hammer2_chain_unlock(*parentp);
5995 hammer2_chain_drop(*parentp);
6000 * Be very careful, this is a backend function and we CANNOT
6001 * lock any frontend inode structure we find. But we have to
6002 * look the inode up this way first in case it exists but is
6003 * detached from the radix tree.
6005 ip = hammer2_inode_lookup(pmp, inum);
6007 *chainp = hammer2_inode_chain_and_parent(ip, clindex,
6010 hammer2_inode_drop(ip);
6013 hammer2_chain_unlock(*chainp);
6014 hammer2_chain_drop(*chainp);
6017 hammer2_chain_unlock(*parentp);
6018 hammer2_chain_drop(*parentp);
6024 * Inodes hang off of the iroot (bit 63 is clear, differentiating
6025 * inodes from root directory entries in the key lookup).
6027 parent = hammer2_inode_chain(pmp->iroot, clindex, resolve_flags);
6030 rchain = hammer2_chain_lookup(&parent, &key_dummy,
6034 error = HAMMER2_ERROR_EIO;
6043 * Used by the bulkscan code to snapshot the synchronized storage for
6044 * a volume, allowing it to be scanned concurrently against normal
6048 hammer2_chain_bulksnap(hammer2_dev_t *hmp)
6050 hammer2_chain_t *copy;
6052 copy = hammer2_chain_alloc(hmp, hmp->spmp, &hmp->vchain.bref);
6053 copy->data = kmalloc(sizeof(copy->data->voldata),
6056 hammer2_voldata_lock(hmp);
6057 copy->data->voldata = hmp->volsync;
6058 hammer2_voldata_unlock(hmp);
6064 hammer2_chain_bulkdrop(hammer2_chain_t *copy)
6066 KKASSERT(copy->bref.type == HAMMER2_BREF_TYPE_VOLUME);
6067 KKASSERT(copy->data);
6068 kfree(copy->data, copy->hmp->mchain);
6070 atomic_add_long(&hammer2_chain_allocs, -1);
6071 hammer2_chain_drop(copy);
6075 * Returns non-zero if the chain (INODE or DIRENT) matches the
6079 hammer2_chain_dirent_test(hammer2_chain_t *chain, const char *name,
6082 const hammer2_inode_data_t *ripdata;
6083 const hammer2_dirent_head_t *den;
6085 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
6086 ripdata = &chain->data->ipdata;
6087 if (ripdata->meta.name_len == name_len &&
6088 bcmp(ripdata->filename, name, name_len) == 0) {
6092 if (chain->bref.type == HAMMER2_BREF_TYPE_DIRENT &&
6093 chain->bref.embed.dirent.namlen == name_len) {
6094 den = &chain->bref.embed.dirent;
6095 if (name_len > sizeof(chain->bref.check.buf) &&
6096 bcmp(chain->data->buf, name, name_len) == 0) {
6099 if (name_len <= sizeof(chain->bref.check.buf) &&
6100 bcmp(chain->bref.check.buf, name, name_len) == 0) {