2 * Copyright (c) 2011-2020 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 int hammer2_chain_delete_obref(hammer2_chain_t *parent,
72 hammer2_chain_t *chain,
73 hammer2_tid_t mtid, int flags,
74 hammer2_blockref_t *obref);
75 static hammer2_chain_t *hammer2_combined_find(
76 hammer2_chain_t *parent,
77 hammer2_blockref_t *base, int count,
78 hammer2_key_t *key_nextp,
79 hammer2_key_t key_beg, hammer2_key_t key_end,
80 hammer2_blockref_t **bresp);
81 static hammer2_chain_t *hammer2_chain_lastdrop(hammer2_chain_t *chain,
83 static void hammer2_chain_lru_flush(hammer2_pfs_t *pmp);
86 * There are many degenerate situations where an extreme rate of console
87 * output can occur from warnings and errors. Make sure this output does
88 * not impede operations.
90 static struct krate krate_h2chk = { .freq = 5 };
91 static struct krate krate_h2me = { .freq = 1 };
92 static struct krate krate_h2em = { .freq = 1 };
95 * Basic RBTree for chains (core.rbtree).
97 RB_GENERATE(hammer2_chain_tree, hammer2_chain, rbnode, hammer2_chain_cmp);
100 hammer2_chain_cmp(hammer2_chain_t *chain1, hammer2_chain_t *chain2)
102 hammer2_key_t c1_beg;
103 hammer2_key_t c1_end;
104 hammer2_key_t c2_beg;
105 hammer2_key_t c2_end;
108 * Compare chains. Overlaps are not supposed to happen and catch
109 * any software issues early we count overlaps as a match.
111 c1_beg = chain1->bref.key;
112 c1_end = c1_beg + ((hammer2_key_t)1 << chain1->bref.keybits) - 1;
113 c2_beg = chain2->bref.key;
114 c2_end = c2_beg + ((hammer2_key_t)1 << chain2->bref.keybits) - 1;
116 if (c1_end < c2_beg) /* fully to the left */
118 if (c1_beg > c2_end) /* fully to the right */
120 return(0); /* overlap (must not cross edge boundary) */
124 * Assert that a chain has no media data associated with it.
127 hammer2_chain_assert_no_data(hammer2_chain_t *chain)
129 KKASSERT(chain->dio == NULL);
130 if (chain->bref.type != HAMMER2_BREF_TYPE_VOLUME &&
131 chain->bref.type != HAMMER2_BREF_TYPE_FREEMAP &&
133 panic("hammer2_chain_assert_no_data: chain %p still has data",
139 * Make a chain visible to the flusher. The flusher operates using a top-down
140 * recursion based on the ONFLUSH flag. It locates MODIFIED and UPDATE chains,
141 * flushes them, and updates blocks back to the volume root.
143 * This routine sets the ONFLUSH flag upward from the triggering chain until
144 * it hits an inode root or the volume root. Inode chains serve as inflection
145 * points, requiring the flusher to bridge across trees. Inodes include
146 * regular inodes, PFS roots (pmp->iroot), and the media super root
150 hammer2_chain_setflush(hammer2_chain_t *chain)
152 hammer2_chain_t *parent;
154 if ((chain->flags & HAMMER2_CHAIN_ONFLUSH) == 0) {
155 hammer2_spin_sh(&chain->core.spin);
156 while ((chain->flags & HAMMER2_CHAIN_ONFLUSH) == 0) {
157 atomic_set_int(&chain->flags, HAMMER2_CHAIN_ONFLUSH);
158 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE)
160 if ((parent = chain->parent) == NULL)
162 hammer2_spin_sh(&parent->core.spin);
163 hammer2_spin_unsh(&chain->core.spin);
166 hammer2_spin_unsh(&chain->core.spin);
171 * Allocate a new disconnected chain element representing the specified
172 * bref. chain->refs is set to 1 and the passed bref is copied to
173 * chain->bref. chain->bytes is derived from the bref.
175 * chain->pmp inherits pmp unless the chain is an inode (other than the
178 * NOTE: Returns a referenced but unlocked (because there is no core) chain.
181 hammer2_chain_alloc(hammer2_dev_t *hmp, hammer2_pfs_t *pmp,
182 hammer2_blockref_t *bref)
184 hammer2_chain_t *chain;
188 * Special case - radix of 0 indicates a chain that does not
189 * need a data reference (context is completely embedded in the
192 if ((int)(bref->data_off & HAMMER2_OFF_MASK_RADIX))
193 bytes = 1U << (int)(bref->data_off & HAMMER2_OFF_MASK_RADIX);
198 case HAMMER2_BREF_TYPE_INODE:
199 case HAMMER2_BREF_TYPE_INDIRECT:
200 case HAMMER2_BREF_TYPE_DATA:
201 case HAMMER2_BREF_TYPE_DIRENT:
202 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
203 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
204 case HAMMER2_BREF_TYPE_FREEMAP:
205 case HAMMER2_BREF_TYPE_VOLUME:
206 chain = kmalloc(sizeof(*chain), hmp->mchain, M_WAITOK | M_ZERO);
207 atomic_add_long(&hammer2_chain_allocs, 1);
209 case HAMMER2_BREF_TYPE_EMPTY:
211 panic("hammer2_chain_alloc: unrecognized blockref type: %d",
217 * Initialize the new chain structure. pmp must be set to NULL for
218 * chains belonging to the super-root topology of a device mount.
220 if (pmp == hmp->spmp)
227 chain->bytes = bytes;
229 chain->flags = HAMMER2_CHAIN_ALLOCATED;
230 lockinit(&chain->diolk, "chdio", 0, 0);
233 * Set the PFS boundary flag if this chain represents a PFS root.
235 if (bref->flags & HAMMER2_BREF_FLAG_PFSROOT)
236 atomic_set_int(&chain->flags, HAMMER2_CHAIN_PFSBOUNDARY);
237 hammer2_chain_core_init(chain);
243 * Initialize a chain's core structure. This structure used to be allocated
244 * but is now embedded.
246 * The core is not locked. No additional refs on the chain are made.
247 * (trans) must not be NULL if (core) is not NULL.
250 hammer2_chain_core_init(hammer2_chain_t *chain)
253 * Fresh core under nchain (no multi-homing of ochain's
256 RB_INIT(&chain->core.rbtree); /* live chains */
257 hammer2_mtx_init(&chain->lock, "h2chain");
261 * Add a reference to a chain element, preventing its destruction.
263 * (can be called with spinlock held)
266 hammer2_chain_ref(hammer2_chain_t *chain)
268 if (atomic_fetchadd_int(&chain->refs, 1) == 0) {
270 * Just flag that the chain was used and should be recycled
271 * on the LRU if it encounters it later.
273 if (chain->flags & HAMMER2_CHAIN_ONLRU)
274 atomic_set_int(&chain->flags, HAMMER2_CHAIN_LRUHINT);
278 * REMOVED - reduces contention, lru_list is more heuristical
281 * 0->non-zero transition must ensure that chain is removed
284 * NOTE: Already holding lru_spin here so we cannot call
285 * hammer2_chain_ref() to get it off lru_list, do
288 if (chain->flags & HAMMER2_CHAIN_ONLRU) {
289 hammer2_pfs_t *pmp = chain->pmp;
290 hammer2_spin_ex(&pmp->lru_spin);
291 if (chain->flags & HAMMER2_CHAIN_ONLRU) {
292 atomic_add_int(&pmp->lru_count, -1);
293 atomic_clear_int(&chain->flags,
294 HAMMER2_CHAIN_ONLRU);
295 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node);
297 hammer2_spin_unex(&pmp->lru_spin);
304 * Ref a locked chain and force the data to be held across an unlock.
305 * Chain must be currently locked. The user of the chain who desires
306 * to release the hold must call hammer2_chain_lock_unhold() to relock
307 * and unhold the chain, then unlock normally, or may simply call
308 * hammer2_chain_drop_unhold() (which is safer against deadlocks).
311 hammer2_chain_ref_hold(hammer2_chain_t *chain)
313 atomic_add_int(&chain->lockcnt, 1);
314 hammer2_chain_ref(chain);
318 * Insert the chain in the core rbtree.
320 * Normal insertions are placed in the live rbtree. Insertion of a deleted
321 * chain is a special case used by the flush code that is placed on the
322 * unstaged deleted list to avoid confusing the live view.
324 #define HAMMER2_CHAIN_INSERT_SPIN 0x0001
325 #define HAMMER2_CHAIN_INSERT_LIVE 0x0002
326 #define HAMMER2_CHAIN_INSERT_RACE 0x0004
330 hammer2_chain_insert(hammer2_chain_t *parent, hammer2_chain_t *chain,
331 int flags, int generation)
333 hammer2_chain_t *xchain;
336 if (flags & HAMMER2_CHAIN_INSERT_SPIN)
337 hammer2_spin_ex(&parent->core.spin);
340 * Interlocked by spinlock, check for race
342 if ((flags & HAMMER2_CHAIN_INSERT_RACE) &&
343 parent->core.generation != generation) {
344 error = HAMMER2_ERROR_EAGAIN;
351 xchain = RB_INSERT(hammer2_chain_tree, &parent->core.rbtree, chain);
352 KASSERT(xchain == NULL,
353 ("hammer2_chain_insert: collision %p %p (key=%016jx)",
354 chain, xchain, chain->bref.key));
355 atomic_set_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE);
356 chain->parent = parent;
357 ++parent->core.chain_count;
358 ++parent->core.generation; /* XXX incs for _get() too, XXX */
361 * We have to keep track of the effective live-view blockref count
362 * so the create code knows when to push an indirect block.
364 if (flags & HAMMER2_CHAIN_INSERT_LIVE)
365 atomic_add_int(&parent->core.live_count, 1);
367 if (flags & HAMMER2_CHAIN_INSERT_SPIN)
368 hammer2_spin_unex(&parent->core.spin);
373 * Drop the caller's reference to the chain. When the ref count drops to
374 * zero this function will try to disassociate the chain from its parent and
375 * deallocate it, then recursely drop the parent using the implied ref
376 * from the chain's chain->parent.
378 * Nobody should own chain's mutex on the 1->0 transition, unless this drop
379 * races an acquisition by another cpu. Therefore we can loop if we are
380 * unable to acquire the mutex, and refs is unlikely to be 1 unless we again
381 * race against another drop.
384 hammer2_chain_drop(hammer2_chain_t *chain)
388 KKASSERT(chain->refs > 0);
396 if (hammer2_mtx_ex_try(&chain->lock) == 0)
397 chain = hammer2_chain_lastdrop(chain, 0);
398 /* retry the same chain, or chain from lastdrop */
400 if (atomic_cmpset_int(&chain->refs, refs, refs - 1))
402 /* retry the same chain */
409 * Unhold a held and probably not-locked chain, ensure that the data is
410 * dropped on the 1->0 transition of lockcnt by obtaining an exclusive
411 * lock and then simply unlocking the chain.
414 hammer2_chain_unhold(hammer2_chain_t *chain)
420 lockcnt = chain->lockcnt;
423 if (atomic_cmpset_int(&chain->lockcnt,
424 lockcnt, lockcnt - 1)) {
427 } else if (hammer2_mtx_ex_try(&chain->lock) == 0) {
428 hammer2_chain_unlock(chain);
432 * This situation can easily occur on SMP due to
433 * the gap inbetween the 1->0 transition and the
434 * final unlock. We cannot safely block on the
435 * mutex because lockcnt might go above 1.
437 * XXX Sleep for one tick if it takes too long.
440 if (iter > 1000 + hz) {
441 kprintf("hammer2: h2race1 %p\n", chain);
444 tsleep(&iter, 0, "h2race1", 1);
452 hammer2_chain_drop_unhold(hammer2_chain_t *chain)
454 hammer2_chain_unhold(chain);
455 hammer2_chain_drop(chain);
459 hammer2_chain_rehold(hammer2_chain_t *chain)
461 hammer2_chain_lock(chain, HAMMER2_RESOLVE_SHARED);
462 atomic_add_int(&chain->lockcnt, 1);
463 hammer2_chain_unlock(chain);
467 * Handles the (potential) last drop of chain->refs from 1->0. Called with
468 * the mutex exclusively locked, refs == 1, and lockcnt 0. SMP races are
469 * possible against refs and lockcnt. We must dispose of the mutex on chain.
471 * This function returns an unlocked chain for recursive drop or NULL. It
472 * can return the same chain if it determines it has raced another ref.
476 * When two chains need to be recursively dropped we use the chain we
477 * would otherwise free to placehold the additional chain. It's a bit
478 * convoluted but we can't just recurse without potentially blowing out
481 * The chain cannot be freed if it has any children.
482 * The chain cannot be freed if flagged MODIFIED unless we can dispose of it.
483 * The chain cannot be freed if flagged UPDATE unless we can dispose of it.
484 * Any dedup registration can remain intact.
486 * The core spinlock is allowed to nest child-to-parent (not parent-to-child).
490 hammer2_chain_lastdrop(hammer2_chain_t *chain, int depth)
494 hammer2_chain_t *parent;
495 hammer2_chain_t *rdrop;
498 * We need chain's spinlock to interlock the sub-tree test.
499 * We already have chain's mutex, protecting chain->parent.
501 * Remember that chain->refs can be in flux.
503 hammer2_spin_ex(&chain->core.spin);
505 if (chain->parent != NULL) {
507 * If the chain has a parent the UPDATE bit prevents scrapping
508 * as the chain is needed to properly flush the parent. Try
509 * to complete the 1->0 transition and return NULL. Retry
510 * (return chain) if we are unable to complete the 1->0
511 * transition, else return NULL (nothing more to do).
513 * If the chain has a parent the MODIFIED bit prevents
516 * Chains with UPDATE/MODIFIED are *not* put on the LRU list!
518 if (chain->flags & (HAMMER2_CHAIN_UPDATE |
519 HAMMER2_CHAIN_MODIFIED)) {
520 if (atomic_cmpset_int(&chain->refs, 1, 0)) {
521 hammer2_spin_unex(&chain->core.spin);
522 hammer2_chain_assert_no_data(chain);
523 hammer2_mtx_unlock(&chain->lock);
526 hammer2_spin_unex(&chain->core.spin);
527 hammer2_mtx_unlock(&chain->lock);
531 /* spinlock still held */
532 } else if (chain->bref.type == HAMMER2_BREF_TYPE_VOLUME ||
533 chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP) {
535 * Retain the static vchain and fchain. Clear bits that
536 * are not relevant. Do not clear the MODIFIED bit,
537 * and certainly do not put it on the delayed-flush queue.
539 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
542 * The chain has no parent and can be flagged for destruction.
543 * Since it has no parent, UPDATE can also be cleared.
545 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DESTROY);
546 if (chain->flags & HAMMER2_CHAIN_UPDATE)
547 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
550 * If the chain has children we must propagate the DESTROY
551 * flag downward and rip the disconnected topology apart.
552 * This is accomplished by calling hammer2_flush() on the
555 * Any dedup is already handled by the underlying DIO, so
556 * we do not have to specifically flush it here.
558 if (chain->core.chain_count) {
559 hammer2_spin_unex(&chain->core.spin);
560 hammer2_flush(chain, HAMMER2_FLUSH_TOP |
562 hammer2_mtx_unlock(&chain->lock);
564 return(chain); /* retry drop */
568 * Otherwise we can scrap the MODIFIED bit if it is set,
569 * and continue along the freeing path.
571 * Be sure to clean-out any dedup bits. Without a parent
572 * this chain will no longer be visible to the flush code.
573 * Easy check data_off to avoid the volume root.
575 if (chain->flags & HAMMER2_CHAIN_MODIFIED) {
576 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
577 atomic_add_long(&hammer2_count_modified_chains, -1);
579 hammer2_pfs_memory_wakeup(chain->pmp, -1);
581 /* spinlock still held */
584 /* spinlock still held */
587 * If any children exist we must leave the chain intact with refs == 0.
588 * They exist because chains are retained below us which have refs or
589 * may require flushing.
591 * Retry (return chain) if we fail to transition the refs to 0, else
592 * return NULL indication nothing more to do.
594 * Chains with children are NOT put on the LRU list.
596 if (chain->core.chain_count) {
597 if (atomic_cmpset_int(&chain->refs, 1, 0)) {
598 hammer2_spin_unex(&chain->core.spin);
599 hammer2_chain_assert_no_data(chain);
600 hammer2_mtx_unlock(&chain->lock);
603 hammer2_spin_unex(&chain->core.spin);
604 hammer2_mtx_unlock(&chain->lock);
608 /* spinlock still held */
609 /* no chains left under us */
612 * chain->core has no children left so no accessors can get to our
613 * chain from there. Now we have to lock the parent core to interlock
614 * remaining possible accessors that might bump chain's refs before
615 * we can safely drop chain's refs with intent to free the chain.
618 pmp = chain->pmp; /* can be NULL */
621 parent = chain->parent;
624 * WARNING! chain's spin lock is still held here, and other spinlocks
625 * will be acquired and released in the code below. We
626 * cannot be making fancy procedure calls!
630 * We can cache the chain if it is associated with a pmp
631 * and not flagged as being destroyed or requesting a full
632 * release. In this situation the chain is not removed
633 * from its parent, i.e. it can still be looked up.
635 * We intentionally do not cache DATA chains because these
636 * were likely used to load data into the logical buffer cache
637 * and will not be accessed again for some time.
640 (HAMMER2_CHAIN_DESTROY | HAMMER2_CHAIN_RELEASE)) == 0 &&
642 chain->bref.type != HAMMER2_BREF_TYPE_DATA) {
644 hammer2_spin_ex(&parent->core.spin);
645 if (atomic_cmpset_int(&chain->refs, 1, 0) == 0) {
647 * 1->0 transition failed, retry. Do not drop
648 * the chain's data yet!
651 hammer2_spin_unex(&parent->core.spin);
652 hammer2_spin_unex(&chain->core.spin);
653 hammer2_mtx_unlock(&chain->lock);
661 hammer2_chain_assert_no_data(chain);
664 * Make sure we are on the LRU list, clean up excessive
665 * LRU entries. We can only really drop one but there might
666 * be other entries that we can remove from the lru_list
669 * NOTE: HAMMER2_CHAIN_ONLRU may only be safely set when
670 * chain->core.spin AND pmp->lru_spin are held, but
671 * can be safely cleared only holding pmp->lru_spin.
673 if ((chain->flags & HAMMER2_CHAIN_ONLRU) == 0) {
674 hammer2_spin_ex(&pmp->lru_spin);
675 if ((chain->flags & HAMMER2_CHAIN_ONLRU) == 0) {
676 atomic_set_int(&chain->flags,
677 HAMMER2_CHAIN_ONLRU);
678 TAILQ_INSERT_TAIL(&pmp->lru_list,
680 atomic_add_int(&pmp->lru_count, 1);
682 if (pmp->lru_count < HAMMER2_LRU_LIMIT)
683 depth = 1; /* disable lru_list flush */
684 hammer2_spin_unex(&pmp->lru_spin);
686 /* disable lru flush */
691 hammer2_spin_unex(&parent->core.spin);
692 parent = NULL; /* safety */
694 hammer2_spin_unex(&chain->core.spin);
695 hammer2_mtx_unlock(&chain->lock);
698 * lru_list hysteresis (see above for depth overrides).
699 * Note that depth also prevents excessive lastdrop recursion.
702 hammer2_chain_lru_flush(pmp);
709 * Make sure we are not on the LRU list.
711 if (chain->flags & HAMMER2_CHAIN_ONLRU) {
712 hammer2_spin_ex(&pmp->lru_spin);
713 if (chain->flags & HAMMER2_CHAIN_ONLRU) {
714 atomic_add_int(&pmp->lru_count, -1);
715 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU);
716 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node);
718 hammer2_spin_unex(&pmp->lru_spin);
722 * Spinlock the parent and try to drop the last ref on chain.
723 * On success determine if we should dispose of the chain
724 * (remove the chain from its parent, etc).
726 * (normal core locks are top-down recursive but we define
727 * core spinlocks as bottom-up recursive, so this is safe).
730 hammer2_spin_ex(&parent->core.spin);
731 if (atomic_cmpset_int(&chain->refs, 1, 0) == 0) {
733 * 1->0 transition failed, retry.
735 hammer2_spin_unex(&parent->core.spin);
736 hammer2_spin_unex(&chain->core.spin);
737 hammer2_mtx_unlock(&chain->lock);
743 * 1->0 transition successful, parent spin held to prevent
744 * new lookups, chain spinlock held to protect parent field.
745 * Remove chain from the parent.
747 * If the chain is being removed from the parent's btree but
748 * is not bmapped, we have to adjust live_count downward. If
749 * it is bmapped then the blockref is retained in the parent
750 * as is its associated live_count. This case can occur when
751 * a chain added to the topology is unable to flush and is
752 * then later deleted.
754 if (chain->flags & HAMMER2_CHAIN_ONRBTREE) {
755 if ((parent->flags & HAMMER2_CHAIN_COUNTEDBREFS) &&
756 (chain->flags & HAMMER2_CHAIN_BMAPPED) == 0) {
757 atomic_add_int(&parent->core.live_count, -1);
759 RB_REMOVE(hammer2_chain_tree,
760 &parent->core.rbtree, chain);
761 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE);
762 --parent->core.chain_count;
763 chain->parent = NULL;
767 * If our chain was the last chain in the parent's core the
768 * core is now empty and its parent might have to be
769 * re-dropped if it has 0 refs.
771 if (parent->core.chain_count == 0) {
773 atomic_add_int(&rdrop->refs, 1);
775 if (atomic_cmpset_int(&rdrop->refs, 0, 1) == 0)
779 hammer2_spin_unex(&parent->core.spin);
780 parent = NULL; /* safety */
786 if (atomic_cmpset_int(&chain->refs, 1, 0) == 0) {
788 * 1->0 transition failed, retry.
790 hammer2_spin_unex(&parent->core.spin);
791 hammer2_spin_unex(&chain->core.spin);
792 hammer2_mtx_unlock(&chain->lock);
799 * Successful 1->0 transition, no parent, no children... no way for
800 * anyone to ref this chain any more. We can clean-up and free it.
802 * We still have the core spinlock, and core's chain_count is 0.
803 * Any parent spinlock is gone.
805 hammer2_spin_unex(&chain->core.spin);
806 hammer2_chain_assert_no_data(chain);
807 hammer2_mtx_unlock(&chain->lock);
808 KKASSERT(RB_EMPTY(&chain->core.rbtree) &&
809 chain->core.chain_count == 0);
812 * All locks are gone, no pointers remain to the chain, finish
815 KKASSERT((chain->flags & (HAMMER2_CHAIN_UPDATE |
816 HAMMER2_CHAIN_MODIFIED)) == 0);
819 * Once chain resources are gone we can use the now dead chain
820 * structure to placehold what might otherwise require a recursive
821 * drop, because we have potentially two things to drop and can only
822 * return one directly.
824 if (chain->flags & HAMMER2_CHAIN_ALLOCATED) {
825 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ALLOCATED);
827 kfree(chain, hmp->mchain);
831 * Possible chaining loop when parent re-drop needed.
837 * Heuristical flush of the LRU, try to reduce the number of entries
838 * on the LRU to (HAMMER2_LRU_LIMIT * 2 / 3). This procedure is called
839 * only when lru_count exceeds HAMMER2_LRU_LIMIT.
843 hammer2_chain_lru_flush(hammer2_pfs_t *pmp)
845 hammer2_chain_t *chain;
849 hammer2_spin_ex(&pmp->lru_spin);
850 while (pmp->lru_count > HAMMER2_LRU_LIMIT * 2 / 3) {
852 * Pick a chain off the lru_list, just recycle it quickly
853 * if LRUHINT is set (the chain was ref'd but left on
854 * the lru_list, so cycle to the end).
856 chain = TAILQ_FIRST(&pmp->lru_list);
857 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node);
859 if (chain->flags & HAMMER2_CHAIN_LRUHINT) {
860 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_LRUHINT);
861 TAILQ_INSERT_TAIL(&pmp->lru_list, chain, lru_node);
867 * Ok, we are off the LRU. We must adjust refs before we
868 * can safely clear the ONLRU flag.
870 atomic_add_int(&pmp->lru_count, -1);
871 if (atomic_cmpset_int(&chain->refs, 0, 1)) {
872 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU);
873 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
876 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU);
879 hammer2_spin_unex(&pmp->lru_spin);
884 * If we picked a chain off the lru list we may be able to lastdrop
885 * it. Use a depth of 1 to prevent excessive lastdrop recursion.
895 if (hammer2_mtx_ex_try(&chain->lock) == 0)
896 chain = hammer2_chain_lastdrop(chain, 1);
897 /* retry the same chain, or chain from lastdrop */
899 if (atomic_cmpset_int(&chain->refs, refs, refs - 1))
901 /* retry the same chain */
909 * On last lock release.
911 static hammer2_io_t *
912 hammer2_chain_drop_data(hammer2_chain_t *chain)
916 if ((dio = chain->dio) != NULL) {
920 switch(chain->bref.type) {
921 case HAMMER2_BREF_TYPE_VOLUME:
922 case HAMMER2_BREF_TYPE_FREEMAP:
925 if (chain->data != NULL) {
926 hammer2_spin_unex(&chain->core.spin);
927 panic("chain data not null: "
928 "chain %p bref %016jx.%02x "
929 "refs %d parent %p dio %p data %p",
930 chain, chain->bref.data_off,
931 chain->bref.type, chain->refs,
933 chain->dio, chain->data);
935 KKASSERT(chain->data == NULL);
943 * Lock a referenced chain element, acquiring its data with I/O if necessary,
944 * and specify how you would like the data to be resolved.
946 * If an I/O or other fatal error occurs, chain->error will be set to non-zero.
948 * The lock is allowed to recurse, multiple locking ops will aggregate
949 * the requested resolve types. Once data is assigned it will not be
950 * removed until the last unlock.
952 * HAMMER2_RESOLVE_NEVER - Do not resolve the data element.
953 * (typically used to avoid device/logical buffer
956 * HAMMER2_RESOLVE_MAYBE - Do not resolve data elements for chains in
957 * the INITIAL-create state (indirect blocks only).
959 * Do not resolve data elements for DATA chains.
960 * (typically used to avoid device/logical buffer
963 * HAMMER2_RESOLVE_ALWAYS- Always resolve the data element.
965 * HAMMER2_RESOLVE_SHARED- (flag) The chain is locked shared, otherwise
966 * it will be locked exclusive.
968 * HAMMER2_RESOLVE_NONBLOCK- (flag) The chain is locked non-blocking. If
969 * the lock fails, EAGAIN is returned.
971 * NOTE: Embedded elements (volume header, inodes) are always resolved
974 * NOTE: Specifying HAMMER2_RESOLVE_ALWAYS on a newly-created non-embedded
975 * element will instantiate and zero its buffer, and flush it on
978 * NOTE: (data) elements are normally locked RESOLVE_NEVER or RESOLVE_MAYBE
979 * so as not to instantiate a device buffer, which could alias against
980 * a logical file buffer. However, if ALWAYS is specified the
981 * device buffer will be instantiated anyway.
983 * NOTE: The return value is always 0 unless NONBLOCK is specified, in which
984 * case it can be either 0 or EAGAIN.
986 * WARNING! This function blocks on I/O if data needs to be fetched. This
987 * blocking can run concurrent with other compatible lock holders
988 * who do not need data returning. The lock is not upgraded to
989 * exclusive during a data fetch, a separate bit is used to
990 * interlock I/O. However, an exclusive lock holder can still count
991 * on being interlocked against an I/O fetch managed by a shared
995 hammer2_chain_lock(hammer2_chain_t *chain, int how)
997 KKASSERT(chain->refs > 0);
999 if (how & HAMMER2_RESOLVE_NONBLOCK) {
1001 * We still have to bump lockcnt before acquiring the lock,
1002 * even for non-blocking operation, because the unlock code
1003 * live-loops on lockcnt == 1 when dropping the last lock.
1005 * If the non-blocking operation fails we have to use an
1006 * unhold sequence to undo the mess.
1008 * NOTE: LOCKAGAIN must always succeed without blocking,
1009 * even if NONBLOCK is specified.
1011 atomic_add_int(&chain->lockcnt, 1);
1012 if (how & HAMMER2_RESOLVE_SHARED) {
1013 if (how & HAMMER2_RESOLVE_LOCKAGAIN) {
1014 hammer2_mtx_sh_again(&chain->lock);
1016 if (hammer2_mtx_sh_try(&chain->lock) != 0) {
1017 hammer2_chain_unhold(chain);
1022 if (hammer2_mtx_ex_try(&chain->lock) != 0) {
1023 hammer2_chain_unhold(chain);
1029 * Get the appropriate lock. If LOCKAGAIN is flagged with
1030 * SHARED the caller expects a shared lock to already be
1031 * present and we are giving it another ref. This case must
1032 * importantly not block if there is a pending exclusive lock
1035 atomic_add_int(&chain->lockcnt, 1);
1036 if (how & HAMMER2_RESOLVE_SHARED) {
1037 if (how & HAMMER2_RESOLVE_LOCKAGAIN) {
1038 hammer2_mtx_sh_again(&chain->lock);
1040 hammer2_mtx_sh(&chain->lock);
1043 hammer2_mtx_ex(&chain->lock);
1048 * If we already have a valid data pointer make sure the data is
1049 * synchronized to the current cpu, and then no further action is
1054 hammer2_io_bkvasync(chain->dio);
1059 * Do we have to resolve the data? This is generally only
1060 * applicable to HAMMER2_BREF_TYPE_DATA which is special-cased.
1061 * Other BREF types expects the data to be there.
1063 switch(how & HAMMER2_RESOLVE_MASK) {
1064 case HAMMER2_RESOLVE_NEVER:
1066 case HAMMER2_RESOLVE_MAYBE:
1067 if (chain->flags & HAMMER2_CHAIN_INITIAL)
1069 if (chain->bref.type == HAMMER2_BREF_TYPE_DATA)
1072 if (chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE)
1074 if (chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_LEAF)
1078 case HAMMER2_RESOLVE_ALWAYS:
1084 * Caller requires data
1086 hammer2_chain_load_data(chain);
1092 * Lock the chain, retain the hold, and drop the data persistence count.
1093 * The data should remain valid because we never transitioned lockcnt
1097 hammer2_chain_lock_unhold(hammer2_chain_t *chain, int how)
1099 hammer2_chain_lock(chain, how);
1100 atomic_add_int(&chain->lockcnt, -1);
1105 * Downgrade an exclusive chain lock to a shared chain lock.
1107 * NOTE: There is no upgrade equivalent due to the ease of
1108 * deadlocks in that direction.
1111 hammer2_chain_lock_downgrade(hammer2_chain_t *chain)
1113 hammer2_mtx_downgrade(&chain->lock);
1118 * Issue I/O and install chain->data. Caller must hold a chain lock, lock
1119 * may be of any type.
1121 * Once chain->data is set it cannot be disposed of until all locks are
1124 * Make sure the data is synchronized to the current cpu.
1127 hammer2_chain_load_data(hammer2_chain_t *chain)
1129 hammer2_blockref_t *bref;
1136 * Degenerate case, data already present, or chain has no media
1137 * reference to load.
1139 KKASSERT(chain->lock.mtx_lock & MTX_MASK);
1142 hammer2_io_bkvasync(chain->dio);
1145 if ((chain->bref.data_off & ~HAMMER2_OFF_MASK_RADIX) == 0)
1149 KKASSERT(hmp != NULL);
1152 * Gain the IOINPROG bit, interlocked block.
1158 oflags = chain->flags;
1160 if (oflags & HAMMER2_CHAIN_IOINPROG) {
1161 nflags = oflags | HAMMER2_CHAIN_IOSIGNAL;
1162 tsleep_interlock(&chain->flags, 0);
1163 if (atomic_cmpset_int(&chain->flags, oflags, nflags)) {
1164 tsleep(&chain->flags, PINTERLOCKED,
1169 nflags = oflags | HAMMER2_CHAIN_IOINPROG;
1170 if (atomic_cmpset_int(&chain->flags, oflags, nflags)) {
1178 * We own CHAIN_IOINPROG
1180 * Degenerate case if we raced another load.
1184 hammer2_io_bkvasync(chain->dio);
1189 * We must resolve to a device buffer, either by issuing I/O or
1190 * by creating a zero-fill element. We do not mark the buffer
1191 * dirty when creating a zero-fill element (the hammer2_chain_modify()
1192 * API must still be used to do that).
1194 * The device buffer is variable-sized in powers of 2 down
1195 * to HAMMER2_MIN_ALLOC (typically 1K). A 64K physical storage
1196 * chunk always contains buffers of the same size. (XXX)
1198 * The minimum physical IO size may be larger than the variable
1201 bref = &chain->bref;
1204 * The getblk() optimization can only be used on newly created
1205 * elements if the physical block size matches the request.
1207 if (chain->flags & HAMMER2_CHAIN_INITIAL) {
1208 error = hammer2_io_new(hmp, bref->type,
1209 bref->data_off, chain->bytes,
1212 error = hammer2_io_bread(hmp, bref->type,
1213 bref->data_off, chain->bytes,
1215 hammer2_adjreadcounter(chain->bref.type, chain->bytes);
1218 chain->error = HAMMER2_ERROR_EIO;
1219 kprintf("hammer2_chain_load_data: I/O error %016jx: %d\n",
1220 (intmax_t)bref->data_off, error);
1221 hammer2_io_bqrelse(&chain->dio);
1227 * This isn't perfect and can be ignored on OSs which do not have
1228 * an indication as to whether a buffer is coming from cache or
1229 * if I/O was actually issued for the read. TESTEDGOOD will work
1230 * pretty well without the B_IOISSUED logic because chains are
1231 * cached, but in that situation (without B_IOISSUED) it will not
1232 * detect whether a re-read via I/O is corrupted verses the original
1235 * We can't re-run the CRC on every fresh lock. That would be
1236 * insanely expensive.
1238 * If the underlying kernel buffer covers the entire chain we can
1239 * use the B_IOISSUED indication to determine if we have to re-run
1240 * the CRC on chain data for chains that managed to stay cached
1241 * across the kernel disposal of the original buffer.
1243 if ((dio = chain->dio) != NULL && dio->bp) {
1244 struct buf *bp = dio->bp;
1246 if (dio->psize == chain->bytes &&
1247 (bp->b_flags & B_IOISSUED)) {
1248 atomic_clear_int(&chain->flags,
1249 HAMMER2_CHAIN_TESTEDGOOD);
1250 bp->b_flags &= ~B_IOISSUED;
1255 * NOTE: A locked chain's data cannot be modified without first
1256 * calling hammer2_chain_modify().
1260 * NOTE: hammer2_io_data() call issues bkvasync()
1262 bdata = hammer2_io_data(chain->dio, chain->bref.data_off);
1264 if (chain->flags & HAMMER2_CHAIN_INITIAL) {
1266 * Clear INITIAL. In this case we used io_new() and the
1267 * buffer has been zero'd and marked dirty.
1269 * CHAIN_MODIFIED has not been set yet, and we leave it
1270 * that way for now. Set a temporary CHAIN_NOTTESTED flag
1271 * to prevent hammer2_chain_testcheck() from trying to match
1272 * a check code that has not yet been generated. This bit
1273 * should NOT end up on the actual media.
1275 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1276 atomic_set_int(&chain->flags, HAMMER2_CHAIN_NOTTESTED);
1277 } else if (chain->flags & HAMMER2_CHAIN_MODIFIED) {
1279 * check data not currently synchronized due to
1280 * modification. XXX assumes data stays in the buffer
1281 * cache, which might not be true (need biodep on flush
1282 * to calculate crc? or simple crc?).
1284 } else if ((chain->flags & HAMMER2_CHAIN_TESTEDGOOD) == 0) {
1285 if (hammer2_chain_testcheck(chain, bdata) == 0) {
1286 chain->error = HAMMER2_ERROR_CHECK;
1288 atomic_set_int(&chain->flags, HAMMER2_CHAIN_TESTEDGOOD);
1293 * Setup the data pointer, either pointing it to an embedded data
1294 * structure and copying the data from the buffer, or pointing it
1297 * The buffer is not retained when copying to an embedded data
1298 * structure in order to avoid potential deadlocks or recursions
1299 * on the same physical buffer.
1301 * WARNING! Other threads can start using the data the instant we
1302 * set chain->data non-NULL.
1304 switch (bref->type) {
1305 case HAMMER2_BREF_TYPE_VOLUME:
1306 case HAMMER2_BREF_TYPE_FREEMAP:
1308 * Copy data from bp to embedded buffer
1310 panic("hammer2_chain_load_data: unresolved volume header");
1312 case HAMMER2_BREF_TYPE_DIRENT:
1313 KKASSERT(chain->bytes != 0);
1315 case HAMMER2_BREF_TYPE_INODE:
1316 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
1317 case HAMMER2_BREF_TYPE_INDIRECT:
1318 case HAMMER2_BREF_TYPE_DATA:
1319 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1322 * Point data at the device buffer and leave dio intact.
1324 chain->data = (void *)bdata;
1329 * Release HAMMER2_CHAIN_IOINPROG and signal waiters if requested.
1336 oflags = chain->flags;
1337 nflags = oflags & ~(HAMMER2_CHAIN_IOINPROG |
1338 HAMMER2_CHAIN_IOSIGNAL);
1339 KKASSERT(oflags & HAMMER2_CHAIN_IOINPROG);
1340 if (atomic_cmpset_int(&chain->flags, oflags, nflags)) {
1341 if (oflags & HAMMER2_CHAIN_IOSIGNAL)
1342 wakeup(&chain->flags);
1349 * Unlock and deref a chain element.
1351 * Remember that the presence of children under chain prevent the chain's
1352 * destruction but do not add additional references, so the dio will still
1356 hammer2_chain_unlock(hammer2_chain_t *chain)
1363 * If multiple locks are present (or being attempted) on this
1364 * particular chain we can just unlock, drop refs, and return.
1366 * Otherwise fall-through on the 1->0 transition.
1369 lockcnt = chain->lockcnt;
1370 KKASSERT(lockcnt > 0);
1373 if (atomic_cmpset_int(&chain->lockcnt,
1374 lockcnt, lockcnt - 1)) {
1375 hammer2_mtx_unlock(&chain->lock);
1378 } else if (hammer2_mtx_upgrade_try(&chain->lock) == 0) {
1379 /* while holding the mutex exclusively */
1380 if (atomic_cmpset_int(&chain->lockcnt, 1, 0))
1384 * This situation can easily occur on SMP due to
1385 * the gap inbetween the 1->0 transition and the
1386 * final unlock. We cannot safely block on the
1387 * mutex because lockcnt might go above 1.
1389 * XXX Sleep for one tick if it takes too long.
1391 if (++iter > 1000) {
1392 if (iter > 1000 + hz) {
1393 kprintf("hammer2: h2race2 %p\n", chain);
1396 tsleep(&iter, 0, "h2race2", 1);
1404 * Last unlock / mutex upgraded to exclusive. Drop the data
1407 dio = hammer2_chain_drop_data(chain);
1409 hammer2_io_bqrelse(&dio);
1410 hammer2_mtx_unlock(&chain->lock);
1414 * Unlock and hold chain data intact
1417 hammer2_chain_unlock_hold(hammer2_chain_t *chain)
1419 atomic_add_int(&chain->lockcnt, 1);
1420 hammer2_chain_unlock(chain);
1424 * Helper to obtain the blockref[] array base and count for a chain.
1426 * XXX Not widely used yet, various use cases need to be validated and
1427 * converted to use this function.
1430 hammer2_blockref_t *
1431 hammer2_chain_base_and_count(hammer2_chain_t *parent, int *countp)
1433 hammer2_blockref_t *base;
1436 if (parent->flags & HAMMER2_CHAIN_INITIAL) {
1439 switch(parent->bref.type) {
1440 case HAMMER2_BREF_TYPE_INODE:
1441 count = HAMMER2_SET_COUNT;
1443 case HAMMER2_BREF_TYPE_INDIRECT:
1444 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1445 count = parent->bytes / sizeof(hammer2_blockref_t);
1447 case HAMMER2_BREF_TYPE_VOLUME:
1448 count = HAMMER2_SET_COUNT;
1450 case HAMMER2_BREF_TYPE_FREEMAP:
1451 count = HAMMER2_SET_COUNT;
1454 panic("hammer2_chain_base_and_count: "
1455 "unrecognized blockref type: %d",
1461 switch(parent->bref.type) {
1462 case HAMMER2_BREF_TYPE_INODE:
1463 base = &parent->data->ipdata.u.blockset.blockref[0];
1464 count = HAMMER2_SET_COUNT;
1466 case HAMMER2_BREF_TYPE_INDIRECT:
1467 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1468 base = &parent->data->npdata[0];
1469 count = parent->bytes / sizeof(hammer2_blockref_t);
1471 case HAMMER2_BREF_TYPE_VOLUME:
1472 base = &parent->data->voldata.
1473 sroot_blockset.blockref[0];
1474 count = HAMMER2_SET_COUNT;
1476 case HAMMER2_BREF_TYPE_FREEMAP:
1477 base = &parent->data->blkset.blockref[0];
1478 count = HAMMER2_SET_COUNT;
1481 panic("hammer2_chain_base_and_count: "
1482 "unrecognized blockref type: %d",
1495 * This counts the number of live blockrefs in a block array and
1496 * also calculates the point at which all remaining blockrefs are empty.
1497 * This routine can only be called on a live chain.
1499 * Caller holds the chain locked, but possibly with a shared lock. We
1500 * must use an exclusive spinlock to prevent corruption.
1502 * NOTE: Flag is not set until after the count is complete, allowing
1503 * callers to test the flag without holding the spinlock.
1505 * NOTE: If base is NULL the related chain is still in the INITIAL
1506 * state and there are no blockrefs to count.
1508 * NOTE: live_count may already have some counts accumulated due to
1509 * creation and deletion and could even be initially negative.
1512 hammer2_chain_countbrefs(hammer2_chain_t *chain,
1513 hammer2_blockref_t *base, int count)
1515 hammer2_spin_ex(&chain->core.spin);
1516 if ((chain->flags & HAMMER2_CHAIN_COUNTEDBREFS) == 0) {
1518 while (--count >= 0) {
1519 if (base[count].type != HAMMER2_BREF_TYPE_EMPTY)
1522 chain->core.live_zero = count + 1;
1523 while (count >= 0) {
1524 if (base[count].type != HAMMER2_BREF_TYPE_EMPTY)
1525 atomic_add_int(&chain->core.live_count,
1530 chain->core.live_zero = 0;
1532 /* else do not modify live_count */
1533 atomic_set_int(&chain->flags, HAMMER2_CHAIN_COUNTEDBREFS);
1535 hammer2_spin_unex(&chain->core.spin);
1539 * Resize the chain's physical storage allocation in-place. This function does
1540 * not usually adjust the data pointer and must be followed by (typically) a
1541 * hammer2_chain_modify() call to copy any old data over and adjust the
1544 * Chains can be resized smaller without reallocating the storage. Resizing
1545 * larger will reallocate the storage. Excess or prior storage is reclaimed
1546 * asynchronously at a later time.
1548 * An nradix value of 0 is special-cased to mean that the storage should
1549 * be disassociated, that is the chain is being resized to 0 bytes (not 1
1552 * Must be passed an exclusively locked parent and chain.
1554 * This function is mostly used with DATA blocks locked RESOLVE_NEVER in order
1555 * to avoid instantiating a device buffer that conflicts with the vnode data
1556 * buffer. However, because H2 can compress or encrypt data, the chain may
1557 * have a dio assigned to it in those situations, and they do not conflict.
1559 * XXX return error if cannot resize.
1562 hammer2_chain_resize(hammer2_chain_t *chain,
1563 hammer2_tid_t mtid, hammer2_off_t dedup_off,
1564 int nradix, int flags)
1574 * Only data and indirect blocks can be resized for now.
1575 * (The volu root, inodes, and freemap elements use a fixed size).
1577 KKASSERT(chain != &hmp->vchain);
1578 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_DATA ||
1579 chain->bref.type == HAMMER2_BREF_TYPE_INDIRECT ||
1580 chain->bref.type == HAMMER2_BREF_TYPE_DIRENT);
1583 * Nothing to do if the element is already the proper size
1585 obytes = chain->bytes;
1586 nbytes = (nradix) ? (1U << nradix) : 0;
1587 if (obytes == nbytes)
1588 return (chain->error);
1591 * Make sure the old data is instantiated so we can copy it. If this
1592 * is a data block, the device data may be superfluous since the data
1593 * might be in a logical block, but compressed or encrypted data is
1596 * NOTE: The modify will set BMAPUPD for us if BMAPPED is set.
1598 error = hammer2_chain_modify(chain, mtid, dedup_off, 0);
1603 * Reallocate the block, even if making it smaller (because different
1604 * block sizes may be in different regions).
1606 * NOTE: Operation does not copy the data and may only be used
1607 * to resize data blocks in-place, or directory entry blocks
1608 * which are about to be modified in some manner.
1610 error = hammer2_freemap_alloc(chain, nbytes);
1614 chain->bytes = nbytes;
1617 * We don't want the followup chain_modify() to try to copy data
1618 * from the old (wrong-sized) buffer. It won't know how much to
1619 * copy. This case should only occur during writes when the
1620 * originator already has the data to write in-hand.
1623 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_DATA ||
1624 chain->bref.type == HAMMER2_BREF_TYPE_DIRENT);
1625 hammer2_io_brelse(&chain->dio);
1628 return (chain->error);
1632 * Set the chain modified so its data can be changed by the caller, or
1633 * install deduplicated data. The caller must call this routine for each
1634 * set of modifications it makes, even if the chain is already flagged
1637 * Sets bref.modify_tid to mtid only if mtid != 0. Note that bref.modify_tid
1638 * is a CLC (cluster level change) field and is not updated by parent
1639 * propagation during a flush.
1641 * Returns an appropriate HAMMER2_ERROR_* code, which will generally reflect
1642 * chain->error except for HAMMER2_ERROR_ENOSPC. If the allocation fails
1643 * due to no space available, HAMMER2_ERROR_ENOSPC is returned and the chain
1644 * remains unmodified with its old data ref intact and chain->error
1649 * If the DEDUPABLE flag is set in the chain the storage must be reallocated
1650 * even if the chain is still flagged MODIFIED. In this case the chain's
1651 * DEDUPABLE flag will be cleared once the new storage has been assigned.
1653 * If the caller passes a non-zero dedup_off we will use it to assign the
1654 * new storage. The MODIFIED flag will be *CLEARED* in this case, and
1655 * DEDUPABLE will be set (NOTE: the UPDATE flag is always set). The caller
1656 * must not modify the data content upon return.
1659 hammer2_chain_modify(hammer2_chain_t *chain, hammer2_tid_t mtid,
1660 hammer2_off_t dedup_off, int flags)
1662 hammer2_blockref_t obref;
1673 obref = chain->bref;
1674 KKASSERT(chain->lock.mtx_lock & MTX_EXCLUSIVE);
1677 * Data is not optional for freemap chains (we must always be sure
1678 * to copy the data on COW storage allocations).
1680 if (chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE ||
1681 chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) {
1682 KKASSERT((chain->flags & HAMMER2_CHAIN_INITIAL) ||
1683 (flags & HAMMER2_MODIFY_OPTDATA) == 0);
1687 * Data must be resolved if already assigned, unless explicitly
1688 * flagged otherwise. If we cannot safety load the data the
1689 * modification fails and we return early.
1691 if (chain->data == NULL && chain->bytes != 0 &&
1692 (flags & HAMMER2_MODIFY_OPTDATA) == 0 &&
1693 (chain->bref.data_off & ~HAMMER2_OFF_MASK_RADIX)) {
1694 hammer2_chain_load_data(chain);
1696 return (chain->error);
1701 * Set MODIFIED to indicate that the chain has been modified. A new
1702 * allocation is required when modifying a chain.
1704 * Set UPDATE to ensure that the blockref is updated in the parent.
1706 * If MODIFIED is already set determine if we can reuse the assigned
1707 * data block or if we need a new data block.
1709 if ((chain->flags & HAMMER2_CHAIN_MODIFIED) == 0) {
1711 * Must set modified bit.
1713 atomic_add_long(&hammer2_count_modified_chains, 1);
1714 atomic_set_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
1715 hammer2_pfs_memory_inc(chain->pmp); /* can be NULL */
1719 * We may be able to avoid a copy-on-write if the chain's
1720 * check mode is set to NONE and the chain's current
1721 * modify_tid is beyond the last explicit snapshot tid.
1723 * This implements HAMMER2's overwrite-in-place feature.
1725 * NOTE! This data-block cannot be used as a de-duplication
1726 * source when the check mode is set to NONE.
1728 if ((chain->bref.type == HAMMER2_BREF_TYPE_DATA ||
1729 chain->bref.type == HAMMER2_BREF_TYPE_DIRENT) &&
1730 (chain->flags & HAMMER2_CHAIN_INITIAL) == 0 &&
1731 (chain->flags & HAMMER2_CHAIN_DEDUPABLE) == 0 &&
1732 HAMMER2_DEC_CHECK(chain->bref.methods) ==
1733 HAMMER2_CHECK_NONE &&
1735 chain->bref.modify_tid >
1736 chain->pmp->iroot->meta.pfs_lsnap_tid) {
1738 * Sector overwrite allowed.
1741 } else if ((hmp->hflags & HMNT2_EMERG) &&
1743 chain->bref.modify_tid >
1744 chain->pmp->iroot->meta.pfs_lsnap_tid) {
1746 * If in emergency delete mode then do a modify-in-
1747 * place on any chain type belonging to the PFS as
1748 * long as it doesn't mess up a snapshot. We might
1749 * be forced to do this anyway a little further down
1750 * in the code if the allocation fails.
1752 * Also note that in emergency mode, these modify-in-
1753 * place operations are NOT SAFE. A storage failure,
1754 * power failure, or panic can corrupt the filesystem.
1759 * Sector overwrite not allowed, must copy-on-write.
1763 } else if (chain->flags & HAMMER2_CHAIN_DEDUPABLE) {
1765 * If the modified chain was registered for dedup we need
1766 * a new allocation. This only happens for delayed-flush
1767 * chains (i.e. which run through the front-end buffer
1774 * Already flagged modified, no new allocation is needed.
1781 * Flag parent update required.
1783 if ((chain->flags & HAMMER2_CHAIN_UPDATE) == 0) {
1784 atomic_set_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
1791 * The XOP code returns held but unlocked focus chains. This
1792 * prevents the chain from being destroyed but does not prevent
1793 * it from being modified. diolk is used to interlock modifications
1794 * against XOP frontend accesses to the focus.
1796 * This allows us to theoretically avoid deadlocking the frontend
1797 * if one of the backends lock up by not formally locking the
1798 * focused chain in the frontend. In addition, the synchronization
1799 * code relies on this mechanism to avoid deadlocking concurrent
1800 * synchronization threads.
1802 lockmgr(&chain->diolk, LK_EXCLUSIVE);
1805 * The modification or re-modification requires an allocation and
1806 * possible COW. If an error occurs, the previous content and data
1807 * reference is retained and the modification fails.
1809 * If dedup_off is non-zero, the caller is requesting a deduplication
1810 * rather than a modification. The MODIFIED bit is not set and the
1811 * data offset is set to the deduplication offset. The data cannot
1814 * NOTE: The dedup offset is allowed to be in a partially free state
1815 * and we must be sure to reset it to a fully allocated state
1816 * to force two bulkfree passes to free it again.
1818 * NOTE: Only applicable when chain->bytes != 0.
1820 * XXX can a chain already be marked MODIFIED without a data
1821 * assignment? If not, assert here instead of testing the case.
1823 if (chain != &hmp->vchain && chain != &hmp->fchain &&
1825 if ((chain->bref.data_off & ~HAMMER2_OFF_MASK_RADIX) == 0 ||
1829 * NOTE: We do not have to remove the dedup
1830 * registration because the area is still
1831 * allocated and the underlying DIO will
1835 chain->bref.data_off = dedup_off;
1836 chain->bytes = 1 << (dedup_off &
1837 HAMMER2_OFF_MASK_RADIX);
1839 atomic_clear_int(&chain->flags,
1840 HAMMER2_CHAIN_MODIFIED);
1841 atomic_add_long(&hammer2_count_modified_chains,
1844 hammer2_pfs_memory_wakeup(
1847 hammer2_freemap_adjust(hmp, &chain->bref,
1848 HAMMER2_FREEMAP_DORECOVER);
1849 atomic_set_int(&chain->flags,
1850 HAMMER2_CHAIN_DEDUPABLE);
1852 error = hammer2_freemap_alloc(chain,
1854 atomic_clear_int(&chain->flags,
1855 HAMMER2_CHAIN_DEDUPABLE);
1858 * If we are unable to allocate a new block
1859 * but we are in emergency mode, issue a
1860 * warning to the console and reuse the same
1863 * We behave as if the allocation were
1866 * THIS IS IMPORTANT: These modifications
1867 * are virtually guaranteed to corrupt any
1868 * snapshots related to this filesystem.
1870 if (error && (hmp->hflags & HMNT2_EMERG)) {
1872 chain->bref.flags |=
1873 HAMMER2_BREF_FLAG_EMERG_MIP;
1875 krateprintf(&krate_h2em,
1876 "hammer2: Emergency Mode WARNING: "
1877 "Operation will likely corrupt "
1878 "related snapshot: "
1879 "%016jx.%02x key=%016jx\n",
1880 chain->bref.data_off,
1883 } else if (error == 0) {
1884 chain->bref.flags &=
1885 ~HAMMER2_BREF_FLAG_EMERG_MIP;
1892 * Stop here if error. We have to undo any flag bits we might
1897 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
1898 atomic_add_long(&hammer2_count_modified_chains, -1);
1900 hammer2_pfs_memory_wakeup(chain->pmp, -1);
1903 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
1905 lockmgr(&chain->diolk, LK_RELEASE);
1911 * Update mirror_tid and modify_tid. modify_tid is only updated
1912 * if not passed as zero (during flushes, parent propagation passes
1915 * NOTE: chain->pmp could be the device spmp.
1917 chain->bref.mirror_tid = hmp->voldata.mirror_tid + 1;
1919 chain->bref.modify_tid = mtid;
1922 * Set BMAPUPD to tell the flush code that an existing blockmap entry
1923 * requires updating as well as to tell the delete code that the
1924 * chain's blockref might not exactly match (in terms of physical size
1925 * or block offset) the one in the parent's blocktable. The base key
1926 * of course will still match.
1928 if (chain->flags & HAMMER2_CHAIN_BMAPPED)
1929 atomic_set_int(&chain->flags, HAMMER2_CHAIN_BMAPUPD);
1932 * Short-cut data block handling when the caller does not need an
1933 * actual data reference to (aka OPTDATA), as long as the chain does
1934 * not already have a data pointer to the data and no de-duplication
1937 * This generally means that the modifications are being done via the
1938 * logical buffer cache.
1940 * NOTE: If deduplication occurred we have to run through the data
1941 * stuff to clear INITIAL, and the caller will likely want to
1942 * assign the check code anyway. Leaving INITIAL set on a
1943 * dedup can be deadly (it can cause the block to be zero'd!).
1945 * This code also handles bytes == 0 (most dirents).
1947 if (chain->bref.type == HAMMER2_BREF_TYPE_DATA &&
1948 (flags & HAMMER2_MODIFY_OPTDATA) &&
1949 chain->data == NULL) {
1950 if (dedup_off == 0) {
1951 KKASSERT(chain->dio == NULL);
1957 * Clearing the INITIAL flag (for indirect blocks) indicates that
1958 * we've processed the uninitialized storage allocation.
1960 * If this flag is already clear we are likely in a copy-on-write
1961 * situation but we have to be sure NOT to bzero the storage if
1962 * no data is present.
1964 * Clearing of NOTTESTED is allowed if the MODIFIED bit is set,
1966 if (chain->flags & HAMMER2_CHAIN_INITIAL) {
1967 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1974 * Instantiate data buffer and possibly execute COW operation
1976 switch(chain->bref.type) {
1977 case HAMMER2_BREF_TYPE_VOLUME:
1978 case HAMMER2_BREF_TYPE_FREEMAP:
1980 * The data is embedded, no copy-on-write operation is
1983 KKASSERT(chain->dio == NULL);
1985 case HAMMER2_BREF_TYPE_DIRENT:
1987 * The data might be fully embedded.
1989 if (chain->bytes == 0) {
1990 KKASSERT(chain->dio == NULL);
1994 case HAMMER2_BREF_TYPE_INODE:
1995 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
1996 case HAMMER2_BREF_TYPE_DATA:
1997 case HAMMER2_BREF_TYPE_INDIRECT:
1998 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2000 * Perform the copy-on-write operation
2002 * zero-fill or copy-on-write depending on whether
2003 * chain->data exists or not and set the dirty state for
2004 * the new buffer. hammer2_io_new() will handle the
2007 * If a dedup_off was supplied this is an existing block
2008 * and no COW, copy, or further modification is required.
2010 KKASSERT(chain != &hmp->vchain && chain != &hmp->fchain);
2012 if (wasinitial && dedup_off == 0) {
2013 error = hammer2_io_new(hmp, chain->bref.type,
2014 chain->bref.data_off,
2015 chain->bytes, &dio);
2017 error = hammer2_io_bread(hmp, chain->bref.type,
2018 chain->bref.data_off,
2019 chain->bytes, &dio);
2021 hammer2_adjreadcounter(chain->bref.type, chain->bytes);
2024 * If an I/O error occurs make sure callers cannot accidently
2025 * modify the old buffer's contents and corrupt the filesystem.
2027 * NOTE: hammer2_io_data() call issues bkvasync()
2030 kprintf("hammer2_chain_modify: hmp=%p I/O error\n",
2032 chain->error = HAMMER2_ERROR_EIO;
2033 hammer2_io_brelse(&dio);
2034 hammer2_io_brelse(&chain->dio);
2039 bdata = hammer2_io_data(dio, chain->bref.data_off);
2043 * COW (unless a dedup).
2045 KKASSERT(chain->dio != NULL);
2046 if (chain->data != (void *)bdata && dedup_off == 0) {
2047 bcopy(chain->data, bdata, chain->bytes);
2049 } else if (wasinitial == 0 && dedup_off == 0) {
2051 * We have a problem. We were asked to COW but
2052 * we don't have any data to COW with!
2054 panic("hammer2_chain_modify: having a COW %p\n",
2059 * Retire the old buffer, replace with the new. Dirty or
2060 * redirty the new buffer.
2062 * WARNING! The system buffer cache may have already flushed
2063 * the buffer, so we must be sure to [re]dirty it
2064 * for further modification.
2066 * If dedup_off was supplied, the caller is not
2067 * expected to make any further modification to the
2070 * WARNING! hammer2_get_gdata() assumes dio never transitions
2071 * through NULL in order to optimize away unnecessary
2077 if ((tio = chain->dio) != NULL)
2078 hammer2_io_bqrelse(&tio);
2079 chain->data = (void *)bdata;
2082 hammer2_io_setdirty(dio);
2086 panic("hammer2_chain_modify: illegal non-embedded type %d",
2093 * setflush on parent indicating that the parent must recurse down
2094 * to us. Do not call on chain itself which might already have it
2098 hammer2_chain_setflush(chain->parent);
2099 lockmgr(&chain->diolk, LK_RELEASE);
2101 return (chain->error);
2105 * Modify the chain associated with an inode.
2108 hammer2_chain_modify_ip(hammer2_inode_t *ip, hammer2_chain_t *chain,
2109 hammer2_tid_t mtid, int flags)
2113 hammer2_inode_modify(ip);
2114 error = hammer2_chain_modify(chain, mtid, 0, flags);
2120 * Volume header data locks
2123 hammer2_voldata_lock(hammer2_dev_t *hmp)
2125 lockmgr(&hmp->vollk, LK_EXCLUSIVE);
2129 hammer2_voldata_unlock(hammer2_dev_t *hmp)
2131 lockmgr(&hmp->vollk, LK_RELEASE);
2135 hammer2_voldata_modify(hammer2_dev_t *hmp)
2137 if ((hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) == 0) {
2138 atomic_add_long(&hammer2_count_modified_chains, 1);
2139 atomic_set_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
2140 hammer2_pfs_memory_inc(hmp->vchain.pmp);
2145 * This function returns the chain at the nearest key within the specified
2146 * range. The returned chain will be referenced but not locked.
2148 * This function will recurse through chain->rbtree as necessary and will
2149 * return a *key_nextp suitable for iteration. *key_nextp is only set if
2150 * the iteration value is less than the current value of *key_nextp.
2152 * The caller should use (*key_nextp) to calculate the actual range of
2153 * the returned element, which will be (key_beg to *key_nextp - 1), because
2154 * there might be another element which is superior to the returned element
2157 * (*key_nextp) can be passed as key_beg in an iteration only while non-NULL
2158 * chains continue to be returned. On EOF (*key_nextp) may overflow since
2159 * it will wind up being (key_end + 1).
2161 * WARNING! Must be called with child's spinlock held. Spinlock remains
2162 * held through the operation.
2164 struct hammer2_chain_find_info {
2165 hammer2_chain_t *best;
2166 hammer2_key_t key_beg;
2167 hammer2_key_t key_end;
2168 hammer2_key_t key_next;
2171 static int hammer2_chain_find_cmp(hammer2_chain_t *child, void *data);
2172 static int hammer2_chain_find_callback(hammer2_chain_t *child, void *data);
2176 hammer2_chain_find(hammer2_chain_t *parent, hammer2_key_t *key_nextp,
2177 hammer2_key_t key_beg, hammer2_key_t key_end)
2179 struct hammer2_chain_find_info info;
2182 info.key_beg = key_beg;
2183 info.key_end = key_end;
2184 info.key_next = *key_nextp;
2186 RB_SCAN(hammer2_chain_tree, &parent->core.rbtree,
2187 hammer2_chain_find_cmp, hammer2_chain_find_callback,
2189 *key_nextp = info.key_next;
2191 kprintf("chain_find %p %016jx:%016jx next=%016jx\n",
2192 parent, key_beg, key_end, *key_nextp);
2200 hammer2_chain_find_cmp(hammer2_chain_t *child, void *data)
2202 struct hammer2_chain_find_info *info = data;
2203 hammer2_key_t child_beg;
2204 hammer2_key_t child_end;
2206 child_beg = child->bref.key;
2207 child_end = child_beg + ((hammer2_key_t)1 << child->bref.keybits) - 1;
2209 if (child_end < info->key_beg)
2211 if (child_beg > info->key_end)
2218 hammer2_chain_find_callback(hammer2_chain_t *child, void *data)
2220 struct hammer2_chain_find_info *info = data;
2221 hammer2_chain_t *best;
2222 hammer2_key_t child_end;
2225 * WARNING! Layerq is scanned forwards, exact matches should keep
2226 * the existing info->best.
2228 if ((best = info->best) == NULL) {
2230 * No previous best. Assign best
2233 } else if (best->bref.key <= info->key_beg &&
2234 child->bref.key <= info->key_beg) {
2239 /*info->best = child;*/
2240 } else if (child->bref.key < best->bref.key) {
2242 * Child has a nearer key and best is not flush with key_beg.
2243 * Set best to child. Truncate key_next to the old best key.
2246 if (info->key_next > best->bref.key || info->key_next == 0)
2247 info->key_next = best->bref.key;
2248 } else if (child->bref.key == best->bref.key) {
2250 * If our current best is flush with the child then this
2251 * is an illegal overlap.
2253 * key_next will automatically be limited to the smaller of
2254 * the two end-points.
2260 * Keep the current best but truncate key_next to the child's
2263 * key_next will also automatically be limited to the smaller
2264 * of the two end-points (probably not necessary for this case
2265 * but we do it anyway).
2267 if (info->key_next > child->bref.key || info->key_next == 0)
2268 info->key_next = child->bref.key;
2272 * Always truncate key_next based on child's end-of-range.
2274 child_end = child->bref.key + ((hammer2_key_t)1 << child->bref.keybits);
2275 if (child_end && (info->key_next > child_end || info->key_next == 0))
2276 info->key_next = child_end;
2282 * Retrieve the specified chain from a media blockref, creating the
2283 * in-memory chain structure which reflects it. The returned chain is
2284 * held and locked according to (how) (HAMMER2_RESOLVE_*). The caller must
2285 * handle crc-checks and so forth, and should check chain->error before
2286 * assuming that the data is good.
2288 * To handle insertion races pass the INSERT_RACE flag along with the
2289 * generation number of the core. NULL will be returned if the generation
2290 * number changes before we have a chance to insert the chain. Insert
2291 * races can occur because the parent might be held shared.
2293 * Caller must hold the parent locked shared or exclusive since we may
2294 * need the parent's bref array to find our block.
2296 * WARNING! chain->pmp is always set to NULL for any chain representing
2297 * part of the super-root topology.
2300 hammer2_chain_get(hammer2_chain_t *parent, int generation,
2301 hammer2_blockref_t *bref, int how)
2303 hammer2_dev_t *hmp = parent->hmp;
2304 hammer2_chain_t *chain;
2308 * Allocate a chain structure representing the existing media
2309 * entry. Resulting chain has one ref and is not locked.
2311 if (bref->flags & HAMMER2_BREF_FLAG_PFSROOT)
2312 chain = hammer2_chain_alloc(hmp, NULL, bref);
2314 chain = hammer2_chain_alloc(hmp, parent->pmp, bref);
2315 /* ref'd chain returned */
2318 * Flag that the chain is in the parent's blockmap so delete/flush
2319 * knows what to do with it.
2321 atomic_set_int(&chain->flags, HAMMER2_CHAIN_BMAPPED);
2324 * chain must be locked to avoid unexpected ripouts
2326 hammer2_chain_lock(chain, how);
2329 * Link the chain into its parent. A spinlock is required to safely
2330 * access the RBTREE, and it is possible to collide with another
2331 * hammer2_chain_get() operation because the caller might only hold
2332 * a shared lock on the parent.
2334 * NOTE: Get races can occur quite often when we distribute
2335 * asynchronous read-aheads across multiple threads.
2337 KKASSERT(parent->refs > 0);
2338 error = hammer2_chain_insert(parent, chain,
2339 HAMMER2_CHAIN_INSERT_SPIN |
2340 HAMMER2_CHAIN_INSERT_RACE,
2343 KKASSERT((chain->flags & HAMMER2_CHAIN_ONRBTREE) == 0);
2344 /*kprintf("chain %p get race\n", chain);*/
2345 hammer2_chain_unlock(chain);
2346 hammer2_chain_drop(chain);
2349 KKASSERT(chain->flags & HAMMER2_CHAIN_ONRBTREE);
2353 * Return our new chain referenced but not locked, or NULL if
2360 * Lookup initialization/completion API
2363 hammer2_chain_lookup_init(hammer2_chain_t *parent, int flags)
2365 hammer2_chain_ref(parent);
2366 if (flags & HAMMER2_LOOKUP_SHARED) {
2367 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS |
2368 HAMMER2_RESOLVE_SHARED);
2370 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2376 hammer2_chain_lookup_done(hammer2_chain_t *parent)
2379 hammer2_chain_unlock(parent);
2380 hammer2_chain_drop(parent);
2385 * Take the locked chain and return a locked parent. The chain remains
2386 * locked on return, but may have to be temporarily unlocked to acquire
2387 * the parent. Because of this, (chain) must be stable and cannot be
2388 * deleted while it was temporarily unlocked (typically means that (chain)
2391 * Pass HAMMER2_RESOLVE_* flags in flags.
2393 * This will work even if the chain is errored, and the caller can check
2394 * parent->error on return if desired since the parent will be locked.
2396 * This function handles the lock order reversal.
2399 hammer2_chain_getparent(hammer2_chain_t *chain, int flags)
2401 hammer2_chain_t *parent;
2404 * Be careful of order, chain must be unlocked before parent
2405 * is locked below to avoid a deadlock. Try it trivially first.
2407 parent = chain->parent;
2409 panic("hammer2_chain_getparent: no parent");
2410 hammer2_chain_ref(parent);
2411 if (hammer2_chain_lock(parent, flags|HAMMER2_RESOLVE_NONBLOCK) == 0)
2415 hammer2_chain_unlock(chain);
2416 hammer2_chain_lock(parent, flags);
2417 hammer2_chain_lock(chain, flags);
2420 * Parent relinking races are quite common. We have to get
2421 * it right or we will blow up the block table.
2423 if (chain->parent == parent)
2425 hammer2_chain_unlock(parent);
2426 hammer2_chain_drop(parent);
2428 parent = chain->parent;
2430 panic("hammer2_chain_getparent: no parent");
2431 hammer2_chain_ref(parent);
2437 * Take the locked chain and return a locked parent. The chain is unlocked
2438 * and dropped. *chainp is set to the returned parent as a convenience.
2439 * Pass HAMMER2_RESOLVE_* flags in flags.
2441 * This will work even if the chain is errored, and the caller can check
2442 * parent->error on return if desired since the parent will be locked.
2444 * The chain does NOT need to be stable. We use a tracking structure
2445 * to track the expected parent if the chain is deleted out from under us.
2447 * This function handles the lock order reversal.
2450 hammer2_chain_repparent(hammer2_chain_t **chainp, int flags)
2452 hammer2_chain_t *chain;
2453 hammer2_chain_t *parent;
2454 struct hammer2_reptrack reptrack;
2455 struct hammer2_reptrack **repp;
2458 * Be careful of order, chain must be unlocked before parent
2459 * is locked below to avoid a deadlock. Try it trivially first.
2462 parent = chain->parent;
2463 if (parent == NULL) {
2464 hammer2_spin_unex(&chain->core.spin);
2465 panic("hammer2_chain_repparent: no parent");
2467 hammer2_chain_ref(parent);
2468 if (hammer2_chain_lock(parent, flags|HAMMER2_RESOLVE_NONBLOCK) == 0) {
2469 hammer2_chain_unlock(chain);
2470 hammer2_chain_drop(chain);
2477 * Ok, now it gets a bit nasty. There are multiple situations where
2478 * the parent might be in the middle of a deletion, or where the child
2479 * (chain) might be deleted the instant we let go of its lock.
2480 * We can potentially end up in a no-win situation!
2482 * In particular, the indirect_maintenance() case can cause these
2485 * To deal with this we install a reptrack structure in the parent
2486 * This reptrack structure 'owns' the parent ref and will automatically
2487 * migrate to the parent's parent if the parent is deleted permanently.
2489 hammer2_spin_init(&reptrack.spin, "h2reptrk");
2490 reptrack.chain = parent;
2491 hammer2_chain_ref(parent); /* for the reptrack */
2493 hammer2_spin_ex(&parent->core.spin);
2494 reptrack.next = parent->core.reptrack;
2495 parent->core.reptrack = &reptrack;
2496 hammer2_spin_unex(&parent->core.spin);
2498 hammer2_chain_unlock(chain);
2499 hammer2_chain_drop(chain);
2500 chain = NULL; /* gone */
2503 * At the top of this loop, chain is gone and parent is refd both
2504 * by us explicitly AND via our reptrack. We are attempting to
2508 hammer2_chain_lock(parent, flags);
2510 if (reptrack.chain == parent)
2512 hammer2_chain_unlock(parent);
2513 hammer2_chain_drop(parent);
2515 kprintf("hammer2: debug REPTRACK %p->%p\n",
2516 parent, reptrack.chain);
2517 hammer2_spin_ex(&reptrack.spin);
2518 parent = reptrack.chain;
2519 hammer2_chain_ref(parent);
2520 hammer2_spin_unex(&reptrack.spin);
2524 * Once parent is locked and matches our reptrack, our reptrack
2525 * will be stable and we have our parent. We can unlink our
2528 * WARNING! Remember that the chain lock might be shared. Chains
2529 * locked shared have stable parent linkages.
2531 hammer2_spin_ex(&parent->core.spin);
2532 repp = &parent->core.reptrack;
2533 while (*repp != &reptrack)
2534 repp = &(*repp)->next;
2535 *repp = reptrack.next;
2536 hammer2_spin_unex(&parent->core.spin);
2538 hammer2_chain_drop(parent); /* reptrack ref */
2539 *chainp = parent; /* return parent lock+ref */
2545 * Dispose of any linked reptrack structures in (chain) by shifting them to
2546 * (parent). Both (chain) and (parent) must be exclusively locked.
2548 * This is interlocked against any children of (chain) on the other side.
2549 * No children so remain as-of when this is called so we can test
2550 * core.reptrack without holding the spin-lock.
2552 * Used whenever the caller intends to permanently delete chains related
2553 * to topological recursions (BREF_TYPE_INDIRECT, BREF_TYPE_FREEMAP_NODE),
2554 * where the chains underneath the node being deleted are given a new parent
2555 * above the node being deleted.
2559 hammer2_chain_repchange(hammer2_chain_t *parent, hammer2_chain_t *chain)
2561 struct hammer2_reptrack *reptrack;
2563 KKASSERT(chain->core.live_count == 0 && RB_EMPTY(&chain->core.rbtree));
2564 while (chain->core.reptrack) {
2565 hammer2_spin_ex(&parent->core.spin);
2566 hammer2_spin_ex(&chain->core.spin);
2567 reptrack = chain->core.reptrack;
2568 if (reptrack == NULL) {
2569 hammer2_spin_unex(&chain->core.spin);
2570 hammer2_spin_unex(&parent->core.spin);
2573 hammer2_spin_ex(&reptrack->spin);
2574 chain->core.reptrack = reptrack->next;
2575 reptrack->chain = parent;
2576 reptrack->next = parent->core.reptrack;
2577 parent->core.reptrack = reptrack;
2578 hammer2_chain_ref(parent); /* reptrack */
2580 hammer2_spin_unex(&chain->core.spin);
2581 hammer2_spin_unex(&parent->core.spin);
2582 kprintf("hammer2: debug repchange %p %p->%p\n",
2583 reptrack, chain, parent);
2584 hammer2_chain_drop(chain); /* reptrack */
2589 * Locate the first chain whos key range overlaps (key_beg, key_end) inclusive.
2590 * (*parentp) typically points to an inode but can also point to a related
2591 * indirect block and this function will recurse upwards and find the inode
2592 * or the nearest undeleted indirect block covering the key range.
2594 * This function unconditionally sets *errorp, replacing any previous value.
2596 * (*parentp) must be exclusive or shared locked (depending on flags) and
2597 * referenced and can be an inode or an existing indirect block within the
2600 * If (*parent) is errored out, this function will not attempt to recurse
2601 * the radix tree and will return NULL along with an appropriate *errorp.
2602 * If NULL is returned and *errorp is 0, the requested lookup could not be
2605 * On return (*parentp) will be modified to point at the deepest parent chain
2606 * element encountered during the search, as a helper for an insertion or
2609 * The new (*parentp) will be locked shared or exclusive (depending on flags),
2610 * and referenced, and the old will be unlocked and dereferenced (no change
2611 * if they are both the same). This is particularly important if the caller
2612 * wishes to insert a new chain, (*parentp) will be set properly even if NULL
2613 * is returned, as long as no error occurred.
2615 * The matching chain will be returned locked according to flags.
2619 * NULL is returned if no match was found, but (*parentp) will still
2620 * potentially be adjusted.
2622 * On return (*key_nextp) will point to an iterative value for key_beg.
2623 * (If NULL is returned (*key_nextp) is set to (key_end + 1)).
2625 * This function will also recurse up the chain if the key is not within the
2626 * current parent's range. (*parentp) can never be set to NULL. An iteration
2627 * can simply allow (*parentp) to float inside the loop.
2629 * NOTE! chain->data is not always resolved. By default it will not be
2630 * resolved for BREF_TYPE_DATA, FREEMAP_NODE, or FREEMAP_LEAF. Use
2631 * HAMMER2_LOOKUP_ALWAYS to force resolution (but be careful w/
2632 * BREF_TYPE_DATA as the device buffer can alias the logical file
2637 hammer2_chain_lookup(hammer2_chain_t **parentp, hammer2_key_t *key_nextp,
2638 hammer2_key_t key_beg, hammer2_key_t key_end,
2639 int *errorp, int flags)
2642 hammer2_chain_t *parent;
2643 hammer2_chain_t *chain;
2644 hammer2_blockref_t *base;
2645 hammer2_blockref_t *bref;
2646 hammer2_blockref_t bsave;
2647 hammer2_key_t scan_beg;
2648 hammer2_key_t scan_end;
2650 int how_always = HAMMER2_RESOLVE_ALWAYS;
2651 int how_maybe = HAMMER2_RESOLVE_MAYBE;
2654 int maxloops = 300000;
2655 volatile hammer2_mtx_t save_mtx;
2657 if (flags & HAMMER2_LOOKUP_ALWAYS) {
2658 how_maybe = how_always;
2659 how = HAMMER2_RESOLVE_ALWAYS;
2660 } else if (flags & HAMMER2_LOOKUP_NODATA) {
2661 how = HAMMER2_RESOLVE_NEVER;
2663 how = HAMMER2_RESOLVE_MAYBE;
2665 if (flags & HAMMER2_LOOKUP_SHARED) {
2666 how_maybe |= HAMMER2_RESOLVE_SHARED;
2667 how_always |= HAMMER2_RESOLVE_SHARED;
2668 how |= HAMMER2_RESOLVE_SHARED;
2672 * Recurse (*parentp) upward if necessary until the parent completely
2673 * encloses the key range or we hit the inode.
2675 * Handle races against the flusher deleting indirect nodes on its
2676 * way back up by continuing to recurse upward past the deletion.
2682 while (parent->bref.type == HAMMER2_BREF_TYPE_INDIRECT ||
2683 parent->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2684 scan_beg = parent->bref.key;
2685 scan_end = scan_beg +
2686 ((hammer2_key_t)1 << parent->bref.keybits) - 1;
2687 if ((parent->flags & HAMMER2_CHAIN_DELETED) == 0) {
2688 if (key_beg >= scan_beg && key_end <= scan_end)
2691 parent = hammer2_chain_repparent(parentp, how_maybe);
2694 if (--maxloops == 0)
2695 panic("hammer2_chain_lookup: maxloops");
2698 * MATCHIND case that does not require parent->data (do prior to
2699 * parent->error check).
2701 switch(parent->bref.type) {
2702 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2703 case HAMMER2_BREF_TYPE_INDIRECT:
2704 if (flags & HAMMER2_LOOKUP_MATCHIND) {
2705 scan_beg = parent->bref.key;
2706 scan_end = scan_beg +
2707 ((hammer2_key_t)1 << parent->bref.keybits) - 1;
2708 if (key_beg == scan_beg && key_end == scan_end) {
2710 hammer2_chain_ref(chain);
2711 hammer2_chain_lock(chain, how_maybe);
2712 *key_nextp = scan_end + 1;
2722 * No lookup is possible if the parent is errored. We delayed
2723 * this check as long as we could to ensure that the parent backup,
2724 * embedded data, and MATCHIND code could still execute.
2726 if (parent->error) {
2727 *errorp = parent->error;
2732 * Locate the blockref array. Currently we do a fully associative
2733 * search through the array.
2735 switch(parent->bref.type) {
2736 case HAMMER2_BREF_TYPE_INODE:
2738 * Special shortcut for embedded data returns the inode
2739 * itself. Callers must detect this condition and access
2740 * the embedded data (the strategy code does this for us).
2742 * This is only applicable to regular files and softlinks.
2744 * We need a second lock on parent. Since we already have
2745 * a lock we must pass LOCKAGAIN to prevent unexpected
2746 * blocking (we don't want to block on a second shared
2747 * ref if an exclusive lock is pending)
2749 if (parent->data->ipdata.meta.op_flags &
2750 HAMMER2_OPFLAG_DIRECTDATA) {
2751 if (flags & HAMMER2_LOOKUP_NODIRECT) {
2753 *key_nextp = key_end + 1;
2756 hammer2_chain_ref(parent);
2757 hammer2_chain_lock(parent, how_always |
2758 HAMMER2_RESOLVE_LOCKAGAIN);
2759 *key_nextp = key_end + 1;
2762 base = &parent->data->ipdata.u.blockset.blockref[0];
2763 count = HAMMER2_SET_COUNT;
2765 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2766 case HAMMER2_BREF_TYPE_INDIRECT:
2768 * Optimize indirect blocks in the INITIAL state to avoid
2771 * Debugging: Enter permanent wait state instead of
2772 * panicing on unexpectedly NULL data for the moment.
2774 if (parent->flags & HAMMER2_CHAIN_INITIAL) {
2777 if (parent->data == NULL) {
2778 kprintf("hammer2: unexpected NULL data "
2781 tsleep(parent, 0, "xxx", 0);
2783 base = &parent->data->npdata[0];
2785 count = parent->bytes / sizeof(hammer2_blockref_t);
2787 case HAMMER2_BREF_TYPE_VOLUME:
2788 base = &parent->data->voldata.sroot_blockset.blockref[0];
2789 count = HAMMER2_SET_COUNT;
2791 case HAMMER2_BREF_TYPE_FREEMAP:
2792 base = &parent->data->blkset.blockref[0];
2793 count = HAMMER2_SET_COUNT;
2796 panic("hammer2_chain_lookup: unrecognized "
2797 "blockref(B) type: %d",
2799 base = NULL; /* safety */
2800 count = 0; /* safety */
2805 * Merged scan to find next candidate.
2807 * hammer2_base_*() functions require the parent->core.live_* fields
2808 * to be synchronized.
2810 * We need to hold the spinlock to access the block array and RB tree
2811 * and to interlock chain creation.
2813 if ((parent->flags & HAMMER2_CHAIN_COUNTEDBREFS) == 0)
2814 hammer2_chain_countbrefs(parent, base, count);
2819 hammer2_spin_ex(&parent->core.spin);
2820 chain = hammer2_combined_find(parent, base, count,
2824 generation = parent->core.generation;
2827 * Exhausted parent chain, iterate.
2830 KKASSERT(chain == NULL);
2831 hammer2_spin_unex(&parent->core.spin);
2832 if (key_beg == key_end) /* short cut single-key case */
2836 * Stop if we reached the end of the iteration.
2838 if (parent->bref.type != HAMMER2_BREF_TYPE_INDIRECT &&
2839 parent->bref.type != HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2844 * Calculate next key, stop if we reached the end of the
2845 * iteration, otherwise go up one level and loop.
2847 key_beg = parent->bref.key +
2848 ((hammer2_key_t)1 << parent->bref.keybits);
2849 if (key_beg == 0 || key_beg > key_end)
2851 parent = hammer2_chain_repparent(parentp, how_maybe);
2856 * Selected from blockref or in-memory chain.
2859 if (chain == NULL) {
2860 hammer2_spin_unex(&parent->core.spin);
2861 if (bsave.type == HAMMER2_BREF_TYPE_INDIRECT ||
2862 bsave.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2863 chain = hammer2_chain_get(parent, generation,
2866 chain = hammer2_chain_get(parent, generation,
2872 hammer2_chain_ref(chain);
2873 hammer2_spin_unex(&parent->core.spin);
2876 * chain is referenced but not locked. We must lock the
2877 * chain to obtain definitive state.
2879 if (bsave.type == HAMMER2_BREF_TYPE_INDIRECT ||
2880 bsave.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2881 hammer2_chain_lock(chain, how_maybe);
2883 hammer2_chain_lock(chain, how);
2885 KKASSERT(chain->parent == parent);
2887 if (bcmp(&bsave, &chain->bref, sizeof(bsave)) ||
2888 chain->parent != parent) {
2889 hammer2_chain_unlock(chain);
2890 hammer2_chain_drop(chain);
2891 chain = NULL; /* SAFETY */
2897 * Skip deleted chains (XXX cache 'i' end-of-block-array? XXX)
2899 * NOTE: Chain's key range is not relevant as there might be
2900 * one-offs within the range that are not deleted.
2902 * NOTE: Lookups can race delete-duplicate because
2903 * delete-duplicate does not lock the parent's core
2904 * (they just use the spinlock on the core).
2906 if (chain->flags & HAMMER2_CHAIN_DELETED) {
2907 kprintf("skip deleted chain %016jx.%02x key=%016jx\n",
2908 chain->bref.data_off, chain->bref.type,
2910 hammer2_chain_unlock(chain);
2911 hammer2_chain_drop(chain);
2912 chain = NULL; /* SAFETY */
2913 key_beg = *key_nextp;
2914 if (key_beg == 0 || key_beg > key_end)
2920 * If the chain element is an indirect block it becomes the new
2921 * parent and we loop on it. We must maintain our top-down locks
2922 * to prevent the flusher from interfering (i.e. doing a
2923 * delete-duplicate and leaving us recursing down a deleted chain).
2925 * The parent always has to be locked with at least RESOLVE_MAYBE
2926 * so we can access its data. It might need a fixup if the caller
2927 * passed incompatible flags. Be careful not to cause a deadlock
2928 * as a data-load requires an exclusive lock.
2930 * If HAMMER2_LOOKUP_MATCHIND is set and the indirect block's key
2931 * range is within the requested key range we return the indirect
2932 * block and do NOT loop. This is usually only used to acquire
2935 if (chain->bref.type == HAMMER2_BREF_TYPE_INDIRECT ||
2936 chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2937 save_mtx = parent->lock;
2938 hammer2_chain_unlock(parent);
2939 hammer2_chain_drop(parent);
2940 *parentp = parent = chain;
2941 chain = NULL; /* SAFETY */
2946 * All done, return the locked chain.
2948 * If the caller does not want a locked chain, replace the lock with
2949 * a ref. Perhaps this can eventually be optimized to not obtain the
2950 * lock in the first place for situations where the data does not
2951 * need to be resolved.
2953 * NOTE! A chain->error must be tested by the caller upon return.
2954 * *errorp is only set based on issues which occur while
2955 * trying to reach the chain.
2961 * After having issued a lookup we can iterate all matching keys.
2963 * If chain is non-NULL we continue the iteration from just after it's index.
2965 * If chain is NULL we assume the parent was exhausted and continue the
2966 * iteration at the next parent.
2968 * If a fatal error occurs (typically an I/O error), a dummy chain is
2969 * returned with chain->error and error-identifying information set. This
2970 * chain will assert if you try to do anything fancy with it.
2972 * XXX Depending on where the error occurs we should allow continued iteration.
2974 * parent must be locked on entry and remains locked throughout. chain's
2975 * lock status must match flags. Chain is always at least referenced.
2977 * WARNING! The MATCHIND flag does not apply to this function.
2980 hammer2_chain_next(hammer2_chain_t **parentp, hammer2_chain_t *chain,
2981 hammer2_key_t *key_nextp,
2982 hammer2_key_t key_beg, hammer2_key_t key_end,
2983 int *errorp, int flags)
2985 hammer2_chain_t *parent;
2989 * Calculate locking flags for upward recursion.
2991 how_maybe = HAMMER2_RESOLVE_MAYBE;
2992 if (flags & HAMMER2_LOOKUP_SHARED)
2993 how_maybe |= HAMMER2_RESOLVE_SHARED;
2999 * Calculate the next index and recalculate the parent if necessary.
3002 key_beg = chain->bref.key +
3003 ((hammer2_key_t)1 << chain->bref.keybits);
3004 hammer2_chain_unlock(chain);
3005 hammer2_chain_drop(chain);
3008 * chain invalid past this point, but we can still do a
3009 * pointer comparison w/parent.
3011 * Any scan where the lookup returned degenerate data embedded
3012 * in the inode has an invalid index and must terminate.
3014 if (chain == parent)
3016 if (key_beg == 0 || key_beg > key_end)
3019 } else if (parent->bref.type != HAMMER2_BREF_TYPE_INDIRECT &&
3020 parent->bref.type != HAMMER2_BREF_TYPE_FREEMAP_NODE) {
3022 * We reached the end of the iteration.
3027 * Continue iteration with next parent unless the current
3028 * parent covers the range.
3030 * (This also handles the case of a deleted, empty indirect
3033 key_beg = parent->bref.key +
3034 ((hammer2_key_t)1 << parent->bref.keybits);
3035 if (key_beg == 0 || key_beg > key_end)
3037 parent = hammer2_chain_repparent(parentp, how_maybe);
3043 return (hammer2_chain_lookup(parentp, key_nextp,
3049 * Caller wishes to iterate chains under parent, loading new chains into
3050 * chainp. Caller must initialize *chainp to NULL and *firstp to 1, and
3051 * then call hammer2_chain_scan() repeatedly until a non-zero return.
3052 * During the scan, *firstp will be set to 0 and (*chainp) will be replaced
3053 * with the returned chain for the scan. The returned *chainp will be
3054 * locked and referenced. Any prior contents will be unlocked and dropped.
3056 * Caller should check the return value. A normal scan EOF will return
3057 * exactly HAMMER2_ERROR_EOF. Any other non-zero value indicates an
3058 * error trying to access parent data. Any error in the returned chain
3059 * must be tested separately by the caller.
3061 * (*chainp) is dropped on each scan, but will only be set if the returned
3062 * element itself can recurse. Leaf elements are NOT resolved, loaded, or
3063 * returned via *chainp. The caller will get their bref only.
3065 * The raw scan function is similar to lookup/next but does not seek to a key.
3066 * Blockrefs are iterated via first_bref = (parent, NULL) and
3067 * next_chain = (parent, bref).
3069 * The passed-in parent must be locked and its data resolved. The function
3070 * nominally returns a locked and referenced *chainp != NULL for chains
3071 * the caller might need to recurse on (and will dipose of any *chainp passed
3072 * in). The caller must check the chain->bref.type either way.
3075 hammer2_chain_scan(hammer2_chain_t *parent, hammer2_chain_t **chainp,
3076 hammer2_blockref_t *bref, int *firstp,
3080 hammer2_blockref_t *base;
3081 hammer2_blockref_t *bref_ptr;
3083 hammer2_key_t next_key;
3084 hammer2_chain_t *chain = NULL;
3086 int how_always = HAMMER2_RESOLVE_ALWAYS;
3087 int how_maybe = HAMMER2_RESOLVE_MAYBE;
3090 int maxloops = 300000;
3097 * Scan flags borrowed from lookup.
3099 if (flags & HAMMER2_LOOKUP_ALWAYS) {
3100 how_maybe = how_always;
3101 how = HAMMER2_RESOLVE_ALWAYS;
3102 } else if (flags & HAMMER2_LOOKUP_NODATA) {
3103 how = HAMMER2_RESOLVE_NEVER;
3105 how = HAMMER2_RESOLVE_MAYBE;
3107 if (flags & HAMMER2_LOOKUP_SHARED) {
3108 how_maybe |= HAMMER2_RESOLVE_SHARED;
3109 how_always |= HAMMER2_RESOLVE_SHARED;
3110 how |= HAMMER2_RESOLVE_SHARED;
3114 * Calculate key to locate first/next element, unlocking the previous
3115 * element as we go. Be careful, the key calculation can overflow.
3117 * (also reset bref to NULL)
3123 key = bref->key + ((hammer2_key_t)1 << bref->keybits);
3124 if ((chain = *chainp) != NULL) {
3126 hammer2_chain_unlock(chain);
3127 hammer2_chain_drop(chain);
3131 error |= HAMMER2_ERROR_EOF;
3137 if (parent->error) {
3138 error = parent->error;
3141 if (--maxloops == 0)
3142 panic("hammer2_chain_scan: maxloops");
3145 * Locate the blockref array. Currently we do a fully associative
3146 * search through the array.
3148 switch(parent->bref.type) {
3149 case HAMMER2_BREF_TYPE_INODE:
3151 * An inode with embedded data has no sub-chains.
3153 * WARNING! Bulk scan code may pass a static chain marked
3154 * as BREF_TYPE_INODE with a copy of the volume
3155 * root blockset to snapshot the volume.
3157 if (parent->data->ipdata.meta.op_flags &
3158 HAMMER2_OPFLAG_DIRECTDATA) {
3159 error |= HAMMER2_ERROR_EOF;
3162 base = &parent->data->ipdata.u.blockset.blockref[0];
3163 count = HAMMER2_SET_COUNT;
3165 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3166 case HAMMER2_BREF_TYPE_INDIRECT:
3168 * Optimize indirect blocks in the INITIAL state to avoid
3171 if (parent->flags & HAMMER2_CHAIN_INITIAL) {
3174 if (parent->data == NULL)
3175 panic("parent->data is NULL");
3176 base = &parent->data->npdata[0];
3178 count = parent->bytes / sizeof(hammer2_blockref_t);
3180 case HAMMER2_BREF_TYPE_VOLUME:
3181 base = &parent->data->voldata.sroot_blockset.blockref[0];
3182 count = HAMMER2_SET_COUNT;
3184 case HAMMER2_BREF_TYPE_FREEMAP:
3185 base = &parent->data->blkset.blockref[0];
3186 count = HAMMER2_SET_COUNT;
3189 panic("hammer2_chain_scan: unrecognized blockref type: %d",
3191 base = NULL; /* safety */
3192 count = 0; /* safety */
3197 * Merged scan to find next candidate.
3199 * hammer2_base_*() functions require the parent->core.live_* fields
3200 * to be synchronized.
3202 * We need to hold the spinlock to access the block array and RB tree
3203 * and to interlock chain creation.
3205 if ((parent->flags & HAMMER2_CHAIN_COUNTEDBREFS) == 0)
3206 hammer2_chain_countbrefs(parent, base, count);
3210 hammer2_spin_ex(&parent->core.spin);
3211 chain = hammer2_combined_find(parent, base, count,
3213 key, HAMMER2_KEY_MAX,
3215 generation = parent->core.generation;
3218 * Exhausted parent chain, we're done.
3220 if (bref_ptr == NULL) {
3221 hammer2_spin_unex(&parent->core.spin);
3222 KKASSERT(chain == NULL);
3223 error |= HAMMER2_ERROR_EOF;
3228 * Copy into the supplied stack-based blockref.
3233 * Selected from blockref or in-memory chain.
3235 if (chain == NULL) {
3236 switch(bref->type) {
3237 case HAMMER2_BREF_TYPE_INODE:
3238 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3239 case HAMMER2_BREF_TYPE_INDIRECT:
3240 case HAMMER2_BREF_TYPE_VOLUME:
3241 case HAMMER2_BREF_TYPE_FREEMAP:
3243 * Recursion, always get the chain
3245 hammer2_spin_unex(&parent->core.spin);
3246 chain = hammer2_chain_get(parent, generation,
3253 * No recursion, do not waste time instantiating
3254 * a chain, just iterate using the bref.
3256 hammer2_spin_unex(&parent->core.spin);
3261 * Recursion or not we need the chain in order to supply
3264 hammer2_chain_ref(chain);
3265 hammer2_spin_unex(&parent->core.spin);
3266 hammer2_chain_lock(chain, how);
3269 (bcmp(bref, &chain->bref, sizeof(*bref)) ||
3270 chain->parent != parent)) {
3271 hammer2_chain_unlock(chain);
3272 hammer2_chain_drop(chain);
3278 * Skip deleted chains (XXX cache 'i' end-of-block-array? XXX)
3280 * NOTE: chain's key range is not relevant as there might be
3281 * one-offs within the range that are not deleted.
3283 * NOTE: XXX this could create problems with scans used in
3284 * situations other than mount-time recovery.
3286 * NOTE: Lookups can race delete-duplicate because
3287 * delete-duplicate does not lock the parent's core
3288 * (they just use the spinlock on the core).
3290 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) {
3291 hammer2_chain_unlock(chain);
3292 hammer2_chain_drop(chain);
3297 error |= HAMMER2_ERROR_EOF;
3305 * All done, return the bref or NULL, supply chain if necessary.
3313 * Create and return a new hammer2 system memory structure of the specified
3314 * key, type and size and insert it under (*parentp). This is a full
3315 * insertion, based on the supplied key/keybits, and may involve creating
3316 * indirect blocks and moving other chains around via delete/duplicate.
3318 * This call can be made with parent == NULL as long as a non -1 methods
3319 * is supplied. hmp must also be supplied in this situation (otherwise
3320 * hmp is extracted from the supplied parent). The chain will be detached
3321 * from the topology. A later call with both parent and chain can be made
3324 * THE CALLER MUST HAVE ALREADY PROPERLY SEEKED (*parentp) TO THE INSERTION
3325 * POINT SANS ANY REQUIRED INDIRECT BLOCK CREATIONS DUE TO THE ARRAY BEING
3326 * FULL. This typically means that the caller is creating the chain after
3327 * doing a hammer2_chain_lookup().
3329 * (*parentp) must be exclusive locked and may be replaced on return
3330 * depending on how much work the function had to do.
3332 * (*parentp) must not be errored or this function will assert.
3334 * (*chainp) usually starts out NULL and returns the newly created chain,
3335 * but if the caller desires the caller may allocate a disconnected chain
3336 * and pass it in instead.
3338 * This function should NOT be used to insert INDIRECT blocks. It is
3339 * typically used to create/insert inodes and data blocks.
3341 * Caller must pass-in an exclusively locked parent the new chain is to
3342 * be inserted under, and optionally pass-in a disconnected, exclusively
3343 * locked chain to insert (else we create a new chain). The function will
3344 * adjust (*parentp) as necessary, create or connect the chain, and
3345 * return an exclusively locked chain in *chainp.
3347 * When creating a PFSROOT inode under the super-root, pmp is typically NULL
3348 * and will be reassigned.
3350 * NOTE: returns HAMMER_ERROR_* flags
3353 hammer2_chain_create(hammer2_chain_t **parentp, hammer2_chain_t **chainp,
3354 hammer2_dev_t *hmp, hammer2_pfs_t *pmp, int methods,
3355 hammer2_key_t key, int keybits, int type, size_t bytes,
3356 hammer2_tid_t mtid, hammer2_off_t dedup_off, int flags)
3358 hammer2_chain_t *chain;
3359 hammer2_chain_t *parent;
3360 hammer2_blockref_t *base;
3361 hammer2_blockref_t dummy;
3365 int maxloops = 300000;
3368 * Topology may be crossing a PFS boundary.
3372 KKASSERT(hammer2_mtx_owned(&parent->lock));
3373 KKASSERT(parent->error == 0);
3378 if (chain == NULL) {
3380 * First allocate media space and construct the dummy bref,
3381 * then allocate the in-memory chain structure. Set the
3382 * INITIAL flag for fresh chains which do not have embedded
3385 bzero(&dummy, sizeof(dummy));
3388 dummy.keybits = keybits;
3389 dummy.data_off = hammer2_getradix(bytes);
3392 * Inherit methods from parent by default. Primarily used
3393 * for BREF_TYPE_DATA. Non-data types *must* be set to
3394 * a non-NONE check algorithm.
3397 dummy.methods = parent->bref.methods;
3399 dummy.methods = (uint8_t)methods;
3401 if (type != HAMMER2_BREF_TYPE_DATA &&
3402 HAMMER2_DEC_CHECK(dummy.methods) == HAMMER2_CHECK_NONE) {
3404 HAMMER2_ENC_CHECK(HAMMER2_CHECK_DEFAULT);
3407 chain = hammer2_chain_alloc(hmp, pmp, &dummy);
3410 * Lock the chain manually, chain_lock will load the chain
3411 * which we do NOT want to do. (note: chain->refs is set
3412 * to 1 by chain_alloc() for us, but lockcnt is not).
3415 hammer2_mtx_ex(&chain->lock);
3419 * Set INITIAL to optimize I/O. The flag will generally be
3420 * processed when we call hammer2_chain_modify().
3423 case HAMMER2_BREF_TYPE_VOLUME:
3424 case HAMMER2_BREF_TYPE_FREEMAP:
3425 panic("hammer2_chain_create: called with volume type");
3427 case HAMMER2_BREF_TYPE_INDIRECT:
3428 panic("hammer2_chain_create: cannot be used to"
3429 "create indirect block");
3431 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3432 panic("hammer2_chain_create: cannot be used to"
3433 "create freemap root or node");
3435 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
3436 KKASSERT(bytes == sizeof(chain->data->bmdata));
3438 case HAMMER2_BREF_TYPE_DIRENT:
3439 case HAMMER2_BREF_TYPE_INODE:
3440 case HAMMER2_BREF_TYPE_DATA:
3443 * leave chain->data NULL, set INITIAL
3445 KKASSERT(chain->data == NULL);
3446 atomic_set_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
3451 * We are reattaching a previously deleted chain, possibly
3452 * under a new parent and possibly with a new key/keybits.
3453 * The chain does not have to be in a modified state. The
3454 * UPDATE flag will be set later on in this routine.
3456 * Do NOT mess with the current state of the INITIAL flag.
3458 chain->bref.key = key;
3459 chain->bref.keybits = keybits;
3460 if (chain->flags & HAMMER2_CHAIN_DELETED)
3461 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_DELETED);
3462 KKASSERT(chain->parent == NULL);
3466 * Set the appropriate bref flag if requested.
3468 * NOTE! Callers can call this function to move chains without
3469 * knowing about special flags, so don't clear bref flags
3472 if (flags & HAMMER2_INSERT_PFSROOT)
3473 chain->bref.flags |= HAMMER2_BREF_FLAG_PFSROOT;
3479 * Calculate how many entries we have in the blockref array and
3480 * determine if an indirect block is required when inserting into
3484 if (--maxloops == 0)
3485 panic("hammer2_chain_create: maxloops");
3487 switch(parent->bref.type) {
3488 case HAMMER2_BREF_TYPE_INODE:
3489 if ((parent->data->ipdata.meta.op_flags &
3490 HAMMER2_OPFLAG_DIRECTDATA) != 0) {
3491 kprintf("hammer2: parent set for direct-data! "
3492 "pkey=%016jx ckey=%016jx\n",
3496 KKASSERT((parent->data->ipdata.meta.op_flags &
3497 HAMMER2_OPFLAG_DIRECTDATA) == 0);
3498 KKASSERT(parent->data != NULL);
3499 base = &parent->data->ipdata.u.blockset.blockref[0];
3500 count = HAMMER2_SET_COUNT;
3502 case HAMMER2_BREF_TYPE_INDIRECT:
3503 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3504 if (parent->flags & HAMMER2_CHAIN_INITIAL)
3507 base = &parent->data->npdata[0];
3508 count = parent->bytes / sizeof(hammer2_blockref_t);
3510 case HAMMER2_BREF_TYPE_VOLUME:
3511 KKASSERT(parent->data != NULL);
3512 base = &parent->data->voldata.sroot_blockset.blockref[0];
3513 count = HAMMER2_SET_COUNT;
3515 case HAMMER2_BREF_TYPE_FREEMAP:
3516 KKASSERT(parent->data != NULL);
3517 base = &parent->data->blkset.blockref[0];
3518 count = HAMMER2_SET_COUNT;
3521 panic("hammer2_chain_create: unrecognized blockref type: %d",
3529 * Make sure we've counted the brefs
3531 if ((parent->flags & HAMMER2_CHAIN_COUNTEDBREFS) == 0)
3532 hammer2_chain_countbrefs(parent, base, count);
3534 KASSERT(parent->core.live_count >= 0 &&
3535 parent->core.live_count <= count,
3536 ("bad live_count %d/%d (%02x, %d)",
3537 parent->core.live_count, count,
3538 parent->bref.type, parent->bytes));
3541 * If no free blockref could be found we must create an indirect
3542 * block and move a number of blockrefs into it. With the parent
3543 * locked we can safely lock each child in order to delete+duplicate
3544 * it without causing a deadlock.
3546 * This may return the new indirect block or the old parent depending
3547 * on where the key falls. NULL is returned on error.
3549 if (parent->core.live_count == count) {
3550 hammer2_chain_t *nparent;
3552 KKASSERT((flags & HAMMER2_INSERT_SAMEPARENT) == 0);
3554 nparent = hammer2_chain_create_indirect(parent, key, keybits,
3555 mtid, type, &error);
3556 if (nparent == NULL) {
3558 hammer2_chain_drop(chain);
3562 if (parent != nparent) {
3563 hammer2_chain_unlock(parent);
3564 hammer2_chain_drop(parent);
3565 parent = *parentp = nparent;
3571 * fall through if parent, or skip to here if no parent.
3574 if (chain->flags & HAMMER2_CHAIN_DELETED)
3575 kprintf("Inserting deleted chain @%016jx\n",
3579 * Link the chain into its parent.
3581 if (chain->parent != NULL)
3582 panic("hammer2: hammer2_chain_create: chain already connected");
3583 KKASSERT(chain->parent == NULL);
3585 KKASSERT(parent->core.live_count < count);
3586 hammer2_chain_insert(parent, chain,
3587 HAMMER2_CHAIN_INSERT_SPIN |
3588 HAMMER2_CHAIN_INSERT_LIVE,
3594 * Mark the newly created chain modified. This will cause
3595 * UPDATE to be set and process the INITIAL flag.
3597 * Device buffers are not instantiated for DATA elements
3598 * as these are handled by logical buffers.
3600 * Indirect and freemap node indirect blocks are handled
3601 * by hammer2_chain_create_indirect() and not by this
3604 * Data for all other bref types is expected to be
3605 * instantiated (INODE, LEAF).
3607 switch(chain->bref.type) {
3608 case HAMMER2_BREF_TYPE_DATA:
3609 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
3610 case HAMMER2_BREF_TYPE_DIRENT:
3611 case HAMMER2_BREF_TYPE_INODE:
3612 error = hammer2_chain_modify(chain, mtid, dedup_off,
3613 HAMMER2_MODIFY_OPTDATA);
3617 * Remaining types are not supported by this function.
3618 * In particular, INDIRECT and LEAF_NODE types are
3619 * handled by create_indirect().
3621 panic("hammer2_chain_create: bad type: %d",
3628 * When reconnecting a chain we must set UPDATE and
3629 * setflush so the flush recognizes that it must update
3630 * the bref in the parent.
3632 if ((chain->flags & HAMMER2_CHAIN_UPDATE) == 0)
3633 atomic_set_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
3637 * We must setflush(parent) to ensure that it recurses through to
3638 * chain. setflush(chain) might not work because ONFLUSH is possibly
3639 * already set in the chain (so it won't recurse up to set it in the
3643 hammer2_chain_setflush(parent);
3652 * Move the chain from its old parent to a new parent. The chain must have
3653 * already been deleted or already disconnected (or never associated) with
3654 * a parent. The chain is reassociated with the new parent and the deleted
3655 * flag will be cleared (no longer deleted). The chain's modification state
3658 * THE CALLER MUST HAVE ALREADY PROPERLY SEEKED (parent) TO THE INSERTION
3659 * POINT SANS ANY REQUIRED INDIRECT BLOCK CREATIONS DUE TO THE ARRAY BEING
3660 * FULL. This typically means that the caller is creating the chain after
3661 * doing a hammer2_chain_lookup().
3663 * Neither (parent) or (chain) can be errored.
3665 * If (parent) is non-NULL then the chain is inserted under the parent.
3667 * If (parent) is NULL then the newly duplicated chain is not inserted
3668 * anywhere, similar to if it had just been chain_alloc()'d (suitable for
3669 * passing into hammer2_chain_create() after this function returns).
3671 * WARNING! This function calls create which means it can insert indirect
3672 * blocks. This can cause other unrelated chains in the parent to
3673 * be moved to a newly inserted indirect block in addition to the
3677 hammer2_chain_rename(hammer2_chain_t **parentp, hammer2_chain_t *chain,
3678 hammer2_tid_t mtid, int flags)
3680 hammer2_blockref_t *bref;
3682 hammer2_chain_t *parent;
3685 * WARNING! We should never resolve DATA to device buffers
3686 * (XXX allow it if the caller did?), and since
3687 * we currently do not have the logical buffer cache
3688 * buffer in-hand to fix its cached physical offset
3689 * we also force the modify code to not COW it. XXX
3691 * NOTE! We allow error'd chains to be renamed. The bref itself
3692 * is good and can be renamed. The content, however, may
3696 KKASSERT(chain->parent == NULL);
3697 /*KKASSERT(chain->error == 0); allow */
3698 bref = &chain->bref;
3701 * If parent is not NULL the duplicated chain will be entered under
3702 * the parent and the UPDATE bit set to tell flush to update
3705 * We must setflush(parent) to ensure that it recurses through to
3706 * chain. setflush(chain) might not work because ONFLUSH is possibly
3707 * already set in the chain (so it won't recurse up to set it in the
3710 * Having both chains locked is extremely important for atomicy.
3712 if (parentp && (parent = *parentp) != NULL) {
3713 KKASSERT(hammer2_mtx_owned(&parent->lock));
3714 KKASSERT(parent->refs > 0);
3715 KKASSERT(parent->error == 0);
3717 hammer2_chain_create(parentp, &chain, NULL, chain->pmp,
3718 HAMMER2_METH_DEFAULT,
3719 bref->key, bref->keybits, bref->type,
3720 chain->bytes, mtid, 0, flags);
3721 KKASSERT(chain->flags & HAMMER2_CHAIN_UPDATE);
3722 hammer2_chain_setflush(*parentp);
3727 * This works in tandem with delete_obref() to install a blockref in
3728 * (typically) an indirect block that is associated with the chain being
3729 * moved to *parentp.
3731 * The reason we need this function is that the caller needs to maintain
3732 * the blockref as it was, and not generate a new blockref for what might
3733 * be a modified chain. Otherwise stuff will leak into the flush that
3734 * the flush code's FLUSH_INODE_STOP flag is unable to catch.
3736 * It is EXTREMELY important that we properly set CHAIN_BMAPUPD and
3737 * CHAIN_UPDATE. We must set BMAPUPD if the bref does not match, and
3738 * we must clear CHAIN_UPDATE (that was likely set by the chain_rename) if
3739 * it does. Otherwise we can end up in a situation where H2 is unable to
3740 * clean up the in-memory chain topology.
3742 * The reason for this is that flushes do not generally flush through
3743 * BREF_TYPE_INODE chains and depend on a hammer2_inode_t queued to syncq
3744 * or sideq to properly flush and dispose of the related inode chain's flags.
3745 * Situations where the inode is not actually modified by the frontend,
3746 * but where we have to move the related chains around as we insert or cleanup
3747 * indirect blocks, can leave us with a 'dirty' (non-disposable) in-memory
3748 * inode chain that does not have a hammer2_inode_t associated with it.
3751 hammer2_chain_rename_obref(hammer2_chain_t **parentp, hammer2_chain_t *chain,
3752 hammer2_tid_t mtid, int flags,
3753 hammer2_blockref_t *obref)
3755 hammer2_chain_rename(parentp, chain, mtid, flags);
3757 if (obref->type != HAMMER2_BREF_TYPE_EMPTY) {
3758 hammer2_blockref_t *tbase;
3761 KKASSERT((chain->flags & HAMMER2_CHAIN_BMAPPED) == 0);
3762 hammer2_chain_modify(*parentp, mtid, 0, 0);
3763 tbase = hammer2_chain_base_and_count(*parentp, &tcount);
3764 hammer2_base_insert(*parentp, tbase, tcount, chain, obref);
3765 if (bcmp(obref, &chain->bref, sizeof(chain->bref))) {
3766 atomic_set_int(&chain->flags, HAMMER2_CHAIN_BMAPUPD |
3767 HAMMER2_CHAIN_UPDATE);
3769 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
3775 * Helper function for deleting chains.
3777 * The chain is removed from the live view (the RBTREE) as well as the parent's
3778 * blockmap. Both chain and its parent must be locked.
3780 * parent may not be errored. chain can be errored.
3783 _hammer2_chain_delete_helper(hammer2_chain_t *parent, hammer2_chain_t *chain,
3784 hammer2_tid_t mtid, int flags,
3785 hammer2_blockref_t *obref)
3790 KKASSERT((chain->flags & HAMMER2_CHAIN_DELETED) == 0);
3791 KKASSERT(chain->parent == parent);
3794 if (chain->flags & HAMMER2_CHAIN_BMAPPED) {
3796 * Chain is blockmapped, so there must be a parent.
3797 * Atomically remove the chain from the parent and remove
3798 * the blockmap entry. The parent must be set modified
3799 * to remove the blockmap entry.
3801 hammer2_blockref_t *base;
3804 KKASSERT(parent != NULL);
3805 KKASSERT(parent->error == 0);
3806 KKASSERT((parent->flags & HAMMER2_CHAIN_INITIAL) == 0);
3807 error = hammer2_chain_modify(parent, mtid, 0, 0);
3812 * Calculate blockmap pointer
3814 KKASSERT(chain->flags & HAMMER2_CHAIN_ONRBTREE);
3815 hammer2_spin_ex(&chain->core.spin);
3816 hammer2_spin_ex(&parent->core.spin);
3818 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DELETED);
3819 atomic_add_int(&parent->core.live_count, -1);
3820 ++parent->core.generation;
3821 RB_REMOVE(hammer2_chain_tree, &parent->core.rbtree, chain);
3822 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE);
3823 --parent->core.chain_count;
3824 chain->parent = NULL;
3826 switch(parent->bref.type) {
3827 case HAMMER2_BREF_TYPE_INODE:
3829 * Access the inode's block array. However, there
3830 * is no block array if the inode is flagged
3834 (parent->data->ipdata.meta.op_flags &
3835 HAMMER2_OPFLAG_DIRECTDATA) == 0) {
3837 &parent->data->ipdata.u.blockset.blockref[0];
3841 count = HAMMER2_SET_COUNT;
3843 case HAMMER2_BREF_TYPE_INDIRECT:
3844 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3846 base = &parent->data->npdata[0];
3849 count = parent->bytes / sizeof(hammer2_blockref_t);
3851 case HAMMER2_BREF_TYPE_VOLUME:
3852 base = &parent->data->voldata.
3853 sroot_blockset.blockref[0];
3854 count = HAMMER2_SET_COUNT;
3856 case HAMMER2_BREF_TYPE_FREEMAP:
3857 base = &parent->data->blkset.blockref[0];
3858 count = HAMMER2_SET_COUNT;
3863 panic("_hammer2_chain_delete_helper: "
3864 "unrecognized blockref type: %d",
3870 * delete blockmapped chain from its parent.
3872 * The parent is not affected by any statistics in chain
3873 * which are pending synchronization. That is, there is
3874 * nothing to undo in the parent since they have not yet
3875 * been incorporated into the parent.
3877 * The parent is affected by statistics stored in inodes.
3878 * Those have already been synchronized, so they must be
3879 * undone. XXX split update possible w/delete in middle?
3882 hammer2_base_delete(parent, base, count, chain, obref);
3884 hammer2_spin_unex(&parent->core.spin);
3885 hammer2_spin_unex(&chain->core.spin);
3886 } else if (chain->flags & HAMMER2_CHAIN_ONRBTREE) {
3888 * Chain is not blockmapped but a parent is present.
3889 * Atomically remove the chain from the parent. There is
3890 * no blockmap entry to remove.
3892 * Because chain was associated with a parent but not
3893 * synchronized, the chain's *_count_up fields contain
3894 * inode adjustment statistics which must be undone.
3896 hammer2_spin_ex(&chain->core.spin);
3897 hammer2_spin_ex(&parent->core.spin);
3898 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DELETED);
3899 atomic_add_int(&parent->core.live_count, -1);
3900 ++parent->core.generation;
3901 RB_REMOVE(hammer2_chain_tree, &parent->core.rbtree, chain);
3902 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE);
3903 --parent->core.chain_count;
3904 chain->parent = NULL;
3905 hammer2_spin_unex(&parent->core.spin);
3906 hammer2_spin_unex(&chain->core.spin);
3909 * Chain is not blockmapped and has no parent. This
3910 * is a degenerate case.
3912 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DELETED);
3919 * Create an indirect block that covers one or more of the elements in the
3920 * current parent. Either returns the existing parent with no locking or
3921 * ref changes or returns the new indirect block locked and referenced
3922 * and leaving the original parent lock/ref intact as well.
3924 * If an error occurs, NULL is returned and *errorp is set to the H2 error.
3926 * The returned chain depends on where the specified key falls.
3928 * The key/keybits for the indirect mode only needs to follow three rules:
3930 * (1) That all elements underneath it fit within its key space and
3932 * (2) That all elements outside it are outside its key space.
3934 * (3) When creating the new indirect block any elements in the current
3935 * parent that fit within the new indirect block's keyspace must be
3936 * moved into the new indirect block.
3938 * (4) The keyspace chosen for the inserted indirect block CAN cover a wider
3939 * keyspace the the current parent, but lookup/iteration rules will
3940 * ensure (and must ensure) that rule (2) for all parents leading up
3941 * to the nearest inode or the root volume header is adhered to. This
3942 * is accomplished by always recursing through matching keyspaces in
3943 * the hammer2_chain_lookup() and hammer2_chain_next() API.
3945 * The current implementation calculates the current worst-case keyspace by
3946 * iterating the current parent and then divides it into two halves, choosing
3947 * whichever half has the most elements (not necessarily the half containing
3948 * the requested key).
3950 * We can also opt to use the half with the least number of elements. This
3951 * causes lower-numbered keys (aka logical file offsets) to recurse through
3952 * fewer indirect blocks and higher-numbered keys to recurse through more.
3953 * This also has the risk of not moving enough elements to the new indirect
3954 * block and being forced to create several indirect blocks before the element
3957 * Must be called with an exclusively locked parent.
3959 * NOTE: *errorp set to HAMMER_ERROR_* flags
3961 static int hammer2_chain_indkey_freemap(hammer2_chain_t *parent,
3962 hammer2_key_t *keyp, int keybits,
3963 hammer2_blockref_t *base, int count);
3964 static int hammer2_chain_indkey_file(hammer2_chain_t *parent,
3965 hammer2_key_t *keyp, int keybits,
3966 hammer2_blockref_t *base, int count,
3968 static int hammer2_chain_indkey_dir(hammer2_chain_t *parent,
3969 hammer2_key_t *keyp, int keybits,
3970 hammer2_blockref_t *base, int count,
3974 hammer2_chain_create_indirect(hammer2_chain_t *parent,
3975 hammer2_key_t create_key, int create_bits,
3976 hammer2_tid_t mtid, int for_type, int *errorp)
3979 hammer2_blockref_t *base;
3980 hammer2_blockref_t *bref;
3981 hammer2_blockref_t bsave;
3982 hammer2_blockref_t dummy;
3983 hammer2_chain_t *chain;
3984 hammer2_chain_t *ichain;
3985 hammer2_key_t key = create_key;
3986 hammer2_key_t key_beg;
3987 hammer2_key_t key_end;
3988 hammer2_key_t key_next;
3989 int keybits = create_bits;
3997 int maxloops = 300000;
4000 * Calculate the base blockref pointer or NULL if the chain
4001 * is known to be empty. We need to calculate the array count
4002 * for RB lookups either way.
4005 KKASSERT(hammer2_mtx_owned(&parent->lock));
4008 * Pre-modify the parent now to avoid having to deal with error
4009 * processing if we tried to later (in the middle of our loop).
4011 * We are going to be moving bref's around, the indirect blocks
4012 * cannot be in an initial state. Do not pass MODIFY_OPTDATA.
4014 *errorp = hammer2_chain_modify(parent, mtid, 0, 0);
4016 kprintf("hammer2_chain_create_indirect: error %08x %s\n",
4017 *errorp, hammer2_error_str(*errorp));
4020 KKASSERT((parent->flags & HAMMER2_CHAIN_INITIAL) == 0);
4022 /*hammer2_chain_modify(&parent, HAMMER2_MODIFY_OPTDATA);*/
4023 base = hammer2_chain_base_and_count(parent, &count);
4026 * How big should our new indirect block be? It has to be at least
4027 * as large as its parent for splits to work properly.
4029 * The freemap uses a specific indirect block size. The number of
4030 * levels are built dynamically and ultimately depend on the size
4031 * volume. Because freemap blocks are taken from the reserved areas
4032 * of the volume our goal is efficiency (fewer levels) and not so
4033 * much to save disk space.
4035 * The first indirect block level for a directory usually uses
4036 * HAMMER2_IND_BYTES_MIN (4KB = 32 directory entries). Due to
4037 * the hash mechanism, this typically gives us a nominal
4038 * 32 * 4 entries with one level of indirection.
4040 * We use HAMMER2_IND_BYTES_NOM (16KB = 128 blockrefs) for FILE
4041 * indirect blocks. The initial 4 entries in the inode gives us
4042 * 256KB. Up to 4 indirect blocks gives us 32MB. Three levels
4043 * of indirection gives us 137GB, and so forth. H2 can support
4044 * huge file sizes but they are not typical, so we try to stick
4045 * with compactness and do not use a larger indirect block size.
4047 * We could use 64KB (PBUFSIZE), giving us 512 blockrefs, but
4048 * due to the way indirect blocks are created this usually winds
4049 * up being extremely inefficient for small files. Even though
4050 * 16KB requires more levels of indirection for very large files,
4051 * the 16KB records can be ganged together into 64KB DIOs.
4053 if (for_type == HAMMER2_BREF_TYPE_FREEMAP_NODE ||
4054 for_type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) {
4055 nbytes = HAMMER2_FREEMAP_LEVELN_PSIZE;
4056 } else if (parent->bref.type == HAMMER2_BREF_TYPE_INODE) {
4057 if (parent->data->ipdata.meta.type ==
4058 HAMMER2_OBJTYPE_DIRECTORY)
4059 nbytes = HAMMER2_IND_BYTES_MIN; /* 4KB = 32 entries */
4061 nbytes = HAMMER2_IND_BYTES_NOM; /* 16KB = ~8MB file */
4064 nbytes = HAMMER2_IND_BYTES_NOM;
4066 if (nbytes < count * sizeof(hammer2_blockref_t)) {
4067 KKASSERT(for_type != HAMMER2_BREF_TYPE_FREEMAP_NODE &&
4068 for_type != HAMMER2_BREF_TYPE_FREEMAP_LEAF);
4069 nbytes = count * sizeof(hammer2_blockref_t);
4071 ncount = nbytes / sizeof(hammer2_blockref_t);
4074 * When creating an indirect block for a freemap node or leaf
4075 * the key/keybits must be fitted to static radix levels because
4076 * particular radix levels use particular reserved blocks in the
4079 * This routine calculates the key/radix of the indirect block
4080 * we need to create, and whether it is on the high-side or the
4084 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
4085 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
4086 keybits = hammer2_chain_indkey_freemap(parent, &key, keybits,
4089 case HAMMER2_BREF_TYPE_DATA:
4090 keybits = hammer2_chain_indkey_file(parent, &key, keybits,
4091 base, count, ncount);
4093 case HAMMER2_BREF_TYPE_DIRENT:
4094 case HAMMER2_BREF_TYPE_INODE:
4095 keybits = hammer2_chain_indkey_dir(parent, &key, keybits,
4096 base, count, ncount);
4099 panic("illegal indirect block for bref type %d", for_type);
4104 * Normalize the key for the radix being represented, keeping the
4105 * high bits and throwing away the low bits.
4107 key &= ~(((hammer2_key_t)1 << keybits) - 1);
4110 * Ok, create our new indirect block
4112 bzero(&dummy, sizeof(dummy));
4113 if (for_type == HAMMER2_BREF_TYPE_FREEMAP_NODE ||
4114 for_type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) {
4115 dummy.type = HAMMER2_BREF_TYPE_FREEMAP_NODE;
4117 dummy.type = HAMMER2_BREF_TYPE_INDIRECT;
4120 dummy.keybits = keybits;
4121 dummy.data_off = hammer2_getradix(nbytes);
4123 HAMMER2_ENC_CHECK(HAMMER2_DEC_CHECK(parent->bref.methods)) |
4124 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
4126 ichain = hammer2_chain_alloc(hmp, parent->pmp, &dummy);
4127 atomic_set_int(&ichain->flags, HAMMER2_CHAIN_INITIAL);
4128 hammer2_chain_lock(ichain, HAMMER2_RESOLVE_MAYBE);
4129 /* ichain has one ref at this point */
4132 * We have to mark it modified to allocate its block, but use
4133 * OPTDATA to allow it to remain in the INITIAL state. Otherwise
4134 * it won't be acted upon by the flush code.
4136 * XXX remove OPTDATA, we need a fully initialized indirect block to
4137 * be able to move the original blockref.
4139 *errorp = hammer2_chain_modify(ichain, mtid, 0, 0);
4141 kprintf("hammer2_chain_create_indirect: error %08x %s\n",
4142 *errorp, hammer2_error_str(*errorp));
4143 hammer2_chain_unlock(ichain);
4144 hammer2_chain_drop(ichain);
4147 KKASSERT((ichain->flags & HAMMER2_CHAIN_INITIAL) == 0);
4150 * Iterate the original parent and move the matching brefs into
4151 * the new indirect block.
4153 * XXX handle flushes.
4156 key_end = HAMMER2_KEY_MAX;
4157 key_next = 0; /* avoid gcc warnings */
4158 hammer2_spin_ex(&parent->core.spin);
4164 * Parent may have been modified, relocating its block array.
4165 * Reload the base pointer.
4167 base = hammer2_chain_base_and_count(parent, &count);
4169 if (++loops > 100000) {
4170 hammer2_spin_unex(&parent->core.spin);
4171 panic("excessive loops r=%d p=%p base/count %p:%d %016jx\n",
4172 reason, parent, base, count, key_next);
4176 * NOTE: spinlock stays intact, returned chain (if not NULL)
4177 * is not referenced or locked which means that we
4178 * cannot safely check its flagged / deletion status
4181 chain = hammer2_combined_find(parent, base, count,
4185 generation = parent->core.generation;
4188 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
4191 * Skip keys that are not within the key/radix of the new
4192 * indirect block. They stay in the parent.
4194 if (rounddown2(key ^ bref->key, (hammer2_key_t)1 << keybits) != 0) {
4195 goto next_key_spinlocked;
4199 * Load the new indirect block by acquiring the related
4200 * chains (potentially from media as it might not be
4201 * in-memory). Then move it to the new parent (ichain).
4203 * chain is referenced but not locked. We must lock the
4204 * chain to obtain definitive state.
4209 * Use chain already present in the RBTREE
4211 hammer2_chain_ref(chain);
4212 hammer2_spin_unex(&parent->core.spin);
4213 hammer2_chain_lock(chain, HAMMER2_RESOLVE_NEVER);
4216 * Get chain for blockref element. _get returns NULL
4217 * on insertion race.
4219 hammer2_spin_unex(&parent->core.spin);
4220 chain = hammer2_chain_get(parent, generation, &bsave,
4221 HAMMER2_RESOLVE_NEVER);
4222 if (chain == NULL) {
4224 hammer2_spin_ex(&parent->core.spin);
4230 * This is always live so if the chain has been deleted
4231 * we raced someone and we have to retry.
4233 * NOTE: Lookups can race delete-duplicate because
4234 * delete-duplicate does not lock the parent's core
4235 * (they just use the spinlock on the core).
4237 * (note reversed logic for this one)
4239 if (bcmp(&bsave, &chain->bref, sizeof(bsave)) ||
4240 chain->parent != parent ||
4241 (chain->flags & HAMMER2_CHAIN_DELETED)) {
4242 hammer2_chain_unlock(chain);
4243 hammer2_chain_drop(chain);
4244 if (hammer2_debug & 0x0040) {
4245 kprintf("LOST PARENT RETRY "
4246 "RETRY (%p,%p)->%p %08x\n",
4247 parent, chain->parent, chain, chain->flags);
4249 hammer2_spin_ex(&parent->core.spin);
4254 * Shift the chain to the indirect block.
4256 * WARNING! No reason for us to load chain data, pass NOSTATS
4257 * to prevent delete/insert from trying to access
4258 * inode stats (and thus asserting if there is no
4259 * chain->data loaded).
4261 * WARNING! The (parent, chain) deletion may modify the parent
4262 * and invalidate the base pointer.
4264 * WARNING! Parent must already be marked modified, so we
4265 * can assume that chain_delete always suceeds.
4267 * WARNING! hammer2_chain_repchange() does not have to be
4268 * called (and doesn't work anyway because we are
4269 * only doing a partial shift). A recursion that is
4270 * in-progress can continue at the current parent
4271 * and will be able to properly find its next key.
4273 error = hammer2_chain_delete_obref(parent, chain, mtid, 0,
4275 KKASSERT(error == 0);
4276 hammer2_chain_rename_obref(&ichain, chain, mtid, 0, &bsave);
4277 hammer2_chain_unlock(chain);
4278 hammer2_chain_drop(chain);
4279 KKASSERT(parent->refs > 0);
4281 base = NULL; /* safety */
4282 hammer2_spin_ex(&parent->core.spin);
4283 next_key_spinlocked:
4284 if (--maxloops == 0)
4285 panic("hammer2_chain_create_indirect: maxloops");
4287 if (key_next == 0 || key_next > key_end)
4292 hammer2_spin_unex(&parent->core.spin);
4295 * Insert the new indirect block into the parent now that we've
4296 * cleared out some entries in the parent. We calculated a good
4297 * insertion index in the loop above (ichain->index).
4299 * We don't have to set UPDATE here because we mark ichain
4300 * modified down below (so the normal modified -> flush -> set-moved
4301 * sequence applies).
4303 * The insertion shouldn't race as this is a completely new block
4304 * and the parent is locked.
4306 base = NULL; /* safety, parent modify may change address */
4307 KKASSERT((ichain->flags & HAMMER2_CHAIN_ONRBTREE) == 0);
4308 KKASSERT(parent->core.live_count < count);
4309 hammer2_chain_insert(parent, ichain,
4310 HAMMER2_CHAIN_INSERT_SPIN |
4311 HAMMER2_CHAIN_INSERT_LIVE,
4315 * Make sure flushes propogate after our manual insertion.
4317 hammer2_chain_setflush(ichain);
4318 hammer2_chain_setflush(parent);
4321 * Figure out what to return.
4323 if (rounddown2(create_key ^ key, (hammer2_key_t)1 << keybits)) {
4325 * Key being created is outside the key range,
4326 * return the original parent.
4328 hammer2_chain_unlock(ichain);
4329 hammer2_chain_drop(ichain);
4332 * Otherwise its in the range, return the new parent.
4333 * (leave both the new and old parent locked).
4342 * Do maintenance on an indirect chain. Both parent and chain are locked.
4344 * Returns non-zero if (chain) is deleted, either due to being empty or
4345 * because its children were safely moved into the parent.
4348 hammer2_chain_indirect_maintenance(hammer2_chain_t *parent,
4349 hammer2_chain_t *chain)
4351 hammer2_blockref_t *chain_base;
4352 hammer2_blockref_t *base;
4353 hammer2_blockref_t *bref;
4354 hammer2_blockref_t bsave;
4355 hammer2_key_t key_next;
4356 hammer2_key_t key_beg;
4357 hammer2_key_t key_end;
4358 hammer2_chain_t *sub;
4365 * Make sure we have an accurate live_count
4367 if ((chain->flags & (HAMMER2_CHAIN_INITIAL |
4368 HAMMER2_CHAIN_COUNTEDBREFS)) == 0) {
4369 base = &chain->data->npdata[0];
4370 count = chain->bytes / sizeof(hammer2_blockref_t);
4371 hammer2_chain_countbrefs(chain, base, count);
4375 * If the indirect block is empty we can delete it.
4376 * (ignore deletion error)
4378 if (chain->core.live_count == 0 && RB_EMPTY(&chain->core.rbtree)) {
4379 hammer2_chain_delete(parent, chain,
4380 chain->bref.modify_tid,
4381 HAMMER2_DELETE_PERMANENT);
4382 hammer2_chain_repchange(parent, chain);
4386 base = hammer2_chain_base_and_count(parent, &count);
4388 if ((parent->flags & (HAMMER2_CHAIN_INITIAL |
4389 HAMMER2_CHAIN_COUNTEDBREFS)) == 0) {
4390 hammer2_chain_countbrefs(parent, base, count);
4394 * Determine if we can collapse chain into parent, calculate
4395 * hysteresis for chain emptiness.
4397 if (parent->core.live_count + chain->core.live_count - 1 > count)
4399 chain_count = chain->bytes / sizeof(hammer2_blockref_t);
4400 if (chain->core.live_count > chain_count * 3 / 4)
4404 * Ok, theoretically we can collapse chain's contents into
4405 * parent. chain is locked, but any in-memory children of chain
4406 * are not. For this to work, we must be able to dispose of any
4407 * in-memory children of chain.
4409 * For now require that there are no in-memory children of chain.
4411 * WARNING! Both chain and parent must remain locked across this
4416 * Parent must be marked modified. Don't try to collapse it if we
4417 * can't mark it modified. Once modified, destroy chain to make room
4418 * and to get rid of what will be a conflicting key (this is included
4419 * in the calculation above). Finally, move the children of chain
4420 * into chain's parent.
4422 * This order creates an accounting problem for bref.embed.stats
4423 * because we destroy chain before we remove its children. Any
4424 * elements whos blockref is already synchronized will be counted
4425 * twice. To deal with the problem we clean out chain's stats prior
4428 error = hammer2_chain_modify(parent, 0, 0, 0);
4430 krateprintf(&krate_h2me, "hammer2: indirect_maint: %s\n",
4431 hammer2_error_str(error));
4434 error = hammer2_chain_modify(chain, chain->bref.modify_tid, 0, 0);
4436 krateprintf(&krate_h2me, "hammer2: indirect_maint: %s\n",
4437 hammer2_error_str(error));
4441 chain->bref.embed.stats.inode_count = 0;
4442 chain->bref.embed.stats.data_count = 0;
4443 error = hammer2_chain_delete(parent, chain,
4444 chain->bref.modify_tid,
4445 HAMMER2_DELETE_PERMANENT);
4446 KKASSERT(error == 0);
4449 * The combined_find call requires core.spin to be held. One would
4450 * think there wouldn't be any conflicts since we hold chain
4451 * exclusively locked, but the caching mechanism for 0-ref children
4452 * does not require a chain lock.
4454 hammer2_spin_ex(&chain->core.spin);
4458 key_end = HAMMER2_KEY_MAX;
4460 chain_base = &chain->data->npdata[0];
4461 chain_count = chain->bytes / sizeof(hammer2_blockref_t);
4462 sub = hammer2_combined_find(chain, chain_base, chain_count,
4466 generation = chain->core.generation;
4469 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
4473 hammer2_chain_ref(sub);
4474 hammer2_spin_unex(&chain->core.spin);
4475 hammer2_chain_lock(sub, HAMMER2_RESOLVE_NEVER);
4477 hammer2_spin_unex(&chain->core.spin);
4478 sub = hammer2_chain_get(chain, generation, &bsave,
4479 HAMMER2_RESOLVE_NEVER);
4481 hammer2_spin_ex(&chain->core.spin);
4485 if (bcmp(&bsave, &sub->bref, sizeof(bsave)) ||
4486 sub->parent != chain ||
4487 (sub->flags & HAMMER2_CHAIN_DELETED)) {
4488 hammer2_chain_unlock(sub);
4489 hammer2_chain_drop(sub);
4490 hammer2_spin_ex(&chain->core.spin);
4491 sub = NULL; /* safety */
4494 error = hammer2_chain_delete_obref(chain, sub,
4495 sub->bref.modify_tid, 0,
4497 KKASSERT(error == 0);
4498 hammer2_chain_rename_obref(&parent, sub,
4499 sub->bref.modify_tid,
4500 HAMMER2_INSERT_SAMEPARENT, &bsave);
4501 hammer2_chain_unlock(sub);
4502 hammer2_chain_drop(sub);
4503 hammer2_spin_ex(&chain->core.spin);
4509 hammer2_spin_unex(&chain->core.spin);
4511 hammer2_chain_repchange(parent, chain);
4517 * Freemap indirect blocks
4519 * Calculate the keybits and highside/lowside of the freemap node the
4520 * caller is creating.
4522 * This routine will specify the next higher-level freemap key/radix
4523 * representing the lowest-ordered set. By doing so, eventually all
4524 * low-ordered sets will be moved one level down.
4526 * We have to be careful here because the freemap reserves a limited
4527 * number of blocks for a limited number of levels. So we can't just
4528 * push indiscriminately.
4531 hammer2_chain_indkey_freemap(hammer2_chain_t *parent, hammer2_key_t *keyp,
4532 int keybits, hammer2_blockref_t *base, int count)
4534 hammer2_chain_t *chain;
4535 hammer2_blockref_t *bref;
4537 hammer2_key_t key_beg;
4538 hammer2_key_t key_end;
4539 hammer2_key_t key_next;
4542 int maxloops = 300000;
4550 * Calculate the range of keys in the array being careful to skip
4551 * slots which are overridden with a deletion.
4554 key_end = HAMMER2_KEY_MAX;
4555 hammer2_spin_ex(&parent->core.spin);
4558 if (--maxloops == 0) {
4559 panic("indkey_freemap shit %p %p:%d\n",
4560 parent, base, count);
4562 chain = hammer2_combined_find(parent, base, count,
4574 * Skip deleted chains.
4576 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) {
4577 if (key_next == 0 || key_next > key_end)
4584 * Use the full live (not deleted) element for the scan
4585 * iteration. HAMMER2 does not allow partial replacements.
4587 * XXX should be built into hammer2_combined_find().
4589 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
4591 if (keybits > bref->keybits) {
4593 keybits = bref->keybits;
4594 } else if (keybits == bref->keybits && bref->key < key) {
4601 hammer2_spin_unex(&parent->core.spin);
4604 * Return the keybits for a higher-level FREEMAP_NODE covering
4608 case HAMMER2_FREEMAP_LEVEL0_RADIX:
4609 keybits = HAMMER2_FREEMAP_LEVEL1_RADIX;
4611 case HAMMER2_FREEMAP_LEVEL1_RADIX:
4612 keybits = HAMMER2_FREEMAP_LEVEL2_RADIX;
4614 case HAMMER2_FREEMAP_LEVEL2_RADIX:
4615 keybits = HAMMER2_FREEMAP_LEVEL3_RADIX;
4617 case HAMMER2_FREEMAP_LEVEL3_RADIX:
4618 keybits = HAMMER2_FREEMAP_LEVEL4_RADIX;
4620 case HAMMER2_FREEMAP_LEVEL4_RADIX:
4621 keybits = HAMMER2_FREEMAP_LEVEL5_RADIX;
4623 case HAMMER2_FREEMAP_LEVEL5_RADIX:
4624 panic("hammer2_chain_indkey_freemap: level too high");
4627 panic("hammer2_chain_indkey_freemap: bad radix");
4636 * File indirect blocks
4638 * Calculate the key/keybits for the indirect block to create by scanning
4639 * existing keys. The key being created is also passed in *keyp and can be
4640 * inside or outside the indirect block. Regardless, the indirect block
4641 * must hold at least two keys in order to guarantee sufficient space.
4643 * We use a modified version of the freemap's fixed radix tree, but taylored
4644 * for file data. Basically we configure an indirect block encompassing the
4648 hammer2_chain_indkey_file(hammer2_chain_t *parent, hammer2_key_t *keyp,
4649 int keybits, hammer2_blockref_t *base, int count,
4652 hammer2_chain_t *chain;
4653 hammer2_blockref_t *bref;
4655 hammer2_key_t key_beg;
4656 hammer2_key_t key_end;
4657 hammer2_key_t key_next;
4661 int maxloops = 300000;
4669 * Calculate the range of keys in the array being careful to skip
4670 * slots which are overridden with a deletion.
4672 * Locate the smallest key.
4675 key_end = HAMMER2_KEY_MAX;
4676 hammer2_spin_ex(&parent->core.spin);
4679 if (--maxloops == 0) {
4680 panic("indkey_freemap shit %p %p:%d\n",
4681 parent, base, count);
4683 chain = hammer2_combined_find(parent, base, count,
4695 * Skip deleted chains.
4697 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) {
4698 if (key_next == 0 || key_next > key_end)
4705 * Use the full live (not deleted) element for the scan
4706 * iteration. HAMMER2 does not allow partial replacements.
4708 * XXX should be built into hammer2_combined_find().
4710 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
4712 if (keybits > bref->keybits) {
4714 keybits = bref->keybits;
4715 } else if (keybits == bref->keybits && bref->key < key) {
4722 hammer2_spin_unex(&parent->core.spin);
4725 * Calculate the static keybits for a higher-level indirect block
4726 * that contains the key.
4731 case HAMMER2_IND_BYTES_MIN / sizeof(hammer2_blockref_t):
4732 nradix = HAMMER2_IND_RADIX_MIN - HAMMER2_BLOCKREF_RADIX;
4734 case HAMMER2_IND_BYTES_NOM / sizeof(hammer2_blockref_t):
4735 nradix = HAMMER2_IND_RADIX_NOM - HAMMER2_BLOCKREF_RADIX;
4737 case HAMMER2_IND_BYTES_MAX / sizeof(hammer2_blockref_t):
4738 nradix = HAMMER2_IND_RADIX_MAX - HAMMER2_BLOCKREF_RADIX;
4741 panic("bad ncount %d\n", ncount);
4747 * The largest radix that can be returned for an indirect block is
4748 * 63 bits. (The largest practical indirect block radix is actually
4749 * 62 bits because the top-level inode or volume root contains four
4750 * entries, but allow 63 to be returned).
4755 return keybits + nradix;
4761 * Directory indirect blocks.
4763 * Covers both the inode index (directory of inodes), and directory contents
4764 * (filenames hardlinked to inodes).
4766 * Because directory keys are hashed we generally try to cut the space in
4767 * half. We accomodate the inode index (which tends to have linearly
4768 * increasing inode numbers) by ensuring that the keyspace is at least large
4769 * enough to fill up the indirect block being created.
4772 hammer2_chain_indkey_dir(hammer2_chain_t *parent, hammer2_key_t *keyp,
4773 int keybits, hammer2_blockref_t *base, int count,
4776 hammer2_blockref_t *bref;
4777 hammer2_chain_t *chain;
4778 hammer2_key_t key_beg;
4779 hammer2_key_t key_end;
4780 hammer2_key_t key_next;
4785 int maxloops = 300000;
4788 * NOTE: We can't take a shortcut here anymore for inodes because
4789 * the root directory can contain a mix of inodes and directory
4790 * entries (we used to just return 63 if parent->bref.type was
4791 * HAMMER2_BREF_TYPE_INODE.
4798 * Calculate the range of keys in the array being careful to skip
4799 * slots which are overridden with a deletion.
4802 key_end = HAMMER2_KEY_MAX;
4803 hammer2_spin_ex(&parent->core.spin);
4806 if (--maxloops == 0) {
4807 panic("indkey_freemap shit %p %p:%d\n",
4808 parent, base, count);
4810 chain = hammer2_combined_find(parent, base, count,
4824 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) {
4825 if (key_next == 0 || key_next > key_end)
4832 * Use the full live (not deleted) element for the scan
4833 * iteration. HAMMER2 does not allow partial replacements.
4835 * XXX should be built into hammer2_combined_find().
4837 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
4840 * Expand our calculated key range (key, keybits) to fit
4841 * the scanned key. nkeybits represents the full range
4842 * that we will later cut in half (two halves @ nkeybits - 1).
4845 if (nkeybits < bref->keybits) {
4846 if (bref->keybits > 64) {
4847 kprintf("bad bref chain %p bref %p\n",
4851 nkeybits = bref->keybits;
4853 while (nkeybits < 64 &&
4854 rounddown2(key ^ bref->key, (hammer2_key_t)1 << nkeybits) != 0) {
4859 * If the new key range is larger we have to determine
4860 * which side of the new key range the existing keys fall
4861 * under by checking the high bit, then collapsing the
4862 * locount into the hicount or vise-versa.
4864 if (keybits != nkeybits) {
4865 if (((hammer2_key_t)1 << (nkeybits - 1)) & key) {
4876 * The newly scanned key will be in the lower half or the
4877 * upper half of the (new) key range.
4879 if (((hammer2_key_t)1 << (nkeybits - 1)) & bref->key)
4888 hammer2_spin_unex(&parent->core.spin);
4889 bref = NULL; /* now invalid (safety) */
4892 * Adjust keybits to represent half of the full range calculated
4893 * above (radix 63 max) for our new indirect block.
4898 * Expand keybits to hold at least ncount elements. ncount will be
4899 * a power of 2. This is to try to completely fill leaf nodes (at
4900 * least for keys which are not hashes).
4902 * We aren't counting 'in' or 'out', we are counting 'high side'
4903 * and 'low side' based on the bit at (1LL << keybits). We want
4904 * everything to be inside in these cases so shift it all to
4905 * the low or high side depending on the new high bit.
4907 while (((hammer2_key_t)1 << keybits) < ncount) {
4909 if (key & ((hammer2_key_t)1 << keybits)) {
4918 if (hicount > locount)
4919 key |= (hammer2_key_t)1 << keybits;
4921 key &= ~(hammer2_key_t)1 << keybits;
4931 * Directory indirect blocks.
4933 * Covers both the inode index (directory of inodes), and directory contents
4934 * (filenames hardlinked to inodes).
4936 * Because directory keys are hashed we generally try to cut the space in
4937 * half. We accomodate the inode index (which tends to have linearly
4938 * increasing inode numbers) by ensuring that the keyspace is at least large
4939 * enough to fill up the indirect block being created.
4942 hammer2_chain_indkey_dir(hammer2_chain_t *parent, hammer2_key_t *keyp,
4943 int keybits, hammer2_blockref_t *base, int count,
4946 hammer2_blockref_t *bref;
4947 hammer2_chain_t *chain;
4948 hammer2_key_t key_beg;
4949 hammer2_key_t key_end;
4950 hammer2_key_t key_next;
4955 int maxloops = 300000;
4958 * Shortcut if the parent is the inode. In this situation the
4959 * parent has 4+1 directory entries and we are creating an indirect
4960 * block capable of holding many more.
4962 if (parent->bref.type == HAMMER2_BREF_TYPE_INODE) {
4971 * Calculate the range of keys in the array being careful to skip
4972 * slots which are overridden with a deletion.
4975 key_end = HAMMER2_KEY_MAX;
4976 hammer2_spin_ex(&parent->core.spin);
4979 if (--maxloops == 0) {
4980 panic("indkey_freemap shit %p %p:%d\n",
4981 parent, base, count);
4983 chain = hammer2_combined_find(parent, base, count,
4997 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) {
4998 if (key_next == 0 || key_next > key_end)
5005 * Use the full live (not deleted) element for the scan
5006 * iteration. HAMMER2 does not allow partial replacements.
5008 * XXX should be built into hammer2_combined_find().
5010 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
5013 * Expand our calculated key range (key, keybits) to fit
5014 * the scanned key. nkeybits represents the full range
5015 * that we will later cut in half (two halves @ nkeybits - 1).
5018 if (nkeybits < bref->keybits) {
5019 if (bref->keybits > 64) {
5020 kprintf("bad bref chain %p bref %p\n",
5024 nkeybits = bref->keybits;
5026 while (nkeybits < 64 &&
5027 (~(((hammer2_key_t)1 << nkeybits) - 1) &
5028 (key ^ bref->key)) != 0) {
5033 * If the new key range is larger we have to determine
5034 * which side of the new key range the existing keys fall
5035 * under by checking the high bit, then collapsing the
5036 * locount into the hicount or vise-versa.
5038 if (keybits != nkeybits) {
5039 if (((hammer2_key_t)1 << (nkeybits - 1)) & key) {
5050 * The newly scanned key will be in the lower half or the
5051 * upper half of the (new) key range.
5053 if (((hammer2_key_t)1 << (nkeybits - 1)) & bref->key)
5062 hammer2_spin_unex(&parent->core.spin);
5063 bref = NULL; /* now invalid (safety) */
5066 * Adjust keybits to represent half of the full range calculated
5067 * above (radix 63 max) for our new indirect block.
5072 * Expand keybits to hold at least ncount elements. ncount will be
5073 * a power of 2. This is to try to completely fill leaf nodes (at
5074 * least for keys which are not hashes).
5076 * We aren't counting 'in' or 'out', we are counting 'high side'
5077 * and 'low side' based on the bit at (1LL << keybits). We want
5078 * everything to be inside in these cases so shift it all to
5079 * the low or high side depending on the new high bit.
5081 while (((hammer2_key_t)1 << keybits) < ncount) {
5083 if (key & ((hammer2_key_t)1 << keybits)) {
5092 if (hicount > locount)
5093 key |= (hammer2_key_t)1 << keybits;
5095 key &= ~(hammer2_key_t)1 << keybits;
5105 * Sets CHAIN_DELETED and remove the chain's blockref from the parent if
5108 * Both parent and chain must be locked exclusively.
5110 * This function will modify the parent if the blockref requires removal
5111 * from the parent's block table.
5113 * This function is NOT recursive. Any entity already pushed into the
5114 * chain (such as an inode) may still need visibility into its contents,
5115 * as well as the ability to read and modify the contents. For example,
5116 * for an unlinked file which is still open.
5118 * Also note that the flusher is responsible for cleaning up empty
5122 hammer2_chain_delete(hammer2_chain_t *parent, hammer2_chain_t *chain,
5123 hammer2_tid_t mtid, int flags)
5127 KKASSERT(hammer2_mtx_owned(&chain->lock));
5130 * Nothing to do if already marked.
5132 * We need the spinlock on the core whos RBTREE contains chain
5133 * to protect against races.
5135 if ((chain->flags & HAMMER2_CHAIN_DELETED) == 0) {
5136 KKASSERT((chain->flags & HAMMER2_CHAIN_DELETED) == 0 &&
5137 chain->parent == parent);
5138 error = _hammer2_chain_delete_helper(parent, chain,
5143 * Permanent deletions mark the chain as destroyed.
5145 * NOTE: We do not setflush the chain unless the deletion is
5146 * permanent, since the deletion of a chain does not actually
5147 * require it to be flushed.
5150 if (flags & HAMMER2_DELETE_PERMANENT) {
5151 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DESTROY);
5152 hammer2_chain_setflush(chain);
5160 hammer2_chain_delete_obref(hammer2_chain_t *parent, hammer2_chain_t *chain,
5161 hammer2_tid_t mtid, int flags,
5162 hammer2_blockref_t *obref)
5166 KKASSERT(hammer2_mtx_owned(&chain->lock));
5169 * Nothing to do if already marked.
5171 * We need the spinlock on the core whos RBTREE contains chain
5172 * to protect against races.
5174 obref->type = HAMMER2_BREF_TYPE_EMPTY;
5175 if ((chain->flags & HAMMER2_CHAIN_DELETED) == 0) {
5176 KKASSERT((chain->flags & HAMMER2_CHAIN_DELETED) == 0 &&
5177 chain->parent == parent);
5178 error = _hammer2_chain_delete_helper(parent, chain,
5179 mtid, flags, obref);
5183 * Permanent deletions mark the chain as destroyed.
5185 * NOTE: We do not setflush the chain unless the deletion is
5186 * permanent, since the deletion of a chain does not actually
5187 * require it to be flushed.
5190 if (flags & HAMMER2_DELETE_PERMANENT) {
5191 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DESTROY);
5192 hammer2_chain_setflush(chain);
5200 * Returns the index of the nearest element in the blockref array >= elm.
5201 * Returns (count) if no element could be found.
5203 * Sets *key_nextp to the next key for loop purposes but does not modify
5204 * it if the next key would be higher than the current value of *key_nextp.
5205 * Note that *key_nexp can overflow to 0, which should be tested by the
5208 * WARNING! Must be called with parent's spinlock held. Spinlock remains
5209 * held through the operation.
5212 hammer2_base_find(hammer2_chain_t *parent,
5213 hammer2_blockref_t *base, int count,
5214 hammer2_key_t *key_nextp,
5215 hammer2_key_t key_beg, hammer2_key_t key_end)
5217 hammer2_blockref_t *scan;
5218 hammer2_key_t scan_end;
5223 * Require the live chain's already have their core's counted
5224 * so we can optimize operations.
5226 KKASSERT(parent->flags & HAMMER2_CHAIN_COUNTEDBREFS);
5231 if (count == 0 || base == NULL)
5235 * Sequential optimization using parent->cache_index. This is
5236 * the most likely scenario.
5238 * We can avoid trailing empty entries on live chains, otherwise
5239 * we might have to check the whole block array.
5241 i = parent->cache_index; /* SMP RACE OK */
5243 limit = parent->core.live_zero;
5248 KKASSERT(i < count);
5254 while (i > 0 && (scan->type == HAMMER2_BREF_TYPE_EMPTY ||
5255 scan->key > key_beg)) {
5259 parent->cache_index = i;
5262 * Search forwards, stop when we find a scan element which
5263 * encloses the key or until we know that there are no further
5267 if (scan->type != HAMMER2_BREF_TYPE_EMPTY) {
5268 scan_end = scan->key +
5269 ((hammer2_key_t)1 << scan->keybits) - 1;
5270 if (scan->key > key_beg || scan_end >= key_beg)
5279 parent->cache_index = i;
5283 scan_end = scan->key +
5284 ((hammer2_key_t)1 << scan->keybits);
5285 if (scan_end && (*key_nextp > scan_end ||
5287 *key_nextp = scan_end;
5295 * Do a combined search and return the next match either from the blockref
5296 * array or from the in-memory chain. Sets *bresp to the returned bref in
5297 * both cases, or sets it to NULL if the search exhausted. Only returns
5298 * a non-NULL chain if the search matched from the in-memory chain.
5300 * When no in-memory chain has been found and a non-NULL bref is returned
5304 * The returned chain is not locked or referenced. Use the returned bref
5305 * to determine if the search exhausted or not. Iterate if the base find
5306 * is chosen but matches a deleted chain.
5308 * WARNING! Must be called with parent's spinlock held. Spinlock remains
5309 * held through the operation.
5312 hammer2_combined_find(hammer2_chain_t *parent,
5313 hammer2_blockref_t *base, int count,
5314 hammer2_key_t *key_nextp,
5315 hammer2_key_t key_beg, hammer2_key_t key_end,
5316 hammer2_blockref_t **bresp)
5318 hammer2_blockref_t *bref;
5319 hammer2_chain_t *chain;
5323 * Lookup in block array and in rbtree.
5325 *key_nextp = key_end + 1;
5326 i = hammer2_base_find(parent, base, count, key_nextp,
5328 chain = hammer2_chain_find(parent, key_nextp, key_beg, key_end);
5333 if (i == count && chain == NULL) {
5339 * Only chain matched.
5342 bref = &chain->bref;
5347 * Only blockref matched.
5349 if (chain == NULL) {
5355 * Both in-memory and blockref matched, select the nearer element.
5357 * If both are flush with the left-hand side or both are the
5358 * same distance away, select the chain. In this situation the
5359 * chain must have been loaded from the matching blockmap.
5361 if ((chain->bref.key <= key_beg && base[i].key <= key_beg) ||
5362 chain->bref.key == base[i].key) {
5363 KKASSERT(chain->bref.key == base[i].key);
5364 bref = &chain->bref;
5369 * Select the nearer key
5371 if (chain->bref.key < base[i].key) {
5372 bref = &chain->bref;
5379 * If the bref is out of bounds we've exhausted our search.
5382 if (bref->key > key_end) {
5392 * Locate the specified block array element and delete it. The element
5395 * The spin lock on the related chain must be held.
5397 * NOTE: live_count was adjusted when the chain was deleted, so it does not
5398 * need to be adjusted when we commit the media change.
5401 hammer2_base_delete(hammer2_chain_t *parent,
5402 hammer2_blockref_t *base, int count,
5403 hammer2_chain_t *chain,
5404 hammer2_blockref_t *obref)
5406 hammer2_blockref_t *elm = &chain->bref;
5407 hammer2_blockref_t *scan;
5408 hammer2_key_t key_next;
5412 * Delete element. Expect the element to exist.
5414 * XXX see caller, flush code not yet sophisticated enough to prevent
5415 * re-flushed in some cases.
5417 key_next = 0; /* max range */
5418 i = hammer2_base_find(parent, base, count, &key_next,
5419 elm->key, elm->key);
5421 if (i == count || scan->type == HAMMER2_BREF_TYPE_EMPTY ||
5422 scan->key != elm->key ||
5423 ((chain->flags & HAMMER2_CHAIN_BMAPUPD) == 0 &&
5424 scan->keybits != elm->keybits)) {
5425 hammer2_spin_unex(&parent->core.spin);
5426 panic("delete base %p element not found at %d/%d elm %p\n",
5427 base, i, count, elm);
5432 * Update stats and zero the entry.
5434 * NOTE: Handle radix == 0 (0 bytes) case.
5436 if ((int)(scan->data_off & HAMMER2_OFF_MASK_RADIX)) {
5437 parent->bref.embed.stats.data_count -= (hammer2_off_t)1 <<
5438 (int)(scan->data_off & HAMMER2_OFF_MASK_RADIX);
5440 switch(scan->type) {
5441 case HAMMER2_BREF_TYPE_INODE:
5442 --parent->bref.embed.stats.inode_count;
5444 case HAMMER2_BREF_TYPE_DATA:
5445 if (parent->bref.leaf_count == HAMMER2_BLOCKREF_LEAF_MAX) {
5446 atomic_set_int(&chain->flags,
5447 HAMMER2_CHAIN_HINT_LEAF_COUNT);
5449 if (parent->bref.leaf_count)
5450 --parent->bref.leaf_count;
5453 case HAMMER2_BREF_TYPE_INDIRECT:
5454 if (scan->type != HAMMER2_BREF_TYPE_DATA) {
5455 parent->bref.embed.stats.data_count -=
5456 scan->embed.stats.data_count;
5457 parent->bref.embed.stats.inode_count -=
5458 scan->embed.stats.inode_count;
5460 if (scan->type == HAMMER2_BREF_TYPE_INODE)
5462 if (parent->bref.leaf_count == HAMMER2_BLOCKREF_LEAF_MAX) {
5463 atomic_set_int(&chain->flags,
5464 HAMMER2_CHAIN_HINT_LEAF_COUNT);
5466 if (parent->bref.leaf_count <= scan->leaf_count)
5467 parent->bref.leaf_count = 0;
5469 parent->bref.leaf_count -= scan->leaf_count;
5472 case HAMMER2_BREF_TYPE_DIRENT:
5473 if (parent->bref.leaf_count == HAMMER2_BLOCKREF_LEAF_MAX) {
5474 atomic_set_int(&chain->flags,
5475 HAMMER2_CHAIN_HINT_LEAF_COUNT);
5477 if (parent->bref.leaf_count)
5478 --parent->bref.leaf_count;
5486 bzero(scan, sizeof(*scan));
5489 * We can only optimize parent->core.live_zero for live chains.
5491 if (parent->core.live_zero == i + 1) {
5492 while (--i >= 0 && base[i].type == HAMMER2_BREF_TYPE_EMPTY)
5494 parent->core.live_zero = i + 1;
5498 * Clear appropriate blockmap flags in chain.
5500 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_BMAPPED |
5501 HAMMER2_CHAIN_BMAPUPD);
5505 * Insert the specified element. The block array must not already have the
5506 * element and must have space available for the insertion.
5508 * The spin lock on the related chain must be held.
5510 * NOTE: live_count was adjusted when the chain was deleted, so it does not
5511 * need to be adjusted when we commit the media change.
5514 hammer2_base_insert(hammer2_chain_t *parent,
5515 hammer2_blockref_t *base, int count,
5516 hammer2_chain_t *chain, hammer2_blockref_t *elm)
5518 hammer2_key_t key_next;
5527 * Insert new element. Expect the element to not already exist
5528 * unless we are replacing it.
5530 * XXX see caller, flush code not yet sophisticated enough to prevent
5531 * re-flushed in some cases.
5533 key_next = 0; /* max range */
5534 i = hammer2_base_find(parent, base, count, &key_next,
5535 elm->key, elm->key);
5538 * Shortcut fill optimization, typical ordered insertion(s) may not
5541 KKASSERT(i >= 0 && i <= count);
5544 * Set appropriate blockmap flags in chain (if not NULL)
5547 atomic_set_int(&chain->flags, HAMMER2_CHAIN_BMAPPED);
5550 * Update stats and zero the entry
5552 if ((int)(elm->data_off & HAMMER2_OFF_MASK_RADIX)) {
5553 parent->bref.embed.stats.data_count += (hammer2_off_t)1 <<
5554 (int)(elm->data_off & HAMMER2_OFF_MASK_RADIX);
5557 case HAMMER2_BREF_TYPE_INODE:
5558 ++parent->bref.embed.stats.inode_count;
5560 case HAMMER2_BREF_TYPE_DATA:
5561 if (parent->bref.leaf_count != HAMMER2_BLOCKREF_LEAF_MAX)
5562 ++parent->bref.leaf_count;
5564 case HAMMER2_BREF_TYPE_INDIRECT:
5565 if (elm->type != HAMMER2_BREF_TYPE_DATA) {
5566 parent->bref.embed.stats.data_count +=
5567 elm->embed.stats.data_count;
5568 parent->bref.embed.stats.inode_count +=
5569 elm->embed.stats.inode_count;
5571 if (elm->type == HAMMER2_BREF_TYPE_INODE)
5573 if (parent->bref.leaf_count + elm->leaf_count <
5574 HAMMER2_BLOCKREF_LEAF_MAX) {
5575 parent->bref.leaf_count += elm->leaf_count;
5577 parent->bref.leaf_count = HAMMER2_BLOCKREF_LEAF_MAX;
5580 case HAMMER2_BREF_TYPE_DIRENT:
5581 if (parent->bref.leaf_count != HAMMER2_BLOCKREF_LEAF_MAX)
5582 ++parent->bref.leaf_count;
5590 * We can only optimize parent->core.live_zero for live chains.
5592 if (i == count && parent->core.live_zero < count) {
5593 i = parent->core.live_zero++;
5598 xkey = elm->key + ((hammer2_key_t)1 << elm->keybits) - 1;
5599 if (i != count && (base[i].key < elm->key || xkey >= base[i].key)) {
5600 hammer2_spin_unex(&parent->core.spin);
5601 panic("insert base %p overlapping elements at %d elm %p\n",
5606 * Try to find an empty slot before or after.
5610 while (j > 0 || k < count) {
5612 if (j >= 0 && base[j].type == HAMMER2_BREF_TYPE_EMPTY) {
5616 bcopy(&base[j+1], &base[j],
5617 (i - j - 1) * sizeof(*base));
5623 if (k < count && base[k].type == HAMMER2_BREF_TYPE_EMPTY) {
5624 bcopy(&base[i], &base[i+1],
5625 (k - i) * sizeof(hammer2_blockref_t));
5629 * We can only update parent->core.live_zero for live
5632 if (parent->core.live_zero <= k)
5633 parent->core.live_zero = k + 1;
5638 panic("hammer2_base_insert: no room!");
5645 for (l = 0; l < count; ++l) {
5646 if (base[l].type != HAMMER2_BREF_TYPE_EMPTY) {
5647 key_next = base[l].key +
5648 ((hammer2_key_t)1 << base[l].keybits) - 1;
5652 while (++l < count) {
5653 if (base[l].type != HAMMER2_BREF_TYPE_EMPTY) {
5654 if (base[l].key <= key_next)
5655 panic("base_insert %d %d,%d,%d fail %p:%d", u, i, j, k, base, l);
5656 key_next = base[l].key +
5657 ((hammer2_key_t)1 << base[l].keybits) - 1;
5667 * Sort the blockref array for the chain. Used by the flush code to
5668 * sort the blockref[] array.
5670 * The chain must be exclusively locked AND spin-locked.
5672 typedef hammer2_blockref_t *hammer2_blockref_p;
5676 hammer2_base_sort_callback(const void *v1, const void *v2)
5678 hammer2_blockref_p bref1 = *(const hammer2_blockref_p *)v1;
5679 hammer2_blockref_p bref2 = *(const hammer2_blockref_p *)v2;
5682 * Make sure empty elements are placed at the end of the array
5684 if (bref1->type == HAMMER2_BREF_TYPE_EMPTY) {
5685 if (bref2->type == HAMMER2_BREF_TYPE_EMPTY)
5688 } else if (bref2->type == HAMMER2_BREF_TYPE_EMPTY) {
5695 if (bref1->key < bref2->key)
5697 if (bref1->key > bref2->key)
5703 hammer2_base_sort(hammer2_chain_t *chain)
5705 hammer2_blockref_t *base;
5708 switch(chain->bref.type) {
5709 case HAMMER2_BREF_TYPE_INODE:
5711 * Special shortcut for embedded data returns the inode
5712 * itself. Callers must detect this condition and access
5713 * the embedded data (the strategy code does this for us).
5715 * This is only applicable to regular files and softlinks.
5717 if (chain->data->ipdata.meta.op_flags &
5718 HAMMER2_OPFLAG_DIRECTDATA) {
5721 base = &chain->data->ipdata.u.blockset.blockref[0];
5722 count = HAMMER2_SET_COUNT;
5724 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
5725 case HAMMER2_BREF_TYPE_INDIRECT:
5727 * Optimize indirect blocks in the INITIAL state to avoid
5730 KKASSERT((chain->flags & HAMMER2_CHAIN_INITIAL) == 0);
5731 base = &chain->data->npdata[0];
5732 count = chain->bytes / sizeof(hammer2_blockref_t);
5734 case HAMMER2_BREF_TYPE_VOLUME:
5735 base = &chain->data->voldata.sroot_blockset.blockref[0];
5736 count = HAMMER2_SET_COUNT;
5738 case HAMMER2_BREF_TYPE_FREEMAP:
5739 base = &chain->data->blkset.blockref[0];
5740 count = HAMMER2_SET_COUNT;
5743 panic("hammer2_base_sort: unrecognized "
5744 "blockref(A) type: %d",
5746 base = NULL; /* safety */
5747 count = 0; /* safety */
5750 kqsort(base, count, sizeof(*base), hammer2_base_sort_callback);
5756 * Chain memory management
5758 const hammer2_media_data_t *
5759 hammer2_chain_rdata(hammer2_chain_t *chain)
5761 KKASSERT(chain->data != NULL);
5762 return (chain->data);
5765 hammer2_media_data_t *
5766 hammer2_chain_wdata(hammer2_chain_t *chain)
5768 KKASSERT(chain->data != NULL);
5769 return (chain->data);
5773 * Set the check data for a chain. This can be a heavy-weight operation
5774 * and typically only runs on-flush. For file data check data is calculated
5775 * when the logical buffers are flushed.
5778 hammer2_chain_setcheck(hammer2_chain_t *chain, void *bdata)
5780 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_NOTTESTED);
5782 switch(HAMMER2_DEC_CHECK(chain->bref.methods)) {
5783 case HAMMER2_CHECK_NONE:
5785 case HAMMER2_CHECK_DISABLED:
5787 case HAMMER2_CHECK_ISCSI32:
5788 chain->bref.check.iscsi32.value =
5789 hammer2_icrc32(bdata, chain->bytes);
5791 case HAMMER2_CHECK_XXHASH64:
5792 chain->bref.check.xxhash64.value =
5793 XXH64(bdata, chain->bytes, XXH_HAMMER2_SEED);
5795 case HAMMER2_CHECK_SHA192:
5797 SHA256_CTX hash_ctx;
5799 uint8_t digest[SHA256_DIGEST_LENGTH];
5800 uint64_t digest64[SHA256_DIGEST_LENGTH/8];
5803 SHA256_Init(&hash_ctx);
5804 SHA256_Update(&hash_ctx, bdata, chain->bytes);
5805 SHA256_Final(u.digest, &hash_ctx);
5806 u.digest64[2] ^= u.digest64[3];
5808 chain->bref.check.sha192.data,
5809 sizeof(chain->bref.check.sha192.data));
5812 case HAMMER2_CHECK_FREEMAP:
5813 chain->bref.check.freemap.icrc32 =
5814 hammer2_icrc32(bdata, chain->bytes);
5817 kprintf("hammer2_chain_setcheck: unknown check type %02x\n",
5818 chain->bref.methods);
5824 * Characterize a failed check code and try to trace back to the inode.
5827 hammer2_characterize_failed_chain(hammer2_chain_t *chain, uint64_t check,
5830 hammer2_chain_t *lchain;
5831 hammer2_chain_t *ochain;
5834 did = krateprintf(&krate_h2chk,
5835 "chain %016jx.%02x (%s) meth=%02x CHECK FAIL "
5836 "(flags=%08x, bref/data ",
5837 chain->bref.data_off,
5839 hammer2_bref_type_str(chain->bref.type),
5840 chain->bref.methods,
5846 kprintf("%08x/%08x)\n",
5847 chain->bref.check.iscsi32.value,
5850 kprintf("%016jx/%016jx)\n",
5851 chain->bref.check.xxhash64.value,
5856 * Run up the chains to try to find the governing inode so we
5859 * XXX This error reporting is not really MPSAFE
5863 while (chain && chain->bref.type != HAMMER2_BREF_TYPE_INODE) {
5865 chain = chain->parent;
5868 if (chain && chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
5869 ((chain->bref.flags & HAMMER2_BREF_FLAG_PFSROOT) == 0 ||
5870 (lchain->bref.key & HAMMER2_DIRHASH_VISIBLE))) {
5871 kprintf(" Resides at/in inode %ld\n",
5873 } else if (chain && chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
5874 kprintf(" Resides in inode index - CRITICAL!!!\n");
5876 kprintf(" Resides in root index - CRITICAL!!!\n");
5879 const char *pfsname = "UNKNOWN";
5883 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
5884 if (ochain->pmp->pfs_hmps[i] == ochain->hmp &&
5885 ochain->pmp->pfs_names[i]) {
5886 pfsname = ochain->pmp->pfs_names[i];
5891 kprintf(" In pfs %s on device %s\n",
5892 pfsname, ochain->hmp->devrepname);
5897 * Returns non-zero on success, 0 on failure.
5900 hammer2_chain_testcheck(hammer2_chain_t *chain, void *bdata)
5906 if (chain->flags & HAMMER2_CHAIN_NOTTESTED)
5909 switch(HAMMER2_DEC_CHECK(chain->bref.methods)) {
5910 case HAMMER2_CHECK_NONE:
5913 case HAMMER2_CHECK_DISABLED:
5916 case HAMMER2_CHECK_ISCSI32:
5917 check32 = hammer2_icrc32(bdata, chain->bytes);
5918 r = (chain->bref.check.iscsi32.value == check32);
5920 hammer2_characterize_failed_chain(chain, check32, 32);
5922 hammer2_process_icrc32 += chain->bytes;
5924 case HAMMER2_CHECK_XXHASH64:
5925 check64 = XXH64(bdata, chain->bytes, XXH_HAMMER2_SEED);
5926 r = (chain->bref.check.xxhash64.value == check64);
5928 hammer2_characterize_failed_chain(chain, check64, 64);
5930 hammer2_process_xxhash64 += chain->bytes;
5932 case HAMMER2_CHECK_SHA192:
5934 SHA256_CTX hash_ctx;
5936 uint8_t digest[SHA256_DIGEST_LENGTH];
5937 uint64_t digest64[SHA256_DIGEST_LENGTH/8];
5940 SHA256_Init(&hash_ctx);
5941 SHA256_Update(&hash_ctx, bdata, chain->bytes);
5942 SHA256_Final(u.digest, &hash_ctx);
5943 u.digest64[2] ^= u.digest64[3];
5945 chain->bref.check.sha192.data,
5946 sizeof(chain->bref.check.sha192.data)) == 0) {
5950 krateprintf(&krate_h2chk,
5951 "chain %016jx.%02x meth=%02x "
5953 chain->bref.data_off,
5955 chain->bref.methods);
5959 case HAMMER2_CHECK_FREEMAP:
5960 r = (chain->bref.check.freemap.icrc32 ==
5961 hammer2_icrc32(bdata, chain->bytes));
5965 did = krateprintf(&krate_h2chk,
5966 "chain %016jx.%02x meth=%02x "
5968 chain->bref.data_off,
5970 chain->bref.methods);
5972 kprintf("freemap.icrc %08x icrc32 %08x (%d)\n",
5973 chain->bref.check.freemap.icrc32,
5974 hammer2_icrc32(bdata, chain->bytes),
5977 kprintf("dio %p buf %016jx,%d "
5980 chain->dio->bp->b_loffset,
5981 chain->dio->bp->b_bufsize,
5983 chain->dio->bp->b_data);
5989 kprintf("hammer2_chain_testcheck: unknown check type %02x\n",
5990 chain->bref.methods);
5998 * Acquire the chain and parent representing the specified inode for the
5999 * device at the specified cluster index.
6001 * The flags passed in are LOOKUP flags, not RESOLVE flags.
6003 * If we are unable to locate the inode, HAMMER2_ERROR_EIO is returned and
6004 * *chainp will be NULL. *parentp may still be set error or not, or NULL
6005 * if the parent itself could not be resolved.
6007 * The caller may pass-in a locked *parentp and/or *chainp, or neither.
6008 * They will be unlocked and released by this function. The *parentp and
6009 * *chainp representing the located inode are returned locked.
6012 hammer2_chain_inode_find(hammer2_pfs_t *pmp, hammer2_key_t inum,
6013 int clindex, int flags,
6014 hammer2_chain_t **parentp, hammer2_chain_t **chainp)
6016 hammer2_chain_t *parent;
6017 hammer2_chain_t *rchain;
6018 hammer2_key_t key_dummy;
6019 hammer2_inode_t *ip;
6023 resolve_flags = (flags & HAMMER2_LOOKUP_SHARED) ?
6024 HAMMER2_RESOLVE_SHARED : 0;
6027 * Caller expects us to replace these.
6030 hammer2_chain_unlock(*chainp);
6031 hammer2_chain_drop(*chainp);
6035 hammer2_chain_unlock(*parentp);
6036 hammer2_chain_drop(*parentp);
6041 * Be very careful, this is a backend function and we CANNOT
6042 * lock any frontend inode structure we find. But we have to
6043 * look the inode up this way first in case it exists but is
6044 * detached from the radix tree.
6046 ip = hammer2_inode_lookup(pmp, inum);
6048 *chainp = hammer2_inode_chain_and_parent(ip, clindex,
6051 hammer2_inode_drop(ip);
6054 hammer2_chain_unlock(*chainp);
6055 hammer2_chain_drop(*chainp);
6058 hammer2_chain_unlock(*parentp);
6059 hammer2_chain_drop(*parentp);
6065 * Inodes hang off of the iroot (bit 63 is clear, differentiating
6066 * inodes from root directory entries in the key lookup).
6068 parent = hammer2_inode_chain(pmp->iroot, clindex, resolve_flags);
6071 rchain = hammer2_chain_lookup(&parent, &key_dummy,
6075 error = HAMMER2_ERROR_EIO;
6084 * Used by the bulkscan code to snapshot the synchronized storage for
6085 * a volume, allowing it to be scanned concurrently against normal
6089 hammer2_chain_bulksnap(hammer2_dev_t *hmp)
6091 hammer2_chain_t *copy;
6093 copy = hammer2_chain_alloc(hmp, hmp->spmp, &hmp->vchain.bref);
6094 copy->data = kmalloc(sizeof(copy->data->voldata),
6097 hammer2_voldata_lock(hmp);
6098 copy->data->voldata = hmp->volsync;
6099 hammer2_voldata_unlock(hmp);
6105 hammer2_chain_bulkdrop(hammer2_chain_t *copy)
6107 KKASSERT(copy->bref.type == HAMMER2_BREF_TYPE_VOLUME);
6108 KKASSERT(copy->data);
6109 kfree(copy->data, copy->hmp->mchain);
6111 atomic_add_long(&hammer2_chain_allocs, -1);
6112 hammer2_chain_drop(copy);
6116 * Returns non-zero if the chain (INODE or DIRENT) matches the
6120 hammer2_chain_dirent_test(hammer2_chain_t *chain, const char *name,
6123 const hammer2_inode_data_t *ripdata;
6124 const hammer2_dirent_head_t *den;
6126 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
6127 ripdata = &chain->data->ipdata;
6128 if (ripdata->meta.name_len == name_len &&
6129 bcmp(ripdata->filename, name, name_len) == 0) {
6133 if (chain->bref.type == HAMMER2_BREF_TYPE_DIRENT &&
6134 chain->bref.embed.dirent.namlen == name_len) {
6135 den = &chain->bref.embed.dirent;
6136 if (name_len > sizeof(chain->bref.check.buf) &&
6137 bcmp(chain->data->buf, name, name_len) == 0) {
6140 if (name_len <= sizeof(chain->bref.check.buf) &&
6141 bcmp(chain->bref.check.buf, name, name_len) == 0) {