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);
197 atomic_add_long(&hammer2_chain_allocs, 1);
200 * Construct the appropriate system structure.
203 case HAMMER2_BREF_TYPE_DIRENT:
204 case HAMMER2_BREF_TYPE_INODE:
205 case HAMMER2_BREF_TYPE_INDIRECT:
206 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
207 case HAMMER2_BREF_TYPE_DATA:
208 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
209 chain = kmalloc(sizeof(*chain), hmp->mchain, M_WAITOK | M_ZERO);
211 case HAMMER2_BREF_TYPE_VOLUME:
212 case HAMMER2_BREF_TYPE_FREEMAP:
214 * Only hammer2_chain_bulksnap() calls this function with these
217 chain = kmalloc(sizeof(*chain), hmp->mchain, M_WAITOK | M_ZERO);
221 panic("hammer2_chain_alloc: unrecognized blockref type: %d",
226 * Initialize the new chain structure. pmp must be set to NULL for
227 * chains belonging to the super-root topology of a device mount.
229 if (pmp == hmp->spmp)
236 chain->bytes = bytes;
238 chain->flags = HAMMER2_CHAIN_ALLOCATED;
239 lockinit(&chain->diolk, "chdio", 0, 0);
242 * Set the PFS boundary flag if this chain represents a PFS root.
244 if (bref->flags & HAMMER2_BREF_FLAG_PFSROOT)
245 atomic_set_int(&chain->flags, HAMMER2_CHAIN_PFSBOUNDARY);
246 hammer2_chain_core_init(chain);
252 * Initialize a chain's core structure. This structure used to be allocated
253 * but is now embedded.
255 * The core is not locked. No additional refs on the chain are made.
256 * (trans) must not be NULL if (core) is not NULL.
259 hammer2_chain_core_init(hammer2_chain_t *chain)
262 * Fresh core under nchain (no multi-homing of ochain's
265 RB_INIT(&chain->core.rbtree); /* live chains */
266 hammer2_mtx_init(&chain->lock, "h2chain");
270 * Add a reference to a chain element, preventing its destruction.
272 * (can be called with spinlock held)
275 hammer2_chain_ref(hammer2_chain_t *chain)
277 if (atomic_fetchadd_int(&chain->refs, 1) == 0) {
279 * Just flag that the chain was used and should be recycled
280 * on the LRU if it encounters it later.
282 if (chain->flags & HAMMER2_CHAIN_ONLRU)
283 atomic_set_int(&chain->flags, HAMMER2_CHAIN_LRUHINT);
287 * REMOVED - reduces contention, lru_list is more heuristical
290 * 0->non-zero transition must ensure that chain is removed
293 * NOTE: Already holding lru_spin here so we cannot call
294 * hammer2_chain_ref() to get it off lru_list, do
297 if (chain->flags & HAMMER2_CHAIN_ONLRU) {
298 hammer2_pfs_t *pmp = chain->pmp;
299 hammer2_spin_ex(&pmp->lru_spin);
300 if (chain->flags & HAMMER2_CHAIN_ONLRU) {
301 atomic_add_int(&pmp->lru_count, -1);
302 atomic_clear_int(&chain->flags,
303 HAMMER2_CHAIN_ONLRU);
304 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node);
306 hammer2_spin_unex(&pmp->lru_spin);
313 * Ref a locked chain and force the data to be held across an unlock.
314 * Chain must be currently locked. The user of the chain who desires
315 * to release the hold must call hammer2_chain_lock_unhold() to relock
316 * and unhold the chain, then unlock normally, or may simply call
317 * hammer2_chain_drop_unhold() (which is safer against deadlocks).
320 hammer2_chain_ref_hold(hammer2_chain_t *chain)
322 atomic_add_int(&chain->lockcnt, 1);
323 hammer2_chain_ref(chain);
327 * Insert the chain in the core rbtree.
329 * Normal insertions are placed in the live rbtree. Insertion of a deleted
330 * chain is a special case used by the flush code that is placed on the
331 * unstaged deleted list to avoid confusing the live view.
333 #define HAMMER2_CHAIN_INSERT_SPIN 0x0001
334 #define HAMMER2_CHAIN_INSERT_LIVE 0x0002
335 #define HAMMER2_CHAIN_INSERT_RACE 0x0004
339 hammer2_chain_insert(hammer2_chain_t *parent, hammer2_chain_t *chain,
340 int flags, int generation)
342 hammer2_chain_t *xchain;
345 if (flags & HAMMER2_CHAIN_INSERT_SPIN)
346 hammer2_spin_ex(&parent->core.spin);
349 * Interlocked by spinlock, check for race
351 if ((flags & HAMMER2_CHAIN_INSERT_RACE) &&
352 parent->core.generation != generation) {
353 error = HAMMER2_ERROR_EAGAIN;
360 xchain = RB_INSERT(hammer2_chain_tree, &parent->core.rbtree, chain);
361 KASSERT(xchain == NULL,
362 ("hammer2_chain_insert: collision %p %p (key=%016jx)",
363 chain, xchain, chain->bref.key));
364 atomic_set_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE);
365 chain->parent = parent;
366 ++parent->core.chain_count;
367 ++parent->core.generation; /* XXX incs for _get() too, XXX */
370 * We have to keep track of the effective live-view blockref count
371 * so the create code knows when to push an indirect block.
373 if (flags & HAMMER2_CHAIN_INSERT_LIVE)
374 atomic_add_int(&parent->core.live_count, 1);
376 if (flags & HAMMER2_CHAIN_INSERT_SPIN)
377 hammer2_spin_unex(&parent->core.spin);
382 * Drop the caller's reference to the chain. When the ref count drops to
383 * zero this function will try to disassociate the chain from its parent and
384 * deallocate it, then recursely drop the parent using the implied ref
385 * from the chain's chain->parent.
387 * Nobody should own chain's mutex on the 1->0 transition, unless this drop
388 * races an acquisition by another cpu. Therefore we can loop if we are
389 * unable to acquire the mutex, and refs is unlikely to be 1 unless we again
390 * race against another drop.
393 hammer2_chain_drop(hammer2_chain_t *chain)
397 KKASSERT(chain->refs > 0);
405 if (hammer2_mtx_ex_try(&chain->lock) == 0)
406 chain = hammer2_chain_lastdrop(chain, 0);
407 /* retry the same chain, or chain from lastdrop */
409 if (atomic_cmpset_int(&chain->refs, refs, refs - 1))
411 /* retry the same chain */
418 * Unhold a held and probably not-locked chain, ensure that the data is
419 * dropped on the 1->0 transition of lockcnt by obtaining an exclusive
420 * lock and then simply unlocking the chain.
423 hammer2_chain_unhold(hammer2_chain_t *chain)
429 lockcnt = chain->lockcnt;
432 if (atomic_cmpset_int(&chain->lockcnt,
433 lockcnt, lockcnt - 1)) {
436 } else if (hammer2_mtx_ex_try(&chain->lock) == 0) {
437 hammer2_chain_unlock(chain);
441 * This situation can easily occur on SMP due to
442 * the gap inbetween the 1->0 transition and the
443 * final unlock. We cannot safely block on the
444 * mutex because lockcnt might go above 1.
446 * XXX Sleep for one tick if it takes too long.
449 if (iter > 1000 + hz) {
450 kprintf("hammer2: h2race1 %p\n", chain);
453 tsleep(&iter, 0, "h2race1", 1);
461 hammer2_chain_drop_unhold(hammer2_chain_t *chain)
463 hammer2_chain_unhold(chain);
464 hammer2_chain_drop(chain);
468 hammer2_chain_rehold(hammer2_chain_t *chain)
470 hammer2_chain_lock(chain, HAMMER2_RESOLVE_SHARED);
471 atomic_add_int(&chain->lockcnt, 1);
472 hammer2_chain_unlock(chain);
476 * Handles the (potential) last drop of chain->refs from 1->0. Called with
477 * the mutex exclusively locked, refs == 1, and lockcnt 0. SMP races are
478 * possible against refs and lockcnt. We must dispose of the mutex on chain.
480 * This function returns an unlocked chain for recursive drop or NULL. It
481 * can return the same chain if it determines it has raced another ref.
485 * When two chains need to be recursively dropped we use the chain we
486 * would otherwise free to placehold the additional chain. It's a bit
487 * convoluted but we can't just recurse without potentially blowing out
490 * The chain cannot be freed if it has any children.
491 * The chain cannot be freed if flagged MODIFIED unless we can dispose of it.
492 * The chain cannot be freed if flagged UPDATE unless we can dispose of it.
493 * Any dedup registration can remain intact.
495 * The core spinlock is allowed to nest child-to-parent (not parent-to-child).
499 hammer2_chain_lastdrop(hammer2_chain_t *chain, int depth)
503 hammer2_chain_t *parent;
504 hammer2_chain_t *rdrop;
507 * We need chain's spinlock to interlock the sub-tree test.
508 * We already have chain's mutex, protecting chain->parent.
510 * Remember that chain->refs can be in flux.
512 hammer2_spin_ex(&chain->core.spin);
514 if (chain->parent != NULL) {
516 * If the chain has a parent the UPDATE bit prevents scrapping
517 * as the chain is needed to properly flush the parent. Try
518 * to complete the 1->0 transition and return NULL. Retry
519 * (return chain) if we are unable to complete the 1->0
520 * transition, else return NULL (nothing more to do).
522 * If the chain has a parent the MODIFIED bit prevents
525 * Chains with UPDATE/MODIFIED are *not* put on the LRU list!
527 if (chain->flags & (HAMMER2_CHAIN_UPDATE |
528 HAMMER2_CHAIN_MODIFIED)) {
529 if (atomic_cmpset_int(&chain->refs, 1, 0)) {
530 hammer2_spin_unex(&chain->core.spin);
531 hammer2_chain_assert_no_data(chain);
532 hammer2_mtx_unlock(&chain->lock);
535 hammer2_spin_unex(&chain->core.spin);
536 hammer2_mtx_unlock(&chain->lock);
540 /* spinlock still held */
541 } else if (chain->bref.type == HAMMER2_BREF_TYPE_VOLUME ||
542 chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP) {
544 * Retain the static vchain and fchain. Clear bits that
545 * are not relevant. Do not clear the MODIFIED bit,
546 * and certainly do not put it on the delayed-flush queue.
548 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
551 * The chain has no parent and can be flagged for destruction.
552 * Since it has no parent, UPDATE can also be cleared.
554 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DESTROY);
555 if (chain->flags & HAMMER2_CHAIN_UPDATE)
556 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
559 * If the chain has children we must propagate the DESTROY
560 * flag downward and rip the disconnected topology apart.
561 * This is accomplished by calling hammer2_flush() on the
564 * Any dedup is already handled by the underlying DIO, so
565 * we do not have to specifically flush it here.
567 if (chain->core.chain_count) {
568 hammer2_spin_unex(&chain->core.spin);
569 hammer2_flush(chain, HAMMER2_FLUSH_TOP |
571 hammer2_mtx_unlock(&chain->lock);
573 return(chain); /* retry drop */
577 * Otherwise we can scrap the MODIFIED bit if it is set,
578 * and continue along the freeing path.
580 * Be sure to clean-out any dedup bits. Without a parent
581 * this chain will no longer be visible to the flush code.
582 * Easy check data_off to avoid the volume root.
584 if (chain->flags & HAMMER2_CHAIN_MODIFIED) {
585 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
586 atomic_add_long(&hammer2_count_modified_chains, -1);
588 hammer2_pfs_memory_wakeup(chain->pmp, -1);
590 /* spinlock still held */
593 /* spinlock still held */
596 * If any children exist we must leave the chain intact with refs == 0.
597 * They exist because chains are retained below us which have refs or
598 * may require flushing.
600 * Retry (return chain) if we fail to transition the refs to 0, else
601 * return NULL indication nothing more to do.
603 * Chains with children are NOT put on the LRU list.
605 if (chain->core.chain_count) {
606 if (atomic_cmpset_int(&chain->refs, 1, 0)) {
607 hammer2_spin_unex(&chain->core.spin);
608 hammer2_chain_assert_no_data(chain);
609 hammer2_mtx_unlock(&chain->lock);
612 hammer2_spin_unex(&chain->core.spin);
613 hammer2_mtx_unlock(&chain->lock);
617 /* spinlock still held */
618 /* no chains left under us */
621 * chain->core has no children left so no accessors can get to our
622 * chain from there. Now we have to lock the parent core to interlock
623 * remaining possible accessors that might bump chain's refs before
624 * we can safely drop chain's refs with intent to free the chain.
627 pmp = chain->pmp; /* can be NULL */
630 parent = chain->parent;
633 * WARNING! chain's spin lock is still held here, and other spinlocks
634 * will be acquired and released in the code below. We
635 * cannot be making fancy procedure calls!
639 * We can cache the chain if it is associated with a pmp
640 * and not flagged as being destroyed or requesting a full
641 * release. In this situation the chain is not removed
642 * from its parent, i.e. it can still be looked up.
644 * We intentionally do not cache DATA chains because these
645 * were likely used to load data into the logical buffer cache
646 * and will not be accessed again for some time.
649 (HAMMER2_CHAIN_DESTROY | HAMMER2_CHAIN_RELEASE)) == 0 &&
651 chain->bref.type != HAMMER2_BREF_TYPE_DATA) {
653 hammer2_spin_ex(&parent->core.spin);
654 if (atomic_cmpset_int(&chain->refs, 1, 0) == 0) {
656 * 1->0 transition failed, retry. Do not drop
657 * the chain's data yet!
660 hammer2_spin_unex(&parent->core.spin);
661 hammer2_spin_unex(&chain->core.spin);
662 hammer2_mtx_unlock(&chain->lock);
670 hammer2_chain_assert_no_data(chain);
673 * Make sure we are on the LRU list, clean up excessive
674 * LRU entries. We can only really drop one but there might
675 * be other entries that we can remove from the lru_list
678 * NOTE: HAMMER2_CHAIN_ONLRU may only be safely set when
679 * chain->core.spin AND pmp->lru_spin are held, but
680 * can be safely cleared only holding pmp->lru_spin.
682 if ((chain->flags & HAMMER2_CHAIN_ONLRU) == 0) {
683 hammer2_spin_ex(&pmp->lru_spin);
684 if ((chain->flags & HAMMER2_CHAIN_ONLRU) == 0) {
685 atomic_set_int(&chain->flags,
686 HAMMER2_CHAIN_ONLRU);
687 TAILQ_INSERT_TAIL(&pmp->lru_list,
689 atomic_add_int(&pmp->lru_count, 1);
691 if (pmp->lru_count < HAMMER2_LRU_LIMIT)
692 depth = 1; /* disable lru_list flush */
693 hammer2_spin_unex(&pmp->lru_spin);
695 /* disable lru flush */
700 hammer2_spin_unex(&parent->core.spin);
701 parent = NULL; /* safety */
703 hammer2_spin_unex(&chain->core.spin);
704 hammer2_mtx_unlock(&chain->lock);
707 * lru_list hysteresis (see above for depth overrides).
708 * Note that depth also prevents excessive lastdrop recursion.
711 hammer2_chain_lru_flush(pmp);
718 * Make sure we are not on the LRU list.
720 if (chain->flags & HAMMER2_CHAIN_ONLRU) {
721 hammer2_spin_ex(&pmp->lru_spin);
722 if (chain->flags & HAMMER2_CHAIN_ONLRU) {
723 atomic_add_int(&pmp->lru_count, -1);
724 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU);
725 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node);
727 hammer2_spin_unex(&pmp->lru_spin);
731 * Spinlock the parent and try to drop the last ref on chain.
732 * On success determine if we should dispose of the chain
733 * (remove the chain from its parent, etc).
735 * (normal core locks are top-down recursive but we define
736 * core spinlocks as bottom-up recursive, so this is safe).
739 hammer2_spin_ex(&parent->core.spin);
740 if (atomic_cmpset_int(&chain->refs, 1, 0) == 0) {
742 * 1->0 transition failed, retry.
744 hammer2_spin_unex(&parent->core.spin);
745 hammer2_spin_unex(&chain->core.spin);
746 hammer2_mtx_unlock(&chain->lock);
752 * 1->0 transition successful, parent spin held to prevent
753 * new lookups, chain spinlock held to protect parent field.
754 * Remove chain from the parent.
756 * If the chain is being removed from the parent's btree but
757 * is not bmapped, we have to adjust live_count downward. If
758 * it is bmapped then the blockref is retained in the parent
759 * as is its associated live_count. This case can occur when
760 * a chain added to the topology is unable to flush and is
761 * then later deleted.
763 if (chain->flags & HAMMER2_CHAIN_ONRBTREE) {
764 if ((parent->flags & HAMMER2_CHAIN_COUNTEDBREFS) &&
765 (chain->flags & HAMMER2_CHAIN_BMAPPED) == 0) {
766 atomic_add_int(&parent->core.live_count, -1);
768 RB_REMOVE(hammer2_chain_tree,
769 &parent->core.rbtree, chain);
770 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE);
771 --parent->core.chain_count;
772 chain->parent = NULL;
776 * If our chain was the last chain in the parent's core the
777 * core is now empty and its parent might have to be
778 * re-dropped if it has 0 refs.
780 if (parent->core.chain_count == 0) {
782 atomic_add_int(&rdrop->refs, 1);
784 if (atomic_cmpset_int(&rdrop->refs, 0, 1) == 0)
788 hammer2_spin_unex(&parent->core.spin);
789 parent = NULL; /* safety */
795 if (atomic_cmpset_int(&chain->refs, 1, 0) == 0) {
797 * 1->0 transition failed, retry.
799 hammer2_spin_unex(&parent->core.spin);
800 hammer2_spin_unex(&chain->core.spin);
801 hammer2_mtx_unlock(&chain->lock);
808 * Successful 1->0 transition, no parent, no children... no way for
809 * anyone to ref this chain any more. We can clean-up and free it.
811 * We still have the core spinlock, and core's chain_count is 0.
812 * Any parent spinlock is gone.
814 hammer2_spin_unex(&chain->core.spin);
815 hammer2_chain_assert_no_data(chain);
816 hammer2_mtx_unlock(&chain->lock);
817 KKASSERT(RB_EMPTY(&chain->core.rbtree) &&
818 chain->core.chain_count == 0);
821 * All locks are gone, no pointers remain to the chain, finish
824 KKASSERT((chain->flags & (HAMMER2_CHAIN_UPDATE |
825 HAMMER2_CHAIN_MODIFIED)) == 0);
828 * Once chain resources are gone we can use the now dead chain
829 * structure to placehold what might otherwise require a recursive
830 * drop, because we have potentially two things to drop and can only
831 * return one directly.
833 if (chain->flags & HAMMER2_CHAIN_ALLOCATED) {
834 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ALLOCATED);
836 kfree(chain, hmp->mchain);
840 * Possible chaining loop when parent re-drop needed.
846 * Heuristical flush of the LRU, try to reduce the number of entries
847 * on the LRU to (HAMMER2_LRU_LIMIT * 2 / 3). This procedure is called
848 * only when lru_count exceeds HAMMER2_LRU_LIMIT.
852 hammer2_chain_lru_flush(hammer2_pfs_t *pmp)
854 hammer2_chain_t *chain;
858 hammer2_spin_ex(&pmp->lru_spin);
859 while (pmp->lru_count > HAMMER2_LRU_LIMIT * 2 / 3) {
861 * Pick a chain off the lru_list, just recycle it quickly
862 * if LRUHINT is set (the chain was ref'd but left on
863 * the lru_list, so cycle to the end).
865 chain = TAILQ_FIRST(&pmp->lru_list);
866 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node);
868 if (chain->flags & HAMMER2_CHAIN_LRUHINT) {
869 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_LRUHINT);
870 TAILQ_INSERT_TAIL(&pmp->lru_list, chain, lru_node);
876 * Ok, we are off the LRU. We must adjust refs before we
877 * can safely clear the ONLRU flag.
879 atomic_add_int(&pmp->lru_count, -1);
880 if (atomic_cmpset_int(&chain->refs, 0, 1)) {
881 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU);
882 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
885 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU);
888 hammer2_spin_unex(&pmp->lru_spin);
893 * If we picked a chain off the lru list we may be able to lastdrop
894 * it. Use a depth of 1 to prevent excessive lastdrop recursion.
904 if (hammer2_mtx_ex_try(&chain->lock) == 0)
905 chain = hammer2_chain_lastdrop(chain, 1);
906 /* retry the same chain, or chain from lastdrop */
908 if (atomic_cmpset_int(&chain->refs, refs, refs - 1))
910 /* retry the same chain */
918 * On last lock release.
920 static hammer2_io_t *
921 hammer2_chain_drop_data(hammer2_chain_t *chain)
925 if ((dio = chain->dio) != NULL) {
929 switch(chain->bref.type) {
930 case HAMMER2_BREF_TYPE_VOLUME:
931 case HAMMER2_BREF_TYPE_FREEMAP:
934 if (chain->data != NULL) {
935 hammer2_spin_unex(&chain->core.spin);
936 panic("chain data not null: "
937 "chain %p bref %016jx.%02x "
938 "refs %d parent %p dio %p data %p",
939 chain, chain->bref.data_off,
940 chain->bref.type, chain->refs,
942 chain->dio, chain->data);
944 KKASSERT(chain->data == NULL);
952 * Lock a referenced chain element, acquiring its data with I/O if necessary,
953 * and specify how you would like the data to be resolved.
955 * If an I/O or other fatal error occurs, chain->error will be set to non-zero.
957 * The lock is allowed to recurse, multiple locking ops will aggregate
958 * the requested resolve types. Once data is assigned it will not be
959 * removed until the last unlock.
961 * HAMMER2_RESOLVE_NEVER - Do not resolve the data element.
962 * (typically used to avoid device/logical buffer
965 * HAMMER2_RESOLVE_MAYBE - Do not resolve data elements for chains in
966 * the INITIAL-create state (indirect blocks only).
968 * Do not resolve data elements for DATA chains.
969 * (typically used to avoid device/logical buffer
972 * HAMMER2_RESOLVE_ALWAYS- Always resolve the data element.
974 * HAMMER2_RESOLVE_SHARED- (flag) The chain is locked shared, otherwise
975 * it will be locked exclusive.
977 * HAMMER2_RESOLVE_NONBLOCK- (flag) The chain is locked non-blocking. If
978 * the lock fails, EAGAIN is returned.
980 * NOTE: Embedded elements (volume header, inodes) are always resolved
983 * NOTE: Specifying HAMMER2_RESOLVE_ALWAYS on a newly-created non-embedded
984 * element will instantiate and zero its buffer, and flush it on
987 * NOTE: (data) elements are normally locked RESOLVE_NEVER or RESOLVE_MAYBE
988 * so as not to instantiate a device buffer, which could alias against
989 * a logical file buffer. However, if ALWAYS is specified the
990 * device buffer will be instantiated anyway.
992 * NOTE: The return value is always 0 unless NONBLOCK is specified, in which
993 * case it can be either 0 or EAGAIN.
995 * WARNING! This function blocks on I/O if data needs to be fetched. This
996 * blocking can run concurrent with other compatible lock holders
997 * who do not need data returning. The lock is not upgraded to
998 * exclusive during a data fetch, a separate bit is used to
999 * interlock I/O. However, an exclusive lock holder can still count
1000 * on being interlocked against an I/O fetch managed by a shared
1004 hammer2_chain_lock(hammer2_chain_t *chain, int how)
1006 KKASSERT(chain->refs > 0);
1008 if (how & HAMMER2_RESOLVE_NONBLOCK) {
1010 * We still have to bump lockcnt before acquiring the lock,
1011 * even for non-blocking operation, because the unlock code
1012 * live-loops on lockcnt == 1 when dropping the last lock.
1014 * If the non-blocking operation fails we have to use an
1015 * unhold sequence to undo the mess.
1017 * NOTE: LOCKAGAIN must always succeed without blocking,
1018 * even if NONBLOCK is specified.
1020 atomic_add_int(&chain->lockcnt, 1);
1021 if (how & HAMMER2_RESOLVE_SHARED) {
1022 if (how & HAMMER2_RESOLVE_LOCKAGAIN) {
1023 hammer2_mtx_sh_again(&chain->lock);
1025 if (hammer2_mtx_sh_try(&chain->lock) != 0) {
1026 hammer2_chain_unhold(chain);
1031 if (hammer2_mtx_ex_try(&chain->lock) != 0) {
1032 hammer2_chain_unhold(chain);
1038 * Get the appropriate lock. If LOCKAGAIN is flagged with
1039 * SHARED the caller expects a shared lock to already be
1040 * present and we are giving it another ref. This case must
1041 * importantly not block if there is a pending exclusive lock
1044 atomic_add_int(&chain->lockcnt, 1);
1045 if (how & HAMMER2_RESOLVE_SHARED) {
1046 if (how & HAMMER2_RESOLVE_LOCKAGAIN) {
1047 hammer2_mtx_sh_again(&chain->lock);
1049 hammer2_mtx_sh(&chain->lock);
1052 hammer2_mtx_ex(&chain->lock);
1057 * If we already have a valid data pointer make sure the data is
1058 * synchronized to the current cpu, and then no further action is
1063 hammer2_io_bkvasync(chain->dio);
1068 * Do we have to resolve the data? This is generally only
1069 * applicable to HAMMER2_BREF_TYPE_DATA which is special-cased.
1070 * Other BREF types expects the data to be there.
1072 switch(how & HAMMER2_RESOLVE_MASK) {
1073 case HAMMER2_RESOLVE_NEVER:
1075 case HAMMER2_RESOLVE_MAYBE:
1076 if (chain->flags & HAMMER2_CHAIN_INITIAL)
1078 if (chain->bref.type == HAMMER2_BREF_TYPE_DATA)
1081 if (chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE)
1083 if (chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_LEAF)
1087 case HAMMER2_RESOLVE_ALWAYS:
1093 * Caller requires data
1095 hammer2_chain_load_data(chain);
1101 * Lock the chain, retain the hold, and drop the data persistence count.
1102 * The data should remain valid because we never transitioned lockcnt
1106 hammer2_chain_lock_unhold(hammer2_chain_t *chain, int how)
1108 hammer2_chain_lock(chain, how);
1109 atomic_add_int(&chain->lockcnt, -1);
1114 * Downgrade an exclusive chain lock to a shared chain lock.
1116 * NOTE: There is no upgrade equivalent due to the ease of
1117 * deadlocks in that direction.
1120 hammer2_chain_lock_downgrade(hammer2_chain_t *chain)
1122 hammer2_mtx_downgrade(&chain->lock);
1127 * Issue I/O and install chain->data. Caller must hold a chain lock, lock
1128 * may be of any type.
1130 * Once chain->data is set it cannot be disposed of until all locks are
1133 * Make sure the data is synchronized to the current cpu.
1136 hammer2_chain_load_data(hammer2_chain_t *chain)
1138 hammer2_blockref_t *bref;
1145 * Degenerate case, data already present, or chain has no media
1146 * reference to load.
1148 KKASSERT(chain->lock.mtx_lock & MTX_MASK);
1151 hammer2_io_bkvasync(chain->dio);
1154 if ((chain->bref.data_off & ~HAMMER2_OFF_MASK_RADIX) == 0)
1158 KKASSERT(hmp != NULL);
1161 * Gain the IOINPROG bit, interlocked block.
1167 oflags = chain->flags;
1169 if (oflags & HAMMER2_CHAIN_IOINPROG) {
1170 nflags = oflags | HAMMER2_CHAIN_IOSIGNAL;
1171 tsleep_interlock(&chain->flags, 0);
1172 if (atomic_cmpset_int(&chain->flags, oflags, nflags)) {
1173 tsleep(&chain->flags, PINTERLOCKED,
1178 nflags = oflags | HAMMER2_CHAIN_IOINPROG;
1179 if (atomic_cmpset_int(&chain->flags, oflags, nflags)) {
1187 * We own CHAIN_IOINPROG
1189 * Degenerate case if we raced another load.
1193 hammer2_io_bkvasync(chain->dio);
1198 * We must resolve to a device buffer, either by issuing I/O or
1199 * by creating a zero-fill element. We do not mark the buffer
1200 * dirty when creating a zero-fill element (the hammer2_chain_modify()
1201 * API must still be used to do that).
1203 * The device buffer is variable-sized in powers of 2 down
1204 * to HAMMER2_MIN_ALLOC (typically 1K). A 64K physical storage
1205 * chunk always contains buffers of the same size. (XXX)
1207 * The minimum physical IO size may be larger than the variable
1210 bref = &chain->bref;
1213 * The getblk() optimization can only be used on newly created
1214 * elements if the physical block size matches the request.
1216 if (chain->flags & HAMMER2_CHAIN_INITIAL) {
1217 error = hammer2_io_new(hmp, bref->type,
1218 bref->data_off, chain->bytes,
1221 error = hammer2_io_bread(hmp, bref->type,
1222 bref->data_off, chain->bytes,
1224 hammer2_adjreadcounter(chain->bref.type, chain->bytes);
1227 chain->error = HAMMER2_ERROR_EIO;
1228 kprintf("hammer2_chain_load_data: I/O error %016jx: %d\n",
1229 (intmax_t)bref->data_off, error);
1230 hammer2_io_bqrelse(&chain->dio);
1236 * This isn't perfect and can be ignored on OSs which do not have
1237 * an indication as to whether a buffer is coming from cache or
1238 * if I/O was actually issued for the read. TESTEDGOOD will work
1239 * pretty well without the B_IOISSUED logic because chains are
1240 * cached, but in that situation (without B_IOISSUED) it will not
1241 * detect whether a re-read via I/O is corrupted verses the original
1244 * We can't re-run the CRC on every fresh lock. That would be
1245 * insanely expensive.
1247 * If the underlying kernel buffer covers the entire chain we can
1248 * use the B_IOISSUED indication to determine if we have to re-run
1249 * the CRC on chain data for chains that managed to stay cached
1250 * across the kernel disposal of the original buffer.
1252 if ((dio = chain->dio) != NULL && dio->bp) {
1253 struct buf *bp = dio->bp;
1255 if (dio->psize == chain->bytes &&
1256 (bp->b_flags & B_IOISSUED)) {
1257 atomic_clear_int(&chain->flags,
1258 HAMMER2_CHAIN_TESTEDGOOD);
1259 bp->b_flags &= ~B_IOISSUED;
1264 * NOTE: A locked chain's data cannot be modified without first
1265 * calling hammer2_chain_modify().
1269 * NOTE: hammer2_io_data() call issues bkvasync()
1271 bdata = hammer2_io_data(chain->dio, chain->bref.data_off);
1273 if (chain->flags & HAMMER2_CHAIN_INITIAL) {
1275 * Clear INITIAL. In this case we used io_new() and the
1276 * buffer has been zero'd and marked dirty.
1278 * CHAIN_MODIFIED has not been set yet, and we leave it
1279 * that way for now. Set a temporary CHAIN_NOTTESTED flag
1280 * to prevent hammer2_chain_testcheck() from trying to match
1281 * a check code that has not yet been generated. This bit
1282 * should NOT end up on the actual media.
1284 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1285 atomic_set_int(&chain->flags, HAMMER2_CHAIN_NOTTESTED);
1286 } else if (chain->flags & HAMMER2_CHAIN_MODIFIED) {
1288 * check data not currently synchronized due to
1289 * modification. XXX assumes data stays in the buffer
1290 * cache, which might not be true (need biodep on flush
1291 * to calculate crc? or simple crc?).
1293 } else if ((chain->flags & HAMMER2_CHAIN_TESTEDGOOD) == 0) {
1294 if (hammer2_chain_testcheck(chain, bdata) == 0) {
1295 chain->error = HAMMER2_ERROR_CHECK;
1297 atomic_set_int(&chain->flags, HAMMER2_CHAIN_TESTEDGOOD);
1302 * Setup the data pointer, either pointing it to an embedded data
1303 * structure and copying the data from the buffer, or pointing it
1306 * The buffer is not retained when copying to an embedded data
1307 * structure in order to avoid potential deadlocks or recursions
1308 * on the same physical buffer.
1310 * WARNING! Other threads can start using the data the instant we
1311 * set chain->data non-NULL.
1313 switch (bref->type) {
1314 case HAMMER2_BREF_TYPE_VOLUME:
1315 case HAMMER2_BREF_TYPE_FREEMAP:
1317 * Copy data from bp to embedded buffer
1319 panic("hammer2_chain_load_data: unresolved volume header");
1321 case HAMMER2_BREF_TYPE_DIRENT:
1322 KKASSERT(chain->bytes != 0);
1324 case HAMMER2_BREF_TYPE_INODE:
1325 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
1326 case HAMMER2_BREF_TYPE_INDIRECT:
1327 case HAMMER2_BREF_TYPE_DATA:
1328 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1331 * Point data at the device buffer and leave dio intact.
1333 chain->data = (void *)bdata;
1338 * Release HAMMER2_CHAIN_IOINPROG and signal waiters if requested.
1345 oflags = chain->flags;
1346 nflags = oflags & ~(HAMMER2_CHAIN_IOINPROG |
1347 HAMMER2_CHAIN_IOSIGNAL);
1348 KKASSERT(oflags & HAMMER2_CHAIN_IOINPROG);
1349 if (atomic_cmpset_int(&chain->flags, oflags, nflags)) {
1350 if (oflags & HAMMER2_CHAIN_IOSIGNAL)
1351 wakeup(&chain->flags);
1358 * Unlock and deref a chain element.
1360 * Remember that the presence of children under chain prevent the chain's
1361 * destruction but do not add additional references, so the dio will still
1365 hammer2_chain_unlock(hammer2_chain_t *chain)
1372 * If multiple locks are present (or being attempted) on this
1373 * particular chain we can just unlock, drop refs, and return.
1375 * Otherwise fall-through on the 1->0 transition.
1378 lockcnt = chain->lockcnt;
1379 KKASSERT(lockcnt > 0);
1382 if (atomic_cmpset_int(&chain->lockcnt,
1383 lockcnt, lockcnt - 1)) {
1384 hammer2_mtx_unlock(&chain->lock);
1387 } else if (hammer2_mtx_upgrade_try(&chain->lock) == 0) {
1388 /* while holding the mutex exclusively */
1389 if (atomic_cmpset_int(&chain->lockcnt, 1, 0))
1393 * This situation can easily occur on SMP due to
1394 * the gap inbetween the 1->0 transition and the
1395 * final unlock. We cannot safely block on the
1396 * mutex because lockcnt might go above 1.
1398 * XXX Sleep for one tick if it takes too long.
1400 if (++iter > 1000) {
1401 if (iter > 1000 + hz) {
1402 kprintf("hammer2: h2race2 %p\n", chain);
1405 tsleep(&iter, 0, "h2race2", 1);
1413 * Last unlock / mutex upgraded to exclusive. Drop the data
1416 dio = hammer2_chain_drop_data(chain);
1418 hammer2_io_bqrelse(&dio);
1419 hammer2_mtx_unlock(&chain->lock);
1423 * Unlock and hold chain data intact
1426 hammer2_chain_unlock_hold(hammer2_chain_t *chain)
1428 atomic_add_int(&chain->lockcnt, 1);
1429 hammer2_chain_unlock(chain);
1433 * Helper to obtain the blockref[] array base and count for a chain.
1435 * XXX Not widely used yet, various use cases need to be validated and
1436 * converted to use this function.
1439 hammer2_blockref_t *
1440 hammer2_chain_base_and_count(hammer2_chain_t *parent, int *countp)
1442 hammer2_blockref_t *base;
1445 if (parent->flags & HAMMER2_CHAIN_INITIAL) {
1448 switch(parent->bref.type) {
1449 case HAMMER2_BREF_TYPE_INODE:
1450 count = HAMMER2_SET_COUNT;
1452 case HAMMER2_BREF_TYPE_INDIRECT:
1453 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1454 count = parent->bytes / sizeof(hammer2_blockref_t);
1456 case HAMMER2_BREF_TYPE_VOLUME:
1457 count = HAMMER2_SET_COUNT;
1459 case HAMMER2_BREF_TYPE_FREEMAP:
1460 count = HAMMER2_SET_COUNT;
1463 panic("hammer2_chain_base_and_count: "
1464 "unrecognized blockref type: %d",
1470 switch(parent->bref.type) {
1471 case HAMMER2_BREF_TYPE_INODE:
1472 base = &parent->data->ipdata.u.blockset.blockref[0];
1473 count = HAMMER2_SET_COUNT;
1475 case HAMMER2_BREF_TYPE_INDIRECT:
1476 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1477 base = &parent->data->npdata[0];
1478 count = parent->bytes / sizeof(hammer2_blockref_t);
1480 case HAMMER2_BREF_TYPE_VOLUME:
1481 base = &parent->data->voldata.
1482 sroot_blockset.blockref[0];
1483 count = HAMMER2_SET_COUNT;
1485 case HAMMER2_BREF_TYPE_FREEMAP:
1486 base = &parent->data->blkset.blockref[0];
1487 count = HAMMER2_SET_COUNT;
1490 panic("hammer2_chain_base_and_count: "
1491 "unrecognized blockref type: %d",
1503 * This counts the number of live blockrefs in a block array and
1504 * also calculates the point at which all remaining blockrefs are empty.
1505 * This routine can only be called on a live chain.
1507 * Caller holds the chain locked, but possibly with a shared lock. We
1508 * must use an exclusive spinlock to prevent corruption.
1510 * NOTE: Flag is not set until after the count is complete, allowing
1511 * callers to test the flag without holding the spinlock.
1513 * NOTE: If base is NULL the related chain is still in the INITIAL
1514 * state and there are no blockrefs to count.
1516 * NOTE: live_count may already have some counts accumulated due to
1517 * creation and deletion and could even be initially negative.
1520 hammer2_chain_countbrefs(hammer2_chain_t *chain,
1521 hammer2_blockref_t *base, int count)
1523 hammer2_spin_ex(&chain->core.spin);
1524 if ((chain->flags & HAMMER2_CHAIN_COUNTEDBREFS) == 0) {
1526 while (--count >= 0) {
1527 if (base[count].type != HAMMER2_BREF_TYPE_EMPTY)
1530 chain->core.live_zero = count + 1;
1531 while (count >= 0) {
1532 if (base[count].type != HAMMER2_BREF_TYPE_EMPTY)
1533 atomic_add_int(&chain->core.live_count,
1538 chain->core.live_zero = 0;
1540 /* else do not modify live_count */
1541 atomic_set_int(&chain->flags, HAMMER2_CHAIN_COUNTEDBREFS);
1543 hammer2_spin_unex(&chain->core.spin);
1547 * Resize the chain's physical storage allocation in-place. This function does
1548 * not usually adjust the data pointer and must be followed by (typically) a
1549 * hammer2_chain_modify() call to copy any old data over and adjust the
1552 * Chains can be resized smaller without reallocating the storage. Resizing
1553 * larger will reallocate the storage. Excess or prior storage is reclaimed
1554 * asynchronously at a later time.
1556 * An nradix value of 0 is special-cased to mean that the storage should
1557 * be disassociated, that is the chain is being resized to 0 bytes (not 1
1560 * Must be passed an exclusively locked parent and chain.
1562 * This function is mostly used with DATA blocks locked RESOLVE_NEVER in order
1563 * to avoid instantiating a device buffer that conflicts with the vnode data
1564 * buffer. However, because H2 can compress or encrypt data, the chain may
1565 * have a dio assigned to it in those situations, and they do not conflict.
1567 * XXX return error if cannot resize.
1570 hammer2_chain_resize(hammer2_chain_t *chain,
1571 hammer2_tid_t mtid, hammer2_off_t dedup_off,
1572 int nradix, int flags)
1582 * Only data and indirect blocks can be resized for now.
1583 * (The volu root, inodes, and freemap elements use a fixed size).
1585 KKASSERT(chain != &hmp->vchain);
1586 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_DATA ||
1587 chain->bref.type == HAMMER2_BREF_TYPE_INDIRECT ||
1588 chain->bref.type == HAMMER2_BREF_TYPE_DIRENT);
1591 * Nothing to do if the element is already the proper size
1593 obytes = chain->bytes;
1594 nbytes = (nradix) ? (1U << nradix) : 0;
1595 if (obytes == nbytes)
1596 return (chain->error);
1599 * Make sure the old data is instantiated so we can copy it. If this
1600 * is a data block, the device data may be superfluous since the data
1601 * might be in a logical block, but compressed or encrypted data is
1604 * NOTE: The modify will set BMAPUPD for us if BMAPPED is set.
1606 error = hammer2_chain_modify(chain, mtid, dedup_off, 0);
1611 * Reallocate the block, even if making it smaller (because different
1612 * block sizes may be in different regions).
1614 * NOTE: Operation does not copy the data and may only be used
1615 * to resize data blocks in-place, or directory entry blocks
1616 * which are about to be modified in some manner.
1618 error = hammer2_freemap_alloc(chain, nbytes);
1622 chain->bytes = nbytes;
1625 * We don't want the followup chain_modify() to try to copy data
1626 * from the old (wrong-sized) buffer. It won't know how much to
1627 * copy. This case should only occur during writes when the
1628 * originator already has the data to write in-hand.
1631 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_DATA ||
1632 chain->bref.type == HAMMER2_BREF_TYPE_DIRENT);
1633 hammer2_io_brelse(&chain->dio);
1636 return (chain->error);
1640 * Set the chain modified so its data can be changed by the caller, or
1641 * install deduplicated data. The caller must call this routine for each
1642 * set of modifications it makes, even if the chain is already flagged
1645 * Sets bref.modify_tid to mtid only if mtid != 0. Note that bref.modify_tid
1646 * is a CLC (cluster level change) field and is not updated by parent
1647 * propagation during a flush.
1649 * Returns an appropriate HAMMER2_ERROR_* code, which will generally reflect
1650 * chain->error except for HAMMER2_ERROR_ENOSPC. If the allocation fails
1651 * due to no space available, HAMMER2_ERROR_ENOSPC is returned and the chain
1652 * remains unmodified with its old data ref intact and chain->error
1657 * If the DEDUPABLE flag is set in the chain the storage must be reallocated
1658 * even if the chain is still flagged MODIFIED. In this case the chain's
1659 * DEDUPABLE flag will be cleared once the new storage has been assigned.
1661 * If the caller passes a non-zero dedup_off we will use it to assign the
1662 * new storage. The MODIFIED flag will be *CLEARED* in this case, and
1663 * DEDUPABLE will be set (NOTE: the UPDATE flag is always set). The caller
1664 * must not modify the data content upon return.
1667 hammer2_chain_modify(hammer2_chain_t *chain, hammer2_tid_t mtid,
1668 hammer2_off_t dedup_off, int flags)
1670 hammer2_blockref_t obref;
1681 obref = chain->bref;
1682 KKASSERT(chain->lock.mtx_lock & MTX_EXCLUSIVE);
1685 * Data is not optional for freemap chains (we must always be sure
1686 * to copy the data on COW storage allocations).
1688 if (chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE ||
1689 chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) {
1690 KKASSERT((chain->flags & HAMMER2_CHAIN_INITIAL) ||
1691 (flags & HAMMER2_MODIFY_OPTDATA) == 0);
1695 * Data must be resolved if already assigned, unless explicitly
1696 * flagged otherwise. If we cannot safety load the data the
1697 * modification fails and we return early.
1699 if (chain->data == NULL && chain->bytes != 0 &&
1700 (flags & HAMMER2_MODIFY_OPTDATA) == 0 &&
1701 (chain->bref.data_off & ~HAMMER2_OFF_MASK_RADIX)) {
1702 hammer2_chain_load_data(chain);
1704 return (chain->error);
1709 * Set MODIFIED to indicate that the chain has been modified. A new
1710 * allocation is required when modifying a chain.
1712 * Set UPDATE to ensure that the blockref is updated in the parent.
1714 * If MODIFIED is already set determine if we can reuse the assigned
1715 * data block or if we need a new data block.
1717 if ((chain->flags & HAMMER2_CHAIN_MODIFIED) == 0) {
1719 * Must set modified bit.
1721 atomic_add_long(&hammer2_count_modified_chains, 1);
1722 atomic_set_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
1723 hammer2_pfs_memory_inc(chain->pmp); /* can be NULL */
1727 * We may be able to avoid a copy-on-write if the chain's
1728 * check mode is set to NONE and the chain's current
1729 * modify_tid is beyond the last explicit snapshot tid.
1731 * This implements HAMMER2's overwrite-in-place feature.
1733 * NOTE! This data-block cannot be used as a de-duplication
1734 * source when the check mode is set to NONE.
1736 if ((chain->bref.type == HAMMER2_BREF_TYPE_DATA ||
1737 chain->bref.type == HAMMER2_BREF_TYPE_DIRENT) &&
1738 (chain->flags & HAMMER2_CHAIN_INITIAL) == 0 &&
1739 (chain->flags & HAMMER2_CHAIN_DEDUPABLE) == 0 &&
1740 HAMMER2_DEC_CHECK(chain->bref.methods) ==
1741 HAMMER2_CHECK_NONE &&
1743 chain->bref.modify_tid >
1744 chain->pmp->iroot->meta.pfs_lsnap_tid) {
1746 * Sector overwrite allowed.
1749 } else if ((hmp->hflags & HMNT2_EMERG) &&
1751 chain->bref.modify_tid >
1752 chain->pmp->iroot->meta.pfs_lsnap_tid) {
1754 * If in emergency delete mode then do a modify-in-
1755 * place on any chain type belonging to the PFS as
1756 * long as it doesn't mess up a snapshot. We might
1757 * be forced to do this anyway a little further down
1758 * in the code if the allocation fails.
1760 * Also note that in emergency mode, these modify-in-
1761 * place operations are NOT SAFE. A storage failure,
1762 * power failure, or panic can corrupt the filesystem.
1767 * Sector overwrite not allowed, must copy-on-write.
1771 } else if (chain->flags & HAMMER2_CHAIN_DEDUPABLE) {
1773 * If the modified chain was registered for dedup we need
1774 * a new allocation. This only happens for delayed-flush
1775 * chains (i.e. which run through the front-end buffer
1782 * Already flagged modified, no new allocation is needed.
1789 * Flag parent update required.
1791 if ((chain->flags & HAMMER2_CHAIN_UPDATE) == 0) {
1792 atomic_set_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
1799 * The XOP code returns held but unlocked focus chains. This
1800 * prevents the chain from being destroyed but does not prevent
1801 * it from being modified. diolk is used to interlock modifications
1802 * against XOP frontend accesses to the focus.
1804 * This allows us to theoretically avoid deadlocking the frontend
1805 * if one of the backends lock up by not formally locking the
1806 * focused chain in the frontend. In addition, the synchronization
1807 * code relies on this mechanism to avoid deadlocking concurrent
1808 * synchronization threads.
1810 lockmgr(&chain->diolk, LK_EXCLUSIVE);
1813 * The modification or re-modification requires an allocation and
1814 * possible COW. If an error occurs, the previous content and data
1815 * reference is retained and the modification fails.
1817 * If dedup_off is non-zero, the caller is requesting a deduplication
1818 * rather than a modification. The MODIFIED bit is not set and the
1819 * data offset is set to the deduplication offset. The data cannot
1822 * NOTE: The dedup offset is allowed to be in a partially free state
1823 * and we must be sure to reset it to a fully allocated state
1824 * to force two bulkfree passes to free it again.
1826 * NOTE: Only applicable when chain->bytes != 0.
1828 * XXX can a chain already be marked MODIFIED without a data
1829 * assignment? If not, assert here instead of testing the case.
1831 if (chain != &hmp->vchain && chain != &hmp->fchain &&
1833 if ((chain->bref.data_off & ~HAMMER2_OFF_MASK_RADIX) == 0 ||
1837 * NOTE: We do not have to remove the dedup
1838 * registration because the area is still
1839 * allocated and the underlying DIO will
1843 chain->bref.data_off = dedup_off;
1844 chain->bytes = 1 << (dedup_off &
1845 HAMMER2_OFF_MASK_RADIX);
1847 atomic_clear_int(&chain->flags,
1848 HAMMER2_CHAIN_MODIFIED);
1849 atomic_add_long(&hammer2_count_modified_chains,
1852 hammer2_pfs_memory_wakeup(
1855 hammer2_freemap_adjust(hmp, &chain->bref,
1856 HAMMER2_FREEMAP_DORECOVER);
1857 atomic_set_int(&chain->flags,
1858 HAMMER2_CHAIN_DEDUPABLE);
1860 error = hammer2_freemap_alloc(chain,
1862 atomic_clear_int(&chain->flags,
1863 HAMMER2_CHAIN_DEDUPABLE);
1866 * If we are unable to allocate a new block
1867 * but we are in emergency mode, issue a
1868 * warning to the console and reuse the same
1871 * We behave as if the allocation were
1874 * THIS IS IMPORTANT: These modifications
1875 * are virtually guaranteed to corrupt any
1876 * snapshots related to this filesystem.
1878 if (error && (hmp->hflags & HMNT2_EMERG)) {
1880 chain->bref.flags |=
1881 HAMMER2_BREF_FLAG_EMERG_MIP;
1883 krateprintf(&krate_h2em,
1884 "hammer2: Emergency Mode WARNING: "
1885 "Operation will likely corrupt "
1886 "related snapshot: "
1887 "%016jx.%02x key=%016jx\n",
1888 chain->bref.data_off,
1891 } else if (error == 0) {
1892 chain->bref.flags &=
1893 ~HAMMER2_BREF_FLAG_EMERG_MIP;
1900 * Stop here if error. We have to undo any flag bits we might
1905 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
1906 atomic_add_long(&hammer2_count_modified_chains, -1);
1908 hammer2_pfs_memory_wakeup(chain->pmp, -1);
1911 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
1913 lockmgr(&chain->diolk, LK_RELEASE);
1919 * Update mirror_tid and modify_tid. modify_tid is only updated
1920 * if not passed as zero (during flushes, parent propagation passes
1923 * NOTE: chain->pmp could be the device spmp.
1925 chain->bref.mirror_tid = hmp->voldata.mirror_tid + 1;
1927 chain->bref.modify_tid = mtid;
1930 * Set BMAPUPD to tell the flush code that an existing blockmap entry
1931 * requires updating as well as to tell the delete code that the
1932 * chain's blockref might not exactly match (in terms of physical size
1933 * or block offset) the one in the parent's blocktable. The base key
1934 * of course will still match.
1936 if (chain->flags & HAMMER2_CHAIN_BMAPPED)
1937 atomic_set_int(&chain->flags, HAMMER2_CHAIN_BMAPUPD);
1940 * Short-cut data block handling when the caller does not need an
1941 * actual data reference to (aka OPTDATA), as long as the chain does
1942 * not already have a data pointer to the data and no de-duplication
1945 * This generally means that the modifications are being done via the
1946 * logical buffer cache.
1948 * NOTE: If deduplication occurred we have to run through the data
1949 * stuff to clear INITIAL, and the caller will likely want to
1950 * assign the check code anyway. Leaving INITIAL set on a
1951 * dedup can be deadly (it can cause the block to be zero'd!).
1953 * This code also handles bytes == 0 (most dirents).
1955 if (chain->bref.type == HAMMER2_BREF_TYPE_DATA &&
1956 (flags & HAMMER2_MODIFY_OPTDATA) &&
1957 chain->data == NULL) {
1958 if (dedup_off == 0) {
1959 KKASSERT(chain->dio == NULL);
1965 * Clearing the INITIAL flag (for indirect blocks) indicates that
1966 * we've processed the uninitialized storage allocation.
1968 * If this flag is already clear we are likely in a copy-on-write
1969 * situation but we have to be sure NOT to bzero the storage if
1970 * no data is present.
1972 * Clearing of NOTTESTED is allowed if the MODIFIED bit is set,
1974 if (chain->flags & HAMMER2_CHAIN_INITIAL) {
1975 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1982 * Instantiate data buffer and possibly execute COW operation
1984 switch(chain->bref.type) {
1985 case HAMMER2_BREF_TYPE_VOLUME:
1986 case HAMMER2_BREF_TYPE_FREEMAP:
1988 * The data is embedded, no copy-on-write operation is
1991 KKASSERT(chain->dio == NULL);
1993 case HAMMER2_BREF_TYPE_DIRENT:
1995 * The data might be fully embedded.
1997 if (chain->bytes == 0) {
1998 KKASSERT(chain->dio == NULL);
2002 case HAMMER2_BREF_TYPE_INODE:
2003 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2004 case HAMMER2_BREF_TYPE_DATA:
2005 case HAMMER2_BREF_TYPE_INDIRECT:
2006 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2008 * Perform the copy-on-write operation
2010 * zero-fill or copy-on-write depending on whether
2011 * chain->data exists or not and set the dirty state for
2012 * the new buffer. hammer2_io_new() will handle the
2015 * If a dedup_off was supplied this is an existing block
2016 * and no COW, copy, or further modification is required.
2018 KKASSERT(chain != &hmp->vchain && chain != &hmp->fchain);
2020 if (wasinitial && dedup_off == 0) {
2021 error = hammer2_io_new(hmp, chain->bref.type,
2022 chain->bref.data_off,
2023 chain->bytes, &dio);
2025 error = hammer2_io_bread(hmp, chain->bref.type,
2026 chain->bref.data_off,
2027 chain->bytes, &dio);
2029 hammer2_adjreadcounter(chain->bref.type, chain->bytes);
2032 * If an I/O error occurs make sure callers cannot accidently
2033 * modify the old buffer's contents and corrupt the filesystem.
2035 * NOTE: hammer2_io_data() call issues bkvasync()
2038 kprintf("hammer2_chain_modify: hmp=%p I/O error\n",
2040 chain->error = HAMMER2_ERROR_EIO;
2041 hammer2_io_brelse(&dio);
2042 hammer2_io_brelse(&chain->dio);
2047 bdata = hammer2_io_data(dio, chain->bref.data_off);
2051 * COW (unless a dedup).
2053 KKASSERT(chain->dio != NULL);
2054 if (chain->data != (void *)bdata && dedup_off == 0) {
2055 bcopy(chain->data, bdata, chain->bytes);
2057 } else if (wasinitial == 0 && dedup_off == 0) {
2059 * We have a problem. We were asked to COW but
2060 * we don't have any data to COW with!
2062 panic("hammer2_chain_modify: having a COW %p\n",
2067 * Retire the old buffer, replace with the new. Dirty or
2068 * redirty the new buffer.
2070 * WARNING! The system buffer cache may have already flushed
2071 * the buffer, so we must be sure to [re]dirty it
2072 * for further modification.
2074 * If dedup_off was supplied, the caller is not
2075 * expected to make any further modification to the
2078 * WARNING! hammer2_get_gdata() assumes dio never transitions
2079 * through NULL in order to optimize away unnecessary
2085 if ((tio = chain->dio) != NULL)
2086 hammer2_io_bqrelse(&tio);
2087 chain->data = (void *)bdata;
2090 hammer2_io_setdirty(dio);
2094 panic("hammer2_chain_modify: illegal non-embedded type %d",
2101 * setflush on parent indicating that the parent must recurse down
2102 * to us. Do not call on chain itself which might already have it
2106 hammer2_chain_setflush(chain->parent);
2107 lockmgr(&chain->diolk, LK_RELEASE);
2109 return (chain->error);
2113 * Modify the chain associated with an inode.
2116 hammer2_chain_modify_ip(hammer2_inode_t *ip, hammer2_chain_t *chain,
2117 hammer2_tid_t mtid, int flags)
2121 hammer2_inode_modify(ip);
2122 error = hammer2_chain_modify(chain, mtid, 0, flags);
2128 * Volume header data locks
2131 hammer2_voldata_lock(hammer2_dev_t *hmp)
2133 lockmgr(&hmp->vollk, LK_EXCLUSIVE);
2137 hammer2_voldata_unlock(hammer2_dev_t *hmp)
2139 lockmgr(&hmp->vollk, LK_RELEASE);
2143 hammer2_voldata_modify(hammer2_dev_t *hmp)
2145 if ((hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) == 0) {
2146 atomic_add_long(&hammer2_count_modified_chains, 1);
2147 atomic_set_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
2148 hammer2_pfs_memory_inc(hmp->vchain.pmp);
2153 * This function returns the chain at the nearest key within the specified
2154 * range. The returned chain will be referenced but not locked.
2156 * This function will recurse through chain->rbtree as necessary and will
2157 * return a *key_nextp suitable for iteration. *key_nextp is only set if
2158 * the iteration value is less than the current value of *key_nextp.
2160 * The caller should use (*key_nextp) to calculate the actual range of
2161 * the returned element, which will be (key_beg to *key_nextp - 1), because
2162 * there might be another element which is superior to the returned element
2165 * (*key_nextp) can be passed as key_beg in an iteration only while non-NULL
2166 * chains continue to be returned. On EOF (*key_nextp) may overflow since
2167 * it will wind up being (key_end + 1).
2169 * WARNING! Must be called with child's spinlock held. Spinlock remains
2170 * held through the operation.
2172 struct hammer2_chain_find_info {
2173 hammer2_chain_t *best;
2174 hammer2_key_t key_beg;
2175 hammer2_key_t key_end;
2176 hammer2_key_t key_next;
2179 static int hammer2_chain_find_cmp(hammer2_chain_t *child, void *data);
2180 static int hammer2_chain_find_callback(hammer2_chain_t *child, void *data);
2184 hammer2_chain_find(hammer2_chain_t *parent, hammer2_key_t *key_nextp,
2185 hammer2_key_t key_beg, hammer2_key_t key_end)
2187 struct hammer2_chain_find_info info;
2190 info.key_beg = key_beg;
2191 info.key_end = key_end;
2192 info.key_next = *key_nextp;
2194 RB_SCAN(hammer2_chain_tree, &parent->core.rbtree,
2195 hammer2_chain_find_cmp, hammer2_chain_find_callback,
2197 *key_nextp = info.key_next;
2199 kprintf("chain_find %p %016jx:%016jx next=%016jx\n",
2200 parent, key_beg, key_end, *key_nextp);
2208 hammer2_chain_find_cmp(hammer2_chain_t *child, void *data)
2210 struct hammer2_chain_find_info *info = data;
2211 hammer2_key_t child_beg;
2212 hammer2_key_t child_end;
2214 child_beg = child->bref.key;
2215 child_end = child_beg + ((hammer2_key_t)1 << child->bref.keybits) - 1;
2217 if (child_end < info->key_beg)
2219 if (child_beg > info->key_end)
2226 hammer2_chain_find_callback(hammer2_chain_t *child, void *data)
2228 struct hammer2_chain_find_info *info = data;
2229 hammer2_chain_t *best;
2230 hammer2_key_t child_end;
2233 * WARNING! Layerq is scanned forwards, exact matches should keep
2234 * the existing info->best.
2236 if ((best = info->best) == NULL) {
2238 * No previous best. Assign best
2241 } else if (best->bref.key <= info->key_beg &&
2242 child->bref.key <= info->key_beg) {
2247 /*info->best = child;*/
2248 } else if (child->bref.key < best->bref.key) {
2250 * Child has a nearer key and best is not flush with key_beg.
2251 * Set best to child. Truncate key_next to the old best key.
2254 if (info->key_next > best->bref.key || info->key_next == 0)
2255 info->key_next = best->bref.key;
2256 } else if (child->bref.key == best->bref.key) {
2258 * If our current best is flush with the child then this
2259 * is an illegal overlap.
2261 * key_next will automatically be limited to the smaller of
2262 * the two end-points.
2268 * Keep the current best but truncate key_next to the child's
2271 * key_next will also automatically be limited to the smaller
2272 * of the two end-points (probably not necessary for this case
2273 * but we do it anyway).
2275 if (info->key_next > child->bref.key || info->key_next == 0)
2276 info->key_next = child->bref.key;
2280 * Always truncate key_next based on child's end-of-range.
2282 child_end = child->bref.key + ((hammer2_key_t)1 << child->bref.keybits);
2283 if (child_end && (info->key_next > child_end || info->key_next == 0))
2284 info->key_next = child_end;
2290 * Retrieve the specified chain from a media blockref, creating the
2291 * in-memory chain structure which reflects it. The returned chain is
2292 * held and locked according to (how) (HAMMER2_RESOLVE_*). The caller must
2293 * handle crc-checks and so forth, and should check chain->error before
2294 * assuming that the data is good.
2296 * To handle insertion races pass the INSERT_RACE flag along with the
2297 * generation number of the core. NULL will be returned if the generation
2298 * number changes before we have a chance to insert the chain. Insert
2299 * races can occur because the parent might be held shared.
2301 * Caller must hold the parent locked shared or exclusive since we may
2302 * need the parent's bref array to find our block.
2304 * WARNING! chain->pmp is always set to NULL for any chain representing
2305 * part of the super-root topology.
2308 hammer2_chain_get(hammer2_chain_t *parent, int generation,
2309 hammer2_blockref_t *bref, int how)
2311 hammer2_dev_t *hmp = parent->hmp;
2312 hammer2_chain_t *chain;
2316 * Allocate a chain structure representing the existing media
2317 * entry. Resulting chain has one ref and is not locked.
2319 if (bref->flags & HAMMER2_BREF_FLAG_PFSROOT)
2320 chain = hammer2_chain_alloc(hmp, NULL, bref);
2322 chain = hammer2_chain_alloc(hmp, parent->pmp, bref);
2323 /* ref'd chain returned */
2326 * Flag that the chain is in the parent's blockmap so delete/flush
2327 * knows what to do with it.
2329 atomic_set_int(&chain->flags, HAMMER2_CHAIN_BMAPPED);
2332 * chain must be locked to avoid unexpected ripouts
2334 hammer2_chain_lock(chain, how);
2337 * Link the chain into its parent. A spinlock is required to safely
2338 * access the RBTREE, and it is possible to collide with another
2339 * hammer2_chain_get() operation because the caller might only hold
2340 * a shared lock on the parent.
2342 * NOTE: Get races can occur quite often when we distribute
2343 * asynchronous read-aheads across multiple threads.
2345 KKASSERT(parent->refs > 0);
2346 error = hammer2_chain_insert(parent, chain,
2347 HAMMER2_CHAIN_INSERT_SPIN |
2348 HAMMER2_CHAIN_INSERT_RACE,
2351 KKASSERT((chain->flags & HAMMER2_CHAIN_ONRBTREE) == 0);
2352 /*kprintf("chain %p get race\n", chain);*/
2353 hammer2_chain_unlock(chain);
2354 hammer2_chain_drop(chain);
2357 KKASSERT(chain->flags & HAMMER2_CHAIN_ONRBTREE);
2361 * Return our new chain referenced but not locked, or NULL if
2368 * Lookup initialization/completion API
2371 hammer2_chain_lookup_init(hammer2_chain_t *parent, int flags)
2373 hammer2_chain_ref(parent);
2374 if (flags & HAMMER2_LOOKUP_SHARED) {
2375 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS |
2376 HAMMER2_RESOLVE_SHARED);
2378 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2384 hammer2_chain_lookup_done(hammer2_chain_t *parent)
2387 hammer2_chain_unlock(parent);
2388 hammer2_chain_drop(parent);
2393 * Take the locked chain and return a locked parent. The chain remains
2394 * locked on return, but may have to be temporarily unlocked to acquire
2395 * the parent. Because of this, (chain) must be stable and cannot be
2396 * deleted while it was temporarily unlocked (typically means that (chain)
2399 * Pass HAMMER2_RESOLVE_* flags in flags.
2401 * This will work even if the chain is errored, and the caller can check
2402 * parent->error on return if desired since the parent will be locked.
2404 * This function handles the lock order reversal.
2407 hammer2_chain_getparent(hammer2_chain_t *chain, int flags)
2409 hammer2_chain_t *parent;
2412 * Be careful of order, chain must be unlocked before parent
2413 * is locked below to avoid a deadlock. Try it trivially first.
2415 parent = chain->parent;
2417 panic("hammer2_chain_getparent: no parent");
2418 hammer2_chain_ref(parent);
2419 if (hammer2_chain_lock(parent, flags|HAMMER2_RESOLVE_NONBLOCK) == 0)
2423 hammer2_chain_unlock(chain);
2424 hammer2_chain_lock(parent, flags);
2425 hammer2_chain_lock(chain, flags);
2428 * Parent relinking races are quite common. We have to get
2429 * it right or we will blow up the block table.
2431 if (chain->parent == parent)
2433 hammer2_chain_unlock(parent);
2434 hammer2_chain_drop(parent);
2436 parent = chain->parent;
2438 panic("hammer2_chain_getparent: no parent");
2439 hammer2_chain_ref(parent);
2445 * Take the locked chain and return a locked parent. The chain is unlocked
2446 * and dropped. *chainp is set to the returned parent as a convenience.
2447 * Pass HAMMER2_RESOLVE_* flags in flags.
2449 * This will work even if the chain is errored, and the caller can check
2450 * parent->error on return if desired since the parent will be locked.
2452 * The chain does NOT need to be stable. We use a tracking structure
2453 * to track the expected parent if the chain is deleted out from under us.
2455 * This function handles the lock order reversal.
2458 hammer2_chain_repparent(hammer2_chain_t **chainp, int flags)
2460 hammer2_chain_t *chain;
2461 hammer2_chain_t *parent;
2462 struct hammer2_reptrack reptrack;
2463 struct hammer2_reptrack **repp;
2466 * Be careful of order, chain must be unlocked before parent
2467 * is locked below to avoid a deadlock. Try it trivially first.
2470 parent = chain->parent;
2471 if (parent == NULL) {
2472 hammer2_spin_unex(&chain->core.spin);
2473 panic("hammer2_chain_repparent: no parent");
2475 hammer2_chain_ref(parent);
2476 if (hammer2_chain_lock(parent, flags|HAMMER2_RESOLVE_NONBLOCK) == 0) {
2477 hammer2_chain_unlock(chain);
2478 hammer2_chain_drop(chain);
2485 * Ok, now it gets a bit nasty. There are multiple situations where
2486 * the parent might be in the middle of a deletion, or where the child
2487 * (chain) might be deleted the instant we let go of its lock.
2488 * We can potentially end up in a no-win situation!
2490 * In particular, the indirect_maintenance() case can cause these
2493 * To deal with this we install a reptrack structure in the parent
2494 * This reptrack structure 'owns' the parent ref and will automatically
2495 * migrate to the parent's parent if the parent is deleted permanently.
2497 hammer2_spin_init(&reptrack.spin, "h2reptrk");
2498 reptrack.chain = parent;
2499 hammer2_chain_ref(parent); /* for the reptrack */
2501 hammer2_spin_ex(&parent->core.spin);
2502 reptrack.next = parent->core.reptrack;
2503 parent->core.reptrack = &reptrack;
2504 hammer2_spin_unex(&parent->core.spin);
2506 hammer2_chain_unlock(chain);
2507 hammer2_chain_drop(chain);
2508 chain = NULL; /* gone */
2511 * At the top of this loop, chain is gone and parent is refd both
2512 * by us explicitly AND via our reptrack. We are attempting to
2516 hammer2_chain_lock(parent, flags);
2518 if (reptrack.chain == parent)
2520 hammer2_chain_unlock(parent);
2521 hammer2_chain_drop(parent);
2523 kprintf("hammer2: debug REPTRACK %p->%p\n",
2524 parent, reptrack.chain);
2525 hammer2_spin_ex(&reptrack.spin);
2526 parent = reptrack.chain;
2527 hammer2_chain_ref(parent);
2528 hammer2_spin_unex(&reptrack.spin);
2532 * Once parent is locked and matches our reptrack, our reptrack
2533 * will be stable and we have our parent. We can unlink our
2536 * WARNING! Remember that the chain lock might be shared. Chains
2537 * locked shared have stable parent linkages.
2539 hammer2_spin_ex(&parent->core.spin);
2540 repp = &parent->core.reptrack;
2541 while (*repp != &reptrack)
2542 repp = &(*repp)->next;
2543 *repp = reptrack.next;
2544 hammer2_spin_unex(&parent->core.spin);
2546 hammer2_chain_drop(parent); /* reptrack ref */
2547 *chainp = parent; /* return parent lock+ref */
2553 * Dispose of any linked reptrack structures in (chain) by shifting them to
2554 * (parent). Both (chain) and (parent) must be exclusively locked.
2556 * This is interlocked against any children of (chain) on the other side.
2557 * No children so remain as-of when this is called so we can test
2558 * core.reptrack without holding the spin-lock.
2560 * Used whenever the caller intends to permanently delete chains related
2561 * to topological recursions (BREF_TYPE_INDIRECT, BREF_TYPE_FREEMAP_NODE),
2562 * where the chains underneath the node being deleted are given a new parent
2563 * above the node being deleted.
2567 hammer2_chain_repchange(hammer2_chain_t *parent, hammer2_chain_t *chain)
2569 struct hammer2_reptrack *reptrack;
2571 KKASSERT(chain->core.live_count == 0 && RB_EMPTY(&chain->core.rbtree));
2572 while (chain->core.reptrack) {
2573 hammer2_spin_ex(&parent->core.spin);
2574 hammer2_spin_ex(&chain->core.spin);
2575 reptrack = chain->core.reptrack;
2576 if (reptrack == NULL) {
2577 hammer2_spin_unex(&chain->core.spin);
2578 hammer2_spin_unex(&parent->core.spin);
2581 hammer2_spin_ex(&reptrack->spin);
2582 chain->core.reptrack = reptrack->next;
2583 reptrack->chain = parent;
2584 reptrack->next = parent->core.reptrack;
2585 parent->core.reptrack = reptrack;
2586 hammer2_chain_ref(parent); /* reptrack */
2588 hammer2_spin_unex(&chain->core.spin);
2589 hammer2_spin_unex(&parent->core.spin);
2590 kprintf("hammer2: debug repchange %p %p->%p\n",
2591 reptrack, chain, parent);
2592 hammer2_chain_drop(chain); /* reptrack */
2597 * Locate the first chain whos key range overlaps (key_beg, key_end) inclusive.
2598 * (*parentp) typically points to an inode but can also point to a related
2599 * indirect block and this function will recurse upwards and find the inode
2600 * or the nearest undeleted indirect block covering the key range.
2602 * This function unconditionally sets *errorp, replacing any previous value.
2604 * (*parentp) must be exclusive or shared locked (depending on flags) and
2605 * referenced and can be an inode or an existing indirect block within the
2608 * If (*parent) is errored out, this function will not attempt to recurse
2609 * the radix tree and will return NULL along with an appropriate *errorp.
2610 * If NULL is returned and *errorp is 0, the requested lookup could not be
2613 * On return (*parentp) will be modified to point at the deepest parent chain
2614 * element encountered during the search, as a helper for an insertion or
2617 * The new (*parentp) will be locked shared or exclusive (depending on flags),
2618 * and referenced, and the old will be unlocked and dereferenced (no change
2619 * if they are both the same). This is particularly important if the caller
2620 * wishes to insert a new chain, (*parentp) will be set properly even if NULL
2621 * is returned, as long as no error occurred.
2623 * The matching chain will be returned locked according to flags.
2627 * NULL is returned if no match was found, but (*parentp) will still
2628 * potentially be adjusted.
2630 * On return (*key_nextp) will point to an iterative value for key_beg.
2631 * (If NULL is returned (*key_nextp) is set to (key_end + 1)).
2633 * This function will also recurse up the chain if the key is not within the
2634 * current parent's range. (*parentp) can never be set to NULL. An iteration
2635 * can simply allow (*parentp) to float inside the loop.
2637 * NOTE! chain->data is not always resolved. By default it will not be
2638 * resolved for BREF_TYPE_DATA, FREEMAP_NODE, or FREEMAP_LEAF. Use
2639 * HAMMER2_LOOKUP_ALWAYS to force resolution (but be careful w/
2640 * BREF_TYPE_DATA as the device buffer can alias the logical file
2645 hammer2_chain_lookup(hammer2_chain_t **parentp, hammer2_key_t *key_nextp,
2646 hammer2_key_t key_beg, hammer2_key_t key_end,
2647 int *errorp, int flags)
2650 hammer2_chain_t *parent;
2651 hammer2_chain_t *chain;
2652 hammer2_blockref_t *base;
2653 hammer2_blockref_t *bref;
2654 hammer2_blockref_t bsave;
2655 hammer2_key_t scan_beg;
2656 hammer2_key_t scan_end;
2658 int how_always = HAMMER2_RESOLVE_ALWAYS;
2659 int how_maybe = HAMMER2_RESOLVE_MAYBE;
2662 int maxloops = 300000;
2663 volatile hammer2_mtx_t save_mtx;
2665 if (flags & HAMMER2_LOOKUP_ALWAYS) {
2666 how_maybe = how_always;
2667 how = HAMMER2_RESOLVE_ALWAYS;
2668 } else if (flags & HAMMER2_LOOKUP_NODATA) {
2669 how = HAMMER2_RESOLVE_NEVER;
2671 how = HAMMER2_RESOLVE_MAYBE;
2673 if (flags & HAMMER2_LOOKUP_SHARED) {
2674 how_maybe |= HAMMER2_RESOLVE_SHARED;
2675 how_always |= HAMMER2_RESOLVE_SHARED;
2676 how |= HAMMER2_RESOLVE_SHARED;
2680 * Recurse (*parentp) upward if necessary until the parent completely
2681 * encloses the key range or we hit the inode.
2683 * Handle races against the flusher deleting indirect nodes on its
2684 * way back up by continuing to recurse upward past the deletion.
2690 while (parent->bref.type == HAMMER2_BREF_TYPE_INDIRECT ||
2691 parent->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2692 scan_beg = parent->bref.key;
2693 scan_end = scan_beg +
2694 ((hammer2_key_t)1 << parent->bref.keybits) - 1;
2695 if ((parent->flags & HAMMER2_CHAIN_DELETED) == 0) {
2696 if (key_beg >= scan_beg && key_end <= scan_end)
2699 parent = hammer2_chain_repparent(parentp, how_maybe);
2702 if (--maxloops == 0)
2703 panic("hammer2_chain_lookup: maxloops");
2706 * MATCHIND case that does not require parent->data (do prior to
2707 * parent->error check).
2709 switch(parent->bref.type) {
2710 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2711 case HAMMER2_BREF_TYPE_INDIRECT:
2712 if (flags & HAMMER2_LOOKUP_MATCHIND) {
2713 scan_beg = parent->bref.key;
2714 scan_end = scan_beg +
2715 ((hammer2_key_t)1 << parent->bref.keybits) - 1;
2716 if (key_beg == scan_beg && key_end == scan_end) {
2718 hammer2_chain_ref(chain);
2719 hammer2_chain_lock(chain, how_maybe);
2720 *key_nextp = scan_end + 1;
2730 * No lookup is possible if the parent is errored. We delayed
2731 * this check as long as we could to ensure that the parent backup,
2732 * embedded data, and MATCHIND code could still execute.
2734 if (parent->error) {
2735 *errorp = parent->error;
2740 * Locate the blockref array. Currently we do a fully associative
2741 * search through the array.
2743 switch(parent->bref.type) {
2744 case HAMMER2_BREF_TYPE_INODE:
2746 * Special shortcut for embedded data returns the inode
2747 * itself. Callers must detect this condition and access
2748 * the embedded data (the strategy code does this for us).
2750 * This is only applicable to regular files and softlinks.
2752 * We need a second lock on parent. Since we already have
2753 * a lock we must pass LOCKAGAIN to prevent unexpected
2754 * blocking (we don't want to block on a second shared
2755 * ref if an exclusive lock is pending)
2757 if (parent->data->ipdata.meta.op_flags &
2758 HAMMER2_OPFLAG_DIRECTDATA) {
2759 if (flags & HAMMER2_LOOKUP_NODIRECT) {
2761 *key_nextp = key_end + 1;
2764 hammer2_chain_ref(parent);
2765 hammer2_chain_lock(parent, how_always |
2766 HAMMER2_RESOLVE_LOCKAGAIN);
2767 *key_nextp = key_end + 1;
2770 base = &parent->data->ipdata.u.blockset.blockref[0];
2771 count = HAMMER2_SET_COUNT;
2773 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2774 case HAMMER2_BREF_TYPE_INDIRECT:
2776 * Optimize indirect blocks in the INITIAL state to avoid
2779 * Debugging: Enter permanent wait state instead of
2780 * panicing on unexpectedly NULL data for the moment.
2782 if (parent->flags & HAMMER2_CHAIN_INITIAL) {
2785 if (parent->data == NULL) {
2786 kprintf("hammer2: unexpected NULL data "
2789 tsleep(parent, 0, "xxx", 0);
2791 base = &parent->data->npdata[0];
2793 count = parent->bytes / sizeof(hammer2_blockref_t);
2795 case HAMMER2_BREF_TYPE_VOLUME:
2796 base = &parent->data->voldata.sroot_blockset.blockref[0];
2797 count = HAMMER2_SET_COUNT;
2799 case HAMMER2_BREF_TYPE_FREEMAP:
2800 base = &parent->data->blkset.blockref[0];
2801 count = HAMMER2_SET_COUNT;
2804 kprintf("hammer2_chain_lookup: unrecognized "
2805 "blockref(B) type: %d",
2808 tsleep(&base, 0, "dead", 0);
2809 panic("hammer2_chain_lookup: unrecognized "
2810 "blockref(B) type: %d",
2812 base = NULL; /* safety */
2813 count = 0; /* safety */
2817 * Merged scan to find next candidate.
2819 * hammer2_base_*() functions require the parent->core.live_* fields
2820 * to be synchronized.
2822 * We need to hold the spinlock to access the block array and RB tree
2823 * and to interlock chain creation.
2825 if ((parent->flags & HAMMER2_CHAIN_COUNTEDBREFS) == 0)
2826 hammer2_chain_countbrefs(parent, base, count);
2831 hammer2_spin_ex(&parent->core.spin);
2832 chain = hammer2_combined_find(parent, base, count,
2836 generation = parent->core.generation;
2839 * Exhausted parent chain, iterate.
2842 KKASSERT(chain == NULL);
2843 hammer2_spin_unex(&parent->core.spin);
2844 if (key_beg == key_end) /* short cut single-key case */
2848 * Stop if we reached the end of the iteration.
2850 if (parent->bref.type != HAMMER2_BREF_TYPE_INDIRECT &&
2851 parent->bref.type != HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2856 * Calculate next key, stop if we reached the end of the
2857 * iteration, otherwise go up one level and loop.
2859 key_beg = parent->bref.key +
2860 ((hammer2_key_t)1 << parent->bref.keybits);
2861 if (key_beg == 0 || key_beg > key_end)
2863 parent = hammer2_chain_repparent(parentp, how_maybe);
2868 * Selected from blockref or in-memory chain.
2871 if (chain == NULL) {
2872 hammer2_spin_unex(&parent->core.spin);
2873 if (bsave.type == HAMMER2_BREF_TYPE_INDIRECT ||
2874 bsave.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2875 chain = hammer2_chain_get(parent, generation,
2878 chain = hammer2_chain_get(parent, generation,
2884 hammer2_chain_ref(chain);
2885 hammer2_spin_unex(&parent->core.spin);
2888 * chain is referenced but not locked. We must lock the
2889 * chain to obtain definitive state.
2891 if (bsave.type == HAMMER2_BREF_TYPE_INDIRECT ||
2892 bsave.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2893 hammer2_chain_lock(chain, how_maybe);
2895 hammer2_chain_lock(chain, how);
2897 KKASSERT(chain->parent == parent);
2899 if (bcmp(&bsave, &chain->bref, sizeof(bsave)) ||
2900 chain->parent != parent) {
2901 hammer2_chain_unlock(chain);
2902 hammer2_chain_drop(chain);
2903 chain = NULL; /* SAFETY */
2909 * Skip deleted chains (XXX cache 'i' end-of-block-array? XXX)
2911 * NOTE: Chain's key range is not relevant as there might be
2912 * one-offs within the range that are not deleted.
2914 * NOTE: Lookups can race delete-duplicate because
2915 * delete-duplicate does not lock the parent's core
2916 * (they just use the spinlock on the core).
2918 if (chain->flags & HAMMER2_CHAIN_DELETED) {
2919 kprintf("skip deleted chain %016jx.%02x key=%016jx\n",
2920 chain->bref.data_off, chain->bref.type,
2922 hammer2_chain_unlock(chain);
2923 hammer2_chain_drop(chain);
2924 chain = NULL; /* SAFETY */
2925 key_beg = *key_nextp;
2926 if (key_beg == 0 || key_beg > key_end)
2932 * If the chain element is an indirect block it becomes the new
2933 * parent and we loop on it. We must maintain our top-down locks
2934 * to prevent the flusher from interfering (i.e. doing a
2935 * delete-duplicate and leaving us recursing down a deleted chain).
2937 * The parent always has to be locked with at least RESOLVE_MAYBE
2938 * so we can access its data. It might need a fixup if the caller
2939 * passed incompatible flags. Be careful not to cause a deadlock
2940 * as a data-load requires an exclusive lock.
2942 * If HAMMER2_LOOKUP_MATCHIND is set and the indirect block's key
2943 * range is within the requested key range we return the indirect
2944 * block and do NOT loop. This is usually only used to acquire
2947 if (chain->bref.type == HAMMER2_BREF_TYPE_INDIRECT ||
2948 chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2949 save_mtx = parent->lock;
2950 hammer2_chain_unlock(parent);
2951 hammer2_chain_drop(parent);
2952 *parentp = parent = chain;
2953 chain = NULL; /* SAFETY */
2958 * All done, return the locked chain.
2960 * If the caller does not want a locked chain, replace the lock with
2961 * a ref. Perhaps this can eventually be optimized to not obtain the
2962 * lock in the first place for situations where the data does not
2963 * need to be resolved.
2965 * NOTE! A chain->error must be tested by the caller upon return.
2966 * *errorp is only set based on issues which occur while
2967 * trying to reach the chain.
2973 * After having issued a lookup we can iterate all matching keys.
2975 * If chain is non-NULL we continue the iteration from just after it's index.
2977 * If chain is NULL we assume the parent was exhausted and continue the
2978 * iteration at the next parent.
2980 * If a fatal error occurs (typically an I/O error), a dummy chain is
2981 * returned with chain->error and error-identifying information set. This
2982 * chain will assert if you try to do anything fancy with it.
2984 * XXX Depending on where the error occurs we should allow continued iteration.
2986 * parent must be locked on entry and remains locked throughout. chain's
2987 * lock status must match flags. Chain is always at least referenced.
2989 * WARNING! The MATCHIND flag does not apply to this function.
2992 hammer2_chain_next(hammer2_chain_t **parentp, hammer2_chain_t *chain,
2993 hammer2_key_t *key_nextp,
2994 hammer2_key_t key_beg, hammer2_key_t key_end,
2995 int *errorp, int flags)
2997 hammer2_chain_t *parent;
3001 * Calculate locking flags for upward recursion.
3003 how_maybe = HAMMER2_RESOLVE_MAYBE;
3004 if (flags & HAMMER2_LOOKUP_SHARED)
3005 how_maybe |= HAMMER2_RESOLVE_SHARED;
3011 * Calculate the next index and recalculate the parent if necessary.
3014 key_beg = chain->bref.key +
3015 ((hammer2_key_t)1 << chain->bref.keybits);
3016 hammer2_chain_unlock(chain);
3017 hammer2_chain_drop(chain);
3020 * chain invalid past this point, but we can still do a
3021 * pointer comparison w/parent.
3023 * Any scan where the lookup returned degenerate data embedded
3024 * in the inode has an invalid index and must terminate.
3026 if (chain == parent)
3028 if (key_beg == 0 || key_beg > key_end)
3031 } else if (parent->bref.type != HAMMER2_BREF_TYPE_INDIRECT &&
3032 parent->bref.type != HAMMER2_BREF_TYPE_FREEMAP_NODE) {
3034 * We reached the end of the iteration.
3039 * Continue iteration with next parent unless the current
3040 * parent covers the range.
3042 * (This also handles the case of a deleted, empty indirect
3045 key_beg = parent->bref.key +
3046 ((hammer2_key_t)1 << parent->bref.keybits);
3047 if (key_beg == 0 || key_beg > key_end)
3049 parent = hammer2_chain_repparent(parentp, how_maybe);
3055 return (hammer2_chain_lookup(parentp, key_nextp,
3061 * Caller wishes to iterate chains under parent, loading new chains into
3062 * chainp. Caller must initialize *chainp to NULL and *firstp to 1, and
3063 * then call hammer2_chain_scan() repeatedly until a non-zero return.
3064 * During the scan, *firstp will be set to 0 and (*chainp) will be replaced
3065 * with the returned chain for the scan. The returned *chainp will be
3066 * locked and referenced. Any prior contents will be unlocked and dropped.
3068 * Caller should check the return value. A normal scan EOF will return
3069 * exactly HAMMER2_ERROR_EOF. Any other non-zero value indicates an
3070 * error trying to access parent data. Any error in the returned chain
3071 * must be tested separately by the caller.
3073 * (*chainp) is dropped on each scan, but will only be set if the returned
3074 * element itself can recurse. Leaf elements are NOT resolved, loaded, or
3075 * returned via *chainp. The caller will get their bref only.
3077 * The raw scan function is similar to lookup/next but does not seek to a key.
3078 * Blockrefs are iterated via first_bref = (parent, NULL) and
3079 * next_chain = (parent, bref).
3081 * The passed-in parent must be locked and its data resolved. The function
3082 * nominally returns a locked and referenced *chainp != NULL for chains
3083 * the caller might need to recurse on (and will dipose of any *chainp passed
3084 * in). The caller must check the chain->bref.type either way.
3087 hammer2_chain_scan(hammer2_chain_t *parent, hammer2_chain_t **chainp,
3088 hammer2_blockref_t *bref, int *firstp,
3092 hammer2_blockref_t *base;
3093 hammer2_blockref_t *bref_ptr;
3095 hammer2_key_t next_key;
3096 hammer2_chain_t *chain = NULL;
3098 int how_always = HAMMER2_RESOLVE_ALWAYS;
3099 int how_maybe = HAMMER2_RESOLVE_MAYBE;
3102 int maxloops = 300000;
3109 * Scan flags borrowed from lookup.
3111 if (flags & HAMMER2_LOOKUP_ALWAYS) {
3112 how_maybe = how_always;
3113 how = HAMMER2_RESOLVE_ALWAYS;
3114 } else if (flags & HAMMER2_LOOKUP_NODATA) {
3115 how = HAMMER2_RESOLVE_NEVER;
3117 how = HAMMER2_RESOLVE_MAYBE;
3119 if (flags & HAMMER2_LOOKUP_SHARED) {
3120 how_maybe |= HAMMER2_RESOLVE_SHARED;
3121 how_always |= HAMMER2_RESOLVE_SHARED;
3122 how |= HAMMER2_RESOLVE_SHARED;
3126 * Calculate key to locate first/next element, unlocking the previous
3127 * element as we go. Be careful, the key calculation can overflow.
3129 * (also reset bref to NULL)
3135 key = bref->key + ((hammer2_key_t)1 << bref->keybits);
3136 if ((chain = *chainp) != NULL) {
3138 hammer2_chain_unlock(chain);
3139 hammer2_chain_drop(chain);
3143 error |= HAMMER2_ERROR_EOF;
3149 if (parent->error) {
3150 error = parent->error;
3153 if (--maxloops == 0)
3154 panic("hammer2_chain_scan: maxloops");
3157 * Locate the blockref array. Currently we do a fully associative
3158 * search through the array.
3160 switch(parent->bref.type) {
3161 case HAMMER2_BREF_TYPE_INODE:
3163 * An inode with embedded data has no sub-chains.
3165 * WARNING! Bulk scan code may pass a static chain marked
3166 * as BREF_TYPE_INODE with a copy of the volume
3167 * root blockset to snapshot the volume.
3169 if (parent->data->ipdata.meta.op_flags &
3170 HAMMER2_OPFLAG_DIRECTDATA) {
3171 error |= HAMMER2_ERROR_EOF;
3174 base = &parent->data->ipdata.u.blockset.blockref[0];
3175 count = HAMMER2_SET_COUNT;
3177 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3178 case HAMMER2_BREF_TYPE_INDIRECT:
3180 * Optimize indirect blocks in the INITIAL state to avoid
3183 if (parent->flags & HAMMER2_CHAIN_INITIAL) {
3186 if (parent->data == NULL)
3187 panic("parent->data is NULL");
3188 base = &parent->data->npdata[0];
3190 count = parent->bytes / sizeof(hammer2_blockref_t);
3192 case HAMMER2_BREF_TYPE_VOLUME:
3193 base = &parent->data->voldata.sroot_blockset.blockref[0];
3194 count = HAMMER2_SET_COUNT;
3196 case HAMMER2_BREF_TYPE_FREEMAP:
3197 base = &parent->data->blkset.blockref[0];
3198 count = HAMMER2_SET_COUNT;
3201 panic("hammer2_chain_scan: unrecognized blockref type: %d",
3203 base = NULL; /* safety */
3204 count = 0; /* safety */
3208 * Merged scan to find next candidate.
3210 * hammer2_base_*() functions require the parent->core.live_* fields
3211 * to be synchronized.
3213 * We need to hold the spinlock to access the block array and RB tree
3214 * and to interlock chain creation.
3216 if ((parent->flags & HAMMER2_CHAIN_COUNTEDBREFS) == 0)
3217 hammer2_chain_countbrefs(parent, base, count);
3221 hammer2_spin_ex(&parent->core.spin);
3222 chain = hammer2_combined_find(parent, base, count,
3224 key, HAMMER2_KEY_MAX,
3226 generation = parent->core.generation;
3229 * Exhausted parent chain, we're done.
3231 if (bref_ptr == NULL) {
3232 hammer2_spin_unex(&parent->core.spin);
3233 KKASSERT(chain == NULL);
3234 error |= HAMMER2_ERROR_EOF;
3239 * Copy into the supplied stack-based blockref.
3244 * Selected from blockref or in-memory chain.
3246 if (chain == NULL) {
3247 switch(bref->type) {
3248 case HAMMER2_BREF_TYPE_INODE:
3249 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3250 case HAMMER2_BREF_TYPE_INDIRECT:
3251 case HAMMER2_BREF_TYPE_VOLUME:
3252 case HAMMER2_BREF_TYPE_FREEMAP:
3254 * Recursion, always get the chain
3256 hammer2_spin_unex(&parent->core.spin);
3257 chain = hammer2_chain_get(parent, generation,
3264 * No recursion, do not waste time instantiating
3265 * a chain, just iterate using the bref.
3267 hammer2_spin_unex(&parent->core.spin);
3272 * Recursion or not we need the chain in order to supply
3275 hammer2_chain_ref(chain);
3276 hammer2_spin_unex(&parent->core.spin);
3277 hammer2_chain_lock(chain, how);
3280 (bcmp(bref, &chain->bref, sizeof(*bref)) ||
3281 chain->parent != parent)) {
3282 hammer2_chain_unlock(chain);
3283 hammer2_chain_drop(chain);
3289 * Skip deleted chains (XXX cache 'i' end-of-block-array? XXX)
3291 * NOTE: chain's key range is not relevant as there might be
3292 * one-offs within the range that are not deleted.
3294 * NOTE: XXX this could create problems with scans used in
3295 * situations other than mount-time recovery.
3297 * NOTE: Lookups can race delete-duplicate because
3298 * delete-duplicate does not lock the parent's core
3299 * (they just use the spinlock on the core).
3301 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) {
3302 hammer2_chain_unlock(chain);
3303 hammer2_chain_drop(chain);
3308 error |= HAMMER2_ERROR_EOF;
3316 * All done, return the bref or NULL, supply chain if necessary.
3324 * Create and return a new hammer2 system memory structure of the specified
3325 * key, type and size and insert it under (*parentp). This is a full
3326 * insertion, based on the supplied key/keybits, and may involve creating
3327 * indirect blocks and moving other chains around via delete/duplicate.
3329 * This call can be made with parent == NULL as long as a non -1 methods
3330 * is supplied. hmp must also be supplied in this situation (otherwise
3331 * hmp is extracted from the supplied parent). The chain will be detached
3332 * from the topology. A later call with both parent and chain can be made
3335 * THE CALLER MUST HAVE ALREADY PROPERLY SEEKED (*parentp) TO THE INSERTION
3336 * POINT SANS ANY REQUIRED INDIRECT BLOCK CREATIONS DUE TO THE ARRAY BEING
3337 * FULL. This typically means that the caller is creating the chain after
3338 * doing a hammer2_chain_lookup().
3340 * (*parentp) must be exclusive locked and may be replaced on return
3341 * depending on how much work the function had to do.
3343 * (*parentp) must not be errored or this function will assert.
3345 * (*chainp) usually starts out NULL and returns the newly created chain,
3346 * but if the caller desires the caller may allocate a disconnected chain
3347 * and pass it in instead.
3349 * This function should NOT be used to insert INDIRECT blocks. It is
3350 * typically used to create/insert inodes and data blocks.
3352 * Caller must pass-in an exclusively locked parent the new chain is to
3353 * be inserted under, and optionally pass-in a disconnected, exclusively
3354 * locked chain to insert (else we create a new chain). The function will
3355 * adjust (*parentp) as necessary, create or connect the chain, and
3356 * return an exclusively locked chain in *chainp.
3358 * When creating a PFSROOT inode under the super-root, pmp is typically NULL
3359 * and will be reassigned.
3361 * NOTE: returns HAMMER_ERROR_* flags
3364 hammer2_chain_create(hammer2_chain_t **parentp, hammer2_chain_t **chainp,
3365 hammer2_dev_t *hmp, hammer2_pfs_t *pmp, int methods,
3366 hammer2_key_t key, int keybits, int type, size_t bytes,
3367 hammer2_tid_t mtid, hammer2_off_t dedup_off, int flags)
3369 hammer2_chain_t *chain;
3370 hammer2_chain_t *parent;
3371 hammer2_blockref_t *base;
3372 hammer2_blockref_t dummy;
3376 int maxloops = 300000;
3379 * Topology may be crossing a PFS boundary.
3383 KKASSERT(hammer2_mtx_owned(&parent->lock));
3384 KKASSERT(parent->error == 0);
3389 if (chain == NULL) {
3391 * First allocate media space and construct the dummy bref,
3392 * then allocate the in-memory chain structure. Set the
3393 * INITIAL flag for fresh chains which do not have embedded
3396 * XXX for now set the check mode of the child based on
3397 * the parent or, if the parent is an inode, the
3398 * specification in the inode.
3400 bzero(&dummy, sizeof(dummy));
3403 dummy.keybits = keybits;
3404 dummy.data_off = hammer2_getradix(bytes);
3407 * Inherit methods from parent by default. Primarily used
3408 * for BREF_TYPE_DATA. Non-data types *must* be set to
3409 * a non-NONE check algorithm.
3412 dummy.methods = parent->bref.methods;
3414 dummy.methods = (uint8_t)methods;
3416 if (type != HAMMER2_BREF_TYPE_DATA &&
3417 HAMMER2_DEC_CHECK(dummy.methods) == HAMMER2_CHECK_NONE) {
3419 HAMMER2_ENC_CHECK(HAMMER2_CHECK_DEFAULT);
3422 chain = hammer2_chain_alloc(hmp, pmp, &dummy);
3425 * Lock the chain manually, chain_lock will load the chain
3426 * which we do NOT want to do. (note: chain->refs is set
3427 * to 1 by chain_alloc() for us, but lockcnt is not).
3430 hammer2_mtx_ex(&chain->lock);
3434 * Set INITIAL to optimize I/O. The flag will generally be
3435 * processed when we call hammer2_chain_modify().
3437 * Recalculate bytes to reflect the actual media block
3438 * allocation. Handle special case radix 0 == 0 bytes.
3440 bytes = (size_t)(chain->bref.data_off & HAMMER2_OFF_MASK_RADIX);
3442 bytes = (hammer2_off_t)1 << bytes;
3443 chain->bytes = bytes;
3446 case HAMMER2_BREF_TYPE_VOLUME:
3447 case HAMMER2_BREF_TYPE_FREEMAP:
3448 panic("hammer2_chain_create: called with volume type");
3450 case HAMMER2_BREF_TYPE_INDIRECT:
3451 panic("hammer2_chain_create: cannot be used to"
3452 "create indirect block");
3454 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3455 panic("hammer2_chain_create: cannot be used to"
3456 "create freemap root or node");
3458 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
3459 KKASSERT(bytes == sizeof(chain->data->bmdata));
3461 case HAMMER2_BREF_TYPE_DIRENT:
3462 case HAMMER2_BREF_TYPE_INODE:
3463 case HAMMER2_BREF_TYPE_DATA:
3466 * leave chain->data NULL, set INITIAL
3468 KKASSERT(chain->data == NULL);
3469 atomic_set_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
3474 * We are reattaching a previously deleted chain, possibly
3475 * under a new parent and possibly with a new key/keybits.
3476 * The chain does not have to be in a modified state. The
3477 * UPDATE flag will be set later on in this routine.
3479 * Do NOT mess with the current state of the INITIAL flag.
3481 chain->bref.key = key;
3482 chain->bref.keybits = keybits;
3483 if (chain->flags & HAMMER2_CHAIN_DELETED)
3484 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_DELETED);
3485 KKASSERT(chain->parent == NULL);
3489 * Set the appropriate bref flag if requested.
3491 * NOTE! Callers can call this function to move chains without
3492 * knowing about special flags, so don't clear bref flags
3495 if (flags & HAMMER2_INSERT_PFSROOT)
3496 chain->bref.flags |= HAMMER2_BREF_FLAG_PFSROOT;
3502 * Calculate how many entries we have in the blockref array and
3503 * determine if an indirect block is required when inserting into
3507 if (--maxloops == 0)
3508 panic("hammer2_chain_create: maxloops");
3510 switch(parent->bref.type) {
3511 case HAMMER2_BREF_TYPE_INODE:
3512 if ((parent->data->ipdata.meta.op_flags &
3513 HAMMER2_OPFLAG_DIRECTDATA) != 0) {
3514 kprintf("hammer2: parent set for direct-data! "
3515 "pkey=%016jx ckey=%016jx\n",
3519 KKASSERT((parent->data->ipdata.meta.op_flags &
3520 HAMMER2_OPFLAG_DIRECTDATA) == 0);
3521 KKASSERT(parent->data != NULL);
3522 base = &parent->data->ipdata.u.blockset.blockref[0];
3523 count = HAMMER2_SET_COUNT;
3525 case HAMMER2_BREF_TYPE_INDIRECT:
3526 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3527 if (parent->flags & HAMMER2_CHAIN_INITIAL)
3530 base = &parent->data->npdata[0];
3531 count = parent->bytes / sizeof(hammer2_blockref_t);
3533 case HAMMER2_BREF_TYPE_VOLUME:
3534 KKASSERT(parent->data != NULL);
3535 base = &parent->data->voldata.sroot_blockset.blockref[0];
3536 count = HAMMER2_SET_COUNT;
3538 case HAMMER2_BREF_TYPE_FREEMAP:
3539 KKASSERT(parent->data != NULL);
3540 base = &parent->data->blkset.blockref[0];
3541 count = HAMMER2_SET_COUNT;
3544 panic("hammer2_chain_create: unrecognized blockref type: %d",
3552 * Make sure we've counted the brefs
3554 if ((parent->flags & HAMMER2_CHAIN_COUNTEDBREFS) == 0)
3555 hammer2_chain_countbrefs(parent, base, count);
3557 KASSERT(parent->core.live_count >= 0 &&
3558 parent->core.live_count <= count,
3559 ("bad live_count %d/%d (%02x, %d)",
3560 parent->core.live_count, count,
3561 parent->bref.type, parent->bytes));
3564 * If no free blockref could be found we must create an indirect
3565 * block and move a number of blockrefs into it. With the parent
3566 * locked we can safely lock each child in order to delete+duplicate
3567 * it without causing a deadlock.
3569 * This may return the new indirect block or the old parent depending
3570 * on where the key falls. NULL is returned on error.
3572 if (parent->core.live_count == count) {
3573 hammer2_chain_t *nparent;
3575 KKASSERT((flags & HAMMER2_INSERT_SAMEPARENT) == 0);
3577 nparent = hammer2_chain_create_indirect(parent, key, keybits,
3578 mtid, type, &error);
3579 if (nparent == NULL) {
3581 hammer2_chain_drop(chain);
3585 if (parent != nparent) {
3586 hammer2_chain_unlock(parent);
3587 hammer2_chain_drop(parent);
3588 parent = *parentp = nparent;
3594 * fall through if parent, or skip to here if no parent.
3597 if (chain->flags & HAMMER2_CHAIN_DELETED)
3598 kprintf("Inserting deleted chain @%016jx\n",
3602 * Link the chain into its parent.
3604 if (chain->parent != NULL)
3605 panic("hammer2: hammer2_chain_create: chain already connected");
3606 KKASSERT(chain->parent == NULL);
3608 KKASSERT(parent->core.live_count < count);
3609 hammer2_chain_insert(parent, chain,
3610 HAMMER2_CHAIN_INSERT_SPIN |
3611 HAMMER2_CHAIN_INSERT_LIVE,
3617 * Mark the newly created chain modified. This will cause
3618 * UPDATE to be set and process the INITIAL flag.
3620 * Device buffers are not instantiated for DATA elements
3621 * as these are handled by logical buffers.
3623 * Indirect and freemap node indirect blocks are handled
3624 * by hammer2_chain_create_indirect() and not by this
3627 * Data for all other bref types is expected to be
3628 * instantiated (INODE, LEAF).
3630 switch(chain->bref.type) {
3631 case HAMMER2_BREF_TYPE_DATA:
3632 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
3633 case HAMMER2_BREF_TYPE_DIRENT:
3634 case HAMMER2_BREF_TYPE_INODE:
3635 error = hammer2_chain_modify(chain, mtid, dedup_off,
3636 HAMMER2_MODIFY_OPTDATA);
3640 * Remaining types are not supported by this function.
3641 * In particular, INDIRECT and LEAF_NODE types are
3642 * handled by create_indirect().
3644 panic("hammer2_chain_create: bad type: %d",
3651 * When reconnecting a chain we must set UPDATE and
3652 * setflush so the flush recognizes that it must update
3653 * the bref in the parent.
3655 if ((chain->flags & HAMMER2_CHAIN_UPDATE) == 0)
3656 atomic_set_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
3660 * We must setflush(parent) to ensure that it recurses through to
3661 * chain. setflush(chain) might not work because ONFLUSH is possibly
3662 * already set in the chain (so it won't recurse up to set it in the
3666 hammer2_chain_setflush(parent);
3675 * Move the chain from its old parent to a new parent. The chain must have
3676 * already been deleted or already disconnected (or never associated) with
3677 * a parent. The chain is reassociated with the new parent and the deleted
3678 * flag will be cleared (no longer deleted). The chain's modification state
3681 * THE CALLER MUST HAVE ALREADY PROPERLY SEEKED (parent) TO THE INSERTION
3682 * POINT SANS ANY REQUIRED INDIRECT BLOCK CREATIONS DUE TO THE ARRAY BEING
3683 * FULL. This typically means that the caller is creating the chain after
3684 * doing a hammer2_chain_lookup().
3686 * Neither (parent) or (chain) can be errored.
3688 * If (parent) is non-NULL then the chain is inserted under the parent.
3690 * If (parent) is NULL then the newly duplicated chain is not inserted
3691 * anywhere, similar to if it had just been chain_alloc()'d (suitable for
3692 * passing into hammer2_chain_create() after this function returns).
3694 * WARNING! This function calls create which means it can insert indirect
3695 * blocks. This can cause other unrelated chains in the parent to
3696 * be moved to a newly inserted indirect block in addition to the
3700 hammer2_chain_rename(hammer2_chain_t **parentp, hammer2_chain_t *chain,
3701 hammer2_tid_t mtid, int flags)
3703 hammer2_blockref_t *bref;
3705 hammer2_chain_t *parent;
3709 * WARNING! We should never resolve DATA to device buffers
3710 * (XXX allow it if the caller did?), and since
3711 * we currently do not have the logical buffer cache
3712 * buffer in-hand to fix its cached physical offset
3713 * we also force the modify code to not COW it. XXX
3715 * NOTE! We allow error'd chains to be renamed. The bref itself
3716 * is good and can be renamed. The content, however, may
3720 KKASSERT(chain->parent == NULL);
3721 /*KKASSERT(chain->error == 0); allow */
3724 * Now create a duplicate of the chain structure, associating
3725 * it with the same core, making it the same size, pointing it
3726 * to the same bref (the same media block).
3728 * NOTE: Handle special radix == 0 case (means 0 bytes).
3730 bref = &chain->bref;
3731 bytes = (size_t)(bref->data_off & HAMMER2_OFF_MASK_RADIX);
3733 bytes = (hammer2_off_t)1 << bytes;
3736 * If parent is not NULL the duplicated chain will be entered under
3737 * the parent and the UPDATE bit set to tell flush to update
3740 * We must setflush(parent) to ensure that it recurses through to
3741 * chain. setflush(chain) might not work because ONFLUSH is possibly
3742 * already set in the chain (so it won't recurse up to set it in the
3745 * Having both chains locked is extremely important for atomicy.
3747 if (parentp && (parent = *parentp) != NULL) {
3748 KKASSERT(hammer2_mtx_owned(&parent->lock));
3749 KKASSERT(parent->refs > 0);
3750 KKASSERT(parent->error == 0);
3752 hammer2_chain_create(parentp, &chain, NULL, chain->pmp,
3753 HAMMER2_METH_DEFAULT,
3754 bref->key, bref->keybits, bref->type,
3755 chain->bytes, mtid, 0, flags);
3756 KKASSERT(chain->flags & HAMMER2_CHAIN_UPDATE);
3757 hammer2_chain_setflush(*parentp);
3762 * This works in tandem with delete_obref() to install a blockref in
3763 * (typically) an indirect block that is associated with the chain being
3764 * moved to *parentp.
3766 * The reason we need this function is that the caller needs to maintain
3767 * the blockref as it was, and not generate a new blockref for what might
3768 * be a modified chain. Otherwise stuff will leak into the flush that
3769 * the flush code's FLUSH_INODE_STOP flag is unable to catch.
3771 * It is EXTREMELY important that we properly set CHAIN_BMAPUPD and
3772 * CHAIN_UPDATE. We must set BMAPUPD if the bref does not match, and
3773 * we must clear CHAIN_UPDATE (that was likely set by the chain_rename) if
3774 * it does. Otherwise we can end up in a situation where H2 is unable to
3775 * clean up the in-memory chain topology.
3777 * The reason for this is that flushes do not generally flush through
3778 * BREF_TYPE_INODE chains and depend on a hammer2_inode_t queued to syncq
3779 * or sideq to properly flush and dispose of the related inode chain's flags.
3780 * Situations where the inode is not actually modified by the frontend,
3781 * but where we have to move the related chains around as we insert or cleanup
3782 * indirect blocks, can leave us with a 'dirty' (non-disposable) in-memory
3783 * inode chain that does not have a hammer2_inode_t associated with it.
3786 hammer2_chain_rename_obref(hammer2_chain_t **parentp, hammer2_chain_t *chain,
3787 hammer2_tid_t mtid, int flags,
3788 hammer2_blockref_t *obref)
3790 hammer2_chain_rename(parentp, chain, mtid, flags);
3792 if (obref->type != HAMMER2_BREF_TYPE_EMPTY) {
3793 hammer2_blockref_t *tbase;
3796 KKASSERT((chain->flags & HAMMER2_CHAIN_BMAPPED) == 0);
3797 hammer2_chain_modify(*parentp, mtid, 0, 0);
3798 tbase = hammer2_chain_base_and_count(*parentp, &tcount);
3799 hammer2_base_insert(*parentp, tbase, tcount, chain, obref);
3800 if (bcmp(obref, &chain->bref, sizeof(chain->bref))) {
3801 atomic_set_int(&chain->flags, HAMMER2_CHAIN_BMAPUPD |
3802 HAMMER2_CHAIN_UPDATE);
3804 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
3810 * Helper function for deleting chains.
3812 * The chain is removed from the live view (the RBTREE) as well as the parent's
3813 * blockmap. Both chain and its parent must be locked.
3815 * parent may not be errored. chain can be errored.
3818 _hammer2_chain_delete_helper(hammer2_chain_t *parent, hammer2_chain_t *chain,
3819 hammer2_tid_t mtid, int flags,
3820 hammer2_blockref_t *obref)
3825 KKASSERT((chain->flags & HAMMER2_CHAIN_DELETED) == 0);
3826 KKASSERT(chain->parent == parent);
3829 if (chain->flags & HAMMER2_CHAIN_BMAPPED) {
3831 * Chain is blockmapped, so there must be a parent.
3832 * Atomically remove the chain from the parent and remove
3833 * the blockmap entry. The parent must be set modified
3834 * to remove the blockmap entry.
3836 hammer2_blockref_t *base;
3839 KKASSERT(parent != NULL);
3840 KKASSERT(parent->error == 0);
3841 KKASSERT((parent->flags & HAMMER2_CHAIN_INITIAL) == 0);
3842 error = hammer2_chain_modify(parent, mtid, 0, 0);
3847 * Calculate blockmap pointer
3849 KKASSERT(chain->flags & HAMMER2_CHAIN_ONRBTREE);
3850 hammer2_spin_ex(&chain->core.spin);
3851 hammer2_spin_ex(&parent->core.spin);
3853 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DELETED);
3854 atomic_add_int(&parent->core.live_count, -1);
3855 ++parent->core.generation;
3856 RB_REMOVE(hammer2_chain_tree, &parent->core.rbtree, chain);
3857 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE);
3858 --parent->core.chain_count;
3859 chain->parent = NULL;
3861 switch(parent->bref.type) {
3862 case HAMMER2_BREF_TYPE_INODE:
3864 * Access the inode's block array. However, there
3865 * is no block array if the inode is flagged
3869 (parent->data->ipdata.meta.op_flags &
3870 HAMMER2_OPFLAG_DIRECTDATA) == 0) {
3872 &parent->data->ipdata.u.blockset.blockref[0];
3876 count = HAMMER2_SET_COUNT;
3878 case HAMMER2_BREF_TYPE_INDIRECT:
3879 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3881 base = &parent->data->npdata[0];
3884 count = parent->bytes / sizeof(hammer2_blockref_t);
3886 case HAMMER2_BREF_TYPE_VOLUME:
3887 base = &parent->data->voldata.
3888 sroot_blockset.blockref[0];
3889 count = HAMMER2_SET_COUNT;
3891 case HAMMER2_BREF_TYPE_FREEMAP:
3892 base = &parent->data->blkset.blockref[0];
3893 count = HAMMER2_SET_COUNT;
3898 panic("_hammer2_chain_delete_helper: "
3899 "unrecognized blockref type: %d",
3904 * delete blockmapped chain from its parent.
3906 * The parent is not affected by any statistics in chain
3907 * which are pending synchronization. That is, there is
3908 * nothing to undo in the parent since they have not yet
3909 * been incorporated into the parent.
3911 * The parent is affected by statistics stored in inodes.
3912 * Those have already been synchronized, so they must be
3913 * undone. XXX split update possible w/delete in middle?
3916 hammer2_base_delete(parent, base, count, chain, obref);
3918 hammer2_spin_unex(&parent->core.spin);
3919 hammer2_spin_unex(&chain->core.spin);
3920 } else if (chain->flags & HAMMER2_CHAIN_ONRBTREE) {
3922 * Chain is not blockmapped but a parent is present.
3923 * Atomically remove the chain from the parent. There is
3924 * no blockmap entry to remove.
3926 * Because chain was associated with a parent but not
3927 * synchronized, the chain's *_count_up fields contain
3928 * inode adjustment statistics which must be undone.
3930 hammer2_spin_ex(&chain->core.spin);
3931 hammer2_spin_ex(&parent->core.spin);
3932 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DELETED);
3933 atomic_add_int(&parent->core.live_count, -1);
3934 ++parent->core.generation;
3935 RB_REMOVE(hammer2_chain_tree, &parent->core.rbtree, chain);
3936 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE);
3937 --parent->core.chain_count;
3938 chain->parent = NULL;
3939 hammer2_spin_unex(&parent->core.spin);
3940 hammer2_spin_unex(&chain->core.spin);
3943 * Chain is not blockmapped and has no parent. This
3944 * is a degenerate case.
3946 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DELETED);
3953 * Create an indirect block that covers one or more of the elements in the
3954 * current parent. Either returns the existing parent with no locking or
3955 * ref changes or returns the new indirect block locked and referenced
3956 * and leaving the original parent lock/ref intact as well.
3958 * If an error occurs, NULL is returned and *errorp is set to the H2 error.
3960 * The returned chain depends on where the specified key falls.
3962 * The key/keybits for the indirect mode only needs to follow three rules:
3964 * (1) That all elements underneath it fit within its key space and
3966 * (2) That all elements outside it are outside its key space.
3968 * (3) When creating the new indirect block any elements in the current
3969 * parent that fit within the new indirect block's keyspace must be
3970 * moved into the new indirect block.
3972 * (4) The keyspace chosen for the inserted indirect block CAN cover a wider
3973 * keyspace the the current parent, but lookup/iteration rules will
3974 * ensure (and must ensure) that rule (2) for all parents leading up
3975 * to the nearest inode or the root volume header is adhered to. This
3976 * is accomplished by always recursing through matching keyspaces in
3977 * the hammer2_chain_lookup() and hammer2_chain_next() API.
3979 * The current implementation calculates the current worst-case keyspace by
3980 * iterating the current parent and then divides it into two halves, choosing
3981 * whichever half has the most elements (not necessarily the half containing
3982 * the requested key).
3984 * We can also opt to use the half with the least number of elements. This
3985 * causes lower-numbered keys (aka logical file offsets) to recurse through
3986 * fewer indirect blocks and higher-numbered keys to recurse through more.
3987 * This also has the risk of not moving enough elements to the new indirect
3988 * block and being forced to create several indirect blocks before the element
3991 * Must be called with an exclusively locked parent.
3993 * NOTE: *errorp set to HAMMER_ERROR_* flags
3995 static int hammer2_chain_indkey_freemap(hammer2_chain_t *parent,
3996 hammer2_key_t *keyp, int keybits,
3997 hammer2_blockref_t *base, int count);
3998 static int hammer2_chain_indkey_file(hammer2_chain_t *parent,
3999 hammer2_key_t *keyp, int keybits,
4000 hammer2_blockref_t *base, int count,
4002 static int hammer2_chain_indkey_dir(hammer2_chain_t *parent,
4003 hammer2_key_t *keyp, int keybits,
4004 hammer2_blockref_t *base, int count,
4008 hammer2_chain_create_indirect(hammer2_chain_t *parent,
4009 hammer2_key_t create_key, int create_bits,
4010 hammer2_tid_t mtid, int for_type, int *errorp)
4013 hammer2_blockref_t *base;
4014 hammer2_blockref_t *bref;
4015 hammer2_blockref_t bsave;
4016 hammer2_blockref_t dummy;
4017 hammer2_chain_t *chain;
4018 hammer2_chain_t *ichain;
4019 hammer2_key_t key = create_key;
4020 hammer2_key_t key_beg;
4021 hammer2_key_t key_end;
4022 hammer2_key_t key_next;
4023 int keybits = create_bits;
4031 int maxloops = 300000;
4034 * Calculate the base blockref pointer or NULL if the chain
4035 * is known to be empty. We need to calculate the array count
4036 * for RB lookups either way.
4039 KKASSERT(hammer2_mtx_owned(&parent->lock));
4042 * Pre-modify the parent now to avoid having to deal with error
4043 * processing if we tried to later (in the middle of our loop).
4045 * We are going to be moving bref's around, the indirect blocks
4046 * cannot be in an initial state. Do not pass MODIFY_OPTDATA.
4048 *errorp = hammer2_chain_modify(parent, mtid, 0, 0);
4050 kprintf("hammer2_chain_create_indirect: error %08x %s\n",
4051 *errorp, hammer2_error_str(*errorp));
4054 KKASSERT((parent->flags & HAMMER2_CHAIN_INITIAL) == 0);
4056 /*hammer2_chain_modify(&parent, HAMMER2_MODIFY_OPTDATA);*/
4057 base = hammer2_chain_base_and_count(parent, &count);
4060 * How big should our new indirect block be? It has to be at least
4061 * as large as its parent for splits to work properly.
4063 * The freemap uses a specific indirect block size. The number of
4064 * levels are built dynamically and ultimately depend on the size
4065 * volume. Because freemap blocks are taken from the reserved areas
4066 * of the volume our goal is efficiency (fewer levels) and not so
4067 * much to save disk space.
4069 * The first indirect block level for a directory usually uses
4070 * HAMMER2_IND_BYTES_MIN (4KB = 32 directory entries). Due to
4071 * the hash mechanism, this typically gives us a nominal
4072 * 32 * 4 entries with one level of indirection.
4074 * We use HAMMER2_IND_BYTES_NOM (16KB = 128 blockrefs) for FILE
4075 * indirect blocks. The initial 4 entries in the inode gives us
4076 * 256KB. Up to 4 indirect blocks gives us 32MB. Three levels
4077 * of indirection gives us 137GB, and so forth. H2 can support
4078 * huge file sizes but they are not typical, so we try to stick
4079 * with compactness and do not use a larger indirect block size.
4081 * We could use 64KB (PBUFSIZE), giving us 512 blockrefs, but
4082 * due to the way indirect blocks are created this usually winds
4083 * up being extremely inefficient for small files. Even though
4084 * 16KB requires more levels of indirection for very large files,
4085 * the 16KB records can be ganged together into 64KB DIOs.
4087 if (for_type == HAMMER2_BREF_TYPE_FREEMAP_NODE ||
4088 for_type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) {
4089 nbytes = HAMMER2_FREEMAP_LEVELN_PSIZE;
4090 } else if (parent->bref.type == HAMMER2_BREF_TYPE_INODE) {
4091 if (parent->data->ipdata.meta.type ==
4092 HAMMER2_OBJTYPE_DIRECTORY)
4093 nbytes = HAMMER2_IND_BYTES_MIN; /* 4KB = 32 entries */
4095 nbytes = HAMMER2_IND_BYTES_NOM; /* 16KB = ~8MB file */
4098 nbytes = HAMMER2_IND_BYTES_NOM;
4100 if (nbytes < count * sizeof(hammer2_blockref_t)) {
4101 KKASSERT(for_type != HAMMER2_BREF_TYPE_FREEMAP_NODE &&
4102 for_type != HAMMER2_BREF_TYPE_FREEMAP_LEAF);
4103 nbytes = count * sizeof(hammer2_blockref_t);
4105 ncount = nbytes / sizeof(hammer2_blockref_t);
4108 * When creating an indirect block for a freemap node or leaf
4109 * the key/keybits must be fitted to static radix levels because
4110 * particular radix levels use particular reserved blocks in the
4113 * This routine calculates the key/radix of the indirect block
4114 * we need to create, and whether it is on the high-side or the
4118 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
4119 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
4120 keybits = hammer2_chain_indkey_freemap(parent, &key, keybits,
4123 case HAMMER2_BREF_TYPE_DATA:
4124 keybits = hammer2_chain_indkey_file(parent, &key, keybits,
4125 base, count, ncount);
4127 case HAMMER2_BREF_TYPE_DIRENT:
4128 case HAMMER2_BREF_TYPE_INODE:
4129 keybits = hammer2_chain_indkey_dir(parent, &key, keybits,
4130 base, count, ncount);
4133 panic("illegal indirect block for bref type %d", for_type);
4138 * Normalize the key for the radix being represented, keeping the
4139 * high bits and throwing away the low bits.
4141 key &= ~(((hammer2_key_t)1 << keybits) - 1);
4144 * Ok, create our new indirect block
4146 bzero(&dummy, sizeof(dummy));
4147 if (for_type == HAMMER2_BREF_TYPE_FREEMAP_NODE ||
4148 for_type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) {
4149 dummy.type = HAMMER2_BREF_TYPE_FREEMAP_NODE;
4151 dummy.type = HAMMER2_BREF_TYPE_INDIRECT;
4154 dummy.keybits = keybits;
4155 dummy.data_off = hammer2_getradix(nbytes);
4157 HAMMER2_ENC_CHECK(HAMMER2_DEC_CHECK(parent->bref.methods)) |
4158 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
4160 ichain = hammer2_chain_alloc(hmp, parent->pmp, &dummy);
4161 atomic_set_int(&ichain->flags, HAMMER2_CHAIN_INITIAL);
4162 hammer2_chain_lock(ichain, HAMMER2_RESOLVE_MAYBE);
4163 /* ichain has one ref at this point */
4166 * We have to mark it modified to allocate its block, but use
4167 * OPTDATA to allow it to remain in the INITIAL state. Otherwise
4168 * it won't be acted upon by the flush code.
4170 * XXX remove OPTDATA, we need a fully initialized indirect block to
4171 * be able to move the original blockref.
4173 *errorp = hammer2_chain_modify(ichain, mtid, 0, 0);
4175 kprintf("hammer2_chain_create_indirect: error %08x %s\n",
4176 *errorp, hammer2_error_str(*errorp));
4177 hammer2_chain_unlock(ichain);
4178 hammer2_chain_drop(ichain);
4181 KKASSERT((ichain->flags & HAMMER2_CHAIN_INITIAL) == 0);
4184 * Iterate the original parent and move the matching brefs into
4185 * the new indirect block.
4187 * XXX handle flushes.
4190 key_end = HAMMER2_KEY_MAX;
4191 key_next = 0; /* avoid gcc warnings */
4192 hammer2_spin_ex(&parent->core.spin);
4198 * Parent may have been modified, relocating its block array.
4199 * Reload the base pointer.
4201 base = hammer2_chain_base_and_count(parent, &count);
4203 if (++loops > 100000) {
4204 hammer2_spin_unex(&parent->core.spin);
4205 panic("excessive loops r=%d p=%p base/count %p:%d %016jx\n",
4206 reason, parent, base, count, key_next);
4210 * NOTE: spinlock stays intact, returned chain (if not NULL)
4211 * is not referenced or locked which means that we
4212 * cannot safely check its flagged / deletion status
4215 chain = hammer2_combined_find(parent, base, count,
4219 generation = parent->core.generation;
4222 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
4225 * Skip keys that are not within the key/radix of the new
4226 * indirect block. They stay in the parent.
4228 if (rounddown2(key ^ bref->key, (hammer2_key_t)1 << keybits) != 0) {
4229 goto next_key_spinlocked;
4233 * Load the new indirect block by acquiring the related
4234 * chains (potentially from media as it might not be
4235 * in-memory). Then move it to the new parent (ichain).
4237 * chain is referenced but not locked. We must lock the
4238 * chain to obtain definitive state.
4243 * Use chain already present in the RBTREE
4245 hammer2_chain_ref(chain);
4246 hammer2_spin_unex(&parent->core.spin);
4247 hammer2_chain_lock(chain, HAMMER2_RESOLVE_NEVER);
4250 * Get chain for blockref element. _get returns NULL
4251 * on insertion race.
4253 hammer2_spin_unex(&parent->core.spin);
4254 chain = hammer2_chain_get(parent, generation, &bsave,
4255 HAMMER2_RESOLVE_NEVER);
4256 if (chain == NULL) {
4258 hammer2_spin_ex(&parent->core.spin);
4264 * This is always live so if the chain has been deleted
4265 * we raced someone and we have to retry.
4267 * NOTE: Lookups can race delete-duplicate because
4268 * delete-duplicate does not lock the parent's core
4269 * (they just use the spinlock on the core).
4271 * (note reversed logic for this one)
4273 if (bcmp(&bsave, &chain->bref, sizeof(bsave)) ||
4274 chain->parent != parent ||
4275 (chain->flags & HAMMER2_CHAIN_DELETED)) {
4276 hammer2_chain_unlock(chain);
4277 hammer2_chain_drop(chain);
4278 if (hammer2_debug & 0x0040) {
4279 kprintf("LOST PARENT RETRY "
4280 "RETRY (%p,%p)->%p %08x\n",
4281 parent, chain->parent, chain, chain->flags);
4283 hammer2_spin_ex(&parent->core.spin);
4288 * Shift the chain to the indirect block.
4290 * WARNING! No reason for us to load chain data, pass NOSTATS
4291 * to prevent delete/insert from trying to access
4292 * inode stats (and thus asserting if there is no
4293 * chain->data loaded).
4295 * WARNING! The (parent, chain) deletion may modify the parent
4296 * and invalidate the base pointer.
4298 * WARNING! Parent must already be marked modified, so we
4299 * can assume that chain_delete always suceeds.
4301 * WARNING! hammer2_chain_repchange() does not have to be
4302 * called (and doesn't work anyway because we are
4303 * only doing a partial shift). A recursion that is
4304 * in-progress can continue at the current parent
4305 * and will be able to properly find its next key.
4307 error = hammer2_chain_delete_obref(parent, chain, mtid, 0,
4309 KKASSERT(error == 0);
4310 hammer2_chain_rename_obref(&ichain, chain, mtid, 0, &bsave);
4311 hammer2_chain_unlock(chain);
4312 hammer2_chain_drop(chain);
4313 KKASSERT(parent->refs > 0);
4315 base = NULL; /* safety */
4316 hammer2_spin_ex(&parent->core.spin);
4317 next_key_spinlocked:
4318 if (--maxloops == 0)
4319 panic("hammer2_chain_create_indirect: maxloops");
4321 if (key_next == 0 || key_next > key_end)
4326 hammer2_spin_unex(&parent->core.spin);
4329 * Insert the new indirect block into the parent now that we've
4330 * cleared out some entries in the parent. We calculated a good
4331 * insertion index in the loop above (ichain->index).
4333 * We don't have to set UPDATE here because we mark ichain
4334 * modified down below (so the normal modified -> flush -> set-moved
4335 * sequence applies).
4337 * The insertion shouldn't race as this is a completely new block
4338 * and the parent is locked.
4340 base = NULL; /* safety, parent modify may change address */
4341 KKASSERT((ichain->flags & HAMMER2_CHAIN_ONRBTREE) == 0);
4342 KKASSERT(parent->core.live_count < count);
4343 hammer2_chain_insert(parent, ichain,
4344 HAMMER2_CHAIN_INSERT_SPIN |
4345 HAMMER2_CHAIN_INSERT_LIVE,
4349 * Make sure flushes propogate after our manual insertion.
4351 hammer2_chain_setflush(ichain);
4352 hammer2_chain_setflush(parent);
4355 * Figure out what to return.
4357 if (rounddown2(create_key ^ key, (hammer2_key_t)1 << keybits)) {
4359 * Key being created is outside the key range,
4360 * return the original parent.
4362 hammer2_chain_unlock(ichain);
4363 hammer2_chain_drop(ichain);
4366 * Otherwise its in the range, return the new parent.
4367 * (leave both the new and old parent locked).
4376 * Do maintenance on an indirect chain. Both parent and chain are locked.
4378 * Returns non-zero if (chain) is deleted, either due to being empty or
4379 * because its children were safely moved into the parent.
4382 hammer2_chain_indirect_maintenance(hammer2_chain_t *parent,
4383 hammer2_chain_t *chain)
4385 hammer2_blockref_t *chain_base;
4386 hammer2_blockref_t *base;
4387 hammer2_blockref_t *bref;
4388 hammer2_blockref_t bsave;
4389 hammer2_key_t key_next;
4390 hammer2_key_t key_beg;
4391 hammer2_key_t key_end;
4392 hammer2_chain_t *sub;
4399 * Make sure we have an accurate live_count
4401 if ((chain->flags & (HAMMER2_CHAIN_INITIAL |
4402 HAMMER2_CHAIN_COUNTEDBREFS)) == 0) {
4403 base = &chain->data->npdata[0];
4404 count = chain->bytes / sizeof(hammer2_blockref_t);
4405 hammer2_chain_countbrefs(chain, base, count);
4409 * If the indirect block is empty we can delete it.
4410 * (ignore deletion error)
4412 if (chain->core.live_count == 0 && RB_EMPTY(&chain->core.rbtree)) {
4413 hammer2_chain_delete(parent, chain,
4414 chain->bref.modify_tid,
4415 HAMMER2_DELETE_PERMANENT);
4416 hammer2_chain_repchange(parent, chain);
4420 base = hammer2_chain_base_and_count(parent, &count);
4422 if ((parent->flags & (HAMMER2_CHAIN_INITIAL |
4423 HAMMER2_CHAIN_COUNTEDBREFS)) == 0) {
4424 hammer2_chain_countbrefs(parent, base, count);
4428 * Determine if we can collapse chain into parent, calculate
4429 * hysteresis for chain emptiness.
4431 if (parent->core.live_count + chain->core.live_count - 1 > count)
4433 chain_count = chain->bytes / sizeof(hammer2_blockref_t);
4434 if (chain->core.live_count > chain_count * 3 / 4)
4438 * Ok, theoretically we can collapse chain's contents into
4439 * parent. chain is locked, but any in-memory children of chain
4440 * are not. For this to work, we must be able to dispose of any
4441 * in-memory children of chain.
4443 * For now require that there are no in-memory children of chain.
4445 * WARNING! Both chain and parent must remain locked across this
4450 * Parent must be marked modified. Don't try to collapse it if we
4451 * can't mark it modified. Once modified, destroy chain to make room
4452 * and to get rid of what will be a conflicting key (this is included
4453 * in the calculation above). Finally, move the children of chain
4454 * into chain's parent.
4456 * This order creates an accounting problem for bref.embed.stats
4457 * because we destroy chain before we remove its children. Any
4458 * elements whos blockref is already synchronized will be counted
4459 * twice. To deal with the problem we clean out chain's stats prior
4462 error = hammer2_chain_modify(parent, 0, 0, 0);
4464 krateprintf(&krate_h2me, "hammer2: indirect_maint: %s\n",
4465 hammer2_error_str(error));
4468 error = hammer2_chain_modify(chain, chain->bref.modify_tid, 0, 0);
4470 krateprintf(&krate_h2me, "hammer2: indirect_maint: %s\n",
4471 hammer2_error_str(error));
4475 chain->bref.embed.stats.inode_count = 0;
4476 chain->bref.embed.stats.data_count = 0;
4477 error = hammer2_chain_delete(parent, chain,
4478 chain->bref.modify_tid,
4479 HAMMER2_DELETE_PERMANENT);
4480 KKASSERT(error == 0);
4483 * The combined_find call requires core.spin to be held. One would
4484 * think there wouldn't be any conflicts since we hold chain
4485 * exclusively locked, but the caching mechanism for 0-ref children
4486 * does not require a chain lock.
4488 hammer2_spin_ex(&chain->core.spin);
4492 key_end = HAMMER2_KEY_MAX;
4494 chain_base = &chain->data->npdata[0];
4495 chain_count = chain->bytes / sizeof(hammer2_blockref_t);
4496 sub = hammer2_combined_find(chain, chain_base, chain_count,
4500 generation = chain->core.generation;
4503 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
4507 hammer2_chain_ref(sub);
4508 hammer2_spin_unex(&chain->core.spin);
4509 hammer2_chain_lock(sub, HAMMER2_RESOLVE_NEVER);
4511 hammer2_spin_unex(&chain->core.spin);
4512 sub = hammer2_chain_get(chain, generation, &bsave,
4513 HAMMER2_RESOLVE_NEVER);
4515 hammer2_spin_ex(&chain->core.spin);
4519 if (bcmp(&bsave, &sub->bref, sizeof(bsave)) ||
4520 sub->parent != chain ||
4521 (sub->flags & HAMMER2_CHAIN_DELETED)) {
4522 hammer2_chain_unlock(sub);
4523 hammer2_chain_drop(sub);
4524 hammer2_spin_ex(&chain->core.spin);
4525 sub = NULL; /* safety */
4528 error = hammer2_chain_delete_obref(chain, sub,
4529 sub->bref.modify_tid, 0,
4531 KKASSERT(error == 0);
4532 hammer2_chain_rename_obref(&parent, sub,
4533 sub->bref.modify_tid,
4534 HAMMER2_INSERT_SAMEPARENT, &bsave);
4535 hammer2_chain_unlock(sub);
4536 hammer2_chain_drop(sub);
4537 hammer2_spin_ex(&chain->core.spin);
4543 hammer2_spin_unex(&chain->core.spin);
4545 hammer2_chain_repchange(parent, chain);
4551 * Freemap indirect blocks
4553 * Calculate the keybits and highside/lowside of the freemap node the
4554 * caller is creating.
4556 * This routine will specify the next higher-level freemap key/radix
4557 * representing the lowest-ordered set. By doing so, eventually all
4558 * low-ordered sets will be moved one level down.
4560 * We have to be careful here because the freemap reserves a limited
4561 * number of blocks for a limited number of levels. So we can't just
4562 * push indiscriminately.
4565 hammer2_chain_indkey_freemap(hammer2_chain_t *parent, hammer2_key_t *keyp,
4566 int keybits, hammer2_blockref_t *base, int count)
4568 hammer2_chain_t *chain;
4569 hammer2_blockref_t *bref;
4571 hammer2_key_t key_beg;
4572 hammer2_key_t key_end;
4573 hammer2_key_t key_next;
4576 int maxloops = 300000;
4584 * Calculate the range of keys in the array being careful to skip
4585 * slots which are overridden with a deletion.
4588 key_end = HAMMER2_KEY_MAX;
4589 hammer2_spin_ex(&parent->core.spin);
4592 if (--maxloops == 0) {
4593 panic("indkey_freemap shit %p %p:%d\n",
4594 parent, base, count);
4596 chain = hammer2_combined_find(parent, base, count,
4608 * Skip deleted chains.
4610 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) {
4611 if (key_next == 0 || key_next > key_end)
4618 * Use the full live (not deleted) element for the scan
4619 * iteration. HAMMER2 does not allow partial replacements.
4621 * XXX should be built into hammer2_combined_find().
4623 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
4625 if (keybits > bref->keybits) {
4627 keybits = bref->keybits;
4628 } else if (keybits == bref->keybits && bref->key < key) {
4635 hammer2_spin_unex(&parent->core.spin);
4638 * Return the keybits for a higher-level FREEMAP_NODE covering
4642 case HAMMER2_FREEMAP_LEVEL0_RADIX:
4643 keybits = HAMMER2_FREEMAP_LEVEL1_RADIX;
4645 case HAMMER2_FREEMAP_LEVEL1_RADIX:
4646 keybits = HAMMER2_FREEMAP_LEVEL2_RADIX;
4648 case HAMMER2_FREEMAP_LEVEL2_RADIX:
4649 keybits = HAMMER2_FREEMAP_LEVEL3_RADIX;
4651 case HAMMER2_FREEMAP_LEVEL3_RADIX:
4652 keybits = HAMMER2_FREEMAP_LEVEL4_RADIX;
4654 case HAMMER2_FREEMAP_LEVEL4_RADIX:
4655 keybits = HAMMER2_FREEMAP_LEVEL5_RADIX;
4657 case HAMMER2_FREEMAP_LEVEL5_RADIX:
4658 panic("hammer2_chain_indkey_freemap: level too high");
4661 panic("hammer2_chain_indkey_freemap: bad radix");
4670 * File indirect blocks
4672 * Calculate the key/keybits for the indirect block to create by scanning
4673 * existing keys. The key being created is also passed in *keyp and can be
4674 * inside or outside the indirect block. Regardless, the indirect block
4675 * must hold at least two keys in order to guarantee sufficient space.
4677 * We use a modified version of the freemap's fixed radix tree, but taylored
4678 * for file data. Basically we configure an indirect block encompassing the
4682 hammer2_chain_indkey_file(hammer2_chain_t *parent, hammer2_key_t *keyp,
4683 int keybits, hammer2_blockref_t *base, int count,
4686 hammer2_chain_t *chain;
4687 hammer2_blockref_t *bref;
4689 hammer2_key_t key_beg;
4690 hammer2_key_t key_end;
4691 hammer2_key_t key_next;
4695 int maxloops = 300000;
4703 * Calculate the range of keys in the array being careful to skip
4704 * slots which are overridden with a deletion.
4706 * Locate the smallest key.
4709 key_end = HAMMER2_KEY_MAX;
4710 hammer2_spin_ex(&parent->core.spin);
4713 if (--maxloops == 0) {
4714 panic("indkey_freemap shit %p %p:%d\n",
4715 parent, base, count);
4717 chain = hammer2_combined_find(parent, base, count,
4729 * Skip deleted chains.
4731 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) {
4732 if (key_next == 0 || key_next > key_end)
4739 * Use the full live (not deleted) element for the scan
4740 * iteration. HAMMER2 does not allow partial replacements.
4742 * XXX should be built into hammer2_combined_find().
4744 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
4746 if (keybits > bref->keybits) {
4748 keybits = bref->keybits;
4749 } else if (keybits == bref->keybits && bref->key < key) {
4756 hammer2_spin_unex(&parent->core.spin);
4759 * Calculate the static keybits for a higher-level indirect block
4760 * that contains the key.
4765 case HAMMER2_IND_BYTES_MIN / sizeof(hammer2_blockref_t):
4766 nradix = HAMMER2_IND_RADIX_MIN - HAMMER2_BLOCKREF_RADIX;
4768 case HAMMER2_IND_BYTES_NOM / sizeof(hammer2_blockref_t):
4769 nradix = HAMMER2_IND_RADIX_NOM - HAMMER2_BLOCKREF_RADIX;
4771 case HAMMER2_IND_BYTES_MAX / sizeof(hammer2_blockref_t):
4772 nradix = HAMMER2_IND_RADIX_MAX - HAMMER2_BLOCKREF_RADIX;
4775 panic("bad ncount %d\n", ncount);
4781 * The largest radix that can be returned for an indirect block is
4782 * 63 bits. (The largest practical indirect block radix is actually
4783 * 62 bits because the top-level inode or volume root contains four
4784 * entries, but allow 63 to be returned).
4789 return keybits + nradix;
4795 * Directory indirect blocks.
4797 * Covers both the inode index (directory of inodes), and directory contents
4798 * (filenames hardlinked to inodes).
4800 * Because directory keys are hashed we generally try to cut the space in
4801 * half. We accomodate the inode index (which tends to have linearly
4802 * increasing inode numbers) by ensuring that the keyspace is at least large
4803 * enough to fill up the indirect block being created.
4806 hammer2_chain_indkey_dir(hammer2_chain_t *parent, hammer2_key_t *keyp,
4807 int keybits, hammer2_blockref_t *base, int count,
4810 hammer2_blockref_t *bref;
4811 hammer2_chain_t *chain;
4812 hammer2_key_t key_beg;
4813 hammer2_key_t key_end;
4814 hammer2_key_t key_next;
4819 int maxloops = 300000;
4822 * NOTE: We can't take a shortcut here anymore for inodes because
4823 * the root directory can contain a mix of inodes and directory
4824 * entries (we used to just return 63 if parent->bref.type was
4825 * HAMMER2_BREF_TYPE_INODE.
4832 * Calculate the range of keys in the array being careful to skip
4833 * slots which are overridden with a deletion.
4836 key_end = HAMMER2_KEY_MAX;
4837 hammer2_spin_ex(&parent->core.spin);
4840 if (--maxloops == 0) {
4841 panic("indkey_freemap shit %p %p:%d\n",
4842 parent, base, count);
4844 chain = hammer2_combined_find(parent, base, count,
4858 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) {
4859 if (key_next == 0 || key_next > key_end)
4866 * Use the full live (not deleted) element for the scan
4867 * iteration. HAMMER2 does not allow partial replacements.
4869 * XXX should be built into hammer2_combined_find().
4871 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
4874 * Expand our calculated key range (key, keybits) to fit
4875 * the scanned key. nkeybits represents the full range
4876 * that we will later cut in half (two halves @ nkeybits - 1).
4879 if (nkeybits < bref->keybits) {
4880 if (bref->keybits > 64) {
4881 kprintf("bad bref chain %p bref %p\n",
4885 nkeybits = bref->keybits;
4887 while (nkeybits < 64 &&
4888 rounddown2(key ^ bref->key, (hammer2_key_t)1 << nkeybits) != 0) {
4893 * If the new key range is larger we have to determine
4894 * which side of the new key range the existing keys fall
4895 * under by checking the high bit, then collapsing the
4896 * locount into the hicount or vise-versa.
4898 if (keybits != nkeybits) {
4899 if (((hammer2_key_t)1 << (nkeybits - 1)) & key) {
4910 * The newly scanned key will be in the lower half or the
4911 * upper half of the (new) key range.
4913 if (((hammer2_key_t)1 << (nkeybits - 1)) & bref->key)
4922 hammer2_spin_unex(&parent->core.spin);
4923 bref = NULL; /* now invalid (safety) */
4926 * Adjust keybits to represent half of the full range calculated
4927 * above (radix 63 max) for our new indirect block.
4932 * Expand keybits to hold at least ncount elements. ncount will be
4933 * a power of 2. This is to try to completely fill leaf nodes (at
4934 * least for keys which are not hashes).
4936 * We aren't counting 'in' or 'out', we are counting 'high side'
4937 * and 'low side' based on the bit at (1LL << keybits). We want
4938 * everything to be inside in these cases so shift it all to
4939 * the low or high side depending on the new high bit.
4941 while (((hammer2_key_t)1 << keybits) < ncount) {
4943 if (key & ((hammer2_key_t)1 << keybits)) {
4952 if (hicount > locount)
4953 key |= (hammer2_key_t)1 << keybits;
4955 key &= ~(hammer2_key_t)1 << keybits;
4965 * Directory indirect blocks.
4967 * Covers both the inode index (directory of inodes), and directory contents
4968 * (filenames hardlinked to inodes).
4970 * Because directory keys are hashed we generally try to cut the space in
4971 * half. We accomodate the inode index (which tends to have linearly
4972 * increasing inode numbers) by ensuring that the keyspace is at least large
4973 * enough to fill up the indirect block being created.
4976 hammer2_chain_indkey_dir(hammer2_chain_t *parent, hammer2_key_t *keyp,
4977 int keybits, hammer2_blockref_t *base, int count,
4980 hammer2_blockref_t *bref;
4981 hammer2_chain_t *chain;
4982 hammer2_key_t key_beg;
4983 hammer2_key_t key_end;
4984 hammer2_key_t key_next;
4989 int maxloops = 300000;
4992 * Shortcut if the parent is the inode. In this situation the
4993 * parent has 4+1 directory entries and we are creating an indirect
4994 * block capable of holding many more.
4996 if (parent->bref.type == HAMMER2_BREF_TYPE_INODE) {
5005 * Calculate the range of keys in the array being careful to skip
5006 * slots which are overridden with a deletion.
5009 key_end = HAMMER2_KEY_MAX;
5010 hammer2_spin_ex(&parent->core.spin);
5013 if (--maxloops == 0) {
5014 panic("indkey_freemap shit %p %p:%d\n",
5015 parent, base, count);
5017 chain = hammer2_combined_find(parent, base, count,
5031 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) {
5032 if (key_next == 0 || key_next > key_end)
5039 * Use the full live (not deleted) element for the scan
5040 * iteration. HAMMER2 does not allow partial replacements.
5042 * XXX should be built into hammer2_combined_find().
5044 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
5047 * Expand our calculated key range (key, keybits) to fit
5048 * the scanned key. nkeybits represents the full range
5049 * that we will later cut in half (two halves @ nkeybits - 1).
5052 if (nkeybits < bref->keybits) {
5053 if (bref->keybits > 64) {
5054 kprintf("bad bref chain %p bref %p\n",
5058 nkeybits = bref->keybits;
5060 while (nkeybits < 64 &&
5061 (~(((hammer2_key_t)1 << nkeybits) - 1) &
5062 (key ^ bref->key)) != 0) {
5067 * If the new key range is larger we have to determine
5068 * which side of the new key range the existing keys fall
5069 * under by checking the high bit, then collapsing the
5070 * locount into the hicount or vise-versa.
5072 if (keybits != nkeybits) {
5073 if (((hammer2_key_t)1 << (nkeybits - 1)) & key) {
5084 * The newly scanned key will be in the lower half or the
5085 * upper half of the (new) key range.
5087 if (((hammer2_key_t)1 << (nkeybits - 1)) & bref->key)
5096 hammer2_spin_unex(&parent->core.spin);
5097 bref = NULL; /* now invalid (safety) */
5100 * Adjust keybits to represent half of the full range calculated
5101 * above (radix 63 max) for our new indirect block.
5106 * Expand keybits to hold at least ncount elements. ncount will be
5107 * a power of 2. This is to try to completely fill leaf nodes (at
5108 * least for keys which are not hashes).
5110 * We aren't counting 'in' or 'out', we are counting 'high side'
5111 * and 'low side' based on the bit at (1LL << keybits). We want
5112 * everything to be inside in these cases so shift it all to
5113 * the low or high side depending on the new high bit.
5115 while (((hammer2_key_t)1 << keybits) < ncount) {
5117 if (key & ((hammer2_key_t)1 << keybits)) {
5126 if (hicount > locount)
5127 key |= (hammer2_key_t)1 << keybits;
5129 key &= ~(hammer2_key_t)1 << keybits;
5139 * Sets CHAIN_DELETED and remove the chain's blockref from the parent if
5142 * Both parent and chain must be locked exclusively.
5144 * This function will modify the parent if the blockref requires removal
5145 * from the parent's block table.
5147 * This function is NOT recursive. Any entity already pushed into the
5148 * chain (such as an inode) may still need visibility into its contents,
5149 * as well as the ability to read and modify the contents. For example,
5150 * for an unlinked file which is still open.
5152 * Also note that the flusher is responsible for cleaning up empty
5156 hammer2_chain_delete(hammer2_chain_t *parent, hammer2_chain_t *chain,
5157 hammer2_tid_t mtid, int flags)
5161 KKASSERT(hammer2_mtx_owned(&chain->lock));
5164 * Nothing to do if already marked.
5166 * We need the spinlock on the core whos RBTREE contains chain
5167 * to protect against races.
5169 if ((chain->flags & HAMMER2_CHAIN_DELETED) == 0) {
5170 KKASSERT((chain->flags & HAMMER2_CHAIN_DELETED) == 0 &&
5171 chain->parent == parent);
5172 error = _hammer2_chain_delete_helper(parent, chain,
5177 * Permanent deletions mark the chain as destroyed.
5179 * NOTE: We do not setflush the chain unless the deletion is
5180 * permanent, since the deletion of a chain does not actually
5181 * require it to be flushed.
5184 if (flags & HAMMER2_DELETE_PERMANENT) {
5185 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DESTROY);
5186 hammer2_chain_setflush(chain);
5194 hammer2_chain_delete_obref(hammer2_chain_t *parent, hammer2_chain_t *chain,
5195 hammer2_tid_t mtid, int flags,
5196 hammer2_blockref_t *obref)
5200 KKASSERT(hammer2_mtx_owned(&chain->lock));
5203 * Nothing to do if already marked.
5205 * We need the spinlock on the core whos RBTREE contains chain
5206 * to protect against races.
5208 obref->type = HAMMER2_BREF_TYPE_EMPTY;
5209 if ((chain->flags & HAMMER2_CHAIN_DELETED) == 0) {
5210 KKASSERT((chain->flags & HAMMER2_CHAIN_DELETED) == 0 &&
5211 chain->parent == parent);
5212 error = _hammer2_chain_delete_helper(parent, chain,
5213 mtid, flags, obref);
5217 * Permanent deletions mark the chain as destroyed.
5219 * NOTE: We do not setflush the chain unless the deletion is
5220 * permanent, since the deletion of a chain does not actually
5221 * require it to be flushed.
5224 if (flags & HAMMER2_DELETE_PERMANENT) {
5225 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DESTROY);
5226 hammer2_chain_setflush(chain);
5234 * Returns the index of the nearest element in the blockref array >= elm.
5235 * Returns (count) if no element could be found.
5237 * Sets *key_nextp to the next key for loop purposes but does not modify
5238 * it if the next key would be higher than the current value of *key_nextp.
5239 * Note that *key_nexp can overflow to 0, which should be tested by the
5242 * WARNING! Must be called with parent's spinlock held. Spinlock remains
5243 * held through the operation.
5246 hammer2_base_find(hammer2_chain_t *parent,
5247 hammer2_blockref_t *base, int count,
5248 hammer2_key_t *key_nextp,
5249 hammer2_key_t key_beg, hammer2_key_t key_end)
5251 hammer2_blockref_t *scan;
5252 hammer2_key_t scan_end;
5257 * Require the live chain's already have their core's counted
5258 * so we can optimize operations.
5260 KKASSERT(parent->flags & HAMMER2_CHAIN_COUNTEDBREFS);
5265 if (count == 0 || base == NULL)
5269 * Sequential optimization using parent->cache_index. This is
5270 * the most likely scenario.
5272 * We can avoid trailing empty entries on live chains, otherwise
5273 * we might have to check the whole block array.
5275 i = parent->cache_index; /* SMP RACE OK */
5277 limit = parent->core.live_zero;
5282 KKASSERT(i < count);
5288 while (i > 0 && (scan->type == HAMMER2_BREF_TYPE_EMPTY ||
5289 scan->key > key_beg)) {
5293 parent->cache_index = i;
5296 * Search forwards, stop when we find a scan element which
5297 * encloses the key or until we know that there are no further
5301 if (scan->type != HAMMER2_BREF_TYPE_EMPTY) {
5302 scan_end = scan->key +
5303 ((hammer2_key_t)1 << scan->keybits) - 1;
5304 if (scan->key > key_beg || scan_end >= key_beg)
5313 parent->cache_index = i;
5317 scan_end = scan->key +
5318 ((hammer2_key_t)1 << scan->keybits);
5319 if (scan_end && (*key_nextp > scan_end ||
5321 *key_nextp = scan_end;
5329 * Do a combined search and return the next match either from the blockref
5330 * array or from the in-memory chain. Sets *bresp to the returned bref in
5331 * both cases, or sets it to NULL if the search exhausted. Only returns
5332 * a non-NULL chain if the search matched from the in-memory chain.
5334 * When no in-memory chain has been found and a non-NULL bref is returned
5338 * The returned chain is not locked or referenced. Use the returned bref
5339 * to determine if the search exhausted or not. Iterate if the base find
5340 * is chosen but matches a deleted chain.
5342 * WARNING! Must be called with parent's spinlock held. Spinlock remains
5343 * held through the operation.
5346 hammer2_combined_find(hammer2_chain_t *parent,
5347 hammer2_blockref_t *base, int count,
5348 hammer2_key_t *key_nextp,
5349 hammer2_key_t key_beg, hammer2_key_t key_end,
5350 hammer2_blockref_t **bresp)
5352 hammer2_blockref_t *bref;
5353 hammer2_chain_t *chain;
5357 * Lookup in block array and in rbtree.
5359 *key_nextp = key_end + 1;
5360 i = hammer2_base_find(parent, base, count, key_nextp,
5362 chain = hammer2_chain_find(parent, key_nextp, key_beg, key_end);
5367 if (i == count && chain == NULL) {
5373 * Only chain matched.
5376 bref = &chain->bref;
5381 * Only blockref matched.
5383 if (chain == NULL) {
5389 * Both in-memory and blockref matched, select the nearer element.
5391 * If both are flush with the left-hand side or both are the
5392 * same distance away, select the chain. In this situation the
5393 * chain must have been loaded from the matching blockmap.
5395 if ((chain->bref.key <= key_beg && base[i].key <= key_beg) ||
5396 chain->bref.key == base[i].key) {
5397 KKASSERT(chain->bref.key == base[i].key);
5398 bref = &chain->bref;
5403 * Select the nearer key
5405 if (chain->bref.key < base[i].key) {
5406 bref = &chain->bref;
5413 * If the bref is out of bounds we've exhausted our search.
5416 if (bref->key > key_end) {
5426 * Locate the specified block array element and delete it. The element
5429 * The spin lock on the related chain must be held.
5431 * NOTE: live_count was adjusted when the chain was deleted, so it does not
5432 * need to be adjusted when we commit the media change.
5435 hammer2_base_delete(hammer2_chain_t *parent,
5436 hammer2_blockref_t *base, int count,
5437 hammer2_chain_t *chain,
5438 hammer2_blockref_t *obref)
5440 hammer2_blockref_t *elm = &chain->bref;
5441 hammer2_blockref_t *scan;
5442 hammer2_key_t key_next;
5446 * Delete element. Expect the element to exist.
5448 * XXX see caller, flush code not yet sophisticated enough to prevent
5449 * re-flushed in some cases.
5451 key_next = 0; /* max range */
5452 i = hammer2_base_find(parent, base, count, &key_next,
5453 elm->key, elm->key);
5455 if (i == count || scan->type == HAMMER2_BREF_TYPE_EMPTY ||
5456 scan->key != elm->key ||
5457 ((chain->flags & HAMMER2_CHAIN_BMAPUPD) == 0 &&
5458 scan->keybits != elm->keybits)) {
5459 hammer2_spin_unex(&parent->core.spin);
5460 panic("delete base %p element not found at %d/%d elm %p\n",
5461 base, i, count, elm);
5466 * Update stats and zero the entry.
5468 * NOTE: Handle radix == 0 (0 bytes) case.
5470 if ((int)(scan->data_off & HAMMER2_OFF_MASK_RADIX)) {
5471 parent->bref.embed.stats.data_count -= (hammer2_off_t)1 <<
5472 (int)(scan->data_off & HAMMER2_OFF_MASK_RADIX);
5474 switch(scan->type) {
5475 case HAMMER2_BREF_TYPE_INODE:
5476 --parent->bref.embed.stats.inode_count;
5478 case HAMMER2_BREF_TYPE_DATA:
5479 if (parent->bref.leaf_count == HAMMER2_BLOCKREF_LEAF_MAX) {
5480 atomic_set_int(&chain->flags,
5481 HAMMER2_CHAIN_HINT_LEAF_COUNT);
5483 if (parent->bref.leaf_count)
5484 --parent->bref.leaf_count;
5487 case HAMMER2_BREF_TYPE_INDIRECT:
5488 if (scan->type != HAMMER2_BREF_TYPE_DATA) {
5489 parent->bref.embed.stats.data_count -=
5490 scan->embed.stats.data_count;
5491 parent->bref.embed.stats.inode_count -=
5492 scan->embed.stats.inode_count;
5494 if (scan->type == HAMMER2_BREF_TYPE_INODE)
5496 if (parent->bref.leaf_count == HAMMER2_BLOCKREF_LEAF_MAX) {
5497 atomic_set_int(&chain->flags,
5498 HAMMER2_CHAIN_HINT_LEAF_COUNT);
5500 if (parent->bref.leaf_count <= scan->leaf_count)
5501 parent->bref.leaf_count = 0;
5503 parent->bref.leaf_count -= scan->leaf_count;
5506 case HAMMER2_BREF_TYPE_DIRENT:
5507 if (parent->bref.leaf_count == HAMMER2_BLOCKREF_LEAF_MAX) {
5508 atomic_set_int(&chain->flags,
5509 HAMMER2_CHAIN_HINT_LEAF_COUNT);
5511 if (parent->bref.leaf_count)
5512 --parent->bref.leaf_count;
5520 bzero(scan, sizeof(*scan));
5523 * We can only optimize parent->core.live_zero for live chains.
5525 if (parent->core.live_zero == i + 1) {
5526 while (--i >= 0 && base[i].type == HAMMER2_BREF_TYPE_EMPTY)
5528 parent->core.live_zero = i + 1;
5532 * Clear appropriate blockmap flags in chain.
5534 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_BMAPPED |
5535 HAMMER2_CHAIN_BMAPUPD);
5539 * Insert the specified element. The block array must not already have the
5540 * element and must have space available for the insertion.
5542 * The spin lock on the related chain must be held.
5544 * NOTE: live_count was adjusted when the chain was deleted, so it does not
5545 * need to be adjusted when we commit the media change.
5548 hammer2_base_insert(hammer2_chain_t *parent,
5549 hammer2_blockref_t *base, int count,
5550 hammer2_chain_t *chain, hammer2_blockref_t *elm)
5552 hammer2_key_t key_next;
5561 * Insert new element. Expect the element to not already exist
5562 * unless we are replacing it.
5564 * XXX see caller, flush code not yet sophisticated enough to prevent
5565 * re-flushed in some cases.
5567 key_next = 0; /* max range */
5568 i = hammer2_base_find(parent, base, count, &key_next,
5569 elm->key, elm->key);
5572 * Shortcut fill optimization, typical ordered insertion(s) may not
5575 KKASSERT(i >= 0 && i <= count);
5578 * Set appropriate blockmap flags in chain (if not NULL)
5581 atomic_set_int(&chain->flags, HAMMER2_CHAIN_BMAPPED);
5584 * Update stats and zero the entry
5586 if ((int)(elm->data_off & HAMMER2_OFF_MASK_RADIX)) {
5587 parent->bref.embed.stats.data_count += (hammer2_off_t)1 <<
5588 (int)(elm->data_off & HAMMER2_OFF_MASK_RADIX);
5591 case HAMMER2_BREF_TYPE_INODE:
5592 ++parent->bref.embed.stats.inode_count;
5594 case HAMMER2_BREF_TYPE_DATA:
5595 if (parent->bref.leaf_count != HAMMER2_BLOCKREF_LEAF_MAX)
5596 ++parent->bref.leaf_count;
5598 case HAMMER2_BREF_TYPE_INDIRECT:
5599 if (elm->type != HAMMER2_BREF_TYPE_DATA) {
5600 parent->bref.embed.stats.data_count +=
5601 elm->embed.stats.data_count;
5602 parent->bref.embed.stats.inode_count +=
5603 elm->embed.stats.inode_count;
5605 if (elm->type == HAMMER2_BREF_TYPE_INODE)
5607 if (parent->bref.leaf_count + elm->leaf_count <
5608 HAMMER2_BLOCKREF_LEAF_MAX) {
5609 parent->bref.leaf_count += elm->leaf_count;
5611 parent->bref.leaf_count = HAMMER2_BLOCKREF_LEAF_MAX;
5614 case HAMMER2_BREF_TYPE_DIRENT:
5615 if (parent->bref.leaf_count != HAMMER2_BLOCKREF_LEAF_MAX)
5616 ++parent->bref.leaf_count;
5624 * We can only optimize parent->core.live_zero for live chains.
5626 if (i == count && parent->core.live_zero < count) {
5627 i = parent->core.live_zero++;
5632 xkey = elm->key + ((hammer2_key_t)1 << elm->keybits) - 1;
5633 if (i != count && (base[i].key < elm->key || xkey >= base[i].key)) {
5634 hammer2_spin_unex(&parent->core.spin);
5635 panic("insert base %p overlapping elements at %d elm %p\n",
5640 * Try to find an empty slot before or after.
5644 while (j > 0 || k < count) {
5646 if (j >= 0 && base[j].type == HAMMER2_BREF_TYPE_EMPTY) {
5650 bcopy(&base[j+1], &base[j],
5651 (i - j - 1) * sizeof(*base));
5657 if (k < count && base[k].type == HAMMER2_BREF_TYPE_EMPTY) {
5658 bcopy(&base[i], &base[i+1],
5659 (k - i) * sizeof(hammer2_blockref_t));
5663 * We can only update parent->core.live_zero for live
5666 if (parent->core.live_zero <= k)
5667 parent->core.live_zero = k + 1;
5672 panic("hammer2_base_insert: no room!");
5679 for (l = 0; l < count; ++l) {
5680 if (base[l].type != HAMMER2_BREF_TYPE_EMPTY) {
5681 key_next = base[l].key +
5682 ((hammer2_key_t)1 << base[l].keybits) - 1;
5686 while (++l < count) {
5687 if (base[l].type != HAMMER2_BREF_TYPE_EMPTY) {
5688 if (base[l].key <= key_next)
5689 panic("base_insert %d %d,%d,%d fail %p:%d", u, i, j, k, base, l);
5690 key_next = base[l].key +
5691 ((hammer2_key_t)1 << base[l].keybits) - 1;
5701 * Sort the blockref array for the chain. Used by the flush code to
5702 * sort the blockref[] array.
5704 * The chain must be exclusively locked AND spin-locked.
5706 typedef hammer2_blockref_t *hammer2_blockref_p;
5710 hammer2_base_sort_callback(const void *v1, const void *v2)
5712 hammer2_blockref_p bref1 = *(const hammer2_blockref_p *)v1;
5713 hammer2_blockref_p bref2 = *(const hammer2_blockref_p *)v2;
5716 * Make sure empty elements are placed at the end of the array
5718 if (bref1->type == HAMMER2_BREF_TYPE_EMPTY) {
5719 if (bref2->type == HAMMER2_BREF_TYPE_EMPTY)
5722 } else if (bref2->type == HAMMER2_BREF_TYPE_EMPTY) {
5729 if (bref1->key < bref2->key)
5731 if (bref1->key > bref2->key)
5737 hammer2_base_sort(hammer2_chain_t *chain)
5739 hammer2_blockref_t *base;
5742 switch(chain->bref.type) {
5743 case HAMMER2_BREF_TYPE_INODE:
5745 * Special shortcut for embedded data returns the inode
5746 * itself. Callers must detect this condition and access
5747 * the embedded data (the strategy code does this for us).
5749 * This is only applicable to regular files and softlinks.
5751 if (chain->data->ipdata.meta.op_flags &
5752 HAMMER2_OPFLAG_DIRECTDATA) {
5755 base = &chain->data->ipdata.u.blockset.blockref[0];
5756 count = HAMMER2_SET_COUNT;
5758 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
5759 case HAMMER2_BREF_TYPE_INDIRECT:
5761 * Optimize indirect blocks in the INITIAL state to avoid
5764 KKASSERT((chain->flags & HAMMER2_CHAIN_INITIAL) == 0);
5765 base = &chain->data->npdata[0];
5766 count = chain->bytes / sizeof(hammer2_blockref_t);
5768 case HAMMER2_BREF_TYPE_VOLUME:
5769 base = &chain->data->voldata.sroot_blockset.blockref[0];
5770 count = HAMMER2_SET_COUNT;
5772 case HAMMER2_BREF_TYPE_FREEMAP:
5773 base = &chain->data->blkset.blockref[0];
5774 count = HAMMER2_SET_COUNT;
5777 kprintf("hammer2_base_sort: unrecognized "
5778 "blockref(A) type: %d",
5781 tsleep(&base, 0, "dead", 0);
5782 panic("hammer2_base_sort: unrecognized "
5783 "blockref(A) type: %d",
5785 base = NULL; /* safety */
5786 count = 0; /* safety */
5788 kqsort(base, count, sizeof(*base), hammer2_base_sort_callback);
5794 * Chain memory management
5797 hammer2_chain_wait(hammer2_chain_t *chain)
5799 tsleep(chain, 0, "chnflw", 1);
5802 const hammer2_media_data_t *
5803 hammer2_chain_rdata(hammer2_chain_t *chain)
5805 KKASSERT(chain->data != NULL);
5806 return (chain->data);
5809 hammer2_media_data_t *
5810 hammer2_chain_wdata(hammer2_chain_t *chain)
5812 KKASSERT(chain->data != NULL);
5813 return (chain->data);
5817 * Set the check data for a chain. This can be a heavy-weight operation
5818 * and typically only runs on-flush. For file data check data is calculated
5819 * when the logical buffers are flushed.
5822 hammer2_chain_setcheck(hammer2_chain_t *chain, void *bdata)
5824 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_NOTTESTED);
5826 switch(HAMMER2_DEC_CHECK(chain->bref.methods)) {
5827 case HAMMER2_CHECK_NONE:
5829 case HAMMER2_CHECK_DISABLED:
5831 case HAMMER2_CHECK_ISCSI32:
5832 chain->bref.check.iscsi32.value =
5833 hammer2_icrc32(bdata, chain->bytes);
5835 case HAMMER2_CHECK_XXHASH64:
5836 chain->bref.check.xxhash64.value =
5837 XXH64(bdata, chain->bytes, XXH_HAMMER2_SEED);
5839 case HAMMER2_CHECK_SHA192:
5841 SHA256_CTX hash_ctx;
5843 uint8_t digest[SHA256_DIGEST_LENGTH];
5844 uint64_t digest64[SHA256_DIGEST_LENGTH/8];
5847 SHA256_Init(&hash_ctx);
5848 SHA256_Update(&hash_ctx, bdata, chain->bytes);
5849 SHA256_Final(u.digest, &hash_ctx);
5850 u.digest64[2] ^= u.digest64[3];
5852 chain->bref.check.sha192.data,
5853 sizeof(chain->bref.check.sha192.data));
5856 case HAMMER2_CHECK_FREEMAP:
5857 chain->bref.check.freemap.icrc32 =
5858 hammer2_icrc32(bdata, chain->bytes);
5861 kprintf("hammer2_chain_setcheck: unknown check type %02x\n",
5862 chain->bref.methods);
5868 * Characterize a failed check code and try to trace back to the inode.
5871 hammer2_characterize_failed_chain(hammer2_chain_t *chain, uint64_t check,
5874 hammer2_chain_t *lchain;
5875 hammer2_chain_t *ochain;
5878 did = krateprintf(&krate_h2chk,
5879 "chain %016jx.%02x (%s) meth=%02x CHECK FAIL "
5880 "(flags=%08x, bref/data ",
5881 chain->bref.data_off,
5883 hammer2_bref_type_str(chain->bref.type),
5884 chain->bref.methods,
5890 kprintf("%08x/%08x)\n",
5891 chain->bref.check.iscsi32.value,
5894 kprintf("%016jx/%016jx)\n",
5895 chain->bref.check.xxhash64.value,
5900 * Run up the chains to try to find the governing inode so we
5903 * XXX This error reporting is not really MPSAFE
5907 while (chain && chain->bref.type != HAMMER2_BREF_TYPE_INODE) {
5909 chain = chain->parent;
5912 if (chain && chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
5913 ((chain->bref.flags & HAMMER2_BREF_FLAG_PFSROOT) == 0 ||
5914 (lchain->bref.key & HAMMER2_DIRHASH_VISIBLE))) {
5915 kprintf(" Resides at/in inode %ld\n",
5917 } else if (chain && chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
5918 kprintf(" Resides in inode index - CRITICAL!!!\n");
5920 kprintf(" Resides in root index - CRITICAL!!!\n");
5923 const char *pfsname = "UNKNOWN";
5927 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
5928 if (ochain->pmp->pfs_hmps[i] == ochain->hmp &&
5929 ochain->pmp->pfs_names[i]) {
5930 pfsname = ochain->pmp->pfs_names[i];
5935 kprintf(" In pfs %s on device %s\n",
5936 pfsname, ochain->hmp->devrepname);
5941 * Returns non-zero on success, 0 on failure.
5944 hammer2_chain_testcheck(hammer2_chain_t *chain, void *bdata)
5950 if (chain->flags & HAMMER2_CHAIN_NOTTESTED)
5953 switch(HAMMER2_DEC_CHECK(chain->bref.methods)) {
5954 case HAMMER2_CHECK_NONE:
5957 case HAMMER2_CHECK_DISABLED:
5960 case HAMMER2_CHECK_ISCSI32:
5961 check32 = hammer2_icrc32(bdata, chain->bytes);
5962 r = (chain->bref.check.iscsi32.value == check32);
5964 hammer2_characterize_failed_chain(chain, check32, 32);
5966 hammer2_process_icrc32 += chain->bytes;
5968 case HAMMER2_CHECK_XXHASH64:
5969 check64 = XXH64(bdata, chain->bytes, XXH_HAMMER2_SEED);
5970 r = (chain->bref.check.xxhash64.value == check64);
5972 hammer2_characterize_failed_chain(chain, check64, 64);
5974 hammer2_process_xxhash64 += chain->bytes;
5976 case HAMMER2_CHECK_SHA192:
5978 SHA256_CTX hash_ctx;
5980 uint8_t digest[SHA256_DIGEST_LENGTH];
5981 uint64_t digest64[SHA256_DIGEST_LENGTH/8];
5984 SHA256_Init(&hash_ctx);
5985 SHA256_Update(&hash_ctx, bdata, chain->bytes);
5986 SHA256_Final(u.digest, &hash_ctx);
5987 u.digest64[2] ^= u.digest64[3];
5989 chain->bref.check.sha192.data,
5990 sizeof(chain->bref.check.sha192.data)) == 0) {
5994 krateprintf(&krate_h2chk,
5995 "chain %016jx.%02x meth=%02x "
5997 chain->bref.data_off,
5999 chain->bref.methods);
6003 case HAMMER2_CHECK_FREEMAP:
6004 r = (chain->bref.check.freemap.icrc32 ==
6005 hammer2_icrc32(bdata, chain->bytes));
6009 did = krateprintf(&krate_h2chk,
6010 "chain %016jx.%02x meth=%02x "
6012 chain->bref.data_off,
6014 chain->bref.methods);
6016 kprintf("freemap.icrc %08x icrc32 %08x (%d)\n",
6017 chain->bref.check.freemap.icrc32,
6018 hammer2_icrc32(bdata, chain->bytes),
6021 kprintf("dio %p buf %016jx,%d "
6024 chain->dio->bp->b_loffset,
6025 chain->dio->bp->b_bufsize,
6027 chain->dio->bp->b_data);
6033 kprintf("hammer2_chain_testcheck: unknown check type %02x\n",
6034 chain->bref.methods);
6042 * Acquire the chain and parent representing the specified inode for the
6043 * device at the specified cluster index.
6045 * The flags passed in are LOOKUP flags, not RESOLVE flags.
6047 * If we are unable to locate the inode, HAMMER2_ERROR_EIO is returned and
6048 * *chainp will be NULL. *parentp may still be set error or not, or NULL
6049 * if the parent itself could not be resolved.
6051 * The caller may pass-in a locked *parentp and/or *chainp, or neither.
6052 * They will be unlocked and released by this function. The *parentp and
6053 * *chainp representing the located inode are returned locked.
6056 hammer2_chain_inode_find(hammer2_pfs_t *pmp, hammer2_key_t inum,
6057 int clindex, int flags,
6058 hammer2_chain_t **parentp, hammer2_chain_t **chainp)
6060 hammer2_chain_t *parent;
6061 hammer2_chain_t *rchain;
6062 hammer2_key_t key_dummy;
6063 hammer2_inode_t *ip;
6067 resolve_flags = (flags & HAMMER2_LOOKUP_SHARED) ?
6068 HAMMER2_RESOLVE_SHARED : 0;
6071 * Caller expects us to replace these.
6074 hammer2_chain_unlock(*chainp);
6075 hammer2_chain_drop(*chainp);
6079 hammer2_chain_unlock(*parentp);
6080 hammer2_chain_drop(*parentp);
6085 * Be very careful, this is a backend function and we CANNOT
6086 * lock any frontend inode structure we find. But we have to
6087 * look the inode up this way first in case it exists but is
6088 * detached from the radix tree.
6090 ip = hammer2_inode_lookup(pmp, inum);
6092 *chainp = hammer2_inode_chain_and_parent(ip, clindex,
6095 hammer2_inode_drop(ip);
6098 hammer2_chain_unlock(*chainp);
6099 hammer2_chain_drop(*chainp);
6102 hammer2_chain_unlock(*parentp);
6103 hammer2_chain_drop(*parentp);
6109 * Inodes hang off of the iroot (bit 63 is clear, differentiating
6110 * inodes from root directory entries in the key lookup).
6112 parent = hammer2_inode_chain(pmp->iroot, clindex, resolve_flags);
6115 rchain = hammer2_chain_lookup(&parent, &key_dummy,
6119 error = HAMMER2_ERROR_EIO;
6128 * Used by the bulkscan code to snapshot the synchronized storage for
6129 * a volume, allowing it to be scanned concurrently against normal
6133 hammer2_chain_bulksnap(hammer2_dev_t *hmp)
6135 hammer2_chain_t *copy;
6137 copy = hammer2_chain_alloc(hmp, hmp->spmp, &hmp->vchain.bref);
6138 copy->data = kmalloc(sizeof(copy->data->voldata),
6141 hammer2_voldata_lock(hmp);
6142 copy->data->voldata = hmp->volsync;
6143 hammer2_voldata_unlock(hmp);
6149 hammer2_chain_bulkdrop(hammer2_chain_t *copy)
6151 KKASSERT(copy->bref.type == HAMMER2_BREF_TYPE_VOLUME);
6152 KKASSERT(copy->data);
6153 kfree(copy->data, copy->hmp->mchain);
6155 atomic_add_long(&hammer2_chain_allocs, -1);
6156 hammer2_chain_drop(copy);
6160 * Returns non-zero if the chain (INODE or DIRENT) matches the
6164 hammer2_chain_dirent_test(hammer2_chain_t *chain, const char *name,
6167 const hammer2_inode_data_t *ripdata;
6168 const hammer2_dirent_head_t *den;
6170 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
6171 ripdata = &chain->data->ipdata;
6172 if (ripdata->meta.name_len == name_len &&
6173 bcmp(ripdata->filename, name, name_len) == 0) {
6177 if (chain->bref.type == HAMMER2_BREF_TYPE_DIRENT &&
6178 chain->bref.embed.dirent.namlen == name_len) {
6179 den = &chain->bref.embed.dirent;
6180 if (name_len > sizeof(chain->bref.check.buf) &&
6181 bcmp(chain->data->buf, name, name_len) == 0) {
6184 if (name_len <= sizeof(chain->bref.check.buf) &&
6185 bcmp(chain->bref.check.buf, name, name_len) == 0) {