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 void hammer2_chain_rename_obref(hammer2_chain_t **parentp,
72 hammer2_chain_t *chain, hammer2_tid_t mtid,
73 int flags, hammer2_blockref_t *obref);
74 static int hammer2_chain_delete_obref(hammer2_chain_t *parent,
75 hammer2_chain_t *chain,
76 hammer2_tid_t mtid, int flags,
77 hammer2_blockref_t *obref);
78 static hammer2_io_t *hammer2_chain_drop_data(hammer2_chain_t *chain);
79 static hammer2_chain_t *hammer2_combined_find(
80 hammer2_chain_t *parent,
81 hammer2_blockref_t *base, int count,
82 hammer2_key_t *key_nextp,
83 hammer2_key_t key_beg, hammer2_key_t key_end,
84 hammer2_blockref_t **bresp);
87 * There are many degenerate situations where an extreme rate of console
88 * output can occur from warnings and errors. Make sure this output does
89 * not impede operations.
91 static struct krate krate_h2chk = { .freq = 5 };
92 static struct krate krate_h2me = { .freq = 1 };
93 static struct krate krate_h2em = { .freq = 1 };
96 * Basic RBTree for chains (core.rbtree).
98 RB_GENERATE(hammer2_chain_tree, hammer2_chain, rbnode, hammer2_chain_cmp);
101 hammer2_chain_cmp(hammer2_chain_t *chain1, hammer2_chain_t *chain2)
103 hammer2_key_t c1_beg;
104 hammer2_key_t c1_end;
105 hammer2_key_t c2_beg;
106 hammer2_key_t c2_end;
109 * Compare chains. Overlaps are not supposed to happen and catch
110 * any software issues early we count overlaps as a match.
112 c1_beg = chain1->bref.key;
113 c1_end = c1_beg + ((hammer2_key_t)1 << chain1->bref.keybits) - 1;
114 c2_beg = chain2->bref.key;
115 c2_end = c2_beg + ((hammer2_key_t)1 << chain2->bref.keybits) - 1;
117 if (c1_end < c2_beg) /* fully to the left */
119 if (c1_beg > c2_end) /* fully to the right */
121 return(0); /* overlap (must not cross edge boundary) */
125 * Assert that a chain has no media data associated with it.
128 hammer2_chain_assert_no_data(hammer2_chain_t *chain)
130 KKASSERT(chain->dio == NULL);
131 if (chain->bref.type != HAMMER2_BREF_TYPE_VOLUME &&
132 chain->bref.type != HAMMER2_BREF_TYPE_FREEMAP &&
134 panic("hammer2_chain_assert_no_data: chain %p still has data",
140 * Make a chain visible to the flusher. The flusher operates using a top-down
141 * recursion based on the ONFLUSH flag. It locates MODIFIED and UPDATE chains,
142 * flushes them, and updates blocks back to the volume root.
144 * This routine sets the ONFLUSH flag upward from the triggering chain until
145 * it hits an inode root or the volume root. Inode chains serve as inflection
146 * points, requiring the flusher to bridge across trees. Inodes include
147 * regular inodes, PFS roots (pmp->iroot), and the media super root
151 hammer2_chain_setflush(hammer2_chain_t *chain)
153 hammer2_chain_t *parent;
155 if ((chain->flags & HAMMER2_CHAIN_ONFLUSH) == 0) {
156 hammer2_spin_sh(&chain->core.spin);
157 while ((chain->flags & HAMMER2_CHAIN_ONFLUSH) == 0) {
158 atomic_set_int(&chain->flags, HAMMER2_CHAIN_ONFLUSH);
159 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE)
161 if ((parent = chain->parent) == NULL)
163 hammer2_spin_sh(&parent->core.spin);
164 hammer2_spin_unsh(&chain->core.spin);
167 hammer2_spin_unsh(&chain->core.spin);
172 * Allocate a new disconnected chain element representing the specified
173 * bref. chain->refs is set to 1 and the passed bref is copied to
174 * chain->bref. chain->bytes is derived from the bref.
176 * chain->pmp inherits pmp unless the chain is an inode (other than the
179 * NOTE: Returns a referenced but unlocked (because there is no core) chain.
182 hammer2_chain_alloc(hammer2_dev_t *hmp, hammer2_pfs_t *pmp,
183 hammer2_blockref_t *bref)
185 hammer2_chain_t *chain;
189 * Special case - radix of 0 indicates a chain that does not
190 * need a data reference (context is completely embedded in the
193 if ((int)(bref->data_off & HAMMER2_OFF_MASK_RADIX))
194 bytes = 1U << (int)(bref->data_off & HAMMER2_OFF_MASK_RADIX);
198 atomic_add_long(&hammer2_chain_allocs, 1);
201 * Construct the appropriate system structure.
204 case HAMMER2_BREF_TYPE_DIRENT:
205 case HAMMER2_BREF_TYPE_INODE:
206 case HAMMER2_BREF_TYPE_INDIRECT:
207 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
208 case HAMMER2_BREF_TYPE_DATA:
209 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
210 chain = kmalloc(sizeof(*chain), hmp->mchain, M_WAITOK | M_ZERO);
212 case HAMMER2_BREF_TYPE_VOLUME:
213 case HAMMER2_BREF_TYPE_FREEMAP:
215 * Only hammer2_chain_bulksnap() calls this function with these
218 chain = kmalloc(sizeof(*chain), hmp->mchain, M_WAITOK | M_ZERO);
222 panic("hammer2_chain_alloc: unrecognized blockref type: %d",
227 * Initialize the new chain structure. pmp must be set to NULL for
228 * chains belonging to the super-root topology of a device mount.
230 if (pmp == hmp->spmp)
237 chain->bytes = bytes;
239 chain->flags = HAMMER2_CHAIN_ALLOCATED;
240 lockinit(&chain->diolk, "chdio", 0, 0);
243 * Set the PFS boundary flag if this chain represents a PFS root.
245 if (bref->flags & HAMMER2_BREF_FLAG_PFSROOT)
246 atomic_set_int(&chain->flags, HAMMER2_CHAIN_PFSBOUNDARY);
247 hammer2_chain_core_init(chain);
253 * Initialize a chain's core structure. This structure used to be allocated
254 * but is now embedded.
256 * The core is not locked. No additional refs on the chain are made.
257 * (trans) must not be NULL if (core) is not NULL.
260 hammer2_chain_core_init(hammer2_chain_t *chain)
263 * Fresh core under nchain (no multi-homing of ochain's
266 RB_INIT(&chain->core.rbtree); /* live chains */
267 hammer2_mtx_init(&chain->lock, "h2chain");
271 * Add a reference to a chain element, preventing its destruction.
273 * (can be called with spinlock held)
276 hammer2_chain_ref(hammer2_chain_t *chain)
278 if (atomic_fetchadd_int(&chain->refs, 1) == 0) {
280 * Just flag that the chain was used and should be recycled
281 * on the LRU if it encounters it later.
283 if (chain->flags & HAMMER2_CHAIN_ONLRU)
284 atomic_set_int(&chain->flags, HAMMER2_CHAIN_LRUHINT);
288 * REMOVED - reduces contention, lru_list is more heuristical
291 * 0->non-zero transition must ensure that chain is removed
294 * NOTE: Already holding lru_spin here so we cannot call
295 * hammer2_chain_ref() to get it off lru_list, do
298 if (chain->flags & HAMMER2_CHAIN_ONLRU) {
299 hammer2_pfs_t *pmp = chain->pmp;
300 hammer2_spin_ex(&pmp->lru_spin);
301 if (chain->flags & HAMMER2_CHAIN_ONLRU) {
302 atomic_add_int(&pmp->lru_count, -1);
303 atomic_clear_int(&chain->flags,
304 HAMMER2_CHAIN_ONLRU);
305 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node);
307 hammer2_spin_unex(&pmp->lru_spin);
314 * Ref a locked chain and force the data to be held across an unlock.
315 * Chain must be currently locked. The user of the chain who desires
316 * to release the hold must call hammer2_chain_lock_unhold() to relock
317 * and unhold the chain, then unlock normally, or may simply call
318 * hammer2_chain_drop_unhold() (which is safer against deadlocks).
321 hammer2_chain_ref_hold(hammer2_chain_t *chain)
323 atomic_add_int(&chain->lockcnt, 1);
324 hammer2_chain_ref(chain);
328 * Insert the chain in the core rbtree.
330 * Normal insertions are placed in the live rbtree. Insertion of a deleted
331 * chain is a special case used by the flush code that is placed on the
332 * unstaged deleted list to avoid confusing the live view.
334 #define HAMMER2_CHAIN_INSERT_SPIN 0x0001
335 #define HAMMER2_CHAIN_INSERT_LIVE 0x0002
336 #define HAMMER2_CHAIN_INSERT_RACE 0x0004
340 hammer2_chain_insert(hammer2_chain_t *parent, hammer2_chain_t *chain,
341 int flags, int generation)
343 hammer2_chain_t *xchain;
346 if (flags & HAMMER2_CHAIN_INSERT_SPIN)
347 hammer2_spin_ex(&parent->core.spin);
350 * Interlocked by spinlock, check for race
352 if ((flags & HAMMER2_CHAIN_INSERT_RACE) &&
353 parent->core.generation != generation) {
354 error = HAMMER2_ERROR_EAGAIN;
361 xchain = RB_INSERT(hammer2_chain_tree, &parent->core.rbtree, chain);
362 KASSERT(xchain == NULL,
363 ("hammer2_chain_insert: collision %p %p (key=%016jx)",
364 chain, xchain, chain->bref.key));
365 atomic_set_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE);
366 chain->parent = parent;
367 ++parent->core.chain_count;
368 ++parent->core.generation; /* XXX incs for _get() too, XXX */
371 * We have to keep track of the effective live-view blockref count
372 * so the create code knows when to push an indirect block.
374 if (flags & HAMMER2_CHAIN_INSERT_LIVE)
375 atomic_add_int(&parent->core.live_count, 1);
377 if (flags & HAMMER2_CHAIN_INSERT_SPIN)
378 hammer2_spin_unex(&parent->core.spin);
383 * Drop the caller's reference to the chain. When the ref count drops to
384 * zero this function will try to disassociate the chain from its parent and
385 * deallocate it, then recursely drop the parent using the implied ref
386 * from the chain's chain->parent.
388 * Nobody should own chain's mutex on the 1->0 transition, unless this drop
389 * races an acquisition by another cpu. Therefore we can loop if we are
390 * unable to acquire the mutex, and refs is unlikely to be 1 unless we again
391 * race against another drop.
393 static hammer2_chain_t *hammer2_chain_lastdrop(hammer2_chain_t *chain,
395 static void hammer2_chain_lru_flush(hammer2_pfs_t *pmp);
398 hammer2_chain_drop(hammer2_chain_t *chain)
402 KKASSERT(chain->refs > 0);
410 if (hammer2_mtx_ex_try(&chain->lock) == 0)
411 chain = hammer2_chain_lastdrop(chain, 0);
412 /* retry the same chain, or chain from lastdrop */
414 if (atomic_cmpset_int(&chain->refs, refs, refs - 1))
416 /* retry the same chain */
423 * Unhold a held and probably not-locked chain, ensure that the data is
424 * dropped on the 1->0 transition of lockcnt by obtaining an exclusive
425 * lock and then simply unlocking the chain.
428 hammer2_chain_unhold(hammer2_chain_t *chain)
434 lockcnt = chain->lockcnt;
437 if (atomic_cmpset_int(&chain->lockcnt,
438 lockcnt, lockcnt - 1)) {
441 } else if (hammer2_mtx_ex_try(&chain->lock) == 0) {
442 hammer2_chain_unlock(chain);
446 * This situation can easily occur on SMP due to
447 * the gap inbetween the 1->0 transition and the
448 * final unlock. We cannot safely block on the
449 * mutex because lockcnt might go above 1.
451 * XXX Sleep for one tick if it takes too long.
454 if (iter > 1000 + hz) {
455 kprintf("hammer2: h2race1 %p\n", chain);
458 tsleep(&iter, 0, "h2race1", 1);
466 hammer2_chain_drop_unhold(hammer2_chain_t *chain)
468 hammer2_chain_unhold(chain);
469 hammer2_chain_drop(chain);
473 hammer2_chain_rehold(hammer2_chain_t *chain)
475 hammer2_chain_lock(chain, HAMMER2_RESOLVE_SHARED);
476 atomic_add_int(&chain->lockcnt, 1);
477 hammer2_chain_unlock(chain);
481 * Handles the (potential) last drop of chain->refs from 1->0. Called with
482 * the mutex exclusively locked, refs == 1, and lockcnt 0. SMP races are
483 * possible against refs and lockcnt. We must dispose of the mutex on chain.
485 * This function returns an unlocked chain for recursive drop or NULL. It
486 * can return the same chain if it determines it has raced another ref.
490 * When two chains need to be recursively dropped we use the chain we
491 * would otherwise free to placehold the additional chain. It's a bit
492 * convoluted but we can't just recurse without potentially blowing out
495 * The chain cannot be freed if it has any children.
496 * The chain cannot be freed if flagged MODIFIED unless we can dispose of it.
497 * The chain cannot be freed if flagged UPDATE unless we can dispose of it.
498 * Any dedup registration can remain intact.
500 * The core spinlock is allowed to nest child-to-parent (not parent-to-child).
504 hammer2_chain_lastdrop(hammer2_chain_t *chain, int depth)
508 hammer2_chain_t *parent;
509 hammer2_chain_t *rdrop;
512 * We need chain's spinlock to interlock the sub-tree test.
513 * We already have chain's mutex, protecting chain->parent.
515 * Remember that chain->refs can be in flux.
517 hammer2_spin_ex(&chain->core.spin);
519 if (chain->parent != NULL) {
521 * If the chain has a parent the UPDATE bit prevents scrapping
522 * as the chain is needed to properly flush the parent. Try
523 * to complete the 1->0 transition and return NULL. Retry
524 * (return chain) if we are unable to complete the 1->0
525 * transition, else return NULL (nothing more to do).
527 * If the chain has a parent the MODIFIED bit prevents
530 * Chains with UPDATE/MODIFIED are *not* put on the LRU list!
532 if (chain->flags & (HAMMER2_CHAIN_UPDATE |
533 HAMMER2_CHAIN_MODIFIED)) {
534 if (atomic_cmpset_int(&chain->refs, 1, 0)) {
535 hammer2_spin_unex(&chain->core.spin);
536 hammer2_chain_assert_no_data(chain);
537 hammer2_mtx_unlock(&chain->lock);
540 hammer2_spin_unex(&chain->core.spin);
541 hammer2_mtx_unlock(&chain->lock);
545 /* spinlock still held */
546 } else if (chain->bref.type == HAMMER2_BREF_TYPE_VOLUME ||
547 chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP) {
549 * Retain the static vchain and fchain. Clear bits that
550 * are not relevant. Do not clear the MODIFIED bit,
551 * and certainly do not put it on the delayed-flush queue.
553 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
556 * The chain has no parent and can be flagged for destruction.
557 * Since it has no parent, UPDATE can also be cleared.
559 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DESTROY);
560 if (chain->flags & HAMMER2_CHAIN_UPDATE)
561 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
564 * If the chain has children we must propagate the DESTROY
565 * flag downward and rip the disconnected topology apart.
566 * This is accomplished by calling hammer2_flush() on the
569 * Any dedup is already handled by the underlying DIO, so
570 * we do not have to specifically flush it here.
572 if (chain->core.chain_count) {
573 hammer2_spin_unex(&chain->core.spin);
574 hammer2_flush(chain, HAMMER2_FLUSH_TOP |
576 hammer2_mtx_unlock(&chain->lock);
578 return(chain); /* retry drop */
582 * Otherwise we can scrap the MODIFIED bit if it is set,
583 * and continue along the freeing path.
585 * Be sure to clean-out any dedup bits. Without a parent
586 * this chain will no longer be visible to the flush code.
587 * Easy check data_off to avoid the volume root.
589 if (chain->flags & HAMMER2_CHAIN_MODIFIED) {
590 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
591 atomic_add_long(&hammer2_count_modified_chains, -1);
593 hammer2_pfs_memory_wakeup(chain->pmp, -1);
595 /* spinlock still held */
598 /* spinlock still held */
601 * If any children exist we must leave the chain intact with refs == 0.
602 * They exist because chains are retained below us which have refs or
603 * may require flushing.
605 * Retry (return chain) if we fail to transition the refs to 0, else
606 * return NULL indication nothing more to do.
608 * Chains with children are NOT put on the LRU list.
610 if (chain->core.chain_count) {
611 if (atomic_cmpset_int(&chain->refs, 1, 0)) {
612 hammer2_spin_unex(&chain->core.spin);
613 hammer2_chain_assert_no_data(chain);
614 hammer2_mtx_unlock(&chain->lock);
617 hammer2_spin_unex(&chain->core.spin);
618 hammer2_mtx_unlock(&chain->lock);
622 /* spinlock still held */
623 /* no chains left under us */
626 * chain->core has no children left so no accessors can get to our
627 * chain from there. Now we have to lock the parent core to interlock
628 * remaining possible accessors that might bump chain's refs before
629 * we can safely drop chain's refs with intent to free the chain.
632 pmp = chain->pmp; /* can be NULL */
635 parent = chain->parent;
638 * WARNING! chain's spin lock is still held here, and other spinlocks
639 * will be acquired and released in the code below. We
640 * cannot be making fancy procedure calls!
644 * We can cache the chain if it is associated with a pmp
645 * and not flagged as being destroyed or requesting a full
646 * release. In this situation the chain is not removed
647 * from its parent, i.e. it can still be looked up.
649 * We intentionally do not cache DATA chains because these
650 * were likely used to load data into the logical buffer cache
651 * and will not be accessed again for some time.
654 (HAMMER2_CHAIN_DESTROY | HAMMER2_CHAIN_RELEASE)) == 0 &&
656 chain->bref.type != HAMMER2_BREF_TYPE_DATA) {
658 hammer2_spin_ex(&parent->core.spin);
659 if (atomic_cmpset_int(&chain->refs, 1, 0) == 0) {
661 * 1->0 transition failed, retry. Do not drop
662 * the chain's data yet!
665 hammer2_spin_unex(&parent->core.spin);
666 hammer2_spin_unex(&chain->core.spin);
667 hammer2_mtx_unlock(&chain->lock);
675 hammer2_chain_assert_no_data(chain);
678 * Make sure we are on the LRU list, clean up excessive
679 * LRU entries. We can only really drop one but there might
680 * be other entries that we can remove from the lru_list
683 * NOTE: HAMMER2_CHAIN_ONLRU may only be safely set when
684 * chain->core.spin AND pmp->lru_spin are held, but
685 * can be safely cleared only holding pmp->lru_spin.
687 if ((chain->flags & HAMMER2_CHAIN_ONLRU) == 0) {
688 hammer2_spin_ex(&pmp->lru_spin);
689 if ((chain->flags & HAMMER2_CHAIN_ONLRU) == 0) {
690 atomic_set_int(&chain->flags,
691 HAMMER2_CHAIN_ONLRU);
692 TAILQ_INSERT_TAIL(&pmp->lru_list,
694 atomic_add_int(&pmp->lru_count, 1);
696 if (pmp->lru_count < HAMMER2_LRU_LIMIT)
697 depth = 1; /* disable lru_list flush */
698 hammer2_spin_unex(&pmp->lru_spin);
700 /* disable lru flush */
705 hammer2_spin_unex(&parent->core.spin);
706 parent = NULL; /* safety */
708 hammer2_spin_unex(&chain->core.spin);
709 hammer2_mtx_unlock(&chain->lock);
712 * lru_list hysteresis (see above for depth overrides).
713 * Note that depth also prevents excessive lastdrop recursion.
716 hammer2_chain_lru_flush(pmp);
723 * Make sure we are not on the LRU list.
725 if (chain->flags & HAMMER2_CHAIN_ONLRU) {
726 hammer2_spin_ex(&pmp->lru_spin);
727 if (chain->flags & HAMMER2_CHAIN_ONLRU) {
728 atomic_add_int(&pmp->lru_count, -1);
729 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU);
730 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node);
732 hammer2_spin_unex(&pmp->lru_spin);
736 * Spinlock the parent and try to drop the last ref on chain.
737 * On success determine if we should dispose of the chain
738 * (remove the chain from its parent, etc).
740 * (normal core locks are top-down recursive but we define
741 * core spinlocks as bottom-up recursive, so this is safe).
744 hammer2_spin_ex(&parent->core.spin);
745 if (atomic_cmpset_int(&chain->refs, 1, 0) == 0) {
747 * 1->0 transition failed, retry.
749 hammer2_spin_unex(&parent->core.spin);
750 hammer2_spin_unex(&chain->core.spin);
751 hammer2_mtx_unlock(&chain->lock);
757 * 1->0 transition successful, parent spin held to prevent
758 * new lookups, chain spinlock held to protect parent field.
759 * Remove chain from the parent.
761 * If the chain is being removed from the parent's btree but
762 * is not bmapped, we have to adjust live_count downward. If
763 * it is bmapped then the blockref is retained in the parent
764 * as is its associated live_count. This case can occur when
765 * a chain added to the topology is unable to flush and is
766 * then later deleted.
768 if (chain->flags & HAMMER2_CHAIN_ONRBTREE) {
769 if ((parent->flags & HAMMER2_CHAIN_COUNTEDBREFS) &&
770 (chain->flags & HAMMER2_CHAIN_BMAPPED) == 0) {
771 atomic_add_int(&parent->core.live_count, -1);
773 RB_REMOVE(hammer2_chain_tree,
774 &parent->core.rbtree, chain);
775 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE);
776 --parent->core.chain_count;
777 chain->parent = NULL;
781 * If our chain was the last chain in the parent's core the
782 * core is now empty and its parent might have to be
783 * re-dropped if it has 0 refs.
785 if (parent->core.chain_count == 0) {
787 atomic_add_int(&rdrop->refs, 1);
789 if (atomic_cmpset_int(&rdrop->refs, 0, 1) == 0)
793 hammer2_spin_unex(&parent->core.spin);
794 parent = NULL; /* safety */
800 if (atomic_cmpset_int(&chain->refs, 1, 0) == 0) {
802 * 1->0 transition failed, retry.
804 hammer2_spin_unex(&parent->core.spin);
805 hammer2_spin_unex(&chain->core.spin);
806 hammer2_mtx_unlock(&chain->lock);
813 * Successful 1->0 transition, no parent, no children... no way for
814 * anyone to ref this chain any more. We can clean-up and free it.
816 * We still have the core spinlock, and core's chain_count is 0.
817 * Any parent spinlock is gone.
819 hammer2_spin_unex(&chain->core.spin);
820 hammer2_chain_assert_no_data(chain);
821 hammer2_mtx_unlock(&chain->lock);
822 KKASSERT(RB_EMPTY(&chain->core.rbtree) &&
823 chain->core.chain_count == 0);
826 * All locks are gone, no pointers remain to the chain, finish
829 KKASSERT((chain->flags & (HAMMER2_CHAIN_UPDATE |
830 HAMMER2_CHAIN_MODIFIED)) == 0);
833 * Once chain resources are gone we can use the now dead chain
834 * structure to placehold what might otherwise require a recursive
835 * drop, because we have potentially two things to drop and can only
836 * return one directly.
838 if (chain->flags & HAMMER2_CHAIN_ALLOCATED) {
839 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ALLOCATED);
841 kfree(chain, hmp->mchain);
845 * Possible chaining loop when parent re-drop needed.
851 * Heuristical flush of the LRU, try to reduce the number of entries
852 * on the LRU to (HAMMER2_LRU_LIMIT * 2 / 3). This procedure is called
853 * only when lru_count exceeds HAMMER2_LRU_LIMIT.
857 hammer2_chain_lru_flush(hammer2_pfs_t *pmp)
859 hammer2_chain_t *chain;
863 hammer2_spin_ex(&pmp->lru_spin);
864 while (pmp->lru_count > HAMMER2_LRU_LIMIT * 2 / 3) {
866 * Pick a chain off the lru_list, just recycle it quickly
867 * if LRUHINT is set (the chain was ref'd but left on
868 * the lru_list, so cycle to the end).
870 chain = TAILQ_FIRST(&pmp->lru_list);
871 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node);
873 if (chain->flags & HAMMER2_CHAIN_LRUHINT) {
874 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_LRUHINT);
875 TAILQ_INSERT_TAIL(&pmp->lru_list, chain, lru_node);
881 * Ok, we are off the LRU. We must adjust refs before we
882 * can safely clear the ONLRU flag.
884 atomic_add_int(&pmp->lru_count, -1);
885 if (atomic_cmpset_int(&chain->refs, 0, 1)) {
886 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU);
887 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
890 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU);
893 hammer2_spin_unex(&pmp->lru_spin);
898 * If we picked a chain off the lru list we may be able to lastdrop
899 * it. Use a depth of 1 to prevent excessive lastdrop recursion.
909 if (hammer2_mtx_ex_try(&chain->lock) == 0)
910 chain = hammer2_chain_lastdrop(chain, 1);
911 /* retry the same chain, or chain from lastdrop */
913 if (atomic_cmpset_int(&chain->refs, refs, refs - 1))
915 /* retry the same chain */
923 * On last lock release.
925 static hammer2_io_t *
926 hammer2_chain_drop_data(hammer2_chain_t *chain)
930 if ((dio = chain->dio) != NULL) {
934 switch(chain->bref.type) {
935 case HAMMER2_BREF_TYPE_VOLUME:
936 case HAMMER2_BREF_TYPE_FREEMAP:
939 if (chain->data != NULL) {
940 hammer2_spin_unex(&chain->core.spin);
941 panic("chain data not null: "
942 "chain %p bref %016jx.%02x "
943 "refs %d parent %p dio %p data %p",
944 chain, chain->bref.data_off,
945 chain->bref.type, chain->refs,
947 chain->dio, chain->data);
949 KKASSERT(chain->data == NULL);
957 * Lock a referenced chain element, acquiring its data with I/O if necessary,
958 * and specify how you would like the data to be resolved.
960 * If an I/O or other fatal error occurs, chain->error will be set to non-zero.
962 * The lock is allowed to recurse, multiple locking ops will aggregate
963 * the requested resolve types. Once data is assigned it will not be
964 * removed until the last unlock.
966 * HAMMER2_RESOLVE_NEVER - Do not resolve the data element.
967 * (typically used to avoid device/logical buffer
970 * HAMMER2_RESOLVE_MAYBE - Do not resolve data elements for chains in
971 * the INITIAL-create state (indirect blocks only).
973 * Do not resolve data elements for DATA chains.
974 * (typically used to avoid device/logical buffer
977 * HAMMER2_RESOLVE_ALWAYS- Always resolve the data element.
979 * HAMMER2_RESOLVE_SHARED- (flag) The chain is locked shared, otherwise
980 * it will be locked exclusive.
982 * HAMMER2_RESOLVE_NONBLOCK- (flag) The chain is locked non-blocking. If
983 * the lock fails, EAGAIN is returned.
985 * NOTE: Embedded elements (volume header, inodes) are always resolved
988 * NOTE: Specifying HAMMER2_RESOLVE_ALWAYS on a newly-created non-embedded
989 * element will instantiate and zero its buffer, and flush it on
992 * NOTE: (data) elements are normally locked RESOLVE_NEVER or RESOLVE_MAYBE
993 * so as not to instantiate a device buffer, which could alias against
994 * a logical file buffer. However, if ALWAYS is specified the
995 * device buffer will be instantiated anyway.
997 * NOTE: The return value is always 0 unless NONBLOCK is specified, in which
998 * case it can be either 0 or EAGAIN.
1000 * WARNING! This function blocks on I/O if data needs to be fetched. This
1001 * blocking can run concurrent with other compatible lock holders
1002 * who do not need data returning. The lock is not upgraded to
1003 * exclusive during a data fetch, a separate bit is used to
1004 * interlock I/O. However, an exclusive lock holder can still count
1005 * on being interlocked against an I/O fetch managed by a shared
1009 hammer2_chain_lock(hammer2_chain_t *chain, int how)
1011 KKASSERT(chain->refs > 0);
1013 if (how & HAMMER2_RESOLVE_NONBLOCK) {
1015 * We still have to bump lockcnt before acquiring the lock,
1016 * even for non-blocking operation, because the unlock code
1017 * live-loops on lockcnt == 1 when dropping the last lock.
1019 * If the non-blocking operation fails we have to use an
1020 * unhold sequence to undo the mess.
1022 * NOTE: LOCKAGAIN must always succeed without blocking,
1023 * even if NONBLOCK is specified.
1025 atomic_add_int(&chain->lockcnt, 1);
1026 if (how & HAMMER2_RESOLVE_SHARED) {
1027 if (how & HAMMER2_RESOLVE_LOCKAGAIN) {
1028 hammer2_mtx_sh_again(&chain->lock);
1030 if (hammer2_mtx_sh_try(&chain->lock) != 0) {
1031 hammer2_chain_unhold(chain);
1036 if (hammer2_mtx_ex_try(&chain->lock) != 0) {
1037 hammer2_chain_unhold(chain);
1043 * Get the appropriate lock. If LOCKAGAIN is flagged with
1044 * SHARED the caller expects a shared lock to already be
1045 * present and we are giving it another ref. This case must
1046 * importantly not block if there is a pending exclusive lock
1049 atomic_add_int(&chain->lockcnt, 1);
1050 if (how & HAMMER2_RESOLVE_SHARED) {
1051 if (how & HAMMER2_RESOLVE_LOCKAGAIN) {
1052 hammer2_mtx_sh_again(&chain->lock);
1054 hammer2_mtx_sh(&chain->lock);
1057 hammer2_mtx_ex(&chain->lock);
1062 * If we already have a valid data pointer make sure the data is
1063 * synchronized to the current cpu, and then no further action is
1068 hammer2_io_bkvasync(chain->dio);
1073 * Do we have to resolve the data? This is generally only
1074 * applicable to HAMMER2_BREF_TYPE_DATA which is special-cased.
1075 * Other BREF types expects the data to be there.
1077 switch(how & HAMMER2_RESOLVE_MASK) {
1078 case HAMMER2_RESOLVE_NEVER:
1080 case HAMMER2_RESOLVE_MAYBE:
1081 if (chain->flags & HAMMER2_CHAIN_INITIAL)
1083 if (chain->bref.type == HAMMER2_BREF_TYPE_DATA)
1086 if (chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE)
1088 if (chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_LEAF)
1092 case HAMMER2_RESOLVE_ALWAYS:
1098 * Caller requires data
1100 hammer2_chain_load_data(chain);
1106 * Lock the chain, retain the hold, and drop the data persistence count.
1107 * The data should remain valid because we never transitioned lockcnt
1111 hammer2_chain_lock_unhold(hammer2_chain_t *chain, int how)
1113 hammer2_chain_lock(chain, how);
1114 atomic_add_int(&chain->lockcnt, -1);
1119 * Downgrade an exclusive chain lock to a shared chain lock.
1121 * NOTE: There is no upgrade equivalent due to the ease of
1122 * deadlocks in that direction.
1125 hammer2_chain_lock_downgrade(hammer2_chain_t *chain)
1127 hammer2_mtx_downgrade(&chain->lock);
1132 * Issue I/O and install chain->data. Caller must hold a chain lock, lock
1133 * may be of any type.
1135 * Once chain->data is set it cannot be disposed of until all locks are
1138 * Make sure the data is synchronized to the current cpu.
1141 hammer2_chain_load_data(hammer2_chain_t *chain)
1143 hammer2_blockref_t *bref;
1150 * Degenerate case, data already present, or chain has no media
1151 * reference to load.
1153 KKASSERT(chain->lock.mtx_lock & MTX_MASK);
1156 hammer2_io_bkvasync(chain->dio);
1159 if ((chain->bref.data_off & ~HAMMER2_OFF_MASK_RADIX) == 0)
1163 KKASSERT(hmp != NULL);
1166 * Gain the IOINPROG bit, interlocked block.
1172 oflags = chain->flags;
1174 if (oflags & HAMMER2_CHAIN_IOINPROG) {
1175 nflags = oflags | HAMMER2_CHAIN_IOSIGNAL;
1176 tsleep_interlock(&chain->flags, 0);
1177 if (atomic_cmpset_int(&chain->flags, oflags, nflags)) {
1178 tsleep(&chain->flags, PINTERLOCKED,
1183 nflags = oflags | HAMMER2_CHAIN_IOINPROG;
1184 if (atomic_cmpset_int(&chain->flags, oflags, nflags)) {
1192 * We own CHAIN_IOINPROG
1194 * Degenerate case if we raced another load.
1198 hammer2_io_bkvasync(chain->dio);
1203 * We must resolve to a device buffer, either by issuing I/O or
1204 * by creating a zero-fill element. We do not mark the buffer
1205 * dirty when creating a zero-fill element (the hammer2_chain_modify()
1206 * API must still be used to do that).
1208 * The device buffer is variable-sized in powers of 2 down
1209 * to HAMMER2_MIN_ALLOC (typically 1K). A 64K physical storage
1210 * chunk always contains buffers of the same size. (XXX)
1212 * The minimum physical IO size may be larger than the variable
1215 bref = &chain->bref;
1218 * The getblk() optimization can only be used on newly created
1219 * elements if the physical block size matches the request.
1221 if (chain->flags & HAMMER2_CHAIN_INITIAL) {
1222 error = hammer2_io_new(hmp, bref->type,
1223 bref->data_off, chain->bytes,
1226 error = hammer2_io_bread(hmp, bref->type,
1227 bref->data_off, chain->bytes,
1229 hammer2_adjreadcounter(&chain->bref, chain->bytes);
1232 chain->error = HAMMER2_ERROR_EIO;
1233 kprintf("hammer2_chain_lock: I/O error %016jx: %d\n",
1234 (intmax_t)bref->data_off, error);
1235 hammer2_io_bqrelse(&chain->dio);
1241 * This isn't perfect and can be ignored on OSs which do not have
1242 * an indication as to whether a buffer is coming from cache or
1243 * if I/O was actually issued for the read. TESTEDGOOD will work
1244 * pretty well without the B_IOISSUED logic because chains are
1245 * cached, but in that situation (without B_IOISSUED) it will not
1246 * detect whether a re-read via I/O is corrupted verses the original
1249 * We can't re-run the CRC on every fresh lock. That would be
1250 * insanely expensive.
1252 * If the underlying kernel buffer covers the entire chain we can
1253 * use the B_IOISSUED indication to determine if we have to re-run
1254 * the CRC on chain data for chains that managed to stay cached
1255 * across the kernel disposal of the original buffer.
1257 if ((dio = chain->dio) != NULL && dio->bp) {
1258 struct buf *bp = dio->bp;
1260 if (dio->psize == chain->bytes &&
1261 (bp->b_flags & B_IOISSUED)) {
1262 atomic_clear_int(&chain->flags,
1263 HAMMER2_CHAIN_TESTEDGOOD);
1264 bp->b_flags &= ~B_IOISSUED;
1269 * NOTE: A locked chain's data cannot be modified without first
1270 * calling hammer2_chain_modify().
1274 * NOTE: hammer2_io_data() call issues bkvasync()
1276 bdata = hammer2_io_data(chain->dio, chain->bref.data_off);
1278 if (chain->flags & HAMMER2_CHAIN_INITIAL) {
1280 * Clear INITIAL. In this case we used io_new() and the
1281 * buffer has been zero'd and marked dirty.
1283 * CHAIN_MODIFIED has not been set yet, and we leave it
1284 * that way for now. Set a temporary CHAIN_NOTTESTED flag
1285 * to prevent hammer2_chain_testcheck() from trying to match
1286 * a check code that has not yet been generated. This bit
1287 * should NOT end up on the actual media.
1289 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1290 atomic_set_int(&chain->flags, HAMMER2_CHAIN_NOTTESTED);
1291 } else if (chain->flags & HAMMER2_CHAIN_MODIFIED) {
1293 * check data not currently synchronized due to
1294 * modification. XXX assumes data stays in the buffer
1295 * cache, which might not be true (need biodep on flush
1296 * to calculate crc? or simple crc?).
1298 } else if ((chain->flags & HAMMER2_CHAIN_TESTEDGOOD) == 0) {
1299 if (hammer2_chain_testcheck(chain, bdata) == 0) {
1300 chain->error = HAMMER2_ERROR_CHECK;
1302 atomic_set_int(&chain->flags, HAMMER2_CHAIN_TESTEDGOOD);
1307 * Setup the data pointer, either pointing it to an embedded data
1308 * structure and copying the data from the buffer, or pointing it
1311 * The buffer is not retained when copying to an embedded data
1312 * structure in order to avoid potential deadlocks or recursions
1313 * on the same physical buffer.
1315 * WARNING! Other threads can start using the data the instant we
1316 * set chain->data non-NULL.
1318 switch (bref->type) {
1319 case HAMMER2_BREF_TYPE_VOLUME:
1320 case HAMMER2_BREF_TYPE_FREEMAP:
1322 * Copy data from bp to embedded buffer
1324 panic("hammer2_chain_load_data: unresolved volume header");
1326 case HAMMER2_BREF_TYPE_DIRENT:
1327 KKASSERT(chain->bytes != 0);
1329 case HAMMER2_BREF_TYPE_INODE:
1330 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
1331 case HAMMER2_BREF_TYPE_INDIRECT:
1332 case HAMMER2_BREF_TYPE_DATA:
1333 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1336 * Point data at the device buffer and leave dio intact.
1338 chain->data = (void *)bdata;
1343 * Release HAMMER2_CHAIN_IOINPROG and signal waiters if requested.
1350 oflags = chain->flags;
1351 nflags = oflags & ~(HAMMER2_CHAIN_IOINPROG |
1352 HAMMER2_CHAIN_IOSIGNAL);
1353 KKASSERT(oflags & HAMMER2_CHAIN_IOINPROG);
1354 if (atomic_cmpset_int(&chain->flags, oflags, nflags)) {
1355 if (oflags & HAMMER2_CHAIN_IOSIGNAL)
1356 wakeup(&chain->flags);
1363 * Unlock and deref a chain element.
1365 * Remember that the presence of children under chain prevent the chain's
1366 * destruction but do not add additional references, so the dio will still
1370 hammer2_chain_unlock(hammer2_chain_t *chain)
1377 * If multiple locks are present (or being attempted) on this
1378 * particular chain we can just unlock, drop refs, and return.
1380 * Otherwise fall-through on the 1->0 transition.
1383 lockcnt = chain->lockcnt;
1384 KKASSERT(lockcnt > 0);
1387 if (atomic_cmpset_int(&chain->lockcnt,
1388 lockcnt, lockcnt - 1)) {
1389 hammer2_mtx_unlock(&chain->lock);
1392 } else if (hammer2_mtx_upgrade_try(&chain->lock) == 0) {
1393 /* while holding the mutex exclusively */
1394 if (atomic_cmpset_int(&chain->lockcnt, 1, 0))
1398 * This situation can easily occur on SMP due to
1399 * the gap inbetween the 1->0 transition and the
1400 * final unlock. We cannot safely block on the
1401 * mutex because lockcnt might go above 1.
1403 * XXX Sleep for one tick if it takes too long.
1405 if (++iter > 1000) {
1406 if (iter > 1000 + hz) {
1407 kprintf("hammer2: h2race2 %p\n", chain);
1410 tsleep(&iter, 0, "h2race2", 1);
1418 * Last unlock / mutex upgraded to exclusive. Drop the data
1421 dio = hammer2_chain_drop_data(chain);
1423 hammer2_io_bqrelse(&dio);
1424 hammer2_mtx_unlock(&chain->lock);
1428 * Unlock and hold chain data intact
1431 hammer2_chain_unlock_hold(hammer2_chain_t *chain)
1433 atomic_add_int(&chain->lockcnt, 1);
1434 hammer2_chain_unlock(chain);
1438 * Helper to obtain the blockref[] array base and count for a chain.
1440 * XXX Not widely used yet, various use cases need to be validated and
1441 * converted to use this function.
1444 hammer2_blockref_t *
1445 hammer2_chain_base_and_count(hammer2_chain_t *parent, int *countp)
1447 hammer2_blockref_t *base;
1450 if (parent->flags & HAMMER2_CHAIN_INITIAL) {
1453 switch(parent->bref.type) {
1454 case HAMMER2_BREF_TYPE_INODE:
1455 count = HAMMER2_SET_COUNT;
1457 case HAMMER2_BREF_TYPE_INDIRECT:
1458 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1459 count = parent->bytes / sizeof(hammer2_blockref_t);
1461 case HAMMER2_BREF_TYPE_VOLUME:
1462 count = HAMMER2_SET_COUNT;
1464 case HAMMER2_BREF_TYPE_FREEMAP:
1465 count = HAMMER2_SET_COUNT;
1468 panic("hammer2_chain_base_and_count: "
1469 "unrecognized blockref type: %d",
1475 switch(parent->bref.type) {
1476 case HAMMER2_BREF_TYPE_INODE:
1477 base = &parent->data->ipdata.u.blockset.blockref[0];
1478 count = HAMMER2_SET_COUNT;
1480 case HAMMER2_BREF_TYPE_INDIRECT:
1481 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1482 base = &parent->data->npdata[0];
1483 count = parent->bytes / sizeof(hammer2_blockref_t);
1485 case HAMMER2_BREF_TYPE_VOLUME:
1486 base = &parent->data->voldata.
1487 sroot_blockset.blockref[0];
1488 count = HAMMER2_SET_COUNT;
1490 case HAMMER2_BREF_TYPE_FREEMAP:
1491 base = &parent->data->blkset.blockref[0];
1492 count = HAMMER2_SET_COUNT;
1495 panic("hammer2_chain_base_and_count: "
1496 "unrecognized blockref type: %d",
1508 * This counts the number of live blockrefs in a block array and
1509 * also calculates the point at which all remaining blockrefs are empty.
1510 * This routine can only be called on a live chain.
1512 * Caller holds the chain locked, but possibly with a shared lock. We
1513 * must use an exclusive spinlock to prevent corruption.
1515 * NOTE: Flag is not set until after the count is complete, allowing
1516 * callers to test the flag without holding the spinlock.
1518 * NOTE: If base is NULL the related chain is still in the INITIAL
1519 * state and there are no blockrefs to count.
1521 * NOTE: live_count may already have some counts accumulated due to
1522 * creation and deletion and could even be initially negative.
1525 hammer2_chain_countbrefs(hammer2_chain_t *chain,
1526 hammer2_blockref_t *base, int count)
1528 hammer2_spin_ex(&chain->core.spin);
1529 if ((chain->flags & HAMMER2_CHAIN_COUNTEDBREFS) == 0) {
1531 while (--count >= 0) {
1532 if (base[count].type != HAMMER2_BREF_TYPE_EMPTY)
1535 chain->core.live_zero = count + 1;
1536 while (count >= 0) {
1537 if (base[count].type != HAMMER2_BREF_TYPE_EMPTY)
1538 atomic_add_int(&chain->core.live_count,
1543 chain->core.live_zero = 0;
1545 /* else do not modify live_count */
1546 atomic_set_int(&chain->flags, HAMMER2_CHAIN_COUNTEDBREFS);
1548 hammer2_spin_unex(&chain->core.spin);
1552 * Resize the chain's physical storage allocation in-place. This function does
1553 * not usually adjust the data pointer and must be followed by (typically) a
1554 * hammer2_chain_modify() call to copy any old data over and adjust the
1557 * Chains can be resized smaller without reallocating the storage. Resizing
1558 * larger will reallocate the storage. Excess or prior storage is reclaimed
1559 * asynchronously at a later time.
1561 * An nradix value of 0 is special-cased to mean that the storage should
1562 * be disassociated, that is the chain is being resized to 0 bytes (not 1
1565 * Must be passed an exclusively locked parent and chain.
1567 * This function is mostly used with DATA blocks locked RESOLVE_NEVER in order
1568 * to avoid instantiating a device buffer that conflicts with the vnode data
1569 * buffer. However, because H2 can compress or encrypt data, the chain may
1570 * have a dio assigned to it in those situations, and they do not conflict.
1572 * XXX return error if cannot resize.
1575 hammer2_chain_resize(hammer2_chain_t *chain,
1576 hammer2_tid_t mtid, hammer2_off_t dedup_off,
1577 int nradix, int flags)
1587 * Only data and indirect blocks can be resized for now.
1588 * (The volu root, inodes, and freemap elements use a fixed size).
1590 KKASSERT(chain != &hmp->vchain);
1591 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_DATA ||
1592 chain->bref.type == HAMMER2_BREF_TYPE_INDIRECT ||
1593 chain->bref.type == HAMMER2_BREF_TYPE_DIRENT);
1596 * Nothing to do if the element is already the proper size
1598 obytes = chain->bytes;
1599 nbytes = (nradix) ? (1U << nradix) : 0;
1600 if (obytes == nbytes)
1601 return (chain->error);
1604 * Make sure the old data is instantiated so we can copy it. If this
1605 * is a data block, the device data may be superfluous since the data
1606 * might be in a logical block, but compressed or encrypted data is
1609 * NOTE: The modify will set BMAPUPD for us if BMAPPED is set.
1611 error = hammer2_chain_modify(chain, mtid, dedup_off, 0);
1616 * Reallocate the block, even if making it smaller (because different
1617 * block sizes may be in different regions).
1619 * NOTE: Operation does not copy the data and may only be used
1620 * to resize data blocks in-place, or directory entry blocks
1621 * which are about to be modified in some manner.
1623 error = hammer2_freemap_alloc(chain, nbytes);
1627 chain->bytes = nbytes;
1630 * We don't want the followup chain_modify() to try to copy data
1631 * from the old (wrong-sized) buffer. It won't know how much to
1632 * copy. This case should only occur during writes when the
1633 * originator already has the data to write in-hand.
1636 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_DATA ||
1637 chain->bref.type == HAMMER2_BREF_TYPE_DIRENT);
1638 hammer2_io_brelse(&chain->dio);
1641 return (chain->error);
1645 * Set the chain modified so its data can be changed by the caller, or
1646 * install deduplicated data. The caller must call this routine for each
1647 * set of modifications it makes, even if the chain is already flagged
1650 * Sets bref.modify_tid to mtid only if mtid != 0. Note that bref.modify_tid
1651 * is a CLC (cluster level change) field and is not updated by parent
1652 * propagation during a flush.
1654 * Returns an appropriate HAMMER2_ERROR_* code, which will generally reflect
1655 * chain->error except for HAMMER2_ERROR_ENOSPC. If the allocation fails
1656 * due to no space available, HAMMER2_ERROR_ENOSPC is returned and the chain
1657 * remains unmodified with its old data ref intact and chain->error
1662 * If the DEDUPABLE flag is set in the chain the storage must be reallocated
1663 * even if the chain is still flagged MODIFIED. In this case the chain's
1664 * DEDUPABLE flag will be cleared once the new storage has been assigned.
1666 * If the caller passes a non-zero dedup_off we will use it to assign the
1667 * new storage. The MODIFIED flag will be *CLEARED* in this case, and
1668 * DEDUPABLE will be set (NOTE: the UPDATE flag is always set). The caller
1669 * must not modify the data content upon return.
1672 hammer2_chain_modify(hammer2_chain_t *chain, hammer2_tid_t mtid,
1673 hammer2_off_t dedup_off, int flags)
1675 hammer2_blockref_t obref;
1686 obref = chain->bref;
1687 KKASSERT((chain->flags & HAMMER2_CHAIN_FICTITIOUS) == 0);
1688 KKASSERT(chain->lock.mtx_lock & MTX_EXCLUSIVE);
1691 * Data is not optional for freemap chains (we must always be sure
1692 * to copy the data on COW storage allocations).
1694 if (chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE ||
1695 chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) {
1696 KKASSERT((chain->flags & HAMMER2_CHAIN_INITIAL) ||
1697 (flags & HAMMER2_MODIFY_OPTDATA) == 0);
1701 * Data must be resolved if already assigned, unless explicitly
1702 * flagged otherwise. If we cannot safety load the data the
1703 * modification fails and we return early.
1705 if (chain->data == NULL && chain->bytes != 0 &&
1706 (flags & HAMMER2_MODIFY_OPTDATA) == 0 &&
1707 (chain->bref.data_off & ~HAMMER2_OFF_MASK_RADIX)) {
1708 hammer2_chain_load_data(chain);
1710 return (chain->error);
1715 * Set MODIFIED to indicate that the chain has been modified. A new
1716 * allocation is required when modifying a chain.
1718 * Set UPDATE to ensure that the blockref is updated in the parent.
1720 * If MODIFIED is already set determine if we can reuse the assigned
1721 * data block or if we need a new data block.
1723 if ((chain->flags & HAMMER2_CHAIN_MODIFIED) == 0) {
1725 * Must set modified bit.
1727 atomic_add_long(&hammer2_count_modified_chains, 1);
1728 atomic_set_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
1729 hammer2_pfs_memory_inc(chain->pmp); /* can be NULL */
1733 * We may be able to avoid a copy-on-write if the chain's
1734 * check mode is set to NONE and the chain's current
1735 * modify_tid is beyond the last explicit snapshot tid.
1737 * This implements HAMMER2's overwrite-in-place feature.
1739 * NOTE! This data-block cannot be used as a de-duplication
1740 * source when the check mode is set to NONE.
1742 if ((chain->bref.type == HAMMER2_BREF_TYPE_DATA ||
1743 chain->bref.type == HAMMER2_BREF_TYPE_DIRENT) &&
1744 (chain->flags & HAMMER2_CHAIN_INITIAL) == 0 &&
1745 (chain->flags & HAMMER2_CHAIN_DEDUPABLE) == 0 &&
1746 HAMMER2_DEC_CHECK(chain->bref.methods) ==
1747 HAMMER2_CHECK_NONE &&
1749 chain->bref.modify_tid >
1750 chain->pmp->iroot->meta.pfs_lsnap_tid) {
1752 * Sector overwrite allowed.
1755 } else if ((hmp->hflags & HMNT2_EMERG) &&
1757 chain->bref.modify_tid >
1758 chain->pmp->iroot->meta.pfs_lsnap_tid) {
1760 * If in emergency delete mode then do a modify-in-
1761 * place on any chain type belonging to the PFS as
1762 * long as it doesn't mess up a snapshot. We might
1763 * be forced to do this anyway a little further down
1764 * in the code if the allocation fails.
1766 * Also note that in emergency mode, these modify-in-
1767 * place operations are NOT SAFE. A storage failure,
1768 * power failure, or panic can corrupt the filesystem.
1773 * Sector overwrite not allowed, must copy-on-write.
1777 } else if (chain->flags & HAMMER2_CHAIN_DEDUPABLE) {
1779 * If the modified chain was registered for dedup we need
1780 * a new allocation. This only happens for delayed-flush
1781 * chains (i.e. which run through the front-end buffer
1788 * Already flagged modified, no new allocation is needed.
1795 * Flag parent update required.
1797 if ((chain->flags & HAMMER2_CHAIN_UPDATE) == 0) {
1798 atomic_set_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
1805 * The XOP code returns held but unlocked focus chains. This
1806 * prevents the chain from being destroyed but does not prevent
1807 * it from being modified. diolk is used to interlock modifications
1808 * against XOP frontend accesses to the focus.
1810 * This allows us to theoretically avoid deadlocking the frontend
1811 * if one of the backends lock up by not formally locking the
1812 * focused chain in the frontend. In addition, the synchronization
1813 * code relies on this mechanism to avoid deadlocking concurrent
1814 * synchronization threads.
1816 lockmgr(&chain->diolk, LK_EXCLUSIVE);
1819 * The modification or re-modification requires an allocation and
1820 * possible COW. If an error occurs, the previous content and data
1821 * reference is retained and the modification fails.
1823 * If dedup_off is non-zero, the caller is requesting a deduplication
1824 * rather than a modification. The MODIFIED bit is not set and the
1825 * data offset is set to the deduplication offset. The data cannot
1828 * NOTE: The dedup offset is allowed to be in a partially free state
1829 * and we must be sure to reset it to a fully allocated state
1830 * to force two bulkfree passes to free it again.
1832 * NOTE: Only applicable when chain->bytes != 0.
1834 * XXX can a chain already be marked MODIFIED without a data
1835 * assignment? If not, assert here instead of testing the case.
1837 if (chain != &hmp->vchain && chain != &hmp->fchain &&
1839 if ((chain->bref.data_off & ~HAMMER2_OFF_MASK_RADIX) == 0 ||
1843 * NOTE: We do not have to remove the dedup
1844 * registration because the area is still
1845 * allocated and the underlying DIO will
1849 chain->bref.data_off = dedup_off;
1850 chain->bytes = 1 << (dedup_off &
1851 HAMMER2_OFF_MASK_RADIX);
1853 atomic_clear_int(&chain->flags,
1854 HAMMER2_CHAIN_MODIFIED);
1855 atomic_add_long(&hammer2_count_modified_chains,
1858 hammer2_pfs_memory_wakeup(
1861 hammer2_freemap_adjust(hmp, &chain->bref,
1862 HAMMER2_FREEMAP_DORECOVER);
1863 atomic_set_int(&chain->flags,
1864 HAMMER2_CHAIN_DEDUPABLE);
1866 error = hammer2_freemap_alloc(chain,
1868 atomic_clear_int(&chain->flags,
1869 HAMMER2_CHAIN_DEDUPABLE);
1872 * If we are unable to allocate a new block
1873 * but we are in emergency mode, issue a
1874 * warning to the console and reuse the same
1877 * We behave as if the allocation were
1880 * THIS IS IMPORTANT: These modifications
1881 * are virtually guaranteed to corrupt any
1882 * snapshots related to this filesystem.
1884 if (error && (hmp->hflags & HMNT2_EMERG)) {
1886 chain->bref.flags |=
1887 HAMMER2_BREF_FLAG_EMERG_MIP;
1889 krateprintf(&krate_h2em,
1890 "hammer2: Emergency Mode WARNING: "
1891 "Operation will likely corrupt "
1892 "related snapshot: "
1893 "%016jx.%02x key=%016jx\n",
1894 chain->bref.data_off,
1897 } else if (error == 0) {
1898 chain->bref.flags &=
1899 ~HAMMER2_BREF_FLAG_EMERG_MIP;
1906 * Stop here if error. We have to undo any flag bits we might
1911 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
1912 atomic_add_long(&hammer2_count_modified_chains, -1);
1914 hammer2_pfs_memory_wakeup(chain->pmp, -1);
1917 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
1919 lockmgr(&chain->diolk, LK_RELEASE);
1925 * Update mirror_tid and modify_tid. modify_tid is only updated
1926 * if not passed as zero (during flushes, parent propagation passes
1929 * NOTE: chain->pmp could be the device spmp.
1931 chain->bref.mirror_tid = hmp->voldata.mirror_tid + 1;
1933 chain->bref.modify_tid = mtid;
1936 * Set BMAPUPD to tell the flush code that an existing blockmap entry
1937 * requires updating as well as to tell the delete code that the
1938 * chain's blockref might not exactly match (in terms of physical size
1939 * or block offset) the one in the parent's blocktable. The base key
1940 * of course will still match.
1942 if (chain->flags & HAMMER2_CHAIN_BMAPPED)
1943 atomic_set_int(&chain->flags, HAMMER2_CHAIN_BMAPUPD);
1946 * Short-cut data block handling when the caller does not need an
1947 * actual data reference to (aka OPTDATA), as long as the chain does
1948 * not already have a data pointer to the data and no de-duplication
1951 * This generally means that the modifications are being done via the
1952 * logical buffer cache.
1954 * NOTE: If deduplication occurred we have to run through the data
1955 * stuff to clear INITIAL, and the caller will likely want to
1956 * assign the check code anyway. Leaving INITIAL set on a
1957 * dedup can be deadly (it can cause the block to be zero'd!).
1959 * This code also handles bytes == 0 (most dirents).
1961 if (chain->bref.type == HAMMER2_BREF_TYPE_DATA &&
1962 (flags & HAMMER2_MODIFY_OPTDATA) &&
1963 chain->data == NULL) {
1964 if (dedup_off == 0) {
1965 KKASSERT(chain->dio == NULL);
1971 * Clearing the INITIAL flag (for indirect blocks) indicates that
1972 * we've processed the uninitialized storage allocation.
1974 * If this flag is already clear we are likely in a copy-on-write
1975 * situation but we have to be sure NOT to bzero the storage if
1976 * no data is present.
1978 * Clearing of NOTTESTED is allowed if the MODIFIED bit is set,
1980 if (chain->flags & HAMMER2_CHAIN_INITIAL) {
1981 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1988 * Instantiate data buffer and possibly execute COW operation
1990 switch(chain->bref.type) {
1991 case HAMMER2_BREF_TYPE_VOLUME:
1992 case HAMMER2_BREF_TYPE_FREEMAP:
1994 * The data is embedded, no copy-on-write operation is
1997 KKASSERT(chain->dio == NULL);
1999 case HAMMER2_BREF_TYPE_DIRENT:
2001 * The data might be fully embedded.
2003 if (chain->bytes == 0) {
2004 KKASSERT(chain->dio == NULL);
2008 case HAMMER2_BREF_TYPE_INODE:
2009 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2010 case HAMMER2_BREF_TYPE_DATA:
2011 case HAMMER2_BREF_TYPE_INDIRECT:
2012 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2014 * Perform the copy-on-write operation
2016 * zero-fill or copy-on-write depending on whether
2017 * chain->data exists or not and set the dirty state for
2018 * the new buffer. hammer2_io_new() will handle the
2021 * If a dedup_off was supplied this is an existing block
2022 * and no COW, copy, or further modification is required.
2024 KKASSERT(chain != &hmp->vchain && chain != &hmp->fchain);
2026 if (wasinitial && dedup_off == 0) {
2027 error = hammer2_io_new(hmp, chain->bref.type,
2028 chain->bref.data_off,
2029 chain->bytes, &dio);
2031 error = hammer2_io_bread(hmp, chain->bref.type,
2032 chain->bref.data_off,
2033 chain->bytes, &dio);
2035 hammer2_adjreadcounter(&chain->bref, chain->bytes);
2038 * If an I/O error occurs make sure callers cannot accidently
2039 * modify the old buffer's contents and corrupt the filesystem.
2041 * NOTE: hammer2_io_data() call issues bkvasync()
2044 kprintf("hammer2_chain_modify: hmp=%p I/O error\n",
2046 chain->error = HAMMER2_ERROR_EIO;
2047 hammer2_io_brelse(&dio);
2048 hammer2_io_brelse(&chain->dio);
2053 bdata = hammer2_io_data(dio, chain->bref.data_off);
2057 * COW (unless a dedup).
2059 KKASSERT(chain->dio != NULL);
2060 if (chain->data != (void *)bdata && dedup_off == 0) {
2061 bcopy(chain->data, bdata, chain->bytes);
2063 } else if (wasinitial == 0 && dedup_off == 0) {
2065 * We have a problem. We were asked to COW but
2066 * we don't have any data to COW with!
2068 panic("hammer2_chain_modify: having a COW %p\n",
2073 * Retire the old buffer, replace with the new. Dirty or
2074 * redirty the new buffer.
2076 * WARNING! The system buffer cache may have already flushed
2077 * the buffer, so we must be sure to [re]dirty it
2078 * for further modification.
2080 * If dedup_off was supplied, the caller is not
2081 * expected to make any further modification to the
2084 * WARNING! hammer2_get_gdata() assumes dio never transitions
2085 * through NULL in order to optimize away unnecessary
2091 if ((tio = chain->dio) != NULL)
2092 hammer2_io_bqrelse(&tio);
2093 chain->data = (void *)bdata;
2096 hammer2_io_setdirty(dio);
2100 panic("hammer2_chain_modify: illegal non-embedded type %d",
2107 * setflush on parent indicating that the parent must recurse down
2108 * to us. Do not call on chain itself which might already have it
2112 hammer2_chain_setflush(chain->parent);
2113 lockmgr(&chain->diolk, LK_RELEASE);
2115 return (chain->error);
2119 * Modify the chain associated with an inode.
2122 hammer2_chain_modify_ip(hammer2_inode_t *ip, hammer2_chain_t *chain,
2123 hammer2_tid_t mtid, int flags)
2127 hammer2_inode_modify(ip);
2128 error = hammer2_chain_modify(chain, mtid, 0, flags);
2134 * Volume header data locks
2137 hammer2_voldata_lock(hammer2_dev_t *hmp)
2139 lockmgr(&hmp->vollk, LK_EXCLUSIVE);
2143 hammer2_voldata_unlock(hammer2_dev_t *hmp)
2145 lockmgr(&hmp->vollk, LK_RELEASE);
2149 hammer2_voldata_modify(hammer2_dev_t *hmp)
2151 if ((hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) == 0) {
2152 atomic_add_long(&hammer2_count_modified_chains, 1);
2153 atomic_set_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
2154 hammer2_pfs_memory_inc(hmp->vchain.pmp);
2159 * This function returns the chain at the nearest key within the specified
2160 * range. The returned chain will be referenced but not locked.
2162 * This function will recurse through chain->rbtree as necessary and will
2163 * return a *key_nextp suitable for iteration. *key_nextp is only set if
2164 * the iteration value is less than the current value of *key_nextp.
2166 * The caller should use (*key_nextp) to calculate the actual range of
2167 * the returned element, which will be (key_beg to *key_nextp - 1), because
2168 * there might be another element which is superior to the returned element
2171 * (*key_nextp) can be passed as key_beg in an iteration only while non-NULL
2172 * chains continue to be returned. On EOF (*key_nextp) may overflow since
2173 * it will wind up being (key_end + 1).
2175 * WARNING! Must be called with child's spinlock held. Spinlock remains
2176 * held through the operation.
2178 struct hammer2_chain_find_info {
2179 hammer2_chain_t *best;
2180 hammer2_key_t key_beg;
2181 hammer2_key_t key_end;
2182 hammer2_key_t key_next;
2185 static int hammer2_chain_find_cmp(hammer2_chain_t *child, void *data);
2186 static int hammer2_chain_find_callback(hammer2_chain_t *child, void *data);
2190 hammer2_chain_find(hammer2_chain_t *parent, hammer2_key_t *key_nextp,
2191 hammer2_key_t key_beg, hammer2_key_t key_end)
2193 struct hammer2_chain_find_info info;
2196 info.key_beg = key_beg;
2197 info.key_end = key_end;
2198 info.key_next = *key_nextp;
2200 RB_SCAN(hammer2_chain_tree, &parent->core.rbtree,
2201 hammer2_chain_find_cmp, hammer2_chain_find_callback,
2203 *key_nextp = info.key_next;
2205 kprintf("chain_find %p %016jx:%016jx next=%016jx\n",
2206 parent, key_beg, key_end, *key_nextp);
2214 hammer2_chain_find_cmp(hammer2_chain_t *child, void *data)
2216 struct hammer2_chain_find_info *info = data;
2217 hammer2_key_t child_beg;
2218 hammer2_key_t child_end;
2220 child_beg = child->bref.key;
2221 child_end = child_beg + ((hammer2_key_t)1 << child->bref.keybits) - 1;
2223 if (child_end < info->key_beg)
2225 if (child_beg > info->key_end)
2232 hammer2_chain_find_callback(hammer2_chain_t *child, void *data)
2234 struct hammer2_chain_find_info *info = data;
2235 hammer2_chain_t *best;
2236 hammer2_key_t child_end;
2239 * WARNING! Layerq is scanned forwards, exact matches should keep
2240 * the existing info->best.
2242 if ((best = info->best) == NULL) {
2244 * No previous best. Assign best
2247 } else if (best->bref.key <= info->key_beg &&
2248 child->bref.key <= info->key_beg) {
2253 /*info->best = child;*/
2254 } else if (child->bref.key < best->bref.key) {
2256 * Child has a nearer key and best is not flush with key_beg.
2257 * Set best to child. Truncate key_next to the old best key.
2260 if (info->key_next > best->bref.key || info->key_next == 0)
2261 info->key_next = best->bref.key;
2262 } else if (child->bref.key == best->bref.key) {
2264 * If our current best is flush with the child then this
2265 * is an illegal overlap.
2267 * key_next will automatically be limited to the smaller of
2268 * the two end-points.
2274 * Keep the current best but truncate key_next to the child's
2277 * key_next will also automatically be limited to the smaller
2278 * of the two end-points (probably not necessary for this case
2279 * but we do it anyway).
2281 if (info->key_next > child->bref.key || info->key_next == 0)
2282 info->key_next = child->bref.key;
2286 * Always truncate key_next based on child's end-of-range.
2288 child_end = child->bref.key + ((hammer2_key_t)1 << child->bref.keybits);
2289 if (child_end && (info->key_next > child_end || info->key_next == 0))
2290 info->key_next = child_end;
2296 * Retrieve the specified chain from a media blockref, creating the
2297 * in-memory chain structure which reflects it. The returned chain is
2298 * held and locked according to (how) (HAMMER2_RESOLVE_*). The caller must
2299 * handle crc-checks and so forth, and should check chain->error before
2300 * assuming that the data is good.
2302 * To handle insertion races pass the INSERT_RACE flag along with the
2303 * generation number of the core. NULL will be returned if the generation
2304 * number changes before we have a chance to insert the chain. Insert
2305 * races can occur because the parent might be held shared.
2307 * Caller must hold the parent locked shared or exclusive since we may
2308 * need the parent's bref array to find our block.
2310 * WARNING! chain->pmp is always set to NULL for any chain representing
2311 * part of the super-root topology.
2314 hammer2_chain_get(hammer2_chain_t *parent, int generation,
2315 hammer2_blockref_t *bref, int how)
2317 hammer2_dev_t *hmp = parent->hmp;
2318 hammer2_chain_t *chain;
2322 * Allocate a chain structure representing the existing media
2323 * entry. Resulting chain has one ref and is not locked.
2325 if (bref->flags & HAMMER2_BREF_FLAG_PFSROOT)
2326 chain = hammer2_chain_alloc(hmp, NULL, bref);
2328 chain = hammer2_chain_alloc(hmp, parent->pmp, bref);
2329 /* ref'd chain returned */
2332 * Flag that the chain is in the parent's blockmap so delete/flush
2333 * knows what to do with it.
2335 atomic_set_int(&chain->flags, HAMMER2_CHAIN_BMAPPED);
2338 * chain must be locked to avoid unexpected ripouts
2340 hammer2_chain_lock(chain, how);
2343 * Link the chain into its parent. A spinlock is required to safely
2344 * access the RBTREE, and it is possible to collide with another
2345 * hammer2_chain_get() operation because the caller might only hold
2346 * a shared lock on the parent.
2348 * NOTE: Get races can occur quite often when we distribute
2349 * asynchronous read-aheads across multiple threads.
2351 KKASSERT(parent->refs > 0);
2352 error = hammer2_chain_insert(parent, chain,
2353 HAMMER2_CHAIN_INSERT_SPIN |
2354 HAMMER2_CHAIN_INSERT_RACE,
2357 KKASSERT((chain->flags & HAMMER2_CHAIN_ONRBTREE) == 0);
2358 /*kprintf("chain %p get race\n", chain);*/
2359 hammer2_chain_unlock(chain);
2360 hammer2_chain_drop(chain);
2363 KKASSERT(chain->flags & HAMMER2_CHAIN_ONRBTREE);
2367 * Return our new chain referenced but not locked, or NULL if
2374 * Lookup initialization/completion API
2377 hammer2_chain_lookup_init(hammer2_chain_t *parent, int flags)
2379 hammer2_chain_ref(parent);
2380 if (flags & HAMMER2_LOOKUP_SHARED) {
2381 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS |
2382 HAMMER2_RESOLVE_SHARED);
2384 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2390 hammer2_chain_lookup_done(hammer2_chain_t *parent)
2393 hammer2_chain_unlock(parent);
2394 hammer2_chain_drop(parent);
2399 * Take the locked chain and return a locked parent. The chain remains
2400 * locked on return, but may have to be temporarily unlocked to acquire
2401 * the parent. Because of this, (chain) must be stable and cannot be
2402 * deleted while it was temporarily unlocked (typically means that (chain)
2405 * Pass HAMMER2_RESOLVE_* flags in flags.
2407 * This will work even if the chain is errored, and the caller can check
2408 * parent->error on return if desired since the parent will be locked.
2410 * This function handles the lock order reversal.
2413 hammer2_chain_getparent(hammer2_chain_t *chain, int flags)
2415 hammer2_chain_t *parent;
2418 * Be careful of order, chain must be unlocked before parent
2419 * is locked below to avoid a deadlock. Try it trivially first.
2421 parent = chain->parent;
2423 panic("hammer2_chain_getparent: no parent");
2424 hammer2_chain_ref(parent);
2425 if (hammer2_chain_lock(parent, flags|HAMMER2_RESOLVE_NONBLOCK) == 0)
2429 hammer2_chain_unlock(chain);
2430 hammer2_chain_lock(parent, flags);
2431 hammer2_chain_lock(chain, flags);
2434 * Parent relinking races are quite common. We have to get
2435 * it right or we will blow up the block table.
2437 if (chain->parent == parent)
2439 hammer2_chain_unlock(parent);
2440 hammer2_chain_drop(parent);
2442 parent = chain->parent;
2444 panic("hammer2_chain_getparent: no parent");
2445 hammer2_chain_ref(parent);
2451 * Take the locked chain and return a locked parent. The chain is unlocked
2452 * and dropped. *chainp is set to the returned parent as a convenience.
2453 * Pass HAMMER2_RESOLVE_* flags in flags.
2455 * This will work even if the chain is errored, and the caller can check
2456 * parent->error on return if desired since the parent will be locked.
2458 * The chain does NOT need to be stable. We use a tracking structure
2459 * to track the expected parent if the chain is deleted out from under us.
2461 * This function handles the lock order reversal.
2464 hammer2_chain_repparent(hammer2_chain_t **chainp, int flags)
2466 hammer2_chain_t *chain;
2467 hammer2_chain_t *parent;
2468 struct hammer2_reptrack reptrack;
2469 struct hammer2_reptrack **repp;
2472 * Be careful of order, chain must be unlocked before parent
2473 * is locked below to avoid a deadlock. Try it trivially first.
2476 parent = chain->parent;
2477 if (parent == NULL) {
2478 hammer2_spin_unex(&chain->core.spin);
2479 panic("hammer2_chain_repparent: no parent");
2481 hammer2_chain_ref(parent);
2482 if (hammer2_chain_lock(parent, flags|HAMMER2_RESOLVE_NONBLOCK) == 0) {
2483 hammer2_chain_unlock(chain);
2484 hammer2_chain_drop(chain);
2491 * Ok, now it gets a bit nasty. There are multiple situations where
2492 * the parent might be in the middle of a deletion, or where the child
2493 * (chain) might be deleted the instant we let go of its lock.
2494 * We can potentially end up in a no-win situation!
2496 * In particular, the indirect_maintenance() case can cause these
2499 * To deal with this we install a reptrack structure in the parent
2500 * This reptrack structure 'owns' the parent ref and will automatically
2501 * migrate to the parent's parent if the parent is deleted permanently.
2503 hammer2_spin_init(&reptrack.spin, "h2reptrk");
2504 reptrack.chain = parent;
2505 hammer2_chain_ref(parent); /* for the reptrack */
2507 hammer2_spin_ex(&parent->core.spin);
2508 reptrack.next = parent->core.reptrack;
2509 parent->core.reptrack = &reptrack;
2510 hammer2_spin_unex(&parent->core.spin);
2512 hammer2_chain_unlock(chain);
2513 hammer2_chain_drop(chain);
2514 chain = NULL; /* gone */
2517 * At the top of this loop, chain is gone and parent is refd both
2518 * by us explicitly AND via our reptrack. We are attempting to
2522 hammer2_chain_lock(parent, flags);
2524 if (reptrack.chain == parent)
2526 hammer2_chain_unlock(parent);
2527 hammer2_chain_drop(parent);
2529 kprintf("hammer2: debug REPTRACK %p->%p\n",
2530 parent, reptrack.chain);
2531 hammer2_spin_ex(&reptrack.spin);
2532 parent = reptrack.chain;
2533 hammer2_chain_ref(parent);
2534 hammer2_spin_unex(&reptrack.spin);
2538 * Once parent is locked and matches our reptrack, our reptrack
2539 * will be stable and we have our parent. We can unlink our
2542 * WARNING! Remember that the chain lock might be shared. Chains
2543 * locked shared have stable parent linkages.
2545 hammer2_spin_ex(&parent->core.spin);
2546 repp = &parent->core.reptrack;
2547 while (*repp != &reptrack)
2548 repp = &(*repp)->next;
2549 *repp = reptrack.next;
2550 hammer2_spin_unex(&parent->core.spin);
2552 hammer2_chain_drop(parent); /* reptrack ref */
2553 *chainp = parent; /* return parent lock+ref */
2559 * Dispose of any linked reptrack structures in (chain) by shifting them to
2560 * (parent). Both (chain) and (parent) must be exclusively locked.
2562 * This is interlocked against any children of (chain) on the other side.
2563 * No children so remain as-of when this is called so we can test
2564 * core.reptrack without holding the spin-lock.
2566 * Used whenever the caller intends to permanently delete chains related
2567 * to topological recursions (BREF_TYPE_INDIRECT, BREF_TYPE_FREEMAP_NODE),
2568 * where the chains underneath the node being deleted are given a new parent
2569 * above the node being deleted.
2573 hammer2_chain_repchange(hammer2_chain_t *parent, hammer2_chain_t *chain)
2575 struct hammer2_reptrack *reptrack;
2577 KKASSERT(chain->core.live_count == 0 && RB_EMPTY(&chain->core.rbtree));
2578 while (chain->core.reptrack) {
2579 hammer2_spin_ex(&parent->core.spin);
2580 hammer2_spin_ex(&chain->core.spin);
2581 reptrack = chain->core.reptrack;
2582 if (reptrack == NULL) {
2583 hammer2_spin_unex(&chain->core.spin);
2584 hammer2_spin_unex(&parent->core.spin);
2587 hammer2_spin_ex(&reptrack->spin);
2588 chain->core.reptrack = reptrack->next;
2589 reptrack->chain = parent;
2590 reptrack->next = parent->core.reptrack;
2591 parent->core.reptrack = reptrack;
2592 hammer2_chain_ref(parent); /* reptrack */
2594 hammer2_spin_unex(&chain->core.spin);
2595 hammer2_spin_unex(&parent->core.spin);
2596 kprintf("hammer2: debug repchange %p %p->%p\n",
2597 reptrack, chain, parent);
2598 hammer2_chain_drop(chain); /* reptrack */
2603 * Locate the first chain whos key range overlaps (key_beg, key_end) inclusive.
2604 * (*parentp) typically points to an inode but can also point to a related
2605 * indirect block and this function will recurse upwards and find the inode
2606 * or the nearest undeleted indirect block covering the key range.
2608 * This function unconditionally sets *errorp, replacing any previous value.
2610 * (*parentp) must be exclusive or shared locked (depending on flags) and
2611 * referenced and can be an inode or an existing indirect block within the
2614 * If (*parent) is errored out, this function will not attempt to recurse
2615 * the radix tree and will return NULL along with an appropriate *errorp.
2616 * If NULL is returned and *errorp is 0, the requested lookup could not be
2619 * On return (*parentp) will be modified to point at the deepest parent chain
2620 * element encountered during the search, as a helper for an insertion or
2623 * The new (*parentp) will be locked shared or exclusive (depending on flags),
2624 * and referenced, and the old will be unlocked and dereferenced (no change
2625 * if they are both the same). This is particularly important if the caller
2626 * wishes to insert a new chain, (*parentp) will be set properly even if NULL
2627 * is returned, as long as no error occurred.
2629 * The matching chain will be returned locked according to flags.
2633 * NULL is returned if no match was found, but (*parentp) will still
2634 * potentially be adjusted.
2636 * On return (*key_nextp) will point to an iterative value for key_beg.
2637 * (If NULL is returned (*key_nextp) is set to (key_end + 1)).
2639 * This function will also recurse up the chain if the key is not within the
2640 * current parent's range. (*parentp) can never be set to NULL. An iteration
2641 * can simply allow (*parentp) to float inside the loop.
2643 * NOTE! chain->data is not always resolved. By default it will not be
2644 * resolved for BREF_TYPE_DATA, FREEMAP_NODE, or FREEMAP_LEAF. Use
2645 * HAMMER2_LOOKUP_ALWAYS to force resolution (but be careful w/
2646 * BREF_TYPE_DATA as the device buffer can alias the logical file
2651 hammer2_chain_lookup(hammer2_chain_t **parentp, hammer2_key_t *key_nextp,
2652 hammer2_key_t key_beg, hammer2_key_t key_end,
2653 int *errorp, int flags)
2656 hammer2_chain_t *parent;
2657 hammer2_chain_t *chain;
2658 hammer2_blockref_t *base;
2659 hammer2_blockref_t *bref;
2660 hammer2_blockref_t bsave;
2661 hammer2_key_t scan_beg;
2662 hammer2_key_t scan_end;
2664 int how_always = HAMMER2_RESOLVE_ALWAYS;
2665 int how_maybe = HAMMER2_RESOLVE_MAYBE;
2668 int maxloops = 300000;
2669 volatile hammer2_mtx_t save_mtx;
2671 if (flags & HAMMER2_LOOKUP_ALWAYS) {
2672 how_maybe = how_always;
2673 how = HAMMER2_RESOLVE_ALWAYS;
2674 } else if (flags & HAMMER2_LOOKUP_NODATA) {
2675 how = HAMMER2_RESOLVE_NEVER;
2677 how = HAMMER2_RESOLVE_MAYBE;
2679 if (flags & HAMMER2_LOOKUP_SHARED) {
2680 how_maybe |= HAMMER2_RESOLVE_SHARED;
2681 how_always |= HAMMER2_RESOLVE_SHARED;
2682 how |= HAMMER2_RESOLVE_SHARED;
2686 * Recurse (*parentp) upward if necessary until the parent completely
2687 * encloses the key range or we hit the inode.
2689 * Handle races against the flusher deleting indirect nodes on its
2690 * way back up by continuing to recurse upward past the deletion.
2696 while (parent->bref.type == HAMMER2_BREF_TYPE_INDIRECT ||
2697 parent->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2698 scan_beg = parent->bref.key;
2699 scan_end = scan_beg +
2700 ((hammer2_key_t)1 << parent->bref.keybits) - 1;
2701 if ((parent->flags & HAMMER2_CHAIN_DELETED) == 0) {
2702 if (key_beg >= scan_beg && key_end <= scan_end)
2705 parent = hammer2_chain_repparent(parentp, how_maybe);
2708 if (--maxloops == 0)
2709 panic("hammer2_chain_lookup: maxloops");
2712 * MATCHIND case that does not require parent->data (do prior to
2713 * parent->error check).
2715 switch(parent->bref.type) {
2716 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2717 case HAMMER2_BREF_TYPE_INDIRECT:
2718 if (flags & HAMMER2_LOOKUP_MATCHIND) {
2719 scan_beg = parent->bref.key;
2720 scan_end = scan_beg +
2721 ((hammer2_key_t)1 << parent->bref.keybits) - 1;
2722 if (key_beg == scan_beg && key_end == scan_end) {
2724 hammer2_chain_ref(chain);
2725 hammer2_chain_lock(chain, how_maybe);
2726 *key_nextp = scan_end + 1;
2736 * No lookup is possible if the parent is errored. We delayed
2737 * this check as long as we could to ensure that the parent backup,
2738 * embedded data, and MATCHIND code could still execute.
2740 if (parent->error) {
2741 *errorp = parent->error;
2746 * Locate the blockref array. Currently we do a fully associative
2747 * search through the array.
2749 switch(parent->bref.type) {
2750 case HAMMER2_BREF_TYPE_INODE:
2752 * Special shortcut for embedded data returns the inode
2753 * itself. Callers must detect this condition and access
2754 * the embedded data (the strategy code does this for us).
2756 * This is only applicable to regular files and softlinks.
2758 * We need a second lock on parent. Since we already have
2759 * a lock we must pass LOCKAGAIN to prevent unexpected
2760 * blocking (we don't want to block on a second shared
2761 * ref if an exclusive lock is pending)
2763 if (parent->data->ipdata.meta.op_flags &
2764 HAMMER2_OPFLAG_DIRECTDATA) {
2765 if (flags & HAMMER2_LOOKUP_NODIRECT) {
2767 *key_nextp = key_end + 1;
2770 hammer2_chain_ref(parent);
2771 hammer2_chain_lock(parent, how_always |
2772 HAMMER2_RESOLVE_LOCKAGAIN);
2773 *key_nextp = key_end + 1;
2776 base = &parent->data->ipdata.u.blockset.blockref[0];
2777 count = HAMMER2_SET_COUNT;
2779 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2780 case HAMMER2_BREF_TYPE_INDIRECT:
2782 * Optimize indirect blocks in the INITIAL state to avoid
2785 * Debugging: Enter permanent wait state instead of
2786 * panicing on unexpectedly NULL data for the moment.
2788 if (parent->flags & HAMMER2_CHAIN_INITIAL) {
2791 if (parent->data == NULL) {
2792 kprintf("hammer2: unexpected NULL data "
2795 tsleep(parent, 0, "xxx", 0);
2797 base = &parent->data->npdata[0];
2799 count = parent->bytes / sizeof(hammer2_blockref_t);
2801 case HAMMER2_BREF_TYPE_VOLUME:
2802 base = &parent->data->voldata.sroot_blockset.blockref[0];
2803 count = HAMMER2_SET_COUNT;
2805 case HAMMER2_BREF_TYPE_FREEMAP:
2806 base = &parent->data->blkset.blockref[0];
2807 count = HAMMER2_SET_COUNT;
2810 kprintf("hammer2_chain_lookup: unrecognized "
2811 "blockref(B) type: %d",
2814 tsleep(&base, 0, "dead", 0);
2815 panic("hammer2_chain_lookup: unrecognized "
2816 "blockref(B) type: %d",
2818 base = NULL; /* safety */
2819 count = 0; /* safety */
2823 * Merged scan to find next candidate.
2825 * hammer2_base_*() functions require the parent->core.live_* fields
2826 * to be synchronized.
2828 * We need to hold the spinlock to access the block array and RB tree
2829 * and to interlock chain creation.
2831 if ((parent->flags & HAMMER2_CHAIN_COUNTEDBREFS) == 0)
2832 hammer2_chain_countbrefs(parent, base, count);
2837 hammer2_spin_ex(&parent->core.spin);
2838 chain = hammer2_combined_find(parent, base, count,
2842 generation = parent->core.generation;
2845 * Exhausted parent chain, iterate.
2848 KKASSERT(chain == NULL);
2849 hammer2_spin_unex(&parent->core.spin);
2850 if (key_beg == key_end) /* short cut single-key case */
2854 * Stop if we reached the end of the iteration.
2856 if (parent->bref.type != HAMMER2_BREF_TYPE_INDIRECT &&
2857 parent->bref.type != HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2862 * Calculate next key, stop if we reached the end of the
2863 * iteration, otherwise go up one level and loop.
2865 key_beg = parent->bref.key +
2866 ((hammer2_key_t)1 << parent->bref.keybits);
2867 if (key_beg == 0 || key_beg > key_end)
2869 parent = hammer2_chain_repparent(parentp, how_maybe);
2874 * Selected from blockref or in-memory chain.
2877 if (chain == NULL) {
2878 hammer2_spin_unex(&parent->core.spin);
2879 if (bsave.type == HAMMER2_BREF_TYPE_INDIRECT ||
2880 bsave.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2881 chain = hammer2_chain_get(parent, generation,
2884 chain = hammer2_chain_get(parent, generation,
2890 hammer2_chain_ref(chain);
2891 hammer2_spin_unex(&parent->core.spin);
2894 * chain is referenced but not locked. We must lock the
2895 * chain to obtain definitive state.
2897 if (bsave.type == HAMMER2_BREF_TYPE_INDIRECT ||
2898 bsave.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2899 hammer2_chain_lock(chain, how_maybe);
2901 hammer2_chain_lock(chain, how);
2903 KKASSERT(chain->parent == parent);
2905 if (bcmp(&bsave, &chain->bref, sizeof(bsave)) ||
2906 chain->parent != parent) {
2907 hammer2_chain_unlock(chain);
2908 hammer2_chain_drop(chain);
2909 chain = NULL; /* SAFETY */
2915 * Skip deleted chains (XXX cache 'i' end-of-block-array? XXX)
2917 * NOTE: Chain's key range is not relevant as there might be
2918 * one-offs within the range that are not deleted.
2920 * NOTE: Lookups can race delete-duplicate because
2921 * delete-duplicate does not lock the parent's core
2922 * (they just use the spinlock on the core).
2924 if (chain->flags & HAMMER2_CHAIN_DELETED) {
2925 kprintf("skip deleted chain %016jx.%02x key=%016jx\n",
2926 chain->bref.data_off, chain->bref.type,
2928 hammer2_chain_unlock(chain);
2929 hammer2_chain_drop(chain);
2930 chain = NULL; /* SAFETY */
2931 key_beg = *key_nextp;
2932 if (key_beg == 0 || key_beg > key_end)
2938 * If the chain element is an indirect block it becomes the new
2939 * parent and we loop on it. We must maintain our top-down locks
2940 * to prevent the flusher from interfering (i.e. doing a
2941 * delete-duplicate and leaving us recursing down a deleted chain).
2943 * The parent always has to be locked with at least RESOLVE_MAYBE
2944 * so we can access its data. It might need a fixup if the caller
2945 * passed incompatible flags. Be careful not to cause a deadlock
2946 * as a data-load requires an exclusive lock.
2948 * If HAMMER2_LOOKUP_MATCHIND is set and the indirect block's key
2949 * range is within the requested key range we return the indirect
2950 * block and do NOT loop. This is usually only used to acquire
2953 if (chain->bref.type == HAMMER2_BREF_TYPE_INDIRECT ||
2954 chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2955 save_mtx = parent->lock;
2956 hammer2_chain_unlock(parent);
2957 hammer2_chain_drop(parent);
2958 *parentp = parent = chain;
2959 chain = NULL; /* SAFETY */
2964 * All done, return the locked chain.
2966 * If the caller does not want a locked chain, replace the lock with
2967 * a ref. Perhaps this can eventually be optimized to not obtain the
2968 * lock in the first place for situations where the data does not
2969 * need to be resolved.
2971 * NOTE! A chain->error must be tested by the caller upon return.
2972 * *errorp is only set based on issues which occur while
2973 * trying to reach the chain.
2979 * After having issued a lookup we can iterate all matching keys.
2981 * If chain is non-NULL we continue the iteration from just after it's index.
2983 * If chain is NULL we assume the parent was exhausted and continue the
2984 * iteration at the next parent.
2986 * If a fatal error occurs (typically an I/O error), a dummy chain is
2987 * returned with chain->error and error-identifying information set. This
2988 * chain will assert if you try to do anything fancy with it.
2990 * XXX Depending on where the error occurs we should allow continued iteration.
2992 * parent must be locked on entry and remains locked throughout. chain's
2993 * lock status must match flags. Chain is always at least referenced.
2995 * WARNING! The MATCHIND flag does not apply to this function.
2998 hammer2_chain_next(hammer2_chain_t **parentp, hammer2_chain_t *chain,
2999 hammer2_key_t *key_nextp,
3000 hammer2_key_t key_beg, hammer2_key_t key_end,
3001 int *errorp, int flags)
3003 hammer2_chain_t *parent;
3007 * Calculate locking flags for upward recursion.
3009 how_maybe = HAMMER2_RESOLVE_MAYBE;
3010 if (flags & HAMMER2_LOOKUP_SHARED)
3011 how_maybe |= HAMMER2_RESOLVE_SHARED;
3017 * Calculate the next index and recalculate the parent if necessary.
3020 key_beg = chain->bref.key +
3021 ((hammer2_key_t)1 << chain->bref.keybits);
3022 hammer2_chain_unlock(chain);
3023 hammer2_chain_drop(chain);
3026 * chain invalid past this point, but we can still do a
3027 * pointer comparison w/parent.
3029 * Any scan where the lookup returned degenerate data embedded
3030 * in the inode has an invalid index and must terminate.
3032 if (chain == parent)
3034 if (key_beg == 0 || key_beg > key_end)
3037 } else if (parent->bref.type != HAMMER2_BREF_TYPE_INDIRECT &&
3038 parent->bref.type != HAMMER2_BREF_TYPE_FREEMAP_NODE) {
3040 * We reached the end of the iteration.
3045 * Continue iteration with next parent unless the current
3046 * parent covers the range.
3048 * (This also handles the case of a deleted, empty indirect
3051 key_beg = parent->bref.key +
3052 ((hammer2_key_t)1 << parent->bref.keybits);
3053 if (key_beg == 0 || key_beg > key_end)
3055 parent = hammer2_chain_repparent(parentp, how_maybe);
3061 return (hammer2_chain_lookup(parentp, key_nextp,
3067 * Caller wishes to iterate chains under parent, loading new chains into
3068 * chainp. Caller must initialize *chainp to NULL and *firstp to 1, and
3069 * then call hammer2_chain_scan() repeatedly until a non-zero return.
3070 * During the scan, *firstp will be set to 0 and (*chainp) will be replaced
3071 * with the returned chain for the scan. The returned *chainp will be
3072 * locked and referenced. Any prior contents will be unlocked and dropped.
3074 * Caller should check the return value. A normal scan EOF will return
3075 * exactly HAMMER2_ERROR_EOF. Any other non-zero value indicates an
3076 * error trying to access parent data. Any error in the returned chain
3077 * must be tested separately by the caller.
3079 * (*chainp) is dropped on each scan, but will only be set if the returned
3080 * element itself can recurse. Leaf elements are NOT resolved, loaded, or
3081 * returned via *chainp. The caller will get their bref only.
3083 * The raw scan function is similar to lookup/next but does not seek to a key.
3084 * Blockrefs are iterated via first_bref = (parent, NULL) and
3085 * next_chain = (parent, bref).
3087 * The passed-in parent must be locked and its data resolved. The function
3088 * nominally returns a locked and referenced *chainp != NULL for chains
3089 * the caller might need to recurse on (and will dipose of any *chainp passed
3090 * in). The caller must check the chain->bref.type either way.
3093 hammer2_chain_scan(hammer2_chain_t *parent, hammer2_chain_t **chainp,
3094 hammer2_blockref_t *bref, int *firstp,
3098 hammer2_blockref_t *base;
3099 hammer2_blockref_t *bref_ptr;
3101 hammer2_key_t next_key;
3102 hammer2_chain_t *chain = NULL;
3104 int how_always = HAMMER2_RESOLVE_ALWAYS;
3105 int how_maybe = HAMMER2_RESOLVE_MAYBE;
3108 int maxloops = 300000;
3115 * Scan flags borrowed from lookup.
3117 if (flags & HAMMER2_LOOKUP_ALWAYS) {
3118 how_maybe = how_always;
3119 how = HAMMER2_RESOLVE_ALWAYS;
3120 } else if (flags & HAMMER2_LOOKUP_NODATA) {
3121 how = HAMMER2_RESOLVE_NEVER;
3123 how = HAMMER2_RESOLVE_MAYBE;
3125 if (flags & HAMMER2_LOOKUP_SHARED) {
3126 how_maybe |= HAMMER2_RESOLVE_SHARED;
3127 how_always |= HAMMER2_RESOLVE_SHARED;
3128 how |= HAMMER2_RESOLVE_SHARED;
3132 * Calculate key to locate first/next element, unlocking the previous
3133 * element as we go. Be careful, the key calculation can overflow.
3135 * (also reset bref to NULL)
3141 key = bref->key + ((hammer2_key_t)1 << bref->keybits);
3142 if ((chain = *chainp) != NULL) {
3144 hammer2_chain_unlock(chain);
3145 hammer2_chain_drop(chain);
3149 error |= HAMMER2_ERROR_EOF;
3155 if (parent->error) {
3156 error = parent->error;
3159 if (--maxloops == 0)
3160 panic("hammer2_chain_scan: maxloops");
3163 * Locate the blockref array. Currently we do a fully associative
3164 * search through the array.
3166 switch(parent->bref.type) {
3167 case HAMMER2_BREF_TYPE_INODE:
3169 * An inode with embedded data has no sub-chains.
3171 * WARNING! Bulk scan code may pass a static chain marked
3172 * as BREF_TYPE_INODE with a copy of the volume
3173 * root blockset to snapshot the volume.
3175 if (parent->data->ipdata.meta.op_flags &
3176 HAMMER2_OPFLAG_DIRECTDATA) {
3177 error |= HAMMER2_ERROR_EOF;
3180 base = &parent->data->ipdata.u.blockset.blockref[0];
3181 count = HAMMER2_SET_COUNT;
3183 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3184 case HAMMER2_BREF_TYPE_INDIRECT:
3186 * Optimize indirect blocks in the INITIAL state to avoid
3189 if (parent->flags & HAMMER2_CHAIN_INITIAL) {
3192 if (parent->data == NULL)
3193 panic("parent->data is NULL");
3194 base = &parent->data->npdata[0];
3196 count = parent->bytes / sizeof(hammer2_blockref_t);
3198 case HAMMER2_BREF_TYPE_VOLUME:
3199 base = &parent->data->voldata.sroot_blockset.blockref[0];
3200 count = HAMMER2_SET_COUNT;
3202 case HAMMER2_BREF_TYPE_FREEMAP:
3203 base = &parent->data->blkset.blockref[0];
3204 count = HAMMER2_SET_COUNT;
3207 panic("hammer2_chain_scan: unrecognized blockref type: %d",
3209 base = NULL; /* safety */
3210 count = 0; /* safety */
3214 * Merged scan to find next candidate.
3216 * hammer2_base_*() functions require the parent->core.live_* fields
3217 * to be synchronized.
3219 * We need to hold the spinlock to access the block array and RB tree
3220 * and to interlock chain creation.
3222 if ((parent->flags & HAMMER2_CHAIN_COUNTEDBREFS) == 0)
3223 hammer2_chain_countbrefs(parent, base, count);
3227 hammer2_spin_ex(&parent->core.spin);
3228 chain = hammer2_combined_find(parent, base, count,
3230 key, HAMMER2_KEY_MAX,
3232 generation = parent->core.generation;
3235 * Exhausted parent chain, we're done.
3237 if (bref_ptr == NULL) {
3238 hammer2_spin_unex(&parent->core.spin);
3239 KKASSERT(chain == NULL);
3240 error |= HAMMER2_ERROR_EOF;
3245 * Copy into the supplied stack-based blockref.
3250 * Selected from blockref or in-memory chain.
3252 if (chain == NULL) {
3253 switch(bref->type) {
3254 case HAMMER2_BREF_TYPE_INODE:
3255 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3256 case HAMMER2_BREF_TYPE_INDIRECT:
3257 case HAMMER2_BREF_TYPE_VOLUME:
3258 case HAMMER2_BREF_TYPE_FREEMAP:
3260 * Recursion, always get the chain
3262 hammer2_spin_unex(&parent->core.spin);
3263 chain = hammer2_chain_get(parent, generation,
3270 * No recursion, do not waste time instantiating
3271 * a chain, just iterate using the bref.
3273 hammer2_spin_unex(&parent->core.spin);
3278 * Recursion or not we need the chain in order to supply
3281 hammer2_chain_ref(chain);
3282 hammer2_spin_unex(&parent->core.spin);
3283 hammer2_chain_lock(chain, how);
3286 (bcmp(bref, &chain->bref, sizeof(*bref)) ||
3287 chain->parent != parent)) {
3288 hammer2_chain_unlock(chain);
3289 hammer2_chain_drop(chain);
3295 * Skip deleted chains (XXX cache 'i' end-of-block-array? XXX)
3297 * NOTE: chain's key range is not relevant as there might be
3298 * one-offs within the range that are not deleted.
3300 * NOTE: XXX this could create problems with scans used in
3301 * situations other than mount-time recovery.
3303 * NOTE: Lookups can race delete-duplicate because
3304 * delete-duplicate does not lock the parent's core
3305 * (they just use the spinlock on the core).
3307 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) {
3308 hammer2_chain_unlock(chain);
3309 hammer2_chain_drop(chain);
3314 error |= HAMMER2_ERROR_EOF;
3322 * All done, return the bref or NULL, supply chain if necessary.
3330 * Create and return a new hammer2 system memory structure of the specified
3331 * key, type and size and insert it under (*parentp). This is a full
3332 * insertion, based on the supplied key/keybits, and may involve creating
3333 * indirect blocks and moving other chains around via delete/duplicate.
3335 * This call can be made with parent == NULL as long as a non -1 methods
3336 * is supplied. hmp must also be supplied in this situation (otherwise
3337 * hmp is extracted from the supplied parent). The chain will be detached
3338 * from the topology. A later call with both parent and chain can be made
3341 * THE CALLER MUST HAVE ALREADY PROPERLY SEEKED (*parentp) TO THE INSERTION
3342 * POINT SANS ANY REQUIRED INDIRECT BLOCK CREATIONS DUE TO THE ARRAY BEING
3343 * FULL. This typically means that the caller is creating the chain after
3344 * doing a hammer2_chain_lookup().
3346 * (*parentp) must be exclusive locked and may be replaced on return
3347 * depending on how much work the function had to do.
3349 * (*parentp) must not be errored or this function will assert.
3351 * (*chainp) usually starts out NULL and returns the newly created chain,
3352 * but if the caller desires the caller may allocate a disconnected chain
3353 * and pass it in instead.
3355 * This function should NOT be used to insert INDIRECT blocks. It is
3356 * typically used to create/insert inodes and data blocks.
3358 * Caller must pass-in an exclusively locked parent the new chain is to
3359 * be inserted under, and optionally pass-in a disconnected, exclusively
3360 * locked chain to insert (else we create a new chain). The function will
3361 * adjust (*parentp) as necessary, create or connect the chain, and
3362 * return an exclusively locked chain in *chainp.
3364 * When creating a PFSROOT inode under the super-root, pmp is typically NULL
3365 * and will be reassigned.
3367 * NOTE: returns HAMMER_ERROR_* flags
3370 hammer2_chain_create(hammer2_chain_t **parentp, hammer2_chain_t **chainp,
3371 hammer2_dev_t *hmp, hammer2_pfs_t *pmp, int methods,
3372 hammer2_key_t key, int keybits, int type, size_t bytes,
3373 hammer2_tid_t mtid, hammer2_off_t dedup_off, int flags)
3375 hammer2_chain_t *chain;
3376 hammer2_chain_t *parent;
3377 hammer2_blockref_t *base;
3378 hammer2_blockref_t dummy;
3382 int maxloops = 300000;
3385 * Topology may be crossing a PFS boundary.
3389 KKASSERT(hammer2_mtx_owned(&parent->lock));
3390 KKASSERT(parent->error == 0);
3395 if (chain == NULL) {
3397 * First allocate media space and construct the dummy bref,
3398 * then allocate the in-memory chain structure. Set the
3399 * INITIAL flag for fresh chains which do not have embedded
3402 * XXX for now set the check mode of the child based on
3403 * the parent or, if the parent is an inode, the
3404 * specification in the inode.
3406 bzero(&dummy, sizeof(dummy));
3409 dummy.keybits = keybits;
3410 dummy.data_off = hammer2_getradix(bytes);
3413 * Inherit methods from parent by default. Primarily used
3414 * for BREF_TYPE_DATA. Non-data types *must* be set to
3415 * a non-NONE check algorithm.
3418 dummy.methods = parent->bref.methods;
3420 dummy.methods = (uint8_t)methods;
3422 if (type != HAMMER2_BREF_TYPE_DATA &&
3423 HAMMER2_DEC_CHECK(dummy.methods) == HAMMER2_CHECK_NONE) {
3425 HAMMER2_ENC_CHECK(HAMMER2_CHECK_DEFAULT);
3428 chain = hammer2_chain_alloc(hmp, pmp, &dummy);
3431 * Lock the chain manually, chain_lock will load the chain
3432 * which we do NOT want to do. (note: chain->refs is set
3433 * to 1 by chain_alloc() for us, but lockcnt is not).
3436 hammer2_mtx_ex(&chain->lock);
3440 * Set INITIAL to optimize I/O. The flag will generally be
3441 * processed when we call hammer2_chain_modify().
3443 * Recalculate bytes to reflect the actual media block
3444 * allocation. Handle special case radix 0 == 0 bytes.
3446 bytes = (size_t)(chain->bref.data_off & HAMMER2_OFF_MASK_RADIX);
3448 bytes = (hammer2_off_t)1 << bytes;
3449 chain->bytes = bytes;
3452 case HAMMER2_BREF_TYPE_VOLUME:
3453 case HAMMER2_BREF_TYPE_FREEMAP:
3454 panic("hammer2_chain_create: called with volume type");
3456 case HAMMER2_BREF_TYPE_INDIRECT:
3457 panic("hammer2_chain_create: cannot be used to"
3458 "create indirect block");
3460 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3461 panic("hammer2_chain_create: cannot be used to"
3462 "create freemap root or node");
3464 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
3465 KKASSERT(bytes == sizeof(chain->data->bmdata));
3467 case HAMMER2_BREF_TYPE_DIRENT:
3468 case HAMMER2_BREF_TYPE_INODE:
3469 case HAMMER2_BREF_TYPE_DATA:
3472 * leave chain->data NULL, set INITIAL
3474 KKASSERT(chain->data == NULL);
3475 atomic_set_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
3480 * We are reattaching a previously deleted chain, possibly
3481 * under a new parent and possibly with a new key/keybits.
3482 * The chain does not have to be in a modified state. The
3483 * UPDATE flag will be set later on in this routine.
3485 * Do NOT mess with the current state of the INITIAL flag.
3487 chain->bref.key = key;
3488 chain->bref.keybits = keybits;
3489 if (chain->flags & HAMMER2_CHAIN_DELETED)
3490 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_DELETED);
3491 KKASSERT(chain->parent == NULL);
3495 * Set the appropriate bref flag if requested.
3497 * NOTE! Callers can call this function to move chains without
3498 * knowing about special flags, so don't clear bref flags
3501 if (flags & HAMMER2_INSERT_PFSROOT)
3502 chain->bref.flags |= HAMMER2_BREF_FLAG_PFSROOT;
3508 * Calculate how many entries we have in the blockref array and
3509 * determine if an indirect block is required when inserting into
3513 if (--maxloops == 0)
3514 panic("hammer2_chain_create: maxloops");
3516 switch(parent->bref.type) {
3517 case HAMMER2_BREF_TYPE_INODE:
3518 if ((parent->data->ipdata.meta.op_flags &
3519 HAMMER2_OPFLAG_DIRECTDATA) != 0) {
3520 kprintf("hammer2: parent set for direct-data! "
3521 "pkey=%016jx ckey=%016jx\n",
3525 KKASSERT((parent->data->ipdata.meta.op_flags &
3526 HAMMER2_OPFLAG_DIRECTDATA) == 0);
3527 KKASSERT(parent->data != NULL);
3528 base = &parent->data->ipdata.u.blockset.blockref[0];
3529 count = HAMMER2_SET_COUNT;
3531 case HAMMER2_BREF_TYPE_INDIRECT:
3532 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3533 if (parent->flags & HAMMER2_CHAIN_INITIAL)
3536 base = &parent->data->npdata[0];
3537 count = parent->bytes / sizeof(hammer2_blockref_t);
3539 case HAMMER2_BREF_TYPE_VOLUME:
3540 KKASSERT(parent->data != NULL);
3541 base = &parent->data->voldata.sroot_blockset.blockref[0];
3542 count = HAMMER2_SET_COUNT;
3544 case HAMMER2_BREF_TYPE_FREEMAP:
3545 KKASSERT(parent->data != NULL);
3546 base = &parent->data->blkset.blockref[0];
3547 count = HAMMER2_SET_COUNT;
3550 panic("hammer2_chain_create: unrecognized blockref type: %d",
3558 * Make sure we've counted the brefs
3560 if ((parent->flags & HAMMER2_CHAIN_COUNTEDBREFS) == 0)
3561 hammer2_chain_countbrefs(parent, base, count);
3563 KASSERT(parent->core.live_count >= 0 &&
3564 parent->core.live_count <= count,
3565 ("bad live_count %d/%d (%02x, %d)",
3566 parent->core.live_count, count,
3567 parent->bref.type, parent->bytes));
3570 * If no free blockref could be found we must create an indirect
3571 * block and move a number of blockrefs into it. With the parent
3572 * locked we can safely lock each child in order to delete+duplicate
3573 * it without causing a deadlock.
3575 * This may return the new indirect block or the old parent depending
3576 * on where the key falls. NULL is returned on error.
3578 if (parent->core.live_count == count) {
3579 hammer2_chain_t *nparent;
3581 KKASSERT((flags & HAMMER2_INSERT_SAMEPARENT) == 0);
3583 nparent = hammer2_chain_create_indirect(parent, key, keybits,
3584 mtid, type, &error);
3585 if (nparent == NULL) {
3587 hammer2_chain_drop(chain);
3591 if (parent != nparent) {
3592 hammer2_chain_unlock(parent);
3593 hammer2_chain_drop(parent);
3594 parent = *parentp = nparent;
3600 * fall through if parent, or skip to here if no parent.
3603 if (chain->flags & HAMMER2_CHAIN_DELETED)
3604 kprintf("Inserting deleted chain @%016jx\n",
3608 * Link the chain into its parent.
3610 if (chain->parent != NULL)
3611 panic("hammer2: hammer2_chain_create: chain already connected");
3612 KKASSERT(chain->parent == NULL);
3614 KKASSERT(parent->core.live_count < count);
3615 hammer2_chain_insert(parent, chain,
3616 HAMMER2_CHAIN_INSERT_SPIN |
3617 HAMMER2_CHAIN_INSERT_LIVE,
3623 * Mark the newly created chain modified. This will cause
3624 * UPDATE to be set and process the INITIAL flag.
3626 * Device buffers are not instantiated for DATA elements
3627 * as these are handled by logical buffers.
3629 * Indirect and freemap node indirect blocks are handled
3630 * by hammer2_chain_create_indirect() and not by this
3633 * Data for all other bref types is expected to be
3634 * instantiated (INODE, LEAF).
3636 switch(chain->bref.type) {
3637 case HAMMER2_BREF_TYPE_DATA:
3638 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
3639 case HAMMER2_BREF_TYPE_DIRENT:
3640 case HAMMER2_BREF_TYPE_INODE:
3641 error = hammer2_chain_modify(chain, mtid, dedup_off,
3642 HAMMER2_MODIFY_OPTDATA);
3646 * Remaining types are not supported by this function.
3647 * In particular, INDIRECT and LEAF_NODE types are
3648 * handled by create_indirect().
3650 panic("hammer2_chain_create: bad type: %d",
3657 * When reconnecting a chain we must set UPDATE and
3658 * setflush so the flush recognizes that it must update
3659 * the bref in the parent.
3661 if ((chain->flags & HAMMER2_CHAIN_UPDATE) == 0)
3662 atomic_set_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
3666 * We must setflush(parent) to ensure that it recurses through to
3667 * chain. setflush(chain) might not work because ONFLUSH is possibly
3668 * already set in the chain (so it won't recurse up to set it in the
3672 hammer2_chain_setflush(parent);
3681 * Move the chain from its old parent to a new parent. The chain must have
3682 * already been deleted or already disconnected (or never associated) with
3683 * a parent. The chain is reassociated with the new parent and the deleted
3684 * flag will be cleared (no longer deleted). The chain's modification state
3687 * THE CALLER MUST HAVE ALREADY PROPERLY SEEKED (parent) TO THE INSERTION
3688 * POINT SANS ANY REQUIRED INDIRECT BLOCK CREATIONS DUE TO THE ARRAY BEING
3689 * FULL. This typically means that the caller is creating the chain after
3690 * doing a hammer2_chain_lookup().
3692 * Neither (parent) or (chain) can be errored.
3694 * If (parent) is non-NULL then the chain is inserted under the parent.
3696 * If (parent) is NULL then the newly duplicated chain is not inserted
3697 * anywhere, similar to if it had just been chain_alloc()'d (suitable for
3698 * passing into hammer2_chain_create() after this function returns).
3700 * WARNING! This function calls create which means it can insert indirect
3701 * blocks. This can cause other unrelated chains in the parent to
3702 * be moved to a newly inserted indirect block in addition to the
3706 hammer2_chain_rename(hammer2_chain_t **parentp, hammer2_chain_t *chain,
3707 hammer2_tid_t mtid, int flags)
3709 hammer2_blockref_t *bref;
3711 hammer2_chain_t *parent;
3715 * WARNING! We should never resolve DATA to device buffers
3716 * (XXX allow it if the caller did?), and since
3717 * we currently do not have the logical buffer cache
3718 * buffer in-hand to fix its cached physical offset
3719 * we also force the modify code to not COW it. XXX
3721 * NOTE! We allow error'd chains to be renamed. The bref itself
3722 * is good and can be renamed. The content, however, may
3726 KKASSERT(chain->parent == NULL);
3727 /*KKASSERT(chain->error == 0); allow */
3730 * Now create a duplicate of the chain structure, associating
3731 * it with the same core, making it the same size, pointing it
3732 * to the same bref (the same media block).
3734 * NOTE: Handle special radix == 0 case (means 0 bytes).
3736 bref = &chain->bref;
3737 bytes = (size_t)(bref->data_off & HAMMER2_OFF_MASK_RADIX);
3739 bytes = (hammer2_off_t)1 << bytes;
3742 * If parent is not NULL the duplicated chain will be entered under
3743 * the parent and the UPDATE bit set to tell flush to update
3746 * We must setflush(parent) to ensure that it recurses through to
3747 * chain. setflush(chain) might not work because ONFLUSH is possibly
3748 * already set in the chain (so it won't recurse up to set it in the
3751 * Having both chains locked is extremely important for atomicy.
3753 if (parentp && (parent = *parentp) != NULL) {
3754 KKASSERT(hammer2_mtx_owned(&parent->lock));
3755 KKASSERT(parent->refs > 0);
3756 KKASSERT(parent->error == 0);
3758 hammer2_chain_create(parentp, &chain, NULL, chain->pmp,
3759 HAMMER2_METH_DEFAULT,
3760 bref->key, bref->keybits, bref->type,
3761 chain->bytes, mtid, 0, flags);
3762 KKASSERT(chain->flags & HAMMER2_CHAIN_UPDATE);
3763 hammer2_chain_setflush(*parentp);
3768 * This works in tandem with delete_obref() to install a blockref in
3769 * (typically) an indirect block that is associated with the chain being
3770 * moved to *parentp.
3772 * The reason we need this function is that the caller needs to maintain
3773 * the blockref as it was, and not generate a new blockref for what might
3774 * be a modified chain. Otherwise stuff will leak into the flush that
3775 * the flush code's FLUSH_INODE_STOP flag is unable to catch.
3777 * It is EXTREMELY important that we properly set CHAIN_BMAPUPD and
3778 * CHAIN_UPDATE. We must set BMAPUPD if the bref does not match, and
3779 * we must clear CHAIN_UPDATE (that was likely set by the chain_rename) if
3780 * it does. Otherwise we can end up in a situation where H2 is unable to
3781 * clean up the in-memory chain topology.
3783 * The reason for this is that flushes do not generally flush through
3784 * BREF_TYPE_INODE chains and depend on a hammer2_inode_t queued to syncq
3785 * or sideq to properly flush and dispose of the related inode chain's flags.
3786 * Situations where the inode is not actually modified by the frontend,
3787 * but where we have to move the related chains around as we insert or cleanup
3788 * indirect blocks, can leave us with a 'dirty' (non-disposable) in-memory
3789 * inode chain that does not have a hammer2_inode_t associated with it.
3792 hammer2_chain_rename_obref(hammer2_chain_t **parentp, hammer2_chain_t *chain,
3793 hammer2_tid_t mtid, int flags,
3794 hammer2_blockref_t *obref)
3796 hammer2_chain_rename(parentp, chain, mtid, flags);
3798 if (obref->type != HAMMER2_BREF_TYPE_EMPTY) {
3799 hammer2_blockref_t *tbase;
3802 KKASSERT((chain->flags & HAMMER2_CHAIN_BMAPPED) == 0);
3803 hammer2_chain_modify(*parentp, mtid, 0, 0);
3804 tbase = hammer2_chain_base_and_count(*parentp, &tcount);
3805 hammer2_base_insert(*parentp, tbase, tcount, chain, obref);
3806 if (bcmp(obref, &chain->bref, sizeof(chain->bref))) {
3807 atomic_set_int(&chain->flags, HAMMER2_CHAIN_BMAPUPD |
3808 HAMMER2_CHAIN_UPDATE);
3810 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
3816 * Helper function for deleting chains.
3818 * The chain is removed from the live view (the RBTREE) as well as the parent's
3819 * blockmap. Both chain and its parent must be locked.
3821 * parent may not be errored. chain can be errored.
3824 _hammer2_chain_delete_helper(hammer2_chain_t *parent, hammer2_chain_t *chain,
3825 hammer2_tid_t mtid, int flags,
3826 hammer2_blockref_t *obref)
3831 KKASSERT((chain->flags & (HAMMER2_CHAIN_DELETED |
3832 HAMMER2_CHAIN_FICTITIOUS)) == 0);
3833 KKASSERT(chain->parent == parent);
3836 if (chain->flags & HAMMER2_CHAIN_BMAPPED) {
3838 * Chain is blockmapped, so there must be a parent.
3839 * Atomically remove the chain from the parent and remove
3840 * the blockmap entry. The parent must be set modified
3841 * to remove the blockmap entry.
3843 hammer2_blockref_t *base;
3846 KKASSERT(parent != NULL);
3847 KKASSERT(parent->error == 0);
3848 KKASSERT((parent->flags & HAMMER2_CHAIN_INITIAL) == 0);
3849 error = hammer2_chain_modify(parent, mtid, 0, 0);
3854 * Calculate blockmap pointer
3856 KKASSERT(chain->flags & HAMMER2_CHAIN_ONRBTREE);
3857 hammer2_spin_ex(&chain->core.spin);
3858 hammer2_spin_ex(&parent->core.spin);
3860 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DELETED);
3861 atomic_add_int(&parent->core.live_count, -1);
3862 ++parent->core.generation;
3863 RB_REMOVE(hammer2_chain_tree, &parent->core.rbtree, chain);
3864 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE);
3865 --parent->core.chain_count;
3866 chain->parent = NULL;
3868 switch(parent->bref.type) {
3869 case HAMMER2_BREF_TYPE_INODE:
3871 * Access the inode's block array. However, there
3872 * is no block array if the inode is flagged
3876 (parent->data->ipdata.meta.op_flags &
3877 HAMMER2_OPFLAG_DIRECTDATA) == 0) {
3879 &parent->data->ipdata.u.blockset.blockref[0];
3883 count = HAMMER2_SET_COUNT;
3885 case HAMMER2_BREF_TYPE_INDIRECT:
3886 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3888 base = &parent->data->npdata[0];
3891 count = parent->bytes / sizeof(hammer2_blockref_t);
3893 case HAMMER2_BREF_TYPE_VOLUME:
3894 base = &parent->data->voldata.
3895 sroot_blockset.blockref[0];
3896 count = HAMMER2_SET_COUNT;
3898 case HAMMER2_BREF_TYPE_FREEMAP:
3899 base = &parent->data->blkset.blockref[0];
3900 count = HAMMER2_SET_COUNT;
3905 panic("_hammer2_chain_delete_helper: "
3906 "unrecognized blockref type: %d",
3911 * delete blockmapped chain from its parent.
3913 * The parent is not affected by any statistics in chain
3914 * which are pending synchronization. That is, there is
3915 * nothing to undo in the parent since they have not yet
3916 * been incorporated into the parent.
3918 * The parent is affected by statistics stored in inodes.
3919 * Those have already been synchronized, so they must be
3920 * undone. XXX split update possible w/delete in middle?
3923 hammer2_base_delete(parent, base, count, chain, obref);
3925 hammer2_spin_unex(&parent->core.spin);
3926 hammer2_spin_unex(&chain->core.spin);
3927 } else if (chain->flags & HAMMER2_CHAIN_ONRBTREE) {
3929 * Chain is not blockmapped but a parent is present.
3930 * Atomically remove the chain from the parent. There is
3931 * no blockmap entry to remove.
3933 * Because chain was associated with a parent but not
3934 * synchronized, the chain's *_count_up fields contain
3935 * inode adjustment statistics which must be undone.
3937 hammer2_spin_ex(&chain->core.spin);
3938 hammer2_spin_ex(&parent->core.spin);
3939 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DELETED);
3940 atomic_add_int(&parent->core.live_count, -1);
3941 ++parent->core.generation;
3942 RB_REMOVE(hammer2_chain_tree, &parent->core.rbtree, chain);
3943 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE);
3944 --parent->core.chain_count;
3945 chain->parent = NULL;
3946 hammer2_spin_unex(&parent->core.spin);
3947 hammer2_spin_unex(&chain->core.spin);
3950 * Chain is not blockmapped and has no parent. This
3951 * is a degenerate case.
3953 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DELETED);
3960 * Create an indirect block that covers one or more of the elements in the
3961 * current parent. Either returns the existing parent with no locking or
3962 * ref changes or returns the new indirect block locked and referenced
3963 * and leaving the original parent lock/ref intact as well.
3965 * If an error occurs, NULL is returned and *errorp is set to the H2 error.
3967 * The returned chain depends on where the specified key falls.
3969 * The key/keybits for the indirect mode only needs to follow three rules:
3971 * (1) That all elements underneath it fit within its key space and
3973 * (2) That all elements outside it are outside its key space.
3975 * (3) When creating the new indirect block any elements in the current
3976 * parent that fit within the new indirect block's keyspace must be
3977 * moved into the new indirect block.
3979 * (4) The keyspace chosen for the inserted indirect block CAN cover a wider
3980 * keyspace the the current parent, but lookup/iteration rules will
3981 * ensure (and must ensure) that rule (2) for all parents leading up
3982 * to the nearest inode or the root volume header is adhered to. This
3983 * is accomplished by always recursing through matching keyspaces in
3984 * the hammer2_chain_lookup() and hammer2_chain_next() API.
3986 * The current implementation calculates the current worst-case keyspace by
3987 * iterating the current parent and then divides it into two halves, choosing
3988 * whichever half has the most elements (not necessarily the half containing
3989 * the requested key).
3991 * We can also opt to use the half with the least number of elements. This
3992 * causes lower-numbered keys (aka logical file offsets) to recurse through
3993 * fewer indirect blocks and higher-numbered keys to recurse through more.
3994 * This also has the risk of not moving enough elements to the new indirect
3995 * block and being forced to create several indirect blocks before the element
3998 * Must be called with an exclusively locked parent.
4000 * NOTE: *errorp set to HAMMER_ERROR_* flags
4002 static int hammer2_chain_indkey_freemap(hammer2_chain_t *parent,
4003 hammer2_key_t *keyp, int keybits,
4004 hammer2_blockref_t *base, int count);
4005 static int hammer2_chain_indkey_file(hammer2_chain_t *parent,
4006 hammer2_key_t *keyp, int keybits,
4007 hammer2_blockref_t *base, int count,
4009 static int hammer2_chain_indkey_dir(hammer2_chain_t *parent,
4010 hammer2_key_t *keyp, int keybits,
4011 hammer2_blockref_t *base, int count,
4015 hammer2_chain_create_indirect(hammer2_chain_t *parent,
4016 hammer2_key_t create_key, int create_bits,
4017 hammer2_tid_t mtid, int for_type, int *errorp)
4020 hammer2_blockref_t *base;
4021 hammer2_blockref_t *bref;
4022 hammer2_blockref_t bsave;
4023 hammer2_blockref_t dummy;
4024 hammer2_chain_t *chain;
4025 hammer2_chain_t *ichain;
4026 hammer2_key_t key = create_key;
4027 hammer2_key_t key_beg;
4028 hammer2_key_t key_end;
4029 hammer2_key_t key_next;
4030 int keybits = create_bits;
4038 int maxloops = 300000;
4041 * Calculate the base blockref pointer or NULL if the chain
4042 * is known to be empty. We need to calculate the array count
4043 * for RB lookups either way.
4046 KKASSERT(hammer2_mtx_owned(&parent->lock));
4049 * Pre-modify the parent now to avoid having to deal with error
4050 * processing if we tried to later (in the middle of our loop).
4052 * We are going to be moving bref's around, the indirect blocks
4053 * cannot be in an initial state. Do not pass MODIFY_OPTDATA.
4055 *errorp = hammer2_chain_modify(parent, mtid, 0, 0);
4057 kprintf("hammer2_create_indirect: error %08x %s\n",
4058 *errorp, hammer2_error_str(*errorp));
4061 KKASSERT((parent->flags & HAMMER2_CHAIN_INITIAL) == 0);
4063 /*hammer2_chain_modify(&parent, HAMMER2_MODIFY_OPTDATA);*/
4064 base = hammer2_chain_base_and_count(parent, &count);
4067 * How big should our new indirect block be? It has to be at least
4068 * as large as its parent for splits to work properly.
4070 * The freemap uses a specific indirect block size. The number of
4071 * levels are built dynamically and ultimately depend on the size
4072 * volume. Because freemap blocks are taken from the reserved areas
4073 * of the volume our goal is efficiency (fewer levels) and not so
4074 * much to save disk space.
4076 * The first indirect block level for a directory usually uses
4077 * HAMMER2_IND_BYTES_MIN (4KB = 32 directory entries). Due to
4078 * the hash mechanism, this typically gives us a nominal
4079 * 32 * 4 entries with one level of indirection.
4081 * We use HAMMER2_IND_BYTES_NOM (16KB = 128 blockrefs) for FILE
4082 * indirect blocks. The initial 4 entries in the inode gives us
4083 * 256KB. Up to 4 indirect blocks gives us 32MB. Three levels
4084 * of indirection gives us 137GB, and so forth. H2 can support
4085 * huge file sizes but they are not typical, so we try to stick
4086 * with compactness and do not use a larger indirect block size.
4088 * We could use 64KB (PBUFSIZE), giving us 512 blockrefs, but
4089 * due to the way indirect blocks are created this usually winds
4090 * up being extremely inefficient for small files. Even though
4091 * 16KB requires more levels of indirection for very large files,
4092 * the 16KB records can be ganged together into 64KB DIOs.
4094 if (for_type == HAMMER2_BREF_TYPE_FREEMAP_NODE ||
4095 for_type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) {
4096 nbytes = HAMMER2_FREEMAP_LEVELN_PSIZE;
4097 } else if (parent->bref.type == HAMMER2_BREF_TYPE_INODE) {
4098 if (parent->data->ipdata.meta.type ==
4099 HAMMER2_OBJTYPE_DIRECTORY)
4100 nbytes = HAMMER2_IND_BYTES_MIN; /* 4KB = 32 entries */
4102 nbytes = HAMMER2_IND_BYTES_NOM; /* 16KB = ~8MB file */
4105 nbytes = HAMMER2_IND_BYTES_NOM;
4107 if (nbytes < count * sizeof(hammer2_blockref_t)) {
4108 KKASSERT(for_type != HAMMER2_BREF_TYPE_FREEMAP_NODE &&
4109 for_type != HAMMER2_BREF_TYPE_FREEMAP_LEAF);
4110 nbytes = count * sizeof(hammer2_blockref_t);
4112 ncount = nbytes / sizeof(hammer2_blockref_t);
4115 * When creating an indirect block for a freemap node or leaf
4116 * the key/keybits must be fitted to static radix levels because
4117 * particular radix levels use particular reserved blocks in the
4120 * This routine calculates the key/radix of the indirect block
4121 * we need to create, and whether it is on the high-side or the
4125 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
4126 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
4127 keybits = hammer2_chain_indkey_freemap(parent, &key, keybits,
4130 case HAMMER2_BREF_TYPE_DATA:
4131 keybits = hammer2_chain_indkey_file(parent, &key, keybits,
4132 base, count, ncount);
4134 case HAMMER2_BREF_TYPE_DIRENT:
4135 case HAMMER2_BREF_TYPE_INODE:
4136 keybits = hammer2_chain_indkey_dir(parent, &key, keybits,
4137 base, count, ncount);
4140 panic("illegal indirect block for bref type %d", for_type);
4145 * Normalize the key for the radix being represented, keeping the
4146 * high bits and throwing away the low bits.
4148 key &= ~(((hammer2_key_t)1 << keybits) - 1);
4151 * Ok, create our new indirect block
4153 bzero(&dummy, sizeof(dummy));
4154 if (for_type == HAMMER2_BREF_TYPE_FREEMAP_NODE ||
4155 for_type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) {
4156 dummy.type = HAMMER2_BREF_TYPE_FREEMAP_NODE;
4158 dummy.type = HAMMER2_BREF_TYPE_INDIRECT;
4161 dummy.keybits = keybits;
4162 dummy.data_off = hammer2_getradix(nbytes);
4164 HAMMER2_ENC_CHECK(HAMMER2_DEC_CHECK(parent->bref.methods)) |
4165 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
4167 ichain = hammer2_chain_alloc(hmp, parent->pmp, &dummy);
4168 atomic_set_int(&ichain->flags, HAMMER2_CHAIN_INITIAL);
4169 hammer2_chain_lock(ichain, HAMMER2_RESOLVE_MAYBE);
4170 /* ichain has one ref at this point */
4173 * We have to mark it modified to allocate its block, but use
4174 * OPTDATA to allow it to remain in the INITIAL state. Otherwise
4175 * it won't be acted upon by the flush code.
4177 * XXX remove OPTDATA, we need a fully initialized indirect block to
4178 * be able to move the original blockref.
4180 *errorp = hammer2_chain_modify(ichain, mtid, 0, 0);
4182 kprintf("hammer2_alloc_indirect: error %08x %s\n",
4183 *errorp, hammer2_error_str(*errorp));
4184 hammer2_chain_unlock(ichain);
4185 hammer2_chain_drop(ichain);
4188 KKASSERT((ichain->flags & HAMMER2_CHAIN_INITIAL) == 0);
4191 * Iterate the original parent and move the matching brefs into
4192 * the new indirect block.
4194 * XXX handle flushes.
4197 key_end = HAMMER2_KEY_MAX;
4198 key_next = 0; /* avoid gcc warnings */
4199 hammer2_spin_ex(&parent->core.spin);
4205 * Parent may have been modified, relocating its block array.
4206 * Reload the base pointer.
4208 base = hammer2_chain_base_and_count(parent, &count);
4210 if (++loops > 100000) {
4211 hammer2_spin_unex(&parent->core.spin);
4212 panic("excessive loops r=%d p=%p base/count %p:%d %016jx\n",
4213 reason, parent, base, count, key_next);
4217 * NOTE: spinlock stays intact, returned chain (if not NULL)
4218 * is not referenced or locked which means that we
4219 * cannot safely check its flagged / deletion status
4222 chain = hammer2_combined_find(parent, base, count,
4226 generation = parent->core.generation;
4229 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
4232 * Skip keys that are not within the key/radix of the new
4233 * indirect block. They stay in the parent.
4235 if (rounddown2(key ^ bref->key, (hammer2_key_t)1 << keybits) != 0) {
4236 goto next_key_spinlocked;
4240 * Load the new indirect block by acquiring the related
4241 * chains (potentially from media as it might not be
4242 * in-memory). Then move it to the new parent (ichain).
4244 * chain is referenced but not locked. We must lock the
4245 * chain to obtain definitive state.
4250 * Use chain already present in the RBTREE
4252 hammer2_chain_ref(chain);
4253 hammer2_spin_unex(&parent->core.spin);
4254 hammer2_chain_lock(chain, HAMMER2_RESOLVE_NEVER);
4257 * Get chain for blockref element. _get returns NULL
4258 * on insertion race.
4260 hammer2_spin_unex(&parent->core.spin);
4261 chain = hammer2_chain_get(parent, generation, &bsave,
4262 HAMMER2_RESOLVE_NEVER);
4263 if (chain == NULL) {
4265 hammer2_spin_ex(&parent->core.spin);
4271 * This is always live so if the chain has been deleted
4272 * we raced someone and we have to retry.
4274 * NOTE: Lookups can race delete-duplicate because
4275 * delete-duplicate does not lock the parent's core
4276 * (they just use the spinlock on the core).
4278 * (note reversed logic for this one)
4280 if (bcmp(&bsave, &chain->bref, sizeof(bsave)) ||
4281 chain->parent != parent ||
4282 (chain->flags & HAMMER2_CHAIN_DELETED)) {
4283 hammer2_chain_unlock(chain);
4284 hammer2_chain_drop(chain);
4285 if (hammer2_debug & 0x0040) {
4286 kprintf("LOST PARENT RETRY "
4287 "RETRY (%p,%p)->%p %08x\n",
4288 parent, chain->parent, chain, chain->flags);
4290 hammer2_spin_ex(&parent->core.spin);
4295 * Shift the chain to the indirect block.
4297 * WARNING! No reason for us to load chain data, pass NOSTATS
4298 * to prevent delete/insert from trying to access
4299 * inode stats (and thus asserting if there is no
4300 * chain->data loaded).
4302 * WARNING! The (parent, chain) deletion may modify the parent
4303 * and invalidate the base pointer.
4305 * WARNING! Parent must already be marked modified, so we
4306 * can assume that chain_delete always suceeds.
4308 * WARNING! hammer2_chain_repchange() does not have to be
4309 * called (and doesn't work anyway because we are
4310 * only doing a partial shift). A recursion that is
4311 * in-progress can continue at the current parent
4312 * and will be able to properly find its next key.
4314 error = hammer2_chain_delete_obref(parent, chain, mtid, 0,
4316 KKASSERT(error == 0);
4317 hammer2_chain_rename_obref(&ichain, chain, mtid, 0, &bsave);
4318 hammer2_chain_unlock(chain);
4319 hammer2_chain_drop(chain);
4320 KKASSERT(parent->refs > 0);
4322 base = NULL; /* safety */
4323 hammer2_spin_ex(&parent->core.spin);
4324 next_key_spinlocked:
4325 if (--maxloops == 0)
4326 panic("hammer2_chain_create_indirect: maxloops");
4328 if (key_next == 0 || key_next > key_end)
4333 hammer2_spin_unex(&parent->core.spin);
4336 * Insert the new indirect block into the parent now that we've
4337 * cleared out some entries in the parent. We calculated a good
4338 * insertion index in the loop above (ichain->index).
4340 * We don't have to set UPDATE here because we mark ichain
4341 * modified down below (so the normal modified -> flush -> set-moved
4342 * sequence applies).
4344 * The insertion shouldn't race as this is a completely new block
4345 * and the parent is locked.
4347 base = NULL; /* safety, parent modify may change address */
4348 KKASSERT((ichain->flags & HAMMER2_CHAIN_ONRBTREE) == 0);
4349 KKASSERT(parent->core.live_count < count);
4350 hammer2_chain_insert(parent, ichain,
4351 HAMMER2_CHAIN_INSERT_SPIN |
4352 HAMMER2_CHAIN_INSERT_LIVE,
4356 * Make sure flushes propogate after our manual insertion.
4358 hammer2_chain_setflush(ichain);
4359 hammer2_chain_setflush(parent);
4362 * Figure out what to return.
4364 if (rounddown2(create_key ^ key, (hammer2_key_t)1 << keybits)) {
4366 * Key being created is outside the key range,
4367 * return the original parent.
4369 hammer2_chain_unlock(ichain);
4370 hammer2_chain_drop(ichain);
4373 * Otherwise its in the range, return the new parent.
4374 * (leave both the new and old parent locked).
4383 * Do maintenance on an indirect chain. Both parent and chain are locked.
4385 * Returns non-zero if (chain) is deleted, either due to being empty or
4386 * because its children were safely moved into the parent.
4389 hammer2_chain_indirect_maintenance(hammer2_chain_t *parent,
4390 hammer2_chain_t *chain)
4392 hammer2_blockref_t *chain_base;
4393 hammer2_blockref_t *base;
4394 hammer2_blockref_t *bref;
4395 hammer2_blockref_t bsave;
4396 hammer2_key_t key_next;
4397 hammer2_key_t key_beg;
4398 hammer2_key_t key_end;
4399 hammer2_chain_t *sub;
4406 * Make sure we have an accurate live_count
4408 if ((chain->flags & (HAMMER2_CHAIN_INITIAL |
4409 HAMMER2_CHAIN_COUNTEDBREFS)) == 0) {
4410 base = &chain->data->npdata[0];
4411 count = chain->bytes / sizeof(hammer2_blockref_t);
4412 hammer2_chain_countbrefs(chain, base, count);
4416 * If the indirect block is empty we can delete it.
4417 * (ignore deletion error)
4419 if (chain->core.live_count == 0 && RB_EMPTY(&chain->core.rbtree)) {
4420 hammer2_chain_delete(parent, chain,
4421 chain->bref.modify_tid,
4422 HAMMER2_DELETE_PERMANENT);
4423 hammer2_chain_repchange(parent, chain);
4427 base = hammer2_chain_base_and_count(parent, &count);
4429 if ((parent->flags & (HAMMER2_CHAIN_INITIAL |
4430 HAMMER2_CHAIN_COUNTEDBREFS)) == 0) {
4431 hammer2_chain_countbrefs(parent, base, count);
4435 * Determine if we can collapse chain into parent, calculate
4436 * hysteresis for chain emptiness.
4438 if (parent->core.live_count + chain->core.live_count - 1 > count)
4440 chain_count = chain->bytes / sizeof(hammer2_blockref_t);
4441 if (chain->core.live_count > chain_count * 3 / 4)
4445 * Ok, theoretically we can collapse chain's contents into
4446 * parent. chain is locked, but any in-memory children of chain
4447 * are not. For this to work, we must be able to dispose of any
4448 * in-memory children of chain.
4450 * For now require that there are no in-memory children of chain.
4452 * WARNING! Both chain and parent must remain locked across this
4457 * Parent must be marked modified. Don't try to collapse it if we
4458 * can't mark it modified. Once modified, destroy chain to make room
4459 * and to get rid of what will be a conflicting key (this is included
4460 * in the calculation above). Finally, move the children of chain
4461 * into chain's parent.
4463 * This order creates an accounting problem for bref.embed.stats
4464 * because we destroy chain before we remove its children. Any
4465 * elements whos blockref is already synchronized will be counted
4466 * twice. To deal with the problem we clean out chain's stats prior
4469 error = hammer2_chain_modify(parent, 0, 0, 0);
4471 krateprintf(&krate_h2me, "hammer2: indirect_maint: %s\n",
4472 hammer2_error_str(error));
4475 error = hammer2_chain_modify(chain, chain->bref.modify_tid, 0, 0);
4477 krateprintf(&krate_h2me, "hammer2: indirect_maint: %s\n",
4478 hammer2_error_str(error));
4482 chain->bref.embed.stats.inode_count = 0;
4483 chain->bref.embed.stats.data_count = 0;
4484 error = hammer2_chain_delete(parent, chain,
4485 chain->bref.modify_tid,
4486 HAMMER2_DELETE_PERMANENT);
4487 KKASSERT(error == 0);
4490 * The combined_find call requires core.spin to be held. One would
4491 * think there wouldn't be any conflicts since we hold chain
4492 * exclusively locked, but the caching mechanism for 0-ref children
4493 * does not require a chain lock.
4495 hammer2_spin_ex(&chain->core.spin);
4499 key_end = HAMMER2_KEY_MAX;
4501 chain_base = &chain->data->npdata[0];
4502 chain_count = chain->bytes / sizeof(hammer2_blockref_t);
4503 sub = hammer2_combined_find(chain, chain_base, chain_count,
4507 generation = chain->core.generation;
4510 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
4514 hammer2_chain_ref(sub);
4515 hammer2_spin_unex(&chain->core.spin);
4516 hammer2_chain_lock(sub, HAMMER2_RESOLVE_NEVER);
4518 hammer2_spin_unex(&chain->core.spin);
4519 sub = hammer2_chain_get(chain, generation, &bsave,
4520 HAMMER2_RESOLVE_NEVER);
4522 hammer2_spin_ex(&chain->core.spin);
4526 if (bcmp(&bsave, &sub->bref, sizeof(bsave)) ||
4527 sub->parent != chain ||
4528 (sub->flags & HAMMER2_CHAIN_DELETED)) {
4529 hammer2_chain_unlock(sub);
4530 hammer2_chain_drop(sub);
4531 hammer2_spin_ex(&chain->core.spin);
4532 sub = NULL; /* safety */
4535 error = hammer2_chain_delete_obref(chain, sub,
4536 sub->bref.modify_tid, 0,
4538 KKASSERT(error == 0);
4539 hammer2_chain_rename_obref(&parent, sub,
4540 sub->bref.modify_tid,
4541 HAMMER2_INSERT_SAMEPARENT, &bsave);
4542 hammer2_chain_unlock(sub);
4543 hammer2_chain_drop(sub);
4544 hammer2_spin_ex(&chain->core.spin);
4550 hammer2_spin_unex(&chain->core.spin);
4552 hammer2_chain_repchange(parent, chain);
4558 * Freemap indirect blocks
4560 * Calculate the keybits and highside/lowside of the freemap node the
4561 * caller is creating.
4563 * This routine will specify the next higher-level freemap key/radix
4564 * representing the lowest-ordered set. By doing so, eventually all
4565 * low-ordered sets will be moved one level down.
4567 * We have to be careful here because the freemap reserves a limited
4568 * number of blocks for a limited number of levels. So we can't just
4569 * push indiscriminately.
4572 hammer2_chain_indkey_freemap(hammer2_chain_t *parent, hammer2_key_t *keyp,
4573 int keybits, hammer2_blockref_t *base, int count)
4575 hammer2_chain_t *chain;
4576 hammer2_blockref_t *bref;
4578 hammer2_key_t key_beg;
4579 hammer2_key_t key_end;
4580 hammer2_key_t key_next;
4583 int maxloops = 300000;
4591 * Calculate the range of keys in the array being careful to skip
4592 * slots which are overridden with a deletion.
4595 key_end = HAMMER2_KEY_MAX;
4596 hammer2_spin_ex(&parent->core.spin);
4599 if (--maxloops == 0) {
4600 panic("indkey_freemap shit %p %p:%d\n",
4601 parent, base, count);
4603 chain = hammer2_combined_find(parent, base, count,
4615 * Skip deleted chains.
4617 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) {
4618 if (key_next == 0 || key_next > key_end)
4625 * Use the full live (not deleted) element for the scan
4626 * iteration. HAMMER2 does not allow partial replacements.
4628 * XXX should be built into hammer2_combined_find().
4630 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
4632 if (keybits > bref->keybits) {
4634 keybits = bref->keybits;
4635 } else if (keybits == bref->keybits && bref->key < key) {
4642 hammer2_spin_unex(&parent->core.spin);
4645 * Return the keybits for a higher-level FREEMAP_NODE covering
4649 case HAMMER2_FREEMAP_LEVEL0_RADIX:
4650 keybits = HAMMER2_FREEMAP_LEVEL1_RADIX;
4652 case HAMMER2_FREEMAP_LEVEL1_RADIX:
4653 keybits = HAMMER2_FREEMAP_LEVEL2_RADIX;
4655 case HAMMER2_FREEMAP_LEVEL2_RADIX:
4656 keybits = HAMMER2_FREEMAP_LEVEL3_RADIX;
4658 case HAMMER2_FREEMAP_LEVEL3_RADIX:
4659 keybits = HAMMER2_FREEMAP_LEVEL4_RADIX;
4661 case HAMMER2_FREEMAP_LEVEL4_RADIX:
4662 keybits = HAMMER2_FREEMAP_LEVEL5_RADIX;
4664 case HAMMER2_FREEMAP_LEVEL5_RADIX:
4665 panic("hammer2_chain_indkey_freemap: level too high");
4668 panic("hammer2_chain_indkey_freemap: bad radix");
4677 * File indirect blocks
4679 * Calculate the key/keybits for the indirect block to create by scanning
4680 * existing keys. The key being created is also passed in *keyp and can be
4681 * inside or outside the indirect block. Regardless, the indirect block
4682 * must hold at least two keys in order to guarantee sufficient space.
4684 * We use a modified version of the freemap's fixed radix tree, but taylored
4685 * for file data. Basically we configure an indirect block encompassing the
4689 hammer2_chain_indkey_file(hammer2_chain_t *parent, hammer2_key_t *keyp,
4690 int keybits, hammer2_blockref_t *base, int count,
4693 hammer2_chain_t *chain;
4694 hammer2_blockref_t *bref;
4696 hammer2_key_t key_beg;
4697 hammer2_key_t key_end;
4698 hammer2_key_t key_next;
4702 int maxloops = 300000;
4710 * Calculate the range of keys in the array being careful to skip
4711 * slots which are overridden with a deletion.
4713 * Locate the smallest key.
4716 key_end = HAMMER2_KEY_MAX;
4717 hammer2_spin_ex(&parent->core.spin);
4720 if (--maxloops == 0) {
4721 panic("indkey_freemap shit %p %p:%d\n",
4722 parent, base, count);
4724 chain = hammer2_combined_find(parent, base, count,
4736 * Skip deleted chains.
4738 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) {
4739 if (key_next == 0 || key_next > key_end)
4746 * Use the full live (not deleted) element for the scan
4747 * iteration. HAMMER2 does not allow partial replacements.
4749 * XXX should be built into hammer2_combined_find().
4751 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
4753 if (keybits > bref->keybits) {
4755 keybits = bref->keybits;
4756 } else if (keybits == bref->keybits && bref->key < key) {
4763 hammer2_spin_unex(&parent->core.spin);
4766 * Calculate the static keybits for a higher-level indirect block
4767 * that contains the key.
4772 case HAMMER2_IND_BYTES_MIN / sizeof(hammer2_blockref_t):
4773 nradix = HAMMER2_IND_RADIX_MIN - HAMMER2_BLOCKREF_RADIX;
4775 case HAMMER2_IND_BYTES_NOM / sizeof(hammer2_blockref_t):
4776 nradix = HAMMER2_IND_RADIX_NOM - HAMMER2_BLOCKREF_RADIX;
4778 case HAMMER2_IND_BYTES_MAX / sizeof(hammer2_blockref_t):
4779 nradix = HAMMER2_IND_RADIX_MAX - HAMMER2_BLOCKREF_RADIX;
4782 panic("bad ncount %d\n", ncount);
4788 * The largest radix that can be returned for an indirect block is
4789 * 63 bits. (The largest practical indirect block radix is actually
4790 * 62 bits because the top-level inode or volume root contains four
4791 * entries, but allow 63 to be returned).
4796 return keybits + nradix;
4802 * Directory indirect blocks.
4804 * Covers both the inode index (directory of inodes), and directory contents
4805 * (filenames hardlinked to inodes).
4807 * Because directory keys are hashed we generally try to cut the space in
4808 * half. We accomodate the inode index (which tends to have linearly
4809 * increasing inode numbers) by ensuring that the keyspace is at least large
4810 * enough to fill up the indirect block being created.
4813 hammer2_chain_indkey_dir(hammer2_chain_t *parent, hammer2_key_t *keyp,
4814 int keybits, hammer2_blockref_t *base, int count,
4817 hammer2_blockref_t *bref;
4818 hammer2_chain_t *chain;
4819 hammer2_key_t key_beg;
4820 hammer2_key_t key_end;
4821 hammer2_key_t key_next;
4826 int maxloops = 300000;
4829 * NOTE: We can't take a shortcut here anymore for inodes because
4830 * the root directory can contain a mix of inodes and directory
4831 * entries (we used to just return 63 if parent->bref.type was
4832 * HAMMER2_BREF_TYPE_INODE.
4839 * Calculate the range of keys in the array being careful to skip
4840 * slots which are overridden with a deletion.
4843 key_end = HAMMER2_KEY_MAX;
4844 hammer2_spin_ex(&parent->core.spin);
4847 if (--maxloops == 0) {
4848 panic("indkey_freemap shit %p %p:%d\n",
4849 parent, base, count);
4851 chain = hammer2_combined_find(parent, base, count,
4865 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) {
4866 if (key_next == 0 || key_next > key_end)
4873 * Use the full live (not deleted) element for the scan
4874 * iteration. HAMMER2 does not allow partial replacements.
4876 * XXX should be built into hammer2_combined_find().
4878 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
4881 * Expand our calculated key range (key, keybits) to fit
4882 * the scanned key. nkeybits represents the full range
4883 * that we will later cut in half (two halves @ nkeybits - 1).
4886 if (nkeybits < bref->keybits) {
4887 if (bref->keybits > 64) {
4888 kprintf("bad bref chain %p bref %p\n",
4892 nkeybits = bref->keybits;
4894 while (nkeybits < 64 &&
4895 rounddown2(key ^ bref->key, (hammer2_key_t)1 << nkeybits) != 0) {
4900 * If the new key range is larger we have to determine
4901 * which side of the new key range the existing keys fall
4902 * under by checking the high bit, then collapsing the
4903 * locount into the hicount or vise-versa.
4905 if (keybits != nkeybits) {
4906 if (((hammer2_key_t)1 << (nkeybits - 1)) & key) {
4917 * The newly scanned key will be in the lower half or the
4918 * upper half of the (new) key range.
4920 if (((hammer2_key_t)1 << (nkeybits - 1)) & bref->key)
4929 hammer2_spin_unex(&parent->core.spin);
4930 bref = NULL; /* now invalid (safety) */
4933 * Adjust keybits to represent half of the full range calculated
4934 * above (radix 63 max) for our new indirect block.
4939 * Expand keybits to hold at least ncount elements. ncount will be
4940 * a power of 2. This is to try to completely fill leaf nodes (at
4941 * least for keys which are not hashes).
4943 * We aren't counting 'in' or 'out', we are counting 'high side'
4944 * and 'low side' based on the bit at (1LL << keybits). We want
4945 * everything to be inside in these cases so shift it all to
4946 * the low or high side depending on the new high bit.
4948 while (((hammer2_key_t)1 << keybits) < ncount) {
4950 if (key & ((hammer2_key_t)1 << keybits)) {
4959 if (hicount > locount)
4960 key |= (hammer2_key_t)1 << keybits;
4962 key &= ~(hammer2_key_t)1 << keybits;
4972 * Directory indirect blocks.
4974 * Covers both the inode index (directory of inodes), and directory contents
4975 * (filenames hardlinked to inodes).
4977 * Because directory keys are hashed we generally try to cut the space in
4978 * half. We accomodate the inode index (which tends to have linearly
4979 * increasing inode numbers) by ensuring that the keyspace is at least large
4980 * enough to fill up the indirect block being created.
4983 hammer2_chain_indkey_dir(hammer2_chain_t *parent, hammer2_key_t *keyp,
4984 int keybits, hammer2_blockref_t *base, int count,
4987 hammer2_blockref_t *bref;
4988 hammer2_chain_t *chain;
4989 hammer2_key_t key_beg;
4990 hammer2_key_t key_end;
4991 hammer2_key_t key_next;
4996 int maxloops = 300000;
4999 * Shortcut if the parent is the inode. In this situation the
5000 * parent has 4+1 directory entries and we are creating an indirect
5001 * block capable of holding many more.
5003 if (parent->bref.type == HAMMER2_BREF_TYPE_INODE) {
5012 * Calculate the range of keys in the array being careful to skip
5013 * slots which are overridden with a deletion.
5016 key_end = HAMMER2_KEY_MAX;
5017 hammer2_spin_ex(&parent->core.spin);
5020 if (--maxloops == 0) {
5021 panic("indkey_freemap shit %p %p:%d\n",
5022 parent, base, count);
5024 chain = hammer2_combined_find(parent, base, count,
5038 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) {
5039 if (key_next == 0 || key_next > key_end)
5046 * Use the full live (not deleted) element for the scan
5047 * iteration. HAMMER2 does not allow partial replacements.
5049 * XXX should be built into hammer2_combined_find().
5051 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
5054 * Expand our calculated key range (key, keybits) to fit
5055 * the scanned key. nkeybits represents the full range
5056 * that we will later cut in half (two halves @ nkeybits - 1).
5059 if (nkeybits < bref->keybits) {
5060 if (bref->keybits > 64) {
5061 kprintf("bad bref chain %p bref %p\n",
5065 nkeybits = bref->keybits;
5067 while (nkeybits < 64 &&
5068 (~(((hammer2_key_t)1 << nkeybits) - 1) &
5069 (key ^ bref->key)) != 0) {
5074 * If the new key range is larger we have to determine
5075 * which side of the new key range the existing keys fall
5076 * under by checking the high bit, then collapsing the
5077 * locount into the hicount or vise-versa.
5079 if (keybits != nkeybits) {
5080 if (((hammer2_key_t)1 << (nkeybits - 1)) & key) {
5091 * The newly scanned key will be in the lower half or the
5092 * upper half of the (new) key range.
5094 if (((hammer2_key_t)1 << (nkeybits - 1)) & bref->key)
5103 hammer2_spin_unex(&parent->core.spin);
5104 bref = NULL; /* now invalid (safety) */
5107 * Adjust keybits to represent half of the full range calculated
5108 * above (radix 63 max) for our new indirect block.
5113 * Expand keybits to hold at least ncount elements. ncount will be
5114 * a power of 2. This is to try to completely fill leaf nodes (at
5115 * least for keys which are not hashes).
5117 * We aren't counting 'in' or 'out', we are counting 'high side'
5118 * and 'low side' based on the bit at (1LL << keybits). We want
5119 * everything to be inside in these cases so shift it all to
5120 * the low or high side depending on the new high bit.
5122 while (((hammer2_key_t)1 << keybits) < ncount) {
5124 if (key & ((hammer2_key_t)1 << keybits)) {
5133 if (hicount > locount)
5134 key |= (hammer2_key_t)1 << keybits;
5136 key &= ~(hammer2_key_t)1 << keybits;
5146 * Sets CHAIN_DELETED and remove the chain's blockref from the parent if
5149 * Both parent and chain must be locked exclusively.
5151 * This function will modify the parent if the blockref requires removal
5152 * from the parent's block table.
5154 * This function is NOT recursive. Any entity already pushed into the
5155 * chain (such as an inode) may still need visibility into its contents,
5156 * as well as the ability to read and modify the contents. For example,
5157 * for an unlinked file which is still open.
5159 * Also note that the flusher is responsible for cleaning up empty
5163 hammer2_chain_delete(hammer2_chain_t *parent, hammer2_chain_t *chain,
5164 hammer2_tid_t mtid, int flags)
5168 KKASSERT(hammer2_mtx_owned(&chain->lock));
5171 * Nothing to do if already marked.
5173 * We need the spinlock on the core whos RBTREE contains chain
5174 * to protect against races.
5176 if ((chain->flags & HAMMER2_CHAIN_DELETED) == 0) {
5177 KKASSERT((chain->flags & HAMMER2_CHAIN_DELETED) == 0 &&
5178 chain->parent == parent);
5179 error = _hammer2_chain_delete_helper(parent, chain,
5184 * Permanent deletions mark the chain as destroyed.
5186 * NOTE: We do not setflush the chain unless the deletion is
5187 * permanent, since the deletion of a chain does not actually
5188 * require it to be flushed.
5191 if (flags & HAMMER2_DELETE_PERMANENT) {
5192 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DESTROY);
5193 hammer2_chain_setflush(chain);
5201 hammer2_chain_delete_obref(hammer2_chain_t *parent, hammer2_chain_t *chain,
5202 hammer2_tid_t mtid, int flags,
5203 hammer2_blockref_t *obref)
5207 KKASSERT(hammer2_mtx_owned(&chain->lock));
5210 * Nothing to do if already marked.
5212 * We need the spinlock on the core whos RBTREE contains chain
5213 * to protect against races.
5215 obref->type = HAMMER2_BREF_TYPE_EMPTY;
5216 if ((chain->flags & HAMMER2_CHAIN_DELETED) == 0) {
5217 KKASSERT((chain->flags & HAMMER2_CHAIN_DELETED) == 0 &&
5218 chain->parent == parent);
5219 error = _hammer2_chain_delete_helper(parent, chain,
5220 mtid, flags, obref);
5224 * Permanent deletions mark the chain as destroyed.
5226 * NOTE: We do not setflush the chain unless the deletion is
5227 * permanent, since the deletion of a chain does not actually
5228 * require it to be flushed.
5231 if (flags & HAMMER2_DELETE_PERMANENT) {
5232 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DESTROY);
5233 hammer2_chain_setflush(chain);
5241 * Returns the index of the nearest element in the blockref array >= elm.
5242 * Returns (count) if no element could be found.
5244 * Sets *key_nextp to the next key for loop purposes but does not modify
5245 * it if the next key would be higher than the current value of *key_nextp.
5246 * Note that *key_nexp can overflow to 0, which should be tested by the
5249 * WARNING! Must be called with parent's spinlock held. Spinlock remains
5250 * held through the operation.
5253 hammer2_base_find(hammer2_chain_t *parent,
5254 hammer2_blockref_t *base, int count,
5255 hammer2_key_t *key_nextp,
5256 hammer2_key_t key_beg, hammer2_key_t key_end)
5258 hammer2_blockref_t *scan;
5259 hammer2_key_t scan_end;
5264 * Require the live chain's already have their core's counted
5265 * so we can optimize operations.
5267 KKASSERT(parent->flags & HAMMER2_CHAIN_COUNTEDBREFS);
5272 if (count == 0 || base == NULL)
5276 * Sequential optimization using parent->cache_index. This is
5277 * the most likely scenario.
5279 * We can avoid trailing empty entries on live chains, otherwise
5280 * we might have to check the whole block array.
5282 i = parent->cache_index; /* SMP RACE OK */
5284 limit = parent->core.live_zero;
5289 KKASSERT(i < count);
5295 while (i > 0 && (scan->type == HAMMER2_BREF_TYPE_EMPTY ||
5296 scan->key > key_beg)) {
5300 parent->cache_index = i;
5303 * Search forwards, stop when we find a scan element which
5304 * encloses the key or until we know that there are no further
5308 if (scan->type != HAMMER2_BREF_TYPE_EMPTY) {
5309 scan_end = scan->key +
5310 ((hammer2_key_t)1 << scan->keybits) - 1;
5311 if (scan->key > key_beg || scan_end >= key_beg)
5320 parent->cache_index = i;
5324 scan_end = scan->key +
5325 ((hammer2_key_t)1 << scan->keybits);
5326 if (scan_end && (*key_nextp > scan_end ||
5328 *key_nextp = scan_end;
5336 * Do a combined search and return the next match either from the blockref
5337 * array or from the in-memory chain. Sets *bresp to the returned bref in
5338 * both cases, or sets it to NULL if the search exhausted. Only returns
5339 * a non-NULL chain if the search matched from the in-memory chain.
5341 * When no in-memory chain has been found and a non-NULL bref is returned
5345 * The returned chain is not locked or referenced. Use the returned bref
5346 * to determine if the search exhausted or not. Iterate if the base find
5347 * is chosen but matches a deleted chain.
5349 * WARNING! Must be called with parent's spinlock held. Spinlock remains
5350 * held through the operation.
5353 hammer2_combined_find(hammer2_chain_t *parent,
5354 hammer2_blockref_t *base, int count,
5355 hammer2_key_t *key_nextp,
5356 hammer2_key_t key_beg, hammer2_key_t key_end,
5357 hammer2_blockref_t **bresp)
5359 hammer2_blockref_t *bref;
5360 hammer2_chain_t *chain;
5364 * Lookup in block array and in rbtree.
5366 *key_nextp = key_end + 1;
5367 i = hammer2_base_find(parent, base, count, key_nextp,
5369 chain = hammer2_chain_find(parent, key_nextp, key_beg, key_end);
5374 if (i == count && chain == NULL) {
5380 * Only chain matched.
5383 bref = &chain->bref;
5388 * Only blockref matched.
5390 if (chain == NULL) {
5396 * Both in-memory and blockref matched, select the nearer element.
5398 * If both are flush with the left-hand side or both are the
5399 * same distance away, select the chain. In this situation the
5400 * chain must have been loaded from the matching blockmap.
5402 if ((chain->bref.key <= key_beg && base[i].key <= key_beg) ||
5403 chain->bref.key == base[i].key) {
5404 KKASSERT(chain->bref.key == base[i].key);
5405 bref = &chain->bref;
5410 * Select the nearer key
5412 if (chain->bref.key < base[i].key) {
5413 bref = &chain->bref;
5420 * If the bref is out of bounds we've exhausted our search.
5423 if (bref->key > key_end) {
5433 * Locate the specified block array element and delete it. The element
5436 * The spin lock on the related chain must be held.
5438 * NOTE: live_count was adjusted when the chain was deleted, so it does not
5439 * need to be adjusted when we commit the media change.
5442 hammer2_base_delete(hammer2_chain_t *parent,
5443 hammer2_blockref_t *base, int count,
5444 hammer2_chain_t *chain,
5445 hammer2_blockref_t *obref)
5447 hammer2_blockref_t *elm = &chain->bref;
5448 hammer2_blockref_t *scan;
5449 hammer2_key_t key_next;
5453 * Delete element. Expect the element to exist.
5455 * XXX see caller, flush code not yet sophisticated enough to prevent
5456 * re-flushed in some cases.
5458 key_next = 0; /* max range */
5459 i = hammer2_base_find(parent, base, count, &key_next,
5460 elm->key, elm->key);
5462 if (i == count || scan->type == HAMMER2_BREF_TYPE_EMPTY ||
5463 scan->key != elm->key ||
5464 ((chain->flags & HAMMER2_CHAIN_BMAPUPD) == 0 &&
5465 scan->keybits != elm->keybits)) {
5466 hammer2_spin_unex(&parent->core.spin);
5467 panic("delete base %p element not found at %d/%d elm %p\n",
5468 base, i, count, elm);
5473 * Update stats and zero the entry.
5475 * NOTE: Handle radix == 0 (0 bytes) case.
5477 if ((int)(scan->data_off & HAMMER2_OFF_MASK_RADIX)) {
5478 parent->bref.embed.stats.data_count -= (hammer2_off_t)1 <<
5479 (int)(scan->data_off & HAMMER2_OFF_MASK_RADIX);
5481 switch(scan->type) {
5482 case HAMMER2_BREF_TYPE_INODE:
5483 --parent->bref.embed.stats.inode_count;
5485 case HAMMER2_BREF_TYPE_DATA:
5486 if (parent->bref.leaf_count == HAMMER2_BLOCKREF_LEAF_MAX) {
5487 atomic_set_int(&chain->flags,
5488 HAMMER2_CHAIN_HINT_LEAF_COUNT);
5490 if (parent->bref.leaf_count)
5491 --parent->bref.leaf_count;
5494 case HAMMER2_BREF_TYPE_INDIRECT:
5495 if (scan->type != HAMMER2_BREF_TYPE_DATA) {
5496 parent->bref.embed.stats.data_count -=
5497 scan->embed.stats.data_count;
5498 parent->bref.embed.stats.inode_count -=
5499 scan->embed.stats.inode_count;
5501 if (scan->type == HAMMER2_BREF_TYPE_INODE)
5503 if (parent->bref.leaf_count == HAMMER2_BLOCKREF_LEAF_MAX) {
5504 atomic_set_int(&chain->flags,
5505 HAMMER2_CHAIN_HINT_LEAF_COUNT);
5507 if (parent->bref.leaf_count <= scan->leaf_count)
5508 parent->bref.leaf_count = 0;
5510 parent->bref.leaf_count -= scan->leaf_count;
5513 case HAMMER2_BREF_TYPE_DIRENT:
5514 if (parent->bref.leaf_count == HAMMER2_BLOCKREF_LEAF_MAX) {
5515 atomic_set_int(&chain->flags,
5516 HAMMER2_CHAIN_HINT_LEAF_COUNT);
5518 if (parent->bref.leaf_count)
5519 --parent->bref.leaf_count;
5527 bzero(scan, sizeof(*scan));
5530 * We can only optimize parent->core.live_zero for live chains.
5532 if (parent->core.live_zero == i + 1) {
5533 while (--i >= 0 && base[i].type == HAMMER2_BREF_TYPE_EMPTY)
5535 parent->core.live_zero = i + 1;
5539 * Clear appropriate blockmap flags in chain.
5541 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_BMAPPED |
5542 HAMMER2_CHAIN_BMAPUPD);
5546 * Insert the specified element. The block array must not already have the
5547 * element and must have space available for the insertion.
5549 * The spin lock on the related chain must be held.
5551 * NOTE: live_count was adjusted when the chain was deleted, so it does not
5552 * need to be adjusted when we commit the media change.
5555 hammer2_base_insert(hammer2_chain_t *parent,
5556 hammer2_blockref_t *base, int count,
5557 hammer2_chain_t *chain, hammer2_blockref_t *elm)
5559 hammer2_key_t key_next;
5568 * Insert new element. Expect the element to not already exist
5569 * unless we are replacing it.
5571 * XXX see caller, flush code not yet sophisticated enough to prevent
5572 * re-flushed in some cases.
5574 key_next = 0; /* max range */
5575 i = hammer2_base_find(parent, base, count, &key_next,
5576 elm->key, elm->key);
5579 * Shortcut fill optimization, typical ordered insertion(s) may not
5582 KKASSERT(i >= 0 && i <= count);
5585 * Set appropriate blockmap flags in chain (if not NULL)
5588 atomic_set_int(&chain->flags, HAMMER2_CHAIN_BMAPPED);
5591 * Update stats and zero the entry
5593 if ((int)(elm->data_off & HAMMER2_OFF_MASK_RADIX)) {
5594 parent->bref.embed.stats.data_count += (hammer2_off_t)1 <<
5595 (int)(elm->data_off & HAMMER2_OFF_MASK_RADIX);
5598 case HAMMER2_BREF_TYPE_INODE:
5599 ++parent->bref.embed.stats.inode_count;
5601 case HAMMER2_BREF_TYPE_DATA:
5602 if (parent->bref.leaf_count != HAMMER2_BLOCKREF_LEAF_MAX)
5603 ++parent->bref.leaf_count;
5605 case HAMMER2_BREF_TYPE_INDIRECT:
5606 if (elm->type != HAMMER2_BREF_TYPE_DATA) {
5607 parent->bref.embed.stats.data_count +=
5608 elm->embed.stats.data_count;
5609 parent->bref.embed.stats.inode_count +=
5610 elm->embed.stats.inode_count;
5612 if (elm->type == HAMMER2_BREF_TYPE_INODE)
5614 if (parent->bref.leaf_count + elm->leaf_count <
5615 HAMMER2_BLOCKREF_LEAF_MAX) {
5616 parent->bref.leaf_count += elm->leaf_count;
5618 parent->bref.leaf_count = HAMMER2_BLOCKREF_LEAF_MAX;
5621 case HAMMER2_BREF_TYPE_DIRENT:
5622 if (parent->bref.leaf_count != HAMMER2_BLOCKREF_LEAF_MAX)
5623 ++parent->bref.leaf_count;
5631 * We can only optimize parent->core.live_zero for live chains.
5633 if (i == count && parent->core.live_zero < count) {
5634 i = parent->core.live_zero++;
5639 xkey = elm->key + ((hammer2_key_t)1 << elm->keybits) - 1;
5640 if (i != count && (base[i].key < elm->key || xkey >= base[i].key)) {
5641 hammer2_spin_unex(&parent->core.spin);
5642 panic("insert base %p overlapping elements at %d elm %p\n",
5647 * Try to find an empty slot before or after.
5651 while (j > 0 || k < count) {
5653 if (j >= 0 && base[j].type == HAMMER2_BREF_TYPE_EMPTY) {
5657 bcopy(&base[j+1], &base[j],
5658 (i - j - 1) * sizeof(*base));
5664 if (k < count && base[k].type == HAMMER2_BREF_TYPE_EMPTY) {
5665 bcopy(&base[i], &base[i+1],
5666 (k - i) * sizeof(hammer2_blockref_t));
5670 * We can only update parent->core.live_zero for live
5673 if (parent->core.live_zero <= k)
5674 parent->core.live_zero = k + 1;
5679 panic("hammer2_base_insert: no room!");
5686 for (l = 0; l < count; ++l) {
5687 if (base[l].type != HAMMER2_BREF_TYPE_EMPTY) {
5688 key_next = base[l].key +
5689 ((hammer2_key_t)1 << base[l].keybits) - 1;
5693 while (++l < count) {
5694 if (base[l].type != HAMMER2_BREF_TYPE_EMPTY) {
5695 if (base[l].key <= key_next)
5696 panic("base_insert %d %d,%d,%d fail %p:%d", u, i, j, k, base, l);
5697 key_next = base[l].key +
5698 ((hammer2_key_t)1 << base[l].keybits) - 1;
5708 * Sort the blockref array for the chain. Used by the flush code to
5709 * sort the blockref[] array.
5711 * The chain must be exclusively locked AND spin-locked.
5713 typedef hammer2_blockref_t *hammer2_blockref_p;
5717 hammer2_base_sort_callback(const void *v1, const void *v2)
5719 hammer2_blockref_p bref1 = *(const hammer2_blockref_p *)v1;
5720 hammer2_blockref_p bref2 = *(const hammer2_blockref_p *)v2;
5723 * Make sure empty elements are placed at the end of the array
5725 if (bref1->type == HAMMER2_BREF_TYPE_EMPTY) {
5726 if (bref2->type == HAMMER2_BREF_TYPE_EMPTY)
5729 } else if (bref2->type == HAMMER2_BREF_TYPE_EMPTY) {
5736 if (bref1->key < bref2->key)
5738 if (bref1->key > bref2->key)
5744 hammer2_base_sort(hammer2_chain_t *chain)
5746 hammer2_blockref_t *base;
5749 switch(chain->bref.type) {
5750 case HAMMER2_BREF_TYPE_INODE:
5752 * Special shortcut for embedded data returns the inode
5753 * itself. Callers must detect this condition and access
5754 * the embedded data (the strategy code does this for us).
5756 * This is only applicable to regular files and softlinks.
5758 if (chain->data->ipdata.meta.op_flags &
5759 HAMMER2_OPFLAG_DIRECTDATA) {
5762 base = &chain->data->ipdata.u.blockset.blockref[0];
5763 count = HAMMER2_SET_COUNT;
5765 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
5766 case HAMMER2_BREF_TYPE_INDIRECT:
5768 * Optimize indirect blocks in the INITIAL state to avoid
5771 KKASSERT((chain->flags & HAMMER2_CHAIN_INITIAL) == 0);
5772 base = &chain->data->npdata[0];
5773 count = chain->bytes / sizeof(hammer2_blockref_t);
5775 case HAMMER2_BREF_TYPE_VOLUME:
5776 base = &chain->data->voldata.sroot_blockset.blockref[0];
5777 count = HAMMER2_SET_COUNT;
5779 case HAMMER2_BREF_TYPE_FREEMAP:
5780 base = &chain->data->blkset.blockref[0];
5781 count = HAMMER2_SET_COUNT;
5784 kprintf("hammer2_chain_lookup: unrecognized "
5785 "blockref(A) type: %d",
5788 tsleep(&base, 0, "dead", 0);
5789 panic("hammer2_base_sort: unrecognized "
5790 "blockref(A) type: %d",
5792 base = NULL; /* safety */
5793 count = 0; /* safety */
5795 kqsort(base, count, sizeof(*base), hammer2_base_sort_callback);
5801 * Chain memory management
5804 hammer2_chain_wait(hammer2_chain_t *chain)
5806 tsleep(chain, 0, "chnflw", 1);
5809 const hammer2_media_data_t *
5810 hammer2_chain_rdata(hammer2_chain_t *chain)
5812 KKASSERT(chain->data != NULL);
5813 return (chain->data);
5816 hammer2_media_data_t *
5817 hammer2_chain_wdata(hammer2_chain_t *chain)
5819 KKASSERT(chain->data != NULL);
5820 return (chain->data);
5824 * Set the check data for a chain. This can be a heavy-weight operation
5825 * and typically only runs on-flush. For file data check data is calculated
5826 * when the logical buffers are flushed.
5829 hammer2_chain_setcheck(hammer2_chain_t *chain, void *bdata)
5831 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_NOTTESTED);
5833 switch(HAMMER2_DEC_CHECK(chain->bref.methods)) {
5834 case HAMMER2_CHECK_NONE:
5836 case HAMMER2_CHECK_DISABLED:
5838 case HAMMER2_CHECK_ISCSI32:
5839 chain->bref.check.iscsi32.value =
5840 hammer2_icrc32(bdata, chain->bytes);
5842 case HAMMER2_CHECK_XXHASH64:
5843 chain->bref.check.xxhash64.value =
5844 XXH64(bdata, chain->bytes, XXH_HAMMER2_SEED);
5846 case HAMMER2_CHECK_SHA192:
5848 SHA256_CTX hash_ctx;
5850 uint8_t digest[SHA256_DIGEST_LENGTH];
5851 uint64_t digest64[SHA256_DIGEST_LENGTH/8];
5854 SHA256_Init(&hash_ctx);
5855 SHA256_Update(&hash_ctx, bdata, chain->bytes);
5856 SHA256_Final(u.digest, &hash_ctx);
5857 u.digest64[2] ^= u.digest64[3];
5859 chain->bref.check.sha192.data,
5860 sizeof(chain->bref.check.sha192.data));
5863 case HAMMER2_CHECK_FREEMAP:
5864 chain->bref.check.freemap.icrc32 =
5865 hammer2_icrc32(bdata, chain->bytes);
5868 kprintf("hammer2_chain_setcheck: unknown check type %02x\n",
5869 chain->bref.methods);
5875 * Characterize a failed check code and try to trace back to the inode.
5878 hammer2_characterize_failed_chain(hammer2_chain_t *chain, uint64_t check,
5881 hammer2_chain_t *lchain;
5882 hammer2_chain_t *ochain;
5885 did = krateprintf(&krate_h2chk,
5886 "chain %016jx.%02x (%s) meth=%02x CHECK FAIL "
5887 "(flags=%08x, bref/data ",
5888 chain->bref.data_off,
5890 hammer2_bref_type_str(&chain->bref),
5891 chain->bref.methods,
5897 kprintf("%08x/%08x)\n",
5898 chain->bref.check.iscsi32.value,
5901 kprintf("%016jx/%016jx)\n",
5902 chain->bref.check.xxhash64.value,
5907 * Run up the chains to try to find the governing inode so we
5910 * XXX This error reporting is not really MPSAFE
5914 while (chain && chain->bref.type != HAMMER2_BREF_TYPE_INODE) {
5916 chain = chain->parent;
5919 if (chain && chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
5920 ((chain->bref.flags & HAMMER2_BREF_FLAG_PFSROOT) == 0 ||
5921 (lchain->bref.key & HAMMER2_DIRHASH_VISIBLE))) {
5922 kprintf(" Resides at/in inode %ld\n",
5924 } else if (chain && chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
5925 kprintf(" Resides in inode index - CRITICAL!!!\n");
5927 kprintf(" Resides in root index - CRITICAL!!!\n");
5930 const char *pfsname = "UNKNOWN";
5934 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
5935 if (ochain->pmp->pfs_hmps[i] == ochain->hmp &&
5936 ochain->pmp->pfs_names[i]) {
5937 pfsname = ochain->pmp->pfs_names[i];
5942 kprintf(" In pfs %s on device %s\n",
5943 pfsname, ochain->hmp->devrepname);
5948 * Returns non-zero on success, 0 on failure.
5951 hammer2_chain_testcheck(hammer2_chain_t *chain, void *bdata)
5957 if (chain->flags & HAMMER2_CHAIN_NOTTESTED)
5960 switch(HAMMER2_DEC_CHECK(chain->bref.methods)) {
5961 case HAMMER2_CHECK_NONE:
5964 case HAMMER2_CHECK_DISABLED:
5967 case HAMMER2_CHECK_ISCSI32:
5968 check32 = hammer2_icrc32(bdata, chain->bytes);
5969 r = (chain->bref.check.iscsi32.value == check32);
5971 hammer2_characterize_failed_chain(chain, check32, 32);
5973 hammer2_process_icrc32 += chain->bytes;
5975 case HAMMER2_CHECK_XXHASH64:
5976 check64 = XXH64(bdata, chain->bytes, XXH_HAMMER2_SEED);
5977 r = (chain->bref.check.xxhash64.value == check64);
5979 hammer2_characterize_failed_chain(chain, check64, 64);
5981 hammer2_process_xxhash64 += chain->bytes;
5983 case HAMMER2_CHECK_SHA192:
5985 SHA256_CTX hash_ctx;
5987 uint8_t digest[SHA256_DIGEST_LENGTH];
5988 uint64_t digest64[SHA256_DIGEST_LENGTH/8];
5991 SHA256_Init(&hash_ctx);
5992 SHA256_Update(&hash_ctx, bdata, chain->bytes);
5993 SHA256_Final(u.digest, &hash_ctx);
5994 u.digest64[2] ^= u.digest64[3];
5996 chain->bref.check.sha192.data,
5997 sizeof(chain->bref.check.sha192.data)) == 0) {
6001 krateprintf(&krate_h2chk,
6002 "chain %016jx.%02x meth=%02x "
6004 chain->bref.data_off,
6006 chain->bref.methods);
6010 case HAMMER2_CHECK_FREEMAP:
6011 r = (chain->bref.check.freemap.icrc32 ==
6012 hammer2_icrc32(bdata, chain->bytes));
6016 did = krateprintf(&krate_h2chk,
6017 "chain %016jx.%02x meth=%02x "
6019 chain->bref.data_off,
6021 chain->bref.methods);
6023 kprintf("freemap.icrc %08x icrc32 %08x (%d)\n",
6024 chain->bref.check.freemap.icrc32,
6025 hammer2_icrc32(bdata, chain->bytes),
6028 kprintf("dio %p buf %016jx,%d "
6031 chain->dio->bp->b_loffset,
6032 chain->dio->bp->b_bufsize,
6034 chain->dio->bp->b_data);
6040 kprintf("hammer2_chain_setcheck: unknown check type %02x\n",
6041 chain->bref.methods);
6049 * Acquire the chain and parent representing the specified inode for the
6050 * device at the specified cluster index.
6052 * The flags passed in are LOOKUP flags, not RESOLVE flags.
6054 * If we are unable to locate the inode, HAMMER2_ERROR_EIO is returned and
6055 * *chainp will be NULL. *parentp may still be set error or not, or NULL
6056 * if the parent itself could not be resolved.
6058 * The caller may pass-in a locked *parentp and/or *chainp, or neither.
6059 * They will be unlocked and released by this function. The *parentp and
6060 * *chainp representing the located inode are returned locked.
6063 hammer2_chain_inode_find(hammer2_pfs_t *pmp, hammer2_key_t inum,
6064 int clindex, int flags,
6065 hammer2_chain_t **parentp, hammer2_chain_t **chainp)
6067 hammer2_chain_t *parent;
6068 hammer2_chain_t *rchain;
6069 hammer2_key_t key_dummy;
6070 hammer2_inode_t *ip;
6074 resolve_flags = (flags & HAMMER2_LOOKUP_SHARED) ?
6075 HAMMER2_RESOLVE_SHARED : 0;
6078 * Caller expects us to replace these.
6081 hammer2_chain_unlock(*chainp);
6082 hammer2_chain_drop(*chainp);
6086 hammer2_chain_unlock(*parentp);
6087 hammer2_chain_drop(*parentp);
6092 * Be very careful, this is a backend function and we CANNOT
6093 * lock any frontend inode structure we find. But we have to
6094 * look the inode up this way first in case it exists but is
6095 * detached from the radix tree.
6097 ip = hammer2_inode_lookup(pmp, inum);
6099 *chainp = hammer2_inode_chain_and_parent(ip, clindex,
6102 hammer2_inode_drop(ip);
6105 hammer2_chain_unlock(*chainp);
6106 hammer2_chain_drop(*chainp);
6109 hammer2_chain_unlock(*parentp);
6110 hammer2_chain_drop(*parentp);
6116 * Inodes hang off of the iroot (bit 63 is clear, differentiating
6117 * inodes from root directory entries in the key lookup).
6119 parent = hammer2_inode_chain(pmp->iroot, clindex, resolve_flags);
6122 rchain = hammer2_chain_lookup(&parent, &key_dummy,
6126 error = HAMMER2_ERROR_EIO;
6135 * Used by the bulkscan code to snapshot the synchronized storage for
6136 * a volume, allowing it to be scanned concurrently against normal
6140 hammer2_chain_bulksnap(hammer2_dev_t *hmp)
6142 hammer2_chain_t *copy;
6144 copy = hammer2_chain_alloc(hmp, hmp->spmp, &hmp->vchain.bref);
6145 copy->data = kmalloc(sizeof(copy->data->voldata),
6148 hammer2_voldata_lock(hmp);
6149 copy->data->voldata = hmp->volsync;
6150 hammer2_voldata_unlock(hmp);
6156 hammer2_chain_bulkdrop(hammer2_chain_t *copy)
6158 KKASSERT(copy->bref.type == HAMMER2_BREF_TYPE_VOLUME);
6159 KKASSERT(copy->data);
6160 kfree(copy->data, copy->hmp->mchain);
6162 atomic_add_long(&hammer2_chain_allocs, -1);
6163 hammer2_chain_drop(copy);
6167 * Returns non-zero if the chain (INODE or DIRENT) matches the
6171 hammer2_chain_dirent_test(hammer2_chain_t *chain, const char *name,
6174 const hammer2_inode_data_t *ripdata;
6175 const hammer2_dirent_head_t *den;
6177 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
6178 ripdata = &chain->data->ipdata;
6179 if (ripdata->meta.name_len == name_len &&
6180 bcmp(ripdata->filename, name, name_len) == 0) {
6184 if (chain->bref.type == HAMMER2_BREF_TYPE_DIRENT &&
6185 chain->bref.embed.dirent.namlen == name_len) {
6186 den = &chain->bref.embed.dirent;
6187 if (name_len > sizeof(chain->bref.check.buf) &&
6188 bcmp(chain->data->buf, name, name_len) == 0) {
6191 if (name_len <= sizeof(chain->bref.check.buf) &&
6192 bcmp(chain->bref.check.buf, name, name_len) == 0) {