2 * Copyright (c) 2011-2018 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@dragonflybsd.org>
6 * and Venkatesh Srinivas <vsrinivas@dragonflybsd.org>
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in
16 * the documentation and/or other materials provided with the
18 * 3. Neither the name of The DragonFly Project nor the names of its
19 * contributors may be used to endorse or promote products derived
20 * from this software without specific, prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
23 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
24 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
25 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
26 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
27 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
28 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
29 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
30 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
31 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
32 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * This subsystem implements most of the core support functions for
37 * the hammer2_chain structure.
39 * Chains are the in-memory version on media objects (volume header, inodes,
40 * indirect blocks, data blocks, etc). Chains represent a portion of the
43 * Chains are no-longer delete-duplicated. Instead, the original in-memory
44 * chain will be moved along with its block reference (e.g. for things like
45 * renames, hardlink operations, modifications, etc), and will be indexed
46 * on a secondary list for flush handling instead of propagating a flag
49 * Concurrent front-end operations can still run against backend flushes
50 * as long as they do not cross the current flush boundary. An operation
51 * running above the current flush (in areas not yet flushed) can become
52 * part of the current flush while ano peration running below the current
53 * flush can become part of the next flush.
55 #include <sys/cdefs.h>
56 #include <sys/param.h>
57 #include <sys/systm.h>
58 #include <sys/types.h>
60 #include <sys/kern_syscall.h>
63 #include <crypto/sha2/sha2.h>
67 static hammer2_chain_t
*hammer2_chain_create_indirect(
68 hammer2_chain_t
*parent
,
69 hammer2_key_t key
, int keybits
,
70 hammer2_tid_t mtid
, int for_type
, int *errorp
);
71 static hammer2_io_t
*hammer2_chain_drop_data(hammer2_chain_t
*chain
);
72 static hammer2_chain_t
*hammer2_combined_find(
73 hammer2_chain_t
*parent
,
74 hammer2_blockref_t
*base
, int count
,
75 hammer2_key_t
*key_nextp
,
76 hammer2_key_t key_beg
, hammer2_key_t key_end
,
77 hammer2_blockref_t
**bresp
);
79 static struct krate krate_h2me
= { .freq
= 1 };
82 * Basic RBTree for chains (core->rbtree and core->dbtree). Chains cannot
83 * overlap in the RB trees. Deleted chains are moved from rbtree to either
86 * Chains in delete-duplicate sequences can always iterate through core_entry
87 * to locate the live version of the chain.
89 RB_GENERATE(hammer2_chain_tree
, hammer2_chain
, rbnode
, hammer2_chain_cmp
);
92 hammer2_chain_cmp(hammer2_chain_t
*chain1
, hammer2_chain_t
*chain2
)
100 * Compare chains. Overlaps are not supposed to happen and catch
101 * any software issues early we count overlaps as a match.
103 c1_beg
= chain1
->bref
.key
;
104 c1_end
= c1_beg
+ ((hammer2_key_t
)1 << chain1
->bref
.keybits
) - 1;
105 c2_beg
= chain2
->bref
.key
;
106 c2_end
= c2_beg
+ ((hammer2_key_t
)1 << chain2
->bref
.keybits
) - 1;
108 if (c1_end
< c2_beg
) /* fully to the left */
110 if (c1_beg
> c2_end
) /* fully to the right */
112 return(0); /* overlap (must not cross edge boundary) */
116 * Assert that a chain has no media data associated with it.
119 hammer2_chain_assert_no_data(hammer2_chain_t
*chain
)
121 KKASSERT(chain
->dio
== NULL
);
122 if (chain
->bref
.type
!= HAMMER2_BREF_TYPE_VOLUME
&&
123 chain
->bref
.type
!= HAMMER2_BREF_TYPE_FREEMAP
&&
125 panic("hammer2_assert_no_data: chain %p still has data", chain
);
130 * Make a chain visible to the flusher. The flusher needs to be able to
131 * do flushes of subdirectory chains or single files so it does a top-down
132 * recursion using the ONFLUSH flag for the recursion. It locates MODIFIED
133 * or UPDATE chains and flushes back up the chain to the volume root.
135 * This routine sets ONFLUSH upward until it hits the volume root. For
136 * simplicity we ignore PFSROOT boundaries whos rules can be complex.
137 * Extra ONFLUSH flagging doesn't hurt the filesystem.
140 hammer2_chain_setflush(hammer2_chain_t
*chain
)
142 hammer2_chain_t
*parent
;
144 if ((chain
->flags
& HAMMER2_CHAIN_ONFLUSH
) == 0) {
145 hammer2_spin_sh(&chain
->core
.spin
);
146 while ((chain
->flags
& HAMMER2_CHAIN_ONFLUSH
) == 0) {
147 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_ONFLUSH
);
148 if ((parent
= chain
->parent
) == NULL
)
150 hammer2_spin_sh(&parent
->core
.spin
);
151 hammer2_spin_unsh(&chain
->core
.spin
);
154 hammer2_spin_unsh(&chain
->core
.spin
);
159 * Allocate a new disconnected chain element representing the specified
160 * bref. chain->refs is set to 1 and the passed bref is copied to
161 * chain->bref. chain->bytes is derived from the bref.
163 * chain->pmp inherits pmp unless the chain is an inode (other than the
166 * NOTE: Returns a referenced but unlocked (because there is no core) chain.
169 hammer2_chain_alloc(hammer2_dev_t
*hmp
, hammer2_pfs_t
*pmp
,
170 hammer2_blockref_t
*bref
)
172 hammer2_chain_t
*chain
;
176 * Special case - radix of 0 indicates a chain that does not
177 * need a data reference (context is completely embedded in the
180 if ((int)(bref
->data_off
& HAMMER2_OFF_MASK_RADIX
))
181 bytes
= 1U << (int)(bref
->data_off
& HAMMER2_OFF_MASK_RADIX
);
185 atomic_add_long(&hammer2_chain_allocs
, 1);
188 * Construct the appropriate system structure.
191 case HAMMER2_BREF_TYPE_DIRENT
:
192 case HAMMER2_BREF_TYPE_INODE
:
193 case HAMMER2_BREF_TYPE_INDIRECT
:
194 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
195 case HAMMER2_BREF_TYPE_DATA
:
196 case HAMMER2_BREF_TYPE_FREEMAP_LEAF
:
198 * Chain's are really only associated with the hmp but we
199 * maintain a pmp association for per-mount memory tracking
200 * purposes. The pmp can be NULL.
202 chain
= kmalloc(sizeof(*chain
), hmp
->mchain
, M_WAITOK
| M_ZERO
);
204 case HAMMER2_BREF_TYPE_VOLUME
:
205 case HAMMER2_BREF_TYPE_FREEMAP
:
207 * Only hammer2_chain_bulksnap() calls this function with these
210 chain
= kmalloc(sizeof(*chain
), hmp
->mchain
, M_WAITOK
| M_ZERO
);
214 panic("hammer2_chain_alloc: unrecognized blockref type: %d",
219 * Initialize the new chain structure. pmp must be set to NULL for
220 * chains belonging to the super-root topology of a device mount.
222 if (pmp
== hmp
->spmp
)
229 chain
->bytes
= bytes
;
231 chain
->flags
= HAMMER2_CHAIN_ALLOCATED
;
232 lockinit(&chain
->diolk
, "chdio", 0, 0);
235 * Set the PFS boundary flag if this chain represents a PFS root.
237 if (bref
->flags
& HAMMER2_BREF_FLAG_PFSROOT
)
238 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_PFSBOUNDARY
);
239 hammer2_chain_core_init(chain
);
245 * Initialize a chain's core structure. This structure used to be allocated
246 * but is now embedded.
248 * The core is not locked. No additional refs on the chain are made.
249 * (trans) must not be NULL if (core) is not NULL.
252 hammer2_chain_core_init(hammer2_chain_t
*chain
)
255 * Fresh core under nchain (no multi-homing of ochain's
258 RB_INIT(&chain
->core
.rbtree
); /* live chains */
259 hammer2_mtx_init(&chain
->lock
, "h2chain");
263 * Add a reference to a chain element, preventing its destruction.
265 * (can be called with spinlock held)
268 hammer2_chain_ref(hammer2_chain_t
*chain
)
270 if (atomic_fetchadd_int(&chain
->refs
, 1) == 0) {
272 * Just flag that the chain was used and should be recycled
273 * on the LRU if it encounters it later.
275 if (chain
->flags
& HAMMER2_CHAIN_ONLRU
)
276 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_LRUHINT
);
280 * REMOVED - reduces contention, lru_list is more heuristical
283 * 0->non-zero transition must ensure that chain is removed
286 * NOTE: Already holding lru_spin here so we cannot call
287 * hammer2_chain_ref() to get it off lru_list, do
290 if (chain
->flags
& HAMMER2_CHAIN_ONLRU
) {
291 hammer2_pfs_t
*pmp
= chain
->pmp
;
292 hammer2_spin_ex(&pmp
->lru_spin
);
293 if (chain
->flags
& HAMMER2_CHAIN_ONLRU
) {
294 atomic_add_int(&pmp
->lru_count
, -1);
295 atomic_clear_int(&chain
->flags
,
296 HAMMER2_CHAIN_ONLRU
);
297 TAILQ_REMOVE(&pmp
->lru_list
, chain
, lru_node
);
299 hammer2_spin_unex(&pmp
->lru_spin
);
306 * Ref a locked chain and force the data to be held across an unlock.
307 * Chain must be currently locked. The user of the chain who desires
308 * to release the hold must call hammer2_chain_lock_unhold() to relock
309 * and unhold the chain, then unlock normally, or may simply call
310 * hammer2_chain_drop_unhold() (which is safer against deadlocks).
313 hammer2_chain_ref_hold(hammer2_chain_t
*chain
)
315 atomic_add_int(&chain
->lockcnt
, 1);
316 hammer2_chain_ref(chain
);
320 * Insert the chain in the core rbtree.
322 * Normal insertions are placed in the live rbtree. Insertion of a deleted
323 * chain is a special case used by the flush code that is placed on the
324 * unstaged deleted list to avoid confusing the live view.
326 #define HAMMER2_CHAIN_INSERT_SPIN 0x0001
327 #define HAMMER2_CHAIN_INSERT_LIVE 0x0002
328 #define HAMMER2_CHAIN_INSERT_RACE 0x0004
332 hammer2_chain_insert(hammer2_chain_t
*parent
, hammer2_chain_t
*chain
,
333 int flags
, int generation
)
335 hammer2_chain_t
*xchain
;
338 if (flags
& HAMMER2_CHAIN_INSERT_SPIN
)
339 hammer2_spin_ex(&parent
->core
.spin
);
342 * Interlocked by spinlock, check for race
344 if ((flags
& HAMMER2_CHAIN_INSERT_RACE
) &&
345 parent
->core
.generation
!= generation
) {
346 error
= HAMMER2_ERROR_EAGAIN
;
353 xchain
= RB_INSERT(hammer2_chain_tree
, &parent
->core
.rbtree
, chain
);
354 KASSERT(xchain
== NULL
,
355 ("hammer2_chain_insert: collision %p %p (key=%016jx)",
356 chain
, xchain
, chain
->bref
.key
));
357 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_ONRBTREE
);
358 chain
->parent
= parent
;
359 ++parent
->core
.chain_count
;
360 ++parent
->core
.generation
; /* XXX incs for _get() too, XXX */
363 * We have to keep track of the effective live-view blockref count
364 * so the create code knows when to push an indirect block.
366 if (flags
& HAMMER2_CHAIN_INSERT_LIVE
)
367 atomic_add_int(&parent
->core
.live_count
, 1);
369 if (flags
& HAMMER2_CHAIN_INSERT_SPIN
)
370 hammer2_spin_unex(&parent
->core
.spin
);
375 * Drop the caller's reference to the chain. When the ref count drops to
376 * zero this function will try to disassociate the chain from its parent and
377 * deallocate it, then recursely drop the parent using the implied ref
378 * from the chain's chain->parent.
380 * Nobody should own chain's mutex on the 1->0 transition, unless this drop
381 * races an acquisition by another cpu. Therefore we can loop if we are
382 * unable to acquire the mutex, and refs is unlikely to be 1 unless we again
383 * race against another drop.
385 static hammer2_chain_t
*hammer2_chain_lastdrop(hammer2_chain_t
*chain
,
387 static void hammer2_chain_lru_flush(hammer2_pfs_t
*pmp
);
390 hammer2_chain_drop(hammer2_chain_t
*chain
)
394 if (hammer2_debug
& 0x200000)
397 KKASSERT(chain
->refs
> 0);
405 if (hammer2_mtx_ex_try(&chain
->lock
) == 0)
406 chain
= hammer2_chain_lastdrop(chain
, 0);
407 /* retry the same chain, or chain from lastdrop */
409 if (atomic_cmpset_int(&chain
->refs
, refs
, refs
- 1))
411 /* retry the same chain */
418 * Unhold a held and probably not-locked chain, ensure that the data is
419 * dropped on the 1->0 transition of lockcnt by obtaining an exclusive
420 * lock and then simply unlocking the chain.
423 hammer2_chain_drop_unhold(hammer2_chain_t
*chain
)
429 lockcnt
= chain
->lockcnt
;
432 if (atomic_cmpset_int(&chain
->lockcnt
,
433 lockcnt
, lockcnt
- 1)) {
436 } else if (hammer2_mtx_ex_try(&chain
->lock
) == 0) {
437 hammer2_chain_unlock(chain
);
441 * This situation can easily occur on SMP due to
442 * the gap inbetween the 1->0 transition and the
443 * final unlock. We cannot safely block on the
444 * mutex because lockcnt might go above 1.
446 * XXX Sleep for one tick if it takes too long.
449 if (iter
> 1000 + hz
) {
450 kprintf("hammer2: h2race1 %p\n", chain
);
453 tsleep(&iter
, 0, "h2race1", 1);
458 hammer2_chain_drop(chain
);
462 * Handles the (potential) last drop of chain->refs from 1->0. Called with
463 * the mutex exclusively locked, refs == 1, and lockcnt 0. SMP races are
464 * possible against refs and lockcnt. We must dispose of the mutex on chain.
466 * This function returns an unlocked chain for recursive drop or NULL. It
467 * can return the same chain if it determines it has raced another ref.
471 * When two chains need to be recursively dropped we use the chain we
472 * would otherwise free to placehold the additional chain. It's a bit
473 * convoluted but we can't just recurse without potentially blowing out
476 * The chain cannot be freed if it has any children.
477 * The chain cannot be freed if flagged MODIFIED unless we can dispose of it.
478 * The chain cannot be freed if flagged UPDATE unless we can dispose of it.
479 * Any dedup registration can remain intact.
481 * The core spinlock is allowed to nest child-to-parent (not parent-to-child).
485 hammer2_chain_lastdrop(hammer2_chain_t
*chain
, int depth
)
489 hammer2_chain_t
*parent
;
490 hammer2_chain_t
*rdrop
;
497 * On last drop if there is no parent and data_off is good (at
498 * least does not represent the volume root), the modified chain
499 * is probably going to be destroyed. We have to make sure that
500 * the data area is not registered for dedup.
502 * XXX removed. In fact, we do not have to make sure that the
503 * data area is not registered for dedup. The data area
504 * can, in fact, still be used for dedup because it is
505 * still allocated in the freemap and the underlying I/O
506 * will still be flushed.
508 if (chain
->parent
== NULL
&&
509 (chain
->flags
& HAMMER2_CHAIN_MODIFIED
) &&
510 (chain
->bref
.data_off
& ~HAMMER2_OFF_MASK_RADIX
)) {
512 hammer2_io_dedup_delete(hmp
, chain
->bref
.type
,
513 chain
->bref
.data_off
, chain
->bytes
);
517 * We need chain's spinlock to interlock the sub-tree test.
518 * We already have chain's mutex, protecting chain->parent.
520 * Remember that chain->refs can be in flux.
522 hammer2_spin_ex(&chain
->core
.spin
);
524 if (chain
->parent
!= NULL
) {
526 * If the chain has a parent the UPDATE bit prevents scrapping
527 * as the chain is needed to properly flush the parent. Try
528 * to complete the 1->0 transition and return NULL. Retry
529 * (return chain) if we are unable to complete the 1->0
530 * transition, else return NULL (nothing more to do).
532 * If the chain has a parent the MODIFIED bit prevents
535 * Chains with UPDATE/MODIFIED are *not* put on the LRU list!
537 if (chain
->flags
& (HAMMER2_CHAIN_UPDATE
|
538 HAMMER2_CHAIN_MODIFIED
)) {
539 if (atomic_cmpset_int(&chain
->refs
, 1, 0)) {
540 hammer2_spin_unex(&chain
->core
.spin
);
542 dio
= hammer2_chain_drop_data(chain
, 0);
544 hammer2_io_bqrelse(&dio
);
546 hammer2_chain_assert_no_data(chain
);
547 hammer2_mtx_unlock(&chain
->lock
);
550 hammer2_spin_unex(&chain
->core
.spin
);
551 hammer2_mtx_unlock(&chain
->lock
);
555 /* spinlock still held */
556 } else if (chain
->bref
.type
== HAMMER2_BREF_TYPE_VOLUME
||
557 chain
->bref
.type
== HAMMER2_BREF_TYPE_FREEMAP
) {
559 * Retain the static vchain and fchain. Clear bits that
560 * are not relevant. Do not clear the MODIFIED bit,
561 * and certainly do not put it on the delayed-flush queue.
563 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_UPDATE
);
566 * The chain has no parent and can be flagged for destruction.
567 * Since it has no parent, UPDATE can also be cleared.
569 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_DESTROY
);
570 if (chain
->flags
& HAMMER2_CHAIN_UPDATE
)
571 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_UPDATE
);
574 * If the chain has children we must still flush the chain.
575 * Any dedup is already handled by the underlying DIO, so
576 * we do not have to specifically flush it here.
578 * In the case where it has children, the DESTROY flag test
579 * in the flush code will prevent unnecessary flushes of
580 * MODIFIED chains that are not flagged DEDUP so don't worry
583 if (chain
->core
.chain_count
) {
585 * Put on flushq (should ensure refs > 1), retry
588 hammer2_spin_unex(&chain
->core
.spin
);
589 hammer2_delayed_flush(chain
);
590 hammer2_mtx_unlock(&chain
->lock
);
592 return(chain
); /* retry drop */
596 * Otherwise we can scrap the MODIFIED bit if it is set,
597 * and continue along the freeing path.
599 * Be sure to clean-out any dedup bits. Without a parent
600 * this chain will no longer be visible to the flush code.
601 * Easy check data_off to avoid the volume root.
603 if (chain
->flags
& HAMMER2_CHAIN_MODIFIED
) {
604 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_MODIFIED
);
605 atomic_add_long(&hammer2_count_modified_chains
, -1);
607 hammer2_pfs_memory_wakeup(chain
->pmp
);
609 /* spinlock still held */
612 /* spinlock still held */
618 * If any children exist we must leave the chain intact with refs == 0.
619 * They exist because chains are retained below us which have refs or
620 * may require flushing.
622 * Retry (return chain) if we fail to transition the refs to 0, else
623 * return NULL indication nothing more to do.
625 * Chains with children are NOT put on the LRU list.
627 if (chain
->core
.chain_count
) {
628 if (atomic_cmpset_int(&chain
->refs
, 1, 0)) {
629 hammer2_spin_unex(&chain
->core
.spin
);
630 hammer2_chain_assert_no_data(chain
);
631 hammer2_mtx_unlock(&chain
->lock
);
634 hammer2_spin_unex(&chain
->core
.spin
);
635 hammer2_mtx_unlock(&chain
->lock
);
639 /* spinlock still held */
640 /* no chains left under us */
643 * chain->core has no children left so no accessors can get to our
644 * chain from there. Now we have to lock the parent core to interlock
645 * remaining possible accessors that might bump chain's refs before
646 * we can safely drop chain's refs with intent to free the chain.
649 pmp
= chain
->pmp
; /* can be NULL */
652 parent
= chain
->parent
;
655 * WARNING! chain's spin lock is still held here, and other spinlocks
656 * will be acquired and released in the code below. We
657 * cannot be making fancy procedure calls!
661 * We can cache the chain if it is associated with a pmp
662 * and not flagged as being destroyed or requesting a full
663 * release. In this situation the chain is not removed
664 * from its parent, i.e. it can still be looked up.
666 * We intentionally do not cache DATA chains because these
667 * were likely used to load data into the logical buffer cache
668 * and will not be accessed again for some time.
671 (HAMMER2_CHAIN_DESTROY
| HAMMER2_CHAIN_RELEASE
)) == 0 &&
673 chain
->bref
.type
!= HAMMER2_BREF_TYPE_DATA
) {
675 hammer2_spin_ex(&parent
->core
.spin
);
676 if (atomic_cmpset_int(&chain
->refs
, 1, 0) == 0) {
678 * 1->0 transition failed, retry. Do not drop
679 * the chain's data yet!
682 hammer2_spin_unex(&parent
->core
.spin
);
683 hammer2_spin_unex(&chain
->core
.spin
);
684 hammer2_mtx_unlock(&chain
->lock
);
693 dio
= hammer2_chain_drop_data(chain
, 1);
695 hammer2_chain_assert_no_data(chain
);
698 * Make sure we are on the LRU list, clean up excessive
699 * LRU entries. We can only really drop one but there might
700 * be other entries that we can remove from the lru_list
703 * NOTE: HAMMER2_CHAIN_ONLRU may only be safely set when
704 * chain->core.spin AND pmp->lru_spin are held, but
705 * can be safely cleared only holding pmp->lru_spin.
707 if ((chain
->flags
& HAMMER2_CHAIN_ONLRU
) == 0) {
708 hammer2_spin_ex(&pmp
->lru_spin
);
709 if ((chain
->flags
& HAMMER2_CHAIN_ONLRU
) == 0) {
710 atomic_set_int(&chain
->flags
,
711 HAMMER2_CHAIN_ONLRU
);
712 TAILQ_INSERT_TAIL(&pmp
->lru_list
,
714 atomic_add_int(&pmp
->lru_count
, 1);
716 if (pmp
->lru_count
< HAMMER2_LRU_LIMIT
)
717 depth
= 1; /* disable lru_list flush */
718 hammer2_spin_unex(&pmp
->lru_spin
);
720 /* disable lru flush */
725 hammer2_spin_unex(&parent
->core
.spin
);
726 parent
= NULL
; /* safety */
728 hammer2_spin_unex(&chain
->core
.spin
);
729 hammer2_mtx_unlock(&chain
->lock
);
732 hammer2_io_bqrelse(&dio
);
736 * lru_list hysteresis (see above for depth overrides).
737 * Note that depth also prevents excessive lastdrop recursion.
740 hammer2_chain_lru_flush(pmp
);
747 * Make sure we are not on the LRU list.
749 if (chain
->flags
& HAMMER2_CHAIN_ONLRU
) {
750 hammer2_spin_ex(&pmp
->lru_spin
);
751 if (chain
->flags
& HAMMER2_CHAIN_ONLRU
) {
752 atomic_add_int(&pmp
->lru_count
, -1);
753 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_ONLRU
);
754 TAILQ_REMOVE(&pmp
->lru_list
, chain
, lru_node
);
756 hammer2_spin_unex(&pmp
->lru_spin
);
760 * Spinlock the parent and try to drop the last ref on chain.
761 * On success determine if we should dispose of the chain
762 * (remove the chain from its parent, etc).
764 * (normal core locks are top-down recursive but we define
765 * core spinlocks as bottom-up recursive, so this is safe).
768 hammer2_spin_ex(&parent
->core
.spin
);
769 if (atomic_cmpset_int(&chain
->refs
, 1, 0) == 0) {
771 /* XXX remove, don't try to drop data on fail */
772 hammer2_spin_unex(&parent
->core
.spin
);
773 dio
= hammer2_chain_drop_data(chain
, 0);
774 hammer2_spin_unex(&chain
->core
.spin
);
776 hammer2_io_bqrelse(&dio
);
779 * 1->0 transition failed, retry.
781 hammer2_spin_unex(&parent
->core
.spin
);
782 hammer2_spin_unex(&chain
->core
.spin
);
783 hammer2_mtx_unlock(&chain
->lock
);
789 * 1->0 transition successful, parent spin held to prevent
790 * new lookups, chain spinlock held to protect parent field.
791 * Remove chain from the parent.
793 * If the chain is being removed from the parent's btree but
794 * is not bmapped, we have to adjust live_count downward. If
795 * it is bmapped then the blockref is retained in the parent
796 * as is its associated live_count. This case can occur when
797 * a chain added to the topology is unable to flush and is
798 * then later deleted.
800 if (chain
->flags
& HAMMER2_CHAIN_ONRBTREE
) {
801 if ((parent
->flags
& HAMMER2_CHAIN_COUNTEDBREFS
) &&
802 (chain
->flags
& HAMMER2_CHAIN_BMAPPED
) == 0) {
803 atomic_add_int(&parent
->core
.live_count
, -1);
805 RB_REMOVE(hammer2_chain_tree
,
806 &parent
->core
.rbtree
, chain
);
807 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_ONRBTREE
);
808 --parent
->core
.chain_count
;
809 chain
->parent
= NULL
;
813 * If our chain was the last chain in the parent's core the
814 * core is now empty and its parent might have to be
815 * re-dropped if it has 0 refs.
817 if (parent
->core
.chain_count
== 0) {
819 atomic_add_int(&rdrop
->refs
, 1);
821 if (atomic_cmpset_int(&rdrop->refs, 0, 1) == 0)
825 hammer2_spin_unex(&parent
->core
.spin
);
826 parent
= NULL
; /* safety */
832 if (atomic_cmpset_int(&chain
->refs
, 1, 0) == 0) {
834 * 1->0 transition failed, retry.
836 hammer2_spin_unex(&parent
->core
.spin
);
837 hammer2_spin_unex(&chain
->core
.spin
);
838 hammer2_mtx_unlock(&chain
->lock
);
845 * Successful 1->0 transition, no parent, no children... no way for
846 * anyone to ref this chain any more. We can clean-up and free it.
848 * We still have the core spinlock, and core's chain_count is 0.
849 * Any parent spinlock is gone.
851 hammer2_spin_unex(&chain
->core
.spin
);
852 hammer2_chain_assert_no_data(chain
);
853 hammer2_mtx_unlock(&chain
->lock
);
854 KKASSERT(RB_EMPTY(&chain
->core
.rbtree
) &&
855 chain
->core
.chain_count
== 0);
858 * All locks are gone, no pointers remain to the chain, finish
861 KKASSERT((chain
->flags
& (HAMMER2_CHAIN_UPDATE
|
862 HAMMER2_CHAIN_MODIFIED
)) == 0);
864 dio
= hammer2_chain_drop_data(chain
, 1);
866 hammer2_io_bqrelse(&dio
);
870 * Once chain resources are gone we can use the now dead chain
871 * structure to placehold what might otherwise require a recursive
872 * drop, because we have potentially two things to drop and can only
873 * return one directly.
875 if (chain
->flags
& HAMMER2_CHAIN_ALLOCATED
) {
876 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_ALLOCATED
);
878 kfree(chain
, hmp
->mchain
);
882 * Possible chaining loop when parent re-drop needed.
888 * Heuristical flush of the LRU, try to reduce the number of entries
889 * on the LRU to (HAMMER2_LRU_LIMIT * 2 / 3). This procedure is called
890 * only when lru_count exceeds HAMMER2_LRU_LIMIT.
894 hammer2_chain_lru_flush(hammer2_pfs_t
*pmp
)
896 hammer2_chain_t
*chain
;
900 hammer2_spin_ex(&pmp
->lru_spin
);
901 while (pmp
->lru_count
> HAMMER2_LRU_LIMIT
* 2 / 3) {
903 * Pick a chain off the lru_list, just recycle it quickly
904 * if LRUHINT is set (the chain was ref'd but left on
905 * the lru_list, so cycle to the end).
907 chain
= TAILQ_FIRST(&pmp
->lru_list
);
908 TAILQ_REMOVE(&pmp
->lru_list
, chain
, lru_node
);
910 if (chain
->flags
& HAMMER2_CHAIN_LRUHINT
) {
911 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_LRUHINT
);
912 TAILQ_INSERT_TAIL(&pmp
->lru_list
, chain
, lru_node
);
918 * Ok, we are off the LRU. We must adjust refs before we
919 * can safely clear the ONLRU flag.
921 atomic_add_int(&pmp
->lru_count
, -1);
922 if (atomic_cmpset_int(&chain
->refs
, 0, 1)) {
923 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_ONLRU
);
924 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_RELEASE
);
927 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_ONLRU
);
930 hammer2_spin_unex(&pmp
->lru_spin
);
935 * If we picked a chain off the lru list we may be able to lastdrop
936 * it. Use a depth of 1 to prevent excessive lastdrop recursion.
946 if (hammer2_mtx_ex_try(&chain
->lock
) == 0)
947 chain
= hammer2_chain_lastdrop(chain
, 1);
948 /* retry the same chain, or chain from lastdrop */
950 if (atomic_cmpset_int(&chain
->refs
, refs
, refs
- 1))
952 /* retry the same chain */
960 * On last lock release.
962 static hammer2_io_t
*
963 hammer2_chain_drop_data(hammer2_chain_t
*chain
)
967 if ((dio
= chain
->dio
) != NULL
) {
971 switch(chain
->bref
.type
) {
972 case HAMMER2_BREF_TYPE_VOLUME
:
973 case HAMMER2_BREF_TYPE_FREEMAP
:
976 if (chain
->data
!= NULL
) {
977 hammer2_spin_unex(&chain
->core
.spin
);
978 panic("chain data not null: "
979 "chain %p bref %016jx.%02x "
980 "refs %d parent %p dio %p data %p",
981 chain
, chain
->bref
.data_off
,
982 chain
->bref
.type
, chain
->refs
,
984 chain
->dio
, chain
->data
);
986 KKASSERT(chain
->data
== NULL
);
994 * Lock a referenced chain element, acquiring its data with I/O if necessary,
995 * and specify how you would like the data to be resolved.
997 * If an I/O or other fatal error occurs, chain->error will be set to non-zero.
999 * The lock is allowed to recurse, multiple locking ops will aggregate
1000 * the requested resolve types. Once data is assigned it will not be
1001 * removed until the last unlock.
1003 * HAMMER2_RESOLVE_NEVER - Do not resolve the data element.
1004 * (typically used to avoid device/logical buffer
1005 * aliasing for data)
1007 * HAMMER2_RESOLVE_MAYBE - Do not resolve data elements for chains in
1008 * the INITIAL-create state (indirect blocks only).
1010 * Do not resolve data elements for DATA chains.
1011 * (typically used to avoid device/logical buffer
1012 * aliasing for data)
1014 * HAMMER2_RESOLVE_ALWAYS- Always resolve the data element.
1016 * HAMMER2_RESOLVE_SHARED- (flag) The chain is locked shared, otherwise
1017 * it will be locked exclusive.
1019 * HAMMER2_RESOLVE_NONBLOCK- (flag) The chain is locked non-blocking. If
1020 * the lock fails, EAGAIN is returned.
1022 * NOTE: Embedded elements (volume header, inodes) are always resolved
1025 * NOTE: Specifying HAMMER2_RESOLVE_ALWAYS on a newly-created non-embedded
1026 * element will instantiate and zero its buffer, and flush it on
1029 * NOTE: (data) elements are normally locked RESOLVE_NEVER or RESOLVE_MAYBE
1030 * so as not to instantiate a device buffer, which could alias against
1031 * a logical file buffer. However, if ALWAYS is specified the
1032 * device buffer will be instantiated anyway.
1034 * NOTE: The return value is always 0 unless NONBLOCK is specified, in which
1035 * case it can be either 0 or EAGAIN.
1037 * WARNING! This function blocks on I/O if data needs to be fetched. This
1038 * blocking can run concurrent with other compatible lock holders
1039 * who do not need data returning. The lock is not upgraded to
1040 * exclusive during a data fetch, a separate bit is used to
1041 * interlock I/O. However, an exclusive lock holder can still count
1042 * on being interlocked against an I/O fetch managed by a shared
1046 hammer2_chain_lock(hammer2_chain_t
*chain
, int how
)
1048 KKASSERT(chain
->refs
> 0);
1050 if (how
& HAMMER2_RESOLVE_NONBLOCK
) {
1052 * For non-blocking operation attempt to get the lock
1053 * before bumping lockcnt, just so we don't have to deal
1054 * with dropping lockcnt (and dealing with the underlying
1057 * NOTE: LOCKAGAIN must always succeed without blocking.
1059 if (how
& HAMMER2_RESOLVE_SHARED
) {
1060 if (how
& HAMMER2_RESOLVE_LOCKAGAIN
) {
1061 hammer2_mtx_sh_again(&chain
->lock
);
1063 if (hammer2_mtx_sh_try(&chain
->lock
) != 0)
1067 if (hammer2_mtx_ex_try(&chain
->lock
) != 0)
1070 atomic_add_int(&chain
->lockcnt
, 1);
1071 ++curthread
->td_tracker
;
1074 * Lock the element. Recursive locks are allowed. lockcnt
1075 * ensures that data is left intact.
1077 atomic_add_int(&chain
->lockcnt
, 1);
1080 * Get the appropriate lock. If LOCKAGAIN is flagged with
1081 * SHARED the caller expects a shared lock to already be
1082 * present and we are giving it another ref. This case must
1083 * importantly not block if there is a pending exclusive lock
1086 if (how
& HAMMER2_RESOLVE_SHARED
) {
1087 if (how
& HAMMER2_RESOLVE_LOCKAGAIN
) {
1088 hammer2_mtx_sh_again(&chain
->lock
);
1090 hammer2_mtx_sh(&chain
->lock
);
1093 hammer2_mtx_ex(&chain
->lock
);
1095 ++curthread
->td_tracker
;
1099 * If we already have a valid data pointer make sure the data is
1100 * synchronized to the current cpu, and then no further action is
1105 hammer2_io_bkvasync(chain
->dio
);
1110 * Do we have to resolve the data? This is generally only
1111 * applicable to HAMMER2_BREF_TYPE_DATA which is special-cased.
1112 * Other BREF types expects the data to be there.
1114 switch(how
& HAMMER2_RESOLVE_MASK
) {
1115 case HAMMER2_RESOLVE_NEVER
:
1117 case HAMMER2_RESOLVE_MAYBE
:
1118 if (chain
->flags
& HAMMER2_CHAIN_INITIAL
)
1120 if (chain
->bref
.type
== HAMMER2_BREF_TYPE_DATA
)
1123 if (chain
->bref
.type
== HAMMER2_BREF_TYPE_FREEMAP_NODE
)
1125 if (chain
->bref
.type
== HAMMER2_BREF_TYPE_FREEMAP_LEAF
)
1129 case HAMMER2_RESOLVE_ALWAYS
:
1135 * Caller requires data
1137 hammer2_chain_load_data(chain
);
1143 * Lock the chain, retain the hold, and drop the data persistence count.
1144 * The data should remain valid because we never transitioned lockcnt
1148 hammer2_chain_lock_unhold(hammer2_chain_t
*chain
, int how
)
1150 hammer2_chain_lock(chain
, how
);
1151 atomic_add_int(&chain
->lockcnt
, -1);
1156 * Downgrade an exclusive chain lock to a shared chain lock.
1158 * NOTE: There is no upgrade equivalent due to the ease of
1159 * deadlocks in that direction.
1162 hammer2_chain_lock_downgrade(hammer2_chain_t
*chain
)
1164 hammer2_mtx_downgrade(&chain
->lock
);
1169 * Issue I/O and install chain->data. Caller must hold a chain lock, lock
1170 * may be of any type.
1172 * Once chain->data is set it cannot be disposed of until all locks are
1175 * Make sure the data is synchronized to the current cpu.
1178 hammer2_chain_load_data(hammer2_chain_t
*chain
)
1180 hammer2_blockref_t
*bref
;
1187 * Degenerate case, data already present, or chain has no media
1188 * reference to load.
1192 hammer2_io_bkvasync(chain
->dio
);
1195 if ((chain
->bref
.data_off
& ~HAMMER2_OFF_MASK_RADIX
) == 0)
1199 KKASSERT(hmp
!= NULL
);
1202 * Gain the IOINPROG bit, interlocked block.
1208 oflags
= chain
->flags
;
1210 if (oflags
& HAMMER2_CHAIN_IOINPROG
) {
1211 nflags
= oflags
| HAMMER2_CHAIN_IOSIGNAL
;
1212 tsleep_interlock(&chain
->flags
, 0);
1213 if (atomic_cmpset_int(&chain
->flags
, oflags
, nflags
)) {
1214 tsleep(&chain
->flags
, PINTERLOCKED
,
1219 nflags
= oflags
| HAMMER2_CHAIN_IOINPROG
;
1220 if (atomic_cmpset_int(&chain
->flags
, oflags
, nflags
)) {
1228 * We own CHAIN_IOINPROG
1230 * Degenerate case if we raced another load.
1234 hammer2_io_bkvasync(chain
->dio
);
1239 * We must resolve to a device buffer, either by issuing I/O or
1240 * by creating a zero-fill element. We do not mark the buffer
1241 * dirty when creating a zero-fill element (the hammer2_chain_modify()
1242 * API must still be used to do that).
1244 * The device buffer is variable-sized in powers of 2 down
1245 * to HAMMER2_MIN_ALLOC (typically 1K). A 64K physical storage
1246 * chunk always contains buffers of the same size. (XXX)
1248 * The minimum physical IO size may be larger than the variable
1251 bref
= &chain
->bref
;
1254 * The getblk() optimization can only be used on newly created
1255 * elements if the physical block size matches the request.
1257 if (chain
->flags
& HAMMER2_CHAIN_INITIAL
) {
1258 error
= hammer2_io_new(hmp
, bref
->type
,
1259 bref
->data_off
, chain
->bytes
,
1262 error
= hammer2_io_bread(hmp
, bref
->type
,
1263 bref
->data_off
, chain
->bytes
,
1265 hammer2_adjreadcounter(&chain
->bref
, chain
->bytes
);
1268 chain
->error
= HAMMER2_ERROR_EIO
;
1269 kprintf("hammer2_chain_lock: I/O error %016jx: %d\n",
1270 (intmax_t)bref
->data_off
, error
);
1271 hammer2_io_bqrelse(&chain
->dio
);
1277 * This isn't perfect and can be ignored on OSs which do not have
1278 * an indication as to whether a buffer is coming from cache or
1279 * if I/O was actually issued for the read. TESTEDGOOD will work
1280 * pretty well without the B_IOISSUED logic because chains are
1281 * cached, but in that situation (without B_IOISSUED) it will not
1282 * detect whether a re-read via I/O is corrupted verses the original
1285 * We can't re-run the CRC on every fresh lock. That would be
1286 * insanely expensive.
1288 * If the underlying kernel buffer covers the entire chain we can
1289 * use the B_IOISSUED indication to determine if we have to re-run
1290 * the CRC on chain data for chains that managed to stay cached
1291 * across the kernel disposal of the original buffer.
1293 if ((dio
= chain
->dio
) != NULL
&& dio
->bp
) {
1294 struct buf
*bp
= dio
->bp
;
1296 if (dio
->psize
== chain
->bytes
&&
1297 (bp
->b_flags
& B_IOISSUED
)) {
1298 atomic_clear_int(&chain
->flags
,
1299 HAMMER2_CHAIN_TESTEDGOOD
);
1300 bp
->b_flags
&= ~B_IOISSUED
;
1305 * NOTE: A locked chain's data cannot be modified without first
1306 * calling hammer2_chain_modify().
1310 * Clear INITIAL. In this case we used io_new() and the buffer has
1311 * been zero'd and marked dirty.
1313 * NOTE: hammer2_io_data() call issues bkvasync()
1315 bdata
= hammer2_io_data(chain
->dio
, chain
->bref
.data_off
);
1317 if (chain
->flags
& HAMMER2_CHAIN_INITIAL
) {
1318 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_INITIAL
);
1319 chain
->bref
.flags
|= HAMMER2_BREF_FLAG_ZERO
;
1320 } else if (chain
->flags
& HAMMER2_CHAIN_MODIFIED
) {
1322 * check data not currently synchronized due to
1323 * modification. XXX assumes data stays in the buffer
1324 * cache, which might not be true (need biodep on flush
1325 * to calculate crc? or simple crc?).
1327 } else if ((chain
->flags
& HAMMER2_CHAIN_TESTEDGOOD
) == 0) {
1328 if (hammer2_chain_testcheck(chain
, bdata
) == 0) {
1329 chain
->error
= HAMMER2_ERROR_CHECK
;
1331 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_TESTEDGOOD
);
1336 * Setup the data pointer, either pointing it to an embedded data
1337 * structure and copying the data from the buffer, or pointing it
1340 * The buffer is not retained when copying to an embedded data
1341 * structure in order to avoid potential deadlocks or recursions
1342 * on the same physical buffer.
1344 * WARNING! Other threads can start using the data the instant we
1345 * set chain->data non-NULL.
1347 switch (bref
->type
) {
1348 case HAMMER2_BREF_TYPE_VOLUME
:
1349 case HAMMER2_BREF_TYPE_FREEMAP
:
1351 * Copy data from bp to embedded buffer
1353 panic("hammer2_chain_load_data: unresolved volume header");
1355 case HAMMER2_BREF_TYPE_DIRENT
:
1356 KKASSERT(chain
->bytes
!= 0);
1358 case HAMMER2_BREF_TYPE_INODE
:
1359 case HAMMER2_BREF_TYPE_FREEMAP_LEAF
:
1360 case HAMMER2_BREF_TYPE_INDIRECT
:
1361 case HAMMER2_BREF_TYPE_DATA
:
1362 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
1365 * Point data at the device buffer and leave dio intact.
1367 chain
->data
= (void *)bdata
;
1372 * Release HAMMER2_CHAIN_IOINPROG and signal waiters if requested.
1379 oflags
= chain
->flags
;
1380 nflags
= oflags
& ~(HAMMER2_CHAIN_IOINPROG
|
1381 HAMMER2_CHAIN_IOSIGNAL
);
1382 KKASSERT(oflags
& HAMMER2_CHAIN_IOINPROG
);
1383 if (atomic_cmpset_int(&chain
->flags
, oflags
, nflags
)) {
1384 if (oflags
& HAMMER2_CHAIN_IOSIGNAL
)
1385 wakeup(&chain
->flags
);
1392 * Unlock and deref a chain element.
1394 * Remember that the presence of children under chain prevent the chain's
1395 * destruction but do not add additional references, so the dio will still
1399 hammer2_chain_unlock(hammer2_chain_t
*chain
)
1405 --curthread
->td_tracker
;
1408 * If multiple locks are present (or being attempted) on this
1409 * particular chain we can just unlock, drop refs, and return.
1411 * Otherwise fall-through on the 1->0 transition.
1414 lockcnt
= chain
->lockcnt
;
1415 KKASSERT(lockcnt
> 0);
1418 if (atomic_cmpset_int(&chain
->lockcnt
,
1419 lockcnt
, lockcnt
- 1)) {
1420 hammer2_mtx_unlock(&chain
->lock
);
1423 } else if (hammer2_mtx_upgrade_try(&chain
->lock
) == 0) {
1424 /* while holding the mutex exclusively */
1425 if (atomic_cmpset_int(&chain
->lockcnt
, 1, 0))
1429 * This situation can easily occur on SMP due to
1430 * the gap inbetween the 1->0 transition and the
1431 * final unlock. We cannot safely block on the
1432 * mutex because lockcnt might go above 1.
1434 * XXX Sleep for one tick if it takes too long.
1436 if (++iter
> 1000) {
1437 if (iter
> 1000 + hz
) {
1438 kprintf("hammer2: h2race2 %p\n", chain
);
1441 tsleep(&iter
, 0, "h2race2", 1);
1449 * Last unlock / mutex upgraded to exclusive. Drop the data
1452 dio
= hammer2_chain_drop_data(chain
);
1454 hammer2_io_bqrelse(&dio
);
1455 hammer2_mtx_unlock(&chain
->lock
);
1459 * Unlock and hold chain data intact
1462 hammer2_chain_unlock_hold(hammer2_chain_t
*chain
)
1464 atomic_add_int(&chain
->lockcnt
, 1);
1465 hammer2_chain_unlock(chain
);
1469 * Helper to obtain the blockref[] array base and count for a chain.
1471 * XXX Not widely used yet, various use cases need to be validated and
1472 * converted to use this function.
1475 hammer2_blockref_t
*
1476 hammer2_chain_base_and_count(hammer2_chain_t
*parent
, int *countp
)
1478 hammer2_blockref_t
*base
;
1481 if (parent
->flags
& HAMMER2_CHAIN_INITIAL
) {
1484 switch(parent
->bref
.type
) {
1485 case HAMMER2_BREF_TYPE_INODE
:
1486 count
= HAMMER2_SET_COUNT
;
1488 case HAMMER2_BREF_TYPE_INDIRECT
:
1489 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
1490 count
= parent
->bytes
/ sizeof(hammer2_blockref_t
);
1492 case HAMMER2_BREF_TYPE_VOLUME
:
1493 count
= HAMMER2_SET_COUNT
;
1495 case HAMMER2_BREF_TYPE_FREEMAP
:
1496 count
= HAMMER2_SET_COUNT
;
1499 panic("hammer2_chain_create_indirect: "
1500 "unrecognized blockref type: %d",
1506 switch(parent
->bref
.type
) {
1507 case HAMMER2_BREF_TYPE_INODE
:
1508 base
= &parent
->data
->ipdata
.u
.blockset
.blockref
[0];
1509 count
= HAMMER2_SET_COUNT
;
1511 case HAMMER2_BREF_TYPE_INDIRECT
:
1512 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
1513 base
= &parent
->data
->npdata
[0];
1514 count
= parent
->bytes
/ sizeof(hammer2_blockref_t
);
1516 case HAMMER2_BREF_TYPE_VOLUME
:
1517 base
= &parent
->data
->voldata
.
1518 sroot_blockset
.blockref
[0];
1519 count
= HAMMER2_SET_COUNT
;
1521 case HAMMER2_BREF_TYPE_FREEMAP
:
1522 base
= &parent
->data
->blkset
.blockref
[0];
1523 count
= HAMMER2_SET_COUNT
;
1526 panic("hammer2_chain_create_indirect: "
1527 "unrecognized blockref type: %d",
1539 * This counts the number of live blockrefs in a block array and
1540 * also calculates the point at which all remaining blockrefs are empty.
1541 * This routine can only be called on a live chain.
1543 * Caller holds the chain locked, but possibly with a shared lock. We
1544 * must use an exclusive spinlock to prevent corruption.
1546 * NOTE: Flag is not set until after the count is complete, allowing
1547 * callers to test the flag without holding the spinlock.
1549 * NOTE: If base is NULL the related chain is still in the INITIAL
1550 * state and there are no blockrefs to count.
1552 * NOTE: live_count may already have some counts accumulated due to
1553 * creation and deletion and could even be initially negative.
1556 hammer2_chain_countbrefs(hammer2_chain_t
*chain
,
1557 hammer2_blockref_t
*base
, int count
)
1559 hammer2_spin_ex(&chain
->core
.spin
);
1560 if ((chain
->flags
& HAMMER2_CHAIN_COUNTEDBREFS
) == 0) {
1562 while (--count
>= 0) {
1563 if (base
[count
].type
)
1566 chain
->core
.live_zero
= count
+ 1;
1567 while (count
>= 0) {
1568 if (base
[count
].type
)
1569 atomic_add_int(&chain
->core
.live_count
,
1574 chain
->core
.live_zero
= 0;
1576 /* else do not modify live_count */
1577 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_COUNTEDBREFS
);
1579 hammer2_spin_unex(&chain
->core
.spin
);
1583 * Resize the chain's physical storage allocation in-place. This function does
1584 * not usually adjust the data pointer and must be followed by (typically) a
1585 * hammer2_chain_modify() call to copy any old data over and adjust the
1588 * Chains can be resized smaller without reallocating the storage. Resizing
1589 * larger will reallocate the storage. Excess or prior storage is reclaimed
1590 * asynchronously at a later time.
1592 * An nradix value of 0 is special-cased to mean that the storage should
1593 * be disassociated, that is the chain is being resized to 0 bytes (not 1
1596 * Must be passed an exclusively locked parent and chain.
1598 * This function is mostly used with DATA blocks locked RESOLVE_NEVER in order
1599 * to avoid instantiating a device buffer that conflicts with the vnode data
1600 * buffer. However, because H2 can compress or encrypt data, the chain may
1601 * have a dio assigned to it in those situations, and they do not conflict.
1603 * XXX return error if cannot resize.
1606 hammer2_chain_resize(hammer2_chain_t
*chain
,
1607 hammer2_tid_t mtid
, hammer2_off_t dedup_off
,
1608 int nradix
, int flags
)
1618 * Only data and indirect blocks can be resized for now.
1619 * (The volu root, inodes, and freemap elements use a fixed size).
1621 KKASSERT(chain
!= &hmp
->vchain
);
1622 KKASSERT(chain
->bref
.type
== HAMMER2_BREF_TYPE_DATA
||
1623 chain
->bref
.type
== HAMMER2_BREF_TYPE_INDIRECT
||
1624 chain
->bref
.type
== HAMMER2_BREF_TYPE_DIRENT
);
1627 * Nothing to do if the element is already the proper size
1629 obytes
= chain
->bytes
;
1630 nbytes
= (nradix
) ? (1U << nradix
) : 0;
1631 if (obytes
== nbytes
)
1632 return (chain
->error
);
1635 * Make sure the old data is instantiated so we can copy it. If this
1636 * is a data block, the device data may be superfluous since the data
1637 * might be in a logical block, but compressed or encrypted data is
1640 * NOTE: The modify will set BMAPUPD for us if BMAPPED is set.
1642 error
= hammer2_chain_modify(chain
, mtid
, dedup_off
, 0);
1647 * Relocate the block, even if making it smaller (because different
1648 * block sizes may be in different regions).
1650 * NOTE: Operation does not copy the data and may only be used
1651 * to resize data blocks in-place, or directory entry blocks
1652 * which are about to be modified in some manner.
1654 error
= hammer2_freemap_alloc(chain
, nbytes
);
1658 chain
->bytes
= nbytes
;
1661 * We don't want the followup chain_modify() to try to copy data
1662 * from the old (wrong-sized) buffer. It won't know how much to
1663 * copy. This case should only occur during writes when the
1664 * originator already has the data to write in-hand.
1667 KKASSERT(chain
->bref
.type
== HAMMER2_BREF_TYPE_DATA
||
1668 chain
->bref
.type
== HAMMER2_BREF_TYPE_DIRENT
);
1669 hammer2_io_brelse(&chain
->dio
);
1672 return (chain
->error
);
1676 * Set the chain modified so its data can be changed by the caller, or
1677 * install deduplicated data. The caller must call this routine for each
1678 * set of modifications it makes, even if the chain is already flagged
1681 * Sets bref.modify_tid to mtid only if mtid != 0. Note that bref.modify_tid
1682 * is a CLC (cluster level change) field and is not updated by parent
1683 * propagation during a flush.
1685 * Returns an appropriate HAMMER2_ERROR_* code, which will generally reflect
1686 * chain->error except for HAMMER2_ERROR_ENOSPC. If the allocation fails
1687 * due to no space available, HAMMER2_ERROR_ENOSPC is returned and the chain
1688 * remains unmodified with its old data ref intact and chain->error
1693 * If the DEDUPABLE flag is set in the chain the storage must be reallocated
1694 * even if the chain is still flagged MODIFIED. In this case the chain's
1695 * DEDUPABLE flag will be cleared once the new storage has been assigned.
1697 * If the caller passes a non-zero dedup_off we will use it to assign the
1698 * new storage. The MODIFIED flag will be *CLEARED* in this case, and
1699 * DEDUPABLE will be set (NOTE: the UPDATE flag is always set). The caller
1700 * must not modify the data content upon return.
1703 hammer2_chain_modify(hammer2_chain_t
*chain
, hammer2_tid_t mtid
,
1704 hammer2_off_t dedup_off
, int flags
)
1706 hammer2_blockref_t obref
;
1717 obref
= chain
->bref
;
1718 KKASSERT((chain
->flags
& HAMMER2_CHAIN_FICTITIOUS
) == 0);
1721 * Data is not optional for freemap chains (we must always be sure
1722 * to copy the data on COW storage allocations).
1724 if (chain
->bref
.type
== HAMMER2_BREF_TYPE_FREEMAP_NODE
||
1725 chain
->bref
.type
== HAMMER2_BREF_TYPE_FREEMAP_LEAF
) {
1726 KKASSERT((chain
->flags
& HAMMER2_CHAIN_INITIAL
) ||
1727 (flags
& HAMMER2_MODIFY_OPTDATA
) == 0);
1731 * Data must be resolved if already assigned, unless explicitly
1732 * flagged otherwise. If we cannot safety load the data the
1733 * modification fails and we return early.
1735 if (chain
->data
== NULL
&& chain
->bytes
!= 0 &&
1736 (flags
& HAMMER2_MODIFY_OPTDATA
) == 0 &&
1737 (chain
->bref
.data_off
& ~HAMMER2_OFF_MASK_RADIX
)) {
1738 hammer2_chain_load_data(chain
);
1740 return (chain
->error
);
1745 * Set MODIFIED to indicate that the chain has been modified. A new
1746 * allocation is required when modifying a chain.
1748 * Set UPDATE to ensure that the blockref is updated in the parent.
1750 * If MODIFIED is already set determine if we can reuse the assigned
1751 * data block or if we need a new data block.
1753 if ((chain
->flags
& HAMMER2_CHAIN_MODIFIED
) == 0) {
1755 * Must set modified bit.
1757 atomic_add_long(&hammer2_count_modified_chains
, 1);
1758 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_MODIFIED
);
1759 hammer2_pfs_memory_inc(chain
->pmp
); /* can be NULL */
1763 * We may be able to avoid a copy-on-write if the chain's
1764 * check mode is set to NONE and the chain's current
1765 * modify_tid is beyond the last explicit snapshot tid.
1767 * This implements HAMMER2's overwrite-in-place feature.
1769 * NOTE! This data-block cannot be used as a de-duplication
1770 * source when the check mode is set to NONE.
1772 if ((chain
->bref
.type
== HAMMER2_BREF_TYPE_DATA
||
1773 chain
->bref
.type
== HAMMER2_BREF_TYPE_DIRENT
) &&
1774 (chain
->flags
& HAMMER2_CHAIN_INITIAL
) == 0 &&
1775 (chain
->flags
& HAMMER2_CHAIN_DEDUPABLE
) == 0 &&
1776 HAMMER2_DEC_CHECK(chain
->bref
.methods
) ==
1777 HAMMER2_CHECK_NONE
&&
1779 chain
->bref
.modify_tid
>
1780 chain
->pmp
->iroot
->meta
.pfs_lsnap_tid
) {
1782 * Sector overwrite allowed.
1787 * Sector overwrite not allowed, must copy-on-write.
1791 } else if (chain
->flags
& HAMMER2_CHAIN_DEDUPABLE
) {
1793 * If the modified chain was registered for dedup we need
1794 * a new allocation. This only happens for delayed-flush
1795 * chains (i.e. which run through the front-end buffer
1802 * Already flagged modified, no new allocation is needed.
1809 * Flag parent update required.
1811 if ((chain
->flags
& HAMMER2_CHAIN_UPDATE
) == 0) {
1812 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_UPDATE
);
1819 * The XOP code returns held but unlocked focus chains. This
1820 * prevents the chain from being destroyed but does not prevent
1821 * it from being modified. diolk is used to interlock modifications
1822 * against XOP frontend accesses to the focus.
1824 * This allows us to theoretically avoid deadlocking the frontend
1825 * if one of the backends lock up by not formally locking the
1826 * focused chain in the frontend. In addition, the synchronization
1827 * code relies on this mechanism to avoid deadlocking concurrent
1828 * synchronization threads.
1830 lockmgr(&chain
->diolk
, LK_EXCLUSIVE
);
1833 * The modification or re-modification requires an allocation and
1834 * possible COW. If an error occurs, the previous content and data
1835 * reference is retained and the modification fails.
1837 * If dedup_off is non-zero, the caller is requesting a deduplication
1838 * rather than a modification. The MODIFIED bit is not set and the
1839 * data offset is set to the deduplication offset. The data cannot
1842 * NOTE: The dedup offset is allowed to be in a partially free state
1843 * and we must be sure to reset it to a fully allocated state
1844 * to force two bulkfree passes to free it again.
1846 * NOTE: Only applicable when chain->bytes != 0.
1848 * XXX can a chain already be marked MODIFIED without a data
1849 * assignment? If not, assert here instead of testing the case.
1851 if (chain
!= &hmp
->vchain
&& chain
!= &hmp
->fchain
&&
1853 if ((chain
->bref
.data_off
& ~HAMMER2_OFF_MASK_RADIX
) == 0 ||
1857 * NOTE: We do not have to remove the dedup
1858 * registration because the area is still
1859 * allocated and the underlying DIO will
1863 chain
->bref
.data_off
= dedup_off
;
1864 chain
->bytes
= 1 << (dedup_off
&
1865 HAMMER2_OFF_MASK_RADIX
);
1867 atomic_clear_int(&chain
->flags
,
1868 HAMMER2_CHAIN_MODIFIED
);
1869 atomic_add_long(&hammer2_count_modified_chains
,
1872 hammer2_pfs_memory_wakeup(chain
->pmp
);
1873 hammer2_freemap_adjust(hmp
, &chain
->bref
,
1874 HAMMER2_FREEMAP_DORECOVER
);
1875 atomic_set_int(&chain
->flags
,
1876 HAMMER2_CHAIN_DEDUPABLE
);
1878 error
= hammer2_freemap_alloc(chain
,
1880 atomic_clear_int(&chain
->flags
,
1881 HAMMER2_CHAIN_DEDUPABLE
);
1887 * Stop here if error. We have to undo any flag bits we might
1892 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_MODIFIED
);
1893 atomic_add_long(&hammer2_count_modified_chains
, -1);
1895 hammer2_pfs_memory_wakeup(chain
->pmp
);
1898 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_UPDATE
);
1900 lockmgr(&chain
->diolk
, LK_RELEASE
);
1906 * Update mirror_tid and modify_tid. modify_tid is only updated
1907 * if not passed as zero (during flushes, parent propagation passes
1910 * NOTE: chain->pmp could be the device spmp.
1912 chain
->bref
.mirror_tid
= hmp
->voldata
.mirror_tid
+ 1;
1914 chain
->bref
.modify_tid
= mtid
;
1917 * Set BMAPUPD to tell the flush code that an existing blockmap entry
1918 * requires updating as well as to tell the delete code that the
1919 * chain's blockref might not exactly match (in terms of physical size
1920 * or block offset) the one in the parent's blocktable. The base key
1921 * of course will still match.
1923 if (chain
->flags
& HAMMER2_CHAIN_BMAPPED
)
1924 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_BMAPUPD
);
1927 * Short-cut data blocks which the caller does not need an actual
1928 * data reference to (aka OPTDATA), as long as the chain does not
1929 * already have a data pointer to the data. This generally means
1930 * that the modifications are being done via the logical buffer cache.
1931 * The INITIAL flag relates only to the device data buffer and thus
1932 * remains unchange in this situation.
1934 * This code also handles bytes == 0 (most dirents).
1936 if (chain
->bref
.type
== HAMMER2_BREF_TYPE_DATA
&&
1937 (flags
& HAMMER2_MODIFY_OPTDATA
) &&
1938 chain
->data
== NULL
) {
1939 KKASSERT(chain
->dio
== NULL
);
1944 * Clearing the INITIAL flag (for indirect blocks) indicates that
1945 * we've processed the uninitialized storage allocation.
1947 * If this flag is already clear we are likely in a copy-on-write
1948 * situation but we have to be sure NOT to bzero the storage if
1949 * no data is present.
1951 if (chain
->flags
& HAMMER2_CHAIN_INITIAL
) {
1952 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_INITIAL
);
1959 * Instantiate data buffer and possibly execute COW operation
1961 switch(chain
->bref
.type
) {
1962 case HAMMER2_BREF_TYPE_VOLUME
:
1963 case HAMMER2_BREF_TYPE_FREEMAP
:
1965 * The data is embedded, no copy-on-write operation is
1968 KKASSERT(chain
->dio
== NULL
);
1970 case HAMMER2_BREF_TYPE_DIRENT
:
1972 * The data might be fully embedded.
1974 if (chain
->bytes
== 0) {
1975 KKASSERT(chain
->dio
== NULL
);
1979 case HAMMER2_BREF_TYPE_INODE
:
1980 case HAMMER2_BREF_TYPE_FREEMAP_LEAF
:
1981 case HAMMER2_BREF_TYPE_DATA
:
1982 case HAMMER2_BREF_TYPE_INDIRECT
:
1983 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
1985 * Perform the copy-on-write operation
1987 * zero-fill or copy-on-write depending on whether
1988 * chain->data exists or not and set the dirty state for
1989 * the new buffer. hammer2_io_new() will handle the
1992 * If a dedup_off was supplied this is an existing block
1993 * and no COW, copy, or further modification is required.
1995 KKASSERT(chain
!= &hmp
->vchain
&& chain
!= &hmp
->fchain
);
1997 if (wasinitial
&& dedup_off
== 0) {
1998 error
= hammer2_io_new(hmp
, chain
->bref
.type
,
1999 chain
->bref
.data_off
,
2000 chain
->bytes
, &dio
);
2002 error
= hammer2_io_bread(hmp
, chain
->bref
.type
,
2003 chain
->bref
.data_off
,
2004 chain
->bytes
, &dio
);
2006 hammer2_adjreadcounter(&chain
->bref
, chain
->bytes
);
2009 * If an I/O error occurs make sure callers cannot accidently
2010 * modify the old buffer's contents and corrupt the filesystem.
2012 * NOTE: hammer2_io_data() call issues bkvasync()
2015 kprintf("hammer2_chain_modify: hmp=%p I/O error\n",
2017 chain
->error
= HAMMER2_ERROR_EIO
;
2018 hammer2_io_brelse(&dio
);
2019 hammer2_io_brelse(&chain
->dio
);
2024 bdata
= hammer2_io_data(dio
, chain
->bref
.data_off
);
2028 * COW (unless a dedup).
2030 KKASSERT(chain
->dio
!= NULL
);
2031 if (chain
->data
!= (void *)bdata
&& dedup_off
== 0) {
2032 bcopy(chain
->data
, bdata
, chain
->bytes
);
2034 } else if (wasinitial
== 0) {
2036 * We have a problem. We were asked to COW but
2037 * we don't have any data to COW with!
2039 panic("hammer2_chain_modify: having a COW %p\n",
2044 * Retire the old buffer, replace with the new. Dirty or
2045 * redirty the new buffer.
2047 * WARNING! The system buffer cache may have already flushed
2048 * the buffer, so we must be sure to [re]dirty it
2049 * for further modification.
2051 * If dedup_off was supplied, the caller is not
2052 * expected to make any further modification to the
2055 * WARNING! hammer2_get_gdata() assumes dio never transitions
2056 * through NULL in order to optimize away unnecessary
2062 if ((tio
= chain
->dio
) != NULL
)
2063 hammer2_io_bqrelse(&tio
);
2064 chain
->data
= (void *)bdata
;
2067 hammer2_io_setdirty(dio
);
2071 panic("hammer2_chain_modify: illegal non-embedded type %d",
2078 * setflush on parent indicating that the parent must recurse down
2079 * to us. Do not call on chain itself which might already have it
2083 hammer2_chain_setflush(chain
->parent
);
2084 lockmgr(&chain
->diolk
, LK_RELEASE
);
2086 return (chain
->error
);
2090 * Modify the chain associated with an inode.
2093 hammer2_chain_modify_ip(hammer2_inode_t
*ip
, hammer2_chain_t
*chain
,
2094 hammer2_tid_t mtid
, int flags
)
2098 hammer2_inode_modify(ip
);
2099 error
= hammer2_chain_modify(chain
, mtid
, 0, flags
);
2105 * Volume header data locks
2108 hammer2_voldata_lock(hammer2_dev_t
*hmp
)
2110 lockmgr(&hmp
->vollk
, LK_EXCLUSIVE
);
2114 hammer2_voldata_unlock(hammer2_dev_t
*hmp
)
2116 lockmgr(&hmp
->vollk
, LK_RELEASE
);
2120 hammer2_voldata_modify(hammer2_dev_t
*hmp
)
2122 if ((hmp
->vchain
.flags
& HAMMER2_CHAIN_MODIFIED
) == 0) {
2123 atomic_add_long(&hammer2_count_modified_chains
, 1);
2124 atomic_set_int(&hmp
->vchain
.flags
, HAMMER2_CHAIN_MODIFIED
);
2125 hammer2_pfs_memory_inc(hmp
->vchain
.pmp
);
2130 * This function returns the chain at the nearest key within the specified
2131 * range. The returned chain will be referenced but not locked.
2133 * This function will recurse through chain->rbtree as necessary and will
2134 * return a *key_nextp suitable for iteration. *key_nextp is only set if
2135 * the iteration value is less than the current value of *key_nextp.
2137 * The caller should use (*key_nextp) to calculate the actual range of
2138 * the returned element, which will be (key_beg to *key_nextp - 1), because
2139 * there might be another element which is superior to the returned element
2142 * (*key_nextp) can be passed as key_beg in an iteration only while non-NULL
2143 * chains continue to be returned. On EOF (*key_nextp) may overflow since
2144 * it will wind up being (key_end + 1).
2146 * WARNING! Must be called with child's spinlock held. Spinlock remains
2147 * held through the operation.
2149 struct hammer2_chain_find_info
{
2150 hammer2_chain_t
*best
;
2151 hammer2_key_t key_beg
;
2152 hammer2_key_t key_end
;
2153 hammer2_key_t key_next
;
2156 static int hammer2_chain_find_cmp(hammer2_chain_t
*child
, void *data
);
2157 static int hammer2_chain_find_callback(hammer2_chain_t
*child
, void *data
);
2161 hammer2_chain_find(hammer2_chain_t
*parent
, hammer2_key_t
*key_nextp
,
2162 hammer2_key_t key_beg
, hammer2_key_t key_end
)
2164 struct hammer2_chain_find_info info
;
2167 info
.key_beg
= key_beg
;
2168 info
.key_end
= key_end
;
2169 info
.key_next
= *key_nextp
;
2171 RB_SCAN(hammer2_chain_tree
, &parent
->core
.rbtree
,
2172 hammer2_chain_find_cmp
, hammer2_chain_find_callback
,
2174 *key_nextp
= info
.key_next
;
2176 kprintf("chain_find %p %016jx:%016jx next=%016jx\n",
2177 parent
, key_beg
, key_end
, *key_nextp
);
2185 hammer2_chain_find_cmp(hammer2_chain_t
*child
, void *data
)
2187 struct hammer2_chain_find_info
*info
= data
;
2188 hammer2_key_t child_beg
;
2189 hammer2_key_t child_end
;
2191 child_beg
= child
->bref
.key
;
2192 child_end
= child_beg
+ ((hammer2_key_t
)1 << child
->bref
.keybits
) - 1;
2194 if (child_end
< info
->key_beg
)
2196 if (child_beg
> info
->key_end
)
2203 hammer2_chain_find_callback(hammer2_chain_t
*child
, void *data
)
2205 struct hammer2_chain_find_info
*info
= data
;
2206 hammer2_chain_t
*best
;
2207 hammer2_key_t child_end
;
2210 * WARNING! Layerq is scanned forwards, exact matches should keep
2211 * the existing info->best.
2213 if ((best
= info
->best
) == NULL
) {
2215 * No previous best. Assign best
2218 } else if (best
->bref
.key
<= info
->key_beg
&&
2219 child
->bref
.key
<= info
->key_beg
) {
2224 /*info->best = child;*/
2225 } else if (child
->bref
.key
< best
->bref
.key
) {
2227 * Child has a nearer key and best is not flush with key_beg.
2228 * Set best to child. Truncate key_next to the old best key.
2231 if (info
->key_next
> best
->bref
.key
|| info
->key_next
== 0)
2232 info
->key_next
= best
->bref
.key
;
2233 } else if (child
->bref
.key
== best
->bref
.key
) {
2235 * If our current best is flush with the child then this
2236 * is an illegal overlap.
2238 * key_next will automatically be limited to the smaller of
2239 * the two end-points.
2245 * Keep the current best but truncate key_next to the child's
2248 * key_next will also automatically be limited to the smaller
2249 * of the two end-points (probably not necessary for this case
2250 * but we do it anyway).
2252 if (info
->key_next
> child
->bref
.key
|| info
->key_next
== 0)
2253 info
->key_next
= child
->bref
.key
;
2257 * Always truncate key_next based on child's end-of-range.
2259 child_end
= child
->bref
.key
+ ((hammer2_key_t
)1 << child
->bref
.keybits
);
2260 if (child_end
&& (info
->key_next
> child_end
|| info
->key_next
== 0))
2261 info
->key_next
= child_end
;
2267 * Retrieve the specified chain from a media blockref, creating the
2268 * in-memory chain structure which reflects it. The returned chain is
2269 * held and locked according to (how) (HAMMER2_RESOLVE_*). The caller must
2270 * handle crc-checks and so forth, and should check chain->error before
2271 * assuming that the data is good.
2273 * To handle insertion races pass the INSERT_RACE flag along with the
2274 * generation number of the core. NULL will be returned if the generation
2275 * number changes before we have a chance to insert the chain. Insert
2276 * races can occur because the parent might be held shared.
2278 * Caller must hold the parent locked shared or exclusive since we may
2279 * need the parent's bref array to find our block.
2281 * WARNING! chain->pmp is always set to NULL for any chain representing
2282 * part of the super-root topology.
2285 hammer2_chain_get(hammer2_chain_t
*parent
, int generation
,
2286 hammer2_blockref_t
*bref
, int how
)
2288 hammer2_dev_t
*hmp
= parent
->hmp
;
2289 hammer2_chain_t
*chain
;
2293 * Allocate a chain structure representing the existing media
2294 * entry. Resulting chain has one ref and is not locked.
2296 if (bref
->flags
& HAMMER2_BREF_FLAG_PFSROOT
)
2297 chain
= hammer2_chain_alloc(hmp
, NULL
, bref
);
2299 chain
= hammer2_chain_alloc(hmp
, parent
->pmp
, bref
);
2300 /* ref'd chain returned */
2303 * Flag that the chain is in the parent's blockmap so delete/flush
2304 * knows what to do with it.
2306 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_BMAPPED
);
2309 * chain must be locked to avoid unexpected ripouts
2311 hammer2_chain_lock(chain
, how
);
2314 * Link the chain into its parent. A spinlock is required to safely
2315 * access the RBTREE, and it is possible to collide with another
2316 * hammer2_chain_get() operation because the caller might only hold
2317 * a shared lock on the parent.
2319 * NOTE: Get races can occur quite often when we distribute
2320 * asynchronous read-aheads across multiple threads.
2322 KKASSERT(parent
->refs
> 0);
2323 error
= hammer2_chain_insert(parent
, chain
,
2324 HAMMER2_CHAIN_INSERT_SPIN
|
2325 HAMMER2_CHAIN_INSERT_RACE
,
2328 KKASSERT((chain
->flags
& HAMMER2_CHAIN_ONRBTREE
) == 0);
2329 /*kprintf("chain %p get race\n", chain);*/
2330 hammer2_chain_unlock(chain
);
2331 hammer2_chain_drop(chain
);
2334 KKASSERT(chain
->flags
& HAMMER2_CHAIN_ONRBTREE
);
2338 * Return our new chain referenced but not locked, or NULL if
2345 * Lookup initialization/completion API
2348 hammer2_chain_lookup_init(hammer2_chain_t
*parent
, int flags
)
2350 hammer2_chain_ref(parent
);
2351 if (flags
& HAMMER2_LOOKUP_SHARED
) {
2352 hammer2_chain_lock(parent
, HAMMER2_RESOLVE_ALWAYS
|
2353 HAMMER2_RESOLVE_SHARED
);
2355 hammer2_chain_lock(parent
, HAMMER2_RESOLVE_ALWAYS
);
2361 hammer2_chain_lookup_done(hammer2_chain_t
*parent
)
2364 hammer2_chain_unlock(parent
);
2365 hammer2_chain_drop(parent
);
2370 * Take the locked chain and return a locked parent. The chain remains
2371 * locked on return, but may have to be temporarily unlocked to acquire
2372 * the parent. Because of this, (chain) must be stable and cannot be
2373 * deleted while it was temporarily unlocked (typically means that (chain)
2376 * Pass HAMMER2_RESOLVE_* flags in flags.
2378 * This will work even if the chain is errored, and the caller can check
2379 * parent->error on return if desired since the parent will be locked.
2381 * This function handles the lock order reversal.
2384 hammer2_chain_getparent(hammer2_chain_t
*chain
, int flags
)
2386 hammer2_chain_t
*parent
;
2389 * Be careful of order, chain must be unlocked before parent
2390 * is locked below to avoid a deadlock. Try it trivially first.
2392 parent
= chain
->parent
;
2394 panic("hammer2_chain_getparent: no parent");
2395 hammer2_chain_ref(parent
);
2396 if (hammer2_chain_lock(parent
, flags
|HAMMER2_RESOLVE_NONBLOCK
) == 0)
2400 hammer2_chain_unlock(chain
);
2401 hammer2_chain_lock(parent
, flags
);
2402 hammer2_chain_lock(chain
, flags
);
2405 * Parent relinking races are quite common. We have to get
2406 * it right or we will blow up the block table.
2408 if (chain
->parent
== parent
)
2410 hammer2_chain_unlock(parent
);
2411 hammer2_chain_drop(parent
);
2413 parent
= chain
->parent
;
2415 panic("hammer2_chain_getparent: no parent");
2416 hammer2_chain_ref(parent
);
2422 * Take the locked chain and return a locked parent. The chain is unlocked
2423 * and dropped. *chainp is set to the returned parent as a convenience.
2424 * Pass HAMMER2_RESOLVE_* flags in flags.
2426 * This will work even if the chain is errored, and the caller can check
2427 * parent->error on return if desired since the parent will be locked.
2429 * The chain does NOT need to be stable. We use a tracking structure
2430 * to track the expected parent if the chain is deleted out from under us.
2432 * This function handles the lock order reversal.
2435 hammer2_chain_repparent(hammer2_chain_t
**chainp
, int flags
)
2437 hammer2_chain_t
*chain
;
2438 hammer2_chain_t
*parent
;
2439 struct hammer2_reptrack reptrack
;
2440 struct hammer2_reptrack
**repp
;
2443 * Be careful of order, chain must be unlocked before parent
2444 * is locked below to avoid a deadlock. Try it trivially first.
2447 parent
= chain
->parent
;
2448 if (parent
== NULL
) {
2449 hammer2_spin_unex(&chain
->core
.spin
);
2450 panic("hammer2_chain_repparent: no parent");
2452 hammer2_chain_ref(parent
);
2453 if (hammer2_chain_lock(parent
, flags
|HAMMER2_RESOLVE_NONBLOCK
) == 0) {
2454 hammer2_chain_unlock(chain
);
2455 hammer2_chain_drop(chain
);
2462 * Ok, now it gets a bit nasty. There are multiple situations where
2463 * the parent might be in the middle of a deletion, or where the child
2464 * (chain) might be deleted the instant we let go of its lock.
2465 * We can potentially end up in a no-win situation!
2467 * In particular, the indirect_maintenance() case can cause these
2470 * To deal with this we install a reptrack structure in the parent
2471 * This reptrack structure 'owns' the parent ref and will automatically
2472 * migrate to the parent's parent if the parent is deleted permanently.
2474 hammer2_spin_init(&reptrack
.spin
, "h2reptrk");
2475 reptrack
.chain
= parent
;
2476 hammer2_chain_ref(parent
); /* for the reptrack */
2478 hammer2_spin_ex(&parent
->core
.spin
);
2479 reptrack
.next
= parent
->core
.reptrack
;
2480 parent
->core
.reptrack
= &reptrack
;
2481 hammer2_spin_unex(&parent
->core
.spin
);
2483 hammer2_chain_unlock(chain
);
2484 hammer2_chain_drop(chain
);
2485 chain
= NULL
; /* gone */
2488 * At the top of this loop, chain is gone and parent is refd both
2489 * by us explicitly AND via our reptrack. We are attempting to
2493 hammer2_chain_lock(parent
, flags
);
2495 if (reptrack
.chain
== parent
)
2497 hammer2_chain_unlock(parent
);
2498 hammer2_chain_drop(parent
);
2500 kprintf("hammer2: debug REPTRACK %p->%p\n",
2501 parent
, reptrack
.chain
);
2502 hammer2_spin_ex(&reptrack
.spin
);
2503 parent
= reptrack
.chain
;
2504 hammer2_chain_ref(parent
);
2505 hammer2_spin_unex(&reptrack
.spin
);
2509 * Once parent is locked and matches our reptrack, our reptrack
2510 * will be stable and we have our parent. We can unlink our
2513 * WARNING! Remember that the chain lock might be shared. Chains
2514 * locked shared have stable parent linkages.
2516 hammer2_spin_ex(&parent
->core
.spin
);
2517 repp
= &parent
->core
.reptrack
;
2518 while (*repp
!= &reptrack
)
2519 repp
= &(*repp
)->next
;
2520 *repp
= reptrack
.next
;
2521 hammer2_spin_unex(&parent
->core
.spin
);
2523 hammer2_chain_drop(parent
); /* reptrack ref */
2524 *chainp
= parent
; /* return parent lock+ref */
2530 * Dispose of any linked reptrack structures in (chain) by shifting them to
2531 * (parent). Both (chain) and (parent) must be exclusively locked.
2533 * This is interlocked against any children of (chain) on the other side.
2534 * No children so remain as-of when this is called so we can test
2535 * core.reptrack without holding the spin-lock.
2537 * Used whenever the caller intends to permanently delete chains related
2538 * to topological recursions (BREF_TYPE_INDIRECT, BREF_TYPE_FREEMAP_NODE),
2539 * where the chains underneath the node being deleted are given a new parent
2540 * above the node being deleted.
2544 hammer2_chain_repchange(hammer2_chain_t
*parent
, hammer2_chain_t
*chain
)
2546 struct hammer2_reptrack
*reptrack
;
2548 KKASSERT(chain
->core
.live_count
== 0 && RB_EMPTY(&chain
->core
.rbtree
));
2549 while (chain
->core
.reptrack
) {
2550 hammer2_spin_ex(&parent
->core
.spin
);
2551 hammer2_spin_ex(&chain
->core
.spin
);
2552 reptrack
= chain
->core
.reptrack
;
2553 if (reptrack
== NULL
) {
2554 hammer2_spin_unex(&chain
->core
.spin
);
2555 hammer2_spin_unex(&parent
->core
.spin
);
2558 hammer2_spin_ex(&reptrack
->spin
);
2559 chain
->core
.reptrack
= reptrack
->next
;
2560 reptrack
->chain
= parent
;
2561 reptrack
->next
= parent
->core
.reptrack
;
2562 parent
->core
.reptrack
= reptrack
;
2563 hammer2_chain_ref(parent
); /* reptrack */
2565 hammer2_spin_unex(&chain
->core
.spin
);
2566 hammer2_spin_unex(&parent
->core
.spin
);
2567 kprintf("hammer2: debug repchange %p %p->%p\n",
2568 reptrack
, chain
, parent
);
2569 hammer2_chain_drop(chain
); /* reptrack */
2574 * Locate the first chain whos key range overlaps (key_beg, key_end) inclusive.
2575 * (*parentp) typically points to an inode but can also point to a related
2576 * indirect block and this function will recurse upwards and find the inode
2577 * or the nearest undeleted indirect block covering the key range.
2579 * This function unconditionally sets *errorp, replacing any previous value.
2581 * (*parentp) must be exclusive or shared locked (depending on flags) and
2582 * referenced and can be an inode or an existing indirect block within the
2585 * If (*parent) is errored out, this function will not attempt to recurse
2586 * the radix tree and will return NULL along with an appropriate *errorp.
2587 * If NULL is returned and *errorp is 0, the requested lookup could not be
2590 * On return (*parentp) will be modified to point at the deepest parent chain
2591 * element encountered during the search, as a helper for an insertion or
2594 * The new (*parentp) will be locked shared or exclusive (depending on flags),
2595 * and referenced, and the old will be unlocked and dereferenced (no change
2596 * if they are both the same). This is particularly important if the caller
2597 * wishes to insert a new chain, (*parentp) will be set properly even if NULL
2598 * is returned, as long as no error occurred.
2600 * The matching chain will be returned locked according to flags.
2604 * NULL is returned if no match was found, but (*parentp) will still
2605 * potentially be adjusted.
2607 * On return (*key_nextp) will point to an iterative value for key_beg.
2608 * (If NULL is returned (*key_nextp) is set to (key_end + 1)).
2610 * This function will also recurse up the chain if the key is not within the
2611 * current parent's range. (*parentp) can never be set to NULL. An iteration
2612 * can simply allow (*parentp) to float inside the loop.
2614 * NOTE! chain->data is not always resolved. By default it will not be
2615 * resolved for BREF_TYPE_DATA, FREEMAP_NODE, or FREEMAP_LEAF. Use
2616 * HAMMER2_LOOKUP_ALWAYS to force resolution (but be careful w/
2617 * BREF_TYPE_DATA as the device buffer can alias the logical file
2622 hammer2_chain_lookup(hammer2_chain_t
**parentp
, hammer2_key_t
*key_nextp
,
2623 hammer2_key_t key_beg
, hammer2_key_t key_end
,
2624 int *errorp
, int flags
)
2627 hammer2_chain_t
*parent
;
2628 hammer2_chain_t
*chain
;
2629 hammer2_blockref_t
*base
;
2630 hammer2_blockref_t
*bref
;
2631 hammer2_blockref_t bcopy
;
2632 hammer2_key_t scan_beg
;
2633 hammer2_key_t scan_end
;
2635 int how_always
= HAMMER2_RESOLVE_ALWAYS
;
2636 int how_maybe
= HAMMER2_RESOLVE_MAYBE
;
2639 int maxloops
= 300000;
2640 volatile hammer2_mtx_t save_mtx
;
2642 if (flags
& HAMMER2_LOOKUP_ALWAYS
) {
2643 how_maybe
= how_always
;
2644 how
= HAMMER2_RESOLVE_ALWAYS
;
2645 } else if (flags
& HAMMER2_LOOKUP_NODATA
) {
2646 how
= HAMMER2_RESOLVE_NEVER
;
2648 how
= HAMMER2_RESOLVE_MAYBE
;
2650 if (flags
& HAMMER2_LOOKUP_SHARED
) {
2651 how_maybe
|= HAMMER2_RESOLVE_SHARED
;
2652 how_always
|= HAMMER2_RESOLVE_SHARED
;
2653 how
|= HAMMER2_RESOLVE_SHARED
;
2657 * Recurse (*parentp) upward if necessary until the parent completely
2658 * encloses the key range or we hit the inode.
2660 * Handle races against the flusher deleting indirect nodes on its
2661 * way back up by continuing to recurse upward past the deletion.
2667 while (parent
->bref
.type
== HAMMER2_BREF_TYPE_INDIRECT
||
2668 parent
->bref
.type
== HAMMER2_BREF_TYPE_FREEMAP_NODE
) {
2669 scan_beg
= parent
->bref
.key
;
2670 scan_end
= scan_beg
+
2671 ((hammer2_key_t
)1 << parent
->bref
.keybits
) - 1;
2672 if ((parent
->flags
& HAMMER2_CHAIN_DELETED
) == 0) {
2673 if (key_beg
>= scan_beg
&& key_end
<= scan_end
)
2676 parent
= hammer2_chain_repparent(parentp
, how_maybe
);
2679 if (--maxloops
== 0)
2680 panic("hammer2_chain_lookup: maxloops");
2682 * Locate the blockref array. Currently we do a fully associative
2683 * search through the array.
2685 switch(parent
->bref
.type
) {
2686 case HAMMER2_BREF_TYPE_INODE
:
2688 * Special shortcut for embedded data returns the inode
2689 * itself. Callers must detect this condition and access
2690 * the embedded data (the strategy code does this for us).
2692 * This is only applicable to regular files and softlinks.
2694 * We need a second lock on parent. Since we already have
2695 * a lock we must pass LOCKAGAIN to prevent unexpected
2696 * blocking (we don't want to block on a second shared
2697 * ref if an exclusive lock is pending)
2699 if (parent
->data
->ipdata
.meta
.op_flags
&
2700 HAMMER2_OPFLAG_DIRECTDATA
) {
2701 if (flags
& HAMMER2_LOOKUP_NODIRECT
) {
2703 *key_nextp
= key_end
+ 1;
2706 hammer2_chain_ref(parent
);
2707 hammer2_chain_lock(parent
, how_always
|
2708 HAMMER2_RESOLVE_LOCKAGAIN
);
2709 *key_nextp
= key_end
+ 1;
2712 base
= &parent
->data
->ipdata
.u
.blockset
.blockref
[0];
2713 count
= HAMMER2_SET_COUNT
;
2715 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
2716 case HAMMER2_BREF_TYPE_INDIRECT
:
2718 * Handle MATCHIND on the parent
2720 if (flags
& HAMMER2_LOOKUP_MATCHIND
) {
2721 scan_beg
= parent
->bref
.key
;
2722 scan_end
= scan_beg
+
2723 ((hammer2_key_t
)1 << parent
->bref
.keybits
) - 1;
2724 if (key_beg
== scan_beg
&& key_end
== scan_end
) {
2726 hammer2_chain_ref(chain
);
2727 hammer2_chain_lock(chain
, how_maybe
);
2728 *key_nextp
= scan_end
+ 1;
2734 * Optimize indirect blocks in the INITIAL state to avoid
2737 if (parent
->flags
& HAMMER2_CHAIN_INITIAL
) {
2740 if (parent
->data
== NULL
) {
2741 kprintf("parent->data is NULL %p\n", parent
);
2743 tsleep(parent
, 0, "xxx", 0);
2745 base
= &parent
->data
->npdata
[0];
2747 count
= parent
->bytes
/ sizeof(hammer2_blockref_t
);
2749 case HAMMER2_BREF_TYPE_VOLUME
:
2750 base
= &parent
->data
->voldata
.sroot_blockset
.blockref
[0];
2751 count
= HAMMER2_SET_COUNT
;
2753 case HAMMER2_BREF_TYPE_FREEMAP
:
2754 base
= &parent
->data
->blkset
.blockref
[0];
2755 count
= HAMMER2_SET_COUNT
;
2758 kprintf("hammer2_chain_lookup: unrecognized "
2759 "blockref(B) type: %d",
2762 tsleep(&base
, 0, "dead", 0);
2763 panic("hammer2_chain_lookup: unrecognized "
2764 "blockref(B) type: %d",
2766 base
= NULL
; /* safety */
2767 count
= 0; /* safety */
2771 * No lookup is possible if the parent is errored. We delayed
2772 * this check as long as we could to ensure that the parent backup,
2773 * embedded data, and MATCHIND code could still execute.
2775 if (parent
->error
) {
2776 *errorp
= parent
->error
;
2781 * Merged scan to find next candidate.
2783 * hammer2_base_*() functions require the parent->core.live_* fields
2784 * to be synchronized.
2786 * We need to hold the spinlock to access the block array and RB tree
2787 * and to interlock chain creation.
2789 if ((parent
->flags
& HAMMER2_CHAIN_COUNTEDBREFS
) == 0)
2790 hammer2_chain_countbrefs(parent
, base
, count
);
2795 hammer2_spin_ex(&parent
->core
.spin
);
2796 chain
= hammer2_combined_find(parent
, base
, count
,
2800 generation
= parent
->core
.generation
;
2803 * Exhausted parent chain, iterate.
2806 KKASSERT(chain
== NULL
);
2807 hammer2_spin_unex(&parent
->core
.spin
);
2808 if (key_beg
== key_end
) /* short cut single-key case */
2812 * Stop if we reached the end of the iteration.
2814 if (parent
->bref
.type
!= HAMMER2_BREF_TYPE_INDIRECT
&&
2815 parent
->bref
.type
!= HAMMER2_BREF_TYPE_FREEMAP_NODE
) {
2820 * Calculate next key, stop if we reached the end of the
2821 * iteration, otherwise go up one level and loop.
2823 key_beg
= parent
->bref
.key
+
2824 ((hammer2_key_t
)1 << parent
->bref
.keybits
);
2825 if (key_beg
== 0 || key_beg
> key_end
)
2827 parent
= hammer2_chain_repparent(parentp
, how_maybe
);
2832 * Selected from blockref or in-memory chain.
2835 if (chain
== NULL
) {
2836 hammer2_spin_unex(&parent
->core
.spin
);
2837 if (bcopy
.type
== HAMMER2_BREF_TYPE_INDIRECT
||
2838 bcopy
.type
== HAMMER2_BREF_TYPE_FREEMAP_NODE
) {
2839 chain
= hammer2_chain_get(parent
, generation
,
2842 chain
= hammer2_chain_get(parent
, generation
,
2848 hammer2_chain_ref(chain
);
2849 hammer2_spin_unex(&parent
->core
.spin
);
2852 * chain is referenced but not locked. We must lock the
2853 * chain to obtain definitive state.
2855 if (bcopy
.type
== HAMMER2_BREF_TYPE_INDIRECT
||
2856 bcopy
.type
== HAMMER2_BREF_TYPE_FREEMAP_NODE
) {
2857 hammer2_chain_lock(chain
, how_maybe
);
2859 hammer2_chain_lock(chain
, how
);
2861 KKASSERT(chain
->parent
== parent
);
2863 if (bcmp(&bcopy
, &chain
->bref
, sizeof(bcopy
)) ||
2864 chain
->parent
!= parent
) {
2865 hammer2_chain_unlock(chain
);
2866 hammer2_chain_drop(chain
);
2867 chain
= NULL
; /* SAFETY */
2873 * Skip deleted chains (XXX cache 'i' end-of-block-array? XXX)
2875 * NOTE: Chain's key range is not relevant as there might be
2876 * one-offs within the range that are not deleted.
2878 * NOTE: Lookups can race delete-duplicate because
2879 * delete-duplicate does not lock the parent's core
2880 * (they just use the spinlock on the core).
2882 if (chain
->flags
& HAMMER2_CHAIN_DELETED
) {
2883 kprintf("skip deleted chain %016jx.%02x key=%016jx\n",
2884 chain
->bref
.data_off
, chain
->bref
.type
,
2886 hammer2_chain_unlock(chain
);
2887 hammer2_chain_drop(chain
);
2888 chain
= NULL
; /* SAFETY */
2889 key_beg
= *key_nextp
;
2890 if (key_beg
== 0 || key_beg
> key_end
)
2896 * If the chain element is an indirect block it becomes the new
2897 * parent and we loop on it. We must maintain our top-down locks
2898 * to prevent the flusher from interfering (i.e. doing a
2899 * delete-duplicate and leaving us recursing down a deleted chain).
2901 * The parent always has to be locked with at least RESOLVE_MAYBE
2902 * so we can access its data. It might need a fixup if the caller
2903 * passed incompatible flags. Be careful not to cause a deadlock
2904 * as a data-load requires an exclusive lock.
2906 * If HAMMER2_LOOKUP_MATCHIND is set and the indirect block's key
2907 * range is within the requested key range we return the indirect
2908 * block and do NOT loop. This is usually only used to acquire
2911 if (chain
->bref
.type
== HAMMER2_BREF_TYPE_INDIRECT
||
2912 chain
->bref
.type
== HAMMER2_BREF_TYPE_FREEMAP_NODE
) {
2913 save_mtx
= parent
->lock
;
2914 hammer2_chain_unlock(parent
);
2915 hammer2_chain_drop(parent
);
2916 *parentp
= parent
= chain
;
2917 chain
= NULL
; /* SAFETY */
2922 * All done, return the locked chain.
2924 * If the caller does not want a locked chain, replace the lock with
2925 * a ref. Perhaps this can eventually be optimized to not obtain the
2926 * lock in the first place for situations where the data does not
2927 * need to be resolved.
2929 * NOTE! A chain->error must be tested by the caller upon return.
2930 * *errorp is only set based on issues which occur while
2931 * trying to reach the chain.
2937 * After having issued a lookup we can iterate all matching keys.
2939 * If chain is non-NULL we continue the iteration from just after it's index.
2941 * If chain is NULL we assume the parent was exhausted and continue the
2942 * iteration at the next parent.
2944 * If a fatal error occurs (typically an I/O error), a dummy chain is
2945 * returned with chain->error and error-identifying information set. This
2946 * chain will assert if you try to do anything fancy with it.
2948 * XXX Depending on where the error occurs we should allow continued iteration.
2950 * parent must be locked on entry and remains locked throughout. chain's
2951 * lock status must match flags. Chain is always at least referenced.
2953 * WARNING! The MATCHIND flag does not apply to this function.
2956 hammer2_chain_next(hammer2_chain_t
**parentp
, hammer2_chain_t
*chain
,
2957 hammer2_key_t
*key_nextp
,
2958 hammer2_key_t key_beg
, hammer2_key_t key_end
,
2959 int *errorp
, int flags
)
2961 hammer2_chain_t
*parent
;
2965 * Calculate locking flags for upward recursion.
2967 how_maybe
= HAMMER2_RESOLVE_MAYBE
;
2968 if (flags
& HAMMER2_LOOKUP_SHARED
)
2969 how_maybe
|= HAMMER2_RESOLVE_SHARED
;
2975 * Calculate the next index and recalculate the parent if necessary.
2978 key_beg
= chain
->bref
.key
+
2979 ((hammer2_key_t
)1 << chain
->bref
.keybits
);
2980 hammer2_chain_unlock(chain
);
2981 hammer2_chain_drop(chain
);
2984 * chain invalid past this point, but we can still do a
2985 * pointer comparison w/parent.
2987 * Any scan where the lookup returned degenerate data embedded
2988 * in the inode has an invalid index and must terminate.
2990 if (chain
== parent
)
2992 if (key_beg
== 0 || key_beg
> key_end
)
2995 } else if (parent
->bref
.type
!= HAMMER2_BREF_TYPE_INDIRECT
&&
2996 parent
->bref
.type
!= HAMMER2_BREF_TYPE_FREEMAP_NODE
) {
2998 * We reached the end of the iteration.
3003 * Continue iteration with next parent unless the current
3004 * parent covers the range.
3006 * (This also handles the case of a deleted, empty indirect
3009 key_beg
= parent
->bref
.key
+
3010 ((hammer2_key_t
)1 << parent
->bref
.keybits
);
3011 if (key_beg
== 0 || key_beg
> key_end
)
3013 parent
= hammer2_chain_repparent(parentp
, how_maybe
);
3019 return (hammer2_chain_lookup(parentp
, key_nextp
,
3025 * Caller wishes to iterate chains under parent, loading new chains into
3026 * chainp. Caller must initialize *chainp to NULL and *firstp to 1, and
3027 * then call hammer2_chain_scan() repeatedly until a non-zero return.
3028 * During the scan, *firstp will be set to 0 and (*chainp) will be replaced
3029 * with the returned chain for the scan. The returned *chainp will be
3030 * locked and referenced. Any prior contents will be unlocked and dropped.
3032 * Caller should check the return value. A normal scan EOF will return
3033 * exactly HAMMER2_ERROR_EOF. Any other non-zero value indicates an
3034 * error trying to access parent data. Any error in the returned chain
3035 * must be tested separately by the caller.
3037 * (*chainp) is dropped on each scan, but will only be set if the returned
3038 * element itself can recurse. Leaf elements are NOT resolved, loaded, or
3039 * returned via *chainp. The caller will get their bref only.
3041 * The raw scan function is similar to lookup/next but does not seek to a key.
3042 * Blockrefs are iterated via first_bref = (parent, NULL) and
3043 * next_chain = (parent, bref).
3045 * The passed-in parent must be locked and its data resolved. The function
3046 * nominally returns a locked and referenced *chainp != NULL for chains
3047 * the caller might need to recurse on (and will dipose of any *chainp passed
3048 * in). The caller must check the chain->bref.type either way.
3051 hammer2_chain_scan(hammer2_chain_t
*parent
, hammer2_chain_t
**chainp
,
3052 hammer2_blockref_t
*bref
, int *firstp
,
3056 hammer2_blockref_t
*base
;
3057 hammer2_blockref_t
*bref_ptr
;
3059 hammer2_key_t next_key
;
3060 hammer2_chain_t
*chain
= NULL
;
3062 int how_always
= HAMMER2_RESOLVE_ALWAYS
;
3063 int how_maybe
= HAMMER2_RESOLVE_MAYBE
;
3066 int maxloops
= 300000;
3073 * Scan flags borrowed from lookup.
3075 if (flags
& HAMMER2_LOOKUP_ALWAYS
) {
3076 how_maybe
= how_always
;
3077 how
= HAMMER2_RESOLVE_ALWAYS
;
3078 } else if (flags
& HAMMER2_LOOKUP_NODATA
) {
3079 how
= HAMMER2_RESOLVE_NEVER
;
3081 how
= HAMMER2_RESOLVE_MAYBE
;
3083 if (flags
& HAMMER2_LOOKUP_SHARED
) {
3084 how_maybe
|= HAMMER2_RESOLVE_SHARED
;
3085 how_always
|= HAMMER2_RESOLVE_SHARED
;
3086 how
|= HAMMER2_RESOLVE_SHARED
;
3090 * Calculate key to locate first/next element, unlocking the previous
3091 * element as we go. Be careful, the key calculation can overflow.
3093 * (also reset bref to NULL)
3099 key
= bref
->key
+ ((hammer2_key_t
)1 << bref
->keybits
);
3100 if ((chain
= *chainp
) != NULL
) {
3102 hammer2_chain_unlock(chain
);
3103 hammer2_chain_drop(chain
);
3107 error
|= HAMMER2_ERROR_EOF
;
3113 if (parent
->error
) {
3114 error
= parent
->error
;
3117 if (--maxloops
== 0)
3118 panic("hammer2_chain_scan: maxloops");
3121 * Locate the blockref array. Currently we do a fully associative
3122 * search through the array.
3124 switch(parent
->bref
.type
) {
3125 case HAMMER2_BREF_TYPE_INODE
:
3127 * An inode with embedded data has no sub-chains.
3129 * WARNING! Bulk scan code may pass a static chain marked
3130 * as BREF_TYPE_INODE with a copy of the volume
3131 * root blockset to snapshot the volume.
3133 if (parent
->data
->ipdata
.meta
.op_flags
&
3134 HAMMER2_OPFLAG_DIRECTDATA
) {
3135 error
|= HAMMER2_ERROR_EOF
;
3138 base
= &parent
->data
->ipdata
.u
.blockset
.blockref
[0];
3139 count
= HAMMER2_SET_COUNT
;
3141 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
3142 case HAMMER2_BREF_TYPE_INDIRECT
:
3144 * Optimize indirect blocks in the INITIAL state to avoid
3147 if (parent
->flags
& HAMMER2_CHAIN_INITIAL
) {
3150 if (parent
->data
== NULL
)
3151 panic("parent->data is NULL");
3152 base
= &parent
->data
->npdata
[0];
3154 count
= parent
->bytes
/ sizeof(hammer2_blockref_t
);
3156 case HAMMER2_BREF_TYPE_VOLUME
:
3157 base
= &parent
->data
->voldata
.sroot_blockset
.blockref
[0];
3158 count
= HAMMER2_SET_COUNT
;
3160 case HAMMER2_BREF_TYPE_FREEMAP
:
3161 base
= &parent
->data
->blkset
.blockref
[0];
3162 count
= HAMMER2_SET_COUNT
;
3165 panic("hammer2_chain_lookup: unrecognized blockref type: %d",
3167 base
= NULL
; /* safety */
3168 count
= 0; /* safety */
3172 * Merged scan to find next candidate.
3174 * hammer2_base_*() functions require the parent->core.live_* fields
3175 * to be synchronized.
3177 * We need to hold the spinlock to access the block array and RB tree
3178 * and to interlock chain creation.
3180 if ((parent
->flags
& HAMMER2_CHAIN_COUNTEDBREFS
) == 0)
3181 hammer2_chain_countbrefs(parent
, base
, count
);
3185 hammer2_spin_ex(&parent
->core
.spin
);
3186 chain
= hammer2_combined_find(parent
, base
, count
,
3188 key
, HAMMER2_KEY_MAX
,
3190 generation
= parent
->core
.generation
;
3193 * Exhausted parent chain, we're done.
3195 if (bref_ptr
== NULL
) {
3196 hammer2_spin_unex(&parent
->core
.spin
);
3197 KKASSERT(chain
== NULL
);
3198 error
|= HAMMER2_ERROR_EOF
;
3203 * Copy into the supplied stack-based blockref.
3208 * Selected from blockref or in-memory chain.
3210 if (chain
== NULL
) {
3211 switch(bref
->type
) {
3212 case HAMMER2_BREF_TYPE_INODE
:
3213 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
3214 case HAMMER2_BREF_TYPE_INDIRECT
:
3215 case HAMMER2_BREF_TYPE_VOLUME
:
3216 case HAMMER2_BREF_TYPE_FREEMAP
:
3218 * Recursion, always get the chain
3220 hammer2_spin_unex(&parent
->core
.spin
);
3221 chain
= hammer2_chain_get(parent
, generation
,
3228 * No recursion, do not waste time instantiating
3229 * a chain, just iterate using the bref.
3231 hammer2_spin_unex(&parent
->core
.spin
);
3236 * Recursion or not we need the chain in order to supply
3239 hammer2_chain_ref(chain
);
3240 hammer2_spin_unex(&parent
->core
.spin
);
3241 hammer2_chain_lock(chain
, how
);
3244 (bcmp(bref
, &chain
->bref
, sizeof(*bref
)) ||
3245 chain
->parent
!= parent
)) {
3246 hammer2_chain_unlock(chain
);
3247 hammer2_chain_drop(chain
);
3253 * Skip deleted chains (XXX cache 'i' end-of-block-array? XXX)
3255 * NOTE: chain's key range is not relevant as there might be
3256 * one-offs within the range that are not deleted.
3258 * NOTE: XXX this could create problems with scans used in
3259 * situations other than mount-time recovery.
3261 * NOTE: Lookups can race delete-duplicate because
3262 * delete-duplicate does not lock the parent's core
3263 * (they just use the spinlock on the core).
3265 if (chain
&& (chain
->flags
& HAMMER2_CHAIN_DELETED
)) {
3266 hammer2_chain_unlock(chain
);
3267 hammer2_chain_drop(chain
);
3272 error
|= HAMMER2_ERROR_EOF
;
3280 * All done, return the bref or NULL, supply chain if necessary.
3288 * Create and return a new hammer2 system memory structure of the specified
3289 * key, type and size and insert it under (*parentp). This is a full
3290 * insertion, based on the supplied key/keybits, and may involve creating
3291 * indirect blocks and moving other chains around via delete/duplicate.
3293 * THE CALLER MUST HAVE ALREADY PROPERLY SEEKED (*parentp) TO THE INSERTION
3294 * POINT SANS ANY REQUIRED INDIRECT BLOCK CREATIONS DUE TO THE ARRAY BEING
3295 * FULL. This typically means that the caller is creating the chain after
3296 * doing a hammer2_chain_lookup().
3298 * (*parentp) must be exclusive locked and may be replaced on return
3299 * depending on how much work the function had to do.
3301 * (*parentp) must not be errored or this function will assert.
3303 * (*chainp) usually starts out NULL and returns the newly created chain,
3304 * but if the caller desires the caller may allocate a disconnected chain
3305 * and pass it in instead.
3307 * This function should NOT be used to insert INDIRECT blocks. It is
3308 * typically used to create/insert inodes and data blocks.
3310 * Caller must pass-in an exclusively locked parent the new chain is to
3311 * be inserted under, and optionally pass-in a disconnected, exclusively
3312 * locked chain to insert (else we create a new chain). The function will
3313 * adjust (*parentp) as necessary, create or connect the chain, and
3314 * return an exclusively locked chain in *chainp.
3316 * When creating a PFSROOT inode under the super-root, pmp is typically NULL
3317 * and will be reassigned.
3319 * NOTE: returns HAMMER_ERROR_* flags
3322 hammer2_chain_create(hammer2_chain_t
**parentp
, hammer2_chain_t
**chainp
,
3323 hammer2_pfs_t
*pmp
, int methods
,
3324 hammer2_key_t key
, int keybits
, int type
, size_t bytes
,
3325 hammer2_tid_t mtid
, hammer2_off_t dedup_off
, int flags
)
3328 hammer2_chain_t
*chain
;
3329 hammer2_chain_t
*parent
;
3330 hammer2_blockref_t
*base
;
3331 hammer2_blockref_t dummy
;
3335 int maxloops
= 300000;
3338 * Topology may be crossing a PFS boundary.
3341 KKASSERT(hammer2_mtx_owned(&parent
->lock
));
3342 KKASSERT(parent
->error
== 0);
3346 if (chain
== NULL
) {
3348 * First allocate media space and construct the dummy bref,
3349 * then allocate the in-memory chain structure. Set the
3350 * INITIAL flag for fresh chains which do not have embedded
3353 * XXX for now set the check mode of the child based on
3354 * the parent or, if the parent is an inode, the
3355 * specification in the inode.
3357 bzero(&dummy
, sizeof(dummy
));
3360 dummy
.keybits
= keybits
;
3361 dummy
.data_off
= hammer2_getradix(bytes
);
3364 * Inherit methods from parent by default. Primarily used
3365 * for BREF_TYPE_DATA. Non-data types *must* be set to
3366 * a non-NONE check algorithm.
3369 dummy
.methods
= parent
->bref
.methods
;
3371 dummy
.methods
= (uint8_t)methods
;
3373 if (type
!= HAMMER2_BREF_TYPE_DATA
&&
3374 HAMMER2_DEC_CHECK(dummy
.methods
) == HAMMER2_CHECK_NONE
) {
3376 HAMMER2_ENC_CHECK(HAMMER2_CHECK_DEFAULT
);
3379 chain
= hammer2_chain_alloc(hmp
, pmp
, &dummy
);
3382 * Lock the chain manually, chain_lock will load the chain
3383 * which we do NOT want to do. (note: chain->refs is set
3384 * to 1 by chain_alloc() for us, but lockcnt is not).
3387 hammer2_mtx_ex(&chain
->lock
);
3389 ++curthread
->td_tracker
;
3392 * Set INITIAL to optimize I/O. The flag will generally be
3393 * processed when we call hammer2_chain_modify().
3395 * Recalculate bytes to reflect the actual media block
3396 * allocation. Handle special case radix 0 == 0 bytes.
3398 bytes
= (size_t)(chain
->bref
.data_off
& HAMMER2_OFF_MASK_RADIX
);
3400 bytes
= (hammer2_off_t
)1 << bytes
;
3401 chain
->bytes
= bytes
;
3404 case HAMMER2_BREF_TYPE_VOLUME
:
3405 case HAMMER2_BREF_TYPE_FREEMAP
:
3406 panic("hammer2_chain_create: called with volume type");
3408 case HAMMER2_BREF_TYPE_INDIRECT
:
3409 panic("hammer2_chain_create: cannot be used to"
3410 "create indirect block");
3412 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
3413 panic("hammer2_chain_create: cannot be used to"
3414 "create freemap root or node");
3416 case HAMMER2_BREF_TYPE_FREEMAP_LEAF
:
3417 KKASSERT(bytes
== sizeof(chain
->data
->bmdata
));
3419 case HAMMER2_BREF_TYPE_DIRENT
:
3420 case HAMMER2_BREF_TYPE_INODE
:
3421 case HAMMER2_BREF_TYPE_DATA
:
3424 * leave chain->data NULL, set INITIAL
3426 KKASSERT(chain
->data
== NULL
);
3427 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_INITIAL
);
3432 * We are reattaching a previously deleted chain, possibly
3433 * under a new parent and possibly with a new key/keybits.
3434 * The chain does not have to be in a modified state. The
3435 * UPDATE flag will be set later on in this routine.
3437 * Do NOT mess with the current state of the INITIAL flag.
3439 chain
->bref
.key
= key
;
3440 chain
->bref
.keybits
= keybits
;
3441 if (chain
->flags
& HAMMER2_CHAIN_DELETED
)
3442 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_DELETED
);
3443 KKASSERT(chain
->parent
== NULL
);
3447 * Set the appropriate bref flag if requested.
3449 * NOTE! Callers can call this function to move chains without
3450 * knowing about special flags, so don't clear bref flags
3453 if (flags
& HAMMER2_INSERT_PFSROOT
)
3454 chain
->bref
.flags
|= HAMMER2_BREF_FLAG_PFSROOT
;
3457 * Calculate how many entries we have in the blockref array and
3458 * determine if an indirect block is required.
3461 if (--maxloops
== 0)
3462 panic("hammer2_chain_create: maxloops");
3464 switch(parent
->bref
.type
) {
3465 case HAMMER2_BREF_TYPE_INODE
:
3466 if ((parent
->data
->ipdata
.meta
.op_flags
&
3467 HAMMER2_OPFLAG_DIRECTDATA
) != 0) {
3468 kprintf("hammer2: parent set for direct-data! "
3469 "pkey=%016jx ckey=%016jx\n",
3473 KKASSERT((parent
->data
->ipdata
.meta
.op_flags
&
3474 HAMMER2_OPFLAG_DIRECTDATA
) == 0);
3475 KKASSERT(parent
->data
!= NULL
);
3476 base
= &parent
->data
->ipdata
.u
.blockset
.blockref
[0];
3477 count
= HAMMER2_SET_COUNT
;
3479 case HAMMER2_BREF_TYPE_INDIRECT
:
3480 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
3481 if (parent
->flags
& HAMMER2_CHAIN_INITIAL
)
3484 base
= &parent
->data
->npdata
[0];
3485 count
= parent
->bytes
/ sizeof(hammer2_blockref_t
);
3487 case HAMMER2_BREF_TYPE_VOLUME
:
3488 KKASSERT(parent
->data
!= NULL
);
3489 base
= &parent
->data
->voldata
.sroot_blockset
.blockref
[0];
3490 count
= HAMMER2_SET_COUNT
;
3492 case HAMMER2_BREF_TYPE_FREEMAP
:
3493 KKASSERT(parent
->data
!= NULL
);
3494 base
= &parent
->data
->blkset
.blockref
[0];
3495 count
= HAMMER2_SET_COUNT
;
3498 panic("hammer2_chain_create: unrecognized blockref type: %d",
3506 * Make sure we've counted the brefs
3508 if ((parent
->flags
& HAMMER2_CHAIN_COUNTEDBREFS
) == 0)
3509 hammer2_chain_countbrefs(parent
, base
, count
);
3511 KASSERT(parent
->core
.live_count
>= 0 &&
3512 parent
->core
.live_count
<= count
,
3513 ("bad live_count %d/%d (%02x, %d)",
3514 parent
->core
.live_count
, count
,
3515 parent
->bref
.type
, parent
->bytes
));
3518 * If no free blockref could be found we must create an indirect
3519 * block and move a number of blockrefs into it. With the parent
3520 * locked we can safely lock each child in order to delete+duplicate
3521 * it without causing a deadlock.
3523 * This may return the new indirect block or the old parent depending
3524 * on where the key falls. NULL is returned on error.
3526 if (parent
->core
.live_count
== count
) {
3527 hammer2_chain_t
*nparent
;
3529 KKASSERT((flags
& HAMMER2_INSERT_SAMEPARENT
) == 0);
3531 nparent
= hammer2_chain_create_indirect(parent
, key
, keybits
,
3532 mtid
, type
, &error
);
3533 if (nparent
== NULL
) {
3535 hammer2_chain_drop(chain
);
3539 if (parent
!= nparent
) {
3540 hammer2_chain_unlock(parent
);
3541 hammer2_chain_drop(parent
);
3542 parent
= *parentp
= nparent
;
3547 if (chain
->flags
& HAMMER2_CHAIN_DELETED
)
3548 kprintf("Inserting deleted chain @%016jx\n",
3552 * Link the chain into its parent.
3554 if (chain
->parent
!= NULL
)
3555 panic("hammer2: hammer2_chain_create: chain already connected");
3556 KKASSERT(chain
->parent
== NULL
);
3557 KKASSERT(parent
->core
.live_count
< count
);
3558 hammer2_chain_insert(parent
, chain
,
3559 HAMMER2_CHAIN_INSERT_SPIN
|
3560 HAMMER2_CHAIN_INSERT_LIVE
,
3565 * Mark the newly created chain modified. This will cause
3566 * UPDATE to be set and process the INITIAL flag.
3568 * Device buffers are not instantiated for DATA elements
3569 * as these are handled by logical buffers.
3571 * Indirect and freemap node indirect blocks are handled
3572 * by hammer2_chain_create_indirect() and not by this
3575 * Data for all other bref types is expected to be
3576 * instantiated (INODE, LEAF).
3578 switch(chain
->bref
.type
) {
3579 case HAMMER2_BREF_TYPE_DATA
:
3580 case HAMMER2_BREF_TYPE_FREEMAP_LEAF
:
3581 case HAMMER2_BREF_TYPE_DIRENT
:
3582 case HAMMER2_BREF_TYPE_INODE
:
3583 error
= hammer2_chain_modify(chain
, mtid
, dedup_off
,
3584 HAMMER2_MODIFY_OPTDATA
);
3588 * Remaining types are not supported by this function.
3589 * In particular, INDIRECT and LEAF_NODE types are
3590 * handled by create_indirect().
3592 panic("hammer2_chain_create: bad type: %d",
3599 * When reconnecting a chain we must set UPDATE and
3600 * setflush so the flush recognizes that it must update
3601 * the bref in the parent.
3603 if ((chain
->flags
& HAMMER2_CHAIN_UPDATE
) == 0)
3604 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_UPDATE
);
3608 * We must setflush(parent) to ensure that it recurses through to
3609 * chain. setflush(chain) might not work because ONFLUSH is possibly
3610 * already set in the chain (so it won't recurse up to set it in the
3613 hammer2_chain_setflush(parent
);
3622 * Move the chain from its old parent to a new parent. The chain must have
3623 * already been deleted or already disconnected (or never associated) with
3624 * a parent. The chain is reassociated with the new parent and the deleted
3625 * flag will be cleared (no longer deleted). The chain's modification state
3628 * THE CALLER MUST HAVE ALREADY PROPERLY SEEKED (parent) TO THE INSERTION
3629 * POINT SANS ANY REQUIRED INDIRECT BLOCK CREATIONS DUE TO THE ARRAY BEING
3630 * FULL. This typically means that the caller is creating the chain after
3631 * doing a hammer2_chain_lookup().
3633 * Neither (parent) or (chain) can be errored.
3635 * If (parent) is non-NULL then the chain is inserted under the parent.
3637 * If (parent) is NULL then the newly duplicated chain is not inserted
3638 * anywhere, similar to if it had just been chain_alloc()'d (suitable for
3639 * passing into hammer2_chain_create() after this function returns).
3641 * WARNING! This function calls create which means it can insert indirect
3642 * blocks. This can cause other unrelated chains in the parent to
3643 * be moved to a newly inserted indirect block in addition to the
3647 hammer2_chain_rename(hammer2_chain_t
**parentp
, hammer2_chain_t
*chain
,
3648 hammer2_tid_t mtid
, int flags
)
3650 hammer2_blockref_t
*bref
;
3652 hammer2_chain_t
*parent
;
3656 * WARNING! We should never resolve DATA to device buffers
3657 * (XXX allow it if the caller did?), and since
3658 * we currently do not have the logical buffer cache
3659 * buffer in-hand to fix its cached physical offset
3660 * we also force the modify code to not COW it. XXX
3663 KKASSERT(chain
->parent
== NULL
);
3664 KKASSERT(chain
->error
== 0);
3667 * Now create a duplicate of the chain structure, associating
3668 * it with the same core, making it the same size, pointing it
3669 * to the same bref (the same media block).
3671 * NOTE: Handle special radix == 0 case (means 0 bytes).
3673 bref
= &chain
->bref
;
3674 bytes
= (size_t)(bref
->data_off
& HAMMER2_OFF_MASK_RADIX
);
3676 bytes
= (hammer2_off_t
)1 << bytes
;
3679 * If parent is not NULL the duplicated chain will be entered under
3680 * the parent and the UPDATE bit set to tell flush to update
3683 * We must setflush(parent) to ensure that it recurses through to
3684 * chain. setflush(chain) might not work because ONFLUSH is possibly
3685 * already set in the chain (so it won't recurse up to set it in the
3688 * Having both chains locked is extremely important for atomicy.
3690 if (parentp
&& (parent
= *parentp
) != NULL
) {
3691 KKASSERT(hammer2_mtx_owned(&parent
->lock
));
3692 KKASSERT(parent
->refs
> 0);
3693 KKASSERT(parent
->error
== 0);
3695 hammer2_chain_create(parentp
, &chain
,
3696 chain
->pmp
, HAMMER2_METH_DEFAULT
,
3697 bref
->key
, bref
->keybits
, bref
->type
,
3698 chain
->bytes
, mtid
, 0, flags
);
3699 KKASSERT(chain
->flags
& HAMMER2_CHAIN_UPDATE
);
3700 hammer2_chain_setflush(*parentp
);
3705 * Helper function for deleting chains.
3707 * The chain is removed from the live view (the RBTREE) as well as the parent's
3708 * blockmap. Both chain and its parent must be locked.
3710 * parent may not be errored. chain can be errored.
3713 _hammer2_chain_delete_helper(hammer2_chain_t
*parent
, hammer2_chain_t
*chain
,
3714 hammer2_tid_t mtid
, int flags
)
3719 KKASSERT((chain
->flags
& (HAMMER2_CHAIN_DELETED
|
3720 HAMMER2_CHAIN_FICTITIOUS
)) == 0);
3721 KKASSERT(chain
->parent
== parent
);
3724 if (chain
->flags
& HAMMER2_CHAIN_BMAPPED
) {
3726 * Chain is blockmapped, so there must be a parent.
3727 * Atomically remove the chain from the parent and remove
3728 * the blockmap entry. The parent must be set modified
3729 * to remove the blockmap entry.
3731 hammer2_blockref_t
*base
;
3734 KKASSERT(parent
!= NULL
);
3735 KKASSERT(parent
->error
== 0);
3736 KKASSERT((parent
->flags
& HAMMER2_CHAIN_INITIAL
) == 0);
3737 error
= hammer2_chain_modify(parent
, mtid
, 0, 0);
3742 * Calculate blockmap pointer
3744 KKASSERT(chain
->flags
& HAMMER2_CHAIN_ONRBTREE
);
3745 hammer2_spin_ex(&chain
->core
.spin
);
3746 hammer2_spin_ex(&parent
->core
.spin
);
3748 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_DELETED
);
3749 atomic_add_int(&parent
->core
.live_count
, -1);
3750 ++parent
->core
.generation
;
3751 RB_REMOVE(hammer2_chain_tree
, &parent
->core
.rbtree
, chain
);
3752 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_ONRBTREE
);
3753 --parent
->core
.chain_count
;
3754 chain
->parent
= NULL
;
3756 switch(parent
->bref
.type
) {
3757 case HAMMER2_BREF_TYPE_INODE
:
3759 * Access the inode's block array. However, there
3760 * is no block array if the inode is flagged
3764 (parent
->data
->ipdata
.meta
.op_flags
&
3765 HAMMER2_OPFLAG_DIRECTDATA
) == 0) {
3767 &parent
->data
->ipdata
.u
.blockset
.blockref
[0];
3771 count
= HAMMER2_SET_COUNT
;
3773 case HAMMER2_BREF_TYPE_INDIRECT
:
3774 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
3776 base
= &parent
->data
->npdata
[0];
3779 count
= parent
->bytes
/ sizeof(hammer2_blockref_t
);
3781 case HAMMER2_BREF_TYPE_VOLUME
:
3782 base
= &parent
->data
->voldata
.
3783 sroot_blockset
.blockref
[0];
3784 count
= HAMMER2_SET_COUNT
;
3786 case HAMMER2_BREF_TYPE_FREEMAP
:
3787 base
= &parent
->data
->blkset
.blockref
[0];
3788 count
= HAMMER2_SET_COUNT
;
3793 panic("hammer2_flush_pass2: "
3794 "unrecognized blockref type: %d",
3799 * delete blockmapped chain from its parent.
3801 * The parent is not affected by any statistics in chain
3802 * which are pending synchronization. That is, there is
3803 * nothing to undo in the parent since they have not yet
3804 * been incorporated into the parent.
3806 * The parent is affected by statistics stored in inodes.
3807 * Those have already been synchronized, so they must be
3808 * undone. XXX split update possible w/delete in middle?
3811 hammer2_base_delete(parent
, base
, count
, chain
);
3813 hammer2_spin_unex(&parent
->core
.spin
);
3814 hammer2_spin_unex(&chain
->core
.spin
);
3815 } else if (chain
->flags
& HAMMER2_CHAIN_ONRBTREE
) {
3817 * Chain is not blockmapped but a parent is present.
3818 * Atomically remove the chain from the parent. There is
3819 * no blockmap entry to remove.
3821 * Because chain was associated with a parent but not
3822 * synchronized, the chain's *_count_up fields contain
3823 * inode adjustment statistics which must be undone.
3825 hammer2_spin_ex(&chain
->core
.spin
);
3826 hammer2_spin_ex(&parent
->core
.spin
);
3827 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_DELETED
);
3828 atomic_add_int(&parent
->core
.live_count
, -1);
3829 ++parent
->core
.generation
;
3830 RB_REMOVE(hammer2_chain_tree
, &parent
->core
.rbtree
, chain
);
3831 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_ONRBTREE
);
3832 --parent
->core
.chain_count
;
3833 chain
->parent
= NULL
;
3834 hammer2_spin_unex(&parent
->core
.spin
);
3835 hammer2_spin_unex(&chain
->core
.spin
);
3838 * Chain is not blockmapped and has no parent. This
3839 * is a degenerate case.
3841 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_DELETED
);
3848 * Create an indirect block that covers one or more of the elements in the
3849 * current parent. Either returns the existing parent with no locking or
3850 * ref changes or returns the new indirect block locked and referenced
3851 * and leaving the original parent lock/ref intact as well.
3853 * If an error occurs, NULL is returned and *errorp is set to the H2 error.
3855 * The returned chain depends on where the specified key falls.
3857 * The key/keybits for the indirect mode only needs to follow three rules:
3859 * (1) That all elements underneath it fit within its key space and
3861 * (2) That all elements outside it are outside its key space.
3863 * (3) When creating the new indirect block any elements in the current
3864 * parent that fit within the new indirect block's keyspace must be
3865 * moved into the new indirect block.
3867 * (4) The keyspace chosen for the inserted indirect block CAN cover a wider
3868 * keyspace the the current parent, but lookup/iteration rules will
3869 * ensure (and must ensure) that rule (2) for all parents leading up
3870 * to the nearest inode or the root volume header is adhered to. This
3871 * is accomplished by always recursing through matching keyspaces in
3872 * the hammer2_chain_lookup() and hammer2_chain_next() API.
3874 * The current implementation calculates the current worst-case keyspace by
3875 * iterating the current parent and then divides it into two halves, choosing
3876 * whichever half has the most elements (not necessarily the half containing
3877 * the requested key).
3879 * We can also opt to use the half with the least number of elements. This
3880 * causes lower-numbered keys (aka logical file offsets) to recurse through
3881 * fewer indirect blocks and higher-numbered keys to recurse through more.
3882 * This also has the risk of not moving enough elements to the new indirect
3883 * block and being forced to create several indirect blocks before the element
3886 * Must be called with an exclusively locked parent.
3888 * NOTE: *errorp set to HAMMER_ERROR_* flags
3890 static int hammer2_chain_indkey_freemap(hammer2_chain_t
*parent
,
3891 hammer2_key_t
*keyp
, int keybits
,
3892 hammer2_blockref_t
*base
, int count
);
3893 static int hammer2_chain_indkey_file(hammer2_chain_t
*parent
,
3894 hammer2_key_t
*keyp
, int keybits
,
3895 hammer2_blockref_t
*base
, int count
,
3897 static int hammer2_chain_indkey_dir(hammer2_chain_t
*parent
,
3898 hammer2_key_t
*keyp
, int keybits
,
3899 hammer2_blockref_t
*base
, int count
,
3903 hammer2_chain_create_indirect(hammer2_chain_t
*parent
,
3904 hammer2_key_t create_key
, int create_bits
,
3905 hammer2_tid_t mtid
, int for_type
, int *errorp
)
3908 hammer2_blockref_t
*base
;
3909 hammer2_blockref_t
*bref
;
3910 hammer2_blockref_t bcopy
;
3911 hammer2_chain_t
*chain
;
3912 hammer2_chain_t
*ichain
;
3913 hammer2_chain_t dummy
;
3914 hammer2_key_t key
= create_key
;
3915 hammer2_key_t key_beg
;
3916 hammer2_key_t key_end
;
3917 hammer2_key_t key_next
;
3918 int keybits
= create_bits
;
3926 int maxloops
= 300000;
3929 * Calculate the base blockref pointer or NULL if the chain
3930 * is known to be empty. We need to calculate the array count
3931 * for RB lookups either way.
3934 KKASSERT(hammer2_mtx_owned(&parent
->lock
));
3937 * Pre-modify the parent now to avoid having to deal with error
3938 * processing if we tried to later (in the middle of our loop).
3940 *errorp
= hammer2_chain_modify(parent
, mtid
, 0, 0);
3942 kprintf("hammer2_create_indirect: error %08x %s\n",
3943 *errorp
, hammer2_error_str(*errorp
));
3947 /*hammer2_chain_modify(&parent, HAMMER2_MODIFY_OPTDATA);*/
3948 base
= hammer2_chain_base_and_count(parent
, &count
);
3951 * dummy used in later chain allocation (no longer used for lookups).
3953 bzero(&dummy
, sizeof(dummy
));
3956 * How big should our new indirect block be? It has to be at least
3957 * as large as its parent for splits to work properly.
3959 * The freemap uses a specific indirect block size. The number of
3960 * levels are built dynamically and ultimately depend on the size
3961 * volume. Because freemap blocks are taken from the reserved areas
3962 * of the volume our goal is efficiency (fewer levels) and not so
3963 * much to save disk space.
3965 * The first indirect block level for a directory usually uses
3966 * HAMMER2_IND_BYTES_MIN (4KB = 32 directory entries). Due to
3967 * the hash mechanism, this typically gives us a nominal
3968 * 32 * 4 entries with one level of indirection.
3970 * We use HAMMER2_IND_BYTES_NOM (16KB = 128 blockrefs) for FILE
3971 * indirect blocks. The initial 4 entries in the inode gives us
3972 * 256KB. Up to 4 indirect blocks gives us 32MB. Three levels
3973 * of indirection gives us 137GB, and so forth. H2 can support
3974 * huge file sizes but they are not typical, so we try to stick
3975 * with compactness and do not use a larger indirect block size.
3977 * We could use 64KB (PBUFSIZE), giving us 512 blockrefs, but
3978 * due to the way indirect blocks are created this usually winds
3979 * up being extremely inefficient for small files. Even though
3980 * 16KB requires more levels of indirection for very large files,
3981 * the 16KB records can be ganged together into 64KB DIOs.
3983 if (for_type
== HAMMER2_BREF_TYPE_FREEMAP_NODE
||
3984 for_type
== HAMMER2_BREF_TYPE_FREEMAP_LEAF
) {
3985 nbytes
= HAMMER2_FREEMAP_LEVELN_PSIZE
;
3986 } else if (parent
->bref
.type
== HAMMER2_BREF_TYPE_INODE
) {
3987 if (parent
->data
->ipdata
.meta
.type
==
3988 HAMMER2_OBJTYPE_DIRECTORY
)
3989 nbytes
= HAMMER2_IND_BYTES_MIN
; /* 4KB = 32 entries */
3991 nbytes
= HAMMER2_IND_BYTES_NOM
; /* 16KB = ~8MB file */
3994 nbytes
= HAMMER2_IND_BYTES_NOM
;
3996 if (nbytes
< count
* sizeof(hammer2_blockref_t
)) {
3997 KKASSERT(for_type
!= HAMMER2_BREF_TYPE_FREEMAP_NODE
&&
3998 for_type
!= HAMMER2_BREF_TYPE_FREEMAP_LEAF
);
3999 nbytes
= count
* sizeof(hammer2_blockref_t
);
4001 ncount
= nbytes
/ sizeof(hammer2_blockref_t
);
4004 * When creating an indirect block for a freemap node or leaf
4005 * the key/keybits must be fitted to static radix levels because
4006 * particular radix levels use particular reserved blocks in the
4009 * This routine calculates the key/radix of the indirect block
4010 * we need to create, and whether it is on the high-side or the
4014 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
4015 case HAMMER2_BREF_TYPE_FREEMAP_LEAF
:
4016 keybits
= hammer2_chain_indkey_freemap(parent
, &key
, keybits
,
4019 case HAMMER2_BREF_TYPE_DATA
:
4020 keybits
= hammer2_chain_indkey_file(parent
, &key
, keybits
,
4021 base
, count
, ncount
);
4023 case HAMMER2_BREF_TYPE_DIRENT
:
4024 case HAMMER2_BREF_TYPE_INODE
:
4025 keybits
= hammer2_chain_indkey_dir(parent
, &key
, keybits
,
4026 base
, count
, ncount
);
4029 panic("illegal indirect block for bref type %d", for_type
);
4034 * Normalize the key for the radix being represented, keeping the
4035 * high bits and throwing away the low bits.
4037 key
&= ~(((hammer2_key_t
)1 << keybits
) - 1);
4040 * Ok, create our new indirect block
4042 if (for_type
== HAMMER2_BREF_TYPE_FREEMAP_NODE
||
4043 for_type
== HAMMER2_BREF_TYPE_FREEMAP_LEAF
) {
4044 dummy
.bref
.type
= HAMMER2_BREF_TYPE_FREEMAP_NODE
;
4046 dummy
.bref
.type
= HAMMER2_BREF_TYPE_INDIRECT
;
4048 dummy
.bref
.key
= key
;
4049 dummy
.bref
.keybits
= keybits
;
4050 dummy
.bref
.data_off
= hammer2_getradix(nbytes
);
4051 dummy
.bref
.methods
=
4052 HAMMER2_ENC_CHECK(HAMMER2_DEC_CHECK(parent
->bref
.methods
)) |
4053 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE
);
4055 ichain
= hammer2_chain_alloc(hmp
, parent
->pmp
, &dummy
.bref
);
4056 atomic_set_int(&ichain
->flags
, HAMMER2_CHAIN_INITIAL
);
4057 hammer2_chain_lock(ichain
, HAMMER2_RESOLVE_MAYBE
);
4058 /* ichain has one ref at this point */
4061 * We have to mark it modified to allocate its block, but use
4062 * OPTDATA to allow it to remain in the INITIAL state. Otherwise
4063 * it won't be acted upon by the flush code.
4065 *errorp
= hammer2_chain_modify(ichain
, mtid
, 0, HAMMER2_MODIFY_OPTDATA
);
4067 kprintf("hammer2_alloc_indirect: error %08x %s\n",
4068 *errorp
, hammer2_error_str(*errorp
));
4069 hammer2_chain_unlock(ichain
);
4070 hammer2_chain_drop(ichain
);
4075 * Iterate the original parent and move the matching brefs into
4076 * the new indirect block.
4078 * XXX handle flushes.
4081 key_end
= HAMMER2_KEY_MAX
;
4082 key_next
= 0; /* avoid gcc warnings */
4083 hammer2_spin_ex(&parent
->core
.spin
);
4089 * Parent may have been modified, relocating its block array.
4090 * Reload the base pointer.
4092 base
= hammer2_chain_base_and_count(parent
, &count
);
4094 if (++loops
> 100000) {
4095 hammer2_spin_unex(&parent
->core
.spin
);
4096 panic("excessive loops r=%d p=%p base/count %p:%d %016jx\n",
4097 reason
, parent
, base
, count
, key_next
);
4101 * NOTE: spinlock stays intact, returned chain (if not NULL)
4102 * is not referenced or locked which means that we
4103 * cannot safely check its flagged / deletion status
4106 chain
= hammer2_combined_find(parent
, base
, count
,
4110 generation
= parent
->core
.generation
;
4113 key_next
= bref
->key
+ ((hammer2_key_t
)1 << bref
->keybits
);
4116 * Skip keys that are not within the key/radix of the new
4117 * indirect block. They stay in the parent.
4119 if ((~(((hammer2_key_t
)1 << keybits
) - 1) &
4120 (key
^ bref
->key
)) != 0) {
4121 goto next_key_spinlocked
;
4125 * Load the new indirect block by acquiring the related
4126 * chains (potentially from media as it might not be
4127 * in-memory). Then move it to the new parent (ichain).
4129 * chain is referenced but not locked. We must lock the
4130 * chain to obtain definitive state.
4135 * Use chain already present in the RBTREE
4137 hammer2_chain_ref(chain
);
4138 hammer2_spin_unex(&parent
->core
.spin
);
4139 hammer2_chain_lock(chain
, HAMMER2_RESOLVE_NEVER
);
4142 * Get chain for blockref element. _get returns NULL
4143 * on insertion race.
4145 hammer2_spin_unex(&parent
->core
.spin
);
4146 chain
= hammer2_chain_get(parent
, generation
, &bcopy
,
4147 HAMMER2_RESOLVE_NEVER
);
4148 if (chain
== NULL
) {
4150 hammer2_spin_ex(&parent
->core
.spin
);
4156 * This is always live so if the chain has been deleted
4157 * we raced someone and we have to retry.
4159 * NOTE: Lookups can race delete-duplicate because
4160 * delete-duplicate does not lock the parent's core
4161 * (they just use the spinlock on the core).
4163 * (note reversed logic for this one)
4165 if (bcmp(&bcopy
, &chain
->bref
, sizeof(bcopy
)) ||
4166 chain
->parent
!= parent
||
4167 (chain
->flags
& HAMMER2_CHAIN_DELETED
)) {
4168 hammer2_chain_unlock(chain
);
4169 hammer2_chain_drop(chain
);
4170 if (hammer2_debug
& 0x0040) {
4171 kprintf("LOST PARENT RETRY "
4172 "RETRY (%p,%p)->%p %08x\n",
4173 parent
, chain
->parent
, chain
, chain
->flags
);
4175 hammer2_spin_ex(&parent
->core
.spin
);
4180 * Shift the chain to the indirect block.
4182 * WARNING! No reason for us to load chain data, pass NOSTATS
4183 * to prevent delete/insert from trying to access
4184 * inode stats (and thus asserting if there is no
4185 * chain->data loaded).
4187 * WARNING! The (parent, chain) deletion may modify the parent
4188 * and invalidate the base pointer.
4190 * WARNING! Parent must already be marked modified, so we
4191 * can assume that chain_delete always suceeds.
4193 * WARNING! hammer2_chain_repchange() does not have to be
4194 * called (and doesn't work anyway because we are
4195 * only doing a partial shift). A recursion that is
4196 * in-progress can continue at the current parent
4197 * and will be able to properly find its next key.
4199 error
= hammer2_chain_delete(parent
, chain
, mtid
, 0);
4200 KKASSERT(error
== 0);
4201 hammer2_chain_rename(&ichain
, chain
, mtid
, 0);
4202 hammer2_chain_unlock(chain
);
4203 hammer2_chain_drop(chain
);
4204 KKASSERT(parent
->refs
> 0);
4206 base
= NULL
; /* safety */
4207 hammer2_spin_ex(&parent
->core
.spin
);
4208 next_key_spinlocked
:
4209 if (--maxloops
== 0)
4210 panic("hammer2_chain_create_indirect: maxloops");
4212 if (key_next
== 0 || key_next
> key_end
)
4217 hammer2_spin_unex(&parent
->core
.spin
);
4220 * Insert the new indirect block into the parent now that we've
4221 * cleared out some entries in the parent. We calculated a good
4222 * insertion index in the loop above (ichain->index).
4224 * We don't have to set UPDATE here because we mark ichain
4225 * modified down below (so the normal modified -> flush -> set-moved
4226 * sequence applies).
4228 * The insertion shouldn't race as this is a completely new block
4229 * and the parent is locked.
4231 base
= NULL
; /* safety, parent modify may change address */
4232 KKASSERT((ichain
->flags
& HAMMER2_CHAIN_ONRBTREE
) == 0);
4233 KKASSERT(parent
->core
.live_count
< count
);
4234 hammer2_chain_insert(parent
, ichain
,
4235 HAMMER2_CHAIN_INSERT_SPIN
|
4236 HAMMER2_CHAIN_INSERT_LIVE
,
4240 * Make sure flushes propogate after our manual insertion.
4242 hammer2_chain_setflush(ichain
);
4243 hammer2_chain_setflush(parent
);
4246 * Figure out what to return.
4248 if (~(((hammer2_key_t
)1 << keybits
) - 1) &
4249 (create_key
^ key
)) {
4251 * Key being created is outside the key range,
4252 * return the original parent.
4254 hammer2_chain_unlock(ichain
);
4255 hammer2_chain_drop(ichain
);
4258 * Otherwise its in the range, return the new parent.
4259 * (leave both the new and old parent locked).
4268 * Do maintenance on an indirect chain. Both parent and chain are locked.
4270 * Returns non-zero if (chain) is deleted, either due to being empty or
4271 * because its children were safely moved into the parent.
4274 hammer2_chain_indirect_maintenance(hammer2_chain_t
*parent
,
4275 hammer2_chain_t
*chain
)
4277 hammer2_blockref_t
*chain_base
;
4278 hammer2_blockref_t
*base
;
4279 hammer2_blockref_t
*bref
;
4280 hammer2_blockref_t bcopy
;
4281 hammer2_key_t key_next
;
4282 hammer2_key_t key_beg
;
4283 hammer2_key_t key_end
;
4284 hammer2_chain_t
*sub
;
4291 * Make sure we have an accurate live_count
4293 if ((chain
->flags
& (HAMMER2_CHAIN_INITIAL
|
4294 HAMMER2_CHAIN_COUNTEDBREFS
)) == 0) {
4295 base
= &chain
->data
->npdata
[0];
4296 count
= chain
->bytes
/ sizeof(hammer2_blockref_t
);
4297 hammer2_chain_countbrefs(chain
, base
, count
);
4301 * If the indirect block is empty we can delete it.
4302 * (ignore deletion error)
4304 if (chain
->core
.live_count
== 0 && RB_EMPTY(&chain
->core
.rbtree
)) {
4305 hammer2_chain_delete(parent
, chain
,
4306 chain
->bref
.modify_tid
,
4307 HAMMER2_DELETE_PERMANENT
);
4308 hammer2_chain_repchange(parent
, chain
);
4312 base
= hammer2_chain_base_and_count(parent
, &count
);
4314 if ((parent
->flags
& (HAMMER2_CHAIN_INITIAL
|
4315 HAMMER2_CHAIN_COUNTEDBREFS
)) == 0) {
4316 hammer2_chain_countbrefs(parent
, base
, count
);
4320 * Determine if we can collapse chain into parent, calculate
4321 * hysteresis for chain emptiness.
4323 if (parent
->core
.live_count
+ chain
->core
.live_count
- 1 > count
)
4325 chain_count
= chain
->bytes
/ sizeof(hammer2_blockref_t
);
4326 if (chain
->core
.live_count
> chain_count
* 3 / 4)
4330 * Ok, theoretically we can collapse chain's contents into
4331 * parent. chain is locked, but any in-memory children of chain
4332 * are not. For this to work, we must be able to dispose of any
4333 * in-memory children of chain.
4335 * For now require that there are no in-memory children of chain.
4337 * WARNING! Both chain and parent must remain locked across this
4342 * Parent must be marked modified. Don't try to collapse it if we
4343 * can't mark it modified. Once modified, destroy chain to make room
4344 * and to get rid of what will be a conflicting key (this is included
4345 * in the calculation above). Finally, move the children of chain
4346 * into chain's parent.
4348 * This order creates an accounting problem for bref.embed.stats
4349 * because we destroy chain before we remove its children. Any
4350 * elements whos blockref is already synchronized will be counted
4351 * twice. To deal with the problem we clean out chain's stats prior
4354 error
= hammer2_chain_modify(parent
, 0, 0, 0);
4356 krateprintf(&krate_h2me
, "hammer2: indirect_maint: %s\n",
4357 hammer2_error_str(error
));
4360 error
= hammer2_chain_modify(chain
, chain
->bref
.modify_tid
, 0, 0);
4362 krateprintf(&krate_h2me
, "hammer2: indirect_maint: %s\n",
4363 hammer2_error_str(error
));
4367 chain
->bref
.embed
.stats
.inode_count
= 0;
4368 chain
->bref
.embed
.stats
.data_count
= 0;
4369 error
= hammer2_chain_delete(parent
, chain
,
4370 chain
->bref
.modify_tid
,
4371 HAMMER2_DELETE_PERMANENT
);
4372 KKASSERT(error
== 0);
4375 * The combined_find call requires core.spin to be held. One would
4376 * think there wouldn't be any conflicts since we hold chain
4377 * exclusively locked, but the caching mechanism for 0-ref children
4378 * does not require a chain lock.
4380 hammer2_spin_ex(&chain
->core
.spin
);
4384 key_end
= HAMMER2_KEY_MAX
;
4386 chain_base
= &chain
->data
->npdata
[0];
4387 chain_count
= chain
->bytes
/ sizeof(hammer2_blockref_t
);
4388 sub
= hammer2_combined_find(chain
, chain_base
, chain_count
,
4392 generation
= chain
->core
.generation
;
4395 key_next
= bref
->key
+ ((hammer2_key_t
)1 << bref
->keybits
);
4399 hammer2_chain_ref(sub
);
4400 hammer2_spin_unex(&chain
->core
.spin
);
4401 hammer2_chain_lock(sub
, HAMMER2_RESOLVE_NEVER
);
4403 hammer2_spin_unex(&chain
->core
.spin
);
4404 sub
= hammer2_chain_get(chain
, generation
, &bcopy
,
4405 HAMMER2_RESOLVE_NEVER
);
4407 hammer2_spin_ex(&chain
->core
.spin
);
4411 if (bcmp(&bcopy
, &sub
->bref
, sizeof(bcopy
)) ||
4412 sub
->parent
!= chain
||
4413 (sub
->flags
& HAMMER2_CHAIN_DELETED
)) {
4414 hammer2_chain_unlock(sub
);
4415 hammer2_chain_drop(sub
);
4416 hammer2_spin_ex(&chain
->core
.spin
);
4417 sub
= NULL
; /* safety */
4420 error
= hammer2_chain_delete(chain
, sub
,
4421 sub
->bref
.modify_tid
, 0);
4422 KKASSERT(error
== 0);
4423 hammer2_chain_rename(&parent
, sub
,
4424 sub
->bref
.modify_tid
,
4425 HAMMER2_INSERT_SAMEPARENT
);
4426 hammer2_chain_unlock(sub
);
4427 hammer2_chain_drop(sub
);
4428 hammer2_spin_ex(&chain
->core
.spin
);
4434 hammer2_spin_unex(&chain
->core
.spin
);
4436 hammer2_chain_repchange(parent
, chain
);
4442 * Freemap indirect blocks
4444 * Calculate the keybits and highside/lowside of the freemap node the
4445 * caller is creating.
4447 * This routine will specify the next higher-level freemap key/radix
4448 * representing the lowest-ordered set. By doing so, eventually all
4449 * low-ordered sets will be moved one level down.
4451 * We have to be careful here because the freemap reserves a limited
4452 * number of blocks for a limited number of levels. So we can't just
4453 * push indiscriminately.
4456 hammer2_chain_indkey_freemap(hammer2_chain_t
*parent
, hammer2_key_t
*keyp
,
4457 int keybits
, hammer2_blockref_t
*base
, int count
)
4459 hammer2_chain_t
*chain
;
4460 hammer2_blockref_t
*bref
;
4462 hammer2_key_t key_beg
;
4463 hammer2_key_t key_end
;
4464 hammer2_key_t key_next
;
4467 int maxloops
= 300000;
4475 * Calculate the range of keys in the array being careful to skip
4476 * slots which are overridden with a deletion.
4479 key_end
= HAMMER2_KEY_MAX
;
4480 hammer2_spin_ex(&parent
->core
.spin
);
4483 if (--maxloops
== 0) {
4484 panic("indkey_freemap shit %p %p:%d\n",
4485 parent
, base
, count
);
4487 chain
= hammer2_combined_find(parent
, base
, count
,
4499 * Skip deleted chains.
4501 if (chain
&& (chain
->flags
& HAMMER2_CHAIN_DELETED
)) {
4502 if (key_next
== 0 || key_next
> key_end
)
4509 * Use the full live (not deleted) element for the scan
4510 * iteration. HAMMER2 does not allow partial replacements.
4512 * XXX should be built into hammer2_combined_find().
4514 key_next
= bref
->key
+ ((hammer2_key_t
)1 << bref
->keybits
);
4516 if (keybits
> bref
->keybits
) {
4518 keybits
= bref
->keybits
;
4519 } else if (keybits
== bref
->keybits
&& bref
->key
< key
) {
4526 hammer2_spin_unex(&parent
->core
.spin
);
4529 * Return the keybits for a higher-level FREEMAP_NODE covering
4533 case HAMMER2_FREEMAP_LEVEL0_RADIX
:
4534 keybits
= HAMMER2_FREEMAP_LEVEL1_RADIX
;
4536 case HAMMER2_FREEMAP_LEVEL1_RADIX
:
4537 keybits
= HAMMER2_FREEMAP_LEVEL2_RADIX
;
4539 case HAMMER2_FREEMAP_LEVEL2_RADIX
:
4540 keybits
= HAMMER2_FREEMAP_LEVEL3_RADIX
;
4542 case HAMMER2_FREEMAP_LEVEL3_RADIX
:
4543 keybits
= HAMMER2_FREEMAP_LEVEL4_RADIX
;
4545 case HAMMER2_FREEMAP_LEVEL4_RADIX
:
4546 keybits
= HAMMER2_FREEMAP_LEVEL5_RADIX
;
4548 case HAMMER2_FREEMAP_LEVEL5_RADIX
:
4549 panic("hammer2_chain_indkey_freemap: level too high");
4552 panic("hammer2_chain_indkey_freemap: bad radix");
4561 * File indirect blocks
4563 * Calculate the key/keybits for the indirect block to create by scanning
4564 * existing keys. The key being created is also passed in *keyp and can be
4565 * inside or outside the indirect block. Regardless, the indirect block
4566 * must hold at least two keys in order to guarantee sufficient space.
4568 * We use a modified version of the freemap's fixed radix tree, but taylored
4569 * for file data. Basically we configure an indirect block encompassing the
4573 hammer2_chain_indkey_file(hammer2_chain_t
*parent
, hammer2_key_t
*keyp
,
4574 int keybits
, hammer2_blockref_t
*base
, int count
,
4577 hammer2_chain_t
*chain
;
4578 hammer2_blockref_t
*bref
;
4580 hammer2_key_t key_beg
;
4581 hammer2_key_t key_end
;
4582 hammer2_key_t key_next
;
4586 int maxloops
= 300000;
4594 * Calculate the range of keys in the array being careful to skip
4595 * slots which are overridden with a deletion.
4597 * Locate the smallest key.
4600 key_end
= HAMMER2_KEY_MAX
;
4601 hammer2_spin_ex(&parent
->core
.spin
);
4604 if (--maxloops
== 0) {
4605 panic("indkey_freemap shit %p %p:%d\n",
4606 parent
, base
, count
);
4608 chain
= hammer2_combined_find(parent
, base
, count
,
4620 * Skip deleted chains.
4622 if (chain
&& (chain
->flags
& HAMMER2_CHAIN_DELETED
)) {
4623 if (key_next
== 0 || key_next
> key_end
)
4630 * Use the full live (not deleted) element for the scan
4631 * iteration. HAMMER2 does not allow partial replacements.
4633 * XXX should be built into hammer2_combined_find().
4635 key_next
= bref
->key
+ ((hammer2_key_t
)1 << bref
->keybits
);
4637 if (keybits
> bref
->keybits
) {
4639 keybits
= bref
->keybits
;
4640 } else if (keybits
== bref
->keybits
&& bref
->key
< key
) {
4647 hammer2_spin_unex(&parent
->core
.spin
);
4650 * Calculate the static keybits for a higher-level indirect block
4651 * that contains the key.
4656 case HAMMER2_IND_BYTES_MIN
/ sizeof(hammer2_blockref_t
):
4657 nradix
= HAMMER2_IND_RADIX_MIN
- HAMMER2_BLOCKREF_RADIX
;
4659 case HAMMER2_IND_BYTES_NOM
/ sizeof(hammer2_blockref_t
):
4660 nradix
= HAMMER2_IND_RADIX_NOM
- HAMMER2_BLOCKREF_RADIX
;
4662 case HAMMER2_IND_BYTES_MAX
/ sizeof(hammer2_blockref_t
):
4663 nradix
= HAMMER2_IND_RADIX_MAX
- HAMMER2_BLOCKREF_RADIX
;
4666 panic("bad ncount %d\n", ncount
);
4672 * The largest radix that can be returned for an indirect block is
4673 * 63 bits. (The largest practical indirect block radix is actually
4674 * 62 bits because the top-level inode or volume root contains four
4675 * entries, but allow 63 to be returned).
4680 return keybits
+ nradix
;
4686 * Directory indirect blocks.
4688 * Covers both the inode index (directory of inodes), and directory contents
4689 * (filenames hardlinked to inodes).
4691 * Because directory keys are hashed we generally try to cut the space in
4692 * half. We accomodate the inode index (which tends to have linearly
4693 * increasing inode numbers) by ensuring that the keyspace is at least large
4694 * enough to fill up the indirect block being created.
4697 hammer2_chain_indkey_dir(hammer2_chain_t
*parent
, hammer2_key_t
*keyp
,
4698 int keybits
, hammer2_blockref_t
*base
, int count
,
4701 hammer2_blockref_t
*bref
;
4702 hammer2_chain_t
*chain
;
4703 hammer2_key_t key_beg
;
4704 hammer2_key_t key_end
;
4705 hammer2_key_t key_next
;
4710 int maxloops
= 300000;
4713 * NOTE: We can't take a shortcut here anymore for inodes because
4714 * the root directory can contain a mix of inodes and directory
4715 * entries (we used to just return 63 if parent->bref.type was
4716 * HAMMER2_BREF_TYPE_INODE.
4723 * Calculate the range of keys in the array being careful to skip
4724 * slots which are overridden with a deletion.
4727 key_end
= HAMMER2_KEY_MAX
;
4728 hammer2_spin_ex(&parent
->core
.spin
);
4731 if (--maxloops
== 0) {
4732 panic("indkey_freemap shit %p %p:%d\n",
4733 parent
, base
, count
);
4735 chain
= hammer2_combined_find(parent
, base
, count
,
4749 if (chain
&& (chain
->flags
& HAMMER2_CHAIN_DELETED
)) {
4750 if (key_next
== 0 || key_next
> key_end
)
4757 * Use the full live (not deleted) element for the scan
4758 * iteration. HAMMER2 does not allow partial replacements.
4760 * XXX should be built into hammer2_combined_find().
4762 key_next
= bref
->key
+ ((hammer2_key_t
)1 << bref
->keybits
);
4765 * Expand our calculated key range (key, keybits) to fit
4766 * the scanned key. nkeybits represents the full range
4767 * that we will later cut in half (two halves @ nkeybits - 1).
4770 if (nkeybits
< bref
->keybits
) {
4771 if (bref
->keybits
> 64) {
4772 kprintf("bad bref chain %p bref %p\n",
4776 nkeybits
= bref
->keybits
;
4778 while (nkeybits
< 64 &&
4779 (~(((hammer2_key_t
)1 << nkeybits
) - 1) &
4780 (key
^ bref
->key
)) != 0) {
4785 * If the new key range is larger we have to determine
4786 * which side of the new key range the existing keys fall
4787 * under by checking the high bit, then collapsing the
4788 * locount into the hicount or vise-versa.
4790 if (keybits
!= nkeybits
) {
4791 if (((hammer2_key_t
)1 << (nkeybits
- 1)) & key
) {
4802 * The newly scanned key will be in the lower half or the
4803 * upper half of the (new) key range.
4805 if (((hammer2_key_t
)1 << (nkeybits
- 1)) & bref
->key
)
4814 hammer2_spin_unex(&parent
->core
.spin
);
4815 bref
= NULL
; /* now invalid (safety) */
4818 * Adjust keybits to represent half of the full range calculated
4819 * above (radix 63 max) for our new indirect block.
4824 * Expand keybits to hold at least ncount elements. ncount will be
4825 * a power of 2. This is to try to completely fill leaf nodes (at
4826 * least for keys which are not hashes).
4828 * We aren't counting 'in' or 'out', we are counting 'high side'
4829 * and 'low side' based on the bit at (1LL << keybits). We want
4830 * everything to be inside in these cases so shift it all to
4831 * the low or high side depending on the new high bit.
4833 while (((hammer2_key_t
)1 << keybits
) < ncount
) {
4835 if (key
& ((hammer2_key_t
)1 << keybits
)) {
4844 if (hicount
> locount
)
4845 key
|= (hammer2_key_t
)1 << keybits
;
4847 key
&= ~(hammer2_key_t
)1 << keybits
;
4857 * Directory indirect blocks.
4859 * Covers both the inode index (directory of inodes), and directory contents
4860 * (filenames hardlinked to inodes).
4862 * Because directory keys are hashed we generally try to cut the space in
4863 * half. We accomodate the inode index (which tends to have linearly
4864 * increasing inode numbers) by ensuring that the keyspace is at least large
4865 * enough to fill up the indirect block being created.
4868 hammer2_chain_indkey_dir(hammer2_chain_t
*parent
, hammer2_key_t
*keyp
,
4869 int keybits
, hammer2_blockref_t
*base
, int count
,
4872 hammer2_blockref_t
*bref
;
4873 hammer2_chain_t
*chain
;
4874 hammer2_key_t key_beg
;
4875 hammer2_key_t key_end
;
4876 hammer2_key_t key_next
;
4881 int maxloops
= 300000;
4884 * Shortcut if the parent is the inode. In this situation the
4885 * parent has 4+1 directory entries and we are creating an indirect
4886 * block capable of holding many more.
4888 if (parent
->bref
.type
== HAMMER2_BREF_TYPE_INODE
) {
4897 * Calculate the range of keys in the array being careful to skip
4898 * slots which are overridden with a deletion.
4901 key_end
= HAMMER2_KEY_MAX
;
4902 hammer2_spin_ex(&parent
->core
.spin
);
4905 if (--maxloops
== 0) {
4906 panic("indkey_freemap shit %p %p:%d\n",
4907 parent
, base
, count
);
4909 chain
= hammer2_combined_find(parent
, base
, count
,
4923 if (chain
&& (chain
->flags
& HAMMER2_CHAIN_DELETED
)) {
4924 if (key_next
== 0 || key_next
> key_end
)
4931 * Use the full live (not deleted) element for the scan
4932 * iteration. HAMMER2 does not allow partial replacements.
4934 * XXX should be built into hammer2_combined_find().
4936 key_next
= bref
->key
+ ((hammer2_key_t
)1 << bref
->keybits
);
4939 * Expand our calculated key range (key, keybits) to fit
4940 * the scanned key. nkeybits represents the full range
4941 * that we will later cut in half (two halves @ nkeybits - 1).
4944 if (nkeybits
< bref
->keybits
) {
4945 if (bref
->keybits
> 64) {
4946 kprintf("bad bref chain %p bref %p\n",
4950 nkeybits
= bref
->keybits
;
4952 while (nkeybits
< 64 &&
4953 (~(((hammer2_key_t
)1 << nkeybits
) - 1) &
4954 (key
^ bref
->key
)) != 0) {
4959 * If the new key range is larger we have to determine
4960 * which side of the new key range the existing keys fall
4961 * under by checking the high bit, then collapsing the
4962 * locount into the hicount or vise-versa.
4964 if (keybits
!= nkeybits
) {
4965 if (((hammer2_key_t
)1 << (nkeybits
- 1)) & key
) {
4976 * The newly scanned key will be in the lower half or the
4977 * upper half of the (new) key range.
4979 if (((hammer2_key_t
)1 << (nkeybits
- 1)) & bref
->key
)
4988 hammer2_spin_unex(&parent
->core
.spin
);
4989 bref
= NULL
; /* now invalid (safety) */
4992 * Adjust keybits to represent half of the full range calculated
4993 * above (radix 63 max) for our new indirect block.
4998 * Expand keybits to hold at least ncount elements. ncount will be
4999 * a power of 2. This is to try to completely fill leaf nodes (at
5000 * least for keys which are not hashes).
5002 * We aren't counting 'in' or 'out', we are counting 'high side'
5003 * and 'low side' based on the bit at (1LL << keybits). We want
5004 * everything to be inside in these cases so shift it all to
5005 * the low or high side depending on the new high bit.
5007 while (((hammer2_key_t
)1 << keybits
) < ncount
) {
5009 if (key
& ((hammer2_key_t
)1 << keybits
)) {
5018 if (hicount
> locount
)
5019 key
|= (hammer2_key_t
)1 << keybits
;
5021 key
&= ~(hammer2_key_t
)1 << keybits
;
5031 * Sets CHAIN_DELETED and remove the chain's blockref from the parent if
5034 * Both parent and chain must be locked exclusively.
5036 * This function will modify the parent if the blockref requires removal
5037 * from the parent's block table.
5039 * This function is NOT recursive. Any entity already pushed into the
5040 * chain (such as an inode) may still need visibility into its contents,
5041 * as well as the ability to read and modify the contents. For example,
5042 * for an unlinked file which is still open.
5044 * Also note that the flusher is responsible for cleaning up empty
5048 hammer2_chain_delete(hammer2_chain_t
*parent
, hammer2_chain_t
*chain
,
5049 hammer2_tid_t mtid
, int flags
)
5053 KKASSERT(hammer2_mtx_owned(&chain
->lock
));
5056 * Nothing to do if already marked.
5058 * We need the spinlock on the core whos RBTREE contains chain
5059 * to protect against races.
5061 if ((chain
->flags
& HAMMER2_CHAIN_DELETED
) == 0) {
5062 KKASSERT((chain
->flags
& HAMMER2_CHAIN_DELETED
) == 0 &&
5063 chain
->parent
== parent
);
5064 error
= _hammer2_chain_delete_helper(parent
, chain
,
5069 * Permanent deletions mark the chain as destroyed.
5071 * NOTE: We do not setflush the chain unless the deletion is
5072 * permanent, since the deletion of a chain does not actually
5073 * require it to be flushed.
5076 if (flags
& HAMMER2_DELETE_PERMANENT
) {
5077 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_DESTROY
);
5078 hammer2_chain_setflush(chain
);
5086 * Returns the index of the nearest element in the blockref array >= elm.
5087 * Returns (count) if no element could be found.
5089 * Sets *key_nextp to the next key for loop purposes but does not modify
5090 * it if the next key would be higher than the current value of *key_nextp.
5091 * Note that *key_nexp can overflow to 0, which should be tested by the
5094 * WARNING! Must be called with parent's spinlock held. Spinlock remains
5095 * held through the operation.
5098 hammer2_base_find(hammer2_chain_t
*parent
,
5099 hammer2_blockref_t
*base
, int count
,
5100 hammer2_key_t
*key_nextp
,
5101 hammer2_key_t key_beg
, hammer2_key_t key_end
)
5103 hammer2_blockref_t
*scan
;
5104 hammer2_key_t scan_end
;
5109 * Require the live chain's already have their core's counted
5110 * so we can optimize operations.
5112 KKASSERT(parent
->flags
& HAMMER2_CHAIN_COUNTEDBREFS
);
5117 if (count
== 0 || base
== NULL
)
5121 * Sequential optimization using parent->cache_index. This is
5122 * the most likely scenario.
5124 * We can avoid trailing empty entries on live chains, otherwise
5125 * we might have to check the whole block array.
5127 i
= parent
->cache_index
; /* SMP RACE OK */
5129 limit
= parent
->core
.live_zero
;
5134 KKASSERT(i
< count
);
5140 while (i
> 0 && (scan
->type
== 0 || scan
->key
> key_beg
)) {
5144 parent
->cache_index
= i
;
5147 * Search forwards, stop when we find a scan element which
5148 * encloses the key or until we know that there are no further
5152 if (scan
->type
!= 0) {
5153 scan_end
= scan
->key
+
5154 ((hammer2_key_t
)1 << scan
->keybits
) - 1;
5155 if (scan
->key
> key_beg
|| scan_end
>= key_beg
)
5164 parent
->cache_index
= i
;
5168 scan_end
= scan
->key
+
5169 ((hammer2_key_t
)1 << scan
->keybits
);
5170 if (scan_end
&& (*key_nextp
> scan_end
||
5172 *key_nextp
= scan_end
;
5180 * Do a combined search and return the next match either from the blockref
5181 * array or from the in-memory chain. Sets *bresp to the returned bref in
5182 * both cases, or sets it to NULL if the search exhausted. Only returns
5183 * a non-NULL chain if the search matched from the in-memory chain.
5185 * When no in-memory chain has been found and a non-NULL bref is returned
5189 * The returned chain is not locked or referenced. Use the returned bref
5190 * to determine if the search exhausted or not. Iterate if the base find
5191 * is chosen but matches a deleted chain.
5193 * WARNING! Must be called with parent's spinlock held. Spinlock remains
5194 * held through the operation.
5197 hammer2_combined_find(hammer2_chain_t
*parent
,
5198 hammer2_blockref_t
*base
, int count
,
5199 hammer2_key_t
*key_nextp
,
5200 hammer2_key_t key_beg
, hammer2_key_t key_end
,
5201 hammer2_blockref_t
**bresp
)
5203 hammer2_blockref_t
*bref
;
5204 hammer2_chain_t
*chain
;
5208 * Lookup in block array and in rbtree.
5210 *key_nextp
= key_end
+ 1;
5211 i
= hammer2_base_find(parent
, base
, count
, key_nextp
,
5213 chain
= hammer2_chain_find(parent
, key_nextp
, key_beg
, key_end
);
5218 if (i
== count
&& chain
== NULL
) {
5224 * Only chain matched.
5227 bref
= &chain
->bref
;
5232 * Only blockref matched.
5234 if (chain
== NULL
) {
5240 * Both in-memory and blockref matched, select the nearer element.
5242 * If both are flush with the left-hand side or both are the
5243 * same distance away, select the chain. In this situation the
5244 * chain must have been loaded from the matching blockmap.
5246 if ((chain
->bref
.key
<= key_beg
&& base
[i
].key
<= key_beg
) ||
5247 chain
->bref
.key
== base
[i
].key
) {
5248 KKASSERT(chain
->bref
.key
== base
[i
].key
);
5249 bref
= &chain
->bref
;
5254 * Select the nearer key
5256 if (chain
->bref
.key
< base
[i
].key
) {
5257 bref
= &chain
->bref
;
5264 * If the bref is out of bounds we've exhausted our search.
5267 if (bref
->key
> key_end
) {
5277 * Locate the specified block array element and delete it. The element
5280 * The spin lock on the related chain must be held.
5282 * NOTE: live_count was adjusted when the chain was deleted, so it does not
5283 * need to be adjusted when we commit the media change.
5286 hammer2_base_delete(hammer2_chain_t
*parent
,
5287 hammer2_blockref_t
*base
, int count
,
5288 hammer2_chain_t
*chain
)
5290 hammer2_blockref_t
*elm
= &chain
->bref
;
5291 hammer2_blockref_t
*scan
;
5292 hammer2_key_t key_next
;
5296 * Delete element. Expect the element to exist.
5298 * XXX see caller, flush code not yet sophisticated enough to prevent
5299 * re-flushed in some cases.
5301 key_next
= 0; /* max range */
5302 i
= hammer2_base_find(parent
, base
, count
, &key_next
,
5303 elm
->key
, elm
->key
);
5305 if (i
== count
|| scan
->type
== 0 ||
5306 scan
->key
!= elm
->key
||
5307 ((chain
->flags
& HAMMER2_CHAIN_BMAPUPD
) == 0 &&
5308 scan
->keybits
!= elm
->keybits
)) {
5309 hammer2_spin_unex(&parent
->core
.spin
);
5310 panic("delete base %p element not found at %d/%d elm %p\n",
5311 base
, i
, count
, elm
);
5316 * Update stats and zero the entry.
5318 * NOTE: Handle radix == 0 (0 bytes) case.
5320 if ((int)(scan
->data_off
& HAMMER2_OFF_MASK_RADIX
)) {
5321 parent
->bref
.embed
.stats
.data_count
-= (hammer2_off_t
)1 <<
5322 (int)(scan
->data_off
& HAMMER2_OFF_MASK_RADIX
);
5324 switch(scan
->type
) {
5325 case HAMMER2_BREF_TYPE_INODE
:
5326 --parent
->bref
.embed
.stats
.inode_count
;
5328 case HAMMER2_BREF_TYPE_DATA
:
5329 if (parent
->bref
.leaf_count
== HAMMER2_BLOCKREF_LEAF_MAX
) {
5330 atomic_set_int(&chain
->flags
,
5331 HAMMER2_CHAIN_HINT_LEAF_COUNT
);
5333 if (parent
->bref
.leaf_count
)
5334 --parent
->bref
.leaf_count
;
5337 case HAMMER2_BREF_TYPE_INDIRECT
:
5338 if (scan
->type
!= HAMMER2_BREF_TYPE_DATA
) {
5339 parent
->bref
.embed
.stats
.data_count
-=
5340 scan
->embed
.stats
.data_count
;
5341 parent
->bref
.embed
.stats
.inode_count
-=
5342 scan
->embed
.stats
.inode_count
;
5344 if (scan
->type
== HAMMER2_BREF_TYPE_INODE
)
5346 if (parent
->bref
.leaf_count
== HAMMER2_BLOCKREF_LEAF_MAX
) {
5347 atomic_set_int(&chain
->flags
,
5348 HAMMER2_CHAIN_HINT_LEAF_COUNT
);
5350 if (parent
->bref
.leaf_count
<= scan
->leaf_count
)
5351 parent
->bref
.leaf_count
= 0;
5353 parent
->bref
.leaf_count
-= scan
->leaf_count
;
5356 case HAMMER2_BREF_TYPE_DIRENT
:
5357 if (parent
->bref
.leaf_count
== HAMMER2_BLOCKREF_LEAF_MAX
) {
5358 atomic_set_int(&chain
->flags
,
5359 HAMMER2_CHAIN_HINT_LEAF_COUNT
);
5361 if (parent
->bref
.leaf_count
)
5362 --parent
->bref
.leaf_count
;
5368 bzero(scan
, sizeof(*scan
));
5371 * We can only optimize parent->core.live_zero for live chains.
5373 if (parent
->core
.live_zero
== i
+ 1) {
5374 while (--i
>= 0 && base
[i
].type
== 0)
5376 parent
->core
.live_zero
= i
+ 1;
5380 * Clear appropriate blockmap flags in chain.
5382 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_BMAPPED
|
5383 HAMMER2_CHAIN_BMAPUPD
);
5387 * Insert the specified element. The block array must not already have the
5388 * element and must have space available for the insertion.
5390 * The spin lock on the related chain must be held.
5392 * NOTE: live_count was adjusted when the chain was deleted, so it does not
5393 * need to be adjusted when we commit the media change.
5396 hammer2_base_insert(hammer2_chain_t
*parent
,
5397 hammer2_blockref_t
*base
, int count
,
5398 hammer2_chain_t
*chain
, hammer2_blockref_t
*elm
)
5400 hammer2_key_t key_next
;
5409 * Insert new element. Expect the element to not already exist
5410 * unless we are replacing it.
5412 * XXX see caller, flush code not yet sophisticated enough to prevent
5413 * re-flushed in some cases.
5415 key_next
= 0; /* max range */
5416 i
= hammer2_base_find(parent
, base
, count
, &key_next
,
5417 elm
->key
, elm
->key
);
5420 * Shortcut fill optimization, typical ordered insertion(s) may not
5423 KKASSERT(i
>= 0 && i
<= count
);
5426 * Set appropriate blockmap flags in chain (if not NULL)
5429 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_BMAPPED
);
5432 * Update stats and zero the entry
5434 if ((int)(elm
->data_off
& HAMMER2_OFF_MASK_RADIX
)) {
5435 parent
->bref
.embed
.stats
.data_count
+= (hammer2_off_t
)1 <<
5436 (int)(elm
->data_off
& HAMMER2_OFF_MASK_RADIX
);
5439 case HAMMER2_BREF_TYPE_INODE
:
5440 ++parent
->bref
.embed
.stats
.inode_count
;
5442 case HAMMER2_BREF_TYPE_DATA
:
5443 if (parent
->bref
.leaf_count
!= HAMMER2_BLOCKREF_LEAF_MAX
)
5444 ++parent
->bref
.leaf_count
;
5446 case HAMMER2_BREF_TYPE_INDIRECT
:
5447 if (elm
->type
!= HAMMER2_BREF_TYPE_DATA
) {
5448 parent
->bref
.embed
.stats
.data_count
+=
5449 elm
->embed
.stats
.data_count
;
5450 parent
->bref
.embed
.stats
.inode_count
+=
5451 elm
->embed
.stats
.inode_count
;
5453 if (elm
->type
== HAMMER2_BREF_TYPE_INODE
)
5455 if (parent
->bref
.leaf_count
+ elm
->leaf_count
<
5456 HAMMER2_BLOCKREF_LEAF_MAX
) {
5457 parent
->bref
.leaf_count
+= elm
->leaf_count
;
5459 parent
->bref
.leaf_count
= HAMMER2_BLOCKREF_LEAF_MAX
;
5462 case HAMMER2_BREF_TYPE_DIRENT
:
5463 if (parent
->bref
.leaf_count
!= HAMMER2_BLOCKREF_LEAF_MAX
)
5464 ++parent
->bref
.leaf_count
;
5472 * We can only optimize parent->core.live_zero for live chains.
5474 if (i
== count
&& parent
->core
.live_zero
< count
) {
5475 i
= parent
->core
.live_zero
++;
5480 xkey
= elm
->key
+ ((hammer2_key_t
)1 << elm
->keybits
) - 1;
5481 if (i
!= count
&& (base
[i
].key
< elm
->key
|| xkey
>= base
[i
].key
)) {
5482 hammer2_spin_unex(&parent
->core
.spin
);
5483 panic("insert base %p overlapping elements at %d elm %p\n",
5488 * Try to find an empty slot before or after.
5492 while (j
> 0 || k
< count
) {
5494 if (j
>= 0 && base
[j
].type
== 0) {
5498 bcopy(&base
[j
+1], &base
[j
],
5499 (i
- j
- 1) * sizeof(*base
));
5505 if (k
< count
&& base
[k
].type
== 0) {
5506 bcopy(&base
[i
], &base
[i
+1],
5507 (k
- i
) * sizeof(hammer2_blockref_t
));
5511 * We can only update parent->core.live_zero for live
5514 if (parent
->core
.live_zero
<= k
)
5515 parent
->core
.live_zero
= k
+ 1;
5520 panic("hammer2_base_insert: no room!");
5527 for (l
= 0; l
< count
; ++l
) {
5529 key_next
= base
[l
].key
+
5530 ((hammer2_key_t
)1 << base
[l
].keybits
) - 1;
5534 while (++l
< count
) {
5536 if (base
[l
].key
<= key_next
)
5537 panic("base_insert %d %d,%d,%d fail %p:%d", u
, i
, j
, k
, base
, l
);
5538 key_next
= base
[l
].key
+
5539 ((hammer2_key_t
)1 << base
[l
].keybits
) - 1;
5549 * Sort the blockref array for the chain. Used by the flush code to
5550 * sort the blockref[] array.
5552 * The chain must be exclusively locked AND spin-locked.
5554 typedef hammer2_blockref_t
*hammer2_blockref_p
;
5558 hammer2_base_sort_callback(const void *v1
, const void *v2
)
5560 hammer2_blockref_p bref1
= *(const hammer2_blockref_p
*)v1
;
5561 hammer2_blockref_p bref2
= *(const hammer2_blockref_p
*)v2
;
5564 * Make sure empty elements are placed at the end of the array
5566 if (bref1
->type
== 0) {
5567 if (bref2
->type
== 0)
5570 } else if (bref2
->type
== 0) {
5577 if (bref1
->key
< bref2
->key
)
5579 if (bref1
->key
> bref2
->key
)
5585 hammer2_base_sort(hammer2_chain_t
*chain
)
5587 hammer2_blockref_t
*base
;
5590 switch(chain
->bref
.type
) {
5591 case HAMMER2_BREF_TYPE_INODE
:
5593 * Special shortcut for embedded data returns the inode
5594 * itself. Callers must detect this condition and access
5595 * the embedded data (the strategy code does this for us).
5597 * This is only applicable to regular files and softlinks.
5599 if (chain
->data
->ipdata
.meta
.op_flags
&
5600 HAMMER2_OPFLAG_DIRECTDATA
) {
5603 base
= &chain
->data
->ipdata
.u
.blockset
.blockref
[0];
5604 count
= HAMMER2_SET_COUNT
;
5606 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
5607 case HAMMER2_BREF_TYPE_INDIRECT
:
5609 * Optimize indirect blocks in the INITIAL state to avoid
5612 KKASSERT((chain
->flags
& HAMMER2_CHAIN_INITIAL
) == 0);
5613 base
= &chain
->data
->npdata
[0];
5614 count
= chain
->bytes
/ sizeof(hammer2_blockref_t
);
5616 case HAMMER2_BREF_TYPE_VOLUME
:
5617 base
= &chain
->data
->voldata
.sroot_blockset
.blockref
[0];
5618 count
= HAMMER2_SET_COUNT
;
5620 case HAMMER2_BREF_TYPE_FREEMAP
:
5621 base
= &chain
->data
->blkset
.blockref
[0];
5622 count
= HAMMER2_SET_COUNT
;
5625 kprintf("hammer2_chain_lookup: unrecognized "
5626 "blockref(A) type: %d",
5629 tsleep(&base
, 0, "dead", 0);
5630 panic("hammer2_chain_lookup: unrecognized "
5631 "blockref(A) type: %d",
5633 base
= NULL
; /* safety */
5634 count
= 0; /* safety */
5636 kqsort(base
, count
, sizeof(*base
), hammer2_base_sort_callback
);
5642 * Chain memory management
5645 hammer2_chain_wait(hammer2_chain_t
*chain
)
5647 tsleep(chain
, 0, "chnflw", 1);
5650 const hammer2_media_data_t
*
5651 hammer2_chain_rdata(hammer2_chain_t
*chain
)
5653 KKASSERT(chain
->data
!= NULL
);
5654 return (chain
->data
);
5657 hammer2_media_data_t
*
5658 hammer2_chain_wdata(hammer2_chain_t
*chain
)
5660 KKASSERT(chain
->data
!= NULL
);
5661 return (chain
->data
);
5665 * Set the check data for a chain. This can be a heavy-weight operation
5666 * and typically only runs on-flush. For file data check data is calculated
5667 * when the logical buffers are flushed.
5670 hammer2_chain_setcheck(hammer2_chain_t
*chain
, void *bdata
)
5672 chain
->bref
.flags
&= ~HAMMER2_BREF_FLAG_ZERO
;
5674 switch(HAMMER2_DEC_CHECK(chain
->bref
.methods
)) {
5675 case HAMMER2_CHECK_NONE
:
5677 case HAMMER2_CHECK_DISABLED
:
5679 case HAMMER2_CHECK_ISCSI32
:
5680 chain
->bref
.check
.iscsi32
.value
=
5681 hammer2_icrc32(bdata
, chain
->bytes
);
5683 case HAMMER2_CHECK_XXHASH64
:
5684 chain
->bref
.check
.xxhash64
.value
=
5685 XXH64(bdata
, chain
->bytes
, XXH_HAMMER2_SEED
);
5687 case HAMMER2_CHECK_SHA192
:
5689 SHA256_CTX hash_ctx
;
5691 uint8_t digest
[SHA256_DIGEST_LENGTH
];
5692 uint64_t digest64
[SHA256_DIGEST_LENGTH
/8];
5695 SHA256_Init(&hash_ctx
);
5696 SHA256_Update(&hash_ctx
, bdata
, chain
->bytes
);
5697 SHA256_Final(u
.digest
, &hash_ctx
);
5698 u
.digest64
[2] ^= u
.digest64
[3];
5700 chain
->bref
.check
.sha192
.data
,
5701 sizeof(chain
->bref
.check
.sha192
.data
));
5704 case HAMMER2_CHECK_FREEMAP
:
5705 chain
->bref
.check
.freemap
.icrc32
=
5706 hammer2_icrc32(bdata
, chain
->bytes
);
5709 kprintf("hammer2_chain_setcheck: unknown check type %02x\n",
5710 chain
->bref
.methods
);
5716 hammer2_chain_testcheck(hammer2_chain_t
*chain
, void *bdata
)
5722 if (chain
->bref
.flags
& HAMMER2_BREF_FLAG_ZERO
)
5725 switch(HAMMER2_DEC_CHECK(chain
->bref
.methods
)) {
5726 case HAMMER2_CHECK_NONE
:
5729 case HAMMER2_CHECK_DISABLED
:
5732 case HAMMER2_CHECK_ISCSI32
:
5733 check32
= hammer2_icrc32(bdata
, chain
->bytes
);
5734 r
= (chain
->bref
.check
.iscsi32
.value
== check32
);
5736 kprintf("chain %016jx.%02x meth=%02x CHECK FAIL "
5737 "(flags=%08x, bref/data %08x/%08x)\n",
5738 chain
->bref
.data_off
,
5740 chain
->bref
.methods
,
5742 chain
->bref
.check
.iscsi32
.value
,
5745 hammer2_process_icrc32
+= chain
->bytes
;
5747 case HAMMER2_CHECK_XXHASH64
:
5748 check64
= XXH64(bdata
, chain
->bytes
, XXH_HAMMER2_SEED
);
5749 r
= (chain
->bref
.check
.xxhash64
.value
== check64
);
5751 kprintf("chain %016jx.%02x key=%016jx "
5752 "meth=%02x CHECK FAIL "
5753 "(flags=%08x, bref/data %016jx/%016jx)\n",
5754 chain
->bref
.data_off
,
5757 chain
->bref
.methods
,
5759 chain
->bref
.check
.xxhash64
.value
,
5762 hammer2_process_xxhash64
+= chain
->bytes
;
5764 case HAMMER2_CHECK_SHA192
:
5766 SHA256_CTX hash_ctx
;
5768 uint8_t digest
[SHA256_DIGEST_LENGTH
];
5769 uint64_t digest64
[SHA256_DIGEST_LENGTH
/8];
5772 SHA256_Init(&hash_ctx
);
5773 SHA256_Update(&hash_ctx
, bdata
, chain
->bytes
);
5774 SHA256_Final(u
.digest
, &hash_ctx
);
5775 u
.digest64
[2] ^= u
.digest64
[3];
5777 chain
->bref
.check
.sha192
.data
,
5778 sizeof(chain
->bref
.check
.sha192
.data
)) == 0) {
5782 kprintf("chain %016jx.%02x meth=%02x "
5784 chain
->bref
.data_off
,
5786 chain
->bref
.methods
);
5790 case HAMMER2_CHECK_FREEMAP
:
5791 r
= (chain
->bref
.check
.freemap
.icrc32
==
5792 hammer2_icrc32(bdata
, chain
->bytes
));
5794 kprintf("chain %016jx.%02x meth=%02x "
5796 chain
->bref
.data_off
,
5798 chain
->bref
.methods
);
5799 kprintf("freemap.icrc %08x icrc32 %08x (%d)\n",
5800 chain
->bref
.check
.freemap
.icrc32
,
5801 hammer2_icrc32(bdata
, chain
->bytes
),
5804 kprintf("dio %p buf %016jx,%d bdata %p/%p\n",
5805 chain
->dio
, chain
->dio
->bp
->b_loffset
,
5806 chain
->dio
->bp
->b_bufsize
, bdata
,
5807 chain
->dio
->bp
->b_data
);
5812 kprintf("hammer2_chain_setcheck: unknown check type %02x\n",
5813 chain
->bref
.methods
);
5821 * Acquire the chain and parent representing the specified inode for the
5822 * device at the specified cluster index.
5824 * The flags passed in are LOOKUP flags, not RESOLVE flags.
5826 * If we are unable to locate the hardlink, INVAL is returned and *chainp
5827 * will be NULL. *parentp may still be set error or not, or NULL if the
5828 * parent itself could not be resolved.
5830 * Caller must pass-in a valid (and locked), or NULL *parentp or *chainp.
5831 * This function replaces *parentp and *chainp. Generally speaking, if
5832 * the caller found a directory entry and wants the inode, the caller should
5833 * pass the parent,chain representing the directory entry so this function
5834 * can dispose of it properly to avoid any possible lock order reversals.
5837 hammer2_chain_inode_find(hammer2_pfs_t
*pmp
, hammer2_key_t inum
,
5838 int clindex
, int flags
,
5839 hammer2_chain_t
**parentp
, hammer2_chain_t
**chainp
)
5841 hammer2_chain_t
*parent
;
5842 hammer2_chain_t
*rchain
;
5843 hammer2_key_t key_dummy
;
5847 resolve_flags
= (flags
& HAMMER2_LOOKUP_SHARED
) ?
5848 HAMMER2_RESOLVE_SHARED
: 0;
5851 * Caller expects us to replace these.
5854 hammer2_chain_unlock(*chainp
);
5855 hammer2_chain_drop(*chainp
);
5859 hammer2_chain_unlock(*parentp
);
5860 hammer2_chain_drop(*parentp
);
5865 * Inodes hang off of the iroot (bit 63 is clear, differentiating
5866 * inodes from root directory entries in the key lookup).
5868 parent
= hammer2_inode_chain(pmp
->iroot
, clindex
, resolve_flags
);
5871 rchain
= hammer2_chain_lookup(&parent
, &key_dummy
,
5875 error
= HAMMER2_ERROR_EIO
;
5884 * Used by the bulkscan code to snapshot the synchronized storage for
5885 * a volume, allowing it to be scanned concurrently against normal
5889 hammer2_chain_bulksnap(hammer2_dev_t
*hmp
)
5891 hammer2_chain_t
*copy
;
5893 copy
= hammer2_chain_alloc(hmp
, hmp
->spmp
, &hmp
->vchain
.bref
);
5894 copy
->data
= kmalloc(sizeof(copy
->data
->voldata
),
5897 hammer2_voldata_lock(hmp
);
5898 copy
->data
->voldata
= hmp
->volsync
;
5899 hammer2_voldata_unlock(hmp
);
5905 hammer2_chain_bulkdrop(hammer2_chain_t
*copy
)
5907 KKASSERT(copy
->bref
.type
== HAMMER2_BREF_TYPE_VOLUME
);
5908 KKASSERT(copy
->data
);
5909 kfree(copy
->data
, copy
->hmp
->mchain
);
5911 atomic_add_long(&hammer2_chain_allocs
, -1);
5912 hammer2_chain_drop(copy
);
5916 * Returns non-zero if the chain (INODE or DIRENT) matches the
5920 hammer2_chain_dirent_test(hammer2_chain_t
*chain
, const char *name
,
5923 const hammer2_inode_data_t
*ripdata
;
5924 const hammer2_dirent_head_t
*den
;
5926 if (chain
->bref
.type
== HAMMER2_BREF_TYPE_INODE
) {
5927 ripdata
= &chain
->data
->ipdata
;
5928 if (ripdata
->meta
.name_len
== name_len
&&
5929 bcmp(ripdata
->filename
, name
, name_len
) == 0) {
5933 if (chain
->bref
.type
== HAMMER2_BREF_TYPE_DIRENT
&&
5934 chain
->bref
.embed
.dirent
.namlen
== name_len
) {
5935 den
= &chain
->bref
.embed
.dirent
;
5936 if (name_len
> sizeof(chain
->bref
.check
.buf
) &&
5937 bcmp(chain
->data
->buf
, name
, name_len
) == 0) {
5940 if (name_len
<= sizeof(chain
->bref
.check
.buf
) &&
5941 bcmp(chain
->bref
.check
.buf
, name
, name_len
) == 0) {