2 * Copyright (c) 2011-2014 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 int hammer2_indirect_optimize
; /* XXX SYSCTL */
69 static hammer2_chain_t
*hammer2_chain_create_indirect(
70 hammer2_trans_t
*trans
, hammer2_chain_t
*parent
,
71 hammer2_key_t key
, int keybits
, int for_type
, int *errorp
);
72 static void hammer2_chain_drop_data(hammer2_chain_t
*chain
, int lastdrop
);
73 static hammer2_chain_t
*hammer2_combined_find(
74 hammer2_chain_t
*parent
,
75 hammer2_blockref_t
*base
, int count
,
76 int *cache_indexp
, hammer2_key_t
*key_nextp
,
77 hammer2_key_t key_beg
, hammer2_key_t key_end
,
78 hammer2_blockref_t
**bresp
);
81 * Basic RBTree for chains (core->rbtree and core->dbtree). Chains cannot
82 * overlap in the RB trees. Deleted chains are moved from rbtree to either
85 * Chains in delete-duplicate sequences can always iterate through core_entry
86 * to locate the live version of the chain.
88 RB_GENERATE(hammer2_chain_tree
, hammer2_chain
, rbnode
, hammer2_chain_cmp
);
91 hammer2_chain_cmp(hammer2_chain_t
*chain1
, hammer2_chain_t
*chain2
)
99 * Compare chains. Overlaps are not supposed to happen and catch
100 * any software issues early we count overlaps as a match.
102 c1_beg
= chain1
->bref
.key
;
103 c1_end
= c1_beg
+ ((hammer2_key_t
)1 << chain1
->bref
.keybits
) - 1;
104 c2_beg
= chain2
->bref
.key
;
105 c2_end
= c2_beg
+ ((hammer2_key_t
)1 << chain2
->bref
.keybits
) - 1;
107 if (c1_end
< c2_beg
) /* fully to the left */
109 if (c1_beg
> c2_end
) /* fully to the right */
111 return(0); /* overlap (must not cross edge boundary) */
116 hammer2_isclusterable(hammer2_chain_t
*chain
)
118 if (hammer2_cluster_enable
) {
119 if (chain
->bref
.type
== HAMMER2_BREF_TYPE_INDIRECT
||
120 chain
->bref
.type
== HAMMER2_BREF_TYPE_INODE
||
121 chain
->bref
.type
== HAMMER2_BREF_TYPE_DATA
) {
129 * Make a chain visible to the flusher. The flusher needs to be able to
130 * do flushes of subdirectory chains or single files so it does a top-down
131 * recursion using the ONFLUSH flag for the recursion. It locates MODIFIED
132 * or UPDATE chains and flushes back up the chain to the volume root.
134 * This routine sets ONFLUSH upward until it hits the volume root. For
135 * simplicity we ignore PFSROOT boundaries whos rules can be complex.
136 * Extra ONFLUSH flagging doesn't hurt the filesystem.
139 hammer2_chain_setflush(hammer2_trans_t
*trans
, hammer2_chain_t
*chain
)
141 hammer2_chain_t
*parent
;
143 if ((chain
->flags
& HAMMER2_CHAIN_ONFLUSH
) == 0) {
144 hammer2_spin_sh(&chain
->core
.spin
);
145 while ((chain
->flags
& HAMMER2_CHAIN_ONFLUSH
) == 0) {
146 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_ONFLUSH
);
147 if ((parent
= chain
->parent
) == NULL
)
149 hammer2_spin_sh(&parent
->core
.spin
);
150 hammer2_spin_unsh(&chain
->core
.spin
);
153 hammer2_spin_unsh(&chain
->core
.spin
);
158 * Allocate a new disconnected chain element representing the specified
159 * bref. chain->refs is set to 1 and the passed bref is copied to
160 * chain->bref. chain->bytes is derived from the bref.
162 * chain->pmp inherits pmp unless the chain is an inode (other than the
165 * NOTE: Returns a referenced but unlocked (because there is no core) chain.
168 hammer2_chain_alloc(hammer2_dev_t
*hmp
, hammer2_pfs_t
*pmp
,
169 hammer2_trans_t
*trans
, hammer2_blockref_t
*bref
)
171 hammer2_chain_t
*chain
;
172 u_int bytes
= 1U << (int)(bref
->data_off
& HAMMER2_OFF_MASK_RADIX
);
175 * Construct the appropriate system structure.
178 case HAMMER2_BREF_TYPE_INODE
:
179 case HAMMER2_BREF_TYPE_INDIRECT
:
180 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
181 case HAMMER2_BREF_TYPE_DATA
:
182 case HAMMER2_BREF_TYPE_FREEMAP_LEAF
:
184 * Chain's are really only associated with the hmp but we
185 * maintain a pmp association for per-mount memory tracking
186 * purposes. The pmp can be NULL.
188 chain
= kmalloc(sizeof(*chain
), hmp
->mchain
, M_WAITOK
| M_ZERO
);
190 case HAMMER2_BREF_TYPE_VOLUME
:
191 case HAMMER2_BREF_TYPE_FREEMAP
:
193 panic("hammer2_chain_alloc volume type illegal for op");
196 panic("hammer2_chain_alloc: unrecognized blockref type: %d",
201 * Initialize the new chain structure. pmp must be set to NULL for
202 * chains belonging to the super-root topology of a device mount.
204 if (pmp
== hmp
->spmp
)
210 chain
->bytes
= bytes
;
212 chain
->flags
= HAMMER2_CHAIN_ALLOCATED
;
215 * Set the PFS boundary flag if this chain represents a PFS root.
217 if (bref
->flags
& HAMMER2_BREF_FLAG_PFSROOT
)
218 chain
->flags
|= HAMMER2_CHAIN_PFSBOUNDARY
;
219 hammer2_chain_core_init(chain
);
225 * Initialize a chain's core structure. This structure used to be allocated
226 * but is now embedded.
228 * The core is not locked. No additional refs on the chain are made.
229 * (trans) must not be NULL if (core) is not NULL.
232 hammer2_chain_core_init(hammer2_chain_t
*chain
)
235 * Fresh core under nchain (no multi-homing of ochain's
238 RB_INIT(&chain
->core
.rbtree
); /* live chains */
239 hammer2_mtx_init(&chain
->lock
, "h2chain");
243 * Add a reference to a chain element, preventing its destruction.
245 * (can be called with spinlock held)
248 hammer2_chain_ref(hammer2_chain_t
*chain
)
250 atomic_add_int(&chain
->refs
, 1);
252 kprintf("REFC %p %d %08x\n", chain
, chain
->refs
- 1, chain
->flags
);
258 * Insert the chain in the core rbtree.
260 * Normal insertions are placed in the live rbtree. Insertion of a deleted
261 * chain is a special case used by the flush code that is placed on the
262 * unstaged deleted list to avoid confusing the live view.
264 #define HAMMER2_CHAIN_INSERT_SPIN 0x0001
265 #define HAMMER2_CHAIN_INSERT_LIVE 0x0002
266 #define HAMMER2_CHAIN_INSERT_RACE 0x0004
270 hammer2_chain_insert(hammer2_chain_t
*parent
, hammer2_chain_t
*chain
,
271 int flags
, int generation
)
273 hammer2_chain_t
*xchain
;
276 if (flags
& HAMMER2_CHAIN_INSERT_SPIN
)
277 hammer2_spin_ex(&parent
->core
.spin
);
280 * Interlocked by spinlock, check for race
282 if ((flags
& HAMMER2_CHAIN_INSERT_RACE
) &&
283 parent
->core
.generation
!= generation
) {
291 xchain
= RB_INSERT(hammer2_chain_tree
, &parent
->core
.rbtree
, chain
);
292 KASSERT(xchain
== NULL
,
293 ("hammer2_chain_insert: collision %p %p", chain
, xchain
));
294 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_ONRBTREE
);
295 chain
->parent
= parent
;
296 ++parent
->core
.chain_count
;
297 ++parent
->core
.generation
; /* XXX incs for _get() too, XXX */
300 * We have to keep track of the effective live-view blockref count
301 * so the create code knows when to push an indirect block.
303 if (flags
& HAMMER2_CHAIN_INSERT_LIVE
)
304 atomic_add_int(&parent
->core
.live_count
, 1);
306 if (flags
& HAMMER2_CHAIN_INSERT_SPIN
)
307 hammer2_spin_unex(&parent
->core
.spin
);
312 * Drop the caller's reference to the chain. When the ref count drops to
313 * zero this function will try to disassociate the chain from its parent and
314 * deallocate it, then recursely drop the parent using the implied ref
315 * from the chain's chain->parent.
317 static hammer2_chain_t
*hammer2_chain_lastdrop(hammer2_chain_t
*chain
);
320 hammer2_chain_drop(hammer2_chain_t
*chain
)
325 if (hammer2_debug
& 0x200000)
328 kprintf("DROP %p %d %08x\n", chain
, chain
->refs
- 1, chain
->flags
);
332 if (chain
->flags
& HAMMER2_CHAIN_UPDATE
)
334 if (chain
->flags
& HAMMER2_CHAIN_MODIFIED
)
336 KKASSERT(chain
->refs
> need
);
344 chain
= hammer2_chain_lastdrop(chain
);
346 if (atomic_cmpset_int(&chain
->refs
, refs
, refs
- 1))
348 /* retry the same chain */
354 * Safe handling of the 1->0 transition on chain. Returns a chain for
355 * recursive drop or NULL, possibly returning the same chain if the atomic
358 * Whem two chains need to be recursively dropped we use the chain
359 * we would otherwise free to placehold the additional chain. It's a bit
360 * convoluted but we can't just recurse without potentially blowing out
363 * The chain cannot be freed if it has any children.
365 * The core spinlock is allowed nest child-to-parent (not parent-to-child).
369 hammer2_chain_lastdrop(hammer2_chain_t
*chain
)
373 hammer2_chain_t
*parent
;
374 hammer2_chain_t
*rdrop
;
377 * Spinlock the core and check to see if it is empty. If it is
378 * not empty we leave chain intact with refs == 0. The elements
379 * in core->rbtree are associated with other chains contemporary
380 * with ours but not with our chain directly.
382 hammer2_spin_ex(&chain
->core
.spin
);
385 * We can't free non-stale chains with children until we are
386 * able to free the children because there might be a flush
387 * dependency. Flushes of stale children (which should also
388 * have their deleted flag set) short-cut recursive flush
389 * dependencies and can be freed here. Any flushes which run
390 * through stale children due to the flush synchronization
391 * point should have a FLUSH_* bit set in the chain and not
392 * reach lastdrop at this time.
394 * NOTE: We return (chain) on failure to retry.
396 if (chain
->core
.chain_count
) {
397 if (atomic_cmpset_int(&chain
->refs
, 1, 0)) {
398 hammer2_spin_unex(&chain
->core
.spin
);
399 chain
= NULL
; /* success */
401 hammer2_spin_unex(&chain
->core
.spin
);
405 /* no chains left under us */
408 * chain->core has no children left so no accessors can get to our
409 * chain from there. Now we have to lock the parent core to interlock
410 * remaining possible accessors that might bump chain's refs before
411 * we can safely drop chain's refs with intent to free the chain.
414 pmp
= chain
->pmp
; /* can be NULL */
418 * Spinlock the parent and try to drop the last ref on chain.
419 * On success remove chain from its parent, otherwise return NULL.
421 * (normal core locks are top-down recursive but we define core
422 * spinlocks as bottom-up recursive, so this is safe).
424 if ((parent
= chain
->parent
) != NULL
) {
425 hammer2_spin_ex(&parent
->core
.spin
);
426 if (atomic_cmpset_int(&chain
->refs
, 1, 0) == 0) {
427 /* 1->0 transition failed */
428 hammer2_spin_unex(&parent
->core
.spin
);
429 hammer2_spin_unex(&chain
->core
.spin
);
430 return(chain
); /* retry */
434 * 1->0 transition successful, remove chain from its
437 if (chain
->flags
& HAMMER2_CHAIN_ONRBTREE
) {
438 RB_REMOVE(hammer2_chain_tree
,
439 &parent
->core
.rbtree
, chain
);
440 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_ONRBTREE
);
441 --parent
->core
.chain_count
;
442 chain
->parent
= NULL
;
446 * If our chain was the last chain in the parent's core the
447 * core is now empty and its parent might have to be
448 * re-dropped if it has 0 refs.
450 if (parent
->core
.chain_count
== 0) {
452 if (atomic_cmpset_int(&rdrop
->refs
, 0, 1) == 0) {
456 hammer2_spin_unex(&parent
->core
.spin
);
457 parent
= NULL
; /* safety */
461 * Successful 1->0 transition and the chain can be destroyed now.
463 * We still have the core spinlock, and core's chain_count is 0.
464 * Any parent spinlock is gone.
466 hammer2_spin_unex(&chain
->core
.spin
);
467 KKASSERT(RB_EMPTY(&chain
->core
.rbtree
) &&
468 chain
->core
.chain_count
== 0);
471 * All spin locks are gone, finish freeing stuff.
473 KKASSERT((chain
->flags
& (HAMMER2_CHAIN_UPDATE
|
474 HAMMER2_CHAIN_MODIFIED
)) == 0);
475 hammer2_chain_drop_data(chain
, 1);
477 KKASSERT(chain
->dio
== NULL
);
480 * Once chain resources are gone we can use the now dead chain
481 * structure to placehold what might otherwise require a recursive
482 * drop, because we have potentially two things to drop and can only
483 * return one directly.
485 if (chain
->flags
& HAMMER2_CHAIN_ALLOCATED
) {
486 chain
->flags
&= ~HAMMER2_CHAIN_ALLOCATED
;
488 kfree(chain
, hmp
->mchain
);
492 * Possible chaining loop when parent re-drop needed.
498 * On either last lock release or last drop
501 hammer2_chain_drop_data(hammer2_chain_t
*chain
, int lastdrop
)
503 /*hammer2_dev_t *hmp = chain->hmp;*/
505 switch(chain
->bref
.type
) {
506 case HAMMER2_BREF_TYPE_VOLUME
:
507 case HAMMER2_BREF_TYPE_FREEMAP
:
512 KKASSERT(chain
->data
== NULL
);
518 * Lock a referenced chain element, acquiring its data with I/O if necessary,
519 * and specify how you would like the data to be resolved.
521 * If an I/O or other fatal error occurs, chain->error will be set to non-zero.
523 * The lock is allowed to recurse, multiple locking ops will aggregate
524 * the requested resolve types. Once data is assigned it will not be
525 * removed until the last unlock.
527 * HAMMER2_RESOLVE_NEVER - Do not resolve the data element.
528 * (typically used to avoid device/logical buffer
531 * HAMMER2_RESOLVE_MAYBE - Do not resolve data elements for chains in
532 * the INITIAL-create state (indirect blocks only).
534 * Do not resolve data elements for DATA chains.
535 * (typically used to avoid device/logical buffer
538 * HAMMER2_RESOLVE_ALWAYS- Always resolve the data element.
540 * HAMMER2_RESOLVE_SHARED- (flag) The chain is locked shared, otherwise
541 * it will be locked exclusive.
543 * NOTE: Embedded elements (volume header, inodes) are always resolved
546 * NOTE: Specifying HAMMER2_RESOLVE_ALWAYS on a newly-created non-embedded
547 * element will instantiate and zero its buffer, and flush it on
550 * NOTE: (data) elements are normally locked RESOLVE_NEVER or RESOLVE_MAYBE
551 * so as not to instantiate a device buffer, which could alias against
552 * a logical file buffer. However, if ALWAYS is specified the
553 * device buffer will be instantiated anyway.
555 * WARNING! This function blocks on I/O if data needs to be fetched. This
556 * blocking can run concurrent with other compatible lock holders
557 * who do not need data returning. The lock is not upgraded to
558 * exclusive during a data fetch, a separate bit is used to
559 * interlock I/O. However, an exclusive lock holder can still count
560 * on being interlocked against an I/O fetch managed by a shared
564 hammer2_chain_lock(hammer2_chain_t
*chain
, int how
)
567 * Ref and lock the element. Recursive locks are allowed.
569 KKASSERT(chain
->refs
> 0);
570 atomic_add_int(&chain
->lockcnt
, 1);
573 * Get the appropriate lock.
575 if (how
& HAMMER2_RESOLVE_SHARED
)
576 hammer2_mtx_sh(&chain
->lock
);
578 hammer2_mtx_ex(&chain
->lock
);
581 * If we already have a valid data pointer no further action is
588 * Do we have to resolve the data?
590 switch(how
& HAMMER2_RESOLVE_MASK
) {
591 case HAMMER2_RESOLVE_NEVER
:
593 case HAMMER2_RESOLVE_MAYBE
:
594 if (chain
->flags
& HAMMER2_CHAIN_INITIAL
)
596 if (chain
->bref
.type
== HAMMER2_BREF_TYPE_DATA
)
599 if (chain
->bref
.type
== HAMMER2_BREF_TYPE_FREEMAP_NODE
)
601 if (chain
->bref
.type
== HAMMER2_BREF_TYPE_FREEMAP_LEAF
)
605 case HAMMER2_RESOLVE_ALWAYS
:
610 * Caller requires data
612 hammer2_chain_load_data(chain
);
616 * Issue I/O and install chain->data. Caller must hold a chain lock, lock
617 * may be of any type.
619 * Once chain->data is set it cannot be disposed of until all locks are
623 hammer2_chain_load_data(hammer2_chain_t
*chain
)
625 hammer2_blockref_t
*bref
;
631 * Degenerate case, data already present.
637 KKASSERT(hmp
!= NULL
);
640 * Gain the IOINPROG bit, interlocked block.
646 oflags
= chain
->flags
;
648 if (oflags
& HAMMER2_CHAIN_IOINPROG
) {
649 nflags
= oflags
| HAMMER2_CHAIN_IOSIGNAL
;
650 tsleep_interlock(&chain
->flags
, 0);
651 if (atomic_cmpset_int(&chain
->flags
, oflags
, nflags
)) {
652 tsleep(&chain
->flags
, PINTERLOCKED
,
657 nflags
= oflags
| HAMMER2_CHAIN_IOINPROG
;
658 if (atomic_cmpset_int(&chain
->flags
, oflags
, nflags
)) {
666 * We own CHAIN_IOINPROG
668 * Degenerate case if we raced another load.
674 * We must resolve to a device buffer, either by issuing I/O or
675 * by creating a zero-fill element. We do not mark the buffer
676 * dirty when creating a zero-fill element (the hammer2_chain_modify()
677 * API must still be used to do that).
679 * The device buffer is variable-sized in powers of 2 down
680 * to HAMMER2_MIN_ALLOC (typically 1K). A 64K physical storage
681 * chunk always contains buffers of the same size. (XXX)
683 * The minimum physical IO size may be larger than the variable
689 * The getblk() optimization can only be used on newly created
690 * elements if the physical block size matches the request.
692 if (chain
->flags
& HAMMER2_CHAIN_INITIAL
) {
693 error
= hammer2_io_new(hmp
, bref
->data_off
, chain
->bytes
,
696 error
= hammer2_io_bread(hmp
, bref
->data_off
, chain
->bytes
,
698 hammer2_adjreadcounter(&chain
->bref
, chain
->bytes
);
701 chain
->error
= HAMMER2_ERROR_IO
;
702 kprintf("hammer2_chain_lock: I/O error %016jx: %d\n",
703 (intmax_t)bref
->data_off
, error
);
704 hammer2_io_bqrelse(&chain
->dio
);
710 * NOTE: A locked chain's data cannot be modified without first
711 * calling hammer2_chain_modify().
715 * Clear INITIAL. In this case we used io_new() and the buffer has
716 * been zero'd and marked dirty.
718 bdata
= hammer2_io_data(chain
->dio
, chain
->bref
.data_off
);
719 if (chain
->flags
& HAMMER2_CHAIN_INITIAL
) {
720 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_INITIAL
);
721 chain
->bref
.flags
|= HAMMER2_BREF_FLAG_ZERO
;
722 } else if (chain
->flags
& HAMMER2_CHAIN_MODIFIED
) {
724 * check data not currently synchronized due to
725 * modification. XXX assumes data stays in the buffer
726 * cache, which might not be true (need biodep on flush
727 * to calculate crc? or simple crc?).
730 if (hammer2_chain_testcheck(chain
, bdata
) == 0) {
731 kprintf("chain %016jx.%02x meth=%02x "
732 "CHECK FAIL %08x (flags=%08x)\n",
733 chain
->bref
.data_off
,
736 hammer2_icrc32(bdata
, chain
->bytes
),
738 chain
->error
= HAMMER2_ERROR_CHECK
;
743 * Setup the data pointer, either pointing it to an embedded data
744 * structure and copying the data from the buffer, or pointing it
747 * The buffer is not retained when copying to an embedded data
748 * structure in order to avoid potential deadlocks or recursions
749 * on the same physical buffer.
751 * WARNING! Other threads can start using the data the instant we
752 * set chain->data non-NULL.
754 switch (bref
->type
) {
755 case HAMMER2_BREF_TYPE_VOLUME
:
756 case HAMMER2_BREF_TYPE_FREEMAP
:
758 * Copy data from bp to embedded buffer
760 panic("hammer2_chain_lock: called on unresolved volume header");
762 case HAMMER2_BREF_TYPE_INODE
:
763 case HAMMER2_BREF_TYPE_FREEMAP_LEAF
:
764 case HAMMER2_BREF_TYPE_INDIRECT
:
765 case HAMMER2_BREF_TYPE_DATA
:
766 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
769 * Point data at the device buffer and leave dio intact.
771 chain
->data
= (void *)bdata
;
776 * Release HAMMER2_CHAIN_IOINPROG and signal waiters if requested.
783 oflags
= chain
->flags
;
784 nflags
= oflags
& ~(HAMMER2_CHAIN_IOINPROG
|
785 HAMMER2_CHAIN_IOSIGNAL
);
786 KKASSERT(oflags
& HAMMER2_CHAIN_IOINPROG
);
787 if (atomic_cmpset_int(&chain
->flags
, oflags
, nflags
)) {
788 if (oflags
& HAMMER2_CHAIN_IOSIGNAL
)
789 wakeup(&chain
->flags
);
796 * Unlock and deref a chain element.
798 * On the last lock release any non-embedded data (chain->dio) will be
802 hammer2_chain_unlock(hammer2_chain_t
*chain
)
804 hammer2_mtx_state_t ostate
;
809 * If multiple locks are present (or being attempted) on this
810 * particular chain we can just unlock, drop refs, and return.
812 * Otherwise fall-through on the 1->0 transition.
815 lockcnt
= chain
->lockcnt
;
816 KKASSERT(lockcnt
> 0);
819 if (atomic_cmpset_int(&chain
->lockcnt
,
820 lockcnt
, lockcnt
- 1)) {
821 hammer2_mtx_unlock(&chain
->lock
);
825 if (atomic_cmpset_int(&chain
->lockcnt
, 1, 0))
832 * On the 1->0 transition we upgrade the core lock (if necessary)
833 * to exclusive for terminal processing. If after upgrading we find
834 * that lockcnt is non-zero, another thread is racing us and will
835 * handle the unload for us later on, so just cleanup and return
836 * leaving the data/io intact
838 * Otherwise if lockcnt is still 0 it is possible for it to become
839 * non-zero and race, but since we hold the core->lock exclusively
840 * all that will happen is that the chain will be reloaded after we
843 ostate
= hammer2_mtx_upgrade(&chain
->lock
);
844 if (chain
->lockcnt
) {
845 hammer2_mtx_unlock(&chain
->lock
);
850 * Shortcut the case if the data is embedded or not resolved.
852 * Do NOT NULL out chain->data (e.g. inode data), it might be
855 if (chain
->dio
== NULL
) {
856 if ((chain
->flags
& HAMMER2_CHAIN_MODIFIED
) == 0)
857 hammer2_chain_drop_data(chain
, 0);
858 hammer2_mtx_unlock(&chain
->lock
);
865 if (hammer2_io_isdirty(chain
->dio
) == 0) {
867 } else if (chain
->flags
& HAMMER2_CHAIN_IOFLUSH
) {
868 switch(chain
->bref
.type
) {
869 case HAMMER2_BREF_TYPE_DATA
:
870 counterp
= &hammer2_ioa_file_write
;
872 case HAMMER2_BREF_TYPE_INODE
:
873 counterp
= &hammer2_ioa_meta_write
;
875 case HAMMER2_BREF_TYPE_INDIRECT
:
876 counterp
= &hammer2_ioa_indr_write
;
878 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
879 case HAMMER2_BREF_TYPE_FREEMAP_LEAF
:
880 counterp
= &hammer2_ioa_fmap_write
;
883 counterp
= &hammer2_ioa_volu_write
;
886 *counterp
+= chain
->bytes
;
888 switch(chain
->bref
.type
) {
889 case HAMMER2_BREF_TYPE_DATA
:
890 counterp
= &hammer2_iod_file_write
;
892 case HAMMER2_BREF_TYPE_INODE
:
893 counterp
= &hammer2_iod_meta_write
;
895 case HAMMER2_BREF_TYPE_INDIRECT
:
896 counterp
= &hammer2_iod_indr_write
;
898 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
899 case HAMMER2_BREF_TYPE_FREEMAP_LEAF
:
900 counterp
= &hammer2_iod_fmap_write
;
903 counterp
= &hammer2_iod_volu_write
;
906 *counterp
+= chain
->bytes
;
912 * If a device buffer was used for data be sure to destroy the
913 * buffer when we are done to avoid aliases (XXX what about the
914 * underlying VM pages?).
916 * NOTE: Freemap leaf's use reserved blocks and thus no aliasing
919 * NOTE: The isdirty check tracks whether we have to bdwrite() the
920 * buffer or not. The buffer might already be dirty. The
921 * flag is re-set when chain_modify() is called, even if
922 * MODIFIED is already set, allowing the OS to retire the
923 * buffer independent of a hammer2 flush.
926 if ((chain
->flags
& HAMMER2_CHAIN_IOFLUSH
) &&
927 hammer2_io_isdirty(chain
->dio
)) {
928 hammer2_io_bawrite(&chain
->dio
);
930 hammer2_io_bqrelse(&chain
->dio
);
932 hammer2_mtx_unlock(&chain
->lock
);
936 * This counts the number of live blockrefs in a block array and
937 * also calculates the point at which all remaining blockrefs are empty.
938 * This routine can only be called on a live chain (DUPLICATED flag not set).
940 * NOTE: Flag is not set until after the count is complete, allowing
941 * callers to test the flag without holding the spinlock.
943 * NOTE: If base is NULL the related chain is still in the INITIAL
944 * state and there are no blockrefs to count.
946 * NOTE: live_count may already have some counts accumulated due to
947 * creation and deletion and could even be initially negative.
950 hammer2_chain_countbrefs(hammer2_chain_t
*chain
,
951 hammer2_blockref_t
*base
, int count
)
953 hammer2_spin_ex(&chain
->core
.spin
);
954 if ((chain
->flags
& HAMMER2_CHAIN_COUNTEDBREFS
) == 0) {
956 while (--count
>= 0) {
957 if (base
[count
].type
)
960 chain
->core
.live_zero
= count
+ 1;
962 if (base
[count
].type
)
963 atomic_add_int(&chain
->core
.live_count
,
968 chain
->core
.live_zero
= 0;
970 /* else do not modify live_count */
971 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_COUNTEDBREFS
);
973 hammer2_spin_unex(&chain
->core
.spin
);
977 * Resize the chain's physical storage allocation in-place. This function does
978 * not adjust the data pointer and must be followed by (typically) a
979 * hammer2_chain_modify() call to copy any old data over and adjust the
982 * Chains can be resized smaller without reallocating the storage. Resizing
983 * larger will reallocate the storage. Excess or prior storage is reclaimed
984 * asynchronously at a later time.
986 * Must be passed an exclusively locked parent and chain.
988 * This function is mostly used with DATA blocks locked RESOLVE_NEVER in order
989 * to avoid instantiating a device buffer that conflicts with the vnode data
990 * buffer. However, because H2 can compress or encrypt data, the chain may
991 * have a dio assigned to it in those situations, and they do not conflict.
993 * XXX return error if cannot resize.
996 hammer2_chain_resize(hammer2_trans_t
*trans
, hammer2_inode_t
*ip
,
997 hammer2_chain_t
*parent
, hammer2_chain_t
*chain
,
998 int nradix
, int flags
)
1007 * Only data and indirect blocks can be resized for now.
1008 * (The volu root, inodes, and freemap elements use a fixed size).
1010 KKASSERT(chain
!= &hmp
->vchain
);
1011 KKASSERT(chain
->bref
.type
== HAMMER2_BREF_TYPE_DATA
||
1012 chain
->bref
.type
== HAMMER2_BREF_TYPE_INDIRECT
);
1015 * Nothing to do if the element is already the proper size
1017 obytes
= chain
->bytes
;
1018 nbytes
= 1U << nradix
;
1019 if (obytes
== nbytes
)
1023 * Make sure the old data is instantiated so we can copy it. If this
1024 * is a data block, the device data may be superfluous since the data
1025 * might be in a logical block, but compressed or encrypted data is
1028 * NOTE: The modify will set BMAPUPD for us if BMAPPED is set.
1030 hammer2_chain_modify(trans
, chain
, 0);
1033 * Relocate the block, even if making it smaller (because different
1034 * block sizes may be in different regions).
1036 * (data blocks only, we aren't copying the storage here).
1038 hammer2_freemap_alloc(trans
, chain
, nbytes
);
1039 chain
->bytes
= nbytes
;
1040 /*ip->delta_dcount += (ssize_t)(nbytes - obytes);*/ /* XXX atomic */
1043 * We don't want the followup chain_modify() to try to copy data
1044 * from the old (wrong-sized) buffer. It won't know how much to
1045 * copy. This case should only occur during writes when the
1046 * originator already has the data to write in-hand.
1049 KKASSERT(chain
->bref
.type
== HAMMER2_BREF_TYPE_DATA
);
1050 hammer2_io_brelse(&chain
->dio
);
1056 hammer2_chain_modify(hammer2_trans_t
*trans
, hammer2_chain_t
*chain
, int flags
)
1058 hammer2_blockref_t obref
;
1067 obref
= chain
->bref
;
1068 KKASSERT((chain
->flags
& HAMMER2_CHAIN_FICTITIOUS
) == 0);
1071 * Data is not optional for freemap chains (we must always be sure
1072 * to copy the data on COW storage allocations).
1074 if (chain
->bref
.type
== HAMMER2_BREF_TYPE_FREEMAP_NODE
||
1075 chain
->bref
.type
== HAMMER2_BREF_TYPE_FREEMAP_LEAF
) {
1076 KKASSERT((chain
->flags
& HAMMER2_CHAIN_INITIAL
) ||
1077 (flags
& HAMMER2_MODIFY_OPTDATA
) == 0);
1081 * Data must be resolved if already assigned, unless explicitly
1082 * flagged otherwise.
1084 if (chain
->data
== NULL
&& (flags
& HAMMER2_MODIFY_OPTDATA
) == 0 &&
1085 (chain
->bref
.data_off
& ~HAMMER2_OFF_MASK_RADIX
)) {
1086 hammer2_chain_load_data(chain
);
1090 * Set MODIFIED to indicate that the chain has been modified.
1091 * Set UPDATE to ensure that the blockref is updated in the parent.
1093 if ((chain
->flags
& HAMMER2_CHAIN_MODIFIED
) == 0) {
1094 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_MODIFIED
);
1095 hammer2_chain_ref(chain
);
1096 hammer2_pfs_memory_inc(chain
->pmp
); /* can be NULL */
1101 if ((chain
->flags
& HAMMER2_CHAIN_UPDATE
) == 0) {
1102 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_UPDATE
);
1103 hammer2_chain_ref(chain
);
1107 * The modification or re-modification requires an allocation and
1110 * We normally always allocate new storage here. If storage exists
1111 * and MODIFY_NOREALLOC is passed in, we do not allocate new storage.
1113 if (chain
!= &hmp
->vchain
&& chain
!= &hmp
->fchain
) {
1114 if ((chain
->bref
.data_off
& ~HAMMER2_OFF_MASK_RADIX
) == 0 ||
1115 ((flags
& HAMMER2_MODIFY_NOREALLOC
) == 0 && newmod
)
1117 hammer2_freemap_alloc(trans
, chain
, chain
->bytes
);
1118 /* XXX failed allocation */
1123 * Update mirror_tid and modify_tid. modify_tid is only updated
1124 * automatically by this function when used from the frontend.
1125 * Flushes and synchronization adjust the flag manually.
1127 * NOTE: chain->pmp could be the device spmp.
1129 chain
->bref
.mirror_tid
= hmp
->voldata
.mirror_tid
+ 1;
1130 if (chain
->pmp
&& (trans
->flags
& (HAMMER2_TRANS_KEEPMODIFY
|
1131 HAMMER2_TRANS_ISFLUSH
)) == 0) {
1132 chain
->bref
.modify_tid
= chain
->pmp
->modify_tid
+ 1;
1136 * Set BMAPUPD to tell the flush code that an existing blockmap entry
1137 * requires updating as well as to tell the delete code that the
1138 * chain's blockref might not exactly match (in terms of physical size
1139 * or block offset) the one in the parent's blocktable. The base key
1140 * of course will still match.
1142 if (chain
->flags
& HAMMER2_CHAIN_BMAPPED
)
1143 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_BMAPUPD
);
1146 * Short-cut data blocks which the caller does not need an actual
1147 * data reference to (aka OPTDATA), as long as the chain does not
1148 * already have a data pointer to the data. This generally means
1149 * that the modifications are being done via the logical buffer cache.
1150 * The INITIAL flag relates only to the device data buffer and thus
1151 * remains unchange in this situation.
1153 if (chain
->bref
.type
== HAMMER2_BREF_TYPE_DATA
&&
1154 (flags
& HAMMER2_MODIFY_OPTDATA
) &&
1155 chain
->data
== NULL
) {
1160 * Clearing the INITIAL flag (for indirect blocks) indicates that
1161 * we've processed the uninitialized storage allocation.
1163 * If this flag is already clear we are likely in a copy-on-write
1164 * situation but we have to be sure NOT to bzero the storage if
1165 * no data is present.
1167 if (chain
->flags
& HAMMER2_CHAIN_INITIAL
) {
1168 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_INITIAL
);
1175 * Instantiate data buffer and possibly execute COW operation
1177 switch(chain
->bref
.type
) {
1178 case HAMMER2_BREF_TYPE_VOLUME
:
1179 case HAMMER2_BREF_TYPE_FREEMAP
:
1181 * The data is embedded, no copy-on-write operation is
1184 KKASSERT(chain
->dio
== NULL
);
1186 case HAMMER2_BREF_TYPE_INODE
:
1187 case HAMMER2_BREF_TYPE_FREEMAP_LEAF
:
1188 case HAMMER2_BREF_TYPE_DATA
:
1189 case HAMMER2_BREF_TYPE_INDIRECT
:
1190 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
1192 * Perform the copy-on-write operation
1194 * zero-fill or copy-on-write depending on whether
1195 * chain->data exists or not and set the dirty state for
1196 * the new buffer. hammer2_io_new() will handle the
1199 KKASSERT(chain
!= &hmp
->vchain
&& chain
!= &hmp
->fchain
);
1202 error
= hammer2_io_new(hmp
, chain
->bref
.data_off
,
1203 chain
->bytes
, &dio
);
1205 error
= hammer2_io_bread(hmp
, chain
->bref
.data_off
,
1206 chain
->bytes
, &dio
);
1208 hammer2_adjreadcounter(&chain
->bref
, chain
->bytes
);
1211 * If an I/O error occurs make sure callers cannot accidently
1212 * modify the old buffer's contents and corrupt the filesystem.
1215 kprintf("hammer2_chain_modify: hmp=%p I/O error\n",
1217 chain
->error
= HAMMER2_ERROR_IO
;
1218 hammer2_io_brelse(&dio
);
1219 hammer2_io_brelse(&chain
->dio
);
1224 bdata
= hammer2_io_data(dio
, chain
->bref
.data_off
);
1227 KKASSERT(chain
->dio
!= NULL
);
1228 if (chain
->data
!= (void *)bdata
) {
1229 bcopy(chain
->data
, bdata
, chain
->bytes
);
1231 } else if (wasinitial
== 0) {
1233 * We have a problem. We were asked to COW but
1234 * we don't have any data to COW with!
1236 panic("hammer2_chain_modify: having a COW %p\n",
1241 * Retire the old buffer, replace with the new. Dirty or
1242 * redirty the new buffer.
1244 * WARNING! The system buffer cache may have already flushed
1245 * the buffer, so we must be sure to [re]dirty it
1246 * for further modification.
1249 hammer2_io_brelse(&chain
->dio
);
1250 chain
->data
= (void *)bdata
;
1252 hammer2_io_setdirty(dio
); /* modified by bcopy above */
1255 panic("hammer2_chain_modify: illegal non-embedded type %d",
1262 * setflush on parent indicating that the parent must recurse down
1263 * to us. Do not call on chain itself which might already have it
1267 hammer2_chain_setflush(trans
, chain
->parent
);
1271 * Volume header data locks
1274 hammer2_voldata_lock(hammer2_dev_t
*hmp
)
1276 lockmgr(&hmp
->vollk
, LK_EXCLUSIVE
);
1280 hammer2_voldata_unlock(hammer2_dev_t
*hmp
)
1282 lockmgr(&hmp
->vollk
, LK_RELEASE
);
1286 hammer2_voldata_modify(hammer2_dev_t
*hmp
)
1288 if ((hmp
->vchain
.flags
& HAMMER2_CHAIN_MODIFIED
) == 0) {
1289 atomic_set_int(&hmp
->vchain
.flags
, HAMMER2_CHAIN_MODIFIED
);
1290 hammer2_chain_ref(&hmp
->vchain
);
1291 hammer2_pfs_memory_inc(hmp
->vchain
.pmp
);
1296 * This function returns the chain at the nearest key within the specified
1297 * range. The returned chain will be referenced but not locked.
1299 * This function will recurse through chain->rbtree as necessary and will
1300 * return a *key_nextp suitable for iteration. *key_nextp is only set if
1301 * the iteration value is less than the current value of *key_nextp.
1303 * The caller should use (*key_nextp) to calculate the actual range of
1304 * the returned element, which will be (key_beg to *key_nextp - 1), because
1305 * there might be another element which is superior to the returned element
1308 * (*key_nextp) can be passed as key_beg in an iteration only while non-NULL
1309 * chains continue to be returned. On EOF (*key_nextp) may overflow since
1310 * it will wind up being (key_end + 1).
1312 * WARNING! Must be called with child's spinlock held. Spinlock remains
1313 * held through the operation.
1315 struct hammer2_chain_find_info
{
1316 hammer2_chain_t
*best
;
1317 hammer2_key_t key_beg
;
1318 hammer2_key_t key_end
;
1319 hammer2_key_t key_next
;
1322 static int hammer2_chain_find_cmp(hammer2_chain_t
*child
, void *data
);
1323 static int hammer2_chain_find_callback(hammer2_chain_t
*child
, void *data
);
1327 hammer2_chain_find(hammer2_chain_t
*parent
, hammer2_key_t
*key_nextp
,
1328 hammer2_key_t key_beg
, hammer2_key_t key_end
)
1330 struct hammer2_chain_find_info info
;
1333 info
.key_beg
= key_beg
;
1334 info
.key_end
= key_end
;
1335 info
.key_next
= *key_nextp
;
1337 RB_SCAN(hammer2_chain_tree
, &parent
->core
.rbtree
,
1338 hammer2_chain_find_cmp
, hammer2_chain_find_callback
,
1340 *key_nextp
= info
.key_next
;
1342 kprintf("chain_find %p %016jx:%016jx next=%016jx\n",
1343 parent
, key_beg
, key_end
, *key_nextp
);
1351 hammer2_chain_find_cmp(hammer2_chain_t
*child
, void *data
)
1353 struct hammer2_chain_find_info
*info
= data
;
1354 hammer2_key_t child_beg
;
1355 hammer2_key_t child_end
;
1357 child_beg
= child
->bref
.key
;
1358 child_end
= child_beg
+ ((hammer2_key_t
)1 << child
->bref
.keybits
) - 1;
1360 if (child_end
< info
->key_beg
)
1362 if (child_beg
> info
->key_end
)
1369 hammer2_chain_find_callback(hammer2_chain_t
*child
, void *data
)
1371 struct hammer2_chain_find_info
*info
= data
;
1372 hammer2_chain_t
*best
;
1373 hammer2_key_t child_end
;
1376 * WARNING! Do not discard DUPLICATED chains, it is possible that
1377 * we are catching an insertion half-way done. If a
1378 * duplicated chain turns out to be the best choice the
1379 * caller will re-check its flags after locking it.
1381 * WARNING! Layerq is scanned forwards, exact matches should keep
1382 * the existing info->best.
1384 if ((best
= info
->best
) == NULL
) {
1386 * No previous best. Assign best
1389 } else if (best
->bref
.key
<= info
->key_beg
&&
1390 child
->bref
.key
<= info
->key_beg
) {
1395 /*info->best = child;*/
1396 } else if (child
->bref
.key
< best
->bref
.key
) {
1398 * Child has a nearer key and best is not flush with key_beg.
1399 * Set best to child. Truncate key_next to the old best key.
1402 if (info
->key_next
> best
->bref
.key
|| info
->key_next
== 0)
1403 info
->key_next
= best
->bref
.key
;
1404 } else if (child
->bref
.key
== best
->bref
.key
) {
1406 * If our current best is flush with the child then this
1407 * is an illegal overlap.
1409 * key_next will automatically be limited to the smaller of
1410 * the two end-points.
1416 * Keep the current best but truncate key_next to the child's
1419 * key_next will also automatically be limited to the smaller
1420 * of the two end-points (probably not necessary for this case
1421 * but we do it anyway).
1423 if (info
->key_next
> child
->bref
.key
|| info
->key_next
== 0)
1424 info
->key_next
= child
->bref
.key
;
1428 * Always truncate key_next based on child's end-of-range.
1430 child_end
= child
->bref
.key
+ ((hammer2_key_t
)1 << child
->bref
.keybits
);
1431 if (child_end
&& (info
->key_next
> child_end
|| info
->key_next
== 0))
1432 info
->key_next
= child_end
;
1438 * Retrieve the specified chain from a media blockref, creating the
1439 * in-memory chain structure which reflects it.
1441 * To handle insertion races pass the INSERT_RACE flag along with the
1442 * generation number of the core. NULL will be returned if the generation
1443 * number changes before we have a chance to insert the chain. Insert
1444 * races can occur because the parent might be held shared.
1446 * Caller must hold the parent locked shared or exclusive since we may
1447 * need the parent's bref array to find our block.
1449 * WARNING! chain->pmp is always set to NULL for any chain representing
1450 * part of the super-root topology.
1453 hammer2_chain_get(hammer2_chain_t
*parent
, int generation
,
1454 hammer2_blockref_t
*bref
)
1456 hammer2_dev_t
*hmp
= parent
->hmp
;
1457 hammer2_chain_t
*chain
;
1461 * Allocate a chain structure representing the existing media
1462 * entry. Resulting chain has one ref and is not locked.
1464 if (bref
->flags
& HAMMER2_BREF_FLAG_PFSROOT
)
1465 chain
= hammer2_chain_alloc(hmp
, NULL
, NULL
, bref
);
1467 chain
= hammer2_chain_alloc(hmp
, parent
->pmp
, NULL
, bref
);
1468 /* ref'd chain returned */
1471 * Flag that the chain is in the parent's blockmap so delete/flush
1472 * knows what to do with it.
1474 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_BMAPPED
);
1477 * Link the chain into its parent. A spinlock is required to safely
1478 * access the RBTREE, and it is possible to collide with another
1479 * hammer2_chain_get() operation because the caller might only hold
1480 * a shared lock on the parent.
1482 KKASSERT(parent
->refs
> 0);
1483 error
= hammer2_chain_insert(parent
, chain
,
1484 HAMMER2_CHAIN_INSERT_SPIN
|
1485 HAMMER2_CHAIN_INSERT_RACE
,
1488 KKASSERT((chain
->flags
& HAMMER2_CHAIN_ONRBTREE
) == 0);
1489 kprintf("chain %p get race\n", chain
);
1490 hammer2_chain_drop(chain
);
1493 KKASSERT(chain
->flags
& HAMMER2_CHAIN_ONRBTREE
);
1497 * Return our new chain referenced but not locked, or NULL if
1504 * Lookup initialization/completion API
1507 hammer2_chain_lookup_init(hammer2_chain_t
*parent
, int flags
)
1509 hammer2_chain_ref(parent
);
1510 if (flags
& HAMMER2_LOOKUP_SHARED
) {
1511 hammer2_chain_lock(parent
, HAMMER2_RESOLVE_ALWAYS
|
1512 HAMMER2_RESOLVE_SHARED
);
1514 hammer2_chain_lock(parent
, HAMMER2_RESOLVE_ALWAYS
);
1520 hammer2_chain_lookup_done(hammer2_chain_t
*parent
)
1523 hammer2_chain_unlock(parent
);
1524 hammer2_chain_drop(parent
);
1530 hammer2_chain_getparent(hammer2_chain_t
**parentp
, int how
)
1532 hammer2_chain_t
*oparent
;
1533 hammer2_chain_t
*nparent
;
1536 * Be careful of order, oparent must be unlocked before nparent
1537 * is locked below to avoid a deadlock.
1540 hammer2_spin_ex(&oparent
->core
.spin
);
1541 nparent
= oparent
->parent
;
1542 hammer2_chain_ref(nparent
);
1543 hammer2_spin_unex(&oparent
->core
.spin
);
1545 hammer2_chain_unlock(oparent
);
1546 hammer2_chain_drop(oparent
);
1550 hammer2_chain_lock(nparent
, how
);
1557 * Locate the first chain whos key range overlaps (key_beg, key_end) inclusive.
1558 * (*parentp) typically points to an inode but can also point to a related
1559 * indirect block and this function will recurse upwards and find the inode
1562 * (*parentp) must be exclusively locked and referenced and can be an inode
1563 * or an existing indirect block within the inode.
1565 * On return (*parentp) will be modified to point at the deepest parent chain
1566 * element encountered during the search, as a helper for an insertion or
1567 * deletion. The new (*parentp) will be locked and referenced and the old
1568 * will be unlocked and dereferenced (no change if they are both the same).
1570 * The matching chain will be returned exclusively locked. If NOLOCK is
1571 * requested the chain will be returned only referenced. Note that the
1572 * parent chain must always be locked shared or exclusive, matching the
1573 * HAMMER2_LOOKUP_SHARED flag. We can conceivably lock it SHARED temporarily
1574 * when NOLOCK is specified but that complicates matters if *parentp must
1575 * inherit the chain.
1577 * NOLOCK also implies NODATA, since an unlocked chain usually has a NULL
1578 * data pointer or can otherwise be in flux.
1580 * NULL is returned if no match was found, but (*parentp) will still
1581 * potentially be adjusted.
1583 * If a fatal error occurs (typically an I/O error), a dummy chain is
1584 * returned with chain->error and error-identifying information set. This
1585 * chain will assert if you try to do anything fancy with it.
1587 * XXX Depending on where the error occurs we should allow continued iteration.
1589 * On return (*key_nextp) will point to an iterative value for key_beg.
1590 * (If NULL is returned (*key_nextp) is set to (key_end + 1)).
1592 * This function will also recurse up the chain if the key is not within the
1593 * current parent's range. (*parentp) can never be set to NULL. An iteration
1594 * can simply allow (*parentp) to float inside the loop.
1596 * NOTE! chain->data is not always resolved. By default it will not be
1597 * resolved for BREF_TYPE_DATA, FREEMAP_NODE, or FREEMAP_LEAF. Use
1598 * HAMMER2_LOOKUP_ALWAYS to force resolution (but be careful w/
1599 * BREF_TYPE_DATA as the device buffer can alias the logical file
1603 hammer2_chain_lookup(hammer2_chain_t
**parentp
, hammer2_key_t
*key_nextp
,
1604 hammer2_key_t key_beg
, hammer2_key_t key_end
,
1605 int *cache_indexp
, int flags
)
1608 hammer2_chain_t
*parent
;
1609 hammer2_chain_t
*chain
;
1610 hammer2_blockref_t
*base
;
1611 hammer2_blockref_t
*bref
;
1612 hammer2_blockref_t bcopy
;
1613 hammer2_key_t scan_beg
;
1614 hammer2_key_t scan_end
;
1616 int how_always
= HAMMER2_RESOLVE_ALWAYS
;
1617 int how_maybe
= HAMMER2_RESOLVE_MAYBE
;
1620 int maxloops
= 300000;
1622 if (flags
& HAMMER2_LOOKUP_ALWAYS
) {
1623 how_maybe
= how_always
;
1624 how
= HAMMER2_RESOLVE_ALWAYS
;
1625 } else if (flags
& (HAMMER2_LOOKUP_NODATA
| HAMMER2_LOOKUP_NOLOCK
)) {
1626 how
= HAMMER2_RESOLVE_NEVER
;
1628 how
= HAMMER2_RESOLVE_MAYBE
;
1630 if (flags
& HAMMER2_LOOKUP_SHARED
) {
1631 how_maybe
|= HAMMER2_RESOLVE_SHARED
;
1632 how_always
|= HAMMER2_RESOLVE_SHARED
;
1633 how
|= HAMMER2_RESOLVE_SHARED
;
1637 * Recurse (*parentp) upward if necessary until the parent completely
1638 * encloses the key range or we hit the inode.
1640 * This function handles races against the flusher doing a delete-
1641 * duplicate above us and re-homes the parent to the duplicate in
1642 * that case, otherwise we'd wind up recursing down a stale chain.
1647 while (parent
->bref
.type
== HAMMER2_BREF_TYPE_INDIRECT
||
1648 parent
->bref
.type
== HAMMER2_BREF_TYPE_FREEMAP_NODE
) {
1649 scan_beg
= parent
->bref
.key
;
1650 scan_end
= scan_beg
+
1651 ((hammer2_key_t
)1 << parent
->bref
.keybits
) - 1;
1652 if (key_beg
>= scan_beg
&& key_end
<= scan_end
)
1654 parent
= hammer2_chain_getparent(parentp
, how_maybe
);
1658 if (--maxloops
== 0)
1659 panic("hammer2_chain_lookup: maxloops");
1661 * Locate the blockref array. Currently we do a fully associative
1662 * search through the array.
1664 switch(parent
->bref
.type
) {
1665 case HAMMER2_BREF_TYPE_INODE
:
1667 * Special shortcut for embedded data returns the inode
1668 * itself. Callers must detect this condition and access
1669 * the embedded data (the strategy code does this for us).
1671 * This is only applicable to regular files and softlinks.
1673 if (parent
->data
->ipdata
.meta
.op_flags
&
1674 HAMMER2_OPFLAG_DIRECTDATA
) {
1675 if (flags
& HAMMER2_LOOKUP_NODIRECT
) {
1677 *key_nextp
= key_end
+ 1;
1680 hammer2_chain_ref(parent
);
1681 if ((flags
& HAMMER2_LOOKUP_NOLOCK
) == 0)
1682 hammer2_chain_lock(parent
, how_always
);
1683 *key_nextp
= key_end
+ 1;
1686 base
= &parent
->data
->ipdata
.u
.blockset
.blockref
[0];
1687 count
= HAMMER2_SET_COUNT
;
1689 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
1690 case HAMMER2_BREF_TYPE_INDIRECT
:
1692 * Handle MATCHIND on the parent
1694 if (flags
& HAMMER2_LOOKUP_MATCHIND
) {
1695 scan_beg
= parent
->bref
.key
;
1696 scan_end
= scan_beg
+
1697 ((hammer2_key_t
)1 << parent
->bref
.keybits
) - 1;
1698 if (key_beg
== scan_beg
&& key_end
== scan_end
) {
1700 hammer2_chain_ref(chain
);
1701 hammer2_chain_lock(chain
, how_maybe
);
1702 *key_nextp
= scan_end
+ 1;
1707 * Optimize indirect blocks in the INITIAL state to avoid
1710 if (parent
->flags
& HAMMER2_CHAIN_INITIAL
) {
1713 if (parent
->data
== NULL
)
1714 panic("parent->data is NULL");
1715 base
= &parent
->data
->npdata
[0];
1717 count
= parent
->bytes
/ sizeof(hammer2_blockref_t
);
1719 case HAMMER2_BREF_TYPE_VOLUME
:
1720 base
= &hmp
->voldata
.sroot_blockset
.blockref
[0];
1721 count
= HAMMER2_SET_COUNT
;
1723 case HAMMER2_BREF_TYPE_FREEMAP
:
1724 base
= &hmp
->voldata
.freemap_blockset
.blockref
[0];
1725 count
= HAMMER2_SET_COUNT
;
1728 kprintf("hammer2_chain_lookup: unrecognized "
1729 "blockref(B) type: %d",
1732 tsleep(&base
, 0, "dead", 0);
1733 panic("hammer2_chain_lookup: unrecognized "
1734 "blockref(B) type: %d",
1736 base
= NULL
; /* safety */
1737 count
= 0; /* safety */
1741 * Merged scan to find next candidate.
1743 * hammer2_base_*() functions require the parent->core.live_* fields
1744 * to be synchronized.
1746 * We need to hold the spinlock to access the block array and RB tree
1747 * and to interlock chain creation.
1749 if ((parent
->flags
& HAMMER2_CHAIN_COUNTEDBREFS
) == 0)
1750 hammer2_chain_countbrefs(parent
, base
, count
);
1755 hammer2_spin_ex(&parent
->core
.spin
);
1756 chain
= hammer2_combined_find(parent
, base
, count
,
1757 cache_indexp
, key_nextp
,
1760 generation
= parent
->core
.generation
;
1763 * Exhausted parent chain, iterate.
1766 hammer2_spin_unex(&parent
->core
.spin
);
1767 if (key_beg
== key_end
) /* short cut single-key case */
1771 * Stop if we reached the end of the iteration.
1773 if (parent
->bref
.type
!= HAMMER2_BREF_TYPE_INDIRECT
&&
1774 parent
->bref
.type
!= HAMMER2_BREF_TYPE_FREEMAP_NODE
) {
1779 * Calculate next key, stop if we reached the end of the
1780 * iteration, otherwise go up one level and loop.
1782 key_beg
= parent
->bref
.key
+
1783 ((hammer2_key_t
)1 << parent
->bref
.keybits
);
1784 if (key_beg
== 0 || key_beg
> key_end
)
1786 parent
= hammer2_chain_getparent(parentp
, how_maybe
);
1791 * Selected from blockref or in-memory chain.
1793 if (chain
== NULL
) {
1795 hammer2_spin_unex(&parent
->core
.spin
);
1796 chain
= hammer2_chain_get(parent
, generation
,
1798 if (chain
== NULL
) {
1799 kprintf("retry lookup parent %p keys %016jx:%016jx\n",
1800 parent
, key_beg
, key_end
);
1803 if (bcmp(&bcopy
, bref
, sizeof(bcopy
))) {
1804 hammer2_chain_drop(chain
);
1808 hammer2_chain_ref(chain
);
1809 hammer2_spin_unex(&parent
->core
.spin
);
1813 * chain is referenced but not locked. We must lock the chain
1814 * to obtain definitive DUPLICATED/DELETED state
1816 if (chain
->bref
.type
== HAMMER2_BREF_TYPE_INDIRECT
||
1817 chain
->bref
.type
== HAMMER2_BREF_TYPE_FREEMAP_NODE
) {
1818 hammer2_chain_lock(chain
, how_maybe
);
1820 hammer2_chain_lock(chain
, how
);
1824 * Skip deleted chains (XXX cache 'i' end-of-block-array? XXX)
1826 * NOTE: Chain's key range is not relevant as there might be
1827 * one-offs within the range that are not deleted.
1829 * NOTE: Lookups can race delete-duplicate because
1830 * delete-duplicate does not lock the parent's core
1831 * (they just use the spinlock on the core). We must
1832 * check for races by comparing the DUPLICATED flag before
1833 * releasing the spinlock with the flag after locking the
1836 if (chain
->flags
& HAMMER2_CHAIN_DELETED
) {
1837 hammer2_chain_unlock(chain
);
1838 hammer2_chain_drop(chain
);
1839 key_beg
= *key_nextp
;
1840 if (key_beg
== 0 || key_beg
> key_end
)
1846 * If the chain element is an indirect block it becomes the new
1847 * parent and we loop on it. We must maintain our top-down locks
1848 * to prevent the flusher from interfering (i.e. doing a
1849 * delete-duplicate and leaving us recursing down a deleted chain).
1851 * The parent always has to be locked with at least RESOLVE_MAYBE
1852 * so we can access its data. It might need a fixup if the caller
1853 * passed incompatible flags. Be careful not to cause a deadlock
1854 * as a data-load requires an exclusive lock.
1856 * If HAMMER2_LOOKUP_MATCHIND is set and the indirect block's key
1857 * range is within the requested key range we return the indirect
1858 * block and do NOT loop. This is usually only used to acquire
1861 if (chain
->bref
.type
== HAMMER2_BREF_TYPE_INDIRECT
||
1862 chain
->bref
.type
== HAMMER2_BREF_TYPE_FREEMAP_NODE
) {
1863 hammer2_chain_unlock(parent
);
1864 hammer2_chain_drop(parent
);
1865 *parentp
= parent
= chain
;
1870 * All done, return the chain.
1872 * If the caller does not want a locked chain, replace the lock with
1873 * a ref. Perhaps this can eventually be optimized to not obtain the
1874 * lock in the first place for situations where the data does not
1875 * need to be resolved.
1878 if (flags
& HAMMER2_LOOKUP_NOLOCK
)
1879 hammer2_chain_unlock(chain
);
1886 * After having issued a lookup we can iterate all matching keys.
1888 * If chain is non-NULL we continue the iteration from just after it's index.
1890 * If chain is NULL we assume the parent was exhausted and continue the
1891 * iteration at the next parent.
1893 * If a fatal error occurs (typically an I/O error), a dummy chain is
1894 * returned with chain->error and error-identifying information set. This
1895 * chain will assert if you try to do anything fancy with it.
1897 * XXX Depending on where the error occurs we should allow continued iteration.
1899 * parent must be locked on entry and remains locked throughout. chain's
1900 * lock status must match flags. Chain is always at least referenced.
1902 * WARNING! The MATCHIND flag does not apply to this function.
1905 hammer2_chain_next(hammer2_chain_t
**parentp
, hammer2_chain_t
*chain
,
1906 hammer2_key_t
*key_nextp
,
1907 hammer2_key_t key_beg
, hammer2_key_t key_end
,
1908 int *cache_indexp
, int flags
)
1910 hammer2_chain_t
*parent
;
1914 * Calculate locking flags for upward recursion.
1916 how_maybe
= HAMMER2_RESOLVE_MAYBE
;
1917 if (flags
& HAMMER2_LOOKUP_SHARED
)
1918 how_maybe
|= HAMMER2_RESOLVE_SHARED
;
1923 * Calculate the next index and recalculate the parent if necessary.
1926 key_beg
= chain
->bref
.key
+
1927 ((hammer2_key_t
)1 << chain
->bref
.keybits
);
1928 if ((flags
& (HAMMER2_LOOKUP_NOLOCK
|
1929 HAMMER2_LOOKUP_NOUNLOCK
)) == 0) {
1930 hammer2_chain_unlock(chain
);
1932 hammer2_chain_drop(chain
);
1935 * chain invalid past this point, but we can still do a
1936 * pointer comparison w/parent.
1938 * Any scan where the lookup returned degenerate data embedded
1939 * in the inode has an invalid index and must terminate.
1941 if (chain
== parent
)
1943 if (key_beg
== 0 || key_beg
> key_end
)
1946 } else if (parent
->bref
.type
!= HAMMER2_BREF_TYPE_INDIRECT
&&
1947 parent
->bref
.type
!= HAMMER2_BREF_TYPE_FREEMAP_NODE
) {
1949 * We reached the end of the iteration.
1954 * Continue iteration with next parent unless the current
1955 * parent covers the range.
1957 key_beg
= parent
->bref
.key
+
1958 ((hammer2_key_t
)1 << parent
->bref
.keybits
);
1959 if (key_beg
== 0 || key_beg
> key_end
)
1961 parent
= hammer2_chain_getparent(parentp
, how_maybe
);
1967 return (hammer2_chain_lookup(parentp
, key_nextp
,
1969 cache_indexp
, flags
));
1973 * The raw scan function is similar to lookup/next but does not seek to a key.
1974 * Blockrefs are iterated via first_chain = (parent, NULL) and
1975 * next_chain = (parent, chain).
1977 * The passed-in parent must be locked and its data resolved. The returned
1978 * chain will be locked. Pass chain == NULL to acquire the first sub-chain
1979 * under parent and then iterate with the passed-in chain (which this
1980 * function will unlock).
1983 hammer2_chain_scan(hammer2_chain_t
*parent
, hammer2_chain_t
*chain
,
1984 int *cache_indexp
, int flags
)
1987 hammer2_blockref_t
*base
;
1988 hammer2_blockref_t
*bref
;
1989 hammer2_blockref_t bcopy
;
1991 hammer2_key_t next_key
;
1993 int how_always
= HAMMER2_RESOLVE_ALWAYS
;
1994 int how_maybe
= HAMMER2_RESOLVE_MAYBE
;
1997 int maxloops
= 300000;
2002 * Scan flags borrowed from lookup.
2004 if (flags
& HAMMER2_LOOKUP_ALWAYS
) {
2005 how_maybe
= how_always
;
2006 how
= HAMMER2_RESOLVE_ALWAYS
;
2007 } else if (flags
& (HAMMER2_LOOKUP_NODATA
| HAMMER2_LOOKUP_NOLOCK
)) {
2008 how
= HAMMER2_RESOLVE_NEVER
;
2010 how
= HAMMER2_RESOLVE_MAYBE
;
2012 if (flags
& HAMMER2_LOOKUP_SHARED
) {
2013 how_maybe
|= HAMMER2_RESOLVE_SHARED
;
2014 how_always
|= HAMMER2_RESOLVE_SHARED
;
2015 how
|= HAMMER2_RESOLVE_SHARED
;
2019 * Calculate key to locate first/next element, unlocking the previous
2020 * element as we go. Be careful, the key calculation can overflow.
2023 key
= chain
->bref
.key
+
2024 ((hammer2_key_t
)1 << chain
->bref
.keybits
);
2025 hammer2_chain_unlock(chain
);
2026 hammer2_chain_drop(chain
);
2035 KKASSERT(parent
->error
== 0); /* XXX case not handled yet */
2036 if (--maxloops
== 0)
2037 panic("hammer2_chain_scan: maxloops");
2039 * Locate the blockref array. Currently we do a fully associative
2040 * search through the array.
2042 switch(parent
->bref
.type
) {
2043 case HAMMER2_BREF_TYPE_INODE
:
2045 * An inode with embedded data has no sub-chains.
2047 if (parent
->data
->ipdata
.meta
.op_flags
&
2048 HAMMER2_OPFLAG_DIRECTDATA
) {
2051 base
= &parent
->data
->ipdata
.u
.blockset
.blockref
[0];
2052 count
= HAMMER2_SET_COUNT
;
2054 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
2055 case HAMMER2_BREF_TYPE_INDIRECT
:
2057 * Optimize indirect blocks in the INITIAL state to avoid
2060 if (parent
->flags
& HAMMER2_CHAIN_INITIAL
) {
2063 if (parent
->data
== NULL
)
2064 panic("parent->data is NULL");
2065 base
= &parent
->data
->npdata
[0];
2067 count
= parent
->bytes
/ sizeof(hammer2_blockref_t
);
2069 case HAMMER2_BREF_TYPE_VOLUME
:
2070 base
= &hmp
->voldata
.sroot_blockset
.blockref
[0];
2071 count
= HAMMER2_SET_COUNT
;
2073 case HAMMER2_BREF_TYPE_FREEMAP
:
2074 base
= &hmp
->voldata
.freemap_blockset
.blockref
[0];
2075 count
= HAMMER2_SET_COUNT
;
2078 panic("hammer2_chain_lookup: unrecognized blockref type: %d",
2080 base
= NULL
; /* safety */
2081 count
= 0; /* safety */
2085 * Merged scan to find next candidate.
2087 * hammer2_base_*() functions require the parent->core.live_* fields
2088 * to be synchronized.
2090 * We need to hold the spinlock to access the block array and RB tree
2091 * and to interlock chain creation.
2093 if ((parent
->flags
& HAMMER2_CHAIN_COUNTEDBREFS
) == 0)
2094 hammer2_chain_countbrefs(parent
, base
, count
);
2097 hammer2_spin_ex(&parent
->core
.spin
);
2098 chain
= hammer2_combined_find(parent
, base
, count
,
2099 cache_indexp
, &next_key
,
2100 key
, HAMMER2_KEY_MAX
,
2102 generation
= parent
->core
.generation
;
2105 * Exhausted parent chain, we're done.
2108 hammer2_spin_unex(&parent
->core
.spin
);
2109 KKASSERT(chain
== NULL
);
2114 * Selected from blockref or in-memory chain.
2116 if (chain
== NULL
) {
2118 hammer2_spin_unex(&parent
->core
.spin
);
2119 chain
= hammer2_chain_get(parent
, generation
, &bcopy
);
2120 if (chain
== NULL
) {
2121 kprintf("retry scan parent %p keys %016jx\n",
2125 if (bcmp(&bcopy
, bref
, sizeof(bcopy
))) {
2126 hammer2_chain_drop(chain
);
2131 hammer2_chain_ref(chain
);
2132 hammer2_spin_unex(&parent
->core
.spin
);
2136 * chain is referenced but not locked. We must lock the chain
2137 * to obtain definitive DUPLICATED/DELETED state
2139 hammer2_chain_lock(chain
, how
);
2142 * Skip deleted chains (XXX cache 'i' end-of-block-array? XXX)
2144 * NOTE: chain's key range is not relevant as there might be
2145 * one-offs within the range that are not deleted.
2147 * NOTE: XXX this could create problems with scans used in
2148 * situations other than mount-time recovery.
2150 * NOTE: Lookups can race delete-duplicate because
2151 * delete-duplicate does not lock the parent's core
2152 * (they just use the spinlock on the core). We must
2153 * check for races by comparing the DUPLICATED flag before
2154 * releasing the spinlock with the flag after locking the
2157 if (chain
->flags
& HAMMER2_CHAIN_DELETED
) {
2158 hammer2_chain_unlock(chain
);
2159 hammer2_chain_drop(chain
);
2170 * All done, return the chain or NULL
2176 * Create and return a new hammer2 system memory structure of the specified
2177 * key, type and size and insert it under (*parentp). This is a full
2178 * insertion, based on the supplied key/keybits, and may involve creating
2179 * indirect blocks and moving other chains around via delete/duplicate.
2181 * THE CALLER MUST HAVE ALREADY PROPERLY SEEKED (*parentp) TO THE INSERTION
2182 * POINT SANS ANY REQUIRED INDIRECT BLOCK CREATIONS DUE TO THE ARRAY BEING
2183 * FULL. This typically means that the caller is creating the chain after
2184 * doing a hammer2_chain_lookup().
2186 * (*parentp) must be exclusive locked and may be replaced on return
2187 * depending on how much work the function had to do.
2189 * (*parentp) must not be errored or this function will assert.
2191 * (*chainp) usually starts out NULL and returns the newly created chain,
2192 * but if the caller desires the caller may allocate a disconnected chain
2193 * and pass it in instead.
2195 * This function should NOT be used to insert INDIRECT blocks. It is
2196 * typically used to create/insert inodes and data blocks.
2198 * Caller must pass-in an exclusively locked parent the new chain is to
2199 * be inserted under, and optionally pass-in a disconnected, exclusively
2200 * locked chain to insert (else we create a new chain). The function will
2201 * adjust (*parentp) as necessary, create or connect the chain, and
2202 * return an exclusively locked chain in *chainp.
2204 * When creating a PFSROOT inode under the super-root, pmp is typically NULL
2205 * and will be reassigned.
2208 hammer2_chain_create(hammer2_trans_t
*trans
, hammer2_chain_t
**parentp
,
2209 hammer2_chain_t
**chainp
, hammer2_pfs_t
*pmp
,
2210 hammer2_key_t key
, int keybits
, int type
, size_t bytes
,
2214 hammer2_chain_t
*chain
;
2215 hammer2_chain_t
*parent
;
2216 hammer2_blockref_t
*base
;
2217 hammer2_blockref_t dummy
;
2221 int maxloops
= 300000;
2224 * Topology may be crossing a PFS boundary.
2227 KKASSERT(hammer2_mtx_owned(&parent
->lock
));
2228 KKASSERT(parent
->error
== 0);
2232 if (chain
== NULL
) {
2234 * First allocate media space and construct the dummy bref,
2235 * then allocate the in-memory chain structure. Set the
2236 * INITIAL flag for fresh chains which do not have embedded
2239 bzero(&dummy
, sizeof(dummy
));
2242 dummy
.keybits
= keybits
;
2243 dummy
.data_off
= hammer2_getradix(bytes
);
2244 dummy
.methods
= parent
->bref
.methods
;
2245 chain
= hammer2_chain_alloc(hmp
, pmp
, trans
, &dummy
);
2248 * Lock the chain manually, chain_lock will load the chain
2249 * which we do NOT want to do. (note: chain->refs is set
2250 * to 1 by chain_alloc() for us, but lockcnt is not).
2253 hammer2_mtx_ex(&chain
->lock
);
2257 * Set INITIAL to optimize I/O. The flag will generally be
2258 * processed when we call hammer2_chain_modify().
2260 * Recalculate bytes to reflect the actual media block
2263 bytes
= (hammer2_off_t
)1 <<
2264 (int)(chain
->bref
.data_off
& HAMMER2_OFF_MASK_RADIX
);
2265 chain
->bytes
= bytes
;
2268 case HAMMER2_BREF_TYPE_VOLUME
:
2269 case HAMMER2_BREF_TYPE_FREEMAP
:
2270 panic("hammer2_chain_create: called with volume type");
2272 case HAMMER2_BREF_TYPE_INDIRECT
:
2273 panic("hammer2_chain_create: cannot be used to"
2274 "create indirect block");
2276 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
2277 panic("hammer2_chain_create: cannot be used to"
2278 "create freemap root or node");
2280 case HAMMER2_BREF_TYPE_FREEMAP_LEAF
:
2281 KKASSERT(bytes
== sizeof(chain
->data
->bmdata
));
2283 case HAMMER2_BREF_TYPE_INODE
:
2284 case HAMMER2_BREF_TYPE_DATA
:
2287 * leave chain->data NULL, set INITIAL
2289 KKASSERT(chain
->data
== NULL
);
2290 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_INITIAL
);
2295 * We are reattaching a previously deleted chain, possibly
2296 * under a new parent and possibly with a new key/keybits.
2297 * The chain does not have to be in a modified state. The
2298 * UPDATE flag will be set later on in this routine.
2300 * Do NOT mess with the current state of the INITIAL flag.
2302 chain
->bref
.key
= key
;
2303 chain
->bref
.keybits
= keybits
;
2304 if (chain
->flags
& HAMMER2_CHAIN_DELETED
)
2305 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_DELETED
);
2306 KKASSERT(chain
->parent
== NULL
);
2308 if (flags
& HAMMER2_INSERT_PFSROOT
)
2309 chain
->bref
.flags
|= HAMMER2_BREF_FLAG_PFSROOT
;
2311 chain
->bref
.flags
&= ~HAMMER2_BREF_FLAG_PFSROOT
;
2314 * Calculate how many entries we have in the blockref array and
2315 * determine if an indirect block is required.
2318 if (--maxloops
== 0)
2319 panic("hammer2_chain_create: maxloops");
2321 switch(parent
->bref
.type
) {
2322 case HAMMER2_BREF_TYPE_INODE
:
2323 KKASSERT((parent
->data
->ipdata
.meta
.op_flags
&
2324 HAMMER2_OPFLAG_DIRECTDATA
) == 0);
2325 KKASSERT(parent
->data
!= NULL
);
2326 base
= &parent
->data
->ipdata
.u
.blockset
.blockref
[0];
2327 count
= HAMMER2_SET_COUNT
;
2329 case HAMMER2_BREF_TYPE_INDIRECT
:
2330 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
2331 if (parent
->flags
& HAMMER2_CHAIN_INITIAL
)
2334 base
= &parent
->data
->npdata
[0];
2335 count
= parent
->bytes
/ sizeof(hammer2_blockref_t
);
2337 case HAMMER2_BREF_TYPE_VOLUME
:
2338 KKASSERT(parent
->data
!= NULL
);
2339 base
= &hmp
->voldata
.sroot_blockset
.blockref
[0];
2340 count
= HAMMER2_SET_COUNT
;
2342 case HAMMER2_BREF_TYPE_FREEMAP
:
2343 KKASSERT(parent
->data
!= NULL
);
2344 base
= &hmp
->voldata
.freemap_blockset
.blockref
[0];
2345 count
= HAMMER2_SET_COUNT
;
2348 panic("hammer2_chain_create: unrecognized blockref type: %d",
2356 * Make sure we've counted the brefs
2358 if ((parent
->flags
& HAMMER2_CHAIN_COUNTEDBREFS
) == 0)
2359 hammer2_chain_countbrefs(parent
, base
, count
);
2361 KKASSERT(parent
->core
.live_count
>= 0 &&
2362 parent
->core
.live_count
<= count
);
2365 * If no free blockref could be found we must create an indirect
2366 * block and move a number of blockrefs into it. With the parent
2367 * locked we can safely lock each child in order to delete+duplicate
2368 * it without causing a deadlock.
2370 * This may return the new indirect block or the old parent depending
2371 * on where the key falls. NULL is returned on error.
2373 if (parent
->core
.live_count
== count
) {
2374 hammer2_chain_t
*nparent
;
2376 nparent
= hammer2_chain_create_indirect(trans
, parent
,
2379 if (nparent
== NULL
) {
2381 hammer2_chain_drop(chain
);
2385 if (parent
!= nparent
) {
2386 hammer2_chain_unlock(parent
);
2387 hammer2_chain_drop(parent
);
2388 parent
= *parentp
= nparent
;
2394 * Link the chain into its parent.
2396 if (chain
->parent
!= NULL
)
2397 panic("hammer2: hammer2_chain_create: chain already connected");
2398 KKASSERT(chain
->parent
== NULL
);
2399 hammer2_chain_insert(parent
, chain
,
2400 HAMMER2_CHAIN_INSERT_SPIN
|
2401 HAMMER2_CHAIN_INSERT_LIVE
,
2406 * Mark the newly created chain modified. This will cause
2407 * UPDATE to be set and process the INITIAL flag.
2409 * Device buffers are not instantiated for DATA elements
2410 * as these are handled by logical buffers.
2412 * Indirect and freemap node indirect blocks are handled
2413 * by hammer2_chain_create_indirect() and not by this
2416 * Data for all other bref types is expected to be
2417 * instantiated (INODE, LEAF).
2419 switch(chain
->bref
.type
) {
2420 case HAMMER2_BREF_TYPE_DATA
:
2421 case HAMMER2_BREF_TYPE_FREEMAP_LEAF
:
2422 case HAMMER2_BREF_TYPE_INODE
:
2423 hammer2_chain_modify(trans
, chain
,
2424 HAMMER2_MODIFY_OPTDATA
);
2428 * Remaining types are not supported by this function.
2429 * In particular, INDIRECT and LEAF_NODE types are
2430 * handled by create_indirect().
2432 panic("hammer2_chain_create: bad type: %d",
2439 * When reconnecting a chain we must set UPDATE and
2440 * setflush so the flush recognizes that it must update
2441 * the bref in the parent.
2443 if ((chain
->flags
& HAMMER2_CHAIN_UPDATE
) == 0) {
2444 hammer2_chain_ref(chain
);
2445 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_UPDATE
);
2450 * We must setflush(parent) to ensure that it recurses through to
2451 * chain. setflush(chain) might not work because ONFLUSH is possibly
2452 * already set in the chain (so it won't recurse up to set it in the
2455 hammer2_chain_setflush(trans
, parent
);
2464 * Move the chain from its old parent to a new parent. The chain must have
2465 * already been deleted or already disconnected (or never associated) with
2466 * a parent. The chain is reassociated with the new parent and the deleted
2467 * flag will be cleared (no longer deleted). The chain's modification state
2470 * THE CALLER MUST HAVE ALREADY PROPERLY SEEKED (parent) TO THE INSERTION
2471 * POINT SANS ANY REQUIRED INDIRECT BLOCK CREATIONS DUE TO THE ARRAY BEING
2472 * FULL. This typically means that the caller is creating the chain after
2473 * doing a hammer2_chain_lookup().
2475 * A non-NULL bref is typically passed when key and keybits must be overridden.
2476 * Note that hammer2_cluster_duplicate() *ONLY* uses the key and keybits fields
2477 * from a passed-in bref and uses the old chain's bref for everything else.
2479 * Neither (parent) or (chain) can be errored.
2481 * If (parent) is non-NULL then the new duplicated chain is inserted under
2484 * If (parent) is NULL then the newly duplicated chain is not inserted
2485 * anywhere, similar to if it had just been chain_alloc()'d (suitable for
2486 * passing into hammer2_chain_create() after this function returns).
2488 * WARNING! This function calls create which means it can insert indirect
2489 * blocks. This can cause other unrelated chains in the parent to
2490 * be moved to a newly inserted indirect block in addition to the
2494 hammer2_chain_rename(hammer2_trans_t
*trans
, hammer2_blockref_t
*bref
,
2495 hammer2_chain_t
**parentp
, hammer2_chain_t
*chain
,
2499 hammer2_chain_t
*parent
;
2503 * WARNING! We should never resolve DATA to device buffers
2504 * (XXX allow it if the caller did?), and since
2505 * we currently do not have the logical buffer cache
2506 * buffer in-hand to fix its cached physical offset
2507 * we also force the modify code to not COW it. XXX
2510 KKASSERT(chain
->parent
== NULL
);
2511 KKASSERT(chain
->error
== 0);
2514 * Now create a duplicate of the chain structure, associating
2515 * it with the same core, making it the same size, pointing it
2516 * to the same bref (the same media block).
2519 bref
= &chain
->bref
;
2520 bytes
= (hammer2_off_t
)1 <<
2521 (int)(bref
->data_off
& HAMMER2_OFF_MASK_RADIX
);
2524 * If parent is not NULL the duplicated chain will be entered under
2525 * the parent and the UPDATE bit set to tell flush to update
2528 * We must setflush(parent) to ensure that it recurses through to
2529 * chain. setflush(chain) might not work because ONFLUSH is possibly
2530 * already set in the chain (so it won't recurse up to set it in the
2533 * Having both chains locked is extremely important for atomicy.
2535 if (parentp
&& (parent
= *parentp
) != NULL
) {
2536 KKASSERT(hammer2_mtx_owned(&parent
->lock
));
2537 KKASSERT(parent
->refs
> 0);
2538 KKASSERT(parent
->error
== 0);
2540 hammer2_chain_create(trans
, parentp
, &chain
, chain
->pmp
,
2541 bref
->key
, bref
->keybits
, bref
->type
,
2542 chain
->bytes
, flags
);
2543 KKASSERT(chain
->flags
& HAMMER2_CHAIN_UPDATE
);
2544 hammer2_chain_setflush(trans
, *parentp
);
2549 * Helper function for deleting chains.
2551 * The chain is removed from the live view (the RBTREE) as well as the parent's
2552 * blockmap. Both chain and its parent must be locked.
2554 * parent may not be errored. chain can be errored.
2557 _hammer2_chain_delete_helper(hammer2_trans_t
*trans
,
2558 hammer2_chain_t
*parent
, hammer2_chain_t
*chain
,
2563 KKASSERT((chain
->flags
& (HAMMER2_CHAIN_DELETED
|
2564 HAMMER2_CHAIN_FICTITIOUS
)) == 0);
2567 if (chain
->flags
& HAMMER2_CHAIN_BMAPPED
) {
2569 * Chain is blockmapped, so there must be a parent.
2570 * Atomically remove the chain from the parent and remove
2571 * the blockmap entry.
2573 hammer2_blockref_t
*base
;
2576 KKASSERT(parent
!= NULL
);
2577 KKASSERT(parent
->error
== 0);
2578 KKASSERT((parent
->flags
& HAMMER2_CHAIN_INITIAL
) == 0);
2579 hammer2_chain_modify(trans
, parent
,
2580 HAMMER2_MODIFY_OPTDATA
);
2583 * Calculate blockmap pointer
2585 KKASSERT(chain
->flags
& HAMMER2_CHAIN_ONRBTREE
);
2586 hammer2_spin_ex(&parent
->core
.spin
);
2588 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_DELETED
);
2589 atomic_add_int(&parent
->core
.live_count
, -1);
2590 ++parent
->core
.generation
;
2591 RB_REMOVE(hammer2_chain_tree
, &parent
->core
.rbtree
, chain
);
2592 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_ONRBTREE
);
2593 --parent
->core
.chain_count
;
2594 chain
->parent
= NULL
;
2596 switch(parent
->bref
.type
) {
2597 case HAMMER2_BREF_TYPE_INODE
:
2599 * Access the inode's block array. However, there
2600 * is no block array if the inode is flagged
2601 * DIRECTDATA. The DIRECTDATA case typicaly only
2602 * occurs when a hardlink has been shifted up the
2603 * tree and the original inode gets replaced with
2604 * an OBJTYPE_HARDLINK placeholding inode.
2607 (parent
->data
->ipdata
.meta
.op_flags
&
2608 HAMMER2_OPFLAG_DIRECTDATA
) == 0) {
2610 &parent
->data
->ipdata
.u
.blockset
.blockref
[0];
2614 count
= HAMMER2_SET_COUNT
;
2616 case HAMMER2_BREF_TYPE_INDIRECT
:
2617 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
2619 base
= &parent
->data
->npdata
[0];
2622 count
= parent
->bytes
/ sizeof(hammer2_blockref_t
);
2624 case HAMMER2_BREF_TYPE_VOLUME
:
2625 base
= &hmp
->voldata
.sroot_blockset
.blockref
[0];
2626 count
= HAMMER2_SET_COUNT
;
2628 case HAMMER2_BREF_TYPE_FREEMAP
:
2629 base
= &parent
->data
->npdata
[0];
2630 count
= HAMMER2_SET_COUNT
;
2635 panic("hammer2_flush_pass2: "
2636 "unrecognized blockref type: %d",
2641 * delete blockmapped chain from its parent.
2643 * The parent is not affected by any statistics in chain
2644 * which are pending synchronization. That is, there is
2645 * nothing to undo in the parent since they have not yet
2646 * been incorporated into the parent.
2648 * The parent is affected by statistics stored in inodes.
2649 * Those have already been synchronized, so they must be
2650 * undone. XXX split update possible w/delete in middle?
2653 int cache_index
= -1;
2654 hammer2_base_delete(trans
, parent
, base
, count
,
2655 &cache_index
, chain
);
2657 hammer2_spin_unex(&parent
->core
.spin
);
2658 } else if (chain
->flags
& HAMMER2_CHAIN_ONRBTREE
) {
2660 * Chain is not blockmapped but a parent is present.
2661 * Atomically remove the chain from the parent. There is
2662 * no blockmap entry to remove.
2664 * Because chain was associated with a parent but not
2665 * synchronized, the chain's *_count_up fields contain
2666 * inode adjustment statistics which must be undone.
2668 hammer2_spin_ex(&parent
->core
.spin
);
2669 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_DELETED
);
2670 atomic_add_int(&parent
->core
.live_count
, -1);
2671 ++parent
->core
.generation
;
2672 RB_REMOVE(hammer2_chain_tree
, &parent
->core
.rbtree
, chain
);
2673 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_ONRBTREE
);
2674 --parent
->core
.chain_count
;
2675 chain
->parent
= NULL
;
2676 hammer2_spin_unex(&parent
->core
.spin
);
2679 * Chain is not blockmapped and has no parent. This
2680 * is a degenerate case.
2682 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_DELETED
);
2687 * Create an indirect block that covers one or more of the elements in the
2688 * current parent. Either returns the existing parent with no locking or
2689 * ref changes or returns the new indirect block locked and referenced
2690 * and leaving the original parent lock/ref intact as well.
2692 * If an error occurs, NULL is returned and *errorp is set to the error.
2694 * The returned chain depends on where the specified key falls.
2696 * The key/keybits for the indirect mode only needs to follow three rules:
2698 * (1) That all elements underneath it fit within its key space and
2700 * (2) That all elements outside it are outside its key space.
2702 * (3) When creating the new indirect block any elements in the current
2703 * parent that fit within the new indirect block's keyspace must be
2704 * moved into the new indirect block.
2706 * (4) The keyspace chosen for the inserted indirect block CAN cover a wider
2707 * keyspace the the current parent, but lookup/iteration rules will
2708 * ensure (and must ensure) that rule (2) for all parents leading up
2709 * to the nearest inode or the root volume header is adhered to. This
2710 * is accomplished by always recursing through matching keyspaces in
2711 * the hammer2_chain_lookup() and hammer2_chain_next() API.
2713 * The current implementation calculates the current worst-case keyspace by
2714 * iterating the current parent and then divides it into two halves, choosing
2715 * whichever half has the most elements (not necessarily the half containing
2716 * the requested key).
2718 * We can also opt to use the half with the least number of elements. This
2719 * causes lower-numbered keys (aka logical file offsets) to recurse through
2720 * fewer indirect blocks and higher-numbered keys to recurse through more.
2721 * This also has the risk of not moving enough elements to the new indirect
2722 * block and being forced to create several indirect blocks before the element
2725 * Must be called with an exclusively locked parent.
2727 static int hammer2_chain_indkey_freemap(hammer2_chain_t
*parent
,
2728 hammer2_key_t
*keyp
, int keybits
,
2729 hammer2_blockref_t
*base
, int count
);
2730 static int hammer2_chain_indkey_normal(hammer2_chain_t
*parent
,
2731 hammer2_key_t
*keyp
, int keybits
,
2732 hammer2_blockref_t
*base
, int count
);
2735 hammer2_chain_create_indirect(hammer2_trans_t
*trans
, hammer2_chain_t
*parent
,
2736 hammer2_key_t create_key
, int create_bits
,
2737 int for_type
, int *errorp
)
2740 hammer2_blockref_t
*base
;
2741 hammer2_blockref_t
*bref
;
2742 hammer2_blockref_t bcopy
;
2743 hammer2_chain_t
*chain
;
2744 hammer2_chain_t
*ichain
;
2745 hammer2_chain_t dummy
;
2746 hammer2_key_t key
= create_key
;
2747 hammer2_key_t key_beg
;
2748 hammer2_key_t key_end
;
2749 hammer2_key_t key_next
;
2750 int keybits
= create_bits
;
2757 int maxloops
= 300000;
2760 * Calculate the base blockref pointer or NULL if the chain
2761 * is known to be empty. We need to calculate the array count
2762 * for RB lookups either way.
2766 KKASSERT(hammer2_mtx_owned(&parent
->lock
));
2768 /*hammer2_chain_modify(trans, &parent, HAMMER2_MODIFY_OPTDATA);*/
2769 if (parent
->flags
& HAMMER2_CHAIN_INITIAL
) {
2772 switch(parent
->bref
.type
) {
2773 case HAMMER2_BREF_TYPE_INODE
:
2774 count
= HAMMER2_SET_COUNT
;
2776 case HAMMER2_BREF_TYPE_INDIRECT
:
2777 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
2778 count
= parent
->bytes
/ sizeof(hammer2_blockref_t
);
2780 case HAMMER2_BREF_TYPE_VOLUME
:
2781 count
= HAMMER2_SET_COUNT
;
2783 case HAMMER2_BREF_TYPE_FREEMAP
:
2784 count
= HAMMER2_SET_COUNT
;
2787 panic("hammer2_chain_create_indirect: "
2788 "unrecognized blockref type: %d",
2794 switch(parent
->bref
.type
) {
2795 case HAMMER2_BREF_TYPE_INODE
:
2796 base
= &parent
->data
->ipdata
.u
.blockset
.blockref
[0];
2797 count
= HAMMER2_SET_COUNT
;
2799 case HAMMER2_BREF_TYPE_INDIRECT
:
2800 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
2801 base
= &parent
->data
->npdata
[0];
2802 count
= parent
->bytes
/ sizeof(hammer2_blockref_t
);
2804 case HAMMER2_BREF_TYPE_VOLUME
:
2805 base
= &hmp
->voldata
.sroot_blockset
.blockref
[0];
2806 count
= HAMMER2_SET_COUNT
;
2808 case HAMMER2_BREF_TYPE_FREEMAP
:
2809 base
= &hmp
->voldata
.freemap_blockset
.blockref
[0];
2810 count
= HAMMER2_SET_COUNT
;
2813 panic("hammer2_chain_create_indirect: "
2814 "unrecognized blockref type: %d",
2822 * dummy used in later chain allocation (no longer used for lookups).
2824 bzero(&dummy
, sizeof(dummy
));
2827 * When creating an indirect block for a freemap node or leaf
2828 * the key/keybits must be fitted to static radix levels because
2829 * particular radix levels use particular reserved blocks in the
2832 * This routine calculates the key/radix of the indirect block
2833 * we need to create, and whether it is on the high-side or the
2836 if (for_type
== HAMMER2_BREF_TYPE_FREEMAP_NODE
||
2837 for_type
== HAMMER2_BREF_TYPE_FREEMAP_LEAF
) {
2838 keybits
= hammer2_chain_indkey_freemap(parent
, &key
, keybits
,
2841 keybits
= hammer2_chain_indkey_normal(parent
, &key
, keybits
,
2846 * Normalize the key for the radix being represented, keeping the
2847 * high bits and throwing away the low bits.
2849 key
&= ~(((hammer2_key_t
)1 << keybits
) - 1);
2852 * How big should our new indirect block be? It has to be at least
2853 * as large as its parent.
2855 * The freemap uses a specific indirect block size.
2857 * The first indirect block level down from an inode typically
2858 * uses LBUFSIZE (16384), else it uses PBUFSIZE (65536).
2860 if (for_type
== HAMMER2_BREF_TYPE_FREEMAP_NODE
||
2861 for_type
== HAMMER2_BREF_TYPE_FREEMAP_LEAF
) {
2862 nbytes
= HAMMER2_FREEMAP_LEVELN_PSIZE
;
2863 } else if (parent
->bref
.type
== HAMMER2_BREF_TYPE_INODE
) {
2864 nbytes
= HAMMER2_IND_BYTES_MIN
;
2866 nbytes
= HAMMER2_IND_BYTES_MAX
;
2868 if (nbytes
< count
* sizeof(hammer2_blockref_t
)) {
2869 KKASSERT(for_type
!= HAMMER2_BREF_TYPE_FREEMAP_NODE
&&
2870 for_type
!= HAMMER2_BREF_TYPE_FREEMAP_LEAF
);
2871 nbytes
= count
* sizeof(hammer2_blockref_t
);
2875 * Ok, create our new indirect block
2877 if (for_type
== HAMMER2_BREF_TYPE_FREEMAP_NODE
||
2878 for_type
== HAMMER2_BREF_TYPE_FREEMAP_LEAF
) {
2879 dummy
.bref
.type
= HAMMER2_BREF_TYPE_FREEMAP_NODE
;
2881 dummy
.bref
.type
= HAMMER2_BREF_TYPE_INDIRECT
;
2883 dummy
.bref
.key
= key
;
2884 dummy
.bref
.keybits
= keybits
;
2885 dummy
.bref
.data_off
= hammer2_getradix(nbytes
);
2886 dummy
.bref
.methods
= parent
->bref
.methods
;
2888 ichain
= hammer2_chain_alloc(hmp
, parent
->pmp
, trans
, &dummy
.bref
);
2889 atomic_set_int(&ichain
->flags
, HAMMER2_CHAIN_INITIAL
);
2890 hammer2_chain_lock(ichain
, HAMMER2_RESOLVE_MAYBE
);
2891 /* ichain has one ref at this point */
2894 * We have to mark it modified to allocate its block, but use
2895 * OPTDATA to allow it to remain in the INITIAL state. Otherwise
2896 * it won't be acted upon by the flush code.
2898 hammer2_chain_modify(trans
, ichain
, HAMMER2_MODIFY_OPTDATA
);
2901 * Iterate the original parent and move the matching brefs into
2902 * the new indirect block.
2904 * XXX handle flushes.
2907 key_end
= HAMMER2_KEY_MAX
;
2909 hammer2_spin_ex(&parent
->core
.spin
);
2914 if (++loops
> 100000) {
2915 hammer2_spin_unex(&parent
->core
.spin
);
2916 panic("excessive loops r=%d p=%p base/count %p:%d %016jx\n",
2917 reason
, parent
, base
, count
, key_next
);
2921 * NOTE: spinlock stays intact, returned chain (if not NULL)
2922 * is not referenced or locked which means that we
2923 * cannot safely check its flagged / deletion status
2926 chain
= hammer2_combined_find(parent
, base
, count
,
2927 &cache_index
, &key_next
,
2930 generation
= parent
->core
.generation
;
2933 key_next
= bref
->key
+ ((hammer2_key_t
)1 << bref
->keybits
);
2936 * Skip keys that are not within the key/radix of the new
2937 * indirect block. They stay in the parent.
2939 if ((~(((hammer2_key_t
)1 << keybits
) - 1) &
2940 (key
^ bref
->key
)) != 0) {
2941 goto next_key_spinlocked
;
2945 * Load the new indirect block by acquiring the related
2946 * chains (potentially from media as it might not be
2947 * in-memory). Then move it to the new parent (ichain)
2948 * via DELETE-DUPLICATE.
2950 * chain is referenced but not locked. We must lock the
2951 * chain to obtain definitive DUPLICATED/DELETED state
2955 * Use chain already present in the RBTREE
2957 hammer2_chain_ref(chain
);
2958 hammer2_spin_unex(&parent
->core
.spin
);
2959 hammer2_chain_lock(chain
, HAMMER2_RESOLVE_NEVER
);
2962 * Get chain for blockref element. _get returns NULL
2963 * on insertion race.
2966 hammer2_spin_unex(&parent
->core
.spin
);
2967 chain
= hammer2_chain_get(parent
, generation
, &bcopy
);
2968 if (chain
== NULL
) {
2970 hammer2_spin_ex(&parent
->core
.spin
);
2973 if (bcmp(&bcopy
, bref
, sizeof(bcopy
))) {
2974 kprintf("REASON 2\n");
2976 hammer2_chain_drop(chain
);
2977 hammer2_spin_ex(&parent
->core
.spin
);
2980 hammer2_chain_lock(chain
, HAMMER2_RESOLVE_NEVER
);
2984 * This is always live so if the chain has been deleted
2985 * we raced someone and we have to retry.
2987 * NOTE: Lookups can race delete-duplicate because
2988 * delete-duplicate does not lock the parent's core
2989 * (they just use the spinlock on the core). We must
2990 * check for races by comparing the DUPLICATED flag before
2991 * releasing the spinlock with the flag after locking the
2994 * (note reversed logic for this one)
2996 if (chain
->flags
& HAMMER2_CHAIN_DELETED
) {
2997 hammer2_chain_unlock(chain
);
2998 hammer2_chain_drop(chain
);
3003 * Shift the chain to the indirect block.
3005 * WARNING! No reason for us to load chain data, pass NOSTATS
3006 * to prevent delete/insert from trying to access
3007 * inode stats (and thus asserting if there is no
3008 * chain->data loaded).
3010 hammer2_chain_delete(trans
, parent
, chain
,
3011 HAMMER2_DELETE_NOSTATS
);
3012 hammer2_chain_rename(trans
, NULL
, &ichain
, chain
,
3013 HAMMER2_INSERT_NOSTATS
);
3014 hammer2_chain_unlock(chain
);
3015 hammer2_chain_drop(chain
);
3016 KKASSERT(parent
->refs
> 0);
3019 hammer2_spin_ex(&parent
->core
.spin
);
3020 next_key_spinlocked
:
3021 if (--maxloops
== 0)
3022 panic("hammer2_chain_create_indirect: maxloops");
3024 if (key_next
== 0 || key_next
> key_end
)
3029 hammer2_spin_unex(&parent
->core
.spin
);
3032 * Insert the new indirect block into the parent now that we've
3033 * cleared out some entries in the parent. We calculated a good
3034 * insertion index in the loop above (ichain->index).
3036 * We don't have to set UPDATE here because we mark ichain
3037 * modified down below (so the normal modified -> flush -> set-moved
3038 * sequence applies).
3040 * The insertion shouldn't race as this is a completely new block
3041 * and the parent is locked.
3043 KKASSERT((ichain
->flags
& HAMMER2_CHAIN_ONRBTREE
) == 0);
3044 hammer2_chain_insert(parent
, ichain
,
3045 HAMMER2_CHAIN_INSERT_SPIN
|
3046 HAMMER2_CHAIN_INSERT_LIVE
,
3050 * Make sure flushes propogate after our manual insertion.
3052 hammer2_chain_setflush(trans
, ichain
);
3053 hammer2_chain_setflush(trans
, parent
);
3056 * Figure out what to return.
3058 if (~(((hammer2_key_t
)1 << keybits
) - 1) &
3059 (create_key
^ key
)) {
3061 * Key being created is outside the key range,
3062 * return the original parent.
3064 hammer2_chain_unlock(ichain
);
3065 hammer2_chain_drop(ichain
);
3068 * Otherwise its in the range, return the new parent.
3069 * (leave both the new and old parent locked).
3078 * Calculate the keybits and highside/lowside of the freemap node the
3079 * caller is creating.
3081 * This routine will specify the next higher-level freemap key/radix
3082 * representing the lowest-ordered set. By doing so, eventually all
3083 * low-ordered sets will be moved one level down.
3085 * We have to be careful here because the freemap reserves a limited
3086 * number of blocks for a limited number of levels. So we can't just
3087 * push indiscriminately.
3090 hammer2_chain_indkey_freemap(hammer2_chain_t
*parent
, hammer2_key_t
*keyp
,
3091 int keybits
, hammer2_blockref_t
*base
, int count
)
3093 hammer2_chain_t
*chain
;
3094 hammer2_blockref_t
*bref
;
3096 hammer2_key_t key_beg
;
3097 hammer2_key_t key_end
;
3098 hammer2_key_t key_next
;
3102 int maxloops
= 300000;
3110 * Calculate the range of keys in the array being careful to skip
3111 * slots which are overridden with a deletion.
3114 key_end
= HAMMER2_KEY_MAX
;
3116 hammer2_spin_ex(&parent
->core
.spin
);
3119 if (--maxloops
== 0) {
3120 panic("indkey_freemap shit %p %p:%d\n",
3121 parent
, base
, count
);
3123 chain
= hammer2_combined_find(parent
, base
, count
,
3124 &cache_index
, &key_next
,
3135 * Skip deleted chains.
3137 if (chain
&& (chain
->flags
& HAMMER2_CHAIN_DELETED
)) {
3138 if (key_next
== 0 || key_next
> key_end
)
3145 * Use the full live (not deleted) element for the scan
3146 * iteration. HAMMER2 does not allow partial replacements.
3148 * XXX should be built into hammer2_combined_find().
3150 key_next
= bref
->key
+ ((hammer2_key_t
)1 << bref
->keybits
);
3152 if (keybits
> bref
->keybits
) {
3154 keybits
= bref
->keybits
;
3155 } else if (keybits
== bref
->keybits
&& bref
->key
< key
) {
3162 hammer2_spin_unex(&parent
->core
.spin
);
3165 * Return the keybits for a higher-level FREEMAP_NODE covering
3169 case HAMMER2_FREEMAP_LEVEL0_RADIX
:
3170 keybits
= HAMMER2_FREEMAP_LEVEL1_RADIX
;
3172 case HAMMER2_FREEMAP_LEVEL1_RADIX
:
3173 keybits
= HAMMER2_FREEMAP_LEVEL2_RADIX
;
3175 case HAMMER2_FREEMAP_LEVEL2_RADIX
:
3176 keybits
= HAMMER2_FREEMAP_LEVEL3_RADIX
;
3178 case HAMMER2_FREEMAP_LEVEL3_RADIX
:
3179 keybits
= HAMMER2_FREEMAP_LEVEL4_RADIX
;
3181 case HAMMER2_FREEMAP_LEVEL4_RADIX
:
3182 keybits
= HAMMER2_FREEMAP_LEVEL5_RADIX
;
3184 case HAMMER2_FREEMAP_LEVEL5_RADIX
:
3185 panic("hammer2_chain_indkey_freemap: level too high");
3188 panic("hammer2_chain_indkey_freemap: bad radix");
3197 * Calculate the keybits and highside/lowside of the indirect block the
3198 * caller is creating.
3201 hammer2_chain_indkey_normal(hammer2_chain_t
*parent
, hammer2_key_t
*keyp
,
3202 int keybits
, hammer2_blockref_t
*base
, int count
)
3204 hammer2_blockref_t
*bref
;
3205 hammer2_chain_t
*chain
;
3206 hammer2_key_t key_beg
;
3207 hammer2_key_t key_end
;
3208 hammer2_key_t key_next
;
3214 int maxloops
= 300000;
3221 * Calculate the range of keys in the array being careful to skip
3222 * slots which are overridden with a deletion. Once the scan
3223 * completes we will cut the key range in half and shift half the
3224 * range into the new indirect block.
3227 key_end
= HAMMER2_KEY_MAX
;
3229 hammer2_spin_ex(&parent
->core
.spin
);
3232 if (--maxloops
== 0) {
3233 panic("indkey_freemap shit %p %p:%d\n",
3234 parent
, base
, count
);
3236 chain
= hammer2_combined_find(parent
, base
, count
,
3237 &cache_index
, &key_next
,
3248 * NOTE: No need to check DUPLICATED here because we do
3249 * not release the spinlock.
3251 if (chain
&& (chain
->flags
& HAMMER2_CHAIN_DELETED
)) {
3252 if (key_next
== 0 || key_next
> key_end
)
3259 * Use the full live (not deleted) element for the scan
3260 * iteration. HAMMER2 does not allow partial replacements.
3262 * XXX should be built into hammer2_combined_find().
3264 key_next
= bref
->key
+ ((hammer2_key_t
)1 << bref
->keybits
);
3267 * Expand our calculated key range (key, keybits) to fit
3268 * the scanned key. nkeybits represents the full range
3269 * that we will later cut in half (two halves @ nkeybits - 1).
3272 if (nkeybits
< bref
->keybits
) {
3273 if (bref
->keybits
> 64) {
3274 kprintf("bad bref chain %p bref %p\n",
3278 nkeybits
= bref
->keybits
;
3280 while (nkeybits
< 64 &&
3281 (~(((hammer2_key_t
)1 << nkeybits
) - 1) &
3282 (key
^ bref
->key
)) != 0) {
3287 * If the new key range is larger we have to determine
3288 * which side of the new key range the existing keys fall
3289 * under by checking the high bit, then collapsing the
3290 * locount into the hicount or vise-versa.
3292 if (keybits
!= nkeybits
) {
3293 if (((hammer2_key_t
)1 << (nkeybits
- 1)) & key
) {
3304 * The newly scanned key will be in the lower half or the
3305 * upper half of the (new) key range.
3307 if (((hammer2_key_t
)1 << (nkeybits
- 1)) & bref
->key
)
3316 hammer2_spin_unex(&parent
->core
.spin
);
3317 bref
= NULL
; /* now invalid (safety) */
3320 * Adjust keybits to represent half of the full range calculated
3321 * above (radix 63 max)
3326 * Select whichever half contains the most elements. Theoretically
3327 * we can select either side as long as it contains at least one
3328 * element (in order to ensure that a free slot is present to hold
3329 * the indirect block).
3331 if (hammer2_indirect_optimize
) {
3333 * Insert node for least number of keys, this will arrange
3334 * the first few blocks of a large file or the first few
3335 * inodes in a directory with fewer indirect blocks when
3338 if (hicount
< locount
&& hicount
!= 0)
3339 key
|= (hammer2_key_t
)1 << keybits
;
3341 key
&= ~(hammer2_key_t
)1 << keybits
;
3344 * Insert node for most number of keys, best for heavily
3347 if (hicount
> locount
)
3348 key
|= (hammer2_key_t
)1 << keybits
;
3350 key
&= ~(hammer2_key_t
)1 << keybits
;
3358 * Sets CHAIN_DELETED and remove the chain's blockref from the parent if
3361 * Both parent and chain must be locked exclusively.
3363 * This function will modify the parent if the blockref requires removal
3364 * from the parent's block table.
3366 * This function is NOT recursive. Any entity already pushed into the
3367 * chain (such as an inode) may still need visibility into its contents,
3368 * as well as the ability to read and modify the contents. For example,
3369 * for an unlinked file which is still open.
3372 hammer2_chain_delete(hammer2_trans_t
*trans
, hammer2_chain_t
*parent
,
3373 hammer2_chain_t
*chain
, int flags
)
3375 KKASSERT(hammer2_mtx_owned(&chain
->lock
));
3378 * Nothing to do if already marked.
3380 * We need the spinlock on the core whos RBTREE contains chain
3381 * to protect against races.
3383 if ((chain
->flags
& HAMMER2_CHAIN_DELETED
) == 0) {
3384 KKASSERT((chain
->flags
& HAMMER2_CHAIN_DELETED
) == 0 &&
3385 chain
->parent
== parent
);
3386 _hammer2_chain_delete_helper(trans
, parent
, chain
, flags
);
3390 * To avoid losing track of a permanent deletion we add the chain
3391 * to the delayed flush queue. If were to flush it right now the
3392 * parent would end up in a modified state and generate I/O.
3393 * The delayed queue gives the parent a chance to be deleted to
3396 if (flags
& HAMMER2_DELETE_PERMANENT
) {
3397 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_DESTROY
);
3398 hammer2_delayed_flush(trans
, chain
);
3400 /* XXX might not be needed */
3401 hammer2_chain_setflush(trans
, chain
);
3406 * Returns the index of the nearest element in the blockref array >= elm.
3407 * Returns (count) if no element could be found.
3409 * Sets *key_nextp to the next key for loop purposes but does not modify
3410 * it if the next key would be higher than the current value of *key_nextp.
3411 * Note that *key_nexp can overflow to 0, which should be tested by the
3414 * (*cache_indexp) is a heuristic and can be any value without effecting
3417 * WARNING! Must be called with parent's spinlock held. Spinlock remains
3418 * held through the operation.
3421 hammer2_base_find(hammer2_chain_t
*parent
,
3422 hammer2_blockref_t
*base
, int count
,
3423 int *cache_indexp
, hammer2_key_t
*key_nextp
,
3424 hammer2_key_t key_beg
, hammer2_key_t key_end
)
3426 hammer2_blockref_t
*scan
;
3427 hammer2_key_t scan_end
;
3432 * Require the live chain's already have their core's counted
3433 * so we can optimize operations.
3435 KKASSERT(parent
->flags
& HAMMER2_CHAIN_COUNTEDBREFS
);
3440 if (count
== 0 || base
== NULL
)
3444 * Sequential optimization using *cache_indexp. This is the most
3447 * We can avoid trailing empty entries on live chains, otherwise
3448 * we might have to check the whole block array.
3452 limit
= parent
->core
.live_zero
;
3457 KKASSERT(i
< count
);
3463 while (i
> 0 && (scan
->type
== 0 || scan
->key
> key_beg
)) {
3470 * Search forwards, stop when we find a scan element which
3471 * encloses the key or until we know that there are no further
3475 if (scan
->type
!= 0) {
3476 scan_end
= scan
->key
+
3477 ((hammer2_key_t
)1 << scan
->keybits
) - 1;
3478 if (scan
->key
> key_beg
|| scan_end
>= key_beg
)
3491 scan_end
= scan
->key
+
3492 ((hammer2_key_t
)1 << scan
->keybits
);
3493 if (scan_end
&& (*key_nextp
> scan_end
||
3495 *key_nextp
= scan_end
;
3503 * Do a combined search and return the next match either from the blockref
3504 * array or from the in-memory chain. Sets *bresp to the returned bref in
3505 * both cases, or sets it to NULL if the search exhausted. Only returns
3506 * a non-NULL chain if the search matched from the in-memory chain.
3508 * When no in-memory chain has been found and a non-NULL bref is returned
3512 * The returned chain is not locked or referenced. Use the returned bref
3513 * to determine if the search exhausted or not. Iterate if the base find
3514 * is chosen but matches a deleted chain.
3516 * WARNING! Must be called with parent's spinlock held. Spinlock remains
3517 * held through the operation.
3519 static hammer2_chain_t
*
3520 hammer2_combined_find(hammer2_chain_t
*parent
,
3521 hammer2_blockref_t
*base
, int count
,
3522 int *cache_indexp
, hammer2_key_t
*key_nextp
,
3523 hammer2_key_t key_beg
, hammer2_key_t key_end
,
3524 hammer2_blockref_t
**bresp
)
3526 hammer2_blockref_t
*bref
;
3527 hammer2_chain_t
*chain
;
3531 * Lookup in block array and in rbtree.
3533 *key_nextp
= key_end
+ 1;
3534 i
= hammer2_base_find(parent
, base
, count
, cache_indexp
,
3535 key_nextp
, key_beg
, key_end
);
3536 chain
= hammer2_chain_find(parent
, key_nextp
, key_beg
, key_end
);
3541 if (i
== count
&& chain
== NULL
) {
3547 * Only chain matched.
3550 bref
= &chain
->bref
;
3555 * Only blockref matched.
3557 if (chain
== NULL
) {
3563 * Both in-memory and blockref matched, select the nearer element.
3565 * If both are flush with the left-hand side or both are the
3566 * same distance away, select the chain. In this situation the
3567 * chain must have been loaded from the matching blockmap.
3569 if ((chain
->bref
.key
<= key_beg
&& base
[i
].key
<= key_beg
) ||
3570 chain
->bref
.key
== base
[i
].key
) {
3571 KKASSERT(chain
->bref
.key
== base
[i
].key
);
3572 bref
= &chain
->bref
;
3577 * Select the nearer key
3579 if (chain
->bref
.key
< base
[i
].key
) {
3580 bref
= &chain
->bref
;
3587 * If the bref is out of bounds we've exhausted our search.
3590 if (bref
->key
> key_end
) {
3600 * Locate the specified block array element and delete it. The element
3603 * The spin lock on the related chain must be held.
3605 * NOTE: live_count was adjusted when the chain was deleted, so it does not
3606 * need to be adjusted when we commit the media change.
3609 hammer2_base_delete(hammer2_trans_t
*trans
, hammer2_chain_t
*parent
,
3610 hammer2_blockref_t
*base
, int count
,
3611 int *cache_indexp
, hammer2_chain_t
*chain
)
3613 hammer2_blockref_t
*elm
= &chain
->bref
;
3614 hammer2_key_t key_next
;
3618 * Delete element. Expect the element to exist.
3620 * XXX see caller, flush code not yet sophisticated enough to prevent
3621 * re-flushed in some cases.
3623 key_next
= 0; /* max range */
3624 i
= hammer2_base_find(parent
, base
, count
, cache_indexp
,
3625 &key_next
, elm
->key
, elm
->key
);
3626 if (i
== count
|| base
[i
].type
== 0 ||
3627 base
[i
].key
!= elm
->key
||
3628 ((chain
->flags
& HAMMER2_CHAIN_BMAPUPD
) == 0 &&
3629 base
[i
].keybits
!= elm
->keybits
)) {
3630 hammer2_spin_unex(&parent
->core
.spin
);
3631 panic("delete base %p element not found at %d/%d elm %p\n",
3632 base
, i
, count
, elm
);
3637 * Update stats and zero the entry
3639 parent
->bref
.data_count
-= base
[i
].data_count
;
3640 parent
->bref
.data_count
-= (hammer2_off_t
)1 <<
3641 (int)(base
[i
].data_off
& HAMMER2_OFF_MASK_RADIX
);
3642 parent
->bref
.inode_count
-= base
[i
].inode_count
;
3643 if (base
[i
].type
== HAMMER2_BREF_TYPE_INODE
)
3644 parent
->bref
.inode_count
-= 1;
3646 bzero(&base
[i
], sizeof(*base
));
3649 * We can only optimize parent->core.live_zero for live chains.
3651 if (parent
->core
.live_zero
== i
+ 1) {
3652 while (--i
>= 0 && base
[i
].type
== 0)
3654 parent
->core
.live_zero
= i
+ 1;
3658 * Clear appropriate blockmap flags in chain.
3660 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_BMAPPED
|
3661 HAMMER2_CHAIN_BMAPUPD
);
3665 * Insert the specified element. The block array must not already have the
3666 * element and must have space available for the insertion.
3668 * The spin lock on the related chain must be held.
3670 * NOTE: live_count was adjusted when the chain was deleted, so it does not
3671 * need to be adjusted when we commit the media change.
3674 hammer2_base_insert(hammer2_trans_t
*trans __unused
, hammer2_chain_t
*parent
,
3675 hammer2_blockref_t
*base
, int count
,
3676 int *cache_indexp
, hammer2_chain_t
*chain
)
3678 hammer2_blockref_t
*elm
= &chain
->bref
;
3679 hammer2_key_t key_next
;
3688 * Insert new element. Expect the element to not already exist
3689 * unless we are replacing it.
3691 * XXX see caller, flush code not yet sophisticated enough to prevent
3692 * re-flushed in some cases.
3694 key_next
= 0; /* max range */
3695 i
= hammer2_base_find(parent
, base
, count
, cache_indexp
,
3696 &key_next
, elm
->key
, elm
->key
);
3699 * Shortcut fill optimization, typical ordered insertion(s) may not
3702 KKASSERT(i
>= 0 && i
<= count
);
3705 * Set appropriate blockmap flags in chain.
3707 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_BMAPPED
);
3710 * Update stats and zero the entry
3712 parent
->bref
.data_count
+= elm
->data_count
;
3713 parent
->bref
.data_count
+= (hammer2_off_t
)1 <<
3714 (int)(elm
->data_off
& HAMMER2_OFF_MASK_RADIX
);
3715 parent
->bref
.inode_count
+= elm
->inode_count
;
3716 if (elm
->type
== HAMMER2_BREF_TYPE_INODE
)
3717 parent
->bref
.inode_count
+= 1;
3721 * We can only optimize parent->core.live_zero for live chains.
3723 if (i
== count
&& parent
->core
.live_zero
< count
) {
3724 i
= parent
->core
.live_zero
++;
3729 xkey
= elm
->key
+ ((hammer2_key_t
)1 << elm
->keybits
) - 1;
3730 if (i
!= count
&& (base
[i
].key
< elm
->key
|| xkey
>= base
[i
].key
)) {
3731 hammer2_spin_unex(&parent
->core
.spin
);
3732 panic("insert base %p overlapping elements at %d elm %p\n",
3737 * Try to find an empty slot before or after.
3741 while (j
> 0 || k
< count
) {
3743 if (j
>= 0 && base
[j
].type
== 0) {
3747 bcopy(&base
[j
+1], &base
[j
],
3748 (i
- j
- 1) * sizeof(*base
));
3754 if (k
< count
&& base
[k
].type
== 0) {
3755 bcopy(&base
[i
], &base
[i
+1],
3756 (k
- i
) * sizeof(hammer2_blockref_t
));
3760 * We can only update parent->core.live_zero for live
3763 if (parent
->core
.live_zero
<= k
)
3764 parent
->core
.live_zero
= k
+ 1;
3769 panic("hammer2_base_insert: no room!");
3776 for (l
= 0; l
< count
; ++l
) {
3778 key_next
= base
[l
].key
+
3779 ((hammer2_key_t
)1 << base
[l
].keybits
) - 1;
3783 while (++l
< count
) {
3785 if (base
[l
].key
<= key_next
)
3786 panic("base_insert %d %d,%d,%d fail %p:%d", u
, i
, j
, k
, base
, l
);
3787 key_next
= base
[l
].key
+
3788 ((hammer2_key_t
)1 << base
[l
].keybits
) - 1;
3798 * Sort the blockref array for the chain. Used by the flush code to
3799 * sort the blockref[] array.
3801 * The chain must be exclusively locked AND spin-locked.
3803 typedef hammer2_blockref_t
*hammer2_blockref_p
;
3807 hammer2_base_sort_callback(const void *v1
, const void *v2
)
3809 hammer2_blockref_p bref1
= *(const hammer2_blockref_p
*)v1
;
3810 hammer2_blockref_p bref2
= *(const hammer2_blockref_p
*)v2
;
3813 * Make sure empty elements are placed at the end of the array
3815 if (bref1
->type
== 0) {
3816 if (bref2
->type
== 0)
3819 } else if (bref2
->type
== 0) {
3826 if (bref1
->key
< bref2
->key
)
3828 if (bref1
->key
> bref2
->key
)
3834 hammer2_base_sort(hammer2_chain_t
*chain
)
3836 hammer2_blockref_t
*base
;
3839 switch(chain
->bref
.type
) {
3840 case HAMMER2_BREF_TYPE_INODE
:
3842 * Special shortcut for embedded data returns the inode
3843 * itself. Callers must detect this condition and access
3844 * the embedded data (the strategy code does this for us).
3846 * This is only applicable to regular files and softlinks.
3848 if (chain
->data
->ipdata
.meta
.op_flags
&
3849 HAMMER2_OPFLAG_DIRECTDATA
) {
3852 base
= &chain
->data
->ipdata
.u
.blockset
.blockref
[0];
3853 count
= HAMMER2_SET_COUNT
;
3855 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
3856 case HAMMER2_BREF_TYPE_INDIRECT
:
3858 * Optimize indirect blocks in the INITIAL state to avoid
3861 KKASSERT((chain
->flags
& HAMMER2_CHAIN_INITIAL
) == 0);
3862 base
= &chain
->data
->npdata
[0];
3863 count
= chain
->bytes
/ sizeof(hammer2_blockref_t
);
3865 case HAMMER2_BREF_TYPE_VOLUME
:
3866 base
= &chain
->hmp
->voldata
.sroot_blockset
.blockref
[0];
3867 count
= HAMMER2_SET_COUNT
;
3869 case HAMMER2_BREF_TYPE_FREEMAP
:
3870 base
= &chain
->hmp
->voldata
.freemap_blockset
.blockref
[0];
3871 count
= HAMMER2_SET_COUNT
;
3874 kprintf("hammer2_chain_lookup: unrecognized "
3875 "blockref(A) type: %d",
3878 tsleep(&base
, 0, "dead", 0);
3879 panic("hammer2_chain_lookup: unrecognized "
3880 "blockref(A) type: %d",
3882 base
= NULL
; /* safety */
3883 count
= 0; /* safety */
3885 kqsort(base
, count
, sizeof(*base
), hammer2_base_sort_callback
);
3891 * Chain memory management
3894 hammer2_chain_wait(hammer2_chain_t
*chain
)
3896 tsleep(chain
, 0, "chnflw", 1);
3899 const hammer2_media_data_t
*
3900 hammer2_chain_rdata(hammer2_chain_t
*chain
)
3902 KKASSERT(chain
->data
!= NULL
);
3903 return (chain
->data
);
3906 hammer2_media_data_t
*
3907 hammer2_chain_wdata(hammer2_chain_t
*chain
)
3909 KKASSERT(chain
->data
!= NULL
);
3910 return (chain
->data
);
3914 * Set the check data for a chain. This can be a heavy-weight operation
3915 * and typically only runs on-flush. For file data check data is calculated
3916 * when the logical buffers are flushed.
3919 hammer2_chain_setcheck(hammer2_chain_t
*chain
, void *bdata
)
3921 chain
->bref
.flags
&= ~HAMMER2_BREF_FLAG_ZERO
;
3923 switch(HAMMER2_DEC_CHECK(chain
->bref
.methods
)) {
3924 case HAMMER2_CHECK_NONE
:
3926 case HAMMER2_CHECK_DISABLED
:
3928 case HAMMER2_CHECK_ISCSI32
:
3929 chain
->bref
.check
.iscsi32
.value
=
3930 hammer2_icrc32(bdata
, chain
->bytes
);
3932 case HAMMER2_CHECK_CRC64
:
3933 chain
->bref
.check
.crc64
.value
= 0;
3936 case HAMMER2_CHECK_SHA192
:
3938 SHA256_CTX hash_ctx
;
3940 uint8_t digest
[SHA256_DIGEST_LENGTH
];
3941 uint64_t digest64
[SHA256_DIGEST_LENGTH
/8];
3944 SHA256_Init(&hash_ctx
);
3945 SHA256_Update(&hash_ctx
, bdata
, chain
->bytes
);
3946 SHA256_Final(u
.digest
, &hash_ctx
);
3947 u
.digest64
[2] ^= u
.digest64
[3];
3949 chain
->bref
.check
.sha192
.data
,
3950 sizeof(chain
->bref
.check
.sha192
.data
));
3953 case HAMMER2_CHECK_FREEMAP
:
3954 chain
->bref
.check
.freemap
.icrc32
=
3955 hammer2_icrc32(bdata
, chain
->bytes
);
3958 kprintf("hammer2_chain_setcheck: unknown check type %02x\n",
3959 chain
->bref
.methods
);
3965 hammer2_chain_testcheck(hammer2_chain_t
*chain
, void *bdata
)
3969 if (chain
->bref
.flags
& HAMMER2_BREF_FLAG_ZERO
)
3972 switch(HAMMER2_DEC_CHECK(chain
->bref
.methods
)) {
3973 case HAMMER2_CHECK_NONE
:
3976 case HAMMER2_CHECK_DISABLED
:
3979 case HAMMER2_CHECK_ISCSI32
:
3980 r
= (chain
->bref
.check
.iscsi32
.value
==
3981 hammer2_icrc32(bdata
, chain
->bytes
));
3983 case HAMMER2_CHECK_CRC64
:
3984 r
= (chain
->bref
.check
.crc64
.value
== 0);
3987 case HAMMER2_CHECK_SHA192
:
3989 SHA256_CTX hash_ctx
;
3991 uint8_t digest
[SHA256_DIGEST_LENGTH
];
3992 uint64_t digest64
[SHA256_DIGEST_LENGTH
/8];
3995 SHA256_Init(&hash_ctx
);
3996 SHA256_Update(&hash_ctx
, bdata
, chain
->bytes
);
3997 SHA256_Final(u
.digest
, &hash_ctx
);
3998 u
.digest64
[2] ^= u
.digest64
[3];
4000 chain
->bref
.check
.sha192
.data
,
4001 sizeof(chain
->bref
.check
.sha192
.data
)) == 0) {
4008 case HAMMER2_CHECK_FREEMAP
:
4009 r
= (chain
->bref
.check
.freemap
.icrc32
==
4010 hammer2_icrc32(bdata
, chain
->bytes
));
4012 kprintf("freemap.icrc %08x icrc32 %08x (%d)\n",
4013 chain
->bref
.check
.freemap
.icrc32
,
4014 hammer2_icrc32(bdata
, chain
->bytes
), chain
->bytes
);
4016 kprintf("dio %p buf %016jx,%d bdata %p/%p\n",
4017 chain
->dio
, chain
->dio
->bp
->b_loffset
, chain
->dio
->bp
->b_bufsize
, bdata
, chain
->dio
->bp
->b_data
);
4022 kprintf("hammer2_chain_setcheck: unknown check type %02x\n",
4023 chain
->bref
.methods
);