| blob | 4caa4153278e53ea05538ecb19ef5f4420d3c7c2 |
1 /*
2 * SPDX-License-Identifier: BSD-3-Clause
3 *
4 * Copyright (c) 2022 Tomohiro Kusumi <tkusumi@netbsd.org>
5 * Copyright (c) 2011-2022 The DragonFly Project. All rights reserved.
6 *
7 * This code is derived from software contributed to The DragonFly Project
8 * by Matthew Dillon <dillon@dragonflybsd.org>
9 *
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
12 * are met:
13 *
14 * 1. Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
16 * 2. Redistributions in binary form must reproduce the above copyright
17 * notice, this list of conditions and the following disclaimer in
18 * the documentation and/or other materials provided with the
19 * distribution.
20 * 3. Neither the name of The DragonFly Project nor the names of its
21 * contributors may be used to endorse or promote products derived
22 * from this software without specific, prior written permission.
23 *
24 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
25 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
26 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
27 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
28 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
29 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
30 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
31 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
32 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
33 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
34 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35 * SUCH DAMAGE.
36 */
37 /*
38 * This subsystem implements most of the core support functions for
39 * the hammer2_chain structure.
40 *
41 * Chains are the in-memory version on media objects (volume header, inodes,
42 * indirect blocks, data blocks, etc). Chains represent a portion of the
43 * HAMMER2 topology.
44 *
45 * Chains are no-longer delete-duplicated. Instead, the original in-memory
46 * chain will be moved along with its block reference (e.g. for things like
47 * renames, hardlink operations, modifications, etc), and will be indexed
48 * on a secondary list for flush handling instead of propagating a flag
49 * upward to the root.
50 *
51 * Concurrent front-end operations can still run against backend flushes
52 * as long as they do not cross the current flush boundary. An operation
53 * running above the current flush (in areas not yet flushed) can become
54 * part of the current flush while ano peration running below the current
55 * flush can become part of the next flush.
56 */
57 /*
58 #include <sys/cdefs.h>
59 #include <sys/param.h>
60 #include <sys/systm.h>
61 #include <sys/types.h>
62 #include <sys/lock.h>
63 #include <sys/buf.h>
65 #include <crypto/sha2/sha2.h>
66 */
88 /*
89 * There are many degenerate situations where an extreme rate of console
90 * output can occur from warnings and errors. Make sure this output does
91 * not impede operations.
92 */
93 /*
94 static struct krate krate_h2chk = { .freq = 5 };
95 static struct krate krate_h2me = { .freq = 1 };
96 static struct krate krate_h2em = { .freq = 1 };
97 */
99 /*
100 * Basic RBTree for chains (core.rbtree).
101 */
104 int
106 {
107 hammer2_key_t c1_beg;
108 hammer2_key_t c1_end;
109 hammer2_key_t c2_beg;
110 hammer2_key_t c2_end;
112 /*
113 * Compare chains. Overlaps are not supposed to happen and catch
114 * any software issues early we count overlaps as a match.
115 */
126 }
128 /*
129 * Assert that a chain has no media data associated with it.
130 */
133 {
139 chain);
140 }
141 }
143 /*
144 * Make a chain visible to the flusher. The flusher operates using a top-down
145 * recursion based on the ONFLUSH flag. It locates MODIFIED and UPDATE chains,
146 * flushes them, and updates blocks back to the volume root.
147 *
148 * This routine sets the ONFLUSH flag upward from the triggering chain until
149 * it hits an inode root or the volume root. Inode chains serve as inflection
150 * points, requiring the flusher to bridge across trees. Inodes include
151 * regular inodes, PFS roots (pmp->iroot), and the media super root
152 * (spmp->iroot).
153 */
154 void
156 {
170 }
172 }
173 }
175 /*
176 * Allocate a new disconnected chain element representing the specified
177 * bref. chain->refs is set to 1 and the passed bref is copied to
178 * chain->bref. chain->bytes is derived from the bref.
179 *
180 * chain->pmp inherits pmp unless the chain is an inode (other than the
181 * super-root inode).
182 *
183 * NOTE: Returns a referenced but unlocked (because there is no core) chain.
184 */
185 hammer2_chain_t *
188 {
190 u_int bytes;
192 /*
193 * Special case - radix of 0 indicates a chain that does not
194 * need a data reference (context is completely embedded in the
195 * bref).
196 */
199 else
220 }
222 /*
223 * Initialize the new chain structure. pmp must be set to NULL for
224 * chains belonging to the super-root topology of a device mount.
225 */
228 else
238 /*
239 * Set the PFS boundary flag if this chain represents a PFS root.
240 */
246 }
248 /*
249 * A common function to initialize chains including fchain and vchain.
250 */
251 void
253 {
257 }
259 /*
260 * Add a reference to a chain element, preventing its destruction.
261 *
262 * (can be called with spinlock held)
263 */
264 void
266 {
268 /*
269 * Just flag that the chain was used and should be recycled
270 * on the LRU if it encounters it later.
271 */
275 #if 0
276 /*
277 * REMOVED - reduces contention, lru_list is more heuristical
278 * now.
279 *
280 * 0->non-zero transition must ensure that chain is removed
281 * from the LRU list.
282 *
283 * NOTE: Already holding lru_spin here so we cannot call
284 * hammer2_chain_ref() to get it off lru_list, do
285 * it manually.
286 */
293 HAMMER2_CHAIN_ONLRU);
295 }
297 }
298 #endif
299 }
300 }
302 /*
303 * Ref a locked chain and force the data to be held across an unlock.
304 * Chain must be currently locked. The user of the chain who desires
305 * to release the hold must call hammer2_chain_lock_unhold() to relock
306 * and unhold the chain, then unlock normally, or may simply call
307 * hammer2_chain_drop_unhold() (which is safer against deadlocks).
308 */
309 void
311 {
314 }
316 /*
317 * Insert the chain in the core rbtree.
318 *
319 * Normal insertions are placed in the live rbtree. Insertion of a deleted
320 * chain is a special case used by the flush code that is placed on the
321 * unstaged deleted list to avoid confusing the live view.
322 */
323 #define HAMMER2_CHAIN_INSERT_SPIN 0x0001
324 #define HAMMER2_CHAIN_INSERT_LIVE 0x0002
325 #define HAMMER2_CHAIN_INSERT_RACE 0x0004
327 static
328 int
331 {
338 /*
339 * Interlocked by spinlock, check for race
340 */
345 }
347 /*
348 * Insert chain
349 */
359 /*
360 * We have to keep track of the effective live-view blockref count
361 * so the create code knows when to push an indirect block.
362 */
365 failed:
369 }
371 /*
372 * Drop the caller's reference to the chain. When the ref count drops to
373 * zero this function will try to disassociate the chain from its parent and
374 * deallocate it, then recursely drop the parent using the implied ref
375 * from the chain's chain->parent.
376 *
377 * Nobody should own chain's mutex on the 1->0 transition, unless this drop
378 * races an acquisition by another cpu. Therefore we can loop if we are
379 * unable to acquire the mutex, and refs is unlikely to be 1 unless we again
380 * race against another drop.
381 */
382 void
384 {
385 u_int refs;
397 /* retry the same chain, or chain from lastdrop */
401 /* retry the same chain */
402 }
404 }
405 }
407 /*
408 * Unhold a held and probably not-locked chain, ensure that the data is
409 * dropped on the 1->0 transition of lockcnt by obtaining an exclusive
410 * lock and then simply unlocking the chain.
411 */
412 void
414 {
415 u_int lockcnt;
425 }
430 /*
431 * This situation can easily occur on SMP due to
432 * the gap inbetween the 1->0 transition and the
433 * final unlock. We cannot safely block on the
434 * mutex because lockcnt might go above 1.
435 *
436 * XXX Sleep for one tick if it takes too long.
437 */
442 }
444 }
446 }
447 }
448 }
450 void
452 {
455 }
457 void
459 {
463 }
465 /*
466 * Handles the (potential) last drop of chain->refs from 1->0. Called with
467 * the mutex exclusively locked, refs == 1, and lockcnt 0. SMP races are
468 * possible against refs and lockcnt. We must dispose of the mutex on chain.
469 *
470 * This function returns an unlocked chain for recursive drop or NULL. It
471 * can return the same chain if it determines it has raced another ref.
472 *
473 * --
474 *
475 * When two chains need to be recursively dropped we use the chain we
476 * would otherwise free to placehold the additional chain. It's a bit
477 * convoluted but we can't just recurse without potentially blowing out
478 * the kernel stack.
479 *
480 * The chain cannot be freed if it has any children.
481 * The chain cannot be freed if flagged MODIFIED unless we can dispose of it.
482 * The chain cannot be freed if flagged UPDATE unless we can dispose of it.
483 * Any dedup registration can remain intact.
484 *
485 * The core spinlock is allowed to nest child-to-parent (not parent-to-child).
486 */
487 static
488 hammer2_chain_t *
490 {
496 /*
497 * We need chain's spinlock to interlock the sub-tree test.
498 * We already have chain's mutex, protecting chain->parent.
499 *
500 * Remember that chain->refs can be in flux.
501 */
505 /*
506 * If the chain has a parent the UPDATE bit prevents scrapping
507 * as the chain is needed to properly flush the parent. Try
508 * to complete the 1->0 transition and return NULL. Retry
509 * (return chain) if we are unable to complete the 1->0
510 * transition, else return NULL (nothing more to do).
511 *
512 * If the chain has a parent the MODIFIED bit prevents
513 * scrapping.
514 *
515 * Chains with UPDATE/MODIFIED are *not* put on the LRU list!
516 */
518 HAMMER2_CHAIN_MODIFIED)) {
527 }
529 }
530 /* spinlock still held */
533 /*
534 * Retain the static vchain and fchain. Clear bits that
535 * are not relevant. Do not clear the MODIFIED bit,
536 * and certainly do not put it on the delayed-flush queue.
537 */
540 /*
541 * The chain has no parent and can be flagged for destruction.
542 * Since it has no parent, UPDATE can also be cleared.
543 */
548 /*
549 * If the chain has children we must propagate the DESTROY
550 * flag downward and rip the disconnected topology apart.
551 * This is accomplished by calling hammer2_flush() on the
552 * chain.
553 *
554 * Any dedup is already handled by the underlying DIO, so
555 * we do not have to specifically flush it here.
556 */
560 HAMMER2_FLUSH_ALL);
564 }
566 /*
567 * Otherwise we can scrap the MODIFIED bit if it is set,
568 * and continue along the freeing path.
569 *
570 * Be sure to clean-out any dedup bits. Without a parent
571 * this chain will no longer be visible to the flush code.
572 * Easy check data_off to avoid the volume root.
573 */
579 }
580 /* spinlock still held */
581 }
583 /* spinlock still held */
585 /*
586 * If any children exist we must leave the chain intact with refs == 0.
587 * They exist because chains are retained below us which have refs or
588 * may require flushing.
589 *
590 * Retry (return chain) if we fail to transition the refs to 0, else
591 * return NULL indication nothing more to do.
592 *
593 * Chains with children are NOT put on the LRU list.
594 */
604 }
606 }
607 /* spinlock still held */
608 /* no chains left under us */
610 /*
611 * chain->core has no children left so no accessors can get to our
612 * chain from there. Now we have to lock the parent core to interlock
613 * remaining possible accessors that might bump chain's refs before
614 * we can safely drop chain's refs with intent to free the chain.
615 */
622 /*
623 * WARNING! chain's spin lock is still held here, and other spinlocks
624 * will be acquired and released in the code below. We
625 * cannot be making fancy procedure calls!
626 */
628 /*
629 * We can cache the chain if it is associated with a pmp
630 * and not flagged as being destroyed or requesting a full
631 * release. In this situation the chain is not removed
632 * from its parent, i.e. it can still be looked up.
633 *
634 * We intentionally do not cache DATA chains because these
635 * were likely used to load data into the logical buffer cache
636 * and will not be accessed again for some time.
637 */
645 /*
646 * 1->0 transition failed, retry. Do not drop
647 * the chain's data yet!
648 */
655 }
657 /*
658 * Success
659 */
662 /*
663 * Make sure we are on the LRU list, clean up excessive
664 * LRU entries. We can only really drop one but there might
665 * be other entries that we can remove from the lru_list
666 * without dropping.
667 *
668 * NOTE: HAMMER2_CHAIN_ONLRU may only be safely set when
669 * chain->core.spin AND pmp->lru_spin are held, but
670 * can be safely cleared only holding pmp->lru_spin.
671 */
676 HAMMER2_CHAIN_ONLRU);
680 }
685 /* disable lru flush */
687 }
692 }
696 /*
697 * lru_list hysteresis (see above for depth overrides).
698 * Note that depth also prevents excessive lastdrop recursion.
699 */
704 /* NOT REACHED */
705 }
707 /*
708 * Make sure we are not on the LRU list.
709 */
716 }
718 }
720 /*
721 * Spinlock the parent and try to drop the last ref on chain.
722 * On success determine if we should dispose of the chain
723 * (remove the chain from its parent, etc).
724 *
725 * (normal core locks are top-down recursive but we define
726 * core spinlocks as bottom-up recursive, so this is safe).
727 */
731 /*
732 * 1->0 transition failed, retry.
733 */
739 }
741 /*
742 * 1->0 transition successful, parent spin held to prevent
743 * new lookups, chain spinlock held to protect parent field.
744 * Remove chain from the parent.
745 *
746 * If the chain is being removed from the parent's rbtree but
747 * is not blkmapped, we have to adjust live_count downward. If
748 * it is blkmapped then the blockref is retained in the parent
749 * as is its associated live_count. This case can occur when
750 * a chain added to the topology is unable to flush and is
751 * then later deleted.
752 */
757 }
763 }
765 /*
766 * If our chain was the last chain in the parent's core the
767 * core is now empty and its parent might have to be
768 * re-dropped if it has 0 refs.
769 */
773 /*
774 if (atomic_cmpset_int(&rdrop->refs, 0, 1) == 0)
775 rdrop = NULL;
776 */
777 }
780 /* FALL THROUGH */
782 /*
783 * No-parent case.
784 */
786 /*
787 * 1->0 transition failed, retry.
788 */
794 }
795 }
797 /*
798 * Successful 1->0 transition, no parent, no children... no way for
799 * anyone to ref this chain any more. We can clean-up and free it.
800 *
801 * We still have the core spinlock, and core's chain_count is 0.
802 * Any parent spinlock is gone.
803 */
810 /*
811 * All locks are gone, no pointers remain to the chain, finish
812 * freeing it.
813 */
817 /*
818 * Once chain resources are gone we can use the now dead chain
819 * structure to placehold what might otherwise require a recursive
820 * drop, because we have potentially two things to drop and can only
821 * return one directly.
822 */
828 }
830 /*
831 * Possible chaining loop when parent re-drop needed.
832 */
834 }
836 /*
837 * Heuristical flush of the LRU, try to reduce the number of entries
838 * on the LRU to (HAMMER2_LRU_LIMIT * 2 / 3). This procedure is called
839 * only when lru_count exceeds HAMMER2_LRU_LIMIT.
840 */
841 static
842 void
844 {
847 again:
851 /*
852 * Pick a chain off the lru_list, just recycle it quickly
853 * if LRUHINT is set (the chain was ref'd but left on
854 * the lru_list, so cycle to the end).
855 */
864 }
866 /*
867 * Ok, we are off the LRU. We must adjust refs before we
868 * can safely clear the ONLRU flag.
869 */
875 }
878 }
883 /*
884 * If we picked a chain off the lru list we may be able to lastdrop
885 * it. Use a depth of 1 to prevent excessive lastdrop recursion.
886 */
888 u_int refs;
897 /* retry the same chain, or chain from lastdrop */
901 /* retry the same chain */
902 }
904 }
906 }
908 /*
909 * On last lock release.
910 */
913 {
928 "chain %p bref %016jx.%02x "
934 }
937 }
938 }
940 }
942 /*
943 * Lock a referenced chain element, acquiring its data with I/O if necessary,
944 * and specify how you would like the data to be resolved.
945 *
946 * If an I/O or other fatal error occurs, chain->error will be set to non-zero.
947 *
948 * The lock is allowed to recurse, multiple locking ops will aggregate
949 * the requested resolve types. Once data is assigned it will not be
950 * removed until the last unlock.
951 *
952 * HAMMER2_RESOLVE_NEVER - Do not resolve the data element.
953 * (typically used to avoid device/logical buffer
954 * aliasing for data)
955 *
956 * HAMMER2_RESOLVE_MAYBE - Do not resolve data elements for chains in
957 * the INITIAL-create state (indirect blocks only).
958 *
959 * Do not resolve data elements for DATA chains.
960 * (typically used to avoid device/logical buffer
961 * aliasing for data)
962 *
963 * HAMMER2_RESOLVE_ALWAYS- Always resolve the data element.
964 *
965 * HAMMER2_RESOLVE_SHARED- (flag) The chain is locked shared, otherwise
966 * it will be locked exclusive.
967 *
968 * HAMMER2_RESOLVE_NONBLOCK- (flag) The chain is locked non-blocking. If
969 * the lock fails, EAGAIN is returned.
970 *
971 * NOTE: Embedded elements (volume header, inodes) are always resolved
972 * regardless.
973 *
974 * NOTE: Specifying HAMMER2_RESOLVE_ALWAYS on a newly-created non-embedded
975 * element will instantiate and zero its buffer, and flush it on
976 * release.
977 *
978 * NOTE: (data) elements are normally locked RESOLVE_NEVER or RESOLVE_MAYBE
979 * so as not to instantiate a device buffer, which could alias against
980 * a logical file buffer. However, if ALWAYS is specified the
981 * device buffer will be instantiated anyway.
982 *
983 * NOTE: The return value is always 0 unless NONBLOCK is specified, in which
984 * case it can be either 0 or EAGAIN.
985 *
986 * WARNING! This function blocks on I/O if data needs to be fetched. This
987 * blocking can run concurrent with other compatible lock holders
988 * who do not need data returning. The lock is not upgraded to
989 * exclusive during a data fetch, a separate bit is used to
990 * interlock I/O. However, an exclusive lock holder can still count
991 * on being interlocked against an I/O fetch managed by a shared
992 * lock holder.
993 */
994 int
996 {
1000 /*
1001 * We still have to bump lockcnt before acquiring the lock,
1002 * even for non-blocking operation, because the unlock code
1003 * live-loops on lockcnt == 1 when dropping the last lock.
1004 *
1005 * If the non-blocking operation fails we have to use an
1006 * unhold sequence to undo the mess.
1007 *
1008 * NOTE: LOCKAGAIN must always succeed without blocking,
1009 * even if NONBLOCK is specified.
1010 */
1019 }
1020 }
1025 }
1026 }
1028 /*
1029 * Get the appropriate lock. If LOCKAGAIN is flagged with
1030 * SHARED the caller expects a shared lock to already be
1031 * present and we are giving it another ref. This case must
1032 * importantly not block if there is a pending exclusive lock
1033 * request.
1034 */
1041 }
1044 }
1045 }
1047 /*
1048 * If we already have a valid data pointer make sure the data is
1049 * synchronized to the current cpu, and then no further action is
1050 * necessary.
1051 */
1056 }
1058 /*
1059 * Do we have to resolve the data? This is generally only
1060 * applicable to HAMMER2_BREF_TYPE_DATA which is special-cased.
1061 * Other BREF types expects the data to be there.
1062 */
1071 #if 0
1076 #endif
1077 /* fall through */
1081 }
1083 /*
1084 * Caller requires data
1085 */
1089 }
1091 #if 0
1092 /*
1093 * Lock the chain, retain the hold, and drop the data persistence count.
1094 * The data should remain valid because we never transitioned lockcnt
1095 * through 0.
1096 */
1097 void
1099 {
1102 }
1104 /*
1105 * Downgrade an exclusive chain lock to a shared chain lock.
1106 *
1107 * NOTE: There is no upgrade equivalent due to the ease of
1108 * deadlocks in that direction.
1109 */
1110 void
1112 {
1114 }
1115 #endif
1117 /*
1118 * Issue I/O and install chain->data. Caller must hold a chain lock, lock
1119 * may be of any type.
1120 *
1121 * Once chain->data is set it cannot be disposed of until all locks are
1122 * released.
1123 *
1124 * Make sure the data is synchronized to the current cpu.
1125 */
1126 void
1128 {
1135 /*
1136 * Degenerate case, data already present, or chain has no media
1137 * reference to load.
1138 */
1144 }
1151 /*
1152 * Gain the IOINPROG bit, interlocked block.
1153 */
1155 u_int oflags;
1156 u_int nflags;
1166 }
1167 /* retry */
1172 }
1173 /* retry */
1174 }
1175 }
1177 /*
1178 * We own CHAIN_IOINPROG
1179 *
1180 * Degenerate case if we raced another load.
1181 */
1186 }
1188 /*
1189 * We must resolve to a device buffer, either by issuing I/O or
1190 * by creating a zero-fill element. We do not mark the buffer
1191 * dirty when creating a zero-fill element (the hammer2_chain_modify()
1192 * API must still be used to do that).
1193 *
1194 * The device buffer is variable-sized in powers of 2 down
1195 * to HAMMER2_MIN_ALLOC (typically 1K). A 64K physical storage
1196 * chunk always contains buffers of the same size. (XXX)
1197 *
1198 * The minimum physical IO size may be larger than the variable
1199 * block size.
1200 */
1203 /*
1204 * The getblk() optimization can only be used on newly created
1205 * elements if the physical block size matches the request.
1206 */
1216 }
1223 }
1226 /*
1227 * This isn't perfect and can be ignored on OSs which do not have
1228 * an indication as to whether a buffer is coming from cache or
1229 * if I/O was actually issued for the read. TESTEDGOOD will work
1230 * pretty well without the B_IOISSUED logic because chains are
1231 * cached, but in that situation (without B_IOISSUED) it will not
1232 * detect whether a re-read via I/O is corrupted verses the original
1233 * read.
1234 *
1235 * We can't re-run the CRC on every fresh lock. That would be
1236 * insanely expensive.
1237 *
1238 * If the underlying kernel buffer covers the entire chain we can
1239 * use the B_IOISSUED indication to determine if we have to re-run
1240 * the CRC on chain data for chains that managed to stay cached
1241 * across the kernel disposal of the original buffer.
1242 */
1244 //struct m_buf *bp = dio->bp;
1249 HAMMER2_CHAIN_TESTEDGOOD);
1250 //bp->b_flags &= ~B_IOISSUED;
1251 }
1252 }
1254 /*
1255 * NOTE: A locked chain's data cannot be modified without first
1256 * calling hammer2_chain_modify().
1257 */
1259 /*
1260 * NOTE: hammer2_io_data() call issues bkvasync()
1261 */
1265 /*
1266 * Clear INITIAL. In this case we used io_new() and the
1267 * buffer has been zero'd and marked dirty.
1268 *
1269 * CHAIN_MODIFIED has not been set yet, and we leave it
1270 * that way for now. Set a temporary CHAIN_NOTTESTED flag
1271 * to prevent hammer2_chain_testcheck() from trying to match
1272 * a check code that has not yet been generated. This bit
1273 * should NOT end up on the actual media.
1274 */
1278 /*
1279 * check data not currently synchronized due to
1280 * modification. XXX assumes data stays in the buffer
1281 * cache, which might not be true (need biodep on flush
1282 * to calculate crc? or simple crc?).
1283 */
1289 }
1290 }
1292 /*
1293 * Setup the data pointer by pointing it into the buffer.
1294 * WARNING! Other threads can start using the data the instant we
1295 * set chain->data non-NULL.
1296 */
1304 /* fall through */
1311 /*
1312 * Point data at the device buffer and leave dio intact.
1313 */
1316 }
1318 /*
1319 * Release HAMMER2_CHAIN_IOINPROG and signal waiters if requested.
1320 */
1321 done:
1323 u_int oflags;
1324 u_int nflags;
1328 HAMMER2_CHAIN_IOSIGNAL);
1334 }
1335 }
1336 }
1338 /*
1339 * Unlock and deref a chain element.
1340 *
1341 * Remember that the presence of children under chain prevent the chain's
1342 * destruction but do not add additional references, so the dio will still
1343 * be dropped.
1344 */
1345 void
1347 {
1349 u_int lockcnt;
1352 /*
1353 * If multiple locks are present (or being attempted) on this
1354 * particular chain we can just unlock, drop refs, and return.
1355 *
1356 * Otherwise fall-through on the 1->0 transition.
1357 */
1367 }
1369 /* while holding the mutex exclusively */
1373 /*
1374 * This situation can easily occur on SMP due to
1375 * the gap inbetween the 1->0 transition and the
1376 * final unlock. We cannot safely block on the
1377 * mutex because lockcnt might go above 1.
1378 *
1379 * XXX Sleep for one tick if it takes too long.
1380 */
1385 }
1387 }
1389 }
1390 /* retry */
1391 }
1393 /*
1394 * Last unlock / mutex upgraded to exclusive. Drop the data
1395 * reference.
1396 */
1401 }
1403 #if 0
1404 /*
1405 * Unlock and hold chain data intact
1406 */
1407 void
1409 {
1412 }
1413 #endif
1415 /*
1416 * Helper to obtain the blockref[] array base and count for a chain.
1417 *
1418 * XXX Not widely used yet, various use cases need to be validated and
1419 * converted to use this function.
1420 */
1421 static
1422 hammer2_blockref_t *
1424 {
1451 }
1479 }
1480 }
1484 }
1486 /*
1487 * This counts the number of live blockrefs in a block array and
1488 * also calculates the point at which all remaining blockrefs are empty.
1489 * This routine can only be called on a live chain.
1490 *
1491 * Caller holds the chain locked, but possibly with a shared lock. We
1492 * must use an exclusive spinlock to prevent corruption.
1493 *
1494 * NOTE: Flag is not set until after the count is complete, allowing
1495 * callers to test the flag without holding the spinlock.
1496 *
1497 * NOTE: If base is NULL the related chain is still in the INITIAL
1498 * state and there are no blockrefs to count.
1499 *
1500 * NOTE: live_count may already have some counts accumulated due to
1501 * creation and deletion and could even be initially negative.
1502 */
1503 void
1506 {
1513 }
1520 }
1523 }
1524 /* else do not modify live_count */
1526 }
1528 }
1530 /*
1531 * Resize the chain's physical storage allocation in-place. This function does
1532 * not usually adjust the data pointer and must be followed by (typically) a
1533 * hammer2_chain_modify() call to copy any old data over and adjust the
1534 * data pointer.
1535 *
1536 * Chains can be resized smaller without reallocating the storage. Resizing
1537 * larger will reallocate the storage. Excess or prior storage is reclaimed
1538 * asynchronously at a later time.
1539 *
1540 * An nradix value of 0 is special-cased to mean that the storage should
1541 * be disassociated, that is the chain is being resized to 0 bytes (not 1
1542 * byte).
1543 *
1544 * Must be passed an exclusively locked parent and chain.
1545 *
1546 * This function is mostly used with DATA blocks locked RESOLVE_NEVER in order
1547 * to avoid instantiating a device buffer that conflicts with the vnode data
1548 * buffer. However, because H2 can compress or encrypt data, the chain may
1549 * have a dio assigned to it in those situations, and they do not conflict.
1550 *
1551 * XXX return error if cannot resize.
1552 */
1553 int
1557 {
1565 /*
1566 * Only data and indirect blocks can be resized for now.
1567 * (The volu root, inodes, and freemap elements use a fixed size).
1568 */
1574 /*
1575 * Nothing to do if the element is already the proper size
1576 */
1582 /*
1583 * Make sure the old data is instantiated so we can copy it. If this
1584 * is a data block, the device data may be superfluous since the data
1585 * might be in a logical block, but compressed or encrypted data is
1586 * another matter.
1587 *
1588 * NOTE: The modify will set BLKMAPUPD for us if BLKMAPPED is set.
1589 */
1594 /*
1595 * Reallocate the block, even if making it smaller (because different
1596 * block sizes may be in different regions).
1597 *
1598 * NOTE: Operation does not copy the data and may only be used
1599 * to resize data blocks in-place, or directory entry blocks
1600 * which are about to be modified in some manner.
1601 */
1608 /*
1609 * We don't want the followup chain_modify() to try to copy data
1610 * from the old (wrong-sized) buffer. It won't know how much to
1611 * copy. This case should only occur during writes when the
1612 * originator already has the data to write in-hand.
1613 */
1619 }
1621 }
1623 /*
1624 * Set the chain modified so its data can be changed by the caller, or
1625 * install deduplicated data. The caller must call this routine for each
1626 * set of modifications it makes, even if the chain is already flagged
1627 * MODIFIED.
1628 *
1629 * Sets bref.modify_tid to mtid only if mtid != 0. Note that bref.modify_tid
1630 * is a CLC (cluster level change) field and is not updated by parent
1631 * propagation during a flush.
1632 *
1633 * Returns an appropriate HAMMER2_ERROR_* code, which will generally reflect
1634 * chain->error except for HAMMER2_ERROR_ENOSPC. If the allocation fails
1635 * due to no space available, HAMMER2_ERROR_ENOSPC is returned and the chain
1636 * remains unmodified with its old data ref intact and chain->error
1637 * unchanged.
1638 *
1639 * Dedup Handling
1640 *
1641 * If the DEDUPABLE flag is set in the chain the storage must be reallocated
1642 * even if the chain is still flagged MODIFIED. In this case the chain's
1643 * DEDUPABLE flag will be cleared once the new storage has been assigned.
1644 *
1645 * If the caller passes a non-zero dedup_off we will use it to assign the
1646 * new storage. The MODIFIED flag will be *CLEARED* in this case, and
1647 * DEDUPABLE will be set (NOTE: the UPDATE flag is always set). The caller
1648 * must not modify the data content upon return.
1649 */
1650 int
1653 {
1666 /*
1667 * Data is not optional for freemap chains (we must always be sure
1668 * to copy the data on COW storage allocations).
1669 */
1674 }
1676 /*
1677 * Data must be resolved if already assigned, unless explicitly
1678 * flagged otherwise. If we cannot safety load the data the
1679 * modification fails and we return early.
1680 */
1687 }
1690 /*
1691 * Set MODIFIED to indicate that the chain has been modified. A new
1692 * allocation is required when modifying a chain.
1693 *
1694 * Set UPDATE to ensure that the blockref is updated in the parent.
1695 *
1696 * If MODIFIED is already set determine if we can reuse the assigned
1697 * data block or if we need a new data block.
1698 */
1700 /*
1701 * Must set modified bit.
1702 */
1708 /*
1709 * We may be able to avoid a copy-on-write if the chain's
1710 * check mode is set to NONE and the chain's current
1711 * modify_tid is beyond the last explicit snapshot tid.
1712 *
1713 * This implements HAMMER2's overwrite-in-place feature.
1714 *
1715 * NOTE! This data-block cannot be used as a de-duplication
1716 * source when the check mode is set to NONE.
1717 */
1723 HAMMER2_CHECK_NONE &&
1727 /*
1728 * Sector overwrite allowed.
1729 */
1735 /*
1736 * If in emergency delete mode then do a modify-in-
1737 * place on any chain type belonging to the PFS as
1738 * long as it doesn't mess up a snapshot. We might
1739 * be forced to do this anyway a little further down
1740 * in the code if the allocation fails.
1741 *
1742 * Also note that in emergency mode, these modify-in-
1743 * place operations are NOT SAFE. A storage failure,
1744 * power failure, or panic can corrupt the filesystem.
1745 */
1748 /*
1749 * Sector overwrite not allowed, must copy-on-write.
1750 */
1752 }
1754 /*
1755 * If the modified chain was registered for dedup we need
1756 * a new allocation. This only happens for delayed-flush
1757 * chains (i.e. which run through the front-end buffer
1758 * cache).
1759 */
1763 /*
1764 * Already flagged modified, no new allocation is needed.
1765 */
1768 }
1770 /*
1771 * Flag parent update required.
1772 */
1778 }
1780 /*
1781 * The XOP code returns held but unlocked focus chains. This
1782 * prevents the chain from being destroyed but does not prevent
1783 * it from being modified. diolk is used to interlock modifications
1784 * against XOP frontend accesses to the focus.
1785 *
1786 * This allows us to theoretically avoid deadlocking the frontend
1787 * if one of the backends lock up by not formally locking the
1788 * focused chain in the frontend. In addition, the synchronization
1789 * code relies on this mechanism to avoid deadlocking concurrent
1790 * synchronization threads.
1791 */
1794 /*
1795 * The modification or re-modification requires an allocation and
1796 * possible COW. If an error occurs, the previous content and data
1797 * reference is retained and the modification fails.
1798 *
1799 * If dedup_off is non-zero, the caller is requesting a deduplication
1800 * rather than a modification. The MODIFIED bit is not set and the
1801 * data offset is set to the deduplication offset. The data cannot
1802 * be modified.
1803 *
1804 * NOTE: The dedup offset is allowed to be in a partially free state
1805 * and we must be sure to reset it to a fully allocated state
1806 * to force two bulkfree passes to free it again.
1807 *
1808 * NOTE: Only applicable when chain->bytes != 0.
1809 *
1810 * XXX can a chain already be marked MODIFIED without a data
1811 * assignment? If not, assert here instead of testing the case.
1812 */
1816 newmod
1817 ) {
1818 /*
1819 * NOTE: We do not have to remove the dedup
1820 * registration because the area is still
1821 * allocated and the underlying DIO will
1822 * still be flushed.
1823 */
1829 HAMMER2_OFF_MASK_RADIX);
1830 else
1834 HAMMER2_CHAIN_MODIFIED);
1840 }
1842 HAMMER2_FREEMAP_DORECOVER);
1844 HAMMER2_CHAIN_DEDUPABLE);
1849 HAMMER2_CHAIN_DEDUPABLE);
1851 /*
1852 * If we are unable to allocate a new block
1853 * but we are in emergency mode, issue a
1854 * warning to the console and reuse the same
1855 * block.
1856 *
1857 * We behave as if the allocation were
1858 * successful.
1859 *
1860 * THIS IS IMPORTANT: These modifications
1861 * are virtually guaranteed to corrupt any
1862 * snapshots related to this filesystem.
1863 */
1867 HAMMER2_BREF_FLAG_EMERG_MIP;
1870 "hammer2: Emergency Mode WARNING: "
1871 "Operation will likely corrupt "
1872 "related snapshot: "
1880 }
1881 }
1882 }
1883 }
1885 /*
1886 * Stop here if error. We have to undo any flag bits we might
1887 * have set above.
1888 */
1895 }
1898 }
1902 }
1904 /*
1905 * Update mirror_tid and modify_tid. modify_tid is only updated
1906 * if not passed as zero (during flushes, parent propagation passes
1907 * the value 0).
1908 *
1909 * NOTE: chain->pmp could be the device spmp.
1910 */
1915 /*
1916 * Set BLKMAPUPD to tell the flush code that an existing blockmap entry
1917 * requires updating as well as to tell the delete code that the
1918 * chain's blockref might not exactly match (in terms of physical size
1919 * or block offset) the one in the parent's blocktable. The base key
1920 * of course will still match.
1921 */
1925 /*
1926 * Short-cut data block handling when the caller does not need an
1927 * actual data reference to (aka OPTDATA), as long as the chain does
1928 * not already have a data pointer to the data and no de-duplication
1929 * occurred.
1930 *
1931 * This generally means that the modifications are being done via the
1932 * logical buffer cache.
1933 *
1934 * NOTE: If deduplication occurred we have to run through the data
1935 * stuff to clear INITIAL, and the caller will likely want to
1936 * assign the check code anyway. Leaving INITIAL set on a
1937 * dedup can be deadly (it can cause the block to be zero'd!).
1938 *
1939 * This code also handles bytes == 0 (most dirents).
1940 */
1947 }
1948 }
1950 /*
1951 * Clearing the INITIAL flag (for indirect blocks) indicates that
1952 * we've processed the uninitialized storage allocation.
1953 *
1954 * If this flag is already clear we are likely in a copy-on-write
1955 * situation but we have to be sure NOT to bzero the storage if
1956 * no data is present.
1957 *
1958 * Clearing of NOTTESTED is allowed if the MODIFIED bit is set,
1959 */
1965 }
1967 /*
1968 * Instantiate data buffer and possibly execute COW operation
1969 */
1973 /*
1974 * The data is embedded, no copy-on-write operation is
1975 * needed.
1976 */
1980 /*
1981 * The data might be fully embedded.
1982 */
1986 }
1987 /* fall through */
1993 /*
1994 * Perform the copy-on-write operation
1995 *
1996 * zero-fill or copy-on-write depending on whether
1997 * chain->data exists or not and set the dirty state for
1998 * the new buffer. hammer2_io_new() will handle the
1999 * zero-fill.
2000 *
2001 * If a dedup_off was supplied this is an existing block
2002 * and no COW, copy, or further modification is required.
2003 */
2014 }
2017 /*
2018 * If an I/O error occurs make sure callers cannot accidently
2019 * modify the old buffer's contents and corrupt the filesystem.
2020 *
2021 * NOTE: hammer2_io_data() call issues bkvasync()
2022 */
2025 hmp);
2031 }
2036 /*
2037 * COW (unless a dedup).
2038 */
2042 }
2044 /*
2045 * We have a problem. We were asked to COW but
2046 * we don't have any data to COW with!
2047 */
2049 chain);
2050 }
2052 /*
2053 * Retire the old buffer, replace with the new. Dirty or
2054 * redirty the new buffer.
2055 *
2056 * WARNING! The system buffer cache may have already flushed
2057 * the buffer, so we must be sure to [re]dirty it
2058 * for further modification.
2059 *
2060 * If dedup_off was supplied, the caller is not
2061 * expected to make any further modification to the
2062 * buffer.
2063 *
2064 * WARNING! hammer2_get_gdata() assumes dio never transitions
2065 * through NULL in order to optimize away unnecessary
2066 * diolk operations.
2067 */
2068 {
2077 }
2084 }
2085 skip2:
2086 /*
2087 * setflush on parent indicating that the parent must recurse down
2088 * to us. Do not call on chain itself which might already have it
2089 * set.
2090 */
2096 }
2098 /*
2099 * Modify the chain associated with an inode.
2100 */
2101 int
2104 {
2111 }
2113 /*
2114 * This function returns the chain at the nearest key within the specified
2115 * range. The returned chain will be referenced but not locked.
2116 *
2117 * This function will recurse through chain->rbtree as necessary and will
2118 * return a *key_nextp suitable for iteration. *key_nextp is only set if
2119 * the iteration value is less than the current value of *key_nextp.
2120 *
2121 * The caller should use (*key_nextp) to calculate the actual range of
2122 * the returned element, which will be (key_beg to *key_nextp - 1), because
2123 * there might be another element which is superior to the returned element
2124 * and overlaps it.
2125 *
2126 * (*key_nextp) can be passed as key_beg in an iteration only while non-NULL
2127 * chains continue to be returned. On EOF (*key_nextp) may overflow since
2128 * it will wind up being (key_end + 1).
2129 *
2130 * WARNING! Must be called with child's spinlock held. Spinlock remains
2131 * held through the operation.
2132 */
2135 hammer2_key_t key_beg;
2136 hammer2_key_t key_end;
2137 hammer2_key_t key_next;
2138 };
2143 static
2144 hammer2_chain_t *
2147 {
2159 #if 0
2162 #endif
2165 }
2167 static
2168 int
2170 {
2172 hammer2_key_t child_beg;
2173 hammer2_key_t child_end;
2183 }
2185 static
2186 int
2188 {
2191 hammer2_key_t child_end;
2194 /*
2195 * No previous best. Assign best
2196 */
2200 /*
2201 * Illegal overlap.
2202 */
2204 /*info->best = child;*/
2206 /*
2207 * Child has a nearer key and best is not flush with key_beg.
2208 * Set best to child. Truncate key_next to the old best key.
2209 */
2214 /*
2215 * If our current best is flush with the child then this
2216 * is an illegal overlap.
2217 *
2218 * key_next will automatically be limited to the smaller of
2219 * the two end-points.
2220 */
2224 /*
2225 * Keep the current best but truncate key_next to the child's
2226 * base.
2227 *
2228 * key_next will also automatically be limited to the smaller
2229 * of the two end-points (probably not necessary for this case
2230 * but we do it anyway).
2231 */
2234 }
2236 /*
2237 * Always truncate key_next based on child's end-of-range.
2238 */
2244 }
2246 /*
2247 * Retrieve the specified chain from a media blockref, creating the
2248 * in-memory chain structure which reflects it. The returned chain is
2249 * held and locked according to (how) (HAMMER2_RESOLVE_*). The caller must
2250 * handle crc-checks and so forth, and should check chain->error before
2251 * assuming that the data is good.
2252 *
2253 * To handle insertion races pass the INSERT_RACE flag along with the
2254 * generation number of the core. NULL will be returned if the generation
2255 * number changes before we have a chance to insert the chain. Insert
2256 * races can occur because the parent might be held shared.
2257 *
2258 * Caller must hold the parent locked shared or exclusive since we may
2259 * need the parent's bref array to find our block.
2260 *
2261 * WARNING! chain->pmp is always set to NULL for any chain representing
2262 * part of the super-root topology.
2263 */
2264 hammer2_chain_t *
2267 {
2272 /*
2273 * Allocate a chain structure representing the existing media
2274 * entry. Resulting chain has one ref and is not locked.
2275 */
2278 else
2280 /* ref'd chain returned */
2282 /*
2283 * Flag that the chain is in the parent's blockmap so delete/flush
2284 * knows what to do with it.
2285 */
2288 /*
2289 * chain must be locked to avoid unexpected ripouts
2290 */
2293 /*
2294 * Link the chain into its parent. A spinlock is required to safely
2295 * access the RBTREE, and it is possible to collide with another
2296 * hammer2_chain_get() operation because the caller might only hold
2297 * a shared lock on the parent.
2298 *
2299 * NOTE: Get races can occur quite often when we distribute
2300 * asynchronous read-aheads across multiple threads.
2301 */
2304 HAMMER2_CHAIN_INSERT_SPIN |
2305 HAMMER2_CHAIN_INSERT_RACE,
2306 generation);
2309 /*kprintf("chain %p get race\n", chain);*/
2315 }
2317 /*
2318 * Return our new chain referenced but not locked, or NULL if
2319 * a race occurred.
2320 */
2322 }
2324 /*
2325 * Lookup initialization/completion API
2326 */
2327 hammer2_chain_t *
2329 {
2333 HAMMER2_RESOLVE_SHARED);
2336 }
2338 }
2340 void
2342 {
2346 }
2347 }
2349 /*
2350 * Take the locked chain and return a locked parent. The chain remains
2351 * locked on return, but may have to be temporarily unlocked to acquire
2352 * the parent. Because of this, (chain) must be stable and cannot be
2353 * deleted while it was temporarily unlocked (typically means that (chain)
2354 * is an inode).
2355 *
2356 * Pass HAMMER2_RESOLVE_* flags in flags.
2357 *
2358 * This will work even if the chain is errored, and the caller can check
2359 * parent->error on return if desired since the parent will be locked.
2360 *
2361 * This function handles the lock order reversal.
2362 */
2363 hammer2_chain_t *
2365 {
2368 /*
2369 * Be careful of order, chain must be unlocked before parent
2370 * is locked below to avoid a deadlock. Try it trivially first.
2371 */
2384 /*
2385 * Parent relinking races are quite common. We have to get
2386 * it right or we will blow up the block table.
2387 */
2397 }
2399 }
2401 /*
2402 * Take the locked chain and return a locked parent. The chain is unlocked
2403 * and dropped. *chainp is set to the returned parent as a convenience.
2404 * Pass HAMMER2_RESOLVE_* flags in flags.
2405 *
2406 * This will work even if the chain is errored, and the caller can check
2407 * parent->error on return if desired since the parent will be locked.
2408 *
2409 * The chain does NOT need to be stable. We use a tracking structure
2410 * to track the expected parent if the chain is deleted out from under us.
2411 *
2412 * This function handles the lock order reversal.
2413 */
2414 hammer2_chain_t *
2416 {
2422 /*
2423 * Be careful of order, chain must be unlocked before parent
2424 * is locked below to avoid a deadlock. Try it trivially first.
2425 */
2431 }
2439 }
2441 /*
2442 * Ok, now it gets a bit nasty. There are multiple situations where
2443 * the parent might be in the middle of a deletion, or where the child
2444 * (chain) might be deleted the instant we let go of its lock.
2445 * We can potentially end up in a no-win situation!
2446 *
2447 * In particular, the indirect_maintenance() case can cause these
2448 * situations.
2449 *
2450 * To deal with this we install a reptrack structure in the parent
2451 * This reptrack structure 'owns' the parent ref and will automatically
2452 * migrate to the parent's parent if the parent is deleted permanently.
2453 */
2467 /*
2468 * At the top of this loop, chain is gone and parent is refd both
2469 * by us explicitly AND via our reptrack. We are attempting to
2470 * lock parent.
2471 */
2486 }
2488 /*
2489 * Once parent is locked and matches our reptrack, our reptrack
2490 * will be stable and we have our parent. We can unlink our
2491 * reptrack.
2492 *
2493 * WARNING! Remember that the chain lock might be shared. Chains
2494 * locked shared have stable parent linkages.
2495 */
2507 }
2509 /*
2510 * Dispose of any linked reptrack structures in (chain) by shifting them to
2511 * (parent). Both (chain) and (parent) must be exclusively locked.
2512 *
2513 * This is interlocked against any children of (chain) on the other side.
2514 * No children so remain as-of when this is called so we can test
2515 * core.reptrack without holding the spin-lock.
2516 *
2517 * Used whenever the caller intends to permanently delete chains related
2518 * to topological recursions (BREF_TYPE_INDIRECT, BREF_TYPE_FREEMAP_NODE),
2519 * where the chains underneath the node being deleted are given a new parent
2520 * above the node being deleted.
2521 */
2522 static
2523 void
2525 {
2537 }
2550 }
2551 }
2553 /*
2554 * Locate the first chain whos key range overlaps (key_beg, key_end) inclusive.
2555 * (*parentp) typically points to an inode but can also point to a related
2556 * indirect block and this function will recurse upwards and find the inode
2557 * or the nearest undeleted indirect block covering the key range.
2558 *
2559 * This function unconditionally sets *errorp, replacing any previous value.
2560 *
2561 * (*parentp) must be exclusive or shared locked (depending on flags) and
2562 * referenced and can be an inode or an existing indirect block within the
2563 * inode.
2564 *
2565 * If (*parent) is errored out, this function will not attempt to recurse
2566 * the radix tree and will return NULL along with an appropriate *errorp.
2567 * If NULL is returned and *errorp is 0, the requested lookup could not be
2568 * located.
2569 *
2570 * On return (*parentp) will be modified to point at the deepest parent chain
2571 * element encountered during the search, as a helper for an insertion or
2572 * deletion.
2573 *
2574 * The new (*parentp) will be locked shared or exclusive (depending on flags),
2575 * and referenced, and the old will be unlocked and dereferenced (no change
2576 * if they are both the same). This is particularly important if the caller
2577 * wishes to insert a new chain, (*parentp) will be set properly even if NULL
2578 * is returned, as long as no error occurred.
2579 *
2580 * The matching chain will be returned locked according to flags.
2581 *
2582 * --
2583 *
2584 * NULL is returned if no match was found, but (*parentp) will still
2585 * potentially be adjusted.
2586 *
2587 * On return (*key_nextp) will point to an iterative value for key_beg.
2588 * (If NULL is returned (*key_nextp) is set to (key_end + 1)).
2589 *
2590 * This function will also recurse up the chain if the key is not within the
2591 * current parent's range. (*parentp) can never be set to NULL. An iteration
2592 * can simply allow (*parentp) to float inside the loop.
2593 *
2594 * NOTE! chain->data is not always resolved. By default it will not be
2595 * resolved for BREF_TYPE_DATA, FREEMAP_NODE, or FREEMAP_LEAF. Use
2596 * HAMMER2_LOOKUP_ALWAYS to force resolution (but be careful w/
2597 * BREF_TYPE_DATA as the device buffer can alias the logical file
2598 * buffer).
2599 */
2600 hammer2_chain_t *
2604 {
2609 hammer2_blockref_t bsave;
2610 hammer2_key_t scan_beg;
2611 hammer2_key_t scan_end;
2626 }
2631 }
2633 /*
2634 * Recurse (*parentp) upward if necessary until the parent completely
2635 * encloses the key range or we hit the inode.
2636 *
2637 * Handle races against the flusher deleting indirect nodes on its
2638 * way back up by continuing to recurse upward past the deletion.
2639 */
2651 }
2653 }
2654 again:
2658 /*
2659 * MATCHIND case that does not require parent->data (do prior to
2660 * parent->error check).
2661 */
2675 }
2676 }
2680 }
2682 /*
2683 * No lookup is possible if the parent is errored. We delayed
2684 * this check as long as we could to ensure that the parent backup,
2685 * embedded data, and MATCHIND code could still execute.
2686 */
2690 }
2692 /*
2693 * Locate the blockref array. Currently we do a fully associative
2694 * search through the array.
2695 */
2698 /*
2699 * Special shortcut for embedded data returns the inode
2700 * itself. Callers must detect this condition and access
2701 * the embedded data (the strategy code does this for us).
2702 *
2703 * This is only applicable to regular files and softlinks.
2704 *
2705 * We need a second lock on parent. Since we already have
2706 * a lock we must pass LOCKAGAIN to prevent unexpected
2707 * blocking (we don't want to block on a second shared
2708 * ref if an exclusive lock is pending)
2709 */
2711 HAMMER2_OPFLAG_DIRECTDATA) {
2716 }
2719 HAMMER2_RESOLVE_LOCKAGAIN);
2722 }
2728 /*
2729 * Optimize indirect blocks in the INITIAL state to avoid
2730 * I/O.
2731 *
2732 * Debugging: Enter permanent wait state instead of
2733 * panicing on unexpectedly NULL data for the moment.
2734 */
2743 }
2745 }
2763 }
2765 /*
2766 * Merged scan to find next candidate.
2767 *
2768 * hammer2_base_*() functions require the parent->core.live_* fields
2769 * to be synchronized.
2770 *
2771 * We need to hold the spinlock to access the block array and RB tree
2772 * and to interlock chain creation.
2773 */
2777 /*
2778 * Combined search
2779 */
2782 key_nextp,
2787 /*
2788 * Exhausted parent chain, iterate.
2789 */
2796 /*
2797 * Stop if we reached the end of the iteration.
2798 */
2802 }
2804 /*
2805 * Calculate next key, stop if we reached the end of the
2806 * iteration, otherwise go up one level and loop.
2807 */
2814 }
2816 /*
2817 * Selected from blockref or in-memory chain.
2818 */
2829 }
2836 /*
2837 * chain is referenced but not locked. We must lock the
2838 * chain to obtain definitive state.
2839 */
2845 }
2847 }
2854 }
2857 /*
2858 * Skip deleted chains (XXX cache 'i' end-of-block-array? XXX)
2859 *
2860 * NOTE: Chain's key range is not relevant as there might be
2861 * one-offs within the range that are not deleted.
2862 *
2863 * NOTE: Lookups can race delete-duplicate because
2864 * delete-duplicate does not lock the parent's core
2865 * (they just use the spinlock on the core).
2866 */
2878 }
2880 /*
2881 * If the chain element is an indirect block it becomes the new
2882 * parent and we loop on it. We must maintain our top-down locks
2883 * to prevent the flusher from interfering (i.e. doing a
2884 * delete-duplicate and leaving us recursing down a deleted chain).
2885 *
2886 * The parent always has to be locked with at least RESOLVE_MAYBE
2887 * so we can access its data. It might need a fixup if the caller
2888 * passed incompatible flags. Be careful not to cause a deadlock
2889 * as a data-load requires an exclusive lock.
2890 *
2891 * If HAMMER2_LOOKUP_MATCHIND is set and the indirect block's key
2892 * range is within the requested key range we return the indirect
2893 * block and do NOT loop. This is usually only used to acquire
2894 * freemap nodes.
2895 */
2903 }
2904 done:
2905 /*
2906 * All done, return the locked chain.
2907 *
2908 * If the caller does not want a locked chain, replace the lock with
2909 * a ref. Perhaps this can eventually be optimized to not obtain the
2910 * lock in the first place for situations where the data does not
2911 * need to be resolved.
2912 *
2913 * NOTE! A chain->error must be tested by the caller upon return.
2914 * *errorp is only set based on issues which occur while
2915 * trying to reach the chain.
2916 */
2918 }
2920 /*
2921 * After having issued a lookup we can iterate all matching keys.
2922 *
2923 * If chain is non-NULL we continue the iteration from just after it's index.
2924 *
2925 * If chain is NULL we assume the parent was exhausted and continue the
2926 * iteration at the next parent.
2927 *
2928 * If a fatal error occurs (typically an I/O error), a dummy chain is
2929 * returned with chain->error and error-identifying information set. This
2930 * chain will assert if you try to do anything fancy with it.
2931 *
2932 * XXX Depending on where the error occurs we should allow continued iteration.
2933 *
2934 * parent must be locked on entry and remains locked throughout. chain's
2935 * lock status must match flags. Chain is always at least referenced.
2936 *
2937 * WARNING! The MATCHIND flag does not apply to this function.
2938 */
2939 hammer2_chain_t *
2944 {
2948 /*
2949 * Calculate locking flags for upward recursion.
2950 */
2958 /*
2959 * Calculate the next index and recalculate the parent if necessary.
2960 */
2967 /*
2968 * chain invalid past this point, but we can still do a
2969 * pointer comparison w/parent.
2970 *
2971 * Any scan where the lookup returned degenerate data embedded
2972 * in the inode has an invalid index and must terminate.
2973 */
2981 /*
2982 * We reached the end of the iteration.
2983 */
2986 /*
2987 * Continue iteration with next parent unless the current
2988 * parent covers the range.
2989 *
2990 * (This also handles the case of a deleted, empty indirect
2991 * node).
2992 */
2998 }
3000 /*
3001 * And execute
3002 */
3006 }
3008 /*
3009 * Caller wishes to iterate chains under parent, loading new chains into
3010 * chainp. Caller must initialize *chainp to NULL and *firstp to 1, and
3011 * then call hammer2_chain_scan() repeatedly until a non-zero return.
3012 * During the scan, *firstp will be set to 0 and (*chainp) will be replaced
3013 * with the returned chain for the scan. The returned *chainp will be
3014 * locked and referenced. Any prior contents will be unlocked and dropped.
3015 *
3016 * Caller should check the return value. A normal scan EOF will return
3017 * exactly HAMMER2_ERROR_EOF. Any other non-zero value indicates an
3018 * error trying to access parent data. Any error in the returned chain
3019 * must be tested separately by the caller.
3020 *
3021 * (*chainp) is dropped on each scan, but will only be set if the returned
3022 * element itself can recurse. Leaf elements are NOT resolved, loaded, or
3023 * returned via *chainp. The caller will get their bref only.
3024 *
3025 * The raw scan function is similar to lookup/next but does not seek to a key.
3026 * Blockrefs are iterated via first_bref = (parent, NULL) and
3027 * next_chain = (parent, bref).
3028 *
3029 * The passed-in parent must be locked and its data resolved. The function
3030 * nominally returns a locked and referenced *chainp != NULL for chains
3031 * the caller might need to recurse on (and will dipose of any *chainp passed
3032 * in). The caller must check the chain->bref.type either way.
3033 */
3034 int
3038 {
3041 hammer2_key_t key;
3042 hammer2_key_t next_key;
3052 /*
3053 * Scan flags borrowed from lookup.
3054 */
3061 }
3064 }
3066 /*
3067 * Calculate key to locate first/next element, unlocking the previous
3068 * element as we go. Be careful, the key calculation can overflow.
3069 *
3070 * (also reset bref to NULL)
3071 */
3082 }
3086 }
3087 }
3089 again:
3093 }
3097 /*
3098 * Locate the blockref array. Currently we do a fully associative
3099 * search through the array.
3100 */
3103 /*
3104 * An inode with embedded data has no sub-chains.
3105 *
3106 * WARNING! Bulk scan code may pass a static chain marked
3107 * as BREF_TYPE_INODE with a copy of the volume
3108 * root blockset to snapshot the volume.
3109 */
3111 HAMMER2_OPFLAG_DIRECTDATA) {
3114 }
3120 /*
3121 * Optimize indirect blocks in the INITIAL state to avoid
3122 * I/O.
3123 */
3130 }
3147 }
3149 /*
3150 * Merged scan to find next candidate.
3151 *
3152 * hammer2_base_*() functions require the parent->core.live_* fields
3153 * to be synchronized.
3154 *
3155 * We need to hold the spinlock to access the block array and RB tree
3156 * and to interlock chain creation.
3157 */
3170 /*
3171 * Exhausted parent chain, we're done.
3172 */
3178 }
3180 /*
3181 * Copy into the supplied stack-based blockref.
3182 */
3185 /*
3186 * Selected from blockref or in-memory chain.
3187 */
3195 /*
3196 * Recursion, always get the chain
3197 */
3205 /*
3206 * No recursion, do not waste time instantiating
3207 * a chain, just iterate using the bref.
3208 */
3211 }
3213 /*
3214 * Recursion or not we need the chain in order to supply
3215 * the bref.
3216 */
3220 }
3228 }
3230 /*
3231 * Skip deleted chains (XXX cache 'i' end-of-block-array? XXX)
3232 *
3233 * NOTE: chain's key range is not relevant as there might be
3234 * one-offs within the range that are not deleted.
3235 *
3236 * NOTE: XXX this could create problems with scans used in
3237 * situations other than mount-time recovery.
3238 *
3239 * NOTE: Lookups can race delete-duplicate because
3240 * delete-duplicate does not lock the parent's core
3241 * (they just use the spinlock on the core).
3242 */
3252 }
3254 }
3256 done:
3257 /*
3258 * All done, return the bref or NULL, supply chain if necessary.
3259 */
3263 }
3265 /*
3266 * Create and return a new hammer2 system memory structure of the specified
3267 * key, type and size and insert it under (*parentp). This is a full
3268 * insertion, based on the supplied key/keybits, and may involve creating
3269 * indirect blocks and moving other chains around via delete/duplicate.
3270 *
3271 * This call can be made with parent == NULL as long as a non -1 methods
3272 * is supplied. hmp must also be supplied in this situation (otherwise
3273 * hmp is extracted from the supplied parent). The chain will be detached
3274 * from the topology. A later call with both parent and chain can be made
3275 * to attach it.
3276 *
3277 * THE CALLER MUST HAVE ALREADY PROPERLY SEEKED (*parentp) TO THE INSERTION
3278 * POINT SANS ANY REQUIRED INDIRECT BLOCK CREATIONS DUE TO THE ARRAY BEING
3279 * FULL. This typically means that the caller is creating the chain after
3280 * doing a hammer2_chain_lookup().
3281 *
3282 * (*parentp) must be exclusive locked and may be replaced on return
3283 * depending on how much work the function had to do.
3284 *
3285 * (*parentp) must not be errored or this function will assert.
3286 *
3287 * (*chainp) usually starts out NULL and returns the newly created chain,
3288 * but if the caller desires the caller may allocate a disconnected chain
3289 * and pass it in instead.
3290 *
3291 * This function should NOT be used to insert INDIRECT blocks. It is
3292 * typically used to create/insert inodes and data blocks.
3293 *
3294 * Caller must pass-in an exclusively locked parent the new chain is to
3295 * be inserted under, and optionally pass-in a disconnected, exclusively
3296 * locked chain to insert (else we create a new chain). The function will
3297 * adjust (*parentp) as necessary, create or connect the chain, and
3298 * return an exclusively locked chain in *chainp.
3299 *
3300 * When creating a PFSROOT inode under the super-root, pmp is typically NULL
3301 * and will be reassigned.
3302 *
3303 * NOTE: returns HAMMER_ERROR_* flags
3304 */
3305 int
3310 {
3314 hammer2_blockref_t dummy;
3320 /*
3321 * Topology may be crossing a PFS boundary.
3322 */
3328 }
3332 /*
3333 * First allocate media space and construct the dummy bref,
3334 * then allocate the in-memory chain structure. Set the
3335 * INITIAL flag for fresh chains which do not have embedded
3336 * data.
3337 */
3344 /*
3345 * Inherit methods from parent by default. Primarily used
3346 * for BREF_TYPE_DATA. Non-data types *must* be set to
3347 * a non-NONE check algorithm.
3348 */
3351 else
3358 }
3362 /*
3363 * Lock the chain manually, chain_lock will load the chain
3364 * which we do NOT want to do. (note: chain->refs is set
3365 * to 1 by chain_alloc() for us, but lockcnt is not).
3366 */
3371 /*
3372 * Set INITIAL to optimize I/O. The flag will generally be
3373 * processed when we call hammer2_chain_modify().
3374 */
3390 /* fall through */
3395 /*
3396 * leave chain->data NULL, set INITIAL
3397 */
3401 }
3403 /*
3404 * We are reattaching a previously deleted chain, possibly
3405 * under a new parent and possibly with a new key/keybits.
3406 * The chain does not have to be in a modified state. The
3407 * UPDATE flag will be set later on in this routine.
3408 *
3409 * Do NOT mess with the current state of the INITIAL flag.
3410 */
3416 }
3418 /*
3419 * Set the appropriate bref flag if requested.
3420 *
3421 * NOTE! Callers can call this function to move chains without
3422 * knowing about special flags, so don't clear bref flags
3423 * here!
3424 */
3431 /*
3432 * Calculate how many entries we have in the blockref array and
3433 * determine if an indirect block is required when inserting into
3434 * the parent.
3435 */
3436 again:
3448 }
3459 else
3479 }
3481 /*
3482 * Make sure we've counted the brefs
3483 */
3493 /*
3494 * If no free blockref could be found we must create an indirect
3495 * block and move a number of blockrefs into it. With the parent
3496 * locked we can safely lock each child in order to delete+duplicate
3497 * it without causing a deadlock.
3498 *
3499 * This may return the new indirect block or the old parent depending
3500 * on where the key falls. NULL is returned on error.
3501 */
3514 }
3519 }
3521 }
3523 /*
3524 * fall through if parent, or skip to here if no parent.
3525 */
3526 skip:
3531 /*
3532 * Link the chain into its parent.
3533 */
3540 HAMMER2_CHAIN_INSERT_SPIN |
3541 HAMMER2_CHAIN_INSERT_LIVE,
3543 }
3546 /*
3547 * Mark the newly created chain modified. This will cause
3548 * UPDATE to be set and process the INITIAL flag.
3549 *
3550 * Device buffers are not instantiated for DATA elements
3551 * as these are handled by logical buffers.
3552 *
3553 * Indirect and freemap node indirect blocks are handled
3554 * by hammer2_chain_create_indirect() and not by this
3555 * function.
3556 *
3557 * Data for all other bref types is expected to be
3558 * instantiated (INODE, LEAF).
3559 */
3566 HAMMER2_MODIFY_OPTDATA);
3569 /*
3570 * Remaining types are not supported by this function.
3571 * In particular, INDIRECT and LEAF_NODE types are
3572 * handled by create_indirect().
3573 */
3576 /* NOT REACHED */
3578 }
3580 /*
3581 * When reconnecting a chain we must set UPDATE and
3582 * setflush so the flush recognizes that it must update
3583 * the bref in the parent.
3584 */
3587 }
3589 /*
3590 * We must setflush(parent) to ensure that it recurses through to
3591 * chain. setflush(chain) might not work because ONFLUSH is possibly
3592 * already set in the chain (so it won't recurse up to set it in the
3593 * parent).
3594 */
3598 done:
3602 }
3604 /*
3605 * Move the chain from its old parent to a new parent. The chain must have
3606 * already been deleted or already disconnected (or never associated) with
3607 * a parent. The chain is reassociated with the new parent and the deleted
3608 * flag will be cleared (no longer deleted). The chain's modification state
3609 * is not altered.
3610 *
3611 * THE CALLER MUST HAVE ALREADY PROPERLY SEEKED (parent) TO THE INSERTION
3612 * POINT SANS ANY REQUIRED INDIRECT BLOCK CREATIONS DUE TO THE ARRAY BEING
3613 * FULL. This typically means that the caller is creating the chain after
3614 * doing a hammer2_chain_lookup().
3615 *
3616 * Neither (parent) or (chain) can be errored.
3617 *
3618 * If (parent) is non-NULL then the chain is inserted under the parent.
3619 *
3620 * If (parent) is NULL then the newly duplicated chain is not inserted
3621 * anywhere, similar to if it had just been chain_alloc()'d (suitable for
3622 * passing into hammer2_chain_create() after this function returns).
3623 *
3624 * WARNING! This function calls create which means it can insert indirect
3625 * blocks. This can cause other unrelated chains in the parent to
3626 * be moved to a newly inserted indirect block in addition to the
3627 * specific chain.
3628 */
3629 void
3632 {
3636 /*
3637 * WARNING! We should never resolve DATA to device buffers
3638 * (XXX allow it if the caller did?), and since
3639 * we currently do not have the logical buffer cache
3640 * buffer in-hand to fix its cached physical offset
3641 * we also force the modify code to not COW it. XXX
3642 *
3643 * NOTE! We allow error'd chains to be renamed. The bref itself
3644 * is good and can be renamed. The content, however, may
3645 * be inaccessible.
3646 */
3648 /*KKASSERT(chain->error == 0); allow */
3651 /*
3652 * If parent is not NULL the duplicated chain will be entered under
3653 * the parent and the UPDATE bit set to tell flush to update
3654 * the blockref.
3655 *
3656 * We must setflush(parent) to ensure that it recurses through to
3657 * chain. setflush(chain) might not work because ONFLUSH is possibly
3658 * already set in the chain (so it won't recurse up to set it in the
3659 * parent).
3660 *
3661 * Having both chains locked is extremely important for atomicy.
3662 */
3669 HAMMER2_METH_DEFAULT,
3674 }
3675 }
3677 /*
3678 * This works in tandem with delete_obref() to install a blockref in
3679 * (typically) an indirect block that is associated with the chain being
3680 * moved to *parentp.
3681 *
3682 * The reason we need this function is that the caller needs to maintain
3683 * the blockref as it was, and not generate a new blockref for what might
3684 * be a modified chain. Otherwise stuff will leak into the flush that
3685 * the flush code's FLUSH_INODE_STOP flag is unable to catch.
3686 *
3687 * It is EXTREMELY important that we properly set CHAIN_BLKMAPUPD and
3688 * CHAIN_UPDATE. We must set BLKMAPUPD if the bref does not match, and
3689 * we must clear CHAIN_UPDATE (that was likely set by the chain_rename) if
3690 * it does. Otherwise we can end up in a situation where H2 is unable to
3691 * clean up the in-memory chain topology.
3692 *
3693 * The reason for this is that flushes do not generally flush through
3694 * BREF_TYPE_INODE chains and depend on a hammer2_inode_t queued to syncq
3695 * or sideq to properly flush and dispose of the related inode chain's flags.
3696 * Situations where the inode is not actually modified by the frontend,
3697 * but where we have to move the related chains around as we insert or cleanup
3698 * indirect blocks, can leave us with a 'dirty' (non-disposable) in-memory
3699 * inode chain that does not have a hammer2_inode_t associated with it.
3700 */
3701 static void
3705 {
3718 HAMMER2_CHAIN_UPDATE);
3721 }
3722 }
3723 }
3725 /*
3726 * Helper function for deleting chains.
3727 *
3728 * The chain is removed from the live view (the RBTREE) as well as the parent's
3729 * blockmap. Both chain and its parent must be locked.
3730 *
3731 * parent may not be errored. chain can be errored.
3732 */
3733 static int
3737 {
3744 /*
3745 * Chain is blockmapped, so there must be a parent.
3746 * Atomically remove the chain from the parent and remove
3747 * the blockmap entry. The parent must be set modified
3748 * to remove the blockmap entry.
3749 */
3760 /*
3761 * Calculate blockmap pointer
3762 */
3777 /*
3778 * Access the inode's block array. However, there
3779 * is no block array if the inode is flagged
3780 * DIRECTDATA.
3781 */
3785 base =
3789 }
3796 else
3816 }
3818 /*
3819 * delete blockmapped chain from its parent.
3820 *
3821 * The parent is not affected by any statistics in chain
3822 * which are pending synchronization. That is, there is
3823 * nothing to undo in the parent since they have not yet
3824 * been incorporated into the parent.
3825 *
3826 * The parent is affected by statistics stored in inodes.
3827 * Those have already been synchronized, so they must be
3828 * undone. XXX split update possible w/delete in middle?
3829 */
3832 }
3836 /*
3837 * Chain is not blockmapped but a parent is present.
3838 * Atomically remove the chain from the parent. There is
3839 * no blockmap entry to remove.
3840 *
3841 * Because chain was associated with a parent but not
3842 * synchronized, the chain's *_count_up fields contain
3843 * inode adjustment statistics which must be undone.
3844 */
3857 /*
3858 * Chain is not blockmapped and has no parent. This
3859 * is a degenerate case.
3860 */
3862 }
3863 done:
3865 }
3867 /*
3868 * Create an indirect block that covers one or more of the elements in the
3869 * current parent. Either returns the existing parent with no locking or
3870 * ref changes or returns the new indirect block locked and referenced
3871 * and leaving the original parent lock/ref intact as well.
3872 *
3873 * If an error occurs, NULL is returned and *errorp is set to the H2 error.
3874 *
3875 * The returned chain depends on where the specified key falls.
3876 *
3877 * The key/keybits for the indirect mode only needs to follow three rules:
3878 *
3879 * (1) That all elements underneath it fit within its key space and
3880 *
3881 * (2) That all elements outside it are outside its key space.
3882 *
3883 * (3) When creating the new indirect block any elements in the current
3884 * parent that fit within the new indirect block's keyspace must be
3885 * moved into the new indirect block.
3886 *
3887 * (4) The keyspace chosen for the inserted indirect block CAN cover a wider
3888 * keyspace the the current parent, but lookup/iteration rules will
3889 * ensure (and must ensure) that rule (2) for all parents leading up
3890 * to the nearest inode or the root volume header is adhered to. This
3891 * is accomplished by always recursing through matching keyspaces in
3892 * the hammer2_chain_lookup() and hammer2_chain_next() API.
3893 *
3894 * The current implementation calculates the current worst-case keyspace by
3895 * iterating the current parent and then divides it into two halves, choosing
3896 * whichever half has the most elements (not necessarily the half containing
3897 * the requested key).
3898 *
3899 * We can also opt to use the half with the least number of elements. This
3900 * causes lower-numbered keys (aka logical file offsets) to recurse through
3901 * fewer indirect blocks and higher-numbered keys to recurse through more.
3902 * This also has the risk of not moving enough elements to the new indirect
3903 * block and being forced to create several indirect blocks before the element
3904 * can be inserted.
3905 *
3906 * Must be called with an exclusively locked parent.
3907 *
3908 * NOTE: *errorp set to HAMMER_ERROR_* flags
3909 */
3921 static
3922 hammer2_chain_t *
3926 {
3930 hammer2_blockref_t bsave;
3931 hammer2_blockref_t dummy;
3935 hammer2_key_t key_beg;
3936 hammer2_key_t key_end;
3937 hammer2_key_t key_next;
3948 /*
3949 * Calculate the base blockref pointer or NULL if the chain
3950 * is known to be empty. We need to calculate the array count
3951 * for RB lookups either way.
3952 */
3956 /*
3957 * Pre-modify the parent now to avoid having to deal with error
3958 * processing if we tried to later (in the middle of our loop).
3959 *
3960 * We are going to be moving bref's around, the indirect blocks
3961 * cannot be in an initial state. Do not pass MODIFY_OPTDATA.
3962 */
3968 }
3971 /*hammer2_chain_modify(&parent, HAMMER2_MODIFY_OPTDATA);*/
3974 /*
3975 * How big should our new indirect block be? It has to be at least
3976 * as large as its parent for splits to work properly.
3977 *
3978 * The freemap uses a specific indirect block size. The number of
3979 * levels are built dynamically and ultimately depend on the size
3980 * volume. Because freemap blocks are taken from the reserved areas
3981 * of the volume our goal is efficiency (fewer levels) and not so
3982 * much to save disk space.
3983 *
3984 * The first indirect block level for a directory usually uses
3985 * HAMMER2_IND_BYTES_MIN (4KB = 32 directory entries). Due to
3986 * the hash mechanism, this typically gives us a nominal
3987 * 32 * 4 entries with one level of indirection.
3988 *
3989 * We use HAMMER2_IND_BYTES_NOM (16KB = 128 blockrefs) for FILE
3990 * indirect blocks. The initial 4 entries in the inode gives us
3991 * 256KB. Up to 4 indirect blocks gives us 32MB. Three levels
3992 * of indirection gives us 137GB, and so forth. H2 can support
3993 * huge file sizes but they are not typical, so we try to stick
3994 * with compactness and do not use a larger indirect block size.
3995 *
3996 * We could use 64KB (PBUFSIZE), giving us 512 blockrefs, but
3997 * due to the way indirect blocks are created this usually winds
3998 * up being extremely inefficient for small files. Even though
3999 * 16KB requires more levels of indirection for very large files,
4000 * the 16KB records can be ganged together into 64KB DIOs.
4001 */
4007 HAMMER2_OBJTYPE_DIRECTORY)
4009 else
4014 }
4019 }
4022 /*
4023 * When creating an indirect block for a freemap node or leaf
4024 * the key/keybits must be fitted to static radix levels because
4025 * particular radix levels use particular reserved blocks in the
4026 * related zone.
4027 *
4028 * This routine calculates the key/radix of the indirect block
4029 * we need to create, and whether it is on the high-side or the
4030 * low-side.
4031 */
4050 }
4052 /*
4053 * Normalize the key for the radix being represented, keeping the
4054 * high bits and throwing away the low bits.
4055 */
4058 /*
4059 * Ok, create our new indirect block
4060 */
4067 }
4078 /* ichain has one ref at this point */
4080 /*
4081 * We have to mark it modified to allocate its block, but use
4082 * OPTDATA to allow it to remain in the INITIAL state. Otherwise
4083 * it won't be acted upon by the flush code.
4084 *
4085 * XXX remove OPTDATA, we need a fully initialized indirect block to
4086 * be able to move the original blockref.
4087 */
4095 }
4098 /*
4099 * Iterate the original parent and move the matching brefs into
4100 * the new indirect block.
4101 *
4102 * XXX handle flushes.
4103 */
4112 /*
4113 * Parent may have been modified, relocating its block array.
4114 * Reload the base pointer.
4115 */
4122 }
4124 /*
4125 * NOTE: spinlock stays intact, returned chain (if not NULL)
4126 * is not referenced or locked which means that we
4127 * cannot safely check its flagged / deletion status
4128 * until we lock it.
4129 */
4139 /*
4140 * Skip keys that are not within the key/radix of the new
4141 * indirect block. They stay in the parent.
4142 */
4145 }
4147 /*
4148 * Load the new indirect block by acquiring the related
4149 * chains (potentially from media as it might not be
4150 * in-memory). Then move it to the new parent (ichain).
4151 *
4152 * chain is referenced but not locked. We must lock the
4153 * chain to obtain definitive state.
4154 */
4157 /*
4158 * Use chain already present in the RBTREE
4159 */
4164 /*
4165 * Get chain for blockref element. _get returns NULL
4166 * on insertion race.
4167 */
4170 HAMMER2_RESOLVE_NEVER);
4175 }
4176 }
4178 /*
4179 * This is always live so if the chain has been deleted
4180 * we raced someone and we have to retry.
4181 *
4182 * NOTE: Lookups can race delete-duplicate because
4183 * delete-duplicate does not lock the parent's core
4184 * (they just use the spinlock on the core).
4185 *
4186 * (note reversed logic for this one)
4187 */
4197 }
4200 }
4202 /*
4203 * Shift the chain to the indirect block.
4204 *
4205 * WARNING! No reason for us to load chain data, pass NOSTATS
4206 * to prevent delete/insert from trying to access
4207 * inode stats (and thus asserting if there is no
4208 * chain->data loaded).
4209 *
4210 * WARNING! The (parent, chain) deletion may modify the parent
4211 * and invalidate the base pointer.
4212 *
4213 * WARNING! Parent must already be marked modified, so we
4214 * can assume that chain_delete always suceeds.
4215 *
4216 * WARNING! hammer2_chain_repchange() does not have to be
4217 * called (and doesn't work anyway because we are
4218 * only doing a partial shift). A recursion that is
4219 * in-progress can continue at the current parent
4220 * and will be able to properly find its next key.
4221 */
4232 next_key_spinlocked:
4239 /* loop */
4240 }
4243 /*
4244 * Insert the new indirect block into the parent now that we've
4245 * cleared out some entries in the parent. We calculated a good
4246 * insertion index in the loop above (ichain->index).
4247 *
4248 * We don't have to set UPDATE here because we mark ichain
4249 * modified down below (so the normal modified -> flush -> set-moved
4250 * sequence applies).
4251 *
4252 * The insertion shouldn't race as this is a completely new block
4253 * and the parent is locked.
4254 */
4259 HAMMER2_CHAIN_INSERT_SPIN |
4260 HAMMER2_CHAIN_INSERT_LIVE,
4263 /*
4264 * Make sure flushes propogate after our manual insertion.
4265 */
4269 /*
4270 * Figure out what to return.
4271 */
4273 /*
4274 * Key being created is outside the key range,
4275 * return the original parent.
4276 */
4280 /*
4281 * Otherwise its in the range, return the new parent.
4282 * (leave both the new and old parent locked).
4283 */
4285 }
4288 }
4290 /*
4291 * Do maintenance on an indirect chain. Both parent and chain are locked.
4292 *
4293 * Returns non-zero if (chain) is deleted, either due to being empty or
4294 * because its children were safely moved into the parent.
4295 */
4296 int
4299 {
4303 hammer2_blockref_t bsave;
4304 hammer2_key_t key_next;
4305 hammer2_key_t key_beg;
4306 hammer2_key_t key_end;
4313 /*
4314 * Make sure we have an accurate live_count
4315 */
4321 }
4323 /*
4324 * If the indirect block is empty we can delete it.
4325 * (ignore deletion error)
4326 */
4330 HAMMER2_DELETE_PERMANENT);
4333 }
4340 }
4342 /*
4343 * Determine if we can collapse chain into parent, calculate
4344 * hysteresis for chain emptiness.
4345 */
4352 /*
4353 * Ok, theoretically we can collapse chain's contents into
4354 * parent. chain is locked, but any in-memory children of chain
4355 * are not. For this to work, we must be able to dispose of any
4356 * in-memory children of chain.
4357 *
4358 * For now require that there are no in-memory children of chain.
4359 *
4360 * WARNING! Both chain and parent must remain locked across this
4361 * entire operation.
4362 */
4364 /*
4365 * Parent must be marked modified. Don't try to collapse it if we
4366 * can't mark it modified. Once modified, destroy chain to make room
4367 * and to get rid of what will be a conflicting key (this is included
4368 * in the calculation above). Finally, move the children of chain
4369 * into chain's parent.
4370 *
4371 * This order creates an accounting problem for bref.embed.stats
4372 * because we destroy chain before we remove its children. Any
4373 * elements whos blockref is already synchronized will be counted
4374 * twice. To deal with the problem we clean out chain's stats prior
4375 * to deleting it.
4376 */
4382 }
4388 }
4394 HAMMER2_DELETE_PERMANENT);
4397 /*
4398 * The combined_find call requires core.spin to be held. One would
4399 * think there wouldn't be any conflicts since we hold chain
4400 * exclusively locked, but the caching mechanism for 0-ref children
4401 * does not require a chain lock.
4402 */
4428 HAMMER2_RESOLVE_NEVER);
4432 }
4433 }
4442 }
4457 }
4463 }
4465 /*
4466 * Freemap indirect blocks
4467 *
4468 * Calculate the keybits and highside/lowside of the freemap node the
4469 * caller is creating.
4470 *
4471 * This routine will specify the next higher-level freemap key/radix
4472 * representing the lowest-ordered set. By doing so, eventually all
4473 * low-ordered sets will be moved one level down.
4474 *
4475 * We have to be careful here because the freemap reserves a limited
4476 * number of blocks for a limited number of levels. So we can't just
4477 * push indiscriminately.
4478 */
4479 int
4482 {
4485 hammer2_key_t key;
4486 hammer2_key_t key_beg;
4487 hammer2_key_t key_end;
4488 hammer2_key_t key_next;
4494 /*
4495 * Calculate the range of keys in the array being careful to skip
4496 * slots which are overridden with a deletion.
4497 */
4506 }
4512 /*
4513 * Exhausted search
4514 */
4518 /*
4519 * Skip deleted chains.
4520 */
4526 }
4528 /*
4529 * Use the full live (not deleted) element for the scan
4530 * iteration. HAMMER2 does not allow partial replacements.
4531 *
4532 * XXX should be built into hammer2_combined_find().
4533 */
4541 }
4545 }
4548 /*
4549 * Return the keybits for a higher-level FREEMAP_NODE covering
4550 * this node.
4551 */
4574 }
4578 }
4580 /*
4581 * File indirect blocks
4582 *
4583 * Calculate the key/keybits for the indirect block to create by scanning
4584 * existing keys. The key being created is also passed in *keyp and can be
4585 * inside or outside the indirect block. Regardless, the indirect block
4586 * must hold at least two keys in order to guarantee sufficient space.
4587 *
4588 * We use a modified version of the freemap's fixed radix tree, but taylored
4589 * for file data. Basically we configure an indirect block encompassing the
4590 * smallest key.
4591 */
4592 static int
4596 {
4599 hammer2_key_t key;
4600 hammer2_key_t key_beg;
4601 hammer2_key_t key_end;
4602 hammer2_key_t key_next;
4609 /*
4610 * Calculate the range of keys in the array being careful to skip
4611 * slots which are overridden with a deletion.
4612 *
4613 * Locate the smallest key.
4614 */
4623 }
4629 /*
4630 * Exhausted search
4631 */
4635 /*
4636 * Skip deleted chains.
4637 */
4643 }
4645 /*
4646 * Use the full live (not deleted) element for the scan
4647 * iteration. HAMMER2 does not allow partial replacements.
4648 *
4649 * XXX should be built into hammer2_combined_find().
4650 */
4658 }
4662 }
4665 /*
4666 * Calculate the static keybits for a higher-level indirect block
4667 * that contains the key.
4668 */
4685 }
4687 /*
4688 * The largest radix that can be returned for an indirect block is
4689 * 63 bits. (The largest practical indirect block radix is actually
4690 * 62 bits because the top-level inode or volume root contains four
4691 * entries, but allow 63 to be returned).
4692 */
4697 }
4699 #if 1
4701 /*
4702 * Directory indirect blocks.
4703 *
4704 * Covers both the inode index (directory of inodes), and directory contents
4705 * (filenames hardlinked to inodes).
4706 *
4707 * Because directory keys are hashed we generally try to cut the space in
4708 * half. We accomodate the inode index (which tends to have linearly
4709 * increasing inode numbers) by ensuring that the keyspace is at least large
4710 * enough to fill up the indirect block being created.
4711 */
4712 static int
4716 {
4719 hammer2_key_t key_beg;
4720 hammer2_key_t key_end;
4721 hammer2_key_t key_next;
4722 hammer2_key_t key;
4728 /*
4729 * NOTE: We can't take a shortcut here anymore for inodes because
4730 * the root directory can contain a mix of inodes and directory
4731 * entries (we used to just return 63 if parent->bref.type was
4732 * HAMMER2_BREF_TYPE_INODE.
4733 */
4738 /*
4739 * Calculate the range of keys in the array being careful to skip
4740 * slots which are overridden with a deletion.
4741 */
4750 }
4756 /*
4757 * Exhausted search
4758 */
4762 /*
4763 * Deleted object
4764 */
4770 }
4772 /*
4773 * Use the full live (not deleted) element for the scan
4774 * iteration. HAMMER2 does not allow partial replacements.
4775 *
4776 * XXX should be built into hammer2_combined_find().
4777 */
4780 /*
4781 * Expand our calculated key range (key, keybits) to fit
4782 * the scanned key. nkeybits represents the full range
4783 * that we will later cut in half (two halves @ nkeybits - 1).
4784 */
4791 }
4793 }
4797 }
4799 /*
4800 * If the new key range is larger we have to determine
4801 * which side of the new key range the existing keys fall
4802 * under by checking the high bit, then collapsing the
4803 * locount into the hicount or vise-versa.
4804 */
4812 }
4814 }
4816 /*
4817 * The newly scanned key will be in the lower half or the
4818 * upper half of the (new) key range.
4819 */
4822 else
4828 }
4832 /*
4833 * Adjust keybits to represent half of the full range calculated
4834 * above (radix 63 max) for our new indirect block.
4835 */
4838 /*
4839 * Expand keybits to hold at least ncount elements. ncount will be
4840 * a power of 2. This is to try to completely fill leaf nodes (at
4841 * least for keys which are not hashes).
4842 *
4843 * We aren't counting 'in' or 'out', we are counting 'high side'
4844 * and 'low side' based on the bit at (1LL << keybits). We want
4845 * everything to be inside in these cases so shift it all to
4846 * the low or high side depending on the new high bit.
4847 */
4856 }
4857 }
4861 else
4867 }
4869 #else
4871 /*
4872 * Directory indirect blocks.
4873 *
4874 * Covers both the inode index (directory of inodes), and directory contents
4875 * (filenames hardlinked to inodes).
4876 *
4877 * Because directory keys are hashed we generally try to cut the space in
4878 * half. We accomodate the inode index (which tends to have linearly
4879 * increasing inode numbers) by ensuring that the keyspace is at least large
4880 * enough to fill up the indirect block being created.
4881 */
4882 static int
4886 {
4889 hammer2_key_t key_beg;
4890 hammer2_key_t key_end;
4891 hammer2_key_t key_next;
4892 hammer2_key_t key;
4898 /*
4899 * Shortcut if the parent is the inode. In this situation the
4900 * parent has 4+1 directory entries and we are creating an indirect
4901 * block capable of holding many more.
4902 */
4905 }
4911 /*
4912 * Calculate the range of keys in the array being careful to skip
4913 * slots which are overridden with a deletion.
4914 */
4923 }
4929 /*
4930 * Exhausted search
4931 */
4935 /*
4936 * Deleted object
4937 */
4943 }
4945 /*
4946 * Use the full live (not deleted) element for the scan
4947 * iteration. HAMMER2 does not allow partial replacements.
4948 *
4949 * XXX should be built into hammer2_combined_find().
4950 */
4953 /*
4954 * Expand our calculated key range (key, keybits) to fit
4955 * the scanned key. nkeybits represents the full range
4956 * that we will later cut in half (two halves @ nkeybits - 1).
4957 */
4964 }
4966 }
4971 }
4973 /*
4974 * If the new key range is larger we have to determine
4975 * which side of the new key range the existing keys fall
4976 * under by checking the high bit, then collapsing the
4977 * locount into the hicount or vise-versa.
4978 */
4986 }
4988 }
4990 /*
4991 * The newly scanned key will be in the lower half or the
4992 * upper half of the (new) key range.
4993 */
4996 else
5002 }
5006 /*
5007 * Adjust keybits to represent half of the full range calculated
5008 * above (radix 63 max) for our new indirect block.
5009 */
5012 /*
5013 * Expand keybits to hold at least ncount elements. ncount will be
5014 * a power of 2. This is to try to completely fill leaf nodes (at
5015 * least for keys which are not hashes).
5016 *
5017 * We aren't counting 'in' or 'out', we are counting 'high side'
5018 * and 'low side' based on the bit at (1LL << keybits). We want
5019 * everything to be inside in these cases so shift it all to
5020 * the low or high side depending on the new high bit.
5021 */
5030 }
5031 }
5035 else
5041 }
5043 #endif
5045 /*
5046 * Sets CHAIN_DELETED and remove the chain's blockref from the parent if
5047 * it exists.
5048 *
5049 * Both parent and chain must be locked exclusively.
5050 *
5051 * This function will modify the parent if the blockref requires removal
5052 * from the parent's block table.
5053 *
5054 * This function is NOT recursive. Any entity already pushed into the
5055 * chain (such as an inode) may still need visibility into its contents,
5056 * as well as the ability to read and modify the contents. For example,
5057 * for an unlinked file which is still open.
5058 *
5059 * Also note that the flusher is responsible for cleaning up empty
5060 * indirect blocks.
5061 */
5062 int
5065 {
5070 /*
5071 * Nothing to do if already marked.
5072 *
5073 * We need the spinlock on the core whos RBTREE contains chain
5074 * to protect against races.
5075 */
5081 }
5083 /*
5084 * Permanent deletions mark the chain as destroyed.
5085 *
5086 * NOTE: We do not setflush the chain unless the deletion is
5087 * permanent, since the deletion of a chain does not actually
5088 * require it to be flushed.
5089 */
5094 }
5095 }
5098 }
5100 static int
5104 {
5109 /*
5110 * Nothing to do if already marked.
5111 *
5112 * We need the spinlock on the core whos RBTREE contains chain
5113 * to protect against races.
5114 */
5121 }
5123 /*
5124 * Permanent deletions mark the chain as destroyed.
5125 *
5126 * NOTE: We do not setflush the chain unless the deletion is
5127 * permanent, since the deletion of a chain does not actually
5128 * require it to be flushed.
5129 */
5134 }
5135 }
5138 }
5140 /*
5141 * Returns the index of the nearest element in the blockref array >= elm.
5142 * Returns (count) if no element could be found.
5143 *
5144 * Sets *key_nextp to the next key for loop purposes but does not modify
5145 * it if the next key would be higher than the current value of *key_nextp.
5146 * Note that *key_nexp can overflow to 0, which should be tested by the
5147 * caller.
5148 *
5149 * WARNING! Must be called with parent's spinlock held. Spinlock remains
5150 * held through the operation.
5151 */
5152 static int
5157 {
5159 hammer2_key_t scan_end;
5163 /*
5164 * Require the live chain's already have their core's counted
5165 * so we can optimize operations.
5166 */
5169 /*
5170 * Degenerate case
5171 */
5175 /*
5176 * Sequential optimization using parent->cache_index. This is
5177 * the most likely scenario.
5178 *
5179 * We can avoid trailing empty entries on live chains, otherwise
5180 * we might have to check the whole block array.
5181 */
5191 /*
5192 * Search backwards
5193 */
5199 }
5202 /*
5203 * Search forwards, stop when we find a scan element which
5204 * encloses the key or until we know that there are no further
5205 * elements.
5206 */
5213 }
5218 }
5229 }
5230 }
5231 }
5233 }
5235 /*
5236 * Do a combined search and return the next match either from the blockref
5237 * array or from the in-memory chain. Sets *brefp to the returned bref in
5238 * both cases, or sets it to NULL if the search exhausted. Only returns
5239 * a non-NULL chain if the search matched from the in-memory chain.
5240 *
5241 * When no in-memory chain has been found and a non-NULL bref is returned
5242 * in *brefp.
5243 *
5244 *
5245 * The returned chain is not locked or referenced. Use the returned bref
5246 * to determine if the search exhausted or not. Iterate if the base find
5247 * is chosen but matches a deleted chain.
5248 *
5249 * WARNING! Must be called with parent's spinlock held. Spinlock remains
5250 * held through the operation.
5251 */
5258 {
5263 /*
5264 * Lookup in block array and in rbtree.
5265 */
5271 /*
5272 * Neither matched
5273 */
5277 }
5279 /*
5280 * Only chain matched.
5281 */
5285 }
5287 /*
5288 * Only blockref matched.
5289 */
5293 }
5295 /*
5296 * Both in-memory and blockref matched, select the nearer element.
5297 *
5298 * If both are flush with the left-hand side or both are the
5299 * same distance away, select the chain. In this situation the
5300 * chain must have been loaded from the matching blockmap.
5301 */
5307 }
5309 /*
5310 * Select the nearer key
5311 */
5317 }
5319 /*
5320 * If the bref is out of bounds we've exhausted our search.
5321 */
5322 found:
5328 }
5330 }
5332 /*
5333 * Locate the specified block array element and delete it. The element
5334 * must exist.
5335 *
5336 * The spin lock on the related chain must be held.
5337 *
5338 * NOTE: live_count was adjusted when the chain was deleted, so it does not
5339 * need to be adjusted when we commit the media change.
5340 */
5341 void
5346 {
5349 hammer2_key_t key_next;
5352 /*
5353 * Delete element. Expect the element to exist.
5354 *
5355 * XXX see caller, flush code not yet sophisticated enough to prevent
5356 * re-flushed in some cases.
5357 */
5370 }
5372 /*
5373 * Update stats and zero the entry.
5374 *
5375 * NOTE: Handle radix == 0 (0 bytes) case.
5376 */
5380 }
5384 /* fall through */
5388 HAMMER2_CHAIN_HINT_LEAF_COUNT);
5392 }
5393 /* fall through */
5400 }
5405 HAMMER2_CHAIN_HINT_LEAF_COUNT);
5409 else
5411 }
5416 HAMMER2_CHAIN_HINT_LEAF_COUNT);
5420 }
5423 }
5429 /*
5430 * We can only optimize parent->core.live_zero for live chains.
5431 */
5434 ;
5436 }
5438 /*
5439 * Clear appropriate blockmap flags in chain.
5440 */
5442 HAMMER2_CHAIN_BLKMAPUPD);
5443 }
5445 /*
5446 * Insert the specified element. The block array must not already have the
5447 * element and must have space available for the insertion.
5448 *
5449 * The spin lock on the related chain must be held.
5450 *
5451 * NOTE: live_count was adjusted when the chain was deleted, so it does not
5452 * need to be adjusted when we commit the media change.
5453 */
5454 void
5458 {
5459 hammer2_key_t key_next;
5460 hammer2_key_t xkey;
5467 /*
5468 * Insert new element. Expect the element to not already exist
5469 * unless we are replacing it.
5470 *
5471 * XXX see caller, flush code not yet sophisticated enough to prevent
5472 * re-flushed in some cases.
5473 */
5478 /*
5479 * Shortcut fill optimization, typical ordered insertion(s) may not
5480 * require a search.
5481 */
5484 /*
5485 * Set appropriate blockmap flags in chain (if not NULL)
5486 */
5490 /*
5491 * Update stats and zero the entry
5492 */
5496 }
5500 /* fall through */
5504 /* fall through */
5511 }
5515 HAMMER2_BLOCKREF_LEAF_MAX) {
5519 }
5527 }
5530 /*
5531 * We can only optimize parent->core.live_zero for live chains.
5532 */
5537 }
5544 }
5546 /*
5547 * Try to find an empty slot before or after.
5548 */
5560 }
5562 }
5569 /*
5570 * We can only update parent->core.live_zero for live
5571 * chains.
5572 */
5577 }
5578 }
5581 /*
5582 * Debugging
5583 */
5584 validate:
5591 }
5592 }
5600 }
5601 }
5603 }
5605 #if 0
5607 /*
5608 * Sort the blockref array for the chain. Used by the flush code to
5609 * sort the blockref[] array.
5610 *
5611 * The chain must be exclusively locked AND spin-locked.
5612 */
5615 static
5616 int
5618 {
5622 /*
5623 * Make sure empty elements are placed at the end of the array
5624 */
5631 }
5633 /*
5634 * Sort by key
5635 */
5641 }
5643 void
5645 {
5651 /*
5652 * Special shortcut for embedded data returns the inode
5653 * itself. Callers must detect this condition and access
5654 * the embedded data (the strategy code does this for us).
5655 *
5656 * This is only applicable to regular files and softlinks.
5657 */
5659 HAMMER2_OPFLAG_DIRECTDATA) {
5661 }
5667 /*
5668 * Optimize indirect blocks in the INITIAL state to avoid
5669 * I/O.
5670 */
5690 }
5692 }
5694 #endif
5696 /*
5697 * Set the check data for a chain. This can be a heavy-weight operation
5698 * and typically only runs on-flush. For file data check data is calculated
5699 * when the logical buffers are flushed.
5700 */
5701 void
5703 {
5721 /*
5722 {
5723 SHA256_CTX hash_ctx;
5724 union {
5725 uint8_t digest[SHA256_DIGEST_LENGTH];
5726 uint64_t digest64[SHA256_DIGEST_LENGTH/8];
5727 } u;
5729 SHA256_Init(&hash_ctx);
5730 SHA256_Update(&hash_ctx, bdata, chain->bytes);
5731 SHA256_Final(u.digest, &hash_ctx);
5732 u.digest64[2] ^= u.digest64[3];
5733 bcopy(u.digest,
5734 chain->bref.check.sha192.data,
5735 sizeof(chain->bref.check.sha192.data));
5736 }
5737 */
5747 }
5748 }
5750 /*
5751 * Characterize a failed check code and try to trace back to the inode.
5752 */
5753 static void
5756 {
5762 "chain %016jx.%02x (%s) meth=%02x CHECK FAIL "
5779 check);
5780 }
5782 /*
5783 * Run up the chains to try to find the governing inode so we
5784 * can report it.
5785 *
5786 * XXX This error reporting is not really MPSAFE
5787 */
5793 }
5804 }
5815 }
5816 }
5817 }
5820 }
5821 }
5823 /*
5824 * Returns non-zero on success, 0 on failure.
5825 */
5826 int
5828 {
5848 }
5856 }
5861 /*
5862 {
5863 SHA256_CTX hash_ctx;
5864 union {
5865 uint8_t digest[SHA256_DIGEST_LENGTH];
5866 uint64_t digest64[SHA256_DIGEST_LENGTH/8];
5867 } u;
5869 SHA256_Init(&hash_ctx);
5870 SHA256_Update(&hash_ctx, bdata, chain->bytes);
5871 SHA256_Final(u.digest, &hash_ctx);
5872 u.digest64[2] ^= u.digest64[3];
5873 if (bcmp(u.digest,
5874 chain->bref.check.sha192.data,
5875 sizeof(chain->bref.check.sha192.data)) == 0) {
5876 r = 1;
5877 } else {
5878 r = 0;
5879 krateprintf(&krate_h2chk,
5880 "chain %016jx.%02x meth=%02x "
5881 "CHECK FAIL\n",
5882 chain->bref.data_off,
5883 chain->bref.type,
5884 chain->bref.methods);
5885 }
5886 }
5887 */
5896 "chain %016jx.%02x meth=%02x "
5912 bdata,
5914 }
5915 }
5916 }
5923 }
5925 }
5927 /*
5928 * Acquire the chain and parent representing the specified inode for the
5929 * device at the specified cluster index.
5930 *
5931 * The flags passed in are LOOKUP flags, not RESOLVE flags.
5932 *
5933 * If we are unable to locate the inode, HAMMER2_ERROR_EIO or HAMMER2_ERROR_CHECK
5934 * is returned. In case of error, *chainp and/or *parentp may still be returned
5935 * non-NULL.
5936 *
5937 * The caller may pass-in a locked *parentp and/or *chainp, or neither.
5938 * They will be unlocked and released by this function. The *parentp and
5939 * *chainp representing the located inode are returned locked.
5940 *
5941 * The returned error includes any error on the returned chain in addition to
5942 * errors incurred while trying to lookup the inode. However, a chain->error
5943 * might not be recognized if HAMMER2_LOOKUP_NODATA is passed. This flag may
5944 * not be passed to this function.
5945 */
5946 int
5950 {
5953 hammer2_key_t key_dummy;
5963 /*
5964 * Caller expects us to replace these.
5965 */
5970 }
5975 }
5977 /*
5978 * Be very careful, this is a backend function and we CANNOT
5979 * lock any frontend inode structure we find. But we have to
5980 * look the inode up this way first in case it exists but is
5981 * detached from the radix tree.
5982 */
5986 parentp,
5987 resolve_flags);
5998 }
5999 }
6001 /*
6002 * Inodes hang off of the iroot (bit 63 is clear, differentiating
6003 * inodes from root directory entries in the key lookup).
6004 */
6008 /*
6009 * NOTE: rchain can be returned as NULL even if error == 0
6010 * (i.e. not found)
6011 */
6015 /*
6016 * Propagate a chain-specific error to caller.
6017 *
6018 * If the chain is not errored, we must still validate that the inode
6019 * number is correct, because all hell will break loose if it isn't
6020 * correct. It should always be correct so print to the console and
6021 * simulate a CHECK error if it is not.
6022 */
6032 }
6033 }
6034 }
6037 }
6042 }
6044 /*
6045 * Used by the bulkscan code to snapshot the synchronized storage for
6046 * a volume, allowing it to be scanned concurrently against normal
6047 * operation.
6048 */
6049 hammer2_chain_t *
6051 {
6062 }
6064 void
6066 {
6072 }
6074 /*
6075 * Returns non-zero if the chain (INODE or DIRENT) matches the
6076 * filename.
6077 */
6078 int
6081 {
6089 }
6090 }
6096 }
6100 }
6101 }
6103 }
6105 /*
6106 * Debugging
6107 */
6108 void
6111 {
6119 }
6149 }
6150 bi++;
6151 }
6155 else
6157 }
6158 }
6160 void
6162 {
6170 }