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>
64
65 #include <crypto/sha2/sha2.h>
66 */
67
68 #include "hammer2.h"
69
70 static hammer2_chain_t *hammer2_chain_create_indirect(
71 hammer2_chain_t *parent,
72 hammer2_key_t key, int keybits,
73 hammer2_tid_t mtid, int for_type, int *errorp);
74 static int hammer2_chain_delete_obref(hammer2_chain_t *parent,
75 hammer2_chain_t *chain,
76 hammer2_tid_t mtid, int flags,
77 hammer2_blockref_t *obref);
78 static hammer2_chain_t *hammer2_combined_find(
79 hammer2_chain_t *parent,
80 hammer2_blockref_t *base, int count,
81 hammer2_key_t *key_nextp,
82 hammer2_key_t key_beg, hammer2_key_t key_end,
83 hammer2_blockref_t **brefp);
84 static hammer2_chain_t *hammer2_chain_lastdrop(hammer2_chain_t *chain,
85 int depth);
86 /*
87 * There are many degenerate situations where an extreme rate of console
88 * output can occur from warnings and errors. Make sure this output does
89 * not impede operations.
90 */
91 /*
92 static struct krate krate_h2chk = { .freq = 5 };
93 static struct krate krate_h2me = { .freq = 1 };
94 static struct krate krate_h2em = { .freq = 1 };
95 */
96
97 /*
98 * Basic RBTree for chains (core.rbtree).
99 */
100 RB_GENERATE(hammer2_chain_tree, hammer2_chain, rbnode, hammer2_chain_cmp);
101
102 int
hammer2_chain_cmp(hammer2_chain_t * chain1,hammer2_chain_t * chain2)103 hammer2_chain_cmp(hammer2_chain_t *chain1, hammer2_chain_t *chain2)
104 {
105 hammer2_key_t c1_beg;
106 hammer2_key_t c1_end;
107 hammer2_key_t c2_beg;
108 hammer2_key_t c2_end;
109
110 /*
111 * Compare chains. Overlaps are not supposed to happen and catch
112 * any software issues early we count overlaps as a match.
113 */
114 c1_beg = chain1->bref.key;
115 c1_end = c1_beg + ((hammer2_key_t)1 << chain1->bref.keybits) - 1;
116 c2_beg = chain2->bref.key;
117 c2_end = c2_beg + ((hammer2_key_t)1 << chain2->bref.keybits) - 1;
118
119 if (c1_end < c2_beg) /* fully to the left */
120 return(-1);
121 if (c1_beg > c2_end) /* fully to the right */
122 return(1);
123 return(0); /* overlap (must not cross edge boundary) */
124 }
125
126 /*
127 * Assert that a chain has no media data associated with it.
128 */
129 static __inline void
hammer2_chain_assert_no_data(hammer2_chain_t * chain)130 hammer2_chain_assert_no_data(hammer2_chain_t *chain)
131 {
132 KKASSERT(chain->dio == NULL);
133 if (chain->bref.type != HAMMER2_BREF_TYPE_VOLUME &&
134 chain->bref.type != HAMMER2_BREF_TYPE_FREEMAP &&
135 chain->data) {
136 panic("hammer2_chain_assert_no_data: chain %p still has data",
137 chain);
138 }
139 }
140
141 /*
142 * Make a chain visible to the flusher. The flusher operates using a top-down
143 * recursion based on the ONFLUSH flag. It locates MODIFIED and UPDATE chains,
144 * flushes them, and updates blocks back to the volume root.
145 *
146 * This routine sets the ONFLUSH flag upward from the triggering chain until
147 * it hits an inode root or the volume root. Inode chains serve as inflection
148 * points, requiring the flusher to bridge across trees. Inodes include
149 * regular inodes, PFS roots (pmp->iroot), and the media super root
150 * (spmp->iroot).
151 */
152 void
hammer2_chain_setflush(hammer2_chain_t * chain)153 hammer2_chain_setflush(hammer2_chain_t *chain)
154 {
155 hammer2_chain_t *parent;
156
157 if ((chain->flags & HAMMER2_CHAIN_ONFLUSH) == 0) {
158 hammer2_spin_sh(&chain->core.spin);
159 while ((chain->flags & HAMMER2_CHAIN_ONFLUSH) == 0) {
160 atomic_set_int(&chain->flags, HAMMER2_CHAIN_ONFLUSH);
161 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE)
162 break;
163 if ((parent = chain->parent) == NULL)
164 break;
165 hammer2_spin_sh(&parent->core.spin);
166 hammer2_spin_unsh(&chain->core.spin);
167 chain = parent;
168 }
169 hammer2_spin_unsh(&chain->core.spin);
170 }
171 }
172
173 /*
174 * Allocate a new disconnected chain element representing the specified
175 * bref. chain->refs is set to 1 and the passed bref is copied to
176 * chain->bref. chain->bytes is derived from the bref.
177 *
178 * chain->pmp inherits pmp unless the chain is an inode (other than the
179 * super-root inode).
180 *
181 * NOTE: Returns a referenced but unlocked (because there is no core) chain.
182 */
183 hammer2_chain_t *
hammer2_chain_alloc(hammer2_dev_t * hmp,hammer2_pfs_t * pmp,hammer2_blockref_t * bref)184 hammer2_chain_alloc(hammer2_dev_t *hmp, hammer2_pfs_t *pmp,
185 hammer2_blockref_t *bref)
186 {
187 hammer2_chain_t *chain;
188 u_int bytes;
189
190 /*
191 * Special case - radix of 0 indicates a chain that does not
192 * need a data reference (context is completely embedded in the
193 * bref).
194 */
195 if ((int)(bref->data_off & HAMMER2_OFF_MASK_RADIX))
196 bytes = 1U << (int)(bref->data_off & HAMMER2_OFF_MASK_RADIX);
197 else
198 bytes = 0;
199
200 switch(bref->type) {
201 case HAMMER2_BREF_TYPE_INODE:
202 case HAMMER2_BREF_TYPE_INDIRECT:
203 case HAMMER2_BREF_TYPE_DATA:
204 case HAMMER2_BREF_TYPE_DIRENT:
205 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
206 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
207 case HAMMER2_BREF_TYPE_FREEMAP:
208 case HAMMER2_BREF_TYPE_VOLUME:
209 chain = kmalloc_obj(sizeof(*chain), hmp->mchain,
210 M_WAITOK | M_ZERO);
211 atomic_add_long(&hammer2_chain_allocs, 1);
212 break;
213 case HAMMER2_BREF_TYPE_EMPTY:
214 default:
215 panic("hammer2_chain_alloc: unrecognized blockref type: %d",
216 bref->type);
217 break;
218 }
219
220 /*
221 * Initialize the new chain structure. pmp must be set to NULL for
222 * chains belonging to the super-root topology of a device mount.
223 */
224 if (pmp == hmp->spmp)
225 chain->pmp = NULL;
226 else
227 chain->pmp = pmp;
228
229 chain->hmp = hmp;
230 chain->bref = *bref;
231 chain->bytes = bytes;
232 chain->refs = 1;
233 chain->flags = HAMMER2_CHAIN_ALLOCATED;
234
235 /*
236 * Set the PFS boundary flag if this chain represents a PFS root.
237 */
238 if (bref->flags & HAMMER2_BREF_FLAG_PFSROOT)
239 atomic_set_int(&chain->flags, HAMMER2_CHAIN_PFSBOUNDARY);
240 hammer2_chain_init(chain);
241
242 return (chain);
243 }
244
245 /*
246 * A common function to initialize chains including fchain and vchain.
247 */
248 void
hammer2_chain_init(hammer2_chain_t * chain)249 hammer2_chain_init(hammer2_chain_t *chain)
250 {
251 RB_INIT(&chain->core.rbtree); /* live chains */
252 hammer2_mtx_init(&chain->lock, "h2chain");
253 hammer2_spin_init(&chain->core.spin, "h2chain");
254 lockinit(&chain->diolk, "chdio", 0, 0);
255 }
256
257 /*
258 * Add a reference to a chain element, preventing its destruction.
259 * Undone via hammer2_chain_drop()
260 *
261 * (can be called with spinlock held)
262 */
263 void
hammer2_chain_ref(hammer2_chain_t * chain)264 hammer2_chain_ref(hammer2_chain_t *chain)
265 {
266 if (atomic_fetchadd_int(&chain->refs, 1) == 0) {
267 /* NOP */
268 }
269 }
270
271 /*
272 * Ref a locked chain and force the data to be held across an unlock.
273 * Chain must be currently locked. The user of the chain who desires
274 * to release the hold must call hammer2_chain_lock_unhold() to relock
275 * and unhold the chain, then unlock normally, or may simply call
276 * hammer2_chain_drop_unhold() (which is safer against deadlocks).
277 */
278 void
hammer2_chain_ref_hold(hammer2_chain_t * chain)279 hammer2_chain_ref_hold(hammer2_chain_t *chain)
280 {
281 atomic_add_int(&chain->lockcnt, 1);
282 hammer2_chain_ref(chain);
283 }
284
285 /*
286 * Insert the chain in the core rbtree.
287 *
288 * Normal insertions are placed in the live rbtree. Insertion of a deleted
289 * chain is a special case used by the flush code that is placed on the
290 * unstaged deleted list to avoid confusing the live view.
291 */
292 #define HAMMER2_CHAIN_INSERT_SPIN 0x0001
293 #define HAMMER2_CHAIN_INSERT_LIVE 0x0002
294 #define HAMMER2_CHAIN_INSERT_RACE 0x0004
295
296 static
297 int
hammer2_chain_insert(hammer2_chain_t * parent,hammer2_chain_t * chain,int flags,int generation)298 hammer2_chain_insert(hammer2_chain_t *parent, hammer2_chain_t *chain,
299 int flags, int generation)
300 {
301 hammer2_chain_t *xchain __debugvar;
302 int error = 0;
303
304 if (flags & HAMMER2_CHAIN_INSERT_SPIN)
305 hammer2_spin_ex(&parent->core.spin);
306
307 /*
308 * Interlocked by spinlock, check for race
309 */
310 if ((flags & HAMMER2_CHAIN_INSERT_RACE) &&
311 parent->core.generation != generation) {
312 error = HAMMER2_ERROR_EAGAIN;
313 goto failed;
314 }
315
316 /*
317 * Insert chain
318 */
319 xchain = RB_INSERT(hammer2_chain_tree, &parent->core.rbtree, chain);
320 KASSERT(xchain == NULL,
321 ("hammer2_chain_insert: collision %p %p (key=%016jx)",
322 chain, xchain, chain->bref.key));
323 atomic_set_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE);
324 chain->parent = parent;
325 ++parent->core.chain_count;
326 ++parent->core.generation; /* XXX incs for _get() too, XXX */
327
328 /*
329 * We have to keep track of the effective live-view blockref count
330 * so the create code knows when to push an indirect block.
331 */
332 if (flags & HAMMER2_CHAIN_INSERT_LIVE)
333 atomic_add_int(&parent->core.live_count, 1);
334 failed:
335 if (flags & HAMMER2_CHAIN_INSERT_SPIN)
336 hammer2_spin_unex(&parent->core.spin);
337 return error;
338 }
339
340 /*
341 * Drop the caller's reference to the chain. When the ref count drops to
342 * zero this function will try to disassociate the chain from its parent and
343 * deallocate it, then recursely drop the parent using the implied ref
344 * from the chain's chain->parent.
345 *
346 * Nobody should own chain's mutex on the 1->0 transition, unless this drop
347 * races an acquisition by another cpu. Therefore we can loop if we are
348 * unable to acquire the mutex, and refs is unlikely to be 1 unless we again
349 * race against another drop.
350 */
351 void
hammer2_chain_drop(hammer2_chain_t * chain)352 hammer2_chain_drop(hammer2_chain_t *chain)
353 {
354 u_int refs;
355
356 KKASSERT(chain->refs > 0);
357
358 while (chain) {
359 refs = chain->refs;
360 cpu_ccfence();
361 KKASSERT(refs > 0);
362
363 if (refs == 1) {
364 if (hammer2_mtx_ex_try(&chain->lock) == 0)
365 chain = hammer2_chain_lastdrop(chain, 0);
366 /* retry the same chain, or chain from lastdrop */
367 } else {
368 if (atomic_cmpset_int(&chain->refs, refs, refs - 1))
369 break;
370 /* retry the same chain */
371 }
372 cpu_pause();
373 }
374 }
375
376 /*
377 * Unhold a held and probably not-locked chain, ensure that the data is
378 * dropped on the 1->0 transition of lockcnt by obtaining an exclusive
379 * lock and then simply unlocking the chain.
380 */
381 void
hammer2_chain_unhold(hammer2_chain_t * chain)382 hammer2_chain_unhold(hammer2_chain_t *chain)
383 {
384 u_int lockcnt;
385 int iter = 0;
386
387 for (;;) {
388 lockcnt = chain->lockcnt;
389 cpu_ccfence();
390 if (lockcnt > 1) {
391 if (atomic_cmpset_int(&chain->lockcnt,
392 lockcnt, lockcnt - 1)) {
393 break;
394 }
395 } else if (hammer2_mtx_ex_try(&chain->lock) == 0) {
396 hammer2_chain_unlock(chain);
397 break;
398 } else {
399 /*
400 * This situation can easily occur on SMP due to
401 * the gap inbetween the 1->0 transition and the
402 * final unlock. We cannot safely block on the
403 * mutex because lockcnt might go above 1.
404 *
405 * XXX Sleep for one tick if it takes too long.
406 */
407 if (++iter > 1000) {
408 if (iter > 1000 + hz) {
409 kprintf("hammer2: h2race1 %p\n", chain);
410 iter = 1000;
411 }
412 tsleep(&iter, 0, "h2race1", 1);
413 }
414 cpu_pause();
415 }
416 }
417 }
418
419 void
hammer2_chain_drop_unhold(hammer2_chain_t * chain)420 hammer2_chain_drop_unhold(hammer2_chain_t *chain)
421 {
422 hammer2_chain_unhold(chain);
423 hammer2_chain_drop(chain);
424 }
425
426 void
hammer2_chain_rehold(hammer2_chain_t * chain)427 hammer2_chain_rehold(hammer2_chain_t *chain)
428 {
429 hammer2_chain_lock(chain, HAMMER2_RESOLVE_SHARED);
430 atomic_add_int(&chain->lockcnt, 1);
431 hammer2_chain_unlock(chain);
432 }
433
434 /*
435 * Handles the (potential) last drop of chain->refs from 1->0. Called with
436 * the mutex exclusively locked, refs == 1, and lockcnt 0. SMP races are
437 * possible against refs and lockcnt. We must dispose of the mutex on chain.
438 *
439 * This function returns an unlocked chain for recursive drop or NULL. It
440 * can return the same chain if it determines it has raced another ref.
441 *
442 * --
443 *
444 * When two chains need to be recursively dropped we use the chain we
445 * would otherwise free to placehold the additional chain. It's a bit
446 * convoluted but we can't just recurse without potentially blowing out
447 * the kernel stack.
448 *
449 * The chain cannot be freed if it has any children.
450 * The chain cannot be freed if flagged MODIFIED unless we can dispose of it.
451 * The chain cannot be freed if flagged UPDATE unless we can dispose of it.
452 * Any dedup registration can remain intact.
453 *
454 * The core spinlock is allowed to nest child-to-parent (not parent-to-child).
455 */
456 static
457 hammer2_chain_t *
hammer2_chain_lastdrop(hammer2_chain_t * chain,int depth)458 hammer2_chain_lastdrop(hammer2_chain_t *chain, int depth)
459 {
460 hammer2_dev_t *hmp;
461 hammer2_chain_t *parent;
462 hammer2_chain_t *rdrop;
463
464 /*
465 * We need chain's spinlock to interlock the sub-tree test.
466 * We already have chain's mutex, protecting chain->parent.
467 *
468 * Remember that chain->refs can be in flux.
469 */
470 hammer2_spin_ex(&chain->core.spin);
471
472 if (chain->parent != NULL) {
473 /*
474 * If the chain has a parent the UPDATE bit prevents scrapping
475 * as the chain is needed to properly flush the parent. Try
476 * to complete the 1->0 transition and return NULL. Retry
477 * (return chain) if we are unable to complete the 1->0
478 * transition, else return NULL (nothing more to do).
479 *
480 * If the chain has a parent the MODIFIED bit prevents
481 * scrapping.
482 */
483 if (chain->flags & (HAMMER2_CHAIN_UPDATE |
484 HAMMER2_CHAIN_MODIFIED)) {
485 if (atomic_cmpset_int(&chain->refs, 1, 0)) {
486 hammer2_spin_unex(&chain->core.spin);
487 hammer2_chain_assert_no_data(chain);
488 hammer2_mtx_unlock(&chain->lock);
489 chain = NULL;
490 } else {
491 hammer2_spin_unex(&chain->core.spin);
492 hammer2_mtx_unlock(&chain->lock);
493 }
494 return (chain);
495 }
496 /* spinlock still held */
497 } else if (chain->bref.type == HAMMER2_BREF_TYPE_VOLUME ||
498 chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP) {
499 /*
500 * Retain the static vchain and fchain. Clear bits that
501 * are not relevant. Do not clear the MODIFIED bit,
502 * and certainly do not put it on the delayed-flush queue.
503 */
504 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
505 } else {
506 /*
507 * The chain has no parent and can be flagged for destruction.
508 * Since it has no parent, UPDATE can also be cleared.
509 */
510 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DESTROY);
511 if (chain->flags & HAMMER2_CHAIN_UPDATE)
512 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
513
514 /*
515 * If the chain has children we must propagate the DESTROY
516 * flag downward and rip the disconnected topology apart.
517 * This is accomplished by calling hammer2_flush() on the
518 * chain.
519 *
520 * Any dedup is already handled by the underlying DIO, so
521 * we do not have to specifically flush it here.
522 */
523 if (chain->core.chain_count) {
524 hammer2_spin_unex(&chain->core.spin);
525 hammer2_flush(chain, HAMMER2_FLUSH_TOP |
526 HAMMER2_FLUSH_ALL);
527 hammer2_mtx_unlock(&chain->lock);
528
529 return(chain); /* retry drop */
530 }
531
532 /*
533 * Otherwise we can scrap the MODIFIED bit if it is set,
534 * and continue along the freeing path.
535 *
536 * Be sure to clean-out any dedup bits. Without a parent
537 * this chain will no longer be visible to the flush code.
538 * Easy check data_off to avoid the volume root.
539 */
540 if (chain->flags & HAMMER2_CHAIN_MODIFIED) {
541 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
542 atomic_add_long(&hammer2_count_modified_chains, -1);
543 if (chain->pmp)
544 hammer2_pfs_memory_wakeup(chain->pmp, -1);
545 }
546 /* spinlock still held */
547 }
548
549 /* spinlock still held */
550
551 /*
552 * If any children exist we must leave the chain intact with refs == 0.
553 * They exist because chains are retained below us which have refs or
554 * may require flushing.
555 *
556 * Retry (return chain) if we fail to transition the refs to 0, else
557 * return NULL indication nothing more to do.
558 */
559 if (chain->core.chain_count) {
560 if (atomic_cmpset_int(&chain->refs, 1, 0)) {
561 hammer2_spin_unex(&chain->core.spin);
562 hammer2_chain_assert_no_data(chain);
563 hammer2_mtx_unlock(&chain->lock);
564 chain = NULL;
565 } else {
566 hammer2_spin_unex(&chain->core.spin);
567 hammer2_mtx_unlock(&chain->lock);
568 }
569 return (chain);
570 }
571 /* spinlock still held */
572 /* no chains left under us */
573
574 /*
575 * chain->core has no children left so no accessors can get to our
576 * chain from there. Now we have to lock the parent core to interlock
577 * remaining possible accessors that might bump chain's refs before
578 * we can safely drop chain's refs with intent to free the chain.
579 */
580 hmp = chain->hmp;
581 rdrop = NULL;
582
583 parent = chain->parent;
584
585 /*
586 * WARNING! chain's spin lock is still held here, and other spinlocks
587 * will be acquired and released in the code below. We
588 * cannot be making fancy procedure calls!
589 */
590
591 /*
592 * Spinlock the parent and try to drop the last ref on chain.
593 * On success determine if we should dispose of the chain
594 * (remove the chain from its parent, etc).
595 *
596 * (normal core locks are top-down recursive but we define
597 * core spinlocks as bottom-up recursive, so this is safe).
598 */
599 if (parent) {
600 hammer2_spin_ex(&parent->core.spin);
601 if (atomic_cmpset_int(&chain->refs, 1, 0) == 0) {
602 /*
603 * 1->0 transition failed, retry.
604 */
605 hammer2_spin_unex(&parent->core.spin);
606 hammer2_spin_unex(&chain->core.spin);
607 hammer2_mtx_unlock(&chain->lock);
608
609 return(chain);
610 }
611
612 /*
613 * 1->0 transition successful, parent spin held to prevent
614 * new lookups, chain spinlock held to protect parent field.
615 * Remove chain from the parent.
616 *
617 * If the chain is being removed from the parent's rbtree but
618 * is not blkmapped, we have to adjust live_count downward. If
619 * it is blkmapped then the blockref is retained in the parent
620 * as is its associated live_count. This case can occur when
621 * a chain added to the topology is unable to flush and is
622 * then later deleted.
623 */
624 if (chain->flags & HAMMER2_CHAIN_ONRBTREE) {
625 if ((parent->flags & HAMMER2_CHAIN_COUNTEDBREFS) &&
626 (chain->flags & HAMMER2_CHAIN_BLKMAPPED) == 0) {
627 atomic_add_int(&parent->core.live_count, -1);
628 }
629 RB_REMOVE(hammer2_chain_tree,
630 &parent->core.rbtree, chain);
631 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE);
632 --parent->core.chain_count;
633 chain->parent = NULL;
634 }
635
636 /*
637 * If our chain was the last chain in the parent's core the
638 * core is now empty and its parent might have to be
639 * re-dropped if it has 0 refs.
640 */
641 if (parent->core.chain_count == 0) {
642 rdrop = parent;
643 atomic_add_int(&rdrop->refs, 1);
644 /*
645 if (atomic_cmpset_int(&rdrop->refs, 0, 1) == 0)
646 rdrop = NULL;
647 */
648 }
649 hammer2_spin_unex(&parent->core.spin);
650 parent = NULL; /* safety */
651 /* FALL THROUGH */
652 } else {
653 /*
654 * No-parent case.
655 */
656 if (atomic_cmpset_int(&chain->refs, 1, 0) == 0) {
657 /*
658 * 1->0 transition failed, retry.
659 */
660 hammer2_spin_unex(&parent->core.spin);
661 hammer2_spin_unex(&chain->core.spin);
662 hammer2_mtx_unlock(&chain->lock);
663
664 return(chain);
665 }
666 }
667
668 /*
669 * Successful 1->0 transition, no parent, no children... no way for
670 * anyone to ref this chain any more. We can clean-up and free it.
671 *
672 * We still have the core spinlock, and core's chain_count is 0.
673 * Any parent spinlock is gone.
674 */
675 hammer2_spin_unex(&chain->core.spin);
676 hammer2_chain_assert_no_data(chain);
677 hammer2_mtx_unlock(&chain->lock);
678 KKASSERT(RB_EMPTY(&chain->core.rbtree) &&
679 chain->core.chain_count == 0);
680
681 /*
682 * All locks are gone, no pointers remain to the chain, finish
683 * freeing it.
684 */
685 KKASSERT((chain->flags & (HAMMER2_CHAIN_UPDATE |
686 HAMMER2_CHAIN_MODIFIED)) == 0);
687
688 /*
689 * Once chain resources are gone we can use the now dead chain
690 * structure to placehold what might otherwise require a recursive
691 * drop, because we have potentially two things to drop and can only
692 * return one directly.
693 */
694 if (chain->flags & HAMMER2_CHAIN_ALLOCATED) {
695 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ALLOCATED);
696 chain->hmp = NULL;
697 kfree_obj(chain, hmp->mchain);
698 atomic_add_long(&hammer2_chain_allocs, -1);
699 }
700
701 /*
702 * Possible chaining loop when parent re-drop needed.
703 */
704 return(rdrop);
705 }
706
707 /*
708 * On last lock release.
709 */
710 static hammer2_io_t *
hammer2_chain_drop_data(hammer2_chain_t * chain)711 hammer2_chain_drop_data(hammer2_chain_t *chain)
712 {
713 hammer2_io_t *dio;
714
715 if ((dio = chain->dio) != NULL) {
716 chain->dio = NULL;
717 chain->data = NULL;
718 } else {
719 switch(chain->bref.type) {
720 case HAMMER2_BREF_TYPE_VOLUME:
721 case HAMMER2_BREF_TYPE_FREEMAP:
722 break;
723 default:
724 if (chain->data != NULL) {
725 hammer2_spin_unex(&chain->core.spin);
726 panic("chain data not null: "
727 "chain %p bref %016jx.%02x "
728 "refs %d parent %p dio %p data %p",
729 chain, chain->bref.data_off,
730 chain->bref.type, chain->refs,
731 chain->parent,
732 chain->dio, chain->data);
733 }
734 KKASSERT(chain->data == NULL);
735 break;
736 }
737 }
738 return dio;
739 }
740
741 /*
742 * Lock a referenced chain element, acquiring its data with I/O if necessary,
743 * and specify how you would like the data to be resolved.
744 *
745 * If an I/O or other fatal error occurs, chain->error will be set to non-zero.
746 *
747 * The lock is allowed to recurse, multiple locking ops will aggregate
748 * the requested resolve types. Once data is assigned it will not be
749 * removed until the last unlock.
750 *
751 * HAMMER2_RESOLVE_NEVER - Do not resolve the data element.
752 * (typically used to avoid device/logical buffer
753 * aliasing for data)
754 *
755 * HAMMER2_RESOLVE_MAYBE - Do not resolve data elements for chains in
756 * the INITIAL-create state (indirect blocks only).
757 *
758 * Do not resolve data elements for DATA chains.
759 * (typically used to avoid device/logical buffer
760 * aliasing for data)
761 *
762 * HAMMER2_RESOLVE_ALWAYS- Always resolve the data element.
763 *
764 * HAMMER2_RESOLVE_SHARED- (flag) The chain is locked shared, otherwise
765 * it will be locked exclusive.
766 *
767 * HAMMER2_RESOLVE_NONBLOCK- (flag) The chain is locked non-blocking. If
768 * the lock fails, EAGAIN is returned.
769 *
770 * NOTE: Embedded elements (volume header, inodes) are always resolved
771 * regardless.
772 *
773 * NOTE: Specifying HAMMER2_RESOLVE_ALWAYS on a newly-created non-embedded
774 * element will instantiate and zero its buffer, and flush it on
775 * release.
776 *
777 * NOTE: (data) elements are normally locked RESOLVE_NEVER or RESOLVE_MAYBE
778 * so as not to instantiate a device buffer, which could alias against
779 * a logical file buffer. However, if ALWAYS is specified the
780 * device buffer will be instantiated anyway.
781 *
782 * NOTE: The return value is always 0 unless NONBLOCK is specified, in which
783 * case it can be either 0 or EAGAIN.
784 *
785 * WARNING! This function blocks on I/O if data needs to be fetched. This
786 * blocking can run concurrent with other compatible lock holders
787 * who do not need data returning. The lock is not upgraded to
788 * exclusive during a data fetch, a separate bit is used to
789 * interlock I/O. However, an exclusive lock holder can still count
790 * on being interlocked against an I/O fetch managed by a shared
791 * lock holder.
792 */
793 int
hammer2_chain_lock(hammer2_chain_t * chain,int how)794 hammer2_chain_lock(hammer2_chain_t *chain, int how)
795 {
796 KKASSERT(chain->refs > 0);
797
798 if (how & HAMMER2_RESOLVE_NONBLOCK) {
799 /*
800 * We still have to bump lockcnt before acquiring the lock,
801 * even for non-blocking operation, because the unlock code
802 * live-loops on lockcnt == 1 when dropping the last lock.
803 *
804 * If the non-blocking operation fails we have to use an
805 * unhold sequence to undo the mess.
806 *
807 * NOTE: LOCKAGAIN must always succeed without blocking,
808 * even if NONBLOCK is specified.
809 */
810 atomic_add_int(&chain->lockcnt, 1);
811 if (how & HAMMER2_RESOLVE_SHARED) {
812 if (how & HAMMER2_RESOLVE_LOCKAGAIN) {
813 hammer2_mtx_sh_again(&chain->lock);
814 } else {
815 if (hammer2_mtx_sh_try(&chain->lock) != 0) {
816 hammer2_chain_unhold(chain);
817 return EAGAIN;
818 }
819 }
820 } else {
821 if (hammer2_mtx_ex_try(&chain->lock) != 0) {
822 hammer2_chain_unhold(chain);
823 return EAGAIN;
824 }
825 }
826 } else {
827 /*
828 * Get the appropriate lock. If LOCKAGAIN is flagged with
829 * SHARED the caller expects a shared lock to already be
830 * present and we are giving it another ref. This case must
831 * importantly not block if there is a pending exclusive lock
832 * request.
833 */
834 atomic_add_int(&chain->lockcnt, 1);
835 if (how & HAMMER2_RESOLVE_SHARED) {
836 if (how & HAMMER2_RESOLVE_LOCKAGAIN) {
837 hammer2_mtx_sh_again(&chain->lock);
838 } else {
839 hammer2_mtx_sh(&chain->lock);
840 }
841 } else {
842 hammer2_mtx_ex(&chain->lock);
843 }
844 }
845
846 /*
847 * If we already have a valid data pointer make sure the data is
848 * synchronized to the current cpu, and then no further action is
849 * necessary.
850 */
851 if (chain->data) {
852 if (chain->dio)
853 hammer2_io_bkvasync(chain->dio);
854 return 0;
855 }
856
857 /*
858 * Do we have to resolve the data? This is generally only
859 * applicable to HAMMER2_BREF_TYPE_DATA which is special-cased.
860 * Other BREF types expects the data to be there.
861 */
862 switch(how & HAMMER2_RESOLVE_MASK) {
863 case HAMMER2_RESOLVE_NEVER:
864 return 0;
865 case HAMMER2_RESOLVE_MAYBE:
866 if (chain->flags & HAMMER2_CHAIN_INITIAL)
867 return 0;
868 if (chain->bref.type == HAMMER2_BREF_TYPE_DATA)
869 return 0;
870 #if 0
871 if (chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE)
872 return 0;
873 if (chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_LEAF)
874 return 0;
875 #endif
876 /* fall through */
877 case HAMMER2_RESOLVE_ALWAYS:
878 default:
879 break;
880 }
881
882 /*
883 * Caller requires data
884 */
885 hammer2_chain_load_data(chain);
886
887 return 0;
888 }
889
890 #if 0
891 /*
892 * Lock the chain, retain the hold, and drop the data persistence count.
893 * The data should remain valid because we never transitioned lockcnt
894 * through 0.
895 */
896 void
897 hammer2_chain_lock_unhold(hammer2_chain_t *chain, int how)
898 {
899 hammer2_chain_lock(chain, how);
900 atomic_add_int(&chain->lockcnt, -1);
901 }
902
903 /*
904 * Downgrade an exclusive chain lock to a shared chain lock.
905 *
906 * NOTE: There is no upgrade equivalent due to the ease of
907 * deadlocks in that direction.
908 */
909 void
910 hammer2_chain_lock_downgrade(hammer2_chain_t *chain)
911 {
912 hammer2_mtx_downgrade(&chain->lock);
913 }
914 #endif
915
916 /*
917 * Issue I/O and install chain->data. Caller must hold a chain lock, lock
918 * may be of any type.
919 *
920 * Once chain->data is set it cannot be disposed of until all locks are
921 * released.
922 *
923 * Make sure the data is synchronized to the current cpu.
924 */
925 void
hammer2_chain_load_data(hammer2_chain_t * chain)926 hammer2_chain_load_data(hammer2_chain_t *chain)
927 {
928 hammer2_blockref_t *bref;
929 hammer2_dev_t *hmp;
930 hammer2_io_t *dio;
931 char *bdata;
932 int error;
933
934 /*
935 * Degenerate case, data already present, or chain has no media
936 * reference to load.
937 */
938 KKASSERT(chain->lock.mtx_lock & MTX_MASK);
939 if (chain->data) {
940 if (chain->dio)
941 hammer2_io_bkvasync(chain->dio);
942 return;
943 }
944 if ((chain->bref.data_off & ~HAMMER2_OFF_MASK_RADIX) == 0)
945 return;
946
947 hmp = chain->hmp;
948 KKASSERT(hmp != NULL);
949
950 /*
951 * Gain the IOINPROG bit, interlocked block.
952 */
953 for (;;) {
954 u_int oflags;
955 u_int nflags;
956
957 oflags = chain->flags;
958 cpu_ccfence();
959 if (oflags & HAMMER2_CHAIN_IOINPROG) {
960 nflags = oflags | HAMMER2_CHAIN_IOSIGNAL;
961 tsleep_interlock(&chain->flags, 0);
962 if (atomic_cmpset_int(&chain->flags, oflags, nflags)) {
963 tsleep(&chain->flags, PINTERLOCKED,
964 "h2iocw", 0);
965 }
966 /* retry */
967 } else {
968 nflags = oflags | HAMMER2_CHAIN_IOINPROG;
969 if (atomic_cmpset_int(&chain->flags, oflags, nflags)) {
970 break;
971 }
972 /* retry */
973 }
974 }
975
976 /*
977 * We own CHAIN_IOINPROG
978 *
979 * Degenerate case if we raced another load.
980 */
981 if (chain->data) {
982 if (chain->dio)
983 hammer2_io_bkvasync(chain->dio);
984 goto done;
985 }
986
987 /*
988 * We must resolve to a device buffer, either by issuing I/O or
989 * by creating a zero-fill element. We do not mark the buffer
990 * dirty when creating a zero-fill element (the hammer2_chain_modify()
991 * API must still be used to do that).
992 */
993 bref = &chain->bref;
994
995 /*
996 * The getblk() optimization can only be used on newly created
997 * elements if the physical block size matches the request.
998 */
999 if (chain->flags & HAMMER2_CHAIN_INITIAL) {
1000 error = hammer2_io_new(hmp, bref->type,
1001 bref->data_off, chain->bytes,
1002 &chain->dio);
1003 } else {
1004 error = hammer2_io_bread(hmp, bref->type,
1005 bref->data_off, chain->bytes,
1006 &chain->dio);
1007 hammer2_adjreadcounter(chain->bref.type, chain->bytes);
1008 }
1009 if (error) {
1010 chain->error = HAMMER2_ERROR_EIO;
1011 kprintf("hammer2_chain_load_data: I/O error %016jx: %d\n",
1012 (intmax_t)bref->data_off, error);
1013 hammer2_io_bqrelse(&chain->dio);
1014 goto done;
1015 }
1016 chain->error = 0;
1017
1018 /*
1019 * This isn't perfect and can be ignored on OSs which do not have
1020 * an indication as to whether a buffer is coming from cache or
1021 * if I/O was actually issued for the read. TESTEDGOOD will work
1022 * pretty well without the B_IOISSUED logic because chains are
1023 * cached, but in that situation (without B_IOISSUED) it will not
1024 * detect whether a re-read via I/O is corrupted verses the original
1025 * read.
1026 *
1027 * We can't re-run the CRC on every fresh lock. That would be
1028 * insanely expensive.
1029 *
1030 * If the underlying kernel buffer covers the entire chain we can
1031 * use the B_IOISSUED indication to determine if we have to re-run
1032 * the CRC on chain data for chains that managed to stay cached
1033 * across the kernel disposal of the original buffer.
1034 */
1035 if ((dio = chain->dio) != NULL && dio->bp) {
1036 //struct m_buf *bp = dio->bp;
1037
1038 if (dio->psize == chain->bytes //&&
1039 /*(bp->b_flags & B_IOISSUED)*/) {
1040 atomic_clear_int(&chain->flags,
1041 HAMMER2_CHAIN_TESTEDGOOD);
1042 //bp->b_flags &= ~B_IOISSUED;
1043 }
1044 }
1045
1046 /*
1047 * NOTE: A locked chain's data cannot be modified without first
1048 * calling hammer2_chain_modify().
1049 */
1050
1051 /*
1052 * NOTE: hammer2_io_data() call issues bkvasync()
1053 */
1054 bdata = hammer2_io_data(chain->dio, chain->bref.data_off);
1055
1056 if (chain->flags & HAMMER2_CHAIN_INITIAL) {
1057 /*
1058 * Clear INITIAL. In this case we used io_new() and the
1059 * buffer has been zero'd and marked dirty.
1060 *
1061 * CHAIN_MODIFIED has not been set yet, and we leave it
1062 * that way for now. Set a temporary CHAIN_NOTTESTED flag
1063 * to prevent hammer2_chain_testcheck() from trying to match
1064 * a check code that has not yet been generated. This bit
1065 * should NOT end up on the actual media.
1066 */
1067 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1068 atomic_set_int(&chain->flags, HAMMER2_CHAIN_NOTTESTED);
1069 } else if (chain->flags & HAMMER2_CHAIN_MODIFIED) {
1070 /*
1071 * check data not currently synchronized due to
1072 * modification. XXX assumes data stays in the buffer
1073 * cache, which might not be true (need biodep on flush
1074 * to calculate crc? or simple crc?).
1075 */
1076 } else if ((chain->flags & HAMMER2_CHAIN_TESTEDGOOD) == 0) {
1077 if (hammer2_chain_testcheck(chain, bdata) == 0) {
1078 chain->error = HAMMER2_ERROR_CHECK;
1079 } else {
1080 atomic_set_int(&chain->flags, HAMMER2_CHAIN_TESTEDGOOD);
1081 }
1082 }
1083
1084 /*
1085 * Setup the data pointer by pointing it into the buffer.
1086 * WARNING! Other threads can start using the data the instant we
1087 * set chain->data non-NULL.
1088 */
1089 switch (bref->type) {
1090 case HAMMER2_BREF_TYPE_VOLUME:
1091 case HAMMER2_BREF_TYPE_FREEMAP:
1092 panic("hammer2_chain_load_data: unresolved volume header");
1093 break;
1094 case HAMMER2_BREF_TYPE_DIRENT:
1095 KKASSERT(chain->bytes != 0);
1096 /* fall through */
1097 case HAMMER2_BREF_TYPE_INODE:
1098 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
1099 case HAMMER2_BREF_TYPE_INDIRECT:
1100 case HAMMER2_BREF_TYPE_DATA:
1101 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1102 default:
1103 /*
1104 * Point data at the device buffer and leave dio intact.
1105 */
1106 chain->data = (void *)bdata;
1107 break;
1108 }
1109
1110 /*
1111 * Release HAMMER2_CHAIN_IOINPROG and signal waiters if requested.
1112 */
1113 done:
1114 for (;;) {
1115 u_int oflags;
1116 u_int nflags;
1117
1118 oflags = chain->flags;
1119 nflags = oflags & ~(HAMMER2_CHAIN_IOINPROG |
1120 HAMMER2_CHAIN_IOSIGNAL);
1121 KKASSERT(oflags & HAMMER2_CHAIN_IOINPROG);
1122 if (atomic_cmpset_int(&chain->flags, oflags, nflags)) {
1123 if (oflags & HAMMER2_CHAIN_IOSIGNAL)
1124 wakeup(&chain->flags);
1125 break;
1126 }
1127 }
1128 }
1129
1130 /*
1131 * Unlock and deref a chain element.
1132 *
1133 * Remember that the presence of children under chain prevent the chain's
1134 * destruction but do not add additional references, so the dio will still
1135 * be dropped.
1136 */
1137 void
hammer2_chain_unlock(hammer2_chain_t * chain)1138 hammer2_chain_unlock(hammer2_chain_t *chain)
1139 {
1140 hammer2_io_t *dio;
1141 u_int lockcnt;
1142 int iter = 0;
1143
1144 /*
1145 * If multiple locks are present (or being attempted) on this
1146 * particular chain we can just unlock, drop refs, and return.
1147 *
1148 * Otherwise fall-through on the 1->0 transition.
1149 */
1150 for (;;) {
1151 lockcnt = chain->lockcnt;
1152 KKASSERT(lockcnt > 0);
1153 cpu_ccfence();
1154 if (lockcnt > 1) {
1155 if (atomic_cmpset_int(&chain->lockcnt,
1156 lockcnt, lockcnt - 1)) {
1157 hammer2_mtx_unlock(&chain->lock);
1158 return;
1159 }
1160 } else if (hammer2_mtx_upgrade_try(&chain->lock) == 0) {
1161 /* while holding the mutex exclusively */
1162 if (atomic_cmpset_int(&chain->lockcnt, 1, 0))
1163 break;
1164 } else {
1165 /*
1166 * This situation can easily occur on SMP due to
1167 * the gap inbetween the 1->0 transition and the
1168 * final unlock. We cannot safely block on the
1169 * mutex because lockcnt might go above 1.
1170 *
1171 * XXX Sleep for one tick if it takes too long.
1172 */
1173 if (++iter > 1000) {
1174 if (iter > 1000 + hz) {
1175 kprintf("hammer2: h2race2 %p\n", chain);
1176 iter = 1000;
1177 }
1178 tsleep(&iter, 0, "h2race2", 1);
1179 }
1180 cpu_pause();
1181 }
1182 /* retry */
1183 }
1184
1185 /*
1186 * Last unlock / mutex upgraded to exclusive. Drop the data
1187 * reference.
1188 */
1189 dio = hammer2_chain_drop_data(chain);
1190 if (dio)
1191 hammer2_io_bqrelse(&dio);
1192 hammer2_mtx_unlock(&chain->lock);
1193 }
1194
1195 #if 0
1196 /*
1197 * Unlock and hold chain data intact
1198 */
1199 void
1200 hammer2_chain_unlock_hold(hammer2_chain_t *chain)
1201 {
1202 atomic_add_int(&chain->lockcnt, 1);
1203 hammer2_chain_unlock(chain);
1204 }
1205 #endif
1206
1207 /*
1208 * Helper to obtain the blockref[] array base and count for a chain.
1209 *
1210 * XXX Not widely used yet, various use cases need to be validated and
1211 * converted to use this function.
1212 */
1213 static
1214 hammer2_blockref_t *
hammer2_chain_base_and_count(hammer2_chain_t * parent,int * countp)1215 hammer2_chain_base_and_count(hammer2_chain_t *parent, int *countp)
1216 {
1217 hammer2_blockref_t *base;
1218 int count;
1219
1220 if (parent->flags & HAMMER2_CHAIN_INITIAL) {
1221 base = NULL;
1222
1223 switch(parent->bref.type) {
1224 case HAMMER2_BREF_TYPE_INODE:
1225 count = HAMMER2_SET_COUNT;
1226 break;
1227 case HAMMER2_BREF_TYPE_INDIRECT:
1228 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1229 count = parent->bytes / sizeof(hammer2_blockref_t);
1230 break;
1231 case HAMMER2_BREF_TYPE_VOLUME:
1232 count = HAMMER2_SET_COUNT;
1233 break;
1234 case HAMMER2_BREF_TYPE_FREEMAP:
1235 count = HAMMER2_SET_COUNT;
1236 break;
1237 default:
1238 panic("hammer2_chain_base_and_count: "
1239 "unrecognized blockref type: %d",
1240 parent->bref.type);
1241 count = 0;
1242 break;
1243 }
1244 } else {
1245 switch(parent->bref.type) {
1246 case HAMMER2_BREF_TYPE_INODE:
1247 base = &parent->data->ipdata.u.blockset.blockref[0];
1248 count = HAMMER2_SET_COUNT;
1249 break;
1250 case HAMMER2_BREF_TYPE_INDIRECT:
1251 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1252 base = &parent->data->npdata[0];
1253 count = parent->bytes / sizeof(hammer2_blockref_t);
1254 break;
1255 case HAMMER2_BREF_TYPE_VOLUME:
1256 base = &parent->data->voldata.
1257 sroot_blockset.blockref[0];
1258 count = HAMMER2_SET_COUNT;
1259 break;
1260 case HAMMER2_BREF_TYPE_FREEMAP:
1261 base = &parent->data->blkset.blockref[0];
1262 count = HAMMER2_SET_COUNT;
1263 break;
1264 default:
1265 panic("hammer2_chain_base_and_count: "
1266 "unrecognized blockref type: %d",
1267 parent->bref.type);
1268 base = NULL;
1269 count = 0;
1270 break;
1271 }
1272 }
1273 *countp = count;
1274
1275 return base;
1276 }
1277
1278 /*
1279 * This counts the number of live blockrefs in a block array and
1280 * also calculates the point at which all remaining blockrefs are empty.
1281 * This routine can only be called on a live chain.
1282 *
1283 * Caller holds the chain locked, but possibly with a shared lock. We
1284 * must use an exclusive spinlock to prevent corruption.
1285 *
1286 * NOTE: Flag is not set until after the count is complete, allowing
1287 * callers to test the flag without holding the spinlock.
1288 *
1289 * NOTE: If base is NULL the related chain is still in the INITIAL
1290 * state and there are no blockrefs to count.
1291 *
1292 * NOTE: live_count may already have some counts accumulated due to
1293 * creation and deletion and could even be initially negative.
1294 */
1295 void
hammer2_chain_countbrefs(hammer2_chain_t * chain,hammer2_blockref_t * base,int count)1296 hammer2_chain_countbrefs(hammer2_chain_t *chain,
1297 hammer2_blockref_t *base, int count)
1298 {
1299 hammer2_spin_ex(&chain->core.spin);
1300 if ((chain->flags & HAMMER2_CHAIN_COUNTEDBREFS) == 0) {
1301 if (base) {
1302 while (--count >= 0) {
1303 if (base[count].type != HAMMER2_BREF_TYPE_EMPTY)
1304 break;
1305 }
1306 chain->core.live_zero = count + 1;
1307 while (count >= 0) {
1308 if (base[count].type != HAMMER2_BREF_TYPE_EMPTY)
1309 atomic_add_int(&chain->core.live_count,
1310 1);
1311 --count;
1312 }
1313 } else {
1314 chain->core.live_zero = 0;
1315 }
1316 /* else do not modify live_count */
1317 atomic_set_int(&chain->flags, HAMMER2_CHAIN_COUNTEDBREFS);
1318 }
1319 hammer2_spin_unex(&chain->core.spin);
1320 }
1321
1322 /*
1323 * Resize the chain's physical storage allocation in-place. This function does
1324 * not usually adjust the data pointer and must be followed by (typically) a
1325 * hammer2_chain_modify() call to copy any old data over and adjust the
1326 * data pointer.
1327 *
1328 * Chains can be resized smaller without reallocating the storage. Resizing
1329 * larger will reallocate the storage. Excess or prior storage is reclaimed
1330 * asynchronously at a later time.
1331 *
1332 * An nradix value of 0 is special-cased to mean that the storage should
1333 * be disassociated, that is the chain is being resized to 0 bytes (not 1
1334 * byte).
1335 *
1336 * Must be passed an exclusively locked parent and chain.
1337 *
1338 * This function is mostly used with DATA blocks locked RESOLVE_NEVER in order
1339 * to avoid instantiating a device buffer that conflicts with the vnode data
1340 * buffer. However, because H2 can compress or encrypt data, the chain may
1341 * have a dio assigned to it in those situations, and they do not conflict.
1342 *
1343 * XXX return error if cannot resize.
1344 */
1345 int
hammer2_chain_resize(hammer2_chain_t * chain,hammer2_tid_t mtid,hammer2_off_t dedup_off,int nradix,int flags)1346 hammer2_chain_resize(hammer2_chain_t *chain,
1347 hammer2_tid_t mtid, hammer2_off_t dedup_off,
1348 int nradix, int flags)
1349 {
1350 hammer2_dev_t *hmp;
1351 size_t obytes;
1352 size_t nbytes;
1353 int error;
1354
1355 hmp = chain->hmp;
1356
1357 /*
1358 * Only data and indirect blocks can be resized for now.
1359 * (The volu root, inodes, and freemap elements use a fixed size).
1360 */
1361 KKASSERT(chain != &hmp->vchain);
1362 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_DATA ||
1363 chain->bref.type == HAMMER2_BREF_TYPE_INDIRECT ||
1364 chain->bref.type == HAMMER2_BREF_TYPE_DIRENT);
1365
1366 /*
1367 * Nothing to do if the element is already the proper size
1368 */
1369 obytes = chain->bytes;
1370 nbytes = (nradix) ? (1U << nradix) : 0;
1371 if (obytes == nbytes)
1372 return (chain->error);
1373
1374 /*
1375 * Make sure the old data is instantiated so we can copy it. If this
1376 * is a data block, the device data may be superfluous since the data
1377 * might be in a logical block, but compressed or encrypted data is
1378 * another matter.
1379 *
1380 * NOTE: The modify will set BLKMAPUPD for us if BLKMAPPED is set.
1381 */
1382 error = hammer2_chain_modify(chain, mtid, dedup_off, 0);
1383 if (error)
1384 return error;
1385
1386 /*
1387 * Reallocate the block, even if making it smaller (because different
1388 * block sizes may be in different regions).
1389 *
1390 * NOTE: Operation does not copy the data and may only be used
1391 * to resize data blocks in-place, or directory entry blocks
1392 * which are about to be modified in some manner.
1393 */
1394 error = hammer2_freemap_alloc(chain, nbytes);
1395 if (error)
1396 return error;
1397
1398 chain->bytes = nbytes;
1399
1400 /*
1401 * We don't want the followup chain_modify() to try to copy data
1402 * from the old (wrong-sized) buffer. It won't know how much to
1403 * copy. This case should only occur during writes when the
1404 * originator already has the data to write in-hand.
1405 */
1406 if (chain->dio) {
1407 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_DATA ||
1408 chain->bref.type == HAMMER2_BREF_TYPE_DIRENT);
1409 hammer2_io_brelse(&chain->dio);
1410 chain->data = NULL;
1411 }
1412 return (chain->error);
1413 }
1414
1415 /*
1416 * Set the chain modified so its data can be changed by the caller, or
1417 * install deduplicated data. The caller must call this routine for each
1418 * set of modifications it makes, even if the chain is already flagged
1419 * MODIFIED.
1420 *
1421 * Sets bref.modify_tid to mtid only if mtid != 0. Note that bref.modify_tid
1422 * is a CLC (cluster level change) field and is not updated by parent
1423 * propagation during a flush.
1424 *
1425 * Returns an appropriate HAMMER2_ERROR_* code, which will generally reflect
1426 * chain->error except for HAMMER2_ERROR_ENOSPC. If the allocation fails
1427 * due to no space available, HAMMER2_ERROR_ENOSPC is returned and the chain
1428 * remains unmodified with its old data ref intact and chain->error
1429 * unchanged.
1430 *
1431 * Dedup Handling
1432 *
1433 * If the DEDUPABLE flag is set in the chain the storage must be reallocated
1434 * even if the chain is still flagged MODIFIED. In this case the chain's
1435 * DEDUPABLE flag will be cleared once the new storage has been assigned.
1436 *
1437 * If the caller passes a non-zero dedup_off we will use it to assign the
1438 * new storage. The MODIFIED flag will be *CLEARED* in this case, and
1439 * DEDUPABLE will be set (NOTE: the UPDATE flag is always set). The caller
1440 * must not modify the data content upon return.
1441 */
1442 int
hammer2_chain_modify(hammer2_chain_t * chain,hammer2_tid_t mtid,hammer2_off_t dedup_off,int flags)1443 hammer2_chain_modify(hammer2_chain_t *chain, hammer2_tid_t mtid,
1444 hammer2_off_t dedup_off, int flags)
1445 {
1446 hammer2_dev_t *hmp;
1447 hammer2_io_t *dio;
1448 int error;
1449 int wasinitial;
1450 int setmodified;
1451 int setupdate;
1452 int newmod;
1453 char *bdata;
1454
1455 hmp = chain->hmp;
1456 KKASSERT(chain->lock.mtx_lock & MTX_EXCLUSIVE);
1457
1458 /*
1459 * Data is not optional for freemap chains (we must always be sure
1460 * to copy the data on COW storage allocations).
1461 */
1462 if (chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE ||
1463 chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) {
1464 KKASSERT((chain->flags & HAMMER2_CHAIN_INITIAL) ||
1465 (flags & HAMMER2_MODIFY_OPTDATA) == 0);
1466 }
1467
1468 /*
1469 * Data must be resolved if already assigned, unless explicitly
1470 * flagged otherwise. If we cannot safety load the data the
1471 * modification fails and we return early.
1472 */
1473 if (chain->data == NULL && chain->bytes != 0 &&
1474 (flags & HAMMER2_MODIFY_OPTDATA) == 0 &&
1475 (chain->bref.data_off & ~HAMMER2_OFF_MASK_RADIX)) {
1476 hammer2_chain_load_data(chain);
1477 if (chain->error)
1478 return (chain->error);
1479 }
1480 error = 0;
1481
1482 /*
1483 * Set MODIFIED to indicate that the chain has been modified. A new
1484 * allocation is required when modifying a chain.
1485 *
1486 * Set UPDATE to ensure that the blockref is updated in the parent.
1487 *
1488 * If MODIFIED is already set determine if we can reuse the assigned
1489 * data block or if we need a new data block.
1490 */
1491 if ((chain->flags & HAMMER2_CHAIN_MODIFIED) == 0) {
1492 /*
1493 * Must set modified bit.
1494 */
1495 atomic_add_long(&hammer2_count_modified_chains, 1);
1496 atomic_set_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
1497 hammer2_pfs_memory_inc(chain->pmp); /* can be NULL */
1498 setmodified = 1;
1499
1500 /*
1501 * We may be able to avoid a copy-on-write if the chain's
1502 * check mode is set to NONE and the chain's current
1503 * modify_tid is beyond the last explicit snapshot tid.
1504 *
1505 * This implements HAMMER2's overwrite-in-place feature.
1506 *
1507 * NOTE! This data-block cannot be used as a de-duplication
1508 * source when the check mode is set to NONE.
1509 */
1510 if ((chain->bref.type == HAMMER2_BREF_TYPE_DATA ||
1511 chain->bref.type == HAMMER2_BREF_TYPE_DIRENT) &&
1512 (chain->flags & HAMMER2_CHAIN_INITIAL) == 0 &&
1513 (chain->flags & HAMMER2_CHAIN_DEDUPABLE) == 0 &&
1514 HAMMER2_DEC_CHECK(chain->bref.methods) ==
1515 HAMMER2_CHECK_NONE &&
1516 chain->pmp &&
1517 chain->bref.modify_tid >
1518 chain->pmp->iroot->meta.pfs_lsnap_tid) {
1519 /*
1520 * Sector overwrite allowed.
1521 */
1522 newmod = 0;
1523 } else if ((hmp->hflags & HMNT2_EMERG) &&
1524 chain->pmp &&
1525 chain->bref.modify_tid >
1526 chain->pmp->iroot->meta.pfs_lsnap_tid) {
1527 /*
1528 * If in emergency delete mode then do a modify-in-
1529 * place on any chain type belonging to the PFS as
1530 * long as it doesn't mess up a snapshot. We might
1531 * be forced to do this anyway a little further down
1532 * in the code if the allocation fails.
1533 *
1534 * Also note that in emergency mode, these modify-in-
1535 * place operations are NOT SAFE. A storage failure,
1536 * power failure, or panic can corrupt the filesystem.
1537 */
1538 newmod = 0;
1539 } else {
1540 /*
1541 * Sector overwrite not allowed, must copy-on-write.
1542 */
1543 newmod = 1;
1544 }
1545 } else if (chain->flags & HAMMER2_CHAIN_DEDUPABLE) {
1546 /*
1547 * If the modified chain was registered for dedup we need
1548 * a new allocation. This only happens for delayed-flush
1549 * chains (i.e. which run through the front-end buffer
1550 * cache).
1551 */
1552 newmod = 1;
1553 setmodified = 0;
1554 } else {
1555 /*
1556 * Already flagged modified, no new allocation is needed.
1557 */
1558 newmod = 0;
1559 setmodified = 0;
1560 }
1561
1562 /*
1563 * Flag parent update required.
1564 */
1565 if ((chain->flags & HAMMER2_CHAIN_UPDATE) == 0) {
1566 atomic_set_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
1567 setupdate = 1;
1568 } else {
1569 setupdate = 0;
1570 }
1571
1572 /*
1573 * The XOP code returns held but unlocked focus chains. This
1574 * prevents the chain from being destroyed but does not prevent
1575 * it from being modified. diolk is used to interlock modifications
1576 * against XOP frontend accesses to the focus.
1577 *
1578 * This allows us to theoretically avoid deadlocking the frontend
1579 * if one of the backends lock up by not formally locking the
1580 * focused chain in the frontend. In addition, the synchronization
1581 * code relies on this mechanism to avoid deadlocking concurrent
1582 * synchronization threads.
1583 */
1584 lockmgr(&chain->diolk, LK_EXCLUSIVE);
1585
1586 /*
1587 * The modification or re-modification requires an allocation and
1588 * possible COW. If an error occurs, the previous content and data
1589 * reference is retained and the modification fails.
1590 *
1591 * If dedup_off is non-zero, the caller is requesting a deduplication
1592 * rather than a modification. The MODIFIED bit is not set and the
1593 * data offset is set to the deduplication offset. The data cannot
1594 * be modified.
1595 *
1596 * NOTE: The dedup offset is allowed to be in a partially free state
1597 * and we must be sure to reset it to a fully allocated state
1598 * to force two bulkfree passes to free it again.
1599 *
1600 * NOTE: Only applicable when chain->bytes != 0.
1601 *
1602 * XXX can a chain already be marked MODIFIED without a data
1603 * assignment? If not, assert here instead of testing the case.
1604 */
1605 if (chain != &hmp->vchain && chain != &hmp->fchain &&
1606 chain->bytes) {
1607 if ((chain->bref.data_off & ~HAMMER2_OFF_MASK_RADIX) == 0 ||
1608 newmod
1609 ) {
1610 /*
1611 * NOTE: We do not have to remove the dedup
1612 * registration because the area is still
1613 * allocated and the underlying DIO will
1614 * still be flushed.
1615 */
1616 if (dedup_off) {
1617 chain->bref.data_off = dedup_off;
1618 if ((int)(dedup_off & HAMMER2_OFF_MASK_RADIX))
1619 chain->bytes = 1 <<
1620 (int)(dedup_off &
1621 HAMMER2_OFF_MASK_RADIX);
1622 else
1623 chain->bytes = 0;
1624 chain->error = 0;
1625 atomic_clear_int(&chain->flags,
1626 HAMMER2_CHAIN_MODIFIED);
1627 atomic_add_long(&hammer2_count_modified_chains,
1628 -1);
1629 if (chain->pmp) {
1630 hammer2_pfs_memory_wakeup(
1631 chain->pmp, -1);
1632 }
1633 hammer2_freemap_adjust(hmp, &chain->bref,
1634 HAMMER2_FREEMAP_DORECOVER);
1635 atomic_set_int(&chain->flags,
1636 HAMMER2_CHAIN_DEDUPABLE);
1637 } else {
1638 error = hammer2_freemap_alloc(chain,
1639 chain->bytes);
1640 atomic_clear_int(&chain->flags,
1641 HAMMER2_CHAIN_DEDUPABLE);
1642
1643 /*
1644 * If we are unable to allocate a new block
1645 * but we are in emergency mode, issue a
1646 * warning to the console and reuse the same
1647 * block.
1648 *
1649 * We behave as if the allocation were
1650 * successful.
1651 *
1652 * THIS IS IMPORTANT: These modifications
1653 * are virtually guaranteed to corrupt any
1654 * snapshots related to this filesystem.
1655 */
1656 if (error && (hmp->hflags & HMNT2_EMERG)) {
1657 error = 0;
1658 chain->bref.flags |=
1659 HAMMER2_BREF_FLAG_EMERG_MIP;
1660
1661 krateprintf(&krate_h2em,
1662 "hammer2: Emergency Mode WARNING: "
1663 "Operation will likely corrupt "
1664 "related snapshot: "
1665 "%016jx.%02x key=%016jx\n",
1666 chain->bref.data_off,
1667 chain->bref.type,
1668 chain->bref.key);
1669 } else if (error == 0) {
1670 chain->bref.flags &=
1671 ~HAMMER2_BREF_FLAG_EMERG_MIP;
1672 }
1673 }
1674 }
1675 }
1676
1677 /*
1678 * Stop here if error. We have to undo any flag bits we might
1679 * have set above.
1680 */
1681 if (error) {
1682 if (setmodified) {
1683 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
1684 atomic_add_long(&hammer2_count_modified_chains, -1);
1685 if (chain->pmp)
1686 hammer2_pfs_memory_wakeup(chain->pmp, -1);
1687 }
1688 if (setupdate) {
1689 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
1690 }
1691 lockmgr(&chain->diolk, LK_RELEASE);
1692
1693 return error;
1694 }
1695
1696 /*
1697 * Update mirror_tid and modify_tid. modify_tid is only updated
1698 * if not passed as zero (during flushes, parent propagation passes
1699 * the value 0).
1700 *
1701 * NOTE: chain->pmp could be the device spmp.
1702 */
1703 chain->bref.mirror_tid = hmp->voldata.mirror_tid + 1;
1704 if (mtid)
1705 chain->bref.modify_tid = mtid;
1706
1707 /*
1708 * Set BLKMAPUPD to tell the flush code that an existing blockmap entry
1709 * requires updating as well as to tell the delete code that the
1710 * chain's blockref might not exactly match (in terms of physical size
1711 * or block offset) the one in the parent's blocktable. The base key
1712 * of course will still match.
1713 */
1714 if (chain->flags & HAMMER2_CHAIN_BLKMAPPED)
1715 atomic_set_int(&chain->flags, HAMMER2_CHAIN_BLKMAPUPD);
1716
1717 /*
1718 * Short-cut data block handling when the caller does not need an
1719 * actual data reference to (aka OPTDATA), as long as the chain does
1720 * not already have a data pointer to the data and no de-duplication
1721 * occurred.
1722 *
1723 * This generally means that the modifications are being done via the
1724 * logical buffer cache.
1725 *
1726 * NOTE: If deduplication occurred we have to run through the data
1727 * stuff to clear INITIAL, and the caller will likely want to
1728 * assign the check code anyway. Leaving INITIAL set on a
1729 * dedup can be deadly (it can cause the block to be zero'd!).
1730 *
1731 * This code also handles bytes == 0 (most dirents).
1732 */
1733 if (chain->bref.type == HAMMER2_BREF_TYPE_DATA &&
1734 (flags & HAMMER2_MODIFY_OPTDATA) &&
1735 chain->data == NULL) {
1736 if (dedup_off == 0) {
1737 KKASSERT(chain->dio == NULL);
1738 goto skip2;
1739 }
1740 }
1741
1742 /*
1743 * Clearing the INITIAL flag (for indirect blocks) indicates that
1744 * we've processed the uninitialized storage allocation.
1745 *
1746 * If this flag is already clear we are likely in a copy-on-write
1747 * situation but we have to be sure NOT to bzero the storage if
1748 * no data is present.
1749 *
1750 * Clearing of NOTTESTED is allowed if the MODIFIED bit is set,
1751 */
1752 if (chain->flags & HAMMER2_CHAIN_INITIAL) {
1753 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1754 wasinitial = 1;
1755 } else {
1756 wasinitial = 0;
1757 }
1758
1759 /*
1760 * Instantiate data buffer and possibly execute COW operation
1761 */
1762 switch(chain->bref.type) {
1763 case HAMMER2_BREF_TYPE_VOLUME:
1764 case HAMMER2_BREF_TYPE_FREEMAP:
1765 /*
1766 * The data is embedded, no copy-on-write operation is
1767 * needed.
1768 */
1769 KKASSERT(chain->dio == NULL);
1770 break;
1771 case HAMMER2_BREF_TYPE_DIRENT:
1772 /*
1773 * The data might be fully embedded.
1774 */
1775 if (chain->bytes == 0) {
1776 KKASSERT(chain->dio == NULL);
1777 break;
1778 }
1779 /* fall through */
1780 case HAMMER2_BREF_TYPE_INODE:
1781 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
1782 case HAMMER2_BREF_TYPE_DATA:
1783 case HAMMER2_BREF_TYPE_INDIRECT:
1784 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1785 /*
1786 * Perform the copy-on-write operation
1787 *
1788 * zero-fill or copy-on-write depending on whether
1789 * chain->data exists or not and set the dirty state for
1790 * the new buffer. hammer2_io_new() will handle the
1791 * zero-fill.
1792 *
1793 * If a dedup_off was supplied this is an existing block
1794 * and no COW, copy, or further modification is required.
1795 */
1796 KKASSERT(chain != &hmp->vchain && chain != &hmp->fchain);
1797
1798 if (wasinitial && dedup_off == 0) {
1799 error = hammer2_io_new(hmp, chain->bref.type,
1800 chain->bref.data_off,
1801 chain->bytes, &dio);
1802 } else {
1803 error = hammer2_io_bread(hmp, chain->bref.type,
1804 chain->bref.data_off,
1805 chain->bytes, &dio);
1806 }
1807 hammer2_adjreadcounter(chain->bref.type, chain->bytes);
1808
1809 /*
1810 * If an I/O error occurs make sure callers cannot accidently
1811 * modify the old buffer's contents and corrupt the filesystem.
1812 *
1813 * NOTE: hammer2_io_data() call issues bkvasync()
1814 */
1815 if (error) {
1816 kprintf("hammer2_chain_modify: hmp=%p I/O error\n",
1817 hmp);
1818 chain->error = HAMMER2_ERROR_EIO;
1819 hammer2_io_brelse(&dio);
1820 hammer2_io_brelse(&chain->dio);
1821 chain->data = NULL;
1822 break;
1823 }
1824 chain->error = 0;
1825 bdata = hammer2_io_data(dio, chain->bref.data_off);
1826
1827 if (chain->data) {
1828 /*
1829 * COW (unless a dedup).
1830 */
1831 KKASSERT(chain->dio != NULL);
1832 if (chain->data != (void *)bdata && dedup_off == 0) {
1833 bcopy(chain->data, bdata, chain->bytes);
1834 }
1835 } else if (wasinitial == 0 && dedup_off == 0) {
1836 /*
1837 * We have a problem. We were asked to COW but
1838 * we don't have any data to COW with!
1839 */
1840 panic("hammer2_chain_modify: having a COW %p\n",
1841 chain);
1842 }
1843
1844 /*
1845 * Retire the old buffer, replace with the new. Dirty or
1846 * redirty the new buffer.
1847 *
1848 * WARNING! The system buffer cache may have already flushed
1849 * the buffer, so we must be sure to [re]dirty it
1850 * for further modification.
1851 *
1852 * If dedup_off was supplied, the caller is not
1853 * expected to make any further modification to the
1854 * buffer.
1855 *
1856 * WARNING! hammer2_get_gdata() assumes dio never transitions
1857 * through NULL in order to optimize away unnecessary
1858 * diolk operations.
1859 */
1860 {
1861 hammer2_io_t *tio;
1862
1863 if ((tio = chain->dio) != NULL)
1864 hammer2_io_bqrelse(&tio);
1865 chain->data = (void *)bdata;
1866 chain->dio = dio;
1867 if (dedup_off == 0)
1868 hammer2_io_setdirty(dio);
1869 }
1870 break;
1871 default:
1872 panic("hammer2_chain_modify: illegal non-embedded type %d",
1873 chain->bref.type);
1874 break;
1875
1876 }
1877 skip2:
1878 /*
1879 * setflush on parent indicating that the parent must recurse down
1880 * to us. Do not call on chain itself which might already have it
1881 * set.
1882 */
1883 if (chain->parent)
1884 hammer2_chain_setflush(chain->parent);
1885 lockmgr(&chain->diolk, LK_RELEASE);
1886
1887 return (chain->error);
1888 }
1889
1890 /*
1891 * Modify the chain associated with an inode.
1892 */
1893 int
hammer2_chain_modify_ip(hammer2_inode_t * ip,hammer2_chain_t * chain,hammer2_tid_t mtid,int flags)1894 hammer2_chain_modify_ip(hammer2_inode_t *ip, hammer2_chain_t *chain,
1895 hammer2_tid_t mtid, int flags)
1896 {
1897 int error;
1898
1899 hammer2_inode_modify(ip);
1900 error = hammer2_chain_modify(chain, mtid, 0, flags);
1901
1902 return error;
1903 }
1904
1905 /*
1906 * This function returns the chain at the nearest key within the specified
1907 * range. The returned chain will be referenced but not locked.
1908 *
1909 * This function will recurse through chain->rbtree as necessary and will
1910 * return a *key_nextp suitable for iteration. *key_nextp is only set if
1911 * the iteration value is less than the current value of *key_nextp.
1912 *
1913 * The caller should use (*key_nextp) to calculate the actual range of
1914 * the returned element, which will be (key_beg to *key_nextp - 1), because
1915 * there might be another element which is superior to the returned element
1916 * and overlaps it.
1917 *
1918 * (*key_nextp) can be passed as key_beg in an iteration only while non-NULL
1919 * chains continue to be returned. On EOF (*key_nextp) may overflow since
1920 * it will wind up being (key_end + 1).
1921 *
1922 * WARNING! Must be called with child's spinlock held. Spinlock remains
1923 * held through the operation.
1924 */
1925 struct hammer2_chain_find_info {
1926 hammer2_chain_t *best;
1927 hammer2_key_t key_beg;
1928 hammer2_key_t key_end;
1929 hammer2_key_t key_next;
1930 };
1931
1932 static int hammer2_chain_find_cmp(hammer2_chain_t *child, void *data);
1933 static int hammer2_chain_find_callback(hammer2_chain_t *child, void *data);
1934
1935 static
1936 hammer2_chain_t *
hammer2_chain_find(hammer2_chain_t * parent,hammer2_key_t * key_nextp,hammer2_key_t key_beg,hammer2_key_t key_end)1937 hammer2_chain_find(hammer2_chain_t *parent, hammer2_key_t *key_nextp,
1938 hammer2_key_t key_beg, hammer2_key_t key_end)
1939 {
1940 struct hammer2_chain_find_info info;
1941
1942 info.best = NULL;
1943 info.key_beg = key_beg;
1944 info.key_end = key_end;
1945 info.key_next = *key_nextp;
1946
1947 RB_SCAN(hammer2_chain_tree, &parent->core.rbtree,
1948 hammer2_chain_find_cmp, hammer2_chain_find_callback,
1949 &info);
1950 *key_nextp = info.key_next;
1951 #if 0
1952 kprintf("chain_find %p %016jx:%016jx next=%016jx\n",
1953 parent, key_beg, key_end, *key_nextp);
1954 #endif
1955
1956 return (info.best);
1957 }
1958
1959 static
1960 int
hammer2_chain_find_cmp(hammer2_chain_t * child,void * data)1961 hammer2_chain_find_cmp(hammer2_chain_t *child, void *data)
1962 {
1963 struct hammer2_chain_find_info *info = data;
1964 hammer2_key_t child_beg;
1965 hammer2_key_t child_end;
1966
1967 child_beg = child->bref.key;
1968 child_end = child_beg + ((hammer2_key_t)1 << child->bref.keybits) - 1;
1969
1970 if (child_end < info->key_beg)
1971 return(-1);
1972 if (child_beg > info->key_end)
1973 return(1);
1974 return(0);
1975 }
1976
1977 static
1978 int
hammer2_chain_find_callback(hammer2_chain_t * child,void * data)1979 hammer2_chain_find_callback(hammer2_chain_t *child, void *data)
1980 {
1981 struct hammer2_chain_find_info *info = data;
1982 hammer2_chain_t *best;
1983 hammer2_key_t child_end;
1984
1985 if ((best = info->best) == NULL) {
1986 /*
1987 * No previous best. Assign best
1988 */
1989 info->best = child;
1990 } else if (best->bref.key <= info->key_beg &&
1991 child->bref.key <= info->key_beg) {
1992 /*
1993 * Illegal overlap.
1994 */
1995 KKASSERT(0);
1996 /*info->best = child;*/
1997 } else if (child->bref.key < best->bref.key) {
1998 /*
1999 * Child has a nearer key and best is not flush with key_beg.
2000 * Set best to child. Truncate key_next to the old best key.
2001 */
2002 info->best = child;
2003 if (info->key_next > best->bref.key || info->key_next == 0)
2004 info->key_next = best->bref.key;
2005 } else if (child->bref.key == best->bref.key) {
2006 /*
2007 * If our current best is flush with the child then this
2008 * is an illegal overlap.
2009 *
2010 * key_next will automatically be limited to the smaller of
2011 * the two end-points.
2012 */
2013 KKASSERT(0);
2014 info->best = child;
2015 } else {
2016 /*
2017 * Keep the current best but truncate key_next to the child's
2018 * base.
2019 *
2020 * key_next will also automatically be limited to the smaller
2021 * of the two end-points (probably not necessary for this case
2022 * but we do it anyway).
2023 */
2024 if (info->key_next > child->bref.key || info->key_next == 0)
2025 info->key_next = child->bref.key;
2026 }
2027
2028 /*
2029 * Always truncate key_next based on child's end-of-range.
2030 */
2031 child_end = child->bref.key + ((hammer2_key_t)1 << child->bref.keybits);
2032 if (child_end && (info->key_next > child_end || info->key_next == 0))
2033 info->key_next = child_end;
2034
2035 return(0);
2036 }
2037
2038 /*
2039 * Retrieve the specified chain from a media blockref, creating the
2040 * in-memory chain structure which reflects it. The returned chain is
2041 * held and locked according to (how) (HAMMER2_RESOLVE_*). The caller must
2042 * handle crc-checks and so forth, and should check chain->error before
2043 * assuming that the data is good.
2044 *
2045 * To handle insertion races pass the INSERT_RACE flag along with the
2046 * generation number of the core. NULL will be returned if the generation
2047 * number changes before we have a chance to insert the chain. Insert
2048 * races can occur because the parent might be held shared.
2049 *
2050 * Caller must hold the parent locked shared or exclusive since we may
2051 * need the parent's bref array to find our block.
2052 *
2053 * WARNING! chain->pmp is always set to NULL for any chain representing
2054 * part of the super-root topology.
2055 */
2056 hammer2_chain_t *
hammer2_chain_get(hammer2_chain_t * parent,int generation,hammer2_blockref_t * bref,int how)2057 hammer2_chain_get(hammer2_chain_t *parent, int generation,
2058 hammer2_blockref_t *bref, int how)
2059 {
2060 hammer2_dev_t *hmp = parent->hmp;
2061 hammer2_chain_t *chain;
2062 int error;
2063
2064 /*
2065 * Allocate a chain structure representing the existing media
2066 * entry. Resulting chain has one ref and is not locked.
2067 */
2068 if (bref->flags & HAMMER2_BREF_FLAG_PFSROOT)
2069 chain = hammer2_chain_alloc(hmp, NULL, bref);
2070 else
2071 chain = hammer2_chain_alloc(hmp, parent->pmp, bref);
2072 /* ref'd chain returned */
2073
2074 /*
2075 * Flag that the chain is in the parent's blockmap so delete/flush
2076 * knows what to do with it.
2077 */
2078 atomic_set_int(&chain->flags, HAMMER2_CHAIN_BLKMAPPED);
2079
2080 /*
2081 * chain must be locked to avoid unexpected ripouts
2082 */
2083 hammer2_chain_lock(chain, how);
2084
2085 /*
2086 * Link the chain into its parent. A spinlock is required to safely
2087 * access the RBTREE, and it is possible to collide with another
2088 * hammer2_chain_get() operation because the caller might only hold
2089 * a shared lock on the parent.
2090 *
2091 * NOTE: Get races can occur quite often when we distribute
2092 * asynchronous read-aheads across multiple threads.
2093 */
2094 KKASSERT(parent->refs > 0);
2095 error = hammer2_chain_insert(parent, chain,
2096 HAMMER2_CHAIN_INSERT_SPIN |
2097 HAMMER2_CHAIN_INSERT_RACE,
2098 generation);
2099 if (error) {
2100 KKASSERT((chain->flags & HAMMER2_CHAIN_ONRBTREE) == 0);
2101 /*kprintf("chain %p get race\n", chain);*/
2102 hammer2_chain_unlock(chain);
2103 hammer2_chain_drop(chain);
2104 chain = NULL;
2105 } else {
2106 KKASSERT(chain->flags & HAMMER2_CHAIN_ONRBTREE);
2107 }
2108
2109 /*
2110 * Return our new chain referenced but not locked, or NULL if
2111 * a race occurred.
2112 */
2113 return (chain);
2114 }
2115
2116 /*
2117 * Lookup initialization/completion API
2118 */
2119 hammer2_chain_t *
hammer2_chain_lookup_init(hammer2_chain_t * parent,int flags)2120 hammer2_chain_lookup_init(hammer2_chain_t *parent, int flags)
2121 {
2122 hammer2_chain_ref(parent);
2123 if (flags & HAMMER2_LOOKUP_SHARED) {
2124 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS |
2125 HAMMER2_RESOLVE_SHARED);
2126 } else {
2127 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2128 }
2129 return (parent);
2130 }
2131
2132 void
hammer2_chain_lookup_done(hammer2_chain_t * parent)2133 hammer2_chain_lookup_done(hammer2_chain_t *parent)
2134 {
2135 if (parent) {
2136 hammer2_chain_unlock(parent);
2137 hammer2_chain_drop(parent);
2138 }
2139 }
2140
2141 /*
2142 * Take the locked chain and return a locked parent. The chain remains
2143 * locked on return, but may have to be temporarily unlocked to acquire
2144 * the parent. Because of this, (chain) must be stable and cannot be
2145 * deleted while it was temporarily unlocked (typically means that (chain)
2146 * is an inode).
2147 *
2148 * Pass HAMMER2_RESOLVE_* flags in flags.
2149 *
2150 * This will work even if the chain is errored, and the caller can check
2151 * parent->error on return if desired since the parent will be locked.
2152 *
2153 * This function handles the lock order reversal.
2154 */
2155 hammer2_chain_t *
hammer2_chain_getparent(hammer2_chain_t * chain,int flags)2156 hammer2_chain_getparent(hammer2_chain_t *chain, int flags)
2157 {
2158 hammer2_chain_t *parent;
2159
2160 /*
2161 * Be careful of order, chain must be unlocked before parent
2162 * is locked below to avoid a deadlock. Try it trivially first.
2163 */
2164 parent = chain->parent;
2165 if (parent == NULL)
2166 panic("hammer2_chain_getparent: no parent");
2167 hammer2_chain_ref(parent);
2168 if (hammer2_chain_lock(parent, flags|HAMMER2_RESOLVE_NONBLOCK) == 0)
2169 return parent;
2170
2171 for (;;) {
2172 hammer2_chain_unlock(chain);
2173 hammer2_chain_lock(parent, flags);
2174 hammer2_chain_lock(chain, flags);
2175
2176 /*
2177 * Parent relinking races are quite common. We have to get
2178 * it right or we will blow up the block table.
2179 */
2180 if (chain->parent == parent)
2181 break;
2182 hammer2_chain_unlock(parent);
2183 hammer2_chain_drop(parent);
2184 cpu_ccfence();
2185 parent = chain->parent;
2186 if (parent == NULL)
2187 panic("hammer2_chain_getparent: no parent");
2188 hammer2_chain_ref(parent);
2189 }
2190 return parent;
2191 }
2192
2193 /*
2194 * Take the locked chain and return a locked parent. The chain is unlocked
2195 * and dropped. *chainp is set to the returned parent as a convenience.
2196 * Pass HAMMER2_RESOLVE_* flags in flags.
2197 *
2198 * This will work even if the chain is errored, and the caller can check
2199 * parent->error on return if desired since the parent will be locked.
2200 *
2201 * The chain does NOT need to be stable. We use a tracking structure
2202 * to track the expected parent if the chain is deleted out from under us.
2203 *
2204 * This function handles the lock order reversal.
2205 */
2206 hammer2_chain_t *
hammer2_chain_repparent(hammer2_chain_t ** chainp,int flags)2207 hammer2_chain_repparent(hammer2_chain_t **chainp, int flags)
2208 {
2209 hammer2_chain_t *chain;
2210 hammer2_chain_t *parent;
2211 struct hammer2_reptrack reptrack;
2212 struct hammer2_reptrack **repp;
2213
2214 /*
2215 * Be careful of order, chain must be unlocked before parent
2216 * is locked below to avoid a deadlock. Try it trivially first.
2217 */
2218 chain = *chainp;
2219 parent = chain->parent;
2220 if (parent == NULL) {
2221 hammer2_spin_unex(&chain->core.spin);
2222 panic("hammer2_chain_repparent: no parent");
2223 }
2224 hammer2_chain_ref(parent);
2225 if (hammer2_chain_lock(parent, flags|HAMMER2_RESOLVE_NONBLOCK) == 0) {
2226 hammer2_chain_unlock(chain);
2227 hammer2_chain_drop(chain);
2228 *chainp = parent;
2229
2230 return parent;
2231 }
2232
2233 /*
2234 * Ok, now it gets a bit nasty. There are multiple situations where
2235 * the parent might be in the middle of a deletion, or where the child
2236 * (chain) might be deleted the instant we let go of its lock.
2237 * We can potentially end up in a no-win situation!
2238 *
2239 * In particular, the indirect_maintenance() case can cause these
2240 * situations.
2241 *
2242 * To deal with this we install a reptrack structure in the parent
2243 * This reptrack structure 'owns' the parent ref and will automatically
2244 * migrate to the parent's parent if the parent is deleted permanently.
2245 */
2246 hammer2_spin_init(&reptrack.spin, "h2reptrk");
2247 reptrack.chain = parent;
2248 hammer2_chain_ref(parent); /* for the reptrack */
2249
2250 hammer2_spin_ex(&parent->core.spin);
2251 reptrack.next = parent->core.reptrack;
2252 parent->core.reptrack = &reptrack;
2253 hammer2_spin_unex(&parent->core.spin);
2254
2255 hammer2_chain_unlock(chain);
2256 hammer2_chain_drop(chain);
2257 chain = NULL; /* gone */
2258
2259 /*
2260 * At the top of this loop, chain is gone and parent is refd both
2261 * by us explicitly AND via our reptrack. We are attempting to
2262 * lock parent.
2263 */
2264 for (;;) {
2265 hammer2_chain_lock(parent, flags);
2266
2267 if (reptrack.chain == parent)
2268 break;
2269 hammer2_chain_unlock(parent);
2270 hammer2_chain_drop(parent);
2271
2272 kprintf("hammer2: debug REPTRACK %p->%p\n",
2273 parent, reptrack.chain);
2274 hammer2_spin_ex(&reptrack.spin);
2275 parent = reptrack.chain;
2276 hammer2_chain_ref(parent);
2277 hammer2_spin_unex(&reptrack.spin);
2278 }
2279
2280 /*
2281 * Once parent is locked and matches our reptrack, our reptrack
2282 * will be stable and we have our parent. We can unlink our
2283 * reptrack.
2284 *
2285 * WARNING! Remember that the chain lock might be shared. Chains
2286 * locked shared have stable parent linkages.
2287 */
2288 hammer2_spin_ex(&parent->core.spin);
2289 repp = &parent->core.reptrack;
2290 while (*repp != &reptrack)
2291 repp = &(*repp)->next;
2292 *repp = reptrack.next;
2293 hammer2_spin_unex(&parent->core.spin);
2294
2295 hammer2_chain_drop(parent); /* reptrack ref */
2296 *chainp = parent; /* return parent lock+ref */
2297
2298 return parent;
2299 }
2300
2301 /*
2302 * Dispose of any linked reptrack structures in (chain) by shifting them to
2303 * (parent). Both (chain) and (parent) must be exclusively locked.
2304 *
2305 * This is interlocked against any children of (chain) on the other side.
2306 * No children so remain as-of when this is called so we can test
2307 * core.reptrack without holding the spin-lock.
2308 *
2309 * Used whenever the caller intends to permanently delete chains related
2310 * to topological recursions (BREF_TYPE_INDIRECT, BREF_TYPE_FREEMAP_NODE),
2311 * where the chains underneath the node being deleted are given a new parent
2312 * above the node being deleted.
2313 */
2314 static
2315 void
hammer2_chain_repchange(hammer2_chain_t * parent,hammer2_chain_t * chain)2316 hammer2_chain_repchange(hammer2_chain_t *parent, hammer2_chain_t *chain)
2317 {
2318 struct hammer2_reptrack *reptrack;
2319
2320 KKASSERT(chain->core.live_count == 0 && RB_EMPTY(&chain->core.rbtree));
2321 while (chain->core.reptrack) {
2322 hammer2_spin_ex(&parent->core.spin);
2323 hammer2_spin_ex(&chain->core.spin);
2324 reptrack = chain->core.reptrack;
2325 if (reptrack == NULL) {
2326 hammer2_spin_unex(&chain->core.spin);
2327 hammer2_spin_unex(&parent->core.spin);
2328 break;
2329 }
2330 hammer2_spin_ex(&reptrack->spin);
2331 chain->core.reptrack = reptrack->next;
2332 reptrack->chain = parent;
2333 reptrack->next = parent->core.reptrack;
2334 parent->core.reptrack = reptrack;
2335 hammer2_chain_ref(parent); /* reptrack */
2336
2337 hammer2_spin_unex(&chain->core.spin);
2338 hammer2_spin_unex(&parent->core.spin);
2339 kprintf("hammer2: debug repchange %p %p->%p\n",
2340 reptrack, chain, parent);
2341 hammer2_chain_drop(chain); /* reptrack */
2342 }
2343 }
2344
2345 /*
2346 * Locate the first chain whos key range overlaps (key_beg, key_end) inclusive.
2347 * (*parentp) typically points to an inode but can also point to a related
2348 * indirect block and this function will recurse upwards and find the inode
2349 * or the nearest undeleted indirect block covering the key range.
2350 *
2351 * This function unconditionally sets *errorp, replacing any previous value.
2352 *
2353 * (*parentp) must be exclusive or shared locked (depending on flags) and
2354 * referenced and can be an inode or an existing indirect block within the
2355 * inode.
2356 *
2357 * If (*parent) is errored out, this function will not attempt to recurse
2358 * the radix tree and will return NULL along with an appropriate *errorp.
2359 * If NULL is returned and *errorp is 0, the requested lookup could not be
2360 * located.
2361 *
2362 * On return (*parentp) will be modified to point at the deepest parent chain
2363 * element encountered during the search, as a helper for an insertion or
2364 * deletion.
2365 *
2366 * The new (*parentp) will be locked shared or exclusive (depending on flags),
2367 * and referenced, and the old will be unlocked and dereferenced (no change
2368 * if they are both the same). This is particularly important if the caller
2369 * wishes to insert a new chain, (*parentp) will be set properly even if NULL
2370 * is returned, as long as no error occurred.
2371 *
2372 * The matching chain will be returned locked according to flags.
2373 *
2374 * --
2375 *
2376 * NULL is returned if no match was found, but (*parentp) will still
2377 * potentially be adjusted.
2378 *
2379 * On return (*key_nextp) will point to an iterative value for key_beg.
2380 * (If NULL is returned (*key_nextp) is set to (key_end + 1)).
2381 *
2382 * This function will also recurse up the chain if the key is not within the
2383 * current parent's range. (*parentp) can never be set to NULL. An iteration
2384 * can simply allow (*parentp) to float inside the loop.
2385 *
2386 * NOTE! chain->data is not always resolved. By default it will not be
2387 * resolved for BREF_TYPE_DATA, FREEMAP_NODE, or FREEMAP_LEAF. Use
2388 * HAMMER2_LOOKUP_ALWAYS to force resolution (but be careful w/
2389 * BREF_TYPE_DATA as the device buffer can alias the logical file
2390 * buffer).
2391 */
2392 hammer2_chain_t *
hammer2_chain_lookup(hammer2_chain_t ** parentp,hammer2_key_t * key_nextp,hammer2_key_t key_beg,hammer2_key_t key_end,int * errorp,int flags)2393 hammer2_chain_lookup(hammer2_chain_t **parentp, hammer2_key_t *key_nextp,
2394 hammer2_key_t key_beg, hammer2_key_t key_end,
2395 int *errorp, int flags)
2396 {
2397 hammer2_chain_t *parent;
2398 hammer2_chain_t *chain;
2399 hammer2_blockref_t *base;
2400 hammer2_blockref_t *bref;
2401 hammer2_blockref_t bsave;
2402 hammer2_key_t scan_beg;
2403 hammer2_key_t scan_end;
2404 int count = 0;
2405 int how_always = HAMMER2_RESOLVE_ALWAYS;
2406 int how_maybe = HAMMER2_RESOLVE_MAYBE;
2407 int how;
2408 int generation;
2409 int maxloops = 300000;
2410
2411 if (flags & HAMMER2_LOOKUP_ALWAYS) {
2412 how_maybe = how_always;
2413 how = HAMMER2_RESOLVE_ALWAYS;
2414 } else if (flags & HAMMER2_LOOKUP_NODATA) {
2415 how = HAMMER2_RESOLVE_NEVER;
2416 } else {
2417 how = HAMMER2_RESOLVE_MAYBE;
2418 }
2419 if (flags & HAMMER2_LOOKUP_SHARED) {
2420 how_maybe |= HAMMER2_RESOLVE_SHARED;
2421 how_always |= HAMMER2_RESOLVE_SHARED;
2422 how |= HAMMER2_RESOLVE_SHARED;
2423 }
2424
2425 /*
2426 * Recurse (*parentp) upward if necessary until the parent completely
2427 * encloses the key range or we hit the inode.
2428 *
2429 * Handle races against the flusher deleting indirect nodes on its
2430 * way back up by continuing to recurse upward past the deletion.
2431 */
2432 parent = *parentp;
2433 *errorp = 0;
2434
2435 while (parent->bref.type == HAMMER2_BREF_TYPE_INDIRECT ||
2436 parent->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2437 scan_beg = parent->bref.key;
2438 scan_end = scan_beg +
2439 ((hammer2_key_t)1 << parent->bref.keybits) - 1;
2440 if ((parent->flags & HAMMER2_CHAIN_DELETED) == 0) {
2441 if (key_beg >= scan_beg && key_end <= scan_end)
2442 break;
2443 }
2444 parent = hammer2_chain_repparent(parentp, how_maybe);
2445 }
2446 again:
2447 if (--maxloops == 0)
2448 panic("hammer2_chain_lookup: maxloops");
2449
2450 /*
2451 * MATCHIND case that does not require parent->data (do prior to
2452 * parent->error check).
2453 */
2454 switch(parent->bref.type) {
2455 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2456 case HAMMER2_BREF_TYPE_INDIRECT:
2457 if (flags & HAMMER2_LOOKUP_MATCHIND) {
2458 scan_beg = parent->bref.key;
2459 scan_end = scan_beg +
2460 ((hammer2_key_t)1 << parent->bref.keybits) - 1;
2461 if (key_beg == scan_beg && key_end == scan_end) {
2462 chain = parent;
2463 hammer2_chain_ref(chain);
2464 hammer2_chain_lock(chain, how_maybe);
2465 *key_nextp = scan_end + 1;
2466 goto done;
2467 }
2468 }
2469 break;
2470 default:
2471 break;
2472 }
2473
2474 /*
2475 * No lookup is possible if the parent is errored. We delayed
2476 * this check as long as we could to ensure that the parent backup,
2477 * embedded data, and MATCHIND code could still execute.
2478 */
2479 if (parent->error) {
2480 *errorp = parent->error;
2481 return NULL;
2482 }
2483
2484 /*
2485 * Locate the blockref array. Currently we do a fully associative
2486 * search through the array.
2487 */
2488 switch(parent->bref.type) {
2489 case HAMMER2_BREF_TYPE_INODE:
2490 /*
2491 * Special shortcut for embedded data returns the inode
2492 * itself. Callers must detect this condition and access
2493 * the embedded data (the strategy code does this for us).
2494 *
2495 * This is only applicable to regular files and softlinks.
2496 *
2497 * We need a second lock on parent. Since we already have
2498 * a lock we must pass LOCKAGAIN to prevent unexpected
2499 * blocking (we don't want to block on a second shared
2500 * ref if an exclusive lock is pending)
2501 */
2502 if (parent->data->ipdata.meta.op_flags &
2503 HAMMER2_OPFLAG_DIRECTDATA) {
2504 if (flags & HAMMER2_LOOKUP_NODIRECT) {
2505 chain = NULL;
2506 *key_nextp = key_end + 1;
2507 goto done;
2508 }
2509 hammer2_chain_ref(parent);
2510 hammer2_chain_lock(parent, how_always |
2511 HAMMER2_RESOLVE_LOCKAGAIN);
2512 *key_nextp = key_end + 1;
2513 return (parent);
2514 }
2515 base = &parent->data->ipdata.u.blockset.blockref[0];
2516 count = HAMMER2_SET_COUNT;
2517 break;
2518 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2519 case HAMMER2_BREF_TYPE_INDIRECT:
2520 /*
2521 * Optimize indirect blocks in the INITIAL state to avoid
2522 * I/O.
2523 *
2524 * Debugging: Enter permanent wait state instead of
2525 * panicing on unexpectedly NULL data for the moment.
2526 */
2527 if (parent->flags & HAMMER2_CHAIN_INITIAL) {
2528 base = NULL;
2529 } else {
2530 if (parent->data == NULL) {
2531 kprintf("hammer2: unexpected NULL data "
2532 "on %p\n", parent);
2533 while (1)
2534 tsleep(parent, 0, "xxx", 0);
2535 }
2536 base = &parent->data->npdata[0];
2537 }
2538 count = parent->bytes / sizeof(hammer2_blockref_t);
2539 break;
2540 case HAMMER2_BREF_TYPE_VOLUME:
2541 base = &parent->data->voldata.sroot_blockset.blockref[0];
2542 count = HAMMER2_SET_COUNT;
2543 break;
2544 case HAMMER2_BREF_TYPE_FREEMAP:
2545 base = &parent->data->blkset.blockref[0];
2546 count = HAMMER2_SET_COUNT;
2547 break;
2548 default:
2549 panic("hammer2_chain_lookup: unrecognized "
2550 "blockref(B) type: %d",
2551 parent->bref.type);
2552 base = NULL; /* safety */
2553 count = 0; /* safety */
2554 break;
2555 }
2556
2557 /*
2558 * Merged scan to find next candidate.
2559 *
2560 * hammer2_base_*() functions require the parent->core.live_* fields
2561 * to be synchronized.
2562 *
2563 * We need to hold the spinlock to access the block array and RB tree
2564 * and to interlock chain creation.
2565 */
2566 if ((parent->flags & HAMMER2_CHAIN_COUNTEDBREFS) == 0)
2567 hammer2_chain_countbrefs(parent, base, count);
2568
2569 /*
2570 * Combined search
2571 */
2572 hammer2_spin_ex(&parent->core.spin);
2573 chain = hammer2_combined_find(parent, base, count,
2574 key_nextp,
2575 key_beg, key_end,
2576 &bref);
2577 generation = parent->core.generation;
2578
2579 /*
2580 * Exhausted parent chain, iterate.
2581 */
2582 if (bref == NULL) {
2583 KKASSERT(chain == NULL);
2584 hammer2_spin_unex(&parent->core.spin);
2585 if (key_beg == key_end) /* short cut single-key case */
2586 return (NULL);
2587
2588 /*
2589 * Stop if we reached the end of the iteration.
2590 */
2591 if (parent->bref.type != HAMMER2_BREF_TYPE_INDIRECT &&
2592 parent->bref.type != HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2593 return (NULL);
2594 }
2595
2596 /*
2597 * Calculate next key, stop if we reached the end of the
2598 * iteration, otherwise go up one level and loop.
2599 */
2600 key_beg = parent->bref.key +
2601 ((hammer2_key_t)1 << parent->bref.keybits);
2602 if (key_beg == 0 || key_beg > key_end)
2603 return (NULL);
2604 parent = hammer2_chain_repparent(parentp, how_maybe);
2605 goto again;
2606 }
2607
2608 /*
2609 * Selected from blockref or in-memory chain.
2610 */
2611 bsave = *bref;
2612 if (chain == NULL) {
2613 hammer2_spin_unex(&parent->core.spin);
2614 if (bsave.type == HAMMER2_BREF_TYPE_INDIRECT ||
2615 bsave.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2616 chain = hammer2_chain_get(parent, generation,
2617 &bsave, how_maybe);
2618 } else {
2619 chain = hammer2_chain_get(parent, generation,
2620 &bsave, how);
2621 }
2622 if (chain == NULL)
2623 goto again;
2624 } else {
2625 hammer2_chain_ref(chain);
2626 hammer2_spin_unex(&parent->core.spin);
2627
2628 /*
2629 * chain is referenced but not locked. We must lock the
2630 * chain to obtain definitive state.
2631 */
2632 if (bsave.type == HAMMER2_BREF_TYPE_INDIRECT ||
2633 bsave.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2634 hammer2_chain_lock(chain, how_maybe);
2635 } else {
2636 hammer2_chain_lock(chain, how);
2637 }
2638 KKASSERT(chain->parent == parent);
2639 }
2640 if (bcmp(&bsave, &chain->bref, sizeof(bsave)) ||
2641 chain->parent != parent) {
2642 hammer2_chain_unlock(chain);
2643 hammer2_chain_drop(chain);
2644 chain = NULL; /* SAFETY */
2645 goto again;
2646 }
2647
2648
2649 /*
2650 * Skip deleted chains (XXX cache 'i' end-of-block-array? XXX)
2651 *
2652 * NOTE: Chain's key range is not relevant as there might be
2653 * one-offs within the range that are not deleted.
2654 *
2655 * NOTE: Lookups can race delete-duplicate because
2656 * delete-duplicate does not lock the parent's core
2657 * (they just use the spinlock on the core).
2658 */
2659 if (chain->flags & HAMMER2_CHAIN_DELETED) {
2660 kprintf("skip deleted chain %016jx.%02x key=%016jx\n",
2661 chain->bref.data_off, chain->bref.type,
2662 chain->bref.key);
2663 hammer2_chain_unlock(chain);
2664 hammer2_chain_drop(chain);
2665 chain = NULL; /* SAFETY */
2666 key_beg = *key_nextp;
2667 if (key_beg == 0 || key_beg > key_end)
2668 return(NULL);
2669 goto again;
2670 }
2671
2672 /*
2673 * If the chain element is an indirect block it becomes the new
2674 * parent and we loop on it. We must maintain our top-down locks
2675 * to prevent the flusher from interfering (i.e. doing a
2676 * delete-duplicate and leaving us recursing down a deleted chain).
2677 *
2678 * The parent always has to be locked with at least RESOLVE_MAYBE
2679 * so we can access its data. It might need a fixup if the caller
2680 * passed incompatible flags. Be careful not to cause a deadlock
2681 * as a data-load requires an exclusive lock.
2682 *
2683 * If HAMMER2_LOOKUP_MATCHIND is set and the indirect block's key
2684 * range is within the requested key range we return the indirect
2685 * block and do NOT loop. This is usually only used to acquire
2686 * freemap nodes.
2687 */
2688 if (chain->bref.type == HAMMER2_BREF_TYPE_INDIRECT ||
2689 chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2690 hammer2_chain_unlock(parent);
2691 hammer2_chain_drop(parent);
2692 *parentp = parent = chain;
2693 chain = NULL; /* SAFETY */
2694 goto again;
2695 }
2696 done:
2697 /*
2698 * All done, return the locked chain.
2699 *
2700 * If the caller does not want a locked chain, replace the lock with
2701 * a ref. Perhaps this can eventually be optimized to not obtain the
2702 * lock in the first place for situations where the data does not
2703 * need to be resolved.
2704 *
2705 * NOTE! A chain->error must be tested by the caller upon return.
2706 * *errorp is only set based on issues which occur while
2707 * trying to reach the chain.
2708 */
2709 return (chain);
2710 }
2711
2712 /*
2713 * After having issued a lookup we can iterate all matching keys.
2714 *
2715 * If chain is non-NULL we continue the iteration from just after it's index.
2716 *
2717 * If chain is NULL we assume the parent was exhausted and continue the
2718 * iteration at the next parent.
2719 *
2720 * If a fatal error occurs (typically an I/O error), a dummy chain is
2721 * returned with chain->error and error-identifying information set. This
2722 * chain will assert if you try to do anything fancy with it.
2723 *
2724 * XXX Depending on where the error occurs we should allow continued iteration.
2725 *
2726 * parent must be locked on entry and remains locked throughout. chain's
2727 * lock status must match flags. Chain is always at least referenced.
2728 *
2729 * WARNING! The MATCHIND flag does not apply to this function.
2730 */
2731 hammer2_chain_t *
hammer2_chain_next(hammer2_chain_t ** parentp,hammer2_chain_t * chain,hammer2_key_t * key_nextp,hammer2_key_t key_beg,hammer2_key_t key_end,int * errorp,int flags)2732 hammer2_chain_next(hammer2_chain_t **parentp, hammer2_chain_t *chain,
2733 hammer2_key_t *key_nextp,
2734 hammer2_key_t key_beg, hammer2_key_t key_end,
2735 int *errorp, int flags)
2736 {
2737 hammer2_chain_t *parent;
2738 int how_maybe;
2739
2740 /*
2741 * Calculate locking flags for upward recursion.
2742 */
2743 how_maybe = HAMMER2_RESOLVE_MAYBE;
2744 if (flags & HAMMER2_LOOKUP_SHARED)
2745 how_maybe |= HAMMER2_RESOLVE_SHARED;
2746
2747 parent = *parentp;
2748 *errorp = 0;
2749
2750 /*
2751 * Calculate the next index and recalculate the parent if necessary.
2752 */
2753 if (chain) {
2754 key_beg = chain->bref.key +
2755 ((hammer2_key_t)1 << chain->bref.keybits);
2756 hammer2_chain_unlock(chain);
2757 hammer2_chain_drop(chain);
2758
2759 /*
2760 * chain invalid past this point, but we can still do a
2761 * pointer comparison w/parent.
2762 *
2763 * Any scan where the lookup returned degenerate data embedded
2764 * in the inode has an invalid index and must terminate.
2765 */
2766 if (chain == parent)
2767 return(NULL);
2768 if (key_beg == 0 || key_beg > key_end)
2769 return(NULL);
2770 chain = NULL;
2771 } else if (parent->bref.type != HAMMER2_BREF_TYPE_INDIRECT &&
2772 parent->bref.type != HAMMER2_BREF_TYPE_FREEMAP_NODE) {
2773 /*
2774 * We reached the end of the iteration.
2775 */
2776 return (NULL);
2777 } else {
2778 /*
2779 * Continue iteration with next parent unless the current
2780 * parent covers the range.
2781 *
2782 * (This also handles the case of a deleted, empty indirect
2783 * node).
2784 */
2785 key_beg = parent->bref.key +
2786 ((hammer2_key_t)1 << parent->bref.keybits);
2787 if (key_beg == 0 || key_beg > key_end)
2788 return (NULL);
2789 parent = hammer2_chain_repparent(parentp, how_maybe);
2790 }
2791
2792 /*
2793 * And execute
2794 */
2795 return (hammer2_chain_lookup(parentp, key_nextp,
2796 key_beg, key_end,
2797 errorp, flags));
2798 }
2799
2800 /*
2801 * Caller wishes to iterate chains under parent, loading new chains into
2802 * chainp. Caller must initialize *chainp to NULL and *firstp to 1, and
2803 * then call hammer2_chain_scan() repeatedly until a non-zero return.
2804 * During the scan, *firstp will be set to 0 and (*chainp) will be replaced
2805 * with the returned chain for the scan. The returned *chainp will be
2806 * locked and referenced. Any prior contents will be unlocked and dropped.
2807 *
2808 * Caller should check the return value. A normal scan EOF will return
2809 * exactly HAMMER2_ERROR_EOF. Any other non-zero value indicates an
2810 * error trying to access parent data. Any error in the returned chain
2811 * must be tested separately by the caller.
2812 *
2813 * (*chainp) is dropped on each scan, but will only be set if the returned
2814 * element itself can recurse. Leaf elements are NOT resolved, loaded, or
2815 * returned via *chainp. The caller will get their bref only.
2816 *
2817 * The raw scan function is similar to lookup/next but does not seek to a key.
2818 * Blockrefs are iterated via first_bref = (parent, NULL) and
2819 * next_chain = (parent, bref).
2820 *
2821 * The passed-in parent must be locked and its data resolved. The function
2822 * nominally returns a locked and referenced *chainp != NULL for chains
2823 * the caller might need to recurse on (and will dipose of any *chainp passed
2824 * in). The caller must check the chain->bref.type either way.
2825 */
2826 int
hammer2_chain_scan(hammer2_chain_t * parent,hammer2_chain_t ** chainp,hammer2_blockref_t * bref,int * firstp,int flags)2827 hammer2_chain_scan(hammer2_chain_t *parent, hammer2_chain_t **chainp,
2828 hammer2_blockref_t *bref, int *firstp,
2829 int flags)
2830 {
2831 hammer2_blockref_t *base;
2832 hammer2_blockref_t *bref_ptr;
2833 hammer2_key_t key;
2834 hammer2_key_t next_key;
2835 hammer2_chain_t *chain = NULL;
2836 int count = 0;
2837 int how;
2838 int generation;
2839 int maxloops = 300000;
2840 int error;
2841
2842 error = 0;
2843
2844 /*
2845 * Scan flags borrowed from lookup.
2846 */
2847 if (flags & HAMMER2_LOOKUP_ALWAYS) {
2848 how = HAMMER2_RESOLVE_ALWAYS;
2849 } else if (flags & HAMMER2_LOOKUP_NODATA) {
2850 how = HAMMER2_RESOLVE_NEVER;
2851 } else {
2852 how = HAMMER2_RESOLVE_MAYBE;
2853 }
2854 if (flags & HAMMER2_LOOKUP_SHARED) {
2855 how |= HAMMER2_RESOLVE_SHARED;
2856 }
2857
2858 /*
2859 * Calculate key to locate first/next element, unlocking the previous
2860 * element as we go. Be careful, the key calculation can overflow.
2861 *
2862 * (also reset bref to NULL)
2863 */
2864 if (*firstp) {
2865 key = 0;
2866 *firstp = 0;
2867 } else {
2868 key = bref->key + ((hammer2_key_t)1 << bref->keybits);
2869 if ((chain = *chainp) != NULL) {
2870 *chainp = NULL;
2871 hammer2_chain_unlock(chain);
2872 hammer2_chain_drop(chain);
2873 chain = NULL;
2874 }
2875 if (key == 0) {
2876 error |= HAMMER2_ERROR_EOF;
2877 goto done;
2878 }
2879 }
2880
2881 again:
2882 if (parent->error) {
2883 error = parent->error;
2884 goto done;
2885 }
2886 if (--maxloops == 0)
2887 panic("hammer2_chain_scan: maxloops");
2888
2889 /*
2890 * Locate the blockref array. Currently we do a fully associative
2891 * search through the array.
2892 */
2893 switch(parent->bref.type) {
2894 case HAMMER2_BREF_TYPE_INODE:
2895 /*
2896 * An inode with embedded data has no sub-chains.
2897 *
2898 * WARNING! Bulk scan code may pass a static chain marked
2899 * as BREF_TYPE_INODE with a copy of the volume
2900 * root blockset to snapshot the volume.
2901 */
2902 if (parent->data->ipdata.meta.op_flags &
2903 HAMMER2_OPFLAG_DIRECTDATA) {
2904 error |= HAMMER2_ERROR_EOF;
2905 goto done;
2906 }
2907 base = &parent->data->ipdata.u.blockset.blockref[0];
2908 count = HAMMER2_SET_COUNT;
2909 break;
2910 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2911 case HAMMER2_BREF_TYPE_INDIRECT:
2912 /*
2913 * Optimize indirect blocks in the INITIAL state to avoid
2914 * I/O.
2915 */
2916 if (parent->flags & HAMMER2_CHAIN_INITIAL) {
2917 base = NULL;
2918 } else {
2919 if (parent->data == NULL)
2920 panic("parent->data is NULL");
2921 base = &parent->data->npdata[0];
2922 }
2923 count = parent->bytes / sizeof(hammer2_blockref_t);
2924 break;
2925 case HAMMER2_BREF_TYPE_VOLUME:
2926 base = &parent->data->voldata.sroot_blockset.blockref[0];
2927 count = HAMMER2_SET_COUNT;
2928 break;
2929 case HAMMER2_BREF_TYPE_FREEMAP:
2930 base = &parent->data->blkset.blockref[0];
2931 count = HAMMER2_SET_COUNT;
2932 break;
2933 default:
2934 panic("hammer2_chain_scan: unrecognized blockref type: %d",
2935 parent->bref.type);
2936 base = NULL; /* safety */
2937 count = 0; /* safety */
2938 break;
2939 }
2940
2941 /*
2942 * Merged scan to find next candidate.
2943 *
2944 * hammer2_base_*() functions require the parent->core.live_* fields
2945 * to be synchronized.
2946 *
2947 * We need to hold the spinlock to access the block array and RB tree
2948 * and to interlock chain creation.
2949 */
2950 if ((parent->flags & HAMMER2_CHAIN_COUNTEDBREFS) == 0)
2951 hammer2_chain_countbrefs(parent, base, count);
2952
2953 next_key = 0;
2954 bref_ptr = NULL;
2955 hammer2_spin_ex(&parent->core.spin);
2956 chain = hammer2_combined_find(parent, base, count,
2957 &next_key,
2958 key, HAMMER2_KEY_MAX,
2959 &bref_ptr);
2960 generation = parent->core.generation;
2961
2962 /*
2963 * Exhausted parent chain, we're done.
2964 */
2965 if (bref_ptr == NULL) {
2966 hammer2_spin_unex(&parent->core.spin);
2967 KKASSERT(chain == NULL);
2968 error |= HAMMER2_ERROR_EOF;
2969 goto done;
2970 }
2971
2972 /*
2973 * Copy into the supplied stack-based blockref.
2974 */
2975 *bref = *bref_ptr;
2976
2977 /*
2978 * Selected from blockref or in-memory chain.
2979 */
2980 if (chain == NULL) {
2981 switch(bref->type) {
2982 case HAMMER2_BREF_TYPE_INODE:
2983 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2984 case HAMMER2_BREF_TYPE_INDIRECT:
2985 case HAMMER2_BREF_TYPE_VOLUME:
2986 case HAMMER2_BREF_TYPE_FREEMAP:
2987 /*
2988 * Recursion, always get the chain
2989 */
2990 hammer2_spin_unex(&parent->core.spin);
2991 chain = hammer2_chain_get(parent, generation,
2992 bref, how);
2993 if (chain == NULL)
2994 goto again;
2995 break;
2996 default:
2997 /*
2998 * No recursion, do not waste time instantiating
2999 * a chain, just iterate using the bref.
3000 */
3001 hammer2_spin_unex(&parent->core.spin);
3002 break;
3003 }
3004 } else {
3005 /*
3006 * Recursion or not we need the chain in order to supply
3007 * the bref.
3008 */
3009 hammer2_chain_ref(chain);
3010 hammer2_spin_unex(&parent->core.spin);
3011 hammer2_chain_lock(chain, how);
3012 }
3013 if (chain &&
3014 (bcmp(bref, &chain->bref, sizeof(*bref)) ||
3015 chain->parent != parent)) {
3016 hammer2_chain_unlock(chain);
3017 hammer2_chain_drop(chain);
3018 chain = NULL;
3019 goto again;
3020 }
3021
3022 /*
3023 * Skip deleted chains (XXX cache 'i' end-of-block-array? XXX)
3024 *
3025 * NOTE: chain's key range is not relevant as there might be
3026 * one-offs within the range that are not deleted.
3027 *
3028 * NOTE: XXX this could create problems with scans used in
3029 * situations other than mount-time recovery.
3030 *
3031 * NOTE: Lookups can race delete-duplicate because
3032 * delete-duplicate does not lock the parent's core
3033 * (they just use the spinlock on the core).
3034 */
3035 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) {
3036 hammer2_chain_unlock(chain);
3037 hammer2_chain_drop(chain);
3038 chain = NULL;
3039
3040 key = next_key;
3041 if (key == 0) {
3042 error |= HAMMER2_ERROR_EOF;
3043 goto done;
3044 }
3045 goto again;
3046 }
3047
3048 done:
3049 /*
3050 * All done, return the bref or NULL, supply chain if necessary.
3051 */
3052 if (chain)
3053 *chainp = chain;
3054 return (error);
3055 }
3056
3057 /*
3058 * Create and return a new hammer2 system memory structure of the specified
3059 * key, type and size and insert it under (*parentp). This is a full
3060 * insertion, based on the supplied key/keybits, and may involve creating
3061 * indirect blocks and moving other chains around via delete/duplicate.
3062 *
3063 * This call can be made with parent == NULL as long as a non -1 methods
3064 * is supplied. hmp must also be supplied in this situation (otherwise
3065 * hmp is extracted from the supplied parent). The chain will be detached
3066 * from the topology. A later call with both parent and chain can be made
3067 * to attach it.
3068 *
3069 * THE CALLER MUST HAVE ALREADY PROPERLY SEEKED (*parentp) TO THE INSERTION
3070 * POINT SANS ANY REQUIRED INDIRECT BLOCK CREATIONS DUE TO THE ARRAY BEING
3071 * FULL. This typically means that the caller is creating the chain after
3072 * doing a hammer2_chain_lookup().
3073 *
3074 * (*parentp) must be exclusive locked and may be replaced on return
3075 * depending on how much work the function had to do.
3076 *
3077 * (*parentp) must not be errored or this function will assert.
3078 *
3079 * (*chainp) usually starts out NULL and returns the newly created chain,
3080 * but if the caller desires the caller may allocate a disconnected chain
3081 * and pass it in instead.
3082 *
3083 * This function should NOT be used to insert INDIRECT blocks. It is
3084 * typically used to create/insert inodes and data blocks.
3085 *
3086 * Caller must pass-in an exclusively locked parent the new chain is to
3087 * be inserted under, and optionally pass-in a disconnected, exclusively
3088 * locked chain to insert (else we create a new chain). The function will
3089 * adjust (*parentp) as necessary, create or connect the chain, and
3090 * return an exclusively locked chain in *chainp.
3091 *
3092 * When creating a PFSROOT inode under the super-root, pmp is typically NULL
3093 * and will be reassigned.
3094 *
3095 * NOTE: returns HAMMER_ERROR_* flags
3096 */
3097 int
hammer2_chain_create(hammer2_chain_t ** parentp,hammer2_chain_t ** chainp,hammer2_dev_t * hmp,hammer2_pfs_t * pmp,int methods,hammer2_key_t key,int keybits,int type,size_t bytes,hammer2_tid_t mtid,hammer2_off_t dedup_off,int flags)3098 hammer2_chain_create(hammer2_chain_t **parentp, hammer2_chain_t **chainp,
3099 hammer2_dev_t *hmp, hammer2_pfs_t *pmp, int methods,
3100 hammer2_key_t key, int keybits, int type, size_t bytes,
3101 hammer2_tid_t mtid, hammer2_off_t dedup_off, int flags)
3102 {
3103 hammer2_chain_t *chain;
3104 hammer2_chain_t *parent;
3105 hammer2_blockref_t *base;
3106 hammer2_blockref_t dummy;
3107 int allocated = 0;
3108 int error = 0;
3109 int count;
3110 int maxloops = 300000;
3111
3112 /*
3113 * Topology may be crossing a PFS boundary.
3114 */
3115 parent = *parentp;
3116 if (parent) {
3117 KKASSERT(hammer2_mtx_owned(&parent->lock));
3118 KKASSERT(parent->error == 0);
3119 hmp = parent->hmp;
3120 }
3121 chain = *chainp;
3122
3123 if (chain == NULL) {
3124 /*
3125 * First allocate media space and construct the dummy bref,
3126 * then allocate the in-memory chain structure. Set the
3127 * INITIAL flag for fresh chains which do not have embedded
3128 * data.
3129 */
3130 bzero(&dummy, sizeof(dummy));
3131 dummy.type = type;
3132 dummy.key = key;
3133 dummy.keybits = keybits;
3134 dummy.data_off = hammer2_getradix(bytes);
3135
3136 /*
3137 * Inherit methods from parent by default. Primarily used
3138 * for BREF_TYPE_DATA. Non-data types *must* be set to
3139 * a non-NONE check algorithm.
3140 */
3141 if (methods == HAMMER2_METH_DEFAULT)
3142 dummy.methods = parent->bref.methods;
3143 else
3144 dummy.methods = (uint8_t)methods;
3145
3146 if (type != HAMMER2_BREF_TYPE_DATA &&
3147 HAMMER2_DEC_CHECK(dummy.methods) == HAMMER2_CHECK_NONE) {
3148 dummy.methods |=
3149 HAMMER2_ENC_CHECK(HAMMER2_CHECK_DEFAULT);
3150 }
3151
3152 chain = hammer2_chain_alloc(hmp, pmp, &dummy);
3153
3154 /*
3155 * Lock the chain manually, chain_lock will load the chain
3156 * which we do NOT want to do. (note: chain->refs is set
3157 * to 1 by chain_alloc() for us, but lockcnt is not).
3158 */
3159 chain->lockcnt = 1;
3160 hammer2_mtx_ex(&chain->lock);
3161 allocated = 1;
3162
3163 /*
3164 * Set INITIAL to optimize I/O. The flag will generally be
3165 * processed when we call hammer2_chain_modify().
3166 */
3167 switch(type) {
3168 case HAMMER2_BREF_TYPE_VOLUME:
3169 case HAMMER2_BREF_TYPE_FREEMAP:
3170 panic("hammer2_chain_create: called with volume type");
3171 break;
3172 case HAMMER2_BREF_TYPE_INDIRECT:
3173 panic("hammer2_chain_create: cannot be used to"
3174 "create indirect block");
3175 break;
3176 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3177 panic("hammer2_chain_create: cannot be used to"
3178 "create freemap root or node");
3179 break;
3180 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
3181 KKASSERT(bytes == sizeof(chain->data->bmdata));
3182 /* fall through */
3183 case HAMMER2_BREF_TYPE_DIRENT:
3184 case HAMMER2_BREF_TYPE_INODE:
3185 case HAMMER2_BREF_TYPE_DATA:
3186 default:
3187 /*
3188 * leave chain->data NULL, set INITIAL
3189 */
3190 KKASSERT(chain->data == NULL);
3191 atomic_set_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
3192 break;
3193 }
3194 } else {
3195 /*
3196 * We are reattaching a previously deleted chain, possibly
3197 * under a new parent and possibly with a new key/keybits.
3198 * The chain does not have to be in a modified state. The
3199 * UPDATE flag will be set later on in this routine.
3200 *
3201 * Do NOT mess with the current state of the INITIAL flag.
3202 */
3203 chain->bref.key = key;
3204 chain->bref.keybits = keybits;
3205 if (chain->flags & HAMMER2_CHAIN_DELETED)
3206 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_DELETED);
3207 KKASSERT(chain->parent == NULL);
3208 }
3209
3210 /*
3211 * Set the appropriate bref flag if requested.
3212 *
3213 * NOTE! Callers can call this function to move chains without
3214 * knowing about special flags, so don't clear bref flags
3215 * here!
3216 */
3217 if (flags & HAMMER2_INSERT_PFSROOT)
3218 chain->bref.flags |= HAMMER2_BREF_FLAG_PFSROOT;
3219
3220 if (parent == NULL)
3221 goto skip;
3222
3223 /*
3224 * Calculate how many entries we have in the blockref array and
3225 * determine if an indirect block is required when inserting into
3226 * the parent.
3227 */
3228 again:
3229 if (--maxloops == 0)
3230 panic("hammer2_chain_create: maxloops");
3231
3232 switch(parent->bref.type) {
3233 case HAMMER2_BREF_TYPE_INODE:
3234 if ((parent->data->ipdata.meta.op_flags &
3235 HAMMER2_OPFLAG_DIRECTDATA) != 0) {
3236 kprintf("hammer2: parent set for direct-data! "
3237 "pkey=%016jx ckey=%016jx\n",
3238 parent->bref.key,
3239 chain->bref.key);
3240 }
3241 KKASSERT((parent->data->ipdata.meta.op_flags &
3242 HAMMER2_OPFLAG_DIRECTDATA) == 0);
3243 KKASSERT(parent->data != NULL);
3244 base = &parent->data->ipdata.u.blockset.blockref[0];
3245 count = HAMMER2_SET_COUNT;
3246 break;
3247 case HAMMER2_BREF_TYPE_INDIRECT:
3248 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3249 if (parent->flags & HAMMER2_CHAIN_INITIAL)
3250 base = NULL;
3251 else
3252 base = &parent->data->npdata[0];
3253 count = parent->bytes / sizeof(hammer2_blockref_t);
3254 break;
3255 case HAMMER2_BREF_TYPE_VOLUME:
3256 KKASSERT(parent->data != NULL);
3257 base = &parent->data->voldata.sroot_blockset.blockref[0];
3258 count = HAMMER2_SET_COUNT;
3259 break;
3260 case HAMMER2_BREF_TYPE_FREEMAP:
3261 KKASSERT(parent->data != NULL);
3262 base = &parent->data->blkset.blockref[0];
3263 count = HAMMER2_SET_COUNT;
3264 break;
3265 default:
3266 panic("hammer2_chain_create: unrecognized blockref type: %d",
3267 parent->bref.type);
3268 base = NULL;
3269 count = 0;
3270 break;
3271 }
3272
3273 /*
3274 * Make sure we've counted the brefs
3275 */
3276 if ((parent->flags & HAMMER2_CHAIN_COUNTEDBREFS) == 0)
3277 hammer2_chain_countbrefs(parent, base, count);
3278
3279 KASSERT(parent->core.live_count >= 0 &&
3280 parent->core.live_count <= count,
3281 ("bad live_count %d/%d (%02x, %d)",
3282 parent->core.live_count, count,
3283 parent->bref.type, parent->bytes));
3284
3285 /*
3286 * If no free blockref could be found we must create an indirect
3287 * block and move a number of blockrefs into it. With the parent
3288 * locked we can safely lock each child in order to delete+duplicate
3289 * it without causing a deadlock.
3290 *
3291 * This may return the new indirect block or the old parent depending
3292 * on where the key falls. NULL is returned on error.
3293 */
3294 if (parent->core.live_count == count) {
3295 hammer2_chain_t *nparent;
3296
3297 KKASSERT((flags & HAMMER2_INSERT_SAMEPARENT) == 0);
3298
3299 nparent = hammer2_chain_create_indirect(parent, key, keybits,
3300 mtid, type, &error);
3301 if (nparent == NULL) {
3302 if (allocated)
3303 hammer2_chain_drop(chain);
3304 chain = NULL;
3305 goto done;
3306 }
3307 if (parent != nparent) {
3308 hammer2_chain_unlock(parent);
3309 hammer2_chain_drop(parent);
3310 parent = *parentp = nparent;
3311 }
3312 goto again;
3313 }
3314
3315 /*
3316 * fall through if parent, or skip to here if no parent.
3317 */
3318 skip:
3319 if (chain->flags & HAMMER2_CHAIN_DELETED)
3320 kprintf("Inserting deleted chain @%016jx\n",
3321 chain->bref.key);
3322
3323 /*
3324 * Link the chain into its parent.
3325 */
3326 if (chain->parent != NULL)
3327 panic("hammer2: hammer2_chain_create: chain already connected");
3328 KKASSERT(chain->parent == NULL);
3329 if (parent) {
3330 KKASSERT(parent->core.live_count < count);
3331 hammer2_chain_insert(parent, chain,
3332 HAMMER2_CHAIN_INSERT_SPIN |
3333 HAMMER2_CHAIN_INSERT_LIVE,
3334 0);
3335 }
3336
3337 if (allocated) {
3338 /*
3339 * Mark the newly created chain modified. This will cause
3340 * UPDATE to be set and process the INITIAL flag.
3341 *
3342 * Device buffers are not instantiated for DATA elements
3343 * as these are handled by logical buffers.
3344 *
3345 * Indirect and freemap node indirect blocks are handled
3346 * by hammer2_chain_create_indirect() and not by this
3347 * function.
3348 *
3349 * Data for all other bref types is expected to be
3350 * instantiated (INODE, LEAF).
3351 */
3352 switch(chain->bref.type) {
3353 case HAMMER2_BREF_TYPE_DATA:
3354 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
3355 case HAMMER2_BREF_TYPE_DIRENT:
3356 case HAMMER2_BREF_TYPE_INODE:
3357 error = hammer2_chain_modify(chain, mtid, dedup_off,
3358 HAMMER2_MODIFY_OPTDATA);
3359 break;
3360 default:
3361 /*
3362 * Remaining types are not supported by this function.
3363 * In particular, INDIRECT and LEAF_NODE types are
3364 * handled by create_indirect().
3365 */
3366 panic("hammer2_chain_create: bad type: %d",
3367 chain->bref.type);
3368 /* NOT REACHED */
3369 break;
3370 }
3371 } else {
3372 /*
3373 * When reconnecting a chain we must set UPDATE and
3374 * setflush so the flush recognizes that it must update
3375 * the bref in the parent.
3376 */
3377 if ((chain->flags & HAMMER2_CHAIN_UPDATE) == 0)
3378 atomic_set_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
3379 }
3380
3381 /*
3382 * We must setflush(parent) to ensure that it recurses through to
3383 * chain. setflush(chain) might not work because ONFLUSH is possibly
3384 * already set in the chain (so it won't recurse up to set it in the
3385 * parent).
3386 */
3387 if (parent)
3388 hammer2_chain_setflush(parent);
3389
3390 done:
3391 *chainp = chain;
3392
3393 return (error);
3394 }
3395
3396 /*
3397 * Move the chain from its old parent to a new parent. The chain must have
3398 * already been deleted or already disconnected (or never associated) with
3399 * a parent. The chain is reassociated with the new parent and the deleted
3400 * flag will be cleared (no longer deleted). The chain's modification state
3401 * is not altered.
3402 *
3403 * THE CALLER MUST HAVE ALREADY PROPERLY SEEKED (parent) TO THE INSERTION
3404 * POINT SANS ANY REQUIRED INDIRECT BLOCK CREATIONS DUE TO THE ARRAY BEING
3405 * FULL. This typically means that the caller is creating the chain after
3406 * doing a hammer2_chain_lookup().
3407 *
3408 * Neither (parent) or (chain) can be errored.
3409 *
3410 * If (parent) is non-NULL then the chain is inserted under the parent.
3411 *
3412 * If (parent) is NULL then the newly duplicated chain is not inserted
3413 * anywhere, similar to if it had just been chain_alloc()'d (suitable for
3414 * passing into hammer2_chain_create() after this function returns).
3415 *
3416 * WARNING! This function calls create which means it can insert indirect
3417 * blocks. This can cause other unrelated chains in the parent to
3418 * be moved to a newly inserted indirect block in addition to the
3419 * specific chain.
3420 */
3421 void
hammer2_chain_rename(hammer2_chain_t ** parentp,hammer2_chain_t * chain,hammer2_tid_t mtid,int flags)3422 hammer2_chain_rename(hammer2_chain_t **parentp, hammer2_chain_t *chain,
3423 hammer2_tid_t mtid, int flags)
3424 {
3425 hammer2_blockref_t *bref;
3426 hammer2_chain_t *parent;
3427
3428 /*
3429 * WARNING! We should never resolve DATA to device buffers
3430 * (XXX allow it if the caller did?), and since
3431 * we currently do not have the logical buffer cache
3432 * buffer in-hand to fix its cached physical offset
3433 * we also force the modify code to not COW it. XXX
3434 *
3435 * NOTE! We allow error'd chains to be renamed. The bref itself
3436 * is good and can be renamed. The content, however, may
3437 * be inaccessible.
3438 */
3439 KKASSERT(chain->parent == NULL);
3440 /*KKASSERT(chain->error == 0); allow */
3441 bref = &chain->bref;
3442
3443 /*
3444 * If parent is not NULL the duplicated chain will be entered under
3445 * the parent and the UPDATE bit set to tell flush to update
3446 * the blockref.
3447 *
3448 * We must setflush(parent) to ensure that it recurses through to
3449 * chain. setflush(chain) might not work because ONFLUSH is possibly
3450 * already set in the chain (so it won't recurse up to set it in the
3451 * parent).
3452 *
3453 * Having both chains locked is extremely important for atomicy.
3454 */
3455 if (parentp && (parent = *parentp) != NULL) {
3456 KKASSERT(hammer2_mtx_owned(&parent->lock));
3457 KKASSERT(parent->refs > 0);
3458 KKASSERT(parent->error == 0);
3459
3460 hammer2_chain_create(parentp, &chain, NULL, chain->pmp,
3461 HAMMER2_METH_DEFAULT,
3462 bref->key, bref->keybits, bref->type,
3463 chain->bytes, mtid, 0, flags);
3464 KKASSERT(chain->flags & HAMMER2_CHAIN_UPDATE);
3465 hammer2_chain_setflush(*parentp);
3466 }
3467 }
3468
3469 /*
3470 * This works in tandem with delete_obref() to install a blockref in
3471 * (typically) an indirect block that is associated with the chain being
3472 * moved to *parentp.
3473 *
3474 * The reason we need this function is that the caller needs to maintain
3475 * the blockref as it was, and not generate a new blockref for what might
3476 * be a modified chain. Otherwise stuff will leak into the flush that
3477 * the flush code's FLUSH_INODE_STOP flag is unable to catch.
3478 *
3479 * It is EXTREMELY important that we properly set CHAIN_BLKMAPUPD and
3480 * CHAIN_UPDATE. We must set BLKMAPUPD if the bref does not match, and
3481 * we must clear CHAIN_UPDATE (that was likely set by the chain_rename) if
3482 * it does. Otherwise we can end up in a situation where H2 is unable to
3483 * clean up the in-memory chain topology.
3484 *
3485 * The reason for this is that flushes do not generally flush through
3486 * BREF_TYPE_INODE chains and depend on a hammer2_inode_t queued to syncq
3487 * or sideq to properly flush and dispose of the related inode chain's flags.
3488 * Situations where the inode is not actually modified by the frontend,
3489 * but where we have to move the related chains around as we insert or cleanup
3490 * indirect blocks, can leave us with a 'dirty' (non-disposable) in-memory
3491 * inode chain that does not have a hammer2_inode_t associated with it.
3492 */
3493 static void
hammer2_chain_rename_obref(hammer2_chain_t ** parentp,hammer2_chain_t * chain,hammer2_tid_t mtid,int flags,hammer2_blockref_t * obref)3494 hammer2_chain_rename_obref(hammer2_chain_t **parentp, hammer2_chain_t *chain,
3495 hammer2_tid_t mtid, int flags,
3496 hammer2_blockref_t *obref)
3497 {
3498 hammer2_chain_rename(parentp, chain, mtid, flags);
3499
3500 if (obref->type != HAMMER2_BREF_TYPE_EMPTY) {
3501 hammer2_blockref_t *tbase;
3502 int tcount;
3503
3504 KKASSERT((chain->flags & HAMMER2_CHAIN_BLKMAPPED) == 0);
3505 hammer2_chain_modify(*parentp, mtid, 0, 0);
3506 tbase = hammer2_chain_base_and_count(*parentp, &tcount);
3507 hammer2_base_insert(*parentp, tbase, tcount, chain, obref);
3508 if (bcmp(obref, &chain->bref, sizeof(chain->bref))) {
3509 atomic_set_int(&chain->flags, HAMMER2_CHAIN_BLKMAPUPD |
3510 HAMMER2_CHAIN_UPDATE);
3511 } else {
3512 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
3513 }
3514 }
3515 }
3516
3517 /*
3518 * Helper function for deleting chains.
3519 *
3520 * The chain is removed from the live view (the RBTREE) as well as the parent's
3521 * blockmap. Both chain and its parent must be locked.
3522 *
3523 * parent may not be errored. chain can be errored.
3524 */
3525 static int
_hammer2_chain_delete_helper(hammer2_chain_t * parent,hammer2_chain_t * chain,hammer2_tid_t mtid,int flags,hammer2_blockref_t * obref)3526 _hammer2_chain_delete_helper(hammer2_chain_t *parent, hammer2_chain_t *chain,
3527 hammer2_tid_t mtid, int flags,
3528 hammer2_blockref_t *obref)
3529 {
3530 int error = 0;
3531
3532 KKASSERT((chain->flags & HAMMER2_CHAIN_DELETED) == 0);
3533 KKASSERT(chain->parent == parent);
3534
3535 if (chain->flags & HAMMER2_CHAIN_BLKMAPPED) {
3536 /*
3537 * Chain is blockmapped, so there must be a parent.
3538 * Atomically remove the chain from the parent and remove
3539 * the blockmap entry. The parent must be set modified
3540 * to remove the blockmap entry.
3541 */
3542 hammer2_blockref_t *base;
3543 int count;
3544
3545 KKASSERT(parent != NULL);
3546 KKASSERT(parent->error == 0);
3547 KKASSERT((parent->flags & HAMMER2_CHAIN_INITIAL) == 0);
3548 error = hammer2_chain_modify(parent, mtid, 0, 0);
3549 if (error)
3550 goto done;
3551
3552 /*
3553 * Calculate blockmap pointer
3554 */
3555 KKASSERT(chain->flags & HAMMER2_CHAIN_ONRBTREE);
3556 hammer2_spin_ex(&chain->core.spin);
3557 hammer2_spin_ex(&parent->core.spin);
3558
3559 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DELETED);
3560 atomic_add_int(&parent->core.live_count, -1);
3561 ++parent->core.generation;
3562 RB_REMOVE(hammer2_chain_tree, &parent->core.rbtree, chain);
3563 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE);
3564 --parent->core.chain_count;
3565 chain->parent = NULL;
3566
3567 switch(parent->bref.type) {
3568 case HAMMER2_BREF_TYPE_INODE:
3569 /*
3570 * Access the inode's block array. However, there
3571 * is no block array if the inode is flagged
3572 * DIRECTDATA.
3573 */
3574 if (parent->data &&
3575 (parent->data->ipdata.meta.op_flags &
3576 HAMMER2_OPFLAG_DIRECTDATA) == 0) {
3577 base =
3578 &parent->data->ipdata.u.blockset.blockref[0];
3579 } else {
3580 base = NULL;
3581 }
3582 count = HAMMER2_SET_COUNT;
3583 break;
3584 case HAMMER2_BREF_TYPE_INDIRECT:
3585 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3586 if (parent->data)
3587 base = &parent->data->npdata[0];
3588 else
3589 base = NULL;
3590 count = parent->bytes / sizeof(hammer2_blockref_t);
3591 break;
3592 case HAMMER2_BREF_TYPE_VOLUME:
3593 base = &parent->data->voldata.
3594 sroot_blockset.blockref[0];
3595 count = HAMMER2_SET_COUNT;
3596 break;
3597 case HAMMER2_BREF_TYPE_FREEMAP:
3598 base = &parent->data->blkset.blockref[0];
3599 count = HAMMER2_SET_COUNT;
3600 break;
3601 default:
3602 base = NULL;
3603 count = 0;
3604 panic("_hammer2_chain_delete_helper: "
3605 "unrecognized blockref type: %d",
3606 parent->bref.type);
3607 break;
3608 }
3609
3610 /*
3611 * delete blockmapped chain from its parent.
3612 */
3613 if (base) {
3614 hammer2_base_delete(parent, base, count, chain, obref);
3615 }
3616 hammer2_spin_unex(&parent->core.spin);
3617 hammer2_spin_unex(&chain->core.spin);
3618 } else if (chain->flags & HAMMER2_CHAIN_ONRBTREE) {
3619 /*
3620 * Chain is not blockmapped but a parent is present.
3621 * Atomically remove the chain from the parent. There is
3622 * no blockmap entry to remove.
3623 */
3624 hammer2_spin_ex(&chain->core.spin);
3625 hammer2_spin_ex(&parent->core.spin);
3626 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DELETED);
3627 atomic_add_int(&parent->core.live_count, -1);
3628 ++parent->core.generation;
3629 RB_REMOVE(hammer2_chain_tree, &parent->core.rbtree, chain);
3630 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE);
3631 --parent->core.chain_count;
3632 chain->parent = NULL;
3633 hammer2_spin_unex(&parent->core.spin);
3634 hammer2_spin_unex(&chain->core.spin);
3635 } else {
3636 /*
3637 * Chain is not blockmapped and has no parent. This
3638 * is a degenerate case.
3639 */
3640 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DELETED);
3641 }
3642 done:
3643 return error;
3644 }
3645
3646 /*
3647 * Create an indirect block that covers one or more of the elements in the
3648 * current parent. Either returns the existing parent with no locking or
3649 * ref changes or returns the new indirect block locked and referenced
3650 * and leaving the original parent lock/ref intact as well.
3651 *
3652 * If an error occurs, NULL is returned and *errorp is set to the H2 error.
3653 *
3654 * The returned chain depends on where the specified key falls.
3655 *
3656 * The key/keybits for the indirect mode only needs to follow three rules:
3657 *
3658 * (1) That all elements underneath it fit within its key space and
3659 *
3660 * (2) That all elements outside it are outside its key space.
3661 *
3662 * (3) When creating the new indirect block any elements in the current
3663 * parent that fit within the new indirect block's keyspace must be
3664 * moved into the new indirect block.
3665 *
3666 * (4) The keyspace chosen for the inserted indirect block CAN cover a wider
3667 * keyspace the the current parent, but lookup/iteration rules will
3668 * ensure (and must ensure) that rule (2) for all parents leading up
3669 * to the nearest inode or the root volume header is adhered to. This
3670 * is accomplished by always recursing through matching keyspaces in
3671 * the hammer2_chain_lookup() and hammer2_chain_next() API.
3672 *
3673 * The current implementation calculates the current worst-case keyspace by
3674 * iterating the current parent and then divides it into two halves, choosing
3675 * whichever half has the most elements (not necessarily the half containing
3676 * the requested key).
3677 *
3678 * We can also opt to use the half with the least number of elements. This
3679 * causes lower-numbered keys (aka logical file offsets) to recurse through
3680 * fewer indirect blocks and higher-numbered keys to recurse through more.
3681 * This also has the risk of not moving enough elements to the new indirect
3682 * block and being forced to create several indirect blocks before the element
3683 * can be inserted.
3684 *
3685 * Must be called with an exclusively locked parent.
3686 *
3687 * NOTE: *errorp set to HAMMER_ERROR_* flags
3688 */
3689 static int hammer2_chain_indkey_freemap(hammer2_chain_t *parent,
3690 hammer2_key_t *keyp, int keybits,
3691 hammer2_blockref_t *base, int count);
3692 static int hammer2_chain_indkey_file(hammer2_chain_t *parent,
3693 hammer2_key_t *keyp, int keybits,
3694 hammer2_blockref_t *base, int count,
3695 int ncount);
3696 static int hammer2_chain_indkey_dir(hammer2_chain_t *parent,
3697 hammer2_key_t *keyp, int keybits,
3698 hammer2_blockref_t *base, int count,
3699 int ncount);
3700 static
3701 hammer2_chain_t *
hammer2_chain_create_indirect(hammer2_chain_t * parent,hammer2_key_t create_key,int create_bits,hammer2_tid_t mtid,int for_type,int * errorp)3702 hammer2_chain_create_indirect(hammer2_chain_t *parent,
3703 hammer2_key_t create_key, int create_bits,
3704 hammer2_tid_t mtid, int for_type, int *errorp)
3705 {
3706 hammer2_dev_t *hmp;
3707 hammer2_blockref_t *base;
3708 hammer2_blockref_t *bref;
3709 hammer2_blockref_t bsave;
3710 hammer2_blockref_t dummy;
3711 hammer2_chain_t *chain;
3712 hammer2_chain_t *ichain;
3713 hammer2_key_t key = create_key;
3714 hammer2_key_t key_beg;
3715 hammer2_key_t key_end;
3716 hammer2_key_t key_next;
3717 int keybits = create_bits;
3718 int count;
3719 int ncount;
3720 int nbytes;
3721 int loops;
3722 int error;
3723 int reason;
3724 int generation;
3725 int maxloops = 300000;
3726
3727 /*
3728 * Calculate the base blockref pointer or NULL if the chain
3729 * is known to be empty. We need to calculate the array count
3730 * for RB lookups either way.
3731 */
3732 hmp = parent->hmp;
3733 KKASSERT(hammer2_mtx_owned(&parent->lock));
3734
3735 /*
3736 * Pre-modify the parent now to avoid having to deal with error
3737 * processing if we tried to later (in the middle of our loop).
3738 *
3739 * We are going to be moving bref's around, the indirect blocks
3740 * cannot be in an initial state. Do not pass MODIFY_OPTDATA.
3741 */
3742 *errorp = hammer2_chain_modify(parent, mtid, 0, 0);
3743 if (*errorp) {
3744 kprintf("hammer2_chain_create_indirect: error %08x %s\n",
3745 *errorp, hammer2_error_str(*errorp));
3746 return NULL;
3747 }
3748 KKASSERT((parent->flags & HAMMER2_CHAIN_INITIAL) == 0);
3749
3750 /*hammer2_chain_modify(&parent, HAMMER2_MODIFY_OPTDATA);*/
3751 base = hammer2_chain_base_and_count(parent, &count);
3752
3753 /*
3754 * How big should our new indirect block be? It has to be at least
3755 * as large as its parent for splits to work properly.
3756 *
3757 * The freemap uses a specific indirect block size. The number of
3758 * levels are built dynamically and ultimately depend on the size
3759 * volume. Because freemap blocks are taken from the reserved areas
3760 * of the volume our goal is efficiency (fewer levels) and not so
3761 * much to save disk space.
3762 *
3763 * The first indirect block level for a directory usually uses
3764 * HAMMER2_IND_BYTES_MIN (4KB = 32 directory entries). Due to
3765 * the hash mechanism, this typically gives us a nominal
3766 * 32 * 4 entries with one level of indirection.
3767 *
3768 * We use HAMMER2_IND_BYTES_NOM (16KB = 128 blockrefs) for FILE
3769 * indirect blocks. The initial 4 entries in the inode gives us
3770 * 256KB. Up to 4 indirect blocks gives us 32MB. Three levels
3771 * of indirection gives us 137GB, and so forth. H2 can support
3772 * huge file sizes but they are not typical, so we try to stick
3773 * with compactness and do not use a larger indirect block size.
3774 *
3775 * We could use 64KB (PBUFSIZE), giving us 512 blockrefs, but
3776 * due to the way indirect blocks are created this usually winds
3777 * up being extremely inefficient for small files. Even though
3778 * 16KB requires more levels of indirection for very large files,
3779 * the 16KB records can be ganged together into 64KB DIOs.
3780 */
3781 if (for_type == HAMMER2_BREF_TYPE_FREEMAP_NODE ||
3782 for_type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) {
3783 nbytes = HAMMER2_FREEMAP_LEVELN_PSIZE;
3784 } else if (parent->bref.type == HAMMER2_BREF_TYPE_INODE) {
3785 if (parent->data->ipdata.meta.type ==
3786 HAMMER2_OBJTYPE_DIRECTORY)
3787 nbytes = HAMMER2_IND_BYTES_MIN; /* 4KB = 32 entries */
3788 else
3789 nbytes = HAMMER2_IND_BYTES_NOM; /* 16KB = ~8MB file */
3790
3791 } else {
3792 nbytes = HAMMER2_IND_BYTES_NOM;
3793 }
3794 if (nbytes < count * sizeof(hammer2_blockref_t)) {
3795 KKASSERT(for_type != HAMMER2_BREF_TYPE_FREEMAP_NODE &&
3796 for_type != HAMMER2_BREF_TYPE_FREEMAP_LEAF);
3797 nbytes = count * sizeof(hammer2_blockref_t);
3798 }
3799 ncount = nbytes / sizeof(hammer2_blockref_t);
3800
3801 /*
3802 * When creating an indirect block for a freemap node or leaf
3803 * the key/keybits must be fitted to static radix levels because
3804 * particular radix levels use particular reserved blocks in the
3805 * related zone.
3806 *
3807 * This routine calculates the key/radix of the indirect block
3808 * we need to create, and whether it is on the high-side or the
3809 * low-side.
3810 */
3811 switch(for_type) {
3812 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
3813 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
3814 keybits = hammer2_chain_indkey_freemap(parent, &key, keybits,
3815 base, count);
3816 break;
3817 case HAMMER2_BREF_TYPE_DATA:
3818 keybits = hammer2_chain_indkey_file(parent, &key, keybits,
3819 base, count, ncount);
3820 break;
3821 case HAMMER2_BREF_TYPE_DIRENT:
3822 case HAMMER2_BREF_TYPE_INODE:
3823 keybits = hammer2_chain_indkey_dir(parent, &key, keybits,
3824 base, count, ncount);
3825 break;
3826 default:
3827 panic("illegal indirect block for bref type %d", for_type);
3828 break;
3829 }
3830
3831 /*
3832 * Normalize the key for the radix being represented, keeping the
3833 * high bits and throwing away the low bits.
3834 */
3835 key &= ~(((hammer2_key_t)1 << keybits) - 1);
3836
3837 /*
3838 * Ok, create our new indirect block
3839 */
3840 bzero(&dummy, sizeof(dummy));
3841 if (for_type == HAMMER2_BREF_TYPE_FREEMAP_NODE ||
3842 for_type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) {
3843 dummy.type = HAMMER2_BREF_TYPE_FREEMAP_NODE;
3844 } else {
3845 dummy.type = HAMMER2_BREF_TYPE_INDIRECT;
3846 }
3847 dummy.key = key;
3848 dummy.keybits = keybits;
3849 dummy.data_off = hammer2_getradix(nbytes);
3850 dummy.methods =
3851 HAMMER2_ENC_CHECK(HAMMER2_DEC_CHECK(parent->bref.methods)) |
3852 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
3853
3854 ichain = hammer2_chain_alloc(hmp, parent->pmp, &dummy);
3855 atomic_set_int(&ichain->flags, HAMMER2_CHAIN_INITIAL);
3856 hammer2_chain_lock(ichain, HAMMER2_RESOLVE_MAYBE);
3857 /* ichain has one ref at this point */
3858
3859 /*
3860 * We have to mark it modified to allocate its block, but use
3861 * OPTDATA to allow it to remain in the INITIAL state. Otherwise
3862 * it won't be acted upon by the flush code.
3863 *
3864 * XXX remove OPTDATA, we need a fully initialized indirect block to
3865 * be able to move the original blockref.
3866 */
3867 *errorp = hammer2_chain_modify(ichain, mtid, 0, 0);
3868 if (*errorp) {
3869 kprintf("hammer2_chain_create_indirect: error %08x %s\n",
3870 *errorp, hammer2_error_str(*errorp));
3871 hammer2_chain_unlock(ichain);
3872 hammer2_chain_drop(ichain);
3873 return NULL;
3874 }
3875 KKASSERT((ichain->flags & HAMMER2_CHAIN_INITIAL) == 0);
3876
3877 /*
3878 * Iterate the original parent and move the matching brefs into
3879 * the new indirect block.
3880 *
3881 * XXX handle flushes.
3882 */
3883 key_beg = 0;
3884 key_end = HAMMER2_KEY_MAX;
3885 key_next = 0; /* avoid gcc warnings */
3886 hammer2_spin_ex(&parent->core.spin);
3887 loops = 0;
3888 reason = 0;
3889
3890 for (;;) {
3891 /*
3892 * Parent may have been modified, relocating its block array.
3893 * Reload the base pointer.
3894 */
3895 base = hammer2_chain_base_and_count(parent, &count);
3896
3897 if (++loops > 100000) {
3898 hammer2_spin_unex(&parent->core.spin);
3899 panic("excessive loops r=%d p=%p base/count %p:%d %016jx\n",
3900 reason, parent, base, count, key_next);
3901 }
3902
3903 /*
3904 * NOTE: spinlock stays intact, returned chain (if not NULL)
3905 * is not referenced or locked which means that we
3906 * cannot safely check its flagged / deletion status
3907 * until we lock it.
3908 */
3909 chain = hammer2_combined_find(parent, base, count,
3910 &key_next,
3911 key_beg, key_end,
3912 &bref);
3913 generation = parent->core.generation;
3914 if (bref == NULL)
3915 break;
3916 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
3917
3918 /*
3919 * Skip keys that are not within the key/radix of the new
3920 * indirect block. They stay in the parent.
3921 */
3922 if (rounddown2(key ^ bref->key, (hammer2_key_t)1 << keybits) != 0) {
3923 goto next_key_spinlocked;
3924 }
3925
3926 /*
3927 * Load the new indirect block by acquiring the related
3928 * chains (potentially from media as it might not be
3929 * in-memory). Then move it to the new parent (ichain).
3930 *
3931 * chain is referenced but not locked. We must lock the
3932 * chain to obtain definitive state.
3933 */
3934 bsave = *bref;
3935 if (chain) {
3936 /*
3937 * Use chain already present in the RBTREE
3938 */
3939 hammer2_chain_ref(chain);
3940 hammer2_spin_unex(&parent->core.spin);
3941 hammer2_chain_lock(chain, HAMMER2_RESOLVE_NEVER);
3942 } else {
3943 /*
3944 * Get chain for blockref element. _get returns NULL
3945 * on insertion race.
3946 */
3947 hammer2_spin_unex(&parent->core.spin);
3948 chain = hammer2_chain_get(parent, generation, &bsave,
3949 HAMMER2_RESOLVE_NEVER);
3950 if (chain == NULL) {
3951 reason = 1;
3952 hammer2_spin_ex(&parent->core.spin);
3953 continue;
3954 }
3955 }
3956
3957 /*
3958 * This is always live so if the chain has been deleted
3959 * we raced someone and we have to retry.
3960 *
3961 * NOTE: Lookups can race delete-duplicate because
3962 * delete-duplicate does not lock the parent's core
3963 * (they just use the spinlock on the core).
3964 *
3965 * (note reversed logic for this one)
3966 */
3967 if (bcmp(&bsave, &chain->bref, sizeof(bsave)) ||
3968 chain->parent != parent ||
3969 (chain->flags & HAMMER2_CHAIN_DELETED)) {
3970 hammer2_chain_unlock(chain);
3971 hammer2_chain_drop(chain);
3972 if (hammer2_debug & 0x0040) {
3973 kprintf("LOST PARENT RETRY "
3974 "RETRY (%p,%p)->%p %08x\n",
3975 parent, chain->parent, chain, chain->flags);
3976 }
3977 hammer2_spin_ex(&parent->core.spin);
3978 continue;
3979 }
3980
3981 /*
3982 * Shift the chain to the indirect block.
3983 *
3984 * WARNING! The (parent, chain) deletion may modify the parent
3985 * and invalidate the base pointer.
3986 *
3987 * WARNING! Parent must already be marked modified, so we
3988 * can assume that chain_delete always suceeds.
3989 *
3990 * WARNING! hammer2_chain_repchange() does not have to be
3991 * called (and doesn't work anyway because we are
3992 * only doing a partial shift). A recursion that is
3993 * in-progress can continue at the current parent
3994 * and will be able to properly find its next key.
3995 */
3996 error = hammer2_chain_delete_obref(parent, chain, mtid, 0,
3997 &bsave);
3998 KKASSERT(error == 0);
3999 hammer2_chain_rename_obref(&ichain, chain, mtid, 0, &bsave);
4000 hammer2_chain_unlock(chain);
4001 hammer2_chain_drop(chain);
4002 KKASSERT(parent->refs > 0);
4003 chain = NULL;
4004 base = NULL; /* safety */
4005 hammer2_spin_ex(&parent->core.spin);
4006 next_key_spinlocked:
4007 if (--maxloops == 0)
4008 panic("hammer2_chain_create_indirect: maxloops");
4009 reason = 4;
4010 if (key_next == 0 || key_next > key_end)
4011 break;
4012 key_beg = key_next;
4013 /* loop */
4014 }
4015 hammer2_spin_unex(&parent->core.spin);
4016
4017 /*
4018 * Insert the new indirect block into the parent now that we've
4019 * cleared out some entries in the parent. We calculated a good
4020 * insertion index in the loop above (ichain->index).
4021 *
4022 * We don't have to set UPDATE here because we mark ichain
4023 * modified down below (so the normal modified -> flush -> set-moved
4024 * sequence applies).
4025 *
4026 * The insertion shouldn't race as this is a completely new block
4027 * and the parent is locked.
4028 */
4029 base = NULL; /* safety, parent modify may change address */
4030 KKASSERT((ichain->flags & HAMMER2_CHAIN_ONRBTREE) == 0);
4031 KKASSERT(parent->core.live_count < count);
4032 hammer2_chain_insert(parent, ichain,
4033 HAMMER2_CHAIN_INSERT_SPIN |
4034 HAMMER2_CHAIN_INSERT_LIVE,
4035 0);
4036
4037 /*
4038 * Make sure flushes propogate after our manual insertion.
4039 */
4040 hammer2_chain_setflush(ichain);
4041 hammer2_chain_setflush(parent);
4042
4043 /*
4044 * Figure out what to return.
4045 */
4046 if (rounddown2(create_key ^ key, (hammer2_key_t)1 << keybits) != 0) {
4047 /*
4048 * Key being created is outside the key range,
4049 * return the original parent.
4050 */
4051 hammer2_chain_unlock(ichain);
4052 hammer2_chain_drop(ichain);
4053 } else {
4054 /*
4055 * Otherwise its in the range, return the new parent.
4056 * (leave both the new and old parent locked).
4057 */
4058 parent = ichain;
4059 }
4060
4061 return(parent);
4062 }
4063
4064 /*
4065 * Do maintenance on an indirect chain. Both parent and chain are locked.
4066 *
4067 * Returns non-zero if (chain) is deleted, either due to being empty or
4068 * because its children were safely moved into the parent.
4069 */
4070 int
hammer2_chain_indirect_maintenance(hammer2_chain_t * parent,hammer2_chain_t * chain)4071 hammer2_chain_indirect_maintenance(hammer2_chain_t *parent,
4072 hammer2_chain_t *chain)
4073 {
4074 hammer2_blockref_t *chain_base;
4075 hammer2_blockref_t *base;
4076 hammer2_blockref_t *bref;
4077 hammer2_blockref_t bsave;
4078 hammer2_key_t key_next;
4079 hammer2_key_t key_beg;
4080 hammer2_key_t key_end;
4081 hammer2_chain_t *sub;
4082 int chain_count;
4083 int count;
4084 int error;
4085 int generation;
4086
4087 /*
4088 * Make sure we have an accurate live_count
4089 */
4090 if ((chain->flags & (HAMMER2_CHAIN_INITIAL |
4091 HAMMER2_CHAIN_COUNTEDBREFS)) == 0) {
4092 base = &chain->data->npdata[0];
4093 count = chain->bytes / sizeof(hammer2_blockref_t);
4094 hammer2_chain_countbrefs(chain, base, count);
4095 }
4096
4097 /*
4098 * If the indirect block is empty we can delete it.
4099 * (ignore deletion error)
4100 */
4101 if (chain->core.live_count == 0 && RB_EMPTY(&chain->core.rbtree)) {
4102 hammer2_chain_delete(parent, chain,
4103 chain->bref.modify_tid,
4104 HAMMER2_DELETE_PERMANENT);
4105 hammer2_chain_repchange(parent, chain);
4106 return 1;
4107 }
4108
4109 base = hammer2_chain_base_and_count(parent, &count);
4110
4111 if ((parent->flags & (HAMMER2_CHAIN_INITIAL |
4112 HAMMER2_CHAIN_COUNTEDBREFS)) == 0) {
4113 hammer2_chain_countbrefs(parent, base, count);
4114 }
4115
4116 /*
4117 * Determine if we can collapse chain into parent, calculate
4118 * hysteresis for chain emptiness.
4119 */
4120 if (parent->core.live_count + chain->core.live_count - 1 > count)
4121 return 0;
4122 chain_count = chain->bytes / sizeof(hammer2_blockref_t);
4123 if (chain->core.live_count > chain_count * 3 / 4)
4124 return 0;
4125
4126 /*
4127 * Ok, theoretically we can collapse chain's contents into
4128 * parent. chain is locked, but any in-memory children of chain
4129 * are not. For this to work, we must be able to dispose of any
4130 * in-memory children of chain.
4131 *
4132 * For now require that there are no in-memory children of chain.
4133 *
4134 * WARNING! Both chain and parent must remain locked across this
4135 * entire operation.
4136 */
4137
4138 /*
4139 * Parent must be marked modified. Don't try to collapse it if we
4140 * can't mark it modified. Once modified, destroy chain to make room
4141 * and to get rid of what will be a conflicting key (this is included
4142 * in the calculation above). Finally, move the children of chain
4143 * into chain's parent.
4144 *
4145 * This order creates an accounting problem for bref.embed.stats
4146 * because we destroy chain before we remove its children. Any
4147 * elements whos blockref is already synchronized will be counted
4148 * twice. To deal with the problem we clean out chain's stats prior
4149 * to deleting it.
4150 */
4151 error = hammer2_chain_modify(parent, 0, 0, 0);
4152 if (error) {
4153 krateprintf(&krate_h2me, "hammer2: indirect_maint: %s\n",
4154 hammer2_error_str(error));
4155 return 0;
4156 }
4157 error = hammer2_chain_modify(chain, chain->bref.modify_tid, 0, 0);
4158 if (error) {
4159 krateprintf(&krate_h2me, "hammer2: indirect_maint: %s\n",
4160 hammer2_error_str(error));
4161 return 0;
4162 }
4163
4164 chain->bref.embed.stats.inode_count = 0;
4165 chain->bref.embed.stats.data_count = 0;
4166 error = hammer2_chain_delete(parent, chain,
4167 chain->bref.modify_tid,
4168 HAMMER2_DELETE_PERMANENT);
4169 KKASSERT(error == 0);
4170
4171 /*
4172 * The combined_find call requires core.spin to be held. One would
4173 * think there wouldn't be any conflicts since we hold chain
4174 * exclusively locked, but the caching mechanism for 0-ref children
4175 * does not require a chain lock.
4176 */
4177 hammer2_spin_ex(&chain->core.spin);
4178
4179 key_next = 0;
4180 key_beg = 0;
4181 key_end = HAMMER2_KEY_MAX;
4182 for (;;) {
4183 chain_base = &chain->data->npdata[0];
4184 chain_count = chain->bytes / sizeof(hammer2_blockref_t);
4185 sub = hammer2_combined_find(chain, chain_base, chain_count,
4186 &key_next,
4187 key_beg, key_end,
4188 &bref);
4189 generation = chain->core.generation;
4190 if (bref == NULL)
4191 break;
4192 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
4193
4194 bsave = *bref;
4195 if (sub) {
4196 hammer2_chain_ref(sub);
4197 hammer2_spin_unex(&chain->core.spin);
4198 hammer2_chain_lock(sub, HAMMER2_RESOLVE_NEVER);
4199 } else {
4200 hammer2_spin_unex(&chain->core.spin);
4201 sub = hammer2_chain_get(chain, generation, &bsave,
4202 HAMMER2_RESOLVE_NEVER);
4203 if (sub == NULL) {
4204 hammer2_spin_ex(&chain->core.spin);
4205 continue;
4206 }
4207 }
4208 if (bcmp(&bsave, &sub->bref, sizeof(bsave)) ||
4209 sub->parent != chain ||
4210 (sub->flags & HAMMER2_CHAIN_DELETED)) {
4211 hammer2_chain_unlock(sub);
4212 hammer2_chain_drop(sub);
4213 hammer2_spin_ex(&chain->core.spin);
4214 sub = NULL; /* safety */
4215 continue;
4216 }
4217 error = hammer2_chain_delete_obref(chain, sub,
4218 sub->bref.modify_tid, 0,
4219 &bsave);
4220 KKASSERT(error == 0);
4221 hammer2_chain_rename_obref(&parent, sub,
4222 sub->bref.modify_tid,
4223 HAMMER2_INSERT_SAMEPARENT, &bsave);
4224 hammer2_chain_unlock(sub);
4225 hammer2_chain_drop(sub);
4226 hammer2_spin_ex(&chain->core.spin);
4227
4228 if (key_next == 0)
4229 break;
4230 key_beg = key_next;
4231 }
4232 hammer2_spin_unex(&chain->core.spin);
4233
4234 hammer2_chain_repchange(parent, chain);
4235
4236 return 1;
4237 }
4238
4239 /*
4240 * Freemap indirect blocks
4241 *
4242 * Calculate the keybits and highside/lowside of the freemap node the
4243 * caller is creating.
4244 *
4245 * This routine will specify the next higher-level freemap key/radix
4246 * representing the lowest-ordered set. By doing so, eventually all
4247 * low-ordered sets will be moved one level down.
4248 *
4249 * We have to be careful here because the freemap reserves a limited
4250 * number of blocks for a limited number of levels. So we can't just
4251 * push indiscriminately.
4252 */
4253 int
hammer2_chain_indkey_freemap(hammer2_chain_t * parent,hammer2_key_t * keyp,int keybits,hammer2_blockref_t * base,int count)4254 hammer2_chain_indkey_freemap(hammer2_chain_t *parent, hammer2_key_t *keyp,
4255 int keybits, hammer2_blockref_t *base, int count)
4256 {
4257 hammer2_chain_t *chain;
4258 hammer2_blockref_t *bref;
4259 hammer2_key_t key;
4260 hammer2_key_t key_beg;
4261 hammer2_key_t key_end;
4262 hammer2_key_t key_next;
4263 int maxloops = 300000;
4264
4265 key = *keyp;
4266 keybits = 64;
4267
4268 /*
4269 * Calculate the range of keys in the array being careful to skip
4270 * slots which are overridden with a deletion.
4271 */
4272 key_beg = 0;
4273 key_end = HAMMER2_KEY_MAX;
4274 hammer2_spin_ex(&parent->core.spin);
4275
4276 for (;;) {
4277 if (--maxloops == 0) {
4278 panic("indkey_freemap shit %p %p:%d\n",
4279 parent, base, count);
4280 }
4281 chain = hammer2_combined_find(parent, base, count,
4282 &key_next,
4283 key_beg, key_end,
4284 &bref);
4285
4286 /*
4287 * Exhausted search
4288 */
4289 if (bref == NULL)
4290 break;
4291
4292 /*
4293 * Skip deleted chains.
4294 */
4295 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) {
4296 if (key_next == 0 || key_next > key_end)
4297 break;
4298 key_beg = key_next;
4299 continue;
4300 }
4301
4302 /*
4303 * Use the full live (not deleted) element for the scan
4304 * iteration. HAMMER2 does not allow partial replacements.
4305 *
4306 * XXX should be built into hammer2_combined_find().
4307 */
4308 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
4309
4310 if (keybits > bref->keybits) {
4311 key = bref->key;
4312 keybits = bref->keybits;
4313 } else if (keybits == bref->keybits && bref->key < key) {
4314 key = bref->key;
4315 }
4316 if (key_next == 0)
4317 break;
4318 key_beg = key_next;
4319 }
4320 hammer2_spin_unex(&parent->core.spin);
4321
4322 /*
4323 * Return the keybits for a higher-level FREEMAP_NODE covering
4324 * this node.
4325 */
4326 switch(keybits) {
4327 case HAMMER2_FREEMAP_LEVEL0_RADIX:
4328 keybits = HAMMER2_FREEMAP_LEVEL1_RADIX;
4329 break;
4330 case HAMMER2_FREEMAP_LEVEL1_RADIX:
4331 keybits = HAMMER2_FREEMAP_LEVEL2_RADIX;
4332 break;
4333 case HAMMER2_FREEMAP_LEVEL2_RADIX:
4334 keybits = HAMMER2_FREEMAP_LEVEL3_RADIX;
4335 break;
4336 case HAMMER2_FREEMAP_LEVEL3_RADIX:
4337 keybits = HAMMER2_FREEMAP_LEVEL4_RADIX;
4338 break;
4339 case HAMMER2_FREEMAP_LEVEL4_RADIX:
4340 keybits = HAMMER2_FREEMAP_LEVEL5_RADIX;
4341 break;
4342 case HAMMER2_FREEMAP_LEVEL5_RADIX:
4343 panic("hammer2_chain_indkey_freemap: level too high");
4344 break;
4345 default:
4346 panic("hammer2_chain_indkey_freemap: bad radix");
4347 break;
4348 }
4349 *keyp = key;
4350
4351 return (keybits);
4352 }
4353
4354 /*
4355 * File indirect blocks
4356 *
4357 * Calculate the key/keybits for the indirect block to create by scanning
4358 * existing keys. The key being created is also passed in *keyp and can be
4359 * inside or outside the indirect block. Regardless, the indirect block
4360 * must hold at least two keys in order to guarantee sufficient space.
4361 *
4362 * We use a modified version of the freemap's fixed radix tree, but taylored
4363 * for file data. Basically we configure an indirect block encompassing the
4364 * smallest key.
4365 */
4366 static int
hammer2_chain_indkey_file(hammer2_chain_t * parent,hammer2_key_t * keyp,int keybits,hammer2_blockref_t * base,int count,int ncount)4367 hammer2_chain_indkey_file(hammer2_chain_t *parent, hammer2_key_t *keyp,
4368 int keybits, hammer2_blockref_t *base, int count,
4369 int ncount)
4370 {
4371 hammer2_chain_t *chain;
4372 hammer2_blockref_t *bref;
4373 hammer2_key_t key;
4374 hammer2_key_t key_beg;
4375 hammer2_key_t key_end;
4376 hammer2_key_t key_next;
4377 int nradix;
4378 int maxloops = 300000;
4379
4380 key = *keyp;
4381 keybits = 64;
4382
4383 /*
4384 * Calculate the range of keys in the array being careful to skip
4385 * slots which are overridden with a deletion.
4386 *
4387 * Locate the smallest key.
4388 */
4389 key_beg = 0;
4390 key_end = HAMMER2_KEY_MAX;
4391 hammer2_spin_ex(&parent->core.spin);
4392
4393 for (;;) {
4394 if (--maxloops == 0) {
4395 panic("indkey_freemap shit %p %p:%d\n",
4396 parent, base, count);
4397 }
4398 chain = hammer2_combined_find(parent, base, count,
4399 &key_next,
4400 key_beg, key_end,
4401 &bref);
4402
4403 /*
4404 * Exhausted search
4405 */
4406 if (bref == NULL)
4407 break;
4408
4409 /*
4410 * Skip deleted chains.
4411 */
4412 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) {
4413 if (key_next == 0 || key_next > key_end)
4414 break;
4415 key_beg = key_next;
4416 continue;
4417 }
4418
4419 /*
4420 * Use the full live (not deleted) element for the scan
4421 * iteration. HAMMER2 does not allow partial replacements.
4422 *
4423 * XXX should be built into hammer2_combined_find().
4424 */
4425 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
4426
4427 if (keybits > bref->keybits) {
4428 key = bref->key;
4429 keybits = bref->keybits;
4430 } else if (keybits == bref->keybits && bref->key < key) {
4431 key = bref->key;
4432 }
4433 if (key_next == 0)
4434 break;
4435 key_beg = key_next;
4436 }
4437 hammer2_spin_unex(&parent->core.spin);
4438
4439 /*
4440 * Calculate the static keybits for a higher-level indirect block
4441 * that contains the key.
4442 */
4443 *keyp = key;
4444
4445 switch(ncount) {
4446 case HAMMER2_IND_COUNT_MIN:
4447 nradix = HAMMER2_IND_RADIX_MIN - HAMMER2_BLOCKREF_RADIX;
4448 break;
4449 case HAMMER2_IND_COUNT_NOM:
4450 nradix = HAMMER2_IND_RADIX_NOM - HAMMER2_BLOCKREF_RADIX;
4451 break;
4452 case HAMMER2_IND_COUNT_MAX:
4453 nradix = HAMMER2_IND_RADIX_MAX - HAMMER2_BLOCKREF_RADIX;
4454 break;
4455 default:
4456 panic("bad ncount %d\n", ncount);
4457 nradix = 0;
4458 break;
4459 }
4460
4461 /*
4462 * The largest radix that can be returned for an indirect block is
4463 * 63 bits. (The largest practical indirect block radix is actually
4464 * 62 bits because the top-level inode or volume root contains four
4465 * entries, but allow 63 to be returned).
4466 */
4467 if (nradix >= 64)
4468 nradix = 63;
4469
4470 return keybits + nradix;
4471 }
4472
4473 #if 1
4474
4475 /*
4476 * Directory indirect blocks.
4477 *
4478 * Covers both the inode index (directory of inodes), and directory contents
4479 * (filenames hardlinked to inodes).
4480 *
4481 * Because directory keys are hashed we generally try to cut the space in
4482 * half. We accomodate the inode index (which tends to have linearly
4483 * increasing inode numbers) by ensuring that the keyspace is at least large
4484 * enough to fill up the indirect block being created.
4485 */
4486 static int
hammer2_chain_indkey_dir(hammer2_chain_t * parent,hammer2_key_t * keyp,int keybits,hammer2_blockref_t * base,int count,int ncount)4487 hammer2_chain_indkey_dir(hammer2_chain_t *parent, hammer2_key_t *keyp,
4488 int keybits, hammer2_blockref_t *base, int count,
4489 int ncount)
4490 {
4491 hammer2_blockref_t *bref;
4492 hammer2_chain_t *chain;
4493 hammer2_key_t key_beg;
4494 hammer2_key_t key_end;
4495 hammer2_key_t key_next;
4496 hammer2_key_t key;
4497 int nkeybits;
4498 int locount;
4499 int hicount;
4500 int maxloops = 300000;
4501
4502 /*
4503 * NOTE: We can't take a shortcut here anymore for inodes because
4504 * the root directory can contain a mix of inodes and directory
4505 * entries (we used to just return 63 if parent->bref.type was
4506 * HAMMER2_BREF_TYPE_INODE.
4507 */
4508 key = *keyp;
4509 locount = 0;
4510 hicount = 0;
4511
4512 /*
4513 * Calculate the range of keys in the array being careful to skip
4514 * slots which are overridden with a deletion.
4515 */
4516 key_beg = 0;
4517 key_end = HAMMER2_KEY_MAX;
4518 hammer2_spin_ex(&parent->core.spin);
4519
4520 for (;;) {
4521 if (--maxloops == 0) {
4522 panic("indkey_freemap shit %p %p:%d\n",
4523 parent, base, count);
4524 }
4525 chain = hammer2_combined_find(parent, base, count,
4526 &key_next,
4527 key_beg, key_end,
4528 &bref);
4529
4530 /*
4531 * Exhausted search
4532 */
4533 if (bref == NULL)
4534 break;
4535
4536 /*
4537 * Deleted object
4538 */
4539 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) {
4540 if (key_next == 0 || key_next > key_end)
4541 break;
4542 key_beg = key_next;
4543 continue;
4544 }
4545
4546 /*
4547 * Use the full live (not deleted) element for the scan
4548 * iteration. HAMMER2 does not allow partial replacements.
4549 *
4550 * XXX should be built into hammer2_combined_find().
4551 */
4552 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
4553
4554 /*
4555 * Expand our calculated key range (key, keybits) to fit
4556 * the scanned key. nkeybits represents the full range
4557 * that we will later cut in half (two halves @ nkeybits - 1).
4558 */
4559 nkeybits = keybits;
4560 if (nkeybits < bref->keybits) {
4561 if (bref->keybits > 64) {
4562 kprintf("bad bref chain %p bref %p\n",
4563 chain, bref);
4564 Debugger("fubar");
4565 }
4566 nkeybits = bref->keybits;
4567 }
4568 while (nkeybits < 64 &&
4569 rounddown2(key ^ bref->key, (hammer2_key_t)1 << nkeybits) != 0) {
4570 ++nkeybits;
4571 }
4572
4573 /*
4574 * If the new key range is larger we have to determine
4575 * which side of the new key range the existing keys fall
4576 * under by checking the high bit, then collapsing the
4577 * locount into the hicount or vise-versa.
4578 */
4579 if (keybits != nkeybits) {
4580 if (((hammer2_key_t)1 << (nkeybits - 1)) & key) {
4581 hicount += locount;
4582 locount = 0;
4583 } else {
4584 locount += hicount;
4585 hicount = 0;
4586 }
4587 keybits = nkeybits;
4588 }
4589
4590 /*
4591 * The newly scanned key will be in the lower half or the
4592 * upper half of the (new) key range.
4593 */
4594 if (((hammer2_key_t)1 << (nkeybits - 1)) & bref->key)
4595 ++hicount;
4596 else
4597 ++locount;
4598
4599 if (key_next == 0)
4600 break;
4601 key_beg = key_next;
4602 }
4603 hammer2_spin_unex(&parent->core.spin);
4604 bref = NULL; /* now invalid (safety) */
4605
4606 /*
4607 * Adjust keybits to represent half of the full range calculated
4608 * above (radix 63 max) for our new indirect block.
4609 */
4610 --keybits;
4611
4612 /*
4613 * Expand keybits to hold at least ncount elements. ncount will be
4614 * a power of 2. This is to try to completely fill leaf nodes (at
4615 * least for keys which are not hashes).
4616 *
4617 * We aren't counting 'in' or 'out', we are counting 'high side'
4618 * and 'low side' based on the bit at (1LL << keybits). We want
4619 * everything to be inside in these cases so shift it all to
4620 * the low or high side depending on the new high bit.
4621 */
4622 while (((hammer2_key_t)1 << keybits) < ncount) {
4623 ++keybits;
4624 if (key & ((hammer2_key_t)1 << keybits)) {
4625 hicount += locount;
4626 locount = 0;
4627 } else {
4628 locount += hicount;
4629 hicount = 0;
4630 }
4631 }
4632
4633 if (hicount > locount)
4634 key |= (hammer2_key_t)1 << keybits;
4635 else
4636 key &= ~(hammer2_key_t)1 << keybits;
4637
4638 *keyp = key;
4639
4640 return (keybits);
4641 }
4642
4643 #else
4644
4645 /*
4646 * Directory indirect blocks.
4647 *
4648 * Covers both the inode index (directory of inodes), and directory contents
4649 * (filenames hardlinked to inodes).
4650 *
4651 * Because directory keys are hashed we generally try to cut the space in
4652 * half. We accomodate the inode index (which tends to have linearly
4653 * increasing inode numbers) by ensuring that the keyspace is at least large
4654 * enough to fill up the indirect block being created.
4655 */
4656 static int
hammer2_chain_indkey_dir(hammer2_chain_t * parent,hammer2_key_t * keyp,int keybits,hammer2_blockref_t * base,int count,int ncount)4657 hammer2_chain_indkey_dir(hammer2_chain_t *parent, hammer2_key_t *keyp,
4658 int keybits, hammer2_blockref_t *base, int count,
4659 int ncount)
4660 {
4661 hammer2_blockref_t *bref;
4662 hammer2_chain_t *chain;
4663 hammer2_key_t key_beg;
4664 hammer2_key_t key_end;
4665 hammer2_key_t key_next;
4666 hammer2_key_t key;
4667 int nkeybits;
4668 int locount;
4669 int hicount;
4670 int maxloops = 300000;
4671
4672 /*
4673 * Shortcut if the parent is the inode. In this situation the
4674 * parent has 4+1 directory entries and we are creating an indirect
4675 * block capable of holding many more.
4676 */
4677 if (parent->bref.type == HAMMER2_BREF_TYPE_INODE) {
4678 return 63;
4679 }
4680
4681 key = *keyp;
4682 locount = 0;
4683 hicount = 0;
4684
4685 /*
4686 * Calculate the range of keys in the array being careful to skip
4687 * slots which are overridden with a deletion.
4688 */
4689 key_beg = 0;
4690 key_end = HAMMER2_KEY_MAX;
4691 hammer2_spin_ex(&parent->core.spin);
4692
4693 for (;;) {
4694 if (--maxloops == 0) {
4695 panic("indkey_freemap shit %p %p:%d\n",
4696 parent, base, count);
4697 }
4698 chain = hammer2_combined_find(parent, base, count,
4699 &key_next,
4700 key_beg, key_end,
4701 &bref);
4702
4703 /*
4704 * Exhausted search
4705 */
4706 if (bref == NULL)
4707 break;
4708
4709 /*
4710 * Deleted object
4711 */
4712 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) {
4713 if (key_next == 0 || key_next > key_end)
4714 break;
4715 key_beg = key_next;
4716 continue;
4717 }
4718
4719 /*
4720 * Use the full live (not deleted) element for the scan
4721 * iteration. HAMMER2 does not allow partial replacements.
4722 *
4723 * XXX should be built into hammer2_combined_find().
4724 */
4725 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);
4726
4727 /*
4728 * Expand our calculated key range (key, keybits) to fit
4729 * the scanned key. nkeybits represents the full range
4730 * that we will later cut in half (two halves @ nkeybits - 1).
4731 */
4732 nkeybits = keybits;
4733 if (nkeybits < bref->keybits) {
4734 if (bref->keybits > 64) {
4735 kprintf("bad bref chain %p bref %p\n",
4736 chain, bref);
4737 Debugger("fubar");
4738 }
4739 nkeybits = bref->keybits;
4740 }
4741 while (nkeybits < 64 &&
4742 (~(((hammer2_key_t)1 << nkeybits) - 1) &
4743 (key ^ bref->key)) != 0) {
4744 ++nkeybits;
4745 }
4746
4747 /*
4748 * If the new key range is larger we have to determine
4749 * which side of the new key range the existing keys fall
4750 * under by checking the high bit, then collapsing the
4751 * locount into the hicount or vise-versa.
4752 */
4753 if (keybits != nkeybits) {
4754 if (((hammer2_key_t)1 << (nkeybits - 1)) & key) {
4755 hicount += locount;
4756 locount = 0;
4757 } else {
4758 locount += hicount;
4759 hicount = 0;
4760 }
4761 keybits = nkeybits;
4762 }
4763
4764 /*
4765 * The newly scanned key will be in the lower half or the
4766 * upper half of the (new) key range.
4767 */
4768 if (((hammer2_key_t)1 << (nkeybits - 1)) & bref->key)
4769 ++hicount;
4770 else
4771 ++locount;
4772
4773 if (key_next == 0)
4774 break;
4775 key_beg = key_next;
4776 }
4777 hammer2_spin_unex(&parent->core.spin);
4778 bref = NULL; /* now invalid (safety) */
4779
4780 /*
4781 * Adjust keybits to represent half of the full range calculated
4782 * above (radix 63 max) for our new indirect block.
4783 */
4784 --keybits;
4785
4786 /*
4787 * Expand keybits to hold at least ncount elements. ncount will be
4788 * a power of 2. This is to try to completely fill leaf nodes (at
4789 * least for keys which are not hashes).
4790 *
4791 * We aren't counting 'in' or 'out', we are counting 'high side'
4792 * and 'low side' based on the bit at (1LL << keybits). We want
4793 * everything to be inside in these cases so shift it all to
4794 * the low or high side depending on the new high bit.
4795 */
4796 while (((hammer2_key_t)1 << keybits) < ncount) {
4797 ++keybits;
4798 if (key & ((hammer2_key_t)1 << keybits)) {
4799 hicount += locount;
4800 locount = 0;
4801 } else {
4802 locount += hicount;
4803 hicount = 0;
4804 }
4805 }
4806
4807 if (hicount > locount)
4808 key |= (hammer2_key_t)1 << keybits;
4809 else
4810 key &= ~(hammer2_key_t)1 << keybits;
4811
4812 *keyp = key;
4813
4814 return (keybits);
4815 }
4816
4817 #endif
4818
4819 /*
4820 * Sets CHAIN_DELETED and remove the chain's blockref from the parent if
4821 * it exists.
4822 *
4823 * Both parent and chain must be locked exclusively.
4824 *
4825 * This function will modify the parent if the blockref requires removal
4826 * from the parent's block table.
4827 *
4828 * This function is NOT recursive. Any entity already pushed into the
4829 * chain (such as an inode) may still need visibility into its contents,
4830 * as well as the ability to read and modify the contents. For example,
4831 * for an unlinked file which is still open.
4832 *
4833 * Also note that the flusher is responsible for cleaning up empty
4834 * indirect blocks.
4835 */
4836 int
hammer2_chain_delete(hammer2_chain_t * parent,hammer2_chain_t * chain,hammer2_tid_t mtid,int flags)4837 hammer2_chain_delete(hammer2_chain_t *parent, hammer2_chain_t *chain,
4838 hammer2_tid_t mtid, int flags)
4839 {
4840 int error = 0;
4841
4842 KKASSERT(hammer2_mtx_owned(&chain->lock));
4843
4844 /*
4845 * Nothing to do if already marked.
4846 *
4847 * We need the spinlock on the core whos RBTREE contains chain
4848 * to protect against races.
4849 */
4850 if ((chain->flags & HAMMER2_CHAIN_DELETED) == 0) {
4851 KKASSERT((chain->flags & HAMMER2_CHAIN_DELETED) == 0 &&
4852 chain->parent == parent);
4853 error = _hammer2_chain_delete_helper(parent, chain,
4854 mtid, flags, NULL);
4855 }
4856
4857 /*
4858 * Permanent deletions mark the chain as destroyed.
4859 *
4860 * NOTE: We do not setflush the chain unless the deletion is
4861 * permanent, since the deletion of a chain does not actually
4862 * require it to be flushed.
4863 */
4864 if (error == 0) {
4865 if (flags & HAMMER2_DELETE_PERMANENT) {
4866 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DESTROY);
4867 hammer2_chain_setflush(chain);
4868 }
4869 }
4870
4871 return error;
4872 }
4873
4874 static int
hammer2_chain_delete_obref(hammer2_chain_t * parent,hammer2_chain_t * chain,hammer2_tid_t mtid,int flags,hammer2_blockref_t * obref)4875 hammer2_chain_delete_obref(hammer2_chain_t *parent, hammer2_chain_t *chain,
4876 hammer2_tid_t mtid, int flags,
4877 hammer2_blockref_t *obref)
4878 {
4879 int error = 0;
4880
4881 KKASSERT(hammer2_mtx_owned(&chain->lock));
4882
4883 /*
4884 * Nothing to do if already marked.
4885 *
4886 * We need the spinlock on the core whos RBTREE contains chain
4887 * to protect against races.
4888 */
4889 obref->type = HAMMER2_BREF_TYPE_EMPTY;
4890 if ((chain->flags & HAMMER2_CHAIN_DELETED) == 0) {
4891 KKASSERT((chain->flags & HAMMER2_CHAIN_DELETED) == 0 &&
4892 chain->parent == parent);
4893 error = _hammer2_chain_delete_helper(parent, chain,
4894 mtid, flags, obref);
4895 }
4896
4897 /*
4898 * Permanent deletions mark the chain as destroyed.
4899 *
4900 * NOTE: We do not setflush the chain unless the deletion is
4901 * permanent, since the deletion of a chain does not actually
4902 * require it to be flushed.
4903 */
4904 if (error == 0) {
4905 if (flags & HAMMER2_DELETE_PERMANENT) {
4906 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DESTROY);
4907 hammer2_chain_setflush(chain);
4908 }
4909 }
4910
4911 return error;
4912 }
4913
4914 /*
4915 * Returns the index of the nearest element in the blockref array >= elm.
4916 * Returns (count) if no element could be found.
4917 *
4918 * Sets *key_nextp to the next key for loop purposes but does not modify
4919 * it if the next key would be higher than the current value of *key_nextp.
4920 * Note that *key_nexp can overflow to 0, which should be tested by the
4921 * caller.
4922 *
4923 * WARNING! Must be called with parent's spinlock held. Spinlock remains
4924 * held through the operation.
4925 */
4926 static int
hammer2_base_find(hammer2_chain_t * parent,hammer2_blockref_t * base,int count,hammer2_key_t * key_nextp,hammer2_key_t key_beg,hammer2_key_t key_end)4927 hammer2_base_find(hammer2_chain_t *parent,
4928 hammer2_blockref_t *base, int count,
4929 hammer2_key_t *key_nextp,
4930 hammer2_key_t key_beg, hammer2_key_t key_end)
4931 {
4932 hammer2_blockref_t *scan;
4933 hammer2_key_t scan_end;
4934 int i;
4935 int limit;
4936
4937 /*
4938 * Require the live chain's already have their core's counted
4939 * so we can optimize operations.
4940 */
4941 KKASSERT(parent->flags & HAMMER2_CHAIN_COUNTEDBREFS);
4942
4943 /*
4944 * Degenerate case
4945 */
4946 if (count == 0 || base == NULL)
4947 return(count);
4948
4949 /*
4950 * Sequential optimization using parent->cache_index. This is
4951 * the most likely scenario.
4952 *
4953 * We can avoid trailing empty entries on live chains, otherwise
4954 * we might have to check the whole block array.
4955 */
4956 i = parent->cache_index; /* SMP RACE OK */
4957 cpu_ccfence();
4958 limit = parent->core.live_zero;
4959 if (i >= limit)
4960 i = limit - 1;
4961 if (i < 0)
4962 i = 0;
4963 KKASSERT(i < count);
4964
4965 /*
4966 * Search backwards
4967 */
4968 scan = &base[i];
4969 while (i > 0 && (scan->type == HAMMER2_BREF_TYPE_EMPTY ||
4970 scan->key > key_beg)) {
4971 --scan;
4972 --i;
4973 }
4974 parent->cache_index = i;
4975
4976 /*
4977 * Search forwards, stop when we find a scan element which
4978 * encloses the key or until we know that there are no further
4979 * elements.
4980 */
4981 while (i < count) {
4982 if (scan->type != HAMMER2_BREF_TYPE_EMPTY) {
4983 scan_end = scan->key +
4984 ((hammer2_key_t)1 << scan->keybits) - 1;
4985 if (scan->key > key_beg || scan_end >= key_beg)
4986 break;
4987 }
4988 if (i >= limit)
4989 return (count);
4990 ++scan;
4991 ++i;
4992 }
4993 if (i != count) {
4994 parent->cache_index = i;
4995 if (i >= limit) {
4996 i = count;
4997 } else {
4998 scan_end = scan->key +
4999 ((hammer2_key_t)1 << scan->keybits);
5000 if (scan_end && (*key_nextp > scan_end ||
5001 *key_nextp == 0)) {
5002 *key_nextp = scan_end;
5003 }
5004 }
5005 }
5006 return (i);
5007 }
5008
5009 /*
5010 * Do a combined search and return the next match either from the blockref
5011 * array or from the in-memory chain. Sets *brefp to the returned bref in
5012 * both cases, or sets it to NULL if the search exhausted. Only returns
5013 * a non-NULL chain if the search matched from the in-memory chain.
5014 *
5015 * When no in-memory chain has been found and a non-NULL bref is returned
5016 * in *brefp.
5017 *
5018 *
5019 * The returned chain is not locked or referenced. Use the returned bref
5020 * to determine if the search exhausted or not. Iterate if the base find
5021 * is chosen but matches a deleted chain.
5022 *
5023 * WARNING! Must be called with parent's spinlock held. Spinlock remains
5024 * held through the operation.
5025 */
5026 static hammer2_chain_t *
hammer2_combined_find(hammer2_chain_t * parent,hammer2_blockref_t * base,int count,hammer2_key_t * key_nextp,hammer2_key_t key_beg,hammer2_key_t key_end,hammer2_blockref_t ** brefp)5027 hammer2_combined_find(hammer2_chain_t *parent,
5028 hammer2_blockref_t *base, int count,
5029 hammer2_key_t *key_nextp,
5030 hammer2_key_t key_beg, hammer2_key_t key_end,
5031 hammer2_blockref_t **brefp)
5032 {
5033 hammer2_blockref_t *bref;
5034 hammer2_chain_t *chain;
5035 int i;
5036
5037 /*
5038 * Lookup in block array and in rbtree.
5039 */
5040 *key_nextp = key_end + 1;
5041 i = hammer2_base_find(parent, base, count, key_nextp,
5042 key_beg, key_end);
5043 chain = hammer2_chain_find(parent, key_nextp, key_beg, key_end);
5044
5045 /*
5046 * Neither matched
5047 */
5048 if (i == count && chain == NULL) {
5049 *brefp = NULL;
5050 return(NULL);
5051 }
5052
5053 /*
5054 * Only chain matched.
5055 */
5056 if (i == count) {
5057 bref = &chain->bref;
5058 goto found;
5059 }
5060
5061 /*
5062 * Only blockref matched.
5063 */
5064 if (chain == NULL) {
5065 bref = &base[i];
5066 goto found;
5067 }
5068
5069 /*
5070 * Both in-memory and blockref matched, select the nearer element.
5071 *
5072 * If both are flush with the left-hand side or both are the
5073 * same distance away, select the chain. In this situation the
5074 * chain must have been loaded from the matching blockmap.
5075 */
5076 if ((chain->bref.key <= key_beg && base[i].key <= key_beg) ||
5077 chain->bref.key == base[i].key) {
5078 KKASSERT(chain->bref.key == base[i].key);
5079 bref = &chain->bref;
5080 goto found;
5081 }
5082
5083 /*
5084 * Select the nearer key
5085 */
5086 if (chain->bref.key < base[i].key) {
5087 bref = &chain->bref;
5088 } else {
5089 bref = &base[i];
5090 chain = NULL;
5091 }
5092
5093 /*
5094 * If the bref is out of bounds we've exhausted our search.
5095 */
5096 found:
5097 if (bref->key > key_end) {
5098 *brefp = NULL;
5099 chain = NULL;
5100 } else {
5101 *brefp = bref;
5102 }
5103 return(chain);
5104 }
5105
5106 /*
5107 * Locate the specified block array element and delete it. The element
5108 * must exist.
5109 *
5110 * The spin lock on the related chain must be held.
5111 *
5112 * NOTE: live_count was adjusted when the chain was deleted, so it does not
5113 * need to be adjusted when we commit the media change.
5114 */
5115 void
hammer2_base_delete(hammer2_chain_t * parent,hammer2_blockref_t * base,int count,hammer2_chain_t * chain,hammer2_blockref_t * obref)5116 hammer2_base_delete(hammer2_chain_t *parent,
5117 hammer2_blockref_t *base, int count,
5118 hammer2_chain_t *chain,
5119 hammer2_blockref_t *obref)
5120 {
5121 hammer2_blockref_t *elm = &chain->bref;
5122 hammer2_blockref_t *scan;
5123 hammer2_key_t key_next;
5124 int i;
5125
5126 /*
5127 * Delete element. Expect the element to exist.
5128 *
5129 * XXX see caller, flush code not yet sophisticated enough to prevent
5130 * re-flushed in some cases.
5131 */
5132 key_next = 0; /* max range */
5133 i = hammer2_base_find(parent, base, count, &key_next,
5134 elm->key, elm->key);
5135 scan = &base[i];
5136 if (i == count || scan->type == HAMMER2_BREF_TYPE_EMPTY ||
5137 scan->key != elm->key ||
5138 ((chain->flags & HAMMER2_CHAIN_BLKMAPUPD) == 0 &&
5139 scan->keybits != elm->keybits)) {
5140 hammer2_spin_unex(&parent->core.spin);
5141 panic("delete base %p element not found at %d/%d elm %p\n",
5142 base, i, count, elm);
5143 return;
5144 }
5145
5146 /*
5147 * Update stats and zero the entry.
5148 *
5149 * NOTE: Handle radix == 0 (0 bytes) case.
5150 */
5151 if ((int)(scan->data_off & HAMMER2_OFF_MASK_RADIX)) {
5152 parent->bref.embed.stats.data_count -= (hammer2_off_t)1 <<
5153 (int)(scan->data_off & HAMMER2_OFF_MASK_RADIX);
5154 }
5155 switch(scan->type) {
5156 case HAMMER2_BREF_TYPE_INODE:
5157 --parent->bref.embed.stats.inode_count;
5158 /* fall through */
5159 case HAMMER2_BREF_TYPE_DATA:
5160 if (parent->bref.leaf_count == HAMMER2_BLOCKREF_LEAF_MAX) {
5161 atomic_set_int(&chain->flags,
5162 HAMMER2_CHAIN_HINT_LEAF_COUNT);
5163 } else {
5164 if (parent->bref.leaf_count)
5165 --parent->bref.leaf_count;
5166 }
5167 /* fall through */
5168 case HAMMER2_BREF_TYPE_INDIRECT:
5169 if (scan->type != HAMMER2_BREF_TYPE_DATA) {
5170 parent->bref.embed.stats.data_count -=
5171 scan->embed.stats.data_count;
5172 parent->bref.embed.stats.inode_count -=
5173 scan->embed.stats.inode_count;
5174 }
5175 if (scan->type == HAMMER2_BREF_TYPE_INODE)
5176 break;
5177 if (parent->bref.leaf_count == HAMMER2_BLOCKREF_LEAF_MAX) {
5178 atomic_set_int(&chain->flags,
5179 HAMMER2_CHAIN_HINT_LEAF_COUNT);
5180 } else {
5181 if (parent->bref.leaf_count <= scan->leaf_count)
5182 parent->bref.leaf_count = 0;
5183 else
5184 parent->bref.leaf_count -= scan->leaf_count;
5185 }
5186 break;
5187 case HAMMER2_BREF_TYPE_DIRENT:
5188 if (parent->bref.leaf_count == HAMMER2_BLOCKREF_LEAF_MAX) {
5189 atomic_set_int(&chain->flags,
5190 HAMMER2_CHAIN_HINT_LEAF_COUNT);
5191 } else {
5192 if (parent->bref.leaf_count)
5193 --parent->bref.leaf_count;
5194 }
5195 default:
5196 break;
5197 }
5198
5199 if (obref)
5200 *obref = *scan;
5201 bzero(scan, sizeof(*scan));
5202
5203 /*
5204 * We can only optimize parent->core.live_zero for live chains.
5205 */
5206 if (parent->core.live_zero == i + 1) {
5207 while (--i >= 0 && base[i].type == HAMMER2_BREF_TYPE_EMPTY)
5208 ;
5209 parent->core.live_zero = i + 1;
5210 }
5211
5212 /*
5213 * Clear appropriate blockmap flags in chain.
5214 */
5215 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_BLKMAPPED |
5216 HAMMER2_CHAIN_BLKMAPUPD);
5217 }
5218
5219 /*
5220 * Insert the specified element. The block array must not already have the
5221 * element and must have space available for the insertion.
5222 *
5223 * The spin lock on the related chain must be held.
5224 *
5225 * NOTE: live_count was adjusted when the chain was deleted, so it does not
5226 * need to be adjusted when we commit the media change.
5227 */
5228 void
hammer2_base_insert(hammer2_chain_t * parent,hammer2_blockref_t * base,int count,hammer2_chain_t * chain,hammer2_blockref_t * elm)5229 hammer2_base_insert(hammer2_chain_t *parent,
5230 hammer2_blockref_t *base, int count,
5231 hammer2_chain_t *chain, hammer2_blockref_t *elm)
5232 {
5233 hammer2_key_t key_next;
5234 hammer2_key_t xkey;
5235 int i;
5236 int j;
5237 int k;
5238 int l;
5239 int u = 1;
5240
5241 /*
5242 * Insert new element. Expect the element to not already exist
5243 * unless we are replacing it.
5244 *
5245 * XXX see caller, flush code not yet sophisticated enough to prevent
5246 * re-flushed in some cases.
5247 */
5248 key_next = 0; /* max range */
5249 i = hammer2_base_find(parent, base, count, &key_next,
5250 elm->key, elm->key);
5251
5252 /*
5253 * Shortcut fill optimization, typical ordered insertion(s) may not
5254 * require a search.
5255 */
5256 KKASSERT(i >= 0 && i <= count);
5257
5258 /*
5259 * Set appropriate blockmap flags in chain (if not NULL)
5260 */
5261 if (chain)
5262 atomic_set_int(&chain->flags, HAMMER2_CHAIN_BLKMAPPED);
5263
5264 /*
5265 * Update stats and zero the entry
5266 */
5267 if ((int)(elm->data_off & HAMMER2_OFF_MASK_RADIX)) {
5268 parent->bref.embed.stats.data_count += (hammer2_off_t)1 <<
5269 (int)(elm->data_off & HAMMER2_OFF_MASK_RADIX);
5270 }
5271 switch(elm->type) {
5272 case HAMMER2_BREF_TYPE_INODE:
5273 ++parent->bref.embed.stats.inode_count;
5274 /* fall through */
5275 case HAMMER2_BREF_TYPE_DATA:
5276 if (parent->bref.leaf_count != HAMMER2_BLOCKREF_LEAF_MAX)
5277 ++parent->bref.leaf_count;
5278 /* fall through */
5279 case HAMMER2_BREF_TYPE_INDIRECT:
5280 if (elm->type != HAMMER2_BREF_TYPE_DATA) {
5281 parent->bref.embed.stats.data_count +=
5282 elm->embed.stats.data_count;
5283 parent->bref.embed.stats.inode_count +=
5284 elm->embed.stats.inode_count;
5285 }
5286 if (elm->type == HAMMER2_BREF_TYPE_INODE)
5287 break;
5288 if (parent->bref.leaf_count + elm->leaf_count <
5289 HAMMER2_BLOCKREF_LEAF_MAX) {
5290 parent->bref.leaf_count += elm->leaf_count;
5291 } else {
5292 parent->bref.leaf_count = HAMMER2_BLOCKREF_LEAF_MAX;
5293 }
5294 break;
5295 case HAMMER2_BREF_TYPE_DIRENT:
5296 if (parent->bref.leaf_count != HAMMER2_BLOCKREF_LEAF_MAX)
5297 ++parent->bref.leaf_count;
5298 break;
5299 default:
5300 break;
5301 }
5302
5303
5304 /*
5305 * We can only optimize parent->core.live_zero for live chains.
5306 */
5307 if (i == count && parent->core.live_zero < count) {
5308 i = parent->core.live_zero++;
5309 base[i] = *elm;
5310 return;
5311 }
5312
5313 xkey = elm->key + ((hammer2_key_t)1 << elm->keybits) - 1;
5314 if (i != count && (base[i].key < elm->key || xkey >= base[i].key)) {
5315 hammer2_spin_unex(&parent->core.spin);
5316 panic("insert base %p overlapping elements at %d elm %p\n",
5317 base, i, elm);
5318 }
5319
5320 /*
5321 * Try to find an empty slot before or after.
5322 */
5323 j = i;
5324 k = i;
5325 while (j > 0 || k < count) {
5326 --j;
5327 if (j >= 0 && base[j].type == HAMMER2_BREF_TYPE_EMPTY) {
5328 if (j == i - 1) {
5329 base[j] = *elm;
5330 } else {
5331 bcopy(&base[j+1], &base[j],
5332 (i - j - 1) * sizeof(*base));
5333 base[i - 1] = *elm;
5334 }
5335 goto validate;
5336 }
5337 ++k;
5338 if (k < count && base[k].type == HAMMER2_BREF_TYPE_EMPTY) {
5339 bcopy(&base[i], &base[i+1],
5340 (k - i) * sizeof(hammer2_blockref_t));
5341 base[i] = *elm;
5342
5343 /*
5344 * We can only update parent->core.live_zero for live
5345 * chains.
5346 */
5347 if (parent->core.live_zero <= k)
5348 parent->core.live_zero = k + 1;
5349 u = 2;
5350 goto validate;
5351 }
5352 }
5353 panic("hammer2_base_insert: no room!");
5354
5355 /*
5356 * Debugging
5357 */
5358 validate:
5359 key_next = 0;
5360 for (l = 0; l < count; ++l) {
5361 if (base[l].type != HAMMER2_BREF_TYPE_EMPTY) {
5362 key_next = base[l].key +
5363 ((hammer2_key_t)1 << base[l].keybits) - 1;
5364 break;
5365 }
5366 }
5367 while (++l < count) {
5368 if (base[l].type != HAMMER2_BREF_TYPE_EMPTY) {
5369 if (base[l].key <= key_next)
5370 panic("base_insert %d %d,%d,%d fail %p:%d", u, i, j, k, base, l);
5371 key_next = base[l].key +
5372 ((hammer2_key_t)1 << base[l].keybits) - 1;
5373
5374 }
5375 }
5376
5377 }
5378
5379 #if 0
5380
5381 /*
5382 * Sort the blockref array for the chain. Used by the flush code to
5383 * sort the blockref[] array.
5384 *
5385 * The chain must be exclusively locked AND spin-locked.
5386 */
5387 typedef hammer2_blockref_t *hammer2_blockref_p;
5388
5389 static
5390 int
5391 hammer2_base_sort_callback(const void *v1, const void *v2)
5392 {
5393 hammer2_blockref_p bref1 = *(const hammer2_blockref_p *)v1;
5394 hammer2_blockref_p bref2 = *(const hammer2_blockref_p *)v2;
5395
5396 /*
5397 * Make sure empty elements are placed at the end of the array
5398 */
5399 if (bref1->type == HAMMER2_BREF_TYPE_EMPTY) {
5400 if (bref2->type == HAMMER2_BREF_TYPE_EMPTY)
5401 return(0);
5402 return(1);
5403 } else if (bref2->type == HAMMER2_BREF_TYPE_EMPTY) {
5404 return(-1);
5405 }
5406
5407 /*
5408 * Sort by key
5409 */
5410 if (bref1->key < bref2->key)
5411 return(-1);
5412 if (bref1->key > bref2->key)
5413 return(1);
5414 return(0);
5415 }
5416
5417 void
5418 hammer2_base_sort(hammer2_chain_t *chain)
5419 {
5420 hammer2_blockref_t *base;
5421 int count;
5422
5423 switch(chain->bref.type) {
5424 case HAMMER2_BREF_TYPE_INODE:
5425 /*
5426 * Special shortcut for embedded data returns the inode
5427 * itself. Callers must detect this condition and access
5428 * the embedded data (the strategy code does this for us).
5429 *
5430 * This is only applicable to regular files and softlinks.
5431 */
5432 if (chain->data->ipdata.meta.op_flags &
5433 HAMMER2_OPFLAG_DIRECTDATA) {
5434 return;
5435 }
5436 base = &chain->data->ipdata.u.blockset.blockref[0];
5437 count = HAMMER2_SET_COUNT;
5438 break;
5439 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
5440 case HAMMER2_BREF_TYPE_INDIRECT:
5441 /*
5442 * Optimize indirect blocks in the INITIAL state to avoid
5443 * I/O.
5444 */
5445 KKASSERT((chain->flags & HAMMER2_CHAIN_INITIAL) == 0);
5446 base = &chain->data->npdata[0];
5447 count = chain->bytes / sizeof(hammer2_blockref_t);
5448 break;
5449 case HAMMER2_BREF_TYPE_VOLUME:
5450 base = &chain->data->voldata.sroot_blockset.blockref[0];
5451 count = HAMMER2_SET_COUNT;
5452 break;
5453 case HAMMER2_BREF_TYPE_FREEMAP:
5454 base = &chain->data->blkset.blockref[0];
5455 count = HAMMER2_SET_COUNT;
5456 break;
5457 default:
5458 panic("hammer2_base_sort: unrecognized "
5459 "blockref(A) type: %d",
5460 chain->bref.type);
5461 base = NULL; /* safety */
5462 count = 0; /* safety */
5463 break;
5464 }
5465 kqsort(base, count, sizeof(*base), hammer2_base_sort_callback);
5466 }
5467
5468 #endif
5469
5470 /*
5471 * Set the check data for a chain. This can be a heavy-weight operation
5472 * and typically only runs on-flush. For file data check data is calculated
5473 * when the logical buffers are flushed.
5474 */
5475 void
hammer2_chain_setcheck(hammer2_chain_t * chain,void * bdata)5476 hammer2_chain_setcheck(hammer2_chain_t *chain, void *bdata)
5477 {
5478 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_NOTTESTED);
5479
5480 switch(HAMMER2_DEC_CHECK(chain->bref.methods)) {
5481 case HAMMER2_CHECK_NONE:
5482 break;
5483 case HAMMER2_CHECK_DISABLED:
5484 break;
5485 case HAMMER2_CHECK_ISCSI32:
5486 chain->bref.check.iscsi32.value =
5487 hammer2_icrc32(bdata, chain->bytes);
5488 break;
5489 case HAMMER2_CHECK_XXHASH64:
5490 chain->bref.check.xxhash64.value =
5491 XXH64(bdata, chain->bytes, XXH_HAMMER2_SEED);
5492 break;
5493 case HAMMER2_CHECK_SHA192:
5494 assert(0); /* XXX unsupported */
5495 /*
5496 {
5497 SHA256_CTX hash_ctx;
5498 union {
5499 uint8_t digest[SHA256_DIGEST_LENGTH];
5500 uint64_t digest64[SHA256_DIGEST_LENGTH/8];
5501 } u;
5502
5503 SHA256_Init(&hash_ctx);
5504 SHA256_Update(&hash_ctx, bdata, chain->bytes);
5505 SHA256_Final(u.digest, &hash_ctx);
5506 u.digest64[2] ^= u.digest64[3];
5507 bcopy(u.digest,
5508 chain->bref.check.sha192.data,
5509 sizeof(chain->bref.check.sha192.data));
5510 }
5511 */
5512 break;
5513 case HAMMER2_CHECK_FREEMAP:
5514 chain->bref.check.freemap.icrc32 =
5515 hammer2_icrc32(bdata, chain->bytes);
5516 break;
5517 default:
5518 kprintf("hammer2_chain_setcheck: unknown check type %02x\n",
5519 chain->bref.methods);
5520 break;
5521 }
5522 }
5523
5524 /*
5525 * Characterize a failed check code and try to trace back to the inode.
5526 */
5527 static void
hammer2_characterize_failed_chain(hammer2_chain_t * chain,uint64_t check,int bits)5528 hammer2_characterize_failed_chain(hammer2_chain_t *chain, uint64_t check,
5529 int bits)
5530 {
5531 hammer2_chain_t *lchain;
5532 hammer2_chain_t *ochain;
5533 int did;
5534
5535 did = krateprintf(&krate_h2chk,
5536 "chain %016jx.%02x (%s) meth=%02x CHECK FAIL "
5537 "(flags=%08x, bref/data ",
5538 chain->bref.data_off,
5539 chain->bref.type,
5540 hammer2_bref_type_str(chain->bref.type),
5541 chain->bref.methods,
5542 chain->flags);
5543 if (did == 0)
5544 return;
5545
5546 if (bits == 32) {
5547 kprintf("%08x/%08x)\n",
5548 chain->bref.check.iscsi32.value,
5549 (uint32_t)check);
5550 } else {
5551 kprintf("%016jx/%016jx)\n",
5552 chain->bref.check.xxhash64.value,
5553 check);
5554 }
5555
5556 /*
5557 * Run up the chains to try to find the governing inode so we
5558 * can report it.
5559 *
5560 * XXX This error reporting is not really MPSAFE
5561 */
5562 ochain = chain;
5563 lchain = chain;
5564 while (chain && chain->bref.type != HAMMER2_BREF_TYPE_INODE) {
5565 lchain = chain;
5566 chain = chain->parent;
5567 }
5568
5569 if (chain && chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
5570 ((chain->bref.flags & HAMMER2_BREF_FLAG_PFSROOT) == 0 ||
5571 (lchain->bref.key & HAMMER2_DIRHASH_VISIBLE))) {
5572 kprintf(" Resides at/in inode %ld\n",
5573 (long)chain->bref.key);
5574 } else if (chain && chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
5575 kprintf(" Resides in inode index - CRITICAL!!!\n");
5576 } else {
5577 kprintf(" Resides in root index - CRITICAL!!!\n");
5578 }
5579 if (ochain->hmp) {
5580 const char *pfsname = "UNKNOWN";
5581 int i;
5582
5583 if (ochain->pmp) {
5584 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
5585 if (ochain->pmp->pfs_hmps[i] == ochain->hmp &&
5586 ochain->pmp->pfs_names[i]) {
5587 pfsname = ochain->pmp->pfs_names[i];
5588 break;
5589 }
5590 }
5591 }
5592 kprintf(" In pfs %s on device %s\n",
5593 pfsname, ochain->hmp->devrepname);
5594 }
5595 }
5596
5597 /*
5598 * Returns non-zero on success, 0 on failure.
5599 */
5600 int
hammer2_chain_testcheck(hammer2_chain_t * chain,void * bdata)5601 hammer2_chain_testcheck(hammer2_chain_t *chain, void *bdata)
5602 {
5603 uint32_t check32;
5604 uint64_t check64;
5605 int r;
5606
5607 if (chain->flags & HAMMER2_CHAIN_NOTTESTED)
5608 return 1;
5609
5610 switch(HAMMER2_DEC_CHECK(chain->bref.methods)) {
5611 case HAMMER2_CHECK_NONE:
5612 r = 1;
5613 break;
5614 case HAMMER2_CHECK_DISABLED:
5615 r = 1;
5616 break;
5617 case HAMMER2_CHECK_ISCSI32:
5618 check32 = hammer2_icrc32(bdata, chain->bytes);
5619 r = (chain->bref.check.iscsi32.value == check32);
5620 if (r == 0) {
5621 hammer2_characterize_failed_chain(chain, check32, 32);
5622 }
5623 hammer2_process_icrc32 += chain->bytes;
5624 break;
5625 case HAMMER2_CHECK_XXHASH64:
5626 check64 = XXH64(bdata, chain->bytes, XXH_HAMMER2_SEED);
5627 r = (chain->bref.check.xxhash64.value == check64);
5628 if (r == 0) {
5629 hammer2_characterize_failed_chain(chain, check64, 64);
5630 }
5631 hammer2_process_xxhash64 += chain->bytes;
5632 break;
5633 case HAMMER2_CHECK_SHA192:
5634 assert(0); /* XXX unsupported */
5635 /*
5636 {
5637 SHA256_CTX hash_ctx;
5638 union {
5639 uint8_t digest[SHA256_DIGEST_LENGTH];
5640 uint64_t digest64[SHA256_DIGEST_LENGTH/8];
5641 } u;
5642
5643 SHA256_Init(&hash_ctx);
5644 SHA256_Update(&hash_ctx, bdata, chain->bytes);
5645 SHA256_Final(u.digest, &hash_ctx);
5646 u.digest64[2] ^= u.digest64[3];
5647 if (bcmp(u.digest,
5648 chain->bref.check.sha192.data,
5649 sizeof(chain->bref.check.sha192.data)) == 0) {
5650 r = 1;
5651 } else {
5652 r = 0;
5653 krateprintf(&krate_h2chk,
5654 "chain %016jx.%02x meth=%02x "
5655 "CHECK FAIL\n",
5656 chain->bref.data_off,
5657 chain->bref.type,
5658 chain->bref.methods);
5659 }
5660 }
5661 */
5662 break;
5663 case HAMMER2_CHECK_FREEMAP:
5664 r = (chain->bref.check.freemap.icrc32 ==
5665 hammer2_icrc32(bdata, chain->bytes));
5666 if (r == 0) {
5667 int did;
5668
5669 did = krateprintf(&krate_h2chk,
5670 "chain %016jx.%02x meth=%02x "
5671 "CHECK FAIL\n",
5672 chain->bref.data_off,
5673 chain->bref.type,
5674 chain->bref.methods);
5675 if (did) {
5676 kprintf("freemap.icrc %08x icrc32 %08x (%d)\n",
5677 chain->bref.check.freemap.icrc32,
5678 hammer2_icrc32(bdata, chain->bytes),
5679 chain->bytes);
5680 if (chain->dio) {
5681 kprintf("dio %p buf %016jx,%ld "
5682 "bdata %p/%p\n",
5683 chain->dio,
5684 (intmax_t)chain->dio->bp->b_loffset,
5685 chain->dio->bp->b_bufsize,
5686 bdata,
5687 chain->dio->bp->b_data);
5688 }
5689 }
5690 }
5691 break;
5692 default:
5693 kprintf("hammer2_chain_testcheck: unknown check type %02x\n",
5694 chain->bref.methods);
5695 r = 1;
5696 break;
5697 }
5698 return r;
5699 }
5700
5701 /*
5702 * Acquire the chain and parent representing the specified inode for the
5703 * device at the specified cluster index.
5704 *
5705 * The flags passed in are LOOKUP flags, not RESOLVE flags.
5706 *
5707 * If we are unable to locate the inode, HAMMER2_ERROR_EIO or HAMMER2_ERROR_CHECK
5708 * is returned. In case of error, *chainp and/or *parentp may still be returned
5709 * non-NULL.
5710 *
5711 * The caller may pass-in a locked *parentp and/or *chainp, or neither.
5712 * They will be unlocked and released by this function. The *parentp and
5713 * *chainp representing the located inode are returned locked.
5714 *
5715 * The returned error includes any error on the returned chain in addition to
5716 * errors incurred while trying to lookup the inode. However, a chain->error
5717 * might not be recognized if HAMMER2_LOOKUP_NODATA is passed. This flag may
5718 * not be passed to this function.
5719 */
5720 int
hammer2_chain_inode_find(hammer2_pfs_t * pmp,hammer2_key_t inum,int clindex,int flags,hammer2_chain_t ** parentp,hammer2_chain_t ** chainp)5721 hammer2_chain_inode_find(hammer2_pfs_t *pmp, hammer2_key_t inum,
5722 int clindex, int flags,
5723 hammer2_chain_t **parentp, hammer2_chain_t **chainp)
5724 {
5725 hammer2_chain_t *parent;
5726 hammer2_chain_t *rchain;
5727 hammer2_key_t key_dummy;
5728 hammer2_inode_t *ip;
5729 int resolve_flags;
5730 int error;
5731
5732 KKASSERT((flags & HAMMER2_LOOKUP_NODATA) == 0);
5733
5734 resolve_flags = (flags & HAMMER2_LOOKUP_SHARED) ?
5735 HAMMER2_RESOLVE_SHARED : 0;
5736
5737 /*
5738 * Caller expects us to replace these.
5739 */
5740 if (*chainp) {
5741 hammer2_chain_unlock(*chainp);
5742 hammer2_chain_drop(*chainp);
5743 *chainp = NULL;
5744 }
5745 if (*parentp) {
5746 hammer2_chain_unlock(*parentp);
5747 hammer2_chain_drop(*parentp);
5748 *parentp = NULL;
5749 }
5750
5751 /*
5752 * Be very careful, this is a backend function and we CANNOT
5753 * lock any frontend inode structure we find. But we have to
5754 * look the inode up this way first in case it exists but is
5755 * detached from the radix tree.
5756 */
5757 ip = hammer2_inode_lookup(pmp, inum);
5758 if (ip) {
5759 *chainp = hammer2_inode_chain_and_parent(ip, clindex,
5760 parentp,
5761 resolve_flags);
5762 hammer2_inode_drop(ip);
5763 if (*chainp)
5764 return (*chainp)->error;
5765 hammer2_chain_unlock(*chainp);
5766 hammer2_chain_drop(*chainp);
5767 *chainp = NULL;
5768 if (*parentp) {
5769 hammer2_chain_unlock(*parentp);
5770 hammer2_chain_drop(*parentp);
5771 *parentp = NULL;
5772 }
5773 }
5774
5775 /*
5776 * Inodes hang off of the iroot (bit 63 is clear, differentiating
5777 * inodes from root directory entries in the key lookup).
5778 */
5779 parent = hammer2_inode_chain(pmp->iroot, clindex, resolve_flags);
5780 rchain = NULL;
5781 if (parent) {
5782 /*
5783 * NOTE: rchain can be returned as NULL even if error == 0
5784 * (i.e. not found)
5785 */
5786 rchain = hammer2_chain_lookup(&parent, &key_dummy,
5787 inum, inum,
5788 &error, flags);
5789 /*
5790 * Propagate a chain-specific error to caller.
5791 *
5792 * If the chain is not errored, we must still validate that the inode
5793 * number is correct, because all hell will break loose if it isn't
5794 * correct. It should always be correct so print to the console and
5795 * simulate a CHECK error if it is not.
5796 */
5797 if (error == 0 && rchain) {
5798 error = rchain->error;
5799 if (error == 0 && rchain->data) {
5800 if (inum != rchain->data->ipdata.meta.inum) {
5801 kprintf("hammer2_chain_inode_find: lookup inum %ld, "
5802 "got valid inode but with inum %ld\n",
5803 (long)inum, (long)rchain->data->ipdata.meta.inum);
5804 error = HAMMER2_ERROR_CHECK;
5805 rchain->error = error;
5806 }
5807 }
5808 }
5809 } else {
5810 error = HAMMER2_ERROR_EIO;
5811 }
5812 *parentp = parent;
5813 *chainp = rchain;
5814
5815 return error;
5816 }
5817
5818 /*
5819 * Used by the bulkscan code to snapshot the synchronized storage for
5820 * a volume, allowing it to be scanned concurrently against normal
5821 * operation.
5822 */
5823 hammer2_chain_t *
hammer2_chain_bulksnap(hammer2_dev_t * hmp)5824 hammer2_chain_bulksnap(hammer2_dev_t *hmp)
5825 {
5826 hammer2_chain_t *copy;
5827
5828 copy = hammer2_chain_alloc(hmp, hmp->spmp, &hmp->vchain.bref);
5829 copy->data = kmalloc(sizeof(copy->data->voldata),
5830 hmp->mmsg, M_WAITOK | M_ZERO);
5831 hammer2_voldata_lock(hmp);
5832 copy->data->voldata = hmp->volsync;
5833 hammer2_voldata_unlock(hmp);
5834
5835 return copy;
5836 }
5837
5838 void
hammer2_chain_bulkdrop(hammer2_chain_t * copy)5839 hammer2_chain_bulkdrop(hammer2_chain_t *copy)
5840 {
5841 KKASSERT(copy->bref.type == HAMMER2_BREF_TYPE_VOLUME);
5842 KKASSERT(copy->data);
5843 kfree(copy->data, copy->hmp->mmsg);
5844 copy->data = NULL;
5845 hammer2_chain_drop(copy);
5846 }
5847
5848 /*
5849 * Returns non-zero if the chain (INODE or DIRENT) matches the
5850 * filename.
5851 */
5852 int
hammer2_chain_dirent_test(hammer2_chain_t * chain,const char * name,size_t name_len)5853 hammer2_chain_dirent_test(hammer2_chain_t *chain, const char *name,
5854 size_t name_len)
5855 {
5856 const hammer2_inode_data_t *ripdata;
5857
5858 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
5859 ripdata = &chain->data->ipdata;
5860 if (ripdata->meta.name_len == name_len &&
5861 bcmp(ripdata->filename, name, name_len) == 0) {
5862 return 1;
5863 }
5864 }
5865 if (chain->bref.type == HAMMER2_BREF_TYPE_DIRENT &&
5866 chain->bref.embed.dirent.namlen == name_len) {
5867 if (name_len > sizeof(chain->bref.check.buf) &&
5868 bcmp(chain->data->buf, name, name_len) == 0) {
5869 return 1;
5870 }
5871 if (name_len <= sizeof(chain->bref.check.buf) &&
5872 bcmp(chain->bref.check.buf, name, name_len) == 0) {
5873 return 1;
5874 }
5875 }
5876 return 0;
5877 }
5878
5879 /*
5880 * Debugging
5881 */
5882 void
hammer2_dump_chain(hammer2_chain_t * chain,int tab,int bi,int * countp,char pfx,u_int flags)5883 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int bi, int *countp,
5884 char pfx, u_int flags)
5885 {
5886 hammer2_chain_t *scan;
5887 hammer2_chain_t *parent;
5888
5889 --*countp;
5890 if (*countp == 0) {
5891 kprintf("%*.*s...\n", tab, tab, "");
5892 return;
5893 }
5894 if (*countp < 0)
5895 return;
5896 kprintf("%*.*s%c-chain %p %s.%-3d %016jx %016jx/%-2d mir=%016jx\n",
5897 tab, tab, "", pfx, chain,
5898 hammer2_bref_type_str(chain->bref.type), bi,
5899 chain->bref.data_off, chain->bref.key, chain->bref.keybits,
5900 chain->bref.mirror_tid);
5901
5902 kprintf("%*.*s [%08x] (%s) refs=%d",
5903 tab, tab, "",
5904 chain->flags,
5905 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
5906 chain->data) ? (char *)chain->data->ipdata.filename : "?"),
5907 chain->refs);
5908
5909 parent = chain->parent;
5910 if (parent)
5911 kprintf("\n%*.*s p=%p [pflags %08x prefs %d]",
5912 tab, tab, "",
5913 parent, parent->flags, parent->refs);
5914 if (RB_EMPTY(&chain->core.rbtree)) {
5915 kprintf("\n");
5916 } else {
5917 int bi = 0;
5918 kprintf(" {\n");
5919 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree) {
5920 if ((scan->flags & flags) || flags == (u_int)-1) {
5921 hammer2_dump_chain(scan, tab + 4, bi, countp,
5922 'a', flags);
5923 }
5924 bi++;
5925 }
5926 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
5927 kprintf("%*.*s}(%s)\n", tab, tab, "",
5928 chain->data->ipdata.filename);
5929 else
5930 kprintf("%*.*s}\n", tab, tab, "");
5931 }
5932 }
5933
5934 void
hammer2_dump_chains(hammer2_dev_t * hmp,char vpfx,char fpfx)5935 hammer2_dump_chains(hammer2_dev_t *hmp, char vpfx, char fpfx)
5936 {
5937 int dumpcnt;
5938
5939 dumpcnt = 50;
5940 hammer2_dump_chain(&hmp->vchain, 0, 0, &dumpcnt, vpfx, (u_int)-1);
5941
5942 dumpcnt = 50;
5943 hammer2_dump_chain(&hmp->fchain, 0, 0, &dumpcnt, fpfx, (u_int)-1);
5944 }
5945