2 * Copyright (c) 2013-2015 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@dragonflybsd.org>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 #include <sys/param.h>
35 #include <sys/systm.h>
36 #include <sys/kernel.h>
37 #include <sys/fcntl.h>
40 #include <sys/namei.h>
41 #include <sys/mount.h>
42 #include <sys/vnode.h>
43 #include <sys/mountctl.h>
44 #include <vm/vm_kern.h>
45 #include <vm/vm_extern.h>
50 * XXX I made a mistake and made the reserved area begin at each LEVEL1 zone,
51 * which is on a 1GB demark. This will eat a little more space but for
52 * now we retain compatibility and make FMZONEBASE every 1GB
54 #define H2FMZONEBASE(key) ((key) & ~HAMMER2_FREEMAP_LEVEL1_MASK)
55 #define H2FMBASE(key, radix) ((key) & ~(((hammer2_off_t)1 << (radix)) - 1))
56 #define H2FMSHIFT(radix) ((hammer2_off_t)1 << (radix))
59 * breadth-first search
61 typedef struct hammer2_chain_save
{
62 TAILQ_ENTRY(hammer2_chain_save
) entry
;
63 hammer2_chain_t
*chain
;
65 } hammer2_chain_save_t
;
67 TAILQ_HEAD(hammer2_chain_save_list
, hammer2_chain_save
);
68 typedef struct hammer2_chain_save_list hammer2_chain_save_list_t
;
70 typedef struct hammer2_bulkfree_info
{
73 hammer2_off_t sbase
; /* sub-loop iteration */
75 hammer2_bmap_data_t
*bmap
;
77 long count_10_00
; /* staged->free */
78 long count_11_10
; /* allocated->staged */
79 long count_00_11
; /* (should not happen) */
80 long count_01_11
; /* (should not happen) */
81 long count_10_11
; /* staged->allocated */
83 long count_linadjusts
;
84 long count_inodes_scanned
;
85 long count_dedup_factor
;
87 hammer2_off_t adj_free
;
89 hammer2_tid_t saved_mirror_tid
;
91 hammer2_chain_save_list_t list
;
92 hammer2_dedup_t
*dedup
;
94 } hammer2_bulkfree_info_t
;
96 static int h2_bulkfree_test(hammer2_bulkfree_info_t
*info
,
97 hammer2_blockref_t
*bref
, int pri
);
100 * General bulk scan function with callback. Called with a referenced
101 * but UNLOCKED parent. The parent is returned in the same state.
105 hammer2_bulk_scan(hammer2_chain_t
*parent
,
106 int (*func
)(hammer2_bulkfree_info_t
*info
,
107 hammer2_blockref_t
*bref
),
108 hammer2_bulkfree_info_t
*info
)
110 hammer2_blockref_t bref
;
111 hammer2_chain_t
*chain
;
118 hammer2_chain_lock(parent
, HAMMER2_RESOLVE_ALWAYS
|
119 HAMMER2_RESOLVE_SHARED
);
125 * Generally loop on the contents if we have not been flagged
128 * Remember that these chains are completely isolated from
129 * the frontend, so we can release locks temporarily without
133 error
|= hammer2_chain_scan(parent
, &chain
, &bref
, &first
,
134 HAMMER2_LOOKUP_NODATA
|
135 HAMMER2_LOOKUP_SHARED
);
138 * Handle EOF or other error at current level. This stops
145 * Process bref, chain is only non-NULL if the bref
146 * might be recursable (its possible that we sometimes get
147 * a non-NULL chain where the bref cannot be recursed).
150 if (h2_bulkfree_test(info
, &bref
, 1))
153 error
|= func(info
, &bref
);
158 * A non-null chain is always returned if it is
159 * recursive, otherwise a non-null chain might be
160 * returned but usually is not when not recursive.
166 * Else check type and setup depth-first scan.
168 * Account for bytes actually read.
170 info
->bytes_scanned
+= chain
->bytes
;
172 switch(chain
->bref
.type
) {
173 case HAMMER2_BREF_TYPE_INODE
:
174 case HAMMER2_BREF_TYPE_FREEMAP_NODE
:
175 case HAMMER2_BREF_TYPE_INDIRECT
:
176 case HAMMER2_BREF_TYPE_VOLUME
:
177 case HAMMER2_BREF_TYPE_FREEMAP
:
179 if (info
->depth
> 16) {
180 hammer2_chain_save_t
*save
;
181 save
= kmalloc(sizeof(*save
), M_HAMMER2
,
184 hammer2_chain_ref(chain
);
185 TAILQ_INSERT_TAIL(&info
->list
, save
, entry
);
190 int savepri
= info
->pri
;
192 hammer2_chain_unlock(chain
);
193 hammer2_chain_unlock(parent
);
196 hammer2_bulk_scan(chain
, func
, info
);
197 info
->pri
+= savepri
;
198 hammer2_chain_lock(parent
,
199 HAMMER2_RESOLVE_ALWAYS
|
200 HAMMER2_RESOLVE_SHARED
);
201 hammer2_chain_lock(chain
,
202 HAMMER2_RESOLVE_ALWAYS
|
203 HAMMER2_RESOLVE_SHARED
);
208 /* does not recurse */
211 if (rup_error
& HAMMER2_ERROR_ABORTED
)
215 hammer2_chain_unlock(chain
);
216 hammer2_chain_drop(chain
);
220 * Save with higher pri now that we know what it is.
222 h2_bulkfree_test(info
, &parent
->bref
, info
->pri
+ 1);
224 hammer2_chain_unlock(parent
);
226 return ((error
| rup_error
) & ~HAMMER2_ERROR_EOF
);
233 * Chain flush (partial synchronization) XXX removed
234 * Scan the whole topology - build in-memory freemap (mark 11)
235 * Reconcile the in-memory freemap against the on-disk freemap.
236 * ondisk xx -> ondisk 11 (if allocated)
237 * ondisk 11 -> ondisk 10 (if free in-memory)
238 * ondisk 10 -> ondisk 00 (if free in-memory) - on next pass
241 * The topology scan may have to be performed multiple times to window
242 * freemaps which are too large to fit in kernel memory.
244 * Races are handled using a double-transition (11->10, 10->00). The bulkfree
245 * scan snapshots the volume root's blockset and thus can run concurrent with
246 * normal operations, as long as a full flush is made between each pass to
247 * synchronize any modified chains (otherwise their blocks might be improperly
250 * Temporary memory in multiples of 64KB is required to reconstruct the leaf
251 * hammer2_bmap_data blocks so they can later be compared against the live
252 * freemap. Each 64KB block represents 128 x 16KB x 1024 = ~2 GB of storage.
253 * A 32MB save area thus represents around ~1 TB. The temporary memory
254 * allocated can be specified. If it is not sufficient multiple topology
255 * passes will be made.
259 * Bulkfree callback info
261 static void hammer2_bulkfree_thread(void *arg __unused
);
262 static void cbinfo_bmap_init(hammer2_bulkfree_info_t
*cbinfo
, size_t size
);
263 static int h2_bulkfree_callback(hammer2_bulkfree_info_t
*cbinfo
,
264 hammer2_blockref_t
*bref
);
265 static int h2_bulkfree_sync(hammer2_bulkfree_info_t
*cbinfo
);
266 static void h2_bulkfree_sync_adjust(hammer2_bulkfree_info_t
*cbinfo
,
267 hammer2_off_t data_off
, hammer2_bmap_data_t
*live
,
268 hammer2_bmap_data_t
*bmap
, hammer2_key_t alloc_base
);
271 hammer2_bulkfree_init(hammer2_dev_t
*hmp
)
273 hammer2_thr_create(&hmp
->bfthr
, NULL
, hmp
,
274 hmp
->devrepname
, -1, -1,
275 hammer2_bulkfree_thread
);
279 hammer2_bulkfree_uninit(hammer2_dev_t
*hmp
)
281 hammer2_thr_delete(&hmp
->bfthr
);
285 hammer2_bulkfree_thread(void *arg
)
287 hammer2_thread_t
*thr
= arg
;
288 hammer2_ioc_bulkfree_t bfi
;
292 hammer2_thr_wait_any(thr
,
293 HAMMER2_THREAD_STOP
|
294 HAMMER2_THREAD_FREEZE
|
295 HAMMER2_THREAD_UNFREEZE
|
296 HAMMER2_THREAD_REMASTER
,
301 if (flags
& HAMMER2_THREAD_STOP
)
303 if (flags
& HAMMER2_THREAD_FREEZE
) {
304 hammer2_thr_signal2(thr
, HAMMER2_THREAD_FROZEN
,
305 HAMMER2_THREAD_FREEZE
);
308 if (flags
& HAMMER2_THREAD_UNFREEZE
) {
309 hammer2_thr_signal2(thr
, 0,
310 HAMMER2_THREAD_FROZEN
|
311 HAMMER2_THREAD_UNFREEZE
);
314 if (flags
& HAMMER2_THREAD_FROZEN
)
316 if (flags
& HAMMER2_THREAD_REMASTER
) {
317 hammer2_thr_signal2(thr
, 0, HAMMER2_THREAD_REMASTER
);
318 bzero(&bfi
, sizeof(bfi
));
319 bfi
.size
= 8192 * 1024;
320 /* hammer2_bulkfree_pass(thr->hmp, &bfi); */
324 hammer2_thr_signal(thr
, HAMMER2_THREAD_STOPPED
);
325 /* structure can go invalid at this point */
329 hammer2_bulkfree_pass(hammer2_dev_t
*hmp
, hammer2_chain_t
*vchain
,
330 hammer2_ioc_bulkfree_t
*bfi
)
332 hammer2_bulkfree_info_t cbinfo
;
333 hammer2_chain_save_t
*save
;
339 * We have to clear the live dedup cache as it might have entries
340 * that are freeable as of now. Any new entries in the dedup cache
341 * made after this point, even if they become freeable, will have
342 * previously been fully allocated and will be protected by the
345 hammer2_dedup_clear(hmp
);
348 * Setup for free pass
350 bzero(&cbinfo
, sizeof(cbinfo
));
351 size
= (bfi
->size
+ HAMMER2_FREEMAP_LEVELN_PSIZE
- 1) &
352 ~(size_t)(HAMMER2_FREEMAP_LEVELN_PSIZE
- 1);
354 cbinfo
.bmap
= kmem_alloc_swapbacked(&cbinfo
.kp
, size
, VM_SUBSYS_HAMMER
);
355 cbinfo
.saved_mirror_tid
= hmp
->voldata
.mirror_tid
;
357 cbinfo
.dedup
= kmalloc(sizeof(*cbinfo
.dedup
) * HAMMER2_DEDUP_HEUR_SIZE
,
358 M_HAMMER2
, M_WAITOK
| M_ZERO
);
361 * Normalize start point to a 2GB boundary. We operate on a
362 * 64KB leaf bitmap boundary which represents 2GB of storage.
364 cbinfo
.sbase
= bfi
->sbase
;
365 if (cbinfo
.sbase
> hmp
->voldata
.volu_size
)
366 cbinfo
.sbase
= hmp
->voldata
.volu_size
;
367 cbinfo
.sbase
&= ~HAMMER2_FREEMAP_LEVEL1_MASK
;
368 TAILQ_INIT(&cbinfo
.list
);
371 * Loop on a full meta-data scan as many times as required to
372 * get through all available storage.
375 while (cbinfo
.sbase
< hmp
->voldata
.volu_size
) {
377 * We have enough ram to represent (incr) bytes of storage.
378 * Each 64KB of ram represents 2GB of storage.
380 * We must also clean out our de-duplication heuristic for
381 * each (incr) bytes of storage, otherwise we wind up not
382 * scanning meta-data for later areas of storage because
383 * they had already been scanned in earlier areas of storage.
384 * Since the ranging is different, we have to restart
385 * the dedup heuristic too.
387 cbinfo_bmap_init(&cbinfo
, size
);
388 bzero(cbinfo
.dedup
, sizeof(*cbinfo
.dedup
) *
389 HAMMER2_DEDUP_HEUR_SIZE
);
390 incr
= size
/ HAMMER2_FREEMAP_LEVELN_PSIZE
*
391 HAMMER2_FREEMAP_LEVEL1_SIZE
;
392 if (hmp
->voldata
.volu_size
- cbinfo
.sbase
< incr
)
393 cbinfo
.sstop
= hmp
->voldata
.volu_size
;
395 cbinfo
.sstop
= cbinfo
.sbase
+ incr
;
396 if (hammer2_debug
& 1) {
397 kprintf("bulkfree pass %016jx/%jdGB\n",
398 (intmax_t)cbinfo
.sbase
,
399 (intmax_t)incr
/ HAMMER2_FREEMAP_LEVEL1_SIZE
);
403 * Scan topology for stuff inside this range.
405 * NOTE - By not using a transaction the operation can
406 * run concurrent with the frontend as well as
409 * We cannot safely set a mtid without a transaction,
410 * and in fact we don't want to set one anyway. We
411 * want the bulkfree to be passive and no interfere
412 * with crash recovery.
414 #undef HAMMER2_BULKFREE_TRANS /* undef - don't use transaction */
415 #ifdef HAMMER2_BULKFREE_TRANS
416 hammer2_trans_init(hmp
->spmp
, 0);
417 cbinfo
.mtid
= hammer2_trans_sub(hmp
->spmp
);
422 error
|= hammer2_bulk_scan(vchain
, h2_bulkfree_callback
,
425 while ((save
= TAILQ_FIRST(&cbinfo
.list
)) != NULL
&&
427 TAILQ_REMOVE(&cbinfo
.list
, save
, entry
);
429 error
|= hammer2_bulk_scan(save
->chain
,
430 h2_bulkfree_callback
,
432 hammer2_chain_drop(save
->chain
);
433 kfree(save
, M_HAMMER2
);
436 TAILQ_REMOVE(&cbinfo
.list
, save
, entry
);
437 hammer2_chain_drop(save
->chain
);
438 kfree(save
, M_HAMMER2
);
439 save
= TAILQ_FIRST(&cbinfo
.list
);
442 kprintf("bulkfree lastdrop %d %d error=0x%04x\n",
443 vchain
->refs
, vchain
->core
.chain_count
, error
);
446 * If complete scan succeeded we can synchronize our
447 * in-memory freemap against live storage. If an abort
448 * did occur we cannot safely synchronize our partially
449 * filled-out in-memory freemap.
452 error
= h2_bulkfree_sync(&cbinfo
);
454 hammer2_voldata_lock(hmp
);
455 hammer2_voldata_modify(hmp
);
456 hmp
->voldata
.allocator_free
+= cbinfo
.adj_free
;
457 hammer2_voldata_unlock(hmp
);
461 * Cleanup for next loop.
463 #ifdef HAMMER2_BULKFREE_TRANS
464 hammer2_trans_done(hmp
->spmp
);
468 cbinfo
.sbase
= cbinfo
.sstop
;
471 kmem_free_swapbacked(&cbinfo
.kp
);
472 kfree(cbinfo
.dedup
, M_HAMMER2
);
475 bfi
->sstop
= cbinfo
.sbase
;
477 incr
= bfi
->sstop
/ (hmp
->voldata
.volu_size
/ 10000);
481 kprintf("bulkfree pass statistics (%d.%02d%% storage processed):\n",
486 kprintf(" bulkfree was aborted\n");
488 kprintf(" transition->free %ld\n", cbinfo
.count_10_00
);
489 kprintf(" transition->staged %ld\n", cbinfo
.count_11_10
);
490 kprintf(" ERR(00)->allocated %ld\n", cbinfo
.count_00_11
);
491 kprintf(" ERR(01)->allocated %ld\n", cbinfo
.count_01_11
);
492 kprintf(" staged->allocated %ld\n", cbinfo
.count_10_11
);
493 kprintf(" ~2MB segs cleaned %ld\n", cbinfo
.count_l0cleans
);
494 kprintf(" linear adjusts %ld\n",
495 cbinfo
.count_linadjusts
);
496 kprintf(" dedup factor %ld\n",
497 cbinfo
.count_dedup_factor
);
504 cbinfo_bmap_init(hammer2_bulkfree_info_t
*cbinfo
, size_t size
)
506 hammer2_bmap_data_t
*bmap
= cbinfo
->bmap
;
507 hammer2_key_t key
= cbinfo
->sbase
;
511 lokey
= (cbinfo
->hmp
->voldata
.allocator_beg
+ HAMMER2_SEGMASK64
) &
513 hikey
= cbinfo
->hmp
->voldata
.volu_size
& ~HAMMER2_SEGMASK64
;
517 bzero(bmap
, sizeof(*bmap
));
518 if (lokey
< H2FMBASE(key
, HAMMER2_FREEMAP_LEVEL1_RADIX
))
519 lokey
= H2FMBASE(key
, HAMMER2_FREEMAP_LEVEL1_RADIX
);
520 if (lokey
< H2FMZONEBASE(key
) + HAMMER2_ZONE_SEG64
)
521 lokey
= H2FMZONEBASE(key
) + HAMMER2_ZONE_SEG64
;
522 if (key
< lokey
|| key
>= hikey
) {
523 memset(bmap
->bitmapq
, -1,
524 sizeof(bmap
->bitmapq
));
526 bmap
->linear
= HAMMER2_SEGSIZE
;
528 bmap
->avail
= H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX
);
530 size
-= sizeof(*bmap
);
531 key
+= HAMMER2_FREEMAP_LEVEL0_SIZE
;
537 h2_bulkfree_callback(hammer2_bulkfree_info_t
*cbinfo
, hammer2_blockref_t
*bref
)
539 hammer2_bmap_data_t
*bmap
;
540 hammer2_off_t data_off
;
546 * Check for signal and allow yield to userland during scan
548 if (hammer2_signal_check(&cbinfo
->save_time
))
549 return HAMMER2_ERROR_ABORTED
;
551 if (bref
->type
== HAMMER2_BREF_TYPE_INODE
) {
552 ++cbinfo
->count_inodes_scanned
;
553 if ((cbinfo
->count_inodes_scanned
& 65535) == 0)
554 kprintf(" inodes %6ld bytes %9ld\n",
555 cbinfo
->count_inodes_scanned
,
556 cbinfo
->bytes_scanned
);
560 * Calculate the data offset and determine if it is within
561 * the current freemap range being gathered.
563 data_off
= bref
->data_off
& ~HAMMER2_OFF_MASK_RADIX
;
564 if (data_off
< cbinfo
->sbase
|| data_off
>= cbinfo
->sstop
)
566 if (data_off
< cbinfo
->hmp
->voldata
.allocator_beg
)
568 if (data_off
>= cbinfo
->hmp
->voldata
.volu_size
)
572 * Calculate the information needed to generate the in-memory
575 * Hammer2 does not allow allocations to cross the L1 (2GB) boundary,
576 * it's a problem if it does. (Or L0 (2MB) for that matter).
578 radix
= (int)(bref
->data_off
& HAMMER2_OFF_MASK_RADIX
);
579 KKASSERT(radix
!= 0);
580 bytes
= (size_t)1 << radix
;
581 class = (bref
->type
<< 8) | hammer2_devblkradix(radix
);
583 if (data_off
+ bytes
> cbinfo
->sstop
) {
584 kprintf("hammer2_bulkfree_scan: illegal 2GB boundary "
585 "%016jx %016jx/%d\n",
586 (intmax_t)bref
->data_off
,
589 bytes
= cbinfo
->sstop
- data_off
; /* XXX */
593 * Convert to a storage offset relative to the beginning of the
594 * storage range we are collecting. Then lookup the level0 bmap entry.
596 data_off
-= cbinfo
->sbase
;
597 bmap
= cbinfo
->bmap
+ (data_off
>> HAMMER2_FREEMAP_LEVEL0_RADIX
);
600 * Convert data_off to a bmap-relative value (~4MB storage range).
601 * Adjust linear, class, and avail.
603 * Hammer2 does not allow allocations to cross the L0 (4MB) boundary,
605 data_off
&= HAMMER2_FREEMAP_LEVEL0_MASK
;
606 if (data_off
+ bytes
> HAMMER2_FREEMAP_LEVEL0_SIZE
) {
607 kprintf("hammer2_bulkfree_scan: illegal 4MB boundary "
608 "%016jx %016jx/%d\n",
609 (intmax_t)bref
->data_off
,
612 bytes
= HAMMER2_FREEMAP_LEVEL0_SIZE
- data_off
;
615 if (bmap
->class == 0) {
617 bmap
->avail
= HAMMER2_FREEMAP_LEVEL0_SIZE
;
621 * NOTE: bmap->class does not have to match class. Classification
622 * is relaxed when free space is low, so some mixing can occur.
628 if (bmap
->class != class) {
629 kprintf("hammer2_bulkfree_scan: illegal mixed class "
630 "%016jx %016jx/%d (%04x vs %04x)\n",
631 (intmax_t)bref
->data_off
,
639 * Just record the highest byte-granular offset for now. Do not
640 * match against allocations which are in multiples of whole blocks.
642 * Make sure that any in-block linear offset at least covers the
643 * data range. This can cause bmap->linear to become block-aligned.
645 if (bytes
& HAMMER2_FREEMAP_BLOCK_MASK
) {
646 if (bmap
->linear
< (int32_t)data_off
+ (int32_t)bytes
)
647 bmap
->linear
= (int32_t)data_off
+ (int32_t)bytes
;
648 } else if (bmap
->linear
>= (int32_t)data_off
&&
649 bmap
->linear
< (int32_t)data_off
+ (int32_t)bytes
) {
650 bmap
->linear
= (int32_t)data_off
+ (int32_t)bytes
;
654 * Adjust the hammer2_bitmap_t bitmap[HAMMER2_BMAP_ELEMENTS].
655 * 64-bit entries, 2 bits per entry, to code 11.
657 * NOTE: data_off mask to 524288, shift right by 14 (radix for 16384),
658 * and multiply shift amount by 2 for sets of 2 bits.
660 * NOTE: The allocation can be smaller than HAMMER2_FREEMAP_BLOCK_SIZE.
661 * also, data_off may not be FREEMAP_BLOCK_SIZE aligned.
664 hammer2_bitmap_t bmask
;
667 bindex
= (int)data_off
>> (HAMMER2_FREEMAP_BLOCK_RADIX
+
668 HAMMER2_BMAP_INDEX_RADIX
);
669 bmask
= (hammer2_bitmap_t
)3 <<
670 ((((int)data_off
& HAMMER2_BMAP_INDEX_MASK
) >>
671 HAMMER2_FREEMAP_BLOCK_RADIX
) << 1);
674 * NOTE! The (avail) calculation is bitmap-granular. Multiple
675 * sub-granular records can wind up at the same bitmap
678 if ((bmap
->bitmapq
[bindex
] & bmask
) == 0) {
679 if (bytes
< HAMMER2_FREEMAP_BLOCK_SIZE
) {
680 bmap
->avail
-= HAMMER2_FREEMAP_BLOCK_SIZE
;
682 bmap
->avail
-= bytes
;
684 bmap
->bitmapq
[bindex
] |= bmask
;
686 data_off
+= HAMMER2_FREEMAP_BLOCK_SIZE
;
687 if (bytes
< HAMMER2_FREEMAP_BLOCK_SIZE
)
690 bytes
-= HAMMER2_FREEMAP_BLOCK_SIZE
;
696 * Synchronize the in-memory bitmap with the live freemap. This is not a
697 * direct copy. Instead the bitmaps must be compared:
699 * In-memory Live-freemap
700 * 00 11 -> 10 (do nothing if live modified)
701 * 10 -> 00 (do nothing if live modified)
702 * 11 10 -> 11 handles race against live
703 * ** -> 11 nominally warn of corruption
705 * We must also fixup the hints in HAMMER2_BREF_TYPE_FREEMAP_LEAF.
708 h2_bulkfree_sync(hammer2_bulkfree_info_t
*cbinfo
)
710 hammer2_off_t data_off
;
712 hammer2_key_t key_dummy
;
713 hammer2_bmap_data_t
*bmap
;
714 hammer2_bmap_data_t
*live
;
715 hammer2_chain_t
*live_parent
;
716 hammer2_chain_t
*live_chain
;
720 kprintf("hammer2_bulkfree - range ");
722 if (cbinfo
->sbase
< cbinfo
->hmp
->voldata
.allocator_beg
)
724 (intmax_t)cbinfo
->hmp
->voldata
.allocator_beg
);
727 (intmax_t)cbinfo
->sbase
);
729 if (cbinfo
->sstop
> cbinfo
->hmp
->voldata
.volu_size
)
731 (intmax_t)cbinfo
->hmp
->voldata
.volu_size
);
734 (intmax_t)cbinfo
->sstop
);
736 data_off
= cbinfo
->sbase
;
739 live_parent
= &cbinfo
->hmp
->fchain
;
740 hammer2_chain_ref(live_parent
);
741 hammer2_chain_lock(live_parent
, HAMMER2_RESOLVE_ALWAYS
);
746 * Iterate each hammer2_bmap_data_t line (128 bytes) managing
749 while (data_off
< cbinfo
->sstop
) {
751 * The freemap is not used below allocator_beg or beyond
755 if (data_off
< cbinfo
->hmp
->voldata
.allocator_beg
)
757 if (data_off
>= cbinfo
->hmp
->voldata
.volu_size
)
761 * Locate the freemap leaf on the live filesystem
763 key
= (data_off
& ~HAMMER2_FREEMAP_LEVEL1_MASK
);
765 if (live_chain
== NULL
|| live_chain
->bref
.key
!= key
) {
767 hammer2_chain_unlock(live_chain
);
768 hammer2_chain_drop(live_chain
);
770 live_chain
= hammer2_chain_lookup(
774 key
+ HAMMER2_FREEMAP_LEVEL1_MASK
,
776 HAMMER2_LOOKUP_ALWAYS
);
778 kprintf("hammer2_bulkfree: freemap lookup "
779 "error near %016jx, error %s\n",
781 hammer2_error_str(live_chain
->error
));
785 if (live_chain
== NULL
) {
787 * XXX if we implement a full recovery mode we need
788 * to create/recreate missing freemap chains if our
789 * bmap has any allocated blocks.
792 bmap
->avail
!= HAMMER2_FREEMAP_LEVEL0_SIZE
) {
793 kprintf("hammer2_bulkfree: cannot locate "
794 "live leaf for allocated data "
800 if (live_chain
->error
) {
801 kprintf("hammer2_bulkfree: unable to access freemap "
802 "near %016jx, error %s\n",
804 hammer2_error_str(live_chain
->error
));
805 hammer2_chain_unlock(live_chain
);
806 hammer2_chain_drop(live_chain
);
811 bmapindex
= (data_off
& HAMMER2_FREEMAP_LEVEL1_MASK
) >>
812 HAMMER2_FREEMAP_LEVEL0_RADIX
;
813 live
= &live_chain
->data
->bmdata
[bmapindex
];
816 * Shortcut if the bitmaps match and the live linear
817 * indicator is sane. We can't do a perfect check of
818 * live->linear because the only real requirement is that
819 * if it is not block-aligned, that it not cover the space
820 * within its current block which overlaps one of the data
821 * ranges we scan. We don't retain enough fine-grained
822 * data in our scan to be able to set it exactly.
824 * TODO - we could shortcut this by testing that both
825 * live->class and bmap->class are 0, and both avails are
826 * set to HAMMER2_FREEMAP_LEVEL0_SIZE (4MB).
828 if (bcmp(live
->bitmapq
, bmap
->bitmapq
,
829 sizeof(bmap
->bitmapq
)) == 0 &&
830 live
->linear
>= bmap
->linear
) {
833 if (hammer2_debug
& 1) {
834 kprintf("live %016jx %04d.%04x (avail=%d)\n",
835 data_off
, bmapindex
, live
->class, live
->avail
);
838 hammer2_chain_modify(live_chain
, cbinfo
->mtid
, 0, 0);
839 live_chain
->bref
.check
.freemap
.bigmask
= -1;
840 cbinfo
->hmp
->freemap_relaxed
= 0; /* reset heuristic */
841 live
= &live_chain
->data
->bmdata
[bmapindex
];
843 h2_bulkfree_sync_adjust(cbinfo
, data_off
, live
, bmap
,
844 live_chain
->bref
.key
+
846 HAMMER2_FREEMAP_LEVEL0_SIZE
);
848 data_off
+= HAMMER2_FREEMAP_LEVEL0_SIZE
;
852 hammer2_chain_unlock(live_chain
);
853 hammer2_chain_drop(live_chain
);
856 hammer2_chain_unlock(live_parent
);
857 hammer2_chain_drop(live_parent
);
863 * Merge the bulkfree bitmap against the existing bitmap.
867 h2_bulkfree_sync_adjust(hammer2_bulkfree_info_t
*cbinfo
,
868 hammer2_off_t data_off
, hammer2_bmap_data_t
*live
,
869 hammer2_bmap_data_t
*bmap
, hammer2_key_t alloc_base
)
873 hammer2_off_t tmp_off
;
874 hammer2_bitmap_t lmask
;
875 hammer2_bitmap_t mmask
;
879 for (bindex
= 0; bindex
< HAMMER2_BMAP_ELEMENTS
; ++bindex
) {
880 lmask
= live
->bitmapq
[bindex
]; /* live */
881 mmask
= bmap
->bitmapq
[bindex
]; /* snapshotted bulkfree */
882 if (lmask
== mmask
) {
883 tmp_off
+= HAMMER2_BMAP_INDEX_SIZE
;
888 scount
< HAMMER2_BMAP_BITS_PER_ELEMENT
;
890 if ((mmask
& 3) == 0) {
892 * in-memory 00 live 11 -> 10
895 * Storage might be marked allocated or
896 * staged and must be remarked staged or
903 kprintf("hammer2_bulkfree: cannot "
904 "transition m=00/l=01\n");
906 case 2: /* 10 -> 00 */
907 live
->bitmapq
[bindex
] &=
908 ~((hammer2_bitmap_t
)2 << scount
);
910 HAMMER2_FREEMAP_BLOCK_SIZE
;
912 HAMMER2_FREEMAP_LEVEL0_SIZE
) {
914 HAMMER2_FREEMAP_LEVEL0_SIZE
;
917 HAMMER2_FREEMAP_BLOCK_SIZE
;
918 ++cbinfo
->count_10_00
;
919 hammer2_io_dedup_assert(
922 HAMMER2_FREEMAP_BLOCK_RADIX
,
923 HAMMER2_FREEMAP_BLOCK_SIZE
);
925 case 3: /* 11 -> 10 */
926 live
->bitmapq
[bindex
] &=
927 ~((hammer2_bitmap_t
)1 << scount
);
928 ++cbinfo
->count_11_10
;
929 hammer2_io_dedup_delete(
931 HAMMER2_BREF_TYPE_DATA
,
933 HAMMER2_FREEMAP_BLOCK_RADIX
,
934 HAMMER2_FREEMAP_BLOCK_SIZE
);
937 } else if ((mmask
& 3) == 3) {
939 * in-memory 11 live 10 -> 11
942 * Storage might be incorrectly marked free
943 * or staged and must be remarked fully
948 ++cbinfo
->count_00_11
;
950 HAMMER2_FREEMAP_BLOCK_SIZE
;
952 HAMMER2_FREEMAP_BLOCK_SIZE
;
953 if ((int32_t)live
->avail
< 0)
957 ++cbinfo
->count_01_11
;
959 case 2: /* 10 -> 11 */
960 ++cbinfo
->count_10_11
;
965 live
->bitmapq
[bindex
] |=
966 ((hammer2_bitmap_t
)3 << scount
);
970 tmp_off
+= HAMMER2_FREEMAP_BLOCK_SIZE
;
975 * Determine if the live bitmap is completely free and reset its
976 * fields if so. Otherwise check to see if we can reduce the linear
979 for (bindex
= HAMMER2_BMAP_ELEMENTS
- 1; bindex
>= 0; --bindex
) {
980 if (live
->bitmapq
[bindex
] != 0)
985 * Completely empty, reset entire segment
988 kprintf("hammer2: cleanseg %016jx.%04x (%d)\n",
989 alloc_base
, live
->class, live
->avail
);
991 live
->avail
= HAMMER2_FREEMAP_LEVEL0_SIZE
;
994 ++cbinfo
->count_l0cleans
;
995 } else if (bindex
< 7) {
997 * Partially full, bitmapq[bindex] != 0. The live->linear
998 * offset can legitimately be just about anything, but
999 * our bulkfree pass doesn't record enough information to
1000 * set it exactly. Just make sure that it is set to a
1001 * safe value that also works in our match code above (the
1002 * bcmp and linear test).
1004 * We cannot safely leave live->linear at a sub-block offset
1005 * unless it is already in the same block as bmap->linear.
1007 * If it is not in the same block, we cannot assume that
1008 * we can set it to bmap->linear on a sub-block boundary,
1009 * because the live system could have bounced it around.
1010 * In that situation we satisfy our bcmp/skip requirement
1011 * above by setting it to the nearest higher block boundary.
1012 * This alignment effectively kills any partial allocation it
1013 * might have been tracking before.
1015 if (live
->linear
< bmap
->linear
&&
1016 ((live
->linear
^ bmap
->linear
) &
1017 ~HAMMER2_FREEMAP_BLOCK_MASK
) == 0) {
1018 live
->linear
= bmap
->linear
;
1019 ++cbinfo
->count_linadjusts
;
1022 (bmap
->linear
+ HAMMER2_FREEMAP_BLOCK_MASK
) &
1023 ~HAMMER2_FREEMAP_BLOCK_MASK
;
1024 ++cbinfo
->count_linadjusts
;
1028 * Completely full, effectively disable the linear iterator
1030 live
->linear
= HAMMER2_SEGSIZE
;
1035 kprintf("%016jx %04d.%04x (avail=%7d) "
1036 "%08x %08x %08x %08x %08x %08x %08x %08x\n",
1039 HAMMER2_FREEMAP_LEVEL1_MASK
) >>
1040 HAMMER2_FREEMAP_LEVEL0_RADIX
),
1043 bmap
->bitmap
[0], bmap
->bitmap
[1],
1044 bmap
->bitmap
[2], bmap
->bitmap
[3],
1045 bmap
->bitmap
[4], bmap
->bitmap
[5],
1046 bmap
->bitmap
[6], bmap
->bitmap
[7]);
1052 * BULKFREE DEDUP HEURISTIC
1054 * WARNING! This code is SMP safe but the heuristic allows SMP collisions.
1055 * All fields must be loaded into locals and validated.
1059 h2_bulkfree_test(hammer2_bulkfree_info_t
*cbinfo
, hammer2_blockref_t
*bref
,
1062 hammer2_dedup_t
*dedup
;
1067 n
= hammer2_icrc32(&bref
->data_off
, sizeof(bref
->data_off
));
1068 dedup
= cbinfo
->dedup
+ (n
& (HAMMER2_DEDUP_HEUR_MASK
& ~7));
1070 for (i
= best
= 0; i
< 8; ++i
) {
1071 if (dedup
[i
].data_off
== bref
->data_off
) {
1072 if (dedup
[i
].ticks
< pri
)
1073 dedup
[i
].ticks
= pri
;
1075 cbinfo
->count_dedup_factor
+= dedup
[i
].ticks
;
1078 if (dedup
[i
].ticks
< dedup
[best
].ticks
)
1081 dedup
[best
].data_off
= bref
->data_off
;
1082 dedup
[best
].ticks
= pri
;