2 * Copyright (c) 2013-2018 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@dragonflybsd.org>
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_dirents_scanned;
86 long count_dedup_factor;
87 long count_bytes_scanned;
88 long count_chains_scanned;
89 long count_chains_reported;
92 hammer2_off_t adj_free;
94 hammer2_tid_t saved_mirror_tid;
96 hammer2_chain_save_list_t list;
97 hammer2_dedup_t *dedup;
99 } hammer2_bulkfree_info_t;
101 static int h2_bulkfree_test(hammer2_bulkfree_info_t *info,
102 hammer2_blockref_t *bref, int pri);
105 * General bulk scan function with callback. Called with a referenced
106 * but UNLOCKED parent. The parent is returned in the same state.
110 hammer2_bulk_scan(hammer2_chain_t *parent,
111 int (*func)(hammer2_bulkfree_info_t *info,
112 hammer2_blockref_t *bref),
113 hammer2_bulkfree_info_t *info)
115 hammer2_blockref_t bref;
116 hammer2_chain_t *chain;
123 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS |
124 HAMMER2_RESOLVE_SHARED);
130 * Generally loop on the contents if we have not been flagged
133 * Remember that these chains are completely isolated from
134 * the frontend, so we can release locks temporarily without
138 error |= hammer2_chain_scan(parent, &chain, &bref, &first,
139 HAMMER2_LOOKUP_NODATA |
140 HAMMER2_LOOKUP_SHARED);
143 * Handle EOF or other error at current level. This stops
150 * Account for dirents before thre data_off test, since most
151 * dirents do not need a data reference.
153 if (bref.type == HAMMER2_BREF_TYPE_DIRENT)
154 ++info->count_dirents_scanned;
157 * Ignore brefs without data (typically dirents)
159 if ((bref.data_off & ~HAMMER2_OFF_MASK_RADIX) == 0)
163 * Process bref, chain is only non-NULL if the bref
164 * might be recursable (its possible that we sometimes get
165 * a non-NULL chain where the bref cannot be recursed).
168 if (h2_bulkfree_test(info, &bref, 1))
171 if (bref.type == HAMMER2_BREF_TYPE_INODE)
172 ++info->count_inodes_scanned;
174 error |= func(info, &bref);
179 * A non-null chain is always returned if it is
180 * recursive, otherwise a non-null chain might be
181 * returned but usually is not when not recursive.
187 info->count_bytes_scanned += chain->bytes;
188 ++info->count_chains_scanned;
190 if (info->count_chains_scanned >=
191 info->count_chains_reported + 1000000 ||
192 (info->count_chains_scanned < 1000000 &&
193 info->count_chains_scanned >=
194 info->count_chains_reported + 100000)) {
195 kprintf(" chains %-7ld inodes %-7ld "
196 "dirents %-7ld bytes %5ldMB\n",
197 info->count_chains_scanned,
198 info->count_inodes_scanned,
199 info->count_dirents_scanned,
200 info->count_bytes_scanned / 1000000);
201 info->count_chains_reported =
202 info->count_chains_scanned;
208 * Else check type and setup depth-first scan.
210 * Account for bytes actually read.
212 switch(chain->bref.type) {
213 case HAMMER2_BREF_TYPE_INODE:
214 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
215 case HAMMER2_BREF_TYPE_INDIRECT:
216 case HAMMER2_BREF_TYPE_VOLUME:
217 case HAMMER2_BREF_TYPE_FREEMAP:
219 if (info->depth > 16) {
220 hammer2_chain_save_t *save;
221 save = kmalloc(sizeof(*save), M_HAMMER2,
224 hammer2_chain_ref(chain);
225 TAILQ_INSERT_TAIL(&info->list, save, entry);
230 int savepri = info->pri;
232 hammer2_chain_unlock(chain);
233 hammer2_chain_unlock(parent);
236 hammer2_bulk_scan(chain, func, info);
237 info->pri += savepri;
238 hammer2_chain_lock(parent,
239 HAMMER2_RESOLVE_ALWAYS |
240 HAMMER2_RESOLVE_SHARED);
241 hammer2_chain_lock(chain,
242 HAMMER2_RESOLVE_ALWAYS |
243 HAMMER2_RESOLVE_SHARED);
247 case HAMMER2_BREF_TYPE_DATA:
250 /* does not recurse */
253 if (rup_error & HAMMER2_ERROR_ABORTED)
257 hammer2_chain_unlock(chain);
258 hammer2_chain_drop(chain);
262 * Save with higher pri now that we know what it is.
264 h2_bulkfree_test(info, &parent->bref, info->pri + 1);
266 hammer2_chain_unlock(parent);
268 return ((error | rup_error) & ~HAMMER2_ERROR_EOF);
275 * Chain flush (partial synchronization) XXX removed
276 * Scan the whole topology - build in-memory freemap (mark 11)
277 * Reconcile the in-memory freemap against the on-disk freemap.
278 * ondisk xx -> ondisk 11 (if allocated)
279 * ondisk 11 -> ondisk 10 (if free in-memory)
280 * ondisk 10 -> ondisk 00 (if free in-memory) - on next pass
283 * The topology scan may have to be performed multiple times to window
284 * freemaps which are too large to fit in kernel memory.
286 * Races are handled using a double-transition (11->10, 10->00). The bulkfree
287 * scan snapshots the volume root's blockset and thus can run concurrent with
288 * normal operations, as long as a full flush is made between each pass to
289 * synchronize any modified chains (otherwise their blocks might be improperly
292 * Temporary memory in multiples of 64KB is required to reconstruct the leaf
293 * hammer2_bmap_data blocks so they can later be compared against the live
294 * freemap. Each 64KB block represents 128 x 16KB x 1024 = ~2 GB of storage.
295 * A 32MB save area thus represents around ~1 TB. The temporary memory
296 * allocated can be specified. If it is not sufficient multiple topology
297 * passes will be made.
301 * Bulkfree callback info
303 static void hammer2_bulkfree_thread(void *arg __unused);
304 static void cbinfo_bmap_init(hammer2_bulkfree_info_t *cbinfo, size_t size);
305 static int h2_bulkfree_callback(hammer2_bulkfree_info_t *cbinfo,
306 hammer2_blockref_t *bref);
307 static int h2_bulkfree_sync(hammer2_bulkfree_info_t *cbinfo);
308 static void h2_bulkfree_sync_adjust(hammer2_bulkfree_info_t *cbinfo,
309 hammer2_off_t data_off, hammer2_bmap_data_t *live,
310 hammer2_bmap_data_t *bmap, hammer2_key_t alloc_base);
313 hammer2_bulkfree_init(hammer2_dev_t *hmp)
315 hammer2_thr_create(&hmp->bfthr, NULL, hmp,
316 hmp->devrepname, -1, -1,
317 hammer2_bulkfree_thread);
321 hammer2_bulkfree_uninit(hammer2_dev_t *hmp)
323 hammer2_thr_delete(&hmp->bfthr);
327 hammer2_bulkfree_thread(void *arg)
329 hammer2_thread_t *thr = arg;
330 hammer2_ioc_bulkfree_t bfi;
334 hammer2_thr_wait_any(thr,
335 HAMMER2_THREAD_STOP |
336 HAMMER2_THREAD_FREEZE |
337 HAMMER2_THREAD_UNFREEZE |
338 HAMMER2_THREAD_REMASTER,
343 if (flags & HAMMER2_THREAD_STOP)
345 if (flags & HAMMER2_THREAD_FREEZE) {
346 hammer2_thr_signal2(thr, HAMMER2_THREAD_FROZEN,
347 HAMMER2_THREAD_FREEZE);
350 if (flags & HAMMER2_THREAD_UNFREEZE) {
351 hammer2_thr_signal2(thr, 0,
352 HAMMER2_THREAD_FROZEN |
353 HAMMER2_THREAD_UNFREEZE);
356 if (flags & HAMMER2_THREAD_FROZEN)
358 if (flags & HAMMER2_THREAD_REMASTER) {
359 hammer2_thr_signal2(thr, 0, HAMMER2_THREAD_REMASTER);
360 bzero(&bfi, sizeof(bfi));
361 bfi.size = 8192 * 1024;
362 /* hammer2_bulkfree_pass(thr->hmp, &bfi); */
366 hammer2_thr_signal(thr, HAMMER2_THREAD_STOPPED);
367 /* structure can go invalid at this point */
371 hammer2_bulkfree_pass(hammer2_dev_t *hmp, hammer2_chain_t *vchain,
372 hammer2_ioc_bulkfree_t *bfi)
374 hammer2_bulkfree_info_t cbinfo;
375 hammer2_chain_save_t *save;
381 * We have to clear the live dedup cache as it might have entries
382 * that are freeable as of now. Any new entries in the dedup cache
383 * made after this point, even if they become freeable, will have
384 * previously been fully allocated and will be protected by the
387 hammer2_dedup_clear(hmp);
390 * Setup for free pass using the buffer size specified by the
391 * hammer2 utility, 32K-aligned.
393 bzero(&cbinfo, sizeof(cbinfo));
394 size = (bfi->size + HAMMER2_FREEMAP_LEVELN_PSIZE - 1) &
395 ~(size_t)(HAMMER2_FREEMAP_LEVELN_PSIZE - 1);
398 * Cap at 1/4 physical memory (hammer2 utility will not normally
399 * ever specify a buffer this big, but leave the option available).
401 if (size > kmem_lim_size() * 1024 * 1024 / 4) {
402 size = kmem_lim_size() * 1024 * 1024 / 4;
403 kprintf("hammer2: Warning: capping bulkfree buffer at %jdM\n",
404 (intmax_t)size / (1024 * 1024));
407 #define HAMMER2_FREEMAP_SIZEDIV \
408 (HAMMER2_FREEMAP_LEVEL1_SIZE / HAMMER2_FREEMAP_LEVELN_PSIZE)
409 #define HAMMER2_FREEMAP_SIZEMASK (HAMMER2_FREEMAP_SIZEDIV - 1)
412 * Cap at the size needed to cover the whole volume to avoid
413 * making an unnecessarily large allocation.
415 if (size > hmp->voldata.volu_size / HAMMER2_FREEMAP_SIZEDIV) {
416 size = (hmp->voldata.volu_size + HAMMER2_FREEMAP_SIZEMASK) /
417 HAMMER2_FREEMAP_SIZEDIV;
421 * Minimum bitmap buffer size, then align to a LEVELN_PSIZE (32K)
424 if (size < 1024 * 1024)
426 size = (size + HAMMER2_FREEMAP_LEVELN_PSIZE - 1) &
427 ~(size_t)(HAMMER2_FREEMAP_LEVELN_PSIZE - 1);
430 cbinfo.bmap = kmem_alloc_swapbacked(&cbinfo.kp, size, VM_SUBSYS_HAMMER);
431 cbinfo.saved_mirror_tid = hmp->voldata.mirror_tid;
433 cbinfo.dedup = kmalloc(sizeof(*cbinfo.dedup) * HAMMER2_DEDUP_HEUR_SIZE,
434 M_HAMMER2, M_WAITOK | M_ZERO);
436 kprintf("hammer2: bulkfree buf=%jdM\n",
437 (intmax_t)size / (1024 * 1024));
440 * Normalize start point to a 2GB boundary. We operate on a
441 * 64KB leaf bitmap boundary which represents 2GB of storage.
443 cbinfo.sbase = bfi->sbase;
444 if (cbinfo.sbase > hmp->voldata.volu_size)
445 cbinfo.sbase = hmp->voldata.volu_size;
446 cbinfo.sbase &= ~HAMMER2_FREEMAP_LEVEL1_MASK;
447 TAILQ_INIT(&cbinfo.list);
449 cbinfo.bulkfree_ticks = ticks;
452 * Loop on a full meta-data scan as many times as required to
453 * get through all available storage.
456 while (cbinfo.sbase < hmp->voldata.volu_size) {
458 * We have enough ram to represent (incr) bytes of storage.
459 * Each 64KB of ram represents 2GB of storage.
461 * We must also clean out our de-duplication heuristic for
462 * each (incr) bytes of storage, otherwise we wind up not
463 * scanning meta-data for later areas of storage because
464 * they had already been scanned in earlier areas of storage.
465 * Since the ranging is different, we have to restart
466 * the dedup heuristic too.
470 cbinfo_bmap_init(&cbinfo, size);
471 bzero(cbinfo.dedup, sizeof(*cbinfo.dedup) *
472 HAMMER2_DEDUP_HEUR_SIZE);
473 cbinfo.count_inodes_scanned = 0;
474 cbinfo.count_dirents_scanned = 0;
475 cbinfo.count_bytes_scanned = 0;
476 cbinfo.count_chains_scanned = 0;
477 cbinfo.count_chains_reported = 0;
479 incr = size / HAMMER2_FREEMAP_LEVELN_PSIZE *
480 HAMMER2_FREEMAP_LEVEL1_SIZE;
481 if (hmp->voldata.volu_size - cbinfo.sbase <= incr) {
482 cbinfo.sstop = hmp->voldata.volu_size;
485 cbinfo.sstop = cbinfo.sbase + incr;
488 kprintf("hammer2: pass %016jx-%016jx ",
489 (intmax_t)cbinfo.sbase,
490 (intmax_t)cbinfo.sstop);
491 if (allmedia && cbinfo.sbase == 0)
492 kprintf("(all media)\n");
494 kprintf("(remaining media)\n");
496 kprintf("(%jdGB of media)\n",
497 (intmax_t)incr / (1024L*1024*1024));
500 * Scan topology for stuff inside this range.
502 * NOTE - By not using a transaction the operation can
503 * run concurrent with the frontend as well as
506 * We cannot safely set a mtid without a transaction,
507 * and in fact we don't want to set one anyway. We
508 * want the bulkfree to be passive and no interfere
509 * with crash recovery.
511 #undef HAMMER2_BULKFREE_TRANS /* undef - don't use transaction */
512 #ifdef HAMMER2_BULKFREE_TRANS
513 hammer2_trans_init(hmp->spmp, 0);
514 cbinfo.mtid = hammer2_trans_sub(hmp->spmp);
519 error |= hammer2_bulk_scan(vchain, h2_bulkfree_callback,
522 while ((save = TAILQ_FIRST(&cbinfo.list)) != NULL &&
524 TAILQ_REMOVE(&cbinfo.list, save, entry);
526 error |= hammer2_bulk_scan(save->chain,
527 h2_bulkfree_callback,
529 hammer2_chain_drop(save->chain);
530 kfree(save, M_HAMMER2);
533 TAILQ_REMOVE(&cbinfo.list, save, entry);
534 hammer2_chain_drop(save->chain);
535 kfree(save, M_HAMMER2);
536 save = TAILQ_FIRST(&cbinfo.list);
540 * If the complete scan succeeded we can synchronize our
541 * in-memory freemap against live storage. If an abort
542 * occured we cannot safely synchronize our partially
543 * filled-out in-memory freemap.
546 kprintf("bulkfree lastdrop %d %d error=0x%04x\n",
547 vchain->refs, vchain->core.chain_count, error);
549 kprintf("bulkfree lastdrop %d %d\n",
550 vchain->refs, vchain->core.chain_count);
552 error = h2_bulkfree_sync(&cbinfo);
554 hammer2_voldata_lock(hmp);
555 hammer2_voldata_modify(hmp);
556 hmp->voldata.allocator_free += cbinfo.adj_free;
557 hammer2_voldata_unlock(hmp);
561 * Cleanup for next loop.
563 #ifdef HAMMER2_BULKFREE_TRANS
564 hammer2_trans_done(hmp->spmp, 0);
568 cbinfo.sbase = cbinfo.sstop;
571 kmem_free_swapbacked(&cbinfo.kp);
572 kfree(cbinfo.dedup, M_HAMMER2);
575 bfi->sstop = cbinfo.sbase;
577 incr = bfi->sstop / (hmp->voldata.volu_size / 10000);
581 kprintf("bulkfree pass statistics (%d.%02d%% storage processed):\n",
586 kprintf(" bulkfree was aborted\n");
588 kprintf(" transition->free %ld\n", cbinfo.count_10_00);
589 kprintf(" transition->staged %ld\n", cbinfo.count_11_10);
590 kprintf(" ERR(00)->allocated %ld\n", cbinfo.count_00_11);
591 kprintf(" ERR(01)->allocated %ld\n", cbinfo.count_01_11);
592 kprintf(" staged->allocated %ld\n", cbinfo.count_10_11);
593 kprintf(" ~2MB segs cleaned %ld\n", cbinfo.count_l0cleans);
594 kprintf(" linear adjusts %ld\n",
595 cbinfo.count_linadjusts);
596 kprintf(" dedup factor %ld\n",
597 cbinfo.count_dedup_factor);
604 cbinfo_bmap_init(hammer2_bulkfree_info_t *cbinfo, size_t size)
606 hammer2_bmap_data_t *bmap = cbinfo->bmap;
607 hammer2_key_t key = cbinfo->sbase;
611 lokey = (cbinfo->hmp->voldata.allocator_beg + HAMMER2_SEGMASK64) &
613 hikey = cbinfo->hmp->voldata.volu_size & ~HAMMER2_SEGMASK64;
617 bzero(bmap, sizeof(*bmap));
618 if (lokey < H2FMBASE(key, HAMMER2_FREEMAP_LEVEL1_RADIX))
619 lokey = H2FMBASE(key, HAMMER2_FREEMAP_LEVEL1_RADIX);
620 if (lokey < H2FMZONEBASE(key) + HAMMER2_ZONE_SEG64)
621 lokey = H2FMZONEBASE(key) + HAMMER2_ZONE_SEG64;
622 if (key < lokey || key >= hikey) {
623 memset(bmap->bitmapq, -1,
624 sizeof(bmap->bitmapq));
626 bmap->linear = HAMMER2_SEGSIZE;
628 bmap->avail = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
630 size -= sizeof(*bmap);
631 key += HAMMER2_FREEMAP_LEVEL0_SIZE;
637 h2_bulkfree_callback(hammer2_bulkfree_info_t *cbinfo, hammer2_blockref_t *bref)
639 hammer2_bmap_data_t *bmap;
640 hammer2_off_t data_off;
646 * Check for signal and allow yield to userland during scan.
648 if (hammer2_signal_check(&cbinfo->save_time))
649 return HAMMER2_ERROR_ABORTED;
652 * Deal with kernel thread cpu or I/O hogging by limiting the
653 * number of chains scanned per second to hammer2_bulkfree_tps.
654 * Ignore leaf records (DIRENT and DATA), no per-record I/O is
655 * involved for those since we don't load their data.
657 if (bref->type != HAMMER2_BREF_TYPE_DATA &&
658 bref->type != HAMMER2_BREF_TYPE_DIRENT) {
659 ++cbinfo->bulkfree_calls;
660 if (cbinfo->bulkfree_calls > hammer2_bulkfree_tps) {
661 int dticks = ticks - cbinfo->bulkfree_ticks;
665 tsleep(&cbinfo->bulkfree_ticks, 0,
666 "h2bw", hz - dticks);
668 cbinfo->bulkfree_calls = 0;
669 cbinfo->bulkfree_ticks = ticks;
674 * Calculate the data offset and determine if it is within
675 * the current freemap range being gathered.
677 data_off = bref->data_off & ~HAMMER2_OFF_MASK_RADIX;
678 if (data_off < cbinfo->sbase || data_off >= cbinfo->sstop)
680 if (data_off < cbinfo->hmp->voldata.allocator_beg)
682 if (data_off >= cbinfo->hmp->voldata.volu_size)
686 * Calculate the information needed to generate the in-memory
689 * Hammer2 does not allow allocations to cross the L1 (2GB) boundary,
690 * it's a problem if it does. (Or L0 (2MB) for that matter).
692 radix = (int)(bref->data_off & HAMMER2_OFF_MASK_RADIX);
693 KKASSERT(radix != 0);
694 bytes = (size_t)1 << radix;
695 class = (bref->type << 8) | hammer2_devblkradix(radix);
697 if (data_off + bytes > cbinfo->sstop) {
698 kprintf("hammer2_bulkfree_scan: illegal 2GB boundary "
699 "%016jx %016jx/%d\n",
700 (intmax_t)bref->data_off,
703 bytes = cbinfo->sstop - data_off; /* XXX */
707 * Convert to a storage offset relative to the beginning of the
708 * storage range we are collecting. Then lookup the level0 bmap entry.
710 data_off -= cbinfo->sbase;
711 bmap = cbinfo->bmap + (data_off >> HAMMER2_FREEMAP_LEVEL0_RADIX);
714 * Convert data_off to a bmap-relative value (~4MB storage range).
715 * Adjust linear, class, and avail.
717 * Hammer2 does not allow allocations to cross the L0 (4MB) boundary,
719 data_off &= HAMMER2_FREEMAP_LEVEL0_MASK;
720 if (data_off + bytes > HAMMER2_FREEMAP_LEVEL0_SIZE) {
721 kprintf("hammer2_bulkfree_scan: illegal 4MB boundary "
722 "%016jx %016jx/%d\n",
723 (intmax_t)bref->data_off,
726 bytes = HAMMER2_FREEMAP_LEVEL0_SIZE - data_off;
729 if (bmap->class == 0) {
731 bmap->avail = HAMMER2_FREEMAP_LEVEL0_SIZE;
735 * NOTE: bmap->class does not have to match class. Classification
736 * is relaxed when free space is low, so some mixing can occur.
742 if (bmap->class != class) {
743 kprintf("hammer2_bulkfree_scan: illegal mixed class "
744 "%016jx %016jx/%d (%04x vs %04x)\n",
745 (intmax_t)bref->data_off,
753 * Just record the highest byte-granular offset for now. Do not
754 * match against allocations which are in multiples of whole blocks.
756 * Make sure that any in-block linear offset at least covers the
757 * data range. This can cause bmap->linear to become block-aligned.
759 if (bytes & HAMMER2_FREEMAP_BLOCK_MASK) {
760 if (bmap->linear < (int32_t)data_off + (int32_t)bytes)
761 bmap->linear = (int32_t)data_off + (int32_t)bytes;
762 } else if (bmap->linear >= (int32_t)data_off &&
763 bmap->linear < (int32_t)data_off + (int32_t)bytes) {
764 bmap->linear = (int32_t)data_off + (int32_t)bytes;
768 * Adjust the hammer2_bitmap_t bitmap[HAMMER2_BMAP_ELEMENTS].
769 * 64-bit entries, 2 bits per entry, to code 11.
771 * NOTE: data_off mask to 524288, shift right by 14 (radix for 16384),
772 * and multiply shift amount by 2 for sets of 2 bits.
774 * NOTE: The allocation can be smaller than HAMMER2_FREEMAP_BLOCK_SIZE.
775 * also, data_off may not be FREEMAP_BLOCK_SIZE aligned.
778 hammer2_bitmap_t bmask;
781 bindex = (int)data_off >> (HAMMER2_FREEMAP_BLOCK_RADIX +
782 HAMMER2_BMAP_INDEX_RADIX);
783 bmask = (hammer2_bitmap_t)3 <<
784 ((((int)data_off & HAMMER2_BMAP_INDEX_MASK) >>
785 HAMMER2_FREEMAP_BLOCK_RADIX) << 1);
788 * NOTE! The (avail) calculation is bitmap-granular. Multiple
789 * sub-granular records can wind up at the same bitmap
792 if ((bmap->bitmapq[bindex] & bmask) == 0) {
793 if (bytes < HAMMER2_FREEMAP_BLOCK_SIZE) {
794 bmap->avail -= HAMMER2_FREEMAP_BLOCK_SIZE;
796 bmap->avail -= bytes;
798 bmap->bitmapq[bindex] |= bmask;
800 data_off += HAMMER2_FREEMAP_BLOCK_SIZE;
801 if (bytes < HAMMER2_FREEMAP_BLOCK_SIZE)
804 bytes -= HAMMER2_FREEMAP_BLOCK_SIZE;
810 * Synchronize the in-memory bitmap with the live freemap. This is not a
811 * direct copy. Instead the bitmaps must be compared:
813 * In-memory Live-freemap
814 * 00 11 -> 10 (do nothing if live modified)
815 * 10 -> 00 (do nothing if live modified)
816 * 11 10 -> 11 handles race against live
817 * ** -> 11 nominally warn of corruption
819 * We must also fixup the hints in HAMMER2_BREF_TYPE_FREEMAP_LEAF.
822 h2_bulkfree_sync(hammer2_bulkfree_info_t *cbinfo)
824 hammer2_off_t data_off;
826 hammer2_key_t key_dummy;
827 hammer2_bmap_data_t *bmap;
828 hammer2_bmap_data_t *live;
829 hammer2_chain_t *live_parent;
830 hammer2_chain_t *live_chain;
834 kprintf("hammer2_bulkfree - range ");
836 if (cbinfo->sbase < cbinfo->hmp->voldata.allocator_beg)
838 (intmax_t)cbinfo->hmp->voldata.allocator_beg);
841 (intmax_t)cbinfo->sbase);
843 if (cbinfo->sstop > cbinfo->hmp->voldata.volu_size)
845 (intmax_t)cbinfo->hmp->voldata.volu_size);
848 (intmax_t)cbinfo->sstop);
850 data_off = cbinfo->sbase;
853 live_parent = &cbinfo->hmp->fchain;
854 hammer2_chain_ref(live_parent);
855 hammer2_chain_lock(live_parent, HAMMER2_RESOLVE_ALWAYS);
860 * Iterate each hammer2_bmap_data_t line (128 bytes) managing
863 while (data_off < cbinfo->sstop) {
865 * The freemap is not used below allocator_beg or beyond
869 if (data_off < cbinfo->hmp->voldata.allocator_beg)
871 if (data_off >= cbinfo->hmp->voldata.volu_size)
875 * Locate the freemap leaf on the live filesystem
877 key = (data_off & ~HAMMER2_FREEMAP_LEVEL1_MASK);
879 if (live_chain == NULL || live_chain->bref.key != key) {
881 hammer2_chain_unlock(live_chain);
882 hammer2_chain_drop(live_chain);
884 live_chain = hammer2_chain_lookup(
888 key + HAMMER2_FREEMAP_LEVEL1_MASK,
890 HAMMER2_LOOKUP_ALWAYS);
892 kprintf("hammer2_bulkfree: freemap lookup "
893 "error near %016jx, error %s\n",
895 hammer2_error_str(live_chain->error));
899 if (live_chain == NULL) {
901 * XXX if we implement a full recovery mode we need
902 * to create/recreate missing freemap chains if our
903 * bmap has any allocated blocks.
906 bmap->avail != HAMMER2_FREEMAP_LEVEL0_SIZE) {
907 kprintf("hammer2_bulkfree: cannot locate "
908 "live leaf for allocated data "
914 if (live_chain->error) {
915 kprintf("hammer2_bulkfree: unable to access freemap "
916 "near %016jx, error %s\n",
918 hammer2_error_str(live_chain->error));
919 hammer2_chain_unlock(live_chain);
920 hammer2_chain_drop(live_chain);
925 bmapindex = (data_off & HAMMER2_FREEMAP_LEVEL1_MASK) >>
926 HAMMER2_FREEMAP_LEVEL0_RADIX;
927 live = &live_chain->data->bmdata[bmapindex];
930 * Shortcut if the bitmaps match and the live linear
931 * indicator is sane. We can't do a perfect check of
932 * live->linear because the only real requirement is that
933 * if it is not block-aligned, that it not cover the space
934 * within its current block which overlaps one of the data
935 * ranges we scan. We don't retain enough fine-grained
936 * data in our scan to be able to set it exactly.
938 * TODO - we could shortcut this by testing that both
939 * live->class and bmap->class are 0, and both avails are
940 * set to HAMMER2_FREEMAP_LEVEL0_SIZE (4MB).
942 if (bcmp(live->bitmapq, bmap->bitmapq,
943 sizeof(bmap->bitmapq)) == 0 &&
944 live->linear >= bmap->linear) {
947 if (hammer2_debug & 1) {
948 kprintf("live %016jx %04d.%04x (avail=%d)\n",
949 data_off, bmapindex, live->class, live->avail);
952 hammer2_chain_modify(live_chain, cbinfo->mtid, 0, 0);
953 live_chain->bref.check.freemap.bigmask = -1;
954 cbinfo->hmp->freemap_relaxed = 0; /* reset heuristic */
955 live = &live_chain->data->bmdata[bmapindex];
957 h2_bulkfree_sync_adjust(cbinfo, data_off, live, bmap,
958 live_chain->bref.key +
960 HAMMER2_FREEMAP_LEVEL0_SIZE);
962 data_off += HAMMER2_FREEMAP_LEVEL0_SIZE;
966 hammer2_chain_unlock(live_chain);
967 hammer2_chain_drop(live_chain);
970 hammer2_chain_unlock(live_parent);
971 hammer2_chain_drop(live_parent);
977 * Merge the bulkfree bitmap against the existing bitmap.
981 h2_bulkfree_sync_adjust(hammer2_bulkfree_info_t *cbinfo,
982 hammer2_off_t data_off, hammer2_bmap_data_t *live,
983 hammer2_bmap_data_t *bmap, hammer2_key_t alloc_base)
987 hammer2_off_t tmp_off;
988 hammer2_bitmap_t lmask;
989 hammer2_bitmap_t mmask;
993 for (bindex = 0; bindex < HAMMER2_BMAP_ELEMENTS; ++bindex) {
994 lmask = live->bitmapq[bindex]; /* live */
995 mmask = bmap->bitmapq[bindex]; /* snapshotted bulkfree */
996 if (lmask == mmask) {
997 tmp_off += HAMMER2_BMAP_INDEX_SIZE;
1002 scount < HAMMER2_BMAP_BITS_PER_ELEMENT;
1004 if ((mmask & 3) == 0) {
1006 * in-memory 00 live 11 -> 10
1009 * Storage might be marked allocated or
1010 * staged and must be remarked staged or
1013 switch (lmask & 3) {
1017 kprintf("hammer2_bulkfree: cannot "
1018 "transition m=00/l=01\n");
1020 case 2: /* 10 -> 00 */
1021 live->bitmapq[bindex] &=
1022 ~((hammer2_bitmap_t)2 << scount);
1024 HAMMER2_FREEMAP_BLOCK_SIZE;
1026 HAMMER2_FREEMAP_LEVEL0_SIZE) {
1028 HAMMER2_FREEMAP_LEVEL0_SIZE;
1031 HAMMER2_FREEMAP_BLOCK_SIZE;
1032 ++cbinfo->count_10_00;
1033 hammer2_io_dedup_assert(
1036 HAMMER2_FREEMAP_BLOCK_RADIX,
1037 HAMMER2_FREEMAP_BLOCK_SIZE);
1039 case 3: /* 11 -> 10 */
1040 live->bitmapq[bindex] &=
1041 ~((hammer2_bitmap_t)1 << scount);
1042 ++cbinfo->count_11_10;
1043 hammer2_io_dedup_delete(
1045 HAMMER2_BREF_TYPE_DATA,
1047 HAMMER2_FREEMAP_BLOCK_RADIX,
1048 HAMMER2_FREEMAP_BLOCK_SIZE);
1051 } else if ((mmask & 3) == 3) {
1053 * in-memory 11 live 10 -> 11
1056 * Storage might be incorrectly marked free
1057 * or staged and must be remarked fully
1060 switch (lmask & 3) {
1062 ++cbinfo->count_00_11;
1064 HAMMER2_FREEMAP_BLOCK_SIZE;
1066 HAMMER2_FREEMAP_BLOCK_SIZE;
1067 if ((int32_t)live->avail < 0)
1071 ++cbinfo->count_01_11;
1073 case 2: /* 10 -> 11 */
1074 ++cbinfo->count_10_11;
1079 live->bitmapq[bindex] |=
1080 ((hammer2_bitmap_t)3 << scount);
1084 tmp_off += HAMMER2_FREEMAP_BLOCK_SIZE;
1089 * Determine if the live bitmap is completely free and reset its
1090 * fields if so. Otherwise check to see if we can reduce the linear
1093 for (bindex = HAMMER2_BMAP_ELEMENTS - 1; bindex >= 0; --bindex) {
1094 if (live->bitmapq[bindex] != 0)
1099 * Completely empty, reset entire segment
1102 kprintf("hammer2: cleanseg %016jx.%04x (%d)\n",
1103 alloc_base, live->class, live->avail);
1105 live->avail = HAMMER2_FREEMAP_LEVEL0_SIZE;
1108 ++cbinfo->count_l0cleans;
1109 } else if (bindex < 7) {
1111 * Partially full, bitmapq[bindex] != 0. Our bulkfree pass
1112 * does not record enough information to set live->linear
1115 * NOTE: Setting live->linear to a sub-block (16K) boundary
1116 * forces the live code to iterate to the next fully
1117 * free block. It does NOT mean that all blocks above
1118 * live->linear are available.
1120 * Setting live->linear to a fragmentary (less than
1121 * 16K) boundary allows allocations to iterate within
1124 if (live->linear < bmap->linear &&
1125 ((live->linear ^ bmap->linear) &
1126 ~HAMMER2_FREEMAP_BLOCK_MASK) == 0) {
1128 * If greater than but still within the same
1129 * sub-block as live we can adjust linear upward.
1131 live->linear = bmap->linear;
1132 ++cbinfo->count_linadjusts;
1135 * Otherwise adjust to the nearest higher or same
1136 * sub-block boundary. The live system may have
1137 * bounced live->linear around so we cannot make any
1138 * assumptions with regards to available fragmentary
1142 (bmap->linear + HAMMER2_FREEMAP_BLOCK_MASK) &
1143 ~HAMMER2_FREEMAP_BLOCK_MASK;
1144 ++cbinfo->count_linadjusts;
1148 * Completely full, effectively disable the linear iterator
1150 live->linear = HAMMER2_SEGSIZE;
1155 kprintf("%016jx %04d.%04x (avail=%7d) "
1156 "%08x %08x %08x %08x %08x %08x %08x %08x\n",
1159 HAMMER2_FREEMAP_LEVEL1_MASK) >>
1160 HAMMER2_FREEMAP_LEVEL0_RADIX),
1163 bmap->bitmap[0], bmap->bitmap[1],
1164 bmap->bitmap[2], bmap->bitmap[3],
1165 bmap->bitmap[4], bmap->bitmap[5],
1166 bmap->bitmap[6], bmap->bitmap[7]);
1172 * BULKFREE DEDUP HEURISTIC
1174 * WARNING! This code is SMP safe but the heuristic allows SMP collisions.
1175 * All fields must be loaded into locals and validated.
1179 h2_bulkfree_test(hammer2_bulkfree_info_t *cbinfo, hammer2_blockref_t *bref,
1182 hammer2_dedup_t *dedup;
1187 n = hammer2_icrc32(&bref->data_off, sizeof(bref->data_off));
1188 dedup = cbinfo->dedup + (n & (HAMMER2_DEDUP_HEUR_MASK & ~7));
1190 for (i = best = 0; i < 8; ++i) {
1191 if (dedup[i].data_off == bref->data_off) {
1192 if (dedup[i].ticks < pri)
1193 dedup[i].ticks = pri;
1195 cbinfo->count_dedup_factor += dedup[i].ticks;
1198 if (dedup[i].ticks < dedup[best].ticks)
1201 dedup[best].data_off = bref->data_off;
1202 dedup[best].ticks = pri;