2 * Copyright (c) 2011-2018 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>
6 * by Venkatesh Srinivas <vsrinivas@dragonflybsd.org>
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in
16 * the documentation and/or other materials provided with the
18 * 3. Neither the name of The DragonFly Project nor the names of its
19 * contributors may be used to endorse or promote products derived
20 * from this software without specific, prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
23 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
24 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
25 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
26 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
27 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
28 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
29 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
30 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
31 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
32 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/kernel.h>
38 #include <sys/fcntl.h>
41 #include <sys/namei.h>
42 #include <sys/mount.h>
43 #include <sys/vnode.h>
44 #include <sys/mountctl.h>
48 #define FREEMAP_DEBUG 0
50 struct hammer2_fiterate {
57 typedef struct hammer2_fiterate hammer2_fiterate_t;
59 static int hammer2_freemap_try_alloc(hammer2_chain_t **parentp,
60 hammer2_blockref_t *bref, int radix,
61 hammer2_fiterate_t *iter, hammer2_tid_t mtid);
62 static void hammer2_freemap_init(hammer2_dev_t *hmp,
63 hammer2_key_t key, hammer2_chain_t *chain);
64 static int hammer2_bmap_alloc(hammer2_dev_t *hmp,
65 hammer2_bmap_data_t *bmap, uint16_t class,
66 int n, int sub_key, int radix, hammer2_key_t *basep);
67 static int hammer2_freemap_iterate(hammer2_chain_t **parentp,
68 hammer2_chain_t **chainp,
69 hammer2_fiterate_t *iter);
72 * Calculate the device offset for the specified FREEMAP_NODE or FREEMAP_LEAF
73 * bref. Return a combined media offset and physical size radix. Freemap
74 * chains use fixed storage offsets in the 4MB reserved area at the
75 * beginning of each 2GB zone
77 * Rotate between four possibilities. Theoretically this means we have three
78 * good freemaps in case of a crash which we can use as a base for the fixup
83 hammer2_freemap_reserve(hammer2_chain_t *chain, int radix)
85 hammer2_blockref_t *bref = &chain->bref;
92 * Physical allocation size.
94 bytes = (size_t)1 << radix;
97 * Calculate block selection index 0..7 of current block. If this
98 * is the first allocation of the block (verses a modification of an
99 * existing block), we use index 0, otherwise we use the next rotating
102 if ((bref->data_off & ~HAMMER2_OFF_MASK_RADIX) == 0) {
105 off = bref->data_off & ~HAMMER2_OFF_MASK_RADIX &
106 HAMMER2_FREEMAP_LEVEL1_MASK;
107 off = off / HAMMER2_PBUFSIZE;
108 KKASSERT(off >= HAMMER2_ZONE_FREEMAP_00 &&
109 off < HAMMER2_ZONE_FREEMAP_END);
110 index = (int)(off - HAMMER2_ZONE_FREEMAP_00) /
111 HAMMER2_ZONE_FREEMAP_INC;
112 KKASSERT(index >= 0 && index < HAMMER2_NFREEMAPS);
113 if (++index == HAMMER2_NFREEMAPS)
118 * Calculate the block offset of the reserved block. This will
119 * point into the 4MB reserved area at the base of the appropriate
120 * 2GB zone, once added to the FREEMAP_x selection above.
122 index_inc = index * HAMMER2_ZONE_FREEMAP_INC;
124 switch(bref->keybits) {
125 /* case HAMMER2_FREEMAP_LEVEL6_RADIX: not applicable */
126 case HAMMER2_FREEMAP_LEVEL5_RADIX: /* 4EB */
127 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE);
128 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
129 off = H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL5_RADIX) +
130 (index_inc + HAMMER2_ZONE_FREEMAP_00 +
131 HAMMER2_ZONEFM_LEVEL5) * HAMMER2_PBUFSIZE;
133 case HAMMER2_FREEMAP_LEVEL4_RADIX: /* 16PB */
134 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE);
135 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
136 off = H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL4_RADIX) +
137 (index_inc + HAMMER2_ZONE_FREEMAP_00 +
138 HAMMER2_ZONEFM_LEVEL4) * HAMMER2_PBUFSIZE;
140 case HAMMER2_FREEMAP_LEVEL3_RADIX: /* 64TB */
141 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE);
142 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
143 off = H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL3_RADIX) +
144 (index_inc + HAMMER2_ZONE_FREEMAP_00 +
145 HAMMER2_ZONEFM_LEVEL3) * HAMMER2_PBUFSIZE;
147 case HAMMER2_FREEMAP_LEVEL2_RADIX: /* 256GB */
148 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE);
149 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
150 off = H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL2_RADIX) +
151 (index_inc + HAMMER2_ZONE_FREEMAP_00 +
152 HAMMER2_ZONEFM_LEVEL2) * HAMMER2_PBUFSIZE;
154 case HAMMER2_FREEMAP_LEVEL1_RADIX: /* 1GB */
155 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_LEAF);
156 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
157 off = H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL1_RADIX) +
158 (index_inc + HAMMER2_ZONE_FREEMAP_00 +
159 HAMMER2_ZONEFM_LEVEL1) * HAMMER2_PBUFSIZE;
162 panic("freemap: bad radix(2) %p %d\n", bref, bref->keybits);
164 off = (hammer2_off_t)-1;
167 bref->data_off = off | radix;
169 kprintf("FREEMAP BLOCK TYPE %d %016jx/%d DATA_OFF=%016jx\n",
170 bref->type, bref->key, bref->keybits, bref->data_off);
176 * Normal freemap allocator
178 * Use available hints to allocate space using the freemap. Create missing
179 * freemap infrastructure on-the-fly as needed (including marking initial
180 * allocations using the iterator as allocated, instantiating new 2GB zones,
181 * and dealing with the end-of-media edge case).
183 * ip and bpref are only used as a heuristic to determine locality of
184 * reference. bref->key may also be used heuristically.
186 * This function is a NOP if bytes is 0.
189 hammer2_freemap_alloc(hammer2_chain_t *chain, size_t bytes)
191 hammer2_dev_t *hmp = chain->hmp;
192 hammer2_blockref_t *bref = &chain->bref;
193 hammer2_chain_t *parent;
198 hammer2_fiterate_t iter;
201 * If allocating or downsizing to zero we just get rid of whatever
205 chain->bref.data_off = 0;
210 mtid = hammer2_trans_sub(hmp->spmp);
213 * Validate the allocation size. It must be a power of 2.
215 * For now require that the caller be aware of the minimum
218 radix = hammer2_getradix(bytes);
219 KKASSERT((size_t)1 << radix == bytes);
221 if (bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE ||
222 bref->type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) {
224 * Freemap blocks themselves are assigned from the reserve
225 * area, not allocated from the freemap.
227 error = hammer2_freemap_reserve(chain, radix);
232 KKASSERT(bytes >= HAMMER2_ALLOC_MIN && bytes <= HAMMER2_ALLOC_MAX);
235 * Calculate the starting point for our allocation search.
237 * Each freemap leaf is dedicated to a specific freemap_radix.
238 * The freemap_radix can be more fine-grained than the device buffer
239 * radix which results in inodes being grouped together in their
240 * own segment, terminal-data (16K or less) and initial indirect
241 * block being grouped together, and then full-indirect and full-data
242 * blocks (64K) being grouped together.
244 * The single most important aspect of this is the inode grouping
245 * because that is what allows 'find' and 'ls' and other filesystem
246 * topology operations to run fast.
249 if (bref->data_off & ~HAMMER2_OFF_MASK_RADIX)
250 bpref = bref->data_off & ~HAMMER2_OFF_MASK_RADIX;
251 else if (trans->tmp_bpref)
252 bpref = trans->tmp_bpref;
253 else if (trans->tmp_ip)
254 bpref = trans->tmp_ip->chain->bref.data_off;
258 * Heuristic tracking index. We would like one for each distinct
259 * bref type if possible. heur_freemap[] has room for two classes
260 * for each type. At a minimum we have to break-up our heuristic
261 * by device block sizes.
263 hindex = hammer2_devblkradix(radix) - HAMMER2_MINIORADIX;
264 KKASSERT(hindex < HAMMER2_FREEMAP_HEUR_NRADIX);
265 hindex += bref->type * HAMMER2_FREEMAP_HEUR_NRADIX;
266 hindex &= HAMMER2_FREEMAP_HEUR_TYPES * HAMMER2_FREEMAP_HEUR_NRADIX - 1;
267 KKASSERT(hindex < HAMMER2_FREEMAP_HEUR_SIZE);
269 iter.bpref = hmp->heur_freemap[hindex];
270 iter.relaxed = hmp->freemap_relaxed;
273 * Make sure bpref is in-bounds. It's ok if bpref covers a zone's
274 * reserved area, the try code will iterate past it.
276 if (iter.bpref > hmp->voldata.volu_size)
277 iter.bpref = hmp->voldata.volu_size - 1;
280 * Iterate the freemap looking for free space before and after.
282 parent = &hmp->fchain;
283 hammer2_chain_ref(parent);
284 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
285 error = HAMMER2_ERROR_EAGAIN;
286 iter.bnext = iter.bpref;
289 while (error == HAMMER2_ERROR_EAGAIN) {
290 error = hammer2_freemap_try_alloc(&parent, bref, radix,
293 hmp->freemap_relaxed |= iter.relaxed; /* heuristical, SMP race ok */
294 hmp->heur_freemap[hindex] = iter.bnext;
295 hammer2_chain_unlock(parent);
296 hammer2_chain_drop(parent);
302 hammer2_freemap_try_alloc(hammer2_chain_t **parentp,
303 hammer2_blockref_t *bref, int radix,
304 hammer2_fiterate_t *iter, hammer2_tid_t mtid)
306 hammer2_dev_t *hmp = (*parentp)->hmp;
307 hammer2_off_t l0size;
308 hammer2_off_t l1size;
309 hammer2_off_t l1mask;
310 hammer2_key_t key_dummy;
311 hammer2_chain_t *chain;
318 * Calculate the number of bytes being allocated, the number
319 * of contiguous bits of bitmap being allocated, and the bitmap
322 * WARNING! cpu hardware may mask bits == 64 -> 0 and blow up the
325 bytes = (size_t)1 << radix;
326 class = (bref->type << 8) | hammer2_devblkradix(radix);
329 * Lookup the level1 freemap chain, creating and initializing one
330 * if necessary. Intermediate levels will be created automatically
331 * when necessary by hammer2_chain_create().
333 key = H2FMBASE(iter->bnext, HAMMER2_FREEMAP_LEVEL1_RADIX);
334 l0size = HAMMER2_FREEMAP_LEVEL0_SIZE;
335 l1size = HAMMER2_FREEMAP_LEVEL1_SIZE;
338 chain = hammer2_chain_lookup(parentp, &key_dummy, key, key + l1mask,
340 HAMMER2_LOOKUP_ALWAYS |
341 HAMMER2_LOOKUP_MATCHIND);
345 * Create the missing leaf, be sure to initialize
346 * the auxillary freemap tracking information in
347 * the bref.check.freemap structure.
350 kprintf("freemap create L1 @ %016jx bpref %016jx\n",
353 error = hammer2_chain_create(parentp, &chain, NULL, hmp->spmp,
354 HAMMER2_METH_DEFAULT,
355 key, HAMMER2_FREEMAP_LEVEL1_RADIX,
356 HAMMER2_BREF_TYPE_FREEMAP_LEAF,
357 HAMMER2_FREEMAP_LEVELN_PSIZE,
359 KKASSERT(error == 0);
361 hammer2_chain_modify(chain, mtid, 0, 0);
362 bzero(&chain->data->bmdata[0],
363 HAMMER2_FREEMAP_LEVELN_PSIZE);
364 chain->bref.check.freemap.bigmask = (uint32_t)-1;
365 chain->bref.check.freemap.avail = l1size;
366 /* bref.methods should already be inherited */
368 hammer2_freemap_init(hmp, key, chain);
370 } else if (chain->error) {
372 * Error during lookup.
374 kprintf("hammer2_freemap_try_alloc: %016jx: error %s\n",
375 (intmax_t)bref->data_off,
376 hammer2_error_str(chain->error));
377 error = HAMMER2_ERROR_EIO;
378 } else if ((chain->bref.check.freemap.bigmask &
379 ((size_t)1 << radix)) == 0) {
381 * Already flagged as not having enough space
383 error = HAMMER2_ERROR_ENOSPC;
386 * Modify existing chain to setup for adjustment.
388 hammer2_chain_modify(chain, mtid, 0, 0);
395 hammer2_bmap_data_t *bmap;
396 hammer2_key_t base_key;
401 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_LEAF);
402 start = (int)((iter->bnext - key) >>
403 HAMMER2_FREEMAP_LEVEL0_RADIX);
404 KKASSERT(start >= 0 && start < HAMMER2_FREEMAP_COUNT);
405 hammer2_chain_modify(chain, mtid, 0, 0);
407 error = HAMMER2_ERROR_ENOSPC;
408 for (count = 0; count < HAMMER2_FREEMAP_COUNT; ++count) {
411 if (start + count >= HAMMER2_FREEMAP_COUNT &&
417 * Calculate bmap pointer from thart starting index
420 * NOTE: bmap pointer is invalid if n >= FREEMAP_COUNT.
423 bmap = &chain->data->bmdata[n];
425 if (n >= HAMMER2_FREEMAP_COUNT) {
427 } else if (bmap->avail) {
429 } else if (radix < HAMMER2_FREEMAP_BLOCK_RADIX &&
430 (bmap->linear & HAMMER2_FREEMAP_BLOCK_MASK)) {
437 * Try to allocate from a matching freemap class
438 * superblock. If we are in relaxed mode we allocate
439 * from any freemap class superblock.
442 (bmap->class == 0 || bmap->class == class ||
444 base_key = key + n * l0size;
445 error = hammer2_bmap_alloc(hmp, bmap,
450 if (error != HAMMER2_ERROR_ENOSPC) {
457 * Calculate bmap pointer from thart starting index
458 * backwards (locality).
460 * Must recalculate after potentially having called
461 * hammer2_bmap_alloc() above in case chain was
464 * NOTE: bmap pointer is invalid if n < 0.
467 bmap = &chain->data->bmdata[n];
470 } else if (bmap->avail) {
472 } else if (radix < HAMMER2_FREEMAP_BLOCK_RADIX &&
473 (bmap->linear & HAMMER2_FREEMAP_BLOCK_MASK)) {
480 * Try to allocate from a matching freemap class
481 * superblock. If we are in relaxed mode we allocate
482 * from any freemap class superblock.
485 (bmap->class == 0 || bmap->class == class ||
487 base_key = key + n * l0size;
488 error = hammer2_bmap_alloc(hmp, bmap,
493 if (error != HAMMER2_ERROR_ENOSPC) {
501 * We only know for sure that we can clear the bitmap bit
502 * if we scanned the entire array (start == 0).
504 if (error == HAMMER2_ERROR_ENOSPC && start == 0) {
505 chain->bref.check.freemap.bigmask &=
506 (uint32_t)~((size_t)1 << radix);
508 /* XXX also scan down from original count */
513 * Assert validity. Must be beyond the static allocator used
514 * by newfs_hammer2 (and thus also beyond the aux area),
515 * not go past the volume size, and must not be in the
516 * reserved segment area for a zone.
518 KKASSERT(key >= hmp->voldata.allocator_beg &&
519 key + bytes <= hmp->voldata.volu_size);
520 KKASSERT((key & HAMMER2_ZONE_MASK64) >= HAMMER2_ZONE_SEG);
521 bref->data_off = key | radix;
524 * Record dedupability. The dedup bits are cleared
525 * when bulkfree transitions the freemap from 11->10,
526 * and asserted to be clear on the 10->00 transition.
528 * We must record the bitmask with the chain locked
529 * at the time we set the allocation bits to avoid
532 if (bref->type == HAMMER2_BREF_TYPE_DATA)
533 hammer2_io_dedup_set(hmp, bref);
535 kprintf("alloc cp=%p %016jx %016jx using %016jx\n",
537 bref->key, bref->data_off, chain->bref.data_off);
539 } else if (error == HAMMER2_ERROR_ENOSPC) {
541 * Return EAGAIN with next iteration in iter->bnext, or
542 * return ENOSPC if the allocation map has been exhausted.
544 error = hammer2_freemap_iterate(parentp, &chain, iter);
551 hammer2_chain_unlock(chain);
552 hammer2_chain_drop(chain);
558 * Allocate (1<<radix) bytes from the bmap whos base data offset is (*basep).
560 * If the linear iterator is mid-block we use it directly (the bitmap should
561 * already be marked allocated), otherwise we search for a block in the
562 * bitmap that fits the allocation request.
564 * A partial bitmap allocation sets the minimum bitmap granularity (16KB)
565 * to fully allocated and adjusts the linear allocator to allow the
566 * remaining space to be allocated.
568 * sub_key is the lower 32 bits of the chain->bref.key for the chain whos
569 * bref is being allocated. If the radix represents an allocation >= 16KB
570 * (aka HAMMER2_FREEMAP_BLOCK_RADIX) we try to use this key to select the
571 * blocks directly out of the bmap.
575 hammer2_bmap_alloc(hammer2_dev_t *hmp, hammer2_bmap_data_t *bmap,
576 uint16_t class, int n, int sub_key,
577 int radix, hammer2_key_t *basep)
582 hammer2_bitmap_t bmmask;
588 * Take into account 2-bits per block when calculating bmradix.
590 size = (size_t)1 << radix;
592 if (radix <= HAMMER2_FREEMAP_BLOCK_RADIX) {
594 /* (16K) 2 bits per allocation block */
596 bmradix = (hammer2_bitmap_t)2 <<
597 (radix - HAMMER2_FREEMAP_BLOCK_RADIX);
598 /* (32K-256K) 4, 8, 16, 32 bits per allocation block */
602 * Use the linear iterator to pack small allocations, otherwise
603 * fall-back to finding a free 16KB chunk. The linear iterator
604 * is only valid when *NOT* on a freemap chunking boundary (16KB).
605 * If it is the bitmap must be scanned. It can become invalid
606 * once we pack to the boundary. We adjust it after a bitmap
607 * allocation only for sub-16KB allocations (so the perfectly good
608 * previous value can still be used for fragments when 16KB+
609 * allocations are made inbetween fragmentary allocations).
611 * Beware of hardware artifacts when bmradix == 64 (intermediate
612 * result can wind up being '1' instead of '0' if hardware masks
615 * NOTE: j needs to be even in the j= calculation. As an artifact
616 * of the /2 division, our bitmask has to clear bit 0.
618 * NOTE: TODO this can leave little unallocatable fragments lying
621 if (((uint32_t)bmap->linear & HAMMER2_FREEMAP_BLOCK_MASK) + size <=
622 HAMMER2_FREEMAP_BLOCK_SIZE &&
623 (bmap->linear & HAMMER2_FREEMAP_BLOCK_MASK) &&
624 bmap->linear < HAMMER2_SEGSIZE) {
626 * Use linear iterator if it is not block-aligned to avoid
629 * Calculate the bitmapq[] index (i) and calculate the
630 * shift count within the 64-bit bitmapq[] entry.
632 * The freemap block size is 16KB, but each bitmap
633 * entry is two bits so use a little trick to get
634 * a (j) shift of 0, 2, 4, ... 62 in 16KB chunks.
636 KKASSERT(bmap->linear >= 0 &&
637 bmap->linear + size <= HAMMER2_SEGSIZE &&
638 (bmap->linear & (HAMMER2_ALLOC_MIN - 1)) == 0);
639 offset = bmap->linear;
640 i = offset / (HAMMER2_SEGSIZE / HAMMER2_BMAP_ELEMENTS);
641 j = (offset / (HAMMER2_FREEMAP_BLOCK_SIZE / 2)) & 62;
642 bmmask = (bmradix == HAMMER2_BMAP_BITS_PER_ELEMENT) ?
643 HAMMER2_BMAP_ALLONES :
644 ((hammer2_bitmap_t)1 << bmradix) - 1;
646 bmap->linear = offset + size;
649 * Try to index a starting point based on sub_key. This
650 * attempts to restore sequential block ordering on-disk
651 * whenever possible, even if data is committed out of
654 * i - Index bitmapq[], full data range represented is
657 * j - Index within bitmapq[i], full data range represented is
658 * HAMMER2_BMAP_INDEX_SIZE.
666 case HAMMER2_BREF_TYPE_DATA:
667 if (radix >= HAMMER2_FREEMAP_BLOCK_RADIX) {
668 i = (sub_key & HAMMER2_BMAP_MASK) /
669 (HAMMER2_BMAP_SIZE / HAMMER2_BMAP_ELEMENTS);
670 j = (sub_key & HAMMER2_BMAP_INDEX_MASK) /
671 (HAMMER2_BMAP_INDEX_SIZE /
672 HAMMER2_BMAP_BLOCKS_PER_ELEMENT);
676 case HAMMER2_BREF_TYPE_INODE:
682 KKASSERT(i < HAMMER2_BMAP_ELEMENTS &&
683 j < 2 * HAMMER2_BMAP_BLOCKS_PER_ELEMENT);
684 KKASSERT(j + bmradix <= HAMMER2_BMAP_BITS_PER_ELEMENT);
685 bmmask = (bmradix == HAMMER2_BMAP_BITS_PER_ELEMENT) ?
686 HAMMER2_BMAP_ALLONES :
687 ((hammer2_bitmap_t)1 << bmradix) - 1;
690 if ((bmap->bitmapq[i] & bmmask) == 0)
695 * General element scan.
697 * WARNING: (j) is iterating a bit index (by 2's)
699 for (i = 0; i < HAMMER2_BMAP_ELEMENTS; ++i) {
700 bmmask = (bmradix == HAMMER2_BMAP_BITS_PER_ELEMENT) ?
701 HAMMER2_BMAP_ALLONES :
702 ((hammer2_bitmap_t)1 << bmradix) - 1;
704 j < HAMMER2_BMAP_BITS_PER_ELEMENT;
706 if ((bmap->bitmapq[i] & bmmask) == 0)
711 /*fragments might remain*/
712 /*KKASSERT(bmap->avail == 0);*/
713 return (HAMMER2_ERROR_ENOSPC);
715 offset = i * (HAMMER2_SEGSIZE / HAMMER2_BMAP_ELEMENTS) +
716 (j * (HAMMER2_FREEMAP_BLOCK_SIZE / 2));
717 if (size & HAMMER2_FREEMAP_BLOCK_MASK)
718 bmap->linear = offset + size;
721 /* 8 x (64/2) -> 256 x 16K -> 4MB */
722 KKASSERT(i >= 0 && i < HAMMER2_BMAP_ELEMENTS);
725 * Optimize the buffer cache to avoid unnecessary read-before-write
728 * The device block size could be larger than the allocation size
729 * so the actual bitmap test is somewhat more involved. We have
730 * to use a compatible buffer size for this operation.
732 if ((bmap->bitmapq[i] & bmmask) == 0 &&
733 hammer2_devblksize(size) != size) {
734 size_t psize = hammer2_devblksize(size);
735 hammer2_off_t pmask = (hammer2_off_t)psize - 1;
736 int pbmradix = (hammer2_bitmap_t)2 <<
737 (hammer2_devblkradix(radix) -
738 HAMMER2_FREEMAP_BLOCK_RADIX);
739 hammer2_bitmap_t pbmmask;
740 int pradix = hammer2_getradix(psize);
742 pbmmask = (pbmradix == HAMMER2_BMAP_BITS_PER_ELEMENT) ?
743 HAMMER2_BMAP_ALLONES :
744 ((hammer2_bitmap_t)1 << pbmradix) - 1;
745 while ((pbmmask & bmmask) == 0)
746 pbmmask <<= pbmradix;
749 kprintf("%016jx mask %016jx %016jx %016jx (%zd/%zd)\n",
750 *basep + offset, bmap->bitmapq[i],
751 pbmmask, bmmask, size, psize);
754 if ((bmap->bitmapq[i] & pbmmask) == 0) {
757 hammer2_io_newnz(hmp, class >> 8,
758 (*basep + (offset & ~pmask)) |
759 pradix, psize, &dio);
760 hammer2_io_putblk(&dio);
766 * When initializing a new inode segment also attempt to initialize
767 * an adjacent segment. Be careful not to index beyond the array
770 * We do this to try to localize inode accesses to improve
771 * directory scan rates. XXX doesn't improve scan rates.
773 if (size == HAMMER2_INODE_BYTES) {
775 if (bmap[-1].radix == 0 && bmap[-1].avail)
776 bmap[-1].radix = radix;
778 if (bmap[1].radix == 0 && bmap[1].avail)
779 bmap[1].radix = radix;
784 * Calculate the bitmap-granular change in bgsize for the volume
785 * header. We cannot use the fine-grained change here because
786 * the bulkfree code can't undo it. If the bitmap element is already
787 * marked allocated it has already been accounted for.
789 if (radix < HAMMER2_FREEMAP_BLOCK_RADIX) {
790 if (bmap->bitmapq[i] & bmmask)
793 bgsize = HAMMER2_FREEMAP_BLOCK_SIZE;
799 * Adjust the bitmap, set the class (it might have been 0),
800 * and available bytes, update the allocation offset (*basep)
801 * from the L0 base to the actual offset.
803 * Do not override the class if doing a relaxed class allocation.
805 * avail must reflect the bitmap-granular availability. The allocator
806 * tests will also check the linear iterator.
808 bmap->bitmapq[i] |= bmmask;
809 if (bmap->class == 0)
811 bmap->avail -= bgsize;
815 * Adjust the volume header's allocator_free parameter. This
816 * parameter has to be fixed up by bulkfree which has no way to
817 * figure out sub-16K chunking, so it must be adjusted by the
818 * bitmap-granular size.
821 hammer2_voldata_lock(hmp);
822 hammer2_voldata_modify(hmp);
823 hmp->voldata.allocator_free -= bgsize;
824 hammer2_voldata_unlock(hmp);
831 * Initialize a freemap for the storage area (in bytes) that begins at (key).
835 hammer2_freemap_init(hammer2_dev_t *hmp, hammer2_key_t key,
836 hammer2_chain_t *chain)
838 hammer2_off_t l1size;
841 hammer2_bmap_data_t *bmap;
845 * LEVEL1 is 1GB, there are two level1 1GB freemaps per 2GB zone.
847 l1size = HAMMER2_FREEMAP_LEVEL1_SIZE;
850 * Calculate the portion of the 1GB map that should be initialized
851 * as free. Portions below or after will be initialized as allocated.
852 * SEGMASK-align the areas so we don't have to worry about sub-scans
853 * or endianess when using memset.
855 * WARNING! It is possible for lokey to be larger than hikey if the
856 * entire 2GB segment is within the static allocation.
859 * (1) Ensure that all statically allocated space from newfs_hammer2
860 * is marked allocated, and take it up to the level1 base for
863 lokey = (hmp->voldata.allocator_beg + HAMMER2_SEGMASK64) &
865 if (lokey < H2FMBASE(key, HAMMER2_FREEMAP_LEVEL1_RADIX))
866 lokey = H2FMBASE(key, HAMMER2_FREEMAP_LEVEL1_RADIX);
869 * (2) Ensure that the reserved area is marked allocated (typically
870 * the first 4MB of each 2GB area being represented). Since
871 * each LEAF represents 1GB of storage and the zone is 2GB, we
872 * have to adjust lowkey upward every other LEAF sequentially.
874 if (lokey < H2FMZONEBASE(key) + HAMMER2_ZONE_SEG64)
875 lokey = H2FMZONEBASE(key) + HAMMER2_ZONE_SEG64;
878 * (3) Ensure that any trailing space at the end-of-volume is marked
881 hikey = key + HAMMER2_FREEMAP_LEVEL1_SIZE;
882 if (hikey > hmp->voldata.volu_size) {
883 hikey = hmp->voldata.volu_size & ~HAMMER2_SEGMASK64;
887 * Heuristic highest possible value
889 chain->bref.check.freemap.avail = HAMMER2_FREEMAP_LEVEL1_SIZE;
890 bmap = &chain->data->bmdata[0];
893 * Initialize bitmap (bzero'd by caller)
895 for (count = 0; count < HAMMER2_FREEMAP_COUNT; ++count) {
896 if (key < lokey || key >= hikey) {
897 memset(bmap->bitmapq, -1,
898 sizeof(bmap->bitmapq));
900 bmap->linear = HAMMER2_SEGSIZE;
901 chain->bref.check.freemap.avail -=
902 HAMMER2_FREEMAP_LEVEL0_SIZE;
904 bmap->avail = HAMMER2_FREEMAP_LEVEL0_SIZE;
906 key += HAMMER2_FREEMAP_LEVEL0_SIZE;
912 * The current Level 1 freemap has been exhausted, iterate to the next
913 * one, return ENOSPC if no freemaps remain.
915 * At least two loops are required. If we are not in relaxed mode and
916 * we run out of storage we enter relaxed mode and do a third loop.
917 * The relaxed mode is recorded back in the hmp so once we enter the mode
918 * we remain relaxed until stuff begins to get freed and only do 2 loops.
920 * XXX this should rotate back to the beginning to handle freed-up space
921 * XXX or use intermediate entries to locate free space. TODO
924 hammer2_freemap_iterate(hammer2_chain_t **parentp, hammer2_chain_t **chainp,
925 hammer2_fiterate_t *iter)
927 hammer2_dev_t *hmp = (*parentp)->hmp;
929 iter->bnext &= ~HAMMER2_FREEMAP_LEVEL1_MASK;
930 iter->bnext += HAMMER2_FREEMAP_LEVEL1_SIZE;
931 if (iter->bnext >= hmp->voldata.volu_size) {
933 if (++iter->loops >= 2) {
934 if (iter->relaxed == 0)
937 return (HAMMER2_ERROR_ENOSPC);
940 return(HAMMER2_ERROR_EAGAIN);
944 * Adjust the bit-pattern for data in the freemap bitmap according to
945 * (how). This code is called from on-mount recovery to fixup (mark
946 * as allocated) blocks whos freemap upates might not have been committed
947 * in the last crash and is used by the bulk freemap scan to stage frees.
949 * WARNING! Cannot be called with a empty-data bref (radix == 0).
951 * XXX currently disabled when how == 0 (the normal real-time case). At
952 * the moment we depend on the bulk freescan to actually free blocks. It
953 * will still call this routine with a non-zero how to stage possible frees
954 * and to do the actual free.
957 hammer2_freemap_adjust(hammer2_dev_t *hmp, hammer2_blockref_t *bref,
960 hammer2_off_t data_off = bref->data_off;
961 hammer2_chain_t *chain;
962 hammer2_chain_t *parent;
963 hammer2_bmap_data_t *bmap;
965 hammer2_key_t key_dummy;
966 hammer2_off_t l0size;
967 hammer2_off_t l1size;
968 hammer2_off_t l1mask;
970 hammer2_bitmap_t *bitmap;
971 const hammer2_bitmap_t bmmask00 = 0;
972 hammer2_bitmap_t bmmask01;
973 hammer2_bitmap_t bmmask10;
974 hammer2_bitmap_t bmmask11;
984 KKASSERT(how == HAMMER2_FREEMAP_DORECOVER);
987 mtid = hammer2_trans_sub(hmp->spmp);
989 radix = (int)data_off & HAMMER2_OFF_MASK_RADIX;
990 KKASSERT(radix != 0);
991 data_off &= ~HAMMER2_OFF_MASK_RADIX;
992 KKASSERT(radix <= HAMMER2_RADIX_MAX);
995 bytes = (size_t)1 << radix;
998 class = (bref->type << 8) | hammer2_devblkradix(radix);
1001 * We can't adjust the freemap for data allocations made by
1004 if (data_off < hmp->voldata.allocator_beg)
1007 KKASSERT((data_off & HAMMER2_ZONE_MASK64) >= HAMMER2_ZONE_SEG);
1010 * Lookup the level1 freemap chain. The chain must exist.
1012 key = H2FMBASE(data_off, HAMMER2_FREEMAP_LEVEL1_RADIX);
1013 l0size = HAMMER2_FREEMAP_LEVEL0_SIZE;
1014 l1size = HAMMER2_FREEMAP_LEVEL1_SIZE;
1015 l1mask = l1size - 1;
1017 parent = &hmp->fchain;
1018 hammer2_chain_ref(parent);
1019 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1021 chain = hammer2_chain_lookup(&parent, &key_dummy, key, key + l1mask,
1023 HAMMER2_LOOKUP_ALWAYS |
1024 HAMMER2_LOOKUP_MATCHIND);
1027 * Stop early if we are trying to free something but no leaf exists.
1029 if (chain == NULL && how != HAMMER2_FREEMAP_DORECOVER) {
1030 kprintf("hammer2_freemap_adjust: %016jx: no chain\n",
1031 (intmax_t)bref->data_off);
1035 kprintf("hammer2_freemap_adjust: %016jx: error %s\n",
1036 (intmax_t)bref->data_off,
1037 hammer2_error_str(chain->error));
1038 hammer2_chain_unlock(chain);
1039 hammer2_chain_drop(chain);
1045 * Create any missing leaf(s) if we are doing a recovery (marking
1046 * the block(s) as being allocated instead of being freed). Be sure
1047 * to initialize the auxillary freemap tracking info in the
1048 * bref.check.freemap structure.
1050 if (chain == NULL && how == HAMMER2_FREEMAP_DORECOVER) {
1051 error = hammer2_chain_create(&parent, &chain, NULL, hmp->spmp,
1052 HAMMER2_METH_DEFAULT,
1053 key, HAMMER2_FREEMAP_LEVEL1_RADIX,
1054 HAMMER2_BREF_TYPE_FREEMAP_LEAF,
1055 HAMMER2_FREEMAP_LEVELN_PSIZE,
1058 if (hammer2_debug & 0x0040) {
1059 kprintf("fixup create chain %p %016jx:%d\n",
1060 chain, chain->bref.key, chain->bref.keybits);
1064 error = hammer2_chain_modify(chain, mtid, 0, 0);
1065 KKASSERT(error == 0);
1066 bzero(&chain->data->bmdata[0],
1067 HAMMER2_FREEMAP_LEVELN_PSIZE);
1068 chain->bref.check.freemap.bigmask = (uint32_t)-1;
1069 chain->bref.check.freemap.avail = l1size;
1070 /* bref.methods should already be inherited */
1072 hammer2_freemap_init(hmp, key, chain);
1074 /* XXX handle error */
1078 kprintf("FREEMAP ADJUST TYPE %d %016jx/%d DATA_OFF=%016jx\n",
1079 chain->bref.type, chain->bref.key,
1080 chain->bref.keybits, chain->bref.data_off);
1084 * Calculate the bitmask (runs in 2-bit pairs).
1086 start = ((int)(data_off >> HAMMER2_FREEMAP_BLOCK_RADIX) & 15) * 2;
1087 bmmask01 = (hammer2_bitmap_t)1 << start;
1088 bmmask10 = (hammer2_bitmap_t)2 << start;
1089 bmmask11 = (hammer2_bitmap_t)3 << start;
1092 * Fixup the bitmap. Partial blocks cannot be fully freed unless
1093 * a bulk scan is able to roll them up.
1095 if (radix < HAMMER2_FREEMAP_BLOCK_RADIX) {
1097 if (how == HAMMER2_FREEMAP_DOREALFREE)
1098 how = HAMMER2_FREEMAP_DOMAYFREE;
1100 count = 1 << (radix - HAMMER2_FREEMAP_BLOCK_RADIX);
1104 * [re]load the bmap and bitmap pointers. Each bmap entry covers
1105 * a 4MB swath. The bmap itself (LEVEL1) covers 2GB.
1107 * Be sure to reset the linear iterator to ensure that the adjustment
1111 bmap = &chain->data->bmdata[(int)(data_off >> HAMMER2_SEGRADIX) &
1112 (HAMMER2_FREEMAP_COUNT - 1)];
1113 bitmap = &bmap->bitmapq[(int)(data_off >> (HAMMER2_SEGRADIX - 3)) & 7];
1120 if (how == HAMMER2_FREEMAP_DORECOVER) {
1122 * Recovery request, mark as allocated.
1124 if ((*bitmap & bmmask11) != bmmask11) {
1125 if (modified == 0) {
1126 hammer2_chain_modify(chain, mtid, 0, 0);
1130 if ((*bitmap & bmmask11) == bmmask00) {
1132 HAMMER2_FREEMAP_BLOCK_SIZE;
1133 bgsize += HAMMER2_FREEMAP_BLOCK_SIZE;
1135 if (bmap->class == 0)
1136 bmap->class = class;
1137 *bitmap |= bmmask11;
1138 if (hammer2_debug & 0x0040) {
1139 kprintf("hammer2_freemap_recover: "
1141 "block=%016jx/%zd\n",
1142 bref->type, data_off, bytes);
1146 kprintf("hammer2_freemap_recover: good "
1147 "type=%02x block=%016jx/%zd\n",
1148 bref->type, data_off, bytes);
1154 * XXX this stuff doesn't work, avail is miscalculated and
1155 * code 10 means something else now.
1157 else if ((*bitmap & bmmask11) == bmmask11) {
1159 * Mayfree/Realfree request and bitmap is currently
1160 * marked as being fully allocated.
1163 hammer2_chain_modify(chain, 0);
1167 if (how == HAMMER2_FREEMAP_DOREALFREE)
1168 *bitmap &= ~bmmask11;
1170 *bitmap = (*bitmap & ~bmmask11) | bmmask10;
1171 } else if ((*bitmap & bmmask11) == bmmask10) {
1173 * Mayfree/Realfree request and bitmap is currently
1174 * marked as being possibly freeable.
1176 if (how == HAMMER2_FREEMAP_DOREALFREE) {
1178 hammer2_chain_modify(chain, 0);
1182 *bitmap &= ~bmmask11;
1186 * 01 - Not implemented, currently illegal state
1187 * 00 - Not allocated at all, illegal free.
1189 panic("hammer2_freemap_adjust: "
1190 "Illegal state %08x(%08x)",
1191 *bitmap, *bitmap & bmmask11);
1199 #if HAMMER2_BMAP_ELEMENTS != 8
1200 #error "hammer2_freemap.c: HAMMER2_BMAP_ELEMENTS expected to be 8"
1202 if (how == HAMMER2_FREEMAP_DOREALFREE && modified) {
1203 bmap->avail += 1 << radix;
1204 KKASSERT(bmap->avail <= HAMMER2_SEGSIZE);
1205 if (bmap->avail == HAMMER2_SEGSIZE &&
1206 bmap->bitmapq[0] == 0 &&
1207 bmap->bitmapq[1] == 0 &&
1208 bmap->bitmapq[2] == 0 &&
1209 bmap->bitmapq[3] == 0 &&
1210 bmap->bitmapq[4] == 0 &&
1211 bmap->bitmapq[5] == 0 &&
1212 bmap->bitmapq[6] == 0 &&
1213 bmap->bitmapq[7] == 0) {
1214 key = H2FMBASE(data_off, HAMMER2_FREEMAP_LEVEL0_RADIX);
1215 kprintf("Freeseg %016jx\n", (intmax_t)key);
1221 * chain->bref.check.freemap.bigmask (XXX)
1223 * Setting bigmask is a hint to the allocation code that there might
1224 * be something allocatable. We also set this in recovery... it
1225 * doesn't hurt and we might want to use the hint for other validation
1226 * operations later on.
1228 * We could calculate the largest possible allocation and set the
1229 * radii that could fit, but its easier just to set bigmask to -1.
1232 chain->bref.check.freemap.bigmask = -1;
1233 hmp->freemap_relaxed = 0; /* reset heuristic */
1236 hammer2_chain_unlock(chain);
1237 hammer2_chain_drop(chain);
1239 hammer2_chain_unlock(parent);
1240 hammer2_chain_drop(parent);
1243 hammer2_voldata_lock(hmp);
1244 hammer2_voldata_modify(hmp);
1245 hmp->voldata.allocator_free -= bgsize;
1246 hammer2_voldata_unlock(hmp);
1251 * Validate the freemap, in three stages.
1253 * stage-1 ALLOCATED -> POSSIBLY FREE
1254 * POSSIBLY FREE -> POSSIBLY FREE (type corrected)
1256 * This transitions bitmap entries from ALLOCATED to POSSIBLY FREE.
1257 * The POSSIBLY FREE state does not mean that a block is actually free
1258 * and may be transitioned back to ALLOCATED in stage-2.
1260 * This is typically done during normal filesystem operations when
1261 * something is deleted or a block is replaced.
1263 * This is done by bulkfree in-bulk after a memory-bounded meta-data
1264 * scan to try to determine what might be freeable.
1266 * This can be done unconditionally through a freemap scan when the
1267 * intention is to brute-force recover the proper state of the freemap.
1269 * stage-2 POSSIBLY FREE -> ALLOCATED (scan metadata topology)
1271 * This is done by bulkfree during a meta-data scan to ensure that
1272 * all blocks still actually allocated by the filesystem are marked
1275 * NOTE! Live filesystem transitions to POSSIBLY FREE can occur while
1276 * the bulkfree stage-2 and stage-3 is running. The live filesystem
1277 * will use the alternative POSSIBLY FREE type (2) to prevent
1278 * stage-3 from improperly transitioning unvetted possibly-free
1281 * stage-3 POSSIBLY FREE (type 1) -> FREE (scan freemap)
1283 * This is done by bulkfree to finalize POSSIBLY FREE states.