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 * XXX I made a mistake and made the reserved area begin at each LEVEL1 zone,
78 * which is on a 1GB demark. This will eat a little more space but for
79 * now we retain compatibility and make FMZONEBASE every 1GB
81 * (see same thing in hammer2_bulkfree.c near the top, as well as in
84 * Rotate between four possibilities. Theoretically this means we have three
85 * good freemaps in case of a crash which we can use as a base for the fixup
88 #define H2FMZONEBASE(key) ((key) & ~HAMMER2_FREEMAP_LEVEL1_MASK)
89 #define H2FMBASE(key, radix) ((key) & ~(((hammer2_off_t)1 << (radix)) - 1))
90 #define H2FMSHIFT(radix) ((hammer2_off_t)1 << (radix))
94 hammer2_freemap_reserve(hammer2_chain_t *chain, int radix)
96 hammer2_blockref_t *bref = &chain->bref;
103 * Physical allocation size.
105 bytes = (size_t)1 << radix;
108 * Calculate block selection index 0..7 of current block. If this
109 * is the first allocation of the block (verses a modification of an
110 * existing block), we use index 0, otherwise we use the next rotating
113 if ((bref->data_off & ~HAMMER2_OFF_MASK_RADIX) == 0) {
116 off = bref->data_off & ~HAMMER2_OFF_MASK_RADIX &
117 (HAMMER2_FREEMAP_LEVEL1_SIZE - 1);
118 off = off / HAMMER2_PBUFSIZE;
119 KKASSERT(off >= HAMMER2_ZONE_FREEMAP_00 &&
120 off < HAMMER2_ZONE_FREEMAP_END);
121 index = (int)(off - HAMMER2_ZONE_FREEMAP_00) /
122 HAMMER2_ZONE_FREEMAP_INC;
123 KKASSERT(index >= 0 && index < HAMMER2_NFREEMAPS);
124 if (++index == HAMMER2_NFREEMAPS)
129 * Calculate the block offset of the reserved block. This will
130 * point into the 4MB reserved area at the base of the appropriate
131 * 2GB zone, once added to the FREEMAP_x selection above.
133 index_inc = index * HAMMER2_ZONE_FREEMAP_INC;
135 switch(bref->keybits) {
136 /* case HAMMER2_FREEMAP_LEVEL6_RADIX: not applicable */
137 case HAMMER2_FREEMAP_LEVEL5_RADIX: /* 4EB */
138 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE);
139 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
140 off = H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL5_RADIX) +
141 (index_inc + HAMMER2_ZONE_FREEMAP_00 +
142 HAMMER2_ZONEFM_LEVEL5) * HAMMER2_PBUFSIZE;
144 case HAMMER2_FREEMAP_LEVEL4_RADIX: /* 16PB */
145 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE);
146 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
147 off = H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL4_RADIX) +
148 (index_inc + HAMMER2_ZONE_FREEMAP_00 +
149 HAMMER2_ZONEFM_LEVEL4) * HAMMER2_PBUFSIZE;
151 case HAMMER2_FREEMAP_LEVEL3_RADIX: /* 64TB */
152 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE);
153 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
154 off = H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL3_RADIX) +
155 (index_inc + HAMMER2_ZONE_FREEMAP_00 +
156 HAMMER2_ZONEFM_LEVEL3) * HAMMER2_PBUFSIZE;
158 case HAMMER2_FREEMAP_LEVEL2_RADIX: /* 256GB */
159 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE);
160 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
161 off = H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL2_RADIX) +
162 (index_inc + HAMMER2_ZONE_FREEMAP_00 +
163 HAMMER2_ZONEFM_LEVEL2) * HAMMER2_PBUFSIZE;
165 case HAMMER2_FREEMAP_LEVEL1_RADIX: /* 1GB */
166 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_LEAF);
167 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
168 off = H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL1_RADIX) +
169 (index_inc + HAMMER2_ZONE_FREEMAP_00 +
170 HAMMER2_ZONEFM_LEVEL1) * HAMMER2_PBUFSIZE;
173 panic("freemap: bad radix(2) %p %d\n", bref, bref->keybits);
175 off = (hammer2_off_t)-1;
178 bref->data_off = off | radix;
180 kprintf("FREEMAP BLOCK TYPE %d %016jx/%d DATA_OFF=%016jx\n",
181 bref->type, bref->key, bref->keybits, bref->data_off);
187 * Normal freemap allocator
189 * Use available hints to allocate space using the freemap. Create missing
190 * freemap infrastructure on-the-fly as needed (including marking initial
191 * allocations using the iterator as allocated, instantiating new 2GB zones,
192 * and dealing with the end-of-media edge case).
194 * ip and bpref are only used as a heuristic to determine locality of
195 * reference. bref->key may also be used heuristically.
197 * This function is a NOP if bytes is 0.
200 hammer2_freemap_alloc(hammer2_chain_t *chain, size_t bytes)
202 hammer2_dev_t *hmp = chain->hmp;
203 hammer2_blockref_t *bref = &chain->bref;
204 hammer2_chain_t *parent;
209 hammer2_fiterate_t iter;
212 * If allocating or downsizing to zero we just get rid of whatever
216 chain->bref.data_off = 0;
221 mtid = hammer2_trans_sub(hmp->spmp);
224 * Validate the allocation size. It must be a power of 2.
226 * For now require that the caller be aware of the minimum
229 radix = hammer2_getradix(bytes);
230 KKASSERT((size_t)1 << radix == bytes);
232 if (bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE ||
233 bref->type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) {
235 * Freemap blocks themselves are assigned from the reserve
236 * area, not allocated from the freemap.
238 error = hammer2_freemap_reserve(chain, radix);
243 KKASSERT(bytes >= HAMMER2_ALLOC_MIN && bytes <= HAMMER2_ALLOC_MAX);
246 * Calculate the starting point for our allocation search.
248 * Each freemap leaf is dedicated to a specific freemap_radix.
249 * The freemap_radix can be more fine-grained than the device buffer
250 * radix which results in inodes being grouped together in their
251 * own segment, terminal-data (16K or less) and initial indirect
252 * block being grouped together, and then full-indirect and full-data
253 * blocks (64K) being grouped together.
255 * The single most important aspect of this is the inode grouping
256 * because that is what allows 'find' and 'ls' and other filesystem
257 * topology operations to run fast.
260 if (bref->data_off & ~HAMMER2_OFF_MASK_RADIX)
261 bpref = bref->data_off & ~HAMMER2_OFF_MASK_RADIX;
262 else if (trans->tmp_bpref)
263 bpref = trans->tmp_bpref;
264 else if (trans->tmp_ip)
265 bpref = trans->tmp_ip->chain->bref.data_off;
269 * Heuristic tracking index. We would like one for each distinct
270 * bref type if possible. heur_freemap[] has room for two classes
271 * for each type. At a minimum we have to break-up our heuristic
272 * by device block sizes.
274 hindex = hammer2_devblkradix(radix) - HAMMER2_MINIORADIX;
275 KKASSERT(hindex < HAMMER2_FREEMAP_HEUR_NRADIX);
276 hindex += bref->type * HAMMER2_FREEMAP_HEUR_NRADIX;
277 hindex &= HAMMER2_FREEMAP_HEUR_TYPES * HAMMER2_FREEMAP_HEUR_NRADIX - 1;
278 KKASSERT(hindex < HAMMER2_FREEMAP_HEUR_SIZE);
280 iter.bpref = hmp->heur_freemap[hindex];
281 iter.relaxed = hmp->freemap_relaxed;
284 * Make sure bpref is in-bounds. It's ok if bpref covers a zone's
285 * reserved area, the try code will iterate past it.
287 if (iter.bpref > hmp->voldata.volu_size)
288 iter.bpref = hmp->voldata.volu_size - 1;
291 * Iterate the freemap looking for free space before and after.
293 parent = &hmp->fchain;
294 hammer2_chain_ref(parent);
295 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
296 error = HAMMER2_ERROR_EAGAIN;
297 iter.bnext = iter.bpref;
300 while (error == HAMMER2_ERROR_EAGAIN) {
301 error = hammer2_freemap_try_alloc(&parent, bref, radix,
304 hmp->freemap_relaxed |= iter.relaxed; /* heuristical, SMP race ok */
305 hmp->heur_freemap[hindex] = iter.bnext;
306 hammer2_chain_unlock(parent);
307 hammer2_chain_drop(parent);
313 hammer2_freemap_try_alloc(hammer2_chain_t **parentp,
314 hammer2_blockref_t *bref, int radix,
315 hammer2_fiterate_t *iter, hammer2_tid_t mtid)
317 hammer2_dev_t *hmp = (*parentp)->hmp;
318 hammer2_off_t l0size;
319 hammer2_off_t l1size;
320 hammer2_off_t l1mask;
321 hammer2_key_t key_dummy;
322 hammer2_chain_t *chain;
329 * Calculate the number of bytes being allocated, the number
330 * of contiguous bits of bitmap being allocated, and the bitmap
333 * WARNING! cpu hardware may mask bits == 64 -> 0 and blow up the
336 bytes = (size_t)1 << radix;
337 class = (bref->type << 8) | hammer2_devblkradix(radix);
340 * Lookup the level1 freemap chain, creating and initializing one
341 * if necessary. Intermediate levels will be created automatically
342 * when necessary by hammer2_chain_create().
344 key = H2FMBASE(iter->bnext, HAMMER2_FREEMAP_LEVEL1_RADIX);
345 l0size = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
346 l1size = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
349 chain = hammer2_chain_lookup(parentp, &key_dummy, key, key + l1mask,
351 HAMMER2_LOOKUP_ALWAYS |
352 HAMMER2_LOOKUP_MATCHIND);
356 * Create the missing leaf, be sure to initialize
357 * the auxillary freemap tracking information in
358 * the bref.check.freemap structure.
361 kprintf("freemap create L1 @ %016jx bpref %016jx\n",
364 error = hammer2_chain_create(parentp, &chain, NULL, hmp->spmp,
365 HAMMER2_METH_DEFAULT,
366 key, HAMMER2_FREEMAP_LEVEL1_RADIX,
367 HAMMER2_BREF_TYPE_FREEMAP_LEAF,
368 HAMMER2_FREEMAP_LEVELN_PSIZE,
370 KKASSERT(error == 0);
372 hammer2_chain_modify(chain, mtid, 0, 0);
373 bzero(&chain->data->bmdata[0],
374 HAMMER2_FREEMAP_LEVELN_PSIZE);
375 chain->bref.check.freemap.bigmask = (uint32_t)-1;
376 chain->bref.check.freemap.avail = l1size;
377 /* bref.methods should already be inherited */
379 hammer2_freemap_init(hmp, key, chain);
381 } else if (chain->error) {
383 * Error during lookup.
385 kprintf("hammer2_freemap_try_alloc: %016jx: error %s\n",
386 (intmax_t)bref->data_off,
387 hammer2_error_str(chain->error));
388 error = HAMMER2_ERROR_EIO;
389 } else if ((chain->bref.check.freemap.bigmask &
390 ((size_t)1 << radix)) == 0) {
392 * Already flagged as not having enough space
394 error = HAMMER2_ERROR_ENOSPC;
397 * Modify existing chain to setup for adjustment.
399 hammer2_chain_modify(chain, mtid, 0, 0);
406 hammer2_bmap_data_t *bmap;
407 hammer2_key_t base_key;
412 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_LEAF);
413 start = (int)((iter->bnext - key) >>
414 HAMMER2_FREEMAP_LEVEL0_RADIX);
415 KKASSERT(start >= 0 && start < HAMMER2_FREEMAP_COUNT);
416 hammer2_chain_modify(chain, mtid, 0, 0);
418 error = HAMMER2_ERROR_ENOSPC;
419 for (count = 0; count < HAMMER2_FREEMAP_COUNT; ++count) {
422 if (start + count >= HAMMER2_FREEMAP_COUNT &&
428 * Calculate bmap pointer from thart starting index
431 * NOTE: bmap pointer is invalid if n >= FREEMAP_COUNT.
434 bmap = &chain->data->bmdata[n];
436 if (n >= HAMMER2_FREEMAP_COUNT) {
438 } else if (bmap->avail) {
440 } else if (radix < HAMMER2_FREEMAP_BLOCK_RADIX &&
441 (bmap->linear & HAMMER2_FREEMAP_BLOCK_MASK)) {
448 * Try to allocate from a matching freemap class
449 * superblock. If we are in relaxed mode we allocate
450 * from any freemap class superblock.
453 (bmap->class == 0 || bmap->class == class ||
455 base_key = key + n * l0size;
456 error = hammer2_bmap_alloc(hmp, bmap,
461 if (error != HAMMER2_ERROR_ENOSPC) {
468 * Calculate bmap pointer from thart starting index
469 * backwards (locality).
471 * Must recalculate after potentially having called
472 * hammer2_bmap_alloc() above in case chain was
475 * NOTE: bmap pointer is invalid if n < 0.
478 bmap = &chain->data->bmdata[n];
481 } else if (bmap->avail) {
483 } else if (radix < HAMMER2_FREEMAP_BLOCK_RADIX &&
484 (bmap->linear & HAMMER2_FREEMAP_BLOCK_MASK)) {
491 * Try to allocate from a matching freemap class
492 * superblock. If we are in relaxed mode we allocate
493 * from any freemap class superblock.
496 (bmap->class == 0 || bmap->class == class ||
498 base_key = key + n * l0size;
499 error = hammer2_bmap_alloc(hmp, bmap,
504 if (error != HAMMER2_ERROR_ENOSPC) {
512 * We only know for sure that we can clear the bitmap bit
513 * if we scanned the entire array (start == 0).
515 if (error == HAMMER2_ERROR_ENOSPC && start == 0) {
516 chain->bref.check.freemap.bigmask &=
517 (uint32_t)~((size_t)1 << radix);
519 /* XXX also scan down from original count */
524 * Assert validity. Must be beyond the static allocator used
525 * by newfs_hammer2 (and thus also beyond the aux area),
526 * not go past the volume size, and must not be in the
527 * reserved segment area for a zone.
529 KKASSERT(key >= hmp->voldata.allocator_beg &&
530 key + bytes <= hmp->voldata.volu_size);
531 KKASSERT((key & HAMMER2_ZONE_MASK64) >= HAMMER2_ZONE_SEG);
532 bref->data_off = key | radix;
535 * Record dedupability. The dedup bits are cleared
536 * when bulkfree transitions the freemap from 11->10,
537 * and asserted to be clear on the 10->00 transition.
539 * We must record the bitmask with the chain locked
540 * at the time we set the allocation bits to avoid
543 if (bref->type == HAMMER2_BREF_TYPE_DATA)
544 hammer2_io_dedup_set(hmp, bref);
546 kprintf("alloc cp=%p %016jx %016jx using %016jx\n",
548 bref->key, bref->data_off, chain->bref.data_off);
550 } else if (error == HAMMER2_ERROR_ENOSPC) {
552 * Return EAGAIN with next iteration in iter->bnext, or
553 * return ENOSPC if the allocation map has been exhausted.
555 error = hammer2_freemap_iterate(parentp, &chain, iter);
562 hammer2_chain_unlock(chain);
563 hammer2_chain_drop(chain);
569 * Allocate (1<<radix) bytes from the bmap whos base data offset is (*basep).
571 * If the linear iterator is mid-block we use it directly (the bitmap should
572 * already be marked allocated), otherwise we search for a block in the
573 * bitmap that fits the allocation request.
575 * A partial bitmap allocation sets the minimum bitmap granularity (16KB)
576 * to fully allocated and adjusts the linear allocator to allow the
577 * remaining space to be allocated.
579 * sub_key is the lower 32 bits of the chain->bref.key for the chain whos
580 * bref is being allocated. If the radix represents an allocation >= 16KB
581 * (aka HAMMER2_FREEMAP_BLOCK_RADIX) we try to use this key to select the
582 * blocks directly out of the bmap.
586 hammer2_bmap_alloc(hammer2_dev_t *hmp, hammer2_bmap_data_t *bmap,
587 uint16_t class, int n, int sub_key,
588 int radix, hammer2_key_t *basep)
593 hammer2_bitmap_t bmmask;
599 * Take into account 2-bits per block when calculating bmradix.
601 size = (size_t)1 << radix;
603 if (radix <= HAMMER2_FREEMAP_BLOCK_RADIX) {
605 /* (16K) 2 bits per allocation block */
607 bmradix = (hammer2_bitmap_t)2 <<
608 (radix - HAMMER2_FREEMAP_BLOCK_RADIX);
609 /* (32K-256K) 4, 8, 16, 32 bits per allocation block */
613 * Use the linear iterator to pack small allocations, otherwise
614 * fall-back to finding a free 16KB chunk. The linear iterator
615 * is only valid when *NOT* on a freemap chunking boundary (16KB).
616 * If it is the bitmap must be scanned. It can become invalid
617 * once we pack to the boundary. We adjust it after a bitmap
618 * allocation only for sub-16KB allocations (so the perfectly good
619 * previous value can still be used for fragments when 16KB+
620 * allocations are made inbetween fragmentary allocations).
622 * Beware of hardware artifacts when bmradix == 64 (intermediate
623 * result can wind up being '1' instead of '0' if hardware masks
626 * NOTE: j needs to be even in the j= calculation. As an artifact
627 * of the /2 division, our bitmask has to clear bit 0.
629 * NOTE: TODO this can leave little unallocatable fragments lying
632 if (((uint32_t)bmap->linear & HAMMER2_FREEMAP_BLOCK_MASK) + size <=
633 HAMMER2_FREEMAP_BLOCK_SIZE &&
634 (bmap->linear & HAMMER2_FREEMAP_BLOCK_MASK) &&
635 bmap->linear < HAMMER2_SEGSIZE) {
637 * Use linear iterator if it is not block-aligned to avoid
640 * Calculate the bitmapq[] index (i) and calculate the
641 * shift count within the 64-bit bitmapq[] entry.
643 * The freemap block size is 16KB, but each bitmap
644 * entry is two bits so use a little trick to get
645 * a (j) shift of 0, 2, 4, ... 62 in 16KB chunks.
647 KKASSERT(bmap->linear >= 0 &&
648 bmap->linear + size <= HAMMER2_SEGSIZE &&
649 (bmap->linear & (HAMMER2_ALLOC_MIN - 1)) == 0);
650 offset = bmap->linear;
651 i = offset / (HAMMER2_SEGSIZE / HAMMER2_BMAP_ELEMENTS);
652 j = (offset / (HAMMER2_FREEMAP_BLOCK_SIZE / 2)) & 62;
653 bmmask = (bmradix == HAMMER2_BMAP_BITS_PER_ELEMENT) ?
654 HAMMER2_BMAP_ALLONES :
655 ((hammer2_bitmap_t)1 << bmradix) - 1;
657 bmap->linear = offset + size;
660 * Try to index a starting point based on sub_key. This
661 * attempts to restore sequential block ordering on-disk
662 * whenever possible, even if data is committed out of
665 * i - Index bitmapq[], full data range represented is
668 * j - Index within bitmapq[i], full data range represented is
669 * HAMMER2_BMAP_INDEX_SIZE.
677 case HAMMER2_BREF_TYPE_DATA:
678 if (radix >= HAMMER2_FREEMAP_BLOCK_RADIX) {
679 i = (sub_key & HAMMER2_BMAP_MASK) /
680 (HAMMER2_BMAP_SIZE / HAMMER2_BMAP_ELEMENTS);
681 j = (sub_key & HAMMER2_BMAP_INDEX_MASK) /
682 (HAMMER2_BMAP_INDEX_SIZE /
683 HAMMER2_BMAP_BLOCKS_PER_ELEMENT);
687 case HAMMER2_BREF_TYPE_INODE:
693 KKASSERT(i < HAMMER2_BMAP_ELEMENTS &&
694 j < 2 * HAMMER2_BMAP_BLOCKS_PER_ELEMENT);
695 KKASSERT(j + bmradix <= HAMMER2_BMAP_BITS_PER_ELEMENT);
696 bmmask = (bmradix == HAMMER2_BMAP_BITS_PER_ELEMENT) ?
697 HAMMER2_BMAP_ALLONES :
698 ((hammer2_bitmap_t)1 << bmradix) - 1;
701 if ((bmap->bitmapq[i] & bmmask) == 0)
706 * General element scan.
708 * WARNING: (j) is iterating a bit index (by 2's)
710 for (i = 0; i < HAMMER2_BMAP_ELEMENTS; ++i) {
711 bmmask = (bmradix == HAMMER2_BMAP_BITS_PER_ELEMENT) ?
712 HAMMER2_BMAP_ALLONES :
713 ((hammer2_bitmap_t)1 << bmradix) - 1;
715 j < HAMMER2_BMAP_BITS_PER_ELEMENT;
717 if ((bmap->bitmapq[i] & bmmask) == 0)
722 /*fragments might remain*/
723 /*KKASSERT(bmap->avail == 0);*/
724 return (HAMMER2_ERROR_ENOSPC);
726 offset = i * (HAMMER2_SEGSIZE / HAMMER2_BMAP_ELEMENTS) +
727 (j * (HAMMER2_FREEMAP_BLOCK_SIZE / 2));
728 if (size & HAMMER2_FREEMAP_BLOCK_MASK)
729 bmap->linear = offset + size;
732 /* 8 x (64/2) -> 256 x 16K -> 4MB */
733 KKASSERT(i >= 0 && i < HAMMER2_BMAP_ELEMENTS);
736 * Optimize the buffer cache to avoid unnecessary read-before-write
739 * The device block size could be larger than the allocation size
740 * so the actual bitmap test is somewhat more involved. We have
741 * to use a compatible buffer size for this operation.
743 if ((bmap->bitmapq[i] & bmmask) == 0 &&
744 hammer2_devblksize(size) != size) {
745 size_t psize = hammer2_devblksize(size);
746 hammer2_off_t pmask = (hammer2_off_t)psize - 1;
747 int pbmradix = (hammer2_bitmap_t)2 <<
748 (hammer2_devblkradix(radix) -
749 HAMMER2_FREEMAP_BLOCK_RADIX);
750 hammer2_bitmap_t pbmmask;
751 int pradix = hammer2_getradix(psize);
753 pbmmask = (pbmradix == HAMMER2_BMAP_BITS_PER_ELEMENT) ?
754 HAMMER2_BMAP_ALLONES :
755 ((hammer2_bitmap_t)1 << pbmradix) - 1;
756 while ((pbmmask & bmmask) == 0)
757 pbmmask <<= pbmradix;
760 kprintf("%016jx mask %016jx %016jx %016jx (%zd/%zd)\n",
761 *basep + offset, bmap->bitmapq[i],
762 pbmmask, bmmask, size, psize);
765 if ((bmap->bitmapq[i] & pbmmask) == 0) {
768 hammer2_io_newnz(hmp, class >> 8,
769 (*basep + (offset & ~pmask)) |
770 pradix, psize, &dio);
771 hammer2_io_putblk(&dio);
777 * When initializing a new inode segment also attempt to initialize
778 * an adjacent segment. Be careful not to index beyond the array
781 * We do this to try to localize inode accesses to improve
782 * directory scan rates. XXX doesn't improve scan rates.
784 if (size == HAMMER2_INODE_BYTES) {
786 if (bmap[-1].radix == 0 && bmap[-1].avail)
787 bmap[-1].radix = radix;
789 if (bmap[1].radix == 0 && bmap[1].avail)
790 bmap[1].radix = radix;
795 * Calculate the bitmap-granular change in bgsize for the volume
796 * header. We cannot use the fine-grained change here because
797 * the bulkfree code can't undo it. If the bitmap element is already
798 * marked allocated it has already been accounted for.
800 if (radix < HAMMER2_FREEMAP_BLOCK_RADIX) {
801 if (bmap->bitmapq[i] & bmmask)
804 bgsize = HAMMER2_FREEMAP_BLOCK_SIZE;
810 * Adjust the bitmap, set the class (it might have been 0),
811 * and available bytes, update the allocation offset (*basep)
812 * from the L0 base to the actual offset.
814 * Do not override the class if doing a relaxed class allocation.
816 * avail must reflect the bitmap-granular availability. The allocator
817 * tests will also check the linear iterator.
819 bmap->bitmapq[i] |= bmmask;
820 if (bmap->class == 0)
822 bmap->avail -= bgsize;
826 * Adjust the volume header's allocator_free parameter. This
827 * parameter has to be fixed up by bulkfree which has no way to
828 * figure out sub-16K chunking, so it must be adjusted by the
829 * bitmap-granular size.
832 hammer2_voldata_lock(hmp);
833 hammer2_voldata_modify(hmp);
834 hmp->voldata.allocator_free -= bgsize;
835 hammer2_voldata_unlock(hmp);
842 * Initialize a freemap for the storage area (in bytes) that begins at (key).
846 hammer2_freemap_init(hammer2_dev_t *hmp, hammer2_key_t key,
847 hammer2_chain_t *chain)
849 hammer2_off_t l1size;
852 hammer2_bmap_data_t *bmap;
856 * LEVEL1 is 1GB, there are two level1 1GB freemaps per 2GB zone.
858 l1size = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
861 * Calculate the portion of the 1GB map that should be initialized
862 * as free. Portions below or after will be initialized as allocated.
863 * SEGMASK-align the areas so we don't have to worry about sub-scans
864 * or endianess when using memset.
866 * WARNING! It is possible for lokey to be larger than hikey if the
867 * entire 2GB segment is within the static allocation.
870 * (1) Ensure that all statically allocated space from newfs_hammer2
871 * is marked allocated, and take it up to the level1 base for
874 lokey = (hmp->voldata.allocator_beg + HAMMER2_SEGMASK64) &
876 if (lokey < H2FMBASE(key, HAMMER2_FREEMAP_LEVEL1_RADIX))
877 lokey = H2FMBASE(key, HAMMER2_FREEMAP_LEVEL1_RADIX);
880 * (2) Ensure that the reserved area is marked allocated (typically
881 * the first 4MB of each 2GB area being represented). Since
882 * each LEAF represents 1GB of storage and the zone is 2GB, we
883 * have to adjust lowkey upward every other LEAF sequentially.
885 if (lokey < H2FMZONEBASE(key) + HAMMER2_ZONE_SEG64)
886 lokey = H2FMZONEBASE(key) + HAMMER2_ZONE_SEG64;
889 * (3) Ensure that any trailing space at the end-of-volume is marked
892 hikey = key + H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
893 if (hikey > hmp->voldata.volu_size) {
894 hikey = hmp->voldata.volu_size & ~HAMMER2_SEGMASK64;
898 * Heuristic highest possible value
900 chain->bref.check.freemap.avail =
901 H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
902 bmap = &chain->data->bmdata[0];
905 * Initialize bitmap (bzero'd by caller)
907 for (count = 0; count < HAMMER2_FREEMAP_COUNT; ++count) {
908 if (key < lokey || key >= hikey) {
909 memset(bmap->bitmapq, -1,
910 sizeof(bmap->bitmapq));
912 bmap->linear = HAMMER2_SEGSIZE;
913 chain->bref.check.freemap.avail -=
914 H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
916 bmap->avail = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
918 key += H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
924 * The current Level 1 freemap has been exhausted, iterate to the next
925 * one, return ENOSPC if no freemaps remain.
927 * At least two loops are required. If we are not in relaxed mode and
928 * we run out of storage we enter relaxed mode and do a third loop.
929 * The relaxed mode is recorded back in the hmp so once we enter the mode
930 * we remain relaxed until stuff begins to get freed and only do 2 loops.
932 * XXX this should rotate back to the beginning to handle freed-up space
933 * XXX or use intermediate entries to locate free space. TODO
936 hammer2_freemap_iterate(hammer2_chain_t **parentp, hammer2_chain_t **chainp,
937 hammer2_fiterate_t *iter)
939 hammer2_dev_t *hmp = (*parentp)->hmp;
941 iter->bnext &= ~(H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX) - 1);
942 iter->bnext += H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
943 if (iter->bnext >= hmp->voldata.volu_size) {
945 if (++iter->loops >= 2) {
946 if (iter->relaxed == 0)
949 return (HAMMER2_ERROR_ENOSPC);
952 return(HAMMER2_ERROR_EAGAIN);
956 * Adjust the bit-pattern for data in the freemap bitmap according to
957 * (how). This code is called from on-mount recovery to fixup (mark
958 * as allocated) blocks whos freemap upates might not have been committed
959 * in the last crash and is used by the bulk freemap scan to stage frees.
961 * WARNING! Cannot be called with a empty-data bref (radix == 0).
963 * XXX currently disabled when how == 0 (the normal real-time case). At
964 * the moment we depend on the bulk freescan to actually free blocks. It
965 * will still call this routine with a non-zero how to stage possible frees
966 * and to do the actual free.
969 hammer2_freemap_adjust(hammer2_dev_t *hmp, hammer2_blockref_t *bref,
972 hammer2_off_t data_off = bref->data_off;
973 hammer2_chain_t *chain;
974 hammer2_chain_t *parent;
975 hammer2_bmap_data_t *bmap;
977 hammer2_key_t key_dummy;
978 hammer2_off_t l0size;
979 hammer2_off_t l1size;
980 hammer2_off_t l1mask;
982 hammer2_bitmap_t *bitmap;
983 const hammer2_bitmap_t bmmask00 = 0;
984 hammer2_bitmap_t bmmask01;
985 hammer2_bitmap_t bmmask10;
986 hammer2_bitmap_t bmmask11;
996 KKASSERT(how == HAMMER2_FREEMAP_DORECOVER);
999 mtid = hammer2_trans_sub(hmp->spmp);
1001 radix = (int)data_off & HAMMER2_OFF_MASK_RADIX;
1002 KKASSERT(radix != 0);
1003 data_off &= ~HAMMER2_OFF_MASK_RADIX;
1004 KKASSERT(radix <= HAMMER2_RADIX_MAX);
1007 bytes = (size_t)1 << radix;
1010 class = (bref->type << 8) | hammer2_devblkradix(radix);
1013 * We can't adjust the freemap for data allocations made by
1016 if (data_off < hmp->voldata.allocator_beg)
1019 KKASSERT((data_off & HAMMER2_ZONE_MASK64) >= HAMMER2_ZONE_SEG);
1022 * Lookup the level1 freemap chain. The chain must exist.
1024 key = H2FMBASE(data_off, HAMMER2_FREEMAP_LEVEL1_RADIX);
1025 l0size = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
1026 l1size = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
1027 l1mask = l1size - 1;
1029 parent = &hmp->fchain;
1030 hammer2_chain_ref(parent);
1031 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1033 chain = hammer2_chain_lookup(&parent, &key_dummy, key, key + l1mask,
1035 HAMMER2_LOOKUP_ALWAYS |
1036 HAMMER2_LOOKUP_MATCHIND);
1039 * Stop early if we are trying to free something but no leaf exists.
1041 if (chain == NULL && how != HAMMER2_FREEMAP_DORECOVER) {
1042 kprintf("hammer2_freemap_adjust: %016jx: no chain\n",
1043 (intmax_t)bref->data_off);
1047 kprintf("hammer2_freemap_adjust: %016jx: error %s\n",
1048 (intmax_t)bref->data_off,
1049 hammer2_error_str(chain->error));
1050 hammer2_chain_unlock(chain);
1051 hammer2_chain_drop(chain);
1057 * Create any missing leaf(s) if we are doing a recovery (marking
1058 * the block(s) as being allocated instead of being freed). Be sure
1059 * to initialize the auxillary freemap tracking info in the
1060 * bref.check.freemap structure.
1062 if (chain == NULL && how == HAMMER2_FREEMAP_DORECOVER) {
1063 error = hammer2_chain_create(&parent, &chain, NULL, hmp->spmp,
1064 HAMMER2_METH_DEFAULT,
1065 key, HAMMER2_FREEMAP_LEVEL1_RADIX,
1066 HAMMER2_BREF_TYPE_FREEMAP_LEAF,
1067 HAMMER2_FREEMAP_LEVELN_PSIZE,
1070 if (hammer2_debug & 0x0040) {
1071 kprintf("fixup create chain %p %016jx:%d\n",
1072 chain, chain->bref.key, chain->bref.keybits);
1076 error = hammer2_chain_modify(chain, mtid, 0, 0);
1077 KKASSERT(error == 0);
1078 bzero(&chain->data->bmdata[0],
1079 HAMMER2_FREEMAP_LEVELN_PSIZE);
1080 chain->bref.check.freemap.bigmask = (uint32_t)-1;
1081 chain->bref.check.freemap.avail = l1size;
1082 /* bref.methods should already be inherited */
1084 hammer2_freemap_init(hmp, key, chain);
1086 /* XXX handle error */
1090 kprintf("FREEMAP ADJUST TYPE %d %016jx/%d DATA_OFF=%016jx\n",
1091 chain->bref.type, chain->bref.key,
1092 chain->bref.keybits, chain->bref.data_off);
1096 * Calculate the bitmask (runs in 2-bit pairs).
1098 start = ((int)(data_off >> HAMMER2_FREEMAP_BLOCK_RADIX) & 15) * 2;
1099 bmmask01 = (hammer2_bitmap_t)1 << start;
1100 bmmask10 = (hammer2_bitmap_t)2 << start;
1101 bmmask11 = (hammer2_bitmap_t)3 << start;
1104 * Fixup the bitmap. Partial blocks cannot be fully freed unless
1105 * a bulk scan is able to roll them up.
1107 if (radix < HAMMER2_FREEMAP_BLOCK_RADIX) {
1109 if (how == HAMMER2_FREEMAP_DOREALFREE)
1110 how = HAMMER2_FREEMAP_DOMAYFREE;
1112 count = 1 << (radix - HAMMER2_FREEMAP_BLOCK_RADIX);
1116 * [re]load the bmap and bitmap pointers. Each bmap entry covers
1117 * a 4MB swath. The bmap itself (LEVEL1) covers 2GB.
1119 * Be sure to reset the linear iterator to ensure that the adjustment
1123 bmap = &chain->data->bmdata[(int)(data_off >> HAMMER2_SEGRADIX) &
1124 (HAMMER2_FREEMAP_COUNT - 1)];
1125 bitmap = &bmap->bitmapq[(int)(data_off >> (HAMMER2_SEGRADIX - 3)) & 7];
1132 if (how == HAMMER2_FREEMAP_DORECOVER) {
1134 * Recovery request, mark as allocated.
1136 if ((*bitmap & bmmask11) != bmmask11) {
1137 if (modified == 0) {
1138 hammer2_chain_modify(chain, mtid, 0, 0);
1142 if ((*bitmap & bmmask11) == bmmask00) {
1144 HAMMER2_FREEMAP_BLOCK_SIZE;
1145 bgsize += HAMMER2_FREEMAP_BLOCK_SIZE;
1147 if (bmap->class == 0)
1148 bmap->class = class;
1149 *bitmap |= bmmask11;
1150 if (hammer2_debug & 0x0040) {
1151 kprintf("hammer2_freemap_recover: "
1153 "block=%016jx/%zd\n",
1154 bref->type, data_off, bytes);
1158 kprintf("hammer2_freemap_recover: good "
1159 "type=%02x block=%016jx/%zd\n",
1160 bref->type, data_off, bytes);
1166 * XXX this stuff doesn't work, avail is miscalculated and
1167 * code 10 means something else now.
1169 else if ((*bitmap & bmmask11) == bmmask11) {
1171 * Mayfree/Realfree request and bitmap is currently
1172 * marked as being fully allocated.
1175 hammer2_chain_modify(chain, 0);
1179 if (how == HAMMER2_FREEMAP_DOREALFREE)
1180 *bitmap &= ~bmmask11;
1182 *bitmap = (*bitmap & ~bmmask11) | bmmask10;
1183 } else if ((*bitmap & bmmask11) == bmmask10) {
1185 * Mayfree/Realfree request and bitmap is currently
1186 * marked as being possibly freeable.
1188 if (how == HAMMER2_FREEMAP_DOREALFREE) {
1190 hammer2_chain_modify(chain, 0);
1194 *bitmap &= ~bmmask11;
1198 * 01 - Not implemented, currently illegal state
1199 * 00 - Not allocated at all, illegal free.
1201 panic("hammer2_freemap_adjust: "
1202 "Illegal state %08x(%08x)",
1203 *bitmap, *bitmap & bmmask11);
1211 #if HAMMER2_BMAP_ELEMENTS != 8
1212 #error "hammer2_freemap.c: HAMMER2_BMAP_ELEMENTS expected to be 8"
1214 if (how == HAMMER2_FREEMAP_DOREALFREE && modified) {
1215 bmap->avail += 1 << radix;
1216 KKASSERT(bmap->avail <= HAMMER2_SEGSIZE);
1217 if (bmap->avail == HAMMER2_SEGSIZE &&
1218 bmap->bitmapq[0] == 0 &&
1219 bmap->bitmapq[1] == 0 &&
1220 bmap->bitmapq[2] == 0 &&
1221 bmap->bitmapq[3] == 0 &&
1222 bmap->bitmapq[4] == 0 &&
1223 bmap->bitmapq[5] == 0 &&
1224 bmap->bitmapq[6] == 0 &&
1225 bmap->bitmapq[7] == 0) {
1226 key = H2FMBASE(data_off, HAMMER2_FREEMAP_LEVEL0_RADIX);
1227 kprintf("Freeseg %016jx\n", (intmax_t)key);
1233 * chain->bref.check.freemap.bigmask (XXX)
1235 * Setting bigmask is a hint to the allocation code that there might
1236 * be something allocatable. We also set this in recovery... it
1237 * doesn't hurt and we might want to use the hint for other validation
1238 * operations later on.
1240 * We could calculate the largest possible allocation and set the
1241 * radii that could fit, but its easier just to set bigmask to -1.
1244 chain->bref.check.freemap.bigmask = -1;
1245 hmp->freemap_relaxed = 0; /* reset heuristic */
1248 hammer2_chain_unlock(chain);
1249 hammer2_chain_drop(chain);
1251 hammer2_chain_unlock(parent);
1252 hammer2_chain_drop(parent);
1255 hammer2_voldata_lock(hmp);
1256 hammer2_voldata_modify(hmp);
1257 hmp->voldata.allocator_free -= bgsize;
1258 hammer2_voldata_unlock(hmp);
1263 * Validate the freemap, in three stages.
1265 * stage-1 ALLOCATED -> POSSIBLY FREE
1266 * POSSIBLY FREE -> POSSIBLY FREE (type corrected)
1268 * This transitions bitmap entries from ALLOCATED to POSSIBLY FREE.
1269 * The POSSIBLY FREE state does not mean that a block is actually free
1270 * and may be transitioned back to ALLOCATED in stage-2.
1272 * This is typically done during normal filesystem operations when
1273 * something is deleted or a block is replaced.
1275 * This is done by bulkfree in-bulk after a memory-bounded meta-data
1276 * scan to try to determine what might be freeable.
1278 * This can be done unconditionally through a freemap scan when the
1279 * intention is to brute-force recover the proper state of the freemap.
1281 * stage-2 POSSIBLY FREE -> ALLOCATED (scan metadata topology)
1283 * This is done by bulkfree during a meta-data scan to ensure that
1284 * all blocks still actually allocated by the filesystem are marked
1287 * NOTE! Live filesystem transitions to POSSIBLY FREE can occur while
1288 * the bulkfree stage-2 and stage-3 is running. The live filesystem
1289 * will use the alternative POSSIBLY FREE type (2) to prevent
1290 * stage-3 from improperly transitioning unvetted possibly-free
1293 * stage-3 POSSIBLY FREE (type 1) -> FREE (scan freemap)
1295 * This is done by bulkfree to finalize POSSIBLY FREE states.