2 * Copyright (c) 2011-2015 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@dragonflybsd.org>
6 * by Venkatesh Srinivas <vsrinivas@dragonflybsd.org>
7 * by Daniel Flores (GSOC 2013 - mentored by Matthew Dillon, compression)
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in
17 * the documentation and/or other materials provided with the
19 * 3. Neither the name of The DragonFly Project nor the names of its
20 * contributors may be used to endorse or promote products derived
21 * from this software without specific, prior written permission.
23 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
24 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
25 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
26 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
27 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
28 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
29 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
30 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
31 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
32 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
33 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
37 * This module handles low level logical file I/O (strategy) which backs
38 * the logical buffer cache.
40 * [De]compression, zero-block, check codes, and buffer cache operations
41 * for file data is handled here.
43 * Live dedup makes its home here as well.
46 #include <sys/param.h>
47 #include <sys/systm.h>
48 #include <sys/kernel.h>
49 #include <sys/fcntl.h>
52 #include <sys/namei.h>
53 #include <sys/mount.h>
54 #include <sys/vnode.h>
55 #include <sys/mountctl.h>
56 #include <sys/dirent.h>
58 #include <sys/objcache.h>
59 #include <sys/event.h>
61 #include <vfs/fifofs/fifo.h>
64 #include "hammer2_lz4.h"
66 #include "zlib/hammer2_zlib.h"
68 struct objcache *cache_buffer_read;
69 struct objcache *cache_buffer_write;
72 * Strategy code (async logical file buffer I/O from system)
74 * Except for the transaction init (which should normally not block),
75 * we essentially run the strategy operation asynchronously via a XOP.
77 * XXX This isn't supposed to be able to deadlock against vfs_sync vfsync()
78 * calls but it has in the past when multiple flushes are queued.
80 * XXX We currently terminate the transaction once we get a quorum, otherwise
81 * the frontend can stall, but this can leave the remaining nodes with
82 * a potential flush conflict. We need to delay flushes on those nodes
83 * until running transactions complete separately from the normal
84 * transaction sequencing. FIXME TODO.
86 static void hammer2_strategy_xop_read(hammer2_thread_t *thr,
88 static void hammer2_strategy_xop_write(hammer2_thread_t *thr,
90 static int hammer2_strategy_read(struct vop_strategy_args *ap);
91 static int hammer2_strategy_write(struct vop_strategy_args *ap);
92 static void hammer2_strategy_read_completion(hammer2_chain_t *chain,
93 char *data, struct bio *bio);
95 static hammer2_off_t hammer2_dedup_lookup(hammer2_dev_t *hmp,
96 char **datap, int pblksize);
105 #define TIMER(which) do { \
107 h2timer[h2lid] += (int)(ticks - h2last);\
114 hammer2_vop_strategy(struct vop_strategy_args *ap)
125 error = hammer2_strategy_read(ap);
126 ++hammer2_iod_file_read;
129 error = hammer2_strategy_write(ap);
130 ++hammer2_iod_file_write;
133 bp->b_error = error = EINVAL;
134 bp->b_flags |= B_ERROR;
142 * Return the largest contiguous physical disk range for the logical
145 * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
147 * Basically disabled, the logical buffer write thread has to deal with
148 * buffers one-at-a-time. Note that this should not prevent cluster_read()
149 * from reading-ahead, it simply prevents it from trying form a single
150 * cluster buffer for the logical request. H2 already uses 64KB buffers!
153 hammer2_vop_bmap(struct vop_bmap_args *ap)
155 *ap->a_doffsetp = NOOFFSET;
163 /****************************************************************************
165 ****************************************************************************/
167 * Callback used in read path in case that a block is compressed with LZ4.
171 hammer2_decompress_LZ4_callback(const char *data, u_int bytes, struct bio *bio)
174 char *compressed_buffer;
181 if bio->bio_caller_info2.index &&
182 bio->bio_caller_info1.uvalue32 !=
183 crc32(bp->b_data, bp->b_bufsize) --- return error
186 KKASSERT(bp->b_bufsize <= HAMMER2_PBUFSIZE);
187 compressed_size = *(const int *)data;
188 KKASSERT((uint32_t)compressed_size <= bytes - sizeof(int));
190 compressed_buffer = objcache_get(cache_buffer_read, M_INTWAIT);
191 result = LZ4_decompress_safe(__DECONST(char *, &data[sizeof(int)]),
196 kprintf("READ PATH: Error during decompression."
198 (intmax_t)bio->bio_offset, bytes);
199 /* make sure it isn't random garbage */
200 bzero(compressed_buffer, bp->b_bufsize);
202 KKASSERT(result <= bp->b_bufsize);
203 bcopy(compressed_buffer, bp->b_data, bp->b_bufsize);
204 if (result < bp->b_bufsize)
205 bzero(bp->b_data + result, bp->b_bufsize - result);
206 objcache_put(cache_buffer_read, compressed_buffer);
208 bp->b_flags |= B_AGE;
212 * Callback used in read path in case that a block is compressed with ZLIB.
213 * It is almost identical to LZ4 callback, so in theory they can be unified,
214 * but we didn't want to make changes in bio structure for that.
218 hammer2_decompress_ZLIB_callback(const char *data, u_int bytes, struct bio *bio)
221 char *compressed_buffer;
222 z_stream strm_decompress;
228 KKASSERT(bp->b_bufsize <= HAMMER2_PBUFSIZE);
229 strm_decompress.avail_in = 0;
230 strm_decompress.next_in = Z_NULL;
232 ret = inflateInit(&strm_decompress);
235 kprintf("HAMMER2 ZLIB: Fatal error in inflateInit.\n");
237 compressed_buffer = objcache_get(cache_buffer_read, M_INTWAIT);
238 strm_decompress.next_in = __DECONST(char *, data);
240 /* XXX supply proper size, subset of device bp */
241 strm_decompress.avail_in = bytes;
242 strm_decompress.next_out = compressed_buffer;
243 strm_decompress.avail_out = bp->b_bufsize;
245 ret = inflate(&strm_decompress, Z_FINISH);
246 if (ret != Z_STREAM_END) {
247 kprintf("HAMMER2 ZLIB: Fatar error during decompression.\n");
248 bzero(compressed_buffer, bp->b_bufsize);
250 bcopy(compressed_buffer, bp->b_data, bp->b_bufsize);
251 result = bp->b_bufsize - strm_decompress.avail_out;
252 if (result < bp->b_bufsize)
253 bzero(bp->b_data + result, strm_decompress.avail_out);
254 objcache_put(cache_buffer_read, compressed_buffer);
255 ret = inflateEnd(&strm_decompress);
258 bp->b_flags |= B_AGE;
262 * Logical buffer I/O, async read.
266 hammer2_strategy_read(struct vop_strategy_args *ap)
268 hammer2_xop_strategy_t *xop;
278 nbio = push_bio(bio);
280 lbase = bio->bio_offset;
281 KKASSERT(((int)lbase & HAMMER2_PBUFMASK) == 0);
283 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_STRATEGY);
287 hammer2_mtx_init(&xop->lock, "h2bior");
288 hammer2_xop_start(&xop->head, hammer2_strategy_xop_read);
289 /* asynchronous completion */
295 * Per-node XOP (threaded), do a synchronous lookup of the chain and
296 * its data. The frontend is asynchronous, so we are also responsible
297 * for racing to terminate the frontend.
301 hammer2_strategy_xop_read(hammer2_thread_t *thr, hammer2_xop_t *arg)
303 hammer2_xop_strategy_t *xop = &arg->xop_strategy;
304 hammer2_chain_t *parent;
305 hammer2_chain_t *chain;
306 hammer2_key_t key_dummy;
310 int cache_index = -1;
314 * Note that we can race completion of the bio supplied by
315 * the front-end so we cannot access it until we determine
316 * that we are the ones finishing it up.
322 * This is difficult to optimize. The logical buffer might be
323 * partially dirty (contain dummy zero-fill pages), which would
324 * mess up our crc calculation if we were to try a direct read.
325 * So for now we always double-buffer through the underlying
328 * If not for the above problem we could conditionalize on
329 * (1) 64KB buffer, (2) one chain (not multi-master) and
330 * (3) !hammer2_double_buffer, and issue a direct read into the
333 parent = hammer2_inode_chain(xop->head.ip1, thr->clindex,
334 HAMMER2_RESOLVE_ALWAYS |
335 HAMMER2_RESOLVE_SHARED);
338 chain = hammer2_chain_lookup(&parent, &key_dummy,
341 HAMMER2_LOOKUP_ALWAYS |
342 HAMMER2_LOOKUP_SHARED);
343 error = chain ? chain->error : 0;
349 error = hammer2_xop_feed(&xop->head, chain, thr->clindex, error);
352 hammer2_chain_unlock(chain);
353 hammer2_chain_drop(chain);
356 hammer2_chain_unlock(parent);
357 hammer2_chain_drop(parent);
359 chain = NULL; /* safety */
360 parent = NULL; /* safety */
364 * Race to finish the frontend. First-to-complete. bio is only
365 * valid if we are determined to be the ones able to complete
370 hammer2_mtx_ex(&xop->lock);
372 hammer2_mtx_unlock(&xop->lock);
379 * Async operation has not completed and we now own the lock.
380 * Determine if we can complete the operation by issuing the
381 * frontend collection non-blocking.
383 * H2 double-buffers the data, setting B_NOTMETA on the logical
384 * buffer hints to the OS that the logical buffer should not be
385 * swapcached (since the device buffer can be).
387 * Also note that even for compressed data we would rather the
388 * kernel cache/swapcache device buffers more and (decompressed)
389 * logical buffers less, since that will significantly improve
390 * the amount of end-user data that can be cached.
392 error = hammer2_xop_collect(&xop->head, HAMMER2_XOP_COLLECT_NOWAIT);
398 hammer2_mtx_unlock(&xop->lock);
399 bp->b_flags |= B_NOTMETA;
400 chain = xop->head.cluster.focus;
401 hammer2_strategy_read_completion(chain, (char *)chain->data,
404 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
408 hammer2_mtx_unlock(&xop->lock);
409 bp->b_flags |= B_NOTMETA;
412 bzero(bp->b_data, bp->b_bcount);
414 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
417 hammer2_mtx_unlock(&xop->lock);
420 kprintf("strategy_xop_read: error %d loff=%016jx\n",
421 error, bp->b_loffset);
423 hammer2_mtx_unlock(&xop->lock);
424 bp->b_flags |= B_ERROR;
427 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
435 hammer2_strategy_read_completion(hammer2_chain_t *chain, char *data,
438 struct buf *bp = bio->bio_buf;
440 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
442 * Copy from in-memory inode structure.
444 bcopy(((hammer2_inode_data_t *)data)->u.data,
445 bp->b_data, HAMMER2_EMBEDDED_BYTES);
446 bzero(bp->b_data + HAMMER2_EMBEDDED_BYTES,
447 bp->b_bcount - HAMMER2_EMBEDDED_BYTES);
450 } else if (chain->bref.type == HAMMER2_BREF_TYPE_DATA) {
452 * Data is on-media, record for live dedup. Release the
453 * chain (try to free it) when done. The data is still
454 * cached by both the buffer cache in front and the
455 * block device behind us. This leaves more room in the
456 * LRU chain cache for meta-data chains which we really
459 * NOTE: Deduplication cannot be safely recorded for
460 * records without a check code.
462 hammer2_dedup_record(chain, NULL, data);
463 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
466 * Decompression and copy.
468 switch (HAMMER2_DEC_COMP(chain->bref.methods)) {
469 case HAMMER2_COMP_LZ4:
470 hammer2_decompress_LZ4_callback(data, chain->bytes,
472 /* b_resid set by call */
474 case HAMMER2_COMP_ZLIB:
475 hammer2_decompress_ZLIB_callback(data, chain->bytes,
477 /* b_resid set by call */
479 case HAMMER2_COMP_NONE:
480 KKASSERT(chain->bytes <= bp->b_bcount);
481 bcopy(data, bp->b_data, chain->bytes);
482 if (chain->bytes < bp->b_bcount) {
483 bzero(bp->b_data + chain->bytes,
484 bp->b_bcount - chain->bytes);
490 panic("hammer2_strategy_read: "
491 "unknown compression type");
494 panic("hammer2_strategy_read: unknown bref type");
498 /****************************************************************************
500 ****************************************************************************/
503 * Functions for compression in threads,
504 * from hammer2_vnops.c
506 static void hammer2_write_file_core(char *data, hammer2_inode_t *ip,
507 hammer2_chain_t **parentp,
508 hammer2_key_t lbase, int ioflag, int pblksize,
509 hammer2_tid_t mtid, int *errorp);
510 static void hammer2_compress_and_write(char *data, hammer2_inode_t *ip,
511 hammer2_chain_t **parentp,
512 hammer2_key_t lbase, int ioflag, int pblksize,
513 hammer2_tid_t mtid, int *errorp,
514 int comp_algo, int check_algo);
515 static void hammer2_zero_check_and_write(char *data, hammer2_inode_t *ip,
516 hammer2_chain_t **parentp,
517 hammer2_key_t lbase, int ioflag, int pblksize,
518 hammer2_tid_t mtid, int *errorp,
520 static int test_block_zeros(const char *buf, size_t bytes);
521 static void zero_write(char *data, hammer2_inode_t *ip,
522 hammer2_chain_t **parentp,
524 hammer2_tid_t mtid, int *errorp);
525 static void hammer2_write_bp(hammer2_chain_t *chain, char *data,
526 int ioflag, int pblksize,
527 hammer2_tid_t mtid, int *errorp,
532 hammer2_strategy_write(struct vop_strategy_args *ap)
534 hammer2_xop_strategy_t *xop;
545 hammer2_lwinprog_ref(pmp);
546 hammer2_trans_assert_strategy(pmp);
547 hammer2_trans_init(pmp, HAMMER2_TRANS_BUFCACHE);
549 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING |
550 HAMMER2_XOP_STRATEGY);
553 xop->lbase = bio->bio_offset;
554 hammer2_mtx_init(&xop->lock, "h2biow");
555 hammer2_xop_start(&xop->head, hammer2_strategy_xop_write);
556 /* asynchronous completion */
558 hammer2_lwinprog_wait(pmp, hammer2_flush_pipe);
564 * Per-node XOP (threaded). Write the logical buffer to the media.
566 * This is a bit problematic because there may be multiple target and
567 * any of them may be able to release the bp. In addition, if our
568 * particulr target is offline we don't want to block the bp (and thus
569 * the frontend). To accomplish this we copy the data to the per-thr
574 hammer2_strategy_xop_write(hammer2_thread_t *thr, hammer2_xop_t *arg)
576 hammer2_xop_strategy_t *xop = &arg->xop_strategy;
577 hammer2_chain_t *parent;
585 hammer2_off_t bio_offset;
589 * We can only access the bp/bio if the frontend has not yet
594 hammer2_mtx_sh(&xop->lock);
596 hammer2_mtx_unlock(&xop->lock);
601 bio = xop->bio; /* ephermal */
602 bp = bio->bio_buf; /* ephermal */
603 ip = xop->head.ip1; /* retained by ref */
604 bio_offset = bio->bio_offset;
605 bio_data = thr->scratch;
607 /* hammer2_trans_init(parent->hmp->spmp, HAMMER2_TRANS_BUFCACHE); */
609 lblksize = hammer2_calc_logical(ip, bio->bio_offset, &lbase, NULL);
610 pblksize = hammer2_calc_physical(ip, lbase);
611 bcopy(bp->b_data, bio_data, lblksize);
613 hammer2_mtx_unlock(&xop->lock);
614 bp = NULL; /* safety, illegal to access after unlock */
615 bio = NULL; /* safety, illegal to access after unlock */
620 parent = hammer2_inode_chain(ip, thr->clindex, HAMMER2_RESOLVE_ALWAYS);
621 hammer2_write_file_core(bio_data, ip, &parent,
622 lbase, IO_ASYNC, pblksize,
623 xop->head.mtid, &error);
625 hammer2_chain_unlock(parent);
626 hammer2_chain_drop(parent);
627 parent = NULL; /* safety */
629 hammer2_xop_feed(&xop->head, NULL, thr->clindex, error);
632 * Try to complete the operation on behalf of the front-end.
636 hammer2_mtx_ex(&xop->lock);
638 hammer2_mtx_unlock(&xop->lock);
643 * Async operation has not completed and we now own the lock.
644 * Determine if we can complete the operation by issuing the
645 * frontend collection non-blocking.
647 * H2 double-buffers the data, setting B_NOTMETA on the logical
648 * buffer hints to the OS that the logical buffer should not be
649 * swapcached (since the device buffer can be).
651 error = hammer2_xop_collect(&xop->head, HAMMER2_XOP_COLLECT_NOWAIT);
653 if (error == EINPROGRESS) {
654 hammer2_mtx_unlock(&xop->lock);
659 * Async operation has completed.
662 hammer2_mtx_unlock(&xop->lock);
664 bio = xop->bio; /* now owned by us */
665 bp = bio->bio_buf; /* now owned by us */
667 if (error == ENOENT || error == 0) {
668 bp->b_flags |= B_NOTMETA;
673 kprintf("strategy_xop_write: error %d loff=%016jx\n",
674 error, bp->b_loffset);
675 bp->b_flags |= B_ERROR;
679 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
680 hammer2_trans_assert_strategy(ip->pmp);
681 hammer2_lwinprog_drop(ip->pmp);
682 hammer2_trans_done(ip->pmp);
686 * Wait for pending I/O to complete
689 hammer2_bioq_sync(hammer2_pfs_t *pmp)
691 hammer2_lwinprog_wait(pmp, 0);
695 * Create a new cluster at (cparent, lbase) and assign physical storage,
696 * returning a cluster suitable for I/O. The cluster will be in a modified
699 * cparent can wind up being anything.
701 * If datap is not NULL, *datap points to the real data we intend to write.
702 * If we can dedup the storage location we set *datap to NULL to indicate
703 * to the caller that a dedup occurred.
705 * NOTE: Special case for data embedded in inode.
709 hammer2_assign_physical(hammer2_inode_t *ip, hammer2_chain_t **parentp,
710 hammer2_key_t lbase, int pblksize,
711 hammer2_tid_t mtid, char **datap, int *errorp)
713 hammer2_chain_t *chain;
714 hammer2_key_t key_dummy;
715 hammer2_off_t dedup_off;
716 int pradix = hammer2_getradix(pblksize);
717 int cache_index = -1;
720 * Locate the chain associated with lbase, return a locked chain.
721 * However, do not instantiate any data reference (which utilizes a
722 * device buffer) because we will be using direct IO via the
723 * logical buffer cache buffer.
726 KKASSERT(pblksize >= HAMMER2_ALLOC_MIN);
729 chain = hammer2_chain_lookup(parentp, &key_dummy,
732 HAMMER2_LOOKUP_NODATA);
735 * The lookup code should not return a DELETED chain to us, unless
736 * its a short-file embedded in the inode. Then it is possible for
737 * the lookup to return a deleted inode.
739 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED) &&
740 chain->bref.type != HAMMER2_BREF_TYPE_INODE) {
741 kprintf("assign physical deleted chain @ "
742 "%016jx (%016jx.%02x) ip %016jx\n",
743 lbase, chain->bref.data_off, chain->bref.type,
750 * We found a hole, create a new chain entry.
752 * NOTE: DATA chains are created without device backing
753 * store (nor do we want any).
755 dedup_off = hammer2_dedup_lookup((*parentp)->hmp, datap,
757 *errorp = hammer2_chain_create(parentp, &chain,
759 HAMMER2_ENC_CHECK(ip->meta.check_algo) |
760 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE),
761 lbase, HAMMER2_PBUFRADIX,
762 HAMMER2_BREF_TYPE_DATA,
766 panic("hammer2_chain_create: par=%p error=%d\n",
770 /*ip->delta_dcount += pblksize;*/
772 switch (chain->bref.type) {
773 case HAMMER2_BREF_TYPE_INODE:
775 * The data is embedded in the inode, which requires
778 hammer2_chain_modify_ip(ip, chain, mtid, 0);
780 case HAMMER2_BREF_TYPE_DATA:
781 dedup_off = hammer2_dedup_lookup(chain->hmp, datap,
783 if (chain->bytes != pblksize) {
784 hammer2_chain_resize(chain,
787 HAMMER2_MODIFY_OPTDATA);
791 * DATA buffers must be marked modified whether the
792 * data is in a logical buffer or not. We also have
793 * to make this call to fixup the chain data pointers
794 * after resizing in case this is an encrypted or
797 hammer2_chain_modify(chain, mtid, dedup_off,
798 HAMMER2_MODIFY_OPTDATA);
801 panic("hammer2_assign_physical: bad type");
811 * hammer2_write_file_core() - hammer2_write_thread() helper
813 * The core write function which determines which path to take
814 * depending on compression settings. We also have to locate the
815 * related chains so we can calculate and set the check data for
820 hammer2_write_file_core(char *data, hammer2_inode_t *ip,
821 hammer2_chain_t **parentp,
822 hammer2_key_t lbase, int ioflag, int pblksize,
823 hammer2_tid_t mtid, int *errorp)
825 hammer2_chain_t *chain;
830 switch(HAMMER2_DEC_ALGO(ip->meta.comp_algo)) {
831 case HAMMER2_COMP_NONE:
833 * We have to assign physical storage to the buffer
834 * we intend to dirty or write now to avoid deadlocks
835 * in the strategy code later.
837 * This can return NOOFFSET for inode-embedded data.
838 * The strategy code will take care of it in that case.
841 chain = hammer2_assign_physical(ip, parentp, lbase, pblksize,
842 mtid, &bdata, errorp);
843 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
844 hammer2_inode_data_t *wipdata;
846 wipdata = &chain->data->ipdata;
847 KKASSERT(wipdata->meta.op_flags &
848 HAMMER2_OPFLAG_DIRECTDATA);
849 bcopy(data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
850 ++hammer2_iod_file_wembed;
851 } else if (bdata == NULL) {
853 * Copy of data already present on-media.
855 chain->bref.methods =
856 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) +
857 HAMMER2_ENC_CHECK(ip->meta.check_algo);
858 hammer2_chain_setcheck(chain, data);
860 hammer2_write_bp(chain, data, ioflag, pblksize,
861 mtid, errorp, ip->meta.check_algo);
864 hammer2_chain_unlock(chain);
865 hammer2_chain_drop(chain);
868 case HAMMER2_COMP_AUTOZERO:
870 * Check for zero-fill only
872 hammer2_zero_check_and_write(data, ip, parentp,
873 lbase, ioflag, pblksize,
875 ip->meta.check_algo);
877 case HAMMER2_COMP_LZ4:
878 case HAMMER2_COMP_ZLIB:
881 * Check for zero-fill and attempt compression.
883 hammer2_compress_and_write(data, ip, parentp,
884 lbase, ioflag, pblksize,
887 ip->meta.check_algo);
895 * Generic function that will perform the compression in compression
896 * write path. The compression algorithm is determined by the settings
897 * obtained from inode.
901 hammer2_compress_and_write(char *data, hammer2_inode_t *ip,
902 hammer2_chain_t **parentp,
903 hammer2_key_t lbase, int ioflag, int pblksize,
904 hammer2_tid_t mtid, int *errorp, int comp_algo, int check_algo)
906 hammer2_chain_t *chain;
913 * An all-zeros write creates a hole unless the check code
914 * is disabled. When the check code is disabled all writes
915 * are done in-place, including any all-zeros writes.
917 * NOTE: A snapshot will still force a copy-on-write
918 * (see the HAMMER2_CHECK_NONE in hammer2_chain.c).
920 if (check_algo != HAMMER2_CHECK_NONE &&
921 test_block_zeros(data, pblksize)) {
922 zero_write(data, ip, parentp, lbase, mtid, errorp);
927 * Compression requested. Try to compress the block. We store
928 * the data normally if we cannot sufficiently compress it.
930 * We have a heuristic to detect files which are mostly
931 * uncompressable and avoid the compression attempt in that
932 * case. If the compression heuristic is turned off, we always
938 KKASSERT(pblksize / 2 <= 32768);
940 if (ip->comp_heuristic < 8 || (ip->comp_heuristic & 7) == 0 ||
941 hammer2_always_compress) {
942 z_stream strm_compress;
946 switch(HAMMER2_DEC_ALGO(comp_algo)) {
947 case HAMMER2_COMP_LZ4:
948 comp_buffer = objcache_get(cache_buffer_write,
950 comp_size = LZ4_compress_limitedOutput(
952 &comp_buffer[sizeof(int)],
954 pblksize / 2 - sizeof(int));
956 * We need to prefix with the size, LZ4
957 * doesn't do it for us. Add the related
960 *(int *)comp_buffer = comp_size;
962 comp_size += sizeof(int);
964 case HAMMER2_COMP_ZLIB:
965 comp_level = HAMMER2_DEC_LEVEL(comp_algo);
967 comp_level = 6; /* default zlib compression */
968 else if (comp_level < 6)
970 else if (comp_level > 9)
972 ret = deflateInit(&strm_compress, comp_level);
974 kprintf("HAMMER2 ZLIB: fatal error "
975 "on deflateInit.\n");
978 comp_buffer = objcache_get(cache_buffer_write,
980 strm_compress.next_in = data;
981 strm_compress.avail_in = pblksize;
982 strm_compress.next_out = comp_buffer;
983 strm_compress.avail_out = pblksize / 2;
984 ret = deflate(&strm_compress, Z_FINISH);
985 if (ret == Z_STREAM_END) {
986 comp_size = pblksize / 2 -
987 strm_compress.avail_out;
991 ret = deflateEnd(&strm_compress);
994 kprintf("Error: Unknown compression method.\n");
995 kprintf("Comp_method = %d.\n", comp_algo);
1000 if (comp_size == 0) {
1002 * compression failed or turned off
1004 comp_block_size = pblksize; /* safety */
1005 if (++ip->comp_heuristic > 128)
1006 ip->comp_heuristic = 8;
1009 * compression succeeded
1011 ip->comp_heuristic = 0;
1012 if (comp_size <= 1024) {
1013 comp_block_size = 1024;
1014 } else if (comp_size <= 2048) {
1015 comp_block_size = 2048;
1016 } else if (comp_size <= 4096) {
1017 comp_block_size = 4096;
1018 } else if (comp_size <= 8192) {
1019 comp_block_size = 8192;
1020 } else if (comp_size <= 16384) {
1021 comp_block_size = 16384;
1022 } else if (comp_size <= 32768) {
1023 comp_block_size = 32768;
1025 panic("hammer2: WRITE PATH: "
1026 "Weird comp_size value.");
1028 comp_block_size = pblksize;
1032 * Must zero the remainder or dedup (which operates on a
1033 * physical block basis) will not find matches.
1035 if (comp_size < comp_block_size) {
1036 bzero(comp_buffer + comp_size,
1037 comp_block_size - comp_size);
1042 * Assign physical storage, data will be set to NULL if a live-dedup
1045 bdata = comp_size ? comp_buffer : data;
1046 chain = hammer2_assign_physical(ip, parentp, lbase, comp_block_size,
1047 mtid, &bdata, errorp);
1050 kprintf("WRITE PATH: An error occurred while "
1051 "assigning physical space.\n");
1052 KKASSERT(chain == NULL);
1056 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
1057 hammer2_inode_data_t *wipdata;
1059 hammer2_chain_modify_ip(ip, chain, mtid, 0);
1060 wipdata = &chain->data->ipdata;
1061 KKASSERT(wipdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA);
1062 bcopy(data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
1063 ++hammer2_iod_file_wembed;
1064 } else if (bdata == NULL) {
1066 * Live deduplication, a copy of the data is already present
1070 chain->bref.methods =
1071 HAMMER2_ENC_COMP(comp_algo) +
1072 HAMMER2_ENC_CHECK(check_algo);
1074 chain->bref.methods =
1076 HAMMER2_COMP_NONE) +
1077 HAMMER2_ENC_CHECK(check_algo);
1079 bdata = comp_size ? comp_buffer : data;
1080 hammer2_chain_setcheck(chain, bdata);
1081 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1085 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
1087 switch(chain->bref.type) {
1088 case HAMMER2_BREF_TYPE_INODE:
1089 panic("hammer2_write_bp: unexpected inode\n");
1091 case HAMMER2_BREF_TYPE_DATA:
1093 * Optimize out the read-before-write
1096 *errorp = hammer2_io_newnz(chain->hmp,
1098 chain->bref.data_off,
1102 hammer2_io_brelse(&dio);
1103 kprintf("hammer2: WRITE PATH: "
1104 "dbp bread error\n");
1107 bdata = hammer2_io_data(dio, chain->bref.data_off);
1110 * When loading the block make sure we don't
1111 * leave garbage after the compressed data.
1114 chain->bref.methods =
1115 HAMMER2_ENC_COMP(comp_algo) +
1116 HAMMER2_ENC_CHECK(check_algo);
1117 bcopy(comp_buffer, bdata, comp_size);
1119 chain->bref.methods =
1121 HAMMER2_COMP_NONE) +
1122 HAMMER2_ENC_CHECK(check_algo);
1123 bcopy(data, bdata, pblksize);
1127 * The flush code doesn't calculate check codes for
1128 * file data (doing so can result in excessive I/O),
1131 hammer2_chain_setcheck(chain, bdata);
1134 * Device buffer is now valid, chain is no longer in
1135 * the initial state.
1137 * (No blockref table worries with file data)
1139 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1140 hammer2_dedup_record(chain, dio, bdata);
1142 /* Now write the related bdp. */
1143 if (ioflag & IO_SYNC) {
1145 * Synchronous I/O requested.
1147 hammer2_io_bwrite(&dio);
1149 } else if ((ioflag & IO_DIRECT) &&
1150 loff + n == pblksize) {
1151 hammer2_io_bdwrite(&dio);
1153 } else if (ioflag & IO_ASYNC) {
1154 hammer2_io_bawrite(&dio);
1156 hammer2_io_bdwrite(&dio);
1160 panic("hammer2_write_bp: bad chain type %d\n",
1168 hammer2_chain_unlock(chain);
1169 hammer2_chain_drop(chain);
1172 objcache_put(cache_buffer_write, comp_buffer);
1178 * Function that performs zero-checking and writing without compression,
1179 * it corresponds to default zero-checking path.
1183 hammer2_zero_check_and_write(char *data, hammer2_inode_t *ip,
1184 hammer2_chain_t **parentp,
1185 hammer2_key_t lbase, int ioflag, int pblksize,
1186 hammer2_tid_t mtid, int *errorp,
1189 hammer2_chain_t *chain;
1192 if (check_algo != HAMMER2_CHECK_NONE &&
1193 test_block_zeros(data, pblksize)) {
1195 * An all-zeros write creates a hole unless the check code
1196 * is disabled. When the check code is disabled all writes
1197 * are done in-place, including any all-zeros writes.
1199 * NOTE: A snapshot will still force a copy-on-write
1200 * (see the HAMMER2_CHECK_NONE in hammer2_chain.c).
1202 zero_write(data, ip, parentp, lbase, mtid, errorp);
1208 chain = hammer2_assign_physical(ip, parentp, lbase, pblksize,
1209 mtid, &bdata, errorp);
1211 hammer2_write_bp(chain, data, ioflag, pblksize,
1212 mtid, errorp, check_algo);
1214 /* dedup occurred */
1215 chain->bref.methods =
1216 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) +
1217 HAMMER2_ENC_CHECK(check_algo);
1218 hammer2_chain_setcheck(chain, data);
1221 hammer2_chain_unlock(chain);
1222 hammer2_chain_drop(chain);
1230 * A function to test whether a block of data contains only zeros,
1231 * returns TRUE (non-zero) if the block is all zeros.
1235 test_block_zeros(const char *buf, size_t bytes)
1239 for (i = 0; i < bytes; i += sizeof(long)) {
1240 if (*(const long *)(buf + i) != 0)
1249 * Function to "write" a block that contains only zeros.
1253 zero_write(char *data, hammer2_inode_t *ip,
1254 hammer2_chain_t **parentp,
1255 hammer2_key_t lbase, hammer2_tid_t mtid, int *errorp)
1257 hammer2_chain_t *chain;
1258 hammer2_key_t key_dummy;
1259 int cache_index = -1;
1262 chain = hammer2_chain_lookup(parentp, &key_dummy,
1265 HAMMER2_LOOKUP_NODATA);
1267 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
1268 hammer2_inode_data_t *wipdata;
1270 hammer2_chain_modify_ip(ip, chain, mtid, 0);
1271 wipdata = &chain->data->ipdata;
1272 KKASSERT(wipdata->meta.op_flags &
1273 HAMMER2_OPFLAG_DIRECTDATA);
1274 bzero(wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
1275 ++hammer2_iod_file_wembed;
1277 hammer2_chain_delete(*parentp, chain,
1278 mtid, HAMMER2_DELETE_PERMANENT);
1279 ++hammer2_iod_file_wzero;
1281 hammer2_chain_unlock(chain);
1282 hammer2_chain_drop(chain);
1284 ++hammer2_iod_file_wzero;
1291 * Function to write the data as it is, without performing any sort of
1292 * compression. This function is used in path without compression and
1293 * default zero-checking path.
1297 hammer2_write_bp(hammer2_chain_t *chain, char *data, int ioflag,
1299 hammer2_tid_t mtid, int *errorp, int check_algo)
1301 hammer2_inode_data_t *wipdata;
1306 error = 0; /* XXX TODO below */
1308 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
1310 switch(chain->bref.type) {
1311 case HAMMER2_BREF_TYPE_INODE:
1312 wipdata = &chain->data->ipdata;
1313 KKASSERT(wipdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA);
1314 bcopy(data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
1316 ++hammer2_iod_file_wembed;
1318 case HAMMER2_BREF_TYPE_DATA:
1319 error = hammer2_io_newnz(chain->hmp,
1321 chain->bref.data_off,
1322 chain->bytes, &dio);
1324 hammer2_io_bqrelse(&dio);
1325 kprintf("hammer2: WRITE PATH: "
1326 "dbp bread error\n");
1329 bdata = hammer2_io_data(dio, chain->bref.data_off);
1331 chain->bref.methods = HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) +
1332 HAMMER2_ENC_CHECK(check_algo);
1333 bcopy(data, bdata, chain->bytes);
1336 * The flush code doesn't calculate check codes for
1337 * file data (doing so can result in excessive I/O),
1340 hammer2_chain_setcheck(chain, bdata);
1343 * Device buffer is now valid, chain is no longer in
1344 * the initial state.
1346 * (No blockref table worries with file data)
1348 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1349 hammer2_dedup_record(chain, dio, bdata);
1351 if (ioflag & IO_SYNC) {
1353 * Synchronous I/O requested.
1355 hammer2_io_bwrite(&dio);
1357 } else if ((ioflag & IO_DIRECT) &&
1358 loff + n == pblksize) {
1359 hammer2_io_bdwrite(&dio);
1361 } else if (ioflag & IO_ASYNC) {
1362 hammer2_io_bawrite(&dio);
1364 hammer2_io_bdwrite(&dio);
1368 panic("hammer2_write_bp: bad chain type %d\n",
1374 KKASSERT(error == 0); /* XXX TODO */
1379 * LIVE DEDUP HEURISTICS
1381 * Record media and crc information for possible dedup operation. Note
1382 * that the dedup mask bits must also be set in the related DIO for a dedup
1383 * to be fully validated (which is handled in the freemap allocation code).
1385 * WARNING! This code is SMP safe but the heuristic allows SMP collisions.
1386 * All fields must be loaded into locals and validated.
1388 * WARNING! Should only be used for file data and directory entries,
1389 * hammer2_chain_modify() only checks for the dedup case on data
1390 * chains. Also, dedup data can only be recorded for committed
1391 * chains (so NOT strategy writes which can undergo further
1392 * modification after the fact!).
1395 hammer2_dedup_record(hammer2_chain_t *chain, hammer2_io_t *dio, char *data)
1398 hammer2_dedup_t *dedup;
1406 * We can only record a dedup if we have media data to test against.
1407 * If dedup is not enabled, return early, which allows a chain to
1408 * remain marked MODIFIED (which might have benefits in special
1409 * situations, though typically it does not).
1411 if (hammer2_dedup_enable == 0)
1421 switch(HAMMER2_DEC_CHECK(chain->bref.methods)) {
1422 case HAMMER2_CHECK_ISCSI32:
1424 * XXX use the built-in crc (the dedup lookup sequencing
1425 * needs to be fixed so the check code is already present
1426 * when dedup_lookup is called)
1429 crc = (uint64_t)(uint32_t)chain->bref.check.iscsi32.value;
1431 crc = XXH64(data, chain->bytes, XXH_HAMMER2_SEED);
1433 case HAMMER2_CHECK_XXHASH64:
1434 crc = chain->bref.check.xxhash64.value;
1436 case HAMMER2_CHECK_SHA192:
1438 * XXX use the built-in crc (the dedup lookup sequencing
1439 * needs to be fixed so the check code is already present
1440 * when dedup_lookup is called)
1443 crc = ((uint64_t *)chain->bref.check.sha192.data)[0] ^
1444 ((uint64_t *)chain->bref.check.sha192.data)[1] ^
1445 ((uint64_t *)chain->bref.check.sha192.data)[2];
1447 crc = XXH64(data, chain->bytes, XXH_HAMMER2_SEED);
1451 * Cannot dedup without a check code
1453 * NOTE: In particular, CHECK_NONE allows a sector to be
1454 * overwritten without copy-on-write, recording
1455 * a dedup block for a CHECK_NONE object would be
1461 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DEDUPABLE);
1463 dedup = &hmp->heur_dedup[crc & (HAMMER2_DEDUP_HEUR_MASK & ~3)];
1464 for (i = 0; i < 4; ++i) {
1465 if (dedup[i].data_crc == crc) {
1469 dticks = (int)(dedup[i].ticks - dedup[best].ticks);
1470 if (dticks < 0 || dticks > hz * 60 * 30)
1474 if (hammer2_debug & 0x40000) {
1475 kprintf("REC %04x %016jx %016jx\n",
1476 (int)(dedup - hmp->heur_dedup),
1478 chain->bref.data_off);
1480 dedup->ticks = ticks;
1481 dedup->data_off = chain->bref.data_off;
1482 dedup->data_crc = crc;
1485 * Set the valid bits for the dedup only after we know the data
1486 * buffer has been updated. The alloc bits were set (and the valid
1487 * bits cleared) when the media was allocated.
1489 * This is done in two stages becuase the bulkfree code can race
1490 * the gap between allocation and data population. Both masks must
1491 * be set before a bcmp/dedup operation is able to use the block.
1493 mask = hammer2_dedup_mask(dio, chain->bref.data_off, chain->bytes);
1494 atomic_set_64(&dio->dedup_valid, mask);
1498 * XXX removed. MODIFIED is an integral part of the flush code,
1499 * lets not just clear it
1502 * Once we record the dedup the chain must be marked clean to
1503 * prevent reuse of the underlying block. Remember that this
1504 * write occurs when the buffer cache is flushed (i.e. on sync(),
1505 * fsync(), filesystem periodic sync, or when the kernel needs to
1506 * flush a buffer), and not whenever the user write()s.
1508 if (chain->flags & HAMMER2_CHAIN_MODIFIED) {
1509 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
1510 atomic_add_long(&hammer2_count_modified_chains, -1);
1512 hammer2_pfs_memory_wakeup(chain->pmp);
1519 hammer2_dedup_lookup(hammer2_dev_t *hmp, char **datap, int pblksize)
1521 hammer2_dedup_t *dedup;
1530 if (hammer2_dedup_enable == 0)
1537 * XXX use the built-in crc (the dedup lookup sequencing
1538 * needs to be fixed so the check code is already present
1539 * when dedup_lookup is called)
1541 crc = XXH64(data, pblksize, XXH_HAMMER2_SEED);
1542 dedup = &hmp->heur_dedup[crc & (HAMMER2_DEDUP_HEUR_MASK & ~3)];
1544 if (hammer2_debug & 0x40000) {
1545 kprintf("LOC %04x/4 %016jx\n",
1546 (int)(dedup - hmp->heur_dedup),
1550 for (i = 0; i < 4; ++i) {
1551 off = dedup[i].data_off;
1553 if (dedup[i].data_crc != crc)
1555 if ((1 << (int)(off & HAMMER2_OFF_MASK_RADIX)) != pblksize)
1557 dio = hammer2_io_getquick(hmp, off, pblksize, 0);
1559 dtmp = hammer2_io_data(dio, off),
1560 mask = hammer2_dedup_mask(dio, off, pblksize);
1561 if ((dio->dedup_alloc & mask) == mask &&
1562 (dio->dedup_valid & mask) == mask &&
1563 bcmp(data, dtmp, pblksize) == 0) {
1564 if (hammer2_debug & 0x40000) {
1565 kprintf("DEDUP SUCCESS %016jx\n",
1568 hammer2_io_putblk(&dio);
1570 dedup[i].ticks = ticks; /* update use */
1571 atomic_add_long(&hammer2_iod_file_wdedup,
1574 return off; /* RETURN */
1576 hammer2_io_putblk(&dio);
1583 * Poof. Races are ok, if someone gets in and reuses a dedup offset
1584 * before or while we are clearing it they will also recover the freemap
1585 * entry (set it to fully allocated), so a bulkfree race can only set it
1586 * to a possibly-free state.
1588 * XXX ok, well, not really sure races are ok but going to run with it
1592 hammer2_dedup_clear(hammer2_dev_t *hmp)
1596 for (i = 0; i < HAMMER2_DEDUP_HEUR_SIZE; ++i) {
1597 hmp->heur_dedup[i].data_off = 0;
1598 hmp->heur_dedup[i].ticks = ticks - 1;