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 * WARNING: The strategy code cannot safely use hammer2 transactions
75 * as this can deadlock against vfs_sync's vfsync() call
76 * if multiple flushes are queued. All H2 structures must
77 * already be present and ready for the DIO.
79 * Reads can be initiated asynchronously, writes have to be
80 * spooled to a separate thread for action to avoid deadlocks.
82 static void hammer2_strategy_xop_read(hammer2_xop_t *arg, int clindex);
83 static void hammer2_strategy_xop_write(hammer2_xop_t *arg, int clindex);
84 static int hammer2_strategy_read(struct vop_strategy_args *ap);
85 static int hammer2_strategy_write(struct vop_strategy_args *ap);
86 static void hammer2_strategy_read_completion(hammer2_chain_t *chain,
87 char *data, struct bio *bio);
89 static void hammer2_dedup_record(hammer2_chain_t *chain, char *data);
90 static hammer2_off_t hammer2_dedup_lookup(hammer2_dev_t *hmp,
91 char **datap, int pblksize);
97 #define TIMER(which) do { \
99 h2timer[h2lid] += (int)(ticks - h2last);\
105 hammer2_vop_strategy(struct vop_strategy_args *ap)
116 error = hammer2_strategy_read(ap);
117 ++hammer2_iod_file_read;
120 error = hammer2_strategy_write(ap);
121 ++hammer2_iod_file_write;
124 bp->b_error = error = EINVAL;
125 bp->b_flags |= B_ERROR;
133 * Return the largest contiguous physical disk range for the logical
136 * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
138 * Basically disabled, the logical buffer write thread has to deal with
139 * buffers one-at-a-time. Note that this should not prevent cluster_read()
140 * from reading-ahead, it simply prevents it from trying form a single
141 * cluster buffer for the logical request. H2 already uses 64KB buffers!
144 hammer2_vop_bmap(struct vop_bmap_args *ap)
146 *ap->a_doffsetp = NOOFFSET;
154 /****************************************************************************
156 ****************************************************************************/
158 * Callback used in read path in case that a block is compressed with LZ4.
162 hammer2_decompress_LZ4_callback(const char *data, u_int bytes, struct bio *bio)
165 char *compressed_buffer;
172 if bio->bio_caller_info2.index &&
173 bio->bio_caller_info1.uvalue32 !=
174 crc32(bp->b_data, bp->b_bufsize) --- return error
177 KKASSERT(bp->b_bufsize <= HAMMER2_PBUFSIZE);
178 compressed_size = *(const int *)data;
179 KKASSERT(compressed_size <= bytes - sizeof(int));
181 compressed_buffer = objcache_get(cache_buffer_read, M_INTWAIT);
182 result = LZ4_decompress_safe(__DECONST(char *, &data[sizeof(int)]),
187 kprintf("READ PATH: Error during decompression."
189 (intmax_t)bio->bio_offset, bytes);
190 /* make sure it isn't random garbage */
191 bzero(compressed_buffer, bp->b_bufsize);
193 KKASSERT(result <= bp->b_bufsize);
194 bcopy(compressed_buffer, bp->b_data, bp->b_bufsize);
195 if (result < bp->b_bufsize)
196 bzero(bp->b_data + result, bp->b_bufsize - result);
197 objcache_put(cache_buffer_read, compressed_buffer);
199 bp->b_flags |= B_AGE;
203 * Callback used in read path in case that a block is compressed with ZLIB.
204 * It is almost identical to LZ4 callback, so in theory they can be unified,
205 * but we didn't want to make changes in bio structure for that.
209 hammer2_decompress_ZLIB_callback(const char *data, u_int bytes, struct bio *bio)
212 char *compressed_buffer;
213 z_stream strm_decompress;
219 KKASSERT(bp->b_bufsize <= HAMMER2_PBUFSIZE);
220 strm_decompress.avail_in = 0;
221 strm_decompress.next_in = Z_NULL;
223 ret = inflateInit(&strm_decompress);
226 kprintf("HAMMER2 ZLIB: Fatal error in inflateInit.\n");
228 compressed_buffer = objcache_get(cache_buffer_read, M_INTWAIT);
229 strm_decompress.next_in = __DECONST(char *, data);
231 /* XXX supply proper size, subset of device bp */
232 strm_decompress.avail_in = bytes;
233 strm_decompress.next_out = compressed_buffer;
234 strm_decompress.avail_out = bp->b_bufsize;
236 ret = inflate(&strm_decompress, Z_FINISH);
237 if (ret != Z_STREAM_END) {
238 kprintf("HAMMER2 ZLIB: Fatar error during decompression.\n");
239 bzero(compressed_buffer, bp->b_bufsize);
241 bcopy(compressed_buffer, bp->b_data, bp->b_bufsize);
242 result = bp->b_bufsize - strm_decompress.avail_out;
243 if (result < bp->b_bufsize)
244 bzero(bp->b_data + result, strm_decompress.avail_out);
245 objcache_put(cache_buffer_read, compressed_buffer);
246 ret = inflateEnd(&strm_decompress);
249 bp->b_flags |= B_AGE;
253 * Logical buffer I/O, async read.
257 hammer2_strategy_read(struct vop_strategy_args *ap)
259 hammer2_xop_strategy_t *xop;
269 nbio = push_bio(bio);
271 lbase = bio->bio_offset;
272 KKASSERT(((int)lbase & HAMMER2_PBUFMASK) == 0);
274 if (bp->b_bio1.bio_flags & BIO_SYNC) {
275 xop = hammer2_xop_alloc(ip, 0);
277 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_ITERATOR);
282 hammer2_mtx_init(&xop->lock, "h2bior");
283 hammer2_xop_start(&xop->head, hammer2_strategy_xop_read);
284 /* asynchronous completion */
290 * Per-node XOP (threaded), do a synchronous lookup of the chain and
291 * its data. The frontend is asynchronous, so we are also responsible
292 * for racing to terminate the frontend.
296 hammer2_strategy_xop_read(hammer2_xop_t *arg, int clindex)
298 hammer2_xop_strategy_t *xop = &arg->xop_strategy;
299 hammer2_chain_t *parent;
300 hammer2_chain_t *chain;
301 hammer2_key_t key_dummy;
305 int cache_index = -1;
313 parent = hammer2_inode_chain(xop->head.ip1, clindex,
314 HAMMER2_RESOLVE_ALWAYS |
315 HAMMER2_RESOLVE_SHARED);
318 chain = hammer2_chain_lookup(&parent, &key_dummy,
321 HAMMER2_LOOKUP_ALWAYS |
322 HAMMER2_LOOKUP_SHARED);
323 error = chain ? chain->error : 0;
329 error = hammer2_xop_feed(&xop->head, chain, clindex, error);
332 hammer2_chain_unlock(chain);
333 hammer2_chain_drop(chain);
336 hammer2_chain_unlock(parent);
337 hammer2_chain_drop(parent);
339 chain = NULL; /* safety */
340 parent = NULL; /* safety */
344 * Race to finish the frontend
348 hammer2_mtx_ex(&xop->lock);
350 hammer2_mtx_unlock(&xop->lock);
355 * Async operation has not completed and we now own the lock.
356 * Determine if we can complete the operation by issuing the
357 * frontend collection non-blocking.
359 error = hammer2_xop_collect(&xop->head, HAMMER2_XOP_COLLECT_NOWAIT);
365 hammer2_mtx_unlock(&xop->lock);
366 chain = xop->head.cluster.focus;
367 hammer2_strategy_read_completion(chain, (char *)chain->data,
369 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
374 hammer2_mtx_unlock(&xop->lock);
377 bzero(bp->b_data, bp->b_bcount);
379 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
382 hammer2_mtx_unlock(&xop->lock);
386 hammer2_mtx_unlock(&xop->lock);
387 bp->b_flags |= B_ERROR;
390 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
398 hammer2_strategy_read_completion(hammer2_chain_t *chain, char *data,
401 struct buf *bp = bio->bio_buf;
403 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
405 * Data is embedded in the inode (copy from inode).
407 bcopy(((hammer2_inode_data_t *)data)->u.data,
408 bp->b_data, HAMMER2_EMBEDDED_BYTES);
409 bzero(bp->b_data + HAMMER2_EMBEDDED_BYTES,
410 bp->b_bcount - HAMMER2_EMBEDDED_BYTES);
413 } else if (chain->bref.type == HAMMER2_BREF_TYPE_DATA) {
415 * Data is on-media, record for live dedup.
417 hammer2_dedup_record(chain, data);
420 * Decopmression and copy.
422 switch (HAMMER2_DEC_COMP(chain->bref.methods)) {
423 case HAMMER2_COMP_LZ4:
424 hammer2_decompress_LZ4_callback(data, chain->bytes,
427 case HAMMER2_COMP_ZLIB:
428 hammer2_decompress_ZLIB_callback(data, chain->bytes,
431 case HAMMER2_COMP_NONE:
432 KKASSERT(chain->bytes <= bp->b_bcount);
433 bcopy(data, bp->b_data, chain->bytes);
434 if (chain->bytes < bp->b_bcount) {
435 bzero(bp->b_data + chain->bytes,
436 bp->b_bcount - chain->bytes);
438 bp->b_flags |= B_NOTMETA;
443 panic("hammer2_strategy_read: "
444 "unknown compression type");
447 panic("hammer2_strategy_read: unknown bref type");
451 /****************************************************************************
453 ****************************************************************************/
456 * Functions for compression in threads,
457 * from hammer2_vnops.c
459 static void hammer2_write_file_core(struct buf *bp, hammer2_inode_t *ip,
460 hammer2_chain_t **parentp,
461 hammer2_key_t lbase, int ioflag, int pblksize,
462 hammer2_tid_t mtid, int *errorp);
463 static void hammer2_compress_and_write(struct buf *bp, hammer2_inode_t *ip,
464 hammer2_chain_t **parentp,
465 hammer2_key_t lbase, int ioflag, int pblksize,
466 hammer2_tid_t mtid, int *errorp,
467 int comp_algo, int check_algo);
468 static void hammer2_zero_check_and_write(struct buf *bp, hammer2_inode_t *ip,
469 hammer2_chain_t **parentp,
470 hammer2_key_t lbase, int ioflag, int pblksize,
471 hammer2_tid_t mtid, int *errorp,
473 static int test_block_zeros(const char *buf, size_t bytes);
474 static void zero_write(struct buf *bp, hammer2_inode_t *ip,
475 hammer2_chain_t **parentp,
477 hammer2_tid_t mtid, int *errorp);
478 static void hammer2_write_bp(hammer2_chain_t *chain, struct buf *bp,
479 int ioflag, int pblksize,
480 hammer2_tid_t mtid, int *errorp,
485 hammer2_strategy_write(struct vop_strategy_args *ap)
487 hammer2_xop_strategy_t *xop;
498 hammer2_lwinprog_ref(pmp);
499 hammer2_trans_assert_strategy(pmp);
501 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING);
504 xop->lbase = bio->bio_offset;
505 hammer2_mtx_init(&xop->lock, "h2biow");
506 hammer2_xop_start(&xop->head, hammer2_strategy_xop_write);
507 /* asynchronous completion */
509 hammer2_lwinprog_wait(pmp, hammer2_flush_pipe);
515 * Per-node XOP (threaded). Write the logical buffer to the media.
519 hammer2_strategy_xop_write(hammer2_xop_t *arg, int clindex)
521 hammer2_xop_strategy_t *xop = &arg->xop_strategy;
522 hammer2_chain_t *parent;
536 /* hammer2_trans_init(parent->hmp->spmp, HAMMER2_TRANS_BUFCACHE); */
538 lblksize = hammer2_calc_logical(ip, bio->bio_offset, &lbase, NULL);
539 pblksize = hammer2_calc_physical(ip, lbase);
540 parent = hammer2_inode_chain(ip, clindex, HAMMER2_RESOLVE_ALWAYS);
541 hammer2_write_file_core(bp, ip, &parent,
542 lbase, IO_ASYNC, pblksize,
543 xop->head.mtid, &error);
545 hammer2_chain_unlock(parent);
546 hammer2_chain_drop(parent);
547 parent = NULL; /* safety */
549 error = hammer2_xop_feed(&xop->head, NULL, clindex, error);
552 * Race to finish the frontend
556 hammer2_mtx_ex(&xop->lock);
558 hammer2_mtx_unlock(&xop->lock);
563 * Async operation has not completed and we now own the lock.
564 * Determine if we can complete the operation by issuing the
565 * frontend collection non-blocking.
567 error = hammer2_xop_collect(&xop->head, HAMMER2_XOP_COLLECT_NOWAIT);
573 hammer2_mtx_unlock(&xop->lock);
577 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
578 hammer2_lwinprog_drop(ip->pmp);
581 hammer2_mtx_unlock(&xop->lock);
585 hammer2_mtx_unlock(&xop->lock);
586 bp->b_flags |= B_ERROR;
589 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
590 hammer2_lwinprog_drop(ip->pmp);
596 * Wait for pending I/O to complete
599 hammer2_bioq_sync(hammer2_pfs_t *pmp)
601 hammer2_lwinprog_wait(pmp, 0);
605 * Create a new cluster at (cparent, lbase) and assign physical storage,
606 * returning a cluster suitable for I/O. The cluster will be in a modified
609 * cparent can wind up being anything.
611 * If datap is not NULL, *datap points to the real data we intend to write.
612 * If we can dedup the storage location we set *datap to NULL to indicate
613 * to the caller that a dedup occurred.
615 * NOTE: Special case for data embedded in inode.
619 hammer2_assign_physical(hammer2_inode_t *ip, hammer2_chain_t **parentp,
620 hammer2_key_t lbase, int pblksize,
621 hammer2_tid_t mtid, char **datap, int *errorp)
623 hammer2_chain_t *chain;
624 hammer2_key_t key_dummy;
625 hammer2_off_t dedup_off;
626 int pradix = hammer2_getradix(pblksize);
627 int cache_index = -1;
630 * Locate the chain associated with lbase, return a locked chain.
631 * However, do not instantiate any data reference (which utilizes a
632 * device buffer) because we will be using direct IO via the
633 * logical buffer cache buffer.
636 KKASSERT(pblksize >= HAMMER2_ALLOC_MIN);
639 chain = hammer2_chain_lookup(parentp, &key_dummy,
642 HAMMER2_LOOKUP_NODATA);
645 * We found a hole, create a new chain entry.
647 * NOTE: DATA chains are created without device backing
648 * store (nor do we want any).
650 dedup_off = hammer2_dedup_lookup((*parentp)->hmp, datap,
652 *errorp = hammer2_chain_create(parentp, &chain, ip->pmp,
653 lbase, HAMMER2_PBUFRADIX,
654 HAMMER2_BREF_TYPE_DATA,
658 panic("hammer2_chain_create: par=%p error=%d\n",
662 /*ip->delta_dcount += pblksize;*/
664 switch (chain->bref.type) {
665 case HAMMER2_BREF_TYPE_INODE:
667 * The data is embedded in the inode, which requires
670 hammer2_chain_modify_ip(ip, chain, mtid, 0);
672 case HAMMER2_BREF_TYPE_DATA:
673 dedup_off = hammer2_dedup_lookup(chain->hmp, datap,
675 if (chain->bytes != pblksize) {
676 hammer2_chain_resize(ip, *parentp, chain,
679 HAMMER2_MODIFY_OPTDATA);
683 * DATA buffers must be marked modified whether the
684 * data is in a logical buffer or not. We also have
685 * to make this call to fixup the chain data pointers
686 * after resizing in case this is an encrypted or
689 hammer2_chain_modify(chain, mtid, dedup_off,
690 HAMMER2_MODIFY_OPTDATA);
693 panic("hammer2_assign_physical: bad type");
703 * hammer2_write_file_core() - hammer2_write_thread() helper
705 * The core write function which determines which path to take
706 * depending on compression settings. We also have to locate the
707 * related chains so we can calculate and set the check data for
712 hammer2_write_file_core(struct buf *bp, hammer2_inode_t *ip,
713 hammer2_chain_t **parentp,
714 hammer2_key_t lbase, int ioflag, int pblksize,
715 hammer2_tid_t mtid, int *errorp)
717 hammer2_chain_t *chain;
718 char *data = bp->b_data;
720 switch(HAMMER2_DEC_ALGO(ip->meta.comp_algo)) {
721 case HAMMER2_COMP_NONE:
723 * We have to assign physical storage to the buffer
724 * we intend to dirty or write now to avoid deadlocks
725 * in the strategy code later.
727 * This can return NOOFFSET for inode-embedded data.
728 * The strategy code will take care of it in that case.
730 chain = hammer2_assign_physical(ip, parentp, lbase, pblksize,
731 mtid, &data, errorp);
732 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
733 hammer2_inode_data_t *wipdata;
735 wipdata = &chain->data->ipdata;
736 KKASSERT(wipdata->meta.op_flags &
737 HAMMER2_OPFLAG_DIRECTDATA);
738 KKASSERT(bp->b_loffset == 0);
739 bcopy(bp->b_data, wipdata->u.data,
740 HAMMER2_EMBEDDED_BYTES);
741 ++hammer2_iod_file_wembed;
742 } else if (data == NULL) {
744 * Copy of data already present on-media.
746 chain->bref.methods =
747 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) +
748 HAMMER2_ENC_CHECK(ip->meta.check_algo);
749 hammer2_chain_setcheck(chain, bp->b_data);
751 hammer2_write_bp(chain, bp, ioflag, pblksize,
752 mtid, errorp, ip->meta.check_algo);
755 hammer2_chain_unlock(chain);
756 hammer2_chain_drop(chain);
759 case HAMMER2_COMP_AUTOZERO:
761 * Check for zero-fill only
763 hammer2_zero_check_and_write(bp, ip, parentp,
764 lbase, ioflag, pblksize,
766 ip->meta.check_algo);
768 case HAMMER2_COMP_LZ4:
769 case HAMMER2_COMP_ZLIB:
772 * Check for zero-fill and attempt compression.
774 hammer2_compress_and_write(bp, ip, parentp,
775 lbase, ioflag, pblksize,
778 ip->meta.check_algo);
786 * Generic function that will perform the compression in compression
787 * write path. The compression algorithm is determined by the settings
788 * obtained from inode.
792 hammer2_compress_and_write(struct buf *bp, hammer2_inode_t *ip,
793 hammer2_chain_t **parentp,
794 hammer2_key_t lbase, int ioflag, int pblksize,
795 hammer2_tid_t mtid, int *errorp, int comp_algo, int check_algo)
797 hammer2_chain_t *chain;
803 if (test_block_zeros(bp->b_data, pblksize)) {
804 zero_write(bp, ip, parentp, lbase, mtid, errorp);
811 KKASSERT(pblksize / 2 <= 32768);
813 if (ip->comp_heuristic < 8 || (ip->comp_heuristic & 7) == 0) {
814 z_stream strm_compress;
818 switch(HAMMER2_DEC_ALGO(comp_algo)) {
819 case HAMMER2_COMP_LZ4:
820 comp_buffer = objcache_get(cache_buffer_write,
822 comp_size = LZ4_compress_limitedOutput(
824 &comp_buffer[sizeof(int)],
826 pblksize / 2 - sizeof(int));
828 * We need to prefix with the size, LZ4
829 * doesn't do it for us. Add the related
832 *(int *)comp_buffer = comp_size;
834 comp_size += sizeof(int);
836 case HAMMER2_COMP_ZLIB:
837 comp_level = HAMMER2_DEC_LEVEL(comp_algo);
839 comp_level = 6; /* default zlib compression */
840 else if (comp_level < 6)
842 else if (comp_level > 9)
844 ret = deflateInit(&strm_compress, comp_level);
846 kprintf("HAMMER2 ZLIB: fatal error "
847 "on deflateInit.\n");
850 comp_buffer = objcache_get(cache_buffer_write,
852 strm_compress.next_in = bp->b_data;
853 strm_compress.avail_in = pblksize;
854 strm_compress.next_out = comp_buffer;
855 strm_compress.avail_out = pblksize / 2;
856 ret = deflate(&strm_compress, Z_FINISH);
857 if (ret == Z_STREAM_END) {
858 comp_size = pblksize / 2 -
859 strm_compress.avail_out;
863 ret = deflateEnd(&strm_compress);
866 kprintf("Error: Unknown compression method.\n");
867 kprintf("Comp_method = %d.\n", comp_algo);
872 if (comp_size == 0) {
874 * compression failed or turned off
876 comp_block_size = pblksize; /* safety */
877 if (++ip->comp_heuristic > 128)
878 ip->comp_heuristic = 8;
881 * compression succeeded
883 ip->comp_heuristic = 0;
884 if (comp_size <= 1024) {
885 comp_block_size = 1024;
886 } else if (comp_size <= 2048) {
887 comp_block_size = 2048;
888 } else if (comp_size <= 4096) {
889 comp_block_size = 4096;
890 } else if (comp_size <= 8192) {
891 comp_block_size = 8192;
892 } else if (comp_size <= 16384) {
893 comp_block_size = 16384;
894 } else if (comp_size <= 32768) {
895 comp_block_size = 32768;
897 panic("hammer2: WRITE PATH: "
898 "Weird comp_size value.");
900 comp_block_size = pblksize;
904 * Must zero the remainder or dedup (which operates on a
905 * physical block basis) will not find matches.
907 if (comp_size < comp_block_size) {
908 bzero(comp_buffer + comp_size,
909 comp_block_size - comp_size);
914 * Assign physical storage, data will be set to NULL if a live-dedup
917 data = comp_size ? comp_buffer : bp->b_data;
918 chain = hammer2_assign_physical(ip, parentp, lbase, comp_block_size,
919 mtid, &data, errorp);
922 kprintf("WRITE PATH: An error occurred while "
923 "assigning physical space.\n");
924 KKASSERT(chain == NULL);
928 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
929 hammer2_inode_data_t *wipdata;
931 hammer2_chain_modify_ip(ip, chain, mtid, 0);
932 wipdata = &chain->data->ipdata;
933 KKASSERT(wipdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA);
934 KKASSERT(bp->b_loffset == 0);
935 bcopy(bp->b_data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
936 ++hammer2_iod_file_wembed;
937 } else if (data == NULL) {
939 * Live deduplication, a copy of the data is already present
945 chain->bref.methods =
946 HAMMER2_ENC_COMP(comp_algo) +
947 HAMMER2_ENC_CHECK(check_algo);
949 chain->bref.methods =
952 HAMMER2_ENC_CHECK(check_algo);
954 bdata = comp_size ? comp_buffer : bp->b_data;
955 hammer2_chain_setcheck(chain, bdata);
956 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
961 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
963 switch(chain->bref.type) {
964 case HAMMER2_BREF_TYPE_INODE:
965 panic("hammer2_write_bp: unexpected inode\n");
967 case HAMMER2_BREF_TYPE_DATA:
969 * Optimize out the read-before-write
972 *errorp = hammer2_io_newnz(chain->hmp,
973 chain->bref.data_off,
977 hammer2_io_brelse(&dio);
978 kprintf("hammer2: WRITE PATH: "
979 "dbp bread error\n");
982 bdata = hammer2_io_data(dio, chain->bref.data_off);
985 * When loading the block make sure we don't
986 * leave garbage after the compressed data.
989 chain->bref.methods =
990 HAMMER2_ENC_COMP(comp_algo) +
991 HAMMER2_ENC_CHECK(check_algo);
992 bcopy(comp_buffer, bdata, comp_size);
994 chain->bref.methods =
997 HAMMER2_ENC_CHECK(check_algo);
998 bcopy(bp->b_data, bdata, pblksize);
1002 * The flush code doesn't calculate check codes for
1003 * file data (doing so can result in excessive I/O),
1006 * Record for dedup only after the DIO's buffer cache
1007 * buffer has been updated.
1009 hammer2_chain_setcheck(chain, bdata);
1010 hammer2_dedup_record(chain, bdata);
1013 * Device buffer is now valid, chain is no longer in
1014 * the initial state.
1016 * (No blockref table worries with file data)
1018 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1020 /* Now write the related bdp. */
1021 if (ioflag & IO_SYNC) {
1023 * Synchronous I/O requested.
1025 hammer2_io_bwrite(&dio);
1027 } else if ((ioflag & IO_DIRECT) &&
1028 loff + n == pblksize) {
1029 hammer2_io_bdwrite(&dio);
1031 } else if (ioflag & IO_ASYNC) {
1032 hammer2_io_bawrite(&dio);
1034 hammer2_io_bdwrite(&dio);
1038 panic("hammer2_write_bp: bad chain type %d\n",
1046 hammer2_chain_unlock(chain);
1047 hammer2_chain_drop(chain);
1050 objcache_put(cache_buffer_write, comp_buffer);
1056 * Function that performs zero-checking and writing without compression,
1057 * it corresponds to default zero-checking path.
1061 hammer2_zero_check_and_write(struct buf *bp, hammer2_inode_t *ip,
1062 hammer2_chain_t **parentp,
1063 hammer2_key_t lbase, int ioflag, int pblksize,
1064 hammer2_tid_t mtid, int *errorp,
1067 hammer2_chain_t *chain;
1068 char *data = bp->b_data;
1070 if (test_block_zeros(bp->b_data, pblksize)) {
1071 zero_write(bp, ip, parentp, lbase, mtid, errorp);
1073 chain = hammer2_assign_physical(ip, parentp, lbase, pblksize,
1074 mtid, &data, errorp);
1076 hammer2_write_bp(chain, bp, ioflag, pblksize,
1077 mtid, errorp, check_algo);
1078 } /* else dedup occurred */
1080 hammer2_chain_unlock(chain);
1081 hammer2_chain_drop(chain);
1089 * A function to test whether a block of data contains only zeros,
1090 * returns TRUE (non-zero) if the block is all zeros.
1094 test_block_zeros(const char *buf, size_t bytes)
1098 for (i = 0; i < bytes; i += sizeof(long)) {
1099 if (*(const long *)(buf + i) != 0)
1108 * Function to "write" a block that contains only zeros.
1112 zero_write(struct buf *bp, hammer2_inode_t *ip,
1113 hammer2_chain_t **parentp,
1114 hammer2_key_t lbase, hammer2_tid_t mtid, int *errorp __unused)
1116 hammer2_chain_t *chain;
1117 hammer2_key_t key_dummy;
1118 int cache_index = -1;
1120 chain = hammer2_chain_lookup(parentp, &key_dummy,
1123 HAMMER2_LOOKUP_NODATA);
1125 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
1126 hammer2_inode_data_t *wipdata;
1128 hammer2_chain_modify_ip(ip, chain, mtid, 0);
1129 wipdata = &chain->data->ipdata;
1130 KKASSERT(wipdata->meta.op_flags &
1131 HAMMER2_OPFLAG_DIRECTDATA);
1132 KKASSERT(bp->b_loffset == 0);
1133 bzero(wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
1134 ++hammer2_iod_file_wembed;
1136 hammer2_chain_delete(*parentp, chain,
1137 mtid, HAMMER2_DELETE_PERMANENT);
1138 ++hammer2_iod_file_wzero;
1140 hammer2_chain_unlock(chain);
1141 hammer2_chain_drop(chain);
1143 ++hammer2_iod_file_wzero;
1150 * Function to write the data as it is, without performing any sort of
1151 * compression. This function is used in path without compression and
1152 * default zero-checking path.
1156 hammer2_write_bp(hammer2_chain_t *chain, struct buf *bp, int ioflag,
1158 hammer2_tid_t mtid, int *errorp, int check_algo)
1160 hammer2_inode_data_t *wipdata;
1165 error = 0; /* XXX TODO below */
1167 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
1169 switch(chain->bref.type) {
1170 case HAMMER2_BREF_TYPE_INODE:
1171 wipdata = &chain->data->ipdata;
1172 KKASSERT(wipdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA);
1173 KKASSERT(bp->b_loffset == 0);
1174 bcopy(bp->b_data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
1176 ++hammer2_iod_file_wembed;
1178 case HAMMER2_BREF_TYPE_DATA:
1179 error = hammer2_io_newnz(chain->hmp,
1180 chain->bref.data_off,
1181 chain->bytes, &dio);
1183 hammer2_io_bqrelse(&dio);
1184 kprintf("hammer2: WRITE PATH: "
1185 "dbp bread error\n");
1188 bdata = hammer2_io_data(dio, chain->bref.data_off);
1190 chain->bref.methods = HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) +
1191 HAMMER2_ENC_CHECK(check_algo);
1192 bcopy(bp->b_data, bdata, chain->bytes);
1195 * The flush code doesn't calculate check codes for
1196 * file data (doing so can result in excessive I/O),
1199 * Record for dedup only after the DIO's buffer cache
1200 * buffer has been updated.
1202 hammer2_chain_setcheck(chain, bdata);
1203 hammer2_dedup_record(chain, bdata);
1206 * Device buffer is now valid, chain is no longer in
1207 * the initial state.
1209 * (No blockref table worries with file data)
1211 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1213 if (ioflag & IO_SYNC) {
1215 * Synchronous I/O requested.
1217 hammer2_io_bwrite(&dio);
1219 } else if ((ioflag & IO_DIRECT) &&
1220 loff + n == pblksize) {
1221 hammer2_io_bdwrite(&dio);
1223 } else if (ioflag & IO_ASYNC) {
1224 hammer2_io_bawrite(&dio);
1226 hammer2_io_bdwrite(&dio);
1230 panic("hammer2_write_bp: bad chain type %d\n",
1236 KKASSERT(error == 0); /* XXX TODO */
1241 * LIVE DEDUP HEURISTIC
1243 * WARNING! This code is SMP safe but the heuristic allows SMP collisions.
1244 * All fields must be loaded into locals and validated.
1248 hammer2_dedup_record(hammer2_chain_t *chain, char *data)
1251 hammer2_dedup_t *dedup;
1259 switch(HAMMER2_DEC_CHECK(chain->bref.methods)) {
1260 case HAMMER2_CHECK_ISCSI32:
1262 * XXX use the built-in crc (the dedup lookup sequencing
1263 * needs to be fixed so the check code is already present
1264 * when dedup_lookup is called)
1267 crc = (uint64_t)(uint32_t)chain->bref.check.iscsi32.value;
1269 crc = XXH64(data, chain->bytes, XXH_HAMMER2_SEED);
1271 case HAMMER2_CHECK_XXHASH64:
1272 crc = chain->bref.check.xxhash64.value;
1274 case HAMMER2_CHECK_SHA192:
1276 * XXX use the built-in crc (the dedup lookup sequencing
1277 * needs to be fixed so the check code is already present
1278 * when dedup_lookup is called)
1281 crc = ((uint64_t *)chain->bref.check.sha192.data)[0] ^
1282 ((uint64_t *)chain->bref.check.sha192.data)[1] ^
1283 ((uint64_t *)chain->bref.check.sha192.data)[2];
1285 crc = XXH64(data, chain->bytes, XXH_HAMMER2_SEED);
1289 * Cannot dedup without a check code
1293 dedup = &hmp->heur_dedup[crc & (HAMMER2_DEDUP_HEUR_MASK & ~3)];
1294 for (i = 0; i < 4; ++i) {
1295 if (dedup[i].data_crc == crc) {
1299 dticks = (int)(dedup[i].ticks - dedup[best].ticks);
1300 if (dticks < 0 || dticks > hz * 60 * 30)
1304 if (hammer2_debug & 0x40000) {
1305 kprintf("REC %04x %016jx %016jx\n",
1306 (int)(dedup - hmp->heur_dedup),
1308 chain->bref.data_off);
1310 dedup->ticks = ticks;
1311 dedup->data_off = chain->bref.data_off;
1312 dedup->data_crc = crc;
1313 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DEDUP);
1318 hammer2_dedup_lookup(hammer2_dev_t *hmp, char **datap, int pblksize)
1320 hammer2_dedup_t *dedup;
1332 * XXX use the built-in crc (the dedup lookup sequencing
1333 * needs to be fixed so the check code is already present
1334 * when dedup_lookup is called)
1336 crc = XXH64(data, pblksize, XXH_HAMMER2_SEED);
1337 dedup = &hmp->heur_dedup[crc & (HAMMER2_DEDUP_HEUR_MASK & ~3)];
1339 if (hammer2_debug & 0x40000) {
1340 kprintf("LOC %04x/4 %016jx\n",
1341 (int)(dedup - hmp->heur_dedup),
1345 for (i = 0; i < 4; ++i) {
1346 off = dedup[i].data_off;
1348 if (dedup[i].data_crc != crc)
1350 if ((1 << (int)(off & HAMMER2_OFF_MASK_RADIX)) != pblksize)
1352 dio = hammer2_io_getquick(hmp, off, pblksize);
1354 bcmp(data, hammer2_io_data(dio, off), pblksize) == 0) {
1355 if (hammer2_debug & 0x40000) {
1356 kprintf("DEDUP SUCCESS %016jx\n",
1359 hammer2_io_putblk(&dio);
1361 dedup[i].ticks = ticks; /* update use */
1362 ++hammer2_iod_file_wdedup;
1363 return off; /* RETURN */
1366 hammer2_io_putblk(&dio);
1372 * Poof. Races are ok, if someone gets in and reuses a dedup offset
1373 * before or while we are clearing it they will also recover the freemap
1374 * entry (set it to fully allocated), so a bulkfree race can only set it
1375 * to a possibly-free state.
1377 * XXX ok, well, not really sure races are ok but going to run with it
1381 hammer2_dedup_clear(hammer2_dev_t *hmp)
1385 for (i = 0; i < HAMMER2_DEDUP_HEUR_SIZE; ++i) {
1386 hmp->heur_dedup[i].data_off = 0;
1387 hmp->heur_dedup[i].ticks = ticks - 1;