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((uint32_t)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;
314 * This is difficult to optimize. The logical buffer might be
315 * partially dirty (contain dummy zero-fill pages), which would
316 * mess up our crc calculation if we were to try a direct read.
317 * So for now we always double-buffer through the underlying
320 * If not for the above problem we could conditionalize on
321 * (1) 64KB buffer, (2) one chain (not multi-master) and
322 * (3) !hammer2_double_buffer, and issue a direct read into the
325 parent = hammer2_inode_chain(xop->head.ip1, clindex,
326 HAMMER2_RESOLVE_ALWAYS |
327 HAMMER2_RESOLVE_SHARED);
330 chain = hammer2_chain_lookup(&parent, &key_dummy,
333 HAMMER2_LOOKUP_ALWAYS |
334 HAMMER2_LOOKUP_SHARED);
335 error = chain ? chain->error : 0;
341 error = hammer2_xop_feed(&xop->head, chain, clindex, error);
344 hammer2_chain_unlock(chain);
345 hammer2_chain_drop(chain);
348 hammer2_chain_unlock(parent);
349 hammer2_chain_drop(parent);
351 chain = NULL; /* safety */
352 parent = NULL; /* safety */
356 * Race to finish the frontend
360 hammer2_mtx_ex(&xop->lock);
362 hammer2_mtx_unlock(&xop->lock);
367 * Async operation has not completed and we now own the lock.
368 * Determine if we can complete the operation by issuing the
369 * frontend collection non-blocking.
371 error = hammer2_xop_collect(&xop->head, HAMMER2_XOP_COLLECT_NOWAIT);
377 hammer2_mtx_unlock(&xop->lock);
378 chain = xop->head.cluster.focus;
379 hammer2_strategy_read_completion(chain, (char *)chain->data,
382 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
386 hammer2_mtx_unlock(&xop->lock);
389 bzero(bp->b_data, bp->b_bcount);
391 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
394 hammer2_mtx_unlock(&xop->lock);
397 kprintf("strategy_xop_read: error %d loff=%016jx\n",
398 error, bp->b_loffset);
400 hammer2_mtx_unlock(&xop->lock);
401 bp->b_flags |= B_ERROR;
404 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
412 hammer2_strategy_read_completion(hammer2_chain_t *chain, char *data,
415 struct buf *bp = bio->bio_buf;
417 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
419 * Data is embedded in the inode (copy from inode).
421 bcopy(((hammer2_inode_data_t *)data)->u.data,
422 bp->b_data, HAMMER2_EMBEDDED_BYTES);
423 bzero(bp->b_data + HAMMER2_EMBEDDED_BYTES,
424 bp->b_bcount - HAMMER2_EMBEDDED_BYTES);
427 } else if (chain->bref.type == HAMMER2_BREF_TYPE_DATA) {
429 * Data is on-media, record for live dedup.
431 hammer2_dedup_record(chain, data);
434 * Decompression and copy.
436 switch (HAMMER2_DEC_COMP(chain->bref.methods)) {
437 case HAMMER2_COMP_LZ4:
438 hammer2_decompress_LZ4_callback(data, chain->bytes,
440 /* b_resid set by call */
442 case HAMMER2_COMP_ZLIB:
443 hammer2_decompress_ZLIB_callback(data, chain->bytes,
445 /* b_resid set by call */
447 case HAMMER2_COMP_NONE:
448 KKASSERT(chain->bytes <= bp->b_bcount);
449 bcopy(data, bp->b_data, chain->bytes);
450 if (chain->bytes < bp->b_bcount) {
451 bzero(bp->b_data + chain->bytes,
452 bp->b_bcount - chain->bytes);
454 bp->b_flags |= B_NOTMETA;
459 panic("hammer2_strategy_read: "
460 "unknown compression type");
463 panic("hammer2_strategy_read: unknown bref type");
467 /****************************************************************************
469 ****************************************************************************/
472 * Functions for compression in threads,
473 * from hammer2_vnops.c
475 static void hammer2_write_file_core(struct buf *bp, hammer2_inode_t *ip,
476 hammer2_chain_t **parentp,
477 hammer2_key_t lbase, int ioflag, int pblksize,
478 hammer2_tid_t mtid, int *errorp);
479 static void hammer2_compress_and_write(struct buf *bp, hammer2_inode_t *ip,
480 hammer2_chain_t **parentp,
481 hammer2_key_t lbase, int ioflag, int pblksize,
482 hammer2_tid_t mtid, int *errorp,
483 int comp_algo, int check_algo);
484 static void hammer2_zero_check_and_write(struct buf *bp, hammer2_inode_t *ip,
485 hammer2_chain_t **parentp,
486 hammer2_key_t lbase, int ioflag, int pblksize,
487 hammer2_tid_t mtid, int *errorp,
489 static int test_block_zeros(const char *buf, size_t bytes);
490 static void zero_write(struct buf *bp, hammer2_inode_t *ip,
491 hammer2_chain_t **parentp,
493 hammer2_tid_t mtid, int *errorp);
494 static void hammer2_write_bp(hammer2_chain_t *chain, struct buf *bp,
495 int ioflag, int pblksize,
496 hammer2_tid_t mtid, int *errorp,
501 hammer2_strategy_write(struct vop_strategy_args *ap)
503 hammer2_xop_strategy_t *xop;
514 hammer2_lwinprog_ref(pmp);
515 hammer2_trans_assert_strategy(pmp);
517 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING);
520 xop->lbase = bio->bio_offset;
521 hammer2_mtx_init(&xop->lock, "h2biow");
522 hammer2_xop_start(&xop->head, hammer2_strategy_xop_write);
523 /* asynchronous completion */
525 hammer2_lwinprog_wait(pmp, hammer2_flush_pipe);
531 * Per-node XOP (threaded). Write the logical buffer to the media.
535 hammer2_strategy_xop_write(hammer2_xop_t *arg, int clindex)
537 hammer2_xop_strategy_t *xop = &arg->xop_strategy;
538 hammer2_chain_t *parent;
552 /* hammer2_trans_init(parent->hmp->spmp, HAMMER2_TRANS_BUFCACHE); */
554 lblksize = hammer2_calc_logical(ip, bio->bio_offset, &lbase, NULL);
555 pblksize = hammer2_calc_physical(ip, lbase);
556 parent = hammer2_inode_chain(ip, clindex, HAMMER2_RESOLVE_ALWAYS);
557 hammer2_write_file_core(bp, ip, &parent,
558 lbase, IO_ASYNC, pblksize,
559 xop->head.mtid, &error);
561 hammer2_chain_unlock(parent);
562 hammer2_chain_drop(parent);
563 parent = NULL; /* safety */
565 hammer2_xop_feed(&xop->head, NULL, clindex, error);
568 * Race to finish the frontend
572 hammer2_mtx_ex(&xop->lock);
574 hammer2_mtx_unlock(&xop->lock);
579 * Async operation has not completed and we now own the lock.
580 * Determine if we can complete the operation by issuing the
581 * frontend collection non-blocking.
583 error = hammer2_xop_collect(&xop->head, HAMMER2_XOP_COLLECT_NOWAIT);
589 hammer2_mtx_unlock(&xop->lock);
593 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
594 hammer2_lwinprog_drop(ip->pmp);
597 hammer2_mtx_unlock(&xop->lock);
600 kprintf("strategy_xop_write: error %d loff=%016jx\n",
601 error, bp->b_loffset);
603 hammer2_mtx_unlock(&xop->lock);
604 bp->b_flags |= B_ERROR;
607 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
608 hammer2_lwinprog_drop(ip->pmp);
614 * Wait for pending I/O to complete
617 hammer2_bioq_sync(hammer2_pfs_t *pmp)
619 hammer2_lwinprog_wait(pmp, 0);
623 * Create a new cluster at (cparent, lbase) and assign physical storage,
624 * returning a cluster suitable for I/O. The cluster will be in a modified
627 * cparent can wind up being anything.
629 * If datap is not NULL, *datap points to the real data we intend to write.
630 * If we can dedup the storage location we set *datap to NULL to indicate
631 * to the caller that a dedup occurred.
633 * NOTE: Special case for data embedded in inode.
637 hammer2_assign_physical(hammer2_inode_t *ip, hammer2_chain_t **parentp,
638 hammer2_key_t lbase, int pblksize,
639 hammer2_tid_t mtid, char **datap, int *errorp)
641 hammer2_chain_t *chain;
642 hammer2_key_t key_dummy;
643 hammer2_off_t dedup_off;
644 int pradix = hammer2_getradix(pblksize);
645 int cache_index = -1;
648 * Locate the chain associated with lbase, return a locked chain.
649 * However, do not instantiate any data reference (which utilizes a
650 * device buffer) because we will be using direct IO via the
651 * logical buffer cache buffer.
654 KKASSERT(pblksize >= HAMMER2_ALLOC_MIN);
657 chain = hammer2_chain_lookup(parentp, &key_dummy,
660 HAMMER2_LOOKUP_NODATA);
662 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED))
663 kprintf("assign physical deleted chain @ %016jx\n",
668 * We found a hole, create a new chain entry.
670 * NOTE: DATA chains are created without device backing
671 * store (nor do we want any).
673 dedup_off = hammer2_dedup_lookup((*parentp)->hmp, datap,
675 *errorp = hammer2_chain_create(parentp, &chain, ip->pmp,
676 lbase, HAMMER2_PBUFRADIX,
677 HAMMER2_BREF_TYPE_DATA,
681 panic("hammer2_chain_create: par=%p error=%d\n",
685 /*ip->delta_dcount += pblksize;*/
687 switch (chain->bref.type) {
688 case HAMMER2_BREF_TYPE_INODE:
690 * The data is embedded in the inode, which requires
693 hammer2_chain_modify_ip(ip, chain, mtid, 0);
695 case HAMMER2_BREF_TYPE_DATA:
696 dedup_off = hammer2_dedup_lookup(chain->hmp, datap,
698 if (chain->bytes != pblksize) {
699 hammer2_chain_resize(ip, *parentp, chain,
702 HAMMER2_MODIFY_OPTDATA);
706 * DATA buffers must be marked modified whether the
707 * data is in a logical buffer or not. We also have
708 * to make this call to fixup the chain data pointers
709 * after resizing in case this is an encrypted or
712 hammer2_chain_modify(chain, mtid, dedup_off,
713 HAMMER2_MODIFY_OPTDATA);
716 panic("hammer2_assign_physical: bad type");
726 * hammer2_write_file_core() - hammer2_write_thread() helper
728 * The core write function which determines which path to take
729 * depending on compression settings. We also have to locate the
730 * related chains so we can calculate and set the check data for
735 hammer2_write_file_core(struct buf *bp, hammer2_inode_t *ip,
736 hammer2_chain_t **parentp,
737 hammer2_key_t lbase, int ioflag, int pblksize,
738 hammer2_tid_t mtid, int *errorp)
740 hammer2_chain_t *chain;
741 char *data = bp->b_data;
745 switch(HAMMER2_DEC_ALGO(ip->meta.comp_algo)) {
746 case HAMMER2_COMP_NONE:
748 * We have to assign physical storage to the buffer
749 * we intend to dirty or write now to avoid deadlocks
750 * in the strategy code later.
752 * This can return NOOFFSET for inode-embedded data.
753 * The strategy code will take care of it in that case.
755 chain = hammer2_assign_physical(ip, parentp, lbase, pblksize,
756 mtid, &data, errorp);
757 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
758 hammer2_inode_data_t *wipdata;
760 wipdata = &chain->data->ipdata;
761 KKASSERT(wipdata->meta.op_flags &
762 HAMMER2_OPFLAG_DIRECTDATA);
763 KKASSERT(bp->b_loffset == 0);
764 bcopy(bp->b_data, wipdata->u.data,
765 HAMMER2_EMBEDDED_BYTES);
766 ++hammer2_iod_file_wembed;
767 } else if (data == NULL) {
769 * Copy of data already present on-media.
771 chain->bref.methods =
772 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) +
773 HAMMER2_ENC_CHECK(ip->meta.check_algo);
774 hammer2_chain_setcheck(chain, bp->b_data);
776 hammer2_write_bp(chain, bp, ioflag, pblksize,
777 mtid, errorp, ip->meta.check_algo);
780 hammer2_chain_unlock(chain);
781 hammer2_chain_drop(chain);
784 case HAMMER2_COMP_AUTOZERO:
786 * Check for zero-fill only
788 hammer2_zero_check_and_write(bp, ip, parentp,
789 lbase, ioflag, pblksize,
791 ip->meta.check_algo);
793 case HAMMER2_COMP_LZ4:
794 case HAMMER2_COMP_ZLIB:
797 * Check for zero-fill and attempt compression.
799 hammer2_compress_and_write(bp, ip, parentp,
800 lbase, ioflag, pblksize,
803 ip->meta.check_algo);
811 * Generic function that will perform the compression in compression
812 * write path. The compression algorithm is determined by the settings
813 * obtained from inode.
817 hammer2_compress_and_write(struct buf *bp, hammer2_inode_t *ip,
818 hammer2_chain_t **parentp,
819 hammer2_key_t lbase, int ioflag, int pblksize,
820 hammer2_tid_t mtid, int *errorp, int comp_algo, int check_algo)
822 hammer2_chain_t *chain;
828 if (test_block_zeros(bp->b_data, pblksize)) {
829 zero_write(bp, ip, parentp, lbase, mtid, errorp);
836 KKASSERT(pblksize / 2 <= 32768);
838 if (ip->comp_heuristic < 8 || (ip->comp_heuristic & 7) == 0) {
839 z_stream strm_compress;
843 switch(HAMMER2_DEC_ALGO(comp_algo)) {
844 case HAMMER2_COMP_LZ4:
845 comp_buffer = objcache_get(cache_buffer_write,
847 comp_size = LZ4_compress_limitedOutput(
849 &comp_buffer[sizeof(int)],
851 pblksize / 2 - sizeof(int));
853 * We need to prefix with the size, LZ4
854 * doesn't do it for us. Add the related
857 *(int *)comp_buffer = comp_size;
859 comp_size += sizeof(int);
861 case HAMMER2_COMP_ZLIB:
862 comp_level = HAMMER2_DEC_LEVEL(comp_algo);
864 comp_level = 6; /* default zlib compression */
865 else if (comp_level < 6)
867 else if (comp_level > 9)
869 ret = deflateInit(&strm_compress, comp_level);
871 kprintf("HAMMER2 ZLIB: fatal error "
872 "on deflateInit.\n");
875 comp_buffer = objcache_get(cache_buffer_write,
877 strm_compress.next_in = bp->b_data;
878 strm_compress.avail_in = pblksize;
879 strm_compress.next_out = comp_buffer;
880 strm_compress.avail_out = pblksize / 2;
881 ret = deflate(&strm_compress, Z_FINISH);
882 if (ret == Z_STREAM_END) {
883 comp_size = pblksize / 2 -
884 strm_compress.avail_out;
888 ret = deflateEnd(&strm_compress);
891 kprintf("Error: Unknown compression method.\n");
892 kprintf("Comp_method = %d.\n", comp_algo);
897 if (comp_size == 0) {
899 * compression failed or turned off
901 comp_block_size = pblksize; /* safety */
902 if (++ip->comp_heuristic > 128)
903 ip->comp_heuristic = 8;
906 * compression succeeded
908 ip->comp_heuristic = 0;
909 if (comp_size <= 1024) {
910 comp_block_size = 1024;
911 } else if (comp_size <= 2048) {
912 comp_block_size = 2048;
913 } else if (comp_size <= 4096) {
914 comp_block_size = 4096;
915 } else if (comp_size <= 8192) {
916 comp_block_size = 8192;
917 } else if (comp_size <= 16384) {
918 comp_block_size = 16384;
919 } else if (comp_size <= 32768) {
920 comp_block_size = 32768;
922 panic("hammer2: WRITE PATH: "
923 "Weird comp_size value.");
925 comp_block_size = pblksize;
929 * Must zero the remainder or dedup (which operates on a
930 * physical block basis) will not find matches.
932 if (comp_size < comp_block_size) {
933 bzero(comp_buffer + comp_size,
934 comp_block_size - comp_size);
939 * Assign physical storage, data will be set to NULL if a live-dedup
942 data = comp_size ? comp_buffer : bp->b_data;
943 chain = hammer2_assign_physical(ip, parentp, lbase, comp_block_size,
944 mtid, &data, errorp);
947 kprintf("WRITE PATH: An error occurred while "
948 "assigning physical space.\n");
949 KKASSERT(chain == NULL);
953 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
954 hammer2_inode_data_t *wipdata;
956 hammer2_chain_modify_ip(ip, chain, mtid, 0);
957 wipdata = &chain->data->ipdata;
958 KKASSERT(wipdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA);
959 KKASSERT(bp->b_loffset == 0);
960 bcopy(bp->b_data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
961 ++hammer2_iod_file_wembed;
962 } else if (data == NULL) {
964 * Live deduplication, a copy of the data is already present
970 chain->bref.methods =
971 HAMMER2_ENC_COMP(comp_algo) +
972 HAMMER2_ENC_CHECK(check_algo);
974 chain->bref.methods =
977 HAMMER2_ENC_CHECK(check_algo);
979 bdata = comp_size ? comp_buffer : bp->b_data;
980 hammer2_chain_setcheck(chain, bdata);
981 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
986 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
988 switch(chain->bref.type) {
989 case HAMMER2_BREF_TYPE_INODE:
990 panic("hammer2_write_bp: unexpected inode\n");
992 case HAMMER2_BREF_TYPE_DATA:
994 * Optimize out the read-before-write
997 *errorp = hammer2_io_newnz(chain->hmp,
999 chain->bref.data_off,
1003 hammer2_io_brelse(&dio);
1004 kprintf("hammer2: WRITE PATH: "
1005 "dbp bread error\n");
1008 bdata = hammer2_io_data(dio, chain->bref.data_off);
1011 * When loading the block make sure we don't
1012 * leave garbage after the compressed data.
1015 chain->bref.methods =
1016 HAMMER2_ENC_COMP(comp_algo) +
1017 HAMMER2_ENC_CHECK(check_algo);
1018 bcopy(comp_buffer, bdata, comp_size);
1020 chain->bref.methods =
1022 HAMMER2_COMP_NONE) +
1023 HAMMER2_ENC_CHECK(check_algo);
1024 bcopy(bp->b_data, bdata, pblksize);
1028 * The flush code doesn't calculate check codes for
1029 * file data (doing so can result in excessive I/O),
1032 * Record for dedup only after the DIO's buffer cache
1033 * buffer has been updated.
1035 hammer2_chain_setcheck(chain, bdata);
1036 hammer2_dedup_record(chain, bdata);
1039 * Device buffer is now valid, chain is no longer in
1040 * the initial state.
1042 * (No blockref table worries with file data)
1044 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1046 /* Now write the related bdp. */
1047 if (ioflag & IO_SYNC) {
1049 * Synchronous I/O requested.
1051 hammer2_io_bwrite(&dio);
1053 } else if ((ioflag & IO_DIRECT) &&
1054 loff + n == pblksize) {
1055 hammer2_io_bdwrite(&dio);
1057 } else if (ioflag & IO_ASYNC) {
1058 hammer2_io_bawrite(&dio);
1060 hammer2_io_bdwrite(&dio);
1064 panic("hammer2_write_bp: bad chain type %d\n",
1072 hammer2_chain_unlock(chain);
1073 hammer2_chain_drop(chain);
1076 objcache_put(cache_buffer_write, comp_buffer);
1082 * Function that performs zero-checking and writing without compression,
1083 * it corresponds to default zero-checking path.
1087 hammer2_zero_check_and_write(struct buf *bp, hammer2_inode_t *ip,
1088 hammer2_chain_t **parentp,
1089 hammer2_key_t lbase, int ioflag, int pblksize,
1090 hammer2_tid_t mtid, int *errorp,
1093 hammer2_chain_t *chain;
1094 char *data = bp->b_data;
1096 if (test_block_zeros(bp->b_data, pblksize)) {
1097 zero_write(bp, ip, parentp, lbase, mtid, errorp);
1099 chain = hammer2_assign_physical(ip, parentp, lbase, pblksize,
1100 mtid, &data, errorp);
1102 hammer2_write_bp(chain, bp, ioflag, pblksize,
1103 mtid, errorp, check_algo);
1104 } /* else dedup occurred */
1106 hammer2_chain_unlock(chain);
1107 hammer2_chain_drop(chain);
1115 * A function to test whether a block of data contains only zeros,
1116 * returns TRUE (non-zero) if the block is all zeros.
1120 test_block_zeros(const char *buf, size_t bytes)
1124 for (i = 0; i < bytes; i += sizeof(long)) {
1125 if (*(const long *)(buf + i) != 0)
1134 * Function to "write" a block that contains only zeros.
1138 zero_write(struct buf *bp, hammer2_inode_t *ip,
1139 hammer2_chain_t **parentp,
1140 hammer2_key_t lbase, hammer2_tid_t mtid, int *errorp)
1142 hammer2_chain_t *chain;
1143 hammer2_key_t key_dummy;
1144 int cache_index = -1;
1147 chain = hammer2_chain_lookup(parentp, &key_dummy,
1150 HAMMER2_LOOKUP_NODATA);
1152 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
1153 hammer2_inode_data_t *wipdata;
1155 hammer2_chain_modify_ip(ip, chain, mtid, 0);
1156 wipdata = &chain->data->ipdata;
1157 KKASSERT(wipdata->meta.op_flags &
1158 HAMMER2_OPFLAG_DIRECTDATA);
1159 KKASSERT(bp->b_loffset == 0);
1160 bzero(wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
1161 ++hammer2_iod_file_wembed;
1163 hammer2_chain_delete(*parentp, chain,
1164 mtid, HAMMER2_DELETE_PERMANENT);
1165 ++hammer2_iod_file_wzero;
1167 hammer2_chain_unlock(chain);
1168 hammer2_chain_drop(chain);
1170 ++hammer2_iod_file_wzero;
1177 * Function to write the data as it is, without performing any sort of
1178 * compression. This function is used in path without compression and
1179 * default zero-checking path.
1183 hammer2_write_bp(hammer2_chain_t *chain, struct buf *bp, int ioflag,
1185 hammer2_tid_t mtid, int *errorp, int check_algo)
1187 hammer2_inode_data_t *wipdata;
1192 error = 0; /* XXX TODO below */
1194 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
1196 switch(chain->bref.type) {
1197 case HAMMER2_BREF_TYPE_INODE:
1198 wipdata = &chain->data->ipdata;
1199 KKASSERT(wipdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA);
1200 KKASSERT(bp->b_loffset == 0);
1201 bcopy(bp->b_data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
1203 ++hammer2_iod_file_wembed;
1205 case HAMMER2_BREF_TYPE_DATA:
1206 error = hammer2_io_newnz(chain->hmp,
1208 chain->bref.data_off,
1209 chain->bytes, &dio);
1211 hammer2_io_bqrelse(&dio);
1212 kprintf("hammer2: WRITE PATH: "
1213 "dbp bread error\n");
1216 bdata = hammer2_io_data(dio, chain->bref.data_off);
1218 chain->bref.methods = HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) +
1219 HAMMER2_ENC_CHECK(check_algo);
1220 bcopy(bp->b_data, bdata, chain->bytes);
1223 * The flush code doesn't calculate check codes for
1224 * file data (doing so can result in excessive I/O),
1227 * Record for dedup only after the DIO's buffer cache
1228 * buffer has been updated.
1230 hammer2_chain_setcheck(chain, bdata);
1231 hammer2_dedup_record(chain, bdata);
1234 * Device buffer is now valid, chain is no longer in
1235 * the initial state.
1237 * (No blockref table worries with file data)
1239 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1241 if (ioflag & IO_SYNC) {
1243 * Synchronous I/O requested.
1245 hammer2_io_bwrite(&dio);
1247 } else if ((ioflag & IO_DIRECT) &&
1248 loff + n == pblksize) {
1249 hammer2_io_bdwrite(&dio);
1251 } else if (ioflag & IO_ASYNC) {
1252 hammer2_io_bawrite(&dio);
1254 hammer2_io_bdwrite(&dio);
1258 panic("hammer2_write_bp: bad chain type %d\n",
1264 KKASSERT(error == 0); /* XXX TODO */
1269 * LIVE DEDUP HEURISTIC
1271 * WARNING! This code is SMP safe but the heuristic allows SMP collisions.
1272 * All fields must be loaded into locals and validated.
1276 hammer2_dedup_record(hammer2_chain_t *chain, char *data)
1279 hammer2_dedup_t *dedup;
1287 switch(HAMMER2_DEC_CHECK(chain->bref.methods)) {
1288 case HAMMER2_CHECK_ISCSI32:
1290 * XXX use the built-in crc (the dedup lookup sequencing
1291 * needs to be fixed so the check code is already present
1292 * when dedup_lookup is called)
1295 crc = (uint64_t)(uint32_t)chain->bref.check.iscsi32.value;
1297 crc = XXH64(data, chain->bytes, XXH_HAMMER2_SEED);
1299 case HAMMER2_CHECK_XXHASH64:
1300 crc = chain->bref.check.xxhash64.value;
1302 case HAMMER2_CHECK_SHA192:
1304 * XXX use the built-in crc (the dedup lookup sequencing
1305 * needs to be fixed so the check code is already present
1306 * when dedup_lookup is called)
1309 crc = ((uint64_t *)chain->bref.check.sha192.data)[0] ^
1310 ((uint64_t *)chain->bref.check.sha192.data)[1] ^
1311 ((uint64_t *)chain->bref.check.sha192.data)[2];
1313 crc = XXH64(data, chain->bytes, XXH_HAMMER2_SEED);
1317 * Cannot dedup without a check code
1321 dedup = &hmp->heur_dedup[crc & (HAMMER2_DEDUP_HEUR_MASK & ~3)];
1322 for (i = 0; i < 4; ++i) {
1323 if (dedup[i].data_crc == crc) {
1327 dticks = (int)(dedup[i].ticks - dedup[best].ticks);
1328 if (dticks < 0 || dticks > hz * 60 * 30)
1332 if (hammer2_debug & 0x40000) {
1333 kprintf("REC %04x %016jx %016jx\n",
1334 (int)(dedup - hmp->heur_dedup),
1336 chain->bref.data_off);
1338 dedup->ticks = ticks;
1339 dedup->data_off = chain->bref.data_off;
1340 dedup->data_crc = crc;
1341 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DEDUP);
1346 hammer2_dedup_lookup(hammer2_dev_t *hmp, char **datap, int pblksize)
1348 hammer2_dedup_t *dedup;
1360 * XXX use the built-in crc (the dedup lookup sequencing
1361 * needs to be fixed so the check code is already present
1362 * when dedup_lookup is called)
1364 crc = XXH64(data, pblksize, XXH_HAMMER2_SEED);
1365 dedup = &hmp->heur_dedup[crc & (HAMMER2_DEDUP_HEUR_MASK & ~3)];
1367 if (hammer2_debug & 0x40000) {
1368 kprintf("LOC %04x/4 %016jx\n",
1369 (int)(dedup - hmp->heur_dedup),
1373 for (i = 0; i < 4; ++i) {
1374 off = dedup[i].data_off;
1376 if (dedup[i].data_crc != crc)
1378 if ((1 << (int)(off & HAMMER2_OFF_MASK_RADIX)) != pblksize)
1380 dio = hammer2_io_getquick(hmp, off, pblksize);
1382 bcmp(data, hammer2_io_data(dio, off), pblksize) == 0) {
1383 if (hammer2_debug & 0x40000) {
1384 kprintf("DEDUP SUCCESS %016jx\n",
1387 hammer2_io_putblk(&dio);
1389 dedup[i].ticks = ticks; /* update use */
1390 ++hammer2_iod_file_wdedup;
1391 return off; /* RETURN */
1394 hammer2_io_putblk(&dio);
1400 * Poof. Races are ok, if someone gets in and reuses a dedup offset
1401 * before or while we are clearing it they will also recover the freemap
1402 * entry (set it to fully allocated), so a bulkfree race can only set it
1403 * to a possibly-free state.
1405 * XXX ok, well, not really sure races are ok but going to run with it
1409 hammer2_dedup_clear(hammer2_dev_t *hmp)
1413 for (i = 0; i < HAMMER2_DEDUP_HEUR_SIZE; ++i) {
1414 hmp->heur_dedup[i].data_off = 0;
1415 hmp->heur_dedup[i].ticks = ticks - 1;