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 * H2 double-buffers the data, setting B_NOTMETA on the logical
372 * buffer hints to the OS that the logical buffer should not be
373 * swapcached (since the device buffer can be).
375 * Also note that even for compressed data we would rather the
376 * kernel cache/swapcache device buffers more and (decompressed)
377 * logical buffers less, since that will significantly improve
378 * the amount of end-user data that can be cached.
380 error = hammer2_xop_collect(&xop->head, HAMMER2_XOP_COLLECT_NOWAIT);
386 hammer2_mtx_unlock(&xop->lock);
387 bp->b_flags |= B_NOTMETA;
388 chain = xop->head.cluster.focus;
389 hammer2_strategy_read_completion(chain, (char *)chain->data,
392 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
396 hammer2_mtx_unlock(&xop->lock);
397 bp->b_flags |= B_NOTMETA;
400 bzero(bp->b_data, bp->b_bcount);
402 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
405 hammer2_mtx_unlock(&xop->lock);
408 kprintf("strategy_xop_read: error %d loff=%016jx\n",
409 error, bp->b_loffset);
411 hammer2_mtx_unlock(&xop->lock);
412 bp->b_flags |= B_ERROR;
415 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
423 hammer2_strategy_read_completion(hammer2_chain_t *chain, char *data,
426 struct buf *bp = bio->bio_buf;
428 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
430 * Data is embedded in the inode (copy from inode).
432 bcopy(((hammer2_inode_data_t *)data)->u.data,
433 bp->b_data, HAMMER2_EMBEDDED_BYTES);
434 bzero(bp->b_data + HAMMER2_EMBEDDED_BYTES,
435 bp->b_bcount - HAMMER2_EMBEDDED_BYTES);
438 } else if (chain->bref.type == HAMMER2_BREF_TYPE_DATA) {
440 * Data is on-media, record for live dedup. Release the
441 * chain (try to free it) when done. The data is still
442 * cached by both the buffer cache in front and the
443 * block device behind us. This leaves more room in the
444 * LRU chain cache for meta-data chains which we really
447 hammer2_dedup_record(chain, data);
448 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
451 * Decompression and copy.
453 switch (HAMMER2_DEC_COMP(chain->bref.methods)) {
454 case HAMMER2_COMP_LZ4:
455 hammer2_decompress_LZ4_callback(data, chain->bytes,
457 /* b_resid set by call */
459 case HAMMER2_COMP_ZLIB:
460 hammer2_decompress_ZLIB_callback(data, chain->bytes,
462 /* b_resid set by call */
464 case HAMMER2_COMP_NONE:
465 KKASSERT(chain->bytes <= bp->b_bcount);
466 bcopy(data, bp->b_data, chain->bytes);
467 if (chain->bytes < bp->b_bcount) {
468 bzero(bp->b_data + chain->bytes,
469 bp->b_bcount - chain->bytes);
475 panic("hammer2_strategy_read: "
476 "unknown compression type");
479 panic("hammer2_strategy_read: unknown bref type");
483 /****************************************************************************
485 ****************************************************************************/
488 * Functions for compression in threads,
489 * from hammer2_vnops.c
491 static void hammer2_write_file_core(struct buf *bp, hammer2_inode_t *ip,
492 hammer2_chain_t **parentp,
493 hammer2_key_t lbase, int ioflag, int pblksize,
494 hammer2_tid_t mtid, int *errorp);
495 static void hammer2_compress_and_write(struct buf *bp, hammer2_inode_t *ip,
496 hammer2_chain_t **parentp,
497 hammer2_key_t lbase, int ioflag, int pblksize,
498 hammer2_tid_t mtid, int *errorp,
499 int comp_algo, int check_algo);
500 static void hammer2_zero_check_and_write(struct buf *bp, hammer2_inode_t *ip,
501 hammer2_chain_t **parentp,
502 hammer2_key_t lbase, int ioflag, int pblksize,
503 hammer2_tid_t mtid, int *errorp,
505 static int test_block_zeros(const char *buf, size_t bytes);
506 static void zero_write(struct buf *bp, hammer2_inode_t *ip,
507 hammer2_chain_t **parentp,
509 hammer2_tid_t mtid, int *errorp);
510 static void hammer2_write_bp(hammer2_chain_t *chain, struct buf *bp,
511 int ioflag, int pblksize,
512 hammer2_tid_t mtid, int *errorp,
517 hammer2_strategy_write(struct vop_strategy_args *ap)
519 hammer2_xop_strategy_t *xop;
530 hammer2_lwinprog_ref(pmp);
531 hammer2_trans_assert_strategy(pmp);
533 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING);
536 xop->lbase = bio->bio_offset;
537 hammer2_mtx_init(&xop->lock, "h2biow");
538 hammer2_xop_start(&xop->head, hammer2_strategy_xop_write);
539 /* asynchronous completion */
541 hammer2_lwinprog_wait(pmp, hammer2_flush_pipe);
547 * Per-node XOP (threaded). Write the logical buffer to the media.
551 hammer2_strategy_xop_write(hammer2_xop_t *arg, int clindex)
553 hammer2_xop_strategy_t *xop = &arg->xop_strategy;
554 hammer2_chain_t *parent;
568 /* hammer2_trans_init(parent->hmp->spmp, HAMMER2_TRANS_BUFCACHE); */
570 lblksize = hammer2_calc_logical(ip, bio->bio_offset, &lbase, NULL);
571 pblksize = hammer2_calc_physical(ip, lbase);
572 parent = hammer2_inode_chain(ip, clindex, HAMMER2_RESOLVE_ALWAYS);
573 hammer2_write_file_core(bp, ip, &parent,
574 lbase, IO_ASYNC, pblksize,
575 xop->head.mtid, &error);
577 hammer2_chain_unlock(parent);
578 hammer2_chain_drop(parent);
579 parent = NULL; /* safety */
581 hammer2_xop_feed(&xop->head, NULL, clindex, error);
584 * Race to finish the frontend
588 hammer2_mtx_ex(&xop->lock);
590 hammer2_mtx_unlock(&xop->lock);
595 * Async operation has not completed and we now own the lock.
596 * Determine if we can complete the operation by issuing the
597 * frontend collection non-blocking.
599 * H2 double-buffers the data, setting B_NOTMETA on the logical
600 * buffer hints to the OS that the logical buffer should not be
601 * swapcached (since the device buffer can be).
603 error = hammer2_xop_collect(&xop->head, HAMMER2_XOP_COLLECT_NOWAIT);
609 hammer2_mtx_unlock(&xop->lock);
610 bp->b_flags |= B_NOTMETA;
614 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
615 hammer2_lwinprog_drop(ip->pmp);
618 hammer2_mtx_unlock(&xop->lock);
621 kprintf("strategy_xop_write: error %d loff=%016jx\n",
622 error, bp->b_loffset);
624 hammer2_mtx_unlock(&xop->lock);
625 bp->b_flags |= B_ERROR;
628 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
629 hammer2_lwinprog_drop(ip->pmp);
635 * Wait for pending I/O to complete
638 hammer2_bioq_sync(hammer2_pfs_t *pmp)
640 hammer2_lwinprog_wait(pmp, 0);
644 * Create a new cluster at (cparent, lbase) and assign physical storage,
645 * returning a cluster suitable for I/O. The cluster will be in a modified
648 * cparent can wind up being anything.
650 * If datap is not NULL, *datap points to the real data we intend to write.
651 * If we can dedup the storage location we set *datap to NULL to indicate
652 * to the caller that a dedup occurred.
654 * NOTE: Special case for data embedded in inode.
658 hammer2_assign_physical(hammer2_inode_t *ip, hammer2_chain_t **parentp,
659 hammer2_key_t lbase, int pblksize,
660 hammer2_tid_t mtid, char **datap, int *errorp)
662 hammer2_chain_t *chain;
663 hammer2_key_t key_dummy;
664 hammer2_off_t dedup_off;
665 int pradix = hammer2_getradix(pblksize);
666 int cache_index = -1;
669 * Locate the chain associated with lbase, return a locked chain.
670 * However, do not instantiate any data reference (which utilizes a
671 * device buffer) because we will be using direct IO via the
672 * logical buffer cache buffer.
675 KKASSERT(pblksize >= HAMMER2_ALLOC_MIN);
678 chain = hammer2_chain_lookup(parentp, &key_dummy,
681 HAMMER2_LOOKUP_NODATA);
683 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED))
684 kprintf("assign physical deleted chain @ %016jx\n",
689 * We found a hole, create a new chain entry.
691 * NOTE: DATA chains are created without device backing
692 * store (nor do we want any).
694 dedup_off = hammer2_dedup_lookup((*parentp)->hmp, datap,
696 *errorp = hammer2_chain_create(parentp, &chain, ip->pmp,
697 lbase, HAMMER2_PBUFRADIX,
698 HAMMER2_BREF_TYPE_DATA,
702 panic("hammer2_chain_create: par=%p error=%d\n",
706 /*ip->delta_dcount += pblksize;*/
708 switch (chain->bref.type) {
709 case HAMMER2_BREF_TYPE_INODE:
711 * The data is embedded in the inode, which requires
714 hammer2_chain_modify_ip(ip, chain, mtid, 0);
716 case HAMMER2_BREF_TYPE_DATA:
717 dedup_off = hammer2_dedup_lookup(chain->hmp, datap,
719 if (chain->bytes != pblksize) {
720 hammer2_chain_resize(ip, *parentp, chain,
723 HAMMER2_MODIFY_OPTDATA);
727 * DATA buffers must be marked modified whether the
728 * data is in a logical buffer or not. We also have
729 * to make this call to fixup the chain data pointers
730 * after resizing in case this is an encrypted or
733 hammer2_chain_modify(chain, mtid, dedup_off,
734 HAMMER2_MODIFY_OPTDATA);
737 panic("hammer2_assign_physical: bad type");
747 * hammer2_write_file_core() - hammer2_write_thread() helper
749 * The core write function which determines which path to take
750 * depending on compression settings. We also have to locate the
751 * related chains so we can calculate and set the check data for
756 hammer2_write_file_core(struct buf *bp, hammer2_inode_t *ip,
757 hammer2_chain_t **parentp,
758 hammer2_key_t lbase, int ioflag, int pblksize,
759 hammer2_tid_t mtid, int *errorp)
761 hammer2_chain_t *chain;
762 char *data = bp->b_data;
766 switch(HAMMER2_DEC_ALGO(ip->meta.comp_algo)) {
767 case HAMMER2_COMP_NONE:
769 * We have to assign physical storage to the buffer
770 * we intend to dirty or write now to avoid deadlocks
771 * in the strategy code later.
773 * This can return NOOFFSET for inode-embedded data.
774 * The strategy code will take care of it in that case.
776 chain = hammer2_assign_physical(ip, parentp, lbase, pblksize,
777 mtid, &data, errorp);
778 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
779 hammer2_inode_data_t *wipdata;
781 wipdata = &chain->data->ipdata;
782 KKASSERT(wipdata->meta.op_flags &
783 HAMMER2_OPFLAG_DIRECTDATA);
784 KKASSERT(bp->b_loffset == 0);
785 bcopy(bp->b_data, wipdata->u.data,
786 HAMMER2_EMBEDDED_BYTES);
787 ++hammer2_iod_file_wembed;
788 } else if (data == NULL) {
790 * Copy of data already present on-media.
792 chain->bref.methods =
793 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) +
794 HAMMER2_ENC_CHECK(ip->meta.check_algo);
795 hammer2_chain_setcheck(chain, bp->b_data);
797 hammer2_write_bp(chain, bp, ioflag, pblksize,
798 mtid, errorp, ip->meta.check_algo);
801 hammer2_chain_unlock(chain);
802 hammer2_chain_drop(chain);
805 case HAMMER2_COMP_AUTOZERO:
807 * Check for zero-fill only
809 hammer2_zero_check_and_write(bp, ip, parentp,
810 lbase, ioflag, pblksize,
812 ip->meta.check_algo);
814 case HAMMER2_COMP_LZ4:
815 case HAMMER2_COMP_ZLIB:
818 * Check for zero-fill and attempt compression.
820 hammer2_compress_and_write(bp, ip, parentp,
821 lbase, ioflag, pblksize,
824 ip->meta.check_algo);
832 * Generic function that will perform the compression in compression
833 * write path. The compression algorithm is determined by the settings
834 * obtained from inode.
838 hammer2_compress_and_write(struct buf *bp, hammer2_inode_t *ip,
839 hammer2_chain_t **parentp,
840 hammer2_key_t lbase, int ioflag, int pblksize,
841 hammer2_tid_t mtid, int *errorp, int comp_algo, int check_algo)
843 hammer2_chain_t *chain;
849 if (test_block_zeros(bp->b_data, pblksize)) {
850 zero_write(bp, ip, parentp, lbase, mtid, errorp);
857 KKASSERT(pblksize / 2 <= 32768);
859 if (ip->comp_heuristic < 8 || (ip->comp_heuristic & 7) == 0) {
860 z_stream strm_compress;
864 switch(HAMMER2_DEC_ALGO(comp_algo)) {
865 case HAMMER2_COMP_LZ4:
866 comp_buffer = objcache_get(cache_buffer_write,
868 comp_size = LZ4_compress_limitedOutput(
870 &comp_buffer[sizeof(int)],
872 pblksize / 2 - sizeof(int));
874 * We need to prefix with the size, LZ4
875 * doesn't do it for us. Add the related
878 *(int *)comp_buffer = comp_size;
880 comp_size += sizeof(int);
882 case HAMMER2_COMP_ZLIB:
883 comp_level = HAMMER2_DEC_LEVEL(comp_algo);
885 comp_level = 6; /* default zlib compression */
886 else if (comp_level < 6)
888 else if (comp_level > 9)
890 ret = deflateInit(&strm_compress, comp_level);
892 kprintf("HAMMER2 ZLIB: fatal error "
893 "on deflateInit.\n");
896 comp_buffer = objcache_get(cache_buffer_write,
898 strm_compress.next_in = bp->b_data;
899 strm_compress.avail_in = pblksize;
900 strm_compress.next_out = comp_buffer;
901 strm_compress.avail_out = pblksize / 2;
902 ret = deflate(&strm_compress, Z_FINISH);
903 if (ret == Z_STREAM_END) {
904 comp_size = pblksize / 2 -
905 strm_compress.avail_out;
909 ret = deflateEnd(&strm_compress);
912 kprintf("Error: Unknown compression method.\n");
913 kprintf("Comp_method = %d.\n", comp_algo);
918 if (comp_size == 0) {
920 * compression failed or turned off
922 comp_block_size = pblksize; /* safety */
923 if (++ip->comp_heuristic > 128)
924 ip->comp_heuristic = 8;
927 * compression succeeded
929 ip->comp_heuristic = 0;
930 if (comp_size <= 1024) {
931 comp_block_size = 1024;
932 } else if (comp_size <= 2048) {
933 comp_block_size = 2048;
934 } else if (comp_size <= 4096) {
935 comp_block_size = 4096;
936 } else if (comp_size <= 8192) {
937 comp_block_size = 8192;
938 } else if (comp_size <= 16384) {
939 comp_block_size = 16384;
940 } else if (comp_size <= 32768) {
941 comp_block_size = 32768;
943 panic("hammer2: WRITE PATH: "
944 "Weird comp_size value.");
946 comp_block_size = pblksize;
950 * Must zero the remainder or dedup (which operates on a
951 * physical block basis) will not find matches.
953 if (comp_size < comp_block_size) {
954 bzero(comp_buffer + comp_size,
955 comp_block_size - comp_size);
960 * Assign physical storage, data will be set to NULL if a live-dedup
963 data = comp_size ? comp_buffer : bp->b_data;
964 chain = hammer2_assign_physical(ip, parentp, lbase, comp_block_size,
965 mtid, &data, errorp);
968 kprintf("WRITE PATH: An error occurred while "
969 "assigning physical space.\n");
970 KKASSERT(chain == NULL);
974 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
975 hammer2_inode_data_t *wipdata;
977 hammer2_chain_modify_ip(ip, chain, mtid, 0);
978 wipdata = &chain->data->ipdata;
979 KKASSERT(wipdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA);
980 KKASSERT(bp->b_loffset == 0);
981 bcopy(bp->b_data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
982 ++hammer2_iod_file_wembed;
983 } else if (data == NULL) {
985 * Live deduplication, a copy of the data is already present
991 chain->bref.methods =
992 HAMMER2_ENC_COMP(comp_algo) +
993 HAMMER2_ENC_CHECK(check_algo);
995 chain->bref.methods =
998 HAMMER2_ENC_CHECK(check_algo);
1000 bdata = comp_size ? comp_buffer : bp->b_data;
1001 hammer2_chain_setcheck(chain, bdata);
1002 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1007 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
1009 switch(chain->bref.type) {
1010 case HAMMER2_BREF_TYPE_INODE:
1011 panic("hammer2_write_bp: unexpected inode\n");
1013 case HAMMER2_BREF_TYPE_DATA:
1015 * Optimize out the read-before-write
1018 *errorp = hammer2_io_newnz(chain->hmp,
1020 chain->bref.data_off,
1024 hammer2_io_brelse(&dio);
1025 kprintf("hammer2: WRITE PATH: "
1026 "dbp bread error\n");
1029 bdata = hammer2_io_data(dio, chain->bref.data_off);
1032 * When loading the block make sure we don't
1033 * leave garbage after the compressed data.
1036 chain->bref.methods =
1037 HAMMER2_ENC_COMP(comp_algo) +
1038 HAMMER2_ENC_CHECK(check_algo);
1039 bcopy(comp_buffer, bdata, comp_size);
1041 chain->bref.methods =
1043 HAMMER2_COMP_NONE) +
1044 HAMMER2_ENC_CHECK(check_algo);
1045 bcopy(bp->b_data, bdata, pblksize);
1049 * The flush code doesn't calculate check codes for
1050 * file data (doing so can result in excessive I/O),
1053 * Record for dedup only after the DIO's buffer cache
1054 * buffer has been updated.
1056 hammer2_chain_setcheck(chain, bdata);
1057 hammer2_dedup_record(chain, bdata);
1060 * Device buffer is now valid, chain is no longer in
1061 * the initial state.
1063 * (No blockref table worries with file data)
1065 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1067 /* Now write the related bdp. */
1068 if (ioflag & IO_SYNC) {
1070 * Synchronous I/O requested.
1072 hammer2_io_bwrite(&dio);
1074 } else if ((ioflag & IO_DIRECT) &&
1075 loff + n == pblksize) {
1076 hammer2_io_bdwrite(&dio);
1078 } else if (ioflag & IO_ASYNC) {
1079 hammer2_io_bawrite(&dio);
1081 hammer2_io_bdwrite(&dio);
1085 panic("hammer2_write_bp: bad chain type %d\n",
1093 hammer2_chain_unlock(chain);
1094 hammer2_chain_drop(chain);
1097 objcache_put(cache_buffer_write, comp_buffer);
1103 * Function that performs zero-checking and writing without compression,
1104 * it corresponds to default zero-checking path.
1108 hammer2_zero_check_and_write(struct buf *bp, hammer2_inode_t *ip,
1109 hammer2_chain_t **parentp,
1110 hammer2_key_t lbase, int ioflag, int pblksize,
1111 hammer2_tid_t mtid, int *errorp,
1114 hammer2_chain_t *chain;
1115 char *data = bp->b_data;
1117 if (test_block_zeros(bp->b_data, pblksize)) {
1118 zero_write(bp, ip, parentp, lbase, mtid, errorp);
1120 chain = hammer2_assign_physical(ip, parentp, lbase, pblksize,
1121 mtid, &data, errorp);
1123 hammer2_write_bp(chain, bp, ioflag, pblksize,
1124 mtid, errorp, check_algo);
1125 } /* else dedup occurred */
1127 hammer2_chain_unlock(chain);
1128 hammer2_chain_drop(chain);
1136 * A function to test whether a block of data contains only zeros,
1137 * returns TRUE (non-zero) if the block is all zeros.
1141 test_block_zeros(const char *buf, size_t bytes)
1145 for (i = 0; i < bytes; i += sizeof(long)) {
1146 if (*(const long *)(buf + i) != 0)
1155 * Function to "write" a block that contains only zeros.
1159 zero_write(struct buf *bp, hammer2_inode_t *ip,
1160 hammer2_chain_t **parentp,
1161 hammer2_key_t lbase, hammer2_tid_t mtid, int *errorp)
1163 hammer2_chain_t *chain;
1164 hammer2_key_t key_dummy;
1165 int cache_index = -1;
1168 chain = hammer2_chain_lookup(parentp, &key_dummy,
1171 HAMMER2_LOOKUP_NODATA);
1173 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
1174 hammer2_inode_data_t *wipdata;
1176 hammer2_chain_modify_ip(ip, chain, mtid, 0);
1177 wipdata = &chain->data->ipdata;
1178 KKASSERT(wipdata->meta.op_flags &
1179 HAMMER2_OPFLAG_DIRECTDATA);
1180 KKASSERT(bp->b_loffset == 0);
1181 bzero(wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
1182 ++hammer2_iod_file_wembed;
1184 hammer2_chain_delete(*parentp, chain,
1185 mtid, HAMMER2_DELETE_PERMANENT);
1186 ++hammer2_iod_file_wzero;
1188 hammer2_chain_unlock(chain);
1189 hammer2_chain_drop(chain);
1191 ++hammer2_iod_file_wzero;
1198 * Function to write the data as it is, without performing any sort of
1199 * compression. This function is used in path without compression and
1200 * default zero-checking path.
1204 hammer2_write_bp(hammer2_chain_t *chain, struct buf *bp, int ioflag,
1206 hammer2_tid_t mtid, int *errorp, int check_algo)
1208 hammer2_inode_data_t *wipdata;
1213 error = 0; /* XXX TODO below */
1215 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
1217 switch(chain->bref.type) {
1218 case HAMMER2_BREF_TYPE_INODE:
1219 wipdata = &chain->data->ipdata;
1220 KKASSERT(wipdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA);
1221 KKASSERT(bp->b_loffset == 0);
1222 bcopy(bp->b_data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
1224 ++hammer2_iod_file_wembed;
1226 case HAMMER2_BREF_TYPE_DATA:
1227 error = hammer2_io_newnz(chain->hmp,
1229 chain->bref.data_off,
1230 chain->bytes, &dio);
1232 hammer2_io_bqrelse(&dio);
1233 kprintf("hammer2: WRITE PATH: "
1234 "dbp bread error\n");
1237 bdata = hammer2_io_data(dio, chain->bref.data_off);
1239 chain->bref.methods = HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) +
1240 HAMMER2_ENC_CHECK(check_algo);
1241 bcopy(bp->b_data, bdata, chain->bytes);
1244 * The flush code doesn't calculate check codes for
1245 * file data (doing so can result in excessive I/O),
1248 * Record for dedup only after the DIO's buffer cache
1249 * buffer has been updated.
1251 hammer2_chain_setcheck(chain, bdata);
1252 hammer2_dedup_record(chain, bdata);
1255 * Device buffer is now valid, chain is no longer in
1256 * the initial state.
1258 * (No blockref table worries with file data)
1260 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1262 if (ioflag & IO_SYNC) {
1264 * Synchronous I/O requested.
1266 hammer2_io_bwrite(&dio);
1268 } else if ((ioflag & IO_DIRECT) &&
1269 loff + n == pblksize) {
1270 hammer2_io_bdwrite(&dio);
1272 } else if (ioflag & IO_ASYNC) {
1273 hammer2_io_bawrite(&dio);
1275 hammer2_io_bdwrite(&dio);
1279 panic("hammer2_write_bp: bad chain type %d\n",
1285 KKASSERT(error == 0); /* XXX TODO */
1290 * LIVE DEDUP HEURISTIC
1292 * WARNING! This code is SMP safe but the heuristic allows SMP collisions.
1293 * All fields must be loaded into locals and validated.
1295 * WARNING! Should only be used for file data, hammer2_chain_modify() only
1296 * checks for the dedup case on data chains.
1300 hammer2_dedup_record(hammer2_chain_t *chain, char *data)
1303 hammer2_dedup_t *dedup;
1311 switch(HAMMER2_DEC_CHECK(chain->bref.methods)) {
1312 case HAMMER2_CHECK_ISCSI32:
1314 * XXX use the built-in crc (the dedup lookup sequencing
1315 * needs to be fixed so the check code is already present
1316 * when dedup_lookup is called)
1319 crc = (uint64_t)(uint32_t)chain->bref.check.iscsi32.value;
1321 crc = XXH64(data, chain->bytes, XXH_HAMMER2_SEED);
1323 case HAMMER2_CHECK_XXHASH64:
1324 crc = chain->bref.check.xxhash64.value;
1326 case HAMMER2_CHECK_SHA192:
1328 * XXX use the built-in crc (the dedup lookup sequencing
1329 * needs to be fixed so the check code is already present
1330 * when dedup_lookup is called)
1333 crc = ((uint64_t *)chain->bref.check.sha192.data)[0] ^
1334 ((uint64_t *)chain->bref.check.sha192.data)[1] ^
1335 ((uint64_t *)chain->bref.check.sha192.data)[2];
1337 crc = XXH64(data, chain->bytes, XXH_HAMMER2_SEED);
1341 * Cannot dedup without a check code
1345 dedup = &hmp->heur_dedup[crc & (HAMMER2_DEDUP_HEUR_MASK & ~3)];
1346 for (i = 0; i < 4; ++i) {
1347 if (dedup[i].data_crc == crc) {
1351 dticks = (int)(dedup[i].ticks - dedup[best].ticks);
1352 if (dticks < 0 || dticks > hz * 60 * 30)
1356 if (hammer2_debug & 0x40000) {
1357 kprintf("REC %04x %016jx %016jx\n",
1358 (int)(dedup - hmp->heur_dedup),
1360 chain->bref.data_off);
1362 dedup->ticks = ticks;
1363 dedup->data_off = chain->bref.data_off;
1364 dedup->data_crc = crc;
1365 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DEDUP);
1370 hammer2_dedup_lookup(hammer2_dev_t *hmp, char **datap, int pblksize)
1372 hammer2_dedup_t *dedup;
1384 * XXX use the built-in crc (the dedup lookup sequencing
1385 * needs to be fixed so the check code is already present
1386 * when dedup_lookup is called)
1388 crc = XXH64(data, pblksize, XXH_HAMMER2_SEED);
1389 dedup = &hmp->heur_dedup[crc & (HAMMER2_DEDUP_HEUR_MASK & ~3)];
1391 if (hammer2_debug & 0x40000) {
1392 kprintf("LOC %04x/4 %016jx\n",
1393 (int)(dedup - hmp->heur_dedup),
1397 for (i = 0; i < 4; ++i) {
1398 off = dedup[i].data_off;
1400 if (dedup[i].data_crc != crc)
1402 if ((1 << (int)(off & HAMMER2_OFF_MASK_RADIX)) != pblksize)
1404 dio = hammer2_io_getquick(hmp, off, pblksize);
1406 bcmp(data, hammer2_io_data(dio, off), pblksize) == 0) {
1408 * Make sure the INVALOK flag is cleared to prevent
1409 * the possibly-dirty bp from being invalidated now
1410 * that we are using it as part of a de-dup operation.
1412 if (hammer2_debug & 0x40000) {
1413 kprintf("DEDUP SUCCESS %016jx\n",
1416 atomic_clear_64(&dio->refs, HAMMER2_DIO_INVALOK);
1417 hammer2_io_putblk(&dio);
1419 dedup[i].ticks = ticks; /* update use */
1420 ++hammer2_iod_file_wdedup;
1422 return off; /* RETURN */
1425 hammer2_io_putblk(&dio);
1431 * Poof. Races are ok, if someone gets in and reuses a dedup offset
1432 * before or while we are clearing it they will also recover the freemap
1433 * entry (set it to fully allocated), so a bulkfree race can only set it
1434 * to a possibly-free state.
1436 * XXX ok, well, not really sure races are ok but going to run with it
1440 hammer2_dedup_clear(hammer2_dev_t *hmp)
1444 for (i = 0; i < HAMMER2_DEDUP_HEUR_SIZE; ++i) {
1445 hmp->heur_dedup[i].data_off = 0;
1446 hmp->heur_dedup[i].ticks = ticks - 1;