2 * Copyright (c) 2011-2018 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@dragonflybsd.org>
6 * by Venkatesh Srinivas <vsrinivas@dragonflybsd.org>
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>
51 #include <sys/mount.h>
52 #include <sys/vnode.h>
53 #include <sys/objcache.h>
56 #include "hammer2_lz4.h"
58 #include "zlib/hammer2_zlib.h"
60 struct objcache *cache_buffer_read;
61 struct objcache *cache_buffer_write;
64 * Strategy code (async logical file buffer I/O from system)
66 * Except for the transaction init (which should normally not block),
67 * we essentially run the strategy operation asynchronously via a XOP.
69 * WARNING! The XOP deals with buffer synchronization. It is not synchronized
72 * XXX This isn't supposed to be able to deadlock against vfs_sync vfsync()
73 * calls but it has in the past when multiple flushes are queued.
75 * XXX We currently terminate the transaction once we get a quorum, otherwise
76 * the frontend can stall, but this can leave the remaining nodes with
77 * a potential flush conflict. We need to delay flushes on those nodes
78 * until running transactions complete separately from the normal
79 * transaction sequencing. FIXME TODO.
81 static int hammer2_strategy_read(struct vop_strategy_args *ap);
82 static int hammer2_strategy_write(struct vop_strategy_args *ap);
83 static void hammer2_strategy_read_completion(hammer2_chain_t *focus,
84 const char *data, struct bio *bio);
86 static hammer2_off_t hammer2_dedup_lookup(hammer2_dev_t *hmp,
87 char **datap, int pblksize);
90 hammer2_vop_strategy(struct vop_strategy_args *ap)
101 error = hammer2_strategy_read(ap);
104 error = hammer2_strategy_write(ap);
107 bp->b_error = error = EINVAL;
108 bp->b_flags |= B_ERROR;
116 * Return the largest contiguous physical disk range for the logical
119 * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
121 * Basically disabled, the logical buffer write thread has to deal with
122 * buffers one-at-a-time. Note that this should not prevent cluster_read()
123 * from reading-ahead, it simply prevents it from trying form a single
124 * cluster buffer for the logical request. H2 already uses 64KB buffers!
127 hammer2_vop_bmap(struct vop_bmap_args *ap)
129 *ap->a_doffsetp = NOOFFSET;
137 /****************************************************************************
139 ****************************************************************************/
141 * Callback used in read path in case that a block is compressed with LZ4.
145 hammer2_decompress_LZ4_callback(const char *data, u_int bytes, struct bio *bio)
148 char *compressed_buffer;
155 if bio->bio_caller_info2.index &&
156 bio->bio_caller_info1.uvalue32 !=
157 crc32(bp->b_data, bp->b_bufsize) --- return error
160 KKASSERT(bp->b_bufsize <= HAMMER2_PBUFSIZE);
161 compressed_size = *(const int *)data;
162 KKASSERT((uint32_t)compressed_size <= bytes - sizeof(int));
164 compressed_buffer = objcache_get(cache_buffer_read, M_INTWAIT);
165 result = LZ4_decompress_safe(__DECONST(char *, &data[sizeof(int)]),
170 kprintf("READ PATH: Error during decompression."
172 (intmax_t)bio->bio_offset, bytes);
173 /* make sure it isn't random garbage */
174 bzero(compressed_buffer, bp->b_bufsize);
176 KKASSERT(result <= bp->b_bufsize);
177 bcopy(compressed_buffer, bp->b_data, bp->b_bufsize);
178 if (result < bp->b_bufsize)
179 bzero(bp->b_data + result, bp->b_bufsize - result);
180 objcache_put(cache_buffer_read, compressed_buffer);
182 bp->b_flags |= B_AGE;
186 * Callback used in read path in case that a block is compressed with ZLIB.
187 * It is almost identical to LZ4 callback, so in theory they can be unified,
188 * but we didn't want to make changes in bio structure for that.
192 hammer2_decompress_ZLIB_callback(const char *data, u_int bytes, struct bio *bio)
195 char *compressed_buffer;
196 z_stream strm_decompress;
202 KKASSERT(bp->b_bufsize <= HAMMER2_PBUFSIZE);
203 strm_decompress.avail_in = 0;
204 strm_decompress.next_in = Z_NULL;
206 ret = inflateInit(&strm_decompress);
209 kprintf("HAMMER2 ZLIB: Fatal error in inflateInit.\n");
211 compressed_buffer = objcache_get(cache_buffer_read, M_INTWAIT);
212 strm_decompress.next_in = __DECONST(char *, data);
214 /* XXX supply proper size, subset of device bp */
215 strm_decompress.avail_in = bytes;
216 strm_decompress.next_out = compressed_buffer;
217 strm_decompress.avail_out = bp->b_bufsize;
219 ret = inflate(&strm_decompress, Z_FINISH);
220 if (ret != Z_STREAM_END) {
221 kprintf("HAMMER2 ZLIB: Fatal error during decompression.\n");
222 bzero(compressed_buffer, bp->b_bufsize);
224 bcopy(compressed_buffer, bp->b_data, bp->b_bufsize);
225 result = bp->b_bufsize - strm_decompress.avail_out;
226 if (result < bp->b_bufsize)
227 bzero(bp->b_data + result, strm_decompress.avail_out);
228 objcache_put(cache_buffer_read, compressed_buffer);
229 ret = inflateEnd(&strm_decompress);
232 bp->b_flags |= B_AGE;
236 * Logical buffer I/O, async read.
240 hammer2_strategy_read(struct vop_strategy_args *ap)
242 hammer2_xop_strategy_t *xop;
250 lbase = bio->bio_offset;
251 KKASSERT(((int)lbase & HAMMER2_PBUFMASK) == 0);
253 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_STRATEGY);
257 hammer2_mtx_init(&xop->lock, "h2bior");
258 hammer2_xop_start(&xop->head, &hammer2_strategy_read_desc);
259 /* asynchronous completion */
265 * Per-node XOP (threaded), do a synchronous lookup of the chain and
266 * its data. The frontend is asynchronous, so we are also responsible
267 * for racing to terminate the frontend.
270 hammer2_xop_strategy_read(hammer2_xop_t *arg, void *scratch, int clindex)
272 hammer2_xop_strategy_t *xop = &arg->xop_strategy;
273 hammer2_chain_t *parent;
274 hammer2_chain_t *chain;
275 hammer2_chain_t *focus;
276 hammer2_key_t key_dummy;
284 * Note that we can race completion of the bio supplied by
285 * the front-end so we cannot access it until we determine
286 * that we are the ones finishing it up.
291 * This is difficult to optimize. The logical buffer might be
292 * partially dirty (contain dummy zero-fill pages), which would
293 * mess up our crc calculation if we were to try a direct read.
294 * So for now we always double-buffer through the underlying
297 * If not for the above problem we could conditionalize on
298 * (1) 64KB buffer, (2) one chain (not multi-master) and
299 * (3) !hammer2_double_buffer, and issue a direct read into the
302 parent = hammer2_inode_chain(xop->head.ip1, clindex,
303 HAMMER2_RESOLVE_ALWAYS |
304 HAMMER2_RESOLVE_SHARED);
306 chain = hammer2_chain_lookup(&parent, &key_dummy,
309 HAMMER2_LOOKUP_ALWAYS |
310 HAMMER2_LOOKUP_SHARED);
312 error = chain->error;
314 error = HAMMER2_ERROR_EIO;
317 error = hammer2_xop_feed(&xop->head, chain, clindex, error);
319 hammer2_chain_unlock(chain);
320 hammer2_chain_drop(chain);
323 hammer2_chain_unlock(parent);
324 hammer2_chain_drop(parent);
326 chain = NULL; /* safety */
327 parent = NULL; /* safety */
330 * Race to finish the frontend. First-to-complete. bio is only
331 * valid if we are determined to be the ones able to complete
336 hammer2_mtx_ex(&xop->lock);
338 hammer2_mtx_unlock(&xop->lock);
346 * Async operation has not completed and we now own the lock.
347 * Determine if we can complete the operation by issuing the
348 * frontend collection non-blocking.
350 * H2 double-buffers the data, setting B_NOTMETA on the logical
351 * buffer hints to the OS that the logical buffer should not be
352 * swapcached (since the device buffer can be).
354 * Also note that even for compressed data we would rather the
355 * kernel cache/swapcache device buffers more and (decompressed)
356 * logical buffers less, since that will significantly improve
357 * the amount of end-user data that can be cached.
359 * NOTE: The chain->data for xop->head.cluster.focus will be
360 * synchronized to the current cpu by xop_collect(),
361 * but other chains in the cluster might not be.
363 error = hammer2_xop_collect(&xop->head, HAMMER2_XOP_COLLECT_NOWAIT);
368 hammer2_mtx_unlock(&xop->lock);
369 bp->b_flags |= B_NOTMETA;
370 focus = xop->head.cluster.focus;
371 data = hammer2_xop_gdata(&xop->head)->buf;
372 hammer2_strategy_read_completion(focus, data, xop->bio);
373 hammer2_xop_pdata(&xop->head);
375 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
377 case HAMMER2_ERROR_ENOENT:
379 hammer2_mtx_unlock(&xop->lock);
380 bp->b_flags |= B_NOTMETA;
383 bzero(bp->b_data, bp->b_bcount);
385 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
387 case HAMMER2_ERROR_EINPROGRESS:
388 hammer2_mtx_unlock(&xop->lock);
391 kprintf("xop_strategy_read: error %08x loff=%016jx\n",
392 error, (intmax_t)bp->b_loffset);
394 hammer2_mtx_unlock(&xop->lock);
395 bp->b_flags |= B_ERROR;
398 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
405 hammer2_strategy_read_completion(hammer2_chain_t *focus, const char *data,
408 struct buf *bp = bio->bio_buf;
410 if (focus->bref.type == HAMMER2_BREF_TYPE_INODE) {
412 * Copy from in-memory inode structure.
414 bcopy(((const hammer2_inode_data_t *)data)->u.data,
415 bp->b_data, HAMMER2_EMBEDDED_BYTES);
416 bzero(bp->b_data + HAMMER2_EMBEDDED_BYTES,
417 bp->b_bcount - HAMMER2_EMBEDDED_BYTES);
420 } else if (focus->bref.type == HAMMER2_BREF_TYPE_DATA) {
422 * Data is on-media, record for live dedup. Release the
423 * chain (try to free it) when done. The data is still
424 * cached by both the buffer cache in front and the
425 * block device behind us.
427 * NOTE: Deduplication cannot be safely recorded for
428 * records without a check code.
430 hammer2_dedup_record(focus, NULL, data);
431 atomic_set_int(&focus->flags, HAMMER2_CHAIN_RELEASE);
434 * Decompression and copy.
436 switch (HAMMER2_DEC_COMP(focus->bref.methods)) {
437 case HAMMER2_COMP_LZ4:
438 hammer2_decompress_LZ4_callback(data, focus->bytes,
440 /* b_resid set by call */
442 case HAMMER2_COMP_ZLIB:
443 hammer2_decompress_ZLIB_callback(data, focus->bytes,
445 /* b_resid set by call */
447 case HAMMER2_COMP_NONE:
448 KKASSERT(focus->bytes <= bp->b_bcount);
449 bcopy(data, bp->b_data, focus->bytes);
450 if (focus->bytes < bp->b_bcount) {
451 bzero(bp->b_data + focus->bytes,
452 bp->b_bcount - focus->bytes);
458 panic("hammer2_strategy_read_completion: "
459 "unknown compression type");
462 panic("hammer2_strategy_read_completion: unknown bref type");
466 /****************************************************************************
468 ****************************************************************************/
471 * Functions for compression in threads,
472 * from hammer2_vnops.c
474 static void hammer2_write_file_core(char *data, hammer2_inode_t *ip,
475 hammer2_chain_t **parentp,
476 hammer2_key_t lbase, int ioflag, int pblksize,
477 hammer2_tid_t mtid, int *errorp);
478 static void hammer2_compress_and_write(char *data, hammer2_inode_t *ip,
479 hammer2_chain_t **parentp,
480 hammer2_key_t lbase, int ioflag, int pblksize,
481 hammer2_tid_t mtid, int *errorp,
482 int comp_algo, int check_algo);
483 static void hammer2_zero_check_and_write(char *data, hammer2_inode_t *ip,
484 hammer2_chain_t **parentp,
485 hammer2_key_t lbase, int ioflag, int pblksize,
486 hammer2_tid_t mtid, int *errorp,
488 static int test_block_zeros(const char *buf, size_t bytes);
489 static void zero_write(char *data, hammer2_inode_t *ip,
490 hammer2_chain_t **parentp,
492 hammer2_tid_t mtid, int *errorp);
493 static void hammer2_write_bp(hammer2_chain_t *chain, char *data,
494 int ioflag, int pblksize,
495 hammer2_tid_t mtid, int *errorp,
499 hammer2_strategy_write(struct vop_strategy_args *ap)
501 hammer2_xop_strategy_t *xop;
510 atomic_set_int(&ip->flags, HAMMER2_INODE_DIRTYDATA);
511 hammer2_lwinprog_ref(pmp);
512 hammer2_trans_assert_strategy(pmp);
513 hammer2_trans_init(pmp, HAMMER2_TRANS_BUFCACHE);
515 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING |
516 HAMMER2_XOP_STRATEGY);
519 xop->lbase = bio->bio_offset;
520 hammer2_mtx_init(&xop->lock, "h2biow");
521 hammer2_xop_start(&xop->head, &hammer2_strategy_write_desc);
522 /* asynchronous completion */
524 hammer2_lwinprog_wait(pmp, hammer2_flush_pipe);
530 * Per-node XOP (threaded). Write the logical buffer to the media.
532 * This is a bit problematic because there may be multiple target and
533 * any of them may be able to release the bp. In addition, if our
534 * particulr target is offline we don't want to block the bp (and thus
535 * the frontend). To accomplish this we copy the data to the per-thr
539 hammer2_xop_strategy_write(hammer2_xop_t *arg, void *scratch, int clindex)
541 hammer2_xop_strategy_t *xop = &arg->xop_strategy;
542 hammer2_chain_t *parent;
553 * We can only access the bp/bio if the frontend has not yet
558 hammer2_mtx_sh(&xop->lock);
560 hammer2_mtx_unlock(&xop->lock);
565 bio = xop->bio; /* ephermal */
566 bp = bio->bio_buf; /* ephermal */
567 ip = xop->head.ip1; /* retained by ref */
570 /* hammer2_trans_init(parent->hmp->spmp, HAMMER2_TRANS_BUFCACHE); */
572 lblksize = hammer2_calc_logical(ip, bio->bio_offset, &lbase, NULL);
573 pblksize = hammer2_calc_physical(ip, lbase);
575 KKASSERT(lblksize <= MAXPHYS);
576 bcopy(bp->b_data, bio_data, lblksize);
578 hammer2_mtx_unlock(&xop->lock);
579 bp = NULL; /* safety, illegal to access after unlock */
580 bio = NULL; /* safety, illegal to access after unlock */
585 parent = hammer2_inode_chain(ip, clindex, HAMMER2_RESOLVE_ALWAYS);
586 hammer2_write_file_core(bio_data, ip, &parent,
587 lbase, IO_ASYNC, pblksize,
588 xop->head.mtid, &error);
590 hammer2_chain_unlock(parent);
591 hammer2_chain_drop(parent);
592 parent = NULL; /* safety */
594 hammer2_xop_feed(&xop->head, NULL, clindex, error);
597 * Try to complete the operation on behalf of the front-end.
601 hammer2_mtx_ex(&xop->lock);
603 hammer2_mtx_unlock(&xop->lock);
608 * Async operation has not completed and we now own the lock.
609 * Determine if we can complete the operation by issuing the
610 * frontend collection non-blocking.
612 * H2 double-buffers the data, setting B_NOTMETA on the logical
613 * buffer hints to the OS that the logical buffer should not be
614 * swapcached (since the device buffer can be).
616 error = hammer2_xop_collect(&xop->head, HAMMER2_XOP_COLLECT_NOWAIT);
618 if (error == HAMMER2_ERROR_EINPROGRESS) {
619 hammer2_mtx_unlock(&xop->lock);
624 * Async operation has completed.
627 hammer2_mtx_unlock(&xop->lock);
629 bio = xop->bio; /* now owned by us */
630 bp = bio->bio_buf; /* now owned by us */
632 if (error == HAMMER2_ERROR_ENOENT || error == 0) {
633 bp->b_flags |= B_NOTMETA;
638 kprintf("xop_strategy_write: error %d loff=%016jx\n",
639 error, (intmax_t)bp->b_loffset);
640 bp->b_flags |= B_ERROR;
644 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
645 hammer2_trans_assert_strategy(ip->pmp);
646 hammer2_lwinprog_drop(ip->pmp);
647 hammer2_trans_done(ip->pmp, HAMMER2_TRANS_BUFCACHE);
651 * Wait for pending I/O to complete
654 hammer2_bioq_sync(hammer2_pfs_t *pmp)
656 hammer2_lwinprog_wait(pmp, 0);
660 * Assign physical storage at (cparent, lbase), returning a suitable chain
661 * and setting *errorp appropriately.
663 * If no error occurs, the returned chain will be in a modified state.
665 * If an error occurs, the returned chain may or may not be NULL. If
666 * not-null any chain->error (if not 0) will also be rolled up into *errorp.
667 * So the caller only needs to test *errorp.
669 * cparent can wind up being anything.
671 * If datap is not NULL, *datap points to the real data we intend to write.
672 * If we can dedup the storage location we set *datap to NULL to indicate
673 * to the caller that a dedup occurred.
675 * NOTE: Special case for data embedded in inode.
679 hammer2_assign_physical(hammer2_inode_t *ip, hammer2_chain_t **parentp,
680 hammer2_key_t lbase, int pblksize,
681 hammer2_tid_t mtid, char **datap, int *errorp)
683 hammer2_chain_t *chain;
684 hammer2_key_t key_dummy;
685 hammer2_off_t dedup_off;
686 int pradix = hammer2_getradix(pblksize);
689 * Locate the chain associated with lbase, return a locked chain.
690 * However, do not instantiate any data reference (which utilizes a
691 * device buffer) because we will be using direct IO via the
692 * logical buffer cache buffer.
694 KKASSERT(pblksize >= HAMMER2_ALLOC_MIN);
696 chain = hammer2_chain_lookup(parentp, &key_dummy,
699 HAMMER2_LOOKUP_NODATA);
702 * The lookup code should not return a DELETED chain to us, unless
703 * its a short-file embedded in the inode. Then it is possible for
704 * the lookup to return a deleted inode.
706 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED) &&
707 chain->bref.type != HAMMER2_BREF_TYPE_INODE) {
708 kprintf("assign physical deleted chain @ "
709 "%016jx (%016jx.%02x) ip %016jx\n",
710 lbase, chain->bref.data_off, chain->bref.type,
717 * We found a hole, create a new chain entry.
719 * NOTE: DATA chains are created without device backing
720 * store (nor do we want any).
722 dedup_off = hammer2_dedup_lookup((*parentp)->hmp, datap,
724 *errorp |= hammer2_chain_create(parentp, &chain, NULL, ip->pmp,
725 HAMMER2_ENC_CHECK(ip->meta.check_algo) |
726 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE),
727 lbase, HAMMER2_PBUFRADIX,
728 HAMMER2_BREF_TYPE_DATA,
733 /*ip->delta_dcount += pblksize;*/
734 } else if (chain->error == 0) {
735 switch (chain->bref.type) {
736 case HAMMER2_BREF_TYPE_INODE:
738 * The data is embedded in the inode, which requires
741 *errorp |= hammer2_chain_modify_ip(ip, chain, mtid, 0);
743 case HAMMER2_BREF_TYPE_DATA:
744 dedup_off = hammer2_dedup_lookup(chain->hmp, datap,
746 if (chain->bytes != pblksize) {
747 *errorp |= hammer2_chain_resize(chain,
750 HAMMER2_MODIFY_OPTDATA);
756 * DATA buffers must be marked modified whether the
757 * data is in a logical buffer or not. We also have
758 * to make this call to fixup the chain data pointers
759 * after resizing in case this is an encrypted or
762 *errorp |= hammer2_chain_modify(chain, mtid, dedup_off,
763 HAMMER2_MODIFY_OPTDATA);
766 panic("hammer2_assign_physical: bad type");
771 *errorp = chain->error;
773 atomic_set_int(&ip->flags, HAMMER2_INODE_DIRTYDATA);
779 * hammer2_write_file_core()
781 * The core write function which determines which path to take
782 * depending on compression settings. We also have to locate the
783 * related chains so we can calculate and set the check data for
788 hammer2_write_file_core(char *data, hammer2_inode_t *ip,
789 hammer2_chain_t **parentp,
790 hammer2_key_t lbase, int ioflag, int pblksize,
791 hammer2_tid_t mtid, int *errorp)
793 hammer2_chain_t *chain;
798 switch(HAMMER2_DEC_ALGO(ip->meta.comp_algo)) {
799 case HAMMER2_COMP_NONE:
801 * We have to assign physical storage to the buffer
802 * we intend to dirty or write now to avoid deadlocks
803 * in the strategy code later.
805 * This can return NOOFFSET for inode-embedded data.
806 * The strategy code will take care of it in that case.
809 chain = hammer2_assign_physical(ip, parentp, lbase, pblksize,
810 mtid, &bdata, errorp);
812 /* skip modifications */
813 } else if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
814 hammer2_inode_data_t *wipdata;
816 wipdata = &chain->data->ipdata;
817 KKASSERT(wipdata->meta.op_flags &
818 HAMMER2_OPFLAG_DIRECTDATA);
819 bcopy(data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
820 ++hammer2_iod_file_wembed;
821 } else if (bdata == NULL) {
823 * Copy of data already present on-media.
825 chain->bref.methods =
826 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) +
827 HAMMER2_ENC_CHECK(ip->meta.check_algo);
828 hammer2_chain_setcheck(chain, data);
829 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
831 hammer2_write_bp(chain, data, ioflag, pblksize,
832 mtid, errorp, ip->meta.check_algo);
835 hammer2_chain_unlock(chain);
836 hammer2_chain_drop(chain);
839 case HAMMER2_COMP_AUTOZERO:
841 * Check for zero-fill only
843 hammer2_zero_check_and_write(data, ip, parentp,
844 lbase, ioflag, pblksize,
846 ip->meta.check_algo);
848 case HAMMER2_COMP_LZ4:
849 case HAMMER2_COMP_ZLIB:
852 * Check for zero-fill and attempt compression.
854 hammer2_compress_and_write(data, ip, parentp,
855 lbase, ioflag, pblksize,
858 ip->meta.check_algo);
866 * Generic function that will perform the compression in compression
867 * write path. The compression algorithm is determined by the settings
868 * obtained from inode.
872 hammer2_compress_and_write(char *data, hammer2_inode_t *ip,
873 hammer2_chain_t **parentp,
874 hammer2_key_t lbase, int ioflag, int pblksize,
875 hammer2_tid_t mtid, int *errorp, int comp_algo, int check_algo)
877 hammer2_chain_t *chain;
884 * An all-zeros write creates a hole unless the check code
885 * is disabled. When the check code is disabled all writes
886 * are done in-place, including any all-zeros writes.
888 * NOTE: A snapshot will still force a copy-on-write
889 * (see the HAMMER2_CHECK_NONE in hammer2_chain.c).
891 if (check_algo != HAMMER2_CHECK_NONE &&
892 test_block_zeros(data, pblksize)) {
893 zero_write(data, ip, parentp, lbase, mtid, errorp);
898 * Compression requested. Try to compress the block. We store
899 * the data normally if we cannot sufficiently compress it.
901 * We have a heuristic to detect files which are mostly
902 * uncompressable and avoid the compression attempt in that
903 * case. If the compression heuristic is turned off, we always
909 KKASSERT(pblksize / 2 <= 32768);
911 if (ip->comp_heuristic < 8 || (ip->comp_heuristic & 7) == 0 ||
912 hammer2_always_compress) {
913 z_stream strm_compress;
917 switch(HAMMER2_DEC_ALGO(comp_algo)) {
918 case HAMMER2_COMP_LZ4:
920 * We need to prefix with the size, LZ4
921 * doesn't do it for us. Add the related
924 * NOTE: The LZ4 code seems to assume at least an
925 * 8-byte buffer size granularity and may
926 * overrun the buffer if given a 4-byte
929 comp_buffer = objcache_get(cache_buffer_write,
931 comp_size = LZ4_compress_limitedOutput(
933 &comp_buffer[sizeof(int)],
935 pblksize / 2 - sizeof(int64_t));
936 *(int *)comp_buffer = comp_size;
938 comp_size += sizeof(int);
940 case HAMMER2_COMP_ZLIB:
941 comp_level = HAMMER2_DEC_LEVEL(comp_algo);
943 comp_level = 6; /* default zlib compression */
944 else if (comp_level < 6)
946 else if (comp_level > 9)
948 ret = deflateInit(&strm_compress, comp_level);
950 kprintf("HAMMER2 ZLIB: fatal error "
951 "on deflateInit.\n");
954 comp_buffer = objcache_get(cache_buffer_write,
956 strm_compress.next_in = data;
957 strm_compress.avail_in = pblksize;
958 strm_compress.next_out = comp_buffer;
959 strm_compress.avail_out = pblksize / 2;
960 ret = deflate(&strm_compress, Z_FINISH);
961 if (ret == Z_STREAM_END) {
962 comp_size = pblksize / 2 -
963 strm_compress.avail_out;
967 ret = deflateEnd(&strm_compress);
970 kprintf("Error: Unknown compression method.\n");
971 kprintf("Comp_method = %d.\n", comp_algo);
976 if (comp_size == 0) {
978 * compression failed or turned off
980 comp_block_size = pblksize; /* safety */
981 if (++ip->comp_heuristic > 128)
982 ip->comp_heuristic = 8;
985 * compression succeeded
987 ip->comp_heuristic = 0;
988 if (comp_size <= 1024) {
989 comp_block_size = 1024;
990 } else if (comp_size <= 2048) {
991 comp_block_size = 2048;
992 } else if (comp_size <= 4096) {
993 comp_block_size = 4096;
994 } else if (comp_size <= 8192) {
995 comp_block_size = 8192;
996 } else if (comp_size <= 16384) {
997 comp_block_size = 16384;
998 } else if (comp_size <= 32768) {
999 comp_block_size = 32768;
1001 panic("hammer2: WRITE PATH: "
1002 "Weird comp_size value.");
1004 comp_block_size = pblksize;
1008 * Must zero the remainder or dedup (which operates on a
1009 * physical block basis) will not find matches.
1011 if (comp_size < comp_block_size) {
1012 bzero(comp_buffer + comp_size,
1013 comp_block_size - comp_size);
1018 * Assign physical storage, bdata will be set to NULL if a live-dedup
1021 bdata = comp_size ? comp_buffer : data;
1022 chain = hammer2_assign_physical(ip, parentp, lbase, comp_block_size,
1023 mtid, &bdata, errorp);
1029 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
1030 hammer2_inode_data_t *wipdata;
1032 *errorp = hammer2_chain_modify_ip(ip, chain, mtid, 0);
1034 wipdata = &chain->data->ipdata;
1035 KKASSERT(wipdata->meta.op_flags &
1036 HAMMER2_OPFLAG_DIRECTDATA);
1037 bcopy(data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
1038 ++hammer2_iod_file_wembed;
1040 } else if (bdata == NULL) {
1042 * Live deduplication, a copy of the data is already present
1046 chain->bref.methods =
1047 HAMMER2_ENC_COMP(comp_algo) +
1048 HAMMER2_ENC_CHECK(check_algo);
1050 chain->bref.methods =
1052 HAMMER2_COMP_NONE) +
1053 HAMMER2_ENC_CHECK(check_algo);
1055 bdata = comp_size ? comp_buffer : data;
1056 hammer2_chain_setcheck(chain, bdata);
1057 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1061 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
1063 switch(chain->bref.type) {
1064 case HAMMER2_BREF_TYPE_INODE:
1065 panic("hammer2_compress_and_write: unexpected inode\n");
1067 case HAMMER2_BREF_TYPE_DATA:
1069 * Optimize out the read-before-write
1072 *errorp = hammer2_io_newnz(chain->hmp,
1074 chain->bref.data_off,
1078 hammer2_io_brelse(&dio);
1079 kprintf("hammer2: WRITE PATH: "
1080 "dbp bread error\n");
1083 bdata = hammer2_io_data(dio, chain->bref.data_off);
1086 * When loading the block make sure we don't
1087 * leave garbage after the compressed data.
1090 chain->bref.methods =
1091 HAMMER2_ENC_COMP(comp_algo) +
1092 HAMMER2_ENC_CHECK(check_algo);
1093 bcopy(comp_buffer, bdata, comp_block_size);
1095 chain->bref.methods =
1097 HAMMER2_COMP_NONE) +
1098 HAMMER2_ENC_CHECK(check_algo);
1099 bcopy(data, bdata, pblksize);
1103 * The flush code doesn't calculate check codes for
1104 * file data (doing so can result in excessive I/O),
1107 hammer2_chain_setcheck(chain, bdata);
1110 * Device buffer is now valid, chain is no longer in
1111 * the initial state.
1113 * (No blockref table worries with file data)
1115 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1116 hammer2_dedup_record(chain, dio, bdata);
1118 /* Now write the related bdp. */
1119 if (ioflag & IO_SYNC) {
1121 * Synchronous I/O requested.
1123 hammer2_io_bwrite(&dio);
1125 } else if ((ioflag & IO_DIRECT) &&
1126 loff + n == pblksize) {
1127 hammer2_io_bdwrite(&dio);
1129 } else if (ioflag & IO_ASYNC) {
1130 hammer2_io_bawrite(&dio);
1132 hammer2_io_bdwrite(&dio);
1136 panic("hammer2_compress_and_write: bad chain type %d\n",
1144 hammer2_chain_unlock(chain);
1145 hammer2_chain_drop(chain);
1148 objcache_put(cache_buffer_write, comp_buffer);
1154 * Function that performs zero-checking and writing without compression,
1155 * it corresponds to default zero-checking path.
1159 hammer2_zero_check_and_write(char *data, hammer2_inode_t *ip,
1160 hammer2_chain_t **parentp,
1161 hammer2_key_t lbase, int ioflag, int pblksize,
1162 hammer2_tid_t mtid, int *errorp,
1165 hammer2_chain_t *chain;
1168 if (check_algo != HAMMER2_CHECK_NONE &&
1169 test_block_zeros(data, pblksize)) {
1171 * An all-zeros write creates a hole unless the check code
1172 * is disabled. When the check code is disabled all writes
1173 * are done in-place, including any all-zeros writes.
1175 * NOTE: A snapshot will still force a copy-on-write
1176 * (see the HAMMER2_CHECK_NONE in hammer2_chain.c).
1178 zero_write(data, ip, parentp, lbase, mtid, errorp);
1181 * Normal write (bdata set to NULL if de-duplicated)
1184 chain = hammer2_assign_physical(ip, parentp, lbase, pblksize,
1185 mtid, &bdata, errorp);
1189 hammer2_write_bp(chain, data, ioflag, pblksize,
1190 mtid, errorp, check_algo);
1192 /* dedup occurred */
1193 chain->bref.methods =
1194 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) +
1195 HAMMER2_ENC_CHECK(check_algo);
1196 hammer2_chain_setcheck(chain, data);
1197 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1200 hammer2_chain_unlock(chain);
1201 hammer2_chain_drop(chain);
1209 * A function to test whether a block of data contains only zeros,
1210 * returns TRUE (non-zero) if the block is all zeros.
1214 test_block_zeros(const char *buf, size_t bytes)
1218 for (i = 0; i < bytes; i += sizeof(long)) {
1219 if (*(const long *)(buf + i) != 0)
1228 * Function to "write" a block that contains only zeros.
1232 zero_write(char *data, hammer2_inode_t *ip,
1233 hammer2_chain_t **parentp,
1234 hammer2_key_t lbase, hammer2_tid_t mtid, int *errorp)
1236 hammer2_chain_t *chain;
1237 hammer2_key_t key_dummy;
1239 chain = hammer2_chain_lookup(parentp, &key_dummy,
1242 HAMMER2_LOOKUP_NODATA);
1244 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
1245 hammer2_inode_data_t *wipdata;
1248 *errorp = hammer2_chain_modify_ip(ip, chain,
1252 wipdata = &chain->data->ipdata;
1253 KKASSERT(wipdata->meta.op_flags &
1254 HAMMER2_OPFLAG_DIRECTDATA);
1255 bzero(wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
1256 ++hammer2_iod_file_wembed;
1259 /* chain->error ok for deletion */
1260 hammer2_chain_delete(*parentp, chain,
1261 mtid, HAMMER2_DELETE_PERMANENT);
1262 ++hammer2_iod_file_wzero;
1264 atomic_set_int(&ip->flags, HAMMER2_INODE_DIRTYDATA);
1265 hammer2_chain_unlock(chain);
1266 hammer2_chain_drop(chain);
1268 ++hammer2_iod_file_wzero;
1275 * Function to write the data as it is, without performing any sort of
1276 * compression. This function is used in path without compression and
1277 * default zero-checking path.
1281 hammer2_write_bp(hammer2_chain_t *chain, char *data, int ioflag,
1283 hammer2_tid_t mtid, int *errorp, int check_algo)
1285 hammer2_inode_data_t *wipdata;
1290 error = 0; /* XXX TODO below */
1292 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
1294 switch(chain->bref.type) {
1295 case HAMMER2_BREF_TYPE_INODE:
1296 wipdata = &chain->data->ipdata;
1297 KKASSERT(wipdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA);
1298 bcopy(data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
1300 ++hammer2_iod_file_wembed;
1302 case HAMMER2_BREF_TYPE_DATA:
1303 error = hammer2_io_newnz(chain->hmp,
1305 chain->bref.data_off,
1306 chain->bytes, &dio);
1308 hammer2_io_bqrelse(&dio);
1309 kprintf("hammer2: WRITE PATH: "
1310 "dbp bread error\n");
1313 bdata = hammer2_io_data(dio, chain->bref.data_off);
1315 chain->bref.methods = HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) +
1316 HAMMER2_ENC_CHECK(check_algo);
1317 bcopy(data, bdata, chain->bytes);
1320 * The flush code doesn't calculate check codes for
1321 * file data (doing so can result in excessive I/O),
1324 hammer2_chain_setcheck(chain, bdata);
1327 * Device buffer is now valid, chain is no longer in
1328 * the initial state.
1330 * (No blockref table worries with file data)
1332 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1333 hammer2_dedup_record(chain, dio, bdata);
1335 if (ioflag & IO_SYNC) {
1337 * Synchronous I/O requested.
1339 hammer2_io_bwrite(&dio);
1341 } else if ((ioflag & IO_DIRECT) &&
1342 loff + n == pblksize) {
1343 hammer2_io_bdwrite(&dio);
1345 } else if (ioflag & IO_ASYNC) {
1346 hammer2_io_bawrite(&dio);
1348 hammer2_io_bdwrite(&dio);
1352 panic("hammer2_write_bp: bad chain type %d\n",
1362 * LIVE DEDUP HEURISTICS
1364 * Record media and crc information for possible dedup operation. Note
1365 * that the dedup mask bits must also be set in the related DIO for a dedup
1366 * to be fully validated (which is handled in the freemap allocation code).
1368 * WARNING! This code is SMP safe but the heuristic allows SMP collisions.
1369 * All fields must be loaded into locals and validated.
1371 * WARNING! Should only be used for file data and directory entries,
1372 * hammer2_chain_modify() only checks for the dedup case on data
1373 * chains. Also, dedup data can only be recorded for committed
1374 * chains (so NOT strategy writes which can undergo further
1375 * modification after the fact!).
1378 hammer2_dedup_record(hammer2_chain_t *chain, hammer2_io_t *dio,
1382 hammer2_dedup_t *dedup;
1390 * We can only record a dedup if we have media data to test against.
1391 * If dedup is not enabled, return early, which allows a chain to
1392 * remain marked MODIFIED (which might have benefits in special
1393 * situations, though typically it does not).
1395 if (hammer2_dedup_enable == 0)
1405 switch(HAMMER2_DEC_CHECK(chain->bref.methods)) {
1406 case HAMMER2_CHECK_ISCSI32:
1408 * XXX use the built-in crc (the dedup lookup sequencing
1409 * needs to be fixed so the check code is already present
1410 * when dedup_lookup is called)
1413 crc = (uint64_t)(uint32_t)chain->bref.check.iscsi32.value;
1415 crc = XXH64(data, chain->bytes, XXH_HAMMER2_SEED);
1417 case HAMMER2_CHECK_XXHASH64:
1418 crc = chain->bref.check.xxhash64.value;
1420 case HAMMER2_CHECK_SHA192:
1422 * XXX use the built-in crc (the dedup lookup sequencing
1423 * needs to be fixed so the check code is already present
1424 * when dedup_lookup is called)
1427 crc = ((uint64_t *)chain->bref.check.sha192.data)[0] ^
1428 ((uint64_t *)chain->bref.check.sha192.data)[1] ^
1429 ((uint64_t *)chain->bref.check.sha192.data)[2];
1431 crc = XXH64(data, chain->bytes, XXH_HAMMER2_SEED);
1435 * Cannot dedup without a check code
1437 * NOTE: In particular, CHECK_NONE allows a sector to be
1438 * overwritten without copy-on-write, recording
1439 * a dedup block for a CHECK_NONE object would be
1445 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DEDUPABLE);
1447 dedup = &hmp->heur_dedup[crc & (HAMMER2_DEDUP_HEUR_MASK & ~3)];
1448 for (i = 0; i < 4; ++i) {
1449 if (dedup[i].data_crc == crc) {
1453 dticks = (int)(dedup[i].ticks - dedup[best].ticks);
1454 if (dticks < 0 || dticks > hz * 60 * 30)
1458 if (hammer2_debug & 0x40000) {
1459 kprintf("REC %04x %016jx %016jx\n",
1460 (int)(dedup - hmp->heur_dedup),
1462 chain->bref.data_off);
1464 dedup->ticks = ticks;
1465 dedup->data_off = chain->bref.data_off;
1466 dedup->data_crc = crc;
1469 * Set the valid bits for the dedup only after we know the data
1470 * buffer has been updated. The alloc bits were set (and the valid
1471 * bits cleared) when the media was allocated.
1473 * This is done in two stages becuase the bulkfree code can race
1474 * the gap between allocation and data population. Both masks must
1475 * be set before a bcmp/dedup operation is able to use the block.
1477 mask = hammer2_dedup_mask(dio, chain->bref.data_off, chain->bytes);
1478 atomic_set_64(&dio->dedup_valid, mask);
1482 * XXX removed. MODIFIED is an integral part of the flush code,
1483 * lets not just clear it
1486 * Once we record the dedup the chain must be marked clean to
1487 * prevent reuse of the underlying block. Remember that this
1488 * write occurs when the buffer cache is flushed (i.e. on sync(),
1489 * fsync(), filesystem periodic sync, or when the kernel needs to
1490 * flush a buffer), and not whenever the user write()s.
1492 if (chain->flags & HAMMER2_CHAIN_MODIFIED) {
1493 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
1494 atomic_add_long(&hammer2_count_modified_chains, -1);
1496 hammer2_pfs_memory_wakeup(chain->pmp, -1);
1503 hammer2_dedup_lookup(hammer2_dev_t *hmp, char **datap, int pblksize)
1505 hammer2_dedup_t *dedup;
1514 if (hammer2_dedup_enable == 0)
1521 * XXX use the built-in crc (the dedup lookup sequencing
1522 * needs to be fixed so the check code is already present
1523 * when dedup_lookup is called)
1525 crc = XXH64(data, pblksize, XXH_HAMMER2_SEED);
1526 dedup = &hmp->heur_dedup[crc & (HAMMER2_DEDUP_HEUR_MASK & ~3)];
1528 if (hammer2_debug & 0x40000) {
1529 kprintf("LOC %04x/4 %016jx\n",
1530 (int)(dedup - hmp->heur_dedup),
1534 for (i = 0; i < 4; ++i) {
1535 off = dedup[i].data_off;
1537 if (dedup[i].data_crc != crc)
1539 if ((1 << (int)(off & HAMMER2_OFF_MASK_RADIX)) != pblksize)
1541 dio = hammer2_io_getquick(hmp, off, pblksize);
1543 dtmp = hammer2_io_data(dio, off),
1544 mask = hammer2_dedup_mask(dio, off, pblksize);
1545 if ((dio->dedup_alloc & mask) == mask &&
1546 (dio->dedup_valid & mask) == mask &&
1547 bcmp(data, dtmp, pblksize) == 0) {
1548 if (hammer2_debug & 0x40000) {
1549 kprintf("DEDUP SUCCESS %016jx\n",
1552 hammer2_io_putblk(&dio);
1554 dedup[i].ticks = ticks; /* update use */
1555 atomic_add_long(&hammer2_iod_file_wdedup,
1558 return off; /* RETURN */
1560 hammer2_io_putblk(&dio);
1567 * Poof. Races are ok, if someone gets in and reuses a dedup offset
1568 * before or while we are clearing it they will also recover the freemap
1569 * entry (set it to fully allocated), so a bulkfree race can only set it
1570 * to a possibly-free state.
1572 * XXX ok, well, not really sure races are ok but going to run with it
1576 hammer2_dedup_clear(hammer2_dev_t *hmp)
1580 for (i = 0; i < HAMMER2_DEDUP_HEUR_SIZE; ++i) {
1581 hmp->heur_dedup[i].data_off = 0;
1582 hmp->heur_dedup[i].ticks = ticks - 1;