2 * Copyright (c) 2011-2015 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@dragonflybsd.org>
6 * by Venkatesh Srinivas <vsrinivas@dragonflybsd.org>
7 * by Daniel Flores (GSOC 2013 - mentored by Matthew Dillon, compression)
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in
17 * the documentation and/or other materials provided with the
19 * 3. Neither the name of The DragonFly Project nor the names of its
20 * contributors may be used to endorse or promote products derived
21 * from this software without specific, prior written permission.
23 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
24 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
25 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
26 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
27 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
28 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
29 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
30 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
31 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
32 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
33 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
37 * This module handles low level logical file I/O (strategy) which backs
38 * the logical buffer cache.
40 * [De]compression, zero-block, check codes, and buffer cache operations
41 * for file data is handled here.
43 * Live dedup makes its home here as well.
46 #include <sys/param.h>
47 #include <sys/systm.h>
48 #include <sys/kernel.h>
49 #include <sys/fcntl.h>
52 #include <sys/namei.h>
53 #include <sys/mount.h>
54 #include <sys/vnode.h>
55 #include <sys/mountctl.h>
56 #include <sys/dirent.h>
58 #include <sys/objcache.h>
59 #include <sys/event.h>
61 #include <vfs/fifofs/fifo.h>
64 #include "hammer2_lz4.h"
66 #include "zlib/hammer2_zlib.h"
68 struct objcache *cache_buffer_read;
69 struct objcache *cache_buffer_write;
72 * Strategy code (async logical file buffer I/O from system)
74 * Except for the transaction init (which should normally not block),
75 * we essentially run the strategy operation asynchronously via a XOP.
77 * XXX This isn't supposed to be able to deadlock against vfs_sync vfsync()
78 * calls but it has in the past when multiple flushes are queued.
80 * XXX We currently terminate the transaction once we get a quorum, otherwise
81 * the frontend can stall, but this can leave the remaining nodes with
82 * a potential flush conflict. We need to delay flushes on those nodes
83 * until running transactions complete separately from the normal
84 * transaction sequencing. FIXME TODO.
86 static void hammer2_strategy_xop_read(hammer2_thread_t *thr,
88 static void hammer2_strategy_xop_write(hammer2_thread_t *thr,
90 static int hammer2_strategy_read(struct vop_strategy_args *ap);
91 static int hammer2_strategy_write(struct vop_strategy_args *ap);
92 static void hammer2_strategy_read_completion(hammer2_chain_t *chain,
93 char *data, struct bio *bio);
95 static hammer2_off_t hammer2_dedup_lookup(hammer2_dev_t *hmp,
96 char **datap, int pblksize);
99 hammer2_vop_strategy(struct vop_strategy_args *ap)
110 error = hammer2_strategy_read(ap);
111 ++hammer2_iod_file_read;
114 error = hammer2_strategy_write(ap);
115 ++hammer2_iod_file_write;
118 bp->b_error = error = EINVAL;
119 bp->b_flags |= B_ERROR;
127 * Return the largest contiguous physical disk range for the logical
130 * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
132 * Basically disabled, the logical buffer write thread has to deal with
133 * buffers one-at-a-time. Note that this should not prevent cluster_read()
134 * from reading-ahead, it simply prevents it from trying form a single
135 * cluster buffer for the logical request. H2 already uses 64KB buffers!
138 hammer2_vop_bmap(struct vop_bmap_args *ap)
140 *ap->a_doffsetp = NOOFFSET;
148 /****************************************************************************
150 ****************************************************************************/
152 * Callback used in read path in case that a block is compressed with LZ4.
156 hammer2_decompress_LZ4_callback(const char *data, u_int bytes, struct bio *bio)
159 char *compressed_buffer;
166 if bio->bio_caller_info2.index &&
167 bio->bio_caller_info1.uvalue32 !=
168 crc32(bp->b_data, bp->b_bufsize) --- return error
171 KKASSERT(bp->b_bufsize <= HAMMER2_PBUFSIZE);
172 compressed_size = *(const int *)data;
173 KKASSERT((uint32_t)compressed_size <= bytes - sizeof(int));
175 compressed_buffer = objcache_get(cache_buffer_read, M_INTWAIT);
176 result = LZ4_decompress_safe(__DECONST(char *, &data[sizeof(int)]),
181 kprintf("READ PATH: Error during decompression."
183 (intmax_t)bio->bio_offset, bytes);
184 /* make sure it isn't random garbage */
185 bzero(compressed_buffer, bp->b_bufsize);
187 KKASSERT(result <= bp->b_bufsize);
188 bcopy(compressed_buffer, bp->b_data, bp->b_bufsize);
189 if (result < bp->b_bufsize)
190 bzero(bp->b_data + result, bp->b_bufsize - result);
191 objcache_put(cache_buffer_read, compressed_buffer);
193 bp->b_flags |= B_AGE;
197 * Callback used in read path in case that a block is compressed with ZLIB.
198 * It is almost identical to LZ4 callback, so in theory they can be unified,
199 * but we didn't want to make changes in bio structure for that.
203 hammer2_decompress_ZLIB_callback(const char *data, u_int bytes, struct bio *bio)
206 char *compressed_buffer;
207 z_stream strm_decompress;
213 KKASSERT(bp->b_bufsize <= HAMMER2_PBUFSIZE);
214 strm_decompress.avail_in = 0;
215 strm_decompress.next_in = Z_NULL;
217 ret = inflateInit(&strm_decompress);
220 kprintf("HAMMER2 ZLIB: Fatal error in inflateInit.\n");
222 compressed_buffer = objcache_get(cache_buffer_read, M_INTWAIT);
223 strm_decompress.next_in = __DECONST(char *, data);
225 /* XXX supply proper size, subset of device bp */
226 strm_decompress.avail_in = bytes;
227 strm_decompress.next_out = compressed_buffer;
228 strm_decompress.avail_out = bp->b_bufsize;
230 ret = inflate(&strm_decompress, Z_FINISH);
231 if (ret != Z_STREAM_END) {
232 kprintf("HAMMER2 ZLIB: Fatar error during decompression.\n");
233 bzero(compressed_buffer, bp->b_bufsize);
235 bcopy(compressed_buffer, bp->b_data, bp->b_bufsize);
236 result = bp->b_bufsize - strm_decompress.avail_out;
237 if (result < bp->b_bufsize)
238 bzero(bp->b_data + result, strm_decompress.avail_out);
239 objcache_put(cache_buffer_read, compressed_buffer);
240 ret = inflateEnd(&strm_decompress);
243 bp->b_flags |= B_AGE;
247 * Logical buffer I/O, async read.
251 hammer2_strategy_read(struct vop_strategy_args *ap)
253 hammer2_xop_strategy_t *xop;
263 nbio = push_bio(bio);
265 lbase = bio->bio_offset;
266 KKASSERT(((int)lbase & HAMMER2_PBUFMASK) == 0);
268 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_STRATEGY);
272 hammer2_mtx_init(&xop->lock, "h2bior");
273 hammer2_xop_start(&xop->head, hammer2_strategy_xop_read);
274 /* asynchronous completion */
280 * Per-node XOP (threaded), do a synchronous lookup of the chain and
281 * its data. The frontend is asynchronous, so we are also responsible
282 * for racing to terminate the frontend.
286 hammer2_strategy_xop_read(hammer2_thread_t *thr, hammer2_xop_t *arg)
288 hammer2_xop_strategy_t *xop = &arg->xop_strategy;
289 hammer2_chain_t *parent;
290 hammer2_chain_t *chain;
291 hammer2_key_t key_dummy;
298 * Note that we can race completion of the bio supplied by
299 * the front-end so we cannot access it until we determine
300 * that we are the ones finishing it up.
305 * This is difficult to optimize. The logical buffer might be
306 * partially dirty (contain dummy zero-fill pages), which would
307 * mess up our crc calculation if we were to try a direct read.
308 * So for now we always double-buffer through the underlying
311 * If not for the above problem we could conditionalize on
312 * (1) 64KB buffer, (2) one chain (not multi-master) and
313 * (3) !hammer2_double_buffer, and issue a direct read into the
316 parent = hammer2_inode_chain(xop->head.ip1, thr->clindex,
317 HAMMER2_RESOLVE_ALWAYS |
318 HAMMER2_RESOLVE_SHARED);
320 chain = hammer2_chain_lookup(&parent, &key_dummy,
323 HAMMER2_LOOKUP_ALWAYS |
324 HAMMER2_LOOKUP_SHARED);
326 error = chain->error;
327 error = hammer2_error_to_errno(error);
332 error = hammer2_xop_feed(&xop->head, chain, thr->clindex, error);
334 hammer2_chain_unlock(chain);
335 hammer2_chain_drop(chain);
338 hammer2_chain_unlock(parent);
339 hammer2_chain_drop(parent);
341 chain = NULL; /* safety */
342 parent = NULL; /* safety */
345 * Race to finish the frontend. First-to-complete. bio is only
346 * valid if we are determined to be the ones able to complete
351 hammer2_mtx_ex(&xop->lock);
353 hammer2_mtx_unlock(&xop->lock);
360 * Async operation has not completed and we now own the lock.
361 * Determine if we can complete the operation by issuing the
362 * frontend collection non-blocking.
364 * H2 double-buffers the data, setting B_NOTMETA on the logical
365 * buffer hints to the OS that the logical buffer should not be
366 * swapcached (since the device buffer can be).
368 * Also note that even for compressed data we would rather the
369 * kernel cache/swapcache device buffers more and (decompressed)
370 * logical buffers less, since that will significantly improve
371 * the amount of end-user data that can be cached.
373 error = hammer2_xop_collect(&xop->head, HAMMER2_XOP_COLLECT_NOWAIT);
378 hammer2_mtx_unlock(&xop->lock);
379 bp->b_flags |= B_NOTMETA;
380 chain = xop->head.cluster.focus;
381 hammer2_strategy_read_completion(chain, (char *)chain->data,
384 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
388 hammer2_mtx_unlock(&xop->lock);
389 bp->b_flags |= B_NOTMETA;
392 bzero(bp->b_data, bp->b_bcount);
394 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
397 hammer2_mtx_unlock(&xop->lock);
400 kprintf("strategy_xop_read: error %d loff=%016jx\n",
401 error, bp->b_loffset);
403 hammer2_mtx_unlock(&xop->lock);
404 bp->b_flags |= B_ERROR;
407 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
414 hammer2_strategy_read_completion(hammer2_chain_t *chain, char *data,
417 struct buf *bp = bio->bio_buf;
419 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
421 * Copy from in-memory inode structure.
423 bcopy(((hammer2_inode_data_t *)data)->u.data,
424 bp->b_data, HAMMER2_EMBEDDED_BYTES);
425 bzero(bp->b_data + HAMMER2_EMBEDDED_BYTES,
426 bp->b_bcount - HAMMER2_EMBEDDED_BYTES);
429 } else if (chain->bref.type == HAMMER2_BREF_TYPE_DATA) {
431 * Data is on-media, record for live dedup. Release the
432 * chain (try to free it) when done. The data is still
433 * cached by both the buffer cache in front and the
434 * block device behind us. This leaves more room in the
435 * LRU chain cache for meta-data chains which we really
438 * NOTE: Deduplication cannot be safely recorded for
439 * records without a check code.
441 hammer2_dedup_record(chain, NULL, data);
442 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
445 * Decompression and copy.
447 switch (HAMMER2_DEC_COMP(chain->bref.methods)) {
448 case HAMMER2_COMP_LZ4:
449 hammer2_decompress_LZ4_callback(data, chain->bytes,
451 /* b_resid set by call */
453 case HAMMER2_COMP_ZLIB:
454 hammer2_decompress_ZLIB_callback(data, chain->bytes,
456 /* b_resid set by call */
458 case HAMMER2_COMP_NONE:
459 KKASSERT(chain->bytes <= bp->b_bcount);
460 bcopy(data, bp->b_data, chain->bytes);
461 if (chain->bytes < bp->b_bcount) {
462 bzero(bp->b_data + chain->bytes,
463 bp->b_bcount - chain->bytes);
469 panic("hammer2_strategy_read: "
470 "unknown compression type");
473 panic("hammer2_strategy_read: unknown bref type");
477 /****************************************************************************
479 ****************************************************************************/
482 * Functions for compression in threads,
483 * from hammer2_vnops.c
485 static void hammer2_write_file_core(char *data, hammer2_inode_t *ip,
486 hammer2_chain_t **parentp,
487 hammer2_key_t lbase, int ioflag, int pblksize,
488 hammer2_tid_t mtid, int *errorp);
489 static void hammer2_compress_and_write(char *data, hammer2_inode_t *ip,
490 hammer2_chain_t **parentp,
491 hammer2_key_t lbase, int ioflag, int pblksize,
492 hammer2_tid_t mtid, int *errorp,
493 int comp_algo, int check_algo);
494 static void hammer2_zero_check_and_write(char *data, hammer2_inode_t *ip,
495 hammer2_chain_t **parentp,
496 hammer2_key_t lbase, int ioflag, int pblksize,
497 hammer2_tid_t mtid, int *errorp,
499 static int test_block_zeros(const char *buf, size_t bytes);
500 static void zero_write(char *data, hammer2_inode_t *ip,
501 hammer2_chain_t **parentp,
503 hammer2_tid_t mtid, int *errorp);
504 static void hammer2_write_bp(hammer2_chain_t *chain, char *data,
505 int ioflag, int pblksize,
506 hammer2_tid_t mtid, int *errorp,
511 hammer2_strategy_write(struct vop_strategy_args *ap)
513 hammer2_xop_strategy_t *xop;
524 hammer2_lwinprog_ref(pmp);
525 hammer2_trans_assert_strategy(pmp);
526 hammer2_trans_init(pmp, HAMMER2_TRANS_BUFCACHE);
528 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING |
529 HAMMER2_XOP_STRATEGY);
532 xop->lbase = bio->bio_offset;
533 hammer2_mtx_init(&xop->lock, "h2biow");
534 hammer2_xop_start(&xop->head, hammer2_strategy_xop_write);
535 /* asynchronous completion */
537 hammer2_lwinprog_wait(pmp, hammer2_flush_pipe);
543 * Per-node XOP (threaded). Write the logical buffer to the media.
545 * This is a bit problematic because there may be multiple target and
546 * any of them may be able to release the bp. In addition, if our
547 * particulr target is offline we don't want to block the bp (and thus
548 * the frontend). To accomplish this we copy the data to the per-thr
553 hammer2_strategy_xop_write(hammer2_thread_t *thr, hammer2_xop_t *arg)
555 hammer2_xop_strategy_t *xop = &arg->xop_strategy;
556 hammer2_chain_t *parent;
564 hammer2_off_t bio_offset;
568 * We can only access the bp/bio if the frontend has not yet
573 hammer2_mtx_sh(&xop->lock);
575 hammer2_mtx_unlock(&xop->lock);
580 bio = xop->bio; /* ephermal */
581 bp = bio->bio_buf; /* ephermal */
582 ip = xop->head.ip1; /* retained by ref */
583 bio_offset = bio->bio_offset;
584 bio_data = thr->scratch;
586 /* hammer2_trans_init(parent->hmp->spmp, HAMMER2_TRANS_BUFCACHE); */
588 lblksize = hammer2_calc_logical(ip, bio->bio_offset, &lbase, NULL);
589 pblksize = hammer2_calc_physical(ip, lbase);
590 bcopy(bp->b_data, bio_data, lblksize);
592 hammer2_mtx_unlock(&xop->lock);
593 bp = NULL; /* safety, illegal to access after unlock */
594 bio = NULL; /* safety, illegal to access after unlock */
599 parent = hammer2_inode_chain(ip, thr->clindex, HAMMER2_RESOLVE_ALWAYS);
600 hammer2_write_file_core(bio_data, ip, &parent,
601 lbase, IO_ASYNC, pblksize,
602 xop->head.mtid, &error);
603 error = hammer2_error_to_errno(error);
605 hammer2_chain_unlock(parent);
606 hammer2_chain_drop(parent);
607 parent = NULL; /* safety */
609 hammer2_xop_feed(&xop->head, NULL, thr->clindex, error);
612 * Try to complete the operation on behalf of the front-end.
616 hammer2_mtx_ex(&xop->lock);
618 hammer2_mtx_unlock(&xop->lock);
623 * Async operation has not completed and we now own the lock.
624 * Determine if we can complete the operation by issuing the
625 * frontend collection non-blocking.
627 * H2 double-buffers the data, setting B_NOTMETA on the logical
628 * buffer hints to the OS that the logical buffer should not be
629 * swapcached (since the device buffer can be).
631 error = hammer2_xop_collect(&xop->head, HAMMER2_XOP_COLLECT_NOWAIT);
633 if (error == EINPROGRESS) {
634 hammer2_mtx_unlock(&xop->lock);
639 * Async operation has completed.
642 hammer2_mtx_unlock(&xop->lock);
644 bio = xop->bio; /* now owned by us */
645 bp = bio->bio_buf; /* now owned by us */
647 if (error == ENOENT || error == 0) {
648 bp->b_flags |= B_NOTMETA;
653 kprintf("strategy_xop_write: error %d loff=%016jx\n",
654 error, bp->b_loffset);
655 bp->b_flags |= B_ERROR;
659 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
660 hammer2_trans_assert_strategy(ip->pmp);
661 hammer2_lwinprog_drop(ip->pmp);
662 hammer2_trans_done(ip->pmp);
666 * Wait for pending I/O to complete
669 hammer2_bioq_sync(hammer2_pfs_t *pmp)
671 hammer2_lwinprog_wait(pmp, 0);
675 * Create a new cluster at (cparent, lbase) and assign physical storage,
676 * returning a cluster suitable for I/O. The cluster will be in a modified
677 * state. Any chain->error will be rolled up into *errorp, but still
678 * returned. Caller must check *errorp. Caller need not check chain->error.
680 * cparent can wind up being anything.
682 * If datap is not NULL, *datap points to the real data we intend to write.
683 * If we can dedup the storage location we set *datap to NULL to indicate
684 * to the caller that a dedup occurred.
686 * NOTE: Special case for data embedded in inode.
690 hammer2_assign_physical(hammer2_inode_t *ip, hammer2_chain_t **parentp,
691 hammer2_key_t lbase, int pblksize,
692 hammer2_tid_t mtid, char **datap, int *errorp)
694 hammer2_chain_t *chain;
695 hammer2_key_t key_dummy;
696 hammer2_off_t dedup_off;
697 int pradix = hammer2_getradix(pblksize);
700 * Locate the chain associated with lbase, return a locked chain.
701 * However, do not instantiate any data reference (which utilizes a
702 * device buffer) because we will be using direct IO via the
703 * logical buffer cache buffer.
705 KKASSERT(pblksize >= HAMMER2_ALLOC_MIN);
707 chain = hammer2_chain_lookup(parentp, &key_dummy,
710 HAMMER2_LOOKUP_NODATA);
713 * The lookup code should not return a DELETED chain to us, unless
714 * its a short-file embedded in the inode. Then it is possible for
715 * the lookup to return a deleted inode.
717 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED) &&
718 chain->bref.type != HAMMER2_BREF_TYPE_INODE) {
719 kprintf("assign physical deleted chain @ "
720 "%016jx (%016jx.%02x) ip %016jx\n",
721 lbase, chain->bref.data_off, chain->bref.type,
728 * We found a hole, create a new chain entry.
730 * NOTE: DATA chains are created without device backing
731 * store (nor do we want any).
733 dedup_off = hammer2_dedup_lookup((*parentp)->hmp, datap,
735 *errorp = hammer2_chain_create(parentp, &chain,
737 HAMMER2_ENC_CHECK(ip->meta.check_algo) |
738 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE),
739 lbase, HAMMER2_PBUFRADIX,
740 HAMMER2_BREF_TYPE_DATA,
744 panic("hammer2_chain_create: par=%p error=%d\n",
748 /*ip->delta_dcount += pblksize;*/
749 } else if (chain->error == 0) {
750 switch (chain->bref.type) {
751 case HAMMER2_BREF_TYPE_INODE:
753 * The data is embedded in the inode, which requires
756 hammer2_chain_modify_ip(ip, chain, mtid, 0);
758 case HAMMER2_BREF_TYPE_DATA:
759 dedup_off = hammer2_dedup_lookup(chain->hmp, datap,
761 if (chain->bytes != pblksize) {
762 hammer2_chain_resize(chain,
765 HAMMER2_MODIFY_OPTDATA);
769 * DATA buffers must be marked modified whether the
770 * data is in a logical buffer or not. We also have
771 * to make this call to fixup the chain data pointers
772 * after resizing in case this is an encrypted or
775 hammer2_chain_modify(chain, mtid, dedup_off,
776 HAMMER2_MODIFY_OPTDATA);
779 panic("hammer2_assign_physical: bad type");
784 *errorp = chain->error;
790 * hammer2_write_file_core() - hammer2_write_thread() helper
792 * The core write function which determines which path to take
793 * depending on compression settings. We also have to locate the
794 * related chains so we can calculate and set the check data for
799 hammer2_write_file_core(char *data, hammer2_inode_t *ip,
800 hammer2_chain_t **parentp,
801 hammer2_key_t lbase, int ioflag, int pblksize,
802 hammer2_tid_t mtid, int *errorp)
804 hammer2_chain_t *chain;
809 switch(HAMMER2_DEC_ALGO(ip->meta.comp_algo)) {
810 case HAMMER2_COMP_NONE:
812 * We have to assign physical storage to the buffer
813 * we intend to dirty or write now to avoid deadlocks
814 * in the strategy code later.
816 * This can return NOOFFSET for inode-embedded data.
817 * The strategy code will take care of it in that case.
820 chain = hammer2_assign_physical(ip, parentp, lbase, pblksize,
821 mtid, &bdata, errorp);
823 /* skip modifications */
824 } else if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
825 hammer2_inode_data_t *wipdata;
827 wipdata = &chain->data->ipdata;
828 KKASSERT(wipdata->meta.op_flags &
829 HAMMER2_OPFLAG_DIRECTDATA);
830 bcopy(data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
831 ++hammer2_iod_file_wembed;
832 } else if (bdata == NULL) {
834 * Copy of data already present on-media.
836 chain->bref.methods =
837 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) +
838 HAMMER2_ENC_CHECK(ip->meta.check_algo);
839 hammer2_chain_setcheck(chain, data);
841 hammer2_write_bp(chain, data, ioflag, pblksize,
842 mtid, errorp, ip->meta.check_algo);
845 hammer2_chain_unlock(chain);
846 hammer2_chain_drop(chain);
849 case HAMMER2_COMP_AUTOZERO:
851 * Check for zero-fill only
853 hammer2_zero_check_and_write(data, ip, parentp,
854 lbase, ioflag, pblksize,
856 ip->meta.check_algo);
858 case HAMMER2_COMP_LZ4:
859 case HAMMER2_COMP_ZLIB:
862 * Check for zero-fill and attempt compression.
864 hammer2_compress_and_write(data, ip, parentp,
865 lbase, ioflag, pblksize,
868 ip->meta.check_algo);
876 * Generic function that will perform the compression in compression
877 * write path. The compression algorithm is determined by the settings
878 * obtained from inode.
882 hammer2_compress_and_write(char *data, hammer2_inode_t *ip,
883 hammer2_chain_t **parentp,
884 hammer2_key_t lbase, int ioflag, int pblksize,
885 hammer2_tid_t mtid, int *errorp, int comp_algo, int check_algo)
887 hammer2_chain_t *chain;
894 * An all-zeros write creates a hole unless the check code
895 * is disabled. When the check code is disabled all writes
896 * are done in-place, including any all-zeros writes.
898 * NOTE: A snapshot will still force a copy-on-write
899 * (see the HAMMER2_CHECK_NONE in hammer2_chain.c).
901 if (check_algo != HAMMER2_CHECK_NONE &&
902 test_block_zeros(data, pblksize)) {
903 zero_write(data, ip, parentp, lbase, mtid, errorp);
908 * Compression requested. Try to compress the block. We store
909 * the data normally if we cannot sufficiently compress it.
911 * We have a heuristic to detect files which are mostly
912 * uncompressable and avoid the compression attempt in that
913 * case. If the compression heuristic is turned off, we always
919 KKASSERT(pblksize / 2 <= 32768);
921 if (ip->comp_heuristic < 8 || (ip->comp_heuristic & 7) == 0 ||
922 hammer2_always_compress) {
923 z_stream strm_compress;
927 switch(HAMMER2_DEC_ALGO(comp_algo)) {
928 case HAMMER2_COMP_LZ4:
929 comp_buffer = objcache_get(cache_buffer_write,
931 comp_size = LZ4_compress_limitedOutput(
933 &comp_buffer[sizeof(int)],
935 pblksize / 2 - sizeof(int));
937 * We need to prefix with the size, LZ4
938 * doesn't do it for us. Add the related
941 *(int *)comp_buffer = comp_size;
943 comp_size += sizeof(int);
945 case HAMMER2_COMP_ZLIB:
946 comp_level = HAMMER2_DEC_LEVEL(comp_algo);
948 comp_level = 6; /* default zlib compression */
949 else if (comp_level < 6)
951 else if (comp_level > 9)
953 ret = deflateInit(&strm_compress, comp_level);
955 kprintf("HAMMER2 ZLIB: fatal error "
956 "on deflateInit.\n");
959 comp_buffer = objcache_get(cache_buffer_write,
961 strm_compress.next_in = data;
962 strm_compress.avail_in = pblksize;
963 strm_compress.next_out = comp_buffer;
964 strm_compress.avail_out = pblksize / 2;
965 ret = deflate(&strm_compress, Z_FINISH);
966 if (ret == Z_STREAM_END) {
967 comp_size = pblksize / 2 -
968 strm_compress.avail_out;
972 ret = deflateEnd(&strm_compress);
975 kprintf("Error: Unknown compression method.\n");
976 kprintf("Comp_method = %d.\n", comp_algo);
981 if (comp_size == 0) {
983 * compression failed or turned off
985 comp_block_size = pblksize; /* safety */
986 if (++ip->comp_heuristic > 128)
987 ip->comp_heuristic = 8;
990 * compression succeeded
992 ip->comp_heuristic = 0;
993 if (comp_size <= 1024) {
994 comp_block_size = 1024;
995 } else if (comp_size <= 2048) {
996 comp_block_size = 2048;
997 } else if (comp_size <= 4096) {
998 comp_block_size = 4096;
999 } else if (comp_size <= 8192) {
1000 comp_block_size = 8192;
1001 } else if (comp_size <= 16384) {
1002 comp_block_size = 16384;
1003 } else if (comp_size <= 32768) {
1004 comp_block_size = 32768;
1006 panic("hammer2: WRITE PATH: "
1007 "Weird comp_size value.");
1009 comp_block_size = pblksize;
1013 * Must zero the remainder or dedup (which operates on a
1014 * physical block basis) will not find matches.
1016 if (comp_size < comp_block_size) {
1017 bzero(comp_buffer + comp_size,
1018 comp_block_size - comp_size);
1023 * Assign physical storage, data will be set to NULL if a live-dedup
1026 bdata = comp_size ? comp_buffer : data;
1027 chain = hammer2_assign_physical(ip, parentp, lbase, comp_block_size,
1028 mtid, &bdata, errorp);
1031 kprintf("WRITE PATH: An error occurred while "
1032 "assigning physical space.\n");
1033 KKASSERT(chain == NULL);
1037 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
1038 hammer2_inode_data_t *wipdata;
1040 hammer2_chain_modify_ip(ip, chain, mtid, 0);
1041 wipdata = &chain->data->ipdata;
1042 KKASSERT(wipdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA);
1043 bcopy(data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
1044 ++hammer2_iod_file_wembed;
1045 } else if (bdata == NULL) {
1047 * Live deduplication, a copy of the data is already present
1051 chain->bref.methods =
1052 HAMMER2_ENC_COMP(comp_algo) +
1053 HAMMER2_ENC_CHECK(check_algo);
1055 chain->bref.methods =
1057 HAMMER2_COMP_NONE) +
1058 HAMMER2_ENC_CHECK(check_algo);
1060 bdata = comp_size ? comp_buffer : data;
1061 hammer2_chain_setcheck(chain, bdata);
1062 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1066 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
1068 switch(chain->bref.type) {
1069 case HAMMER2_BREF_TYPE_INODE:
1070 panic("hammer2_write_bp: unexpected inode\n");
1072 case HAMMER2_BREF_TYPE_DATA:
1074 * Optimize out the read-before-write
1077 *errorp = hammer2_io_newnz(chain->hmp,
1079 chain->bref.data_off,
1083 hammer2_io_brelse(&dio);
1084 kprintf("hammer2: WRITE PATH: "
1085 "dbp bread error\n");
1088 bdata = hammer2_io_data(dio, chain->bref.data_off);
1091 * When loading the block make sure we don't
1092 * leave garbage after the compressed data.
1095 chain->bref.methods =
1096 HAMMER2_ENC_COMP(comp_algo) +
1097 HAMMER2_ENC_CHECK(check_algo);
1098 bcopy(comp_buffer, bdata, comp_size);
1100 chain->bref.methods =
1102 HAMMER2_COMP_NONE) +
1103 HAMMER2_ENC_CHECK(check_algo);
1104 bcopy(data, bdata, pblksize);
1108 * The flush code doesn't calculate check codes for
1109 * file data (doing so can result in excessive I/O),
1112 hammer2_chain_setcheck(chain, bdata);
1115 * Device buffer is now valid, chain is no longer in
1116 * the initial state.
1118 * (No blockref table worries with file data)
1120 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1121 hammer2_dedup_record(chain, dio, bdata);
1123 /* Now write the related bdp. */
1124 if (ioflag & IO_SYNC) {
1126 * Synchronous I/O requested.
1128 hammer2_io_bwrite(&dio);
1130 } else if ((ioflag & IO_DIRECT) &&
1131 loff + n == pblksize) {
1132 hammer2_io_bdwrite(&dio);
1134 } else if (ioflag & IO_ASYNC) {
1135 hammer2_io_bawrite(&dio);
1137 hammer2_io_bdwrite(&dio);
1141 panic("hammer2_write_bp: bad chain type %d\n",
1149 hammer2_chain_unlock(chain);
1150 hammer2_chain_drop(chain);
1153 objcache_put(cache_buffer_write, comp_buffer);
1159 * Function that performs zero-checking and writing without compression,
1160 * it corresponds to default zero-checking path.
1164 hammer2_zero_check_and_write(char *data, hammer2_inode_t *ip,
1165 hammer2_chain_t **parentp,
1166 hammer2_key_t lbase, int ioflag, int pblksize,
1167 hammer2_tid_t mtid, int *errorp,
1170 hammer2_chain_t *chain;
1173 if (check_algo != HAMMER2_CHECK_NONE &&
1174 test_block_zeros(data, pblksize)) {
1176 * An all-zeros write creates a hole unless the check code
1177 * is disabled. When the check code is disabled all writes
1178 * are done in-place, including any all-zeros writes.
1180 * NOTE: A snapshot will still force a copy-on-write
1181 * (see the HAMMER2_CHECK_NONE in hammer2_chain.c).
1183 zero_write(data, ip, parentp, lbase, mtid, errorp);
1189 chain = hammer2_assign_physical(ip, parentp, lbase, pblksize,
1190 mtid, &bdata, errorp);
1194 hammer2_write_bp(chain, data, ioflag, pblksize,
1195 mtid, errorp, check_algo);
1197 /* dedup occurred */
1198 chain->bref.methods =
1199 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) +
1200 HAMMER2_ENC_CHECK(check_algo);
1201 hammer2_chain_setcheck(chain, data);
1204 hammer2_chain_unlock(chain);
1205 hammer2_chain_drop(chain);
1213 * A function to test whether a block of data contains only zeros,
1214 * returns TRUE (non-zero) if the block is all zeros.
1218 test_block_zeros(const char *buf, size_t bytes)
1222 for (i = 0; i < bytes; i += sizeof(long)) {
1223 if (*(const long *)(buf + i) != 0)
1232 * Function to "write" a block that contains only zeros.
1236 zero_write(char *data, hammer2_inode_t *ip,
1237 hammer2_chain_t **parentp,
1238 hammer2_key_t lbase, hammer2_tid_t mtid, int *errorp)
1240 hammer2_chain_t *chain;
1241 hammer2_key_t key_dummy;
1243 chain = hammer2_chain_lookup(parentp, &key_dummy,
1246 HAMMER2_LOOKUP_NODATA);
1248 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
1249 hammer2_inode_data_t *wipdata;
1251 hammer2_chain_modify_ip(ip, chain, mtid, 0);
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;
1258 hammer2_chain_delete(*parentp, chain,
1259 mtid, HAMMER2_DELETE_PERMANENT);
1260 ++hammer2_iod_file_wzero;
1262 hammer2_chain_unlock(chain);
1263 hammer2_chain_drop(chain);
1265 ++hammer2_iod_file_wzero;
1272 * Function to write the data as it is, without performing any sort of
1273 * compression. This function is used in path without compression and
1274 * default zero-checking path.
1278 hammer2_write_bp(hammer2_chain_t *chain, char *data, int ioflag,
1280 hammer2_tid_t mtid, int *errorp, int check_algo)
1282 hammer2_inode_data_t *wipdata;
1287 error = 0; /* XXX TODO below */
1289 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
1291 switch(chain->bref.type) {
1292 case HAMMER2_BREF_TYPE_INODE:
1293 wipdata = &chain->data->ipdata;
1294 KKASSERT(wipdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA);
1295 bcopy(data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
1297 ++hammer2_iod_file_wembed;
1299 case HAMMER2_BREF_TYPE_DATA:
1300 error = hammer2_io_newnz(chain->hmp,
1302 chain->bref.data_off,
1303 chain->bytes, &dio);
1305 hammer2_io_bqrelse(&dio);
1306 kprintf("hammer2: WRITE PATH: "
1307 "dbp bread error\n");
1310 bdata = hammer2_io_data(dio, chain->bref.data_off);
1312 chain->bref.methods = HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) +
1313 HAMMER2_ENC_CHECK(check_algo);
1314 bcopy(data, bdata, chain->bytes);
1317 * The flush code doesn't calculate check codes for
1318 * file data (doing so can result in excessive I/O),
1321 hammer2_chain_setcheck(chain, bdata);
1324 * Device buffer is now valid, chain is no longer in
1325 * the initial state.
1327 * (No blockref table worries with file data)
1329 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1330 hammer2_dedup_record(chain, dio, bdata);
1332 if (ioflag & IO_SYNC) {
1334 * Synchronous I/O requested.
1336 hammer2_io_bwrite(&dio);
1338 } else if ((ioflag & IO_DIRECT) &&
1339 loff + n == pblksize) {
1340 hammer2_io_bdwrite(&dio);
1342 } else if (ioflag & IO_ASYNC) {
1343 hammer2_io_bawrite(&dio);
1345 hammer2_io_bdwrite(&dio);
1349 panic("hammer2_write_bp: bad chain type %d\n",
1355 KKASSERT(error == 0); /* XXX TODO */
1360 * LIVE DEDUP HEURISTICS
1362 * Record media and crc information for possible dedup operation. Note
1363 * that the dedup mask bits must also be set in the related DIO for a dedup
1364 * to be fully validated (which is handled in the freemap allocation code).
1366 * WARNING! This code is SMP safe but the heuristic allows SMP collisions.
1367 * All fields must be loaded into locals and validated.
1369 * WARNING! Should only be used for file data and directory entries,
1370 * hammer2_chain_modify() only checks for the dedup case on data
1371 * chains. Also, dedup data can only be recorded for committed
1372 * chains (so NOT strategy writes which can undergo further
1373 * modification after the fact!).
1376 hammer2_dedup_record(hammer2_chain_t *chain, hammer2_io_t *dio, char *data)
1379 hammer2_dedup_t *dedup;
1387 * We can only record a dedup if we have media data to test against.
1388 * If dedup is not enabled, return early, which allows a chain to
1389 * remain marked MODIFIED (which might have benefits in special
1390 * situations, though typically it does not).
1392 if (hammer2_dedup_enable == 0)
1402 switch(HAMMER2_DEC_CHECK(chain->bref.methods)) {
1403 case HAMMER2_CHECK_ISCSI32:
1405 * XXX use the built-in crc (the dedup lookup sequencing
1406 * needs to be fixed so the check code is already present
1407 * when dedup_lookup is called)
1410 crc = (uint64_t)(uint32_t)chain->bref.check.iscsi32.value;
1412 crc = XXH64(data, chain->bytes, XXH_HAMMER2_SEED);
1414 case HAMMER2_CHECK_XXHASH64:
1415 crc = chain->bref.check.xxhash64.value;
1417 case HAMMER2_CHECK_SHA192:
1419 * XXX use the built-in crc (the dedup lookup sequencing
1420 * needs to be fixed so the check code is already present
1421 * when dedup_lookup is called)
1424 crc = ((uint64_t *)chain->bref.check.sha192.data)[0] ^
1425 ((uint64_t *)chain->bref.check.sha192.data)[1] ^
1426 ((uint64_t *)chain->bref.check.sha192.data)[2];
1428 crc = XXH64(data, chain->bytes, XXH_HAMMER2_SEED);
1432 * Cannot dedup without a check code
1434 * NOTE: In particular, CHECK_NONE allows a sector to be
1435 * overwritten without copy-on-write, recording
1436 * a dedup block for a CHECK_NONE object would be
1442 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DEDUPABLE);
1444 dedup = &hmp->heur_dedup[crc & (HAMMER2_DEDUP_HEUR_MASK & ~3)];
1445 for (i = 0; i < 4; ++i) {
1446 if (dedup[i].data_crc == crc) {
1450 dticks = (int)(dedup[i].ticks - dedup[best].ticks);
1451 if (dticks < 0 || dticks > hz * 60 * 30)
1455 if (hammer2_debug & 0x40000) {
1456 kprintf("REC %04x %016jx %016jx\n",
1457 (int)(dedup - hmp->heur_dedup),
1459 chain->bref.data_off);
1461 dedup->ticks = ticks;
1462 dedup->data_off = chain->bref.data_off;
1463 dedup->data_crc = crc;
1466 * Set the valid bits for the dedup only after we know the data
1467 * buffer has been updated. The alloc bits were set (and the valid
1468 * bits cleared) when the media was allocated.
1470 * This is done in two stages becuase the bulkfree code can race
1471 * the gap between allocation and data population. Both masks must
1472 * be set before a bcmp/dedup operation is able to use the block.
1474 mask = hammer2_dedup_mask(dio, chain->bref.data_off, chain->bytes);
1475 atomic_set_64(&dio->dedup_valid, mask);
1479 * XXX removed. MODIFIED is an integral part of the flush code,
1480 * lets not just clear it
1483 * Once we record the dedup the chain must be marked clean to
1484 * prevent reuse of the underlying block. Remember that this
1485 * write occurs when the buffer cache is flushed (i.e. on sync(),
1486 * fsync(), filesystem periodic sync, or when the kernel needs to
1487 * flush a buffer), and not whenever the user write()s.
1489 if (chain->flags & HAMMER2_CHAIN_MODIFIED) {
1490 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
1491 atomic_add_long(&hammer2_count_modified_chains, -1);
1493 hammer2_pfs_memory_wakeup(chain->pmp);
1500 hammer2_dedup_lookup(hammer2_dev_t *hmp, char **datap, int pblksize)
1502 hammer2_dedup_t *dedup;
1511 if (hammer2_dedup_enable == 0)
1518 * XXX use the built-in crc (the dedup lookup sequencing
1519 * needs to be fixed so the check code is already present
1520 * when dedup_lookup is called)
1522 crc = XXH64(data, pblksize, XXH_HAMMER2_SEED);
1523 dedup = &hmp->heur_dedup[crc & (HAMMER2_DEDUP_HEUR_MASK & ~3)];
1525 if (hammer2_debug & 0x40000) {
1526 kprintf("LOC %04x/4 %016jx\n",
1527 (int)(dedup - hmp->heur_dedup),
1531 for (i = 0; i < 4; ++i) {
1532 off = dedup[i].data_off;
1534 if (dedup[i].data_crc != crc)
1536 if ((1 << (int)(off & HAMMER2_OFF_MASK_RADIX)) != pblksize)
1538 dio = hammer2_io_getquick(hmp, off, pblksize);
1540 dtmp = hammer2_io_data(dio, off),
1541 mask = hammer2_dedup_mask(dio, off, pblksize);
1542 if ((dio->dedup_alloc & mask) == mask &&
1543 (dio->dedup_valid & mask) == mask &&
1544 bcmp(data, dtmp, pblksize) == 0) {
1545 if (hammer2_debug & 0x40000) {
1546 kprintf("DEDUP SUCCESS %016jx\n",
1549 hammer2_io_putblk(&dio);
1551 dedup[i].ticks = ticks; /* update use */
1552 atomic_add_long(&hammer2_iod_file_wdedup,
1555 return off; /* RETURN */
1557 hammer2_io_putblk(&dio);
1564 * Poof. Races are ok, if someone gets in and reuses a dedup offset
1565 * before or while we are clearing it they will also recover the freemap
1566 * entry (set it to fully allocated), so a bulkfree race can only set it
1567 * to a possibly-free state.
1569 * XXX ok, well, not really sure races are ok but going to run with it
1573 hammer2_dedup_clear(hammer2_dev_t *hmp)
1577 for (i = 0; i < HAMMER2_DEDUP_HEUR_SIZE; ++i) {
1578 hmp->heur_dedup[i].data_off = 0;
1579 hmp->heur_dedup[i].ticks = ticks - 1;