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;
328 error = HAMMER2_ERROR_EIO;
331 error = hammer2_xop_feed(&xop->head, chain, thr->clindex, error);
333 hammer2_chain_unlock(chain);
334 hammer2_chain_drop(chain);
337 hammer2_chain_unlock(parent);
338 hammer2_chain_drop(parent);
340 chain = NULL; /* safety */
341 parent = NULL; /* safety */
344 * Race to finish the frontend. First-to-complete. bio is only
345 * valid if we are determined to be the ones able to complete
350 hammer2_mtx_ex(&xop->lock);
352 hammer2_mtx_unlock(&xop->lock);
359 * Async operation has not completed and we now own the lock.
360 * Determine if we can complete the operation by issuing the
361 * frontend collection non-blocking.
363 * H2 double-buffers the data, setting B_NOTMETA on the logical
364 * buffer hints to the OS that the logical buffer should not be
365 * swapcached (since the device buffer can be).
367 * Also note that even for compressed data we would rather the
368 * kernel cache/swapcache device buffers more and (decompressed)
369 * logical buffers less, since that will significantly improve
370 * the amount of end-user data that can be cached.
372 error = hammer2_xop_collect(&xop->head, HAMMER2_XOP_COLLECT_NOWAIT);
377 hammer2_mtx_unlock(&xop->lock);
378 bp->b_flags |= B_NOTMETA;
379 chain = xop->head.cluster.focus;
380 hammer2_strategy_read_completion(chain, (char *)chain->data,
383 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
385 case HAMMER2_ERROR_ENOENT:
387 hammer2_mtx_unlock(&xop->lock);
388 bp->b_flags |= B_NOTMETA;
391 bzero(bp->b_data, bp->b_bcount);
393 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
395 case HAMMER2_ERROR_EINPROGRESS:
396 hammer2_mtx_unlock(&xop->lock);
399 kprintf("strategy_xop_read: error %08x loff=%016jx\n",
400 error, bp->b_loffset);
402 hammer2_mtx_unlock(&xop->lock);
403 bp->b_flags |= B_ERROR;
406 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
413 hammer2_strategy_read_completion(hammer2_chain_t *chain, char *data,
416 struct buf *bp = bio->bio_buf;
418 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
420 * Copy from in-memory inode structure.
422 bcopy(((hammer2_inode_data_t *)data)->u.data,
423 bp->b_data, HAMMER2_EMBEDDED_BYTES);
424 bzero(bp->b_data + HAMMER2_EMBEDDED_BYTES,
425 bp->b_bcount - HAMMER2_EMBEDDED_BYTES);
428 } else if (chain->bref.type == HAMMER2_BREF_TYPE_DATA) {
430 * Data is on-media, record for live dedup. Release the
431 * chain (try to free it) when done. The data is still
432 * cached by both the buffer cache in front and the
433 * block device behind us. This leaves more room in the
434 * LRU chain cache for meta-data chains which we really
437 * NOTE: Deduplication cannot be safely recorded for
438 * records without a check code.
440 hammer2_dedup_record(chain, NULL, data);
441 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
444 * Decompression and copy.
446 switch (HAMMER2_DEC_COMP(chain->bref.methods)) {
447 case HAMMER2_COMP_LZ4:
448 hammer2_decompress_LZ4_callback(data, chain->bytes,
450 /* b_resid set by call */
452 case HAMMER2_COMP_ZLIB:
453 hammer2_decompress_ZLIB_callback(data, chain->bytes,
455 /* b_resid set by call */
457 case HAMMER2_COMP_NONE:
458 KKASSERT(chain->bytes <= bp->b_bcount);
459 bcopy(data, bp->b_data, chain->bytes);
460 if (chain->bytes < bp->b_bcount) {
461 bzero(bp->b_data + chain->bytes,
462 bp->b_bcount - chain->bytes);
468 panic("hammer2_strategy_read: "
469 "unknown compression type");
472 panic("hammer2_strategy_read: unknown bref type");
476 /****************************************************************************
478 ****************************************************************************/
481 * Functions for compression in threads,
482 * from hammer2_vnops.c
484 static void hammer2_write_file_core(char *data, hammer2_inode_t *ip,
485 hammer2_chain_t **parentp,
486 hammer2_key_t lbase, int ioflag, int pblksize,
487 hammer2_tid_t mtid, int *errorp);
488 static void hammer2_compress_and_write(char *data, hammer2_inode_t *ip,
489 hammer2_chain_t **parentp,
490 hammer2_key_t lbase, int ioflag, int pblksize,
491 hammer2_tid_t mtid, int *errorp,
492 int comp_algo, int check_algo);
493 static void hammer2_zero_check_and_write(char *data, hammer2_inode_t *ip,
494 hammer2_chain_t **parentp,
495 hammer2_key_t lbase, int ioflag, int pblksize,
496 hammer2_tid_t mtid, int *errorp,
498 static int test_block_zeros(const char *buf, size_t bytes);
499 static void zero_write(char *data, hammer2_inode_t *ip,
500 hammer2_chain_t **parentp,
502 hammer2_tid_t mtid, int *errorp);
503 static void hammer2_write_bp(hammer2_chain_t *chain, char *data,
504 int ioflag, int pblksize,
505 hammer2_tid_t mtid, int *errorp,
510 hammer2_strategy_write(struct vop_strategy_args *ap)
512 hammer2_xop_strategy_t *xop;
523 hammer2_lwinprog_ref(pmp);
524 hammer2_trans_assert_strategy(pmp);
525 hammer2_trans_init(pmp, HAMMER2_TRANS_BUFCACHE);
527 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING |
528 HAMMER2_XOP_STRATEGY);
531 xop->lbase = bio->bio_offset;
532 hammer2_mtx_init(&xop->lock, "h2biow");
533 hammer2_xop_start(&xop->head, hammer2_strategy_xop_write);
534 /* asynchronous completion */
536 hammer2_lwinprog_wait(pmp, hammer2_flush_pipe);
542 * Per-node XOP (threaded). Write the logical buffer to the media.
544 * This is a bit problematic because there may be multiple target and
545 * any of them may be able to release the bp. In addition, if our
546 * particulr target is offline we don't want to block the bp (and thus
547 * the frontend). To accomplish this we copy the data to the per-thr
552 hammer2_strategy_xop_write(hammer2_thread_t *thr, hammer2_xop_t *arg)
554 hammer2_xop_strategy_t *xop = &arg->xop_strategy;
555 hammer2_chain_t *parent;
563 hammer2_off_t bio_offset;
567 * We can only access the bp/bio if the frontend has not yet
572 hammer2_mtx_sh(&xop->lock);
574 hammer2_mtx_unlock(&xop->lock);
579 bio = xop->bio; /* ephermal */
580 bp = bio->bio_buf; /* ephermal */
581 ip = xop->head.ip1; /* retained by ref */
582 bio_offset = bio->bio_offset;
583 bio_data = thr->scratch;
585 /* hammer2_trans_init(parent->hmp->spmp, HAMMER2_TRANS_BUFCACHE); */
587 lblksize = hammer2_calc_logical(ip, bio->bio_offset, &lbase, NULL);
588 pblksize = hammer2_calc_physical(ip, lbase);
589 bcopy(bp->b_data, bio_data, lblksize);
591 hammer2_mtx_unlock(&xop->lock);
592 bp = NULL; /* safety, illegal to access after unlock */
593 bio = NULL; /* safety, illegal to access after unlock */
598 parent = hammer2_inode_chain(ip, thr->clindex, HAMMER2_RESOLVE_ALWAYS);
599 hammer2_write_file_core(bio_data, ip, &parent,
600 lbase, IO_ASYNC, pblksize,
601 xop->head.mtid, &error);
603 hammer2_chain_unlock(parent);
604 hammer2_chain_drop(parent);
605 parent = NULL; /* safety */
607 hammer2_xop_feed(&xop->head, NULL, thr->clindex, error);
610 * Try to complete the operation on behalf of the front-end.
614 hammer2_mtx_ex(&xop->lock);
616 hammer2_mtx_unlock(&xop->lock);
621 * Async operation has not completed and we now own the lock.
622 * Determine if we can complete the operation by issuing the
623 * frontend collection non-blocking.
625 * H2 double-buffers the data, setting B_NOTMETA on the logical
626 * buffer hints to the OS that the logical buffer should not be
627 * swapcached (since the device buffer can be).
629 error = hammer2_xop_collect(&xop->head, HAMMER2_XOP_COLLECT_NOWAIT);
631 if (error == HAMMER2_ERROR_EINPROGRESS) {
632 hammer2_mtx_unlock(&xop->lock);
637 * Async operation has completed.
640 hammer2_mtx_unlock(&xop->lock);
642 bio = xop->bio; /* now owned by us */
643 bp = bio->bio_buf; /* now owned by us */
645 if (error == HAMMER2_ERROR_ENOENT || error == 0) {
646 bp->b_flags |= B_NOTMETA;
651 kprintf("strategy_xop_write: error %d loff=%016jx\n",
652 error, bp->b_loffset);
653 bp->b_flags |= B_ERROR;
657 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
658 hammer2_trans_assert_strategy(ip->pmp);
659 hammer2_lwinprog_drop(ip->pmp);
660 hammer2_trans_done(ip->pmp);
664 * Wait for pending I/O to complete
667 hammer2_bioq_sync(hammer2_pfs_t *pmp)
669 hammer2_lwinprog_wait(pmp, 0);
673 * Assign physical storage at (cparent, lbase), returning a suitable chain
674 * and setting *errorp appropriately.
676 * If no error occurs, the returned chain will be in a modified state.
678 * If an error occurs, the returned chain may or may not be NULL. If
679 * not-null any chain->error (if not 0) will also be rolled up into *errorp.
680 * So the caller only needs to test *errorp.
682 * cparent can wind up being anything.
684 * If datap is not NULL, *datap points to the real data we intend to write.
685 * If we can dedup the storage location we set *datap to NULL to indicate
686 * to the caller that a dedup occurred.
688 * NOTE: Special case for data embedded in inode.
692 hammer2_assign_physical(hammer2_inode_t *ip, hammer2_chain_t **parentp,
693 hammer2_key_t lbase, int pblksize,
694 hammer2_tid_t mtid, char **datap, int *errorp)
696 hammer2_chain_t *chain;
697 hammer2_key_t key_dummy;
698 hammer2_off_t dedup_off;
699 int pradix = hammer2_getradix(pblksize);
702 * Locate the chain associated with lbase, return a locked chain.
703 * However, do not instantiate any data reference (which utilizes a
704 * device buffer) because we will be using direct IO via the
705 * logical buffer cache buffer.
707 KKASSERT(pblksize >= HAMMER2_ALLOC_MIN);
709 chain = hammer2_chain_lookup(parentp, &key_dummy,
712 HAMMER2_LOOKUP_NODATA);
715 * The lookup code should not return a DELETED chain to us, unless
716 * its a short-file embedded in the inode. Then it is possible for
717 * the lookup to return a deleted inode.
719 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED) &&
720 chain->bref.type != HAMMER2_BREF_TYPE_INODE) {
721 kprintf("assign physical deleted chain @ "
722 "%016jx (%016jx.%02x) ip %016jx\n",
723 lbase, chain->bref.data_off, chain->bref.type,
730 * We found a hole, create a new chain entry.
732 * NOTE: DATA chains are created without device backing
733 * store (nor do we want any).
735 dedup_off = hammer2_dedup_lookup((*parentp)->hmp, datap,
737 *errorp |= hammer2_chain_create(parentp, &chain,
739 HAMMER2_ENC_CHECK(ip->meta.check_algo) |
740 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE),
741 lbase, HAMMER2_PBUFRADIX,
742 HAMMER2_BREF_TYPE_DATA,
747 /*ip->delta_dcount += pblksize;*/
748 } else if (chain->error == 0) {
749 switch (chain->bref.type) {
750 case HAMMER2_BREF_TYPE_INODE:
752 * The data is embedded in the inode, which requires
755 *errorp |= hammer2_chain_modify_ip(ip, chain, mtid, 0);
757 case HAMMER2_BREF_TYPE_DATA:
758 dedup_off = hammer2_dedup_lookup(chain->hmp, datap,
760 if (chain->bytes != pblksize) {
761 *errorp |= hammer2_chain_resize(chain,
764 HAMMER2_MODIFY_OPTDATA);
770 * DATA buffers must be marked modified whether the
771 * data is in a logical buffer or not. We also have
772 * to make this call to fixup the chain data pointers
773 * after resizing in case this is an encrypted or
776 *errorp |= hammer2_chain_modify(chain, mtid, dedup_off,
777 HAMMER2_MODIFY_OPTDATA);
780 panic("hammer2_assign_physical: bad type");
785 *errorp = chain->error;
792 * hammer2_write_file_core() - hammer2_write_thread() helper
794 * The core write function which determines which path to take
795 * depending on compression settings. We also have to locate the
796 * related chains so we can calculate and set the check data for
801 hammer2_write_file_core(char *data, hammer2_inode_t *ip,
802 hammer2_chain_t **parentp,
803 hammer2_key_t lbase, int ioflag, int pblksize,
804 hammer2_tid_t mtid, int *errorp)
806 hammer2_chain_t *chain;
811 switch(HAMMER2_DEC_ALGO(ip->meta.comp_algo)) {
812 case HAMMER2_COMP_NONE:
814 * We have to assign physical storage to the buffer
815 * we intend to dirty or write now to avoid deadlocks
816 * in the strategy code later.
818 * This can return NOOFFSET for inode-embedded data.
819 * The strategy code will take care of it in that case.
822 chain = hammer2_assign_physical(ip, parentp, lbase, pblksize,
823 mtid, &bdata, errorp);
825 /* skip modifications */
826 } else if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
827 hammer2_inode_data_t *wipdata;
829 wipdata = &chain->data->ipdata;
830 KKASSERT(wipdata->meta.op_flags &
831 HAMMER2_OPFLAG_DIRECTDATA);
832 bcopy(data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
833 ++hammer2_iod_file_wembed;
834 } else if (bdata == NULL) {
836 * Copy of data already present on-media.
838 chain->bref.methods =
839 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) +
840 HAMMER2_ENC_CHECK(ip->meta.check_algo);
841 hammer2_chain_setcheck(chain, data);
843 hammer2_write_bp(chain, data, ioflag, pblksize,
844 mtid, errorp, ip->meta.check_algo);
847 hammer2_chain_unlock(chain);
848 hammer2_chain_drop(chain);
851 case HAMMER2_COMP_AUTOZERO:
853 * Check for zero-fill only
855 hammer2_zero_check_and_write(data, ip, parentp,
856 lbase, ioflag, pblksize,
858 ip->meta.check_algo);
860 case HAMMER2_COMP_LZ4:
861 case HAMMER2_COMP_ZLIB:
864 * Check for zero-fill and attempt compression.
866 hammer2_compress_and_write(data, ip, parentp,
867 lbase, ioflag, pblksize,
870 ip->meta.check_algo);
878 * Generic function that will perform the compression in compression
879 * write path. The compression algorithm is determined by the settings
880 * obtained from inode.
884 hammer2_compress_and_write(char *data, hammer2_inode_t *ip,
885 hammer2_chain_t **parentp,
886 hammer2_key_t lbase, int ioflag, int pblksize,
887 hammer2_tid_t mtid, int *errorp, int comp_algo, int check_algo)
889 hammer2_chain_t *chain;
896 * An all-zeros write creates a hole unless the check code
897 * is disabled. When the check code is disabled all writes
898 * are done in-place, including any all-zeros writes.
900 * NOTE: A snapshot will still force a copy-on-write
901 * (see the HAMMER2_CHECK_NONE in hammer2_chain.c).
903 if (check_algo != HAMMER2_CHECK_NONE &&
904 test_block_zeros(data, pblksize)) {
905 zero_write(data, ip, parentp, lbase, mtid, errorp);
910 * Compression requested. Try to compress the block. We store
911 * the data normally if we cannot sufficiently compress it.
913 * We have a heuristic to detect files which are mostly
914 * uncompressable and avoid the compression attempt in that
915 * case. If the compression heuristic is turned off, we always
921 KKASSERT(pblksize / 2 <= 32768);
923 if (ip->comp_heuristic < 8 || (ip->comp_heuristic & 7) == 0 ||
924 hammer2_always_compress) {
925 z_stream strm_compress;
929 switch(HAMMER2_DEC_ALGO(comp_algo)) {
930 case HAMMER2_COMP_LZ4:
932 * We need to prefix with the size, LZ4
933 * doesn't do it for us. Add the related
936 * NOTE: The LZ4 code seems to assume at least an
937 * 8-byte buffer size granularity and may
938 * overrun the buffer if given a 4-byte
941 comp_buffer = objcache_get(cache_buffer_write,
943 comp_size = LZ4_compress_limitedOutput(
945 &comp_buffer[sizeof(int)],
947 pblksize / 2 - sizeof(int64_t));
948 *(int *)comp_buffer = comp_size;
950 comp_size += sizeof(int);
952 case HAMMER2_COMP_ZLIB:
953 comp_level = HAMMER2_DEC_LEVEL(comp_algo);
955 comp_level = 6; /* default zlib compression */
956 else if (comp_level < 6)
958 else if (comp_level > 9)
960 ret = deflateInit(&strm_compress, comp_level);
962 kprintf("HAMMER2 ZLIB: fatal error "
963 "on deflateInit.\n");
966 comp_buffer = objcache_get(cache_buffer_write,
968 strm_compress.next_in = data;
969 strm_compress.avail_in = pblksize;
970 strm_compress.next_out = comp_buffer;
971 strm_compress.avail_out = pblksize / 2;
972 ret = deflate(&strm_compress, Z_FINISH);
973 if (ret == Z_STREAM_END) {
974 comp_size = pblksize / 2 -
975 strm_compress.avail_out;
979 ret = deflateEnd(&strm_compress);
982 kprintf("Error: Unknown compression method.\n");
983 kprintf("Comp_method = %d.\n", comp_algo);
988 if (comp_size == 0) {
990 * compression failed or turned off
992 comp_block_size = pblksize; /* safety */
993 if (++ip->comp_heuristic > 128)
994 ip->comp_heuristic = 8;
997 * compression succeeded
999 ip->comp_heuristic = 0;
1000 if (comp_size <= 1024) {
1001 comp_block_size = 1024;
1002 } else if (comp_size <= 2048) {
1003 comp_block_size = 2048;
1004 } else if (comp_size <= 4096) {
1005 comp_block_size = 4096;
1006 } else if (comp_size <= 8192) {
1007 comp_block_size = 8192;
1008 } else if (comp_size <= 16384) {
1009 comp_block_size = 16384;
1010 } else if (comp_size <= 32768) {
1011 comp_block_size = 32768;
1013 panic("hammer2: WRITE PATH: "
1014 "Weird comp_size value.");
1016 comp_block_size = pblksize;
1020 * Must zero the remainder or dedup (which operates on a
1021 * physical block basis) will not find matches.
1023 if (comp_size < comp_block_size) {
1024 bzero(comp_buffer + comp_size,
1025 comp_block_size - comp_size);
1030 * Assign physical storage, data will be set to NULL if a live-dedup
1033 bdata = comp_size ? comp_buffer : data;
1034 chain = hammer2_assign_physical(ip, parentp, lbase, comp_block_size,
1035 mtid, &bdata, errorp);
1041 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
1042 hammer2_inode_data_t *wipdata;
1044 *errorp = hammer2_chain_modify_ip(ip, chain, mtid, 0);
1046 wipdata = &chain->data->ipdata;
1047 KKASSERT(wipdata->meta.op_flags &
1048 HAMMER2_OPFLAG_DIRECTDATA);
1049 bcopy(data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
1050 ++hammer2_iod_file_wembed;
1052 } else if (bdata == NULL) {
1054 * Live deduplication, a copy of the data is already present
1058 chain->bref.methods =
1059 HAMMER2_ENC_COMP(comp_algo) +
1060 HAMMER2_ENC_CHECK(check_algo);
1062 chain->bref.methods =
1064 HAMMER2_COMP_NONE) +
1065 HAMMER2_ENC_CHECK(check_algo);
1067 bdata = comp_size ? comp_buffer : data;
1068 hammer2_chain_setcheck(chain, bdata);
1069 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1073 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
1075 switch(chain->bref.type) {
1076 case HAMMER2_BREF_TYPE_INODE:
1077 panic("hammer2_write_bp: unexpected inode\n");
1079 case HAMMER2_BREF_TYPE_DATA:
1081 * Optimize out the read-before-write
1084 *errorp = hammer2_io_newnz(chain->hmp,
1086 chain->bref.data_off,
1090 hammer2_io_brelse(&dio);
1091 kprintf("hammer2: WRITE PATH: "
1092 "dbp bread error\n");
1095 bdata = hammer2_io_data(dio, chain->bref.data_off);
1098 * When loading the block make sure we don't
1099 * leave garbage after the compressed data.
1102 chain->bref.methods =
1103 HAMMER2_ENC_COMP(comp_algo) +
1104 HAMMER2_ENC_CHECK(check_algo);
1105 bcopy(comp_buffer, bdata, comp_size);
1107 chain->bref.methods =
1109 HAMMER2_COMP_NONE) +
1110 HAMMER2_ENC_CHECK(check_algo);
1111 bcopy(data, bdata, pblksize);
1115 * The flush code doesn't calculate check codes for
1116 * file data (doing so can result in excessive I/O),
1119 hammer2_chain_setcheck(chain, bdata);
1122 * Device buffer is now valid, chain is no longer in
1123 * the initial state.
1125 * (No blockref table worries with file data)
1127 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1128 hammer2_dedup_record(chain, dio, bdata);
1130 /* Now write the related bdp. */
1131 if (ioflag & IO_SYNC) {
1133 * Synchronous I/O requested.
1135 hammer2_io_bwrite(&dio);
1137 } else if ((ioflag & IO_DIRECT) &&
1138 loff + n == pblksize) {
1139 hammer2_io_bdwrite(&dio);
1141 } else if (ioflag & IO_ASYNC) {
1142 hammer2_io_bawrite(&dio);
1144 hammer2_io_bdwrite(&dio);
1148 panic("hammer2_write_bp: bad chain type %d\n",
1156 hammer2_chain_unlock(chain);
1157 hammer2_chain_drop(chain);
1160 objcache_put(cache_buffer_write, comp_buffer);
1166 * Function that performs zero-checking and writing without compression,
1167 * it corresponds to default zero-checking path.
1171 hammer2_zero_check_and_write(char *data, hammer2_inode_t *ip,
1172 hammer2_chain_t **parentp,
1173 hammer2_key_t lbase, int ioflag, int pblksize,
1174 hammer2_tid_t mtid, int *errorp,
1177 hammer2_chain_t *chain;
1180 if (check_algo != HAMMER2_CHECK_NONE &&
1181 test_block_zeros(data, pblksize)) {
1183 * An all-zeros write creates a hole unless the check code
1184 * is disabled. When the check code is disabled all writes
1185 * are done in-place, including any all-zeros writes.
1187 * NOTE: A snapshot will still force a copy-on-write
1188 * (see the HAMMER2_CHECK_NONE in hammer2_chain.c).
1190 zero_write(data, ip, parentp, lbase, mtid, errorp);
1196 chain = hammer2_assign_physical(ip, parentp, lbase, pblksize,
1197 mtid, &bdata, errorp);
1201 hammer2_write_bp(chain, data, ioflag, pblksize,
1202 mtid, errorp, check_algo);
1204 /* dedup occurred */
1205 chain->bref.methods =
1206 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) +
1207 HAMMER2_ENC_CHECK(check_algo);
1208 hammer2_chain_setcheck(chain, data);
1211 hammer2_chain_unlock(chain);
1212 hammer2_chain_drop(chain);
1220 * A function to test whether a block of data contains only zeros,
1221 * returns TRUE (non-zero) if the block is all zeros.
1225 test_block_zeros(const char *buf, size_t bytes)
1229 for (i = 0; i < bytes; i += sizeof(long)) {
1230 if (*(const long *)(buf + i) != 0)
1239 * Function to "write" a block that contains only zeros.
1243 zero_write(char *data, hammer2_inode_t *ip,
1244 hammer2_chain_t **parentp,
1245 hammer2_key_t lbase, hammer2_tid_t mtid, int *errorp)
1247 hammer2_chain_t *chain;
1248 hammer2_key_t key_dummy;
1250 chain = hammer2_chain_lookup(parentp, &key_dummy,
1253 HAMMER2_LOOKUP_NODATA);
1255 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
1256 hammer2_inode_data_t *wipdata;
1259 *errorp = hammer2_chain_modify_ip(ip, chain,
1263 wipdata = &chain->data->ipdata;
1264 KKASSERT(wipdata->meta.op_flags &
1265 HAMMER2_OPFLAG_DIRECTDATA);
1266 bzero(wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
1267 ++hammer2_iod_file_wembed;
1270 /* chain->error ok for deletion */
1271 hammer2_chain_delete(*parentp, chain,
1272 mtid, HAMMER2_DELETE_PERMANENT);
1273 ++hammer2_iod_file_wzero;
1275 hammer2_chain_unlock(chain);
1276 hammer2_chain_drop(chain);
1278 ++hammer2_iod_file_wzero;
1285 * Function to write the data as it is, without performing any sort of
1286 * compression. This function is used in path without compression and
1287 * default zero-checking path.
1291 hammer2_write_bp(hammer2_chain_t *chain, char *data, int ioflag,
1293 hammer2_tid_t mtid, int *errorp, int check_algo)
1295 hammer2_inode_data_t *wipdata;
1300 error = 0; /* XXX TODO below */
1302 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
1304 switch(chain->bref.type) {
1305 case HAMMER2_BREF_TYPE_INODE:
1306 wipdata = &chain->data->ipdata;
1307 KKASSERT(wipdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA);
1308 bcopy(data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
1310 ++hammer2_iod_file_wembed;
1312 case HAMMER2_BREF_TYPE_DATA:
1313 error = hammer2_io_newnz(chain->hmp,
1315 chain->bref.data_off,
1316 chain->bytes, &dio);
1318 hammer2_io_bqrelse(&dio);
1319 kprintf("hammer2: WRITE PATH: "
1320 "dbp bread error\n");
1323 bdata = hammer2_io_data(dio, chain->bref.data_off);
1325 chain->bref.methods = HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) +
1326 HAMMER2_ENC_CHECK(check_algo);
1327 bcopy(data, bdata, chain->bytes);
1330 * The flush code doesn't calculate check codes for
1331 * file data (doing so can result in excessive I/O),
1334 hammer2_chain_setcheck(chain, bdata);
1337 * Device buffer is now valid, chain is no longer in
1338 * the initial state.
1340 * (No blockref table worries with file data)
1342 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1343 hammer2_dedup_record(chain, dio, bdata);
1345 if (ioflag & IO_SYNC) {
1347 * Synchronous I/O requested.
1349 hammer2_io_bwrite(&dio);
1351 } else if ((ioflag & IO_DIRECT) &&
1352 loff + n == pblksize) {
1353 hammer2_io_bdwrite(&dio);
1355 } else if (ioflag & IO_ASYNC) {
1356 hammer2_io_bawrite(&dio);
1358 hammer2_io_bdwrite(&dio);
1362 panic("hammer2_write_bp: bad chain type %d\n",
1372 * LIVE DEDUP HEURISTICS
1374 * Record media and crc information for possible dedup operation. Note
1375 * that the dedup mask bits must also be set in the related DIO for a dedup
1376 * to be fully validated (which is handled in the freemap allocation code).
1378 * WARNING! This code is SMP safe but the heuristic allows SMP collisions.
1379 * All fields must be loaded into locals and validated.
1381 * WARNING! Should only be used for file data and directory entries,
1382 * hammer2_chain_modify() only checks for the dedup case on data
1383 * chains. Also, dedup data can only be recorded for committed
1384 * chains (so NOT strategy writes which can undergo further
1385 * modification after the fact!).
1388 hammer2_dedup_record(hammer2_chain_t *chain, hammer2_io_t *dio, char *data)
1391 hammer2_dedup_t *dedup;
1399 * We can only record a dedup if we have media data to test against.
1400 * If dedup is not enabled, return early, which allows a chain to
1401 * remain marked MODIFIED (which might have benefits in special
1402 * situations, though typically it does not).
1404 if (hammer2_dedup_enable == 0)
1414 switch(HAMMER2_DEC_CHECK(chain->bref.methods)) {
1415 case HAMMER2_CHECK_ISCSI32:
1417 * XXX use the built-in crc (the dedup lookup sequencing
1418 * needs to be fixed so the check code is already present
1419 * when dedup_lookup is called)
1422 crc = (uint64_t)(uint32_t)chain->bref.check.iscsi32.value;
1424 crc = XXH64(data, chain->bytes, XXH_HAMMER2_SEED);
1426 case HAMMER2_CHECK_XXHASH64:
1427 crc = chain->bref.check.xxhash64.value;
1429 case HAMMER2_CHECK_SHA192:
1431 * XXX use the built-in crc (the dedup lookup sequencing
1432 * needs to be fixed so the check code is already present
1433 * when dedup_lookup is called)
1436 crc = ((uint64_t *)chain->bref.check.sha192.data)[0] ^
1437 ((uint64_t *)chain->bref.check.sha192.data)[1] ^
1438 ((uint64_t *)chain->bref.check.sha192.data)[2];
1440 crc = XXH64(data, chain->bytes, XXH_HAMMER2_SEED);
1444 * Cannot dedup without a check code
1446 * NOTE: In particular, CHECK_NONE allows a sector to be
1447 * overwritten without copy-on-write, recording
1448 * a dedup block for a CHECK_NONE object would be
1454 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DEDUPABLE);
1456 dedup = &hmp->heur_dedup[crc & (HAMMER2_DEDUP_HEUR_MASK & ~3)];
1457 for (i = 0; i < 4; ++i) {
1458 if (dedup[i].data_crc == crc) {
1462 dticks = (int)(dedup[i].ticks - dedup[best].ticks);
1463 if (dticks < 0 || dticks > hz * 60 * 30)
1467 if (hammer2_debug & 0x40000) {
1468 kprintf("REC %04x %016jx %016jx\n",
1469 (int)(dedup - hmp->heur_dedup),
1471 chain->bref.data_off);
1473 dedup->ticks = ticks;
1474 dedup->data_off = chain->bref.data_off;
1475 dedup->data_crc = crc;
1478 * Set the valid bits for the dedup only after we know the data
1479 * buffer has been updated. The alloc bits were set (and the valid
1480 * bits cleared) when the media was allocated.
1482 * This is done in two stages becuase the bulkfree code can race
1483 * the gap between allocation and data population. Both masks must
1484 * be set before a bcmp/dedup operation is able to use the block.
1486 mask = hammer2_dedup_mask(dio, chain->bref.data_off, chain->bytes);
1487 atomic_set_64(&dio->dedup_valid, mask);
1491 * XXX removed. MODIFIED is an integral part of the flush code,
1492 * lets not just clear it
1495 * Once we record the dedup the chain must be marked clean to
1496 * prevent reuse of the underlying block. Remember that this
1497 * write occurs when the buffer cache is flushed (i.e. on sync(),
1498 * fsync(), filesystem periodic sync, or when the kernel needs to
1499 * flush a buffer), and not whenever the user write()s.
1501 if (chain->flags & HAMMER2_CHAIN_MODIFIED) {
1502 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
1503 atomic_add_long(&hammer2_count_modified_chains, -1);
1505 hammer2_pfs_memory_wakeup(chain->pmp);
1512 hammer2_dedup_lookup(hammer2_dev_t *hmp, char **datap, int pblksize)
1514 hammer2_dedup_t *dedup;
1523 if (hammer2_dedup_enable == 0)
1530 * XXX use the built-in crc (the dedup lookup sequencing
1531 * needs to be fixed so the check code is already present
1532 * when dedup_lookup is called)
1534 crc = XXH64(data, pblksize, XXH_HAMMER2_SEED);
1535 dedup = &hmp->heur_dedup[crc & (HAMMER2_DEDUP_HEUR_MASK & ~3)];
1537 if (hammer2_debug & 0x40000) {
1538 kprintf("LOC %04x/4 %016jx\n",
1539 (int)(dedup - hmp->heur_dedup),
1543 for (i = 0; i < 4; ++i) {
1544 off = dedup[i].data_off;
1546 if (dedup[i].data_crc != crc)
1548 if ((1 << (int)(off & HAMMER2_OFF_MASK_RADIX)) != pblksize)
1550 dio = hammer2_io_getquick(hmp, off, pblksize);
1552 dtmp = hammer2_io_data(dio, off),
1553 mask = hammer2_dedup_mask(dio, off, pblksize);
1554 if ((dio->dedup_alloc & mask) == mask &&
1555 (dio->dedup_valid & mask) == mask &&
1556 bcmp(data, dtmp, pblksize) == 0) {
1557 if (hammer2_debug & 0x40000) {
1558 kprintf("DEDUP SUCCESS %016jx\n",
1561 hammer2_io_putblk(&dio);
1563 dedup[i].ticks = ticks; /* update use */
1564 atomic_add_long(&hammer2_iod_file_wdedup,
1567 return off; /* RETURN */
1569 hammer2_io_putblk(&dio);
1576 * Poof. Races are ok, if someone gets in and reuses a dedup offset
1577 * before or while we are clearing it they will also recover the freemap
1578 * entry (set it to fully allocated), so a bulkfree race can only set it
1579 * to a possibly-free state.
1581 * XXX ok, well, not really sure races are ok but going to run with it
1585 hammer2_dedup_clear(hammer2_dev_t *hmp)
1589 for (i = 0; i < HAMMER2_DEDUP_HEUR_SIZE; ++i) {
1590 hmp->heur_dedup[i].data_off = 0;
1591 hmp->heur_dedup[i].ticks = ticks - 1;