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 * WARNING! The XOP deals with buffer synchronization. It is not synchronized
80 * XXX This isn't supposed to be able to deadlock against vfs_sync vfsync()
81 * calls but it has in the past when multiple flushes are queued.
83 * XXX We currently terminate the transaction once we get a quorum, otherwise
84 * the frontend can stall, but this can leave the remaining nodes with
85 * a potential flush conflict. We need to delay flushes on those nodes
86 * until running transactions complete separately from the normal
87 * transaction sequencing. FIXME TODO.
89 static void hammer2_strategy_xop_read(hammer2_thread_t
*thr
,
91 static void hammer2_strategy_xop_write(hammer2_thread_t
*thr
,
93 static int hammer2_strategy_read(struct vop_strategy_args
*ap
);
94 static int hammer2_strategy_write(struct vop_strategy_args
*ap
);
95 static void hammer2_strategy_read_completion(hammer2_chain_t
*chain
,
96 char *data
, struct bio
*bio
);
98 static hammer2_off_t
hammer2_dedup_lookup(hammer2_dev_t
*hmp
,
99 char **datap
, int pblksize
);
102 hammer2_vop_strategy(struct vop_strategy_args
*ap
)
113 error
= hammer2_strategy_read(ap
);
114 ++hammer2_iod_file_read
;
117 error
= hammer2_strategy_write(ap
);
118 ++hammer2_iod_file_write
;
121 bp
->b_error
= error
= EINVAL
;
122 bp
->b_flags
|= B_ERROR
;
130 * Return the largest contiguous physical disk range for the logical
133 * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
135 * Basically disabled, the logical buffer write thread has to deal with
136 * buffers one-at-a-time. Note that this should not prevent cluster_read()
137 * from reading-ahead, it simply prevents it from trying form a single
138 * cluster buffer for the logical request. H2 already uses 64KB buffers!
141 hammer2_vop_bmap(struct vop_bmap_args
*ap
)
143 *ap
->a_doffsetp
= NOOFFSET
;
151 /****************************************************************************
153 ****************************************************************************/
155 * Callback used in read path in case that a block is compressed with LZ4.
159 hammer2_decompress_LZ4_callback(const char *data
, u_int bytes
, struct bio
*bio
)
162 char *compressed_buffer
;
169 if bio
->bio_caller_info2
.index
&&
170 bio
->bio_caller_info1
.uvalue32
!=
171 crc32(bp
->b_data
, bp
->b_bufsize
) --- return error
174 KKASSERT(bp
->b_bufsize
<= HAMMER2_PBUFSIZE
);
175 compressed_size
= *(const int *)data
;
176 KKASSERT((uint32_t)compressed_size
<= bytes
- sizeof(int));
178 compressed_buffer
= objcache_get(cache_buffer_read
, M_INTWAIT
);
179 result
= LZ4_decompress_safe(__DECONST(char *, &data
[sizeof(int)]),
184 kprintf("READ PATH: Error during decompression."
186 (intmax_t)bio
->bio_offset
, bytes
);
187 /* make sure it isn't random garbage */
188 bzero(compressed_buffer
, bp
->b_bufsize
);
190 KKASSERT(result
<= bp
->b_bufsize
);
191 bcopy(compressed_buffer
, bp
->b_data
, bp
->b_bufsize
);
192 if (result
< bp
->b_bufsize
)
193 bzero(bp
->b_data
+ result
, bp
->b_bufsize
- result
);
194 objcache_put(cache_buffer_read
, compressed_buffer
);
196 bp
->b_flags
|= B_AGE
;
200 * Callback used in read path in case that a block is compressed with ZLIB.
201 * It is almost identical to LZ4 callback, so in theory they can be unified,
202 * but we didn't want to make changes in bio structure for that.
206 hammer2_decompress_ZLIB_callback(const char *data
, u_int bytes
, struct bio
*bio
)
209 char *compressed_buffer
;
210 z_stream strm_decompress
;
216 KKASSERT(bp
->b_bufsize
<= HAMMER2_PBUFSIZE
);
217 strm_decompress
.avail_in
= 0;
218 strm_decompress
.next_in
= Z_NULL
;
220 ret
= inflateInit(&strm_decompress
);
223 kprintf("HAMMER2 ZLIB: Fatal error in inflateInit.\n");
225 compressed_buffer
= objcache_get(cache_buffer_read
, M_INTWAIT
);
226 strm_decompress
.next_in
= __DECONST(char *, data
);
228 /* XXX supply proper size, subset of device bp */
229 strm_decompress
.avail_in
= bytes
;
230 strm_decompress
.next_out
= compressed_buffer
;
231 strm_decompress
.avail_out
= bp
->b_bufsize
;
233 ret
= inflate(&strm_decompress
, Z_FINISH
);
234 if (ret
!= Z_STREAM_END
) {
235 kprintf("HAMMER2 ZLIB: Fatar error during decompression.\n");
236 bzero(compressed_buffer
, bp
->b_bufsize
);
238 bcopy(compressed_buffer
, bp
->b_data
, bp
->b_bufsize
);
239 result
= bp
->b_bufsize
- strm_decompress
.avail_out
;
240 if (result
< bp
->b_bufsize
)
241 bzero(bp
->b_data
+ result
, strm_decompress
.avail_out
);
242 objcache_put(cache_buffer_read
, compressed_buffer
);
243 ret
= inflateEnd(&strm_decompress
);
246 bp
->b_flags
|= B_AGE
;
250 * Logical buffer I/O, async read.
254 hammer2_strategy_read(struct vop_strategy_args
*ap
)
256 hammer2_xop_strategy_t
*xop
;
266 nbio
= push_bio(bio
);
268 lbase
= bio
->bio_offset
;
269 KKASSERT(((int)lbase
& HAMMER2_PBUFMASK
) == 0);
271 xop
= hammer2_xop_alloc(ip
, HAMMER2_XOP_STRATEGY
);
275 hammer2_mtx_init(&xop
->lock
, "h2bior");
276 hammer2_xop_start(&xop
->head
, hammer2_strategy_xop_read
);
277 /* asynchronous completion */
283 * Per-node XOP (threaded), do a synchronous lookup of the chain and
284 * its data. The frontend is asynchronous, so we are also responsible
285 * for racing to terminate the frontend.
289 hammer2_strategy_xop_read(hammer2_thread_t
*thr
, hammer2_xop_t
*arg
)
291 hammer2_xop_strategy_t
*xop
= &arg
->xop_strategy
;
292 hammer2_chain_t
*parent
;
293 hammer2_chain_t
*chain
;
294 hammer2_key_t key_dummy
;
301 * Note that we can race completion of the bio supplied by
302 * the front-end so we cannot access it until we determine
303 * that we are the ones finishing it up.
308 * This is difficult to optimize. The logical buffer might be
309 * partially dirty (contain dummy zero-fill pages), which would
310 * mess up our crc calculation if we were to try a direct read.
311 * So for now we always double-buffer through the underlying
314 * If not for the above problem we could conditionalize on
315 * (1) 64KB buffer, (2) one chain (not multi-master) and
316 * (3) !hammer2_double_buffer, and issue a direct read into the
319 parent
= hammer2_inode_chain(xop
->head
.ip1
, thr
->clindex
,
320 HAMMER2_RESOLVE_ALWAYS
|
321 HAMMER2_RESOLVE_SHARED
);
323 chain
= hammer2_chain_lookup(&parent
, &key_dummy
,
326 HAMMER2_LOOKUP_ALWAYS
|
327 HAMMER2_LOOKUP_SHARED
);
329 error
= chain
->error
;
331 error
= HAMMER2_ERROR_EIO
;
334 error
= hammer2_xop_feed(&xop
->head
, chain
, thr
->clindex
, error
);
336 hammer2_chain_unlock(chain
);
337 hammer2_chain_drop(chain
);
340 hammer2_chain_unlock(parent
);
341 hammer2_chain_drop(parent
);
343 chain
= NULL
; /* safety */
344 parent
= NULL
; /* safety */
347 * Race to finish the frontend. First-to-complete. bio is only
348 * valid if we are determined to be the ones able to complete
353 hammer2_mtx_ex(&xop
->lock
);
355 hammer2_mtx_unlock(&xop
->lock
);
363 * Async operation has not completed and we now own the lock.
364 * Determine if we can complete the operation by issuing the
365 * frontend collection non-blocking.
367 * H2 double-buffers the data, setting B_NOTMETA on the logical
368 * buffer hints to the OS that the logical buffer should not be
369 * swapcached (since the device buffer can be).
371 * Also note that even for compressed data we would rather the
372 * kernel cache/swapcache device buffers more and (decompressed)
373 * logical buffers less, since that will significantly improve
374 * the amount of end-user data that can be cached.
376 error
= hammer2_xop_collect(&xop
->head
, HAMMER2_XOP_COLLECT_NOWAIT
);
381 hammer2_mtx_unlock(&xop
->lock
);
382 bp
->b_flags
|= B_NOTMETA
;
383 chain
= xop
->head
.cluster
.focus
;
384 hammer2_strategy_read_completion(chain
, (char *)chain
->data
,
387 hammer2_xop_retire(&xop
->head
, HAMMER2_XOPMASK_VOP
);
389 case HAMMER2_ERROR_ENOENT
:
391 hammer2_mtx_unlock(&xop
->lock
);
392 bp
->b_flags
|= B_NOTMETA
;
395 bzero(bp
->b_data
, bp
->b_bcount
);
397 hammer2_xop_retire(&xop
->head
, HAMMER2_XOPMASK_VOP
);
399 case HAMMER2_ERROR_EINPROGRESS
:
400 hammer2_mtx_unlock(&xop
->lock
);
403 kprintf("strategy_xop_read: error %08x loff=%016jx\n",
404 error
, bp
->b_loffset
);
406 hammer2_mtx_unlock(&xop
->lock
);
407 bp
->b_flags
|= B_ERROR
;
410 hammer2_xop_retire(&xop
->head
, HAMMER2_XOPMASK_VOP
);
417 hammer2_strategy_read_completion(hammer2_chain_t
*chain
, char *data
,
420 struct buf
*bp
= bio
->bio_buf
;
422 if (chain
->bref
.type
== HAMMER2_BREF_TYPE_INODE
) {
424 * Copy from in-memory inode structure.
426 bcopy(((hammer2_inode_data_t
*)data
)->u
.data
,
427 bp
->b_data
, HAMMER2_EMBEDDED_BYTES
);
428 bzero(bp
->b_data
+ HAMMER2_EMBEDDED_BYTES
,
429 bp
->b_bcount
- HAMMER2_EMBEDDED_BYTES
);
432 } else if (chain
->bref
.type
== HAMMER2_BREF_TYPE_DATA
) {
434 * Data is on-media, record for live dedup. Release the
435 * chain (try to free it) when done. The data is still
436 * cached by both the buffer cache in front and the
437 * block device behind us. This leaves more room in the
438 * LRU chain cache for meta-data chains which we really
441 * NOTE: Deduplication cannot be safely recorded for
442 * records without a check code.
444 hammer2_dedup_record(chain
, NULL
, data
);
445 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_RELEASE
);
448 * Decompression and copy.
450 switch (HAMMER2_DEC_COMP(chain
->bref
.methods
)) {
451 case HAMMER2_COMP_LZ4
:
452 hammer2_decompress_LZ4_callback(data
, chain
->bytes
,
454 /* b_resid set by call */
456 case HAMMER2_COMP_ZLIB
:
457 hammer2_decompress_ZLIB_callback(data
, chain
->bytes
,
459 /* b_resid set by call */
461 case HAMMER2_COMP_NONE
:
462 KKASSERT(chain
->bytes
<= bp
->b_bcount
);
463 bcopy(data
, bp
->b_data
, chain
->bytes
);
464 if (chain
->bytes
< bp
->b_bcount
) {
465 bzero(bp
->b_data
+ chain
->bytes
,
466 bp
->b_bcount
- chain
->bytes
);
472 panic("hammer2_strategy_read: "
473 "unknown compression type");
476 panic("hammer2_strategy_read: unknown bref type");
480 /****************************************************************************
482 ****************************************************************************/
485 * Functions for compression in threads,
486 * from hammer2_vnops.c
488 static void hammer2_write_file_core(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 static void hammer2_compress_and_write(char *data
, hammer2_inode_t
*ip
,
493 hammer2_chain_t
**parentp
,
494 hammer2_key_t lbase
, int ioflag
, int pblksize
,
495 hammer2_tid_t mtid
, int *errorp
,
496 int comp_algo
, int check_algo
);
497 static void hammer2_zero_check_and_write(char *data
, hammer2_inode_t
*ip
,
498 hammer2_chain_t
**parentp
,
499 hammer2_key_t lbase
, int ioflag
, int pblksize
,
500 hammer2_tid_t mtid
, int *errorp
,
502 static int test_block_zeros(const char *buf
, size_t bytes
);
503 static void zero_write(char *data
, hammer2_inode_t
*ip
,
504 hammer2_chain_t
**parentp
,
506 hammer2_tid_t mtid
, int *errorp
);
507 static void hammer2_write_bp(hammer2_chain_t
*chain
, char *data
,
508 int ioflag
, int pblksize
,
509 hammer2_tid_t mtid
, int *errorp
,
514 hammer2_strategy_write(struct vop_strategy_args
*ap
)
516 hammer2_xop_strategy_t
*xop
;
527 hammer2_lwinprog_ref(pmp
);
528 hammer2_trans_assert_strategy(pmp
);
529 hammer2_trans_init(pmp
, HAMMER2_TRANS_BUFCACHE
);
531 xop
= hammer2_xop_alloc(ip
, HAMMER2_XOP_MODIFYING
|
532 HAMMER2_XOP_STRATEGY
);
535 xop
->lbase
= bio
->bio_offset
;
536 hammer2_mtx_init(&xop
->lock
, "h2biow");
537 hammer2_xop_start(&xop
->head
, hammer2_strategy_xop_write
);
538 /* asynchronous completion */
540 hammer2_lwinprog_wait(pmp
, hammer2_flush_pipe
);
546 * Per-node XOP (threaded). Write the logical buffer to the media.
548 * This is a bit problematic because there may be multiple target and
549 * any of them may be able to release the bp. In addition, if our
550 * particulr target is offline we don't want to block the bp (and thus
551 * the frontend). To accomplish this we copy the data to the per-thr
556 hammer2_strategy_xop_write(hammer2_thread_t
*thr
, hammer2_xop_t
*arg
)
558 hammer2_xop_strategy_t
*xop
= &arg
->xop_strategy
;
559 hammer2_chain_t
*parent
;
567 hammer2_off_t bio_offset
;
571 * We can only access the bp/bio if the frontend has not yet
576 hammer2_mtx_sh(&xop
->lock
);
578 hammer2_mtx_unlock(&xop
->lock
);
583 bio
= xop
->bio
; /* ephermal */
584 bp
= bio
->bio_buf
; /* ephermal */
585 ip
= xop
->head
.ip1
; /* retained by ref */
586 bio_offset
= bio
->bio_offset
;
587 bio_data
= thr
->scratch
;
589 /* hammer2_trans_init(parent->hmp->spmp, HAMMER2_TRANS_BUFCACHE); */
591 lblksize
= hammer2_calc_logical(ip
, bio
->bio_offset
, &lbase
, NULL
);
592 pblksize
= hammer2_calc_physical(ip
, lbase
);
594 bcopy(bp
->b_data
, bio_data
, lblksize
);
596 hammer2_mtx_unlock(&xop
->lock
);
597 bp
= NULL
; /* safety, illegal to access after unlock */
598 bio
= NULL
; /* safety, illegal to access after unlock */
603 parent
= hammer2_inode_chain(ip
, thr
->clindex
, HAMMER2_RESOLVE_ALWAYS
);
604 hammer2_write_file_core(bio_data
, ip
, &parent
,
605 lbase
, IO_ASYNC
, pblksize
,
606 xop
->head
.mtid
, &error
);
608 hammer2_chain_unlock(parent
);
609 hammer2_chain_drop(parent
);
610 parent
= NULL
; /* safety */
612 hammer2_xop_feed(&xop
->head
, NULL
, thr
->clindex
, error
);
615 * Try to complete the operation on behalf of the front-end.
619 hammer2_mtx_ex(&xop
->lock
);
621 hammer2_mtx_unlock(&xop
->lock
);
626 * Async operation has not completed and we now own the lock.
627 * Determine if we can complete the operation by issuing the
628 * frontend collection non-blocking.
630 * H2 double-buffers the data, setting B_NOTMETA on the logical
631 * buffer hints to the OS that the logical buffer should not be
632 * swapcached (since the device buffer can be).
634 error
= hammer2_xop_collect(&xop
->head
, HAMMER2_XOP_COLLECT_NOWAIT
);
636 if (error
== HAMMER2_ERROR_EINPROGRESS
) {
637 hammer2_mtx_unlock(&xop
->lock
);
642 * Async operation has completed.
645 hammer2_mtx_unlock(&xop
->lock
);
647 bio
= xop
->bio
; /* now owned by us */
648 bp
= bio
->bio_buf
; /* now owned by us */
650 if (error
== HAMMER2_ERROR_ENOENT
|| error
== 0) {
651 bp
->b_flags
|= B_NOTMETA
;
656 kprintf("strategy_xop_write: error %d loff=%016jx\n",
657 error
, bp
->b_loffset
);
658 bp
->b_flags
|= B_ERROR
;
662 hammer2_xop_retire(&xop
->head
, HAMMER2_XOPMASK_VOP
);
663 hammer2_trans_assert_strategy(ip
->pmp
);
664 hammer2_lwinprog_drop(ip
->pmp
);
665 hammer2_trans_done(ip
->pmp
);
669 * Wait for pending I/O to complete
672 hammer2_bioq_sync(hammer2_pfs_t
*pmp
)
674 hammer2_lwinprog_wait(pmp
, 0);
678 * Assign physical storage at (cparent, lbase), returning a suitable chain
679 * and setting *errorp appropriately.
681 * If no error occurs, the returned chain will be in a modified state.
683 * If an error occurs, the returned chain may or may not be NULL. If
684 * not-null any chain->error (if not 0) will also be rolled up into *errorp.
685 * So the caller only needs to test *errorp.
687 * cparent can wind up being anything.
689 * If datap is not NULL, *datap points to the real data we intend to write.
690 * If we can dedup the storage location we set *datap to NULL to indicate
691 * to the caller that a dedup occurred.
693 * NOTE: Special case for data embedded in inode.
697 hammer2_assign_physical(hammer2_inode_t
*ip
, hammer2_chain_t
**parentp
,
698 hammer2_key_t lbase
, int pblksize
,
699 hammer2_tid_t mtid
, char **datap
, int *errorp
)
701 hammer2_chain_t
*chain
;
702 hammer2_key_t key_dummy
;
703 hammer2_off_t dedup_off
;
704 int pradix
= hammer2_getradix(pblksize
);
707 * Locate the chain associated with lbase, return a locked chain.
708 * However, do not instantiate any data reference (which utilizes a
709 * device buffer) because we will be using direct IO via the
710 * logical buffer cache buffer.
712 KKASSERT(pblksize
>= HAMMER2_ALLOC_MIN
);
714 chain
= hammer2_chain_lookup(parentp
, &key_dummy
,
717 HAMMER2_LOOKUP_NODATA
);
720 * The lookup code should not return a DELETED chain to us, unless
721 * its a short-file embedded in the inode. Then it is possible for
722 * the lookup to return a deleted inode.
724 if (chain
&& (chain
->flags
& HAMMER2_CHAIN_DELETED
) &&
725 chain
->bref
.type
!= HAMMER2_BREF_TYPE_INODE
) {
726 kprintf("assign physical deleted chain @ "
727 "%016jx (%016jx.%02x) ip %016jx\n",
728 lbase
, chain
->bref
.data_off
, chain
->bref
.type
,
735 * We found a hole, create a new chain entry.
737 * NOTE: DATA chains are created without device backing
738 * store (nor do we want any).
740 dedup_off
= hammer2_dedup_lookup((*parentp
)->hmp
, datap
,
742 *errorp
|= hammer2_chain_create(parentp
, &chain
,
744 HAMMER2_ENC_CHECK(ip
->meta
.check_algo
) |
745 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE
),
746 lbase
, HAMMER2_PBUFRADIX
,
747 HAMMER2_BREF_TYPE_DATA
,
752 /*ip->delta_dcount += pblksize;*/
753 } else if (chain
->error
== 0) {
754 switch (chain
->bref
.type
) {
755 case HAMMER2_BREF_TYPE_INODE
:
757 * The data is embedded in the inode, which requires
760 *errorp
|= hammer2_chain_modify_ip(ip
, chain
, mtid
, 0);
762 case HAMMER2_BREF_TYPE_DATA
:
763 dedup_off
= hammer2_dedup_lookup(chain
->hmp
, datap
,
765 if (chain
->bytes
!= pblksize
) {
766 *errorp
|= hammer2_chain_resize(chain
,
769 HAMMER2_MODIFY_OPTDATA
);
775 * DATA buffers must be marked modified whether the
776 * data is in a logical buffer or not. We also have
777 * to make this call to fixup the chain data pointers
778 * after resizing in case this is an encrypted or
781 *errorp
|= hammer2_chain_modify(chain
, mtid
, dedup_off
,
782 HAMMER2_MODIFY_OPTDATA
);
785 panic("hammer2_assign_physical: bad type");
790 *errorp
= chain
->error
;
797 * hammer2_write_file_core() - hammer2_write_thread() helper
799 * The core write function which determines which path to take
800 * depending on compression settings. We also have to locate the
801 * related chains so we can calculate and set the check data for
806 hammer2_write_file_core(char *data
, hammer2_inode_t
*ip
,
807 hammer2_chain_t
**parentp
,
808 hammer2_key_t lbase
, int ioflag
, int pblksize
,
809 hammer2_tid_t mtid
, int *errorp
)
811 hammer2_chain_t
*chain
;
816 switch(HAMMER2_DEC_ALGO(ip
->meta
.comp_algo
)) {
817 case HAMMER2_COMP_NONE
:
819 * We have to assign physical storage to the buffer
820 * we intend to dirty or write now to avoid deadlocks
821 * in the strategy code later.
823 * This can return NOOFFSET for inode-embedded data.
824 * The strategy code will take care of it in that case.
827 chain
= hammer2_assign_physical(ip
, parentp
, lbase
, pblksize
,
828 mtid
, &bdata
, errorp
);
830 /* skip modifications */
831 } else if (chain
->bref
.type
== HAMMER2_BREF_TYPE_INODE
) {
832 hammer2_inode_data_t
*wipdata
;
834 wipdata
= &chain
->data
->ipdata
;
835 KKASSERT(wipdata
->meta
.op_flags
&
836 HAMMER2_OPFLAG_DIRECTDATA
);
837 bcopy(data
, wipdata
->u
.data
, HAMMER2_EMBEDDED_BYTES
);
838 ++hammer2_iod_file_wembed
;
839 } else if (bdata
== NULL
) {
841 * Copy of data already present on-media.
843 chain
->bref
.methods
=
844 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE
) +
845 HAMMER2_ENC_CHECK(ip
->meta
.check_algo
);
846 hammer2_chain_setcheck(chain
, data
);
848 hammer2_write_bp(chain
, data
, ioflag
, pblksize
,
849 mtid
, errorp
, ip
->meta
.check_algo
);
852 hammer2_chain_unlock(chain
);
853 hammer2_chain_drop(chain
);
856 case HAMMER2_COMP_AUTOZERO
:
858 * Check for zero-fill only
860 hammer2_zero_check_and_write(data
, ip
, parentp
,
861 lbase
, ioflag
, pblksize
,
863 ip
->meta
.check_algo
);
865 case HAMMER2_COMP_LZ4
:
866 case HAMMER2_COMP_ZLIB
:
869 * Check for zero-fill and attempt compression.
871 hammer2_compress_and_write(data
, ip
, parentp
,
872 lbase
, ioflag
, pblksize
,
875 ip
->meta
.check_algo
);
883 * Generic function that will perform the compression in compression
884 * write path. The compression algorithm is determined by the settings
885 * obtained from inode.
889 hammer2_compress_and_write(char *data
, hammer2_inode_t
*ip
,
890 hammer2_chain_t
**parentp
,
891 hammer2_key_t lbase
, int ioflag
, int pblksize
,
892 hammer2_tid_t mtid
, int *errorp
, int comp_algo
, int check_algo
)
894 hammer2_chain_t
*chain
;
901 * An all-zeros write creates a hole unless the check code
902 * is disabled. When the check code is disabled all writes
903 * are done in-place, including any all-zeros writes.
905 * NOTE: A snapshot will still force a copy-on-write
906 * (see the HAMMER2_CHECK_NONE in hammer2_chain.c).
908 if (check_algo
!= HAMMER2_CHECK_NONE
&&
909 test_block_zeros(data
, pblksize
)) {
910 zero_write(data
, ip
, parentp
, lbase
, mtid
, errorp
);
915 * Compression requested. Try to compress the block. We store
916 * the data normally if we cannot sufficiently compress it.
918 * We have a heuristic to detect files which are mostly
919 * uncompressable and avoid the compression attempt in that
920 * case. If the compression heuristic is turned off, we always
926 KKASSERT(pblksize
/ 2 <= 32768);
928 if (ip
->comp_heuristic
< 8 || (ip
->comp_heuristic
& 7) == 0 ||
929 hammer2_always_compress
) {
930 z_stream strm_compress
;
934 switch(HAMMER2_DEC_ALGO(comp_algo
)) {
935 case HAMMER2_COMP_LZ4
:
937 * We need to prefix with the size, LZ4
938 * doesn't do it for us. Add the related
941 * NOTE: The LZ4 code seems to assume at least an
942 * 8-byte buffer size granularity and may
943 * overrun the buffer if given a 4-byte
946 comp_buffer
= objcache_get(cache_buffer_write
,
948 comp_size
= LZ4_compress_limitedOutput(
950 &comp_buffer
[sizeof(int)],
952 pblksize
/ 2 - sizeof(int64_t));
953 *(int *)comp_buffer
= comp_size
;
955 comp_size
+= sizeof(int);
957 case HAMMER2_COMP_ZLIB
:
958 comp_level
= HAMMER2_DEC_LEVEL(comp_algo
);
960 comp_level
= 6; /* default zlib compression */
961 else if (comp_level
< 6)
963 else if (comp_level
> 9)
965 ret
= deflateInit(&strm_compress
, comp_level
);
967 kprintf("HAMMER2 ZLIB: fatal error "
968 "on deflateInit.\n");
971 comp_buffer
= objcache_get(cache_buffer_write
,
973 strm_compress
.next_in
= data
;
974 strm_compress
.avail_in
= pblksize
;
975 strm_compress
.next_out
= comp_buffer
;
976 strm_compress
.avail_out
= pblksize
/ 2;
977 ret
= deflate(&strm_compress
, Z_FINISH
);
978 if (ret
== Z_STREAM_END
) {
979 comp_size
= pblksize
/ 2 -
980 strm_compress
.avail_out
;
984 ret
= deflateEnd(&strm_compress
);
987 kprintf("Error: Unknown compression method.\n");
988 kprintf("Comp_method = %d.\n", comp_algo
);
993 if (comp_size
== 0) {
995 * compression failed or turned off
997 comp_block_size
= pblksize
; /* safety */
998 if (++ip
->comp_heuristic
> 128)
999 ip
->comp_heuristic
= 8;
1002 * compression succeeded
1004 ip
->comp_heuristic
= 0;
1005 if (comp_size
<= 1024) {
1006 comp_block_size
= 1024;
1007 } else if (comp_size
<= 2048) {
1008 comp_block_size
= 2048;
1009 } else if (comp_size
<= 4096) {
1010 comp_block_size
= 4096;
1011 } else if (comp_size
<= 8192) {
1012 comp_block_size
= 8192;
1013 } else if (comp_size
<= 16384) {
1014 comp_block_size
= 16384;
1015 } else if (comp_size
<= 32768) {
1016 comp_block_size
= 32768;
1018 panic("hammer2: WRITE PATH: "
1019 "Weird comp_size value.");
1021 comp_block_size
= pblksize
;
1025 * Must zero the remainder or dedup (which operates on a
1026 * physical block basis) will not find matches.
1028 if (comp_size
< comp_block_size
) {
1029 bzero(comp_buffer
+ comp_size
,
1030 comp_block_size
- comp_size
);
1035 * Assign physical storage, data will be set to NULL if a live-dedup
1038 bdata
= comp_size
? comp_buffer
: data
;
1039 chain
= hammer2_assign_physical(ip
, parentp
, lbase
, comp_block_size
,
1040 mtid
, &bdata
, errorp
);
1046 if (chain
->bref
.type
== HAMMER2_BREF_TYPE_INODE
) {
1047 hammer2_inode_data_t
*wipdata
;
1049 *errorp
= hammer2_chain_modify_ip(ip
, chain
, mtid
, 0);
1051 wipdata
= &chain
->data
->ipdata
;
1052 KKASSERT(wipdata
->meta
.op_flags
&
1053 HAMMER2_OPFLAG_DIRECTDATA
);
1054 bcopy(data
, wipdata
->u
.data
, HAMMER2_EMBEDDED_BYTES
);
1055 ++hammer2_iod_file_wembed
;
1057 } else if (bdata
== NULL
) {
1059 * Live deduplication, a copy of the data is already present
1063 chain
->bref
.methods
=
1064 HAMMER2_ENC_COMP(comp_algo
) +
1065 HAMMER2_ENC_CHECK(check_algo
);
1067 chain
->bref
.methods
=
1069 HAMMER2_COMP_NONE
) +
1070 HAMMER2_ENC_CHECK(check_algo
);
1072 bdata
= comp_size
? comp_buffer
: data
;
1073 hammer2_chain_setcheck(chain
, bdata
);
1074 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_INITIAL
);
1078 KKASSERT(chain
->flags
& HAMMER2_CHAIN_MODIFIED
);
1080 switch(chain
->bref
.type
) {
1081 case HAMMER2_BREF_TYPE_INODE
:
1082 panic("hammer2_write_bp: unexpected inode\n");
1084 case HAMMER2_BREF_TYPE_DATA
:
1086 * Optimize out the read-before-write
1089 *errorp
= hammer2_io_newnz(chain
->hmp
,
1091 chain
->bref
.data_off
,
1095 hammer2_io_brelse(&dio
);
1096 kprintf("hammer2: WRITE PATH: "
1097 "dbp bread error\n");
1100 bdata
= hammer2_io_data(dio
, chain
->bref
.data_off
);
1103 * When loading the block make sure we don't
1104 * leave garbage after the compressed data.
1107 chain
->bref
.methods
=
1108 HAMMER2_ENC_COMP(comp_algo
) +
1109 HAMMER2_ENC_CHECK(check_algo
);
1110 bcopy(comp_buffer
, bdata
, comp_size
);
1112 chain
->bref
.methods
=
1114 HAMMER2_COMP_NONE
) +
1115 HAMMER2_ENC_CHECK(check_algo
);
1116 bcopy(data
, bdata
, pblksize
);
1120 * The flush code doesn't calculate check codes for
1121 * file data (doing so can result in excessive I/O),
1124 hammer2_chain_setcheck(chain
, bdata
);
1127 * Device buffer is now valid, chain is no longer in
1128 * the initial state.
1130 * (No blockref table worries with file data)
1132 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_INITIAL
);
1133 hammer2_dedup_record(chain
, dio
, bdata
);
1135 /* Now write the related bdp. */
1136 if (ioflag
& IO_SYNC
) {
1138 * Synchronous I/O requested.
1140 hammer2_io_bwrite(&dio
);
1142 } else if ((ioflag & IO_DIRECT) &&
1143 loff + n == pblksize) {
1144 hammer2_io_bdwrite(&dio);
1146 } else if (ioflag
& IO_ASYNC
) {
1147 hammer2_io_bawrite(&dio
);
1149 hammer2_io_bdwrite(&dio
);
1153 panic("hammer2_write_bp: bad chain type %d\n",
1161 hammer2_chain_unlock(chain
);
1162 hammer2_chain_drop(chain
);
1165 objcache_put(cache_buffer_write
, comp_buffer
);
1171 * Function that performs zero-checking and writing without compression,
1172 * it corresponds to default zero-checking path.
1176 hammer2_zero_check_and_write(char *data
, hammer2_inode_t
*ip
,
1177 hammer2_chain_t
**parentp
,
1178 hammer2_key_t lbase
, int ioflag
, int pblksize
,
1179 hammer2_tid_t mtid
, int *errorp
,
1182 hammer2_chain_t
*chain
;
1185 if (check_algo
!= HAMMER2_CHECK_NONE
&&
1186 test_block_zeros(data
, pblksize
)) {
1188 * An all-zeros write creates a hole unless the check code
1189 * is disabled. When the check code is disabled all writes
1190 * are done in-place, including any all-zeros writes.
1192 * NOTE: A snapshot will still force a copy-on-write
1193 * (see the HAMMER2_CHECK_NONE in hammer2_chain.c).
1195 zero_write(data
, ip
, parentp
, lbase
, mtid
, errorp
);
1201 chain
= hammer2_assign_physical(ip
, parentp
, lbase
, pblksize
,
1202 mtid
, &bdata
, errorp
);
1206 hammer2_write_bp(chain
, data
, ioflag
, pblksize
,
1207 mtid
, errorp
, check_algo
);
1209 /* dedup occurred */
1210 chain
->bref
.methods
=
1211 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE
) +
1212 HAMMER2_ENC_CHECK(check_algo
);
1213 hammer2_chain_setcheck(chain
, data
);
1216 hammer2_chain_unlock(chain
);
1217 hammer2_chain_drop(chain
);
1225 * A function to test whether a block of data contains only zeros,
1226 * returns TRUE (non-zero) if the block is all zeros.
1230 test_block_zeros(const char *buf
, size_t bytes
)
1234 for (i
= 0; i
< bytes
; i
+= sizeof(long)) {
1235 if (*(const long *)(buf
+ i
) != 0)
1244 * Function to "write" a block that contains only zeros.
1248 zero_write(char *data
, hammer2_inode_t
*ip
,
1249 hammer2_chain_t
**parentp
,
1250 hammer2_key_t lbase
, hammer2_tid_t mtid
, int *errorp
)
1252 hammer2_chain_t
*chain
;
1253 hammer2_key_t key_dummy
;
1255 chain
= hammer2_chain_lookup(parentp
, &key_dummy
,
1258 HAMMER2_LOOKUP_NODATA
);
1260 if (chain
->bref
.type
== HAMMER2_BREF_TYPE_INODE
) {
1261 hammer2_inode_data_t
*wipdata
;
1264 *errorp
= hammer2_chain_modify_ip(ip
, chain
,
1268 wipdata
= &chain
->data
->ipdata
;
1269 KKASSERT(wipdata
->meta
.op_flags
&
1270 HAMMER2_OPFLAG_DIRECTDATA
);
1271 bzero(wipdata
->u
.data
, HAMMER2_EMBEDDED_BYTES
);
1272 ++hammer2_iod_file_wembed
;
1275 /* chain->error ok for deletion */
1276 hammer2_chain_delete(*parentp
, chain
,
1277 mtid
, HAMMER2_DELETE_PERMANENT
);
1278 ++hammer2_iod_file_wzero
;
1280 hammer2_chain_unlock(chain
);
1281 hammer2_chain_drop(chain
);
1283 ++hammer2_iod_file_wzero
;
1290 * Function to write the data as it is, without performing any sort of
1291 * compression. This function is used in path without compression and
1292 * default zero-checking path.
1296 hammer2_write_bp(hammer2_chain_t
*chain
, char *data
, int ioflag
,
1298 hammer2_tid_t mtid
, int *errorp
, int check_algo
)
1300 hammer2_inode_data_t
*wipdata
;
1305 error
= 0; /* XXX TODO below */
1307 KKASSERT(chain
->flags
& HAMMER2_CHAIN_MODIFIED
);
1309 switch(chain
->bref
.type
) {
1310 case HAMMER2_BREF_TYPE_INODE
:
1311 wipdata
= &chain
->data
->ipdata
;
1312 KKASSERT(wipdata
->meta
.op_flags
& HAMMER2_OPFLAG_DIRECTDATA
);
1313 bcopy(data
, wipdata
->u
.data
, HAMMER2_EMBEDDED_BYTES
);
1315 ++hammer2_iod_file_wembed
;
1317 case HAMMER2_BREF_TYPE_DATA
:
1318 error
= hammer2_io_newnz(chain
->hmp
,
1320 chain
->bref
.data_off
,
1321 chain
->bytes
, &dio
);
1323 hammer2_io_bqrelse(&dio
);
1324 kprintf("hammer2: WRITE PATH: "
1325 "dbp bread error\n");
1328 bdata
= hammer2_io_data(dio
, chain
->bref
.data_off
);
1330 chain
->bref
.methods
= HAMMER2_ENC_COMP(HAMMER2_COMP_NONE
) +
1331 HAMMER2_ENC_CHECK(check_algo
);
1332 bcopy(data
, bdata
, chain
->bytes
);
1335 * The flush code doesn't calculate check codes for
1336 * file data (doing so can result in excessive I/O),
1339 hammer2_chain_setcheck(chain
, bdata
);
1342 * Device buffer is now valid, chain is no longer in
1343 * the initial state.
1345 * (No blockref table worries with file data)
1347 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_INITIAL
);
1348 hammer2_dedup_record(chain
, dio
, bdata
);
1350 if (ioflag
& IO_SYNC
) {
1352 * Synchronous I/O requested.
1354 hammer2_io_bwrite(&dio
);
1356 } else if ((ioflag & IO_DIRECT) &&
1357 loff + n == pblksize) {
1358 hammer2_io_bdwrite(&dio);
1360 } else if (ioflag
& IO_ASYNC
) {
1361 hammer2_io_bawrite(&dio
);
1363 hammer2_io_bdwrite(&dio
);
1367 panic("hammer2_write_bp: bad chain type %d\n",
1377 * LIVE DEDUP HEURISTICS
1379 * Record media and crc information for possible dedup operation. Note
1380 * that the dedup mask bits must also be set in the related DIO for a dedup
1381 * to be fully validated (which is handled in the freemap allocation code).
1383 * WARNING! This code is SMP safe but the heuristic allows SMP collisions.
1384 * All fields must be loaded into locals and validated.
1386 * WARNING! Should only be used for file data and directory entries,
1387 * hammer2_chain_modify() only checks for the dedup case on data
1388 * chains. Also, dedup data can only be recorded for committed
1389 * chains (so NOT strategy writes which can undergo further
1390 * modification after the fact!).
1393 hammer2_dedup_record(hammer2_chain_t
*chain
, hammer2_io_t
*dio
, char *data
)
1396 hammer2_dedup_t
*dedup
;
1404 * We can only record a dedup if we have media data to test against.
1405 * If dedup is not enabled, return early, which allows a chain to
1406 * remain marked MODIFIED (which might have benefits in special
1407 * situations, though typically it does not).
1409 if (hammer2_dedup_enable
== 0)
1419 switch(HAMMER2_DEC_CHECK(chain
->bref
.methods
)) {
1420 case HAMMER2_CHECK_ISCSI32
:
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)(uint32_t)chain
->bref
.check
.iscsi32
.value
;
1429 crc
= XXH64(data
, chain
->bytes
, XXH_HAMMER2_SEED
);
1431 case HAMMER2_CHECK_XXHASH64
:
1432 crc
= chain
->bref
.check
.xxhash64
.value
;
1434 case HAMMER2_CHECK_SHA192
:
1436 * XXX use the built-in crc (the dedup lookup sequencing
1437 * needs to be fixed so the check code is already present
1438 * when dedup_lookup is called)
1441 crc
= ((uint64_t *)chain
->bref
.check
.sha192
.data
)[0] ^
1442 ((uint64_t *)chain
->bref
.check
.sha192
.data
)[1] ^
1443 ((uint64_t *)chain
->bref
.check
.sha192
.data
)[2];
1445 crc
= XXH64(data
, chain
->bytes
, XXH_HAMMER2_SEED
);
1449 * Cannot dedup without a check code
1451 * NOTE: In particular, CHECK_NONE allows a sector to be
1452 * overwritten without copy-on-write, recording
1453 * a dedup block for a CHECK_NONE object would be
1459 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_DEDUPABLE
);
1461 dedup
= &hmp
->heur_dedup
[crc
& (HAMMER2_DEDUP_HEUR_MASK
& ~3)];
1462 for (i
= 0; i
< 4; ++i
) {
1463 if (dedup
[i
].data_crc
== crc
) {
1467 dticks
= (int)(dedup
[i
].ticks
- dedup
[best
].ticks
);
1468 if (dticks
< 0 || dticks
> hz
* 60 * 30)
1472 if (hammer2_debug
& 0x40000) {
1473 kprintf("REC %04x %016jx %016jx\n",
1474 (int)(dedup
- hmp
->heur_dedup
),
1476 chain
->bref
.data_off
);
1478 dedup
->ticks
= ticks
;
1479 dedup
->data_off
= chain
->bref
.data_off
;
1480 dedup
->data_crc
= crc
;
1483 * Set the valid bits for the dedup only after we know the data
1484 * buffer has been updated. The alloc bits were set (and the valid
1485 * bits cleared) when the media was allocated.
1487 * This is done in two stages becuase the bulkfree code can race
1488 * the gap between allocation and data population. Both masks must
1489 * be set before a bcmp/dedup operation is able to use the block.
1491 mask
= hammer2_dedup_mask(dio
, chain
->bref
.data_off
, chain
->bytes
);
1492 atomic_set_64(&dio
->dedup_valid
, mask
);
1496 * XXX removed. MODIFIED is an integral part of the flush code,
1497 * lets not just clear it
1500 * Once we record the dedup the chain must be marked clean to
1501 * prevent reuse of the underlying block. Remember that this
1502 * write occurs when the buffer cache is flushed (i.e. on sync(),
1503 * fsync(), filesystem periodic sync, or when the kernel needs to
1504 * flush a buffer), and not whenever the user write()s.
1506 if (chain
->flags
& HAMMER2_CHAIN_MODIFIED
) {
1507 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_MODIFIED
);
1508 atomic_add_long(&hammer2_count_modified_chains
, -1);
1510 hammer2_pfs_memory_wakeup(chain
->pmp
);
1517 hammer2_dedup_lookup(hammer2_dev_t
*hmp
, char **datap
, int pblksize
)
1519 hammer2_dedup_t
*dedup
;
1528 if (hammer2_dedup_enable
== 0)
1535 * XXX use the built-in crc (the dedup lookup sequencing
1536 * needs to be fixed so the check code is already present
1537 * when dedup_lookup is called)
1539 crc
= XXH64(data
, pblksize
, XXH_HAMMER2_SEED
);
1540 dedup
= &hmp
->heur_dedup
[crc
& (HAMMER2_DEDUP_HEUR_MASK
& ~3)];
1542 if (hammer2_debug
& 0x40000) {
1543 kprintf("LOC %04x/4 %016jx\n",
1544 (int)(dedup
- hmp
->heur_dedup
),
1548 for (i
= 0; i
< 4; ++i
) {
1549 off
= dedup
[i
].data_off
;
1551 if (dedup
[i
].data_crc
!= crc
)
1553 if ((1 << (int)(off
& HAMMER2_OFF_MASK_RADIX
)) != pblksize
)
1555 dio
= hammer2_io_getquick(hmp
, off
, pblksize
);
1557 dtmp
= hammer2_io_data(dio
, off
),
1558 mask
= hammer2_dedup_mask(dio
, off
, pblksize
);
1559 if ((dio
->dedup_alloc
& mask
) == mask
&&
1560 (dio
->dedup_valid
& mask
) == mask
&&
1561 bcmp(data
, dtmp
, pblksize
) == 0) {
1562 if (hammer2_debug
& 0x40000) {
1563 kprintf("DEDUP SUCCESS %016jx\n",
1566 hammer2_io_putblk(&dio
);
1568 dedup
[i
].ticks
= ticks
; /* update use */
1569 atomic_add_long(&hammer2_iod_file_wdedup
,
1572 return off
; /* RETURN */
1574 hammer2_io_putblk(&dio
);
1581 * Poof. Races are ok, if someone gets in and reuses a dedup offset
1582 * before or while we are clearing it they will also recover the freemap
1583 * entry (set it to fully allocated), so a bulkfree race can only set it
1584 * to a possibly-free state.
1586 * XXX ok, well, not really sure races are ok but going to run with it
1590 hammer2_dedup_clear(hammer2_dev_t
*hmp
)
1594 for (i
= 0; i
< HAMMER2_DEDUP_HEUR_SIZE
; ++i
) {
1595 hmp
->heur_dedup
[i
].data_off
= 0;
1596 hmp
->heur_dedup
[i
].ticks
= ticks
- 1;