2 * Copyright (c) 2011-2018 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@dragonflybsd.org>
6 * by Venkatesh Srinivas <vsrinivas@dragonflybsd.org>
7 * by Daniel Flores (GSOC 2013 - mentored by Matthew Dillon, compression)
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in
17 * the documentation and/or other materials provided with the
19 * 3. Neither the name of The DragonFly Project nor the names of its
20 * contributors may be used to endorse or promote products derived
21 * from this software without specific, prior written permission.
23 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
24 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
25 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
26 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
27 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
28 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
29 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
30 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
31 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
32 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
33 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
37 * This module handles low level logical file I/O (strategy) which backs
38 * the logical buffer cache.
40 * [De]compression, zero-block, check codes, and buffer cache operations
41 * for file data is handled here.
43 * Live dedup makes its home here as well.
46 #include <sys/param.h>
47 #include <sys/systm.h>
48 #include <sys/kernel.h>
51 #include <sys/mount.h>
52 #include <sys/vnode.h>
53 #include <sys/objcache.h>
56 #include "hammer2_lz4.h"
58 #include "zlib/hammer2_zlib.h"
60 struct objcache
*cache_buffer_read
;
61 struct objcache
*cache_buffer_write
;
64 * Strategy code (async logical file buffer I/O from system)
66 * Except for the transaction init (which should normally not block),
67 * we essentially run the strategy operation asynchronously via a XOP.
69 * WARNING! The XOP deals with buffer synchronization. It is not synchronized
72 * XXX This isn't supposed to be able to deadlock against vfs_sync vfsync()
73 * calls but it has in the past when multiple flushes are queued.
75 * XXX We currently terminate the transaction once we get a quorum, otherwise
76 * the frontend can stall, but this can leave the remaining nodes with
77 * a potential flush conflict. We need to delay flushes on those nodes
78 * until running transactions complete separately from the normal
79 * transaction sequencing. FIXME TODO.
81 static int hammer2_strategy_read(struct vop_strategy_args
*ap
);
82 static int hammer2_strategy_write(struct vop_strategy_args
*ap
);
83 static void hammer2_strategy_read_completion(hammer2_chain_t
*focus
,
84 const char *data
, struct bio
*bio
);
86 static hammer2_off_t
hammer2_dedup_lookup(hammer2_dev_t
*hmp
,
87 char **datap
, int pblksize
);
90 hammer2_vop_strategy(struct vop_strategy_args
*ap
)
101 error
= hammer2_strategy_read(ap
);
104 error
= hammer2_strategy_write(ap
);
107 bp
->b_error
= error
= EINVAL
;
108 bp
->b_flags
|= B_ERROR
;
116 * Return the largest contiguous physical disk range for the logical
119 * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
121 * Basically disabled, the logical buffer write thread has to deal with
122 * buffers one-at-a-time. Note that this should not prevent cluster_read()
123 * from reading-ahead, it simply prevents it from trying form a single
124 * cluster buffer for the logical request. H2 already uses 64KB buffers!
127 hammer2_vop_bmap(struct vop_bmap_args
*ap
)
129 *ap
->a_doffsetp
= NOOFFSET
;
137 /****************************************************************************
139 ****************************************************************************/
141 * Callback used in read path in case that a block is compressed with LZ4.
145 hammer2_decompress_LZ4_callback(const char *data
, u_int bytes
, struct bio
*bio
)
148 char *compressed_buffer
;
155 if bio
->bio_caller_info2
.index
&&
156 bio
->bio_caller_info1
.uvalue32
!=
157 crc32(bp
->b_data
, bp
->b_bufsize
) --- return error
160 KKASSERT(bp
->b_bufsize
<= HAMMER2_PBUFSIZE
);
161 compressed_size
= *(const int *)data
;
162 KKASSERT((uint32_t)compressed_size
<= bytes
- sizeof(int));
164 compressed_buffer
= objcache_get(cache_buffer_read
, M_INTWAIT
);
165 result
= LZ4_decompress_safe(__DECONST(char *, &data
[sizeof(int)]),
170 kprintf("READ PATH: Error during decompression."
172 (intmax_t)bio
->bio_offset
, bytes
);
173 /* make sure it isn't random garbage */
174 bzero(compressed_buffer
, bp
->b_bufsize
);
176 KKASSERT(result
<= bp
->b_bufsize
);
177 bcopy(compressed_buffer
, bp
->b_data
, bp
->b_bufsize
);
178 if (result
< bp
->b_bufsize
)
179 bzero(bp
->b_data
+ result
, bp
->b_bufsize
- result
);
180 objcache_put(cache_buffer_read
, compressed_buffer
);
182 bp
->b_flags
|= B_AGE
;
186 * Callback used in read path in case that a block is compressed with ZLIB.
187 * It is almost identical to LZ4 callback, so in theory they can be unified,
188 * but we didn't want to make changes in bio structure for that.
192 hammer2_decompress_ZLIB_callback(const char *data
, u_int bytes
, struct bio
*bio
)
195 char *compressed_buffer
;
196 z_stream strm_decompress
;
202 KKASSERT(bp
->b_bufsize
<= HAMMER2_PBUFSIZE
);
203 strm_decompress
.avail_in
= 0;
204 strm_decompress
.next_in
= Z_NULL
;
206 ret
= inflateInit(&strm_decompress
);
209 kprintf("HAMMER2 ZLIB: Fatal error in inflateInit.\n");
211 compressed_buffer
= objcache_get(cache_buffer_read
, M_INTWAIT
);
212 strm_decompress
.next_in
= __DECONST(char *, data
);
214 /* XXX supply proper size, subset of device bp */
215 strm_decompress
.avail_in
= bytes
;
216 strm_decompress
.next_out
= compressed_buffer
;
217 strm_decompress
.avail_out
= bp
->b_bufsize
;
219 ret
= inflate(&strm_decompress
, Z_FINISH
);
220 if (ret
!= Z_STREAM_END
) {
221 kprintf("HAMMER2 ZLIB: Fatal error during decompression.\n");
222 bzero(compressed_buffer
, bp
->b_bufsize
);
224 bcopy(compressed_buffer
, bp
->b_data
, bp
->b_bufsize
);
225 result
= bp
->b_bufsize
- strm_decompress
.avail_out
;
226 if (result
< bp
->b_bufsize
)
227 bzero(bp
->b_data
+ result
, strm_decompress
.avail_out
);
228 objcache_put(cache_buffer_read
, compressed_buffer
);
229 ret
= inflateEnd(&strm_decompress
);
232 bp
->b_flags
|= B_AGE
;
236 * Logical buffer I/O, async read.
240 hammer2_strategy_read(struct vop_strategy_args
*ap
)
242 hammer2_xop_strategy_t
*xop
;
250 lbase
= bio
->bio_offset
;
251 KKASSERT(((int)lbase
& HAMMER2_PBUFMASK
) == 0);
253 xop
= hammer2_xop_alloc(ip
, HAMMER2_XOP_STRATEGY
);
257 hammer2_mtx_init(&xop
->lock
, "h2bior");
258 hammer2_xop_start(&xop
->head
, &hammer2_strategy_read_desc
);
259 /* asynchronous completion */
265 * Per-node XOP (threaded), do a synchronous lookup of the chain and
266 * its data. The frontend is asynchronous, so we are also responsible
267 * for racing to terminate the frontend.
270 hammer2_xop_strategy_read(hammer2_xop_t
*arg
, void *scratch
, int clindex
)
272 hammer2_xop_strategy_t
*xop
= &arg
->xop_strategy
;
273 hammer2_chain_t
*parent
;
274 hammer2_chain_t
*chain
;
275 hammer2_chain_t
*focus
;
276 hammer2_key_t key_dummy
;
284 * Note that we can race completion of the bio supplied by
285 * the front-end so we cannot access it until we determine
286 * that we are the ones finishing it up.
291 * This is difficult to optimize. The logical buffer might be
292 * partially dirty (contain dummy zero-fill pages), which would
293 * mess up our crc calculation if we were to try a direct read.
294 * So for now we always double-buffer through the underlying
297 * If not for the above problem we could conditionalize on
298 * (1) 64KB buffer, (2) one chain (not multi-master) and
299 * (3) !hammer2_double_buffer, and issue a direct read into the
302 parent
= hammer2_inode_chain(xop
->head
.ip1
, clindex
,
303 HAMMER2_RESOLVE_ALWAYS
|
304 HAMMER2_RESOLVE_SHARED
);
306 chain
= hammer2_chain_lookup(&parent
, &key_dummy
,
309 HAMMER2_LOOKUP_ALWAYS
|
310 HAMMER2_LOOKUP_SHARED
);
312 error
= chain
->error
;
314 error
= HAMMER2_ERROR_EIO
;
317 error
= hammer2_xop_feed(&xop
->head
, chain
, clindex
, error
);
319 hammer2_chain_unlock(chain
);
320 hammer2_chain_drop(chain
);
323 hammer2_chain_unlock(parent
);
324 hammer2_chain_drop(parent
);
326 chain
= NULL
; /* safety */
327 parent
= NULL
; /* safety */
330 * Race to finish the frontend. First-to-complete. bio is only
331 * valid if we are determined to be the ones able to complete
336 hammer2_mtx_ex(&xop
->lock
);
338 hammer2_mtx_unlock(&xop
->lock
);
346 * Async operation has not completed and we now own the lock.
347 * Determine if we can complete the operation by issuing the
348 * frontend collection non-blocking.
350 * H2 double-buffers the data, setting B_NOTMETA on the logical
351 * buffer hints to the OS that the logical buffer should not be
352 * swapcached (since the device buffer can be).
354 * Also note that even for compressed data we would rather the
355 * kernel cache/swapcache device buffers more and (decompressed)
356 * logical buffers less, since that will significantly improve
357 * the amount of end-user data that can be cached.
359 * NOTE: The chain->data for xop->head.cluster.focus will be
360 * synchronized to the current cpu by xop_collect(),
361 * but other chains in the cluster might not be.
363 error
= hammer2_xop_collect(&xop
->head
, HAMMER2_XOP_COLLECT_NOWAIT
);
368 hammer2_mtx_unlock(&xop
->lock
);
369 bp
->b_flags
|= B_NOTMETA
;
370 focus
= xop
->head
.cluster
.focus
;
371 data
= hammer2_xop_gdata(&xop
->head
)->buf
;
372 hammer2_strategy_read_completion(focus
, data
, xop
->bio
);
373 hammer2_xop_pdata(&xop
->head
);
375 hammer2_xop_retire(&xop
->head
, HAMMER2_XOPMASK_VOP
);
377 case HAMMER2_ERROR_ENOENT
:
379 hammer2_mtx_unlock(&xop
->lock
);
380 bp
->b_flags
|= B_NOTMETA
;
383 bzero(bp
->b_data
, bp
->b_bcount
);
385 hammer2_xop_retire(&xop
->head
, HAMMER2_XOPMASK_VOP
);
387 case HAMMER2_ERROR_EINPROGRESS
:
388 hammer2_mtx_unlock(&xop
->lock
);
391 kprintf("xop_strategy_read: error %08x loff=%016jx\n",
392 error
, (intmax_t)bp
->b_loffset
);
394 hammer2_mtx_unlock(&xop
->lock
);
395 bp
->b_flags
|= B_ERROR
;
398 hammer2_xop_retire(&xop
->head
, HAMMER2_XOPMASK_VOP
);
405 hammer2_strategy_read_completion(hammer2_chain_t
*focus
, const char *data
,
408 struct buf
*bp
= bio
->bio_buf
;
410 if (focus
->bref
.type
== HAMMER2_BREF_TYPE_INODE
) {
412 * Copy from in-memory inode structure.
414 bcopy(((const hammer2_inode_data_t
*)data
)->u
.data
,
415 bp
->b_data
, HAMMER2_EMBEDDED_BYTES
);
416 bzero(bp
->b_data
+ HAMMER2_EMBEDDED_BYTES
,
417 bp
->b_bcount
- HAMMER2_EMBEDDED_BYTES
);
420 } else if (focus
->bref
.type
== HAMMER2_BREF_TYPE_DATA
) {
422 * Data is on-media, record for live dedup. Release the
423 * chain (try to free it) when done. The data is still
424 * cached by both the buffer cache in front and the
425 * block device behind us. This leaves more room in the
426 * LRU chain cache for meta-data chains which we really
429 * NOTE: Deduplication cannot be safely recorded for
430 * records without a check code.
432 hammer2_dedup_record(focus
, NULL
, data
);
433 atomic_set_int(&focus
->flags
, HAMMER2_CHAIN_RELEASE
);
436 * Decompression and copy.
438 switch (HAMMER2_DEC_COMP(focus
->bref
.methods
)) {
439 case HAMMER2_COMP_LZ4
:
440 hammer2_decompress_LZ4_callback(data
, focus
->bytes
,
442 /* b_resid set by call */
444 case HAMMER2_COMP_ZLIB
:
445 hammer2_decompress_ZLIB_callback(data
, focus
->bytes
,
447 /* b_resid set by call */
449 case HAMMER2_COMP_NONE
:
450 KKASSERT(focus
->bytes
<= bp
->b_bcount
);
451 bcopy(data
, bp
->b_data
, focus
->bytes
);
452 if (focus
->bytes
< bp
->b_bcount
) {
453 bzero(bp
->b_data
+ focus
->bytes
,
454 bp
->b_bcount
- focus
->bytes
);
460 panic("hammer2_strategy_read_completion: "
461 "unknown compression type");
464 panic("hammer2_strategy_read_completion: unknown bref type");
468 /****************************************************************************
470 ****************************************************************************/
473 * Functions for compression in threads,
474 * from hammer2_vnops.c
476 static void hammer2_write_file_core(char *data
, hammer2_inode_t
*ip
,
477 hammer2_chain_t
**parentp
,
478 hammer2_key_t lbase
, int ioflag
, int pblksize
,
479 hammer2_tid_t mtid
, int *errorp
);
480 static void hammer2_compress_and_write(char *data
, hammer2_inode_t
*ip
,
481 hammer2_chain_t
**parentp
,
482 hammer2_key_t lbase
, int ioflag
, int pblksize
,
483 hammer2_tid_t mtid
, int *errorp
,
484 int comp_algo
, int check_algo
);
485 static void hammer2_zero_check_and_write(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
,
490 static int test_block_zeros(const char *buf
, size_t bytes
);
491 static void zero_write(char *data
, hammer2_inode_t
*ip
,
492 hammer2_chain_t
**parentp
,
494 hammer2_tid_t mtid
, int *errorp
);
495 static void hammer2_write_bp(hammer2_chain_t
*chain
, char *data
,
496 int ioflag
, int pblksize
,
497 hammer2_tid_t mtid
, int *errorp
,
501 hammer2_strategy_write(struct vop_strategy_args
*ap
)
503 hammer2_xop_strategy_t
*xop
;
512 atomic_set_int(&ip
->flags
, HAMMER2_INODE_DIRTYDATA
);
513 hammer2_lwinprog_ref(pmp
);
514 hammer2_trans_assert_strategy(pmp
);
515 hammer2_trans_init(pmp
, HAMMER2_TRANS_BUFCACHE
);
517 xop
= hammer2_xop_alloc(ip
, HAMMER2_XOP_MODIFYING
|
518 HAMMER2_XOP_STRATEGY
);
521 xop
->lbase
= bio
->bio_offset
;
522 hammer2_mtx_init(&xop
->lock
, "h2biow");
523 hammer2_xop_start(&xop
->head
, &hammer2_strategy_write_desc
);
524 /* asynchronous completion */
526 hammer2_lwinprog_wait(pmp
, hammer2_flush_pipe
);
532 * Per-node XOP (threaded). Write the logical buffer to the media.
534 * This is a bit problematic because there may be multiple target and
535 * any of them may be able to release the bp. In addition, if our
536 * particulr target is offline we don't want to block the bp (and thus
537 * the frontend). To accomplish this we copy the data to the per-thr
541 hammer2_xop_strategy_write(hammer2_xop_t
*arg
, void *scratch
, int clindex
)
543 hammer2_xop_strategy_t
*xop
= &arg
->xop_strategy
;
544 hammer2_chain_t
*parent
;
555 * We can only access the bp/bio if the frontend has not yet
560 hammer2_mtx_sh(&xop
->lock
);
562 hammer2_mtx_unlock(&xop
->lock
);
567 bio
= xop
->bio
; /* ephermal */
568 bp
= bio
->bio_buf
; /* ephermal */
569 ip
= xop
->head
.ip1
; /* retained by ref */
572 /* hammer2_trans_init(parent->hmp->spmp, HAMMER2_TRANS_BUFCACHE); */
574 lblksize
= hammer2_calc_logical(ip
, bio
->bio_offset
, &lbase
, NULL
);
575 pblksize
= hammer2_calc_physical(ip
, lbase
);
577 KKASSERT(lblksize
<= MAXPHYS
);
578 bcopy(bp
->b_data
, bio_data
, lblksize
);
580 hammer2_mtx_unlock(&xop
->lock
);
581 bp
= NULL
; /* safety, illegal to access after unlock */
582 bio
= NULL
; /* safety, illegal to access after unlock */
587 parent
= hammer2_inode_chain(ip
, clindex
, HAMMER2_RESOLVE_ALWAYS
);
588 hammer2_write_file_core(bio_data
, ip
, &parent
,
589 lbase
, IO_ASYNC
, pblksize
,
590 xop
->head
.mtid
, &error
);
592 hammer2_chain_unlock(parent
);
593 hammer2_chain_drop(parent
);
594 parent
= NULL
; /* safety */
596 hammer2_xop_feed(&xop
->head
, NULL
, clindex
, error
);
599 * Try to complete the operation on behalf of the front-end.
603 hammer2_mtx_ex(&xop
->lock
);
605 hammer2_mtx_unlock(&xop
->lock
);
610 * Async operation has not completed and we now own the lock.
611 * Determine if we can complete the operation by issuing the
612 * frontend collection non-blocking.
614 * H2 double-buffers the data, setting B_NOTMETA on the logical
615 * buffer hints to the OS that the logical buffer should not be
616 * swapcached (since the device buffer can be).
618 error
= hammer2_xop_collect(&xop
->head
, HAMMER2_XOP_COLLECT_NOWAIT
);
620 if (error
== HAMMER2_ERROR_EINPROGRESS
) {
621 hammer2_mtx_unlock(&xop
->lock
);
626 * Async operation has completed.
629 hammer2_mtx_unlock(&xop
->lock
);
631 bio
= xop
->bio
; /* now owned by us */
632 bp
= bio
->bio_buf
; /* now owned by us */
634 if (error
== HAMMER2_ERROR_ENOENT
|| error
== 0) {
635 bp
->b_flags
|= B_NOTMETA
;
640 kprintf("xop_strategy_write: error %d loff=%016jx\n",
641 error
, (intmax_t)bp
->b_loffset
);
642 bp
->b_flags
|= B_ERROR
;
646 hammer2_xop_retire(&xop
->head
, HAMMER2_XOPMASK_VOP
);
647 hammer2_trans_assert_strategy(ip
->pmp
);
648 hammer2_lwinprog_drop(ip
->pmp
);
649 hammer2_trans_done(ip
->pmp
, HAMMER2_TRANS_BUFCACHE
);
653 * Wait for pending I/O to complete
656 hammer2_bioq_sync(hammer2_pfs_t
*pmp
)
658 hammer2_lwinprog_wait(pmp
, 0);
662 * Assign physical storage at (cparent, lbase), returning a suitable chain
663 * and setting *errorp appropriately.
665 * If no error occurs, the returned chain will be in a modified state.
667 * If an error occurs, the returned chain may or may not be NULL. If
668 * not-null any chain->error (if not 0) will also be rolled up into *errorp.
669 * So the caller only needs to test *errorp.
671 * cparent can wind up being anything.
673 * If datap is not NULL, *datap points to the real data we intend to write.
674 * If we can dedup the storage location we set *datap to NULL to indicate
675 * to the caller that a dedup occurred.
677 * NOTE: Special case for data embedded in inode.
681 hammer2_assign_physical(hammer2_inode_t
*ip
, hammer2_chain_t
**parentp
,
682 hammer2_key_t lbase
, int pblksize
,
683 hammer2_tid_t mtid
, char **datap
, int *errorp
)
685 hammer2_chain_t
*chain
;
686 hammer2_key_t key_dummy
;
687 hammer2_off_t dedup_off
;
688 int pradix
= hammer2_getradix(pblksize
);
691 * Locate the chain associated with lbase, return a locked chain.
692 * However, do not instantiate any data reference (which utilizes a
693 * device buffer) because we will be using direct IO via the
694 * logical buffer cache buffer.
696 KKASSERT(pblksize
>= HAMMER2_ALLOC_MIN
);
698 chain
= hammer2_chain_lookup(parentp
, &key_dummy
,
701 HAMMER2_LOOKUP_NODATA
);
704 * The lookup code should not return a DELETED chain to us, unless
705 * its a short-file embedded in the inode. Then it is possible for
706 * the lookup to return a deleted inode.
708 if (chain
&& (chain
->flags
& HAMMER2_CHAIN_DELETED
) &&
709 chain
->bref
.type
!= HAMMER2_BREF_TYPE_INODE
) {
710 kprintf("assign physical deleted chain @ "
711 "%016jx (%016jx.%02x) ip %016jx\n",
712 lbase
, chain
->bref
.data_off
, chain
->bref
.type
,
719 * We found a hole, create a new chain entry.
721 * NOTE: DATA chains are created without device backing
722 * store (nor do we want any).
724 dedup_off
= hammer2_dedup_lookup((*parentp
)->hmp
, datap
,
726 *errorp
|= hammer2_chain_create(parentp
, &chain
, NULL
, ip
->pmp
,
727 HAMMER2_ENC_CHECK(ip
->meta
.check_algo
) |
728 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE
),
729 lbase
, HAMMER2_PBUFRADIX
,
730 HAMMER2_BREF_TYPE_DATA
,
735 /*ip->delta_dcount += pblksize;*/
736 } else if (chain
->error
== 0) {
737 switch (chain
->bref
.type
) {
738 case HAMMER2_BREF_TYPE_INODE
:
740 * The data is embedded in the inode, which requires
743 *errorp
|= hammer2_chain_modify_ip(ip
, chain
, mtid
, 0);
745 case HAMMER2_BREF_TYPE_DATA
:
746 dedup_off
= hammer2_dedup_lookup(chain
->hmp
, datap
,
748 if (chain
->bytes
!= pblksize
) {
749 *errorp
|= hammer2_chain_resize(chain
,
752 HAMMER2_MODIFY_OPTDATA
);
758 * DATA buffers must be marked modified whether the
759 * data is in a logical buffer or not. We also have
760 * to make this call to fixup the chain data pointers
761 * after resizing in case this is an encrypted or
764 *errorp
|= hammer2_chain_modify(chain
, mtid
, dedup_off
,
765 HAMMER2_MODIFY_OPTDATA
);
768 panic("hammer2_assign_physical: bad type");
773 *errorp
= chain
->error
;
775 atomic_set_int(&ip
->flags
, HAMMER2_INODE_DIRTYDATA
);
781 * hammer2_write_file_core()
783 * The core write function which determines which path to take
784 * depending on compression settings. We also have to locate the
785 * related chains so we can calculate and set the check data for
790 hammer2_write_file_core(char *data
, hammer2_inode_t
*ip
,
791 hammer2_chain_t
**parentp
,
792 hammer2_key_t lbase
, int ioflag
, int pblksize
,
793 hammer2_tid_t mtid
, int *errorp
)
795 hammer2_chain_t
*chain
;
800 switch(HAMMER2_DEC_ALGO(ip
->meta
.comp_algo
)) {
801 case HAMMER2_COMP_NONE
:
803 * We have to assign physical storage to the buffer
804 * we intend to dirty or write now to avoid deadlocks
805 * in the strategy code later.
807 * This can return NOOFFSET for inode-embedded data.
808 * The strategy code will take care of it in that case.
811 chain
= hammer2_assign_physical(ip
, parentp
, lbase
, pblksize
,
812 mtid
, &bdata
, errorp
);
814 /* skip modifications */
815 } else if (chain
->bref
.type
== HAMMER2_BREF_TYPE_INODE
) {
816 hammer2_inode_data_t
*wipdata
;
818 wipdata
= &chain
->data
->ipdata
;
819 KKASSERT(wipdata
->meta
.op_flags
&
820 HAMMER2_OPFLAG_DIRECTDATA
);
821 bcopy(data
, wipdata
->u
.data
, HAMMER2_EMBEDDED_BYTES
);
822 ++hammer2_iod_file_wembed
;
823 } else if (bdata
== NULL
) {
825 * Copy of data already present on-media.
827 chain
->bref
.methods
=
828 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE
) +
829 HAMMER2_ENC_CHECK(ip
->meta
.check_algo
);
830 hammer2_chain_setcheck(chain
, data
);
831 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_INITIAL
);
833 hammer2_write_bp(chain
, data
, ioflag
, pblksize
,
834 mtid
, errorp
, ip
->meta
.check_algo
);
837 hammer2_chain_unlock(chain
);
838 hammer2_chain_drop(chain
);
841 case HAMMER2_COMP_AUTOZERO
:
843 * Check for zero-fill only
845 hammer2_zero_check_and_write(data
, ip
, parentp
,
846 lbase
, ioflag
, pblksize
,
848 ip
->meta
.check_algo
);
850 case HAMMER2_COMP_LZ4
:
851 case HAMMER2_COMP_ZLIB
:
854 * Check for zero-fill and attempt compression.
856 hammer2_compress_and_write(data
, ip
, parentp
,
857 lbase
, ioflag
, pblksize
,
860 ip
->meta
.check_algo
);
868 * Generic function that will perform the compression in compression
869 * write path. The compression algorithm is determined by the settings
870 * obtained from inode.
874 hammer2_compress_and_write(char *data
, hammer2_inode_t
*ip
,
875 hammer2_chain_t
**parentp
,
876 hammer2_key_t lbase
, int ioflag
, int pblksize
,
877 hammer2_tid_t mtid
, int *errorp
, int comp_algo
, int check_algo
)
879 hammer2_chain_t
*chain
;
886 * An all-zeros write creates a hole unless the check code
887 * is disabled. When the check code is disabled all writes
888 * are done in-place, including any all-zeros writes.
890 * NOTE: A snapshot will still force a copy-on-write
891 * (see the HAMMER2_CHECK_NONE in hammer2_chain.c).
893 if (check_algo
!= HAMMER2_CHECK_NONE
&&
894 test_block_zeros(data
, pblksize
)) {
895 zero_write(data
, ip
, parentp
, lbase
, mtid
, errorp
);
900 * Compression requested. Try to compress the block. We store
901 * the data normally if we cannot sufficiently compress it.
903 * We have a heuristic to detect files which are mostly
904 * uncompressable and avoid the compression attempt in that
905 * case. If the compression heuristic is turned off, we always
911 KKASSERT(pblksize
/ 2 <= 32768);
913 if (ip
->comp_heuristic
< 8 || (ip
->comp_heuristic
& 7) == 0 ||
914 hammer2_always_compress
) {
915 z_stream strm_compress
;
919 switch(HAMMER2_DEC_ALGO(comp_algo
)) {
920 case HAMMER2_COMP_LZ4
:
922 * We need to prefix with the size, LZ4
923 * doesn't do it for us. Add the related
926 * NOTE: The LZ4 code seems to assume at least an
927 * 8-byte buffer size granularity and may
928 * overrun the buffer if given a 4-byte
931 comp_buffer
= objcache_get(cache_buffer_write
,
933 comp_size
= LZ4_compress_limitedOutput(
935 &comp_buffer
[sizeof(int)],
937 pblksize
/ 2 - sizeof(int64_t));
938 *(int *)comp_buffer
= comp_size
;
940 comp_size
+= sizeof(int);
942 case HAMMER2_COMP_ZLIB
:
943 comp_level
= HAMMER2_DEC_LEVEL(comp_algo
);
945 comp_level
= 6; /* default zlib compression */
946 else if (comp_level
< 6)
948 else if (comp_level
> 9)
950 ret
= deflateInit(&strm_compress
, comp_level
);
952 kprintf("HAMMER2 ZLIB: fatal error "
953 "on deflateInit.\n");
956 comp_buffer
= objcache_get(cache_buffer_write
,
958 strm_compress
.next_in
= data
;
959 strm_compress
.avail_in
= pblksize
;
960 strm_compress
.next_out
= comp_buffer
;
961 strm_compress
.avail_out
= pblksize
/ 2;
962 ret
= deflate(&strm_compress
, Z_FINISH
);
963 if (ret
== Z_STREAM_END
) {
964 comp_size
= pblksize
/ 2 -
965 strm_compress
.avail_out
;
969 ret
= deflateEnd(&strm_compress
);
972 kprintf("Error: Unknown compression method.\n");
973 kprintf("Comp_method = %d.\n", comp_algo
);
978 if (comp_size
== 0) {
980 * compression failed or turned off
982 comp_block_size
= pblksize
; /* safety */
983 if (++ip
->comp_heuristic
> 128)
984 ip
->comp_heuristic
= 8;
987 * compression succeeded
989 ip
->comp_heuristic
= 0;
990 if (comp_size
<= 1024) {
991 comp_block_size
= 1024;
992 } else if (comp_size
<= 2048) {
993 comp_block_size
= 2048;
994 } else if (comp_size
<= 4096) {
995 comp_block_size
= 4096;
996 } else if (comp_size
<= 8192) {
997 comp_block_size
= 8192;
998 } else if (comp_size
<= 16384) {
999 comp_block_size
= 16384;
1000 } else if (comp_size
<= 32768) {
1001 comp_block_size
= 32768;
1003 panic("hammer2: WRITE PATH: "
1004 "Weird comp_size value.");
1006 comp_block_size
= pblksize
;
1010 * Must zero the remainder or dedup (which operates on a
1011 * physical block basis) will not find matches.
1013 if (comp_size
< comp_block_size
) {
1014 bzero(comp_buffer
+ comp_size
,
1015 comp_block_size
- comp_size
);
1020 * Assign physical storage, bdata will be set to NULL if a live-dedup
1023 bdata
= comp_size
? comp_buffer
: data
;
1024 chain
= hammer2_assign_physical(ip
, parentp
, lbase
, comp_block_size
,
1025 mtid
, &bdata
, errorp
);
1031 if (chain
->bref
.type
== HAMMER2_BREF_TYPE_INODE
) {
1032 hammer2_inode_data_t
*wipdata
;
1034 *errorp
= hammer2_chain_modify_ip(ip
, chain
, mtid
, 0);
1036 wipdata
= &chain
->data
->ipdata
;
1037 KKASSERT(wipdata
->meta
.op_flags
&
1038 HAMMER2_OPFLAG_DIRECTDATA
);
1039 bcopy(data
, wipdata
->u
.data
, HAMMER2_EMBEDDED_BYTES
);
1040 ++hammer2_iod_file_wembed
;
1042 } else if (bdata
== NULL
) {
1044 * Live deduplication, a copy of the data is already present
1048 chain
->bref
.methods
=
1049 HAMMER2_ENC_COMP(comp_algo
) +
1050 HAMMER2_ENC_CHECK(check_algo
);
1052 chain
->bref
.methods
=
1054 HAMMER2_COMP_NONE
) +
1055 HAMMER2_ENC_CHECK(check_algo
);
1057 bdata
= comp_size
? comp_buffer
: data
;
1058 hammer2_chain_setcheck(chain
, bdata
);
1059 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_INITIAL
);
1063 KKASSERT(chain
->flags
& HAMMER2_CHAIN_MODIFIED
);
1065 switch(chain
->bref
.type
) {
1066 case HAMMER2_BREF_TYPE_INODE
:
1067 panic("hammer2_compress_and_write: unexpected inode\n");
1069 case HAMMER2_BREF_TYPE_DATA
:
1071 * Optimize out the read-before-write
1074 *errorp
= hammer2_io_newnz(chain
->hmp
,
1076 chain
->bref
.data_off
,
1080 hammer2_io_brelse(&dio
);
1081 kprintf("hammer2: WRITE PATH: "
1082 "dbp bread error\n");
1085 bdata
= hammer2_io_data(dio
, chain
->bref
.data_off
);
1088 * When loading the block make sure we don't
1089 * leave garbage after the compressed data.
1092 chain
->bref
.methods
=
1093 HAMMER2_ENC_COMP(comp_algo
) +
1094 HAMMER2_ENC_CHECK(check_algo
);
1095 bcopy(comp_buffer
, bdata
, comp_block_size
);
1097 chain
->bref
.methods
=
1099 HAMMER2_COMP_NONE
) +
1100 HAMMER2_ENC_CHECK(check_algo
);
1101 bcopy(data
, bdata
, pblksize
);
1105 * The flush code doesn't calculate check codes for
1106 * file data (doing so can result in excessive I/O),
1109 hammer2_chain_setcheck(chain
, bdata
);
1112 * Device buffer is now valid, chain is no longer in
1113 * the initial state.
1115 * (No blockref table worries with file data)
1117 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_INITIAL
);
1118 hammer2_dedup_record(chain
, dio
, bdata
);
1120 /* Now write the related bdp. */
1121 if (ioflag
& IO_SYNC
) {
1123 * Synchronous I/O requested.
1125 hammer2_io_bwrite(&dio
);
1127 } else if ((ioflag & IO_DIRECT) &&
1128 loff + n == pblksize) {
1129 hammer2_io_bdwrite(&dio);
1131 } else if (ioflag
& IO_ASYNC
) {
1132 hammer2_io_bawrite(&dio
);
1134 hammer2_io_bdwrite(&dio
);
1138 panic("hammer2_compress_and_write: bad chain type %d\n",
1146 hammer2_chain_unlock(chain
);
1147 hammer2_chain_drop(chain
);
1150 objcache_put(cache_buffer_write
, comp_buffer
);
1156 * Function that performs zero-checking and writing without compression,
1157 * it corresponds to default zero-checking path.
1161 hammer2_zero_check_and_write(char *data
, hammer2_inode_t
*ip
,
1162 hammer2_chain_t
**parentp
,
1163 hammer2_key_t lbase
, int ioflag
, int pblksize
,
1164 hammer2_tid_t mtid
, int *errorp
,
1167 hammer2_chain_t
*chain
;
1170 if (check_algo
!= HAMMER2_CHECK_NONE
&&
1171 test_block_zeros(data
, pblksize
)) {
1173 * An all-zeros write creates a hole unless the check code
1174 * is disabled. When the check code is disabled all writes
1175 * are done in-place, including any all-zeros writes.
1177 * NOTE: A snapshot will still force a copy-on-write
1178 * (see the HAMMER2_CHECK_NONE in hammer2_chain.c).
1180 zero_write(data
, ip
, parentp
, lbase
, mtid
, errorp
);
1183 * Normal write (bdata set to NULL if de-duplicated)
1186 chain
= hammer2_assign_physical(ip
, parentp
, lbase
, pblksize
,
1187 mtid
, &bdata
, errorp
);
1191 hammer2_write_bp(chain
, data
, ioflag
, pblksize
,
1192 mtid
, errorp
, check_algo
);
1194 /* dedup occurred */
1195 chain
->bref
.methods
=
1196 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE
) +
1197 HAMMER2_ENC_CHECK(check_algo
);
1198 hammer2_chain_setcheck(chain
, data
);
1199 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_INITIAL
);
1202 hammer2_chain_unlock(chain
);
1203 hammer2_chain_drop(chain
);
1211 * A function to test whether a block of data contains only zeros,
1212 * returns TRUE (non-zero) if the block is all zeros.
1216 test_block_zeros(const char *buf
, size_t bytes
)
1220 for (i
= 0; i
< bytes
; i
+= sizeof(long)) {
1221 if (*(const long *)(buf
+ i
) != 0)
1230 * Function to "write" a block that contains only zeros.
1234 zero_write(char *data
, hammer2_inode_t
*ip
,
1235 hammer2_chain_t
**parentp
,
1236 hammer2_key_t lbase
, hammer2_tid_t mtid
, int *errorp
)
1238 hammer2_chain_t
*chain
;
1239 hammer2_key_t key_dummy
;
1241 chain
= hammer2_chain_lookup(parentp
, &key_dummy
,
1244 HAMMER2_LOOKUP_NODATA
);
1246 if (chain
->bref
.type
== HAMMER2_BREF_TYPE_INODE
) {
1247 hammer2_inode_data_t
*wipdata
;
1250 *errorp
= hammer2_chain_modify_ip(ip
, chain
,
1254 wipdata
= &chain
->data
->ipdata
;
1255 KKASSERT(wipdata
->meta
.op_flags
&
1256 HAMMER2_OPFLAG_DIRECTDATA
);
1257 bzero(wipdata
->u
.data
, HAMMER2_EMBEDDED_BYTES
);
1258 ++hammer2_iod_file_wembed
;
1261 /* chain->error ok for deletion */
1262 hammer2_chain_delete(*parentp
, chain
,
1263 mtid
, HAMMER2_DELETE_PERMANENT
);
1264 ++hammer2_iod_file_wzero
;
1266 atomic_set_int(&ip
->flags
, HAMMER2_INODE_DIRTYDATA
);
1267 hammer2_chain_unlock(chain
);
1268 hammer2_chain_drop(chain
);
1270 ++hammer2_iod_file_wzero
;
1277 * Function to write the data as it is, without performing any sort of
1278 * compression. This function is used in path without compression and
1279 * default zero-checking path.
1283 hammer2_write_bp(hammer2_chain_t
*chain
, char *data
, int ioflag
,
1285 hammer2_tid_t mtid
, int *errorp
, int check_algo
)
1287 hammer2_inode_data_t
*wipdata
;
1292 error
= 0; /* XXX TODO below */
1294 KKASSERT(chain
->flags
& HAMMER2_CHAIN_MODIFIED
);
1296 switch(chain
->bref
.type
) {
1297 case HAMMER2_BREF_TYPE_INODE
:
1298 wipdata
= &chain
->data
->ipdata
;
1299 KKASSERT(wipdata
->meta
.op_flags
& HAMMER2_OPFLAG_DIRECTDATA
);
1300 bcopy(data
, wipdata
->u
.data
, HAMMER2_EMBEDDED_BYTES
);
1302 ++hammer2_iod_file_wembed
;
1304 case HAMMER2_BREF_TYPE_DATA
:
1305 error
= hammer2_io_newnz(chain
->hmp
,
1307 chain
->bref
.data_off
,
1308 chain
->bytes
, &dio
);
1310 hammer2_io_bqrelse(&dio
);
1311 kprintf("hammer2: WRITE PATH: "
1312 "dbp bread error\n");
1315 bdata
= hammer2_io_data(dio
, chain
->bref
.data_off
);
1317 chain
->bref
.methods
= HAMMER2_ENC_COMP(HAMMER2_COMP_NONE
) +
1318 HAMMER2_ENC_CHECK(check_algo
);
1319 bcopy(data
, bdata
, chain
->bytes
);
1322 * The flush code doesn't calculate check codes for
1323 * file data (doing so can result in excessive I/O),
1326 hammer2_chain_setcheck(chain
, bdata
);
1329 * Device buffer is now valid, chain is no longer in
1330 * the initial state.
1332 * (No blockref table worries with file data)
1334 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_INITIAL
);
1335 hammer2_dedup_record(chain
, dio
, bdata
);
1337 if (ioflag
& IO_SYNC
) {
1339 * Synchronous I/O requested.
1341 hammer2_io_bwrite(&dio
);
1343 } else if ((ioflag & IO_DIRECT) &&
1344 loff + n == pblksize) {
1345 hammer2_io_bdwrite(&dio);
1347 } else if (ioflag
& IO_ASYNC
) {
1348 hammer2_io_bawrite(&dio
);
1350 hammer2_io_bdwrite(&dio
);
1354 panic("hammer2_write_bp: bad chain type %d\n",
1364 * LIVE DEDUP HEURISTICS
1366 * Record media and crc information for possible dedup operation. Note
1367 * that the dedup mask bits must also be set in the related DIO for a dedup
1368 * to be fully validated (which is handled in the freemap allocation code).
1370 * WARNING! This code is SMP safe but the heuristic allows SMP collisions.
1371 * All fields must be loaded into locals and validated.
1373 * WARNING! Should only be used for file data and directory entries,
1374 * hammer2_chain_modify() only checks for the dedup case on data
1375 * chains. Also, dedup data can only be recorded for committed
1376 * chains (so NOT strategy writes which can undergo further
1377 * modification after the fact!).
1380 hammer2_dedup_record(hammer2_chain_t
*chain
, hammer2_io_t
*dio
,
1384 hammer2_dedup_t
*dedup
;
1392 * We can only record a dedup if we have media data to test against.
1393 * If dedup is not enabled, return early, which allows a chain to
1394 * remain marked MODIFIED (which might have benefits in special
1395 * situations, though typically it does not).
1397 if (hammer2_dedup_enable
== 0)
1407 switch(HAMMER2_DEC_CHECK(chain
->bref
.methods
)) {
1408 case HAMMER2_CHECK_ISCSI32
:
1410 * XXX use the built-in crc (the dedup lookup sequencing
1411 * needs to be fixed so the check code is already present
1412 * when dedup_lookup is called)
1415 crc
= (uint64_t)(uint32_t)chain
->bref
.check
.iscsi32
.value
;
1417 crc
= XXH64(data
, chain
->bytes
, XXH_HAMMER2_SEED
);
1419 case HAMMER2_CHECK_XXHASH64
:
1420 crc
= chain
->bref
.check
.xxhash64
.value
;
1422 case HAMMER2_CHECK_SHA192
:
1424 * XXX use the built-in crc (the dedup lookup sequencing
1425 * needs to be fixed so the check code is already present
1426 * when dedup_lookup is called)
1429 crc
= ((uint64_t *)chain
->bref
.check
.sha192
.data
)[0] ^
1430 ((uint64_t *)chain
->bref
.check
.sha192
.data
)[1] ^
1431 ((uint64_t *)chain
->bref
.check
.sha192
.data
)[2];
1433 crc
= XXH64(data
, chain
->bytes
, XXH_HAMMER2_SEED
);
1437 * Cannot dedup without a check code
1439 * NOTE: In particular, CHECK_NONE allows a sector to be
1440 * overwritten without copy-on-write, recording
1441 * a dedup block for a CHECK_NONE object would be
1447 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_DEDUPABLE
);
1449 dedup
= &hmp
->heur_dedup
[crc
& (HAMMER2_DEDUP_HEUR_MASK
& ~3)];
1450 for (i
= 0; i
< 4; ++i
) {
1451 if (dedup
[i
].data_crc
== crc
) {
1455 dticks
= (int)(dedup
[i
].ticks
- dedup
[best
].ticks
);
1456 if (dticks
< 0 || dticks
> hz
* 60 * 30)
1460 if (hammer2_debug
& 0x40000) {
1461 kprintf("REC %04x %016jx %016jx\n",
1462 (int)(dedup
- hmp
->heur_dedup
),
1464 chain
->bref
.data_off
);
1466 dedup
->ticks
= ticks
;
1467 dedup
->data_off
= chain
->bref
.data_off
;
1468 dedup
->data_crc
= crc
;
1471 * Set the valid bits for the dedup only after we know the data
1472 * buffer has been updated. The alloc bits were set (and the valid
1473 * bits cleared) when the media was allocated.
1475 * This is done in two stages becuase the bulkfree code can race
1476 * the gap between allocation and data population. Both masks must
1477 * be set before a bcmp/dedup operation is able to use the block.
1479 mask
= hammer2_dedup_mask(dio
, chain
->bref
.data_off
, chain
->bytes
);
1480 atomic_set_64(&dio
->dedup_valid
, mask
);
1484 * XXX removed. MODIFIED is an integral part of the flush code,
1485 * lets not just clear it
1488 * Once we record the dedup the chain must be marked clean to
1489 * prevent reuse of the underlying block. Remember that this
1490 * write occurs when the buffer cache is flushed (i.e. on sync(),
1491 * fsync(), filesystem periodic sync, or when the kernel needs to
1492 * flush a buffer), and not whenever the user write()s.
1494 if (chain
->flags
& HAMMER2_CHAIN_MODIFIED
) {
1495 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_MODIFIED
);
1496 atomic_add_long(&hammer2_count_modified_chains
, -1);
1498 hammer2_pfs_memory_wakeup(chain
->pmp
, -1);
1505 hammer2_dedup_lookup(hammer2_dev_t
*hmp
, char **datap
, int pblksize
)
1507 hammer2_dedup_t
*dedup
;
1516 if (hammer2_dedup_enable
== 0)
1523 * XXX use the built-in crc (the dedup lookup sequencing
1524 * needs to be fixed so the check code is already present
1525 * when dedup_lookup is called)
1527 crc
= XXH64(data
, pblksize
, XXH_HAMMER2_SEED
);
1528 dedup
= &hmp
->heur_dedup
[crc
& (HAMMER2_DEDUP_HEUR_MASK
& ~3)];
1530 if (hammer2_debug
& 0x40000) {
1531 kprintf("LOC %04x/4 %016jx\n",
1532 (int)(dedup
- hmp
->heur_dedup
),
1536 for (i
= 0; i
< 4; ++i
) {
1537 off
= dedup
[i
].data_off
;
1539 if (dedup
[i
].data_crc
!= crc
)
1541 if ((1 << (int)(off
& HAMMER2_OFF_MASK_RADIX
)) != pblksize
)
1543 dio
= hammer2_io_getquick(hmp
, off
, pblksize
);
1545 dtmp
= hammer2_io_data(dio
, off
),
1546 mask
= hammer2_dedup_mask(dio
, off
, pblksize
);
1547 if ((dio
->dedup_alloc
& mask
) == mask
&&
1548 (dio
->dedup_valid
& mask
) == mask
&&
1549 bcmp(data
, dtmp
, pblksize
) == 0) {
1550 if (hammer2_debug
& 0x40000) {
1551 kprintf("DEDUP SUCCESS %016jx\n",
1554 hammer2_io_putblk(&dio
);
1556 dedup
[i
].ticks
= ticks
; /* update use */
1557 atomic_add_long(&hammer2_iod_file_wdedup
,
1560 return off
; /* RETURN */
1562 hammer2_io_putblk(&dio
);
1569 * Poof. Races are ok, if someone gets in and reuses a dedup offset
1570 * before or while we are clearing it they will also recover the freemap
1571 * entry (set it to fully allocated), so a bulkfree race can only set it
1572 * to a possibly-free state.
1574 * XXX ok, well, not really sure races are ok but going to run with it
1578 hammer2_dedup_clear(hammer2_dev_t
*hmp
)
1582 for (i
= 0; i
< HAMMER2_DEDUP_HEUR_SIZE
; ++i
) {
1583 hmp
->heur_dedup
[i
].data_off
= 0;
1584 hmp
->heur_dedup
[i
].ticks
= ticks
- 1;