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>
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 * NOTE: The chain->data for xop->head.cluster.focus will be
377 * synchronized to the current cpu by xop_collect(),
378 * but other chains in the cluster might not be.
380 error
= hammer2_xop_collect(&xop
->head
, HAMMER2_XOP_COLLECT_NOWAIT
);
385 hammer2_mtx_unlock(&xop
->lock
);
386 bp
->b_flags
|= B_NOTMETA
;
387 chain
= xop
->head
.cluster
.focus
;
388 hammer2_strategy_read_completion(chain
, (char *)chain
->data
,
391 hammer2_xop_retire(&xop
->head
, HAMMER2_XOPMASK_VOP
);
393 case HAMMER2_ERROR_ENOENT
:
395 hammer2_mtx_unlock(&xop
->lock
);
396 bp
->b_flags
|= B_NOTMETA
;
399 bzero(bp
->b_data
, bp
->b_bcount
);
401 hammer2_xop_retire(&xop
->head
, HAMMER2_XOPMASK_VOP
);
403 case HAMMER2_ERROR_EINPROGRESS
:
404 hammer2_mtx_unlock(&xop
->lock
);
407 kprintf("strategy_xop_read: error %08x loff=%016jx\n",
408 error
, bp
->b_loffset
);
410 hammer2_mtx_unlock(&xop
->lock
);
411 bp
->b_flags
|= B_ERROR
;
414 hammer2_xop_retire(&xop
->head
, HAMMER2_XOPMASK_VOP
);
421 hammer2_strategy_read_completion(hammer2_chain_t
*chain
, char *data
,
424 struct buf
*bp
= bio
->bio_buf
;
426 if (chain
->bref
.type
== HAMMER2_BREF_TYPE_INODE
) {
428 * Copy from in-memory inode structure.
430 bcopy(((hammer2_inode_data_t
*)data
)->u
.data
,
431 bp
->b_data
, HAMMER2_EMBEDDED_BYTES
);
432 bzero(bp
->b_data
+ HAMMER2_EMBEDDED_BYTES
,
433 bp
->b_bcount
- HAMMER2_EMBEDDED_BYTES
);
436 } else if (chain
->bref
.type
== HAMMER2_BREF_TYPE_DATA
) {
438 * Data is on-media, record for live dedup. Release the
439 * chain (try to free it) when done. The data is still
440 * cached by both the buffer cache in front and the
441 * block device behind us. This leaves more room in the
442 * LRU chain cache for meta-data chains which we really
445 * NOTE: Deduplication cannot be safely recorded for
446 * records without a check code.
448 hammer2_dedup_record(chain
, NULL
, data
);
449 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_RELEASE
);
452 * Decompression and copy.
454 switch (HAMMER2_DEC_COMP(chain
->bref
.methods
)) {
455 case HAMMER2_COMP_LZ4
:
456 hammer2_decompress_LZ4_callback(data
, chain
->bytes
,
458 /* b_resid set by call */
460 case HAMMER2_COMP_ZLIB
:
461 hammer2_decompress_ZLIB_callback(data
, chain
->bytes
,
463 /* b_resid set by call */
465 case HAMMER2_COMP_NONE
:
466 KKASSERT(chain
->bytes
<= bp
->b_bcount
);
467 bcopy(data
, bp
->b_data
, chain
->bytes
);
468 if (chain
->bytes
< bp
->b_bcount
) {
469 bzero(bp
->b_data
+ chain
->bytes
,
470 bp
->b_bcount
- chain
->bytes
);
476 panic("hammer2_strategy_read: "
477 "unknown compression type");
480 panic("hammer2_strategy_read: unknown bref type");
484 /****************************************************************************
486 ****************************************************************************/
489 * Functions for compression in threads,
490 * from hammer2_vnops.c
492 static void hammer2_write_file_core(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 static void hammer2_compress_and_write(char *data
, hammer2_inode_t
*ip
,
497 hammer2_chain_t
**parentp
,
498 hammer2_key_t lbase
, int ioflag
, int pblksize
,
499 hammer2_tid_t mtid
, int *errorp
,
500 int comp_algo
, int check_algo
);
501 static void hammer2_zero_check_and_write(char *data
, hammer2_inode_t
*ip
,
502 hammer2_chain_t
**parentp
,
503 hammer2_key_t lbase
, int ioflag
, int pblksize
,
504 hammer2_tid_t mtid
, int *errorp
,
506 static int test_block_zeros(const char *buf
, size_t bytes
);
507 static void zero_write(char *data
, hammer2_inode_t
*ip
,
508 hammer2_chain_t
**parentp
,
510 hammer2_tid_t mtid
, int *errorp
);
511 static void hammer2_write_bp(hammer2_chain_t
*chain
, char *data
,
512 int ioflag
, int pblksize
,
513 hammer2_tid_t mtid
, int *errorp
,
518 hammer2_strategy_write(struct vop_strategy_args
*ap
)
520 hammer2_xop_strategy_t
*xop
;
531 atomic_set_int(&ip
->flags
, HAMMER2_INODE_DIRTYDATA
);
532 hammer2_lwinprog_ref(pmp
);
533 hammer2_trans_assert_strategy(pmp
);
534 hammer2_trans_init(pmp
, HAMMER2_TRANS_BUFCACHE
);
536 xop
= hammer2_xop_alloc(ip
, HAMMER2_XOP_MODIFYING
|
537 HAMMER2_XOP_STRATEGY
);
540 xop
->lbase
= bio
->bio_offset
;
541 hammer2_mtx_init(&xop
->lock
, "h2biow");
542 hammer2_xop_start(&xop
->head
, hammer2_strategy_xop_write
);
543 /* asynchronous completion */
545 hammer2_lwinprog_wait(pmp
, hammer2_flush_pipe
);
551 * Per-node XOP (threaded). Write the logical buffer to the media.
553 * This is a bit problematic because there may be multiple target and
554 * any of them may be able to release the bp. In addition, if our
555 * particulr target is offline we don't want to block the bp (and thus
556 * the frontend). To accomplish this we copy the data to the per-thr
561 hammer2_strategy_xop_write(hammer2_thread_t
*thr
, hammer2_xop_t
*arg
)
563 hammer2_xop_strategy_t
*xop
= &arg
->xop_strategy
;
564 hammer2_chain_t
*parent
;
572 hammer2_off_t bio_offset
;
576 * We can only access the bp/bio if the frontend has not yet
581 hammer2_mtx_sh(&xop
->lock
);
583 hammer2_mtx_unlock(&xop
->lock
);
588 bio
= xop
->bio
; /* ephermal */
589 bp
= bio
->bio_buf
; /* ephermal */
590 ip
= xop
->head
.ip1
; /* retained by ref */
591 bio_offset
= bio
->bio_offset
;
592 bio_data
= thr
->scratch
;
594 /* hammer2_trans_init(parent->hmp->spmp, HAMMER2_TRANS_BUFCACHE); */
596 lblksize
= hammer2_calc_logical(ip
, bio
->bio_offset
, &lbase
, NULL
);
597 pblksize
= hammer2_calc_physical(ip
, lbase
);
599 KKASSERT(lblksize
<= MAXPHYS
);
600 bcopy(bp
->b_data
, bio_data
, lblksize
);
602 hammer2_mtx_unlock(&xop
->lock
);
603 bp
= NULL
; /* safety, illegal to access after unlock */
604 bio
= NULL
; /* safety, illegal to access after unlock */
609 parent
= hammer2_inode_chain(ip
, thr
->clindex
, HAMMER2_RESOLVE_ALWAYS
);
610 hammer2_write_file_core(bio_data
, ip
, &parent
,
611 lbase
, IO_ASYNC
, pblksize
,
612 xop
->head
.mtid
, &error
);
614 hammer2_chain_unlock(parent
);
615 hammer2_chain_drop(parent
);
616 parent
= NULL
; /* safety */
618 hammer2_xop_feed(&xop
->head
, NULL
, thr
->clindex
, error
);
621 * Try to complete the operation on behalf of the front-end.
625 hammer2_mtx_ex(&xop
->lock
);
627 hammer2_mtx_unlock(&xop
->lock
);
632 * Async operation has not completed and we now own the lock.
633 * Determine if we can complete the operation by issuing the
634 * frontend collection non-blocking.
636 * H2 double-buffers the data, setting B_NOTMETA on the logical
637 * buffer hints to the OS that the logical buffer should not be
638 * swapcached (since the device buffer can be).
640 error
= hammer2_xop_collect(&xop
->head
, HAMMER2_XOP_COLLECT_NOWAIT
);
642 if (error
== HAMMER2_ERROR_EINPROGRESS
) {
643 hammer2_mtx_unlock(&xop
->lock
);
648 * Async operation has completed.
651 hammer2_mtx_unlock(&xop
->lock
);
653 bio
= xop
->bio
; /* now owned by us */
654 bp
= bio
->bio_buf
; /* now owned by us */
656 if (error
== HAMMER2_ERROR_ENOENT
|| error
== 0) {
657 bp
->b_flags
|= B_NOTMETA
;
662 kprintf("strategy_xop_write: error %d loff=%016jx\n",
663 error
, bp
->b_loffset
);
664 bp
->b_flags
|= B_ERROR
;
668 hammer2_xop_retire(&xop
->head
, HAMMER2_XOPMASK_VOP
);
669 hammer2_trans_assert_strategy(ip
->pmp
);
670 hammer2_lwinprog_drop(ip
->pmp
);
671 hammer2_trans_done(ip
->pmp
);
675 * Wait for pending I/O to complete
678 hammer2_bioq_sync(hammer2_pfs_t
*pmp
)
680 hammer2_lwinprog_wait(pmp
, 0);
684 * Assign physical storage at (cparent, lbase), returning a suitable chain
685 * and setting *errorp appropriately.
687 * If no error occurs, the returned chain will be in a modified state.
689 * If an error occurs, the returned chain may or may not be NULL. If
690 * not-null any chain->error (if not 0) will also be rolled up into *errorp.
691 * So the caller only needs to test *errorp.
693 * cparent can wind up being anything.
695 * If datap is not NULL, *datap points to the real data we intend to write.
696 * If we can dedup the storage location we set *datap to NULL to indicate
697 * to the caller that a dedup occurred.
699 * NOTE: Special case for data embedded in inode.
703 hammer2_assign_physical(hammer2_inode_t
*ip
, hammer2_chain_t
**parentp
,
704 hammer2_key_t lbase
, int pblksize
,
705 hammer2_tid_t mtid
, char **datap
, int *errorp
)
707 hammer2_chain_t
*chain
;
708 hammer2_key_t key_dummy
;
709 hammer2_off_t dedup_off
;
710 int pradix
= hammer2_getradix(pblksize
);
713 * Locate the chain associated with lbase, return a locked chain.
714 * However, do not instantiate any data reference (which utilizes a
715 * device buffer) because we will be using direct IO via the
716 * logical buffer cache buffer.
718 KKASSERT(pblksize
>= HAMMER2_ALLOC_MIN
);
720 chain
= hammer2_chain_lookup(parentp
, &key_dummy
,
723 HAMMER2_LOOKUP_NODATA
);
726 * The lookup code should not return a DELETED chain to us, unless
727 * its a short-file embedded in the inode. Then it is possible for
728 * the lookup to return a deleted inode.
730 if (chain
&& (chain
->flags
& HAMMER2_CHAIN_DELETED
) &&
731 chain
->bref
.type
!= HAMMER2_BREF_TYPE_INODE
) {
732 kprintf("assign physical deleted chain @ "
733 "%016jx (%016jx.%02x) ip %016jx\n",
734 lbase
, chain
->bref
.data_off
, chain
->bref
.type
,
741 * We found a hole, create a new chain entry.
743 * NOTE: DATA chains are created without device backing
744 * store (nor do we want any).
746 dedup_off
= hammer2_dedup_lookup((*parentp
)->hmp
, datap
,
748 *errorp
|= hammer2_chain_create(parentp
, &chain
,
750 HAMMER2_ENC_CHECK(ip
->meta
.check_algo
) |
751 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE
),
752 lbase
, HAMMER2_PBUFRADIX
,
753 HAMMER2_BREF_TYPE_DATA
,
758 /*ip->delta_dcount += pblksize;*/
759 } else if (chain
->error
== 0) {
760 switch (chain
->bref
.type
) {
761 case HAMMER2_BREF_TYPE_INODE
:
763 * The data is embedded in the inode, which requires
766 *errorp
|= hammer2_chain_modify_ip(ip
, chain
, mtid
, 0);
768 case HAMMER2_BREF_TYPE_DATA
:
769 dedup_off
= hammer2_dedup_lookup(chain
->hmp
, datap
,
771 if (chain
->bytes
!= pblksize
) {
772 *errorp
|= hammer2_chain_resize(chain
,
775 HAMMER2_MODIFY_OPTDATA
);
781 * DATA buffers must be marked modified whether the
782 * data is in a logical buffer or not. We also have
783 * to make this call to fixup the chain data pointers
784 * after resizing in case this is an encrypted or
787 *errorp
|= hammer2_chain_modify(chain
, mtid
, dedup_off
,
788 HAMMER2_MODIFY_OPTDATA
);
791 panic("hammer2_assign_physical: bad type");
796 *errorp
= chain
->error
;
798 atomic_set_int(&ip
->flags
, HAMMER2_INODE_DIRTYDATA
);
804 * hammer2_write_file_core() - hammer2_write_thread() helper
806 * The core write function which determines which path to take
807 * depending on compression settings. We also have to locate the
808 * related chains so we can calculate and set the check data for
813 hammer2_write_file_core(char *data
, hammer2_inode_t
*ip
,
814 hammer2_chain_t
**parentp
,
815 hammer2_key_t lbase
, int ioflag
, int pblksize
,
816 hammer2_tid_t mtid
, int *errorp
)
818 hammer2_chain_t
*chain
;
823 switch(HAMMER2_DEC_ALGO(ip
->meta
.comp_algo
)) {
824 case HAMMER2_COMP_NONE
:
826 * We have to assign physical storage to the buffer
827 * we intend to dirty or write now to avoid deadlocks
828 * in the strategy code later.
830 * This can return NOOFFSET for inode-embedded data.
831 * The strategy code will take care of it in that case.
834 chain
= hammer2_assign_physical(ip
, parentp
, lbase
, pblksize
,
835 mtid
, &bdata
, errorp
);
837 /* skip modifications */
838 } else if (chain
->bref
.type
== HAMMER2_BREF_TYPE_INODE
) {
839 hammer2_inode_data_t
*wipdata
;
841 wipdata
= &chain
->data
->ipdata
;
842 KKASSERT(wipdata
->meta
.op_flags
&
843 HAMMER2_OPFLAG_DIRECTDATA
);
844 bcopy(data
, wipdata
->u
.data
, HAMMER2_EMBEDDED_BYTES
);
845 ++hammer2_iod_file_wembed
;
846 } else if (bdata
== NULL
) {
848 * Copy of data already present on-media.
850 chain
->bref
.methods
=
851 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE
) +
852 HAMMER2_ENC_CHECK(ip
->meta
.check_algo
);
853 hammer2_chain_setcheck(chain
, data
);
855 hammer2_write_bp(chain
, data
, ioflag
, pblksize
,
856 mtid
, errorp
, ip
->meta
.check_algo
);
859 hammer2_chain_unlock(chain
);
860 hammer2_chain_drop(chain
);
863 case HAMMER2_COMP_AUTOZERO
:
865 * Check for zero-fill only
867 hammer2_zero_check_and_write(data
, ip
, parentp
,
868 lbase
, ioflag
, pblksize
,
870 ip
->meta
.check_algo
);
872 case HAMMER2_COMP_LZ4
:
873 case HAMMER2_COMP_ZLIB
:
876 * Check for zero-fill and attempt compression.
878 hammer2_compress_and_write(data
, ip
, parentp
,
879 lbase
, ioflag
, pblksize
,
882 ip
->meta
.check_algo
);
890 * Generic function that will perform the compression in compression
891 * write path. The compression algorithm is determined by the settings
892 * obtained from inode.
896 hammer2_compress_and_write(char *data
, hammer2_inode_t
*ip
,
897 hammer2_chain_t
**parentp
,
898 hammer2_key_t lbase
, int ioflag
, int pblksize
,
899 hammer2_tid_t mtid
, int *errorp
, int comp_algo
, int check_algo
)
901 hammer2_chain_t
*chain
;
908 * An all-zeros write creates a hole unless the check code
909 * is disabled. When the check code is disabled all writes
910 * are done in-place, including any all-zeros writes.
912 * NOTE: A snapshot will still force a copy-on-write
913 * (see the HAMMER2_CHECK_NONE in hammer2_chain.c).
915 if (check_algo
!= HAMMER2_CHECK_NONE
&&
916 test_block_zeros(data
, pblksize
)) {
917 zero_write(data
, ip
, parentp
, lbase
, mtid
, errorp
);
922 * Compression requested. Try to compress the block. We store
923 * the data normally if we cannot sufficiently compress it.
925 * We have a heuristic to detect files which are mostly
926 * uncompressable and avoid the compression attempt in that
927 * case. If the compression heuristic is turned off, we always
933 KKASSERT(pblksize
/ 2 <= 32768);
935 if (ip
->comp_heuristic
< 8 || (ip
->comp_heuristic
& 7) == 0 ||
936 hammer2_always_compress
) {
937 z_stream strm_compress
;
941 switch(HAMMER2_DEC_ALGO(comp_algo
)) {
942 case HAMMER2_COMP_LZ4
:
944 * We need to prefix with the size, LZ4
945 * doesn't do it for us. Add the related
948 * NOTE: The LZ4 code seems to assume at least an
949 * 8-byte buffer size granularity and may
950 * overrun the buffer if given a 4-byte
953 comp_buffer
= objcache_get(cache_buffer_write
,
955 comp_size
= LZ4_compress_limitedOutput(
957 &comp_buffer
[sizeof(int)],
959 pblksize
/ 2 - sizeof(int64_t));
960 *(int *)comp_buffer
= comp_size
;
962 comp_size
+= sizeof(int);
964 case HAMMER2_COMP_ZLIB
:
965 comp_level
= HAMMER2_DEC_LEVEL(comp_algo
);
967 comp_level
= 6; /* default zlib compression */
968 else if (comp_level
< 6)
970 else if (comp_level
> 9)
972 ret
= deflateInit(&strm_compress
, comp_level
);
974 kprintf("HAMMER2 ZLIB: fatal error "
975 "on deflateInit.\n");
978 comp_buffer
= objcache_get(cache_buffer_write
,
980 strm_compress
.next_in
= data
;
981 strm_compress
.avail_in
= pblksize
;
982 strm_compress
.next_out
= comp_buffer
;
983 strm_compress
.avail_out
= pblksize
/ 2;
984 ret
= deflate(&strm_compress
, Z_FINISH
);
985 if (ret
== Z_STREAM_END
) {
986 comp_size
= pblksize
/ 2 -
987 strm_compress
.avail_out
;
991 ret
= deflateEnd(&strm_compress
);
994 kprintf("Error: Unknown compression method.\n");
995 kprintf("Comp_method = %d.\n", comp_algo
);
1000 if (comp_size
== 0) {
1002 * compression failed or turned off
1004 comp_block_size
= pblksize
; /* safety */
1005 if (++ip
->comp_heuristic
> 128)
1006 ip
->comp_heuristic
= 8;
1009 * compression succeeded
1011 ip
->comp_heuristic
= 0;
1012 if (comp_size
<= 1024) {
1013 comp_block_size
= 1024;
1014 } else if (comp_size
<= 2048) {
1015 comp_block_size
= 2048;
1016 } else if (comp_size
<= 4096) {
1017 comp_block_size
= 4096;
1018 } else if (comp_size
<= 8192) {
1019 comp_block_size
= 8192;
1020 } else if (comp_size
<= 16384) {
1021 comp_block_size
= 16384;
1022 } else if (comp_size
<= 32768) {
1023 comp_block_size
= 32768;
1025 panic("hammer2: WRITE PATH: "
1026 "Weird comp_size value.");
1028 comp_block_size
= pblksize
;
1032 * Must zero the remainder or dedup (which operates on a
1033 * physical block basis) will not find matches.
1035 if (comp_size
< comp_block_size
) {
1036 bzero(comp_buffer
+ comp_size
,
1037 comp_block_size
- comp_size
);
1042 * Assign physical storage, data will be set to NULL if a live-dedup
1045 bdata
= comp_size
? comp_buffer
: data
;
1046 chain
= hammer2_assign_physical(ip
, parentp
, lbase
, comp_block_size
,
1047 mtid
, &bdata
, errorp
);
1053 if (chain
->bref
.type
== HAMMER2_BREF_TYPE_INODE
) {
1054 hammer2_inode_data_t
*wipdata
;
1056 *errorp
= hammer2_chain_modify_ip(ip
, chain
, mtid
, 0);
1058 wipdata
= &chain
->data
->ipdata
;
1059 KKASSERT(wipdata
->meta
.op_flags
&
1060 HAMMER2_OPFLAG_DIRECTDATA
);
1061 bcopy(data
, wipdata
->u
.data
, HAMMER2_EMBEDDED_BYTES
);
1062 ++hammer2_iod_file_wembed
;
1064 } else if (bdata
== NULL
) {
1066 * Live deduplication, a copy of the data is already present
1070 chain
->bref
.methods
=
1071 HAMMER2_ENC_COMP(comp_algo
) +
1072 HAMMER2_ENC_CHECK(check_algo
);
1074 chain
->bref
.methods
=
1076 HAMMER2_COMP_NONE
) +
1077 HAMMER2_ENC_CHECK(check_algo
);
1079 bdata
= comp_size
? comp_buffer
: data
;
1080 hammer2_chain_setcheck(chain
, bdata
);
1081 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_INITIAL
);
1085 KKASSERT(chain
->flags
& HAMMER2_CHAIN_MODIFIED
);
1087 switch(chain
->bref
.type
) {
1088 case HAMMER2_BREF_TYPE_INODE
:
1089 panic("hammer2_write_bp: unexpected inode\n");
1091 case HAMMER2_BREF_TYPE_DATA
:
1093 * Optimize out the read-before-write
1096 *errorp
= hammer2_io_newnz(chain
->hmp
,
1098 chain
->bref
.data_off
,
1102 hammer2_io_brelse(&dio
);
1103 kprintf("hammer2: WRITE PATH: "
1104 "dbp bread error\n");
1107 bdata
= hammer2_io_data(dio
, chain
->bref
.data_off
);
1110 * When loading the block make sure we don't
1111 * leave garbage after the compressed data.
1114 chain
->bref
.methods
=
1115 HAMMER2_ENC_COMP(comp_algo
) +
1116 HAMMER2_ENC_CHECK(check_algo
);
1117 bcopy(comp_buffer
, bdata
, comp_size
);
1119 chain
->bref
.methods
=
1121 HAMMER2_COMP_NONE
) +
1122 HAMMER2_ENC_CHECK(check_algo
);
1123 bcopy(data
, bdata
, pblksize
);
1127 * The flush code doesn't calculate check codes for
1128 * file data (doing so can result in excessive I/O),
1131 hammer2_chain_setcheck(chain
, bdata
);
1134 * Device buffer is now valid, chain is no longer in
1135 * the initial state.
1137 * (No blockref table worries with file data)
1139 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_INITIAL
);
1140 hammer2_dedup_record(chain
, dio
, bdata
);
1142 /* Now write the related bdp. */
1143 if (ioflag
& IO_SYNC
) {
1145 * Synchronous I/O requested.
1147 hammer2_io_bwrite(&dio
);
1149 } else if ((ioflag & IO_DIRECT) &&
1150 loff + n == pblksize) {
1151 hammer2_io_bdwrite(&dio);
1153 } else if (ioflag
& IO_ASYNC
) {
1154 hammer2_io_bawrite(&dio
);
1156 hammer2_io_bdwrite(&dio
);
1160 panic("hammer2_write_bp: bad chain type %d\n",
1168 hammer2_chain_unlock(chain
);
1169 hammer2_chain_drop(chain
);
1172 objcache_put(cache_buffer_write
, comp_buffer
);
1178 * Function that performs zero-checking and writing without compression,
1179 * it corresponds to default zero-checking path.
1183 hammer2_zero_check_and_write(char *data
, hammer2_inode_t
*ip
,
1184 hammer2_chain_t
**parentp
,
1185 hammer2_key_t lbase
, int ioflag
, int pblksize
,
1186 hammer2_tid_t mtid
, int *errorp
,
1189 hammer2_chain_t
*chain
;
1192 if (check_algo
!= HAMMER2_CHECK_NONE
&&
1193 test_block_zeros(data
, pblksize
)) {
1195 * An all-zeros write creates a hole unless the check code
1196 * is disabled. When the check code is disabled all writes
1197 * are done in-place, including any all-zeros writes.
1199 * NOTE: A snapshot will still force a copy-on-write
1200 * (see the HAMMER2_CHECK_NONE in hammer2_chain.c).
1202 zero_write(data
, ip
, parentp
, lbase
, mtid
, errorp
);
1208 chain
= hammer2_assign_physical(ip
, parentp
, lbase
, pblksize
,
1209 mtid
, &bdata
, errorp
);
1213 hammer2_write_bp(chain
, data
, ioflag
, pblksize
,
1214 mtid
, errorp
, check_algo
);
1216 /* dedup occurred */
1217 chain
->bref
.methods
=
1218 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE
) +
1219 HAMMER2_ENC_CHECK(check_algo
);
1220 hammer2_chain_setcheck(chain
, data
);
1223 hammer2_chain_unlock(chain
);
1224 hammer2_chain_drop(chain
);
1232 * A function to test whether a block of data contains only zeros,
1233 * returns TRUE (non-zero) if the block is all zeros.
1237 test_block_zeros(const char *buf
, size_t bytes
)
1241 for (i
= 0; i
< bytes
; i
+= sizeof(long)) {
1242 if (*(const long *)(buf
+ i
) != 0)
1251 * Function to "write" a block that contains only zeros.
1255 zero_write(char *data
, hammer2_inode_t
*ip
,
1256 hammer2_chain_t
**parentp
,
1257 hammer2_key_t lbase
, hammer2_tid_t mtid
, int *errorp
)
1259 hammer2_chain_t
*chain
;
1260 hammer2_key_t key_dummy
;
1262 chain
= hammer2_chain_lookup(parentp
, &key_dummy
,
1265 HAMMER2_LOOKUP_NODATA
);
1267 if (chain
->bref
.type
== HAMMER2_BREF_TYPE_INODE
) {
1268 hammer2_inode_data_t
*wipdata
;
1271 *errorp
= hammer2_chain_modify_ip(ip
, chain
,
1275 wipdata
= &chain
->data
->ipdata
;
1276 KKASSERT(wipdata
->meta
.op_flags
&
1277 HAMMER2_OPFLAG_DIRECTDATA
);
1278 bzero(wipdata
->u
.data
, HAMMER2_EMBEDDED_BYTES
);
1279 ++hammer2_iod_file_wembed
;
1282 /* chain->error ok for deletion */
1283 hammer2_chain_delete(*parentp
, chain
,
1284 mtid
, HAMMER2_DELETE_PERMANENT
);
1285 ++hammer2_iod_file_wzero
;
1287 atomic_set_int(&ip
->flags
, HAMMER2_INODE_DIRTYDATA
);
1288 hammer2_chain_unlock(chain
);
1289 hammer2_chain_drop(chain
);
1291 ++hammer2_iod_file_wzero
;
1298 * Function to write the data as it is, without performing any sort of
1299 * compression. This function is used in path without compression and
1300 * default zero-checking path.
1304 hammer2_write_bp(hammer2_chain_t
*chain
, char *data
, int ioflag
,
1306 hammer2_tid_t mtid
, int *errorp
, int check_algo
)
1308 hammer2_inode_data_t
*wipdata
;
1313 error
= 0; /* XXX TODO below */
1315 KKASSERT(chain
->flags
& HAMMER2_CHAIN_MODIFIED
);
1317 switch(chain
->bref
.type
) {
1318 case HAMMER2_BREF_TYPE_INODE
:
1319 wipdata
= &chain
->data
->ipdata
;
1320 KKASSERT(wipdata
->meta
.op_flags
& HAMMER2_OPFLAG_DIRECTDATA
);
1321 bcopy(data
, wipdata
->u
.data
, HAMMER2_EMBEDDED_BYTES
);
1323 ++hammer2_iod_file_wembed
;
1325 case HAMMER2_BREF_TYPE_DATA
:
1326 error
= hammer2_io_newnz(chain
->hmp
,
1328 chain
->bref
.data_off
,
1329 chain
->bytes
, &dio
);
1331 hammer2_io_bqrelse(&dio
);
1332 kprintf("hammer2: WRITE PATH: "
1333 "dbp bread error\n");
1336 bdata
= hammer2_io_data(dio
, chain
->bref
.data_off
);
1338 chain
->bref
.methods
= HAMMER2_ENC_COMP(HAMMER2_COMP_NONE
) +
1339 HAMMER2_ENC_CHECK(check_algo
);
1340 bcopy(data
, bdata
, chain
->bytes
);
1343 * The flush code doesn't calculate check codes for
1344 * file data (doing so can result in excessive I/O),
1347 hammer2_chain_setcheck(chain
, bdata
);
1350 * Device buffer is now valid, chain is no longer in
1351 * the initial state.
1353 * (No blockref table worries with file data)
1355 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_INITIAL
);
1356 hammer2_dedup_record(chain
, dio
, bdata
);
1358 if (ioflag
& IO_SYNC
) {
1360 * Synchronous I/O requested.
1362 hammer2_io_bwrite(&dio
);
1364 } else if ((ioflag & IO_DIRECT) &&
1365 loff + n == pblksize) {
1366 hammer2_io_bdwrite(&dio);
1368 } else if (ioflag
& IO_ASYNC
) {
1369 hammer2_io_bawrite(&dio
);
1371 hammer2_io_bdwrite(&dio
);
1375 panic("hammer2_write_bp: bad chain type %d\n",
1385 * LIVE DEDUP HEURISTICS
1387 * Record media and crc information for possible dedup operation. Note
1388 * that the dedup mask bits must also be set in the related DIO for a dedup
1389 * to be fully validated (which is handled in the freemap allocation code).
1391 * WARNING! This code is SMP safe but the heuristic allows SMP collisions.
1392 * All fields must be loaded into locals and validated.
1394 * WARNING! Should only be used for file data and directory entries,
1395 * hammer2_chain_modify() only checks for the dedup case on data
1396 * chains. Also, dedup data can only be recorded for committed
1397 * chains (so NOT strategy writes which can undergo further
1398 * modification after the fact!).
1401 hammer2_dedup_record(hammer2_chain_t
*chain
, hammer2_io_t
*dio
, char *data
)
1404 hammer2_dedup_t
*dedup
;
1412 * We can only record a dedup if we have media data to test against.
1413 * If dedup is not enabled, return early, which allows a chain to
1414 * remain marked MODIFIED (which might have benefits in special
1415 * situations, though typically it does not).
1417 if (hammer2_dedup_enable
== 0)
1427 switch(HAMMER2_DEC_CHECK(chain
->bref
.methods
)) {
1428 case HAMMER2_CHECK_ISCSI32
:
1430 * XXX use the built-in crc (the dedup lookup sequencing
1431 * needs to be fixed so the check code is already present
1432 * when dedup_lookup is called)
1435 crc
= (uint64_t)(uint32_t)chain
->bref
.check
.iscsi32
.value
;
1437 crc
= XXH64(data
, chain
->bytes
, XXH_HAMMER2_SEED
);
1439 case HAMMER2_CHECK_XXHASH64
:
1440 crc
= chain
->bref
.check
.xxhash64
.value
;
1442 case HAMMER2_CHECK_SHA192
:
1444 * XXX use the built-in crc (the dedup lookup sequencing
1445 * needs to be fixed so the check code is already present
1446 * when dedup_lookup is called)
1449 crc
= ((uint64_t *)chain
->bref
.check
.sha192
.data
)[0] ^
1450 ((uint64_t *)chain
->bref
.check
.sha192
.data
)[1] ^
1451 ((uint64_t *)chain
->bref
.check
.sha192
.data
)[2];
1453 crc
= XXH64(data
, chain
->bytes
, XXH_HAMMER2_SEED
);
1457 * Cannot dedup without a check code
1459 * NOTE: In particular, CHECK_NONE allows a sector to be
1460 * overwritten without copy-on-write, recording
1461 * a dedup block for a CHECK_NONE object would be
1467 atomic_set_int(&chain
->flags
, HAMMER2_CHAIN_DEDUPABLE
);
1469 dedup
= &hmp
->heur_dedup
[crc
& (HAMMER2_DEDUP_HEUR_MASK
& ~3)];
1470 for (i
= 0; i
< 4; ++i
) {
1471 if (dedup
[i
].data_crc
== crc
) {
1475 dticks
= (int)(dedup
[i
].ticks
- dedup
[best
].ticks
);
1476 if (dticks
< 0 || dticks
> hz
* 60 * 30)
1480 if (hammer2_debug
& 0x40000) {
1481 kprintf("REC %04x %016jx %016jx\n",
1482 (int)(dedup
- hmp
->heur_dedup
),
1484 chain
->bref
.data_off
);
1486 dedup
->ticks
= ticks
;
1487 dedup
->data_off
= chain
->bref
.data_off
;
1488 dedup
->data_crc
= crc
;
1491 * Set the valid bits for the dedup only after we know the data
1492 * buffer has been updated. The alloc bits were set (and the valid
1493 * bits cleared) when the media was allocated.
1495 * This is done in two stages becuase the bulkfree code can race
1496 * the gap between allocation and data population. Both masks must
1497 * be set before a bcmp/dedup operation is able to use the block.
1499 mask
= hammer2_dedup_mask(dio
, chain
->bref
.data_off
, chain
->bytes
);
1500 atomic_set_64(&dio
->dedup_valid
, mask
);
1504 * XXX removed. MODIFIED is an integral part of the flush code,
1505 * lets not just clear it
1508 * Once we record the dedup the chain must be marked clean to
1509 * prevent reuse of the underlying block. Remember that this
1510 * write occurs when the buffer cache is flushed (i.e. on sync(),
1511 * fsync(), filesystem periodic sync, or when the kernel needs to
1512 * flush a buffer), and not whenever the user write()s.
1514 if (chain
->flags
& HAMMER2_CHAIN_MODIFIED
) {
1515 atomic_clear_int(&chain
->flags
, HAMMER2_CHAIN_MODIFIED
);
1516 atomic_add_long(&hammer2_count_modified_chains
, -1);
1518 hammer2_pfs_memory_wakeup(chain
->pmp
);
1525 hammer2_dedup_lookup(hammer2_dev_t
*hmp
, char **datap
, int pblksize
)
1527 hammer2_dedup_t
*dedup
;
1536 if (hammer2_dedup_enable
== 0)
1543 * XXX use the built-in crc (the dedup lookup sequencing
1544 * needs to be fixed so the check code is already present
1545 * when dedup_lookup is called)
1547 crc
= XXH64(data
, pblksize
, XXH_HAMMER2_SEED
);
1548 dedup
= &hmp
->heur_dedup
[crc
& (HAMMER2_DEDUP_HEUR_MASK
& ~3)];
1550 if (hammer2_debug
& 0x40000) {
1551 kprintf("LOC %04x/4 %016jx\n",
1552 (int)(dedup
- hmp
->heur_dedup
),
1556 for (i
= 0; i
< 4; ++i
) {
1557 off
= dedup
[i
].data_off
;
1559 if (dedup
[i
].data_crc
!= crc
)
1561 if ((1 << (int)(off
& HAMMER2_OFF_MASK_RADIX
)) != pblksize
)
1563 dio
= hammer2_io_getquick(hmp
, off
, pblksize
);
1565 dtmp
= hammer2_io_data(dio
, off
),
1566 mask
= hammer2_dedup_mask(dio
, off
, pblksize
);
1567 if ((dio
->dedup_alloc
& mask
) == mask
&&
1568 (dio
->dedup_valid
& mask
) == mask
&&
1569 bcmp(data
, dtmp
, pblksize
) == 0) {
1570 if (hammer2_debug
& 0x40000) {
1571 kprintf("DEDUP SUCCESS %016jx\n",
1574 hammer2_io_putblk(&dio
);
1576 dedup
[i
].ticks
= ticks
; /* update use */
1577 atomic_add_long(&hammer2_iod_file_wdedup
,
1580 return off
; /* RETURN */
1582 hammer2_io_putblk(&dio
);
1589 * Poof. Races are ok, if someone gets in and reuses a dedup offset
1590 * before or while we are clearing it they will also recover the freemap
1591 * entry (set it to fully allocated), so a bulkfree race can only set it
1592 * to a possibly-free state.
1594 * XXX ok, well, not really sure races are ok but going to run with it
1598 hammer2_dedup_clear(hammer2_dev_t
*hmp
)
1602 for (i
= 0; i
< HAMMER2_DEDUP_HEUR_SIZE
; ++i
) {
1603 hmp
->heur_dedup
[i
].data_off
= 0;
1604 hmp
->heur_dedup
[i
].ticks
= ticks
- 1;