4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2015, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 * Lustre Lite I/O page cache routines shared by different kernel revs
41 #include <linux/kernel.h>
43 #include <linux/string.h>
44 #include <linux/stat.h>
45 #include <linux/errno.h>
46 #include <linux/unistd.h>
47 #include <linux/writeback.h>
48 #include <linux/uaccess.h>
51 #include <linux/pagemap.h>
52 /* current_is_kswapd() */
53 #include <linux/swap.h>
55 #define DEBUG_SUBSYSTEM S_LLITE
57 #include "../include/lustre_lite.h"
58 #include "../include/obd_cksum.h"
59 #include "llite_internal.h"
60 #include "../include/linux/lustre_compat25.h"
63 * Finalizes cl-data before exiting typical address_space operation. Dual to
66 static void ll_cl_fini(struct ll_cl_context
*lcc
)
68 struct lu_env
*env
= lcc
->lcc_env
;
69 struct cl_io
*io
= lcc
->lcc_io
;
70 struct cl_page
*page
= lcc
->lcc_page
;
72 LASSERT(lcc
->lcc_cookie
== current
);
76 lu_ref_del(&page
->cp_reference
, "cl_io", io
);
77 cl_page_put(env
, page
);
80 cl_env_put(env
, &lcc
->lcc_refcheck
);
84 * Initializes common cl-data at the typical address_space operation entry
87 static struct ll_cl_context
*ll_cl_init(struct file
*file
,
88 struct page
*vmpage
, int create
)
90 struct ll_cl_context
*lcc
;
93 struct cl_object
*clob
;
99 clob
= ll_i2info(vmpage
->mapping
->host
)->lli_clob
;
100 LASSERT(clob
!= NULL
);
102 env
= cl_env_get(&refcheck
);
104 return ERR_CAST(env
);
106 lcc
= &vvp_env_info(env
)->vti_io_ctx
;
107 memset(lcc
, 0, sizeof(*lcc
));
109 lcc
->lcc_refcheck
= refcheck
;
110 lcc
->lcc_cookie
= current
;
112 cio
= ccc_env_io(env
);
113 io
= cio
->cui_cl
.cis_io
;
114 if (io
== NULL
&& create
) {
115 struct inode
*inode
= vmpage
->mapping
->host
;
118 if (inode_trylock(inode
)) {
119 inode_unlock((inode
));
121 /* this is too bad. Someone is trying to write the
122 * page w/o holding inode mutex. This means we can
123 * add dirty pages into cache during truncate */
124 CERROR("Proc %s is dirtying page w/o inode lock, this will break truncate\n",
128 return ERR_PTR(-EIO
);
132 * Loop-back driver calls ->prepare_write().
133 * methods directly, bypassing file system ->write() operation,
134 * so cl_io has to be created here.
136 io
= ccc_env_thread_io(env
);
137 ll_io_init(io
, file
, 1);
139 /* No lock at all for this kind of IO - we can't do it because
140 * we have held page lock, it would cause deadlock.
141 * XXX: This causes poor performance to loop device - One page
143 * In order to get better performance, users should use
144 * lloop driver instead.
146 io
->ci_lockreq
= CILR_NEVER
;
148 pos
= vmpage
->index
<< PAGE_CACHE_SHIFT
;
150 /* Create a temp IO to serve write. */
151 result
= cl_io_rw_init(env
, io
, CIT_WRITE
, pos
, PAGE_CACHE_SIZE
);
153 cio
->cui_fd
= LUSTRE_FPRIVATE(file
);
154 cio
->cui_iter
= NULL
;
155 result
= cl_io_iter_init(env
, io
);
157 result
= cl_io_lock(env
, io
);
159 result
= cl_io_start(env
, io
);
162 result
= io
->ci_result
;
169 struct cl_page
*page
;
172 LASSERT(io
->ci_state
== CIS_IO_GOING
);
173 LASSERT(cio
->cui_fd
== LUSTRE_FPRIVATE(file
));
174 page
= cl_page_find(env
, clob
, vmpage
->index
, vmpage
,
177 lcc
->lcc_page
= page
;
178 lu_ref_add(&page
->cp_reference
, "cl_io", io
);
181 result
= PTR_ERR(page
);
185 lcc
= ERR_PTR(result
);
188 CDEBUG(D_VFSTRACE
, "%lu@"DFID
" -> %d %p %p\n",
189 vmpage
->index
, PFID(lu_object_fid(&clob
->co_lu
)), result
,
194 static struct ll_cl_context
*ll_cl_get(void)
196 struct ll_cl_context
*lcc
;
200 env
= cl_env_get(&refcheck
);
201 LASSERT(!IS_ERR(env
));
202 lcc
= &vvp_env_info(env
)->vti_io_ctx
;
203 LASSERT(env
== lcc
->lcc_env
);
204 LASSERT(current
== lcc
->lcc_cookie
);
205 cl_env_put(env
, &refcheck
);
207 /* env has got in ll_cl_init, so it is still usable. */
212 * ->prepare_write() address space operation called by generic_file_write()
213 * for every page during write.
215 int ll_prepare_write(struct file
*file
, struct page
*vmpage
, unsigned from
,
218 struct ll_cl_context
*lcc
;
221 lcc
= ll_cl_init(file
, vmpage
, 1);
223 struct lu_env
*env
= lcc
->lcc_env
;
224 struct cl_io
*io
= lcc
->lcc_io
;
225 struct cl_page
*page
= lcc
->lcc_page
;
227 cl_page_assume(env
, io
, page
);
229 result
= cl_io_prepare_write(env
, io
, page
, from
, to
);
232 * Add a reference, so that page is not evicted from
233 * the cache until ->commit_write() is called.
236 lu_ref_add(&page
->cp_reference
, "prepare_write",
239 cl_page_unassume(env
, io
, page
);
242 /* returning 0 in prepare assumes commit must be called
245 result
= PTR_ERR(lcc
);
250 int ll_commit_write(struct file
*file
, struct page
*vmpage
, unsigned from
,
253 struct ll_cl_context
*lcc
;
256 struct cl_page
*page
;
261 page
= lcc
->lcc_page
;
264 LASSERT(cl_page_is_owned(page
, io
));
266 if (from
!= to
) /* handle short write case. */
267 result
= cl_io_commit_write(env
, io
, page
, from
, to
);
268 if (cl_page_is_owned(page
, io
))
269 cl_page_unassume(env
, io
, page
);
272 * Release reference acquired by ll_prepare_write().
274 lu_ref_del(&page
->cp_reference
, "prepare_write", current
);
275 cl_page_put(env
, page
);
280 static void ll_ra_stats_inc_sbi(struct ll_sb_info
*sbi
, enum ra_stat which
);
283 * Get readahead pages from the filesystem readahead pool of the client for a
286 * /param sbi superblock for filesystem readahead state ll_ra_info
287 * /param ria per-thread readahead state
288 * /param pages number of pages requested for readahead for the thread.
290 * WARNING: This algorithm is used to reduce contention on sbi->ll_lock.
291 * It should work well if the ra_max_pages is much greater than the single
292 * file's read-ahead window, and not too many threads contending for
293 * these readahead pages.
295 * TODO: There may be a 'global sync problem' if many threads are trying
296 * to get an ra budget that is larger than the remaining readahead pages
297 * and reach here at exactly the same time. They will compute /a ret to
298 * consume the remaining pages, but will fail at atomic_add_return() and
299 * get a zero ra window, although there is still ra space remaining. - Jay */
301 static unsigned long ll_ra_count_get(struct ll_sb_info
*sbi
,
302 struct ra_io_arg
*ria
,
305 struct ll_ra_info
*ra
= &sbi
->ll_ra_info
;
308 /* If read-ahead pages left are less than 1M, do not do read-ahead,
309 * otherwise it will form small read RPC(< 1M), which hurt server
310 * performance a lot. */
311 ret
= min(ra
->ra_max_pages
- atomic_read(&ra
->ra_cur_pages
), pages
);
312 if (ret
< 0 || ret
< min_t(long, PTLRPC_MAX_BRW_PAGES
, pages
)) {
317 /* If the non-strided (ria_pages == 0) readahead window
318 * (ria_start + ret) has grown across an RPC boundary, then trim
319 * readahead size by the amount beyond the RPC so it ends on an
320 * RPC boundary. If the readahead window is already ending on
321 * an RPC boundary (beyond_rpc == 0), or smaller than a full
322 * RPC (beyond_rpc < ret) the readahead size is unchanged.
323 * The (beyond_rpc != 0) check is skipped since the conditional
324 * branch is more expensive than subtracting zero from the result.
326 * Strided read is left unaligned to avoid small fragments beyond
327 * the RPC boundary from needing an extra read RPC. */
328 if (ria
->ria_pages
== 0) {
329 long beyond_rpc
= (ria
->ria_start
+ ret
) % PTLRPC_MAX_BRW_PAGES
;
331 if (/* beyond_rpc != 0 && */ beyond_rpc
< ret
)
335 if (atomic_add_return(ret
, &ra
->ra_cur_pages
) > ra
->ra_max_pages
) {
336 atomic_sub(ret
, &ra
->ra_cur_pages
);
344 void ll_ra_count_put(struct ll_sb_info
*sbi
, unsigned long len
)
346 struct ll_ra_info
*ra
= &sbi
->ll_ra_info
;
348 atomic_sub(len
, &ra
->ra_cur_pages
);
351 static void ll_ra_stats_inc_sbi(struct ll_sb_info
*sbi
, enum ra_stat which
)
353 LASSERTF(which
>= 0 && which
< _NR_RA_STAT
, "which: %u\n", which
);
354 lprocfs_counter_incr(sbi
->ll_ra_stats
, which
);
357 void ll_ra_stats_inc(struct address_space
*mapping
, enum ra_stat which
)
359 struct ll_sb_info
*sbi
= ll_i2sbi(mapping
->host
);
361 ll_ra_stats_inc_sbi(sbi
, which
);
364 #define RAS_CDEBUG(ras) \
366 "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu r %lu ri %lu" \
367 "csr %lu sf %lu sp %lu sl %lu \n", \
368 ras->ras_last_readpage, ras->ras_consecutive_requests, \
369 ras->ras_consecutive_pages, ras->ras_window_start, \
370 ras->ras_window_len, ras->ras_next_readahead, \
371 ras->ras_requests, ras->ras_request_index, \
372 ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
373 ras->ras_stride_pages, ras->ras_stride_length)
375 static int index_in_window(unsigned long index
, unsigned long point
,
376 unsigned long before
, unsigned long after
)
378 unsigned long start
= point
- before
, end
= point
+ after
;
385 return start
<= index
&& index
<= end
;
388 static struct ll_readahead_state
*ll_ras_get(struct file
*f
)
390 struct ll_file_data
*fd
;
392 fd
= LUSTRE_FPRIVATE(f
);
396 void ll_ra_read_in(struct file
*f
, struct ll_ra_read
*rar
)
398 struct ll_readahead_state
*ras
;
402 spin_lock(&ras
->ras_lock
);
404 ras
->ras_request_index
= 0;
405 ras
->ras_consecutive_requests
++;
406 rar
->lrr_reader
= current
;
408 list_add(&rar
->lrr_linkage
, &ras
->ras_read_beads
);
409 spin_unlock(&ras
->ras_lock
);
412 void ll_ra_read_ex(struct file
*f
, struct ll_ra_read
*rar
)
414 struct ll_readahead_state
*ras
;
418 spin_lock(&ras
->ras_lock
);
419 list_del_init(&rar
->lrr_linkage
);
420 spin_unlock(&ras
->ras_lock
);
423 static int cl_read_ahead_page(const struct lu_env
*env
, struct cl_io
*io
,
424 struct cl_page_list
*queue
, struct cl_page
*page
,
431 cl_page_assume(env
, io
, page
);
432 lu_ref_add(&page
->cp_reference
, "ra", current
);
433 cp
= cl2ccc_page(cl_page_at(page
, &vvp_device_type
));
434 if (!cp
->cpg_defer_uptodate
&& !PageUptodate(vmpage
)) {
435 rc
= cl_page_is_under_lock(env
, io
, page
);
437 cp
->cpg_defer_uptodate
= 1;
439 cl_page_list_add(queue
, page
);
442 cl_page_delete(env
, page
);
446 /* skip completed pages */
447 cl_page_unassume(env
, io
, page
);
449 lu_ref_del(&page
->cp_reference
, "ra", current
);
450 cl_page_put(env
, page
);
455 * Initiates read-ahead of a page with given index.
457 * \retval +ve: page was added to \a queue.
459 * \retval -ENOLCK: there is no extent lock for this part of a file, stop
462 * \retval -ve, 0: page wasn't added to \a queue for other reason.
464 static int ll_read_ahead_page(const struct lu_env
*env
, struct cl_io
*io
,
465 struct cl_page_list
*queue
,
466 pgoff_t index
, struct address_space
*mapping
)
469 struct cl_object
*clob
= ll_i2info(mapping
->host
)->lli_clob
;
470 struct cl_page
*page
;
471 enum ra_stat which
= _NR_RA_STAT
; /* keep gcc happy */
473 const char *msg
= NULL
;
475 vmpage
= grab_cache_page_nowait(mapping
, index
);
476 if (vmpage
!= NULL
) {
477 /* Check if vmpage was truncated or reclaimed */
478 if (vmpage
->mapping
== mapping
) {
479 page
= cl_page_find(env
, clob
, vmpage
->index
,
480 vmpage
, CPT_CACHEABLE
);
482 rc
= cl_read_ahead_page(env
, io
, queue
,
485 which
= RA_STAT_FAILED_MATCH
;
486 msg
= "lock match failed";
489 which
= RA_STAT_FAILED_GRAB_PAGE
;
490 msg
= "cl_page_find failed";
493 which
= RA_STAT_WRONG_GRAB_PAGE
;
494 msg
= "g_c_p_n returned invalid page";
498 page_cache_release(vmpage
);
500 which
= RA_STAT_FAILED_GRAB_PAGE
;
501 msg
= "g_c_p_n failed";
504 ll_ra_stats_inc(mapping
, which
);
505 CDEBUG(D_READA
, "%s\n", msg
);
510 #define RIA_DEBUG(ria) \
511 CDEBUG(D_READA, "rs %lu re %lu ro %lu rl %lu rp %lu\n", \
512 ria->ria_start, ria->ria_end, ria->ria_stoff, ria->ria_length,\
515 /* Limit this to the blocksize instead of PTLRPC_BRW_MAX_SIZE, since we don't
516 * know what the actual RPC size is. If this needs to change, it makes more
517 * sense to tune the i_blkbits value for the file based on the OSTs it is
518 * striped over, rather than having a constant value for all files here. */
520 /* RAS_INCREASE_STEP should be (1UL << (inode->i_blkbits - PAGE_CACHE_SHIFT)).
521 * Temporarily set RAS_INCREASE_STEP to 1MB. After 4MB RPC is enabled
522 * by default, this should be adjusted corresponding with max_read_ahead_mb
523 * and max_read_ahead_per_file_mb otherwise the readahead budget can be used
524 * up quickly which will affect read performance significantly. See LU-2816 */
525 #define RAS_INCREASE_STEP(inode) (ONE_MB_BRW_SIZE >> PAGE_CACHE_SHIFT)
527 static inline int stride_io_mode(struct ll_readahead_state
*ras
)
529 return ras
->ras_consecutive_stride_requests
> 1;
532 /* The function calculates how much pages will be read in
533 * [off, off + length], in such stride IO area,
534 * stride_offset = st_off, stride_length = st_len,
535 * stride_pages = st_pgs
537 * |------------------|*****|------------------|*****|------------|*****|....
540 * |----- st_len -----|
542 * How many pages it should read in such pattern
543 * |-------------------------------------------------------------|
545 * |<------ length ------->|
547 * = |<----->| + |-------------------------------------| + |---|
548 * start_left st_pgs * i end_left
551 stride_pg_count(pgoff_t st_off
, unsigned long st_len
, unsigned long st_pgs
,
552 unsigned long off
, unsigned long length
)
554 __u64 start
= off
> st_off
? off
- st_off
: 0;
555 __u64 end
= off
+ length
> st_off
? off
+ length
- st_off
: 0;
556 unsigned long start_left
= 0;
557 unsigned long end_left
= 0;
558 unsigned long pg_count
;
560 if (st_len
== 0 || length
== 0 || end
== 0)
563 start_left
= do_div(start
, st_len
);
564 if (start_left
< st_pgs
)
565 start_left
= st_pgs
- start_left
;
569 end_left
= do_div(end
, st_len
);
570 if (end_left
> st_pgs
)
573 CDEBUG(D_READA
, "start %llu, end %llu start_left %lu end_left %lu \n",
574 start
, end
, start_left
, end_left
);
577 pg_count
= end_left
- (st_pgs
- start_left
);
579 pg_count
= start_left
+ st_pgs
* (end
- start
- 1) + end_left
;
581 CDEBUG(D_READA
, "st_off %lu, st_len %lu st_pgs %lu off %lu length %lu pgcount %lu\n",
582 st_off
, st_len
, st_pgs
, off
, length
, pg_count
);
587 static int ria_page_count(struct ra_io_arg
*ria
)
589 __u64 length
= ria
->ria_end
>= ria
->ria_start
?
590 ria
->ria_end
- ria
->ria_start
+ 1 : 0;
592 return stride_pg_count(ria
->ria_stoff
, ria
->ria_length
,
593 ria
->ria_pages
, ria
->ria_start
,
597 /*Check whether the index is in the defined ra-window */
598 static int ras_inside_ra_window(unsigned long idx
, struct ra_io_arg
*ria
)
600 /* If ria_length == ria_pages, it means non-stride I/O mode,
601 * idx should always inside read-ahead window in this case
602 * For stride I/O mode, just check whether the idx is inside
604 return ria
->ria_length
== 0 || ria
->ria_length
== ria
->ria_pages
||
605 (idx
>= ria
->ria_stoff
&& (idx
- ria
->ria_stoff
) %
606 ria
->ria_length
< ria
->ria_pages
);
609 static int ll_read_ahead_pages(const struct lu_env
*env
,
610 struct cl_io
*io
, struct cl_page_list
*queue
,
611 struct ra_io_arg
*ria
,
612 unsigned long *reserved_pages
,
613 struct address_space
*mapping
,
614 unsigned long *ra_end
)
616 int rc
, count
= 0, stride_ria
;
617 unsigned long page_idx
;
619 LASSERT(ria
!= NULL
);
622 stride_ria
= ria
->ria_length
> ria
->ria_pages
&& ria
->ria_pages
> 0;
623 for (page_idx
= ria
->ria_start
; page_idx
<= ria
->ria_end
&&
624 *reserved_pages
> 0; page_idx
++) {
625 if (ras_inside_ra_window(page_idx
, ria
)) {
626 /* If the page is inside the read-ahead window*/
627 rc
= ll_read_ahead_page(env
, io
, queue
,
632 } else if (rc
== -ENOLCK
)
634 } else if (stride_ria
) {
635 /* If it is not in the read-ahead window, and it is
636 * read-ahead mode, then check whether it should skip
639 /* FIXME: This assertion only is valid when it is for
640 * forward read-ahead, it will be fixed when backward
641 * read-ahead is implemented */
642 LASSERTF(page_idx
> ria
->ria_stoff
, "Invalid page_idx %lu rs %lu re %lu ro %lu rl %lu rp %lu\n",
644 ria
->ria_start
, ria
->ria_end
, ria
->ria_stoff
,
645 ria
->ria_length
, ria
->ria_pages
);
646 offset
= page_idx
- ria
->ria_stoff
;
647 offset
= offset
% (ria
->ria_length
);
648 if (offset
> ria
->ria_pages
) {
649 page_idx
+= ria
->ria_length
- offset
;
650 CDEBUG(D_READA
, "i %lu skip %lu \n", page_idx
,
651 ria
->ria_length
- offset
);
660 int ll_readahead(const struct lu_env
*env
, struct cl_io
*io
,
661 struct ll_readahead_state
*ras
, struct address_space
*mapping
,
662 struct cl_page_list
*queue
, int flags
)
664 struct vvp_io
*vio
= vvp_env_io(env
);
665 struct vvp_thread_info
*vti
= vvp_env_info(env
);
666 struct cl_attr
*attr
= ccc_env_thread_attr(env
);
667 unsigned long start
= 0, end
= 0, reserved
;
668 unsigned long ra_end
, len
;
670 struct ll_ra_read
*bead
;
671 struct ra_io_arg
*ria
= &vti
->vti_ria
;
672 struct ll_inode_info
*lli
;
673 struct cl_object
*clob
;
677 inode
= mapping
->host
;
678 lli
= ll_i2info(inode
);
679 clob
= lli
->lli_clob
;
681 memset(ria
, 0, sizeof(*ria
));
683 cl_object_attr_lock(clob
);
684 ret
= cl_object_attr_get(env
, clob
, attr
);
685 cl_object_attr_unlock(clob
);
691 ll_ra_stats_inc(mapping
, RA_STAT_ZERO_LEN
);
695 spin_lock(&ras
->ras_lock
);
696 if (vio
->cui_ra_window_set
)
697 bead
= &vio
->cui_bead
;
701 /* Enlarge the RA window to encompass the full read */
702 if (bead
!= NULL
&& ras
->ras_window_start
+ ras
->ras_window_len
<
703 bead
->lrr_start
+ bead
->lrr_count
) {
704 ras
->ras_window_len
= bead
->lrr_start
+ bead
->lrr_count
-
705 ras
->ras_window_start
;
707 /* Reserve a part of the read-ahead window that we'll be issuing */
708 if (ras
->ras_window_len
) {
709 start
= ras
->ras_next_readahead
;
710 end
= ras
->ras_window_start
+ ras
->ras_window_len
- 1;
713 unsigned long rpc_boundary
;
715 * Align RA window to an optimal boundary.
717 * XXX This would be better to align to cl_max_pages_per_rpc
718 * instead of PTLRPC_MAX_BRW_PAGES, because the RPC size may
719 * be aligned to the RAID stripe size in the future and that
720 * is more important than the RPC size.
722 /* Note: we only trim the RPC, instead of extending the RPC
723 * to the boundary, so to avoid reading too much pages during
725 rpc_boundary
= (end
+ 1) & (~(PTLRPC_MAX_BRW_PAGES
- 1));
726 if (rpc_boundary
> 0)
729 if (rpc_boundary
> start
)
732 /* Truncate RA window to end of file */
733 end
= min(end
, (unsigned long)((kms
- 1) >> PAGE_CACHE_SHIFT
));
735 ras
->ras_next_readahead
= max(end
, end
+ 1);
738 ria
->ria_start
= start
;
740 /* If stride I/O mode is detected, get stride window*/
741 if (stride_io_mode(ras
)) {
742 ria
->ria_stoff
= ras
->ras_stride_offset
;
743 ria
->ria_length
= ras
->ras_stride_length
;
744 ria
->ria_pages
= ras
->ras_stride_pages
;
746 spin_unlock(&ras
->ras_lock
);
749 ll_ra_stats_inc(mapping
, RA_STAT_ZERO_WINDOW
);
752 len
= ria_page_count(ria
);
756 reserved
= ll_ra_count_get(ll_i2sbi(inode
), ria
, len
);
758 ll_ra_stats_inc(mapping
, RA_STAT_MAX_IN_FLIGHT
);
760 CDEBUG(D_READA
, "reserved page %lu ra_cur %d ra_max %lu\n", reserved
,
761 atomic_read(&ll_i2sbi(inode
)->ll_ra_info
.ra_cur_pages
),
762 ll_i2sbi(inode
)->ll_ra_info
.ra_max_pages
);
764 ret
= ll_read_ahead_pages(env
, io
, queue
,
765 ria
, &reserved
, mapping
, &ra_end
);
767 LASSERTF(reserved
>= 0, "reserved %lu\n", reserved
);
769 ll_ra_count_put(ll_i2sbi(inode
), reserved
);
771 if (ra_end
== end
+ 1 && ra_end
== (kms
>> PAGE_CACHE_SHIFT
))
772 ll_ra_stats_inc(mapping
, RA_STAT_EOF
);
774 /* if we didn't get to the end of the region we reserved from
775 * the ras we need to go back and update the ras so that the
776 * next read-ahead tries from where we left off. we only do so
777 * if the region we failed to issue read-ahead on is still ahead
778 * of the app and behind the next index to start read-ahead from */
779 CDEBUG(D_READA
, "ra_end %lu end %lu stride end %lu \n",
780 ra_end
, end
, ria
->ria_end
);
782 if (ra_end
!= end
+ 1) {
783 spin_lock(&ras
->ras_lock
);
784 if (ra_end
< ras
->ras_next_readahead
&&
785 index_in_window(ra_end
, ras
->ras_window_start
, 0,
786 ras
->ras_window_len
)) {
787 ras
->ras_next_readahead
= ra_end
;
790 spin_unlock(&ras
->ras_lock
);
796 static void ras_set_start(struct inode
*inode
, struct ll_readahead_state
*ras
,
799 ras
->ras_window_start
= index
& (~(RAS_INCREASE_STEP(inode
) - 1));
802 /* called with the ras_lock held or from places where it doesn't matter */
803 static void ras_reset(struct inode
*inode
, struct ll_readahead_state
*ras
,
806 ras
->ras_last_readpage
= index
;
807 ras
->ras_consecutive_requests
= 0;
808 ras
->ras_consecutive_pages
= 0;
809 ras
->ras_window_len
= 0;
810 ras_set_start(inode
, ras
, index
);
811 ras
->ras_next_readahead
= max(ras
->ras_window_start
, index
);
816 /* called with the ras_lock held or from places where it doesn't matter */
817 static void ras_stride_reset(struct ll_readahead_state
*ras
)
819 ras
->ras_consecutive_stride_requests
= 0;
820 ras
->ras_stride_length
= 0;
821 ras
->ras_stride_pages
= 0;
825 void ll_readahead_init(struct inode
*inode
, struct ll_readahead_state
*ras
)
827 spin_lock_init(&ras
->ras_lock
);
828 ras_reset(inode
, ras
, 0);
829 ras
->ras_requests
= 0;
830 INIT_LIST_HEAD(&ras
->ras_read_beads
);
834 * Check whether the read request is in the stride window.
835 * If it is in the stride window, return 1, otherwise return 0.
837 static int index_in_stride_window(struct ll_readahead_state
*ras
,
840 unsigned long stride_gap
;
842 if (ras
->ras_stride_length
== 0 || ras
->ras_stride_pages
== 0 ||
843 ras
->ras_stride_pages
== ras
->ras_stride_length
)
846 stride_gap
= index
- ras
->ras_last_readpage
- 1;
848 /* If it is contiguous read */
850 return ras
->ras_consecutive_pages
+ 1 <= ras
->ras_stride_pages
;
852 /* Otherwise check the stride by itself */
853 return (ras
->ras_stride_length
- ras
->ras_stride_pages
) == stride_gap
&&
854 ras
->ras_consecutive_pages
== ras
->ras_stride_pages
;
857 static void ras_update_stride_detector(struct ll_readahead_state
*ras
,
860 unsigned long stride_gap
= index
- ras
->ras_last_readpage
- 1;
862 if (!stride_io_mode(ras
) && (stride_gap
!= 0 ||
863 ras
->ras_consecutive_stride_requests
== 0)) {
864 ras
->ras_stride_pages
= ras
->ras_consecutive_pages
;
865 ras
->ras_stride_length
= stride_gap
+ras
->ras_consecutive_pages
;
867 LASSERT(ras
->ras_request_index
== 0);
868 LASSERT(ras
->ras_consecutive_stride_requests
== 0);
870 if (index
<= ras
->ras_last_readpage
) {
871 /*Reset stride window for forward read*/
872 ras_stride_reset(ras
);
876 ras
->ras_stride_pages
= ras
->ras_consecutive_pages
;
877 ras
->ras_stride_length
= stride_gap
+ras
->ras_consecutive_pages
;
883 /* Stride Read-ahead window will be increased inc_len according to
884 * stride I/O pattern */
885 static void ras_stride_increase_window(struct ll_readahead_state
*ras
,
886 struct ll_ra_info
*ra
,
887 unsigned long inc_len
)
889 unsigned long left
, step
, window_len
;
890 unsigned long stride_len
;
892 LASSERT(ras
->ras_stride_length
> 0);
893 LASSERTF(ras
->ras_window_start
+ ras
->ras_window_len
894 >= ras
->ras_stride_offset
, "window_start %lu, window_len %lu stride_offset %lu\n",
895 ras
->ras_window_start
,
896 ras
->ras_window_len
, ras
->ras_stride_offset
);
898 stride_len
= ras
->ras_window_start
+ ras
->ras_window_len
-
899 ras
->ras_stride_offset
;
901 left
= stride_len
% ras
->ras_stride_length
;
902 window_len
= ras
->ras_window_len
- left
;
904 if (left
< ras
->ras_stride_pages
)
907 left
= ras
->ras_stride_pages
+ inc_len
;
909 LASSERT(ras
->ras_stride_pages
!= 0);
911 step
= left
/ ras
->ras_stride_pages
;
912 left
%= ras
->ras_stride_pages
;
914 window_len
+= step
* ras
->ras_stride_length
+ left
;
916 if (stride_pg_count(ras
->ras_stride_offset
, ras
->ras_stride_length
,
917 ras
->ras_stride_pages
, ras
->ras_stride_offset
,
918 window_len
) <= ra
->ra_max_pages_per_file
)
919 ras
->ras_window_len
= window_len
;
924 static void ras_increase_window(struct inode
*inode
,
925 struct ll_readahead_state
*ras
,
926 struct ll_ra_info
*ra
)
928 /* The stretch of ra-window should be aligned with max rpc_size
929 * but current clio architecture does not support retrieve such
930 * information from lower layer. FIXME later
932 if (stride_io_mode(ras
))
933 ras_stride_increase_window(ras
, ra
, RAS_INCREASE_STEP(inode
));
935 ras
->ras_window_len
= min(ras
->ras_window_len
+
936 RAS_INCREASE_STEP(inode
),
937 ra
->ra_max_pages_per_file
);
940 void ras_update(struct ll_sb_info
*sbi
, struct inode
*inode
,
941 struct ll_readahead_state
*ras
, unsigned long index
,
944 struct ll_ra_info
*ra
= &sbi
->ll_ra_info
;
945 int zero
= 0, stride_detect
= 0, ra_miss
= 0;
947 spin_lock(&ras
->ras_lock
);
949 ll_ra_stats_inc_sbi(sbi
, hit
? RA_STAT_HIT
: RA_STAT_MISS
);
951 /* reset the read-ahead window in two cases. First when the app seeks
952 * or reads to some other part of the file. Secondly if we get a
953 * read-ahead miss that we think we've previously issued. This can
954 * be a symptom of there being so many read-ahead pages that the VM is
955 * reclaiming it before we get to it. */
956 if (!index_in_window(index
, ras
->ras_last_readpage
, 8, 8)) {
958 ll_ra_stats_inc_sbi(sbi
, RA_STAT_DISTANT_READPAGE
);
959 } else if (!hit
&& ras
->ras_window_len
&&
960 index
< ras
->ras_next_readahead
&&
961 index_in_window(index
, ras
->ras_window_start
, 0,
962 ras
->ras_window_len
)) {
964 ll_ra_stats_inc_sbi(sbi
, RA_STAT_MISS_IN_WINDOW
);
967 /* On the second access to a file smaller than the tunable
968 * ra_max_read_ahead_whole_pages trigger RA on all pages in the
969 * file up to ra_max_pages_per_file. This is simply a best effort
970 * and only occurs once per open file. Normal RA behavior is reverted
971 * to for subsequent IO. The mmap case does not increment
972 * ras_requests and thus can never trigger this behavior. */
973 if (ras
->ras_requests
== 2 && !ras
->ras_request_index
) {
976 kms_pages
= (i_size_read(inode
) + PAGE_CACHE_SIZE
- 1) >>
979 CDEBUG(D_READA
, "kmsp %llu mwp %lu mp %lu\n", kms_pages
,
980 ra
->ra_max_read_ahead_whole_pages
, ra
->ra_max_pages_per_file
);
983 kms_pages
<= ra
->ra_max_read_ahead_whole_pages
) {
984 ras
->ras_window_start
= 0;
985 ras
->ras_last_readpage
= 0;
986 ras
->ras_next_readahead
= 0;
987 ras
->ras_window_len
= min(ra
->ra_max_pages_per_file
,
988 ra
->ra_max_read_ahead_whole_pages
);
993 /* check whether it is in stride I/O mode*/
994 if (!index_in_stride_window(ras
, index
)) {
995 if (ras
->ras_consecutive_stride_requests
== 0 &&
996 ras
->ras_request_index
== 0) {
997 ras_update_stride_detector(ras
, index
);
998 ras
->ras_consecutive_stride_requests
++;
1000 ras_stride_reset(ras
);
1002 ras_reset(inode
, ras
, index
);
1003 ras
->ras_consecutive_pages
++;
1006 ras
->ras_consecutive_pages
= 0;
1007 ras
->ras_consecutive_requests
= 0;
1008 if (++ras
->ras_consecutive_stride_requests
> 1)
1014 if (index_in_stride_window(ras
, index
) &&
1015 stride_io_mode(ras
)) {
1016 /*If stride-RA hit cache miss, the stride dector
1017 *will not be reset to avoid the overhead of
1018 *redetecting read-ahead mode */
1019 if (index
!= ras
->ras_last_readpage
+ 1)
1020 ras
->ras_consecutive_pages
= 0;
1021 ras_reset(inode
, ras
, index
);
1024 /* Reset both stride window and normal RA
1026 ras_reset(inode
, ras
, index
);
1027 ras
->ras_consecutive_pages
++;
1028 ras_stride_reset(ras
);
1031 } else if (stride_io_mode(ras
)) {
1032 /* If this is contiguous read but in stride I/O mode
1033 * currently, check whether stride step still is valid,
1034 * if invalid, it will reset the stride ra window*/
1035 if (!index_in_stride_window(ras
, index
)) {
1036 /* Shrink stride read-ahead window to be zero */
1037 ras_stride_reset(ras
);
1038 ras
->ras_window_len
= 0;
1039 ras
->ras_next_readahead
= index
;
1043 ras
->ras_consecutive_pages
++;
1044 ras
->ras_last_readpage
= index
;
1045 ras_set_start(inode
, ras
, index
);
1047 if (stride_io_mode(ras
))
1048 /* Since stride readahead is sensitive to the offset
1049 * of read-ahead, so we use original offset here,
1050 * instead of ras_window_start, which is RPC aligned */
1051 ras
->ras_next_readahead
= max(index
, ras
->ras_next_readahead
);
1053 ras
->ras_next_readahead
= max(ras
->ras_window_start
,
1054 ras
->ras_next_readahead
);
1057 /* Trigger RA in the mmap case where ras_consecutive_requests
1058 * is not incremented and thus can't be used to trigger RA */
1059 if (!ras
->ras_window_len
&& ras
->ras_consecutive_pages
== 4) {
1060 ras
->ras_window_len
= RAS_INCREASE_STEP(inode
);
1064 /* Initially reset the stride window offset to next_readahead*/
1065 if (ras
->ras_consecutive_stride_requests
== 2 && stride_detect
) {
1067 * Once stride IO mode is detected, next_readahead should be
1068 * reset to make sure next_readahead > stride offset
1070 ras
->ras_next_readahead
= max(index
, ras
->ras_next_readahead
);
1071 ras
->ras_stride_offset
= index
;
1072 ras
->ras_window_len
= RAS_INCREASE_STEP(inode
);
1075 /* The initial ras_window_len is set to the request size. To avoid
1076 * uselessly reading and discarding pages for random IO the window is
1077 * only increased once per consecutive request received. */
1078 if ((ras
->ras_consecutive_requests
> 1 || stride_detect
) &&
1079 !ras
->ras_request_index
)
1080 ras_increase_window(inode
, ras
, ra
);
1083 ras
->ras_request_index
++;
1084 spin_unlock(&ras
->ras_lock
);
1088 int ll_writepage(struct page
*vmpage
, struct writeback_control
*wbc
)
1090 struct inode
*inode
= vmpage
->mapping
->host
;
1091 struct ll_inode_info
*lli
= ll_i2info(inode
);
1094 struct cl_page
*page
;
1095 struct cl_object
*clob
;
1096 struct cl_env_nest nest
;
1097 bool redirtied
= false;
1098 bool unlocked
= false;
1101 LASSERT(PageLocked(vmpage
));
1102 LASSERT(!PageWriteback(vmpage
));
1104 LASSERT(ll_i2dtexp(inode
) != NULL
);
1106 env
= cl_env_nested_get(&nest
);
1108 result
= PTR_ERR(env
);
1112 clob
= ll_i2info(inode
)->lli_clob
;
1113 LASSERT(clob
!= NULL
);
1115 io
= ccc_env_thread_io(env
);
1117 io
->ci_ignore_layout
= 1;
1118 result
= cl_io_init(env
, io
, CIT_MISC
, clob
);
1120 page
= cl_page_find(env
, clob
, vmpage
->index
,
1121 vmpage
, CPT_CACHEABLE
);
1122 if (!IS_ERR(page
)) {
1123 lu_ref_add(&page
->cp_reference
, "writepage",
1125 cl_page_assume(env
, io
, page
);
1126 result
= cl_page_flush(env
, io
, page
);
1129 * Re-dirty page on error so it retries write,
1130 * but not in case when IO has actually
1131 * occurred and completed with an error.
1133 if (!PageError(vmpage
)) {
1134 redirty_page_for_writepage(wbc
, vmpage
);
1139 cl_page_disown(env
, io
, page
);
1141 lu_ref_del(&page
->cp_reference
,
1142 "writepage", current
);
1143 cl_page_put(env
, page
);
1145 result
= PTR_ERR(page
);
1148 cl_io_fini(env
, io
);
1150 if (redirtied
&& wbc
->sync_mode
== WB_SYNC_ALL
) {
1151 loff_t offset
= cl_offset(clob
, vmpage
->index
);
1153 /* Flush page failed because the extent is being written out.
1154 * Wait for the write of extent to be finished to avoid
1155 * breaking kernel which assumes ->writepage should mark
1156 * PageWriteback or clean the page. */
1157 result
= cl_sync_file_range(inode
, offset
,
1158 offset
+ PAGE_CACHE_SIZE
- 1,
1161 /* actually we may have written more than one page.
1162 * decreasing this page because the caller will count
1164 wbc
->nr_to_write
-= result
- 1;
1169 cl_env_nested_put(&nest
, env
);
1174 if (!lli
->lli_async_rc
)
1175 lli
->lli_async_rc
= result
;
1176 SetPageError(vmpage
);
1178 unlock_page(vmpage
);
1183 int ll_writepages(struct address_space
*mapping
, struct writeback_control
*wbc
)
1185 struct inode
*inode
= mapping
->host
;
1186 struct ll_sb_info
*sbi
= ll_i2sbi(inode
);
1189 enum cl_fsync_mode mode
;
1190 int range_whole
= 0;
1192 int ignore_layout
= 0;
1194 if (wbc
->range_cyclic
) {
1195 start
= mapping
->writeback_index
<< PAGE_CACHE_SHIFT
;
1196 end
= OBD_OBJECT_EOF
;
1198 start
= wbc
->range_start
;
1199 end
= wbc
->range_end
;
1200 if (end
== LLONG_MAX
) {
1201 end
= OBD_OBJECT_EOF
;
1202 range_whole
= start
== 0;
1206 mode
= CL_FSYNC_NONE
;
1207 if (wbc
->sync_mode
== WB_SYNC_ALL
)
1208 mode
= CL_FSYNC_LOCAL
;
1210 if (sbi
->ll_umounting
)
1211 /* if the mountpoint is being umounted, all pages have to be
1212 * evicted to avoid hitting LBUG when truncate_inode_pages()
1213 * is called later on. */
1215 result
= cl_sync_file_range(inode
, start
, end
, mode
, ignore_layout
);
1217 wbc
->nr_to_write
-= result
;
1221 if (wbc
->range_cyclic
|| (range_whole
&& wbc
->nr_to_write
> 0)) {
1222 if (end
== OBD_OBJECT_EOF
)
1223 end
= i_size_read(inode
);
1224 mapping
->writeback_index
= (end
>> PAGE_CACHE_SHIFT
) + 1;
1229 int ll_readpage(struct file
*file
, struct page
*vmpage
)
1231 struct ll_cl_context
*lcc
;
1234 lcc
= ll_cl_init(file
, vmpage
, 0);
1236 struct lu_env
*env
= lcc
->lcc_env
;
1237 struct cl_io
*io
= lcc
->lcc_io
;
1238 struct cl_page
*page
= lcc
->lcc_page
;
1240 LASSERT(page
->cp_type
== CPT_CACHEABLE
);
1241 if (likely(!PageUptodate(vmpage
))) {
1242 cl_page_assume(env
, io
, page
);
1243 result
= cl_io_read_page(env
, io
, page
);
1245 /* Page from a non-object file. */
1246 unlock_page(vmpage
);
1251 unlock_page(vmpage
);
1252 result
= PTR_ERR(lcc
);