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;
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;
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
125 CERROR("Proc %s is dirtying page w/o inode lock, this will break truncate\n",
129 return ERR_PTR(-EIO);
133 * Loop-back driver calls ->prepare_write().
134 * methods directly, bypassing file system ->write() operation,
135 * so cl_io has to be created here.
137 io = ccc_env_thread_io(env);
138 ll_io_init(io, file, 1);
140 /* No lock at all for this kind of IO - we can't do it because
141 * we have held page lock, it would cause deadlock.
142 * XXX: This causes poor performance to loop device - One page
144 * In order to get better performance, users should use
145 * lloop driver instead.
147 io->ci_lockreq = CILR_NEVER;
149 pos = vmpage->index << PAGE_CACHE_SHIFT;
151 /* Create a temp IO to serve write. */
152 result = cl_io_rw_init(env, io, CIT_WRITE, pos, PAGE_CACHE_SIZE);
154 cio->cui_fd = LUSTRE_FPRIVATE(file);
155 cio->cui_iter = NULL;
156 result = cl_io_iter_init(env, io);
158 result = cl_io_lock(env, io);
160 result = cl_io_start(env, io);
163 result = io->ci_result;
170 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
246 result = PTR_ERR(lcc);
251 int ll_commit_write(struct file *file, struct page *vmpage, unsigned from,
254 struct ll_cl_context *lcc;
257 struct cl_page *page;
262 page = lcc->lcc_page;
265 LASSERT(cl_page_is_owned(page, io));
267 if (from != to) /* handle short write case. */
268 result = cl_io_commit_write(env, io, page, from, to);
269 if (cl_page_is_owned(page, io))
270 cl_page_unassume(env, io, page);
273 * Release reference acquired by ll_prepare_write().
275 lu_ref_del(&page->cp_reference, "prepare_write", current);
276 cl_page_put(env, page);
281 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which);
284 * Get readahead pages from the filesystem readahead pool of the client for a
287 * /param sbi superblock for filesystem readahead state ll_ra_info
288 * /param ria per-thread readahead state
289 * /param pages number of pages requested for readahead for the thread.
291 * WARNING: This algorithm is used to reduce contention on sbi->ll_lock.
292 * It should work well if the ra_max_pages is much greater than the single
293 * file's read-ahead window, and not too many threads contending for
294 * these readahead pages.
296 * TODO: There may be a 'global sync problem' if many threads are trying
297 * to get an ra budget that is larger than the remaining readahead pages
298 * and reach here at exactly the same time. They will compute /a ret to
299 * consume the remaining pages, but will fail at atomic_add_return() and
300 * get a zero ra window, although there is still ra space remaining. - Jay
302 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi,
303 struct ra_io_arg *ria,
306 struct ll_ra_info *ra = &sbi->ll_ra_info;
309 /* If read-ahead pages left are less than 1M, do not do read-ahead,
310 * otherwise it will form small read RPC(< 1M), which hurt server
313 ret = min(ra->ra_max_pages - atomic_read(&ra->ra_cur_pages), pages);
314 if (ret < 0 || ret < min_t(long, PTLRPC_MAX_BRW_PAGES, pages)) {
319 /* If the non-strided (ria_pages == 0) readahead window
320 * (ria_start + ret) has grown across an RPC boundary, then trim
321 * readahead size by the amount beyond the RPC so it ends on an
322 * RPC boundary. If the readahead window is already ending on
323 * an RPC boundary (beyond_rpc == 0), or smaller than a full
324 * RPC (beyond_rpc < ret) the readahead size is unchanged.
325 * The (beyond_rpc != 0) check is skipped since the conditional
326 * branch is more expensive than subtracting zero from the result.
328 * Strided read is left unaligned to avoid small fragments beyond
329 * the RPC boundary from needing an extra read RPC.
331 if (ria->ria_pages == 0) {
332 long beyond_rpc = (ria->ria_start + ret) % PTLRPC_MAX_BRW_PAGES;
334 if (/* beyond_rpc != 0 && */ beyond_rpc < ret)
338 if (atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) {
339 atomic_sub(ret, &ra->ra_cur_pages);
347 void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
349 struct ll_ra_info *ra = &sbi->ll_ra_info;
351 atomic_sub(len, &ra->ra_cur_pages);
354 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which)
356 LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
357 lprocfs_counter_incr(sbi->ll_ra_stats, which);
360 void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which)
362 struct ll_sb_info *sbi = ll_i2sbi(mapping->host);
364 ll_ra_stats_inc_sbi(sbi, which);
367 #define RAS_CDEBUG(ras) \
369 "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu r %lu ri %lu" \
370 "csr %lu sf %lu sp %lu sl %lu\n", \
371 ras->ras_last_readpage, ras->ras_consecutive_requests, \
372 ras->ras_consecutive_pages, ras->ras_window_start, \
373 ras->ras_window_len, ras->ras_next_readahead, \
374 ras->ras_requests, ras->ras_request_index, \
375 ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
376 ras->ras_stride_pages, ras->ras_stride_length)
378 static int index_in_window(unsigned long index, unsigned long point,
379 unsigned long before, unsigned long after)
381 unsigned long start = point - before, end = point + after;
388 return start <= index && index <= end;
391 static struct ll_readahead_state *ll_ras_get(struct file *f)
393 struct ll_file_data *fd;
395 fd = LUSTRE_FPRIVATE(f);
399 void ll_ra_read_in(struct file *f, struct ll_ra_read *rar)
401 struct ll_readahead_state *ras;
405 spin_lock(&ras->ras_lock);
407 ras->ras_request_index = 0;
408 ras->ras_consecutive_requests++;
409 rar->lrr_reader = current;
411 list_add(&rar->lrr_linkage, &ras->ras_read_beads);
412 spin_unlock(&ras->ras_lock);
415 void ll_ra_read_ex(struct file *f, struct ll_ra_read *rar)
417 struct ll_readahead_state *ras;
421 spin_lock(&ras->ras_lock);
422 list_del_init(&rar->lrr_linkage);
423 spin_unlock(&ras->ras_lock);
426 static int cl_read_ahead_page(const struct lu_env *env, struct cl_io *io,
427 struct cl_page_list *queue, struct cl_page *page,
434 cl_page_assume(env, io, page);
435 lu_ref_add(&page->cp_reference, "ra", current);
436 cp = cl2ccc_page(cl_page_at(page, &vvp_device_type));
437 if (!cp->cpg_defer_uptodate && !PageUptodate(vmpage)) {
438 rc = cl_page_is_under_lock(env, io, page);
440 cp->cpg_defer_uptodate = 1;
442 cl_page_list_add(queue, page);
445 cl_page_delete(env, page);
449 /* skip completed pages */
450 cl_page_unassume(env, io, page);
452 lu_ref_del(&page->cp_reference, "ra", current);
453 cl_page_put(env, page);
458 * Initiates read-ahead of a page with given index.
460 * \retval +ve: page was added to \a queue.
462 * \retval -ENOLCK: there is no extent lock for this part of a file, stop
465 * \retval -ve, 0: page wasn't added to \a queue for other reason.
467 static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io,
468 struct cl_page_list *queue,
469 pgoff_t index, struct address_space *mapping)
472 struct cl_object *clob = ll_i2info(mapping->host)->lli_clob;
473 struct cl_page *page;
474 enum ra_stat which = _NR_RA_STAT; /* keep gcc happy */
476 const char *msg = NULL;
478 vmpage = grab_cache_page_nowait(mapping, index);
480 /* Check if vmpage was truncated or reclaimed */
481 if (vmpage->mapping == mapping) {
482 page = cl_page_find(env, clob, vmpage->index,
483 vmpage, CPT_CACHEABLE);
485 rc = cl_read_ahead_page(env, io, queue,
488 which = RA_STAT_FAILED_MATCH;
489 msg = "lock match failed";
492 which = RA_STAT_FAILED_GRAB_PAGE;
493 msg = "cl_page_find failed";
496 which = RA_STAT_WRONG_GRAB_PAGE;
497 msg = "g_c_p_n returned invalid page";
501 page_cache_release(vmpage);
503 which = RA_STAT_FAILED_GRAB_PAGE;
504 msg = "g_c_p_n failed";
507 ll_ra_stats_inc(mapping, which);
508 CDEBUG(D_READA, "%s\n", msg);
513 #define RIA_DEBUG(ria) \
514 CDEBUG(D_READA, "rs %lu re %lu ro %lu rl %lu rp %lu\n", \
515 ria->ria_start, ria->ria_end, ria->ria_stoff, ria->ria_length,\
518 /* Limit this to the blocksize instead of PTLRPC_BRW_MAX_SIZE, since we don't
519 * know what the actual RPC size is. If this needs to change, it makes more
520 * sense to tune the i_blkbits value for the file based on the OSTs it is
521 * striped over, rather than having a constant value for all files here.
524 /* RAS_INCREASE_STEP should be (1UL << (inode->i_blkbits - PAGE_CACHE_SHIFT)).
525 * Temporarily set RAS_INCREASE_STEP to 1MB. After 4MB RPC is enabled
526 * by default, this should be adjusted corresponding with max_read_ahead_mb
527 * and max_read_ahead_per_file_mb otherwise the readahead budget can be used
528 * up quickly which will affect read performance significantly. See LU-2816
530 #define RAS_INCREASE_STEP(inode) (ONE_MB_BRW_SIZE >> PAGE_CACHE_SHIFT)
532 static inline int stride_io_mode(struct ll_readahead_state *ras)
534 return ras->ras_consecutive_stride_requests > 1;
537 /* The function calculates how much pages will be read in
538 * [off, off + length], in such stride IO area,
539 * stride_offset = st_off, stride_length = st_len,
540 * stride_pages = st_pgs
542 * |------------------|*****|------------------|*****|------------|*****|....
545 * |----- st_len -----|
547 * How many pages it should read in such pattern
548 * |-------------------------------------------------------------|
550 * |<------ length ------->|
552 * = |<----->| + |-------------------------------------| + |---|
553 * start_left st_pgs * i end_left
556 stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs,
557 unsigned long off, unsigned long length)
559 __u64 start = off > st_off ? off - st_off : 0;
560 __u64 end = off + length > st_off ? off + length - st_off : 0;
561 unsigned long start_left = 0;
562 unsigned long end_left = 0;
563 unsigned long pg_count;
565 if (st_len == 0 || length == 0 || end == 0)
568 start_left = do_div(start, st_len);
569 if (start_left < st_pgs)
570 start_left = st_pgs - start_left;
574 end_left = do_div(end, st_len);
575 if (end_left > st_pgs)
578 CDEBUG(D_READA, "start %llu, end %llu start_left %lu end_left %lu\n",
579 start, end, start_left, end_left);
582 pg_count = end_left - (st_pgs - start_left);
584 pg_count = start_left + st_pgs * (end - start - 1) + end_left;
586 CDEBUG(D_READA, "st_off %lu, st_len %lu st_pgs %lu off %lu length %lu pgcount %lu\n",
587 st_off, st_len, st_pgs, off, length, pg_count);
592 static int ria_page_count(struct ra_io_arg *ria)
594 __u64 length = ria->ria_end >= ria->ria_start ?
595 ria->ria_end - ria->ria_start + 1 : 0;
597 return stride_pg_count(ria->ria_stoff, ria->ria_length,
598 ria->ria_pages, ria->ria_start,
602 /*Check whether the index is in the defined ra-window */
603 static int ras_inside_ra_window(unsigned long idx, struct ra_io_arg *ria)
605 /* If ria_length == ria_pages, it means non-stride I/O mode,
606 * idx should always inside read-ahead window in this case
607 * For stride I/O mode, just check whether the idx is inside
610 return ria->ria_length == 0 || ria->ria_length == ria->ria_pages ||
611 (idx >= ria->ria_stoff && (idx - ria->ria_stoff) %
612 ria->ria_length < ria->ria_pages);
615 static int ll_read_ahead_pages(const struct lu_env *env,
616 struct cl_io *io, struct cl_page_list *queue,
617 struct ra_io_arg *ria,
618 unsigned long *reserved_pages,
619 struct address_space *mapping,
620 unsigned long *ra_end)
622 int rc, count = 0, stride_ria;
623 unsigned long page_idx;
628 stride_ria = ria->ria_length > ria->ria_pages && ria->ria_pages > 0;
629 for (page_idx = ria->ria_start;
630 page_idx <= ria->ria_end && *reserved_pages > 0; page_idx++) {
631 if (ras_inside_ra_window(page_idx, ria)) {
632 /* If the page is inside the read-ahead window*/
633 rc = ll_read_ahead_page(env, io, queue,
638 } else if (rc == -ENOLCK)
640 } else if (stride_ria) {
641 /* If it is not in the read-ahead window, and it is
642 * read-ahead mode, then check whether it should skip
646 /* FIXME: This assertion only is valid when it is for
647 * forward read-ahead, it will be fixed when backward
648 * read-ahead is implemented
650 LASSERTF(page_idx > ria->ria_stoff, "Invalid page_idx %lu rs %lu re %lu ro %lu rl %lu rp %lu\n",
652 ria->ria_start, ria->ria_end, ria->ria_stoff,
653 ria->ria_length, ria->ria_pages);
654 offset = page_idx - ria->ria_stoff;
655 offset = offset % (ria->ria_length);
656 if (offset > ria->ria_pages) {
657 page_idx += ria->ria_length - offset;
658 CDEBUG(D_READA, "i %lu skip %lu\n", page_idx,
659 ria->ria_length - offset);
668 int ll_readahead(const struct lu_env *env, struct cl_io *io,
669 struct ll_readahead_state *ras, struct address_space *mapping,
670 struct cl_page_list *queue, int flags)
672 struct vvp_io *vio = vvp_env_io(env);
673 struct vvp_thread_info *vti = vvp_env_info(env);
674 struct cl_attr *attr = ccc_env_thread_attr(env);
675 unsigned long start = 0, end = 0, reserved;
676 unsigned long ra_end, len;
678 struct ll_ra_read *bead;
679 struct ra_io_arg *ria = &vti->vti_ria;
680 struct ll_inode_info *lli;
681 struct cl_object *clob;
685 inode = mapping->host;
686 lli = ll_i2info(inode);
687 clob = lli->lli_clob;
689 memset(ria, 0, sizeof(*ria));
691 cl_object_attr_lock(clob);
692 ret = cl_object_attr_get(env, clob, attr);
693 cl_object_attr_unlock(clob);
699 ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN);
703 spin_lock(&ras->ras_lock);
704 if (vio->cui_ra_window_set)
705 bead = &vio->cui_bead;
709 /* Enlarge the RA window to encompass the full read */
710 if (bead && ras->ras_window_start + ras->ras_window_len <
711 bead->lrr_start + bead->lrr_count) {
712 ras->ras_window_len = bead->lrr_start + bead->lrr_count -
713 ras->ras_window_start;
715 /* Reserve a part of the read-ahead window that we'll be issuing */
716 if (ras->ras_window_len) {
717 start = ras->ras_next_readahead;
718 end = ras->ras_window_start + ras->ras_window_len - 1;
721 unsigned long rpc_boundary;
723 * Align RA window to an optimal boundary.
725 * XXX This would be better to align to cl_max_pages_per_rpc
726 * instead of PTLRPC_MAX_BRW_PAGES, because the RPC size may
727 * be aligned to the RAID stripe size in the future and that
728 * is more important than the RPC size.
730 /* Note: we only trim the RPC, instead of extending the RPC
731 * to the boundary, so to avoid reading too much pages during
734 rpc_boundary = (end + 1) & (~(PTLRPC_MAX_BRW_PAGES - 1));
735 if (rpc_boundary > 0)
738 if (rpc_boundary > start)
741 /* Truncate RA window to end of file */
742 end = min(end, (unsigned long)((kms - 1) >> PAGE_CACHE_SHIFT));
744 ras->ras_next_readahead = max(end, end + 1);
747 ria->ria_start = start;
749 /* If stride I/O mode is detected, get stride window*/
750 if (stride_io_mode(ras)) {
751 ria->ria_stoff = ras->ras_stride_offset;
752 ria->ria_length = ras->ras_stride_length;
753 ria->ria_pages = ras->ras_stride_pages;
755 spin_unlock(&ras->ras_lock);
758 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
761 len = ria_page_count(ria);
765 reserved = ll_ra_count_get(ll_i2sbi(inode), ria, len);
767 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
769 CDEBUG(D_READA, "reserved page %lu ra_cur %d ra_max %lu\n", reserved,
770 atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
771 ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
773 ret = ll_read_ahead_pages(env, io, queue,
774 ria, &reserved, mapping, &ra_end);
777 ll_ra_count_put(ll_i2sbi(inode), reserved);
779 if (ra_end == end + 1 && ra_end == (kms >> PAGE_CACHE_SHIFT))
780 ll_ra_stats_inc(mapping, RA_STAT_EOF);
782 /* if we didn't get to the end of the region we reserved from
783 * the ras we need to go back and update the ras so that the
784 * next read-ahead tries from where we left off. we only do so
785 * if the region we failed to issue read-ahead on is still ahead
786 * of the app and behind the next index to start read-ahead from
788 CDEBUG(D_READA, "ra_end %lu end %lu stride end %lu\n",
789 ra_end, end, ria->ria_end);
791 if (ra_end != end + 1) {
792 spin_lock(&ras->ras_lock);
793 if (ra_end < ras->ras_next_readahead &&
794 index_in_window(ra_end, ras->ras_window_start, 0,
795 ras->ras_window_len)) {
796 ras->ras_next_readahead = ra_end;
799 spin_unlock(&ras->ras_lock);
805 static void ras_set_start(struct inode *inode, struct ll_readahead_state *ras,
808 ras->ras_window_start = index & (~(RAS_INCREASE_STEP(inode) - 1));
811 /* called with the ras_lock held or from places where it doesn't matter */
812 static void ras_reset(struct inode *inode, struct ll_readahead_state *ras,
815 ras->ras_last_readpage = index;
816 ras->ras_consecutive_requests = 0;
817 ras->ras_consecutive_pages = 0;
818 ras->ras_window_len = 0;
819 ras_set_start(inode, ras, index);
820 ras->ras_next_readahead = max(ras->ras_window_start, index);
825 /* called with the ras_lock held or from places where it doesn't matter */
826 static void ras_stride_reset(struct ll_readahead_state *ras)
828 ras->ras_consecutive_stride_requests = 0;
829 ras->ras_stride_length = 0;
830 ras->ras_stride_pages = 0;
834 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
836 spin_lock_init(&ras->ras_lock);
837 ras_reset(inode, ras, 0);
838 ras->ras_requests = 0;
839 INIT_LIST_HEAD(&ras->ras_read_beads);
843 * Check whether the read request is in the stride window.
844 * If it is in the stride window, return 1, otherwise return 0.
846 static int index_in_stride_window(struct ll_readahead_state *ras,
849 unsigned long stride_gap;
851 if (ras->ras_stride_length == 0 || ras->ras_stride_pages == 0 ||
852 ras->ras_stride_pages == ras->ras_stride_length)
855 stride_gap = index - ras->ras_last_readpage - 1;
857 /* If it is contiguous read */
859 return ras->ras_consecutive_pages + 1 <= ras->ras_stride_pages;
861 /* Otherwise check the stride by itself */
862 return (ras->ras_stride_length - ras->ras_stride_pages) == stride_gap &&
863 ras->ras_consecutive_pages == ras->ras_stride_pages;
866 static void ras_update_stride_detector(struct ll_readahead_state *ras,
869 unsigned long stride_gap = index - ras->ras_last_readpage - 1;
871 if (!stride_io_mode(ras) && (stride_gap != 0 ||
872 ras->ras_consecutive_stride_requests == 0)) {
873 ras->ras_stride_pages = ras->ras_consecutive_pages;
874 ras->ras_stride_length = stride_gap+ras->ras_consecutive_pages;
876 LASSERT(ras->ras_request_index == 0);
877 LASSERT(ras->ras_consecutive_stride_requests == 0);
879 if (index <= ras->ras_last_readpage) {
880 /*Reset stride window for forward read*/
881 ras_stride_reset(ras);
885 ras->ras_stride_pages = ras->ras_consecutive_pages;
886 ras->ras_stride_length = stride_gap+ras->ras_consecutive_pages;
892 /* Stride Read-ahead window will be increased inc_len according to
895 static void ras_stride_increase_window(struct ll_readahead_state *ras,
896 struct ll_ra_info *ra,
897 unsigned long inc_len)
899 unsigned long left, step, window_len;
900 unsigned long stride_len;
902 LASSERT(ras->ras_stride_length > 0);
903 LASSERTF(ras->ras_window_start + ras->ras_window_len
904 >= ras->ras_stride_offset, "window_start %lu, window_len %lu stride_offset %lu\n",
905 ras->ras_window_start,
906 ras->ras_window_len, ras->ras_stride_offset);
908 stride_len = ras->ras_window_start + ras->ras_window_len -
909 ras->ras_stride_offset;
911 left = stride_len % ras->ras_stride_length;
912 window_len = ras->ras_window_len - left;
914 if (left < ras->ras_stride_pages)
917 left = ras->ras_stride_pages + inc_len;
919 LASSERT(ras->ras_stride_pages != 0);
921 step = left / ras->ras_stride_pages;
922 left %= ras->ras_stride_pages;
924 window_len += step * ras->ras_stride_length + left;
926 if (stride_pg_count(ras->ras_stride_offset, ras->ras_stride_length,
927 ras->ras_stride_pages, ras->ras_stride_offset,
928 window_len) <= ra->ra_max_pages_per_file)
929 ras->ras_window_len = window_len;
934 static void ras_increase_window(struct inode *inode,
935 struct ll_readahead_state *ras,
936 struct ll_ra_info *ra)
938 /* The stretch of ra-window should be aligned with max rpc_size
939 * but current clio architecture does not support retrieve such
940 * information from lower layer. FIXME later
942 if (stride_io_mode(ras))
943 ras_stride_increase_window(ras, ra, RAS_INCREASE_STEP(inode));
945 ras->ras_window_len = min(ras->ras_window_len +
946 RAS_INCREASE_STEP(inode),
947 ra->ra_max_pages_per_file);
950 void ras_update(struct ll_sb_info *sbi, struct inode *inode,
951 struct ll_readahead_state *ras, unsigned long index,
954 struct ll_ra_info *ra = &sbi->ll_ra_info;
955 int zero = 0, stride_detect = 0, ra_miss = 0;
957 spin_lock(&ras->ras_lock);
959 ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
961 /* reset the read-ahead window in two cases. First when the app seeks
962 * or reads to some other part of the file. Secondly if we get a
963 * read-ahead miss that we think we've previously issued. This can
964 * be a symptom of there being so many read-ahead pages that the VM is
965 * reclaiming it before we get to it.
967 if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
969 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
970 } else if (!hit && ras->ras_window_len &&
971 index < ras->ras_next_readahead &&
972 index_in_window(index, ras->ras_window_start, 0,
973 ras->ras_window_len)) {
975 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
978 /* On the second access to a file smaller than the tunable
979 * ra_max_read_ahead_whole_pages trigger RA on all pages in the
980 * file up to ra_max_pages_per_file. This is simply a best effort
981 * and only occurs once per open file. Normal RA behavior is reverted
982 * to for subsequent IO. The mmap case does not increment
983 * ras_requests and thus can never trigger this behavior.
985 if (ras->ras_requests == 2 && !ras->ras_request_index) {
988 kms_pages = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >>
991 CDEBUG(D_READA, "kmsp %llu mwp %lu mp %lu\n", kms_pages,
992 ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages_per_file);
995 kms_pages <= ra->ra_max_read_ahead_whole_pages) {
996 ras->ras_window_start = 0;
997 ras->ras_last_readpage = 0;
998 ras->ras_next_readahead = 0;
999 ras->ras_window_len = min(ra->ra_max_pages_per_file,
1000 ra->ra_max_read_ahead_whole_pages);
1005 /* check whether it is in stride I/O mode*/
1006 if (!index_in_stride_window(ras, index)) {
1007 if (ras->ras_consecutive_stride_requests == 0 &&
1008 ras->ras_request_index == 0) {
1009 ras_update_stride_detector(ras, index);
1010 ras->ras_consecutive_stride_requests++;
1012 ras_stride_reset(ras);
1014 ras_reset(inode, ras, index);
1015 ras->ras_consecutive_pages++;
1018 ras->ras_consecutive_pages = 0;
1019 ras->ras_consecutive_requests = 0;
1020 if (++ras->ras_consecutive_stride_requests > 1)
1026 if (index_in_stride_window(ras, index) &&
1027 stride_io_mode(ras)) {
1028 /*If stride-RA hit cache miss, the stride dector
1029 *will not be reset to avoid the overhead of
1030 *redetecting read-ahead mode
1032 if (index != ras->ras_last_readpage + 1)
1033 ras->ras_consecutive_pages = 0;
1034 ras_reset(inode, ras, index);
1037 /* Reset both stride window and normal RA
1040 ras_reset(inode, ras, index);
1041 ras->ras_consecutive_pages++;
1042 ras_stride_reset(ras);
1045 } else if (stride_io_mode(ras)) {
1046 /* If this is contiguous read but in stride I/O mode
1047 * currently, check whether stride step still is valid,
1048 * if invalid, it will reset the stride ra window
1050 if (!index_in_stride_window(ras, index)) {
1051 /* Shrink stride read-ahead window to be zero */
1052 ras_stride_reset(ras);
1053 ras->ras_window_len = 0;
1054 ras->ras_next_readahead = index;
1058 ras->ras_consecutive_pages++;
1059 ras->ras_last_readpage = index;
1060 ras_set_start(inode, ras, index);
1062 if (stride_io_mode(ras))
1063 /* Since stride readahead is sensitive to the offset
1064 * of read-ahead, so we use original offset here,
1065 * instead of ras_window_start, which is RPC aligned
1067 ras->ras_next_readahead = max(index, ras->ras_next_readahead);
1069 ras->ras_next_readahead = max(ras->ras_window_start,
1070 ras->ras_next_readahead);
1073 /* Trigger RA in the mmap case where ras_consecutive_requests
1074 * is not incremented and thus can't be used to trigger RA
1076 if (!ras->ras_window_len && ras->ras_consecutive_pages == 4) {
1077 ras->ras_window_len = RAS_INCREASE_STEP(inode);
1081 /* Initially reset the stride window offset to next_readahead*/
1082 if (ras->ras_consecutive_stride_requests == 2 && stride_detect) {
1084 * Once stride IO mode is detected, next_readahead should be
1085 * reset to make sure next_readahead > stride offset
1087 ras->ras_next_readahead = max(index, ras->ras_next_readahead);
1088 ras->ras_stride_offset = index;
1089 ras->ras_window_len = RAS_INCREASE_STEP(inode);
1092 /* The initial ras_window_len is set to the request size. To avoid
1093 * uselessly reading and discarding pages for random IO the window is
1094 * only increased once per consecutive request received. */
1095 if ((ras->ras_consecutive_requests > 1 || stride_detect) &&
1096 !ras->ras_request_index)
1097 ras_increase_window(inode, ras, ra);
1100 ras->ras_request_index++;
1101 spin_unlock(&ras->ras_lock);
1105 int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
1107 struct inode *inode = vmpage->mapping->host;
1108 struct ll_inode_info *lli = ll_i2info(inode);
1111 struct cl_page *page;
1112 struct cl_object *clob;
1113 struct cl_env_nest nest;
1114 bool redirtied = false;
1115 bool unlocked = false;
1118 LASSERT(PageLocked(vmpage));
1119 LASSERT(!PageWriteback(vmpage));
1121 LASSERT(ll_i2dtexp(inode));
1123 env = cl_env_nested_get(&nest);
1125 result = PTR_ERR(env);
1129 clob = ll_i2info(inode)->lli_clob;
1132 io = ccc_env_thread_io(env);
1134 io->ci_ignore_layout = 1;
1135 result = cl_io_init(env, io, CIT_MISC, clob);
1137 page = cl_page_find(env, clob, vmpage->index,
1138 vmpage, CPT_CACHEABLE);
1139 if (!IS_ERR(page)) {
1140 lu_ref_add(&page->cp_reference, "writepage",
1142 cl_page_assume(env, io, page);
1143 result = cl_page_flush(env, io, page);
1146 * Re-dirty page on error so it retries write,
1147 * but not in case when IO has actually
1148 * occurred and completed with an error.
1150 if (!PageError(vmpage)) {
1151 redirty_page_for_writepage(wbc, vmpage);
1156 cl_page_disown(env, io, page);
1158 lu_ref_del(&page->cp_reference,
1159 "writepage", current);
1160 cl_page_put(env, page);
1162 result = PTR_ERR(page);
1165 cl_io_fini(env, io);
1167 if (redirtied && wbc->sync_mode == WB_SYNC_ALL) {
1168 loff_t offset = cl_offset(clob, vmpage->index);
1170 /* Flush page failed because the extent is being written out.
1171 * Wait for the write of extent to be finished to avoid
1172 * breaking kernel which assumes ->writepage should mark
1173 * PageWriteback or clean the page.
1175 result = cl_sync_file_range(inode, offset,
1176 offset + PAGE_CACHE_SIZE - 1,
1179 /* actually we may have written more than one page.
1180 * decreasing this page because the caller will count
1183 wbc->nr_to_write -= result - 1;
1188 cl_env_nested_put(&nest, env);
1193 if (!lli->lli_async_rc)
1194 lli->lli_async_rc = result;
1195 SetPageError(vmpage);
1197 unlock_page(vmpage);
1202 int ll_writepages(struct address_space *mapping, struct writeback_control *wbc)
1204 struct inode *inode = mapping->host;
1205 struct ll_sb_info *sbi = ll_i2sbi(inode);
1208 enum cl_fsync_mode mode;
1209 int range_whole = 0;
1211 int ignore_layout = 0;
1213 if (wbc->range_cyclic) {
1214 start = mapping->writeback_index << PAGE_CACHE_SHIFT;
1215 end = OBD_OBJECT_EOF;
1217 start = wbc->range_start;
1218 end = wbc->range_end;
1219 if (end == LLONG_MAX) {
1220 end = OBD_OBJECT_EOF;
1221 range_whole = start == 0;
1225 mode = CL_FSYNC_NONE;
1226 if (wbc->sync_mode == WB_SYNC_ALL)
1227 mode = CL_FSYNC_LOCAL;
1229 if (sbi->ll_umounting)
1230 /* if the mountpoint is being umounted, all pages have to be
1231 * evicted to avoid hitting LBUG when truncate_inode_pages()
1232 * is called later on.
1235 result = cl_sync_file_range(inode, start, end, mode, ignore_layout);
1237 wbc->nr_to_write -= result;
1241 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) {
1242 if (end == OBD_OBJECT_EOF)
1243 end = i_size_read(inode);
1244 mapping->writeback_index = (end >> PAGE_CACHE_SHIFT) + 1;
1249 int ll_readpage(struct file *file, struct page *vmpage)
1251 struct ll_cl_context *lcc;
1254 lcc = ll_cl_init(file, vmpage, 0);
1256 struct lu_env *env = lcc->lcc_env;
1257 struct cl_io *io = lcc->lcc_io;
1258 struct cl_page *page = lcc->lcc_page;
1260 LASSERT(page->cp_type == CPT_CACHEABLE);
1261 if (likely(!PageUptodate(vmpage))) {
1262 cl_page_assume(env, io, page);
1263 result = cl_io_read_page(env, io, page);
1265 /* Page from a non-object file. */
1266 unlock_page(vmpage);
1271 unlock_page(vmpage);
1272 result = PTR_ERR(lcc);