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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
34 * Lustre Lite I/O page cache routines shared by different kernel revs
37 #include <linux/kernel.h>
39 #include <linux/string.h>
40 #include <linux/stat.h>
41 #include <linux/errno.h>
42 #include <linux/unistd.h>
43 #include <linux/writeback.h>
44 #include <asm/uaccess.h>
47 #include <linux/file.h>
48 #include <linux/stat.h>
49 #include <asm/uaccess.h>
51 #include <linux/pagemap.h>
52 /* current_is_kswapd() */
53 #include <linux/swap.h>
54 #include <linux/task_io_accounting_ops.h>
56 #define DEBUG_SUBSYSTEM S_LLITE
58 #include <obd_cksum.h>
59 #include "llite_internal.h"
60 #include <lustre_compat.h>
62 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which);
65 * Get readahead pages from the filesystem readahead pool of the client for a
68 * /param sbi superblock for filesystem readahead state ll_ra_info
69 * /param ria per-thread readahead state
70 * /param pages number of pages requested for readahead for the thread.
72 * WARNING: This algorithm is used to reduce contention on sbi->ll_lock.
73 * It should work well if the ra_max_pages is much greater than the single
74 * file's read-ahead window, and not too many threads contending for
75 * these readahead pages.
77 * TODO: There may be a 'global sync problem' if many threads are trying
78 * to get an ra budget that is larger than the remaining readahead pages
79 * and reach here at exactly the same time. They will compute /a ret to
80 * consume the remaining pages, but will fail at atomic_add_return() and
81 * get a zero ra window, although there is still ra space remaining. - Jay */
83 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi,
84 struct ra_io_arg *ria,
86 unsigned long pages_min)
88 struct ll_ra_info *ra = &sbi->ll_ra_info;
92 /* If read-ahead pages left are less than 1M, do not do read-ahead,
93 * otherwise it will form small read RPC(< 1M), which hurt server
94 * performance a lot. */
95 ret = min(ra->ra_max_pages - atomic_read(&ra->ra_cur_pages),
97 if (ret < 0 || ret < min_t(long, PTLRPC_MAX_BRW_PAGES, pages))
100 if (atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) {
101 atomic_sub(ret, &ra->ra_cur_pages);
106 if (ret < pages_min) {
107 /* override ra limit for maximum performance */
108 atomic_add(pages_min - ret, &ra->ra_cur_pages);
114 void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long pages)
116 struct ll_ra_info *ra = &sbi->ll_ra_info;
117 atomic_sub(pages, &ra->ra_cur_pages);
120 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which)
122 LASSERTF(which < _NR_RA_STAT, "which: %u\n", which);
123 lprocfs_counter_incr(sbi->ll_ra_stats, which);
126 void ll_ra_stats_inc(struct inode *inode, enum ra_stat which)
128 struct ll_sb_info *sbi = ll_i2sbi(inode);
129 ll_ra_stats_inc_sbi(sbi, which);
132 #define RAS_CDEBUG(ras) \
134 "lre %llu cr %lu cb %llu wsi %lu wp %lu nra %lu rpc %lu " \
135 "r %lu csr %lu so %llu sb %llu sl %llu lr %lu\n", \
136 ras->ras_last_read_end_bytes, ras->ras_consecutive_requests, \
137 ras->ras_consecutive_bytes, ras->ras_window_start_idx, \
138 ras->ras_window_pages, ras->ras_next_readahead_idx, \
139 ras->ras_rpc_pages, ras->ras_requests, \
140 ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
141 ras->ras_stride_bytes, ras->ras_stride_length, \
142 ras->ras_async_last_readpage_idx)
144 static bool pos_in_window(loff_t pos, loff_t point,
145 unsigned long before, unsigned long after)
147 loff_t start = point - before;
148 loff_t end = point + after;
155 return start <= pos && pos <= end;
159 * Initiates read-ahead of a page with given index.
161 * \retval +ve: page was already uptodate so it will be skipped
163 * \retval -ve: page wasn't added to \a queue for error;
164 * \retval 0: page was added into \a queue for read ahead.
166 static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io,
167 struct cl_page_list *queue, pgoff_t index)
169 struct cl_object *clob = io->ci_obj;
170 struct inode *inode = vvp_object_inode(clob);
172 struct cl_page *page;
173 struct vvp_page *vpg;
174 enum ra_stat which = _NR_RA_STAT; /* keep gcc happy */
176 const char *msg = NULL;
179 vmpage = grab_cache_page_nowait(inode->i_mapping, index);
180 if (vmpage == NULL) {
181 which = RA_STAT_FAILED_GRAB_PAGE;
182 msg = "g_c_p_n failed";
183 GOTO(out, rc = -EBUSY);
186 /* Check if vmpage was truncated or reclaimed */
187 if (vmpage->mapping != inode->i_mapping) {
188 which = RA_STAT_WRONG_GRAB_PAGE;
189 msg = "g_c_p_n returned invalid page";
190 GOTO(out, rc = -EBUSY);
193 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
195 which = RA_STAT_FAILED_GRAB_PAGE;
196 msg = "cl_page_find failed";
197 GOTO(out, rc = PTR_ERR(page));
200 lu_ref_add(&page->cp_reference, "ra", current);
201 cl_page_assume(env, io, page);
202 vpg = cl2vvp_page(cl_object_page_slice(clob, page));
203 if (!vpg->vpg_defer_uptodate && !PageUptodate(vmpage)) {
204 vpg->vpg_defer_uptodate = 1;
205 vpg->vpg_ra_used = 0;
206 cl_page_list_add(queue, page);
208 /* skip completed pages */
209 cl_page_unassume(env, io, page);
210 /* This page is already uptodate, returning a positive number
211 * to tell the callers about this */
215 lu_ref_del(&page->cp_reference, "ra", current);
216 cl_page_put(env, page);
219 if (vmpage != NULL) {
225 ll_ra_stats_inc(inode, which);
226 CDEBUG(D_READA, "%s\n", msg);
233 #define RIA_DEBUG(ria) \
234 CDEBUG(D_READA, "rs %lu re %lu ro %llu rl %llu rb %llu\n", \
235 ria->ria_start_idx, ria->ria_end_idx, ria->ria_stoff, \
236 ria->ria_length, ria->ria_bytes)
238 static inline int stride_io_mode(struct ll_readahead_state *ras)
240 return ras->ras_consecutive_stride_requests > 1;
243 /* The function calculates how many bytes will be read in
244 * [off, off + length], in such stride IO area,
245 * stride_offset = st_off, stride_lengh = st_len,
246 * stride_bytes = st_bytes
248 * |------------------|*****|------------------|*****|------------|*****|....
251 * |----- st_len -----|
253 * How many bytes it should read in such pattern
254 * |-------------------------------------------------------------|
256 * |<------ length ------->|
258 * = |<----->| + |-------------------------------------| + |---|
259 * start_left st_bytes * i end_left
261 static loff_t stride_byte_count(loff_t st_off, loff_t st_len, loff_t st_bytes,
262 loff_t off, loff_t length)
264 u64 start = off > st_off ? off - st_off : 0;
265 u64 end = off + length > st_off ? off + length - st_off : 0;
270 if (st_len == 0 || length == 0 || end == 0)
273 start = div64_u64_rem(start, st_len, &start_left);
274 if (start_left < st_bytes)
275 start_left = st_bytes - start_left;
279 end = div64_u64_rem(end, st_len, &end_left);
280 if (end_left > st_bytes)
283 CDEBUG(D_READA, "start %llu, end %llu start_left %llu end_left %llu\n",
284 start, end, start_left, end_left);
287 bytes_count = end_left - (st_bytes - start_left);
289 bytes_count = start_left +
290 st_bytes * (end - start - 1) + end_left;
293 "st_off %llu, st_len %llu st_bytes %llu off %llu length %llu bytescount %llu\n",
294 st_off, st_len, st_bytes, off, length, bytes_count);
299 static unsigned long ria_page_count(struct ra_io_arg *ria)
301 loff_t length_bytes = ria->ria_end_idx >= ria->ria_start_idx ?
302 (loff_t)(ria->ria_end_idx -
303 ria->ria_start_idx + 1) << PAGE_SHIFT : 0;
306 if (ria->ria_length > ria->ria_bytes && ria->ria_bytes &&
307 (ria->ria_length & ~PAGE_MASK || ria->ria_bytes & ~PAGE_MASK ||
308 ria->ria_stoff & ~PAGE_MASK)) {
309 /* Over-estimate un-aligned page stride read */
310 unsigned long pg_count = ((ria->ria_bytes +
311 PAGE_SIZE - 1) >> PAGE_SHIFT) + 1;
312 pg_count *= length_bytes / ria->ria_length + 1;
316 bytes_count = stride_byte_count(ria->ria_stoff, ria->ria_length,
318 (loff_t)ria->ria_start_idx<<PAGE_SHIFT,
320 return (bytes_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
323 static pgoff_t ras_align(struct ll_readahead_state *ras, pgoff_t index)
325 return index - (index % ras->ras_rpc_pages);
328 /* Check whether the index is in the defined ra-window */
329 static bool ras_inside_ra_window(pgoff_t idx, struct ra_io_arg *ria)
331 loff_t pos = (loff_t)idx << PAGE_SHIFT;
333 /* If ria_length == ria_bytes, it means non-stride I/O mode,
334 * idx should always inside read-ahead window in this case
335 * For stride I/O mode, just check whether the idx is inside
338 if (ria->ria_length == 0 || ria->ria_length == ria->ria_bytes)
341 if (pos >= ria->ria_stoff) {
344 div64_u64_rem(pos - ria->ria_stoff, ria->ria_length, &offset);
346 if (offset < ria->ria_bytes ||
347 (ria->ria_length - offset) < PAGE_SIZE)
349 } else if (pos + PAGE_SIZE > ria->ria_stoff) {
357 ll_read_ahead_pages(const struct lu_env *env, struct cl_io *io,
358 struct cl_page_list *queue, struct ll_readahead_state *ras,
359 struct ra_io_arg *ria, pgoff_t *ra_end)
361 struct cl_read_ahead ra = { 0 };
362 /* busy page count is per stride */
363 int rc = 0, count = 0, busy_page_count = 0;
366 LASSERT(ria != NULL);
369 for (page_idx = ria->ria_start_idx;
370 page_idx <= ria->ria_end_idx && ria->ria_reserved > 0;
372 if (ras_inside_ra_window(page_idx, ria)) {
373 if (ra.cra_end_idx == 0 || ra.cra_end_idx < page_idx) {
377 * Do not shrink ria_end_idx at any case until
378 * the minimum end of current read is covered.
380 * Do not extend read lock accross stripe if
381 * lock contention detected.
383 if (ra.cra_contention &&
384 page_idx > ria->ria_end_idx_min) {
385 ria->ria_end_idx = *ra_end;
389 cl_read_ahead_release(env, &ra);
391 rc = cl_io_read_ahead(env, io, page_idx, &ra);
396 * Only shrink ria_end_idx if the matched
397 * LDLM lock doesn't cover more.
399 if (page_idx > ra.cra_end_idx) {
400 ria->ria_end_idx = ra.cra_end_idx;
404 CDEBUG(D_READA, "idx: %lu, ra: %lu, rpc: %lu\n",
405 page_idx, ra.cra_end_idx,
407 LASSERTF(ra.cra_end_idx >= page_idx,
408 "object: %p, indcies %lu / %lu\n",
409 io->ci_obj, ra.cra_end_idx, page_idx);
410 /* update read ahead RPC size.
411 * NB: it's racy but doesn't matter */
412 if (ras->ras_rpc_pages != ra.cra_rpc_pages &&
413 ra.cra_rpc_pages > 0)
414 ras->ras_rpc_pages = ra.cra_rpc_pages;
415 /* trim it to align with optimal RPC size */
416 end_idx = ras_align(ras, ria->ria_end_idx + 1);
417 if (end_idx > 0 && !ria->ria_eof)
418 ria->ria_end_idx = end_idx - 1;
419 if (ria->ria_end_idx < ria->ria_end_idx_min)
420 ria->ria_end_idx = ria->ria_end_idx_min;
422 if (page_idx > ria->ria_end_idx)
425 /* If the page is inside the read-ahead window */
426 rc = ll_read_ahead_page(env, io, queue, page_idx);
427 if (rc < 0 && rc != -EBUSY)
432 "skip busy page: %lu\n", page_idx);
433 /* For page unaligned readahead the first
434 * last pages of each region can be read by
435 * another reader on the same node, and so
436 * may be busy. So only stop for > 2 busy
438 if (busy_page_count > 2)
443 /* Only subtract from reserve & count the page if we
444 * really did readahead on that page. */
449 } else if (stride_io_mode(ras)) {
450 /* If it is not in the read-ahead window, and it is
451 * read-ahead mode, then check whether it should skip
454 loff_t pos = (loff_t)page_idx << PAGE_SHIFT;
457 div64_u64_rem(pos - ria->ria_stoff, ria->ria_length,
459 if (offset >= ria->ria_bytes) {
460 pos += (ria->ria_length - offset);
461 if ((pos >> PAGE_SHIFT) >= page_idx + 1)
462 page_idx = (pos >> PAGE_SHIFT) - 1;
465 "Stride: jump %llu pages to %lu\n",
466 ria->ria_length - offset, page_idx);
472 cl_read_ahead_release(env, &ra);
477 static void ll_readahead_work_free(struct ll_readahead_work *work)
479 fput(work->lrw_file);
483 static void ll_readahead_handle_work(struct work_struct *wq);
484 static void ll_readahead_work_add(struct inode *inode,
485 struct ll_readahead_work *work)
487 INIT_WORK(&work->lrw_readahead_work, ll_readahead_handle_work);
488 queue_work(ll_i2sbi(inode)->ll_ra_info.ll_readahead_wq,
489 &work->lrw_readahead_work);
492 static int ll_readahead_file_kms(const struct lu_env *env,
493 struct cl_io *io, __u64 *kms)
495 struct cl_object *clob;
497 struct cl_attr *attr = vvp_env_thread_attr(env);
501 inode = vvp_object_inode(clob);
503 cl_object_attr_lock(clob);
504 ret = cl_object_attr_get(env, clob, attr);
505 cl_object_attr_unlock(clob);
510 *kms = attr->cat_kms;
514 static void ll_readahead_handle_work(struct work_struct *wq)
516 struct ll_readahead_work *work;
519 struct ra_io_arg *ria;
521 struct ll_file_data *fd;
522 struct ll_readahead_state *ras;
524 struct cl_2queue *queue;
525 pgoff_t ra_end_idx = 0;
526 unsigned long pages, pages_min = 0;
532 work = container_of(wq, struct ll_readahead_work,
534 fd = work->lrw_file->private_data;
536 file = work->lrw_file;
537 inode = file_inode(file);
539 env = cl_env_alloc(&refcheck, LCT_NOREF);
541 GOTO(out_free_work, rc = PTR_ERR(env));
543 io = vvp_env_thread_io(env);
544 ll_io_init(io, file, CIT_READ, NULL);
546 rc = ll_readahead_file_kms(env, io, &kms);
548 GOTO(out_put_env, rc);
551 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
552 GOTO(out_put_env, rc = 0);
555 ria = &ll_env_info(env)->lti_ria;
556 memset(ria, 0, sizeof(*ria));
558 ria->ria_start_idx = work->lrw_start_idx;
559 /* Truncate RA window to end of file */
560 eof_index = (pgoff_t)(kms - 1) >> PAGE_SHIFT;
561 if (eof_index <= work->lrw_end_idx) {
562 work->lrw_end_idx = eof_index;
565 if (work->lrw_end_idx <= work->lrw_start_idx)
566 GOTO(out_put_env, rc = 0);
568 ria->ria_end_idx = work->lrw_end_idx;
569 pages = ria->ria_end_idx - ria->ria_start_idx + 1;
570 ria->ria_reserved = ll_ra_count_get(ll_i2sbi(inode), ria,
571 ria_page_count(ria), pages_min);
574 "async reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
575 ria->ria_reserved, pages, pages_min,
576 atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
577 ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
579 if (ria->ria_reserved < pages) {
580 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
581 if (PAGES_TO_MiB(ria->ria_reserved) < 1) {
582 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
583 GOTO(out_put_env, rc = 0);
587 rc = cl_io_rw_init(env, io, CIT_READ, ria->ria_start_idx, pages);
589 GOTO(out_put_env, rc);
591 vvp_env_io(env)->vui_io_subtype = IO_NORMAL;
592 vvp_env_io(env)->vui_fd = fd;
593 io->ci_state = CIS_LOCKED;
594 io->ci_async_readahead = true;
595 rc = cl_io_start(env, io);
597 GOTO(out_io_fini, rc);
599 queue = &io->ci_queue;
600 cl_2queue_init(queue);
602 rc = ll_read_ahead_pages(env, io, &queue->c2_qin, ras, ria,
604 if (ria->ria_reserved != 0)
605 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
606 if (queue->c2_qin.pl_nr > 0) {
607 int count = queue->c2_qin.pl_nr;
609 rc = cl_io_submit_rw(env, io, CRT_READ, queue);
611 task_io_account_read(PAGE_SIZE * count);
613 if (ria->ria_end_idx == ra_end_idx && ra_end_idx == (kms >> PAGE_SHIFT))
614 ll_ra_stats_inc(inode, RA_STAT_EOF);
616 if (ra_end_idx != ria->ria_end_idx)
617 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
619 /* TODO: discard all pages until page reinit route is implemented */
620 cl_page_list_discard(env, io, &queue->c2_qin);
622 /* Unlock unsent read pages in case of error. */
623 cl_page_list_disown(env, io, &queue->c2_qin);
625 cl_2queue_fini(env, queue);
630 cl_env_put(env, &refcheck);
633 ll_ra_stats_inc_sbi(ll_i2sbi(inode), RA_STAT_ASYNC);
634 ll_readahead_work_free(work);
637 static int ll_readahead(const struct lu_env *env, struct cl_io *io,
638 struct cl_page_list *queue,
639 struct ll_readahead_state *ras, bool hit,
642 struct vvp_io *vio = vvp_env_io(env);
643 struct ll_thread_info *lti = ll_env_info(env);
644 unsigned long pages, pages_min = 0;
645 pgoff_t ra_end_idx = 0, start_idx = 0, end_idx = 0;
647 struct ra_io_arg *ria = <i->lti_ria;
648 struct cl_object *clob;
654 inode = vvp_object_inode(clob);
656 memset(ria, 0, sizeof(*ria));
657 ret = ll_readahead_file_kms(env, io, &kms);
662 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
666 spin_lock(&ras->ras_lock);
669 * Note: other thread might rollback the ras_next_readahead_idx,
670 * if it can not get the full size of prepared pages, see the
671 * end of this function. For stride read ahead, it needs to
672 * make sure the offset is no less than ras_stride_offset,
673 * so that stride read ahead can work correctly.
675 if (stride_io_mode(ras))
676 start_idx = max_t(pgoff_t, ras->ras_next_readahead_idx,
677 ras->ras_stride_offset >> PAGE_SHIFT);
679 start_idx = ras->ras_next_readahead_idx;
681 if (ras->ras_window_pages > 0)
682 end_idx = ras->ras_window_start_idx + ras->ras_window_pages - 1;
684 /* Enlarge the RA window to encompass the full read */
685 if (vio->vui_ra_valid &&
686 end_idx < vio->vui_ra_start_idx + vio->vui_ra_pages - 1)
687 end_idx = vio->vui_ra_start_idx + vio->vui_ra_pages - 1;
692 /* Truncate RA window to end of file */
693 eof_index = (pgoff_t)((kms - 1) >> PAGE_SHIFT);
694 if (eof_index <= end_idx) {
699 ria->ria_start_idx = start_idx;
700 ria->ria_end_idx = end_idx;
701 /* If stride I/O mode is detected, get stride window*/
702 if (stride_io_mode(ras)) {
703 ria->ria_stoff = ras->ras_stride_offset;
704 ria->ria_length = ras->ras_stride_length;
705 ria->ria_bytes = ras->ras_stride_bytes;
707 spin_unlock(&ras->ras_lock);
710 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
713 pages = ria_page_count(ria);
715 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
720 CDEBUG(D_READA, DFID": ria: %lu/%lu, bead: %lu/%lu, hit: %d\n",
721 PFID(lu_object_fid(&clob->co_lu)),
722 ria->ria_start_idx, ria->ria_end_idx,
723 vio->vui_ra_valid ? vio->vui_ra_start_idx : 0,
724 vio->vui_ra_valid ? vio->vui_ra_pages : 0,
727 /* at least to extend the readahead window to cover current read */
728 if (!hit && vio->vui_ra_valid &&
729 vio->vui_ra_start_idx + vio->vui_ra_pages > ria->ria_start_idx)
730 ria->ria_end_idx_min =
731 vio->vui_ra_start_idx + vio->vui_ra_pages - 1;
733 ria->ria_reserved = ll_ra_count_get(ll_i2sbi(inode), ria, pages,
735 if (ria->ria_reserved < pages)
736 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
738 CDEBUG(D_READA, "reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
739 ria->ria_reserved, pages, pages_min,
740 atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
741 ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
743 ret = ll_read_ahead_pages(env, io, queue, ras, ria, &ra_end_idx);
745 if (ria->ria_reserved != 0)
746 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
748 if (ra_end_idx == end_idx && ra_end_idx == (kms >> PAGE_SHIFT))
749 ll_ra_stats_inc(inode, RA_STAT_EOF);
752 "ra_end_idx = %lu end_idx = %lu stride end = %lu pages = %d\n",
753 ra_end_idx, end_idx, ria->ria_end_idx, ret);
755 if (ra_end_idx != end_idx)
756 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
757 if (ra_end_idx > 0) {
758 /* update the ras so that the next read-ahead tries from
759 * where we left off. */
760 spin_lock(&ras->ras_lock);
761 ras->ras_next_readahead_idx = ra_end_idx + 1;
762 spin_unlock(&ras->ras_lock);
769 static void ras_set_start(struct ll_readahead_state *ras, pgoff_t index)
771 ras->ras_window_start_idx = ras_align(ras, index);
774 /* called with the ras_lock held or from places where it doesn't matter */
775 static void ras_reset(struct ll_readahead_state *ras, pgoff_t index)
777 ras->ras_consecutive_requests = 0;
778 ras->ras_consecutive_bytes = 0;
779 ras->ras_window_pages = 0;
780 ras_set_start(ras, index);
781 ras->ras_next_readahead_idx = max(ras->ras_window_start_idx, index + 1);
786 /* called with the ras_lock held or from places where it doesn't matter */
787 static void ras_stride_reset(struct ll_readahead_state *ras)
789 ras->ras_consecutive_stride_requests = 0;
790 ras->ras_stride_length = 0;
791 ras->ras_stride_bytes = 0;
795 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
797 spin_lock_init(&ras->ras_lock);
798 ras->ras_rpc_pages = PTLRPC_MAX_BRW_PAGES;
800 ras->ras_last_read_end_bytes = 0;
801 ras->ras_requests = 0;
805 * Check whether the read request is in the stride window.
806 * If it is in the stride window, return true, otherwise return false.
808 static bool read_in_stride_window(struct ll_readahead_state *ras,
809 loff_t pos, loff_t count)
813 if (ras->ras_stride_length == 0 || ras->ras_stride_bytes == 0 ||
814 ras->ras_stride_bytes == ras->ras_stride_length)
817 stride_gap = pos - ras->ras_last_read_end_bytes - 1;
819 /* If it is contiguous read */
821 return ras->ras_consecutive_bytes + count <=
822 ras->ras_stride_bytes;
824 /* Otherwise check the stride by itself */
825 return (ras->ras_stride_length - ras->ras_stride_bytes) == stride_gap &&
826 ras->ras_consecutive_bytes == ras->ras_stride_bytes &&
827 count <= ras->ras_stride_bytes;
830 static void ras_init_stride_detector(struct ll_readahead_state *ras,
831 loff_t pos, loff_t count)
833 loff_t stride_gap = pos - ras->ras_last_read_end_bytes - 1;
835 LASSERT(ras->ras_consecutive_stride_requests == 0);
837 if (pos <= ras->ras_last_read_end_bytes) {
838 /*Reset stride window for forward read*/
839 ras_stride_reset(ras);
843 ras->ras_stride_bytes = ras->ras_consecutive_bytes;
844 ras->ras_stride_length = stride_gap + ras->ras_consecutive_bytes;
845 ras->ras_consecutive_stride_requests++;
846 ras->ras_stride_offset = pos;
852 stride_page_count(struct ll_readahead_state *ras, loff_t len)
855 stride_byte_count(ras->ras_stride_offset,
856 ras->ras_stride_length, ras->ras_stride_bytes,
857 ras->ras_window_start_idx << PAGE_SHIFT, len);
859 return (bytes_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
862 /* Stride Read-ahead window will be increased inc_len according to
863 * stride I/O pattern */
864 static void ras_stride_increase_window(struct ll_readahead_state *ras,
865 struct ll_ra_info *ra, loff_t inc_bytes)
867 loff_t window_bytes, stride_bytes;
872 /* temporarily store in page units to reduce LASSERT() cost below */
873 end = ras->ras_window_start_idx + ras->ras_window_pages;
875 LASSERT(ras->ras_stride_length > 0);
876 LASSERTF(end >= (ras->ras_stride_offset >> PAGE_SHIFT),
877 "window_start_idx %lu, window_pages %lu stride_offset %llu\n",
878 ras->ras_window_start_idx, ras->ras_window_pages,
879 ras->ras_stride_offset);
882 if (end <= ras->ras_stride_offset)
885 stride_bytes = end - ras->ras_stride_offset;
887 div64_u64_rem(stride_bytes, ras->ras_stride_length, &left_bytes);
888 window_bytes = (ras->ras_window_pages << PAGE_SHIFT);
889 if (left_bytes < ras->ras_stride_bytes) {
890 if (ras->ras_stride_bytes - left_bytes >= inc_bytes) {
891 window_bytes += inc_bytes;
894 window_bytes += (ras->ras_stride_bytes - left_bytes);
895 inc_bytes -= (ras->ras_stride_bytes - left_bytes);
898 window_bytes += (ras->ras_stride_length - left_bytes);
901 LASSERT(ras->ras_stride_bytes != 0);
903 step = div64_u64_rem(inc_bytes, ras->ras_stride_bytes, &left_bytes);
905 window_bytes += step * ras->ras_stride_length + left_bytes;
906 LASSERT(window_bytes > 0);
909 if (stride_page_count(ras, window_bytes) <=
910 ra->ra_max_pages_per_file || ras->ras_window_pages == 0)
911 ras->ras_window_pages = (window_bytes >> PAGE_SHIFT);
913 LASSERT(ras->ras_window_pages > 0);
918 static void ras_increase_window(struct inode *inode,
919 struct ll_readahead_state *ras,
920 struct ll_ra_info *ra)
922 /* The stretch of ra-window should be aligned with max rpc_size
923 * but current clio architecture does not support retrieve such
924 * information from lower layer. FIXME later
926 if (stride_io_mode(ras)) {
927 ras_stride_increase_window(ras, ra,
928 (loff_t)ras->ras_rpc_pages << PAGE_SHIFT);
930 pgoff_t window_pages;
932 window_pages = min(ras->ras_window_pages + ras->ras_rpc_pages,
933 ra->ra_max_pages_per_file);
934 if (window_pages < ras->ras_rpc_pages)
935 ras->ras_window_pages = window_pages;
937 ras->ras_window_pages = ras_align(ras, window_pages);
942 * Seek within 8 pages are considered as sequential read for now.
944 static inline bool is_loose_seq_read(struct ll_readahead_state *ras, loff_t pos)
946 return pos_in_window(pos, ras->ras_last_read_end_bytes,
947 8UL << PAGE_SHIFT, 8UL << PAGE_SHIFT);
950 static void ras_detect_read_pattern(struct ll_readahead_state *ras,
951 struct ll_sb_info *sbi,
952 loff_t pos, size_t count, bool mmap)
954 bool stride_detect = false;
955 pgoff_t index = pos >> PAGE_SHIFT;
958 * Reset the read-ahead window in two cases. First when the app seeks
959 * or reads to some other part of the file. Secondly if we get a
960 * read-ahead miss that we think we've previously issued. This can
961 * be a symptom of there being so many read-ahead pages that the VM
962 * is reclaiming it before we get to it.
964 if (!is_loose_seq_read(ras, pos)) {
965 /* Check whether it is in stride I/O mode */
966 if (!read_in_stride_window(ras, pos, count)) {
967 if (ras->ras_consecutive_stride_requests == 0)
968 ras_init_stride_detector(ras, pos, count);
970 ras_stride_reset(ras);
971 ras->ras_consecutive_bytes = 0;
972 ras_reset(ras, index);
974 ras->ras_consecutive_bytes = 0;
975 ras->ras_consecutive_requests = 0;
976 if (++ras->ras_consecutive_stride_requests > 1)
977 stride_detect = true;
980 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
981 } else if (stride_io_mode(ras)) {
983 * If this is contiguous read but in stride I/O mode
984 * currently, check whether stride step still is valid,
985 * if invalid, it will reset the stride ra window to
988 if (!read_in_stride_window(ras, pos, count)) {
989 ras_stride_reset(ras);
990 ras->ras_window_pages = 0;
991 ras->ras_next_readahead_idx = index;
995 ras->ras_consecutive_bytes += count;
997 pgoff_t idx = ras->ras_consecutive_bytes >> PAGE_SHIFT;
999 if ((idx >= 4 && (idx & 3UL) == 0) || stride_detect)
1000 ras->ras_need_increase_window = true;
1001 } else if ((ras->ras_consecutive_requests > 1 || stride_detect)) {
1002 ras->ras_need_increase_window = true;
1005 ras->ras_last_read_end_bytes = pos + count - 1;
1008 void ll_ras_enter(struct file *f, loff_t pos, size_t count)
1010 struct ll_file_data *fd = f->private_data;
1011 struct ll_readahead_state *ras = &fd->fd_ras;
1012 struct inode *inode = file_inode(f);
1013 unsigned long index = pos >> PAGE_SHIFT;
1014 struct ll_sb_info *sbi = ll_i2sbi(inode);
1016 spin_lock(&ras->ras_lock);
1017 ras->ras_requests++;
1018 ras->ras_consecutive_requests++;
1019 ras->ras_need_increase_window = false;
1020 ras->ras_no_miss_check = false;
1022 * On the second access to a file smaller than the tunable
1023 * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1024 * file up to ra_max_pages_per_file. This is simply a best effort
1025 * and only occurs once per open file. Normal RA behavior is reverted
1026 * to for subsequent IO.
1028 if (ras->ras_requests >= 2) {
1030 struct ll_ra_info *ra = &sbi->ll_ra_info;
1032 kms_pages = (i_size_read(inode) + PAGE_SIZE - 1) >>
1035 CDEBUG(D_READA, "kmsp %llu mwp %lu mp %lu\n", kms_pages,
1036 ra->ra_max_read_ahead_whole_pages,
1037 ra->ra_max_pages_per_file);
1040 kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1041 ras->ras_window_start_idx = 0;
1042 ras->ras_next_readahead_idx = index + 1;
1043 ras->ras_window_pages = min(ra->ra_max_pages_per_file,
1044 ra->ra_max_read_ahead_whole_pages);
1045 ras->ras_no_miss_check = true;
1046 GOTO(out_unlock, 0);
1049 ras_detect_read_pattern(ras, sbi, pos, count, false);
1051 spin_unlock(&ras->ras_lock);
1054 static bool index_in_stride_window(struct ll_readahead_state *ras,
1057 loff_t pos = (loff_t)index << PAGE_SHIFT;
1059 if (ras->ras_stride_length == 0 || ras->ras_stride_bytes == 0 ||
1060 ras->ras_stride_bytes == ras->ras_stride_length)
1063 if (pos >= ras->ras_stride_offset) {
1066 div64_u64_rem(pos - ras->ras_stride_offset,
1067 ras->ras_stride_length, &offset);
1068 if (offset < ras->ras_stride_bytes ||
1069 ras->ras_stride_length - offset < PAGE_SIZE)
1071 } else if (ras->ras_stride_offset - pos < PAGE_SIZE) {
1079 * ll_ras_enter() is used to detect read pattern according to pos and count.
1081 * ras_update() is used to detect cache miss and
1082 * reset window or increase window accordingly
1084 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
1085 struct ll_readahead_state *ras, pgoff_t index,
1086 enum ras_update_flags flags)
1088 struct ll_ra_info *ra = &sbi->ll_ra_info;
1089 bool hit = flags & LL_RAS_HIT;
1092 spin_lock(&ras->ras_lock);
1095 CDEBUG(D_READA, DFID " pages at %lu miss.\n",
1096 PFID(ll_inode2fid(inode)), index);
1097 ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
1100 * The readahead window has been expanded to cover whole
1101 * file size, we don't care whether ra miss happen or not.
1102 * Because we will read whole file to page cache even if
1103 * some pages missed.
1105 if (ras->ras_no_miss_check)
1106 GOTO(out_unlock, 0);
1108 if (flags & LL_RAS_MMAP)
1109 ras_detect_read_pattern(ras, sbi, (loff_t)index << PAGE_SHIFT,
1112 if (!hit && ras->ras_window_pages &&
1113 index < ras->ras_next_readahead_idx &&
1114 pos_in_window(index, ras->ras_window_start_idx, 0,
1115 ras->ras_window_pages)) {
1116 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
1117 ras->ras_need_increase_window = false;
1119 if (index_in_stride_window(ras, index) &&
1120 stride_io_mode(ras)) {
1122 * if (index != ras->ras_last_readpage + 1)
1123 * ras->ras_consecutive_pages = 0;
1125 ras_reset(ras, index);
1128 * If stride-RA hit cache miss, the stride
1129 * detector will not be reset to avoid the
1130 * overhead of redetecting read-ahead mode,
1131 * but on the condition that the stride window
1132 * is still intersect with normal sequential
1133 * read-ahead window.
1135 if (ras->ras_window_start_idx < ras->ras_stride_offset)
1136 ras_stride_reset(ras);
1140 * Reset both stride window and normal RA
1143 ras_reset(ras, index);
1144 /* ras->ras_consecutive_pages++; */
1145 ras->ras_consecutive_bytes = 0;
1146 ras_stride_reset(ras);
1147 GOTO(out_unlock, 0);
1150 ras_set_start(ras, index);
1152 if (stride_io_mode(ras)) {
1153 /* Since stride readahead is sentivite to the offset
1154 * of read-ahead, so we use original offset here,
1155 * instead of ras_window_start_idx, which is RPC aligned.
1157 ras->ras_next_readahead_idx = max(index + 1,
1158 ras->ras_next_readahead_idx);
1159 ras->ras_window_start_idx =
1160 max_t(pgoff_t, ras->ras_window_start_idx,
1161 ras->ras_stride_offset >> PAGE_SHIFT);
1163 if (ras->ras_next_readahead_idx < ras->ras_window_start_idx)
1164 ras->ras_next_readahead_idx = ras->ras_window_start_idx;
1166 ras->ras_next_readahead_idx = index + 1;
1169 if (ras->ras_need_increase_window) {
1170 ras_increase_window(inode, ras, ra);
1171 ras->ras_need_increase_window = false;
1176 spin_unlock(&ras->ras_lock);
1179 int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
1181 struct inode *inode = vmpage->mapping->host;
1182 struct ll_inode_info *lli = ll_i2info(inode);
1185 struct cl_page *page;
1186 struct cl_object *clob;
1187 bool redirtied = false;
1188 bool unlocked = false;
1193 LASSERT(PageLocked(vmpage));
1194 LASSERT(!PageWriteback(vmpage));
1196 LASSERT(ll_i2dtexp(inode) != NULL);
1198 env = cl_env_get(&refcheck);
1200 GOTO(out, result = PTR_ERR(env));
1202 clob = ll_i2info(inode)->lli_clob;
1203 LASSERT(clob != NULL);
1205 io = vvp_env_thread_io(env);
1207 io->ci_ignore_layout = 1;
1208 result = cl_io_init(env, io, CIT_MISC, clob);
1210 page = cl_page_find(env, clob, vmpage->index,
1211 vmpage, CPT_CACHEABLE);
1212 if (!IS_ERR(page)) {
1213 lu_ref_add(&page->cp_reference, "writepage",
1215 cl_page_assume(env, io, page);
1216 result = cl_page_flush(env, io, page);
1219 * Re-dirty page on error so it retries write,
1220 * but not in case when IO has actually
1221 * occurred and completed with an error.
1223 if (!PageError(vmpage)) {
1224 redirty_page_for_writepage(wbc, vmpage);
1229 cl_page_disown(env, io, page);
1231 lu_ref_del(&page->cp_reference,
1232 "writepage", current);
1233 cl_page_put(env, page);
1235 result = PTR_ERR(page);
1238 cl_io_fini(env, io);
1240 if (redirtied && wbc->sync_mode == WB_SYNC_ALL) {
1241 loff_t offset = cl_offset(clob, vmpage->index);
1243 /* Flush page failed because the extent is being written out.
1244 * Wait for the write of extent to be finished to avoid
1245 * breaking kernel which assumes ->writepage should mark
1246 * PageWriteback or clean the page. */
1247 result = cl_sync_file_range(inode, offset,
1248 offset + PAGE_SIZE - 1,
1251 /* actually we may have written more than one page.
1252 * decreasing this page because the caller will count
1254 wbc->nr_to_write -= result - 1;
1259 cl_env_put(env, &refcheck);
1264 if (!lli->lli_async_rc)
1265 lli->lli_async_rc = result;
1266 SetPageError(vmpage);
1268 unlock_page(vmpage);
1273 int ll_writepages(struct address_space *mapping, struct writeback_control *wbc)
1275 struct inode *inode = mapping->host;
1278 enum cl_fsync_mode mode;
1279 int range_whole = 0;
1283 if (wbc->range_cyclic) {
1284 start = (loff_t)mapping->writeback_index << PAGE_SHIFT;
1285 end = OBD_OBJECT_EOF;
1287 start = wbc->range_start;
1288 end = wbc->range_end;
1289 if (end == LLONG_MAX) {
1290 end = OBD_OBJECT_EOF;
1291 range_whole = start == 0;
1295 mode = CL_FSYNC_NONE;
1296 if (wbc->sync_mode == WB_SYNC_ALL)
1297 mode = CL_FSYNC_LOCAL;
1299 if (ll_i2info(inode)->lli_clob == NULL)
1302 /* for directio, it would call writepages() to evict cached pages
1303 * inside the IO context of write, which will cause deadlock at
1304 * layout_conf since it waits for active IOs to complete. */
1305 result = cl_sync_file_range(inode, start, end, mode, 1);
1307 wbc->nr_to_write -= result;
1311 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) {
1312 if (end == OBD_OBJECT_EOF)
1313 mapping->writeback_index = 0;
1315 mapping->writeback_index = (end >> PAGE_SHIFT) + 1;
1320 struct ll_cl_context *ll_cl_find(struct file *file)
1322 struct ll_file_data *fd = file->private_data;
1323 struct ll_cl_context *lcc;
1324 struct ll_cl_context *found = NULL;
1326 read_lock(&fd->fd_lock);
1327 list_for_each_entry(lcc, &fd->fd_lccs, lcc_list) {
1328 if (lcc->lcc_cookie == current) {
1333 read_unlock(&fd->fd_lock);
1338 void ll_cl_add(struct file *file, const struct lu_env *env, struct cl_io *io,
1341 struct ll_file_data *fd = file->private_data;
1342 struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1344 memset(lcc, 0, sizeof(*lcc));
1345 INIT_LIST_HEAD(&lcc->lcc_list);
1346 lcc->lcc_cookie = current;
1349 lcc->lcc_type = type;
1351 write_lock(&fd->fd_lock);
1352 list_add(&lcc->lcc_list, &fd->fd_lccs);
1353 write_unlock(&fd->fd_lock);
1356 void ll_cl_remove(struct file *file, const struct lu_env *env)
1358 struct ll_file_data *fd = file->private_data;
1359 struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1361 write_lock(&fd->fd_lock);
1362 list_del_init(&lcc->lcc_list);
1363 write_unlock(&fd->fd_lock);
1366 int ll_io_read_page(const struct lu_env *env, struct cl_io *io,
1367 struct cl_page *page, struct file *file)
1369 struct inode *inode = vvp_object_inode(page->cp_obj);
1370 struct ll_sb_info *sbi = ll_i2sbi(inode);
1371 struct ll_file_data *fd = file->private_data;
1372 struct ll_readahead_state *ras = &fd->fd_ras;
1373 struct cl_2queue *queue = &io->ci_queue;
1374 struct cl_sync_io *anchor = NULL;
1375 struct vvp_page *vpg;
1380 vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1381 uptodate = vpg->vpg_defer_uptodate;
1383 if (sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
1384 sbi->ll_ra_info.ra_max_pages > 0 &&
1385 !vpg->vpg_ra_updated) {
1386 struct vvp_io *vio = vvp_env_io(env);
1387 enum ras_update_flags flags = 0;
1390 flags |= LL_RAS_HIT;
1391 if (!vio->vui_ra_valid)
1392 flags |= LL_RAS_MMAP;
1393 ras_update(sbi, inode, ras, vvp_index(vpg), flags);
1396 cl_2queue_init(queue);
1398 vpg->vpg_ra_used = 1;
1399 cl_page_export(env, page, 1);
1400 cl_page_disown(env, io, page);
1402 anchor = &vvp_env_info(env)->vti_anchor;
1403 cl_sync_io_init(anchor, 1);
1404 page->cp_sync_io = anchor;
1406 cl_2queue_add(queue, page);
1409 if (sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
1410 sbi->ll_ra_info.ra_max_pages > 0) {
1413 rc2 = ll_readahead(env, io, &queue->c2_qin, ras,
1415 CDEBUG(D_READA, DFID "%d pages read ahead at %lu\n",
1416 PFID(ll_inode2fid(inode)), rc2, vvp_index(vpg));
1419 if (queue->c2_qin.pl_nr > 0) {
1420 int count = queue->c2_qin.pl_nr;
1421 rc = cl_io_submit_rw(env, io, CRT_READ, queue);
1423 task_io_account_read(PAGE_SIZE * count);
1427 if (anchor != NULL && !cl_page_is_owned(page, io)) { /* have sent */
1428 rc = cl_sync_io_wait(env, anchor, 0);
1430 cl_page_assume(env, io, page);
1431 cl_page_list_del(env, &queue->c2_qout, page);
1433 if (!PageUptodate(cl_page_vmpage(page))) {
1434 /* Failed to read a mirror, discard this page so that
1435 * new page can be created with new mirror.
1437 * TODO: this is not needed after page reinit
1438 * route is implemented */
1439 cl_page_discard(env, io, page);
1441 cl_page_disown(env, io, page);
1444 /* TODO: discard all pages until page reinit route is implemented */
1445 cl_page_list_discard(env, io, &queue->c2_qin);
1447 /* Unlock unsent read pages in case of error. */
1448 cl_page_list_disown(env, io, &queue->c2_qin);
1450 cl_2queue_fini(env, queue);
1456 * Possible return value:
1457 * 0 no async readahead triggered and fast read could not be used.
1458 * 1 no async readahead, but fast read could be used.
1459 * 2 async readahead triggered and fast read could be used too.
1462 static int kickoff_async_readahead(struct file *file, unsigned long pages)
1464 struct ll_readahead_work *lrw;
1465 struct inode *inode = file_inode(file);
1466 struct ll_sb_info *sbi = ll_i2sbi(inode);
1467 struct ll_file_data *fd = file->private_data;
1468 struct ll_readahead_state *ras = &fd->fd_ras;
1469 struct ll_ra_info *ra = &sbi->ll_ra_info;
1470 unsigned long throttle;
1471 pgoff_t start_idx = ras_align(ras, ras->ras_next_readahead_idx);
1472 pgoff_t end_idx = start_idx + pages - 1;
1474 throttle = min(ra->ra_async_pages_per_file_threshold,
1475 ra->ra_max_pages_per_file);
1477 * If this is strided i/o or the window is smaller than the
1478 * throttle limit, we do not do async readahead. Otherwise,
1479 * we do async readahead, allowing the user thread to do fast i/o.
1481 if (stride_io_mode(ras) || !throttle ||
1482 ras->ras_window_pages < throttle)
1485 if ((atomic_read(&ra->ra_cur_pages) + pages) > ra->ra_max_pages)
1488 if (ras->ras_async_last_readpage_idx == start_idx)
1491 /* ll_readahead_work_free() free it */
1494 lrw->lrw_file = get_file(file);
1495 lrw->lrw_start_idx = start_idx;
1496 lrw->lrw_end_idx = end_idx;
1497 spin_lock(&ras->ras_lock);
1498 ras->ras_next_readahead_idx = end_idx + 1;
1499 ras->ras_async_last_readpage_idx = start_idx;
1500 spin_unlock(&ras->ras_lock);
1501 ll_readahead_work_add(inode, lrw);
1510 * Check if we can issue a readahead RPC, if that is
1511 * the case, we can't do fast IO because we will need
1512 * a cl_io to issue the RPC.
1514 static bool ll_use_fast_io(struct file *file,
1515 struct ll_readahead_state *ras, pgoff_t index)
1517 unsigned long fast_read_pages =
1518 max(RA_REMAIN_WINDOW_MIN, ras->ras_rpc_pages);
1521 if (stride_io_mode(ras)) {
1522 skip_pages = (ras->ras_stride_length +
1523 ras->ras_stride_bytes - 1) / ras->ras_stride_bytes;
1524 skip_pages *= fast_read_pages;
1526 skip_pages = fast_read_pages;
1529 if (ras->ras_window_start_idx + ras->ras_window_pages <
1530 ras->ras_next_readahead_idx + skip_pages ||
1531 kickoff_async_readahead(file, fast_read_pages) > 0)
1537 int ll_readpage(struct file *file, struct page *vmpage)
1539 struct inode *inode = file_inode(file);
1540 struct cl_object *clob = ll_i2info(inode)->lli_clob;
1541 struct ll_cl_context *lcc;
1542 const struct lu_env *env = NULL;
1543 struct cl_io *io = NULL;
1544 struct cl_page *page;
1545 struct ll_sb_info *sbi = ll_i2sbi(inode);
1549 lcc = ll_cl_find(file);
1555 if (io == NULL) { /* fast read */
1556 struct inode *inode = file_inode(file);
1557 struct ll_file_data *fd = file->private_data;
1558 struct ll_readahead_state *ras = &fd->fd_ras;
1559 struct lu_env *local_env = NULL;
1560 struct vvp_page *vpg;
1564 /* TODO: need to verify the layout version to make sure
1565 * the page is not invalid due to layout change. */
1566 page = cl_vmpage_page(vmpage, clob);
1568 unlock_page(vmpage);
1569 ll_ra_stats_inc_sbi(sbi, RA_STAT_FAILED_FAST_READ);
1573 vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1574 if (vpg->vpg_defer_uptodate) {
1575 enum ras_update_flags flags = LL_RAS_HIT;
1577 if (lcc && lcc->lcc_type == LCC_MMAP)
1578 flags |= LL_RAS_MMAP;
1580 /* For fast read, it updates read ahead state only
1581 * if the page is hit in cache because non cache page
1582 * case will be handled by slow read later. */
1583 ras_update(sbi, inode, ras, vvp_index(vpg), flags);
1584 /* avoid duplicate ras_update() call */
1585 vpg->vpg_ra_updated = 1;
1587 if (ll_use_fast_io(file, ras, vvp_index(vpg)))
1592 local_env = cl_env_percpu_get();
1596 /* export the page and skip io stack */
1598 vpg->vpg_ra_used = 1;
1599 cl_page_export(env, page, 1);
1601 ll_ra_stats_inc_sbi(sbi, RA_STAT_FAILED_FAST_READ);
1603 /* release page refcount before unlocking the page to ensure
1604 * the object won't be destroyed in the calling path of
1605 * cl_page_put(). Please see comment in ll_releasepage(). */
1606 cl_page_put(env, page);
1607 unlock_page(vmpage);
1609 cl_env_percpu_put(local_env);
1615 * Direct read can fall back to buffered read, but DIO is done
1616 * with lockless i/o, and buffered requires LDLM locking, so in
1617 * this case we must restart without lockless.
1619 if (file->f_flags & O_DIRECT &&
1620 lcc && lcc->lcc_type == LCC_RW &&
1621 !io->ci_ignore_lockless) {
1622 unlock_page(vmpage);
1623 io->ci_ignore_lockless = 1;
1624 io->ci_need_restart = 1;
1628 LASSERT(io->ci_state == CIS_IO_GOING);
1629 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
1630 if (!IS_ERR(page)) {
1631 LASSERT(page->cp_type == CPT_CACHEABLE);
1632 if (likely(!PageUptodate(vmpage))) {
1633 cl_page_assume(env, io, page);
1635 result = ll_io_read_page(env, io, page, file);
1637 /* Page from a non-object file. */
1638 unlock_page(vmpage);
1641 cl_page_put(env, page);
1643 unlock_page(vmpage);
1644 result = PTR_ERR(page);