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 static inline bool ll_readahead_enabled(struct ll_sb_info *sbi)
128 return sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
129 sbi->ll_ra_info.ra_max_pages > 0;
132 void ll_ra_stats_inc(struct inode *inode, enum ra_stat which)
134 struct ll_sb_info *sbi = ll_i2sbi(inode);
135 ll_ra_stats_inc_sbi(sbi, which);
138 #define RAS_CDEBUG(ras) \
140 "lre %llu cr %lu cb %llu wsi %lu wp %lu nra %lu rpc %lu " \
141 "r %lu csr %lu so %llu sb %llu sl %llu lr %lu\n", \
142 ras->ras_last_read_end_bytes, ras->ras_consecutive_requests, \
143 ras->ras_consecutive_bytes, ras->ras_window_start_idx, \
144 ras->ras_window_pages, ras->ras_next_readahead_idx, \
145 ras->ras_rpc_pages, ras->ras_requests, \
146 ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
147 ras->ras_stride_bytes, ras->ras_stride_length, \
148 ras->ras_async_last_readpage_idx)
150 static bool pos_in_window(loff_t pos, loff_t point,
151 unsigned long before, unsigned long after)
153 loff_t start = point - before;
154 loff_t end = point + after;
161 return start <= pos && pos <= end;
164 enum ll_ra_page_hint {
165 MAYNEED = 0, /* this page possibly accessed soon */
166 WILLNEED /* this page is gurateed to be needed */
170 * Initiates read-ahead of a page with given index.
172 * \retval +ve: page was already uptodate so it will be skipped
174 * \retval -ve: page wasn't added to \a queue for error;
175 * \retval 0: page was added into \a queue for read ahead.
177 static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io,
178 struct cl_page_list *queue, pgoff_t index,
179 enum ll_ra_page_hint hint)
181 struct cl_object *clob = io->ci_obj;
182 struct inode *inode = vvp_object_inode(clob);
183 struct page *vmpage = NULL;
184 struct cl_page *page;
185 struct vvp_page *vpg;
186 enum ra_stat which = _NR_RA_STAT; /* keep gcc happy */
188 const char *msg = NULL;
194 vmpage = grab_cache_page_nowait(inode->i_mapping, index);
195 if (vmpage == NULL) {
196 which = RA_STAT_FAILED_GRAB_PAGE;
197 msg = "g_c_p_n failed";
198 GOTO(out, rc = -EBUSY);
202 vmpage = find_or_create_page(inode->i_mapping, index,
205 GOTO(out, rc = -ENOMEM);
208 /* should not come here */
209 GOTO(out, rc = -EINVAL);
212 /* Check if vmpage was truncated or reclaimed */
213 if (vmpage->mapping != inode->i_mapping) {
214 which = RA_STAT_WRONG_GRAB_PAGE;
215 msg = "g_c_p_n returned invalid page";
216 GOTO(out, rc = -EBUSY);
219 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
221 which = RA_STAT_FAILED_GRAB_PAGE;
222 msg = "cl_page_find failed";
223 GOTO(out, rc = PTR_ERR(page));
226 lu_ref_add(&page->cp_reference, "ra", current);
227 cl_page_assume(env, io, page);
228 vpg = cl2vvp_page(cl_object_page_slice(clob, page));
229 if (!vpg->vpg_defer_uptodate && !PageUptodate(vmpage)) {
230 vpg->vpg_defer_uptodate = 1;
231 vpg->vpg_ra_used = 0;
232 cl_page_list_add(queue, page);
234 /* skip completed pages */
235 cl_page_unassume(env, io, page);
236 /* This page is already uptodate, returning a positive number
237 * to tell the callers about this */
241 lu_ref_del(&page->cp_reference, "ra", current);
242 cl_page_put(env, page);
245 if (vmpage != NULL) {
250 if (msg != NULL && hint == MAYNEED) {
251 ll_ra_stats_inc(inode, which);
252 CDEBUG(D_READA, "%s\n", msg);
259 #define RIA_DEBUG(ria) \
260 CDEBUG(D_READA, "rs %lu re %lu ro %llu rl %llu rb %llu\n", \
261 ria->ria_start_idx, ria->ria_end_idx, ria->ria_stoff, \
262 ria->ria_length, ria->ria_bytes)
264 static inline int stride_io_mode(struct ll_readahead_state *ras)
266 return ras->ras_consecutive_stride_requests > 1;
269 /* The function calculates how many bytes will be read in
270 * [off, off + length], in such stride IO area,
271 * stride_offset = st_off, stride_lengh = st_len,
272 * stride_bytes = st_bytes
274 * |------------------|*****|------------------|*****|------------|*****|....
277 * |----- st_len -----|
279 * How many bytes it should read in such pattern
280 * |-------------------------------------------------------------|
282 * |<------ length ------->|
284 * = |<----->| + |-------------------------------------| + |---|
285 * start_left st_bytes * i end_left
287 static loff_t stride_byte_count(loff_t st_off, loff_t st_len, loff_t st_bytes,
288 loff_t off, loff_t length)
290 u64 start = off > st_off ? off - st_off : 0;
291 u64 end = off + length > st_off ? off + length - st_off : 0;
296 if (st_len == 0 || length == 0 || end == 0)
299 start = div64_u64_rem(start, st_len, &start_left);
300 if (start_left < st_bytes)
301 start_left = st_bytes - start_left;
305 end = div64_u64_rem(end, st_len, &end_left);
306 if (end_left > st_bytes)
309 CDEBUG(D_READA, "start %llu, end %llu start_left %llu end_left %llu\n",
310 start, end, start_left, end_left);
313 bytes_count = end_left - (st_bytes - start_left);
315 bytes_count = start_left +
316 st_bytes * (end - start - 1) + end_left;
319 "st_off %llu, st_len %llu st_bytes %llu off %llu length %llu bytescount %llu\n",
320 st_off, st_len, st_bytes, off, length, bytes_count);
325 static unsigned long ria_page_count(struct ra_io_arg *ria)
327 loff_t length_bytes = ria->ria_end_idx >= ria->ria_start_idx ?
328 (loff_t)(ria->ria_end_idx -
329 ria->ria_start_idx + 1) << PAGE_SHIFT : 0;
332 if (ria->ria_length > ria->ria_bytes && ria->ria_bytes &&
333 (ria->ria_length & ~PAGE_MASK || ria->ria_bytes & ~PAGE_MASK ||
334 ria->ria_stoff & ~PAGE_MASK)) {
335 /* Over-estimate un-aligned page stride read */
336 unsigned long pg_count = ((ria->ria_bytes +
337 PAGE_SIZE - 1) >> PAGE_SHIFT) + 1;
338 pg_count *= length_bytes / ria->ria_length + 1;
342 bytes_count = stride_byte_count(ria->ria_stoff, ria->ria_length,
344 (loff_t)ria->ria_start_idx<<PAGE_SHIFT,
346 return (bytes_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
349 static pgoff_t ras_align(struct ll_readahead_state *ras, pgoff_t index)
351 return index - (index % ras->ras_rpc_pages);
354 /* Check whether the index is in the defined ra-window */
355 static bool ras_inside_ra_window(pgoff_t idx, struct ra_io_arg *ria)
357 loff_t pos = (loff_t)idx << PAGE_SHIFT;
359 /* If ria_length == ria_bytes, it means non-stride I/O mode,
360 * idx should always inside read-ahead window in this case
361 * For stride I/O mode, just check whether the idx is inside
364 if (ria->ria_length == 0 || ria->ria_length == ria->ria_bytes)
367 if (pos >= ria->ria_stoff) {
370 div64_u64_rem(pos - ria->ria_stoff, ria->ria_length, &offset);
372 if (offset < ria->ria_bytes ||
373 (ria->ria_length - offset) < PAGE_SIZE)
375 } else if (pos + PAGE_SIZE > ria->ria_stoff) {
383 ll_read_ahead_pages(const struct lu_env *env, struct cl_io *io,
384 struct cl_page_list *queue, struct ll_readahead_state *ras,
385 struct ra_io_arg *ria, pgoff_t *ra_end)
387 struct cl_read_ahead ra = { 0 };
388 /* busy page count is per stride */
389 int rc = 0, count = 0, busy_page_count = 0;
392 LASSERT(ria != NULL);
395 for (page_idx = ria->ria_start_idx;
396 page_idx <= ria->ria_end_idx && ria->ria_reserved > 0;
398 if (ras_inside_ra_window(page_idx, ria)) {
399 if (ra.cra_end_idx == 0 || ra.cra_end_idx < page_idx) {
403 * Do not shrink ria_end_idx at any case until
404 * the minimum end of current read is covered.
406 * Do not extend read lock accross stripe if
407 * lock contention detected.
409 if (ra.cra_contention &&
410 page_idx > ria->ria_end_idx_min) {
411 ria->ria_end_idx = *ra_end;
415 cl_read_ahead_release(env, &ra);
417 rc = cl_io_read_ahead(env, io, page_idx, &ra);
422 * Only shrink ria_end_idx if the matched
423 * LDLM lock doesn't cover more.
425 if (page_idx > ra.cra_end_idx) {
426 ria->ria_end_idx = ra.cra_end_idx;
430 CDEBUG(D_READA, "idx: %lu, ra: %lu, rpc: %lu\n",
431 page_idx, ra.cra_end_idx,
433 LASSERTF(ra.cra_end_idx >= page_idx,
434 "object: %p, indcies %lu / %lu\n",
435 io->ci_obj, ra.cra_end_idx, page_idx);
436 /* update read ahead RPC size.
437 * NB: it's racy but doesn't matter */
438 if (ras->ras_rpc_pages != ra.cra_rpc_pages &&
439 ra.cra_rpc_pages > 0)
440 ras->ras_rpc_pages = ra.cra_rpc_pages;
441 /* trim it to align with optimal RPC size */
442 end_idx = ras_align(ras, ria->ria_end_idx + 1);
443 if (end_idx > 0 && !ria->ria_eof)
444 ria->ria_end_idx = end_idx - 1;
445 if (ria->ria_end_idx < ria->ria_end_idx_min)
446 ria->ria_end_idx = ria->ria_end_idx_min;
448 if (page_idx > ria->ria_end_idx)
451 /* If the page is inside the read-ahead window */
452 rc = ll_read_ahead_page(env, io, queue, page_idx,
454 if (rc < 0 && rc != -EBUSY)
459 "skip busy page: %lu\n", page_idx);
460 /* For page unaligned readahead the first
461 * last pages of each region can be read by
462 * another reader on the same node, and so
463 * may be busy. So only stop for > 2 busy
465 if (busy_page_count > 2)
470 /* Only subtract from reserve & count the page if we
471 * really did readahead on that page. */
476 } else if (stride_io_mode(ras)) {
477 /* If it is not in the read-ahead window, and it is
478 * read-ahead mode, then check whether it should skip
481 loff_t pos = (loff_t)page_idx << PAGE_SHIFT;
484 div64_u64_rem(pos - ria->ria_stoff, ria->ria_length,
486 if (offset >= ria->ria_bytes) {
487 pos += (ria->ria_length - offset);
488 if ((pos >> PAGE_SHIFT) >= page_idx + 1)
489 page_idx = (pos >> PAGE_SHIFT) - 1;
492 "Stride: jump %llu pages to %lu\n",
493 ria->ria_length - offset, page_idx);
499 cl_read_ahead_release(env, &ra);
504 static void ll_readahead_work_free(struct ll_readahead_work *work)
506 fput(work->lrw_file);
510 static void ll_readahead_handle_work(struct work_struct *wq);
511 static void ll_readahead_work_add(struct inode *inode,
512 struct ll_readahead_work *work)
514 INIT_WORK(&work->lrw_readahead_work, ll_readahead_handle_work);
515 queue_work(ll_i2sbi(inode)->ll_ra_info.ll_readahead_wq,
516 &work->lrw_readahead_work);
519 static int ll_readahead_file_kms(const struct lu_env *env,
520 struct cl_io *io, __u64 *kms)
522 struct cl_object *clob;
524 struct cl_attr *attr = vvp_env_thread_attr(env);
528 inode = vvp_object_inode(clob);
530 cl_object_attr_lock(clob);
531 ret = cl_object_attr_get(env, clob, attr);
532 cl_object_attr_unlock(clob);
537 *kms = attr->cat_kms;
541 static void ll_readahead_handle_work(struct work_struct *wq)
543 struct ll_readahead_work *work;
546 struct ra_io_arg *ria;
548 struct ll_file_data *fd;
549 struct ll_readahead_state *ras;
551 struct cl_2queue *queue;
552 pgoff_t ra_end_idx = 0;
553 unsigned long pages, pages_min = 0;
558 struct ll_sb_info *sbi;
560 work = container_of(wq, struct ll_readahead_work,
562 fd = work->lrw_file->private_data;
564 file = work->lrw_file;
565 inode = file_inode(file);
566 sbi = ll_i2sbi(inode);
568 env = cl_env_alloc(&refcheck, LCT_NOREF);
570 GOTO(out_free_work, rc = PTR_ERR(env));
572 io = vvp_env_thread_io(env);
573 ll_io_init(io, file, CIT_READ, NULL);
575 rc = ll_readahead_file_kms(env, io, &kms);
577 GOTO(out_put_env, rc);
580 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
581 GOTO(out_put_env, rc = 0);
584 ria = &ll_env_info(env)->lti_ria;
585 memset(ria, 0, sizeof(*ria));
587 ria->ria_start_idx = work->lrw_start_idx;
588 /* Truncate RA window to end of file */
589 eof_index = (pgoff_t)(kms - 1) >> PAGE_SHIFT;
590 if (eof_index <= work->lrw_end_idx) {
591 work->lrw_end_idx = eof_index;
594 if (work->lrw_end_idx <= work->lrw_start_idx)
595 GOTO(out_put_env, rc = 0);
597 ria->ria_end_idx = work->lrw_end_idx;
598 pages = ria->ria_end_idx - ria->ria_start_idx + 1;
599 ria->ria_reserved = ll_ra_count_get(sbi, ria,
600 ria_page_count(ria), pages_min);
603 "async reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
604 ria->ria_reserved, pages, pages_min,
605 atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
606 ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
608 if (ria->ria_reserved < pages) {
609 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
610 if (PAGES_TO_MiB(ria->ria_reserved) < 1) {
611 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
612 GOTO(out_put_env, rc = 0);
616 rc = cl_io_rw_init(env, io, CIT_READ, ria->ria_start_idx, pages);
618 GOTO(out_put_env, rc);
620 vvp_env_io(env)->vui_io_subtype = IO_NORMAL;
621 vvp_env_io(env)->vui_fd = fd;
622 io->ci_state = CIS_LOCKED;
623 io->ci_async_readahead = true;
624 rc = cl_io_start(env, io);
626 GOTO(out_io_fini, rc);
628 queue = &io->ci_queue;
629 cl_2queue_init(queue);
631 rc = ll_read_ahead_pages(env, io, &queue->c2_qin, ras, ria,
633 if (ria->ria_reserved != 0)
634 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
635 if (queue->c2_qin.pl_nr > 0) {
636 int count = queue->c2_qin.pl_nr;
638 rc = cl_io_submit_rw(env, io, CRT_READ, queue);
640 task_io_account_read(PAGE_SIZE * count);
642 if (ria->ria_end_idx == ra_end_idx && ra_end_idx == (kms >> PAGE_SHIFT))
643 ll_ra_stats_inc(inode, RA_STAT_EOF);
645 if (ra_end_idx != ria->ria_end_idx)
646 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
648 /* TODO: discard all pages until page reinit route is implemented */
649 cl_page_list_discard(env, io, &queue->c2_qin);
651 /* Unlock unsent read pages in case of error. */
652 cl_page_list_disown(env, io, &queue->c2_qin);
654 cl_2queue_fini(env, queue);
659 cl_env_put(env, &refcheck);
662 ll_ra_stats_inc_sbi(ll_i2sbi(inode), RA_STAT_ASYNC);
663 atomic_dec(&sbi->ll_ra_info.ra_async_inflight);
664 ll_readahead_work_free(work);
667 static int ll_readahead(const struct lu_env *env, struct cl_io *io,
668 struct cl_page_list *queue,
669 struct ll_readahead_state *ras, bool hit,
672 struct vvp_io *vio = vvp_env_io(env);
673 struct ll_thread_info *lti = ll_env_info(env);
674 unsigned long pages, pages_min = 0;
675 pgoff_t ra_end_idx = 0, start_idx = 0, end_idx = 0;
677 struct ra_io_arg *ria = <i->lti_ria;
678 struct cl_object *clob;
684 inode = vvp_object_inode(clob);
686 memset(ria, 0, sizeof(*ria));
687 ret = ll_readahead_file_kms(env, io, &kms);
692 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
696 spin_lock(&ras->ras_lock);
699 * Note: other thread might rollback the ras_next_readahead_idx,
700 * if it can not get the full size of prepared pages, see the
701 * end of this function. For stride read ahead, it needs to
702 * make sure the offset is no less than ras_stride_offset,
703 * so that stride read ahead can work correctly.
705 if (stride_io_mode(ras))
706 start_idx = max_t(pgoff_t, ras->ras_next_readahead_idx,
707 ras->ras_stride_offset >> PAGE_SHIFT);
709 start_idx = ras->ras_next_readahead_idx;
711 if (ras->ras_window_pages > 0)
712 end_idx = ras->ras_window_start_idx + ras->ras_window_pages - 1;
714 /* Enlarge the RA window to encompass the full read */
715 if (vio->vui_ra_valid &&
716 end_idx < vio->vui_ra_start_idx + vio->vui_ra_pages - 1)
717 end_idx = vio->vui_ra_start_idx + vio->vui_ra_pages - 1;
722 /* Truncate RA window to end of file */
723 eof_index = (pgoff_t)((kms - 1) >> PAGE_SHIFT);
724 if (eof_index <= end_idx) {
729 ria->ria_start_idx = start_idx;
730 ria->ria_end_idx = end_idx;
731 /* If stride I/O mode is detected, get stride window*/
732 if (stride_io_mode(ras)) {
733 ria->ria_stoff = ras->ras_stride_offset;
734 ria->ria_length = ras->ras_stride_length;
735 ria->ria_bytes = ras->ras_stride_bytes;
737 spin_unlock(&ras->ras_lock);
740 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
743 pages = ria_page_count(ria);
745 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
750 CDEBUG(D_READA, DFID": ria: %lu/%lu, bead: %lu/%lu, hit: %d\n",
751 PFID(lu_object_fid(&clob->co_lu)),
752 ria->ria_start_idx, ria->ria_end_idx,
753 vio->vui_ra_valid ? vio->vui_ra_start_idx : 0,
754 vio->vui_ra_valid ? vio->vui_ra_pages : 0,
757 /* at least to extend the readahead window to cover current read */
758 if (!hit && vio->vui_ra_valid &&
759 vio->vui_ra_start_idx + vio->vui_ra_pages > ria->ria_start_idx) {
760 ria->ria_end_idx_min =
761 vio->vui_ra_start_idx + vio->vui_ra_pages - 1;
762 pages_min = vio->vui_ra_start_idx + vio->vui_ra_pages -
766 ria->ria_reserved = ll_ra_count_get(ll_i2sbi(inode), ria, pages,
768 if (ria->ria_reserved < pages)
769 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
771 CDEBUG(D_READA, "reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
772 ria->ria_reserved, pages, pages_min,
773 atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
774 ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
776 ret = ll_read_ahead_pages(env, io, queue, ras, ria, &ra_end_idx);
778 if (ria->ria_reserved != 0)
779 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
781 if (ra_end_idx == end_idx && ra_end_idx == (kms >> PAGE_SHIFT))
782 ll_ra_stats_inc(inode, RA_STAT_EOF);
785 "ra_end_idx = %lu end_idx = %lu stride end = %lu pages = %d\n",
786 ra_end_idx, end_idx, ria->ria_end_idx, ret);
788 if (ra_end_idx != end_idx)
789 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
790 if (ra_end_idx > 0) {
791 /* update the ras so that the next read-ahead tries from
792 * where we left off. */
793 spin_lock(&ras->ras_lock);
794 ras->ras_next_readahead_idx = ra_end_idx + 1;
795 spin_unlock(&ras->ras_lock);
802 static int ll_readpages(const struct lu_env *env, struct cl_io *io,
803 struct cl_page_list *queue,
804 pgoff_t start, pgoff_t end)
813 ret = ll_readahead_file_kms(env, io, &kms);
821 unsigned long end_index;
823 end_index = (unsigned long)((kms - 1) >> PAGE_SHIFT);
824 if (end_index <= end)
828 for (page_idx = start; page_idx <= end; page_idx++) {
829 ret= ll_read_ahead_page(env, io, queue, page_idx,
833 else if (ret == 0) /* ret 1 is already uptodate */
837 RETURN(count > 0 ? count : ret);
840 static void ras_set_start(struct ll_readahead_state *ras, pgoff_t index)
842 ras->ras_window_start_idx = ras_align(ras, index);
845 /* called with the ras_lock held or from places where it doesn't matter */
846 static void ras_reset(struct ll_readahead_state *ras, pgoff_t index)
848 ras->ras_consecutive_requests = 0;
849 ras->ras_consecutive_bytes = 0;
850 ras->ras_window_pages = 0;
851 ras_set_start(ras, index);
852 ras->ras_next_readahead_idx = max(ras->ras_window_start_idx, index + 1);
857 /* called with the ras_lock held or from places where it doesn't matter */
858 static void ras_stride_reset(struct ll_readahead_state *ras)
860 ras->ras_consecutive_stride_requests = 0;
861 ras->ras_stride_length = 0;
862 ras->ras_stride_bytes = 0;
866 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
868 spin_lock_init(&ras->ras_lock);
869 ras->ras_rpc_pages = PTLRPC_MAX_BRW_PAGES;
871 ras->ras_last_read_end_bytes = 0;
872 ras->ras_requests = 0;
876 * Check whether the read request is in the stride window.
877 * If it is in the stride window, return true, otherwise return false.
879 static bool read_in_stride_window(struct ll_readahead_state *ras,
880 loff_t pos, loff_t count)
884 if (ras->ras_stride_length == 0 || ras->ras_stride_bytes == 0 ||
885 ras->ras_stride_bytes == ras->ras_stride_length)
888 stride_gap = pos - ras->ras_last_read_end_bytes - 1;
890 /* If it is contiguous read */
892 return ras->ras_consecutive_bytes + count <=
893 ras->ras_stride_bytes;
895 /* Otherwise check the stride by itself */
896 return (ras->ras_stride_length - ras->ras_stride_bytes) == stride_gap &&
897 ras->ras_consecutive_bytes == ras->ras_stride_bytes &&
898 count <= ras->ras_stride_bytes;
901 static void ras_init_stride_detector(struct ll_readahead_state *ras,
902 loff_t pos, loff_t count)
904 loff_t stride_gap = pos - ras->ras_last_read_end_bytes - 1;
906 LASSERT(ras->ras_consecutive_stride_requests == 0);
908 if (pos <= ras->ras_last_read_end_bytes) {
909 /*Reset stride window for forward read*/
910 ras_stride_reset(ras);
914 ras->ras_stride_bytes = ras->ras_consecutive_bytes;
915 ras->ras_stride_length = stride_gap + ras->ras_consecutive_bytes;
916 ras->ras_consecutive_stride_requests++;
917 ras->ras_stride_offset = pos;
923 stride_page_count(struct ll_readahead_state *ras, loff_t len)
926 stride_byte_count(ras->ras_stride_offset,
927 ras->ras_stride_length, ras->ras_stride_bytes,
928 ras->ras_window_start_idx << PAGE_SHIFT, len);
930 return (bytes_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
933 /* Stride Read-ahead window will be increased inc_len according to
934 * stride I/O pattern */
935 static void ras_stride_increase_window(struct ll_readahead_state *ras,
936 struct ll_ra_info *ra, loff_t inc_bytes)
938 loff_t window_bytes, stride_bytes;
943 /* temporarily store in page units to reduce LASSERT() cost below */
944 end = ras->ras_window_start_idx + ras->ras_window_pages;
946 LASSERT(ras->ras_stride_length > 0);
947 LASSERTF(end >= (ras->ras_stride_offset >> PAGE_SHIFT),
948 "window_start_idx %lu, window_pages %lu stride_offset %llu\n",
949 ras->ras_window_start_idx, ras->ras_window_pages,
950 ras->ras_stride_offset);
953 if (end <= ras->ras_stride_offset)
956 stride_bytes = end - ras->ras_stride_offset;
958 div64_u64_rem(stride_bytes, ras->ras_stride_length, &left_bytes);
959 window_bytes = (ras->ras_window_pages << PAGE_SHIFT);
960 if (left_bytes < ras->ras_stride_bytes) {
961 if (ras->ras_stride_bytes - left_bytes >= inc_bytes) {
962 window_bytes += inc_bytes;
965 window_bytes += (ras->ras_stride_bytes - left_bytes);
966 inc_bytes -= (ras->ras_stride_bytes - left_bytes);
969 window_bytes += (ras->ras_stride_length - left_bytes);
972 LASSERT(ras->ras_stride_bytes != 0);
974 step = div64_u64_rem(inc_bytes, ras->ras_stride_bytes, &left_bytes);
976 window_bytes += step * ras->ras_stride_length + left_bytes;
977 LASSERT(window_bytes > 0);
980 if (stride_page_count(ras, window_bytes) <=
981 ra->ra_max_pages_per_file || ras->ras_window_pages == 0)
982 ras->ras_window_pages = (window_bytes >> PAGE_SHIFT);
984 LASSERT(ras->ras_window_pages > 0);
989 static void ras_increase_window(struct inode *inode,
990 struct ll_readahead_state *ras,
991 struct ll_ra_info *ra)
993 /* The stretch of ra-window should be aligned with max rpc_size
994 * but current clio architecture does not support retrieve such
995 * information from lower layer. FIXME later
997 if (stride_io_mode(ras)) {
998 ras_stride_increase_window(ras, ra,
999 (loff_t)ras->ras_rpc_pages << PAGE_SHIFT);
1001 pgoff_t window_pages;
1003 window_pages = min(ras->ras_window_pages + ras->ras_rpc_pages,
1004 ra->ra_max_pages_per_file);
1005 if (window_pages < ras->ras_rpc_pages)
1006 ras->ras_window_pages = window_pages;
1008 ras->ras_window_pages = ras_align(ras, window_pages);
1013 * Seek within 8 pages are considered as sequential read for now.
1015 static inline bool is_loose_seq_read(struct ll_readahead_state *ras, loff_t pos)
1017 return pos_in_window(pos, ras->ras_last_read_end_bytes,
1018 8UL << PAGE_SHIFT, 8UL << PAGE_SHIFT);
1021 static void ras_detect_read_pattern(struct ll_readahead_state *ras,
1022 struct ll_sb_info *sbi,
1023 loff_t pos, size_t count, bool mmap)
1025 bool stride_detect = false;
1026 pgoff_t index = pos >> PAGE_SHIFT;
1029 * Reset the read-ahead window in two cases. First when the app seeks
1030 * or reads to some other part of the file. Secondly if we get a
1031 * read-ahead miss that we think we've previously issued. This can
1032 * be a symptom of there being so many read-ahead pages that the VM
1033 * is reclaiming it before we get to it.
1035 if (!is_loose_seq_read(ras, pos)) {
1036 /* Check whether it is in stride I/O mode */
1037 if (!read_in_stride_window(ras, pos, count)) {
1038 if (ras->ras_consecutive_stride_requests == 0)
1039 ras_init_stride_detector(ras, pos, count);
1041 ras_stride_reset(ras);
1042 ras->ras_consecutive_bytes = 0;
1043 ras_reset(ras, index);
1045 ras->ras_consecutive_bytes = 0;
1046 ras->ras_consecutive_requests = 0;
1047 if (++ras->ras_consecutive_stride_requests > 1)
1048 stride_detect = true;
1051 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
1052 } else if (stride_io_mode(ras)) {
1054 * If this is contiguous read but in stride I/O mode
1055 * currently, check whether stride step still is valid,
1056 * if invalid, it will reset the stride ra window to
1059 if (!read_in_stride_window(ras, pos, count)) {
1060 ras_stride_reset(ras);
1061 ras->ras_window_pages = 0;
1062 ras->ras_next_readahead_idx = index;
1066 ras->ras_consecutive_bytes += count;
1068 pgoff_t idx = ras->ras_consecutive_bytes >> PAGE_SHIFT;
1070 if ((idx >= 4 && (idx & 3UL) == 0) || stride_detect)
1071 ras->ras_need_increase_window = true;
1072 } else if ((ras->ras_consecutive_requests > 1 || stride_detect)) {
1073 ras->ras_need_increase_window = true;
1076 ras->ras_last_read_end_bytes = pos + count - 1;
1079 void ll_ras_enter(struct file *f, loff_t pos, size_t count)
1081 struct ll_file_data *fd = f->private_data;
1082 struct ll_readahead_state *ras = &fd->fd_ras;
1083 struct inode *inode = file_inode(f);
1084 unsigned long index = pos >> PAGE_SHIFT;
1085 struct ll_sb_info *sbi = ll_i2sbi(inode);
1087 spin_lock(&ras->ras_lock);
1088 ras->ras_requests++;
1089 ras->ras_consecutive_requests++;
1090 ras->ras_need_increase_window = false;
1091 ras->ras_no_miss_check = false;
1093 * On the second access to a file smaller than the tunable
1094 * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1095 * file up to ra_max_pages_per_file. This is simply a best effort
1096 * and only occurs once per open file. Normal RA behavior is reverted
1097 * to for subsequent IO.
1099 if (ras->ras_requests >= 2) {
1101 struct ll_ra_info *ra = &sbi->ll_ra_info;
1103 kms_pages = (i_size_read(inode) + PAGE_SIZE - 1) >>
1106 CDEBUG(D_READA, "kmsp %llu mwp %lu mp %lu\n", kms_pages,
1107 ra->ra_max_read_ahead_whole_pages,
1108 ra->ra_max_pages_per_file);
1111 kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1112 ras->ras_window_start_idx = 0;
1113 ras->ras_next_readahead_idx = index + 1;
1114 ras->ras_window_pages = min(ra->ra_max_pages_per_file,
1115 ra->ra_max_read_ahead_whole_pages);
1116 ras->ras_no_miss_check = true;
1117 GOTO(out_unlock, 0);
1120 ras_detect_read_pattern(ras, sbi, pos, count, false);
1122 spin_unlock(&ras->ras_lock);
1125 static bool index_in_stride_window(struct ll_readahead_state *ras,
1128 loff_t pos = (loff_t)index << PAGE_SHIFT;
1130 if (ras->ras_stride_length == 0 || ras->ras_stride_bytes == 0 ||
1131 ras->ras_stride_bytes == ras->ras_stride_length)
1134 if (pos >= ras->ras_stride_offset) {
1137 div64_u64_rem(pos - ras->ras_stride_offset,
1138 ras->ras_stride_length, &offset);
1139 if (offset < ras->ras_stride_bytes ||
1140 ras->ras_stride_length - offset < PAGE_SIZE)
1142 } else if (ras->ras_stride_offset - pos < PAGE_SIZE) {
1150 * ll_ras_enter() is used to detect read pattern according to pos and count.
1152 * ras_update() is used to detect cache miss and
1153 * reset window or increase window accordingly
1155 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
1156 struct ll_readahead_state *ras, pgoff_t index,
1157 enum ras_update_flags flags)
1159 struct ll_ra_info *ra = &sbi->ll_ra_info;
1160 bool hit = flags & LL_RAS_HIT;
1163 spin_lock(&ras->ras_lock);
1166 CDEBUG(D_READA, DFID " pages at %lu miss.\n",
1167 PFID(ll_inode2fid(inode)), index);
1168 ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
1171 * The readahead window has been expanded to cover whole
1172 * file size, we don't care whether ra miss happen or not.
1173 * Because we will read whole file to page cache even if
1174 * some pages missed.
1176 if (ras->ras_no_miss_check)
1177 GOTO(out_unlock, 0);
1179 if (flags & LL_RAS_MMAP)
1180 ras_detect_read_pattern(ras, sbi, (loff_t)index << PAGE_SHIFT,
1183 if (!hit && ras->ras_window_pages &&
1184 index < ras->ras_next_readahead_idx &&
1185 pos_in_window(index, ras->ras_window_start_idx, 0,
1186 ras->ras_window_pages)) {
1187 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
1188 ras->ras_need_increase_window = false;
1190 if (index_in_stride_window(ras, index) &&
1191 stride_io_mode(ras)) {
1193 * if (index != ras->ras_last_readpage + 1)
1194 * ras->ras_consecutive_pages = 0;
1196 ras_reset(ras, index);
1199 * If stride-RA hit cache miss, the stride
1200 * detector will not be reset to avoid the
1201 * overhead of redetecting read-ahead mode,
1202 * but on the condition that the stride window
1203 * is still intersect with normal sequential
1204 * read-ahead window.
1206 if (ras->ras_window_start_idx < ras->ras_stride_offset)
1207 ras_stride_reset(ras);
1211 * Reset both stride window and normal RA
1214 ras_reset(ras, index);
1215 /* ras->ras_consecutive_pages++; */
1216 ras->ras_consecutive_bytes = 0;
1217 ras_stride_reset(ras);
1218 GOTO(out_unlock, 0);
1221 ras_set_start(ras, index);
1223 if (stride_io_mode(ras)) {
1224 /* Since stride readahead is sentivite to the offset
1225 * of read-ahead, so we use original offset here,
1226 * instead of ras_window_start_idx, which is RPC aligned.
1228 ras->ras_next_readahead_idx = max(index + 1,
1229 ras->ras_next_readahead_idx);
1230 ras->ras_window_start_idx =
1231 max_t(pgoff_t, ras->ras_window_start_idx,
1232 ras->ras_stride_offset >> PAGE_SHIFT);
1234 if (ras->ras_next_readahead_idx < ras->ras_window_start_idx)
1235 ras->ras_next_readahead_idx = ras->ras_window_start_idx;
1237 ras->ras_next_readahead_idx = index + 1;
1240 if (ras->ras_need_increase_window) {
1241 ras_increase_window(inode, ras, ra);
1242 ras->ras_need_increase_window = false;
1247 spin_unlock(&ras->ras_lock);
1250 int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
1252 struct inode *inode = vmpage->mapping->host;
1253 struct ll_inode_info *lli = ll_i2info(inode);
1256 struct cl_page *page;
1257 struct cl_object *clob;
1258 bool redirtied = false;
1259 bool unlocked = false;
1264 LASSERT(PageLocked(vmpage));
1265 LASSERT(!PageWriteback(vmpage));
1267 LASSERT(ll_i2dtexp(inode) != NULL);
1269 env = cl_env_get(&refcheck);
1271 GOTO(out, result = PTR_ERR(env));
1273 clob = ll_i2info(inode)->lli_clob;
1274 LASSERT(clob != NULL);
1276 io = vvp_env_thread_io(env);
1278 io->ci_ignore_layout = 1;
1279 result = cl_io_init(env, io, CIT_MISC, clob);
1281 page = cl_page_find(env, clob, vmpage->index,
1282 vmpage, CPT_CACHEABLE);
1283 if (!IS_ERR(page)) {
1284 lu_ref_add(&page->cp_reference, "writepage",
1286 cl_page_assume(env, io, page);
1287 result = cl_page_flush(env, io, page);
1290 * Re-dirty page on error so it retries write,
1291 * but not in case when IO has actually
1292 * occurred and completed with an error.
1294 if (!PageError(vmpage)) {
1295 redirty_page_for_writepage(wbc, vmpage);
1300 cl_page_disown(env, io, page);
1302 lu_ref_del(&page->cp_reference,
1303 "writepage", current);
1304 cl_page_put(env, page);
1306 result = PTR_ERR(page);
1309 cl_io_fini(env, io);
1311 if (redirtied && wbc->sync_mode == WB_SYNC_ALL) {
1312 loff_t offset = cl_offset(clob, vmpage->index);
1314 /* Flush page failed because the extent is being written out.
1315 * Wait for the write of extent to be finished to avoid
1316 * breaking kernel which assumes ->writepage should mark
1317 * PageWriteback or clean the page. */
1318 result = cl_sync_file_range(inode, offset,
1319 offset + PAGE_SIZE - 1,
1322 /* actually we may have written more than one page.
1323 * decreasing this page because the caller will count
1325 wbc->nr_to_write -= result - 1;
1330 cl_env_put(env, &refcheck);
1335 if (!lli->lli_async_rc)
1336 lli->lli_async_rc = result;
1337 SetPageError(vmpage);
1339 unlock_page(vmpage);
1344 int ll_writepages(struct address_space *mapping, struct writeback_control *wbc)
1346 struct inode *inode = mapping->host;
1349 enum cl_fsync_mode mode;
1350 int range_whole = 0;
1354 if (wbc->range_cyclic) {
1355 start = (loff_t)mapping->writeback_index << PAGE_SHIFT;
1356 end = OBD_OBJECT_EOF;
1358 start = wbc->range_start;
1359 end = wbc->range_end;
1360 if (end == LLONG_MAX) {
1361 end = OBD_OBJECT_EOF;
1362 range_whole = start == 0;
1366 mode = CL_FSYNC_NONE;
1367 if (wbc->sync_mode == WB_SYNC_ALL)
1368 mode = CL_FSYNC_LOCAL;
1370 if (ll_i2info(inode)->lli_clob == NULL)
1373 /* for directio, it would call writepages() to evict cached pages
1374 * inside the IO context of write, which will cause deadlock at
1375 * layout_conf since it waits for active IOs to complete. */
1376 result = cl_sync_file_range(inode, start, end, mode, 1);
1378 wbc->nr_to_write -= result;
1382 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) {
1383 if (end == OBD_OBJECT_EOF)
1384 mapping->writeback_index = 0;
1386 mapping->writeback_index = (end >> PAGE_SHIFT) + 1;
1391 struct ll_cl_context *ll_cl_find(struct file *file)
1393 struct ll_file_data *fd = file->private_data;
1394 struct ll_cl_context *lcc;
1395 struct ll_cl_context *found = NULL;
1397 read_lock(&fd->fd_lock);
1398 list_for_each_entry(lcc, &fd->fd_lccs, lcc_list) {
1399 if (lcc->lcc_cookie == current) {
1404 read_unlock(&fd->fd_lock);
1409 void ll_cl_add(struct file *file, const struct lu_env *env, struct cl_io *io,
1412 struct ll_file_data *fd = file->private_data;
1413 struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1415 memset(lcc, 0, sizeof(*lcc));
1416 INIT_LIST_HEAD(&lcc->lcc_list);
1417 lcc->lcc_cookie = current;
1420 lcc->lcc_type = type;
1422 write_lock(&fd->fd_lock);
1423 list_add(&lcc->lcc_list, &fd->fd_lccs);
1424 write_unlock(&fd->fd_lock);
1427 void ll_cl_remove(struct file *file, const struct lu_env *env)
1429 struct ll_file_data *fd = file->private_data;
1430 struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1432 write_lock(&fd->fd_lock);
1433 list_del_init(&lcc->lcc_list);
1434 write_unlock(&fd->fd_lock);
1437 int ll_io_read_page(const struct lu_env *env, struct cl_io *io,
1438 struct cl_page *page, struct file *file)
1440 struct inode *inode = vvp_object_inode(page->cp_obj);
1441 struct ll_sb_info *sbi = ll_i2sbi(inode);
1442 struct ll_file_data *fd = file->private_data;
1443 struct ll_readahead_state *ras = &fd->fd_ras;
1444 struct cl_2queue *queue = &io->ci_queue;
1445 struct cl_sync_io *anchor = NULL;
1446 struct vvp_page *vpg;
1447 int rc = 0, rc2 = 0;
1449 pgoff_t io_start_index;
1450 pgoff_t io_end_index;
1453 vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1454 uptodate = vpg->vpg_defer_uptodate;
1456 if (ll_readahead_enabled(sbi) && !vpg->vpg_ra_updated) {
1457 struct vvp_io *vio = vvp_env_io(env);
1458 enum ras_update_flags flags = 0;
1461 flags |= LL_RAS_HIT;
1462 if (!vio->vui_ra_valid)
1463 flags |= LL_RAS_MMAP;
1464 ras_update(sbi, inode, ras, vvp_index(vpg), flags);
1467 cl_2queue_init(queue);
1469 vpg->vpg_ra_used = 1;
1470 cl_page_export(env, page, 1);
1471 cl_page_disown(env, io, page);
1473 anchor = &vvp_env_info(env)->vti_anchor;
1474 cl_sync_io_init(anchor, 1);
1475 page->cp_sync_io = anchor;
1477 cl_2queue_add(queue, page);
1480 io_start_index = cl_index(io->ci_obj, io->u.ci_rw.crw_pos);
1481 io_end_index = cl_index(io->ci_obj, io->u.ci_rw.crw_pos +
1482 io->u.ci_rw.crw_count - 1);
1483 if (ll_readahead_enabled(sbi)) {
1484 rc2 = ll_readahead(env, io, &queue->c2_qin, ras,
1486 CDEBUG(D_READA, DFID " %d pages read ahead at %lu\n",
1487 PFID(ll_inode2fid(inode)), rc2, vvp_index(vpg));
1488 } else if (vvp_index(vpg) == io_start_index &&
1489 io_end_index - io_start_index > 0) {
1490 rc2 = ll_readpages(env, io, &queue->c2_qin, io_start_index + 1,
1492 CDEBUG(D_READA, DFID " %d pages read at %lu\n",
1493 PFID(ll_inode2fid(inode)), rc2, vvp_index(vpg));
1496 if (queue->c2_qin.pl_nr > 0) {
1497 int count = queue->c2_qin.pl_nr;
1498 rc = cl_io_submit_rw(env, io, CRT_READ, queue);
1500 task_io_account_read(PAGE_SIZE * count);
1504 if (anchor != NULL && !cl_page_is_owned(page, io)) { /* have sent */
1505 rc = cl_sync_io_wait(env, anchor, 0);
1507 cl_page_assume(env, io, page);
1508 cl_page_list_del(env, &queue->c2_qout, page);
1510 if (!PageUptodate(cl_page_vmpage(page))) {
1511 /* Failed to read a mirror, discard this page so that
1512 * new page can be created with new mirror.
1514 * TODO: this is not needed after page reinit
1515 * route is implemented */
1516 cl_page_discard(env, io, page);
1518 cl_page_disown(env, io, page);
1521 /* TODO: discard all pages until page reinit route is implemented */
1522 cl_page_list_discard(env, io, &queue->c2_qin);
1524 /* Unlock unsent read pages in case of error. */
1525 cl_page_list_disown(env, io, &queue->c2_qin);
1527 cl_2queue_fini(env, queue);
1533 * Possible return value:
1534 * 0 no async readahead triggered and fast read could not be used.
1535 * 1 no async readahead, but fast read could be used.
1536 * 2 async readahead triggered and fast read could be used too.
1539 static int kickoff_async_readahead(struct file *file, unsigned long pages)
1541 struct ll_readahead_work *lrw;
1542 struct inode *inode = file_inode(file);
1543 struct ll_sb_info *sbi = ll_i2sbi(inode);
1544 struct ll_file_data *fd = file->private_data;
1545 struct ll_readahead_state *ras = &fd->fd_ras;
1546 struct ll_ra_info *ra = &sbi->ll_ra_info;
1547 unsigned long throttle;
1548 pgoff_t start_idx = ras_align(ras, ras->ras_next_readahead_idx);
1549 pgoff_t end_idx = start_idx + pages - 1;
1551 throttle = min(ra->ra_async_pages_per_file_threshold,
1552 ra->ra_max_pages_per_file);
1554 * If this is strided i/o or the window is smaller than the
1555 * throttle limit, we do not do async readahead. Otherwise,
1556 * we do async readahead, allowing the user thread to do fast i/o.
1558 if (stride_io_mode(ras) || !throttle ||
1559 ras->ras_window_pages < throttle ||
1560 atomic_read(&ra->ra_async_inflight) > ra->ra_async_max_active)
1563 if ((atomic_read(&ra->ra_cur_pages) + pages) > ra->ra_max_pages)
1566 if (ras->ras_async_last_readpage_idx == start_idx)
1569 /* ll_readahead_work_free() free it */
1572 atomic_inc(&sbi->ll_ra_info.ra_async_inflight);
1573 lrw->lrw_file = get_file(file);
1574 lrw->lrw_start_idx = start_idx;
1575 lrw->lrw_end_idx = end_idx;
1576 spin_lock(&ras->ras_lock);
1577 ras->ras_next_readahead_idx = end_idx + 1;
1578 ras->ras_async_last_readpage_idx = start_idx;
1579 spin_unlock(&ras->ras_lock);
1580 ll_readahead_work_add(inode, lrw);
1589 * Check if we can issue a readahead RPC, if that is
1590 * the case, we can't do fast IO because we will need
1591 * a cl_io to issue the RPC.
1593 static bool ll_use_fast_io(struct file *file,
1594 struct ll_readahead_state *ras, pgoff_t index)
1596 unsigned long fast_read_pages =
1597 max(RA_REMAIN_WINDOW_MIN, ras->ras_rpc_pages);
1600 if (stride_io_mode(ras)) {
1601 skip_pages = (ras->ras_stride_length +
1602 ras->ras_stride_bytes - 1) / ras->ras_stride_bytes;
1603 skip_pages *= fast_read_pages;
1605 skip_pages = fast_read_pages;
1608 if (ras->ras_window_start_idx + ras->ras_window_pages <
1609 ras->ras_next_readahead_idx + skip_pages ||
1610 kickoff_async_readahead(file, fast_read_pages) > 0)
1616 int ll_readpage(struct file *file, struct page *vmpage)
1618 struct inode *inode = file_inode(file);
1619 struct cl_object *clob = ll_i2info(inode)->lli_clob;
1620 struct ll_cl_context *lcc;
1621 const struct lu_env *env = NULL;
1622 struct cl_io *io = NULL;
1623 struct cl_page *page;
1624 struct ll_sb_info *sbi = ll_i2sbi(inode);
1628 lcc = ll_cl_find(file);
1634 if (io == NULL) { /* fast read */
1635 struct inode *inode = file_inode(file);
1636 struct ll_file_data *fd = file->private_data;
1637 struct ll_readahead_state *ras = &fd->fd_ras;
1638 struct lu_env *local_env = NULL;
1639 struct vvp_page *vpg;
1643 /* TODO: need to verify the layout version to make sure
1644 * the page is not invalid due to layout change. */
1645 page = cl_vmpage_page(vmpage, clob);
1647 unlock_page(vmpage);
1648 ll_ra_stats_inc_sbi(sbi, RA_STAT_FAILED_FAST_READ);
1652 vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1653 if (vpg->vpg_defer_uptodate) {
1654 enum ras_update_flags flags = LL_RAS_HIT;
1656 if (lcc && lcc->lcc_type == LCC_MMAP)
1657 flags |= LL_RAS_MMAP;
1659 /* For fast read, it updates read ahead state only
1660 * if the page is hit in cache because non cache page
1661 * case will be handled by slow read later. */
1662 ras_update(sbi, inode, ras, vvp_index(vpg), flags);
1663 /* avoid duplicate ras_update() call */
1664 vpg->vpg_ra_updated = 1;
1666 if (ll_use_fast_io(file, ras, vvp_index(vpg)))
1671 local_env = cl_env_percpu_get();
1675 /* export the page and skip io stack */
1677 vpg->vpg_ra_used = 1;
1678 cl_page_export(env, page, 1);
1680 ll_ra_stats_inc_sbi(sbi, RA_STAT_FAILED_FAST_READ);
1682 /* release page refcount before unlocking the page to ensure
1683 * the object won't be destroyed in the calling path of
1684 * cl_page_put(). Please see comment in ll_releasepage(). */
1685 cl_page_put(env, page);
1686 unlock_page(vmpage);
1688 cl_env_percpu_put(local_env);
1694 * Direct read can fall back to buffered read, but DIO is done
1695 * with lockless i/o, and buffered requires LDLM locking, so in
1696 * this case we must restart without lockless.
1698 if (file->f_flags & O_DIRECT &&
1699 lcc && lcc->lcc_type == LCC_RW &&
1700 !io->ci_ignore_lockless) {
1701 unlock_page(vmpage);
1702 io->ci_ignore_lockless = 1;
1703 io->ci_need_restart = 1;
1707 LASSERT(io->ci_state == CIS_IO_GOING);
1708 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
1709 if (!IS_ERR(page)) {
1710 LASSERT(page->cp_type == CPT_CACHEABLE);
1711 if (likely(!PageUptodate(vmpage))) {
1712 cl_page_assume(env, io, page);
1714 result = ll_io_read_page(env, io, page, file);
1716 /* Page from a non-object file. */
1717 unlock_page(vmpage);
1720 cl_page_put(env, page);
1722 unlock_page(vmpage);
1723 result = PTR_ERR(page);