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;
531 struct ll_sb_info *sbi;
533 work = container_of(wq, struct ll_readahead_work,
535 fd = work->lrw_file->private_data;
537 file = work->lrw_file;
538 inode = file_inode(file);
539 sbi = ll_i2sbi(inode);
541 env = cl_env_alloc(&refcheck, LCT_NOREF);
543 GOTO(out_free_work, rc = PTR_ERR(env));
545 io = vvp_env_thread_io(env);
546 ll_io_init(io, file, CIT_READ, NULL);
548 rc = ll_readahead_file_kms(env, io, &kms);
550 GOTO(out_put_env, rc);
553 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
554 GOTO(out_put_env, rc = 0);
557 ria = &ll_env_info(env)->lti_ria;
558 memset(ria, 0, sizeof(*ria));
560 ria->ria_start_idx = work->lrw_start_idx;
561 /* Truncate RA window to end of file */
562 eof_index = (pgoff_t)(kms - 1) >> PAGE_SHIFT;
563 if (eof_index <= work->lrw_end_idx) {
564 work->lrw_end_idx = eof_index;
567 if (work->lrw_end_idx <= work->lrw_start_idx)
568 GOTO(out_put_env, rc = 0);
570 ria->ria_end_idx = work->lrw_end_idx;
571 pages = ria->ria_end_idx - ria->ria_start_idx + 1;
572 ria->ria_reserved = ll_ra_count_get(sbi, ria,
573 ria_page_count(ria), pages_min);
576 "async reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
577 ria->ria_reserved, pages, pages_min,
578 atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
579 ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
581 if (ria->ria_reserved < pages) {
582 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
583 if (PAGES_TO_MiB(ria->ria_reserved) < 1) {
584 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
585 GOTO(out_put_env, rc = 0);
589 rc = cl_io_rw_init(env, io, CIT_READ, ria->ria_start_idx, pages);
591 GOTO(out_put_env, rc);
593 vvp_env_io(env)->vui_io_subtype = IO_NORMAL;
594 vvp_env_io(env)->vui_fd = fd;
595 io->ci_state = CIS_LOCKED;
596 io->ci_async_readahead = true;
597 rc = cl_io_start(env, io);
599 GOTO(out_io_fini, rc);
601 queue = &io->ci_queue;
602 cl_2queue_init(queue);
604 rc = ll_read_ahead_pages(env, io, &queue->c2_qin, ras, ria,
606 if (ria->ria_reserved != 0)
607 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
608 if (queue->c2_qin.pl_nr > 0) {
609 int count = queue->c2_qin.pl_nr;
611 rc = cl_io_submit_rw(env, io, CRT_READ, queue);
613 task_io_account_read(PAGE_SIZE * count);
615 if (ria->ria_end_idx == ra_end_idx && ra_end_idx == (kms >> PAGE_SHIFT))
616 ll_ra_stats_inc(inode, RA_STAT_EOF);
618 if (ra_end_idx != ria->ria_end_idx)
619 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
621 /* TODO: discard all pages until page reinit route is implemented */
622 cl_page_list_discard(env, io, &queue->c2_qin);
624 /* Unlock unsent read pages in case of error. */
625 cl_page_list_disown(env, io, &queue->c2_qin);
627 cl_2queue_fini(env, queue);
632 cl_env_put(env, &refcheck);
635 ll_ra_stats_inc_sbi(ll_i2sbi(inode), RA_STAT_ASYNC);
636 atomic_dec(&sbi->ll_ra_info.ra_async_inflight);
637 ll_readahead_work_free(work);
640 static int ll_readahead(const struct lu_env *env, struct cl_io *io,
641 struct cl_page_list *queue,
642 struct ll_readahead_state *ras, bool hit,
645 struct vvp_io *vio = vvp_env_io(env);
646 struct ll_thread_info *lti = ll_env_info(env);
647 unsigned long pages, pages_min = 0;
648 pgoff_t ra_end_idx = 0, start_idx = 0, end_idx = 0;
650 struct ra_io_arg *ria = <i->lti_ria;
651 struct cl_object *clob;
657 inode = vvp_object_inode(clob);
659 memset(ria, 0, sizeof(*ria));
660 ret = ll_readahead_file_kms(env, io, &kms);
665 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
669 spin_lock(&ras->ras_lock);
672 * Note: other thread might rollback the ras_next_readahead_idx,
673 * if it can not get the full size of prepared pages, see the
674 * end of this function. For stride read ahead, it needs to
675 * make sure the offset is no less than ras_stride_offset,
676 * so that stride read ahead can work correctly.
678 if (stride_io_mode(ras))
679 start_idx = max_t(pgoff_t, ras->ras_next_readahead_idx,
680 ras->ras_stride_offset >> PAGE_SHIFT);
682 start_idx = ras->ras_next_readahead_idx;
684 if (ras->ras_window_pages > 0)
685 end_idx = ras->ras_window_start_idx + ras->ras_window_pages - 1;
687 /* Enlarge the RA window to encompass the full read */
688 if (vio->vui_ra_valid &&
689 end_idx < vio->vui_ra_start_idx + vio->vui_ra_pages - 1)
690 end_idx = vio->vui_ra_start_idx + vio->vui_ra_pages - 1;
695 /* Truncate RA window to end of file */
696 eof_index = (pgoff_t)((kms - 1) >> PAGE_SHIFT);
697 if (eof_index <= end_idx) {
702 ria->ria_start_idx = start_idx;
703 ria->ria_end_idx = end_idx;
704 /* If stride I/O mode is detected, get stride window*/
705 if (stride_io_mode(ras)) {
706 ria->ria_stoff = ras->ras_stride_offset;
707 ria->ria_length = ras->ras_stride_length;
708 ria->ria_bytes = ras->ras_stride_bytes;
710 spin_unlock(&ras->ras_lock);
713 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
716 pages = ria_page_count(ria);
718 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
723 CDEBUG(D_READA, DFID": ria: %lu/%lu, bead: %lu/%lu, hit: %d\n",
724 PFID(lu_object_fid(&clob->co_lu)),
725 ria->ria_start_idx, ria->ria_end_idx,
726 vio->vui_ra_valid ? vio->vui_ra_start_idx : 0,
727 vio->vui_ra_valid ? vio->vui_ra_pages : 0,
730 /* at least to extend the readahead window to cover current read */
731 if (!hit && vio->vui_ra_valid &&
732 vio->vui_ra_start_idx + vio->vui_ra_pages > ria->ria_start_idx) {
733 ria->ria_end_idx_min =
734 vio->vui_ra_start_idx + vio->vui_ra_pages - 1;
735 pages_min = vio->vui_ra_start_idx + vio->vui_ra_pages -
739 ria->ria_reserved = ll_ra_count_get(ll_i2sbi(inode), ria, pages,
741 if (ria->ria_reserved < pages)
742 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
744 CDEBUG(D_READA, "reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
745 ria->ria_reserved, pages, pages_min,
746 atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
747 ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
749 ret = ll_read_ahead_pages(env, io, queue, ras, ria, &ra_end_idx);
751 if (ria->ria_reserved != 0)
752 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
754 if (ra_end_idx == end_idx && ra_end_idx == (kms >> PAGE_SHIFT))
755 ll_ra_stats_inc(inode, RA_STAT_EOF);
758 "ra_end_idx = %lu end_idx = %lu stride end = %lu pages = %d\n",
759 ra_end_idx, end_idx, ria->ria_end_idx, ret);
761 if (ra_end_idx != end_idx)
762 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
763 if (ra_end_idx > 0) {
764 /* update the ras so that the next read-ahead tries from
765 * where we left off. */
766 spin_lock(&ras->ras_lock);
767 ras->ras_next_readahead_idx = ra_end_idx + 1;
768 spin_unlock(&ras->ras_lock);
775 static void ras_set_start(struct ll_readahead_state *ras, pgoff_t index)
777 ras->ras_window_start_idx = ras_align(ras, index);
780 /* called with the ras_lock held or from places where it doesn't matter */
781 static void ras_reset(struct ll_readahead_state *ras, pgoff_t index)
783 ras->ras_consecutive_requests = 0;
784 ras->ras_consecutive_bytes = 0;
785 ras->ras_window_pages = 0;
786 ras_set_start(ras, index);
787 ras->ras_next_readahead_idx = max(ras->ras_window_start_idx, index + 1);
792 /* called with the ras_lock held or from places where it doesn't matter */
793 static void ras_stride_reset(struct ll_readahead_state *ras)
795 ras->ras_consecutive_stride_requests = 0;
796 ras->ras_stride_length = 0;
797 ras->ras_stride_bytes = 0;
801 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
803 spin_lock_init(&ras->ras_lock);
804 ras->ras_rpc_pages = PTLRPC_MAX_BRW_PAGES;
806 ras->ras_last_read_end_bytes = 0;
807 ras->ras_requests = 0;
811 * Check whether the read request is in the stride window.
812 * If it is in the stride window, return true, otherwise return false.
814 static bool read_in_stride_window(struct ll_readahead_state *ras,
815 loff_t pos, loff_t count)
819 if (ras->ras_stride_length == 0 || ras->ras_stride_bytes == 0 ||
820 ras->ras_stride_bytes == ras->ras_stride_length)
823 stride_gap = pos - ras->ras_last_read_end_bytes - 1;
825 /* If it is contiguous read */
827 return ras->ras_consecutive_bytes + count <=
828 ras->ras_stride_bytes;
830 /* Otherwise check the stride by itself */
831 return (ras->ras_stride_length - ras->ras_stride_bytes) == stride_gap &&
832 ras->ras_consecutive_bytes == ras->ras_stride_bytes &&
833 count <= ras->ras_stride_bytes;
836 static void ras_init_stride_detector(struct ll_readahead_state *ras,
837 loff_t pos, loff_t count)
839 loff_t stride_gap = pos - ras->ras_last_read_end_bytes - 1;
841 LASSERT(ras->ras_consecutive_stride_requests == 0);
843 if (pos <= ras->ras_last_read_end_bytes) {
844 /*Reset stride window for forward read*/
845 ras_stride_reset(ras);
849 ras->ras_stride_bytes = ras->ras_consecutive_bytes;
850 ras->ras_stride_length = stride_gap + ras->ras_consecutive_bytes;
851 ras->ras_consecutive_stride_requests++;
852 ras->ras_stride_offset = pos;
858 stride_page_count(struct ll_readahead_state *ras, loff_t len)
861 stride_byte_count(ras->ras_stride_offset,
862 ras->ras_stride_length, ras->ras_stride_bytes,
863 ras->ras_window_start_idx << PAGE_SHIFT, len);
865 return (bytes_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
868 /* Stride Read-ahead window will be increased inc_len according to
869 * stride I/O pattern */
870 static void ras_stride_increase_window(struct ll_readahead_state *ras,
871 struct ll_ra_info *ra, loff_t inc_bytes)
873 loff_t window_bytes, stride_bytes;
878 /* temporarily store in page units to reduce LASSERT() cost below */
879 end = ras->ras_window_start_idx + ras->ras_window_pages;
881 LASSERT(ras->ras_stride_length > 0);
882 LASSERTF(end >= (ras->ras_stride_offset >> PAGE_SHIFT),
883 "window_start_idx %lu, window_pages %lu stride_offset %llu\n",
884 ras->ras_window_start_idx, ras->ras_window_pages,
885 ras->ras_stride_offset);
888 if (end <= ras->ras_stride_offset)
891 stride_bytes = end - ras->ras_stride_offset;
893 div64_u64_rem(stride_bytes, ras->ras_stride_length, &left_bytes);
894 window_bytes = (ras->ras_window_pages << PAGE_SHIFT);
895 if (left_bytes < ras->ras_stride_bytes) {
896 if (ras->ras_stride_bytes - left_bytes >= inc_bytes) {
897 window_bytes += inc_bytes;
900 window_bytes += (ras->ras_stride_bytes - left_bytes);
901 inc_bytes -= (ras->ras_stride_bytes - left_bytes);
904 window_bytes += (ras->ras_stride_length - left_bytes);
907 LASSERT(ras->ras_stride_bytes != 0);
909 step = div64_u64_rem(inc_bytes, ras->ras_stride_bytes, &left_bytes);
911 window_bytes += step * ras->ras_stride_length + left_bytes;
912 LASSERT(window_bytes > 0);
915 if (stride_page_count(ras, window_bytes) <=
916 ra->ra_max_pages_per_file || ras->ras_window_pages == 0)
917 ras->ras_window_pages = (window_bytes >> PAGE_SHIFT);
919 LASSERT(ras->ras_window_pages > 0);
924 static void ras_increase_window(struct inode *inode,
925 struct ll_readahead_state *ras,
926 struct ll_ra_info *ra)
928 /* The stretch of ra-window should be aligned with max rpc_size
929 * but current clio architecture does not support retrieve such
930 * information from lower layer. FIXME later
932 if (stride_io_mode(ras)) {
933 ras_stride_increase_window(ras, ra,
934 (loff_t)ras->ras_rpc_pages << PAGE_SHIFT);
936 pgoff_t window_pages;
938 window_pages = min(ras->ras_window_pages + ras->ras_rpc_pages,
939 ra->ra_max_pages_per_file);
940 if (window_pages < ras->ras_rpc_pages)
941 ras->ras_window_pages = window_pages;
943 ras->ras_window_pages = ras_align(ras, window_pages);
948 * Seek within 8 pages are considered as sequential read for now.
950 static inline bool is_loose_seq_read(struct ll_readahead_state *ras, loff_t pos)
952 return pos_in_window(pos, ras->ras_last_read_end_bytes,
953 8UL << PAGE_SHIFT, 8UL << PAGE_SHIFT);
956 static void ras_detect_read_pattern(struct ll_readahead_state *ras,
957 struct ll_sb_info *sbi,
958 loff_t pos, size_t count, bool mmap)
960 bool stride_detect = false;
961 pgoff_t index = pos >> PAGE_SHIFT;
964 * Reset the read-ahead window in two cases. First when the app seeks
965 * or reads to some other part of the file. Secondly if we get a
966 * read-ahead miss that we think we've previously issued. This can
967 * be a symptom of there being so many read-ahead pages that the VM
968 * is reclaiming it before we get to it.
970 if (!is_loose_seq_read(ras, pos)) {
971 /* Check whether it is in stride I/O mode */
972 if (!read_in_stride_window(ras, pos, count)) {
973 if (ras->ras_consecutive_stride_requests == 0)
974 ras_init_stride_detector(ras, pos, count);
976 ras_stride_reset(ras);
977 ras->ras_consecutive_bytes = 0;
978 ras_reset(ras, index);
980 ras->ras_consecutive_bytes = 0;
981 ras->ras_consecutive_requests = 0;
982 if (++ras->ras_consecutive_stride_requests > 1)
983 stride_detect = true;
986 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
987 } else if (stride_io_mode(ras)) {
989 * If this is contiguous read but in stride I/O mode
990 * currently, check whether stride step still is valid,
991 * if invalid, it will reset the stride ra window to
994 if (!read_in_stride_window(ras, pos, count)) {
995 ras_stride_reset(ras);
996 ras->ras_window_pages = 0;
997 ras->ras_next_readahead_idx = index;
1001 ras->ras_consecutive_bytes += count;
1003 pgoff_t idx = ras->ras_consecutive_bytes >> PAGE_SHIFT;
1005 if ((idx >= 4 && (idx & 3UL) == 0) || stride_detect)
1006 ras->ras_need_increase_window = true;
1007 } else if ((ras->ras_consecutive_requests > 1 || stride_detect)) {
1008 ras->ras_need_increase_window = true;
1011 ras->ras_last_read_end_bytes = pos + count - 1;
1014 void ll_ras_enter(struct file *f, loff_t pos, size_t count)
1016 struct ll_file_data *fd = f->private_data;
1017 struct ll_readahead_state *ras = &fd->fd_ras;
1018 struct inode *inode = file_inode(f);
1019 unsigned long index = pos >> PAGE_SHIFT;
1020 struct ll_sb_info *sbi = ll_i2sbi(inode);
1022 spin_lock(&ras->ras_lock);
1023 ras->ras_requests++;
1024 ras->ras_consecutive_requests++;
1025 ras->ras_need_increase_window = false;
1026 ras->ras_no_miss_check = false;
1028 * On the second access to a file smaller than the tunable
1029 * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1030 * file up to ra_max_pages_per_file. This is simply a best effort
1031 * and only occurs once per open file. Normal RA behavior is reverted
1032 * to for subsequent IO.
1034 if (ras->ras_requests >= 2) {
1036 struct ll_ra_info *ra = &sbi->ll_ra_info;
1038 kms_pages = (i_size_read(inode) + PAGE_SIZE - 1) >>
1041 CDEBUG(D_READA, "kmsp %llu mwp %lu mp %lu\n", kms_pages,
1042 ra->ra_max_read_ahead_whole_pages,
1043 ra->ra_max_pages_per_file);
1046 kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1047 ras->ras_window_start_idx = 0;
1048 ras->ras_next_readahead_idx = index + 1;
1049 ras->ras_window_pages = min(ra->ra_max_pages_per_file,
1050 ra->ra_max_read_ahead_whole_pages);
1051 ras->ras_no_miss_check = true;
1052 GOTO(out_unlock, 0);
1055 ras_detect_read_pattern(ras, sbi, pos, count, false);
1057 spin_unlock(&ras->ras_lock);
1060 static bool index_in_stride_window(struct ll_readahead_state *ras,
1063 loff_t pos = (loff_t)index << PAGE_SHIFT;
1065 if (ras->ras_stride_length == 0 || ras->ras_stride_bytes == 0 ||
1066 ras->ras_stride_bytes == ras->ras_stride_length)
1069 if (pos >= ras->ras_stride_offset) {
1072 div64_u64_rem(pos - ras->ras_stride_offset,
1073 ras->ras_stride_length, &offset);
1074 if (offset < ras->ras_stride_bytes ||
1075 ras->ras_stride_length - offset < PAGE_SIZE)
1077 } else if (ras->ras_stride_offset - pos < PAGE_SIZE) {
1085 * ll_ras_enter() is used to detect read pattern according to pos and count.
1087 * ras_update() is used to detect cache miss and
1088 * reset window or increase window accordingly
1090 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
1091 struct ll_readahead_state *ras, pgoff_t index,
1092 enum ras_update_flags flags)
1094 struct ll_ra_info *ra = &sbi->ll_ra_info;
1095 bool hit = flags & LL_RAS_HIT;
1098 spin_lock(&ras->ras_lock);
1101 CDEBUG(D_READA, DFID " pages at %lu miss.\n",
1102 PFID(ll_inode2fid(inode)), index);
1103 ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
1106 * The readahead window has been expanded to cover whole
1107 * file size, we don't care whether ra miss happen or not.
1108 * Because we will read whole file to page cache even if
1109 * some pages missed.
1111 if (ras->ras_no_miss_check)
1112 GOTO(out_unlock, 0);
1114 if (flags & LL_RAS_MMAP)
1115 ras_detect_read_pattern(ras, sbi, (loff_t)index << PAGE_SHIFT,
1118 if (!hit && ras->ras_window_pages &&
1119 index < ras->ras_next_readahead_idx &&
1120 pos_in_window(index, ras->ras_window_start_idx, 0,
1121 ras->ras_window_pages)) {
1122 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
1123 ras->ras_need_increase_window = false;
1125 if (index_in_stride_window(ras, index) &&
1126 stride_io_mode(ras)) {
1128 * if (index != ras->ras_last_readpage + 1)
1129 * ras->ras_consecutive_pages = 0;
1131 ras_reset(ras, index);
1134 * If stride-RA hit cache miss, the stride
1135 * detector will not be reset to avoid the
1136 * overhead of redetecting read-ahead mode,
1137 * but on the condition that the stride window
1138 * is still intersect with normal sequential
1139 * read-ahead window.
1141 if (ras->ras_window_start_idx < ras->ras_stride_offset)
1142 ras_stride_reset(ras);
1146 * Reset both stride window and normal RA
1149 ras_reset(ras, index);
1150 /* ras->ras_consecutive_pages++; */
1151 ras->ras_consecutive_bytes = 0;
1152 ras_stride_reset(ras);
1153 GOTO(out_unlock, 0);
1156 ras_set_start(ras, index);
1158 if (stride_io_mode(ras)) {
1159 /* Since stride readahead is sentivite to the offset
1160 * of read-ahead, so we use original offset here,
1161 * instead of ras_window_start_idx, which is RPC aligned.
1163 ras->ras_next_readahead_idx = max(index + 1,
1164 ras->ras_next_readahead_idx);
1165 ras->ras_window_start_idx =
1166 max_t(pgoff_t, ras->ras_window_start_idx,
1167 ras->ras_stride_offset >> PAGE_SHIFT);
1169 if (ras->ras_next_readahead_idx < ras->ras_window_start_idx)
1170 ras->ras_next_readahead_idx = ras->ras_window_start_idx;
1172 ras->ras_next_readahead_idx = index + 1;
1175 if (ras->ras_need_increase_window) {
1176 ras_increase_window(inode, ras, ra);
1177 ras->ras_need_increase_window = false;
1182 spin_unlock(&ras->ras_lock);
1185 int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
1187 struct inode *inode = vmpage->mapping->host;
1188 struct ll_inode_info *lli = ll_i2info(inode);
1191 struct cl_page *page;
1192 struct cl_object *clob;
1193 bool redirtied = false;
1194 bool unlocked = false;
1199 LASSERT(PageLocked(vmpage));
1200 LASSERT(!PageWriteback(vmpage));
1202 LASSERT(ll_i2dtexp(inode) != NULL);
1204 env = cl_env_get(&refcheck);
1206 GOTO(out, result = PTR_ERR(env));
1208 clob = ll_i2info(inode)->lli_clob;
1209 LASSERT(clob != NULL);
1211 io = vvp_env_thread_io(env);
1213 io->ci_ignore_layout = 1;
1214 result = cl_io_init(env, io, CIT_MISC, clob);
1216 page = cl_page_find(env, clob, vmpage->index,
1217 vmpage, CPT_CACHEABLE);
1218 if (!IS_ERR(page)) {
1219 lu_ref_add(&page->cp_reference, "writepage",
1221 cl_page_assume(env, io, page);
1222 result = cl_page_flush(env, io, page);
1225 * Re-dirty page on error so it retries write,
1226 * but not in case when IO has actually
1227 * occurred and completed with an error.
1229 if (!PageError(vmpage)) {
1230 redirty_page_for_writepage(wbc, vmpage);
1235 cl_page_disown(env, io, page);
1237 lu_ref_del(&page->cp_reference,
1238 "writepage", current);
1239 cl_page_put(env, page);
1241 result = PTR_ERR(page);
1244 cl_io_fini(env, io);
1246 if (redirtied && wbc->sync_mode == WB_SYNC_ALL) {
1247 loff_t offset = cl_offset(clob, vmpage->index);
1249 /* Flush page failed because the extent is being written out.
1250 * Wait for the write of extent to be finished to avoid
1251 * breaking kernel which assumes ->writepage should mark
1252 * PageWriteback or clean the page. */
1253 result = cl_sync_file_range(inode, offset,
1254 offset + PAGE_SIZE - 1,
1257 /* actually we may have written more than one page.
1258 * decreasing this page because the caller will count
1260 wbc->nr_to_write -= result - 1;
1265 cl_env_put(env, &refcheck);
1270 if (!lli->lli_async_rc)
1271 lli->lli_async_rc = result;
1272 SetPageError(vmpage);
1274 unlock_page(vmpage);
1279 int ll_writepages(struct address_space *mapping, struct writeback_control *wbc)
1281 struct inode *inode = mapping->host;
1284 enum cl_fsync_mode mode;
1285 int range_whole = 0;
1289 if (wbc->range_cyclic) {
1290 start = (loff_t)mapping->writeback_index << PAGE_SHIFT;
1291 end = OBD_OBJECT_EOF;
1293 start = wbc->range_start;
1294 end = wbc->range_end;
1295 if (end == LLONG_MAX) {
1296 end = OBD_OBJECT_EOF;
1297 range_whole = start == 0;
1301 mode = CL_FSYNC_NONE;
1302 if (wbc->sync_mode == WB_SYNC_ALL)
1303 mode = CL_FSYNC_LOCAL;
1305 if (ll_i2info(inode)->lli_clob == NULL)
1308 /* for directio, it would call writepages() to evict cached pages
1309 * inside the IO context of write, which will cause deadlock at
1310 * layout_conf since it waits for active IOs to complete. */
1311 result = cl_sync_file_range(inode, start, end, mode, 1);
1313 wbc->nr_to_write -= result;
1317 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) {
1318 if (end == OBD_OBJECT_EOF)
1319 mapping->writeback_index = 0;
1321 mapping->writeback_index = (end >> PAGE_SHIFT) + 1;
1326 struct ll_cl_context *ll_cl_find(struct file *file)
1328 struct ll_file_data *fd = file->private_data;
1329 struct ll_cl_context *lcc;
1330 struct ll_cl_context *found = NULL;
1332 read_lock(&fd->fd_lock);
1333 list_for_each_entry(lcc, &fd->fd_lccs, lcc_list) {
1334 if (lcc->lcc_cookie == current) {
1339 read_unlock(&fd->fd_lock);
1344 void ll_cl_add(struct file *file, const struct lu_env *env, struct cl_io *io,
1347 struct ll_file_data *fd = file->private_data;
1348 struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1350 memset(lcc, 0, sizeof(*lcc));
1351 INIT_LIST_HEAD(&lcc->lcc_list);
1352 lcc->lcc_cookie = current;
1355 lcc->lcc_type = type;
1357 write_lock(&fd->fd_lock);
1358 list_add(&lcc->lcc_list, &fd->fd_lccs);
1359 write_unlock(&fd->fd_lock);
1362 void ll_cl_remove(struct file *file, const struct lu_env *env)
1364 struct ll_file_data *fd = file->private_data;
1365 struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1367 write_lock(&fd->fd_lock);
1368 list_del_init(&lcc->lcc_list);
1369 write_unlock(&fd->fd_lock);
1372 int ll_io_read_page(const struct lu_env *env, struct cl_io *io,
1373 struct cl_page *page, struct file *file)
1375 struct inode *inode = vvp_object_inode(page->cp_obj);
1376 struct ll_sb_info *sbi = ll_i2sbi(inode);
1377 struct ll_file_data *fd = file->private_data;
1378 struct ll_readahead_state *ras = &fd->fd_ras;
1379 struct cl_2queue *queue = &io->ci_queue;
1380 struct cl_sync_io *anchor = NULL;
1381 struct vvp_page *vpg;
1386 vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1387 uptodate = vpg->vpg_defer_uptodate;
1389 if (sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
1390 sbi->ll_ra_info.ra_max_pages > 0 &&
1391 !vpg->vpg_ra_updated) {
1392 struct vvp_io *vio = vvp_env_io(env);
1393 enum ras_update_flags flags = 0;
1396 flags |= LL_RAS_HIT;
1397 if (!vio->vui_ra_valid)
1398 flags |= LL_RAS_MMAP;
1399 ras_update(sbi, inode, ras, vvp_index(vpg), flags);
1402 cl_2queue_init(queue);
1404 vpg->vpg_ra_used = 1;
1405 cl_page_export(env, page, 1);
1406 cl_page_disown(env, io, page);
1408 anchor = &vvp_env_info(env)->vti_anchor;
1409 cl_sync_io_init(anchor, 1);
1410 page->cp_sync_io = anchor;
1412 cl_2queue_add(queue, page);
1415 if (sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
1416 sbi->ll_ra_info.ra_max_pages > 0) {
1419 rc2 = ll_readahead(env, io, &queue->c2_qin, ras,
1421 CDEBUG(D_READA, DFID "%d pages read ahead at %lu\n",
1422 PFID(ll_inode2fid(inode)), rc2, vvp_index(vpg));
1425 if (queue->c2_qin.pl_nr > 0) {
1426 int count = queue->c2_qin.pl_nr;
1427 rc = cl_io_submit_rw(env, io, CRT_READ, queue);
1429 task_io_account_read(PAGE_SIZE * count);
1433 if (anchor != NULL && !cl_page_is_owned(page, io)) { /* have sent */
1434 rc = cl_sync_io_wait(env, anchor, 0);
1436 cl_page_assume(env, io, page);
1437 cl_page_list_del(env, &queue->c2_qout, page);
1439 if (!PageUptodate(cl_page_vmpage(page))) {
1440 /* Failed to read a mirror, discard this page so that
1441 * new page can be created with new mirror.
1443 * TODO: this is not needed after page reinit
1444 * route is implemented */
1445 cl_page_discard(env, io, page);
1447 cl_page_disown(env, io, page);
1450 /* TODO: discard all pages until page reinit route is implemented */
1451 cl_page_list_discard(env, io, &queue->c2_qin);
1453 /* Unlock unsent read pages in case of error. */
1454 cl_page_list_disown(env, io, &queue->c2_qin);
1456 cl_2queue_fini(env, queue);
1462 * Possible return value:
1463 * 0 no async readahead triggered and fast read could not be used.
1464 * 1 no async readahead, but fast read could be used.
1465 * 2 async readahead triggered and fast read could be used too.
1468 static int kickoff_async_readahead(struct file *file, unsigned long pages)
1470 struct ll_readahead_work *lrw;
1471 struct inode *inode = file_inode(file);
1472 struct ll_sb_info *sbi = ll_i2sbi(inode);
1473 struct ll_file_data *fd = file->private_data;
1474 struct ll_readahead_state *ras = &fd->fd_ras;
1475 struct ll_ra_info *ra = &sbi->ll_ra_info;
1476 unsigned long throttle;
1477 pgoff_t start_idx = ras_align(ras, ras->ras_next_readahead_idx);
1478 pgoff_t end_idx = start_idx + pages - 1;
1480 throttle = min(ra->ra_async_pages_per_file_threshold,
1481 ra->ra_max_pages_per_file);
1483 * If this is strided i/o or the window is smaller than the
1484 * throttle limit, we do not do async readahead. Otherwise,
1485 * we do async readahead, allowing the user thread to do fast i/o.
1487 if (stride_io_mode(ras) || !throttle ||
1488 ras->ras_window_pages < throttle ||
1489 atomic_read(&ra->ra_async_inflight) > ra->ra_async_max_active)
1492 if ((atomic_read(&ra->ra_cur_pages) + pages) > ra->ra_max_pages)
1495 if (ras->ras_async_last_readpage_idx == start_idx)
1498 /* ll_readahead_work_free() free it */
1501 atomic_inc(&sbi->ll_ra_info.ra_async_inflight);
1502 lrw->lrw_file = get_file(file);
1503 lrw->lrw_start_idx = start_idx;
1504 lrw->lrw_end_idx = end_idx;
1505 spin_lock(&ras->ras_lock);
1506 ras->ras_next_readahead_idx = end_idx + 1;
1507 ras->ras_async_last_readpage_idx = start_idx;
1508 spin_unlock(&ras->ras_lock);
1509 ll_readahead_work_add(inode, lrw);
1518 * Check if we can issue a readahead RPC, if that is
1519 * the case, we can't do fast IO because we will need
1520 * a cl_io to issue the RPC.
1522 static bool ll_use_fast_io(struct file *file,
1523 struct ll_readahead_state *ras, pgoff_t index)
1525 unsigned long fast_read_pages =
1526 max(RA_REMAIN_WINDOW_MIN, ras->ras_rpc_pages);
1529 if (stride_io_mode(ras)) {
1530 skip_pages = (ras->ras_stride_length +
1531 ras->ras_stride_bytes - 1) / ras->ras_stride_bytes;
1532 skip_pages *= fast_read_pages;
1534 skip_pages = fast_read_pages;
1537 if (ras->ras_window_start_idx + ras->ras_window_pages <
1538 ras->ras_next_readahead_idx + skip_pages ||
1539 kickoff_async_readahead(file, fast_read_pages) > 0)
1545 int ll_readpage(struct file *file, struct page *vmpage)
1547 struct inode *inode = file_inode(file);
1548 struct cl_object *clob = ll_i2info(inode)->lli_clob;
1549 struct ll_cl_context *lcc;
1550 const struct lu_env *env = NULL;
1551 struct cl_io *io = NULL;
1552 struct cl_page *page;
1553 struct ll_sb_info *sbi = ll_i2sbi(inode);
1557 lcc = ll_cl_find(file);
1563 if (io == NULL) { /* fast read */
1564 struct inode *inode = file_inode(file);
1565 struct ll_file_data *fd = file->private_data;
1566 struct ll_readahead_state *ras = &fd->fd_ras;
1567 struct lu_env *local_env = NULL;
1568 struct vvp_page *vpg;
1572 /* TODO: need to verify the layout version to make sure
1573 * the page is not invalid due to layout change. */
1574 page = cl_vmpage_page(vmpage, clob);
1576 unlock_page(vmpage);
1577 ll_ra_stats_inc_sbi(sbi, RA_STAT_FAILED_FAST_READ);
1581 vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1582 if (vpg->vpg_defer_uptodate) {
1583 enum ras_update_flags flags = LL_RAS_HIT;
1585 if (lcc && lcc->lcc_type == LCC_MMAP)
1586 flags |= LL_RAS_MMAP;
1588 /* For fast read, it updates read ahead state only
1589 * if the page is hit in cache because non cache page
1590 * case will be handled by slow read later. */
1591 ras_update(sbi, inode, ras, vvp_index(vpg), flags);
1592 /* avoid duplicate ras_update() call */
1593 vpg->vpg_ra_updated = 1;
1595 if (ll_use_fast_io(file, ras, vvp_index(vpg)))
1600 local_env = cl_env_percpu_get();
1604 /* export the page and skip io stack */
1606 vpg->vpg_ra_used = 1;
1607 cl_page_export(env, page, 1);
1609 ll_ra_stats_inc_sbi(sbi, RA_STAT_FAILED_FAST_READ);
1611 /* release page refcount before unlocking the page to ensure
1612 * the object won't be destroyed in the calling path of
1613 * cl_page_put(). Please see comment in ll_releasepage(). */
1614 cl_page_put(env, page);
1615 unlock_page(vmpage);
1617 cl_env_percpu_put(local_env);
1623 * Direct read can fall back to buffered read, but DIO is done
1624 * with lockless i/o, and buffered requires LDLM locking, so in
1625 * this case we must restart without lockless.
1627 if (file->f_flags & O_DIRECT &&
1628 lcc && lcc->lcc_type == LCC_RW &&
1629 !io->ci_ignore_lockless) {
1630 unlock_page(vmpage);
1631 io->ci_ignore_lockless = 1;
1632 io->ci_need_restart = 1;
1636 LASSERT(io->ci_state == CIS_IO_GOING);
1637 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
1638 if (!IS_ERR(page)) {
1639 LASSERT(page->cp_type == CPT_CACHEABLE);
1640 if (likely(!PageUptodate(vmpage))) {
1641 cl_page_assume(env, io, page);
1643 result = ll_io_read_page(env, io, page, file);
1645 /* Page from a non-object file. */
1646 unlock_page(vmpage);
1649 cl_page_put(env, page);
1651 unlock_page(vmpage);
1652 result = PTR_ERR(page);