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/
33 * Lustre Lite I/O page cache routines shared by different kernel revs
36 #include <linux/kernel.h>
38 #include <linux/string.h>
39 #include <linux/stat.h>
40 #include <linux/errno.h>
41 #include <linux/unistd.h>
42 #include <linux/writeback.h>
43 #include <asm/uaccess.h>
46 #include <linux/file.h>
47 #include <linux/stat.h>
48 #include <asm/uaccess.h>
50 #include <linux/pagemap.h>
51 /* current_is_kswapd() */
52 #include <linux/swap.h>
53 #include <linux/task_io_accounting_ops.h>
55 #define DEBUG_SUBSYSTEM S_LLITE
57 #include <obd_cksum.h>
58 #include "llite_internal.h"
59 #include <lustre_compat.h>
61 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which);
64 * Get readahead pages from the filesystem readahead pool of the client for a
67 * /param sbi superblock for filesystem readahead state ll_ra_info
68 * /param ria per-thread readahead state
69 * /param pages number of pages requested for readahead for the thread.
71 * WARNING: This algorithm is used to reduce contention on sbi->ll_lock.
72 * It should work well if the ra_max_pages is much greater than the single
73 * file's read-ahead window, and not too many threads contending for
74 * these readahead pages.
76 * TODO: There may be a 'global sync problem' if many threads are trying
77 * to get an ra budget that is larger than the remaining readahead pages
78 * and reach here at exactly the same time. They will compute /a ret to
79 * consume the remaining pages, but will fail at atomic_add_return() and
80 * get a zero ra window, although there is still ra space remaining. - Jay */
82 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi,
83 struct ra_io_arg *ria,
85 unsigned long pages_min)
87 struct ll_ra_info *ra = &sbi->ll_ra_info;
93 * Don't try readahead agreesively if we are limited
94 * LRU pages, otherwise, it could cause deadlock.
96 pages = min(sbi->ll_cache->ccc_lru_max >> 2, pages);
99 * If read-ahead pages left are less than 1M, do not do read-ahead,
100 * otherwise it will form small read RPC(< 1M), which hurt server
103 ret = min(ra->ra_max_pages - atomic_read(&ra->ra_cur_pages),
105 if (ret < 0 || ret < min_t(long, PTLRPC_MAX_BRW_PAGES, pages))
108 if (atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) {
109 atomic_sub(ret, &ra->ra_cur_pages);
114 if (ret < pages_min) {
115 /* override ra limit for maximum performance */
116 atomic_add(pages_min - ret, &ra->ra_cur_pages);
122 void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long pages)
124 struct ll_ra_info *ra = &sbi->ll_ra_info;
125 atomic_sub(pages, &ra->ra_cur_pages);
128 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which)
130 LASSERTF(which < _NR_RA_STAT, "which: %u\n", which);
131 lprocfs_counter_incr(sbi->ll_ra_stats, which);
134 static inline bool ll_readahead_enabled(struct ll_sb_info *sbi)
136 return sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
137 sbi->ll_ra_info.ra_max_pages > 0;
140 void ll_ra_stats_inc(struct inode *inode, enum ra_stat which)
142 struct ll_sb_info *sbi = ll_i2sbi(inode);
143 ll_ra_stats_inc_sbi(sbi, which);
146 #define RAS_CDEBUG(ras) \
148 "lre %llu cr %lu cb %llu wsi %lu wp %lu nra %lu rpc %lu " \
149 "r %lu csr %lu so %llu sb %llu sl %llu lr %lu\n", \
150 ras->ras_last_read_end_bytes, ras->ras_consecutive_requests, \
151 ras->ras_consecutive_bytes, ras->ras_window_start_idx, \
152 ras->ras_window_pages, ras->ras_next_readahead_idx, \
153 ras->ras_rpc_pages, ras->ras_requests, \
154 ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
155 ras->ras_stride_bytes, ras->ras_stride_length, \
156 ras->ras_async_last_readpage_idx)
158 static bool pos_in_window(loff_t pos, loff_t point,
159 unsigned long before, unsigned long after)
161 loff_t start = point - before;
162 loff_t end = point + after;
169 return start <= pos && pos <= end;
172 enum ll_ra_page_hint {
173 MAYNEED = 0, /* this page possibly accessed soon */
174 WILLNEED /* this page is gurateed to be needed */
178 * Initiates read-ahead of a page with given index.
180 * \retval +ve: page was already uptodate so it will be skipped
182 * \retval -ve: page wasn't added to \a queue for error;
183 * \retval 0: page was added into \a queue for read ahead.
185 static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io,
186 struct cl_page_list *queue, pgoff_t index,
187 enum ll_ra_page_hint hint)
189 struct cl_object *clob = io->ci_obj;
190 struct inode *inode = vvp_object_inode(clob);
191 struct page *vmpage = NULL;
192 struct cl_page *page;
193 struct vvp_page *vpg;
194 enum ra_stat which = _NR_RA_STAT; /* keep gcc happy */
196 const char *msg = NULL;
202 vmpage = grab_cache_page_nowait(inode->i_mapping, index);
203 if (vmpage == NULL) {
204 which = RA_STAT_FAILED_GRAB_PAGE;
205 msg = "g_c_p_n failed";
206 GOTO(out, rc = -EBUSY);
210 vmpage = find_or_create_page(inode->i_mapping, index,
213 GOTO(out, rc = -ENOMEM);
216 /* should not come here */
217 GOTO(out, rc = -EINVAL);
220 /* Check if vmpage was truncated or reclaimed */
221 if (vmpage->mapping != inode->i_mapping) {
222 which = RA_STAT_WRONG_GRAB_PAGE;
223 msg = "g_c_p_n returned invalid page";
224 GOTO(out, rc = -EBUSY);
227 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
229 which = RA_STAT_FAILED_GRAB_PAGE;
230 msg = "cl_page_find failed";
231 GOTO(out, rc = PTR_ERR(page));
234 lu_ref_add(&page->cp_reference, "ra", current);
235 cl_page_assume(env, io, page);
236 vpg = cl2vvp_page(cl_object_page_slice(clob, page));
237 if (!vpg->vpg_defer_uptodate && !PageUptodate(vmpage)) {
238 vpg->vpg_defer_uptodate = 1;
239 vpg->vpg_ra_used = 0;
240 cl_page_list_add(queue, page);
242 /* skip completed pages */
243 cl_page_unassume(env, io, page);
244 /* This page is already uptodate, returning a positive number
245 * to tell the callers about this */
249 lu_ref_del(&page->cp_reference, "ra", current);
250 cl_page_put(env, page);
253 if (vmpage != NULL) {
258 if (msg != NULL && hint == MAYNEED) {
259 ll_ra_stats_inc(inode, which);
260 CDEBUG(D_READA, "%s\n", msg);
267 #define RIA_DEBUG(ria) \
268 CDEBUG(D_READA, "rs %lu re %lu ro %llu rl %llu rb %llu\n", \
269 ria->ria_start_idx, ria->ria_end_idx, ria->ria_stoff, \
270 ria->ria_length, ria->ria_bytes)
272 static inline int stride_io_mode(struct ll_readahead_state *ras)
274 return ras->ras_consecutive_stride_requests > 1;
277 /* The function calculates how many bytes will be read in
278 * [off, off + length], in such stride IO area,
279 * stride_offset = st_off, stride_lengh = st_len,
280 * stride_bytes = st_bytes
282 * |------------------|*****|------------------|*****|------------|*****|....
285 * |----- st_len -----|
287 * How many bytes it should read in such pattern
288 * |-------------------------------------------------------------|
290 * |<------ length ------->|
292 * = |<----->| + |-------------------------------------| + |---|
293 * start_left st_bytes * i end_left
295 static loff_t stride_byte_count(loff_t st_off, loff_t st_len, loff_t st_bytes,
296 loff_t off, loff_t length)
298 u64 start = off > st_off ? off - st_off : 0;
299 u64 end = off + length > st_off ? off + length - st_off : 0;
304 if (st_len == 0 || length == 0 || end == 0)
307 start = div64_u64_rem(start, st_len, &start_left);
308 if (start_left < st_bytes)
309 start_left = st_bytes - start_left;
313 end = div64_u64_rem(end, st_len, &end_left);
314 if (end_left > st_bytes)
317 CDEBUG(D_READA, "start %llu, end %llu start_left %llu end_left %llu\n",
318 start, end, start_left, end_left);
321 bytes_count = end_left - (st_bytes - start_left);
323 bytes_count = start_left +
324 st_bytes * (end - start - 1) + end_left;
327 "st_off %llu, st_len %llu st_bytes %llu off %llu length %llu bytescount %llu\n",
328 st_off, st_len, st_bytes, off, length, bytes_count);
333 static unsigned long ria_page_count(struct ra_io_arg *ria)
335 loff_t length_bytes = ria->ria_end_idx >= ria->ria_start_idx ?
336 (loff_t)(ria->ria_end_idx -
337 ria->ria_start_idx + 1) << PAGE_SHIFT : 0;
340 if (ria->ria_length > ria->ria_bytes && ria->ria_bytes &&
341 (ria->ria_length & ~PAGE_MASK || ria->ria_bytes & ~PAGE_MASK ||
342 ria->ria_stoff & ~PAGE_MASK)) {
343 /* Over-estimate un-aligned page stride read */
344 unsigned long pg_count = ((ria->ria_bytes +
345 PAGE_SIZE - 1) >> PAGE_SHIFT) + 1;
346 pg_count *= length_bytes / ria->ria_length + 1;
350 bytes_count = stride_byte_count(ria->ria_stoff, ria->ria_length,
352 (loff_t)ria->ria_start_idx<<PAGE_SHIFT,
354 return (bytes_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
357 static pgoff_t ras_align(struct ll_readahead_state *ras, pgoff_t index)
359 unsigned opt_size = min(ras->ras_window_pages, ras->ras_rpc_pages);
363 return index - (index % opt_size);
366 /* Check whether the index is in the defined ra-window */
367 static bool ras_inside_ra_window(pgoff_t idx, struct ra_io_arg *ria)
369 loff_t pos = (loff_t)idx << PAGE_SHIFT;
371 /* If ria_length == ria_bytes, it means non-stride I/O mode,
372 * idx should always inside read-ahead window in this case
373 * For stride I/O mode, just check whether the idx is inside
376 if (ria->ria_length == 0 || ria->ria_length == ria->ria_bytes)
379 if (pos >= ria->ria_stoff) {
382 div64_u64_rem(pos - ria->ria_stoff, ria->ria_length, &offset);
384 if (offset < ria->ria_bytes ||
385 (ria->ria_length - offset) < PAGE_SIZE)
387 } else if (pos + PAGE_SIZE > ria->ria_stoff) {
395 ll_read_ahead_pages(const struct lu_env *env, struct cl_io *io,
396 struct cl_page_list *queue, struct ll_readahead_state *ras,
397 struct ra_io_arg *ria, pgoff_t *ra_end, pgoff_t skip_index)
399 struct cl_read_ahead ra = { 0 };
400 /* busy page count is per stride */
401 int rc = 0, count = 0, busy_page_count = 0;
404 LASSERT(ria != NULL);
407 for (page_idx = ria->ria_start_idx;
408 page_idx <= ria->ria_end_idx && ria->ria_reserved > 0;
410 if (skip_index && page_idx == skip_index)
412 if (ras_inside_ra_window(page_idx, ria)) {
413 if (ra.cra_end_idx == 0 || ra.cra_end_idx < page_idx) {
417 * Do not shrink ria_end_idx at any case until
418 * the minimum end of current read is covered.
420 * Do not extend read lock accross stripe if
421 * lock contention detected.
423 if (ra.cra_contention &&
424 page_idx > ria->ria_end_idx_min) {
425 ria->ria_end_idx = *ra_end;
429 cl_read_ahead_release(env, &ra);
431 rc = cl_io_read_ahead(env, io, page_idx, &ra);
436 * Only shrink ria_end_idx if the matched
437 * LDLM lock doesn't cover more.
439 if (page_idx > ra.cra_end_idx) {
440 ria->ria_end_idx = ra.cra_end_idx;
444 CDEBUG(D_READA, "idx: %lu, ra: %lu, rpc: %lu\n",
445 page_idx, ra.cra_end_idx,
447 LASSERTF(ra.cra_end_idx >= page_idx,
448 "object: %p, indcies %lu / %lu\n",
449 io->ci_obj, ra.cra_end_idx, page_idx);
450 /* update read ahead RPC size.
451 * NB: it's racy but doesn't matter */
452 if (ras->ras_rpc_pages != ra.cra_rpc_pages &&
453 ra.cra_rpc_pages > 0)
454 ras->ras_rpc_pages = ra.cra_rpc_pages;
456 /* trim it to align with optimal RPC size */
457 end_idx = ras_align(ras, ria->ria_end_idx + 1);
458 if (end_idx > 0 && !ria->ria_eof)
459 ria->ria_end_idx = end_idx - 1;
461 if (ria->ria_end_idx < ria->ria_end_idx_min)
462 ria->ria_end_idx = ria->ria_end_idx_min;
464 if (page_idx > ria->ria_end_idx)
467 /* If the page is inside the read-ahead window */
468 rc = ll_read_ahead_page(env, io, queue, page_idx,
470 if (rc < 0 && rc != -EBUSY)
475 "skip busy page: %lu\n", page_idx);
476 /* For page unaligned readahead the first
477 * last pages of each region can be read by
478 * another reader on the same node, and so
479 * may be busy. So only stop for > 2 busy
481 if (busy_page_count > 2)
486 /* Only subtract from reserve & count the page if we
487 * really did readahead on that page. */
492 } else if (stride_io_mode(ras)) {
493 /* If it is not in the read-ahead window, and it is
494 * read-ahead mode, then check whether it should skip
497 loff_t pos = (loff_t)page_idx << PAGE_SHIFT;
500 div64_u64_rem(pos - ria->ria_stoff, ria->ria_length,
502 if (offset >= ria->ria_bytes) {
503 pos += (ria->ria_length - offset);
504 if ((pos >> PAGE_SHIFT) >= page_idx + 1)
505 page_idx = (pos >> PAGE_SHIFT) - 1;
508 "Stride: jump %llu pages to %lu\n",
509 ria->ria_length - offset, page_idx);
515 cl_read_ahead_release(env, &ra);
520 static void ll_readahead_work_free(struct ll_readahead_work *work)
522 fput(work->lrw_file);
526 static void ll_readahead_handle_work(struct work_struct *wq);
527 static void ll_readahead_work_add(struct inode *inode,
528 struct ll_readahead_work *work)
530 INIT_WORK(&work->lrw_readahead_work, ll_readahead_handle_work);
531 queue_work(ll_i2sbi(inode)->ll_ra_info.ll_readahead_wq,
532 &work->lrw_readahead_work);
535 static int ll_readahead_file_kms(const struct lu_env *env,
536 struct cl_io *io, __u64 *kms)
538 struct cl_object *clob;
540 struct cl_attr *attr = vvp_env_thread_attr(env);
544 inode = vvp_object_inode(clob);
546 cl_object_attr_lock(clob);
547 ret = cl_object_attr_get(env, clob, attr);
548 cl_object_attr_unlock(clob);
553 *kms = attr->cat_kms;
557 static void ll_readahead_handle_work(struct work_struct *wq)
559 struct ll_readahead_work *work;
562 struct ra_io_arg *ria;
564 struct ll_file_data *fd;
565 struct ll_readahead_state *ras;
567 struct cl_2queue *queue;
568 pgoff_t ra_end_idx = 0;
569 unsigned long pages, pages_min = 0;
574 struct ll_sb_info *sbi;
576 work = container_of(wq, struct ll_readahead_work,
578 fd = work->lrw_file->private_data;
580 file = work->lrw_file;
581 inode = file_inode(file);
582 sbi = ll_i2sbi(inode);
584 env = cl_env_alloc(&refcheck, LCT_NOREF);
586 GOTO(out_free_work, rc = PTR_ERR(env));
588 io = vvp_env_thread_io(env);
589 ll_io_init(io, file, CIT_READ, NULL);
591 rc = ll_readahead_file_kms(env, io, &kms);
593 GOTO(out_put_env, rc);
596 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
597 GOTO(out_put_env, rc = 0);
600 ria = &ll_env_info(env)->lti_ria;
601 memset(ria, 0, sizeof(*ria));
603 ria->ria_start_idx = work->lrw_start_idx;
604 /* Truncate RA window to end of file */
605 eof_index = (pgoff_t)(kms - 1) >> PAGE_SHIFT;
606 if (eof_index <= work->lrw_end_idx) {
607 work->lrw_end_idx = eof_index;
610 if (work->lrw_end_idx <= work->lrw_start_idx)
611 GOTO(out_put_env, rc = 0);
613 ria->ria_end_idx = work->lrw_end_idx;
614 pages = ria->ria_end_idx - ria->ria_start_idx + 1;
615 ria->ria_reserved = ll_ra_count_get(sbi, ria,
616 ria_page_count(ria), pages_min);
619 "async reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
620 ria->ria_reserved, pages, pages_min,
621 atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
622 ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
624 if (ria->ria_reserved < pages) {
625 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
626 if (PAGES_TO_MiB(ria->ria_reserved) < 1) {
627 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
628 GOTO(out_put_env, rc = 0);
632 rc = cl_io_rw_init(env, io, CIT_READ, ria->ria_start_idx, pages);
634 GOTO(out_put_env, rc);
636 /* overwrite jobid inited in vvp_io_init() */
637 if (strncmp(ll_i2info(inode)->lli_jobid, work->lrw_jobid,
638 sizeof(work->lrw_jobid)))
639 memcpy(ll_i2info(inode)->lli_jobid, work->lrw_jobid,
640 sizeof(work->lrw_jobid));
642 vvp_env_io(env)->vui_fd = fd;
643 io->ci_state = CIS_LOCKED;
644 io->ci_async_readahead = true;
645 rc = cl_io_start(env, io);
647 GOTO(out_io_fini, rc);
649 queue = &io->ci_queue;
650 cl_2queue_init(queue);
652 rc = ll_read_ahead_pages(env, io, &queue->c2_qin, ras, ria,
654 if (ria->ria_reserved != 0)
655 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
656 if (queue->c2_qin.pl_nr > 0) {
657 int count = queue->c2_qin.pl_nr;
659 rc = cl_io_submit_rw(env, io, CRT_READ, queue);
661 task_io_account_read(PAGE_SIZE * count);
663 if (ria->ria_end_idx == ra_end_idx && ra_end_idx == (kms >> PAGE_SHIFT))
664 ll_ra_stats_inc(inode, RA_STAT_EOF);
666 if (ra_end_idx != ria->ria_end_idx)
667 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
669 /* TODO: discard all pages until page reinit route is implemented */
670 cl_page_list_discard(env, io, &queue->c2_qin);
672 /* Unlock unsent read pages in case of error. */
673 cl_page_list_disown(env, io, &queue->c2_qin);
675 cl_2queue_fini(env, queue);
680 cl_env_put(env, &refcheck);
683 ll_ra_stats_inc_sbi(ll_i2sbi(inode), RA_STAT_ASYNC);
684 atomic_dec(&sbi->ll_ra_info.ra_async_inflight);
685 ll_readahead_work_free(work);
688 static int ll_readahead(const struct lu_env *env, struct cl_io *io,
689 struct cl_page_list *queue,
690 struct ll_readahead_state *ras, bool hit,
691 struct file *file, pgoff_t skip_index)
693 struct vvp_io *vio = vvp_env_io(env);
694 struct ll_thread_info *lti = ll_env_info(env);
695 unsigned long pages, pages_min = 0;
696 pgoff_t ra_end_idx = 0, start_idx = 0, end_idx = 0;
698 struct ra_io_arg *ria = <i->lti_ria;
699 struct cl_object *clob;
702 struct ll_sb_info *sbi;
703 struct ll_ra_info *ra;
710 inode = vvp_object_inode(clob);
711 sbi = ll_i2sbi(inode);
712 ra = &sbi->ll_ra_info;
715 * In case we have a limited max_cached_mb, readahead
716 * should be stopped if it have run out of all LRU slots.
718 if (atomic_read(&ra->ra_cur_pages) >= sbi->ll_cache->ccc_lru_max) {
719 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
723 memset(ria, 0, sizeof(*ria));
724 ret = ll_readahead_file_kms(env, io, &kms);
729 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
733 spin_lock(&ras->ras_lock);
736 * Note: other thread might rollback the ras_next_readahead_idx,
737 * if it can not get the full size of prepared pages, see the
738 * end of this function. For stride read ahead, it needs to
739 * make sure the offset is no less than ras_stride_offset,
740 * so that stride read ahead can work correctly.
742 if (stride_io_mode(ras))
743 start_idx = max_t(pgoff_t, ras->ras_next_readahead_idx,
744 ras->ras_stride_offset >> PAGE_SHIFT);
746 start_idx = ras->ras_next_readahead_idx;
748 if (ras->ras_window_pages > 0)
749 end_idx = ras->ras_window_start_idx + ras->ras_window_pages - 1;
752 end_idx = start_idx + ras->ras_window_pages - 1;
754 /* Enlarge the RA window to encompass the full read */
755 if (vio->vui_ra_valid &&
756 end_idx < vio->vui_ra_start_idx + vio->vui_ra_pages - 1)
757 end_idx = vio->vui_ra_start_idx + vio->vui_ra_pages - 1;
762 /* Truncate RA window to end of file */
763 eof_index = (pgoff_t)((kms - 1) >> PAGE_SHIFT);
764 if (eof_index <= end_idx) {
769 ria->ria_start_idx = start_idx;
770 ria->ria_end_idx = end_idx;
771 /* If stride I/O mode is detected, get stride window*/
772 if (stride_io_mode(ras)) {
773 ria->ria_stoff = ras->ras_stride_offset;
774 ria->ria_length = ras->ras_stride_length;
775 ria->ria_bytes = ras->ras_stride_bytes;
777 spin_unlock(&ras->ras_lock);
780 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
783 pages = ria_page_count(ria);
785 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
790 CDEBUG(D_READA, DFID": ria: %lu/%lu, bead: %lu/%lu, hit: %d\n",
791 PFID(lu_object_fid(&clob->co_lu)),
792 ria->ria_start_idx, ria->ria_end_idx,
793 vio->vui_ra_valid ? vio->vui_ra_start_idx : 0,
794 vio->vui_ra_valid ? vio->vui_ra_pages : 0,
797 /* at least to extend the readahead window to cover current read */
798 if (!hit && vio->vui_ra_valid &&
799 vio->vui_ra_start_idx + vio->vui_ra_pages > ria->ria_start_idx) {
800 ria->ria_end_idx_min =
801 vio->vui_ra_start_idx + vio->vui_ra_pages - 1;
802 pages_min = vio->vui_ra_start_idx + vio->vui_ra_pages -
805 * For performance reason, exceeding @ra_max_pages
806 * are allowed, but this should be limited with RPC
807 * size in case a large block size read issued. Trim
810 pages_min = min(pages_min, ras->ras_rpc_pages -
811 (ria->ria_start_idx % ras->ras_rpc_pages));
814 /* don't over reserved for mmap range read */
818 ria->ria_reserved = ll_ra_count_get(ll_i2sbi(inode), ria, pages,
820 if (ria->ria_reserved < pages)
821 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
823 CDEBUG(D_READA, "reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
824 ria->ria_reserved, pages, pages_min,
825 atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
826 ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
828 ret = ll_read_ahead_pages(env, io, queue, ras, ria, &ra_end_idx,
830 if (ria->ria_reserved != 0)
831 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
833 if (ra_end_idx == end_idx && ra_end_idx == (kms >> PAGE_SHIFT))
834 ll_ra_stats_inc(inode, RA_STAT_EOF);
837 "ra_end_idx = %lu end_idx = %lu stride end = %lu pages = %d\n",
838 ra_end_idx, end_idx, ria->ria_end_idx, ret);
840 if (ra_end_idx != end_idx)
841 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
842 if (ra_end_idx > 0) {
843 /* update the ras so that the next read-ahead tries from
844 * where we left off. */
845 spin_lock(&ras->ras_lock);
846 ras->ras_next_readahead_idx = ra_end_idx + 1;
847 spin_unlock(&ras->ras_lock);
854 static int ll_readpages(const struct lu_env *env, struct cl_io *io,
855 struct cl_page_list *queue,
856 pgoff_t start, pgoff_t end)
865 ret = ll_readahead_file_kms(env, io, &kms);
873 unsigned long end_index;
875 end_index = (unsigned long)((kms - 1) >> PAGE_SHIFT);
876 if (end_index <= end)
880 for (page_idx = start; page_idx <= end; page_idx++) {
881 ret= ll_read_ahead_page(env, io, queue, page_idx,
885 else if (ret == 0) /* ret 1 is already uptodate */
889 RETURN(count > 0 ? count : ret);
892 static void ras_set_start(struct ll_readahead_state *ras, pgoff_t index)
894 ras->ras_window_start_idx = ras_align(ras, index);
897 /* called with the ras_lock held or from places where it doesn't matter */
898 static void ras_reset(struct ll_readahead_state *ras, pgoff_t index)
900 ras->ras_consecutive_requests = 0;
901 ras->ras_consecutive_bytes = 0;
902 ras->ras_window_pages = 0;
903 ras_set_start(ras, index);
904 ras->ras_next_readahead_idx = max(ras->ras_window_start_idx, index + 1);
909 /* called with the ras_lock held or from places where it doesn't matter */
910 static void ras_stride_reset(struct ll_readahead_state *ras)
912 ras->ras_consecutive_stride_requests = 0;
913 ras->ras_stride_length = 0;
914 ras->ras_stride_bytes = 0;
918 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
920 spin_lock_init(&ras->ras_lock);
921 ras->ras_rpc_pages = PTLRPC_MAX_BRW_PAGES;
923 ras->ras_last_read_end_bytes = 0;
924 ras->ras_requests = 0;
925 ras->ras_range_min_start_idx = 0;
926 ras->ras_range_max_end_idx = 0;
927 ras->ras_range_requests = 0;
928 ras->ras_last_range_pages = 0;
932 * Check whether the read request is in the stride window.
933 * If it is in the stride window, return true, otherwise return false.
935 static bool read_in_stride_window(struct ll_readahead_state *ras,
936 loff_t pos, loff_t count)
940 if (ras->ras_stride_length == 0 || ras->ras_stride_bytes == 0 ||
941 ras->ras_stride_bytes == ras->ras_stride_length)
944 stride_gap = pos - ras->ras_last_read_end_bytes - 1;
946 /* If it is contiguous read */
948 return ras->ras_consecutive_bytes + count <=
949 ras->ras_stride_bytes;
951 /* Otherwise check the stride by itself */
952 return (ras->ras_stride_length - ras->ras_stride_bytes) == stride_gap &&
953 ras->ras_consecutive_bytes == ras->ras_stride_bytes &&
954 count <= ras->ras_stride_bytes;
957 static void ras_init_stride_detector(struct ll_readahead_state *ras,
958 loff_t pos, loff_t count)
960 loff_t stride_gap = pos - ras->ras_last_read_end_bytes - 1;
962 LASSERT(ras->ras_consecutive_stride_requests == 0);
964 if (pos <= ras->ras_last_read_end_bytes) {
965 /*Reset stride window for forward read*/
966 ras_stride_reset(ras);
970 ras->ras_stride_bytes = ras->ras_consecutive_bytes;
971 ras->ras_stride_length = stride_gap + ras->ras_consecutive_bytes;
972 ras->ras_consecutive_stride_requests++;
973 ras->ras_stride_offset = pos;
979 stride_page_count(struct ll_readahead_state *ras, loff_t len)
982 stride_byte_count(ras->ras_stride_offset,
983 ras->ras_stride_length, ras->ras_stride_bytes,
984 ras->ras_window_start_idx << PAGE_SHIFT, len);
986 return (bytes_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
989 /* Stride Read-ahead window will be increased inc_len according to
990 * stride I/O pattern */
991 static void ras_stride_increase_window(struct ll_readahead_state *ras,
992 struct ll_ra_info *ra, loff_t inc_bytes)
994 loff_t window_bytes, stride_bytes;
999 /* temporarily store in page units to reduce LASSERT() cost below */
1000 end = ras->ras_window_start_idx + ras->ras_window_pages;
1002 LASSERT(ras->ras_stride_length > 0);
1003 LASSERTF(end >= (ras->ras_stride_offset >> PAGE_SHIFT),
1004 "window_start_idx %lu, window_pages %lu stride_offset %llu\n",
1005 ras->ras_window_start_idx, ras->ras_window_pages,
1006 ras->ras_stride_offset);
1009 if (end <= ras->ras_stride_offset)
1012 stride_bytes = end - ras->ras_stride_offset;
1014 div64_u64_rem(stride_bytes, ras->ras_stride_length, &left_bytes);
1015 window_bytes = (ras->ras_window_pages << PAGE_SHIFT);
1016 if (left_bytes < ras->ras_stride_bytes) {
1017 if (ras->ras_stride_bytes - left_bytes >= inc_bytes) {
1018 window_bytes += inc_bytes;
1021 window_bytes += (ras->ras_stride_bytes - left_bytes);
1022 inc_bytes -= (ras->ras_stride_bytes - left_bytes);
1025 window_bytes += (ras->ras_stride_length - left_bytes);
1028 LASSERT(ras->ras_stride_bytes != 0);
1030 step = div64_u64_rem(inc_bytes, ras->ras_stride_bytes, &left_bytes);
1032 window_bytes += step * ras->ras_stride_length + left_bytes;
1033 LASSERT(window_bytes > 0);
1036 if (stride_page_count(ras, window_bytes) <=
1037 ra->ra_max_pages_per_file || ras->ras_window_pages == 0)
1038 ras->ras_window_pages = (window_bytes >> PAGE_SHIFT);
1040 LASSERT(ras->ras_window_pages > 0);
1045 static void ras_increase_window(struct inode *inode,
1046 struct ll_readahead_state *ras,
1047 struct ll_ra_info *ra)
1049 /* The stretch of ra-window should be aligned with max rpc_size
1050 * but current clio architecture does not support retrieve such
1051 * information from lower layer. FIXME later
1053 if (stride_io_mode(ras)) {
1054 ras_stride_increase_window(ras, ra,
1055 (loff_t)ras->ras_rpc_pages << PAGE_SHIFT);
1057 pgoff_t window_pages;
1059 window_pages = min(ras->ras_window_pages + ras->ras_rpc_pages,
1060 ra->ra_max_pages_per_file);
1061 if (window_pages < ras->ras_rpc_pages)
1062 ras->ras_window_pages = window_pages;
1064 ras->ras_window_pages = ras_align(ras, window_pages);
1069 * Seek within 8 pages are considered as sequential read for now.
1071 static inline bool is_loose_seq_read(struct ll_readahead_state *ras, loff_t pos)
1073 return pos_in_window(pos, ras->ras_last_read_end_bytes,
1074 8UL << PAGE_SHIFT, 8UL << PAGE_SHIFT);
1077 static inline bool is_loose_mmap_read(struct ll_sb_info *sbi,
1078 struct ll_readahead_state *ras,
1081 unsigned long range_pages = sbi->ll_ra_info.ra_range_pages;
1083 return pos_in_window(pos, ras->ras_last_read_end_bytes,
1084 range_pages << PAGE_SHIFT,
1085 range_pages << PAGE_SHIFT);
1089 * We have observed slow mmap read performances for some
1090 * applications. The problem is if access pattern is neither
1091 * sequential nor stride, but could be still adjacent in a
1092 * small range and then seek a random position.
1094 * So the pattern could be something like this:
1096 * [1M data] [hole] [0.5M data] [hole] [0.7M data] [1M data]
1099 * Every time an application reads mmap data, it may not only
1100 * read a single 4KB page, but aslo a cluster of nearby pages in
1101 * a range(e.g. 1MB) of the first page after a cache miss.
1103 * The readahead engine is modified to track the range size of
1104 * a cluster of mmap reads, so that after a seek and/or cache miss,
1105 * the range size is used to efficiently prefetch multiple pages
1106 * in a single RPC rather than many small RPCs.
1108 static void ras_detect_cluster_range(struct ll_readahead_state *ras,
1109 struct ll_sb_info *sbi,
1110 unsigned long pos, unsigned long count)
1112 pgoff_t last_pages, pages;
1113 pgoff_t end_idx = (pos + count - 1) >> PAGE_SHIFT;
1115 last_pages = ras->ras_range_max_end_idx -
1116 ras->ras_range_min_start_idx + 1;
1117 /* First time come here */
1118 if (!ras->ras_range_max_end_idx)
1121 /* Random or Stride read */
1122 if (!is_loose_mmap_read(sbi, ras, pos))
1125 ras->ras_range_requests++;
1126 if (ras->ras_range_max_end_idx < end_idx)
1127 ras->ras_range_max_end_idx = end_idx;
1129 if (ras->ras_range_min_start_idx > (pos >> PAGE_SHIFT))
1130 ras->ras_range_min_start_idx = pos >> PAGE_SHIFT;
1132 /* Out of range, consider it as random or stride */
1133 pages = ras->ras_range_max_end_idx -
1134 ras->ras_range_min_start_idx + 1;
1135 if (pages <= sbi->ll_ra_info.ra_range_pages)
1138 ras->ras_last_range_pages = last_pages;
1139 ras->ras_range_requests = 0;
1140 ras->ras_range_min_start_idx = pos >> PAGE_SHIFT;
1141 ras->ras_range_max_end_idx = end_idx;
1144 static void ras_detect_read_pattern(struct ll_readahead_state *ras,
1145 struct ll_sb_info *sbi,
1146 loff_t pos, size_t count, bool mmap)
1148 bool stride_detect = false;
1149 pgoff_t index = pos >> PAGE_SHIFT;
1152 * Reset the read-ahead window in two cases. First when the app seeks
1153 * or reads to some other part of the file. Secondly if we get a
1154 * read-ahead miss that we think we've previously issued. This can
1155 * be a symptom of there being so many read-ahead pages that the VM
1156 * is reclaiming it before we get to it.
1158 if (!is_loose_seq_read(ras, pos)) {
1159 /* Check whether it is in stride I/O mode */
1160 if (!read_in_stride_window(ras, pos, count)) {
1161 if (ras->ras_consecutive_stride_requests == 0)
1162 ras_init_stride_detector(ras, pos, count);
1164 ras_stride_reset(ras);
1165 ras->ras_consecutive_bytes = 0;
1166 ras_reset(ras, index);
1168 ras->ras_consecutive_bytes = 0;
1169 ras->ras_consecutive_requests = 0;
1170 if (++ras->ras_consecutive_stride_requests > 1)
1171 stride_detect = true;
1174 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
1175 } else if (stride_io_mode(ras)) {
1177 * If this is contiguous read but in stride I/O mode
1178 * currently, check whether stride step still is valid,
1179 * if invalid, it will reset the stride ra window to
1182 if (!read_in_stride_window(ras, pos, count)) {
1183 ras_stride_reset(ras);
1184 ras->ras_window_pages = 0;
1185 ras->ras_next_readahead_idx = index;
1189 ras->ras_consecutive_bytes += count;
1191 pgoff_t idx = ras->ras_consecutive_bytes >> PAGE_SHIFT;
1192 unsigned long ra_range_pages =
1193 max_t(unsigned long, RA_MIN_MMAP_RANGE_PAGES,
1194 sbi->ll_ra_info.ra_range_pages);
1196 if ((idx >= ra_range_pages &&
1197 idx % ra_range_pages == 0) || stride_detect)
1198 ras->ras_need_increase_window = true;
1199 } else if ((ras->ras_consecutive_requests > 1 || stride_detect)) {
1200 ras->ras_need_increase_window = true;
1203 ras->ras_last_read_end_bytes = pos + count - 1;
1206 void ll_ras_enter(struct file *f, loff_t pos, size_t count)
1208 struct ll_file_data *fd = f->private_data;
1209 struct ll_readahead_state *ras = &fd->fd_ras;
1210 struct inode *inode = file_inode(f);
1211 unsigned long index = pos >> PAGE_SHIFT;
1212 struct ll_sb_info *sbi = ll_i2sbi(inode);
1214 spin_lock(&ras->ras_lock);
1215 ras->ras_requests++;
1216 ras->ras_consecutive_requests++;
1217 ras->ras_need_increase_window = false;
1218 ras->ras_no_miss_check = false;
1220 * On the second access to a file smaller than the tunable
1221 * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1222 * file up to ra_max_pages_per_file. This is simply a best effort
1223 * and only occurs once per open file. Normal RA behavior is reverted
1224 * to for subsequent IO.
1226 if (ras->ras_requests >= 2) {
1228 struct ll_ra_info *ra = &sbi->ll_ra_info;
1230 kms_pages = (i_size_read(inode) + PAGE_SIZE - 1) >>
1233 CDEBUG(D_READA, "kmsp %llu mwp %lu mp %lu\n", kms_pages,
1234 ra->ra_max_read_ahead_whole_pages,
1235 ra->ra_max_pages_per_file);
1238 kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1239 ras->ras_window_start_idx = 0;
1240 ras->ras_next_readahead_idx = index + 1;
1241 ras->ras_window_pages = min(ra->ra_max_pages_per_file,
1242 ra->ra_max_read_ahead_whole_pages);
1243 ras->ras_no_miss_check = true;
1244 GOTO(out_unlock, 0);
1247 ras_detect_read_pattern(ras, sbi, pos, count, false);
1249 spin_unlock(&ras->ras_lock);
1252 static bool index_in_stride_window(struct ll_readahead_state *ras,
1255 loff_t pos = (loff_t)index << PAGE_SHIFT;
1257 if (ras->ras_stride_length == 0 || ras->ras_stride_bytes == 0 ||
1258 ras->ras_stride_bytes == ras->ras_stride_length)
1261 if (pos >= ras->ras_stride_offset) {
1264 div64_u64_rem(pos - ras->ras_stride_offset,
1265 ras->ras_stride_length, &offset);
1266 if (offset < ras->ras_stride_bytes ||
1267 ras->ras_stride_length - offset < PAGE_SIZE)
1269 } else if (ras->ras_stride_offset - pos < PAGE_SIZE) {
1277 * ll_ras_enter() is used to detect read pattern according to pos and count.
1279 * ras_update() is used to detect cache miss and
1280 * reset window or increase window accordingly
1282 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
1283 struct ll_readahead_state *ras, pgoff_t index,
1284 enum ras_update_flags flags, struct cl_io *io)
1286 struct ll_ra_info *ra = &sbi->ll_ra_info;
1287 bool hit = flags & LL_RAS_HIT;
1290 spin_lock(&ras->ras_lock);
1293 CDEBUG(D_READA, DFID " pages at %lu miss.\n",
1294 PFID(ll_inode2fid(inode)), index);
1295 ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
1298 * The readahead window has been expanded to cover whole
1299 * file size, we don't care whether ra miss happen or not.
1300 * Because we will read whole file to page cache even if
1301 * some pages missed.
1303 if (ras->ras_no_miss_check)
1304 GOTO(out_unlock, 0);
1306 if (io && io->ci_rand_read)
1307 GOTO(out_unlock, 0);
1309 if (io && io->ci_seq_read) {
1311 /* to avoid many small read RPC here */
1312 ras->ras_window_pages = sbi->ll_ra_info.ra_range_pages;
1313 ll_ra_stats_inc_sbi(sbi, RA_STAT_MMAP_RANGE_READ);
1318 if (flags & LL_RAS_MMAP) {
1319 unsigned long ra_pages;
1321 ras_detect_cluster_range(ras, sbi, index << PAGE_SHIFT,
1323 ras_detect_read_pattern(ras, sbi, (loff_t)index << PAGE_SHIFT,
1326 /* we did not detect anything but we could prefetch */
1327 if (!ras->ras_need_increase_window &&
1328 ras->ras_window_pages <= sbi->ll_ra_info.ra_range_pages &&
1329 ras->ras_range_requests >= 2) {
1331 ra_pages = max_t(unsigned long,
1332 RA_MIN_MMAP_RANGE_PAGES,
1333 ras->ras_last_range_pages);
1334 if (index < ra_pages / 2)
1337 index -= ra_pages / 2;
1338 ras->ras_window_pages = ra_pages;
1339 ll_ra_stats_inc_sbi(sbi,
1340 RA_STAT_MMAP_RANGE_READ);
1342 ras->ras_window_pages = 0;
1348 if (!hit && ras->ras_window_pages &&
1349 index < ras->ras_next_readahead_idx &&
1350 pos_in_window(index, ras->ras_window_start_idx, 0,
1351 ras->ras_window_pages)) {
1352 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
1353 ras->ras_need_increase_window = false;
1355 if (index_in_stride_window(ras, index) &&
1356 stride_io_mode(ras)) {
1358 * if (index != ras->ras_last_readpage + 1)
1359 * ras->ras_consecutive_pages = 0;
1361 ras_reset(ras, index);
1364 * If stride-RA hit cache miss, the stride
1365 * detector will not be reset to avoid the
1366 * overhead of redetecting read-ahead mode,
1367 * but on the condition that the stride window
1368 * is still intersect with normal sequential
1369 * read-ahead window.
1371 if (ras->ras_window_start_idx < ras->ras_stride_offset)
1372 ras_stride_reset(ras);
1376 * Reset both stride window and normal RA
1379 ras_reset(ras, index);
1380 /* ras->ras_consecutive_pages++; */
1381 ras->ras_consecutive_bytes = 0;
1382 ras_stride_reset(ras);
1383 GOTO(out_unlock, 0);
1388 ras_set_start(ras, index);
1390 if (stride_io_mode(ras)) {
1391 /* Since stride readahead is sentivite to the offset
1392 * of read-ahead, so we use original offset here,
1393 * instead of ras_window_start_idx, which is RPC aligned.
1395 ras->ras_next_readahead_idx = max(index + 1,
1396 ras->ras_next_readahead_idx);
1397 ras->ras_window_start_idx =
1398 max_t(pgoff_t, ras->ras_window_start_idx,
1399 ras->ras_stride_offset >> PAGE_SHIFT);
1401 if (ras->ras_next_readahead_idx < ras->ras_window_start_idx)
1402 ras->ras_next_readahead_idx = ras->ras_window_start_idx;
1404 ras->ras_next_readahead_idx = index + 1;
1407 if (ras->ras_need_increase_window) {
1408 ras_increase_window(inode, ras, ra);
1409 ras->ras_need_increase_window = false;
1414 spin_unlock(&ras->ras_lock);
1417 int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
1419 struct inode *inode = vmpage->mapping->host;
1420 struct ll_inode_info *lli = ll_i2info(inode);
1423 struct cl_page *page;
1424 struct cl_object *clob;
1425 bool redirtied = false;
1426 bool unlocked = false;
1431 LASSERT(PageLocked(vmpage));
1432 LASSERT(!PageWriteback(vmpage));
1434 LASSERT(ll_i2dtexp(inode) != NULL);
1436 env = cl_env_get(&refcheck);
1438 GOTO(out, result = PTR_ERR(env));
1440 clob = ll_i2info(inode)->lli_clob;
1441 LASSERT(clob != NULL);
1443 io = vvp_env_thread_io(env);
1445 io->ci_ignore_layout = 1;
1446 result = cl_io_init(env, io, CIT_MISC, clob);
1448 page = cl_page_find(env, clob, vmpage->index,
1449 vmpage, CPT_CACHEABLE);
1450 if (!IS_ERR(page)) {
1451 lu_ref_add(&page->cp_reference, "writepage",
1453 cl_page_assume(env, io, page);
1454 result = cl_page_flush(env, io, page);
1457 * Re-dirty page on error so it retries write,
1458 * but not in case when IO has actually
1459 * occurred and completed with an error.
1461 if (!PageError(vmpage)) {
1462 redirty_page_for_writepage(wbc, vmpage);
1467 cl_page_disown(env, io, page);
1469 lu_ref_del(&page->cp_reference,
1470 "writepage", current);
1471 cl_page_put(env, page);
1473 result = PTR_ERR(page);
1476 cl_io_fini(env, io);
1478 if (redirtied && wbc->sync_mode == WB_SYNC_ALL) {
1479 loff_t offset = cl_offset(clob, vmpage->index);
1481 /* Flush page failed because the extent is being written out.
1482 * Wait for the write of extent to be finished to avoid
1483 * breaking kernel which assumes ->writepage should mark
1484 * PageWriteback or clean the page. */
1485 result = cl_sync_file_range(inode, offset,
1486 offset + PAGE_SIZE - 1,
1489 /* actually we may have written more than one page.
1490 * decreasing this page because the caller will count
1492 wbc->nr_to_write -= result - 1;
1497 cl_env_put(env, &refcheck);
1502 if (!lli->lli_async_rc)
1503 lli->lli_async_rc = result;
1504 SetPageError(vmpage);
1506 unlock_page(vmpage);
1511 int ll_writepages(struct address_space *mapping, struct writeback_control *wbc)
1513 struct inode *inode = mapping->host;
1516 enum cl_fsync_mode mode;
1517 int range_whole = 0;
1521 if (wbc->range_cyclic) {
1522 start = (loff_t)mapping->writeback_index << PAGE_SHIFT;
1523 end = OBD_OBJECT_EOF;
1525 start = wbc->range_start;
1526 end = wbc->range_end;
1527 if (end == LLONG_MAX) {
1528 end = OBD_OBJECT_EOF;
1529 range_whole = start == 0;
1533 mode = CL_FSYNC_NONE;
1534 if (wbc->sync_mode == WB_SYNC_ALL)
1535 mode = CL_FSYNC_LOCAL;
1537 if (ll_i2info(inode)->lli_clob == NULL)
1540 /* for directio, it would call writepages() to evict cached pages
1541 * inside the IO context of write, which will cause deadlock at
1542 * layout_conf since it waits for active IOs to complete. */
1543 result = cl_sync_file_range(inode, start, end, mode, 1);
1545 wbc->nr_to_write -= result;
1549 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) {
1550 if (end == OBD_OBJECT_EOF)
1551 mapping->writeback_index = 0;
1553 mapping->writeback_index = (end >> PAGE_SHIFT) + 1;
1558 struct ll_cl_context *ll_cl_find(struct file *file)
1560 struct ll_file_data *fd = file->private_data;
1561 struct ll_cl_context *lcc;
1562 struct ll_cl_context *found = NULL;
1564 read_lock(&fd->fd_lock);
1565 list_for_each_entry(lcc, &fd->fd_lccs, lcc_list) {
1566 if (lcc->lcc_cookie == current) {
1571 read_unlock(&fd->fd_lock);
1576 void ll_cl_add(struct file *file, const struct lu_env *env, struct cl_io *io,
1579 struct ll_file_data *fd = file->private_data;
1580 struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1582 memset(lcc, 0, sizeof(*lcc));
1583 INIT_LIST_HEAD(&lcc->lcc_list);
1584 lcc->lcc_cookie = current;
1587 lcc->lcc_type = type;
1589 write_lock(&fd->fd_lock);
1590 list_add(&lcc->lcc_list, &fd->fd_lccs);
1591 write_unlock(&fd->fd_lock);
1594 void ll_cl_remove(struct file *file, const struct lu_env *env)
1596 struct ll_file_data *fd = file->private_data;
1597 struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1599 write_lock(&fd->fd_lock);
1600 list_del_init(&lcc->lcc_list);
1601 write_unlock(&fd->fd_lock);
1604 int ll_io_read_page(const struct lu_env *env, struct cl_io *io,
1605 struct cl_page *page, struct file *file)
1607 struct inode *inode = vvp_object_inode(page->cp_obj);
1608 struct ll_sb_info *sbi = ll_i2sbi(inode);
1609 struct ll_file_data *fd = NULL;
1610 struct ll_readahead_state *ras = NULL;
1611 struct cl_2queue *queue = &io->ci_queue;
1612 struct cl_sync_io *anchor = NULL;
1613 struct vvp_page *vpg;
1614 int rc = 0, rc2 = 0;
1616 pgoff_t io_start_index;
1617 pgoff_t io_end_index;
1621 fd = file->private_data;
1625 vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1626 uptodate = vpg->vpg_defer_uptodate;
1628 if (ll_readahead_enabled(sbi) && !vpg->vpg_ra_updated && ras) {
1629 struct vvp_io *vio = vvp_env_io(env);
1630 enum ras_update_flags flags = 0;
1633 flags |= LL_RAS_HIT;
1634 if (!vio->vui_ra_valid)
1635 flags |= LL_RAS_MMAP;
1636 ras_update(sbi, inode, ras, vvp_index(vpg), flags, io);
1639 cl_2queue_init(queue);
1641 vpg->vpg_ra_used = 1;
1642 cl_page_export(env, page, 1);
1643 cl_page_disown(env, io, page);
1645 anchor = &vvp_env_info(env)->vti_anchor;
1646 cl_sync_io_init(anchor, 1);
1647 page->cp_sync_io = anchor;
1649 cl_2queue_add(queue, page);
1652 io_start_index = cl_index(io->ci_obj, io->u.ci_rw.crw_pos);
1653 io_end_index = cl_index(io->ci_obj, io->u.ci_rw.crw_pos +
1654 io->u.ci_rw.crw_count - 1);
1655 if (ll_readahead_enabled(sbi) && ras && !io->ci_rand_read) {
1656 pgoff_t skip_index = 0;
1658 if (ras->ras_next_readahead_idx < vvp_index(vpg))
1659 skip_index = vvp_index(vpg);
1660 rc2 = ll_readahead(env, io, &queue->c2_qin, ras,
1661 uptodate, file, skip_index);
1662 CDEBUG(D_READA, DFID " %d pages read ahead at %lu\n",
1663 PFID(ll_inode2fid(inode)), rc2, vvp_index(vpg));
1664 } else if (vvp_index(vpg) == io_start_index &&
1665 io_end_index - io_start_index > 0) {
1666 rc2 = ll_readpages(env, io, &queue->c2_qin, io_start_index + 1,
1668 CDEBUG(D_READA, DFID " %d pages read at %lu\n",
1669 PFID(ll_inode2fid(inode)), rc2, vvp_index(vpg));
1672 if (queue->c2_qin.pl_nr > 0) {
1673 int count = queue->c2_qin.pl_nr;
1674 rc = cl_io_submit_rw(env, io, CRT_READ, queue);
1676 task_io_account_read(PAGE_SIZE * count);
1680 if (anchor != NULL && !cl_page_is_owned(page, io)) { /* have sent */
1681 rc = cl_sync_io_wait(env, anchor, 0);
1683 cl_page_assume(env, io, page);
1684 cl_page_list_del(env, &queue->c2_qout, page);
1686 if (!PageUptodate(cl_page_vmpage(page))) {
1687 /* Failed to read a mirror, discard this page so that
1688 * new page can be created with new mirror.
1690 * TODO: this is not needed after page reinit
1691 * route is implemented */
1692 cl_page_discard(env, io, page);
1694 cl_page_disown(env, io, page);
1697 /* TODO: discard all pages until page reinit route is implemented */
1698 cl_page_list_discard(env, io, &queue->c2_qin);
1700 /* Unlock unsent read pages in case of error. */
1701 cl_page_list_disown(env, io, &queue->c2_qin);
1703 cl_2queue_fini(env, queue);
1709 * Possible return value:
1710 * 0 no async readahead triggered and fast read could not be used.
1711 * 1 no async readahead, but fast read could be used.
1712 * 2 async readahead triggered and fast read could be used too.
1715 static int kickoff_async_readahead(struct file *file, unsigned long pages)
1717 struct ll_readahead_work *lrw;
1718 struct inode *inode = file_inode(file);
1719 struct ll_sb_info *sbi = ll_i2sbi(inode);
1720 struct ll_file_data *fd = file->private_data;
1721 struct ll_readahead_state *ras = &fd->fd_ras;
1722 struct ll_ra_info *ra = &sbi->ll_ra_info;
1723 unsigned long throttle;
1724 pgoff_t start_idx = ras_align(ras, ras->ras_next_readahead_idx);
1725 pgoff_t end_idx = start_idx + pages - 1;
1728 * In case we have a limited max_cached_mb, readahead
1729 * should be stopped if it have run out of all LRU slots.
1731 if (atomic_read(&ra->ra_cur_pages) >= sbi->ll_cache->ccc_lru_max) {
1732 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
1736 throttle = min(ra->ra_async_pages_per_file_threshold,
1737 ra->ra_max_pages_per_file);
1739 * If this is strided i/o or the window is smaller than the
1740 * throttle limit, we do not do async readahead. Otherwise,
1741 * we do async readahead, allowing the user thread to do fast i/o.
1743 if (stride_io_mode(ras) || !throttle ||
1744 ras->ras_window_pages < throttle ||
1745 atomic_read(&ra->ra_async_inflight) > ra->ra_async_max_active)
1748 if ((atomic_read(&ra->ra_cur_pages) + pages) > ra->ra_max_pages)
1751 if (ras->ras_async_last_readpage_idx == start_idx)
1754 /* ll_readahead_work_free() free it */
1757 atomic_inc(&sbi->ll_ra_info.ra_async_inflight);
1758 lrw->lrw_file = get_file(file);
1759 lrw->lrw_start_idx = start_idx;
1760 lrw->lrw_end_idx = end_idx;
1761 spin_lock(&ras->ras_lock);
1762 ras->ras_next_readahead_idx = end_idx + 1;
1763 ras->ras_async_last_readpage_idx = start_idx;
1764 spin_unlock(&ras->ras_lock);
1765 memcpy(lrw->lrw_jobid, ll_i2info(inode)->lli_jobid,
1766 sizeof(lrw->lrw_jobid));
1767 ll_readahead_work_add(inode, lrw);
1776 * Check if we can issue a readahead RPC, if that is
1777 * the case, we can't do fast IO because we will need
1778 * a cl_io to issue the RPC.
1780 static bool ll_use_fast_io(struct file *file,
1781 struct ll_readahead_state *ras, pgoff_t index)
1783 unsigned long fast_read_pages =
1784 max(RA_REMAIN_WINDOW_MIN, ras->ras_rpc_pages);
1786 loff_t stride_bytes = ras->ras_stride_bytes;
1788 if (stride_io_mode(ras) && stride_bytes) {
1789 skip_pages = (ras->ras_stride_length +
1790 ras->ras_stride_bytes - 1) / stride_bytes;
1791 skip_pages *= fast_read_pages;
1793 skip_pages = fast_read_pages;
1796 if (ras->ras_window_start_idx + ras->ras_window_pages <
1797 ras->ras_next_readahead_idx + skip_pages ||
1798 kickoff_async_readahead(file, fast_read_pages) > 0)
1804 int ll_readpage(struct file *file, struct page *vmpage)
1806 struct inode *inode = file_inode(file);
1807 struct cl_object *clob = ll_i2info(inode)->lli_clob;
1808 struct ll_cl_context *lcc;
1809 const struct lu_env *env = NULL;
1810 struct cl_io *io = NULL;
1811 struct cl_page *page;
1812 struct ll_sb_info *sbi = ll_i2sbi(inode);
1816 lcc = ll_cl_find(file);
1822 if (io == NULL) { /* fast read */
1823 struct inode *inode = file_inode(file);
1824 struct ll_file_data *fd = file->private_data;
1825 struct ll_readahead_state *ras = &fd->fd_ras;
1826 struct lu_env *local_env = NULL;
1827 struct vvp_page *vpg;
1831 /* TODO: need to verify the layout version to make sure
1832 * the page is not invalid due to layout change. */
1833 page = cl_vmpage_page(vmpage, clob);
1835 unlock_page(vmpage);
1836 ll_ra_stats_inc_sbi(sbi, RA_STAT_FAILED_FAST_READ);
1840 vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1841 if (vpg->vpg_defer_uptodate) {
1842 enum ras_update_flags flags = LL_RAS_HIT;
1844 if (lcc && lcc->lcc_type == LCC_MMAP)
1845 flags |= LL_RAS_MMAP;
1847 /* For fast read, it updates read ahead state only
1848 * if the page is hit in cache because non cache page
1849 * case will be handled by slow read later. */
1850 ras_update(sbi, inode, ras, vvp_index(vpg), flags, io);
1851 /* avoid duplicate ras_update() call */
1852 vpg->vpg_ra_updated = 1;
1854 if (ll_use_fast_io(file, ras, vvp_index(vpg)))
1859 local_env = cl_env_percpu_get();
1863 /* export the page and skip io stack */
1865 vpg->vpg_ra_used = 1;
1866 cl_page_export(env, page, 1);
1868 ll_ra_stats_inc_sbi(sbi, RA_STAT_FAILED_FAST_READ);
1870 /* release page refcount before unlocking the page to ensure
1871 * the object won't be destroyed in the calling path of
1872 * cl_page_put(). Please see comment in ll_releasepage(). */
1873 cl_page_put(env, page);
1874 unlock_page(vmpage);
1876 cl_env_percpu_put(local_env);
1882 * Direct read can fall back to buffered read, but DIO is done
1883 * with lockless i/o, and buffered requires LDLM locking, so in
1884 * this case we must restart without lockless.
1886 if (file->f_flags & O_DIRECT &&
1887 lcc && lcc->lcc_type == LCC_RW &&
1889 unlock_page(vmpage);
1890 io->ci_dio_lock = 1;
1891 io->ci_need_restart = 1;
1895 LASSERT(io->ci_state == CIS_IO_GOING);
1896 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
1897 if (!IS_ERR(page)) {
1898 LASSERT(page->cp_type == CPT_CACHEABLE);
1899 if (likely(!PageUptodate(vmpage))) {
1900 cl_page_assume(env, io, page);
1902 result = ll_io_read_page(env, io, page, file);
1904 /* Page from a non-object file. */
1905 unlock_page(vmpage);
1908 cl_page_put(env, page);
1910 unlock_page(vmpage);
1911 result = PTR_ERR(page);