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;
92 WARN_ON_ONCE(pages_min > pages);
94 * Don't try readahead aggresively if we are limited
95 * LRU pages, otherwise, it could cause deadlock.
97 pages = min(sbi->ll_cache->ccc_lru_max >> 2, pages);
99 * if this happen, we reserve more pages than needed,
100 * this will make us leak @ra_cur_pages, because
101 * ll_ra_count_put() acutally freed @pages.
103 if (unlikely(pages_min > pages))
107 * If read-ahead pages left are less than 1M, do not do read-ahead,
108 * otherwise it will form small read RPC(< 1M), which hurt server
111 ret = min(ra->ra_max_pages - atomic_read(&ra->ra_cur_pages),
113 if (ret < 0 || ret < min_t(long, PTLRPC_MAX_BRW_PAGES, pages))
116 if (atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) {
117 atomic_sub(ret, &ra->ra_cur_pages);
122 if (ret < pages_min) {
123 /* override ra limit for maximum performance */
124 atomic_add(pages_min - ret, &ra->ra_cur_pages);
130 void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long pages)
132 struct ll_ra_info *ra = &sbi->ll_ra_info;
133 atomic_sub(pages, &ra->ra_cur_pages);
136 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which)
138 LASSERTF(which < _NR_RA_STAT, "which: %u\n", which);
139 lprocfs_counter_incr(sbi->ll_ra_stats, which);
142 static inline bool ll_readahead_enabled(struct ll_sb_info *sbi)
144 return sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
145 sbi->ll_ra_info.ra_max_pages > 0;
148 void ll_ra_stats_inc(struct inode *inode, enum ra_stat which)
150 struct ll_sb_info *sbi = ll_i2sbi(inode);
151 ll_ra_stats_inc_sbi(sbi, which);
154 #define RAS_CDEBUG(ras) \
156 "lre %llu cr %lu cb %llu wsi %lu wp %lu nra %lu rpc %lu " \
157 "r %lu csr %lu so %llu sb %llu sl %llu lr %lu\n", \
158 ras->ras_last_read_end_bytes, ras->ras_consecutive_requests, \
159 ras->ras_consecutive_bytes, ras->ras_window_start_idx, \
160 ras->ras_window_pages, ras->ras_next_readahead_idx, \
161 ras->ras_rpc_pages, ras->ras_requests, \
162 ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
163 ras->ras_stride_bytes, ras->ras_stride_length, \
164 ras->ras_async_last_readpage_idx)
166 static bool pos_in_window(loff_t pos, loff_t point,
167 unsigned long before, unsigned long after)
169 loff_t start = point - before;
170 loff_t end = point + after;
177 return start <= pos && pos <= end;
180 enum ll_ra_page_hint {
181 MAYNEED = 0, /* this page possibly accessed soon */
182 WILLNEED /* this page is gurateed to be needed */
186 * Initiates read-ahead of a page with given index.
188 * \retval +ve: page was already uptodate so it will be skipped
190 * \retval -ve: page wasn't added to \a queue for error;
191 * \retval 0: page was added into \a queue for read ahead.
193 static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io,
194 struct cl_page_list *queue, pgoff_t index,
195 enum ll_ra_page_hint hint)
197 struct cl_object *clob = io->ci_obj;
198 struct inode *inode = vvp_object_inode(clob);
199 struct page *vmpage = NULL;
200 struct cl_page *page;
201 struct vvp_page *vpg;
202 enum ra_stat which = _NR_RA_STAT; /* keep gcc happy */
204 const char *msg = NULL;
210 vmpage = grab_cache_page_nowait(inode->i_mapping, index);
211 if (vmpage == NULL) {
212 which = RA_STAT_FAILED_GRAB_PAGE;
213 msg = "g_c_p_n failed";
214 GOTO(out, rc = -EBUSY);
218 vmpage = find_or_create_page(inode->i_mapping, index,
221 GOTO(out, rc = -ENOMEM);
224 /* should not come here */
225 GOTO(out, rc = -EINVAL);
228 /* Check if vmpage was truncated or reclaimed */
229 if (vmpage->mapping != inode->i_mapping) {
230 which = RA_STAT_WRONG_GRAB_PAGE;
231 msg = "g_c_p_n returned invalid page";
232 GOTO(out, rc = -EBUSY);
235 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
237 which = RA_STAT_FAILED_GRAB_PAGE;
238 msg = "cl_page_find failed";
239 GOTO(out, rc = PTR_ERR(page));
242 lu_ref_add(&page->cp_reference, "ra", current);
243 cl_page_assume(env, io, page);
244 vpg = cl2vvp_page(cl_object_page_slice(clob, page));
245 if (!vpg->vpg_defer_uptodate && !PageUptodate(vmpage)) {
246 if (hint == MAYNEED) {
247 vpg->vpg_defer_uptodate = 1;
248 vpg->vpg_ra_used = 0;
250 cl_page_list_add(queue, page, true);
252 /* skip completed pages */
253 cl_page_unassume(env, io, page);
254 /* This page is already uptodate, returning a positive number
255 * to tell the callers about this */
259 lu_ref_del(&page->cp_reference, "ra", current);
260 cl_page_put(env, page);
263 if (vmpage != NULL) {
268 if (msg != NULL && hint == MAYNEED) {
269 ll_ra_stats_inc(inode, which);
270 CDEBUG(D_READA, "%s\n", msg);
277 #define RIA_DEBUG(ria) \
278 CDEBUG(D_READA, "rs %lu re %lu ro %llu rl %llu rb %llu\n", \
279 ria->ria_start_idx, ria->ria_end_idx, ria->ria_stoff, \
280 ria->ria_length, ria->ria_bytes)
282 static inline int stride_io_mode(struct ll_readahead_state *ras)
284 return ras->ras_consecutive_stride_requests > 1;
287 /* The function calculates how many bytes will be read in
288 * [off, off + length], in such stride IO area,
289 * stride_offset = st_off, stride_lengh = st_len,
290 * stride_bytes = st_bytes
292 * |------------------|*****|------------------|*****|------------|*****|....
295 * |----- st_len -----|
297 * How many bytes it should read in such pattern
298 * |-------------------------------------------------------------|
300 * |<------ length ------->|
302 * = |<----->| + |-------------------------------------| + |---|
303 * start_left st_bytes * i end_left
305 static loff_t stride_byte_count(loff_t st_off, loff_t st_len, loff_t st_bytes,
306 loff_t off, loff_t length)
308 u64 start = off > st_off ? off - st_off : 0;
309 u64 end = off + length > st_off ? off + length - st_off : 0;
314 if (st_len == 0 || length == 0 || end == 0)
317 start = div64_u64_rem(start, st_len, &start_left);
318 if (start_left < st_bytes)
319 start_left = st_bytes - start_left;
323 end = div64_u64_rem(end, st_len, &end_left);
324 if (end_left > st_bytes)
327 CDEBUG(D_READA, "start %llu, end %llu start_left %llu end_left %llu\n",
328 start, end, start_left, end_left);
331 bytes_count = end_left - (st_bytes - start_left);
333 bytes_count = start_left +
334 st_bytes * (end - start - 1) + end_left;
337 "st_off %llu, st_len %llu st_bytes %llu off %llu length %llu bytescount %llu\n",
338 st_off, st_len, st_bytes, off, length, bytes_count);
343 static unsigned long ria_page_count(struct ra_io_arg *ria)
345 loff_t length_bytes = ria->ria_end_idx >= ria->ria_start_idx ?
346 (loff_t)(ria->ria_end_idx -
347 ria->ria_start_idx + 1) << PAGE_SHIFT : 0;
350 if (ria->ria_length > ria->ria_bytes && ria->ria_bytes &&
351 (ria->ria_length & ~PAGE_MASK || ria->ria_bytes & ~PAGE_MASK ||
352 ria->ria_stoff & ~PAGE_MASK)) {
353 /* Over-estimate un-aligned page stride read */
354 unsigned long pg_count = ((ria->ria_bytes +
355 PAGE_SIZE - 1) >> PAGE_SHIFT) + 1;
356 pg_count *= length_bytes / ria->ria_length + 1;
360 bytes_count = stride_byte_count(ria->ria_stoff, ria->ria_length,
362 (loff_t)ria->ria_start_idx<<PAGE_SHIFT,
364 return (bytes_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
367 static pgoff_t ras_align(struct ll_readahead_state *ras, pgoff_t index)
369 unsigned opt_size = min(ras->ras_window_pages, ras->ras_rpc_pages);
373 return index - (index % opt_size);
376 /* Check whether the index is in the defined ra-window */
377 static bool ras_inside_ra_window(pgoff_t idx, struct ra_io_arg *ria)
379 loff_t pos = (loff_t)idx << PAGE_SHIFT;
381 /* If ria_length == ria_bytes, it means non-stride I/O mode,
382 * idx should always inside read-ahead window in this case
383 * For stride I/O mode, just check whether the idx is inside
386 if (ria->ria_length == 0 || ria->ria_length == ria->ria_bytes)
389 if (pos >= ria->ria_stoff) {
392 div64_u64_rem(pos - ria->ria_stoff, ria->ria_length, &offset);
394 if (offset < ria->ria_bytes ||
395 (ria->ria_length - offset) < PAGE_SIZE)
397 } else if (pos + PAGE_SIZE > ria->ria_stoff) {
405 ll_read_ahead_pages(const struct lu_env *env, struct cl_io *io,
406 struct cl_page_list *queue, struct ll_readahead_state *ras,
407 struct ra_io_arg *ria, pgoff_t *ra_end, pgoff_t skip_index)
409 struct cl_read_ahead ra = { 0 };
410 /* busy page count is per stride */
411 int rc = 0, count = 0, busy_page_count = 0;
414 LASSERT(ria != NULL);
417 for (page_idx = ria->ria_start_idx;
418 page_idx <= ria->ria_end_idx && ria->ria_reserved > 0;
420 if (skip_index && page_idx == skip_index)
422 if (ras_inside_ra_window(page_idx, ria)) {
423 if (ra.cra_end_idx == 0 || ra.cra_end_idx < page_idx) {
427 * Do not shrink ria_end_idx at any case until
428 * the minimum end of current read is covered.
430 * Do not extend read lock accross stripe if
431 * lock contention detected.
433 if (ra.cra_contention &&
434 page_idx > ria->ria_end_idx_min) {
435 ria->ria_end_idx = *ra_end;
439 cl_read_ahead_release(env, &ra);
441 rc = cl_io_read_ahead(env, io, page_idx, &ra);
446 * Only shrink ria_end_idx if the matched
447 * LDLM lock doesn't cover more.
449 if (page_idx > ra.cra_end_idx) {
450 ria->ria_end_idx = ra.cra_end_idx;
454 CDEBUG(D_READA, "idx: %lu, ra: %lu, rpc: %lu\n",
455 page_idx, ra.cra_end_idx,
457 LASSERTF(ra.cra_end_idx >= page_idx,
458 "object: %p, indcies %lu / %lu\n",
459 io->ci_obj, ra.cra_end_idx, page_idx);
460 /* update read ahead RPC size.
461 * NB: it's racy but doesn't matter */
462 if (ras->ras_rpc_pages != ra.cra_rpc_pages &&
463 ra.cra_rpc_pages > 0)
464 ras->ras_rpc_pages = ra.cra_rpc_pages;
466 /* trim it to align with optimal RPC size */
467 end_idx = ras_align(ras, ria->ria_end_idx + 1);
468 if (end_idx > 0 && !ria->ria_eof)
469 ria->ria_end_idx = end_idx - 1;
471 if (ria->ria_end_idx < ria->ria_end_idx_min)
472 ria->ria_end_idx = ria->ria_end_idx_min;
474 if (page_idx > ria->ria_end_idx)
477 /* If the page is inside the read-ahead window */
478 rc = ll_read_ahead_page(env, io, queue, page_idx,
480 if (rc < 0 && rc != -EBUSY)
485 "skip busy page: %lu\n", page_idx);
486 /* For page unaligned readahead the first
487 * last pages of each region can be read by
488 * another reader on the same node, and so
489 * may be busy. So only stop for > 2 busy
491 if (busy_page_count > 2)
496 /* Only subtract from reserve & count the page if we
497 * really did readahead on that page. */
502 } else if (stride_io_mode(ras)) {
503 /* If it is not in the read-ahead window, and it is
504 * read-ahead mode, then check whether it should skip
507 loff_t pos = (loff_t)page_idx << PAGE_SHIFT;
510 div64_u64_rem(pos - ria->ria_stoff, ria->ria_length,
512 if (offset >= ria->ria_bytes) {
513 pos += (ria->ria_length - offset);
514 if ((pos >> PAGE_SHIFT) >= page_idx + 1)
515 page_idx = (pos >> PAGE_SHIFT) - 1;
518 "Stride: jump %llu pages to %lu\n",
519 ria->ria_length - offset, page_idx);
525 cl_read_ahead_release(env, &ra);
530 static void ll_readahead_work_free(struct ll_readahead_work *work)
532 fput(work->lrw_file);
536 static void ll_readahead_handle_work(struct work_struct *wq);
537 static void ll_readahead_work_add(struct inode *inode,
538 struct ll_readahead_work *work)
540 INIT_WORK(&work->lrw_readahead_work, ll_readahead_handle_work);
541 queue_work(ll_i2sbi(inode)->ll_ra_info.ll_readahead_wq,
542 &work->lrw_readahead_work);
545 static int ll_readahead_file_kms(const struct lu_env *env,
546 struct cl_io *io, __u64 *kms)
548 struct cl_object *clob;
550 struct cl_attr *attr = vvp_env_thread_attr(env);
554 inode = vvp_object_inode(clob);
556 cl_object_attr_lock(clob);
557 ret = cl_object_attr_get(env, clob, attr);
558 cl_object_attr_unlock(clob);
563 *kms = attr->cat_kms;
567 static void ll_readahead_handle_work(struct work_struct *wq)
569 struct ll_readahead_work *work;
572 struct ra_io_arg *ria;
574 struct ll_file_data *fd;
575 struct ll_readahead_state *ras;
577 struct cl_2queue *queue;
578 pgoff_t ra_end_idx = 0;
579 unsigned long pages, pages_min = 0;
584 struct ll_sb_info *sbi;
586 work = container_of(wq, struct ll_readahead_work,
588 fd = work->lrw_file->private_data;
590 file = work->lrw_file;
591 inode = file_inode(file);
592 sbi = ll_i2sbi(inode);
594 env = cl_env_alloc(&refcheck, LCT_NOREF);
596 GOTO(out_free_work, rc = PTR_ERR(env));
598 io = vvp_env_thread_io(env);
599 ll_io_init(io, file, CIT_READ, NULL);
601 rc = ll_readahead_file_kms(env, io, &kms);
603 GOTO(out_put_env, rc);
606 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
607 GOTO(out_put_env, rc = 0);
610 ria = &ll_env_info(env)->lti_ria;
611 memset(ria, 0, sizeof(*ria));
613 ria->ria_start_idx = work->lrw_start_idx;
614 /* Truncate RA window to end of file */
615 eof_index = (pgoff_t)(kms - 1) >> PAGE_SHIFT;
616 if (eof_index <= work->lrw_end_idx) {
617 work->lrw_end_idx = eof_index;
620 if (work->lrw_end_idx <= work->lrw_start_idx)
621 GOTO(out_put_env, rc = 0);
623 ria->ria_end_idx = work->lrw_end_idx;
624 pages = ria->ria_end_idx - ria->ria_start_idx + 1;
625 ria->ria_reserved = ll_ra_count_get(sbi, ria,
626 ria_page_count(ria), pages_min);
629 "async reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
630 ria->ria_reserved, pages, pages_min,
631 atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
632 ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
634 if (ria->ria_reserved < pages) {
635 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
636 if (PAGES_TO_MiB(ria->ria_reserved) < 1) {
637 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
638 GOTO(out_put_env, rc = 0);
642 rc = cl_io_rw_init(env, io, CIT_READ, ria->ria_start_idx, pages);
644 GOTO(out_put_env, rc);
646 /* overwrite jobid inited in vvp_io_init() */
647 if (strncmp(ll_i2info(inode)->lli_jobid, work->lrw_jobid,
648 sizeof(work->lrw_jobid)))
649 memcpy(ll_i2info(inode)->lli_jobid, work->lrw_jobid,
650 sizeof(work->lrw_jobid));
652 vvp_env_io(env)->vui_fd = fd;
653 io->ci_state = CIS_LOCKED;
654 io->ci_async_readahead = true;
655 rc = cl_io_start(env, io);
657 GOTO(out_io_fini, rc);
659 queue = &io->ci_queue;
660 cl_2queue_init(queue);
662 rc = ll_read_ahead_pages(env, io, &queue->c2_qin, ras, ria,
664 if (ria->ria_reserved != 0)
665 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
666 if (queue->c2_qin.pl_nr > 0) {
667 int count = queue->c2_qin.pl_nr;
669 rc = cl_io_submit_rw(env, io, CRT_READ, queue);
671 task_io_account_read(PAGE_SIZE * count);
673 if (ria->ria_end_idx == ra_end_idx && ra_end_idx == (kms >> PAGE_SHIFT))
674 ll_ra_stats_inc(inode, RA_STAT_EOF);
676 if (ra_end_idx != ria->ria_end_idx)
677 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
679 /* TODO: discard all pages until page reinit route is implemented */
680 cl_page_list_discard(env, io, &queue->c2_qin);
682 /* Unlock unsent read pages in case of error. */
683 cl_page_list_disown(env, io, &queue->c2_qin);
685 cl_2queue_fini(env, queue);
690 cl_env_put(env, &refcheck);
693 ll_ra_stats_inc_sbi(ll_i2sbi(inode), RA_STAT_ASYNC);
694 atomic_dec(&sbi->ll_ra_info.ra_async_inflight);
695 ll_readahead_work_free(work);
698 static int ll_readahead(const struct lu_env *env, struct cl_io *io,
699 struct cl_page_list *queue,
700 struct ll_readahead_state *ras, bool hit,
701 struct file *file, pgoff_t skip_index)
703 struct vvp_io *vio = vvp_env_io(env);
704 struct ll_thread_info *lti = ll_env_info(env);
705 unsigned long pages, pages_min = 0;
706 pgoff_t ra_end_idx = 0, start_idx = 0, end_idx = 0;
708 struct ra_io_arg *ria = <i->lti_ria;
709 struct cl_object *clob;
712 struct ll_sb_info *sbi;
713 struct ll_ra_info *ra;
720 inode = vvp_object_inode(clob);
721 sbi = ll_i2sbi(inode);
722 ra = &sbi->ll_ra_info;
725 * In case we have a limited max_cached_mb, readahead
726 * should be stopped if it have run out of all LRU slots.
728 if (atomic_read(&ra->ra_cur_pages) >= sbi->ll_cache->ccc_lru_max) {
729 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
733 memset(ria, 0, sizeof(*ria));
734 ret = ll_readahead_file_kms(env, io, &kms);
739 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
743 spin_lock(&ras->ras_lock);
746 * Note: other thread might rollback the ras_next_readahead_idx,
747 * if it can not get the full size of prepared pages, see the
748 * end of this function. For stride read ahead, it needs to
749 * make sure the offset is no less than ras_stride_offset,
750 * so that stride read ahead can work correctly.
752 if (stride_io_mode(ras))
753 start_idx = max_t(pgoff_t, ras->ras_next_readahead_idx,
754 ras->ras_stride_offset >> PAGE_SHIFT);
756 start_idx = ras->ras_next_readahead_idx;
758 if (ras->ras_window_pages > 0)
759 end_idx = ras->ras_window_start_idx + ras->ras_window_pages - 1;
762 end_idx = start_idx + ras->ras_window_pages - 1;
764 /* Enlarge the RA window to encompass the full read */
765 if (vio->vui_ra_valid &&
766 end_idx < vio->vui_ra_start_idx + vio->vui_ra_pages - 1)
767 end_idx = vio->vui_ra_start_idx + vio->vui_ra_pages - 1;
772 /* Truncate RA window to end of file */
773 eof_index = (pgoff_t)((kms - 1) >> PAGE_SHIFT);
774 if (eof_index <= end_idx) {
779 ria->ria_start_idx = start_idx;
780 ria->ria_end_idx = end_idx;
781 /* If stride I/O mode is detected, get stride window*/
782 if (stride_io_mode(ras)) {
783 ria->ria_stoff = ras->ras_stride_offset;
784 ria->ria_length = ras->ras_stride_length;
785 ria->ria_bytes = ras->ras_stride_bytes;
787 spin_unlock(&ras->ras_lock);
790 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
793 pages = ria_page_count(ria);
795 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
800 CDEBUG(D_READA, DFID": ria: %lu/%lu, bead: %lu/%lu, hit: %d\n",
801 PFID(lu_object_fid(&clob->co_lu)),
802 ria->ria_start_idx, ria->ria_end_idx,
803 vio->vui_ra_valid ? vio->vui_ra_start_idx : 0,
804 vio->vui_ra_valid ? vio->vui_ra_pages : 0,
807 /* at least to extend the readahead window to cover current read */
808 if (!hit && vio->vui_ra_valid &&
809 vio->vui_ra_start_idx + vio->vui_ra_pages > ria->ria_start_idx) {
810 ria->ria_end_idx_min =
811 vio->vui_ra_start_idx + vio->vui_ra_pages - 1;
812 pages_min = vio->vui_ra_start_idx + vio->vui_ra_pages -
815 * For performance reason, exceeding @ra_max_pages
816 * are allowed, but this should be limited with RPC
817 * size in case a large block size read issued. Trim
820 pages_min = min(pages_min, ras->ras_rpc_pages -
821 (ria->ria_start_idx % ras->ras_rpc_pages));
824 /* don't over reserved for mmap range read */
827 if (pages_min > pages)
829 ria->ria_reserved = ll_ra_count_get(ll_i2sbi(inode), ria, pages,
831 if (ria->ria_reserved < pages)
832 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
834 CDEBUG(D_READA, "reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
835 ria->ria_reserved, pages, pages_min,
836 atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
837 ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
839 ret = ll_read_ahead_pages(env, io, queue, ras, ria, &ra_end_idx,
841 if (ria->ria_reserved != 0)
842 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
844 if (ra_end_idx == end_idx && ra_end_idx == (kms >> PAGE_SHIFT))
845 ll_ra_stats_inc(inode, RA_STAT_EOF);
848 "ra_end_idx = %lu end_idx = %lu stride end = %lu pages = %d\n",
849 ra_end_idx, end_idx, ria->ria_end_idx, ret);
851 if (ra_end_idx != end_idx)
852 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
853 if (ra_end_idx > 0) {
854 /* update the ras so that the next read-ahead tries from
855 * where we left off. */
856 spin_lock(&ras->ras_lock);
857 ras->ras_next_readahead_idx = ra_end_idx + 1;
858 spin_unlock(&ras->ras_lock);
865 static int ll_readpages(const struct lu_env *env, struct cl_io *io,
866 struct cl_page_list *queue,
867 pgoff_t start, pgoff_t end)
876 ret = ll_readahead_file_kms(env, io, &kms);
884 unsigned long end_index;
886 end_index = (unsigned long)((kms - 1) >> PAGE_SHIFT);
887 if (end_index <= end)
891 for (page_idx = start; page_idx <= end; page_idx++) {
892 ret= ll_read_ahead_page(env, io, queue, page_idx,
896 else if (ret == 0) /* ret 1 is already uptodate */
900 RETURN(count > 0 ? count : ret);
903 static void ras_set_start(struct ll_readahead_state *ras, pgoff_t index)
905 ras->ras_window_start_idx = ras_align(ras, index);
908 /* called with the ras_lock held or from places where it doesn't matter */
909 static void ras_reset(struct ll_readahead_state *ras, pgoff_t index)
911 ras->ras_consecutive_requests = 0;
912 ras->ras_consecutive_bytes = 0;
913 ras->ras_window_pages = 0;
914 ras_set_start(ras, index);
915 ras->ras_next_readahead_idx = max(ras->ras_window_start_idx, index + 1);
920 /* called with the ras_lock held or from places where it doesn't matter */
921 static void ras_stride_reset(struct ll_readahead_state *ras)
923 ras->ras_consecutive_stride_requests = 0;
924 ras->ras_stride_length = 0;
925 ras->ras_stride_bytes = 0;
929 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
931 spin_lock_init(&ras->ras_lock);
932 ras->ras_rpc_pages = PTLRPC_MAX_BRW_PAGES;
934 ras->ras_last_read_end_bytes = 0;
935 ras->ras_requests = 0;
936 ras->ras_range_min_start_idx = 0;
937 ras->ras_range_max_end_idx = 0;
938 ras->ras_range_requests = 0;
939 ras->ras_last_range_pages = 0;
943 * Check whether the read request is in the stride window.
944 * If it is in the stride window, return true, otherwise return false.
946 static bool read_in_stride_window(struct ll_readahead_state *ras,
947 loff_t pos, loff_t count)
951 if (ras->ras_stride_length == 0 || ras->ras_stride_bytes == 0 ||
952 ras->ras_stride_bytes == ras->ras_stride_length)
955 stride_gap = pos - ras->ras_last_read_end_bytes - 1;
957 /* If it is contiguous read */
959 return ras->ras_consecutive_bytes + count <=
960 ras->ras_stride_bytes;
962 /* Otherwise check the stride by itself */
963 return (ras->ras_stride_length - ras->ras_stride_bytes) == stride_gap &&
964 ras->ras_consecutive_bytes == ras->ras_stride_bytes &&
965 count <= ras->ras_stride_bytes;
968 static void ras_init_stride_detector(struct ll_readahead_state *ras,
969 loff_t pos, loff_t count)
971 loff_t stride_gap = pos - ras->ras_last_read_end_bytes - 1;
973 LASSERT(ras->ras_consecutive_stride_requests == 0);
975 if (pos <= ras->ras_last_read_end_bytes) {
976 /*Reset stride window for forward read*/
977 ras_stride_reset(ras);
981 ras->ras_stride_bytes = ras->ras_consecutive_bytes;
982 ras->ras_stride_length = stride_gap + ras->ras_consecutive_bytes;
983 ras->ras_consecutive_stride_requests++;
984 ras->ras_stride_offset = pos;
990 stride_page_count(struct ll_readahead_state *ras, loff_t len)
993 stride_byte_count(ras->ras_stride_offset,
994 ras->ras_stride_length, ras->ras_stride_bytes,
995 ras->ras_window_start_idx << PAGE_SHIFT, len);
997 return (bytes_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
1000 /* Stride Read-ahead window will be increased inc_len according to
1001 * stride I/O pattern */
1002 static void ras_stride_increase_window(struct ll_readahead_state *ras,
1003 struct ll_ra_info *ra, loff_t inc_bytes)
1005 loff_t window_bytes, stride_bytes;
1010 /* temporarily store in page units to reduce LASSERT() cost below */
1011 end = ras->ras_window_start_idx + ras->ras_window_pages;
1013 LASSERT(ras->ras_stride_length > 0);
1014 LASSERTF(end >= (ras->ras_stride_offset >> PAGE_SHIFT),
1015 "window_start_idx %lu, window_pages %lu stride_offset %llu\n",
1016 ras->ras_window_start_idx, ras->ras_window_pages,
1017 ras->ras_stride_offset);
1020 if (end <= ras->ras_stride_offset)
1023 stride_bytes = end - ras->ras_stride_offset;
1025 div64_u64_rem(stride_bytes, ras->ras_stride_length, &left_bytes);
1026 window_bytes = (ras->ras_window_pages << PAGE_SHIFT);
1027 if (left_bytes < ras->ras_stride_bytes) {
1028 if (ras->ras_stride_bytes - left_bytes >= inc_bytes) {
1029 window_bytes += inc_bytes;
1032 window_bytes += (ras->ras_stride_bytes - left_bytes);
1033 inc_bytes -= (ras->ras_stride_bytes - left_bytes);
1036 window_bytes += (ras->ras_stride_length - left_bytes);
1039 LASSERT(ras->ras_stride_bytes != 0);
1041 step = div64_u64_rem(inc_bytes, ras->ras_stride_bytes, &left_bytes);
1043 window_bytes += step * ras->ras_stride_length + left_bytes;
1044 LASSERT(window_bytes > 0);
1047 if (stride_page_count(ras, window_bytes) <=
1048 ra->ra_max_pages_per_file || ras->ras_window_pages == 0)
1049 ras->ras_window_pages = (window_bytes >> PAGE_SHIFT);
1051 LASSERT(ras->ras_window_pages > 0);
1056 static void ras_increase_window(struct inode *inode,
1057 struct ll_readahead_state *ras,
1058 struct ll_ra_info *ra)
1060 /* The stretch of ra-window should be aligned with max rpc_size
1061 * but current clio architecture does not support retrieve such
1062 * information from lower layer. FIXME later
1064 if (stride_io_mode(ras)) {
1065 ras_stride_increase_window(ras, ra,
1066 (loff_t)ras->ras_rpc_pages << PAGE_SHIFT);
1068 pgoff_t window_pages;
1070 window_pages = min(ras->ras_window_pages + ras->ras_rpc_pages,
1071 ra->ra_max_pages_per_file);
1072 if (window_pages < ras->ras_rpc_pages)
1073 ras->ras_window_pages = window_pages;
1075 ras->ras_window_pages = ras_align(ras, window_pages);
1080 * Seek within 8 pages are considered as sequential read for now.
1082 static inline bool is_loose_seq_read(struct ll_readahead_state *ras, loff_t pos)
1084 return pos_in_window(pos, ras->ras_last_read_end_bytes,
1085 8UL << PAGE_SHIFT, 8UL << PAGE_SHIFT);
1088 static inline bool is_loose_mmap_read(struct ll_sb_info *sbi,
1089 struct ll_readahead_state *ras,
1092 unsigned long range_pages = sbi->ll_ra_info.ra_range_pages;
1094 return pos_in_window(pos, ras->ras_last_read_end_bytes,
1095 range_pages << PAGE_SHIFT,
1096 range_pages << PAGE_SHIFT);
1100 * We have observed slow mmap read performances for some
1101 * applications. The problem is if access pattern is neither
1102 * sequential nor stride, but could be still adjacent in a
1103 * small range and then seek a random position.
1105 * So the pattern could be something like this:
1107 * [1M data] [hole] [0.5M data] [hole] [0.7M data] [1M data]
1110 * Every time an application reads mmap data, it may not only
1111 * read a single 4KB page, but aslo a cluster of nearby pages in
1112 * a range(e.g. 1MB) of the first page after a cache miss.
1114 * The readahead engine is modified to track the range size of
1115 * a cluster of mmap reads, so that after a seek and/or cache miss,
1116 * the range size is used to efficiently prefetch multiple pages
1117 * in a single RPC rather than many small RPCs.
1119 static void ras_detect_cluster_range(struct ll_readahead_state *ras,
1120 struct ll_sb_info *sbi,
1121 unsigned long pos, unsigned long count)
1123 pgoff_t last_pages, pages;
1124 pgoff_t end_idx = (pos + count - 1) >> PAGE_SHIFT;
1126 last_pages = ras->ras_range_max_end_idx -
1127 ras->ras_range_min_start_idx + 1;
1128 /* First time come here */
1129 if (!ras->ras_range_max_end_idx)
1132 /* Random or Stride read */
1133 if (!is_loose_mmap_read(sbi, ras, pos))
1136 ras->ras_range_requests++;
1137 if (ras->ras_range_max_end_idx < end_idx)
1138 ras->ras_range_max_end_idx = end_idx;
1140 if (ras->ras_range_min_start_idx > (pos >> PAGE_SHIFT))
1141 ras->ras_range_min_start_idx = pos >> PAGE_SHIFT;
1143 /* Out of range, consider it as random or stride */
1144 pages = ras->ras_range_max_end_idx -
1145 ras->ras_range_min_start_idx + 1;
1146 if (pages <= sbi->ll_ra_info.ra_range_pages)
1149 ras->ras_last_range_pages = last_pages;
1150 ras->ras_range_requests = 0;
1151 ras->ras_range_min_start_idx = pos >> PAGE_SHIFT;
1152 ras->ras_range_max_end_idx = end_idx;
1155 static void ras_detect_read_pattern(struct ll_readahead_state *ras,
1156 struct ll_sb_info *sbi,
1157 loff_t pos, size_t count, bool mmap)
1159 bool stride_detect = false;
1160 pgoff_t index = pos >> PAGE_SHIFT;
1163 * Reset the read-ahead window in two cases. First when the app seeks
1164 * or reads to some other part of the file. Secondly if we get a
1165 * read-ahead miss that we think we've previously issued. This can
1166 * be a symptom of there being so many read-ahead pages that the VM
1167 * is reclaiming it before we get to it.
1169 if (!is_loose_seq_read(ras, pos)) {
1170 /* Check whether it is in stride I/O mode */
1171 if (!read_in_stride_window(ras, pos, count)) {
1172 if (ras->ras_consecutive_stride_requests == 0)
1173 ras_init_stride_detector(ras, pos, count);
1175 ras_stride_reset(ras);
1176 ras->ras_consecutive_bytes = 0;
1177 ras_reset(ras, index);
1179 ras->ras_consecutive_bytes = 0;
1180 ras->ras_consecutive_requests = 0;
1181 if (++ras->ras_consecutive_stride_requests > 1)
1182 stride_detect = true;
1185 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
1186 } else if (stride_io_mode(ras)) {
1188 * If this is contiguous read but in stride I/O mode
1189 * currently, check whether stride step still is valid,
1190 * if invalid, it will reset the stride ra window to
1193 if (!read_in_stride_window(ras, pos, count)) {
1194 ras_stride_reset(ras);
1195 ras->ras_window_pages = 0;
1196 ras->ras_next_readahead_idx = index;
1200 ras->ras_consecutive_bytes += count;
1202 pgoff_t idx = ras->ras_consecutive_bytes >> PAGE_SHIFT;
1203 unsigned long ra_range_pages =
1204 max_t(unsigned long, RA_MIN_MMAP_RANGE_PAGES,
1205 sbi->ll_ra_info.ra_range_pages);
1207 if ((idx >= ra_range_pages &&
1208 idx % ra_range_pages == 0) || stride_detect)
1209 ras->ras_need_increase_window = true;
1210 } else if ((ras->ras_consecutive_requests > 1 || stride_detect)) {
1211 ras->ras_need_increase_window = true;
1214 ras->ras_last_read_end_bytes = pos + count - 1;
1217 void ll_ras_enter(struct file *f, loff_t pos, size_t count)
1219 struct ll_file_data *fd = f->private_data;
1220 struct ll_readahead_state *ras = &fd->fd_ras;
1221 struct inode *inode = file_inode(f);
1222 unsigned long index = pos >> PAGE_SHIFT;
1223 struct ll_sb_info *sbi = ll_i2sbi(inode);
1225 spin_lock(&ras->ras_lock);
1226 ras->ras_requests++;
1227 ras->ras_consecutive_requests++;
1228 ras->ras_need_increase_window = false;
1229 ras->ras_no_miss_check = false;
1231 * On the second access to a file smaller than the tunable
1232 * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1233 * file up to ra_max_pages_per_file. This is simply a best effort
1234 * and only occurs once per open file. Normal RA behavior is reverted
1235 * to for subsequent IO.
1237 if (ras->ras_requests >= 2) {
1239 struct ll_ra_info *ra = &sbi->ll_ra_info;
1241 kms_pages = (i_size_read(inode) + PAGE_SIZE - 1) >>
1244 CDEBUG(D_READA, "kmsp %llu mwp %lu mp %lu\n", kms_pages,
1245 ra->ra_max_read_ahead_whole_pages,
1246 ra->ra_max_pages_per_file);
1249 kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1250 ras->ras_window_start_idx = 0;
1251 ras->ras_next_readahead_idx = index + 1;
1252 ras->ras_window_pages = min(ra->ra_max_pages_per_file,
1253 ra->ra_max_read_ahead_whole_pages);
1254 ras->ras_no_miss_check = true;
1255 GOTO(out_unlock, 0);
1258 ras_detect_read_pattern(ras, sbi, pos, count, false);
1260 spin_unlock(&ras->ras_lock);
1263 static bool index_in_stride_window(struct ll_readahead_state *ras,
1266 loff_t pos = (loff_t)index << PAGE_SHIFT;
1268 if (ras->ras_stride_length == 0 || ras->ras_stride_bytes == 0 ||
1269 ras->ras_stride_bytes == ras->ras_stride_length)
1272 if (pos >= ras->ras_stride_offset) {
1275 div64_u64_rem(pos - ras->ras_stride_offset,
1276 ras->ras_stride_length, &offset);
1277 if (offset < ras->ras_stride_bytes ||
1278 ras->ras_stride_length - offset < PAGE_SIZE)
1280 } else if (ras->ras_stride_offset - pos < PAGE_SIZE) {
1288 * ll_ras_enter() is used to detect read pattern according to pos and count.
1290 * ras_update() is used to detect cache miss and
1291 * reset window or increase window accordingly
1293 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
1294 struct ll_readahead_state *ras, pgoff_t index,
1295 enum ras_update_flags flags, struct cl_io *io)
1297 struct ll_ra_info *ra = &sbi->ll_ra_info;
1298 bool hit = flags & LL_RAS_HIT;
1301 spin_lock(&ras->ras_lock);
1304 CDEBUG(D_READA, DFID " pages at %lu miss.\n",
1305 PFID(ll_inode2fid(inode)), index);
1306 ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
1309 * The readahead window has been expanded to cover whole
1310 * file size, we don't care whether ra miss happen or not.
1311 * Because we will read whole file to page cache even if
1312 * some pages missed.
1314 if (ras->ras_no_miss_check)
1315 GOTO(out_unlock, 0);
1317 if (io && io->ci_rand_read)
1318 GOTO(out_unlock, 0);
1320 if (io && io->ci_seq_read) {
1322 /* to avoid many small read RPC here */
1323 ras->ras_window_pages = sbi->ll_ra_info.ra_range_pages;
1324 ll_ra_stats_inc_sbi(sbi, RA_STAT_MMAP_RANGE_READ);
1329 if (flags & LL_RAS_MMAP) {
1330 unsigned long ra_pages;
1332 ras_detect_cluster_range(ras, sbi, index << PAGE_SHIFT,
1334 ras_detect_read_pattern(ras, sbi, (loff_t)index << PAGE_SHIFT,
1337 /* we did not detect anything but we could prefetch */
1338 if (!ras->ras_need_increase_window &&
1339 ras->ras_window_pages <= sbi->ll_ra_info.ra_range_pages &&
1340 ras->ras_range_requests >= 2) {
1342 ra_pages = max_t(unsigned long,
1343 RA_MIN_MMAP_RANGE_PAGES,
1344 ras->ras_last_range_pages);
1345 if (index < ra_pages / 2)
1348 index -= ra_pages / 2;
1349 ras->ras_window_pages = ra_pages;
1350 ll_ra_stats_inc_sbi(sbi,
1351 RA_STAT_MMAP_RANGE_READ);
1353 ras->ras_window_pages = 0;
1359 if (!hit && ras->ras_window_pages &&
1360 index < ras->ras_next_readahead_idx &&
1361 pos_in_window(index, ras->ras_window_start_idx, 0,
1362 ras->ras_window_pages)) {
1363 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
1364 ras->ras_need_increase_window = false;
1366 if (index_in_stride_window(ras, index) &&
1367 stride_io_mode(ras)) {
1369 * if (index != ras->ras_last_readpage + 1)
1370 * ras->ras_consecutive_pages = 0;
1372 ras_reset(ras, index);
1375 * If stride-RA hit cache miss, the stride
1376 * detector will not be reset to avoid the
1377 * overhead of redetecting read-ahead mode,
1378 * but on the condition that the stride window
1379 * is still intersect with normal sequential
1380 * read-ahead window.
1382 if (ras->ras_window_start_idx < ras->ras_stride_offset)
1383 ras_stride_reset(ras);
1387 * Reset both stride window and normal RA
1390 ras_reset(ras, index);
1391 /* ras->ras_consecutive_pages++; */
1392 ras->ras_consecutive_bytes = 0;
1393 ras_stride_reset(ras);
1394 GOTO(out_unlock, 0);
1399 ras_set_start(ras, index);
1401 if (stride_io_mode(ras)) {
1402 /* Since stride readahead is sentivite to the offset
1403 * of read-ahead, so we use original offset here,
1404 * instead of ras_window_start_idx, which is RPC aligned.
1406 ras->ras_next_readahead_idx = max(index + 1,
1407 ras->ras_next_readahead_idx);
1408 ras->ras_window_start_idx =
1409 max_t(pgoff_t, ras->ras_window_start_idx,
1410 ras->ras_stride_offset >> PAGE_SHIFT);
1412 if (ras->ras_next_readahead_idx < ras->ras_window_start_idx)
1413 ras->ras_next_readahead_idx = ras->ras_window_start_idx;
1415 ras->ras_next_readahead_idx = index + 1;
1418 if (ras->ras_need_increase_window) {
1419 ras_increase_window(inode, ras, ra);
1420 ras->ras_need_increase_window = false;
1425 spin_unlock(&ras->ras_lock);
1428 int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
1430 struct inode *inode = vmpage->mapping->host;
1431 struct ll_inode_info *lli = ll_i2info(inode);
1434 struct cl_page *page;
1435 struct cl_object *clob;
1436 bool redirtied = false;
1437 bool unlocked = false;
1442 LASSERT(PageLocked(vmpage));
1443 LASSERT(!PageWriteback(vmpage));
1445 LASSERT(ll_i2dtexp(inode) != NULL);
1447 env = cl_env_get(&refcheck);
1449 GOTO(out, result = PTR_ERR(env));
1451 clob = ll_i2info(inode)->lli_clob;
1452 LASSERT(clob != NULL);
1454 io = vvp_env_thread_io(env);
1456 io->ci_ignore_layout = 1;
1457 result = cl_io_init(env, io, CIT_MISC, clob);
1459 page = cl_page_find(env, clob, vmpage->index,
1460 vmpage, CPT_CACHEABLE);
1461 if (!IS_ERR(page)) {
1462 lu_ref_add(&page->cp_reference, "writepage",
1464 cl_page_assume(env, io, page);
1465 result = cl_page_flush(env, io, page);
1468 * Re-dirty page on error so it retries write,
1469 * but not in case when IO has actually
1470 * occurred and completed with an error.
1472 if (!PageError(vmpage)) {
1473 redirty_page_for_writepage(wbc, vmpage);
1478 cl_page_disown(env, io, page);
1480 lu_ref_del(&page->cp_reference,
1481 "writepage", current);
1482 cl_page_put(env, page);
1484 result = PTR_ERR(page);
1487 cl_io_fini(env, io);
1489 if (redirtied && wbc->sync_mode == WB_SYNC_ALL) {
1490 loff_t offset = cl_offset(clob, vmpage->index);
1492 /* Flush page failed because the extent is being written out.
1493 * Wait for the write of extent to be finished to avoid
1494 * breaking kernel which assumes ->writepage should mark
1495 * PageWriteback or clean the page. */
1496 result = cl_sync_file_range(inode, offset,
1497 offset + PAGE_SIZE - 1,
1500 /* actually we may have written more than one page.
1501 * decreasing this page because the caller will count
1503 wbc->nr_to_write -= result - 1;
1508 cl_env_put(env, &refcheck);
1513 if (!lli->lli_async_rc)
1514 lli->lli_async_rc = result;
1515 SetPageError(vmpage);
1517 unlock_page(vmpage);
1522 int ll_writepages(struct address_space *mapping, struct writeback_control *wbc)
1524 struct inode *inode = mapping->host;
1527 enum cl_fsync_mode mode;
1528 int range_whole = 0;
1532 if (wbc->range_cyclic) {
1533 start = (loff_t)mapping->writeback_index << PAGE_SHIFT;
1534 end = OBD_OBJECT_EOF;
1536 start = wbc->range_start;
1537 end = wbc->range_end;
1538 if (end == LLONG_MAX) {
1539 end = OBD_OBJECT_EOF;
1540 range_whole = start == 0;
1544 mode = CL_FSYNC_NONE;
1545 if (wbc->sync_mode == WB_SYNC_ALL)
1546 mode = CL_FSYNC_LOCAL;
1548 if (ll_i2info(inode)->lli_clob == NULL)
1551 /* for directio, it would call writepages() to evict cached pages
1552 * inside the IO context of write, which will cause deadlock at
1553 * layout_conf since it waits for active IOs to complete. */
1554 result = cl_sync_file_range(inode, start, end, mode, 1);
1556 wbc->nr_to_write -= result;
1560 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) {
1561 if (end == OBD_OBJECT_EOF)
1562 mapping->writeback_index = 0;
1564 mapping->writeback_index = (end >> PAGE_SHIFT) + 1;
1569 struct ll_cl_context *ll_cl_find(struct inode *inode)
1571 struct ll_inode_info *lli = ll_i2info(inode);
1572 struct ll_cl_context *lcc;
1573 struct ll_cl_context *found = NULL;
1575 read_lock(&lli->lli_lock);
1576 list_for_each_entry(lcc, &lli->lli_lccs, lcc_list) {
1577 if (lcc->lcc_cookie == current) {
1582 read_unlock(&lli->lli_lock);
1587 void ll_cl_add(struct inode *inode, const struct lu_env *env, struct cl_io *io,
1590 struct ll_inode_info *lli = ll_i2info(inode);
1591 struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1593 memset(lcc, 0, sizeof(*lcc));
1594 INIT_LIST_HEAD(&lcc->lcc_list);
1595 lcc->lcc_cookie = current;
1598 lcc->lcc_type = type;
1600 write_lock(&lli->lli_lock);
1601 list_add(&lcc->lcc_list, &lli->lli_lccs);
1602 write_unlock(&lli->lli_lock);
1605 void ll_cl_remove(struct inode *inode, const struct lu_env *env)
1607 struct ll_inode_info *lli = ll_i2info(inode);
1608 struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1610 write_lock(&lli->lli_lock);
1611 list_del_init(&lcc->lcc_list);
1612 write_unlock(&lli->lli_lock);
1615 int ll_io_read_page(const struct lu_env *env, struct cl_io *io,
1616 struct cl_page *page, struct file *file)
1618 struct inode *inode = vvp_object_inode(page->cp_obj);
1619 struct ll_sb_info *sbi = ll_i2sbi(inode);
1620 struct ll_file_data *fd = NULL;
1621 struct ll_readahead_state *ras = NULL;
1622 struct cl_2queue *queue = &io->ci_queue;
1623 struct cl_sync_io *anchor = NULL;
1624 struct vvp_page *vpg;
1625 int rc = 0, rc2 = 0;
1627 pgoff_t io_start_index;
1628 pgoff_t io_end_index;
1632 fd = file->private_data;
1636 vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1637 uptodate = vpg->vpg_defer_uptodate;
1639 if (ll_readahead_enabled(sbi) && !vpg->vpg_ra_updated && ras) {
1640 struct vvp_io *vio = vvp_env_io(env);
1641 enum ras_update_flags flags = 0;
1644 flags |= LL_RAS_HIT;
1645 if (!vio->vui_ra_valid)
1646 flags |= LL_RAS_MMAP;
1647 ras_update(sbi, inode, ras, vvp_index(vpg), flags, io);
1650 cl_2queue_init(queue);
1652 vpg->vpg_ra_used = 1;
1653 cl_page_export(env, page, 1);
1654 cl_page_disown(env, io, page);
1656 anchor = &vvp_env_info(env)->vti_anchor;
1657 cl_sync_io_init(anchor, 1);
1658 page->cp_sync_io = anchor;
1660 cl_2queue_add(queue, page, true);
1663 io_start_index = cl_index(io->ci_obj, io->u.ci_rw.crw_pos);
1664 io_end_index = cl_index(io->ci_obj, io->u.ci_rw.crw_pos +
1665 io->u.ci_rw.crw_count - 1);
1666 if (ll_readahead_enabled(sbi) && ras && !io->ci_rand_read) {
1667 pgoff_t skip_index = 0;
1669 if (ras->ras_next_readahead_idx < vvp_index(vpg))
1670 skip_index = vvp_index(vpg);
1671 rc2 = ll_readahead(env, io, &queue->c2_qin, ras,
1672 uptodate, file, skip_index);
1673 CDEBUG(D_READA, DFID " %d pages read ahead at %lu\n",
1674 PFID(ll_inode2fid(inode)), rc2, vvp_index(vpg));
1675 } else if (vvp_index(vpg) == io_start_index &&
1676 io_end_index - io_start_index > 0) {
1677 rc2 = ll_readpages(env, io, &queue->c2_qin, io_start_index + 1,
1679 CDEBUG(D_READA, DFID " %d pages read at %lu\n",
1680 PFID(ll_inode2fid(inode)), rc2, vvp_index(vpg));
1683 if (queue->c2_qin.pl_nr > 0) {
1684 int count = queue->c2_qin.pl_nr;
1685 rc = cl_io_submit_rw(env, io, CRT_READ, queue);
1687 task_io_account_read(PAGE_SIZE * count);
1691 if (anchor != NULL && !cl_page_is_owned(page, io)) { /* have sent */
1692 rc = cl_sync_io_wait(env, anchor, 0);
1694 cl_page_assume(env, io, page);
1695 cl_page_list_del(env, &queue->c2_qout, page);
1697 if (!PageUptodate(cl_page_vmpage(page))) {
1698 /* Failed to read a mirror, discard this page so that
1699 * new page can be created with new mirror.
1701 * TODO: this is not needed after page reinit
1702 * route is implemented */
1703 cl_page_discard(env, io, page);
1705 cl_page_disown(env, io, page);
1708 /* TODO: discard all pages until page reinit route is implemented */
1709 cl_page_list_discard(env, io, &queue->c2_qin);
1711 /* Unlock unsent read pages in case of error. */
1712 cl_page_list_disown(env, io, &queue->c2_qin);
1714 cl_2queue_fini(env, queue);
1720 * Possible return value:
1721 * 0 no async readahead triggered and fast read could not be used.
1722 * 1 no async readahead, but fast read could be used.
1723 * 2 async readahead triggered and fast read could be used too.
1726 static int kickoff_async_readahead(struct file *file, unsigned long pages)
1728 struct ll_readahead_work *lrw;
1729 struct inode *inode = file_inode(file);
1730 struct ll_sb_info *sbi = ll_i2sbi(inode);
1731 struct ll_file_data *fd = file->private_data;
1732 struct ll_readahead_state *ras = &fd->fd_ras;
1733 struct ll_ra_info *ra = &sbi->ll_ra_info;
1734 unsigned long throttle;
1735 pgoff_t start_idx = ras_align(ras, ras->ras_next_readahead_idx);
1736 pgoff_t end_idx = start_idx + pages - 1;
1739 * In case we have a limited max_cached_mb, readahead
1740 * should be stopped if it have run out of all LRU slots.
1742 if (atomic_read(&ra->ra_cur_pages) >= sbi->ll_cache->ccc_lru_max) {
1743 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
1747 throttle = min(ra->ra_async_pages_per_file_threshold,
1748 ra->ra_max_pages_per_file);
1750 * If this is strided i/o or the window is smaller than the
1751 * throttle limit, we do not do async readahead. Otherwise,
1752 * we do async readahead, allowing the user thread to do fast i/o.
1754 if (stride_io_mode(ras) || !throttle ||
1755 ras->ras_window_pages < throttle ||
1756 atomic_read(&ra->ra_async_inflight) > ra->ra_async_max_active)
1759 if ((atomic_read(&ra->ra_cur_pages) + pages) > ra->ra_max_pages)
1762 if (ras->ras_async_last_readpage_idx == start_idx)
1765 /* ll_readahead_work_free() free it */
1768 atomic_inc(&sbi->ll_ra_info.ra_async_inflight);
1769 lrw->lrw_file = get_file(file);
1770 lrw->lrw_start_idx = start_idx;
1771 lrw->lrw_end_idx = end_idx;
1772 spin_lock(&ras->ras_lock);
1773 ras->ras_next_readahead_idx = end_idx + 1;
1774 ras->ras_async_last_readpage_idx = start_idx;
1775 spin_unlock(&ras->ras_lock);
1776 memcpy(lrw->lrw_jobid, ll_i2info(inode)->lli_jobid,
1777 sizeof(lrw->lrw_jobid));
1778 ll_readahead_work_add(inode, lrw);
1787 * Check if we can issue a readahead RPC, if that is
1788 * the case, we can't do fast IO because we will need
1789 * a cl_io to issue the RPC.
1791 static bool ll_use_fast_io(struct file *file,
1792 struct ll_readahead_state *ras, pgoff_t index)
1794 unsigned long fast_read_pages =
1795 max(RA_REMAIN_WINDOW_MIN, ras->ras_rpc_pages);
1797 loff_t stride_bytes = ras->ras_stride_bytes;
1799 if (stride_io_mode(ras) && stride_bytes) {
1800 skip_pages = (ras->ras_stride_length +
1801 ras->ras_stride_bytes - 1) / stride_bytes;
1802 skip_pages *= fast_read_pages;
1804 skip_pages = fast_read_pages;
1807 if (ras->ras_window_start_idx + ras->ras_window_pages <
1808 ras->ras_next_readahead_idx + skip_pages ||
1809 kickoff_async_readahead(file, fast_read_pages) > 0)
1815 int ll_readpage(struct file *file, struct page *vmpage)
1817 struct inode *inode = file_inode(file);
1818 struct cl_object *clob = ll_i2info(inode)->lli_clob;
1819 struct ll_cl_context *lcc;
1820 const struct lu_env *env = NULL;
1821 struct cl_io *io = NULL;
1822 struct cl_page *page;
1823 struct ll_sb_info *sbi = ll_i2sbi(inode);
1827 lcc = ll_cl_find(inode);
1833 if (io == NULL) { /* fast read */
1834 struct inode *inode = file_inode(file);
1835 struct ll_file_data *fd = file->private_data;
1836 struct ll_readahead_state *ras = &fd->fd_ras;
1837 struct lu_env *local_env = NULL;
1838 struct vvp_page *vpg;
1842 /* TODO: need to verify the layout version to make sure
1843 * the page is not invalid due to layout change. */
1844 page = cl_vmpage_page(vmpage, clob);
1846 unlock_page(vmpage);
1847 ll_ra_stats_inc_sbi(sbi, RA_STAT_FAILED_FAST_READ);
1851 vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1852 if (vpg->vpg_defer_uptodate) {
1853 enum ras_update_flags flags = LL_RAS_HIT;
1855 if (lcc && lcc->lcc_type == LCC_MMAP)
1856 flags |= LL_RAS_MMAP;
1858 /* For fast read, it updates read ahead state only
1859 * if the page is hit in cache because non cache page
1860 * case will be handled by slow read later. */
1861 ras_update(sbi, inode, ras, vvp_index(vpg), flags, io);
1862 /* avoid duplicate ras_update() call */
1863 vpg->vpg_ra_updated = 1;
1865 if (ll_use_fast_io(file, ras, vvp_index(vpg)))
1870 local_env = cl_env_percpu_get();
1874 /* export the page and skip io stack */
1876 vpg->vpg_ra_used = 1;
1877 cl_page_export(env, page, 1);
1879 ll_ra_stats_inc_sbi(sbi, RA_STAT_FAILED_FAST_READ);
1881 /* release page refcount before unlocking the page to ensure
1882 * the object won't be destroyed in the calling path of
1883 * cl_page_put(). Please see comment in ll_releasepage(). */
1884 cl_page_put(env, page);
1885 unlock_page(vmpage);
1887 cl_env_percpu_put(local_env);
1893 * Direct read can fall back to buffered read, but DIO is done
1894 * with lockless i/o, and buffered requires LDLM locking, so in
1895 * this case we must restart without lockless.
1897 if (file->f_flags & O_DIRECT &&
1898 lcc && lcc->lcc_type == LCC_RW &&
1900 unlock_page(vmpage);
1901 io->ci_dio_lock = 1;
1902 io->ci_need_restart = 1;
1906 LASSERT(io->ci_state == CIS_IO_GOING);
1907 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
1908 if (!IS_ERR(page)) {
1909 LASSERT(page->cp_type == CPT_CACHEABLE);
1910 if (likely(!PageUptodate(vmpage))) {
1911 cl_page_assume(env, io, page);
1913 result = ll_io_read_page(env, io, page, file);
1915 /* Page from a non-object file. */
1916 unlock_page(vmpage);
1919 cl_page_put(env, page);
1921 unlock_page(vmpage);
1922 result = PTR_ERR(page);