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
83 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi,
84 struct ra_io_arg *ria,
86 unsigned long pages_min)
88 struct ll_ra_info *ra = &sbi->ll_ra_info;
93 WARN_ON_ONCE(pages_min > pages);
95 * Don't try readahead aggresively if we are limited
96 * LRU pages, otherwise, it could cause deadlock.
98 pages = min(sbi->ll_cache->ccc_lru_max >> 2, pages);
100 * if this happen, we reserve more pages than needed,
101 * this will make us leak @ra_cur_pages, because
102 * ll_ra_count_put() acutally freed @pages.
104 if (unlikely(pages_min > pages))
108 * If read-ahead pages left are less than 1M, do not do read-ahead,
109 * otherwise it will form small read RPC(< 1M), which hurt server
112 ret = min(ra->ra_max_pages - atomic_read(&ra->ra_cur_pages),
114 if (ret < 0 || ret < min_t(long, PTLRPC_MAX_BRW_PAGES, pages))
117 if (atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) {
118 atomic_sub(ret, &ra->ra_cur_pages);
123 if (ret < pages_min) {
124 /* override ra limit for maximum performance */
125 atomic_add(pages_min - ret, &ra->ra_cur_pages);
131 void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long pages)
133 struct ll_ra_info *ra = &sbi->ll_ra_info;
135 atomic_sub(pages, &ra->ra_cur_pages);
138 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which)
140 LASSERTF(which < _NR_RA_STAT, "which: %u\n", which);
141 lprocfs_counter_incr(sbi->ll_ra_stats, which);
144 static inline bool ll_readahead_enabled(struct ll_sb_info *sbi)
146 return sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
147 sbi->ll_ra_info.ra_max_pages > 0;
150 void ll_ra_stats_inc(struct inode *inode, enum ra_stat which)
152 struct ll_sb_info *sbi = ll_i2sbi(inode);
154 ll_ra_stats_inc_sbi(sbi, which);
157 static void ll_ra_stats_add(struct inode *inode, enum ra_stat which, long count)
159 struct ll_sb_info *sbi = ll_i2sbi(inode);
161 LASSERTF(which < _NR_RA_STAT, "which: %u\n", which);
162 lprocfs_counter_add(sbi->ll_ra_stats, which, count);
165 #define RAS_CDEBUG(ras) \
167 "lre %llu cr %lu cb %llu wsi %lu wp %lu nra %lu rpc %lu " \
168 "r %lu csr %lu so %llu sb %llu sl %llu lr %lu\n", \
169 ras->ras_last_read_end_bytes, ras->ras_consecutive_requests, \
170 ras->ras_consecutive_bytes, ras->ras_window_start_idx, \
171 ras->ras_window_pages, ras->ras_next_readahead_idx, \
172 ras->ras_rpc_pages, ras->ras_requests, \
173 ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
174 ras->ras_stride_bytes, ras->ras_stride_length, \
175 ras->ras_async_last_readpage_idx)
177 static bool pos_in_window(loff_t pos, loff_t point,
178 unsigned long before, unsigned long after)
180 loff_t start = point - before;
181 loff_t end = point + after;
188 return start <= pos && pos <= end;
191 enum ll_ra_page_hint {
192 MAYNEED = 0, /* this page possibly accessed soon */
193 WILLNEED /* this page is gurateed to be needed */
197 * Initiates read-ahead of a page with given index.
199 * \retval +ve: page was already uptodate so it will be skipped
201 * \retval -ve: page wasn't added to \a queue for error;
202 * \retval 0: page was added into \a queue for read ahead.
204 static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io,
205 struct cl_page_list *queue, pgoff_t index,
206 enum ll_ra_page_hint hint)
208 struct cl_object *clob = io->ci_obj;
209 struct inode *inode = vvp_object_inode(clob);
210 struct page *vmpage = NULL;
212 enum ra_stat which = _NR_RA_STAT; /* keep gcc happy */
214 const char *msg = NULL;
221 * We need __GFP_NORETRY here for read-ahead page, otherwise
222 * the process will fail with OOM killed due to memcg limit.
223 * See @readahead_gfp_mask for an example.
225 vmpage = pagecache_get_page(inode->i_mapping, index,
226 FGP_LOCK | FGP_CREAT |
227 FGP_NOFS | FGP_NOWAIT,
228 mapping_gfp_mask(inode->i_mapping) |
229 __GFP_NORETRY | __GFP_NOWARN);
230 if (vmpage == NULL) {
231 which = RA_STAT_FAILED_GRAB_PAGE;
232 msg = "g_c_p_n failed";
233 GOTO(out, rc = -EBUSY);
237 vmpage = find_or_create_page(inode->i_mapping, index,
240 GOTO(out, rc = -ENOMEM);
243 /* should not come here */
244 GOTO(out, rc = -EINVAL);
247 /* Check if vmpage was truncated or reclaimed */
248 if (vmpage->mapping != inode->i_mapping) {
249 which = RA_STAT_WRONG_GRAB_PAGE;
250 msg = "g_c_p_n returned invalid page";
251 GOTO(out, rc = -EBUSY);
254 cp = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
256 which = RA_STAT_FAILED_GRAB_PAGE;
257 msg = "cl_page_find failed";
258 GOTO(out, rc = PTR_ERR(cp));
261 lu_ref_add(&cp->cp_reference, "ra", current);
262 cl_page_assume(env, io, cp);
264 if (!cp->cp_defer_uptodate && !PageUptodate(vmpage)) {
265 if (hint == MAYNEED) {
266 cp->cp_defer_uptodate = 1;
270 cl_page_list_add(queue, cp, true);
272 /* skip completed pages */
273 cl_page_unassume(env, io, cp);
274 /* This page is already uptodate, returning a positive number
275 * to tell the callers about this
280 lu_ref_del(&cp->cp_reference, "ra", current);
281 cl_page_put(env, cp);
284 if (vmpage != NULL) {
289 if (msg != NULL && hint == MAYNEED) {
290 ll_ra_stats_inc(inode, which);
291 CDEBUG(D_READA, "%s\n", msg);
298 #define RIA_DEBUG(ria) \
299 CDEBUG(D_READA, "rs %lu re %lu ro %llu rl %llu rb %llu\n", \
300 ria->ria_start_idx, ria->ria_end_idx, ria->ria_stoff, \
301 ria->ria_length, ria->ria_bytes)
303 static inline int stride_io_mode(struct ll_readahead_state *ras)
305 return ras->ras_consecutive_stride_requests > 1;
308 /* The function calculates how many bytes will be read in
309 * [off, off + length], in such stride IO area,
310 * stride_offset = st_off, stride_lengh = st_len,
311 * stride_bytes = st_bytes
313 * |------------------|*****|------------------|*****|------------|*****|....
316 * |----- st_len -----|
318 * How many bytes it should read in such pattern
319 * |-------------------------------------------------------------|
321 * |<------ length ------->|
323 * = |<----->| + |-------------------------------------| + |---|
324 * start_left st_bytes * i end_left
326 static loff_t stride_byte_count(loff_t st_off, loff_t st_len, loff_t st_bytes,
327 loff_t off, loff_t length)
329 u64 start = off > st_off ? off - st_off : 0;
330 u64 end = off + length > st_off ? off + length - st_off : 0;
335 if (st_len == 0 || length == 0 || end == 0)
338 start = div64_u64_rem(start, st_len, &start_left);
339 if (start_left < st_bytes)
340 start_left = st_bytes - start_left;
344 end = div64_u64_rem(end, st_len, &end_left);
345 if (end_left > st_bytes)
348 CDEBUG(D_READA, "start %llu, end %llu start_left %llu end_left %llu\n",
349 start, end, start_left, end_left);
352 bytes_count = end_left - (st_bytes - start_left);
354 bytes_count = start_left +
355 st_bytes * (end - start - 1) + end_left;
358 "st_off %llu, st_len %llu st_bytes %llu off %llu length %llu bytescount %llu\n",
359 st_off, st_len, st_bytes, off, length, bytes_count);
364 static unsigned long ria_page_count(struct ra_io_arg *ria)
366 loff_t length_bytes = ria->ria_end_idx >= ria->ria_start_idx ?
367 (loff_t)(ria->ria_end_idx -
368 ria->ria_start_idx + 1) << PAGE_SHIFT : 0;
371 if (ria->ria_length > ria->ria_bytes && ria->ria_bytes &&
372 (ria->ria_length & ~PAGE_MASK || ria->ria_bytes & ~PAGE_MASK ||
373 ria->ria_stoff & ~PAGE_MASK)) {
374 /* Over-estimate un-aligned page stride read */
375 unsigned long pg_count = ((ria->ria_bytes +
376 PAGE_SIZE - 1) >> PAGE_SHIFT) + 1;
377 pg_count *= length_bytes / ria->ria_length + 1;
381 bytes_count = stride_byte_count(ria->ria_stoff, ria->ria_length,
383 (loff_t)ria->ria_start_idx<<PAGE_SHIFT,
385 return (bytes_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
388 static pgoff_t ras_align(struct ll_readahead_state *ras, pgoff_t index)
390 unsigned int opt_size = min(ras->ras_window_pages, ras->ras_rpc_pages);
394 return index - (index % opt_size);
397 /* Check whether the index is in the defined ra-window */
398 static bool ras_inside_ra_window(pgoff_t idx, struct ra_io_arg *ria)
400 loff_t pos = (loff_t)idx << PAGE_SHIFT;
402 /* If ria_length == ria_bytes, it means non-stride I/O mode,
403 * idx should always inside read-ahead window in this case
404 * For stride I/O mode, just check whether the idx is inside
407 if (ria->ria_length == 0 || ria->ria_length == ria->ria_bytes)
410 if (pos >= ria->ria_stoff) {
413 div64_u64_rem(pos - ria->ria_stoff, ria->ria_length, &offset);
415 if (offset < ria->ria_bytes ||
416 (ria->ria_length - offset) < PAGE_SIZE)
418 } else if (pos + PAGE_SIZE > ria->ria_stoff) {
426 ll_read_ahead_pages(const struct lu_env *env, struct cl_io *io,
427 struct cl_page_list *queue, struct ll_readahead_state *ras,
428 struct ra_io_arg *ria, pgoff_t *ra_end, pgoff_t skip_index)
430 struct cl_read_ahead ra = { 0 };
431 /* busy page count is per stride */
432 int rc = 0, count = 0, busy_page_count = 0;
435 LASSERT(ria != NULL);
438 for (page_idx = ria->ria_start_idx;
439 page_idx <= ria->ria_end_idx && ria->ria_reserved > 0;
441 if (skip_index && page_idx == skip_index)
443 if (ras_inside_ra_window(page_idx, ria)) {
444 if (ra.cra_end_idx == 0 || ra.cra_end_idx < page_idx) {
448 * Do not shrink ria_end_idx at any case until
449 * the minimum end of current read is covered.
451 * Do not extend read lock accross stripe if
452 * lock contention detected.
454 if (ra.cra_contention &&
455 page_idx > ria->ria_end_idx_min) {
456 ria->ria_end_idx = *ra_end;
460 cl_read_ahead_release(env, &ra);
462 rc = cl_io_read_ahead(env, io, page_idx, &ra);
467 * Only shrink ria_end_idx if the matched
468 * LDLM lock doesn't cover more.
470 if (page_idx > ra.cra_end_idx) {
471 ria->ria_end_idx = ra.cra_end_idx;
475 CDEBUG(D_READA, "idx: %lu, ra: %lu, rpc: %lu\n",
476 page_idx, ra.cra_end_idx,
478 LASSERTF(ra.cra_end_idx >= page_idx,
479 "object: %px, indcies %lu / %lu\n",
480 io->ci_obj, ra.cra_end_idx, page_idx);
481 /* update read ahead RPC size.
482 * NB: it's racy but doesn't matter
484 if (ras->ras_rpc_pages != ra.cra_rpc_pages &&
485 ra.cra_rpc_pages > 0)
486 ras->ras_rpc_pages = ra.cra_rpc_pages;
488 /* trim (align with optimal RPC size) */
489 end_idx = ras_align(ras,
490 ria->ria_end_idx + 1);
491 if (end_idx > 0 && !ria->ria_eof)
492 ria->ria_end_idx = end_idx - 1;
494 if (ria->ria_end_idx < ria->ria_end_idx_min)
495 ria->ria_end_idx = ria->ria_end_idx_min;
497 if (page_idx > ria->ria_end_idx)
500 /* If the page is inside the read-ahead window */
501 rc = ll_read_ahead_page(env, io, queue, page_idx,
503 if (rc < 0 && rc != -EBUSY)
508 "skip busy page: %lu\n", page_idx);
509 /* For page unaligned readahead the first
510 * last pages of each region can be read by
511 * another reader on the same node, and so
512 * may be busy. So only stop for > 2 busy
515 if (busy_page_count > 2)
520 /* Only subtract from reserve & count the page if we
521 * really did readahead on that page.
527 } else if (stride_io_mode(ras)) {
528 /* If it is not in the read-ahead window, and it is
529 * read-ahead mode, then check whether it should skip
532 loff_t pos = (loff_t)page_idx << PAGE_SHIFT;
535 div64_u64_rem(pos - ria->ria_stoff, ria->ria_length,
537 if (offset >= ria->ria_bytes) {
538 pos += (ria->ria_length - offset);
539 if ((pos >> PAGE_SHIFT) >= page_idx + 1)
540 page_idx = (pos >> PAGE_SHIFT) - 1;
543 "Stride: jump %llu pages to %lu\n",
544 ria->ria_length - offset, page_idx);
550 cl_read_ahead_release(env, &ra);
553 ll_ra_stats_add(vvp_object_inode(io->ci_obj),
554 RA_STAT_READAHEAD_PAGES, count);
559 static void ll_readahead_work_free(struct ll_readahead_work *work)
561 fput(work->lrw_file);
565 static void ll_readahead_handle_work(struct work_struct *wq);
566 static void ll_readahead_work_add(struct inode *inode,
567 struct ll_readahead_work *work)
569 INIT_WORK(&work->lrw_readahead_work, ll_readahead_handle_work);
570 queue_work(ll_i2sbi(inode)->ll_ra_info.ll_readahead_wq,
571 &work->lrw_readahead_work);
574 static int ll_readahead_file_kms(const struct lu_env *env,
575 struct cl_io *io, __u64 *kms)
577 struct cl_object *clob;
579 struct cl_attr *attr = vvp_env_thread_attr(env);
583 inode = vvp_object_inode(clob);
585 cl_object_attr_lock(clob);
586 ret = cl_object_attr_get(env, clob, attr);
587 cl_object_attr_unlock(clob);
592 *kms = attr->cat_kms;
596 static void ll_readahead_handle_work(struct work_struct *wq)
598 struct ll_readahead_work *work;
601 struct ra_io_arg *ria;
603 struct ll_file_data *fd;
604 struct ll_readahead_state *ras;
606 struct cl_2queue *queue;
607 pgoff_t ra_end_idx = 0;
608 unsigned long pages, pages_min = 0;
613 struct ll_sb_info *sbi;
615 work = container_of(wq, struct ll_readahead_work,
617 fd = work->lrw_file->private_data;
619 file = work->lrw_file;
620 inode = file_inode(file);
621 sbi = ll_i2sbi(inode);
623 CDEBUG(D_READA|D_IOTRACE,
624 "%s:"DFID": async ra from %lu to %lu triggered by user pid %d\n",
625 file_dentry(file)->d_name.name, PFID(ll_inode2fid(inode)),
626 work->lrw_start_idx, work->lrw_end_idx, work->lrw_user_pid);
628 env = cl_env_alloc(&refcheck, LCT_NOREF);
630 GOTO(out_free_work, rc = PTR_ERR(env));
632 io = vvp_env_thread_io(env);
633 ll_io_init(io, file, CIT_READ, NULL);
635 rc = ll_readahead_file_kms(env, io, &kms);
637 GOTO(out_put_env, rc);
640 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
641 GOTO(out_put_env, rc = 0);
644 ria = &ll_env_info(env)->lti_ria;
645 memset(ria, 0, sizeof(*ria));
647 ria->ria_start_idx = work->lrw_start_idx;
648 /* Truncate RA window to end of file */
649 eof_index = (pgoff_t)(kms - 1) >> PAGE_SHIFT;
650 if (eof_index <= work->lrw_end_idx) {
651 work->lrw_end_idx = eof_index;
654 if (work->lrw_end_idx <= work->lrw_start_idx)
655 GOTO(out_put_env, rc = 0);
657 ria->ria_end_idx = work->lrw_end_idx;
658 pages = ria->ria_end_idx - ria->ria_start_idx + 1;
659 ria->ria_reserved = ll_ra_count_get(sbi, ria,
660 ria_page_count(ria), pages_min);
663 "async reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
664 ria->ria_reserved, pages, pages_min,
665 atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
666 ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
668 if (ria->ria_reserved < pages) {
669 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
670 if (PAGES_TO_MiB(ria->ria_reserved) < 1) {
671 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
672 GOTO(out_put_env, rc = 0);
676 rc = cl_io_rw_init(env, io, CIT_READ, ria->ria_start_idx, pages);
678 GOTO(out_put_env, rc);
680 /* overwrite jobid inited in vvp_io_init() */
681 if (strncmp(ll_i2info(inode)->lli_jobid, work->lrw_jobid,
682 sizeof(work->lrw_jobid)))
683 memcpy(ll_i2info(inode)->lli_jobid, work->lrw_jobid,
684 sizeof(work->lrw_jobid));
686 vvp_env_io(env)->vui_fd = fd;
687 io->ci_state = CIS_LOCKED;
688 io->ci_async_readahead = true;
689 rc = cl_io_start(env, io);
691 GOTO(out_io_fini, rc);
693 queue = &io->ci_queue;
694 cl_2queue_init(queue);
696 rc = ll_read_ahead_pages(env, io, &queue->c2_qin, ras, ria,
698 if (ria->ria_reserved != 0)
699 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
700 if (queue->c2_qin.pl_nr > 0) {
701 int count = queue->c2_qin.pl_nr;
703 rc = cl_io_submit_rw(env, io, CRT_READ, queue);
705 task_io_account_read(PAGE_SIZE * count);
707 if (ria->ria_end_idx == ra_end_idx && ra_end_idx == (kms >> PAGE_SHIFT))
708 ll_ra_stats_inc(inode, RA_STAT_EOF);
710 if (ra_end_idx != ria->ria_end_idx)
711 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
713 /* TODO: discard all pages until page reinit route is implemented */
714 cl_page_list_discard(env, io, &queue->c2_qin);
716 /* Unlock unsent read pages in case of error. */
717 cl_page_list_disown(env, &queue->c2_qin);
719 cl_2queue_fini(env, queue);
724 cl_env_put(env, &refcheck);
727 ll_ra_stats_inc_sbi(ll_i2sbi(inode), RA_STAT_ASYNC);
728 atomic_dec(&sbi->ll_ra_info.ra_async_inflight);
729 ll_readahead_work_free(work);
732 static int ll_readahead(const struct lu_env *env, struct cl_io *io,
733 struct cl_page_list *queue,
734 struct ll_readahead_state *ras, bool hit,
735 struct file *file, pgoff_t skip_index,
738 struct vvp_io *vio = vvp_env_io(env);
739 struct ll_thread_info *lti = ll_env_info(env);
740 unsigned long pages, pages_min = 0;
741 pgoff_t ra_end_idx = 0, end_idx = 0;
743 struct ra_io_arg *ria = <i->lti_ria;
744 struct cl_object *clob;
747 struct ll_sb_info *sbi;
748 struct ll_ra_info *ra;
753 inode = vvp_object_inode(clob);
754 sbi = ll_i2sbi(inode);
755 ra = &sbi->ll_ra_info;
758 * In case we have a limited max_cached_mb, readahead
759 * should be stopped if it have run out of all LRU slots.
761 if (atomic_read(&ra->ra_cur_pages) >= sbi->ll_cache->ccc_lru_max) {
762 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
766 memset(ria, 0, sizeof(*ria));
767 ret = ll_readahead_file_kms(env, io, &kms);
772 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
776 spin_lock(&ras->ras_lock);
779 * Note: other thread might rollback the ras_next_readahead_idx,
780 * if it can not get the full size of prepared pages, see the
781 * end of this function. For stride read ahead, it needs to
782 * make sure the offset is no less than ras_stride_offset,
783 * so that stride read ahead can work correctly.
785 if (stride_io_mode(ras))
786 *start_idx = max_t(pgoff_t, ras->ras_next_readahead_idx,
787 ras->ras_stride_offset >> PAGE_SHIFT);
789 *start_idx = ras->ras_next_readahead_idx;
791 if (ras->ras_window_pages > 0)
792 end_idx = ras->ras_window_start_idx + ras->ras_window_pages - 1;
795 end_idx = *start_idx + ras->ras_window_pages - 1;
797 /* Enlarge the RA window to encompass the full read */
798 if (vio->vui_ra_valid &&
799 end_idx < vio->vui_ra_start_idx + vio->vui_ra_pages - 1)
800 end_idx = vio->vui_ra_start_idx + vio->vui_ra_pages - 1;
805 /* Truncate RA window to end of file */
806 eof_index = (pgoff_t)((kms - 1) >> PAGE_SHIFT);
807 if (eof_index <= end_idx) {
812 ria->ria_start_idx = *start_idx;
813 ria->ria_end_idx = end_idx;
814 /* If stride I/O mode is detected, get stride window*/
815 if (stride_io_mode(ras)) {
816 ria->ria_stoff = ras->ras_stride_offset;
817 ria->ria_length = ras->ras_stride_length;
818 ria->ria_bytes = ras->ras_stride_bytes;
820 spin_unlock(&ras->ras_lock);
822 pages = ria_page_count(ria);
826 DFID": ria: %lu/%lu, bead: %lu/%lu, pages %lu, hit: %d\n",
827 PFID(lu_object_fid(&clob->co_lu)),
828 ria->ria_start_idx, ria->ria_end_idx,
829 vio->vui_ra_valid ? vio->vui_ra_start_idx : 0,
830 vio->vui_ra_valid ? vio->vui_ra_pages : 0,
834 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
838 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
842 /* at least to extend the readahead window to cover current read */
843 if (!hit && vio->vui_ra_valid &&
844 vio->vui_ra_start_idx + vio->vui_ra_pages > ria->ria_start_idx) {
845 ria->ria_end_idx_min =
846 vio->vui_ra_start_idx + vio->vui_ra_pages - 1;
847 pages_min = vio->vui_ra_start_idx + vio->vui_ra_pages -
850 * For performance reason, exceeding @ra_max_pages
851 * are allowed, but this should be limited with RPC
852 * size in case a large block size read issued. Trim
855 pages_min = min(pages_min, ras->ras_rpc_pages -
856 (ria->ria_start_idx % ras->ras_rpc_pages));
859 /* don't over reserved for mmap range read */
862 if (pages_min > pages)
864 ria->ria_reserved = ll_ra_count_get(ll_i2sbi(inode), ria, pages,
866 if (ria->ria_reserved < pages)
867 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
869 CDEBUG(D_READA, "reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
870 ria->ria_reserved, pages, pages_min,
871 atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
872 ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
874 ret = ll_read_ahead_pages(env, io, queue, ras, ria, &ra_end_idx,
876 if (ria->ria_reserved != 0)
877 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
879 if (ra_end_idx == end_idx && ra_end_idx == (kms >> PAGE_SHIFT))
880 ll_ra_stats_inc(inode, RA_STAT_EOF);
883 "ra_end_idx = %lu end_idx = %lu stride end = %lu pages = %d\n",
884 ra_end_idx, end_idx, ria->ria_end_idx, ret);
886 if (ra_end_idx != end_idx)
887 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
888 if (ra_end_idx > 0) {
889 /* update the ras so that the next read-ahead tries from
892 spin_lock(&ras->ras_lock);
893 ras->ras_next_readahead_idx = ra_end_idx + 1;
894 spin_unlock(&ras->ras_lock);
901 static int ll_readpages(const struct lu_env *env, struct cl_io *io,
902 struct cl_page_list *queue,
903 pgoff_t start, pgoff_t end)
912 ret = ll_readahead_file_kms(env, io, &kms);
920 unsigned long end_index;
922 end_index = (unsigned long)((kms - 1) >> PAGE_SHIFT);
923 if (end_index <= end)
927 for (page_idx = start; page_idx <= end; page_idx++) {
928 ret = ll_read_ahead_page(env, io, queue, page_idx,
932 else if (ret == 0) /* ret 1 is already uptodate */
936 RETURN(count > 0 ? count : ret);
939 /* called with the ras_lock held or from places where it doesn't matter */
940 static void ras_reset(struct ll_readahead_state *ras, pgoff_t index)
942 ras->ras_consecutive_requests = 0;
943 ras->ras_consecutive_bytes = 0;
944 ras->ras_window_pages = 0;
945 ras->ras_window_start_idx = ras_align(ras, index);
946 ras->ras_next_readahead_idx = max(ras->ras_window_start_idx, index + 1);
951 /* called with the ras_lock held or from places where it doesn't matter */
952 static void ras_stride_reset(struct ll_readahead_state *ras)
954 ras->ras_consecutive_stride_requests = 0;
955 ras->ras_stride_length = 0;
956 ras->ras_stride_bytes = 0;
960 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
962 spin_lock_init(&ras->ras_lock);
963 ras->ras_rpc_pages = PTLRPC_MAX_BRW_PAGES;
965 ras->ras_last_read_end_bytes = 0;
966 ras->ras_requests = 0;
967 ras->ras_range_min_start_idx = 0;
968 ras->ras_range_max_end_idx = 0;
969 ras->ras_range_requests = 0;
970 ras->ras_last_range_pages = 0;
974 * Check whether the read request is in the stride window.
975 * If it is in the stride window, return true, otherwise return false.
977 static bool read_in_stride_window(struct ll_readahead_state *ras,
978 loff_t pos, loff_t bytes)
982 if (ras->ras_stride_length == 0 || ras->ras_stride_bytes == 0 ||
983 ras->ras_stride_bytes == ras->ras_stride_length)
986 stride_gap = pos - ras->ras_last_read_end_bytes - 1;
988 /* If it is contiguous read */
990 return ras->ras_consecutive_bytes + bytes <=
991 ras->ras_stride_bytes;
993 /* Otherwise check the stride by itself */
994 return (ras->ras_stride_length - ras->ras_stride_bytes) == stride_gap &&
995 ras->ras_consecutive_bytes == ras->ras_stride_bytes &&
996 bytes <= ras->ras_stride_bytes;
999 static void ras_init_stride_detector(struct ll_readahead_state *ras,
1000 loff_t pos, loff_t bytes)
1002 loff_t stride_gap = pos - ras->ras_last_read_end_bytes - 1;
1004 LASSERT(ras->ras_consecutive_stride_requests == 0);
1006 if (pos <= ras->ras_last_read_end_bytes) {
1007 /* Reset stride window for forward read */
1008 ras_stride_reset(ras);
1012 ras->ras_stride_bytes = ras->ras_consecutive_bytes;
1013 ras->ras_stride_length = stride_gap + ras->ras_consecutive_bytes;
1014 ras->ras_consecutive_stride_requests++;
1015 ras->ras_stride_offset = pos;
1020 static unsigned long
1021 stride_page_count(struct ll_readahead_state *ras, loff_t len)
1023 loff_t bytes_count =
1024 stride_byte_count(ras->ras_stride_offset,
1025 ras->ras_stride_length, ras->ras_stride_bytes,
1026 ras->ras_window_start_idx << PAGE_SHIFT, len);
1028 return (bytes_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
1031 /* Stride Read-ahead window will be increased inc_len according to
1032 * stride I/O pattern
1034 static void ras_stride_increase_window(struct ll_readahead_state *ras,
1035 struct ll_ra_info *ra, loff_t inc_bytes)
1037 loff_t window_bytes, stride_bytes;
1042 /* temporarily store in page units to reduce LASSERT() cost below */
1043 end = ras->ras_window_start_idx + ras->ras_window_pages;
1045 LASSERT(ras->ras_stride_length > 0);
1046 LASSERTF(end >= (ras->ras_stride_offset >> PAGE_SHIFT),
1047 "window_start_idx %lu, window_pages %lu stride_offset %llu\n",
1048 ras->ras_window_start_idx, ras->ras_window_pages,
1049 ras->ras_stride_offset);
1052 if (end <= ras->ras_stride_offset)
1055 stride_bytes = end - ras->ras_stride_offset;
1057 div64_u64_rem(stride_bytes, ras->ras_stride_length, &left_bytes);
1058 window_bytes = (ras->ras_window_pages << PAGE_SHIFT);
1059 if (left_bytes < ras->ras_stride_bytes) {
1060 if (ras->ras_stride_bytes - left_bytes >= inc_bytes) {
1061 window_bytes += inc_bytes;
1064 window_bytes += (ras->ras_stride_bytes - left_bytes);
1065 inc_bytes -= (ras->ras_stride_bytes - left_bytes);
1068 window_bytes += (ras->ras_stride_length - left_bytes);
1071 LASSERT(ras->ras_stride_bytes != 0);
1073 step = div64_u64_rem(inc_bytes, ras->ras_stride_bytes, &left_bytes);
1075 window_bytes += step * ras->ras_stride_length + left_bytes;
1076 LASSERT(window_bytes > 0);
1079 if (stride_page_count(ras, window_bytes) <=
1080 ra->ra_max_pages_per_file || ras->ras_window_pages == 0)
1081 ras->ras_window_pages = (window_bytes >> PAGE_SHIFT);
1083 LASSERT(ras->ras_window_pages > 0);
1088 static void ras_increase_window(struct inode *inode,
1089 struct ll_readahead_state *ras,
1090 struct ll_ra_info *ra)
1092 /* The stretch of ra-window should be aligned with max rpc_size
1093 * but current clio architecture does not support retrieve such
1094 * information from lower layer. FIXME later
1096 if (stride_io_mode(ras)) {
1097 ras_stride_increase_window(ras, ra,
1098 (loff_t)ras->ras_rpc_pages << PAGE_SHIFT);
1100 pgoff_t window_pages;
1102 window_pages = min(ras->ras_window_pages + ras->ras_rpc_pages,
1103 ra->ra_max_pages_per_file);
1104 if (window_pages < ras->ras_rpc_pages)
1105 ras->ras_window_pages = window_pages;
1107 ras->ras_window_pages = ras_align(ras, window_pages);
1112 * Seek within 8 pages are considered as sequential read for now.
1114 static inline bool is_loose_seq_read(struct ll_readahead_state *ras, loff_t pos)
1116 return pos_in_window(pos, ras->ras_last_read_end_bytes,
1117 8UL << PAGE_SHIFT, 8UL << PAGE_SHIFT);
1120 static inline bool is_loose_mmap_read(struct ll_sb_info *sbi,
1121 struct ll_readahead_state *ras,
1124 unsigned long range_pages = sbi->ll_ra_info.ra_range_pages;
1126 return pos_in_window(pos, ras->ras_last_read_end_bytes,
1127 range_pages << PAGE_SHIFT,
1128 range_pages << PAGE_SHIFT);
1132 * We have observed slow mmap read performances for some
1133 * applications. The problem is if access pattern is neither
1134 * sequential nor stride, but could be still adjacent in a
1135 * small range and then seek a random position.
1137 * So the pattern could be something like this:
1139 * [1M data] [hole] [0.5M data] [hole] [0.7M data] [1M data]
1142 * Every time an application reads mmap data, it may not only
1143 * read a single 4KB page, but aslo a cluster of nearby pages in
1144 * a range(e.g. 1MB) of the first page after a cache miss.
1146 * The readahead engine is modified to track the range size of
1147 * a cluster of mmap reads, so that after a seek and/or cache miss,
1148 * the range size is used to efficiently prefetch multiple pages
1149 * in a single RPC rather than many small RPCs.
1151 static void ras_detect_cluster_range(struct ll_readahead_state *ras,
1152 struct ll_sb_info *sbi,
1153 unsigned long pos, unsigned long count)
1155 pgoff_t last_pages, pages;
1156 pgoff_t end_idx = (pos + count - 1) >> PAGE_SHIFT;
1158 last_pages = ras->ras_range_max_end_idx -
1159 ras->ras_range_min_start_idx + 1;
1160 /* First time come here */
1161 if (!ras->ras_range_max_end_idx)
1164 /* Random or Stride read */
1165 if (!is_loose_mmap_read(sbi, ras, pos))
1168 ras->ras_range_requests++;
1169 if (ras->ras_range_max_end_idx < end_idx)
1170 ras->ras_range_max_end_idx = end_idx;
1172 if (ras->ras_range_min_start_idx > (pos >> PAGE_SHIFT))
1173 ras->ras_range_min_start_idx = pos >> PAGE_SHIFT;
1175 /* Out of range, consider it as random or stride */
1176 pages = ras->ras_range_max_end_idx -
1177 ras->ras_range_min_start_idx + 1;
1178 if (pages <= sbi->ll_ra_info.ra_range_pages)
1181 ras->ras_last_range_pages = last_pages;
1182 ras->ras_range_requests = 0;
1183 ras->ras_range_min_start_idx = pos >> PAGE_SHIFT;
1184 ras->ras_range_max_end_idx = end_idx;
1187 static void ras_detect_read_pattern(struct ll_readahead_state *ras,
1188 struct ll_sb_info *sbi,
1189 loff_t pos, size_t bytes, bool mmap)
1191 bool stride_detect = false;
1192 pgoff_t index = pos >> PAGE_SHIFT;
1196 * Reset the read-ahead window in two cases. First when the app seeks
1197 * or reads to some other part of the file. Secondly if we get a
1198 * read-ahead miss that we think we've previously issued. This can
1199 * be a symptom of there being so many read-ahead pages that the VM
1200 * is reclaiming it before we get to it.
1202 if (!is_loose_seq_read(ras, pos)) {
1203 /* Check whether it is in stride I/O mode */
1204 if (!read_in_stride_window(ras, pos, bytes)) {
1205 if (ras->ras_consecutive_stride_requests == 0)
1206 ras_init_stride_detector(ras, pos, bytes);
1208 ras_stride_reset(ras);
1209 ras->ras_consecutive_bytes = 0;
1210 ras_reset(ras, index);
1212 ras->ras_consecutive_bytes = 0;
1213 ras->ras_consecutive_requests = 0;
1214 if (++ras->ras_consecutive_stride_requests > 1)
1215 stride_detect = true;
1218 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
1219 } else if (stride_io_mode(ras)) {
1221 * If this is contiguous read but in stride I/O mode
1222 * currently, check whether stride step still is valid,
1223 * if invalid, it will reset the stride ra window to
1226 if (!read_in_stride_window(ras, pos, bytes)) {
1227 ras_stride_reset(ras);
1228 ras->ras_window_pages = 0;
1229 ras->ras_next_readahead_idx = index;
1233 ras->ras_consecutive_bytes += bytes;
1235 pgoff_t idx = ras->ras_consecutive_bytes >> PAGE_SHIFT;
1236 unsigned long ra_range_pages =
1237 max_t(unsigned long, RA_MIN_MMAP_RANGE_PAGES,
1238 sbi->ll_ra_info.ra_range_pages);
1240 if ((idx >= ra_range_pages &&
1241 idx % ra_range_pages == 0) || stride_detect)
1242 ras->ras_need_increase_window = true;
1243 } else if ((ras->ras_consecutive_requests > 1 || stride_detect)) {
1244 ras->ras_need_increase_window = true;
1247 ras->ras_last_read_end_bytes = pos + bytes - 1;
1251 void ll_ras_enter(struct file *f, loff_t pos, size_t bytes)
1253 struct ll_file_data *fd = f->private_data;
1254 struct ll_readahead_state *ras = &fd->fd_ras;
1255 struct inode *inode = file_inode(f);
1256 struct ll_sb_info *sbi = ll_i2sbi(inode);
1258 spin_lock(&ras->ras_lock);
1259 ras->ras_requests++;
1260 ras->ras_consecutive_requests++;
1261 ras->ras_need_increase_window = false;
1262 ras->ras_whole_file_read = false;
1264 * On the second access to a file smaller than the tunable
1265 * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1266 * file up to ra_max_pages_per_file. This is simply a best effort
1267 * and only occurs once per open file. Normal RA behavior is reverted
1268 * to for subsequent IO.
1270 if (ras->ras_requests >= 2) {
1272 struct ll_ra_info *ra = &sbi->ll_ra_info;
1274 kms_pages = (i_size_read(inode) + PAGE_SIZE - 1) >>
1277 CDEBUG(D_READA, "kmsp %llu mwp %lu mp %lu\n", kms_pages,
1278 ra->ra_max_read_ahead_whole_pages,
1279 ra->ra_max_pages_per_file);
1282 kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1283 ras->ras_whole_file_read = true;
1284 ras->ras_window_start_idx = 0;
1285 ras->ras_next_readahead_idx = 0;
1286 ras->ras_window_pages = min(ra->ra_max_pages_per_file,
1287 ra->ra_max_read_ahead_whole_pages);
1288 GOTO(out_unlock, 0);
1291 ras_detect_read_pattern(ras, sbi, pos, bytes, false);
1293 spin_unlock(&ras->ras_lock);
1296 static bool index_in_stride_window(struct ll_readahead_state *ras,
1299 loff_t pos = (loff_t)index << PAGE_SHIFT;
1301 if (ras->ras_stride_length == 0 || ras->ras_stride_bytes == 0 ||
1302 ras->ras_stride_bytes == ras->ras_stride_length)
1305 if (pos >= ras->ras_stride_offset) {
1308 div64_u64_rem(pos - ras->ras_stride_offset,
1309 ras->ras_stride_length, &offset);
1310 if (offset < ras->ras_stride_bytes ||
1311 ras->ras_stride_length - offset < PAGE_SIZE)
1313 } else if (ras->ras_stride_offset - pos < PAGE_SIZE) {
1321 * ll_ras_enter() is used to detect read pattern according to pos and count.
1323 * ras_update() is used to detect cache miss and
1324 * reset window or increase window accordingly
1326 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
1327 struct ll_readahead_state *ras, pgoff_t index,
1328 enum ras_update_flags flags, struct cl_io *io)
1330 struct ll_ra_info *ra = &sbi->ll_ra_info;
1331 bool hit = flags & LL_RAS_HIT;
1334 spin_lock(&ras->ras_lock);
1339 CDEBUG(D_READA|D_IOTRACE, DFID " pages at %lu miss.\n",
1340 PFID(ll_inode2fid(inode)), index);
1341 ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
1344 * The readahead window has been expanded to cover whole
1345 * file size, we don't care whether ra miss happen or not.
1346 * Because we will read whole file to page cache even if
1347 * some pages missed.
1349 if (ras->ras_whole_file_read)
1350 GOTO(out_unlock, 0);
1352 if (io && io->ci_rand_read)
1353 GOTO(out_unlock, 0);
1355 if (io && io->ci_seq_read) {
1357 /* to avoid many small read RPC here */
1358 ras->ras_window_pages = sbi->ll_ra_info.ra_range_pages;
1359 ll_ra_stats_inc_sbi(sbi, RA_STAT_MMAP_RANGE_READ);
1361 goto skip_miss_checking;
1364 if (flags & LL_RAS_MMAP) {
1365 unsigned long ra_pages;
1367 ras_detect_cluster_range(ras, sbi, index << PAGE_SHIFT,
1369 ras_detect_read_pattern(ras, sbi, (loff_t)index << PAGE_SHIFT,
1372 /* we did not detect anything but we could prefetch */
1373 if (!ras->ras_need_increase_window &&
1374 ras->ras_window_pages <= sbi->ll_ra_info.ra_range_pages &&
1375 ras->ras_range_requests >= 2) {
1377 ra_pages = max_t(unsigned long,
1378 RA_MIN_MMAP_RANGE_PAGES,
1379 ras->ras_last_range_pages);
1380 if (index < ra_pages / 2)
1383 index -= ra_pages / 2;
1384 ras->ras_window_pages = ra_pages;
1385 ll_ra_stats_inc_sbi(sbi,
1386 RA_STAT_MMAP_RANGE_READ);
1388 ras->ras_window_pages = 0;
1390 goto skip_miss_checking;
1394 if (!hit && ras->ras_window_pages &&
1395 index < ras->ras_next_readahead_idx &&
1396 pos_in_window(index, ras->ras_window_start_idx, 0,
1397 ras->ras_window_pages)) {
1398 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
1399 ras->ras_need_increase_window = false;
1401 if (index_in_stride_window(ras, index) &&
1402 stride_io_mode(ras)) {
1404 * if (index != ras->ras_last_readpage + 1)
1405 * ras->ras_consecutive_pages = 0;
1407 ras_reset(ras, index);
1410 * If stride-RA hit cache miss, the stride
1411 * detector will not be reset to avoid the
1412 * overhead of redetecting read-ahead mode,
1413 * but on the condition that the stride window
1414 * is still intersect with normal sequential
1415 * read-ahead window.
1417 if (ras->ras_window_start_idx < ras->ras_stride_offset)
1418 ras_stride_reset(ras);
1422 * Reset both stride window and normal RA
1425 ras_reset(ras, index);
1426 /* ras->ras_consecutive_pages++; */
1427 ras->ras_consecutive_bytes = 0;
1428 ras_stride_reset(ras);
1429 GOTO(out_unlock, 0);
1434 ras->ras_window_start_idx = ras_align(ras, index);
1436 if (stride_io_mode(ras)) {
1437 /* Since stride readahead is sentivite to the offset
1438 * of read-ahead, so we use original offset here,
1439 * instead of ras_window_start_idx, which is RPC aligned.
1441 ras->ras_next_readahead_idx = max(index + 1,
1442 ras->ras_next_readahead_idx);
1443 ras->ras_window_start_idx =
1444 max_t(pgoff_t, ras->ras_window_start_idx,
1445 ras->ras_stride_offset >> PAGE_SHIFT);
1447 if (ras->ras_next_readahead_idx < ras->ras_window_start_idx)
1448 ras->ras_next_readahead_idx = ras->ras_window_start_idx;
1450 ras->ras_next_readahead_idx = index + 1;
1453 if (ras->ras_need_increase_window) {
1454 ras_increase_window(inode, ras, ra);
1455 ras->ras_need_increase_window = false;
1461 spin_unlock(&ras->ras_lock);
1464 int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
1466 struct inode *inode = vmpage->mapping->host;
1467 struct ll_inode_info *lli = ll_i2info(inode);
1470 struct cl_page *page;
1471 struct cl_object *clob;
1472 bool redirtied = false;
1473 bool unlocked = false;
1479 LASSERT(PageLocked(vmpage));
1480 LASSERT(!PageWriteback(vmpage));
1482 LASSERT(ll_i2dtexp(inode) != NULL);
1484 env = cl_env_get(&refcheck);
1486 GOTO(out, result = PTR_ERR(env));
1488 clob = ll_i2info(inode)->lli_clob;
1489 LASSERT(clob != NULL);
1491 io = vvp_env_thread_io(env);
1493 io->ci_ignore_layout = 1;
1494 result = cl_io_init(env, io, CIT_MISC, clob);
1496 page = cl_page_find(env, clob, vmpage->index,
1497 vmpage, CPT_CACHEABLE);
1498 if (!IS_ERR(page)) {
1499 lu_ref_add(&page->cp_reference, "writepage",
1501 cl_page_assume(env, io, page);
1502 result = cl_page_flush(env, io, page);
1505 * Re-dirty page on error so it retries write,
1506 * but not in case when IO has actually
1507 * occurred and completed with an error.
1509 if (!PageError(vmpage)) {
1510 redirty_page_for_writepage(wbc, vmpage);
1515 cl_page_disown(env, io, page);
1517 lu_ref_del(&page->cp_reference,
1518 "writepage", current);
1519 cl_page_put(env, page);
1521 result = PTR_ERR(page);
1524 cl_io_fini(env, io);
1526 if (redirtied && wbc->sync_mode == WB_SYNC_ALL) {
1527 loff_t offset = vmpage->index << PAGE_SHIFT;
1529 /* Flush page failed because the extent is being written out.
1530 * Wait for the write of extent to be finished to avoid
1531 * breaking kernel which assumes ->writepage should mark
1532 * PageWriteback or clean the page.
1534 result = cl_sync_file_range(inode, offset,
1535 offset + PAGE_SIZE - 1,
1538 /* May have written more than one page. decreasing this
1539 * page because the caller will count it.
1541 wbc->nr_to_write -= result - 1;
1546 cl_env_put(env, &refcheck);
1551 if (!lli->lli_async_rc)
1552 lli->lli_async_rc = result;
1553 SetPageError(vmpage);
1555 unlock_page(vmpage);
1560 int ll_writepages(struct address_space *mapping, struct writeback_control *wbc)
1562 struct inode *inode = mapping->host;
1565 enum cl_fsync_mode mode;
1566 int range_whole = 0;
1571 if (wbc->range_cyclic) {
1572 start = (loff_t)mapping->writeback_index << PAGE_SHIFT;
1573 end = OBD_OBJECT_EOF;
1575 start = wbc->range_start;
1576 end = wbc->range_end;
1577 if (end == LLONG_MAX) {
1578 end = OBD_OBJECT_EOF;
1579 range_whole = start == 0;
1583 mode = CL_FSYNC_NONE;
1584 if (wbc->sync_mode == WB_SYNC_ALL)
1585 mode = CL_FSYNC_LOCAL;
1587 if (wbc->sync_mode == WB_SYNC_NONE) {
1588 #ifdef SB_I_CGROUPWB
1589 struct bdi_writeback *wb;
1592 * As it may break full stripe writes on the inode,
1593 * disable periodic kupdate writeback (@wbc->for_kupdate)?
1597 * The system is under memory pressure and it is now reclaiming
1600 wb = inode_to_wb(inode);
1601 if (wbc->for_background ||
1602 (wb->start_all_reason == WB_REASON_VMSCAN &&
1603 test_bit(WB_start_all, &wb->state)))
1604 mode = CL_FSYNC_RECLAIM;
1607 * We have no idea about writeback reason for memory reclaim
1608 * WB_REASON_TRY_TO_FREE_PAGES in the old kernel such as rhel7
1609 * (WB_REASON_VMSCAN in the newer kernel) ...
1610 * Here set mode with CL_FSYNC_RECLAIM forcely on the old
1613 if (!wbc->for_kupdate)
1614 mode = CL_FSYNC_RECLAIM;
1618 if (ll_i2info(inode)->lli_clob == NULL || (inode->i_state & I_FREEING))
1621 /* for directio, it would call writepages() to evict cached pages
1622 * inside the IO context of write, which will cause deadlock at
1623 * layout_conf since it waits for active IOs to complete.
1625 result = cl_sync_file_range(inode, start, end, mode, 1);
1627 wbc->nr_to_write -= result;
1631 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) {
1632 if (end == OBD_OBJECT_EOF)
1633 mapping->writeback_index = 0;
1635 mapping->writeback_index = (end >> PAGE_SHIFT) + 1;
1640 struct ll_cl_context *ll_cl_find(struct inode *inode)
1642 struct ll_inode_info *lli = ll_i2info(inode);
1643 struct ll_cl_context *lcc;
1644 struct ll_cl_context *found = NULL;
1646 read_lock(&lli->lli_lock);
1647 list_for_each_entry(lcc, &lli->lli_lccs, lcc_list) {
1648 if (lcc->lcc_cookie == current) {
1653 read_unlock(&lli->lli_lock);
1658 void ll_cl_add(struct inode *inode, const struct lu_env *env, struct cl_io *io,
1661 struct ll_inode_info *lli = ll_i2info(inode);
1662 struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1664 memset(lcc, 0, sizeof(*lcc));
1665 INIT_LIST_HEAD(&lcc->lcc_list);
1666 lcc->lcc_cookie = current;
1669 lcc->lcc_type = type;
1671 write_lock(&lli->lli_lock);
1672 list_add(&lcc->lcc_list, &lli->lli_lccs);
1673 write_unlock(&lli->lli_lock);
1676 void ll_cl_remove(struct inode *inode, const struct lu_env *env)
1678 struct ll_inode_info *lli = ll_i2info(inode);
1679 struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1681 write_lock(&lli->lli_lock);
1682 list_del_init(&lcc->lcc_list);
1683 write_unlock(&lli->lli_lock);
1686 int ll_io_read_page(const struct lu_env *env, struct cl_io *io,
1687 struct cl_page *page, struct file *file)
1689 struct inode *inode = vvp_object_inode(page->cp_obj);
1690 struct ll_sb_info *sbi = ll_i2sbi(inode);
1691 struct ll_file_data *fd = NULL;
1692 struct ll_readahead_state *ras = NULL;
1693 struct cl_2queue *queue = &io->ci_queue;
1694 struct cl_sync_io *anchor = NULL;
1695 int rc = 0, rc2 = 0;
1697 struct vvp_io *vio = vvp_env_io(env);
1698 bool mmap = !vio->vui_ra_valid;
1699 pgoff_t ra_start_index = 0;
1700 pgoff_t io_start_index;
1701 pgoff_t io_end_index;
1702 bool unlockpage = true;
1707 fd = file->private_data;
1711 /* PagePrivate2 is set in ll_io_zero_page() to tell us the vmpage
1712 * must not be unlocked after processing.
1714 if (page->cp_vmpage && PagePrivate2(page->cp_vmpage))
1717 uptodate = page->cp_defer_uptodate;
1719 if (ll_readahead_enabled(sbi) && !page->cp_ra_updated && ras) {
1720 enum ras_update_flags flags = 0;
1723 flags |= LL_RAS_HIT;
1725 flags |= LL_RAS_MMAP;
1726 ras_update(sbi, inode, ras, cl_page_index(page), flags, io);
1729 cl_2queue_init(queue);
1731 page->cp_ra_used = 1;
1732 SetPageUptodate(page->cp_vmpage);
1733 cl_page_disown(env, io, page);
1735 anchor = &vvp_env_info(env)->vti_anchor;
1736 cl_sync_io_init(anchor, 1);
1737 page->cp_sync_io = anchor;
1739 cl_page_list_add(&queue->c2_qin, page, true);
1742 /* mmap does not set the ci_rw fields */
1744 io_start_index = io->u.ci_rw.crw_pos >> PAGE_SHIFT;
1745 io_end_index = (io->u.ci_rw.crw_pos +
1746 io->u.ci_rw.crw_bytes - 1) >> PAGE_SHIFT;
1748 io_start_index = cl_page_index(page);
1749 io_end_index = cl_page_index(page);
1752 if (ll_readahead_enabled(sbi) && ras && !io->ci_rand_read) {
1753 pgoff_t skip_index = 0;
1755 if (ras->ras_next_readahead_idx < cl_page_index(page))
1756 skip_index = cl_page_index(page);
1757 rc2 = ll_readahead(env, io, &queue->c2_qin, ras,
1758 uptodate, file, skip_index,
1760 /* Keep iotrace clean. Print only on actual page read */
1761 CDEBUG(D_READA | (rc2 ? D_IOTRACE : 0),
1762 DFID " %d pages read ahead at %lu, triggered by user read at %lu, stride offset %lld, stride length %lld, stride bytes %lld\n",
1763 PFID(ll_inode2fid(inode)), rc2, ra_start_index,
1764 cl_page_index(page), ras->ras_stride_offset,
1765 ras->ras_stride_length, ras->ras_stride_bytes);
1767 } else if (cl_page_index(page) == io_start_index &&
1768 io_end_index - io_start_index > 0) {
1769 rc2 = ll_readpages(env, io, &queue->c2_qin, io_start_index + 1,
1771 CDEBUG(D_READA, DFID " %d pages read at %lu\n",
1772 PFID(ll_inode2fid(inode)), rc2, cl_page_index(page));
1775 if (queue->c2_qin.pl_nr > 0) {
1776 int count = queue->c2_qin.pl_nr;
1778 rc = cl_io_submit_rw(env, io, CRT_READ, queue);
1780 task_io_account_read(PAGE_SIZE * count);
1784 if (anchor != NULL && !cl_page_is_owned(page, io)) { /* have sent */
1785 rc = cl_sync_io_wait(env, anchor, 0);
1787 cl_page_assume(env, io, page);
1788 cl_page_list_del(env, &queue->c2_qout, page, true);
1790 if (!PageUptodate(cl_page_vmpage(page))) {
1791 /* Failed to read a mirror, discard this page so that
1792 * new page can be created with new mirror.
1794 * TODO: this is not needed after page reinit
1795 * route is implemented
1797 cl_page_discard(env, io, page);
1800 cl_page_disown(env, io, page);
1803 /* TODO: discard all pages until page reinit route is implemented */
1804 cl_page_list_discard(env, io, &queue->c2_qin);
1806 /* Unlock unsent read pages in case of error. */
1807 cl_page_list_disown(env, &queue->c2_qin);
1809 cl_2queue_fini(env, queue);
1815 * Possible return value:
1816 * 0 no async readahead triggered and fast read could not be used.
1817 * 1 no async readahead, but fast read could be used.
1818 * 2 async readahead triggered and fast read could be used too.
1821 static int kickoff_async_readahead(struct file *file, unsigned long pages)
1823 struct ll_readahead_work *lrw;
1824 struct inode *inode = file_inode(file);
1825 struct ll_sb_info *sbi = ll_i2sbi(inode);
1826 struct ll_file_data *fd = file->private_data;
1827 struct ll_readahead_state *ras = &fd->fd_ras;
1828 struct ll_ra_info *ra = &sbi->ll_ra_info;
1829 unsigned long throttle;
1830 pgoff_t start_idx = ras_align(ras, ras->ras_next_readahead_idx);
1831 pgoff_t end_idx = start_idx + pages - 1;
1834 * In case we have a limited max_cached_mb, readahead
1835 * should be stopped if it have run out of all LRU slots.
1837 if (atomic_read(&ra->ra_cur_pages) >= sbi->ll_cache->ccc_lru_max) {
1838 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
1842 throttle = min(ra->ra_async_pages_per_file_threshold,
1843 ra->ra_max_pages_per_file);
1845 * If this is strided i/o or the window is smaller than the
1846 * throttle limit, we do not do async readahead. Otherwise,
1847 * we do async readahead, allowing the user thread to do fast i/o.
1849 if (stride_io_mode(ras) || !throttle ||
1850 ras->ras_window_pages < throttle ||
1851 atomic_read(&ra->ra_async_inflight) > ra->ra_async_max_active)
1854 if ((atomic_read(&ra->ra_cur_pages) + pages) > ra->ra_max_pages)
1857 if (ras->ras_async_last_readpage_idx == start_idx)
1860 /* ll_readahead_work_free() free it */
1863 atomic_inc(&sbi->ll_ra_info.ra_async_inflight);
1864 lrw->lrw_file = get_file(file);
1865 lrw->lrw_start_idx = start_idx;
1866 lrw->lrw_end_idx = end_idx;
1867 lrw->lrw_user_pid = current->pid;
1868 spin_lock(&ras->ras_lock);
1869 ras->ras_next_readahead_idx = end_idx + 1;
1870 ras->ras_async_last_readpage_idx = start_idx;
1871 spin_unlock(&ras->ras_lock);
1872 memcpy(lrw->lrw_jobid, ll_i2info(inode)->lli_jobid,
1873 sizeof(lrw->lrw_jobid));
1874 ll_readahead_work_add(inode, lrw);
1883 * Check if we can issue a readahead RPC, if that is
1884 * the case, we can't do fast IO because we will need
1885 * a cl_io to issue the RPC.
1887 static bool ll_use_fast_io(struct file *file,
1888 struct ll_readahead_state *ras, pgoff_t index)
1890 unsigned long fast_read_pages =
1891 max(RA_REMAIN_WINDOW_MIN, ras->ras_rpc_pages);
1893 loff_t stride_bytes = ras->ras_stride_bytes;
1897 if (stride_io_mode(ras) && stride_bytes) {
1898 skip_pages = (ras->ras_stride_length +
1899 ras->ras_stride_bytes - 1) / stride_bytes;
1900 skip_pages *= fast_read_pages;
1902 skip_pages = fast_read_pages;
1907 if (ras->ras_whole_file_read ||
1908 ras->ras_window_start_idx + ras->ras_window_pages <
1909 ras->ras_next_readahead_idx + skip_pages ||
1910 kickoff_async_readahead(file, fast_read_pages) > 0) {
1917 int ll_readpage(struct file *file, struct page *vmpage)
1919 struct inode *inode = file_inode(file);
1920 struct cl_object *clob = ll_i2info(inode)->lli_clob;
1921 struct ll_sb_info *sbi = ll_i2sbi(inode);
1922 const struct lu_env *env = NULL;
1923 struct cl_read_ahead ra = { 0 };
1924 struct ll_cl_context *lcc;
1925 struct cl_io *io = NULL;
1926 struct cl_page *page;
1933 if (CFS_FAIL_PRECHECK(OBD_FAIL_LLITE_READPAGE_PAUSE)) {
1934 unlock_page(vmpage);
1935 CFS_FAIL_TIMEOUT(OBD_FAIL_LLITE_READPAGE_PAUSE, cfs_fail_val);
1940 * The @vmpage got truncated.
1941 * This is a kernel bug introduced since kernel 5.12:
1942 * comment: cbd59c48ae2bcadc4a7599c29cf32fd3f9b78251
1943 * ("mm/filemap: use head pages in generic_file_buffered_read")
1945 * The page end offset calculation in filemap_get_read_batch() was off
1946 * by one. When a read is submitted with end offset 1048575, then it
1947 * calculates the end page for read of 256 where it should be 255. This
1948 * results in the readpage() for the page with index 256 is over stripe
1949 * boundary and may not covered by a DLM extent lock.
1951 * This happens in a corner race case: filemap_get_read_batch() adds
1952 * the page with index 256 for read which is not in the current read
1953 * I/O context, and this page is being invalidated and will be removed
1954 * from page cache due to the lock protected it being revoken. This
1955 * results in this page in the read path not covered by any DLM lock.
1957 * The solution is simple. Check whether the page was truncated in
1958 * ->readpage(). If so, just return AOP_TRUNCATED_PAGE to the upper
1959 * caller. Then the kernel will retry to batch pages, and it will not
1960 * add the truncated page into batches as it was removed from page
1961 * cache of the file.
1963 if (vmpage->mapping != inode->i_mapping) {
1964 unlock_page(vmpage);
1965 RETURN(AOP_TRUNCATED_PAGE);
1968 lcc = ll_cl_find(inode);
1974 if (io == NULL) { /* fast read */
1975 struct inode *inode = file_inode(file);
1976 struct ll_file_data *fd = file->private_data;
1977 struct ll_readahead_state *ras = &fd->fd_ras;
1978 struct lu_env *local_env = NULL;
1980 CDEBUG(D_VFSTRACE, "fast read pgno: %ld\n", vmpage->index);
1984 /* TODO: need to verify the layout version to make sure
1985 * the page is not invalid due to layout change.
1987 page = cl_vmpage_page(vmpage, clob);
1989 unlock_page(vmpage);
1990 CDEBUG(D_READA, "fast read: failed to find page %ld\n",
1992 ll_ra_stats_inc_sbi(sbi, RA_STAT_FAILED_FAST_READ);
1996 if (page->cp_defer_uptodate) {
1997 enum ras_update_flags flags = LL_RAS_HIT;
1999 if (lcc && lcc->lcc_type == LCC_MMAP)
2000 flags |= LL_RAS_MMAP;
2002 /* For fast read, it updates read ahead state only
2003 * if the page is hit in cache because non cache page
2004 * case will be handled by slow read later.
2006 ras_update(sbi, inode, ras, cl_page_index(page), flags, io);
2007 /* avoid duplicate ras_update() call */
2008 page->cp_ra_updated = 1;
2010 if (ll_use_fast_io(file, ras, cl_page_index(page)))
2015 local_env = cl_env_percpu_get();
2019 /* export the page and skip io stack */
2021 page->cp_ra_used = 1;
2022 SetPageUptodate(vmpage);
2024 ll_ra_stats_inc_sbi(sbi, RA_STAT_FAILED_FAST_READ);
2027 /* release page refcount before unlocking the page to ensure
2028 * the object won't be destroyed in the calling path of
2029 * cl_page_put(). Please see comment in ll_releasepage().
2031 cl_page_put(env, page);
2032 unlock_page(vmpage);
2034 cl_env_percpu_put(local_env);
2039 if (lcc && lcc->lcc_type != LCC_MMAP) {
2041 * This handles a kernel bug introduced in kernel 5.12:
2042 * comment: cbd59c48ae2bcadc4a7599c29cf32fd3f9b78251
2043 * ("mm/filemap: use head pages in generic_file_buffered_read")
2045 * See above in this function for a full description of the
2046 * bug. Briefly, the kernel will try to read 1 more page than
2047 * was actually requested *if that page is already in cache*.
2049 * Because this page is beyond the boundary of the requested
2050 * read, Lustre does not lock it as part of the read. This
2051 * means we must check if there is a valid dlmlock on this
2052 * this page and reference it before we attempt to read in the
2053 * page. If there is not a valid dlmlock, then we are racing
2054 * with dlmlock cancellation and the page is being removed
2057 * That means we should return AOP_TRUNCATED_PAGE, which will
2058 * cause the kernel to retry the read, which should allow the
2059 * page to be removed from cache as the lock is cancelled.
2061 * This should never occur except in kernels with the bug
2064 if (vmpage->index >= lcc->lcc_end_index) {
2066 "pgno:%ld, beyond read end_index:%ld\n",
2067 vmpage->index, lcc->lcc_end_index);
2069 result = cl_io_read_ahead(env, io, vmpage->index, &ra);
2070 if (result < 0 || vmpage->index > ra.cra_end_idx) {
2071 cl_read_ahead_release(env, &ra);
2072 unlock_page(vmpage);
2073 RETURN(AOP_TRUNCATED_PAGE);
2078 vio = vvp_env_io(env);
2080 * Direct read can fall back to buffered read, but DIO is done
2081 * with lockless i/o, and buffered requires LDLM locking, so in
2082 * this case we must restart without lockless.
2084 flags = iocb_ki_flags_get(file, vio->vui_iocb);
2085 if (iocb_ki_flags_check(flags, DIRECT) &&
2086 lcc && lcc->lcc_type == LCC_RW &&
2088 unlock_page(vmpage);
2089 io->ci_dio_lock = 1;
2090 io->ci_need_restart = 1;
2091 GOTO(out, result = -ENOLCK);
2094 LASSERT(io->ci_state == CIS_IO_GOING);
2095 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
2096 if (!IS_ERR(page)) {
2097 LASSERT(page->cp_type == CPT_CACHEABLE);
2098 if (likely(!PageUptodate(vmpage))) {
2099 cl_page_assume(env, io, page);
2101 result = ll_io_read_page(env, io, page, file);
2103 /* Page from a non-object file. */
2104 unlock_page(vmpage);
2107 cl_page_put(env, page);
2109 unlock_page(vmpage);
2110 result = PTR_ERR(page);
2114 if (ra.cra_release != NULL)
2115 cl_read_ahead_release(env, &ra);
2117 /* this delay gives time for the actual read of the page to finish and
2118 * unlock the page in vvp_page_completion_read before we return to our
2119 * caller and the caller tries to use the page, allowing us to test
2120 * races with the page being unlocked after readpage() but before it's
2121 * used by the caller
2123 CFS_FAIL_TIMEOUT(OBD_FAIL_LLITE_READPAGE_PAUSE2, cfs_fail_val);
2128 #ifdef HAVE_AOPS_READ_FOLIO
2129 int ll_read_folio(struct file *file, struct folio *folio)
2131 return ll_readpage(file, folio_page(folio, 0));