4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
33 * Lustre Lite I/O page cache routines shared by different kernel revs
36 #include <linux/kernel.h>
38 #include <linux/string.h>
39 #include <linux/stat.h>
40 #include <linux/errno.h>
41 #include <linux/unistd.h>
42 #include <linux/writeback.h>
43 #include <asm/uaccess.h>
46 #include <linux/file.h>
47 #include <linux/stat.h>
48 #include <asm/uaccess.h>
50 #include <linux/pagemap.h>
51 /* current_is_kswapd() */
52 #include <linux/swap.h>
53 #include <linux/task_io_accounting_ops.h>
55 #define DEBUG_SUBSYSTEM S_LLITE
57 #include <obd_cksum.h>
58 #include "llite_internal.h"
59 #include <lustre_compat.h>
61 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which);
64 * Get readahead pages from the filesystem readahead pool of the client for a
67 * /param sbi superblock for filesystem readahead state ll_ra_info
68 * /param ria per-thread readahead state
69 * /param pages number of pages requested for readahead for the thread.
71 * WARNING: This algorithm is used to reduce contention on sbi->ll_lock.
72 * It should work well if the ra_max_pages is much greater than the single
73 * file's read-ahead window, and not too many threads contending for
74 * these readahead pages.
76 * TODO: There may be a 'global sync problem' if many threads are trying
77 * to get an ra budget that is larger than the remaining readahead pages
78 * and reach here at exactly the same time. They will compute /a ret to
79 * consume the remaining pages, but will fail at atomic_add_return() and
80 * get a zero ra window, although there is still ra space remaining. - Jay */
82 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi,
83 struct ra_io_arg *ria,
85 unsigned long pages_min)
87 struct ll_ra_info *ra = &sbi->ll_ra_info;
93 * Don't try readahead agreesively if we are limited
94 * LRU pages, otherwise, it could cause deadlock.
96 pages = min(sbi->ll_cache->ccc_lru_max >> 2, pages);
98 * if this happen, we reserve more pages than needed,
99 * this will make us leak @ra_cur_pages, because
100 * ll_ra_count_put() acutally freed @pages.
102 if (WARN_ON_ONCE(pages_min > pages))
106 * If read-ahead pages left are less than 1M, do not do read-ahead,
107 * otherwise it will form small read RPC(< 1M), which hurt server
110 ret = min(ra->ra_max_pages - atomic_read(&ra->ra_cur_pages),
112 if (ret < 0 || ret < min_t(long, PTLRPC_MAX_BRW_PAGES, pages))
115 if (atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) {
116 atomic_sub(ret, &ra->ra_cur_pages);
121 if (ret < pages_min) {
122 /* override ra limit for maximum performance */
123 atomic_add(pages_min - ret, &ra->ra_cur_pages);
129 void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long pages)
131 struct ll_ra_info *ra = &sbi->ll_ra_info;
132 atomic_sub(pages, &ra->ra_cur_pages);
135 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which)
137 LASSERTF(which < _NR_RA_STAT, "which: %u\n", which);
138 lprocfs_counter_incr(sbi->ll_ra_stats, which);
141 static inline bool ll_readahead_enabled(struct ll_sb_info *sbi)
143 return sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
144 sbi->ll_ra_info.ra_max_pages > 0;
147 void ll_ra_stats_inc(struct inode *inode, enum ra_stat which)
149 struct ll_sb_info *sbi = ll_i2sbi(inode);
150 ll_ra_stats_inc_sbi(sbi, which);
153 #define RAS_CDEBUG(ras) \
155 "lre %llu cr %lu cb %llu wsi %lu wp %lu nra %lu rpc %lu " \
156 "r %lu csr %lu so %llu sb %llu sl %llu lr %lu\n", \
157 ras->ras_last_read_end_bytes, ras->ras_consecutive_requests, \
158 ras->ras_consecutive_bytes, ras->ras_window_start_idx, \
159 ras->ras_window_pages, ras->ras_next_readahead_idx, \
160 ras->ras_rpc_pages, ras->ras_requests, \
161 ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
162 ras->ras_stride_bytes, ras->ras_stride_length, \
163 ras->ras_async_last_readpage_idx)
165 static bool pos_in_window(loff_t pos, loff_t point,
166 unsigned long before, unsigned long after)
168 loff_t start = point - before;
169 loff_t end = point + after;
176 return start <= pos && pos <= end;
179 enum ll_ra_page_hint {
180 MAYNEED = 0, /* this page possibly accessed soon */
181 WILLNEED /* this page is gurateed to be needed */
185 * Initiates read-ahead of a page with given index.
187 * \retval +ve: page was already uptodate so it will be skipped
189 * \retval -ve: page wasn't added to \a queue for error;
190 * \retval 0: page was added into \a queue for read ahead.
192 static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io,
193 struct cl_page_list *queue, pgoff_t index,
194 enum ll_ra_page_hint hint)
196 struct cl_object *clob = io->ci_obj;
197 struct inode *inode = vvp_object_inode(clob);
198 struct page *vmpage = NULL;
199 struct cl_page *page;
200 struct vvp_page *vpg;
201 enum ra_stat which = _NR_RA_STAT; /* keep gcc happy */
203 const char *msg = NULL;
209 vmpage = grab_cache_page_nowait(inode->i_mapping, index);
210 if (vmpage == NULL) {
211 which = RA_STAT_FAILED_GRAB_PAGE;
212 msg = "g_c_p_n failed";
213 GOTO(out, rc = -EBUSY);
217 vmpage = find_or_create_page(inode->i_mapping, index,
220 GOTO(out, rc = -ENOMEM);
223 /* should not come here */
224 GOTO(out, rc = -EINVAL);
227 /* Check if vmpage was truncated or reclaimed */
228 if (vmpage->mapping != inode->i_mapping) {
229 which = RA_STAT_WRONG_GRAB_PAGE;
230 msg = "g_c_p_n returned invalid page";
231 GOTO(out, rc = -EBUSY);
234 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
236 which = RA_STAT_FAILED_GRAB_PAGE;
237 msg = "cl_page_find failed";
238 GOTO(out, rc = PTR_ERR(page));
241 lu_ref_add(&page->cp_reference, "ra", current);
242 cl_page_assume(env, io, page);
243 vpg = cl2vvp_page(cl_object_page_slice(clob, page));
244 if (!vpg->vpg_defer_uptodate && !PageUptodate(vmpage)) {
245 if (hint == MAYNEED) {
246 vpg->vpg_defer_uptodate = 1;
247 vpg->vpg_ra_used = 0;
249 cl_page_list_add(queue, page);
251 /* skip completed pages */
252 cl_page_unassume(env, io, page);
253 /* This page is already uptodate, returning a positive number
254 * to tell the callers about this */
258 lu_ref_del(&page->cp_reference, "ra", current);
259 cl_page_put(env, page);
262 if (vmpage != NULL) {
267 if (msg != NULL && hint == MAYNEED) {
268 ll_ra_stats_inc(inode, which);
269 CDEBUG(D_READA, "%s\n", msg);
276 #define RIA_DEBUG(ria) \
277 CDEBUG(D_READA, "rs %lu re %lu ro %llu rl %llu rb %llu\n", \
278 ria->ria_start_idx, ria->ria_end_idx, ria->ria_stoff, \
279 ria->ria_length, ria->ria_bytes)
281 static inline int stride_io_mode(struct ll_readahead_state *ras)
283 return ras->ras_consecutive_stride_requests > 1;
286 /* The function calculates how many bytes will be read in
287 * [off, off + length], in such stride IO area,
288 * stride_offset = st_off, stride_lengh = st_len,
289 * stride_bytes = st_bytes
291 * |------------------|*****|------------------|*****|------------|*****|....
294 * |----- st_len -----|
296 * How many bytes it should read in such pattern
297 * |-------------------------------------------------------------|
299 * |<------ length ------->|
301 * = |<----->| + |-------------------------------------| + |---|
302 * start_left st_bytes * i end_left
304 static loff_t stride_byte_count(loff_t st_off, loff_t st_len, loff_t st_bytes,
305 loff_t off, loff_t length)
307 u64 start = off > st_off ? off - st_off : 0;
308 u64 end = off + length > st_off ? off + length - st_off : 0;
313 if (st_len == 0 || length == 0 || end == 0)
316 start = div64_u64_rem(start, st_len, &start_left);
317 if (start_left < st_bytes)
318 start_left = st_bytes - start_left;
322 end = div64_u64_rem(end, st_len, &end_left);
323 if (end_left > st_bytes)
326 CDEBUG(D_READA, "start %llu, end %llu start_left %llu end_left %llu\n",
327 start, end, start_left, end_left);
330 bytes_count = end_left - (st_bytes - start_left);
332 bytes_count = start_left +
333 st_bytes * (end - start - 1) + end_left;
336 "st_off %llu, st_len %llu st_bytes %llu off %llu length %llu bytescount %llu\n",
337 st_off, st_len, st_bytes, off, length, bytes_count);
342 static unsigned long ria_page_count(struct ra_io_arg *ria)
344 loff_t length_bytes = ria->ria_end_idx >= ria->ria_start_idx ?
345 (loff_t)(ria->ria_end_idx -
346 ria->ria_start_idx + 1) << PAGE_SHIFT : 0;
349 if (ria->ria_length > ria->ria_bytes && ria->ria_bytes &&
350 (ria->ria_length & ~PAGE_MASK || ria->ria_bytes & ~PAGE_MASK ||
351 ria->ria_stoff & ~PAGE_MASK)) {
352 /* Over-estimate un-aligned page stride read */
353 unsigned long pg_count = ((ria->ria_bytes +
354 PAGE_SIZE - 1) >> PAGE_SHIFT) + 1;
355 pg_count *= length_bytes / ria->ria_length + 1;
359 bytes_count = stride_byte_count(ria->ria_stoff, ria->ria_length,
361 (loff_t)ria->ria_start_idx<<PAGE_SHIFT,
363 return (bytes_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
366 static pgoff_t ras_align(struct ll_readahead_state *ras, pgoff_t index)
368 unsigned opt_size = min(ras->ras_window_pages, ras->ras_rpc_pages);
372 return index - (index % opt_size);
375 /* Check whether the index is in the defined ra-window */
376 static bool ras_inside_ra_window(pgoff_t idx, struct ra_io_arg *ria)
378 loff_t pos = (loff_t)idx << PAGE_SHIFT;
380 /* If ria_length == ria_bytes, it means non-stride I/O mode,
381 * idx should always inside read-ahead window in this case
382 * For stride I/O mode, just check whether the idx is inside
385 if (ria->ria_length == 0 || ria->ria_length == ria->ria_bytes)
388 if (pos >= ria->ria_stoff) {
391 div64_u64_rem(pos - ria->ria_stoff, ria->ria_length, &offset);
393 if (offset < ria->ria_bytes ||
394 (ria->ria_length - offset) < PAGE_SIZE)
396 } else if (pos + PAGE_SIZE > ria->ria_stoff) {
404 ll_read_ahead_pages(const struct lu_env *env, struct cl_io *io,
405 struct cl_page_list *queue, struct ll_readahead_state *ras,
406 struct ra_io_arg *ria, pgoff_t *ra_end, pgoff_t skip_index)
408 struct cl_read_ahead ra = { 0 };
409 /* busy page count is per stride */
410 int rc = 0, count = 0, busy_page_count = 0;
413 LASSERT(ria != NULL);
416 for (page_idx = ria->ria_start_idx;
417 page_idx <= ria->ria_end_idx && ria->ria_reserved > 0;
419 if (skip_index && page_idx == skip_index)
421 if (ras_inside_ra_window(page_idx, ria)) {
422 if (ra.cra_end_idx == 0 || ra.cra_end_idx < page_idx) {
426 * Do not shrink ria_end_idx at any case until
427 * the minimum end of current read is covered.
429 * Do not extend read lock accross stripe if
430 * lock contention detected.
432 if (ra.cra_contention &&
433 page_idx > ria->ria_end_idx_min) {
434 ria->ria_end_idx = *ra_end;
438 cl_read_ahead_release(env, &ra);
440 rc = cl_io_read_ahead(env, io, page_idx, &ra);
445 * Only shrink ria_end_idx if the matched
446 * LDLM lock doesn't cover more.
448 if (page_idx > ra.cra_end_idx) {
449 ria->ria_end_idx = ra.cra_end_idx;
453 CDEBUG(D_READA, "idx: %lu, ra: %lu, rpc: %lu\n",
454 page_idx, ra.cra_end_idx,
456 LASSERTF(ra.cra_end_idx >= page_idx,
457 "object: %p, indcies %lu / %lu\n",
458 io->ci_obj, ra.cra_end_idx, page_idx);
459 /* update read ahead RPC size.
460 * NB: it's racy but doesn't matter */
461 if (ras->ras_rpc_pages != ra.cra_rpc_pages &&
462 ra.cra_rpc_pages > 0)
463 ras->ras_rpc_pages = ra.cra_rpc_pages;
465 /* trim it to align with optimal RPC size */
466 end_idx = ras_align(ras, ria->ria_end_idx + 1);
467 if (end_idx > 0 && !ria->ria_eof)
468 ria->ria_end_idx = end_idx - 1;
470 if (ria->ria_end_idx < ria->ria_end_idx_min)
471 ria->ria_end_idx = ria->ria_end_idx_min;
473 if (page_idx > ria->ria_end_idx)
476 /* If the page is inside the read-ahead window */
477 rc = ll_read_ahead_page(env, io, queue, page_idx,
479 if (rc < 0 && rc != -EBUSY)
484 "skip busy page: %lu\n", page_idx);
485 /* For page unaligned readahead the first
486 * last pages of each region can be read by
487 * another reader on the same node, and so
488 * may be busy. So only stop for > 2 busy
490 if (busy_page_count > 2)
495 /* Only subtract from reserve & count the page if we
496 * really did readahead on that page. */
501 } else if (stride_io_mode(ras)) {
502 /* If it is not in the read-ahead window, and it is
503 * read-ahead mode, then check whether it should skip
506 loff_t pos = (loff_t)page_idx << PAGE_SHIFT;
509 div64_u64_rem(pos - ria->ria_stoff, ria->ria_length,
511 if (offset >= ria->ria_bytes) {
512 pos += (ria->ria_length - offset);
513 if ((pos >> PAGE_SHIFT) >= page_idx + 1)
514 page_idx = (pos >> PAGE_SHIFT) - 1;
517 "Stride: jump %llu pages to %lu\n",
518 ria->ria_length - offset, page_idx);
524 cl_read_ahead_release(env, &ra);
529 static void ll_readahead_work_free(struct ll_readahead_work *work)
531 fput(work->lrw_file);
535 static void ll_readahead_handle_work(struct work_struct *wq);
536 static void ll_readahead_work_add(struct inode *inode,
537 struct ll_readahead_work *work)
539 INIT_WORK(&work->lrw_readahead_work, ll_readahead_handle_work);
540 queue_work(ll_i2sbi(inode)->ll_ra_info.ll_readahead_wq,
541 &work->lrw_readahead_work);
544 static int ll_readahead_file_kms(const struct lu_env *env,
545 struct cl_io *io, __u64 *kms)
547 struct cl_object *clob;
549 struct cl_attr *attr = vvp_env_thread_attr(env);
553 inode = vvp_object_inode(clob);
555 cl_object_attr_lock(clob);
556 ret = cl_object_attr_get(env, clob, attr);
557 cl_object_attr_unlock(clob);
562 *kms = attr->cat_kms;
566 static void ll_readahead_handle_work(struct work_struct *wq)
568 struct ll_readahead_work *work;
571 struct ra_io_arg *ria;
573 struct ll_file_data *fd;
574 struct ll_readahead_state *ras;
576 struct cl_2queue *queue;
577 pgoff_t ra_end_idx = 0;
578 unsigned long pages, pages_min = 0;
583 struct ll_sb_info *sbi;
585 work = container_of(wq, struct ll_readahead_work,
587 fd = work->lrw_file->private_data;
589 file = work->lrw_file;
590 inode = file_inode(file);
591 sbi = ll_i2sbi(inode);
593 env = cl_env_alloc(&refcheck, LCT_NOREF);
595 GOTO(out_free_work, rc = PTR_ERR(env));
597 io = vvp_env_thread_io(env);
598 ll_io_init(io, file, CIT_READ, NULL);
600 rc = ll_readahead_file_kms(env, io, &kms);
602 GOTO(out_put_env, rc);
605 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
606 GOTO(out_put_env, rc = 0);
609 ria = &ll_env_info(env)->lti_ria;
610 memset(ria, 0, sizeof(*ria));
612 ria->ria_start_idx = work->lrw_start_idx;
613 /* Truncate RA window to end of file */
614 eof_index = (pgoff_t)(kms - 1) >> PAGE_SHIFT;
615 if (eof_index <= work->lrw_end_idx) {
616 work->lrw_end_idx = eof_index;
619 if (work->lrw_end_idx <= work->lrw_start_idx)
620 GOTO(out_put_env, rc = 0);
622 ria->ria_end_idx = work->lrw_end_idx;
623 pages = ria->ria_end_idx - ria->ria_start_idx + 1;
624 ria->ria_reserved = ll_ra_count_get(sbi, ria,
625 ria_page_count(ria), pages_min);
628 "async reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
629 ria->ria_reserved, pages, pages_min,
630 atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
631 ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
633 if (ria->ria_reserved < pages) {
634 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
635 if (PAGES_TO_MiB(ria->ria_reserved) < 1) {
636 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
637 GOTO(out_put_env, rc = 0);
641 rc = cl_io_rw_init(env, io, CIT_READ, ria->ria_start_idx, pages);
643 GOTO(out_put_env, rc);
645 /* overwrite jobid inited in vvp_io_init() */
646 if (strncmp(ll_i2info(inode)->lli_jobid, work->lrw_jobid,
647 sizeof(work->lrw_jobid)))
648 memcpy(ll_i2info(inode)->lli_jobid, work->lrw_jobid,
649 sizeof(work->lrw_jobid));
651 vvp_env_io(env)->vui_fd = fd;
652 io->ci_state = CIS_LOCKED;
653 io->ci_async_readahead = true;
654 rc = cl_io_start(env, io);
656 GOTO(out_io_fini, rc);
658 queue = &io->ci_queue;
659 cl_2queue_init(queue);
661 rc = ll_read_ahead_pages(env, io, &queue->c2_qin, ras, ria,
663 if (ria->ria_reserved != 0)
664 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
665 if (queue->c2_qin.pl_nr > 0) {
666 int count = queue->c2_qin.pl_nr;
668 rc = cl_io_submit_rw(env, io, CRT_READ, queue);
670 task_io_account_read(PAGE_SIZE * count);
672 if (ria->ria_end_idx == ra_end_idx && ra_end_idx == (kms >> PAGE_SHIFT))
673 ll_ra_stats_inc(inode, RA_STAT_EOF);
675 if (ra_end_idx != ria->ria_end_idx)
676 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
678 /* TODO: discard all pages until page reinit route is implemented */
679 cl_page_list_discard(env, io, &queue->c2_qin);
681 /* Unlock unsent read pages in case of error. */
682 cl_page_list_disown(env, io, &queue->c2_qin);
684 cl_2queue_fini(env, queue);
689 cl_env_put(env, &refcheck);
692 ll_ra_stats_inc_sbi(ll_i2sbi(inode), RA_STAT_ASYNC);
693 atomic_dec(&sbi->ll_ra_info.ra_async_inflight);
694 ll_readahead_work_free(work);
697 static int ll_readahead(const struct lu_env *env, struct cl_io *io,
698 struct cl_page_list *queue,
699 struct ll_readahead_state *ras, bool hit,
700 struct file *file, pgoff_t skip_index)
702 struct vvp_io *vio = vvp_env_io(env);
703 struct ll_thread_info *lti = ll_env_info(env);
704 unsigned long pages, pages_min = 0;
705 pgoff_t ra_end_idx = 0, start_idx = 0, end_idx = 0;
707 struct ra_io_arg *ria = <i->lti_ria;
708 struct cl_object *clob;
711 struct ll_sb_info *sbi;
712 struct ll_ra_info *ra;
719 inode = vvp_object_inode(clob);
720 sbi = ll_i2sbi(inode);
721 ra = &sbi->ll_ra_info;
724 * In case we have a limited max_cached_mb, readahead
725 * should be stopped if it have run out of all LRU slots.
727 if (atomic_read(&ra->ra_cur_pages) >= sbi->ll_cache->ccc_lru_max) {
728 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
732 memset(ria, 0, sizeof(*ria));
733 ret = ll_readahead_file_kms(env, io, &kms);
738 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
742 spin_lock(&ras->ras_lock);
745 * Note: other thread might rollback the ras_next_readahead_idx,
746 * if it can not get the full size of prepared pages, see the
747 * end of this function. For stride read ahead, it needs to
748 * make sure the offset is no less than ras_stride_offset,
749 * so that stride read ahead can work correctly.
751 if (stride_io_mode(ras))
752 start_idx = max_t(pgoff_t, ras->ras_next_readahead_idx,
753 ras->ras_stride_offset >> PAGE_SHIFT);
755 start_idx = ras->ras_next_readahead_idx;
757 if (ras->ras_window_pages > 0)
758 end_idx = ras->ras_window_start_idx + ras->ras_window_pages - 1;
761 end_idx = start_idx + ras->ras_window_pages - 1;
763 /* Enlarge the RA window to encompass the full read */
764 if (vio->vui_ra_valid &&
765 end_idx < vio->vui_ra_start_idx + vio->vui_ra_pages - 1)
766 end_idx = vio->vui_ra_start_idx + vio->vui_ra_pages - 1;
771 /* Truncate RA window to end of file */
772 eof_index = (pgoff_t)((kms - 1) >> PAGE_SHIFT);
773 if (eof_index <= end_idx) {
778 ria->ria_start_idx = start_idx;
779 ria->ria_end_idx = end_idx;
780 /* If stride I/O mode is detected, get stride window*/
781 if (stride_io_mode(ras)) {
782 ria->ria_stoff = ras->ras_stride_offset;
783 ria->ria_length = ras->ras_stride_length;
784 ria->ria_bytes = ras->ras_stride_bytes;
786 spin_unlock(&ras->ras_lock);
789 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
792 pages = ria_page_count(ria);
794 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
799 CDEBUG(D_READA, DFID": ria: %lu/%lu, bead: %lu/%lu, hit: %d\n",
800 PFID(lu_object_fid(&clob->co_lu)),
801 ria->ria_start_idx, ria->ria_end_idx,
802 vio->vui_ra_valid ? vio->vui_ra_start_idx : 0,
803 vio->vui_ra_valid ? vio->vui_ra_pages : 0,
806 /* at least to extend the readahead window to cover current read */
807 if (!hit && vio->vui_ra_valid &&
808 vio->vui_ra_start_idx + vio->vui_ra_pages > ria->ria_start_idx) {
809 ria->ria_end_idx_min =
810 vio->vui_ra_start_idx + vio->vui_ra_pages - 1;
811 pages_min = vio->vui_ra_start_idx + vio->vui_ra_pages -
814 * For performance reason, exceeding @ra_max_pages
815 * are allowed, but this should be limited with RPC
816 * size in case a large block size read issued. Trim
819 pages_min = min(pages_min, ras->ras_rpc_pages -
820 (ria->ria_start_idx % ras->ras_rpc_pages));
823 /* don't over reserved for mmap range read */
827 ria->ria_reserved = ll_ra_count_get(ll_i2sbi(inode), ria, pages,
829 if (ria->ria_reserved < pages)
830 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
832 CDEBUG(D_READA, "reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
833 ria->ria_reserved, pages, pages_min,
834 atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
835 ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
837 ret = ll_read_ahead_pages(env, io, queue, ras, ria, &ra_end_idx,
839 if (ria->ria_reserved != 0)
840 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
842 if (ra_end_idx == end_idx && ra_end_idx == (kms >> PAGE_SHIFT))
843 ll_ra_stats_inc(inode, RA_STAT_EOF);
846 "ra_end_idx = %lu end_idx = %lu stride end = %lu pages = %d\n",
847 ra_end_idx, end_idx, ria->ria_end_idx, ret);
849 if (ra_end_idx != end_idx)
850 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
851 if (ra_end_idx > 0) {
852 /* update the ras so that the next read-ahead tries from
853 * where we left off. */
854 spin_lock(&ras->ras_lock);
855 ras->ras_next_readahead_idx = ra_end_idx + 1;
856 spin_unlock(&ras->ras_lock);
863 static int ll_readpages(const struct lu_env *env, struct cl_io *io,
864 struct cl_page_list *queue,
865 pgoff_t start, pgoff_t end)
874 ret = ll_readahead_file_kms(env, io, &kms);
882 unsigned long end_index;
884 end_index = (unsigned long)((kms - 1) >> PAGE_SHIFT);
885 if (end_index <= end)
889 for (page_idx = start; page_idx <= end; page_idx++) {
890 ret= ll_read_ahead_page(env, io, queue, page_idx,
894 else if (ret == 0) /* ret 1 is already uptodate */
898 RETURN(count > 0 ? count : ret);
901 static void ras_set_start(struct ll_readahead_state *ras, pgoff_t index)
903 ras->ras_window_start_idx = ras_align(ras, index);
906 /* called with the ras_lock held or from places where it doesn't matter */
907 static void ras_reset(struct ll_readahead_state *ras, pgoff_t index)
909 ras->ras_consecutive_requests = 0;
910 ras->ras_consecutive_bytes = 0;
911 ras->ras_window_pages = 0;
912 ras_set_start(ras, index);
913 ras->ras_next_readahead_idx = max(ras->ras_window_start_idx, index + 1);
918 /* called with the ras_lock held or from places where it doesn't matter */
919 static void ras_stride_reset(struct ll_readahead_state *ras)
921 ras->ras_consecutive_stride_requests = 0;
922 ras->ras_stride_length = 0;
923 ras->ras_stride_bytes = 0;
927 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
929 spin_lock_init(&ras->ras_lock);
930 ras->ras_rpc_pages = PTLRPC_MAX_BRW_PAGES;
932 ras->ras_last_read_end_bytes = 0;
933 ras->ras_requests = 0;
934 ras->ras_range_min_start_idx = 0;
935 ras->ras_range_max_end_idx = 0;
936 ras->ras_range_requests = 0;
937 ras->ras_last_range_pages = 0;
941 * Check whether the read request is in the stride window.
942 * If it is in the stride window, return true, otherwise return false.
944 static bool read_in_stride_window(struct ll_readahead_state *ras,
945 loff_t pos, loff_t count)
949 if (ras->ras_stride_length == 0 || ras->ras_stride_bytes == 0 ||
950 ras->ras_stride_bytes == ras->ras_stride_length)
953 stride_gap = pos - ras->ras_last_read_end_bytes - 1;
955 /* If it is contiguous read */
957 return ras->ras_consecutive_bytes + count <=
958 ras->ras_stride_bytes;
960 /* Otherwise check the stride by itself */
961 return (ras->ras_stride_length - ras->ras_stride_bytes) == stride_gap &&
962 ras->ras_consecutive_bytes == ras->ras_stride_bytes &&
963 count <= ras->ras_stride_bytes;
966 static void ras_init_stride_detector(struct ll_readahead_state *ras,
967 loff_t pos, loff_t count)
969 loff_t stride_gap = pos - ras->ras_last_read_end_bytes - 1;
971 LASSERT(ras->ras_consecutive_stride_requests == 0);
973 if (pos <= ras->ras_last_read_end_bytes) {
974 /*Reset stride window for forward read*/
975 ras_stride_reset(ras);
979 ras->ras_stride_bytes = ras->ras_consecutive_bytes;
980 ras->ras_stride_length = stride_gap + ras->ras_consecutive_bytes;
981 ras->ras_consecutive_stride_requests++;
982 ras->ras_stride_offset = pos;
988 stride_page_count(struct ll_readahead_state *ras, loff_t len)
991 stride_byte_count(ras->ras_stride_offset,
992 ras->ras_stride_length, ras->ras_stride_bytes,
993 ras->ras_window_start_idx << PAGE_SHIFT, len);
995 return (bytes_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
998 /* Stride Read-ahead window will be increased inc_len according to
999 * stride I/O pattern */
1000 static void ras_stride_increase_window(struct ll_readahead_state *ras,
1001 struct ll_ra_info *ra, loff_t inc_bytes)
1003 loff_t window_bytes, stride_bytes;
1008 /* temporarily store in page units to reduce LASSERT() cost below */
1009 end = ras->ras_window_start_idx + ras->ras_window_pages;
1011 LASSERT(ras->ras_stride_length > 0);
1012 LASSERTF(end >= (ras->ras_stride_offset >> PAGE_SHIFT),
1013 "window_start_idx %lu, window_pages %lu stride_offset %llu\n",
1014 ras->ras_window_start_idx, ras->ras_window_pages,
1015 ras->ras_stride_offset);
1018 if (end <= ras->ras_stride_offset)
1021 stride_bytes = end - ras->ras_stride_offset;
1023 div64_u64_rem(stride_bytes, ras->ras_stride_length, &left_bytes);
1024 window_bytes = (ras->ras_window_pages << PAGE_SHIFT);
1025 if (left_bytes < ras->ras_stride_bytes) {
1026 if (ras->ras_stride_bytes - left_bytes >= inc_bytes) {
1027 window_bytes += inc_bytes;
1030 window_bytes += (ras->ras_stride_bytes - left_bytes);
1031 inc_bytes -= (ras->ras_stride_bytes - left_bytes);
1034 window_bytes += (ras->ras_stride_length - left_bytes);
1037 LASSERT(ras->ras_stride_bytes != 0);
1039 step = div64_u64_rem(inc_bytes, ras->ras_stride_bytes, &left_bytes);
1041 window_bytes += step * ras->ras_stride_length + left_bytes;
1042 LASSERT(window_bytes > 0);
1045 if (stride_page_count(ras, window_bytes) <=
1046 ra->ra_max_pages_per_file || ras->ras_window_pages == 0)
1047 ras->ras_window_pages = (window_bytes >> PAGE_SHIFT);
1049 LASSERT(ras->ras_window_pages > 0);
1054 static void ras_increase_window(struct inode *inode,
1055 struct ll_readahead_state *ras,
1056 struct ll_ra_info *ra)
1058 /* The stretch of ra-window should be aligned with max rpc_size
1059 * but current clio architecture does not support retrieve such
1060 * information from lower layer. FIXME later
1062 if (stride_io_mode(ras)) {
1063 ras_stride_increase_window(ras, ra,
1064 (loff_t)ras->ras_rpc_pages << PAGE_SHIFT);
1066 pgoff_t window_pages;
1068 window_pages = min(ras->ras_window_pages + ras->ras_rpc_pages,
1069 ra->ra_max_pages_per_file);
1070 if (window_pages < ras->ras_rpc_pages)
1071 ras->ras_window_pages = window_pages;
1073 ras->ras_window_pages = ras_align(ras, window_pages);
1078 * Seek within 8 pages are considered as sequential read for now.
1080 static inline bool is_loose_seq_read(struct ll_readahead_state *ras, loff_t pos)
1082 return pos_in_window(pos, ras->ras_last_read_end_bytes,
1083 8UL << PAGE_SHIFT, 8UL << PAGE_SHIFT);
1086 static inline bool is_loose_mmap_read(struct ll_sb_info *sbi,
1087 struct ll_readahead_state *ras,
1090 unsigned long range_pages = sbi->ll_ra_info.ra_range_pages;
1092 return pos_in_window(pos, ras->ras_last_read_end_bytes,
1093 range_pages << PAGE_SHIFT,
1094 range_pages << PAGE_SHIFT);
1098 * We have observed slow mmap read performances for some
1099 * applications. The problem is if access pattern is neither
1100 * sequential nor stride, but could be still adjacent in a
1101 * small range and then seek a random position.
1103 * So the pattern could be something like this:
1105 * [1M data] [hole] [0.5M data] [hole] [0.7M data] [1M data]
1108 * Every time an application reads mmap data, it may not only
1109 * read a single 4KB page, but aslo a cluster of nearby pages in
1110 * a range(e.g. 1MB) of the first page after a cache miss.
1112 * The readahead engine is modified to track the range size of
1113 * a cluster of mmap reads, so that after a seek and/or cache miss,
1114 * the range size is used to efficiently prefetch multiple pages
1115 * in a single RPC rather than many small RPCs.
1117 static void ras_detect_cluster_range(struct ll_readahead_state *ras,
1118 struct ll_sb_info *sbi,
1119 unsigned long pos, unsigned long count)
1121 pgoff_t last_pages, pages;
1122 pgoff_t end_idx = (pos + count - 1) >> PAGE_SHIFT;
1124 last_pages = ras->ras_range_max_end_idx -
1125 ras->ras_range_min_start_idx + 1;
1126 /* First time come here */
1127 if (!ras->ras_range_max_end_idx)
1130 /* Random or Stride read */
1131 if (!is_loose_mmap_read(sbi, ras, pos))
1134 ras->ras_range_requests++;
1135 if (ras->ras_range_max_end_idx < end_idx)
1136 ras->ras_range_max_end_idx = end_idx;
1138 if (ras->ras_range_min_start_idx > (pos >> PAGE_SHIFT))
1139 ras->ras_range_min_start_idx = pos >> PAGE_SHIFT;
1141 /* Out of range, consider it as random or stride */
1142 pages = ras->ras_range_max_end_idx -
1143 ras->ras_range_min_start_idx + 1;
1144 if (pages <= sbi->ll_ra_info.ra_range_pages)
1147 ras->ras_last_range_pages = last_pages;
1148 ras->ras_range_requests = 0;
1149 ras->ras_range_min_start_idx = pos >> PAGE_SHIFT;
1150 ras->ras_range_max_end_idx = end_idx;
1153 static void ras_detect_read_pattern(struct ll_readahead_state *ras,
1154 struct ll_sb_info *sbi,
1155 loff_t pos, size_t count, bool mmap)
1157 bool stride_detect = false;
1158 pgoff_t index = pos >> PAGE_SHIFT;
1161 * Reset the read-ahead window in two cases. First when the app seeks
1162 * or reads to some other part of the file. Secondly if we get a
1163 * read-ahead miss that we think we've previously issued. This can
1164 * be a symptom of there being so many read-ahead pages that the VM
1165 * is reclaiming it before we get to it.
1167 if (!is_loose_seq_read(ras, pos)) {
1168 /* Check whether it is in stride I/O mode */
1169 if (!read_in_stride_window(ras, pos, count)) {
1170 if (ras->ras_consecutive_stride_requests == 0)
1171 ras_init_stride_detector(ras, pos, count);
1173 ras_stride_reset(ras);
1174 ras->ras_consecutive_bytes = 0;
1175 ras_reset(ras, index);
1177 ras->ras_consecutive_bytes = 0;
1178 ras->ras_consecutive_requests = 0;
1179 if (++ras->ras_consecutive_stride_requests > 1)
1180 stride_detect = true;
1183 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
1184 } else if (stride_io_mode(ras)) {
1186 * If this is contiguous read but in stride I/O mode
1187 * currently, check whether stride step still is valid,
1188 * if invalid, it will reset the stride ra window to
1191 if (!read_in_stride_window(ras, pos, count)) {
1192 ras_stride_reset(ras);
1193 ras->ras_window_pages = 0;
1194 ras->ras_next_readahead_idx = index;
1198 ras->ras_consecutive_bytes += count;
1200 pgoff_t idx = ras->ras_consecutive_bytes >> PAGE_SHIFT;
1201 unsigned long ra_range_pages =
1202 max_t(unsigned long, RA_MIN_MMAP_RANGE_PAGES,
1203 sbi->ll_ra_info.ra_range_pages);
1205 if ((idx >= ra_range_pages &&
1206 idx % ra_range_pages == 0) || stride_detect)
1207 ras->ras_need_increase_window = true;
1208 } else if ((ras->ras_consecutive_requests > 1 || stride_detect)) {
1209 ras->ras_need_increase_window = true;
1212 ras->ras_last_read_end_bytes = pos + count - 1;
1215 void ll_ras_enter(struct file *f, loff_t pos, size_t count)
1217 struct ll_file_data *fd = f->private_data;
1218 struct ll_readahead_state *ras = &fd->fd_ras;
1219 struct inode *inode = file_inode(f);
1220 unsigned long index = pos >> PAGE_SHIFT;
1221 struct ll_sb_info *sbi = ll_i2sbi(inode);
1223 spin_lock(&ras->ras_lock);
1224 ras->ras_requests++;
1225 ras->ras_consecutive_requests++;
1226 ras->ras_need_increase_window = false;
1227 ras->ras_no_miss_check = false;
1229 * On the second access to a file smaller than the tunable
1230 * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1231 * file up to ra_max_pages_per_file. This is simply a best effort
1232 * and only occurs once per open file. Normal RA behavior is reverted
1233 * to for subsequent IO.
1235 if (ras->ras_requests >= 2) {
1237 struct ll_ra_info *ra = &sbi->ll_ra_info;
1239 kms_pages = (i_size_read(inode) + PAGE_SIZE - 1) >>
1242 CDEBUG(D_READA, "kmsp %llu mwp %lu mp %lu\n", kms_pages,
1243 ra->ra_max_read_ahead_whole_pages,
1244 ra->ra_max_pages_per_file);
1247 kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1248 ras->ras_window_start_idx = 0;
1249 ras->ras_next_readahead_idx = index + 1;
1250 ras->ras_window_pages = min(ra->ra_max_pages_per_file,
1251 ra->ra_max_read_ahead_whole_pages);
1252 ras->ras_no_miss_check = true;
1253 GOTO(out_unlock, 0);
1256 ras_detect_read_pattern(ras, sbi, pos, count, false);
1258 spin_unlock(&ras->ras_lock);
1261 static bool index_in_stride_window(struct ll_readahead_state *ras,
1264 loff_t pos = (loff_t)index << PAGE_SHIFT;
1266 if (ras->ras_stride_length == 0 || ras->ras_stride_bytes == 0 ||
1267 ras->ras_stride_bytes == ras->ras_stride_length)
1270 if (pos >= ras->ras_stride_offset) {
1273 div64_u64_rem(pos - ras->ras_stride_offset,
1274 ras->ras_stride_length, &offset);
1275 if (offset < ras->ras_stride_bytes ||
1276 ras->ras_stride_length - offset < PAGE_SIZE)
1278 } else if (ras->ras_stride_offset - pos < PAGE_SIZE) {
1286 * ll_ras_enter() is used to detect read pattern according to pos and count.
1288 * ras_update() is used to detect cache miss and
1289 * reset window or increase window accordingly
1291 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
1292 struct ll_readahead_state *ras, pgoff_t index,
1293 enum ras_update_flags flags, struct cl_io *io)
1295 struct ll_ra_info *ra = &sbi->ll_ra_info;
1296 bool hit = flags & LL_RAS_HIT;
1299 spin_lock(&ras->ras_lock);
1302 CDEBUG(D_READA, DFID " pages at %lu miss.\n",
1303 PFID(ll_inode2fid(inode)), index);
1304 ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
1307 * The readahead window has been expanded to cover whole
1308 * file size, we don't care whether ra miss happen or not.
1309 * Because we will read whole file to page cache even if
1310 * some pages missed.
1312 if (ras->ras_no_miss_check)
1313 GOTO(out_unlock, 0);
1315 if (io && io->ci_rand_read)
1316 GOTO(out_unlock, 0);
1318 if (io && io->ci_seq_read) {
1320 /* to avoid many small read RPC here */
1321 ras->ras_window_pages = sbi->ll_ra_info.ra_range_pages;
1322 ll_ra_stats_inc_sbi(sbi, RA_STAT_MMAP_RANGE_READ);
1327 if (flags & LL_RAS_MMAP) {
1328 unsigned long ra_pages;
1330 ras_detect_cluster_range(ras, sbi, index << PAGE_SHIFT,
1332 ras_detect_read_pattern(ras, sbi, (loff_t)index << PAGE_SHIFT,
1335 /* we did not detect anything but we could prefetch */
1336 if (!ras->ras_need_increase_window &&
1337 ras->ras_window_pages <= sbi->ll_ra_info.ra_range_pages &&
1338 ras->ras_range_requests >= 2) {
1340 ra_pages = max_t(unsigned long,
1341 RA_MIN_MMAP_RANGE_PAGES,
1342 ras->ras_last_range_pages);
1343 if (index < ra_pages / 2)
1346 index -= ra_pages / 2;
1347 ras->ras_window_pages = ra_pages;
1348 ll_ra_stats_inc_sbi(sbi,
1349 RA_STAT_MMAP_RANGE_READ);
1351 ras->ras_window_pages = 0;
1357 if (!hit && ras->ras_window_pages &&
1358 index < ras->ras_next_readahead_idx &&
1359 pos_in_window(index, ras->ras_window_start_idx, 0,
1360 ras->ras_window_pages)) {
1361 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
1362 ras->ras_need_increase_window = false;
1364 if (index_in_stride_window(ras, index) &&
1365 stride_io_mode(ras)) {
1367 * if (index != ras->ras_last_readpage + 1)
1368 * ras->ras_consecutive_pages = 0;
1370 ras_reset(ras, index);
1373 * If stride-RA hit cache miss, the stride
1374 * detector will not be reset to avoid the
1375 * overhead of redetecting read-ahead mode,
1376 * but on the condition that the stride window
1377 * is still intersect with normal sequential
1378 * read-ahead window.
1380 if (ras->ras_window_start_idx < ras->ras_stride_offset)
1381 ras_stride_reset(ras);
1385 * Reset both stride window and normal RA
1388 ras_reset(ras, index);
1389 /* ras->ras_consecutive_pages++; */
1390 ras->ras_consecutive_bytes = 0;
1391 ras_stride_reset(ras);
1392 GOTO(out_unlock, 0);
1397 ras_set_start(ras, index);
1399 if (stride_io_mode(ras)) {
1400 /* Since stride readahead is sentivite to the offset
1401 * of read-ahead, so we use original offset here,
1402 * instead of ras_window_start_idx, which is RPC aligned.
1404 ras->ras_next_readahead_idx = max(index + 1,
1405 ras->ras_next_readahead_idx);
1406 ras->ras_window_start_idx =
1407 max_t(pgoff_t, ras->ras_window_start_idx,
1408 ras->ras_stride_offset >> PAGE_SHIFT);
1410 if (ras->ras_next_readahead_idx < ras->ras_window_start_idx)
1411 ras->ras_next_readahead_idx = ras->ras_window_start_idx;
1413 ras->ras_next_readahead_idx = index + 1;
1416 if (ras->ras_need_increase_window) {
1417 ras_increase_window(inode, ras, ra);
1418 ras->ras_need_increase_window = false;
1423 spin_unlock(&ras->ras_lock);
1426 int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
1428 struct inode *inode = vmpage->mapping->host;
1429 struct ll_inode_info *lli = ll_i2info(inode);
1432 struct cl_page *page;
1433 struct cl_object *clob;
1434 bool redirtied = false;
1435 bool unlocked = false;
1440 LASSERT(PageLocked(vmpage));
1441 LASSERT(!PageWriteback(vmpage));
1443 LASSERT(ll_i2dtexp(inode) != NULL);
1445 env = cl_env_get(&refcheck);
1447 GOTO(out, result = PTR_ERR(env));
1449 clob = ll_i2info(inode)->lli_clob;
1450 LASSERT(clob != NULL);
1452 io = vvp_env_thread_io(env);
1454 io->ci_ignore_layout = 1;
1455 result = cl_io_init(env, io, CIT_MISC, clob);
1457 page = cl_page_find(env, clob, vmpage->index,
1458 vmpage, CPT_CACHEABLE);
1459 if (!IS_ERR(page)) {
1460 lu_ref_add(&page->cp_reference, "writepage",
1462 cl_page_assume(env, io, page);
1463 result = cl_page_flush(env, io, page);
1466 * Re-dirty page on error so it retries write,
1467 * but not in case when IO has actually
1468 * occurred and completed with an error.
1470 if (!PageError(vmpage)) {
1471 redirty_page_for_writepage(wbc, vmpage);
1476 cl_page_disown(env, io, page);
1478 lu_ref_del(&page->cp_reference,
1479 "writepage", current);
1480 cl_page_put(env, page);
1482 result = PTR_ERR(page);
1485 cl_io_fini(env, io);
1487 if (redirtied && wbc->sync_mode == WB_SYNC_ALL) {
1488 loff_t offset = cl_offset(clob, vmpage->index);
1490 /* Flush page failed because the extent is being written out.
1491 * Wait for the write of extent to be finished to avoid
1492 * breaking kernel which assumes ->writepage should mark
1493 * PageWriteback or clean the page. */
1494 result = cl_sync_file_range(inode, offset,
1495 offset + PAGE_SIZE - 1,
1498 /* actually we may have written more than one page.
1499 * decreasing this page because the caller will count
1501 wbc->nr_to_write -= result - 1;
1506 cl_env_put(env, &refcheck);
1511 if (!lli->lli_async_rc)
1512 lli->lli_async_rc = result;
1513 SetPageError(vmpage);
1515 unlock_page(vmpage);
1520 int ll_writepages(struct address_space *mapping, struct writeback_control *wbc)
1522 struct inode *inode = mapping->host;
1525 enum cl_fsync_mode mode;
1526 int range_whole = 0;
1530 if (wbc->range_cyclic) {
1531 start = (loff_t)mapping->writeback_index << PAGE_SHIFT;
1532 end = OBD_OBJECT_EOF;
1534 start = wbc->range_start;
1535 end = wbc->range_end;
1536 if (end == LLONG_MAX) {
1537 end = OBD_OBJECT_EOF;
1538 range_whole = start == 0;
1542 mode = CL_FSYNC_NONE;
1543 if (wbc->sync_mode == WB_SYNC_ALL)
1544 mode = CL_FSYNC_LOCAL;
1546 if (ll_i2info(inode)->lli_clob == NULL)
1549 /* for directio, it would call writepages() to evict cached pages
1550 * inside the IO context of write, which will cause deadlock at
1551 * layout_conf since it waits for active IOs to complete. */
1552 result = cl_sync_file_range(inode, start, end, mode, 1);
1554 wbc->nr_to_write -= result;
1558 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) {
1559 if (end == OBD_OBJECT_EOF)
1560 mapping->writeback_index = 0;
1562 mapping->writeback_index = (end >> PAGE_SHIFT) + 1;
1567 struct ll_cl_context *ll_cl_find(struct file *file)
1569 struct ll_file_data *fd = file->private_data;
1570 struct ll_cl_context *lcc;
1571 struct ll_cl_context *found = NULL;
1573 read_lock(&fd->fd_lock);
1574 list_for_each_entry(lcc, &fd->fd_lccs, lcc_list) {
1575 if (lcc->lcc_cookie == current) {
1580 read_unlock(&fd->fd_lock);
1585 void ll_cl_add(struct file *file, const struct lu_env *env, struct cl_io *io,
1588 struct ll_file_data *fd = file->private_data;
1589 struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1591 memset(lcc, 0, sizeof(*lcc));
1592 INIT_LIST_HEAD(&lcc->lcc_list);
1593 lcc->lcc_cookie = current;
1596 lcc->lcc_type = type;
1598 write_lock(&fd->fd_lock);
1599 list_add(&lcc->lcc_list, &fd->fd_lccs);
1600 write_unlock(&fd->fd_lock);
1603 void ll_cl_remove(struct file *file, const struct lu_env *env)
1605 struct ll_file_data *fd = file->private_data;
1606 struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1608 write_lock(&fd->fd_lock);
1609 list_del_init(&lcc->lcc_list);
1610 write_unlock(&fd->fd_lock);
1613 int ll_io_read_page(const struct lu_env *env, struct cl_io *io,
1614 struct cl_page *page, struct file *file)
1616 struct inode *inode = vvp_object_inode(page->cp_obj);
1617 struct ll_sb_info *sbi = ll_i2sbi(inode);
1618 struct ll_file_data *fd = NULL;
1619 struct ll_readahead_state *ras = NULL;
1620 struct cl_2queue *queue = &io->ci_queue;
1621 struct cl_sync_io *anchor = NULL;
1622 struct vvp_page *vpg;
1623 int rc = 0, rc2 = 0;
1625 pgoff_t io_start_index;
1626 pgoff_t io_end_index;
1630 fd = file->private_data;
1634 vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1635 uptodate = vpg->vpg_defer_uptodate;
1637 if (ll_readahead_enabled(sbi) && !vpg->vpg_ra_updated && ras) {
1638 struct vvp_io *vio = vvp_env_io(env);
1639 enum ras_update_flags flags = 0;
1642 flags |= LL_RAS_HIT;
1643 if (!vio->vui_ra_valid)
1644 flags |= LL_RAS_MMAP;
1645 ras_update(sbi, inode, ras, vvp_index(vpg), flags, io);
1648 cl_2queue_init(queue);
1650 vpg->vpg_ra_used = 1;
1651 cl_page_export(env, page, 1);
1652 cl_page_disown(env, io, page);
1654 anchor = &vvp_env_info(env)->vti_anchor;
1655 cl_sync_io_init(anchor, 1);
1656 page->cp_sync_io = anchor;
1658 cl_2queue_add(queue, page);
1661 io_start_index = cl_index(io->ci_obj, io->u.ci_rw.crw_pos);
1662 io_end_index = cl_index(io->ci_obj, io->u.ci_rw.crw_pos +
1663 io->u.ci_rw.crw_count - 1);
1664 if (ll_readahead_enabled(sbi) && ras && !io->ci_rand_read) {
1665 pgoff_t skip_index = 0;
1667 if (ras->ras_next_readahead_idx < vvp_index(vpg))
1668 skip_index = vvp_index(vpg);
1669 rc2 = ll_readahead(env, io, &queue->c2_qin, ras,
1670 uptodate, file, skip_index);
1671 CDEBUG(D_READA, DFID " %d pages read ahead at %lu\n",
1672 PFID(ll_inode2fid(inode)), rc2, vvp_index(vpg));
1673 } else if (vvp_index(vpg) == io_start_index &&
1674 io_end_index - io_start_index > 0) {
1675 rc2 = ll_readpages(env, io, &queue->c2_qin, io_start_index + 1,
1677 CDEBUG(D_READA, DFID " %d pages read at %lu\n",
1678 PFID(ll_inode2fid(inode)), rc2, vvp_index(vpg));
1681 if (queue->c2_qin.pl_nr > 0) {
1682 int count = queue->c2_qin.pl_nr;
1683 rc = cl_io_submit_rw(env, io, CRT_READ, queue);
1685 task_io_account_read(PAGE_SIZE * count);
1689 if (anchor != NULL && !cl_page_is_owned(page, io)) { /* have sent */
1690 rc = cl_sync_io_wait(env, anchor, 0);
1692 cl_page_assume(env, io, page);
1693 cl_page_list_del(env, &queue->c2_qout, page);
1695 if (!PageUptodate(cl_page_vmpage(page))) {
1696 /* Failed to read a mirror, discard this page so that
1697 * new page can be created with new mirror.
1699 * TODO: this is not needed after page reinit
1700 * route is implemented */
1701 cl_page_discard(env, io, page);
1703 cl_page_disown(env, io, page);
1706 /* TODO: discard all pages until page reinit route is implemented */
1707 cl_page_list_discard(env, io, &queue->c2_qin);
1709 /* Unlock unsent read pages in case of error. */
1710 cl_page_list_disown(env, io, &queue->c2_qin);
1712 cl_2queue_fini(env, queue);
1718 * Possible return value:
1719 * 0 no async readahead triggered and fast read could not be used.
1720 * 1 no async readahead, but fast read could be used.
1721 * 2 async readahead triggered and fast read could be used too.
1724 static int kickoff_async_readahead(struct file *file, unsigned long pages)
1726 struct ll_readahead_work *lrw;
1727 struct inode *inode = file_inode(file);
1728 struct ll_sb_info *sbi = ll_i2sbi(inode);
1729 struct ll_file_data *fd = file->private_data;
1730 struct ll_readahead_state *ras = &fd->fd_ras;
1731 struct ll_ra_info *ra = &sbi->ll_ra_info;
1732 unsigned long throttle;
1733 pgoff_t start_idx = ras_align(ras, ras->ras_next_readahead_idx);
1734 pgoff_t end_idx = start_idx + pages - 1;
1737 * In case we have a limited max_cached_mb, readahead
1738 * should be stopped if it have run out of all LRU slots.
1740 if (atomic_read(&ra->ra_cur_pages) >= sbi->ll_cache->ccc_lru_max) {
1741 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
1745 throttle = min(ra->ra_async_pages_per_file_threshold,
1746 ra->ra_max_pages_per_file);
1748 * If this is strided i/o or the window is smaller than the
1749 * throttle limit, we do not do async readahead. Otherwise,
1750 * we do async readahead, allowing the user thread to do fast i/o.
1752 if (stride_io_mode(ras) || !throttle ||
1753 ras->ras_window_pages < throttle ||
1754 atomic_read(&ra->ra_async_inflight) > ra->ra_async_max_active)
1757 if ((atomic_read(&ra->ra_cur_pages) + pages) > ra->ra_max_pages)
1760 if (ras->ras_async_last_readpage_idx == start_idx)
1763 /* ll_readahead_work_free() free it */
1766 atomic_inc(&sbi->ll_ra_info.ra_async_inflight);
1767 lrw->lrw_file = get_file(file);
1768 lrw->lrw_start_idx = start_idx;
1769 lrw->lrw_end_idx = end_idx;
1770 spin_lock(&ras->ras_lock);
1771 ras->ras_next_readahead_idx = end_idx + 1;
1772 ras->ras_async_last_readpage_idx = start_idx;
1773 spin_unlock(&ras->ras_lock);
1774 memcpy(lrw->lrw_jobid, ll_i2info(inode)->lli_jobid,
1775 sizeof(lrw->lrw_jobid));
1776 ll_readahead_work_add(inode, lrw);
1785 * Check if we can issue a readahead RPC, if that is
1786 * the case, we can't do fast IO because we will need
1787 * a cl_io to issue the RPC.
1789 static bool ll_use_fast_io(struct file *file,
1790 struct ll_readahead_state *ras, pgoff_t index)
1792 unsigned long fast_read_pages =
1793 max(RA_REMAIN_WINDOW_MIN, ras->ras_rpc_pages);
1795 loff_t stride_bytes = ras->ras_stride_bytes;
1797 if (stride_io_mode(ras) && stride_bytes) {
1798 skip_pages = (ras->ras_stride_length +
1799 ras->ras_stride_bytes - 1) / stride_bytes;
1800 skip_pages *= fast_read_pages;
1802 skip_pages = fast_read_pages;
1805 if (ras->ras_window_start_idx + ras->ras_window_pages <
1806 ras->ras_next_readahead_idx + skip_pages ||
1807 kickoff_async_readahead(file, fast_read_pages) > 0)
1813 int ll_readpage(struct file *file, struct page *vmpage)
1815 struct inode *inode = file_inode(file);
1816 struct cl_object *clob = ll_i2info(inode)->lli_clob;
1817 struct ll_cl_context *lcc;
1818 const struct lu_env *env = NULL;
1819 struct cl_io *io = NULL;
1820 struct cl_page *page;
1821 struct ll_sb_info *sbi = ll_i2sbi(inode);
1825 lcc = ll_cl_find(file);
1831 if (io == NULL) { /* fast read */
1832 struct inode *inode = file_inode(file);
1833 struct ll_file_data *fd = file->private_data;
1834 struct ll_readahead_state *ras = &fd->fd_ras;
1835 struct lu_env *local_env = NULL;
1836 struct vvp_page *vpg;
1840 /* TODO: need to verify the layout version to make sure
1841 * the page is not invalid due to layout change. */
1842 page = cl_vmpage_page(vmpage, clob);
1844 unlock_page(vmpage);
1845 ll_ra_stats_inc_sbi(sbi, RA_STAT_FAILED_FAST_READ);
1849 vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1850 if (vpg->vpg_defer_uptodate) {
1851 enum ras_update_flags flags = LL_RAS_HIT;
1853 if (lcc && lcc->lcc_type == LCC_MMAP)
1854 flags |= LL_RAS_MMAP;
1856 /* For fast read, it updates read ahead state only
1857 * if the page is hit in cache because non cache page
1858 * case will be handled by slow read later. */
1859 ras_update(sbi, inode, ras, vvp_index(vpg), flags, io);
1860 /* avoid duplicate ras_update() call */
1861 vpg->vpg_ra_updated = 1;
1863 if (ll_use_fast_io(file, ras, vvp_index(vpg)))
1868 local_env = cl_env_percpu_get();
1872 /* export the page and skip io stack */
1874 vpg->vpg_ra_used = 1;
1875 cl_page_export(env, page, 1);
1877 ll_ra_stats_inc_sbi(sbi, RA_STAT_FAILED_FAST_READ);
1879 /* release page refcount before unlocking the page to ensure
1880 * the object won't be destroyed in the calling path of
1881 * cl_page_put(). Please see comment in ll_releasepage(). */
1882 cl_page_put(env, page);
1883 unlock_page(vmpage);
1885 cl_env_percpu_put(local_env);
1891 * Direct read can fall back to buffered read, but DIO is done
1892 * with lockless i/o, and buffered requires LDLM locking, so in
1893 * this case we must restart without lockless.
1895 if (file->f_flags & O_DIRECT &&
1896 lcc && lcc->lcc_type == LCC_RW &&
1898 unlock_page(vmpage);
1899 io->ci_dio_lock = 1;
1900 io->ci_need_restart = 1;
1904 LASSERT(io->ci_state == CIS_IO_GOING);
1905 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
1906 if (!IS_ERR(page)) {
1907 LASSERT(page->cp_type == CPT_CACHEABLE);
1908 if (likely(!PageUptodate(vmpage))) {
1909 cl_page_assume(env, io, page);
1911 result = ll_io_read_page(env, io, page, file);
1913 /* Page from a non-object file. */
1914 unlock_page(vmpage);
1917 cl_page_put(env, page);
1919 unlock_page(vmpage);
1920 result = PTR_ERR(page);