4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
33 * Lustre Lite I/O page cache routines shared by different kernel revs
36 #include <linux/kernel.h>
38 #include <linux/string.h>
39 #include <linux/stat.h>
40 #include <linux/errno.h>
41 #include <linux/unistd.h>
42 #include <linux/writeback.h>
43 #include <asm/uaccess.h>
46 #include <linux/file.h>
47 #include <linux/stat.h>
48 #include <asm/uaccess.h>
50 #include <linux/pagemap.h>
51 /* current_is_kswapd() */
52 #include <linux/swap.h>
53 #include <linux/task_io_accounting_ops.h>
55 #define DEBUG_SUBSYSTEM S_LLITE
57 #include <obd_cksum.h>
58 #include "llite_internal.h"
59 #include <lustre_compat.h>
61 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which);
64 * Get readahead pages from the filesystem readahead pool of the client for a
67 * /param sbi superblock for filesystem readahead state ll_ra_info
68 * /param ria per-thread readahead state
69 * /param pages number of pages requested for readahead for the thread.
71 * WARNING: This algorithm is used to reduce contention on sbi->ll_lock.
72 * It should work well if the ra_max_pages is much greater than the single
73 * file's read-ahead window, and not too many threads contending for
74 * these readahead pages.
76 * TODO: There may be a 'global sync problem' if many threads are trying
77 * to get an ra budget that is larger than the remaining readahead pages
78 * and reach here at exactly the same time. They will compute /a ret to
79 * consume the remaining pages, but will fail at atomic_add_return() and
80 * get a zero ra window, although there is still ra space remaining. - Jay */
82 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi,
83 struct ra_io_arg *ria,
85 unsigned long pages_min)
87 struct ll_ra_info *ra = &sbi->ll_ra_info;
92 WARN_ON_ONCE(pages_min > pages);
94 * Don't try readahead aggresively if we are limited
95 * LRU pages, otherwise, it could cause deadlock.
97 pages = min(sbi->ll_cache->ccc_lru_max >> 2, pages);
99 * if this happen, we reserve more pages than needed,
100 * this will make us leak @ra_cur_pages, because
101 * ll_ra_count_put() acutally freed @pages.
103 if (unlikely(pages_min > pages))
107 * If read-ahead pages left are less than 1M, do not do read-ahead,
108 * otherwise it will form small read RPC(< 1M), which hurt server
111 ret = min(ra->ra_max_pages - atomic_read(&ra->ra_cur_pages),
113 if (ret < 0 || ret < min_t(long, PTLRPC_MAX_BRW_PAGES, pages))
116 if (atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) {
117 atomic_sub(ret, &ra->ra_cur_pages);
122 if (ret < pages_min) {
123 /* override ra limit for maximum performance */
124 atomic_add(pages_min - ret, &ra->ra_cur_pages);
130 void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long pages)
132 struct ll_ra_info *ra = &sbi->ll_ra_info;
133 atomic_sub(pages, &ra->ra_cur_pages);
136 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which)
138 LASSERTF(which < _NR_RA_STAT, "which: %u\n", which);
139 lprocfs_counter_incr(sbi->ll_ra_stats, which);
142 static inline bool ll_readahead_enabled(struct ll_sb_info *sbi)
144 return sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
145 sbi->ll_ra_info.ra_max_pages > 0;
148 void ll_ra_stats_inc(struct inode *inode, enum ra_stat which)
150 struct ll_sb_info *sbi = ll_i2sbi(inode);
151 ll_ra_stats_inc_sbi(sbi, which);
154 #define RAS_CDEBUG(ras) \
156 "lre %llu cr %lu cb %llu wsi %lu wp %lu nra %lu rpc %lu " \
157 "r %lu csr %lu so %llu sb %llu sl %llu lr %lu\n", \
158 ras->ras_last_read_end_bytes, ras->ras_consecutive_requests, \
159 ras->ras_consecutive_bytes, ras->ras_window_start_idx, \
160 ras->ras_window_pages, ras->ras_next_readahead_idx, \
161 ras->ras_rpc_pages, ras->ras_requests, \
162 ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
163 ras->ras_stride_bytes, ras->ras_stride_length, \
164 ras->ras_async_last_readpage_idx)
166 static bool pos_in_window(loff_t pos, loff_t point,
167 unsigned long before, unsigned long after)
169 loff_t start = point - before;
170 loff_t end = point + after;
177 return start <= pos && pos <= end;
180 enum ll_ra_page_hint {
181 MAYNEED = 0, /* this page possibly accessed soon */
182 WILLNEED /* this page is gurateed to be needed */
186 * Initiates read-ahead of a page with given index.
188 * \retval +ve: page was already uptodate so it will be skipped
190 * \retval -ve: page wasn't added to \a queue for error;
191 * \retval 0: page was added into \a queue for read ahead.
193 static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io,
194 struct cl_page_list *queue, pgoff_t index,
195 enum ll_ra_page_hint hint)
197 struct cl_object *clob = io->ci_obj;
198 struct inode *inode = vvp_object_inode(clob);
199 struct page *vmpage = NULL;
200 struct cl_page *page;
201 struct vvp_page *vpg;
202 enum ra_stat which = _NR_RA_STAT; /* keep gcc happy */
204 const char *msg = NULL;
210 vmpage = grab_cache_page_nowait(inode->i_mapping, index);
211 if (vmpage == NULL) {
212 which = RA_STAT_FAILED_GRAB_PAGE;
213 msg = "g_c_p_n failed";
214 GOTO(out, rc = -EBUSY);
218 vmpage = find_or_create_page(inode->i_mapping, index,
221 GOTO(out, rc = -ENOMEM);
224 /* should not come here */
225 GOTO(out, rc = -EINVAL);
228 /* Check if vmpage was truncated or reclaimed */
229 if (vmpage->mapping != inode->i_mapping) {
230 which = RA_STAT_WRONG_GRAB_PAGE;
231 msg = "g_c_p_n returned invalid page";
232 GOTO(out, rc = -EBUSY);
235 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
237 which = RA_STAT_FAILED_GRAB_PAGE;
238 msg = "cl_page_find failed";
239 GOTO(out, rc = PTR_ERR(page));
242 lu_ref_add(&page->cp_reference, "ra", current);
243 cl_page_assume(env, io, page);
244 vpg = cl2vvp_page(cl_object_page_slice(clob, page));
245 if (!vpg->vpg_defer_uptodate && !PageUptodate(vmpage)) {
246 if (hint == MAYNEED) {
247 vpg->vpg_defer_uptodate = 1;
248 vpg->vpg_ra_used = 0;
250 cl_page_list_add(queue, page, true);
252 /* skip completed pages */
253 cl_page_unassume(env, io, page);
254 /* This page is already uptodate, returning a positive number
255 * to tell the callers about this */
259 lu_ref_del(&page->cp_reference, "ra", current);
260 cl_page_put(env, page);
263 if (vmpage != NULL) {
268 if (msg != NULL && hint == MAYNEED) {
269 ll_ra_stats_inc(inode, which);
270 CDEBUG(D_READA, "%s\n", msg);
277 #define RIA_DEBUG(ria) \
278 CDEBUG(D_READA, "rs %lu re %lu ro %llu rl %llu rb %llu\n", \
279 ria->ria_start_idx, ria->ria_end_idx, ria->ria_stoff, \
280 ria->ria_length, ria->ria_bytes)
282 static inline int stride_io_mode(struct ll_readahead_state *ras)
284 return ras->ras_consecutive_stride_requests > 1;
287 /* The function calculates how many bytes will be read in
288 * [off, off + length], in such stride IO area,
289 * stride_offset = st_off, stride_lengh = st_len,
290 * stride_bytes = st_bytes
292 * |------------------|*****|------------------|*****|------------|*****|....
295 * |----- st_len -----|
297 * How many bytes it should read in such pattern
298 * |-------------------------------------------------------------|
300 * |<------ length ------->|
302 * = |<----->| + |-------------------------------------| + |---|
303 * start_left st_bytes * i end_left
305 static loff_t stride_byte_count(loff_t st_off, loff_t st_len, loff_t st_bytes,
306 loff_t off, loff_t length)
308 u64 start = off > st_off ? off - st_off : 0;
309 u64 end = off + length > st_off ? off + length - st_off : 0;
314 if (st_len == 0 || length == 0 || end == 0)
317 start = div64_u64_rem(start, st_len, &start_left);
318 if (start_left < st_bytes)
319 start_left = st_bytes - start_left;
323 end = div64_u64_rem(end, st_len, &end_left);
324 if (end_left > st_bytes)
327 CDEBUG(D_READA, "start %llu, end %llu start_left %llu end_left %llu\n",
328 start, end, start_left, end_left);
331 bytes_count = end_left - (st_bytes - start_left);
333 bytes_count = start_left +
334 st_bytes * (end - start - 1) + end_left;
337 "st_off %llu, st_len %llu st_bytes %llu off %llu length %llu bytescount %llu\n",
338 st_off, st_len, st_bytes, off, length, bytes_count);
343 static unsigned long ria_page_count(struct ra_io_arg *ria)
345 loff_t length_bytes = ria->ria_end_idx >= ria->ria_start_idx ?
346 (loff_t)(ria->ria_end_idx -
347 ria->ria_start_idx + 1) << PAGE_SHIFT : 0;
350 if (ria->ria_length > ria->ria_bytes && ria->ria_bytes &&
351 (ria->ria_length & ~PAGE_MASK || ria->ria_bytes & ~PAGE_MASK ||
352 ria->ria_stoff & ~PAGE_MASK)) {
353 /* Over-estimate un-aligned page stride read */
354 unsigned long pg_count = ((ria->ria_bytes +
355 PAGE_SIZE - 1) >> PAGE_SHIFT) + 1;
356 pg_count *= length_bytes / ria->ria_length + 1;
360 bytes_count = stride_byte_count(ria->ria_stoff, ria->ria_length,
362 (loff_t)ria->ria_start_idx<<PAGE_SHIFT,
364 return (bytes_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
367 static pgoff_t ras_align(struct ll_readahead_state *ras, pgoff_t index)
369 unsigned opt_size = min(ras->ras_window_pages, ras->ras_rpc_pages);
373 return index - (index % opt_size);
376 /* Check whether the index is in the defined ra-window */
377 static bool ras_inside_ra_window(pgoff_t idx, struct ra_io_arg *ria)
379 loff_t pos = (loff_t)idx << PAGE_SHIFT;
381 /* If ria_length == ria_bytes, it means non-stride I/O mode,
382 * idx should always inside read-ahead window in this case
383 * For stride I/O mode, just check whether the idx is inside
386 if (ria->ria_length == 0 || ria->ria_length == ria->ria_bytes)
389 if (pos >= ria->ria_stoff) {
392 div64_u64_rem(pos - ria->ria_stoff, ria->ria_length, &offset);
394 if (offset < ria->ria_bytes ||
395 (ria->ria_length - offset) < PAGE_SIZE)
397 } else if (pos + PAGE_SIZE > ria->ria_stoff) {
405 ll_read_ahead_pages(const struct lu_env *env, struct cl_io *io,
406 struct cl_page_list *queue, struct ll_readahead_state *ras,
407 struct ra_io_arg *ria, pgoff_t *ra_end, pgoff_t skip_index)
409 struct cl_read_ahead ra = { 0 };
410 /* busy page count is per stride */
411 int rc = 0, count = 0, busy_page_count = 0;
414 LASSERT(ria != NULL);
417 for (page_idx = ria->ria_start_idx;
418 page_idx <= ria->ria_end_idx && ria->ria_reserved > 0;
420 if (skip_index && page_idx == skip_index)
422 if (ras_inside_ra_window(page_idx, ria)) {
423 if (ra.cra_end_idx == 0 || ra.cra_end_idx < page_idx) {
427 * Do not shrink ria_end_idx at any case until
428 * the minimum end of current read is covered.
430 * Do not extend read lock accross stripe if
431 * lock contention detected.
433 if (ra.cra_contention &&
434 page_idx > ria->ria_end_idx_min) {
435 ria->ria_end_idx = *ra_end;
439 cl_read_ahead_release(env, &ra);
441 rc = cl_io_read_ahead(env, io, page_idx, &ra);
446 * Only shrink ria_end_idx if the matched
447 * LDLM lock doesn't cover more.
449 if (page_idx > ra.cra_end_idx) {
450 ria->ria_end_idx = ra.cra_end_idx;
454 CDEBUG(D_READA, "idx: %lu, ra: %lu, rpc: %lu\n",
455 page_idx, ra.cra_end_idx,
457 LASSERTF(ra.cra_end_idx >= page_idx,
458 "object: %p, indcies %lu / %lu\n",
459 io->ci_obj, ra.cra_end_idx, page_idx);
460 /* update read ahead RPC size.
461 * NB: it's racy but doesn't matter */
462 if (ras->ras_rpc_pages != ra.cra_rpc_pages &&
463 ra.cra_rpc_pages > 0)
464 ras->ras_rpc_pages = ra.cra_rpc_pages;
466 /* trim it to align with optimal RPC size */
467 end_idx = ras_align(ras, ria->ria_end_idx + 1);
468 if (end_idx > 0 && !ria->ria_eof)
469 ria->ria_end_idx = end_idx - 1;
471 if (ria->ria_end_idx < ria->ria_end_idx_min)
472 ria->ria_end_idx = ria->ria_end_idx_min;
474 if (page_idx > ria->ria_end_idx)
477 /* If the page is inside the read-ahead window */
478 rc = ll_read_ahead_page(env, io, queue, page_idx,
480 if (rc < 0 && rc != -EBUSY)
485 "skip busy page: %lu\n", page_idx);
486 /* For page unaligned readahead the first
487 * last pages of each region can be read by
488 * another reader on the same node, and so
489 * may be busy. So only stop for > 2 busy
491 if (busy_page_count > 2)
496 /* Only subtract from reserve & count the page if we
497 * really did readahead on that page. */
502 } else if (stride_io_mode(ras)) {
503 /* If it is not in the read-ahead window, and it is
504 * read-ahead mode, then check whether it should skip
507 loff_t pos = (loff_t)page_idx << PAGE_SHIFT;
510 div64_u64_rem(pos - ria->ria_stoff, ria->ria_length,
512 if (offset >= ria->ria_bytes) {
513 pos += (ria->ria_length - offset);
514 if ((pos >> PAGE_SHIFT) >= page_idx + 1)
515 page_idx = (pos >> PAGE_SHIFT) - 1;
518 "Stride: jump %llu pages to %lu\n",
519 ria->ria_length - offset, page_idx);
525 cl_read_ahead_release(env, &ra);
530 static void ll_readahead_work_free(struct ll_readahead_work *work)
532 fput(work->lrw_file);
536 static void ll_readahead_handle_work(struct work_struct *wq);
537 static void ll_readahead_work_add(struct inode *inode,
538 struct ll_readahead_work *work)
540 INIT_WORK(&work->lrw_readahead_work, ll_readahead_handle_work);
541 queue_work(ll_i2sbi(inode)->ll_ra_info.ll_readahead_wq,
542 &work->lrw_readahead_work);
545 static int ll_readahead_file_kms(const struct lu_env *env,
546 struct cl_io *io, __u64 *kms)
548 struct cl_object *clob;
550 struct cl_attr *attr = vvp_env_thread_attr(env);
554 inode = vvp_object_inode(clob);
556 cl_object_attr_lock(clob);
557 ret = cl_object_attr_get(env, clob, attr);
558 cl_object_attr_unlock(clob);
563 *kms = attr->cat_kms;
567 static void ll_readahead_handle_work(struct work_struct *wq)
569 struct ll_readahead_work *work;
572 struct ra_io_arg *ria;
574 struct ll_file_data *fd;
575 struct ll_readahead_state *ras;
577 struct cl_2queue *queue;
578 pgoff_t ra_end_idx = 0;
579 unsigned long pages, pages_min = 0;
584 struct ll_sb_info *sbi;
586 work = container_of(wq, struct ll_readahead_work,
588 fd = work->lrw_file->private_data;
590 file = work->lrw_file;
591 inode = file_inode(file);
592 sbi = ll_i2sbi(inode);
594 CDEBUG(D_READA|D_IOTRACE,
595 "%s: async ra from %lu to %lu triggered by user pid %d\n",
596 file_dentry(file)->d_name.name, work->lrw_start_idx,
597 work->lrw_end_idx, work->lrw_user_pid);
599 env = cl_env_alloc(&refcheck, LCT_NOREF);
601 GOTO(out_free_work, rc = PTR_ERR(env));
603 io = vvp_env_thread_io(env);
604 ll_io_init(io, file, CIT_READ, NULL);
606 rc = ll_readahead_file_kms(env, io, &kms);
608 GOTO(out_put_env, rc);
611 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
612 GOTO(out_put_env, rc = 0);
615 ria = &ll_env_info(env)->lti_ria;
616 memset(ria, 0, sizeof(*ria));
618 ria->ria_start_idx = work->lrw_start_idx;
619 /* Truncate RA window to end of file */
620 eof_index = (pgoff_t)(kms - 1) >> PAGE_SHIFT;
621 if (eof_index <= work->lrw_end_idx) {
622 work->lrw_end_idx = eof_index;
625 if (work->lrw_end_idx <= work->lrw_start_idx)
626 GOTO(out_put_env, rc = 0);
628 ria->ria_end_idx = work->lrw_end_idx;
629 pages = ria->ria_end_idx - ria->ria_start_idx + 1;
630 ria->ria_reserved = ll_ra_count_get(sbi, ria,
631 ria_page_count(ria), pages_min);
634 "async reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
635 ria->ria_reserved, pages, pages_min,
636 atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
637 ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
639 if (ria->ria_reserved < pages) {
640 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
641 if (PAGES_TO_MiB(ria->ria_reserved) < 1) {
642 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
643 GOTO(out_put_env, rc = 0);
647 rc = cl_io_rw_init(env, io, CIT_READ, ria->ria_start_idx, pages);
649 GOTO(out_put_env, rc);
651 /* overwrite jobid inited in vvp_io_init() */
652 if (strncmp(ll_i2info(inode)->lli_jobid, work->lrw_jobid,
653 sizeof(work->lrw_jobid)))
654 memcpy(ll_i2info(inode)->lli_jobid, work->lrw_jobid,
655 sizeof(work->lrw_jobid));
657 vvp_env_io(env)->vui_fd = fd;
658 io->ci_state = CIS_LOCKED;
659 io->ci_async_readahead = true;
660 rc = cl_io_start(env, io);
662 GOTO(out_io_fini, rc);
664 queue = &io->ci_queue;
665 cl_2queue_init(queue);
667 rc = ll_read_ahead_pages(env, io, &queue->c2_qin, ras, ria,
669 if (ria->ria_reserved != 0)
670 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
671 if (queue->c2_qin.pl_nr > 0) {
672 int count = queue->c2_qin.pl_nr;
674 rc = cl_io_submit_rw(env, io, CRT_READ, queue);
676 task_io_account_read(PAGE_SIZE * count);
678 if (ria->ria_end_idx == ra_end_idx && ra_end_idx == (kms >> PAGE_SHIFT))
679 ll_ra_stats_inc(inode, RA_STAT_EOF);
681 if (ra_end_idx != ria->ria_end_idx)
682 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
684 /* TODO: discard all pages until page reinit route is implemented */
685 cl_page_list_discard(env, io, &queue->c2_qin);
687 /* Unlock unsent read pages in case of error. */
688 cl_page_list_disown(env, io, &queue->c2_qin);
690 cl_2queue_fini(env, queue);
695 cl_env_put(env, &refcheck);
698 ll_ra_stats_inc_sbi(ll_i2sbi(inode), RA_STAT_ASYNC);
699 atomic_dec(&sbi->ll_ra_info.ra_async_inflight);
700 ll_readahead_work_free(work);
703 static int ll_readahead(const struct lu_env *env, struct cl_io *io,
704 struct cl_page_list *queue,
705 struct ll_readahead_state *ras, bool hit,
706 struct file *file, pgoff_t skip_index,
709 struct vvp_io *vio = vvp_env_io(env);
710 struct ll_thread_info *lti = ll_env_info(env);
711 unsigned long pages, pages_min = 0;
712 pgoff_t ra_end_idx = 0, end_idx = 0;
714 struct ra_io_arg *ria = <i->lti_ria;
715 struct cl_object *clob;
718 struct ll_sb_info *sbi;
719 struct ll_ra_info *ra;
724 inode = vvp_object_inode(clob);
725 sbi = ll_i2sbi(inode);
726 ra = &sbi->ll_ra_info;
729 * In case we have a limited max_cached_mb, readahead
730 * should be stopped if it have run out of all LRU slots.
732 if (atomic_read(&ra->ra_cur_pages) >= sbi->ll_cache->ccc_lru_max) {
733 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
737 memset(ria, 0, sizeof(*ria));
738 ret = ll_readahead_file_kms(env, io, &kms);
743 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
747 spin_lock(&ras->ras_lock);
750 * Note: other thread might rollback the ras_next_readahead_idx,
751 * if it can not get the full size of prepared pages, see the
752 * end of this function. For stride read ahead, it needs to
753 * make sure the offset is no less than ras_stride_offset,
754 * so that stride read ahead can work correctly.
756 if (stride_io_mode(ras))
757 *start_idx = max_t(pgoff_t, ras->ras_next_readahead_idx,
758 ras->ras_stride_offset >> PAGE_SHIFT);
760 *start_idx = ras->ras_next_readahead_idx;
762 if (ras->ras_window_pages > 0)
763 end_idx = ras->ras_window_start_idx + ras->ras_window_pages - 1;
766 end_idx = *start_idx + ras->ras_window_pages - 1;
768 /* Enlarge the RA window to encompass the full read */
769 if (vio->vui_ra_valid &&
770 end_idx < vio->vui_ra_start_idx + vio->vui_ra_pages - 1)
771 end_idx = vio->vui_ra_start_idx + vio->vui_ra_pages - 1;
776 /* Truncate RA window to end of file */
777 eof_index = (pgoff_t)((kms - 1) >> PAGE_SHIFT);
778 if (eof_index <= end_idx) {
783 ria->ria_start_idx = *start_idx;
784 ria->ria_end_idx = end_idx;
785 /* If stride I/O mode is detected, get stride window*/
786 if (stride_io_mode(ras)) {
787 ria->ria_stoff = ras->ras_stride_offset;
788 ria->ria_length = ras->ras_stride_length;
789 ria->ria_bytes = ras->ras_stride_bytes;
791 spin_unlock(&ras->ras_lock);
794 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
797 pages = ria_page_count(ria);
799 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
804 CDEBUG(D_READA, DFID": ria: %lu/%lu, bead: %lu/%lu, hit: %d\n",
805 PFID(lu_object_fid(&clob->co_lu)),
806 ria->ria_start_idx, ria->ria_end_idx,
807 vio->vui_ra_valid ? vio->vui_ra_start_idx : 0,
808 vio->vui_ra_valid ? vio->vui_ra_pages : 0,
811 /* at least to extend the readahead window to cover current read */
812 if (!hit && vio->vui_ra_valid &&
813 vio->vui_ra_start_idx + vio->vui_ra_pages > ria->ria_start_idx) {
814 ria->ria_end_idx_min =
815 vio->vui_ra_start_idx + vio->vui_ra_pages - 1;
816 pages_min = vio->vui_ra_start_idx + vio->vui_ra_pages -
819 * For performance reason, exceeding @ra_max_pages
820 * are allowed, but this should be limited with RPC
821 * size in case a large block size read issued. Trim
824 pages_min = min(pages_min, ras->ras_rpc_pages -
825 (ria->ria_start_idx % ras->ras_rpc_pages));
828 /* don't over reserved for mmap range read */
831 if (pages_min > pages)
833 ria->ria_reserved = ll_ra_count_get(ll_i2sbi(inode), ria, pages,
835 if (ria->ria_reserved < pages)
836 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
838 CDEBUG(D_READA, "reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
839 ria->ria_reserved, pages, pages_min,
840 atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
841 ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
843 ret = ll_read_ahead_pages(env, io, queue, ras, ria, &ra_end_idx,
845 if (ria->ria_reserved != 0)
846 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
848 if (ra_end_idx == end_idx && ra_end_idx == (kms >> PAGE_SHIFT))
849 ll_ra_stats_inc(inode, RA_STAT_EOF);
852 "ra_end_idx = %lu end_idx = %lu stride end = %lu pages = %d\n",
853 ra_end_idx, end_idx, ria->ria_end_idx, ret);
855 if (ra_end_idx != end_idx)
856 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
857 if (ra_end_idx > 0) {
858 /* update the ras so that the next read-ahead tries from
859 * where we left off. */
860 spin_lock(&ras->ras_lock);
861 ras->ras_next_readahead_idx = ra_end_idx + 1;
862 spin_unlock(&ras->ras_lock);
869 static int ll_readpages(const struct lu_env *env, struct cl_io *io,
870 struct cl_page_list *queue,
871 pgoff_t start, pgoff_t end)
880 ret = ll_readahead_file_kms(env, io, &kms);
888 unsigned long end_index;
890 end_index = (unsigned long)((kms - 1) >> PAGE_SHIFT);
891 if (end_index <= end)
895 for (page_idx = start; page_idx <= end; page_idx++) {
896 ret= ll_read_ahead_page(env, io, queue, page_idx,
900 else if (ret == 0) /* ret 1 is already uptodate */
904 RETURN(count > 0 ? count : ret);
907 static void ras_set_start(struct ll_readahead_state *ras, pgoff_t index)
909 ras->ras_window_start_idx = ras_align(ras, index);
912 /* called with the ras_lock held or from places where it doesn't matter */
913 static void ras_reset(struct ll_readahead_state *ras, pgoff_t index)
915 ras->ras_consecutive_requests = 0;
916 ras->ras_consecutive_bytes = 0;
917 ras->ras_window_pages = 0;
918 ras_set_start(ras, index);
919 ras->ras_next_readahead_idx = max(ras->ras_window_start_idx, index + 1);
924 /* called with the ras_lock held or from places where it doesn't matter */
925 static void ras_stride_reset(struct ll_readahead_state *ras)
927 ras->ras_consecutive_stride_requests = 0;
928 ras->ras_stride_length = 0;
929 ras->ras_stride_bytes = 0;
933 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
935 spin_lock_init(&ras->ras_lock);
936 ras->ras_rpc_pages = PTLRPC_MAX_BRW_PAGES;
938 ras->ras_last_read_end_bytes = 0;
939 ras->ras_requests = 0;
940 ras->ras_range_min_start_idx = 0;
941 ras->ras_range_max_end_idx = 0;
942 ras->ras_range_requests = 0;
943 ras->ras_last_range_pages = 0;
947 * Check whether the read request is in the stride window.
948 * If it is in the stride window, return true, otherwise return false.
950 static bool read_in_stride_window(struct ll_readahead_state *ras,
951 loff_t pos, loff_t count)
955 if (ras->ras_stride_length == 0 || ras->ras_stride_bytes == 0 ||
956 ras->ras_stride_bytes == ras->ras_stride_length)
959 stride_gap = pos - ras->ras_last_read_end_bytes - 1;
961 /* If it is contiguous read */
963 return ras->ras_consecutive_bytes + count <=
964 ras->ras_stride_bytes;
966 /* Otherwise check the stride by itself */
967 return (ras->ras_stride_length - ras->ras_stride_bytes) == stride_gap &&
968 ras->ras_consecutive_bytes == ras->ras_stride_bytes &&
969 count <= ras->ras_stride_bytes;
972 static void ras_init_stride_detector(struct ll_readahead_state *ras,
973 loff_t pos, loff_t count)
975 loff_t stride_gap = pos - ras->ras_last_read_end_bytes - 1;
977 LASSERT(ras->ras_consecutive_stride_requests == 0);
979 if (pos <= ras->ras_last_read_end_bytes) {
980 /*Reset stride window for forward read*/
981 ras_stride_reset(ras);
985 ras->ras_stride_bytes = ras->ras_consecutive_bytes;
986 ras->ras_stride_length = stride_gap + ras->ras_consecutive_bytes;
987 ras->ras_consecutive_stride_requests++;
988 ras->ras_stride_offset = pos;
994 stride_page_count(struct ll_readahead_state *ras, loff_t len)
997 stride_byte_count(ras->ras_stride_offset,
998 ras->ras_stride_length, ras->ras_stride_bytes,
999 ras->ras_window_start_idx << PAGE_SHIFT, len);
1001 return (bytes_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
1004 /* Stride Read-ahead window will be increased inc_len according to
1005 * stride I/O pattern */
1006 static void ras_stride_increase_window(struct ll_readahead_state *ras,
1007 struct ll_ra_info *ra, loff_t inc_bytes)
1009 loff_t window_bytes, stride_bytes;
1014 /* temporarily store in page units to reduce LASSERT() cost below */
1015 end = ras->ras_window_start_idx + ras->ras_window_pages;
1017 LASSERT(ras->ras_stride_length > 0);
1018 LASSERTF(end >= (ras->ras_stride_offset >> PAGE_SHIFT),
1019 "window_start_idx %lu, window_pages %lu stride_offset %llu\n",
1020 ras->ras_window_start_idx, ras->ras_window_pages,
1021 ras->ras_stride_offset);
1024 if (end <= ras->ras_stride_offset)
1027 stride_bytes = end - ras->ras_stride_offset;
1029 div64_u64_rem(stride_bytes, ras->ras_stride_length, &left_bytes);
1030 window_bytes = (ras->ras_window_pages << PAGE_SHIFT);
1031 if (left_bytes < ras->ras_stride_bytes) {
1032 if (ras->ras_stride_bytes - left_bytes >= inc_bytes) {
1033 window_bytes += inc_bytes;
1036 window_bytes += (ras->ras_stride_bytes - left_bytes);
1037 inc_bytes -= (ras->ras_stride_bytes - left_bytes);
1040 window_bytes += (ras->ras_stride_length - left_bytes);
1043 LASSERT(ras->ras_stride_bytes != 0);
1045 step = div64_u64_rem(inc_bytes, ras->ras_stride_bytes, &left_bytes);
1047 window_bytes += step * ras->ras_stride_length + left_bytes;
1048 LASSERT(window_bytes > 0);
1051 if (stride_page_count(ras, window_bytes) <=
1052 ra->ra_max_pages_per_file || ras->ras_window_pages == 0)
1053 ras->ras_window_pages = (window_bytes >> PAGE_SHIFT);
1055 LASSERT(ras->ras_window_pages > 0);
1060 static void ras_increase_window(struct inode *inode,
1061 struct ll_readahead_state *ras,
1062 struct ll_ra_info *ra)
1064 /* The stretch of ra-window should be aligned with max rpc_size
1065 * but current clio architecture does not support retrieve such
1066 * information from lower layer. FIXME later
1068 if (stride_io_mode(ras)) {
1069 ras_stride_increase_window(ras, ra,
1070 (loff_t)ras->ras_rpc_pages << PAGE_SHIFT);
1072 pgoff_t window_pages;
1074 window_pages = min(ras->ras_window_pages + ras->ras_rpc_pages,
1075 ra->ra_max_pages_per_file);
1076 if (window_pages < ras->ras_rpc_pages)
1077 ras->ras_window_pages = window_pages;
1079 ras->ras_window_pages = ras_align(ras, window_pages);
1084 * Seek within 8 pages are considered as sequential read for now.
1086 static inline bool is_loose_seq_read(struct ll_readahead_state *ras, loff_t pos)
1088 return pos_in_window(pos, ras->ras_last_read_end_bytes,
1089 8UL << PAGE_SHIFT, 8UL << PAGE_SHIFT);
1092 static inline bool is_loose_mmap_read(struct ll_sb_info *sbi,
1093 struct ll_readahead_state *ras,
1096 unsigned long range_pages = sbi->ll_ra_info.ra_range_pages;
1098 return pos_in_window(pos, ras->ras_last_read_end_bytes,
1099 range_pages << PAGE_SHIFT,
1100 range_pages << PAGE_SHIFT);
1104 * We have observed slow mmap read performances for some
1105 * applications. The problem is if access pattern is neither
1106 * sequential nor stride, but could be still adjacent in a
1107 * small range and then seek a random position.
1109 * So the pattern could be something like this:
1111 * [1M data] [hole] [0.5M data] [hole] [0.7M data] [1M data]
1114 * Every time an application reads mmap data, it may not only
1115 * read a single 4KB page, but aslo a cluster of nearby pages in
1116 * a range(e.g. 1MB) of the first page after a cache miss.
1118 * The readahead engine is modified to track the range size of
1119 * a cluster of mmap reads, so that after a seek and/or cache miss,
1120 * the range size is used to efficiently prefetch multiple pages
1121 * in a single RPC rather than many small RPCs.
1123 static void ras_detect_cluster_range(struct ll_readahead_state *ras,
1124 struct ll_sb_info *sbi,
1125 unsigned long pos, unsigned long count)
1127 pgoff_t last_pages, pages;
1128 pgoff_t end_idx = (pos + count - 1) >> PAGE_SHIFT;
1130 last_pages = ras->ras_range_max_end_idx -
1131 ras->ras_range_min_start_idx + 1;
1132 /* First time come here */
1133 if (!ras->ras_range_max_end_idx)
1136 /* Random or Stride read */
1137 if (!is_loose_mmap_read(sbi, ras, pos))
1140 ras->ras_range_requests++;
1141 if (ras->ras_range_max_end_idx < end_idx)
1142 ras->ras_range_max_end_idx = end_idx;
1144 if (ras->ras_range_min_start_idx > (pos >> PAGE_SHIFT))
1145 ras->ras_range_min_start_idx = pos >> PAGE_SHIFT;
1147 /* Out of range, consider it as random or stride */
1148 pages = ras->ras_range_max_end_idx -
1149 ras->ras_range_min_start_idx + 1;
1150 if (pages <= sbi->ll_ra_info.ra_range_pages)
1153 ras->ras_last_range_pages = last_pages;
1154 ras->ras_range_requests = 0;
1155 ras->ras_range_min_start_idx = pos >> PAGE_SHIFT;
1156 ras->ras_range_max_end_idx = end_idx;
1159 static void ras_detect_read_pattern(struct ll_readahead_state *ras,
1160 struct ll_sb_info *sbi,
1161 loff_t pos, size_t count, bool mmap)
1163 bool stride_detect = false;
1164 pgoff_t index = pos >> PAGE_SHIFT;
1167 * Reset the read-ahead window in two cases. First when the app seeks
1168 * or reads to some other part of the file. Secondly if we get a
1169 * read-ahead miss that we think we've previously issued. This can
1170 * be a symptom of there being so many read-ahead pages that the VM
1171 * is reclaiming it before we get to it.
1173 if (!is_loose_seq_read(ras, pos)) {
1174 /* Check whether it is in stride I/O mode */
1175 if (!read_in_stride_window(ras, pos, count)) {
1176 if (ras->ras_consecutive_stride_requests == 0)
1177 ras_init_stride_detector(ras, pos, count);
1179 ras_stride_reset(ras);
1180 ras->ras_consecutive_bytes = 0;
1181 ras_reset(ras, index);
1183 ras->ras_consecutive_bytes = 0;
1184 ras->ras_consecutive_requests = 0;
1185 if (++ras->ras_consecutive_stride_requests > 1)
1186 stride_detect = true;
1189 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
1190 } else if (stride_io_mode(ras)) {
1192 * If this is contiguous read but in stride I/O mode
1193 * currently, check whether stride step still is valid,
1194 * if invalid, it will reset the stride ra window to
1197 if (!read_in_stride_window(ras, pos, count)) {
1198 ras_stride_reset(ras);
1199 ras->ras_window_pages = 0;
1200 ras->ras_next_readahead_idx = index;
1204 ras->ras_consecutive_bytes += count;
1206 pgoff_t idx = ras->ras_consecutive_bytes >> PAGE_SHIFT;
1207 unsigned long ra_range_pages =
1208 max_t(unsigned long, RA_MIN_MMAP_RANGE_PAGES,
1209 sbi->ll_ra_info.ra_range_pages);
1211 if ((idx >= ra_range_pages &&
1212 idx % ra_range_pages == 0) || stride_detect)
1213 ras->ras_need_increase_window = true;
1214 } else if ((ras->ras_consecutive_requests > 1 || stride_detect)) {
1215 ras->ras_need_increase_window = true;
1218 ras->ras_last_read_end_bytes = pos + count - 1;
1221 void ll_ras_enter(struct file *f, loff_t pos, size_t count)
1223 struct ll_file_data *fd = f->private_data;
1224 struct ll_readahead_state *ras = &fd->fd_ras;
1225 struct inode *inode = file_inode(f);
1226 unsigned long index = pos >> PAGE_SHIFT;
1227 struct ll_sb_info *sbi = ll_i2sbi(inode);
1229 spin_lock(&ras->ras_lock);
1230 ras->ras_requests++;
1231 ras->ras_consecutive_requests++;
1232 ras->ras_need_increase_window = false;
1233 ras->ras_no_miss_check = false;
1235 * On the second access to a file smaller than the tunable
1236 * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1237 * file up to ra_max_pages_per_file. This is simply a best effort
1238 * and only occurs once per open file. Normal RA behavior is reverted
1239 * to for subsequent IO.
1241 if (ras->ras_requests >= 2) {
1243 struct ll_ra_info *ra = &sbi->ll_ra_info;
1245 kms_pages = (i_size_read(inode) + PAGE_SIZE - 1) >>
1248 CDEBUG(D_READA, "kmsp %llu mwp %lu mp %lu\n", kms_pages,
1249 ra->ra_max_read_ahead_whole_pages,
1250 ra->ra_max_pages_per_file);
1253 kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1254 ras->ras_window_start_idx = 0;
1255 ras->ras_next_readahead_idx = index + 1;
1256 ras->ras_window_pages = min(ra->ra_max_pages_per_file,
1257 ra->ra_max_read_ahead_whole_pages);
1258 ras->ras_no_miss_check = true;
1259 GOTO(out_unlock, 0);
1262 ras_detect_read_pattern(ras, sbi, pos, count, false);
1264 spin_unlock(&ras->ras_lock);
1267 static bool index_in_stride_window(struct ll_readahead_state *ras,
1270 loff_t pos = (loff_t)index << PAGE_SHIFT;
1272 if (ras->ras_stride_length == 0 || ras->ras_stride_bytes == 0 ||
1273 ras->ras_stride_bytes == ras->ras_stride_length)
1276 if (pos >= ras->ras_stride_offset) {
1279 div64_u64_rem(pos - ras->ras_stride_offset,
1280 ras->ras_stride_length, &offset);
1281 if (offset < ras->ras_stride_bytes ||
1282 ras->ras_stride_length - offset < PAGE_SIZE)
1284 } else if (ras->ras_stride_offset - pos < PAGE_SIZE) {
1292 * ll_ras_enter() is used to detect read pattern according to pos and count.
1294 * ras_update() is used to detect cache miss and
1295 * reset window or increase window accordingly
1297 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
1298 struct ll_readahead_state *ras, pgoff_t index,
1299 enum ras_update_flags flags, struct cl_io *io)
1301 struct ll_ra_info *ra = &sbi->ll_ra_info;
1302 bool hit = flags & LL_RAS_HIT;
1305 spin_lock(&ras->ras_lock);
1308 CDEBUG(D_READA|D_IOTRACE, DFID " pages at %lu miss.\n",
1309 PFID(ll_inode2fid(inode)), index);
1310 ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
1313 * The readahead window has been expanded to cover whole
1314 * file size, we don't care whether ra miss happen or not.
1315 * Because we will read whole file to page cache even if
1316 * some pages missed.
1318 if (ras->ras_no_miss_check)
1319 GOTO(out_unlock, 0);
1321 if (io && io->ci_rand_read)
1322 GOTO(out_unlock, 0);
1324 if (io && io->ci_seq_read) {
1326 /* to avoid many small read RPC here */
1327 ras->ras_window_pages = sbi->ll_ra_info.ra_range_pages;
1328 ll_ra_stats_inc_sbi(sbi, RA_STAT_MMAP_RANGE_READ);
1333 if (flags & LL_RAS_MMAP) {
1334 unsigned long ra_pages;
1336 ras_detect_cluster_range(ras, sbi, index << PAGE_SHIFT,
1338 ras_detect_read_pattern(ras, sbi, (loff_t)index << PAGE_SHIFT,
1341 /* we did not detect anything but we could prefetch */
1342 if (!ras->ras_need_increase_window &&
1343 ras->ras_window_pages <= sbi->ll_ra_info.ra_range_pages &&
1344 ras->ras_range_requests >= 2) {
1346 ra_pages = max_t(unsigned long,
1347 RA_MIN_MMAP_RANGE_PAGES,
1348 ras->ras_last_range_pages);
1349 if (index < ra_pages / 2)
1352 index -= ra_pages / 2;
1353 ras->ras_window_pages = ra_pages;
1354 ll_ra_stats_inc_sbi(sbi,
1355 RA_STAT_MMAP_RANGE_READ);
1357 ras->ras_window_pages = 0;
1363 if (!hit && ras->ras_window_pages &&
1364 index < ras->ras_next_readahead_idx &&
1365 pos_in_window(index, ras->ras_window_start_idx, 0,
1366 ras->ras_window_pages)) {
1367 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
1368 ras->ras_need_increase_window = false;
1370 if (index_in_stride_window(ras, index) &&
1371 stride_io_mode(ras)) {
1373 * if (index != ras->ras_last_readpage + 1)
1374 * ras->ras_consecutive_pages = 0;
1376 ras_reset(ras, index);
1379 * If stride-RA hit cache miss, the stride
1380 * detector will not be reset to avoid the
1381 * overhead of redetecting read-ahead mode,
1382 * but on the condition that the stride window
1383 * is still intersect with normal sequential
1384 * read-ahead window.
1386 if (ras->ras_window_start_idx < ras->ras_stride_offset)
1387 ras_stride_reset(ras);
1391 * Reset both stride window and normal RA
1394 ras_reset(ras, index);
1395 /* ras->ras_consecutive_pages++; */
1396 ras->ras_consecutive_bytes = 0;
1397 ras_stride_reset(ras);
1398 GOTO(out_unlock, 0);
1403 ras_set_start(ras, index);
1405 if (stride_io_mode(ras)) {
1406 /* Since stride readahead is sentivite to the offset
1407 * of read-ahead, so we use original offset here,
1408 * instead of ras_window_start_idx, which is RPC aligned.
1410 ras->ras_next_readahead_idx = max(index + 1,
1411 ras->ras_next_readahead_idx);
1412 ras->ras_window_start_idx =
1413 max_t(pgoff_t, ras->ras_window_start_idx,
1414 ras->ras_stride_offset >> PAGE_SHIFT);
1416 if (ras->ras_next_readahead_idx < ras->ras_window_start_idx)
1417 ras->ras_next_readahead_idx = ras->ras_window_start_idx;
1419 ras->ras_next_readahead_idx = index + 1;
1422 if (ras->ras_need_increase_window) {
1423 ras_increase_window(inode, ras, ra);
1424 ras->ras_need_increase_window = false;
1429 spin_unlock(&ras->ras_lock);
1432 int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
1434 struct inode *inode = vmpage->mapping->host;
1435 struct ll_inode_info *lli = ll_i2info(inode);
1438 struct cl_page *page;
1439 struct cl_object *clob;
1440 bool redirtied = false;
1441 bool unlocked = false;
1446 LASSERT(PageLocked(vmpage));
1447 LASSERT(!PageWriteback(vmpage));
1449 LASSERT(ll_i2dtexp(inode) != NULL);
1451 env = cl_env_get(&refcheck);
1453 GOTO(out, result = PTR_ERR(env));
1455 clob = ll_i2info(inode)->lli_clob;
1456 LASSERT(clob != NULL);
1458 io = vvp_env_thread_io(env);
1460 io->ci_ignore_layout = 1;
1461 result = cl_io_init(env, io, CIT_MISC, clob);
1463 page = cl_page_find(env, clob, vmpage->index,
1464 vmpage, CPT_CACHEABLE);
1465 if (!IS_ERR(page)) {
1466 lu_ref_add(&page->cp_reference, "writepage",
1468 cl_page_assume(env, io, page);
1469 result = cl_page_flush(env, io, page);
1472 * Re-dirty page on error so it retries write,
1473 * but not in case when IO has actually
1474 * occurred and completed with an error.
1476 if (!PageError(vmpage)) {
1477 redirty_page_for_writepage(wbc, vmpage);
1482 cl_page_disown(env, io, page);
1484 lu_ref_del(&page->cp_reference,
1485 "writepage", current);
1486 cl_page_put(env, page);
1488 result = PTR_ERR(page);
1491 cl_io_fini(env, io);
1493 if (redirtied && wbc->sync_mode == WB_SYNC_ALL) {
1494 loff_t offset = cl_offset(clob, vmpage->index);
1496 /* Flush page failed because the extent is being written out.
1497 * Wait for the write of extent to be finished to avoid
1498 * breaking kernel which assumes ->writepage should mark
1499 * PageWriteback or clean the page. */
1500 result = cl_sync_file_range(inode, offset,
1501 offset + PAGE_SIZE - 1,
1504 /* actually we may have written more than one page.
1505 * decreasing this page because the caller will count
1507 wbc->nr_to_write -= result - 1;
1512 cl_env_put(env, &refcheck);
1517 if (!lli->lli_async_rc)
1518 lli->lli_async_rc = result;
1519 SetPageError(vmpage);
1521 unlock_page(vmpage);
1526 int ll_writepages(struct address_space *mapping, struct writeback_control *wbc)
1528 struct inode *inode = mapping->host;
1531 enum cl_fsync_mode mode;
1532 int range_whole = 0;
1536 if (wbc->range_cyclic) {
1537 start = (loff_t)mapping->writeback_index << PAGE_SHIFT;
1538 end = OBD_OBJECT_EOF;
1540 start = wbc->range_start;
1541 end = wbc->range_end;
1542 if (end == LLONG_MAX) {
1543 end = OBD_OBJECT_EOF;
1544 range_whole = start == 0;
1548 mode = CL_FSYNC_NONE;
1549 if (wbc->sync_mode == WB_SYNC_ALL)
1550 mode = CL_FSYNC_LOCAL;
1552 if (ll_i2info(inode)->lli_clob == NULL)
1555 /* for directio, it would call writepages() to evict cached pages
1556 * inside the IO context of write, which will cause deadlock at
1557 * layout_conf since it waits for active IOs to complete. */
1558 result = cl_sync_file_range(inode, start, end, mode, 1);
1560 wbc->nr_to_write -= result;
1564 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) {
1565 if (end == OBD_OBJECT_EOF)
1566 mapping->writeback_index = 0;
1568 mapping->writeback_index = (end >> PAGE_SHIFT) + 1;
1573 struct ll_cl_context *ll_cl_find(struct inode *inode)
1575 struct ll_inode_info *lli = ll_i2info(inode);
1576 struct ll_cl_context *lcc;
1577 struct ll_cl_context *found = NULL;
1579 read_lock(&lli->lli_lock);
1580 list_for_each_entry(lcc, &lli->lli_lccs, lcc_list) {
1581 if (lcc->lcc_cookie == current) {
1586 read_unlock(&lli->lli_lock);
1591 void ll_cl_add(struct inode *inode, const struct lu_env *env, struct cl_io *io,
1594 struct ll_inode_info *lli = ll_i2info(inode);
1595 struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1597 memset(lcc, 0, sizeof(*lcc));
1598 INIT_LIST_HEAD(&lcc->lcc_list);
1599 lcc->lcc_cookie = current;
1602 lcc->lcc_type = type;
1604 write_lock(&lli->lli_lock);
1605 list_add(&lcc->lcc_list, &lli->lli_lccs);
1606 write_unlock(&lli->lli_lock);
1609 void ll_cl_remove(struct inode *inode, const struct lu_env *env)
1611 struct ll_inode_info *lli = ll_i2info(inode);
1612 struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1614 write_lock(&lli->lli_lock);
1615 list_del_init(&lcc->lcc_list);
1616 write_unlock(&lli->lli_lock);
1619 int ll_io_read_page(const struct lu_env *env, struct cl_io *io,
1620 struct cl_page *page, struct file *file)
1622 struct inode *inode = vvp_object_inode(page->cp_obj);
1623 struct ll_sb_info *sbi = ll_i2sbi(inode);
1624 struct ll_file_data *fd = NULL;
1625 struct ll_readahead_state *ras = NULL;
1626 struct cl_2queue *queue = &io->ci_queue;
1627 struct cl_sync_io *anchor = NULL;
1628 struct vvp_page *vpg;
1629 int rc = 0, rc2 = 0;
1631 struct vvp_io *vio = vvp_env_io(env);
1632 bool mmap = !vio->vui_ra_valid;
1633 pgoff_t ra_start_index = 0;
1634 pgoff_t io_start_index;
1635 pgoff_t io_end_index;
1636 bool unlockpage = true;
1640 fd = file->private_data;
1644 /* PagePrivate2 is set in ll_io_zero_page() to tell us the vmpage
1645 * must not be unlocked after processing.
1647 if (page->cp_vmpage && PagePrivate2(page->cp_vmpage))
1650 vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1651 uptodate = vpg->vpg_defer_uptodate;
1653 if (ll_readahead_enabled(sbi) && !vpg->vpg_ra_updated && ras) {
1654 enum ras_update_flags flags = 0;
1657 flags |= LL_RAS_HIT;
1659 flags |= LL_RAS_MMAP;
1660 ras_update(sbi, inode, ras, vvp_index(vpg), flags, io);
1663 cl_2queue_init(queue);
1665 vpg->vpg_ra_used = 1;
1666 cl_page_export(env, page, 1);
1667 cl_page_disown(env, io, page);
1669 anchor = &vvp_env_info(env)->vti_anchor;
1670 cl_sync_io_init(anchor, 1);
1671 page->cp_sync_io = anchor;
1673 cl_2queue_add(queue, page, true);
1676 /* mmap does not set the ci_rw fields */
1678 io_start_index = cl_index(io->ci_obj, io->u.ci_rw.crw_pos);
1679 io_end_index = cl_index(io->ci_obj, io->u.ci_rw.crw_pos +
1680 io->u.ci_rw.crw_count - 1);
1682 io_start_index = vvp_index(vpg);
1683 io_end_index = vvp_index(vpg);
1686 if (ll_readahead_enabled(sbi) && ras && !io->ci_rand_read) {
1687 pgoff_t skip_index = 0;
1689 if (ras->ras_next_readahead_idx < vvp_index(vpg))
1690 skip_index = vvp_index(vpg);
1691 rc2 = ll_readahead(env, io, &queue->c2_qin, ras,
1692 uptodate, file, skip_index,
1694 CDEBUG(D_READA|D_IOTRACE,
1695 DFID " %d pages read ahead at %lu, triggered by user read at %lu\n",
1696 PFID(ll_inode2fid(inode)), rc2, ra_start_index,
1698 } else if (vvp_index(vpg) == io_start_index &&
1699 io_end_index - io_start_index > 0) {
1700 rc2 = ll_readpages(env, io, &queue->c2_qin, io_start_index + 1,
1702 CDEBUG(D_READA, DFID " %d pages read at %lu\n",
1703 PFID(ll_inode2fid(inode)), rc2, vvp_index(vpg));
1706 if (queue->c2_qin.pl_nr > 0) {
1707 int count = queue->c2_qin.pl_nr;
1708 rc = cl_io_submit_rw(env, io, CRT_READ, queue);
1710 task_io_account_read(PAGE_SIZE * count);
1714 if (anchor != NULL && !cl_page_is_owned(page, io)) { /* have sent */
1715 rc = cl_sync_io_wait(env, anchor, 0);
1717 cl_page_assume(env, io, page);
1718 cl_page_list_del(env, &queue->c2_qout, page);
1720 if (!PageUptodate(cl_page_vmpage(page))) {
1721 /* Failed to read a mirror, discard this page so that
1722 * new page can be created with new mirror.
1724 * TODO: this is not needed after page reinit
1725 * route is implemented */
1726 cl_page_discard(env, io, page);
1729 cl_page_disown(env, io, page);
1732 /* TODO: discard all pages until page reinit route is implemented */
1733 cl_page_list_discard(env, io, &queue->c2_qin);
1735 /* Unlock unsent read pages in case of error. */
1736 cl_page_list_disown(env, io, &queue->c2_qin);
1738 cl_2queue_fini(env, queue);
1744 * Possible return value:
1745 * 0 no async readahead triggered and fast read could not be used.
1746 * 1 no async readahead, but fast read could be used.
1747 * 2 async readahead triggered and fast read could be used too.
1750 static int kickoff_async_readahead(struct file *file, unsigned long pages)
1752 struct ll_readahead_work *lrw;
1753 struct inode *inode = file_inode(file);
1754 struct ll_sb_info *sbi = ll_i2sbi(inode);
1755 struct ll_file_data *fd = file->private_data;
1756 struct ll_readahead_state *ras = &fd->fd_ras;
1757 struct ll_ra_info *ra = &sbi->ll_ra_info;
1758 unsigned long throttle;
1759 pgoff_t start_idx = ras_align(ras, ras->ras_next_readahead_idx);
1760 pgoff_t end_idx = start_idx + pages - 1;
1763 * In case we have a limited max_cached_mb, readahead
1764 * should be stopped if it have run out of all LRU slots.
1766 if (atomic_read(&ra->ra_cur_pages) >= sbi->ll_cache->ccc_lru_max) {
1767 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
1771 throttle = min(ra->ra_async_pages_per_file_threshold,
1772 ra->ra_max_pages_per_file);
1774 * If this is strided i/o or the window is smaller than the
1775 * throttle limit, we do not do async readahead. Otherwise,
1776 * we do async readahead, allowing the user thread to do fast i/o.
1778 if (stride_io_mode(ras) || !throttle ||
1779 ras->ras_window_pages < throttle ||
1780 atomic_read(&ra->ra_async_inflight) > ra->ra_async_max_active)
1783 if ((atomic_read(&ra->ra_cur_pages) + pages) > ra->ra_max_pages)
1786 if (ras->ras_async_last_readpage_idx == start_idx)
1789 /* ll_readahead_work_free() free it */
1792 atomic_inc(&sbi->ll_ra_info.ra_async_inflight);
1793 lrw->lrw_file = get_file(file);
1794 lrw->lrw_start_idx = start_idx;
1795 lrw->lrw_end_idx = end_idx;
1796 lrw->lrw_user_pid = current->pid;
1797 spin_lock(&ras->ras_lock);
1798 ras->ras_next_readahead_idx = end_idx + 1;
1799 ras->ras_async_last_readpage_idx = start_idx;
1800 spin_unlock(&ras->ras_lock);
1801 memcpy(lrw->lrw_jobid, ll_i2info(inode)->lli_jobid,
1802 sizeof(lrw->lrw_jobid));
1803 ll_readahead_work_add(inode, lrw);
1812 * Check if we can issue a readahead RPC, if that is
1813 * the case, we can't do fast IO because we will need
1814 * a cl_io to issue the RPC.
1816 static bool ll_use_fast_io(struct file *file,
1817 struct ll_readahead_state *ras, pgoff_t index)
1819 unsigned long fast_read_pages =
1820 max(RA_REMAIN_WINDOW_MIN, ras->ras_rpc_pages);
1822 loff_t stride_bytes = ras->ras_stride_bytes;
1824 if (stride_io_mode(ras) && stride_bytes) {
1825 skip_pages = (ras->ras_stride_length +
1826 ras->ras_stride_bytes - 1) / stride_bytes;
1827 skip_pages *= fast_read_pages;
1829 skip_pages = fast_read_pages;
1832 if (ras->ras_window_start_idx + ras->ras_window_pages <
1833 ras->ras_next_readahead_idx + skip_pages ||
1834 kickoff_async_readahead(file, fast_read_pages) > 0)
1840 int ll_readpage(struct file *file, struct page *vmpage)
1842 struct inode *inode = file_inode(file);
1843 struct cl_object *clob = ll_i2info(inode)->lli_clob;
1844 struct ll_cl_context *lcc;
1845 const struct lu_env *env = NULL;
1846 struct cl_io *io = NULL;
1847 struct cl_page *page;
1848 struct ll_sb_info *sbi = ll_i2sbi(inode);
1852 lcc = ll_cl_find(inode);
1858 if (io == NULL) { /* fast read */
1859 struct inode *inode = file_inode(file);
1860 struct ll_file_data *fd = file->private_data;
1861 struct ll_readahead_state *ras = &fd->fd_ras;
1862 struct lu_env *local_env = NULL;
1863 struct vvp_page *vpg;
1867 /* TODO: need to verify the layout version to make sure
1868 * the page is not invalid due to layout change. */
1869 page = cl_vmpage_page(vmpage, clob);
1871 unlock_page(vmpage);
1872 ll_ra_stats_inc_sbi(sbi, RA_STAT_FAILED_FAST_READ);
1876 vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1877 if (vpg->vpg_defer_uptodate) {
1878 enum ras_update_flags flags = LL_RAS_HIT;
1880 if (lcc && lcc->lcc_type == LCC_MMAP)
1881 flags |= LL_RAS_MMAP;
1883 /* For fast read, it updates read ahead state only
1884 * if the page is hit in cache because non cache page
1885 * case will be handled by slow read later. */
1886 ras_update(sbi, inode, ras, vvp_index(vpg), flags, io);
1887 /* avoid duplicate ras_update() call */
1888 vpg->vpg_ra_updated = 1;
1890 if (ll_use_fast_io(file, ras, vvp_index(vpg)))
1895 local_env = cl_env_percpu_get();
1899 /* export the page and skip io stack */
1901 vpg->vpg_ra_used = 1;
1902 cl_page_export(env, page, 1);
1904 ll_ra_stats_inc_sbi(sbi, RA_STAT_FAILED_FAST_READ);
1906 /* release page refcount before unlocking the page to ensure
1907 * the object won't be destroyed in the calling path of
1908 * cl_page_put(). Please see comment in ll_releasepage(). */
1909 cl_page_put(env, page);
1910 unlock_page(vmpage);
1912 cl_env_percpu_put(local_env);
1918 * Direct read can fall back to buffered read, but DIO is done
1919 * with lockless i/o, and buffered requires LDLM locking, so in
1920 * this case we must restart without lockless.
1922 if (file->f_flags & O_DIRECT &&
1923 lcc && lcc->lcc_type == LCC_RW &&
1925 unlock_page(vmpage);
1926 io->ci_dio_lock = 1;
1927 io->ci_need_restart = 1;
1931 LASSERT(io->ci_state == CIS_IO_GOING);
1932 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
1933 if (!IS_ERR(page)) {
1934 LASSERT(page->cp_type == CPT_CACHEABLE);
1935 if (likely(!PageUptodate(vmpage))) {
1936 cl_page_assume(env, io, page);
1938 result = ll_io_read_page(env, io, page, file);
1940 /* Page from a non-object file. */
1941 unlock_page(vmpage);
1944 cl_page_put(env, page);
1946 unlock_page(vmpage);
1947 result = PTR_ERR(page);