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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2015, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 * Lustre Lite I/O page cache routines shared by different kernel revs
41 #include <linux/kernel.h>
43 #include <linux/string.h>
44 #include <linux/stat.h>
45 #include <linux/errno.h>
46 #include <linux/unistd.h>
47 #include <linux/writeback.h>
48 #include <asm/uaccess.h>
51 #include <linux/stat.h>
52 #include <asm/uaccess.h>
54 #include <linux/pagemap.h>
55 /* current_is_kswapd() */
56 #include <linux/swap.h>
58 #define DEBUG_SUBSYSTEM S_LLITE
60 #include <obd_cksum.h>
61 #include "llite_internal.h"
62 #include <lustre_compat.h>
64 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which);
67 * Get readahead pages from the filesystem readahead pool of the client for a
70 * /param sbi superblock for filesystem readahead state ll_ra_info
71 * /param ria per-thread readahead state
72 * /param pages number of pages requested for readahead for the thread.
74 * WARNING: This algorithm is used to reduce contention on sbi->ll_lock.
75 * It should work well if the ra_max_pages is much greater than the single
76 * file's read-ahead window, and not too many threads contending for
77 * these readahead pages.
79 * TODO: There may be a 'global sync problem' if many threads are trying
80 * to get an ra budget that is larger than the remaining readahead pages
81 * and reach here at exactly the same time. They will compute /a ret to
82 * consume the remaining pages, but will fail at atomic_add_return() and
83 * get a zero ra window, although there is still ra space remaining. - Jay */
85 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi,
86 struct ra_io_arg *ria,
87 unsigned long pages, unsigned long min)
89 struct ll_ra_info *ra = &sbi->ll_ra_info;
93 /* If read-ahead pages left are less than 1M, do not do read-ahead,
94 * otherwise it will form small read RPC(< 1M), which hurt server
95 * performance a lot. */
96 ret = min(ra->ra_max_pages - atomic_read(&ra->ra_cur_pages),
98 if (ret < 0 || ret < min_t(long, PTLRPC_MAX_BRW_PAGES, pages))
101 if (atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) {
102 atomic_sub(ret, &ra->ra_cur_pages);
108 /* override ra limit for maximum performance */
109 atomic_add(min - ret, &ra->ra_cur_pages);
115 void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
117 struct ll_ra_info *ra = &sbi->ll_ra_info;
118 atomic_sub(len, &ra->ra_cur_pages);
121 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which)
123 LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
124 lprocfs_counter_incr(sbi->ll_ra_stats, which);
127 void ll_ra_stats_inc(struct inode *inode, enum ra_stat which)
129 struct ll_sb_info *sbi = ll_i2sbi(inode);
130 ll_ra_stats_inc_sbi(sbi, which);
133 #define RAS_CDEBUG(ras) \
135 "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu rpc %lu " \
136 "r %lu ri %lu csr %lu sf %lu sp %lu sl %lu\n", \
137 ras->ras_last_readpage, ras->ras_consecutive_requests, \
138 ras->ras_consecutive_pages, ras->ras_window_start, \
139 ras->ras_window_len, ras->ras_next_readahead, \
141 ras->ras_requests, ras->ras_request_index, \
142 ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
143 ras->ras_stride_pages, ras->ras_stride_length)
145 static int index_in_window(unsigned long index, unsigned long point,
146 unsigned long before, unsigned long after)
148 unsigned long start = point - before, end = point + after;
155 return start <= index && index <= end;
158 void ll_ras_enter(struct file *f)
160 struct ll_file_data *fd = LUSTRE_FPRIVATE(f);
161 struct ll_readahead_state *ras = &fd->fd_ras;
163 spin_lock(&ras->ras_lock);
165 ras->ras_request_index = 0;
166 ras->ras_consecutive_requests++;
167 spin_unlock(&ras->ras_lock);
171 * Initiates read-ahead of a page with given index.
173 * \retval +ve: page was already uptodate so it will be skipped
175 * \retval -ve: page wasn't added to \a queue for error;
176 * \retval 0: page was added into \a queue for read ahead.
178 static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io,
179 struct cl_page_list *queue, pgoff_t index)
181 struct cl_object *clob = io->ci_obj;
182 struct inode *inode = vvp_object_inode(clob);
184 struct cl_page *page;
185 struct vvp_page *vpg;
186 enum ra_stat which = _NR_RA_STAT; /* keep gcc happy */
188 const char *msg = NULL;
191 vmpage = grab_cache_page_nowait(inode->i_mapping, index);
192 if (vmpage == NULL) {
193 which = RA_STAT_FAILED_GRAB_PAGE;
194 msg = "g_c_p_n failed";
195 GOTO(out, rc = -EBUSY);
198 /* Check if vmpage was truncated or reclaimed */
199 if (vmpage->mapping != inode->i_mapping) {
200 which = RA_STAT_WRONG_GRAB_PAGE;
201 msg = "g_c_p_n returned invalid page";
202 GOTO(out, rc = -EBUSY);
205 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
207 which = RA_STAT_FAILED_GRAB_PAGE;
208 msg = "cl_page_find failed";
209 GOTO(out, rc = PTR_ERR(page));
212 lu_ref_add(&page->cp_reference, "ra", current);
213 cl_page_assume(env, io, page);
214 vpg = cl2vvp_page(cl_object_page_slice(clob, page));
215 if (!vpg->vpg_defer_uptodate && !PageUptodate(vmpage)) {
216 vpg->vpg_defer_uptodate = 1;
217 vpg->vpg_ra_used = 0;
218 cl_page_list_add(queue, page);
220 /* skip completed pages */
221 cl_page_unassume(env, io, page);
222 /* This page is already uptodate, returning a positive number
223 * to tell the callers about this */
227 lu_ref_del(&page->cp_reference, "ra", current);
228 cl_page_put(env, page);
231 if (vmpage != NULL) {
234 page_cache_release(vmpage);
237 ll_ra_stats_inc(inode, which);
238 CDEBUG(D_READA, "%s\n", msg);
245 #define RIA_DEBUG(ria) \
246 CDEBUG(D_READA, "rs %lu re %lu ro %lu rl %lu rp %lu\n", \
247 ria->ria_start, ria->ria_end, ria->ria_stoff, ria->ria_length,\
250 static inline int stride_io_mode(struct ll_readahead_state *ras)
252 return ras->ras_consecutive_stride_requests > 1;
255 /* The function calculates how much pages will be read in
256 * [off, off + length], in such stride IO area,
257 * stride_offset = st_off, stride_lengh = st_len,
258 * stride_pages = st_pgs
260 * |------------------|*****|------------------|*****|------------|*****|....
263 * |----- st_len -----|
265 * How many pages it should read in such pattern
266 * |-------------------------------------------------------------|
268 * |<------ length ------->|
270 * = |<----->| + |-------------------------------------| + |---|
271 * start_left st_pgs * i end_left
274 stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs,
275 unsigned long off, unsigned long length)
277 __u64 start = off > st_off ? off - st_off : 0;
278 __u64 end = off + length > st_off ? off + length - st_off : 0;
279 unsigned long start_left = 0;
280 unsigned long end_left = 0;
281 unsigned long pg_count;
283 if (st_len == 0 || length == 0 || end == 0)
286 start_left = do_div(start, st_len);
287 if (start_left < st_pgs)
288 start_left = st_pgs - start_left;
292 end_left = do_div(end, st_len);
293 if (end_left > st_pgs)
296 CDEBUG(D_READA, "start "LPU64", end "LPU64" start_left %lu end_left %lu \n",
297 start, end, start_left, end_left);
300 pg_count = end_left - (st_pgs - start_left);
302 pg_count = start_left + st_pgs * (end - start - 1) + end_left;
304 CDEBUG(D_READA, "st_off %lu, st_len %lu st_pgs %lu off %lu length %lu"
305 "pgcount %lu\n", st_off, st_len, st_pgs, off, length, pg_count);
310 static int ria_page_count(struct ra_io_arg *ria)
312 __u64 length = ria->ria_end >= ria->ria_start ?
313 ria->ria_end - ria->ria_start + 1 : 0;
315 return stride_pg_count(ria->ria_stoff, ria->ria_length,
316 ria->ria_pages, ria->ria_start,
320 static unsigned long ras_align(struct ll_readahead_state *ras,
322 unsigned long *remainder)
324 unsigned long rem = index % ras->ras_rpc_size;
325 if (remainder != NULL)
330 /*Check whether the index is in the defined ra-window */
331 static int ras_inside_ra_window(unsigned long idx, struct ra_io_arg *ria)
333 /* If ria_length == ria_pages, it means non-stride I/O mode,
334 * idx should always inside read-ahead window in this case
335 * For stride I/O mode, just check whether the idx is inside
337 return ria->ria_length == 0 || ria->ria_length == ria->ria_pages ||
338 (idx >= ria->ria_stoff && (idx - ria->ria_stoff) %
339 ria->ria_length < ria->ria_pages);
343 ll_read_ahead_pages(const struct lu_env *env, struct cl_io *io,
344 struct cl_page_list *queue, struct ll_readahead_state *ras,
345 struct ra_io_arg *ria)
347 struct cl_read_ahead ra = { 0 };
350 unsigned long ra_end = 0;
353 LASSERT(ria != NULL);
356 stride_ria = ria->ria_length > ria->ria_pages && ria->ria_pages > 0;
357 for (page_idx = ria->ria_start;
358 page_idx <= ria->ria_end && ria->ria_reserved > 0; page_idx++) {
359 if (ras_inside_ra_window(page_idx, ria)) {
360 if (ra.cra_end == 0 || ra.cra_end < page_idx) {
363 cl_read_ahead_release(env, &ra);
365 rc = cl_io_read_ahead(env, io, page_idx, &ra);
369 CDEBUG(D_READA, "idx: %lu, ra: %lu, rpc: %lu\n",
370 page_idx, ra.cra_end, ra.cra_rpc_size);
371 LASSERTF(ra.cra_end >= page_idx,
372 "object: %p, indcies %lu / %lu\n",
373 io->ci_obj, ra.cra_end, page_idx);
374 /* update read ahead RPC size.
375 * NB: it's racy but doesn't matter */
376 if (ras->ras_rpc_size > ra.cra_rpc_size &&
378 ras->ras_rpc_size = ra.cra_rpc_size;
379 /* trim it to align with optimal RPC size */
380 end = ras_align(ras, ria->ria_end + 1, NULL);
381 if (end > 0 && !ria->ria_eof)
382 ria->ria_end = end - 1;
383 if (ria->ria_end < ria->ria_end_min)
384 ria->ria_end = ria->ria_end_min;
385 if (ria->ria_end > ra.cra_end)
386 ria->ria_end = ra.cra_end;
388 if (page_idx > ria->ria_end)
391 /* If the page is inside the read-ahead window */
392 rc = ll_read_ahead_page(env, io, queue, page_idx);
399 } else if (stride_ria) {
400 /* If it is not in the read-ahead window, and it is
401 * read-ahead mode, then check whether it should skip
404 /* FIXME: This assertion only is valid when it is for
405 * forward read-ahead, it will be fixed when backward
406 * read-ahead is implemented */
407 LASSERTF(page_idx >= ria->ria_stoff,
408 "Invalid page_idx %lu rs %lu re %lu ro %lu "
409 "rl %lu rp %lu\n", page_idx,
410 ria->ria_start, ria->ria_end, ria->ria_stoff,
411 ria->ria_length, ria->ria_pages);
412 offset = page_idx - ria->ria_stoff;
413 offset = offset % (ria->ria_length);
414 if (offset > ria->ria_pages) {
415 page_idx += ria->ria_length - offset;
416 CDEBUG(D_READA, "i %lu skip %lu \n", page_idx,
417 ria->ria_length - offset);
423 cl_read_ahead_release(env, &ra);
428 static int ll_readahead(const struct lu_env *env, struct cl_io *io,
429 struct cl_page_list *queue,
430 struct ll_readahead_state *ras, bool hit)
432 struct vvp_io *vio = vvp_env_io(env);
433 struct ll_thread_info *lti = ll_env_info(env);
434 struct cl_attr *attr = vvp_env_thread_attr(env);
435 unsigned long len, mlen = 0;
436 pgoff_t ra_end, start = 0, end = 0;
438 struct ra_io_arg *ria = <i->lti_ria;
439 struct cl_object *clob;
445 inode = vvp_object_inode(clob);
447 memset(ria, 0, sizeof *ria);
449 cl_object_attr_lock(clob);
450 ret = cl_object_attr_get(env, clob, attr);
451 cl_object_attr_unlock(clob);
457 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
461 spin_lock(&ras->ras_lock);
464 * Note: other thread might rollback the ras_next_readahead,
465 * if it can not get the full size of prepared pages, see the
466 * end of this function. For stride read ahead, it needs to
467 * make sure the offset is no less than ras_stride_offset,
468 * so that stride read ahead can work correctly.
470 if (stride_io_mode(ras))
471 start = max(ras->ras_next_readahead, ras->ras_stride_offset);
473 start = ras->ras_next_readahead;
475 if (ras->ras_window_len > 0)
476 end = ras->ras_window_start + ras->ras_window_len - 1;
478 /* Enlarge the RA window to encompass the full read */
479 if (vio->vui_ra_valid &&
480 end < vio->vui_ra_start + vio->vui_ra_count - 1)
481 end = vio->vui_ra_start + vio->vui_ra_count - 1;
484 unsigned long end_index;
486 /* Truncate RA window to end of file */
487 end_index = (unsigned long)((kms - 1) >> PAGE_CACHE_SHIFT);
488 if (end_index <= end) {
493 ras->ras_next_readahead = max(end, end + 1);
496 ria->ria_start = start;
498 /* If stride I/O mode is detected, get stride window*/
499 if (stride_io_mode(ras)) {
500 ria->ria_stoff = ras->ras_stride_offset;
501 ria->ria_length = ras->ras_stride_length;
502 ria->ria_pages = ras->ras_stride_pages;
504 spin_unlock(&ras->ras_lock);
507 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
510 len = ria_page_count(ria);
512 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
516 CDEBUG(D_READA, DFID": ria: %lu/%lu, bead: %lu/%lu, hit: %d\n",
517 PFID(lu_object_fid(&clob->co_lu)),
518 ria->ria_start, ria->ria_end,
519 vio->vui_ra_valid ? vio->vui_ra_start : 0,
520 vio->vui_ra_valid ? vio->vui_ra_count : 0,
523 /* at least to extend the readahead window to cover current read */
524 if (!hit && vio->vui_ra_valid &&
525 vio->vui_ra_start + vio->vui_ra_count > ria->ria_start) {
526 unsigned long remainder;
528 /* to the end of current read window. */
529 mlen = vio->vui_ra_start + vio->vui_ra_count - ria->ria_start;
530 /* trim to RPC boundary */
531 ras_align(ras, ria->ria_start, &remainder);
532 mlen = min(mlen, ras->ras_rpc_size - remainder);
533 ria->ria_end_min = ria->ria_start + mlen;
536 ria->ria_reserved = ll_ra_count_get(ll_i2sbi(inode), ria, len, mlen);
537 if (ria->ria_reserved < len)
538 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
540 CDEBUG(D_READA, "reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
541 ria->ria_reserved, len, mlen,
542 atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
543 ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
545 ra_end = ll_read_ahead_pages(env, io, queue, ras, ria);
547 if (ria->ria_reserved != 0)
548 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
550 if (ra_end == end && ra_end == (kms >> PAGE_CACHE_SHIFT))
551 ll_ra_stats_inc(inode, RA_STAT_EOF);
553 /* if we didn't get to the end of the region we reserved from
554 * the ras we need to go back and update the ras so that the
555 * next read-ahead tries from where we left off. we only do so
556 * if the region we failed to issue read-ahead on is still ahead
557 * of the app and behind the next index to start read-ahead from */
558 CDEBUG(D_READA, "ra_end = %lu end = %lu stride end = %lu pages = %d\n",
559 ra_end, end, ria->ria_end, ret);
561 if (ra_end > 0 && ra_end != end) {
562 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
563 spin_lock(&ras->ras_lock);
564 if (ra_end <= ras->ras_next_readahead &&
565 index_in_window(ra_end, ras->ras_window_start, 0,
566 ras->ras_window_len)) {
567 ras->ras_next_readahead = ra_end + 1;
570 spin_unlock(&ras->ras_lock);
576 static void ras_set_start(struct inode *inode, struct ll_readahead_state *ras,
579 ras->ras_window_start = ras_align(ras, index, NULL);
582 /* called with the ras_lock held or from places where it doesn't matter */
583 static void ras_reset(struct inode *inode, struct ll_readahead_state *ras,
586 ras->ras_last_readpage = index;
587 ras->ras_consecutive_requests = 0;
588 ras->ras_consecutive_pages = 0;
589 ras->ras_window_len = 0;
590 ras_set_start(inode, ras, index);
591 ras->ras_next_readahead = max(ras->ras_window_start, index + 1);
596 /* called with the ras_lock held or from places where it doesn't matter */
597 static void ras_stride_reset(struct ll_readahead_state *ras)
599 ras->ras_consecutive_stride_requests = 0;
600 ras->ras_stride_length = 0;
601 ras->ras_stride_pages = 0;
605 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
607 spin_lock_init(&ras->ras_lock);
608 ras->ras_rpc_size = PTLRPC_MAX_BRW_PAGES;
609 ras_reset(inode, ras, 0);
610 ras->ras_requests = 0;
614 * Check whether the read request is in the stride window.
615 * If it is in the stride window, return 1, otherwise return 0.
617 static int index_in_stride_window(struct ll_readahead_state *ras,
620 unsigned long stride_gap;
622 if (ras->ras_stride_length == 0 || ras->ras_stride_pages == 0 ||
623 ras->ras_stride_pages == ras->ras_stride_length)
626 stride_gap = index - ras->ras_last_readpage - 1;
628 /* If it is contiguous read */
630 return ras->ras_consecutive_pages + 1 <= ras->ras_stride_pages;
632 /* Otherwise check the stride by itself */
633 return (ras->ras_stride_length - ras->ras_stride_pages) == stride_gap &&
634 ras->ras_consecutive_pages == ras->ras_stride_pages;
637 static void ras_update_stride_detector(struct ll_readahead_state *ras,
640 unsigned long stride_gap = index - ras->ras_last_readpage - 1;
642 if (!stride_io_mode(ras) && (stride_gap != 0 ||
643 ras->ras_consecutive_stride_requests == 0)) {
644 ras->ras_stride_pages = ras->ras_consecutive_pages;
645 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
647 LASSERT(ras->ras_request_index == 0);
648 LASSERT(ras->ras_consecutive_stride_requests == 0);
650 if (index <= ras->ras_last_readpage) {
651 /*Reset stride window for forward read*/
652 ras_stride_reset(ras);
656 ras->ras_stride_pages = ras->ras_consecutive_pages;
657 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
664 stride_page_count(struct ll_readahead_state *ras, unsigned long len)
666 return stride_pg_count(ras->ras_stride_offset, ras->ras_stride_length,
667 ras->ras_stride_pages, ras->ras_stride_offset,
671 /* Stride Read-ahead window will be increased inc_len according to
672 * stride I/O pattern */
673 static void ras_stride_increase_window(struct ll_readahead_state *ras,
674 struct ll_ra_info *ra,
675 unsigned long inc_len)
677 unsigned long left, step, window_len;
678 unsigned long stride_len;
680 LASSERT(ras->ras_stride_length > 0);
681 LASSERTF(ras->ras_window_start + ras->ras_window_len
682 >= ras->ras_stride_offset, "window_start %lu, window_len %lu"
683 " stride_offset %lu\n", ras->ras_window_start,
684 ras->ras_window_len, ras->ras_stride_offset);
686 stride_len = ras->ras_window_start + ras->ras_window_len -
687 ras->ras_stride_offset;
689 left = stride_len % ras->ras_stride_length;
690 window_len = ras->ras_window_len - left;
692 if (left < ras->ras_stride_pages)
695 left = ras->ras_stride_pages + inc_len;
697 LASSERT(ras->ras_stride_pages != 0);
699 step = left / ras->ras_stride_pages;
700 left %= ras->ras_stride_pages;
702 window_len += step * ras->ras_stride_length + left;
704 if (stride_page_count(ras, window_len) <= ra->ra_max_pages_per_file)
705 ras->ras_window_len = window_len;
710 static void ras_increase_window(struct inode *inode,
711 struct ll_readahead_state *ras,
712 struct ll_ra_info *ra)
714 /* The stretch of ra-window should be aligned with max rpc_size
715 * but current clio architecture does not support retrieve such
716 * information from lower layer. FIXME later
718 if (stride_io_mode(ras)) {
719 ras_stride_increase_window(ras, ra, ras->ras_rpc_size);
723 wlen = min(ras->ras_window_len + ras->ras_rpc_size,
724 ra->ra_max_pages_per_file);
725 ras->ras_window_len = ras_align(ras, wlen, NULL);
729 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
730 struct ll_readahead_state *ras, unsigned long index,
731 enum ras_update_flags flags)
733 struct ll_ra_info *ra = &sbi->ll_ra_info;
734 bool hit = flags & LL_RAS_HIT;
735 int zero = 0, stride_detect = 0, ra_miss = 0;
738 spin_lock(&ras->ras_lock);
740 ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
742 /* reset the read-ahead window in two cases. First when the app seeks
743 * or reads to some other part of the file. Secondly if we get a
744 * read-ahead miss that we think we've previously issued. This can
745 * be a symptom of there being so many read-ahead pages that the VM is
746 * reclaiming it before we get to it. */
747 if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
749 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
750 } else if (!hit && ras->ras_window_len &&
751 index < ras->ras_next_readahead &&
752 index_in_window(index, ras->ras_window_start, 0,
753 ras->ras_window_len)) {
755 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
758 /* On the second access to a file smaller than the tunable
759 * ra_max_read_ahead_whole_pages trigger RA on all pages in the
760 * file up to ra_max_pages_per_file. This is simply a best effort
761 * and only occurs once per open file. Normal RA behavior is reverted
762 * to for subsequent IO. The mmap case does not increment
763 * ras_requests and thus can never trigger this behavior. */
764 if (ras->ras_requests >= 2 && !ras->ras_request_index) {
767 kms_pages = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >>
770 CDEBUG(D_READA, "kmsp "LPU64" mwp %lu mp %lu\n", kms_pages,
771 ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages_per_file);
774 kms_pages <= ra->ra_max_read_ahead_whole_pages) {
775 ras->ras_window_start = 0;
776 ras->ras_next_readahead = index + 1;
777 ras->ras_window_len = min(ra->ra_max_pages_per_file,
778 ra->ra_max_read_ahead_whole_pages);
783 /* check whether it is in stride I/O mode*/
784 if (!index_in_stride_window(ras, index)) {
785 if (ras->ras_consecutive_stride_requests == 0 &&
786 ras->ras_request_index == 0) {
787 ras_update_stride_detector(ras, index);
788 ras->ras_consecutive_stride_requests++;
790 ras_stride_reset(ras);
792 ras_reset(inode, ras, index);
793 ras->ras_consecutive_pages++;
796 ras->ras_consecutive_pages = 0;
797 ras->ras_consecutive_requests = 0;
798 if (++ras->ras_consecutive_stride_requests > 1)
804 if (index_in_stride_window(ras, index) &&
805 stride_io_mode(ras)) {
806 /*If stride-RA hit cache miss, the stride dector
807 *will not be reset to avoid the overhead of
808 *redetecting read-ahead mode */
809 if (index != ras->ras_last_readpage + 1)
810 ras->ras_consecutive_pages = 0;
811 ras_reset(inode, ras, index);
814 /* Reset both stride window and normal RA
816 ras_reset(inode, ras, index);
817 ras->ras_consecutive_pages++;
818 ras_stride_reset(ras);
821 } else if (stride_io_mode(ras)) {
822 /* If this is contiguous read but in stride I/O mode
823 * currently, check whether stride step still is valid,
824 * if invalid, it will reset the stride ra window*/
825 if (!index_in_stride_window(ras, index)) {
826 /* Shrink stride read-ahead window to be zero */
827 ras_stride_reset(ras);
828 ras->ras_window_len = 0;
829 ras->ras_next_readahead = index;
833 ras->ras_consecutive_pages++;
834 ras->ras_last_readpage = index;
835 ras_set_start(inode, ras, index);
837 if (stride_io_mode(ras)) {
838 /* Since stride readahead is sentivite to the offset
839 * of read-ahead, so we use original offset here,
840 * instead of ras_window_start, which is RPC aligned */
841 ras->ras_next_readahead = max(index, ras->ras_next_readahead);
842 ras->ras_window_start = max(ras->ras_stride_offset,
843 ras->ras_window_start);
845 if (ras->ras_next_readahead < ras->ras_window_start)
846 ras->ras_next_readahead = ras->ras_window_start;
848 ras->ras_next_readahead = index + 1;
852 /* Trigger RA in the mmap case where ras_consecutive_requests
853 * is not incremented and thus can't be used to trigger RA */
854 if (ras->ras_consecutive_pages >= 4 && flags & LL_RAS_MMAP) {
855 ras_increase_window(inode, ras, ra);
856 /* reset consecutive pages so that the readahead window can
858 ras->ras_consecutive_pages = 0;
862 /* Initially reset the stride window offset to next_readahead*/
863 if (ras->ras_consecutive_stride_requests == 2 && stride_detect) {
865 * Once stride IO mode is detected, next_readahead should be
866 * reset to make sure next_readahead > stride offset
868 ras->ras_next_readahead = max(index, ras->ras_next_readahead);
869 ras->ras_stride_offset = index;
870 ras->ras_window_start = max(index, ras->ras_window_start);
873 /* The initial ras_window_len is set to the request size. To avoid
874 * uselessly reading and discarding pages for random IO the window is
875 * only increased once per consecutive request received. */
876 if ((ras->ras_consecutive_requests > 1 || stride_detect) &&
877 !ras->ras_request_index)
878 ras_increase_window(inode, ras, ra);
882 ras->ras_request_index++;
883 spin_unlock(&ras->ras_lock);
887 int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
889 struct inode *inode = vmpage->mapping->host;
890 struct ll_inode_info *lli = ll_i2info(inode);
893 struct cl_page *page;
894 struct cl_object *clob;
895 bool redirtied = false;
896 bool unlocked = false;
901 LASSERT(PageLocked(vmpage));
902 LASSERT(!PageWriteback(vmpage));
904 LASSERT(ll_i2dtexp(inode) != NULL);
906 env = cl_env_get(&refcheck);
908 GOTO(out, result = PTR_ERR(env));
910 clob = ll_i2info(inode)->lli_clob;
911 LASSERT(clob != NULL);
913 io = vvp_env_thread_io(env);
915 io->ci_ignore_layout = 1;
916 result = cl_io_init(env, io, CIT_MISC, clob);
918 page = cl_page_find(env, clob, vmpage->index,
919 vmpage, CPT_CACHEABLE);
921 lu_ref_add(&page->cp_reference, "writepage",
923 cl_page_assume(env, io, page);
924 result = cl_page_flush(env, io, page);
927 * Re-dirty page on error so it retries write,
928 * but not in case when IO has actually
929 * occurred and completed with an error.
931 if (!PageError(vmpage)) {
932 redirty_page_for_writepage(wbc, vmpage);
937 cl_page_disown(env, io, page);
939 lu_ref_del(&page->cp_reference,
940 "writepage", current);
941 cl_page_put(env, page);
943 result = PTR_ERR(page);
948 if (redirtied && wbc->sync_mode == WB_SYNC_ALL) {
949 loff_t offset = cl_offset(clob, vmpage->index);
951 /* Flush page failed because the extent is being written out.
952 * Wait for the write of extent to be finished to avoid
953 * breaking kernel which assumes ->writepage should mark
954 * PageWriteback or clean the page. */
955 result = cl_sync_file_range(inode, offset,
956 offset + PAGE_CACHE_SIZE - 1,
959 /* actually we may have written more than one page.
960 * decreasing this page because the caller will count
962 wbc->nr_to_write -= result - 1;
967 cl_env_put(env, &refcheck);
972 if (!lli->lli_async_rc)
973 lli->lli_async_rc = result;
974 SetPageError(vmpage);
981 int ll_writepages(struct address_space *mapping, struct writeback_control *wbc)
983 struct inode *inode = mapping->host;
984 struct ll_sb_info *sbi = ll_i2sbi(inode);
987 enum cl_fsync_mode mode;
990 int ignore_layout = 0;
993 if (wbc->range_cyclic) {
994 start = mapping->writeback_index << PAGE_CACHE_SHIFT;
995 end = OBD_OBJECT_EOF;
997 start = wbc->range_start;
998 end = wbc->range_end;
999 if (end == LLONG_MAX) {
1000 end = OBD_OBJECT_EOF;
1001 range_whole = start == 0;
1005 mode = CL_FSYNC_NONE;
1006 if (wbc->sync_mode == WB_SYNC_ALL)
1007 mode = CL_FSYNC_LOCAL;
1009 if (sbi->ll_umounting)
1010 /* if the mountpoint is being umounted, all pages have to be
1011 * evicted to avoid hitting LBUG when truncate_inode_pages()
1012 * is called later on. */
1015 if (ll_i2info(inode)->lli_clob == NULL)
1018 result = cl_sync_file_range(inode, start, end, mode, ignore_layout);
1020 wbc->nr_to_write -= result;
1024 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) {
1025 if (end == OBD_OBJECT_EOF)
1026 mapping->writeback_index = 0;
1028 mapping->writeback_index = (end >> PAGE_CACHE_SHIFT) +1;
1033 struct ll_cl_context *ll_cl_find(struct file *file)
1035 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1036 struct ll_cl_context *lcc;
1037 struct ll_cl_context *found = NULL;
1039 read_lock(&fd->fd_lock);
1040 list_for_each_entry(lcc, &fd->fd_lccs, lcc_list) {
1041 if (lcc->lcc_cookie == current) {
1046 read_unlock(&fd->fd_lock);
1051 void ll_cl_add(struct file *file, const struct lu_env *env, struct cl_io *io,
1054 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1055 struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1057 memset(lcc, 0, sizeof(*lcc));
1058 INIT_LIST_HEAD(&lcc->lcc_list);
1059 lcc->lcc_cookie = current;
1062 lcc->lcc_type = type;
1064 write_lock(&fd->fd_lock);
1065 list_add(&lcc->lcc_list, &fd->fd_lccs);
1066 write_unlock(&fd->fd_lock);
1069 void ll_cl_remove(struct file *file, const struct lu_env *env)
1071 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1072 struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1074 write_lock(&fd->fd_lock);
1075 list_del_init(&lcc->lcc_list);
1076 write_unlock(&fd->fd_lock);
1079 static int ll_io_read_page(const struct lu_env *env, struct cl_io *io,
1080 struct cl_page *page, struct file *file)
1082 struct inode *inode = vvp_object_inode(page->cp_obj);
1083 struct ll_sb_info *sbi = ll_i2sbi(inode);
1084 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1085 struct ll_readahead_state *ras = &fd->fd_ras;
1086 struct cl_2queue *queue = &io->ci_queue;
1087 struct vvp_page *vpg;
1092 vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1093 uptodate = vpg->vpg_defer_uptodate;
1095 if (sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
1096 sbi->ll_ra_info.ra_max_pages > 0 &&
1097 !vpg->vpg_ra_updated) {
1098 struct vvp_io *vio = vvp_env_io(env);
1099 enum ras_update_flags flags = 0;
1102 flags |= LL_RAS_HIT;
1103 if (!vio->vui_ra_valid)
1104 flags |= LL_RAS_MMAP;
1105 ras_update(sbi, inode, ras, vvp_index(vpg), flags);
1108 cl_2queue_init(queue);
1110 vpg->vpg_ra_used = 1;
1111 cl_page_export(env, page, 1);
1112 cl_page_disown(env, io, page);
1114 cl_2queue_add(queue, page);
1117 if (sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
1118 sbi->ll_ra_info.ra_max_pages > 0) {
1121 rc2 = ll_readahead(env, io, &queue->c2_qin, ras,
1123 CDEBUG(D_READA, DFID "%d pages read ahead at %lu\n",
1124 PFID(ll_inode2fid(inode)), rc2, vvp_index(vpg));
1127 if (queue->c2_qin.pl_nr > 0)
1128 rc = cl_io_submit_rw(env, io, CRT_READ, queue);
1131 * Unlock unsent pages in case of error.
1133 cl_page_list_disown(env, io, &queue->c2_qin);
1134 cl_2queue_fini(env, queue);
1139 int ll_readpage(struct file *file, struct page *vmpage)
1141 struct inode *inode = file->f_path.dentry->d_inode;
1142 struct cl_object *clob = ll_i2info(inode)->lli_clob;
1143 struct ll_cl_context *lcc;
1144 const struct lu_env *env;
1146 struct cl_page *page;
1150 lcc = ll_cl_find(file);
1152 unlock_page(vmpage);
1158 if (io == NULL) { /* fast read */
1159 struct inode *inode = file_inode(file);
1160 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1161 struct ll_readahead_state *ras = &fd->fd_ras;
1162 struct vvp_page *vpg;
1166 /* TODO: need to verify the layout version to make sure
1167 * the page is not invalid due to layout change. */
1168 page = cl_vmpage_page(vmpage, clob);
1170 unlock_page(vmpage);
1174 vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1175 if (vpg->vpg_defer_uptodate) {
1176 enum ras_update_flags flags = LL_RAS_HIT;
1178 if (lcc->lcc_type == LCC_MMAP)
1179 flags |= LL_RAS_MMAP;
1181 /* For fast read, it updates read ahead state only
1182 * if the page is hit in cache because non cache page
1183 * case will be handled by slow read later. */
1184 ras_update(ll_i2sbi(inode), inode, ras, vvp_index(vpg),
1186 /* avoid duplicate ras_update() call */
1187 vpg->vpg_ra_updated = 1;
1189 /* Check if we can issue a readahead RPC, if that is
1190 * the case, we can't do fast IO because we will need
1191 * a cl_io to issue the RPC. */
1192 if (ras->ras_window_start + ras->ras_window_len <
1193 ras->ras_next_readahead + PTLRPC_MAX_BRW_PAGES) {
1194 /* export the page and skip io stack */
1195 vpg->vpg_ra_used = 1;
1196 cl_page_export(env, page, 1);
1201 unlock_page(vmpage);
1202 cl_page_put(env, page);
1206 LASSERT(io->ci_state == CIS_IO_GOING);
1207 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
1208 if (!IS_ERR(page)) {
1209 LASSERT(page->cp_type == CPT_CACHEABLE);
1210 if (likely(!PageUptodate(vmpage))) {
1211 cl_page_assume(env, io, page);
1212 result = ll_io_read_page(env, io, page, file);
1214 /* Page from a non-object file. */
1215 unlock_page(vmpage);
1218 cl_page_put(env, page);
1220 unlock_page(vmpage);
1221 result = PTR_ERR(page);
1226 int ll_page_sync_io(const struct lu_env *env, struct cl_io *io,
1227 struct cl_page *page, enum cl_req_type crt)
1229 struct cl_2queue *queue;
1232 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
1234 queue = &io->ci_queue;
1235 cl_2queue_init_page(queue, page);
1237 result = cl_io_submit_sync(env, io, crt, queue, 0);
1238 LASSERT(cl_page_is_owned(page, io));
1240 if (crt == CRT_READ)
1242 * in CRT_WRITE case page is left locked even in case of
1245 cl_page_list_disown(env, io, &queue->c2_qin);
1246 cl_2queue_fini(env, queue);