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, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
34 * Lustre Lite I/O page cache routines shared by different kernel revs
37 #include <linux/kernel.h>
39 #include <linux/string.h>
40 #include <linux/stat.h>
41 #include <linux/errno.h>
42 #include <linux/unistd.h>
43 #include <linux/writeback.h>
44 #include <asm/uaccess.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>
54 #define DEBUG_SUBSYSTEM S_LLITE
56 #include <obd_cksum.h>
57 #include "llite_internal.h"
58 #include <lustre_compat.h>
60 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which);
63 * Get readahead pages from the filesystem readahead pool of the client for a
66 * /param sbi superblock for filesystem readahead state ll_ra_info
67 * /param ria per-thread readahead state
68 * /param pages number of pages requested for readahead for the thread.
70 * WARNING: This algorithm is used to reduce contention on sbi->ll_lock.
71 * It should work well if the ra_max_pages is much greater than the single
72 * file's read-ahead window, and not too many threads contending for
73 * these readahead pages.
75 * TODO: There may be a 'global sync problem' if many threads are trying
76 * to get an ra budget that is larger than the remaining readahead pages
77 * and reach here at exactly the same time. They will compute /a ret to
78 * consume the remaining pages, but will fail at atomic_add_return() and
79 * get a zero ra window, although there is still ra space remaining. - Jay */
81 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi,
82 struct ra_io_arg *ria,
83 unsigned long pages, unsigned long min)
85 struct ll_ra_info *ra = &sbi->ll_ra_info;
89 /* If read-ahead pages left are less than 1M, do not do read-ahead,
90 * otherwise it will form small read RPC(< 1M), which hurt server
91 * performance a lot. */
92 ret = min(ra->ra_max_pages - atomic_read(&ra->ra_cur_pages),
94 if (ret < 0 || ret < min_t(long, PTLRPC_MAX_BRW_PAGES, pages))
97 if (atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) {
98 atomic_sub(ret, &ra->ra_cur_pages);
104 /* override ra limit for maximum performance */
105 atomic_add(min - ret, &ra->ra_cur_pages);
111 void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
113 struct ll_ra_info *ra = &sbi->ll_ra_info;
114 atomic_sub(len, &ra->ra_cur_pages);
117 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which)
119 LASSERTF(which < _NR_RA_STAT, "which: %u\n", which);
120 lprocfs_counter_incr(sbi->ll_ra_stats, which);
123 void ll_ra_stats_inc(struct inode *inode, enum ra_stat which)
125 struct ll_sb_info *sbi = ll_i2sbi(inode);
126 ll_ra_stats_inc_sbi(sbi, which);
129 #define RAS_CDEBUG(ras) \
131 "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu rpc %lu " \
132 "r %lu ri %lu csr %lu sf %lu sp %lu sl %lu\n", \
133 ras->ras_last_readpage, ras->ras_consecutive_requests, \
134 ras->ras_consecutive_pages, ras->ras_window_start, \
135 ras->ras_window_len, ras->ras_next_readahead, \
137 ras->ras_requests, ras->ras_request_index, \
138 ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
139 ras->ras_stride_pages, ras->ras_stride_length)
141 static int index_in_window(unsigned long index, unsigned long point,
142 unsigned long before, unsigned long after)
144 unsigned long start = point - before, end = point + after;
151 return start <= index && index <= end;
154 void ll_ras_enter(struct file *f)
156 struct ll_file_data *fd = LUSTRE_FPRIVATE(f);
157 struct ll_readahead_state *ras = &fd->fd_ras;
159 spin_lock(&ras->ras_lock);
161 ras->ras_request_index = 0;
162 ras->ras_consecutive_requests++;
163 spin_unlock(&ras->ras_lock);
167 * Initiates read-ahead of a page with given index.
169 * \retval +ve: page was already uptodate so it will be skipped
171 * \retval -ve: page wasn't added to \a queue for error;
172 * \retval 0: page was added into \a queue for read ahead.
174 static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io,
175 struct cl_page_list *queue, pgoff_t index)
177 struct cl_object *clob = io->ci_obj;
178 struct inode *inode = vvp_object_inode(clob);
180 struct cl_page *page;
181 struct vvp_page *vpg;
182 enum ra_stat which = _NR_RA_STAT; /* keep gcc happy */
184 const char *msg = NULL;
187 vmpage = grab_cache_page_nowait(inode->i_mapping, index);
188 if (vmpage == NULL) {
189 which = RA_STAT_FAILED_GRAB_PAGE;
190 msg = "g_c_p_n failed";
191 GOTO(out, rc = -EBUSY);
194 /* Check if vmpage was truncated or reclaimed */
195 if (vmpage->mapping != inode->i_mapping) {
196 which = RA_STAT_WRONG_GRAB_PAGE;
197 msg = "g_c_p_n returned invalid page";
198 GOTO(out, rc = -EBUSY);
201 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
203 which = RA_STAT_FAILED_GRAB_PAGE;
204 msg = "cl_page_find failed";
205 GOTO(out, rc = PTR_ERR(page));
208 lu_ref_add(&page->cp_reference, "ra", current);
209 cl_page_assume(env, io, page);
210 vpg = cl2vvp_page(cl_object_page_slice(clob, page));
211 if (!vpg->vpg_defer_uptodate && !PageUptodate(vmpage)) {
212 vpg->vpg_defer_uptodate = 1;
213 vpg->vpg_ra_used = 0;
214 cl_page_list_add(queue, page);
216 /* skip completed pages */
217 cl_page_unassume(env, io, page);
218 /* This page is already uptodate, returning a positive number
219 * to tell the callers about this */
223 lu_ref_del(&page->cp_reference, "ra", current);
224 cl_page_put(env, page);
227 if (vmpage != NULL) {
233 ll_ra_stats_inc(inode, which);
234 CDEBUG(D_READA, "%s\n", msg);
241 #define RIA_DEBUG(ria) \
242 CDEBUG(D_READA, "rs %lu re %lu ro %lu rl %lu rp %lu\n", \
243 ria->ria_start, ria->ria_end, ria->ria_stoff, ria->ria_length,\
246 static inline int stride_io_mode(struct ll_readahead_state *ras)
248 return ras->ras_consecutive_stride_requests > 1;
251 /* The function calculates how much pages will be read in
252 * [off, off + length], in such stride IO area,
253 * stride_offset = st_off, stride_lengh = st_len,
254 * stride_pages = st_pgs
256 * |------------------|*****|------------------|*****|------------|*****|....
259 * |----- st_len -----|
261 * How many pages it should read in such pattern
262 * |-------------------------------------------------------------|
264 * |<------ length ------->|
266 * = |<----->| + |-------------------------------------| + |---|
267 * start_left st_pgs * i end_left
270 stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs,
271 unsigned long off, unsigned long length)
273 __u64 start = off > st_off ? off - st_off : 0;
274 __u64 end = off + length > st_off ? off + length - st_off : 0;
275 unsigned long start_left = 0;
276 unsigned long end_left = 0;
277 unsigned long pg_count;
279 if (st_len == 0 || length == 0 || end == 0)
282 start_left = do_div(start, st_len);
283 if (start_left < st_pgs)
284 start_left = st_pgs - start_left;
288 end_left = do_div(end, st_len);
289 if (end_left > st_pgs)
292 CDEBUG(D_READA, "start %llu, end %llu start_left %lu end_left %lu\n",
293 start, end, start_left, end_left);
296 pg_count = end_left - (st_pgs - start_left);
298 pg_count = start_left + st_pgs * (end - start - 1) + end_left;
300 CDEBUG(D_READA, "st_off %lu, st_len %lu st_pgs %lu off %lu length %lu"
301 "pgcount %lu\n", st_off, st_len, st_pgs, off, length, pg_count);
306 static int ria_page_count(struct ra_io_arg *ria)
308 __u64 length = ria->ria_end >= ria->ria_start ?
309 ria->ria_end - ria->ria_start + 1 : 0;
311 return stride_pg_count(ria->ria_stoff, ria->ria_length,
312 ria->ria_pages, ria->ria_start,
316 static unsigned long ras_align(struct ll_readahead_state *ras,
318 unsigned long *remainder)
320 unsigned long rem = index % ras->ras_rpc_size;
321 if (remainder != NULL)
326 /*Check whether the index is in the defined ra-window */
327 static int ras_inside_ra_window(unsigned long idx, struct ra_io_arg *ria)
329 /* If ria_length == ria_pages, it means non-stride I/O mode,
330 * idx should always inside read-ahead window in this case
331 * For stride I/O mode, just check whether the idx is inside
333 return ria->ria_length == 0 || ria->ria_length == ria->ria_pages ||
334 (idx >= ria->ria_stoff && (idx - ria->ria_stoff) %
335 ria->ria_length < ria->ria_pages);
339 ll_read_ahead_pages(const struct lu_env *env, struct cl_io *io,
340 struct cl_page_list *queue, struct ll_readahead_state *ras,
341 struct ra_io_arg *ria, pgoff_t *ra_end)
343 struct cl_read_ahead ra = { 0 };
344 int rc = 0, count = 0;
348 LASSERT(ria != NULL);
351 stride_ria = ria->ria_length > ria->ria_pages && ria->ria_pages > 0;
352 for (page_idx = ria->ria_start;
353 page_idx <= ria->ria_end && ria->ria_reserved > 0; page_idx++) {
354 if (ras_inside_ra_window(page_idx, ria)) {
355 if (ra.cra_end == 0 || ra.cra_end < page_idx) {
358 cl_read_ahead_release(env, &ra);
360 rc = cl_io_read_ahead(env, io, page_idx, &ra);
364 CDEBUG(D_READA, "idx: %lu, ra: %lu, rpc: %lu\n",
365 page_idx, ra.cra_end, ra.cra_rpc_size);
366 LASSERTF(ra.cra_end >= page_idx,
367 "object: %p, indcies %lu / %lu\n",
368 io->ci_obj, ra.cra_end, page_idx);
369 /* update read ahead RPC size.
370 * NB: it's racy but doesn't matter */
371 if (ras->ras_rpc_size > ra.cra_rpc_size &&
373 ras->ras_rpc_size = ra.cra_rpc_size;
374 /* trim it to align with optimal RPC size */
375 end = ras_align(ras, ria->ria_end + 1, NULL);
376 if (end > 0 && !ria->ria_eof)
377 ria->ria_end = end - 1;
378 if (ria->ria_end < ria->ria_end_min)
379 ria->ria_end = ria->ria_end_min;
380 if (ria->ria_end > ra.cra_end)
381 ria->ria_end = ra.cra_end;
383 if (page_idx > ria->ria_end)
386 /* If the page is inside the read-ahead window */
387 rc = ll_read_ahead_page(env, io, queue, page_idx);
392 /* Only subtract from reserve & count the page if we
393 * really did readahead on that page. */
398 } else if (stride_ria) {
399 /* If it is not in the read-ahead window, and it is
400 * read-ahead mode, then check whether it should skip
403 /* FIXME: This assertion only is valid when it is for
404 * forward read-ahead, it will be fixed when backward
405 * read-ahead is implemented */
406 LASSERTF(page_idx >= ria->ria_stoff,
407 "Invalid page_idx %lu rs %lu re %lu ro %lu "
408 "rl %lu rp %lu\n", page_idx,
409 ria->ria_start, ria->ria_end, ria->ria_stoff,
410 ria->ria_length, ria->ria_pages);
411 offset = page_idx - ria->ria_stoff;
412 offset = offset % (ria->ria_length);
413 if (offset > ria->ria_pages) {
414 page_idx += ria->ria_length - offset;
415 CDEBUG(D_READA, "i %lu skip %lu \n", page_idx,
416 ria->ria_length - offset);
422 cl_read_ahead_release(env, &ra);
427 static int ll_readahead(const struct lu_env *env, struct cl_io *io,
428 struct cl_page_list *queue,
429 struct ll_readahead_state *ras, bool hit)
431 struct vvp_io *vio = vvp_env_io(env);
432 struct ll_thread_info *lti = ll_env_info(env);
433 struct cl_attr *attr = vvp_env_thread_attr(env);
434 unsigned long len, mlen = 0;
435 pgoff_t ra_end = 0, start = 0, end = 0;
437 struct ra_io_arg *ria = <i->lti_ria;
438 struct cl_object *clob;
444 inode = vvp_object_inode(clob);
446 memset(ria, 0, sizeof *ria);
448 cl_object_attr_lock(clob);
449 ret = cl_object_attr_get(env, clob, attr);
450 cl_object_attr_unlock(clob);
456 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
460 spin_lock(&ras->ras_lock);
463 * Note: other thread might rollback the ras_next_readahead,
464 * if it can not get the full size of prepared pages, see the
465 * end of this function. For stride read ahead, it needs to
466 * make sure the offset is no less than ras_stride_offset,
467 * so that stride read ahead can work correctly.
469 if (stride_io_mode(ras))
470 start = max(ras->ras_next_readahead, ras->ras_stride_offset);
472 start = ras->ras_next_readahead;
474 if (ras->ras_window_len > 0)
475 end = ras->ras_window_start + ras->ras_window_len - 1;
477 /* Enlarge the RA window to encompass the full read */
478 if (vio->vui_ra_valid &&
479 end < vio->vui_ra_start + vio->vui_ra_count - 1)
480 end = vio->vui_ra_start + vio->vui_ra_count - 1;
483 unsigned long end_index;
485 /* Truncate RA window to end of file */
486 end_index = (unsigned long)((kms - 1) >> PAGE_SHIFT);
487 if (end_index <= end) {
492 ria->ria_start = start;
494 /* If stride I/O mode is detected, get stride window*/
495 if (stride_io_mode(ras)) {
496 ria->ria_stoff = ras->ras_stride_offset;
497 ria->ria_length = ras->ras_stride_length;
498 ria->ria_pages = ras->ras_stride_pages;
500 spin_unlock(&ras->ras_lock);
503 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
506 len = ria_page_count(ria);
508 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
513 CDEBUG(D_READA, DFID": ria: %lu/%lu, bead: %lu/%lu, hit: %d\n",
514 PFID(lu_object_fid(&clob->co_lu)),
515 ria->ria_start, ria->ria_end,
516 vio->vui_ra_valid ? vio->vui_ra_start : 0,
517 vio->vui_ra_valid ? vio->vui_ra_count : 0,
520 /* at least to extend the readahead window to cover current read */
521 if (!hit && vio->vui_ra_valid &&
522 vio->vui_ra_start + vio->vui_ra_count > ria->ria_start) {
523 unsigned long remainder;
525 /* to the end of current read window. */
526 mlen = vio->vui_ra_start + vio->vui_ra_count - ria->ria_start;
527 /* trim to RPC boundary */
528 ras_align(ras, ria->ria_start, &remainder);
529 mlen = min(mlen, ras->ras_rpc_size - remainder);
530 ria->ria_end_min = ria->ria_start + mlen;
533 ria->ria_reserved = ll_ra_count_get(ll_i2sbi(inode), ria, len, mlen);
534 if (ria->ria_reserved < len)
535 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
537 CDEBUG(D_READA, "reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
538 ria->ria_reserved, len, mlen,
539 atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
540 ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
542 ret = ll_read_ahead_pages(env, io, queue, ras, ria, &ra_end);
544 if (ria->ria_reserved != 0)
545 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
547 if (ra_end == end && ra_end == (kms >> PAGE_SHIFT))
548 ll_ra_stats_inc(inode, RA_STAT_EOF);
550 CDEBUG(D_READA, "ra_end = %lu end = %lu stride end = %lu pages = %d\n",
551 ra_end, end, ria->ria_end, ret);
554 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
556 /* update the ras so that the next read-ahead tries from
557 * where we left off. */
558 spin_lock(&ras->ras_lock);
559 ras->ras_next_readahead = ra_end + 1;
560 spin_unlock(&ras->ras_lock);
567 static void ras_set_start(struct inode *inode, struct ll_readahead_state *ras,
570 ras->ras_window_start = ras_align(ras, index, NULL);
573 /* called with the ras_lock held or from places where it doesn't matter */
574 static void ras_reset(struct inode *inode, struct ll_readahead_state *ras,
577 ras->ras_last_readpage = index;
578 ras->ras_consecutive_requests = 0;
579 ras->ras_consecutive_pages = 0;
580 ras->ras_window_len = 0;
581 ras_set_start(inode, ras, index);
582 ras->ras_next_readahead = max(ras->ras_window_start, index + 1);
587 /* called with the ras_lock held or from places where it doesn't matter */
588 static void ras_stride_reset(struct ll_readahead_state *ras)
590 ras->ras_consecutive_stride_requests = 0;
591 ras->ras_stride_length = 0;
592 ras->ras_stride_pages = 0;
596 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
598 spin_lock_init(&ras->ras_lock);
599 ras->ras_rpc_size = PTLRPC_MAX_BRW_PAGES;
600 ras_reset(inode, ras, 0);
601 ras->ras_requests = 0;
605 * Check whether the read request is in the stride window.
606 * If it is in the stride window, return 1, otherwise return 0.
608 static int index_in_stride_window(struct ll_readahead_state *ras,
611 unsigned long stride_gap;
613 if (ras->ras_stride_length == 0 || ras->ras_stride_pages == 0 ||
614 ras->ras_stride_pages == ras->ras_stride_length)
617 stride_gap = index - ras->ras_last_readpage - 1;
619 /* If it is contiguous read */
621 return ras->ras_consecutive_pages + 1 <= ras->ras_stride_pages;
623 /* Otherwise check the stride by itself */
624 return (ras->ras_stride_length - ras->ras_stride_pages) == stride_gap &&
625 ras->ras_consecutive_pages == ras->ras_stride_pages;
628 static void ras_update_stride_detector(struct ll_readahead_state *ras,
631 unsigned long stride_gap = index - ras->ras_last_readpage - 1;
633 if (!stride_io_mode(ras) && (stride_gap != 0 ||
634 ras->ras_consecutive_stride_requests == 0)) {
635 ras->ras_stride_pages = ras->ras_consecutive_pages;
636 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
638 LASSERT(ras->ras_request_index == 0);
639 LASSERT(ras->ras_consecutive_stride_requests == 0);
641 if (index <= ras->ras_last_readpage) {
642 /*Reset stride window for forward read*/
643 ras_stride_reset(ras);
647 ras->ras_stride_pages = ras->ras_consecutive_pages;
648 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
655 stride_page_count(struct ll_readahead_state *ras, unsigned long len)
657 return stride_pg_count(ras->ras_stride_offset, ras->ras_stride_length,
658 ras->ras_stride_pages, ras->ras_stride_offset,
662 /* Stride Read-ahead window will be increased inc_len according to
663 * stride I/O pattern */
664 static void ras_stride_increase_window(struct ll_readahead_state *ras,
665 struct ll_ra_info *ra,
666 unsigned long inc_len)
668 unsigned long left, step, window_len;
669 unsigned long stride_len;
671 LASSERT(ras->ras_stride_length > 0);
672 LASSERTF(ras->ras_window_start + ras->ras_window_len
673 >= ras->ras_stride_offset, "window_start %lu, window_len %lu"
674 " stride_offset %lu\n", ras->ras_window_start,
675 ras->ras_window_len, ras->ras_stride_offset);
677 stride_len = ras->ras_window_start + ras->ras_window_len -
678 ras->ras_stride_offset;
680 left = stride_len % ras->ras_stride_length;
681 window_len = ras->ras_window_len - left;
683 if (left < ras->ras_stride_pages)
686 left = ras->ras_stride_pages + inc_len;
688 LASSERT(ras->ras_stride_pages != 0);
690 step = left / ras->ras_stride_pages;
691 left %= ras->ras_stride_pages;
693 window_len += step * ras->ras_stride_length + left;
695 if (stride_page_count(ras, window_len) <= ra->ra_max_pages_per_file)
696 ras->ras_window_len = window_len;
701 static void ras_increase_window(struct inode *inode,
702 struct ll_readahead_state *ras,
703 struct ll_ra_info *ra)
705 /* The stretch of ra-window should be aligned with max rpc_size
706 * but current clio architecture does not support retrieve such
707 * information from lower layer. FIXME later
709 if (stride_io_mode(ras)) {
710 ras_stride_increase_window(ras, ra, ras->ras_rpc_size);
714 wlen = min(ras->ras_window_len + ras->ras_rpc_size,
715 ra->ra_max_pages_per_file);
716 if (wlen < ras->ras_rpc_size)
717 ras->ras_window_len = wlen;
719 ras->ras_window_len = ras_align(ras, wlen, NULL);
723 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
724 struct ll_readahead_state *ras, unsigned long index,
725 enum ras_update_flags flags)
727 struct ll_ra_info *ra = &sbi->ll_ra_info;
728 bool hit = flags & LL_RAS_HIT;
729 int zero = 0, stride_detect = 0, ra_miss = 0;
732 spin_lock(&ras->ras_lock);
735 CDEBUG(D_READA, DFID " pages at %lu miss.\n",
736 PFID(ll_inode2fid(inode)), index);
737 ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
739 /* reset the read-ahead window in two cases. First when the app seeks
740 * or reads to some other part of the file. Secondly if we get a
741 * read-ahead miss that we think we've previously issued. This can
742 * be a symptom of there being so many read-ahead pages that the VM is
743 * reclaiming it before we get to it. */
744 if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
746 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
747 } else if (!hit && ras->ras_window_len &&
748 index < ras->ras_next_readahead &&
749 index_in_window(index, ras->ras_window_start, 0,
750 ras->ras_window_len)) {
752 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
755 /* On the second access to a file smaller than the tunable
756 * ra_max_read_ahead_whole_pages trigger RA on all pages in the
757 * file up to ra_max_pages_per_file. This is simply a best effort
758 * and only occurs once per open file. Normal RA behavior is reverted
759 * to for subsequent IO. The mmap case does not increment
760 * ras_requests and thus can never trigger this behavior. */
761 if (ras->ras_requests >= 2 && !ras->ras_request_index) {
764 kms_pages = (i_size_read(inode) + PAGE_SIZE - 1) >>
767 CDEBUG(D_READA, "kmsp %llu mwp %lu mp %lu\n", kms_pages,
768 ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages_per_file);
771 kms_pages <= ra->ra_max_read_ahead_whole_pages) {
772 ras->ras_window_start = 0;
773 ras->ras_next_readahead = index + 1;
774 ras->ras_window_len = min(ra->ra_max_pages_per_file,
775 ra->ra_max_read_ahead_whole_pages);
780 /* check whether it is in stride I/O mode*/
781 if (!index_in_stride_window(ras, index)) {
782 if (ras->ras_consecutive_stride_requests == 0 &&
783 ras->ras_request_index == 0) {
784 ras_update_stride_detector(ras, index);
785 ras->ras_consecutive_stride_requests++;
787 ras_stride_reset(ras);
789 ras_reset(inode, ras, index);
790 ras->ras_consecutive_pages++;
793 ras->ras_consecutive_pages = 0;
794 ras->ras_consecutive_requests = 0;
795 if (++ras->ras_consecutive_stride_requests > 1)
801 if (index_in_stride_window(ras, index) &&
802 stride_io_mode(ras)) {
803 if (index != ras->ras_last_readpage + 1)
804 ras->ras_consecutive_pages = 0;
805 ras_reset(inode, ras, index);
807 /* If stride-RA hit cache miss, the stride
808 * detector will not be reset to avoid the
809 * overhead of redetecting read-ahead mode,
810 * but on the condition that the stride window
811 * is still intersect with normal sequential
812 * read-ahead window. */
813 if (ras->ras_window_start <
814 ras->ras_stride_offset)
815 ras_stride_reset(ras);
818 /* Reset both stride window and normal RA
820 ras_reset(inode, ras, index);
821 ras->ras_consecutive_pages++;
822 ras_stride_reset(ras);
825 } else if (stride_io_mode(ras)) {
826 /* If this is contiguous read but in stride I/O mode
827 * currently, check whether stride step still is valid,
828 * if invalid, it will reset the stride ra window*/
829 if (!index_in_stride_window(ras, index)) {
830 /* Shrink stride read-ahead window to be zero */
831 ras_stride_reset(ras);
832 ras->ras_window_len = 0;
833 ras->ras_next_readahead = index;
837 ras->ras_consecutive_pages++;
838 ras->ras_last_readpage = index;
839 ras_set_start(inode, ras, index);
841 if (stride_io_mode(ras)) {
842 /* Since stride readahead is sentivite to the offset
843 * of read-ahead, so we use original offset here,
844 * instead of ras_window_start, which is RPC aligned */
845 ras->ras_next_readahead = max(index + 1,
846 ras->ras_next_readahead);
847 ras->ras_window_start = max(ras->ras_stride_offset,
848 ras->ras_window_start);
850 if (ras->ras_next_readahead < ras->ras_window_start)
851 ras->ras_next_readahead = ras->ras_window_start;
853 ras->ras_next_readahead = index + 1;
857 /* Trigger RA in the mmap case where ras_consecutive_requests
858 * is not incremented and thus can't be used to trigger RA */
859 if (ras->ras_consecutive_pages >= 4 && flags & LL_RAS_MMAP) {
860 ras_increase_window(inode, ras, ra);
861 /* reset consecutive pages so that the readahead window can
863 ras->ras_consecutive_pages = 0;
867 /* Initially reset the stride window offset to next_readahead*/
868 if (ras->ras_consecutive_stride_requests == 2 && stride_detect) {
870 * Once stride IO mode is detected, next_readahead should be
871 * reset to make sure next_readahead > stride offset
873 ras->ras_next_readahead = max(index, ras->ras_next_readahead);
874 ras->ras_stride_offset = index;
875 ras->ras_window_start = max(index, ras->ras_window_start);
878 /* The initial ras_window_len is set to the request size. To avoid
879 * uselessly reading and discarding pages for random IO the window is
880 * only increased once per consecutive request received. */
881 if ((ras->ras_consecutive_requests > 1 || stride_detect) &&
882 !ras->ras_request_index)
883 ras_increase_window(inode, ras, ra);
887 ras->ras_request_index++;
888 spin_unlock(&ras->ras_lock);
892 int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
894 struct inode *inode = vmpage->mapping->host;
895 struct ll_inode_info *lli = ll_i2info(inode);
898 struct cl_page *page;
899 struct cl_object *clob;
900 bool redirtied = false;
901 bool unlocked = false;
906 LASSERT(PageLocked(vmpage));
907 LASSERT(!PageWriteback(vmpage));
909 LASSERT(ll_i2dtexp(inode) != NULL);
911 env = cl_env_get(&refcheck);
913 GOTO(out, result = PTR_ERR(env));
915 clob = ll_i2info(inode)->lli_clob;
916 LASSERT(clob != NULL);
918 io = vvp_env_thread_io(env);
920 io->ci_ignore_layout = 1;
921 result = cl_io_init(env, io, CIT_MISC, clob);
923 page = cl_page_find(env, clob, vmpage->index,
924 vmpage, CPT_CACHEABLE);
926 lu_ref_add(&page->cp_reference, "writepage",
928 cl_page_assume(env, io, page);
929 result = cl_page_flush(env, io, page);
932 * Re-dirty page on error so it retries write,
933 * but not in case when IO has actually
934 * occurred and completed with an error.
936 if (!PageError(vmpage)) {
937 redirty_page_for_writepage(wbc, vmpage);
942 cl_page_disown(env, io, page);
944 lu_ref_del(&page->cp_reference,
945 "writepage", current);
946 cl_page_put(env, page);
948 result = PTR_ERR(page);
953 if (redirtied && wbc->sync_mode == WB_SYNC_ALL) {
954 loff_t offset = cl_offset(clob, vmpage->index);
956 /* Flush page failed because the extent is being written out.
957 * Wait for the write of extent to be finished to avoid
958 * breaking kernel which assumes ->writepage should mark
959 * PageWriteback or clean the page. */
960 result = cl_sync_file_range(inode, offset,
961 offset + PAGE_SIZE - 1,
964 /* actually we may have written more than one page.
965 * decreasing this page because the caller will count
967 wbc->nr_to_write -= result - 1;
972 cl_env_put(env, &refcheck);
977 if (!lli->lli_async_rc)
978 lli->lli_async_rc = result;
979 SetPageError(vmpage);
986 int ll_writepages(struct address_space *mapping, struct writeback_control *wbc)
988 struct inode *inode = mapping->host;
991 enum cl_fsync_mode mode;
996 if (wbc->range_cyclic) {
997 start = mapping->writeback_index << PAGE_SHIFT;
998 end = OBD_OBJECT_EOF;
1000 start = wbc->range_start;
1001 end = wbc->range_end;
1002 if (end == LLONG_MAX) {
1003 end = OBD_OBJECT_EOF;
1004 range_whole = start == 0;
1008 mode = CL_FSYNC_NONE;
1009 if (wbc->sync_mode == WB_SYNC_ALL)
1010 mode = CL_FSYNC_LOCAL;
1012 if (ll_i2info(inode)->lli_clob == NULL)
1015 /* for directio, it would call writepages() to evict cached pages
1016 * inside the IO context of write, which will cause deadlock at
1017 * layout_conf since it waits for active IOs to complete. */
1018 result = cl_sync_file_range(inode, start, end, mode, 1);
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_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_inode(file);
1142 struct cl_object *clob = ll_i2info(inode)->lli_clob;
1143 struct ll_cl_context *lcc;
1144 const struct lu_env *env = NULL;
1145 struct cl_io *io = NULL;
1146 struct cl_page *page;
1150 lcc = ll_cl_find(file);
1156 if (io == NULL) { /* fast read */
1157 struct inode *inode = file_inode(file);
1158 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1159 struct ll_readahead_state *ras = &fd->fd_ras;
1160 struct lu_env *local_env = NULL;
1161 struct vvp_page *vpg;
1165 /* TODO: need to verify the layout version to make sure
1166 * the page is not invalid due to layout change. */
1167 page = cl_vmpage_page(vmpage, clob);
1169 unlock_page(vmpage);
1174 local_env = cl_env_percpu_get();
1178 vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1179 if (vpg->vpg_defer_uptodate) {
1180 enum ras_update_flags flags = LL_RAS_HIT;
1182 if (lcc && lcc->lcc_type == LCC_MMAP)
1183 flags |= LL_RAS_MMAP;
1185 /* For fast read, it updates read ahead state only
1186 * if the page is hit in cache because non cache page
1187 * case will be handled by slow read later. */
1188 ras_update(ll_i2sbi(inode), inode, ras, vvp_index(vpg),
1190 /* avoid duplicate ras_update() call */
1191 vpg->vpg_ra_updated = 1;
1193 /* Check if we can issue a readahead RPC, if that is
1194 * the case, we can't do fast IO because we will need
1195 * a cl_io to issue the RPC. */
1196 if (ras->ras_window_start + ras->ras_window_len <
1197 ras->ras_next_readahead + PTLRPC_MAX_BRW_PAGES) {
1198 /* export the page and skip io stack */
1199 vpg->vpg_ra_used = 1;
1200 cl_page_export(env, page, 1);
1205 /* release page refcount before unlocking the page to ensure
1206 * the object won't be destroyed in the calling path of
1207 * cl_page_put(). Please see comment in ll_releasepage(). */
1208 cl_page_put(env, page);
1209 unlock_page(vmpage);
1211 cl_env_percpu_put(local_env);
1216 LASSERT(io->ci_state == CIS_IO_GOING);
1217 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
1218 if (!IS_ERR(page)) {
1219 LASSERT(page->cp_type == CPT_CACHEABLE);
1220 if (likely(!PageUptodate(vmpage))) {
1221 cl_page_assume(env, io, page);
1222 result = ll_io_read_page(env, io, page, file);
1224 /* Page from a non-object file. */
1225 unlock_page(vmpage);
1228 cl_page_put(env, page);
1230 unlock_page(vmpage);
1231 result = PTR_ERR(page);
1236 int ll_page_sync_io(const struct lu_env *env, struct cl_io *io,
1237 struct cl_page *page, enum cl_req_type crt)
1239 struct cl_2queue *queue;
1242 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
1244 queue = &io->ci_queue;
1245 cl_2queue_init_page(queue, page);
1247 result = cl_io_submit_sync(env, io, crt, queue, 0);
1248 LASSERT(cl_page_is_owned(page, io));
1250 if (crt == CRT_READ)
1252 * in CRT_WRITE case page is left locked even in case of
1255 cl_page_list_disown(env, io, &queue->c2_qin);
1256 cl_2queue_fini(env, queue);