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/
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>
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,
84 unsigned long pages, unsigned long min)
86 struct ll_ra_info *ra = &sbi->ll_ra_info;
90 /* If read-ahead pages left are less than 1M, do not do read-ahead,
91 * otherwise it will form small read RPC(< 1M), which hurt server
92 * performance a lot. */
93 ret = min(ra->ra_max_pages - atomic_read(&ra->ra_cur_pages),
95 if (ret < 0 || ret < min_t(long, PTLRPC_MAX_BRW_PAGES, pages))
98 if (atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) {
99 atomic_sub(ret, &ra->ra_cur_pages);
105 /* override ra limit for maximum performance */
106 atomic_add(min - ret, &ra->ra_cur_pages);
112 void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
114 struct ll_ra_info *ra = &sbi->ll_ra_info;
115 atomic_sub(len, &ra->ra_cur_pages);
118 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which)
120 LASSERTF(which < _NR_RA_STAT, "which: %u\n", which);
121 lprocfs_counter_incr(sbi->ll_ra_stats, which);
124 void ll_ra_stats_inc(struct inode *inode, enum ra_stat which)
126 struct ll_sb_info *sbi = ll_i2sbi(inode);
127 ll_ra_stats_inc_sbi(sbi, which);
130 #define RAS_CDEBUG(ras) \
132 "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu rpc %lu " \
133 "r %lu ri %lu csr %lu sf %lu sp %lu sl %lu\n", \
134 ras->ras_last_readpage, ras->ras_consecutive_requests, \
135 ras->ras_consecutive_pages, ras->ras_window_start, \
136 ras->ras_window_len, ras->ras_next_readahead, \
138 ras->ras_requests, ras->ras_request_index, \
139 ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
140 ras->ras_stride_pages, ras->ras_stride_length)
142 static int index_in_window(unsigned long index, unsigned long point,
143 unsigned long before, unsigned long after)
145 unsigned long start = point - before, end = point + after;
152 return start <= index && index <= end;
155 void ll_ras_enter(struct file *f)
157 struct ll_file_data *fd = LUSTRE_FPRIVATE(f);
158 struct ll_readahead_state *ras = &fd->fd_ras;
160 spin_lock(&ras->ras_lock);
162 ras->ras_request_index = 0;
163 ras->ras_consecutive_requests++;
164 spin_unlock(&ras->ras_lock);
168 * Initiates read-ahead of a page with given index.
170 * \retval +ve: page was already uptodate so it will be skipped
172 * \retval -ve: page wasn't added to \a queue for error;
173 * \retval 0: page was added into \a queue for read ahead.
175 static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io,
176 struct cl_page_list *queue, pgoff_t index)
178 struct cl_object *clob = io->ci_obj;
179 struct inode *inode = vvp_object_inode(clob);
181 struct cl_page *page;
182 struct vvp_page *vpg;
183 enum ra_stat which = _NR_RA_STAT; /* keep gcc happy */
185 const char *msg = NULL;
188 vmpage = grab_cache_page_nowait(inode->i_mapping, index);
189 if (vmpage == NULL) {
190 which = RA_STAT_FAILED_GRAB_PAGE;
191 msg = "g_c_p_n failed";
192 GOTO(out, rc = -EBUSY);
195 /* Check if vmpage was truncated or reclaimed */
196 if (vmpage->mapping != inode->i_mapping) {
197 which = RA_STAT_WRONG_GRAB_PAGE;
198 msg = "g_c_p_n returned invalid page";
199 GOTO(out, rc = -EBUSY);
202 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
204 which = RA_STAT_FAILED_GRAB_PAGE;
205 msg = "cl_page_find failed";
206 GOTO(out, rc = PTR_ERR(page));
209 lu_ref_add(&page->cp_reference, "ra", current);
210 cl_page_assume(env, io, page);
211 vpg = cl2vvp_page(cl_object_page_slice(clob, page));
212 if (!vpg->vpg_defer_uptodate && !PageUptodate(vmpage)) {
213 vpg->vpg_defer_uptodate = 1;
214 vpg->vpg_ra_used = 0;
215 cl_page_list_add(queue, page);
217 /* skip completed pages */
218 cl_page_unassume(env, io, page);
219 /* This page is already uptodate, returning a positive number
220 * to tell the callers about this */
224 lu_ref_del(&page->cp_reference, "ra", current);
225 cl_page_put(env, page);
228 if (vmpage != NULL) {
234 ll_ra_stats_inc(inode, which);
235 CDEBUG(D_READA, "%s\n", msg);
242 #define RIA_DEBUG(ria) \
243 CDEBUG(D_READA, "rs %lu re %lu ro %lu rl %lu rp %lu\n", \
244 ria->ria_start, ria->ria_end, ria->ria_stoff, ria->ria_length,\
247 static inline int stride_io_mode(struct ll_readahead_state *ras)
249 return ras->ras_consecutive_stride_requests > 1;
252 /* The function calculates how much pages will be read in
253 * [off, off + length], in such stride IO area,
254 * stride_offset = st_off, stride_lengh = st_len,
255 * stride_pages = st_pgs
257 * |------------------|*****|------------------|*****|------------|*****|....
260 * |----- st_len -----|
262 * How many pages it should read in such pattern
263 * |-------------------------------------------------------------|
265 * |<------ length ------->|
267 * = |<----->| + |-------------------------------------| + |---|
268 * start_left st_pgs * i end_left
271 stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs,
272 unsigned long off, unsigned long length)
274 __u64 start = off > st_off ? off - st_off : 0;
275 __u64 end = off + length > st_off ? off + length - st_off : 0;
276 unsigned long start_left = 0;
277 unsigned long end_left = 0;
278 unsigned long pg_count;
280 if (st_len == 0 || length == 0 || end == 0)
283 start_left = do_div(start, st_len);
284 if (start_left < st_pgs)
285 start_left = st_pgs - start_left;
289 end_left = do_div(end, st_len);
290 if (end_left > st_pgs)
293 CDEBUG(D_READA, "start %llu, end %llu start_left %lu end_left %lu\n",
294 start, end, start_left, end_left);
297 pg_count = end_left - (st_pgs - start_left);
299 pg_count = start_left + st_pgs * (end - start - 1) + end_left;
301 CDEBUG(D_READA, "st_off %lu, st_len %lu st_pgs %lu off %lu length %lu"
302 "pgcount %lu\n", st_off, st_len, st_pgs, off, length, pg_count);
307 static int ria_page_count(struct ra_io_arg *ria)
309 __u64 length = ria->ria_end >= ria->ria_start ?
310 ria->ria_end - ria->ria_start + 1 : 0;
312 return stride_pg_count(ria->ria_stoff, ria->ria_length,
313 ria->ria_pages, ria->ria_start,
317 static unsigned long ras_align(struct ll_readahead_state *ras,
319 unsigned long *remainder)
321 unsigned long rem = index % ras->ras_rpc_size;
322 if (remainder != NULL)
327 /*Check whether the index is in the defined ra-window */
328 static int ras_inside_ra_window(unsigned long idx, struct ra_io_arg *ria)
330 /* If ria_length == ria_pages, it means non-stride I/O mode,
331 * idx should always inside read-ahead window in this case
332 * For stride I/O mode, just check whether the idx is inside
334 return ria->ria_length == 0 || ria->ria_length == ria->ria_pages ||
335 (idx >= ria->ria_stoff && (idx - ria->ria_stoff) %
336 ria->ria_length < ria->ria_pages);
340 ll_read_ahead_pages(const struct lu_env *env, struct cl_io *io,
341 struct cl_page_list *queue, struct ll_readahead_state *ras,
342 struct ra_io_arg *ria, pgoff_t *ra_end)
344 struct cl_read_ahead ra = { 0 };
345 int rc = 0, count = 0;
349 LASSERT(ria != NULL);
352 stride_ria = ria->ria_length > ria->ria_pages && ria->ria_pages > 0;
353 for (page_idx = ria->ria_start;
354 page_idx <= ria->ria_end && ria->ria_reserved > 0; page_idx++) {
355 if (ras_inside_ra_window(page_idx, ria)) {
356 if (ra.cra_end == 0 || ra.cra_end < page_idx) {
359 cl_read_ahead_release(env, &ra);
361 rc = cl_io_read_ahead(env, io, page_idx, &ra);
365 CDEBUG(D_READA, "idx: %lu, ra: %lu, rpc: %lu\n",
366 page_idx, ra.cra_end, ra.cra_rpc_size);
367 LASSERTF(ra.cra_end >= page_idx,
368 "object: %p, indcies %lu / %lu\n",
369 io->ci_obj, ra.cra_end, page_idx);
370 /* update read ahead RPC size.
371 * NB: it's racy but doesn't matter */
372 if (ras->ras_rpc_size > ra.cra_rpc_size &&
374 ras->ras_rpc_size = ra.cra_rpc_size;
375 /* trim it to align with optimal RPC size */
376 end = ras_align(ras, ria->ria_end + 1, NULL);
377 if (end > 0 && !ria->ria_eof)
378 ria->ria_end = end - 1;
379 if (ria->ria_end < ria->ria_end_min)
380 ria->ria_end = ria->ria_end_min;
381 if (ria->ria_end > ra.cra_end)
382 ria->ria_end = ra.cra_end;
384 if (page_idx > ria->ria_end)
387 /* If the page is inside the read-ahead window */
388 rc = ll_read_ahead_page(env, io, queue, page_idx);
393 /* Only subtract from reserve & count the page if we
394 * really did readahead on that page. */
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 = 0, 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_SHIFT);
488 if (end_index <= end) {
493 ria->ria_start = start;
495 /* If stride I/O mode is detected, get stride window*/
496 if (stride_io_mode(ras)) {
497 ria->ria_stoff = ras->ras_stride_offset;
498 ria->ria_length = ras->ras_stride_length;
499 ria->ria_pages = ras->ras_stride_pages;
501 spin_unlock(&ras->ras_lock);
504 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
507 len = ria_page_count(ria);
509 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
514 CDEBUG(D_READA, DFID": ria: %lu/%lu, bead: %lu/%lu, hit: %d\n",
515 PFID(lu_object_fid(&clob->co_lu)),
516 ria->ria_start, ria->ria_end,
517 vio->vui_ra_valid ? vio->vui_ra_start : 0,
518 vio->vui_ra_valid ? vio->vui_ra_count : 0,
521 /* at least to extend the readahead window to cover current read */
522 if (!hit && vio->vui_ra_valid &&
523 vio->vui_ra_start + vio->vui_ra_count > ria->ria_start) {
524 unsigned long remainder;
526 /* to the end of current read window. */
527 mlen = vio->vui_ra_start + vio->vui_ra_count - ria->ria_start;
528 /* trim to RPC boundary */
529 ras_align(ras, ria->ria_start, &remainder);
530 mlen = min(mlen, ras->ras_rpc_size - remainder);
531 ria->ria_end_min = ria->ria_start + mlen;
534 ria->ria_reserved = ll_ra_count_get(ll_i2sbi(inode), ria, len, mlen);
535 if (ria->ria_reserved < len)
536 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
538 CDEBUG(D_READA, "reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
539 ria->ria_reserved, len, mlen,
540 atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
541 ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
543 ret = ll_read_ahead_pages(env, io, queue, ras, ria, &ra_end);
545 if (ria->ria_reserved != 0)
546 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
548 if (ra_end == end && ra_end == (kms >> PAGE_SHIFT))
549 ll_ra_stats_inc(inode, RA_STAT_EOF);
551 CDEBUG(D_READA, "ra_end = %lu end = %lu stride end = %lu pages = %d\n",
552 ra_end, end, ria->ria_end, ret);
555 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
557 /* update the ras so that the next read-ahead tries from
558 * where we left off. */
559 spin_lock(&ras->ras_lock);
560 ras->ras_next_readahead = ra_end + 1;
561 spin_unlock(&ras->ras_lock);
568 static void ras_set_start(struct inode *inode, struct ll_readahead_state *ras,
571 ras->ras_window_start = ras_align(ras, index, NULL);
574 /* called with the ras_lock held or from places where it doesn't matter */
575 static void ras_reset(struct inode *inode, struct ll_readahead_state *ras,
578 ras->ras_last_readpage = index;
579 ras->ras_consecutive_requests = 0;
580 ras->ras_consecutive_pages = 0;
581 ras->ras_window_len = 0;
582 ras_set_start(inode, ras, index);
583 ras->ras_next_readahead = max(ras->ras_window_start, index + 1);
588 /* called with the ras_lock held or from places where it doesn't matter */
589 static void ras_stride_reset(struct ll_readahead_state *ras)
591 ras->ras_consecutive_stride_requests = 0;
592 ras->ras_stride_length = 0;
593 ras->ras_stride_pages = 0;
597 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
599 spin_lock_init(&ras->ras_lock);
600 ras->ras_rpc_size = PTLRPC_MAX_BRW_PAGES;
601 ras_reset(inode, ras, 0);
602 ras->ras_requests = 0;
606 * Check whether the read request is in the stride window.
607 * If it is in the stride window, return 1, otherwise return 0.
609 static int index_in_stride_window(struct ll_readahead_state *ras,
612 unsigned long stride_gap;
614 if (ras->ras_stride_length == 0 || ras->ras_stride_pages == 0 ||
615 ras->ras_stride_pages == ras->ras_stride_length)
618 stride_gap = index - ras->ras_last_readpage - 1;
620 /* If it is contiguous read */
622 return ras->ras_consecutive_pages + 1 <= ras->ras_stride_pages;
624 /* Otherwise check the stride by itself */
625 return (ras->ras_stride_length - ras->ras_stride_pages) == stride_gap &&
626 ras->ras_consecutive_pages == ras->ras_stride_pages;
629 static void ras_update_stride_detector(struct ll_readahead_state *ras,
632 unsigned long stride_gap = index - ras->ras_last_readpage - 1;
634 if (!stride_io_mode(ras) && (stride_gap != 0 ||
635 ras->ras_consecutive_stride_requests == 0)) {
636 ras->ras_stride_pages = ras->ras_consecutive_pages;
637 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
639 LASSERT(ras->ras_request_index == 0);
640 LASSERT(ras->ras_consecutive_stride_requests == 0);
642 if (index <= ras->ras_last_readpage) {
643 /*Reset stride window for forward read*/
644 ras_stride_reset(ras);
648 ras->ras_stride_pages = ras->ras_consecutive_pages;
649 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
656 stride_page_count(struct ll_readahead_state *ras, unsigned long len)
658 return stride_pg_count(ras->ras_stride_offset, ras->ras_stride_length,
659 ras->ras_stride_pages, ras->ras_stride_offset,
663 /* Stride Read-ahead window will be increased inc_len according to
664 * stride I/O pattern */
665 static void ras_stride_increase_window(struct ll_readahead_state *ras,
666 struct ll_ra_info *ra,
667 unsigned long inc_len)
669 unsigned long left, step, window_len;
670 unsigned long stride_len;
672 LASSERT(ras->ras_stride_length > 0);
673 LASSERTF(ras->ras_window_start + ras->ras_window_len
674 >= ras->ras_stride_offset, "window_start %lu, window_len %lu"
675 " stride_offset %lu\n", ras->ras_window_start,
676 ras->ras_window_len, ras->ras_stride_offset);
678 stride_len = ras->ras_window_start + ras->ras_window_len -
679 ras->ras_stride_offset;
681 left = stride_len % ras->ras_stride_length;
682 window_len = ras->ras_window_len - left;
684 if (left < ras->ras_stride_pages)
687 left = ras->ras_stride_pages + inc_len;
689 LASSERT(ras->ras_stride_pages != 0);
691 step = left / ras->ras_stride_pages;
692 left %= ras->ras_stride_pages;
694 window_len += step * ras->ras_stride_length + left;
696 if (stride_page_count(ras, window_len) <= ra->ra_max_pages_per_file)
697 ras->ras_window_len = window_len;
702 static void ras_increase_window(struct inode *inode,
703 struct ll_readahead_state *ras,
704 struct ll_ra_info *ra)
706 /* The stretch of ra-window should be aligned with max rpc_size
707 * but current clio architecture does not support retrieve such
708 * information from lower layer. FIXME later
710 if (stride_io_mode(ras)) {
711 ras_stride_increase_window(ras, ra, ras->ras_rpc_size);
715 wlen = min(ras->ras_window_len + ras->ras_rpc_size,
716 ra->ra_max_pages_per_file);
717 if (wlen < ras->ras_rpc_size)
718 ras->ras_window_len = wlen;
720 ras->ras_window_len = ras_align(ras, wlen, NULL);
724 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
725 struct ll_readahead_state *ras, unsigned long index,
726 enum ras_update_flags flags)
728 struct ll_ra_info *ra = &sbi->ll_ra_info;
729 bool hit = flags & LL_RAS_HIT;
730 int zero = 0, stride_detect = 0, ra_miss = 0;
733 spin_lock(&ras->ras_lock);
736 CDEBUG(D_READA, DFID " pages at %lu miss.\n",
737 PFID(ll_inode2fid(inode)), index);
738 ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
740 /* reset the read-ahead window in two cases. First when the app seeks
741 * or reads to some other part of the file. Secondly if we get a
742 * read-ahead miss that we think we've previously issued. This can
743 * be a symptom of there being so many read-ahead pages that the VM is
744 * reclaiming it before we get to it. */
745 if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
747 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
748 } else if (!hit && ras->ras_window_len &&
749 index < ras->ras_next_readahead &&
750 index_in_window(index, ras->ras_window_start, 0,
751 ras->ras_window_len)) {
753 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
756 /* On the second access to a file smaller than the tunable
757 * ra_max_read_ahead_whole_pages trigger RA on all pages in the
758 * file up to ra_max_pages_per_file. This is simply a best effort
759 * and only occurs once per open file. Normal RA behavior is reverted
760 * to for subsequent IO. The mmap case does not increment
761 * ras_requests and thus can never trigger this behavior. */
762 if (ras->ras_requests >= 2 && !ras->ras_request_index) {
765 kms_pages = (i_size_read(inode) + PAGE_SIZE - 1) >>
768 CDEBUG(D_READA, "kmsp %llu mwp %lu mp %lu\n", kms_pages,
769 ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages_per_file);
772 kms_pages <= ra->ra_max_read_ahead_whole_pages) {
773 ras->ras_window_start = 0;
774 ras->ras_next_readahead = index + 1;
775 ras->ras_window_len = min(ra->ra_max_pages_per_file,
776 ra->ra_max_read_ahead_whole_pages);
781 /* check whether it is in stride I/O mode*/
782 if (!index_in_stride_window(ras, index)) {
783 if (ras->ras_consecutive_stride_requests == 0 &&
784 ras->ras_request_index == 0) {
785 ras_update_stride_detector(ras, index);
786 ras->ras_consecutive_stride_requests++;
788 ras_stride_reset(ras);
790 ras_reset(inode, ras, index);
791 ras->ras_consecutive_pages++;
794 ras->ras_consecutive_pages = 0;
795 ras->ras_consecutive_requests = 0;
796 if (++ras->ras_consecutive_stride_requests > 1)
802 if (index_in_stride_window(ras, index) &&
803 stride_io_mode(ras)) {
804 if (index != ras->ras_last_readpage + 1)
805 ras->ras_consecutive_pages = 0;
806 ras_reset(inode, ras, index);
808 /* If stride-RA hit cache miss, the stride
809 * detector will not be reset to avoid the
810 * overhead of redetecting read-ahead mode,
811 * but on the condition that the stride window
812 * is still intersect with normal sequential
813 * read-ahead window. */
814 if (ras->ras_window_start <
815 ras->ras_stride_offset)
816 ras_stride_reset(ras);
819 /* Reset both stride window and normal RA
821 ras_reset(inode, ras, index);
822 ras->ras_consecutive_pages++;
823 ras_stride_reset(ras);
826 } else if (stride_io_mode(ras)) {
827 /* If this is contiguous read but in stride I/O mode
828 * currently, check whether stride step still is valid,
829 * if invalid, it will reset the stride ra window*/
830 if (!index_in_stride_window(ras, index)) {
831 /* Shrink stride read-ahead window to be zero */
832 ras_stride_reset(ras);
833 ras->ras_window_len = 0;
834 ras->ras_next_readahead = index;
838 ras->ras_consecutive_pages++;
839 ras->ras_last_readpage = index;
840 ras_set_start(inode, ras, index);
842 if (stride_io_mode(ras)) {
843 /* Since stride readahead is sentivite to the offset
844 * of read-ahead, so we use original offset here,
845 * instead of ras_window_start, which is RPC aligned */
846 ras->ras_next_readahead = max(index + 1,
847 ras->ras_next_readahead);
848 ras->ras_window_start = max(ras->ras_stride_offset,
849 ras->ras_window_start);
851 if (ras->ras_next_readahead < ras->ras_window_start)
852 ras->ras_next_readahead = ras->ras_window_start;
854 ras->ras_next_readahead = index + 1;
858 /* Trigger RA in the mmap case where ras_consecutive_requests
859 * is not incremented and thus can't be used to trigger RA */
860 if (ras->ras_consecutive_pages >= 4 && flags & LL_RAS_MMAP) {
861 ras_increase_window(inode, ras, ra);
862 /* reset consecutive pages so that the readahead window can
864 ras->ras_consecutive_pages = 0;
868 /* Initially reset the stride window offset to next_readahead*/
869 if (ras->ras_consecutive_stride_requests == 2 && stride_detect) {
871 * Once stride IO mode is detected, next_readahead should be
872 * reset to make sure next_readahead > stride offset
874 ras->ras_next_readahead = max(index, ras->ras_next_readahead);
875 ras->ras_stride_offset = index;
876 ras->ras_window_start = max(index, ras->ras_window_start);
879 /* The initial ras_window_len is set to the request size. To avoid
880 * uselessly reading and discarding pages for random IO the window is
881 * only increased once per consecutive request received. */
882 if ((ras->ras_consecutive_requests > 1 || stride_detect) &&
883 !ras->ras_request_index)
884 ras_increase_window(inode, ras, ra);
888 ras->ras_request_index++;
889 spin_unlock(&ras->ras_lock);
893 int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
895 struct inode *inode = vmpage->mapping->host;
896 struct ll_inode_info *lli = ll_i2info(inode);
899 struct cl_page *page;
900 struct cl_object *clob;
901 bool redirtied = false;
902 bool unlocked = false;
907 LASSERT(PageLocked(vmpage));
908 LASSERT(!PageWriteback(vmpage));
910 LASSERT(ll_i2dtexp(inode) != NULL);
912 env = cl_env_get(&refcheck);
914 GOTO(out, result = PTR_ERR(env));
916 clob = ll_i2info(inode)->lli_clob;
917 LASSERT(clob != NULL);
919 io = vvp_env_thread_io(env);
921 io->ci_ignore_layout = 1;
922 result = cl_io_init(env, io, CIT_MISC, clob);
924 page = cl_page_find(env, clob, vmpage->index,
925 vmpage, CPT_CACHEABLE);
927 lu_ref_add(&page->cp_reference, "writepage",
929 cl_page_assume(env, io, page);
930 result = cl_page_flush(env, io, page);
933 * Re-dirty page on error so it retries write,
934 * but not in case when IO has actually
935 * occurred and completed with an error.
937 if (!PageError(vmpage)) {
938 redirty_page_for_writepage(wbc, vmpage);
943 cl_page_disown(env, io, page);
945 lu_ref_del(&page->cp_reference,
946 "writepage", current);
947 cl_page_put(env, page);
949 result = PTR_ERR(page);
954 if (redirtied && wbc->sync_mode == WB_SYNC_ALL) {
955 loff_t offset = cl_offset(clob, vmpage->index);
957 /* Flush page failed because the extent is being written out.
958 * Wait for the write of extent to be finished to avoid
959 * breaking kernel which assumes ->writepage should mark
960 * PageWriteback or clean the page. */
961 result = cl_sync_file_range(inode, offset,
962 offset + PAGE_SIZE - 1,
965 /* actually we may have written more than one page.
966 * decreasing this page because the caller will count
968 wbc->nr_to_write -= result - 1;
973 cl_env_put(env, &refcheck);
978 if (!lli->lli_async_rc)
979 lli->lli_async_rc = result;
980 SetPageError(vmpage);
987 int ll_writepages(struct address_space *mapping, struct writeback_control *wbc)
989 struct inode *inode = mapping->host;
992 enum cl_fsync_mode mode;
997 if (wbc->range_cyclic) {
998 start = mapping->writeback_index << PAGE_SHIFT;
999 end = OBD_OBJECT_EOF;
1001 start = wbc->range_start;
1002 end = wbc->range_end;
1003 if (end == LLONG_MAX) {
1004 end = OBD_OBJECT_EOF;
1005 range_whole = start == 0;
1009 mode = CL_FSYNC_NONE;
1010 if (wbc->sync_mode == WB_SYNC_ALL)
1011 mode = CL_FSYNC_LOCAL;
1013 if (ll_i2info(inode)->lli_clob == NULL)
1016 /* for directio, it would call writepages() to evict cached pages
1017 * inside the IO context of write, which will cause deadlock at
1018 * layout_conf since it waits for active IOs to complete. */
1019 result = cl_sync_file_range(inode, start, end, mode, 1);
1021 wbc->nr_to_write -= result;
1025 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) {
1026 if (end == OBD_OBJECT_EOF)
1027 mapping->writeback_index = 0;
1029 mapping->writeback_index = (end >> PAGE_SHIFT) + 1;
1034 struct ll_cl_context *ll_cl_find(struct file *file)
1036 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1037 struct ll_cl_context *lcc;
1038 struct ll_cl_context *found = NULL;
1040 read_lock(&fd->fd_lock);
1041 list_for_each_entry(lcc, &fd->fd_lccs, lcc_list) {
1042 if (lcc->lcc_cookie == current) {
1047 read_unlock(&fd->fd_lock);
1052 void ll_cl_add(struct file *file, const struct lu_env *env, struct cl_io *io,
1055 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1056 struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1058 memset(lcc, 0, sizeof(*lcc));
1059 INIT_LIST_HEAD(&lcc->lcc_list);
1060 lcc->lcc_cookie = current;
1063 lcc->lcc_type = type;
1065 write_lock(&fd->fd_lock);
1066 list_add(&lcc->lcc_list, &fd->fd_lccs);
1067 write_unlock(&fd->fd_lock);
1070 void ll_cl_remove(struct file *file, const struct lu_env *env)
1072 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1073 struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1075 write_lock(&fd->fd_lock);
1076 list_del_init(&lcc->lcc_list);
1077 write_unlock(&fd->fd_lock);
1080 int ll_io_read_page(const struct lu_env *env, struct cl_io *io,
1081 struct cl_page *page, struct file *file)
1083 struct inode *inode = vvp_object_inode(page->cp_obj);
1084 struct ll_sb_info *sbi = ll_i2sbi(inode);
1085 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1086 struct ll_readahead_state *ras = &fd->fd_ras;
1087 struct cl_2queue *queue = &io->ci_queue;
1088 struct cl_sync_io *anchor = NULL;
1089 struct vvp_page *vpg;
1094 vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1095 uptodate = vpg->vpg_defer_uptodate;
1097 if (sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
1098 sbi->ll_ra_info.ra_max_pages > 0 &&
1099 !vpg->vpg_ra_updated) {
1100 struct vvp_io *vio = vvp_env_io(env);
1101 enum ras_update_flags flags = 0;
1104 flags |= LL_RAS_HIT;
1105 if (!vio->vui_ra_valid)
1106 flags |= LL_RAS_MMAP;
1107 ras_update(sbi, inode, ras, vvp_index(vpg), flags);
1110 cl_2queue_init(queue);
1112 vpg->vpg_ra_used = 1;
1113 cl_page_export(env, page, 1);
1114 cl_page_disown(env, io, page);
1116 anchor = &vvp_env_info(env)->vti_anchor;
1117 cl_sync_io_init(anchor, 1, &cl_sync_io_end);
1118 page->cp_sync_io = anchor;
1120 cl_2queue_add(queue, page);
1123 if (sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
1124 sbi->ll_ra_info.ra_max_pages > 0) {
1127 rc2 = ll_readahead(env, io, &queue->c2_qin, ras,
1129 CDEBUG(D_READA, DFID "%d pages read ahead at %lu\n",
1130 PFID(ll_inode2fid(inode)), rc2, vvp_index(vpg));
1133 if (queue->c2_qin.pl_nr > 0) {
1134 int count = queue->c2_qin.pl_nr;
1135 rc = cl_io_submit_rw(env, io, CRT_READ, queue);
1137 task_io_account_read(PAGE_SIZE * count);
1141 if (anchor != NULL && !cl_page_is_owned(page, io)) { /* have sent */
1142 rc = cl_sync_io_wait(env, anchor, 0);
1144 cl_page_assume(env, io, page);
1145 cl_page_list_del(env, &queue->c2_qout, page);
1147 if (!PageUptodate(cl_page_vmpage(page))) {
1148 /* Failed to read a mirror, discard this page so that
1149 * new page can be created with new mirror.
1151 * TODO: this is not needed after page reinit
1152 * route is implemented */
1153 cl_page_discard(env, io, page);
1155 cl_page_disown(env, io, page);
1158 /* TODO: discard all pages until page reinit route is implemented */
1159 cl_page_list_discard(env, io, &queue->c2_qin);
1161 /* Unlock unsent read pages in case of error. */
1162 cl_page_list_disown(env, io, &queue->c2_qin);
1164 cl_2queue_fini(env, queue);
1169 int ll_readpage(struct file *file, struct page *vmpage)
1171 struct inode *inode = file_inode(file);
1172 struct cl_object *clob = ll_i2info(inode)->lli_clob;
1173 struct ll_cl_context *lcc;
1174 const struct lu_env *env = NULL;
1175 struct cl_io *io = NULL;
1176 struct cl_page *page;
1180 lcc = ll_cl_find(file);
1186 if (io == NULL) { /* fast read */
1187 struct inode *inode = file_inode(file);
1188 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1189 struct ll_readahead_state *ras = &fd->fd_ras;
1190 struct lu_env *local_env = NULL;
1191 struct vvp_page *vpg;
1195 /* TODO: need to verify the layout version to make sure
1196 * the page is not invalid due to layout change. */
1197 page = cl_vmpage_page(vmpage, clob);
1199 unlock_page(vmpage);
1204 local_env = cl_env_percpu_get();
1208 vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1209 if (vpg->vpg_defer_uptodate) {
1210 enum ras_update_flags flags = LL_RAS_HIT;
1212 if (lcc && lcc->lcc_type == LCC_MMAP)
1213 flags |= LL_RAS_MMAP;
1215 /* For fast read, it updates read ahead state only
1216 * if the page is hit in cache because non cache page
1217 * case will be handled by slow read later. */
1218 ras_update(ll_i2sbi(inode), inode, ras, vvp_index(vpg),
1220 /* avoid duplicate ras_update() call */
1221 vpg->vpg_ra_updated = 1;
1223 /* Check if we can issue a readahead RPC, if that is
1224 * the case, we can't do fast IO because we will need
1225 * a cl_io to issue the RPC. */
1226 if (ras->ras_window_start + ras->ras_window_len <
1227 ras->ras_next_readahead + PTLRPC_MAX_BRW_PAGES) {
1228 /* export the page and skip io stack */
1229 vpg->vpg_ra_used = 1;
1230 cl_page_export(env, page, 1);
1235 /* release page refcount before unlocking the page to ensure
1236 * the object won't be destroyed in the calling path of
1237 * cl_page_put(). Please see comment in ll_releasepage(). */
1238 cl_page_put(env, page);
1239 unlock_page(vmpage);
1241 cl_env_percpu_put(local_env);
1246 LASSERT(io->ci_state == CIS_IO_GOING);
1247 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
1248 if (!IS_ERR(page)) {
1249 LASSERT(page->cp_type == CPT_CACHEABLE);
1250 if (likely(!PageUptodate(vmpage))) {
1251 cl_page_assume(env, io, page);
1253 result = ll_io_read_page(env, io, page, file);
1255 /* Page from a non-object file. */
1256 unlock_page(vmpage);
1259 cl_page_put(env, page);
1261 unlock_page(vmpage);
1262 result = PTR_ERR(page);