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/file.h>
48 #include <linux/stat.h>
49 #include <asm/uaccess.h>
51 #include <linux/pagemap.h>
52 /* current_is_kswapd() */
53 #include <linux/swap.h>
54 #include <linux/task_io_accounting_ops.h>
56 #define DEBUG_SUBSYSTEM S_LLITE
58 #include <obd_cksum.h>
59 #include "llite_internal.h"
60 #include <lustre_compat.h>
62 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which);
65 * Get readahead pages from the filesystem readahead pool of the client for a
68 * /param sbi superblock for filesystem readahead state ll_ra_info
69 * /param ria per-thread readahead state
70 * /param pages number of pages requested for readahead for the thread.
72 * WARNING: This algorithm is used to reduce contention on sbi->ll_lock.
73 * It should work well if the ra_max_pages is much greater than the single
74 * file's read-ahead window, and not too many threads contending for
75 * these readahead pages.
77 * TODO: There may be a 'global sync problem' if many threads are trying
78 * to get an ra budget that is larger than the remaining readahead pages
79 * and reach here at exactly the same time. They will compute /a ret to
80 * consume the remaining pages, but will fail at atomic_add_return() and
81 * get a zero ra window, although there is still ra space remaining. - Jay */
83 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi,
84 struct ra_io_arg *ria,
85 unsigned long pages, unsigned long min)
87 struct ll_ra_info *ra = &sbi->ll_ra_info;
91 /* If read-ahead pages left are less than 1M, do not do read-ahead,
92 * otherwise it will form small read RPC(< 1M), which hurt server
93 * performance a lot. */
94 ret = min(ra->ra_max_pages - atomic_read(&ra->ra_cur_pages),
96 if (ret < 0 || ret < min_t(long, PTLRPC_MAX_BRW_PAGES, pages))
99 if (atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) {
100 atomic_sub(ret, &ra->ra_cur_pages);
106 /* override ra limit for maximum performance */
107 atomic_add(min - ret, &ra->ra_cur_pages);
113 void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
115 struct ll_ra_info *ra = &sbi->ll_ra_info;
116 atomic_sub(len, &ra->ra_cur_pages);
119 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which)
121 LASSERTF(which < _NR_RA_STAT, "which: %u\n", which);
122 lprocfs_counter_incr(sbi->ll_ra_stats, which);
125 void ll_ra_stats_inc(struct inode *inode, enum ra_stat which)
127 struct ll_sb_info *sbi = ll_i2sbi(inode);
128 ll_ra_stats_inc_sbi(sbi, which);
131 #define RAS_CDEBUG(ras) \
133 "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu rpc %lu " \
134 "r %lu ri %lu csr %lu sf %lu sp %lu sl %lu lr %lu\n", \
135 ras->ras_last_readpage, ras->ras_consecutive_requests, \
136 ras->ras_consecutive_pages, ras->ras_window_start, \
137 ras->ras_window_len, ras->ras_next_readahead, \
139 ras->ras_requests, ras->ras_request_index, \
140 ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
141 ras->ras_stride_pages, ras->ras_stride_length, \
142 ras->ras_async_last_readpage)
144 static int index_in_window(unsigned long index, unsigned long point,
145 unsigned long before, unsigned long after)
147 unsigned long start = point - before, end = point + after;
154 return start <= index && index <= end;
157 void ll_ras_enter(struct file *f)
159 struct ll_file_data *fd = LUSTRE_FPRIVATE(f);
160 struct ll_readahead_state *ras = &fd->fd_ras;
162 spin_lock(&ras->ras_lock);
164 ras->ras_request_index = 0;
165 ras->ras_consecutive_requests++;
166 spin_unlock(&ras->ras_lock);
170 * Initiates read-ahead of a page with given index.
172 * \retval +ve: page was already uptodate so it will be skipped
174 * \retval -ve: page wasn't added to \a queue for error;
175 * \retval 0: page was added into \a queue for read ahead.
177 static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io,
178 struct cl_page_list *queue, pgoff_t index)
180 struct cl_object *clob = io->ci_obj;
181 struct inode *inode = vvp_object_inode(clob);
183 struct cl_page *page;
184 struct vvp_page *vpg;
185 enum ra_stat which = _NR_RA_STAT; /* keep gcc happy */
187 const char *msg = NULL;
190 vmpage = grab_cache_page_nowait(inode->i_mapping, index);
191 if (vmpage == NULL) {
192 which = RA_STAT_FAILED_GRAB_PAGE;
193 msg = "g_c_p_n failed";
194 GOTO(out, rc = -EBUSY);
197 /* Check if vmpage was truncated or reclaimed */
198 if (vmpage->mapping != inode->i_mapping) {
199 which = RA_STAT_WRONG_GRAB_PAGE;
200 msg = "g_c_p_n returned invalid page";
201 GOTO(out, rc = -EBUSY);
204 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
206 which = RA_STAT_FAILED_GRAB_PAGE;
207 msg = "cl_page_find failed";
208 GOTO(out, rc = PTR_ERR(page));
211 lu_ref_add(&page->cp_reference, "ra", current);
212 cl_page_assume(env, io, page);
213 vpg = cl2vvp_page(cl_object_page_slice(clob, page));
214 if (!vpg->vpg_defer_uptodate && !PageUptodate(vmpage)) {
215 vpg->vpg_defer_uptodate = 1;
216 vpg->vpg_ra_used = 0;
217 cl_page_list_add(queue, page);
219 /* skip completed pages */
220 cl_page_unassume(env, io, page);
221 /* This page is already uptodate, returning a positive number
222 * to tell the callers about this */
226 lu_ref_del(&page->cp_reference, "ra", current);
227 cl_page_put(env, page);
230 if (vmpage != NULL) {
236 ll_ra_stats_inc(inode, which);
237 CDEBUG(D_READA, "%s\n", msg);
244 #define RIA_DEBUG(ria) \
245 CDEBUG(D_READA, "rs %lu re %lu ro %lu rl %lu rp %lu\n", \
246 ria->ria_start, ria->ria_end, ria->ria_stoff, ria->ria_length,\
249 static inline int stride_io_mode(struct ll_readahead_state *ras)
251 return ras->ras_consecutive_stride_requests > 1;
254 /* The function calculates how much pages will be read in
255 * [off, off + length], in such stride IO area,
256 * stride_offset = st_off, stride_lengh = st_len,
257 * stride_pages = st_pgs
259 * |------------------|*****|------------------|*****|------------|*****|....
262 * |----- st_len -----|
264 * How many pages it should read in such pattern
265 * |-------------------------------------------------------------|
267 * |<------ length ------->|
269 * = |<----->| + |-------------------------------------| + |---|
270 * start_left st_pgs * i end_left
273 stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs,
274 unsigned long off, unsigned long length)
276 __u64 start = off > st_off ? off - st_off : 0;
277 __u64 end = off + length > st_off ? off + length - st_off : 0;
278 unsigned long start_left = 0;
279 unsigned long end_left = 0;
280 unsigned long pg_count;
282 if (st_len == 0 || length == 0 || end == 0)
285 start_left = do_div(start, st_len);
286 if (start_left < st_pgs)
287 start_left = st_pgs - start_left;
291 end_left = do_div(end, st_len);
292 if (end_left > st_pgs)
295 CDEBUG(D_READA, "start %llu, end %llu start_left %lu end_left %lu\n",
296 start, end, start_left, end_left);
299 pg_count = end_left - (st_pgs - start_left);
301 pg_count = start_left + st_pgs * (end - start - 1) + end_left;
303 CDEBUG(D_READA, "st_off %lu, st_len %lu st_pgs %lu off %lu length %lu"
304 "pgcount %lu\n", st_off, st_len, st_pgs, off, length, pg_count);
309 static int ria_page_count(struct ra_io_arg *ria)
311 __u64 length = ria->ria_end >= ria->ria_start ?
312 ria->ria_end - ria->ria_start + 1 : 0;
314 return stride_pg_count(ria->ria_stoff, ria->ria_length,
315 ria->ria_pages, ria->ria_start,
319 static unsigned long ras_align(struct ll_readahead_state *ras,
321 unsigned long *remainder)
323 unsigned long rem = index % ras->ras_rpc_size;
324 if (remainder != NULL)
329 /*Check whether the index is in the defined ra-window */
330 static int ras_inside_ra_window(unsigned long idx, struct ra_io_arg *ria)
332 /* If ria_length == ria_pages, it means non-stride I/O mode,
333 * idx should always inside read-ahead window in this case
334 * For stride I/O mode, just check whether the idx is inside
336 return ria->ria_length == 0 || ria->ria_length == ria->ria_pages ||
337 (idx >= ria->ria_stoff && (idx - ria->ria_stoff) %
338 ria->ria_length < ria->ria_pages);
342 ll_read_ahead_pages(const struct lu_env *env, struct cl_io *io,
343 struct cl_page_list *queue, struct ll_readahead_state *ras,
344 struct ra_io_arg *ria, pgoff_t *ra_end)
346 struct cl_read_ahead ra = { 0 };
347 int rc = 0, count = 0;
351 LASSERT(ria != NULL);
354 stride_ria = ria->ria_length > ria->ria_pages && ria->ria_pages > 0;
355 for (page_idx = ria->ria_start;
356 page_idx <= ria->ria_end && ria->ria_reserved > 0; page_idx++) {
357 if (ras_inside_ra_window(page_idx, ria)) {
358 if (ra.cra_end == 0 || ra.cra_end < page_idx) {
361 cl_read_ahead_release(env, &ra);
363 rc = cl_io_read_ahead(env, io, page_idx, &ra);
367 /* Do not shrink the ria_end at any case until
368 * the minimum end of current read is covered.
369 * And only shrink the ria_end if the matched
370 * LDLM lock doesn't cover more. */
371 if (page_idx > ra.cra_end ||
372 (ra.cra_contention &&
373 page_idx > ria->ria_end_min)) {
374 ria->ria_end = ra.cra_end;
378 CDEBUG(D_READA, "idx: %lu, ra: %lu, rpc: %lu\n",
379 page_idx, ra.cra_end, ra.cra_rpc_size);
380 LASSERTF(ra.cra_end >= page_idx,
381 "object: %p, indcies %lu / %lu\n",
382 io->ci_obj, ra.cra_end, page_idx);
383 /* update read ahead RPC size.
384 * NB: it's racy but doesn't matter */
385 if (ras->ras_rpc_size != ra.cra_rpc_size &&
387 ras->ras_rpc_size = ra.cra_rpc_size;
388 /* trim it to align with optimal RPC size */
389 end = ras_align(ras, ria->ria_end + 1, NULL);
390 if (end > 0 && !ria->ria_eof)
391 ria->ria_end = end - 1;
392 if (ria->ria_end < ria->ria_end_min)
393 ria->ria_end = ria->ria_end_min;
395 if (page_idx > ria->ria_end)
398 /* If the page is inside the read-ahead window */
399 rc = ll_read_ahead_page(env, io, queue, page_idx);
404 /* Only subtract from reserve & count the page if we
405 * really did readahead on that page. */
410 } else if (stride_ria) {
411 /* If it is not in the read-ahead window, and it is
412 * read-ahead mode, then check whether it should skip
416 /* NOTE: This assertion only is valid when it is for
417 * forward read-ahead, must adjust if backward
418 * readahead is implemented.
420 LASSERTF(page_idx >= ria->ria_stoff,
421 "Invalid page_idx %lu rs %lu re %lu ro %lu "
422 "rl %lu rp %lu\n", page_idx,
423 ria->ria_start, ria->ria_end, ria->ria_stoff,
424 ria->ria_length, ria->ria_pages);
425 offset = page_idx - ria->ria_stoff;
426 offset = offset % (ria->ria_length);
427 if (offset >= ria->ria_pages) {
428 page_idx += ria->ria_length - offset - 1;
430 "Stride: jump %lu pages to %lu\n",
431 ria->ria_length - offset, page_idx);
437 cl_read_ahead_release(env, &ra);
442 static void ll_readahead_work_free(struct ll_readahead_work *work)
444 fput(work->lrw_file);
448 static void ll_readahead_handle_work(struct work_struct *wq);
449 static void ll_readahead_work_add(struct inode *inode,
450 struct ll_readahead_work *work)
452 INIT_WORK(&work->lrw_readahead_work, ll_readahead_handle_work);
453 queue_work(ll_i2sbi(inode)->ll_ra_info.ll_readahead_wq,
454 &work->lrw_readahead_work);
457 static int ll_readahead_file_kms(const struct lu_env *env,
458 struct cl_io *io, __u64 *kms)
460 struct cl_object *clob;
462 struct cl_attr *attr = vvp_env_thread_attr(env);
466 inode = vvp_object_inode(clob);
468 cl_object_attr_lock(clob);
469 ret = cl_object_attr_get(env, clob, attr);
470 cl_object_attr_unlock(clob);
475 *kms = attr->cat_kms;
479 static void ll_readahead_handle_work(struct work_struct *wq)
481 struct ll_readahead_work *work;
484 struct ra_io_arg *ria;
486 struct ll_file_data *fd;
487 struct ll_readahead_state *ras;
489 struct cl_2queue *queue;
491 unsigned long len, mlen = 0;
495 unsigned long end_index;
497 work = container_of(wq, struct ll_readahead_work,
499 fd = LUSTRE_FPRIVATE(work->lrw_file);
501 file = work->lrw_file;
502 inode = file_inode(file);
504 env = cl_env_alloc(&refcheck, LCT_NOREF);
506 GOTO(out_free_work, rc = PTR_ERR(env));
508 io = vvp_env_thread_io(env);
509 ll_io_init(io, file, CIT_READ, NULL);
511 rc = ll_readahead_file_kms(env, io, &kms);
513 GOTO(out_put_env, rc);
516 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
517 GOTO(out_put_env, rc = 0);
520 ria = &ll_env_info(env)->lti_ria;
521 memset(ria, 0, sizeof(*ria));
523 ria->ria_start = work->lrw_start;
524 /* Truncate RA window to end of file */
525 end_index = (unsigned long)((kms - 1) >> PAGE_SHIFT);
526 if (end_index <= work->lrw_end) {
527 work->lrw_end = end_index;
530 if (work->lrw_end <= work->lrw_start)
531 GOTO(out_put_env, rc = 0);
533 ria->ria_end = work->lrw_end;
534 len = ria->ria_end - ria->ria_start + 1;
535 ria->ria_reserved = ll_ra_count_get(ll_i2sbi(inode), ria,
536 ria_page_count(ria), mlen);
539 "async reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
540 ria->ria_reserved, len, mlen,
541 atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
542 ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
544 if (ria->ria_reserved < len) {
545 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
546 if (PAGES_TO_MiB(ria->ria_reserved) < 1) {
547 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
548 GOTO(out_put_env, rc = 0);
552 rc = cl_io_rw_init(env, io, CIT_READ, ria->ria_start, len);
554 GOTO(out_put_env, rc);
556 vvp_env_io(env)->vui_io_subtype = IO_NORMAL;
557 vvp_env_io(env)->vui_fd = fd;
558 io->ci_state = CIS_LOCKED;
559 io->ci_async_readahead = true;
560 rc = cl_io_start(env, io);
562 GOTO(out_io_fini, rc);
564 queue = &io->ci_queue;
565 cl_2queue_init(queue);
567 rc = ll_read_ahead_pages(env, io, &queue->c2_qin, ras, ria, &ra_end);
568 if (ria->ria_reserved != 0)
569 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
570 if (queue->c2_qin.pl_nr > 0) {
571 int count = queue->c2_qin.pl_nr;
573 rc = cl_io_submit_rw(env, io, CRT_READ, queue);
575 task_io_account_read(PAGE_SIZE * count);
577 if (ria->ria_end == ra_end && ra_end == (kms >> PAGE_SHIFT))
578 ll_ra_stats_inc(inode, RA_STAT_EOF);
580 if (ra_end != ria->ria_end)
581 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
583 /* TODO: discard all pages until page reinit route is implemented */
584 cl_page_list_discard(env, io, &queue->c2_qin);
586 /* Unlock unsent read pages in case of error. */
587 cl_page_list_disown(env, io, &queue->c2_qin);
589 cl_2queue_fini(env, queue);
594 cl_env_put(env, &refcheck);
597 ll_ra_stats_inc_sbi(ll_i2sbi(inode), RA_STAT_ASYNC);
598 ll_readahead_work_free(work);
601 static int ll_readahead(const struct lu_env *env, struct cl_io *io,
602 struct cl_page_list *queue,
603 struct ll_readahead_state *ras, bool hit,
606 struct vvp_io *vio = vvp_env_io(env);
607 struct ll_thread_info *lti = ll_env_info(env);
608 unsigned long len, mlen = 0;
609 pgoff_t ra_end = 0, start = 0, end = 0;
611 struct ra_io_arg *ria = <i->lti_ria;
612 struct cl_object *clob;
618 inode = vvp_object_inode(clob);
620 memset(ria, 0, sizeof *ria);
621 ret = ll_readahead_file_kms(env, io, &kms);
626 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
630 spin_lock(&ras->ras_lock);
633 * Note: other thread might rollback the ras_next_readahead,
634 * if it can not get the full size of prepared pages, see the
635 * end of this function. For stride read ahead, it needs to
636 * make sure the offset is no less than ras_stride_offset,
637 * so that stride read ahead can work correctly.
639 if (stride_io_mode(ras))
640 start = max(ras->ras_next_readahead, ras->ras_stride_offset);
642 start = ras->ras_next_readahead;
644 if (ras->ras_window_len > 0)
645 end = ras->ras_window_start + ras->ras_window_len - 1;
647 /* Enlarge the RA window to encompass the full read */
648 if (vio->vui_ra_valid &&
649 end < vio->vui_ra_start + vio->vui_ra_count - 1)
650 end = vio->vui_ra_start + vio->vui_ra_count - 1;
653 unsigned long end_index;
655 /* Truncate RA window to end of file */
656 end_index = (unsigned long)((kms - 1) >> PAGE_SHIFT);
657 if (end_index <= end) {
662 ria->ria_start = start;
664 /* If stride I/O mode is detected, get stride window*/
665 if (stride_io_mode(ras)) {
666 ria->ria_stoff = ras->ras_stride_offset;
667 ria->ria_length = ras->ras_stride_length;
668 ria->ria_pages = ras->ras_stride_pages;
670 spin_unlock(&ras->ras_lock);
673 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
676 len = ria_page_count(ria);
678 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
683 CDEBUG(D_READA, DFID": ria: %lu/%lu, bead: %lu/%lu, hit: %d\n",
684 PFID(lu_object_fid(&clob->co_lu)),
685 ria->ria_start, ria->ria_end,
686 vio->vui_ra_valid ? vio->vui_ra_start : 0,
687 vio->vui_ra_valid ? vio->vui_ra_count : 0,
690 /* at least to extend the readahead window to cover current read */
691 if (!hit && vio->vui_ra_valid &&
692 vio->vui_ra_start + vio->vui_ra_count > ria->ria_start)
693 ria->ria_end_min = vio->vui_ra_start + vio->vui_ra_count - 1;
695 ria->ria_reserved = ll_ra_count_get(ll_i2sbi(inode), ria, len, mlen);
696 if (ria->ria_reserved < len)
697 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
699 CDEBUG(D_READA, "reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
700 ria->ria_reserved, len, mlen,
701 atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
702 ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
704 ret = ll_read_ahead_pages(env, io, queue, ras, ria, &ra_end);
706 if (ria->ria_reserved != 0)
707 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
709 if (ra_end == end && ra_end == (kms >> PAGE_SHIFT))
710 ll_ra_stats_inc(inode, RA_STAT_EOF);
712 CDEBUG(D_READA, "ra_end = %lu end = %lu stride end = %lu pages = %d\n",
713 ra_end, end, ria->ria_end, ret);
716 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
718 /* update the ras so that the next read-ahead tries from
719 * where we left off. */
720 spin_lock(&ras->ras_lock);
721 ras->ras_next_readahead = ra_end + 1;
722 spin_unlock(&ras->ras_lock);
729 static void ras_set_start(struct inode *inode, struct ll_readahead_state *ras,
732 ras->ras_window_start = ras_align(ras, index, NULL);
735 /* called with the ras_lock held or from places where it doesn't matter */
736 static void ras_reset(struct inode *inode, struct ll_readahead_state *ras,
739 ras->ras_last_readpage = index;
740 ras->ras_consecutive_requests = 0;
741 ras->ras_consecutive_pages = 0;
742 ras->ras_window_len = 0;
743 ras_set_start(inode, ras, index);
744 ras->ras_next_readahead = max(ras->ras_window_start, index + 1);
749 /* called with the ras_lock held or from places where it doesn't matter */
750 static void ras_stride_reset(struct ll_readahead_state *ras)
752 ras->ras_consecutive_stride_requests = 0;
753 ras->ras_stride_length = 0;
754 ras->ras_stride_pages = 0;
758 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
760 spin_lock_init(&ras->ras_lock);
761 ras->ras_rpc_size = PTLRPC_MAX_BRW_PAGES;
762 ras_reset(inode, ras, 0);
763 ras->ras_requests = 0;
767 * Check whether the read request is in the stride window.
768 * If it is in the stride window, return 1, otherwise return 0.
770 static int index_in_stride_window(struct ll_readahead_state *ras,
773 unsigned long stride_gap;
775 if (ras->ras_stride_length == 0 || ras->ras_stride_pages == 0 ||
776 ras->ras_stride_pages == ras->ras_stride_length)
779 stride_gap = index - ras->ras_last_readpage - 1;
781 /* If it is contiguous read */
783 return ras->ras_consecutive_pages + 1 <= ras->ras_stride_pages;
785 /* Otherwise check the stride by itself */
786 return (ras->ras_stride_length - ras->ras_stride_pages) == stride_gap &&
787 ras->ras_consecutive_pages == ras->ras_stride_pages;
790 static void ras_update_stride_detector(struct ll_readahead_state *ras,
793 unsigned long stride_gap = index - ras->ras_last_readpage - 1;
795 if (!stride_io_mode(ras) && (stride_gap != 0 ||
796 ras->ras_consecutive_stride_requests == 0)) {
797 ras->ras_stride_pages = ras->ras_consecutive_pages;
798 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
800 LASSERT(ras->ras_request_index == 0);
801 LASSERT(ras->ras_consecutive_stride_requests == 0);
803 if (index <= ras->ras_last_readpage) {
804 /*Reset stride window for forward read*/
805 ras_stride_reset(ras);
809 ras->ras_stride_pages = ras->ras_consecutive_pages;
810 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
817 stride_page_count(struct ll_readahead_state *ras, unsigned long len)
819 return stride_pg_count(ras->ras_stride_offset, ras->ras_stride_length,
820 ras->ras_stride_pages, ras->ras_stride_offset,
824 /* Stride Read-ahead window will be increased inc_len according to
825 * stride I/O pattern */
826 static void ras_stride_increase_window(struct ll_readahead_state *ras,
827 struct ll_ra_info *ra,
828 unsigned long inc_len)
830 unsigned long left, step, window_len;
831 unsigned long stride_len;
833 LASSERT(ras->ras_stride_length > 0);
834 LASSERTF(ras->ras_window_start + ras->ras_window_len
835 >= ras->ras_stride_offset, "window_start %lu, window_len %lu"
836 " stride_offset %lu\n", ras->ras_window_start,
837 ras->ras_window_len, ras->ras_stride_offset);
839 stride_len = ras->ras_window_start + ras->ras_window_len -
840 ras->ras_stride_offset;
842 left = stride_len % ras->ras_stride_length;
843 window_len = ras->ras_window_len - left;
845 if (left < ras->ras_stride_pages)
848 left = ras->ras_stride_pages + inc_len;
850 LASSERT(ras->ras_stride_pages != 0);
852 step = left / ras->ras_stride_pages;
853 left %= ras->ras_stride_pages;
855 window_len += step * ras->ras_stride_length + left;
857 if (stride_page_count(ras, window_len) <= ra->ra_max_pages_per_file)
858 ras->ras_window_len = window_len;
863 static void ras_increase_window(struct inode *inode,
864 struct ll_readahead_state *ras,
865 struct ll_ra_info *ra)
867 /* The stretch of ra-window should be aligned with max rpc_size
868 * but current clio architecture does not support retrieve such
869 * information from lower layer. FIXME later
871 if (stride_io_mode(ras)) {
872 ras_stride_increase_window(ras, ra, ras->ras_rpc_size);
876 wlen = min(ras->ras_window_len + ras->ras_rpc_size,
877 ra->ra_max_pages_per_file);
878 if (wlen < ras->ras_rpc_size)
879 ras->ras_window_len = wlen;
881 ras->ras_window_len = ras_align(ras, wlen, NULL);
885 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
886 struct ll_readahead_state *ras, unsigned long index,
887 enum ras_update_flags flags)
889 struct ll_ra_info *ra = &sbi->ll_ra_info;
890 bool hit = flags & LL_RAS_HIT;
891 int zero = 0, stride_detect = 0, ra_miss = 0;
894 spin_lock(&ras->ras_lock);
897 CDEBUG(D_READA, DFID " pages at %lu miss.\n",
898 PFID(ll_inode2fid(inode)), index);
899 ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
901 /* reset the read-ahead window in two cases. First when the app seeks
902 * or reads to some other part of the file. Secondly if we get a
903 * read-ahead miss that we think we've previously issued. This can
904 * be a symptom of there being so many read-ahead pages that the VM is
905 * reclaiming it before we get to it. */
906 if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
908 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
909 } else if (!hit && ras->ras_window_len &&
910 index < ras->ras_next_readahead &&
911 index_in_window(index, ras->ras_window_start, 0,
912 ras->ras_window_len)) {
914 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
917 /* On the second access to a file smaller than the tunable
918 * ra_max_read_ahead_whole_pages trigger RA on all pages in the
919 * file up to ra_max_pages_per_file. This is simply a best effort
920 * and only occurs once per open file. Normal RA behavior is reverted
921 * to for subsequent IO. The mmap case does not increment
922 * ras_requests and thus can never trigger this behavior. */
923 if (ras->ras_requests >= 2 && !ras->ras_request_index) {
926 kms_pages = (i_size_read(inode) + PAGE_SIZE - 1) >>
929 CDEBUG(D_READA, "kmsp %llu mwp %lu mp %lu\n", kms_pages,
930 ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages_per_file);
933 kms_pages <= ra->ra_max_read_ahead_whole_pages) {
934 ras->ras_window_start = 0;
935 ras->ras_next_readahead = index + 1;
936 ras->ras_window_len = min(ra->ra_max_pages_per_file,
937 ra->ra_max_read_ahead_whole_pages);
942 /* check whether it is in stride I/O mode*/
943 if (!index_in_stride_window(ras, index)) {
944 if (ras->ras_consecutive_stride_requests == 0 &&
945 ras->ras_request_index == 0) {
946 ras_update_stride_detector(ras, index);
947 ras->ras_consecutive_stride_requests++;
949 ras_stride_reset(ras);
951 ras_reset(inode, ras, index);
952 ras->ras_consecutive_pages++;
955 ras->ras_consecutive_pages = 0;
956 ras->ras_consecutive_requests = 0;
957 if (++ras->ras_consecutive_stride_requests > 1)
963 if (index_in_stride_window(ras, index) &&
964 stride_io_mode(ras)) {
965 if (index != ras->ras_last_readpage + 1)
966 ras->ras_consecutive_pages = 0;
967 ras_reset(inode, ras, index);
969 /* If stride-RA hit cache miss, the stride
970 * detector will not be reset to avoid the
971 * overhead of redetecting read-ahead mode,
972 * but on the condition that the stride window
973 * is still intersect with normal sequential
974 * read-ahead window. */
975 if (ras->ras_window_start <
976 ras->ras_stride_offset)
977 ras_stride_reset(ras);
980 /* Reset both stride window and normal RA
982 ras_reset(inode, ras, index);
983 ras->ras_consecutive_pages++;
984 ras_stride_reset(ras);
987 } else if (stride_io_mode(ras)) {
988 /* If this is contiguous read but in stride I/O mode
989 * currently, check whether stride step still is valid,
990 * if invalid, it will reset the stride ra window*/
991 if (!index_in_stride_window(ras, index)) {
992 /* Shrink stride read-ahead window to be zero */
993 ras_stride_reset(ras);
994 ras->ras_window_len = 0;
995 ras->ras_next_readahead = index;
999 ras->ras_consecutive_pages++;
1000 ras->ras_last_readpage = index;
1001 ras_set_start(inode, ras, index);
1003 if (stride_io_mode(ras)) {
1004 /* Since stride readahead is sentivite to the offset
1005 * of read-ahead, so we use original offset here,
1006 * instead of ras_window_start, which is RPC aligned */
1007 ras->ras_next_readahead = max(index + 1,
1008 ras->ras_next_readahead);
1009 ras->ras_window_start = max(ras->ras_stride_offset,
1010 ras->ras_window_start);
1012 if (ras->ras_next_readahead < ras->ras_window_start)
1013 ras->ras_next_readahead = ras->ras_window_start;
1015 ras->ras_next_readahead = index + 1;
1019 /* Trigger RA in the mmap case where ras_consecutive_requests
1020 * is not incremented and thus can't be used to trigger RA */
1021 if (ras->ras_consecutive_pages >= 4 && flags & LL_RAS_MMAP) {
1022 ras_increase_window(inode, ras, ra);
1023 /* reset consecutive pages so that the readahead window can
1024 * grow gradually. */
1025 ras->ras_consecutive_pages = 0;
1026 GOTO(out_unlock, 0);
1029 /* Initially reset the stride window offset to next_readahead*/
1030 if (ras->ras_consecutive_stride_requests == 2 && stride_detect) {
1032 * Once stride IO mode is detected, next_readahead should be
1033 * reset to make sure next_readahead > stride offset
1035 ras->ras_next_readahead = max(index, ras->ras_next_readahead);
1036 ras->ras_stride_offset = index;
1037 ras->ras_window_start = max(index, ras->ras_window_start);
1040 /* The initial ras_window_len is set to the request size. To avoid
1041 * uselessly reading and discarding pages for random IO the window is
1042 * only increased once per consecutive request received. */
1043 if ((ras->ras_consecutive_requests > 1 || stride_detect) &&
1044 !ras->ras_request_index)
1045 ras_increase_window(inode, ras, ra);
1049 ras->ras_request_index++;
1050 spin_unlock(&ras->ras_lock);
1054 int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
1056 struct inode *inode = vmpage->mapping->host;
1057 struct ll_inode_info *lli = ll_i2info(inode);
1060 struct cl_page *page;
1061 struct cl_object *clob;
1062 bool redirtied = false;
1063 bool unlocked = false;
1068 LASSERT(PageLocked(vmpage));
1069 LASSERT(!PageWriteback(vmpage));
1071 LASSERT(ll_i2dtexp(inode) != NULL);
1073 env = cl_env_get(&refcheck);
1075 GOTO(out, result = PTR_ERR(env));
1077 clob = ll_i2info(inode)->lli_clob;
1078 LASSERT(clob != NULL);
1080 io = vvp_env_thread_io(env);
1082 io->ci_ignore_layout = 1;
1083 result = cl_io_init(env, io, CIT_MISC, clob);
1085 page = cl_page_find(env, clob, vmpage->index,
1086 vmpage, CPT_CACHEABLE);
1087 if (!IS_ERR(page)) {
1088 lu_ref_add(&page->cp_reference, "writepage",
1090 cl_page_assume(env, io, page);
1091 result = cl_page_flush(env, io, page);
1094 * Re-dirty page on error so it retries write,
1095 * but not in case when IO has actually
1096 * occurred and completed with an error.
1098 if (!PageError(vmpage)) {
1099 redirty_page_for_writepage(wbc, vmpage);
1104 cl_page_disown(env, io, page);
1106 lu_ref_del(&page->cp_reference,
1107 "writepage", current);
1108 cl_page_put(env, page);
1110 result = PTR_ERR(page);
1113 cl_io_fini(env, io);
1115 if (redirtied && wbc->sync_mode == WB_SYNC_ALL) {
1116 loff_t offset = cl_offset(clob, vmpage->index);
1118 /* Flush page failed because the extent is being written out.
1119 * Wait for the write of extent to be finished to avoid
1120 * breaking kernel which assumes ->writepage should mark
1121 * PageWriteback or clean the page. */
1122 result = cl_sync_file_range(inode, offset,
1123 offset + PAGE_SIZE - 1,
1126 /* actually we may have written more than one page.
1127 * decreasing this page because the caller will count
1129 wbc->nr_to_write -= result - 1;
1134 cl_env_put(env, &refcheck);
1139 if (!lli->lli_async_rc)
1140 lli->lli_async_rc = result;
1141 SetPageError(vmpage);
1143 unlock_page(vmpage);
1148 int ll_writepages(struct address_space *mapping, struct writeback_control *wbc)
1150 struct inode *inode = mapping->host;
1153 enum cl_fsync_mode mode;
1154 int range_whole = 0;
1158 if (wbc->range_cyclic) {
1159 start = mapping->writeback_index << PAGE_SHIFT;
1160 end = OBD_OBJECT_EOF;
1162 start = wbc->range_start;
1163 end = wbc->range_end;
1164 if (end == LLONG_MAX) {
1165 end = OBD_OBJECT_EOF;
1166 range_whole = start == 0;
1170 mode = CL_FSYNC_NONE;
1171 if (wbc->sync_mode == WB_SYNC_ALL)
1172 mode = CL_FSYNC_LOCAL;
1174 if (ll_i2info(inode)->lli_clob == NULL)
1177 /* for directio, it would call writepages() to evict cached pages
1178 * inside the IO context of write, which will cause deadlock at
1179 * layout_conf since it waits for active IOs to complete. */
1180 result = cl_sync_file_range(inode, start, end, mode, 1);
1182 wbc->nr_to_write -= result;
1186 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) {
1187 if (end == OBD_OBJECT_EOF)
1188 mapping->writeback_index = 0;
1190 mapping->writeback_index = (end >> PAGE_SHIFT) + 1;
1195 struct ll_cl_context *ll_cl_find(struct file *file)
1197 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1198 struct ll_cl_context *lcc;
1199 struct ll_cl_context *found = NULL;
1201 read_lock(&fd->fd_lock);
1202 list_for_each_entry(lcc, &fd->fd_lccs, lcc_list) {
1203 if (lcc->lcc_cookie == current) {
1208 read_unlock(&fd->fd_lock);
1213 void ll_cl_add(struct file *file, const struct lu_env *env, struct cl_io *io,
1216 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1217 struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1219 memset(lcc, 0, sizeof(*lcc));
1220 INIT_LIST_HEAD(&lcc->lcc_list);
1221 lcc->lcc_cookie = current;
1224 lcc->lcc_type = type;
1226 write_lock(&fd->fd_lock);
1227 list_add(&lcc->lcc_list, &fd->fd_lccs);
1228 write_unlock(&fd->fd_lock);
1231 void ll_cl_remove(struct file *file, const struct lu_env *env)
1233 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1234 struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1236 write_lock(&fd->fd_lock);
1237 list_del_init(&lcc->lcc_list);
1238 write_unlock(&fd->fd_lock);
1241 int ll_io_read_page(const struct lu_env *env, struct cl_io *io,
1242 struct cl_page *page, struct file *file)
1244 struct inode *inode = vvp_object_inode(page->cp_obj);
1245 struct ll_sb_info *sbi = ll_i2sbi(inode);
1246 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1247 struct ll_readahead_state *ras = &fd->fd_ras;
1248 struct cl_2queue *queue = &io->ci_queue;
1249 struct cl_sync_io *anchor = NULL;
1250 struct vvp_page *vpg;
1255 vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1256 uptodate = vpg->vpg_defer_uptodate;
1258 if (sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
1259 sbi->ll_ra_info.ra_max_pages > 0 &&
1260 !vpg->vpg_ra_updated) {
1261 struct vvp_io *vio = vvp_env_io(env);
1262 enum ras_update_flags flags = 0;
1265 flags |= LL_RAS_HIT;
1266 if (!vio->vui_ra_valid)
1267 flags |= LL_RAS_MMAP;
1268 ras_update(sbi, inode, ras, vvp_index(vpg), flags);
1271 cl_2queue_init(queue);
1273 vpg->vpg_ra_used = 1;
1274 cl_page_export(env, page, 1);
1275 cl_page_disown(env, io, page);
1277 anchor = &vvp_env_info(env)->vti_anchor;
1278 cl_sync_io_init(anchor, 1);
1279 page->cp_sync_io = anchor;
1281 cl_2queue_add(queue, page);
1284 if (sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
1285 sbi->ll_ra_info.ra_max_pages > 0) {
1288 rc2 = ll_readahead(env, io, &queue->c2_qin, ras,
1290 CDEBUG(D_READA, DFID "%d pages read ahead at %lu\n",
1291 PFID(ll_inode2fid(inode)), rc2, vvp_index(vpg));
1294 if (queue->c2_qin.pl_nr > 0) {
1295 int count = queue->c2_qin.pl_nr;
1296 rc = cl_io_submit_rw(env, io, CRT_READ, queue);
1298 task_io_account_read(PAGE_SIZE * count);
1302 if (anchor != NULL && !cl_page_is_owned(page, io)) { /* have sent */
1303 rc = cl_sync_io_wait(env, anchor, 0);
1305 cl_page_assume(env, io, page);
1306 cl_page_list_del(env, &queue->c2_qout, page);
1308 if (!PageUptodate(cl_page_vmpage(page))) {
1309 /* Failed to read a mirror, discard this page so that
1310 * new page can be created with new mirror.
1312 * TODO: this is not needed after page reinit
1313 * route is implemented */
1314 cl_page_discard(env, io, page);
1316 cl_page_disown(env, io, page);
1319 /* TODO: discard all pages until page reinit route is implemented */
1320 cl_page_list_discard(env, io, &queue->c2_qin);
1322 /* Unlock unsent read pages in case of error. */
1323 cl_page_list_disown(env, io, &queue->c2_qin);
1325 cl_2queue_fini(env, queue);
1331 * Possible return value:
1332 * 0 no async readahead triggered and fast read could not be used.
1333 * 1 no async readahead, but fast read could be used.
1334 * 2 async readahead triggered and fast read could be used too.
1337 static int kickoff_async_readahead(struct file *file, unsigned long pages)
1339 struct ll_readahead_work *lrw;
1340 struct inode *inode = file_inode(file);
1341 struct ll_sb_info *sbi = ll_i2sbi(inode);
1342 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1343 struct ll_readahead_state *ras = &fd->fd_ras;
1344 struct ll_ra_info *ra = &sbi->ll_ra_info;
1345 unsigned long throttle;
1346 unsigned long start = ras_align(ras, ras->ras_next_readahead, NULL);
1347 unsigned long end = start + pages - 1;
1349 throttle = min(ra->ra_async_pages_per_file_threshold,
1350 ra->ra_max_pages_per_file);
1352 * If this is strided i/o or the window is smaller than the
1353 * throttle limit, we do not do async readahead. Otherwise,
1354 * we do async readahead, allowing the user thread to do fast i/o.
1356 if (stride_io_mode(ras) || !throttle ||
1357 ras->ras_window_len < throttle)
1360 if ((atomic_read(&ra->ra_cur_pages) + pages) > ra->ra_max_pages)
1363 if (ras->ras_async_last_readpage == start)
1366 /* ll_readahead_work_free() free it */
1369 lrw->lrw_file = get_file(file);
1370 lrw->lrw_start = start;
1372 spin_lock(&ras->ras_lock);
1373 ras->ras_next_readahead = end + 1;
1374 ras->ras_async_last_readpage = start;
1375 spin_unlock(&ras->ras_lock);
1376 ll_readahead_work_add(inode, lrw);
1384 int ll_readpage(struct file *file, struct page *vmpage)
1386 struct inode *inode = file_inode(file);
1387 struct cl_object *clob = ll_i2info(inode)->lli_clob;
1388 struct ll_cl_context *lcc;
1389 const struct lu_env *env = NULL;
1390 struct cl_io *io = NULL;
1391 struct cl_page *page;
1392 struct ll_sb_info *sbi = ll_i2sbi(inode);
1396 lcc = ll_cl_find(file);
1402 if (io == NULL) { /* fast read */
1403 struct inode *inode = file_inode(file);
1404 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1405 struct ll_readahead_state *ras = &fd->fd_ras;
1406 struct lu_env *local_env = NULL;
1407 unsigned long fast_read_pages =
1408 max(RA_REMAIN_WINDOW_MIN, ras->ras_rpc_size);
1409 struct vvp_page *vpg;
1413 /* TODO: need to verify the layout version to make sure
1414 * the page is not invalid due to layout change. */
1415 page = cl_vmpage_page(vmpage, clob);
1417 unlock_page(vmpage);
1418 ll_ra_stats_inc_sbi(sbi, RA_STAT_FAILED_FAST_READ);
1422 vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1423 if (vpg->vpg_defer_uptodate) {
1424 enum ras_update_flags flags = LL_RAS_HIT;
1426 if (lcc && lcc->lcc_type == LCC_MMAP)
1427 flags |= LL_RAS_MMAP;
1429 /* For fast read, it updates read ahead state only
1430 * if the page is hit in cache because non cache page
1431 * case will be handled by slow read later. */
1432 ras_update(sbi, inode, ras, vvp_index(vpg), flags);
1433 /* avoid duplicate ras_update() call */
1434 vpg->vpg_ra_updated = 1;
1436 /* Check if we can issue a readahead RPC, if that is
1437 * the case, we can't do fast IO because we will need
1438 * a cl_io to issue the RPC. */
1439 if (ras->ras_window_start + ras->ras_window_len <
1440 ras->ras_next_readahead + fast_read_pages ||
1441 kickoff_async_readahead(file, fast_read_pages) > 0)
1446 local_env = cl_env_percpu_get();
1450 /* export the page and skip io stack */
1452 vpg->vpg_ra_used = 1;
1453 cl_page_export(env, page, 1);
1455 ll_ra_stats_inc_sbi(sbi, RA_STAT_FAILED_FAST_READ);
1457 /* release page refcount before unlocking the page to ensure
1458 * the object won't be destroyed in the calling path of
1459 * cl_page_put(). Please see comment in ll_releasepage(). */
1460 cl_page_put(env, page);
1461 unlock_page(vmpage);
1463 cl_env_percpu_put(local_env);
1468 LASSERT(io->ci_state == CIS_IO_GOING);
1469 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
1470 if (!IS_ERR(page)) {
1471 LASSERT(page->cp_type == CPT_CACHEABLE);
1472 if (likely(!PageUptodate(vmpage))) {
1473 cl_page_assume(env, io, page);
1475 result = ll_io_read_page(env, io, page, file);
1477 /* Page from a non-object file. */
1478 unlock_page(vmpage);
1481 cl_page_put(env, page);
1483 unlock_page(vmpage);
1484 result = PTR_ERR(page);