4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2014, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 * Lustre Lite I/O page cache routines shared by different kernel revs
41 #include <linux/kernel.h>
43 #include <linux/string.h>
44 #include <linux/stat.h>
45 #include <linux/errno.h>
46 #include <linux/unistd.h>
47 #include <linux/writeback.h>
48 #include <asm/uaccess.h>
51 #include <linux/stat.h>
52 #include <asm/uaccess.h>
54 #include <linux/pagemap.h>
55 /* current_is_kswapd() */
56 #include <linux/swap.h>
58 #define DEBUG_SUBSYSTEM S_LLITE
60 #include <obd_cksum.h>
61 #include "llite_internal.h"
62 #include <lustre_compat.h>
64 struct obd_capa *cl_capa_lookup(struct inode *inode, enum cl_req_type crt)
68 opc = crt == CRT_WRITE ? CAPA_OPC_OSS_WRITE : CAPA_OPC_OSS_RW;
69 return ll_osscapa_get(inode, opc);
72 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which);
75 * Get readahead pages from the filesystem readahead pool of the client for a
78 * /param sbi superblock for filesystem readahead state ll_ra_info
79 * /param ria per-thread readahead state
80 * /param pages number of pages requested for readahead for the thread.
82 * WARNING: This algorithm is used to reduce contention on sbi->ll_lock.
83 * It should work well if the ra_max_pages is much greater than the single
84 * file's read-ahead window, and not too many threads contending for
85 * these readahead pages.
87 * TODO: There may be a 'global sync problem' if many threads are trying
88 * to get an ra budget that is larger than the remaining readahead pages
89 * and reach here at exactly the same time. They will compute /a ret to
90 * consume the remaining pages, but will fail at atomic_add_return() and
91 * get a zero ra window, although there is still ra space remaining. - Jay */
93 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi,
94 struct ra_io_arg *ria,
95 unsigned long pages, unsigned long min)
97 struct ll_ra_info *ra = &sbi->ll_ra_info;
101 /* If read-ahead pages left are less than 1M, do not do read-ahead,
102 * otherwise it will form small read RPC(< 1M), which hurt server
103 * performance a lot. */
104 ret = min(ra->ra_max_pages - atomic_read(&ra->ra_cur_pages),
106 if (ret < 0 || ret < min_t(long, PTLRPC_MAX_BRW_PAGES, pages))
109 /* If the non-strided (ria_pages == 0) readahead window
110 * (ria_start + ret) has grown across an RPC boundary, then trim
111 * readahead size by the amount beyond the RPC so it ends on an
112 * RPC boundary. If the readahead window is already ending on
113 * an RPC boundary (beyond_rpc == 0), or smaller than a full
114 * RPC (beyond_rpc < ret) the readahead size is unchanged.
115 * The (beyond_rpc != 0) check is skipped since the conditional
116 * branch is more expensive than subtracting zero from the result.
118 * Strided read is left unaligned to avoid small fragments beyond
119 * the RPC boundary from needing an extra read RPC. */
120 if (ria->ria_pages == 0) {
121 long beyond_rpc = (ria->ria_start + ret) % PTLRPC_MAX_BRW_PAGES;
122 if (/* beyond_rpc != 0 && */ beyond_rpc < ret)
126 if (atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) {
127 atomic_sub(ret, &ra->ra_cur_pages);
133 /* override ra limit for maximum performance */
134 atomic_add(min - ret, &ra->ra_cur_pages);
140 void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
142 struct ll_ra_info *ra = &sbi->ll_ra_info;
143 atomic_sub(len, &ra->ra_cur_pages);
146 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which)
148 LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
149 lprocfs_counter_incr(sbi->ll_ra_stats, which);
152 void ll_ra_stats_inc(struct inode *inode, enum ra_stat which)
154 struct ll_sb_info *sbi = ll_i2sbi(inode);
155 ll_ra_stats_inc_sbi(sbi, which);
158 #define RAS_CDEBUG(ras) \
160 "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu r %lu ri %lu" \
161 "csr %lu sf %lu sp %lu sl %lu \n", \
162 ras->ras_last_readpage, ras->ras_consecutive_requests, \
163 ras->ras_consecutive_pages, ras->ras_window_start, \
164 ras->ras_window_len, ras->ras_next_readahead, \
165 ras->ras_requests, ras->ras_request_index, \
166 ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
167 ras->ras_stride_pages, ras->ras_stride_length)
169 static int index_in_window(unsigned long index, unsigned long point,
170 unsigned long before, unsigned long after)
172 unsigned long start = point - before, end = point + after;
179 return start <= index && index <= end;
182 void ll_ras_enter(struct file *f)
184 struct ll_file_data *fd = LUSTRE_FPRIVATE(f);
185 struct ll_readahead_state *ras = &fd->fd_ras;
187 spin_lock(&ras->ras_lock);
189 ras->ras_request_index = 0;
190 ras->ras_consecutive_requests++;
191 spin_unlock(&ras->ras_lock);
194 static int cl_read_ahead_page(const struct lu_env *env, struct cl_io *io,
195 struct cl_page_list *queue, struct cl_page *page,
196 struct cl_object *clob, pgoff_t *max_index)
198 struct page *vmpage = page->cp_vmpage;
199 struct vvp_page *vpg;
205 cl_page_assume(env, io, page);
206 lu_ref_add(&page->cp_reference, "ra", current);
207 vpg = cl2vvp_page(cl_object_page_slice(clob, page));
208 if (!vpg->vpg_defer_uptodate && !PageUptodate(vmpage)) {
209 CDEBUG(D_READA, "page index %lu, max_index: %lu\n",
210 vvp_index(vpg), *max_index);
211 /* Disable the optimization on prefetching maximum readahead
212 * index because there is a race with lock cancellation. This
213 * optimization will be revived later.
214 * if (*max_index == 0 || vvp_index(vpg) > *max_index) */
215 rc = cl_page_is_under_lock(env, io, page, max_index);
217 vpg->vpg_defer_uptodate = 1;
218 vpg->vpg_ra_used = 0;
219 cl_page_list_add(queue, page);
222 cl_page_discard(env, io, page);
226 /* skip completed pages */
227 cl_page_unassume(env, io, page);
229 lu_ref_del(&page->cp_reference, "ra", current);
230 cl_page_put(env, page);
235 * Initiates read-ahead of a page with given index.
237 * \retval +ve: page was added to \a queue.
239 * \retval -ENOLCK: there is no extent lock for this part of a file, stop
242 * \retval -ve, 0: page wasn't added to \a queue for other reason.
244 static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io,
245 struct cl_page_list *queue,
246 pgoff_t index, pgoff_t *max_index)
248 struct cl_object *clob = io->ci_obj;
249 struct inode *inode = vvp_object_inode(clob);
251 struct cl_page *page;
252 enum ra_stat which = _NR_RA_STAT; /* keep gcc happy */
255 const char *msg = NULL;
259 gfp_mask = GFP_HIGHUSER & ~__GFP_WAIT;
261 gfp_mask |= __GFP_NOWARN;
263 vmpage = grab_cache_page_nowait(inode->i_mapping, index);
264 if (vmpage != NULL) {
265 /* Check if vmpage was truncated or reclaimed */
266 if (vmpage->mapping == inode->i_mapping) {
267 page = cl_page_find(env, clob, vmpage->index,
268 vmpage, CPT_CACHEABLE);
270 rc = cl_read_ahead_page(env, io, queue,
271 page, clob, max_index);
273 which = RA_STAT_FAILED_MATCH;
274 msg = "lock match failed";
277 which = RA_STAT_FAILED_GRAB_PAGE;
278 msg = "cl_page_find failed";
281 which = RA_STAT_WRONG_GRAB_PAGE;
282 msg = "g_c_p_n returned invalid page";
286 page_cache_release(vmpage);
288 which = RA_STAT_FAILED_GRAB_PAGE;
289 msg = "g_c_p_n failed";
292 ll_ra_stats_inc(inode, which);
293 CDEBUG(D_READA, "%s\n", msg);
298 #define RIA_DEBUG(ria) \
299 CDEBUG(D_READA, "rs %lu re %lu ro %lu rl %lu rp %lu\n", \
300 ria->ria_start, ria->ria_end, ria->ria_stoff, ria->ria_length,\
303 /* Limit this to the blocksize instead of PTLRPC_BRW_MAX_SIZE, since we don't
304 * know what the actual RPC size is. If this needs to change, it makes more
305 * sense to tune the i_blkbits value for the file based on the OSTs it is
306 * striped over, rather than having a constant value for all files here. */
308 /* RAS_INCREASE_STEP should be (1UL << (inode->i_blkbits - PAGE_CACHE_SHIFT)).
309 * Temprarily set RAS_INCREASE_STEP to 1MB. After 4MB RPC is enabled
310 * by default, this should be adjusted corresponding with max_read_ahead_mb
311 * and max_read_ahead_per_file_mb otherwise the readahead budget can be used
312 * up quickly which will affect read performance siginificantly. See LU-2816 */
313 #define RAS_INCREASE_STEP(inode) (ONE_MB_BRW_SIZE >> PAGE_CACHE_SHIFT)
315 static inline int stride_io_mode(struct ll_readahead_state *ras)
317 return ras->ras_consecutive_stride_requests > 1;
319 /* The function calculates how much pages will be read in
320 * [off, off + length], in such stride IO area,
321 * stride_offset = st_off, stride_lengh = st_len,
322 * stride_pages = st_pgs
324 * |------------------|*****|------------------|*****|------------|*****|....
327 * |----- st_len -----|
329 * How many pages it should read in such pattern
330 * |-------------------------------------------------------------|
332 * |<------ length ------->|
334 * = |<----->| + |-------------------------------------| + |---|
335 * start_left st_pgs * i end_left
338 stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs,
339 unsigned long off, unsigned long length)
341 __u64 start = off > st_off ? off - st_off : 0;
342 __u64 end = off + length > st_off ? off + length - st_off : 0;
343 unsigned long start_left = 0;
344 unsigned long end_left = 0;
345 unsigned long pg_count;
347 if (st_len == 0 || length == 0 || end == 0)
350 start_left = do_div(start, st_len);
351 if (start_left < st_pgs)
352 start_left = st_pgs - start_left;
356 end_left = do_div(end, st_len);
357 if (end_left > st_pgs)
360 CDEBUG(D_READA, "start "LPU64", end "LPU64" start_left %lu end_left %lu \n",
361 start, end, start_left, end_left);
364 pg_count = end_left - (st_pgs - start_left);
366 pg_count = start_left + st_pgs * (end - start - 1) + end_left;
368 CDEBUG(D_READA, "st_off %lu, st_len %lu st_pgs %lu off %lu length %lu"
369 "pgcount %lu\n", st_off, st_len, st_pgs, off, length, pg_count);
374 static int ria_page_count(struct ra_io_arg *ria)
376 __u64 length = ria->ria_end >= ria->ria_start ?
377 ria->ria_end - ria->ria_start + 1 : 0;
379 return stride_pg_count(ria->ria_stoff, ria->ria_length,
380 ria->ria_pages, ria->ria_start,
384 /*Check whether the index is in the defined ra-window */
385 static int ras_inside_ra_window(unsigned long idx, struct ra_io_arg *ria)
387 /* If ria_length == ria_pages, it means non-stride I/O mode,
388 * idx should always inside read-ahead window in this case
389 * For stride I/O mode, just check whether the idx is inside
391 return ria->ria_length == 0 || ria->ria_length == ria->ria_pages ||
392 (idx >= ria->ria_stoff && (idx - ria->ria_stoff) %
393 ria->ria_length < ria->ria_pages);
396 static int ll_read_ahead_pages(const struct lu_env *env,
397 struct cl_io *io, struct cl_page_list *queue,
398 struct ra_io_arg *ria,
399 unsigned long *reserved_pages,
400 unsigned long *ra_end)
405 pgoff_t max_index = 0;
407 LASSERT(ria != NULL);
410 stride_ria = ria->ria_length > ria->ria_pages && ria->ria_pages > 0;
411 for (page_idx = ria->ria_start;
412 page_idx <= ria->ria_end && *reserved_pages > 0; page_idx++) {
413 if (ras_inside_ra_window(page_idx, ria)) {
414 /* If the page is inside the read-ahead window*/
415 rc = ll_read_ahead_page(env, io, queue,
416 page_idx, &max_index);
420 } else if (rc == -ENOLCK)
422 } else if (stride_ria) {
423 /* If it is not in the read-ahead window, and it is
424 * read-ahead mode, then check whether it should skip
427 /* FIXME: This assertion only is valid when it is for
428 * forward read-ahead, it will be fixed when backward
429 * read-ahead is implemented */
430 LASSERTF(page_idx >= ria->ria_stoff,
431 "Invalid page_idx %lu rs %lu re %lu ro %lu "
432 "rl %lu rp %lu\n", page_idx,
433 ria->ria_start, ria->ria_end, ria->ria_stoff,
434 ria->ria_length, ria->ria_pages);
435 offset = page_idx - ria->ria_stoff;
436 offset = offset % (ria->ria_length);
437 if (offset > ria->ria_pages) {
438 page_idx += ria->ria_length - offset;
439 CDEBUG(D_READA, "i %lu skip %lu \n", page_idx,
440 ria->ria_length - offset);
449 int ll_readahead(const struct lu_env *env, struct cl_io *io,
450 struct cl_page_list *queue, struct ll_readahead_state *ras,
453 struct vvp_io *vio = vvp_env_io(env);
454 struct vvp_thread_info *vti = vvp_env_info(env);
455 struct cl_attr *attr = ccc_env_thread_attr(env);
456 unsigned long start = 0, end = 0, reserved;
457 unsigned long ra_end, len, mlen = 0;
459 struct ra_io_arg *ria = &vti->vti_ria;
460 struct cl_object *clob;
466 inode = vvp_object_inode(clob);
468 memset(ria, 0, sizeof *ria);
470 cl_object_attr_lock(clob);
471 ret = cl_object_attr_get(env, clob, attr);
472 cl_object_attr_unlock(clob);
478 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
482 spin_lock(&ras->ras_lock);
484 /* Enlarge the RA window to encompass the full read */
485 if (vio->vui_ra_valid &&
486 ras->ras_window_start + ras->ras_window_len <
487 vio->vui_ra_start + vio->vui_ra_count) {
488 ras->ras_window_len = vio->vui_ra_start + vio->vui_ra_count -
489 ras->ras_window_start;
492 /* Reserve a part of the read-ahead window that we'll be issuing */
493 if (ras->ras_window_len > 0) {
495 * Note: other thread might rollback the ras_next_readahead,
496 * if it can not get the full size of prepared pages, see the
497 * end of this function. For stride read ahead, it needs to
498 * make sure the offset is no less than ras_stride_offset,
499 * so that stride read ahead can work correctly.
501 if (stride_io_mode(ras))
502 start = max(ras->ras_next_readahead,
503 ras->ras_stride_offset);
505 start = ras->ras_next_readahead;
506 end = ras->ras_window_start + ras->ras_window_len - 1;
510 unsigned long rpc_boundary;
512 * Align RA window to an optimal boundary.
514 * XXX This would be better to align to cl_max_pages_per_rpc
515 * instead of PTLRPC_MAX_BRW_PAGES, because the RPC size may
516 * be aligned to the RAID stripe size in the future and that
517 * is more important than the RPC size.
519 /* Note: we only trim the RPC, instead of extending the RPC
520 * to the boundary, so to avoid reading too much pages during
522 rpc_boundary = ((end + 1) & (~(PTLRPC_MAX_BRW_PAGES - 1)));
523 if (rpc_boundary > 0)
526 if (rpc_boundary > start)
529 /* Truncate RA window to end of file */
530 end = min(end, (unsigned long)((kms - 1) >> PAGE_CACHE_SHIFT));
532 ras->ras_next_readahead = max(end, end + 1);
535 ria->ria_start = start;
537 /* If stride I/O mode is detected, get stride window*/
538 if (stride_io_mode(ras)) {
539 ria->ria_stoff = ras->ras_stride_offset;
540 ria->ria_length = ras->ras_stride_length;
541 ria->ria_pages = ras->ras_stride_pages;
543 spin_unlock(&ras->ras_lock);
546 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
549 len = ria_page_count(ria);
551 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
555 CDEBUG(D_READA, DFID": ria: %lu/%lu, bead: %lu/%lu, hit: %d\n",
556 PFID(lu_object_fid(&clob->co_lu)),
557 ria->ria_start, ria->ria_end,
558 vio->vui_ra_valid ? vio->vui_ra_start : 0,
559 vio->vui_ra_valid ? vio->vui_ra_count : 0,
562 /* at least to extend the readahead window to cover current read */
563 if (!hit && vio->vui_ra_valid &&
564 vio->vui_ra_start + vio->vui_ra_count > ria->ria_start) {
565 /* to the end of current read window. */
566 mlen = vio->vui_ra_start + vio->vui_ra_count - ria->ria_start;
567 /* trim to RPC boundary */
568 start = ria->ria_start & (PTLRPC_MAX_BRW_PAGES - 1);
569 mlen = min(mlen, PTLRPC_MAX_BRW_PAGES - start);
572 reserved = ll_ra_count_get(ll_i2sbi(inode), ria, len, mlen);
574 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
576 CDEBUG(D_READA, "reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
578 atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
579 ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
581 ret = ll_read_ahead_pages(env, io, queue, ria, &reserved, &ra_end);
584 ll_ra_count_put(ll_i2sbi(inode), reserved);
586 if (ra_end == end + 1 && ra_end == (kms >> PAGE_CACHE_SHIFT))
587 ll_ra_stats_inc(inode, RA_STAT_EOF);
589 /* if we didn't get to the end of the region we reserved from
590 * the ras we need to go back and update the ras so that the
591 * next read-ahead tries from where we left off. we only do so
592 * if the region we failed to issue read-ahead on is still ahead
593 * of the app and behind the next index to start read-ahead from */
594 CDEBUG(D_READA, "ra_end %lu end %lu stride end %lu \n",
595 ra_end, end, ria->ria_end);
597 if (ra_end != end + 1) {
598 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
599 spin_lock(&ras->ras_lock);
600 if (ra_end < ras->ras_next_readahead &&
601 index_in_window(ra_end, ras->ras_window_start, 0,
602 ras->ras_window_len)) {
603 ras->ras_next_readahead = ra_end;
606 spin_unlock(&ras->ras_lock);
612 static void ras_set_start(struct inode *inode, struct ll_readahead_state *ras,
615 ras->ras_window_start = index & (~(RAS_INCREASE_STEP(inode) - 1));
618 /* called with the ras_lock held or from places where it doesn't matter */
619 static void ras_reset(struct inode *inode, struct ll_readahead_state *ras,
622 ras->ras_last_readpage = index;
623 ras->ras_consecutive_requests = 0;
624 ras->ras_consecutive_pages = 0;
625 ras->ras_window_len = 0;
626 ras_set_start(inode, ras, index);
627 ras->ras_next_readahead = max(ras->ras_window_start, index);
632 /* called with the ras_lock held or from places where it doesn't matter */
633 static void ras_stride_reset(struct ll_readahead_state *ras)
635 ras->ras_consecutive_stride_requests = 0;
636 ras->ras_stride_length = 0;
637 ras->ras_stride_pages = 0;
641 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
643 spin_lock_init(&ras->ras_lock);
644 ras_reset(inode, ras, 0);
645 ras->ras_requests = 0;
649 * Check whether the read request is in the stride window.
650 * If it is in the stride window, return 1, otherwise return 0.
652 static int index_in_stride_window(struct ll_readahead_state *ras,
655 unsigned long stride_gap;
657 if (ras->ras_stride_length == 0 || ras->ras_stride_pages == 0 ||
658 ras->ras_stride_pages == ras->ras_stride_length)
661 stride_gap = index - ras->ras_last_readpage - 1;
663 /* If it is contiguous read */
665 return ras->ras_consecutive_pages + 1 <= ras->ras_stride_pages;
667 /* Otherwise check the stride by itself */
668 return (ras->ras_stride_length - ras->ras_stride_pages) == stride_gap &&
669 ras->ras_consecutive_pages == ras->ras_stride_pages;
672 static void ras_update_stride_detector(struct ll_readahead_state *ras,
675 unsigned long stride_gap = index - ras->ras_last_readpage - 1;
677 if (!stride_io_mode(ras) && (stride_gap != 0 ||
678 ras->ras_consecutive_stride_requests == 0)) {
679 ras->ras_stride_pages = ras->ras_consecutive_pages;
680 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
682 LASSERT(ras->ras_request_index == 0);
683 LASSERT(ras->ras_consecutive_stride_requests == 0);
685 if (index <= ras->ras_last_readpage) {
686 /*Reset stride window for forward read*/
687 ras_stride_reset(ras);
691 ras->ras_stride_pages = ras->ras_consecutive_pages;
692 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
699 stride_page_count(struct ll_readahead_state *ras, unsigned long len)
701 return stride_pg_count(ras->ras_stride_offset, ras->ras_stride_length,
702 ras->ras_stride_pages, ras->ras_stride_offset,
706 /* Stride Read-ahead window will be increased inc_len according to
707 * stride I/O pattern */
708 static void ras_stride_increase_window(struct ll_readahead_state *ras,
709 struct ll_ra_info *ra,
710 unsigned long inc_len)
712 unsigned long left, step, window_len;
713 unsigned long stride_len;
715 LASSERT(ras->ras_stride_length > 0);
716 LASSERTF(ras->ras_window_start + ras->ras_window_len
717 >= ras->ras_stride_offset, "window_start %lu, window_len %lu"
718 " stride_offset %lu\n", ras->ras_window_start,
719 ras->ras_window_len, ras->ras_stride_offset);
721 stride_len = ras->ras_window_start + ras->ras_window_len -
722 ras->ras_stride_offset;
724 left = stride_len % ras->ras_stride_length;
725 window_len = ras->ras_window_len - left;
727 if (left < ras->ras_stride_pages)
730 left = ras->ras_stride_pages + inc_len;
732 LASSERT(ras->ras_stride_pages != 0);
734 step = left / ras->ras_stride_pages;
735 left %= ras->ras_stride_pages;
737 window_len += step * ras->ras_stride_length + left;
739 if (stride_page_count(ras, window_len) <= ra->ra_max_pages_per_file)
740 ras->ras_window_len = window_len;
745 static void ras_increase_window(struct inode *inode,
746 struct ll_readahead_state *ras,
747 struct ll_ra_info *ra)
749 /* The stretch of ra-window should be aligned with max rpc_size
750 * but current clio architecture does not support retrieve such
751 * information from lower layer. FIXME later
753 if (stride_io_mode(ras))
754 ras_stride_increase_window(ras, ra, RAS_INCREASE_STEP(inode));
756 ras->ras_window_len = min(ras->ras_window_len +
757 RAS_INCREASE_STEP(inode),
758 ra->ra_max_pages_per_file);
761 void ras_update(struct ll_sb_info *sbi, struct inode *inode,
762 struct ll_readahead_state *ras, unsigned long index,
765 struct ll_ra_info *ra = &sbi->ll_ra_info;
766 int zero = 0, stride_detect = 0, ra_miss = 0;
769 spin_lock(&ras->ras_lock);
771 ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
773 /* reset the read-ahead window in two cases. First when the app seeks
774 * or reads to some other part of the file. Secondly if we get a
775 * read-ahead miss that we think we've previously issued. This can
776 * be a symptom of there being so many read-ahead pages that the VM is
777 * reclaiming it before we get to it. */
778 if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
780 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
781 } else if (!hit && ras->ras_window_len &&
782 index < ras->ras_next_readahead &&
783 index_in_window(index, ras->ras_window_start, 0,
784 ras->ras_window_len)) {
786 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
789 /* On the second access to a file smaller than the tunable
790 * ra_max_read_ahead_whole_pages trigger RA on all pages in the
791 * file up to ra_max_pages_per_file. This is simply a best effort
792 * and only occurs once per open file. Normal RA behavior is reverted
793 * to for subsequent IO. The mmap case does not increment
794 * ras_requests and thus can never trigger this behavior. */
795 if (ras->ras_requests == 2 && !ras->ras_request_index) {
798 kms_pages = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >>
801 CDEBUG(D_READA, "kmsp "LPU64" mwp %lu mp %lu\n", kms_pages,
802 ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages_per_file);
805 kms_pages <= ra->ra_max_read_ahead_whole_pages) {
806 ras->ras_window_start = 0;
807 ras->ras_last_readpage = 0;
808 ras->ras_next_readahead = 0;
809 ras->ras_window_len = min(ra->ra_max_pages_per_file,
810 ra->ra_max_read_ahead_whole_pages);
815 /* check whether it is in stride I/O mode*/
816 if (!index_in_stride_window(ras, index)) {
817 if (ras->ras_consecutive_stride_requests == 0 &&
818 ras->ras_request_index == 0) {
819 ras_update_stride_detector(ras, index);
820 ras->ras_consecutive_stride_requests++;
822 ras_stride_reset(ras);
824 ras_reset(inode, ras, index);
825 ras->ras_consecutive_pages++;
828 ras->ras_consecutive_pages = 0;
829 ras->ras_consecutive_requests = 0;
830 if (++ras->ras_consecutive_stride_requests > 1)
836 if (index_in_stride_window(ras, index) &&
837 stride_io_mode(ras)) {
838 /*If stride-RA hit cache miss, the stride dector
839 *will not be reset to avoid the overhead of
840 *redetecting read-ahead mode */
841 if (index != ras->ras_last_readpage + 1)
842 ras->ras_consecutive_pages = 0;
843 ras_reset(inode, ras, index);
846 /* Reset both stride window and normal RA
848 ras_reset(inode, ras, index);
849 ras->ras_consecutive_pages++;
850 ras_stride_reset(ras);
853 } else if (stride_io_mode(ras)) {
854 /* If this is contiguous read but in stride I/O mode
855 * currently, check whether stride step still is valid,
856 * if invalid, it will reset the stride ra window*/
857 if (!index_in_stride_window(ras, index)) {
858 /* Shrink stride read-ahead window to be zero */
859 ras_stride_reset(ras);
860 ras->ras_window_len = 0;
861 ras->ras_next_readahead = index;
865 ras->ras_consecutive_pages++;
866 ras->ras_last_readpage = index;
867 ras_set_start(inode, ras, index);
869 if (stride_io_mode(ras)) {
870 /* Since stride readahead is sentivite to the offset
871 * of read-ahead, so we use original offset here,
872 * instead of ras_window_start, which is RPC aligned */
873 ras->ras_next_readahead = max(index, ras->ras_next_readahead);
875 if (ras->ras_next_readahead < ras->ras_window_start)
876 ras->ras_next_readahead = ras->ras_window_start;
878 ras->ras_next_readahead = index + 1;
882 /* Trigger RA in the mmap case where ras_consecutive_requests
883 * is not incremented and thus can't be used to trigger RA */
884 if (!ras->ras_window_len && ras->ras_consecutive_pages == 4) {
885 ras->ras_window_len = RAS_INCREASE_STEP(inode);
889 /* Initially reset the stride window offset to next_readahead*/
890 if (ras->ras_consecutive_stride_requests == 2 && stride_detect) {
892 * Once stride IO mode is detected, next_readahead should be
893 * reset to make sure next_readahead > stride offset
895 ras->ras_next_readahead = max(index, ras->ras_next_readahead);
896 ras->ras_stride_offset = index;
897 ras->ras_window_len = RAS_INCREASE_STEP(inode);
900 /* The initial ras_window_len is set to the request size. To avoid
901 * uselessly reading and discarding pages for random IO the window is
902 * only increased once per consecutive request received. */
903 if ((ras->ras_consecutive_requests > 1 || stride_detect) &&
904 !ras->ras_request_index)
905 ras_increase_window(inode, ras, ra);
909 ras->ras_request_index++;
910 spin_unlock(&ras->ras_lock);
914 int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
916 struct inode *inode = vmpage->mapping->host;
917 struct ll_inode_info *lli = ll_i2info(inode);
920 struct cl_page *page;
921 struct cl_object *clob;
922 struct cl_env_nest nest;
923 bool redirtied = false;
924 bool unlocked = false;
928 LASSERT(PageLocked(vmpage));
929 LASSERT(!PageWriteback(vmpage));
931 LASSERT(ll_i2dtexp(inode) != NULL);
933 env = cl_env_nested_get(&nest);
935 GOTO(out, result = PTR_ERR(env));
937 clob = ll_i2info(inode)->lli_clob;
938 LASSERT(clob != NULL);
940 io = ccc_env_thread_io(env);
942 io->ci_ignore_layout = 1;
943 result = cl_io_init(env, io, CIT_MISC, clob);
945 page = cl_page_find(env, clob, vmpage->index,
946 vmpage, CPT_CACHEABLE);
948 lu_ref_add(&page->cp_reference, "writepage",
950 cl_page_assume(env, io, page);
951 result = cl_page_flush(env, io, page);
954 * Re-dirty page on error so it retries write,
955 * but not in case when IO has actually
956 * occurred and completed with an error.
958 if (!PageError(vmpage)) {
959 redirty_page_for_writepage(wbc, vmpage);
964 cl_page_disown(env, io, page);
966 lu_ref_del(&page->cp_reference,
967 "writepage", current);
968 cl_page_put(env, page);
970 result = PTR_ERR(page);
975 if (redirtied && wbc->sync_mode == WB_SYNC_ALL) {
976 loff_t offset = cl_offset(clob, vmpage->index);
978 /* Flush page failed because the extent is being written out.
979 * Wait for the write of extent to be finished to avoid
980 * breaking kernel which assumes ->writepage should mark
981 * PageWriteback or clean the page. */
982 result = cl_sync_file_range(inode, offset,
983 offset + PAGE_CACHE_SIZE - 1,
986 /* actually we may have written more than one page.
987 * decreasing this page because the caller will count
989 wbc->nr_to_write -= result - 1;
994 cl_env_nested_put(&nest, env);
999 if (!lli->lli_async_rc)
1000 lli->lli_async_rc = result;
1001 SetPageError(vmpage);
1003 unlock_page(vmpage);
1008 int ll_writepages(struct address_space *mapping, struct writeback_control *wbc)
1010 struct inode *inode = mapping->host;
1011 struct ll_sb_info *sbi = ll_i2sbi(inode);
1014 enum cl_fsync_mode mode;
1015 int range_whole = 0;
1017 int ignore_layout = 0;
1020 if (wbc->range_cyclic) {
1021 start = mapping->writeback_index << PAGE_CACHE_SHIFT;
1022 end = OBD_OBJECT_EOF;
1024 start = wbc->range_start;
1025 end = wbc->range_end;
1026 if (end == LLONG_MAX) {
1027 end = OBD_OBJECT_EOF;
1028 range_whole = start == 0;
1032 mode = CL_FSYNC_NONE;
1033 if (wbc->sync_mode == WB_SYNC_ALL)
1034 mode = CL_FSYNC_LOCAL;
1036 if (sbi->ll_umounting)
1037 /* if the mountpoint is being umounted, all pages have to be
1038 * evicted to avoid hitting LBUG when truncate_inode_pages()
1039 * is called later on. */
1042 if (ll_i2info(inode)->lli_clob == NULL)
1045 result = cl_sync_file_range(inode, start, end, mode, ignore_layout);
1047 wbc->nr_to_write -= result;
1051 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) {
1052 if (end == OBD_OBJECT_EOF)
1053 mapping->writeback_index = 0;
1055 mapping->writeback_index = (end >> PAGE_CACHE_SHIFT) +1;
1060 struct ll_cl_context *ll_cl_find(struct file *file)
1062 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1063 struct ll_cl_context *lcc;
1064 struct ll_cl_context *found = NULL;
1066 read_lock(&fd->fd_lock);
1067 list_for_each_entry(lcc, &fd->fd_lccs, lcc_list) {
1068 if (lcc->lcc_cookie == current) {
1073 read_unlock(&fd->fd_lock);
1078 void ll_cl_add(struct file *file, const struct lu_env *env, struct cl_io *io)
1080 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1081 struct ll_cl_context *lcc = &vvp_env_info(env)->vti_io_ctx;
1083 memset(lcc, 0, sizeof(*lcc));
1084 INIT_LIST_HEAD(&lcc->lcc_list);
1085 lcc->lcc_cookie = current;
1089 write_lock(&fd->fd_lock);
1090 list_add(&lcc->lcc_list, &fd->fd_lccs);
1091 write_unlock(&fd->fd_lock);
1094 void ll_cl_remove(struct file *file, const struct lu_env *env)
1096 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1097 struct ll_cl_context *lcc = &vvp_env_info(env)->vti_io_ctx;
1099 write_lock(&fd->fd_lock);
1100 list_del_init(&lcc->lcc_list);
1101 write_unlock(&fd->fd_lock);
1104 int ll_readpage(struct file *file, struct page *vmpage)
1106 struct cl_object *clob = ll_i2info(file->f_dentry->d_inode)->lli_clob;
1107 struct ll_cl_context *lcc;
1108 const struct lu_env *env;
1110 struct cl_page *page;
1114 lcc = ll_cl_find(file);
1116 unlock_page(vmpage);
1122 LASSERT(io != NULL);
1123 LASSERT(io->ci_state == CIS_IO_GOING);
1124 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
1125 if (!IS_ERR(page)) {
1126 LASSERT(page->cp_type == CPT_CACHEABLE);
1127 if (likely(!PageUptodate(vmpage))) {
1128 cl_page_assume(env, io, page);
1129 result = cl_io_read_page(env, io, page);
1131 /* Page from a non-object file. */
1132 unlock_page(vmpage);
1135 cl_page_put(env, page);
1137 unlock_page(vmpage);
1138 result = PTR_ERR(page);
1143 int ll_page_sync_io(const struct lu_env *env, struct cl_io *io,
1144 struct cl_page *page, enum cl_req_type crt)
1146 struct cl_2queue *queue;
1149 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
1151 queue = &io->ci_queue;
1152 cl_2queue_init_page(queue, page);
1154 result = cl_io_submit_sync(env, io, crt, queue, 0);
1155 LASSERT(cl_page_is_owned(page, io));
1157 if (crt == CRT_READ)
1159 * in CRT_WRITE case page is left locked even in case of
1162 cl_page_list_disown(env, io, &queue->c2_qin);
1163 cl_2queue_fini(env, queue);