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, 2013, 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 <linux/lustre_compat25.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), pages);
105 if (ret < 0 || ret < min_t(long, PTLRPC_MAX_BRW_PAGES, pages))
108 /* If the non-strided (ria_pages == 0) readahead window
109 * (ria_start + ret) has grown across an RPC boundary, then trim
110 * readahead size by the amount beyond the RPC so it ends on an
111 * RPC boundary. If the readahead window is already ending on
112 * an RPC boundary (beyond_rpc == 0), or smaller than a full
113 * RPC (beyond_rpc < ret) the readahead size is unchanged.
114 * The (beyond_rpc != 0) check is skipped since the conditional
115 * branch is more expensive than subtracting zero from the result.
117 * Strided read is left unaligned to avoid small fragments beyond
118 * the RPC boundary from needing an extra read RPC. */
119 if (ria->ria_pages == 0) {
120 long beyond_rpc = (ria->ria_start + ret) % PTLRPC_MAX_BRW_PAGES;
121 if (/* beyond_rpc != 0 && */ beyond_rpc < ret)
125 if (atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) {
126 atomic_sub(ret, &ra->ra_cur_pages);
132 /* override ra limit for maximum performance */
133 atomic_add(min - ret, &ra->ra_cur_pages);
139 void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
141 struct ll_ra_info *ra = &sbi->ll_ra_info;
142 atomic_sub(len, &ra->ra_cur_pages);
145 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which)
147 LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
148 lprocfs_counter_incr(sbi->ll_ra_stats, which);
151 void ll_ra_stats_inc(struct inode *inode, enum ra_stat which)
153 struct ll_sb_info *sbi = ll_i2sbi(inode);
154 ll_ra_stats_inc_sbi(sbi, which);
157 #define RAS_CDEBUG(ras) \
159 "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu r %lu ri %lu" \
160 "csr %lu sf %lu sp %lu sl %lu \n", \
161 ras->ras_last_readpage, ras->ras_consecutive_requests, \
162 ras->ras_consecutive_pages, ras->ras_window_start, \
163 ras->ras_window_len, ras->ras_next_readahead, \
164 ras->ras_requests, ras->ras_request_index, \
165 ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
166 ras->ras_stride_pages, ras->ras_stride_length)
168 static int index_in_window(unsigned long index, unsigned long point,
169 unsigned long before, unsigned long after)
171 unsigned long start = point - before, end = point + after;
178 return start <= index && index <= end;
181 static struct ll_readahead_state *ll_ras_get(struct file *f)
183 struct ll_file_data *fd;
185 fd = LUSTRE_FPRIVATE(f);
189 void ll_ra_read_in(struct file *f, struct ll_ra_read *rar)
191 struct ll_readahead_state *ras;
195 spin_lock(&ras->ras_lock);
197 ras->ras_request_index = 0;
198 ras->ras_consecutive_requests++;
199 rar->lrr_reader = current;
201 list_add(&rar->lrr_linkage, &ras->ras_read_beads);
202 spin_unlock(&ras->ras_lock);
205 void ll_ra_read_ex(struct file *f, struct ll_ra_read *rar)
207 struct ll_readahead_state *ras;
211 spin_lock(&ras->ras_lock);
212 list_del_init(&rar->lrr_linkage);
213 spin_unlock(&ras->ras_lock);
216 static struct ll_ra_read *ll_ra_read_get_locked(struct ll_readahead_state *ras)
218 struct ll_ra_read *scan;
220 list_for_each_entry(scan, &ras->ras_read_beads, lrr_linkage) {
221 if (scan->lrr_reader == current)
227 struct ll_ra_read *ll_ra_read_get(struct file *f)
229 struct ll_readahead_state *ras;
230 struct ll_ra_read *bead;
234 spin_lock(&ras->ras_lock);
235 bead = ll_ra_read_get_locked(ras);
236 spin_unlock(&ras->ras_lock);
240 static int cl_read_ahead_page(const struct lu_env *env, struct cl_io *io,
241 struct cl_page_list *queue, struct cl_page *page,
242 struct cl_object *clob, pgoff_t *max_index)
244 struct page *vmpage = page->cp_vmpage;
251 cl_page_assume(env, io, page);
252 lu_ref_add(&page->cp_reference, "ra", current);
253 cp = cl2ccc_page(cl_object_page_slice(clob, page));
254 if (!cp->cpg_defer_uptodate && !PageUptodate(vmpage)) {
255 CDEBUG(D_READA, "page index %lu, max_index: %lu\n",
256 ccc_index(cp), *max_index);
257 /* Disable the optimization on prefetching maximum readahead
258 * index because there is a race with lock cancellation. This
259 * optimization will be revived later.
260 * if (*max_index == 0 || ccc_index(cp) > *max_index) */
261 rc = cl_page_is_under_lock(env, io, page, max_index);
263 cp->cpg_defer_uptodate = 1;
265 cl_page_list_add(queue, page);
268 cl_page_discard(env, io, page);
272 /* skip completed pages */
273 cl_page_unassume(env, io, page);
275 lu_ref_del(&page->cp_reference, "ra", current);
276 cl_page_put(env, page);
281 * Initiates read-ahead of a page with given index.
283 * \retval +ve: page was added to \a queue.
285 * \retval -ENOLCK: there is no extent lock for this part of a file, stop
288 * \retval -ve, 0: page wasn't added to \a queue for other reason.
290 static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io,
291 struct cl_page_list *queue,
292 pgoff_t index, pgoff_t *max_index)
294 struct cl_object *clob = io->ci_obj;
295 struct inode *inode = ccc_object_inode(clob);
297 struct cl_page *page;
298 enum ra_stat which = _NR_RA_STAT; /* keep gcc happy */
301 const char *msg = NULL;
305 gfp_mask = GFP_HIGHUSER & ~__GFP_WAIT;
307 gfp_mask |= __GFP_NOWARN;
309 vmpage = grab_cache_page_nowait(inode->i_mapping, index);
310 if (vmpage != NULL) {
311 /* Check if vmpage was truncated or reclaimed */
312 if (vmpage->mapping == inode->i_mapping) {
313 page = cl_page_find(env, clob, vmpage->index,
314 vmpage, CPT_CACHEABLE);
316 rc = cl_read_ahead_page(env, io, queue,
317 page, clob, max_index);
319 which = RA_STAT_FAILED_MATCH;
320 msg = "lock match failed";
323 which = RA_STAT_FAILED_GRAB_PAGE;
324 msg = "cl_page_find failed";
327 which = RA_STAT_WRONG_GRAB_PAGE;
328 msg = "g_c_p_n returned invalid page";
332 page_cache_release(vmpage);
334 which = RA_STAT_FAILED_GRAB_PAGE;
335 msg = "g_c_p_n failed";
338 ll_ra_stats_inc(inode, which);
339 CDEBUG(D_READA, "%s\n", msg);
344 #define RIA_DEBUG(ria) \
345 CDEBUG(D_READA, "rs %lu re %lu ro %lu rl %lu rp %lu\n", \
346 ria->ria_start, ria->ria_end, ria->ria_stoff, ria->ria_length,\
349 /* Limit this to the blocksize instead of PTLRPC_BRW_MAX_SIZE, since we don't
350 * know what the actual RPC size is. If this needs to change, it makes more
351 * sense to tune the i_blkbits value for the file based on the OSTs it is
352 * striped over, rather than having a constant value for all files here. */
354 /* RAS_INCREASE_STEP should be (1UL << (inode->i_blkbits - PAGE_CACHE_SHIFT)).
355 * Temprarily set RAS_INCREASE_STEP to 1MB. After 4MB RPC is enabled
356 * by default, this should be adjusted corresponding with max_read_ahead_mb
357 * and max_read_ahead_per_file_mb otherwise the readahead budget can be used
358 * up quickly which will affect read performance siginificantly. See LU-2816 */
359 #define RAS_INCREASE_STEP(inode) (ONE_MB_BRW_SIZE >> PAGE_CACHE_SHIFT)
361 static inline int stride_io_mode(struct ll_readahead_state *ras)
363 return ras->ras_consecutive_stride_requests > 1;
365 /* The function calculates how much pages will be read in
366 * [off, off + length], in such stride IO area,
367 * stride_offset = st_off, stride_lengh = st_len,
368 * stride_pages = st_pgs
370 * |------------------|*****|------------------|*****|------------|*****|....
373 * |----- st_len -----|
375 * How many pages it should read in such pattern
376 * |-------------------------------------------------------------|
378 * |<------ length ------->|
380 * = |<----->| + |-------------------------------------| + |---|
381 * start_left st_pgs * i end_left
384 stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs,
385 unsigned long off, unsigned long length)
387 __u64 start = off > st_off ? off - st_off : 0;
388 __u64 end = off + length > st_off ? off + length - st_off : 0;
389 unsigned long start_left = 0;
390 unsigned long end_left = 0;
391 unsigned long pg_count;
393 if (st_len == 0 || length == 0 || end == 0)
396 start_left = do_div(start, st_len);
397 if (start_left < st_pgs)
398 start_left = st_pgs - start_left;
402 end_left = do_div(end, st_len);
403 if (end_left > st_pgs)
406 CDEBUG(D_READA, "start "LPU64", end "LPU64" start_left %lu end_left %lu \n",
407 start, end, start_left, end_left);
410 pg_count = end_left - (st_pgs - start_left);
412 pg_count = start_left + st_pgs * (end - start - 1) + end_left;
414 CDEBUG(D_READA, "st_off %lu, st_len %lu st_pgs %lu off %lu length %lu"
415 "pgcount %lu\n", st_off, st_len, st_pgs, off, length, pg_count);
420 static int ria_page_count(struct ra_io_arg *ria)
422 __u64 length = ria->ria_end >= ria->ria_start ?
423 ria->ria_end - ria->ria_start + 1 : 0;
425 return stride_pg_count(ria->ria_stoff, ria->ria_length,
426 ria->ria_pages, ria->ria_start,
430 /*Check whether the index is in the defined ra-window */
431 static int ras_inside_ra_window(unsigned long idx, struct ra_io_arg *ria)
433 /* If ria_length == ria_pages, it means non-stride I/O mode,
434 * idx should always inside read-ahead window in this case
435 * For stride I/O mode, just check whether the idx is inside
437 return ria->ria_length == 0 || ria->ria_length == ria->ria_pages ||
438 (idx >= ria->ria_stoff && (idx - ria->ria_stoff) %
439 ria->ria_length < ria->ria_pages);
442 static int ll_read_ahead_pages(const struct lu_env *env,
443 struct cl_io *io, struct cl_page_list *queue,
444 struct ra_io_arg *ria,
445 unsigned long *reserved_pages,
446 unsigned long *ra_end)
451 pgoff_t max_index = 0;
453 LASSERT(ria != NULL);
456 stride_ria = ria->ria_length > ria->ria_pages && ria->ria_pages > 0;
457 for (page_idx = ria->ria_start;
458 page_idx <= ria->ria_end && *reserved_pages > 0; page_idx++) {
459 if (ras_inside_ra_window(page_idx, ria)) {
460 /* If the page is inside the read-ahead window*/
461 rc = ll_read_ahead_page(env, io, queue,
462 page_idx, &max_index);
466 } else if (rc == -ENOLCK)
468 } else if (stride_ria) {
469 /* If it is not in the read-ahead window, and it is
470 * read-ahead mode, then check whether it should skip
473 /* FIXME: This assertion only is valid when it is for
474 * forward read-ahead, it will be fixed when backward
475 * read-ahead is implemented */
476 LASSERTF(page_idx >= ria->ria_stoff,
477 "Invalid page_idx %lu rs %lu re %lu ro %lu "
478 "rl %lu rp %lu\n", page_idx,
479 ria->ria_start, ria->ria_end, ria->ria_stoff,
480 ria->ria_length, ria->ria_pages);
481 offset = page_idx - ria->ria_stoff;
482 offset = offset % (ria->ria_length);
483 if (offset > ria->ria_pages) {
484 page_idx += ria->ria_length - offset;
485 CDEBUG(D_READA, "i %lu skip %lu \n", page_idx,
486 ria->ria_length - offset);
495 int ll_readahead(const struct lu_env *env, struct cl_io *io,
496 struct cl_page_list *queue, struct ll_readahead_state *ras,
499 struct vvp_io *vio = vvp_env_io(env);
500 struct vvp_thread_info *vti = vvp_env_info(env);
501 struct cl_attr *attr = ccc_env_thread_attr(env);
502 unsigned long start = 0, end = 0, reserved;
503 unsigned long ra_end, len, mlen = 0;
505 struct ll_ra_read *bead;
506 struct ra_io_arg *ria = &vti->vti_ria;
507 struct cl_object *clob;
513 inode = ccc_object_inode(clob);
515 memset(ria, 0, sizeof *ria);
517 cl_object_attr_lock(clob);
518 ret = cl_object_attr_get(env, clob, attr);
519 cl_object_attr_unlock(clob);
525 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
529 spin_lock(&ras->ras_lock);
530 if (vio->cui_ra_window_set)
531 bead = &vio->cui_bead;
535 /* Enlarge the RA window to encompass the full read */
536 if (bead != NULL && ras->ras_window_start + ras->ras_window_len <
537 bead->lrr_start + bead->lrr_count) {
538 ras->ras_window_len = bead->lrr_start + bead->lrr_count -
539 ras->ras_window_start;
541 /* Reserve a part of the read-ahead window that we'll be issuing */
542 if (ras->ras_window_len > 0) {
544 * Note: other thread might rollback the ras_next_readahead,
545 * if it can not get the full size of prepared pages, see the
546 * end of this function. For stride read ahead, it needs to
547 * make sure the offset is no less than ras_stride_offset,
548 * so that stride read ahead can work correctly.
550 if (stride_io_mode(ras))
551 start = max(ras->ras_next_readahead,
552 ras->ras_stride_offset);
554 start = ras->ras_next_readahead;
555 end = ras->ras_window_start + ras->ras_window_len - 1;
559 unsigned long rpc_boundary;
561 * Align RA window to an optimal boundary.
563 * XXX This would be better to align to cl_max_pages_per_rpc
564 * instead of PTLRPC_MAX_BRW_PAGES, because the RPC size may
565 * be aligned to the RAID stripe size in the future and that
566 * is more important than the RPC size.
568 /* Note: we only trim the RPC, instead of extending the RPC
569 * to the boundary, so to avoid reading too much pages during
571 rpc_boundary = ((end + 1) & (~(PTLRPC_MAX_BRW_PAGES - 1)));
572 if (rpc_boundary > 0)
575 if (rpc_boundary > start)
578 /* Truncate RA window to end of file */
579 end = min(end, (unsigned long)((kms - 1) >> PAGE_CACHE_SHIFT));
581 ras->ras_next_readahead = max(end, end + 1);
584 ria->ria_start = start;
586 /* If stride I/O mode is detected, get stride window*/
587 if (stride_io_mode(ras)) {
588 ria->ria_stoff = ras->ras_stride_offset;
589 ria->ria_length = ras->ras_stride_length;
590 ria->ria_pages = ras->ras_stride_pages;
592 spin_unlock(&ras->ras_lock);
595 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
598 len = ria_page_count(ria);
600 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
604 CDEBUG(D_READA, DFID": ria: %lu/%lu, bead: %lu/%lu, hit: %d\n",
605 PFID(lu_object_fid(&clob->co_lu)),
606 ria->ria_start, ria->ria_end,
607 bead == NULL ? 0 : bead->lrr_start,
608 bead == NULL ? 0 : bead->lrr_count,
611 /* at least to extend the readahead window to cover current read */
612 if (!hit && bead != NULL &&
613 bead->lrr_start + bead->lrr_count > ria->ria_start) {
614 /* to the end of current read window. */
615 mlen = bead->lrr_start + bead->lrr_count - ria->ria_start;
616 /* trim to RPC boundary */
617 start = ria->ria_start & (PTLRPC_MAX_BRW_PAGES - 1);
618 mlen = min(mlen, PTLRPC_MAX_BRW_PAGES - start);
621 reserved = ll_ra_count_get(ll_i2sbi(inode), ria, len, mlen);
623 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
625 CDEBUG(D_READA, "reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
627 atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
628 ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
630 ret = ll_read_ahead_pages(env, io, queue, ria, &reserved, &ra_end);
633 ll_ra_count_put(ll_i2sbi(inode), reserved);
635 if (ra_end == end + 1 && ra_end == (kms >> PAGE_CACHE_SHIFT))
636 ll_ra_stats_inc(inode, RA_STAT_EOF);
638 /* if we didn't get to the end of the region we reserved from
639 * the ras we need to go back and update the ras so that the
640 * next read-ahead tries from where we left off. we only do so
641 * if the region we failed to issue read-ahead on is still ahead
642 * of the app and behind the next index to start read-ahead from */
643 CDEBUG(D_READA, "ra_end %lu end %lu stride end %lu \n",
644 ra_end, end, ria->ria_end);
646 if (ra_end != end + 1) {
647 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
648 spin_lock(&ras->ras_lock);
649 if (ra_end < ras->ras_next_readahead &&
650 index_in_window(ra_end, ras->ras_window_start, 0,
651 ras->ras_window_len)) {
652 ras->ras_next_readahead = ra_end;
655 spin_unlock(&ras->ras_lock);
661 static void ras_set_start(struct inode *inode, struct ll_readahead_state *ras,
664 ras->ras_window_start = index & (~(RAS_INCREASE_STEP(inode) - 1));
667 /* called with the ras_lock held or from places where it doesn't matter */
668 static void ras_reset(struct inode *inode, struct ll_readahead_state *ras,
671 ras->ras_last_readpage = index;
672 ras->ras_consecutive_requests = 0;
673 ras->ras_consecutive_pages = 0;
674 ras->ras_window_len = 0;
675 ras_set_start(inode, ras, index);
676 ras->ras_next_readahead = max(ras->ras_window_start, index);
681 /* called with the ras_lock held or from places where it doesn't matter */
682 static void ras_stride_reset(struct ll_readahead_state *ras)
684 ras->ras_consecutive_stride_requests = 0;
685 ras->ras_stride_length = 0;
686 ras->ras_stride_pages = 0;
690 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
692 spin_lock_init(&ras->ras_lock);
693 ras_reset(inode, ras, 0);
694 ras->ras_requests = 0;
695 INIT_LIST_HEAD(&ras->ras_read_beads);
699 * Check whether the read request is in the stride window.
700 * If it is in the stride window, return 1, otherwise return 0.
702 static int index_in_stride_window(struct ll_readahead_state *ras,
705 unsigned long stride_gap;
707 if (ras->ras_stride_length == 0 || ras->ras_stride_pages == 0 ||
708 ras->ras_stride_pages == ras->ras_stride_length)
711 stride_gap = index - ras->ras_last_readpage - 1;
713 /* If it is contiguous read */
715 return ras->ras_consecutive_pages + 1 <= ras->ras_stride_pages;
717 /* Otherwise check the stride by itself */
718 return (ras->ras_stride_length - ras->ras_stride_pages) == stride_gap &&
719 ras->ras_consecutive_pages == ras->ras_stride_pages;
722 static void ras_update_stride_detector(struct ll_readahead_state *ras,
725 unsigned long stride_gap = index - ras->ras_last_readpage - 1;
727 if (!stride_io_mode(ras) && (stride_gap != 0 ||
728 ras->ras_consecutive_stride_requests == 0)) {
729 ras->ras_stride_pages = ras->ras_consecutive_pages;
730 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
732 LASSERT(ras->ras_request_index == 0);
733 LASSERT(ras->ras_consecutive_stride_requests == 0);
735 if (index <= ras->ras_last_readpage) {
736 /*Reset stride window for forward read*/
737 ras_stride_reset(ras);
741 ras->ras_stride_pages = ras->ras_consecutive_pages;
742 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
749 stride_page_count(struct ll_readahead_state *ras, unsigned long len)
751 return stride_pg_count(ras->ras_stride_offset, ras->ras_stride_length,
752 ras->ras_stride_pages, ras->ras_stride_offset,
756 /* Stride Read-ahead window will be increased inc_len according to
757 * stride I/O pattern */
758 static void ras_stride_increase_window(struct ll_readahead_state *ras,
759 struct ll_ra_info *ra,
760 unsigned long inc_len)
762 unsigned long left, step, window_len;
763 unsigned long stride_len;
765 LASSERT(ras->ras_stride_length > 0);
766 LASSERTF(ras->ras_window_start + ras->ras_window_len
767 >= ras->ras_stride_offset, "window_start %lu, window_len %lu"
768 " stride_offset %lu\n", ras->ras_window_start,
769 ras->ras_window_len, ras->ras_stride_offset);
771 stride_len = ras->ras_window_start + ras->ras_window_len -
772 ras->ras_stride_offset;
774 left = stride_len % ras->ras_stride_length;
775 window_len = ras->ras_window_len - left;
777 if (left < ras->ras_stride_pages)
780 left = ras->ras_stride_pages + inc_len;
782 LASSERT(ras->ras_stride_pages != 0);
784 step = left / ras->ras_stride_pages;
785 left %= ras->ras_stride_pages;
787 window_len += step * ras->ras_stride_length + left;
789 if (stride_page_count(ras, window_len) <= ra->ra_max_pages_per_file)
790 ras->ras_window_len = window_len;
795 static void ras_increase_window(struct inode *inode,
796 struct ll_readahead_state *ras,
797 struct ll_ra_info *ra)
799 /* The stretch of ra-window should be aligned with max rpc_size
800 * but current clio architecture does not support retrieve such
801 * information from lower layer. FIXME later
803 if (stride_io_mode(ras))
804 ras_stride_increase_window(ras, ra, RAS_INCREASE_STEP(inode));
806 ras->ras_window_len = min(ras->ras_window_len +
807 RAS_INCREASE_STEP(inode),
808 ra->ra_max_pages_per_file);
811 void ras_update(struct ll_sb_info *sbi, struct inode *inode,
812 struct ll_readahead_state *ras, unsigned long index,
815 struct ll_ra_info *ra = &sbi->ll_ra_info;
816 int zero = 0, stride_detect = 0, ra_miss = 0;
819 spin_lock(&ras->ras_lock);
821 ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
823 /* reset the read-ahead window in two cases. First when the app seeks
824 * or reads to some other part of the file. Secondly if we get a
825 * read-ahead miss that we think we've previously issued. This can
826 * be a symptom of there being so many read-ahead pages that the VM is
827 * reclaiming it before we get to it. */
828 if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
830 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
831 } else if (!hit && ras->ras_window_len &&
832 index < ras->ras_next_readahead &&
833 index_in_window(index, ras->ras_window_start, 0,
834 ras->ras_window_len)) {
836 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
839 /* On the second access to a file smaller than the tunable
840 * ra_max_read_ahead_whole_pages trigger RA on all pages in the
841 * file up to ra_max_pages_per_file. This is simply a best effort
842 * and only occurs once per open file. Normal RA behavior is reverted
843 * to for subsequent IO. The mmap case does not increment
844 * ras_requests and thus can never trigger this behavior. */
845 if (ras->ras_requests == 2 && !ras->ras_request_index) {
848 kms_pages = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >>
851 CDEBUG(D_READA, "kmsp "LPU64" mwp %lu mp %lu\n", kms_pages,
852 ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages_per_file);
855 kms_pages <= ra->ra_max_read_ahead_whole_pages) {
856 ras->ras_window_start = 0;
857 ras->ras_last_readpage = 0;
858 ras->ras_next_readahead = 0;
859 ras->ras_window_len = min(ra->ra_max_pages_per_file,
860 ra->ra_max_read_ahead_whole_pages);
865 /* check whether it is in stride I/O mode*/
866 if (!index_in_stride_window(ras, index)) {
867 if (ras->ras_consecutive_stride_requests == 0 &&
868 ras->ras_request_index == 0) {
869 ras_update_stride_detector(ras, index);
870 ras->ras_consecutive_stride_requests++;
872 ras_stride_reset(ras);
874 ras_reset(inode, ras, index);
875 ras->ras_consecutive_pages++;
878 ras->ras_consecutive_pages = 0;
879 ras->ras_consecutive_requests = 0;
880 if (++ras->ras_consecutive_stride_requests > 1)
886 if (index_in_stride_window(ras, index) &&
887 stride_io_mode(ras)) {
888 /*If stride-RA hit cache miss, the stride dector
889 *will not be reset to avoid the overhead of
890 *redetecting read-ahead mode */
891 if (index != ras->ras_last_readpage + 1)
892 ras->ras_consecutive_pages = 0;
893 ras_reset(inode, ras, index);
896 /* Reset both stride window and normal RA
898 ras_reset(inode, ras, index);
899 ras->ras_consecutive_pages++;
900 ras_stride_reset(ras);
903 } else if (stride_io_mode(ras)) {
904 /* If this is contiguous read but in stride I/O mode
905 * currently, check whether stride step still is valid,
906 * if invalid, it will reset the stride ra window*/
907 if (!index_in_stride_window(ras, index)) {
908 /* Shrink stride read-ahead window to be zero */
909 ras_stride_reset(ras);
910 ras->ras_window_len = 0;
911 ras->ras_next_readahead = index;
915 ras->ras_consecutive_pages++;
916 ras->ras_last_readpage = index;
917 ras_set_start(inode, ras, index);
919 if (stride_io_mode(ras)) {
920 /* Since stride readahead is sentivite to the offset
921 * of read-ahead, so we use original offset here,
922 * instead of ras_window_start, which is RPC aligned */
923 ras->ras_next_readahead = max(index, ras->ras_next_readahead);
925 if (ras->ras_next_readahead < ras->ras_window_start)
926 ras->ras_next_readahead = ras->ras_window_start;
928 ras->ras_next_readahead = index + 1;
932 /* Trigger RA in the mmap case where ras_consecutive_requests
933 * is not incremented and thus can't be used to trigger RA */
934 if (!ras->ras_window_len && ras->ras_consecutive_pages == 4) {
935 ras->ras_window_len = RAS_INCREASE_STEP(inode);
939 /* Initially reset the stride window offset to next_readahead*/
940 if (ras->ras_consecutive_stride_requests == 2 && stride_detect) {
942 * Once stride IO mode is detected, next_readahead should be
943 * reset to make sure next_readahead > stride offset
945 ras->ras_next_readahead = max(index, ras->ras_next_readahead);
946 ras->ras_stride_offset = index;
947 ras->ras_window_len = RAS_INCREASE_STEP(inode);
950 /* The initial ras_window_len is set to the request size. To avoid
951 * uselessly reading and discarding pages for random IO the window is
952 * only increased once per consecutive request received. */
953 if ((ras->ras_consecutive_requests > 1 || stride_detect) &&
954 !ras->ras_request_index)
955 ras_increase_window(inode, ras, ra);
959 ras->ras_request_index++;
960 spin_unlock(&ras->ras_lock);
964 int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
966 struct inode *inode = vmpage->mapping->host;
967 struct ll_inode_info *lli = ll_i2info(inode);
970 struct cl_page *page;
971 struct cl_object *clob;
972 struct cl_env_nest nest;
973 bool redirtied = false;
974 bool unlocked = false;
978 LASSERT(PageLocked(vmpage));
979 LASSERT(!PageWriteback(vmpage));
981 LASSERT(ll_i2dtexp(inode) != NULL);
983 env = cl_env_nested_get(&nest);
985 GOTO(out, result = PTR_ERR(env));
987 clob = ll_i2info(inode)->lli_clob;
988 LASSERT(clob != NULL);
990 io = ccc_env_thread_io(env);
992 io->ci_ignore_layout = 1;
993 result = cl_io_init(env, io, CIT_MISC, clob);
995 page = cl_page_find(env, clob, vmpage->index,
996 vmpage, CPT_CACHEABLE);
998 lu_ref_add(&page->cp_reference, "writepage",
1000 cl_page_assume(env, io, page);
1001 result = cl_page_flush(env, io, page);
1004 * Re-dirty page on error so it retries write,
1005 * but not in case when IO has actually
1006 * occurred and completed with an error.
1008 if (!PageError(vmpage)) {
1009 redirty_page_for_writepage(wbc, vmpage);
1014 cl_page_disown(env, io, page);
1016 lu_ref_del(&page->cp_reference,
1017 "writepage", current);
1018 cl_page_put(env, page);
1020 result = PTR_ERR(page);
1023 cl_io_fini(env, io);
1025 if (redirtied && wbc->sync_mode == WB_SYNC_ALL) {
1026 loff_t offset = cl_offset(clob, vmpage->index);
1028 /* Flush page failed because the extent is being written out.
1029 * Wait for the write of extent to be finished to avoid
1030 * breaking kernel which assumes ->writepage should mark
1031 * PageWriteback or clean the page. */
1032 result = cl_sync_file_range(inode, offset,
1033 offset + PAGE_CACHE_SIZE - 1,
1036 /* actually we may have written more than one page.
1037 * decreasing this page because the caller will count
1039 wbc->nr_to_write -= result - 1;
1044 cl_env_nested_put(&nest, env);
1049 if (!lli->lli_async_rc)
1050 lli->lli_async_rc = result;
1051 SetPageError(vmpage);
1053 unlock_page(vmpage);
1058 int ll_writepages(struct address_space *mapping, struct writeback_control *wbc)
1060 struct inode *inode = mapping->host;
1061 struct ll_sb_info *sbi = ll_i2sbi(inode);
1064 enum cl_fsync_mode mode;
1065 int range_whole = 0;
1067 int ignore_layout = 0;
1070 if (wbc->range_cyclic) {
1071 start = mapping->writeback_index << PAGE_CACHE_SHIFT;
1072 end = OBD_OBJECT_EOF;
1074 start = wbc->range_start;
1075 end = wbc->range_end;
1076 if (end == LLONG_MAX) {
1077 end = OBD_OBJECT_EOF;
1078 range_whole = start == 0;
1082 mode = CL_FSYNC_NONE;
1083 if (wbc->sync_mode == WB_SYNC_ALL)
1084 mode = CL_FSYNC_LOCAL;
1086 if (sbi->ll_umounting)
1087 /* if the mountpoint is being umounted, all pages have to be
1088 * evicted to avoid hitting LBUG when truncate_inode_pages()
1089 * is called later on. */
1092 if (cl_i2info(inode)->lli_clob == NULL)
1095 result = cl_sync_file_range(inode, start, end, mode, ignore_layout);
1097 wbc->nr_to_write -= result;
1101 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) {
1102 if (end == OBD_OBJECT_EOF)
1103 mapping->writeback_index = 0;
1105 mapping->writeback_index = (end >> PAGE_CACHE_SHIFT) +1;
1110 struct ll_cl_context *ll_cl_find(struct file *file)
1112 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1113 struct ll_cl_context *lcc;
1114 struct ll_cl_context *found = NULL;
1116 read_lock(&fd->fd_lock);
1117 list_for_each_entry(lcc, &fd->fd_lccs, lcc_list) {
1118 if (lcc->lcc_cookie == current) {
1123 read_unlock(&fd->fd_lock);
1128 void ll_cl_add(struct file *file, const struct lu_env *env, struct cl_io *io)
1130 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1131 struct ll_cl_context *lcc = &vvp_env_info(env)->vti_io_ctx;
1133 memset(lcc, 0, sizeof(*lcc));
1134 INIT_LIST_HEAD(&lcc->lcc_list);
1135 lcc->lcc_cookie = current;
1139 write_lock(&fd->fd_lock);
1140 list_add(&lcc->lcc_list, &fd->fd_lccs);
1141 write_unlock(&fd->fd_lock);
1144 void ll_cl_remove(struct file *file, const struct lu_env *env)
1146 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1147 struct ll_cl_context *lcc = &vvp_env_info(env)->vti_io_ctx;
1149 write_lock(&fd->fd_lock);
1150 list_del_init(&lcc->lcc_list);
1151 write_unlock(&fd->fd_lock);
1154 int ll_readpage(struct file *file, struct page *vmpage)
1156 struct cl_object *clob = ll_i2info(file->f_dentry->d_inode)->lli_clob;
1157 struct ll_cl_context *lcc;
1158 const struct lu_env *env;
1160 struct cl_page *page;
1164 lcc = ll_cl_find(file);
1166 unlock_page(vmpage);
1172 LASSERT(io != NULL);
1173 LASSERT(io->ci_state == CIS_IO_GOING);
1174 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
1175 if (!IS_ERR(page)) {
1176 LASSERT(page->cp_type == CPT_CACHEABLE);
1177 if (likely(!PageUptodate(vmpage))) {
1178 cl_page_assume(env, io, page);
1179 result = cl_io_read_page(env, io, page);
1181 /* Page from a non-object file. */
1182 unlock_page(vmpage);
1185 cl_page_put(env, page);
1187 unlock_page(vmpage);
1188 result = PTR_ERR(page);
1193 int ll_page_sync_io(const struct lu_env *env, struct cl_io *io,
1194 struct cl_page *page, enum cl_req_type crt)
1196 struct cl_2queue *queue;
1199 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
1201 queue = &io->ci_queue;
1202 cl_2queue_init_page(queue, page);
1204 result = cl_io_submit_sync(env, io, crt, queue, 0);
1205 LASSERT(cl_page_is_owned(page, io));
1207 if (crt == CRT_READ)
1209 * in CRT_WRITE case page is left locked even in case of
1212 cl_page_list_disown(env, io, &queue->c2_qin);
1213 cl_2queue_fini(env, queue);