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);
195 * Initiates read-ahead of a page with given index.
197 * \retval +ve: page was already uptodate so it will be skipped
199 * \retval -ve: page wasn't added to \a queue for error;
200 * \retval 0: page was added into \a queue for read ahead.
202 static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io,
203 struct cl_page_list *queue, pgoff_t index)
205 struct cl_object *clob = io->ci_obj;
206 struct inode *inode = vvp_object_inode(clob);
208 struct cl_page *page;
209 struct vvp_page *vpg;
210 enum ra_stat which = _NR_RA_STAT; /* keep gcc happy */
212 const char *msg = NULL;
215 vmpage = grab_cache_page_nowait(inode->i_mapping, index);
216 if (vmpage == NULL) {
217 which = RA_STAT_FAILED_GRAB_PAGE;
218 msg = "g_c_p_n failed";
219 GOTO(out, rc = -EBUSY);
222 /* Check if vmpage was truncated or reclaimed */
223 if (vmpage->mapping != inode->i_mapping) {
224 which = RA_STAT_WRONG_GRAB_PAGE;
225 msg = "g_c_p_n returned invalid page";
226 GOTO(out, rc = -EBUSY);
229 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
231 which = RA_STAT_FAILED_GRAB_PAGE;
232 msg = "cl_page_find failed";
233 GOTO(out, rc = PTR_ERR(page));
236 lu_ref_add(&page->cp_reference, "ra", current);
237 cl_page_assume(env, io, page);
238 vpg = cl2vvp_page(cl_object_page_slice(clob, page));
239 if (!vpg->vpg_defer_uptodate && !PageUptodate(vmpage)) {
240 vpg->vpg_defer_uptodate = 1;
241 vpg->vpg_ra_used = 0;
242 cl_page_list_add(queue, page);
244 /* skip completed pages */
245 cl_page_unassume(env, io, page);
246 /* This page is already uptodate, returning a positive number
247 * to tell the callers about this */
251 lu_ref_del(&page->cp_reference, "ra", current);
252 cl_page_put(env, page);
255 if (vmpage != NULL) {
258 page_cache_release(vmpage);
261 ll_ra_stats_inc(inode, which);
262 CDEBUG(D_READA, "%s\n", msg);
269 #define RIA_DEBUG(ria) \
270 CDEBUG(D_READA, "rs %lu re %lu ro %lu rl %lu rp %lu\n", \
271 ria->ria_start, ria->ria_end, ria->ria_stoff, ria->ria_length,\
274 /* Limit this to the blocksize instead of PTLRPC_BRW_MAX_SIZE, since we don't
275 * know what the actual RPC size is. If this needs to change, it makes more
276 * sense to tune the i_blkbits value for the file based on the OSTs it is
277 * striped over, rather than having a constant value for all files here. */
279 /* RAS_INCREASE_STEP should be (1UL << (inode->i_blkbits - PAGE_CACHE_SHIFT)).
280 * Temprarily set RAS_INCREASE_STEP to 1MB. After 4MB RPC is enabled
281 * by default, this should be adjusted corresponding with max_read_ahead_mb
282 * and max_read_ahead_per_file_mb otherwise the readahead budget can be used
283 * up quickly which will affect read performance siginificantly. See LU-2816 */
284 #define RAS_INCREASE_STEP(inode) (ONE_MB_BRW_SIZE >> PAGE_CACHE_SHIFT)
286 static inline int stride_io_mode(struct ll_readahead_state *ras)
288 return ras->ras_consecutive_stride_requests > 1;
290 /* The function calculates how much pages will be read in
291 * [off, off + length], in such stride IO area,
292 * stride_offset = st_off, stride_lengh = st_len,
293 * stride_pages = st_pgs
295 * |------------------|*****|------------------|*****|------------|*****|....
298 * |----- st_len -----|
300 * How many pages it should read in such pattern
301 * |-------------------------------------------------------------|
303 * |<------ length ------->|
305 * = |<----->| + |-------------------------------------| + |---|
306 * start_left st_pgs * i end_left
309 stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs,
310 unsigned long off, unsigned long length)
312 __u64 start = off > st_off ? off - st_off : 0;
313 __u64 end = off + length > st_off ? off + length - st_off : 0;
314 unsigned long start_left = 0;
315 unsigned long end_left = 0;
316 unsigned long pg_count;
318 if (st_len == 0 || length == 0 || end == 0)
321 start_left = do_div(start, st_len);
322 if (start_left < st_pgs)
323 start_left = st_pgs - start_left;
327 end_left = do_div(end, st_len);
328 if (end_left > st_pgs)
331 CDEBUG(D_READA, "start "LPU64", end "LPU64" start_left %lu end_left %lu \n",
332 start, end, start_left, end_left);
335 pg_count = end_left - (st_pgs - start_left);
337 pg_count = start_left + st_pgs * (end - start - 1) + end_left;
339 CDEBUG(D_READA, "st_off %lu, st_len %lu st_pgs %lu off %lu length %lu"
340 "pgcount %lu\n", st_off, st_len, st_pgs, off, length, pg_count);
345 static int ria_page_count(struct ra_io_arg *ria)
347 __u64 length = ria->ria_end >= ria->ria_start ?
348 ria->ria_end - ria->ria_start + 1 : 0;
350 return stride_pg_count(ria->ria_stoff, ria->ria_length,
351 ria->ria_pages, ria->ria_start,
355 /*Check whether the index is in the defined ra-window */
356 static int ras_inside_ra_window(unsigned long idx, struct ra_io_arg *ria)
358 /* If ria_length == ria_pages, it means non-stride I/O mode,
359 * idx should always inside read-ahead window in this case
360 * For stride I/O mode, just check whether the idx is inside
362 return ria->ria_length == 0 || ria->ria_length == ria->ria_pages ||
363 (idx >= ria->ria_stoff && (idx - ria->ria_stoff) %
364 ria->ria_length < ria->ria_pages);
367 static int ll_read_ahead_pages(const struct lu_env *env,
368 struct cl_io *io, struct cl_page_list *queue,
369 struct ra_io_arg *ria,
370 unsigned long *reserved_pages,
373 struct cl_read_ahead ra = { 0 };
378 LASSERT(ria != NULL);
381 stride_ria = ria->ria_length > ria->ria_pages && ria->ria_pages > 0;
382 for (page_idx = ria->ria_start;
383 page_idx <= ria->ria_end && *reserved_pages > 0; page_idx++) {
384 if (ras_inside_ra_window(page_idx, ria)) {
385 if (ra.cra_end == 0 || ra.cra_end < page_idx) {
386 cl_read_ahead_release(env, &ra);
388 rc = cl_io_read_ahead(env, io, page_idx, &ra);
392 LASSERTF(ra.cra_end >= page_idx,
393 "object: %p, indcies %lu / %lu\n",
394 io->ci_obj, ra.cra_end, page_idx);
397 /* If the page is inside the read-ahead window*/
398 rc = ll_read_ahead_page(env, io, queue, page_idx);
403 } else if (stride_ria) {
404 /* If it is not in the read-ahead window, and it is
405 * read-ahead mode, then check whether it should skip
408 /* FIXME: This assertion only is valid when it is for
409 * forward read-ahead, it will be fixed when backward
410 * read-ahead is implemented */
411 LASSERTF(page_idx >= ria->ria_stoff,
412 "Invalid page_idx %lu rs %lu re %lu ro %lu "
413 "rl %lu rp %lu\n", page_idx,
414 ria->ria_start, ria->ria_end, ria->ria_stoff,
415 ria->ria_length, ria->ria_pages);
416 offset = page_idx - ria->ria_stoff;
417 offset = offset % (ria->ria_length);
418 if (offset > ria->ria_pages) {
419 page_idx += ria->ria_length - offset;
420 CDEBUG(D_READA, "i %lu skip %lu \n", page_idx,
421 ria->ria_length - offset);
427 cl_read_ahead_release(env, &ra);
433 static int ll_readahead(const struct lu_env *env, struct cl_io *io,
434 struct cl_page_list *queue,
435 struct ll_readahead_state *ras, bool hit)
437 struct vvp_io *vio = vvp_env_io(env);
438 struct ll_thread_info *lti = ll_env_info(env);
439 struct cl_attr *attr = vvp_env_thread_attr(env);
440 unsigned long len, mlen = 0, reserved;
441 pgoff_t ra_end, start = 0, end = 0;
443 struct ra_io_arg *ria = <i->lti_ria;
444 struct cl_object *clob;
450 inode = vvp_object_inode(clob);
452 memset(ria, 0, sizeof *ria);
454 cl_object_attr_lock(clob);
455 ret = cl_object_attr_get(env, clob, attr);
456 cl_object_attr_unlock(clob);
462 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
466 spin_lock(&ras->ras_lock);
468 /* Enlarge the RA window to encompass the full read */
469 if (vio->vui_ra_valid &&
470 ras->ras_window_start + ras->ras_window_len <
471 vio->vui_ra_start + vio->vui_ra_count) {
472 ras->ras_window_len = vio->vui_ra_start + vio->vui_ra_count -
473 ras->ras_window_start;
476 /* Reserve a part of the read-ahead window that we'll be issuing */
477 if (ras->ras_window_len > 0) {
479 * Note: other thread might rollback the ras_next_readahead,
480 * if it can not get the full size of prepared pages, see the
481 * end of this function. For stride read ahead, it needs to
482 * make sure the offset is no less than ras_stride_offset,
483 * so that stride read ahead can work correctly.
485 if (stride_io_mode(ras))
486 start = max(ras->ras_next_readahead,
487 ras->ras_stride_offset);
489 start = ras->ras_next_readahead;
490 end = ras->ras_window_start + ras->ras_window_len - 1;
494 unsigned long rpc_boundary;
496 * Align RA window to an optimal boundary.
498 * XXX This would be better to align to cl_max_pages_per_rpc
499 * instead of PTLRPC_MAX_BRW_PAGES, because the RPC size may
500 * be aligned to the RAID stripe size in the future and that
501 * is more important than the RPC size.
503 /* Note: we only trim the RPC, instead of extending the RPC
504 * to the boundary, so to avoid reading too much pages during
506 rpc_boundary = ((end + 1) & (~(PTLRPC_MAX_BRW_PAGES - 1)));
507 if (rpc_boundary > 0)
510 if (rpc_boundary > start)
513 /* Truncate RA window to end of file */
514 end = min(end, (unsigned long)((kms - 1) >> PAGE_CACHE_SHIFT));
516 ras->ras_next_readahead = max(end, end + 1);
519 ria->ria_start = start;
521 /* If stride I/O mode is detected, get stride window*/
522 if (stride_io_mode(ras)) {
523 ria->ria_stoff = ras->ras_stride_offset;
524 ria->ria_length = ras->ras_stride_length;
525 ria->ria_pages = ras->ras_stride_pages;
527 spin_unlock(&ras->ras_lock);
530 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
533 len = ria_page_count(ria);
535 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
539 CDEBUG(D_READA, DFID": ria: %lu/%lu, bead: %lu/%lu, hit: %d\n",
540 PFID(lu_object_fid(&clob->co_lu)),
541 ria->ria_start, ria->ria_end,
542 vio->vui_ra_valid ? vio->vui_ra_start : 0,
543 vio->vui_ra_valid ? vio->vui_ra_count : 0,
546 /* at least to extend the readahead window to cover current read */
547 if (!hit && vio->vui_ra_valid &&
548 vio->vui_ra_start + vio->vui_ra_count > ria->ria_start) {
549 /* to the end of current read window. */
550 mlen = vio->vui_ra_start + vio->vui_ra_count - ria->ria_start;
551 /* trim to RPC boundary */
552 start = ria->ria_start & (PTLRPC_MAX_BRW_PAGES - 1);
553 mlen = min(mlen, PTLRPC_MAX_BRW_PAGES - start);
556 reserved = ll_ra_count_get(ll_i2sbi(inode), ria, len, mlen);
558 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
560 CDEBUG(D_READA, "reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
562 atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
563 ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
565 ret = ll_read_ahead_pages(env, io, queue, ria, &reserved, &ra_end);
568 ll_ra_count_put(ll_i2sbi(inode), reserved);
570 if (ra_end == end + 1 && ra_end == (kms >> PAGE_CACHE_SHIFT))
571 ll_ra_stats_inc(inode, RA_STAT_EOF);
573 /* if we didn't get to the end of the region we reserved from
574 * the ras we need to go back and update the ras so that the
575 * next read-ahead tries from where we left off. we only do so
576 * if the region we failed to issue read-ahead on is still ahead
577 * of the app and behind the next index to start read-ahead from */
578 CDEBUG(D_READA, "ra_end = %lu end = %lu stride end = %lu pages = %d\n",
579 ra_end, end, ria->ria_end, ret);
581 if (ra_end != end + 1) {
582 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
583 spin_lock(&ras->ras_lock);
584 if (ra_end < ras->ras_next_readahead &&
585 index_in_window(ra_end, ras->ras_window_start, 0,
586 ras->ras_window_len)) {
587 ras->ras_next_readahead = ra_end;
590 spin_unlock(&ras->ras_lock);
596 static void ras_set_start(struct inode *inode, struct ll_readahead_state *ras,
599 ras->ras_window_start = index & (~(RAS_INCREASE_STEP(inode) - 1));
602 /* called with the ras_lock held or from places where it doesn't matter */
603 static void ras_reset(struct inode *inode, struct ll_readahead_state *ras,
606 ras->ras_last_readpage = index;
607 ras->ras_consecutive_requests = 0;
608 ras->ras_consecutive_pages = 0;
609 ras->ras_window_len = 0;
610 ras_set_start(inode, ras, index);
611 ras->ras_next_readahead = max(ras->ras_window_start, index);
616 /* called with the ras_lock held or from places where it doesn't matter */
617 static void ras_stride_reset(struct ll_readahead_state *ras)
619 ras->ras_consecutive_stride_requests = 0;
620 ras->ras_stride_length = 0;
621 ras->ras_stride_pages = 0;
625 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
627 spin_lock_init(&ras->ras_lock);
628 ras_reset(inode, ras, 0);
629 ras->ras_requests = 0;
633 * Check whether the read request is in the stride window.
634 * If it is in the stride window, return 1, otherwise return 0.
636 static int index_in_stride_window(struct ll_readahead_state *ras,
639 unsigned long stride_gap;
641 if (ras->ras_stride_length == 0 || ras->ras_stride_pages == 0 ||
642 ras->ras_stride_pages == ras->ras_stride_length)
645 stride_gap = index - ras->ras_last_readpage - 1;
647 /* If it is contiguous read */
649 return ras->ras_consecutive_pages + 1 <= ras->ras_stride_pages;
651 /* Otherwise check the stride by itself */
652 return (ras->ras_stride_length - ras->ras_stride_pages) == stride_gap &&
653 ras->ras_consecutive_pages == ras->ras_stride_pages;
656 static void ras_update_stride_detector(struct ll_readahead_state *ras,
659 unsigned long stride_gap = index - ras->ras_last_readpage - 1;
661 if (!stride_io_mode(ras) && (stride_gap != 0 ||
662 ras->ras_consecutive_stride_requests == 0)) {
663 ras->ras_stride_pages = ras->ras_consecutive_pages;
664 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
666 LASSERT(ras->ras_request_index == 0);
667 LASSERT(ras->ras_consecutive_stride_requests == 0);
669 if (index <= ras->ras_last_readpage) {
670 /*Reset stride window for forward read*/
671 ras_stride_reset(ras);
675 ras->ras_stride_pages = ras->ras_consecutive_pages;
676 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
683 stride_page_count(struct ll_readahead_state *ras, unsigned long len)
685 return stride_pg_count(ras->ras_stride_offset, ras->ras_stride_length,
686 ras->ras_stride_pages, ras->ras_stride_offset,
690 /* Stride Read-ahead window will be increased inc_len according to
691 * stride I/O pattern */
692 static void ras_stride_increase_window(struct ll_readahead_state *ras,
693 struct ll_ra_info *ra,
694 unsigned long inc_len)
696 unsigned long left, step, window_len;
697 unsigned long stride_len;
699 LASSERT(ras->ras_stride_length > 0);
700 LASSERTF(ras->ras_window_start + ras->ras_window_len
701 >= ras->ras_stride_offset, "window_start %lu, window_len %lu"
702 " stride_offset %lu\n", ras->ras_window_start,
703 ras->ras_window_len, ras->ras_stride_offset);
705 stride_len = ras->ras_window_start + ras->ras_window_len -
706 ras->ras_stride_offset;
708 left = stride_len % ras->ras_stride_length;
709 window_len = ras->ras_window_len - left;
711 if (left < ras->ras_stride_pages)
714 left = ras->ras_stride_pages + inc_len;
716 LASSERT(ras->ras_stride_pages != 0);
718 step = left / ras->ras_stride_pages;
719 left %= ras->ras_stride_pages;
721 window_len += step * ras->ras_stride_length + left;
723 if (stride_page_count(ras, window_len) <= ra->ra_max_pages_per_file)
724 ras->ras_window_len = window_len;
729 static void ras_increase_window(struct inode *inode,
730 struct ll_readahead_state *ras,
731 struct ll_ra_info *ra)
733 /* The stretch of ra-window should be aligned with max rpc_size
734 * but current clio architecture does not support retrieve such
735 * information from lower layer. FIXME later
737 if (stride_io_mode(ras))
738 ras_stride_increase_window(ras, ra, RAS_INCREASE_STEP(inode));
740 ras->ras_window_len = min(ras->ras_window_len +
741 RAS_INCREASE_STEP(inode),
742 ra->ra_max_pages_per_file);
745 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
746 struct ll_readahead_state *ras, unsigned long index,
749 struct ll_ra_info *ra = &sbi->ll_ra_info;
750 int zero = 0, stride_detect = 0, ra_miss = 0;
753 spin_lock(&ras->ras_lock);
755 ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
757 /* reset the read-ahead window in two cases. First when the app seeks
758 * or reads to some other part of the file. Secondly if we get a
759 * read-ahead miss that we think we've previously issued. This can
760 * be a symptom of there being so many read-ahead pages that the VM is
761 * reclaiming it before we get to it. */
762 if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
764 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
765 } else if (!hit && ras->ras_window_len &&
766 index < ras->ras_next_readahead &&
767 index_in_window(index, ras->ras_window_start, 0,
768 ras->ras_window_len)) {
770 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
773 /* On the second access to a file smaller than the tunable
774 * ra_max_read_ahead_whole_pages trigger RA on all pages in the
775 * file up to ra_max_pages_per_file. This is simply a best effort
776 * and only occurs once per open file. Normal RA behavior is reverted
777 * to for subsequent IO. The mmap case does not increment
778 * ras_requests and thus can never trigger this behavior. */
779 if (ras->ras_requests == 2 && !ras->ras_request_index) {
782 kms_pages = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >>
785 CDEBUG(D_READA, "kmsp "LPU64" mwp %lu mp %lu\n", kms_pages,
786 ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages_per_file);
789 kms_pages <= ra->ra_max_read_ahead_whole_pages) {
790 ras->ras_window_start = 0;
791 ras->ras_last_readpage = 0;
792 ras->ras_next_readahead = 0;
793 ras->ras_window_len = min(ra->ra_max_pages_per_file,
794 ra->ra_max_read_ahead_whole_pages);
799 /* check whether it is in stride I/O mode*/
800 if (!index_in_stride_window(ras, index)) {
801 if (ras->ras_consecutive_stride_requests == 0 &&
802 ras->ras_request_index == 0) {
803 ras_update_stride_detector(ras, index);
804 ras->ras_consecutive_stride_requests++;
806 ras_stride_reset(ras);
808 ras_reset(inode, ras, index);
809 ras->ras_consecutive_pages++;
812 ras->ras_consecutive_pages = 0;
813 ras->ras_consecutive_requests = 0;
814 if (++ras->ras_consecutive_stride_requests > 1)
820 if (index_in_stride_window(ras, index) &&
821 stride_io_mode(ras)) {
822 /*If stride-RA hit cache miss, the stride dector
823 *will not be reset to avoid the overhead of
824 *redetecting read-ahead mode */
825 if (index != ras->ras_last_readpage + 1)
826 ras->ras_consecutive_pages = 0;
827 ras_reset(inode, ras, index);
830 /* Reset both stride window and normal RA
832 ras_reset(inode, ras, index);
833 ras->ras_consecutive_pages++;
834 ras_stride_reset(ras);
837 } else if (stride_io_mode(ras)) {
838 /* If this is contiguous read but in stride I/O mode
839 * currently, check whether stride step still is valid,
840 * if invalid, it will reset the stride ra window*/
841 if (!index_in_stride_window(ras, index)) {
842 /* Shrink stride read-ahead window to be zero */
843 ras_stride_reset(ras);
844 ras->ras_window_len = 0;
845 ras->ras_next_readahead = index;
849 ras->ras_consecutive_pages++;
850 ras->ras_last_readpage = index;
851 ras_set_start(inode, ras, index);
853 if (stride_io_mode(ras)) {
854 /* Since stride readahead is sentivite to the offset
855 * of read-ahead, so we use original offset here,
856 * instead of ras_window_start, which is RPC aligned */
857 ras->ras_next_readahead = max(index, ras->ras_next_readahead);
859 if (ras->ras_next_readahead < ras->ras_window_start)
860 ras->ras_next_readahead = ras->ras_window_start;
862 ras->ras_next_readahead = index + 1;
866 /* Trigger RA in the mmap case where ras_consecutive_requests
867 * is not incremented and thus can't be used to trigger RA */
868 if (!ras->ras_window_len && ras->ras_consecutive_pages == 4) {
869 ras->ras_window_len = RAS_INCREASE_STEP(inode);
873 /* Initially reset the stride window offset to next_readahead*/
874 if (ras->ras_consecutive_stride_requests == 2 && stride_detect) {
876 * Once stride IO mode is detected, next_readahead should be
877 * reset to make sure next_readahead > stride offset
879 ras->ras_next_readahead = max(index, ras->ras_next_readahead);
880 ras->ras_stride_offset = index;
881 ras->ras_window_len = RAS_INCREASE_STEP(inode);
884 /* The initial ras_window_len is set to the request size. To avoid
885 * uselessly reading and discarding pages for random IO the window is
886 * only increased once per consecutive request received. */
887 if ((ras->ras_consecutive_requests > 1 || stride_detect) &&
888 !ras->ras_request_index)
889 ras_increase_window(inode, ras, ra);
893 ras->ras_request_index++;
894 spin_unlock(&ras->ras_lock);
898 int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
900 struct inode *inode = vmpage->mapping->host;
901 struct ll_inode_info *lli = ll_i2info(inode);
904 struct cl_page *page;
905 struct cl_object *clob;
906 struct cl_env_nest nest;
907 bool redirtied = false;
908 bool unlocked = false;
912 LASSERT(PageLocked(vmpage));
913 LASSERT(!PageWriteback(vmpage));
915 LASSERT(ll_i2dtexp(inode) != NULL);
917 env = cl_env_nested_get(&nest);
919 GOTO(out, result = PTR_ERR(env));
921 clob = ll_i2info(inode)->lli_clob;
922 LASSERT(clob != NULL);
924 io = vvp_env_thread_io(env);
926 io->ci_ignore_layout = 1;
927 result = cl_io_init(env, io, CIT_MISC, clob);
929 page = cl_page_find(env, clob, vmpage->index,
930 vmpage, CPT_CACHEABLE);
932 lu_ref_add(&page->cp_reference, "writepage",
934 cl_page_assume(env, io, page);
935 result = cl_page_flush(env, io, page);
938 * Re-dirty page on error so it retries write,
939 * but not in case when IO has actually
940 * occurred and completed with an error.
942 if (!PageError(vmpage)) {
943 redirty_page_for_writepage(wbc, vmpage);
948 cl_page_disown(env, io, page);
950 lu_ref_del(&page->cp_reference,
951 "writepage", current);
952 cl_page_put(env, page);
954 result = PTR_ERR(page);
959 if (redirtied && wbc->sync_mode == WB_SYNC_ALL) {
960 loff_t offset = cl_offset(clob, vmpage->index);
962 /* Flush page failed because the extent is being written out.
963 * Wait for the write of extent to be finished to avoid
964 * breaking kernel which assumes ->writepage should mark
965 * PageWriteback or clean the page. */
966 result = cl_sync_file_range(inode, offset,
967 offset + PAGE_CACHE_SIZE - 1,
970 /* actually we may have written more than one page.
971 * decreasing this page because the caller will count
973 wbc->nr_to_write -= result - 1;
978 cl_env_nested_put(&nest, env);
983 if (!lli->lli_async_rc)
984 lli->lli_async_rc = result;
985 SetPageError(vmpage);
992 int ll_writepages(struct address_space *mapping, struct writeback_control *wbc)
994 struct inode *inode = mapping->host;
995 struct ll_sb_info *sbi = ll_i2sbi(inode);
998 enum cl_fsync_mode mode;
1001 int ignore_layout = 0;
1004 if (wbc->range_cyclic) {
1005 start = mapping->writeback_index << PAGE_CACHE_SHIFT;
1006 end = OBD_OBJECT_EOF;
1008 start = wbc->range_start;
1009 end = wbc->range_end;
1010 if (end == LLONG_MAX) {
1011 end = OBD_OBJECT_EOF;
1012 range_whole = start == 0;
1016 mode = CL_FSYNC_NONE;
1017 if (wbc->sync_mode == WB_SYNC_ALL)
1018 mode = CL_FSYNC_LOCAL;
1020 if (sbi->ll_umounting)
1021 /* if the mountpoint is being umounted, all pages have to be
1022 * evicted to avoid hitting LBUG when truncate_inode_pages()
1023 * is called later on. */
1026 if (ll_i2info(inode)->lli_clob == NULL)
1029 result = cl_sync_file_range(inode, start, end, mode, ignore_layout);
1031 wbc->nr_to_write -= result;
1035 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) {
1036 if (end == OBD_OBJECT_EOF)
1037 mapping->writeback_index = 0;
1039 mapping->writeback_index = (end >> PAGE_CACHE_SHIFT) +1;
1044 struct ll_cl_context *ll_cl_find(struct file *file)
1046 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1047 struct ll_cl_context *lcc;
1048 struct ll_cl_context *found = NULL;
1050 read_lock(&fd->fd_lock);
1051 list_for_each_entry(lcc, &fd->fd_lccs, lcc_list) {
1052 if (lcc->lcc_cookie == current) {
1057 read_unlock(&fd->fd_lock);
1062 void ll_cl_add(struct file *file, const struct lu_env *env, struct cl_io *io)
1064 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1065 struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1067 memset(lcc, 0, sizeof(*lcc));
1068 INIT_LIST_HEAD(&lcc->lcc_list);
1069 lcc->lcc_cookie = current;
1073 write_lock(&fd->fd_lock);
1074 list_add(&lcc->lcc_list, &fd->fd_lccs);
1075 write_unlock(&fd->fd_lock);
1078 void ll_cl_remove(struct file *file, const struct lu_env *env)
1080 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1081 struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1083 write_lock(&fd->fd_lock);
1084 list_del_init(&lcc->lcc_list);
1085 write_unlock(&fd->fd_lock);
1088 static int ll_io_read_page(const struct lu_env *env, struct cl_io *io,
1089 struct cl_page *page)
1091 struct inode *inode = vvp_object_inode(page->cp_obj);
1092 struct ll_sb_info *sbi = ll_i2sbi(inode);
1093 struct ll_file_data *fd = vvp_env_io(env)->vui_fd;
1094 struct ll_readahead_state *ras = &fd->fd_ras;
1095 struct cl_2queue *queue = &io->ci_queue;
1096 struct vvp_page *vpg;
1100 vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1101 if (sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
1102 sbi->ll_ra_info.ra_max_pages > 0)
1103 ras_update(sbi, inode, ras, vvp_index(vpg),
1104 vpg->vpg_defer_uptodate);
1106 if (vpg->vpg_defer_uptodate) {
1107 vpg->vpg_ra_used = 1;
1108 cl_page_export(env, page, 1);
1111 cl_2queue_init(queue);
1113 * Add page into the queue even when it is marked uptodate above.
1114 * this will unlock it automatically as part of cl_page_list_disown().
1116 cl_2queue_add(queue, page);
1117 if (sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
1118 sbi->ll_ra_info.ra_max_pages > 0) {
1121 rc2 = ll_readahead(env, io, &queue->c2_qin, ras,
1122 vpg->vpg_defer_uptodate);
1123 CDEBUG(D_READA, DFID "%d pages read ahead at %lu\n",
1124 PFID(ll_inode2fid(inode)), rc2, vvp_index(vpg));
1127 if (queue->c2_qin.pl_nr > 0)
1128 rc = cl_io_submit_rw(env, io, CRT_READ, queue);
1131 * Unlock unsent pages in case of error.
1133 cl_page_list_disown(env, io, &queue->c2_qin);
1134 cl_2queue_fini(env, queue);
1139 int ll_readpage(struct file *file, struct page *vmpage)
1141 struct cl_object *clob = ll_i2info(file->f_dentry->d_inode)->lli_clob;
1142 struct ll_cl_context *lcc;
1143 const struct lu_env *env;
1145 struct cl_page *page;
1149 lcc = ll_cl_find(file);
1151 unlock_page(vmpage);
1157 LASSERT(io != NULL);
1158 LASSERT(io->ci_state == CIS_IO_GOING);
1159 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
1160 if (!IS_ERR(page)) {
1161 LASSERT(page->cp_type == CPT_CACHEABLE);
1162 if (likely(!PageUptodate(vmpage))) {
1163 cl_page_assume(env, io, page);
1164 result = ll_io_read_page(env, io, page);
1166 /* Page from a non-object file. */
1167 unlock_page(vmpage);
1170 cl_page_put(env, page);
1172 unlock_page(vmpage);
1173 result = PTR_ERR(page);
1178 int ll_page_sync_io(const struct lu_env *env, struct cl_io *io,
1179 struct cl_page *page, enum cl_req_type crt)
1181 struct cl_2queue *queue;
1184 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
1186 queue = &io->ci_queue;
1187 cl_2queue_init_page(queue, page);
1189 result = cl_io_submit_sync(env, io, crt, queue, 0);
1190 LASSERT(cl_page_is_owned(page, io));
1192 if (crt == CRT_READ)
1194 * in CRT_WRITE case page is left locked even in case of
1197 cl_page_list_disown(env, io, &queue->c2_qin);
1198 cl_2queue_fini(env, queue);