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, 2015, 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 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which);
67 * Get readahead pages from the filesystem readahead pool of the client for a
70 * /param sbi superblock for filesystem readahead state ll_ra_info
71 * /param ria per-thread readahead state
72 * /param pages number of pages requested for readahead for the thread.
74 * WARNING: This algorithm is used to reduce contention on sbi->ll_lock.
75 * It should work well if the ra_max_pages is much greater than the single
76 * file's read-ahead window, and not too many threads contending for
77 * these readahead pages.
79 * TODO: There may be a 'global sync problem' if many threads are trying
80 * to get an ra budget that is larger than the remaining readahead pages
81 * and reach here at exactly the same time. They will compute /a ret to
82 * consume the remaining pages, but will fail at atomic_add_return() and
83 * get a zero ra window, although there is still ra space remaining. - Jay */
85 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi,
86 struct ra_io_arg *ria,
87 unsigned long pages, unsigned long min)
89 struct ll_ra_info *ra = &sbi->ll_ra_info;
93 /* If read-ahead pages left are less than 1M, do not do read-ahead,
94 * otherwise it will form small read RPC(< 1M), which hurt server
95 * performance a lot. */
96 ret = min(ra->ra_max_pages - atomic_read(&ra->ra_cur_pages),
98 if (ret < 0 || ret < min_t(long, PTLRPC_MAX_BRW_PAGES, pages))
101 /* If the non-strided (ria_pages == 0) readahead window
102 * (ria_start + ret) has grown across an RPC boundary, then trim
103 * readahead size by the amount beyond the RPC so it ends on an
104 * RPC boundary. If the readahead window is already ending on
105 * an RPC boundary (beyond_rpc == 0), or smaller than a full
106 * RPC (beyond_rpc < ret) the readahead size is unchanged.
107 * The (beyond_rpc != 0) check is skipped since the conditional
108 * branch is more expensive than subtracting zero from the result.
110 * Strided read is left unaligned to avoid small fragments beyond
111 * the RPC boundary from needing an extra read RPC. */
112 if (ria->ria_pages == 0) {
113 long beyond_rpc = (ria->ria_start + ret) % PTLRPC_MAX_BRW_PAGES;
114 if (/* beyond_rpc != 0 && */ beyond_rpc < ret)
118 if (atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) {
119 atomic_sub(ret, &ra->ra_cur_pages);
125 /* override ra limit for maximum performance */
126 atomic_add(min - ret, &ra->ra_cur_pages);
132 void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
134 struct ll_ra_info *ra = &sbi->ll_ra_info;
135 atomic_sub(len, &ra->ra_cur_pages);
138 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which)
140 LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
141 lprocfs_counter_incr(sbi->ll_ra_stats, which);
144 void ll_ra_stats_inc(struct inode *inode, enum ra_stat which)
146 struct ll_sb_info *sbi = ll_i2sbi(inode);
147 ll_ra_stats_inc_sbi(sbi, which);
150 #define RAS_CDEBUG(ras) \
152 "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu r %lu ri %lu" \
153 "csr %lu sf %lu sp %lu sl %lu \n", \
154 ras->ras_last_readpage, ras->ras_consecutive_requests, \
155 ras->ras_consecutive_pages, ras->ras_window_start, \
156 ras->ras_window_len, ras->ras_next_readahead, \
157 ras->ras_requests, ras->ras_request_index, \
158 ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
159 ras->ras_stride_pages, ras->ras_stride_length)
161 static int index_in_window(unsigned long index, unsigned long point,
162 unsigned long before, unsigned long after)
164 unsigned long start = point - before, end = point + after;
171 return start <= index && index <= end;
174 void ll_ras_enter(struct file *f)
176 struct ll_file_data *fd = LUSTRE_FPRIVATE(f);
177 struct ll_readahead_state *ras = &fd->fd_ras;
179 spin_lock(&ras->ras_lock);
181 ras->ras_request_index = 0;
182 ras->ras_consecutive_requests++;
183 spin_unlock(&ras->ras_lock);
187 * Initiates read-ahead of a page with given index.
189 * \retval +ve: page was already uptodate so it will be skipped
191 * \retval -ve: page wasn't added to \a queue for error;
192 * \retval 0: page was added into \a queue for read ahead.
194 static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io,
195 struct cl_page_list *queue, pgoff_t index)
197 struct cl_object *clob = io->ci_obj;
198 struct inode *inode = vvp_object_inode(clob);
200 struct cl_page *page;
201 struct vvp_page *vpg;
202 enum ra_stat which = _NR_RA_STAT; /* keep gcc happy */
204 const char *msg = NULL;
207 vmpage = grab_cache_page_nowait(inode->i_mapping, index);
208 if (vmpage == NULL) {
209 which = RA_STAT_FAILED_GRAB_PAGE;
210 msg = "g_c_p_n failed";
211 GOTO(out, rc = -EBUSY);
214 /* Check if vmpage was truncated or reclaimed */
215 if (vmpage->mapping != inode->i_mapping) {
216 which = RA_STAT_WRONG_GRAB_PAGE;
217 msg = "g_c_p_n returned invalid page";
218 GOTO(out, rc = -EBUSY);
221 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
223 which = RA_STAT_FAILED_GRAB_PAGE;
224 msg = "cl_page_find failed";
225 GOTO(out, rc = PTR_ERR(page));
228 lu_ref_add(&page->cp_reference, "ra", current);
229 cl_page_assume(env, io, page);
230 vpg = cl2vvp_page(cl_object_page_slice(clob, page));
231 if (!vpg->vpg_defer_uptodate && !PageUptodate(vmpage)) {
232 vpg->vpg_defer_uptodate = 1;
233 vpg->vpg_ra_used = 0;
234 cl_page_list_add(queue, page);
236 /* skip completed pages */
237 cl_page_unassume(env, io, page);
238 /* This page is already uptodate, returning a positive number
239 * to tell the callers about this */
243 lu_ref_del(&page->cp_reference, "ra", current);
244 cl_page_put(env, page);
247 if (vmpage != NULL) {
250 page_cache_release(vmpage);
253 ll_ra_stats_inc(inode, which);
254 CDEBUG(D_READA, "%s\n", msg);
261 #define RIA_DEBUG(ria) \
262 CDEBUG(D_READA, "rs %lu re %lu ro %lu rl %lu rp %lu\n", \
263 ria->ria_start, ria->ria_end, ria->ria_stoff, ria->ria_length,\
266 /* Limit this to the blocksize instead of PTLRPC_BRW_MAX_SIZE, since we don't
267 * know what the actual RPC size is. If this needs to change, it makes more
268 * sense to tune the i_blkbits value for the file based on the OSTs it is
269 * striped over, rather than having a constant value for all files here. */
271 /* RAS_INCREASE_STEP should be (1UL << (inode->i_blkbits - PAGE_CACHE_SHIFT)).
272 * Temprarily set RAS_INCREASE_STEP to 1MB. After 4MB RPC is enabled
273 * by default, this should be adjusted corresponding with max_read_ahead_mb
274 * and max_read_ahead_per_file_mb otherwise the readahead budget can be used
275 * up quickly which will affect read performance siginificantly. See LU-2816 */
276 #define RAS_INCREASE_STEP(inode) (ONE_MB_BRW_SIZE >> PAGE_CACHE_SHIFT)
278 static inline int stride_io_mode(struct ll_readahead_state *ras)
280 return ras->ras_consecutive_stride_requests > 1;
282 /* The function calculates how much pages will be read in
283 * [off, off + length], in such stride IO area,
284 * stride_offset = st_off, stride_lengh = st_len,
285 * stride_pages = st_pgs
287 * |------------------|*****|------------------|*****|------------|*****|....
290 * |----- st_len -----|
292 * How many pages it should read in such pattern
293 * |-------------------------------------------------------------|
295 * |<------ length ------->|
297 * = |<----->| + |-------------------------------------| + |---|
298 * start_left st_pgs * i end_left
301 stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs,
302 unsigned long off, unsigned long length)
304 __u64 start = off > st_off ? off - st_off : 0;
305 __u64 end = off + length > st_off ? off + length - st_off : 0;
306 unsigned long start_left = 0;
307 unsigned long end_left = 0;
308 unsigned long pg_count;
310 if (st_len == 0 || length == 0 || end == 0)
313 start_left = do_div(start, st_len);
314 if (start_left < st_pgs)
315 start_left = st_pgs - start_left;
319 end_left = do_div(end, st_len);
320 if (end_left > st_pgs)
323 CDEBUG(D_READA, "start "LPU64", end "LPU64" start_left %lu end_left %lu \n",
324 start, end, start_left, end_left);
327 pg_count = end_left - (st_pgs - start_left);
329 pg_count = start_left + st_pgs * (end - start - 1) + end_left;
331 CDEBUG(D_READA, "st_off %lu, st_len %lu st_pgs %lu off %lu length %lu"
332 "pgcount %lu\n", st_off, st_len, st_pgs, off, length, pg_count);
337 static int ria_page_count(struct ra_io_arg *ria)
339 __u64 length = ria->ria_end >= ria->ria_start ?
340 ria->ria_end - ria->ria_start + 1 : 0;
342 return stride_pg_count(ria->ria_stoff, ria->ria_length,
343 ria->ria_pages, ria->ria_start,
347 /*Check whether the index is in the defined ra-window */
348 static int ras_inside_ra_window(unsigned long idx, struct ra_io_arg *ria)
350 /* If ria_length == ria_pages, it means non-stride I/O mode,
351 * idx should always inside read-ahead window in this case
352 * For stride I/O mode, just check whether the idx is inside
354 return ria->ria_length == 0 || ria->ria_length == ria->ria_pages ||
355 (idx >= ria->ria_stoff && (idx - ria->ria_stoff) %
356 ria->ria_length < ria->ria_pages);
359 static int ll_read_ahead_pages(const struct lu_env *env,
360 struct cl_io *io, struct cl_page_list *queue,
361 struct ra_io_arg *ria,
362 unsigned long *reserved_pages,
365 struct cl_read_ahead ra = { 0 };
370 LASSERT(ria != NULL);
373 stride_ria = ria->ria_length > ria->ria_pages && ria->ria_pages > 0;
374 for (page_idx = ria->ria_start;
375 page_idx <= ria->ria_end && *reserved_pages > 0; page_idx++) {
376 if (ras_inside_ra_window(page_idx, ria)) {
377 if (ra.cra_end == 0 || ra.cra_end < page_idx) {
378 cl_read_ahead_release(env, &ra);
380 rc = cl_io_read_ahead(env, io, page_idx, &ra);
384 LASSERTF(ra.cra_end >= page_idx,
385 "object: %p, indcies %lu / %lu\n",
386 io->ci_obj, ra.cra_end, page_idx);
389 /* If the page is inside the read-ahead window*/
390 rc = ll_read_ahead_page(env, io, queue, page_idx);
395 } else if (stride_ria) {
396 /* If it is not in the read-ahead window, and it is
397 * read-ahead mode, then check whether it should skip
400 /* FIXME: This assertion only is valid when it is for
401 * forward read-ahead, it will be fixed when backward
402 * read-ahead is implemented */
403 LASSERTF(page_idx >= ria->ria_stoff,
404 "Invalid page_idx %lu rs %lu re %lu ro %lu "
405 "rl %lu rp %lu\n", page_idx,
406 ria->ria_start, ria->ria_end, ria->ria_stoff,
407 ria->ria_length, ria->ria_pages);
408 offset = page_idx - ria->ria_stoff;
409 offset = offset % (ria->ria_length);
410 if (offset > ria->ria_pages) {
411 page_idx += ria->ria_length - offset;
412 CDEBUG(D_READA, "i %lu skip %lu \n", page_idx,
413 ria->ria_length - offset);
419 cl_read_ahead_release(env, &ra);
425 static int ll_readahead(const struct lu_env *env, struct cl_io *io,
426 struct cl_page_list *queue,
427 struct ll_readahead_state *ras, bool hit)
429 struct vvp_io *vio = vvp_env_io(env);
430 struct ll_thread_info *lti = ll_env_info(env);
431 struct cl_attr *attr = vvp_env_thread_attr(env);
432 unsigned long len, mlen = 0, reserved;
433 pgoff_t ra_end, start = 0, end = 0;
435 struct ra_io_arg *ria = <i->lti_ria;
436 struct cl_object *clob;
442 inode = vvp_object_inode(clob);
444 memset(ria, 0, sizeof *ria);
446 cl_object_attr_lock(clob);
447 ret = cl_object_attr_get(env, clob, attr);
448 cl_object_attr_unlock(clob);
454 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
458 spin_lock(&ras->ras_lock);
461 * Note: other thread might rollback the ras_next_readahead,
462 * if it can not get the full size of prepared pages, see the
463 * end of this function. For stride read ahead, it needs to
464 * make sure the offset is no less than ras_stride_offset,
465 * so that stride read ahead can work correctly.
467 if (stride_io_mode(ras))
468 start = max(ras->ras_next_readahead, ras->ras_stride_offset);
470 start = ras->ras_next_readahead;
472 if (ras->ras_window_len > 0)
473 end = ras->ras_window_start + ras->ras_window_len - 1;
475 /* Enlarge the RA window to encompass the full read */
476 if (vio->vui_ra_valid &&
477 end < vio->vui_ra_start + vio->vui_ra_count - 1)
478 end = vio->vui_ra_start + vio->vui_ra_count - 1;
481 unsigned long rpc_boundary;
483 * Align RA window to an optimal boundary.
485 * XXX This would be better to align to cl_max_pages_per_rpc
486 * instead of PTLRPC_MAX_BRW_PAGES, because the RPC size may
487 * be aligned to the RAID stripe size in the future and that
488 * is more important than the RPC size.
490 /* Note: we only trim the RPC, instead of extending the RPC
491 * to the boundary, so to avoid reading too much pages during
493 rpc_boundary = ((end + 1) & (~(PTLRPC_MAX_BRW_PAGES - 1)));
494 if (rpc_boundary > 0)
497 if (rpc_boundary > start)
500 /* Truncate RA window to end of file */
501 end = min(end, (unsigned long)((kms - 1) >> PAGE_CACHE_SHIFT));
503 ras->ras_next_readahead = max(end, end + 1);
506 ria->ria_start = start;
508 /* If stride I/O mode is detected, get stride window*/
509 if (stride_io_mode(ras)) {
510 ria->ria_stoff = ras->ras_stride_offset;
511 ria->ria_length = ras->ras_stride_length;
512 ria->ria_pages = ras->ras_stride_pages;
514 spin_unlock(&ras->ras_lock);
517 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
520 len = ria_page_count(ria);
522 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
526 CDEBUG(D_READA, DFID": ria: %lu/%lu, bead: %lu/%lu, hit: %d\n",
527 PFID(lu_object_fid(&clob->co_lu)),
528 ria->ria_start, ria->ria_end,
529 vio->vui_ra_valid ? vio->vui_ra_start : 0,
530 vio->vui_ra_valid ? vio->vui_ra_count : 0,
533 /* at least to extend the readahead window to cover current read */
534 if (!hit && vio->vui_ra_valid &&
535 vio->vui_ra_start + vio->vui_ra_count > ria->ria_start) {
536 /* to the end of current read window. */
537 mlen = vio->vui_ra_start + vio->vui_ra_count - ria->ria_start;
538 /* trim to RPC boundary */
539 start = ria->ria_start & (PTLRPC_MAX_BRW_PAGES - 1);
540 mlen = min(mlen, PTLRPC_MAX_BRW_PAGES - start);
543 reserved = ll_ra_count_get(ll_i2sbi(inode), ria, len, mlen);
545 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
547 CDEBUG(D_READA, "reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
549 atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
550 ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
552 ret = ll_read_ahead_pages(env, io, queue, ria, &reserved, &ra_end);
555 ll_ra_count_put(ll_i2sbi(inode), reserved);
557 if (ra_end == end + 1 && ra_end == (kms >> PAGE_CACHE_SHIFT))
558 ll_ra_stats_inc(inode, RA_STAT_EOF);
560 /* if we didn't get to the end of the region we reserved from
561 * the ras we need to go back and update the ras so that the
562 * next read-ahead tries from where we left off. we only do so
563 * if the region we failed to issue read-ahead on is still ahead
564 * of the app and behind the next index to start read-ahead from */
565 CDEBUG(D_READA, "ra_end = %lu end = %lu stride end = %lu pages = %d\n",
566 ra_end, end, ria->ria_end, ret);
568 if (ra_end != end + 1) {
569 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
570 spin_lock(&ras->ras_lock);
571 if (ra_end < ras->ras_next_readahead &&
572 index_in_window(ra_end, ras->ras_window_start, 0,
573 ras->ras_window_len)) {
574 ras->ras_next_readahead = ra_end;
577 spin_unlock(&ras->ras_lock);
583 static void ras_set_start(struct inode *inode, struct ll_readahead_state *ras,
586 ras->ras_window_start = index & (~(RAS_INCREASE_STEP(inode) - 1));
589 /* called with the ras_lock held or from places where it doesn't matter */
590 static void ras_reset(struct inode *inode, struct ll_readahead_state *ras,
593 ras->ras_last_readpage = index;
594 ras->ras_consecutive_requests = 0;
595 ras->ras_consecutive_pages = 0;
596 ras->ras_window_len = 0;
597 ras_set_start(inode, ras, index);
598 ras->ras_next_readahead = max(ras->ras_window_start, index + 1);
603 /* called with the ras_lock held or from places where it doesn't matter */
604 static void ras_stride_reset(struct ll_readahead_state *ras)
606 ras->ras_consecutive_stride_requests = 0;
607 ras->ras_stride_length = 0;
608 ras->ras_stride_pages = 0;
612 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
614 spin_lock_init(&ras->ras_lock);
615 ras_reset(inode, ras, 0);
616 ras->ras_requests = 0;
620 * Check whether the read request is in the stride window.
621 * If it is in the stride window, return 1, otherwise return 0.
623 static int index_in_stride_window(struct ll_readahead_state *ras,
626 unsigned long stride_gap;
628 if (ras->ras_stride_length == 0 || ras->ras_stride_pages == 0 ||
629 ras->ras_stride_pages == ras->ras_stride_length)
632 stride_gap = index - ras->ras_last_readpage - 1;
634 /* If it is contiguous read */
636 return ras->ras_consecutive_pages + 1 <= ras->ras_stride_pages;
638 /* Otherwise check the stride by itself */
639 return (ras->ras_stride_length - ras->ras_stride_pages) == stride_gap &&
640 ras->ras_consecutive_pages == ras->ras_stride_pages;
643 static void ras_update_stride_detector(struct ll_readahead_state *ras,
646 unsigned long stride_gap = index - ras->ras_last_readpage - 1;
648 if (!stride_io_mode(ras) && (stride_gap != 0 ||
649 ras->ras_consecutive_stride_requests == 0)) {
650 ras->ras_stride_pages = ras->ras_consecutive_pages;
651 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
653 LASSERT(ras->ras_request_index == 0);
654 LASSERT(ras->ras_consecutive_stride_requests == 0);
656 if (index <= ras->ras_last_readpage) {
657 /*Reset stride window for forward read*/
658 ras_stride_reset(ras);
662 ras->ras_stride_pages = ras->ras_consecutive_pages;
663 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
670 stride_page_count(struct ll_readahead_state *ras, unsigned long len)
672 return stride_pg_count(ras->ras_stride_offset, ras->ras_stride_length,
673 ras->ras_stride_pages, ras->ras_stride_offset,
677 /* Stride Read-ahead window will be increased inc_len according to
678 * stride I/O pattern */
679 static void ras_stride_increase_window(struct ll_readahead_state *ras,
680 struct ll_ra_info *ra,
681 unsigned long inc_len)
683 unsigned long left, step, window_len;
684 unsigned long stride_len;
686 LASSERT(ras->ras_stride_length > 0);
687 LASSERTF(ras->ras_window_start + ras->ras_window_len
688 >= ras->ras_stride_offset, "window_start %lu, window_len %lu"
689 " stride_offset %lu\n", ras->ras_window_start,
690 ras->ras_window_len, ras->ras_stride_offset);
692 stride_len = ras->ras_window_start + ras->ras_window_len -
693 ras->ras_stride_offset;
695 left = stride_len % ras->ras_stride_length;
696 window_len = ras->ras_window_len - left;
698 if (left < ras->ras_stride_pages)
701 left = ras->ras_stride_pages + inc_len;
703 LASSERT(ras->ras_stride_pages != 0);
705 step = left / ras->ras_stride_pages;
706 left %= ras->ras_stride_pages;
708 window_len += step * ras->ras_stride_length + left;
710 if (stride_page_count(ras, window_len) <= ra->ra_max_pages_per_file)
711 ras->ras_window_len = window_len;
716 static void ras_increase_window(struct inode *inode,
717 struct ll_readahead_state *ras,
718 struct ll_ra_info *ra)
720 /* The stretch of ra-window should be aligned with max rpc_size
721 * but current clio architecture does not support retrieve such
722 * information from lower layer. FIXME later
724 if (stride_io_mode(ras))
725 ras_stride_increase_window(ras, ra, RAS_INCREASE_STEP(inode));
727 ras->ras_window_len = min(ras->ras_window_len +
728 RAS_INCREASE_STEP(inode),
729 ra->ra_max_pages_per_file);
732 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
733 struct ll_readahead_state *ras, unsigned long index,
734 enum ras_update_flags flags)
736 struct ll_ra_info *ra = &sbi->ll_ra_info;
737 bool hit = flags & LL_RAS_HIT;
738 int zero = 0, stride_detect = 0, ra_miss = 0;
741 spin_lock(&ras->ras_lock);
743 ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
745 /* reset the read-ahead window in two cases. First when the app seeks
746 * or reads to some other part of the file. Secondly if we get a
747 * read-ahead miss that we think we've previously issued. This can
748 * be a symptom of there being so many read-ahead pages that the VM is
749 * reclaiming it before we get to it. */
750 if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
752 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
753 } else if (!hit && ras->ras_window_len &&
754 index < ras->ras_next_readahead &&
755 index_in_window(index, ras->ras_window_start, 0,
756 ras->ras_window_len)) {
758 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
761 /* On the second access to a file smaller than the tunable
762 * ra_max_read_ahead_whole_pages trigger RA on all pages in the
763 * file up to ra_max_pages_per_file. This is simply a best effort
764 * and only occurs once per open file. Normal RA behavior is reverted
765 * to for subsequent IO. The mmap case does not increment
766 * ras_requests and thus can never trigger this behavior. */
767 if (ras->ras_requests >= 2 && !ras->ras_request_index) {
770 kms_pages = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >>
773 CDEBUG(D_READA, "kmsp "LPU64" mwp %lu mp %lu\n", kms_pages,
774 ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages_per_file);
777 kms_pages <= ra->ra_max_read_ahead_whole_pages) {
778 ras->ras_window_start = 0;
779 ras->ras_next_readahead = index + 1;
780 ras->ras_window_len = min(ra->ra_max_pages_per_file,
781 ra->ra_max_read_ahead_whole_pages);
786 /* check whether it is in stride I/O mode*/
787 if (!index_in_stride_window(ras, index)) {
788 if (ras->ras_consecutive_stride_requests == 0 &&
789 ras->ras_request_index == 0) {
790 ras_update_stride_detector(ras, index);
791 ras->ras_consecutive_stride_requests++;
793 ras_stride_reset(ras);
795 ras_reset(inode, ras, index);
796 ras->ras_consecutive_pages++;
799 ras->ras_consecutive_pages = 0;
800 ras->ras_consecutive_requests = 0;
801 if (++ras->ras_consecutive_stride_requests > 1)
807 if (index_in_stride_window(ras, index) &&
808 stride_io_mode(ras)) {
809 /*If stride-RA hit cache miss, the stride dector
810 *will not be reset to avoid the overhead of
811 *redetecting read-ahead mode */
812 if (index != ras->ras_last_readpage + 1)
813 ras->ras_consecutive_pages = 0;
814 ras_reset(inode, ras, index);
817 /* Reset both stride window and normal RA
819 ras_reset(inode, ras, index);
820 ras->ras_consecutive_pages++;
821 ras_stride_reset(ras);
824 } else if (stride_io_mode(ras)) {
825 /* If this is contiguous read but in stride I/O mode
826 * currently, check whether stride step still is valid,
827 * if invalid, it will reset the stride ra window*/
828 if (!index_in_stride_window(ras, index)) {
829 /* Shrink stride read-ahead window to be zero */
830 ras_stride_reset(ras);
831 ras->ras_window_len = 0;
832 ras->ras_next_readahead = index;
836 ras->ras_consecutive_pages++;
837 ras->ras_last_readpage = index;
838 ras_set_start(inode, ras, index);
840 if (stride_io_mode(ras)) {
841 /* Since stride readahead is sentivite to the offset
842 * of read-ahead, so we use original offset here,
843 * instead of ras_window_start, which is RPC aligned */
844 ras->ras_next_readahead = max(index, ras->ras_next_readahead);
846 if (ras->ras_next_readahead < ras->ras_window_start)
847 ras->ras_next_readahead = ras->ras_window_start;
849 ras->ras_next_readahead = index + 1;
853 /* Trigger RA in the mmap case where ras_consecutive_requests
854 * is not incremented and thus can't be used to trigger RA */
855 if (ras->ras_consecutive_pages >= 4 && flags & LL_RAS_MMAP) {
856 ras_increase_window(inode, ras, ra);
857 /* reset consecutive pages so that the readahead window can
859 ras->ras_consecutive_pages = 0;
863 /* Initially reset the stride window offset to next_readahead*/
864 if (ras->ras_consecutive_stride_requests == 2 && stride_detect) {
866 * Once stride IO mode is detected, next_readahead should be
867 * reset to make sure next_readahead > stride offset
869 ras->ras_next_readahead = max(index, ras->ras_next_readahead);
870 ras->ras_stride_offset = index;
871 ras->ras_window_len = RAS_INCREASE_STEP(inode);
874 /* The initial ras_window_len is set to the request size. To avoid
875 * uselessly reading and discarding pages for random IO the window is
876 * only increased once per consecutive request received. */
877 if ((ras->ras_consecutive_requests > 1 || stride_detect) &&
878 !ras->ras_request_index)
879 ras_increase_window(inode, ras, ra);
883 ras->ras_request_index++;
884 spin_unlock(&ras->ras_lock);
888 int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
890 struct inode *inode = vmpage->mapping->host;
891 struct ll_inode_info *lli = ll_i2info(inode);
894 struct cl_page *page;
895 struct cl_object *clob;
896 struct cl_env_nest nest;
897 bool redirtied = false;
898 bool unlocked = false;
902 LASSERT(PageLocked(vmpage));
903 LASSERT(!PageWriteback(vmpage));
905 LASSERT(ll_i2dtexp(inode) != NULL);
907 env = cl_env_nested_get(&nest);
909 GOTO(out, result = PTR_ERR(env));
911 clob = ll_i2info(inode)->lli_clob;
912 LASSERT(clob != NULL);
914 io = vvp_env_thread_io(env);
916 io->ci_ignore_layout = 1;
917 result = cl_io_init(env, io, CIT_MISC, clob);
919 page = cl_page_find(env, clob, vmpage->index,
920 vmpage, CPT_CACHEABLE);
922 lu_ref_add(&page->cp_reference, "writepage",
924 cl_page_assume(env, io, page);
925 result = cl_page_flush(env, io, page);
928 * Re-dirty page on error so it retries write,
929 * but not in case when IO has actually
930 * occurred and completed with an error.
932 if (!PageError(vmpage)) {
933 redirty_page_for_writepage(wbc, vmpage);
938 cl_page_disown(env, io, page);
940 lu_ref_del(&page->cp_reference,
941 "writepage", current);
942 cl_page_put(env, page);
944 result = PTR_ERR(page);
949 if (redirtied && wbc->sync_mode == WB_SYNC_ALL) {
950 loff_t offset = cl_offset(clob, vmpage->index);
952 /* Flush page failed because the extent is being written out.
953 * Wait for the write of extent to be finished to avoid
954 * breaking kernel which assumes ->writepage should mark
955 * PageWriteback or clean the page. */
956 result = cl_sync_file_range(inode, offset,
957 offset + PAGE_CACHE_SIZE - 1,
960 /* actually we may have written more than one page.
961 * decreasing this page because the caller will count
963 wbc->nr_to_write -= result - 1;
968 cl_env_nested_put(&nest, env);
973 if (!lli->lli_async_rc)
974 lli->lli_async_rc = result;
975 SetPageError(vmpage);
982 int ll_writepages(struct address_space *mapping, struct writeback_control *wbc)
984 struct inode *inode = mapping->host;
985 struct ll_sb_info *sbi = ll_i2sbi(inode);
988 enum cl_fsync_mode mode;
991 int ignore_layout = 0;
994 if (wbc->range_cyclic) {
995 start = mapping->writeback_index << PAGE_CACHE_SHIFT;
996 end = OBD_OBJECT_EOF;
998 start = wbc->range_start;
999 end = wbc->range_end;
1000 if (end == LLONG_MAX) {
1001 end = OBD_OBJECT_EOF;
1002 range_whole = start == 0;
1006 mode = CL_FSYNC_NONE;
1007 if (wbc->sync_mode == WB_SYNC_ALL)
1008 mode = CL_FSYNC_LOCAL;
1010 if (sbi->ll_umounting)
1011 /* if the mountpoint is being umounted, all pages have to be
1012 * evicted to avoid hitting LBUG when truncate_inode_pages()
1013 * is called later on. */
1016 if (ll_i2info(inode)->lli_clob == NULL)
1019 result = cl_sync_file_range(inode, start, end, mode, ignore_layout);
1021 wbc->nr_to_write -= result;
1025 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) {
1026 if (end == OBD_OBJECT_EOF)
1027 mapping->writeback_index = 0;
1029 mapping->writeback_index = (end >> PAGE_CACHE_SHIFT) +1;
1034 struct ll_cl_context *ll_cl_find(struct file *file)
1036 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1037 struct ll_cl_context *lcc;
1038 struct ll_cl_context *found = NULL;
1040 read_lock(&fd->fd_lock);
1041 list_for_each_entry(lcc, &fd->fd_lccs, lcc_list) {
1042 if (lcc->lcc_cookie == current) {
1047 read_unlock(&fd->fd_lock);
1052 void ll_cl_add(struct file *file, const struct lu_env *env, struct cl_io *io)
1054 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1055 struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1057 memset(lcc, 0, sizeof(*lcc));
1058 INIT_LIST_HEAD(&lcc->lcc_list);
1059 lcc->lcc_cookie = current;
1063 write_lock(&fd->fd_lock);
1064 list_add(&lcc->lcc_list, &fd->fd_lccs);
1065 write_unlock(&fd->fd_lock);
1068 void ll_cl_remove(struct file *file, const struct lu_env *env)
1070 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1071 struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1073 write_lock(&fd->fd_lock);
1074 list_del_init(&lcc->lcc_list);
1075 write_unlock(&fd->fd_lock);
1078 static int ll_io_read_page(const struct lu_env *env, struct cl_io *io,
1079 struct cl_page *page)
1081 struct inode *inode = vvp_object_inode(page->cp_obj);
1082 struct ll_sb_info *sbi = ll_i2sbi(inode);
1083 struct ll_file_data *fd = vvp_env_io(env)->vui_fd;
1084 struct ll_readahead_state *ras = &fd->fd_ras;
1085 struct cl_2queue *queue = &io->ci_queue;
1086 struct vvp_page *vpg;
1090 vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1091 if (sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
1092 sbi->ll_ra_info.ra_max_pages > 0) {
1093 struct vvp_io *vio = vvp_env_io(env);
1094 enum ras_update_flags flags = 0;
1096 if (vpg->vpg_defer_uptodate)
1097 flags |= LL_RAS_HIT;
1098 if (!vio->vui_ra_valid)
1099 flags |= LL_RAS_MMAP;
1100 ras_update(sbi, inode, ras, vvp_index(vpg), flags);
1103 if (vpg->vpg_defer_uptodate) {
1104 vpg->vpg_ra_used = 1;
1105 cl_page_export(env, page, 1);
1108 cl_2queue_init(queue);
1110 * Add page into the queue even when it is marked uptodate above.
1111 * this will unlock it automatically as part of cl_page_list_disown().
1113 cl_2queue_add(queue, page);
1114 if (sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
1115 sbi->ll_ra_info.ra_max_pages > 0) {
1118 rc2 = ll_readahead(env, io, &queue->c2_qin, ras,
1119 vpg->vpg_defer_uptodate);
1120 CDEBUG(D_READA, DFID "%d pages read ahead at %lu\n",
1121 PFID(ll_inode2fid(inode)), rc2, vvp_index(vpg));
1124 if (queue->c2_qin.pl_nr > 0)
1125 rc = cl_io_submit_rw(env, io, CRT_READ, queue);
1128 * Unlock unsent pages in case of error.
1130 cl_page_list_disown(env, io, &queue->c2_qin);
1131 cl_2queue_fini(env, queue);
1136 int ll_readpage(struct file *file, struct page *vmpage)
1138 struct inode *inode = file->f_path.dentry->d_inode;
1139 struct cl_object *clob = ll_i2info(inode)->lli_clob;
1140 struct ll_cl_context *lcc;
1141 const struct lu_env *env;
1143 struct cl_page *page;
1147 lcc = ll_cl_find(file);
1149 unlock_page(vmpage);
1155 LASSERT(io != NULL);
1156 LASSERT(io->ci_state == CIS_IO_GOING);
1157 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
1158 if (!IS_ERR(page)) {
1159 LASSERT(page->cp_type == CPT_CACHEABLE);
1160 if (likely(!PageUptodate(vmpage))) {
1161 cl_page_assume(env, io, page);
1162 result = ll_io_read_page(env, io, page);
1164 /* Page from a non-object file. */
1165 unlock_page(vmpage);
1168 cl_page_put(env, page);
1170 unlock_page(vmpage);
1171 result = PTR_ERR(page);
1176 int ll_page_sync_io(const struct lu_env *env, struct cl_io *io,
1177 struct cl_page *page, enum cl_req_type crt)
1179 struct cl_2queue *queue;
1182 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
1184 queue = &io->ci_queue;
1185 cl_2queue_init_page(queue, page);
1187 result = cl_io_submit_sync(env, io, crt, queue, 0);
1188 LASSERT(cl_page_is_owned(page, io));
1190 if (crt == CRT_READ)
1192 * in CRT_WRITE case page is left locked even in case of
1195 cl_page_list_disown(env, io, &queue->c2_qin);
1196 cl_2queue_fini(env, queue);