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 <lustre_lite.h>
61 #include <obd_cksum.h>
62 #include "llite_internal.h"
63 #include <linux/lustre_compat25.h>
66 * Finalizes cl-data before exiting typical address_space operation. Dual to
69 void ll_cl_fini(struct ll_cl_context *lcc)
71 struct lu_env *env = lcc->lcc_env;
72 struct cl_io *io = lcc->lcc_io;
73 struct cl_page *page = lcc->lcc_page;
75 LASSERT(lcc->lcc_cookie == current);
79 lu_ref_del(&page->cp_reference, "cl_io", io);
80 cl_page_put(env, page);
83 cl_env_put(env, &lcc->lcc_refcheck);
87 * Initializes common cl-data at the typical address_space operation entry
90 struct ll_cl_context *ll_cl_init(struct file *file, struct page *vmpage)
92 struct ll_cl_context *lcc;
95 struct cl_object *clob;
101 clob = ll_i2info(file->f_dentry->d_inode)->lli_clob;
102 LASSERT(clob != NULL);
104 env = cl_env_get(&refcheck);
106 return ERR_PTR(PTR_ERR(env));
108 lcc = &vvp_env_info(env)->vti_io_ctx;
109 memset(lcc, 0, sizeof(*lcc));
111 lcc->lcc_refcheck = refcheck;
112 lcc->lcc_cookie = current;
114 cio = ccc_env_io(env);
115 io = cio->cui_cl.cis_io;
118 struct inode *inode = file->f_dentry->d_inode;
120 CERROR("%s: " DFID " no active IO, please file a ticket.\n",
121 ll_get_fsname(inode->i_sb, NULL, 0),
122 PFID(ll_inode2fid(inode)));
123 libcfs_debug_dumpstack(NULL);
126 if (result == 0 && vmpage != NULL) {
127 struct cl_page *page;
130 LASSERT(io->ci_state == CIS_IO_GOING);
131 LASSERT(cio->cui_fd == LUSTRE_FPRIVATE(file));
132 page = cl_page_find(env, clob, vmpage->index, vmpage,
135 lcc->lcc_page = page;
136 lu_ref_add(&page->cp_reference, "cl_io", io);
139 result = PTR_ERR(page);
143 lcc = ERR_PTR(result);
149 struct obd_capa *cl_capa_lookup(struct inode *inode, enum cl_req_type crt)
153 opc = crt == CRT_WRITE ? CAPA_OPC_OSS_WRITE : CAPA_OPC_OSS_RW;
154 return ll_osscapa_get(inode, opc);
157 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which);
160 * Get readahead pages from the filesystem readahead pool of the client for a
163 * /param sbi superblock for filesystem readahead state ll_ra_info
164 * /param ria per-thread readahead state
165 * /param pages number of pages requested for readahead for the thread.
167 * WARNING: This algorithm is used to reduce contention on sbi->ll_lock.
168 * It should work well if the ra_max_pages is much greater than the single
169 * file's read-ahead window, and not too many threads contending for
170 * these readahead pages.
172 * TODO: There may be a 'global sync problem' if many threads are trying
173 * to get an ra budget that is larger than the remaining readahead pages
174 * and reach here at exactly the same time. They will compute /a ret to
175 * consume the remaining pages, but will fail at atomic_add_return() and
176 * get a zero ra window, although there is still ra space remaining. - Jay */
178 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi,
179 struct ra_io_arg *ria,
180 unsigned long pages, unsigned long min)
182 struct ll_ra_info *ra = &sbi->ll_ra_info;
186 /* If read-ahead pages left are less than 1M, do not do read-ahead,
187 * otherwise it will form small read RPC(< 1M), which hurt server
188 * performance a lot. */
189 ret = min(ra->ra_max_pages - atomic_read(&ra->ra_cur_pages), pages);
190 if (ret < 0 || ret < min_t(long, PTLRPC_MAX_BRW_PAGES, pages))
193 /* If the non-strided (ria_pages == 0) readahead window
194 * (ria_start + ret) has grown across an RPC boundary, then trim
195 * readahead size by the amount beyond the RPC so it ends on an
196 * RPC boundary. If the readahead window is already ending on
197 * an RPC boundary (beyond_rpc == 0), or smaller than a full
198 * RPC (beyond_rpc < ret) the readahead size is unchanged.
199 * The (beyond_rpc != 0) check is skipped since the conditional
200 * branch is more expensive than subtracting zero from the result.
202 * Strided read is left unaligned to avoid small fragments beyond
203 * the RPC boundary from needing an extra read RPC. */
204 if (ria->ria_pages == 0) {
205 long beyond_rpc = (ria->ria_start + ret) % PTLRPC_MAX_BRW_PAGES;
206 if (/* beyond_rpc != 0 && */ beyond_rpc < ret)
210 if (atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) {
211 atomic_sub(ret, &ra->ra_cur_pages);
217 /* override ra limit for maximum performance */
218 atomic_add(min - ret, &ra->ra_cur_pages);
224 void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
226 struct ll_ra_info *ra = &sbi->ll_ra_info;
227 atomic_sub(len, &ra->ra_cur_pages);
230 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which)
232 LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
233 lprocfs_counter_incr(sbi->ll_ra_stats, which);
236 void ll_ra_stats_inc(struct inode *inode, enum ra_stat which)
238 struct ll_sb_info *sbi = ll_i2sbi(inode);
239 ll_ra_stats_inc_sbi(sbi, which);
242 #define RAS_CDEBUG(ras) \
244 "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu r %lu ri %lu" \
245 "csr %lu sf %lu sp %lu sl %lu \n", \
246 ras->ras_last_readpage, ras->ras_consecutive_requests, \
247 ras->ras_consecutive_pages, ras->ras_window_start, \
248 ras->ras_window_len, ras->ras_next_readahead, \
249 ras->ras_requests, ras->ras_request_index, \
250 ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
251 ras->ras_stride_pages, ras->ras_stride_length)
253 static int index_in_window(unsigned long index, unsigned long point,
254 unsigned long before, unsigned long after)
256 unsigned long start = point - before, end = point + after;
263 return start <= index && index <= end;
266 static struct ll_readahead_state *ll_ras_get(struct file *f)
268 struct ll_file_data *fd;
270 fd = LUSTRE_FPRIVATE(f);
274 void ll_ra_read_in(struct file *f, struct ll_ra_read *rar)
276 struct ll_readahead_state *ras;
280 spin_lock(&ras->ras_lock);
282 ras->ras_request_index = 0;
283 ras->ras_consecutive_requests++;
284 rar->lrr_reader = current;
286 cfs_list_add(&rar->lrr_linkage, &ras->ras_read_beads);
287 spin_unlock(&ras->ras_lock);
290 void ll_ra_read_ex(struct file *f, struct ll_ra_read *rar)
292 struct ll_readahead_state *ras;
296 spin_lock(&ras->ras_lock);
297 cfs_list_del_init(&rar->lrr_linkage);
298 spin_unlock(&ras->ras_lock);
301 static struct ll_ra_read *ll_ra_read_get_locked(struct ll_readahead_state *ras)
303 struct ll_ra_read *scan;
305 cfs_list_for_each_entry(scan, &ras->ras_read_beads, lrr_linkage) {
306 if (scan->lrr_reader == current)
312 struct ll_ra_read *ll_ra_read_get(struct file *f)
314 struct ll_readahead_state *ras;
315 struct ll_ra_read *bead;
319 spin_lock(&ras->ras_lock);
320 bead = ll_ra_read_get_locked(ras);
321 spin_unlock(&ras->ras_lock);
325 static int cl_read_ahead_page(const struct lu_env *env, struct cl_io *io,
326 struct cl_page_list *queue, struct cl_page *page,
327 struct cl_object *clob, pgoff_t *max_index)
329 struct page *vmpage = page->cp_vmpage;
336 cl_page_assume(env, io, page);
337 lu_ref_add(&page->cp_reference, "ra", current);
338 cp = cl2ccc_page(cl_object_page_slice(clob, page));
339 if (!cp->cpg_defer_uptodate && !PageUptodate(vmpage)) {
340 CDEBUG(D_READA, "page index %lu, max_index: %lu\n",
341 ccc_index(cp), *max_index);
342 if (*max_index == 0 || ccc_index(cp) > *max_index)
343 rc = cl_page_is_under_lock(env, io, page, max_index);
345 cp->cpg_defer_uptodate = 1;
347 cl_page_list_add(queue, page);
350 cl_page_discard(env, io, page);
354 /* skip completed pages */
355 cl_page_unassume(env, io, page);
357 lu_ref_del(&page->cp_reference, "ra", current);
358 cl_page_put(env, page);
363 * Initiates read-ahead of a page with given index.
365 * \retval +ve: page was added to \a queue.
367 * \retval -ENOLCK: there is no extent lock for this part of a file, stop
370 * \retval -ve, 0: page wasn't added to \a queue for other reason.
372 static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io,
373 struct cl_page_list *queue,
374 pgoff_t index, pgoff_t *max_index)
376 struct cl_object *clob = io->ci_obj;
377 struct inode *inode = ccc_object_inode(clob);
379 struct cl_page *page;
380 enum ra_stat which = _NR_RA_STAT; /* keep gcc happy */
381 unsigned int gfp_mask;
383 const char *msg = NULL;
387 gfp_mask = GFP_HIGHUSER & ~__GFP_WAIT;
389 gfp_mask |= __GFP_NOWARN;
391 vmpage = grab_cache_page_nowait(inode->i_mapping, index);
392 if (vmpage != NULL) {
393 /* Check if vmpage was truncated or reclaimed */
394 if (vmpage->mapping == inode->i_mapping) {
395 page = cl_page_find(env, clob, vmpage->index,
396 vmpage, CPT_CACHEABLE);
398 rc = cl_read_ahead_page(env, io, queue,
399 page, clob, max_index);
401 which = RA_STAT_FAILED_MATCH;
402 msg = "lock match failed";
405 which = RA_STAT_FAILED_GRAB_PAGE;
406 msg = "cl_page_find failed";
409 which = RA_STAT_WRONG_GRAB_PAGE;
410 msg = "g_c_p_n returned invalid page";
414 page_cache_release(vmpage);
416 which = RA_STAT_FAILED_GRAB_PAGE;
417 msg = "g_c_p_n failed";
420 ll_ra_stats_inc(inode, which);
421 CDEBUG(D_READA, "%s\n", msg);
426 #define RIA_DEBUG(ria) \
427 CDEBUG(D_READA, "rs %lu re %lu ro %lu rl %lu rp %lu\n", \
428 ria->ria_start, ria->ria_end, ria->ria_stoff, ria->ria_length,\
431 /* Limit this to the blocksize instead of PTLRPC_BRW_MAX_SIZE, since we don't
432 * know what the actual RPC size is. If this needs to change, it makes more
433 * sense to tune the i_blkbits value for the file based on the OSTs it is
434 * striped over, rather than having a constant value for all files here. */
436 /* RAS_INCREASE_STEP should be (1UL << (inode->i_blkbits - PAGE_CACHE_SHIFT)).
437 * Temprarily set RAS_INCREASE_STEP to 1MB. After 4MB RPC is enabled
438 * by default, this should be adjusted corresponding with max_read_ahead_mb
439 * and max_read_ahead_per_file_mb otherwise the readahead budget can be used
440 * up quickly which will affect read performance siginificantly. See LU-2816 */
441 #define RAS_INCREASE_STEP(inode) (ONE_MB_BRW_SIZE >> PAGE_CACHE_SHIFT)
443 static inline int stride_io_mode(struct ll_readahead_state *ras)
445 return ras->ras_consecutive_stride_requests > 1;
447 /* The function calculates how much pages will be read in
448 * [off, off + length], in such stride IO area,
449 * stride_offset = st_off, stride_lengh = st_len,
450 * stride_pages = st_pgs
452 * |------------------|*****|------------------|*****|------------|*****|....
455 * |----- st_len -----|
457 * How many pages it should read in such pattern
458 * |-------------------------------------------------------------|
460 * |<------ length ------->|
462 * = |<----->| + |-------------------------------------| + |---|
463 * start_left st_pgs * i end_left
466 stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs,
467 unsigned long off, unsigned long length)
469 __u64 start = off > st_off ? off - st_off : 0;
470 __u64 end = off + length > st_off ? off + length - st_off : 0;
471 unsigned long start_left = 0;
472 unsigned long end_left = 0;
473 unsigned long pg_count;
475 if (st_len == 0 || length == 0 || end == 0)
478 start_left = do_div(start, st_len);
479 if (start_left < st_pgs)
480 start_left = st_pgs - start_left;
484 end_left = do_div(end, st_len);
485 if (end_left > st_pgs)
488 CDEBUG(D_READA, "start "LPU64", end "LPU64" start_left %lu end_left %lu \n",
489 start, end, start_left, end_left);
492 pg_count = end_left - (st_pgs - start_left);
494 pg_count = start_left + st_pgs * (end - start - 1) + end_left;
496 CDEBUG(D_READA, "st_off %lu, st_len %lu st_pgs %lu off %lu length %lu"
497 "pgcount %lu\n", st_off, st_len, st_pgs, off, length, pg_count);
502 static int ria_page_count(struct ra_io_arg *ria)
504 __u64 length = ria->ria_end >= ria->ria_start ?
505 ria->ria_end - ria->ria_start + 1 : 0;
507 return stride_pg_count(ria->ria_stoff, ria->ria_length,
508 ria->ria_pages, ria->ria_start,
512 /*Check whether the index is in the defined ra-window */
513 static int ras_inside_ra_window(unsigned long idx, struct ra_io_arg *ria)
515 /* If ria_length == ria_pages, it means non-stride I/O mode,
516 * idx should always inside read-ahead window in this case
517 * For stride I/O mode, just check whether the idx is inside
519 return ria->ria_length == 0 || ria->ria_length == ria->ria_pages ||
520 (idx >= ria->ria_stoff && (idx - ria->ria_stoff) %
521 ria->ria_length < ria->ria_pages);
524 static int ll_read_ahead_pages(const struct lu_env *env,
525 struct cl_io *io, struct cl_page_list *queue,
526 struct ra_io_arg *ria,
527 unsigned long *reserved_pages,
528 unsigned long *ra_end)
533 pgoff_t max_index = 0;
535 LASSERT(ria != NULL);
538 stride_ria = ria->ria_length > ria->ria_pages && ria->ria_pages > 0;
539 for (page_idx = ria->ria_start;
540 page_idx <= ria->ria_end && *reserved_pages > 0; page_idx++) {
541 if (ras_inside_ra_window(page_idx, ria)) {
542 /* If the page is inside the read-ahead window*/
543 rc = ll_read_ahead_page(env, io, queue,
544 page_idx, &max_index);
548 } else if (rc == -ENOLCK)
550 } else if (stride_ria) {
551 /* If it is not in the read-ahead window, and it is
552 * read-ahead mode, then check whether it should skip
555 /* FIXME: This assertion only is valid when it is for
556 * forward read-ahead, it will be fixed when backward
557 * read-ahead is implemented */
558 LASSERTF(page_idx > ria->ria_stoff, "Invalid page_idx %lu"
559 "rs %lu re %lu ro %lu rl %lu rp %lu\n", page_idx,
560 ria->ria_start, ria->ria_end, ria->ria_stoff,
561 ria->ria_length, ria->ria_pages);
562 offset = page_idx - ria->ria_stoff;
563 offset = offset % (ria->ria_length);
564 if (offset > ria->ria_pages) {
565 page_idx += ria->ria_length - offset;
566 CDEBUG(D_READA, "i %lu skip %lu \n", page_idx,
567 ria->ria_length - offset);
576 int ll_readahead(const struct lu_env *env, struct cl_io *io,
577 struct cl_page_list *queue, struct ll_readahead_state *ras,
580 struct vvp_io *vio = vvp_env_io(env);
581 struct vvp_thread_info *vti = vvp_env_info(env);
582 struct cl_attr *attr = ccc_env_thread_attr(env);
583 unsigned long start = 0, end = 0, reserved;
584 unsigned long ra_end, len, mlen = 0;
586 struct ll_ra_read *bead;
587 struct ra_io_arg *ria = &vti->vti_ria;
588 struct cl_object *clob;
594 inode = ccc_object_inode(clob);
596 memset(ria, 0, sizeof *ria);
598 cl_object_attr_lock(clob);
599 ret = cl_object_attr_get(env, clob, attr);
600 cl_object_attr_unlock(clob);
606 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
610 spin_lock(&ras->ras_lock);
611 if (vio->cui_ra_window_set)
612 bead = &vio->cui_bead;
616 /* Enlarge the RA window to encompass the full read */
617 if (bead != NULL && ras->ras_window_start + ras->ras_window_len <
618 bead->lrr_start + bead->lrr_count) {
619 ras->ras_window_len = bead->lrr_start + bead->lrr_count -
620 ras->ras_window_start;
622 /* Reserve a part of the read-ahead window that we'll be issuing */
623 if (ras->ras_window_len) {
624 start = ras->ras_next_readahead;
625 end = ras->ras_window_start + ras->ras_window_len - 1;
628 unsigned long rpc_boundary;
630 * Align RA window to an optimal boundary.
632 * XXX This would be better to align to cl_max_pages_per_rpc
633 * instead of PTLRPC_MAX_BRW_PAGES, because the RPC size may
634 * be aligned to the RAID stripe size in the future and that
635 * is more important than the RPC size.
637 /* Note: we only trim the RPC, instead of extending the RPC
638 * to the boundary, so to avoid reading too much pages during
640 rpc_boundary = ((end + 1) & (~(PTLRPC_MAX_BRW_PAGES - 1)));
641 if (rpc_boundary > 0)
644 if (rpc_boundary > start)
647 /* Truncate RA window to end of file */
648 end = min(end, (unsigned long)((kms - 1) >> PAGE_CACHE_SHIFT));
650 ras->ras_next_readahead = max(end, end + 1);
653 ria->ria_start = start;
655 /* If stride I/O mode is detected, get stride window*/
656 if (stride_io_mode(ras)) {
657 ria->ria_stoff = ras->ras_stride_offset;
658 ria->ria_length = ras->ras_stride_length;
659 ria->ria_pages = ras->ras_stride_pages;
661 spin_unlock(&ras->ras_lock);
664 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
667 len = ria_page_count(ria);
669 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
673 CDEBUG(D_READA, DFID": ria: %lu/%lu, bead: %lu/%lu, hit: %d\n",
674 PFID(lu_object_fid(&clob->co_lu)),
675 ria->ria_start, ria->ria_end,
676 bead == NULL ? 0 : bead->lrr_start,
677 bead == NULL ? 0 : bead->lrr_count,
680 /* at least to extend the readahead window to cover current read */
681 if (!hit && bead != NULL &&
682 bead->lrr_start + bead->lrr_count > ria->ria_start) {
683 /* to the end of current read window. */
684 mlen = bead->lrr_start + bead->lrr_count - ria->ria_start;
685 /* trim to RPC boundary */
686 start = ria->ria_start & (PTLRPC_MAX_BRW_PAGES - 1);
687 mlen = min(mlen, PTLRPC_MAX_BRW_PAGES - start);
690 reserved = ll_ra_count_get(ll_i2sbi(inode), ria, len, mlen);
692 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
694 CDEBUG(D_READA, "reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
696 atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
697 ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
699 ret = ll_read_ahead_pages(env, io, queue, ria, &reserved, &ra_end);
702 ll_ra_count_put(ll_i2sbi(inode), reserved);
704 if (ra_end == end + 1 && ra_end == (kms >> PAGE_CACHE_SHIFT))
705 ll_ra_stats_inc(inode, RA_STAT_EOF);
707 /* if we didn't get to the end of the region we reserved from
708 * the ras we need to go back and update the ras so that the
709 * next read-ahead tries from where we left off. we only do so
710 * if the region we failed to issue read-ahead on is still ahead
711 * of the app and behind the next index to start read-ahead from */
712 CDEBUG(D_READA, "ra_end %lu end %lu stride end %lu \n",
713 ra_end, end, ria->ria_end);
715 if (ra_end != end + 1) {
716 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
717 spin_lock(&ras->ras_lock);
718 if (ra_end < ras->ras_next_readahead &&
719 index_in_window(ra_end, ras->ras_window_start, 0,
720 ras->ras_window_len)) {
721 ras->ras_next_readahead = ra_end;
724 spin_unlock(&ras->ras_lock);
730 static void ras_set_start(struct inode *inode, struct ll_readahead_state *ras,
733 ras->ras_window_start = index & (~(RAS_INCREASE_STEP(inode) - 1));
736 /* called with the ras_lock held or from places where it doesn't matter */
737 static void ras_reset(struct inode *inode, struct ll_readahead_state *ras,
740 ras->ras_last_readpage = index;
741 ras->ras_consecutive_requests = 0;
742 ras->ras_consecutive_pages = 0;
743 ras->ras_window_len = 0;
744 ras_set_start(inode, ras, index);
745 ras->ras_next_readahead = max(ras->ras_window_start, index);
750 /* called with the ras_lock held or from places where it doesn't matter */
751 static void ras_stride_reset(struct ll_readahead_state *ras)
753 ras->ras_consecutive_stride_requests = 0;
754 ras->ras_stride_length = 0;
755 ras->ras_stride_pages = 0;
759 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
761 spin_lock_init(&ras->ras_lock);
762 ras_reset(inode, ras, 0);
763 ras->ras_requests = 0;
764 CFS_INIT_LIST_HEAD(&ras->ras_read_beads);
768 * Check whether the read request is in the stride window.
769 * If it is in the stride window, return 1, otherwise return 0.
771 static int index_in_stride_window(struct ll_readahead_state *ras,
774 unsigned long stride_gap;
776 if (ras->ras_stride_length == 0 || ras->ras_stride_pages == 0 ||
777 ras->ras_stride_pages == ras->ras_stride_length)
780 stride_gap = index - ras->ras_last_readpage - 1;
782 /* If it is contiguous read */
784 return ras->ras_consecutive_pages + 1 <= ras->ras_stride_pages;
786 /* Otherwise check the stride by itself */
787 return (ras->ras_stride_length - ras->ras_stride_pages) == stride_gap &&
788 ras->ras_consecutive_pages == ras->ras_stride_pages;
791 static void ras_update_stride_detector(struct ll_readahead_state *ras,
794 unsigned long stride_gap = index - ras->ras_last_readpage - 1;
796 if (!stride_io_mode(ras) && (stride_gap != 0 ||
797 ras->ras_consecutive_stride_requests == 0)) {
798 ras->ras_stride_pages = ras->ras_consecutive_pages;
799 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
801 LASSERT(ras->ras_request_index == 0);
802 LASSERT(ras->ras_consecutive_stride_requests == 0);
804 if (index <= ras->ras_last_readpage) {
805 /*Reset stride window for forward read*/
806 ras_stride_reset(ras);
810 ras->ras_stride_pages = ras->ras_consecutive_pages;
811 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
818 stride_page_count(struct ll_readahead_state *ras, unsigned long len)
820 return stride_pg_count(ras->ras_stride_offset, ras->ras_stride_length,
821 ras->ras_stride_pages, ras->ras_stride_offset,
825 /* Stride Read-ahead window will be increased inc_len according to
826 * stride I/O pattern */
827 static void ras_stride_increase_window(struct ll_readahead_state *ras,
828 struct ll_ra_info *ra,
829 unsigned long inc_len)
831 unsigned long left, step, window_len;
832 unsigned long stride_len;
834 LASSERT(ras->ras_stride_length > 0);
835 LASSERTF(ras->ras_window_start + ras->ras_window_len
836 >= ras->ras_stride_offset, "window_start %lu, window_len %lu"
837 " stride_offset %lu\n", ras->ras_window_start,
838 ras->ras_window_len, ras->ras_stride_offset);
840 stride_len = ras->ras_window_start + ras->ras_window_len -
841 ras->ras_stride_offset;
843 left = stride_len % ras->ras_stride_length;
844 window_len = ras->ras_window_len - left;
846 if (left < ras->ras_stride_pages)
849 left = ras->ras_stride_pages + inc_len;
851 LASSERT(ras->ras_stride_pages != 0);
853 step = left / ras->ras_stride_pages;
854 left %= ras->ras_stride_pages;
856 window_len += step * ras->ras_stride_length + left;
858 if (stride_page_count(ras, window_len) <= ra->ra_max_pages_per_file)
859 ras->ras_window_len = window_len;
864 static void ras_increase_window(struct inode *inode,
865 struct ll_readahead_state *ras,
866 struct ll_ra_info *ra)
868 /* The stretch of ra-window should be aligned with max rpc_size
869 * but current clio architecture does not support retrieve such
870 * information from lower layer. FIXME later
872 if (stride_io_mode(ras))
873 ras_stride_increase_window(ras, ra, RAS_INCREASE_STEP(inode));
875 ras->ras_window_len = min(ras->ras_window_len +
876 RAS_INCREASE_STEP(inode),
877 ra->ra_max_pages_per_file);
880 void ras_update(struct ll_sb_info *sbi, struct inode *inode,
881 struct ll_readahead_state *ras, unsigned long index,
884 struct ll_ra_info *ra = &sbi->ll_ra_info;
885 int zero = 0, stride_detect = 0, ra_miss = 0;
888 spin_lock(&ras->ras_lock);
890 ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
892 /* reset the read-ahead window in two cases. First when the app seeks
893 * or reads to some other part of the file. Secondly if we get a
894 * read-ahead miss that we think we've previously issued. This can
895 * be a symptom of there being so many read-ahead pages that the VM is
896 * reclaiming it before we get to it. */
897 if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
899 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
900 } else if (!hit && ras->ras_window_len &&
901 index < ras->ras_next_readahead &&
902 index_in_window(index, ras->ras_window_start, 0,
903 ras->ras_window_len)) {
905 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
908 /* On the second access to a file smaller than the tunable
909 * ra_max_read_ahead_whole_pages trigger RA on all pages in the
910 * file up to ra_max_pages_per_file. This is simply a best effort
911 * and only occurs once per open file. Normal RA behavior is reverted
912 * to for subsequent IO. The mmap case does not increment
913 * ras_requests and thus can never trigger this behavior. */
914 if (ras->ras_requests == 2 && !ras->ras_request_index) {
917 kms_pages = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >>
920 CDEBUG(D_READA, "kmsp "LPU64" mwp %lu mp %lu\n", kms_pages,
921 ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages_per_file);
924 kms_pages <= ra->ra_max_read_ahead_whole_pages) {
925 ras->ras_window_start = 0;
926 ras->ras_last_readpage = 0;
927 ras->ras_next_readahead = 0;
928 ras->ras_window_len = min(ra->ra_max_pages_per_file,
929 ra->ra_max_read_ahead_whole_pages);
934 /* check whether it is in stride I/O mode*/
935 if (!index_in_stride_window(ras, index)) {
936 if (ras->ras_consecutive_stride_requests == 0 &&
937 ras->ras_request_index == 0) {
938 ras_update_stride_detector(ras, index);
939 ras->ras_consecutive_stride_requests++;
941 ras_stride_reset(ras);
943 ras_reset(inode, ras, index);
944 ras->ras_consecutive_pages++;
947 ras->ras_consecutive_pages = 0;
948 ras->ras_consecutive_requests = 0;
949 if (++ras->ras_consecutive_stride_requests > 1)
955 if (index_in_stride_window(ras, index) &&
956 stride_io_mode(ras)) {
957 /*If stride-RA hit cache miss, the stride dector
958 *will not be reset to avoid the overhead of
959 *redetecting read-ahead mode */
960 if (index != ras->ras_last_readpage + 1)
961 ras->ras_consecutive_pages = 0;
962 ras_reset(inode, ras, index);
965 /* Reset both stride window and normal RA
967 ras_reset(inode, ras, index);
968 ras->ras_consecutive_pages++;
969 ras_stride_reset(ras);
972 } else if (stride_io_mode(ras)) {
973 /* If this is contiguous read but in stride I/O mode
974 * currently, check whether stride step still is valid,
975 * if invalid, it will reset the stride ra window*/
976 if (!index_in_stride_window(ras, index)) {
977 /* Shrink stride read-ahead window to be zero */
978 ras_stride_reset(ras);
979 ras->ras_window_len = 0;
980 ras->ras_next_readahead = index;
984 ras->ras_consecutive_pages++;
985 ras->ras_last_readpage = index;
986 ras_set_start(inode, ras, index);
988 if (stride_io_mode(ras)) {
989 /* Since stride readahead is sentivite to the offset
990 * of read-ahead, so we use original offset here,
991 * instead of ras_window_start, which is RPC aligned */
992 ras->ras_next_readahead = max(index, ras->ras_next_readahead);
994 if (ras->ras_next_readahead < ras->ras_window_start)
995 ras->ras_next_readahead = ras->ras_window_start;
997 ras->ras_next_readahead = index + 1;
1001 /* Trigger RA in the mmap case where ras_consecutive_requests
1002 * is not incremented and thus can't be used to trigger RA */
1003 if (!ras->ras_window_len && ras->ras_consecutive_pages == 4) {
1004 ras->ras_window_len = RAS_INCREASE_STEP(inode);
1005 GOTO(out_unlock, 0);
1008 /* Initially reset the stride window offset to next_readahead*/
1009 if (ras->ras_consecutive_stride_requests == 2 && stride_detect) {
1011 * Once stride IO mode is detected, next_readahead should be
1012 * reset to make sure next_readahead > stride offset
1014 ras->ras_next_readahead = max(index, ras->ras_next_readahead);
1015 ras->ras_stride_offset = index;
1016 ras->ras_window_len = RAS_INCREASE_STEP(inode);
1019 /* The initial ras_window_len is set to the request size. To avoid
1020 * uselessly reading and discarding pages for random IO the window is
1021 * only increased once per consecutive request received. */
1022 if ((ras->ras_consecutive_requests > 1 || stride_detect) &&
1023 !ras->ras_request_index)
1024 ras_increase_window(inode, ras, ra);
1028 ras->ras_request_index++;
1029 spin_unlock(&ras->ras_lock);
1033 int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
1035 struct inode *inode = vmpage->mapping->host;
1036 struct ll_inode_info *lli = ll_i2info(inode);
1039 struct cl_page *page;
1040 struct cl_object *clob;
1041 struct cl_env_nest nest;
1042 bool redirtied = false;
1043 bool unlocked = false;
1047 LASSERT(PageLocked(vmpage));
1048 LASSERT(!PageWriteback(vmpage));
1050 LASSERT(ll_i2dtexp(inode) != NULL);
1052 env = cl_env_nested_get(&nest);
1054 GOTO(out, result = PTR_ERR(env));
1056 clob = ll_i2info(inode)->lli_clob;
1057 LASSERT(clob != NULL);
1059 io = ccc_env_thread_io(env);
1061 io->ci_ignore_layout = 1;
1062 result = cl_io_init(env, io, CIT_MISC, clob);
1064 page = cl_page_find(env, clob, vmpage->index,
1065 vmpage, CPT_CACHEABLE);
1066 if (!IS_ERR(page)) {
1067 lu_ref_add(&page->cp_reference, "writepage",
1069 cl_page_assume(env, io, page);
1070 result = cl_page_flush(env, io, page);
1073 * Re-dirty page on error so it retries write,
1074 * but not in case when IO has actually
1075 * occurred and completed with an error.
1077 if (!PageError(vmpage)) {
1078 redirty_page_for_writepage(wbc, vmpage);
1083 cl_page_disown(env, io, page);
1085 lu_ref_del(&page->cp_reference,
1086 "writepage", current);
1087 cl_page_put(env, page);
1089 result = PTR_ERR(page);
1092 cl_io_fini(env, io);
1094 if (redirtied && wbc->sync_mode == WB_SYNC_ALL) {
1095 loff_t offset = cl_offset(clob, vmpage->index);
1097 /* Flush page failed because the extent is being written out.
1098 * Wait for the write of extent to be finished to avoid
1099 * breaking kernel which assumes ->writepage should mark
1100 * PageWriteback or clean the page. */
1101 result = cl_sync_file_range(inode, offset,
1102 offset + PAGE_CACHE_SIZE - 1,
1105 /* actually we may have written more than one page.
1106 * decreasing this page because the caller will count
1108 wbc->nr_to_write -= result - 1;
1113 cl_env_nested_put(&nest, env);
1118 if (!lli->lli_async_rc)
1119 lli->lli_async_rc = result;
1120 SetPageError(vmpage);
1122 unlock_page(vmpage);
1127 int ll_writepages(struct address_space *mapping, struct writeback_control *wbc)
1129 struct inode *inode = mapping->host;
1130 struct ll_sb_info *sbi = ll_i2sbi(inode);
1133 enum cl_fsync_mode mode;
1134 int range_whole = 0;
1136 int ignore_layout = 0;
1139 if (wbc->range_cyclic) {
1140 start = mapping->writeback_index << PAGE_CACHE_SHIFT;
1141 end = OBD_OBJECT_EOF;
1143 start = wbc->range_start;
1144 end = wbc->range_end;
1145 if (end == LLONG_MAX) {
1146 end = OBD_OBJECT_EOF;
1147 range_whole = start == 0;
1151 mode = CL_FSYNC_NONE;
1152 if (wbc->sync_mode == WB_SYNC_ALL)
1153 mode = CL_FSYNC_LOCAL;
1155 if (sbi->ll_umounting)
1156 /* if the mountpoint is being umounted, all pages have to be
1157 * evicted to avoid hitting LBUG when truncate_inode_pages()
1158 * is called later on. */
1161 if (cl_i2info(inode)->lli_clob == NULL)
1164 result = cl_sync_file_range(inode, start, end, mode, ignore_layout);
1166 wbc->nr_to_write -= result;
1170 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) {
1171 if (end == OBD_OBJECT_EOF)
1172 mapping->writeback_index = 0;
1174 mapping->writeback_index = (end >> PAGE_CACHE_SHIFT) +1;
1179 int ll_readpage(struct file *file, struct page *vmpage)
1181 struct ll_cl_context *lcc;
1185 lcc = ll_cl_init(file, vmpage);
1187 struct lu_env *env = lcc->lcc_env;
1188 struct cl_io *io = lcc->lcc_io;
1189 struct cl_page *page = lcc->lcc_page;
1191 LASSERT(page->cp_type == CPT_CACHEABLE);
1192 if (likely(!PageUptodate(vmpage))) {
1193 cl_page_assume(env, io, page);
1194 result = cl_io_read_page(env, io, page);
1196 /* Page from a non-object file. */
1197 unlock_page(vmpage);
1202 unlock_page(vmpage);
1203 result = PTR_ERR(lcc);
1208 int ll_page_sync_io(const struct lu_env *env, struct cl_io *io,
1209 struct cl_page *page, enum cl_req_type crt)
1211 struct cl_2queue *queue;
1214 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
1216 queue = &io->ci_queue;
1217 cl_2queue_init_page(queue, page);
1219 result = cl_io_submit_sync(env, io, crt, queue, 0);
1220 LASSERT(cl_page_is_owned(page, io));
1222 if (crt == CRT_READ)
1224 * in CRT_WRITE case page is left locked even in case of
1227 cl_page_list_disown(env, io, &queue->c2_qin);
1228 cl_2queue_fini(env, queue);