1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
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/autoconf.h>
42 #include <linux/kernel.h>
44 #include <linux/string.h>
45 #include <linux/stat.h>
46 #include <linux/errno.h>
47 #include <linux/smp_lock.h>
48 #include <linux/unistd.h>
49 #include <linux/version.h>
50 #include <asm/system.h>
51 #include <asm/uaccess.h>
54 #include <linux/stat.h>
55 #include <asm/uaccess.h>
57 #include <linux/pagemap.h>
58 #include <linux/smp_lock.h>
59 /* current_is_kswapd() */
60 #include <linux/swap.h>
62 #define DEBUG_SUBSYSTEM S_LLITE
64 #include <lustre_lite.h>
65 #include <obd_cksum.h>
66 #include "llite_internal.h"
67 #include <linux/lustre_compat25.h>
69 /* this isn't where truncate starts. roughly:
70 * sys_truncate->ll_setattr_raw->vmtruncate->ll_truncate. setattr_raw grabs
71 * DLM lock on [size, EOF], i_mutex, ->lli_size_sem, and WRITE_I_ALLOC_SEM to
74 * must be called under ->lli_size_sem */
75 void ll_truncate(struct inode *inode)
77 struct ll_inode_info *lli = ll_i2info(inode);
80 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) to %Lu\n",inode->i_ino,
81 inode->i_generation, inode, i_size_read(inode));
83 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_TRUNC, 1);
84 if (lli->lli_size_sem_owner == cfs_current()) {
85 LASSERT_SEM_LOCKED(&lli->lli_size_sem);
86 ll_inode_size_unlock(inode, 0);
94 * Finalizes cl-data before exiting typical address_space operation. Dual to
97 static void ll_cl_fini(struct ll_cl_context *lcc)
99 struct lu_env *env = lcc->lcc_env;
100 struct cl_io *io = lcc->lcc_io;
101 struct cl_page *page = lcc->lcc_page;
103 LASSERT(lcc->lcc_cookie == current);
104 LASSERT(env != NULL);
107 lu_ref_del(&page->cp_reference, "cl_io", io);
108 cl_page_put(env, page);
111 if (io && lcc->lcc_created) {
113 cl_io_unlock(env, io);
114 cl_io_iter_fini(env, io);
117 cl_env_put(env, &lcc->lcc_refcheck);
121 * Initializes common cl-data at the typical address_space operation entry
124 static struct ll_cl_context *ll_cl_init(struct file *file,
125 struct page *vmpage, int create)
127 struct ll_cl_context *lcc;
130 struct cl_object *clob;
136 clob = ll_i2info(vmpage->mapping->host)->lli_clob;
137 LASSERT(clob != NULL);
139 env = cl_env_get(&refcheck);
141 return ERR_PTR(PTR_ERR(env));
143 lcc = &vvp_env_info(env)->vti_io_ctx;
144 memset(lcc, 0, sizeof(*lcc));
146 lcc->lcc_refcheck = refcheck;
147 lcc->lcc_cookie = current;
149 cio = ccc_env_io(env);
150 io = cio->cui_cl.cis_io;
151 if (io == NULL && create) {
156 * Loop-back driver calls ->prepare_write() and ->sendfile()
157 * methods directly, bypassing file system ->write() operation,
158 * so cl_io has to be created here.
161 io = ccc_env_thread_io(env);
162 vio = vvp_env_io(env);
163 ll_io_init(io, file, 1);
165 /* No lock at all for this kind of IO - we can't do it because
166 * we have held page lock, it would cause deadlock.
167 * XXX: This causes poor performance to loop device - One page
169 * In order to get better performance, users should use
170 * lloop driver instead.
172 io->ci_lockreq = CILR_NEVER;
174 pos = (vmpage->index << CFS_PAGE_SHIFT);
176 /* Create a temp IO to serve write. */
177 result = cl_io_rw_init(env, io, CIT_WRITE, pos, CFS_PAGE_SIZE);
179 cio->cui_fd = LUSTRE_FPRIVATE(file);
182 result = cl_io_iter_init(env, io);
184 result = cl_io_lock(env, io);
186 result = cl_io_start(env, io);
189 result = io->ci_result;
190 lcc->lcc_created = 1;
197 struct cl_page *page;
200 LASSERT(io->ci_state == CIS_IO_GOING);
201 LASSERT(cio->cui_fd == LUSTRE_FPRIVATE(file));
202 page = cl_page_find(env, clob, vmpage->index, vmpage,
205 lcc->lcc_page = page;
206 lu_ref_add(&page->cp_reference, "cl_io", io);
209 result = PTR_ERR(page);
213 lcc = ERR_PTR(result);
216 CDEBUG(D_VFSTRACE, "%lu@"DFID" -> %d %p %p\n",
217 vmpage->index, PFID(lu_object_fid(&clob->co_lu)), result,
222 static struct ll_cl_context *ll_cl_get(void)
224 struct ll_cl_context *lcc;
228 env = cl_env_get(&refcheck);
229 LASSERT(!IS_ERR(env));
230 lcc = &vvp_env_info(env)->vti_io_ctx;
231 LASSERT(env == lcc->lcc_env);
232 LASSERT(current == lcc->lcc_cookie);
233 cl_env_put(env, &refcheck);
235 /* env has got in ll_cl_init, so it is still usable. */
240 * ->prepare_write() address space operation called by generic_file_write()
241 * for every page during write.
243 int ll_prepare_write(struct file *file, struct page *vmpage, unsigned from,
246 struct ll_cl_context *lcc;
250 lcc = ll_cl_init(file, vmpage, 1);
252 struct lu_env *env = lcc->lcc_env;
253 struct cl_io *io = lcc->lcc_io;
254 struct cl_page *page = lcc->lcc_page;
256 cl_page_assume(env, io, page);
257 if (cl_io_is_append(io)) {
258 struct cl_object *obj = io->ci_obj;
259 struct inode *inode = ccc_object_inode(obj);
261 * In VFS file->page write loop, for appending, the
262 * write offset might be reset according to the new
263 * file size before holding i_mutex. So crw_pos should
264 * be reset here. BUG:17711.
266 io->u.ci_wr.wr.crw_pos = i_size_read(inode);
268 result = cl_io_prepare_write(env, io, page, from, to);
271 * Add a reference, so that page is not evicted from
272 * the cache until ->commit_write() is called.
275 lu_ref_add(&page->cp_reference, "prepare_write",
278 cl_page_unassume(env, io, page);
281 /* returning 0 in prepare assumes commit must be called
284 result = PTR_ERR(lcc);
289 int ll_commit_write(struct file *file, struct page *vmpage, unsigned from,
292 struct ll_cl_context *lcc;
295 struct cl_page *page;
301 page = lcc->lcc_page;
304 LASSERT(cl_page_is_owned(page, io));
306 if (from != to) /* handle short write case. */
307 result = cl_io_commit_write(env, io, page, from, to);
308 if (cl_page_is_owned(page, io))
309 cl_page_unassume(env, io, page);
312 * Release reference acquired by ll_prepare_write().
314 lu_ref_del(&page->cp_reference, "prepare_write", cfs_current());
315 cl_page_put(env, page);
320 struct obd_capa *cl_capa_lookup(struct inode *inode, enum cl_req_type crt)
324 opc = crt == CRT_WRITE ? CAPA_OPC_OSS_WRITE : CAPA_OPC_OSS_RW;
325 return ll_osscapa_get(inode, opc);
328 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which);
330 /* WARNING: This algorithm is used to reduce the contention on
331 * sbi->ll_lock. It should work well if the ra_max_pages is much
332 * greater than the single file's read-ahead window.
334 * TODO: There may exist a `global sync problem' in this implementation.
335 * Considering the global ra window is 100M, and each file's ra window is 10M,
336 * there are over 10 files trying to get its ra budget and reach
337 * ll_ra_count_get at the exactly same time. All of them will get a zero ra
338 * window, although the global window is 100M. -jay
340 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, struct ra_io_arg *ria,
343 struct ll_ra_info *ra = &sbi->ll_ra_info;
348 * If read-ahead pages left are less than 1M, do not do read-ahead,
349 * otherwise it will form small read RPC(< 1M), which hurt server
352 if (ra->ra_max_pages < atomic_read(&ra->ra_cur_pages))
355 ret = min(ra->ra_max_pages - cfs_atomic_read(&ra->ra_cur_pages), len);
356 if ((int)ret < 0 || ret < min((unsigned long)PTLRPC_MAX_BRW_PAGES, len))
359 if (ria->ria_pages == 0)
360 /* it needs 1M align again after trimed by ra_max_pages*/
361 if (ret >= ((ria->ria_start + ret) % PTLRPC_MAX_BRW_PAGES))
362 ret -= (ria->ria_start + ret) % PTLRPC_MAX_BRW_PAGES;
364 if (cfs_atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) {
365 cfs_atomic_sub(ret, &ra->ra_cur_pages);
373 void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
375 struct ll_ra_info *ra = &sbi->ll_ra_info;
376 cfs_atomic_sub(len, &ra->ra_cur_pages);
379 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which)
381 LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
382 lprocfs_counter_incr(sbi->ll_ra_stats, which);
385 void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which)
387 struct ll_sb_info *sbi = ll_i2sbi(mapping->host);
388 ll_ra_stats_inc_sbi(sbi, which);
391 #define RAS_CDEBUG(ras) \
393 "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu r %lu ri %lu" \
394 "csr %lu sf %lu sp %lu sl %lu \n", \
395 ras->ras_last_readpage, ras->ras_consecutive_requests, \
396 ras->ras_consecutive_pages, ras->ras_window_start, \
397 ras->ras_window_len, ras->ras_next_readahead, \
398 ras->ras_requests, ras->ras_request_index, \
399 ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
400 ras->ras_stride_pages, ras->ras_stride_length)
402 static int index_in_window(unsigned long index, unsigned long point,
403 unsigned long before, unsigned long after)
405 unsigned long start = point - before, end = point + after;
412 return start <= index && index <= end;
415 static struct ll_readahead_state *ll_ras_get(struct file *f)
417 struct ll_file_data *fd;
419 fd = LUSTRE_FPRIVATE(f);
423 void ll_ra_read_in(struct file *f, struct ll_ra_read *rar)
425 struct ll_readahead_state *ras;
429 cfs_spin_lock(&ras->ras_lock);
431 ras->ras_request_index = 0;
432 ras->ras_consecutive_requests++;
433 rar->lrr_reader = current;
435 cfs_list_add(&rar->lrr_linkage, &ras->ras_read_beads);
436 cfs_spin_unlock(&ras->ras_lock);
439 void ll_ra_read_ex(struct file *f, struct ll_ra_read *rar)
441 struct ll_readahead_state *ras;
445 cfs_spin_lock(&ras->ras_lock);
446 cfs_list_del_init(&rar->lrr_linkage);
447 cfs_spin_unlock(&ras->ras_lock);
450 static struct ll_ra_read *ll_ra_read_get_locked(struct ll_readahead_state *ras)
452 struct ll_ra_read *scan;
454 cfs_list_for_each_entry(scan, &ras->ras_read_beads, lrr_linkage) {
455 if (scan->lrr_reader == current)
461 struct ll_ra_read *ll_ra_read_get(struct file *f)
463 struct ll_readahead_state *ras;
464 struct ll_ra_read *bead;
468 cfs_spin_lock(&ras->ras_lock);
469 bead = ll_ra_read_get_locked(ras);
470 cfs_spin_unlock(&ras->ras_lock);
474 static int cl_read_ahead_page(const struct lu_env *env, struct cl_io *io,
475 struct cl_page_list *queue, struct cl_page *page,
484 cl_page_assume(env, io, page);
485 lu_ref_add(&page->cp_reference, "ra", cfs_current());
486 cp = cl2ccc_page(cl_page_at(page, &vvp_device_type));
487 if (!cp->cpg_defer_uptodate && !Page_Uptodate(vmpage)) {
488 rc = cl_page_is_under_lock(env, io, page);
490 cp->cpg_defer_uptodate = 1;
492 cl_page_list_add(queue, page);
495 cl_page_delete(env, page);
499 /* skip completed pages */
500 cl_page_unassume(env, io, page);
501 lu_ref_del(&page->cp_reference, "ra", cfs_current());
502 cl_page_put(env, page);
507 * Initiates read-ahead of a page with given index.
509 * \retval +ve: page was added to \a queue.
511 * \retval -ENOLCK: there is no extent lock for this part of a file, stop
514 * \retval -ve, 0: page wasn't added to \a queue for other reason.
516 static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io,
517 struct cl_page_list *queue,
518 pgoff_t index, struct address_space *mapping)
521 struct cl_object *clob = ll_i2info(mapping->host)->lli_clob;
522 struct cl_page *page;
523 enum ra_stat which = _NR_RA_STAT; /* keep gcc happy */
524 unsigned int gfp_mask;
526 const char *msg = NULL;
530 gfp_mask = GFP_HIGHUSER & ~__GFP_WAIT;
532 gfp_mask |= __GFP_NOWARN;
534 vmpage = grab_cache_page_nowait_gfp(mapping, index, gfp_mask);
535 if (vmpage != NULL) {
536 /* Check if vmpage was truncated or reclaimed */
537 if (vmpage->mapping == mapping) {
538 page = cl_page_find(env, clob, vmpage->index,
539 vmpage, CPT_CACHEABLE);
541 rc = cl_read_ahead_page(env, io, queue,
544 which = RA_STAT_FAILED_MATCH;
545 msg = "lock match failed";
548 which = RA_STAT_FAILED_GRAB_PAGE;
549 msg = "cl_page_find failed";
552 which = RA_STAT_WRONG_GRAB_PAGE;
553 msg = "g_c_p_n returned invalid page";
557 page_cache_release(vmpage);
559 which = RA_STAT_FAILED_GRAB_PAGE;
560 msg = "g_c_p_n failed";
563 ll_ra_stats_inc(mapping, which);
564 CDEBUG(D_READA, "%s\n", msg);
569 #define RIA_DEBUG(ria) \
570 CDEBUG(D_READA, "rs %lu re %lu ro %lu rl %lu rp %lu\n", \
571 ria->ria_start, ria->ria_end, ria->ria_stoff, ria->ria_length,\
574 #define RAS_INCREASE_STEP PTLRPC_MAX_BRW_PAGES
576 static inline int stride_io_mode(struct ll_readahead_state *ras)
578 return ras->ras_consecutive_stride_requests > 1;
580 /* The function calculates how much pages will be read in
581 * [off, off + length], in such stride IO area,
582 * stride_offset = st_off, stride_lengh = st_len,
583 * stride_pages = st_pgs
585 * |------------------|*****|------------------|*****|------------|*****|....
588 * |----- st_len -----|
590 * How many pages it should read in such pattern
591 * |-------------------------------------------------------------|
593 * |<------ length ------->|
595 * = |<----->| + |-------------------------------------| + |---|
596 * start_left st_pgs * i end_left
599 stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs,
600 unsigned long off, unsigned long length)
602 __u64 start = off > st_off ? off - st_off : 0;
603 __u64 end = off + length > st_off ? off + length - st_off : 0;
604 unsigned long start_left = 0;
605 unsigned long end_left = 0;
606 unsigned long pg_count;
608 if (st_len == 0 || length == 0 || end == 0)
611 start_left = do_div(start, st_len);
612 if (start_left < st_pgs)
613 start_left = st_pgs - start_left;
617 end_left = do_div(end, st_len);
618 if (end_left > st_pgs)
621 CDEBUG(D_READA, "start "LPU64", end "LPU64" start_left %lu end_left %lu \n",
622 start, end, start_left, end_left);
625 pg_count = end_left - (st_pgs - start_left);
627 pg_count = start_left + st_pgs * (end - start - 1) + end_left;
629 CDEBUG(D_READA, "st_off %lu, st_len %lu st_pgs %lu off %lu length %lu"
630 "pgcount %lu\n", st_off, st_len, st_pgs, off, length, pg_count);
635 static int ria_page_count(struct ra_io_arg *ria)
637 __u64 length = ria->ria_end >= ria->ria_start ?
638 ria->ria_end - ria->ria_start + 1 : 0;
640 return stride_pg_count(ria->ria_stoff, ria->ria_length,
641 ria->ria_pages, ria->ria_start,
645 /*Check whether the index is in the defined ra-window */
646 static int ras_inside_ra_window(unsigned long idx, struct ra_io_arg *ria)
648 /* If ria_length == ria_pages, it means non-stride I/O mode,
649 * idx should always inside read-ahead window in this case
650 * For stride I/O mode, just check whether the idx is inside
652 return ria->ria_length == 0 || ria->ria_length == ria->ria_pages ||
653 (idx >= ria->ria_stoff && (idx - ria->ria_stoff) %
654 ria->ria_length < ria->ria_pages);
657 static int ll_read_ahead_pages(const struct lu_env *env,
658 struct cl_io *io, struct cl_page_list *queue,
659 struct ra_io_arg *ria,
660 unsigned long *reserved_pages,
661 struct address_space *mapping,
662 unsigned long *ra_end)
664 int rc, count = 0, stride_ria;
665 unsigned long page_idx;
667 LASSERT(ria != NULL);
670 stride_ria = ria->ria_length > ria->ria_pages && ria->ria_pages > 0;
671 for (page_idx = ria->ria_start; page_idx <= ria->ria_end &&
672 *reserved_pages > 0; page_idx++) {
673 if (ras_inside_ra_window(page_idx, ria)) {
674 /* If the page is inside the read-ahead window*/
675 rc = ll_read_ahead_page(env, io, queue,
680 } else if (rc == -ENOLCK)
682 } else if (stride_ria) {
683 /* If it is not in the read-ahead window, and it is
684 * read-ahead mode, then check whether it should skip
687 /* FIXME: This assertion only is valid when it is for
688 * forward read-ahead, it will be fixed when backward
689 * read-ahead is implemented */
690 LASSERTF(page_idx > ria->ria_stoff, "Invalid page_idx %lu"
691 "rs %lu re %lu ro %lu rl %lu rp %lu\n", page_idx,
692 ria->ria_start, ria->ria_end, ria->ria_stoff,
693 ria->ria_length, ria->ria_pages);
694 offset = page_idx - ria->ria_stoff;
695 offset = offset % (ria->ria_length);
696 if (offset > ria->ria_pages) {
697 page_idx += ria->ria_length - offset;
698 CDEBUG(D_READA, "i %lu skip %lu \n", page_idx,
699 ria->ria_length - offset);
708 int ll_readahead(const struct lu_env *env, struct cl_io *io,
709 struct ll_readahead_state *ras, struct address_space *mapping,
710 struct cl_page_list *queue, int flags)
712 struct vvp_io *vio = vvp_env_io(env);
713 struct vvp_thread_info *vti = vvp_env_info(env);
714 struct cl_attr *attr = ccc_env_thread_attr(env);
715 unsigned long start = 0, end = 0, reserved;
716 unsigned long ra_end, len;
718 struct ll_ra_read *bead;
719 struct ra_io_arg *ria = &vti->vti_ria;
720 struct ll_inode_info *lli;
721 struct cl_object *clob;
726 inode = mapping->host;
727 lli = ll_i2info(inode);
728 clob = lli->lli_clob;
730 memset(ria, 0, sizeof *ria);
732 cl_object_attr_lock(clob);
733 ret = cl_object_attr_get(env, clob, attr);
734 cl_object_attr_unlock(clob);
740 ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN);
744 cfs_spin_lock(&ras->ras_lock);
745 if (vio->cui_ra_window_set)
746 bead = &vio->cui_bead;
750 /* Enlarge the RA window to encompass the full read */
751 if (bead != NULL && ras->ras_window_start + ras->ras_window_len <
752 bead->lrr_start + bead->lrr_count) {
753 ras->ras_window_len = bead->lrr_start + bead->lrr_count -
754 ras->ras_window_start;
756 /* Reserve a part of the read-ahead window that we'll be issuing */
757 if (ras->ras_window_len) {
758 start = ras->ras_next_readahead;
759 end = ras->ras_window_start + ras->ras_window_len - 1;
762 unsigned long tmp_end;
764 * Align RA window to an optimal boundary.
766 * XXX This would be better to align to cl_max_pages_per_rpc
767 * instead of PTLRPC_MAX_BRW_PAGES, because the RPC size may
768 * be aligned to the RAID stripe size in the future and that
769 * is more important than the RPC size.
771 tmp_end = ((end + 1) & (~(PTLRPC_MAX_BRW_PAGES - 1))) - 1;
775 /* Truncate RA window to end of file */
776 end = min(end, (unsigned long)((kms - 1) >> CFS_PAGE_SHIFT));
778 ras->ras_next_readahead = max(end, end + 1);
781 ria->ria_start = start;
783 /* If stride I/O mode is detected, get stride window*/
784 if (stride_io_mode(ras)) {
785 ria->ria_stoff = ras->ras_stride_offset;
786 ria->ria_length = ras->ras_stride_length;
787 ria->ria_pages = ras->ras_stride_pages;
789 cfs_spin_unlock(&ras->ras_lock);
792 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
795 len = ria_page_count(ria);
799 reserved = ll_ra_count_get(ll_i2sbi(inode), ria, len);
801 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
803 CDEBUG(D_READA, "reserved page %lu \n", reserved);
805 ret = ll_read_ahead_pages(env, io, queue,
806 ria, &reserved, mapping, &ra_end);
808 LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
810 ll_ra_count_put(ll_i2sbi(inode), reserved);
812 if (ra_end == end + 1 && ra_end == (kms >> CFS_PAGE_SHIFT))
813 ll_ra_stats_inc(mapping, RA_STAT_EOF);
815 /* if we didn't get to the end of the region we reserved from
816 * the ras we need to go back and update the ras so that the
817 * next read-ahead tries from where we left off. we only do so
818 * if the region we failed to issue read-ahead on is still ahead
819 * of the app and behind the next index to start read-ahead from */
820 CDEBUG(D_READA, "ra_end %lu end %lu stride end %lu \n",
821 ra_end, end, ria->ria_end);
823 if (ra_end != end + 1) {
824 cfs_spin_lock(&ras->ras_lock);
825 if (ra_end < ras->ras_next_readahead &&
826 index_in_window(ra_end, ras->ras_window_start, 0,
827 ras->ras_window_len)) {
828 ras->ras_next_readahead = ra_end;
831 cfs_spin_unlock(&ras->ras_lock);
837 static void ras_set_start(struct ll_readahead_state *ras, unsigned long index)
839 ras->ras_window_start = index & (~(RAS_INCREASE_STEP - 1));
842 /* called with the ras_lock held or from places where it doesn't matter */
843 static void ras_reset(struct ll_readahead_state *ras, unsigned long index)
845 ras->ras_last_readpage = index;
846 ras->ras_consecutive_requests = 0;
847 ras->ras_consecutive_pages = 0;
848 ras->ras_window_len = 0;
849 ras_set_start(ras, index);
850 ras->ras_next_readahead = max(ras->ras_window_start, index);
855 /* called with the ras_lock held or from places where it doesn't matter */
856 static void ras_stride_reset(struct ll_readahead_state *ras)
858 ras->ras_consecutive_stride_requests = 0;
859 ras->ras_stride_length = 0;
860 ras->ras_stride_pages = 0;
864 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
866 cfs_spin_lock_init(&ras->ras_lock);
868 ras->ras_requests = 0;
869 CFS_INIT_LIST_HEAD(&ras->ras_read_beads);
873 * Check whether the read request is in the stride window.
874 * If it is in the stride window, return 1, otherwise return 0.
876 static int index_in_stride_window(unsigned long index,
877 struct ll_readahead_state *ras,
880 unsigned long stride_gap = index - ras->ras_last_readpage - 1;
882 if (ras->ras_stride_length == 0 || ras->ras_stride_pages == 0 ||
883 ras->ras_stride_pages == ras->ras_stride_length)
886 /* If it is contiguous read */
888 return ras->ras_consecutive_pages + 1 <= ras->ras_stride_pages;
890 /*Otherwise check the stride by itself */
891 return (ras->ras_stride_length - ras->ras_stride_pages) == stride_gap &&
892 ras->ras_consecutive_pages == ras->ras_stride_pages;
895 static void ras_update_stride_detector(struct ll_readahead_state *ras,
898 unsigned long stride_gap = index - ras->ras_last_readpage - 1;
900 if (!stride_io_mode(ras) && (stride_gap != 0 ||
901 ras->ras_consecutive_stride_requests == 0)) {
902 ras->ras_stride_pages = ras->ras_consecutive_pages;
903 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
905 LASSERT(ras->ras_request_index == 0);
906 LASSERT(ras->ras_consecutive_stride_requests == 0);
908 if (index <= ras->ras_last_readpage) {
909 /*Reset stride window for forward read*/
910 ras_stride_reset(ras);
914 ras->ras_stride_pages = ras->ras_consecutive_pages;
915 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
922 stride_page_count(struct ll_readahead_state *ras, unsigned long len)
924 return stride_pg_count(ras->ras_stride_offset, ras->ras_stride_length,
925 ras->ras_stride_pages, ras->ras_stride_offset,
929 /* Stride Read-ahead window will be increased inc_len according to
930 * stride I/O pattern */
931 static void ras_stride_increase_window(struct ll_readahead_state *ras,
932 struct ll_ra_info *ra,
933 unsigned long inc_len)
935 unsigned long left, step, window_len;
936 unsigned long stride_len;
938 LASSERT(ras->ras_stride_length > 0);
939 LASSERTF(ras->ras_window_start + ras->ras_window_len
940 >= ras->ras_stride_offset, "window_start %lu, window_len %lu"
941 " stride_offset %lu\n", ras->ras_window_start,
942 ras->ras_window_len, ras->ras_stride_offset);
944 stride_len = ras->ras_window_start + ras->ras_window_len -
945 ras->ras_stride_offset;
947 left = stride_len % ras->ras_stride_length;
948 window_len = ras->ras_window_len - left;
950 if (left < ras->ras_stride_pages)
953 left = ras->ras_stride_pages + inc_len;
955 LASSERT(ras->ras_stride_pages != 0);
957 step = left / ras->ras_stride_pages;
958 left %= ras->ras_stride_pages;
960 window_len += step * ras->ras_stride_length + left;
962 if (stride_page_count(ras, window_len) <= ra->ra_max_pages_per_file)
963 ras->ras_window_len = window_len;
968 static void ras_increase_window(struct ll_readahead_state *ras,
969 struct ll_ra_info *ra, struct inode *inode)
971 /* The stretch of ra-window should be aligned with max rpc_size
972 * but current clio architecture does not support retrieve such
973 * information from lower layer. FIXME later
975 if (stride_io_mode(ras))
976 ras_stride_increase_window(ras, ra, RAS_INCREASE_STEP);
978 ras->ras_window_len = min(ras->ras_window_len +
980 ra->ra_max_pages_per_file);
983 void ras_update(struct ll_sb_info *sbi, struct inode *inode,
984 struct ll_readahead_state *ras, unsigned long index,
987 struct ll_ra_info *ra = &sbi->ll_ra_info;
988 int zero = 0, stride_detect = 0, ra_miss = 0;
991 cfs_spin_lock(&ras->ras_lock);
993 ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
995 /* reset the read-ahead window in two cases. First when the app seeks
996 * or reads to some other part of the file. Secondly if we get a
997 * read-ahead miss that we think we've previously issued. This can
998 * be a symptom of there being so many read-ahead pages that the VM is
999 * reclaiming it before we get to it. */
1000 if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
1002 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
1003 } else if (!hit && ras->ras_window_len &&
1004 index < ras->ras_next_readahead &&
1005 index_in_window(index, ras->ras_window_start, 0,
1006 ras->ras_window_len)) {
1008 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
1011 /* On the second access to a file smaller than the tunable
1012 * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1013 * file up to ra_max_pages_per_file. This is simply a best effort
1014 * and only occurs once per open file. Normal RA behavior is reverted
1015 * to for subsequent IO. The mmap case does not increment
1016 * ras_requests and thus can never trigger this behavior. */
1017 if (ras->ras_requests == 2 && !ras->ras_request_index) {
1020 kms_pages = (i_size_read(inode) + CFS_PAGE_SIZE - 1) >>
1023 CDEBUG(D_READA, "kmsp "LPU64" mwp %lu mp %lu\n", kms_pages,
1024 ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages_per_file);
1027 kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1028 ras->ras_window_start = 0;
1029 ras->ras_last_readpage = 0;
1030 ras->ras_next_readahead = 0;
1031 ras->ras_window_len = min(ra->ra_max_pages_per_file,
1032 ra->ra_max_read_ahead_whole_pages);
1033 GOTO(out_unlock, 0);
1037 /* check whether it is in stride I/O mode*/
1038 if (!index_in_stride_window(index, ras, inode)) {
1039 if (ras->ras_consecutive_stride_requests == 0 &&
1040 ras->ras_request_index == 0) {
1041 ras_update_stride_detector(ras, index);
1042 ras->ras_consecutive_stride_requests ++;
1044 ras_stride_reset(ras);
1046 ras_reset(ras, index);
1047 ras->ras_consecutive_pages++;
1048 GOTO(out_unlock, 0);
1050 ras->ras_consecutive_pages = 0;
1051 ras->ras_consecutive_requests = 0;
1052 if (++ras->ras_consecutive_stride_requests > 1)
1058 if (index_in_stride_window(index, ras, inode) &&
1059 stride_io_mode(ras)) {
1060 /*If stride-RA hit cache miss, the stride dector
1061 *will not be reset to avoid the overhead of
1062 *redetecting read-ahead mode */
1063 if (index != ras->ras_last_readpage + 1)
1064 ras->ras_consecutive_pages = 0;
1065 ras_reset(ras, index);
1068 /* Reset both stride window and normal RA
1070 ras_reset(ras, index);
1071 ras->ras_consecutive_pages++;
1072 ras_stride_reset(ras);
1073 GOTO(out_unlock, 0);
1075 } else if (stride_io_mode(ras)) {
1076 /* If this is contiguous read but in stride I/O mode
1077 * currently, check whether stride step still is valid,
1078 * if invalid, it will reset the stride ra window*/
1079 if (!index_in_stride_window(index, ras, inode)) {
1080 /* Shrink stride read-ahead window to be zero */
1081 ras_stride_reset(ras);
1082 ras->ras_window_len = 0;
1083 ras->ras_next_readahead = index;
1087 ras->ras_consecutive_pages++;
1088 ras->ras_last_readpage = index;
1089 ras_set_start(ras, index);
1091 if (stride_io_mode(ras))
1092 /* Since stride readahead is sentivite to the offset
1093 * of read-ahead, so we use original offset here,
1094 * instead of ras_window_start, which is 1M aligned*/
1095 ras->ras_next_readahead = max(index,
1096 ras->ras_next_readahead);
1098 ras->ras_next_readahead = max(ras->ras_window_start,
1099 ras->ras_next_readahead);
1102 /* Trigger RA in the mmap case where ras_consecutive_requests
1103 * is not incremented and thus can't be used to trigger RA */
1104 if (!ras->ras_window_len && ras->ras_consecutive_pages == 4) {
1105 ras->ras_window_len = RAS_INCREASE_STEP;
1106 GOTO(out_unlock, 0);
1109 /* Initially reset the stride window offset to next_readahead*/
1110 if (ras->ras_consecutive_stride_requests == 2 && stride_detect) {
1112 * Once stride IO mode is detected, next_readahead should be
1113 * reset to make sure next_readahead > stride offset
1115 ras->ras_next_readahead = max(index, ras->ras_next_readahead);
1116 ras->ras_stride_offset = index;
1117 ras->ras_window_len = RAS_INCREASE_STEP;
1120 /* The initial ras_window_len is set to the request size. To avoid
1121 * uselessly reading and discarding pages for random IO the window is
1122 * only increased once per consecutive request received. */
1123 if ((ras->ras_consecutive_requests > 1 || stride_detect) &&
1124 !ras->ras_request_index)
1125 ras_increase_window(ras, ra, inode);
1129 ras->ras_request_index++;
1130 cfs_spin_unlock(&ras->ras_lock);
1134 int ll_writepage(struct page *vmpage, struct writeback_control *unused)
1136 struct inode *inode = vmpage->mapping->host;
1139 struct cl_page *page;
1140 struct cl_object *clob;
1141 struct cl_2queue *queue;
1142 struct cl_env_nest nest;
1146 LASSERT(PageLocked(vmpage));
1147 LASSERT(!PageWriteback(vmpage));
1149 if (ll_i2dtexp(inode) == NULL)
1152 env = cl_env_nested_get(&nest);
1154 RETURN(PTR_ERR(env));
1156 queue = &vvp_env_info(env)->vti_queue;
1157 clob = ll_i2info(inode)->lli_clob;
1158 LASSERT(clob != NULL);
1160 io = ccc_env_thread_io(env);
1162 result = cl_io_init(env, io, CIT_MISC, clob);
1164 page = cl_page_find(env, clob, vmpage->index,
1165 vmpage, CPT_CACHEABLE);
1166 if (!IS_ERR(page)) {
1167 lu_ref_add(&page->cp_reference, "writepage",
1169 cl_page_assume(env, io, page);
1171 * Mark page dirty, because this is what
1172 * ->vio_submit()->cpo_prep_write() assumes.
1174 * XXX better solution is to detect this from within
1175 * cl_io_submit_rw() somehow.
1177 set_page_dirty(vmpage);
1178 cl_2queue_init_page(queue, page);
1179 result = cl_io_submit_rw(env, io, CRT_WRITE,
1181 cl_page_list_disown(env, io, &queue->c2_qin);
1184 * There is no need to clear PG_writeback, as
1185 * cl_io_submit_rw() calls completion callback
1189 * Re-dirty page on error so it retries write,
1190 * but not in case when IO has actually
1191 * occurred and completed with an error.
1193 if (!PageError(vmpage))
1194 set_page_dirty(vmpage);
1196 LASSERT(!cl_page_is_owned(page, io));
1197 lu_ref_del(&page->cp_reference,
1198 "writepage", cfs_current());
1199 cl_page_put(env, page);
1200 cl_2queue_fini(env, queue);
1203 cl_io_fini(env, io);
1204 cl_env_nested_put(&nest, env);
1208 int ll_readpage(struct file *file, struct page *vmpage)
1210 struct ll_cl_context *lcc;
1214 lcc = ll_cl_init(file, vmpage, 0);
1216 struct lu_env *env = lcc->lcc_env;
1217 struct cl_io *io = lcc->lcc_io;
1218 struct cl_page *page = lcc->lcc_page;
1220 LASSERT(page->cp_type == CPT_CACHEABLE);
1221 if (likely(!PageUptodate(vmpage))) {
1222 cl_page_assume(env, io, page);
1223 result = cl_io_read_page(env, io, page);
1225 /* Page from a non-object file. */
1226 LASSERT(!ll_i2info(vmpage->mapping->host)->lli_smd);
1227 unlock_page(vmpage);
1232 unlock_page(vmpage);
1233 result = PTR_ERR(lcc);