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_mdc.h>
65 #include <lustre_lite.h>
66 #include <obd_cksum.h>
67 #include "llite_internal.h"
68 #include <linux/lustre_compat25.h>
70 /* this isn't where truncate starts. roughly:
71 * sys_truncate->ll_setattr_raw->vmtruncate->ll_truncate. setattr_raw grabs
72 * DLM lock on [size, EOF], i_mutex, ->lli_size_sem, and WRITE_I_ALLOC_SEM to
75 * must be called under ->lli_size_sem */
76 void ll_truncate(struct inode *inode)
78 struct ll_inode_info *lli = ll_i2info(inode);
81 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) to %Lu\n",inode->i_ino,
82 inode->i_generation, inode, i_size_read(inode));
84 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_TRUNC, 1);
85 if (lli->lli_size_sem_owner == cfs_current()) {
86 LASSERT_SEM_LOCKED(&lli->lli_size_sem);
87 ll_inode_size_unlock(inode, 0);
95 * Finalizes cl-data before exiting typical address_space operation. Dual to
98 static void ll_cl_fini(struct ll_cl_context *lcc)
100 struct lu_env *env = lcc->lcc_env;
101 struct cl_io *io = lcc->lcc_io;
102 struct cl_page *page = lcc->lcc_page;
104 LASSERT(lcc->lcc_cookie == current);
105 LASSERT(env != NULL);
108 lu_ref_del(&page->cp_reference, "cl_io", io);
109 cl_page_put(env, page);
112 if (io && lcc->lcc_created) {
114 cl_io_unlock(env, io);
115 cl_io_iter_fini(env, io);
118 cl_env_put(env, &lcc->lcc_refcheck);
122 * Initializes common cl-data at the typical address_space operation entry
125 static struct ll_cl_context *ll_cl_init(struct file *file,
126 struct page *vmpage, int create)
128 struct ll_cl_context *lcc;
131 struct cl_object *clob;
137 clob = ll_i2info(vmpage->mapping->host)->lli_clob;
138 LASSERT(clob != NULL);
140 env = cl_env_get(&refcheck);
142 return ERR_PTR(PTR_ERR(env));
144 lcc = &vvp_env_info(env)->vti_io_ctx;
145 memset(lcc, 0, sizeof(*lcc));
147 lcc->lcc_refcheck = refcheck;
148 lcc->lcc_cookie = current;
150 cio = ccc_env_io(env);
151 io = cio->cui_cl.cis_io;
152 if (io == NULL && create) {
157 * Loop-back driver calls ->prepare_write() and ->sendfile()
158 * methods directly, bypassing file system ->write() operation,
159 * so cl_io has to be created here.
162 io = ccc_env_thread_io(env);
163 vio = vvp_env_io(env);
164 ll_io_init(io, file, 1);
166 /* No lock at all for this kind of IO - we can't do it because
167 * we have held page lock, it would cause deadlock.
168 * XXX: This causes poor performance to loop device - One page
170 * In order to get better performance, users should use
171 * lloop driver instead.
173 io->ci_lockreq = CILR_NEVER;
175 pos = (vmpage->index << CFS_PAGE_SHIFT);
177 /* Create a temp IO to serve write. */
178 result = cl_io_rw_init(env, io, CIT_WRITE, pos, CFS_PAGE_SIZE);
180 cio->cui_fd = LUSTRE_FPRIVATE(file);
183 result = cl_io_iter_init(env, io);
185 result = cl_io_lock(env, io);
187 result = cl_io_start(env, io);
190 result = io->ci_result;
191 lcc->lcc_created = 1;
198 struct cl_page *page;
201 LASSERT(io->ci_state == CIS_IO_GOING);
202 LASSERT(cio->cui_fd == LUSTRE_FPRIVATE(file));
203 page = cl_page_find(env, clob, vmpage->index, vmpage,
206 lcc->lcc_page = page;
207 lu_ref_add(&page->cp_reference, "cl_io", io);
210 result = PTR_ERR(page);
214 lcc = ERR_PTR(result);
217 CDEBUG(D_VFSTRACE, "%lu@"DFID" -> %d %p %p\n",
218 vmpage->index, PFID(lu_object_fid(&clob->co_lu)), result,
223 static struct ll_cl_context *ll_cl_get(void)
225 struct ll_cl_context *lcc;
229 env = cl_env_get(&refcheck);
230 LASSERT(!IS_ERR(env));
231 lcc = &vvp_env_info(env)->vti_io_ctx;
232 LASSERT(env == lcc->lcc_env);
233 LASSERT(current == lcc->lcc_cookie);
234 cl_env_put(env, &refcheck);
236 /* env has got in ll_cl_init, so it is still usable. */
241 * ->prepare_write() address space operation called by generic_file_write()
242 * for every page during write.
244 int ll_prepare_write(struct file *file, struct page *vmpage, unsigned from,
247 struct ll_cl_context *lcc;
251 lcc = ll_cl_init(file, vmpage, 1);
253 struct lu_env *env = lcc->lcc_env;
254 struct cl_io *io = lcc->lcc_io;
255 struct cl_page *page = lcc->lcc_page;
257 cl_page_assume(env, io, page);
258 if (cl_io_is_append(io)) {
259 struct cl_object *obj = io->ci_obj;
260 struct inode *inode = ccc_object_inode(obj);
262 * In VFS file->page write loop, for appending, the
263 * write offset might be reset according to the new
264 * file size before holding i_mutex. So crw_pos should
265 * be reset here. BUG:17711.
267 io->u.ci_wr.wr.crw_pos = i_size_read(inode);
269 result = cl_io_prepare_write(env, io, page, from, to);
272 * Add a reference, so that page is not evicted from
273 * the cache until ->commit_write() is called.
276 lu_ref_add(&page->cp_reference, "prepare_write",
279 cl_page_unassume(env, io, page);
282 /* returning 0 in prepare assumes commit must be called
285 result = PTR_ERR(lcc);
290 int ll_commit_write(struct file *file, struct page *vmpage, unsigned from,
293 struct ll_cl_context *lcc;
296 struct cl_page *page;
302 page = lcc->lcc_page;
305 LASSERT(cl_page_is_owned(page, io));
307 if (from != to) /* handle short write case. */
308 result = cl_io_commit_write(env, io, page, from, to);
309 if (cl_page_is_owned(page, io))
310 cl_page_unassume(env, io, page);
313 * Release reference acquired by ll_prepare_write().
315 lu_ref_del(&page->cp_reference, "prepare_write", cfs_current());
316 cl_page_put(env, page);
321 struct obd_capa *cl_capa_lookup(struct inode *inode, enum cl_req_type crt)
325 opc = crt == CRT_WRITE ? CAPA_OPC_OSS_WRITE : CAPA_OPC_OSS_RW;
326 return ll_osscapa_get(inode, opc);
329 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which);
331 /* WARNING: This algorithm is used to reduce the contention on
332 * sbi->ll_lock. It should work well if the ra_max_pages is much
333 * greater than the single file's read-ahead window.
335 * TODO: There may exist a `global sync problem' in this implementation.
336 * Considering the global ra window is 100M, and each file's ra window is 10M,
337 * there are over 10 files trying to get its ra budget and reach
338 * ll_ra_count_get at the exactly same time. All of them will get a zero ra
339 * window, although the global window is 100M. -jay
341 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, struct ra_io_arg *ria,
344 struct ll_ra_info *ra = &sbi->ll_ra_info;
349 * If read-ahead pages left are less than 1M, do not do read-ahead,
350 * otherwise it will form small read RPC(< 1M), which hurt server
353 if (ra->ra_max_pages < atomic_read(&ra->ra_cur_pages))
356 ret = min(ra->ra_max_pages - cfs_atomic_read(&ra->ra_cur_pages), len);
357 if ((int)ret < 0 || ret < min((unsigned long)PTLRPC_MAX_BRW_PAGES, len))
360 if (ria->ria_pages == 0)
361 /* it needs 1M align again after trimed by ra_max_pages*/
362 if (ret >= ((ria->ria_start + ret) % PTLRPC_MAX_BRW_PAGES))
363 ret -= (ria->ria_start + ret) % PTLRPC_MAX_BRW_PAGES;
365 if (cfs_atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) {
366 cfs_atomic_sub(ret, &ra->ra_cur_pages);
374 void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
376 struct ll_ra_info *ra = &sbi->ll_ra_info;
377 cfs_atomic_sub(len, &ra->ra_cur_pages);
380 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which)
382 LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
383 lprocfs_counter_incr(sbi->ll_ra_stats, which);
386 void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which)
388 struct ll_sb_info *sbi = ll_i2sbi(mapping->host);
389 ll_ra_stats_inc_sbi(sbi, which);
392 #define RAS_CDEBUG(ras) \
394 "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu r %lu ri %lu" \
395 "csr %lu sf %lu sp %lu sl %lu \n", \
396 ras->ras_last_readpage, ras->ras_consecutive_requests, \
397 ras->ras_consecutive_pages, ras->ras_window_start, \
398 ras->ras_window_len, ras->ras_next_readahead, \
399 ras->ras_requests, ras->ras_request_index, \
400 ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
401 ras->ras_stride_pages, ras->ras_stride_length)
403 static int index_in_window(unsigned long index, unsigned long point,
404 unsigned long before, unsigned long after)
406 unsigned long start = point - before, end = point + after;
413 return start <= index && index <= end;
416 static struct ll_readahead_state *ll_ras_get(struct file *f)
418 struct ll_file_data *fd;
420 fd = LUSTRE_FPRIVATE(f);
424 void ll_ra_read_in(struct file *f, struct ll_ra_read *rar)
426 struct ll_readahead_state *ras;
430 cfs_spin_lock(&ras->ras_lock);
432 ras->ras_request_index = 0;
433 ras->ras_consecutive_requests++;
434 rar->lrr_reader = current;
436 cfs_list_add(&rar->lrr_linkage, &ras->ras_read_beads);
437 cfs_spin_unlock(&ras->ras_lock);
440 void ll_ra_read_ex(struct file *f, struct ll_ra_read *rar)
442 struct ll_readahead_state *ras;
446 cfs_spin_lock(&ras->ras_lock);
447 cfs_list_del_init(&rar->lrr_linkage);
448 cfs_spin_unlock(&ras->ras_lock);
451 static struct ll_ra_read *ll_ra_read_get_locked(struct ll_readahead_state *ras)
453 struct ll_ra_read *scan;
455 cfs_list_for_each_entry(scan, &ras->ras_read_beads, lrr_linkage) {
456 if (scan->lrr_reader == current)
462 struct ll_ra_read *ll_ra_read_get(struct file *f)
464 struct ll_readahead_state *ras;
465 struct ll_ra_read *bead;
469 cfs_spin_lock(&ras->ras_lock);
470 bead = ll_ra_read_get_locked(ras);
471 cfs_spin_unlock(&ras->ras_lock);
475 static int cl_read_ahead_page(const struct lu_env *env, struct cl_io *io,
476 struct cl_page_list *queue, struct cl_page *page,
485 cl_page_assume(env, io, page);
486 lu_ref_add(&page->cp_reference, "ra", cfs_current());
487 cp = cl2ccc_page(cl_page_at(page, &vvp_device_type));
488 if (!cp->cpg_defer_uptodate && !Page_Uptodate(vmpage)) {
489 rc = cl_page_is_under_lock(env, io, page);
491 cp->cpg_defer_uptodate = 1;
493 cl_page_list_add(queue, page);
496 cl_page_delete(env, page);
500 /* skip completed pages */
501 cl_page_unassume(env, io, page);
502 lu_ref_del(&page->cp_reference, "ra", cfs_current());
503 cl_page_put(env, page);
508 * Initiates read-ahead of a page with given index.
510 * \retval +ve: page was added to \a queue.
512 * \retval -ENOLCK: there is no extent lock for this part of a file, stop
515 * \retval -ve, 0: page wasn't added to \a queue for other reason.
517 static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io,
518 struct cl_page_list *queue,
519 pgoff_t index, struct address_space *mapping)
522 struct cl_object *clob = ll_i2info(mapping->host)->lli_clob;
523 struct cl_page *page;
524 enum ra_stat which = _NR_RA_STAT; /* keep gcc happy */
525 unsigned int gfp_mask;
527 const char *msg = NULL;
531 gfp_mask = GFP_HIGHUSER & ~__GFP_WAIT;
533 gfp_mask |= __GFP_NOWARN;
535 vmpage = grab_cache_page_nowait_gfp(mapping, index, gfp_mask);
536 if (vmpage != NULL) {
537 /* Check if vmpage was truncated or reclaimed */
538 if (vmpage->mapping == mapping) {
539 page = cl_page_find(env, clob, vmpage->index,
540 vmpage, CPT_CACHEABLE);
542 rc = cl_read_ahead_page(env, io, queue,
545 which = RA_STAT_FAILED_MATCH;
546 msg = "lock match failed";
549 which = RA_STAT_FAILED_GRAB_PAGE;
550 msg = "cl_page_find failed";
553 which = RA_STAT_WRONG_GRAB_PAGE;
554 msg = "g_c_p_n returned invalid page";
558 page_cache_release(vmpage);
560 which = RA_STAT_FAILED_GRAB_PAGE;
561 msg = "g_c_p_n failed";
564 ll_ra_stats_inc(mapping, which);
565 CDEBUG(D_READA, "%s\n", msg);
570 #define RIA_DEBUG(ria) \
571 CDEBUG(D_READA, "rs %lu re %lu ro %lu rl %lu rp %lu\n", \
572 ria->ria_start, ria->ria_end, ria->ria_stoff, ria->ria_length,\
575 #define RAS_INCREASE_STEP PTLRPC_MAX_BRW_PAGES
577 static inline int stride_io_mode(struct ll_readahead_state *ras)
579 return ras->ras_consecutive_stride_requests > 1;
581 /* The function calculates how much pages will be read in
582 * [off, off + length], in such stride IO area,
583 * stride_offset = st_off, stride_lengh = st_len,
584 * stride_pages = st_pgs
586 * |------------------|*****|------------------|*****|------------|*****|....
589 * |----- st_len -----|
591 * How many pages it should read in such pattern
592 * |-------------------------------------------------------------|
594 * |<------ length ------->|
596 * = |<----->| + |-------------------------------------| + |---|
597 * start_left st_pgs * i end_left
600 stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs,
601 unsigned long off, unsigned long length)
603 __u64 start = off > st_off ? off - st_off : 0;
604 __u64 end = off + length > st_off ? off + length - st_off : 0;
605 unsigned long start_left = 0;
606 unsigned long end_left = 0;
607 unsigned long pg_count;
609 if (st_len == 0 || length == 0 || end == 0)
612 start_left = do_div(start, st_len);
613 if (start_left < st_pgs)
614 start_left = st_pgs - start_left;
618 end_left = do_div(end, st_len);
619 if (end_left > st_pgs)
622 CDEBUG(D_READA, "start "LPU64", end "LPU64" start_left %lu end_left %lu \n",
623 start, end, start_left, end_left);
626 pg_count = end_left - (st_pgs - start_left);
628 pg_count = start_left + st_pgs * (end - start - 1) + end_left;
630 CDEBUG(D_READA, "st_off %lu, st_len %lu st_pgs %lu off %lu length %lu"
631 "pgcount %lu\n", st_off, st_len, st_pgs, off, length, pg_count);
636 static int ria_page_count(struct ra_io_arg *ria)
638 __u64 length = ria->ria_end >= ria->ria_start ?
639 ria->ria_end - ria->ria_start + 1 : 0;
641 return stride_pg_count(ria->ria_stoff, ria->ria_length,
642 ria->ria_pages, ria->ria_start,
646 /*Check whether the index is in the defined ra-window */
647 static int ras_inside_ra_window(unsigned long idx, struct ra_io_arg *ria)
649 /* If ria_length == ria_pages, it means non-stride I/O mode,
650 * idx should always inside read-ahead window in this case
651 * For stride I/O mode, just check whether the idx is inside
653 return ria->ria_length == 0 || ria->ria_length == ria->ria_pages ||
654 (idx >= ria->ria_stoff && (idx - ria->ria_stoff) %
655 ria->ria_length < ria->ria_pages);
658 static int ll_read_ahead_pages(const struct lu_env *env,
659 struct cl_io *io, struct cl_page_list *queue,
660 struct ra_io_arg *ria,
661 unsigned long *reserved_pages,
662 struct address_space *mapping,
663 unsigned long *ra_end)
665 int rc, count = 0, stride_ria;
666 unsigned long page_idx;
668 LASSERT(ria != NULL);
671 stride_ria = ria->ria_length > ria->ria_pages && ria->ria_pages > 0;
672 for (page_idx = ria->ria_start; page_idx <= ria->ria_end &&
673 *reserved_pages > 0; page_idx++) {
674 if (ras_inside_ra_window(page_idx, ria)) {
675 /* If the page is inside the read-ahead window*/
676 rc = ll_read_ahead_page(env, io, queue,
681 } else if (rc == -ENOLCK)
683 } else if (stride_ria) {
684 /* If it is not in the read-ahead window, and it is
685 * read-ahead mode, then check whether it should skip
688 /* FIXME: This assertion only is valid when it is for
689 * forward read-ahead, it will be fixed when backward
690 * read-ahead is implemented */
691 LASSERTF(page_idx > ria->ria_stoff, "Invalid page_idx %lu"
692 "rs %lu re %lu ro %lu rl %lu rp %lu\n", page_idx,
693 ria->ria_start, ria->ria_end, ria->ria_stoff,
694 ria->ria_length, ria->ria_pages);
695 offset = page_idx - ria->ria_stoff;
696 offset = offset % (ria->ria_length);
697 if (offset > ria->ria_pages) {
698 page_idx += ria->ria_length - offset;
699 CDEBUG(D_READA, "i %lu skip %lu \n", page_idx,
700 ria->ria_length - offset);
709 int ll_readahead(const struct lu_env *env, struct cl_io *io,
710 struct ll_readahead_state *ras, struct address_space *mapping,
711 struct cl_page_list *queue, int flags)
713 struct vvp_io *vio = vvp_env_io(env);
714 struct vvp_thread_info *vti = vvp_env_info(env);
715 struct cl_attr *attr = ccc_env_thread_attr(env);
716 unsigned long start = 0, end = 0, reserved;
717 unsigned long ra_end, len;
719 struct ll_ra_read *bead;
720 struct ra_io_arg *ria = &vti->vti_ria;
721 struct ll_inode_info *lli;
722 struct cl_object *clob;
727 inode = mapping->host;
728 lli = ll_i2info(inode);
729 clob = lli->lli_clob;
731 memset(ria, 0, sizeof *ria);
733 cl_object_attr_lock(clob);
734 ret = cl_object_attr_get(env, clob, attr);
735 cl_object_attr_unlock(clob);
741 ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN);
745 cfs_spin_lock(&ras->ras_lock);
746 if (vio->cui_ra_window_set)
747 bead = &vio->cui_bead;
751 /* Enlarge the RA window to encompass the full read */
752 if (bead != NULL && ras->ras_window_start + ras->ras_window_len <
753 bead->lrr_start + bead->lrr_count) {
754 ras->ras_window_len = bead->lrr_start + bead->lrr_count -
755 ras->ras_window_start;
757 /* Reserve a part of the read-ahead window that we'll be issuing */
758 if (ras->ras_window_len) {
759 start = ras->ras_next_readahead;
760 end = ras->ras_window_start + ras->ras_window_len - 1;
763 unsigned long tmp_end;
765 * Align RA window to an optimal boundary.
767 * XXX This would be better to align to cl_max_pages_per_rpc
768 * instead of PTLRPC_MAX_BRW_PAGES, because the RPC size may
769 * be aligned to the RAID stripe size in the future and that
770 * is more important than the RPC size.
772 tmp_end = ((end + 1) & (~(PTLRPC_MAX_BRW_PAGES - 1))) - 1;
776 /* Truncate RA window to end of file */
777 end = min(end, (unsigned long)((kms - 1) >> CFS_PAGE_SHIFT));
779 ras->ras_next_readahead = max(end, end + 1);
782 ria->ria_start = start;
784 /* If stride I/O mode is detected, get stride window*/
785 if (stride_io_mode(ras)) {
786 ria->ria_stoff = ras->ras_stride_offset;
787 ria->ria_length = ras->ras_stride_length;
788 ria->ria_pages = ras->ras_stride_pages;
790 cfs_spin_unlock(&ras->ras_lock);
793 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
796 len = ria_page_count(ria);
800 reserved = ll_ra_count_get(ll_i2sbi(inode), ria, len);
802 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
804 CDEBUG(D_READA, "reserved page %lu \n", reserved);
806 ret = ll_read_ahead_pages(env, io, queue,
807 ria, &reserved, mapping, &ra_end);
809 LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
811 ll_ra_count_put(ll_i2sbi(inode), reserved);
813 if (ra_end == end + 1 && ra_end == (kms >> CFS_PAGE_SHIFT))
814 ll_ra_stats_inc(mapping, RA_STAT_EOF);
816 /* if we didn't get to the end of the region we reserved from
817 * the ras we need to go back and update the ras so that the
818 * next read-ahead tries from where we left off. we only do so
819 * if the region we failed to issue read-ahead on is still ahead
820 * of the app and behind the next index to start read-ahead from */
821 CDEBUG(D_READA, "ra_end %lu end %lu stride end %lu \n",
822 ra_end, end, ria->ria_end);
824 if (ra_end != end + 1) {
825 cfs_spin_lock(&ras->ras_lock);
826 if (ra_end < ras->ras_next_readahead &&
827 index_in_window(ra_end, ras->ras_window_start, 0,
828 ras->ras_window_len)) {
829 ras->ras_next_readahead = ra_end;
832 cfs_spin_unlock(&ras->ras_lock);
838 static void ras_set_start(struct ll_readahead_state *ras, unsigned long index)
840 ras->ras_window_start = index & (~(RAS_INCREASE_STEP - 1));
843 /* called with the ras_lock held or from places where it doesn't matter */
844 static void ras_reset(struct ll_readahead_state *ras, unsigned long index)
846 ras->ras_last_readpage = index;
847 ras->ras_consecutive_requests = 0;
848 ras->ras_consecutive_pages = 0;
849 ras->ras_window_len = 0;
850 ras_set_start(ras, index);
851 ras->ras_next_readahead = max(ras->ras_window_start, index);
856 /* called with the ras_lock held or from places where it doesn't matter */
857 static void ras_stride_reset(struct ll_readahead_state *ras)
859 ras->ras_consecutive_stride_requests = 0;
860 ras->ras_stride_length = 0;
861 ras->ras_stride_pages = 0;
865 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
867 cfs_spin_lock_init(&ras->ras_lock);
869 ras->ras_requests = 0;
870 CFS_INIT_LIST_HEAD(&ras->ras_read_beads);
874 * Check whether the read request is in the stride window.
875 * If it is in the stride window, return 1, otherwise return 0.
877 static int index_in_stride_window(unsigned long index,
878 struct ll_readahead_state *ras,
881 unsigned long stride_gap = index - ras->ras_last_readpage - 1;
883 if (ras->ras_stride_length == 0 || ras->ras_stride_pages == 0 ||
884 ras->ras_stride_pages == ras->ras_stride_length)
887 /* If it is contiguous read */
889 return ras->ras_consecutive_pages + 1 <= ras->ras_stride_pages;
891 /*Otherwise check the stride by itself */
892 return (ras->ras_stride_length - ras->ras_stride_pages) == stride_gap &&
893 ras->ras_consecutive_pages == ras->ras_stride_pages;
896 static void ras_update_stride_detector(struct ll_readahead_state *ras,
899 unsigned long stride_gap = index - ras->ras_last_readpage - 1;
901 if (!stride_io_mode(ras) && (stride_gap != 0 ||
902 ras->ras_consecutive_stride_requests == 0)) {
903 ras->ras_stride_pages = ras->ras_consecutive_pages;
904 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
906 LASSERT(ras->ras_request_index == 0);
907 LASSERT(ras->ras_consecutive_stride_requests == 0);
909 if (index <= ras->ras_last_readpage) {
910 /*Reset stride window for forward read*/
911 ras_stride_reset(ras);
915 ras->ras_stride_pages = ras->ras_consecutive_pages;
916 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
923 stride_page_count(struct ll_readahead_state *ras, unsigned long len)
925 return stride_pg_count(ras->ras_stride_offset, ras->ras_stride_length,
926 ras->ras_stride_pages, ras->ras_stride_offset,
930 /* Stride Read-ahead window will be increased inc_len according to
931 * stride I/O pattern */
932 static void ras_stride_increase_window(struct ll_readahead_state *ras,
933 struct ll_ra_info *ra,
934 unsigned long inc_len)
936 unsigned long left, step, window_len;
937 unsigned long stride_len;
939 LASSERT(ras->ras_stride_length > 0);
940 LASSERTF(ras->ras_window_start + ras->ras_window_len
941 >= ras->ras_stride_offset, "window_start %lu, window_len %lu"
942 " stride_offset %lu\n", ras->ras_window_start,
943 ras->ras_window_len, ras->ras_stride_offset);
945 stride_len = ras->ras_window_start + ras->ras_window_len -
946 ras->ras_stride_offset;
948 left = stride_len % ras->ras_stride_length;
949 window_len = ras->ras_window_len - left;
951 if (left < ras->ras_stride_pages)
954 left = ras->ras_stride_pages + inc_len;
956 LASSERT(ras->ras_stride_pages != 0);
958 step = left / ras->ras_stride_pages;
959 left %= ras->ras_stride_pages;
961 window_len += step * ras->ras_stride_length + left;
963 if (stride_page_count(ras, window_len) <= ra->ra_max_pages_per_file)
964 ras->ras_window_len = window_len;
969 static void ras_increase_window(struct ll_readahead_state *ras,
970 struct ll_ra_info *ra, struct inode *inode)
972 /* The stretch of ra-window should be aligned with max rpc_size
973 * but current clio architecture does not support retrieve such
974 * information from lower layer. FIXME later
976 if (stride_io_mode(ras))
977 ras_stride_increase_window(ras, ra, RAS_INCREASE_STEP);
979 ras->ras_window_len = min(ras->ras_window_len +
981 ra->ra_max_pages_per_file);
984 void ras_update(struct ll_sb_info *sbi, struct inode *inode,
985 struct ll_readahead_state *ras, unsigned long index,
988 struct ll_ra_info *ra = &sbi->ll_ra_info;
989 int zero = 0, stride_detect = 0, ra_miss = 0;
992 cfs_spin_lock(&ras->ras_lock);
994 ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
996 /* reset the read-ahead window in two cases. First when the app seeks
997 * or reads to some other part of the file. Secondly if we get a
998 * read-ahead miss that we think we've previously issued. This can
999 * be a symptom of there being so many read-ahead pages that the VM is
1000 * reclaiming it before we get to it. */
1001 if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
1003 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
1004 } else if (!hit && ras->ras_window_len &&
1005 index < ras->ras_next_readahead &&
1006 index_in_window(index, ras->ras_window_start, 0,
1007 ras->ras_window_len)) {
1009 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
1012 /* On the second access to a file smaller than the tunable
1013 * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1014 * file up to ra_max_pages_per_file. This is simply a best effort
1015 * and only occurs once per open file. Normal RA behavior is reverted
1016 * to for subsequent IO. The mmap case does not increment
1017 * ras_requests and thus can never trigger this behavior. */
1018 if (ras->ras_requests == 2 && !ras->ras_request_index) {
1021 kms_pages = (i_size_read(inode) + CFS_PAGE_SIZE - 1) >>
1024 CDEBUG(D_READA, "kmsp "LPU64" mwp %lu mp %lu\n", kms_pages,
1025 ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages_per_file);
1028 kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1029 ras->ras_window_start = 0;
1030 ras->ras_last_readpage = 0;
1031 ras->ras_next_readahead = 0;
1032 ras->ras_window_len = min(ra->ra_max_pages_per_file,
1033 ra->ra_max_read_ahead_whole_pages);
1034 GOTO(out_unlock, 0);
1038 /* check whether it is in stride I/O mode*/
1039 if (!index_in_stride_window(index, ras, inode)) {
1040 if (ras->ras_consecutive_stride_requests == 0 &&
1041 ras->ras_request_index == 0) {
1042 ras_update_stride_detector(ras, index);
1043 ras->ras_consecutive_stride_requests ++;
1045 ras_stride_reset(ras);
1047 ras_reset(ras, index);
1048 ras->ras_consecutive_pages++;
1049 GOTO(out_unlock, 0);
1051 ras->ras_consecutive_pages = 0;
1052 ras->ras_consecutive_requests = 0;
1053 if (++ras->ras_consecutive_stride_requests > 1)
1059 if (index_in_stride_window(index, ras, inode) &&
1060 stride_io_mode(ras)) {
1061 /*If stride-RA hit cache miss, the stride dector
1062 *will not be reset to avoid the overhead of
1063 *redetecting read-ahead mode */
1064 if (index != ras->ras_last_readpage + 1)
1065 ras->ras_consecutive_pages = 0;
1066 ras_reset(ras, index);
1069 /* Reset both stride window and normal RA
1071 ras_reset(ras, index);
1072 ras->ras_consecutive_pages++;
1073 ras_stride_reset(ras);
1074 GOTO(out_unlock, 0);
1076 } else if (stride_io_mode(ras)) {
1077 /* If this is contiguous read but in stride I/O mode
1078 * currently, check whether stride step still is valid,
1079 * if invalid, it will reset the stride ra window*/
1080 if (!index_in_stride_window(index, ras, inode)) {
1081 /* Shrink stride read-ahead window to be zero */
1082 ras_stride_reset(ras);
1083 ras->ras_window_len = 0;
1084 ras->ras_next_readahead = index;
1088 ras->ras_consecutive_pages++;
1089 ras->ras_last_readpage = index;
1090 ras_set_start(ras, index);
1092 if (stride_io_mode(ras))
1093 /* Since stride readahead is sentivite to the offset
1094 * of read-ahead, so we use original offset here,
1095 * instead of ras_window_start, which is 1M aligned*/
1096 ras->ras_next_readahead = max(index,
1097 ras->ras_next_readahead);
1099 ras->ras_next_readahead = max(ras->ras_window_start,
1100 ras->ras_next_readahead);
1103 /* Trigger RA in the mmap case where ras_consecutive_requests
1104 * is not incremented and thus can't be used to trigger RA */
1105 if (!ras->ras_window_len && ras->ras_consecutive_pages == 4) {
1106 ras->ras_window_len = RAS_INCREASE_STEP;
1107 GOTO(out_unlock, 0);
1110 /* Initially reset the stride window offset to next_readahead*/
1111 if (ras->ras_consecutive_stride_requests == 2 && stride_detect) {
1113 * Once stride IO mode is detected, next_readahead should be
1114 * reset to make sure next_readahead > stride offset
1116 ras->ras_next_readahead = max(index, ras->ras_next_readahead);
1117 ras->ras_stride_offset = index;
1118 ras->ras_window_len = RAS_INCREASE_STEP;
1121 /* The initial ras_window_len is set to the request size. To avoid
1122 * uselessly reading and discarding pages for random IO the window is
1123 * only increased once per consecutive request received. */
1124 if ((ras->ras_consecutive_requests > 1 || stride_detect) &&
1125 !ras->ras_request_index)
1126 ras_increase_window(ras, ra, inode);
1130 ras->ras_request_index++;
1131 cfs_spin_unlock(&ras->ras_lock);
1135 int ll_writepage(struct page *vmpage, struct writeback_control *unused)
1137 struct inode *inode = vmpage->mapping->host;
1140 struct cl_page *page;
1141 struct cl_object *clob;
1142 struct cl_2queue *queue;
1143 struct cl_env_nest nest;
1147 LASSERT(PageLocked(vmpage));
1148 LASSERT(!PageWriteback(vmpage));
1150 if (ll_i2dtexp(inode) == NULL)
1153 env = cl_env_nested_get(&nest);
1155 RETURN(PTR_ERR(env));
1157 queue = &vvp_env_info(env)->vti_queue;
1158 clob = ll_i2info(inode)->lli_clob;
1159 LASSERT(clob != NULL);
1161 io = ccc_env_thread_io(env);
1163 result = cl_io_init(env, io, CIT_MISC, clob);
1165 page = cl_page_find(env, clob, vmpage->index,
1166 vmpage, CPT_CACHEABLE);
1167 if (!IS_ERR(page)) {
1168 lu_ref_add(&page->cp_reference, "writepage",
1170 cl_page_assume(env, io, page);
1172 * Mark page dirty, because this is what
1173 * ->vio_submit()->cpo_prep_write() assumes.
1175 * XXX better solution is to detect this from within
1176 * cl_io_submit_rw() somehow.
1178 set_page_dirty(vmpage);
1179 cl_2queue_init_page(queue, page);
1180 result = cl_io_submit_rw(env, io, CRT_WRITE,
1182 cl_page_list_disown(env, io, &queue->c2_qin);
1185 * There is no need to clear PG_writeback, as
1186 * cl_io_submit_rw() calls completion callback
1190 * Re-dirty page on error so it retries write,
1191 * but not in case when IO has actually
1192 * occurred and completed with an error.
1194 if (!PageError(vmpage))
1195 set_page_dirty(vmpage);
1197 LASSERT(!cl_page_is_owned(page, io));
1198 lu_ref_del(&page->cp_reference,
1199 "writepage", cfs_current());
1200 cl_page_put(env, page);
1201 cl_2queue_fini(env, queue);
1204 cl_io_fini(env, io);
1205 cl_env_nested_put(&nest, env);
1209 int ll_readpage(struct file *file, struct page *vmpage)
1211 struct ll_cl_context *lcc;
1215 lcc = ll_cl_init(file, vmpage, 0);
1217 struct lu_env *env = lcc->lcc_env;
1218 struct cl_io *io = lcc->lcc_io;
1219 struct cl_page *page = lcc->lcc_page;
1221 LASSERT(page->cp_type == CPT_CACHEABLE);
1222 if (likely(!PageUptodate(vmpage))) {
1223 cl_page_assume(env, io, page);
1224 result = cl_io_read_page(env, io, page);
1226 /* Page from a non-object file. */
1227 LASSERT(!ll_i2info(vmpage->mapping->host)->lli_smd);
1228 unlock_page(vmpage);
1233 unlock_page(vmpage);
1234 result = PTR_ERR(lcc);