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/kernel.h>
43 #include <linux/string.h>
44 #include <linux/stat.h>
45 #include <linux/errno.h>
46 #include <linux/smp_lock.h>
47 #include <linux/unistd.h>
48 #include <linux/version.h>
49 #include <asm/system.h>
50 #include <asm/uaccess.h>
53 #include <linux/stat.h>
54 #include <asm/uaccess.h>
56 #include <linux/pagemap.h>
57 #include <linux/smp_lock.h>
58 /* current_is_kswapd() */
59 #include <linux/swap.h>
61 #define DEBUG_SUBSYSTEM S_LLITE
63 #include <lustre_lite.h>
64 #include <obd_cksum.h>
65 #include "llite_internal.h"
66 #include <linux/lustre_compat25.h>
68 /* this isn't where truncate starts. roughly:
69 * sys_truncate->ll_setattr_raw->vmtruncate->ll_truncate. setattr_raw grabs
70 * DLM lock on [size, EOF], i_mutex, ->lli_size_sem, and WRITE_I_ALLOC_SEM to
73 * must be called under ->lli_size_sem */
74 void ll_truncate(struct inode *inode)
76 struct ll_inode_info *lli = ll_i2info(inode);
79 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) to %llu\n", inode->i_ino,
80 inode->i_generation, inode, i_size_read(inode));
82 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_TRUNC, 1);
83 if (lli->lli_size_sem_owner == cfs_current()) {
84 LASSERT_SEM_LOCKED(&lli->lli_size_sem);
85 ll_inode_size_unlock(inode, 0);
93 * Finalizes cl-data before exiting typical address_space operation. Dual to
96 static void ll_cl_fini(struct ll_cl_context *lcc)
98 struct lu_env *env = lcc->lcc_env;
99 struct cl_io *io = lcc->lcc_io;
100 struct cl_page *page = lcc->lcc_page;
102 LASSERT(lcc->lcc_cookie == current);
103 LASSERT(env != NULL);
106 lu_ref_del(&page->cp_reference, "cl_io", io);
107 cl_page_put(env, page);
110 if (io && lcc->lcc_created) {
112 cl_io_unlock(env, io);
113 cl_io_iter_fini(env, io);
116 cl_env_put(env, &lcc->lcc_refcheck);
120 * Initializes common cl-data at the typical address_space operation entry
123 static struct ll_cl_context *ll_cl_init(struct file *file,
124 struct page *vmpage, int create)
126 struct ll_cl_context *lcc;
129 struct cl_object *clob;
135 clob = ll_i2info(vmpage->mapping->host)->lli_clob;
136 LASSERT(clob != NULL);
138 env = cl_env_get(&refcheck);
140 return ERR_PTR(PTR_ERR(env));
142 lcc = &vvp_env_info(env)->vti_io_ctx;
143 memset(lcc, 0, sizeof(*lcc));
145 lcc->lcc_refcheck = refcheck;
146 lcc->lcc_cookie = current;
148 cio = ccc_env_io(env);
149 io = cio->cui_cl.cis_io;
150 if (io == NULL && create) {
155 * Loop-back driver calls ->prepare_write() and ->sendfile()
156 * methods directly, bypassing file system ->write() operation,
157 * so cl_io has to be created here.
160 io = ccc_env_thread_io(env);
161 vio = vvp_env_io(env);
162 ll_io_init(io, file, 1);
164 /* No lock at all for this kind of IO - we can't do it because
165 * we have held page lock, it would cause deadlock.
166 * XXX: This causes poor performance to loop device - One page
168 * In order to get better performance, users should use
169 * lloop driver instead.
171 io->ci_lockreq = CILR_NEVER;
173 pos = (vmpage->index << CFS_PAGE_SHIFT);
175 /* Create a temp IO to serve write. */
176 result = cl_io_rw_init(env, io, CIT_WRITE, pos, CFS_PAGE_SIZE);
178 cio->cui_fd = LUSTRE_FPRIVATE(file);
181 result = cl_io_iter_init(env, io);
183 result = cl_io_lock(env, io);
185 result = cl_io_start(env, io);
188 result = io->ci_result;
189 lcc->lcc_created = 1;
196 struct cl_page *page;
199 LASSERT(io->ci_state == CIS_IO_GOING);
200 LASSERT(cio->cui_fd == LUSTRE_FPRIVATE(file));
201 page = cl_page_find(env, clob, vmpage->index, vmpage,
204 lcc->lcc_page = page;
205 lu_ref_add(&page->cp_reference, "cl_io", io);
208 result = PTR_ERR(page);
212 lcc = ERR_PTR(result);
215 CDEBUG(D_VFSTRACE, "%lu@"DFID" -> %d %p %p\n",
216 vmpage->index, PFID(lu_object_fid(&clob->co_lu)), result,
221 static struct ll_cl_context *ll_cl_get(void)
223 struct ll_cl_context *lcc;
227 env = cl_env_get(&refcheck);
228 LASSERT(!IS_ERR(env));
229 lcc = &vvp_env_info(env)->vti_io_ctx;
230 LASSERT(env == lcc->lcc_env);
231 LASSERT(current == lcc->lcc_cookie);
232 cl_env_put(env, &refcheck);
234 /* env has got in ll_cl_init, so it is still usable. */
239 * ->prepare_write() address space operation called by generic_file_write()
240 * for every page during write.
242 int ll_prepare_write(struct file *file, struct page *vmpage, unsigned from,
245 struct ll_cl_context *lcc;
249 lcc = ll_cl_init(file, vmpage, 1);
251 struct lu_env *env = lcc->lcc_env;
252 struct cl_io *io = lcc->lcc_io;
253 struct cl_page *page = lcc->lcc_page;
255 cl_page_assume(env, io, page);
256 if (cl_io_is_append(io)) {
257 struct cl_object *obj = io->ci_obj;
258 struct inode *inode = ccc_object_inode(obj);
260 * In VFS file->page write loop, for appending, the
261 * write offset might be reset according to the new
262 * file size before holding i_mutex. So crw_pos should
263 * be reset here. BUG:17711.
265 io->u.ci_wr.wr.crw_pos = i_size_read(inode);
267 result = cl_io_prepare_write(env, io, page, from, to);
270 * Add a reference, so that page is not evicted from
271 * the cache until ->commit_write() is called.
274 lu_ref_add(&page->cp_reference, "prepare_write",
277 cl_page_unassume(env, io, page);
280 /* returning 0 in prepare assumes commit must be called
283 result = PTR_ERR(lcc);
288 int ll_commit_write(struct file *file, struct page *vmpage, unsigned from,
291 struct ll_cl_context *lcc;
294 struct cl_page *page;
300 page = lcc->lcc_page;
303 LASSERT(cl_page_is_owned(page, io));
305 if (from != to) /* handle short write case. */
306 result = cl_io_commit_write(env, io, page, from, to);
307 if (cl_page_is_owned(page, io))
308 cl_page_unassume(env, io, page);
311 * Release reference acquired by ll_prepare_write().
313 lu_ref_del(&page->cp_reference, "prepare_write", cfs_current());
314 cl_page_put(env, page);
319 struct obd_capa *cl_capa_lookup(struct inode *inode, enum cl_req_type crt)
323 opc = crt == CRT_WRITE ? CAPA_OPC_OSS_WRITE : CAPA_OPC_OSS_RW;
324 return ll_osscapa_get(inode, opc);
327 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which);
329 /* WARNING: This algorithm is used to reduce the contention on
330 * sbi->ll_lock. It should work well if the ra_max_pages is much
331 * greater than the single file's read-ahead window.
333 * TODO: There may exist a `global sync problem' in this implementation.
334 * Considering the global ra window is 100M, and each file's ra window is 10M,
335 * there are over 10 files trying to get its ra budget and reach
336 * ll_ra_count_get at the exactly same time. All of them will get a zero ra
337 * window, although the global window is 100M. -jay
339 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, struct ra_io_arg *ria,
342 struct ll_ra_info *ra = &sbi->ll_ra_info;
347 * If read-ahead pages left are less than 1M, do not do read-ahead,
348 * otherwise it will form small read RPC(< 1M), which hurt server
351 if (ra->ra_max_pages < atomic_read(&ra->ra_cur_pages))
354 ret = min(ra->ra_max_pages - cfs_atomic_read(&ra->ra_cur_pages), len);
355 if ((int)ret < 0 || ret < min((unsigned long)PTLRPC_MAX_BRW_PAGES, len))
358 if (ria->ria_pages == 0)
359 /* it needs 1M align again after trimed by ra_max_pages*/
360 if (ret >= ((ria->ria_start + ret) % PTLRPC_MAX_BRW_PAGES))
361 ret -= (ria->ria_start + ret) % PTLRPC_MAX_BRW_PAGES;
363 if (cfs_atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) {
364 cfs_atomic_sub(ret, &ra->ra_cur_pages);
372 void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
374 struct ll_ra_info *ra = &sbi->ll_ra_info;
375 cfs_atomic_sub(len, &ra->ra_cur_pages);
378 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which)
380 LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
381 lprocfs_counter_incr(sbi->ll_ra_stats, which);
384 void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which)
386 struct ll_sb_info *sbi = ll_i2sbi(mapping->host);
387 ll_ra_stats_inc_sbi(sbi, which);
390 #define RAS_CDEBUG(ras) \
392 "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu r %lu ri %lu" \
393 "csr %lu sf %lu sp %lu sl %lu \n", \
394 ras->ras_last_readpage, ras->ras_consecutive_requests, \
395 ras->ras_consecutive_pages, ras->ras_window_start, \
396 ras->ras_window_len, ras->ras_next_readahead, \
397 ras->ras_requests, ras->ras_request_index, \
398 ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
399 ras->ras_stride_pages, ras->ras_stride_length)
401 static int index_in_window(unsigned long index, unsigned long point,
402 unsigned long before, unsigned long after)
404 unsigned long start = point - before, end = point + after;
411 return start <= index && index <= end;
414 static struct ll_readahead_state *ll_ras_get(struct file *f)
416 struct ll_file_data *fd;
418 fd = LUSTRE_FPRIVATE(f);
422 void ll_ra_read_in(struct file *f, struct ll_ra_read *rar)
424 struct ll_readahead_state *ras;
428 cfs_spin_lock(&ras->ras_lock);
430 ras->ras_request_index = 0;
431 ras->ras_consecutive_requests++;
432 rar->lrr_reader = current;
434 cfs_list_add(&rar->lrr_linkage, &ras->ras_read_beads);
435 cfs_spin_unlock(&ras->ras_lock);
438 void ll_ra_read_ex(struct file *f, struct ll_ra_read *rar)
440 struct ll_readahead_state *ras;
444 cfs_spin_lock(&ras->ras_lock);
445 cfs_list_del_init(&rar->lrr_linkage);
446 cfs_spin_unlock(&ras->ras_lock);
449 static struct ll_ra_read *ll_ra_read_get_locked(struct ll_readahead_state *ras)
451 struct ll_ra_read *scan;
453 cfs_list_for_each_entry(scan, &ras->ras_read_beads, lrr_linkage) {
454 if (scan->lrr_reader == current)
460 struct ll_ra_read *ll_ra_read_get(struct file *f)
462 struct ll_readahead_state *ras;
463 struct ll_ra_read *bead;
467 cfs_spin_lock(&ras->ras_lock);
468 bead = ll_ra_read_get_locked(ras);
469 cfs_spin_unlock(&ras->ras_lock);
473 static int cl_read_ahead_page(const struct lu_env *env, struct cl_io *io,
474 struct cl_page_list *queue, struct cl_page *page,
483 cl_page_assume(env, io, page);
484 lu_ref_add(&page->cp_reference, "ra", cfs_current());
485 cp = cl2ccc_page(cl_page_at(page, &vvp_device_type));
486 if (!cp->cpg_defer_uptodate && !Page_Uptodate(vmpage)) {
487 rc = cl_page_is_under_lock(env, io, page);
489 cp->cpg_defer_uptodate = 1;
491 cl_page_list_add(queue, page);
494 cl_page_delete(env, page);
498 /* skip completed pages */
499 cl_page_unassume(env, io, page);
500 lu_ref_del(&page->cp_reference, "ra", cfs_current());
501 cl_page_put(env, page);
506 * Initiates read-ahead of a page with given index.
508 * \retval +ve: page was added to \a queue.
510 * \retval -ENOLCK: there is no extent lock for this part of a file, stop
513 * \retval -ve, 0: page wasn't added to \a queue for other reason.
515 static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io,
516 struct cl_page_list *queue,
517 pgoff_t index, struct address_space *mapping)
520 struct cl_object *clob = ll_i2info(mapping->host)->lli_clob;
521 struct cl_page *page;
522 enum ra_stat which = _NR_RA_STAT; /* keep gcc happy */
523 unsigned int gfp_mask;
525 const char *msg = NULL;
529 gfp_mask = GFP_HIGHUSER & ~__GFP_WAIT;
531 gfp_mask |= __GFP_NOWARN;
533 vmpage = grab_cache_page_nowait_gfp(mapping, index, gfp_mask);
534 if (vmpage != NULL) {
535 /* Check if vmpage was truncated or reclaimed */
536 if (vmpage->mapping == mapping) {
537 page = cl_page_find(env, clob, vmpage->index,
538 vmpage, CPT_CACHEABLE);
540 rc = cl_read_ahead_page(env, io, queue,
543 which = RA_STAT_FAILED_MATCH;
544 msg = "lock match failed";
547 which = RA_STAT_FAILED_GRAB_PAGE;
548 msg = "cl_page_find failed";
551 which = RA_STAT_WRONG_GRAB_PAGE;
552 msg = "g_c_p_n returned invalid page";
556 page_cache_release(vmpage);
558 which = RA_STAT_FAILED_GRAB_PAGE;
559 msg = "g_c_p_n failed";
562 ll_ra_stats_inc(mapping, which);
563 CDEBUG(D_READA, "%s\n", msg);
568 #define RIA_DEBUG(ria) \
569 CDEBUG(D_READA, "rs %lu re %lu ro %lu rl %lu rp %lu\n", \
570 ria->ria_start, ria->ria_end, ria->ria_stoff, ria->ria_length,\
573 #define RAS_INCREASE_STEP PTLRPC_MAX_BRW_PAGES
575 static inline int stride_io_mode(struct ll_readahead_state *ras)
577 return ras->ras_consecutive_stride_requests > 1;
579 /* The function calculates how much pages will be read in
580 * [off, off + length], in such stride IO area,
581 * stride_offset = st_off, stride_lengh = st_len,
582 * stride_pages = st_pgs
584 * |------------------|*****|------------------|*****|------------|*****|....
587 * |----- st_len -----|
589 * How many pages it should read in such pattern
590 * |-------------------------------------------------------------|
592 * |<------ length ------->|
594 * = |<----->| + |-------------------------------------| + |---|
595 * start_left st_pgs * i end_left
598 stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs,
599 unsigned long off, unsigned long length)
601 __u64 start = off > st_off ? off - st_off : 0;
602 __u64 end = off + length > st_off ? off + length - st_off : 0;
603 unsigned long start_left = 0;
604 unsigned long end_left = 0;
605 unsigned long pg_count;
607 if (st_len == 0 || length == 0 || end == 0)
610 start_left = do_div(start, st_len);
611 if (start_left < st_pgs)
612 start_left = st_pgs - start_left;
616 end_left = do_div(end, st_len);
617 if (end_left > st_pgs)
620 CDEBUG(D_READA, "start "LPU64", end "LPU64" start_left %lu end_left %lu \n",
621 start, end, start_left, end_left);
624 pg_count = end_left - (st_pgs - start_left);
626 pg_count = start_left + st_pgs * (end - start - 1) + end_left;
628 CDEBUG(D_READA, "st_off %lu, st_len %lu st_pgs %lu off %lu length %lu"
629 "pgcount %lu\n", st_off, st_len, st_pgs, off, length, pg_count);
634 static int ria_page_count(struct ra_io_arg *ria)
636 __u64 length = ria->ria_end >= ria->ria_start ?
637 ria->ria_end - ria->ria_start + 1 : 0;
639 return stride_pg_count(ria->ria_stoff, ria->ria_length,
640 ria->ria_pages, ria->ria_start,
644 /*Check whether the index is in the defined ra-window */
645 static int ras_inside_ra_window(unsigned long idx, struct ra_io_arg *ria)
647 /* If ria_length == ria_pages, it means non-stride I/O mode,
648 * idx should always inside read-ahead window in this case
649 * For stride I/O mode, just check whether the idx is inside
651 return ria->ria_length == 0 || ria->ria_length == ria->ria_pages ||
652 (idx >= ria->ria_stoff && (idx - ria->ria_stoff) %
653 ria->ria_length < ria->ria_pages);
656 static int ll_read_ahead_pages(const struct lu_env *env,
657 struct cl_io *io, struct cl_page_list *queue,
658 struct ra_io_arg *ria,
659 unsigned long *reserved_pages,
660 struct address_space *mapping,
661 unsigned long *ra_end)
663 int rc, count = 0, stride_ria;
664 unsigned long page_idx;
666 LASSERT(ria != NULL);
669 stride_ria = ria->ria_length > ria->ria_pages && ria->ria_pages > 0;
670 for (page_idx = ria->ria_start; page_idx <= ria->ria_end &&
671 *reserved_pages > 0; page_idx++) {
672 if (ras_inside_ra_window(page_idx, ria)) {
673 /* If the page is inside the read-ahead window*/
674 rc = ll_read_ahead_page(env, io, queue,
679 } else if (rc == -ENOLCK)
681 } else if (stride_ria) {
682 /* If it is not in the read-ahead window, and it is
683 * read-ahead mode, then check whether it should skip
686 /* FIXME: This assertion only is valid when it is for
687 * forward read-ahead, it will be fixed when backward
688 * read-ahead is implemented */
689 LASSERTF(page_idx > ria->ria_stoff, "Invalid page_idx %lu"
690 "rs %lu re %lu ro %lu rl %lu rp %lu\n", page_idx,
691 ria->ria_start, ria->ria_end, ria->ria_stoff,
692 ria->ria_length, ria->ria_pages);
693 offset = page_idx - ria->ria_stoff;
694 offset = offset % (ria->ria_length);
695 if (offset > ria->ria_pages) {
696 page_idx += ria->ria_length - offset;
697 CDEBUG(D_READA, "i %lu skip %lu \n", page_idx,
698 ria->ria_length - offset);
707 int ll_readahead(const struct lu_env *env, struct cl_io *io,
708 struct ll_readahead_state *ras, struct address_space *mapping,
709 struct cl_page_list *queue, int flags)
711 struct vvp_io *vio = vvp_env_io(env);
712 struct vvp_thread_info *vti = vvp_env_info(env);
713 struct cl_attr *attr = ccc_env_thread_attr(env);
714 unsigned long start = 0, end = 0, reserved;
715 unsigned long ra_end, len;
717 struct ll_ra_read *bead;
718 struct ra_io_arg *ria = &vti->vti_ria;
719 struct ll_inode_info *lli;
720 struct cl_object *clob;
725 inode = mapping->host;
726 lli = ll_i2info(inode);
727 clob = lli->lli_clob;
729 memset(ria, 0, sizeof *ria);
731 cl_object_attr_lock(clob);
732 ret = cl_object_attr_get(env, clob, attr);
733 cl_object_attr_unlock(clob);
739 ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN);
743 cfs_spin_lock(&ras->ras_lock);
744 if (vio->cui_ra_window_set)
745 bead = &vio->cui_bead;
749 /* Enlarge the RA window to encompass the full read */
750 if (bead != NULL && ras->ras_window_start + ras->ras_window_len <
751 bead->lrr_start + bead->lrr_count) {
752 ras->ras_window_len = bead->lrr_start + bead->lrr_count -
753 ras->ras_window_start;
755 /* Reserve a part of the read-ahead window that we'll be issuing */
756 if (ras->ras_window_len) {
757 start = ras->ras_next_readahead;
758 end = ras->ras_window_start + ras->ras_window_len - 1;
761 unsigned long tmp_end;
763 * Align RA window to an optimal boundary.
765 * XXX This would be better to align to cl_max_pages_per_rpc
766 * instead of PTLRPC_MAX_BRW_PAGES, because the RPC size may
767 * be aligned to the RAID stripe size in the future and that
768 * is more important than the RPC size.
770 tmp_end = ((end + 1) & (~(PTLRPC_MAX_BRW_PAGES - 1))) - 1;
774 /* Truncate RA window to end of file */
775 end = min(end, (unsigned long)((kms - 1) >> CFS_PAGE_SHIFT));
777 ras->ras_next_readahead = max(end, end + 1);
780 ria->ria_start = start;
782 /* If stride I/O mode is detected, get stride window*/
783 if (stride_io_mode(ras)) {
784 ria->ria_stoff = ras->ras_stride_offset;
785 ria->ria_length = ras->ras_stride_length;
786 ria->ria_pages = ras->ras_stride_pages;
788 cfs_spin_unlock(&ras->ras_lock);
791 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
794 len = ria_page_count(ria);
798 reserved = ll_ra_count_get(ll_i2sbi(inode), ria, len);
800 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
802 CDEBUG(D_READA, "reserved page %lu \n", reserved);
804 ret = ll_read_ahead_pages(env, io, queue,
805 ria, &reserved, mapping, &ra_end);
807 LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
809 ll_ra_count_put(ll_i2sbi(inode), reserved);
811 if (ra_end == end + 1 && ra_end == (kms >> CFS_PAGE_SHIFT))
812 ll_ra_stats_inc(mapping, RA_STAT_EOF);
814 /* if we didn't get to the end of the region we reserved from
815 * the ras we need to go back and update the ras so that the
816 * next read-ahead tries from where we left off. we only do so
817 * if the region we failed to issue read-ahead on is still ahead
818 * of the app and behind the next index to start read-ahead from */
819 CDEBUG(D_READA, "ra_end %lu end %lu stride end %lu \n",
820 ra_end, end, ria->ria_end);
822 if (ra_end != end + 1) {
823 cfs_spin_lock(&ras->ras_lock);
824 if (ra_end < ras->ras_next_readahead &&
825 index_in_window(ra_end, ras->ras_window_start, 0,
826 ras->ras_window_len)) {
827 ras->ras_next_readahead = ra_end;
830 cfs_spin_unlock(&ras->ras_lock);
836 static void ras_set_start(struct ll_readahead_state *ras, unsigned long index)
838 ras->ras_window_start = index & (~(RAS_INCREASE_STEP - 1));
841 /* called with the ras_lock held or from places where it doesn't matter */
842 static void ras_reset(struct ll_readahead_state *ras, unsigned long index)
844 ras->ras_last_readpage = index;
845 ras->ras_consecutive_requests = 0;
846 ras->ras_consecutive_pages = 0;
847 ras->ras_window_len = 0;
848 ras_set_start(ras, index);
849 ras->ras_next_readahead = max(ras->ras_window_start, index);
854 /* called with the ras_lock held or from places where it doesn't matter */
855 static void ras_stride_reset(struct ll_readahead_state *ras)
857 ras->ras_consecutive_stride_requests = 0;
858 ras->ras_stride_length = 0;
859 ras->ras_stride_pages = 0;
863 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
865 cfs_spin_lock_init(&ras->ras_lock);
867 ras->ras_requests = 0;
868 CFS_INIT_LIST_HEAD(&ras->ras_read_beads);
872 * Check whether the read request is in the stride window.
873 * If it is in the stride window, return 1, otherwise return 0.
875 static int index_in_stride_window(unsigned long index,
876 struct ll_readahead_state *ras,
879 unsigned long stride_gap = index - ras->ras_last_readpage - 1;
881 if (ras->ras_stride_length == 0 || ras->ras_stride_pages == 0 ||
882 ras->ras_stride_pages == ras->ras_stride_length)
885 /* If it is contiguous read */
887 return ras->ras_consecutive_pages + 1 <= ras->ras_stride_pages;
889 /*Otherwise check the stride by itself */
890 return (ras->ras_stride_length - ras->ras_stride_pages) == stride_gap &&
891 ras->ras_consecutive_pages == ras->ras_stride_pages;
894 static void ras_update_stride_detector(struct ll_readahead_state *ras,
897 unsigned long stride_gap = index - ras->ras_last_readpage - 1;
899 if (!stride_io_mode(ras) && (stride_gap != 0 ||
900 ras->ras_consecutive_stride_requests == 0)) {
901 ras->ras_stride_pages = ras->ras_consecutive_pages;
902 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
904 LASSERT(ras->ras_request_index == 0);
905 LASSERT(ras->ras_consecutive_stride_requests == 0);
907 if (index <= ras->ras_last_readpage) {
908 /*Reset stride window for forward read*/
909 ras_stride_reset(ras);
913 ras->ras_stride_pages = ras->ras_consecutive_pages;
914 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
921 stride_page_count(struct ll_readahead_state *ras, unsigned long len)
923 return stride_pg_count(ras->ras_stride_offset, ras->ras_stride_length,
924 ras->ras_stride_pages, ras->ras_stride_offset,
928 /* Stride Read-ahead window will be increased inc_len according to
929 * stride I/O pattern */
930 static void ras_stride_increase_window(struct ll_readahead_state *ras,
931 struct ll_ra_info *ra,
932 unsigned long inc_len)
934 unsigned long left, step, window_len;
935 unsigned long stride_len;
937 LASSERT(ras->ras_stride_length > 0);
938 LASSERTF(ras->ras_window_start + ras->ras_window_len
939 >= ras->ras_stride_offset, "window_start %lu, window_len %lu"
940 " stride_offset %lu\n", ras->ras_window_start,
941 ras->ras_window_len, ras->ras_stride_offset);
943 stride_len = ras->ras_window_start + ras->ras_window_len -
944 ras->ras_stride_offset;
946 left = stride_len % ras->ras_stride_length;
947 window_len = ras->ras_window_len - left;
949 if (left < ras->ras_stride_pages)
952 left = ras->ras_stride_pages + inc_len;
954 LASSERT(ras->ras_stride_pages != 0);
956 step = left / ras->ras_stride_pages;
957 left %= ras->ras_stride_pages;
959 window_len += step * ras->ras_stride_length + left;
961 if (stride_page_count(ras, window_len) <= ra->ra_max_pages_per_file)
962 ras->ras_window_len = window_len;
967 static void ras_increase_window(struct ll_readahead_state *ras,
968 struct ll_ra_info *ra, struct inode *inode)
970 /* The stretch of ra-window should be aligned with max rpc_size
971 * but current clio architecture does not support retrieve such
972 * information from lower layer. FIXME later
974 if (stride_io_mode(ras))
975 ras_stride_increase_window(ras, ra, RAS_INCREASE_STEP);
977 ras->ras_window_len = min(ras->ras_window_len +
979 ra->ra_max_pages_per_file);
982 void ras_update(struct ll_sb_info *sbi, struct inode *inode,
983 struct ll_readahead_state *ras, unsigned long index,
986 struct ll_ra_info *ra = &sbi->ll_ra_info;
987 int zero = 0, stride_detect = 0, ra_miss = 0;
990 cfs_spin_lock(&ras->ras_lock);
992 ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
994 /* reset the read-ahead window in two cases. First when the app seeks
995 * or reads to some other part of the file. Secondly if we get a
996 * read-ahead miss that we think we've previously issued. This can
997 * be a symptom of there being so many read-ahead pages that the VM is
998 * reclaiming it before we get to it. */
999 if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
1001 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
1002 } else if (!hit && ras->ras_window_len &&
1003 index < ras->ras_next_readahead &&
1004 index_in_window(index, ras->ras_window_start, 0,
1005 ras->ras_window_len)) {
1007 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
1010 /* On the second access to a file smaller than the tunable
1011 * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1012 * file up to ra_max_pages_per_file. This is simply a best effort
1013 * and only occurs once per open file. Normal RA behavior is reverted
1014 * to for subsequent IO. The mmap case does not increment
1015 * ras_requests and thus can never trigger this behavior. */
1016 if (ras->ras_requests == 2 && !ras->ras_request_index) {
1019 kms_pages = (i_size_read(inode) + CFS_PAGE_SIZE - 1) >>
1022 CDEBUG(D_READA, "kmsp "LPU64" mwp %lu mp %lu\n", kms_pages,
1023 ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages_per_file);
1026 kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1027 ras->ras_window_start = 0;
1028 ras->ras_last_readpage = 0;
1029 ras->ras_next_readahead = 0;
1030 ras->ras_window_len = min(ra->ra_max_pages_per_file,
1031 ra->ra_max_read_ahead_whole_pages);
1032 GOTO(out_unlock, 0);
1036 /* check whether it is in stride I/O mode*/
1037 if (!index_in_stride_window(index, ras, inode)) {
1038 if (ras->ras_consecutive_stride_requests == 0 &&
1039 ras->ras_request_index == 0) {
1040 ras_update_stride_detector(ras, index);
1041 ras->ras_consecutive_stride_requests ++;
1043 ras_stride_reset(ras);
1045 ras_reset(ras, index);
1046 ras->ras_consecutive_pages++;
1047 GOTO(out_unlock, 0);
1049 ras->ras_consecutive_pages = 0;
1050 ras->ras_consecutive_requests = 0;
1051 if (++ras->ras_consecutive_stride_requests > 1)
1057 if (index_in_stride_window(index, ras, inode) &&
1058 stride_io_mode(ras)) {
1059 /*If stride-RA hit cache miss, the stride dector
1060 *will not be reset to avoid the overhead of
1061 *redetecting read-ahead mode */
1062 if (index != ras->ras_last_readpage + 1)
1063 ras->ras_consecutive_pages = 0;
1064 ras_reset(ras, index);
1067 /* Reset both stride window and normal RA
1069 ras_reset(ras, index);
1070 ras->ras_consecutive_pages++;
1071 ras_stride_reset(ras);
1072 GOTO(out_unlock, 0);
1074 } else if (stride_io_mode(ras)) {
1075 /* If this is contiguous read but in stride I/O mode
1076 * currently, check whether stride step still is valid,
1077 * if invalid, it will reset the stride ra window*/
1078 if (!index_in_stride_window(index, ras, inode)) {
1079 /* Shrink stride read-ahead window to be zero */
1080 ras_stride_reset(ras);
1081 ras->ras_window_len = 0;
1082 ras->ras_next_readahead = index;
1086 ras->ras_consecutive_pages++;
1087 ras->ras_last_readpage = index;
1088 ras_set_start(ras, index);
1090 if (stride_io_mode(ras))
1091 /* Since stride readahead is sentivite to the offset
1092 * of read-ahead, so we use original offset here,
1093 * instead of ras_window_start, which is 1M aligned*/
1094 ras->ras_next_readahead = max(index,
1095 ras->ras_next_readahead);
1097 ras->ras_next_readahead = max(ras->ras_window_start,
1098 ras->ras_next_readahead);
1101 /* Trigger RA in the mmap case where ras_consecutive_requests
1102 * is not incremented and thus can't be used to trigger RA */
1103 if (!ras->ras_window_len && ras->ras_consecutive_pages == 4) {
1104 ras->ras_window_len = RAS_INCREASE_STEP;
1105 GOTO(out_unlock, 0);
1108 /* Initially reset the stride window offset to next_readahead*/
1109 if (ras->ras_consecutive_stride_requests == 2 && stride_detect) {
1111 * Once stride IO mode is detected, next_readahead should be
1112 * reset to make sure next_readahead > stride offset
1114 ras->ras_next_readahead = max(index, ras->ras_next_readahead);
1115 ras->ras_stride_offset = index;
1116 ras->ras_window_len = RAS_INCREASE_STEP;
1119 /* The initial ras_window_len is set to the request size. To avoid
1120 * uselessly reading and discarding pages for random IO the window is
1121 * only increased once per consecutive request received. */
1122 if ((ras->ras_consecutive_requests > 1 || stride_detect) &&
1123 !ras->ras_request_index)
1124 ras_increase_window(ras, ra, inode);
1128 ras->ras_request_index++;
1129 cfs_spin_unlock(&ras->ras_lock);
1133 int ll_writepage(struct page *vmpage, struct writeback_control *unused)
1135 struct inode *inode = vmpage->mapping->host;
1138 struct cl_page *page;
1139 struct cl_object *clob;
1140 struct cl_2queue *queue;
1141 struct cl_env_nest nest;
1145 LASSERT(PageLocked(vmpage));
1146 LASSERT(!PageWriteback(vmpage));
1148 if (ll_i2dtexp(inode) == NULL)
1151 env = cl_env_nested_get(&nest);
1153 RETURN(PTR_ERR(env));
1155 queue = &vvp_env_info(env)->vti_queue;
1156 clob = ll_i2info(inode)->lli_clob;
1157 LASSERT(clob != NULL);
1159 io = ccc_env_thread_io(env);
1161 result = cl_io_init(env, io, CIT_MISC, clob);
1163 page = cl_page_find(env, clob, vmpage->index,
1164 vmpage, CPT_CACHEABLE);
1165 if (!IS_ERR(page)) {
1166 lu_ref_add(&page->cp_reference, "writepage",
1168 cl_page_assume(env, io, page);
1170 * Mark page dirty, because this is what
1171 * ->vio_submit()->cpo_prep_write() assumes.
1173 * XXX better solution is to detect this from within
1174 * cl_io_submit_rw() somehow.
1176 set_page_dirty(vmpage);
1177 cl_2queue_init_page(queue, page);
1178 result = cl_io_submit_rw(env, io, CRT_WRITE,
1180 cl_page_list_disown(env, io, &queue->c2_qin);
1183 * There is no need to clear PG_writeback, as
1184 * cl_io_submit_rw() calls completion callback
1188 * Re-dirty page on error so it retries write,
1189 * but not in case when IO has actually
1190 * occurred and completed with an error.
1192 if (!PageError(vmpage))
1193 set_page_dirty(vmpage);
1195 LASSERT(!cl_page_is_owned(page, io));
1196 lu_ref_del(&page->cp_reference,
1197 "writepage", cfs_current());
1198 cl_page_put(env, page);
1199 cl_2queue_fini(env, queue);
1202 cl_io_fini(env, io);
1203 cl_env_nested_put(&nest, env);
1207 int ll_readpage(struct file *file, struct page *vmpage)
1209 struct ll_cl_context *lcc;
1213 lcc = ll_cl_init(file, vmpage, 0);
1215 struct lu_env *env = lcc->lcc_env;
1216 struct cl_io *io = lcc->lcc_io;
1217 struct cl_page *page = lcc->lcc_page;
1219 LASSERT(page->cp_type == CPT_CACHEABLE);
1220 if (likely(!PageUptodate(vmpage))) {
1221 cl_page_assume(env, io, page);
1222 result = cl_io_read_page(env, io, page);
1224 /* Page from a non-object file. */
1225 LASSERT(!ll_i2info(vmpage->mapping->host)->lli_smd);
1226 unlock_page(vmpage);
1231 unlock_page(vmpage);
1232 result = PTR_ERR(lcc);