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) {
154 * Loop-back driver calls ->prepare_write() and ->sendfile()
155 * methods directly, bypassing file system ->write() operation,
156 * so cl_io has to be created here.
158 io = ccc_env_thread_io(env);
159 ll_io_init(io, file, 1);
161 /* No lock at all for this kind of IO - we can't do it because
162 * we have held page lock, it would cause deadlock.
163 * XXX: This causes poor performance to loop device - One page
165 * In order to get better performance, users should use
166 * lloop driver instead.
168 io->ci_lockreq = CILR_NEVER;
170 pos = (vmpage->index << CFS_PAGE_SHIFT);
172 /* Create a temp IO to serve write. */
173 result = cl_io_rw_init(env, io, CIT_WRITE, pos, CFS_PAGE_SIZE);
175 cio->cui_fd = LUSTRE_FPRIVATE(file);
178 result = cl_io_iter_init(env, io);
180 result = cl_io_lock(env, io);
182 result = cl_io_start(env, io);
185 result = io->ci_result;
186 lcc->lcc_created = 1;
193 struct cl_page *page;
196 LASSERT(io->ci_state == CIS_IO_GOING);
197 LASSERT(cio->cui_fd == LUSTRE_FPRIVATE(file));
198 page = cl_page_find(env, clob, vmpage->index, vmpage,
201 lcc->lcc_page = page;
202 lu_ref_add(&page->cp_reference, "cl_io", io);
205 result = PTR_ERR(page);
209 lcc = ERR_PTR(result);
212 CDEBUG(D_VFSTRACE, "%lu@"DFID" -> %d %p %p\n",
213 vmpage->index, PFID(lu_object_fid(&clob->co_lu)), result,
218 static struct ll_cl_context *ll_cl_get(void)
220 struct ll_cl_context *lcc;
224 env = cl_env_get(&refcheck);
225 LASSERT(!IS_ERR(env));
226 lcc = &vvp_env_info(env)->vti_io_ctx;
227 LASSERT(env == lcc->lcc_env);
228 LASSERT(current == lcc->lcc_cookie);
229 cl_env_put(env, &refcheck);
231 /* env has got in ll_cl_init, so it is still usable. */
236 * ->prepare_write() address space operation called by generic_file_write()
237 * for every page during write.
239 int ll_prepare_write(struct file *file, struct page *vmpage, unsigned from,
242 struct ll_cl_context *lcc;
246 lcc = ll_cl_init(file, vmpage, 1);
248 struct lu_env *env = lcc->lcc_env;
249 struct cl_io *io = lcc->lcc_io;
250 struct cl_page *page = lcc->lcc_page;
252 cl_page_assume(env, io, page);
253 if (cl_io_is_append(io)) {
254 struct cl_object *obj = io->ci_obj;
255 struct inode *inode = ccc_object_inode(obj);
257 * In VFS file->page write loop, for appending, the
258 * write offset might be reset according to the new
259 * file size before holding i_mutex. So crw_pos should
260 * be reset here. BUG:17711.
262 io->u.ci_wr.wr.crw_pos = i_size_read(inode);
264 result = cl_io_prepare_write(env, io, page, from, to);
267 * Add a reference, so that page is not evicted from
268 * the cache until ->commit_write() is called.
271 lu_ref_add(&page->cp_reference, "prepare_write",
274 cl_page_unassume(env, io, page);
277 /* returning 0 in prepare assumes commit must be called
280 result = PTR_ERR(lcc);
285 int ll_commit_write(struct file *file, struct page *vmpage, unsigned from,
288 struct ll_cl_context *lcc;
291 struct cl_page *page;
297 page = lcc->lcc_page;
300 LASSERT(cl_page_is_owned(page, io));
302 if (from != to) /* handle short write case. */
303 result = cl_io_commit_write(env, io, page, from, to);
304 if (cl_page_is_owned(page, io))
305 cl_page_unassume(env, io, page);
308 * Release reference acquired by ll_prepare_write().
310 lu_ref_del(&page->cp_reference, "prepare_write", cfs_current());
311 cl_page_put(env, page);
316 struct obd_capa *cl_capa_lookup(struct inode *inode, enum cl_req_type crt)
320 opc = crt == CRT_WRITE ? CAPA_OPC_OSS_WRITE : CAPA_OPC_OSS_RW;
321 return ll_osscapa_get(inode, opc);
324 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which);
326 /* WARNING: This algorithm is used to reduce the contention on
327 * sbi->ll_lock. It should work well if the ra_max_pages is much
328 * greater than the single file's read-ahead window.
330 * TODO: There may exist a `global sync problem' in this implementation.
331 * Considering the global ra window is 100M, and each file's ra window is 10M,
332 * there are over 10 files trying to get its ra budget and reach
333 * ll_ra_count_get at the exactly same time. All of them will get a zero ra
334 * window, although the global window is 100M. -jay
336 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, struct ra_io_arg *ria,
339 struct ll_ra_info *ra = &sbi->ll_ra_info;
344 * If read-ahead pages left are less than 1M, do not do read-ahead,
345 * otherwise it will form small read RPC(< 1M), which hurt server
348 if (ra->ra_max_pages < atomic_read(&ra->ra_cur_pages))
351 ret = min(ra->ra_max_pages - cfs_atomic_read(&ra->ra_cur_pages), len);
352 if ((int)ret < 0 || ret < min((unsigned long)PTLRPC_MAX_BRW_PAGES, len))
355 if (ria->ria_pages == 0)
356 /* it needs 1M align again after trimed by ra_max_pages*/
357 if (ret >= ((ria->ria_start + ret) % PTLRPC_MAX_BRW_PAGES))
358 ret -= (ria->ria_start + ret) % PTLRPC_MAX_BRW_PAGES;
360 if (cfs_atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) {
361 cfs_atomic_sub(ret, &ra->ra_cur_pages);
369 void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
371 struct ll_ra_info *ra = &sbi->ll_ra_info;
372 cfs_atomic_sub(len, &ra->ra_cur_pages);
375 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which)
377 LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
378 lprocfs_counter_incr(sbi->ll_ra_stats, which);
381 void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which)
383 struct ll_sb_info *sbi = ll_i2sbi(mapping->host);
384 ll_ra_stats_inc_sbi(sbi, which);
387 #define RAS_CDEBUG(ras) \
389 "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu r %lu ri %lu" \
390 "csr %lu sf %lu sp %lu sl %lu \n", \
391 ras->ras_last_readpage, ras->ras_consecutive_requests, \
392 ras->ras_consecutive_pages, ras->ras_window_start, \
393 ras->ras_window_len, ras->ras_next_readahead, \
394 ras->ras_requests, ras->ras_request_index, \
395 ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
396 ras->ras_stride_pages, ras->ras_stride_length)
398 static int index_in_window(unsigned long index, unsigned long point,
399 unsigned long before, unsigned long after)
401 unsigned long start = point - before, end = point + after;
408 return start <= index && index <= end;
411 static struct ll_readahead_state *ll_ras_get(struct file *f)
413 struct ll_file_data *fd;
415 fd = LUSTRE_FPRIVATE(f);
419 void ll_ra_read_in(struct file *f, struct ll_ra_read *rar)
421 struct ll_readahead_state *ras;
425 cfs_spin_lock(&ras->ras_lock);
427 ras->ras_request_index = 0;
428 ras->ras_consecutive_requests++;
429 rar->lrr_reader = current;
431 cfs_list_add(&rar->lrr_linkage, &ras->ras_read_beads);
432 cfs_spin_unlock(&ras->ras_lock);
435 void ll_ra_read_ex(struct file *f, struct ll_ra_read *rar)
437 struct ll_readahead_state *ras;
441 cfs_spin_lock(&ras->ras_lock);
442 cfs_list_del_init(&rar->lrr_linkage);
443 cfs_spin_unlock(&ras->ras_lock);
446 static struct ll_ra_read *ll_ra_read_get_locked(struct ll_readahead_state *ras)
448 struct ll_ra_read *scan;
450 cfs_list_for_each_entry(scan, &ras->ras_read_beads, lrr_linkage) {
451 if (scan->lrr_reader == current)
457 struct ll_ra_read *ll_ra_read_get(struct file *f)
459 struct ll_readahead_state *ras;
460 struct ll_ra_read *bead;
464 cfs_spin_lock(&ras->ras_lock);
465 bead = ll_ra_read_get_locked(ras);
466 cfs_spin_unlock(&ras->ras_lock);
470 static int cl_read_ahead_page(const struct lu_env *env, struct cl_io *io,
471 struct cl_page_list *queue, struct cl_page *page,
480 cl_page_assume(env, io, page);
481 lu_ref_add(&page->cp_reference, "ra", cfs_current());
482 cp = cl2ccc_page(cl_page_at(page, &vvp_device_type));
483 if (!cp->cpg_defer_uptodate && !Page_Uptodate(vmpage)) {
484 rc = cl_page_is_under_lock(env, io, page);
486 cp->cpg_defer_uptodate = 1;
488 cl_page_list_add(queue, page);
491 cl_page_delete(env, page);
495 /* skip completed pages */
496 cl_page_unassume(env, io, page);
497 lu_ref_del(&page->cp_reference, "ra", cfs_current());
498 cl_page_put(env, page);
503 * Initiates read-ahead of a page with given index.
505 * \retval +ve: page was added to \a queue.
507 * \retval -ENOLCK: there is no extent lock for this part of a file, stop
510 * \retval -ve, 0: page wasn't added to \a queue for other reason.
512 static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io,
513 struct cl_page_list *queue,
514 pgoff_t index, struct address_space *mapping)
517 struct cl_object *clob = ll_i2info(mapping->host)->lli_clob;
518 struct cl_page *page;
519 enum ra_stat which = _NR_RA_STAT; /* keep gcc happy */
520 unsigned int gfp_mask;
522 const char *msg = NULL;
526 gfp_mask = GFP_HIGHUSER & ~__GFP_WAIT;
528 gfp_mask |= __GFP_NOWARN;
530 vmpage = grab_cache_page_nowait_gfp(mapping, index, gfp_mask);
531 if (vmpage != NULL) {
532 /* Check if vmpage was truncated or reclaimed */
533 if (vmpage->mapping == mapping) {
534 page = cl_page_find(env, clob, vmpage->index,
535 vmpage, CPT_CACHEABLE);
537 rc = cl_read_ahead_page(env, io, queue,
540 which = RA_STAT_FAILED_MATCH;
541 msg = "lock match failed";
544 which = RA_STAT_FAILED_GRAB_PAGE;
545 msg = "cl_page_find failed";
548 which = RA_STAT_WRONG_GRAB_PAGE;
549 msg = "g_c_p_n returned invalid page";
553 page_cache_release(vmpage);
555 which = RA_STAT_FAILED_GRAB_PAGE;
556 msg = "g_c_p_n failed";
559 ll_ra_stats_inc(mapping, which);
560 CDEBUG(D_READA, "%s\n", msg);
565 #define RIA_DEBUG(ria) \
566 CDEBUG(D_READA, "rs %lu re %lu ro %lu rl %lu rp %lu\n", \
567 ria->ria_start, ria->ria_end, ria->ria_stoff, ria->ria_length,\
570 #define RAS_INCREASE_STEP PTLRPC_MAX_BRW_PAGES
572 static inline int stride_io_mode(struct ll_readahead_state *ras)
574 return ras->ras_consecutive_stride_requests > 1;
576 /* The function calculates how much pages will be read in
577 * [off, off + length], in such stride IO area,
578 * stride_offset = st_off, stride_lengh = st_len,
579 * stride_pages = st_pgs
581 * |------------------|*****|------------------|*****|------------|*****|....
584 * |----- st_len -----|
586 * How many pages it should read in such pattern
587 * |-------------------------------------------------------------|
589 * |<------ length ------->|
591 * = |<----->| + |-------------------------------------| + |---|
592 * start_left st_pgs * i end_left
595 stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs,
596 unsigned long off, unsigned long length)
598 __u64 start = off > st_off ? off - st_off : 0;
599 __u64 end = off + length > st_off ? off + length - st_off : 0;
600 unsigned long start_left = 0;
601 unsigned long end_left = 0;
602 unsigned long pg_count;
604 if (st_len == 0 || length == 0 || end == 0)
607 start_left = do_div(start, st_len);
608 if (start_left < st_pgs)
609 start_left = st_pgs - start_left;
613 end_left = do_div(end, st_len);
614 if (end_left > st_pgs)
617 CDEBUG(D_READA, "start "LPU64", end "LPU64" start_left %lu end_left %lu \n",
618 start, end, start_left, end_left);
621 pg_count = end_left - (st_pgs - start_left);
623 pg_count = start_left + st_pgs * (end - start - 1) + end_left;
625 CDEBUG(D_READA, "st_off %lu, st_len %lu st_pgs %lu off %lu length %lu"
626 "pgcount %lu\n", st_off, st_len, st_pgs, off, length, pg_count);
631 static int ria_page_count(struct ra_io_arg *ria)
633 __u64 length = ria->ria_end >= ria->ria_start ?
634 ria->ria_end - ria->ria_start + 1 : 0;
636 return stride_pg_count(ria->ria_stoff, ria->ria_length,
637 ria->ria_pages, ria->ria_start,
641 /*Check whether the index is in the defined ra-window */
642 static int ras_inside_ra_window(unsigned long idx, struct ra_io_arg *ria)
644 /* If ria_length == ria_pages, it means non-stride I/O mode,
645 * idx should always inside read-ahead window in this case
646 * For stride I/O mode, just check whether the idx is inside
648 return ria->ria_length == 0 || ria->ria_length == ria->ria_pages ||
649 (idx >= ria->ria_stoff && (idx - ria->ria_stoff) %
650 ria->ria_length < ria->ria_pages);
653 static int ll_read_ahead_pages(const struct lu_env *env,
654 struct cl_io *io, struct cl_page_list *queue,
655 struct ra_io_arg *ria,
656 unsigned long *reserved_pages,
657 struct address_space *mapping,
658 unsigned long *ra_end)
660 int rc, count = 0, stride_ria;
661 unsigned long page_idx;
663 LASSERT(ria != NULL);
666 stride_ria = ria->ria_length > ria->ria_pages && ria->ria_pages > 0;
667 for (page_idx = ria->ria_start; page_idx <= ria->ria_end &&
668 *reserved_pages > 0; page_idx++) {
669 if (ras_inside_ra_window(page_idx, ria)) {
670 /* If the page is inside the read-ahead window*/
671 rc = ll_read_ahead_page(env, io, queue,
676 } else if (rc == -ENOLCK)
678 } else if (stride_ria) {
679 /* If it is not in the read-ahead window, and it is
680 * read-ahead mode, then check whether it should skip
683 /* FIXME: This assertion only is valid when it is for
684 * forward read-ahead, it will be fixed when backward
685 * read-ahead is implemented */
686 LASSERTF(page_idx > ria->ria_stoff, "Invalid page_idx %lu"
687 "rs %lu re %lu ro %lu rl %lu rp %lu\n", page_idx,
688 ria->ria_start, ria->ria_end, ria->ria_stoff,
689 ria->ria_length, ria->ria_pages);
690 offset = page_idx - ria->ria_stoff;
691 offset = offset % (ria->ria_length);
692 if (offset > ria->ria_pages) {
693 page_idx += ria->ria_length - offset;
694 CDEBUG(D_READA, "i %lu skip %lu \n", page_idx,
695 ria->ria_length - offset);
704 int ll_readahead(const struct lu_env *env, struct cl_io *io,
705 struct ll_readahead_state *ras, struct address_space *mapping,
706 struct cl_page_list *queue, int flags)
708 struct vvp_io *vio = vvp_env_io(env);
709 struct vvp_thread_info *vti = vvp_env_info(env);
710 struct cl_attr *attr = ccc_env_thread_attr(env);
711 unsigned long start = 0, end = 0, reserved;
712 unsigned long ra_end, len;
714 struct ll_ra_read *bead;
715 struct ra_io_arg *ria = &vti->vti_ria;
716 struct ll_inode_info *lli;
717 struct cl_object *clob;
722 inode = mapping->host;
723 lli = ll_i2info(inode);
724 clob = lli->lli_clob;
726 memset(ria, 0, sizeof *ria);
728 cl_object_attr_lock(clob);
729 ret = cl_object_attr_get(env, clob, attr);
730 cl_object_attr_unlock(clob);
736 ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN);
740 cfs_spin_lock(&ras->ras_lock);
741 if (vio->cui_ra_window_set)
742 bead = &vio->cui_bead;
746 /* Enlarge the RA window to encompass the full read */
747 if (bead != NULL && ras->ras_window_start + ras->ras_window_len <
748 bead->lrr_start + bead->lrr_count) {
749 ras->ras_window_len = bead->lrr_start + bead->lrr_count -
750 ras->ras_window_start;
752 /* Reserve a part of the read-ahead window that we'll be issuing */
753 if (ras->ras_window_len) {
754 start = ras->ras_next_readahead;
755 end = ras->ras_window_start + ras->ras_window_len - 1;
758 unsigned long tmp_end;
760 * Align RA window to an optimal boundary.
762 * XXX This would be better to align to cl_max_pages_per_rpc
763 * instead of PTLRPC_MAX_BRW_PAGES, because the RPC size may
764 * be aligned to the RAID stripe size in the future and that
765 * is more important than the RPC size.
767 tmp_end = ((end + 1) & (~(PTLRPC_MAX_BRW_PAGES - 1))) - 1;
771 /* Truncate RA window to end of file */
772 end = min(end, (unsigned long)((kms - 1) >> CFS_PAGE_SHIFT));
774 ras->ras_next_readahead = max(end, end + 1);
777 ria->ria_start = start;
779 /* If stride I/O mode is detected, get stride window*/
780 if (stride_io_mode(ras)) {
781 ria->ria_stoff = ras->ras_stride_offset;
782 ria->ria_length = ras->ras_stride_length;
783 ria->ria_pages = ras->ras_stride_pages;
785 cfs_spin_unlock(&ras->ras_lock);
788 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
791 len = ria_page_count(ria);
795 reserved = ll_ra_count_get(ll_i2sbi(inode), ria, len);
797 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
799 CDEBUG(D_READA, "reserved page %lu \n", reserved);
801 ret = ll_read_ahead_pages(env, io, queue,
802 ria, &reserved, mapping, &ra_end);
804 LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
806 ll_ra_count_put(ll_i2sbi(inode), reserved);
808 if (ra_end == end + 1 && ra_end == (kms >> CFS_PAGE_SHIFT))
809 ll_ra_stats_inc(mapping, RA_STAT_EOF);
811 /* if we didn't get to the end of the region we reserved from
812 * the ras we need to go back and update the ras so that the
813 * next read-ahead tries from where we left off. we only do so
814 * if the region we failed to issue read-ahead on is still ahead
815 * of the app and behind the next index to start read-ahead from */
816 CDEBUG(D_READA, "ra_end %lu end %lu stride end %lu \n",
817 ra_end, end, ria->ria_end);
819 if (ra_end != end + 1) {
820 cfs_spin_lock(&ras->ras_lock);
821 if (ra_end < ras->ras_next_readahead &&
822 index_in_window(ra_end, ras->ras_window_start, 0,
823 ras->ras_window_len)) {
824 ras->ras_next_readahead = ra_end;
827 cfs_spin_unlock(&ras->ras_lock);
833 static void ras_set_start(struct ll_readahead_state *ras, unsigned long index)
835 ras->ras_window_start = index & (~(RAS_INCREASE_STEP - 1));
838 /* called with the ras_lock held or from places where it doesn't matter */
839 static void ras_reset(struct ll_readahead_state *ras, unsigned long index)
841 ras->ras_last_readpage = index;
842 ras->ras_consecutive_requests = 0;
843 ras->ras_consecutive_pages = 0;
844 ras->ras_window_len = 0;
845 ras_set_start(ras, index);
846 ras->ras_next_readahead = max(ras->ras_window_start, index);
851 /* called with the ras_lock held or from places where it doesn't matter */
852 static void ras_stride_reset(struct ll_readahead_state *ras)
854 ras->ras_consecutive_stride_requests = 0;
855 ras->ras_stride_length = 0;
856 ras->ras_stride_pages = 0;
860 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
862 cfs_spin_lock_init(&ras->ras_lock);
864 ras->ras_requests = 0;
865 CFS_INIT_LIST_HEAD(&ras->ras_read_beads);
869 * Check whether the read request is in the stride window.
870 * If it is in the stride window, return 1, otherwise return 0.
872 static int index_in_stride_window(unsigned long index,
873 struct ll_readahead_state *ras,
876 unsigned long stride_gap = index - ras->ras_last_readpage - 1;
878 if (ras->ras_stride_length == 0 || ras->ras_stride_pages == 0 ||
879 ras->ras_stride_pages == ras->ras_stride_length)
882 /* If it is contiguous read */
884 return ras->ras_consecutive_pages + 1 <= ras->ras_stride_pages;
886 /*Otherwise check the stride by itself */
887 return (ras->ras_stride_length - ras->ras_stride_pages) == stride_gap &&
888 ras->ras_consecutive_pages == ras->ras_stride_pages;
891 static void ras_update_stride_detector(struct ll_readahead_state *ras,
894 unsigned long stride_gap = index - ras->ras_last_readpage - 1;
896 if (!stride_io_mode(ras) && (stride_gap != 0 ||
897 ras->ras_consecutive_stride_requests == 0)) {
898 ras->ras_stride_pages = ras->ras_consecutive_pages;
899 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
901 LASSERT(ras->ras_request_index == 0);
902 LASSERT(ras->ras_consecutive_stride_requests == 0);
904 if (index <= ras->ras_last_readpage) {
905 /*Reset stride window for forward read*/
906 ras_stride_reset(ras);
910 ras->ras_stride_pages = ras->ras_consecutive_pages;
911 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
918 stride_page_count(struct ll_readahead_state *ras, unsigned long len)
920 return stride_pg_count(ras->ras_stride_offset, ras->ras_stride_length,
921 ras->ras_stride_pages, ras->ras_stride_offset,
925 /* Stride Read-ahead window will be increased inc_len according to
926 * stride I/O pattern */
927 static void ras_stride_increase_window(struct ll_readahead_state *ras,
928 struct ll_ra_info *ra,
929 unsigned long inc_len)
931 unsigned long left, step, window_len;
932 unsigned long stride_len;
934 LASSERT(ras->ras_stride_length > 0);
935 LASSERTF(ras->ras_window_start + ras->ras_window_len
936 >= ras->ras_stride_offset, "window_start %lu, window_len %lu"
937 " stride_offset %lu\n", ras->ras_window_start,
938 ras->ras_window_len, ras->ras_stride_offset);
940 stride_len = ras->ras_window_start + ras->ras_window_len -
941 ras->ras_stride_offset;
943 left = stride_len % ras->ras_stride_length;
944 window_len = ras->ras_window_len - left;
946 if (left < ras->ras_stride_pages)
949 left = ras->ras_stride_pages + inc_len;
951 LASSERT(ras->ras_stride_pages != 0);
953 step = left / ras->ras_stride_pages;
954 left %= ras->ras_stride_pages;
956 window_len += step * ras->ras_stride_length + left;
958 if (stride_page_count(ras, window_len) <= ra->ra_max_pages_per_file)
959 ras->ras_window_len = window_len;
964 static void ras_increase_window(struct ll_readahead_state *ras,
965 struct ll_ra_info *ra, struct inode *inode)
967 /* The stretch of ra-window should be aligned with max rpc_size
968 * but current clio architecture does not support retrieve such
969 * information from lower layer. FIXME later
971 if (stride_io_mode(ras))
972 ras_stride_increase_window(ras, ra, RAS_INCREASE_STEP);
974 ras->ras_window_len = min(ras->ras_window_len +
976 ra->ra_max_pages_per_file);
979 void ras_update(struct ll_sb_info *sbi, struct inode *inode,
980 struct ll_readahead_state *ras, unsigned long index,
983 struct ll_ra_info *ra = &sbi->ll_ra_info;
984 int zero = 0, stride_detect = 0, ra_miss = 0;
987 cfs_spin_lock(&ras->ras_lock);
989 ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
991 /* reset the read-ahead window in two cases. First when the app seeks
992 * or reads to some other part of the file. Secondly if we get a
993 * read-ahead miss that we think we've previously issued. This can
994 * be a symptom of there being so many read-ahead pages that the VM is
995 * reclaiming it before we get to it. */
996 if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
998 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
999 } else if (!hit && ras->ras_window_len &&
1000 index < ras->ras_next_readahead &&
1001 index_in_window(index, ras->ras_window_start, 0,
1002 ras->ras_window_len)) {
1004 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
1007 /* On the second access to a file smaller than the tunable
1008 * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1009 * file up to ra_max_pages_per_file. This is simply a best effort
1010 * and only occurs once per open file. Normal RA behavior is reverted
1011 * to for subsequent IO. The mmap case does not increment
1012 * ras_requests and thus can never trigger this behavior. */
1013 if (ras->ras_requests == 2 && !ras->ras_request_index) {
1016 kms_pages = (i_size_read(inode) + CFS_PAGE_SIZE - 1) >>
1019 CDEBUG(D_READA, "kmsp "LPU64" mwp %lu mp %lu\n", kms_pages,
1020 ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages_per_file);
1023 kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1024 ras->ras_window_start = 0;
1025 ras->ras_last_readpage = 0;
1026 ras->ras_next_readahead = 0;
1027 ras->ras_window_len = min(ra->ra_max_pages_per_file,
1028 ra->ra_max_read_ahead_whole_pages);
1029 GOTO(out_unlock, 0);
1033 /* check whether it is in stride I/O mode*/
1034 if (!index_in_stride_window(index, ras, inode)) {
1035 if (ras->ras_consecutive_stride_requests == 0 &&
1036 ras->ras_request_index == 0) {
1037 ras_update_stride_detector(ras, index);
1038 ras->ras_consecutive_stride_requests ++;
1040 ras_stride_reset(ras);
1042 ras_reset(ras, index);
1043 ras->ras_consecutive_pages++;
1044 GOTO(out_unlock, 0);
1046 ras->ras_consecutive_pages = 0;
1047 ras->ras_consecutive_requests = 0;
1048 if (++ras->ras_consecutive_stride_requests > 1)
1054 if (index_in_stride_window(index, ras, inode) &&
1055 stride_io_mode(ras)) {
1056 /*If stride-RA hit cache miss, the stride dector
1057 *will not be reset to avoid the overhead of
1058 *redetecting read-ahead mode */
1059 if (index != ras->ras_last_readpage + 1)
1060 ras->ras_consecutive_pages = 0;
1061 ras_reset(ras, index);
1064 /* Reset both stride window and normal RA
1066 ras_reset(ras, index);
1067 ras->ras_consecutive_pages++;
1068 ras_stride_reset(ras);
1069 GOTO(out_unlock, 0);
1071 } else if (stride_io_mode(ras)) {
1072 /* If this is contiguous read but in stride I/O mode
1073 * currently, check whether stride step still is valid,
1074 * if invalid, it will reset the stride ra window*/
1075 if (!index_in_stride_window(index, ras, inode)) {
1076 /* Shrink stride read-ahead window to be zero */
1077 ras_stride_reset(ras);
1078 ras->ras_window_len = 0;
1079 ras->ras_next_readahead = index;
1083 ras->ras_consecutive_pages++;
1084 ras->ras_last_readpage = index;
1085 ras_set_start(ras, index);
1087 if (stride_io_mode(ras))
1088 /* Since stride readahead is sentivite to the offset
1089 * of read-ahead, so we use original offset here,
1090 * instead of ras_window_start, which is 1M aligned*/
1091 ras->ras_next_readahead = max(index,
1092 ras->ras_next_readahead);
1094 ras->ras_next_readahead = max(ras->ras_window_start,
1095 ras->ras_next_readahead);
1098 /* Trigger RA in the mmap case where ras_consecutive_requests
1099 * is not incremented and thus can't be used to trigger RA */
1100 if (!ras->ras_window_len && ras->ras_consecutive_pages == 4) {
1101 ras->ras_window_len = RAS_INCREASE_STEP;
1102 GOTO(out_unlock, 0);
1105 /* Initially reset the stride window offset to next_readahead*/
1106 if (ras->ras_consecutive_stride_requests == 2 && stride_detect) {
1108 * Once stride IO mode is detected, next_readahead should be
1109 * reset to make sure next_readahead > stride offset
1111 ras->ras_next_readahead = max(index, ras->ras_next_readahead);
1112 ras->ras_stride_offset = index;
1113 ras->ras_window_len = RAS_INCREASE_STEP;
1116 /* The initial ras_window_len is set to the request size. To avoid
1117 * uselessly reading and discarding pages for random IO the window is
1118 * only increased once per consecutive request received. */
1119 if ((ras->ras_consecutive_requests > 1 || stride_detect) &&
1120 !ras->ras_request_index)
1121 ras_increase_window(ras, ra, inode);
1125 ras->ras_request_index++;
1126 cfs_spin_unlock(&ras->ras_lock);
1130 int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
1132 struct inode *inode = vmpage->mapping->host;
1135 struct cl_page *page;
1136 struct cl_object *clob;
1137 struct cl_2queue *queue;
1138 struct cl_env_nest nest;
1142 LASSERT(PageLocked(vmpage));
1143 LASSERT(!PageWriteback(vmpage));
1145 if (ll_i2dtexp(inode) == NULL)
1148 env = cl_env_nested_get(&nest);
1150 RETURN(PTR_ERR(env));
1152 queue = &vvp_env_info(env)->vti_queue;
1153 clob = ll_i2info(inode)->lli_clob;
1154 LASSERT(clob != NULL);
1156 io = ccc_env_thread_io(env);
1158 result = cl_io_init(env, io, CIT_MISC, clob);
1160 page = cl_page_find(env, clob, vmpage->index,
1161 vmpage, CPT_CACHEABLE);
1162 if (!IS_ERR(page)) {
1163 lu_ref_add(&page->cp_reference, "writepage",
1165 cl_page_assume(env, io, page);
1167 * Mark page dirty, because this is what
1168 * ->vio_submit()->cpo_prep_write() assumes.
1170 * XXX better solution is to detect this from within
1171 * cl_io_submit_rw() somehow.
1173 set_page_dirty(vmpage);
1174 cl_2queue_init_page(queue, page);
1175 result = cl_io_submit_rw(env, io, CRT_WRITE,
1179 * Re-dirty page on error so it retries write,
1180 * but not in case when IO has actually
1181 * occurred and completed with an error.
1183 if (!PageError(vmpage)) {
1184 redirty_page_for_writepage(wbc, vmpage);
1188 cl_page_list_disown(env, io, &queue->c2_qin);
1189 LASSERT(!cl_page_is_owned(page, io));
1190 lu_ref_del(&page->cp_reference,
1191 "writepage", cfs_current());
1192 cl_page_put(env, page);
1193 cl_2queue_fini(env, queue);
1196 cl_io_fini(env, io);
1197 cl_env_nested_put(&nest, env);
1201 int ll_readpage(struct file *file, struct page *vmpage)
1203 struct ll_cl_context *lcc;
1207 lcc = ll_cl_init(file, vmpage, 0);
1209 struct lu_env *env = lcc->lcc_env;
1210 struct cl_io *io = lcc->lcc_io;
1211 struct cl_page *page = lcc->lcc_page;
1213 LASSERT(page->cp_type == CPT_CACHEABLE);
1214 if (likely(!PageUptodate(vmpage))) {
1215 cl_page_assume(env, io, page);
1216 result = cl_io_read_page(env, io, page);
1218 /* Page from a non-object file. */
1219 LASSERT(!ll_i2info(vmpage->mapping->host)->lli_smd);
1220 unlock_page(vmpage);
1225 unlock_page(vmpage);
1226 result = PTR_ERR(lcc);