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.
32 * Copyright (c) 2011, 2012, Whamcloud, Inc.
35 * This file is part of Lustre, http://www.lustre.org/
36 * Lustre is a trademark of Sun Microsystems, Inc.
40 * Lustre Lite I/O page cache routines shared by different kernel revs
43 #include <linux/kernel.h>
45 #include <linux/string.h>
46 #include <linux/stat.h>
47 #include <linux/errno.h>
48 #include <linux/smp_lock.h>
49 #include <linux/unistd.h>
50 #include <linux/version.h>
51 #include <asm/system.h>
52 #include <asm/uaccess.h>
55 #include <linux/stat.h>
56 #include <asm/uaccess.h>
58 #include <linux/pagemap.h>
59 #include <linux/smp_lock.h>
60 /* current_is_kswapd() */
61 #include <linux/swap.h>
63 #define DEBUG_SUBSYSTEM S_LLITE
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)
80 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) to %llu\n", inode->i_ino,
81 inode->i_generation, inode, i_size_read(inode));
88 * Finalizes cl-data before exiting typical address_space operation. Dual to
91 static void ll_cl_fini(struct ll_cl_context *lcc)
93 struct lu_env *env = lcc->lcc_env;
94 struct cl_io *io = lcc->lcc_io;
95 struct cl_page *page = lcc->lcc_page;
97 LASSERT(lcc->lcc_cookie == current);
101 lu_ref_del(&page->cp_reference, "cl_io", io);
102 cl_page_put(env, page);
105 if (io && lcc->lcc_created) {
107 cl_io_unlock(env, io);
108 cl_io_iter_fini(env, io);
111 cl_env_put(env, &lcc->lcc_refcheck);
115 * Initializes common cl-data at the typical address_space operation entry
118 static struct ll_cl_context *ll_cl_init(struct file *file,
119 struct page *vmpage, int create)
121 struct ll_cl_context *lcc;
124 struct cl_object *clob;
130 clob = ll_i2info(vmpage->mapping->host)->lli_clob;
131 LASSERT(clob != NULL);
133 env = cl_env_get(&refcheck);
135 return ERR_PTR(PTR_ERR(env));
137 lcc = &vvp_env_info(env)->vti_io_ctx;
138 memset(lcc, 0, sizeof(*lcc));
140 lcc->lcc_refcheck = refcheck;
141 lcc->lcc_cookie = current;
143 cio = ccc_env_io(env);
144 io = cio->cui_cl.cis_io;
145 if (io == NULL && create) {
149 * Loop-back driver calls ->prepare_write() and ->sendfile()
150 * methods directly, bypassing file system ->write() operation,
151 * so cl_io has to be created here.
153 io = ccc_env_thread_io(env);
154 ll_io_init(io, file, 1);
156 /* No lock at all for this kind of IO - we can't do it because
157 * we have held page lock, it would cause deadlock.
158 * XXX: This causes poor performance to loop device - One page
160 * In order to get better performance, users should use
161 * lloop driver instead.
163 io->ci_lockreq = CILR_NEVER;
165 pos = (vmpage->index << CFS_PAGE_SHIFT);
167 /* Create a temp IO to serve write. */
168 result = cl_io_rw_init(env, io, CIT_WRITE, pos, CFS_PAGE_SIZE);
170 cio->cui_fd = LUSTRE_FPRIVATE(file);
173 result = cl_io_iter_init(env, io);
175 result = cl_io_lock(env, io);
177 result = cl_io_start(env, io);
180 result = io->ci_result;
181 lcc->lcc_created = 1;
188 struct cl_page *page;
191 LASSERT(io->ci_state == CIS_IO_GOING);
192 LASSERT(cio->cui_fd == LUSTRE_FPRIVATE(file));
193 page = cl_page_find(env, clob, vmpage->index, vmpage,
196 lcc->lcc_page = page;
197 lu_ref_add(&page->cp_reference, "cl_io", io);
200 result = PTR_ERR(page);
204 lcc = ERR_PTR(result);
207 CDEBUG(D_VFSTRACE, "%lu@"DFID" -> %d %p %p\n",
208 vmpage->index, PFID(lu_object_fid(&clob->co_lu)), result,
213 static struct ll_cl_context *ll_cl_get(void)
215 struct ll_cl_context *lcc;
219 env = cl_env_get(&refcheck);
220 LASSERT(!IS_ERR(env));
221 lcc = &vvp_env_info(env)->vti_io_ctx;
222 LASSERT(env == lcc->lcc_env);
223 LASSERT(current == lcc->lcc_cookie);
224 cl_env_put(env, &refcheck);
226 /* env has got in ll_cl_init, so it is still usable. */
231 * ->prepare_write() address space operation called by generic_file_write()
232 * for every page during write.
234 int ll_prepare_write(struct file *file, struct page *vmpage, unsigned from,
237 struct ll_cl_context *lcc;
241 lcc = ll_cl_init(file, vmpage, 1);
243 struct lu_env *env = lcc->lcc_env;
244 struct cl_io *io = lcc->lcc_io;
245 struct cl_page *page = lcc->lcc_page;
247 cl_page_assume(env, io, page);
248 if (cl_io_is_append(io)) {
249 struct cl_object *obj = io->ci_obj;
250 struct inode *inode = ccc_object_inode(obj);
252 * In VFS file->page write loop, for appending, the
253 * write offset might be reset according to the new
254 * file size before holding i_mutex. So crw_pos should
255 * be reset here. BUG:17711.
257 io->u.ci_wr.wr.crw_pos = i_size_read(inode);
259 result = cl_io_prepare_write(env, io, page, from, to);
262 * Add a reference, so that page is not evicted from
263 * the cache until ->commit_write() is called.
266 lu_ref_add(&page->cp_reference, "prepare_write",
269 cl_page_unassume(env, io, page);
272 /* returning 0 in prepare assumes commit must be called
275 result = PTR_ERR(lcc);
280 int ll_commit_write(struct file *file, struct page *vmpage, unsigned from,
283 struct ll_cl_context *lcc;
286 struct cl_page *page;
292 page = lcc->lcc_page;
295 LASSERT(cl_page_is_owned(page, io));
297 if (from != to) /* handle short write case. */
298 result = cl_io_commit_write(env, io, page, from, to);
299 if (cl_page_is_owned(page, io))
300 cl_page_unassume(env, io, page);
303 * Release reference acquired by ll_prepare_write().
305 lu_ref_del(&page->cp_reference, "prepare_write", cfs_current());
306 cl_page_put(env, page);
311 struct obd_capa *cl_capa_lookup(struct inode *inode, enum cl_req_type crt)
315 opc = crt == CRT_WRITE ? CAPA_OPC_OSS_WRITE : CAPA_OPC_OSS_RW;
316 return ll_osscapa_get(inode, opc);
319 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which);
322 * Get readahead pages from the filesystem readahead pool of the client for a
325 * /param sbi superblock for filesystem readahead state ll_ra_info
326 * /param ria per-thread readahead state
327 * /param pages number of pages requested for readahead for the thread.
329 * WARNING: This algorithm is used to reduce contention on sbi->ll_lock.
330 * It should work well if the ra_max_pages is much greater than the single
331 * file's read-ahead window, and not too many threads contending for
332 * these readahead pages.
334 * TODO: There may be a 'global sync problem' if many threads are trying
335 * to get an ra budget that is larger than the remaining readahead pages
336 * and reach here at exactly the same time. They will compute /a ret to
337 * consume the remaining pages, but will fail at atomic_add_return() and
338 * get a zero ra window, although there is still ra space remaining. - Jay */
340 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi,
341 struct ra_io_arg *ria,
344 struct ll_ra_info *ra = &sbi->ll_ra_info;
348 /* If read-ahead pages left are less than 1M, do not do read-ahead,
349 * otherwise it will form small read RPC(< 1M), which hurt server
350 * performance a lot. */
351 ret = min(ra->ra_max_pages - cfs_atomic_read(&ra->ra_cur_pages), pages);
352 if (ret < 0 || ret < min_t(long, PTLRPC_MAX_BRW_PAGES, pages))
355 /* If the non-strided (ria_pages == 0) readahead window
356 * (ria_start + ret) has grown across an RPC boundary, then trim
357 * readahead size by the amount beyond the RPC so it ends on an
358 * RPC boundary. If the readahead window is already ending on
359 * an RPC boundary (beyond_rpc == 0), or smaller than a full
360 * RPC (beyond_rpc < ret) the readahead size is unchanged.
361 * The (beyond_rpc != 0) check is skipped since the conditional
362 * branch is more expensive than subtracting zero from the result.
364 * Strided read is left unaligned to avoid small fragments beyond
365 * the RPC boundary from needing an extra read RPC. */
366 if (ria->ria_pages == 0) {
367 long beyond_rpc = (ria->ria_start + ret) % PTLRPC_MAX_BRW_PAGES;
368 if (/* beyond_rpc != 0 && */ beyond_rpc < ret)
372 if (cfs_atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) {
373 cfs_atomic_sub(ret, &ra->ra_cur_pages);
381 void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
383 struct ll_ra_info *ra = &sbi->ll_ra_info;
384 cfs_atomic_sub(len, &ra->ra_cur_pages);
387 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which)
389 LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
390 lprocfs_counter_incr(sbi->ll_ra_stats, which);
393 void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which)
395 struct ll_sb_info *sbi = ll_i2sbi(mapping->host);
396 ll_ra_stats_inc_sbi(sbi, which);
399 #define RAS_CDEBUG(ras) \
401 "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu r %lu ri %lu" \
402 "csr %lu sf %lu sp %lu sl %lu \n", \
403 ras->ras_last_readpage, ras->ras_consecutive_requests, \
404 ras->ras_consecutive_pages, ras->ras_window_start, \
405 ras->ras_window_len, ras->ras_next_readahead, \
406 ras->ras_requests, ras->ras_request_index, \
407 ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
408 ras->ras_stride_pages, ras->ras_stride_length)
410 static int index_in_window(unsigned long index, unsigned long point,
411 unsigned long before, unsigned long after)
413 unsigned long start = point - before, end = point + after;
420 return start <= index && index <= end;
423 static struct ll_readahead_state *ll_ras_get(struct file *f)
425 struct ll_file_data *fd;
427 fd = LUSTRE_FPRIVATE(f);
431 void ll_ra_read_in(struct file *f, struct ll_ra_read *rar)
433 struct ll_readahead_state *ras;
437 cfs_spin_lock(&ras->ras_lock);
439 ras->ras_request_index = 0;
440 ras->ras_consecutive_requests++;
441 rar->lrr_reader = current;
443 cfs_list_add(&rar->lrr_linkage, &ras->ras_read_beads);
444 cfs_spin_unlock(&ras->ras_lock);
447 void ll_ra_read_ex(struct file *f, struct ll_ra_read *rar)
449 struct ll_readahead_state *ras;
453 cfs_spin_lock(&ras->ras_lock);
454 cfs_list_del_init(&rar->lrr_linkage);
455 cfs_spin_unlock(&ras->ras_lock);
458 static struct ll_ra_read *ll_ra_read_get_locked(struct ll_readahead_state *ras)
460 struct ll_ra_read *scan;
462 cfs_list_for_each_entry(scan, &ras->ras_read_beads, lrr_linkage) {
463 if (scan->lrr_reader == current)
469 struct ll_ra_read *ll_ra_read_get(struct file *f)
471 struct ll_readahead_state *ras;
472 struct ll_ra_read *bead;
476 cfs_spin_lock(&ras->ras_lock);
477 bead = ll_ra_read_get_locked(ras);
478 cfs_spin_unlock(&ras->ras_lock);
482 static int cl_read_ahead_page(const struct lu_env *env, struct cl_io *io,
483 struct cl_page_list *queue, struct cl_page *page,
492 cl_page_assume(env, io, page);
493 lu_ref_add(&page->cp_reference, "ra", cfs_current());
494 cp = cl2ccc_page(cl_page_at(page, &vvp_device_type));
495 if (!cp->cpg_defer_uptodate && !Page_Uptodate(vmpage)) {
496 rc = cl_page_is_under_lock(env, io, page);
498 cp->cpg_defer_uptodate = 1;
500 cl_page_list_add(queue, page);
503 cl_page_delete(env, page);
507 /* skip completed pages */
508 cl_page_unassume(env, io, page);
509 lu_ref_del(&page->cp_reference, "ra", cfs_current());
510 cl_page_put(env, page);
515 * Initiates read-ahead of a page with given index.
517 * \retval +ve: page was added to \a queue.
519 * \retval -ENOLCK: there is no extent lock for this part of a file, stop
522 * \retval -ve, 0: page wasn't added to \a queue for other reason.
524 static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io,
525 struct cl_page_list *queue,
526 pgoff_t index, struct address_space *mapping)
529 struct cl_object *clob = ll_i2info(mapping->host)->lli_clob;
530 struct cl_page *page;
531 enum ra_stat which = _NR_RA_STAT; /* keep gcc happy */
532 unsigned int gfp_mask;
534 const char *msg = NULL;
538 gfp_mask = GFP_HIGHUSER & ~__GFP_WAIT;
540 gfp_mask |= __GFP_NOWARN;
542 vmpage = grab_cache_page_nowait(mapping, index);
543 if (vmpage != NULL) {
544 /* Check if vmpage was truncated or reclaimed */
545 if (vmpage->mapping == mapping) {
546 page = cl_page_find(env, clob, vmpage->index,
547 vmpage, CPT_CACHEABLE);
549 rc = cl_read_ahead_page(env, io, queue,
552 which = RA_STAT_FAILED_MATCH;
553 msg = "lock match failed";
556 which = RA_STAT_FAILED_GRAB_PAGE;
557 msg = "cl_page_find failed";
560 which = RA_STAT_WRONG_GRAB_PAGE;
561 msg = "g_c_p_n returned invalid page";
565 page_cache_release(vmpage);
567 which = RA_STAT_FAILED_GRAB_PAGE;
568 msg = "g_c_p_n failed";
571 ll_ra_stats_inc(mapping, which);
572 CDEBUG(D_READA, "%s\n", msg);
577 #define RIA_DEBUG(ria) \
578 CDEBUG(D_READA, "rs %lu re %lu ro %lu rl %lu rp %lu\n", \
579 ria->ria_start, ria->ria_end, ria->ria_stoff, ria->ria_length,\
582 #define RAS_INCREASE_STEP PTLRPC_MAX_BRW_PAGES
584 static inline int stride_io_mode(struct ll_readahead_state *ras)
586 return ras->ras_consecutive_stride_requests > 1;
588 /* The function calculates how much pages will be read in
589 * [off, off + length], in such stride IO area,
590 * stride_offset = st_off, stride_lengh = st_len,
591 * stride_pages = st_pgs
593 * |------------------|*****|------------------|*****|------------|*****|....
596 * |----- st_len -----|
598 * How many pages it should read in such pattern
599 * |-------------------------------------------------------------|
601 * |<------ length ------->|
603 * = |<----->| + |-------------------------------------| + |---|
604 * start_left st_pgs * i end_left
607 stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs,
608 unsigned long off, unsigned long length)
610 __u64 start = off > st_off ? off - st_off : 0;
611 __u64 end = off + length > st_off ? off + length - st_off : 0;
612 unsigned long start_left = 0;
613 unsigned long end_left = 0;
614 unsigned long pg_count;
616 if (st_len == 0 || length == 0 || end == 0)
619 start_left = do_div(start, st_len);
620 if (start_left < st_pgs)
621 start_left = st_pgs - start_left;
625 end_left = do_div(end, st_len);
626 if (end_left > st_pgs)
629 CDEBUG(D_READA, "start "LPU64", end "LPU64" start_left %lu end_left %lu \n",
630 start, end, start_left, end_left);
633 pg_count = end_left - (st_pgs - start_left);
635 pg_count = start_left + st_pgs * (end - start - 1) + end_left;
637 CDEBUG(D_READA, "st_off %lu, st_len %lu st_pgs %lu off %lu length %lu"
638 "pgcount %lu\n", st_off, st_len, st_pgs, off, length, pg_count);
643 static int ria_page_count(struct ra_io_arg *ria)
645 __u64 length = ria->ria_end >= ria->ria_start ?
646 ria->ria_end - ria->ria_start + 1 : 0;
648 return stride_pg_count(ria->ria_stoff, ria->ria_length,
649 ria->ria_pages, ria->ria_start,
653 /*Check whether the index is in the defined ra-window */
654 static int ras_inside_ra_window(unsigned long idx, struct ra_io_arg *ria)
656 /* If ria_length == ria_pages, it means non-stride I/O mode,
657 * idx should always inside read-ahead window in this case
658 * For stride I/O mode, just check whether the idx is inside
660 return ria->ria_length == 0 || ria->ria_length == ria->ria_pages ||
661 (idx >= ria->ria_stoff && (idx - ria->ria_stoff) %
662 ria->ria_length < ria->ria_pages);
665 static int ll_read_ahead_pages(const struct lu_env *env,
666 struct cl_io *io, struct cl_page_list *queue,
667 struct ra_io_arg *ria,
668 unsigned long *reserved_pages,
669 struct address_space *mapping,
670 unsigned long *ra_end)
672 int rc, count = 0, stride_ria;
673 unsigned long page_idx;
675 LASSERT(ria != NULL);
678 stride_ria = ria->ria_length > ria->ria_pages && ria->ria_pages > 0;
679 for (page_idx = ria->ria_start; page_idx <= ria->ria_end &&
680 *reserved_pages > 0; page_idx++) {
681 if (ras_inside_ra_window(page_idx, ria)) {
682 /* If the page is inside the read-ahead window*/
683 rc = ll_read_ahead_page(env, io, queue,
688 } else if (rc == -ENOLCK)
690 } else if (stride_ria) {
691 /* If it is not in the read-ahead window, and it is
692 * read-ahead mode, then check whether it should skip
695 /* FIXME: This assertion only is valid when it is for
696 * forward read-ahead, it will be fixed when backward
697 * read-ahead is implemented */
698 LASSERTF(page_idx > ria->ria_stoff, "Invalid page_idx %lu"
699 "rs %lu re %lu ro %lu rl %lu rp %lu\n", page_idx,
700 ria->ria_start, ria->ria_end, ria->ria_stoff,
701 ria->ria_length, ria->ria_pages);
702 offset = page_idx - ria->ria_stoff;
703 offset = offset % (ria->ria_length);
704 if (offset > ria->ria_pages) {
705 page_idx += ria->ria_length - offset;
706 CDEBUG(D_READA, "i %lu skip %lu \n", page_idx,
707 ria->ria_length - offset);
716 int ll_readahead(const struct lu_env *env, struct cl_io *io,
717 struct ll_readahead_state *ras, struct address_space *mapping,
718 struct cl_page_list *queue, int flags)
720 struct vvp_io *vio = vvp_env_io(env);
721 struct vvp_thread_info *vti = vvp_env_info(env);
722 struct cl_attr *attr = ccc_env_thread_attr(env);
723 unsigned long start = 0, end = 0, reserved;
724 unsigned long ra_end, len;
726 struct ll_ra_read *bead;
727 struct ra_io_arg *ria = &vti->vti_ria;
728 struct ll_inode_info *lli;
729 struct cl_object *clob;
734 inode = mapping->host;
735 lli = ll_i2info(inode);
736 clob = lli->lli_clob;
738 memset(ria, 0, sizeof *ria);
740 cl_object_attr_lock(clob);
741 ret = cl_object_attr_get(env, clob, attr);
742 cl_object_attr_unlock(clob);
748 ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN);
752 cfs_spin_lock(&ras->ras_lock);
753 if (vio->cui_ra_window_set)
754 bead = &vio->cui_bead;
758 /* Enlarge the RA window to encompass the full read */
759 if (bead != NULL && ras->ras_window_start + ras->ras_window_len <
760 bead->lrr_start + bead->lrr_count) {
761 ras->ras_window_len = bead->lrr_start + bead->lrr_count -
762 ras->ras_window_start;
764 /* Reserve a part of the read-ahead window that we'll be issuing */
765 if (ras->ras_window_len) {
766 start = ras->ras_next_readahead;
767 end = ras->ras_window_start + ras->ras_window_len - 1;
770 unsigned long rpc_boundary;
772 * Align RA window to an optimal boundary.
774 * XXX This would be better to align to cl_max_pages_per_rpc
775 * instead of PTLRPC_MAX_BRW_PAGES, because the RPC size may
776 * be aligned to the RAID stripe size in the future and that
777 * is more important than the RPC size.
779 /* Note: we only trim the RPC, instead of extending the RPC
780 * to the boundary, so to avoid reading too much pages during
782 rpc_boundary = ((end + 1) & (~(PTLRPC_MAX_BRW_PAGES - 1)));
783 if (rpc_boundary > 0)
786 if (rpc_boundary > start)
789 /* Truncate RA window to end of file */
790 end = min(end, (unsigned long)((kms - 1) >> CFS_PAGE_SHIFT));
792 ras->ras_next_readahead = max(end, end + 1);
795 ria->ria_start = start;
797 /* If stride I/O mode is detected, get stride window*/
798 if (stride_io_mode(ras)) {
799 ria->ria_stoff = ras->ras_stride_offset;
800 ria->ria_length = ras->ras_stride_length;
801 ria->ria_pages = ras->ras_stride_pages;
803 cfs_spin_unlock(&ras->ras_lock);
806 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
809 len = ria_page_count(ria);
813 reserved = ll_ra_count_get(ll_i2sbi(inode), ria, len);
815 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
817 CDEBUG(D_READA, "reserved page %lu ra_cur %d ra_max %lu\n", reserved,
818 cfs_atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
819 ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
821 ret = ll_read_ahead_pages(env, io, queue,
822 ria, &reserved, mapping, &ra_end);
824 LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
826 ll_ra_count_put(ll_i2sbi(inode), reserved);
828 if (ra_end == end + 1 && ra_end == (kms >> CFS_PAGE_SHIFT))
829 ll_ra_stats_inc(mapping, RA_STAT_EOF);
831 /* if we didn't get to the end of the region we reserved from
832 * the ras we need to go back and update the ras so that the
833 * next read-ahead tries from where we left off. we only do so
834 * if the region we failed to issue read-ahead on is still ahead
835 * of the app and behind the next index to start read-ahead from */
836 CDEBUG(D_READA, "ra_end %lu end %lu stride end %lu \n",
837 ra_end, end, ria->ria_end);
839 if (ra_end != end + 1) {
840 cfs_spin_lock(&ras->ras_lock);
841 if (ra_end < ras->ras_next_readahead &&
842 index_in_window(ra_end, ras->ras_window_start, 0,
843 ras->ras_window_len)) {
844 ras->ras_next_readahead = ra_end;
847 cfs_spin_unlock(&ras->ras_lock);
853 static void ras_set_start(struct ll_readahead_state *ras, unsigned long index)
855 ras->ras_window_start = index & (~(RAS_INCREASE_STEP - 1));
858 /* called with the ras_lock held or from places where it doesn't matter */
859 static void ras_reset(struct ll_readahead_state *ras, unsigned long index)
861 ras->ras_last_readpage = index;
862 ras->ras_consecutive_requests = 0;
863 ras->ras_consecutive_pages = 0;
864 ras->ras_window_len = 0;
865 ras_set_start(ras, index);
866 ras->ras_next_readahead = max(ras->ras_window_start, index);
871 /* called with the ras_lock held or from places where it doesn't matter */
872 static void ras_stride_reset(struct ll_readahead_state *ras)
874 ras->ras_consecutive_stride_requests = 0;
875 ras->ras_stride_length = 0;
876 ras->ras_stride_pages = 0;
880 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
882 cfs_spin_lock_init(&ras->ras_lock);
884 ras->ras_requests = 0;
885 CFS_INIT_LIST_HEAD(&ras->ras_read_beads);
889 * Check whether the read request is in the stride window.
890 * If it is in the stride window, return 1, otherwise return 0.
892 static int index_in_stride_window(unsigned long index,
893 struct ll_readahead_state *ras,
896 unsigned long stride_gap = index - ras->ras_last_readpage - 1;
898 if (ras->ras_stride_length == 0 || ras->ras_stride_pages == 0 ||
899 ras->ras_stride_pages == ras->ras_stride_length)
902 /* If it is contiguous read */
904 return ras->ras_consecutive_pages + 1 <= ras->ras_stride_pages;
906 /*Otherwise check the stride by itself */
907 return (ras->ras_stride_length - ras->ras_stride_pages) == stride_gap &&
908 ras->ras_consecutive_pages == ras->ras_stride_pages;
911 static void ras_update_stride_detector(struct ll_readahead_state *ras,
914 unsigned long stride_gap = index - ras->ras_last_readpage - 1;
916 if (!stride_io_mode(ras) && (stride_gap != 0 ||
917 ras->ras_consecutive_stride_requests == 0)) {
918 ras->ras_stride_pages = ras->ras_consecutive_pages;
919 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
921 LASSERT(ras->ras_request_index == 0);
922 LASSERT(ras->ras_consecutive_stride_requests == 0);
924 if (index <= ras->ras_last_readpage) {
925 /*Reset stride window for forward read*/
926 ras_stride_reset(ras);
930 ras->ras_stride_pages = ras->ras_consecutive_pages;
931 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
938 stride_page_count(struct ll_readahead_state *ras, unsigned long len)
940 return stride_pg_count(ras->ras_stride_offset, ras->ras_stride_length,
941 ras->ras_stride_pages, ras->ras_stride_offset,
945 /* Stride Read-ahead window will be increased inc_len according to
946 * stride I/O pattern */
947 static void ras_stride_increase_window(struct ll_readahead_state *ras,
948 struct ll_ra_info *ra,
949 unsigned long inc_len)
951 unsigned long left, step, window_len;
952 unsigned long stride_len;
954 LASSERT(ras->ras_stride_length > 0);
955 LASSERTF(ras->ras_window_start + ras->ras_window_len
956 >= ras->ras_stride_offset, "window_start %lu, window_len %lu"
957 " stride_offset %lu\n", ras->ras_window_start,
958 ras->ras_window_len, ras->ras_stride_offset);
960 stride_len = ras->ras_window_start + ras->ras_window_len -
961 ras->ras_stride_offset;
963 left = stride_len % ras->ras_stride_length;
964 window_len = ras->ras_window_len - left;
966 if (left < ras->ras_stride_pages)
969 left = ras->ras_stride_pages + inc_len;
971 LASSERT(ras->ras_stride_pages != 0);
973 step = left / ras->ras_stride_pages;
974 left %= ras->ras_stride_pages;
976 window_len += step * ras->ras_stride_length + left;
978 if (stride_page_count(ras, window_len) <= ra->ra_max_pages_per_file)
979 ras->ras_window_len = window_len;
984 static void ras_increase_window(struct ll_readahead_state *ras,
985 struct ll_ra_info *ra, struct inode *inode)
987 /* The stretch of ra-window should be aligned with max rpc_size
988 * but current clio architecture does not support retrieve such
989 * information from lower layer. FIXME later
991 if (stride_io_mode(ras))
992 ras_stride_increase_window(ras, ra, RAS_INCREASE_STEP);
994 ras->ras_window_len = min(ras->ras_window_len +
996 ra->ra_max_pages_per_file);
999 void ras_update(struct ll_sb_info *sbi, struct inode *inode,
1000 struct ll_readahead_state *ras, unsigned long index,
1003 struct ll_ra_info *ra = &sbi->ll_ra_info;
1004 int zero = 0, stride_detect = 0, ra_miss = 0;
1007 cfs_spin_lock(&ras->ras_lock);
1009 ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
1011 /* reset the read-ahead window in two cases. First when the app seeks
1012 * or reads to some other part of the file. Secondly if we get a
1013 * read-ahead miss that we think we've previously issued. This can
1014 * be a symptom of there being so many read-ahead pages that the VM is
1015 * reclaiming it before we get to it. */
1016 if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
1018 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
1019 } else if (!hit && ras->ras_window_len &&
1020 index < ras->ras_next_readahead &&
1021 index_in_window(index, ras->ras_window_start, 0,
1022 ras->ras_window_len)) {
1024 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
1027 /* On the second access to a file smaller than the tunable
1028 * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1029 * file up to ra_max_pages_per_file. This is simply a best effort
1030 * and only occurs once per open file. Normal RA behavior is reverted
1031 * to for subsequent IO. The mmap case does not increment
1032 * ras_requests and thus can never trigger this behavior. */
1033 if (ras->ras_requests == 2 && !ras->ras_request_index) {
1036 kms_pages = (i_size_read(inode) + CFS_PAGE_SIZE - 1) >>
1039 CDEBUG(D_READA, "kmsp "LPU64" mwp %lu mp %lu\n", kms_pages,
1040 ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages_per_file);
1043 kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1044 ras->ras_window_start = 0;
1045 ras->ras_last_readpage = 0;
1046 ras->ras_next_readahead = 0;
1047 ras->ras_window_len = min(ra->ra_max_pages_per_file,
1048 ra->ra_max_read_ahead_whole_pages);
1049 GOTO(out_unlock, 0);
1053 /* check whether it is in stride I/O mode*/
1054 if (!index_in_stride_window(index, ras, inode)) {
1055 if (ras->ras_consecutive_stride_requests == 0 &&
1056 ras->ras_request_index == 0) {
1057 ras_update_stride_detector(ras, index);
1058 ras->ras_consecutive_stride_requests ++;
1060 ras_stride_reset(ras);
1062 ras_reset(ras, index);
1063 ras->ras_consecutive_pages++;
1064 GOTO(out_unlock, 0);
1066 ras->ras_consecutive_pages = 0;
1067 ras->ras_consecutive_requests = 0;
1068 if (++ras->ras_consecutive_stride_requests > 1)
1074 if (index_in_stride_window(index, ras, inode) &&
1075 stride_io_mode(ras)) {
1076 /*If stride-RA hit cache miss, the stride dector
1077 *will not be reset to avoid the overhead of
1078 *redetecting read-ahead mode */
1079 if (index != ras->ras_last_readpage + 1)
1080 ras->ras_consecutive_pages = 0;
1081 ras_reset(ras, index);
1084 /* Reset both stride window and normal RA
1086 ras_reset(ras, index);
1087 ras->ras_consecutive_pages++;
1088 ras_stride_reset(ras);
1089 GOTO(out_unlock, 0);
1091 } else if (stride_io_mode(ras)) {
1092 /* If this is contiguous read but in stride I/O mode
1093 * currently, check whether stride step still is valid,
1094 * if invalid, it will reset the stride ra window*/
1095 if (!index_in_stride_window(index, ras, inode)) {
1096 /* Shrink stride read-ahead window to be zero */
1097 ras_stride_reset(ras);
1098 ras->ras_window_len = 0;
1099 ras->ras_next_readahead = index;
1103 ras->ras_consecutive_pages++;
1104 ras->ras_last_readpage = index;
1105 ras_set_start(ras, index);
1107 if (stride_io_mode(ras))
1108 /* Since stride readahead is sentivite to the offset
1109 * of read-ahead, so we use original offset here,
1110 * instead of ras_window_start, which is 1M aligned*/
1111 ras->ras_next_readahead = max(index,
1112 ras->ras_next_readahead);
1114 ras->ras_next_readahead = max(ras->ras_window_start,
1115 ras->ras_next_readahead);
1118 /* Trigger RA in the mmap case where ras_consecutive_requests
1119 * is not incremented and thus can't be used to trigger RA */
1120 if (!ras->ras_window_len && ras->ras_consecutive_pages == 4) {
1121 ras->ras_window_len = RAS_INCREASE_STEP;
1122 GOTO(out_unlock, 0);
1125 /* Initially reset the stride window offset to next_readahead*/
1126 if (ras->ras_consecutive_stride_requests == 2 && stride_detect) {
1128 * Once stride IO mode is detected, next_readahead should be
1129 * reset to make sure next_readahead > stride offset
1131 ras->ras_next_readahead = max(index, ras->ras_next_readahead);
1132 ras->ras_stride_offset = index;
1133 ras->ras_window_len = RAS_INCREASE_STEP;
1136 /* The initial ras_window_len is set to the request size. To avoid
1137 * uselessly reading and discarding pages for random IO the window is
1138 * only increased once per consecutive request received. */
1139 if ((ras->ras_consecutive_requests > 1 || stride_detect) &&
1140 !ras->ras_request_index)
1141 ras_increase_window(ras, ra, inode);
1145 ras->ras_request_index++;
1146 cfs_spin_unlock(&ras->ras_lock);
1150 int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
1152 struct inode *inode = vmpage->mapping->host;
1155 struct cl_page *page;
1156 struct cl_object *clob;
1157 struct cl_2queue *queue;
1158 struct cl_env_nest nest;
1162 LASSERT(PageLocked(vmpage));
1163 LASSERT(!PageWriteback(vmpage));
1165 if (ll_i2dtexp(inode) == NULL)
1168 env = cl_env_nested_get(&nest);
1170 RETURN(PTR_ERR(env));
1172 queue = &vvp_env_info(env)->vti_queue;
1173 clob = ll_i2info(inode)->lli_clob;
1174 LASSERT(clob != NULL);
1176 io = ccc_env_thread_io(env);
1178 result = cl_io_init(env, io, CIT_MISC, clob);
1180 page = cl_page_find(env, clob, vmpage->index,
1181 vmpage, CPT_CACHEABLE);
1182 if (!IS_ERR(page)) {
1183 lu_ref_add(&page->cp_reference, "writepage",
1185 cl_page_assume(env, io, page);
1187 * Mark page dirty, because this is what
1188 * ->vio_submit()->cpo_prep_write() assumes.
1190 * XXX better solution is to detect this from within
1191 * cl_io_submit_rw() somehow.
1193 set_page_dirty(vmpage);
1194 cl_2queue_init_page(queue, page);
1195 result = cl_io_submit_rw(env, io, CRT_WRITE,
1199 * Re-dirty page on error so it retries write,
1200 * but not in case when IO has actually
1201 * occurred and completed with an error.
1203 if (!PageError(vmpage)) {
1204 redirty_page_for_writepage(wbc, vmpage);
1208 cl_page_list_disown(env, io, &queue->c2_qin);
1209 LASSERT(!cl_page_is_owned(page, io));
1210 lu_ref_del(&page->cp_reference,
1211 "writepage", cfs_current());
1212 cl_page_put(env, page);
1213 cl_2queue_fini(env, queue);
1216 cl_io_fini(env, io);
1217 cl_env_nested_put(&nest, env);
1221 int ll_readpage(struct file *file, struct page *vmpage)
1223 struct ll_cl_context *lcc;
1227 lcc = ll_cl_init(file, vmpage, 0);
1229 struct lu_env *env = lcc->lcc_env;
1230 struct cl_io *io = lcc->lcc_io;
1231 struct cl_page *page = lcc->lcc_page;
1233 LASSERT(page->cp_type == CPT_CACHEABLE);
1234 if (likely(!PageUptodate(vmpage))) {
1235 cl_page_assume(env, io, page);
1236 result = cl_io_read_page(env, io, page);
1238 /* Page from a non-object file. */
1239 LASSERT(!ll_i2info(vmpage->mapping->host)->lli_smd);
1240 unlock_page(vmpage);
1245 unlock_page(vmpage);
1246 result = PTR_ERR(lcc);