4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2013, Intel Corporation.
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/unistd.h>
47 #include <linux/writeback.h>
48 #include <asm/uaccess.h>
51 #include <linux/stat.h>
52 #include <asm/uaccess.h>
54 #include <linux/pagemap.h>
55 /* current_is_kswapd() */
56 #include <linux/swap.h>
58 #define DEBUG_SUBSYSTEM S_LLITE
60 #include <lustre_lite.h>
61 #include <obd_cksum.h>
62 #include "llite_internal.h"
63 #include <linux/lustre_compat25.h>
66 * Finalizes cl-data before exiting typical address_space operation. Dual to
69 void ll_cl_fini(struct ll_cl_context *lcc)
71 struct lu_env *env = lcc->lcc_env;
72 struct cl_io *io = lcc->lcc_io;
73 struct cl_page *page = lcc->lcc_page;
75 LASSERT(lcc->lcc_cookie == current);
79 lu_ref_del(&page->cp_reference, "cl_io", io);
80 cl_page_put(env, page);
83 if (io && lcc->lcc_created) {
85 cl_io_unlock(env, io);
86 cl_io_iter_fini(env, io);
89 cl_env_put(env, &lcc->lcc_refcheck);
93 * Initializes common cl-data at the typical address_space operation entry
96 struct ll_cl_context *ll_cl_init(struct file *file, struct page *vmpage)
98 struct ll_cl_context *lcc;
101 struct cl_object *clob;
107 clob = ll_i2info(file->f_dentry->d_inode)->lli_clob;
108 LASSERT(clob != NULL);
110 env = cl_env_get(&refcheck);
112 return ERR_PTR(PTR_ERR(env));
114 lcc = &vvp_env_info(env)->vti_io_ctx;
115 memset(lcc, 0, sizeof(*lcc));
117 lcc->lcc_refcheck = refcheck;
118 lcc->lcc_cookie = current;
120 cio = ccc_env_io(env);
121 io = cio->cui_cl.cis_io;
124 struct inode *inode = file->f_dentry->d_inode;
126 CERROR("%s: " DFID " no active IO, please file a ticket.\n",
127 ll_get_fsname(inode->i_sb, NULL, 0),
128 PFID(ll_inode2fid(inode)));
129 libcfs_debug_dumpstack(NULL);
132 if (result == 0 && vmpage != NULL) {
133 struct cl_page *page;
136 LASSERT(io->ci_state == CIS_IO_GOING);
137 LASSERT(cio->cui_fd == LUSTRE_FPRIVATE(file));
138 page = cl_page_find(env, clob, vmpage->index, vmpage,
141 lcc->lcc_page = page;
142 lu_ref_add(&page->cp_reference, "cl_io", io);
145 result = PTR_ERR(page);
149 lcc = ERR_PTR(result);
155 struct obd_capa *cl_capa_lookup(struct inode *inode, enum cl_req_type crt)
159 opc = crt == CRT_WRITE ? CAPA_OPC_OSS_WRITE : CAPA_OPC_OSS_RW;
160 return ll_osscapa_get(inode, opc);
163 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which);
166 * Get readahead pages from the filesystem readahead pool of the client for a
169 * /param sbi superblock for filesystem readahead state ll_ra_info
170 * /param ria per-thread readahead state
171 * /param pages number of pages requested for readahead for the thread.
173 * WARNING: This algorithm is used to reduce contention on sbi->ll_lock.
174 * It should work well if the ra_max_pages is much greater than the single
175 * file's read-ahead window, and not too many threads contending for
176 * these readahead pages.
178 * TODO: There may be a 'global sync problem' if many threads are trying
179 * to get an ra budget that is larger than the remaining readahead pages
180 * and reach here at exactly the same time. They will compute /a ret to
181 * consume the remaining pages, but will fail at atomic_add_return() and
182 * get a zero ra window, although there is still ra space remaining. - Jay */
184 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi,
185 struct ra_io_arg *ria,
186 unsigned long pages, unsigned long min)
188 struct ll_ra_info *ra = &sbi->ll_ra_info;
192 /* If read-ahead pages left are less than 1M, do not do read-ahead,
193 * otherwise it will form small read RPC(< 1M), which hurt server
194 * performance a lot. */
195 ret = min(ra->ra_max_pages - atomic_read(&ra->ra_cur_pages), pages);
196 if (ret < 0 || ret < min_t(long, PTLRPC_MAX_BRW_PAGES, pages))
199 /* If the non-strided (ria_pages == 0) readahead window
200 * (ria_start + ret) has grown across an RPC boundary, then trim
201 * readahead size by the amount beyond the RPC so it ends on an
202 * RPC boundary. If the readahead window is already ending on
203 * an RPC boundary (beyond_rpc == 0), or smaller than a full
204 * RPC (beyond_rpc < ret) the readahead size is unchanged.
205 * The (beyond_rpc != 0) check is skipped since the conditional
206 * branch is more expensive than subtracting zero from the result.
208 * Strided read is left unaligned to avoid small fragments beyond
209 * the RPC boundary from needing an extra read RPC. */
210 if (ria->ria_pages == 0) {
211 long beyond_rpc = (ria->ria_start + ret) % PTLRPC_MAX_BRW_PAGES;
212 if (/* beyond_rpc != 0 && */ beyond_rpc < ret)
216 if (atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) {
217 atomic_sub(ret, &ra->ra_cur_pages);
223 /* override ra limit for maximum performance */
224 atomic_add(min - ret, &ra->ra_cur_pages);
230 void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
232 struct ll_ra_info *ra = &sbi->ll_ra_info;
233 atomic_sub(len, &ra->ra_cur_pages);
236 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which)
238 LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
239 lprocfs_counter_incr(sbi->ll_ra_stats, which);
242 void ll_ra_stats_inc(struct inode *inode, enum ra_stat which)
244 struct ll_sb_info *sbi = ll_i2sbi(inode);
245 ll_ra_stats_inc_sbi(sbi, which);
248 #define RAS_CDEBUG(ras) \
250 "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu r %lu ri %lu" \
251 "csr %lu sf %lu sp %lu sl %lu \n", \
252 ras->ras_last_readpage, ras->ras_consecutive_requests, \
253 ras->ras_consecutive_pages, ras->ras_window_start, \
254 ras->ras_window_len, ras->ras_next_readahead, \
255 ras->ras_requests, ras->ras_request_index, \
256 ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
257 ras->ras_stride_pages, ras->ras_stride_length)
259 static int index_in_window(unsigned long index, unsigned long point,
260 unsigned long before, unsigned long after)
262 unsigned long start = point - before, end = point + after;
269 return start <= index && index <= end;
272 static struct ll_readahead_state *ll_ras_get(struct file *f)
274 struct ll_file_data *fd;
276 fd = LUSTRE_FPRIVATE(f);
280 void ll_ra_read_in(struct file *f, struct ll_ra_read *rar)
282 struct ll_readahead_state *ras;
286 spin_lock(&ras->ras_lock);
288 ras->ras_request_index = 0;
289 ras->ras_consecutive_requests++;
290 rar->lrr_reader = current;
292 cfs_list_add(&rar->lrr_linkage, &ras->ras_read_beads);
293 spin_unlock(&ras->ras_lock);
296 void ll_ra_read_ex(struct file *f, struct ll_ra_read *rar)
298 struct ll_readahead_state *ras;
302 spin_lock(&ras->ras_lock);
303 cfs_list_del_init(&rar->lrr_linkage);
304 spin_unlock(&ras->ras_lock);
307 static struct ll_ra_read *ll_ra_read_get_locked(struct ll_readahead_state *ras)
309 struct ll_ra_read *scan;
311 cfs_list_for_each_entry(scan, &ras->ras_read_beads, lrr_linkage) {
312 if (scan->lrr_reader == current)
318 struct ll_ra_read *ll_ra_read_get(struct file *f)
320 struct ll_readahead_state *ras;
321 struct ll_ra_read *bead;
325 spin_lock(&ras->ras_lock);
326 bead = ll_ra_read_get_locked(ras);
327 spin_unlock(&ras->ras_lock);
331 static int cl_read_ahead_page(const struct lu_env *env, struct cl_io *io,
332 struct cl_page_list *queue, struct cl_page *page,
333 struct cl_object *clob, pgoff_t *max_index)
335 struct page *vmpage = page->cp_vmpage;
342 cl_page_assume(env, io, page);
343 lu_ref_add(&page->cp_reference, "ra", current);
344 cp = cl2ccc_page(cl_object_page_slice(clob, page));
345 if (!cp->cpg_defer_uptodate && !PageUptodate(vmpage)) {
346 CDEBUG(D_READA, "page index %lu, max_index: %lu\n",
347 ccc_index(cp), *max_index);
348 if (*max_index == 0 || ccc_index(cp) > *max_index)
349 rc = cl_page_is_under_lock(env, io, page, max_index);
351 cp->cpg_defer_uptodate = 1;
353 cl_page_list_add(queue, page);
356 cl_page_discard(env, io, page);
360 /* skip completed pages */
361 cl_page_unassume(env, io, page);
363 lu_ref_del(&page->cp_reference, "ra", current);
364 cl_page_put(env, page);
369 * Initiates read-ahead of a page with given index.
371 * \retval +ve: page was added to \a queue.
373 * \retval -ENOLCK: there is no extent lock for this part of a file, stop
376 * \retval -ve, 0: page wasn't added to \a queue for other reason.
378 static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io,
379 struct cl_page_list *queue,
380 pgoff_t index, pgoff_t *max_index)
382 struct cl_object *clob = io->ci_obj;
383 struct inode *inode = ccc_object_inode(clob);
385 struct cl_page *page;
386 enum ra_stat which = _NR_RA_STAT; /* keep gcc happy */
387 unsigned int gfp_mask;
389 const char *msg = NULL;
393 gfp_mask = GFP_HIGHUSER & ~__GFP_WAIT;
395 gfp_mask |= __GFP_NOWARN;
397 vmpage = grab_cache_page_nowait(inode->i_mapping, index);
398 if (vmpage != NULL) {
399 /* Check if vmpage was truncated or reclaimed */
400 if (vmpage->mapping == inode->i_mapping) {
401 page = cl_page_find(env, clob, vmpage->index,
402 vmpage, CPT_CACHEABLE);
404 rc = cl_read_ahead_page(env, io, queue,
405 page, clob, max_index);
407 which = RA_STAT_FAILED_MATCH;
408 msg = "lock match failed";
411 which = RA_STAT_FAILED_GRAB_PAGE;
412 msg = "cl_page_find failed";
415 which = RA_STAT_WRONG_GRAB_PAGE;
416 msg = "g_c_p_n returned invalid page";
420 page_cache_release(vmpage);
422 which = RA_STAT_FAILED_GRAB_PAGE;
423 msg = "g_c_p_n failed";
426 ll_ra_stats_inc(inode, which);
427 CDEBUG(D_READA, "%s\n", msg);
432 #define RIA_DEBUG(ria) \
433 CDEBUG(D_READA, "rs %lu re %lu ro %lu rl %lu rp %lu\n", \
434 ria->ria_start, ria->ria_end, ria->ria_stoff, ria->ria_length,\
437 /* Limit this to the blocksize instead of PTLRPC_BRW_MAX_SIZE, since we don't
438 * know what the actual RPC size is. If this needs to change, it makes more
439 * sense to tune the i_blkbits value for the file based on the OSTs it is
440 * striped over, rather than having a constant value for all files here. */
442 /* RAS_INCREASE_STEP should be (1UL << (inode->i_blkbits - PAGE_CACHE_SHIFT)).
443 * Temprarily set RAS_INCREASE_STEP to 1MB. After 4MB RPC is enabled
444 * by default, this should be adjusted corresponding with max_read_ahead_mb
445 * and max_read_ahead_per_file_mb otherwise the readahead budget can be used
446 * up quickly which will affect read performance siginificantly. See LU-2816 */
447 #define RAS_INCREASE_STEP(inode) (ONE_MB_BRW_SIZE >> PAGE_CACHE_SHIFT)
449 static inline int stride_io_mode(struct ll_readahead_state *ras)
451 return ras->ras_consecutive_stride_requests > 1;
453 /* The function calculates how much pages will be read in
454 * [off, off + length], in such stride IO area,
455 * stride_offset = st_off, stride_lengh = st_len,
456 * stride_pages = st_pgs
458 * |------------------|*****|------------------|*****|------------|*****|....
461 * |----- st_len -----|
463 * How many pages it should read in such pattern
464 * |-------------------------------------------------------------|
466 * |<------ length ------->|
468 * = |<----->| + |-------------------------------------| + |---|
469 * start_left st_pgs * i end_left
472 stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs,
473 unsigned long off, unsigned long length)
475 __u64 start = off > st_off ? off - st_off : 0;
476 __u64 end = off + length > st_off ? off + length - st_off : 0;
477 unsigned long start_left = 0;
478 unsigned long end_left = 0;
479 unsigned long pg_count;
481 if (st_len == 0 || length == 0 || end == 0)
484 start_left = do_div(start, st_len);
485 if (start_left < st_pgs)
486 start_left = st_pgs - start_left;
490 end_left = do_div(end, st_len);
491 if (end_left > st_pgs)
494 CDEBUG(D_READA, "start "LPU64", end "LPU64" start_left %lu end_left %lu \n",
495 start, end, start_left, end_left);
498 pg_count = end_left - (st_pgs - start_left);
500 pg_count = start_left + st_pgs * (end - start - 1) + end_left;
502 CDEBUG(D_READA, "st_off %lu, st_len %lu st_pgs %lu off %lu length %lu"
503 "pgcount %lu\n", st_off, st_len, st_pgs, off, length, pg_count);
508 static int ria_page_count(struct ra_io_arg *ria)
510 __u64 length = ria->ria_end >= ria->ria_start ?
511 ria->ria_end - ria->ria_start + 1 : 0;
513 return stride_pg_count(ria->ria_stoff, ria->ria_length,
514 ria->ria_pages, ria->ria_start,
518 /*Check whether the index is in the defined ra-window */
519 static int ras_inside_ra_window(unsigned long idx, struct ra_io_arg *ria)
521 /* If ria_length == ria_pages, it means non-stride I/O mode,
522 * idx should always inside read-ahead window in this case
523 * For stride I/O mode, just check whether the idx is inside
525 return ria->ria_length == 0 || ria->ria_length == ria->ria_pages ||
526 (idx >= ria->ria_stoff && (idx - ria->ria_stoff) %
527 ria->ria_length < ria->ria_pages);
530 static int ll_read_ahead_pages(const struct lu_env *env,
531 struct cl_io *io, struct cl_page_list *queue,
532 struct ra_io_arg *ria,
533 unsigned long *reserved_pages,
534 unsigned long *ra_end)
539 pgoff_t max_index = 0;
541 LASSERT(ria != NULL);
544 stride_ria = ria->ria_length > ria->ria_pages && ria->ria_pages > 0;
545 for (page_idx = ria->ria_start;
546 page_idx <= ria->ria_end && *reserved_pages > 0; page_idx++) {
547 if (ras_inside_ra_window(page_idx, ria)) {
548 /* If the page is inside the read-ahead window*/
549 rc = ll_read_ahead_page(env, io, queue,
550 page_idx, &max_index);
554 } else if (rc == -ENOLCK)
556 } else if (stride_ria) {
557 /* If it is not in the read-ahead window, and it is
558 * read-ahead mode, then check whether it should skip
561 /* FIXME: This assertion only is valid when it is for
562 * forward read-ahead, it will be fixed when backward
563 * read-ahead is implemented */
564 LASSERTF(page_idx > ria->ria_stoff, "Invalid page_idx %lu"
565 "rs %lu re %lu ro %lu rl %lu rp %lu\n", page_idx,
566 ria->ria_start, ria->ria_end, ria->ria_stoff,
567 ria->ria_length, ria->ria_pages);
568 offset = page_idx - ria->ria_stoff;
569 offset = offset % (ria->ria_length);
570 if (offset > ria->ria_pages) {
571 page_idx += ria->ria_length - offset;
572 CDEBUG(D_READA, "i %lu skip %lu \n", page_idx,
573 ria->ria_length - offset);
582 int ll_readahead(const struct lu_env *env, struct cl_io *io,
583 struct cl_page_list *queue, struct ll_readahead_state *ras,
586 struct vvp_io *vio = vvp_env_io(env);
587 struct vvp_thread_info *vti = vvp_env_info(env);
588 struct cl_attr *attr = ccc_env_thread_attr(env);
589 unsigned long start = 0, end = 0, reserved;
590 unsigned long ra_end, len, mlen = 0;
592 struct ll_ra_read *bead;
593 struct ra_io_arg *ria = &vti->vti_ria;
594 struct cl_object *clob;
600 inode = ccc_object_inode(clob);
602 memset(ria, 0, sizeof *ria);
604 cl_object_attr_lock(clob);
605 ret = cl_object_attr_get(env, clob, attr);
606 cl_object_attr_unlock(clob);
612 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
616 spin_lock(&ras->ras_lock);
617 if (vio->cui_ra_window_set)
618 bead = &vio->cui_bead;
622 /* Enlarge the RA window to encompass the full read */
623 if (bead != NULL && ras->ras_window_start + ras->ras_window_len <
624 bead->lrr_start + bead->lrr_count) {
625 ras->ras_window_len = bead->lrr_start + bead->lrr_count -
626 ras->ras_window_start;
628 /* Reserve a part of the read-ahead window that we'll be issuing */
629 if (ras->ras_window_len) {
630 start = ras->ras_next_readahead;
631 end = ras->ras_window_start + ras->ras_window_len - 1;
634 unsigned long rpc_boundary;
636 * Align RA window to an optimal boundary.
638 * XXX This would be better to align to cl_max_pages_per_rpc
639 * instead of PTLRPC_MAX_BRW_PAGES, because the RPC size may
640 * be aligned to the RAID stripe size in the future and that
641 * is more important than the RPC size.
643 /* Note: we only trim the RPC, instead of extending the RPC
644 * to the boundary, so to avoid reading too much pages during
646 rpc_boundary = ((end + 1) & (~(PTLRPC_MAX_BRW_PAGES - 1)));
647 if (rpc_boundary > 0)
650 if (rpc_boundary > start)
653 /* Truncate RA window to end of file */
654 end = min(end, (unsigned long)((kms - 1) >> PAGE_CACHE_SHIFT));
656 ras->ras_next_readahead = max(end, end + 1);
659 ria->ria_start = start;
661 /* If stride I/O mode is detected, get stride window*/
662 if (stride_io_mode(ras)) {
663 ria->ria_stoff = ras->ras_stride_offset;
664 ria->ria_length = ras->ras_stride_length;
665 ria->ria_pages = ras->ras_stride_pages;
667 spin_unlock(&ras->ras_lock);
670 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
673 len = ria_page_count(ria);
675 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
679 CDEBUG(D_READA, DFID": ria: %lu/%lu, bead: %lu/%lu, hit: %d\n",
680 PFID(lu_object_fid(&clob->co_lu)),
681 ria->ria_start, ria->ria_end,
682 bead == NULL ? 0 : bead->lrr_start,
683 bead == NULL ? 0 : bead->lrr_count,
686 /* at least to extend the readahead window to cover current read */
687 if (!hit && bead != NULL &&
688 bead->lrr_start + bead->lrr_count > ria->ria_start) {
689 /* to the end of current read window. */
690 mlen = bead->lrr_start + bead->lrr_count - ria->ria_start;
691 /* trim to RPC boundary */
692 start = ria->ria_start & (PTLRPC_MAX_BRW_PAGES - 1);
693 mlen = min(mlen, PTLRPC_MAX_BRW_PAGES - start);
696 reserved = ll_ra_count_get(ll_i2sbi(inode), ria, len, mlen);
698 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
700 CDEBUG(D_READA, "reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
702 atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
703 ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
705 ret = ll_read_ahead_pages(env, io, queue, ria, &reserved, &ra_end);
708 ll_ra_count_put(ll_i2sbi(inode), reserved);
710 if (ra_end == end + 1 && ra_end == (kms >> PAGE_CACHE_SHIFT))
711 ll_ra_stats_inc(inode, RA_STAT_EOF);
713 /* if we didn't get to the end of the region we reserved from
714 * the ras we need to go back and update the ras so that the
715 * next read-ahead tries from where we left off. we only do so
716 * if the region we failed to issue read-ahead on is still ahead
717 * of the app and behind the next index to start read-ahead from */
718 CDEBUG(D_READA, "ra_end %lu end %lu stride end %lu \n",
719 ra_end, end, ria->ria_end);
721 if (ra_end != end + 1) {
722 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
723 spin_lock(&ras->ras_lock);
724 if (ra_end < ras->ras_next_readahead &&
725 index_in_window(ra_end, ras->ras_window_start, 0,
726 ras->ras_window_len)) {
727 ras->ras_next_readahead = ra_end;
730 spin_unlock(&ras->ras_lock);
736 static void ras_set_start(struct inode *inode, struct ll_readahead_state *ras,
739 ras->ras_window_start = index & (~(RAS_INCREASE_STEP(inode) - 1));
742 /* called with the ras_lock held or from places where it doesn't matter */
743 static void ras_reset(struct inode *inode, struct ll_readahead_state *ras,
746 ras->ras_last_readpage = index;
747 ras->ras_consecutive_requests = 0;
748 ras->ras_consecutive_pages = 0;
749 ras->ras_window_len = 0;
750 ras_set_start(inode, ras, index);
751 ras->ras_next_readahead = max(ras->ras_window_start, index);
756 /* called with the ras_lock held or from places where it doesn't matter */
757 static void ras_stride_reset(struct ll_readahead_state *ras)
759 ras->ras_consecutive_stride_requests = 0;
760 ras->ras_stride_length = 0;
761 ras->ras_stride_pages = 0;
765 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
767 spin_lock_init(&ras->ras_lock);
768 ras_reset(inode, ras, 0);
769 ras->ras_requests = 0;
770 CFS_INIT_LIST_HEAD(&ras->ras_read_beads);
774 * Check whether the read request is in the stride window.
775 * If it is in the stride window, return 1, otherwise return 0.
777 static int index_in_stride_window(struct ll_readahead_state *ras,
780 unsigned long stride_gap;
782 if (ras->ras_stride_length == 0 || ras->ras_stride_pages == 0 ||
783 ras->ras_stride_pages == ras->ras_stride_length)
786 stride_gap = index - ras->ras_last_readpage - 1;
788 /* If it is contiguous read */
790 return ras->ras_consecutive_pages + 1 <= ras->ras_stride_pages;
792 /* Otherwise check the stride by itself */
793 return (ras->ras_stride_length - ras->ras_stride_pages) == stride_gap &&
794 ras->ras_consecutive_pages == ras->ras_stride_pages;
797 static void ras_update_stride_detector(struct ll_readahead_state *ras,
800 unsigned long stride_gap = index - ras->ras_last_readpage - 1;
802 if (!stride_io_mode(ras) && (stride_gap != 0 ||
803 ras->ras_consecutive_stride_requests == 0)) {
804 ras->ras_stride_pages = ras->ras_consecutive_pages;
805 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
807 LASSERT(ras->ras_request_index == 0);
808 LASSERT(ras->ras_consecutive_stride_requests == 0);
810 if (index <= ras->ras_last_readpage) {
811 /*Reset stride window for forward read*/
812 ras_stride_reset(ras);
816 ras->ras_stride_pages = ras->ras_consecutive_pages;
817 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
824 stride_page_count(struct ll_readahead_state *ras, unsigned long len)
826 return stride_pg_count(ras->ras_stride_offset, ras->ras_stride_length,
827 ras->ras_stride_pages, ras->ras_stride_offset,
831 /* Stride Read-ahead window will be increased inc_len according to
832 * stride I/O pattern */
833 static void ras_stride_increase_window(struct ll_readahead_state *ras,
834 struct ll_ra_info *ra,
835 unsigned long inc_len)
837 unsigned long left, step, window_len;
838 unsigned long stride_len;
840 LASSERT(ras->ras_stride_length > 0);
841 LASSERTF(ras->ras_window_start + ras->ras_window_len
842 >= ras->ras_stride_offset, "window_start %lu, window_len %lu"
843 " stride_offset %lu\n", ras->ras_window_start,
844 ras->ras_window_len, ras->ras_stride_offset);
846 stride_len = ras->ras_window_start + ras->ras_window_len -
847 ras->ras_stride_offset;
849 left = stride_len % ras->ras_stride_length;
850 window_len = ras->ras_window_len - left;
852 if (left < ras->ras_stride_pages)
855 left = ras->ras_stride_pages + inc_len;
857 LASSERT(ras->ras_stride_pages != 0);
859 step = left / ras->ras_stride_pages;
860 left %= ras->ras_stride_pages;
862 window_len += step * ras->ras_stride_length + left;
864 if (stride_page_count(ras, window_len) <= ra->ra_max_pages_per_file)
865 ras->ras_window_len = window_len;
870 static void ras_increase_window(struct inode *inode,
871 struct ll_readahead_state *ras,
872 struct ll_ra_info *ra)
874 /* The stretch of ra-window should be aligned with max rpc_size
875 * but current clio architecture does not support retrieve such
876 * information from lower layer. FIXME later
878 if (stride_io_mode(ras))
879 ras_stride_increase_window(ras, ra, RAS_INCREASE_STEP(inode));
881 ras->ras_window_len = min(ras->ras_window_len +
882 RAS_INCREASE_STEP(inode),
883 ra->ra_max_pages_per_file);
886 void ras_update(struct ll_sb_info *sbi, struct inode *inode,
887 struct ll_readahead_state *ras, unsigned long index,
890 struct ll_ra_info *ra = &sbi->ll_ra_info;
891 int zero = 0, stride_detect = 0, ra_miss = 0;
894 spin_lock(&ras->ras_lock);
896 ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
898 /* reset the read-ahead window in two cases. First when the app seeks
899 * or reads to some other part of the file. Secondly if we get a
900 * read-ahead miss that we think we've previously issued. This can
901 * be a symptom of there being so many read-ahead pages that the VM is
902 * reclaiming it before we get to it. */
903 if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
905 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
906 } else if (!hit && ras->ras_window_len &&
907 index < ras->ras_next_readahead &&
908 index_in_window(index, ras->ras_window_start, 0,
909 ras->ras_window_len)) {
911 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
914 /* On the second access to a file smaller than the tunable
915 * ra_max_read_ahead_whole_pages trigger RA on all pages in the
916 * file up to ra_max_pages_per_file. This is simply a best effort
917 * and only occurs once per open file. Normal RA behavior is reverted
918 * to for subsequent IO. The mmap case does not increment
919 * ras_requests and thus can never trigger this behavior. */
920 if (ras->ras_requests == 2 && !ras->ras_request_index) {
923 kms_pages = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >>
926 CDEBUG(D_READA, "kmsp "LPU64" mwp %lu mp %lu\n", kms_pages,
927 ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages_per_file);
930 kms_pages <= ra->ra_max_read_ahead_whole_pages) {
931 ras->ras_window_start = 0;
932 ras->ras_last_readpage = 0;
933 ras->ras_next_readahead = 0;
934 ras->ras_window_len = min(ra->ra_max_pages_per_file,
935 ra->ra_max_read_ahead_whole_pages);
940 /* check whether it is in stride I/O mode*/
941 if (!index_in_stride_window(ras, index)) {
942 if (ras->ras_consecutive_stride_requests == 0 &&
943 ras->ras_request_index == 0) {
944 ras_update_stride_detector(ras, index);
945 ras->ras_consecutive_stride_requests++;
947 ras_stride_reset(ras);
949 ras_reset(inode, ras, index);
950 ras->ras_consecutive_pages++;
953 ras->ras_consecutive_pages = 0;
954 ras->ras_consecutive_requests = 0;
955 if (++ras->ras_consecutive_stride_requests > 1)
961 if (index_in_stride_window(ras, index) &&
962 stride_io_mode(ras)) {
963 /*If stride-RA hit cache miss, the stride dector
964 *will not be reset to avoid the overhead of
965 *redetecting read-ahead mode */
966 if (index != ras->ras_last_readpage + 1)
967 ras->ras_consecutive_pages = 0;
968 ras_reset(inode, ras, index);
971 /* Reset both stride window and normal RA
973 ras_reset(inode, ras, index);
974 ras->ras_consecutive_pages++;
975 ras_stride_reset(ras);
978 } else if (stride_io_mode(ras)) {
979 /* If this is contiguous read but in stride I/O mode
980 * currently, check whether stride step still is valid,
981 * if invalid, it will reset the stride ra window*/
982 if (!index_in_stride_window(ras, index)) {
983 /* Shrink stride read-ahead window to be zero */
984 ras_stride_reset(ras);
985 ras->ras_window_len = 0;
986 ras->ras_next_readahead = index;
990 ras->ras_consecutive_pages++;
991 ras->ras_last_readpage = index;
992 ras_set_start(inode, ras, index);
994 if (stride_io_mode(ras)) {
995 /* Since stride readahead is sentivite to the offset
996 * of read-ahead, so we use original offset here,
997 * instead of ras_window_start, which is RPC aligned */
998 ras->ras_next_readahead = max(index, ras->ras_next_readahead);
1000 if (ras->ras_next_readahead < ras->ras_window_start)
1001 ras->ras_next_readahead = ras->ras_window_start;
1003 ras->ras_next_readahead = index + 1;
1007 /* Trigger RA in the mmap case where ras_consecutive_requests
1008 * is not incremented and thus can't be used to trigger RA */
1009 if (!ras->ras_window_len && ras->ras_consecutive_pages == 4) {
1010 ras->ras_window_len = RAS_INCREASE_STEP(inode);
1011 GOTO(out_unlock, 0);
1014 /* Initially reset the stride window offset to next_readahead*/
1015 if (ras->ras_consecutive_stride_requests == 2 && stride_detect) {
1017 * Once stride IO mode is detected, next_readahead should be
1018 * reset to make sure next_readahead > stride offset
1020 ras->ras_next_readahead = max(index, ras->ras_next_readahead);
1021 ras->ras_stride_offset = index;
1022 ras->ras_window_len = RAS_INCREASE_STEP(inode);
1025 /* The initial ras_window_len is set to the request size. To avoid
1026 * uselessly reading and discarding pages for random IO the window is
1027 * only increased once per consecutive request received. */
1028 if ((ras->ras_consecutive_requests > 1 || stride_detect) &&
1029 !ras->ras_request_index)
1030 ras_increase_window(inode, ras, ra);
1034 ras->ras_request_index++;
1035 spin_unlock(&ras->ras_lock);
1039 int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
1041 struct inode *inode = vmpage->mapping->host;
1042 struct ll_inode_info *lli = ll_i2info(inode);
1045 struct cl_page *page;
1046 struct cl_object *clob;
1047 struct cl_env_nest nest;
1048 bool redirtied = false;
1049 bool unlocked = false;
1053 LASSERT(PageLocked(vmpage));
1054 LASSERT(!PageWriteback(vmpage));
1056 LASSERT(ll_i2dtexp(inode) != NULL);
1058 env = cl_env_nested_get(&nest);
1060 GOTO(out, result = PTR_ERR(env));
1062 clob = ll_i2info(inode)->lli_clob;
1063 LASSERT(clob != NULL);
1065 io = ccc_env_thread_io(env);
1067 io->ci_ignore_layout = 1;
1068 result = cl_io_init(env, io, CIT_MISC, clob);
1070 page = cl_page_find(env, clob, vmpage->index,
1071 vmpage, CPT_CACHEABLE);
1072 if (!IS_ERR(page)) {
1073 lu_ref_add(&page->cp_reference, "writepage",
1075 cl_page_assume(env, io, page);
1076 result = cl_page_flush(env, io, page);
1079 * Re-dirty page on error so it retries write,
1080 * but not in case when IO has actually
1081 * occurred and completed with an error.
1083 if (!PageError(vmpage)) {
1084 redirty_page_for_writepage(wbc, vmpage);
1089 cl_page_disown(env, io, page);
1091 lu_ref_del(&page->cp_reference,
1092 "writepage", current);
1093 cl_page_put(env, page);
1095 result = PTR_ERR(page);
1098 cl_io_fini(env, io);
1100 if (redirtied && wbc->sync_mode == WB_SYNC_ALL) {
1101 loff_t offset = cl_offset(clob, vmpage->index);
1103 /* Flush page failed because the extent is being written out.
1104 * Wait for the write of extent to be finished to avoid
1105 * breaking kernel which assumes ->writepage should mark
1106 * PageWriteback or clean the page. */
1107 result = cl_sync_file_range(inode, offset,
1108 offset + PAGE_CACHE_SIZE - 1,
1111 /* actually we may have written more than one page.
1112 * decreasing this page because the caller will count
1114 wbc->nr_to_write -= result - 1;
1119 cl_env_nested_put(&nest, env);
1124 if (!lli->lli_async_rc)
1125 lli->lli_async_rc = result;
1126 SetPageError(vmpage);
1128 unlock_page(vmpage);
1133 int ll_writepages(struct address_space *mapping, struct writeback_control *wbc)
1135 struct inode *inode = mapping->host;
1136 struct ll_sb_info *sbi = ll_i2sbi(inode);
1139 enum cl_fsync_mode mode;
1140 int range_whole = 0;
1142 int ignore_layout = 0;
1145 if (wbc->range_cyclic) {
1146 start = mapping->writeback_index << PAGE_CACHE_SHIFT;
1147 end = OBD_OBJECT_EOF;
1149 start = wbc->range_start;
1150 end = wbc->range_end;
1151 if (end == LLONG_MAX) {
1152 end = OBD_OBJECT_EOF;
1153 range_whole = start == 0;
1157 mode = CL_FSYNC_NONE;
1158 if (wbc->sync_mode == WB_SYNC_ALL)
1159 mode = CL_FSYNC_LOCAL;
1161 if (sbi->ll_umounting)
1162 /* if the mountpoint is being umounted, all pages have to be
1163 * evicted to avoid hitting LBUG when truncate_inode_pages()
1164 * is called later on. */
1167 if (cl_i2info(inode)->lli_clob == NULL)
1170 result = cl_sync_file_range(inode, start, end, mode, ignore_layout);
1172 wbc->nr_to_write -= result;
1176 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) {
1177 if (end == OBD_OBJECT_EOF)
1178 mapping->writeback_index = 0;
1180 mapping->writeback_index = (end >> PAGE_CACHE_SHIFT) +1;
1185 int ll_readpage(struct file *file, struct page *vmpage)
1187 struct ll_cl_context *lcc;
1191 lcc = ll_cl_init(file, vmpage);
1193 struct lu_env *env = lcc->lcc_env;
1194 struct cl_io *io = lcc->lcc_io;
1195 struct cl_page *page = lcc->lcc_page;
1197 LASSERT(page->cp_type == CPT_CACHEABLE);
1198 if (likely(!PageUptodate(vmpage))) {
1199 cl_page_assume(env, io, page);
1200 result = cl_io_read_page(env, io, page);
1202 /* Page from a non-object file. */
1203 unlock_page(vmpage);
1208 unlock_page(vmpage);
1209 result = PTR_ERR(lcc);
1214 int ll_page_sync_io(const struct lu_env *env, struct cl_io *io,
1215 struct cl_page *page, enum cl_req_type crt)
1217 struct cl_2queue *queue;
1220 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
1222 queue = &io->ci_queue;
1223 cl_2queue_init_page(queue, page);
1225 result = cl_io_submit_sync(env, io, crt, queue, 0);
1226 LASSERT(cl_page_is_owned(page, io));
1228 if (crt == CRT_READ)
1230 * in CRT_WRITE case page is left locked even in case of
1233 cl_page_list_disown(env, io, &queue->c2_qin);
1234 cl_2queue_fini(env, queue);