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, 2012, 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/pagemap.h>
52 /* current_is_kswapd() */
53 #include <linux/swap.h>
55 #define DEBUG_SUBSYSTEM S_LLITE
57 #include <lustre_lite.h>
58 #include <obd_cksum.h>
59 #include "llite_internal.h"
60 #include <linux/lustre_compat25.h>
63 * Finalizes cl-data before exiting typical address_space operation. Dual to
66 static void ll_cl_fini(struct ll_cl_context *lcc)
68 struct lu_env *env = lcc->lcc_env;
69 struct cl_io *io = lcc->lcc_io;
70 struct cl_page *page = lcc->lcc_page;
72 LASSERT(lcc->lcc_cookie == current);
76 lu_ref_del(&page->cp_reference, "cl_io", io);
77 cl_page_put(env, page);
80 if (io && lcc->lcc_created) {
82 cl_io_unlock(env, io);
83 cl_io_iter_fini(env, io);
86 cl_env_put(env, &lcc->lcc_refcheck);
90 * Initializes common cl-data at the typical address_space operation entry
93 static struct ll_cl_context *ll_cl_init(struct file *file,
94 struct page *vmpage, int create)
96 struct ll_cl_context *lcc;
99 struct cl_object *clob;
105 clob = ll_i2info(vmpage->mapping->host)->lli_clob;
106 LASSERT(clob != NULL);
108 env = cl_env_get(&refcheck);
110 return ERR_CAST(env);
112 lcc = &vvp_env_info(env)->vti_io_ctx;
113 memset(lcc, 0, sizeof(*lcc));
115 lcc->lcc_refcheck = refcheck;
116 lcc->lcc_cookie = current;
118 cio = ccc_env_io(env);
119 io = cio->cui_cl.cis_io;
120 if (io == NULL && create) {
121 struct inode *inode = vmpage->mapping->host;
124 if (mutex_trylock(&inode->i_mutex)) {
125 mutex_unlock(&(inode)->i_mutex);
127 /* this is too bad. Someone is trying to write the
128 * page w/o holding inode mutex. This means we can
129 * add dirty pages into cache during truncate */
130 CERROR("Proc %s is dirting page w/o inode lock, this"
131 "will break truncate.\n", current->comm);
134 return ERR_PTR(-EIO);
138 * Loop-back driver calls ->prepare_write().
139 * methods directly, bypassing file system ->write() operation,
140 * so cl_io has to be created here.
142 io = ccc_env_thread_io(env);
143 ll_io_init(io, file, 1);
145 /* No lock at all for this kind of IO - we can't do it because
146 * we have held page lock, it would cause deadlock.
147 * XXX: This causes poor performance to loop device - One page
149 * In order to get better performance, users should use
150 * lloop driver instead.
152 io->ci_lockreq = CILR_NEVER;
154 pos = (vmpage->index << PAGE_CACHE_SHIFT);
156 /* Create a temp IO to serve write. */
157 result = cl_io_rw_init(env, io, CIT_WRITE, pos, PAGE_CACHE_SIZE);
159 cio->cui_fd = LUSTRE_FPRIVATE(file);
160 cio->cui_iter = NULL;
161 result = cl_io_iter_init(env, io);
163 result = cl_io_lock(env, io);
165 result = cl_io_start(env, io);
168 result = io->ci_result;
169 lcc->lcc_created = 1;
176 struct cl_page *page;
179 LASSERT(io->ci_state == CIS_IO_GOING);
180 LASSERT(cio->cui_fd == LUSTRE_FPRIVATE(file));
181 page = cl_page_find(env, clob, vmpage->index, vmpage,
184 lcc->lcc_page = page;
185 lu_ref_add(&page->cp_reference, "cl_io", io);
188 result = PTR_ERR(page);
192 lcc = ERR_PTR(result);
195 CDEBUG(D_VFSTRACE, "%lu@"DFID" -> %d %p %p\n",
196 vmpage->index, PFID(lu_object_fid(&clob->co_lu)), result,
201 static struct ll_cl_context *ll_cl_get(void)
203 struct ll_cl_context *lcc;
207 env = cl_env_get(&refcheck);
208 LASSERT(!IS_ERR(env));
209 lcc = &vvp_env_info(env)->vti_io_ctx;
210 LASSERT(env == lcc->lcc_env);
211 LASSERT(current == lcc->lcc_cookie);
212 cl_env_put(env, &refcheck);
214 /* env has got in ll_cl_init, so it is still usable. */
219 * ->prepare_write() address space operation called by generic_file_write()
220 * for every page during write.
222 int ll_prepare_write(struct file *file, struct page *vmpage, unsigned from,
225 struct ll_cl_context *lcc;
228 lcc = ll_cl_init(file, vmpage, 1);
230 struct lu_env *env = lcc->lcc_env;
231 struct cl_io *io = lcc->lcc_io;
232 struct cl_page *page = lcc->lcc_page;
234 cl_page_assume(env, io, page);
236 result = cl_io_prepare_write(env, io, page, from, to);
239 * Add a reference, so that page is not evicted from
240 * the cache until ->commit_write() is called.
243 lu_ref_add(&page->cp_reference, "prepare_write",
246 cl_page_unassume(env, io, page);
249 /* returning 0 in prepare assumes commit must be called
252 result = PTR_ERR(lcc);
257 int ll_commit_write(struct file *file, struct page *vmpage, unsigned from,
260 struct ll_cl_context *lcc;
263 struct cl_page *page;
268 page = lcc->lcc_page;
271 LASSERT(cl_page_is_owned(page, io));
273 if (from != to) /* handle short write case. */
274 result = cl_io_commit_write(env, io, page, from, to);
275 if (cl_page_is_owned(page, io))
276 cl_page_unassume(env, io, page);
279 * Release reference acquired by ll_prepare_write().
281 lu_ref_del(&page->cp_reference, "prepare_write", current);
282 cl_page_put(env, page);
287 struct obd_capa *cl_capa_lookup(struct inode *inode, enum cl_req_type crt)
291 opc = crt == CRT_WRITE ? CAPA_OPC_OSS_WRITE : CAPA_OPC_OSS_RW;
292 return ll_osscapa_get(inode, opc);
295 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which);
298 * Get readahead pages from the filesystem readahead pool of the client for a
301 * /param sbi superblock for filesystem readahead state ll_ra_info
302 * /param ria per-thread readahead state
303 * /param pages number of pages requested for readahead for the thread.
305 * WARNING: This algorithm is used to reduce contention on sbi->ll_lock.
306 * It should work well if the ra_max_pages is much greater than the single
307 * file's read-ahead window, and not too many threads contending for
308 * these readahead pages.
310 * TODO: There may be a 'global sync problem' if many threads are trying
311 * to get an ra budget that is larger than the remaining readahead pages
312 * and reach here at exactly the same time. They will compute /a ret to
313 * consume the remaining pages, but will fail at atomic_add_return() and
314 * get a zero ra window, although there is still ra space remaining. - Jay */
316 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi,
317 struct ra_io_arg *ria,
320 struct ll_ra_info *ra = &sbi->ll_ra_info;
323 /* If read-ahead pages left are less than 1M, do not do read-ahead,
324 * otherwise it will form small read RPC(< 1M), which hurt server
325 * performance a lot. */
326 ret = min(ra->ra_max_pages - atomic_read(&ra->ra_cur_pages), pages);
327 if (ret < 0 || ret < min_t(long, PTLRPC_MAX_BRW_PAGES, pages))
330 /* If the non-strided (ria_pages == 0) readahead window
331 * (ria_start + ret) has grown across an RPC boundary, then trim
332 * readahead size by the amount beyond the RPC so it ends on an
333 * RPC boundary. If the readahead window is already ending on
334 * an RPC boundary (beyond_rpc == 0), or smaller than a full
335 * RPC (beyond_rpc < ret) the readahead size is unchanged.
336 * The (beyond_rpc != 0) check is skipped since the conditional
337 * branch is more expensive than subtracting zero from the result.
339 * Strided read is left unaligned to avoid small fragments beyond
340 * the RPC boundary from needing an extra read RPC. */
341 if (ria->ria_pages == 0) {
342 long beyond_rpc = (ria->ria_start + ret) % PTLRPC_MAX_BRW_PAGES;
343 if (/* beyond_rpc != 0 && */ beyond_rpc < ret)
347 if (atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) {
348 atomic_sub(ret, &ra->ra_cur_pages);
356 void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
358 struct ll_ra_info *ra = &sbi->ll_ra_info;
359 atomic_sub(len, &ra->ra_cur_pages);
362 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which)
364 LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
365 lprocfs_counter_incr(sbi->ll_ra_stats, which);
368 void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which)
370 struct ll_sb_info *sbi = ll_i2sbi(mapping->host);
371 ll_ra_stats_inc_sbi(sbi, which);
374 #define RAS_CDEBUG(ras) \
376 "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu r %lu ri %lu" \
377 "csr %lu sf %lu sp %lu sl %lu \n", \
378 ras->ras_last_readpage, ras->ras_consecutive_requests, \
379 ras->ras_consecutive_pages, ras->ras_window_start, \
380 ras->ras_window_len, ras->ras_next_readahead, \
381 ras->ras_requests, ras->ras_request_index, \
382 ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
383 ras->ras_stride_pages, ras->ras_stride_length)
385 static int index_in_window(unsigned long index, unsigned long point,
386 unsigned long before, unsigned long after)
388 unsigned long start = point - before, end = point + after;
395 return start <= index && index <= end;
398 static struct ll_readahead_state *ll_ras_get(struct file *f)
400 struct ll_file_data *fd;
402 fd = LUSTRE_FPRIVATE(f);
406 void ll_ra_read_in(struct file *f, struct ll_ra_read *rar)
408 struct ll_readahead_state *ras;
412 spin_lock(&ras->ras_lock);
414 ras->ras_request_index = 0;
415 ras->ras_consecutive_requests++;
416 rar->lrr_reader = current;
418 list_add(&rar->lrr_linkage, &ras->ras_read_beads);
419 spin_unlock(&ras->ras_lock);
422 void ll_ra_read_ex(struct file *f, struct ll_ra_read *rar)
424 struct ll_readahead_state *ras;
428 spin_lock(&ras->ras_lock);
429 list_del_init(&rar->lrr_linkage);
430 spin_unlock(&ras->ras_lock);
433 static struct ll_ra_read *ll_ra_read_get_locked(struct ll_readahead_state *ras)
435 struct ll_ra_read *scan;
437 list_for_each_entry(scan, &ras->ras_read_beads, lrr_linkage) {
438 if (scan->lrr_reader == current)
444 struct ll_ra_read *ll_ra_read_get(struct file *f)
446 struct ll_readahead_state *ras;
447 struct ll_ra_read *bead;
451 spin_lock(&ras->ras_lock);
452 bead = ll_ra_read_get_locked(ras);
453 spin_unlock(&ras->ras_lock);
457 static int cl_read_ahead_page(const struct lu_env *env, struct cl_io *io,
458 struct cl_page_list *queue, struct cl_page *page,
465 cl_page_assume(env, io, page);
466 lu_ref_add(&page->cp_reference, "ra", current);
467 cp = cl2ccc_page(cl_page_at(page, &vvp_device_type));
468 if (!cp->cpg_defer_uptodate && !PageUptodate(vmpage)) {
469 rc = cl_page_is_under_lock(env, io, page);
471 cp->cpg_defer_uptodate = 1;
473 cl_page_list_add(queue, page);
476 cl_page_delete(env, page);
480 /* skip completed pages */
481 cl_page_unassume(env, io, page);
483 lu_ref_del(&page->cp_reference, "ra", current);
484 cl_page_put(env, page);
489 * Initiates read-ahead of a page with given index.
491 * \retval +ve: page was added to \a queue.
493 * \retval -ENOLCK: there is no extent lock for this part of a file, stop
496 * \retval -ve, 0: page wasn't added to \a queue for other reason.
498 static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io,
499 struct cl_page_list *queue,
500 pgoff_t index, struct address_space *mapping)
503 struct cl_object *clob = ll_i2info(mapping->host)->lli_clob;
504 struct cl_page *page;
505 enum ra_stat which = _NR_RA_STAT; /* keep gcc happy */
506 unsigned int gfp_mask;
508 const char *msg = NULL;
510 gfp_mask = GFP_HIGHUSER & ~__GFP_WAIT;
512 gfp_mask |= __GFP_NOWARN;
514 vmpage = grab_cache_page_nowait(mapping, index);
515 if (vmpage != NULL) {
516 /* Check if vmpage was truncated or reclaimed */
517 if (vmpage->mapping == mapping) {
518 page = cl_page_find(env, clob, vmpage->index,
519 vmpage, CPT_CACHEABLE);
521 rc = cl_read_ahead_page(env, io, queue,
524 which = RA_STAT_FAILED_MATCH;
525 msg = "lock match failed";
528 which = RA_STAT_FAILED_GRAB_PAGE;
529 msg = "cl_page_find failed";
532 which = RA_STAT_WRONG_GRAB_PAGE;
533 msg = "g_c_p_n returned invalid page";
537 page_cache_release(vmpage);
539 which = RA_STAT_FAILED_GRAB_PAGE;
540 msg = "g_c_p_n failed";
543 ll_ra_stats_inc(mapping, which);
544 CDEBUG(D_READA, "%s\n", msg);
549 #define RIA_DEBUG(ria) \
550 CDEBUG(D_READA, "rs %lu re %lu ro %lu rl %lu rp %lu\n", \
551 ria->ria_start, ria->ria_end, ria->ria_stoff, ria->ria_length,\
554 /* Limit this to the blocksize instead of PTLRPC_BRW_MAX_SIZE, since we don't
555 * know what the actual RPC size is. If this needs to change, it makes more
556 * sense to tune the i_blkbits value for the file based on the OSTs it is
557 * striped over, rather than having a constant value for all files here. */
559 /* RAS_INCREASE_STEP should be (1UL << (inode->i_blkbits - PAGE_CACHE_SHIFT)).
560 * Temporarily set RAS_INCREASE_STEP to 1MB. After 4MB RPC is enabled
561 * by default, this should be adjusted corresponding with max_read_ahead_mb
562 * and max_read_ahead_per_file_mb otherwise the readahead budget can be used
563 * up quickly which will affect read performance significantly. See LU-2816 */
564 #define RAS_INCREASE_STEP(inode) (ONE_MB_BRW_SIZE >> PAGE_CACHE_SHIFT)
566 static inline int stride_io_mode(struct ll_readahead_state *ras)
568 return ras->ras_consecutive_stride_requests > 1;
570 /* The function calculates how much pages will be read in
571 * [off, off + length], in such stride IO area,
572 * stride_offset = st_off, stride_length = st_len,
573 * stride_pages = st_pgs
575 * |------------------|*****|------------------|*****|------------|*****|....
578 * |----- st_len -----|
580 * How many pages it should read in such pattern
581 * |-------------------------------------------------------------|
583 * |<------ length ------->|
585 * = |<----->| + |-------------------------------------| + |---|
586 * start_left st_pgs * i end_left
589 stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs,
590 unsigned long off, unsigned long length)
592 __u64 start = off > st_off ? off - st_off : 0;
593 __u64 end = off + length > st_off ? off + length - st_off : 0;
594 unsigned long start_left = 0;
595 unsigned long end_left = 0;
596 unsigned long pg_count;
598 if (st_len == 0 || length == 0 || end == 0)
601 start_left = do_div(start, st_len);
602 if (start_left < st_pgs)
603 start_left = st_pgs - start_left;
607 end_left = do_div(end, st_len);
608 if (end_left > st_pgs)
611 CDEBUG(D_READA, "start "LPU64", end "LPU64" start_left %lu end_left %lu \n",
612 start, end, start_left, end_left);
615 pg_count = end_left - (st_pgs - start_left);
617 pg_count = start_left + st_pgs * (end - start - 1) + end_left;
619 CDEBUG(D_READA, "st_off %lu, st_len %lu st_pgs %lu off %lu length %lu"
620 "pgcount %lu\n", st_off, st_len, st_pgs, off, length, pg_count);
625 static int ria_page_count(struct ra_io_arg *ria)
627 __u64 length = ria->ria_end >= ria->ria_start ?
628 ria->ria_end - ria->ria_start + 1 : 0;
630 return stride_pg_count(ria->ria_stoff, ria->ria_length,
631 ria->ria_pages, ria->ria_start,
635 /*Check whether the index is in the defined ra-window */
636 static int ras_inside_ra_window(unsigned long idx, struct ra_io_arg *ria)
638 /* If ria_length == ria_pages, it means non-stride I/O mode,
639 * idx should always inside read-ahead window in this case
640 * For stride I/O mode, just check whether the idx is inside
642 return ria->ria_length == 0 || ria->ria_length == ria->ria_pages ||
643 (idx >= ria->ria_stoff && (idx - ria->ria_stoff) %
644 ria->ria_length < ria->ria_pages);
647 static int ll_read_ahead_pages(const struct lu_env *env,
648 struct cl_io *io, struct cl_page_list *queue,
649 struct ra_io_arg *ria,
650 unsigned long *reserved_pages,
651 struct address_space *mapping,
652 unsigned long *ra_end)
654 int rc, count = 0, stride_ria;
655 unsigned long page_idx;
657 LASSERT(ria != NULL);
660 stride_ria = ria->ria_length > ria->ria_pages && ria->ria_pages > 0;
661 for (page_idx = ria->ria_start; page_idx <= ria->ria_end &&
662 *reserved_pages > 0; page_idx++) {
663 if (ras_inside_ra_window(page_idx, ria)) {
664 /* If the page is inside the read-ahead window*/
665 rc = ll_read_ahead_page(env, io, queue,
670 } else if (rc == -ENOLCK)
672 } else if (stride_ria) {
673 /* If it is not in the read-ahead window, and it is
674 * read-ahead mode, then check whether it should skip
677 /* FIXME: This assertion only is valid when it is for
678 * forward read-ahead, it will be fixed when backward
679 * read-ahead is implemented */
680 LASSERTF(page_idx > ria->ria_stoff, "Invalid page_idx %lu"
681 "rs %lu re %lu ro %lu rl %lu rp %lu\n", page_idx,
682 ria->ria_start, ria->ria_end, ria->ria_stoff,
683 ria->ria_length, ria->ria_pages);
684 offset = page_idx - ria->ria_stoff;
685 offset = offset % (ria->ria_length);
686 if (offset > ria->ria_pages) {
687 page_idx += ria->ria_length - offset;
688 CDEBUG(D_READA, "i %lu skip %lu \n", page_idx,
689 ria->ria_length - offset);
698 int ll_readahead(const struct lu_env *env, struct cl_io *io,
699 struct ll_readahead_state *ras, struct address_space *mapping,
700 struct cl_page_list *queue, int flags)
702 struct vvp_io *vio = vvp_env_io(env);
703 struct vvp_thread_info *vti = vvp_env_info(env);
704 struct cl_attr *attr = ccc_env_thread_attr(env);
705 unsigned long start = 0, end = 0, reserved;
706 unsigned long ra_end, len;
708 struct ll_ra_read *bead;
709 struct ra_io_arg *ria = &vti->vti_ria;
710 struct ll_inode_info *lli;
711 struct cl_object *clob;
715 inode = mapping->host;
716 lli = ll_i2info(inode);
717 clob = lli->lli_clob;
719 memset(ria, 0, sizeof(*ria));
721 cl_object_attr_lock(clob);
722 ret = cl_object_attr_get(env, clob, attr);
723 cl_object_attr_unlock(clob);
729 ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN);
733 spin_lock(&ras->ras_lock);
734 if (vio->cui_ra_window_set)
735 bead = &vio->cui_bead;
739 /* Enlarge the RA window to encompass the full read */
740 if (bead != NULL && ras->ras_window_start + ras->ras_window_len <
741 bead->lrr_start + bead->lrr_count) {
742 ras->ras_window_len = bead->lrr_start + bead->lrr_count -
743 ras->ras_window_start;
745 /* Reserve a part of the read-ahead window that we'll be issuing */
746 if (ras->ras_window_len) {
747 start = ras->ras_next_readahead;
748 end = ras->ras_window_start + ras->ras_window_len - 1;
751 unsigned long rpc_boundary;
753 * Align RA window to an optimal boundary.
755 * XXX This would be better to align to cl_max_pages_per_rpc
756 * instead of PTLRPC_MAX_BRW_PAGES, because the RPC size may
757 * be aligned to the RAID stripe size in the future and that
758 * is more important than the RPC size.
760 /* Note: we only trim the RPC, instead of extending the RPC
761 * to the boundary, so to avoid reading too much pages during
763 rpc_boundary = ((end + 1) & (~(PTLRPC_MAX_BRW_PAGES - 1)));
764 if (rpc_boundary > 0)
767 if (rpc_boundary > start)
770 /* Truncate RA window to end of file */
771 end = min(end, (unsigned long)((kms - 1) >> PAGE_CACHE_SHIFT));
773 ras->ras_next_readahead = max(end, end + 1);
776 ria->ria_start = start;
778 /* If stride I/O mode is detected, get stride window*/
779 if (stride_io_mode(ras)) {
780 ria->ria_stoff = ras->ras_stride_offset;
781 ria->ria_length = ras->ras_stride_length;
782 ria->ria_pages = ras->ras_stride_pages;
784 spin_unlock(&ras->ras_lock);
787 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
790 len = ria_page_count(ria);
794 reserved = ll_ra_count_get(ll_i2sbi(inode), ria, len);
796 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
798 CDEBUG(D_READA, "reserved page %lu ra_cur %d ra_max %lu\n", reserved,
799 atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
800 ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
802 ret = ll_read_ahead_pages(env, io, queue,
803 ria, &reserved, mapping, &ra_end);
805 LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
807 ll_ra_count_put(ll_i2sbi(inode), reserved);
809 if (ra_end == end + 1 && ra_end == (kms >> PAGE_CACHE_SHIFT))
810 ll_ra_stats_inc(mapping, RA_STAT_EOF);
812 /* if we didn't get to the end of the region we reserved from
813 * the ras we need to go back and update the ras so that the
814 * next read-ahead tries from where we left off. we only do so
815 * if the region we failed to issue read-ahead on is still ahead
816 * of the app and behind the next index to start read-ahead from */
817 CDEBUG(D_READA, "ra_end %lu end %lu stride end %lu \n",
818 ra_end, end, ria->ria_end);
820 if (ra_end != end + 1) {
821 spin_lock(&ras->ras_lock);
822 if (ra_end < ras->ras_next_readahead &&
823 index_in_window(ra_end, ras->ras_window_start, 0,
824 ras->ras_window_len)) {
825 ras->ras_next_readahead = ra_end;
828 spin_unlock(&ras->ras_lock);
834 static void ras_set_start(struct inode *inode, struct ll_readahead_state *ras,
837 ras->ras_window_start = index & (~(RAS_INCREASE_STEP(inode) - 1));
840 /* called with the ras_lock held or from places where it doesn't matter */
841 static void ras_reset(struct inode *inode, struct ll_readahead_state *ras,
844 ras->ras_last_readpage = index;
845 ras->ras_consecutive_requests = 0;
846 ras->ras_consecutive_pages = 0;
847 ras->ras_window_len = 0;
848 ras_set_start(inode, ras, index);
849 ras->ras_next_readahead = max(ras->ras_window_start, index);
854 /* called with the ras_lock held or from places where it doesn't matter */
855 static void ras_stride_reset(struct ll_readahead_state *ras)
857 ras->ras_consecutive_stride_requests = 0;
858 ras->ras_stride_length = 0;
859 ras->ras_stride_pages = 0;
863 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
865 spin_lock_init(&ras->ras_lock);
866 ras_reset(inode, ras, 0);
867 ras->ras_requests = 0;
868 INIT_LIST_HEAD(&ras->ras_read_beads);
872 * Check whether the read request is in the stride window.
873 * If it is in the stride window, return 1, otherwise return 0.
875 static int index_in_stride_window(struct ll_readahead_state *ras,
878 unsigned long stride_gap;
880 if (ras->ras_stride_length == 0 || ras->ras_stride_pages == 0 ||
881 ras->ras_stride_pages == ras->ras_stride_length)
884 stride_gap = index - ras->ras_last_readpage - 1;
886 /* If it is contiguous read */
888 return ras->ras_consecutive_pages + 1 <= ras->ras_stride_pages;
890 /* Otherwise check the stride by itself */
891 return (ras->ras_stride_length - ras->ras_stride_pages) == stride_gap &&
892 ras->ras_consecutive_pages == ras->ras_stride_pages;
895 static void ras_update_stride_detector(struct ll_readahead_state *ras,
898 unsigned long stride_gap = index - ras->ras_last_readpage - 1;
900 if (!stride_io_mode(ras) && (stride_gap != 0 ||
901 ras->ras_consecutive_stride_requests == 0)) {
902 ras->ras_stride_pages = ras->ras_consecutive_pages;
903 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
905 LASSERT(ras->ras_request_index == 0);
906 LASSERT(ras->ras_consecutive_stride_requests == 0);
908 if (index <= ras->ras_last_readpage) {
909 /*Reset stride window for forward read*/
910 ras_stride_reset(ras);
914 ras->ras_stride_pages = ras->ras_consecutive_pages;
915 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
922 stride_page_count(struct ll_readahead_state *ras, unsigned long len)
924 return stride_pg_count(ras->ras_stride_offset, ras->ras_stride_length,
925 ras->ras_stride_pages, ras->ras_stride_offset,
929 /* Stride Read-ahead window will be increased inc_len according to
930 * stride I/O pattern */
931 static void ras_stride_increase_window(struct ll_readahead_state *ras,
932 struct ll_ra_info *ra,
933 unsigned long inc_len)
935 unsigned long left, step, window_len;
936 unsigned long stride_len;
938 LASSERT(ras->ras_stride_length > 0);
939 LASSERTF(ras->ras_window_start + ras->ras_window_len
940 >= ras->ras_stride_offset, "window_start %lu, window_len %lu"
941 " stride_offset %lu\n", ras->ras_window_start,
942 ras->ras_window_len, ras->ras_stride_offset);
944 stride_len = ras->ras_window_start + ras->ras_window_len -
945 ras->ras_stride_offset;
947 left = stride_len % ras->ras_stride_length;
948 window_len = ras->ras_window_len - left;
950 if (left < ras->ras_stride_pages)
953 left = ras->ras_stride_pages + inc_len;
955 LASSERT(ras->ras_stride_pages != 0);
957 step = left / ras->ras_stride_pages;
958 left %= ras->ras_stride_pages;
960 window_len += step * ras->ras_stride_length + left;
962 if (stride_page_count(ras, window_len) <= ra->ra_max_pages_per_file)
963 ras->ras_window_len = window_len;
968 static void ras_increase_window(struct inode *inode,
969 struct ll_readahead_state *ras,
970 struct ll_ra_info *ra)
972 /* The stretch of ra-window should be aligned with max rpc_size
973 * but current clio architecture does not support retrieve such
974 * information from lower layer. FIXME later
976 if (stride_io_mode(ras))
977 ras_stride_increase_window(ras, ra, RAS_INCREASE_STEP(inode));
979 ras->ras_window_len = min(ras->ras_window_len +
980 RAS_INCREASE_STEP(inode),
981 ra->ra_max_pages_per_file);
984 void ras_update(struct ll_sb_info *sbi, struct inode *inode,
985 struct ll_readahead_state *ras, unsigned long index,
988 struct ll_ra_info *ra = &sbi->ll_ra_info;
989 int zero = 0, stride_detect = 0, ra_miss = 0;
991 spin_lock(&ras->ras_lock);
993 ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
995 /* reset the read-ahead window in two cases. First when the app seeks
996 * or reads to some other part of the file. Secondly if we get a
997 * read-ahead miss that we think we've previously issued. This can
998 * be a symptom of there being so many read-ahead pages that the VM is
999 * reclaiming it before we get to it. */
1000 if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
1002 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
1003 } else if (!hit && ras->ras_window_len &&
1004 index < ras->ras_next_readahead &&
1005 index_in_window(index, ras->ras_window_start, 0,
1006 ras->ras_window_len)) {
1008 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
1011 /* On the second access to a file smaller than the tunable
1012 * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1013 * file up to ra_max_pages_per_file. This is simply a best effort
1014 * and only occurs once per open file. Normal RA behavior is reverted
1015 * to for subsequent IO. The mmap case does not increment
1016 * ras_requests and thus can never trigger this behavior. */
1017 if (ras->ras_requests == 2 && !ras->ras_request_index) {
1020 kms_pages = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >>
1023 CDEBUG(D_READA, "kmsp "LPU64" mwp %lu mp %lu\n", kms_pages,
1024 ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages_per_file);
1027 kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1028 ras->ras_window_start = 0;
1029 ras->ras_last_readpage = 0;
1030 ras->ras_next_readahead = 0;
1031 ras->ras_window_len = min(ra->ra_max_pages_per_file,
1032 ra->ra_max_read_ahead_whole_pages);
1033 GOTO(out_unlock, 0);
1037 /* check whether it is in stride I/O mode*/
1038 if (!index_in_stride_window(ras, index)) {
1039 if (ras->ras_consecutive_stride_requests == 0 &&
1040 ras->ras_request_index == 0) {
1041 ras_update_stride_detector(ras, index);
1042 ras->ras_consecutive_stride_requests++;
1044 ras_stride_reset(ras);
1046 ras_reset(inode, ras, index);
1047 ras->ras_consecutive_pages++;
1048 GOTO(out_unlock, 0);
1050 ras->ras_consecutive_pages = 0;
1051 ras->ras_consecutive_requests = 0;
1052 if (++ras->ras_consecutive_stride_requests > 1)
1058 if (index_in_stride_window(ras, index) &&
1059 stride_io_mode(ras)) {
1060 /*If stride-RA hit cache miss, the stride dector
1061 *will not be reset to avoid the overhead of
1062 *redetecting read-ahead mode */
1063 if (index != ras->ras_last_readpage + 1)
1064 ras->ras_consecutive_pages = 0;
1065 ras_reset(inode, ras, index);
1068 /* Reset both stride window and normal RA
1070 ras_reset(inode, ras, index);
1071 ras->ras_consecutive_pages++;
1072 ras_stride_reset(ras);
1073 GOTO(out_unlock, 0);
1075 } else if (stride_io_mode(ras)) {
1076 /* If this is contiguous read but in stride I/O mode
1077 * currently, check whether stride step still is valid,
1078 * if invalid, it will reset the stride ra window*/
1079 if (!index_in_stride_window(ras, index)) {
1080 /* Shrink stride read-ahead window to be zero */
1081 ras_stride_reset(ras);
1082 ras->ras_window_len = 0;
1083 ras->ras_next_readahead = index;
1087 ras->ras_consecutive_pages++;
1088 ras->ras_last_readpage = index;
1089 ras_set_start(inode, ras, index);
1091 if (stride_io_mode(ras))
1092 /* Since stride readahead is sensitive to the offset
1093 * of read-ahead, so we use original offset here,
1094 * instead of ras_window_start, which is RPC aligned */
1095 ras->ras_next_readahead = max(index, ras->ras_next_readahead);
1097 ras->ras_next_readahead = max(ras->ras_window_start,
1098 ras->ras_next_readahead);
1101 /* Trigger RA in the mmap case where ras_consecutive_requests
1102 * is not incremented and thus can't be used to trigger RA */
1103 if (!ras->ras_window_len && ras->ras_consecutive_pages == 4) {
1104 ras->ras_window_len = RAS_INCREASE_STEP(inode);
1105 GOTO(out_unlock, 0);
1108 /* Initially reset the stride window offset to next_readahead*/
1109 if (ras->ras_consecutive_stride_requests == 2 && stride_detect) {
1111 * Once stride IO mode is detected, next_readahead should be
1112 * reset to make sure next_readahead > stride offset
1114 ras->ras_next_readahead = max(index, ras->ras_next_readahead);
1115 ras->ras_stride_offset = index;
1116 ras->ras_window_len = RAS_INCREASE_STEP(inode);
1119 /* The initial ras_window_len is set to the request size. To avoid
1120 * uselessly reading and discarding pages for random IO the window is
1121 * only increased once per consecutive request received. */
1122 if ((ras->ras_consecutive_requests > 1 || stride_detect) &&
1123 !ras->ras_request_index)
1124 ras_increase_window(inode, ras, ra);
1127 ras->ras_request_index++;
1128 spin_unlock(&ras->ras_lock);
1132 int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
1134 struct inode *inode = vmpage->mapping->host;
1135 struct ll_inode_info *lli = ll_i2info(inode);
1138 struct cl_page *page;
1139 struct cl_object *clob;
1140 struct cl_env_nest nest;
1141 bool redirtied = false;
1142 bool unlocked = false;
1145 LASSERT(PageLocked(vmpage));
1146 LASSERT(!PageWriteback(vmpage));
1148 LASSERT(ll_i2dtexp(inode) != NULL);
1150 env = cl_env_nested_get(&nest);
1152 GOTO(out, result = PTR_ERR(env));
1154 clob = ll_i2info(inode)->lli_clob;
1155 LASSERT(clob != NULL);
1157 io = ccc_env_thread_io(env);
1159 io->ci_ignore_layout = 1;
1160 result = cl_io_init(env, io, CIT_MISC, clob);
1162 page = cl_page_find(env, clob, vmpage->index,
1163 vmpage, CPT_CACHEABLE);
1164 if (!IS_ERR(page)) {
1165 lu_ref_add(&page->cp_reference, "writepage",
1167 cl_page_assume(env, io, page);
1168 result = cl_page_flush(env, io, page);
1171 * Re-dirty page on error so it retries write,
1172 * but not in case when IO has actually
1173 * occurred and completed with an error.
1175 if (!PageError(vmpage)) {
1176 redirty_page_for_writepage(wbc, vmpage);
1181 cl_page_disown(env, io, page);
1183 lu_ref_del(&page->cp_reference,
1184 "writepage", current);
1185 cl_page_put(env, page);
1187 result = PTR_ERR(page);
1190 cl_io_fini(env, io);
1192 if (redirtied && wbc->sync_mode == WB_SYNC_ALL) {
1193 loff_t offset = cl_offset(clob, vmpage->index);
1195 /* Flush page failed because the extent is being written out.
1196 * Wait for the write of extent to be finished to avoid
1197 * breaking kernel which assumes ->writepage should mark
1198 * PageWriteback or clean the page. */
1199 result = cl_sync_file_range(inode, offset,
1200 offset + PAGE_CACHE_SIZE - 1,
1203 /* actually we may have written more than one page.
1204 * decreasing this page because the caller will count
1206 wbc->nr_to_write -= result - 1;
1211 cl_env_nested_put(&nest, env);
1216 if (!lli->lli_async_rc)
1217 lli->lli_async_rc = result;
1218 SetPageError(vmpage);
1220 unlock_page(vmpage);
1225 int ll_writepages(struct address_space *mapping, struct writeback_control *wbc)
1227 struct inode *inode = mapping->host;
1228 struct ll_sb_info *sbi = ll_i2sbi(inode);
1231 enum cl_fsync_mode mode;
1232 int range_whole = 0;
1234 int ignore_layout = 0;
1236 if (wbc->range_cyclic) {
1237 start = mapping->writeback_index << PAGE_CACHE_SHIFT;
1238 end = OBD_OBJECT_EOF;
1240 start = wbc->range_start;
1241 end = wbc->range_end;
1242 if (end == LLONG_MAX) {
1243 end = OBD_OBJECT_EOF;
1244 range_whole = start == 0;
1248 mode = CL_FSYNC_NONE;
1249 if (wbc->sync_mode == WB_SYNC_ALL)
1250 mode = CL_FSYNC_LOCAL;
1252 if (sbi->ll_umounting)
1253 /* if the mountpoint is being umounted, all pages have to be
1254 * evicted to avoid hitting LBUG when truncate_inode_pages()
1255 * is called later on. */
1257 result = cl_sync_file_range(inode, start, end, mode, ignore_layout);
1259 wbc->nr_to_write -= result;
1263 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) {
1264 if (end == OBD_OBJECT_EOF)
1265 end = i_size_read(inode);
1266 mapping->writeback_index = (end >> PAGE_CACHE_SHIFT) + 1;
1271 int ll_readpage(struct file *file, struct page *vmpage)
1273 struct ll_cl_context *lcc;
1276 lcc = ll_cl_init(file, vmpage, 0);
1278 struct lu_env *env = lcc->lcc_env;
1279 struct cl_io *io = lcc->lcc_io;
1280 struct cl_page *page = lcc->lcc_page;
1282 LASSERT(page->cp_type == CPT_CACHEABLE);
1283 if (likely(!PageUptodate(vmpage))) {
1284 cl_page_assume(env, io, page);
1285 result = cl_io_read_page(env, io, page);
1287 /* Page from a non-object file. */
1288 unlock_page(vmpage);
1293 unlock_page(vmpage);
1294 result = PTR_ERR(lcc);