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/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 static 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 static struct ll_cl_context *ll_cl_init(struct file *file,
97 struct page *vmpage, int create)
99 struct ll_cl_context *lcc;
102 struct cl_object *clob;
108 clob = ll_i2info(vmpage->mapping->host)->lli_clob;
109 LASSERT(clob != NULL);
111 env = cl_env_get(&refcheck);
113 return ERR_PTR(PTR_ERR(env));
115 lcc = &vvp_env_info(env)->vti_io_ctx;
116 memset(lcc, 0, sizeof(*lcc));
118 lcc->lcc_refcheck = refcheck;
119 lcc->lcc_cookie = current;
121 cio = ccc_env_io(env);
122 io = cio->cui_cl.cis_io;
123 if (io == NULL && create) {
124 struct inode *inode = vmpage->mapping->host;
127 if (mutex_trylock(&inode->i_mutex)) {
128 mutex_unlock(&(inode)->i_mutex);
130 /* this is too bad. Someone is trying to write the
131 * page w/o holding inode mutex. This means we can
132 * add dirty pages into cache during truncate */
133 CERROR("Proc %s is dirting page w/o inode lock, this"
134 "will break truncate.\n", cfs_current()->comm);
135 libcfs_debug_dumpstack(NULL);
137 return ERR_PTR(-EIO);
141 * Loop-back driver calls ->prepare_write() and ->sendfile()
142 * methods directly, bypassing file system ->write() operation,
143 * so cl_io has to be created here.
145 io = ccc_env_thread_io(env);
146 ll_io_init(io, file, 1);
148 /* No lock at all for this kind of IO - we can't do it because
149 * we have held page lock, it would cause deadlock.
150 * XXX: This causes poor performance to loop device - One page
152 * In order to get better performance, users should use
153 * lloop driver instead.
155 io->ci_lockreq = CILR_NEVER;
157 pos = (vmpage->index << CFS_PAGE_SHIFT);
159 /* Create a temp IO to serve write. */
160 result = cl_io_rw_init(env, io, CIT_WRITE, pos, CFS_PAGE_SIZE);
162 cio->cui_fd = LUSTRE_FPRIVATE(file);
165 result = cl_io_iter_init(env, io);
167 result = cl_io_lock(env, io);
169 result = cl_io_start(env, io);
172 result = io->ci_result;
173 lcc->lcc_created = 1;
180 struct cl_page *page;
183 LASSERT(io->ci_state == CIS_IO_GOING);
184 LASSERT(cio->cui_fd == LUSTRE_FPRIVATE(file));
185 page = cl_page_find(env, clob, vmpage->index, vmpage,
188 lcc->lcc_page = page;
189 lu_ref_add(&page->cp_reference, "cl_io", io);
192 result = PTR_ERR(page);
196 lcc = ERR_PTR(result);
199 CDEBUG(D_VFSTRACE, "%lu@"DFID" -> %d %p %p\n",
200 vmpage->index, PFID(lu_object_fid(&clob->co_lu)), result,
205 static struct ll_cl_context *ll_cl_get(void)
207 struct ll_cl_context *lcc;
211 env = cl_env_get(&refcheck);
212 LASSERT(!IS_ERR(env));
213 lcc = &vvp_env_info(env)->vti_io_ctx;
214 LASSERT(env == lcc->lcc_env);
215 LASSERT(current == lcc->lcc_cookie);
216 cl_env_put(env, &refcheck);
218 /* env has got in ll_cl_init, so it is still usable. */
223 * ->prepare_write() address space operation called by generic_file_write()
224 * for every page during write.
226 int ll_prepare_write(struct file *file, struct page *vmpage, unsigned from,
229 struct ll_cl_context *lcc;
233 lcc = ll_cl_init(file, vmpage, 1);
235 struct lu_env *env = lcc->lcc_env;
236 struct cl_io *io = lcc->lcc_io;
237 struct cl_page *page = lcc->lcc_page;
239 cl_page_assume(env, io, page);
240 if (cl_io_is_append(io)) {
241 struct cl_object *obj = io->ci_obj;
242 struct inode *inode = ccc_object_inode(obj);
244 * In VFS file->page write loop, for appending, the
245 * write offset might be reset according to the new
246 * file size before holding i_mutex. So crw_pos should
247 * be reset here. BUG:17711.
249 io->u.ci_wr.wr.crw_pos = i_size_read(inode);
251 result = cl_io_prepare_write(env, io, page, from, to);
254 * Add a reference, so that page is not evicted from
255 * the cache until ->commit_write() is called.
258 lu_ref_add(&page->cp_reference, "prepare_write",
261 cl_page_unassume(env, io, page);
264 /* returning 0 in prepare assumes commit must be called
267 result = PTR_ERR(lcc);
272 int ll_commit_write(struct file *file, struct page *vmpage, unsigned from,
275 struct ll_cl_context *lcc;
278 struct cl_page *page;
284 page = lcc->lcc_page;
287 LASSERT(cl_page_is_owned(page, io));
289 if (from != to) /* handle short write case. */
290 result = cl_io_commit_write(env, io, page, from, to);
291 if (cl_page_is_owned(page, io))
292 cl_page_unassume(env, io, page);
295 * Release reference acquired by ll_prepare_write().
297 lu_ref_del(&page->cp_reference, "prepare_write", cfs_current());
298 cl_page_put(env, page);
303 struct obd_capa *cl_capa_lookup(struct inode *inode, enum cl_req_type crt)
307 opc = crt == CRT_WRITE ? CAPA_OPC_OSS_WRITE : CAPA_OPC_OSS_RW;
308 return ll_osscapa_get(inode, opc);
311 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which);
314 * Get readahead pages from the filesystem readahead pool of the client for a
317 * /param sbi superblock for filesystem readahead state ll_ra_info
318 * /param ria per-thread readahead state
319 * /param pages number of pages requested for readahead for the thread.
321 * WARNING: This algorithm is used to reduce contention on sbi->ll_lock.
322 * It should work well if the ra_max_pages is much greater than the single
323 * file's read-ahead window, and not too many threads contending for
324 * these readahead pages.
326 * TODO: There may be a 'global sync problem' if many threads are trying
327 * to get an ra budget that is larger than the remaining readahead pages
328 * and reach here at exactly the same time. They will compute /a ret to
329 * consume the remaining pages, but will fail at atomic_add_return() and
330 * get a zero ra window, although there is still ra space remaining. - Jay */
332 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi,
333 struct ra_io_arg *ria,
336 struct ll_ra_info *ra = &sbi->ll_ra_info;
340 /* If read-ahead pages left are less than 1M, do not do read-ahead,
341 * otherwise it will form small read RPC(< 1M), which hurt server
342 * performance a lot. */
343 ret = min(ra->ra_max_pages - cfs_atomic_read(&ra->ra_cur_pages), pages);
344 if (ret < 0 || ret < min_t(long, PTLRPC_MAX_BRW_PAGES, pages))
347 /* If the non-strided (ria_pages == 0) readahead window
348 * (ria_start + ret) has grown across an RPC boundary, then trim
349 * readahead size by the amount beyond the RPC so it ends on an
350 * RPC boundary. If the readahead window is already ending on
351 * an RPC boundary (beyond_rpc == 0), or smaller than a full
352 * RPC (beyond_rpc < ret) the readahead size is unchanged.
353 * The (beyond_rpc != 0) check is skipped since the conditional
354 * branch is more expensive than subtracting zero from the result.
356 * Strided read is left unaligned to avoid small fragments beyond
357 * the RPC boundary from needing an extra read RPC. */
358 if (ria->ria_pages == 0) {
359 long beyond_rpc = (ria->ria_start + ret) % PTLRPC_MAX_BRW_PAGES;
360 if (/* beyond_rpc != 0 && */ beyond_rpc < ret)
364 if (cfs_atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) {
365 cfs_atomic_sub(ret, &ra->ra_cur_pages);
373 void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
375 struct ll_ra_info *ra = &sbi->ll_ra_info;
376 cfs_atomic_sub(len, &ra->ra_cur_pages);
379 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which)
381 LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
382 lprocfs_counter_incr(sbi->ll_ra_stats, which);
385 void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which)
387 struct ll_sb_info *sbi = ll_i2sbi(mapping->host);
388 ll_ra_stats_inc_sbi(sbi, which);
391 #define RAS_CDEBUG(ras) \
393 "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu r %lu ri %lu" \
394 "csr %lu sf %lu sp %lu sl %lu \n", \
395 ras->ras_last_readpage, ras->ras_consecutive_requests, \
396 ras->ras_consecutive_pages, ras->ras_window_start, \
397 ras->ras_window_len, ras->ras_next_readahead, \
398 ras->ras_requests, ras->ras_request_index, \
399 ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
400 ras->ras_stride_pages, ras->ras_stride_length)
402 static int index_in_window(unsigned long index, unsigned long point,
403 unsigned long before, unsigned long after)
405 unsigned long start = point - before, end = point + after;
412 return start <= index && index <= end;
415 static struct ll_readahead_state *ll_ras_get(struct file *f)
417 struct ll_file_data *fd;
419 fd = LUSTRE_FPRIVATE(f);
423 void ll_ra_read_in(struct file *f, struct ll_ra_read *rar)
425 struct ll_readahead_state *ras;
429 spin_lock(&ras->ras_lock);
431 ras->ras_request_index = 0;
432 ras->ras_consecutive_requests++;
433 rar->lrr_reader = current;
435 cfs_list_add(&rar->lrr_linkage, &ras->ras_read_beads);
436 spin_unlock(&ras->ras_lock);
439 void ll_ra_read_ex(struct file *f, struct ll_ra_read *rar)
441 struct ll_readahead_state *ras;
445 spin_lock(&ras->ras_lock);
446 cfs_list_del_init(&rar->lrr_linkage);
447 spin_unlock(&ras->ras_lock);
450 static struct ll_ra_read *ll_ra_read_get_locked(struct ll_readahead_state *ras)
452 struct ll_ra_read *scan;
454 cfs_list_for_each_entry(scan, &ras->ras_read_beads, lrr_linkage) {
455 if (scan->lrr_reader == current)
461 struct ll_ra_read *ll_ra_read_get(struct file *f)
463 struct ll_readahead_state *ras;
464 struct ll_ra_read *bead;
468 spin_lock(&ras->ras_lock);
469 bead = ll_ra_read_get_locked(ras);
470 spin_unlock(&ras->ras_lock);
474 static int cl_read_ahead_page(const struct lu_env *env, struct cl_io *io,
475 struct cl_page_list *queue, struct cl_page *page,
484 cl_page_assume(env, io, page);
485 lu_ref_add(&page->cp_reference, "ra", cfs_current());
486 cp = cl2ccc_page(cl_page_at(page, &vvp_device_type));
487 if (!cp->cpg_defer_uptodate && !PageUptodate(vmpage)) {
488 rc = cl_page_is_under_lock(env, io, page);
490 cp->cpg_defer_uptodate = 1;
492 cl_page_list_add(queue, page);
495 cl_page_delete(env, page);
499 /* skip completed pages */
500 cl_page_unassume(env, io, page);
502 lu_ref_del(&page->cp_reference, "ra", cfs_current());
503 cl_page_put(env, page);
508 * Initiates read-ahead of a page with given index.
510 * \retval +ve: page was added to \a queue.
512 * \retval -ENOLCK: there is no extent lock for this part of a file, stop
515 * \retval -ve, 0: page wasn't added to \a queue for other reason.
517 static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io,
518 struct cl_page_list *queue,
519 pgoff_t index, struct address_space *mapping)
522 struct cl_object *clob = ll_i2info(mapping->host)->lli_clob;
523 struct cl_page *page;
524 enum ra_stat which = _NR_RA_STAT; /* keep gcc happy */
525 unsigned int gfp_mask;
527 const char *msg = NULL;
531 gfp_mask = GFP_HIGHUSER & ~__GFP_WAIT;
533 gfp_mask |= __GFP_NOWARN;
535 vmpage = grab_cache_page_nowait(mapping, index);
536 if (vmpage != NULL) {
537 /* Check if vmpage was truncated or reclaimed */
538 if (vmpage->mapping == mapping) {
539 page = cl_page_find(env, clob, vmpage->index,
540 vmpage, CPT_CACHEABLE);
542 rc = cl_read_ahead_page(env, io, queue,
545 which = RA_STAT_FAILED_MATCH;
546 msg = "lock match failed";
549 which = RA_STAT_FAILED_GRAB_PAGE;
550 msg = "cl_page_find failed";
553 which = RA_STAT_WRONG_GRAB_PAGE;
554 msg = "g_c_p_n returned invalid page";
558 page_cache_release(vmpage);
560 which = RA_STAT_FAILED_GRAB_PAGE;
561 msg = "g_c_p_n failed";
564 ll_ra_stats_inc(mapping, which);
565 CDEBUG(D_READA, "%s\n", msg);
570 #define RIA_DEBUG(ria) \
571 CDEBUG(D_READA, "rs %lu re %lu ro %lu rl %lu rp %lu\n", \
572 ria->ria_start, ria->ria_end, ria->ria_stoff, ria->ria_length,\
575 /* Limit this to the blocksize instead of PTLRPC_BRW_MAX_SIZE, since we don't
576 * know what the actual RPC size is. If this needs to change, it makes more
577 * sense to tune the i_blkbits value for the file based on the OSTs it is
578 * striped over, rather than having a constant value for all files here. */
579 #define RAS_INCREASE_STEP(inode) (1UL << inode->i_blkbits)
581 static inline int stride_io_mode(struct ll_readahead_state *ras)
583 return ras->ras_consecutive_stride_requests > 1;
585 /* The function calculates how much pages will be read in
586 * [off, off + length], in such stride IO area,
587 * stride_offset = st_off, stride_lengh = st_len,
588 * stride_pages = st_pgs
590 * |------------------|*****|------------------|*****|------------|*****|....
593 * |----- st_len -----|
595 * How many pages it should read in such pattern
596 * |-------------------------------------------------------------|
598 * |<------ length ------->|
600 * = |<----->| + |-------------------------------------| + |---|
601 * start_left st_pgs * i end_left
604 stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs,
605 unsigned long off, unsigned long length)
607 __u64 start = off > st_off ? off - st_off : 0;
608 __u64 end = off + length > st_off ? off + length - st_off : 0;
609 unsigned long start_left = 0;
610 unsigned long end_left = 0;
611 unsigned long pg_count;
613 if (st_len == 0 || length == 0 || end == 0)
616 start_left = do_div(start, st_len);
617 if (start_left < st_pgs)
618 start_left = st_pgs - start_left;
622 end_left = do_div(end, st_len);
623 if (end_left > st_pgs)
626 CDEBUG(D_READA, "start "LPU64", end "LPU64" start_left %lu end_left %lu \n",
627 start, end, start_left, end_left);
630 pg_count = end_left - (st_pgs - start_left);
632 pg_count = start_left + st_pgs * (end - start - 1) + end_left;
634 CDEBUG(D_READA, "st_off %lu, st_len %lu st_pgs %lu off %lu length %lu"
635 "pgcount %lu\n", st_off, st_len, st_pgs, off, length, pg_count);
640 static int ria_page_count(struct ra_io_arg *ria)
642 __u64 length = ria->ria_end >= ria->ria_start ?
643 ria->ria_end - ria->ria_start + 1 : 0;
645 return stride_pg_count(ria->ria_stoff, ria->ria_length,
646 ria->ria_pages, ria->ria_start,
650 /*Check whether the index is in the defined ra-window */
651 static int ras_inside_ra_window(unsigned long idx, struct ra_io_arg *ria)
653 /* If ria_length == ria_pages, it means non-stride I/O mode,
654 * idx should always inside read-ahead window in this case
655 * For stride I/O mode, just check whether the idx is inside
657 return ria->ria_length == 0 || ria->ria_length == ria->ria_pages ||
658 (idx >= ria->ria_stoff && (idx - ria->ria_stoff) %
659 ria->ria_length < ria->ria_pages);
662 static int ll_read_ahead_pages(const struct lu_env *env,
663 struct cl_io *io, struct cl_page_list *queue,
664 struct ra_io_arg *ria,
665 unsigned long *reserved_pages,
666 struct address_space *mapping,
667 unsigned long *ra_end)
669 int rc, count = 0, stride_ria;
670 unsigned long page_idx;
672 LASSERT(ria != NULL);
675 stride_ria = ria->ria_length > ria->ria_pages && ria->ria_pages > 0;
676 for (page_idx = ria->ria_start; page_idx <= ria->ria_end &&
677 *reserved_pages > 0; page_idx++) {
678 if (ras_inside_ra_window(page_idx, ria)) {
679 /* If the page is inside the read-ahead window*/
680 rc = ll_read_ahead_page(env, io, queue,
685 } else if (rc == -ENOLCK)
687 } else if (stride_ria) {
688 /* If it is not in the read-ahead window, and it is
689 * read-ahead mode, then check whether it should skip
692 /* FIXME: This assertion only is valid when it is for
693 * forward read-ahead, it will be fixed when backward
694 * read-ahead is implemented */
695 LASSERTF(page_idx > ria->ria_stoff, "Invalid page_idx %lu"
696 "rs %lu re %lu ro %lu rl %lu rp %lu\n", page_idx,
697 ria->ria_start, ria->ria_end, ria->ria_stoff,
698 ria->ria_length, ria->ria_pages);
699 offset = page_idx - ria->ria_stoff;
700 offset = offset % (ria->ria_length);
701 if (offset > ria->ria_pages) {
702 page_idx += ria->ria_length - offset;
703 CDEBUG(D_READA, "i %lu skip %lu \n", page_idx,
704 ria->ria_length - offset);
713 int ll_readahead(const struct lu_env *env, struct cl_io *io,
714 struct ll_readahead_state *ras, struct address_space *mapping,
715 struct cl_page_list *queue, int flags)
717 struct vvp_io *vio = vvp_env_io(env);
718 struct vvp_thread_info *vti = vvp_env_info(env);
719 struct cl_attr *attr = ccc_env_thread_attr(env);
720 unsigned long start = 0, end = 0, reserved;
721 unsigned long ra_end, len;
723 struct ll_ra_read *bead;
724 struct ra_io_arg *ria = &vti->vti_ria;
725 struct ll_inode_info *lli;
726 struct cl_object *clob;
731 inode = mapping->host;
732 lli = ll_i2info(inode);
733 clob = lli->lli_clob;
735 memset(ria, 0, sizeof *ria);
737 cl_object_attr_lock(clob);
738 ret = cl_object_attr_get(env, clob, attr);
739 cl_object_attr_unlock(clob);
745 ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN);
749 spin_lock(&ras->ras_lock);
750 if (vio->cui_ra_window_set)
751 bead = &vio->cui_bead;
755 /* Enlarge the RA window to encompass the full read */
756 if (bead != NULL && ras->ras_window_start + ras->ras_window_len <
757 bead->lrr_start + bead->lrr_count) {
758 ras->ras_window_len = bead->lrr_start + bead->lrr_count -
759 ras->ras_window_start;
761 /* Reserve a part of the read-ahead window that we'll be issuing */
762 if (ras->ras_window_len) {
763 start = ras->ras_next_readahead;
764 end = ras->ras_window_start + ras->ras_window_len - 1;
767 unsigned long rpc_boundary;
769 * Align RA window to an optimal boundary.
771 * XXX This would be better to align to cl_max_pages_per_rpc
772 * instead of PTLRPC_MAX_BRW_PAGES, because the RPC size may
773 * be aligned to the RAID stripe size in the future and that
774 * is more important than the RPC size.
776 /* Note: we only trim the RPC, instead of extending the RPC
777 * to the boundary, so to avoid reading too much pages during
779 rpc_boundary = ((end + 1) & (~(PTLRPC_MAX_BRW_PAGES - 1)));
780 if (rpc_boundary > 0)
783 if (rpc_boundary > start)
786 /* Truncate RA window to end of file */
787 end = min(end, (unsigned long)((kms - 1) >> CFS_PAGE_SHIFT));
789 ras->ras_next_readahead = max(end, end + 1);
792 ria->ria_start = start;
794 /* If stride I/O mode is detected, get stride window*/
795 if (stride_io_mode(ras)) {
796 ria->ria_stoff = ras->ras_stride_offset;
797 ria->ria_length = ras->ras_stride_length;
798 ria->ria_pages = ras->ras_stride_pages;
800 spin_unlock(&ras->ras_lock);
803 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
806 len = ria_page_count(ria);
810 reserved = ll_ra_count_get(ll_i2sbi(inode), ria, len);
812 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
814 CDEBUG(D_READA, "reserved page %lu ra_cur %d ra_max %lu\n", reserved,
815 cfs_atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
816 ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
818 ret = ll_read_ahead_pages(env, io, queue,
819 ria, &reserved, mapping, &ra_end);
821 LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
823 ll_ra_count_put(ll_i2sbi(inode), reserved);
825 if (ra_end == end + 1 && ra_end == (kms >> CFS_PAGE_SHIFT))
826 ll_ra_stats_inc(mapping, RA_STAT_EOF);
828 /* if we didn't get to the end of the region we reserved from
829 * the ras we need to go back and update the ras so that the
830 * next read-ahead tries from where we left off. we only do so
831 * if the region we failed to issue read-ahead on is still ahead
832 * of the app and behind the next index to start read-ahead from */
833 CDEBUG(D_READA, "ra_end %lu end %lu stride end %lu \n",
834 ra_end, end, ria->ria_end);
836 if (ra_end != end + 1) {
837 spin_lock(&ras->ras_lock);
838 if (ra_end < ras->ras_next_readahead &&
839 index_in_window(ra_end, ras->ras_window_start, 0,
840 ras->ras_window_len)) {
841 ras->ras_next_readahead = ra_end;
844 spin_unlock(&ras->ras_lock);
850 static void ras_set_start(struct inode *inode, struct ll_readahead_state *ras,
853 ras->ras_window_start = index & (~(RAS_INCREASE_STEP(inode) - 1));
856 /* called with the ras_lock held or from places where it doesn't matter */
857 static void ras_reset(struct inode *inode, struct ll_readahead_state *ras,
860 ras->ras_last_readpage = index;
861 ras->ras_consecutive_requests = 0;
862 ras->ras_consecutive_pages = 0;
863 ras->ras_window_len = 0;
864 ras_set_start(inode, ras, index);
865 ras->ras_next_readahead = max(ras->ras_window_start, index);
870 /* called with the ras_lock held or from places where it doesn't matter */
871 static void ras_stride_reset(struct ll_readahead_state *ras)
873 ras->ras_consecutive_stride_requests = 0;
874 ras->ras_stride_length = 0;
875 ras->ras_stride_pages = 0;
879 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
881 spin_lock_init(&ras->ras_lock);
882 ras_reset(inode, ras, 0);
883 ras->ras_requests = 0;
884 CFS_INIT_LIST_HEAD(&ras->ras_read_beads);
888 * Check whether the read request is in the stride window.
889 * If it is in the stride window, return 1, otherwise return 0.
891 static int index_in_stride_window(struct ll_readahead_state *ras,
894 unsigned long stride_gap;
896 if (ras->ras_stride_length == 0 || ras->ras_stride_pages == 0 ||
897 ras->ras_stride_pages == ras->ras_stride_length)
900 stride_gap = index - ras->ras_last_readpage - 1;
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 inode *inode,
985 struct ll_readahead_state *ras,
986 struct ll_ra_info *ra)
988 /* The stretch of ra-window should be aligned with max rpc_size
989 * but current clio architecture does not support retrieve such
990 * information from lower layer. FIXME later
992 if (stride_io_mode(ras))
993 ras_stride_increase_window(ras, ra, RAS_INCREASE_STEP(inode));
995 ras->ras_window_len = min(ras->ras_window_len +
996 RAS_INCREASE_STEP(inode),
997 ra->ra_max_pages_per_file);
1000 void ras_update(struct ll_sb_info *sbi, struct inode *inode,
1001 struct ll_readahead_state *ras, unsigned long index,
1004 struct ll_ra_info *ra = &sbi->ll_ra_info;
1005 int zero = 0, stride_detect = 0, ra_miss = 0;
1008 spin_lock(&ras->ras_lock);
1010 ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
1012 /* reset the read-ahead window in two cases. First when the app seeks
1013 * or reads to some other part of the file. Secondly if we get a
1014 * read-ahead miss that we think we've previously issued. This can
1015 * be a symptom of there being so many read-ahead pages that the VM is
1016 * reclaiming it before we get to it. */
1017 if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
1019 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
1020 } else if (!hit && ras->ras_window_len &&
1021 index < ras->ras_next_readahead &&
1022 index_in_window(index, ras->ras_window_start, 0,
1023 ras->ras_window_len)) {
1025 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
1028 /* On the second access to a file smaller than the tunable
1029 * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1030 * file up to ra_max_pages_per_file. This is simply a best effort
1031 * and only occurs once per open file. Normal RA behavior is reverted
1032 * to for subsequent IO. The mmap case does not increment
1033 * ras_requests and thus can never trigger this behavior. */
1034 if (ras->ras_requests == 2 && !ras->ras_request_index) {
1037 kms_pages = (i_size_read(inode) + CFS_PAGE_SIZE - 1) >>
1040 CDEBUG(D_READA, "kmsp "LPU64" mwp %lu mp %lu\n", kms_pages,
1041 ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages_per_file);
1044 kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1045 ras->ras_window_start = 0;
1046 ras->ras_last_readpage = 0;
1047 ras->ras_next_readahead = 0;
1048 ras->ras_window_len = min(ra->ra_max_pages_per_file,
1049 ra->ra_max_read_ahead_whole_pages);
1050 GOTO(out_unlock, 0);
1054 /* check whether it is in stride I/O mode*/
1055 if (!index_in_stride_window(ras, index)) {
1056 if (ras->ras_consecutive_stride_requests == 0 &&
1057 ras->ras_request_index == 0) {
1058 ras_update_stride_detector(ras, index);
1059 ras->ras_consecutive_stride_requests++;
1061 ras_stride_reset(ras);
1063 ras_reset(inode, ras, index);
1064 ras->ras_consecutive_pages++;
1065 GOTO(out_unlock, 0);
1067 ras->ras_consecutive_pages = 0;
1068 ras->ras_consecutive_requests = 0;
1069 if (++ras->ras_consecutive_stride_requests > 1)
1075 if (index_in_stride_window(ras, index) &&
1076 stride_io_mode(ras)) {
1077 /*If stride-RA hit cache miss, the stride dector
1078 *will not be reset to avoid the overhead of
1079 *redetecting read-ahead mode */
1080 if (index != ras->ras_last_readpage + 1)
1081 ras->ras_consecutive_pages = 0;
1082 ras_reset(inode, ras, index);
1085 /* Reset both stride window and normal RA
1087 ras_reset(inode, ras, index);
1088 ras->ras_consecutive_pages++;
1089 ras_stride_reset(ras);
1090 GOTO(out_unlock, 0);
1092 } else if (stride_io_mode(ras)) {
1093 /* If this is contiguous read but in stride I/O mode
1094 * currently, check whether stride step still is valid,
1095 * if invalid, it will reset the stride ra window*/
1096 if (!index_in_stride_window(ras, index)) {
1097 /* Shrink stride read-ahead window to be zero */
1098 ras_stride_reset(ras);
1099 ras->ras_window_len = 0;
1100 ras->ras_next_readahead = index;
1104 ras->ras_consecutive_pages++;
1105 ras->ras_last_readpage = index;
1106 ras_set_start(inode, ras, index);
1108 if (stride_io_mode(ras))
1109 /* Since stride readahead is sentivite to the offset
1110 * of read-ahead, so we use original offset here,
1111 * instead of ras_window_start, which is RPC aligned */
1112 ras->ras_next_readahead = max(index, 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(inode);
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(inode);
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(inode, ras, ra);
1145 ras->ras_request_index++;
1146 spin_unlock(&ras->ras_lock);
1150 int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
1152 struct inode *inode = vmpage->mapping->host;
1153 struct ll_inode_info *lli = ll_i2info(inode);
1156 struct cl_page *page;
1157 struct cl_object *clob;
1158 struct cl_env_nest nest;
1159 bool redirtied = false;
1160 bool unlocked = false;
1164 LASSERT(PageLocked(vmpage));
1165 LASSERT(!PageWriteback(vmpage));
1167 LASSERT(ll_i2dtexp(inode) != NULL);
1169 env = cl_env_nested_get(&nest);
1171 GOTO(out, result = PTR_ERR(env));
1173 clob = ll_i2info(inode)->lli_clob;
1174 LASSERT(clob != NULL);
1176 io = ccc_env_thread_io(env);
1178 io->ci_ignore_layout = 1;
1179 result = cl_io_init(env, io, CIT_MISC, clob);
1181 page = cl_page_find(env, clob, vmpage->index,
1182 vmpage, CPT_CACHEABLE);
1183 if (!IS_ERR(page)) {
1184 lu_ref_add(&page->cp_reference, "writepage",
1186 cl_page_assume(env, io, page);
1187 result = cl_page_flush(env, io, page);
1190 * Re-dirty page on error so it retries write,
1191 * but not in case when IO has actually
1192 * occurred and completed with an error.
1194 if (!PageError(vmpage)) {
1195 redirty_page_for_writepage(wbc, vmpage);
1200 cl_page_disown(env, io, page);
1202 lu_ref_del(&page->cp_reference,
1203 "writepage", cfs_current());
1204 cl_page_put(env, page);
1206 result = PTR_ERR(page);
1209 cl_io_fini(env, io);
1211 if (redirtied && wbc->sync_mode == WB_SYNC_ALL) {
1212 loff_t offset = cl_offset(clob, vmpage->index);
1214 /* Flush page failed because the extent is being written out.
1215 * Wait for the write of extent to be finished to avoid
1216 * breaking kernel which assumes ->writepage should mark
1217 * PageWriteback or clean the page. */
1218 result = cl_sync_file_range(inode, offset,
1219 offset + CFS_PAGE_SIZE - 1,
1222 /* actually we may have written more than one page.
1223 * decreasing this page because the caller will count
1225 wbc->nr_to_write -= result - 1;
1230 cl_env_nested_put(&nest, env);
1235 if (!lli->lli_async_rc)
1236 lli->lli_async_rc = result;
1237 SetPageError(vmpage);
1239 unlock_page(vmpage);
1244 int ll_writepages(struct address_space *mapping, struct writeback_control *wbc)
1246 struct inode *inode = mapping->host;
1249 enum cl_fsync_mode mode;
1250 int range_whole = 0;
1254 if (wbc->range_cyclic) {
1255 start = mapping->writeback_index << CFS_PAGE_SHIFT;
1256 end = OBD_OBJECT_EOF;
1258 start = wbc->range_start;
1259 end = wbc->range_end;
1260 if (end == LLONG_MAX) {
1261 end = OBD_OBJECT_EOF;
1262 range_whole = start == 0;
1266 mode = CL_FSYNC_NONE;
1267 if (wbc->sync_mode == WB_SYNC_ALL)
1268 mode = CL_FSYNC_LOCAL;
1270 result = cl_sync_file_range(inode, start, end, mode);
1272 wbc->nr_to_write -= result;
1276 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) {
1277 if (end == OBD_OBJECT_EOF)
1278 end = i_size_read(inode);
1279 mapping->writeback_index = (end >> CFS_PAGE_SHIFT) + 1;
1284 int ll_readpage(struct file *file, struct page *vmpage)
1286 struct ll_cl_context *lcc;
1290 lcc = ll_cl_init(file, vmpage, 0);
1292 struct lu_env *env = lcc->lcc_env;
1293 struct cl_io *io = lcc->lcc_io;
1294 struct cl_page *page = lcc->lcc_page;
1296 LASSERT(page->cp_type == CPT_CACHEABLE);
1297 if (likely(!PageUptodate(vmpage))) {
1298 cl_page_assume(env, io, page);
1299 result = cl_io_read_page(env, io, page);
1301 /* Page from a non-object file. */
1302 LASSERT(!ll_i2info(vmpage->mapping->host)->lli_has_smd);
1303 unlock_page(vmpage);
1308 unlock_page(vmpage);
1309 result = PTR_ERR(lcc);