1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Lustre Lite I/O page cache routines shared by different kernel revs
6 * Copyright (c) 2001-2003 Cluster File Systems, Inc.
8 * This file is part of Lustre, http://www.lustre.org.
10 * Lustre is free software; you can redistribute it and/or
11 * modify it under the terms of version 2 of the GNU General Public
12 * License as published by the Free Software Foundation.
14 * Lustre is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 * GNU General Public License for more details.
19 * You should have received a copy of the GNU General Public License
20 * along with Lustre; if not, write to the Free Software
21 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 #include <linux/config.h>
25 #include <linux/kernel.h>
27 #include <linux/string.h>
28 #include <linux/stat.h>
29 #include <linux/errno.h>
30 #include <linux/smp_lock.h>
31 #include <linux/unistd.h>
32 #include <linux/version.h>
33 #include <asm/system.h>
34 #include <asm/uaccess.h>
37 #include <linux/stat.h>
38 #include <asm/uaccess.h>
39 #include <asm/segment.h>
41 #include <linux/pagemap.h>
42 #include <linux/smp_lock.h>
44 #define DEBUG_SUBSYSTEM S_LLITE
46 #include <linux/lustre_mds.h>
47 #include <linux/lustre_lite.h>
48 #include "llite_internal.h"
49 #include <linux/lustre_compat25.h>
51 #ifndef list_for_each_prev_safe
52 #define list_for_each_prev_safe(pos, n, head) \
53 for (pos = (head)->prev, n = pos->prev; pos != (head); \
54 pos = n, n = pos->prev )
57 /* SYNCHRONOUS I/O to object storage for an inode */
58 static int ll_brw(int cmd, struct inode *inode, struct obdo *oa,
59 struct page *page, int flags)
61 struct ll_inode_info *lli = ll_i2info(inode);
62 struct lov_stripe_md *lsm = lli->lli_smd;
68 do_gettimeofday(&start);
71 pg.disk_offset = pg.page_offset = ((obd_off)page->index) << PAGE_SHIFT;
73 if (cmd == OBD_BRW_WRITE &&
74 (pg.disk_offset + PAGE_SIZE > inode->i_size))
75 pg.count = inode->i_size % PAGE_SIZE;
79 CDEBUG(D_PAGE, "%s %d bytes ino %lu at "LPU64"/"LPX64"\n",
80 cmd & OBD_BRW_WRITE ? "write" : "read", pg.count, inode->i_ino,
81 pg.disk_offset, pg.disk_offset);
83 CERROR("ZERO COUNT: ino %lu: size %p:%Lu(%p:%Lu) idx %lu off "
84 LPU64"\n", inode->i_ino, inode, inode->i_size,
85 page->mapping->host, page->mapping->host->i_size,
86 page->index, pg.disk_offset);
91 if (cmd == OBD_BRW_WRITE)
92 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
93 LPROC_LL_BRW_WRITE, pg.count);
95 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
96 LPROC_LL_BRW_READ, pg.count);
97 rc = obd_brw(cmd, ll_i2dtexp(inode), oa, lsm, 1, &pg, NULL);
99 obdo_to_inode(inode, oa, OBD_MD_FLBLOCKS);
101 CERROR("error from obd_brw: rc = %d\n", rc);
102 ll_stime_record(ll_i2sbi(inode), &start,
103 &ll_i2sbi(inode)->ll_brw_stime);
107 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
110 * this isn't where truncate starts. roughly:
111 * sys_truncate->ll_setattr_raw->vmtruncate->ll_truncate
112 * we grab the lock back in setattr_raw to avoid races.
114 * must be called with lli_size_sem held.
116 void ll_truncate(struct inode *inode)
118 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
119 struct ll_inode_info *lli = ll_i2info(inode);
120 struct obdo *oa = NULL;
124 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) to %llu\n", inode->i_ino,
125 inode->i_generation, inode, inode->i_size);
127 if (lli->lli_size_pid != current->pid) {
133 CDEBUG(D_INODE, "truncate on inode %lu with no objects\n",
138 LASSERT(atomic_read(&lli->lli_size_sem.count) <= 0);
140 if (lov_merge_size(lsm, 0) == inode->i_size) {
141 CDEBUG(D_VFSTRACE, "skipping punch for "LPX64" (size = %llu)\n",
142 lsm->lsm_object_id, inode->i_size);
146 CDEBUG(D_INFO, "calling punch for "LPX64" (new size %llu)\n",
147 lsm->lsm_object_id, inode->i_size);
151 CERROR("cannot alloc oa, error %d\n",
157 oa->o_id = lsm->lsm_object_id;
158 oa->o_gr = lsm->lsm_object_gr;
159 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
160 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLMODE |
161 OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME);
163 obd_adjust_kms(ll_i2dtexp(inode), lsm, inode->i_size, 1);
165 lli->lli_size_pid = 0;
166 up(&lli->lli_size_sem);
168 rc = obd_punch(ll_i2dtexp(inode), oa, lsm, inode->i_size,
169 OBD_OBJECT_EOF, NULL);
171 CERROR("obd_truncate fails (%d) ino %lu\n", rc, inode->i_ino);
173 obdo_to_inode(inode, oa, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
174 OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME);
182 LASSERT(atomic_read(&lli->lli_size_sem.count) <= 0);
183 up(&lli->lli_size_sem);
186 struct ll_async_page *llap_cast_private(struct page *page)
188 struct ll_async_page *llap = (struct ll_async_page *)page->private;
190 LASSERTF(llap == NULL || llap->llap_magic == LLAP_MAGIC,
191 "page %p private %lu gave magic %d which != %d\n",
192 page, page->private, llap->llap_magic, LLAP_MAGIC);
196 int ll_prepare_write(struct file *file, struct page *page,
197 unsigned from, unsigned to)
199 struct inode *inode = page->mapping->host;
200 struct ll_inode_info *lli = ll_i2info(inode);
201 struct lov_stripe_md *lsm = lli->lli_smd;
202 obd_off offset = ((obd_off)page->index) << PAGE_SHIFT;
203 struct obdo *oa = NULL;
209 LASSERT(LLI_DIRTY_HANDLE(inode));
210 LASSERT(PageLocked(page));
211 (void)llap_cast_private(page); /* assertion */
213 /* Check to see if we should return -EIO right away */
215 pga.disk_offset = pga.page_offset = offset;
216 pga.count = PAGE_SIZE;
223 oa->o_id = lsm->lsm_object_id;
224 oa->o_gr = lsm->lsm_object_gr;
225 oa->o_mode = inode->i_mode;
227 oa->o_valid = OBD_MD_FLID | OBD_MD_FLMODE |
228 OBD_MD_FLTYPE | OBD_MD_FLGROUP;
230 rc = obd_brw(OBD_BRW_CHECK, ll_i2dtexp(inode),
231 oa, lsm, 1, &pga, NULL);
233 GOTO(out_free_oa, rc);
235 if (PageUptodate(page))
236 GOTO(out_free_oa, 0);
238 /* We're completely overwriting an existing page, so _don't_ set it up
239 * to date until commit_write */
240 if (from == 0 && to == PAGE_SIZE) {
241 POISON_PAGE(page, 0x11);
242 GOTO(out_free_oa, 0);
245 /* If are writing to a new page, no need to read old data. The extent
246 * locking will have updated the KMS, and for our purposes here we can
247 * treat it like i_size. */
248 down(&lli->lli_size_sem);
249 kms = lov_merge_size(lsm, 1);
250 up(&lli->lli_size_sem);
252 memset(kmap(page), 0, PAGE_SIZE);
254 GOTO(prepare_done, rc = 0);
257 /* XXX could be an async ocp read.. read-ahead? */
258 rc = ll_brw(OBD_BRW_READ, inode, oa, page, 0);
260 /* bug 1598: don't clobber blksize */
261 oa->o_valid &= ~(OBD_MD_FLSIZE | OBD_MD_FLBLKSZ);
262 obdo_refresh_inode(inode, oa, oa->o_valid);
263 } else if (rc == -ENOENT) {
264 /* tolerate no entry error here, cause the objects might
265 * not be created yet */
272 SetPageUptodate(page);
278 static int ll_ap_make_ready(void *data, int cmd)
280 struct ll_async_page *llap;
284 llap = LLAP_FROM_COOKIE(data);
285 page = llap->llap_page;
287 LASSERT(cmd != OBD_BRW_READ);
289 /* we're trying to write, but the page is locked.. come back later */
290 if (TryLockPage(page))
293 LL_CDEBUG_PAGE(D_PAGE, page, "made ready\n");
294 page_cache_get(page);
296 /* if we left PageDirty we might get another writepage call
297 * in the future. list walkers are bright enough
298 * to check page dirty so we can leave it on whatever list
299 * its on. XXX also, we're called with the cli list so if
300 * we got the page cache list we'd create a lock inversion
301 * with the removepage path which gets the page lock then the
303 clear_page_dirty(page);
307 /* We have two reasons for giving llite the opportunity to change the
308 * write length of a given queued page as it builds the RPC containing
311 * 1) Further extending writes may have landed in the page cache
312 * since a partial write first queued this page requiring us
313 * to write more from the page cache. (No further races are possible, since
314 * by the time this is called, the page is locked.)
315 * 2) We might have raced with truncate and want to avoid performing
316 * write RPCs that are just going to be thrown away by the
317 * truncate's punch on the storage targets.
319 * The kms serves these purposes as it is set at both truncate and extending
322 static int ll_ap_refresh_count(void *data, int cmd)
324 struct ll_inode_info *lli;
325 struct ll_async_page *llap;
326 struct lov_stripe_md *lsm;
331 /* readpage queues with _COUNT_STABLE, shouldn't get here. */
332 LASSERT(cmd != OBD_BRW_READ);
334 llap = LLAP_FROM_COOKIE(data);
335 page = llap->llap_page;
336 lli = ll_i2info(page->mapping->host);
340 * this callback is called with client lock taken, thus, it should not
341 * sleep or deadlock is possible. --umka
343 // down(&lli->lli_size_sem);
344 kms = lov_merge_size(lsm, 1);
345 // up(&lli->lli_size_sem);
347 /* catch race with truncate */
348 if (((__u64)page->index << PAGE_SHIFT) >= kms)
351 /* catch sub-page write at end of file */
352 if (((__u64)page->index << PAGE_SHIFT) + PAGE_SIZE > kms)
353 return kms % PAGE_SIZE;
358 void ll_inode_fill_obdo(struct inode *inode, int cmd, struct obdo *oa)
360 struct lov_stripe_md *lsm;
361 obd_valid valid_flags;
363 lsm = ll_i2info(inode)->lli_smd;
365 oa->o_id = lsm->lsm_object_id;
366 oa->o_gr = lsm->lsm_object_gr;
367 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
368 valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME;
369 if (cmd == OBD_BRW_WRITE || cmd == OBD_BRW_READ) {
370 oa->o_valid |= OBD_MD_FLIFID | OBD_MD_FLEPOCH;
371 *(obdo_id(oa)) = ll_i2info(inode)->lli_id;
372 oa->o_easize = ll_i2info(inode)->lli_io_epoch;
373 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME;
376 obdo_from_inode(oa, inode, valid_flags);
379 static void ll_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
381 struct ll_async_page *llap;
384 llap = LLAP_FROM_COOKIE(data);
385 ll_inode_fill_obdo(llap->llap_page->mapping->host, cmd, oa);
389 static struct obd_async_page_ops ll_async_page_ops = {
390 .ap_make_ready = ll_ap_make_ready,
391 .ap_refresh_count = ll_ap_refresh_count,
392 .ap_fill_obdo = ll_ap_fill_obdo,
393 .ap_completion = ll_ap_completion,
397 /* XXX have the exp be an argument? */
398 struct ll_async_page *llap_from_page(struct page *page, unsigned origin)
400 struct ll_async_page *llap;
401 struct obd_export *exp;
402 struct inode *inode = page->mapping->host;
403 struct ll_sb_info *sbi = ll_i2sbi(inode);
407 LASSERTF(origin < LLAP__ORIGIN_MAX, "%u\n", origin);
409 llap = llap_cast_private(page);
413 exp = ll_i2dtexp(page->mapping->host);
415 RETURN(ERR_PTR(-EINVAL));
417 OBD_ALLOC(llap, sizeof(*llap));
419 RETURN(ERR_PTR(-ENOMEM));
421 llap->llap_magic = LLAP_MAGIC;
422 INIT_LIST_HEAD(&llap->llap_pending_write);
423 rc = obd_prep_async_page(exp, ll_i2info(inode)->lli_smd, NULL, page,
424 (obd_off)page->index << PAGE_SHIFT,
425 &ll_async_page_ops, llap, &llap->llap_cookie);
427 OBD_FREE(llap, sizeof(*llap));
431 CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n", llap,
432 page, llap->llap_cookie, (obd_off)page->index << PAGE_SHIFT);
434 __set_page_ll_data(page, llap);
436 /* also zeroing the PRIVBITS low order bitflags */
437 llap->llap_page = page;
439 spin_lock(&sbi->ll_lock);
440 sbi->ll_pglist_gen++;
441 list_add_tail(&llap->llap_proc_item, &sbi->ll_pglist);
442 spin_unlock(&sbi->ll_lock);
445 llap->llap_origin = origin;
449 static int queue_or_sync_write(struct obd_export *exp,
450 struct lov_stripe_md *lsm,
451 struct ll_async_page *llap,
453 obd_flags async_flags)
455 struct obd_io_group *oig;
459 /* _make_ready only sees llap once we've unlocked the page */
460 llap->llap_write_queued = 1;
461 rc = obd_queue_async_io(exp, lsm, NULL, llap->llap_cookie,
462 OBD_BRW_WRITE, 0, 0, 0, async_flags);
464 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "write queued\n");
465 llap_write_pending(llap->llap_page->mapping->host, llap);
469 llap->llap_write_queued = 0;
474 rc = obd_queue_group_io(exp, lsm, NULL, oig, llap->llap_cookie,
475 OBD_BRW_WRITE, 0, to, 0, ASYNC_READY |
476 ASYNC_URGENT | ASYNC_COUNT_STABLE |
481 rc = obd_trigger_group_io(exp, lsm, NULL, oig);
487 if (!rc && async_flags & ASYNC_READY)
488 unlock_page(llap->llap_page);
490 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
491 "sync write returned %d\n", rc);
500 /* be careful not to return success without setting the page Uptodate or
501 * the next pass through prepare_write will read in stale data from disk. */
502 int ll_commit_write(struct file *file, struct page *page, unsigned from,
505 struct inode *inode = page->mapping->host;
506 struct ll_inode_info *lli = ll_i2info(inode);
507 struct lov_stripe_md *lsm = lli->lli_smd;
508 struct obd_export *exp = NULL;
509 struct ll_async_page *llap;
514 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
515 LASSERT(inode == file->f_dentry->d_inode);
516 LASSERT(PageLocked(page));
517 LASSERT(LLI_DIRTY_HANDLE(inode));
519 CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
520 inode, page, from, to, page->index);
522 llap = llap_from_page(page, LLAP_ORIGIN_COMMIT_WRITE);
524 RETURN(PTR_ERR(llap));
526 exp = ll_i2dtexp(inode);
530 /* queue a write for some time in the future the first time we
532 if (!PageDirty(page)) {
533 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
534 LPROC_LL_DIRTY_MISSES);
536 rc = queue_or_sync_write(exp, ll_i2info(inode)->lli_smd,
541 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
542 LPROC_LL_DIRTY_HITS);
545 /* put the page in the page cache, from now on ll_removepage is
546 * responsible for cleaning up the llap.
547 * don't dirty the page if it has been write out in q_o_s_w */
548 if (llap->llap_write_queued)
549 set_page_dirty(page);
552 size = (((obd_off)page->index) << PAGE_SHIFT) + to;
553 down(&lli->lli_size_sem);
555 obd_adjust_kms(exp, lsm, size, 0);
556 if (size > inode->i_size)
557 inode->i_size = size;
558 SetPageUptodate(page);
559 } else if (size > inode->i_size) {
560 /* this page beyond the pales of i_size, so it can't be
561 * truncated in ll_p_r_e during lock revoking. we must
562 * teardown our book-keeping here. */
565 up(&lli->lli_size_sem);
569 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, unsigned long len)
571 struct ll_ra_info *ra = &sbi->ll_ra_info;
575 spin_lock(&sbi->ll_lock);
576 ret = min(ra->ra_max_pages - ra->ra_cur_pages, len);
577 ra->ra_cur_pages += ret;
578 spin_unlock(&sbi->ll_lock);
583 static void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
585 struct ll_ra_info *ra = &sbi->ll_ra_info;
586 spin_lock(&sbi->ll_lock);
587 LASSERTF(ra->ra_cur_pages >= len, "r_c_p %lu len %lu\n",
588 ra->ra_cur_pages, len);
589 ra->ra_cur_pages -= len;
590 spin_unlock(&sbi->ll_lock);
593 int ll_writepage(struct page *page)
595 struct inode *inode = page->mapping->host;
596 struct obd_export *exp;
597 struct ll_async_page *llap;
601 LASSERT(!PageDirty(page));
602 LASSERT(PageLocked(page));
603 LASSERT(LLI_DIRTY_HANDLE(inode));
605 exp = ll_i2dtexp(inode);
607 GOTO(out, rc = -EINVAL);
609 llap = llap_from_page(page, LLAP_ORIGIN_WRITEPAGE);
611 GOTO(out, rc = PTR_ERR(llap));
613 page_cache_get(page);
614 if (llap->llap_write_queued) {
615 LL_CDEBUG_PAGE(D_PAGE, page, "marking urgent\n");
616 rc = obd_set_async_flags(exp, ll_i2info(inode)->lli_smd, NULL,
618 ASYNC_READY | ASYNC_URGENT);
620 rc = queue_or_sync_write(exp, ll_i2info(inode)->lli_smd, llap,
621 PAGE_SIZE, ASYNC_READY |
625 page_cache_release(page);
633 /* called for each page in a completed rpc.*/
634 void ll_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
636 struct ll_async_page *llap;
640 llap = LLAP_FROM_COOKIE(data);
641 page = llap->llap_page;
642 LASSERT(PageLocked(page));
644 LL_CDEBUG_PAGE(D_PAGE, page, "completing cmd %d with %d\n", cmd, rc);
646 if (cmd == OBD_BRW_READ && llap->llap_defer_uptodate)
647 ll_ra_count_put(ll_i2sbi(page->mapping->host), 1);
650 if (cmd == OBD_BRW_READ) {
651 if (!llap->llap_defer_uptodate)
652 SetPageUptodate(page);
654 llap->llap_write_queued = 0;
656 ClearPageError(page);
658 if (cmd == OBD_BRW_READ)
659 llap->llap_defer_uptodate = 0;
665 if (cmd == OBD_BRW_WRITE) {
666 llap_write_complete(page->mapping->host, llap);
667 ll_try_done_writing(page->mapping->host);
670 if (PageWriteback(page)) {
671 end_page_writeback(page);
673 page_cache_release(page);
677 /* the kernel calls us here when a page is unhashed from the page cache.
678 * the page will be locked and the kernel is holding a spinlock, so
679 * we need to be careful. we're just tearing down our book-keeping
681 void ll_removepage(struct page *page)
683 struct inode *inode = page->mapping->host;
684 struct obd_export *exp;
685 struct ll_async_page *llap;
686 struct ll_sb_info *sbi = ll_i2sbi(inode);
690 LASSERT(!in_interrupt());
692 /* sync pages or failed read pages can leave pages in the page
693 * cache that don't have our data associated with them anymore */
694 if (page->private == 0) {
699 LL_CDEBUG_PAGE(D_PAGE, page, "being evicted\n");
701 exp = ll_i2dtexp(inode);
703 CERROR("page %p ind %lu gave null export\n", page, page->index);
707 llap = llap_from_page(page, 0);
709 CERROR("page %p ind %lu couldn't find llap: %ld\n", page,
710 page->index, PTR_ERR(llap));
715 llap_write_complete(inode, llap);
716 rc = obd_teardown_async_page(exp, ll_i2info(inode)->lli_smd, NULL,
719 CERROR("page %p ind %lu failed: %d\n", page, page->index, rc);
721 /* this unconditional free is only safe because the page lock
722 * is providing exclusivity to memory pressure/truncate/writeback..*/
723 __clear_page_ll_data(page);
725 spin_lock(&sbi->ll_lock);
726 if (!list_empty(&llap->llap_proc_item))
727 list_del_init(&llap->llap_proc_item);
728 sbi->ll_pglist_gen++;
729 spin_unlock(&sbi->ll_lock);
730 OBD_FREE(llap, sizeof(*llap));
734 static int ll_page_matches(struct page *page, int fd_flags, int readahead)
736 struct lustre_handle match_lockh = {0};
737 struct inode *inode = page->mapping->host;
738 ldlm_policy_data_t page_extent;
742 if (fd_flags & LL_FILE_GROUP_LOCKED)
745 page_extent.l_extent.start = (__u64)page->index << PAGE_CACHE_SHIFT;
746 page_extent.l_extent.end =
747 page_extent.l_extent.start + PAGE_CACHE_SIZE - 1;
748 flags = LDLM_FL_TEST_LOCK;
750 flags |= LDLM_FL_CBPENDING | LDLM_FL_BLOCK_GRANTED;
751 matches = obd_match(ll_i2sbi(inode)->ll_dt_exp,
752 ll_i2info(inode)->lli_smd, LDLM_EXTENT,
753 &page_extent, LCK_PR | LCK_PW, &flags, inode,
758 static int ll_issue_page_read(struct obd_export *exp,
759 struct ll_async_page *llap,
760 struct obd_io_group *oig, int defer)
762 struct page *page = llap->llap_page;
765 page_cache_get(page);
766 llap->llap_defer_uptodate = defer;
767 llap->llap_ra_used = 0;
769 rc = obd_queue_group_io(exp, ll_i2info(page->mapping->host)->lli_smd,
770 NULL, oig, llap->llap_cookie, OBD_BRW_READ, 0,
771 PAGE_SIZE, 0, ASYNC_COUNT_STABLE | ASYNC_READY
774 LL_CDEBUG_PAGE(D_ERROR, page, "read queue failed: rc %d\n", rc);
775 page_cache_release(page);
780 static void ll_ra_stats_inc_unlocked(struct ll_ra_info *ra, enum ra_stat which)
782 LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
783 ra->ra_stats[which]++;
786 static void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which)
788 struct ll_sb_info *sbi = ll_i2sbi(mapping->host);
789 struct ll_ra_info *ra = &ll_i2sbi(mapping->host)->ll_ra_info;
791 spin_lock(&sbi->ll_lock);
792 ll_ra_stats_inc_unlocked(ra, which);
793 spin_unlock(&sbi->ll_lock);
796 void ll_ra_accounting(struct page *page, struct address_space *mapping)
798 struct ll_async_page *llap;
800 llap = llap_from_page(page, LLAP_ORIGIN_WRITEPAGE);
804 if (!llap->llap_defer_uptodate || llap->llap_ra_used)
807 ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
810 #define RAS_CDEBUG(ras) \
811 CDEBUG(D_READA, "lrp %lu c %lu ws %lu wl %lu nra %lu\n", \
812 ras->ras_last_readpage, ras->ras_consecutive, \
813 ras->ras_window_start, ras->ras_window_len, \
814 ras->ras_next_readahead);
816 static int index_in_window(unsigned long index, unsigned long point,
817 unsigned long before, unsigned long after)
819 unsigned long start = point - before, end = point + after;
826 return start <= index && index <= end;
829 static int ll_readahead(struct ll_readahead_state *ras,
830 struct obd_export *exp, struct address_space *mapping,
831 struct obd_io_group *oig, int flags)
833 unsigned long i, start = 0, end = 0, reserved;
834 struct ll_async_page *llap;
836 int rc, ret = 0, match_failed = 0;
840 kms = lov_merge_size(ll_i2info(mapping->host)->lli_smd, 1);
842 ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN);
845 spin_lock(&ras->ras_lock);
847 /* reserve a part of the read-ahead window that we'll be issuing */
848 if (ras->ras_window_len) {
849 start = ras->ras_next_readahead;
850 end = ras->ras_window_start + ras->ras_window_len - 1;
851 end = min(end, (unsigned long)(kms >> PAGE_CACHE_SHIFT));
852 ras->ras_next_readahead = max(end, end + 1);
857 spin_unlock(&ras->ras_lock);
860 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
864 reserved = ll_ra_count_get(ll_i2sbi(mapping->host), end - start + 1);
865 if (reserved < end - start + 1)
866 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
868 for (i = start; reserved > 0 && !match_failed && i <= end; i++) {
869 /* skip locked pages from previous readpage calls */
870 page = grab_cache_page_nowait(mapping, i);
872 CDEBUG(D_READA, "g_c_p_n failed\n");
876 /* we do this first so that we can see the page in the /proc
878 llap = llap_from_page(page, LLAP_ORIGIN_READAHEAD);
879 if (IS_ERR(llap) || llap->llap_defer_uptodate)
882 /* skip completed pages */
883 if (Page_Uptodate(page))
886 /* bail when we hit the end of the lock. */
887 if ((rc = ll_page_matches(page, flags, 1)) <= 0) {
888 LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
889 "lock match failed: rc %d\n", rc);
890 ll_ra_stats_inc(mapping, RA_STAT_FAILED_MATCH);
895 rc = ll_issue_page_read(exp, llap, oig, 1);
899 LL_CDEBUG_PAGE(D_READA| D_PAGE, page,
900 "started read-ahead\n");
904 LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
905 "skipping read-ahead\n");
909 page_cache_release(page);
912 LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
914 ll_ra_count_put(ll_i2sbi(mapping->host), reserved);
916 if (i == end + 1 && end == (kms >> PAGE_CACHE_SHIFT))
917 ll_ra_stats_inc(mapping, RA_STAT_EOF);
919 /* if we didn't get to the end of the region we reserved from
920 * the ras we need to go back and update the ras so that the
921 * next read-ahead tries from where we left off. we only do so
922 * if the region we failed to issue read-ahead on is still ahead
923 * of the app and behind the next index to start read-ahead from */
925 spin_lock(&ras->ras_lock);
926 if (i < ras->ras_next_readahead &&
927 index_in_window(i, ras->ras_window_start, 0,
928 ras->ras_window_len)) {
929 ras->ras_next_readahead = i;
932 spin_unlock(&ras->ras_lock);
938 static void ras_set_start(struct ll_readahead_state *ras, unsigned long index)
940 ras->ras_window_start = index & (~(PTLRPC_MAX_BRW_PAGES - 1));
943 /* called with the ras_lock held or from places where it doesn't matter */
944 static void ras_reset(struct ll_readahead_state *ras, unsigned long index)
946 ras->ras_last_readpage = index;
947 ras->ras_consecutive = 1;
948 ras->ras_window_len = 0;
949 ras_set_start(ras, index);
950 ras->ras_next_readahead = ras->ras_window_start;
955 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
957 spin_lock_init(&ras->ras_lock);
961 static void ras_update(struct ll_sb_info *sbi, struct ll_readahead_state *ras,
962 unsigned long index, unsigned hit)
964 struct ll_ra_info *ra = &sbi->ll_ra_info;
968 spin_lock(&sbi->ll_lock);
969 spin_lock(&ras->ras_lock);
971 ll_ra_stats_inc_unlocked(ra, hit ? RA_STAT_HIT : RA_STAT_MISS);
973 /* reset the read-ahead window in two cases. First when the app seeks
974 * or reads to some other part of the file. Secondly if we get a
975 * read-ahead miss that we think we've previously issued. This can
976 * be a symptom of there being so many read-ahead pages that the VM is
977 * reclaiming it before we get to it. */
978 if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
980 ll_ra_stats_inc_unlocked(ra, RA_STAT_DISTANT_READPAGE);
981 } else if (!hit && ras->ras_window_len &&
982 index < ras->ras_next_readahead &&
983 index_in_window(index, ras->ras_window_start, 0,
984 ras->ras_window_len)) {
986 ll_ra_stats_inc_unlocked(ra, RA_STAT_MISS_IN_WINDOW);
990 ras_reset(ras, index);
994 ras->ras_last_readpage = index;
995 ras->ras_consecutive++;
996 ras_set_start(ras, index);
997 ras->ras_next_readahead = max(ras->ras_window_start,
998 ras->ras_next_readahead);
1000 /* wait for a few pages to arrive before issuing readahead to avoid
1001 * the worst overutilization */
1002 if (ras->ras_consecutive == 3) {
1003 ras->ras_window_len = PTLRPC_MAX_BRW_PAGES;
1004 GOTO(out_unlock, 0);
1007 /* we need to increase the window sometimes. we'll arbitrarily
1008 * do it half-way through the pages in an rpc */
1009 if ((index & (PTLRPC_MAX_BRW_PAGES - 1)) ==
1010 (PTLRPC_MAX_BRW_PAGES >> 1)) {
1011 ras->ras_window_len += PTLRPC_MAX_BRW_PAGES;
1012 ras->ras_window_len = min(ras->ras_window_len,
1020 spin_unlock(&ras->ras_lock);
1021 spin_unlock(&sbi->ll_lock);
1025 * for now we do our readpage the same on both 2.4 and 2.5. The kernel's
1026 * read-ahead assumes it is valid to issue readpage all the way up to
1027 * i_size, but our dlm locks make that not the case. We disable the
1028 * kernel's read-ahead and do our own by walking ahead in the page cache
1029 * checking for dlm lock coverage. the main difference between 2.4 and
1030 * 2.6 is how read-ahead gets batched and issued, but we're using our own,
1031 * so they look the same.
1033 int ll_readpage(struct file *filp, struct page *page)
1035 struct ll_file_data *fd = filp->private_data;
1036 struct inode *inode = page->mapping->host;
1037 struct obd_export *exp;
1038 struct ll_async_page *llap;
1039 struct obd_io_group *oig = NULL;
1043 LASSERT(PageLocked(page));
1044 LASSERT(!PageUptodate(page));
1045 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset="LPX64"\n",
1046 inode->i_ino, inode->i_generation, inode,
1047 (((obd_off)page->index) << PAGE_SHIFT));
1048 LASSERT(atomic_read(&filp->f_dentry->d_inode->i_count) > 0);
1050 rc = oig_init(&oig);
1054 exp = ll_i2dtexp(inode);
1056 GOTO(out, rc = -EINVAL);
1058 llap = llap_from_page(page, LLAP_ORIGIN_READPAGE);
1060 GOTO(out, rc = PTR_ERR(llap));
1062 if (ll_i2sbi(inode)->ll_flags & LL_SBI_READAHEAD)
1063 ras_update(ll_i2sbi(inode), &fd->fd_ras, page->index,
1064 llap->llap_defer_uptodate);
1066 if (llap->llap_defer_uptodate) {
1067 llap->llap_ra_used = 1;
1068 rc = ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1071 obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd,
1073 LL_CDEBUG_PAGE(D_PAGE, page, "marking uptodate from defer\n");
1074 SetPageUptodate(page);
1076 GOTO(out_oig, rc = 0);
1079 rc = ll_page_matches(page, fd->fd_flags, 0);
1081 LL_CDEBUG_PAGE(D_ERROR, page, "lock match failed: rc %d\n", rc);
1086 CWARN("ino %lu page %lu (%llu) not covered by "
1087 "a lock (mmap?). check debug logs.\n",
1088 inode->i_ino, page->index,
1089 (long long)page->index << PAGE_CACHE_SHIFT);
1092 rc = ll_issue_page_read(exp, llap, oig, 0);
1096 LL_CDEBUG_PAGE(D_PAGE, page, "queued readpage\n");
1097 if (ll_i2sbi(inode)->ll_flags & LL_SBI_READAHEAD)
1098 ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1101 rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);