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 pg.off = ((obd_off)page->index) << PAGE_SHIFT;
70 if (cmd == OBD_BRW_WRITE && (pg.off + PAGE_SIZE > inode->i_size))
71 pg.count = inode->i_size % PAGE_SIZE;
75 CDEBUG(D_PAGE, "%s %d bytes ino %lu at "LPU64"/"LPX64"\n",
76 cmd & OBD_BRW_WRITE ? "write" : "read", pg.count, inode->i_ino,
79 CERROR("ZERO COUNT: ino %lu: size %p:%Lu(%p:%Lu) idx %lu off "
81 inode->i_ino, inode, inode->i_size, page->mapping->host,
82 page->mapping->host->i_size, page->index, pg.off);
87 if (cmd == OBD_BRW_WRITE)
88 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
89 LPROC_LL_BRW_WRITE, pg.count);
91 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
92 LPROC_LL_BRW_READ, pg.count);
93 rc = obd_brw(cmd, ll_i2obdexp(inode), oa, lsm, 1, &pg, NULL);
95 obdo_to_inode(inode, oa, OBD_MD_FLBLOCKS);
97 CERROR("error from obd_brw: rc = %d\n", rc);
101 /* this isn't where truncate starts. roughly:
102 * sys_truncate->ll_setattr_raw->vmtruncate->ll_truncate
103 * we grab the lock back in setattr_raw to avoid races. */
104 void ll_truncate(struct inode *inode)
106 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
110 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
111 inode->i_generation, inode);
114 CDEBUG(D_INODE, "truncate on inode %lu with no objects\n",
120 oa.o_id = lsm->lsm_object_id;
121 oa.o_valid = OBD_MD_FLID;
122 obdo_from_inode(&oa, inode, OBD_MD_FLTYPE|OBD_MD_FLMODE|OBD_MD_FLATIME|
123 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
125 CDEBUG(D_INFO, "calling punch for "LPX64" (all bytes after %Lu)\n",
126 oa.o_id, inode->i_size);
128 /* truncate == punch from new size to absolute end of file */
129 /* NB: obd_punch must be called with i_sem held! It updates the kms! */
130 rc = obd_punch(ll_i2obdexp(inode), &oa, lsm, inode->i_size,
131 OBD_OBJECT_EOF, NULL);
133 CERROR("obd_truncate fails (%d) ino %lu\n", rc, inode->i_ino);
135 obdo_to_inode(inode, &oa, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
136 OBD_MD_FLATIME | OBD_MD_FLMTIME |
143 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
144 int ll_prepare_write(struct file *file, struct page *page, unsigned from,
147 struct inode *inode = page->mapping->host;
148 struct ll_inode_info *lli = ll_i2info(inode);
149 struct lov_stripe_md *lsm = lli->lli_smd;
150 obd_off offset = ((obd_off)page->index) << PAGE_SHIFT;
157 if (!PageLocked(page))
160 /* Check to see if we should return -EIO right away */
163 pga.count = PAGE_SIZE;
166 oa.o_id = lsm->lsm_object_id;
167 oa.o_mode = inode->i_mode;
168 oa.o_valid = OBD_MD_FLID | OBD_MD_FLMODE | OBD_MD_FLTYPE;
170 rc = obd_brw(OBD_BRW_CHECK, ll_i2obdexp(inode), &oa, lsm, 1, &pga,
175 if (PageUptodate(page))
178 /* We're completely overwriting an existing page, so _don't_ set it up
179 * to date until commit_write */
180 if (from == 0 && to == PAGE_SIZE) {
181 POISON_PAGE(page, 0x11);
185 /* If are writing to a new page, no need to read old data. The extent
186 * locking will have updated the KMS, and for our purposes here we can
187 * treat it like i_size. */
188 kms = lov_merge_size(lsm, 1);
190 memset(kmap(page), 0, PAGE_SIZE);
192 GOTO(prepare_done, rc = 0);
195 /* XXX could be an async ocp read.. read-ahead? */
196 rc = ll_brw(OBD_BRW_READ, inode, &oa, page, 0);
198 /* bug 1598: don't clobber blksize */
199 oa.o_valid &= ~(OBD_MD_FLSIZE | OBD_MD_FLBLKSZ);
200 obdo_refresh_inode(inode, &oa, oa.o_valid);
206 SetPageUptodate(page);
211 int ll_write_count(struct page *page)
213 struct inode *inode = page->mapping->host;
215 /* catch race with truncate */
216 if (((loff_t)page->index << PAGE_SHIFT) >= inode->i_size)
219 /* catch sub-page write at end of file */
220 if (((loff_t)page->index << PAGE_SHIFT) + PAGE_SIZE > inode->i_size)
221 return inode->i_size % PAGE_SIZE;
226 struct ll_async_page *llap_from_cookie(void *cookie)
228 struct ll_async_page *llap = cookie;
229 if (llap->llap_magic != LLAP_MAGIC)
230 return ERR_PTR(-EINVAL);
234 static int ll_ap_make_ready(void *data, int cmd)
236 struct ll_async_page *llap;
240 llap = llap_from_cookie(data);
244 page = llap->llap_page;
246 if (cmd == OBD_BRW_READ) {
247 /* _sync_page beat us to it and is about to call
248 * _set_async_flags which will fire off rpcs again */
249 if (!test_and_clear_bit(LL_PRIVBITS_READ, &page->private))
254 /* we're trying to write, but the page is locked.. come back later */
255 if (TryLockPage(page))
258 LL_CDEBUG_PAGE(page, "made ready\n");
259 page_cache_get(page);
261 /* if we left PageDirty we might get another writepage call
262 * in the future. list walkers are bright enough
263 * to check page dirty so we can leave it on whatever list
264 * its on. XXX also, we're called with the cli list so if
265 * we got the page cache list we'd create a lock inversion
266 * with the removepage path which gets the page lock then the
268 clear_page_dirty(page);
272 static int ll_ap_refresh_count(void *data, int cmd)
274 struct ll_async_page *llap;
277 /* readpage queues with _COUNT_STABLE, shouldn't get here. */
278 LASSERT(cmd != OBD_BRW_READ);
280 llap = llap_from_cookie(data);
282 RETURN(PTR_ERR(llap));
284 return ll_write_count(llap->llap_page);
287 void ll_inode_fill_obdo(struct inode *inode, int cmd, struct obdo *oa)
289 struct lov_stripe_md *lsm;
290 obd_flag valid_flags;
292 lsm = ll_i2info(inode)->lli_smd;
294 oa->o_id = lsm->lsm_object_id;
295 oa->o_valid = OBD_MD_FLID;
296 valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME;
297 if (cmd == OBD_BRW_WRITE) {
298 oa->o_valid |= OBD_MD_FLIFID | OBD_MD_FLEPOCH;
299 mdc_pack_fid(obdo_fid(oa), inode->i_ino, 0, inode->i_mode);
300 oa->o_easize = ll_i2info(inode)->lli_io_epoch;
302 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME;
305 obdo_from_inode(oa, inode, valid_flags);
308 static void ll_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
310 struct ll_async_page *llap;
313 llap = llap_from_cookie(data);
319 ll_inode_fill_obdo(llap->llap_page->mapping->host, cmd, oa);
323 static struct obd_async_page_ops ll_async_page_ops = {
324 .ap_make_ready = ll_ap_make_ready,
325 .ap_refresh_count = ll_ap_refresh_count,
326 .ap_fill_obdo = ll_ap_fill_obdo,
327 .ap_completion = ll_ap_completion,
330 #define page_llap(page) \
331 ((struct ll_async_page *)((page)->private & ~LL_PRIVBITS_MASK))
333 /* XXX have the exp be an argument? */
334 struct ll_async_page *llap_from_page(struct page *page)
336 struct ll_async_page *llap;
337 struct obd_export *exp;
338 struct inode *inode = page->mapping->host;
339 struct ll_sb_info *sbi = ll_i2sbi(inode);
343 llap = page_llap(page);
345 if (llap->llap_magic != LLAP_MAGIC)
346 RETURN(ERR_PTR(-EINVAL));
350 exp = ll_i2obdexp(page->mapping->host);
352 RETURN(ERR_PTR(-EINVAL));
354 OBD_ALLOC(llap, sizeof(*llap));
356 RETURN(ERR_PTR(-ENOMEM));
357 llap->llap_magic = LLAP_MAGIC;
358 rc = obd_prep_async_page(exp, ll_i2info(inode)->lli_smd,
360 (obd_off)page->index << PAGE_SHIFT,
361 &ll_async_page_ops, llap, &llap->llap_cookie);
363 OBD_FREE(llap, sizeof(*llap));
367 CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n", llap,
368 page, llap->llap_cookie, (obd_off)page->index << PAGE_SHIFT);
369 /* also zeroing the PRIVBITS low order bitflags */
370 page->private = (unsigned long)llap;
371 llap->llap_page = page;
373 spin_lock(&sbi->ll_pglist_lock);
374 sbi->ll_pglist_gen++;
375 list_add_tail(&llap->llap_proc_item, &sbi->ll_pglist);
376 spin_unlock(&sbi->ll_pglist_lock);
381 void lov_increase_kms(struct obd_export *exp, struct lov_stripe_md *lsm,
383 /* update our write count to account for i_size increases that may have
384 * happened since we've queued the page for io. */
386 /* be careful not to return success without setting the page Uptodate or
387 * the next pass through prepare_write will read in stale data from disk. */
388 int ll_commit_write(struct file *file, struct page *page, unsigned from,
391 struct inode *inode = page->mapping->host;
392 struct ll_inode_info *lli = ll_i2info(inode);
393 struct lov_stripe_md *lsm = lli->lli_smd;
394 struct obd_export *exp = NULL;
395 struct ll_async_page *llap;
400 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
401 LASSERT(inode == file->f_dentry->d_inode);
402 LASSERT(PageLocked(page));
404 CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
405 inode, page, from, to, page->index);
407 llap = llap_from_page(page);
409 RETURN(PTR_ERR(llap));
411 /* queue a write for some time in the future the first time we
413 if (!PageDirty(page)) {
414 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
415 LPROC_LL_DIRTY_MISSES);
417 exp = ll_i2obdexp(inode);
421 /* _make_ready only sees llap once we've unlocked the page */
422 llap->llap_write_queued = 1;
423 rc = obd_queue_async_io(exp, lsm, NULL, llap->llap_cookie,
424 OBD_BRW_WRITE, 0, 0, 0, 0);
425 if (rc != 0) { /* async failed, try sync.. */
426 struct obd_sync_io_container *osic;
429 llap->llap_write_queued = 0;
430 rc = obd_queue_sync_io(exp, lsm, NULL, osic,
432 OBD_BRW_WRITE, 0, to, 0);
436 rc = obd_trigger_sync_io(exp, lsm, NULL, osic);
440 rc = osic_wait(osic);
445 LL_CDEBUG_PAGE(page, "write queued\n");
446 //llap_write_pending(inode, llap);
448 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
449 LPROC_LL_DIRTY_HITS);
452 /* put the page in the page cache, from now on ll_removepage is
453 * responsible for cleaning up the llap */
454 set_page_dirty(page);
458 size = (((obd_off)page->index) << PAGE_SHIFT) + to;
459 lov_increase_kms(exp, lsm, size);
460 if (size > inode->i_size)
461 inode->i_size = size;
462 SetPageUptodate(page);
467 /* the kernel calls us here when a page is unhashed from the page cache.
468 * the page will be locked and the kernel is holding a spinlock, so
469 * we need to be careful. we're just tearing down our book-keeping
471 void ll_removepage(struct page *page)
473 struct inode *inode = page->mapping->host;
474 struct obd_export *exp;
475 struct ll_async_page *llap;
476 struct ll_sb_info *sbi = ll_i2sbi(inode);
480 LASSERT(!in_interrupt());
482 /* sync pages or failed read pages can leave pages in the page
483 * cache that don't have our data associated with them anymore */
484 if (page->private == 0) {
489 LL_CDEBUG_PAGE(page, "being evicted\n");
491 exp = ll_i2obdexp(inode);
493 CERROR("page %p ind %lu gave null export\n", page,
499 llap = llap_from_page(page);
501 CERROR("page %p ind %lu couldn't find llap: %ld\n", page,
502 page->index, PTR_ERR(llap));
507 //llap_write_complete(inode, llap);
508 rc = obd_teardown_async_page(exp, ll_i2info(inode)->lli_smd, NULL,
511 CERROR("page %p ind %lu failed: %d\n", page, page->index, rc);
513 /* this unconditional free is only safe because the page lock
514 * is providing exclusivity to memory pressure/truncate/writeback..*/
517 spin_lock(&sbi->ll_pglist_lock);
518 if (!list_empty(&llap->llap_proc_item))
519 list_del_init(&llap->llap_proc_item);
520 sbi->ll_pglist_gen++;
521 spin_unlock(&sbi->ll_pglist_lock);
522 OBD_FREE(llap, sizeof(*llap));
526 static int ll_page_matches(struct page *page)
528 struct lustre_handle match_lockh = {0};
529 struct inode *inode = page->mapping->host;
530 ldlm_policy_data_t page_extent;
534 page_extent.l_extent.start = (__u64)page->index << PAGE_CACHE_SHIFT;
535 page_extent.l_extent.end =
536 page_extent.l_extent.start + PAGE_CACHE_SIZE - 1;
537 flags = LDLM_FL_CBPENDING | LDLM_FL_BLOCK_GRANTED;
538 matches = obd_match(ll_i2sbi(inode)->ll_osc_exp,
539 ll_i2info(inode)->lli_smd, LDLM_EXTENT,
540 &page_extent, LCK_PR, &flags, inode, &match_lockh);
542 LL_CDEBUG_PAGE(page, "lock match failed\n");
546 obd_cancel(ll_i2sbi(inode)->ll_osc_exp,
547 ll_i2info(inode)->lli_smd, LCK_PR, &match_lockh);
552 static int ll_issue_page_read(struct obd_export *exp,
553 struct ll_async_page *llap, int defer_uptodate)
555 struct page *page = llap->llap_page;
558 /* we don't issue this page as URGENT so that it can be batched
559 * with other pages by the kernel's read-ahead. We have a strong
560 * requirement that readpage() callers must call wait_on_page()
561 * or lock_page() to get into ->sync_page() to trigger the IO */
562 llap->llap_defer_uptodate = defer_uptodate;
563 page_cache_get(page);
564 set_bit(LL_PRIVBITS_READ, &page->private); /* see ll_sync_page() */
565 rc = obd_queue_async_io(exp, ll_i2info(page->mapping->host)->lli_smd,
566 NULL, llap->llap_cookie, OBD_BRW_READ, 0,
567 PAGE_SIZE, 0, ASYNC_COUNT_STABLE);
569 LL_CDEBUG_PAGE(page, "read queueing failed\n");
570 clear_bit(LL_PRIVBITS_READ, &page->private);
571 page_cache_release(page);
576 static void ll_readahead(struct ll_readahead_state *ras,
577 struct obd_export *exp, struct address_space *mapping)
579 unsigned long i, start, end;
580 struct ll_async_page *llap;
584 if (mapping->host->i_size == 0)
587 spin_lock(&ras->ras_lock);
589 /* make sure to issue a window's worth of read-ahead pages */
591 start = end - ras->ras_window;
595 /* but don't iterate over pages that we've already issued. this
596 * will set start to end + 1 if we've already read-ahead up to
597 * ras_last sothe for() won't be entered */
598 if (ras->ras_next_index > start)
599 start = ras->ras_next_index;
601 ras->ras_next_index = end + 1;
603 CDEBUG(D_READA, "ni %lu last %lu win %lu: reading from %lu to %lu\n",
604 ras->ras_next_index, ras->ras_last, ras->ras_window,
607 spin_unlock(&ras->ras_lock);
609 /* clamp to filesize */
610 i = (mapping->host->i_size - 1) >> PAGE_CACHE_SHIFT;
613 for (i = start; i <= end; i++) {
614 /* grab_cache_page_nowait returns null if this races with
615 * truncating the page (page->mapping == NULL) */
616 page = grab_cache_page_nowait(mapping, i);
620 /* the book-keeping above promises that we've tried
621 * all the indices from start to end, so we don't
622 * stop if anyone returns an error. This may not be good. */
623 if (Page_Uptodate(page) || ll_page_matches(page) <= 0)
626 llap = llap_from_page(page);
627 if (IS_ERR(llap) || llap->llap_defer_uptodate)
630 rc = ll_issue_page_read(exp, llap, 1);
632 LL_CDEBUG_PAGE(page, "started read-ahead\n");
635 LL_CDEBUG_PAGE(page, "skipping read-ahead\n");
639 page_cache_release(page);
643 /* XXX this should really bubble up somehow. */
644 #define LL_RA_MIN ((unsigned long)PTL_MD_MAX_PAGES / 2)
645 #define LL_RA_MAX ((unsigned long)(32 * PTL_MD_MAX_PAGES))
647 /* called with the ras_lock held or from places where it doesn't matter */
648 static void ll_readahead_set(struct ll_readahead_state *ras,
651 ras->ras_next_index = index;
652 if (ras->ras_next_index != ~0UL)
653 ras->ras_next_index++;
654 ras->ras_window = LL_RA_MIN;
655 ras->ras_last = ras->ras_next_index + ras->ras_window;
656 if (ras->ras_last < ras->ras_next_index)
657 ras->ras_last = ~0UL;
658 CDEBUG(D_READA, "ni %lu last %lu win %lu: set %lu\n",
659 ras->ras_next_index, ras->ras_last, ras->ras_window,
663 void ll_readahead_init(struct ll_readahead_state *ras)
665 spin_lock_init(&ras->ras_lock);
666 ll_readahead_set(ras, 0);
669 static void ll_readahead_update(struct ll_readahead_state *ras,
670 unsigned long index, int hit)
672 unsigned long issued_start, new_last;
674 spin_lock(&ras->ras_lock);
676 /* we're interested in noticing the index's relation to the
677 * previously issued read-ahead pages */
678 issued_start = ras->ras_next_index - ras->ras_window - 1;
679 if (issued_start > ras->ras_next_index)
682 CDEBUG(D_READA, "ni %lu last %lu win %lu: %s ind %lu start %lu\n",
683 ras->ras_next_index, ras->ras_last, ras->ras_window,
684 hit ? "hit" : "miss", index, issued_start);
686 index == ras->ras_next_index && index == ras->ras_last + 1) {
687 /* special case the kernel's read-ahead running into the
688 * page just beyond our read-ahead window as an extension
689 * of our read-ahead. sigh. wishing it was easier to
690 * turn off 2.4's read-ahead. */
691 ras->ras_window = min(LL_RA_MAX, ras->ras_window + 1);
693 ras->ras_next_index = index + 1;
694 ras->ras_last = index;
696 (index > issued_start || ras->ras_next_index >= index)) {
697 /* deal with a miss way out of the window. we interpret
698 * this as a seek and restart the window */
699 ll_readahead_set(ras, index);
702 issued_start <= index && index < ras->ras_next_index) {
703 /* a miss inside the window? surely its memory pressure
704 * evicting our read pages before the app can see them.
705 * we shrink the window aggressively */
706 unsigned long old_window = ras->ras_window;
708 ras->ras_window = max(ras->ras_window / 2, LL_RA_MIN);
709 ras->ras_last -= old_window - ras->ras_window;
710 if (ras->ras_next_index > ras->ras_last)
711 ras->ras_next_index = ras->ras_last + 1;
712 CDEBUG(D_READA, "ni %lu last %lu win %lu: miss inside\n",
713 ras->ras_next_index, ras->ras_last, ras->ras_window);
716 issued_start <= index && index < ras->ras_next_index) {
717 /* a hit inside the window. grow the window by twice the
718 * number of pages that are satisified within the window. */
719 ras->ras_window = min(LL_RA_MAX, ras->ras_window + 2);
721 /* we want the next readahead pass to issue a windows worth
722 * beyond where the app currently is */
723 new_last = index + ras->ras_window;
724 if (new_last > ras->ras_last)
725 ras->ras_last = new_last;
727 CDEBUG(D_READA, "ni %lu last %lu win %lu: extended window/last\n",
728 ras->ras_next_index, ras->ras_last, ras->ras_window);
731 spin_unlock(&ras->ras_lock);
735 * for now we do our readpage the same on both 2.4 and 2.5. The kernel's
736 * read-ahead assumes it is valid to issue readpage all the way up to
737 * i_size, but our dlm locks make that not the case. We disable the
738 * kernel's read-ahead and do our own by walking ahead in the page cache
739 * checking for dlm lock coverage. the main difference between 2.4 and
740 * 2.6 is how read-ahead gets batched and issued, but we're using our own,
741 * so they look the same.
743 int ll_readpage(struct file *filp, struct page *page)
745 struct ll_file_data *fd = filp->private_data;
746 struct inode *inode = page->mapping->host;
747 struct obd_export *exp;
749 struct ll_async_page *llap;
752 LASSERT(PageLocked(page));
753 LASSERT(!PageUptodate(page));
754 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset="LPX64"\n",
755 inode->i_ino, inode->i_generation, inode,
756 (((obd_off)page->index) << PAGE_SHIFT));
757 LASSERT(atomic_read(&filp->f_dentry->d_inode->i_count) > 0);
759 exp = ll_i2obdexp(inode);
761 GOTO(out, rc = -EINVAL);
763 llap = llap_from_page(page);
765 GOTO(out, rc = PTR_ERR(llap));
767 if (llap->llap_defer_uptodate) {
768 ll_readahead_update(&fd->fd_ras, page->index, 1);
769 LL_CDEBUG_PAGE(page, "marking uptodate from defer\n");
770 SetPageUptodate(page);
771 ll_readahead(&fd->fd_ras, exp, page->mapping);
776 ll_readahead_update(&fd->fd_ras, page->index, 0);
778 rc = ll_page_matches(page);
783 static unsigned long next_print;
784 CDEBUG(D_INODE, "didn't match a lock\n");
785 if (time_after(jiffies, next_print)) {
786 next_print = jiffies + 30 * HZ;
787 CERROR("not covered by a lock (mmap?). check debug "
792 rc = ll_issue_page_read(exp, llap, 0);
794 LL_CDEBUG_PAGE(page, "queued readpage\n");
795 if ((ll_i2sbi(inode)->ll_flags & LL_SBI_READAHEAD))
796 ll_readahead(&fd->fd_ras, exp, page->mapping);
804 /* this is for read pages. we issue them as ready but not urgent. when
805 * someone waits on them we fire them off, hopefully merged with adjacent
806 * reads that were queued by read-ahead. */
807 int ll_sync_page(struct page *page)
809 struct obd_export *exp;
810 struct ll_async_page *llap;
814 /* we're using a low bit flag to signify that a queued read should
815 * be issued once someone goes to lock it. it is also cleared
816 * as the page is built into an RPC */
817 if (!test_and_clear_bit(LL_PRIVBITS_READ, &page->private))
820 /* careful to only deref page->mapping after checking the bit */
821 exp = ll_i2obdexp(page->mapping->host);
825 llap = llap_from_page(page);
827 RETURN(PTR_ERR(llap));
829 LL_CDEBUG_PAGE(page, "setting ready|urgent\n");
831 rc = obd_set_async_flags(exp, ll_i2info(page->mapping->host)->lli_smd,
832 NULL, llap->llap_cookie,
833 ASYNC_READY|ASYNC_URGENT);