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 rc = obd_punch(ll_i2obdexp(inode), &oa, lsm, inode->i_size,
130 OBD_OBJECT_EOF, NULL);
132 CERROR("obd_truncate fails (%d) ino %lu\n", rc, inode->i_ino);
134 obdo_to_inode(inode, &oa, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
135 OBD_MD_FLATIME | OBD_MD_FLMTIME |
142 int ll_prepare_write(struct file *file, struct page *page, unsigned from,
145 struct inode *inode = page->mapping->host;
146 struct ll_inode_info *lli = ll_i2info(inode);
147 struct lov_stripe_md *lsm = lli->lli_smd;
148 obd_off offset = ((obd_off)page->index) << PAGE_SHIFT;
154 if (!PageLocked(page))
157 /* Check to see if we should return -EIO right away */
160 pga.count = PAGE_SIZE;
163 oa.o_id = lsm->lsm_object_id;
164 oa.o_mode = inode->i_mode;
165 oa.o_valid = OBD_MD_FLID | OBD_MD_FLMODE | OBD_MD_FLTYPE;
167 rc = obd_brw(OBD_BRW_CHECK, ll_i2obdexp(inode), &oa, lsm, 1, &pga,
172 if (PageUptodate(page))
175 /* We're completely overwriting an existing page, so _don't_ set it up
176 * to date until commit_write */
177 if (from == 0 && to == PAGE_SIZE) {
178 POISON_PAGE(page, 0x11);
182 /* If are writing to a new page, no need to read old data.
183 * the extent locking and getattr procedures in ll_file_write have
184 * guaranteed that i_size is stable enough for our zeroing needs */
185 if (inode->i_size <= offset) {
186 memset(kmap(page), 0, PAGE_SIZE);
188 GOTO(prepare_done, rc = 0);
191 /* XXX could be an async ocp read.. read-ahead? */
192 rc = ll_brw(OBD_BRW_READ, inode, &oa, page, 0);
194 /* bug 1598: don't clobber blksize */
195 oa.o_valid &= ~(OBD_MD_FLSIZE | OBD_MD_FLBLKSZ);
196 obdo_refresh_inode(inode, &oa, oa.o_valid);
202 SetPageUptodate(page);
207 int ll_write_count(struct page *page)
209 struct inode *inode = page->mapping->host;
211 /* catch race with truncate */
212 if (((loff_t)page->index << PAGE_SHIFT) >= inode->i_size)
215 /* catch sub-page write at end of file */
216 if (((loff_t)page->index << PAGE_SHIFT) + PAGE_SIZE > inode->i_size)
217 return inode->i_size % PAGE_SIZE;
222 struct ll_async_page *llap_from_cookie(void *cookie)
224 struct ll_async_page *llap = cookie;
225 if (llap->llap_magic != LLAP_MAGIC)
226 return ERR_PTR(-EINVAL);
230 static int ll_ap_make_ready(void *data, int cmd)
232 struct ll_async_page *llap;
236 llap = llap_from_cookie(data);
240 page = llap->llap_page;
242 if (cmd == OBD_BRW_READ) {
243 /* _sync_page beat us to it and is about to call
244 * _set_async_flags which will fire off rpcs again */
245 if (!test_and_clear_bit(LL_PRIVBITS_READ, &page->private))
250 /* we're trying to write, but the page is locked.. come back later */
251 if (TryLockPage(page))
254 LL_CDEBUG_PAGE(page, "made ready\n");
255 page_cache_get(page);
257 /* if we left PageDirty we might get another writepage call
258 * in the future. list walkers are bright enough
259 * to check page dirty so we can leave it on whatever list
260 * its on. XXX also, we're called with the cli list so if
261 * we got the page cache list we'd create a lock inversion
262 * with the removepage path which gets the page lock then the
264 clear_page_dirty(page);
268 static int ll_ap_refresh_count(void *data, int cmd)
270 struct ll_async_page *llap;
273 /* readpage queues with _COUNT_STABLE, shouldn't get here. */
274 LASSERT(cmd != OBD_BRW_READ);
276 llap = llap_from_cookie(data);
278 RETURN(PTR_ERR(llap));
280 return ll_write_count(llap->llap_page);
283 void ll_inode_fill_obdo(struct inode *inode, int cmd, struct obdo *oa)
285 struct lov_stripe_md *lsm;
286 obd_flag valid_flags;
288 lsm = ll_i2info(inode)->lli_smd;
290 oa->o_id = lsm->lsm_object_id;
291 oa->o_valid = OBD_MD_FLID;
292 valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME;
293 if (cmd == OBD_BRW_WRITE) {
294 oa->o_valid |= OBD_MD_FLIFID | OBD_MD_FLEPOCH;
295 mdc_pack_fid(obdo_fid(oa), inode->i_ino, 0, inode->i_mode);
296 oa->o_easize = ll_i2info(inode)->lli_io_epoch;
298 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME;
301 obdo_from_inode(oa, inode, valid_flags);
304 static void ll_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
306 struct ll_async_page *llap;
309 llap = llap_from_cookie(data);
315 ll_inode_fill_obdo(llap->llap_page->mapping->host, cmd, oa);
319 static struct obd_async_page_ops ll_async_page_ops = {
320 .ap_make_ready = ll_ap_make_ready,
321 .ap_refresh_count = ll_ap_refresh_count,
322 .ap_fill_obdo = ll_ap_fill_obdo,
323 .ap_completion = ll_ap_completion,
326 #define page_llap(page) \
327 ((struct ll_async_page *)((page)->private & ~LL_PRIVBITS_MASK))
329 /* XXX have the exp be an argument? */
330 struct ll_async_page *llap_from_page(struct page *page)
332 struct ll_async_page *llap;
333 struct obd_export *exp;
334 struct inode *inode = page->mapping->host;
335 struct ll_sb_info *sbi = ll_i2sbi(inode);
339 llap = page_llap(page);
341 if (llap->llap_magic != LLAP_MAGIC)
342 RETURN(ERR_PTR(-EINVAL));
346 exp = ll_i2obdexp(page->mapping->host);
348 RETURN(ERR_PTR(-EINVAL));
350 OBD_ALLOC(llap, sizeof(*llap));
352 RETURN(ERR_PTR(-ENOMEM));
353 llap->llap_magic = LLAP_MAGIC;
354 rc = obd_prep_async_page(exp, ll_i2info(inode)->lli_smd,
356 (obd_off)page->index << PAGE_SHIFT,
357 &ll_async_page_ops, llap, &llap->llap_cookie);
359 OBD_FREE(llap, sizeof(*llap));
363 CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n", llap,
364 page, llap->llap_cookie, (obd_off)page->index << PAGE_SHIFT);
365 /* also zeroing the PRIVBITS low order bitflags */
366 page->private = (unsigned long)llap;
367 llap->llap_page = page;
369 spin_lock(&sbi->ll_pglist_lock);
370 sbi->ll_pglist_gen++;
371 list_add_tail(&llap->llap_proc_item, &sbi->ll_pglist);
372 spin_unlock(&sbi->ll_pglist_lock);
377 /* update our write count to account for i_size increases that may have
378 * happened since we've queued the page for io. */
380 /* be careful not to return success without setting the page Uptodate or
381 * the next pass through prepare_write will read in stale data from disk. */
382 int ll_commit_write(struct file *file, struct page *page, unsigned from,
385 struct inode *inode = page->mapping->host;
386 struct ll_inode_info *lli = ll_i2info(inode);
387 struct lov_stripe_md *lsm = lli->lli_smd;
388 struct obd_export *exp = NULL;
389 struct ll_async_page *llap;
394 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
395 LASSERT(inode == file->f_dentry->d_inode);
396 LASSERT(PageLocked(page));
398 CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
399 inode, page, from, to, page->index);
401 llap = llap_from_page(page);
403 RETURN(PTR_ERR(llap));
405 /* queue a write for some time in the future the first time we
407 if (!PageDirty(page)) {
408 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
409 LPROC_LL_DIRTY_MISSES);
411 exp = ll_i2obdexp(inode);
415 /* _make_ready only sees llap once we've unlocked the page */
416 llap->llap_write_queued = 1;
417 rc = obd_queue_async_io(exp, lsm, NULL, llap->llap_cookie,
418 OBD_BRW_WRITE, 0, 0, 0, 0);
419 if (rc != 0) { /* async failed, try sync.. */
420 struct obd_sync_io_container *osic;
423 llap->llap_write_queued = 0;
424 rc = obd_queue_sync_io(exp, lsm, NULL, osic,
426 OBD_BRW_WRITE, 0, to, 0);
430 rc = obd_trigger_sync_io(exp, lsm, NULL, osic);
434 rc = osic_wait(osic);
439 LL_CDEBUG_PAGE(page, "write queued\n");
440 //llap_write_pending(inode, llap);
442 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
443 LPROC_LL_DIRTY_HITS);
446 /* put the page in the page cache, from now on ll_removepage is
447 * responsible for cleaning up the llap */
448 set_page_dirty(page);
452 /* XXX needs to be pushed down to the OSC as EOC */
453 size = (((obd_off)page->index) << PAGE_SHIFT) + to;
454 if (size > inode->i_size) {
455 inode->i_size = size;
456 /* see commentary in file.c:ll_inode_getattr() */
457 set_bit(LLI_F_PREFER_EXTENDED_SIZE, &lli->lli_flags);
459 SetPageUptodate(page);
464 /* the kernel calls us here when a page is unhashed from the page cache.
465 * the page will be locked and the kernel is holding a spinlock, so
466 * we need to be careful. we're just tearing down our book-keeping
468 void ll_removepage(struct page *page)
470 struct inode *inode = page->mapping->host;
471 struct obd_export *exp;
472 struct ll_async_page *llap;
473 struct ll_sb_info *sbi = ll_i2sbi(inode);
477 LASSERT(!in_interrupt());
479 /* sync pages or failed read pages can leave pages in the page
480 * cache that don't have our data associated with them anymore */
481 if (page->private == 0) {
486 LL_CDEBUG_PAGE(page, "being evicted\n");
488 exp = ll_i2obdexp(inode);
490 CERROR("page %p ind %lu gave null export\n", page,
496 llap = llap_from_page(page);
498 CERROR("page %p ind %lu couldn't find llap: %ld\n", page,
499 page->index, PTR_ERR(llap));
504 //llap_write_complete(inode, llap);
505 rc = obd_teardown_async_page(exp, ll_i2info(inode)->lli_smd, NULL,
508 CERROR("page %p ind %lu failed: %d\n", page, page->index, rc);
510 /* this unconditional free is only safe because the page lock
511 * is providing exclusivity to memory pressure/truncate/writeback..*/
514 spin_lock(&sbi->ll_pglist_lock);
515 if (!list_empty(&llap->llap_proc_item))
516 list_del_init(&llap->llap_proc_item);
517 sbi->ll_pglist_gen++;
518 spin_unlock(&sbi->ll_pglist_lock);
519 OBD_FREE(llap, sizeof(*llap));
523 static int ll_page_matches(struct page *page)
525 struct lustre_handle match_lockh = {0};
526 struct inode *inode = page->mapping->host;
527 struct ldlm_extent page_extent;
531 page_extent.start = (__u64)page->index << PAGE_CACHE_SHIFT;
532 page_extent.end = page_extent.start + PAGE_CACHE_SIZE - 1;
533 flags = LDLM_FL_CBPENDING | LDLM_FL_BLOCK_GRANTED;
534 matches = obd_match(ll_i2sbi(inode)->ll_osc_exp,
535 ll_i2info(inode)->lli_smd, LDLM_EXTENT,
536 &page_extent, sizeof(page_extent),
537 LCK_PR, &flags, inode, &match_lockh);
539 LL_CDEBUG_PAGE(page, "lock match failed\n");
543 obd_cancel(ll_i2sbi(inode)->ll_osc_exp,
544 ll_i2info(inode)->lli_smd, LCK_PR, &match_lockh);
549 static int ll_issue_page_read(struct obd_export *exp,
550 struct ll_async_page *llap, int defer_uptodate)
552 struct page *page = llap->llap_page;
555 /* we don't issue this page as URGENT so that it can be batched
556 * with other pages by the kernel's read-ahead. We have a strong
557 * requirement that readpage() callers must call wait_on_page()
558 * or lock_page() to get into ->sync_page() to trigger the IO */
559 llap->llap_defer_uptodate = defer_uptodate;
560 page_cache_get(page);
561 set_bit(LL_PRIVBITS_READ, &page->private); /* see ll_sync_page() */
562 rc = obd_queue_async_io(exp, ll_i2info(page->mapping->host)->lli_smd,
563 NULL, llap->llap_cookie, OBD_BRW_READ, 0,
564 PAGE_SIZE, 0, ASYNC_COUNT_STABLE);
566 LL_CDEBUG_PAGE(page, "read queueing failed\n");
567 clear_bit(LL_PRIVBITS_READ, &page->private);
568 page_cache_release(page);
573 static void ll_readahead(struct ll_readahead_state *ras,
574 struct obd_export *exp, struct address_space *mapping)
576 unsigned long i, start, end;
577 struct ll_async_page *llap;
581 if (mapping->host->i_size == 0)
584 spin_lock(&ras->ras_lock);
586 /* make sure to issue a window's worth of read-ahead pages */
588 start = end - ras->ras_window;
592 /* but don't iterate over pages that we've already issued. this
593 * will set start to end + 1 if we've already read-ahead up to
594 * ras_last sothe for() won't be entered */
595 if (ras->ras_next_index > start)
596 start = ras->ras_next_index;
598 ras->ras_next_index = end + 1;
600 CDEBUG(D_READA, "ni %lu last %lu win %lu: reading from %lu to %lu\n",
601 ras->ras_next_index, ras->ras_last, ras->ras_window,
604 spin_unlock(&ras->ras_lock);
606 /* clamp to filesize */
607 i = (mapping->host->i_size - 1) >> PAGE_CACHE_SHIFT;
610 for (i = start; i <= end; i++) {
611 /* grab_cache_page_nowait returns null if this races with
612 * truncating the page (page->mapping == NULL) */
613 page = grab_cache_page_nowait(mapping, i);
617 /* the book-keeping above promises that we've tried
618 * all the indices from start to end, so we don't
619 * stop if anyone returns an error. This may not be good. */
620 if (Page_Uptodate(page) || ll_page_matches(page) <= 0)
623 llap = llap_from_page(page);
624 if (IS_ERR(llap) || llap->llap_defer_uptodate)
627 rc = ll_issue_page_read(exp, llap, 1);
629 LL_CDEBUG_PAGE(page, "started read-ahead\n");
632 LL_CDEBUG_PAGE(page, "skipping read-ahead\n");
636 page_cache_release(page);
640 /* XXX this should really bubble up somehow. */
641 #define LL_RA_MIN ((unsigned long)PTL_MD_MAX_PAGES / 2)
642 #define LL_RA_MAX ((unsigned long)(32 * PTL_MD_MAX_PAGES))
644 /* called with the ras_lock held or from places where it doesn't matter */
645 static void ll_readahead_set(struct ll_readahead_state *ras,
648 ras->ras_next_index = index;
649 if (ras->ras_next_index != ~0UL)
650 ras->ras_next_index++;
651 ras->ras_window = LL_RA_MIN;
652 ras->ras_last = ras->ras_next_index + ras->ras_window;
653 if (ras->ras_last < ras->ras_next_index)
654 ras->ras_last = ~0UL;
655 CDEBUG(D_READA, "ni %lu last %lu win %lu: set %lu\n",
656 ras->ras_next_index, ras->ras_last, ras->ras_window,
660 void ll_readahead_init(struct ll_readahead_state *ras)
662 spin_lock_init(&ras->ras_lock);
663 ll_readahead_set(ras, 0);
666 static void ll_readahead_update(struct ll_readahead_state *ras,
667 unsigned long index, int hit)
669 unsigned long issued_start, new_last;
671 spin_lock(&ras->ras_lock);
673 /* we're interested in noticing the index's relation to the
674 * previously issued read-ahead pages */
675 issued_start = ras->ras_next_index - ras->ras_window - 1;
676 if (issued_start > ras->ras_next_index)
679 CDEBUG(D_READA, "ni %lu last %lu win %lu: %s ind %lu start %lu\n",
680 ras->ras_next_index, ras->ras_last, ras->ras_window,
681 hit ? "hit" : "miss", index, issued_start);
683 index == ras->ras_next_index && index == ras->ras_last + 1) {
684 /* special case the kernel's read-ahead running into the
685 * page just beyond our read-ahead window as an extension
686 * of our read-ahead. sigh. wishing it was easier to
687 * turn off 2.4's read-ahead. */
688 ras->ras_window = min(LL_RA_MAX, ras->ras_window + 1);
690 ras->ras_next_index = index + 1;
691 ras->ras_last = index;
693 (index > issued_start || ras->ras_next_index >= index)) {
694 /* deal with a miss way out of the window. we interpret
695 * this as a seek and restart the window */
696 ll_readahead_set(ras, index);
699 issued_start <= index && index < ras->ras_next_index) {
700 /* a miss inside the window? surely its memory pressure
701 * evicting our read pages before the app can see them.
702 * we shrink the window aggressively */
703 unsigned long old_window = ras->ras_window;
705 ras->ras_window = max(ras->ras_window / 2, LL_RA_MIN);
706 ras->ras_last -= old_window - ras->ras_window;
707 if (ras->ras_next_index > ras->ras_last)
708 ras->ras_next_index = ras->ras_last + 1;
709 CDEBUG(D_READA, "ni %lu last %lu win %lu: miss inside\n",
710 ras->ras_next_index, ras->ras_last, ras->ras_window);
713 issued_start <= index && index < ras->ras_next_index) {
714 /* a hit inside the window. grow the window by twice the
715 * number of pages that are satisified within the window. */
716 ras->ras_window = min(LL_RA_MAX, ras->ras_window + 2);
718 /* we want the next readahead pass to issue a windows worth
719 * beyond where the app currently is */
720 new_last = index + ras->ras_window;
721 if (new_last > ras->ras_last)
722 ras->ras_last = new_last;
724 CDEBUG(D_READA, "ni %lu last %lu win %lu: extended window/last\n",
725 ras->ras_next_index, ras->ras_last, ras->ras_window);
728 spin_unlock(&ras->ras_lock);
732 * for now we do our readpage the same on both 2.4 and 2.5. The kernel's
733 * read-ahead assumes it is valid to issue readpage all the way up to
734 * i_size, but our dlm locks make that not the case. We disable the
735 * kernel's read-ahead and do our own by walking ahead in the page cache
736 * checking for dlm lock coverage. the main difference between 2.4 and
737 * 2.6 is how read-ahead gets batched and issued, but we're using our own,
738 * so they look the same.
740 int ll_readpage(struct file *filp, struct page *page)
742 struct ll_file_data *fd = filp->private_data;
743 struct inode *inode = page->mapping->host;
744 struct obd_export *exp;
746 struct ll_async_page *llap;
749 LASSERT(PageLocked(page));
750 LASSERT(!PageUptodate(page));
751 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset="LPX64"\n",
752 inode->i_ino, inode->i_generation, inode,
753 (((obd_off)page->index) << PAGE_SHIFT));
754 LASSERT(atomic_read(&filp->f_dentry->d_inode->i_count) > 0);
756 exp = ll_i2obdexp(inode);
758 GOTO(out, rc = -EINVAL);
760 llap = llap_from_page(page);
762 GOTO(out, rc = PTR_ERR(llap));
764 if (llap->llap_defer_uptodate) {
765 ll_readahead_update(&fd->fd_ras, page->index, 1);
766 LL_CDEBUG_PAGE(page, "marking uptodate from defer\n");
767 SetPageUptodate(page);
768 ll_readahead(&fd->fd_ras, exp, page->mapping);
773 ll_readahead_update(&fd->fd_ras, page->index, 0);
775 rc = ll_page_matches(page);
780 static unsigned long next_print;
781 CDEBUG(D_INODE, "didn't match a lock");
782 if (time_after(jiffies, next_print)) {
783 next_print = jiffies + 30 * HZ;
784 CERROR("not covered by a lock (mmap?). check debug "
789 rc = ll_issue_page_read(exp, llap, 0);
791 LL_CDEBUG_PAGE(page, "queued readpage\n");
792 if ((ll_i2sbi(inode)->ll_flags & LL_SBI_READAHEAD))
793 ll_readahead(&fd->fd_ras, exp, page->mapping);
801 /* this is for read pages. we issue them as ready but not urgent. when
802 * someone waits on them we fire them off, hopefully merged with adjacent
803 * reads that were queued by read-ahead. */
804 int ll_sync_page(struct page *page)
806 struct obd_export *exp;
807 struct ll_async_page *llap;
811 /* we're using a low bit flag to signify that a queued read should
812 * be issued once someone goes to lock it. it is also cleared
813 * as the page is built into an RPC */
814 if (!test_and_clear_bit(LL_PRIVBITS_READ, &page->private))
817 /* careful to only deref page->mapping after checking the bit */
818 exp = ll_i2obdexp(page->mapping->host);
822 llap = llap_from_page(page);
824 RETURN(PTR_ERR(llap));
826 LL_CDEBUG_PAGE(page, "setting ready|urgent\n");
828 rc = obd_set_async_flags(exp, ll_i2info(page->mapping->host)->lli_smd,
829 NULL, llap->llap_cookie,
830 ASYNC_READY|ASYNC_URGENT);