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 struct ll_async_page *llap_from_cookie(void *cookie)
213 struct ll_async_page *llap = cookie;
214 if (llap->llap_magic != LLAP_MAGIC)
215 return ERR_PTR(-EINVAL);
219 static int ll_ap_make_ready(void *data, int cmd)
221 struct ll_async_page *llap;
225 llap = llap_from_cookie(data);
229 page = llap->llap_page;
231 if (cmd == OBD_BRW_READ)
234 /* we're trying to write, but the page is locked.. come back later */
235 if (TryLockPage(page))
238 LL_CDEBUG_PAGE(D_PAGE, page, "made ready\n");
239 page_cache_get(page);
241 /* if we left PageDirty we might get another writepage call
242 * in the future. list walkers are bright enough
243 * to check page dirty so we can leave it on whatever list
244 * its on. XXX also, we're called with the cli list so if
245 * we got the page cache list we'd create a lock inversion
246 * with the removepage path which gets the page lock then the
248 clear_page_dirty(page);
252 /* We have two reasons for giving llite the opportunity to change the
253 * write length of a given queued page as it builds the RPC containing
256 * 1) Further extending writes may have landed in the page cache
257 * since a partial write first queued this page requiring us
258 * to write more from the page cache.
259 * 2) We might have raced with truncate and want to avoid performing
260 * write RPCs that are just going to be thrown away by the
261 * truncate's punch on the storage targets.
263 * The kms serves these purposes as it is set at both truncate and extending
266 static int ll_ap_refresh_count(void *data, int cmd)
268 struct ll_async_page *llap;
269 struct lov_stripe_md *lsm;
274 /* readpage queues with _COUNT_STABLE, shouldn't get here. */
275 LASSERT(cmd != OBD_BRW_READ);
277 llap = llap_from_cookie(data);
279 RETURN(PTR_ERR(llap));
281 page = llap->llap_page;
282 lsm = ll_i2info(page->mapping->host)->lli_smd;
283 kms = lov_merge_size(lsm, 1);
285 /* catch race with truncate */
286 if (((__u64)page->index << PAGE_SHIFT) >= kms)
289 /* catch sub-page write at end of file */
290 if (((__u64)page->index << PAGE_SHIFT) + PAGE_SIZE > kms)
291 return kms % PAGE_SIZE;
296 void ll_inode_fill_obdo(struct inode *inode, int cmd, struct obdo *oa)
298 struct lov_stripe_md *lsm;
299 obd_flag valid_flags;
301 lsm = ll_i2info(inode)->lli_smd;
303 oa->o_id = lsm->lsm_object_id;
304 oa->o_valid = OBD_MD_FLID;
305 valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME;
306 if (cmd == OBD_BRW_WRITE) {
307 oa->o_valid |= OBD_MD_FLIFID | OBD_MD_FLEPOCH;
308 mdc_pack_fid(obdo_fid(oa), inode->i_ino, 0, inode->i_mode);
309 oa->o_easize = ll_i2info(inode)->lli_io_epoch;
311 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME;
314 obdo_from_inode(oa, inode, valid_flags);
317 static void ll_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
319 struct ll_async_page *llap;
322 llap = llap_from_cookie(data);
328 ll_inode_fill_obdo(llap->llap_page->mapping->host, cmd, oa);
332 static struct obd_async_page_ops ll_async_page_ops = {
333 .ap_make_ready = ll_ap_make_ready,
334 .ap_refresh_count = ll_ap_refresh_count,
335 .ap_fill_obdo = ll_ap_fill_obdo,
336 .ap_completion = ll_ap_completion,
339 /* XXX have the exp be an argument? */
340 struct ll_async_page *llap_from_page(struct page *page)
342 struct ll_async_page *llap;
343 struct obd_export *exp;
344 struct inode *inode = page->mapping->host;
345 struct ll_sb_info *sbi = ll_i2sbi(inode);
349 llap = (struct ll_async_page *)page->private;
351 if (llap->llap_magic != LLAP_MAGIC)
352 RETURN(ERR_PTR(-EINVAL));
356 exp = ll_i2obdexp(page->mapping->host);
358 RETURN(ERR_PTR(-EINVAL));
360 OBD_ALLOC(llap, sizeof(*llap));
362 RETURN(ERR_PTR(-ENOMEM));
363 llap->llap_magic = LLAP_MAGIC;
364 rc = obd_prep_async_page(exp, ll_i2info(inode)->lli_smd, NULL, page,
365 (obd_off)page->index << PAGE_SHIFT,
366 &ll_async_page_ops, llap, &llap->llap_cookie);
368 OBD_FREE(llap, sizeof(*llap));
372 CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n", llap,
373 page, llap->llap_cookie, (obd_off)page->index << PAGE_SHIFT);
374 /* also zeroing the PRIVBITS low order bitflags */
375 page->private = (unsigned long)llap;
376 llap->llap_page = page;
378 spin_lock(&sbi->ll_pglist_lock);
379 sbi->ll_pglist_gen++;
380 list_add_tail(&llap->llap_proc_item, &sbi->ll_pglist);
381 spin_unlock(&sbi->ll_pglist_lock);
386 void lov_increase_kms(struct obd_export *exp, struct lov_stripe_md *lsm,
388 /* update our write count to account for i_size increases that may have
389 * happened since we've queued the page for io. */
391 /* be careful not to return success without setting the page Uptodate or
392 * the next pass through prepare_write will read in stale data from disk. */
393 int ll_commit_write(struct file *file, struct page *page, unsigned from,
396 struct inode *inode = page->mapping->host;
397 struct ll_inode_info *lli = ll_i2info(inode);
398 struct lov_stripe_md *lsm = lli->lli_smd;
399 struct obd_export *exp = NULL;
400 struct ll_async_page *llap;
405 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
406 LASSERT(inode == file->f_dentry->d_inode);
407 LASSERT(PageLocked(page));
409 CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
410 inode, page, from, to, page->index);
412 llap = llap_from_page(page);
414 RETURN(PTR_ERR(llap));
416 /* queue a write for some time in the future the first time we
418 if (!PageDirty(page)) {
419 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
420 LPROC_LL_DIRTY_MISSES);
422 exp = ll_i2obdexp(inode);
426 /* _make_ready only sees llap once we've unlocked the page */
427 llap->llap_write_queued = 1;
428 rc = obd_queue_async_io(exp, lsm, NULL, llap->llap_cookie,
429 OBD_BRW_WRITE, 0, 0, 0, 0);
430 if (rc != 0) { /* async failed, try sync.. */
431 struct obd_io_group *oig;
436 llap->llap_write_queued = 0;
437 rc = obd_queue_group_io(exp, lsm, NULL, oig,
439 OBD_BRW_WRITE, 0, to, 0,
440 ASYNC_READY | ASYNC_URGENT |
447 rc = obd_trigger_group_io(exp, lsm, NULL, oig);
456 LL_CDEBUG_PAGE(D_PAGE, page, "write queued\n");
457 //llap_write_pending(inode, llap);
459 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
460 LPROC_LL_DIRTY_HITS);
463 /* put the page in the page cache, from now on ll_removepage is
464 * responsible for cleaning up the llap */
465 set_page_dirty(page);
469 size = (((obd_off)page->index) << PAGE_SHIFT) + to;
470 lov_increase_kms(exp, lsm, size);
471 if (size > inode->i_size)
472 inode->i_size = size;
473 SetPageUptodate(page);
478 /* called for each page in a completed rpc.*/
479 void ll_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
481 struct ll_async_page *llap;
485 llap = llap_from_cookie(data);
491 page = llap->llap_page;
492 LASSERT(PageLocked(page));
494 LL_CDEBUG_PAGE(D_PAGE, page, "completing cmd %d with %d\n", cmd, rc);
497 if (cmd == OBD_BRW_READ) {
498 if (!llap->llap_defer_uptodate)
499 SetPageUptodate(page);
501 llap->llap_write_queued = 0;
503 ClearPageError(page);
505 if (cmd == OBD_BRW_READ)
506 llap->llap_defer_uptodate = 0;
513 if (0 && cmd == OBD_BRW_WRITE) {
514 llap_write_complete(page->mapping->host, llap);
515 ll_try_done_writing(page->mapping->host);
518 page_cache_release(page);
522 /* the kernel calls us here when a page is unhashed from the page cache.
523 * the page will be locked and the kernel is holding a spinlock, so
524 * we need to be careful. we're just tearing down our book-keeping
526 void ll_removepage(struct page *page)
528 struct inode *inode = page->mapping->host;
529 struct obd_export *exp;
530 struct ll_async_page *llap;
531 struct ll_sb_info *sbi = ll_i2sbi(inode);
535 LASSERT(!in_interrupt());
537 /* sync pages or failed read pages can leave pages in the page
538 * cache that don't have our data associated with them anymore */
539 if (page->private == 0) {
544 LL_CDEBUG_PAGE(D_PAGE, page, "being evicted\n");
546 exp = ll_i2obdexp(inode);
548 CERROR("page %p ind %lu gave null export\n", page, page->index);
553 llap = llap_from_page(page);
555 CERROR("page %p ind %lu couldn't find llap: %ld\n", page,
556 page->index, PTR_ERR(llap));
561 //llap_write_complete(inode, llap);
562 rc = obd_teardown_async_page(exp, ll_i2info(inode)->lli_smd, NULL,
565 CERROR("page %p ind %lu failed: %d\n", page, page->index, rc);
567 /* this unconditional free is only safe because the page lock
568 * is providing exclusivity to memory pressure/truncate/writeback..*/
571 spin_lock(&sbi->ll_pglist_lock);
572 if (!list_empty(&llap->llap_proc_item))
573 list_del_init(&llap->llap_proc_item);
574 sbi->ll_pglist_gen++;
575 spin_unlock(&sbi->ll_pglist_lock);
576 OBD_FREE(llap, sizeof(*llap));
580 static int ll_page_matches(struct page *page, int fd_flags)
582 struct lustre_handle match_lockh = {0};
583 struct inode *inode = page->mapping->host;
584 ldlm_policy_data_t page_extent;
588 if (fd_flags & LL_FILE_GROUP_LOCKED)
591 page_extent.l_extent.start = (__u64)page->index << PAGE_CACHE_SHIFT;
592 page_extent.l_extent.end =
593 page_extent.l_extent.start + PAGE_CACHE_SIZE - 1;
594 flags = LDLM_FL_CBPENDING | LDLM_FL_BLOCK_GRANTED | LDLM_FL_TEST_LOCK;
595 matches = obd_match(ll_i2sbi(inode)->ll_osc_exp,
596 ll_i2info(inode)->lli_smd, LDLM_EXTENT,
597 &page_extent, LCK_PR | LCK_PW, &flags, inode,
602 static int ll_issue_page_read(struct obd_export *exp,
603 struct ll_async_page *llap,
604 struct obd_io_group *oig, int defer)
606 struct page *page = llap->llap_page;
609 page_cache_get(page);
610 llap->llap_defer_uptodate = defer;
611 rc = obd_queue_group_io(exp, ll_i2info(page->mapping->host)->lli_smd,
612 NULL, oig, llap->llap_cookie, OBD_BRW_READ, 0,
613 PAGE_SIZE, 0, ASYNC_COUNT_STABLE);
615 LL_CDEBUG_PAGE(D_ERROR, page, "read queue failed: rc %d\n", rc);
616 page_cache_release(page);
621 #define LL_RA_MIN(inode) ((unsigned long)PTLRPC_MAX_BRW_PAGES / 2)
622 #define LL_RA_MAX(inode) ((ll_i2info(inode)->lli_smd->lsm_xfersize * 3) >> \
625 static void ll_readahead(struct ll_readahead_state *ras,
626 struct obd_export *exp, struct address_space *mapping,
627 struct obd_io_group *oig, int flags)
629 unsigned long i, start, end;
630 struct ll_async_page *llap;
634 if (mapping->host->i_size == 0)
637 spin_lock(&ras->ras_lock);
639 /* make sure to issue a window's worth of read-ahead pages */
641 start = end - ras->ras_window;
645 /* but don't iterate over pages that we've already issued. this
646 * will set start to end + 1 if we've already read-ahead up to
647 * ras_last sothe for() won't be entered */
648 if (ras->ras_next_index > start)
649 start = ras->ras_next_index;
651 ras->ras_next_index = end + 1;
653 CDEBUG(D_READA, "ni %lu last %lu win %lu: reading from %lu to %lu\n",
654 ras->ras_next_index, ras->ras_last, ras->ras_window,
657 spin_unlock(&ras->ras_lock);
659 /* clamp to filesize */
660 i = (mapping->host->i_size - 1) >> PAGE_CACHE_SHIFT;
663 for (i = start; i <= end; i++) {
664 /* grab_cache_page_nowait returns null if this races with
665 * truncating the page (page->mapping == NULL) */
666 page = grab_cache_page_nowait(mapping, i);
670 /* the book-keeping above promises that we've tried
671 * all the indices from start to end, so we don't
672 * stop if anyone returns an error. This may not be good. */
673 if (Page_Uptodate(page))
676 if ((rc = ll_page_matches(page, flags)) <= 0) {
677 LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
678 "lock match failed: rc %d\n", rc);
682 llap = llap_from_page(page);
683 if (IS_ERR(llap) || llap->llap_defer_uptodate)
686 rc = ll_issue_page_read(exp, llap, oig, 1);
688 LL_CDEBUG_PAGE(D_PAGE, page, "started read-ahead\n");
691 LL_CDEBUG_PAGE(D_PAGE, page, "skipping read-ahead\n");
695 page_cache_release(page);
699 /* called with the ras_lock held or from places where it doesn't matter */
700 static void ll_readahead_set(struct inode *inode,
701 struct ll_readahead_state *ras,
704 ras->ras_next_index = index;
705 if (ras->ras_next_index != ~0UL)
706 ras->ras_next_index++;
707 ras->ras_window = LL_RA_MIN(inode);
708 ras->ras_last = ras->ras_next_index + ras->ras_window;
709 if (ras->ras_last < ras->ras_next_index)
710 ras->ras_last = ~0UL;
711 CDEBUG(D_READA, "ni %lu last %lu win %lu: set %lu\n",
712 ras->ras_next_index, ras->ras_last, ras->ras_window,
716 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
718 spin_lock_init(&ras->ras_lock);
719 ll_readahead_set(inode, ras, 0);
722 static void ll_readahead_update(struct inode *inode,
723 struct ll_readahead_state *ras,
724 unsigned long index, int hit)
726 unsigned long issued_start, new_last;
728 spin_lock(&ras->ras_lock);
730 /* we're interested in noticing the index's relation to the
731 * previously issued read-ahead pages */
732 issued_start = ras->ras_next_index - ras->ras_window - 1;
733 if (issued_start > ras->ras_next_index)
736 CDEBUG(D_READA, "ni %lu last %lu win %lu: %s ind %lu start %lu\n",
737 ras->ras_next_index, ras->ras_last, ras->ras_window,
738 hit ? "hit" : "miss", index, issued_start);
740 index == ras->ras_next_index && index == ras->ras_last + 1) {
741 /* special case the kernel's read-ahead running into the
742 * page just beyond our read-ahead window as an extension
743 * of our read-ahead. sigh. wishing it was easier to
744 * turn off 2.4's read-ahead. */
745 ras->ras_window = min(LL_RA_MAX(inode), ras->ras_window + 1);
747 ras->ras_next_index = index + 1;
748 ras->ras_last = index;
750 (index > issued_start || ras->ras_next_index >= index)) {
751 /* deal with a miss way out of the window. we interpret
752 * this as a seek and restart the window */
753 ll_readahead_set(inode, ras, index);
756 issued_start <= index && index < ras->ras_next_index) {
757 /* a miss inside the window? surely its memory pressure
758 * evicting our read pages before the app can see them.
759 * we shrink the window aggressively */
760 unsigned long old_window = ras->ras_window;
762 ras->ras_window = max(ras->ras_window / 2, LL_RA_MIN(inode));
763 ras->ras_last -= old_window - ras->ras_window;
764 if (ras->ras_next_index > ras->ras_last)
765 ras->ras_next_index = ras->ras_last + 1;
766 CDEBUG(D_READA, "ni %lu last %lu win %lu: miss inside\n",
767 ras->ras_next_index, ras->ras_last, ras->ras_window);
770 issued_start <= index && index < ras->ras_next_index) {
771 /* a hit inside the window. grow the window by twice the
772 * number of pages that are satisified within the window. */
773 ras->ras_window = min(LL_RA_MAX(inode), ras->ras_window + 2);
775 /* we want the next readahead pass to issue a windows worth
776 * beyond where the app currently is */
777 new_last = index + ras->ras_window;
778 if (new_last > ras->ras_last)
779 ras->ras_last = new_last;
781 CDEBUG(D_READA, "ni %lu last %lu win %lu: extended window/last\n",
782 ras->ras_next_index, ras->ras_last, ras->ras_window);
785 spin_unlock(&ras->ras_lock);
789 * for now we do our readpage the same on both 2.4 and 2.5. The kernel's
790 * read-ahead assumes it is valid to issue readpage all the way up to
791 * i_size, but our dlm locks make that not the case. We disable the
792 * kernel's read-ahead and do our own by walking ahead in the page cache
793 * checking for dlm lock coverage. the main difference between 2.4 and
794 * 2.6 is how read-ahead gets batched and issued, but we're using our own,
795 * so they look the same.
797 int ll_readpage(struct file *filp, struct page *page)
799 struct ll_file_data *fd = filp->private_data;
800 struct inode *inode = page->mapping->host;
801 struct obd_export *exp;
802 struct ll_async_page *llap;
803 struct obd_io_group *oig = NULL;
807 LASSERT(PageLocked(page));
808 LASSERT(!PageUptodate(page));
809 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset="LPX64"\n",
810 inode->i_ino, inode->i_generation, inode,
811 (((obd_off)page->index) << PAGE_SHIFT));
812 LASSERT(atomic_read(&filp->f_dentry->d_inode->i_count) > 0);
818 exp = ll_i2obdexp(inode);
820 GOTO(out, rc = -EINVAL);
822 llap = llap_from_page(page);
824 GOTO(out, rc = PTR_ERR(llap));
826 if (llap->llap_defer_uptodate) {
827 ll_readahead_update(inode, &fd->fd_ras, page->index, 1);
828 ll_readahead(&fd->fd_ras, exp, page->mapping, oig,fd->fd_flags);
829 obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL,
831 LL_CDEBUG_PAGE(D_PAGE, page, "marking uptodate from defer\n");
832 SetPageUptodate(page);
834 GOTO(out_oig, rc = 0);
837 ll_readahead_update(inode, &fd->fd_ras, page->index, 0);
839 rc = ll_page_matches(page, fd->fd_flags);
841 LL_CDEBUG_PAGE(D_ERROR, page, "lock match failed: rc %d\n", rc);
846 static unsigned long next_print;
847 CDEBUG(D_INODE, "ino %lu page %lu (%llu) didn't match a lock\n",
848 inode->i_ino, page->index,
849 (long long)page->index << PAGE_CACHE_SHIFT);
850 if (time_after(jiffies, next_print)) {
851 CERROR("ino %lu page %lu (%llu) not covered by "
852 "a lock (mmap?). check debug logs.\n",
853 inode->i_ino, page->index,
854 (long long)page->index << PAGE_CACHE_SHIFT);
855 ldlm_dump_all_namespaces();
856 if (next_print == 0) {
857 CERROR("%s\n", portals_debug_dumpstack());
858 portals_debug_dumplog();
860 next_print = jiffies + 30 * HZ;
864 rc = ll_issue_page_read(exp, llap, oig, 0);
868 LL_CDEBUG_PAGE(D_PAGE, page, "queued readpage\n");
869 if ((ll_i2sbi(inode)->ll_flags & LL_SBI_READAHEAD))
870 ll_readahead(&fd->fd_ras, exp, page->mapping, oig,fd->fd_flags);
872 rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
884 /* this is for read pages. we issue them as ready but not urgent. when
885 * someone waits on them we fire them off, hopefully merged with adjacent
886 * reads that were queued by read-ahead. */
887 int ll_sync_page(struct page *page)
889 struct obd_export *exp;
890 struct ll_async_page *llap;
894 /* we're using a low bit flag to signify that a queued read should
895 * be issued once someone goes to lock it. it is also cleared
896 * as the page is built into an RPC */
897 if (!test_and_clear_bit(LL_PRIVBITS_READ, &page->private))
900 /* careful to only deref page->mapping after checking the bit */
901 exp = ll_i2obdexp(page->mapping->host);
905 llap = llap_from_page(page);
907 RETURN(PTR_ERR(llap));
909 LL_CDEBUG_PAGE(D_PAGE, page, "setting ready|urgent\n");
911 rc = obd_set_async_flags(exp, ll_i2info(page->mapping->host)->lli_smd,
912 NULL, llap->llap_cookie,
913 ASYNC_READY|ASYNC_URGENT);