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);
128 CDEBUG(D_INODE, "truncate on inode %lu with no objects\n",
133 if (lli->lli_size_pid != current->pid)
135 LASSERT(atomic_read(&lli->lli_size_sem.count) <= 0);
137 if (lov_merge_size(lsm, 0) == inode->i_size) {
138 CDEBUG(D_VFSTRACE, "skipping punch for "LPX64" (size = %llu)\n",
139 lsm->lsm_object_id, inode->i_size);
143 CDEBUG(D_INFO, "calling punch for "LPX64" (new size %llu)\n",
144 lsm->lsm_object_id, inode->i_size);
148 CERROR("cannot alloc oa, error %d\n",
154 oa->o_id = lsm->lsm_object_id;
155 oa->o_gr = lsm->lsm_object_gr;
156 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
157 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLMODE |
158 OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME);
160 obd_adjust_kms(ll_i2dtexp(inode), lsm, inode->i_size, 1);
162 lli->lli_size_pid = 0;
163 up(&lli->lli_size_sem);
165 rc = obd_punch(ll_i2dtexp(inode), oa, lsm, inode->i_size,
166 OBD_OBJECT_EOF, NULL);
168 CERROR("obd_truncate fails (%d) ino %lu\n", rc, inode->i_ino);
170 obdo_to_inode(inode, oa, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
171 OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME);
179 LASSERT(atomic_read(&lli->lli_size_sem.count) <= 0);
180 up(&lli->lli_size_sem);
183 int ll_prepare_write(struct file *file, struct page *page, unsigned from,
186 struct inode *inode = page->mapping->host;
187 struct ll_inode_info *lli = ll_i2info(inode);
188 struct lov_stripe_md *lsm = lli->lli_smd;
189 obd_off offset = ((obd_off)page->index) << PAGE_SHIFT;
190 struct obdo *oa = NULL;
196 LASSERT(PageLocked(page));
197 (void)llap_cast_private(page); /* assertion */
199 /* Check to see if we should return -EIO right away */
201 pga.disk_offset = pga.page_offset = offset;
202 pga.count = PAGE_SIZE;
209 oa->o_id = lsm->lsm_object_id;
210 oa->o_gr = lsm->lsm_object_gr;
211 oa->o_mode = inode->i_mode;
212 oa->o_valid = OBD_MD_FLID | OBD_MD_FLMODE |
213 OBD_MD_FLTYPE | OBD_MD_FLGROUP;
215 rc = obd_brw(OBD_BRW_CHECK, ll_i2dtexp(inode), oa, lsm,
218 GOTO(out_free_oa, rc);
220 if (PageUptodate(page))
221 GOTO(out_free_oa, 0);
223 /* We're completely overwriting an existing page, so _don't_ set it up
224 * to date until commit_write */
225 if (from == 0 && to == PAGE_SIZE) {
226 POISON_PAGE(page, 0x11);
227 GOTO(out_free_oa, 0);
230 /* If are writing to a new page, no need to read old data. The extent
231 * locking will have updated the KMS, and for our purposes here we can
232 * treat it like i_size. */
233 down(&lli->lli_size_sem);
234 kms = lov_merge_size(lsm, 1);
235 up(&lli->lli_size_sem);
237 memset(kmap(page), 0, PAGE_SIZE);
239 GOTO(prepare_done, rc = 0);
242 /* XXX could be an async ocp read.. read-ahead? */
243 rc = ll_brw(OBD_BRW_READ, inode, oa, page, 0);
245 /* bug 1598: don't clobber blksize */
246 oa->o_valid &= ~(OBD_MD_FLSIZE | OBD_MD_FLBLKSZ);
247 obdo_refresh_inode(inode, oa, oa->o_valid);
253 SetPageUptodate(page);
259 static int ll_ap_make_ready(void *data, int cmd)
261 struct ll_async_page *llap;
265 llap = LLAP_FROM_COOKIE(data);
266 page = llap->llap_page;
268 LASSERT(cmd != OBD_BRW_READ);
270 /* we're trying to write, but the page is locked.. come back later */
271 if (TryLockPage(page))
274 LL_CDEBUG_PAGE(D_PAGE, page, "made ready\n");
275 page_cache_get(page);
277 /* if we left PageDirty we might get another writepage call
278 * in the future. list walkers are bright enough
279 * to check page dirty so we can leave it on whatever list
280 * its on. XXX also, we're called with the cli list so if
281 * we got the page cache list we'd create a lock inversion
282 * with the removepage path which gets the page lock then the
284 clear_page_dirty(page);
288 /* We have two reasons for giving llite the opportunity to change the
289 * write length of a given queued page as it builds the RPC containing
292 * 1) Further extending writes may have landed in the page cache
293 * since a partial write first queued this page requiring us
294 * to write more from the page cache. (No further races are possible, since
295 * by the time this is called, the page is locked.)
296 * 2) We might have raced with truncate and want to avoid performing
297 * write RPCs that are just going to be thrown away by the
298 * truncate's punch on the storage targets.
300 * The kms serves these purposes as it is set at both truncate and extending
303 static int ll_ap_refresh_count(void *data, int cmd)
305 struct ll_inode_info *lli;
306 struct ll_async_page *llap;
307 struct lov_stripe_md *lsm;
312 /* readpage queues with _COUNT_STABLE, shouldn't get here. */
313 LASSERT(cmd != OBD_BRW_READ);
315 llap = LLAP_FROM_COOKIE(data);
316 page = llap->llap_page;
317 lli = ll_i2info(page->mapping->host);
320 down(&lli->lli_size_sem);
321 kms = lov_merge_size(lsm, 1);
322 up(&lli->lli_size_sem);
324 /* catch race with truncate */
325 if (((__u64)page->index << PAGE_SHIFT) >= kms)
328 /* catch sub-page write at end of file */
329 if (((__u64)page->index << PAGE_SHIFT) + PAGE_SIZE > kms)
330 return kms % PAGE_SIZE;
335 void ll_inode_fill_obdo(struct inode *inode, int cmd, struct obdo *oa)
337 struct lov_stripe_md *lsm;
338 obd_valid valid_flags;
340 lsm = ll_i2info(inode)->lli_smd;
342 oa->o_id = lsm->lsm_object_id;
343 oa->o_gr = lsm->lsm_object_gr;
344 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
345 valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME;
346 if (cmd == OBD_BRW_WRITE) {
347 oa->o_valid |= OBD_MD_FLIFID | OBD_MD_FLEPOCH;
348 mdc_pack_id(obdo_id(oa), inode->i_ino, 0, inode->i_mode,
349 id_group(&ll_i2info(inode)->lli_id),
350 id_fid(&ll_i2info(inode)->lli_id));
352 oa->o_easize = ll_i2info(inode)->lli_io_epoch;
353 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME;
356 obdo_from_inode(oa, inode, valid_flags);
359 static void ll_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
361 struct ll_async_page *llap;
364 llap = LLAP_FROM_COOKIE(data);
365 ll_inode_fill_obdo(llap->llap_page->mapping->host, cmd, oa);
369 static struct obd_async_page_ops ll_async_page_ops = {
370 .ap_make_ready = ll_ap_make_ready,
371 .ap_refresh_count = ll_ap_refresh_count,
372 .ap_fill_obdo = ll_ap_fill_obdo,
373 .ap_completion = ll_ap_completion,
377 struct ll_async_page *llap_cast_private(struct page *page)
379 struct ll_async_page *llap = (struct ll_async_page *)page->private;
381 LASSERTF(llap == NULL || llap->llap_magic == LLAP_MAGIC,
382 "page %p private %lu gave magic %d which != %d\n",
383 page, page->private, llap->llap_magic, LLAP_MAGIC);
388 /* XXX have the exp be an argument? */
389 struct ll_async_page *llap_from_page(struct page *page, unsigned origin)
391 struct ll_async_page *llap;
392 struct obd_export *exp;
393 struct inode *inode = page->mapping->host;
394 struct ll_sb_info *sbi = ll_i2sbi(inode);
398 LASSERTF(origin < LLAP__ORIGIN_MAX, "%u\n", origin);
400 llap = llap_cast_private(page);
404 exp = ll_i2dtexp(page->mapping->host);
406 RETURN(ERR_PTR(-EINVAL));
408 OBD_ALLOC(llap, sizeof(*llap));
410 RETURN(ERR_PTR(-ENOMEM));
411 llap->llap_magic = LLAP_MAGIC;
412 rc = obd_prep_async_page(exp, ll_i2info(inode)->lli_smd, NULL, page,
413 (obd_off)page->index << PAGE_SHIFT,
414 &ll_async_page_ops, llap, &llap->llap_cookie);
416 OBD_FREE(llap, sizeof(*llap));
420 CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n", llap,
421 page, llap->llap_cookie, (obd_off)page->index << PAGE_SHIFT);
422 /* also zeroing the PRIVBITS low order bitflags */
423 __set_page_ll_data(page, llap);
424 llap->llap_page = page;
426 spin_lock(&sbi->ll_lock);
427 sbi->ll_pglist_gen++;
428 list_add_tail(&llap->llap_proc_item, &sbi->ll_pglist);
429 spin_unlock(&sbi->ll_lock);
432 llap->llap_origin = origin;
436 static int queue_or_sync_write(struct obd_export *exp,
437 struct lov_stripe_md *lsm,
438 struct ll_async_page *llap,
440 obd_flags async_flags)
442 struct obd_io_group *oig;
446 /* _make_ready only sees llap once we've unlocked the page */
447 llap->llap_write_queued = 1;
448 rc = obd_queue_async_io(exp, lsm, NULL, llap->llap_cookie,
449 OBD_BRW_WRITE, 0, 0, 0, async_flags);
451 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "write queued\n");
452 //llap_write_pending(inode, llap);
456 llap->llap_write_queued = 0;
462 rc = obd_queue_group_io(exp, lsm, NULL, oig, llap->llap_cookie,
463 OBD_BRW_WRITE, 0, to, 0, ASYNC_READY |
464 ASYNC_URGENT | ASYNC_COUNT_STABLE |
469 rc = obd_trigger_group_io(exp, lsm, NULL, oig);
475 if (!rc && async_flags & ASYNC_READY)
476 unlock_page(llap->llap_page);
478 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
479 "sync write returned %d\n", rc);
488 /* be careful not to return success without setting the page Uptodate or
489 * the next pass through prepare_write will read in stale data from disk. */
490 int ll_commit_write(struct file *file, struct page *page, unsigned from,
493 struct inode *inode = page->mapping->host;
494 struct ll_inode_info *lli = ll_i2info(inode);
495 struct lov_stripe_md *lsm = lli->lli_smd;
496 struct obd_export *exp = NULL;
497 struct ll_async_page *llap;
502 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
503 LASSERT(inode == file->f_dentry->d_inode);
504 LASSERT(PageLocked(page));
506 CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
507 inode, page, from, to, page->index);
509 llap = llap_from_page(page, LLAP_ORIGIN_COMMIT_WRITE);
511 RETURN(PTR_ERR(llap));
513 exp = ll_i2dtexp(inode);
517 /* queue a write for some time in the future the first time we
519 if (!PageDirty(page)) {
520 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
521 LPROC_LL_DIRTY_MISSES);
523 rc = queue_or_sync_write(exp, ll_i2info(inode)->lli_smd,
528 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
529 LPROC_LL_DIRTY_HITS);
532 /* put the page in the page cache, from now on ll_removepage is
533 * responsible for cleaning up the llap.
534 * don't dirty the page if it has been write out in q_o_s_w */
535 if (llap->llap_write_queued)
536 set_page_dirty(page);
539 size = (((obd_off)page->index) << PAGE_SHIFT) + to;
540 down(&lli->lli_size_sem);
542 obd_adjust_kms(exp, lsm, size, 0);
543 if (size > inode->i_size)
544 inode->i_size = size;
545 SetPageUptodate(page);
546 } else if (size > inode->i_size) {
547 /* this page beyond the pales of i_size, so it can't be
548 * truncated in ll_p_r_e during lock revoking. we must
549 * teardown our book-keeping here. */
552 up(&lli->lli_size_sem);
556 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, unsigned long len)
558 struct ll_ra_info *ra = &sbi->ll_ra_info;
562 spin_lock(&sbi->ll_lock);
563 ret = min(ra->ra_max_pages - ra->ra_cur_pages, len);
564 ra->ra_cur_pages += ret;
565 spin_unlock(&sbi->ll_lock);
570 static void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
572 struct ll_ra_info *ra = &sbi->ll_ra_info;
573 spin_lock(&sbi->ll_lock);
574 LASSERTF(ra->ra_cur_pages >= len, "r_c_p %lu len %lu\n",
575 ra->ra_cur_pages, len);
576 ra->ra_cur_pages -= len;
577 spin_unlock(&sbi->ll_lock);
580 int ll_writepage(struct page *page)
582 struct inode *inode = page->mapping->host;
583 struct obd_export *exp;
584 struct ll_async_page *llap;
588 LASSERT(!PageDirty(page));
589 LASSERT(PageLocked(page));
591 exp = ll_i2dtexp(inode);
593 GOTO(out, rc = -EINVAL);
595 llap = llap_from_page(page, LLAP_ORIGIN_WRITEPAGE);
597 GOTO(out, rc = PTR_ERR(llap));
599 page_cache_get(page);
600 if (llap->llap_write_queued) {
601 LL_CDEBUG_PAGE(D_PAGE, page, "marking urgent\n");
602 rc = obd_set_async_flags(exp, ll_i2info(inode)->lli_smd, NULL,
604 ASYNC_READY | ASYNC_URGENT);
606 rc = queue_or_sync_write(exp, ll_i2info(inode)->lli_smd, llap,
607 PAGE_SIZE, ASYNC_READY |
611 page_cache_release(page);
619 /* called for each page in a completed rpc.*/
620 void ll_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
622 struct ll_async_page *llap;
626 llap = LLAP_FROM_COOKIE(data);
627 page = llap->llap_page;
628 LASSERT(PageLocked(page));
630 LL_CDEBUG_PAGE(D_PAGE, page, "completing cmd %d with %d\n", cmd, rc);
632 if (cmd == OBD_BRW_READ && llap->llap_defer_uptodate)
633 ll_ra_count_put(ll_i2sbi(page->mapping->host), 1);
636 if (cmd == OBD_BRW_READ) {
637 if (!llap->llap_defer_uptodate)
638 SetPageUptodate(page);
640 llap->llap_write_queued = 0;
642 ClearPageError(page);
644 if (cmd == OBD_BRW_READ)
645 llap->llap_defer_uptodate = 0;
651 if (0 && cmd == OBD_BRW_WRITE) {
652 llap_write_complete(page->mapping->host, llap);
653 ll_try_done_writing(page->mapping->host);
656 if (PageWriteback(page)) {
657 end_page_writeback(page);
659 page_cache_release(page);
663 /* the kernel calls us here when a page is unhashed from the page cache.
664 * the page will be locked and the kernel is holding a spinlock, so
665 * we need to be careful. we're just tearing down our book-keeping
667 void ll_removepage(struct page *page)
669 struct inode *inode = page->mapping->host;
670 struct obd_export *exp;
671 struct ll_async_page *llap;
672 struct ll_sb_info *sbi = ll_i2sbi(inode);
676 LASSERT(!in_interrupt());
678 /* sync pages or failed read pages can leave pages in the page
679 * cache that don't have our data associated with them anymore */
680 if (page->private == 0) {
685 LL_CDEBUG_PAGE(D_PAGE, page, "being evicted\n");
687 exp = ll_i2dtexp(inode);
689 CERROR("page %p ind %lu gave null export\n", page, page->index);
694 llap = llap_from_page(page, 0);
696 CERROR("page %p ind %lu couldn't find llap: %ld\n", page,
697 page->index, PTR_ERR(llap));
702 //llap_write_complete(inode, llap);
703 rc = obd_teardown_async_page(exp, ll_i2info(inode)->lli_smd, NULL,
706 CERROR("page %p ind %lu failed: %d\n", page, page->index, rc);
708 /* this unconditional free is only safe because the page lock
709 * is providing exclusivity to memory pressure/truncate/writeback..*/
710 __clear_page_ll_data(page);
712 spin_lock(&sbi->ll_lock);
713 if (!list_empty(&llap->llap_proc_item))
714 list_del_init(&llap->llap_proc_item);
715 sbi->ll_pglist_gen++;
716 spin_unlock(&sbi->ll_lock);
717 OBD_FREE(llap, sizeof(*llap));
721 static int ll_page_matches(struct page *page, int fd_flags, int readahead)
723 struct lustre_handle match_lockh = {0};
724 struct inode *inode = page->mapping->host;
725 ldlm_policy_data_t page_extent;
729 if (fd_flags & LL_FILE_GROUP_LOCKED)
732 page_extent.l_extent.start = (__u64)page->index << PAGE_CACHE_SHIFT;
733 page_extent.l_extent.end =
734 page_extent.l_extent.start + PAGE_CACHE_SIZE - 1;
735 flags = LDLM_FL_TEST_LOCK;
737 flags |= LDLM_FL_CBPENDING | LDLM_FL_BLOCK_GRANTED;
738 matches = obd_match(ll_i2sbi(inode)->ll_dt_exp,
739 ll_i2info(inode)->lli_smd, LDLM_EXTENT,
740 &page_extent, LCK_PR | LCK_PW, &flags, inode,
745 static int ll_issue_page_read(struct obd_export *exp,
746 struct ll_async_page *llap,
747 struct obd_io_group *oig, int defer)
749 struct page *page = llap->llap_page;
752 page_cache_get(page);
753 llap->llap_defer_uptodate = defer;
754 llap->llap_ra_used = 0;
755 rc = obd_queue_group_io(exp, ll_i2info(page->mapping->host)->lli_smd,
756 NULL, oig, llap->llap_cookie, OBD_BRW_READ, 0,
757 PAGE_SIZE, 0, ASYNC_COUNT_STABLE | ASYNC_READY
760 LL_CDEBUG_PAGE(D_ERROR, page, "read queue failed: rc %d\n", rc);
761 page_cache_release(page);
766 static void ll_ra_stats_inc_unlocked(struct ll_ra_info *ra, enum ra_stat which)
768 LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
769 ra->ra_stats[which]++;
772 static void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which)
774 struct ll_sb_info *sbi = ll_i2sbi(mapping->host);
775 struct ll_ra_info *ra = &ll_i2sbi(mapping->host)->ll_ra_info;
777 spin_lock(&sbi->ll_lock);
778 ll_ra_stats_inc_unlocked(ra, which);
779 spin_unlock(&sbi->ll_lock);
782 void ll_ra_accounting(struct page *page, struct address_space *mapping)
784 struct ll_async_page *llap;
786 llap = llap_from_page(page, LLAP_ORIGIN_WRITEPAGE);
790 if (!llap->llap_defer_uptodate || llap->llap_ra_used)
793 ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
796 #define RAS_CDEBUG(ras) \
797 CDEBUG(D_READA, "lrp %lu c %lu ws %lu wl %lu nra %lu\n", \
798 ras->ras_last_readpage, ras->ras_consecutive, \
799 ras->ras_window_start, ras->ras_window_len, \
800 ras->ras_next_readahead);
802 static int index_in_window(unsigned long index, unsigned long point,
803 unsigned long before, unsigned long after)
805 unsigned long start = point - before, end = point + after;
812 return start <= index && index <= end;
815 static int ll_readahead(struct ll_readahead_state *ras,
816 struct obd_export *exp, struct address_space *mapping,
817 struct obd_io_group *oig, int flags)
819 unsigned long i, start = 0, end = 0, reserved;
820 struct ll_async_page *llap;
822 int rc, ret = 0, match_failed = 0;
826 kms = lov_merge_size(ll_i2info(mapping->host)->lli_smd, 1);
828 ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN);
831 spin_lock(&ras->ras_lock);
833 /* reserve a part of the read-ahead window that we'll be issuing */
834 if (ras->ras_window_len) {
835 start = ras->ras_next_readahead;
836 end = ras->ras_window_start + ras->ras_window_len - 1;
837 end = min(end, (unsigned long)(kms >> PAGE_CACHE_SHIFT));
838 ras->ras_next_readahead = max(end, end + 1);
843 spin_unlock(&ras->ras_lock);
846 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
850 reserved = ll_ra_count_get(ll_i2sbi(mapping->host), end - start + 1);
851 if (reserved < end - start + 1)
852 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
854 for (i = start; reserved > 0 && !match_failed && i <= end; i++) {
855 /* skip locked pages from previous readpage calls */
856 page = grab_cache_page_nowait(mapping, i);
858 CDEBUG(D_READA, "g_c_p_n failed\n");
862 /* we do this first so that we can see the page in the /proc
864 llap = llap_from_page(page, LLAP_ORIGIN_READAHEAD);
865 if (IS_ERR(llap) || llap->llap_defer_uptodate)
868 /* skip completed pages */
869 if (Page_Uptodate(page))
872 /* bail when we hit the end of the lock. */
873 if ((rc = ll_page_matches(page, flags, 1)) <= 0) {
874 LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
875 "lock match failed: rc %d\n", rc);
876 ll_ra_stats_inc(mapping, RA_STAT_FAILED_MATCH);
881 rc = ll_issue_page_read(exp, llap, oig, 1);
885 LL_CDEBUG_PAGE(D_READA| D_PAGE, page,
886 "started read-ahead\n");
890 LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
891 "skipping read-ahead\n");
895 page_cache_release(page);
898 LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
900 ll_ra_count_put(ll_i2sbi(mapping->host), reserved);
902 if (i == end + 1 && end == (kms >> PAGE_CACHE_SHIFT))
903 ll_ra_stats_inc(mapping, RA_STAT_EOF);
905 /* if we didn't get to the end of the region we reserved from
906 * the ras we need to go back and update the ras so that the
907 * next read-ahead tries from where we left off. we only do so
908 * if the region we failed to issue read-ahead on is still ahead
909 * of the app and behind the next index to start read-ahead from */
911 spin_lock(&ras->ras_lock);
912 if (i < ras->ras_next_readahead &&
913 index_in_window(i, ras->ras_window_start, 0,
914 ras->ras_window_len)) {
915 ras->ras_next_readahead = i;
918 spin_unlock(&ras->ras_lock);
924 static void ras_set_start(struct ll_readahead_state *ras, unsigned long index)
926 ras->ras_window_start = index & (~(PTLRPC_MAX_BRW_PAGES - 1));
929 /* called with the ras_lock held or from places where it doesn't matter */
930 static void ras_reset(struct ll_readahead_state *ras, unsigned long index)
932 ras->ras_last_readpage = index;
933 ras->ras_consecutive = 1;
934 ras->ras_window_len = 0;
935 ras_set_start(ras, index);
936 ras->ras_next_readahead = ras->ras_window_start;
941 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
943 spin_lock_init(&ras->ras_lock);
947 static void ras_update(struct ll_sb_info *sbi, struct ll_readahead_state *ras,
948 unsigned long index, unsigned hit)
950 struct ll_ra_info *ra = &sbi->ll_ra_info;
954 spin_lock(&sbi->ll_lock);
955 spin_lock(&ras->ras_lock);
957 ll_ra_stats_inc_unlocked(ra, hit ? RA_STAT_HIT : RA_STAT_MISS);
959 /* reset the read-ahead window in two cases. First when the app seeks
960 * or reads to some other part of the file. Secondly if we get a
961 * read-ahead miss that we think we've previously issued. This can
962 * be a symptom of there being so many read-ahead pages that the VM is
963 * reclaiming it before we get to it. */
964 if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
966 ll_ra_stats_inc_unlocked(ra, RA_STAT_DISTANT_READPAGE);
967 } else if (!hit && ras->ras_window_len &&
968 index < ras->ras_next_readahead &&
969 index_in_window(index, ras->ras_window_start, 0,
970 ras->ras_window_len)) {
972 ll_ra_stats_inc_unlocked(ra, RA_STAT_MISS_IN_WINDOW);
976 ras_reset(ras, index);
980 ras->ras_last_readpage = index;
981 ras->ras_consecutive++;
982 ras_set_start(ras, index);
983 ras->ras_next_readahead = max(ras->ras_window_start,
984 ras->ras_next_readahead);
986 /* wait for a few pages to arrive before issuing readahead to avoid
987 * the worst overutilization */
988 if (ras->ras_consecutive == 3) {
989 ras->ras_window_len = PTLRPC_MAX_BRW_PAGES;
993 /* we need to increase the window sometimes. we'll arbitrarily
994 * do it half-way through the pages in an rpc */
995 if ((index & (PTLRPC_MAX_BRW_PAGES - 1)) ==
996 (PTLRPC_MAX_BRW_PAGES >> 1)) {
997 ras->ras_window_len += PTLRPC_MAX_BRW_PAGES;
998 ras->ras_window_len = min(ras->ras_window_len,
1006 spin_unlock(&ras->ras_lock);
1007 spin_unlock(&sbi->ll_lock);
1011 * for now we do our readpage the same on both 2.4 and 2.5. The kernel's
1012 * read-ahead assumes it is valid to issue readpage all the way up to
1013 * i_size, but our dlm locks make that not the case. We disable the
1014 * kernel's read-ahead and do our own by walking ahead in the page cache
1015 * checking for dlm lock coverage. the main difference between 2.4 and
1016 * 2.6 is how read-ahead gets batched and issued, but we're using our own,
1017 * so they look the same.
1019 int ll_readpage(struct file *filp, struct page *page)
1021 struct ll_file_data *fd = filp->private_data;
1022 struct inode *inode = page->mapping->host;
1023 struct obd_export *exp;
1024 struct ll_async_page *llap;
1025 struct obd_io_group *oig = NULL;
1029 LASSERT(PageLocked(page));
1030 LASSERT(!PageUptodate(page));
1031 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset="LPX64"\n",
1032 inode->i_ino, inode->i_generation, inode,
1033 (((obd_off)page->index) << PAGE_SHIFT));
1034 LASSERT(atomic_read(&filp->f_dentry->d_inode->i_count) > 0);
1036 rc = oig_init(&oig);
1040 exp = ll_i2dtexp(inode);
1042 GOTO(out, rc = -EINVAL);
1044 llap = llap_from_page(page, LLAP_ORIGIN_READPAGE);
1046 GOTO(out, rc = PTR_ERR(llap));
1048 if (ll_i2sbi(inode)->ll_flags & LL_SBI_READAHEAD)
1049 ras_update(ll_i2sbi(inode), &fd->fd_ras, page->index,
1050 llap->llap_defer_uptodate);
1052 if (llap->llap_defer_uptodate) {
1053 llap->llap_ra_used = 1;
1054 rc = ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1057 obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd,
1059 LL_CDEBUG_PAGE(D_PAGE, page, "marking uptodate from defer\n");
1060 SetPageUptodate(page);
1062 GOTO(out_oig, rc = 0);
1065 rc = ll_page_matches(page, fd->fd_flags, 0);
1067 LL_CDEBUG_PAGE(D_ERROR, page, "lock match failed: rc %d\n", rc);
1072 CWARN("ino %lu page %lu (%llu) not covered by "
1073 "a lock (mmap?). check debug logs.\n",
1074 inode->i_ino, page->index,
1075 (long long)page->index << PAGE_CACHE_SHIFT);
1078 rc = ll_issue_page_read(exp, llap, oig, 0);
1082 LL_CDEBUG_PAGE(D_PAGE, page, "queued readpage\n");
1083 if (ll_i2sbi(inode)->ll_flags & LL_SBI_READAHEAD)
1084 ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1087 rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);