1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Lustre Lite I/O page cache routines shared by different kernel revs
6 * Copyright (c) 2001-2003 Cluster File Systems, Inc.
8 * This file is part of Lustre, http://www.lustre.org.
10 * Lustre is free software; you can redistribute it and/or
11 * modify it under the terms of version 2 of the GNU General Public
12 * License as published by the Free Software Foundation.
14 * Lustre is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 * GNU General Public License for more details.
19 * You should have received a copy of the GNU General Public License
20 * along with Lustre; if not, write to the Free Software
21 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 #include <linux/config.h>
25 #include <linux/kernel.h>
27 #include <linux/string.h>
28 #include <linux/stat.h>
29 #include <linux/errno.h>
30 #include <linux/smp_lock.h>
31 #include <linux/unistd.h>
32 #include <linux/version.h>
33 #include <asm/system.h>
34 #include <asm/uaccess.h>
37 #include <linux/stat.h>
38 #include <asm/uaccess.h>
39 #include <asm/segment.h>
41 #include <linux/pagemap.h>
42 #include <linux/smp_lock.h>
44 #define DEBUG_SUBSYSTEM S_LLITE
46 #include <linux/lustre_mds.h>
47 #include <linux/lustre_lite.h>
48 #include "llite_internal.h"
49 #include <linux/lustre_compat25.h>
51 #ifndef list_for_each_prev_safe
52 #define list_for_each_prev_safe(pos, n, head) \
53 for (pos = (head)->prev, n = pos->prev; pos != (head); \
54 pos = n, n = pos->prev )
57 /* SYNCHRONOUS I/O to object storage for an inode */
58 static int ll_brw(int cmd, struct inode *inode, struct obdo *oa,
59 struct page *page, int flags)
61 struct ll_inode_info *lli = ll_i2info(inode);
62 struct lov_stripe_md *lsm = lli->lli_smd;
68 do_gettimeofday(&start);
71 pg.disk_offset = pg.page_offset = ((obd_off)page->index) << PAGE_SHIFT;
73 if (cmd == OBD_BRW_WRITE &&
74 (pg.disk_offset + PAGE_SIZE > inode->i_size))
75 pg.count = inode->i_size % PAGE_SIZE;
79 CDEBUG(D_PAGE, "%s %d bytes ino %lu at "LPU64"/"LPX64"\n",
80 cmd & OBD_BRW_WRITE ? "write" : "read", pg.count, inode->i_ino,
81 pg.disk_offset, pg.disk_offset);
83 CERROR("ZERO COUNT: ino %lu: size %p:%Lu(%p:%Lu) idx %lu off "
84 LPU64"\n", inode->i_ino, inode, inode->i_size,
85 page->mapping->host, page->mapping->host->i_size,
86 page->index, pg.disk_offset);
91 if (cmd == OBD_BRW_WRITE)
92 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
93 LPROC_LL_BRW_WRITE, pg.count);
95 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
96 LPROC_LL_BRW_READ, pg.count);
97 rc = obd_brw(cmd, ll_i2dtexp(inode), oa, lsm, 1, &pg, NULL);
99 obdo_to_inode(inode, oa, OBD_MD_FLBLOCKS);
101 CERROR("error from obd_brw: rc = %d\n", rc);
102 ll_stime_record(ll_i2sbi(inode), &start,
103 &ll_i2sbi(inode)->ll_brw_stime);
107 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
110 * this isn't where truncate starts. roughly:
111 * sys_truncate->ll_setattr_raw->vmtruncate->ll_truncate
112 * we grab the lock back in setattr_raw to avoid races.
114 * must be called with lli_size_sem held.
116 void ll_truncate(struct inode *inode)
118 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
119 struct ll_inode_info *lli = ll_i2info(inode);
120 struct obdo *oa = NULL;
124 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) to %llu\n", inode->i_ino,
125 inode->i_generation, inode, inode->i_size);
127 if (lli->lli_size_pid != current->pid) {
133 CDEBUG(D_INODE, "truncate on inode %lu with no objects\n",
138 LASSERT(atomic_read(&lli->lli_size_sem.count) <= 0);
140 if (lov_merge_size(lsm, 0) == inode->i_size) {
141 CDEBUG(D_VFSTRACE, "skipping punch for "LPX64" (size = %llu)\n",
142 lsm->lsm_object_id, inode->i_size);
146 CDEBUG(D_INFO, "calling punch for "LPX64" (new size %llu)\n",
147 lsm->lsm_object_id, inode->i_size);
151 CERROR("cannot alloc oa, error %d\n",
157 oa->o_id = lsm->lsm_object_id;
158 oa->o_gr = lsm->lsm_object_gr;
159 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
160 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLMODE |
161 OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME);
163 obd_adjust_kms(ll_i2dtexp(inode), lsm, inode->i_size, 1);
165 lli->lli_size_pid = 0;
166 up(&lli->lli_size_sem);
168 rc = obd_punch(ll_i2dtexp(inode), oa, lsm, inode->i_size,
169 OBD_OBJECT_EOF, NULL);
171 CERROR("obd_truncate fails (%d) ino %lu\n", rc, inode->i_ino);
173 obdo_to_inode(inode, oa, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
174 OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME);
182 LASSERT(atomic_read(&lli->lli_size_sem.count) <= 0);
183 up(&lli->lli_size_sem);
186 int ll_prepare_write(struct file *file, struct page *page,
187 unsigned from, unsigned to)
189 struct inode *inode = page->mapping->host;
190 struct ll_inode_info *lli = ll_i2info(inode);
191 struct lov_stripe_md *lsm = lli->lli_smd;
192 obd_off offset = ((obd_off)page->index) << PAGE_SHIFT;
193 struct obdo *oa = NULL;
199 LASSERT(PageLocked(page));
200 (void)llap_cast_private(page); /* assertion */
202 /* Check to see if we should return -EIO right away */
204 pga.disk_offset = pga.page_offset = offset;
205 pga.count = PAGE_SIZE;
212 oa->o_id = lsm->lsm_object_id;
213 oa->o_gr = lsm->lsm_object_gr;
214 oa->o_mode = inode->i_mode;
216 oa->o_valid = OBD_MD_FLID | OBD_MD_FLMODE |
217 OBD_MD_FLTYPE | OBD_MD_FLGROUP;
220 * needed for quota to create OSS object on write with correct
223 oa->o_uid = inode->i_uid;
224 oa->o_valid |= OBD_MD_FLUID;
226 oa->o_gid = inode->i_gid;
227 oa->o_valid |= OBD_MD_FLGID;
229 /* putting there also fid, needed for quota too. */
230 memcpy(obdo_id(oa), &lli->lli_id, sizeof(lli->lli_id));
231 oa->o_valid |= OBD_MD_FLINLINE;
233 rc = obd_brw(OBD_BRW_CHECK, ll_i2dtexp(inode),
234 oa, lsm, 1, &pga, NULL);
236 GOTO(out_free_oa, rc);
238 if (PageUptodate(page))
239 GOTO(out_free_oa, 0);
241 /* We're completely overwriting an existing page, so _don't_ set it up
242 * to date until commit_write */
243 if (from == 0 && to == PAGE_SIZE) {
244 POISON_PAGE(page, 0x11);
245 GOTO(out_free_oa, 0);
248 /* If are writing to a new page, no need to read old data. The extent
249 * locking will have updated the KMS, and for our purposes here we can
250 * treat it like i_size. */
251 down(&lli->lli_size_sem);
252 kms = lov_merge_size(lsm, 1);
253 up(&lli->lli_size_sem);
255 memset(kmap(page), 0, PAGE_SIZE);
257 GOTO(prepare_done, rc = 0);
260 /* XXX could be an async ocp read.. read-ahead? */
261 rc = ll_brw(OBD_BRW_READ, inode, oa, page, 0);
263 /* bug 1598: don't clobber blksize */
264 oa->o_valid &= ~(OBD_MD_FLSIZE | OBD_MD_FLBLKSZ);
265 obdo_refresh_inode(inode, oa, oa->o_valid);
271 SetPageUptodate(page);
277 static int ll_ap_make_ready(void *data, int cmd)
279 struct ll_async_page *llap;
283 llap = LLAP_FROM_COOKIE(data);
284 page = llap->llap_page;
286 LASSERT(cmd != OBD_BRW_READ);
288 /* we're trying to write, but the page is locked.. come back later */
289 if (TryLockPage(page))
292 LL_CDEBUG_PAGE(D_PAGE, page, "made ready\n");
293 page_cache_get(page);
295 /* if we left PageDirty we might get another writepage call
296 * in the future. list walkers are bright enough
297 * to check page dirty so we can leave it on whatever list
298 * its on. XXX also, we're called with the cli list so if
299 * we got the page cache list we'd create a lock inversion
300 * with the removepage path which gets the page lock then the
302 clear_page_dirty(page);
306 /* We have two reasons for giving llite the opportunity to change the
307 * write length of a given queued page as it builds the RPC containing
310 * 1) Further extending writes may have landed in the page cache
311 * since a partial write first queued this page requiring us
312 * to write more from the page cache. (No further races are possible, since
313 * by the time this is called, the page is locked.)
314 * 2) We might have raced with truncate and want to avoid performing
315 * write RPCs that are just going to be thrown away by the
316 * truncate's punch on the storage targets.
318 * The kms serves these purposes as it is set at both truncate and extending
321 static int ll_ap_refresh_count(void *data, int cmd)
323 struct ll_inode_info *lli;
324 struct ll_async_page *llap;
325 struct lov_stripe_md *lsm;
330 /* readpage queues with _COUNT_STABLE, shouldn't get here. */
331 LASSERT(cmd != OBD_BRW_READ);
333 llap = LLAP_FROM_COOKIE(data);
334 page = llap->llap_page;
335 lli = ll_i2info(page->mapping->host);
339 * this callback is called with client lock taken, thus, it should not
340 * sleep or deadlock is possible. --umka
342 // down(&lli->lli_size_sem);
343 kms = lov_merge_size(lsm, 1);
344 // up(&lli->lli_size_sem);
346 /* catch race with truncate */
347 if (((__u64)page->index << PAGE_SHIFT) >= kms)
350 /* catch sub-page write at end of file */
351 if (((__u64)page->index << PAGE_SHIFT) + PAGE_SIZE > kms)
352 return kms % PAGE_SIZE;
357 void ll_inode_fill_obdo(struct inode *inode, int cmd, struct obdo *oa)
359 struct lov_stripe_md *lsm;
360 obd_valid valid_flags;
362 lsm = ll_i2info(inode)->lli_smd;
364 oa->o_id = lsm->lsm_object_id;
365 oa->o_gr = lsm->lsm_object_gr;
366 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
367 valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME;
368 if (cmd == OBD_BRW_WRITE) {
369 oa->o_valid |= OBD_MD_FLIFID | OBD_MD_FLEPOCH;
370 mdc_pack_id(obdo_id(oa), inode->i_ino, 0, inode->i_mode,
371 id_group(&ll_i2info(inode)->lli_id),
372 id_fid(&ll_i2info(inode)->lli_id));
374 oa->o_easize = ll_i2info(inode)->lli_io_epoch;
375 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME;
378 obdo_from_inode(oa, inode, valid_flags);
381 static void ll_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
383 struct ll_async_page *llap;
386 llap = LLAP_FROM_COOKIE(data);
387 ll_inode_fill_obdo(llap->llap_page->mapping->host, cmd, oa);
391 static struct obd_async_page_ops ll_async_page_ops = {
392 .ap_make_ready = ll_ap_make_ready,
393 .ap_refresh_count = ll_ap_refresh_count,
394 .ap_fill_obdo = ll_ap_fill_obdo,
395 .ap_completion = ll_ap_completion,
399 struct ll_async_page *llap_cast_private(struct page *page)
401 struct ll_async_page *llap = (struct ll_async_page *)page->private;
403 LASSERTF(llap == NULL || llap->llap_magic == LLAP_MAGIC,
404 "page %p private %lu gave magic %d which != %d\n",
405 page, page->private, llap->llap_magic, LLAP_MAGIC);
410 /* XXX have the exp be an argument? */
411 struct ll_async_page *llap_from_page(struct page *page, unsigned origin)
413 struct ll_async_page *llap;
414 struct obd_export *exp;
415 struct inode *inode = page->mapping->host;
416 struct ll_sb_info *sbi = ll_i2sbi(inode);
420 LASSERTF(origin < LLAP__ORIGIN_MAX, "%u\n", origin);
422 llap = llap_cast_private(page);
426 exp = ll_i2dtexp(page->mapping->host);
428 RETURN(ERR_PTR(-EINVAL));
430 OBD_ALLOC(llap, sizeof(*llap));
432 RETURN(ERR_PTR(-ENOMEM));
433 llap->llap_magic = LLAP_MAGIC;
434 rc = obd_prep_async_page(exp, ll_i2info(inode)->lli_smd, NULL, page,
435 (obd_off)page->index << PAGE_SHIFT,
436 &ll_async_page_ops, llap, &llap->llap_cookie);
438 OBD_FREE(llap, sizeof(*llap));
442 CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n", llap,
443 page, llap->llap_cookie, (obd_off)page->index << PAGE_SHIFT);
444 /* also zeroing the PRIVBITS low order bitflags */
445 __set_page_ll_data(page, llap);
446 llap->llap_page = page;
448 spin_lock(&sbi->ll_lock);
449 sbi->ll_pglist_gen++;
450 list_add_tail(&llap->llap_proc_item, &sbi->ll_pglist);
451 spin_unlock(&sbi->ll_lock);
454 llap->llap_origin = origin;
458 static int queue_or_sync_write(struct obd_export *exp,
459 struct lov_stripe_md *lsm,
460 struct ll_async_page *llap,
462 obd_flags async_flags)
464 struct obd_io_group *oig;
468 /* _make_ready only sees llap once we've unlocked the page */
469 llap->llap_write_queued = 1;
470 rc = obd_queue_async_io(exp, lsm, NULL, llap->llap_cookie,
471 OBD_BRW_WRITE, 0, 0, 0, async_flags);
473 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "write queued\n");
474 //llap_write_pending(inode, llap);
478 llap->llap_write_queued = 0;
484 rc = obd_queue_group_io(exp, lsm, NULL, oig, llap->llap_cookie,
485 OBD_BRW_WRITE, 0, to, 0, ASYNC_READY |
486 ASYNC_URGENT | ASYNC_COUNT_STABLE |
491 rc = obd_trigger_group_io(exp, lsm, NULL, oig);
497 if (!rc && async_flags & ASYNC_READY)
498 unlock_page(llap->llap_page);
500 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
501 "sync write returned %d\n", rc);
510 /* be careful not to return success without setting the page Uptodate or
511 * the next pass through prepare_write will read in stale data from disk. */
512 int ll_commit_write(struct file *file, struct page *page, unsigned from,
515 struct inode *inode = page->mapping->host;
516 struct ll_inode_info *lli = ll_i2info(inode);
517 struct lov_stripe_md *lsm = lli->lli_smd;
518 struct obd_export *exp = NULL;
519 struct ll_async_page *llap;
524 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
525 LASSERT(inode == file->f_dentry->d_inode);
526 LASSERT(PageLocked(page));
528 CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
529 inode, page, from, to, page->index);
531 llap = llap_from_page(page, LLAP_ORIGIN_COMMIT_WRITE);
533 RETURN(PTR_ERR(llap));
535 exp = ll_i2dtexp(inode);
539 /* queue a write for some time in the future the first time we
541 if (!PageDirty(page)) {
542 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
543 LPROC_LL_DIRTY_MISSES);
545 rc = queue_or_sync_write(exp, ll_i2info(inode)->lli_smd,
550 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
551 LPROC_LL_DIRTY_HITS);
554 /* put the page in the page cache, from now on ll_removepage is
555 * responsible for cleaning up the llap.
556 * don't dirty the page if it has been write out in q_o_s_w */
557 if (llap->llap_write_queued)
558 set_page_dirty(page);
561 size = (((obd_off)page->index) << PAGE_SHIFT) + to;
562 down(&lli->lli_size_sem);
564 obd_adjust_kms(exp, lsm, size, 0);
565 if (size > inode->i_size)
566 inode->i_size = size;
567 SetPageUptodate(page);
568 } else if (size > inode->i_size) {
569 /* this page beyond the pales of i_size, so it can't be
570 * truncated in ll_p_r_e during lock revoking. we must
571 * teardown our book-keeping here. */
574 up(&lli->lli_size_sem);
578 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, unsigned long len)
580 struct ll_ra_info *ra = &sbi->ll_ra_info;
584 spin_lock(&sbi->ll_lock);
585 ret = min(ra->ra_max_pages - ra->ra_cur_pages, len);
586 ra->ra_cur_pages += ret;
587 spin_unlock(&sbi->ll_lock);
592 static void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
594 struct ll_ra_info *ra = &sbi->ll_ra_info;
595 spin_lock(&sbi->ll_lock);
596 LASSERTF(ra->ra_cur_pages >= len, "r_c_p %lu len %lu\n",
597 ra->ra_cur_pages, len);
598 ra->ra_cur_pages -= len;
599 spin_unlock(&sbi->ll_lock);
602 int ll_writepage(struct page *page)
604 struct inode *inode = page->mapping->host;
605 struct obd_export *exp;
606 struct ll_async_page *llap;
610 LASSERT(!PageDirty(page));
611 LASSERT(PageLocked(page));
613 exp = ll_i2dtexp(inode);
615 GOTO(out, rc = -EINVAL);
617 llap = llap_from_page(page, LLAP_ORIGIN_WRITEPAGE);
619 GOTO(out, rc = PTR_ERR(llap));
621 page_cache_get(page);
622 if (llap->llap_write_queued) {
623 LL_CDEBUG_PAGE(D_PAGE, page, "marking urgent\n");
624 rc = obd_set_async_flags(exp, ll_i2info(inode)->lli_smd, NULL,
626 ASYNC_READY | ASYNC_URGENT);
628 rc = queue_or_sync_write(exp, ll_i2info(inode)->lli_smd, llap,
629 PAGE_SIZE, ASYNC_READY |
633 page_cache_release(page);
641 /* called for each page in a completed rpc.*/
642 void ll_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
644 struct ll_async_page *llap;
648 llap = LLAP_FROM_COOKIE(data);
649 page = llap->llap_page;
650 LASSERT(PageLocked(page));
652 LL_CDEBUG_PAGE(D_PAGE, page, "completing cmd %d with %d\n", cmd, rc);
654 if (cmd == OBD_BRW_READ && llap->llap_defer_uptodate)
655 ll_ra_count_put(ll_i2sbi(page->mapping->host), 1);
658 if (cmd == OBD_BRW_READ) {
659 if (!llap->llap_defer_uptodate)
660 SetPageUptodate(page);
662 llap->llap_write_queued = 0;
664 ClearPageError(page);
666 if (cmd == OBD_BRW_READ)
667 llap->llap_defer_uptodate = 0;
673 if (0 && cmd == OBD_BRW_WRITE) {
674 llap_write_complete(page->mapping->host, llap);
675 ll_try_done_writing(page->mapping->host);
678 if (PageWriteback(page)) {
679 end_page_writeback(page);
681 page_cache_release(page);
685 /* the kernel calls us here when a page is unhashed from the page cache.
686 * the page will be locked and the kernel is holding a spinlock, so
687 * we need to be careful. we're just tearing down our book-keeping
689 void ll_removepage(struct page *page)
691 struct inode *inode = page->mapping->host;
692 struct obd_export *exp;
693 struct ll_async_page *llap;
694 struct ll_sb_info *sbi = ll_i2sbi(inode);
698 LASSERT(!in_interrupt());
700 /* sync pages or failed read pages can leave pages in the page
701 * cache that don't have our data associated with them anymore */
702 if (page->private == 0) {
707 LL_CDEBUG_PAGE(D_PAGE, page, "being evicted\n");
709 exp = ll_i2dtexp(inode);
711 CERROR("page %p ind %lu gave null export\n", page, page->index);
716 llap = llap_from_page(page, 0);
718 CERROR("page %p ind %lu couldn't find llap: %ld\n", page,
719 page->index, PTR_ERR(llap));
724 //llap_write_complete(inode, llap);
725 rc = obd_teardown_async_page(exp, ll_i2info(inode)->lli_smd, NULL,
728 CERROR("page %p ind %lu failed: %d\n", page, page->index, rc);
730 /* this unconditional free is only safe because the page lock
731 * is providing exclusivity to memory pressure/truncate/writeback..*/
732 __clear_page_ll_data(page);
734 spin_lock(&sbi->ll_lock);
735 if (!list_empty(&llap->llap_proc_item))
736 list_del_init(&llap->llap_proc_item);
737 sbi->ll_pglist_gen++;
738 spin_unlock(&sbi->ll_lock);
739 OBD_FREE(llap, sizeof(*llap));
743 static int ll_page_matches(struct page *page, int fd_flags, int readahead)
745 struct lustre_handle match_lockh = {0};
746 struct inode *inode = page->mapping->host;
747 ldlm_policy_data_t page_extent;
751 if (fd_flags & LL_FILE_GROUP_LOCKED)
754 page_extent.l_extent.start = (__u64)page->index << PAGE_CACHE_SHIFT;
755 page_extent.l_extent.end =
756 page_extent.l_extent.start + PAGE_CACHE_SIZE - 1;
757 flags = LDLM_FL_TEST_LOCK;
759 flags |= LDLM_FL_CBPENDING | LDLM_FL_BLOCK_GRANTED;
760 matches = obd_match(ll_i2sbi(inode)->ll_dt_exp,
761 ll_i2info(inode)->lli_smd, LDLM_EXTENT,
762 &page_extent, LCK_PR | LCK_PW, &flags, inode,
767 static int ll_issue_page_read(struct obd_export *exp,
768 struct ll_async_page *llap,
769 struct obd_io_group *oig, int defer)
771 struct page *page = llap->llap_page;
774 page_cache_get(page);
775 llap->llap_defer_uptodate = defer;
776 llap->llap_ra_used = 0;
777 rc = obd_queue_group_io(exp, ll_i2info(page->mapping->host)->lli_smd,
778 NULL, oig, llap->llap_cookie, OBD_BRW_READ, 0,
779 PAGE_SIZE, 0, ASYNC_COUNT_STABLE | ASYNC_READY
782 LL_CDEBUG_PAGE(D_ERROR, page, "read queue failed: rc %d\n", rc);
783 page_cache_release(page);
788 static void ll_ra_stats_inc_unlocked(struct ll_ra_info *ra, enum ra_stat which)
790 LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
791 ra->ra_stats[which]++;
794 static void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which)
796 struct ll_sb_info *sbi = ll_i2sbi(mapping->host);
797 struct ll_ra_info *ra = &ll_i2sbi(mapping->host)->ll_ra_info;
799 spin_lock(&sbi->ll_lock);
800 ll_ra_stats_inc_unlocked(ra, which);
801 spin_unlock(&sbi->ll_lock);
804 void ll_ra_accounting(struct page *page, struct address_space *mapping)
806 struct ll_async_page *llap;
808 llap = llap_from_page(page, LLAP_ORIGIN_WRITEPAGE);
812 if (!llap->llap_defer_uptodate || llap->llap_ra_used)
815 ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
818 #define RAS_CDEBUG(ras) \
819 CDEBUG(D_READA, "lrp %lu c %lu ws %lu wl %lu nra %lu\n", \
820 ras->ras_last_readpage, ras->ras_consecutive, \
821 ras->ras_window_start, ras->ras_window_len, \
822 ras->ras_next_readahead);
824 static int index_in_window(unsigned long index, unsigned long point,
825 unsigned long before, unsigned long after)
827 unsigned long start = point - before, end = point + after;
834 return start <= index && index <= end;
837 static int ll_readahead(struct ll_readahead_state *ras,
838 struct obd_export *exp, struct address_space *mapping,
839 struct obd_io_group *oig, int flags)
841 unsigned long i, start = 0, end = 0, reserved;
842 struct ll_async_page *llap;
844 int rc, ret = 0, match_failed = 0;
848 kms = lov_merge_size(ll_i2info(mapping->host)->lli_smd, 1);
850 ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN);
853 spin_lock(&ras->ras_lock);
855 /* reserve a part of the read-ahead window that we'll be issuing */
856 if (ras->ras_window_len) {
857 start = ras->ras_next_readahead;
858 end = ras->ras_window_start + ras->ras_window_len - 1;
859 end = min(end, (unsigned long)(kms >> PAGE_CACHE_SHIFT));
860 ras->ras_next_readahead = max(end, end + 1);
865 spin_unlock(&ras->ras_lock);
868 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
872 reserved = ll_ra_count_get(ll_i2sbi(mapping->host), end - start + 1);
873 if (reserved < end - start + 1)
874 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
876 for (i = start; reserved > 0 && !match_failed && i <= end; i++) {
877 /* skip locked pages from previous readpage calls */
878 page = grab_cache_page_nowait(mapping, i);
880 CDEBUG(D_READA, "g_c_p_n failed\n");
884 /* we do this first so that we can see the page in the /proc
886 llap = llap_from_page(page, LLAP_ORIGIN_READAHEAD);
887 if (IS_ERR(llap) || llap->llap_defer_uptodate)
890 /* skip completed pages */
891 if (Page_Uptodate(page))
894 /* bail when we hit the end of the lock. */
895 if ((rc = ll_page_matches(page, flags, 1)) <= 0) {
896 LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
897 "lock match failed: rc %d\n", rc);
898 ll_ra_stats_inc(mapping, RA_STAT_FAILED_MATCH);
903 rc = ll_issue_page_read(exp, llap, oig, 1);
907 LL_CDEBUG_PAGE(D_READA| D_PAGE, page,
908 "started read-ahead\n");
912 LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
913 "skipping read-ahead\n");
917 page_cache_release(page);
920 LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
922 ll_ra_count_put(ll_i2sbi(mapping->host), reserved);
924 if (i == end + 1 && end == (kms >> PAGE_CACHE_SHIFT))
925 ll_ra_stats_inc(mapping, RA_STAT_EOF);
927 /* if we didn't get to the end of the region we reserved from
928 * the ras we need to go back and update the ras so that the
929 * next read-ahead tries from where we left off. we only do so
930 * if the region we failed to issue read-ahead on is still ahead
931 * of the app and behind the next index to start read-ahead from */
933 spin_lock(&ras->ras_lock);
934 if (i < ras->ras_next_readahead &&
935 index_in_window(i, ras->ras_window_start, 0,
936 ras->ras_window_len)) {
937 ras->ras_next_readahead = i;
940 spin_unlock(&ras->ras_lock);
946 static void ras_set_start(struct ll_readahead_state *ras, unsigned long index)
948 ras->ras_window_start = index & (~(PTLRPC_MAX_BRW_PAGES - 1));
951 /* called with the ras_lock held or from places where it doesn't matter */
952 static void ras_reset(struct ll_readahead_state *ras, unsigned long index)
954 ras->ras_last_readpage = index;
955 ras->ras_consecutive = 1;
956 ras->ras_window_len = 0;
957 ras_set_start(ras, index);
958 ras->ras_next_readahead = ras->ras_window_start;
963 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
965 spin_lock_init(&ras->ras_lock);
969 static void ras_update(struct ll_sb_info *sbi, struct ll_readahead_state *ras,
970 unsigned long index, unsigned hit)
972 struct ll_ra_info *ra = &sbi->ll_ra_info;
976 spin_lock(&sbi->ll_lock);
977 spin_lock(&ras->ras_lock);
979 ll_ra_stats_inc_unlocked(ra, hit ? RA_STAT_HIT : RA_STAT_MISS);
981 /* reset the read-ahead window in two cases. First when the app seeks
982 * or reads to some other part of the file. Secondly if we get a
983 * read-ahead miss that we think we've previously issued. This can
984 * be a symptom of there being so many read-ahead pages that the VM is
985 * reclaiming it before we get to it. */
986 if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
988 ll_ra_stats_inc_unlocked(ra, RA_STAT_DISTANT_READPAGE);
989 } else if (!hit && ras->ras_window_len &&
990 index < ras->ras_next_readahead &&
991 index_in_window(index, ras->ras_window_start, 0,
992 ras->ras_window_len)) {
994 ll_ra_stats_inc_unlocked(ra, RA_STAT_MISS_IN_WINDOW);
998 ras_reset(ras, index);
1002 ras->ras_last_readpage = index;
1003 ras->ras_consecutive++;
1004 ras_set_start(ras, index);
1005 ras->ras_next_readahead = max(ras->ras_window_start,
1006 ras->ras_next_readahead);
1008 /* wait for a few pages to arrive before issuing readahead to avoid
1009 * the worst overutilization */
1010 if (ras->ras_consecutive == 3) {
1011 ras->ras_window_len = PTLRPC_MAX_BRW_PAGES;
1012 GOTO(out_unlock, 0);
1015 /* we need to increase the window sometimes. we'll arbitrarily
1016 * do it half-way through the pages in an rpc */
1017 if ((index & (PTLRPC_MAX_BRW_PAGES - 1)) ==
1018 (PTLRPC_MAX_BRW_PAGES >> 1)) {
1019 ras->ras_window_len += PTLRPC_MAX_BRW_PAGES;
1020 ras->ras_window_len = min(ras->ras_window_len,
1028 spin_unlock(&ras->ras_lock);
1029 spin_unlock(&sbi->ll_lock);
1033 * for now we do our readpage the same on both 2.4 and 2.5. The kernel's
1034 * read-ahead assumes it is valid to issue readpage all the way up to
1035 * i_size, but our dlm locks make that not the case. We disable the
1036 * kernel's read-ahead and do our own by walking ahead in the page cache
1037 * checking for dlm lock coverage. the main difference between 2.4 and
1038 * 2.6 is how read-ahead gets batched and issued, but we're using our own,
1039 * so they look the same.
1041 int ll_readpage(struct file *filp, struct page *page)
1043 struct ll_file_data *fd = filp->private_data;
1044 struct inode *inode = page->mapping->host;
1045 struct obd_export *exp;
1046 struct ll_async_page *llap;
1047 struct obd_io_group *oig = NULL;
1051 LASSERT(PageLocked(page));
1052 LASSERT(!PageUptodate(page));
1053 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset="LPX64"\n",
1054 inode->i_ino, inode->i_generation, inode,
1055 (((obd_off)page->index) << PAGE_SHIFT));
1056 LASSERT(atomic_read(&filp->f_dentry->d_inode->i_count) > 0);
1058 rc = oig_init(&oig);
1062 exp = ll_i2dtexp(inode);
1064 GOTO(out, rc = -EINVAL);
1066 llap = llap_from_page(page, LLAP_ORIGIN_READPAGE);
1068 GOTO(out, rc = PTR_ERR(llap));
1070 if (ll_i2sbi(inode)->ll_flags & LL_SBI_READAHEAD)
1071 ras_update(ll_i2sbi(inode), &fd->fd_ras, page->index,
1072 llap->llap_defer_uptodate);
1074 if (llap->llap_defer_uptodate) {
1075 llap->llap_ra_used = 1;
1076 rc = ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1079 obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd,
1081 LL_CDEBUG_PAGE(D_PAGE, page, "marking uptodate from defer\n");
1082 SetPageUptodate(page);
1084 GOTO(out_oig, rc = 0);
1087 rc = ll_page_matches(page, fd->fd_flags, 0);
1089 LL_CDEBUG_PAGE(D_ERROR, page, "lock match failed: rc %d\n", rc);
1094 CWARN("ino %lu page %lu (%llu) not covered by "
1095 "a lock (mmap?). check debug logs.\n",
1096 inode->i_ino, page->index,
1097 (long long)page->index << PAGE_CACHE_SHIFT);
1100 rc = ll_issue_page_read(exp, llap, oig, 0);
1104 LL_CDEBUG_PAGE(D_PAGE, page, "queued readpage\n");
1105 if (ll_i2sbi(inode)->ll_flags & LL_SBI_READAHEAD)
1106 ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1109 rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);