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(LLI_DIRTY_HANDLE(inode));
200 LASSERT(PageLocked(page));
201 (void)llap_cast_private(page); /* assertion */
203 /* Check to see if we should return -EIO right away */
205 pga.disk_offset = pga.page_offset = offset;
206 pga.count = PAGE_SIZE;
213 oa->o_id = lsm->lsm_object_id;
214 oa->o_gr = lsm->lsm_object_gr;
215 oa->o_mode = inode->i_mode;
217 oa->o_valid = OBD_MD_FLID | OBD_MD_FLMODE |
218 OBD_MD_FLTYPE | OBD_MD_FLGROUP;
221 * needed for quota to create OSS object on write with correct
224 oa->o_uid = inode->i_uid;
225 oa->o_valid |= OBD_MD_FLUID;
227 oa->o_gid = inode->i_gid;
228 oa->o_valid |= OBD_MD_FLGID;
230 /* putting there also fid, needed for quota too. */
231 memcpy(obdo_id(oa), &lli->lli_id, sizeof(lli->lli_id));
232 oa->o_valid |= OBD_MD_FLINLINE;
234 rc = obd_brw(OBD_BRW_CHECK, ll_i2dtexp(inode),
235 oa, lsm, 1, &pga, NULL);
237 GOTO(out_free_oa, rc);
239 if (PageUptodate(page))
240 GOTO(out_free_oa, 0);
242 /* We're completely overwriting an existing page, so _don't_ set it up
243 * to date until commit_write */
244 if (from == 0 && to == PAGE_SIZE) {
245 POISON_PAGE(page, 0x11);
246 GOTO(out_free_oa, 0);
249 /* If are writing to a new page, no need to read old data. The extent
250 * locking will have updated the KMS, and for our purposes here we can
251 * treat it like i_size. */
252 down(&lli->lli_size_sem);
253 kms = lov_merge_size(lsm, 1);
254 up(&lli->lli_size_sem);
256 memset(kmap(page), 0, PAGE_SIZE);
258 GOTO(prepare_done, rc = 0);
261 /* XXX could be an async ocp read.. read-ahead? */
262 rc = ll_brw(OBD_BRW_READ, inode, oa, page, 0);
264 /* bug 1598: don't clobber blksize */
265 oa->o_valid &= ~(OBD_MD_FLSIZE | OBD_MD_FLBLKSZ);
266 obdo_refresh_inode(inode, oa, oa->o_valid);
272 SetPageUptodate(page);
278 static int ll_ap_make_ready(void *data, int cmd)
280 struct ll_async_page *llap;
284 llap = LLAP_FROM_COOKIE(data);
285 page = llap->llap_page;
287 LASSERT(cmd != OBD_BRW_READ);
289 /* we're trying to write, but the page is locked.. come back later */
290 if (TryLockPage(page))
293 LL_CDEBUG_PAGE(D_PAGE, page, "made ready\n");
294 page_cache_get(page);
296 /* if we left PageDirty we might get another writepage call
297 * in the future. list walkers are bright enough
298 * to check page dirty so we can leave it on whatever list
299 * its on. XXX also, we're called with the cli list so if
300 * we got the page cache list we'd create a lock inversion
301 * with the removepage path which gets the page lock then the
303 clear_page_dirty(page);
307 /* We have two reasons for giving llite the opportunity to change the
308 * write length of a given queued page as it builds the RPC containing
311 * 1) Further extending writes may have landed in the page cache
312 * since a partial write first queued this page requiring us
313 * to write more from the page cache. (No further races are possible, since
314 * by the time this is called, the page is locked.)
315 * 2) We might have raced with truncate and want to avoid performing
316 * write RPCs that are just going to be thrown away by the
317 * truncate's punch on the storage targets.
319 * The kms serves these purposes as it is set at both truncate and extending
322 static int ll_ap_refresh_count(void *data, int cmd)
324 struct ll_inode_info *lli;
325 struct ll_async_page *llap;
326 struct lov_stripe_md *lsm;
331 /* readpage queues with _COUNT_STABLE, shouldn't get here. */
332 LASSERT(cmd != OBD_BRW_READ);
334 llap = LLAP_FROM_COOKIE(data);
335 page = llap->llap_page;
336 lli = ll_i2info(page->mapping->host);
340 * this callback is called with client lock taken, thus, it should not
341 * sleep or deadlock is possible. --umka
343 // down(&lli->lli_size_sem);
344 kms = lov_merge_size(lsm, 1);
345 // up(&lli->lli_size_sem);
347 /* catch race with truncate */
348 if (((__u64)page->index << PAGE_SHIFT) >= kms)
351 /* catch sub-page write at end of file */
352 if (((__u64)page->index << PAGE_SHIFT) + PAGE_SIZE > kms)
353 return kms % PAGE_SIZE;
358 void ll_inode_fill_obdo(struct inode *inode, int cmd, struct obdo *oa)
360 struct lov_stripe_md *lsm;
361 obd_valid valid_flags;
363 lsm = ll_i2info(inode)->lli_smd;
365 oa->o_id = lsm->lsm_object_id;
366 oa->o_gr = lsm->lsm_object_gr;
367 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
368 valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME;
369 if (cmd == OBD_BRW_WRITE) {
370 oa->o_valid |= OBD_MD_FLIFID | OBD_MD_FLEPOCH;
371 mdc_pack_id(obdo_id(oa), inode->i_ino, 0, inode->i_mode,
372 id_group(&ll_i2info(inode)->lli_id),
373 id_fid(&ll_i2info(inode)->lli_id));
375 oa->o_easize = ll_i2info(inode)->lli_io_epoch;
376 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME;
379 obdo_from_inode(oa, inode, valid_flags);
382 static void ll_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
384 struct ll_async_page *llap;
387 llap = LLAP_FROM_COOKIE(data);
388 ll_inode_fill_obdo(llap->llap_page->mapping->host, cmd, oa);
392 static struct obd_async_page_ops ll_async_page_ops = {
393 .ap_make_ready = ll_ap_make_ready,
394 .ap_refresh_count = ll_ap_refresh_count,
395 .ap_fill_obdo = ll_ap_fill_obdo,
396 .ap_completion = ll_ap_completion,
400 struct ll_async_page *llap_cast_private(struct page *page)
402 struct ll_async_page *llap = (struct ll_async_page *)page->private;
404 LASSERTF(llap == NULL || llap->llap_magic == LLAP_MAGIC,
405 "page %p private %lu gave magic %d which != %d\n",
406 page, page->private, llap->llap_magic, LLAP_MAGIC);
411 /* XXX have the exp be an argument? */
412 struct ll_async_page *llap_from_page(struct page *page, unsigned origin)
414 struct ll_async_page *llap;
415 struct obd_export *exp;
416 struct inode *inode = page->mapping->host;
417 struct ll_sb_info *sbi = ll_i2sbi(inode);
421 LASSERTF(origin < LLAP__ORIGIN_MAX, "%u\n", origin);
423 llap = llap_cast_private(page);
427 exp = ll_i2dtexp(page->mapping->host);
429 RETURN(ERR_PTR(-EINVAL));
431 OBD_ALLOC(llap, sizeof(*llap));
433 RETURN(ERR_PTR(-ENOMEM));
434 llap->llap_magic = LLAP_MAGIC;
435 INIT_LIST_HEAD(&llap->llap_pending_write);
436 rc = obd_prep_async_page(exp, ll_i2info(inode)->lli_smd, NULL, page,
437 (obd_off)page->index << PAGE_SHIFT,
438 &ll_async_page_ops, llap, &llap->llap_cookie);
440 OBD_FREE(llap, sizeof(*llap));
444 CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n", llap,
445 page, llap->llap_cookie, (obd_off)page->index << PAGE_SHIFT);
446 /* also zeroing the PRIVBITS low order bitflags */
447 __set_page_ll_data(page, llap);
448 llap->llap_page = page;
450 spin_lock(&sbi->ll_lock);
451 sbi->ll_pglist_gen++;
452 list_add_tail(&llap->llap_proc_item, &sbi->ll_pglist);
453 spin_unlock(&sbi->ll_lock);
456 llap->llap_origin = origin;
460 static int queue_or_sync_write(struct obd_export *exp,
461 struct lov_stripe_md *lsm,
462 struct ll_async_page *llap,
464 obd_flags async_flags)
466 struct obd_io_group *oig;
470 /* _make_ready only sees llap once we've unlocked the page */
471 llap->llap_write_queued = 1;
472 rc = obd_queue_async_io(exp, lsm, NULL, llap->llap_cookie,
473 OBD_BRW_WRITE, 0, 0, 0, async_flags);
475 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "write queued\n");
476 llap_write_pending(llap->llap_page->mapping->host, llap);
480 llap->llap_write_queued = 0;
486 rc = obd_queue_group_io(exp, lsm, NULL, oig, llap->llap_cookie,
487 OBD_BRW_WRITE, 0, to, 0, ASYNC_READY |
488 ASYNC_URGENT | ASYNC_COUNT_STABLE |
493 rc = obd_trigger_group_io(exp, lsm, NULL, oig);
499 if (!rc && async_flags & ASYNC_READY)
500 unlock_page(llap->llap_page);
502 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
503 "sync write returned %d\n", rc);
512 /* be careful not to return success without setting the page Uptodate or
513 * the next pass through prepare_write will read in stale data from disk. */
514 int ll_commit_write(struct file *file, struct page *page, unsigned from,
517 struct inode *inode = page->mapping->host;
518 struct ll_inode_info *lli = ll_i2info(inode);
519 struct lov_stripe_md *lsm = lli->lli_smd;
520 struct obd_export *exp = NULL;
521 struct ll_async_page *llap;
526 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
527 LASSERT(inode == file->f_dentry->d_inode);
528 LASSERT(PageLocked(page));
529 LASSERT(LLI_DIRTY_HANDLE(inode));
531 CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
532 inode, page, from, to, page->index);
534 llap = llap_from_page(page, LLAP_ORIGIN_COMMIT_WRITE);
536 RETURN(PTR_ERR(llap));
538 exp = ll_i2dtexp(inode);
542 /* queue a write for some time in the future the first time we
544 if (!PageDirty(page)) {
545 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
546 LPROC_LL_DIRTY_MISSES);
548 rc = queue_or_sync_write(exp, ll_i2info(inode)->lli_smd,
553 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
554 LPROC_LL_DIRTY_HITS);
557 /* put the page in the page cache, from now on ll_removepage is
558 * responsible for cleaning up the llap.
559 * don't dirty the page if it has been write out in q_o_s_w */
560 if (llap->llap_write_queued)
561 set_page_dirty(page);
564 size = (((obd_off)page->index) << PAGE_SHIFT) + to;
565 down(&lli->lli_size_sem);
567 obd_adjust_kms(exp, lsm, size, 0);
568 if (size > inode->i_size)
569 inode->i_size = size;
570 SetPageUptodate(page);
571 } else if (size > inode->i_size) {
572 /* this page beyond the pales of i_size, so it can't be
573 * truncated in ll_p_r_e during lock revoking. we must
574 * teardown our book-keeping here. */
577 up(&lli->lli_size_sem);
581 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, unsigned long len)
583 struct ll_ra_info *ra = &sbi->ll_ra_info;
587 spin_lock(&sbi->ll_lock);
588 ret = min(ra->ra_max_pages - ra->ra_cur_pages, len);
589 ra->ra_cur_pages += ret;
590 spin_unlock(&sbi->ll_lock);
595 static void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
597 struct ll_ra_info *ra = &sbi->ll_ra_info;
598 spin_lock(&sbi->ll_lock);
599 LASSERTF(ra->ra_cur_pages >= len, "r_c_p %lu len %lu\n",
600 ra->ra_cur_pages, len);
601 ra->ra_cur_pages -= len;
602 spin_unlock(&sbi->ll_lock);
605 int ll_writepage(struct page *page)
607 struct inode *inode = page->mapping->host;
608 struct obd_export *exp;
609 struct ll_async_page *llap;
613 LASSERT(!PageDirty(page));
614 LASSERT(PageLocked(page));
615 LASSERT(LLI_DIRTY_HANDLE(inode));
617 exp = ll_i2dtexp(inode);
619 GOTO(out, rc = -EINVAL);
621 llap = llap_from_page(page, LLAP_ORIGIN_WRITEPAGE);
623 GOTO(out, rc = PTR_ERR(llap));
625 page_cache_get(page);
626 if (llap->llap_write_queued) {
627 LL_CDEBUG_PAGE(D_PAGE, page, "marking urgent\n");
628 rc = obd_set_async_flags(exp, ll_i2info(inode)->lli_smd, NULL,
630 ASYNC_READY | ASYNC_URGENT);
632 rc = queue_or_sync_write(exp, ll_i2info(inode)->lli_smd, llap,
633 PAGE_SIZE, ASYNC_READY |
637 page_cache_release(page);
645 /* called for each page in a completed rpc.*/
646 void ll_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
648 struct ll_async_page *llap;
652 llap = LLAP_FROM_COOKIE(data);
653 page = llap->llap_page;
654 LASSERT(PageLocked(page));
656 LL_CDEBUG_PAGE(D_PAGE, page, "completing cmd %d with %d\n", cmd, rc);
658 if (cmd == OBD_BRW_READ && llap->llap_defer_uptodate)
659 ll_ra_count_put(ll_i2sbi(page->mapping->host), 1);
662 if (cmd == OBD_BRW_READ) {
663 if (!llap->llap_defer_uptodate)
664 SetPageUptodate(page);
666 llap->llap_write_queued = 0;
668 ClearPageError(page);
670 if (cmd == OBD_BRW_READ)
671 llap->llap_defer_uptodate = 0;
677 if (cmd == OBD_BRW_WRITE) {
678 llap_write_complete(page->mapping->host, llap);
679 ll_try_done_writing(page->mapping->host);
682 if (PageWriteback(page)) {
683 end_page_writeback(page);
685 page_cache_release(page);
689 /* the kernel calls us here when a page is unhashed from the page cache.
690 * the page will be locked and the kernel is holding a spinlock, so
691 * we need to be careful. we're just tearing down our book-keeping
693 void ll_removepage(struct page *page)
695 struct inode *inode = page->mapping->host;
696 struct obd_export *exp;
697 struct ll_async_page *llap;
698 struct ll_sb_info *sbi = ll_i2sbi(inode);
702 LASSERT(!in_interrupt());
704 /* sync pages or failed read pages can leave pages in the page
705 * cache that don't have our data associated with them anymore */
706 if (page->private == 0) {
711 LL_CDEBUG_PAGE(D_PAGE, page, "being evicted\n");
713 exp = ll_i2dtexp(inode);
715 CERROR("page %p ind %lu gave null export\n", page, page->index);
720 llap = llap_from_page(page, 0);
722 CERROR("page %p ind %lu couldn't find llap: %ld\n", page,
723 page->index, PTR_ERR(llap));
728 llap_write_complete(inode, llap);
729 rc = obd_teardown_async_page(exp, ll_i2info(inode)->lli_smd, NULL,
732 CERROR("page %p ind %lu failed: %d\n", page, page->index, rc);
734 /* this unconditional free is only safe because the page lock
735 * is providing exclusivity to memory pressure/truncate/writeback..*/
736 __clear_page_ll_data(page);
738 spin_lock(&sbi->ll_lock);
739 if (!list_empty(&llap->llap_proc_item))
740 list_del_init(&llap->llap_proc_item);
741 sbi->ll_pglist_gen++;
742 spin_unlock(&sbi->ll_lock);
743 OBD_FREE(llap, sizeof(*llap));
747 static int ll_page_matches(struct page *page, int fd_flags, int readahead)
749 struct lustre_handle match_lockh = {0};
750 struct inode *inode = page->mapping->host;
751 ldlm_policy_data_t page_extent;
755 if (fd_flags & LL_FILE_GROUP_LOCKED)
758 page_extent.l_extent.start = (__u64)page->index << PAGE_CACHE_SHIFT;
759 page_extent.l_extent.end =
760 page_extent.l_extent.start + PAGE_CACHE_SIZE - 1;
761 flags = LDLM_FL_TEST_LOCK;
763 flags |= LDLM_FL_CBPENDING | LDLM_FL_BLOCK_GRANTED;
764 matches = obd_match(ll_i2sbi(inode)->ll_dt_exp,
765 ll_i2info(inode)->lli_smd, LDLM_EXTENT,
766 &page_extent, LCK_PR | LCK_PW, &flags, inode,
771 static int ll_issue_page_read(struct obd_export *exp,
772 struct ll_async_page *llap,
773 struct obd_io_group *oig, int defer)
775 struct page *page = llap->llap_page;
778 page_cache_get(page);
779 llap->llap_defer_uptodate = defer;
780 llap->llap_ra_used = 0;
781 rc = obd_queue_group_io(exp, ll_i2info(page->mapping->host)->lli_smd,
782 NULL, oig, llap->llap_cookie, OBD_BRW_READ, 0,
783 PAGE_SIZE, 0, ASYNC_COUNT_STABLE | ASYNC_READY
786 LL_CDEBUG_PAGE(D_ERROR, page, "read queue failed: rc %d\n", rc);
787 page_cache_release(page);
792 static void ll_ra_stats_inc_unlocked(struct ll_ra_info *ra, enum ra_stat which)
794 LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
795 ra->ra_stats[which]++;
798 static void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which)
800 struct ll_sb_info *sbi = ll_i2sbi(mapping->host);
801 struct ll_ra_info *ra = &ll_i2sbi(mapping->host)->ll_ra_info;
803 spin_lock(&sbi->ll_lock);
804 ll_ra_stats_inc_unlocked(ra, which);
805 spin_unlock(&sbi->ll_lock);
808 void ll_ra_accounting(struct page *page, struct address_space *mapping)
810 struct ll_async_page *llap;
812 llap = llap_from_page(page, LLAP_ORIGIN_WRITEPAGE);
816 if (!llap->llap_defer_uptodate || llap->llap_ra_used)
819 ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
822 #define RAS_CDEBUG(ras) \
823 CDEBUG(D_READA, "lrp %lu c %lu ws %lu wl %lu nra %lu\n", \
824 ras->ras_last_readpage, ras->ras_consecutive, \
825 ras->ras_window_start, ras->ras_window_len, \
826 ras->ras_next_readahead);
828 static int index_in_window(unsigned long index, unsigned long point,
829 unsigned long before, unsigned long after)
831 unsigned long start = point - before, end = point + after;
838 return start <= index && index <= end;
841 static int ll_readahead(struct ll_readahead_state *ras,
842 struct obd_export *exp, struct address_space *mapping,
843 struct obd_io_group *oig, int flags)
845 unsigned long i, start = 0, end = 0, reserved;
846 struct ll_async_page *llap;
848 int rc, ret = 0, match_failed = 0;
852 kms = lov_merge_size(ll_i2info(mapping->host)->lli_smd, 1);
854 ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN);
857 spin_lock(&ras->ras_lock);
859 /* reserve a part of the read-ahead window that we'll be issuing */
860 if (ras->ras_window_len) {
861 start = ras->ras_next_readahead;
862 end = ras->ras_window_start + ras->ras_window_len - 1;
863 end = min(end, (unsigned long)(kms >> PAGE_CACHE_SHIFT));
864 ras->ras_next_readahead = max(end, end + 1);
869 spin_unlock(&ras->ras_lock);
872 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
876 reserved = ll_ra_count_get(ll_i2sbi(mapping->host), end - start + 1);
877 if (reserved < end - start + 1)
878 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
880 for (i = start; reserved > 0 && !match_failed && i <= end; i++) {
881 /* skip locked pages from previous readpage calls */
882 page = grab_cache_page_nowait(mapping, i);
884 CDEBUG(D_READA, "g_c_p_n failed\n");
888 /* we do this first so that we can see the page in the /proc
890 llap = llap_from_page(page, LLAP_ORIGIN_READAHEAD);
891 if (IS_ERR(llap) || llap->llap_defer_uptodate)
894 /* skip completed pages */
895 if (Page_Uptodate(page))
898 /* bail when we hit the end of the lock. */
899 if ((rc = ll_page_matches(page, flags, 1)) <= 0) {
900 LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
901 "lock match failed: rc %d\n", rc);
902 ll_ra_stats_inc(mapping, RA_STAT_FAILED_MATCH);
907 rc = ll_issue_page_read(exp, llap, oig, 1);
911 LL_CDEBUG_PAGE(D_READA| D_PAGE, page,
912 "started read-ahead\n");
916 LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
917 "skipping read-ahead\n");
921 page_cache_release(page);
924 LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
926 ll_ra_count_put(ll_i2sbi(mapping->host), reserved);
928 if (i == end + 1 && end == (kms >> PAGE_CACHE_SHIFT))
929 ll_ra_stats_inc(mapping, RA_STAT_EOF);
931 /* if we didn't get to the end of the region we reserved from
932 * the ras we need to go back and update the ras so that the
933 * next read-ahead tries from where we left off. we only do so
934 * if the region we failed to issue read-ahead on is still ahead
935 * of the app and behind the next index to start read-ahead from */
937 spin_lock(&ras->ras_lock);
938 if (i < ras->ras_next_readahead &&
939 index_in_window(i, ras->ras_window_start, 0,
940 ras->ras_window_len)) {
941 ras->ras_next_readahead = i;
944 spin_unlock(&ras->ras_lock);
950 static void ras_set_start(struct ll_readahead_state *ras, unsigned long index)
952 ras->ras_window_start = index & (~(PTLRPC_MAX_BRW_PAGES - 1));
955 /* called with the ras_lock held or from places where it doesn't matter */
956 static void ras_reset(struct ll_readahead_state *ras, unsigned long index)
958 ras->ras_last_readpage = index;
959 ras->ras_consecutive = 1;
960 ras->ras_window_len = 0;
961 ras_set_start(ras, index);
962 ras->ras_next_readahead = ras->ras_window_start;
967 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
969 spin_lock_init(&ras->ras_lock);
973 static void ras_update(struct ll_sb_info *sbi, struct ll_readahead_state *ras,
974 unsigned long index, unsigned hit)
976 struct ll_ra_info *ra = &sbi->ll_ra_info;
980 spin_lock(&sbi->ll_lock);
981 spin_lock(&ras->ras_lock);
983 ll_ra_stats_inc_unlocked(ra, hit ? RA_STAT_HIT : RA_STAT_MISS);
985 /* reset the read-ahead window in two cases. First when the app seeks
986 * or reads to some other part of the file. Secondly if we get a
987 * read-ahead miss that we think we've previously issued. This can
988 * be a symptom of there being so many read-ahead pages that the VM is
989 * reclaiming it before we get to it. */
990 if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
992 ll_ra_stats_inc_unlocked(ra, RA_STAT_DISTANT_READPAGE);
993 } else if (!hit && ras->ras_window_len &&
994 index < ras->ras_next_readahead &&
995 index_in_window(index, ras->ras_window_start, 0,
996 ras->ras_window_len)) {
998 ll_ra_stats_inc_unlocked(ra, RA_STAT_MISS_IN_WINDOW);
1002 ras_reset(ras, index);
1003 GOTO(out_unlock, 0);
1006 ras->ras_last_readpage = index;
1007 ras->ras_consecutive++;
1008 ras_set_start(ras, index);
1009 ras->ras_next_readahead = max(ras->ras_window_start,
1010 ras->ras_next_readahead);
1012 /* wait for a few pages to arrive before issuing readahead to avoid
1013 * the worst overutilization */
1014 if (ras->ras_consecutive == 3) {
1015 ras->ras_window_len = PTLRPC_MAX_BRW_PAGES;
1016 GOTO(out_unlock, 0);
1019 /* we need to increase the window sometimes. we'll arbitrarily
1020 * do it half-way through the pages in an rpc */
1021 if ((index & (PTLRPC_MAX_BRW_PAGES - 1)) ==
1022 (PTLRPC_MAX_BRW_PAGES >> 1)) {
1023 ras->ras_window_len += PTLRPC_MAX_BRW_PAGES;
1024 ras->ras_window_len = min(ras->ras_window_len,
1032 spin_unlock(&ras->ras_lock);
1033 spin_unlock(&sbi->ll_lock);
1037 * for now we do our readpage the same on both 2.4 and 2.5. The kernel's
1038 * read-ahead assumes it is valid to issue readpage all the way up to
1039 * i_size, but our dlm locks make that not the case. We disable the
1040 * kernel's read-ahead and do our own by walking ahead in the page cache
1041 * checking for dlm lock coverage. the main difference between 2.4 and
1042 * 2.6 is how read-ahead gets batched and issued, but we're using our own,
1043 * so they look the same.
1045 int ll_readpage(struct file *filp, struct page *page)
1047 struct ll_file_data *fd = filp->private_data;
1048 struct inode *inode = page->mapping->host;
1049 struct obd_export *exp;
1050 struct ll_async_page *llap;
1051 struct obd_io_group *oig = NULL;
1055 LASSERT(PageLocked(page));
1056 LASSERT(!PageUptodate(page));
1057 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset="LPX64"\n",
1058 inode->i_ino, inode->i_generation, inode,
1059 (((obd_off)page->index) << PAGE_SHIFT));
1060 LASSERT(atomic_read(&filp->f_dentry->d_inode->i_count) > 0);
1062 rc = oig_init(&oig);
1066 exp = ll_i2dtexp(inode);
1068 GOTO(out, rc = -EINVAL);
1070 llap = llap_from_page(page, LLAP_ORIGIN_READPAGE);
1072 GOTO(out, rc = PTR_ERR(llap));
1074 if (ll_i2sbi(inode)->ll_flags & LL_SBI_READAHEAD)
1075 ras_update(ll_i2sbi(inode), &fd->fd_ras, page->index,
1076 llap->llap_defer_uptodate);
1078 if (llap->llap_defer_uptodate) {
1079 llap->llap_ra_used = 1;
1080 rc = ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1083 obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd,
1085 LL_CDEBUG_PAGE(D_PAGE, page, "marking uptodate from defer\n");
1086 SetPageUptodate(page);
1088 GOTO(out_oig, rc = 0);
1091 rc = ll_page_matches(page, fd->fd_flags, 0);
1093 LL_CDEBUG_PAGE(D_ERROR, page, "lock match failed: rc %d\n", rc);
1098 CWARN("ino %lu page %lu (%llu) not covered by "
1099 "a lock (mmap?). check debug logs.\n",
1100 inode->i_ino, page->index,
1101 (long long)page->index << PAGE_CACHE_SHIFT);
1104 rc = ll_issue_page_read(exp, llap, oig, 0);
1108 LL_CDEBUG_PAGE(D_PAGE, page, "queued readpage\n");
1109 if (ll_i2sbi(inode)->ll_flags & LL_SBI_READAHEAD)
1110 ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1113 rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);