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 obd_capa *ocapa;
121 struct lustre_capa *capa = NULL;
122 struct obdo *oa = NULL;
126 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) to %llu\n", inode->i_ino,
127 inode->i_generation, inode, inode->i_size);
129 if (lli->lli_size_pid != current->pid) {
135 CDEBUG(D_INODE, "truncate on inode %lu with no objects\n",
140 LASSERT(atomic_read(&lli->lli_size_sem.count) <= 0);
142 if (lov_merge_size(lsm, 0) == inode->i_size) {
143 CDEBUG(D_VFSTRACE, "skipping punch for "LPX64" (size = %llu)\n",
144 lsm->lsm_object_id, inode->i_size);
148 CDEBUG(D_INFO, "calling punch for "LPX64" (new size %llu)\n",
149 lsm->lsm_object_id, inode->i_size);
153 CERROR("cannot alloc oa, error %d\n",
159 oa->o_id = lsm->lsm_object_id;
160 oa->o_gr = lsm->lsm_object_gr;
161 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP | OBD_MD_FLIFID;
162 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLMODE |
163 OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME);
164 memcpy(obdo_id(oa), &lli->lli_id, sizeof(lli->lli_id));
166 obd_adjust_kms(ll_i2dtexp(inode), lsm, inode->i_size, 1);
168 lli->lli_size_pid = 0;
169 up(&lli->lli_size_sem);
171 ocapa = ll_get_capa(inode, current->fsuid, CAPA_TRUNC);
173 capa = &ocapa->c_capa;
175 rc = obd_punch(ll_i2dtexp(inode), oa, lsm, inode->i_size,
176 OBD_OBJECT_EOF, NULL, capa);
178 CERROR("obd_truncate fails (%d) ino %lu\n", rc, inode->i_ino);
180 obdo_to_inode(inode, oa, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
181 OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME);
189 LASSERT(atomic_read(&lli->lli_size_sem.count) <= 0);
190 up(&lli->lli_size_sem);
193 struct ll_async_page *llap_cast_private(struct page *page)
195 struct ll_async_page *llap = (struct ll_async_page *)page->private;
197 LASSERTF(llap == NULL || llap->llap_magic == LLAP_MAGIC,
198 "page %p private %lu gave magic %d which != %d\n",
199 page, page->private, llap->llap_magic, LLAP_MAGIC);
203 int ll_prepare_write(struct file *file, struct page *page,
204 unsigned from, unsigned to)
206 struct inode *inode = page->mapping->host;
207 struct ll_inode_info *lli = ll_i2info(inode);
208 struct lov_stripe_md *lsm = lli->lli_smd;
209 obd_off offset = ((obd_off)page->index) << PAGE_SHIFT;
210 struct obdo *oa = NULL;
216 LASSERT(LLI_DIRTY_HANDLE(inode));
217 LASSERT(PageLocked(page));
218 (void)llap_cast_private(page); /* assertion */
220 /* Check to see if we should return -EIO right away */
222 pga.disk_offset = pga.page_offset = offset;
223 pga.count = PAGE_SIZE;
230 oa->o_id = lsm->lsm_object_id;
231 oa->o_gr = lsm->lsm_object_gr;
232 oa->o_mode = inode->i_mode;
234 oa->o_valid = OBD_MD_FLID | OBD_MD_FLMODE |
235 OBD_MD_FLTYPE | OBD_MD_FLGROUP;
237 oa->o_fsuid = current->fsuid;
238 oa->o_valid |= OBD_MD_FLFSUID;
239 *(obdo_id(oa)) = ll_i2info(inode)->lli_id;
241 rc = obd_brw(OBD_BRW_CHECK, ll_i2dtexp(inode),
242 oa, lsm, 1, &pga, NULL);
244 GOTO(out_free_oa, rc);
246 if (PageUptodate(page))
247 GOTO(out_free_oa, 0);
249 /* We're completely overwriting an existing page, so _don't_ set it up
250 * to date until commit_write */
251 if (from == 0 && to == PAGE_SIZE) {
252 POISON_PAGE(page, 0x11);
253 GOTO(out_free_oa, 0);
256 /* If are writing to a new page, no need to read old data. The extent
257 * locking will have updated the KMS, and for our purposes here we can
258 * treat it like i_size. */
259 down(&lli->lli_size_sem);
260 kms = lov_merge_size(lsm, 1);
261 up(&lli->lli_size_sem);
263 memset(kmap(page), 0, PAGE_SIZE);
265 GOTO(prepare_done, rc = 0);
268 /* XXX could be an async ocp read.. read-ahead? */
269 rc = ll_brw(OBD_BRW_READ, inode, oa, page, 0);
271 /* bug 1598: don't clobber blksize */
272 oa->o_valid &= ~(OBD_MD_FLSIZE | OBD_MD_FLBLKSZ);
273 obdo_refresh_inode(inode, oa, oa->o_valid);
274 } else if (rc == -ENOENT) {
275 /* tolerate no entry error here, cause the objects might
276 * not be created yet */
283 SetPageUptodate(page);
289 static int ll_ap_make_ready(void *data, int cmd)
291 struct ll_async_page *llap;
295 llap = LLAP_FROM_COOKIE(data);
296 page = llap->llap_page;
298 LASSERT(cmd != OBD_BRW_READ);
300 /* we're trying to write, but the page is locked.. come back later */
301 if (TryLockPage(page))
304 LL_CDEBUG_PAGE(D_PAGE, page, "made ready\n");
305 page_cache_get(page);
307 /* if we left PageDirty we might get another writepage call
308 * in the future. list walkers are bright enough
309 * to check page dirty so we can leave it on whatever list
310 * its on. XXX also, we're called with the cli list so if
311 * we got the page cache list we'd create a lock inversion
312 * with the removepage path which gets the page lock then the
314 clear_page_dirty(page);
318 /* We have two reasons for giving llite the opportunity to change the
319 * write length of a given queued page as it builds the RPC containing
322 * 1) Further extending writes may have landed in the page cache
323 * since a partial write first queued this page requiring us
324 * to write more from the page cache. (No further races are possible, since
325 * by the time this is called, the page is locked.)
326 * 2) We might have raced with truncate and want to avoid performing
327 * write RPCs that are just going to be thrown away by the
328 * truncate's punch on the storage targets.
330 * The kms serves these purposes as it is set at both truncate and extending
333 static int ll_ap_refresh_count(void *data, int cmd)
335 struct ll_inode_info *lli;
336 struct ll_async_page *llap;
337 struct lov_stripe_md *lsm;
342 /* readpage queues with _COUNT_STABLE, shouldn't get here. */
343 LASSERT(cmd != OBD_BRW_READ);
345 llap = LLAP_FROM_COOKIE(data);
346 page = llap->llap_page;
347 lli = ll_i2info(page->mapping->host);
351 * this callback is called with client lock taken, thus, it should not
352 * sleep or deadlock is possible. --umka
354 // down(&lli->lli_size_sem);
355 kms = lov_merge_size(lsm, 1);
356 // up(&lli->lli_size_sem);
358 /* catch race with truncate */
359 if (((__u64)page->index << PAGE_SHIFT) >= kms)
362 /* catch sub-page write at end of file */
363 if (((__u64)page->index << PAGE_SHIFT) + PAGE_SIZE > kms)
364 return kms % PAGE_SIZE;
369 void ll_inode_fill_obdo(struct inode *inode, int cmd, struct obdo *oa)
371 struct lov_stripe_md *lsm;
372 obd_valid valid_flags;
374 lsm = ll_i2info(inode)->lli_smd;
376 oa->o_id = lsm->lsm_object_id;
377 oa->o_gr = lsm->lsm_object_gr;
378 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
379 valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME;
380 if (cmd == OBD_BRW_WRITE || cmd == OBD_BRW_READ) {
381 oa->o_valid |= OBD_MD_FLIFID | OBD_MD_FLEPOCH;
382 *(obdo_id(oa)) = ll_i2info(inode)->lli_id;
383 oa->o_easize = ll_i2info(inode)->lli_io_epoch;
384 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME;
387 obdo_from_inode(oa, inode, valid_flags);
390 static void ll_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
392 struct ll_async_page *llap;
395 llap = LLAP_FROM_COOKIE(data);
396 ll_inode_fill_obdo(llap->llap_page->mapping->host, cmd, oa);
397 oa->o_fsuid = llap->llap_fsuid;
401 static struct obd_async_page_ops ll_async_page_ops = {
402 .ap_make_ready = ll_ap_make_ready,
403 .ap_refresh_count = ll_ap_refresh_count,
404 .ap_fill_obdo = ll_ap_fill_obdo,
405 .ap_completion = ll_ap_completion,
409 /* XXX have the exp be an argument? */
410 struct ll_async_page *llap_from_page(struct page *page, unsigned origin)
412 struct ll_async_page *llap;
413 struct obd_export *exp;
414 struct inode *inode = page->mapping->host;
415 struct ll_sb_info *sbi = ll_i2sbi(inode);
419 LASSERTF(origin < LLAP__ORIGIN_MAX, "%u\n", origin);
421 llap = llap_cast_private(page);
425 exp = ll_i2dtexp(page->mapping->host);
427 RETURN(ERR_PTR(-EINVAL));
429 OBD_ALLOC(llap, sizeof(*llap));
431 RETURN(ERR_PTR(-ENOMEM));
433 llap->llap_magic = LLAP_MAGIC;
434 INIT_LIST_HEAD(&llap->llap_pending_write);
435 rc = obd_prep_async_page(exp, ll_i2info(inode)->lli_smd, NULL, page,
436 (obd_off)page->index << PAGE_SHIFT,
437 &ll_async_page_ops, llap, &llap->llap_cookie);
439 OBD_FREE(llap, sizeof(*llap));
443 CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n", llap,
444 page, llap->llap_cookie, (obd_off)page->index << PAGE_SHIFT);
446 __set_page_ll_data(page, llap);
448 /* also zeroing the PRIVBITS low order bitflags */
449 llap->llap_page = page;
451 spin_lock(&sbi->ll_lock);
452 sbi->ll_pglist_gen++;
453 list_add_tail(&llap->llap_proc_item, &sbi->ll_pglist);
454 spin_unlock(&sbi->ll_lock);
457 llap->llap_origin = origin;
461 static int queue_or_sync_write(struct obd_export *exp,
462 struct lov_stripe_md *lsm,
463 struct ll_async_page *llap,
465 obd_flags async_flags)
467 struct obd_io_group *oig;
471 /* _make_ready only sees llap once we've unlocked the page */
472 llap->llap_write_queued = 1;
473 rc = obd_queue_async_io(exp, lsm, NULL, llap->llap_cookie,
474 OBD_BRW_WRITE, 0, 0, 0, async_flags);
476 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "write queued\n");
477 llap_write_pending(llap->llap_page->mapping->host, llap);
481 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 llap->llap_fsuid = current->fsuid;
544 /* queue a write for some time in the future the first time we
546 if (!PageDirty(page)) {
547 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
548 LPROC_LL_DIRTY_MISSES);
550 rc = queue_or_sync_write(exp, ll_i2info(inode)->lli_smd,
555 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
556 LPROC_LL_DIRTY_HITS);
559 /* put the page in the page cache, from now on ll_removepage is
560 * responsible for cleaning up the llap.
561 * don't dirty the page if it has been write out in q_o_s_w */
562 if (llap->llap_write_queued)
563 set_page_dirty(page);
566 size = (((obd_off)page->index) << PAGE_SHIFT) + to;
567 down(&lli->lli_size_sem);
569 obd_adjust_kms(exp, lsm, size, 0);
570 if (size > inode->i_size)
571 inode->i_size = size;
572 SetPageUptodate(page);
573 } else if (size > inode->i_size) {
574 /* this page beyond the pales of i_size, so it can't be
575 * truncated in ll_p_r_e during lock revoking. we must
576 * teardown our book-keeping here. */
579 up(&lli->lli_size_sem);
583 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, unsigned long len)
585 struct ll_ra_info *ra = &sbi->ll_ra_info;
589 spin_lock(&sbi->ll_lock);
590 ret = min(ra->ra_max_pages - ra->ra_cur_pages, len);
591 ra->ra_cur_pages += ret;
592 spin_unlock(&sbi->ll_lock);
597 static void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
599 struct ll_ra_info *ra = &sbi->ll_ra_info;
600 spin_lock(&sbi->ll_lock);
601 LASSERTF(ra->ra_cur_pages >= len, "r_c_p %lu len %lu\n",
602 ra->ra_cur_pages, len);
603 ra->ra_cur_pages -= len;
604 spin_unlock(&sbi->ll_lock);
607 int ll_writepage(struct page *page)
609 struct inode *inode = page->mapping->host;
610 struct obd_export *exp;
611 struct ll_async_page *llap;
615 LASSERT(!PageDirty(page));
616 LASSERT(PageLocked(page));
617 LASSERT(LLI_DIRTY_HANDLE(inode));
619 exp = ll_i2dtexp(inode);
621 GOTO(out, rc = -EINVAL);
623 llap = llap_from_page(page, LLAP_ORIGIN_WRITEPAGE);
625 GOTO(out, rc = PTR_ERR(llap));
627 page_cache_get(page);
628 if (llap->llap_write_queued) {
629 LL_CDEBUG_PAGE(D_PAGE, page, "marking urgent\n");
630 rc = obd_set_async_flags(exp, ll_i2info(inode)->lli_smd, NULL,
632 ASYNC_READY | ASYNC_URGENT);
634 rc = queue_or_sync_write(exp, ll_i2info(inode)->lli_smd, llap,
635 PAGE_SIZE, ASYNC_READY |
639 page_cache_release(page);
647 /* called for each page in a completed rpc.*/
648 void ll_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
650 struct ll_async_page *llap;
654 llap = LLAP_FROM_COOKIE(data);
655 page = llap->llap_page;
656 LASSERT(PageLocked(page));
658 LL_CDEBUG_PAGE(D_PAGE, page, "completing cmd %d with %d\n", cmd, rc);
660 if (cmd == OBD_BRW_READ && llap->llap_defer_uptodate)
661 ll_ra_count_put(ll_i2sbi(page->mapping->host), 1);
664 if (cmd == OBD_BRW_READ) {
665 if (!llap->llap_defer_uptodate)
666 SetPageUptodate(page);
668 llap->llap_write_queued = 0;
670 ClearPageError(page);
672 if (cmd == OBD_BRW_READ)
673 llap->llap_defer_uptodate = 0;
679 if (cmd == OBD_BRW_WRITE) {
680 llap_write_complete(page->mapping->host, llap);
681 ll_try_done_writing(page->mapping->host);
684 if (PageWriteback(page)) {
685 end_page_writeback(page);
687 page_cache_release(page);
691 /* the kernel calls us here when a page is unhashed from the page cache.
692 * the page will be locked and the kernel is holding a spinlock, so
693 * we need to be careful. we're just tearing down our book-keeping
695 void ll_removepage(struct page *page)
697 struct inode *inode = page->mapping->host;
698 struct obd_export *exp;
699 struct ll_async_page *llap;
700 struct ll_sb_info *sbi = ll_i2sbi(inode);
704 LASSERT(!in_interrupt());
706 /* sync pages or failed read pages can leave pages in the page
707 * cache that don't have our data associated with them anymore */
708 if (page->private == 0) {
713 LL_CDEBUG_PAGE(D_PAGE, page, "being evicted\n");
715 exp = ll_i2dtexp(inode);
717 CERROR("page %p ind %lu gave null export\n", page, page->index);
721 llap = llap_from_page(page, 0);
723 CERROR("page %p ind %lu couldn't find llap: %ld\n", page,
724 page->index, PTR_ERR(llap));
729 llap_write_complete(inode, llap);
730 rc = obd_teardown_async_page(exp, ll_i2info(inode)->lli_smd, NULL,
733 CERROR("page %p ind %lu failed: %d\n", page, page->index, rc);
735 /* this unconditional free is only safe because the page lock
736 * is providing exclusivity to memory pressure/truncate/writeback..*/
737 __clear_page_ll_data(page);
739 spin_lock(&sbi->ll_lock);
740 if (!list_empty(&llap->llap_proc_item))
741 list_del_init(&llap->llap_proc_item);
742 sbi->ll_pglist_gen++;
743 spin_unlock(&sbi->ll_lock);
744 OBD_FREE(llap, sizeof(*llap));
748 static int ll_page_matches(struct page *page, int fd_flags, int readahead)
750 struct lustre_handle match_lockh = {0};
751 struct inode *inode = page->mapping->host;
752 ldlm_policy_data_t page_extent;
756 if (fd_flags & LL_FILE_GROUP_LOCKED)
759 page_extent.l_extent.start = (__u64)page->index << PAGE_CACHE_SHIFT;
760 page_extent.l_extent.end =
761 page_extent.l_extent.start + PAGE_CACHE_SIZE - 1;
762 flags = LDLM_FL_TEST_LOCK;
764 flags |= LDLM_FL_CBPENDING | LDLM_FL_BLOCK_GRANTED;
765 matches = obd_match(ll_i2sbi(inode)->ll_dt_exp,
766 ll_i2info(inode)->lli_smd, LDLM_EXTENT,
767 &page_extent, LCK_PR | LCK_PW, &flags, inode,
772 static int ll_issue_page_read(struct obd_export *exp,
773 struct ll_async_page *llap,
774 struct obd_io_group *oig, int defer)
776 struct page *page = llap->llap_page;
779 page_cache_get(page);
780 llap->llap_defer_uptodate = defer;
781 llap->llap_ra_used = 0;
783 rc = obd_queue_group_io(exp, ll_i2info(page->mapping->host)->lli_smd,
784 NULL, oig, llap->llap_cookie, OBD_BRW_READ, 0,
785 PAGE_SIZE, 0, ASYNC_COUNT_STABLE | ASYNC_READY
788 LL_CDEBUG_PAGE(D_ERROR, page, "read queue failed: rc %d\n", rc);
789 page_cache_release(page);
794 static void ll_ra_stats_inc_unlocked(struct ll_ra_info *ra, enum ra_stat which)
796 LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
797 ra->ra_stats[which]++;
800 static void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which)
802 struct ll_sb_info *sbi = ll_i2sbi(mapping->host);
803 struct ll_ra_info *ra = &ll_i2sbi(mapping->host)->ll_ra_info;
805 spin_lock(&sbi->ll_lock);
806 ll_ra_stats_inc_unlocked(ra, which);
807 spin_unlock(&sbi->ll_lock);
810 void ll_ra_accounting(struct page *page, struct address_space *mapping)
812 struct ll_async_page *llap;
814 llap = llap_from_page(page, LLAP_ORIGIN_WRITEPAGE);
818 if (!llap->llap_defer_uptodate || llap->llap_ra_used)
821 ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
824 #define RAS_CDEBUG(ras) \
825 CDEBUG(D_READA, "lrp %lu c %lu ws %lu wl %lu nra %lu\n", \
826 ras->ras_last_readpage, ras->ras_consecutive, \
827 ras->ras_window_start, ras->ras_window_len, \
828 ras->ras_next_readahead);
830 static int index_in_window(unsigned long index, unsigned long point,
831 unsigned long before, unsigned long after)
833 unsigned long start = point - before, end = point + after;
840 return start <= index && index <= end;
843 static int ll_readahead(struct ll_readahead_state *ras,
844 struct obd_export *exp, struct address_space *mapping,
845 struct obd_io_group *oig, int flags)
847 unsigned long i, start = 0, end = 0, reserved;
848 struct ll_async_page *llap;
850 int rc, ret = 0, match_failed = 0;
854 kms = lov_merge_size(ll_i2info(mapping->host)->lli_smd, 1);
856 ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN);
859 spin_lock(&ras->ras_lock);
861 /* reserve a part of the read-ahead window that we'll be issuing */
862 if (ras->ras_window_len) {
863 start = ras->ras_next_readahead;
864 end = ras->ras_window_start + ras->ras_window_len - 1;
865 end = min(end, (unsigned long)(kms >> PAGE_CACHE_SHIFT));
866 ras->ras_next_readahead = max(end, end + 1);
871 spin_unlock(&ras->ras_lock);
874 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
878 reserved = ll_ra_count_get(ll_i2sbi(mapping->host), end - start + 1);
879 if (reserved < end - start + 1)
880 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
882 for (i = start; reserved > 0 && !match_failed && i <= end; i++) {
883 /* skip locked pages from previous readpage calls */
884 page = grab_cache_page_nowait(mapping, i);
886 CDEBUG(D_READA, "g_c_p_n failed\n");
890 /* we do this first so that we can see the page in the /proc
892 llap = llap_from_page(page, LLAP_ORIGIN_READAHEAD);
893 if (IS_ERR(llap) || llap->llap_defer_uptodate)
896 /* skip completed pages */
897 if (Page_Uptodate(page))
900 /* bail when we hit the end of the lock. */
901 if ((rc = ll_page_matches(page, flags, 1)) <= 0) {
902 LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
903 "lock match failed: rc %d\n", rc);
904 ll_ra_stats_inc(mapping, RA_STAT_FAILED_MATCH);
909 rc = ll_issue_page_read(exp, llap, oig, 1);
913 LL_CDEBUG_PAGE(D_READA| D_PAGE, page,
914 "started read-ahead\n");
918 LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
919 "skipping read-ahead\n");
923 page_cache_release(page);
926 LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
928 ll_ra_count_put(ll_i2sbi(mapping->host), reserved);
930 if (i == end + 1 && end == (kms >> PAGE_CACHE_SHIFT))
931 ll_ra_stats_inc(mapping, RA_STAT_EOF);
933 /* if we didn't get to the end of the region we reserved from
934 * the ras we need to go back and update the ras so that the
935 * next read-ahead tries from where we left off. we only do so
936 * if the region we failed to issue read-ahead on is still ahead
937 * of the app and behind the next index to start read-ahead from */
939 spin_lock(&ras->ras_lock);
940 if (i < ras->ras_next_readahead &&
941 index_in_window(i, ras->ras_window_start, 0,
942 ras->ras_window_len)) {
943 ras->ras_next_readahead = i;
946 spin_unlock(&ras->ras_lock);
952 static void ras_set_start(struct ll_readahead_state *ras, unsigned long index)
954 ras->ras_window_start = index & (~(PTLRPC_MAX_BRW_PAGES - 1));
957 /* called with the ras_lock held or from places where it doesn't matter */
958 static void ras_reset(struct ll_readahead_state *ras, unsigned long index)
960 ras->ras_last_readpage = index;
961 ras->ras_consecutive = 1;
962 ras->ras_window_len = 0;
963 ras_set_start(ras, index);
964 ras->ras_next_readahead = ras->ras_window_start;
969 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
971 spin_lock_init(&ras->ras_lock);
975 static void ras_update(struct ll_sb_info *sbi, struct ll_readahead_state *ras,
976 unsigned long index, unsigned hit)
978 struct ll_ra_info *ra = &sbi->ll_ra_info;
982 spin_lock(&sbi->ll_lock);
983 spin_lock(&ras->ras_lock);
985 ll_ra_stats_inc_unlocked(ra, hit ? RA_STAT_HIT : RA_STAT_MISS);
987 /* reset the read-ahead window in two cases. First when the app seeks
988 * or reads to some other part of the file. Secondly if we get a
989 * read-ahead miss that we think we've previously issued. This can
990 * be a symptom of there being so many read-ahead pages that the VM is
991 * reclaiming it before we get to it. */
992 if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
994 ll_ra_stats_inc_unlocked(ra, RA_STAT_DISTANT_READPAGE);
995 } else if (!hit && ras->ras_window_len &&
996 index < ras->ras_next_readahead &&
997 index_in_window(index, ras->ras_window_start, 0,
998 ras->ras_window_len)) {
1000 ll_ra_stats_inc_unlocked(ra, RA_STAT_MISS_IN_WINDOW);
1004 ras_reset(ras, index);
1005 GOTO(out_unlock, 0);
1008 ras->ras_last_readpage = index;
1009 ras->ras_consecutive++;
1010 ras_set_start(ras, index);
1011 ras->ras_next_readahead = max(ras->ras_window_start,
1012 ras->ras_next_readahead);
1014 /* wait for a few pages to arrive before issuing readahead to avoid
1015 * the worst overutilization */
1016 if (ras->ras_consecutive == 3) {
1017 ras->ras_window_len = PTLRPC_MAX_BRW_PAGES;
1018 GOTO(out_unlock, 0);
1021 /* we need to increase the window sometimes. we'll arbitrarily
1022 * do it half-way through the pages in an rpc */
1023 if ((index & (PTLRPC_MAX_BRW_PAGES - 1)) ==
1024 (PTLRPC_MAX_BRW_PAGES >> 1)) {
1025 ras->ras_window_len += PTLRPC_MAX_BRW_PAGES;
1026 ras->ras_window_len = min(ras->ras_window_len,
1034 spin_unlock(&ras->ras_lock);
1035 spin_unlock(&sbi->ll_lock);
1039 * for now we do our readpage the same on both 2.4 and 2.5. The kernel's
1040 * read-ahead assumes it is valid to issue readpage all the way up to
1041 * i_size, but our dlm locks make that not the case. We disable the
1042 * kernel's read-ahead and do our own by walking ahead in the page cache
1043 * checking for dlm lock coverage. the main difference between 2.4 and
1044 * 2.6 is how read-ahead gets batched and issued, but we're using our own,
1045 * so they look the same.
1047 int ll_readpage(struct file *filp, struct page *page)
1049 struct ll_file_data *fd = filp->private_data;
1050 struct inode *inode = page->mapping->host;
1051 struct obd_export *exp;
1052 struct ll_async_page *llap;
1053 struct obd_io_group *oig = NULL;
1057 LASSERT(PageLocked(page));
1058 LASSERT(!PageUptodate(page));
1059 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset="LPX64"\n",
1060 inode->i_ino, inode->i_generation, inode,
1061 (((obd_off)page->index) << PAGE_SHIFT));
1062 LASSERT(atomic_read(&filp->f_dentry->d_inode->i_count) > 0);
1064 rc = oig_init(&oig);
1068 exp = ll_i2dtexp(inode);
1070 GOTO(out, rc = -EINVAL);
1072 llap = llap_from_page(page, LLAP_ORIGIN_READPAGE);
1074 GOTO(out, rc = PTR_ERR(llap));
1076 /* capability need this */
1077 llap->llap_fsuid = current->fsuid;
1079 if (ll_i2sbi(inode)->ll_flags & LL_SBI_READAHEAD)
1080 ras_update(ll_i2sbi(inode), &fd->fd_ras, page->index,
1081 llap->llap_defer_uptodate);
1083 if (llap->llap_defer_uptodate) {
1084 llap->llap_ra_used = 1;
1085 rc = ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1088 obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd,
1090 LL_CDEBUG_PAGE(D_PAGE, page, "marking uptodate from defer\n");
1091 SetPageUptodate(page);
1093 GOTO(out_oig, rc = 0);
1096 rc = ll_page_matches(page, fd->fd_flags, 0);
1098 LL_CDEBUG_PAGE(D_ERROR, page, "lock match failed: rc %d\n", rc);
1103 CWARN("ino %lu page %lu (%llu) not covered by "
1104 "a lock (mmap?). check debug logs.\n",
1105 inode->i_ino, page->index,
1106 (long long)page->index << PAGE_CACHE_SHIFT);
1109 rc = ll_issue_page_read(exp, llap, oig, 0);
1113 LL_CDEBUG_PAGE(D_PAGE, page, "queued readpage\n");
1114 if (ll_i2sbi(inode)->ll_flags & LL_SBI_READAHEAD)
1115 ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1118 rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);