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);
179 CERROR("obd_truncate fails (%d) ino %lu\n", rc, inode->i_ino);
181 obdo_to_inode(inode, oa, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
182 OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME);
190 LASSERT(atomic_read(&lli->lli_size_sem.count) <= 0);
191 up(&lli->lli_size_sem);
194 struct ll_async_page *llap_cast_private(struct page *page)
196 struct ll_async_page *llap = (struct ll_async_page *)page->private;
198 LASSERTF(llap == NULL || llap->llap_magic == LLAP_MAGIC,
199 "page %p private %lu gave magic %d which != %d\n",
200 page, page->private, llap->llap_magic, LLAP_MAGIC);
204 int ll_prepare_write(struct file *file, struct page *page,
205 unsigned from, unsigned to)
207 struct inode *inode = page->mapping->host;
208 struct ll_inode_info *lli = ll_i2info(inode);
209 struct lov_stripe_md *lsm = lli->lli_smd;
210 obd_off offset = ((obd_off)page->index) << PAGE_SHIFT;
211 struct obdo *oa = NULL;
217 LASSERT(LLI_DIRTY_HANDLE(inode));
218 LASSERT(PageLocked(page));
219 (void)llap_cast_private(page); /* assertion */
221 /* Check to see if we should return -EIO right away */
223 pga.disk_offset = pga.page_offset = offset;
224 pga.count = PAGE_SIZE;
231 oa->o_id = lsm->lsm_object_id;
232 oa->o_gr = lsm->lsm_object_gr;
233 oa->o_mode = inode->i_mode;
235 oa->o_valid = OBD_MD_FLID | OBD_MD_FLMODE |
236 OBD_MD_FLTYPE | OBD_MD_FLGROUP;
238 oa->o_fsuid = current->fsuid;
239 oa->o_valid |= OBD_MD_FLFSUID;
240 *(obdo_id(oa)) = ll_i2info(inode)->lli_id;
242 rc = obd_brw(OBD_BRW_CHECK, ll_i2dtexp(inode),
243 oa, lsm, 1, &pga, NULL);
245 GOTO(out_free_oa, rc);
247 if (PageUptodate(page))
248 GOTO(out_free_oa, 0);
250 /* We're completely overwriting an existing page, so _don't_ set it up
251 * to date until commit_write */
252 if (from == 0 && to == PAGE_SIZE) {
253 POISON_PAGE(page, 0x11);
254 GOTO(out_free_oa, 0);
257 /* If are writing to a new page, no need to read old data. The extent
258 * locking will have updated the KMS, and for our purposes here we can
259 * treat it like i_size. */
260 down(&lli->lli_size_sem);
261 kms = lov_merge_size(lsm, 1);
262 up(&lli->lli_size_sem);
264 memset(kmap(page), 0, PAGE_SIZE);
266 GOTO(prepare_done, rc = 0);
269 /* XXX could be an async ocp read.. read-ahead? */
270 rc = ll_brw(OBD_BRW_READ, inode, oa, page, 0);
272 /* bug 1598: don't clobber blksize */
273 oa->o_valid &= ~(OBD_MD_FLSIZE | OBD_MD_FLBLKSZ);
274 obdo_refresh_inode(inode, oa, oa->o_valid);
275 } else if (rc == -ENOENT) {
276 /* tolerate no entry error here, cause the objects might
277 * not be created yet */
284 SetPageUptodate(page);
290 static int ll_ap_make_ready(void *data, int cmd)
292 struct ll_async_page *llap;
296 llap = LLAP_FROM_COOKIE(data);
297 page = llap->llap_page;
299 LASSERT(cmd != OBD_BRW_READ);
301 /* we're trying to write, but the page is locked.. come back later */
302 if (TryLockPage(page))
305 LL_CDEBUG_PAGE(D_PAGE, page, "made ready\n");
306 page_cache_get(page);
308 /* if we left PageDirty we might get another writepage call
309 * in the future. list walkers are bright enough
310 * to check page dirty so we can leave it on whatever list
311 * its on. XXX also, we're called with the cli list so if
312 * we got the page cache list we'd create a lock inversion
313 * with the removepage path which gets the page lock then the
315 clear_page_dirty(page);
319 /* We have two reasons for giving llite the opportunity to change the
320 * write length of a given queued page as it builds the RPC containing
323 * 1) Further extending writes may have landed in the page cache
324 * since a partial write first queued this page requiring us
325 * to write more from the page cache. (No further races are possible, since
326 * by the time this is called, the page is locked.)
327 * 2) We might have raced with truncate and want to avoid performing
328 * write RPCs that are just going to be thrown away by the
329 * truncate's punch on the storage targets.
331 * The kms serves these purposes as it is set at both truncate and extending
334 static int ll_ap_refresh_count(void *data, int cmd)
336 struct ll_inode_info *lli;
337 struct ll_async_page *llap;
338 struct lov_stripe_md *lsm;
343 /* readpage queues with _COUNT_STABLE, shouldn't get here. */
344 LASSERT(cmd != OBD_BRW_READ);
346 llap = LLAP_FROM_COOKIE(data);
347 page = llap->llap_page;
348 lli = ll_i2info(page->mapping->host);
352 * this callback is called with client lock taken, thus, it should not
353 * sleep or deadlock is possible. --umka
355 // down(&lli->lli_size_sem);
356 kms = lov_merge_size(lsm, 1);
357 // up(&lli->lli_size_sem);
359 /* catch race with truncate */
360 if (((__u64)page->index << PAGE_SHIFT) >= kms)
363 /* catch sub-page write at end of file */
364 if (((__u64)page->index << PAGE_SHIFT) + PAGE_SIZE > kms)
365 return kms % PAGE_SIZE;
370 void ll_inode_fill_obdo(struct inode *inode, int cmd, struct obdo *oa)
372 struct lov_stripe_md *lsm;
373 obd_valid valid_flags;
375 lsm = ll_i2info(inode)->lli_smd;
377 oa->o_id = lsm->lsm_object_id;
378 oa->o_gr = lsm->lsm_object_gr;
379 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
380 valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME;
381 if (cmd == OBD_BRW_WRITE || cmd == OBD_BRW_READ) {
382 oa->o_valid |= OBD_MD_FLIFID | OBD_MD_FLEPOCH;
383 *(obdo_id(oa)) = ll_i2info(inode)->lli_id;
384 oa->o_easize = ll_i2info(inode)->lli_io_epoch;
385 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME;
388 obdo_from_inode(oa, inode, valid_flags);
391 static void ll_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
393 struct ll_async_page *llap;
396 llap = LLAP_FROM_COOKIE(data);
397 ll_inode_fill_obdo(llap->llap_page->mapping->host, cmd, oa);
398 oa->o_fsuid = llap->llap_fsuid;
402 static struct obd_async_page_ops ll_async_page_ops = {
403 .ap_make_ready = ll_ap_make_ready,
404 .ap_refresh_count = ll_ap_refresh_count,
405 .ap_fill_obdo = ll_ap_fill_obdo,
406 .ap_completion = ll_ap_completion,
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));
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);
447 __set_page_ll_data(page, llap);
449 /* also zeroing the PRIVBITS low order bitflags */
450 llap->llap_page = page;
452 spin_lock(&sbi->ll_lock);
453 sbi->ll_pglist_gen++;
454 list_add_tail(&llap->llap_proc_item, &sbi->ll_pglist);
455 spin_unlock(&sbi->ll_lock);
458 llap->llap_origin = origin;
462 static int queue_or_sync_write(struct obd_export *exp,
463 struct lov_stripe_md *lsm,
464 struct ll_async_page *llap,
466 obd_flags async_flags)
468 struct obd_io_group *oig;
472 /* _make_ready only sees llap once we've unlocked the page */
473 llap->llap_write_queued = 1;
474 rc = obd_queue_async_io(exp, lsm, NULL, llap->llap_cookie,
475 OBD_BRW_WRITE, 0, 0, 0, async_flags);
477 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "write queued\n");
478 llap_write_pending(llap->llap_page->mapping->host, llap);
482 llap->llap_write_queued = 0;
487 rc = obd_queue_group_io(exp, lsm, NULL, oig, llap->llap_cookie,
488 OBD_BRW_WRITE, 0, to, 0, ASYNC_READY |
489 ASYNC_URGENT | ASYNC_COUNT_STABLE |
494 rc = obd_trigger_group_io(exp, lsm, NULL, oig);
500 if (!rc && async_flags & ASYNC_READY)
501 unlock_page(llap->llap_page);
503 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
504 "sync write returned %d\n", rc);
513 /* be careful not to return success without setting the page Uptodate or
514 * the next pass through prepare_write will read in stale data from disk. */
515 int ll_commit_write(struct file *file, struct page *page, unsigned from,
518 struct inode *inode = page->mapping->host;
519 struct ll_inode_info *lli = ll_i2info(inode);
520 struct lov_stripe_md *lsm = lli->lli_smd;
521 struct obd_export *exp = NULL;
522 struct ll_async_page *llap;
527 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
528 LASSERT(inode == file->f_dentry->d_inode);
529 LASSERT(PageLocked(page));
530 LASSERT(LLI_DIRTY_HANDLE(inode));
532 CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
533 inode, page, from, to, page->index);
535 llap = llap_from_page(page, LLAP_ORIGIN_COMMIT_WRITE);
537 RETURN(PTR_ERR(llap));
539 exp = ll_i2dtexp(inode);
543 llap->llap_fsuid = current->fsuid;
545 /* queue a write for some time in the future the first time we
547 if (!PageDirty(page)) {
548 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
549 LPROC_LL_DIRTY_MISSES);
551 rc = queue_or_sync_write(exp, ll_i2info(inode)->lli_smd,
556 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
557 LPROC_LL_DIRTY_HITS);
560 /* put the page in the page cache, from now on ll_removepage is
561 * responsible for cleaning up the llap.
562 * don't dirty the page if it has been write out in q_o_s_w */
563 if (llap->llap_write_queued)
564 set_page_dirty(page);
567 size = (((obd_off)page->index) << PAGE_SHIFT) + to;
568 down(&lli->lli_size_sem);
570 obd_adjust_kms(exp, lsm, size, 0);
571 if (size > inode->i_size)
572 inode->i_size = size;
573 SetPageUptodate(page);
574 } else if (size > inode->i_size) {
575 /* this page beyond the pales of i_size, so it can't be
576 * truncated in ll_p_r_e during lock revoking. we must
577 * teardown our book-keeping here. */
580 up(&lli->lli_size_sem);
584 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, unsigned long len)
586 struct ll_ra_info *ra = &sbi->ll_ra_info;
590 spin_lock(&sbi->ll_lock);
591 ret = min(ra->ra_max_pages - ra->ra_cur_pages, len);
592 ra->ra_cur_pages += ret;
593 spin_unlock(&sbi->ll_lock);
598 static void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
600 struct ll_ra_info *ra = &sbi->ll_ra_info;
601 spin_lock(&sbi->ll_lock);
602 LASSERTF(ra->ra_cur_pages >= len, "r_c_p %lu len %lu\n",
603 ra->ra_cur_pages, len);
604 ra->ra_cur_pages -= len;
605 spin_unlock(&sbi->ll_lock);
608 int ll_writepage(struct page *page)
610 struct inode *inode = page->mapping->host;
611 struct obd_export *exp;
612 struct ll_async_page *llap;
616 LASSERT(!PageDirty(page));
617 LASSERT(PageLocked(page));
618 LASSERT(LLI_DIRTY_HANDLE(inode));
620 exp = ll_i2dtexp(inode);
622 GOTO(out, rc = -EINVAL);
624 llap = llap_from_page(page, LLAP_ORIGIN_WRITEPAGE);
626 GOTO(out, rc = PTR_ERR(llap));
628 page_cache_get(page);
629 if (llap->llap_write_queued) {
630 LL_CDEBUG_PAGE(D_PAGE, page, "marking urgent\n");
631 rc = obd_set_async_flags(exp, ll_i2info(inode)->lli_smd, NULL,
633 ASYNC_READY | ASYNC_URGENT);
635 rc = queue_or_sync_write(exp, ll_i2info(inode)->lli_smd, llap,
636 PAGE_SIZE, ASYNC_READY |
640 page_cache_release(page);
648 /* called for each page in a completed rpc.*/
649 void ll_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
651 struct ll_async_page *llap;
655 llap = LLAP_FROM_COOKIE(data);
656 page = llap->llap_page;
657 LASSERT(PageLocked(page));
659 LL_CDEBUG_PAGE(D_PAGE, page, "completing cmd %d with %d\n", cmd, rc);
661 if (cmd == OBD_BRW_READ && llap->llap_defer_uptodate)
662 ll_ra_count_put(ll_i2sbi(page->mapping->host), 1);
665 if (cmd == OBD_BRW_READ) {
666 if (!llap->llap_defer_uptodate)
667 SetPageUptodate(page);
669 llap->llap_write_queued = 0;
671 ClearPageError(page);
673 if (cmd == OBD_BRW_READ)
674 llap->llap_defer_uptodate = 0;
680 if (cmd == OBD_BRW_WRITE) {
681 llap_write_complete(page->mapping->host, llap);
682 ll_try_done_writing(page->mapping->host);
685 if (PageWriteback(page)) {
686 end_page_writeback(page);
688 page_cache_release(page);
692 /* the kernel calls us here when a page is unhashed from the page cache.
693 * the page will be locked and the kernel is holding a spinlock, so
694 * we need to be careful. we're just tearing down our book-keeping
696 void ll_removepage(struct page *page)
698 struct inode *inode = page->mapping->host;
699 struct obd_export *exp;
700 struct ll_async_page *llap;
701 struct ll_sb_info *sbi = ll_i2sbi(inode);
705 LASSERT(!in_interrupt());
707 /* sync pages or failed read pages can leave pages in the page
708 * cache that don't have our data associated with them anymore */
709 if (page->private == 0) {
714 LL_CDEBUG_PAGE(D_PAGE, page, "being evicted\n");
716 exp = ll_i2dtexp(inode);
718 CERROR("page %p ind %lu gave null export\n", page, page->index);
722 llap = llap_from_page(page, 0);
724 CERROR("page %p ind %lu couldn't find llap: %ld\n", page,
725 page->index, PTR_ERR(llap));
730 llap_write_complete(inode, llap);
731 rc = obd_teardown_async_page(exp, ll_i2info(inode)->lli_smd, NULL,
734 CERROR("page %p ind %lu failed: %d\n", page, page->index, rc);
736 /* this unconditional free is only safe because the page lock
737 * is providing exclusivity to memory pressure/truncate/writeback..*/
738 __clear_page_ll_data(page);
740 spin_lock(&sbi->ll_lock);
741 if (!list_empty(&llap->llap_proc_item))
742 list_del_init(&llap->llap_proc_item);
743 sbi->ll_pglist_gen++;
744 spin_unlock(&sbi->ll_lock);
745 OBD_FREE(llap, sizeof(*llap));
749 static int ll_page_matches(struct page *page, int fd_flags, int readahead)
751 struct lustre_handle match_lockh = {0};
752 struct inode *inode = page->mapping->host;
753 ldlm_policy_data_t page_extent;
757 if (fd_flags & LL_FILE_GROUP_LOCKED)
760 page_extent.l_extent.start = (__u64)page->index << PAGE_CACHE_SHIFT;
761 page_extent.l_extent.end =
762 page_extent.l_extent.start + PAGE_CACHE_SIZE - 1;
763 flags = LDLM_FL_TEST_LOCK;
765 flags |= LDLM_FL_CBPENDING | LDLM_FL_BLOCK_GRANTED;
766 matches = obd_match(ll_i2sbi(inode)->ll_dt_exp,
767 ll_i2info(inode)->lli_smd, LDLM_EXTENT,
768 &page_extent, LCK_PR | LCK_PW, &flags, inode,
773 static int ll_issue_page_read(struct obd_export *exp,
774 struct ll_async_page *llap,
775 struct obd_io_group *oig, int defer)
777 struct page *page = llap->llap_page;
780 page_cache_get(page);
781 llap->llap_defer_uptodate = defer;
782 llap->llap_ra_used = 0;
784 rc = obd_queue_group_io(exp, ll_i2info(page->mapping->host)->lli_smd,
785 NULL, oig, llap->llap_cookie, OBD_BRW_READ, 0,
786 PAGE_SIZE, 0, ASYNC_COUNT_STABLE | ASYNC_READY
789 LL_CDEBUG_PAGE(D_ERROR, page, "read queue failed: rc %d\n", rc);
790 page_cache_release(page);
795 static void ll_ra_stats_inc_unlocked(struct ll_ra_info *ra, enum ra_stat which)
797 LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
798 ra->ra_stats[which]++;
801 static void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which)
803 struct ll_sb_info *sbi = ll_i2sbi(mapping->host);
804 struct ll_ra_info *ra = &ll_i2sbi(mapping->host)->ll_ra_info;
806 spin_lock(&sbi->ll_lock);
807 ll_ra_stats_inc_unlocked(ra, which);
808 spin_unlock(&sbi->ll_lock);
811 void ll_ra_accounting(struct page *page, struct address_space *mapping)
813 struct ll_async_page *llap;
815 llap = llap_from_page(page, LLAP_ORIGIN_WRITEPAGE);
819 if (!llap->llap_defer_uptodate || llap->llap_ra_used)
822 ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
825 #define RAS_CDEBUG(ras) \
826 CDEBUG(D_READA, "lrp %lu c %lu ws %lu wl %lu nra %lu\n", \
827 ras->ras_last_readpage, ras->ras_consecutive, \
828 ras->ras_window_start, ras->ras_window_len, \
829 ras->ras_next_readahead);
831 static int index_in_window(unsigned long index, unsigned long point,
832 unsigned long before, unsigned long after)
834 unsigned long start = point - before, end = point + after;
841 return start <= index && index <= end;
844 static int ll_readahead(struct ll_readahead_state *ras,
845 struct obd_export *exp, struct address_space *mapping,
846 struct obd_io_group *oig, int flags)
848 unsigned long i, start = 0, end = 0, reserved;
849 struct ll_async_page *llap;
851 int rc, ret = 0, match_failed = 0;
855 kms = lov_merge_size(ll_i2info(mapping->host)->lli_smd, 1);
857 ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN);
860 spin_lock(&ras->ras_lock);
862 /* reserve a part of the read-ahead window that we'll be issuing */
863 if (ras->ras_window_len) {
864 start = ras->ras_next_readahead;
865 end = ras->ras_window_start + ras->ras_window_len - 1;
866 end = min(end, (unsigned long)(kms >> PAGE_CACHE_SHIFT));
867 ras->ras_next_readahead = max(end, end + 1);
872 spin_unlock(&ras->ras_lock);
875 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
879 reserved = ll_ra_count_get(ll_i2sbi(mapping->host), end - start + 1);
880 if (reserved < end - start + 1)
881 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
883 for (i = start; reserved > 0 && !match_failed && i <= end; i++) {
884 /* skip locked pages from previous readpage calls */
885 page = grab_cache_page_nowait(mapping, i);
887 CDEBUG(D_READA, "g_c_p_n failed\n");
891 /* we do this first so that we can see the page in the /proc
893 llap = llap_from_page(page, LLAP_ORIGIN_READAHEAD);
894 if (IS_ERR(llap) || llap->llap_defer_uptodate)
897 /* skip completed pages */
898 if (Page_Uptodate(page))
901 /* bail when we hit the end of the lock. */
902 if ((rc = ll_page_matches(page, flags, 1)) <= 0) {
903 LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
904 "lock match failed: rc %d\n", rc);
905 ll_ra_stats_inc(mapping, RA_STAT_FAILED_MATCH);
910 rc = ll_issue_page_read(exp, llap, oig, 1);
914 LL_CDEBUG_PAGE(D_READA| D_PAGE, page,
915 "started read-ahead\n");
919 LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
920 "skipping read-ahead\n");
924 page_cache_release(page);
927 LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
929 ll_ra_count_put(ll_i2sbi(mapping->host), reserved);
931 if (i == end + 1 && end == (kms >> PAGE_CACHE_SHIFT))
932 ll_ra_stats_inc(mapping, RA_STAT_EOF);
934 /* if we didn't get to the end of the region we reserved from
935 * the ras we need to go back and update the ras so that the
936 * next read-ahead tries from where we left off. we only do so
937 * if the region we failed to issue read-ahead on is still ahead
938 * of the app and behind the next index to start read-ahead from */
940 spin_lock(&ras->ras_lock);
941 if (i < ras->ras_next_readahead &&
942 index_in_window(i, ras->ras_window_start, 0,
943 ras->ras_window_len)) {
944 ras->ras_next_readahead = i;
947 spin_unlock(&ras->ras_lock);
953 static void ras_set_start(struct ll_readahead_state *ras, unsigned long index)
955 ras->ras_window_start = index & (~(PTLRPC_MAX_BRW_PAGES - 1));
958 /* called with the ras_lock held or from places where it doesn't matter */
959 static void ras_reset(struct ll_readahead_state *ras, unsigned long index)
961 ras->ras_last_readpage = index;
962 ras->ras_consecutive = 1;
963 ras->ras_window_len = 0;
964 ras_set_start(ras, index);
965 ras->ras_next_readahead = ras->ras_window_start;
970 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
972 spin_lock_init(&ras->ras_lock);
976 static void ras_update(struct ll_sb_info *sbi, struct ll_readahead_state *ras,
977 unsigned long index, unsigned hit)
979 struct ll_ra_info *ra = &sbi->ll_ra_info;
983 spin_lock(&sbi->ll_lock);
984 spin_lock(&ras->ras_lock);
986 ll_ra_stats_inc_unlocked(ra, hit ? RA_STAT_HIT : RA_STAT_MISS);
988 /* reset the read-ahead window in two cases. First when the app seeks
989 * or reads to some other part of the file. Secondly if we get a
990 * read-ahead miss that we think we've previously issued. This can
991 * be a symptom of there being so many read-ahead pages that the VM is
992 * reclaiming it before we get to it. */
993 if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
995 ll_ra_stats_inc_unlocked(ra, RA_STAT_DISTANT_READPAGE);
996 } else if (!hit && ras->ras_window_len &&
997 index < ras->ras_next_readahead &&
998 index_in_window(index, ras->ras_window_start, 0,
999 ras->ras_window_len)) {
1001 ll_ra_stats_inc_unlocked(ra, RA_STAT_MISS_IN_WINDOW);
1005 ras_reset(ras, index);
1006 GOTO(out_unlock, 0);
1009 ras->ras_last_readpage = index;
1010 ras->ras_consecutive++;
1011 ras_set_start(ras, index);
1012 ras->ras_next_readahead = max(ras->ras_window_start,
1013 ras->ras_next_readahead);
1015 /* wait for a few pages to arrive before issuing readahead to avoid
1016 * the worst overutilization */
1017 if (ras->ras_consecutive == 3) {
1018 ras->ras_window_len = PTLRPC_MAX_BRW_PAGES;
1019 GOTO(out_unlock, 0);
1022 /* we need to increase the window sometimes. we'll arbitrarily
1023 * do it half-way through the pages in an rpc */
1024 if ((index & (PTLRPC_MAX_BRW_PAGES - 1)) ==
1025 (PTLRPC_MAX_BRW_PAGES >> 1)) {
1026 ras->ras_window_len += PTLRPC_MAX_BRW_PAGES;
1027 ras->ras_window_len = min(ras->ras_window_len,
1035 spin_unlock(&ras->ras_lock);
1036 spin_unlock(&sbi->ll_lock);
1040 * for now we do our readpage the same on both 2.4 and 2.5. The kernel's
1041 * read-ahead assumes it is valid to issue readpage all the way up to
1042 * i_size, but our dlm locks make that not the case. We disable the
1043 * kernel's read-ahead and do our own by walking ahead in the page cache
1044 * checking for dlm lock coverage. the main difference between 2.4 and
1045 * 2.6 is how read-ahead gets batched and issued, but we're using our own,
1046 * so they look the same.
1048 int ll_readpage(struct file *filp, struct page *page)
1050 struct ll_file_data *fd = filp->private_data;
1051 struct inode *inode = page->mapping->host;
1052 struct obd_export *exp;
1053 struct ll_async_page *llap;
1054 struct obd_io_group *oig = NULL;
1058 LASSERT(PageLocked(page));
1059 LASSERT(!PageUptodate(page));
1060 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset="LPX64"\n",
1061 inode->i_ino, inode->i_generation, inode,
1062 (((obd_off)page->index) << PAGE_SHIFT));
1063 LASSERT(atomic_read(&filp->f_dentry->d_inode->i_count) > 0);
1065 rc = oig_init(&oig);
1069 exp = ll_i2dtexp(inode);
1071 GOTO(out, rc = -EINVAL);
1073 llap = llap_from_page(page, LLAP_ORIGIN_READPAGE);
1075 GOTO(out, rc = PTR_ERR(llap));
1077 /* capability need this */
1078 llap->llap_fsuid = current->fsuid;
1080 if (ll_i2sbi(inode)->ll_flags & LL_SBI_READAHEAD)
1081 ras_update(ll_i2sbi(inode), &fd->fd_ras, page->index,
1082 llap->llap_defer_uptodate);
1084 if (llap->llap_defer_uptodate) {
1085 llap->llap_ra_used = 1;
1086 rc = ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1089 obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd,
1091 LL_CDEBUG_PAGE(D_PAGE, page, "marking uptodate from defer\n");
1092 SetPageUptodate(page);
1094 GOTO(out_oig, rc = 0);
1097 rc = ll_page_matches(page, fd->fd_flags, 0);
1099 LL_CDEBUG_PAGE(D_ERROR, page, "lock match failed: rc %d\n", rc);
1104 CWARN("ino %lu page %lu (%llu) not covered by "
1105 "a lock (mmap?). check debug logs.\n",
1106 inode->i_ino, page->index,
1107 (long long)page->index << PAGE_CACHE_SHIFT);
1110 rc = ll_issue_page_read(exp, llap, oig, 0);
1114 LL_CDEBUG_PAGE(D_PAGE, page, "queued readpage\n");
1115 if (ll_i2sbi(inode)->ll_flags & LL_SBI_READAHEAD)
1116 ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1119 rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);