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.
23 #ifndef AUTOCONF_INCLUDED
24 #include <linux/config.h>
26 #include <linux/kernel.h>
28 #include <linux/string.h>
29 #include <linux/stat.h>
30 #include <linux/errno.h>
31 #include <linux/smp_lock.h>
32 #include <linux/unistd.h>
33 #include <linux/version.h>
34 #include <asm/system.h>
35 #include <asm/uaccess.h>
38 #include <linux/stat.h>
39 #include <asm/uaccess.h>
41 #include <linux/pagemap.h>
42 #include <linux/smp_lock.h>
44 #define DEBUG_SUBSYSTEM S_LLITE
46 #include <lustre_lite.h>
47 #include "llite_internal.h"
48 #include <linux/lustre_compat25.h>
50 #ifndef list_for_each_prev_safe
51 #define list_for_each_prev_safe(pos, n, head) \
52 for (pos = (head)->prev, n = pos->prev; pos != (head); \
53 pos = n, n = pos->prev )
56 cfs_mem_cache_t *ll_async_page_slab = NULL;
57 size_t ll_async_page_slab_size = 0;
59 /* SYNCHRONOUS I/O to object storage for an inode */
60 static int ll_brw(int cmd, struct inode *inode, struct obdo *oa,
61 struct page *page, int flags)
63 struct ll_inode_info *lli = ll_i2info(inode);
64 struct lov_stripe_md *lsm = lli->lli_smd;
65 struct obd_info oinfo = { { { 0 } } };
71 pg.off = ((obd_off)page->index) << CFS_PAGE_SHIFT;
73 if ((cmd & OBD_BRW_WRITE) && (pg.off+CFS_PAGE_SIZE>i_size_read(inode)))
74 pg.count = i_size_read(inode) % CFS_PAGE_SIZE;
76 pg.count = CFS_PAGE_SIZE;
78 LL_CDEBUG_PAGE(D_PAGE, page, "%s %d bytes ino %lu at "LPU64"/"LPX64"\n",
79 cmd & OBD_BRW_WRITE ? "write" : "read", pg.count,
80 inode->i_ino, pg.off, pg.off);
82 CERROR("ZERO COUNT: ino %lu: size %p:%Lu(%p:%Lu) idx %lu off "
83 LPU64"\n", inode->i_ino, inode, i_size_read(inode),
84 page->mapping->host, i_size_read(page->mapping->host),
90 if (cmd & OBD_BRW_WRITE)
91 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_BRW_WRITE,
94 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_BRW_READ,
98 rc = obd_brw(cmd, ll_i2obdexp(inode), &oinfo, 1, &pg, NULL);
100 obdo_to_inode(inode, oa, OBD_MD_FLBLOCKS);
102 CERROR("error from obd_brw: rc = %d\n", rc);
106 int ll_file_punch(struct inode * inode, loff_t new_size, int srvlock)
108 struct ll_inode_info *lli = ll_i2info(inode);
109 struct obd_info oinfo = { { { 0 } } };
114 CDEBUG(D_INFO, "calling punch for "LPX64" (new size %Lu=%#Lx)\n",
115 lli->lli_smd->lsm_object_id, new_size, new_size);
117 oinfo.oi_md = lli->lli_smd;
118 oinfo.oi_policy.l_extent.start = new_size;
119 oinfo.oi_policy.l_extent.end = OBD_OBJECT_EOF;
121 oa.o_id = lli->lli_smd->lsm_object_id;
122 oa.o_valid = OBD_MD_FLID;
124 /* set OBD_MD_FLFLAGS in o_valid, only if we
125 * set OBD_FL_TRUNCLOCK, otherwise ost_punch
126 * and filter_setattr get confused, see the comment
128 oa.o_flags = OBD_FL_TRUNCLOCK;
129 oa.o_valid |= OBD_MD_FLFLAGS;
131 obdo_from_inode(&oa, inode, OBD_MD_FLTYPE | OBD_MD_FLMODE |OBD_MD_FLFID|
132 OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME |
133 OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLGENER |
135 rc = obd_punch_rqset(ll_i2obdexp(inode), &oinfo, NULL);
137 CERROR("obd_truncate fails (%d) ino %lu\n", rc, inode->i_ino);
140 obdo_to_inode(inode, &oa, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
141 OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME);
144 /* this isn't where truncate starts. roughly:
145 * sys_truncate->ll_setattr_raw->vmtruncate->ll_truncate. setattr_raw grabs
146 * DLM lock on [size, EOF], i_mutex, ->lli_size_sem, and WRITE_I_ALLOC_SEM to
149 * must be called under ->lli_size_sem */
150 void ll_truncate(struct inode *inode)
152 struct ll_inode_info *lli = ll_i2info(inode);
153 int srvlock = test_bit(LLI_F_SRVLOCK, &lli->lli_flags);
156 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) to %Lu=%#Lx\n",inode->i_ino,
157 inode->i_generation, inode, i_size_read(inode), i_size_read(inode));
159 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_TRUNC, 1);
160 if (lli->lli_size_sem_owner != current) {
166 CDEBUG(D_INODE, "truncate on inode %lu with no objects\n",
171 LASSERT(atomic_read(&lli->lli_size_sem.count) <= 0);
177 /* XXX I'm pretty sure this is a hack to paper over a more fundamental
179 lov_stripe_lock(lli->lli_smd);
180 inode_init_lvb(inode, &lvb);
181 rc = obd_merge_lvb(ll_i2obdexp(inode), lli->lli_smd, &lvb, 0);
182 inode->i_blocks = lvb.lvb_blocks;
183 if (lvb.lvb_size == i_size_read(inode) && rc == 0) {
184 CDEBUG(D_VFSTRACE, "skipping punch for obj "LPX64", %Lu=%#Lx\n",
185 lli->lli_smd->lsm_object_id, i_size_read(inode),
187 lov_stripe_unlock(lli->lli_smd);
191 obd_adjust_kms(ll_i2obdexp(inode), lli->lli_smd,
192 i_size_read(inode), 1);
193 lov_stripe_unlock(lli->lli_smd);
196 if (unlikely((ll_i2sbi(inode)->ll_flags & LL_SBI_LLITE_CHECKSUM) &&
197 (i_size_read(inode) & ~CFS_PAGE_MASK))) {
198 /* If the truncate leaves a partial page, update its checksum */
199 struct page *page = find_get_page(inode->i_mapping,
200 i_size_read(inode) >>
203 struct ll_async_page *llap = llap_cast_private(page);
205 char *kaddr = kmap_atomic(page, KM_USER0);
206 llap->llap_checksum =
207 init_checksum(OSC_DEFAULT_CKSUM);
208 llap->llap_checksum =
209 compute_checksum(llap->llap_checksum,
210 kaddr, CFS_PAGE_SIZE,
212 kunmap_atomic(kaddr, KM_USER0);
214 page_cache_release(page);
218 new_size = i_size_read(inode);
219 ll_inode_size_unlock(inode, 0);
221 ll_file_punch(inode, new_size, 0);
223 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_LOCKLESS_TRUNC, 1);
229 ll_inode_size_unlock(inode, 0);
232 int ll_prepare_write(struct file *file, struct page *page, unsigned from,
235 struct inode *inode = page->mapping->host;
236 struct ll_inode_info *lli = ll_i2info(inode);
237 struct lov_stripe_md *lsm = lli->lli_smd;
238 obd_off offset = ((obd_off)page->index) << CFS_PAGE_SHIFT;
239 struct obd_info oinfo = { { { 0 } } };
246 LASSERT(PageLocked(page));
247 (void)llap_cast_private(page); /* assertion */
249 /* Check to see if we should return -EIO right away */
252 pga.count = CFS_PAGE_SIZE;
255 oa.o_mode = inode->i_mode;
256 oa.o_id = lsm->lsm_object_id;
257 oa.o_valid = OBD_MD_FLID | OBD_MD_FLMODE | OBD_MD_FLTYPE;
258 obdo_from_inode(&oa, inode, OBD_MD_FLFID | OBD_MD_FLGENER);
262 rc = obd_brw(OBD_BRW_CHECK, ll_i2obdexp(inode), &oinfo, 1, &pga, NULL);
266 if (PageUptodate(page)) {
267 LL_CDEBUG_PAGE(D_PAGE, page, "uptodate\n");
271 /* We're completely overwriting an existing page, so _don't_ set it up
272 * to date until commit_write */
273 if (from == 0 && to == CFS_PAGE_SIZE) {
274 LL_CDEBUG_PAGE(D_PAGE, page, "full page write\n");
275 POISON_PAGE(page, 0x11);
279 /* If are writing to a new page, no need to read old data. The extent
280 * locking will have updated the KMS, and for our purposes here we can
281 * treat it like i_size. */
282 lov_stripe_lock(lsm);
283 inode_init_lvb(inode, &lvb);
284 obd_merge_lvb(ll_i2obdexp(inode), lsm, &lvb, 1);
285 lov_stripe_unlock(lsm);
286 if (lvb.lvb_size <= offset) {
287 char *kaddr = kmap_atomic(page, KM_USER0);
288 LL_CDEBUG_PAGE(D_PAGE, page, "kms "LPU64" <= offset "LPU64"\n",
289 lvb.lvb_size, offset);
290 memset(kaddr, 0, CFS_PAGE_SIZE);
291 kunmap_atomic(kaddr, KM_USER0);
292 GOTO(prepare_done, rc = 0);
295 /* XXX could be an async ocp read.. read-ahead? */
296 rc = ll_brw(OBD_BRW_READ, inode, &oa, page, 0);
298 /* bug 1598: don't clobber blksize */
299 oa.o_valid &= ~(OBD_MD_FLSIZE | OBD_MD_FLBLKSZ);
300 obdo_refresh_inode(inode, &oa, oa.o_valid);
306 SetPageUptodate(page);
311 static int ll_ap_make_ready(void *data, int cmd)
313 struct ll_async_page *llap;
317 llap = LLAP_FROM_COOKIE(data);
318 page = llap->llap_page;
320 LASSERTF(!(cmd & OBD_BRW_READ), "cmd %x page %p ino %lu index %lu\n", cmd, page,
321 page->mapping->host->i_ino, page->index);
323 /* we're trying to write, but the page is locked.. come back later */
324 if (TryLockPage(page))
327 LASSERT(!PageWriteback(page));
329 /* if we left PageDirty we might get another writepage call
330 * in the future. list walkers are bright enough
331 * to check page dirty so we can leave it on whatever list
332 * its on. XXX also, we're called with the cli list so if
333 * we got the page cache list we'd create a lock inversion
334 * with the removepage path which gets the page lock then the
336 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,0)
337 clear_page_dirty(page);
339 LASSERTF(!PageWriteback(page),"cmd %x page %p ino %lu index %lu\n", cmd, page,
340 page->mapping->host->i_ino, page->index);
341 clear_page_dirty_for_io(page);
343 /* This actually clears the dirty bit in the radix tree.*/
344 set_page_writeback(page);
347 LL_CDEBUG_PAGE(D_PAGE, page, "made ready\n");
348 page_cache_get(page);
353 /* We have two reasons for giving llite the opportunity to change the
354 * write length of a given queued page as it builds the RPC containing
357 * 1) Further extending writes may have landed in the page cache
358 * since a partial write first queued this page requiring us
359 * to write more from the page cache. (No further races are possible, since
360 * by the time this is called, the page is locked.)
361 * 2) We might have raced with truncate and want to avoid performing
362 * write RPCs that are just going to be thrown away by the
363 * truncate's punch on the storage targets.
365 * The kms serves these purposes as it is set at both truncate and extending
368 static int ll_ap_refresh_count(void *data, int cmd)
370 struct ll_inode_info *lli;
371 struct ll_async_page *llap;
372 struct lov_stripe_md *lsm;
379 /* readpage queues with _COUNT_STABLE, shouldn't get here. */
380 LASSERT(cmd != OBD_BRW_READ);
382 llap = LLAP_FROM_COOKIE(data);
383 page = llap->llap_page;
384 inode = page->mapping->host;
385 lli = ll_i2info(inode);
388 lov_stripe_lock(lsm);
389 inode_init_lvb(inode, &lvb);
390 obd_merge_lvb(ll_i2obdexp(inode), lsm, &lvb, 1);
392 lov_stripe_unlock(lsm);
394 /* catch race with truncate */
395 if (((__u64)page->index << CFS_PAGE_SHIFT) >= kms)
398 /* catch sub-page write at end of file */
399 if (((__u64)page->index << CFS_PAGE_SHIFT) + CFS_PAGE_SIZE > kms)
400 return kms % CFS_PAGE_SIZE;
402 return CFS_PAGE_SIZE;
405 void ll_inode_fill_obdo(struct inode *inode, int cmd, struct obdo *oa)
407 struct lov_stripe_md *lsm;
408 obd_flag valid_flags;
410 lsm = ll_i2info(inode)->lli_smd;
412 oa->o_id = lsm->lsm_object_id;
413 oa->o_valid = OBD_MD_FLID;
414 valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME;
415 if (cmd & OBD_BRW_WRITE) {
416 oa->o_valid |= OBD_MD_FLEPOCH;
417 oa->o_easize = ll_i2info(inode)->lli_io_epoch;
419 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME |
420 OBD_MD_FLUID | OBD_MD_FLGID |
421 OBD_MD_FLFID | OBD_MD_FLGENER;
424 obdo_from_inode(oa, inode, valid_flags);
427 static void ll_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
429 struct ll_async_page *llap;
432 llap = LLAP_FROM_COOKIE(data);
433 ll_inode_fill_obdo(llap->llap_page->mapping->host, cmd, oa);
438 static void ll_ap_update_obdo(void *data, int cmd, struct obdo *oa,
441 struct ll_async_page *llap;
444 llap = LLAP_FROM_COOKIE(data);
445 obdo_from_inode(oa, llap->llap_page->mapping->host, valid);
450 static struct obd_async_page_ops ll_async_page_ops = {
451 .ap_make_ready = ll_ap_make_ready,
452 .ap_refresh_count = ll_ap_refresh_count,
453 .ap_fill_obdo = ll_ap_fill_obdo,
454 .ap_update_obdo = ll_ap_update_obdo,
455 .ap_completion = ll_ap_completion,
458 struct ll_async_page *llap_cast_private(struct page *page)
460 struct ll_async_page *llap = (struct ll_async_page *)page_private(page);
462 LASSERTF(llap == NULL || llap->llap_magic == LLAP_MAGIC,
463 "page %p private %lu gave magic %d which != %d\n",
464 page, page_private(page), llap->llap_magic, LLAP_MAGIC);
469 /* Try to shrink the page cache for the @sbi filesystem by 1/@shrink_fraction.
471 * There is an llap attached onto every page in lustre, linked off @sbi.
472 * We add an llap to the list so we don't lose our place during list walking.
473 * If llaps in the list are being moved they will only move to the end
474 * of the LRU, and we aren't terribly interested in those pages here (we
475 * start at the beginning of the list where the least-used llaps are.
477 int llap_shrink_cache(struct ll_sb_info *sbi, int shrink_fraction)
479 struct ll_async_page *llap, dummy_llap = { .llap_magic = 0xd11ad11a };
480 unsigned long total, want, count = 0;
482 total = sbi->ll_async_page_count;
484 /* There can be a large number of llaps (600k or more in a large
485 * memory machine) so the VM 1/6 shrink ratio is likely too much.
486 * Since we are freeing pages also, we don't necessarily want to
487 * shrink so much. Limit to 40MB of pages + llaps per call. */
488 if (shrink_fraction == 0)
489 want = sbi->ll_async_page_count - sbi->ll_async_page_max + 32;
491 want = (total + shrink_fraction - 1) / shrink_fraction;
493 if (want > 40 << (20 - CFS_PAGE_SHIFT))
494 want = 40 << (20 - CFS_PAGE_SHIFT);
496 CDEBUG(D_CACHE, "shrinking %lu of %lu pages (1/%d)\n",
497 want, total, shrink_fraction);
499 spin_lock(&sbi->ll_lock);
500 list_add(&dummy_llap.llap_pglist_item, &sbi->ll_pglist);
502 while (--total >= 0 && count < want) {
506 if (unlikely(need_resched())) {
507 spin_unlock(&sbi->ll_lock);
509 spin_lock(&sbi->ll_lock);
512 llap = llite_pglist_next_llap(sbi,&dummy_llap.llap_pglist_item);
513 list_del_init(&dummy_llap.llap_pglist_item);
517 page = llap->llap_page;
518 LASSERT(page != NULL);
520 list_add(&dummy_llap.llap_pglist_item, &llap->llap_pglist_item);
522 /* Page needs/undergoing IO */
523 if (TryLockPage(page)) {
524 LL_CDEBUG_PAGE(D_PAGE, page, "can't lock\n");
528 keep = (llap->llap_write_queued || PageDirty(page) ||
529 PageWriteback(page) || (!PageUptodate(page) &&
530 llap->llap_origin != LLAP_ORIGIN_READAHEAD));
532 LL_CDEBUG_PAGE(D_PAGE, page,"%s LRU page: %s%s%s%s%s origin %s\n",
533 keep ? "keep" : "drop",
534 llap->llap_write_queued ? "wq " : "",
535 PageDirty(page) ? "pd " : "",
536 PageUptodate(page) ? "" : "!pu ",
537 PageWriteback(page) ? "wb" : "",
538 llap->llap_defer_uptodate ? "" : "!du",
539 llap_origins[llap->llap_origin]);
541 /* If page is dirty or undergoing IO don't discard it */
547 page_cache_get(page);
548 spin_unlock(&sbi->ll_lock);
550 if (page->mapping != NULL) {
551 ll_teardown_mmaps(page->mapping,
552 (__u64)page->index << CFS_PAGE_SHIFT,
553 ((__u64)page->index << CFS_PAGE_SHIFT)|
555 if (!PageDirty(page) && !page_mapped(page)) {
556 ll_ra_accounting(llap, page->mapping);
557 ll_truncate_complete_page(page);
560 LL_CDEBUG_PAGE(D_PAGE, page, "Not dropping page"
568 page_cache_release(page);
570 spin_lock(&sbi->ll_lock);
572 list_del(&dummy_llap.llap_pglist_item);
573 spin_unlock(&sbi->ll_lock);
575 CDEBUG(D_CACHE, "shrank %lu/%lu and left %lu unscanned\n",
581 static struct ll_async_page *llap_from_page_with_lockh(struct page *page,
583 struct lustre_handle *lockh)
585 struct ll_async_page *llap;
586 struct obd_export *exp;
587 struct inode *inode = page->mapping->host;
588 struct ll_sb_info *sbi;
593 static int triggered;
596 LL_CDEBUG_PAGE(D_ERROR, page, "Bug 10047. Wrong anon "
598 libcfs_debug_dumpstack(NULL);
601 RETURN(ERR_PTR(-EINVAL));
603 sbi = ll_i2sbi(inode);
604 LASSERT(ll_async_page_slab);
605 LASSERTF(origin < LLAP__ORIGIN_MAX, "%u\n", origin);
607 llap = llap_cast_private(page);
609 /* move to end of LRU list, except when page is just about to
611 if (origin != LLAP_ORIGIN_REMOVEPAGE) {
612 spin_lock(&sbi->ll_lock);
613 sbi->ll_pglist_gen++;
614 list_del_init(&llap->llap_pglist_item);
615 list_add_tail(&llap->llap_pglist_item, &sbi->ll_pglist);
616 spin_unlock(&sbi->ll_lock);
621 exp = ll_i2obdexp(page->mapping->host);
623 RETURN(ERR_PTR(-EINVAL));
625 /* limit the number of lustre-cached pages */
626 if (sbi->ll_async_page_count >= sbi->ll_async_page_max)
627 llap_shrink_cache(sbi, 0);
629 OBD_SLAB_ALLOC(llap, ll_async_page_slab, CFS_ALLOC_STD,
630 ll_async_page_slab_size);
632 RETURN(ERR_PTR(-ENOMEM));
633 llap->llap_magic = LLAP_MAGIC;
634 llap->llap_cookie = (void *)llap + size_round(sizeof(*llap));
636 /* XXX: for bug 11270 - check for lockless origin here! */
637 if (origin == LLAP_ORIGIN_LOCKLESS_IO)
638 llap->llap_nocache = 1;
640 rc = obd_prep_async_page(exp, ll_i2info(inode)->lli_smd, NULL, page,
641 (obd_off)page->index << CFS_PAGE_SHIFT,
642 &ll_async_page_ops, llap, &llap->llap_cookie,
643 llap->llap_nocache, lockh);
645 OBD_SLAB_FREE(llap, ll_async_page_slab,
646 ll_async_page_slab_size);
650 CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n", llap,
651 page, llap->llap_cookie, (obd_off)page->index << CFS_PAGE_SHIFT);
652 /* also zeroing the PRIVBITS low order bitflags */
653 __set_page_ll_data(page, llap);
654 llap->llap_page = page;
656 spin_lock(&sbi->ll_lock);
657 sbi->ll_pglist_gen++;
658 sbi->ll_async_page_count++;
659 list_add_tail(&llap->llap_pglist_item, &sbi->ll_pglist);
660 spin_unlock(&sbi->ll_lock);
663 if (unlikely(sbi->ll_flags & LL_SBI_LLITE_CHECKSUM)) {
665 char *kaddr = kmap_atomic(page, KM_USER0);
666 csum = init_checksum(OSC_DEFAULT_CKSUM);
667 csum = compute_checksum(csum, kaddr, CFS_PAGE_SIZE,
669 kunmap_atomic(kaddr, KM_USER0);
670 if (origin == LLAP_ORIGIN_READAHEAD ||
671 origin == LLAP_ORIGIN_READPAGE ||
672 origin == LLAP_ORIGIN_LOCKLESS_IO) {
673 llap->llap_checksum = 0;
674 } else if (origin == LLAP_ORIGIN_COMMIT_WRITE ||
675 llap->llap_checksum == 0) {
676 llap->llap_checksum = csum;
677 CDEBUG(D_PAGE, "page %p cksum %x\n", page, csum);
678 } else if (llap->llap_checksum == csum) {
679 /* origin == LLAP_ORIGIN_WRITEPAGE */
680 CDEBUG(D_PAGE, "page %p cksum %x confirmed\n",
683 /* origin == LLAP_ORIGIN_WRITEPAGE */
684 LL_CDEBUG_PAGE(D_ERROR, page, "old cksum %x != new "
685 "%x!\n", llap->llap_checksum, csum);
689 llap->llap_origin = origin;
693 static inline struct ll_async_page *llap_from_page(struct page *page,
696 return llap_from_page_with_lockh(page, origin, NULL);
699 static int queue_or_sync_write(struct obd_export *exp, struct inode *inode,
700 struct ll_async_page *llap,
701 unsigned to, obd_flag async_flags)
703 unsigned long size_index = i_size_read(inode) >> CFS_PAGE_SHIFT;
704 struct obd_io_group *oig;
705 struct ll_sb_info *sbi = ll_i2sbi(inode);
706 int rc, noquot = llap->llap_ignore_quota ? OBD_BRW_NOQUOTA : 0;
709 /* _make_ready only sees llap once we've unlocked the page */
710 llap->llap_write_queued = 1;
711 rc = obd_queue_async_io(exp, ll_i2info(inode)->lli_smd, NULL,
712 llap->llap_cookie, OBD_BRW_WRITE | noquot,
713 0, 0, 0, async_flags);
715 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "write queued\n");
716 llap_write_pending(inode, llap);
720 llap->llap_write_queued = 0;
726 /* make full-page requests if we are not at EOF (bug 4410) */
727 if (to != CFS_PAGE_SIZE && llap->llap_page->index < size_index) {
728 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
729 "sync write before EOF: size_index %lu, to %d\n",
732 } else if (to != CFS_PAGE_SIZE && llap->llap_page->index == size_index){
733 int size_to = i_size_read(inode) & ~CFS_PAGE_MASK;
734 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
735 "sync write at EOF: size_index %lu, to %d/%d\n",
736 size_index, to, size_to);
741 /* compare the checksum once before the page leaves llite */
742 if (unlikely((sbi->ll_flags & LL_SBI_LLITE_CHECKSUM) &&
743 llap->llap_checksum != 0)) {
745 struct page *page = llap->llap_page;
746 char *kaddr = kmap_atomic(page, KM_USER0);
747 csum = init_checksum(OSC_DEFAULT_CKSUM);
748 csum = compute_checksum(csum, kaddr, CFS_PAGE_SIZE,
750 kunmap_atomic(kaddr, KM_USER0);
751 if (llap->llap_checksum == csum) {
752 CDEBUG(D_PAGE, "page %p cksum %x confirmed\n",
755 CERROR("page %p old cksum %x != new cksum %x!\n",
756 page, llap->llap_checksum, csum);
760 rc = obd_queue_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig,
761 llap->llap_cookie, OBD_BRW_WRITE | noquot,
762 0, to, 0, ASYNC_READY | ASYNC_URGENT |
763 ASYNC_COUNT_STABLE | ASYNC_GROUP_SYNC);
767 rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
773 if (!rc && async_flags & ASYNC_READY) {
774 unlock_page(llap->llap_page);
775 if (PageWriteback(llap->llap_page)) {
776 end_page_writeback(llap->llap_page);
780 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "sync write returned %d\n", rc);
788 /* update our write count to account for i_size increases that may have
789 * happened since we've queued the page for io. */
791 /* be careful not to return success without setting the page Uptodate or
792 * the next pass through prepare_write will read in stale data from disk. */
793 int ll_commit_write(struct file *file, struct page *page, unsigned from,
796 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
797 struct inode *inode = page->mapping->host;
798 struct ll_inode_info *lli = ll_i2info(inode);
799 struct lov_stripe_md *lsm = lli->lli_smd;
800 struct obd_export *exp;
801 struct ll_async_page *llap;
803 struct lustre_handle *lockh = NULL;
807 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
808 LASSERT(inode == file->f_dentry->d_inode);
809 LASSERT(PageLocked(page));
811 CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
812 inode, page, from, to, page->index);
814 if (fd->fd_flags & LL_FILE_GROUP_LOCKED)
815 lockh = &fd->fd_cwlockh;
817 llap = llap_from_page_with_lockh(page, LLAP_ORIGIN_COMMIT_WRITE, lockh);
819 RETURN(PTR_ERR(llap));
821 exp = ll_i2obdexp(inode);
825 llap->llap_ignore_quota = capable(CAP_SYS_RESOURCE);
827 /* queue a write for some time in the future the first time we
829 if (!PageDirty(page)) {
830 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRTY_MISSES, 1);
832 rc = queue_or_sync_write(exp, inode, llap, to, 0);
836 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRTY_HITS, 1);
839 /* put the page in the page cache, from now on ll_removepage is
840 * responsible for cleaning up the llap.
841 * only set page dirty when it's queued to be write out */
842 if (llap->llap_write_queued)
843 set_page_dirty(page);
846 size = (((obd_off)page->index) << CFS_PAGE_SHIFT) + to;
847 ll_inode_size_lock(inode, 0);
849 lov_stripe_lock(lsm);
850 obd_adjust_kms(exp, lsm, size, 0);
851 lov_stripe_unlock(lsm);
852 if (size > i_size_read(inode))
853 i_size_write(inode, size);
854 SetPageUptodate(page);
855 } else if (size > i_size_read(inode)) {
856 /* this page beyond the pales of i_size, so it can't be
857 * truncated in ll_p_r_e during lock revoking. we must
858 * teardown our book-keeping here. */
861 ll_inode_size_unlock(inode, 0);
865 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, unsigned long len)
867 struct ll_ra_info *ra = &sbi->ll_ra_info;
871 spin_lock(&sbi->ll_lock);
872 ret = min(ra->ra_max_pages - ra->ra_cur_pages, len);
873 ra->ra_cur_pages += ret;
874 spin_unlock(&sbi->ll_lock);
879 static void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
881 struct ll_ra_info *ra = &sbi->ll_ra_info;
882 spin_lock(&sbi->ll_lock);
883 LASSERTF(ra->ra_cur_pages >= len, "r_c_p %lu len %lu\n",
884 ra->ra_cur_pages, len);
885 ra->ra_cur_pages -= len;
886 spin_unlock(&sbi->ll_lock);
889 /* called for each page in a completed rpc.*/
890 int ll_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
892 struct ll_async_page *llap;
897 llap = LLAP_FROM_COOKIE(data);
898 page = llap->llap_page;
899 LASSERT(PageLocked(page));
900 LASSERT(CheckWriteback(page,cmd));
902 LL_CDEBUG_PAGE(D_PAGE, page, "completing cmd %d with %d\n", cmd, rc);
904 if (cmd & OBD_BRW_READ && llap->llap_defer_uptodate)
905 ll_ra_count_put(ll_i2sbi(page->mapping->host), 1);
908 if (cmd & OBD_BRW_READ) {
909 if (!llap->llap_defer_uptodate)
910 SetPageUptodate(page);
912 llap->llap_write_queued = 0;
914 ClearPageError(page);
916 if (cmd & OBD_BRW_READ) {
917 llap->llap_defer_uptodate = 0;
920 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
922 set_bit(AS_ENOSPC, &page->mapping->flags);
924 set_bit(AS_EIO, &page->mapping->flags);
926 page->mapping->gfp_mask |= AS_EIO_MASK;
932 if (cmd & OBD_BRW_WRITE) {
933 llap_write_complete(page->mapping->host, llap);
934 ll_try_done_writing(page->mapping->host);
937 if (PageWriteback(page)) {
938 end_page_writeback(page);
940 page_cache_release(page);
945 static void __ll_put_llap(struct page *page)
947 struct inode *inode = page->mapping->host;
948 struct obd_export *exp;
949 struct ll_async_page *llap;
950 struct ll_sb_info *sbi = ll_i2sbi(inode);
954 exp = ll_i2obdexp(inode);
956 CERROR("page %p ind %lu gave null export\n", page, page->index);
961 llap = llap_from_page(page, LLAP_ORIGIN_REMOVEPAGE);
963 CERROR("page %p ind %lu couldn't find llap: %ld\n", page,
964 page->index, PTR_ERR(llap));
969 //llap_write_complete(inode, llap);
970 rc = obd_teardown_async_page(exp, ll_i2info(inode)->lli_smd, NULL,
973 CERROR("page %p ind %lu failed: %d\n", page, page->index, rc);
975 /* this unconditional free is only safe because the page lock
976 * is providing exclusivity to memory pressure/truncate/writeback..*/
977 __clear_page_ll_data(page);
979 spin_lock(&sbi->ll_lock);
980 if (!list_empty(&llap->llap_pglist_item))
981 list_del_init(&llap->llap_pglist_item);
982 sbi->ll_pglist_gen++;
983 sbi->ll_async_page_count--;
984 spin_unlock(&sbi->ll_lock);
985 OBD_SLAB_FREE(llap, ll_async_page_slab, ll_async_page_slab_size);
990 /* the kernel calls us here when a page is unhashed from the page cache.
991 * the page will be locked and the kernel is holding a spinlock, so
992 * we need to be careful. we're just tearing down our book-keeping
994 void ll_removepage(struct page *page)
996 struct ll_async_page *llap = llap_cast_private(page);
999 LASSERT(!in_interrupt());
1001 /* sync pages or failed read pages can leave pages in the page
1002 * cache that don't have our data associated with them anymore */
1003 if (page_private(page) == 0) {
1008 LASSERT(!llap->llap_lockless_io_page);
1009 LASSERT(!llap->llap_nocache);
1011 LL_CDEBUG_PAGE(D_PAGE, page, "being evicted\n");
1012 __ll_put_llap(page);
1017 static int ll_issue_page_read(struct obd_export *exp,
1018 struct ll_async_page *llap,
1019 struct obd_io_group *oig, int defer)
1021 struct page *page = llap->llap_page;
1024 page_cache_get(page);
1025 llap->llap_defer_uptodate = defer;
1026 llap->llap_ra_used = 0;
1027 rc = obd_queue_group_io(exp, ll_i2info(page->mapping->host)->lli_smd,
1028 NULL, oig, llap->llap_cookie, OBD_BRW_READ, 0,
1029 CFS_PAGE_SIZE, 0, ASYNC_COUNT_STABLE | ASYNC_READY |
1032 LL_CDEBUG_PAGE(D_ERROR, page, "read queue failed: rc %d\n", rc);
1033 page_cache_release(page);
1038 static void ll_ra_stats_inc_unlocked(struct ll_ra_info *ra, enum ra_stat which)
1040 LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
1041 ra->ra_stats[which]++;
1044 static void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which)
1046 struct ll_sb_info *sbi = ll_i2sbi(mapping->host);
1047 struct ll_ra_info *ra = &ll_i2sbi(mapping->host)->ll_ra_info;
1049 spin_lock(&sbi->ll_lock);
1050 ll_ra_stats_inc_unlocked(ra, which);
1051 spin_unlock(&sbi->ll_lock);
1054 void ll_ra_accounting(struct ll_async_page *llap, struct address_space *mapping)
1056 if (!llap->llap_defer_uptodate || llap->llap_ra_used)
1059 ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
1062 #define RAS_CDEBUG(ras) \
1064 "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu r %lu ri %lu" \
1065 "csr %lu sf %lu sp %lu sl %lu \n", \
1066 ras->ras_last_readpage, ras->ras_consecutive_requests, \
1067 ras->ras_consecutive_pages, ras->ras_window_start, \
1068 ras->ras_window_len, ras->ras_next_readahead, \
1069 ras->ras_requests, ras->ras_request_index, \
1070 ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
1071 ras->ras_stride_pages, ras->ras_stride_length)
1073 static int index_in_window(unsigned long index, unsigned long point,
1074 unsigned long before, unsigned long after)
1076 unsigned long start = point - before, end = point + after;
1083 return start <= index && index <= end;
1086 static struct ll_readahead_state *ll_ras_get(struct file *f)
1088 struct ll_file_data *fd;
1090 fd = LUSTRE_FPRIVATE(f);
1094 void ll_ra_read_in(struct file *f, struct ll_ra_read *rar)
1096 struct ll_readahead_state *ras;
1098 ras = ll_ras_get(f);
1100 spin_lock(&ras->ras_lock);
1101 ras->ras_requests++;
1102 ras->ras_request_index = 0;
1103 ras->ras_consecutive_requests++;
1104 rar->lrr_reader = current;
1106 list_add(&rar->lrr_linkage, &ras->ras_read_beads);
1107 spin_unlock(&ras->ras_lock);
1110 void ll_ra_read_ex(struct file *f, struct ll_ra_read *rar)
1112 struct ll_readahead_state *ras;
1114 ras = ll_ras_get(f);
1116 spin_lock(&ras->ras_lock);
1117 list_del_init(&rar->lrr_linkage);
1118 spin_unlock(&ras->ras_lock);
1121 static struct ll_ra_read *ll_ra_read_get_locked(struct ll_readahead_state *ras)
1123 struct ll_ra_read *scan;
1125 list_for_each_entry(scan, &ras->ras_read_beads, lrr_linkage) {
1126 if (scan->lrr_reader == current)
1132 struct ll_ra_read *ll_ra_read_get(struct file *f)
1134 struct ll_readahead_state *ras;
1135 struct ll_ra_read *bead;
1137 ras = ll_ras_get(f);
1139 spin_lock(&ras->ras_lock);
1140 bead = ll_ra_read_get_locked(ras);
1141 spin_unlock(&ras->ras_lock);
1145 static int ll_read_ahead_page(struct obd_export *exp, struct obd_io_group *oig,
1146 int index, struct address_space *mapping)
1148 struct ll_async_page *llap;
1150 unsigned int gfp_mask = 0;
1153 gfp_mask = GFP_HIGHUSER & ~__GFP_WAIT;
1155 gfp_mask |= __GFP_NOWARN;
1157 page = grab_cache_page_nowait_gfp(mapping, index, gfp_mask);
1159 ll_ra_stats_inc(mapping, RA_STAT_FAILED_GRAB_PAGE);
1160 CDEBUG(D_READA, "g_c_p_n failed\n");
1164 /* Check if page was truncated or reclaimed */
1165 if (page->mapping != mapping) {
1166 ll_ra_stats_inc(mapping, RA_STAT_WRONG_GRAB_PAGE);
1167 CDEBUG(D_READA, "g_c_p_n returned invalid page\n");
1168 GOTO(unlock_page, rc = 0);
1171 /* we do this first so that we can see the page in the /proc
1173 llap = llap_from_page(page, LLAP_ORIGIN_READAHEAD);
1174 if (IS_ERR(llap) || llap->llap_defer_uptodate) {
1175 if (PTR_ERR(llap) == -ENOLCK) {
1176 ll_ra_stats_inc(mapping, RA_STAT_FAILED_MATCH);
1177 CDEBUG(D_READA | D_PAGE,
1178 "Adding page to cache failed index "
1180 CDEBUG(D_READA, "nolock page\n");
1181 GOTO(unlock_page, rc = -ENOLCK);
1183 CDEBUG(D_READA, "read-ahead page\n");
1184 GOTO(unlock_page, rc = 0);
1187 /* skip completed pages */
1188 if (Page_Uptodate(page))
1189 GOTO(unlock_page, rc = 0);
1191 /* bail out when we hit the end of the lock. */
1192 rc = ll_issue_page_read(exp, llap, oig, 1);
1194 LL_CDEBUG_PAGE(D_READA | D_PAGE, page, "started read-ahead\n");
1199 LL_CDEBUG_PAGE(D_READA | D_PAGE, page, "skipping read-ahead\n");
1201 page_cache_release(page);
1205 /* ra_io_arg will be filled in the beginning of ll_readahead with
1206 * ras_lock, then the following ll_read_ahead_pages will read RA
1207 * pages according to this arg, all the items in this structure are
1208 * counted by page index.
1211 unsigned long ria_start; /* start offset of read-ahead*/
1212 unsigned long ria_end; /* end offset of read-ahead*/
1213 /* If stride read pattern is detected, ria_stoff means where
1214 * stride read is started. Note: for normal read-ahead, the
1215 * value here is meaningless, and also it will not be accessed*/
1217 /* ria_length and ria_pages are the length and pages length in the
1218 * stride I/O mode. And they will also be used to check whether
1219 * it is stride I/O read-ahead in the read-ahead pages*/
1220 unsigned long ria_length;
1221 unsigned long ria_pages;
1224 #define RIA_DEBUG(ria) \
1225 CDEBUG(D_READA, "rs %lu re %lu ro %lu rl %lu rp %lu\n", \
1226 ria->ria_start, ria->ria_end, ria->ria_stoff, ria->ria_length,\
1229 #define RAS_INCREASE_STEP (1024 * 1024 >> CFS_PAGE_SHIFT)
1231 static inline int stride_io_mode(struct ll_readahead_state *ras)
1233 return ras->ras_consecutive_stride_requests > 1;
1236 /* The function calculates how much pages will be read in
1237 * [off, off + length], which will be read by stride I/O mode,
1238 * stride_offset = st_off, stride_lengh = st_len,
1239 * stride_pages = st_pgs
1241 static unsigned long
1242 stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs,
1243 unsigned long off, unsigned length)
1245 unsigned long cont_len = st_off > off ? st_off - off : 0;
1246 unsigned long stride_len = length + off > st_off ?
1247 length + off + 1 - st_off : 0;
1248 unsigned long left, pg_count;
1250 if (st_len == 0 || length == 0)
1253 left = do_div(stride_len, st_len);
1254 left = min(left, st_pgs);
1256 pg_count = left + stride_len * st_pgs + cont_len;
1258 LASSERT(pg_count >= left);
1260 CDEBUG(D_READA, "st_off %lu, st_len %lu st_pgs %lu off %lu length %u"
1261 "pgcount %lu\n", st_off, st_len, st_pgs, off, length, pg_count);
1266 static int ria_page_count(struct ra_io_arg *ria)
1268 __u64 length = ria->ria_end >= ria->ria_start ?
1269 ria->ria_end - ria->ria_start + 1 : 0;
1271 return stride_pg_count(ria->ria_stoff, ria->ria_length,
1272 ria->ria_pages, ria->ria_start,
1276 /*Check whether the index is in the defined ra-window */
1277 static int ras_inside_ra_window(unsigned long idx, struct ra_io_arg *ria)
1279 /* If ria_length == ria_pages, it means non-stride I/O mode,
1280 * idx should always inside read-ahead window in this case
1281 * For stride I/O mode, just check whether the idx is inside
1283 return ria->ria_length == 0 || ria->ria_length == ria->ria_pages ||
1284 (idx - ria->ria_stoff) % ria->ria_length < ria->ria_pages;
1287 static int ll_read_ahead_pages(struct obd_export *exp,
1288 struct obd_io_group *oig,
1289 struct ra_io_arg *ria,
1290 unsigned long *reserved_pages,
1291 struct address_space *mapping,
1292 unsigned long *ra_end)
1294 int rc, count = 0, stride_ria;
1295 unsigned long page_idx;
1297 LASSERT(ria != NULL);
1300 stride_ria = ria->ria_length > ria->ria_pages && ria->ria_pages > 0;
1301 for (page_idx = ria->ria_start; page_idx <= ria->ria_end &&
1302 *reserved_pages > 0; page_idx++) {
1303 if (ras_inside_ra_window(page_idx, ria)) {
1304 /* If the page is inside the read-ahead window*/
1305 rc = ll_read_ahead_page(exp, oig, page_idx, mapping);
1307 (*reserved_pages)--;
1309 } else if (rc == -ENOLCK)
1311 } else if (stride_ria) {
1312 /* If it is not in the read-ahead window, and it is
1313 * read-ahead mode, then check whether it should skip
1316 /* FIXME: This assertion only is valid when it is for
1317 * forward read-ahead, it will be fixed when backward
1318 * read-ahead is implemented */
1319 LASSERTF(page_idx > ria->ria_stoff, "since %lu in the"
1320 " gap of ra window,it should bigger than stride"
1321 " offset %lu \n", page_idx, ria->ria_stoff);
1323 offset = page_idx - ria->ria_stoff;
1324 offset = offset % (ria->ria_length);
1325 if (offset > ria->ria_pages) {
1326 page_idx += ria->ria_length - offset;
1327 CDEBUG(D_READA, "i %lu skip %lu \n", page_idx,
1328 ria->ria_length - offset);
1337 static int ll_readahead(struct ll_readahead_state *ras,
1338 struct obd_export *exp, struct address_space *mapping,
1339 struct obd_io_group *oig, int flags)
1341 unsigned long start = 0, end = 0, reserved;
1342 unsigned long ra_end, len;
1343 struct inode *inode;
1344 struct lov_stripe_md *lsm;
1345 struct ll_ra_read *bead;
1347 struct ra_io_arg ria = { 0 };
1352 inode = mapping->host;
1353 lsm = ll_i2info(inode)->lli_smd;
1355 lov_stripe_lock(lsm);
1356 inode_init_lvb(inode, &lvb);
1357 obd_merge_lvb(ll_i2obdexp(inode), lsm, &lvb, 1);
1359 lov_stripe_unlock(lsm);
1361 ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN);
1365 spin_lock(&ras->ras_lock);
1366 bead = ll_ra_read_get_locked(ras);
1367 /* Enlarge the RA window to encompass the full read */
1368 if (bead != NULL && ras->ras_window_start + ras->ras_window_len <
1369 bead->lrr_start + bead->lrr_count) {
1370 ras->ras_window_len = bead->lrr_start + bead->lrr_count -
1371 ras->ras_window_start;
1373 /* Reserve a part of the read-ahead window that we'll be issuing */
1374 if (ras->ras_window_len) {
1375 start = ras->ras_next_readahead;
1376 end = ras->ras_window_start + ras->ras_window_len - 1;
1379 /* Truncate RA window to end of file */
1380 end = min(end, (unsigned long)((kms - 1) >> CFS_PAGE_SHIFT));
1381 ras->ras_next_readahead = max(end, end + 1);
1384 ria.ria_start = start;
1386 /* If stride I/O mode is detected, get stride window*/
1387 if (stride_io_mode(ras)) {
1388 ria.ria_stoff = ras->ras_stride_offset;
1389 ria.ria_length = ras->ras_stride_length;
1390 ria.ria_pages = ras->ras_stride_pages;
1392 spin_unlock(&ras->ras_lock);
1395 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
1399 len = ria_page_count(&ria);
1403 reserved = ll_ra_count_get(ll_i2sbi(inode), len);
1404 if (reserved < end - start + 1)
1405 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
1407 CDEBUG(D_READA, "reserved page %lu \n", reserved);
1409 ret = ll_read_ahead_pages(exp, oig, &ria, &reserved, mapping, &ra_end);
1411 LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
1413 ll_ra_count_put(ll_i2sbi(inode), reserved);
1415 if (ra_end == end + 1 && ra_end == (kms >> CFS_PAGE_SHIFT))
1416 ll_ra_stats_inc(mapping, RA_STAT_EOF);
1418 /* if we didn't get to the end of the region we reserved from
1419 * the ras we need to go back and update the ras so that the
1420 * next read-ahead tries from where we left off. we only do so
1421 * if the region we failed to issue read-ahead on is still ahead
1422 * of the app and behind the next index to start read-ahead from */
1423 CDEBUG(D_READA, "ra_end %lu end %lu stride end %lu \n",
1424 ra_end, end, ria.ria_end);
1426 if (ra_end != (end + 1)) {
1427 spin_lock(&ras->ras_lock);
1428 if (ra_end < ras->ras_next_readahead &&
1429 index_in_window(ra_end, ras->ras_window_start, 0,
1430 ras->ras_window_len)) {
1431 ras->ras_next_readahead = ra_end;
1434 spin_unlock(&ras->ras_lock);
1440 static void ras_set_start(struct ll_readahead_state *ras, unsigned long index)
1442 ras->ras_window_start = index & (~(RAS_INCREASE_STEP - 1));
1445 /* called with the ras_lock held or from places where it doesn't matter */
1446 static void ras_reset(struct ll_readahead_state *ras, unsigned long index)
1448 ras->ras_last_readpage = index;
1449 ras->ras_consecutive_requests = 0;
1450 ras->ras_consecutive_pages = 0;
1451 ras->ras_window_len = 0;
1452 ras_set_start(ras, index);
1453 ras->ras_next_readahead = max(ras->ras_window_start, index);
1458 /* called with the ras_lock held or from places where it doesn't matter */
1459 static void ras_stride_reset(struct ll_readahead_state *ras)
1461 ras->ras_consecutive_stride_requests = 0;
1465 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
1467 spin_lock_init(&ras->ras_lock);
1469 ras->ras_requests = 0;
1470 INIT_LIST_HEAD(&ras->ras_read_beads);
1473 /* Check whether the read request is in the stride window.
1474 * If it is in the stride window, return 1, otherwise return 0.
1475 * and also update stride_gap and stride_pages.
1477 static int index_in_stride_window(unsigned long index,
1478 struct ll_readahead_state *ras,
1479 struct inode *inode)
1481 int stride_gap = index - ras->ras_last_readpage - 1;
1483 LASSERT(stride_gap != 0);
1485 if (ras->ras_consecutive_pages == 0)
1488 /*Otherwise check the stride by itself */
1489 if ((ras->ras_stride_length - ras->ras_stride_pages) == stride_gap &&
1490 ras->ras_consecutive_pages == ras->ras_stride_pages)
1493 if (stride_gap >= 0) {
1495 * only set stride_pages, stride_length if
1496 * it is forward reading ( stride_gap > 0)
1498 ras->ras_stride_pages = ras->ras_consecutive_pages;
1499 ras->ras_stride_length = stride_gap + ras->ras_consecutive_pages;
1502 * If stride_gap < 0,(back_forward reading),
1503 * reset the stride_pages/length.
1504 * FIXME:back_ward stride I/O read.
1507 ras->ras_stride_pages = 0;
1508 ras->ras_stride_length = 0;
1515 static unsigned long
1516 stride_page_count(struct ll_readahead_state *ras, unsigned long len)
1518 return stride_pg_count(ras->ras_stride_offset, ras->ras_stride_length,
1519 ras->ras_stride_pages, ras->ras_stride_offset,
1523 /* Stride Read-ahead window will be increased inc_len according to
1524 * stride I/O pattern */
1525 static void ras_stride_increase_window(struct ll_readahead_state *ras,
1526 struct ll_ra_info *ra,
1527 unsigned long inc_len)
1529 unsigned long left, step, window_len;
1530 unsigned long stride_len;
1532 LASSERT(ras->ras_stride_length > 0);
1534 stride_len = ras->ras_window_start + ras->ras_window_len -
1535 ras->ras_stride_offset;
1537 LASSERTF(stride_len >= 0, "window_start %lu, window_len %lu"
1538 " stride_offset %lu\n", ras->ras_window_start,
1539 ras->ras_window_len, ras->ras_stride_offset);
1541 left = stride_len % ras->ras_stride_length;
1543 window_len = ras->ras_window_len - left;
1545 if (left < ras->ras_stride_pages)
1548 left = ras->ras_stride_pages + inc_len;
1550 LASSERT(ras->ras_stride_pages != 0);
1552 step = left / ras->ras_stride_pages;
1553 left %= ras->ras_stride_pages;
1555 window_len += step * ras->ras_stride_length + left;
1557 if (stride_page_count(ras, window_len) <= ra->ra_max_pages)
1558 ras->ras_window_len = window_len;
1563 /* Set stride I/O read-ahead window start offset */
1564 static void ras_set_stride_offset(struct ll_readahead_state *ras)
1566 unsigned long window_len = ras->ras_next_readahead -
1567 ras->ras_window_start;
1570 LASSERT(ras->ras_stride_length != 0);
1572 left = window_len % ras->ras_stride_length;
1574 ras->ras_stride_offset = ras->ras_next_readahead - left;
1579 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
1580 struct ll_readahead_state *ras, unsigned long index,
1583 struct ll_ra_info *ra = &sbi->ll_ra_info;
1584 int zero = 0, stride_zero = 0, stride_detect = 0, ra_miss = 0;
1587 spin_lock(&sbi->ll_lock);
1588 spin_lock(&ras->ras_lock);
1590 ll_ra_stats_inc_unlocked(ra, hit ? RA_STAT_HIT : RA_STAT_MISS);
1592 /* reset the read-ahead window in two cases. First when the app seeks
1593 * or reads to some other part of the file. Secondly if we get a
1594 * read-ahead miss that we think we've previously issued. This can
1595 * be a symptom of there being so many read-ahead pages that the VM is
1596 * reclaiming it before we get to it. */
1597 if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
1599 ll_ra_stats_inc_unlocked(ra, RA_STAT_DISTANT_READPAGE);
1600 /* check whether it is in stride I/O mode*/
1601 if (!index_in_stride_window(index, ras, inode))
1603 } else if (!hit && ras->ras_window_len &&
1604 index < ras->ras_next_readahead &&
1605 index_in_window(index, ras->ras_window_start, 0,
1606 ras->ras_window_len)) {
1609 /* If it hits read-ahead miss and the stride I/O is still
1610 * not detected, reset stride stuff to re-detect the whole
1611 * stride I/O mode to avoid complication */
1612 if (!stride_io_mode(ras))
1614 ll_ra_stats_inc_unlocked(ra, RA_STAT_MISS_IN_WINDOW);
1617 /* On the second access to a file smaller than the tunable
1618 * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1619 * file up to ra_max_pages. This is simply a best effort and
1620 * only occurs once per open file. Normal RA behavior is reverted
1621 * to for subsequent IO. The mmap case does not increment
1622 * ras_requests and thus can never trigger this behavior. */
1623 if (ras->ras_requests == 2 && !ras->ras_request_index) {
1626 kms_pages = (i_size_read(inode) + CFS_PAGE_SIZE - 1) >>
1629 CDEBUG(D_READA, "kmsp "LPU64" mwp %lu mp %lu\n", kms_pages,
1630 ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages);
1633 kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1634 ras->ras_window_start = 0;
1635 ras->ras_last_readpage = 0;
1636 ras->ras_next_readahead = 0;
1637 ras->ras_window_len = min(ra->ra_max_pages,
1638 ra->ra_max_read_ahead_whole_pages);
1639 GOTO(out_unlock, 0);
1644 /* If it is discontinuous read, check
1645 * whether it is stride I/O mode*/
1647 ras_reset(ras, index);
1648 ras->ras_consecutive_pages++;
1649 ras_stride_reset(ras);
1651 GOTO(out_unlock, 0);
1653 /* The read is still in stride window or
1654 * it hits read-ahead miss */
1656 /* If ra-window miss is hitted, which probably means VM
1657 * pressure, and some read-ahead pages were reclaimed.So
1658 * the length of ra-window will not increased, but also
1659 * not reset to avoid redetecting the stride I/O mode.*/
1660 ras->ras_consecutive_requests = 0;
1662 ras->ras_consecutive_pages = 0;
1663 if (++ras->ras_consecutive_stride_requests > 1)
1668 } else if (ras->ras_consecutive_stride_requests > 1) {
1669 /* If this is contiguous read but in stride I/O mode
1670 * currently, check whether stride step still is valid,
1671 * if invalid, it will reset the stride ra window*/
1672 if (ras->ras_consecutive_pages + 1 > ras->ras_stride_pages)
1673 ras_stride_reset(ras);
1676 ras->ras_last_readpage = index;
1677 ras->ras_consecutive_pages++;
1678 ras_set_start(ras, index);
1679 ras->ras_next_readahead = max(ras->ras_window_start,
1680 ras->ras_next_readahead);
1683 /* Trigger RA in the mmap case where ras_consecutive_requests
1684 * is not incremented and thus can't be used to trigger RA */
1685 if (!ras->ras_window_len && ras->ras_consecutive_pages == 4) {
1686 ras->ras_window_len = RAS_INCREASE_STEP;
1687 GOTO(out_unlock, 0);
1690 /* Initially reset the stride window offset to next_readahead*/
1691 if (ras->ras_consecutive_stride_requests == 2 && stride_detect)
1692 ras_set_stride_offset(ras);
1694 /* The initial ras_window_len is set to the request size. To avoid
1695 * uselessly reading and discarding pages for random IO the window is
1696 * only increased once per consecutive request received. */
1697 if ((ras->ras_consecutive_requests > 1 &&
1698 !ras->ras_request_index) || stride_detect) {
1699 if (stride_io_mode(ras))
1700 ras_stride_increase_window(ras, ra, RAS_INCREASE_STEP);
1702 ras->ras_window_len = min(ras->ras_window_len +
1709 ras->ras_request_index++;
1710 spin_unlock(&ras->ras_lock);
1711 spin_unlock(&sbi->ll_lock);
1715 int ll_writepage(struct page *page)
1717 struct inode *inode = page->mapping->host;
1718 struct ll_inode_info *lli = ll_i2info(inode);
1719 struct obd_export *exp;
1720 struct ll_async_page *llap;
1724 LASSERT(PageLocked(page));
1726 exp = ll_i2obdexp(inode);
1728 GOTO(out, rc = -EINVAL);
1730 llap = llap_from_page(page, LLAP_ORIGIN_WRITEPAGE);
1732 GOTO(out, rc = PTR_ERR(llap));
1734 LASSERT(!llap->llap_nocache);
1735 LASSERT(!PageWriteback(page));
1736 set_page_writeback(page);
1738 page_cache_get(page);
1739 if (llap->llap_write_queued) {
1740 LL_CDEBUG_PAGE(D_PAGE, page, "marking urgent\n");
1741 rc = obd_set_async_flags(exp, lli->lli_smd, NULL,
1743 ASYNC_READY | ASYNC_URGENT);
1745 rc = queue_or_sync_write(exp, inode, llap, CFS_PAGE_SIZE,
1746 ASYNC_READY | ASYNC_URGENT);
1749 page_cache_release(page);
1752 if (!lli->lli_async_rc)
1753 lli->lli_async_rc = rc;
1754 /* re-dirty page on error so it retries write */
1755 if (PageWriteback(page)) {
1756 end_page_writeback(page);
1758 /* resend page only for not started IO*/
1759 if (!PageError(page))
1760 ll_redirty_page(page);
1767 * for now we do our readpage the same on both 2.4 and 2.5. The kernel's
1768 * read-ahead assumes it is valid to issue readpage all the way up to
1769 * i_size, but our dlm locks make that not the case. We disable the
1770 * kernel's read-ahead and do our own by walking ahead in the page cache
1771 * checking for dlm lock coverage. the main difference between 2.4 and
1772 * 2.6 is how read-ahead gets batched and issued, but we're using our own,
1773 * so they look the same.
1775 int ll_readpage(struct file *filp, struct page *page)
1777 struct ll_file_data *fd = LUSTRE_FPRIVATE(filp);
1778 struct inode *inode = page->mapping->host;
1779 struct obd_export *exp;
1780 struct ll_async_page *llap;
1781 struct obd_io_group *oig = NULL;
1782 struct lustre_handle *lockh = NULL;
1786 LASSERT(PageLocked(page));
1787 LASSERT(!PageUptodate(page));
1788 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset=%Lu=%#Lx\n",
1789 inode->i_ino, inode->i_generation, inode,
1790 (((loff_t)page->index) << CFS_PAGE_SHIFT),
1791 (((loff_t)page->index) << CFS_PAGE_SHIFT));
1792 LASSERT(atomic_read(&filp->f_dentry->d_inode->i_count) > 0);
1794 if (!ll_i2info(inode)->lli_smd) {
1795 /* File with no objects - one big hole */
1796 /* We use this just for remove_from_page_cache that is not
1797 * exported, we'd make page back up to date. */
1798 ll_truncate_complete_page(page);
1799 clear_page(kmap(page));
1801 SetPageUptodate(page);
1806 rc = oig_init(&oig);
1810 exp = ll_i2obdexp(inode);
1812 GOTO(out, rc = -EINVAL);
1814 if (fd->fd_flags & LL_FILE_GROUP_LOCKED)
1815 lockh = &fd->fd_cwlockh;
1817 llap = llap_from_page_with_lockh(page, LLAP_ORIGIN_READPAGE, lockh);
1819 if (PTR_ERR(llap) == -ENOLCK) {
1820 CWARN("ino %lu page %lu (%llu) not covered by "
1821 "a lock (mmap?). check debug logs.\n",
1822 inode->i_ino, page->index,
1823 (long long)page->index << PAGE_CACHE_SHIFT);
1825 GOTO(out, rc = PTR_ERR(llap));
1828 if (ll_i2sbi(inode)->ll_ra_info.ra_max_pages)
1829 ras_update(ll_i2sbi(inode), inode, &fd->fd_ras, page->index,
1830 llap->llap_defer_uptodate);
1833 if (llap->llap_defer_uptodate) {
1834 /* This is the callpath if we got the page from a readahead */
1835 llap->llap_ra_used = 1;
1836 rc = ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1839 obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd,
1841 LL_CDEBUG_PAGE(D_PAGE, page, "marking uptodate from defer\n");
1842 SetPageUptodate(page);
1844 GOTO(out_oig, rc = 0);
1847 rc = ll_issue_page_read(exp, llap, oig, 0);
1851 LL_CDEBUG_PAGE(D_PAGE, page, "queued readpage\n");
1852 /* We have just requested the actual page we want, see if we can tack
1853 * on some readahead to that page's RPC before it is sent. */
1854 if (ll_i2sbi(inode)->ll_ra_info.ra_max_pages)
1855 ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1858 rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
1869 static void ll_file_put_pages(struct page **pages, int numpages)
1875 for (i = 0, pp = pages; i < numpages; i++, pp++) {
1877 LL_CDEBUG_PAGE(D_PAGE, (*pp), "free\n");
1879 if (page_private(*pp))
1880 CERROR("the llap wasn't freed\n");
1881 (*pp)->mapping = NULL;
1882 if (page_count(*pp) != 1)
1883 CERROR("page %p, flags %#lx, count %i, private %p\n",
1884 (*pp), (unsigned long)(*pp)->flags, page_count(*pp),
1885 (void*)page_private(*pp));
1886 __free_pages(*pp, 0);
1889 OBD_FREE(pages, numpages * sizeof(struct page*));
1893 static struct page **ll_file_prepare_pages(int numpages, struct inode *inode,
1894 unsigned long first)
1896 struct page **pages;
1901 OBD_ALLOC(pages, sizeof(struct page *) * numpages);
1903 RETURN(ERR_PTR(-ENOMEM));
1904 for (i = 0; i < numpages; i++) {
1906 struct ll_async_page *llap;
1908 page = alloc_pages(GFP_HIGHUSER, 0);
1910 GOTO(err, rc = -ENOMEM);
1912 /* llap_from_page needs page index and mapping to be set */
1913 page->index = first++;
1914 page->mapping = inode->i_mapping;
1915 llap = llap_from_page(page, LLAP_ORIGIN_LOCKLESS_IO);
1917 GOTO(err, rc = PTR_ERR(llap));
1918 llap->llap_lockless_io_page = 1;
1922 ll_file_put_pages(pages, numpages);
1923 RETURN(ERR_PTR(rc));
1926 static ssize_t ll_file_copy_pages(struct page **pages, int numpages,
1927 const struct iovec *iov, unsigned long nsegs,
1928 ssize_t iov_offset, loff_t pos, size_t count,
1933 int updatechecksum = ll_i2sbi(pages[0]->mapping->host)->ll_flags &
1934 LL_SBI_LLITE_CHECKSUM;
1937 for (i = 0; i < numpages; i++) {
1938 unsigned offset, bytes, left = 0;
1941 vaddr = kmap(pages[i]);
1942 offset = pos & (CFS_PAGE_SIZE - 1);
1943 bytes = min_t(unsigned, CFS_PAGE_SIZE - offset, count);
1944 LL_CDEBUG_PAGE(D_PAGE, pages[i], "op = %s, addr = %p, "
1946 (rw == WRITE) ? "CFU" : "CTU",
1947 vaddr + offset, bytes);
1948 while (bytes > 0 && !left && nsegs) {
1949 unsigned copy = min_t(ssize_t, bytes,
1950 iov->iov_len - iov_offset);
1952 left = copy_from_user(vaddr + offset,
1953 iov->iov_base +iov_offset,
1955 if (updatechecksum) {
1956 struct ll_async_page *llap;
1958 llap = llap_cast_private(pages[i]);
1959 llap->llap_checksum =
1960 init_checksum(OSC_DEFAULT_CKSUM);
1961 llap->llap_checksum =
1962 compute_checksum(llap->llap_checksum,
1963 vaddr,CFS_PAGE_SIZE,
1967 left = copy_to_user(iov->iov_base + iov_offset,
1968 vaddr + offset, copy);
1976 if (iov_offset == iov->iov_len) {
1993 static int ll_file_oig_pages(struct inode * inode, struct page **pages,
1994 int numpages, loff_t pos, size_t count, int rw)
1996 struct obd_io_group *oig;
1997 struct ll_inode_info *lli = ll_i2info(inode);
1998 struct obd_export *exp;
1999 loff_t org_pos = pos;
2005 exp = ll_i2obdexp(inode);
2008 rc = oig_init(&oig);
2011 brw_flags = OBD_BRW_SRVLOCK;
2012 if (capable(CAP_SYS_RESOURCE))
2013 brw_flags |= OBD_BRW_NOQUOTA;
2015 for (i = 0; i < numpages; i++) {
2016 struct ll_async_page *llap;
2017 unsigned from, bytes;
2019 from = pos & (CFS_PAGE_SIZE - 1);
2020 bytes = min_t(unsigned, CFS_PAGE_SIZE - from,
2021 count - pos + org_pos);
2022 llap = llap_cast_private(pages[i]);
2025 lock_page(pages[i]);
2027 LL_CDEBUG_PAGE(D_PAGE, pages[i], "offset "LPU64","
2028 " from %u, bytes = %u\n",
2030 LASSERTF(pos >> CFS_PAGE_SHIFT == pages[i]->index,
2031 "wrong page index %lu (%lu)\n",
2033 (unsigned long)(pos >> CFS_PAGE_SHIFT));
2034 rc = obd_queue_group_io(exp, lli->lli_smd, NULL, oig,
2037 OBD_BRW_WRITE:OBD_BRW_READ,
2038 from, bytes, brw_flags,
2039 ASYNC_READY | ASYNC_URGENT |
2040 ASYNC_COUNT_STABLE | ASYNC_GROUP_SYNC);
2047 rc = obd_trigger_group_io(exp, lli->lli_smd, NULL, oig);
2053 unlock_page(pages[i]);
2058 /* Advance through passed iov, adjust iov pointer as necessary and return
2059 * starting offset in individual entry we are pointing at. Also reduce
2060 * nr_segs as needed */
2061 static ssize_t ll_iov_advance(const struct iovec **iov, unsigned long *nr_segs,
2064 while (*nr_segs > 0) {
2065 if ((*iov)->iov_len > offset)
2066 return ((*iov)->iov_len - offset);
2067 offset -= (*iov)->iov_len;
2074 ssize_t ll_file_lockless_io(struct file *file, const struct iovec *iov,
2075 unsigned long nr_segs,
2076 loff_t *ppos, int rw, ssize_t count)
2079 struct inode *inode = file->f_dentry->d_inode;
2083 unsigned long first, last;
2084 const struct iovec *iv = &iov[0];
2085 unsigned long nsegs = nr_segs;
2086 unsigned long offset = 0;
2092 ll_inode_size_lock(inode, 0);
2093 isize = i_size_read(inode);
2094 ll_inode_size_unlock(inode, 0);
2097 if (*ppos + count >= isize)
2098 count -= *ppos + count - isize;
2102 rc = generic_write_checks(file, ppos, &count, 0);
2105 rc = remove_suid(file->f_dentry);
2111 first = pos >> CFS_PAGE_SHIFT;
2112 last = (pos + count - 1) >> CFS_PAGE_SHIFT;
2113 max_pages = PTLRPC_MAX_BRW_PAGES *
2114 ll_i2info(inode)->lli_smd->lsm_stripe_count;
2115 CDEBUG(D_INFO, "%u, stripe_count = %u\n",
2116 PTLRPC_MAX_BRW_PAGES /* max_pages_per_rpc */,
2117 ll_i2info(inode)->lli_smd->lsm_stripe_count);
2119 while (first <= last && rc >= 0) {
2121 struct page **pages;
2122 size_t bytes = count - amount;
2124 pages_for_io = min_t(int, last - first + 1, max_pages);
2125 pages = ll_file_prepare_pages(pages_for_io, inode, first);
2126 if (IS_ERR(pages)) {
2127 rc = PTR_ERR(pages);
2131 rc = ll_file_copy_pages(pages, pages_for_io, iv, nsegs,
2132 offset, pos + amount, bytes,
2135 GOTO(put_pages, rc);
2136 offset = ll_iov_advance(&iv, &nsegs, offset + rc);
2139 rc = ll_file_oig_pages(inode, pages, pages_for_io,
2140 pos + amount, bytes, rw);
2142 GOTO(put_pages, rc);
2144 rc = ll_file_copy_pages(pages, pages_for_io, iv, nsegs,
2145 offset, pos + amount, bytes, rw);
2147 GOTO(put_pages, rc);
2148 offset = ll_iov_advance(&iv, &nsegs, offset + rc);
2153 ll_file_put_pages(pages, pages_for_io);
2154 first += pages_for_io;
2155 /* a short read/write check */
2156 if (pos + amount < ((loff_t)first << CFS_PAGE_SHIFT))
2158 /* Check if we are out of userspace buffers. (how that could
2163 /* NOTE: don't update i_size and KMS in absence of LDLM locks even
2164 * write makes the file large */
2165 file_accessed(file);
2166 if (rw == READ && amount < count && rc == 0) {
2167 unsigned long not_cleared;
2170 ssize_t to_clear = min_t(ssize_t, count - amount,
2171 iv->iov_len - offset);
2172 not_cleared = clear_user(iv->iov_base + offset,
2174 amount += to_clear - not_cleared;
2185 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
2187 LPROC_LL_LOCKLESS_WRITE :
2188 LPROC_LL_LOCKLESS_READ,