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 <asm/segment.h>
44 #include <linux/pagemap.h>
45 #include <linux/smp_lock.h>
47 #define DEBUG_SUBSYSTEM S_LLITE
49 //#include <lustre_mdc.h>
50 #include <lustre_lite.h>
51 #include "llite_internal.h"
52 #include <linux/lustre_compat25.h>
54 #ifndef list_for_each_prev_safe
55 #define list_for_each_prev_safe(pos, n, head) \
56 for (pos = (head)->prev, n = pos->prev; pos != (head); \
57 pos = n, n = pos->prev )
60 cfs_mem_cache_t *ll_async_page_slab = NULL;
61 size_t ll_async_page_slab_size = 0;
63 /* SYNCHRONOUS I/O to object storage for an inode */
64 static int ll_brw(int cmd, struct inode *inode, struct obdo *oa,
65 struct page *page, int flags)
67 struct ll_inode_info *lli = ll_i2info(inode);
68 struct lov_stripe_md *lsm = lli->lli_smd;
69 struct obd_info oinfo = { { { 0 } } };
75 pg.off = ((obd_off)page->index) << CFS_PAGE_SHIFT;
77 if ((cmd & OBD_BRW_WRITE) && (pg.off+CFS_PAGE_SIZE>i_size_read(inode)))
78 pg.count = i_size_read(inode) % CFS_PAGE_SIZE;
80 pg.count = CFS_PAGE_SIZE;
82 LL_CDEBUG_PAGE(D_PAGE, page, "%s %d bytes ino %lu at "LPU64"/"LPX64"\n",
83 cmd & OBD_BRW_WRITE ? "write" : "read", pg.count,
84 inode->i_ino, pg.off, pg.off);
86 CERROR("ZERO COUNT: ino %lu: size %p:%Lu(%p:%Lu) idx %lu off "
87 LPU64"\n", inode->i_ino, inode, i_size_read(inode),
88 page->mapping->host, i_size_read(page->mapping->host),
94 if (cmd & OBD_BRW_WRITE)
95 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_BRW_WRITE,
98 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_BRW_READ,
102 /* NB partial write, so we might not have CAPA_OPC_OSS_READ capa */
103 opc = cmd & OBD_BRW_WRITE ? CAPA_OPC_OSS_WRITE : CAPA_OPC_OSS_RW;
104 oinfo.oi_capa = ll_osscapa_get(inode, opc);
105 rc = obd_brw(cmd, ll_i2dtexp(inode), &oinfo, 1, &pg, NULL);
106 capa_put(oinfo.oi_capa);
108 obdo_to_inode(inode, oa, OBD_MD_FLBLOCKS);
110 CERROR("error from obd_brw: rc = %d\n", rc);
114 /* this isn't where truncate starts. roughly:
115 * sys_truncate->ll_setattr_raw->vmtruncate->ll_truncate. setattr_raw grabs
116 * DLM lock on [size, EOF], i_mutex, ->lli_size_sem, and WRITE_I_ALLOC_SEM to
119 * must be called under ->lli_size_sem */
120 void ll_truncate(struct inode *inode)
122 struct ll_inode_info *lli = ll_i2info(inode);
123 struct obd_info oinfo = { { { 0 } } };
128 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) to %Lu=%#Lx\n",inode->i_ino,
129 inode->i_generation, inode, i_size_read(inode),
132 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_TRUNC, 1);
133 if (lli->lli_size_sem_owner != current) {
139 CDEBUG(D_INODE, "truncate on inode %lu with no objects\n",
144 LASSERT(atomic_read(&lli->lli_size_sem.count) <= 0);
146 /* XXX I'm pretty sure this is a hack to paper over a more fundamental
148 lov_stripe_lock(lli->lli_smd);
149 inode_init_lvb(inode, &lvb);
150 rc = obd_merge_lvb(ll_i2dtexp(inode), lli->lli_smd, &lvb, 0);
151 if (lvb.lvb_size == i_size_read(inode) && rc == 0) {
152 CDEBUG(D_VFSTRACE, "skipping punch for obj "LPX64", %Lu=%#Lx\n",
153 lli->lli_smd->lsm_object_id, i_size_read(inode),
155 lov_stripe_unlock(lli->lli_smd);
159 obd_adjust_kms(ll_i2dtexp(inode), lli->lli_smd, i_size_read(inode), 1);
160 lov_stripe_unlock(lli->lli_smd);
162 if (unlikely((ll_i2sbi(inode)->ll_flags & LL_SBI_CHECKSUM) &&
163 (i_size_read(inode) & ~CFS_PAGE_MASK))) {
164 /* If the truncate leaves behind a partial page, update its
166 struct page *page = find_get_page(inode->i_mapping,
167 i_size_read(inode) >>
170 struct ll_async_page *llap = llap_cast_private(page);
172 char *kaddr = kmap_atomic(page, KM_USER0);
173 llap->llap_checksum =
174 crc32_le(0, kaddr, CFS_PAGE_SIZE);
175 kunmap_atomic(kaddr, KM_USER0);
177 page_cache_release(page);
181 CDEBUG(D_INFO, "calling punch for "LPX64" (new size %Lu=%#Lx)\n",
182 lli->lli_smd->lsm_object_id, i_size_read(inode), i_size_read(inode));
184 oinfo.oi_md = lli->lli_smd;
185 oinfo.oi_policy.l_extent.start = i_size_read(inode);
186 oinfo.oi_policy.l_extent.end = OBD_OBJECT_EOF;
188 oa.o_id = lli->lli_smd->lsm_object_id;
189 oa.o_gr = lli->lli_smd->lsm_object_gr;
190 oa.o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
192 obdo_from_inode(&oa, inode, OBD_MD_FLTYPE | OBD_MD_FLMODE |
193 OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME |
194 OBD_MD_FLFID | OBD_MD_FLGENER);
196 ll_inode_size_unlock(inode, 0);
198 oinfo.oi_capa = ll_osscapa_get(inode, CAPA_OPC_OSS_TRUNC);
199 rc = obd_punch_rqset(ll_i2dtexp(inode), &oinfo, NULL);
200 ll_truncate_free_capa(oinfo.oi_capa);
202 CERROR("obd_truncate fails (%d) ino %lu\n", rc, inode->i_ino);
204 obdo_to_inode(inode, &oa, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
205 OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME);
210 ll_inode_size_unlock(inode, 0);
213 int ll_prepare_write(struct file *file, struct page *page, unsigned from,
216 struct inode *inode = page->mapping->host;
217 struct ll_inode_info *lli = ll_i2info(inode);
218 struct lov_stripe_md *lsm = lli->lli_smd;
219 obd_off offset = ((obd_off)page->index) << CFS_PAGE_SHIFT;
220 struct obd_info oinfo = { { { 0 } } };
227 LASSERT(PageLocked(page));
228 (void)llap_cast_private(page); /* assertion */
230 /* Check to see if we should return -EIO right away */
233 pga.count = CFS_PAGE_SIZE;
236 oa.o_mode = inode->i_mode;
237 oa.o_id = lsm->lsm_object_id;
238 oa.o_gr = lsm->lsm_object_gr;
239 oa.o_valid = OBD_MD_FLID | OBD_MD_FLMODE |
240 OBD_MD_FLTYPE | OBD_MD_FLGROUP;
241 obdo_from_inode(&oa, inode, OBD_MD_FLFID | OBD_MD_FLGENER);
245 rc = obd_brw(OBD_BRW_CHECK, ll_i2dtexp(inode), &oinfo, 1, &pga, NULL);
249 if (PageUptodate(page)) {
250 LL_CDEBUG_PAGE(D_PAGE, page, "uptodate\n");
254 /* We're completely overwriting an existing page, so _don't_ set it up
255 * to date until commit_write */
256 if (from == 0 && to == CFS_PAGE_SIZE) {
257 LL_CDEBUG_PAGE(D_PAGE, page, "full page write\n");
258 POISON_PAGE(page, 0x11);
262 /* If are writing to a new page, no need to read old data. The extent
263 * locking will have updated the KMS, and for our purposes here we can
264 * treat it like i_size. */
265 lov_stripe_lock(lsm);
266 inode_init_lvb(inode, &lvb);
267 obd_merge_lvb(ll_i2dtexp(inode), lsm, &lvb, 1);
268 lov_stripe_unlock(lsm);
269 if (lvb.lvb_size <= offset) {
270 char *kaddr = kmap_atomic(page, KM_USER0);
271 LL_CDEBUG_PAGE(D_PAGE, page, "kms "LPU64" <= offset "LPU64"\n",
272 lvb.lvb_size, offset);
273 memset(kaddr, 0, CFS_PAGE_SIZE);
274 kunmap_atomic(kaddr, KM_USER0);
275 GOTO(prepare_done, rc = 0);
278 /* XXX could be an async ocp read.. read-ahead? */
279 rc = ll_brw(OBD_BRW_READ, inode, &oa, page, 0);
281 /* bug 1598: don't clobber blksize */
282 oa.o_valid &= ~(OBD_MD_FLSIZE | OBD_MD_FLBLKSZ);
283 obdo_refresh_inode(inode, &oa, oa.o_valid);
289 SetPageUptodate(page);
294 static int ll_ap_make_ready(void *data, int cmd)
296 struct ll_async_page *llap;
300 llap = LLAP_FROM_COOKIE(data);
301 page = llap->llap_page;
303 LASSERTF(!(cmd & OBD_BRW_READ), "cmd %x page %p ino %lu index %lu\n", cmd, page,
304 page->mapping->host->i_ino, page->index);
306 /* we're trying to write, but the page is locked.. come back later */
307 if (TryLockPage(page))
310 LASSERT(!PageWriteback(page));
312 /* if we left PageDirty we might get another writepage call
313 * in the future. list walkers are bright enough
314 * to check page dirty so we can leave it on whatever list
315 * its on. XXX also, we're called with the cli list so if
316 * we got the page cache list we'd create a lock inversion
317 * with the removepage path which gets the page lock then the
319 LASSERTF(!PageWriteback(page),"cmd %x page %p ino %lu index %lu\n", cmd, page,
320 page->mapping->host->i_ino, page->index);
321 clear_page_dirty_for_io(page);
323 /* This actually clears the dirty bit in the radix tree.*/
324 set_page_writeback(page);
326 LL_CDEBUG_PAGE(D_PAGE, page, "made ready\n");
327 page_cache_get(page);
332 /* We have two reasons for giving llite the opportunity to change the
333 * write length of a given queued page as it builds the RPC containing
336 * 1) Further extending writes may have landed in the page cache
337 * since a partial write first queued this page requiring us
338 * to write more from the page cache. (No further races are possible, since
339 * by the time this is called, the page is locked.)
340 * 2) We might have raced with truncate and want to avoid performing
341 * write RPCs that are just going to be thrown away by the
342 * truncate's punch on the storage targets.
344 * The kms serves these purposes as it is set at both truncate and extending
347 static int ll_ap_refresh_count(void *data, int cmd)
349 struct ll_inode_info *lli;
350 struct ll_async_page *llap;
351 struct lov_stripe_md *lsm;
358 /* readpage queues with _COUNT_STABLE, shouldn't get here. */
359 LASSERT(cmd != OBD_BRW_READ);
361 llap = LLAP_FROM_COOKIE(data);
362 page = llap->llap_page;
363 inode = page->mapping->host;
364 lli = ll_i2info(inode);
367 lov_stripe_lock(lsm);
368 inode_init_lvb(inode, &lvb);
369 obd_merge_lvb(ll_i2dtexp(inode), lsm, &lvb, 1);
371 lov_stripe_unlock(lsm);
373 /* catch race with truncate */
374 if (((__u64)page->index << CFS_PAGE_SHIFT) >= kms)
377 /* catch sub-page write at end of file */
378 if (((__u64)page->index << CFS_PAGE_SHIFT) + CFS_PAGE_SIZE > kms)
379 return kms % CFS_PAGE_SIZE;
381 return CFS_PAGE_SIZE;
384 void ll_inode_fill_obdo(struct inode *inode, int cmd, struct obdo *oa)
386 struct lov_stripe_md *lsm;
387 obd_flag valid_flags;
389 lsm = ll_i2info(inode)->lli_smd;
391 oa->o_id = lsm->lsm_object_id;
392 oa->o_gr = lsm->lsm_object_gr;
393 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
394 valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME;
395 if (cmd & OBD_BRW_WRITE) {
396 oa->o_valid |= OBD_MD_FLEPOCH;
397 oa->o_easize = ll_i2info(inode)->lli_ioepoch;
399 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME |
400 OBD_MD_FLUID | OBD_MD_FLGID |
401 OBD_MD_FLFID | OBD_MD_FLGENER;
404 obdo_from_inode(oa, inode, valid_flags);
407 static void ll_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
409 struct ll_async_page *llap;
412 llap = LLAP_FROM_COOKIE(data);
413 ll_inode_fill_obdo(llap->llap_page->mapping->host, cmd, oa);
418 static void ll_ap_update_obdo(void *data, int cmd, struct obdo *oa,
421 struct ll_async_page *llap;
424 llap = LLAP_FROM_COOKIE(data);
425 obdo_from_inode(oa, llap->llap_page->mapping->host, valid);
430 static struct obd_capa *ll_ap_lookup_capa(void *data, int cmd)
432 struct ll_async_page *llap = LLAP_FROM_COOKIE(data);
433 int opc = cmd & OBD_BRW_WRITE ? CAPA_OPC_OSS_WRITE : CAPA_OPC_OSS_RW;
435 return ll_osscapa_get(llap->llap_page->mapping->host, opc);
438 static struct obd_async_page_ops ll_async_page_ops = {
439 .ap_make_ready = ll_ap_make_ready,
440 .ap_refresh_count = ll_ap_refresh_count,
441 .ap_fill_obdo = ll_ap_fill_obdo,
442 .ap_update_obdo = ll_ap_update_obdo,
443 .ap_completion = ll_ap_completion,
444 .ap_lookup_capa = ll_ap_lookup_capa,
447 struct ll_async_page *llap_cast_private(struct page *page)
449 struct ll_async_page *llap = (struct ll_async_page *)page_private(page);
451 LASSERTF(llap == NULL || llap->llap_magic == LLAP_MAGIC,
452 "page %p private %lu gave magic %d which != %d\n",
453 page, page_private(page), llap->llap_magic, LLAP_MAGIC);
458 /* Try to shrink the page cache for the @sbi filesystem by 1/@shrink_fraction.
460 * There is an llap attached onto every page in lustre, linked off @sbi.
461 * We add an llap to the list so we don't lose our place during list walking.
462 * If llaps in the list are being moved they will only move to the end
463 * of the LRU, and we aren't terribly interested in those pages here (we
464 * start at the beginning of the list where the least-used llaps are.
466 int llap_shrink_cache(struct ll_sb_info *sbi, int shrink_fraction)
468 struct ll_async_page *llap, dummy_llap = { .llap_magic = 0xd11ad11a };
469 unsigned long total, want, count = 0;
471 total = sbi->ll_async_page_count;
473 /* There can be a large number of llaps (600k or more in a large
474 * memory machine) so the VM 1/6 shrink ratio is likely too much.
475 * Since we are freeing pages also, we don't necessarily want to
476 * shrink so much. Limit to 40MB of pages + llaps per call. */
477 if (shrink_fraction == 0)
478 want = sbi->ll_async_page_count - sbi->ll_async_page_max + 32;
480 want = (total + shrink_fraction - 1) / shrink_fraction;
482 if (want > 40 << (20 - CFS_PAGE_SHIFT))
483 want = 40 << (20 - CFS_PAGE_SHIFT);
485 CDEBUG(D_CACHE, "shrinking %lu of %lu pages (1/%d)\n",
486 want, total, shrink_fraction);
488 spin_lock(&sbi->ll_lock);
489 list_add(&dummy_llap.llap_pglist_item, &sbi->ll_pglist);
491 while (--total >= 0 && count < want) {
495 if (unlikely(need_resched())) {
496 spin_unlock(&sbi->ll_lock);
498 spin_lock(&sbi->ll_lock);
501 llap = llite_pglist_next_llap(sbi,&dummy_llap.llap_pglist_item);
502 list_del_init(&dummy_llap.llap_pglist_item);
506 page = llap->llap_page;
507 LASSERT(page != NULL);
509 list_add(&dummy_llap.llap_pglist_item, &llap->llap_pglist_item);
511 /* Page needs/undergoing IO */
512 if (TryLockPage(page)) {
513 LL_CDEBUG_PAGE(D_PAGE, page, "can't lock\n");
517 keep = (llap->llap_write_queued || PageDirty(page) ||
518 PageWriteback(page) || (!PageUptodate(page) &&
519 llap->llap_origin != LLAP_ORIGIN_READAHEAD));
521 LL_CDEBUG_PAGE(D_PAGE, page,"%s LRU page: %s%s%s%s%s origin %s\n",
522 keep ? "keep" : "drop",
523 llap->llap_write_queued ? "wq " : "",
524 PageDirty(page) ? "pd " : "",
525 PageUptodate(page) ? "" : "!pu ",
526 PageWriteback(page) ? "wb" : "",
527 llap->llap_defer_uptodate ? "" : "!du",
528 llap_origins[llap->llap_origin]);
530 /* If page is dirty or undergoing IO don't discard it */
536 page_cache_get(page);
537 spin_unlock(&sbi->ll_lock);
539 if (page->mapping != NULL) {
540 ll_teardown_mmaps(page->mapping,
541 (__u64)page->index << CFS_PAGE_SHIFT,
542 ((__u64)page->index << CFS_PAGE_SHIFT)|
544 if (!PageDirty(page) && !page_mapped(page)) {
545 ll_ra_accounting(llap, page->mapping);
546 ll_truncate_complete_page(page);
549 LL_CDEBUG_PAGE(D_PAGE, page, "Not dropping page"
557 page_cache_release(page);
559 spin_lock(&sbi->ll_lock);
561 list_del(&dummy_llap.llap_pglist_item);
562 spin_unlock(&sbi->ll_lock);
564 CDEBUG(D_CACHE, "shrank %lu/%lu and left %lu unscanned\n",
570 struct ll_async_page *llap_from_page(struct page *page, unsigned origin)
572 struct ll_async_page *llap;
573 struct obd_export *exp;
574 struct inode *inode = page->mapping->host;
575 struct ll_sb_info *sbi;
580 static int triggered;
583 LL_CDEBUG_PAGE(D_ERROR, page, "Bug 10047. Wrong anon "
585 libcfs_debug_dumpstack(NULL);
588 RETURN(ERR_PTR(-EINVAL));
590 sbi = ll_i2sbi(inode);
591 LASSERT(ll_async_page_slab);
592 LASSERTF(origin < LLAP__ORIGIN_MAX, "%u\n", origin);
594 llap = llap_cast_private(page);
596 /* move to end of LRU list, except when page is just about to
598 if (origin != LLAP_ORIGIN_REMOVEPAGE) {
599 spin_lock(&sbi->ll_lock);
600 sbi->ll_pglist_gen++;
601 list_del_init(&llap->llap_pglist_item);
602 list_add_tail(&llap->llap_pglist_item, &sbi->ll_pglist);
603 spin_unlock(&sbi->ll_lock);
608 exp = ll_i2dtexp(page->mapping->host);
610 RETURN(ERR_PTR(-EINVAL));
612 /* limit the number of lustre-cached pages */
613 if (sbi->ll_async_page_count >= sbi->ll_async_page_max)
614 llap_shrink_cache(sbi, 0);
616 OBD_SLAB_ALLOC(llap, ll_async_page_slab, CFS_ALLOC_STD,
617 ll_async_page_slab_size);
619 RETURN(ERR_PTR(-ENOMEM));
620 llap->llap_magic = LLAP_MAGIC;
621 llap->llap_cookie = (void *)llap + size_round(sizeof(*llap));
623 rc = obd_prep_async_page(exp, ll_i2info(inode)->lli_smd, NULL, page,
624 (obd_off)page->index << CFS_PAGE_SHIFT,
625 &ll_async_page_ops, llap, &llap->llap_cookie);
627 OBD_SLAB_FREE(llap, ll_async_page_slab,
628 ll_async_page_slab_size);
632 CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n", llap,
633 page, llap->llap_cookie, (obd_off)page->index << CFS_PAGE_SHIFT);
634 /* also zeroing the PRIVBITS low order bitflags */
635 __set_page_ll_data(page, llap);
636 llap->llap_page = page;
637 spin_lock(&sbi->ll_lock);
638 sbi->ll_pglist_gen++;
639 sbi->ll_async_page_count++;
640 list_add_tail(&llap->llap_pglist_item, &sbi->ll_pglist);
641 INIT_LIST_HEAD(&llap->llap_pending_write);
642 spin_unlock(&sbi->ll_lock);
645 if (unlikely(sbi->ll_flags & LL_SBI_CHECKSUM)) {
647 char *kaddr = kmap_atomic(page, KM_USER0);
648 csum = crc32_le(csum, kaddr, CFS_PAGE_SIZE);
649 kunmap_atomic(kaddr, KM_USER0);
650 if (origin == LLAP_ORIGIN_READAHEAD ||
651 origin == LLAP_ORIGIN_READPAGE) {
652 llap->llap_checksum = 0;
653 } else if (origin == LLAP_ORIGIN_COMMIT_WRITE ||
654 llap->llap_checksum == 0) {
655 llap->llap_checksum = csum;
656 CDEBUG(D_PAGE, "page %p cksum %x\n", page, csum);
657 } else if (llap->llap_checksum == csum) {
658 /* origin == LLAP_ORIGIN_WRITEPAGE */
659 CDEBUG(D_PAGE, "page %p cksum %x confirmed\n",
662 /* origin == LLAP_ORIGIN_WRITEPAGE */
663 LL_CDEBUG_PAGE(D_ERROR, page, "old cksum %x != new "
664 "%x!\n", llap->llap_checksum, csum);
668 llap->llap_origin = origin;
672 static int queue_or_sync_write(struct obd_export *exp, struct inode *inode,
673 struct ll_async_page *llap,
674 unsigned to, obd_flag async_flags)
676 unsigned long size_index = i_size_read(inode) >> CFS_PAGE_SHIFT;
677 struct obd_io_group *oig;
678 struct ll_sb_info *sbi = ll_i2sbi(inode);
679 int rc, noquot = llap->llap_ignore_quota ? OBD_BRW_NOQUOTA : 0;
682 /* _make_ready only sees llap once we've unlocked the page */
683 llap->llap_write_queued = 1;
684 rc = obd_queue_async_io(exp, ll_i2info(inode)->lli_smd, NULL,
685 llap->llap_cookie, OBD_BRW_WRITE | noquot,
686 0, 0, 0, async_flags);
688 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "write queued\n");
692 llap->llap_write_queued = 0;
693 /* Do not pass llap here as it is sync write. */
694 llap_write_pending(inode, NULL);
700 /* make full-page requests if we are not at EOF (bug 4410) */
701 if (to != CFS_PAGE_SIZE && llap->llap_page->index < size_index) {
702 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
703 "sync write before EOF: size_index %lu, to %d\n",
706 } else if (to != CFS_PAGE_SIZE && llap->llap_page->index == size_index) {
707 int size_to = i_size_read(inode) & ~CFS_PAGE_MASK;
708 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
709 "sync write at EOF: size_index %lu, to %d/%d\n",
710 size_index, to, size_to);
715 /* compare the checksum once before the page leaves llite */
716 if (unlikely((sbi->ll_flags & LL_SBI_CHECKSUM) &&
717 llap->llap_checksum != 0)) {
719 struct page *page = llap->llap_page;
720 char *kaddr = kmap_atomic(page, KM_USER0);
721 csum = crc32_le(csum, kaddr, CFS_PAGE_SIZE);
722 kunmap_atomic(kaddr, KM_USER0);
723 if (llap->llap_checksum == csum) {
724 CDEBUG(D_PAGE, "page %p cksum %x confirmed\n",
727 CERROR("page %p old cksum %x != new cksum %x!\n",
728 page, llap->llap_checksum, csum);
732 rc = obd_queue_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig,
733 llap->llap_cookie, OBD_BRW_WRITE | noquot,
734 0, to, 0, ASYNC_READY | ASYNC_URGENT |
735 ASYNC_COUNT_STABLE | ASYNC_GROUP_SYNC);
739 rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
745 if (!rc && async_flags & ASYNC_READY) {
746 unlock_page(llap->llap_page);
747 if (PageWriteback(llap->llap_page)) {
748 end_page_writeback(llap->llap_page);
752 if (rc == 0 && llap_write_complete(inode, llap))
753 ll_queue_done_writing(inode, 0);
755 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "sync write returned %d\n", rc);
763 /* update our write count to account for i_size increases that may have
764 * happened since we've queued the page for io. */
766 /* be careful not to return success without setting the page Uptodate or
767 * the next pass through prepare_write will read in stale data from disk. */
768 int ll_commit_write(struct file *file, struct page *page, unsigned from,
771 struct inode *inode = page->mapping->host;
772 struct ll_inode_info *lli = ll_i2info(inode);
773 struct lov_stripe_md *lsm = lli->lli_smd;
774 struct obd_export *exp;
775 struct ll_async_page *llap;
780 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
781 LASSERT(inode == file->f_dentry->d_inode);
782 LASSERT(PageLocked(page));
784 CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
785 inode, page, from, to, page->index);
787 llap = llap_from_page(page, LLAP_ORIGIN_COMMIT_WRITE);
789 RETURN(PTR_ERR(llap));
791 exp = ll_i2dtexp(inode);
795 llap->llap_ignore_quota = capable(CAP_SYS_RESOURCE);
798 * queue a write for some time in the future the first time we
801 * This is different from what other file systems do: they usually
802 * just mark page (and some of its buffers) dirty and rely on
803 * balance_dirty_pages() to start a write-back. Lustre wants write-back
804 * to be started earlier for the following reasons:
806 * (1) with a large number of clients we need to limit the amount
807 * of cached data on the clients a lot;
809 * (2) large compute jobs generally want compute-only then io-only
810 * and the IO should complete as quickly as possible;
812 * (3) IO is batched up to the RPC size and is async until the
813 * client max cache is hit
814 * (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
817 if (!PageDirty(page)) {
818 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRTY_MISSES, 1);
820 rc = queue_or_sync_write(exp, inode, llap, to, 0);
824 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRTY_HITS, 1);
827 /* put the page in the page cache, from now on ll_removepage is
828 * responsible for cleaning up the llap.
829 * only set page dirty when it's queued to be write out */
830 if (llap->llap_write_queued)
831 set_page_dirty(page);
834 size = (((obd_off)page->index) << CFS_PAGE_SHIFT) + to;
835 ll_inode_size_lock(inode, 0);
837 lov_stripe_lock(lsm);
838 obd_adjust_kms(exp, lsm, size, 0);
839 lov_stripe_unlock(lsm);
840 if (size > i_size_read(inode))
841 i_size_write(inode, size);
842 SetPageUptodate(page);
843 } else if (size > i_size_read(inode)) {
844 /* this page beyond the pales of i_size, so it can't be
845 * truncated in ll_p_r_e during lock revoking. we must
846 * teardown our book-keeping here. */
849 ll_inode_size_unlock(inode, 0);
853 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, unsigned long len)
855 struct ll_ra_info *ra = &sbi->ll_ra_info;
859 spin_lock(&sbi->ll_lock);
860 ret = min(ra->ra_max_pages - ra->ra_cur_pages, len);
861 ra->ra_cur_pages += ret;
862 spin_unlock(&sbi->ll_lock);
867 static void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
869 struct ll_ra_info *ra = &sbi->ll_ra_info;
870 spin_lock(&sbi->ll_lock);
871 LASSERTF(ra->ra_cur_pages >= len, "r_c_p %lu len %lu\n",
872 ra->ra_cur_pages, len);
873 ra->ra_cur_pages -= len;
874 spin_unlock(&sbi->ll_lock);
877 /* called for each page in a completed rpc.*/
878 int ll_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
880 struct ll_async_page *llap;
885 llap = LLAP_FROM_COOKIE(data);
886 page = llap->llap_page;
887 LASSERT(PageLocked(page));
888 LASSERT(CheckWriteback(page,cmd));
890 LL_CDEBUG_PAGE(D_PAGE, page, "completing cmd %d with %d\n", cmd, rc);
892 if (cmd & OBD_BRW_READ && llap->llap_defer_uptodate)
893 ll_ra_count_put(ll_i2sbi(page->mapping->host), 1);
896 if (cmd & OBD_BRW_READ) {
897 if (!llap->llap_defer_uptodate)
898 SetPageUptodate(page);
900 llap->llap_write_queued = 0;
902 ClearPageError(page);
904 if (cmd & OBD_BRW_READ) {
905 llap->llap_defer_uptodate = 0;
909 set_bit(AS_ENOSPC, &page->mapping->flags);
911 set_bit(AS_EIO, &page->mapping->flags);
916 if (cmd & OBD_BRW_WRITE) {
917 /* Only rc == 0, write succeed, then this page could be deleted
918 * from the pending_writing list
920 if (rc == 0 && llap_write_complete(page->mapping->host, llap))
921 ll_queue_done_writing(page->mapping->host, 0);
924 if (PageWriteback(page)) {
925 end_page_writeback(page);
927 page_cache_release(page);
932 /* the kernel calls us here when a page is unhashed from the page cache.
933 * the page will be locked and the kernel is holding a spinlock, so
934 * we need to be careful. we're just tearing down our book-keeping
936 void ll_removepage(struct page *page)
938 struct inode *inode = page->mapping->host;
939 struct obd_export *exp;
940 struct ll_async_page *llap;
941 struct ll_sb_info *sbi = ll_i2sbi(inode);
945 LASSERT(!in_interrupt());
947 /* sync pages or failed read pages can leave pages in the page
948 * cache that don't have our data associated with them anymore */
949 if (page_private(page) == 0) {
954 LL_CDEBUG_PAGE(D_PAGE, page, "being evicted\n");
956 exp = ll_i2dtexp(inode);
958 CERROR("page %p ind %lu gave null export\n", page, page->index);
963 llap = llap_from_page(page, LLAP_ORIGIN_REMOVEPAGE);
965 CERROR("page %p ind %lu couldn't find llap: %ld\n", page,
966 page->index, PTR_ERR(llap));
971 if (llap_write_complete(inode, llap))
972 ll_queue_done_writing(inode, 0);
974 rc = obd_teardown_async_page(exp, ll_i2info(inode)->lli_smd, NULL,
977 CERROR("page %p ind %lu failed: %d\n", page, page->index, rc);
979 /* this unconditional free is only safe because the page lock
980 * is providing exclusivity to memory pressure/truncate/writeback..*/
981 __clear_page_ll_data(page);
983 spin_lock(&sbi->ll_lock);
984 if (!list_empty(&llap->llap_pglist_item))
985 list_del_init(&llap->llap_pglist_item);
986 sbi->ll_pglist_gen++;
987 sbi->ll_async_page_count--;
988 spin_unlock(&sbi->ll_lock);
989 OBD_SLAB_FREE(llap, ll_async_page_slab, ll_async_page_slab_size);
993 static int ll_page_matches(struct page *page, int fd_flags)
995 struct lustre_handle match_lockh = {0};
996 struct inode *inode = page->mapping->host;
997 ldlm_policy_data_t page_extent;
1001 if (unlikely(fd_flags & LL_FILE_GROUP_LOCKED))
1004 page_extent.l_extent.start = (__u64)page->index << CFS_PAGE_SHIFT;
1005 page_extent.l_extent.end =
1006 page_extent.l_extent.start + CFS_PAGE_SIZE - 1;
1007 flags = LDLM_FL_TEST_LOCK | LDLM_FL_BLOCK_GRANTED;
1008 if (!(fd_flags & LL_FILE_READAHEAD))
1009 flags |= LDLM_FL_CBPENDING;
1010 matches = obd_match(ll_i2sbi(inode)->ll_dt_exp,
1011 ll_i2info(inode)->lli_smd, LDLM_EXTENT,
1012 &page_extent, LCK_PR | LCK_PW, &flags, inode,
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 |
1030 ASYNC_READY | ASYNC_URGENT);
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_i2dtexp(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_length = ras->ras_stride_length;
1389 ria.ria_pages = ras->ras_stride_pages;
1391 spin_unlock(&ras->ras_lock);
1394 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
1397 len = ria_page_count(&ria);
1401 reserved = ll_ra_count_get(ll_i2sbi(inode), len);
1403 if (reserved < end - start + 1)
1404 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
1406 CDEBUG(D_READA, "reserved page %lu \n", reserved);
1408 ret = ll_read_ahead_pages(exp, oig, &ria, &reserved, mapping, &ra_end);
1410 LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
1412 ll_ra_count_put(ll_i2sbi(inode), reserved);
1414 if (ra_end == end + 1 && ra_end == (kms >> CFS_PAGE_SHIFT))
1415 ll_ra_stats_inc(mapping, RA_STAT_EOF);
1417 /* if we didn't get to the end of the region we reserved from
1418 * the ras we need to go back and update the ras so that the
1419 * next read-ahead tries from where we left off. we only do so
1420 * if the region we failed to issue read-ahead on is still ahead
1421 * of the app and behind the next index to start read-ahead from */
1422 CDEBUG(D_READA, "ra_end %lu end %lu stride end %lu \n",
1423 ra_end, end, ria.ria_end);
1425 if (ra_end != (end + 1)) {
1426 spin_lock(&ras->ras_lock);
1427 if (ra_end < ras->ras_next_readahead &&
1428 index_in_window(ra_end, ras->ras_window_start, 0,
1429 ras->ras_window_len)) {
1430 ras->ras_next_readahead = ra_end;
1433 spin_unlock(&ras->ras_lock);
1439 static void ras_set_start(struct ll_readahead_state *ras, unsigned long index)
1441 ras->ras_window_start = index & (~(RAS_INCREASE_STEP - 1));
1444 /* called with the ras_lock held or from places where it doesn't matter */
1445 static void ras_reset(struct ll_readahead_state *ras, unsigned long index)
1447 ras->ras_last_readpage = index;
1448 ras->ras_consecutive_requests = 0;
1449 ras->ras_consecutive_pages = 0;
1450 ras->ras_window_len = 0;
1451 ras_set_start(ras, index);
1452 ras->ras_next_readahead = max(ras->ras_window_start, index);
1457 /* called with the ras_lock held or from places where it doesn't matter */
1458 static void ras_stride_reset(struct ll_readahead_state *ras)
1460 ras->ras_consecutive_stride_requests = 0;
1464 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
1466 spin_lock_init(&ras->ras_lock);
1468 ras->ras_requests = 0;
1469 INIT_LIST_HEAD(&ras->ras_read_beads);
1472 /* Check whether the read request is in the stride window.
1473 * If it is in the stride window, return 1, otherwise return 0.
1474 * and also update stride_gap and stride_pages.
1476 static int index_in_stride_window(unsigned long index,
1477 struct ll_readahead_state *ras,
1478 struct inode *inode)
1480 int stride_gap = index - ras->ras_last_readpage - 1;
1482 LASSERT(stride_gap != 0);
1484 if (ras->ras_consecutive_pages == 0)
1487 /*Otherwise check the stride by itself */
1488 if ((ras->ras_stride_length - ras->ras_stride_pages) == stride_gap &&
1489 ras->ras_consecutive_pages == ras->ras_stride_pages)
1492 if (stride_gap >= 0) {
1494 * only set stride_pages, stride_length if
1495 * it is forward reading ( stride_gap > 0)
1497 ras->ras_stride_pages = ras->ras_consecutive_pages;
1498 ras->ras_stride_length = stride_gap + ras->ras_consecutive_pages;
1501 * If stride_gap < 0,(back_forward reading),
1502 * reset the stride_pages/length.
1503 * FIXME:back_ward stride I/O read.
1506 ras->ras_stride_pages = 0;
1507 ras->ras_stride_length = 0;
1514 static unsigned long
1515 stride_page_count(struct ll_readahead_state *ras, unsigned long len)
1517 return stride_pg_count(ras->ras_stride_offset, ras->ras_stride_length,
1518 ras->ras_stride_pages, ras->ras_stride_offset,
1522 /* Stride Read-ahead window will be increased inc_len according to
1523 * stride I/O pattern */
1524 static void ras_stride_increase_window(struct ll_readahead_state *ras,
1525 struct ll_ra_info *ra,
1526 unsigned long inc_len)
1528 unsigned long left, step, window_len;
1529 unsigned long stride_len;
1531 LASSERT(ras->ras_stride_length > 0);
1533 stride_len = ras->ras_window_start + ras->ras_window_len -
1534 ras->ras_stride_offset;
1536 LASSERTF(stride_len > 0, "window_start %lu, window_len %lu"
1537 "stride_offset %lu\n", ras->ras_window_start,
1538 ras->ras_window_len, ras->ras_stride_offset);
1540 left = stride_len % ras->ras_stride_length;
1542 window_len = ras->ras_window_len - left;
1544 if (left < ras->ras_stride_pages)
1547 left = ras->ras_stride_pages + inc_len;
1549 LASSERT(ras->ras_stride_pages != 0);
1551 step = left / ras->ras_stride_pages;
1552 left %= ras->ras_stride_pages;
1554 window_len += step * ras->ras_stride_length + left;
1556 if (stride_page_count(ras, window_len) <= ra->ra_max_pages)
1557 ras->ras_window_len = window_len;
1562 /* Set stride I/O read-ahead window start offset */
1563 static void ras_set_stride_offset(struct ll_readahead_state *ras)
1565 unsigned long window_len = ras->ras_next_readahead -
1566 ras->ras_window_start;
1569 LASSERT(ras->ras_stride_length != 0);
1571 left = window_len % ras->ras_stride_length;
1573 ras->ras_stride_offset = ras->ras_next_readahead - left;
1578 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
1579 struct ll_readahead_state *ras, unsigned long index,
1582 struct ll_ra_info *ra = &sbi->ll_ra_info;
1583 int zero = 0, stride_zero = 0, stride_detect = 0, ra_miss = 0;
1586 spin_lock(&sbi->ll_lock);
1587 spin_lock(&ras->ras_lock);
1589 ll_ra_stats_inc_unlocked(ra, hit ? RA_STAT_HIT : RA_STAT_MISS);
1591 /* reset the read-ahead window in two cases. First when the app seeks
1592 * or reads to some other part of the file. Secondly if we get a
1593 * read-ahead miss that we think we've previously issued. This can
1594 * be a symptom of there being so many read-ahead pages that the VM is
1595 * reclaiming it before we get to it. */
1596 if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
1598 ll_ra_stats_inc_unlocked(ra, RA_STAT_DISTANT_READPAGE);
1599 /* check whether it is in stride I/O mode*/
1600 if (!index_in_stride_window(index, ras, inode))
1602 } else if (!hit && ras->ras_window_len &&
1603 index < ras->ras_next_readahead &&
1604 index_in_window(index, ras->ras_window_start, 0,
1605 ras->ras_window_len)) {
1608 /* If it hits read-ahead miss and the stride I/O is still
1609 * not detected, reset stride stuff to re-detect the whole
1610 * stride I/O mode to avoid complication */
1611 if (!stride_io_mode(ras))
1613 ll_ra_stats_inc_unlocked(ra, RA_STAT_MISS_IN_WINDOW);
1616 /* On the second access to a file smaller than the tunable
1617 * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1618 * file up to ra_max_pages. This is simply a best effort and
1619 * only occurs once per open file. Normal RA behavior is reverted
1620 * to for subsequent IO. The mmap case does not increment
1621 * ras_requests and thus can never trigger this behavior. */
1622 if (ras->ras_requests == 2 && !ras->ras_request_index) {
1625 kms_pages = (i_size_read(inode) + CFS_PAGE_SIZE - 1) >>
1628 CDEBUG(D_READA, "kmsp "LPU64" mwp %lu mp %lu\n", kms_pages,
1629 ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages);
1632 kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1633 ras->ras_window_start = 0;
1634 ras->ras_last_readpage = 0;
1635 ras->ras_next_readahead = 0;
1636 ras->ras_window_len = min(ra->ra_max_pages,
1637 ra->ra_max_read_ahead_whole_pages);
1638 GOTO(out_unlock, 0);
1643 /* If it is discontinuous read, check
1644 * whether it is stride I/O mode*/
1646 ras_reset(ras, index);
1647 ras->ras_consecutive_pages++;
1648 ras_stride_reset(ras);
1650 GOTO(out_unlock, 0);
1652 /* The read is still in stride window or
1653 * it hits read-ahead miss */
1655 /* If ra-window miss is hitted, which probably means VM
1656 * pressure, and some read-ahead pages were reclaimed.So
1657 * the length of ra-window will not increased, but also
1658 * not reset to avoid redetecting the stride I/O mode.*/
1659 ras->ras_consecutive_requests = 0;
1661 ras->ras_consecutive_pages = 0;
1662 if (++ras->ras_consecutive_stride_requests > 1)
1667 } else if (ras->ras_consecutive_stride_requests > 1) {
1668 /* If this is contiguous read but in stride I/O mode
1669 * currently, check whether stride step still is valid,
1670 * if invalid, it will reset the stride ra window*/
1671 if (ras->ras_consecutive_pages + 1 > ras->ras_stride_pages)
1672 ras_stride_reset(ras);
1675 ras->ras_last_readpage = index;
1676 ras->ras_consecutive_pages++;
1677 ras_set_start(ras, index);
1678 ras->ras_next_readahead = max(ras->ras_window_start,
1679 ras->ras_next_readahead);
1682 /* Trigger RA in the mmap case where ras_consecutive_requests
1683 * is not incremented and thus can't be used to trigger RA */
1684 if (!ras->ras_window_len && ras->ras_consecutive_pages == 4) {
1685 ras->ras_window_len = RAS_INCREASE_STEP;
1686 GOTO(out_unlock, 0);
1689 /* Initially reset the stride window offset to next_readahead*/
1690 if (ras->ras_consecutive_stride_requests == 2 && stride_detect)
1691 ras_set_stride_offset(ras);
1693 /* The initial ras_window_len is set to the request size. To avoid
1694 * uselessly reading and discarding pages for random IO the window is
1695 * only increased once per consecutive request received. */
1696 if ((ras->ras_consecutive_requests > 1 &&
1697 !ras->ras_request_index) || stride_detect) {
1698 if (stride_io_mode(ras))
1699 ras_stride_increase_window(ras, ra, RAS_INCREASE_STEP);
1701 ras->ras_window_len = min(ras->ras_window_len +
1708 ras->ras_request_index++;
1709 spin_unlock(&ras->ras_lock);
1710 spin_unlock(&sbi->ll_lock);
1714 int ll_writepage(struct page *page)
1716 struct inode *inode = page->mapping->host;
1717 struct ll_inode_info *lli = ll_i2info(inode);
1718 struct obd_export *exp;
1719 struct ll_async_page *llap;
1723 LASSERT(PageLocked(page));
1725 exp = ll_i2dtexp(inode);
1727 GOTO(out, rc = -EINVAL);
1729 llap = llap_from_page(page, LLAP_ORIGIN_WRITEPAGE);
1731 GOTO(out, rc = PTR_ERR(llap));
1733 LASSERT(!PageWriteback(page));
1734 set_page_writeback(page);
1736 page_cache_get(page);
1737 if (llap->llap_write_queued) {
1738 LL_CDEBUG_PAGE(D_PAGE, page, "marking urgent\n");
1739 rc = obd_set_async_flags(exp, lli->lli_smd, NULL,
1741 ASYNC_READY | ASYNC_URGENT);
1743 rc = queue_or_sync_write(exp, inode, llap, CFS_PAGE_SIZE,
1744 ASYNC_READY | ASYNC_URGENT);
1747 page_cache_release(page);
1750 if (!lli->lli_async_rc)
1751 lli->lli_async_rc = rc;
1752 /* re-dirty page on error so it retries write */
1753 if (PageWriteback(page)) {
1754 end_page_writeback(page);
1756 /* resend page only for not started IO*/
1757 if (!PageError(page))
1758 ll_redirty_page(page);
1765 * for now we do our readpage the same on both 2.4 and 2.5. The kernel's
1766 * read-ahead assumes it is valid to issue readpage all the way up to
1767 * i_size, but our dlm locks make that not the case. We disable the
1768 * kernel's read-ahead and do our own by walking ahead in the page cache
1769 * checking for dlm lock coverage. the main difference between 2.4 and
1770 * 2.6 is how read-ahead gets batched and issued, but we're using our own,
1771 * so they look the same.
1773 int ll_readpage(struct file *filp, struct page *page)
1775 struct ll_file_data *fd = LUSTRE_FPRIVATE(filp);
1776 struct inode *inode = page->mapping->host;
1777 struct obd_export *exp;
1778 struct ll_async_page *llap;
1779 struct obd_io_group *oig = NULL;
1783 LASSERT(PageLocked(page));
1784 LASSERT(!PageUptodate(page));
1785 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset=%Lu=%#Lx\n",
1786 inode->i_ino, inode->i_generation, inode,
1787 (((loff_t)page->index) << CFS_PAGE_SHIFT),
1788 (((loff_t)page->index) << CFS_PAGE_SHIFT));
1789 LASSERT(atomic_read(&filp->f_dentry->d_inode->i_count) > 0);
1791 if (!ll_i2info(inode)->lli_smd) {
1792 /* File with no objects - one big hole */
1793 /* We use this just for remove_from_page_cache that is not
1794 * exported, we'd make page back up to date. */
1795 ll_truncate_complete_page(page);
1796 clear_page(kmap(page));
1798 SetPageUptodate(page);
1803 rc = oig_init(&oig);
1807 exp = ll_i2dtexp(inode);
1809 GOTO(out, rc = -EINVAL);
1811 llap = llap_from_page(page, LLAP_ORIGIN_READPAGE);
1813 GOTO(out, rc = PTR_ERR(llap));
1815 if (ll_i2sbi(inode)->ll_ra_info.ra_max_pages)
1816 ras_update(ll_i2sbi(inode), inode, &fd->fd_ras, page->index,
1817 llap->llap_defer_uptodate);
1820 if (llap->llap_defer_uptodate) {
1821 /* This is the callpath if we got the page from a readahead */
1822 llap->llap_ra_used = 1;
1823 rc = ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1826 obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd,
1828 LL_CDEBUG_PAGE(D_PAGE, page, "marking uptodate from defer\n");
1829 SetPageUptodate(page);
1831 GOTO(out_oig, rc = 0);
1834 if (likely((fd->fd_flags & LL_FILE_IGNORE_LOCK) == 0)) {
1835 rc = ll_page_matches(page, fd->fd_flags);
1837 LL_CDEBUG_PAGE(D_ERROR, page,
1838 "lock match failed: rc %d\n", rc);
1843 CWARN("ino %lu page %lu (%llu) not covered by "
1844 "a lock (mmap?). check debug logs.\n",
1845 inode->i_ino, page->index,
1846 (long long)page->index << CFS_PAGE_SHIFT);
1850 rc = ll_issue_page_read(exp, llap, oig, 0);
1854 LL_CDEBUG_PAGE(D_PAGE, page, "queued readpage\n");
1855 /* We have just requested the actual page we want, see if we can tack
1856 * on some readahead to that page's RPC before it is sent. */
1857 if (ll_i2sbi(inode)->ll_ra_info.ra_max_pages)
1858 ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1861 rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);