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 #include <linux/autoconf.h>
24 #include <linux/kernel.h>
26 #include <linux/string.h>
27 #include <linux/stat.h>
28 #include <linux/errno.h>
29 #include <linux/smp_lock.h>
30 #include <linux/unistd.h>
31 #include <linux/version.h>
32 #include <asm/system.h>
33 #include <asm/uaccess.h>
36 #include <linux/stat.h>
37 #include <asm/uaccess.h>
39 #include <linux/pagemap.h>
40 #include <linux/smp_lock.h>
42 #define DEBUG_SUBSYSTEM S_LLITE
44 //#include <lustre_mdc.h>
45 #include <lustre_lite.h>
46 #include <obd_cksum.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 /* NB partial write, so we might not have CAPA_OPC_OSS_READ capa */
99 opc = cmd & OBD_BRW_WRITE ? CAPA_OPC_OSS_WRITE : CAPA_OPC_OSS_RW;
100 oinfo.oi_capa = ll_osscapa_get(inode, opc);
101 rc = obd_brw(cmd, ll_i2dtexp(inode), &oinfo, 1, &pg, NULL);
102 capa_put(oinfo.oi_capa);
104 obdo_to_inode(inode, oa, OBD_MD_FLBLOCKS);
106 CERROR("error from obd_brw: rc = %d\n", rc);
110 int ll_file_punch(struct inode * inode, loff_t new_size, int srvlock)
112 struct ll_inode_info *lli = ll_i2info(inode);
113 struct obd_info oinfo = { { { 0 } } };
118 CDEBUG(D_INFO, "calling punch for "LPX64" (new size %Lu=%#Lx)\n",
119 lli->lli_smd->lsm_object_id, i_size_read(inode), i_size_read(inode));
121 oinfo.oi_md = lli->lli_smd;
122 oinfo.oi_policy.l_extent.start = new_size;
123 oinfo.oi_policy.l_extent.end = OBD_OBJECT_EOF;
125 oa.o_id = lli->lli_smd->lsm_object_id;
126 oa.o_gr = lli->lli_smd->lsm_object_gr;
127 oa.o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
129 /* set OBD_MD_FLFLAGS in o_valid, only if we
130 * set OBD_FL_TRUNCLOCK, otherwise ost_punch
131 * and filter_setattr get confused, see the comment
133 oa.o_flags = OBD_FL_TRUNCLOCK;
134 oa.o_valid |= OBD_MD_FLFLAGS;
136 obdo_from_inode(&oa, inode, OBD_MD_FLTYPE | OBD_MD_FLMODE |
137 OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME |
138 OBD_MD_FLFID | OBD_MD_FLGENER);
140 oinfo.oi_capa = ll_osscapa_get(inode, CAPA_OPC_OSS_TRUNC);
141 rc = obd_punch_rqset(ll_i2dtexp(inode), &oinfo, NULL);
142 ll_truncate_free_capa(oinfo.oi_capa);
144 CERROR("obd_truncate fails (%d) ino %lu\n", rc, inode->i_ino);
146 obdo_to_inode(inode, &oa, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
147 OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME);
151 /* this isn't where truncate starts. roughly:
152 * sys_truncate->ll_setattr_raw->vmtruncate->ll_truncate. setattr_raw grabs
153 * DLM lock on [size, EOF], i_mutex, ->lli_size_sem, and WRITE_I_ALLOC_SEM to
156 * must be called under ->lli_size_sem */
157 void ll_truncate(struct inode *inode)
159 struct ll_inode_info *lli = ll_i2info(inode);
160 int srvlock = !!(lli->lli_flags & LLIF_SRVLOCK);
163 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) to %Lu=%#Lx\n",inode->i_ino,
164 inode->i_generation, inode, i_size_read(inode),
167 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_TRUNC, 1);
168 if (lli->lli_size_sem_owner != current) {
174 CDEBUG(D_INODE, "truncate on inode %lu with no objects\n",
179 LASSERT(atomic_read(&lli->lli_size_sem.count) <= 0);
185 /* XXX I'm pretty sure this is a hack to paper
186 * over a more fundamental race condition. */
187 lov_stripe_lock(lli->lli_smd);
188 inode_init_lvb(inode, &lvb);
189 rc = obd_merge_lvb(ll_i2dtexp(inode), lli->lli_smd, &lvb, 0);
190 if (lvb.lvb_size == i_size_read(inode) && rc == 0) {
191 CDEBUG(D_VFSTRACE, "skipping punch for obj "LPX64
192 ",%Lu=%#Lx\n", lli->lli_smd->lsm_object_id,
193 i_size_read(inode), i_size_read(inode));
194 lov_stripe_unlock(lli->lli_smd);
197 obd_adjust_kms(ll_i2dtexp(inode), lli->lli_smd,
198 i_size_read(inode), 1);
199 lov_stripe_unlock(lli->lli_smd);
202 if (unlikely((ll_i2sbi(inode)->ll_flags & LL_SBI_CHECKSUM) &&
203 (i_size_read(inode) & ~CFS_PAGE_MASK))) {
204 /* If the truncate leaves behind a partial page, update its
206 struct page *page = find_get_page(inode->i_mapping,
207 i_size_read(inode) >>
210 struct ll_async_page *llap = llap_cast_private(page);
212 char *kaddr = kmap_atomic(page, KM_USER0);
213 llap->llap_checksum =
214 init_checksum(OSC_DEFAULT_CKSUM);
215 llap->llap_checksum =
216 compute_checksum(llap->llap_checksum,
217 kaddr, CFS_PAGE_SIZE,
219 kunmap_atomic(kaddr, KM_USER0);
221 page_cache_release(page);
225 new_size = i_size_read(inode);
226 ll_inode_size_unlock(inode, 0);
228 ll_file_punch(inode, new_size, 0);
230 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_LOCKLESS_TRUNC, 1);
236 ll_inode_size_unlock(inode, 0);
239 int ll_prepare_write(struct file *file, struct page *page, unsigned from,
242 struct inode *inode = page->mapping->host;
243 struct ll_inode_info *lli = ll_i2info(inode);
244 struct lov_stripe_md *lsm = lli->lli_smd;
245 obd_off offset = ((obd_off)page->index) << CFS_PAGE_SHIFT;
246 struct obd_info oinfo = { { { 0 } } };
253 LASSERT(PageLocked(page));
254 (void)llap_cast_private(page); /* assertion */
256 /* Check to see if we should return -EIO right away */
259 pga.count = CFS_PAGE_SIZE;
262 oa.o_mode = inode->i_mode;
263 oa.o_id = lsm->lsm_object_id;
264 oa.o_gr = lsm->lsm_object_gr;
265 oa.o_valid = OBD_MD_FLID | OBD_MD_FLMODE |
266 OBD_MD_FLTYPE | OBD_MD_FLGROUP;
267 obdo_from_inode(&oa, inode, OBD_MD_FLFID | OBD_MD_FLGENER);
271 rc = obd_brw(OBD_BRW_CHECK, ll_i2dtexp(inode), &oinfo, 1, &pga, NULL);
275 if (PageUptodate(page)) {
276 LL_CDEBUG_PAGE(D_PAGE, page, "uptodate\n");
280 /* We're completely overwriting an existing page, so _don't_ set it up
281 * to date until commit_write */
282 if (from == 0 && to == CFS_PAGE_SIZE) {
283 LL_CDEBUG_PAGE(D_PAGE, page, "full page write\n");
284 POISON_PAGE(page, 0x11);
288 /* If are writing to a new page, no need to read old data. The extent
289 * locking will have updated the KMS, and for our purposes here we can
290 * treat it like i_size. */
291 lov_stripe_lock(lsm);
292 inode_init_lvb(inode, &lvb);
293 obd_merge_lvb(ll_i2dtexp(inode), lsm, &lvb, 1);
294 lov_stripe_unlock(lsm);
295 if (lvb.lvb_size <= offset) {
296 char *kaddr = kmap_atomic(page, KM_USER0);
297 LL_CDEBUG_PAGE(D_PAGE, page, "kms "LPU64" <= offset "LPU64"\n",
298 lvb.lvb_size, offset);
299 memset(kaddr, 0, CFS_PAGE_SIZE);
300 kunmap_atomic(kaddr, KM_USER0);
301 GOTO(prepare_done, rc = 0);
304 /* XXX could be an async ocp read.. read-ahead? */
305 rc = ll_brw(OBD_BRW_READ, inode, &oa, page, 0);
307 /* bug 1598: don't clobber blksize */
308 oa.o_valid &= ~(OBD_MD_FLSIZE | OBD_MD_FLBLKSZ);
309 obdo_refresh_inode(inode, &oa, oa.o_valid);
315 SetPageUptodate(page);
321 * make page ready for ASYNC write
322 * \param data - pointer to llap cookie
323 * \param cmd - is OBD_BRW_* macroses
325 * \retval 0 is page successfully prepared to send
326 * \retval -EAGAIN is page not need to send
328 static int ll_ap_make_ready(void *data, int cmd)
330 struct ll_async_page *llap;
334 llap = LLAP_FROM_COOKIE(data);
335 page = llap->llap_page;
337 /* we're trying to write, but the page is locked.. come back later */
338 if (TryLockPage(page))
341 LASSERTF(!(cmd & OBD_BRW_READ) || !PageWriteback(page),
342 "cmd %x page %p ino %lu index %lu fl %lx\n", cmd, page,
343 page->mapping->host->i_ino, page->index, page->flags);
345 /* if we left PageDirty we might get another writepage call
346 * in the future. list walkers are bright enough
347 * to check page dirty so we can leave it on whatever list
348 * its on. XXX also, we're called with the cli list so if
349 * we got the page cache list we'd create a lock inversion
350 * with the removepage path which gets the page lock then the
352 LASSERTF(!PageWriteback(page),"cmd %x page %p ino %lu index %lu\n", cmd, page,
353 page->mapping->host->i_ino, page->index);
354 if(!clear_page_dirty_for_io(page)) {
359 /* This actually clears the dirty bit in the radix tree.*/
360 set_page_writeback(page);
362 LL_CDEBUG_PAGE(D_PAGE, page, "made ready\n");
363 page_cache_get(page);
368 /* We have two reasons for giving llite the opportunity to change the
369 * write length of a given queued page as it builds the RPC containing
372 * 1) Further extending writes may have landed in the page cache
373 * since a partial write first queued this page requiring us
374 * to write more from the page cache. (No further races are possible, since
375 * by the time this is called, the page is locked.)
376 * 2) We might have raced with truncate and want to avoid performing
377 * write RPCs that are just going to be thrown away by the
378 * truncate's punch on the storage targets.
380 * The kms serves these purposes as it is set at both truncate and extending
383 static int ll_ap_refresh_count(void *data, int cmd)
385 struct ll_inode_info *lli;
386 struct ll_async_page *llap;
387 struct lov_stripe_md *lsm;
394 /* readpage queues with _COUNT_STABLE, shouldn't get here. */
395 LASSERT(cmd != OBD_BRW_READ);
397 llap = LLAP_FROM_COOKIE(data);
398 page = llap->llap_page;
399 inode = page->mapping->host;
400 lli = ll_i2info(inode);
403 lov_stripe_lock(lsm);
404 inode_init_lvb(inode, &lvb);
405 obd_merge_lvb(ll_i2dtexp(inode), lsm, &lvb, 1);
407 lov_stripe_unlock(lsm);
409 /* catch race with truncate */
410 if (((__u64)page->index << CFS_PAGE_SHIFT) >= kms)
413 /* catch sub-page write at end of file */
414 if (((__u64)page->index << CFS_PAGE_SHIFT) + CFS_PAGE_SIZE > kms)
415 return kms % CFS_PAGE_SIZE;
417 return CFS_PAGE_SIZE;
420 void ll_inode_fill_obdo(struct inode *inode, int cmd, struct obdo *oa)
422 struct lov_stripe_md *lsm;
423 obd_flag valid_flags;
425 lsm = ll_i2info(inode)->lli_smd;
427 oa->o_id = lsm->lsm_object_id;
428 oa->o_gr = lsm->lsm_object_gr;
429 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
430 valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME;
431 if (cmd & OBD_BRW_WRITE) {
432 oa->o_valid |= OBD_MD_FLEPOCH;
433 oa->o_easize = ll_i2info(inode)->lli_ioepoch;
435 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME |
436 OBD_MD_FLUID | OBD_MD_FLGID |
437 OBD_MD_FLFID | OBD_MD_FLGENER;
440 obdo_from_inode(oa, inode, valid_flags);
443 static void ll_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
445 struct ll_async_page *llap;
448 llap = LLAP_FROM_COOKIE(data);
449 ll_inode_fill_obdo(llap->llap_page->mapping->host, cmd, oa);
454 static void ll_ap_update_obdo(void *data, int cmd, struct obdo *oa,
457 struct ll_async_page *llap;
460 llap = LLAP_FROM_COOKIE(data);
461 obdo_from_inode(oa, llap->llap_page->mapping->host, valid);
466 static struct obd_capa *ll_ap_lookup_capa(void *data, int cmd)
468 struct ll_async_page *llap = LLAP_FROM_COOKIE(data);
469 int opc = cmd & OBD_BRW_WRITE ? CAPA_OPC_OSS_WRITE : CAPA_OPC_OSS_RW;
471 return ll_osscapa_get(llap->llap_page->mapping->host, opc);
474 static struct obd_async_page_ops ll_async_page_ops = {
475 .ap_make_ready = ll_ap_make_ready,
476 .ap_refresh_count = ll_ap_refresh_count,
477 .ap_fill_obdo = ll_ap_fill_obdo,
478 .ap_update_obdo = ll_ap_update_obdo,
479 .ap_completion = ll_ap_completion,
480 .ap_lookup_capa = ll_ap_lookup_capa,
483 struct ll_async_page *llap_cast_private(struct page *page)
485 struct ll_async_page *llap = (struct ll_async_page *)page_private(page);
487 LASSERTF(llap == NULL || llap->llap_magic == LLAP_MAGIC,
488 "page %p private %lu gave magic %d which != %d\n",
489 page, page_private(page), llap->llap_magic, LLAP_MAGIC);
494 /* Try to shrink the page cache for the @sbi filesystem by 1/@shrink_fraction.
496 * There is an llap attached onto every page in lustre, linked off @sbi.
497 * We add an llap to the list so we don't lose our place during list walking.
498 * If llaps in the list are being moved they will only move to the end
499 * of the LRU, and we aren't terribly interested in those pages here (we
500 * start at the beginning of the list where the least-used llaps are.
502 int llap_shrink_cache(struct ll_sb_info *sbi, int shrink_fraction)
504 struct ll_async_page *llap, dummy_llap = { .llap_magic = 0xd11ad11a };
505 unsigned long total, want, count = 0;
507 total = sbi->ll_async_page_count;
509 /* There can be a large number of llaps (600k or more in a large
510 * memory machine) so the VM 1/6 shrink ratio is likely too much.
511 * Since we are freeing pages also, we don't necessarily want to
512 * shrink so much. Limit to 40MB of pages + llaps per call. */
513 if (shrink_fraction == 0)
514 want = sbi->ll_async_page_count - sbi->ll_async_page_max + 32;
516 want = (total + shrink_fraction - 1) / shrink_fraction;
518 if (want > 40 << (20 - CFS_PAGE_SHIFT))
519 want = 40 << (20 - CFS_PAGE_SHIFT);
521 CDEBUG(D_CACHE, "shrinking %lu of %lu pages (1/%d)\n",
522 want, total, shrink_fraction);
524 spin_lock(&sbi->ll_lock);
525 list_add(&dummy_llap.llap_pglist_item, &sbi->ll_pglist);
527 while (--total >= 0 && count < want) {
531 if (unlikely(need_resched())) {
532 spin_unlock(&sbi->ll_lock);
534 spin_lock(&sbi->ll_lock);
537 llap = llite_pglist_next_llap(sbi,&dummy_llap.llap_pglist_item);
538 list_del_init(&dummy_llap.llap_pglist_item);
542 page = llap->llap_page;
543 LASSERT(page != NULL);
545 list_add(&dummy_llap.llap_pglist_item, &llap->llap_pglist_item);
547 /* Page needs/undergoing IO */
548 if (TryLockPage(page)) {
549 LL_CDEBUG_PAGE(D_PAGE, page, "can't lock\n");
553 keep = (llap->llap_write_queued || PageDirty(page) ||
554 PageWriteback(page) || (!PageUptodate(page) &&
555 llap->llap_origin != LLAP_ORIGIN_READAHEAD));
557 LL_CDEBUG_PAGE(D_PAGE, page,"%s LRU page: %s%s%s%s%s origin %s\n",
558 keep ? "keep" : "drop",
559 llap->llap_write_queued ? "wq " : "",
560 PageDirty(page) ? "pd " : "",
561 PageUptodate(page) ? "" : "!pu ",
562 PageWriteback(page) ? "wb" : "",
563 llap->llap_defer_uptodate ? "" : "!du",
564 llap_origins[llap->llap_origin]);
566 /* If page is dirty or undergoing IO don't discard it */
572 page_cache_get(page);
573 spin_unlock(&sbi->ll_lock);
575 if (page->mapping != NULL) {
576 ll_teardown_mmaps(page->mapping,
577 (__u64)page->index << CFS_PAGE_SHIFT,
578 ((__u64)page->index << CFS_PAGE_SHIFT)|
580 if (!PageDirty(page) && !page_mapped(page)) {
581 ll_ra_accounting(llap, page->mapping);
582 ll_truncate_complete_page(page);
585 LL_CDEBUG_PAGE(D_PAGE, page, "Not dropping page"
593 page_cache_release(page);
595 spin_lock(&sbi->ll_lock);
597 list_del(&dummy_llap.llap_pglist_item);
598 spin_unlock(&sbi->ll_lock);
600 CDEBUG(D_CACHE, "shrank %lu/%lu and left %lu unscanned\n",
606 static struct ll_async_page *llap_from_page_with_lockh(struct page *page,
608 struct lustre_handle *lockh)
610 struct ll_async_page *llap;
611 struct obd_export *exp;
612 struct inode *inode = page->mapping->host;
613 struct ll_sb_info *sbi;
618 static int triggered;
621 LL_CDEBUG_PAGE(D_ERROR, page, "Bug 10047. Wrong anon "
623 libcfs_debug_dumpstack(NULL);
626 RETURN(ERR_PTR(-EINVAL));
628 sbi = ll_i2sbi(inode);
629 LASSERT(ll_async_page_slab);
630 LASSERTF(origin < LLAP__ORIGIN_MAX, "%u\n", origin);
632 llap = llap_cast_private(page);
634 /* move to end of LRU list, except when page is just about to
636 if (origin != LLAP_ORIGIN_REMOVEPAGE) {
637 spin_lock(&sbi->ll_lock);
638 sbi->ll_pglist_gen++;
639 list_del_init(&llap->llap_pglist_item);
640 list_add_tail(&llap->llap_pglist_item, &sbi->ll_pglist);
641 spin_unlock(&sbi->ll_lock);
646 exp = ll_i2dtexp(page->mapping->host);
648 RETURN(ERR_PTR(-EINVAL));
650 /* limit the number of lustre-cached pages */
651 if (sbi->ll_async_page_count >= sbi->ll_async_page_max)
652 llap_shrink_cache(sbi, 0);
654 OBD_SLAB_ALLOC(llap, ll_async_page_slab, CFS_ALLOC_STD,
655 ll_async_page_slab_size);
657 RETURN(ERR_PTR(-ENOMEM));
658 llap->llap_magic = LLAP_MAGIC;
659 llap->llap_cookie = (void *)llap + size_round(sizeof(*llap));
661 /* XXX: for bug 11270 - check for lockless origin here! */
662 if (origin == LLAP_ORIGIN_LOCKLESS_IO)
663 llap->llap_nocache = 1;
665 rc = obd_prep_async_page(exp, ll_i2info(inode)->lli_smd, NULL, page,
666 (obd_off)page->index << CFS_PAGE_SHIFT,
667 &ll_async_page_ops, llap, &llap->llap_cookie,
668 llap->llap_nocache, lockh);
670 OBD_SLAB_FREE(llap, ll_async_page_slab,
671 ll_async_page_slab_size);
675 CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n", llap,
676 page, llap->llap_cookie, (obd_off)page->index << CFS_PAGE_SHIFT);
677 /* also zeroing the PRIVBITS low order bitflags */
678 __set_page_ll_data(page, llap);
679 llap->llap_page = page;
680 spin_lock(&sbi->ll_lock);
681 sbi->ll_pglist_gen++;
682 sbi->ll_async_page_count++;
683 list_add_tail(&llap->llap_pglist_item, &sbi->ll_pglist);
684 INIT_LIST_HEAD(&llap->llap_pending_write);
685 spin_unlock(&sbi->ll_lock);
688 if (unlikely(sbi->ll_flags & LL_SBI_CHECKSUM)) {
690 char *kaddr = kmap_atomic(page, KM_USER0);
691 csum = init_checksum(OSC_DEFAULT_CKSUM);
692 csum = compute_checksum(csum, kaddr, CFS_PAGE_SIZE,
694 kunmap_atomic(kaddr, KM_USER0);
695 if (origin == LLAP_ORIGIN_READAHEAD ||
696 origin == LLAP_ORIGIN_READPAGE ||
697 origin == LLAP_ORIGIN_LOCKLESS_IO) {
698 llap->llap_checksum = 0;
699 } else if (origin == LLAP_ORIGIN_COMMIT_WRITE ||
700 llap->llap_checksum == 0) {
701 llap->llap_checksum = csum;
702 CDEBUG(D_PAGE, "page %p cksum %x\n", page, csum);
703 } else if (llap->llap_checksum == csum) {
704 /* origin == LLAP_ORIGIN_WRITEPAGE */
705 CDEBUG(D_PAGE, "page %p cksum %x confirmed\n",
708 /* origin == LLAP_ORIGIN_WRITEPAGE */
709 LL_CDEBUG_PAGE(D_ERROR, page, "old cksum %x != new "
710 "%x!\n", llap->llap_checksum, csum);
714 llap->llap_origin = origin;
718 struct ll_async_page *llap_from_page(struct page *page,
721 return llap_from_page_with_lockh(page, origin, NULL);
724 static int queue_or_sync_write(struct obd_export *exp, struct inode *inode,
725 struct ll_async_page *llap,
726 unsigned to, obd_flag async_flags)
728 unsigned long size_index = i_size_read(inode) >> CFS_PAGE_SHIFT;
729 struct obd_io_group *oig;
730 struct ll_sb_info *sbi = ll_i2sbi(inode);
731 int rc, noquot = llap->llap_ignore_quota ? OBD_BRW_NOQUOTA : 0;
734 /* _make_ready only sees llap once we've unlocked the page */
735 llap->llap_write_queued = 1;
736 rc = obd_queue_async_io(exp, ll_i2info(inode)->lli_smd, NULL,
737 llap->llap_cookie, OBD_BRW_WRITE | noquot,
738 0, 0, 0, async_flags);
740 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "write queued\n");
744 llap->llap_write_queued = 0;
745 /* Do not pass llap here as it is sync write. */
746 llap_write_pending(inode, NULL);
752 /* make full-page requests if we are not at EOF (bug 4410) */
753 if (to != CFS_PAGE_SIZE && llap->llap_page->index < size_index) {
754 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
755 "sync write before EOF: size_index %lu, to %d\n",
758 } else if (to != CFS_PAGE_SIZE && llap->llap_page->index == size_index) {
759 int size_to = i_size_read(inode) & ~CFS_PAGE_MASK;
760 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
761 "sync write at EOF: size_index %lu, to %d/%d\n",
762 size_index, to, size_to);
767 /* compare the checksum once before the page leaves llite */
768 if (unlikely((sbi->ll_flags & LL_SBI_CHECKSUM) &&
769 llap->llap_checksum != 0)) {
771 struct page *page = llap->llap_page;
772 char *kaddr = kmap_atomic(page, KM_USER0);
773 csum = init_checksum(OSC_DEFAULT_CKSUM);
774 csum = compute_checksum(csum, kaddr, CFS_PAGE_SIZE,
776 kunmap_atomic(kaddr, KM_USER0);
777 if (llap->llap_checksum == csum) {
778 CDEBUG(D_PAGE, "page %p cksum %x confirmed\n",
781 CERROR("page %p old cksum %x != new cksum %x!\n",
782 page, llap->llap_checksum, csum);
786 rc = obd_queue_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig,
787 llap->llap_cookie, OBD_BRW_WRITE | noquot,
788 0, to, 0, ASYNC_READY | ASYNC_URGENT |
789 ASYNC_COUNT_STABLE | ASYNC_GROUP_SYNC);
793 rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
799 if (!rc && async_flags & ASYNC_READY) {
800 unlock_page(llap->llap_page);
801 if (PageWriteback(llap->llap_page))
802 end_page_writeback(llap->llap_page);
805 if (rc == 0 && llap_write_complete(inode, llap))
806 ll_queue_done_writing(inode, 0);
808 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "sync write returned %d\n", rc);
816 /* update our write count to account for i_size increases that may have
817 * happened since we've queued the page for io. */
819 /* be careful not to return success without setting the page Uptodate or
820 * the next pass through prepare_write will read in stale data from disk. */
821 int ll_commit_write(struct file *file, struct page *page, unsigned from,
824 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
825 struct inode *inode = page->mapping->host;
826 struct ll_inode_info *lli = ll_i2info(inode);
827 struct lov_stripe_md *lsm = lli->lli_smd;
828 struct obd_export *exp;
829 struct ll_async_page *llap;
831 struct lustre_handle *lockh = NULL;
835 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
836 LASSERT(inode == file->f_dentry->d_inode);
837 LASSERT(PageLocked(page));
839 CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
840 inode, page, from, to, page->index);
842 if (fd->fd_flags & LL_FILE_GROUP_LOCKED)
843 lockh = &fd->fd_cwlockh;
845 llap = llap_from_page_with_lockh(page, LLAP_ORIGIN_COMMIT_WRITE, lockh);
847 RETURN(PTR_ERR(llap));
849 exp = ll_i2dtexp(inode);
853 llap->llap_ignore_quota = capable(CAP_SYS_RESOURCE);
856 * queue a write for some time in the future the first time we
859 * This is different from what other file systems do: they usually
860 * just mark page (and some of its buffers) dirty and rely on
861 * balance_dirty_pages() to start a write-back. Lustre wants write-back
862 * to be started earlier for the following reasons:
864 * (1) with a large number of clients we need to limit the amount
865 * of cached data on the clients a lot;
867 * (2) large compute jobs generally want compute-only then io-only
868 * and the IO should complete as quickly as possible;
870 * (3) IO is batched up to the RPC size and is async until the
871 * client max cache is hit
872 * (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
875 if (!PageDirty(page)) {
876 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRTY_MISSES, 1);
878 rc = queue_or_sync_write(exp, inode, llap, to, 0);
882 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRTY_HITS, 1);
885 /* put the page in the page cache, from now on ll_removepage is
886 * responsible for cleaning up the llap.
887 * only set page dirty when it's queued to be write out */
888 if (llap->llap_write_queued)
889 set_page_dirty(page);
892 size = (((obd_off)page->index) << CFS_PAGE_SHIFT) + to;
893 ll_inode_size_lock(inode, 0);
895 lov_stripe_lock(lsm);
896 obd_adjust_kms(exp, lsm, size, 0);
897 lov_stripe_unlock(lsm);
898 if (size > i_size_read(inode))
899 i_size_write(inode, size);
900 SetPageUptodate(page);
901 } else if (size > i_size_read(inode)) {
902 /* this page beyond the pales of i_size, so it can't be
903 * truncated in ll_p_r_e during lock revoking. we must
904 * teardown our book-keeping here. */
907 ll_inode_size_unlock(inode, 0);
911 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, unsigned long len)
913 struct ll_ra_info *ra = &sbi->ll_ra_info;
917 spin_lock(&sbi->ll_lock);
918 ret = min(ra->ra_max_pages - ra->ra_cur_pages, len);
919 ra->ra_cur_pages += ret;
920 spin_unlock(&sbi->ll_lock);
925 static void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
927 struct ll_ra_info *ra = &sbi->ll_ra_info;
928 spin_lock(&sbi->ll_lock);
929 LASSERTF(ra->ra_cur_pages >= len, "r_c_p %lu len %lu\n",
930 ra->ra_cur_pages, len);
931 ra->ra_cur_pages -= len;
932 spin_unlock(&sbi->ll_lock);
935 /* called for each page in a completed rpc.*/
936 int ll_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
938 struct ll_async_page *llap;
943 llap = LLAP_FROM_COOKIE(data);
944 page = llap->llap_page;
945 LASSERT(PageLocked(page));
946 LASSERT(CheckWriteback(page,cmd));
948 LL_CDEBUG_PAGE(D_PAGE, page, "completing cmd %d with %d\n", cmd, rc);
950 if (cmd & OBD_BRW_READ && llap->llap_defer_uptodate)
951 ll_ra_count_put(ll_i2sbi(page->mapping->host), 1);
954 if (cmd & OBD_BRW_READ) {
955 if (!llap->llap_defer_uptodate)
956 SetPageUptodate(page);
958 llap->llap_write_queued = 0;
960 ClearPageError(page);
962 if (cmd & OBD_BRW_READ) {
963 llap->llap_defer_uptodate = 0;
967 set_bit(AS_ENOSPC, &page->mapping->flags);
969 set_bit(AS_EIO, &page->mapping->flags);
972 /* be carefull about clear WB.
973 * if WB will cleared after page lock is released - paralel IO can be
974 * started before ap_make_ready is finished - so we will be have page
975 * with PG_Writeback set from ->writepage() and completed READ which
977 if ((cmd & OBD_BRW_WRITE) && PageWriteback(page))
978 end_page_writeback(page);
982 if (cmd & OBD_BRW_WRITE) {
983 /* Only rc == 0, write succeed, then this page could be deleted
984 * from the pending_writing list
986 if (rc == 0 && llap_write_complete(page->mapping->host, llap))
987 ll_queue_done_writing(page->mapping->host, 0);
990 page_cache_release(page);
995 static void __ll_put_llap(struct page *page)
997 struct inode *inode = page->mapping->host;
998 struct obd_export *exp;
999 struct ll_async_page *llap;
1000 struct ll_sb_info *sbi = ll_i2sbi(inode);
1004 exp = ll_i2dtexp(inode);
1006 CERROR("page %p ind %lu gave null export\n", page, page->index);
1011 llap = llap_from_page(page, LLAP_ORIGIN_REMOVEPAGE);
1013 CERROR("page %p ind %lu couldn't find llap: %ld\n", page,
1014 page->index, PTR_ERR(llap));
1019 if (llap_write_complete(inode, llap))
1020 ll_queue_done_writing(inode, 0);
1022 rc = obd_teardown_async_page(exp, ll_i2info(inode)->lli_smd, NULL,
1025 CERROR("page %p ind %lu failed: %d\n", page, page->index, rc);
1027 /* this unconditional free is only safe because the page lock
1028 * is providing exclusivity to memory pressure/truncate/writeback..*/
1029 __clear_page_ll_data(page);
1031 spin_lock(&sbi->ll_lock);
1032 if (!list_empty(&llap->llap_pglist_item))
1033 list_del_init(&llap->llap_pglist_item);
1034 sbi->ll_pglist_gen++;
1035 sbi->ll_async_page_count--;
1036 spin_unlock(&sbi->ll_lock);
1037 OBD_SLAB_FREE(llap, ll_async_page_slab, ll_async_page_slab_size);
1041 /* the kernel calls us here when a page is unhashed from the page cache.
1042 * the page will be locked and the kernel is holding a spinlock, so
1043 * we need to be careful. we're just tearing down our book-keeping
1045 void ll_removepage(struct page *page)
1047 struct ll_async_page *llap = llap_cast_private(page);
1050 LASSERT(!in_interrupt());
1052 /* sync pages or failed read pages can leave pages in the page
1053 * cache that don't have our data associated with them anymore */
1054 if (page_private(page) == 0) {
1059 LASSERT(!llap->llap_lockless_io_page);
1060 LASSERT(!llap->llap_nocache);
1061 LL_CDEBUG_PAGE(D_PAGE, page, "being evicted\n");
1062 __ll_put_llap(page);
1066 static int ll_issue_page_read(struct obd_export *exp,
1067 struct ll_async_page *llap,
1068 struct obd_io_group *oig, int defer)
1070 struct page *page = llap->llap_page;
1073 page_cache_get(page);
1074 llap->llap_defer_uptodate = defer;
1075 llap->llap_ra_used = 0;
1076 rc = obd_queue_group_io(exp, ll_i2info(page->mapping->host)->lli_smd,
1077 NULL, oig, llap->llap_cookie, OBD_BRW_READ, 0,
1078 CFS_PAGE_SIZE, 0, ASYNC_COUNT_STABLE |
1079 ASYNC_READY | ASYNC_URGENT);
1081 LL_CDEBUG_PAGE(D_ERROR, page, "read queue failed: rc %d\n", rc);
1082 page_cache_release(page);
1087 static void ll_ra_stats_inc_unlocked(struct ll_ra_info *ra, enum ra_stat which)
1089 LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
1090 ra->ra_stats[which]++;
1093 static void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which)
1095 struct ll_sb_info *sbi = ll_i2sbi(mapping->host);
1096 struct ll_ra_info *ra = &ll_i2sbi(mapping->host)->ll_ra_info;
1098 spin_lock(&sbi->ll_lock);
1099 ll_ra_stats_inc_unlocked(ra, which);
1100 spin_unlock(&sbi->ll_lock);
1103 void ll_ra_accounting(struct ll_async_page *llap, struct address_space *mapping)
1105 if (!llap->llap_defer_uptodate || llap->llap_ra_used)
1108 ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
1111 #define RAS_CDEBUG(ras) \
1113 "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu r %lu ri %lu" \
1114 "csr %lu sf %lu sp %lu sl %lu \n", \
1115 ras->ras_last_readpage, ras->ras_consecutive_requests, \
1116 ras->ras_consecutive_pages, ras->ras_window_start, \
1117 ras->ras_window_len, ras->ras_next_readahead, \
1118 ras->ras_requests, ras->ras_request_index, \
1119 ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
1120 ras->ras_stride_pages, ras->ras_stride_length)
1122 static int index_in_window(unsigned long index, unsigned long point,
1123 unsigned long before, unsigned long after)
1125 unsigned long start = point - before, end = point + after;
1132 return start <= index && index <= end;
1135 static struct ll_readahead_state *ll_ras_get(struct file *f)
1137 struct ll_file_data *fd;
1139 fd = LUSTRE_FPRIVATE(f);
1143 void ll_ra_read_in(struct file *f, struct ll_ra_read *rar)
1145 struct ll_readahead_state *ras;
1147 ras = ll_ras_get(f);
1149 spin_lock(&ras->ras_lock);
1150 ras->ras_requests++;
1151 ras->ras_request_index = 0;
1152 ras->ras_consecutive_requests++;
1153 rar->lrr_reader = current;
1155 list_add(&rar->lrr_linkage, &ras->ras_read_beads);
1156 spin_unlock(&ras->ras_lock);
1159 void ll_ra_read_ex(struct file *f, struct ll_ra_read *rar)
1161 struct ll_readahead_state *ras;
1163 ras = ll_ras_get(f);
1165 spin_lock(&ras->ras_lock);
1166 list_del_init(&rar->lrr_linkage);
1167 spin_unlock(&ras->ras_lock);
1170 static struct ll_ra_read *ll_ra_read_get_locked(struct ll_readahead_state *ras)
1172 struct ll_ra_read *scan;
1174 list_for_each_entry(scan, &ras->ras_read_beads, lrr_linkage) {
1175 if (scan->lrr_reader == current)
1181 struct ll_ra_read *ll_ra_read_get(struct file *f)
1183 struct ll_readahead_state *ras;
1184 struct ll_ra_read *bead;
1186 ras = ll_ras_get(f);
1188 spin_lock(&ras->ras_lock);
1189 bead = ll_ra_read_get_locked(ras);
1190 spin_unlock(&ras->ras_lock);
1194 static int ll_read_ahead_page(struct obd_export *exp, struct obd_io_group *oig,
1195 int index, struct address_space *mapping)
1197 struct ll_async_page *llap;
1199 unsigned int gfp_mask = 0;
1202 gfp_mask = GFP_HIGHUSER & ~__GFP_WAIT;
1204 gfp_mask |= __GFP_NOWARN;
1206 page = grab_cache_page_nowait_gfp(mapping, index, gfp_mask);
1208 ll_ra_stats_inc(mapping, RA_STAT_FAILED_GRAB_PAGE);
1209 CDEBUG(D_READA, "g_c_p_n failed\n");
1213 /* Check if page was truncated or reclaimed */
1214 if (page->mapping != mapping) {
1215 ll_ra_stats_inc(mapping, RA_STAT_WRONG_GRAB_PAGE);
1216 CDEBUG(D_READA, "g_c_p_n returned invalid page\n");
1217 GOTO(unlock_page, rc = 0);
1220 /* we do this first so that we can see the page in the /proc
1222 llap = llap_from_page(page, LLAP_ORIGIN_READAHEAD);
1223 if (IS_ERR(llap) || llap->llap_defer_uptodate) {
1224 if (PTR_ERR(llap) == -ENOLCK) {
1225 ll_ra_stats_inc(mapping, RA_STAT_FAILED_MATCH);
1226 CDEBUG(D_READA | D_PAGE,
1227 "Adding page to cache failed index "
1229 CDEBUG(D_READA, "nolock page\n");
1230 GOTO(unlock_page, rc = -ENOLCK);
1232 CDEBUG(D_READA, "read-ahead page\n");
1233 GOTO(unlock_page, rc = 0);
1236 /* skip completed pages */
1237 if (Page_Uptodate(page))
1238 GOTO(unlock_page, rc = 0);
1240 /* bail out when we hit the end of the lock. */
1241 rc = ll_issue_page_read(exp, llap, oig, 1);
1243 LL_CDEBUG_PAGE(D_READA | D_PAGE, page, "started read-ahead\n");
1248 LL_CDEBUG_PAGE(D_READA | D_PAGE, page, "skipping read-ahead\n");
1250 page_cache_release(page);
1254 /* ra_io_arg will be filled in the beginning of ll_readahead with
1255 * ras_lock, then the following ll_read_ahead_pages will read RA
1256 * pages according to this arg, all the items in this structure are
1257 * counted by page index.
1260 unsigned long ria_start; /* start offset of read-ahead*/
1261 unsigned long ria_end; /* end offset of read-ahead*/
1262 /* If stride read pattern is detected, ria_stoff means where
1263 * stride read is started. Note: for normal read-ahead, the
1264 * value here is meaningless, and also it will not be accessed*/
1266 /* ria_length and ria_pages are the length and pages length in the
1267 * stride I/O mode. And they will also be used to check whether
1268 * it is stride I/O read-ahead in the read-ahead pages*/
1269 unsigned long ria_length;
1270 unsigned long ria_pages;
1273 #define RIA_DEBUG(ria) \
1274 CDEBUG(D_READA, "rs %lu re %lu ro %lu rl %lu rp %lu\n", \
1275 ria->ria_start, ria->ria_end, ria->ria_stoff, ria->ria_length,\
1278 #define RAS_INCREASE_STEP (1024 * 1024 >> CFS_PAGE_SHIFT)
1280 static inline int stride_io_mode(struct ll_readahead_state *ras)
1282 return ras->ras_consecutive_stride_requests > 1;
1285 /* The function calculates how much pages will be read in
1286 * [off, off + length], which will be read by stride I/O mode,
1287 * stride_offset = st_off, stride_lengh = st_len,
1288 * stride_pages = st_pgs
1290 static unsigned long
1291 stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs,
1292 unsigned long off, unsigned length)
1294 unsigned long cont_len = st_off > off ? st_off - off : 0;
1295 __u64 stride_len = length + off > st_off ?
1296 length + off + 1 - st_off : 0;
1297 unsigned long left, pg_count;
1299 if (st_len == 0 || length == 0)
1302 left = do_div(stride_len, st_len);
1303 left = min(left, st_pgs);
1305 pg_count = left + stride_len * st_pgs + cont_len;
1307 LASSERT(pg_count >= left);
1309 CDEBUG(D_READA, "st_off %lu, st_len %lu st_pgs %lu off %lu length %u"
1310 "pgcount %lu\n", st_off, st_len, st_pgs, off, length, pg_count);
1315 static int ria_page_count(struct ra_io_arg *ria)
1317 __u64 length = ria->ria_end >= ria->ria_start ?
1318 ria->ria_end - ria->ria_start + 1 : 0;
1320 return stride_pg_count(ria->ria_stoff, ria->ria_length,
1321 ria->ria_pages, ria->ria_start,
1325 /*Check whether the index is in the defined ra-window */
1326 static int ras_inside_ra_window(unsigned long idx, struct ra_io_arg *ria)
1328 /* If ria_length == ria_pages, it means non-stride I/O mode,
1329 * idx should always inside read-ahead window in this case
1330 * For stride I/O mode, just check whether the idx is inside
1332 return ria->ria_length == 0 || ria->ria_length == ria->ria_pages ||
1333 (idx - ria->ria_stoff) % ria->ria_length < ria->ria_pages;
1336 static int ll_read_ahead_pages(struct obd_export *exp,
1337 struct obd_io_group *oig,
1338 struct ra_io_arg *ria,
1339 unsigned long *reserved_pages,
1340 struct address_space *mapping,
1341 unsigned long *ra_end)
1343 int rc, count = 0, stride_ria;
1344 unsigned long page_idx;
1346 LASSERT(ria != NULL);
1349 stride_ria = ria->ria_length > ria->ria_pages && ria->ria_pages > 0;
1350 for (page_idx = ria->ria_start; page_idx <= ria->ria_end &&
1351 *reserved_pages > 0; page_idx++) {
1352 if (ras_inside_ra_window(page_idx, ria)) {
1353 /* If the page is inside the read-ahead window*/
1354 rc = ll_read_ahead_page(exp, oig, page_idx, mapping);
1356 (*reserved_pages)--;
1358 } else if (rc == -ENOLCK)
1360 } else if (stride_ria) {
1361 /* If it is not in the read-ahead window, and it is
1362 * read-ahead mode, then check whether it should skip
1365 /* FIXME: This assertion only is valid when it is for
1366 * forward read-ahead, it will be fixed when backward
1367 * read-ahead is implemented */
1368 LASSERTF(page_idx > ria->ria_stoff, "since %lu in the"
1369 " gap of ra window,it should bigger than stride"
1370 " offset %lu \n", page_idx, ria->ria_stoff);
1372 offset = page_idx - ria->ria_stoff;
1373 offset = offset % (ria->ria_length);
1374 if (offset > ria->ria_pages) {
1375 page_idx += ria->ria_length - offset;
1376 CDEBUG(D_READA, "i %lu skip %lu \n", page_idx,
1377 ria->ria_length - offset);
1386 static int ll_readahead(struct ll_readahead_state *ras,
1387 struct obd_export *exp, struct address_space *mapping,
1388 struct obd_io_group *oig, int flags)
1390 unsigned long start = 0, end = 0, reserved;
1391 unsigned long ra_end, len;
1392 struct inode *inode;
1393 struct lov_stripe_md *lsm;
1394 struct ll_ra_read *bead;
1396 struct ra_io_arg ria = { 0 };
1401 inode = mapping->host;
1402 lsm = ll_i2info(inode)->lli_smd;
1404 lov_stripe_lock(lsm);
1405 inode_init_lvb(inode, &lvb);
1406 obd_merge_lvb(ll_i2dtexp(inode), lsm, &lvb, 1);
1408 lov_stripe_unlock(lsm);
1410 ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN);
1414 spin_lock(&ras->ras_lock);
1415 bead = ll_ra_read_get_locked(ras);
1416 /* Enlarge the RA window to encompass the full read */
1417 if (bead != NULL && ras->ras_window_start + ras->ras_window_len <
1418 bead->lrr_start + bead->lrr_count) {
1419 ras->ras_window_len = bead->lrr_start + bead->lrr_count -
1420 ras->ras_window_start;
1422 /* Reserve a part of the read-ahead window that we'll be issuing */
1423 if (ras->ras_window_len) {
1424 start = ras->ras_next_readahead;
1425 end = ras->ras_window_start + ras->ras_window_len - 1;
1428 /* Truncate RA window to end of file */
1429 end = min(end, (unsigned long)((kms - 1) >> CFS_PAGE_SHIFT));
1430 ras->ras_next_readahead = max(end, end + 1);
1433 ria.ria_start = start;
1435 /* If stride I/O mode is detected, get stride window*/
1436 if (stride_io_mode(ras)) {
1437 ria.ria_stoff = ras->ras_stride_offset;
1438 ria.ria_length = ras->ras_stride_length;
1439 ria.ria_pages = ras->ras_stride_pages;
1441 spin_unlock(&ras->ras_lock);
1444 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
1447 len = ria_page_count(&ria);
1451 reserved = ll_ra_count_get(ll_i2sbi(inode), len);
1454 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
1456 CDEBUG(D_READA, "reserved page %lu \n", reserved);
1458 ret = ll_read_ahead_pages(exp, oig, &ria, &reserved, mapping, &ra_end);
1460 LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
1462 ll_ra_count_put(ll_i2sbi(inode), reserved);
1464 if (ra_end == end + 1 && ra_end == (kms >> CFS_PAGE_SHIFT))
1465 ll_ra_stats_inc(mapping, RA_STAT_EOF);
1467 /* if we didn't get to the end of the region we reserved from
1468 * the ras we need to go back and update the ras so that the
1469 * next read-ahead tries from where we left off. we only do so
1470 * if the region we failed to issue read-ahead on is still ahead
1471 * of the app and behind the next index to start read-ahead from */
1472 CDEBUG(D_READA, "ra_end %lu end %lu stride end %lu \n",
1473 ra_end, end, ria.ria_end);
1475 if (ra_end != (end + 1)) {
1476 spin_lock(&ras->ras_lock);
1477 if (ra_end < ras->ras_next_readahead &&
1478 index_in_window(ra_end, ras->ras_window_start, 0,
1479 ras->ras_window_len)) {
1480 ras->ras_next_readahead = ra_end;
1483 spin_unlock(&ras->ras_lock);
1489 static void ras_set_start(struct ll_readahead_state *ras, unsigned long index)
1491 ras->ras_window_start = index & (~(RAS_INCREASE_STEP - 1));
1494 /* called with the ras_lock held or from places where it doesn't matter */
1495 static void ras_reset(struct ll_readahead_state *ras, unsigned long index)
1497 ras->ras_last_readpage = index;
1498 ras->ras_consecutive_requests = 0;
1499 ras->ras_consecutive_pages = 0;
1500 ras->ras_window_len = 0;
1501 ras_set_start(ras, index);
1502 ras->ras_next_readahead = max(ras->ras_window_start, index);
1507 /* called with the ras_lock held or from places where it doesn't matter */
1508 static void ras_stride_reset(struct ll_readahead_state *ras)
1510 ras->ras_consecutive_stride_requests = 0;
1514 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
1516 spin_lock_init(&ras->ras_lock);
1518 ras->ras_requests = 0;
1519 INIT_LIST_HEAD(&ras->ras_read_beads);
1522 /* Check whether the read request is in the stride window.
1523 * If it is in the stride window, return 1, otherwise return 0.
1524 * and also update stride_gap and stride_pages.
1526 static int index_in_stride_window(unsigned long index,
1527 struct ll_readahead_state *ras,
1528 struct inode *inode)
1530 int stride_gap = index - ras->ras_last_readpage - 1;
1532 LASSERT(stride_gap != 0);
1534 if (ras->ras_consecutive_pages == 0)
1537 /*Otherwise check the stride by itself */
1538 if ((ras->ras_stride_length - ras->ras_stride_pages) == stride_gap &&
1539 ras->ras_consecutive_pages == ras->ras_stride_pages)
1542 if (stride_gap >= 0) {
1544 * only set stride_pages, stride_length if
1545 * it is forward reading ( stride_gap > 0)
1547 ras->ras_stride_pages = ras->ras_consecutive_pages;
1548 ras->ras_stride_length = stride_gap + ras->ras_consecutive_pages;
1551 * If stride_gap < 0,(back_forward reading),
1552 * reset the stride_pages/length.
1553 * FIXME:back_ward stride I/O read.
1556 ras->ras_stride_pages = 0;
1557 ras->ras_stride_length = 0;
1564 static unsigned long
1565 stride_page_count(struct ll_readahead_state *ras, unsigned long len)
1567 return stride_pg_count(ras->ras_stride_offset, ras->ras_stride_length,
1568 ras->ras_stride_pages, ras->ras_stride_offset,
1572 /* Stride Read-ahead window will be increased inc_len according to
1573 * stride I/O pattern */
1574 static void ras_stride_increase_window(struct ll_readahead_state *ras,
1575 struct ll_ra_info *ra,
1576 unsigned long inc_len)
1578 unsigned long left, step, window_len;
1579 unsigned long stride_len;
1581 LASSERT(ras->ras_stride_length > 0);
1583 stride_len = ras->ras_window_start + ras->ras_window_len -
1584 ras->ras_stride_offset;
1586 LASSERTF(stride_len >= 0, "window_start %lu, window_len %lu"
1587 " stride_offset %lu\n", ras->ras_window_start,
1588 ras->ras_window_len, ras->ras_stride_offset);
1590 left = stride_len % ras->ras_stride_length;
1592 window_len = ras->ras_window_len - left;
1594 if (left < ras->ras_stride_pages)
1597 left = ras->ras_stride_pages + inc_len;
1599 LASSERT(ras->ras_stride_pages != 0);
1601 step = left / ras->ras_stride_pages;
1602 left %= ras->ras_stride_pages;
1604 window_len += step * ras->ras_stride_length + left;
1606 if (stride_page_count(ras, window_len) <= ra->ra_max_pages)
1607 ras->ras_window_len = window_len;
1612 /* Set stride I/O read-ahead window start offset */
1613 static void ras_set_stride_offset(struct ll_readahead_state *ras)
1615 unsigned long window_len = ras->ras_next_readahead -
1616 ras->ras_window_start;
1619 LASSERT(ras->ras_stride_length != 0);
1621 left = window_len % ras->ras_stride_length;
1623 ras->ras_stride_offset = ras->ras_next_readahead - left;
1628 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
1629 struct ll_readahead_state *ras, unsigned long index,
1632 struct ll_ra_info *ra = &sbi->ll_ra_info;
1633 int zero = 0, stride_zero = 0, stride_detect = 0, ra_miss = 0;
1636 spin_lock(&sbi->ll_lock);
1637 spin_lock(&ras->ras_lock);
1639 ll_ra_stats_inc_unlocked(ra, hit ? RA_STAT_HIT : RA_STAT_MISS);
1641 /* reset the read-ahead window in two cases. First when the app seeks
1642 * or reads to some other part of the file. Secondly if we get a
1643 * read-ahead miss that we think we've previously issued. This can
1644 * be a symptom of there being so many read-ahead pages that the VM is
1645 * reclaiming it before we get to it. */
1646 if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
1648 ll_ra_stats_inc_unlocked(ra, RA_STAT_DISTANT_READPAGE);
1649 /* check whether it is in stride I/O mode*/
1650 if (!index_in_stride_window(index, ras, inode))
1652 } else if (!hit && ras->ras_window_len &&
1653 index < ras->ras_next_readahead &&
1654 index_in_window(index, ras->ras_window_start, 0,
1655 ras->ras_window_len)) {
1658 /* If it hits read-ahead miss and the stride I/O is still
1659 * not detected, reset stride stuff to re-detect the whole
1660 * stride I/O mode to avoid complication */
1661 if (!stride_io_mode(ras))
1663 ll_ra_stats_inc_unlocked(ra, RA_STAT_MISS_IN_WINDOW);
1666 /* On the second access to a file smaller than the tunable
1667 * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1668 * file up to ra_max_pages. This is simply a best effort and
1669 * only occurs once per open file. Normal RA behavior is reverted
1670 * to for subsequent IO. The mmap case does not increment
1671 * ras_requests and thus can never trigger this behavior. */
1672 if (ras->ras_requests == 2 && !ras->ras_request_index) {
1675 kms_pages = (i_size_read(inode) + CFS_PAGE_SIZE - 1) >>
1678 CDEBUG(D_READA, "kmsp "LPU64" mwp %lu mp %lu\n", kms_pages,
1679 ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages);
1682 kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1683 ras->ras_window_start = 0;
1684 ras->ras_last_readpage = 0;
1685 ras->ras_next_readahead = 0;
1686 ras->ras_window_len = min(ra->ra_max_pages,
1687 ra->ra_max_read_ahead_whole_pages);
1688 GOTO(out_unlock, 0);
1693 /* If it is discontinuous read, check
1694 * whether it is stride I/O mode*/
1696 ras_reset(ras, index);
1697 ras->ras_consecutive_pages++;
1698 ras_stride_reset(ras);
1700 GOTO(out_unlock, 0);
1702 /* The read is still in stride window or
1703 * it hits read-ahead miss */
1705 /* If ra-window miss is hitted, which probably means VM
1706 * pressure, and some read-ahead pages were reclaimed.So
1707 * the length of ra-window will not increased, but also
1708 * not reset to avoid redetecting the stride I/O mode.*/
1709 ras->ras_consecutive_requests = 0;
1711 ras->ras_consecutive_pages = 0;
1712 if (++ras->ras_consecutive_stride_requests > 1)
1717 } else if (ras->ras_consecutive_stride_requests > 1) {
1718 /* If this is contiguous read but in stride I/O mode
1719 * currently, check whether stride step still is valid,
1720 * if invalid, it will reset the stride ra window*/
1721 if (ras->ras_consecutive_pages + 1 > ras->ras_stride_pages)
1722 ras_stride_reset(ras);
1725 ras->ras_last_readpage = index;
1726 ras->ras_consecutive_pages++;
1727 ras_set_start(ras, index);
1728 ras->ras_next_readahead = max(ras->ras_window_start,
1729 ras->ras_next_readahead);
1732 /* Trigger RA in the mmap case where ras_consecutive_requests
1733 * is not incremented and thus can't be used to trigger RA */
1734 if (!ras->ras_window_len && ras->ras_consecutive_pages == 4) {
1735 ras->ras_window_len = RAS_INCREASE_STEP;
1736 GOTO(out_unlock, 0);
1739 /* Initially reset the stride window offset to next_readahead*/
1740 if (ras->ras_consecutive_stride_requests == 2 && stride_detect)
1741 ras_set_stride_offset(ras);
1743 /* The initial ras_window_len is set to the request size. To avoid
1744 * uselessly reading and discarding pages for random IO the window is
1745 * only increased once per consecutive request received. */
1746 if ((ras->ras_consecutive_requests > 1 &&
1747 !ras->ras_request_index) || stride_detect) {
1748 if (stride_io_mode(ras))
1749 ras_stride_increase_window(ras, ra, RAS_INCREASE_STEP);
1751 ras->ras_window_len = min(ras->ras_window_len +
1758 ras->ras_request_index++;
1759 spin_unlock(&ras->ras_lock);
1760 spin_unlock(&sbi->ll_lock);
1764 int ll_writepage(struct page *page)
1766 struct inode *inode = page->mapping->host;
1767 struct ll_inode_info *lli = ll_i2info(inode);
1768 struct obd_export *exp;
1769 struct ll_async_page *llap;
1773 LASSERT(PageLocked(page));
1775 exp = ll_i2dtexp(inode);
1777 GOTO(out, rc = -EINVAL);
1779 llap = llap_from_page(page, LLAP_ORIGIN_WRITEPAGE);
1781 GOTO(out, rc = PTR_ERR(llap));
1783 LASSERT(!llap->llap_nocache);
1784 LASSERT(!PageWriteback(page));
1785 set_page_writeback(page);
1787 page_cache_get(page);
1788 if (llap->llap_write_queued) {
1789 LL_CDEBUG_PAGE(D_PAGE, page, "marking urgent\n");
1790 rc = obd_set_async_flags(exp, lli->lli_smd, NULL,
1792 ASYNC_READY | ASYNC_URGENT);
1794 rc = queue_or_sync_write(exp, inode, llap, CFS_PAGE_SIZE,
1795 ASYNC_READY | ASYNC_URGENT);
1798 /* re-dirty page on error so it retries write */
1799 if (PageWriteback(page))
1800 end_page_writeback(page);
1802 /* resend page only for not started IO*/
1803 if (!PageError(page))
1804 ll_redirty_page(page);
1806 page_cache_release(page);
1810 if (!lli->lli_async_rc)
1811 lli->lli_async_rc = rc;
1812 /* resend page only for not started IO*/
1819 * for now we do our readpage the same on both 2.4 and 2.5. The kernel's
1820 * read-ahead assumes it is valid to issue readpage all the way up to
1821 * i_size, but our dlm locks make that not the case. We disable the
1822 * kernel's read-ahead and do our own by walking ahead in the page cache
1823 * checking for dlm lock coverage. the main difference between 2.4 and
1824 * 2.6 is how read-ahead gets batched and issued, but we're using our own,
1825 * so they look the same.
1827 int ll_readpage(struct file *filp, struct page *page)
1829 struct ll_file_data *fd = LUSTRE_FPRIVATE(filp);
1830 struct inode *inode = page->mapping->host;
1831 struct obd_export *exp;
1832 struct ll_async_page *llap;
1833 struct obd_io_group *oig = NULL;
1834 struct lustre_handle *lockh = NULL;
1838 LASSERT(PageLocked(page));
1839 LASSERT(!PageUptodate(page));
1840 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset=%Lu=%#Lx\n",
1841 inode->i_ino, inode->i_generation, inode,
1842 (((loff_t)page->index) << CFS_PAGE_SHIFT),
1843 (((loff_t)page->index) << CFS_PAGE_SHIFT));
1844 LASSERT(atomic_read(&filp->f_dentry->d_inode->i_count) > 0);
1846 if (!ll_i2info(inode)->lli_smd) {
1847 /* File with no objects - one big hole */
1848 /* We use this just for remove_from_page_cache that is not
1849 * exported, we'd make page back up to date. */
1850 ll_truncate_complete_page(page);
1851 clear_page(kmap(page));
1853 SetPageUptodate(page);
1858 rc = oig_init(&oig);
1862 exp = ll_i2dtexp(inode);
1864 GOTO(out, rc = -EINVAL);
1866 if (fd->fd_flags & LL_FILE_GROUP_LOCKED)
1867 lockh = &fd->fd_cwlockh;
1869 llap = llap_from_page_with_lockh(page, LLAP_ORIGIN_READPAGE, lockh);
1871 if (PTR_ERR(llap) == -ENOLCK) {
1872 CWARN("ino %lu page %lu (%llu) not covered by "
1873 "a lock (mmap?). check debug logs.\n",
1874 inode->i_ino, page->index,
1875 (long long)page->index << PAGE_CACHE_SHIFT);
1877 GOTO(out, rc = PTR_ERR(llap));
1880 if (ll_i2sbi(inode)->ll_ra_info.ra_max_pages)
1881 ras_update(ll_i2sbi(inode), inode, &fd->fd_ras, page->index,
1882 llap->llap_defer_uptodate);
1885 if (llap->llap_defer_uptodate) {
1886 /* This is the callpath if we got the page from a readahead */
1887 llap->llap_ra_used = 1;
1888 rc = ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1891 obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd,
1893 LL_CDEBUG_PAGE(D_PAGE, page, "marking uptodate from defer\n");
1894 SetPageUptodate(page);
1896 GOTO(out_oig, rc = 0);
1899 rc = ll_issue_page_read(exp, llap, oig, 0);
1903 LL_CDEBUG_PAGE(D_PAGE, page, "queued readpage\n");
1904 /* We have just requested the actual page we want, see if we can tack
1905 * on some readahead to that page's RPC before it is sent. */
1906 if (ll_i2sbi(inode)->ll_ra_info.ra_max_pages)
1907 ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1910 rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
1921 static void ll_file_put_pages(struct page **pages, int numpages)
1927 for (i = 0, pp = pages; i < numpages; i++, pp++) {
1929 LL_CDEBUG_PAGE(D_PAGE, (*pp), "free\n");
1931 if (page_private(*pp))
1932 CERROR("the llap wasn't freed\n");
1933 (*pp)->mapping = NULL;
1934 if (page_count(*pp) != 1)
1935 CERROR("page %p, flags %#lx, count %i, private %p\n",
1936 (*pp), (unsigned long)(*pp)->flags, page_count(*pp),
1937 (void*)page_private(*pp));
1938 __free_pages(*pp, 0);
1941 OBD_FREE(pages, numpages * sizeof(struct page*));
1945 static struct page **ll_file_prepare_pages(int numpages, struct inode *inode,
1946 unsigned long first)
1948 struct page **pages;
1953 OBD_ALLOC(pages, sizeof(struct page *) * numpages);
1955 RETURN(ERR_PTR(-ENOMEM));
1956 for (i = 0; i < numpages; i++) {
1958 struct ll_async_page *llap;
1960 page = alloc_pages(GFP_HIGHUSER, 0);
1962 GOTO(err, rc = -ENOMEM);
1964 /* llap_from_page needs page index and mapping to be set */
1965 page->index = first++;
1966 page->mapping = inode->i_mapping;
1967 llap = llap_from_page(page, LLAP_ORIGIN_LOCKLESS_IO);
1969 GOTO(err, rc = PTR_ERR(llap));
1970 llap->llap_lockless_io_page = 1;
1974 ll_file_put_pages(pages, numpages);
1975 RETURN(ERR_PTR(rc));
1978 static ssize_t ll_file_copy_pages(struct page **pages, int numpages,
1979 char *buf, loff_t pos, size_t count, int rw)
1983 int updatechecksum = ll_i2sbi(pages[0]->mapping->host)->ll_flags &
1987 for (i = 0; i < numpages; i++) {
1988 unsigned offset, bytes, left;
1991 vaddr = kmap(pages[i]);
1992 offset = pos & (CFS_PAGE_SIZE - 1);
1993 bytes = min_t(unsigned, CFS_PAGE_SIZE - offset, count);
1994 LL_CDEBUG_PAGE(D_PAGE, pages[i], "op = %s, addr = %p, "
1995 "buf = %p, bytes = %u\n",
1996 (rw == WRITE) ? "CFU" : "CTU",
1997 vaddr + offset, buf, bytes);
1999 left = copy_from_user(vaddr + offset, buf, bytes);
2000 if (updatechecksum) {
2001 struct ll_async_page *llap;
2003 llap = llap_cast_private(pages[i]);
2004 llap->llap_checksum = crc32_le(0, vaddr,
2008 left = copy_to_user(buf, vaddr + offset, bytes);
2025 static int ll_file_oig_pages(struct inode * inode, struct page **pages,
2026 int numpages, loff_t pos, size_t count, int rw)
2028 struct obd_io_group *oig;
2029 struct ll_inode_info *lli = ll_i2info(inode);
2030 struct obd_export *exp;
2031 loff_t org_pos = pos;
2037 exp = ll_i2dtexp(inode);
2040 rc = oig_init(&oig);
2043 brw_flags = OBD_BRW_SRVLOCK;
2044 if (capable(CAP_SYS_RESOURCE))
2045 brw_flags |= OBD_BRW_NOQUOTA;
2047 for (i = 0; i < numpages; i++) {
2048 struct ll_async_page *llap;
2049 unsigned from, bytes;
2051 from = pos & (CFS_PAGE_SIZE - 1);
2052 bytes = min_t(unsigned, CFS_PAGE_SIZE - from,
2053 count - pos + org_pos);
2054 llap = llap_cast_private(pages[i]);
2057 lock_page(pages[i]);
2059 LL_CDEBUG_PAGE(D_PAGE, pages[i], "offset "LPU64","
2060 " from %u, bytes = %u\n",
2062 LASSERTF(pos >> CFS_PAGE_SHIFT == pages[i]->index,
2063 "wrong page index %lu (%lu)\n",
2065 (unsigned long)(pos >> CFS_PAGE_SHIFT));
2066 rc = obd_queue_group_io(exp, lli->lli_smd, NULL, oig,
2069 OBD_BRW_WRITE:OBD_BRW_READ,
2070 from, bytes, brw_flags,
2071 ASYNC_READY | ASYNC_URGENT |
2072 ASYNC_COUNT_STABLE | ASYNC_GROUP_SYNC);
2079 rc = obd_trigger_group_io(exp, lli->lli_smd, NULL, oig);
2085 unlock_page(pages[i]);
2090 ssize_t ll_file_lockless_io(struct file *file, char *buf, size_t count,
2091 loff_t *ppos, int rw)
2094 struct inode *inode = file->f_dentry->d_inode;
2098 unsigned long first, last;
2104 ll_inode_size_lock(inode, 0);
2105 isize = i_size_read(inode);
2106 ll_inode_size_unlock(inode, 0);
2109 if (*ppos + count >= isize)
2110 count -= *ppos + count - isize;
2114 rc = generic_write_checks(file, ppos, &count, 0);
2117 rc = remove_suid(file->f_dentry);
2122 first = pos >> CFS_PAGE_SHIFT;
2123 last = (pos + count - 1) >> CFS_PAGE_SHIFT;
2124 max_pages = PTLRPC_MAX_BRW_PAGES *
2125 ll_i2info(inode)->lli_smd->lsm_stripe_count;
2126 CDEBUG(D_INFO, "%u, stripe_count = %u\n",
2127 PTLRPC_MAX_BRW_PAGES /* max_pages_per_rpc */,
2128 ll_i2info(inode)->lli_smd->lsm_stripe_count);
2130 while (first <= last && rc >= 0) {
2132 struct page **pages;
2133 size_t bytes = count - amount;
2135 pages_for_io = min_t(int, last - first + 1, max_pages);
2136 pages = ll_file_prepare_pages(pages_for_io, inode, first);
2137 if (IS_ERR(pages)) {
2138 rc = PTR_ERR(pages);
2142 rc = ll_file_copy_pages(pages, pages_for_io, buf,
2143 pos + amount, bytes, rw);
2145 GOTO(put_pages, rc);
2148 rc = ll_file_oig_pages(inode, pages, pages_for_io,
2149 pos + amount, bytes, rw);
2151 GOTO(put_pages, rc);
2153 rc = ll_file_copy_pages(pages, pages_for_io, buf,
2154 pos + amount, bytes, rw);
2156 GOTO(put_pages, rc);
2162 ll_file_put_pages(pages, pages_for_io);
2163 first += pages_for_io;
2164 /* a short read/write check */
2165 if (pos + amount < ((loff_t)first << CFS_PAGE_SHIFT))
2168 /* NOTE: don't update i_size and KMS in absence of LDLM locks even
2169 * write makes the file large */
2170 file_accessed(file);
2171 if (rw == READ && amount < count && rc == 0) {
2172 unsigned long not_cleared;
2174 not_cleared = clear_user(buf, count - amount);
2175 amount = count - not_cleared;
2180 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
2182 LPROC_LL_LOCKLESS_WRITE :
2183 LPROC_LL_LOCKLESS_READ,