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);
320 static int ll_ap_make_ready(void *data, int cmd)
322 struct ll_async_page *llap;
326 llap = LLAP_FROM_COOKIE(data);
327 page = llap->llap_page;
329 LASSERTF(!(cmd & OBD_BRW_READ), "cmd %x page %p ino %lu index %lu\n", cmd, page,
330 page->mapping->host->i_ino, page->index);
332 /* we're trying to write, but the page is locked.. come back later */
333 if (TryLockPage(page))
336 LASSERT(!PageWriteback(page));
338 /* if we left PageDirty we might get another writepage call
339 * in the future. list walkers are bright enough
340 * to check page dirty so we can leave it on whatever list
341 * its on. XXX also, we're called with the cli list so if
342 * we got the page cache list we'd create a lock inversion
343 * with the removepage path which gets the page lock then the
345 LASSERTF(!PageWriteback(page),"cmd %x page %p ino %lu index %lu\n", cmd, page,
346 page->mapping->host->i_ino, page->index);
347 clear_page_dirty_for_io(page);
349 /* This actually clears the dirty bit in the radix tree.*/
350 set_page_writeback(page);
352 LL_CDEBUG_PAGE(D_PAGE, page, "made ready\n");
353 page_cache_get(page);
358 /* We have two reasons for giving llite the opportunity to change the
359 * write length of a given queued page as it builds the RPC containing
362 * 1) Further extending writes may have landed in the page cache
363 * since a partial write first queued this page requiring us
364 * to write more from the page cache. (No further races are possible, since
365 * by the time this is called, the page is locked.)
366 * 2) We might have raced with truncate and want to avoid performing
367 * write RPCs that are just going to be thrown away by the
368 * truncate's punch on the storage targets.
370 * The kms serves these purposes as it is set at both truncate and extending
373 static int ll_ap_refresh_count(void *data, int cmd)
375 struct ll_inode_info *lli;
376 struct ll_async_page *llap;
377 struct lov_stripe_md *lsm;
384 /* readpage queues with _COUNT_STABLE, shouldn't get here. */
385 LASSERT(cmd != OBD_BRW_READ);
387 llap = LLAP_FROM_COOKIE(data);
388 page = llap->llap_page;
389 inode = page->mapping->host;
390 lli = ll_i2info(inode);
393 lov_stripe_lock(lsm);
394 inode_init_lvb(inode, &lvb);
395 obd_merge_lvb(ll_i2dtexp(inode), lsm, &lvb, 1);
397 lov_stripe_unlock(lsm);
399 /* catch race with truncate */
400 if (((__u64)page->index << CFS_PAGE_SHIFT) >= kms)
403 /* catch sub-page write at end of file */
404 if (((__u64)page->index << CFS_PAGE_SHIFT) + CFS_PAGE_SIZE > kms)
405 return kms % CFS_PAGE_SIZE;
407 return CFS_PAGE_SIZE;
410 void ll_inode_fill_obdo(struct inode *inode, int cmd, struct obdo *oa)
412 struct lov_stripe_md *lsm;
413 obd_flag valid_flags;
415 lsm = ll_i2info(inode)->lli_smd;
417 oa->o_id = lsm->lsm_object_id;
418 oa->o_gr = lsm->lsm_object_gr;
419 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
420 valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME;
421 if (cmd & OBD_BRW_WRITE) {
422 oa->o_valid |= OBD_MD_FLEPOCH;
423 oa->o_easize = ll_i2info(inode)->lli_ioepoch;
425 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME |
426 OBD_MD_FLUID | OBD_MD_FLGID |
427 OBD_MD_FLFID | OBD_MD_FLGENER;
430 obdo_from_inode(oa, inode, valid_flags);
433 static void ll_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
435 struct ll_async_page *llap;
438 llap = LLAP_FROM_COOKIE(data);
439 ll_inode_fill_obdo(llap->llap_page->mapping->host, cmd, oa);
444 static void ll_ap_update_obdo(void *data, int cmd, struct obdo *oa,
447 struct ll_async_page *llap;
450 llap = LLAP_FROM_COOKIE(data);
451 obdo_from_inode(oa, llap->llap_page->mapping->host, valid);
456 static struct obd_capa *ll_ap_lookup_capa(void *data, int cmd)
458 struct ll_async_page *llap = LLAP_FROM_COOKIE(data);
459 int opc = cmd & OBD_BRW_WRITE ? CAPA_OPC_OSS_WRITE : CAPA_OPC_OSS_RW;
461 return ll_osscapa_get(llap->llap_page->mapping->host, opc);
464 static struct obd_async_page_ops ll_async_page_ops = {
465 .ap_make_ready = ll_ap_make_ready,
466 .ap_refresh_count = ll_ap_refresh_count,
467 .ap_fill_obdo = ll_ap_fill_obdo,
468 .ap_update_obdo = ll_ap_update_obdo,
469 .ap_completion = ll_ap_completion,
470 .ap_lookup_capa = ll_ap_lookup_capa,
473 struct ll_async_page *llap_cast_private(struct page *page)
475 struct ll_async_page *llap = (struct ll_async_page *)page_private(page);
477 LASSERTF(llap == NULL || llap->llap_magic == LLAP_MAGIC,
478 "page %p private %lu gave magic %d which != %d\n",
479 page, page_private(page), llap->llap_magic, LLAP_MAGIC);
484 /* Try to shrink the page cache for the @sbi filesystem by 1/@shrink_fraction.
486 * There is an llap attached onto every page in lustre, linked off @sbi.
487 * We add an llap to the list so we don't lose our place during list walking.
488 * If llaps in the list are being moved they will only move to the end
489 * of the LRU, and we aren't terribly interested in those pages here (we
490 * start at the beginning of the list where the least-used llaps are.
492 int llap_shrink_cache(struct ll_sb_info *sbi, int shrink_fraction)
494 struct ll_async_page *llap, dummy_llap = { .llap_magic = 0xd11ad11a };
495 unsigned long total, want, count = 0;
497 total = sbi->ll_async_page_count;
499 /* There can be a large number of llaps (600k or more in a large
500 * memory machine) so the VM 1/6 shrink ratio is likely too much.
501 * Since we are freeing pages also, we don't necessarily want to
502 * shrink so much. Limit to 40MB of pages + llaps per call. */
503 if (shrink_fraction == 0)
504 want = sbi->ll_async_page_count - sbi->ll_async_page_max + 32;
506 want = (total + shrink_fraction - 1) / shrink_fraction;
508 if (want > 40 << (20 - CFS_PAGE_SHIFT))
509 want = 40 << (20 - CFS_PAGE_SHIFT);
511 CDEBUG(D_CACHE, "shrinking %lu of %lu pages (1/%d)\n",
512 want, total, shrink_fraction);
514 spin_lock(&sbi->ll_lock);
515 list_add(&dummy_llap.llap_pglist_item, &sbi->ll_pglist);
517 while (--total >= 0 && count < want) {
521 if (unlikely(need_resched())) {
522 spin_unlock(&sbi->ll_lock);
524 spin_lock(&sbi->ll_lock);
527 llap = llite_pglist_next_llap(sbi,&dummy_llap.llap_pglist_item);
528 list_del_init(&dummy_llap.llap_pglist_item);
532 page = llap->llap_page;
533 LASSERT(page != NULL);
535 list_add(&dummy_llap.llap_pglist_item, &llap->llap_pglist_item);
537 /* Page needs/undergoing IO */
538 if (TryLockPage(page)) {
539 LL_CDEBUG_PAGE(D_PAGE, page, "can't lock\n");
543 keep = (llap->llap_write_queued || PageDirty(page) ||
544 PageWriteback(page) || (!PageUptodate(page) &&
545 llap->llap_origin != LLAP_ORIGIN_READAHEAD));
547 LL_CDEBUG_PAGE(D_PAGE, page,"%s LRU page: %s%s%s%s%s origin %s\n",
548 keep ? "keep" : "drop",
549 llap->llap_write_queued ? "wq " : "",
550 PageDirty(page) ? "pd " : "",
551 PageUptodate(page) ? "" : "!pu ",
552 PageWriteback(page) ? "wb" : "",
553 llap->llap_defer_uptodate ? "" : "!du",
554 llap_origins[llap->llap_origin]);
556 /* If page is dirty or undergoing IO don't discard it */
562 page_cache_get(page);
563 spin_unlock(&sbi->ll_lock);
565 if (page->mapping != NULL) {
566 ll_teardown_mmaps(page->mapping,
567 (__u64)page->index << CFS_PAGE_SHIFT,
568 ((__u64)page->index << CFS_PAGE_SHIFT)|
570 if (!PageDirty(page) && !page_mapped(page)) {
571 ll_ra_accounting(llap, page->mapping);
572 ll_truncate_complete_page(page);
575 LL_CDEBUG_PAGE(D_PAGE, page, "Not dropping page"
583 page_cache_release(page);
585 spin_lock(&sbi->ll_lock);
587 list_del(&dummy_llap.llap_pglist_item);
588 spin_unlock(&sbi->ll_lock);
590 CDEBUG(D_CACHE, "shrank %lu/%lu and left %lu unscanned\n",
596 struct ll_async_page *llap_from_page(struct page *page, unsigned origin)
598 struct ll_async_page *llap;
599 struct obd_export *exp;
600 struct inode *inode = page->mapping->host;
601 struct ll_sb_info *sbi;
606 static int triggered;
609 LL_CDEBUG_PAGE(D_ERROR, page, "Bug 10047. Wrong anon "
611 libcfs_debug_dumpstack(NULL);
614 RETURN(ERR_PTR(-EINVAL));
616 sbi = ll_i2sbi(inode);
617 LASSERT(ll_async_page_slab);
618 LASSERTF(origin < LLAP__ORIGIN_MAX, "%u\n", origin);
620 llap = llap_cast_private(page);
622 /* move to end of LRU list, except when page is just about to
624 if (origin != LLAP_ORIGIN_REMOVEPAGE) {
625 spin_lock(&sbi->ll_lock);
626 sbi->ll_pglist_gen++;
627 list_del_init(&llap->llap_pglist_item);
628 list_add_tail(&llap->llap_pglist_item, &sbi->ll_pglist);
629 spin_unlock(&sbi->ll_lock);
634 exp = ll_i2dtexp(page->mapping->host);
636 RETURN(ERR_PTR(-EINVAL));
638 /* limit the number of lustre-cached pages */
639 if (sbi->ll_async_page_count >= sbi->ll_async_page_max)
640 llap_shrink_cache(sbi, 0);
642 OBD_SLAB_ALLOC(llap, ll_async_page_slab, CFS_ALLOC_STD,
643 ll_async_page_slab_size);
645 RETURN(ERR_PTR(-ENOMEM));
646 llap->llap_magic = LLAP_MAGIC;
647 llap->llap_cookie = (void *)llap + size_round(sizeof(*llap));
649 rc = obd_prep_async_page(exp, ll_i2info(inode)->lli_smd, NULL, page,
650 (obd_off)page->index << CFS_PAGE_SHIFT,
651 &ll_async_page_ops, llap, &llap->llap_cookie);
653 OBD_SLAB_FREE(llap, ll_async_page_slab,
654 ll_async_page_slab_size);
658 CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n", llap,
659 page, llap->llap_cookie, (obd_off)page->index << CFS_PAGE_SHIFT);
660 /* also zeroing the PRIVBITS low order bitflags */
661 __set_page_ll_data(page, llap);
662 llap->llap_page = page;
663 spin_lock(&sbi->ll_lock);
664 sbi->ll_pglist_gen++;
665 sbi->ll_async_page_count++;
666 list_add_tail(&llap->llap_pglist_item, &sbi->ll_pglist);
667 INIT_LIST_HEAD(&llap->llap_pending_write);
668 spin_unlock(&sbi->ll_lock);
671 if (unlikely(sbi->ll_flags & LL_SBI_CHECKSUM)) {
673 char *kaddr = kmap_atomic(page, KM_USER0);
674 csum = init_checksum(OSC_DEFAULT_CKSUM);
675 csum = compute_checksum(csum, kaddr, CFS_PAGE_SIZE,
677 kunmap_atomic(kaddr, KM_USER0);
678 if (origin == LLAP_ORIGIN_READAHEAD ||
679 origin == LLAP_ORIGIN_READPAGE ||
680 origin == LLAP_ORIGIN_LOCKLESS_IO) {
681 llap->llap_checksum = 0;
682 } else if (origin == LLAP_ORIGIN_COMMIT_WRITE ||
683 llap->llap_checksum == 0) {
684 llap->llap_checksum = csum;
685 CDEBUG(D_PAGE, "page %p cksum %x\n", page, csum);
686 } else if (llap->llap_checksum == csum) {
687 /* origin == LLAP_ORIGIN_WRITEPAGE */
688 CDEBUG(D_PAGE, "page %p cksum %x confirmed\n",
691 /* origin == LLAP_ORIGIN_WRITEPAGE */
692 LL_CDEBUG_PAGE(D_ERROR, page, "old cksum %x != new "
693 "%x!\n", llap->llap_checksum, csum);
697 llap->llap_origin = origin;
701 static int queue_or_sync_write(struct obd_export *exp, struct inode *inode,
702 struct ll_async_page *llap,
703 unsigned to, obd_flag async_flags)
705 unsigned long size_index = i_size_read(inode) >> CFS_PAGE_SHIFT;
706 struct obd_io_group *oig;
707 struct ll_sb_info *sbi = ll_i2sbi(inode);
708 int rc, noquot = llap->llap_ignore_quota ? OBD_BRW_NOQUOTA : 0;
711 /* _make_ready only sees llap once we've unlocked the page */
712 llap->llap_write_queued = 1;
713 rc = obd_queue_async_io(exp, ll_i2info(inode)->lli_smd, NULL,
714 llap->llap_cookie, OBD_BRW_WRITE | noquot,
715 0, 0, 0, async_flags);
717 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "write queued\n");
721 llap->llap_write_queued = 0;
722 /* Do not pass llap here as it is sync write. */
723 llap_write_pending(inode, NULL);
729 /* make full-page requests if we are not at EOF (bug 4410) */
730 if (to != CFS_PAGE_SIZE && llap->llap_page->index < size_index) {
731 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
732 "sync write before EOF: size_index %lu, to %d\n",
735 } else if (to != CFS_PAGE_SIZE && llap->llap_page->index == size_index) {
736 int size_to = i_size_read(inode) & ~CFS_PAGE_MASK;
737 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
738 "sync write at EOF: size_index %lu, to %d/%d\n",
739 size_index, to, size_to);
744 /* compare the checksum once before the page leaves llite */
745 if (unlikely((sbi->ll_flags & LL_SBI_CHECKSUM) &&
746 llap->llap_checksum != 0)) {
748 struct page *page = llap->llap_page;
749 char *kaddr = kmap_atomic(page, KM_USER0);
750 csum = init_checksum(OSC_DEFAULT_CKSUM);
751 csum = compute_checksum(csum, kaddr, CFS_PAGE_SIZE,
753 kunmap_atomic(kaddr, KM_USER0);
754 if (llap->llap_checksum == csum) {
755 CDEBUG(D_PAGE, "page %p cksum %x confirmed\n",
758 CERROR("page %p old cksum %x != new cksum %x!\n",
759 page, llap->llap_checksum, csum);
763 rc = obd_queue_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig,
764 llap->llap_cookie, OBD_BRW_WRITE | noquot,
765 0, to, 0, ASYNC_READY | ASYNC_URGENT |
766 ASYNC_COUNT_STABLE | ASYNC_GROUP_SYNC);
770 rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
776 if (!rc && async_flags & ASYNC_READY) {
777 unlock_page(llap->llap_page);
778 if (PageWriteback(llap->llap_page)) {
779 end_page_writeback(llap->llap_page);
783 if (rc == 0 && llap_write_complete(inode, llap))
784 ll_queue_done_writing(inode, 0);
786 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "sync write returned %d\n", rc);
794 /* update our write count to account for i_size increases that may have
795 * happened since we've queued the page for io. */
797 /* be careful not to return success without setting the page Uptodate or
798 * the next pass through prepare_write will read in stale data from disk. */
799 int ll_commit_write(struct file *file, struct page *page, unsigned from,
802 struct inode *inode = page->mapping->host;
803 struct ll_inode_info *lli = ll_i2info(inode);
804 struct lov_stripe_md *lsm = lli->lli_smd;
805 struct obd_export *exp;
806 struct ll_async_page *llap;
811 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
812 LASSERT(inode == file->f_dentry->d_inode);
813 LASSERT(PageLocked(page));
815 CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
816 inode, page, from, to, page->index);
818 llap = llap_from_page(page, LLAP_ORIGIN_COMMIT_WRITE);
820 RETURN(PTR_ERR(llap));
822 exp = ll_i2dtexp(inode);
826 llap->llap_ignore_quota = capable(CAP_SYS_RESOURCE);
829 * queue a write for some time in the future the first time we
832 * This is different from what other file systems do: they usually
833 * just mark page (and some of its buffers) dirty and rely on
834 * balance_dirty_pages() to start a write-back. Lustre wants write-back
835 * to be started earlier for the following reasons:
837 * (1) with a large number of clients we need to limit the amount
838 * of cached data on the clients a lot;
840 * (2) large compute jobs generally want compute-only then io-only
841 * and the IO should complete as quickly as possible;
843 * (3) IO is batched up to the RPC size and is async until the
844 * client max cache is hit
845 * (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
848 if (!PageDirty(page)) {
849 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRTY_MISSES, 1);
851 rc = queue_or_sync_write(exp, inode, llap, to, 0);
855 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRTY_HITS, 1);
858 /* put the page in the page cache, from now on ll_removepage is
859 * responsible for cleaning up the llap.
860 * only set page dirty when it's queued to be write out */
861 if (llap->llap_write_queued)
862 set_page_dirty(page);
865 size = (((obd_off)page->index) << CFS_PAGE_SHIFT) + to;
866 ll_inode_size_lock(inode, 0);
868 lov_stripe_lock(lsm);
869 obd_adjust_kms(exp, lsm, size, 0);
870 lov_stripe_unlock(lsm);
871 if (size > i_size_read(inode))
872 i_size_write(inode, size);
873 SetPageUptodate(page);
874 } else if (size > i_size_read(inode)) {
875 /* this page beyond the pales of i_size, so it can't be
876 * truncated in ll_p_r_e during lock revoking. we must
877 * teardown our book-keeping here. */
880 ll_inode_size_unlock(inode, 0);
884 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, unsigned long len)
886 struct ll_ra_info *ra = &sbi->ll_ra_info;
890 spin_lock(&sbi->ll_lock);
891 ret = min(ra->ra_max_pages - ra->ra_cur_pages, len);
892 ra->ra_cur_pages += ret;
893 spin_unlock(&sbi->ll_lock);
898 static void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
900 struct ll_ra_info *ra = &sbi->ll_ra_info;
901 spin_lock(&sbi->ll_lock);
902 LASSERTF(ra->ra_cur_pages >= len, "r_c_p %lu len %lu\n",
903 ra->ra_cur_pages, len);
904 ra->ra_cur_pages -= len;
905 spin_unlock(&sbi->ll_lock);
908 /* called for each page in a completed rpc.*/
909 int ll_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
911 struct ll_async_page *llap;
916 llap = LLAP_FROM_COOKIE(data);
917 page = llap->llap_page;
918 LASSERT(PageLocked(page));
919 LASSERT(CheckWriteback(page,cmd));
921 LL_CDEBUG_PAGE(D_PAGE, page, "completing cmd %d with %d\n", cmd, rc);
923 if (cmd & OBD_BRW_READ && llap->llap_defer_uptodate)
924 ll_ra_count_put(ll_i2sbi(page->mapping->host), 1);
927 if (cmd & OBD_BRW_READ) {
928 if (!llap->llap_defer_uptodate)
929 SetPageUptodate(page);
931 llap->llap_write_queued = 0;
933 ClearPageError(page);
935 if (cmd & OBD_BRW_READ) {
936 llap->llap_defer_uptodate = 0;
940 set_bit(AS_ENOSPC, &page->mapping->flags);
942 set_bit(AS_EIO, &page->mapping->flags);
947 if (cmd & OBD_BRW_WRITE) {
948 /* Only rc == 0, write succeed, then this page could be deleted
949 * from the pending_writing list
951 if (rc == 0 && llap_write_complete(page->mapping->host, llap))
952 ll_queue_done_writing(page->mapping->host, 0);
955 if (PageWriteback(page)) {
956 end_page_writeback(page);
958 page_cache_release(page);
963 static void __ll_put_llap(struct page *page)
965 struct inode *inode = page->mapping->host;
966 struct obd_export *exp;
967 struct ll_async_page *llap;
968 struct ll_sb_info *sbi = ll_i2sbi(inode);
972 exp = ll_i2dtexp(inode);
974 CERROR("page %p ind %lu gave null export\n", page, page->index);
979 llap = llap_from_page(page, LLAP_ORIGIN_REMOVEPAGE);
981 CERROR("page %p ind %lu couldn't find llap: %ld\n", page,
982 page->index, PTR_ERR(llap));
987 if (llap_write_complete(inode, llap))
988 ll_queue_done_writing(inode, 0);
990 rc = obd_teardown_async_page(exp, ll_i2info(inode)->lli_smd, NULL,
993 CERROR("page %p ind %lu failed: %d\n", page, page->index, rc);
995 /* this unconditional free is only safe because the page lock
996 * is providing exclusivity to memory pressure/truncate/writeback..*/
997 __clear_page_ll_data(page);
999 spin_lock(&sbi->ll_lock);
1000 if (!list_empty(&llap->llap_pglist_item))
1001 list_del_init(&llap->llap_pglist_item);
1002 sbi->ll_pglist_gen++;
1003 sbi->ll_async_page_count--;
1004 spin_unlock(&sbi->ll_lock);
1005 OBD_SLAB_FREE(llap, ll_async_page_slab, ll_async_page_slab_size);
1009 /* the kernel calls us here when a page is unhashed from the page cache.
1010 * the page will be locked and the kernel is holding a spinlock, so
1011 * we need to be careful. we're just tearing down our book-keeping
1013 void ll_removepage(struct page *page)
1017 LASSERT(!in_interrupt());
1019 /* sync pages or failed read pages can leave pages in the page
1020 * cache that don't have our data associated with them anymore */
1021 if (page_private(page) == 0) {
1026 LASSERT(!llap_cast_private(page)->llap_lockless_io_page);
1027 LL_CDEBUG_PAGE(D_PAGE, page, "being evicted\n");
1028 __ll_put_llap(page);
1032 static int ll_page_matches(struct page *page, int fd_flags)
1034 struct lustre_handle match_lockh = {0};
1035 struct inode *inode = page->mapping->host;
1036 ldlm_policy_data_t page_extent;
1040 if (unlikely(fd_flags & LL_FILE_GROUP_LOCKED))
1043 page_extent.l_extent.start = (__u64)page->index << CFS_PAGE_SHIFT;
1044 page_extent.l_extent.end =
1045 page_extent.l_extent.start + CFS_PAGE_SIZE - 1;
1046 flags = LDLM_FL_TEST_LOCK | LDLM_FL_BLOCK_GRANTED;
1047 if (!(fd_flags & LL_FILE_READAHEAD))
1048 flags |= LDLM_FL_CBPENDING;
1049 matches = obd_match(ll_i2sbi(inode)->ll_dt_exp,
1050 ll_i2info(inode)->lli_smd, LDLM_EXTENT,
1051 &page_extent, LCK_PR | LCK_PW, &flags, inode,
1056 static int ll_issue_page_read(struct obd_export *exp,
1057 struct ll_async_page *llap,
1058 struct obd_io_group *oig, int defer)
1060 struct page *page = llap->llap_page;
1063 page_cache_get(page);
1064 llap->llap_defer_uptodate = defer;
1065 llap->llap_ra_used = 0;
1066 rc = obd_queue_group_io(exp, ll_i2info(page->mapping->host)->lli_smd,
1067 NULL, oig, llap->llap_cookie, OBD_BRW_READ, 0,
1068 CFS_PAGE_SIZE, 0, ASYNC_COUNT_STABLE |
1069 ASYNC_READY | ASYNC_URGENT);
1071 LL_CDEBUG_PAGE(D_ERROR, page, "read queue failed: rc %d\n", rc);
1072 page_cache_release(page);
1077 static void ll_ra_stats_inc_unlocked(struct ll_ra_info *ra, enum ra_stat which)
1079 LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
1080 ra->ra_stats[which]++;
1083 static void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which)
1085 struct ll_sb_info *sbi = ll_i2sbi(mapping->host);
1086 struct ll_ra_info *ra = &ll_i2sbi(mapping->host)->ll_ra_info;
1088 spin_lock(&sbi->ll_lock);
1089 ll_ra_stats_inc_unlocked(ra, which);
1090 spin_unlock(&sbi->ll_lock);
1093 void ll_ra_accounting(struct ll_async_page *llap, struct address_space *mapping)
1095 if (!llap->llap_defer_uptodate || llap->llap_ra_used)
1098 ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
1101 #define RAS_CDEBUG(ras) \
1103 "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu r %lu ri %lu" \
1104 "csr %lu sf %lu sp %lu sl %lu \n", \
1105 ras->ras_last_readpage, ras->ras_consecutive_requests, \
1106 ras->ras_consecutive_pages, ras->ras_window_start, \
1107 ras->ras_window_len, ras->ras_next_readahead, \
1108 ras->ras_requests, ras->ras_request_index, \
1109 ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
1110 ras->ras_stride_pages, ras->ras_stride_length)
1112 static int index_in_window(unsigned long index, unsigned long point,
1113 unsigned long before, unsigned long after)
1115 unsigned long start = point - before, end = point + after;
1122 return start <= index && index <= end;
1125 static struct ll_readahead_state *ll_ras_get(struct file *f)
1127 struct ll_file_data *fd;
1129 fd = LUSTRE_FPRIVATE(f);
1133 void ll_ra_read_in(struct file *f, struct ll_ra_read *rar)
1135 struct ll_readahead_state *ras;
1137 ras = ll_ras_get(f);
1139 spin_lock(&ras->ras_lock);
1140 ras->ras_requests++;
1141 ras->ras_request_index = 0;
1142 ras->ras_consecutive_requests++;
1143 rar->lrr_reader = current;
1145 list_add(&rar->lrr_linkage, &ras->ras_read_beads);
1146 spin_unlock(&ras->ras_lock);
1149 void ll_ra_read_ex(struct file *f, struct ll_ra_read *rar)
1151 struct ll_readahead_state *ras;
1153 ras = ll_ras_get(f);
1155 spin_lock(&ras->ras_lock);
1156 list_del_init(&rar->lrr_linkage);
1157 spin_unlock(&ras->ras_lock);
1160 static struct ll_ra_read *ll_ra_read_get_locked(struct ll_readahead_state *ras)
1162 struct ll_ra_read *scan;
1164 list_for_each_entry(scan, &ras->ras_read_beads, lrr_linkage) {
1165 if (scan->lrr_reader == current)
1171 struct ll_ra_read *ll_ra_read_get(struct file *f)
1173 struct ll_readahead_state *ras;
1174 struct ll_ra_read *bead;
1176 ras = ll_ras_get(f);
1178 spin_lock(&ras->ras_lock);
1179 bead = ll_ra_read_get_locked(ras);
1180 spin_unlock(&ras->ras_lock);
1184 static int ll_read_ahead_page(struct obd_export *exp, struct obd_io_group *oig,
1185 int index, struct address_space *mapping)
1187 struct ll_async_page *llap;
1189 unsigned int gfp_mask = 0;
1192 gfp_mask = GFP_HIGHUSER & ~__GFP_WAIT;
1194 gfp_mask |= __GFP_NOWARN;
1196 page = grab_cache_page_nowait_gfp(mapping, index, gfp_mask);
1198 ll_ra_stats_inc(mapping, RA_STAT_FAILED_GRAB_PAGE);
1199 CDEBUG(D_READA, "g_c_p_n failed\n");
1203 /* Check if page was truncated or reclaimed */
1204 if (page->mapping != mapping) {
1205 ll_ra_stats_inc(mapping, RA_STAT_WRONG_GRAB_PAGE);
1206 CDEBUG(D_READA, "g_c_p_n returned invalid page\n");
1207 GOTO(unlock_page, rc = 0);
1210 /* we do this first so that we can see the page in the /proc
1212 llap = llap_from_page(page, LLAP_ORIGIN_READAHEAD);
1213 if (IS_ERR(llap) || llap->llap_defer_uptodate) {
1214 if (PTR_ERR(llap) == -ENOLCK) {
1215 ll_ra_stats_inc(mapping, RA_STAT_FAILED_MATCH);
1216 CDEBUG(D_READA | D_PAGE,
1217 "Adding page to cache failed index "
1219 CDEBUG(D_READA, "nolock page\n");
1220 GOTO(unlock_page, rc = -ENOLCK);
1222 CDEBUG(D_READA, "read-ahead page\n");
1223 GOTO(unlock_page, rc = 0);
1226 /* skip completed pages */
1227 if (Page_Uptodate(page))
1228 GOTO(unlock_page, rc = 0);
1230 /* bail out when we hit the end of the lock. */
1231 rc = ll_issue_page_read(exp, llap, oig, 1);
1233 LL_CDEBUG_PAGE(D_READA | D_PAGE, page, "started read-ahead\n");
1238 LL_CDEBUG_PAGE(D_READA | D_PAGE, page, "skipping read-ahead\n");
1240 page_cache_release(page);
1244 /* ra_io_arg will be filled in the beginning of ll_readahead with
1245 * ras_lock, then the following ll_read_ahead_pages will read RA
1246 * pages according to this arg, all the items in this structure are
1247 * counted by page index.
1250 unsigned long ria_start; /* start offset of read-ahead*/
1251 unsigned long ria_end; /* end offset of read-ahead*/
1252 /* If stride read pattern is detected, ria_stoff means where
1253 * stride read is started. Note: for normal read-ahead, the
1254 * value here is meaningless, and also it will not be accessed*/
1256 /* ria_length and ria_pages are the length and pages length in the
1257 * stride I/O mode. And they will also be used to check whether
1258 * it is stride I/O read-ahead in the read-ahead pages*/
1259 unsigned long ria_length;
1260 unsigned long ria_pages;
1263 #define RIA_DEBUG(ria) \
1264 CDEBUG(D_READA, "rs %lu re %lu ro %lu rl %lu rp %lu\n", \
1265 ria->ria_start, ria->ria_end, ria->ria_stoff, ria->ria_length,\
1268 #define RAS_INCREASE_STEP (1024 * 1024 >> CFS_PAGE_SHIFT)
1270 static inline int stride_io_mode(struct ll_readahead_state *ras)
1272 return ras->ras_consecutive_stride_requests > 1;
1275 /* The function calculates how much pages will be read in
1276 * [off, off + length], which will be read by stride I/O mode,
1277 * stride_offset = st_off, stride_lengh = st_len,
1278 * stride_pages = st_pgs
1280 static unsigned long
1281 stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs,
1282 unsigned long off, unsigned length)
1284 unsigned long cont_len = st_off > off ? st_off - off : 0;
1285 unsigned long stride_len = length + off > st_off ?
1286 length + off + 1 - st_off : 0;
1287 unsigned long left, pg_count;
1289 if (st_len == 0 || length == 0)
1292 left = do_div(stride_len, st_len);
1293 left = min(left, st_pgs);
1295 pg_count = left + stride_len * st_pgs + cont_len;
1297 LASSERT(pg_count >= left);
1299 CDEBUG(D_READA, "st_off %lu, st_len %lu st_pgs %lu off %lu length %u"
1300 "pgcount %lu\n", st_off, st_len, st_pgs, off, length, pg_count);
1305 static int ria_page_count(struct ra_io_arg *ria)
1307 __u64 length = ria->ria_end >= ria->ria_start ?
1308 ria->ria_end - ria->ria_start + 1 : 0;
1310 return stride_pg_count(ria->ria_stoff, ria->ria_length,
1311 ria->ria_pages, ria->ria_start,
1315 /*Check whether the index is in the defined ra-window */
1316 static int ras_inside_ra_window(unsigned long idx, struct ra_io_arg *ria)
1318 /* If ria_length == ria_pages, it means non-stride I/O mode,
1319 * idx should always inside read-ahead window in this case
1320 * For stride I/O mode, just check whether the idx is inside
1322 return ria->ria_length == 0 || ria->ria_length == ria->ria_pages ||
1323 (idx - ria->ria_stoff) % ria->ria_length < ria->ria_pages;
1326 static int ll_read_ahead_pages(struct obd_export *exp,
1327 struct obd_io_group *oig,
1328 struct ra_io_arg *ria,
1329 unsigned long *reserved_pages,
1330 struct address_space *mapping,
1331 unsigned long *ra_end)
1333 int rc, count = 0, stride_ria;
1334 unsigned long page_idx;
1336 LASSERT(ria != NULL);
1339 stride_ria = ria->ria_length > ria->ria_pages && ria->ria_pages > 0;
1340 for (page_idx = ria->ria_start; page_idx <= ria->ria_end &&
1341 *reserved_pages > 0; page_idx++) {
1342 if (ras_inside_ra_window(page_idx, ria)) {
1343 /* If the page is inside the read-ahead window*/
1344 rc = ll_read_ahead_page(exp, oig, page_idx, mapping);
1346 (*reserved_pages)--;
1348 } else if (rc == -ENOLCK)
1350 } else if (stride_ria) {
1351 /* If it is not in the read-ahead window, and it is
1352 * read-ahead mode, then check whether it should skip
1355 /* FIXME: This assertion only is valid when it is for
1356 * forward read-ahead, it will be fixed when backward
1357 * read-ahead is implemented */
1358 LASSERTF(page_idx > ria->ria_stoff, "since %lu in the"
1359 " gap of ra window,it should bigger than stride"
1360 " offset %lu \n", page_idx, ria->ria_stoff);
1362 offset = page_idx - ria->ria_stoff;
1363 offset = offset % (ria->ria_length);
1364 if (offset > ria->ria_pages) {
1365 page_idx += ria->ria_length - offset;
1366 CDEBUG(D_READA, "i %lu skip %lu \n", page_idx,
1367 ria->ria_length - offset);
1376 static int ll_readahead(struct ll_readahead_state *ras,
1377 struct obd_export *exp, struct address_space *mapping,
1378 struct obd_io_group *oig, int flags)
1380 unsigned long start = 0, end = 0, reserved;
1381 unsigned long ra_end, len;
1382 struct inode *inode;
1383 struct lov_stripe_md *lsm;
1384 struct ll_ra_read *bead;
1386 struct ra_io_arg ria = { 0 };
1391 inode = mapping->host;
1392 lsm = ll_i2info(inode)->lli_smd;
1394 lov_stripe_lock(lsm);
1395 inode_init_lvb(inode, &lvb);
1396 obd_merge_lvb(ll_i2dtexp(inode), lsm, &lvb, 1);
1398 lov_stripe_unlock(lsm);
1400 ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN);
1404 spin_lock(&ras->ras_lock);
1405 bead = ll_ra_read_get_locked(ras);
1406 /* Enlarge the RA window to encompass the full read */
1407 if (bead != NULL && ras->ras_window_start + ras->ras_window_len <
1408 bead->lrr_start + bead->lrr_count) {
1409 ras->ras_window_len = bead->lrr_start + bead->lrr_count -
1410 ras->ras_window_start;
1412 /* Reserve a part of the read-ahead window that we'll be issuing */
1413 if (ras->ras_window_len) {
1414 start = ras->ras_next_readahead;
1415 end = ras->ras_window_start + ras->ras_window_len - 1;
1418 /* Truncate RA window to end of file */
1419 end = min(end, (unsigned long)((kms - 1) >> CFS_PAGE_SHIFT));
1420 ras->ras_next_readahead = max(end, end + 1);
1423 ria.ria_start = start;
1425 /* If stride I/O mode is detected, get stride window*/
1426 if (stride_io_mode(ras)) {
1427 ria.ria_length = ras->ras_stride_length;
1428 ria.ria_pages = ras->ras_stride_pages;
1430 spin_unlock(&ras->ras_lock);
1433 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
1436 len = ria_page_count(&ria);
1440 reserved = ll_ra_count_get(ll_i2sbi(inode), len);
1442 if (reserved < end - start + 1)
1443 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
1445 CDEBUG(D_READA, "reserved page %lu \n", reserved);
1447 ret = ll_read_ahead_pages(exp, oig, &ria, &reserved, mapping, &ra_end);
1449 LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
1451 ll_ra_count_put(ll_i2sbi(inode), reserved);
1453 if (ra_end == end + 1 && ra_end == (kms >> CFS_PAGE_SHIFT))
1454 ll_ra_stats_inc(mapping, RA_STAT_EOF);
1456 /* if we didn't get to the end of the region we reserved from
1457 * the ras we need to go back and update the ras so that the
1458 * next read-ahead tries from where we left off. we only do so
1459 * if the region we failed to issue read-ahead on is still ahead
1460 * of the app and behind the next index to start read-ahead from */
1461 CDEBUG(D_READA, "ra_end %lu end %lu stride end %lu \n",
1462 ra_end, end, ria.ria_end);
1464 if (ra_end != (end + 1)) {
1465 spin_lock(&ras->ras_lock);
1466 if (ra_end < ras->ras_next_readahead &&
1467 index_in_window(ra_end, ras->ras_window_start, 0,
1468 ras->ras_window_len)) {
1469 ras->ras_next_readahead = ra_end;
1472 spin_unlock(&ras->ras_lock);
1478 static void ras_set_start(struct ll_readahead_state *ras, unsigned long index)
1480 ras->ras_window_start = index & (~(RAS_INCREASE_STEP - 1));
1483 /* called with the ras_lock held or from places where it doesn't matter */
1484 static void ras_reset(struct ll_readahead_state *ras, unsigned long index)
1486 ras->ras_last_readpage = index;
1487 ras->ras_consecutive_requests = 0;
1488 ras->ras_consecutive_pages = 0;
1489 ras->ras_window_len = 0;
1490 ras_set_start(ras, index);
1491 ras->ras_next_readahead = max(ras->ras_window_start, index);
1496 /* called with the ras_lock held or from places where it doesn't matter */
1497 static void ras_stride_reset(struct ll_readahead_state *ras)
1499 ras->ras_consecutive_stride_requests = 0;
1503 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
1505 spin_lock_init(&ras->ras_lock);
1507 ras->ras_requests = 0;
1508 INIT_LIST_HEAD(&ras->ras_read_beads);
1511 /* Check whether the read request is in the stride window.
1512 * If it is in the stride window, return 1, otherwise return 0.
1513 * and also update stride_gap and stride_pages.
1515 static int index_in_stride_window(unsigned long index,
1516 struct ll_readahead_state *ras,
1517 struct inode *inode)
1519 int stride_gap = index - ras->ras_last_readpage - 1;
1521 LASSERT(stride_gap != 0);
1523 if (ras->ras_consecutive_pages == 0)
1526 /*Otherwise check the stride by itself */
1527 if ((ras->ras_stride_length - ras->ras_stride_pages) == stride_gap &&
1528 ras->ras_consecutive_pages == ras->ras_stride_pages)
1531 if (stride_gap >= 0) {
1533 * only set stride_pages, stride_length if
1534 * it is forward reading ( stride_gap > 0)
1536 ras->ras_stride_pages = ras->ras_consecutive_pages;
1537 ras->ras_stride_length = stride_gap + ras->ras_consecutive_pages;
1540 * If stride_gap < 0,(back_forward reading),
1541 * reset the stride_pages/length.
1542 * FIXME:back_ward stride I/O read.
1545 ras->ras_stride_pages = 0;
1546 ras->ras_stride_length = 0;
1553 static unsigned long
1554 stride_page_count(struct ll_readahead_state *ras, unsigned long len)
1556 return stride_pg_count(ras->ras_stride_offset, ras->ras_stride_length,
1557 ras->ras_stride_pages, ras->ras_stride_offset,
1561 /* Stride Read-ahead window will be increased inc_len according to
1562 * stride I/O pattern */
1563 static void ras_stride_increase_window(struct ll_readahead_state *ras,
1564 struct ll_ra_info *ra,
1565 unsigned long inc_len)
1567 unsigned long left, step, window_len;
1568 unsigned long stride_len;
1570 LASSERT(ras->ras_stride_length > 0);
1572 stride_len = ras->ras_window_start + ras->ras_window_len -
1573 ras->ras_stride_offset;
1575 LASSERTF(stride_len >= 0, "window_start %lu, window_len %lu"
1576 " stride_offset %lu\n", ras->ras_window_start,
1577 ras->ras_window_len, ras->ras_stride_offset);
1579 left = stride_len % ras->ras_stride_length;
1581 window_len = ras->ras_window_len - left;
1583 if (left < ras->ras_stride_pages)
1586 left = ras->ras_stride_pages + inc_len;
1588 LASSERT(ras->ras_stride_pages != 0);
1590 step = left / ras->ras_stride_pages;
1591 left %= ras->ras_stride_pages;
1593 window_len += step * ras->ras_stride_length + left;
1595 if (stride_page_count(ras, window_len) <= ra->ra_max_pages)
1596 ras->ras_window_len = window_len;
1601 /* Set stride I/O read-ahead window start offset */
1602 static void ras_set_stride_offset(struct ll_readahead_state *ras)
1604 unsigned long window_len = ras->ras_next_readahead -
1605 ras->ras_window_start;
1608 LASSERT(ras->ras_stride_length != 0);
1610 left = window_len % ras->ras_stride_length;
1612 ras->ras_stride_offset = ras->ras_next_readahead - left;
1617 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
1618 struct ll_readahead_state *ras, unsigned long index,
1621 struct ll_ra_info *ra = &sbi->ll_ra_info;
1622 int zero = 0, stride_zero = 0, stride_detect = 0, ra_miss = 0;
1625 spin_lock(&sbi->ll_lock);
1626 spin_lock(&ras->ras_lock);
1628 ll_ra_stats_inc_unlocked(ra, hit ? RA_STAT_HIT : RA_STAT_MISS);
1630 /* reset the read-ahead window in two cases. First when the app seeks
1631 * or reads to some other part of the file. Secondly if we get a
1632 * read-ahead miss that we think we've previously issued. This can
1633 * be a symptom of there being so many read-ahead pages that the VM is
1634 * reclaiming it before we get to it. */
1635 if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
1637 ll_ra_stats_inc_unlocked(ra, RA_STAT_DISTANT_READPAGE);
1638 /* check whether it is in stride I/O mode*/
1639 if (!index_in_stride_window(index, ras, inode))
1641 } else if (!hit && ras->ras_window_len &&
1642 index < ras->ras_next_readahead &&
1643 index_in_window(index, ras->ras_window_start, 0,
1644 ras->ras_window_len)) {
1647 /* If it hits read-ahead miss and the stride I/O is still
1648 * not detected, reset stride stuff to re-detect the whole
1649 * stride I/O mode to avoid complication */
1650 if (!stride_io_mode(ras))
1652 ll_ra_stats_inc_unlocked(ra, RA_STAT_MISS_IN_WINDOW);
1655 /* On the second access to a file smaller than the tunable
1656 * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1657 * file up to ra_max_pages. This is simply a best effort and
1658 * only occurs once per open file. Normal RA behavior is reverted
1659 * to for subsequent IO. The mmap case does not increment
1660 * ras_requests and thus can never trigger this behavior. */
1661 if (ras->ras_requests == 2 && !ras->ras_request_index) {
1664 kms_pages = (i_size_read(inode) + CFS_PAGE_SIZE - 1) >>
1667 CDEBUG(D_READA, "kmsp "LPU64" mwp %lu mp %lu\n", kms_pages,
1668 ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages);
1671 kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1672 ras->ras_window_start = 0;
1673 ras->ras_last_readpage = 0;
1674 ras->ras_next_readahead = 0;
1675 ras->ras_window_len = min(ra->ra_max_pages,
1676 ra->ra_max_read_ahead_whole_pages);
1677 GOTO(out_unlock, 0);
1682 /* If it is discontinuous read, check
1683 * whether it is stride I/O mode*/
1685 ras_reset(ras, index);
1686 ras->ras_consecutive_pages++;
1687 ras_stride_reset(ras);
1689 GOTO(out_unlock, 0);
1691 /* The read is still in stride window or
1692 * it hits read-ahead miss */
1694 /* If ra-window miss is hitted, which probably means VM
1695 * pressure, and some read-ahead pages were reclaimed.So
1696 * the length of ra-window will not increased, but also
1697 * not reset to avoid redetecting the stride I/O mode.*/
1698 ras->ras_consecutive_requests = 0;
1700 ras->ras_consecutive_pages = 0;
1701 if (++ras->ras_consecutive_stride_requests > 1)
1706 } else if (ras->ras_consecutive_stride_requests > 1) {
1707 /* If this is contiguous read but in stride I/O mode
1708 * currently, check whether stride step still is valid,
1709 * if invalid, it will reset the stride ra window*/
1710 if (ras->ras_consecutive_pages + 1 > ras->ras_stride_pages)
1711 ras_stride_reset(ras);
1714 ras->ras_last_readpage = index;
1715 ras->ras_consecutive_pages++;
1716 ras_set_start(ras, index);
1717 ras->ras_next_readahead = max(ras->ras_window_start,
1718 ras->ras_next_readahead);
1721 /* Trigger RA in the mmap case where ras_consecutive_requests
1722 * is not incremented and thus can't be used to trigger RA */
1723 if (!ras->ras_window_len && ras->ras_consecutive_pages == 4) {
1724 ras->ras_window_len = RAS_INCREASE_STEP;
1725 GOTO(out_unlock, 0);
1728 /* Initially reset the stride window offset to next_readahead*/
1729 if (ras->ras_consecutive_stride_requests == 2 && stride_detect)
1730 ras_set_stride_offset(ras);
1732 /* The initial ras_window_len is set to the request size. To avoid
1733 * uselessly reading and discarding pages for random IO the window is
1734 * only increased once per consecutive request received. */
1735 if ((ras->ras_consecutive_requests > 1 &&
1736 !ras->ras_request_index) || stride_detect) {
1737 if (stride_io_mode(ras))
1738 ras_stride_increase_window(ras, ra, RAS_INCREASE_STEP);
1740 ras->ras_window_len = min(ras->ras_window_len +
1747 ras->ras_request_index++;
1748 spin_unlock(&ras->ras_lock);
1749 spin_unlock(&sbi->ll_lock);
1753 int ll_writepage(struct page *page)
1755 struct inode *inode = page->mapping->host;
1756 struct ll_inode_info *lli = ll_i2info(inode);
1757 struct obd_export *exp;
1758 struct ll_async_page *llap;
1762 LASSERT(PageLocked(page));
1764 exp = ll_i2dtexp(inode);
1766 GOTO(out, rc = -EINVAL);
1768 llap = llap_from_page(page, LLAP_ORIGIN_WRITEPAGE);
1770 GOTO(out, rc = PTR_ERR(llap));
1772 LASSERT(!PageWriteback(page));
1773 set_page_writeback(page);
1775 page_cache_get(page);
1776 if (llap->llap_write_queued) {
1777 LL_CDEBUG_PAGE(D_PAGE, page, "marking urgent\n");
1778 rc = obd_set_async_flags(exp, lli->lli_smd, NULL,
1780 ASYNC_READY | ASYNC_URGENT);
1782 rc = queue_or_sync_write(exp, inode, llap, CFS_PAGE_SIZE,
1783 ASYNC_READY | ASYNC_URGENT);
1786 page_cache_release(page);
1789 if (!lli->lli_async_rc)
1790 lli->lli_async_rc = rc;
1791 /* re-dirty page on error so it retries write */
1792 if (PageWriteback(page)) {
1793 end_page_writeback(page);
1795 /* resend page only for not started IO*/
1796 if (!PageError(page))
1797 ll_redirty_page(page);
1804 * for now we do our readpage the same on both 2.4 and 2.5. The kernel's
1805 * read-ahead assumes it is valid to issue readpage all the way up to
1806 * i_size, but our dlm locks make that not the case. We disable the
1807 * kernel's read-ahead and do our own by walking ahead in the page cache
1808 * checking for dlm lock coverage. the main difference between 2.4 and
1809 * 2.6 is how read-ahead gets batched and issued, but we're using our own,
1810 * so they look the same.
1812 int ll_readpage(struct file *filp, struct page *page)
1814 struct ll_file_data *fd = LUSTRE_FPRIVATE(filp);
1815 struct inode *inode = page->mapping->host;
1816 struct obd_export *exp;
1817 struct ll_async_page *llap;
1818 struct obd_io_group *oig = NULL;
1822 LASSERT(PageLocked(page));
1823 LASSERT(!PageUptodate(page));
1824 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset=%Lu=%#Lx\n",
1825 inode->i_ino, inode->i_generation, inode,
1826 (((loff_t)page->index) << CFS_PAGE_SHIFT),
1827 (((loff_t)page->index) << CFS_PAGE_SHIFT));
1828 LASSERT(atomic_read(&filp->f_dentry->d_inode->i_count) > 0);
1830 if (!ll_i2info(inode)->lli_smd) {
1831 /* File with no objects - one big hole */
1832 /* We use this just for remove_from_page_cache that is not
1833 * exported, we'd make page back up to date. */
1834 ll_truncate_complete_page(page);
1835 clear_page(kmap(page));
1837 SetPageUptodate(page);
1842 rc = oig_init(&oig);
1846 exp = ll_i2dtexp(inode);
1848 GOTO(out, rc = -EINVAL);
1850 llap = llap_from_page(page, LLAP_ORIGIN_READPAGE);
1852 GOTO(out, rc = PTR_ERR(llap));
1854 if (ll_i2sbi(inode)->ll_ra_info.ra_max_pages)
1855 ras_update(ll_i2sbi(inode), inode, &fd->fd_ras, page->index,
1856 llap->llap_defer_uptodate);
1859 if (llap->llap_defer_uptodate) {
1860 /* This is the callpath if we got the page from a readahead */
1861 llap->llap_ra_used = 1;
1862 rc = ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1865 obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd,
1867 LL_CDEBUG_PAGE(D_PAGE, page, "marking uptodate from defer\n");
1868 SetPageUptodate(page);
1870 GOTO(out_oig, rc = 0);
1873 if (likely((fd->fd_flags & LL_FILE_IGNORE_LOCK) == 0)) {
1874 rc = ll_page_matches(page, fd->fd_flags);
1876 LL_CDEBUG_PAGE(D_ERROR, page,
1877 "lock match failed: rc %d\n", rc);
1882 CWARN("ino %lu page %lu (%llu) not covered by "
1883 "a lock (mmap?). check debug logs.\n",
1884 inode->i_ino, page->index,
1885 (long long)page->index << CFS_PAGE_SHIFT);
1889 rc = ll_issue_page_read(exp, llap, oig, 0);
1893 LL_CDEBUG_PAGE(D_PAGE, page, "queued readpage\n");
1894 /* We have just requested the actual page we want, see if we can tack
1895 * on some readahead to that page's RPC before it is sent. */
1896 if (ll_i2sbi(inode)->ll_ra_info.ra_max_pages)
1897 ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1900 rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
1911 static void ll_file_put_pages(struct page **pages, int numpages)
1917 for (i = 0, pp = pages; i < numpages; i++, pp++) {
1919 LL_CDEBUG_PAGE(D_PAGE, (*pp), "free\n");
1921 if (page_private(*pp))
1922 CERROR("the llap wasn't freed\n");
1923 (*pp)->mapping = NULL;
1924 if (page_count(*pp) != 1)
1925 CERROR("page %p, flags %#lx, count %i, private %p\n",
1926 (*pp), (unsigned long)(*pp)->flags, page_count(*pp),
1927 (void*)page_private(*pp));
1928 __free_pages(*pp, 0);
1931 OBD_FREE(pages, numpages * sizeof(struct page*));
1935 static struct page **ll_file_prepare_pages(int numpages, struct inode *inode,
1936 unsigned long first)
1938 struct page **pages;
1943 OBD_ALLOC(pages, sizeof(struct page *) * numpages);
1945 RETURN(ERR_PTR(-ENOMEM));
1946 for (i = 0; i < numpages; i++) {
1948 struct ll_async_page *llap;
1950 page = alloc_pages(GFP_HIGHUSER, 0);
1952 GOTO(err, rc = -ENOMEM);
1954 /* llap_from_page needs page index and mapping to be set */
1955 page->index = first++;
1956 page->mapping = inode->i_mapping;
1957 llap = llap_from_page(page, LLAP_ORIGIN_LOCKLESS_IO);
1959 GOTO(err, rc = PTR_ERR(llap));
1960 llap->llap_lockless_io_page = 1;
1964 ll_file_put_pages(pages, numpages);
1965 RETURN(ERR_PTR(rc));
1968 static ssize_t ll_file_copy_pages(struct page **pages, int numpages,
1969 char *buf, loff_t pos, size_t count, int rw)
1973 int updatechecksum = ll_i2sbi(pages[0]->mapping->host)->ll_flags &
1977 for (i = 0; i < numpages; i++) {
1978 unsigned offset, bytes, left;
1981 vaddr = kmap(pages[i]);
1982 offset = pos & (CFS_PAGE_SIZE - 1);
1983 bytes = min_t(unsigned, CFS_PAGE_SIZE - offset, count);
1984 LL_CDEBUG_PAGE(D_PAGE, pages[i], "op = %s, addr = %p, "
1985 "buf = %p, bytes = %u\n",
1986 (rw == WRITE) ? "CFU" : "CTU",
1987 vaddr + offset, buf, bytes);
1989 left = copy_from_user(vaddr + offset, buf, bytes);
1990 if (updatechecksum) {
1991 struct ll_async_page *llap;
1993 llap = llap_cast_private(pages[i]);
1994 llap->llap_checksum = crc32_le(0, vaddr,
1998 left = copy_to_user(buf, vaddr + offset, bytes);
2015 static int ll_file_oig_pages(struct inode * inode, struct page **pages,
2016 int numpages, loff_t pos, size_t count, int rw)
2018 struct obd_io_group *oig;
2019 struct ll_inode_info *lli = ll_i2info(inode);
2020 struct obd_export *exp;
2021 loff_t org_pos = pos;
2027 exp = ll_i2dtexp(inode);
2030 rc = oig_init(&oig);
2033 brw_flags = OBD_BRW_SRVLOCK;
2034 if (capable(CAP_SYS_RESOURCE))
2035 brw_flags |= OBD_BRW_NOQUOTA;
2037 for (i = 0; i < numpages; i++) {
2038 struct ll_async_page *llap;
2039 unsigned from, bytes;
2041 from = pos & (CFS_PAGE_SIZE - 1);
2042 bytes = min_t(unsigned, CFS_PAGE_SIZE - from,
2043 count - pos + org_pos);
2044 llap = llap_cast_private(pages[i]);
2047 lock_page(pages[i]);
2049 LL_CDEBUG_PAGE(D_PAGE, pages[i], "offset "LPU64","
2050 " from %u, bytes = %u\n",
2052 LASSERTF(pos >> CFS_PAGE_SHIFT == pages[i]->index,
2053 "wrong page index %lu (%lu)\n",
2055 (unsigned long)(pos >> CFS_PAGE_SHIFT));
2056 rc = obd_queue_group_io(exp, lli->lli_smd, NULL, oig,
2059 OBD_BRW_WRITE:OBD_BRW_READ,
2060 from, bytes, brw_flags,
2061 ASYNC_READY | ASYNC_URGENT |
2062 ASYNC_COUNT_STABLE | ASYNC_GROUP_SYNC);
2069 rc = obd_trigger_group_io(exp, lli->lli_smd, NULL, oig);
2075 unlock_page(pages[i]);
2080 ssize_t ll_file_lockless_io(struct file *file, char *buf, size_t count,
2081 loff_t *ppos, int rw)
2084 struct inode *inode = file->f_dentry->d_inode;
2088 unsigned long first, last;
2094 ll_inode_size_lock(inode, 0);
2095 isize = i_size_read(inode);
2096 ll_inode_size_unlock(inode, 0);
2099 if (*ppos + count >= isize)
2100 count -= *ppos + count - isize;
2104 rc = generic_write_checks(file, ppos, &count, 0);
2107 rc = remove_suid(file->f_dentry);
2112 first = pos >> CFS_PAGE_SHIFT;
2113 last = (pos + count - 1) >> CFS_PAGE_SHIFT;
2114 max_pages = PTLRPC_MAX_BRW_PAGES *
2115 ll_i2info(inode)->lli_smd->lsm_stripe_count;
2116 CDEBUG(D_INFO, "%u, stripe_count = %u\n",
2117 PTLRPC_MAX_BRW_PAGES /* max_pages_per_rpc */,
2118 ll_i2info(inode)->lli_smd->lsm_stripe_count);
2120 while (first <= last && rc >= 0) {
2122 struct page **pages;
2123 size_t bytes = count - amount;
2125 pages_for_io = min_t(int, last - first + 1, max_pages);
2126 pages = ll_file_prepare_pages(pages_for_io, inode, first);
2127 if (IS_ERR(pages)) {
2128 rc = PTR_ERR(pages);
2132 rc = ll_file_copy_pages(pages, pages_for_io, buf,
2133 pos + amount, bytes, rw);
2135 GOTO(put_pages, rc);
2138 rc = ll_file_oig_pages(inode, pages, pages_for_io,
2139 pos + amount, bytes, rw);
2141 GOTO(put_pages, rc);
2143 rc = ll_file_copy_pages(pages, pages_for_io, buf,
2144 pos + amount, bytes, rw);
2146 GOTO(put_pages, rc);
2152 ll_file_put_pages(pages, pages_for_io);
2153 first += pages_for_io;
2154 /* a short read/write check */
2155 if (pos + amount < ((loff_t)first << CFS_PAGE_SHIFT))
2158 /* NOTE: don't update i_size and KMS in absence of LDLM locks even
2159 * write makes the file large */
2160 file_accessed(file);
2161 if (rw == READ && amount < count && rc == 0) {
2162 unsigned long not_cleared;
2164 not_cleared = clear_user(buf, count - amount);
2165 amount = count - not_cleared;
2170 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
2172 LPROC_LL_LOCKLESS_WRITE :
2173 LPROC_LL_LOCKLESS_READ,