1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/obdfilter/filter_io.c
38 * Author: Peter Braam <braam@clusterfs.com>
39 * Author: Andreas Dilger <adilger@clusterfs.com>
40 * Author: Phil Schwan <phil@clusterfs.com>
43 #define DEBUG_SUBSYSTEM S_FILTER
45 #ifndef AUTOCONF_INCLUDED
46 #include <linux/config.h>
48 #include <linux/module.h>
49 #include <linux/pagemap.h> // XXX kill me soon
50 #include <linux/version.h>
52 #include <obd_class.h>
54 #include <lustre_fsfilt.h>
55 #include "filter_internal.h"
57 int *obdfilter_created_scratchpad;
59 /* Grab the dirty and seen grant announcements from the incoming obdo.
60 * We will later calculate the clients new grant and return it.
61 * Caller must hold osfs lock */
62 void filter_grant_incoming(struct obd_export *exp, struct obdo *oa)
64 struct filter_export_data *fed;
65 struct obd_device *obd = exp->exp_obd;
68 LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
70 if ((oa->o_valid & (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) !=
71 (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) {
72 oa->o_valid &= ~OBD_MD_FLGRANT;
77 fed = &exp->exp_filter_data;
79 /* Add some margin, since there is a small race if other RPCs arrive
80 * out-or-order and have already consumed some grant. We want to
81 * leave this here in case there is a large error in accounting. */
83 "%s: cli %s/%p reports grant: "LPU64" dropped: %u, local: %lu\n",
84 obd->obd_name, exp->exp_client_uuid.uuid, exp, oa->o_grant,
85 oa->o_dropped, fed->fed_grant);
87 /* Update our accounting now so that statfs takes it into account.
88 * Note that fed_dirty is only approximate and can become incorrect
89 * if RPCs arrive out-of-order. No important calculations depend
90 * on fed_dirty however, but we must check sanity to not assert. */
91 if ((long long)oa->o_dirty < 0)
93 else if (oa->o_dirty > fed->fed_grant + 4 * FILTER_GRANT_CHUNK)
94 oa->o_dirty = fed->fed_grant + 4 * FILTER_GRANT_CHUNK;
95 obd->u.filter.fo_tot_dirty += oa->o_dirty - fed->fed_dirty;
96 if (fed->fed_grant < oa->o_dropped) {
97 CDEBUG(D_CACHE,"%s: cli %s/%p reports %u dropped > grant %lu\n",
98 obd->obd_name, exp->exp_client_uuid.uuid, exp,
99 oa->o_dropped, fed->fed_grant);
102 if (obd->u.filter.fo_tot_granted < oa->o_dropped) {
103 CERROR("%s: cli %s/%p reports %u dropped > tot_grant "LPU64"\n",
104 obd->obd_name, exp->exp_client_uuid.uuid, exp,
105 oa->o_dropped, obd->u.filter.fo_tot_granted);
108 obd->u.filter.fo_tot_granted -= oa->o_dropped;
109 fed->fed_grant -= oa->o_dropped;
110 fed->fed_dirty = oa->o_dirty;
112 if (oa->o_flags & OBD_FL_SHRINK_GRANT) {
113 obd_size left_space = filter_grant_space_left(exp);
114 struct filter_obd *filter = &exp->exp_obd->u.filter;
116 /*Only if left_space < fo_tot_clients * 32M,
117 *then the grant space could be shrinked */
118 if (left_space < filter->fo_tot_granted_clients *
119 FILTER_GRANT_SHRINK_LIMIT) {
120 fed->fed_grant -= oa->o_grant;
121 filter->fo_tot_granted -= oa->o_grant;
122 CDEBUG(D_CACHE, "%s: cli %s/%p shrink "LPU64
123 "fed_grant %ld total "LPU64"\n",
124 obd->obd_name, exp->exp_client_uuid.uuid,
125 exp, oa->o_grant, fed->fed_grant,
126 filter->fo_tot_granted);
131 if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
132 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
133 obd->obd_name, exp->exp_client_uuid.uuid, exp,
134 fed->fed_dirty, fed->fed_pending, fed->fed_grant);
135 spin_unlock(&obd->obd_osfs_lock);
141 /* Figure out how much space is available between what we've granted
142 * and what remains in the filesystem. Compensate for ext3 indirect
143 * block overhead when computing how much free space is left ungranted.
145 * Caller must hold obd_osfs_lock. */
146 obd_size filter_grant_space_left(struct obd_export *exp)
148 struct obd_device *obd = exp->exp_obd;
149 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
150 obd_size tot_granted = obd->u.filter.fo_tot_granted, avail, left = 0;
151 int rc, statfs_done = 0;
153 LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
155 if (cfs_time_before_64(obd->obd_osfs_age, cfs_time_current_64() - HZ)) {
157 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb,
158 cfs_time_current_64() + HZ);
159 if (rc) /* N.B. statfs can't really fail */
164 avail = obd->obd_osfs.os_bavail;
165 left = avail - (avail >> (blockbits - 3)); /* (d)indirect */
166 if (left > GRANT_FOR_LLOG(obd)) {
167 left = (left - GRANT_FOR_LLOG(obd)) << blockbits;
169 left = 0 /* << blockbits */;
172 if (!statfs_done && left < 32 * FILTER_GRANT_CHUNK + tot_granted) {
173 CDEBUG(D_CACHE, "fs has no space left and statfs too old\n");
177 if (left >= tot_granted) {
180 if (left < tot_granted - obd->u.filter.fo_tot_pending) {
181 CERROR("%s: cli %s/%p grant "LPU64" > available "
182 LPU64" and pending "LPU64"\n", obd->obd_name,
183 exp->exp_client_uuid.uuid, exp, tot_granted,
184 left, obd->u.filter.fo_tot_pending);
189 CDEBUG(D_CACHE, "%s: cli %s/%p free: "LPU64" avail: "LPU64" grant "LPU64
190 " left: "LPU64" pending: "LPU64"\n", obd->obd_name,
191 exp->exp_client_uuid.uuid, exp,
192 obd->obd_osfs.os_bfree << blockbits, avail << blockbits,
193 tot_granted, left, obd->u.filter.fo_tot_pending);
198 /* Calculate how much grant space to allocate to this client, based on how
199 * much space is currently free and how much of that is already granted.
201 * Caller must hold obd_osfs_lock. */
202 long filter_grant(struct obd_export *exp, obd_size current_grant,
203 obd_size want, obd_size fs_space_left)
205 struct obd_device *obd = exp->exp_obd;
206 struct filter_export_data *fed = &exp->exp_filter_data;
207 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
210 LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
212 /* Grant some fraction of the client's requested grant space so that
213 * they are not always waiting for write credits (not all of it to
214 * avoid overgranting in face of multiple RPCs in flight). This
215 * essentially will be able to control the OSC_MAX_RIF for a client.
217 * If we do have a large disparity between what the client thinks it
218 * has and what we think it has, don't grant very much and let the
219 * client consume its grant first. Either it just has lots of RPCs
220 * in flight, or it was evicted and its grants will soon be used up. */
221 if (want > 0x7fffffff) {
222 CERROR("%s: client %s/%p requesting > 2GB grant "LPU64"\n",
223 obd->obd_name, exp->exp_client_uuid.uuid, exp, want);
224 } else if (current_grant < want &&
225 current_grant < fed->fed_grant + FILTER_GRANT_CHUNK) {
226 grant = min((want >> blockbits),
227 (fs_space_left >> blockbits) / 8);
231 /* Allow >FILTER_GRANT_CHUNK size when clients
232 * reconnect due to a server reboot.
234 if ((grant > FILTER_GRANT_CHUNK) &&
235 (!obd->obd_recovering))
236 grant = FILTER_GRANT_CHUNK;
238 obd->u.filter.fo_tot_granted += grant;
239 fed->fed_grant += grant;
240 if (fed->fed_grant < 0) {
241 CERROR("%s: cli %s/%p grant %ld want "LPU64
243 obd->obd_name, exp->exp_client_uuid.uuid,
244 exp, fed->fed_grant, want,current_grant);
245 spin_unlock(&obd->obd_osfs_lock);
252 "%s: cli %s/%p wants: "LPU64" current grant "LPU64
253 " granting: "LPU64"\n", obd->obd_name, exp->exp_client_uuid.uuid,
254 exp, want, current_grant, grant);
256 "%s: cli %s/%p tot cached:"LPU64" granted:"LPU64
257 " num_exports: %d\n", obd->obd_name, exp->exp_client_uuid.uuid,
258 exp, obd->u.filter.fo_tot_dirty,
259 obd->u.filter.fo_tot_granted, obd->obd_num_exports);
265 * the routine is used to request pages from pagecache
267 * use GFP_NOFS for requests from a local client not allowing to enter FS
268 * as we might end up waiting on a page he sent in the request we're serving.
269 * use __GFP_HIGHMEM so that the pages can use all of the available memory
271 * use more agressive GFP_HIGHUSER flags from non-local clients to be able to
272 * generate more memory pressure, but at the same time use __GFP_NOMEMALLOC
273 * in order not to exhaust emergency reserves.
275 * See Bug 19529 and Bug 19917 for details.
277 static struct page *filter_get_page(struct obd_device *obd,
284 page = find_or_create_page(inode->i_mapping, offset >> CFS_PAGE_SHIFT,
285 (localreq ? (GFP_NOFS | __GFP_HIGHMEM)
286 : (GFP_HIGHUSER | __GFP_NOMEMALLOC)));
287 if (unlikely(page == NULL))
288 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_NO_PAGE, 1);
294 * the routine initializes array of local_niobuf from remote_niobuf
296 static int filter_map_remote_to_local(int objcount, struct obd_ioobj *obj,
297 struct niobuf_remote *nb,
298 int *nrpages, struct niobuf_local *res)
300 struct niobuf_remote *rnb;
301 struct niobuf_local *lnb;
305 /* we don't support multiobject RPC yet
306 * ost_brw_read() and ost_brw_write() check this */
307 LASSERT(objcount == 1);
311 for (i = 0, rnb = nb, lnb = res; i < obj->ioo_bufcnt; i++, rnb++) {
312 obd_off offset = rnb->offset;
313 unsigned int len = rnb->len;
316 int poff = offset & (CFS_PAGE_SIZE - 1);
317 int plen = CFS_PAGE_SIZE - poff;
319 if (*nrpages >= max) {
320 CERROR("small array of local bufs: %d\n", max);
326 lnb->offset = offset;
328 lnb->flags = rnb->flags;
331 lnb->lnb_grant_used = 0;
333 LASSERTF(plen <= len, "plen %u, len %u\n", plen, len);
344 * the function is used to free all pages used for request
345 * just to mimic cacheless OSS which don't occupy much memory
347 void filter_invalidate_cache(struct obd_device *obd, struct obd_ioobj *obj,
348 struct niobuf_remote *nb, struct inode *inode)
350 struct niobuf_remote *rnb;
353 LASSERT(inode != NULL);
355 for (i = 0, rnb = nb; i < obj->ioo_bufcnt; i++, rnb++) {
359 start = rnb->offset >> CFS_PAGE_SHIFT;
360 end = (rnb->offset + rnb->len) >> CFS_PAGE_SHIFT;
361 invalidate_mapping_pages(inode->i_mapping, start, end);
362 /* just to avoid warnings */
369 * the invalidate above doesn't work during read because lnet pins pages.
370 * The truncate is used here instead to drop pages from cache
372 void filter_truncate_cache(struct obd_device *obd, struct obd_ioobj *obj,
373 struct niobuf_remote *nb, int pages,
374 struct niobuf_local *res, struct inode *inode)
376 struct niobuf_remote *rnb;
379 LASSERT(inode != NULL);
380 #ifdef HAVE_TRUNCATE_RANGE
381 for (i = 0, rnb = nb; i < obj->ioo_bufcnt; i++, rnb++) {
382 /* remove pages in which range is fit */
383 truncate_inode_pages_range(inode->i_mapping,
384 rnb->offset & CFS_PAGE_MASK,
385 (rnb->offset + rnb->len - 1) |
388 #elif (defined HAVE_TRUNCATE_COMPLETE)
389 for (i = 0, lnb = res; i < pages; i++, lnb++)
390 truncate_complete_page(inode->i_mapping, lnb->page);
392 #error "Nor truncate_inode_pages_range or truncate_complete_page are supported"
396 static int filter_preprw_read(int cmd, struct obd_export *exp, struct obdo *oa,
397 int objcount, struct obd_ioobj *obj,
398 struct niobuf_remote *nb,
399 int *npages, struct niobuf_local *res,
400 struct obd_trans_info *oti,
401 struct lustre_capa *capa)
403 struct obd_device *obd = exp->exp_obd;
404 struct timeval start, end;
405 struct lvfs_run_ctxt saved;
406 struct niobuf_local *lnb;
407 struct dentry *dentry = NULL;
408 struct inode *inode = NULL;
410 int rc = 0, i, tot_bytes = 0;
411 unsigned long now = jiffies;
416 /* We are currently not supporting multi-obj BRW_READ RPCS at all.
417 * When we do this function's dentry cleanup will need to be fixed.
418 * These values are verified in ost_brw_write() from the wire. */
419 LASSERTF(objcount == 1, "%d\n", objcount);
420 LASSERTF(obj->ioo_bufcnt > 0, "%d\n", obj->ioo_bufcnt);
422 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
423 rc = filter_auth_capa(exp, NULL, oa->o_gr, capa,
428 if (oa && oa->o_valid & OBD_MD_FLGRANT) {
429 spin_lock(&obd->obd_osfs_lock);
430 filter_grant_incoming(exp, oa);
432 if (!(oa->o_flags & OBD_FL_SHRINK_GRANT))
434 spin_unlock(&obd->obd_osfs_lock);
437 iobuf = filter_iobuf_get(&obd->u.filter, oti);
439 RETURN(PTR_ERR(iobuf));
441 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
442 dentry = filter_oa2dentry(obd, oa);
443 if (IS_ERR(dentry)) {
444 rc = PTR_ERR(dentry);
449 inode = dentry->d_inode;
450 isize = i_size_read(inode);
452 obdo_to_inode(inode, oa, OBD_MD_FLATIME);
454 rc = filter_map_remote_to_local(objcount, obj, nb, npages, res);
458 fsfilt_check_slow(obd, now, "preprw_read setup");
460 /* find pages for all segments, fill array with them */
461 do_gettimeofday(&start);
462 for (i = 0, lnb = res; i < *npages; i++, lnb++) {
464 lnb->dentry = dentry;
466 if (isize <= lnb->offset)
467 /* If there's no more data, abort early. lnb->rc == 0,
468 * so it's easy to detect later. */
471 lnb->page = filter_get_page(obd, inode, lnb->offset, 0);
472 if (lnb->page == NULL)
473 GOTO(cleanup, rc = -ENOMEM);
475 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS, 1);
477 if (isize < lnb->offset + lnb->len - 1)
478 lnb->rc = isize - lnb->offset;
482 tot_bytes += lnb->rc;
484 if (PageUptodate(lnb->page)) {
485 lprocfs_counter_add(obd->obd_stats,
486 LPROC_FILTER_CACHE_HIT, 1);
490 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_MISS, 1);
491 filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
493 do_gettimeofday(&end);
494 timediff = cfs_timeval_sub(&end, &start, NULL);
495 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
497 if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
498 GOTO(cleanup, rc = -ENOMEM);
500 fsfilt_check_slow(obd, now, "start_page_read");
502 rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf,
503 exp, NULL, NULL, NULL);
507 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_READ_BYTES, tot_bytes);
509 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
510 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
511 LPROC_FILTER_READ_BYTES, tot_bytes);
516 /* unlock pages to allow access from concurrent OST_READ */
517 for (i = 0, lnb = res; i < *npages; i++, lnb++) {
519 LASSERT(PageLocked(lnb->page));
520 unlock_page(lnb->page);
523 page_cache_release(lnb->page);
534 filter_iobuf_put(&obd->u.filter, iobuf, oti);
536 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
538 CERROR("io error %d\n", rc);
543 /* When clients have dirtied as much space as they've been granted they
544 * fall through to sync writes. These sync writes haven't been expressed
545 * in grants and need to error with ENOSPC when there isn't room in the
546 * filesystem for them after grants are taken into account. However,
547 * writeback of the dirty data that was already granted space can write
550 * Caller must hold obd_osfs_lock. */
551 static int filter_grant_check(struct obd_export *exp, struct obdo *oa,
552 int objcount, struct fsfilt_objinfo *fso,
553 int niocount, struct niobuf_local *lnb,
554 obd_size *left, struct inode *inode)
556 struct filter_export_data *fed = &exp->exp_filter_data;
557 int blocksize = exp->exp_obd->u.obt.obt_sb->s_blocksize;
558 unsigned long used = 0, ungranted = 0, using;
559 int i, rc = -ENOSPC, obj, n = 0;
561 LASSERT_SPIN_LOCKED(&exp->exp_obd->obd_osfs_lock);
563 for (obj = 0; obj < objcount; obj++) {
564 for (i = 0; i < fso[obj].fso_bufcnt; i++, n++) {
567 /* should match the code in osc_exit_cache */
569 bytes += lnb[n].offset & (blocksize - 1);
570 tmp = (lnb[n].offset + lnb[n].len) & (blocksize - 1);
572 bytes += blocksize - tmp;
574 if ((lnb[n].flags & OBD_BRW_FROM_GRANT) &&
575 (oa->o_valid & OBD_MD_FLGRANT)) {
576 if (fed->fed_grant < used + bytes) {
578 "%s: cli %s/%p claims %ld+%d "
579 "GRANT, real grant %lu idx %d\n",
580 exp->exp_obd->obd_name,
581 exp->exp_client_uuid.uuid, exp,
582 used, bytes, fed->fed_grant, n);
585 lnb[n].flags |= OBD_BRW_GRANTED;
586 lnb[n].lnb_grant_used = bytes;
587 CDEBUG(0, "idx %d used=%lu\n", n, used);
592 if (*left > ungranted + bytes) {
593 /* if enough space, pretend it was granted */
595 lnb[n].flags |= OBD_BRW_GRANTED;
596 lnb[n].lnb_grant_used = bytes;
597 CDEBUG(0, "idx %d ungranted=%lu\n",n,ungranted);
602 /* We can't check for already-mapped blocks here, as
603 * it requires dropping the osfs lock to do the bmap.
604 * Instead, we return ENOSPC and in that case we need
605 * to go through and verify if all of the blocks not
606 * marked BRW_GRANTED are already mapped and we can
607 * ignore this error. */
609 lnb[n].flags &= ~OBD_BRW_GRANTED;
610 CDEBUG(D_CACHE,"%s: cli %s/%p idx %d no space for %d\n",
611 exp->exp_obd->obd_name,
612 exp->exp_client_uuid.uuid, exp, n, bytes);
616 /* Now substract what client have used already. We don't subtract
617 * this from the tot_granted yet, so that other client's can't grab
618 * that space before we have actually allocated our blocks. That
619 * happens in filter_grant_commit() after the writes are done. */
621 fed->fed_grant -= used;
622 fed->fed_pending += used + ungranted;
623 exp->exp_obd->u.filter.fo_tot_granted += ungranted;
624 exp->exp_obd->u.filter.fo_tot_pending += used + ungranted;
627 "%s: cli %s/%p used: %lu ungranted: %lu grant: %lu dirty: %lu\n",
628 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp, used,
629 ungranted, fed->fed_grant, fed->fed_dirty);
631 /* Rough calc in case we don't refresh cached statfs data */
632 using = (used + ungranted + 1 ) >>
633 exp->exp_obd->u.obt.obt_sb->s_blocksize_bits;
634 if (exp->exp_obd->obd_osfs.os_bavail > using)
635 exp->exp_obd->obd_osfs.os_bavail -= using;
637 exp->exp_obd->obd_osfs.os_bavail = 0;
639 if (fed->fed_dirty < used) {
640 CERROR("%s: cli %s/%p claims used %lu > fed_dirty %lu\n",
641 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
642 used, fed->fed_dirty);
643 used = fed->fed_dirty;
645 exp->exp_obd->u.filter.fo_tot_dirty -= used;
646 fed->fed_dirty -= used;
648 if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
649 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
650 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
651 fed->fed_dirty, fed->fed_pending, fed->fed_grant);
652 spin_unlock(&exp->exp_obd->obd_osfs_lock);
658 /* If we ever start to support multi-object BRW RPCs, we will need to get locks
659 * on mulitple inodes. That isn't all, because there still exists the
660 * possibility of a truncate starting a new transaction while holding the ext3
661 * rwsem = write while some writes (which have started their transactions here)
662 * blocking on the ext3 rwsem = read => lock inversion.
664 * The handling gets very ugly when dealing with locked pages. It may be easier
665 * to just get rid of the locked page code (which has problems of its own) and
666 * either discover we do not need it anymore (i.e. it was a symptom of another
667 * bug) or ensure we get the page locks in an appropriate order. */
668 static int filter_preprw_write(int cmd, struct obd_export *exp, struct obdo *oa,
669 int objcount, struct obd_ioobj *obj,
670 struct niobuf_remote *nb, int *npages,
671 struct niobuf_local *res,
672 struct obd_trans_info *oti,
673 struct lustre_capa *capa)
675 struct obd_device *obd = exp->exp_obd;
676 struct timeval start, end;
677 struct lvfs_run_ctxt saved;
678 struct niobuf_local *lnb = res;
679 struct fsfilt_objinfo fso;
680 struct filter_mod_data *fmd;
681 struct dentry *dentry = NULL;
684 unsigned long now = jiffies, timediff;
685 int rc = 0, i, tot_bytes = 0, cleanup_phase = 0, localreq = 0;
687 LASSERT(objcount == 1);
688 LASSERT(obj->ioo_bufcnt > 0);
690 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
691 rc = filter_auth_capa(exp, NULL, oa->o_gr, capa,
696 if (exp->exp_connection &&
697 exp->exp_connection->c_peer.nid == exp->exp_connection->c_self)
700 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
701 iobuf = filter_iobuf_get(&obd->u.filter, oti);
703 GOTO(cleanup, rc = PTR_ERR(iobuf));
706 dentry = filter_fid2dentry(obd, NULL, obj->ioo_gr,
709 GOTO(cleanup, rc = PTR_ERR(dentry));
712 if (dentry->d_inode == NULL) {
713 CERROR("%s: trying to BRW to non-existent file "LPU64"\n",
714 obd->obd_name, obj->ioo_id);
715 GOTO(cleanup, rc = -ENOENT);
718 if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID) &&
719 dentry->d_inode->i_mode & (S_ISUID | S_ISGID)) {
720 rc = filter_capa_fixoa(exp, oa, oa->o_gr, capa);
725 rc = filter_map_remote_to_local(objcount, obj, nb, npages, res);
729 fsfilt_check_slow(obd, now, "preprw_write setup");
731 /* Don't update inode timestamps if this write is older than a
732 * setattr which modifies the timestamps. b=10150 */
733 /* XXX when we start having persistent reservations this needs to
734 * be changed to filter_fmd_get() to create the fmd if it doesn't
735 * already exist so we can store the reservation handle there. */
736 fmd = filter_fmd_find(exp, obj->ioo_id, obj->ioo_gr);
739 spin_lock(&obd->obd_osfs_lock);
740 filter_grant_incoming(exp, oa);
741 if (fmd && fmd->fmd_mactime_xid > oti->oti_xid)
742 oa->o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLCTIME |
745 obdo_to_inode(dentry->d_inode, oa, OBD_MD_FLATIME |
746 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
749 left = filter_grant_space_left(exp);
751 fso.fso_dentry = dentry;
752 fso.fso_bufcnt = *npages;
754 rc = filter_grant_check(exp, oa, objcount, &fso, *npages, res,
755 &left, dentry->d_inode);
757 /* do not zero out oa->o_valid as it is used in filter_commitrw_write()
758 * for setting UID/GID and fid EA in first write time. */
759 if (oa->o_valid & OBD_MD_FLGRANT)
760 oa->o_grant = filter_grant(exp,oa->o_grant,oa->o_undirty,left);
762 spin_unlock(&obd->obd_osfs_lock);
763 filter_fmd_put(exp, fmd);
769 /* Filter truncate first locks i_mutex then partally truncated
770 * page, filter write code first locks pages then take
771 * i_mutex. To avoid a deadlock in case of concurrent
772 * punch/write requests from one client, filter writes and
773 * filter truncates are serialized by i_alloc_sem, allowing
774 * multiple writes or single truncate. */
775 down_read(&dentry->d_inode->i_alloc_sem);
777 do_gettimeofday(&start);
778 for (i = 0, lnb = res; i < *npages; i++, lnb++) {
780 /* We still set up for ungranted pages so that granted pages
781 * can be written to disk as they were promised, and portals
782 * needs to keep the pages all aligned properly. */
783 lnb->dentry = dentry;
785 lnb->page = filter_get_page(obd, dentry->d_inode, lnb->offset,
787 if (lnb->page == NULL)
788 GOTO(cleanup, rc = -ENOMEM);
790 /* DLM locking protects us from write and truncate competing
791 * for same region, but truncate can leave dirty page in the
792 * cache. it's possible the writeout on a such a page is in
793 * progress when we access it. it's also possible that during
794 * this writeout we put new (partial) data, but then won't
795 * be able to proceed in filter_commitrw_write(). thus let's
796 * just wait for writeout completion, should be rare enough.
798 if (obd->u.filter.fo_writethrough_cache)
799 wait_on_page_writeback(lnb->page);
800 BUG_ON(PageWriteback(lnb->page));
802 /* If the filter writes a partial page, then has the file
803 * extended, the client will read in the whole page. the
804 * filter has to be careful to zero the rest of the partial
805 * page on disk. we do it by hand for partial extending
806 * writes, send_bio() is responsible for zeroing pages when
807 * asked to read unmapped blocks -- brw_kiovec() does this. */
808 if (lnb->len != CFS_PAGE_SIZE) {
811 maxidx = ((i_size_read(dentry->d_inode) +
812 CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT) - 1;
813 if (maxidx >= lnb->page->index) {
814 LL_CDEBUG_PAGE(D_PAGE, lnb->page, "write %u @ "
815 LPU64" flg %x before EOF %llu\n",
816 lnb->len, lnb->offset,lnb->flags,
817 i_size_read(dentry->d_inode));
818 filter_iobuf_add_page(obd, iobuf,
823 char *p = kmap(lnb->page);
825 off = lnb->offset & ~CFS_PAGE_MASK;
828 off = (lnb->offset + lnb->len) & ~CFS_PAGE_MASK;
830 memset(p + off, 0, CFS_PAGE_SIZE - off);
835 tot_bytes += lnb->len;
837 do_gettimeofday(&end);
838 timediff = cfs_timeval_sub(&end, &start, NULL);
839 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
841 if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
842 GOTO(cleanup, rc = -ENOMEM);
844 /* don't unlock pages to prevent any access */
845 rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf, exp,
848 fsfilt_check_slow(obd, now, "start_page_write");
850 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
851 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
852 LPROC_FILTER_WRITE_BYTES, tot_bytes);
855 switch(cleanup_phase) {
858 for (i = 0, lnb = res; i < *npages; i++, lnb++) {
859 if (lnb->page != NULL) {
860 unlock_page(lnb->page);
861 page_cache_release(lnb->page);
865 up_read(&dentry->d_inode->i_alloc_sem);
868 filter_iobuf_put(&obd->u.filter, iobuf, oti);
870 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
875 filter_iobuf_put(&obd->u.filter, iobuf, oti);
877 spin_lock(&obd->obd_osfs_lock);
879 filter_grant_incoming(exp, oa);
880 spin_unlock(&obd->obd_osfs_lock);
881 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
888 int filter_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
889 int objcount, struct obd_ioobj *obj,
890 struct niobuf_remote *nb, int *npages,
891 struct niobuf_local *res, struct obd_trans_info *oti,
892 struct lustre_capa *capa)
894 if (cmd == OBD_BRW_WRITE)
895 return filter_preprw_write(cmd, exp, oa, objcount, obj,
896 nb, npages, res, oti, capa);
897 if (cmd == OBD_BRW_READ)
898 return filter_preprw_read(cmd, exp, oa, objcount, obj,
899 nb, npages, res, oti, capa);
904 static int filter_commitrw_read(struct obd_export *exp, struct obdo *oa,
905 int objcount, struct obd_ioobj *obj,
906 struct niobuf_remote *rnb,
907 int npages, struct niobuf_local *res,
908 struct obd_trans_info *oti, int rc)
910 struct filter_obd *fo = &exp->exp_obd->u.filter;
911 struct inode *inode = NULL;
912 struct ldlm_res_id res_id;
913 struct ldlm_resource *resource = NULL;
914 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
915 struct niobuf_local *lnb;
919 osc_build_res_name(obj->ioo_id, obj->ioo_gr, &res_id);
920 /* If oa != NULL then filter_preprw_read updated the inode atime
921 * and we should update the lvb so that other glimpses will also
922 * get the updated value. bug 5972 */
923 if (oa && ns && ns->ns_lvbo && ns->ns_lvbo->lvbo_update) {
924 resource = ldlm_resource_get(ns, NULL, &res_id, LDLM_EXTENT, 0);
926 if (resource != NULL) {
927 LDLM_RESOURCE_ADDREF(resource);
928 ns->ns_lvbo->lvbo_update(resource, NULL, 0, 1);
929 LDLM_RESOURCE_DELREF(resource);
930 ldlm_resource_putref(resource);
934 if (res->dentry != NULL)
935 inode = res->dentry->d_inode;
937 for (i = 0, lnb = res; i < npages; i++, lnb++)
938 if (lnb->page != NULL)
939 page_cache_release(lnb->page);
941 if (inode && (fo->fo_read_cache == 0 ||
942 i_size_read(inode) > fo->fo_readcache_max_filesize))
943 filter_truncate_cache(exp->exp_obd, obj, rnb, npages, res,
946 for (i = 0, lnb = res; i < npages; i++, lnb++)
949 if (res->dentry != NULL)
954 void filter_grant_commit(struct obd_export *exp, int niocount,
955 struct niobuf_local *res)
957 struct filter_obd *filter = &exp->exp_obd->u.filter;
958 struct niobuf_local *lnb = res;
959 unsigned long pending = 0;
962 spin_lock(&exp->exp_obd->obd_osfs_lock);
963 for (i = 0, lnb = res; i < niocount; i++, lnb++)
964 pending += lnb->lnb_grant_used;
966 LASSERTF(exp->exp_filter_data.fed_pending >= pending,
967 "%s: cli %s/%p fed_pending: %lu grant_used: %lu\n",
968 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
969 exp->exp_filter_data.fed_pending, pending);
970 exp->exp_filter_data.fed_pending -= pending;
971 LASSERTF(filter->fo_tot_granted >= pending,
972 "%s: cli %s/%p tot_granted: "LPU64" grant_used: %lu\n",
973 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
974 exp->exp_obd->u.filter.fo_tot_granted, pending);
975 filter->fo_tot_granted -= pending;
976 LASSERTF(filter->fo_tot_pending >= pending,
977 "%s: cli %s/%p tot_pending: "LPU64" grant_used: %lu\n",
978 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
979 filter->fo_tot_pending, pending);
980 filter->fo_tot_pending -= pending;
982 spin_unlock(&exp->exp_obd->obd_osfs_lock);
985 int filter_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
986 int objcount, struct obd_ioobj *obj,
987 struct niobuf_remote *nb, int npages,
988 struct niobuf_local *res, struct obd_trans_info *oti,
991 if (cmd == OBD_BRW_WRITE)
992 return filter_commitrw_write(exp, oa, objcount, obj,
993 nb, npages, res, oti, rc);
994 if (cmd == OBD_BRW_READ)
995 return filter_commitrw_read(exp, oa, objcount, obj,
996 nb, npages, res, oti, rc);
1001 int filter_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1002 obd_count oa_bufs, struct brw_page *pga,
1003 struct obd_trans_info *oti)
1005 struct obd_ioobj ioo;
1006 struct niobuf_local *lnb;
1007 struct niobuf_remote *rnb;
1009 int ret = 0, npages;
1012 OBD_ALLOC(lnb, oa_bufs * sizeof(struct niobuf_local));
1013 OBD_ALLOC(rnb, oa_bufs * sizeof(struct niobuf_remote));
1015 if (lnb == NULL || rnb == NULL)
1016 GOTO(out, ret = -ENOMEM);
1018 for (i = 0; i < oa_bufs; i++) {
1019 lnb[i].page = pga[i].pg;
1020 rnb[i].offset = pga[i].off;
1021 rnb[i].len = pga[i].count;
1024 obdo_to_ioobj(oinfo->oi_oa, &ioo);
1025 ioo.ioo_bufcnt = oa_bufs;
1028 ret = filter_preprw(cmd, exp, oinfo->oi_oa, 1, &ioo,
1029 rnb, &npages, lnb, oti, oinfo_capa(oinfo));
1032 LASSERTF(oa_bufs == npages, "%u != %u\n", oa_bufs, npages);
1034 ret = filter_commitrw(cmd, exp, oinfo->oi_oa, 1, &ioo, rnb,
1035 npages, lnb, oti, ret);
1039 OBD_FREE(lnb, oa_bufs * sizeof(struct niobuf_local));
1041 OBD_FREE(rnb, oa_bufs * sizeof(struct niobuf_remote));