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 (c) 2003, 2010, Oracle and/or its affiliates. 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_valid & OBD_MD_FLFLAGS && 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 cfs_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,
156 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS))) {
158 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb,
159 cfs_time_shift_64(OBD_STATFS_CACHE_SECONDS));
160 if (rc) /* N.B. statfs can't really fail */
165 avail = obd->obd_osfs.os_bavail;
166 left = avail - (avail >> (blockbits - 3)); /* (d)indirect */
167 if (left > GRANT_FOR_LLOG(obd)) {
168 left = (left - GRANT_FOR_LLOG(obd)) << blockbits;
170 left = 0 /* << blockbits */;
173 if (!statfs_done && left < 32 * FILTER_GRANT_CHUNK + tot_granted) {
174 CDEBUG(D_CACHE, "fs has no space left and statfs too old\n");
178 if (left >= tot_granted) {
181 if (left < tot_granted - obd->u.filter.fo_tot_pending) {
182 CERROR("%s: cli %s/%p grant "LPU64" > available "
183 LPU64" and pending "LPU64"\n", obd->obd_name,
184 exp->exp_client_uuid.uuid, exp, tot_granted,
185 left, obd->u.filter.fo_tot_pending);
190 CDEBUG(D_CACHE, "%s: cli %s/%p free: "LPU64" avail: "LPU64" grant "LPU64
191 " left: "LPU64" pending: "LPU64"\n", obd->obd_name,
192 exp->exp_client_uuid.uuid, exp,
193 obd->obd_osfs.os_bfree << blockbits, avail << blockbits,
194 tot_granted, left, obd->u.filter.fo_tot_pending);
199 /* Calculate how much grant space to allocate to this client, based on how
200 * much space is currently free and how much of that is already granted.
202 * if @conservative != 0, we limit the maximum grant to FILTER_GRANT_CHUNK;
203 * otherwise we'll satisfy the requested amount as possible as we can, this
204 * is usually due to client reconnect.
206 * Caller must hold obd_osfs_lock. */
207 long filter_grant(struct obd_export *exp, obd_size current_grant,
208 obd_size want, obd_size fs_space_left, int conservative)
210 struct obd_device *obd = exp->exp_obd;
211 struct filter_export_data *fed = &exp->exp_filter_data;
212 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
215 LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
217 /* Grant some fraction of the client's requested grant space so that
218 * they are not always waiting for write credits (not all of it to
219 * avoid overgranting in face of multiple RPCs in flight). This
220 * essentially will be able to control the OSC_MAX_RIF for a client.
222 * If we do have a large disparity between what the client thinks it
223 * has and what we think it has, don't grant very much and let the
224 * client consume its grant first. Either it just has lots of RPCs
225 * in flight, or it was evicted and its grants will soon be used up. */
226 if (want > 0x7fffffff) {
227 CERROR("%s: client %s/%p requesting > 2GB grant "LPU64"\n",
228 obd->obd_name, exp->exp_client_uuid.uuid, exp, want);
229 } else if (current_grant < want &&
230 current_grant < fed->fed_grant + FILTER_GRANT_CHUNK) {
231 grant = min(want + (1 << blockbits) - 1, fs_space_left / 8);
232 grant &= ~((1ULL << blockbits) - 1);
235 if (grant > FILTER_GRANT_CHUNK && conservative)
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 cfs_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 aggressive GFP_HIGHUSER flags from non-local clients to be able to
272 * generate more memory pressure.
274 * See Bug 19529 and Bug 19917 for details.
276 static struct page *filter_get_page(struct obd_device *obd,
283 page = find_or_create_page(inode->i_mapping, offset >> CFS_PAGE_SHIFT,
284 (localreq ? (GFP_NOFS | __GFP_HIGHMEM)
286 if (unlikely(page == NULL))
287 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_NO_PAGE, 1);
293 * the routine initializes array of local_niobuf from remote_niobuf
295 static int filter_map_remote_to_local(int objcount, struct obd_ioobj *obj,
296 struct niobuf_remote *nb,
297 int *nrpages, struct niobuf_local *res)
299 struct niobuf_remote *rnb;
300 struct niobuf_local *lnb;
304 /* we don't support multiobject RPC yet
305 * ost_brw_read() and ost_brw_write() check this */
306 LASSERT(objcount == 1);
310 for (i = 0, rnb = nb, lnb = res; i < obj->ioo_bufcnt; i++, rnb++) {
311 obd_off offset = rnb->offset;
312 unsigned int len = rnb->len;
315 int poff = offset & (CFS_PAGE_SIZE - 1);
316 int plen = CFS_PAGE_SIZE - poff;
318 if (*nrpages >= max) {
319 CERROR("small array of local bufs: %d\n", max);
325 lnb->offset = offset;
327 lnb->flags = rnb->flags;
330 lnb->lnb_grant_used = 0;
332 LASSERTF(plen <= len, "plen %u, len %u\n", plen, len);
343 * the invalidate above doesn't work during read because lnet pins pages.
344 * The truncate is used here instead to drop pages from cache
346 void filter_release_cache(struct obd_device *obd, struct obd_ioobj *obj,
347 struct niobuf_remote *rnb, struct inode *inode)
351 LASSERT(inode != NULL);
352 for (i = 0; i < obj->ioo_bufcnt; i++, rnb++) {
353 #ifdef HAVE_TRUNCATE_RANGE
354 /* remove pages in which range is fit */
355 truncate_inode_pages_range(inode->i_mapping,
356 rnb->offset & CFS_PAGE_MASK,
357 (rnb->offset + rnb->len - 1) |
360 /* use invalidate for old kernels */
361 invalidate_mapping_pages(inode->i_mapping,
362 rnb->offset >> CFS_PAGE_SHIFT,
363 (rnb->offset + rnb->len) >>
369 static int filter_preprw_read(int cmd, struct obd_export *exp, struct obdo *oa,
370 int objcount, struct obd_ioobj *obj,
371 struct niobuf_remote *nb,
372 int *npages, struct niobuf_local *res,
373 struct obd_trans_info *oti,
374 struct lustre_capa *capa)
376 struct obd_device *obd = exp->exp_obd;
377 struct timeval start, end;
378 struct lvfs_run_ctxt saved;
379 struct niobuf_local *lnb;
380 struct dentry *dentry = NULL;
381 struct inode *inode = NULL;
383 int rc = 0, i, tot_bytes = 0;
384 unsigned long now = jiffies;
389 /* We are currently not supporting multi-obj BRW_READ RPCS at all.
390 * When we do this function's dentry cleanup will need to be fixed.
391 * These values are verified in ost_brw_write() from the wire. */
392 LASSERTF(objcount == 1, "%d\n", objcount);
393 LASSERTF(obj->ioo_bufcnt > 0, "%d\n", obj->ioo_bufcnt);
395 rc = filter_auth_capa(exp, NULL, oa->o_seq, capa,
400 if (oa && oa->o_valid & OBD_MD_FLGRANT) {
401 cfs_spin_lock(&obd->obd_osfs_lock);
402 filter_grant_incoming(exp, oa);
404 if (!(oa->o_valid & OBD_MD_FLFLAGS) ||
405 !(oa->o_flags & OBD_FL_SHRINK_GRANT))
407 cfs_spin_unlock(&obd->obd_osfs_lock);
410 iobuf = filter_iobuf_get(&obd->u.filter, oti);
412 RETURN(PTR_ERR(iobuf));
414 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
415 dentry = filter_oa2dentry(obd, &oa->o_oi);
416 if (IS_ERR(dentry)) {
417 rc = PTR_ERR(dentry);
422 inode = dentry->d_inode;
423 isize = i_size_read(inode);
425 obdo_to_inode(inode, oa, OBD_MD_FLATIME);
427 rc = filter_map_remote_to_local(objcount, obj, nb, npages, res);
431 fsfilt_check_slow(obd, now, "preprw_read setup");
433 /* find pages for all segments, fill array with them */
434 cfs_gettimeofday(&start);
435 for (i = 0, lnb = res; i < *npages; i++, lnb++) {
437 lnb->dentry = dentry;
439 if (isize <= lnb->offset)
440 /* If there's no more data, abort early. lnb->rc == 0,
441 * so it's easy to detect later. */
444 lnb->page = filter_get_page(obd, inode, lnb->offset, 0);
445 if (lnb->page == NULL)
446 GOTO(cleanup, rc = -ENOMEM);
448 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS, 1);
450 if (isize < lnb->offset + lnb->len - 1)
451 lnb->rc = isize - lnb->offset;
455 tot_bytes += lnb->rc;
457 if (PageUptodate(lnb->page)) {
458 lprocfs_counter_add(obd->obd_stats,
459 LPROC_FILTER_CACHE_HIT, 1);
463 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_MISS, 1);
464 filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
466 cfs_gettimeofday(&end);
467 timediff = cfs_timeval_sub(&end, &start, NULL);
468 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
470 if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
471 GOTO(cleanup, rc = -ENOMEM);
473 fsfilt_check_slow(obd, now, "start_page_read");
475 rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf,
476 exp, NULL, NULL, NULL);
480 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_READ_BYTES, tot_bytes);
482 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
483 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
484 LPROC_FILTER_READ_BYTES, tot_bytes);
489 /* unlock pages to allow access from concurrent OST_READ */
490 for (i = 0, lnb = res; i < *npages; i++, lnb++) {
492 LASSERT(PageLocked(lnb->page));
493 unlock_page(lnb->page);
496 page_cache_release(lnb->page);
507 filter_iobuf_put(&obd->u.filter, iobuf, oti);
509 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
511 CERROR("io error %d\n", rc);
516 /* When clients have dirtied as much space as they've been granted they
517 * fall through to sync writes. These sync writes haven't been expressed
518 * in grants and need to error with ENOSPC when there isn't room in the
519 * filesystem for them after grants are taken into account. However,
520 * writeback of the dirty data that was already granted space can write
523 * Caller must hold obd_osfs_lock. */
524 static int filter_grant_check(struct obd_export *exp, struct obdo *oa,
525 int objcount, struct fsfilt_objinfo *fso,
526 int niocount, struct niobuf_local *lnb,
527 obd_size *left, struct inode *inode)
529 struct filter_export_data *fed = &exp->exp_filter_data;
530 int blocksize = exp->exp_obd->u.obt.obt_sb->s_blocksize;
531 unsigned long used = 0, ungranted = 0, using;
532 int i, rc = -ENOSPC, obj, n = 0;
535 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
536 (oa->o_flags & OBD_FL_RECOV_RESEND)) {
538 CDEBUG(D_CACHE, "Recoverable resend arrived, skipping "
542 LASSERT_SPIN_LOCKED(&exp->exp_obd->obd_osfs_lock);
544 for (obj = 0; obj < objcount; obj++) {
545 for (i = 0; i < fso[obj].fso_bufcnt; i++, n++) {
548 /* should match the code in osc_exit_cache */
550 bytes += lnb[n].offset & (blocksize - 1);
551 tmp = (lnb[n].offset + lnb[n].len) & (blocksize - 1);
553 bytes += blocksize - tmp;
555 if ((lnb[n].flags & OBD_BRW_FROM_GRANT) &&
556 (oa->o_valid & OBD_MD_FLGRANT)) {
558 /* this is a recoverable resent */
559 lnb[n].flags |= OBD_BRW_GRANTED;
562 } else if (fed->fed_grant < used + bytes) {
564 "%s: cli %s/%p claims %ld+%d "
565 "GRANT, real grant %lu idx %d\n",
566 exp->exp_obd->obd_name,
567 exp->exp_client_uuid.uuid, exp,
568 used, bytes, fed->fed_grant, n);
571 lnb[n].flags |= OBD_BRW_GRANTED;
572 lnb[n].lnb_grant_used = bytes;
573 CDEBUG(0, "idx %d used=%lu\n", n, used);
578 if (*left > ungranted + bytes) {
579 /* if enough space, pretend it was granted */
581 lnb[n].flags |= OBD_BRW_GRANTED;
582 lnb[n].lnb_grant_used = bytes;
583 CDEBUG(0, "idx %d ungranted=%lu\n",n,ungranted);
588 /* We can't check for already-mapped blocks here, as
589 * it requires dropping the osfs lock to do the bmap.
590 * Instead, we return ENOSPC and in that case we need
591 * to go through and verify if all of the blocks not
592 * marked BRW_GRANTED are already mapped and we can
593 * ignore this error. */
595 lnb[n].flags &= ~OBD_BRW_GRANTED;
596 CDEBUG(D_CACHE,"%s: cli %s/%p idx %d no space for %d\n",
597 exp->exp_obd->obd_name,
598 exp->exp_client_uuid.uuid, exp, n, bytes);
602 /* Now substract what client have used already. We don't subtract
603 * this from the tot_granted yet, so that other client's can't grab
604 * that space before we have actually allocated our blocks. That
605 * happens in filter_grant_commit() after the writes are done. */
607 fed->fed_grant -= used;
608 fed->fed_pending += used + ungranted;
609 exp->exp_obd->u.filter.fo_tot_granted += ungranted;
610 exp->exp_obd->u.filter.fo_tot_pending += used + ungranted;
613 "%s: cli %s/%p used: %lu ungranted: %lu grant: %lu dirty: %lu\n",
614 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp, used,
615 ungranted, fed->fed_grant, fed->fed_dirty);
617 /* Rough calc in case we don't refresh cached statfs data */
618 using = (used + ungranted + 1 ) >>
619 exp->exp_obd->u.obt.obt_sb->s_blocksize_bits;
620 if (exp->exp_obd->obd_osfs.os_bavail > using)
621 exp->exp_obd->obd_osfs.os_bavail -= using;
623 exp->exp_obd->obd_osfs.os_bavail = 0;
625 if (fed->fed_dirty < used) {
626 CERROR("%s: cli %s/%p claims used %lu > fed_dirty %lu\n",
627 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
628 used, fed->fed_dirty);
629 used = fed->fed_dirty;
631 exp->exp_obd->u.filter.fo_tot_dirty -= used;
632 fed->fed_dirty -= used;
634 if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
635 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
636 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
637 fed->fed_dirty, fed->fed_pending, fed->fed_grant);
638 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
644 /* If we ever start to support multi-object BRW RPCs, we will need to get locks
645 * on mulitple inodes. That isn't all, because there still exists the
646 * possibility of a truncate starting a new transaction while holding the ext3
647 * rwsem = write while some writes (which have started their transactions here)
648 * blocking on the ext3 rwsem = read => lock inversion.
650 * The handling gets very ugly when dealing with locked pages. It may be easier
651 * to just get rid of the locked page code (which has problems of its own) and
652 * either discover we do not need it anymore (i.e. it was a symptom of another
653 * bug) or ensure we get the page locks in an appropriate order. */
654 static int filter_preprw_write(int cmd, struct obd_export *exp, struct obdo *oa,
655 int objcount, struct obd_ioobj *obj,
656 struct niobuf_remote *nb, int *npages,
657 struct niobuf_local *res,
658 struct obd_trans_info *oti,
659 struct lustre_capa *capa)
661 struct obd_device *obd = exp->exp_obd;
662 struct timeval start, end;
663 struct lvfs_run_ctxt saved;
664 struct niobuf_local *lnb = res;
665 struct fsfilt_objinfo fso;
666 struct filter_mod_data *fmd;
667 struct dentry *dentry = NULL;
670 unsigned long now = jiffies, timediff;
671 int rc = 0, i, tot_bytes = 0, cleanup_phase = 0, localreq = 0;
673 LASSERT(objcount == 1);
674 LASSERT(obj->ioo_bufcnt > 0);
676 rc = filter_auth_capa(exp, NULL, oa->o_seq, capa,
681 if (exp->exp_connection &&
682 exp->exp_connection->c_peer.nid == exp->exp_connection->c_self)
685 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
686 iobuf = filter_iobuf_get(&obd->u.filter, oti);
688 GOTO(cleanup, rc = PTR_ERR(iobuf));
691 dentry = filter_fid2dentry(obd, NULL, obj->ioo_seq,
694 GOTO(cleanup, rc = PTR_ERR(dentry));
697 if (dentry->d_inode == NULL) {
698 if (exp->exp_obd->obd_recovering) {
699 struct obdo *noa = oa;
704 GOTO(recreate_out, rc = -ENOMEM);
705 noa->o_id = obj->ioo_id;
706 noa->o_valid = OBD_MD_FLID;
709 if (filter_create(exp, noa, NULL, oti) == 0) {
711 dentry = filter_fid2dentry(exp->exp_obd, NULL,
719 if (IS_ERR(dentry) || dentry->d_inode == NULL) {
720 CERROR("%s: BRW to missing obj "LPU64"/"LPU64":rc %d\n",
721 exp->exp_obd->obd_name,
722 obj->ioo_id, obj->ioo_seq,
723 IS_ERR(dentry) ? (int)PTR_ERR(dentry) : -ENOENT);
726 GOTO(cleanup, rc = -ENOENT);
730 if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID) &&
731 dentry->d_inode->i_mode & (S_ISUID | S_ISGID)) {
732 rc = filter_capa_fixoa(exp, oa, oa->o_seq, capa);
737 rc = filter_map_remote_to_local(objcount, obj, nb, npages, res);
741 fsfilt_check_slow(obd, now, "preprw_write setup");
743 /* Filter truncate first locks i_mutex then partially truncated
744 * page, filter write code first locks pages then take
745 * i_mutex. To avoid a deadlock in case of concurrent
746 * punch/write requests from one client, filter writes and
747 * filter truncates are serialized by i_alloc_sem, allowing
748 * multiple writes or single truncate. */
749 down_read(&dentry->d_inode->i_alloc_sem);
750 fsfilt_check_slow(obd, now, "i_alloc_sem");
752 /* Don't update inode timestamps if this write is older than a
753 * setattr which modifies the timestamps. b=10150 */
754 /* XXX when we start having persistent reservations this needs to
755 * be changed to filter_fmd_get() to create the fmd if it doesn't
756 * already exist so we can store the reservation handle there. */
757 fmd = filter_fmd_find(exp, obj->ioo_id, obj->ioo_seq);
760 cfs_spin_lock(&obd->obd_osfs_lock);
761 filter_grant_incoming(exp, oa);
762 if (fmd && fmd->fmd_mactime_xid > oti->oti_xid)
763 oa->o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLCTIME |
766 obdo_to_inode(dentry->d_inode, oa, OBD_MD_FLATIME |
767 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
770 left = filter_grant_space_left(exp);
772 fso.fso_dentry = dentry;
773 fso.fso_bufcnt = *npages;
775 rc = filter_grant_check(exp, oa, objcount, &fso, *npages, res,
776 &left, dentry->d_inode);
778 /* do not zero out oa->o_valid as it is used in filter_commitrw_write()
779 * for setting UID/GID and fid EA in first write time. */
780 /* If OBD_FL_SHRINK_GRANT is set, the client just returned us some grant
781 * so no sense in allocating it some more. We either return the grant
782 * back to the client if we have plenty of space or we don't return
783 * anything if we are short. This was decided in filter_grant_incoming*/
784 if ((oa->o_valid & OBD_MD_FLGRANT) &&
785 (!(oa->o_valid & OBD_MD_FLFLAGS) ||
786 !(oa->o_flags & OBD_FL_SHRINK_GRANT)))
787 oa->o_grant = filter_grant(exp, oa->o_grant, oa->o_undirty,
790 cfs_spin_unlock(&obd->obd_osfs_lock);
791 filter_fmd_put(exp, fmd);
793 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK2, (obd_timeout + 1) / 4);
799 cfs_gettimeofday(&start);
800 for (i = 0, lnb = res; i < *npages; i++, lnb++) {
802 /* We still set up for ungranted pages so that granted pages
803 * can be written to disk as they were promised, and portals
804 * needs to keep the pages all aligned properly. */
805 lnb->dentry = dentry;
807 lnb->page = filter_get_page(obd, dentry->d_inode, lnb->offset,
809 if (lnb->page == NULL)
810 GOTO(cleanup, rc = -ENOMEM);
812 /* DLM locking protects us from write and truncate competing
813 * for same region, but truncate can leave dirty page in the
814 * cache. it's possible the writeout on a such a page is in
815 * progress when we access it. it's also possible that during
816 * this writeout we put new (partial) data, but then won't
817 * be able to proceed in filter_commitrw_write(). thus let's
818 * just wait for writeout completion, should be rare enough.
820 wait_on_page_writeback(lnb->page);
821 BUG_ON(PageWriteback(lnb->page));
823 /* If the filter writes a partial page, then has the file
824 * extended, the client will read in the whole page. the
825 * filter has to be careful to zero the rest of the partial
826 * page on disk. we do it by hand for partial extending
827 * writes, send_bio() is responsible for zeroing pages when
828 * asked to read unmapped blocks -- brw_kiovec() does this. */
829 if (lnb->len != CFS_PAGE_SIZE) {
832 maxidx = ((i_size_read(dentry->d_inode) +
833 CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT) - 1;
834 if (maxidx >= lnb->page->index) {
835 LL_CDEBUG_PAGE(D_PAGE, lnb->page, "write %u @ "
836 LPU64" flg %x before EOF %llu\n",
837 lnb->len, lnb->offset,lnb->flags,
838 i_size_read(dentry->d_inode));
839 filter_iobuf_add_page(obd, iobuf,
844 char *p = kmap(lnb->page);
846 off = lnb->offset & ~CFS_PAGE_MASK;
849 off = (lnb->offset + lnb->len) & ~CFS_PAGE_MASK;
851 memset(p + off, 0, CFS_PAGE_SIZE - off);
856 tot_bytes += lnb->len;
858 cfs_gettimeofday(&end);
859 timediff = cfs_timeval_sub(&end, &start, NULL);
860 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
862 if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
863 GOTO(cleanup, rc = -ENOMEM);
865 /* don't unlock pages to prevent any access */
866 rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf, exp,
869 fsfilt_check_slow(obd, now, "start_page_write");
871 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
874 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
875 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
876 LPROC_FILTER_WRITE_BYTES, tot_bytes);
879 switch(cleanup_phase) {
882 for (i = 0, lnb = res; i < *npages; i++, lnb++) {
883 if (lnb->page != NULL) {
884 unlock_page(lnb->page);
885 page_cache_release(lnb->page);
892 up_read(&dentry->d_inode->i_alloc_sem);
894 filter_iobuf_put(&obd->u.filter, iobuf, oti);
896 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
901 filter_iobuf_put(&obd->u.filter, iobuf, oti);
903 cfs_spin_lock(&obd->obd_osfs_lock);
905 filter_grant_incoming(exp, oa);
906 cfs_spin_unlock(&obd->obd_osfs_lock);
907 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
914 int filter_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
915 int objcount, struct obd_ioobj *obj,
916 struct niobuf_remote *nb, int *npages,
917 struct niobuf_local *res, struct obd_trans_info *oti,
918 struct lustre_capa *capa)
920 if (cmd == OBD_BRW_WRITE)
921 return filter_preprw_write(cmd, exp, oa, objcount, obj,
922 nb, npages, res, oti, capa);
923 if (cmd == OBD_BRW_READ)
924 return filter_preprw_read(cmd, exp, oa, objcount, obj,
925 nb, npages, res, oti, capa);
930 static int filter_commitrw_read(struct obd_export *exp, struct obdo *oa,
931 int objcount, struct obd_ioobj *obj,
932 struct niobuf_remote *rnb,
933 int npages, struct niobuf_local *res,
934 struct obd_trans_info *oti, int rc)
936 struct filter_obd *fo = &exp->exp_obd->u.filter;
937 struct inode *inode = NULL;
938 struct ldlm_res_id res_id;
939 struct ldlm_resource *resource = NULL;
940 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
941 struct niobuf_local *lnb;
945 osc_build_res_name(obj->ioo_id, obj->ioo_seq, &res_id);
946 /* If oa != NULL then filter_preprw_read updated the inode atime
947 * and we should update the lvb so that other glimpses will also
948 * get the updated value. bug 5972 */
949 if (oa && ns && ns->ns_lvbo && ns->ns_lvbo->lvbo_update) {
950 resource = ldlm_resource_get(ns, NULL, &res_id, LDLM_EXTENT, 0);
952 if (resource != NULL) {
953 LDLM_RESOURCE_ADDREF(resource);
954 ns->ns_lvbo->lvbo_update(resource, NULL, 1);
955 LDLM_RESOURCE_DELREF(resource);
956 ldlm_resource_putref(resource);
960 if (res->dentry != NULL)
961 inode = res->dentry->d_inode;
963 for (i = 0, lnb = res; i < npages; i++, lnb++) {
964 if (lnb->page != NULL) {
965 page_cache_release(lnb->page);
969 if (inode && (fo->fo_read_cache == 0 ||
970 i_size_read(inode) > fo->fo_readcache_max_filesize))
971 filter_release_cache(exp->exp_obd, obj, rnb, inode);
973 if (res->dentry != NULL)
978 void filter_grant_commit(struct obd_export *exp, int niocount,
979 struct niobuf_local *res)
981 struct filter_obd *filter = &exp->exp_obd->u.filter;
982 struct niobuf_local *lnb = res;
983 unsigned long pending = 0;
986 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
987 for (i = 0, lnb = res; i < niocount; i++, lnb++)
988 pending += lnb->lnb_grant_used;
990 LASSERTF(exp->exp_filter_data.fed_pending >= pending,
991 "%s: cli %s/%p fed_pending: %lu grant_used: %lu\n",
992 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
993 exp->exp_filter_data.fed_pending, pending);
994 exp->exp_filter_data.fed_pending -= pending;
995 LASSERTF(filter->fo_tot_granted >= pending,
996 "%s: cli %s/%p tot_granted: "LPU64" grant_used: %lu\n",
997 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
998 exp->exp_obd->u.filter.fo_tot_granted, pending);
999 filter->fo_tot_granted -= pending;
1000 LASSERTF(filter->fo_tot_pending >= pending,
1001 "%s: cli %s/%p tot_pending: "LPU64" grant_used: %lu\n",
1002 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
1003 filter->fo_tot_pending, pending);
1004 filter->fo_tot_pending -= pending;
1006 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
1009 int filter_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
1010 int objcount, struct obd_ioobj *obj,
1011 struct niobuf_remote *nb, int npages,
1012 struct niobuf_local *res, struct obd_trans_info *oti,
1015 if (cmd == OBD_BRW_WRITE)
1016 return filter_commitrw_write(exp, oa, objcount, obj,
1017 nb, npages, res, oti, rc);
1018 if (cmd == OBD_BRW_READ)
1019 return filter_commitrw_read(exp, oa, objcount, obj,
1020 nb, npages, res, oti, rc);
1025 int filter_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1026 obd_count oa_bufs, struct brw_page *pga,
1027 struct obd_trans_info *oti)
1029 struct obd_ioobj ioo;
1030 struct niobuf_local *lnb;
1031 struct niobuf_remote *rnb;
1033 int ret = 0, npages;
1036 OBD_ALLOC(lnb, oa_bufs * sizeof(struct niobuf_local));
1037 OBD_ALLOC(rnb, oa_bufs * sizeof(struct niobuf_remote));
1039 if (lnb == NULL || rnb == NULL)
1040 GOTO(out, ret = -ENOMEM);
1042 for (i = 0; i < oa_bufs; i++) {
1043 lnb[i].page = pga[i].pg;
1044 rnb[i].offset = pga[i].off;
1045 rnb[i].len = pga[i].count;
1046 lnb[i].flags = rnb[i].flags = pga[i].flag;
1049 obdo_to_ioobj(oinfo->oi_oa, &ioo);
1050 ioo.ioo_bufcnt = oa_bufs;
1053 ret = filter_preprw(cmd, exp, oinfo->oi_oa, 1, &ioo,
1054 rnb, &npages, lnb, oti, oinfo_capa(oinfo));
1057 LASSERTF(oa_bufs == npages, "%u != %u\n", oa_bufs, npages);
1059 ret = filter_commitrw(cmd, exp, oinfo->oi_oa, 1, &ioo, rnb,
1060 npages, lnb, oti, ret);
1064 OBD_FREE(lnb, oa_bufs * sizeof(struct niobuf_local));
1066 OBD_FREE(rnb, oa_bufs * sizeof(struct niobuf_remote));