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 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_current_64() - CFS_HZ)) {
158 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb,
159 cfs_time_current_64() + CFS_HZ);
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 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)
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 invalidate above doesn't work during read because lnet pins pages.
345 * The truncate is used here instead to drop pages from cache
347 void filter_release_cache(struct obd_device *obd, struct obd_ioobj *obj,
348 struct niobuf_remote *rnb, struct inode *inode)
352 LASSERT(inode != NULL);
353 for (i = 0; i < obj->ioo_bufcnt; i++, rnb++) {
354 #ifdef HAVE_TRUNCATE_RANGE
355 /* remove pages in which range is fit */
356 truncate_inode_pages_range(inode->i_mapping,
357 rnb->offset & CFS_PAGE_MASK,
358 (rnb->offset + rnb->len - 1) |
361 /* use invalidate for old kernels */
362 invalidate_mapping_pages(inode->i_mapping,
363 rnb->offset >> CFS_PAGE_SHIFT,
364 (rnb->offset + rnb->len) >>
370 static int filter_preprw_read(int cmd, struct obd_export *exp, struct obdo *oa,
371 int objcount, struct obd_ioobj *obj,
372 struct niobuf_remote *nb,
373 int *npages, struct niobuf_local *res,
374 struct obd_trans_info *oti,
375 struct lustre_capa *capa)
377 struct obd_device *obd = exp->exp_obd;
378 struct timeval start, end;
379 struct lvfs_run_ctxt saved;
380 struct niobuf_local *lnb;
381 struct dentry *dentry = NULL;
382 struct inode *inode = NULL;
384 int rc = 0, i, tot_bytes = 0;
385 unsigned long now = jiffies;
390 /* We are currently not supporting multi-obj BRW_READ RPCS at all.
391 * When we do this function's dentry cleanup will need to be fixed.
392 * These values are verified in ost_brw_write() from the wire. */
393 LASSERTF(objcount == 1, "%d\n", objcount);
394 LASSERTF(obj->ioo_bufcnt > 0, "%d\n", obj->ioo_bufcnt);
396 rc = filter_validate_obdo(oa, exp);
400 rc = filter_auth_capa(exp, NULL, oa->o_seq, capa,
405 if (oa && oa->o_valid & OBD_MD_FLGRANT) {
406 cfs_spin_lock(&obd->obd_osfs_lock);
407 filter_grant_incoming(exp, oa);
409 if (!(oa->o_flags & OBD_FL_SHRINK_GRANT))
411 cfs_spin_unlock(&obd->obd_osfs_lock);
414 iobuf = filter_iobuf_get(&obd->u.filter, oti);
416 RETURN(PTR_ERR(iobuf));
418 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
419 dentry = filter_oa2dentry(obd, &oa->o_oi);
420 if (IS_ERR(dentry)) {
421 rc = PTR_ERR(dentry);
426 inode = dentry->d_inode;
427 isize = i_size_read(inode);
429 obdo_to_inode(inode, oa, OBD_MD_FLATIME);
431 rc = filter_map_remote_to_local(objcount, obj, nb, npages, res);
435 fsfilt_check_slow(obd, now, "preprw_read setup");
437 /* find pages for all segments, fill array with them */
438 cfs_gettimeofday(&start);
439 for (i = 0, lnb = res; i < *npages; i++, lnb++) {
441 lnb->dentry = dentry;
443 if (isize <= lnb->offset)
444 /* If there's no more data, abort early. lnb->rc == 0,
445 * so it's easy to detect later. */
448 lnb->page = filter_get_page(obd, inode, lnb->offset, 0);
449 if (lnb->page == NULL)
450 GOTO(cleanup, rc = -ENOMEM);
452 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS, 1);
454 if (isize < lnb->offset + lnb->len - 1)
455 lnb->rc = isize - lnb->offset;
459 tot_bytes += lnb->rc;
461 if (PageUptodate(lnb->page)) {
462 lprocfs_counter_add(obd->obd_stats,
463 LPROC_FILTER_CACHE_HIT, 1);
467 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_MISS, 1);
468 filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
470 cfs_gettimeofday(&end);
471 timediff = cfs_timeval_sub(&end, &start, NULL);
472 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
474 if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
475 GOTO(cleanup, rc = -ENOMEM);
477 fsfilt_check_slow(obd, now, "start_page_read");
479 rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf,
480 exp, NULL, NULL, NULL);
484 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_READ_BYTES, tot_bytes);
486 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
487 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
488 LPROC_FILTER_READ_BYTES, tot_bytes);
493 /* unlock pages to allow access from concurrent OST_READ */
494 for (i = 0, lnb = res; i < *npages; i++, lnb++) {
496 LASSERT(PageLocked(lnb->page));
497 unlock_page(lnb->page);
500 page_cache_release(lnb->page);
511 filter_iobuf_put(&obd->u.filter, iobuf, oti);
513 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
515 CERROR("io error %d\n", rc);
520 /* When clients have dirtied as much space as they've been granted they
521 * fall through to sync writes. These sync writes haven't been expressed
522 * in grants and need to error with ENOSPC when there isn't room in the
523 * filesystem for them after grants are taken into account. However,
524 * writeback of the dirty data that was already granted space can write
527 * Caller must hold obd_osfs_lock. */
528 static int filter_grant_check(struct obd_export *exp, struct obdo *oa,
529 int objcount, struct fsfilt_objinfo *fso,
530 int niocount, struct niobuf_local *lnb,
531 obd_size *left, struct inode *inode)
533 struct filter_export_data *fed = &exp->exp_filter_data;
534 int blocksize = exp->exp_obd->u.obt.obt_sb->s_blocksize;
535 unsigned long used = 0, ungranted = 0, using;
536 int i, rc = -ENOSPC, obj, n = 0;
538 LASSERT_SPIN_LOCKED(&exp->exp_obd->obd_osfs_lock);
540 for (obj = 0; obj < objcount; obj++) {
541 for (i = 0; i < fso[obj].fso_bufcnt; i++, n++) {
544 /* should match the code in osc_exit_cache */
546 bytes += lnb[n].offset & (blocksize - 1);
547 tmp = (lnb[n].offset + lnb[n].len) & (blocksize - 1);
549 bytes += blocksize - tmp;
551 if ((lnb[n].flags & OBD_BRW_FROM_GRANT) &&
552 (oa->o_valid & OBD_MD_FLGRANT)) {
553 if (fed->fed_grant < used + bytes) {
555 "%s: cli %s/%p claims %ld+%d "
556 "GRANT, real grant %lu idx %d\n",
557 exp->exp_obd->obd_name,
558 exp->exp_client_uuid.uuid, exp,
559 used, bytes, fed->fed_grant, n);
562 lnb[n].flags |= OBD_BRW_GRANTED;
563 lnb[n].lnb_grant_used = bytes;
564 CDEBUG(0, "idx %d used=%lu\n", n, used);
569 if (*left > ungranted + bytes) {
570 /* if enough space, pretend it was granted */
572 lnb[n].flags |= OBD_BRW_GRANTED;
573 lnb[n].lnb_grant_used = bytes;
574 CDEBUG(0, "idx %d ungranted=%lu\n",n,ungranted);
579 /* We can't check for already-mapped blocks here, as
580 * it requires dropping the osfs lock to do the bmap.
581 * Instead, we return ENOSPC and in that case we need
582 * to go through and verify if all of the blocks not
583 * marked BRW_GRANTED are already mapped and we can
584 * ignore this error. */
586 lnb[n].flags &= ~OBD_BRW_GRANTED;
587 CDEBUG(D_CACHE,"%s: cli %s/%p idx %d no space for %d\n",
588 exp->exp_obd->obd_name,
589 exp->exp_client_uuid.uuid, exp, n, bytes);
593 /* Now substract what client have used already. We don't subtract
594 * this from the tot_granted yet, so that other client's can't grab
595 * that space before we have actually allocated our blocks. That
596 * happens in filter_grant_commit() after the writes are done. */
598 fed->fed_grant -= used;
599 fed->fed_pending += used + ungranted;
600 exp->exp_obd->u.filter.fo_tot_granted += ungranted;
601 exp->exp_obd->u.filter.fo_tot_pending += used + ungranted;
604 "%s: cli %s/%p used: %lu ungranted: %lu grant: %lu dirty: %lu\n",
605 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp, used,
606 ungranted, fed->fed_grant, fed->fed_dirty);
608 /* Rough calc in case we don't refresh cached statfs data */
609 using = (used + ungranted + 1 ) >>
610 exp->exp_obd->u.obt.obt_sb->s_blocksize_bits;
611 if (exp->exp_obd->obd_osfs.os_bavail > using)
612 exp->exp_obd->obd_osfs.os_bavail -= using;
614 exp->exp_obd->obd_osfs.os_bavail = 0;
616 if (fed->fed_dirty < used) {
617 CERROR("%s: cli %s/%p claims used %lu > fed_dirty %lu\n",
618 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
619 used, fed->fed_dirty);
620 used = fed->fed_dirty;
622 exp->exp_obd->u.filter.fo_tot_dirty -= used;
623 fed->fed_dirty -= used;
625 if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
626 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
627 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
628 fed->fed_dirty, fed->fed_pending, fed->fed_grant);
629 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
635 /* If we ever start to support multi-object BRW RPCs, we will need to get locks
636 * on mulitple inodes. That isn't all, because there still exists the
637 * possibility of a truncate starting a new transaction while holding the ext3
638 * rwsem = write while some writes (which have started their transactions here)
639 * blocking on the ext3 rwsem = read => lock inversion.
641 * The handling gets very ugly when dealing with locked pages. It may be easier
642 * to just get rid of the locked page code (which has problems of its own) and
643 * either discover we do not need it anymore (i.e. it was a symptom of another
644 * bug) or ensure we get the page locks in an appropriate order. */
645 static int filter_preprw_write(int cmd, struct obd_export *exp, struct obdo *oa,
646 int objcount, struct obd_ioobj *obj,
647 struct niobuf_remote *nb, int *npages,
648 struct niobuf_local *res,
649 struct obd_trans_info *oti,
650 struct lustre_capa *capa)
652 struct obd_device *obd = exp->exp_obd;
653 struct timeval start, end;
654 struct lvfs_run_ctxt saved;
655 struct niobuf_local *lnb = res;
656 struct fsfilt_objinfo fso;
657 struct filter_mod_data *fmd;
658 struct dentry *dentry = NULL;
661 unsigned long now = jiffies, timediff;
662 int rc = 0, i, tot_bytes = 0, cleanup_phase = 0, localreq = 0;
664 LASSERT(objcount == 1);
665 LASSERT(obj->ioo_bufcnt > 0);
667 rc = filter_validate_obdo(oa, exp);
671 rc = filter_auth_capa(exp, NULL, oa->o_seq, capa,
676 if (exp->exp_connection &&
677 exp->exp_connection->c_peer.nid == exp->exp_connection->c_self)
680 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
681 iobuf = filter_iobuf_get(&obd->u.filter, oti);
683 GOTO(cleanup, rc = PTR_ERR(iobuf));
686 dentry = filter_fid2dentry(obd, NULL, obj->ioo_seq,
689 GOTO(cleanup, rc = PTR_ERR(dentry));
692 if (dentry->d_inode == NULL) {
693 CERROR("%s: trying to BRW to non-existent file "LPU64"\n",
694 obd->obd_name, obj->ioo_id);
695 GOTO(cleanup, rc = -ENOENT);
698 if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID) &&
699 dentry->d_inode->i_mode & (S_ISUID | S_ISGID)) {
700 rc = filter_capa_fixoa(exp, oa, oa->o_seq, capa);
705 rc = filter_map_remote_to_local(objcount, obj, nb, npages, res);
709 fsfilt_check_slow(obd, now, "preprw_write setup");
711 /* Filter truncate first locks i_mutex then partially truncated
712 * page, filter write code first locks pages then take
713 * i_mutex. To avoid a deadlock in case of concurrent
714 * punch/write requests from one client, filter writes and
715 * filter truncates are serialized by i_alloc_sem, allowing
716 * multiple writes or single truncate. */
717 down_read(&dentry->d_inode->i_alloc_sem);
719 /* Don't update inode timestamps if this write is older than a
720 * setattr which modifies the timestamps. b=10150 */
721 /* XXX when we start having persistent reservations this needs to
722 * be changed to filter_fmd_get() to create the fmd if it doesn't
723 * already exist so we can store the reservation handle there. */
724 fmd = filter_fmd_find(exp, obj->ioo_id, obj->ioo_seq);
727 cfs_spin_lock(&obd->obd_osfs_lock);
728 filter_grant_incoming(exp, oa);
729 if (fmd && fmd->fmd_mactime_xid > oti->oti_xid)
730 oa->o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLCTIME |
733 obdo_to_inode(dentry->d_inode, oa, OBD_MD_FLATIME |
734 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
737 left = filter_grant_space_left(exp);
739 fso.fso_dentry = dentry;
740 fso.fso_bufcnt = *npages;
742 rc = filter_grant_check(exp, oa, objcount, &fso, *npages, res,
743 &left, dentry->d_inode);
745 /* do not zero out oa->o_valid as it is used in filter_commitrw_write()
746 * for setting UID/GID and fid EA in first write time. */
747 if (oa->o_valid & OBD_MD_FLGRANT)
748 oa->o_grant = filter_grant(exp, oa->o_grant, oa->o_undirty,
751 cfs_spin_unlock(&obd->obd_osfs_lock);
752 filter_fmd_put(exp, fmd);
754 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK2, (obd_timeout + 1) / 4);
760 cfs_gettimeofday(&start);
761 for (i = 0, lnb = res; i < *npages; i++, lnb++) {
763 /* We still set up for ungranted pages so that granted pages
764 * can be written to disk as they were promised, and portals
765 * needs to keep the pages all aligned properly. */
766 lnb->dentry = dentry;
768 lnb->page = filter_get_page(obd, dentry->d_inode, lnb->offset,
770 if (lnb->page == NULL)
771 GOTO(cleanup, rc = -ENOMEM);
773 /* DLM locking protects us from write and truncate competing
774 * for same region, but truncate can leave dirty page in the
775 * cache. it's possible the writeout on a such a page is in
776 * progress when we access it. it's also possible that during
777 * this writeout we put new (partial) data, but then won't
778 * be able to proceed in filter_commitrw_write(). thus let's
779 * just wait for writeout completion, should be rare enough.
781 wait_on_page_writeback(lnb->page);
782 BUG_ON(PageWriteback(lnb->page));
784 /* If the filter writes a partial page, then has the file
785 * extended, the client will read in the whole page. the
786 * filter has to be careful to zero the rest of the partial
787 * page on disk. we do it by hand for partial extending
788 * writes, send_bio() is responsible for zeroing pages when
789 * asked to read unmapped blocks -- brw_kiovec() does this. */
790 if (lnb->len != CFS_PAGE_SIZE) {
793 maxidx = ((i_size_read(dentry->d_inode) +
794 CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT) - 1;
795 if (maxidx >= lnb->page->index) {
796 LL_CDEBUG_PAGE(D_PAGE, lnb->page, "write %u @ "
797 LPU64" flg %x before EOF %llu\n",
798 lnb->len, lnb->offset,lnb->flags,
799 i_size_read(dentry->d_inode));
800 filter_iobuf_add_page(obd, iobuf,
805 char *p = kmap(lnb->page);
807 off = lnb->offset & ~CFS_PAGE_MASK;
810 off = (lnb->offset + lnb->len) & ~CFS_PAGE_MASK;
812 memset(p + off, 0, CFS_PAGE_SIZE - off);
817 tot_bytes += lnb->len;
819 cfs_gettimeofday(&end);
820 timediff = cfs_timeval_sub(&end, &start, NULL);
821 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
823 if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
824 GOTO(cleanup, rc = -ENOMEM);
826 /* don't unlock pages to prevent any access */
827 rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf, exp,
830 fsfilt_check_slow(obd, now, "start_page_write");
832 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
835 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
836 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
837 LPROC_FILTER_WRITE_BYTES, tot_bytes);
840 switch(cleanup_phase) {
843 for (i = 0, lnb = res; i < *npages; i++, lnb++) {
844 if (lnb->page != NULL) {
845 unlock_page(lnb->page);
846 page_cache_release(lnb->page);
853 up_read(&dentry->d_inode->i_alloc_sem);
855 filter_iobuf_put(&obd->u.filter, iobuf, oti);
857 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
862 filter_iobuf_put(&obd->u.filter, iobuf, oti);
864 cfs_spin_lock(&obd->obd_osfs_lock);
866 filter_grant_incoming(exp, oa);
867 cfs_spin_unlock(&obd->obd_osfs_lock);
868 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
875 int filter_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
876 int objcount, struct obd_ioobj *obj,
877 struct niobuf_remote *nb, int *npages,
878 struct niobuf_local *res, struct obd_trans_info *oti,
879 struct lustre_capa *capa)
881 if (cmd == OBD_BRW_WRITE)
882 return filter_preprw_write(cmd, exp, oa, objcount, obj,
883 nb, npages, res, oti, capa);
884 if (cmd == OBD_BRW_READ)
885 return filter_preprw_read(cmd, exp, oa, objcount, obj,
886 nb, npages, res, oti, capa);
891 static int filter_commitrw_read(struct obd_export *exp, struct obdo *oa,
892 int objcount, struct obd_ioobj *obj,
893 struct niobuf_remote *rnb,
894 int npages, struct niobuf_local *res,
895 struct obd_trans_info *oti, int rc)
897 struct filter_obd *fo = &exp->exp_obd->u.filter;
898 struct inode *inode = NULL;
899 struct ldlm_res_id res_id;
900 struct ldlm_resource *resource = NULL;
901 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
902 struct niobuf_local *lnb;
906 osc_build_res_name(obj->ioo_id, obj->ioo_seq, &res_id);
907 /* If oa != NULL then filter_preprw_read updated the inode atime
908 * and we should update the lvb so that other glimpses will also
909 * get the updated value. bug 5972 */
910 if (oa && ns && ns->ns_lvbo && ns->ns_lvbo->lvbo_update) {
911 resource = ldlm_resource_get(ns, NULL, &res_id, LDLM_EXTENT, 0);
913 if (resource != NULL) {
914 LDLM_RESOURCE_ADDREF(resource);
915 ns->ns_lvbo->lvbo_update(resource, NULL, 1);
916 LDLM_RESOURCE_DELREF(resource);
917 ldlm_resource_putref(resource);
921 if (res->dentry != NULL)
922 inode = res->dentry->d_inode;
924 for (i = 0, lnb = res; i < npages; i++, lnb++) {
925 if (lnb->page != NULL) {
926 page_cache_release(lnb->page);
930 if (inode && (fo->fo_read_cache == 0 ||
931 i_size_read(inode) > fo->fo_readcache_max_filesize))
932 filter_release_cache(exp->exp_obd, obj, rnb, inode);
934 if (res->dentry != NULL)
939 void filter_grant_commit(struct obd_export *exp, int niocount,
940 struct niobuf_local *res)
942 struct filter_obd *filter = &exp->exp_obd->u.filter;
943 struct niobuf_local *lnb = res;
944 unsigned long pending = 0;
947 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
948 for (i = 0, lnb = res; i < niocount; i++, lnb++)
949 pending += lnb->lnb_grant_used;
951 LASSERTF(exp->exp_filter_data.fed_pending >= pending,
952 "%s: cli %s/%p fed_pending: %lu grant_used: %lu\n",
953 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
954 exp->exp_filter_data.fed_pending, pending);
955 exp->exp_filter_data.fed_pending -= pending;
956 LASSERTF(filter->fo_tot_granted >= pending,
957 "%s: cli %s/%p tot_granted: "LPU64" grant_used: %lu\n",
958 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
959 exp->exp_obd->u.filter.fo_tot_granted, pending);
960 filter->fo_tot_granted -= pending;
961 LASSERTF(filter->fo_tot_pending >= pending,
962 "%s: cli %s/%p tot_pending: "LPU64" grant_used: %lu\n",
963 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
964 filter->fo_tot_pending, pending);
965 filter->fo_tot_pending -= pending;
967 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
970 int filter_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
971 int objcount, struct obd_ioobj *obj,
972 struct niobuf_remote *nb, int npages,
973 struct niobuf_local *res, struct obd_trans_info *oti,
976 if (cmd == OBD_BRW_WRITE)
977 return filter_commitrw_write(exp, oa, objcount, obj,
978 nb, npages, res, oti, rc);
979 if (cmd == OBD_BRW_READ)
980 return filter_commitrw_read(exp, oa, objcount, obj,
981 nb, npages, res, oti, rc);
986 int filter_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
987 obd_count oa_bufs, struct brw_page *pga,
988 struct obd_trans_info *oti)
990 struct obd_ioobj ioo;
991 struct niobuf_local *lnb;
992 struct niobuf_remote *rnb;
997 OBD_ALLOC(lnb, oa_bufs * sizeof(struct niobuf_local));
998 OBD_ALLOC(rnb, oa_bufs * sizeof(struct niobuf_remote));
1000 if (lnb == NULL || rnb == NULL)
1001 GOTO(out, ret = -ENOMEM);
1003 for (i = 0; i < oa_bufs; i++) {
1004 lnb[i].page = pga[i].pg;
1005 rnb[i].offset = pga[i].off;
1006 rnb[i].len = pga[i].count;
1009 obdo_to_ioobj(oinfo->oi_oa, &ioo);
1010 ioo.ioo_bufcnt = oa_bufs;
1013 ret = filter_preprw(cmd, exp, oinfo->oi_oa, 1, &ioo,
1014 rnb, &npages, lnb, oti, oinfo_capa(oinfo));
1017 LASSERTF(oa_bufs == npages, "%u != %u\n", oa_bufs, npages);
1019 ret = filter_commitrw(cmd, exp, oinfo->oi_oa, 1, &ioo, rnb,
1020 npages, lnb, oti, ret);
1024 OBD_FREE(lnb, oa_bufs * sizeof(struct niobuf_local));
1026 OBD_FREE(rnb, oa_bufs * sizeof(struct niobuf_remote));