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 static 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;
111 if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
112 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
113 obd->obd_name, exp->exp_client_uuid.uuid, exp,
114 fed->fed_dirty, fed->fed_pending, fed->fed_grant);
115 spin_unlock(&obd->obd_osfs_lock);
121 /* Figure out how much space is available between what we've granted
122 * and what remains in the filesystem. Compensate for ext3 indirect
123 * block overhead when computing how much free space is left ungranted.
125 * Caller must hold obd_osfs_lock. */
126 obd_size filter_grant_space_left(struct obd_export *exp)
128 struct obd_device *obd = exp->exp_obd;
129 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
130 obd_size tot_granted = obd->u.filter.fo_tot_granted, avail, left = 0;
131 int rc, statfs_done = 0;
133 LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
135 if (cfs_time_before_64(obd->obd_osfs_age, cfs_time_current_64() - HZ)) {
137 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb,
138 cfs_time_current_64() + HZ);
139 if (rc) /* N.B. statfs can't really fail */
144 avail = obd->obd_osfs.os_bavail;
145 left = avail - (avail >> (blockbits - 3)); /* (d)indirect */
146 if (left > GRANT_FOR_LLOG(obd)) {
147 left = (left - GRANT_FOR_LLOG(obd)) << blockbits;
149 left = 0 /* << blockbits */;
152 if (!statfs_done && left < 32 * FILTER_GRANT_CHUNK + tot_granted) {
153 CDEBUG(D_CACHE, "fs has no space left and statfs too old\n");
157 if (left >= tot_granted) {
160 if (left < tot_granted - obd->u.filter.fo_tot_pending) {
161 CERROR("%s: cli %s/%p grant "LPU64" > available "
162 LPU64" and pending "LPU64"\n", obd->obd_name,
163 exp->exp_client_uuid.uuid, exp, tot_granted,
164 left, obd->u.filter.fo_tot_pending);
169 CDEBUG(D_CACHE, "%s: cli %s/%p free: "LPU64" avail: "LPU64" grant "LPU64
170 " left: "LPU64" pending: "LPU64"\n", obd->obd_name,
171 exp->exp_client_uuid.uuid, exp,
172 obd->obd_osfs.os_bfree << blockbits, avail << blockbits,
173 tot_granted, left, obd->u.filter.fo_tot_pending);
178 /* Calculate how much grant space to allocate to this client, based on how
179 * much space is currently free and how much of that is already granted.
181 * Caller must hold obd_osfs_lock. */
182 long filter_grant(struct obd_export *exp, obd_size current_grant,
183 obd_size want, obd_size fs_space_left)
185 struct obd_device *obd = exp->exp_obd;
186 struct filter_export_data *fed = &exp->exp_filter_data;
187 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
190 LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
192 /* Grant some fraction of the client's requested grant space so that
193 * they are not always waiting for write credits (not all of it to
194 * avoid overgranting in face of multiple RPCs in flight). This
195 * essentially will be able to control the OSC_MAX_RIF for a client.
197 * If we do have a large disparity between what the client thinks it
198 * has and what we think it has, don't grant very much and let the
199 * client consume its grant first. Either it just has lots of RPCs
200 * in flight, or it was evicted and its grants will soon be used up. */
201 if (want > 0x7fffffff) {
202 CERROR("%s: client %s/%p requesting > 2GB grant "LPU64"\n",
203 obd->obd_name, exp->exp_client_uuid.uuid, exp, want);
204 } else if (current_grant < want &&
205 current_grant < fed->fed_grant + FILTER_GRANT_CHUNK) {
206 grant = min((want >> blockbits),
207 (fs_space_left >> blockbits) / 8);
211 /* Allow >FILTER_GRANT_CHUNK size when clients
212 * reconnect due to a server reboot.
214 if ((grant > FILTER_GRANT_CHUNK) &&
215 (!obd->obd_recovering))
216 grant = FILTER_GRANT_CHUNK;
218 obd->u.filter.fo_tot_granted += grant;
219 fed->fed_grant += grant;
220 if (fed->fed_grant < 0) {
221 CERROR("%s: cli %s/%p grant %ld want "LPU64
223 obd->obd_name, exp->exp_client_uuid.uuid,
224 exp, fed->fed_grant, want,current_grant);
225 spin_unlock(&obd->obd_osfs_lock);
232 "%s: cli %s/%p wants: "LPU64" current grant "LPU64
233 " granting: "LPU64"\n", obd->obd_name, exp->exp_client_uuid.uuid,
234 exp, want, current_grant, grant);
236 "%s: cli %s/%p tot cached:"LPU64" granted:"LPU64
237 " num_exports: %d\n", obd->obd_name, exp->exp_client_uuid.uuid,
238 exp, obd->u.filter.fo_tot_dirty,
239 obd->u.filter.fo_tot_granted, obd->obd_num_exports);
245 * the routine is used to request pages from pagecache
247 * use GFP_NOFS not allowing to enter FS as the client can run on this node
248 * and we might end waiting on a page he sent in the request we're serving.
250 * use NORETRY so that the allocator doesn't go crazy: chance to more lucky
251 * thread have enough memory to complete his request. for our request client
252 * will do resend hopefully -bzzz
254 static struct page * filter_get_page(struct obd_device *obd,
260 page = find_or_create_page(inode->i_mapping, offset >> CFS_PAGE_SHIFT,
261 GFP_NOFS | __GFP_NORETRY);
262 if (unlikely(page == NULL))
263 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_NO_PAGE, 1);
269 * the routine initializes array of local_niobuf from remote_niobuf
271 static int filter_map_remote_to_local(int objcount, struct obd_ioobj *obj,
272 struct niobuf_remote *nb,
273 int *nrpages, struct niobuf_local *res)
275 struct niobuf_remote *rnb;
276 struct niobuf_local *lnb;
280 /* we don't support multiobject RPC yet
281 * ost_brw_read() and ost_brw_write() check this */
282 LASSERT(objcount == 1);
286 for (i = 0, rnb = nb, lnb = res; i < obj->ioo_bufcnt; i++, rnb++) {
287 obd_off offset = rnb->offset;
288 unsigned int len = rnb->len;
291 int poff = offset & (CFS_PAGE_SIZE - 1);
292 int plen = CFS_PAGE_SIZE - poff;
294 if (*nrpages >= max) {
295 CERROR("small array of local bufs: %d\n", max);
301 lnb->offset = offset;
303 lnb->flags = rnb->flags;
306 lnb->lnb_grant_used = 0;
308 LASSERTF(plen <= len, "plen %u, len %u\n", plen, len);
319 * the function is used to free all pages used for request
320 * just to mimic cacheless OSS which don't occupy much memory
322 void filter_invalidate_cache(struct obd_device *obd, struct obd_ioobj *obj,
323 struct niobuf_remote *nb, struct inode *inode)
325 struct niobuf_remote *rnb;
328 LASSERT(inode != NULL);
330 for (i = 0, rnb = nb; i < obj->ioo_bufcnt; i++, rnb++) {
331 obd_off start = rnb->offset >> CFS_PAGE_SHIFT;
332 obd_off end = (rnb->offset + rnb->len) >> CFS_PAGE_SHIFT;
333 invalidate_mapping_pages(inode->i_mapping, start, end);
334 /* just to avoid warnings */
340 static int filter_preprw_read(int cmd, struct obd_export *exp, struct obdo *oa,
341 int objcount, struct obd_ioobj *obj,
342 struct niobuf_remote *nb,
343 int *npages, struct niobuf_local *res,
344 struct obd_trans_info *oti,
345 struct lustre_capa *capa)
347 struct obd_device *obd = exp->exp_obd;
348 struct timeval start, end;
349 struct lvfs_run_ctxt saved;
350 struct niobuf_local *lnb;
351 struct dentry *dentry = NULL;
352 struct inode *inode = NULL;
354 int rc = 0, i, tot_bytes = 0;
355 unsigned long now = jiffies;
359 /* We are currently not supporting multi-obj BRW_READ RPCS at all.
360 * When we do this function's dentry cleanup will need to be fixed.
361 * These values are verified in ost_brw_write() from the wire. */
362 LASSERTF(objcount == 1, "%d\n", objcount);
363 LASSERTF(obj->ioo_bufcnt > 0, "%d\n", obj->ioo_bufcnt);
365 rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa), capa,
370 if (oa && oa->o_valid & OBD_MD_FLGRANT) {
371 spin_lock(&obd->obd_osfs_lock);
372 filter_grant_incoming(exp, oa);
375 spin_unlock(&obd->obd_osfs_lock);
378 iobuf = filter_iobuf_get(&obd->u.filter, oti);
380 RETURN(PTR_ERR(iobuf));
382 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
383 dentry = filter_oa2dentry(obd, oa);
384 if (IS_ERR(dentry)) {
385 rc = PTR_ERR(dentry);
390 inode = dentry->d_inode;
392 obdo_to_inode(inode, oa, OBD_MD_FLATIME);
394 rc = filter_map_remote_to_local(objcount, obj, nb, npages, res);
398 fsfilt_check_slow(obd, now, "preprw_read setup");
400 /* find pages for all segments, fill array with them */
401 do_gettimeofday(&start);
402 for (i = 0, lnb = res; i < *npages; i++, lnb++) {
404 lnb->dentry = dentry;
406 if (i_size_read(inode) <= lnb->offset)
407 /* If there's no more data, abort early. lnb->rc == 0,
408 * so it's easy to detect later. */
411 lnb->page = filter_get_page(obd, inode, lnb->offset);
412 if (lnb->page == NULL)
413 GOTO(cleanup, rc = -ENOMEM);
415 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS, 1);
417 if (i_size_read(inode) < lnb->offset + lnb->len - 1)
418 lnb->rc = i_size_read(inode) - lnb->offset;
422 tot_bytes += lnb->rc;
424 if (PageUptodate(lnb->page)) {
425 lprocfs_counter_add(obd->obd_stats,
426 LPROC_FILTER_CACHE_HIT, 1);
430 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_MISS, 1);
431 filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
433 do_gettimeofday(&end);
434 timediff = cfs_timeval_sub(&end, &start, NULL);
435 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
437 if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
438 GOTO(cleanup, rc = -ENOMEM);
440 fsfilt_check_slow(obd, now, "start_page_read");
442 rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf,
443 exp, NULL, NULL, NULL);
447 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_READ_BYTES, tot_bytes);
449 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
450 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
451 LPROC_FILTER_READ_BYTES, tot_bytes);
456 /* unlock pages to allow access from concurrent OST_READ */
457 for (i = 0, lnb = res; i < *npages; i++, lnb++) {
459 LASSERT(PageLocked(lnb->page));
460 unlock_page(lnb->page);
463 page_cache_release(lnb->page);
474 filter_iobuf_put(&obd->u.filter, iobuf, oti);
476 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
478 CERROR("io error %d\n", rc);
483 /* When clients have dirtied as much space as they've been granted they
484 * fall through to sync writes. These sync writes haven't been expressed
485 * in grants and need to error with ENOSPC when there isn't room in the
486 * filesystem for them after grants are taken into account. However,
487 * writeback of the dirty data that was already granted space can write
490 * Caller must hold obd_osfs_lock. */
491 static int filter_grant_check(struct obd_export *exp, struct obdo *oa,
492 int objcount, struct fsfilt_objinfo *fso,
493 int niocount, struct niobuf_local *lnb,
494 obd_size *left, struct inode *inode)
496 struct filter_export_data *fed = &exp->exp_filter_data;
497 int blocksize = exp->exp_obd->u.obt.obt_sb->s_blocksize;
498 unsigned long used = 0, ungranted = 0, using;
499 int i, rc = -ENOSPC, obj, n = 0;
501 LASSERT_SPIN_LOCKED(&exp->exp_obd->obd_osfs_lock);
503 for (obj = 0; obj < objcount; obj++) {
504 for (i = 0; i < fso[obj].fso_bufcnt; i++, n++) {
507 /* should match the code in osc_exit_cache */
509 bytes += lnb[n].offset & (blocksize - 1);
510 tmp = (lnb[n].offset + lnb[n].len) & (blocksize - 1);
512 bytes += blocksize - tmp;
514 if ((lnb[n].flags & OBD_BRW_FROM_GRANT) &&
515 (oa->o_valid & OBD_MD_FLGRANT)) {
516 if (fed->fed_grant < used + bytes) {
518 "%s: cli %s/%p claims %ld+%d "
519 "GRANT, real grant %lu idx %d\n",
520 exp->exp_obd->obd_name,
521 exp->exp_client_uuid.uuid, exp,
522 used, bytes, fed->fed_grant, n);
525 lnb[n].flags |= OBD_BRW_GRANTED;
526 lnb[n].lnb_grant_used = bytes;
527 CDEBUG(0, "idx %d used=%lu\n", n, used);
532 if (*left > ungranted + bytes) {
533 /* if enough space, pretend it was granted */
535 lnb[n].flags |= OBD_BRW_GRANTED;
536 lnb[n].lnb_grant_used = bytes;
537 CDEBUG(0, "idx %d ungranted=%lu\n",n,ungranted);
542 /* We can't check for already-mapped blocks here, as
543 * it requires dropping the osfs lock to do the bmap.
544 * Instead, we return ENOSPC and in that case we need
545 * to go through and verify if all of the blocks not
546 * marked BRW_GRANTED are already mapped and we can
547 * ignore this error. */
549 lnb[n].flags &= ~OBD_BRW_GRANTED;
550 CDEBUG(D_CACHE,"%s: cli %s/%p idx %d no space for %d\n",
551 exp->exp_obd->obd_name,
552 exp->exp_client_uuid.uuid, exp, n, bytes);
556 /* Now substract what client have used already. We don't subtract
557 * this from the tot_granted yet, so that other client's can't grab
558 * that space before we have actually allocated our blocks. That
559 * happens in filter_grant_commit() after the writes are done. */
561 fed->fed_grant -= used;
562 fed->fed_pending += used + ungranted;
563 exp->exp_obd->u.filter.fo_tot_granted += ungranted;
564 exp->exp_obd->u.filter.fo_tot_pending += used + ungranted;
567 "%s: cli %s/%p used: %lu ungranted: %lu grant: %lu dirty: %lu\n",
568 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp, used,
569 ungranted, fed->fed_grant, fed->fed_dirty);
571 /* Rough calc in case we don't refresh cached statfs data */
572 using = (used + ungranted + 1 ) >>
573 exp->exp_obd->u.obt.obt_sb->s_blocksize_bits;
574 if (exp->exp_obd->obd_osfs.os_bavail > using)
575 exp->exp_obd->obd_osfs.os_bavail -= using;
577 exp->exp_obd->obd_osfs.os_bavail = 0;
579 if (fed->fed_dirty < used) {
580 CERROR("%s: cli %s/%p claims used %lu > fed_dirty %lu\n",
581 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
582 used, fed->fed_dirty);
583 used = fed->fed_dirty;
585 exp->exp_obd->u.filter.fo_tot_dirty -= used;
586 fed->fed_dirty -= used;
588 if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
589 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
590 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
591 fed->fed_dirty, fed->fed_pending, fed->fed_grant);
592 spin_unlock(&exp->exp_obd->obd_osfs_lock);
598 /* If we ever start to support multi-object BRW RPCs, we will need to get locks
599 * on mulitple inodes. That isn't all, because there still exists the
600 * possibility of a truncate starting a new transaction while holding the ext3
601 * rwsem = write while some writes (which have started their transactions here)
602 * blocking on the ext3 rwsem = read => lock inversion.
604 * The handling gets very ugly when dealing with locked pages. It may be easier
605 * to just get rid of the locked page code (which has problems of its own) and
606 * either discover we do not need it anymore (i.e. it was a symptom of another
607 * bug) or ensure we get the page locks in an appropriate order. */
608 static int filter_preprw_write(int cmd, struct obd_export *exp, struct obdo *oa,
609 int objcount, struct obd_ioobj *obj,
610 struct niobuf_remote *nb, int *npages,
611 struct niobuf_local *res,
612 struct obd_trans_info *oti,
613 struct lustre_capa *capa)
615 struct obd_device *obd = exp->exp_obd;
616 struct timeval start, end;
617 struct lvfs_run_ctxt saved;
618 struct niobuf_local *lnb = res;
619 struct fsfilt_objinfo fso;
620 struct filter_mod_data *fmd;
621 struct dentry *dentry = NULL;
624 unsigned long now = jiffies, timediff;
625 int rc = 0, i, tot_bytes = 0, cleanup_phase = 0;
627 LASSERT(objcount == 1);
628 LASSERT(obj->ioo_bufcnt > 0);
630 rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa), capa,
635 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
636 iobuf = filter_iobuf_get(&exp->exp_obd->u.filter, oti);
638 GOTO(cleanup, rc = PTR_ERR(iobuf));
641 dentry = filter_fid2dentry(exp->exp_obd, NULL, obj->ioo_gr,
644 GOTO(cleanup, rc = PTR_ERR(dentry));
647 if (dentry->d_inode == NULL) {
648 CERROR("%s: trying to BRW to non-existent file "LPU64"\n",
649 exp->exp_obd->obd_name, obj->ioo_id);
650 GOTO(cleanup, rc = -ENOENT);
653 rc = filter_map_remote_to_local(objcount, obj, nb, npages, res);
657 fsfilt_check_slow(exp->exp_obd, now, "preprw_write setup");
659 /* Don't update inode timestamps if this write is older than a
660 * setattr which modifies the timestamps. b=10150 */
661 /* XXX when we start having persistent reservations this needs to
662 * be changed to filter_fmd_get() to create the fmd if it doesn't
663 * already exist so we can store the reservation handle there. */
664 fmd = filter_fmd_find(exp, obj->ioo_id, obj->ioo_gr);
667 spin_lock(&exp->exp_obd->obd_osfs_lock);
668 filter_grant_incoming(exp, oa);
669 if (fmd && fmd->fmd_mactime_xid > oti->oti_xid)
670 oa->o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLCTIME |
673 obdo_to_inode(dentry->d_inode, oa, OBD_MD_FLATIME |
674 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
677 left = filter_grant_space_left(exp);
679 fso.fso_dentry = dentry;
680 fso.fso_bufcnt = *npages;
682 rc = filter_grant_check(exp, oa, objcount, &fso, *npages, res,
683 &left, dentry->d_inode);
685 /* do not zero out oa->o_valid as it is used in filter_commitrw_write()
686 * for setting UID/GID and fid EA in first write time. */
687 if (oa->o_valid & OBD_MD_FLGRANT)
688 oa->o_grant = filter_grant(exp,oa->o_grant,oa->o_undirty,left);
690 spin_unlock(&exp->exp_obd->obd_osfs_lock);
691 filter_fmd_put(exp, fmd);
696 do_gettimeofday(&start);
697 for (i = 0, lnb = res; i < *npages; i++, lnb++) {
699 /* We still set up for ungranted pages so that granted pages
700 * can be written to disk as they were promised, and portals
701 * needs to keep the pages all aligned properly. */
702 lnb->dentry = dentry;
704 lnb->page = filter_get_page(obd, dentry->d_inode, lnb->offset);
705 if (lnb->page == NULL)
706 GOTO(cleanup, rc = -ENOMEM);
709 /* DLM locking protects us from write and truncate competing
710 * for same region, but truncate can leave dirty page in the
711 * cache. it's possible the writeout on a such a page is in
712 * progress when we access it. it's also possible that during
713 * this writeout we put new (partial) data, but then won't
714 * be able to proceed in filter_commitrw_write(). thus let's
715 * just wait for writeout completion, should be rare enough.
717 if (obd->u.filter.fo_writethrough_cache)
718 wait_on_page_writeback(lnb->page);
719 BUG_ON(PageWriteback(lnb->page));
721 /* If the filter writes a partial page, then has the file
722 * extended, the client will read in the whole page. the
723 * filter has to be careful to zero the rest of the partial
724 * page on disk. we do it by hand for partial extending
725 * writes, send_bio() is responsible for zeroing pages when
726 * asked to read unmapped blocks -- brw_kiovec() does this. */
727 if (lnb->len != CFS_PAGE_SIZE) {
730 maxidx = ((i_size_read(dentry->d_inode) +
731 CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT) - 1;
732 if (maxidx >= lnb->page->index) {
733 LL_CDEBUG_PAGE(D_PAGE, lnb->page, "write %u @ "
734 LPU64" flg %x before EOF %llu\n",
735 lnb->len, lnb->offset,lnb->flags,
736 i_size_read(dentry->d_inode));
737 filter_iobuf_add_page(exp->exp_obd, iobuf,
742 char *p = kmap(lnb->page);
744 off = lnb->offset & ~CFS_PAGE_MASK;
747 off = (lnb->offset + lnb->len) & ~CFS_PAGE_MASK;
749 memset(p + off, 0, CFS_PAGE_SIZE - off);
754 tot_bytes += lnb->len;
756 do_gettimeofday(&end);
757 timediff = cfs_timeval_sub(&end, &start, NULL);
758 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
760 if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
761 GOTO(cleanup, rc = -ENOMEM);
763 /* don't unlock pages to prevent any access */
764 rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf, exp,
767 fsfilt_check_slow(exp->exp_obd, now, "start_page_write");
769 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
770 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
771 LPROC_FILTER_WRITE_BYTES, tot_bytes);
774 switch(cleanup_phase) {
777 for (i = 0, lnb = res; i < *npages; i++, lnb++) {
778 if (lnb->page != NULL) {
779 unlock_page(lnb->page);
780 page_cache_release(lnb->page);
786 filter_iobuf_put(&exp->exp_obd->u.filter, iobuf, oti);
788 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
793 filter_iobuf_put(&exp->exp_obd->u.filter, iobuf, oti);
795 spin_lock(&exp->exp_obd->obd_osfs_lock);
797 filter_grant_incoming(exp, oa);
798 spin_unlock(&exp->exp_obd->obd_osfs_lock);
799 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
806 int filter_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
807 int objcount, struct obd_ioobj *obj,
808 struct niobuf_remote *nb, int *npages,
809 struct niobuf_local *res, struct obd_trans_info *oti,
810 struct lustre_capa *capa)
812 if (cmd == OBD_BRW_WRITE)
813 return filter_preprw_write(cmd, exp, oa, objcount, obj,
814 nb, npages, res, oti, capa);
815 if (cmd == OBD_BRW_READ)
816 return filter_preprw_read(cmd, exp, oa, objcount, obj,
817 nb, npages, res, oti, capa);
822 static int filter_commitrw_read(struct obd_export *exp, struct obdo *oa,
823 int objcount, struct obd_ioobj *obj,
824 struct niobuf_remote *rnb,
825 int npages, struct niobuf_local *res,
826 struct obd_trans_info *oti, int rc)
828 struct filter_obd *fo = &exp->exp_obd->u.filter;
829 struct inode *inode = NULL;
830 struct ldlm_res_id res_id;
831 struct ldlm_resource *resource = NULL;
832 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
833 struct niobuf_local *lnb;
837 osc_build_res_name(obj->ioo_id, obj->ioo_gr, &res_id);
838 /* If oa != NULL then filter_preprw_read updated the inode atime
839 * and we should update the lvb so that other glimpses will also
840 * get the updated value. bug 5972 */
841 if (oa && ns && ns->ns_lvbo && ns->ns_lvbo->lvbo_update) {
842 resource = ldlm_resource_get(ns, NULL, &res_id, LDLM_EXTENT, 0);
844 if (resource != NULL) {
845 LDLM_RESOURCE_ADDREF(resource);
846 ns->ns_lvbo->lvbo_update(resource, NULL, 0, 1);
847 LDLM_RESOURCE_DELREF(resource);
848 ldlm_resource_putref(resource);
852 if (res->dentry != NULL)
853 inode = res->dentry->d_inode;
855 for (i = 0, lnb = res; i < npages; i++, lnb++) {
856 if (lnb->page != NULL) {
857 page_cache_release(lnb->page);
862 if (inode && (fo->fo_read_cache == 0 ||
863 i_size_read(inode) > fo->fo_readcache_max_filesize))
864 filter_invalidate_cache(exp->exp_obd, obj, rnb, inode);
866 if (res->dentry != NULL)
871 void filter_grant_commit(struct obd_export *exp, int niocount,
872 struct niobuf_local *res)
874 struct filter_obd *filter = &exp->exp_obd->u.filter;
875 struct niobuf_local *lnb = res;
876 unsigned long pending = 0;
879 spin_lock(&exp->exp_obd->obd_osfs_lock);
880 for (i = 0, lnb = res; i < niocount; i++, lnb++)
881 pending += lnb->lnb_grant_used;
883 LASSERTF(exp->exp_filter_data.fed_pending >= pending,
884 "%s: cli %s/%p fed_pending: %lu grant_used: %lu\n",
885 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
886 exp->exp_filter_data.fed_pending, pending);
887 exp->exp_filter_data.fed_pending -= pending;
888 LASSERTF(filter->fo_tot_granted >= pending,
889 "%s: cli %s/%p tot_granted: "LPU64" grant_used: %lu\n",
890 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
891 exp->exp_obd->u.filter.fo_tot_granted, pending);
892 filter->fo_tot_granted -= pending;
893 LASSERTF(filter->fo_tot_pending >= pending,
894 "%s: cli %s/%p tot_pending: "LPU64" grant_used: %lu\n",
895 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
896 filter->fo_tot_pending, pending);
897 filter->fo_tot_pending -= pending;
899 spin_unlock(&exp->exp_obd->obd_osfs_lock);
902 int filter_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
903 int objcount, struct obd_ioobj *obj,
904 struct niobuf_remote *nb, int npages,
905 struct niobuf_local *res, struct obd_trans_info *oti,
908 if (cmd == OBD_BRW_WRITE)
909 return filter_commitrw_write(exp, oa, objcount, obj,
910 nb, npages, res, oti, rc);
911 if (cmd == OBD_BRW_READ)
912 return filter_commitrw_read(exp, oa, objcount, obj,
913 nb, npages, res, oti, rc);
918 int filter_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
919 obd_count oa_bufs, struct brw_page *pga,
920 struct obd_trans_info *oti)
922 struct obd_ioobj ioo;
923 struct niobuf_local *lnb;
924 struct niobuf_remote *rnb;
929 OBD_ALLOC(lnb, oa_bufs * sizeof(struct niobuf_local));
930 OBD_ALLOC(rnb, oa_bufs * sizeof(struct niobuf_remote));
932 if (lnb == NULL || rnb == NULL)
933 GOTO(out, ret = -ENOMEM);
935 for (i = 0; i < oa_bufs; i++) {
936 lnb[i].page = pga[i].pg;
937 rnb[i].offset = pga[i].off;
938 rnb[i].len = pga[i].count;
941 obdo_to_ioobj(oinfo->oi_oa, &ioo);
942 ioo.ioo_bufcnt = oa_bufs;
945 ret = filter_preprw(cmd, exp, oinfo->oi_oa, 1, &ioo,
946 rnb, &npages, lnb, oti, oinfo_capa(oinfo));
949 LASSERTF(oa_bufs == npages, "%u != %u\n", oa_bufs, npages);
951 ret = filter_commitrw(cmd, exp, oinfo->oi_oa, 1, &ioo, rnb,
952 npages, lnb, oti, ret);
956 OBD_FREE(lnb, oa_bufs * sizeof(struct niobuf_local));
958 OBD_FREE(rnb, oa_bufs * sizeof(struct niobuf_remote));