Whamcloud - gitweb
Branch: HEAD
[fs/lustre-release.git] / lustre / obdfilter / filter_io.c
1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2  * vim:expandtab:shiftwidth=8:tabstop=8:
3  *
4  * GPL HEADER START
5  *
6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
7  *
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.
11  *
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).
17  *
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
21  *
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
24  * have any questions.
25  *
26  * GPL HEADER END
27  */
28 /*
29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
30  * Use is subject to license terms.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  *
36  * lustre/obdfilter/filter_io.c
37  *
38  * Author: Peter Braam <braam@clusterfs.com>
39  * Author: Andreas Dilger <adilger@clusterfs.com>
40  * Author: Phil Schwan <phil@clusterfs.com>
41  */
42
43 #define DEBUG_SUBSYSTEM S_FILTER
44
45 #ifndef AUTOCONF_INCLUDED
46 #include <linux/config.h>
47 #endif
48 #include <linux/module.h>
49 #include <linux/pagemap.h> // XXX kill me soon
50 #include <linux/version.h>
51
52 #include <obd_class.h>
53 #include <obd_ost.h>
54 #include <lustre_fsfilt.h>
55 #include "filter_internal.h"
56
57 int *obdfilter_created_scratchpad;
58
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)
63 {
64         struct filter_export_data *fed;
65         struct obd_device *obd = exp->exp_obd;
66         ENTRY;
67
68         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
69
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;
73                 EXIT;
74                 return;
75         }
76
77         fed = &exp->exp_filter_data;
78
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. */
82         CDEBUG(D_CACHE,
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);
86
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)
92                 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);
100                 oa->o_dropped = 0;
101         }
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);
106                 oa->o_dropped = 0;
107         }
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
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;
115
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);
127                         oa->o_grant = 0;
128                 }
129         }
130
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);
136                 LBUG();
137         }
138         EXIT;
139 }
140
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.
144  *
145  * Caller must hold obd_osfs_lock. */
146 obd_size filter_grant_space_left(struct obd_export *exp)
147 {
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;
152
153         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
154
155         if (cfs_time_before_64(obd->obd_osfs_age, cfs_time_current_64() - HZ)) {
156 restat:
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 */
160                         RETURN(0);
161                 statfs_done = 1;
162         }
163
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;
168         } else {
169                 left = 0 /* << blockbits */;
170         }
171
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");
174                 goto restat;
175         }
176
177         if (left >= tot_granted) {
178                 left -= tot_granted;
179         } else {
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);
185                 }
186                 left = 0;
187         }
188
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);
194
195         return left;
196 }
197
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.
200  *
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)
204 {
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;
208         __u64 grant = 0;
209
210         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
211
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.
216          *
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);
228                 grant <<= blockbits;
229
230                 if (grant) {
231                         /* Allow >FILTER_GRANT_CHUNK size when clients
232                          * reconnect due to a server reboot.
233                          */
234                         if ((grant > FILTER_GRANT_CHUNK) &&
235                             (!obd->obd_recovering))
236                                 grant = FILTER_GRANT_CHUNK;
237
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
242                                        "current"LPU64"\n",
243                                        obd->obd_name, exp->exp_client_uuid.uuid,
244                                        exp, fed->fed_grant, want,current_grant);
245                                 spin_unlock(&obd->obd_osfs_lock);
246                                 LBUG();
247                         }
248                 }
249         }
250
251         CDEBUG(D_CACHE,
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);
255         CDEBUG(D_CACHE,
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);
260
261         return grant;
262 }
263
264 /*
265  * the routine is used to request pages from pagecache
266  *
267  * use GFP_NOFS not allowing to enter FS as the client can run on this node
268  * and we might end waiting on a page he sent in the request we're serving.
269  *
270  * use __GFP_HIGHMEM so that the pages can use all of the available memory
271  * on 32-bit machines
272  */
273 static struct page *filter_get_page(struct obd_device *obd,
274                                     struct inode *inode,
275                                     obd_off offset)
276 {
277         struct page *page;
278
279         page = find_or_create_page(inode->i_mapping, offset >> CFS_PAGE_SHIFT,
280                                    GFP_NOFS | __GFP_HIGHMEM);
281         if (unlikely(page == NULL))
282                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_NO_PAGE, 1);
283
284         return page;
285 }
286
287 /*
288  * the routine initializes array of local_niobuf from remote_niobuf
289  */
290 static int filter_map_remote_to_local(int objcount, struct obd_ioobj *obj,
291                                       struct niobuf_remote *nb,
292                                       int *nrpages, struct niobuf_local *res)
293 {
294         struct niobuf_remote *rnb;
295         struct niobuf_local *lnb;
296         int i, max;
297         ENTRY;
298
299         /* we don't support multiobject RPC yet
300          * ost_brw_read() and ost_brw_write() check this */
301         LASSERT(objcount == 1);
302
303         max = *nrpages;
304         *nrpages = 0;
305         for (i = 0, rnb = nb, lnb = res; i < obj->ioo_bufcnt; i++, rnb++) {
306                 obd_off offset = rnb->offset;
307                 unsigned int len = rnb->len;
308
309                 while (len > 0) {
310                         int poff = offset & (CFS_PAGE_SIZE - 1);
311                         int plen = CFS_PAGE_SIZE - poff;
312
313                         if (*nrpages >= max) {
314                                 CERROR("small array of local bufs: %d\n", max);
315                                 RETURN(-EINVAL);
316                         }
317
318                         if (plen > len)
319                                 plen = len;
320                         lnb->offset = offset;
321                         lnb->len = plen;
322                         lnb->flags = rnb->flags;
323                         lnb->page = NULL;
324                         lnb->rc = 0;
325                         lnb->lnb_grant_used = 0;
326
327                         LASSERTF(plen <= len, "plen %u, len %u\n", plen, len);
328                         offset += plen;
329                         len -= plen;
330                         lnb++;
331                         (*nrpages)++;
332                 }
333         }
334         RETURN(0);
335 }
336
337 /*
338  * the function is used to free all pages used for request
339  * just to mimic cacheless OSS which don't occupy much memory
340  */
341 void filter_invalidate_cache(struct obd_device *obd, struct obd_ioobj *obj,
342                              struct niobuf_remote *nb, struct inode *inode)
343 {
344         struct niobuf_remote *rnb;
345         int i;
346
347         LASSERT(inode != NULL);
348
349         for (i = 0, rnb = nb; i < obj->ioo_bufcnt; i++, rnb++) {
350                 obd_off start;
351                 obd_off end;
352
353                 start = rnb->offset >> CFS_PAGE_SHIFT;
354                 end = (rnb->offset + rnb->len) >> CFS_PAGE_SHIFT;
355                 invalidate_mapping_pages(inode->i_mapping, start, end);
356                 /* just to avoid warnings */
357                 start = 0;
358                 end = 0;
359         }
360 }
361
362 static int filter_preprw_read(int cmd, struct obd_export *exp, struct obdo *oa,
363                               int objcount, struct obd_ioobj *obj,
364                               struct niobuf_remote *nb,
365                               int *npages, struct niobuf_local *res,
366                               struct obd_trans_info *oti,
367                               struct lustre_capa *capa)
368 {
369         struct obd_device *obd = exp->exp_obd;
370         struct timeval start, end;
371         struct lvfs_run_ctxt saved;
372         struct niobuf_local *lnb;
373         struct dentry *dentry = NULL;
374         struct inode *inode = NULL;
375         void *iobuf = NULL;
376         int rc = 0, i, tot_bytes = 0;
377         unsigned long now = jiffies;
378         long timediff;
379         ENTRY;
380
381         /* We are currently not supporting multi-obj BRW_READ RPCS at all.
382          * When we do this function's dentry cleanup will need to be fixed.
383          * These values are verified in ost_brw_write() from the wire. */
384         LASSERTF(objcount == 1, "%d\n", objcount);
385         LASSERTF(obj->ioo_bufcnt > 0, "%d\n", obj->ioo_bufcnt);
386
387         rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa), capa,
388                               CAPA_OPC_OSS_READ);
389         if (rc)
390                 RETURN(rc);
391
392         if (oa && oa->o_valid & OBD_MD_FLGRANT) {
393                 spin_lock(&obd->obd_osfs_lock);
394                 filter_grant_incoming(exp, oa);
395
396                 if (!(oa->o_flags & OBD_FL_SHRINK_GRANT))
397                         oa->o_grant = 0;
398                 spin_unlock(&obd->obd_osfs_lock);
399         }
400
401         iobuf = filter_iobuf_get(&obd->u.filter, oti);
402         if (IS_ERR(iobuf))
403                 RETURN(PTR_ERR(iobuf));
404
405         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
406         dentry = filter_oa2dentry(obd, oa);
407         if (IS_ERR(dentry)) {
408                 rc = PTR_ERR(dentry);
409                 dentry = NULL;
410                 GOTO(cleanup, rc);
411         }
412
413         inode = dentry->d_inode;
414
415         obdo_to_inode(inode, oa, OBD_MD_FLATIME);
416
417         rc = filter_map_remote_to_local(objcount, obj, nb, npages, res);
418         if (rc)
419                 GOTO(cleanup, rc);
420
421         fsfilt_check_slow(obd, now, "preprw_read setup");
422
423         /* find pages for all segments, fill array with them */
424         do_gettimeofday(&start);
425         for (i = 0, lnb = res; i < *npages; i++, lnb++) {
426
427                 lnb->dentry = dentry;
428
429                 if (i_size_read(inode) <= lnb->offset)
430                         /* If there's no more data, abort early.  lnb->rc == 0,
431                          * so it's easy to detect later. */
432                         break;
433
434                 lnb->page = filter_get_page(obd, inode, lnb->offset);
435                 if (lnb->page == NULL)
436                         GOTO(cleanup, rc = -ENOMEM);
437
438                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS, 1);
439
440                 if (i_size_read(inode) < lnb->offset + lnb->len - 1)
441                         lnb->rc = i_size_read(inode) - lnb->offset;
442                 else
443                         lnb->rc = lnb->len;
444
445                 tot_bytes += lnb->rc;
446
447                 if (PageUptodate(lnb->page)) {
448                         lprocfs_counter_add(obd->obd_stats,
449                                             LPROC_FILTER_CACHE_HIT, 1);
450                         continue;
451                 }
452
453                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_MISS, 1);
454                 filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
455         }
456         do_gettimeofday(&end);
457         timediff = cfs_timeval_sub(&end, &start, NULL);
458         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
459
460         if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
461                 GOTO(cleanup, rc = -ENOMEM);
462
463         fsfilt_check_slow(obd, now, "start_page_read");
464
465         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf,
466                               exp, NULL, NULL, NULL);
467         if (rc)
468                 GOTO(cleanup, rc);
469
470         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_READ_BYTES, tot_bytes);
471
472         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
473                 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
474                                     LPROC_FILTER_READ_BYTES, tot_bytes);
475
476         EXIT;
477
478  cleanup:
479         /* unlock pages to allow access from concurrent OST_READ */
480         for (i = 0, lnb = res; i < *npages; i++, lnb++) {
481                 if (lnb->page) {
482                         LASSERT(PageLocked(lnb->page));
483                         unlock_page(lnb->page);
484
485                         if (rc) {
486                                 page_cache_release(lnb->page);
487                                 lnb->page = NULL;
488                         }
489                 }
490         }
491
492         if (rc != 0) {
493                 if (dentry != NULL)
494                         f_dput(dentry);
495         }
496
497         filter_iobuf_put(&obd->u.filter, iobuf, oti);
498
499         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
500         if (rc)
501                 CERROR("io error %d\n", rc);
502
503         return rc;
504 }
505
506 /* When clients have dirtied as much space as they've been granted they
507  * fall through to sync writes.  These sync writes haven't been expressed
508  * in grants and need to error with ENOSPC when there isn't room in the
509  * filesystem for them after grants are taken into account.  However,
510  * writeback of the dirty data that was already granted space can write
511  * right on through.
512  *
513  * Caller must hold obd_osfs_lock. */
514 static int filter_grant_check(struct obd_export *exp, struct obdo *oa,
515                               int objcount, struct fsfilt_objinfo *fso,
516                               int niocount, struct niobuf_local *lnb,
517                               obd_size *left, struct inode *inode)
518 {
519         struct filter_export_data *fed = &exp->exp_filter_data;
520         int blocksize = exp->exp_obd->u.obt.obt_sb->s_blocksize;
521         unsigned long used = 0, ungranted = 0, using;
522         int i, rc = -ENOSPC, obj, n = 0;
523
524         LASSERT_SPIN_LOCKED(&exp->exp_obd->obd_osfs_lock);
525
526         for (obj = 0; obj < objcount; obj++) {
527                 for (i = 0; i < fso[obj].fso_bufcnt; i++, n++) {
528                         int tmp, bytes;
529
530                         /* should match the code in osc_exit_cache */
531                         bytes = lnb[n].len;
532                         bytes += lnb[n].offset & (blocksize - 1);
533                         tmp = (lnb[n].offset + lnb[n].len) & (blocksize - 1);
534                         if (tmp)
535                                 bytes += blocksize - tmp;
536
537                         if ((lnb[n].flags & OBD_BRW_FROM_GRANT) &&
538                             (oa->o_valid & OBD_MD_FLGRANT)) {
539                                 if (fed->fed_grant < used + bytes) {
540                                         CDEBUG(D_CACHE,
541                                                "%s: cli %s/%p claims %ld+%d "
542                                                "GRANT, real grant %lu idx %d\n",
543                                                exp->exp_obd->obd_name,
544                                                exp->exp_client_uuid.uuid, exp,
545                                                used, bytes, fed->fed_grant, n);
546                                 } else {
547                                         used += bytes;
548                                         lnb[n].flags |= OBD_BRW_GRANTED;
549                                         lnb[n].lnb_grant_used = bytes;
550                                         CDEBUG(0, "idx %d used=%lu\n", n, used);
551                                         rc = 0;
552                                         continue;
553                                 }
554                         }
555                         if (*left > ungranted + bytes) {
556                                 /* if enough space, pretend it was granted */
557                                 ungranted += bytes;
558                                 lnb[n].flags |= OBD_BRW_GRANTED;
559                                 lnb[n].lnb_grant_used = bytes;
560                                 CDEBUG(0, "idx %d ungranted=%lu\n",n,ungranted);
561                                 rc = 0;
562                                 continue;
563                         }
564
565                         /* We can't check for already-mapped blocks here, as
566                          * it requires dropping the osfs lock to do the bmap.
567                          * Instead, we return ENOSPC and in that case we need
568                          * to go through and verify if all of the blocks not
569                          * marked BRW_GRANTED are already mapped and we can
570                          * ignore this error. */
571                         lnb[n].rc = -ENOSPC;
572                         lnb[n].flags &= ~OBD_BRW_GRANTED;
573                         CDEBUG(D_CACHE,"%s: cli %s/%p idx %d no space for %d\n",
574                                exp->exp_obd->obd_name,
575                                exp->exp_client_uuid.uuid, exp, n, bytes);
576                 }
577         }
578
579         /* Now substract what client have used already.  We don't subtract
580          * this from the tot_granted yet, so that other client's can't grab
581          * that space before we have actually allocated our blocks.  That
582          * happens in filter_grant_commit() after the writes are done. */
583         *left -= ungranted;
584         fed->fed_grant -= used;
585         fed->fed_pending += used + ungranted;
586         exp->exp_obd->u.filter.fo_tot_granted += ungranted;
587         exp->exp_obd->u.filter.fo_tot_pending += used + ungranted;
588
589         CDEBUG(D_CACHE,
590                "%s: cli %s/%p used: %lu ungranted: %lu grant: %lu dirty: %lu\n",
591                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp, used,
592                ungranted, fed->fed_grant, fed->fed_dirty);
593
594         /* Rough calc in case we don't refresh cached statfs data */
595         using = (used + ungranted + 1 ) >>
596                 exp->exp_obd->u.obt.obt_sb->s_blocksize_bits;
597         if (exp->exp_obd->obd_osfs.os_bavail > using)
598                 exp->exp_obd->obd_osfs.os_bavail -= using;
599         else
600                 exp->exp_obd->obd_osfs.os_bavail = 0;
601
602         if (fed->fed_dirty < used) {
603                 CERROR("%s: cli %s/%p claims used %lu > fed_dirty %lu\n",
604                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
605                        used, fed->fed_dirty);
606                 used = fed->fed_dirty;
607         }
608         exp->exp_obd->u.filter.fo_tot_dirty -= used;
609         fed->fed_dirty -= used;
610
611         if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
612                 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
613                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
614                        fed->fed_dirty, fed->fed_pending, fed->fed_grant);
615                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
616                 LBUG();
617         }
618         return rc;
619 }
620
621 /* If we ever start to support multi-object BRW RPCs, we will need to get locks
622  * on mulitple inodes.  That isn't all, because there still exists the
623  * possibility of a truncate starting a new transaction while holding the ext3
624  * rwsem = write while some writes (which have started their transactions here)
625  * blocking on the ext3 rwsem = read => lock inversion.
626  *
627  * The handling gets very ugly when dealing with locked pages.  It may be easier
628  * to just get rid of the locked page code (which has problems of its own) and
629  * either discover we do not need it anymore (i.e. it was a symptom of another
630  * bug) or ensure we get the page locks in an appropriate order. */
631 static int filter_preprw_write(int cmd, struct obd_export *exp, struct obdo *oa,
632                                int objcount, struct obd_ioobj *obj,
633                                struct niobuf_remote *nb, int *npages,
634                                struct niobuf_local *res,
635                                struct obd_trans_info *oti,
636                                struct lustre_capa *capa)
637 {
638         struct obd_device *obd = exp->exp_obd;
639         struct timeval start, end;
640         struct lvfs_run_ctxt saved;
641         struct niobuf_local *lnb = res;
642         struct fsfilt_objinfo fso;
643         struct filter_mod_data *fmd;
644         struct dentry *dentry = NULL;
645         void *iobuf;
646         obd_size left;
647         unsigned long now = jiffies, timediff;
648         int rc = 0, i, tot_bytes = 0, cleanup_phase = 0;
649         ENTRY;
650         LASSERT(objcount == 1);
651         LASSERT(obj->ioo_bufcnt > 0);
652
653         rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa), capa,
654                               CAPA_OPC_OSS_WRITE);
655         if (rc)
656                 RETURN(rc);
657
658         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
659         iobuf = filter_iobuf_get(&obd->u.filter, oti);
660         if (IS_ERR(iobuf))
661                 GOTO(cleanup, rc = PTR_ERR(iobuf));
662         cleanup_phase = 1;
663
664         dentry = filter_fid2dentry(obd, NULL, obj->ioo_gr,
665                                    obj->ioo_id);
666         if (IS_ERR(dentry))
667                 GOTO(cleanup, rc = PTR_ERR(dentry));
668         cleanup_phase = 2;
669
670         if (dentry->d_inode == NULL) {
671                 CERROR("%s: trying to BRW to non-existent file "LPU64"\n",
672                        obd->obd_name, obj->ioo_id);
673                 GOTO(cleanup, rc = -ENOENT);
674         }
675
676         if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID) &&
677             dentry->d_inode->i_mode & (S_ISUID | S_ISGID)) {
678                 rc = filter_capa_fixoa(exp, oa, obdo_mdsno(oa), capa);
679                 if (rc)
680                         GOTO(cleanup, rc);
681         }
682
683         rc = filter_map_remote_to_local(objcount, obj, nb, npages, res);
684         if (rc)
685                 GOTO(cleanup, rc);
686
687         fsfilt_check_slow(obd, now, "preprw_write setup");
688
689         /* Don't update inode timestamps if this write is older than a
690          * setattr which modifies the timestamps. b=10150 */
691         /* XXX when we start having persistent reservations this needs to
692          * be changed to filter_fmd_get() to create the fmd if it doesn't
693          * already exist so we can store the reservation handle there. */
694         fmd = filter_fmd_find(exp, obj->ioo_id, obj->ioo_gr);
695
696         LASSERT(oa != NULL);
697         spin_lock(&obd->obd_osfs_lock);
698         filter_grant_incoming(exp, oa);
699         if (fmd && fmd->fmd_mactime_xid > oti->oti_xid)
700                 oa->o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLCTIME |
701                                  OBD_MD_FLATIME);
702         else
703                 obdo_to_inode(dentry->d_inode, oa, OBD_MD_FLATIME |
704                               OBD_MD_FLMTIME | OBD_MD_FLCTIME);
705         cleanup_phase = 3;
706
707         left = filter_grant_space_left(exp);
708
709         fso.fso_dentry = dentry;
710         fso.fso_bufcnt = *npages;
711
712         rc = filter_grant_check(exp, oa, objcount, &fso, *npages, res,
713                                 &left, dentry->d_inode);
714
715         /* do not zero out oa->o_valid as it is used in filter_commitrw_write()
716          * for setting UID/GID and fid EA in first write time. */
717         if (oa->o_valid & OBD_MD_FLGRANT)
718                 oa->o_grant = filter_grant(exp,oa->o_grant,oa->o_undirty,left);
719
720         spin_unlock(&obd->obd_osfs_lock);
721         filter_fmd_put(exp, fmd);
722
723         if (rc)
724                 GOTO(cleanup, rc);
725
726         do_gettimeofday(&start);
727         for (i = 0, lnb = res; i < *npages; i++, lnb++) {
728
729                 /* We still set up for ungranted pages so that granted pages
730                  * can be written to disk as they were promised, and portals
731                  * needs to keep the pages all aligned properly. */
732                 lnb->dentry = dentry;
733
734                 lnb->page = filter_get_page(obd, dentry->d_inode, lnb->offset);
735                 if (lnb->page == NULL)
736                         GOTO(cleanup, rc = -ENOMEM);
737                 cleanup_phase = 4;
738
739                 /* DLM locking protects us from write and truncate competing
740                  * for same region, but truncate can leave dirty page in the
741                  * cache. it's possible the writeout on a such a page is in
742                  * progress when we access it. it's also possible that during
743                  * this writeout we put new (partial) data, but then won't
744                  * be able to proceed in filter_commitrw_write(). thus let's
745                  * just wait for writeout completion, should be rare enough.
746                  * -bzzz */
747                 if (obd->u.filter.fo_writethrough_cache)
748                         wait_on_page_writeback(lnb->page);
749                 BUG_ON(PageWriteback(lnb->page));
750
751                 /* If the filter writes a partial page, then has the file
752                  * extended, the client will read in the whole page.  the
753                  * filter has to be careful to zero the rest of the partial
754                  * page on disk.  we do it by hand for partial extending
755                  * writes, send_bio() is responsible for zeroing pages when
756                  * asked to read unmapped blocks -- brw_kiovec() does this. */
757                 if (lnb->len != CFS_PAGE_SIZE) {
758                         __s64 maxidx;
759
760                         maxidx = ((i_size_read(dentry->d_inode) +
761                                    CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT) - 1;
762                         if (maxidx >= lnb->page->index) {
763                                 LL_CDEBUG_PAGE(D_PAGE, lnb->page, "write %u @ "
764                                                LPU64" flg %x before EOF %llu\n",
765                                                lnb->len, lnb->offset,lnb->flags,
766                                                i_size_read(dentry->d_inode));
767                                 filter_iobuf_add_page(obd, iobuf,
768                                                       dentry->d_inode,
769                                                       lnb->page);
770                         } else {
771                                 long off;
772                                 char *p = kmap(lnb->page);
773
774                                 off = lnb->offset & ~CFS_PAGE_MASK;
775                                 if (off)
776                                         memset(p, 0, off);
777                                 off = (lnb->offset + lnb->len) & ~CFS_PAGE_MASK;
778                                 if (off)
779                                         memset(p + off, 0, CFS_PAGE_SIZE - off);
780                                 kunmap(lnb->page);
781                         }
782                 }
783                 if (lnb->rc == 0)
784                         tot_bytes += lnb->len;
785         }
786         do_gettimeofday(&end);
787         timediff = cfs_timeval_sub(&end, &start, NULL);
788         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
789
790         if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
791                 GOTO(cleanup, rc = -ENOMEM);
792
793         /* don't unlock pages to prevent any access */
794         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf, exp,
795                               NULL, NULL, NULL);
796
797         fsfilt_check_slow(obd, now, "start_page_write");
798
799         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
800                 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
801                                     LPROC_FILTER_WRITE_BYTES, tot_bytes);
802         EXIT;
803 cleanup:
804         switch(cleanup_phase) {
805         case 4:
806                 if (rc) {
807                         for (i = 0, lnb = res; i < *npages; i++, lnb++) {
808                                 if (lnb->page != NULL) {
809                                         unlock_page(lnb->page);
810                                         page_cache_release(lnb->page);
811                                         lnb->page = NULL;
812                                 }
813                         }
814                 }
815         case 3:
816                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
817         case 2:
818                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
819                 if (rc)
820                         f_dput(dentry);
821                 break;
822         case 1:
823                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
824         case 0:
825                 spin_lock(&obd->obd_osfs_lock);
826                 if (oa)
827                         filter_grant_incoming(exp, oa);
828                 spin_unlock(&obd->obd_osfs_lock);
829                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
830                 break;
831         default:;
832         }
833         return rc;
834 }
835
836 int filter_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
837                   int objcount, struct obd_ioobj *obj,
838                   struct niobuf_remote *nb, int *npages,
839                   struct niobuf_local *res, struct obd_trans_info *oti,
840                   struct lustre_capa *capa)
841 {
842         if (cmd == OBD_BRW_WRITE)
843                 return filter_preprw_write(cmd, exp, oa, objcount, obj,
844                                            nb, npages, res, oti, capa);
845         if (cmd == OBD_BRW_READ)
846                 return filter_preprw_read(cmd, exp, oa, objcount, obj,
847                                           nb, npages, res, oti, capa);
848         LBUG();
849         return -EPROTO;
850 }
851
852 static int filter_commitrw_read(struct obd_export *exp, struct obdo *oa,
853                                 int objcount, struct obd_ioobj *obj,
854                                 struct niobuf_remote *rnb,
855                                 int npages, struct niobuf_local *res,
856                                 struct obd_trans_info *oti, int rc)
857 {
858         struct filter_obd *fo = &exp->exp_obd->u.filter;
859         struct inode *inode = NULL;
860         struct ldlm_res_id res_id;
861         struct ldlm_resource *resource = NULL;
862         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
863         struct niobuf_local *lnb;
864         int i;
865         ENTRY;
866
867         osc_build_res_name(obj->ioo_id, obj->ioo_gr, &res_id);
868         /* If oa != NULL then filter_preprw_read updated the inode atime
869          * and we should update the lvb so that other glimpses will also
870          * get the updated value. bug 5972 */
871         if (oa && ns && ns->ns_lvbo && ns->ns_lvbo->lvbo_update) {
872                 resource = ldlm_resource_get(ns, NULL, &res_id, LDLM_EXTENT, 0);
873
874                 if (resource != NULL) {
875                         LDLM_RESOURCE_ADDREF(resource);
876                         ns->ns_lvbo->lvbo_update(resource, NULL, 0, 1);
877                         LDLM_RESOURCE_DELREF(resource);
878                         ldlm_resource_putref(resource);
879                 }
880         }
881
882         if (res->dentry != NULL)
883                 inode = res->dentry->d_inode;
884
885         for (i = 0, lnb = res; i < npages; i++, lnb++) {
886                 if (lnb->page != NULL) {
887                         page_cache_release(lnb->page);
888                         lnb->page = NULL;
889                 }
890         }
891
892         if (inode && (fo->fo_read_cache == 0 ||
893                         i_size_read(inode) > fo->fo_readcache_max_filesize))
894                 filter_invalidate_cache(exp->exp_obd, obj, rnb, inode);
895
896         if (res->dentry != NULL)
897                 f_dput(res->dentry);
898         RETURN(rc);
899 }
900
901 void filter_grant_commit(struct obd_export *exp, int niocount,
902                          struct niobuf_local *res)
903 {
904         struct filter_obd *filter = &exp->exp_obd->u.filter;
905         struct niobuf_local *lnb = res;
906         unsigned long pending = 0;
907         int i;
908
909         spin_lock(&exp->exp_obd->obd_osfs_lock);
910         for (i = 0, lnb = res; i < niocount; i++, lnb++)
911                 pending += lnb->lnb_grant_used;
912
913         LASSERTF(exp->exp_filter_data.fed_pending >= pending,
914                  "%s: cli %s/%p fed_pending: %lu grant_used: %lu\n",
915                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
916                  exp->exp_filter_data.fed_pending, pending);
917         exp->exp_filter_data.fed_pending -= pending;
918         LASSERTF(filter->fo_tot_granted >= pending,
919                  "%s: cli %s/%p tot_granted: "LPU64" grant_used: %lu\n",
920                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
921                  exp->exp_obd->u.filter.fo_tot_granted, pending);
922         filter->fo_tot_granted -= pending;
923         LASSERTF(filter->fo_tot_pending >= pending,
924                  "%s: cli %s/%p tot_pending: "LPU64" grant_used: %lu\n",
925                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
926                  filter->fo_tot_pending, pending);
927         filter->fo_tot_pending -= pending;
928
929         spin_unlock(&exp->exp_obd->obd_osfs_lock);
930 }
931
932 int filter_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
933                     int objcount, struct obd_ioobj *obj,
934                     struct niobuf_remote *nb, int npages,
935                     struct niobuf_local *res, struct obd_trans_info *oti,
936                     int rc)
937 {
938         if (cmd == OBD_BRW_WRITE)
939                 return filter_commitrw_write(exp, oa, objcount, obj,
940                                              nb, npages, res, oti, rc);
941         if (cmd == OBD_BRW_READ)
942                 return filter_commitrw_read(exp, oa, objcount, obj,
943                                             nb, npages, res, oti, rc);
944         LBUG();
945         return -EPROTO;
946 }
947
948 int filter_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
949                obd_count oa_bufs, struct brw_page *pga,
950                struct obd_trans_info *oti)
951 {
952         struct obd_ioobj ioo;
953         struct niobuf_local *lnb;
954         struct niobuf_remote *rnb;
955         obd_count i;
956         int ret = 0, npages;
957         ENTRY;
958
959         OBD_ALLOC(lnb, oa_bufs * sizeof(struct niobuf_local));
960         OBD_ALLOC(rnb, oa_bufs * sizeof(struct niobuf_remote));
961
962         if (lnb == NULL || rnb == NULL)
963                 GOTO(out, ret = -ENOMEM);
964
965         for (i = 0; i < oa_bufs; i++) {
966                 lnb[i].page = pga[i].pg;
967                 rnb[i].offset = pga[i].off;
968                 rnb[i].len = pga[i].count;
969         }
970
971         obdo_to_ioobj(oinfo->oi_oa, &ioo);
972         ioo.ioo_bufcnt = oa_bufs;
973
974         npages = oa_bufs;
975         ret = filter_preprw(cmd, exp, oinfo->oi_oa, 1, &ioo,
976                             rnb, &npages, lnb, oti, oinfo_capa(oinfo));
977         if (ret != 0)
978                 GOTO(out, ret);
979         LASSERTF(oa_bufs == npages, "%u != %u\n", oa_bufs, npages);
980
981         ret = filter_commitrw(cmd, exp, oinfo->oi_oa, 1, &ioo, rnb,
982                               npages, lnb, oti, ret);
983
984 out:
985         if (lnb)
986                 OBD_FREE(lnb, oa_bufs * sizeof(struct niobuf_local));
987         if (rnb)
988                 OBD_FREE(rnb, oa_bufs * sizeof(struct niobuf_remote));
989         RETURN(ret);
990 }