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 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
270  * on 32-bit machines
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.
274  *
275  * See Bug 19529 and Bug 19917 for details. 
276  */
277 static struct page *filter_get_page(struct obd_device *obd,
278                                     struct inode *inode,
279                                     obd_off offset,
280                                     int localreq)
281 {
282         struct page *page;
283
284         page = find_or_create_page(inode->i_mapping, offset >> CFS_PAGE_SHIFT,
285                                    (localreq ? (GFP_NOFS | __GFP_HIGHMEM) 
286                                              : (GFP_HIGHUSER | __GFP_NOMEMALLOC)));
287         if (unlikely(page == NULL))
288                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_NO_PAGE, 1);
289
290         return page;
291 }
292
293 /*
294  * the routine initializes array of local_niobuf from remote_niobuf
295  */
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)
299 {
300         struct niobuf_remote *rnb;
301         struct niobuf_local *lnb;
302         int i, max;
303         ENTRY;
304
305         /* we don't support multiobject RPC yet
306          * ost_brw_read() and ost_brw_write() check this */
307         LASSERT(objcount == 1);
308
309         max = *nrpages;
310         *nrpages = 0;
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;
314
315                 while (len > 0) {
316                         int poff = offset & (CFS_PAGE_SIZE - 1);
317                         int plen = CFS_PAGE_SIZE - poff;
318
319                         if (*nrpages >= max) {
320                                 CERROR("small array of local bufs: %d\n", max);
321                                 RETURN(-EINVAL);
322                         }
323
324                         if (plen > len)
325                                 plen = len;
326                         lnb->offset = offset;
327                         lnb->len = plen;
328                         lnb->flags = rnb->flags;
329                         lnb->page = NULL;
330                         lnb->rc = 0;
331                         lnb->lnb_grant_used = 0;
332
333                         LASSERTF(plen <= len, "plen %u, len %u\n", plen, len);
334                         offset += plen;
335                         len -= plen;
336                         lnb++;
337                         (*nrpages)++;
338                 }
339         }
340         RETURN(0);
341 }
342
343 /*
344  * the function is used to free all pages used for request
345  * just to mimic cacheless OSS which don't occupy much memory
346  */
347 void filter_invalidate_cache(struct obd_device *obd, struct obd_ioobj *obj,
348                              struct niobuf_remote *nb, struct inode *inode)
349 {
350         struct niobuf_remote *rnb;
351         int i;
352
353         LASSERT(inode != NULL);
354
355         for (i = 0, rnb = nb; i < obj->ioo_bufcnt; i++, rnb++) {
356                 obd_off start;
357                 obd_off end;
358
359                 start = rnb->offset >> CFS_PAGE_SHIFT;
360                 end = (rnb->offset + rnb->len) >> CFS_PAGE_SHIFT;
361                 invalidate_mapping_pages(inode->i_mapping, start, end);
362                 /* just to avoid warnings */
363                 start = 0;
364                 end = 0;
365         }
366 }
367
368 /*
369  * the invalidate above doesn't work during read because lnet pins pages.
370  * The truncate is used here instead to drop pages from cache
371  */
372 void filter_truncate_cache(struct obd_device *obd, struct obd_ioobj *obj,
373                            struct niobuf_remote *nb, int pages,
374                            struct niobuf_local *res, struct inode *inode)
375 {
376         struct niobuf_remote *rnb;
377         int i;
378
379         LASSERT(inode != NULL);
380 #ifdef HAVE_TRUNCATE_RANGE
381         for (i = 0, rnb = nb; i < obj->ioo_bufcnt; i++, rnb++) {
382                 /* remove pages in which range is fit */
383                 truncate_inode_pages_range(inode->i_mapping,
384                                            rnb->offset & CFS_PAGE_MASK,
385                                            (rnb->offset + rnb->len - 1) |
386                                            ~CFS_PAGE_MASK);
387         }
388 #elif (defined HAVE_TRUNCATE_COMPLETE)
389         for (i = 0, lnb = res; i < pages; i++, lnb++)
390                 truncate_complete_page(inode->i_mapping, lnb->page);
391 #else
392 #error "Nor truncate_inode_pages_range or truncate_complete_page are supported"
393 #endif
394 }
395
396 static int filter_preprw_read(int cmd, struct obd_export *exp, struct obdo *oa,
397                               int objcount, struct obd_ioobj *obj,
398                               struct niobuf_remote *nb,
399                               int *npages, struct niobuf_local *res,
400                               struct obd_trans_info *oti,
401                               struct lustre_capa *capa)
402 {
403         struct obd_device *obd = exp->exp_obd;
404         struct timeval start, end;
405         struct lvfs_run_ctxt saved;
406         struct niobuf_local *lnb;
407         struct dentry *dentry = NULL;
408         struct inode *inode = NULL;
409         void *iobuf = NULL;
410         int rc = 0, i, tot_bytes = 0;
411         unsigned long now = jiffies;
412         long timediff;
413         ENTRY;
414
415         /* We are currently not supporting multi-obj BRW_READ RPCS at all.
416          * When we do this function's dentry cleanup will need to be fixed.
417          * These values are verified in ost_brw_write() from the wire. */
418         LASSERTF(objcount == 1, "%d\n", objcount);
419         LASSERTF(obj->ioo_bufcnt > 0, "%d\n", obj->ioo_bufcnt);
420
421         rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa), capa,
422                               CAPA_OPC_OSS_READ);
423         if (rc)
424                 RETURN(rc);
425
426         if (oa && oa->o_valid & OBD_MD_FLGRANT) {
427                 spin_lock(&obd->obd_osfs_lock);
428                 filter_grant_incoming(exp, oa);
429
430                 if (!(oa->o_flags & OBD_FL_SHRINK_GRANT))
431                         oa->o_grant = 0;
432                 spin_unlock(&obd->obd_osfs_lock);
433         }
434
435         iobuf = filter_iobuf_get(&obd->u.filter, oti);
436         if (IS_ERR(iobuf))
437                 RETURN(PTR_ERR(iobuf));
438
439         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
440         dentry = filter_oa2dentry(obd, oa);
441         if (IS_ERR(dentry)) {
442                 rc = PTR_ERR(dentry);
443                 dentry = NULL;
444                 GOTO(cleanup, rc);
445         }
446
447         inode = dentry->d_inode;
448
449         obdo_to_inode(inode, oa, OBD_MD_FLATIME);
450
451         rc = filter_map_remote_to_local(objcount, obj, nb, npages, res);
452         if (rc)
453                 GOTO(cleanup, rc);
454
455         fsfilt_check_slow(obd, now, "preprw_read setup");
456
457         /* find pages for all segments, fill array with them */
458         do_gettimeofday(&start);
459         for (i = 0, lnb = res; i < *npages; i++, lnb++) {
460
461                 lnb->dentry = dentry;
462
463                 if (i_size_read(inode) <= lnb->offset)
464                         /* If there's no more data, abort early.  lnb->rc == 0,
465                          * so it's easy to detect later. */
466                         break;
467
468                 lnb->page = filter_get_page(obd, inode, lnb->offset, 0);
469                 if (lnb->page == NULL)
470                         GOTO(cleanup, rc = -ENOMEM);
471
472                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS, 1);
473
474                 if (i_size_read(inode) < lnb->offset + lnb->len - 1)
475                         lnb->rc = i_size_read(inode) - lnb->offset;
476                 else
477                         lnb->rc = lnb->len;
478
479                 tot_bytes += lnb->rc;
480
481                 if (PageUptodate(lnb->page)) {
482                         lprocfs_counter_add(obd->obd_stats,
483                                             LPROC_FILTER_CACHE_HIT, 1);
484                         continue;
485                 }
486
487                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_MISS, 1);
488                 filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
489         }
490         do_gettimeofday(&end);
491         timediff = cfs_timeval_sub(&end, &start, NULL);
492         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
493
494         if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
495                 GOTO(cleanup, rc = -ENOMEM);
496
497         fsfilt_check_slow(obd, now, "start_page_read");
498
499         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf,
500                               exp, NULL, NULL, NULL);
501         if (rc)
502                 GOTO(cleanup, rc);
503
504         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_READ_BYTES, tot_bytes);
505
506         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
507                 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
508                                     LPROC_FILTER_READ_BYTES, tot_bytes);
509
510         EXIT;
511
512  cleanup:
513         /* unlock pages to allow access from concurrent OST_READ */
514         for (i = 0, lnb = res; i < *npages; i++, lnb++) {
515                 if (lnb->page) {
516                         LASSERT(PageLocked(lnb->page));
517                         unlock_page(lnb->page);
518
519                         if (rc) {
520                                 page_cache_release(lnb->page);
521                                 lnb->page = NULL;
522                         }
523                 }
524         }
525
526         if (rc != 0) {
527                 if (dentry != NULL)
528                         f_dput(dentry);
529         }
530
531         filter_iobuf_put(&obd->u.filter, iobuf, oti);
532
533         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
534         if (rc)
535                 CERROR("io error %d\n", rc);
536
537         return rc;
538 }
539
540 /* When clients have dirtied as much space as they've been granted they
541  * fall through to sync writes.  These sync writes haven't been expressed
542  * in grants and need to error with ENOSPC when there isn't room in the
543  * filesystem for them after grants are taken into account.  However,
544  * writeback of the dirty data that was already granted space can write
545  * right on through.
546  *
547  * Caller must hold obd_osfs_lock. */
548 static int filter_grant_check(struct obd_export *exp, struct obdo *oa,
549                               int objcount, struct fsfilt_objinfo *fso,
550                               int niocount, struct niobuf_local *lnb,
551                               obd_size *left, struct inode *inode)
552 {
553         struct filter_export_data *fed = &exp->exp_filter_data;
554         int blocksize = exp->exp_obd->u.obt.obt_sb->s_blocksize;
555         unsigned long used = 0, ungranted = 0, using;
556         int i, rc = -ENOSPC, obj, n = 0;
557
558         LASSERT_SPIN_LOCKED(&exp->exp_obd->obd_osfs_lock);
559
560         for (obj = 0; obj < objcount; obj++) {
561                 for (i = 0; i < fso[obj].fso_bufcnt; i++, n++) {
562                         int tmp, bytes;
563
564                         /* should match the code in osc_exit_cache */
565                         bytes = lnb[n].len;
566                         bytes += lnb[n].offset & (blocksize - 1);
567                         tmp = (lnb[n].offset + lnb[n].len) & (blocksize - 1);
568                         if (tmp)
569                                 bytes += blocksize - tmp;
570
571                         if ((lnb[n].flags & OBD_BRW_FROM_GRANT) &&
572                             (oa->o_valid & OBD_MD_FLGRANT)) {
573                                 if (fed->fed_grant < used + bytes) {
574                                         CDEBUG(D_CACHE,
575                                                "%s: cli %s/%p claims %ld+%d "
576                                                "GRANT, real grant %lu idx %d\n",
577                                                exp->exp_obd->obd_name,
578                                                exp->exp_client_uuid.uuid, exp,
579                                                used, bytes, fed->fed_grant, n);
580                                 } else {
581                                         used += bytes;
582                                         lnb[n].flags |= OBD_BRW_GRANTED;
583                                         lnb[n].lnb_grant_used = bytes;
584                                         CDEBUG(0, "idx %d used=%lu\n", n, used);
585                                         rc = 0;
586                                         continue;
587                                 }
588                         }
589                         if (*left > ungranted + bytes) {
590                                 /* if enough space, pretend it was granted */
591                                 ungranted += bytes;
592                                 lnb[n].flags |= OBD_BRW_GRANTED;
593                                 lnb[n].lnb_grant_used = bytes;
594                                 CDEBUG(0, "idx %d ungranted=%lu\n",n,ungranted);
595                                 rc = 0;
596                                 continue;
597                         }
598
599                         /* We can't check for already-mapped blocks here, as
600                          * it requires dropping the osfs lock to do the bmap.
601                          * Instead, we return ENOSPC and in that case we need
602                          * to go through and verify if all of the blocks not
603                          * marked BRW_GRANTED are already mapped and we can
604                          * ignore this error. */
605                         lnb[n].rc = -ENOSPC;
606                         lnb[n].flags &= ~OBD_BRW_GRANTED;
607                         CDEBUG(D_CACHE,"%s: cli %s/%p idx %d no space for %d\n",
608                                exp->exp_obd->obd_name,
609                                exp->exp_client_uuid.uuid, exp, n, bytes);
610                 }
611         }
612
613         /* Now substract what client have used already.  We don't subtract
614          * this from the tot_granted yet, so that other client's can't grab
615          * that space before we have actually allocated our blocks.  That
616          * happens in filter_grant_commit() after the writes are done. */
617         *left -= ungranted;
618         fed->fed_grant -= used;
619         fed->fed_pending += used + ungranted;
620         exp->exp_obd->u.filter.fo_tot_granted += ungranted;
621         exp->exp_obd->u.filter.fo_tot_pending += used + ungranted;
622
623         CDEBUG(D_CACHE,
624                "%s: cli %s/%p used: %lu ungranted: %lu grant: %lu dirty: %lu\n",
625                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp, used,
626                ungranted, fed->fed_grant, fed->fed_dirty);
627
628         /* Rough calc in case we don't refresh cached statfs data */
629         using = (used + ungranted + 1 ) >>
630                 exp->exp_obd->u.obt.obt_sb->s_blocksize_bits;
631         if (exp->exp_obd->obd_osfs.os_bavail > using)
632                 exp->exp_obd->obd_osfs.os_bavail -= using;
633         else
634                 exp->exp_obd->obd_osfs.os_bavail = 0;
635
636         if (fed->fed_dirty < used) {
637                 CERROR("%s: cli %s/%p claims used %lu > fed_dirty %lu\n",
638                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
639                        used, fed->fed_dirty);
640                 used = fed->fed_dirty;
641         }
642         exp->exp_obd->u.filter.fo_tot_dirty -= used;
643         fed->fed_dirty -= used;
644
645         if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
646                 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
647                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
648                        fed->fed_dirty, fed->fed_pending, fed->fed_grant);
649                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
650                 LBUG();
651         }
652         return rc;
653 }
654
655 /* If we ever start to support multi-object BRW RPCs, we will need to get locks
656  * on mulitple inodes.  That isn't all, because there still exists the
657  * possibility of a truncate starting a new transaction while holding the ext3
658  * rwsem = write while some writes (which have started their transactions here)
659  * blocking on the ext3 rwsem = read => lock inversion.
660  *
661  * The handling gets very ugly when dealing with locked pages.  It may be easier
662  * to just get rid of the locked page code (which has problems of its own) and
663  * either discover we do not need it anymore (i.e. it was a symptom of another
664  * bug) or ensure we get the page locks in an appropriate order. */
665 static int filter_preprw_write(int cmd, struct obd_export *exp, struct obdo *oa,
666                                int objcount, struct obd_ioobj *obj,
667                                struct niobuf_remote *nb, int *npages,
668                                struct niobuf_local *res,
669                                struct obd_trans_info *oti,
670                                struct lustre_capa *capa)
671 {
672         struct obd_device *obd = exp->exp_obd;
673         struct timeval start, end;
674         struct lvfs_run_ctxt saved;
675         struct niobuf_local *lnb = res;
676         struct fsfilt_objinfo fso;
677         struct filter_mod_data *fmd;
678         struct dentry *dentry = NULL;
679         void *iobuf;
680         obd_size left;
681         unsigned long now = jiffies, timediff;
682         int rc = 0, i, tot_bytes = 0, cleanup_phase = 0, localreq = 0;
683         ENTRY;
684         LASSERT(objcount == 1);
685         LASSERT(obj->ioo_bufcnt > 0);
686
687         rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa), capa,
688                               CAPA_OPC_OSS_WRITE);
689         if (rc)
690                 RETURN(rc);
691
692         if (exp->exp_connection->c_peer.nid == exp->exp_connection->c_self) 
693                 localreq = 1;
694
695         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
696         iobuf = filter_iobuf_get(&obd->u.filter, oti);
697         if (IS_ERR(iobuf))
698                 GOTO(cleanup, rc = PTR_ERR(iobuf));
699         cleanup_phase = 1;
700
701         dentry = filter_fid2dentry(obd, NULL, obj->ioo_gr,
702                                    obj->ioo_id);
703         if (IS_ERR(dentry))
704                 GOTO(cleanup, rc = PTR_ERR(dentry));
705         cleanup_phase = 2;
706
707         if (dentry->d_inode == NULL) {
708                 CERROR("%s: trying to BRW to non-existent file "LPU64"\n",
709                        obd->obd_name, obj->ioo_id);
710                 GOTO(cleanup, rc = -ENOENT);
711         }
712
713         if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID) &&
714             dentry->d_inode->i_mode & (S_ISUID | S_ISGID)) {
715                 rc = filter_capa_fixoa(exp, oa, obdo_mdsno(oa), capa);
716                 if (rc)
717                         GOTO(cleanup, rc);
718         }
719
720         rc = filter_map_remote_to_local(objcount, obj, nb, npages, res);
721         if (rc)
722                 GOTO(cleanup, rc);
723
724         fsfilt_check_slow(obd, now, "preprw_write setup");
725
726         /* Don't update inode timestamps if this write is older than a
727          * setattr which modifies the timestamps. b=10150 */
728         /* XXX when we start having persistent reservations this needs to
729          * be changed to filter_fmd_get() to create the fmd if it doesn't
730          * already exist so we can store the reservation handle there. */
731         fmd = filter_fmd_find(exp, obj->ioo_id, obj->ioo_gr);
732
733         LASSERT(oa != NULL);
734         spin_lock(&obd->obd_osfs_lock);
735         filter_grant_incoming(exp, oa);
736         if (fmd && fmd->fmd_mactime_xid > oti->oti_xid)
737                 oa->o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLCTIME |
738                                  OBD_MD_FLATIME);
739         else
740                 obdo_to_inode(dentry->d_inode, oa, OBD_MD_FLATIME |
741                               OBD_MD_FLMTIME | OBD_MD_FLCTIME);
742         cleanup_phase = 3;
743
744         left = filter_grant_space_left(exp);
745
746         fso.fso_dentry = dentry;
747         fso.fso_bufcnt = *npages;
748
749         rc = filter_grant_check(exp, oa, objcount, &fso, *npages, res,
750                                 &left, dentry->d_inode);
751
752         /* do not zero out oa->o_valid as it is used in filter_commitrw_write()
753          * for setting UID/GID and fid EA in first write time. */
754         if (oa->o_valid & OBD_MD_FLGRANT)
755                 oa->o_grant = filter_grant(exp,oa->o_grant,oa->o_undirty,left);
756
757         spin_unlock(&obd->obd_osfs_lock);
758         filter_fmd_put(exp, fmd);
759
760         if (rc)
761                 GOTO(cleanup, rc);
762         cleanup_phase = 4;
763
764         /* Filter truncate first locks i_mutex then partally truncated
765          * page, filter write code first locks pages then take
766          * i_mutex.  To avoid a deadlock in case of concurrent
767          * punch/write requests from one client, filter writes and
768          * filter truncates are serialized by i_alloc_sem, allowing
769          * multiple writes or single truncate. */
770         down_read(&dentry->d_inode->i_alloc_sem);
771
772         do_gettimeofday(&start);
773         for (i = 0, lnb = res; i < *npages; i++, lnb++) {
774
775                 /* We still set up for ungranted pages so that granted pages
776                  * can be written to disk as they were promised, and portals
777                  * needs to keep the pages all aligned properly. */
778                 lnb->dentry = dentry;
779
780                 lnb->page = filter_get_page(obd, dentry->d_inode, lnb->offset,
781                                             localreq);
782                 if (lnb->page == NULL)
783                         GOTO(cleanup, rc = -ENOMEM);
784
785                 /* DLM locking protects us from write and truncate competing
786                  * for same region, but truncate can leave dirty page in the
787                  * cache. it's possible the writeout on a such a page is in
788                  * progress when we access it. it's also possible that during
789                  * this writeout we put new (partial) data, but then won't
790                  * be able to proceed in filter_commitrw_write(). thus let's
791                  * just wait for writeout completion, should be rare enough.
792                  * -bzzz */
793                 if (obd->u.filter.fo_writethrough_cache)
794                         wait_on_page_writeback(lnb->page);
795                 BUG_ON(PageWriteback(lnb->page));
796
797                 /* If the filter writes a partial page, then has the file
798                  * extended, the client will read in the whole page.  the
799                  * filter has to be careful to zero the rest of the partial
800                  * page on disk.  we do it by hand for partial extending
801                  * writes, send_bio() is responsible for zeroing pages when
802                  * asked to read unmapped blocks -- brw_kiovec() does this. */
803                 if (lnb->len != CFS_PAGE_SIZE) {
804                         __s64 maxidx;
805
806                         maxidx = ((i_size_read(dentry->d_inode) +
807                                    CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT) - 1;
808                         if (maxidx >= lnb->page->index) {
809                                 LL_CDEBUG_PAGE(D_PAGE, lnb->page, "write %u @ "
810                                                LPU64" flg %x before EOF %llu\n",
811                                                lnb->len, lnb->offset,lnb->flags,
812                                                i_size_read(dentry->d_inode));
813                                 filter_iobuf_add_page(obd, iobuf,
814                                                       dentry->d_inode,
815                                                       lnb->page);
816                         } else {
817                                 long off;
818                                 char *p = kmap(lnb->page);
819
820                                 off = lnb->offset & ~CFS_PAGE_MASK;
821                                 if (off)
822                                         memset(p, 0, off);
823                                 off = (lnb->offset + lnb->len) & ~CFS_PAGE_MASK;
824                                 if (off)
825                                         memset(p + off, 0, CFS_PAGE_SIZE - off);
826                                 kunmap(lnb->page);
827                         }
828                 }
829                 if (lnb->rc == 0)
830                         tot_bytes += lnb->len;
831         }
832         do_gettimeofday(&end);
833         timediff = cfs_timeval_sub(&end, &start, NULL);
834         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
835
836         if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
837                 GOTO(cleanup, rc = -ENOMEM);
838
839         /* don't unlock pages to prevent any access */
840         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf, exp,
841                               NULL, NULL, NULL);
842
843         fsfilt_check_slow(obd, now, "start_page_write");
844
845         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
846                 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
847                                     LPROC_FILTER_WRITE_BYTES, tot_bytes);
848         EXIT;
849 cleanup:
850         switch(cleanup_phase) {
851         case 4:
852                 if (rc) {
853                         for (i = 0, lnb = res; i < *npages; i++, lnb++) {
854                                 if (lnb->page != NULL) {
855                                         unlock_page(lnb->page);
856                                         page_cache_release(lnb->page);
857                                         lnb->page = NULL;
858                                 }
859                         }
860                         up_read(&dentry->d_inode->i_alloc_sem);
861                 }
862         case 3:
863                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
864         case 2:
865                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
866                 if (rc)
867                         f_dput(dentry);
868                 break;
869         case 1:
870                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
871         case 0:
872                 spin_lock(&obd->obd_osfs_lock);
873                 if (oa)
874                         filter_grant_incoming(exp, oa);
875                 spin_unlock(&obd->obd_osfs_lock);
876                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
877                 break;
878         default:;
879         }
880         return rc;
881 }
882
883 int filter_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
884                   int objcount, struct obd_ioobj *obj,
885                   struct niobuf_remote *nb, int *npages,
886                   struct niobuf_local *res, struct obd_trans_info *oti,
887                   struct lustre_capa *capa)
888 {
889         if (cmd == OBD_BRW_WRITE)
890                 return filter_preprw_write(cmd, exp, oa, objcount, obj,
891                                            nb, npages, res, oti, capa);
892         if (cmd == OBD_BRW_READ)
893                 return filter_preprw_read(cmd, exp, oa, objcount, obj,
894                                           nb, npages, res, oti, capa);
895         LBUG();
896         return -EPROTO;
897 }
898
899 static int filter_commitrw_read(struct obd_export *exp, struct obdo *oa,
900                                 int objcount, struct obd_ioobj *obj,
901                                 struct niobuf_remote *rnb,
902                                 int npages, struct niobuf_local *res,
903                                 struct obd_trans_info *oti, int rc)
904 {
905         struct filter_obd *fo = &exp->exp_obd->u.filter;
906         struct inode *inode = NULL;
907         struct ldlm_res_id res_id;
908         struct ldlm_resource *resource = NULL;
909         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
910         struct niobuf_local *lnb;
911         int i;
912         ENTRY;
913
914         osc_build_res_name(obj->ioo_id, obj->ioo_gr, &res_id);
915         /* If oa != NULL then filter_preprw_read updated the inode atime
916          * and we should update the lvb so that other glimpses will also
917          * get the updated value. bug 5972 */
918         if (oa && ns && ns->ns_lvbo && ns->ns_lvbo->lvbo_update) {
919                 resource = ldlm_resource_get(ns, NULL, &res_id, LDLM_EXTENT, 0);
920
921                 if (resource != NULL) {
922                         LDLM_RESOURCE_ADDREF(resource);
923                         ns->ns_lvbo->lvbo_update(resource, NULL, 0, 1);
924                         LDLM_RESOURCE_DELREF(resource);
925                         ldlm_resource_putref(resource);
926                 }
927         }
928
929         if (res->dentry != NULL)
930                 inode = res->dentry->d_inode;
931
932         for (i = 0, lnb = res; i < npages; i++, lnb++)
933                 if (lnb->page != NULL)
934                         page_cache_release(lnb->page);
935
936         if (inode && (fo->fo_read_cache == 0 ||
937                       i_size_read(inode) > fo->fo_readcache_max_filesize))
938                 filter_truncate_cache(exp->exp_obd, obj, rnb, npages, res,
939                                       inode);
940
941         for (i = 0, lnb = res; i < npages; i++, lnb++)
942                 lnb->page = NULL;
943
944         if (res->dentry != NULL)
945                 f_dput(res->dentry);
946         RETURN(rc);
947 }
948
949 void filter_grant_commit(struct obd_export *exp, int niocount,
950                          struct niobuf_local *res)
951 {
952         struct filter_obd *filter = &exp->exp_obd->u.filter;
953         struct niobuf_local *lnb = res;
954         unsigned long pending = 0;
955         int i;
956
957         spin_lock(&exp->exp_obd->obd_osfs_lock);
958         for (i = 0, lnb = res; i < niocount; i++, lnb++)
959                 pending += lnb->lnb_grant_used;
960
961         LASSERTF(exp->exp_filter_data.fed_pending >= pending,
962                  "%s: cli %s/%p fed_pending: %lu grant_used: %lu\n",
963                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
964                  exp->exp_filter_data.fed_pending, pending);
965         exp->exp_filter_data.fed_pending -= pending;
966         LASSERTF(filter->fo_tot_granted >= pending,
967                  "%s: cli %s/%p tot_granted: "LPU64" grant_used: %lu\n",
968                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
969                  exp->exp_obd->u.filter.fo_tot_granted, pending);
970         filter->fo_tot_granted -= pending;
971         LASSERTF(filter->fo_tot_pending >= pending,
972                  "%s: cli %s/%p tot_pending: "LPU64" grant_used: %lu\n",
973                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
974                  filter->fo_tot_pending, pending);
975         filter->fo_tot_pending -= pending;
976
977         spin_unlock(&exp->exp_obd->obd_osfs_lock);
978 }
979
980 int filter_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
981                     int objcount, struct obd_ioobj *obj,
982                     struct niobuf_remote *nb, int npages,
983                     struct niobuf_local *res, struct obd_trans_info *oti,
984                     int rc)
985 {
986         if (cmd == OBD_BRW_WRITE)
987                 return filter_commitrw_write(exp, oa, objcount, obj,
988                                              nb, npages, res, oti, rc);
989         if (cmd == OBD_BRW_READ)
990                 return filter_commitrw_read(exp, oa, objcount, obj,
991                                             nb, npages, res, oti, rc);
992         LBUG();
993         return -EPROTO;
994 }
995
996 int filter_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
997                obd_count oa_bufs, struct brw_page *pga,
998                struct obd_trans_info *oti)
999 {
1000         struct obd_ioobj ioo;
1001         struct niobuf_local *lnb;
1002         struct niobuf_remote *rnb;
1003         obd_count i;
1004         int ret = 0, npages;
1005         ENTRY;
1006
1007         OBD_ALLOC(lnb, oa_bufs * sizeof(struct niobuf_local));
1008         OBD_ALLOC(rnb, oa_bufs * sizeof(struct niobuf_remote));
1009
1010         if (lnb == NULL || rnb == NULL)
1011                 GOTO(out, ret = -ENOMEM);
1012
1013         for (i = 0; i < oa_bufs; i++) {
1014                 lnb[i].page = pga[i].pg;
1015                 rnb[i].offset = pga[i].off;
1016                 rnb[i].len = pga[i].count;
1017         }
1018
1019         obdo_to_ioobj(oinfo->oi_oa, &ioo);
1020         ioo.ioo_bufcnt = oa_bufs;
1021
1022         npages = oa_bufs;
1023         ret = filter_preprw(cmd, exp, oinfo->oi_oa, 1, &ioo,
1024                             rnb, &npages, lnb, oti, oinfo_capa(oinfo));
1025         if (ret != 0)
1026                 GOTO(out, ret);
1027         LASSERTF(oa_bufs == npages, "%u != %u\n", oa_bufs, npages);
1028
1029         ret = filter_commitrw(cmd, exp, oinfo->oi_oa, 1, &ioo, rnb,
1030                               npages, lnb, oti, ret);
1031
1032 out:
1033         if (lnb)
1034                 OBD_FREE(lnb, oa_bufs * sizeof(struct niobuf_local));
1035         if (rnb)
1036                 OBD_FREE(rnb, oa_bufs * sizeof(struct niobuf_remote));
1037         RETURN(ret);
1038 }