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 static 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         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);
 116                 LBUG();
 117         }
 118         EXIT;
 119 }
 120
 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.
 124  *
 125  * Caller must hold obd_osfs_lock. */
 126 obd_size filter_grant_space_left(struct obd_export *exp)
 127 {
 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;
 132
 133         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
 134
 135         if (cfs_time_before_64(obd->obd_osfs_age, cfs_time_current_64() - HZ)) {
 136 restat:
 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 */
 140                         RETURN(0);
 141                 statfs_done = 1;
 142         }
 143
 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;
 148         } else {
 149                 left = 0 /* << blockbits */;
 150         }
 151
 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");
 154                 goto restat;
 155         }
 156
 157         if (left >= tot_granted) {
 158                 left -= tot_granted;
 159         } else {
 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);
 165                 }
 166                 left = 0;
 167         }
 168
 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);
 174
 175         return left;
 176 }
 177
 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.
 180  *
 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)
 184 {
 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;
 188         __u64 grant = 0;
 189
 190         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
 191
 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.
 196          *
 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);
 208                 grant <<= blockbits;
 209
 210                 if (grant) {
 211                         /* Allow >FILTER_GRANT_CHUNK size when clients
 212                          * reconnect due to a server reboot.
 213                          */
 214                         if ((grant > FILTER_GRANT_CHUNK) &&
 215                             (!obd->obd_recovering))
 216                                 grant = FILTER_GRANT_CHUNK;
 217
 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
 222                                        "current"LPU64"\n",
 223                                        obd->obd_name, exp->exp_client_uuid.uuid,
 224                                        exp, fed->fed_grant, want,current_grant);
 225                                 spin_unlock(&obd->obd_osfs_lock);
 226                                 LBUG();
 227                         }
 228                 }
 229         }
 230
 231         CDEBUG(D_CACHE,
 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);
 235         CDEBUG(D_CACHE,
 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);
 240
 241         return grant;
 242 }
 243
 244 /*
 245  * the routine is used to request pages from pagecache
 246  *
 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.
 249  *
 250  * use __GFP_HIGHMEM so that the pages can use all of the available memory
 251  * on 32-bit machines
 252  */
 253 static struct page *filter_get_page(struct obd_device *obd,
 254                                     struct inode *inode,
 255                                     obd_off offset)
 256 {
 257         struct page *page;
 258
 259         page = find_or_create_page(inode->i_mapping, offset >> CFS_PAGE_SHIFT,
 260                                    GFP_NOFS | __GFP_HIGHMEM);
 261         if (unlikely(page == NULL))
 262                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_NO_PAGE, 1);
 263
 264         return page;
 265 }
 266
 267 /*
 268  * the routine initializes array of local_niobuf from remote_niobuf
 269  */
 270 static int filter_map_remote_to_local(int objcount, struct obd_ioobj *obj,
 271                                       struct niobuf_remote *nb,
 272                                       int *nrpages, struct niobuf_local *res)
 273 {
 274         struct niobuf_remote *rnb;
 275         struct niobuf_local *lnb;
 276         int i, max;
 277         ENTRY;
 278
 279         /* we don't support multiobject RPC yet
 280          * ost_brw_read() and ost_brw_write() check this */
 281         LASSERT(objcount == 1);
 282
 283         max = *nrpages;
 284         *nrpages = 0;
 285         for (i = 0, rnb = nb, lnb = res; i < obj->ioo_bufcnt; i++, rnb++) {
 286                 obd_off offset = rnb->offset;
 287                 unsigned int len = rnb->len;
 288
 289                 while (len > 0) {
 290                         int poff = offset & (CFS_PAGE_SIZE - 1);
 291                         int plen = CFS_PAGE_SIZE - poff;
 292
 293                         if (*nrpages >= max) {
 294                                 CERROR("small array of local bufs: %d\n", max);
 295                                 RETURN(-EINVAL);
 296                         }
 297
 298                         if (plen > len)
 299                                 plen = len;
 300                         lnb->offset = offset;
 301                         lnb->len = plen;
 302                         lnb->flags = rnb->flags;
 303                         lnb->page = NULL;
 304                         lnb->rc = 0;
 305                         lnb->lnb_grant_used = 0;
 306
 307                         LASSERTF(plen <= len, "plen %u, len %u\n", plen, len);
 308                         offset += plen;
 309                         len -= plen;
 310                         lnb++;
 311                         (*nrpages)++;
 312                 }
 313         }
 314         RETURN(0);
 315 }
 316
 317 /*
 318  * the function is used to free all pages used for request
 319  * just to mimic cacheless OSS which don't occupy much memory
 320  */
 321 void filter_invalidate_cache(struct obd_device *obd, struct obd_ioobj *obj,
 322                              struct niobuf_remote *nb, struct inode *inode)
 323 {
 324         struct niobuf_remote *rnb;
 325         int i;
 326
 327         LASSERT(inode != NULL);
 328
 329         for (i = 0, rnb = nb; i < obj->ioo_bufcnt; i++, rnb++) {
 330                 obd_off start;
 331                 obd_off end;
 332
 333                 start = rnb->offset >> CFS_PAGE_SHIFT;
 334                 end = (rnb->offset + rnb->len) >> CFS_PAGE_SHIFT;
 335                 invalidate_mapping_pages(inode->i_mapping, start, end);
 336                 /* just to avoid warnings */
 337                 start = 0;
 338                 end = 0;
 339         }
 340 }
 341
 342 static int filter_preprw_read(int cmd, struct obd_export *exp, struct obdo *oa,
 343                               int objcount, struct obd_ioobj *obj,
 344                               struct niobuf_remote *nb,
 345                               int *npages, struct niobuf_local *res,
 346                               struct obd_trans_info *oti,
 347                               struct lustre_capa *capa)
 348 {
 349         struct obd_device *obd = exp->exp_obd;
 350         struct timeval start, end;
 351         struct lvfs_run_ctxt saved;
 352         struct niobuf_local *lnb;
 353         struct dentry *dentry = NULL;
 354         struct inode *inode = NULL;
 355         void *iobuf = NULL;
 356         int rc = 0, i, tot_bytes = 0;
 357         unsigned long now = jiffies;
 358         long timediff;
 359         ENTRY;
 360
 361         /* We are currently not supporting multi-obj BRW_READ RPCS at all.
 362          * When we do this function's dentry cleanup will need to be fixed.
 363          * These values are verified in ost_brw_write() from the wire. */
 364         LASSERTF(objcount == 1, "%d\n", objcount);
 365         LASSERTF(obj->ioo_bufcnt > 0, "%d\n", obj->ioo_bufcnt);
 366
 367         rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa), capa,
 368                               CAPA_OPC_OSS_READ);
 369         if (rc)
 370                 RETURN(rc);
 371
 372         if (oa && oa->o_valid & OBD_MD_FLGRANT) {
 373                 spin_lock(&obd->obd_osfs_lock);
 374                 filter_grant_incoming(exp, oa);
 375
 376                 oa->o_grant = 0;
 377                 spin_unlock(&obd->obd_osfs_lock);
 378         }
 379
 380         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 381         if (IS_ERR(iobuf))
 382                 RETURN(PTR_ERR(iobuf));
 383
 384         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 385         dentry = filter_oa2dentry(obd, oa);
 386         if (IS_ERR(dentry)) {
 387                 rc = PTR_ERR(dentry);
 388                 dentry = NULL;
 389                 GOTO(cleanup, rc);
 390         }
 391
 392         inode = dentry->d_inode;
 393
 394         obdo_to_inode(inode, oa, OBD_MD_FLATIME);
 395
 396         rc = filter_map_remote_to_local(objcount, obj, nb, npages, res);
 397         if (rc)
 398                 GOTO(cleanup, rc);
 399
 400         fsfilt_check_slow(obd, now, "preprw_read setup");
 401
 402         /* find pages for all segments, fill array with them */
 403         do_gettimeofday(&start);
 404         for (i = 0, lnb = res; i < *npages; i++, lnb++) {
 405
 406                 lnb->dentry = dentry;
 407
 408                 if (i_size_read(inode) <= lnb->offset)
 409                         /* If there's no more data, abort early.  lnb->rc == 0,
 410                          * so it's easy to detect later. */
 411                         break;
 412
 413                 lnb->page = filter_get_page(obd, inode, lnb->offset);
 414                 if (lnb->page == NULL)
 415                         GOTO(cleanup, rc = -ENOMEM);
 416
 417                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS, 1);
 418
 419                 if (i_size_read(inode) < lnb->offset + lnb->len - 1)
 420                         lnb->rc = i_size_read(inode) - lnb->offset;
 421                 else
 422                         lnb->rc = lnb->len;
 423
 424                 tot_bytes += lnb->rc;
 425
 426                 if (PageUptodate(lnb->page)) {
 427                         lprocfs_counter_add(obd->obd_stats,
 428                                             LPROC_FILTER_CACHE_HIT, 1);
 429                         continue;
 430                 }
 431
 432                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_MISS, 1);
 433                 filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 434         }
 435         do_gettimeofday(&end);
 436         timediff = cfs_timeval_sub(&end, &start, NULL);
 437         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
 438
 439         if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
 440                 GOTO(cleanup, rc = -ENOMEM);
 441
 442         fsfilt_check_slow(obd, now, "start_page_read");
 443
 444         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf,
 445                               exp, NULL, NULL, NULL);
 446         if (rc)
 447                 GOTO(cleanup, rc);
 448
 449         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_READ_BYTES, tot_bytes);
 450
 451         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
 452                 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
 453                                     LPROC_FILTER_READ_BYTES, tot_bytes);
 454
 455         EXIT;
 456
 457  cleanup:
 458         /* unlock pages to allow access from concurrent OST_READ */
 459         for (i = 0, lnb = res; i < *npages; i++, lnb++) {
 460                 if (lnb->page) {
 461                         LASSERT(PageLocked(lnb->page));
 462                         unlock_page(lnb->page);
 463
 464                         if (rc) {
 465                                 page_cache_release(lnb->page);
 466                                 lnb->page = NULL;
 467                         }
 468                 }
 469         }
 470
 471         if (rc != 0) {
 472                 if (dentry != NULL)
 473                         f_dput(dentry);
 474         }
 475
 476         filter_iobuf_put(&obd->u.filter, iobuf, oti);
 477
 478         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 479         if (rc)
 480                 CERROR("io error %d\n", rc);
 481
 482         return rc;
 483 }
 484
 485 /* When clients have dirtied as much space as they've been granted they
 486  * fall through to sync writes.  These sync writes haven't been expressed
 487  * in grants and need to error with ENOSPC when there isn't room in the
 488  * filesystem for them after grants are taken into account.  However,
 489  * writeback of the dirty data that was already granted space can write
 490  * right on through.
 491  *
 492  * Caller must hold obd_osfs_lock. */
 493 static int filter_grant_check(struct obd_export *exp, struct obdo *oa,
 494                               int objcount, struct fsfilt_objinfo *fso,
 495                               int niocount, struct niobuf_local *lnb,
 496                               obd_size *left, struct inode *inode)
 497 {
 498         struct filter_export_data *fed = &exp->exp_filter_data;
 499         int blocksize = exp->exp_obd->u.obt.obt_sb->s_blocksize;
 500         unsigned long used = 0, ungranted = 0, using;
 501         int i, rc = -ENOSPC, obj, n = 0;
 502
 503         LASSERT_SPIN_LOCKED(&exp->exp_obd->obd_osfs_lock);
 504
 505         for (obj = 0; obj < objcount; obj++) {
 506                 for (i = 0; i < fso[obj].fso_bufcnt; i++, n++) {
 507                         int tmp, bytes;
 508
 509                         /* should match the code in osc_exit_cache */
 510                         bytes = lnb[n].len;
 511                         bytes += lnb[n].offset & (blocksize - 1);
 512                         tmp = (lnb[n].offset + lnb[n].len) & (blocksize - 1);
 513                         if (tmp)
 514                                 bytes += blocksize - tmp;
 515
 516                         if ((lnb[n].flags & OBD_BRW_FROM_GRANT) &&
 517                             (oa->o_valid & OBD_MD_FLGRANT)) {
 518                                 if (fed->fed_grant < used + bytes) {
 519                                         CDEBUG(D_CACHE,
 520                                                "%s: cli %s/%p claims %ld+%d "
 521                                                "GRANT, real grant %lu idx %d\n",
 522                                                exp->exp_obd->obd_name,
 523                                                exp->exp_client_uuid.uuid, exp,
 524                                                used, bytes, fed->fed_grant, n);
 525                                 } else {
 526                                         used += bytes;
 527                                         lnb[n].flags |= OBD_BRW_GRANTED;
 528                                         lnb[n].lnb_grant_used = bytes;
 529                                         CDEBUG(0, "idx %d used=%lu\n", n, used);
 530                                         rc = 0;
 531                                         continue;
 532                                 }
 533                         }
 534                         if (*left > ungranted + bytes) {
 535                                 /* if enough space, pretend it was granted */
 536                                 ungranted += bytes;
 537                                 lnb[n].flags |= OBD_BRW_GRANTED;
 538                                 lnb[n].lnb_grant_used = bytes;
 539                                 CDEBUG(0, "idx %d ungranted=%lu\n",n,ungranted);
 540                                 rc = 0;
 541                                 continue;
 542                         }
 543
 544                         /* We can't check for already-mapped blocks here, as
 545                          * it requires dropping the osfs lock to do the bmap.
 546                          * Instead, we return ENOSPC and in that case we need
 547                          * to go through and verify if all of the blocks not
 548                          * marked BRW_GRANTED are already mapped and we can
 549                          * ignore this error. */
 550                         lnb[n].rc = -ENOSPC;
 551                         lnb[n].flags &= ~OBD_BRW_GRANTED;
 552                         CDEBUG(D_CACHE,"%s: cli %s/%p idx %d no space for %d\n",
 553                                exp->exp_obd->obd_name,
 554                                exp->exp_client_uuid.uuid, exp, n, bytes);
 555                 }
 556         }
 557
 558         /* Now substract what client have used already.  We don't subtract
 559          * this from the tot_granted yet, so that other client's can't grab
 560          * that space before we have actually allocated our blocks.  That
 561          * happens in filter_grant_commit() after the writes are done. */
 562         *left -= ungranted;
 563         fed->fed_grant -= used;
 564         fed->fed_pending += used + ungranted;
 565         exp->exp_obd->u.filter.fo_tot_granted += ungranted;
 566         exp->exp_obd->u.filter.fo_tot_pending += used + ungranted;
 567
 568         CDEBUG(D_CACHE,
 569                "%s: cli %s/%p used: %lu ungranted: %lu grant: %lu dirty: %lu\n",
 570                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp, used,
 571                ungranted, fed->fed_grant, fed->fed_dirty);
 572
 573         /* Rough calc in case we don't refresh cached statfs data */
 574         using = (used + ungranted + 1 ) >>
 575                 exp->exp_obd->u.obt.obt_sb->s_blocksize_bits;
 576         if (exp->exp_obd->obd_osfs.os_bavail > using)
 577                 exp->exp_obd->obd_osfs.os_bavail -= using;
 578         else
 579                 exp->exp_obd->obd_osfs.os_bavail = 0;
 580
 581         if (fed->fed_dirty < used) {
 582                 CERROR("%s: cli %s/%p claims used %lu > fed_dirty %lu\n",
 583                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 584                        used, fed->fed_dirty);
 585                 used = fed->fed_dirty;
 586         }
 587         exp->exp_obd->u.filter.fo_tot_dirty -= used;
 588         fed->fed_dirty -= used;
 589
 590         if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
 591                 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
 592                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 593                        fed->fed_dirty, fed->fed_pending, fed->fed_grant);
 594                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
 595                 LBUG();
 596         }
 597         return rc;
 598 }
 599
 600 /* If we ever start to support multi-object BRW RPCs, we will need to get locks
 601  * on mulitple inodes.  That isn't all, because there still exists the
 602  * possibility of a truncate starting a new transaction while holding the ext3
 603  * rwsem = write while some writes (which have started their transactions here)
 604  * blocking on the ext3 rwsem = read => lock inversion.
 605  *
 606  * The handling gets very ugly when dealing with locked pages.  It may be easier
 607  * to just get rid of the locked page code (which has problems of its own) and
 608  * either discover we do not need it anymore (i.e. it was a symptom of another
 609  * bug) or ensure we get the page locks in an appropriate order. */
 610 static int filter_preprw_write(int cmd, struct obd_export *exp, struct obdo *oa,
 611                                int objcount, struct obd_ioobj *obj,
 612                                struct niobuf_remote *nb, int *npages,
 613                                struct niobuf_local *res,
 614                                struct obd_trans_info *oti,
 615                                struct lustre_capa *capa)
 616 {
 617         struct obd_device *obd = exp->exp_obd;
 618         struct timeval start, end;
 619         struct lvfs_run_ctxt saved;
 620         struct niobuf_local *lnb = res;
 621         struct fsfilt_objinfo fso;
 622         struct filter_mod_data *fmd;
 623         struct dentry *dentry = NULL;
 624         void *iobuf;
 625         obd_size left;
 626         unsigned long now = jiffies, timediff;
 627         int rc = 0, i, tot_bytes = 0, cleanup_phase = 0;
 628         ENTRY;
 629         LASSERT(objcount == 1);
 630         LASSERT(obj->ioo_bufcnt > 0);
 631
 632         rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa), capa,
 633                               CAPA_OPC_OSS_WRITE);
 634         if (rc)
 635                 RETURN(rc);
 636
 637         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 638         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 639         if (IS_ERR(iobuf))
 640                 GOTO(cleanup, rc = PTR_ERR(iobuf));
 641         cleanup_phase = 1;
 642
 643         dentry = filter_fid2dentry(obd, NULL, obj->ioo_gr,
 644                                    obj->ioo_id);
 645         if (IS_ERR(dentry))
 646                 GOTO(cleanup, rc = PTR_ERR(dentry));
 647         cleanup_phase = 2;
 648
 649         if (dentry->d_inode == NULL) {
 650                 CERROR("%s: trying to BRW to non-existent file "LPU64"\n",
 651                        obd->obd_name, obj->ioo_id);
 652                 GOTO(cleanup, rc = -ENOENT);
 653         }
 654
 655         if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID) &&
 656             dentry->d_inode->i_mode & (S_ISUID | S_ISGID)) {
 657                 rc = filter_capa_fixoa(exp, oa, obdo_mdsno(oa), capa);
 658                 if (rc)
 659                         GOTO(cleanup, rc);
 660         }
 661
 662         rc = filter_map_remote_to_local(objcount, obj, nb, npages, res);
 663         if (rc)
 664                 GOTO(cleanup, rc);
 665
 666         fsfilt_check_slow(obd, now, "preprw_write setup");
 667
 668         /* Don't update inode timestamps if this write is older than a
 669          * setattr which modifies the timestamps. b=10150 */
 670         /* XXX when we start having persistent reservations this needs to
 671          * be changed to filter_fmd_get() to create the fmd if it doesn't
 672          * already exist so we can store the reservation handle there. */
 673         fmd = filter_fmd_find(exp, obj->ioo_id, obj->ioo_gr);
 674
 675         LASSERT(oa != NULL);
 676         spin_lock(&obd->obd_osfs_lock);
 677         filter_grant_incoming(exp, oa);
 678         if (fmd && fmd->fmd_mactime_xid > oti->oti_xid)
 679                 oa->o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 680                                  OBD_MD_FLATIME);
 681         else
 682                 obdo_to_inode(dentry->d_inode, oa, OBD_MD_FLATIME |
 683                               OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 684         cleanup_phase = 3;
 685
 686         left = filter_grant_space_left(exp);
 687
 688         fso.fso_dentry = dentry;
 689         fso.fso_bufcnt = *npages;
 690
 691         rc = filter_grant_check(exp, oa, objcount, &fso, *npages, res,
 692                                 &left, dentry->d_inode);
 693
 694         /* do not zero out oa->o_valid as it is used in filter_commitrw_write()
 695          * for setting UID/GID and fid EA in first write time. */
 696         if (oa->o_valid & OBD_MD_FLGRANT)
 697                 oa->o_grant = filter_grant(exp,oa->o_grant,oa->o_undirty,left);
 698
 699         spin_unlock(&obd->obd_osfs_lock);
 700         filter_fmd_put(exp, fmd);
 701
 702         if (rc)
 703                 GOTO(cleanup, rc);
 704
 705         do_gettimeofday(&start);
 706         for (i = 0, lnb = res; i < *npages; i++, lnb++) {
 707
 708                 /* We still set up for ungranted pages so that granted pages
 709                  * can be written to disk as they were promised, and portals
 710                  * needs to keep the pages all aligned properly. */
 711                 lnb->dentry = dentry;
 712
 713                 lnb->page = filter_get_page(obd, dentry->d_inode, lnb->offset);
 714                 if (lnb->page == NULL)
 715                         GOTO(cleanup, rc = -ENOMEM);
 716                 cleanup_phase = 4;
 717
 718                 /* DLM locking protects us from write and truncate competing
 719                  * for same region, but truncate can leave dirty page in the
 720                  * cache. it's possible the writeout on a such a page is in
 721                  * progress when we access it. it's also possible that during
 722                  * this writeout we put new (partial) data, but then won't
 723                  * be able to proceed in filter_commitrw_write(). thus let's
 724                  * just wait for writeout completion, should be rare enough.
 725                  * -bzzz */
 726                 if (obd->u.filter.fo_writethrough_cache)
 727                         wait_on_page_writeback(lnb->page);
 728                 BUG_ON(PageWriteback(lnb->page));
 729
 730                 /* If the filter writes a partial page, then has the file
 731                  * extended, the client will read in the whole page.  the
 732                  * filter has to be careful to zero the rest of the partial
 733                  * page on disk.  we do it by hand for partial extending
 734                  * writes, send_bio() is responsible for zeroing pages when
 735                  * asked to read unmapped blocks -- brw_kiovec() does this. */
 736                 if (lnb->len != CFS_PAGE_SIZE) {
 737                         __s64 maxidx;
 738
 739                         maxidx = ((i_size_read(dentry->d_inode) +
 740                                    CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT) - 1;
 741                         if (maxidx >= lnb->page->index) {
 742                                 LL_CDEBUG_PAGE(D_PAGE, lnb->page, "write %u @ "
 743                                                LPU64" flg %x before EOF %llu\n",
 744                                                lnb->len, lnb->offset,lnb->flags,
 745                                                i_size_read(dentry->d_inode));
 746                                 filter_iobuf_add_page(obd, iobuf,
 747                                                       dentry->d_inode,
 748                                                       lnb->page);
 749                         } else {
 750                                 long off;
 751                                 char *p = kmap(lnb->page);
 752
 753                                 off = lnb->offset & ~CFS_PAGE_MASK;
 754                                 if (off)
 755                                         memset(p, 0, off);
 756                                 off = (lnb->offset + lnb->len) & ~CFS_PAGE_MASK;
 757                                 if (off)
 758                                         memset(p + off, 0, CFS_PAGE_SIZE - off);
 759                                 kunmap(lnb->page);
 760                         }
 761                 }
 762                 if (lnb->rc == 0)
 763                         tot_bytes += lnb->len;
 764         }
 765         do_gettimeofday(&end);
 766         timediff = cfs_timeval_sub(&end, &start, NULL);
 767         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
 768
 769         if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
 770                 GOTO(cleanup, rc = -ENOMEM);
 771
 772         /* don't unlock pages to prevent any access */
 773         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf, exp,
 774                               NULL, NULL, NULL);
 775
 776         fsfilt_check_slow(obd, now, "start_page_write");
 777
 778         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
 779                 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
 780                                     LPROC_FILTER_WRITE_BYTES, tot_bytes);
 781         EXIT;
 782 cleanup:
 783         switch(cleanup_phase) {
 784         case 4:
 785                 if (rc) {
 786                         for (i = 0, lnb = res; i < *npages; i++, lnb++) {
 787                                 if (lnb->page != NULL) {
 788                                         unlock_page(lnb->page);
 789                                         page_cache_release(lnb->page);
 790                                         lnb->page = NULL;
 791                                 }
 792                         }
 793                 }
 794         case 3:
 795                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 796         case 2:
 797                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 798                 if (rc)
 799                         f_dput(dentry);
 800                 break;
 801         case 1:
 802                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 803         case 0:
 804                 spin_lock(&obd->obd_osfs_lock);
 805                 if (oa)
 806                         filter_grant_incoming(exp, oa);
 807                 spin_unlock(&obd->obd_osfs_lock);
 808                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 809                 break;
 810         default:;
 811         }
 812         return rc;
 813 }
 814
 815 int filter_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
 816                   int objcount, struct obd_ioobj *obj,
 817                   struct niobuf_remote *nb, int *npages,
 818                   struct niobuf_local *res, struct obd_trans_info *oti,
 819                   struct lustre_capa *capa)
 820 {
 821         if (cmd == OBD_BRW_WRITE)
 822                 return filter_preprw_write(cmd, exp, oa, objcount, obj,
 823                                            nb, npages, res, oti, capa);
 824         if (cmd == OBD_BRW_READ)
 825                 return filter_preprw_read(cmd, exp, oa, objcount, obj,
 826                                           nb, npages, res, oti, capa);
 827         LBUG();
 828         return -EPROTO;
 829 }
 830
 831 static int filter_commitrw_read(struct obd_export *exp, struct obdo *oa,
 832                                 int objcount, struct obd_ioobj *obj,
 833                                 struct niobuf_remote *rnb,
 834                                 int npages, struct niobuf_local *res,
 835                                 struct obd_trans_info *oti, int rc)
 836 {
 837         struct filter_obd *fo = &exp->exp_obd->u.filter;
 838         struct inode *inode = NULL;
 839         struct ldlm_res_id res_id;
 840         struct ldlm_resource *resource = NULL;
 841         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 842         struct niobuf_local *lnb;
 843         int i;
 844         ENTRY;
 845
 846         osc_build_res_name(obj->ioo_id, obj->ioo_gr, &res_id);
 847         /* If oa != NULL then filter_preprw_read updated the inode atime
 848          * and we should update the lvb so that other glimpses will also
 849          * get the updated value. bug 5972 */
 850         if (oa && ns && ns->ns_lvbo && ns->ns_lvbo->lvbo_update) {
 851                 resource = ldlm_resource_get(ns, NULL, &res_id, LDLM_EXTENT, 0);
 852
 853                 if (resource != NULL) {
 854                         LDLM_RESOURCE_ADDREF(resource);
 855                         ns->ns_lvbo->lvbo_update(resource, NULL, 0, 1);
 856                         LDLM_RESOURCE_DELREF(resource);
 857                         ldlm_resource_putref(resource);
 858                 }
 859         }
 860
 861         if (res->dentry != NULL)
 862                 inode = res->dentry->d_inode;
 863
 864         for (i = 0, lnb = res; i < npages; i++, lnb++) {
 865                 if (lnb->page != NULL) {
 866                         page_cache_release(lnb->page);
 867                         lnb->page = NULL;
 868                 }
 869         }
 870
 871         if (inode && (fo->fo_read_cache == 0 ||
 872                         i_size_read(inode) > fo->fo_readcache_max_filesize))
 873                 filter_invalidate_cache(exp->exp_obd, obj, rnb, inode);
 874
 875         if (res->dentry != NULL)
 876                 f_dput(res->dentry);
 877         RETURN(rc);
 878 }
 879
 880 void filter_grant_commit(struct obd_export *exp, int niocount,
 881                          struct niobuf_local *res)
 882 {
 883         struct filter_obd *filter = &exp->exp_obd->u.filter;
 884         struct niobuf_local *lnb = res;
 885         unsigned long pending = 0;
 886         int i;
 887
 888         spin_lock(&exp->exp_obd->obd_osfs_lock);
 889         for (i = 0, lnb = res; i < niocount; i++, lnb++)
 890                 pending += lnb->lnb_grant_used;
 891
 892         LASSERTF(exp->exp_filter_data.fed_pending >= pending,
 893                  "%s: cli %s/%p fed_pending: %lu grant_used: %lu\n",
 894                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 895                  exp->exp_filter_data.fed_pending, pending);
 896         exp->exp_filter_data.fed_pending -= pending;
 897         LASSERTF(filter->fo_tot_granted >= pending,
 898                  "%s: cli %s/%p tot_granted: "LPU64" grant_used: %lu\n",
 899                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 900                  exp->exp_obd->u.filter.fo_tot_granted, pending);
 901         filter->fo_tot_granted -= pending;
 902         LASSERTF(filter->fo_tot_pending >= pending,
 903                  "%s: cli %s/%p tot_pending: "LPU64" grant_used: %lu\n",
 904                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 905                  filter->fo_tot_pending, pending);
 906         filter->fo_tot_pending -= pending;
 907
 908         spin_unlock(&exp->exp_obd->obd_osfs_lock);
 909 }
 910
 911 int filter_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
 912                     int objcount, struct obd_ioobj *obj,
 913                     struct niobuf_remote *nb, int npages,
 914                     struct niobuf_local *res, struct obd_trans_info *oti,
 915                     int rc)
 916 {
 917         if (cmd == OBD_BRW_WRITE)
 918                 return filter_commitrw_write(exp, oa, objcount, obj,
 919                                              nb, npages, res, oti, rc);
 920         if (cmd == OBD_BRW_READ)
 921                 return filter_commitrw_read(exp, oa, objcount, obj,
 922                                             nb, npages, res, oti, rc);
 923         LBUG();
 924         return -EPROTO;
 925 }
 926
 927 int filter_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
 928                obd_count oa_bufs, struct brw_page *pga,
 929                struct obd_trans_info *oti)
 930 {
 931         struct obd_ioobj ioo;
 932         struct niobuf_local *lnb;
 933         struct niobuf_remote *rnb;
 934         obd_count i;
 935         int ret = 0, npages;
 936         ENTRY;
 937
 938         OBD_ALLOC(lnb, oa_bufs * sizeof(struct niobuf_local));
 939         OBD_ALLOC(rnb, oa_bufs * sizeof(struct niobuf_remote));
 940
 941         if (lnb == NULL || rnb == NULL)
 942                 GOTO(out, ret = -ENOMEM);
 943
 944         for (i = 0; i < oa_bufs; i++) {
 945                 lnb[i].page = pga[i].pg;
 946                 rnb[i].offset = pga[i].off;
 947                 rnb[i].len = pga[i].count;
 948         }
 949
 950         obdo_to_ioobj(oinfo->oi_oa, &ioo);
 951         ioo.ioo_bufcnt = oa_bufs;
 952
 953         npages = oa_bufs;
 954         ret = filter_preprw(cmd, exp, oinfo->oi_oa, 1, &ioo,
 955                             rnb, &npages, lnb, oti, oinfo_capa(oinfo));
 956         if (ret != 0)
 957                 GOTO(out, ret);
 958         LASSERTF(oa_bufs == npages, "%u != %u\n", oa_bufs, npages);
 959
 960         ret = filter_commitrw(cmd, exp, oinfo->oi_oa, 1, &ioo, rnb,
 961                               npages, lnb, oti, ret);
 962
 963 out:
 964         if (lnb)
 965                 OBD_FREE(lnb, oa_bufs * sizeof(struct niobuf_local));
 966         if (rnb)
 967                 OBD_FREE(rnb, oa_bufs * sizeof(struct niobuf_remote));
 968         RETURN(ret);
 969 }