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  *  linux/fs/obdfilter/filter_io.c
   5  *
   6  *  Copyright (c) 2001-2003 Cluster File Systems, Inc.
   7  *   Author: Peter Braam <braam@clusterfs.com>
   8  *   Author: Andreas Dilger <adilger@clusterfs.com>
   9  *   Author: Phil Schwan <phil@clusterfs.com>
  10  *
  11  *   This file is part of the Lustre file system, http://www.lustre.org
  12  *   Lustre is a trademark of Cluster File Systems, Inc.
  13  *
  14  *   You may have signed or agreed to another license before downloading
  15  *   this software.  If so, you are bound by the terms and conditions
  16  *   of that agreement, and the following does not apply to you.  See the
  17  *   LICENSE file included with this distribution for more information.
  18  *
  19  *   If you did not agree to a different license, then this copy of Lustre
  20  *   is open source software; you can redistribute it and/or modify it
  21  *   under the terms of version 2 of the GNU General Public License as
  22  *   published by the Free Software Foundation.
  23  *
  24  *   In either case, Lustre is distributed in the hope that it will be
  25  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  26  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  27  *   license text for more details.
  28  */
  29
  30 #define DEBUG_SUBSYSTEM S_FILTER
  31
  32 #ifndef AUTOCONF_INCLUDED
  33 #include <linux/config.h>
  34 #endif
  35 #include <linux/module.h>
  36 #include <linux/pagemap.h> // XXX kill me soon
  37 #include <linux/version.h>
  38
  39 #include <obd_class.h>
  40 #include <lustre_fsfilt.h>
  41 #include "filter_internal.h"
  42
  43 int *obdfilter_created_scratchpad;
  44
  45 static int filter_alloc_dio_page(struct obd_device *obd, struct inode *inode,
  46                                  struct niobuf_local *lnb)
  47 {
  48         struct page *page;
  49
  50         LASSERT(lnb->page != NULL);
  51
  52         page = lnb->page;
  53 #if 0
  54         POISON_PAGE(page, 0xf1);
  55         if (lnb->len != CFS_PAGE_SIZE) {
  56                 memset(kmap(page) + lnb->len, 0, CFS_PAGE_SIZE - lnb->len);
  57                 kunmap(page);
  58         }
  59 #endif
  60         page->index = lnb->offset >> CFS_PAGE_SHIFT;
  61
  62         RETURN(0);
  63 }
  64
  65 static void filter_free_dio_pages(int objcount, struct obd_ioobj *obj,
  66                            int niocount, struct niobuf_local *res)
  67 {
  68         int i, j;
  69
  70         for (i = 0; i < objcount; i++, obj++) {
  71                 for (j = 0 ; j < obj->ioo_bufcnt ; j++, res++)
  72                                 res->page = NULL;
  73         }
  74 }
  75
  76 /* Grab the dirty and seen grant announcements from the incoming obdo.
  77  * We will later calculate the clients new grant and return it.
  78  * Caller must hold osfs lock */
  79 static void filter_grant_incoming(struct obd_export *exp, struct obdo *oa)
  80 {
  81         struct filter_export_data *fed;
  82         struct obd_device *obd = exp->exp_obd;
  83         ENTRY;
  84
  85         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
  86
  87         if ((oa->o_valid & (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) !=
  88                                         (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) {
  89                 oa->o_valid &= ~OBD_MD_FLGRANT;
  90                 EXIT;
  91                 return;
  92         }
  93
  94         fed = &exp->exp_filter_data;
  95
  96         /* Add some margin, since there is a small race if other RPCs arrive
  97          * out-or-order and have already consumed some grant.  We want to
  98          * leave this here in case there is a large error in accounting. */
  99         CDEBUG(D_CACHE,
 100                "%s: cli %s/%p reports grant: "LPU64" dropped: %u, local: %lu\n",
 101                obd->obd_name, exp->exp_client_uuid.uuid, exp, oa->o_grant,
 102                oa->o_dropped, fed->fed_grant);
 103
 104         /* Update our accounting now so that statfs takes it into account.
 105          * Note that fed_dirty is only approximate and can become incorrect
 106          * if RPCs arrive out-of-order.  No important calculations depend
 107          * on fed_dirty however, but we must check sanity to not assert. */
 108         if ((long long)oa->o_dirty < 0)
 109                 oa->o_dirty = 0;
 110         else if (oa->o_dirty > fed->fed_grant + 4 * FILTER_GRANT_CHUNK)
 111                 oa->o_dirty = fed->fed_grant + 4 * FILTER_GRANT_CHUNK;
 112         obd->u.filter.fo_tot_dirty += oa->o_dirty - fed->fed_dirty;
 113         if (fed->fed_grant < oa->o_dropped) {
 114                 CDEBUG(D_CACHE,"%s: cli %s/%p reports %u dropped > grant %lu\n",
 115                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
 116                        oa->o_dropped, fed->fed_grant);
 117                 oa->o_dropped = 0;
 118         }
 119         if (obd->u.filter.fo_tot_granted < oa->o_dropped) {
 120                 CERROR("%s: cli %s/%p reports %u dropped > tot_grant "LPU64"\n",
 121                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
 122                        oa->o_dropped, obd->u.filter.fo_tot_granted);
 123                 oa->o_dropped = 0;
 124         }
 125         obd->u.filter.fo_tot_granted -= oa->o_dropped;
 126         fed->fed_grant -= oa->o_dropped;
 127         fed->fed_dirty = oa->o_dirty;
 128         if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
 129                 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
 130                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
 131                        fed->fed_dirty, fed->fed_pending, fed->fed_grant);
 132                 spin_unlock(&obd->obd_osfs_lock);
 133                 LBUG();
 134         }
 135         EXIT;
 136 }
 137
 138 /* Figure out how much space is available between what we've granted
 139  * and what remains in the filesystem.  Compensate for ext3 indirect
 140  * block overhead when computing how much free space is left ungranted.
 141  *
 142  * Caller must hold obd_osfs_lock. */
 143 obd_size filter_grant_space_left(struct obd_export *exp)
 144 {
 145         struct obd_device *obd = exp->exp_obd;
 146         int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
 147         obd_size tot_granted = obd->u.filter.fo_tot_granted, avail, left = 0;
 148         int rc, statfs_done = 0;
 149
 150         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
 151
 152         if (cfs_time_before_64(obd->obd_osfs_age, cfs_time_current_64() - HZ)) {
 153 restat:
 154                 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb,
 155                                    cfs_time_current_64() + HZ);
 156                 if (rc) /* N.B. statfs can't really fail */
 157                         RETURN(0);
 158                 statfs_done = 1;
 159         }
 160
 161         avail = obd->obd_osfs.os_bavail;
 162         left = avail - (avail >> (blockbits - 3)); /* (d)indirect */
 163         if (left > GRANT_FOR_LLOG(obd)) {
 164                 left = (left - GRANT_FOR_LLOG(obd)) << blockbits;
 165         } else {
 166                 left = 0 /* << blockbits */;
 167         }
 168
 169         if (!statfs_done && left < 32 * FILTER_GRANT_CHUNK + tot_granted) {
 170                 CDEBUG(D_CACHE, "fs has no space left and statfs too old\n");
 171                 goto restat;
 172         }
 173
 174         if (left >= tot_granted) {
 175                 left -= tot_granted;
 176         } else {
 177                 if (left < tot_granted - obd->u.filter.fo_tot_pending) {
 178                         CERROR("%s: cli %s/%p grant "LPU64" > available "
 179                                LPU64" and pending "LPU64"\n", obd->obd_name,
 180                                exp->exp_client_uuid.uuid, exp, tot_granted,
 181                                left, obd->u.filter.fo_tot_pending);
 182                 }
 183                 left = 0;
 184         }
 185
 186         CDEBUG(D_CACHE, "%s: cli %s/%p free: "LPU64" avail: "LPU64" grant "LPU64
 187                " left: "LPU64" pending: "LPU64"\n", obd->obd_name,
 188                exp->exp_client_uuid.uuid, exp,
 189                obd->obd_osfs.os_bfree << blockbits, avail << blockbits,
 190                tot_granted, left, obd->u.filter.fo_tot_pending);
 191
 192         return left;
 193 }
 194
 195 /* Calculate how much grant space to allocate to this client, based on how
 196  * much space is currently free and how much of that is already granted.
 197  *
 198  * Caller must hold obd_osfs_lock. */
 199 long filter_grant(struct obd_export *exp, obd_size current_grant,
 200                   obd_size want, obd_size fs_space_left)
 201 {
 202         struct obd_device *obd = exp->exp_obd;
 203         struct filter_export_data *fed = &exp->exp_filter_data;
 204         int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
 205         __u64 grant = 0;
 206
 207         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
 208
 209         /* Grant some fraction of the client's requested grant space so that
 210          * they are not always waiting for write credits (not all of it to
 211          * avoid overgranting in face of multiple RPCs in flight).  This
 212          * essentially will be able to control the OSC_MAX_RIF for a client.
 213          *
 214          * If we do have a large disparity between what the client thinks it
 215          * has and what we think it has, don't grant very much and let the
 216          * client consume its grant first.  Either it just has lots of RPCs
 217          * in flight, or it was evicted and its grants will soon be used up. */
 218         if (want > 0x7fffffff) {
 219                 CERROR("%s: client %s/%p requesting > 2GB grant "LPU64"\n",
 220                        obd->obd_name, exp->exp_client_uuid.uuid, exp, want);
 221         } else if (current_grant < want &&
 222                    current_grant < fed->fed_grant + FILTER_GRANT_CHUNK) {
 223                 grant = min((want >> blockbits),
 224                             (fs_space_left >> blockbits) / 8);
 225                 grant <<= blockbits;
 226
 227                 if (grant) {
 228                         /* Allow >FILTER_GRANT_CHUNK size when clients
 229                          * reconnect due to a server reboot.
 230                          */
 231                         if ((grant > FILTER_GRANT_CHUNK) &&
 232                             (!obd->obd_recovering))
 233                                 grant = FILTER_GRANT_CHUNK;
 234
 235                         obd->u.filter.fo_tot_granted += grant;
 236                         fed->fed_grant += grant;
 237                         if (fed->fed_grant < 0) {
 238                                 CERROR("%s: cli %s/%p grant %ld want "LPU64
 239                                        "current"LPU64"\n",
 240                                        obd->obd_name, exp->exp_client_uuid.uuid,
 241                                        exp, fed->fed_grant, want,current_grant);
 242                                 spin_unlock(&obd->obd_osfs_lock);
 243                                 LBUG();
 244                         }
 245                 }
 246         }
 247
 248         CDEBUG(D_CACHE,
 249                "%s: cli %s/%p wants: "LPU64" current grant "LPU64
 250                " granting: "LPU64"\n", obd->obd_name, exp->exp_client_uuid.uuid,
 251                exp, want, current_grant, grant);
 252         CDEBUG(D_CACHE,
 253                "%s: cli %s/%p tot cached:"LPU64" granted:"LPU64
 254                " num_exports: %d\n", obd->obd_name, exp->exp_client_uuid.uuid,
 255                exp, obd->u.filter.fo_tot_dirty,
 256                obd->u.filter.fo_tot_granted, obd->obd_num_exports);
 257
 258         return grant;
 259 }
 260
 261 static int filter_preprw_read(int cmd, struct obd_export *exp, struct obdo *oa,
 262                               int objcount, struct obd_ioobj *obj,
 263                               int niocount, struct niobuf_remote *nb,
 264                               struct niobuf_local *res,
 265                               struct obd_trans_info *oti,
 266                               struct lustre_capa *capa)
 267 {
 268         struct obd_device *obd = exp->exp_obd;
 269         struct lvfs_run_ctxt saved;
 270         struct niobuf_remote *rnb;
 271         struct niobuf_local *lnb;
 272         struct dentry *dentry = NULL;
 273         struct inode *inode;
 274         void *iobuf = NULL;
 275         int rc = 0, i, tot_bytes = 0;
 276         unsigned long now = jiffies;
 277         ENTRY;
 278
 279         /* We are currently not supporting multi-obj BRW_READ RPCS at all.
 280          * When we do this function's dentry cleanup will need to be fixed.
 281          * These values are verified in ost_brw_write() from the wire. */
 282         LASSERTF(objcount == 1, "%d\n", objcount);
 283         LASSERTF(obj->ioo_bufcnt > 0, "%d\n", obj->ioo_bufcnt);
 284
 285         rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa), capa,
 286                               CAPA_OPC_OSS_READ);
 287         if (rc)
 288                 RETURN(rc);
 289
 290         if (oa && oa->o_valid & OBD_MD_FLGRANT) {
 291                 spin_lock(&obd->obd_osfs_lock);
 292                 filter_grant_incoming(exp, oa);
 293
 294                 oa->o_grant = 0;
 295                 spin_unlock(&obd->obd_osfs_lock);
 296         }
 297
 298         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 299         if (IS_ERR(iobuf))
 300                 RETURN(PTR_ERR(iobuf));
 301
 302         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 303         dentry = filter_oa2dentry(obd, oa);
 304         if (IS_ERR(dentry)) {
 305                 rc = PTR_ERR(dentry);
 306                 dentry = NULL;
 307                 GOTO(cleanup, rc);
 308         }
 309
 310         inode = dentry->d_inode;
 311
 312         obdo_to_inode(inode, oa, OBD_MD_FLATIME);
 313         fsfilt_check_slow(obd, now, obd_timeout, "preprw_read setup");
 314
 315         for (i = 0, lnb = res, rnb = nb; i < obj->ioo_bufcnt;
 316              i++, rnb++, lnb++) {
 317                 lnb->dentry = dentry;
 318                 lnb->offset = rnb->offset;
 319                 lnb->len    = rnb->len;
 320                 lnb->flags  = rnb->flags;
 321
 322                 /*
 323                  * ost_brw_write()->ost_nio_pages_get() already initialized
 324                  * lnb->page to point to the page from the per-thread page
 325                  * pool (bug 5137), initialize page.
 326                  */
 327                 LASSERT(lnb->page != NULL);
 328
 329                 if (i_size_read(inode) <= rnb->offset)
 330                         /* If there's no more data, abort early.  lnb->rc == 0,
 331                          * so it's easy to detect later. */
 332                         break;
 333                 else
 334                         filter_alloc_dio_page(obd, inode, lnb);
 335
 336                 if (i_size_read(inode) < lnb->offset + lnb->len - 1)
 337                         lnb->rc = i_size_read(inode) - lnb->offset;
 338                 else
 339                         lnb->rc = lnb->len;
 340
 341                 tot_bytes += lnb->rc;
 342
 343                 filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 344         }
 345
 346         fsfilt_check_slow(obd, now, obd_timeout, "start_page_read");
 347
 348         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf,
 349                               exp, NULL, NULL, NULL);
 350         if (rc)
 351                 GOTO(cleanup, rc);
 352
 353         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_READ_BYTES, tot_bytes);
 354
 355         lprocfs_counter_add(exp->exp_ops_stats, LPROC_FILTER_READ_BYTES,
 356                             tot_bytes);
 357
 358         EXIT;
 359
 360  cleanup:
 361         if (rc != 0) {
 362                 filter_free_dio_pages(objcount, obj, niocount, res);
 363
 364                 if (dentry != NULL)
 365                         f_dput(dentry);
 366         }
 367
 368         filter_iobuf_put(&obd->u.filter, iobuf, oti);
 369
 370         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 371         if (rc)
 372                 CERROR("io error %d\n", rc);
 373
 374         return rc;
 375 }
 376
 377 /* When clients have dirtied as much space as they've been granted they
 378  * fall through to sync writes.  These sync writes haven't been expressed
 379  * in grants and need to error with ENOSPC when there isn't room in the
 380  * filesystem for them after grants are taken into account.  However,
 381  * writeback of the dirty data that was already granted space can write
 382  * right on through.
 383  *
 384  * Caller must hold obd_osfs_lock. */
 385 static int filter_grant_check(struct obd_export *exp, struct obdo *oa,
 386                               int objcount, struct fsfilt_objinfo *fso,
 387                               int niocount, struct niobuf_remote *rnb,
 388                               struct niobuf_local *lnb, obd_size *left,
 389                               struct inode *inode)
 390 {
 391         struct filter_export_data *fed = &exp->exp_filter_data;
 392         int blocksize = exp->exp_obd->u.obt.obt_sb->s_blocksize;
 393         unsigned long used = 0, ungranted = 0, using;
 394         int i, rc = -ENOSPC, obj, n = 0;
 395
 396         LASSERT_SPIN_LOCKED(&exp->exp_obd->obd_osfs_lock);
 397
 398         for (obj = 0; obj < objcount; obj++) {
 399                 for (i = 0; i < fso[obj].fso_bufcnt; i++, n++) {
 400                         int tmp, bytes;
 401
 402                         /* should match the code in osc_exit_cache */
 403                         bytes = rnb[n].len;
 404                         bytes += rnb[n].offset & (blocksize - 1);
 405                         tmp = (rnb[n].offset + rnb[n].len) & (blocksize - 1);
 406                         if (tmp)
 407                                 bytes += blocksize - tmp;
 408
 409                         if ((rnb[n].flags & OBD_BRW_FROM_GRANT) &&
 410                             (oa->o_valid & OBD_MD_FLGRANT)) {
 411                                 if (fed->fed_grant < used + bytes) {
 412                                         CDEBUG(D_CACHE,
 413                                                "%s: cli %s/%p claims %ld+%d "
 414                                                "GRANT, real grant %lu idx %d\n",
 415                                                exp->exp_obd->obd_name,
 416                                                exp->exp_client_uuid.uuid, exp,
 417                                                used, bytes, fed->fed_grant, n);
 418                                 } else {
 419                                         used += bytes;
 420                                         rnb[n].flags |= OBD_BRW_GRANTED;
 421                                         lnb[n].lnb_grant_used = bytes;
 422                                         CDEBUG(0, "idx %d used=%lu\n", n, used);
 423                                         rc = 0;
 424                                         continue;
 425                                 }
 426                         }
 427                         if (*left > ungranted + bytes) {
 428                                 /* if enough space, pretend it was granted */
 429                                 ungranted += bytes;
 430                                 rnb[n].flags |= OBD_BRW_GRANTED;
 431                                 lnb[n].lnb_grant_used = bytes;
 432                                 CDEBUG(0, "idx %d ungranted=%lu\n",n,ungranted);
 433                                 rc = 0;
 434                                 continue;
 435                         }
 436
 437                         /* We can't check for already-mapped blocks here, as
 438                          * it requires dropping the osfs lock to do the bmap.
 439                          * Instead, we return ENOSPC and in that case we need
 440                          * to go through and verify if all of the blocks not
 441                          * marked BRW_GRANTED are already mapped and we can
 442                          * ignore this error. */
 443                         lnb[n].rc = -ENOSPC;
 444                         rnb[n].flags &= ~OBD_BRW_GRANTED;
 445                         CDEBUG(D_CACHE,"%s: cli %s/%p idx %d no space for %d\n",
 446                                exp->exp_obd->obd_name,
 447                                exp->exp_client_uuid.uuid, exp, n, bytes);
 448                 }
 449         }
 450
 451         /* Now substract what client have used already.  We don't subtract
 452          * this from the tot_granted yet, so that other client's can't grab
 453          * that space before we have actually allocated our blocks.  That
 454          * happens in filter_grant_commit() after the writes are done. */
 455         *left -= ungranted;
 456         fed->fed_grant -= used;
 457         fed->fed_pending += used + ungranted;
 458         exp->exp_obd->u.filter.fo_tot_granted += ungranted;
 459         exp->exp_obd->u.filter.fo_tot_pending += used + ungranted;
 460
 461         CDEBUG(D_CACHE,
 462                "%s: cli %s/%p used: %lu ungranted: %lu grant: %lu dirty: %lu\n",
 463                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp, used,
 464                ungranted, fed->fed_grant, fed->fed_dirty);
 465
 466         /* Rough calc in case we don't refresh cached statfs data */
 467         using = (used + ungranted + 1 ) >>
 468                 exp->exp_obd->u.obt.obt_sb->s_blocksize_bits;
 469         if (exp->exp_obd->obd_osfs.os_bavail > using)
 470                 exp->exp_obd->obd_osfs.os_bavail -= using;
 471         else
 472                 exp->exp_obd->obd_osfs.os_bavail = 0;
 473
 474         if (fed->fed_dirty < used) {
 475                 CERROR("%s: cli %s/%p claims used %lu > fed_dirty %lu\n",
 476                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 477                        used, fed->fed_dirty);
 478                 used = fed->fed_dirty;
 479         }
 480         exp->exp_obd->u.filter.fo_tot_dirty -= used;
 481         fed->fed_dirty -= used;
 482
 483         if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
 484                 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
 485                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 486                        fed->fed_dirty, fed->fed_pending, fed->fed_grant);
 487                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
 488                 LBUG();
 489         }
 490         return rc;
 491 }
 492
 493 /* If we ever start to support multi-object BRW RPCs, we will need to get locks
 494  * on mulitple inodes.  That isn't all, because there still exists the
 495  * possibility of a truncate starting a new transaction while holding the ext3
 496  * rwsem = write while some writes (which have started their transactions here)
 497  * blocking on the ext3 rwsem = read => lock inversion.
 498  *
 499  * The handling gets very ugly when dealing with locked pages.  It may be easier
 500  * to just get rid of the locked page code (which has problems of its own) and
 501  * either discover we do not need it anymore (i.e. it was a symptom of another
 502  * bug) or ensure we get the page locks in an appropriate order. */
 503 static int filter_preprw_write(int cmd, struct obd_export *exp, struct obdo *oa,
 504                                int objcount, struct obd_ioobj *obj,
 505                                int niocount, struct niobuf_remote *nb,
 506                                struct niobuf_local *res,
 507                                struct obd_trans_info *oti,
 508                                struct lustre_capa *capa)
 509 {
 510         struct lvfs_run_ctxt saved;
 511         struct niobuf_remote *rnb;
 512         struct niobuf_local *lnb = res;
 513         struct fsfilt_objinfo fso;
 514         struct filter_mod_data *fmd;
 515         struct dentry *dentry = NULL;
 516         void *iobuf;
 517         obd_size left;
 518         unsigned long now = jiffies;
 519         int rc = 0, i, tot_bytes = 0, cleanup_phase = 0;
 520         ENTRY;
 521         LASSERT(objcount == 1);
 522         LASSERT(obj->ioo_bufcnt > 0);
 523
 524         rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa), capa,
 525                               CAPA_OPC_OSS_WRITE);
 526         if (rc)
 527                 RETURN(rc);
 528
 529         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 530         iobuf = filter_iobuf_get(&exp->exp_obd->u.filter, oti);
 531         if (IS_ERR(iobuf))
 532                 GOTO(cleanup, rc = PTR_ERR(iobuf));
 533         cleanup_phase = 1;
 534
 535         dentry = filter_fid2dentry(exp->exp_obd, NULL, obj->ioo_gr,
 536                                    obj->ioo_id);
 537         if (IS_ERR(dentry))
 538                 GOTO(cleanup, rc = PTR_ERR(dentry));
 539         cleanup_phase = 2;
 540
 541         if (dentry->d_inode == NULL) {
 542                 CERROR("%s: trying to BRW to non-existent file "LPU64"\n",
 543                        exp->exp_obd->obd_name, obj->ioo_id);
 544                 GOTO(cleanup, rc = -ENOENT);
 545         }
 546
 547         fso.fso_dentry = dentry;
 548         fso.fso_bufcnt = obj->ioo_bufcnt;
 549
 550         fsfilt_check_slow(exp->exp_obd, now, obd_timeout, "preprw_write setup");
 551
 552         /* Don't update inode timestamps if this write is older than a
 553          * setattr which modifies the timestamps. b=10150 */
 554         /* XXX when we start having persistent reservations this needs to
 555          * be changed to filter_fmd_get() to create the fmd if it doesn't
 556          * already exist so we can store the reservation handle there. */
 557         fmd = filter_fmd_find(exp, obj->ioo_id, obj->ioo_gr);
 558
 559         LASSERT(oa != NULL);
 560         spin_lock(&exp->exp_obd->obd_osfs_lock);
 561         filter_grant_incoming(exp, oa);
 562         if (fmd && fmd->fmd_mactime_xid > oti->oti_xid)
 563                 oa->o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 564                                  OBD_MD_FLATIME);
 565         else
 566                 obdo_to_inode(dentry->d_inode, oa, OBD_MD_FLATIME |
 567                               OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 568         cleanup_phase = 3;
 569
 570         left = filter_grant_space_left(exp);
 571
 572         rc = filter_grant_check(exp, oa, objcount, &fso, niocount, nb, res,
 573                                 &left, dentry->d_inode);
 574
 575         /* do not zero out oa->o_valid as it is used in filter_commitrw_write()
 576          * for setting UID/GID and fid EA in first write time. */
 577         if (oa->o_valid & OBD_MD_FLGRANT)
 578                 oa->o_grant = filter_grant(exp,oa->o_grant,oa->o_undirty,left);
 579
 580         spin_unlock(&exp->exp_obd->obd_osfs_lock);
 581         filter_fmd_put(exp, fmd);
 582
 583         if (rc)
 584                 GOTO(cleanup, rc);
 585
 586         for (i = 0, rnb = nb, lnb = res; i < obj->ioo_bufcnt;
 587              i++, lnb++, rnb++) {
 588                 /* We still set up for ungranted pages so that granted pages
 589                  * can be written to disk as they were promised, and portals
 590                  * needs to keep the pages all aligned properly. */
 591                 lnb->dentry = dentry;
 592                 lnb->offset = rnb->offset;
 593                 lnb->len    = rnb->len;
 594                 lnb->flags  = rnb->flags;
 595
 596                 /*
 597                  * ost_brw_write()->ost_nio_pages_get() already initialized
 598                  * lnb->page to point to the page from the per-thread page
 599                  * pool (bug 5137), initialize page.
 600                  */
 601                 LASSERT(lnb->page != NULL);
 602                 if (lnb->len != CFS_PAGE_SIZE) {
 603                         memset(kmap(lnb->page) + lnb->len,
 604                                0, CFS_PAGE_SIZE - lnb->len);
 605                         kunmap(lnb->page);
 606                 }
 607                 lnb->page->index = lnb->offset >> CFS_PAGE_SHIFT;
 608
 609                 cleanup_phase = 4;
 610
 611                 /* If the filter writes a partial page, then has the file
 612                  * extended, the client will read in the whole page.  the
 613                  * filter has to be careful to zero the rest of the partial
 614                  * page on disk.  we do it by hand for partial extending
 615                  * writes, send_bio() is responsible for zeroing pages when
 616                  * asked to read unmapped blocks -- brw_kiovec() does this. */
 617                 if (lnb->len != CFS_PAGE_SIZE) {
 618                         __s64 maxidx;
 619
 620                         maxidx = ((i_size_read(dentry->d_inode) +
 621                                    CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT) - 1;
 622                         if (maxidx >= lnb->page->index) {
 623                                 LL_CDEBUG_PAGE(D_PAGE, lnb->page, "write %u @ "
 624                                                LPU64" flg %x before EOF %llu\n",
 625                                                lnb->len, lnb->offset,lnb->flags,
 626                                                i_size_read(dentry->d_inode));
 627                                 filter_iobuf_add_page(exp->exp_obd, iobuf,
 628                                                       dentry->d_inode,
 629                                                       lnb->page);
 630                         } else {
 631                                 long off;
 632                                 char *p = kmap(lnb->page);
 633
 634                                 off = lnb->offset & ~CFS_PAGE_MASK;
 635                                 if (off)
 636                                         memset(p, 0, off);
 637                                 off = (lnb->offset + lnb->len) & ~CFS_PAGE_MASK;
 638                                 if (off)
 639                                         memset(p + off, 0, CFS_PAGE_SIZE - off);
 640                                 kunmap(lnb->page);
 641                         }
 642                 }
 643                 if (lnb->rc == 0)
 644                         tot_bytes += lnb->len;
 645         }
 646
 647         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf, exp,
 648                               NULL, NULL, NULL);
 649
 650         fsfilt_check_slow(exp->exp_obd, now, obd_timeout, "start_page_write");
 651
 652         lprocfs_counter_add(exp->exp_ops_stats, LPROC_FILTER_WRITE_BYTES,
 653                             tot_bytes);
 654         EXIT;
 655 cleanup:
 656         switch(cleanup_phase) {
 657         case 4:
 658         case 3:
 659                 filter_iobuf_put(&exp->exp_obd->u.filter, iobuf, oti);
 660         case 2:
 661                 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 662                 if (rc)
 663                         f_dput(dentry);
 664                 break;
 665         case 1:
 666                 filter_iobuf_put(&exp->exp_obd->u.filter, iobuf, oti);
 667         case 0:
 668                 spin_lock(&exp->exp_obd->obd_osfs_lock);
 669                 if (oa)
 670                         filter_grant_incoming(exp, oa);
 671                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
 672                 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 673                 break;
 674         default:;
 675         }
 676         return rc;
 677 }
 678
 679 int filter_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
 680                   int objcount, struct obd_ioobj *obj, int niocount,
 681                   struct niobuf_remote *nb, struct niobuf_local *res,
 682                   struct obd_trans_info *oti, struct lustre_capa *capa)
 683 {
 684         if (cmd == OBD_BRW_WRITE)
 685                 return filter_preprw_write(cmd, exp, oa, objcount, obj,
 686                                            niocount, nb, res, oti, capa);
 687         if (cmd == OBD_BRW_READ)
 688                 return filter_preprw_read(cmd, exp, oa, objcount, obj,
 689                                           niocount, nb, res, oti, capa);
 690         LBUG();
 691         return -EPROTO;
 692 }
 693
 694 void filter_release_read_page(struct filter_obd *filter, struct inode *inode,
 695                               struct page *page)
 696 {
 697         int drop = 0;
 698
 699         if (inode != NULL &&
 700             (i_size_read(inode) > filter->fo_readcache_max_filesize))
 701                 drop = 1;
 702
 703         /* drop from cache like truncate_list_pages() */
 704         if (drop && !TryLockPage(page)) {
 705                 if (page->mapping)
 706                         ll_truncate_complete_page(page);
 707                 unlock_page(page);
 708         }
 709         page_cache_release(page);
 710 }
 711
 712 static int filter_commitrw_read(struct obd_export *exp, struct obdo *oa,
 713                                 int objcount, struct obd_ioobj *obj,
 714                                 int niocount, struct niobuf_local *res,
 715                                 struct obd_trans_info *oti, int rc)
 716 {
 717         struct inode *inode = NULL;
 718         struct ldlm_res_id res_id = { .name = { obj->ioo_id, 0,
 719                                                 obj->ioo_gr, 0} };
 720         struct ldlm_resource *resource = NULL;
 721         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 722         ENTRY;
 723
 724         /* If oa != NULL then filter_preprw_read updated the inode atime
 725          * and we should update the lvb so that other glimpses will also
 726          * get the updated value. bug 5972 */
 727         if (oa && ns && ns->ns_lvbo && ns->ns_lvbo->lvbo_update) {
 728                 resource = ldlm_resource_get(ns, NULL, &res_id, LDLM_EXTENT, 0);
 729
 730                 if (resource != NULL) {
 731                         ns->ns_lvbo->lvbo_update(resource, NULL, 0, 1);
 732                         ldlm_resource_putref(resource);
 733                 }
 734         }
 735
 736         if (res->dentry != NULL)
 737                 inode = res->dentry->d_inode;
 738
 739         filter_free_dio_pages(objcount, obj, niocount, res);
 740
 741         if (res->dentry != NULL)
 742                 f_dput(res->dentry);
 743         RETURN(rc);
 744 }
 745
 746 void flip_into_page_cache(struct inode *inode, struct page *new_page)
 747 {
 748         struct page *old_page;
 749         int rc;
 750
 751         do {
 752                 /* the dlm is protecting us from read/write concurrency, so we
 753                  * expect this find_lock_page to return quickly.  even if we
 754                  * race with another writer it won't be doing much work with
 755                  * the page locked.  we do this 'cause t_c_p expects a
 756                  * locked page, and it wants to grab the pagecache lock
 757                  * as well. */
 758                 old_page = find_lock_page(inode->i_mapping, new_page->index);
 759                 if (old_page) {
 760                         ll_truncate_complete_page(old_page);
 761                         unlock_page(old_page);
 762                         page_cache_release(old_page);
 763                 }
 764
 765 #if 0 /* this should be a /proc tunable someday */
 766                 /* racing o_directs (no locking ioctl) could race adding
 767                  * their pages, so we repeat the page invalidation unless
 768                  * we successfully added our new page */
 769                 rc = add_to_page_cache_unique(new_page, inode->i_mapping,
 770                                               new_page->index,
 771                                               page_hash(inode->i_mapping,
 772                                                         new_page->index));
 773                 if (rc == 0) {
 774                         /* add_to_page_cache clears uptodate|dirty and locks
 775                          * the page */
 776                         SetPageUptodate(new_page);
 777                         unlock_page(new_page);
 778                 }
 779 #else
 780                 rc = 0;
 781 #endif
 782         } while (rc != 0);
 783 }
 784
 785 void filter_grant_commit(struct obd_export *exp, int niocount,
 786                          struct niobuf_local *res)
 787 {
 788         struct filter_obd *filter = &exp->exp_obd->u.filter;
 789         struct niobuf_local *lnb = res;
 790         unsigned long pending = 0;
 791         int i;
 792
 793         spin_lock(&exp->exp_obd->obd_osfs_lock);
 794         for (i = 0, lnb = res; i < niocount; i++, lnb++)
 795                 pending += lnb->lnb_grant_used;
 796
 797         LASSERTF(exp->exp_filter_data.fed_pending >= pending,
 798                  "%s: cli %s/%p fed_pending: %lu grant_used: %lu\n",
 799                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 800                  exp->exp_filter_data.fed_pending, pending);
 801         exp->exp_filter_data.fed_pending -= pending;
 802         LASSERTF(filter->fo_tot_granted >= pending,
 803                  "%s: cli %s/%p tot_granted: "LPU64" grant_used: %lu\n",
 804                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 805                  exp->exp_obd->u.filter.fo_tot_granted, pending);
 806         filter->fo_tot_granted -= pending;
 807         LASSERTF(filter->fo_tot_pending >= pending,
 808                  "%s: cli %s/%p tot_pending: "LPU64" grant_used: %lu\n",
 809                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 810                  filter->fo_tot_pending, pending);
 811         filter->fo_tot_pending -= pending;
 812
 813         spin_unlock(&exp->exp_obd->obd_osfs_lock);
 814 }
 815
 816 int filter_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
 817                     int objcount, struct obd_ioobj *obj, int niocount,
 818                     struct niobuf_local *res, struct obd_trans_info *oti,
 819                     int rc)
 820 {
 821         if (cmd == OBD_BRW_WRITE)
 822                 return filter_commitrw_write(exp, oa, objcount, obj, niocount,
 823                                              res, oti, rc);
 824         if (cmd == OBD_BRW_READ)
 825                 return filter_commitrw_read(exp, oa, objcount, obj, niocount,
 826                                             res, oti, rc);
 827         LBUG();
 828         return -EPROTO;
 829 }
 830
 831 int filter_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
 832                obd_count oa_bufs, struct brw_page *pga,
 833                struct obd_trans_info *oti)
 834 {
 835         struct obd_ioobj ioo;
 836         struct niobuf_local *lnb;
 837         struct niobuf_remote *rnb;
 838         obd_count i;
 839         int ret = 0;
 840         ENTRY;
 841
 842         OBD_ALLOC(lnb, oa_bufs * sizeof(struct niobuf_local));
 843         OBD_ALLOC(rnb, oa_bufs * sizeof(struct niobuf_remote));
 844
 845         if (lnb == NULL || rnb == NULL)
 846                 GOTO(out, ret = -ENOMEM);
 847
 848         for (i = 0; i < oa_bufs; i++) {
 849                 lnb[i].page = pga[i].pg;
 850                 rnb[i].offset = pga[i].off;
 851                 rnb[i].len = pga[i].count;
 852         }
 853
 854         obdo_to_ioobj(oinfo->oi_oa, &ioo);
 855         ioo.ioo_bufcnt = oa_bufs;
 856
 857         ret = filter_preprw(cmd, exp, oinfo->oi_oa, 1, &ioo,
 858                             oa_bufs, rnb, lnb, oti, oinfo_capa(oinfo));
 859         if (ret != 0)
 860                 GOTO(out, ret);
 861
 862         ret = filter_commitrw(cmd, exp, oinfo->oi_oa, 1, &ioo,
 863                               oa_bufs, lnb, oti, ret);
 864
 865 out:
 866         if (lnb)
 867                 OBD_FREE(lnb, oa_bufs * sizeof(struct niobuf_local));
 868         if (rnb)
 869                 OBD_FREE(rnb, oa_bufs * sizeof(struct niobuf_remote));
 870         RETURN(ret);
 871 }