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 #ifdef HAVE_KERNEL_CONFIG_H
  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         static unsigned long last_msg;
  84         static int last_count;
  85         int mask = D_CACHE;
  86         ENTRY;
  87
  88         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
  89
  90         if ((oa->o_valid & (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) !=
  91                                         (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) {
  92                 oa->o_valid &= ~OBD_MD_FLGRANT;
  93                 EXIT;
  94                 return;
  95         }
  96
  97         fed = &exp->exp_filter_data;
  98
  99         /* Don't print this to the console the first time it happens, since
 100          * it can happen legitimately on occasion, but only rarely. */
 101         if (time_after(jiffies, last_msg + 60 * HZ)) {
 102                 last_count = 0;
 103                 last_msg = jiffies;
 104         }
 105         if ((last_count & (-last_count)) == last_count)
 106                 mask = D_HA /* until bug 3273 is fixed D_WARNING */;
 107         last_count++;
 108
 109         /* Add some margin, since there is a small race if other RPCs arrive
 110          * out-or-order and have already consumed some grant.  We want to
 111          * leave this here in case there is a large error in accounting. */
 112         CDEBUG(oa->o_grant > fed->fed_grant + FILTER_GRANT_CHUNK ? mask:D_CACHE,
 113                "%s: cli %s/%p reports grant: "LPU64" dropped: %u, local: %lu\n",
 114                obd->obd_name, exp->exp_client_uuid.uuid, exp, oa->o_grant,
 115                oa->o_dropped, fed->fed_grant);
 116
 117         /* Update our accounting now so that statfs takes it into account.
 118          * Note that fed_dirty is only approximate and can become incorrect
 119          * if RPCs arrive out-of-order.  No important calculations depend
 120          * on fed_dirty however, but we must check sanity to not assert. */
 121         if ((long long)oa->o_dirty < 0)
 122                 oa->o_dirty = 0;
 123         else if (oa->o_dirty > fed->fed_grant + 4 * FILTER_GRANT_CHUNK)
 124                 oa->o_dirty = fed->fed_grant + 4 * FILTER_GRANT_CHUNK;
 125         obd->u.filter.fo_tot_dirty += oa->o_dirty - fed->fed_dirty;
 126         if (fed->fed_grant < oa->o_dropped) {
 127                 CDEBUG(D_HA,"%s: cli %s/%p reports %u dropped > fedgrant %lu\n",
 128                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
 129                        oa->o_dropped, fed->fed_grant);
 130                 oa->o_dropped = 0;
 131         }
 132         if (obd->u.filter.fo_tot_granted < oa->o_dropped) {
 133                 CERROR("%s: cli %s/%p reports %u dropped > tot_grant "LPU64"\n",
 134                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
 135                        oa->o_dropped, obd->u.filter.fo_tot_granted);
 136                 oa->o_dropped = 0;
 137         }
 138         obd->u.filter.fo_tot_granted -= oa->o_dropped;
 139         fed->fed_grant -= oa->o_dropped;
 140         fed->fed_dirty = oa->o_dirty;
 141         if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
 142                 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
 143                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
 144                        fed->fed_dirty, fed->fed_pending, fed->fed_grant);
 145                 spin_unlock(&obd->obd_osfs_lock);
 146                 LBUG();
 147         }
 148         EXIT;
 149 }
 150
 151 /* Figure out how much space is available between what we've granted
 152  * and what remains in the filesystem.  Compensate for ext3 indirect
 153  * block overhead when computing how much free space is left ungranted.
 154  *
 155  * Caller must hold obd_osfs_lock. */
 156 obd_size filter_grant_space_left(struct obd_export *exp)
 157 {
 158         struct obd_device *obd = exp->exp_obd;
 159         int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
 160         obd_size tot_granted = obd->u.filter.fo_tot_granted, avail, left = 0;
 161         int rc, statfs_done = 0;
 162
 163         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
 164
 165         if (cfs_time_before_64(obd->obd_osfs_age, cfs_time_current_64() - HZ)) {
 166 restat:
 167                 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb,
 168                                    cfs_time_current_64() + HZ);
 169                 if (rc) /* N.B. statfs can't really fail */
 170                         RETURN(0);
 171                 statfs_done = 1;
 172         }
 173
 174         avail = obd->obd_osfs.os_bavail;
 175         left = avail - (avail >> (blockbits - 3)); /* (d)indirect */
 176         if (left > GRANT_FOR_LLOG(obd)) {
 177                 left = (left - GRANT_FOR_LLOG(obd)) << blockbits;
 178         } else {
 179                 left = 0 /* << blockbits */;
 180         }
 181
 182         if (!statfs_done && left < 32 * FILTER_GRANT_CHUNK + tot_granted) {
 183                 CDEBUG(D_CACHE, "fs has no space left and statfs too old\n");
 184                 goto restat;
 185         }
 186
 187         if (left >= tot_granted) {
 188                 left -= tot_granted;
 189         } else {
 190                 if (left < tot_granted - obd->u.filter.fo_tot_pending) {
 191                         CERROR("%s: cli %s/%p grant "LPU64" > available "
 192                                LPU64" and pending "LPU64"\n", obd->obd_name,
 193                                exp->exp_client_uuid.uuid, exp, tot_granted,
 194                                left, obd->u.filter.fo_tot_pending);
 195                 }
 196                 left = 0;
 197         }
 198
 199         CDEBUG(D_CACHE, "%s: cli %s/%p free: "LPU64" avail: "LPU64" grant "LPU64
 200                " left: "LPU64" pending: "LPU64"\n", obd->obd_name,
 201                exp->exp_client_uuid.uuid, exp,
 202                obd->obd_osfs.os_bfree << blockbits, avail << blockbits,
 203                tot_granted, left, obd->u.filter.fo_tot_pending);
 204
 205         return left;
 206 }
 207
 208 /* Calculate how much grant space to allocate to this client, based on how
 209  * much space is currently free and how much of that is already granted.
 210  *
 211  * Caller must hold obd_osfs_lock. */
 212 long filter_grant(struct obd_export *exp, obd_size current_grant,
 213                   obd_size want, obd_size fs_space_left)
 214 {
 215         struct obd_device *obd = exp->exp_obd;
 216         struct filter_export_data *fed = &exp->exp_filter_data;
 217         int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
 218         __u64 grant = 0;
 219
 220         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
 221
 222         /* Grant some fraction of the client's requested grant space so that
 223          * they are not always waiting for write credits (not all of it to
 224          * avoid overgranting in face of multiple RPCs in flight).  This
 225          * essentially will be able to control the OSC_MAX_RIF for a client.
 226          *
 227          * If we do have a large disparity between what the client thinks it
 228          * has and what we think it has, don't grant very much and let the
 229          * client consume its grant first.  Either it just has lots of RPCs
 230          * in flight, or it was evicted and its grants will soon be used up. */
 231         if (want > 0x7fffffff) {
 232                 CERROR("%s: client %s/%p requesting > 2GB grant "LPU64"\n",
 233                        obd->obd_name, exp->exp_client_uuid.uuid, exp, want);
 234         } else if (current_grant < want &&
 235                    current_grant < fed->fed_grant + FILTER_GRANT_CHUNK) {
 236                 grant = min((want >> blockbits),
 237                             (fs_space_left >> blockbits) / 8);
 238                 grant <<= blockbits;
 239
 240                 if (grant) {
 241                         /* Allow >FILTER_GRANT_CHUNK size when clients
 242                          * reconnect due to a server reboot.
 243                          */
 244                         if ((grant > FILTER_GRANT_CHUNK) &&
 245                             (!obd->obd_recovering))
 246                                 grant = FILTER_GRANT_CHUNK;
 247
 248                         obd->u.filter.fo_tot_granted += grant;
 249                         fed->fed_grant += grant;
 250                         if (fed->fed_grant < 0) {
 251                                 CERROR("%s: cli %s/%p grant %ld want "LPU64
 252                                        "current"LPU64"\n",
 253                                        obd->obd_name, exp->exp_client_uuid.uuid,
 254                                        exp, fed->fed_grant, want,current_grant);
 255                                 spin_unlock(&obd->obd_osfs_lock);
 256                                 LBUG();
 257                         }
 258                 }
 259         }
 260
 261         CDEBUG(D_CACHE,"%s: cli %s/%p wants: "LPU64" granting: "LPU64"\n",
 262                obd->obd_name, exp->exp_client_uuid.uuid, exp, want, grant);
 263         CDEBUG(D_CACHE,
 264                "%s: cli %s/%p tot cached:"LPU64" granted:"LPU64
 265                " num_exports: %d\n", obd->obd_name, exp->exp_client_uuid.uuid,
 266                exp, obd->u.filter.fo_tot_dirty,
 267                obd->u.filter.fo_tot_granted, obd->obd_num_exports);
 268
 269         return grant;
 270 }
 271
 272 static int filter_preprw_read(int cmd, struct obd_export *exp, struct obdo *oa,
 273                               int objcount, struct obd_ioobj *obj,
 274                               int niocount, struct niobuf_remote *nb,
 275                               struct niobuf_local *res,
 276                               struct obd_trans_info *oti)
 277 {
 278         struct obd_device *obd = exp->exp_obd;
 279         struct lvfs_run_ctxt saved;
 280         struct niobuf_remote *rnb;
 281         struct niobuf_local *lnb;
 282         struct dentry *dentry = NULL;
 283         struct inode *inode;
 284         void *iobuf = NULL;
 285         int rc = 0, i, tot_bytes = 0;
 286         unsigned long now = jiffies;
 287         ENTRY;
 288
 289         /* We are currently not supporting multi-obj BRW_READ RPCS at all.
 290          * When we do this function's dentry cleanup will need to be fixed.
 291          * These values are verified in ost_brw_write() from the wire. */
 292         LASSERTF(objcount == 1, "%d\n", objcount);
 293         LASSERTF(obj->ioo_bufcnt > 0, "%d\n", obj->ioo_bufcnt);
 294
 295         if (oa->o_valid & OBD_MD_FLGRANT) {
 296                 spin_lock(&obd->obd_osfs_lock);
 297                 filter_grant_incoming(exp, oa);
 298
 299                 oa->o_grant = 0;
 300                 spin_unlock(&obd->obd_osfs_lock);
 301         }
 302
 303         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 304         if (IS_ERR(iobuf))
 305                 RETURN(PTR_ERR(iobuf));
 306
 307         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 308         dentry = filter_oa2dentry(obd, oa);
 309         if (IS_ERR(dentry)) {
 310                 rc = PTR_ERR(dentry);
 311                 dentry = NULL;
 312                 GOTO(cleanup, rc);
 313         }
 314
 315         inode = dentry->d_inode;
 316
 317         obdo_to_inode(inode, oa, OBD_MD_FLATIME);
 318         fsfilt_check_slow(obd, now, obd_timeout, "preprw_read setup");
 319
 320         for (i = 0, lnb = res, rnb = nb; i < obj->ioo_bufcnt;
 321              i++, rnb++, lnb++) {
 322                 lnb->dentry = dentry;
 323                 lnb->offset = rnb->offset;
 324                 lnb->len    = rnb->len;
 325                 lnb->flags  = rnb->flags;
 326
 327                 /*
 328                  * ost_brw_write()->ost_nio_pages_get() already initialized
 329                  * lnb->page to point to the page from the per-thread page
 330                  * pool (bug 5137), initialize page.
 331                  */
 332                 LASSERT(lnb->page != NULL);
 333
 334                 if (inode->i_size <= rnb->offset)
 335                         /* If there's no more data, abort early.  lnb->rc == 0,
 336                          * so it's easy to detect later. */
 337                         break;
 338                 else
 339                         filter_alloc_dio_page(obd, inode, lnb);
 340
 341                 if (inode->i_size < lnb->offset + lnb->len - 1)
 342                         lnb->rc = inode->i_size - lnb->offset;
 343                 else
 344                         lnb->rc = lnb->len;
 345
 346                 tot_bytes += lnb->rc;
 347
 348                 filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 349         }
 350
 351         fsfilt_check_slow(obd, now, obd_timeout, "start_page_read");
 352
 353         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf,
 354                               exp, NULL, NULL, NULL);
 355         if (rc)
 356                 GOTO(cleanup, rc);
 357
 358         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_READ_BYTES, tot_bytes);
 359         lprocfs_counter_add(exp->exp_ops_stats, LPROC_FILTER_READ_BYTES,
 360                             tot_bytes);
 361
 362         EXIT;
 363
 364  cleanup:
 365         if (rc != 0) {
 366                 filter_free_dio_pages(objcount, obj, niocount, res);
 367
 368                 if (dentry != NULL)
 369                         f_dput(dentry);
 370         }
 371
 372         filter_iobuf_put(&obd->u.filter, iobuf, oti);
 373
 374         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 375         if (rc)
 376                 CERROR("io error %d\n", rc);
 377
 378         return rc;
 379 }
 380
 381 /* When clients have dirtied as much space as they've been granted they
 382  * fall through to sync writes.  These sync writes haven't been expressed
 383  * in grants and need to error with ENOSPC when there isn't room in the
 384  * filesystem for them after grants are taken into account.  However,
 385  * writeback of the dirty data that was already granted space can write
 386  * right on through.
 387  *
 388  * Caller must hold obd_osfs_lock. */
 389 static int filter_grant_check(struct obd_export *exp, int objcount,
 390                               struct fsfilt_objinfo *fso, int niocount,
 391                               struct niobuf_remote *rnb,
 392                               struct niobuf_local *lnb, obd_size *left,
 393                               struct inode *inode)
 394 {
 395         struct filter_export_data *fed = &exp->exp_filter_data;
 396         int blocksize = exp->exp_obd->u.obt.obt_sb->s_blocksize;
 397         unsigned long used = 0, ungranted = 0, using;
 398         int i, rc = -ENOSPC, obj, n = 0, mask = D_CACHE;
 399
 400         LASSERT_SPIN_LOCKED(&exp->exp_obd->obd_osfs_lock);
 401
 402         for (obj = 0; obj < objcount; obj++) {
 403                 for (i = 0; i < fso[obj].fso_bufcnt; i++, n++) {
 404                         int tmp, bytes;
 405
 406                         /* should match the code in osc_exit_cache */
 407                         bytes = rnb[n].len;
 408                         bytes += rnb[n].offset & (blocksize - 1);
 409                         tmp = (rnb[n].offset + rnb[n].len) & (blocksize - 1);
 410                         if (tmp)
 411                                 bytes += blocksize - tmp;
 412
 413                         if (rnb[n].flags & OBD_BRW_FROM_GRANT) {
 414                                 if (fed->fed_grant < used + bytes) {
 415                                         CDEBUG(D_CACHE,
 416                                                "%s: cli %s/%p claims %ld+%d "
 417                                                "GRANT, real grant %lu idx %d\n",
 418                                                exp->exp_obd->obd_name,
 419                                                exp->exp_client_uuid.uuid, exp,
 420                                                used, bytes, fed->fed_grant, n);
 421                                         mask = D_RPCTRACE;
 422                                 } else {
 423                                         used += bytes;
 424                                         rnb[n].flags |= OBD_BRW_GRANTED;
 425                                         lnb[n].lnb_grant_used = bytes;
 426                                         CDEBUG(0, "idx %d used=%lu\n", n, used);
 427                                         rc = 0;
 428                                         continue;
 429                                 }
 430                         }
 431                         if (*left > ungranted) {
 432                                 /* if enough space, pretend it was granted */
 433                                 ungranted += bytes;
 434                                 rnb[n].flags |= OBD_BRW_GRANTED;
 435                                 CDEBUG(0, "idx %d ungranted=%lu\n",n,ungranted);
 436                                 rc = 0;
 437                                 continue;
 438                         }
 439
 440                         /* We can't check for already-mapped blocks here, as
 441                          * it requires dropping the osfs lock to do the bmap.
 442                          * Instead, we return ENOSPC and in that case we need
 443                          * to go through and verify if all of the blocks not
 444                          * marked BRW_GRANTED are already mapped and we can
 445                          * ignore this error. */
 446                         lnb[n].rc = -ENOSPC;
 447                         rnb[n].flags &= ~OBD_BRW_GRANTED;
 448                         CDEBUG(D_CACHE,"%s: cli %s/%p idx %d no space for %d\n",
 449                                exp->exp_obd->obd_name,
 450                                exp->exp_client_uuid.uuid, exp, n, bytes);
 451                 }
 452         }
 453
 454         /* Now substract what client have used already.  We don't subtract
 455          * this from the tot_granted yet, so that other client's can't grab
 456          * that space before we have actually allocated our blocks.  That
 457          * happens in filter_grant_commit() after the writes are done. */
 458         *left -= ungranted;
 459         fed->fed_grant -= used;
 460         fed->fed_pending += used;
 461         exp->exp_obd->u.filter.fo_tot_pending += used;
 462
 463         CDEBUG(mask,
 464                "%s: cli %s/%p used: %lu ungranted: %lu grant: %lu dirty: %lu\n",
 465                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp, used,
 466                ungranted, fed->fed_grant, fed->fed_dirty);
 467
 468         /* Rough calc in case we don't refresh cached statfs data */
 469         using = (used + ungranted + 1 ) >>
 470                 exp->exp_obd->u.obt.obt_sb->s_blocksize_bits;
 471         if (exp->exp_obd->obd_osfs.os_bavail > using)
 472                 exp->exp_obd->obd_osfs.os_bavail -= using;
 473         else
 474                 exp->exp_obd->obd_osfs.os_bavail = 0;
 475
 476         if (fed->fed_dirty < used) {
 477                 CERROR("%s: cli %s/%p claims used %lu > fed_dirty %lu\n",
 478                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 479                        used, fed->fed_dirty);
 480                 used = fed->fed_dirty;
 481         }
 482         exp->exp_obd->u.filter.fo_tot_dirty -= used;
 483         fed->fed_dirty -= used;
 484
 485         if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
 486                 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
 487                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 488                        fed->fed_dirty, fed->fed_pending, fed->fed_grant);
 489                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
 490                 LBUG();
 491         }
 492         return rc;
 493 }
 494
 495 /* If we ever start to support multi-object BRW RPCs, we will need to get locks
 496  * on mulitple inodes.  That isn't all, because there still exists the
 497  * possibility of a truncate starting a new transaction while holding the ext3
 498  * rwsem = write while some writes (which have started their transactions here)
 499  * blocking on the ext3 rwsem = read => lock inversion.
 500  *
 501  * The handling gets very ugly when dealing with locked pages.  It may be easier
 502  * to just get rid of the locked page code (which has problems of its own) and
 503  * either discover we do not need it anymore (i.e. it was a symptom of another
 504  * bug) or ensure we get the page locks in an appropriate order. */
 505 static int filter_preprw_write(int cmd, struct obd_export *exp, struct obdo *oa,
 506                                int objcount, struct obd_ioobj *obj,
 507                                int niocount, struct niobuf_remote *nb,
 508                                struct niobuf_local *res,
 509                                struct obd_trans_info *oti)
 510 {
 511         struct lvfs_run_ctxt saved;
 512         struct niobuf_remote *rnb;
 513         struct niobuf_local *lnb = res;
 514         struct fsfilt_objinfo fso;
 515         struct filter_mod_data *fmd;
 516         struct dentry *dentry = NULL;
 517         void *iobuf;
 518         obd_size left;
 519         unsigned long now = jiffies;
 520         int rc = 0, i, tot_bytes = 0, cleanup_phase = 0;
 521         ENTRY;
 522         LASSERT(objcount == 1);
 523         LASSERT(obj->ioo_bufcnt > 0);
 524
 525         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 526         iobuf = filter_iobuf_get(&exp->exp_obd->u.filter, oti);
 527         if (IS_ERR(iobuf))
 528                 GOTO(cleanup, rc = PTR_ERR(iobuf));
 529         cleanup_phase = 1;
 530
 531         dentry = filter_fid2dentry(exp->exp_obd, NULL, obj->ioo_gr,
 532                                    obj->ioo_id);
 533         if (IS_ERR(dentry))
 534                 GOTO(cleanup, rc = PTR_ERR(dentry));
 535         cleanup_phase = 2;
 536
 537         if (dentry->d_inode == NULL) {
 538                 CERROR("%s: trying to BRW to non-existent file "LPU64"\n",
 539                        exp->exp_obd->obd_name, obj->ioo_id);
 540                 GOTO(cleanup, rc = -ENOENT);
 541         }
 542
 543         fso.fso_dentry = dentry;
 544         fso.fso_bufcnt = obj->ioo_bufcnt;
 545
 546         fsfilt_check_slow(exp->exp_obd, now, obd_timeout, "preprw_write setup");
 547
 548         /* Don't update inode timestamps if this write is older than a
 549          * setattr which modifies the timestamps. b=10150 */
 550         /* XXX when we start having persistent reservations this needs to
 551          * be changed to filter_fmd_get() to create the fmd if it doesn't
 552          * already exist so we can store the reservation handle there. */
 553         fmd = filter_fmd_find(exp, obj->ioo_id, obj->ioo_gr);
 554
 555         spin_lock(&exp->exp_obd->obd_osfs_lock);
 556         if (oa) {
 557                 filter_grant_incoming(exp, oa);
 558                 if (fmd && fmd->fmd_mactime_xid > oti->oti_xid)
 559                         oa->o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 560                                          OBD_MD_FLATIME);
 561                 else
 562                         obdo_to_inode(dentry->d_inode, oa, OBD_MD_FLATIME |
 563                                       OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 564         }
 565         cleanup_phase = 3;
 566
 567         left = filter_grant_space_left(exp);
 568
 569         rc = filter_grant_check(exp, objcount, &fso, niocount, nb, res,
 570                                 &left, dentry->d_inode);
 571
 572         /* do not zero out oa->o_valid as it is used in filter_commitrw_write()
 573          * for setting UID/GID and fid EA in first write time. */
 574         if (oa && oa->o_valid & OBD_MD_FLGRANT) {
 575                 oa->o_grant = filter_grant(exp,oa->o_grant,oa->o_undirty,left);
 576                 oa->o_valid |= OBD_MD_FLGRANT;
 577         }
 578
 579         spin_unlock(&exp->exp_obd->obd_osfs_lock);
 580         filter_fmd_put(exp, fmd);
 581
 582         if (rc)
 583                 GOTO(cleanup, rc);
 584
 585         for (i = 0, rnb = nb, lnb = res; i < obj->ioo_bufcnt;
 586              i++, lnb++, rnb++) {
 587                 /* We still set up for ungranted pages so that granted pages
 588                  * can be written to disk as they were promised, and portals
 589                  * needs to keep the pages all aligned properly. */
 590                 lnb->dentry = dentry;
 591                 lnb->offset = rnb->offset;
 592                 lnb->len    = rnb->len;
 593                 lnb->flags  = rnb->flags;
 594
 595                 /*
 596                  * ost_brw_write()->ost_nio_pages_get() already initialized
 597                  * lnb->page to point to the page from the per-thread page
 598                  * pool (bug 5137), initialize page.
 599                  */
 600                 LASSERT(lnb->page != NULL);
 601                 if (lnb->len != CFS_PAGE_SIZE) {
 602                         memset(kmap(lnb->page) + lnb->len,
 603                                0, CFS_PAGE_SIZE - lnb->len);
 604                         kunmap(lnb->page);
 605                 }
 606                 lnb->page->index = lnb->offset >> CFS_PAGE_SHIFT;
 607
 608                 cleanup_phase = 4;
 609
 610                 /* If the filter writes a partial page, then has the file
 611                  * extended, the client will read in the whole page.  the
 612                  * filter has to be careful to zero the rest of the partial
 613                  * page on disk.  we do it by hand for partial extending
 614                  * writes, send_bio() is responsible for zeroing pages when
 615                  * asked to read unmapped blocks -- brw_kiovec() does this. */
 616                 if (lnb->len != CFS_PAGE_SIZE) {
 617                         __s64 maxidx;
 618
 619                         maxidx = ((dentry->d_inode->i_size + CFS_PAGE_SIZE - 1) >>
 620                                  CFS_PAGE_SHIFT) - 1;
 621                         if (maxidx >= lnb->page->index) {
 622                                 LL_CDEBUG_PAGE(D_PAGE, lnb->page, "write %u @ "
 623                                                LPU64" flg %x before EOF %llu\n",
 624                                                lnb->len, lnb->offset,lnb->flags,
 625                                                dentry->d_inode->i_size);
 626                                 filter_iobuf_add_page(exp->exp_obd, iobuf,
 627                                                       dentry->d_inode,
 628                                                       lnb->page);
 629                         } else {
 630                                 long off;
 631                                 char *p = kmap(lnb->page);
 632
 633                                 off = lnb->offset & ~CFS_PAGE_MASK;
 634                                 if (off)
 635                                         memset(p, 0, off);
 636                                 off = (lnb->offset + lnb->len) & ~CFS_PAGE_MASK;
 637                                 if (off)
 638                                         memset(p + off, 0, CFS_PAGE_SIZE - off);
 639                                 kunmap(lnb->page);
 640                         }
 641                 }
 642                 if (lnb->rc == 0)
 643                         tot_bytes += lnb->len;
 644         }
 645
 646         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf, exp,
 647                               NULL, NULL, NULL);
 648
 649         fsfilt_check_slow(exp->exp_obd, now, obd_timeout, "start_page_write");
 650
 651         lprocfs_counter_add(exp->exp_obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
 652                             tot_bytes);
 653         lprocfs_counter_add(exp->exp_ops_stats, LPROC_FILTER_WRITE_BYTES,
 654                             tot_bytes);
 655         EXIT;
 656 cleanup:
 657         switch(cleanup_phase) {
 658         case 4:
 659         case 3:
 660                 filter_iobuf_put(&exp->exp_obd->u.filter, iobuf, oti);
 661         case 2:
 662                 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 663                 if (rc)
 664                         f_dput(dentry);
 665                 break;
 666         case 1:
 667                 filter_iobuf_put(&exp->exp_obd->u.filter, iobuf, oti);
 668         case 0:
 669                 spin_lock(&exp->exp_obd->obd_osfs_lock);
 670                 if (oa)
 671                         filter_grant_incoming(exp, oa);
 672                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
 673                 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 674                 break;
 675         default:;
 676         }
 677         return rc;
 678 }
 679
 680 int filter_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
 681                   int objcount, struct obd_ioobj *obj, int niocount,
 682                   struct niobuf_remote *nb, struct niobuf_local *res,
 683                   struct obd_trans_info *oti)
 684 {
 685         if (cmd == OBD_BRW_WRITE)
 686                 return filter_preprw_write(cmd, exp, oa, objcount, obj,
 687                                            niocount, nb, res, oti);
 688         if (cmd == OBD_BRW_READ)
 689                 return filter_preprw_read(cmd, exp, oa, objcount, obj,
 690                                           niocount, nb, res, oti);
 691         LBUG();
 692         return -EPROTO;
 693 }
 694
 695 void filter_release_read_page(struct filter_obd *filter, struct inode *inode,
 696                               struct page *page)
 697 {
 698         int drop = 0;
 699
 700         if (inode != NULL &&
 701             (inode->i_size > filter->fo_readcache_max_filesize))
 702                 drop = 1;
 703
 704         /* drop from cache like truncate_list_pages() */
 705         if (drop && !TryLockPage(page)) {
 706                 if (page->mapping)
 707                         ll_truncate_complete_page(page);
 708                 unlock_page(page);
 709         }
 710         page_cache_release(page);
 711 }
 712
 713 static int filter_commitrw_read(struct obd_export *exp, struct obdo *oa,
 714                                 int objcount, struct obd_ioobj *obj,
 715                                 int niocount, struct niobuf_local *res,
 716                                 struct obd_trans_info *oti, int rc)
 717 {
 718         struct inode *inode = NULL;
 719         struct ldlm_res_id res_id = { .name = { obj->ioo_id } };
 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);
 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 }