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 <obd_ost.h>
  41 #include <lustre_fsfilt.h>
  42 #include "filter_internal.h"
  43
  44 int *obdfilter_created_scratchpad;
  45
  46 static int filter_alloc_dio_page(struct obd_device *obd, struct inode *inode,
  47                                  struct niobuf_local *lnb)
  48 {
  49         struct page *page;
  50
  51         LASSERT(lnb->page != NULL);
  52
  53         page = lnb->page;
  54 #if 0
  55         POISON_PAGE(page, 0xf1);
  56         if (lnb->len != CFS_PAGE_SIZE) {
  57                 memset(kmap(page) + lnb->len, 0, CFS_PAGE_SIZE - lnb->len);
  58                 kunmap(page);
  59         }
  60 #endif
  61         page->index = lnb->offset >> CFS_PAGE_SHIFT;
  62
  63         RETURN(0);
  64 }
  65
  66 static void filter_free_dio_pages(int objcount, struct obd_ioobj *obj,
  67                            int niocount, struct niobuf_local *res)
  68 {
  69         int i, j;
  70
  71         for (i = 0; i < objcount; i++, obj++) {
  72                 for (j = 0 ; j < obj->ioo_bufcnt ; j++, res++)
  73                                 res->page = NULL;
  74         }
  75 }
  76
  77 /* Grab the dirty and seen grant announcements from the incoming obdo.
  78  * We will later calculate the clients new grant and return it.
  79  * Caller must hold osfs lock */
  80 static void filter_grant_incoming(struct obd_export *exp, struct obdo *oa)
  81 {
  82         struct filter_export_data *fed;
  83         struct obd_device *obd = exp->exp_obd;
  84         ENTRY;
  85
  86         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
  87
  88         if ((oa->o_valid & (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) !=
  89                                         (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) {
  90                 oa->o_valid &= ~OBD_MD_FLGRANT;
  91                 EXIT;
  92                 return;
  93         }
  94
  95         fed = &exp->exp_filter_data;
  96
  97         /* Add some margin, since there is a small race if other RPCs arrive
  98          * out-or-order and have already consumed some grant.  We want to
  99          * leave this here in case there is a large error in accounting. */
 100         CDEBUG(D_CACHE,
 101                "%s: cli %s/%p reports grant: "LPU64" dropped: %u, local: %lu\n",
 102                obd->obd_name, exp->exp_client_uuid.uuid, exp, oa->o_grant,
 103                oa->o_dropped, fed->fed_grant);
 104
 105         /* Update our accounting now so that statfs takes it into account.
 106          * Note that fed_dirty is only approximate and can become incorrect
 107          * if RPCs arrive out-of-order.  No important calculations depend
 108          * on fed_dirty however, but we must check sanity to not assert. */
 109         if ((long long)oa->o_dirty < 0)
 110                 oa->o_dirty = 0;
 111         else if (oa->o_dirty > fed->fed_grant + 4 * FILTER_GRANT_CHUNK)
 112                 oa->o_dirty = fed->fed_grant + 4 * FILTER_GRANT_CHUNK;
 113         obd->u.filter.fo_tot_dirty += oa->o_dirty - fed->fed_dirty;
 114         if (fed->fed_grant < oa->o_dropped) {
 115                 CDEBUG(D_CACHE,"%s: cli %s/%p reports %u dropped > grant %lu\n",
 116                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
 117                        oa->o_dropped, fed->fed_grant);
 118                 oa->o_dropped = 0;
 119         }
 120         if (obd->u.filter.fo_tot_granted < oa->o_dropped) {
 121                 CERROR("%s: cli %s/%p reports %u dropped > tot_grant "LPU64"\n",
 122                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
 123                        oa->o_dropped, obd->u.filter.fo_tot_granted);
 124                 oa->o_dropped = 0;
 125         }
 126         obd->u.filter.fo_tot_granted -= oa->o_dropped;
 127         fed->fed_grant -= oa->o_dropped;
 128         fed->fed_dirty = oa->o_dirty;
 129         if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
 130                 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
 131                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
 132                        fed->fed_dirty, fed->fed_pending, fed->fed_grant);
 133                 spin_unlock(&obd->obd_osfs_lock);
 134                 LBUG();
 135         }
 136         EXIT;
 137 }
 138
 139 /* Figure out how much space is available between what we've granted
 140  * and what remains in the filesystem.  Compensate for ext3 indirect
 141  * block overhead when computing how much free space is left ungranted.
 142  *
 143  * Caller must hold obd_osfs_lock. */
 144 obd_size filter_grant_space_left(struct obd_export *exp)
 145 {
 146         struct obd_device *obd = exp->exp_obd;
 147         int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
 148         obd_size tot_granted = obd->u.filter.fo_tot_granted, avail, left = 0;
 149         int rc, statfs_done = 0;
 150
 151         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
 152
 153         if (cfs_time_before_64(obd->obd_osfs_age, cfs_time_current_64() - HZ)) {
 154 restat:
 155                 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb,
 156                                    cfs_time_current_64() + HZ);
 157                 if (rc) /* N.B. statfs can't really fail */
 158                         RETURN(0);
 159                 statfs_done = 1;
 160         }
 161
 162         avail = obd->obd_osfs.os_bavail;
 163         left = avail - (avail >> (blockbits - 3)); /* (d)indirect */
 164         if (left > GRANT_FOR_LLOG(obd)) {
 165                 left = (left - GRANT_FOR_LLOG(obd)) << blockbits;
 166         } else {
 167                 left = 0 /* << blockbits */;
 168         }
 169
 170         if (!statfs_done && left < 32 * FILTER_GRANT_CHUNK + tot_granted) {
 171                 CDEBUG(D_CACHE, "fs has no space left and statfs too old\n");
 172                 goto restat;
 173         }
 174
 175         if (left >= tot_granted) {
 176                 left -= tot_granted;
 177         } else {
 178                 if (left < tot_granted - obd->u.filter.fo_tot_pending) {
 179                         CERROR("%s: cli %s/%p grant "LPU64" > available "
 180                                LPU64" and pending "LPU64"\n", obd->obd_name,
 181                                exp->exp_client_uuid.uuid, exp, tot_granted,
 182                                left, obd->u.filter.fo_tot_pending);
 183                 }
 184                 left = 0;
 185         }
 186
 187         CDEBUG(D_CACHE, "%s: cli %s/%p free: "LPU64" avail: "LPU64" grant "LPU64
 188                " left: "LPU64" pending: "LPU64"\n", obd->obd_name,
 189                exp->exp_client_uuid.uuid, exp,
 190                obd->obd_osfs.os_bfree << blockbits, avail << blockbits,
 191                tot_granted, left, obd->u.filter.fo_tot_pending);
 192
 193         return left;
 194 }
 195
 196 /* Calculate how much grant space to allocate to this client, based on how
 197  * much space is currently free and how much of that is already granted.
 198  *
 199  * Caller must hold obd_osfs_lock. */
 200 long filter_grant(struct obd_export *exp, obd_size current_grant,
 201                   obd_size want, obd_size fs_space_left)
 202 {
 203         struct obd_device *obd = exp->exp_obd;
 204         struct filter_export_data *fed = &exp->exp_filter_data;
 205         int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
 206         __u64 grant = 0;
 207
 208         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
 209
 210         /* Grant some fraction of the client's requested grant space so that
 211          * they are not always waiting for write credits (not all of it to
 212          * avoid overgranting in face of multiple RPCs in flight).  This
 213          * essentially will be able to control the OSC_MAX_RIF for a client.
 214          *
 215          * If we do have a large disparity between what the client thinks it
 216          * has and what we think it has, don't grant very much and let the
 217          * client consume its grant first.  Either it just has lots of RPCs
 218          * in flight, or it was evicted and its grants will soon be used up. */
 219         if (want > 0x7fffffff) {
 220                 CERROR("%s: client %s/%p requesting > 2GB grant "LPU64"\n",
 221                        obd->obd_name, exp->exp_client_uuid.uuid, exp, want);
 222         } else if (current_grant < want &&
 223                    current_grant < fed->fed_grant + FILTER_GRANT_CHUNK) {
 224                 grant = min((want >> blockbits),
 225                             (fs_space_left >> blockbits) / 8);
 226                 grant <<= blockbits;
 227
 228                 if (grant) {
 229                         /* Allow >FILTER_GRANT_CHUNK size when clients
 230                          * reconnect due to a server reboot.
 231                          */
 232                         if ((grant > FILTER_GRANT_CHUNK) &&
 233                             (!obd->obd_recovering))
 234                                 grant = FILTER_GRANT_CHUNK;
 235
 236                         obd->u.filter.fo_tot_granted += grant;
 237                         fed->fed_grant += grant;
 238                         if (fed->fed_grant < 0) {
 239                                 CERROR("%s: cli %s/%p grant %ld want "LPU64
 240                                        "current"LPU64"\n",
 241                                        obd->obd_name, exp->exp_client_uuid.uuid,
 242                                        exp, fed->fed_grant, want,current_grant);
 243                                 spin_unlock(&obd->obd_osfs_lock);
 244                                 LBUG();
 245                         }
 246                 }
 247         }
 248
 249         CDEBUG(D_CACHE,
 250                "%s: cli %s/%p wants: "LPU64" current grant "LPU64
 251                " granting: "LPU64"\n", obd->obd_name, exp->exp_client_uuid.uuid,
 252                exp, want, current_grant, grant);
 253         CDEBUG(D_CACHE,
 254                "%s: cli %s/%p tot cached:"LPU64" granted:"LPU64
 255                " num_exports: %d\n", obd->obd_name, exp->exp_client_uuid.uuid,
 256                exp, obd->u.filter.fo_tot_dirty,
 257                obd->u.filter.fo_tot_granted, obd->obd_num_exports);
 258
 259         return grant;
 260 }
 261
 262 static int filter_preprw_read(int cmd, struct obd_export *exp, struct obdo *oa,
 263                               int objcount, struct obd_ioobj *obj,
 264                               int niocount, struct niobuf_remote *nb,
 265                               struct niobuf_local *res,
 266                               struct obd_trans_info *oti,
 267                               struct lustre_capa *capa)
 268 {
 269         struct obd_device *obd = exp->exp_obd;
 270         struct lvfs_run_ctxt saved;
 271         struct niobuf_remote *rnb;
 272         struct niobuf_local *lnb;
 273         struct dentry *dentry = NULL;
 274         struct inode *inode;
 275         void *iobuf = NULL;
 276         int rc = 0, i, tot_bytes = 0;
 277         unsigned long now = jiffies;
 278         ENTRY;
 279
 280         /* We are currently not supporting multi-obj BRW_READ RPCS at all.
 281          * When we do this function's dentry cleanup will need to be fixed.
 282          * These values are verified in ost_brw_write() from the wire. */
 283         LASSERTF(objcount == 1, "%d\n", objcount);
 284         LASSERTF(obj->ioo_bufcnt > 0, "%d\n", obj->ioo_bufcnt);
 285
 286         rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa), capa,
 287                               CAPA_OPC_OSS_READ);
 288         if (rc)
 289                 RETURN(rc);
 290
 291         if (oa && oa->o_valid & OBD_MD_FLGRANT) {
 292                 spin_lock(&obd->obd_osfs_lock);
 293                 filter_grant_incoming(exp, oa);
 294
 295                 oa->o_grant = 0;
 296                 spin_unlock(&obd->obd_osfs_lock);
 297         }
 298
 299         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 300         if (IS_ERR(iobuf))
 301                 RETURN(PTR_ERR(iobuf));
 302
 303         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 304         dentry = filter_oa2dentry(obd, oa);
 305         if (IS_ERR(dentry)) {
 306                 rc = PTR_ERR(dentry);
 307                 dentry = NULL;
 308                 GOTO(cleanup, rc);
 309         }
 310
 311         inode = dentry->d_inode;
 312
 313         obdo_to_inode(inode, oa, OBD_MD_FLATIME);
 314         fsfilt_check_slow(obd, now, "preprw_read setup");
 315
 316         for (i = 0, lnb = res, rnb = nb; i < obj->ioo_bufcnt;
 317              i++, rnb++, lnb++) {
 318                 lnb->dentry = dentry;
 319                 lnb->offset = rnb->offset;
 320                 lnb->len    = rnb->len;
 321                 lnb->flags  = rnb->flags;
 322
 323                 /*
 324                  * ost_brw_write()->ost_nio_pages_get() already initialized
 325                  * lnb->page to point to the page from the per-thread page
 326                  * pool (bug 5137), initialize page.
 327                  */
 328                 LASSERT(lnb->page != NULL);
 329
 330                 if (i_size_read(inode) <= rnb->offset)
 331                         /* If there's no more data, abort early.  lnb->rc == 0,
 332                          * so it's easy to detect later. */
 333                         break;
 334                 else
 335                         filter_alloc_dio_page(obd, inode, lnb);
 336
 337                 if (i_size_read(inode) < lnb->offset + lnb->len - 1)
 338                         lnb->rc = i_size_read(inode) - lnb->offset;
 339                 else
 340                         lnb->rc = lnb->len;
 341
 342                 tot_bytes += lnb->rc;
 343
 344                 filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 345         }
 346
 347         fsfilt_check_slow(obd, now, "start_page_read");
 348
 349         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf,
 350                               exp, NULL, NULL, NULL);
 351         if (rc)
 352                 GOTO(cleanup, rc);
 353
 354         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_READ_BYTES, tot_bytes);
 355
 356         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
 357                 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
 358                                     LPROC_FILTER_READ_BYTES, tot_bytes);
 359
 360         EXIT;
 361
 362  cleanup:
 363         if (rc != 0) {
 364                 filter_free_dio_pages(objcount, obj, niocount, res);
 365
 366                 if (dentry != NULL)
 367                         f_dput(dentry);
 368         }
 369
 370         filter_iobuf_put(&obd->u.filter, iobuf, oti);
 371
 372         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 373         if (rc)
 374                 CERROR("io error %d\n", rc);
 375
 376         return rc;
 377 }
 378
 379 /* When clients have dirtied as much space as they've been granted they
 380  * fall through to sync writes.  These sync writes haven't been expressed
 381  * in grants and need to error with ENOSPC when there isn't room in the
 382  * filesystem for them after grants are taken into account.  However,
 383  * writeback of the dirty data that was already granted space can write
 384  * right on through.
 385  *
 386  * Caller must hold obd_osfs_lock. */
 387 static int filter_grant_check(struct obd_export *exp, struct obdo *oa,
 388                               int objcount, struct fsfilt_objinfo *fso,
 389                               int niocount, struct niobuf_remote *rnb,
 390                               struct niobuf_local *lnb, obd_size *left,
 391                               struct inode *inode)
 392 {
 393         struct filter_export_data *fed = &exp->exp_filter_data;
 394         int blocksize = exp->exp_obd->u.obt.obt_sb->s_blocksize;
 395         unsigned long used = 0, ungranted = 0, using;
 396         int i, rc = -ENOSPC, obj, n = 0;
 397
 398         LASSERT_SPIN_LOCKED(&exp->exp_obd->obd_osfs_lock);
 399
 400         for (obj = 0; obj < objcount; obj++) {
 401                 for (i = 0; i < fso[obj].fso_bufcnt; i++, n++) {
 402                         int tmp, bytes;
 403
 404                         /* should match the code in osc_exit_cache */
 405                         bytes = rnb[n].len;
 406                         bytes += rnb[n].offset & (blocksize - 1);
 407                         tmp = (rnb[n].offset + rnb[n].len) & (blocksize - 1);
 408                         if (tmp)
 409                                 bytes += blocksize - tmp;
 410
 411                         if ((rnb[n].flags & OBD_BRW_FROM_GRANT) &&
 412                             (oa->o_valid & OBD_MD_FLGRANT)) {
 413                                 if (fed->fed_grant < used + bytes) {
 414                                         CDEBUG(D_CACHE,
 415                                                "%s: cli %s/%p claims %ld+%d "
 416                                                "GRANT, real grant %lu idx %d\n",
 417                                                exp->exp_obd->obd_name,
 418                                                exp->exp_client_uuid.uuid, exp,
 419                                                used, bytes, fed->fed_grant, n);
 420                                 } else {
 421                                         used += bytes;
 422                                         rnb[n].flags |= OBD_BRW_GRANTED;
 423                                         lnb[n].lnb_grant_used = bytes;
 424                                         CDEBUG(0, "idx %d used=%lu\n", n, used);
 425                                         rc = 0;
 426                                         continue;
 427                                 }
 428                         }
 429                         if (*left > ungranted + bytes) {
 430                                 /* if enough space, pretend it was granted */
 431                                 ungranted += bytes;
 432                                 rnb[n].flags |= OBD_BRW_GRANTED;
 433                                 lnb[n].lnb_grant_used = bytes;
 434                                 CDEBUG(0, "idx %d ungranted=%lu\n",n,ungranted);
 435                                 rc = 0;
 436                                 continue;
 437                         }
 438
 439                         /* We can't check for already-mapped blocks here, as
 440                          * it requires dropping the osfs lock to do the bmap.
 441                          * Instead, we return ENOSPC and in that case we need
 442                          * to go through and verify if all of the blocks not
 443                          * marked BRW_GRANTED are already mapped and we can
 444                          * ignore this error. */
 445                         lnb[n].rc = -ENOSPC;
 446                         rnb[n].flags &= ~OBD_BRW_GRANTED;
 447                         CDEBUG(D_CACHE,"%s: cli %s/%p idx %d no space for %d\n",
 448                                exp->exp_obd->obd_name,
 449                                exp->exp_client_uuid.uuid, exp, n, bytes);
 450                 }
 451         }
 452
 453         /* Now substract what client have used already.  We don't subtract
 454          * this from the tot_granted yet, so that other client's can't grab
 455          * that space before we have actually allocated our blocks.  That
 456          * happens in filter_grant_commit() after the writes are done. */
 457         *left -= ungranted;
 458         fed->fed_grant -= used;
 459         fed->fed_pending += used + ungranted;
 460         exp->exp_obd->u.filter.fo_tot_granted += ungranted;
 461         exp->exp_obd->u.filter.fo_tot_pending += used + ungranted;
 462
 463         CDEBUG(D_CACHE,
 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                                struct lustre_capa *capa)
 511 {
 512         struct lvfs_run_ctxt saved;
 513         struct niobuf_remote *rnb;
 514         struct niobuf_local *lnb = res;
 515         struct fsfilt_objinfo fso;
 516         struct filter_mod_data *fmd;
 517         struct dentry *dentry = NULL;
 518         void *iobuf;
 519         obd_size left;
 520         unsigned long now = jiffies;
 521         int rc = 0, i, tot_bytes = 0, cleanup_phase = 0;
 522         ENTRY;
 523         LASSERT(objcount == 1);
 524         LASSERT(obj->ioo_bufcnt > 0);
 525
 526         rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa), capa,
 527                               CAPA_OPC_OSS_WRITE);
 528         if (rc)
 529                 RETURN(rc);
 530
 531         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 532         iobuf = filter_iobuf_get(&exp->exp_obd->u.filter, oti);
 533         if (IS_ERR(iobuf))
 534                 GOTO(cleanup, rc = PTR_ERR(iobuf));
 535         cleanup_phase = 1;
 536
 537         dentry = filter_fid2dentry(exp->exp_obd, NULL, obj->ioo_gr,
 538                                    obj->ioo_id);
 539         if (IS_ERR(dentry))
 540                 GOTO(cleanup, rc = PTR_ERR(dentry));
 541         cleanup_phase = 2;
 542
 543         if (dentry->d_inode == NULL) {
 544                 CERROR("%s: trying to BRW to non-existent file "LPU64"\n",
 545                        exp->exp_obd->obd_name, obj->ioo_id);
 546                 GOTO(cleanup, rc = -ENOENT);
 547         }
 548
 549         fso.fso_dentry = dentry;
 550         fso.fso_bufcnt = obj->ioo_bufcnt;
 551
 552         fsfilt_check_slow(exp->exp_obd, now, "preprw_write setup");
 553
 554         /* Don't update inode timestamps if this write is older than a
 555          * setattr which modifies the timestamps. b=10150 */
 556         /* XXX when we start having persistent reservations this needs to
 557          * be changed to filter_fmd_get() to create the fmd if it doesn't
 558          * already exist so we can store the reservation handle there. */
 559         fmd = filter_fmd_find(exp, obj->ioo_id, obj->ioo_gr);
 560
 561         LASSERT(oa != NULL);
 562         spin_lock(&exp->exp_obd->obd_osfs_lock);
 563         filter_grant_incoming(exp, oa);
 564         if (fmd && fmd->fmd_mactime_xid > oti->oti_xid)
 565                 oa->o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 566                                  OBD_MD_FLATIME);
 567         else
 568                 obdo_to_inode(dentry->d_inode, oa, OBD_MD_FLATIME |
 569                               OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 570         cleanup_phase = 3;
 571
 572         left = filter_grant_space_left(exp);
 573
 574         rc = filter_grant_check(exp, oa, objcount, &fso, niocount, nb, res,
 575                                 &left, dentry->d_inode);
 576
 577         /* do not zero out oa->o_valid as it is used in filter_commitrw_write()
 578          * for setting UID/GID and fid EA in first write time. */
 579         if (oa->o_valid & OBD_MD_FLGRANT)
 580                 oa->o_grant = filter_grant(exp,oa->o_grant,oa->o_undirty,left);
 581
 582         spin_unlock(&exp->exp_obd->obd_osfs_lock);
 583         filter_fmd_put(exp, fmd);
 584
 585         if (rc)
 586                 GOTO(cleanup, rc);
 587
 588         for (i = 0, rnb = nb, lnb = res; i < obj->ioo_bufcnt;
 589              i++, lnb++, rnb++) {
 590                 /* We still set up for ungranted pages so that granted pages
 591                  * can be written to disk as they were promised, and portals
 592                  * needs to keep the pages all aligned properly. */
 593                 lnb->dentry = dentry;
 594                 lnb->offset = rnb->offset;
 595                 lnb->len    = rnb->len;
 596                 lnb->flags  = rnb->flags;
 597
 598                 /*
 599                  * ost_brw_write()->ost_nio_pages_get() already initialized
 600                  * lnb->page to point to the page from the per-thread page
 601                  * pool (bug 5137), initialize page.
 602                  */
 603                 LASSERT(lnb->page != NULL);
 604                 if (lnb->len != CFS_PAGE_SIZE) {
 605                         memset(kmap(lnb->page) + lnb->len,
 606                                0, CFS_PAGE_SIZE - lnb->len);
 607                         kunmap(lnb->page);
 608                 }
 609                 lnb->page->index = lnb->offset >> CFS_PAGE_SHIFT;
 610
 611                 cleanup_phase = 4;
 612
 613                 /* If the filter writes a partial page, then has the file
 614                  * extended, the client will read in the whole page.  the
 615                  * filter has to be careful to zero the rest of the partial
 616                  * page on disk.  we do it by hand for partial extending
 617                  * writes, send_bio() is responsible for zeroing pages when
 618                  * asked to read unmapped blocks -- brw_kiovec() does this. */
 619                 if (lnb->len != CFS_PAGE_SIZE) {
 620                         __s64 maxidx;
 621
 622                         maxidx = ((i_size_read(dentry->d_inode) +
 623                                    CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT) - 1;
 624                         if (maxidx >= lnb->page->index) {
 625                                 LL_CDEBUG_PAGE(D_PAGE, lnb->page, "write %u @ "
 626                                                LPU64" flg %x before EOF %llu\n",
 627                                                lnb->len, lnb->offset,lnb->flags,
 628                                                i_size_read(dentry->d_inode));
 629                                 filter_iobuf_add_page(exp->exp_obd, iobuf,
 630                                                       dentry->d_inode,
 631                                                       lnb->page);
 632                         } else {
 633                                 long off;
 634                                 char *p = kmap(lnb->page);
 635
 636                                 off = lnb->offset & ~CFS_PAGE_MASK;
 637                                 if (off)
 638                                         memset(p, 0, off);
 639                                 off = (lnb->offset + lnb->len) & ~CFS_PAGE_MASK;
 640                                 if (off)
 641                                         memset(p + off, 0, CFS_PAGE_SIZE - off);
 642                                 kunmap(lnb->page);
 643                         }
 644                 }
 645                 if (lnb->rc == 0)
 646                         tot_bytes += lnb->len;
 647         }
 648
 649         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf, exp,
 650                               NULL, NULL, NULL);
 651
 652         fsfilt_check_slow(exp->exp_obd, now, "start_page_write");
 653
 654         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
 655                 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
 656                                     LPROC_FILTER_WRITE_BYTES, tot_bytes);
 657         EXIT;
 658 cleanup:
 659         switch(cleanup_phase) {
 660         case 4:
 661         case 3:
 662                 filter_iobuf_put(&exp->exp_obd->u.filter, iobuf, oti);
 663         case 2:
 664                 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 665                 if (rc)
 666                         f_dput(dentry);
 667                 break;
 668         case 1:
 669                 filter_iobuf_put(&exp->exp_obd->u.filter, iobuf, oti);
 670         case 0:
 671                 spin_lock(&exp->exp_obd->obd_osfs_lock);
 672                 if (oa)
 673                         filter_grant_incoming(exp, oa);
 674                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
 675                 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 676                 break;
 677         default:;
 678         }
 679         return rc;
 680 }
 681
 682 int filter_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
 683                   int objcount, struct obd_ioobj *obj, int niocount,
 684                   struct niobuf_remote *nb, struct niobuf_local *res,
 685                   struct obd_trans_info *oti, struct lustre_capa *capa)
 686 {
 687         if (cmd == OBD_BRW_WRITE)
 688                 return filter_preprw_write(cmd, exp, oa, objcount, obj,
 689                                            niocount, nb, res, oti, capa);
 690         if (cmd == OBD_BRW_READ)
 691                 return filter_preprw_read(cmd, exp, oa, objcount, obj,
 692                                           niocount, nb, res, oti, capa);
 693         LBUG();
 694         return -EPROTO;
 695 }
 696
 697 void filter_release_read_page(struct filter_obd *filter, struct inode *inode,
 698                               struct page *page)
 699 {
 700         int drop = 0;
 701
 702         if (inode != NULL &&
 703             (i_size_read(inode) > filter->fo_readcache_max_filesize))
 704                 drop = 1;
 705
 706         /* drop from cache like truncate_list_pages() */
 707         if (drop && !TryLockPage(page)) {
 708                 if (page->mapping)
 709                         ll_truncate_complete_page(page);
 710                 unlock_page(page);
 711         }
 712         page_cache_release(page);
 713 }
 714
 715 static int filter_commitrw_read(struct obd_export *exp, struct obdo *oa,
 716                                 int objcount, struct obd_ioobj *obj,
 717                                 int niocount, struct niobuf_local *res,
 718                                 struct obd_trans_info *oti, int rc)
 719 {
 720         struct inode *inode = NULL;
 721         struct ldlm_res_id res_id;
 722         struct ldlm_resource *resource = NULL;
 723         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 724         ENTRY;
 725
 726         osc_build_res_name(obj->ioo_id, obj->ioo_gr, &res_id);
 727         /* If oa != NULL then filter_preprw_read updated the inode atime
 728          * and we should update the lvb so that other glimpses will also
 729          * get the updated value. bug 5972 */
 730         if (oa && ns && ns->ns_lvbo && ns->ns_lvbo->lvbo_update) {
 731                 resource = ldlm_resource_get(ns, NULL, &res_id, LDLM_EXTENT, 0);
 732
 733                 if (resource != NULL) {
 734                         ns->ns_lvbo->lvbo_update(resource, NULL, 0, 1);
 735                         ldlm_resource_putref(resource);
 736                 }
 737         }
 738
 739         if (res->dentry != NULL)
 740                 inode = res->dentry->d_inode;
 741
 742         filter_free_dio_pages(objcount, obj, niocount, res);
 743
 744         if (res->dentry != NULL)
 745                 f_dput(res->dentry);
 746         RETURN(rc);
 747 }
 748
 749 void flip_into_page_cache(struct inode *inode, struct page *new_page)
 750 {
 751         struct page *old_page;
 752         int rc;
 753
 754         do {
 755                 /* the dlm is protecting us from read/write concurrency, so we
 756                  * expect this find_lock_page to return quickly.  even if we
 757                  * race with another writer it won't be doing much work with
 758                  * the page locked.  we do this 'cause t_c_p expects a
 759                  * locked page, and it wants to grab the pagecache lock
 760                  * as well. */
 761                 old_page = find_lock_page(inode->i_mapping, new_page->index);
 762                 if (old_page) {
 763                         ll_truncate_complete_page(old_page);
 764                         unlock_page(old_page);
 765                         page_cache_release(old_page);
 766                 }
 767
 768 #if 0 /* this should be a /proc tunable someday */
 769                 /* racing o_directs (no locking ioctl) could race adding
 770                  * their pages, so we repeat the page invalidation unless
 771                  * we successfully added our new page */
 772                 rc = add_to_page_cache_unique(new_page, inode->i_mapping,
 773                                               new_page->index,
 774                                               page_hash(inode->i_mapping,
 775                                                         new_page->index));
 776                 if (rc == 0) {
 777                         /* add_to_page_cache clears uptodate|dirty and locks
 778                          * the page */
 779                         SetPageUptodate(new_page);
 780                         unlock_page(new_page);
 781                 }
 782 #else
 783                 rc = 0;
 784 #endif
 785         } while (rc != 0);
 786 }
 787
 788 void filter_grant_commit(struct obd_export *exp, int niocount,
 789                          struct niobuf_local *res)
 790 {
 791         struct filter_obd *filter = &exp->exp_obd->u.filter;
 792         struct niobuf_local *lnb = res;
 793         unsigned long pending = 0;
 794         int i;
 795
 796         spin_lock(&exp->exp_obd->obd_osfs_lock);
 797         for (i = 0, lnb = res; i < niocount; i++, lnb++)
 798                 pending += lnb->lnb_grant_used;
 799
 800         LASSERTF(exp->exp_filter_data.fed_pending >= pending,
 801                  "%s: cli %s/%p fed_pending: %lu grant_used: %lu\n",
 802                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 803                  exp->exp_filter_data.fed_pending, pending);
 804         exp->exp_filter_data.fed_pending -= pending;
 805         LASSERTF(filter->fo_tot_granted >= pending,
 806                  "%s: cli %s/%p tot_granted: "LPU64" grant_used: %lu\n",
 807                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 808                  exp->exp_obd->u.filter.fo_tot_granted, pending);
 809         filter->fo_tot_granted -= pending;
 810         LASSERTF(filter->fo_tot_pending >= pending,
 811                  "%s: cli %s/%p tot_pending: "LPU64" grant_used: %lu\n",
 812                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 813                  filter->fo_tot_pending, pending);
 814         filter->fo_tot_pending -= pending;
 815
 816         spin_unlock(&exp->exp_obd->obd_osfs_lock);
 817 }
 818
 819 int filter_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
 820                     int objcount, struct obd_ioobj *obj, int niocount,
 821                     struct niobuf_local *res, struct obd_trans_info *oti,
 822                     int rc)
 823 {
 824         if (cmd == OBD_BRW_WRITE)
 825                 return filter_commitrw_write(exp, oa, objcount, obj, niocount,
 826                                              res, oti, rc);
 827         if (cmd == OBD_BRW_READ)
 828                 return filter_commitrw_read(exp, oa, objcount, obj, niocount,
 829                                             res, oti, rc);
 830         LBUG();
 831         return -EPROTO;
 832 }
 833
 834 int filter_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
 835                obd_count oa_bufs, struct brw_page *pga,
 836                struct obd_trans_info *oti)
 837 {
 838         struct obd_ioobj ioo;
 839         struct niobuf_local *lnb;
 840         struct niobuf_remote *rnb;
 841         obd_count i;
 842         int ret = 0;
 843         ENTRY;
 844
 845         OBD_ALLOC(lnb, oa_bufs * sizeof(struct niobuf_local));
 846         OBD_ALLOC(rnb, oa_bufs * sizeof(struct niobuf_remote));
 847
 848         if (lnb == NULL || rnb == NULL)
 849                 GOTO(out, ret = -ENOMEM);
 850
 851         for (i = 0; i < oa_bufs; i++) {
 852                 lnb[i].page = pga[i].pg;
 853                 rnb[i].offset = pga[i].off;
 854                 rnb[i].len = pga[i].count;
 855         }
 856
 857         obdo_to_ioobj(oinfo->oi_oa, &ioo);
 858         ioo.ioo_bufcnt = oa_bufs;
 859
 860         ret = filter_preprw(cmd, exp, oinfo->oi_oa, 1, &ioo,
 861                             oa_bufs, rnb, lnb, oti, oinfo_capa(oinfo));
 862         if (ret != 0)
 863                 GOTO(out, ret);
 864
 865         ret = filter_commitrw(cmd, exp, oinfo->oi_oa, 1, &ioo,
 866                               oa_bufs, lnb, oti, ret);
 867
 868 out:
 869         if (lnb)
 870                 OBD_FREE(lnb, oa_bufs * sizeof(struct niobuf_local));
 871         if (rnb)
 872                 OBD_FREE(rnb, oa_bufs * sizeof(struct niobuf_remote));
 873         RETURN(ret);
 874 }