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 #include <linux/config.h>
  33 #include <linux/module.h>
  34 #include <linux/pagemap.h> // XXX kill me soon
  35 #include <linux/version.h>
  36
  37 #include <obd_class.h>
  38 #include <lustre_fsfilt.h>
  39 #include "filter_internal.h"
  40
  41 int *obdfilter_created_scratchpad;
  42
  43 static int filter_alloc_dio_page(struct obd_device *obd, struct inode *inode,
  44                                  struct niobuf_local *lnb)
  45 {
  46         struct page *page;
  47
  48         LASSERT(lnb->page != NULL);
  49
  50         page = lnb->page;
  51 #if 0
  52         POISON_PAGE(page, 0xf1);
  53         if (lnb->len != PAGE_SIZE) {
  54                 memset(kmap(page) + lnb->len, 0, PAGE_SIZE - lnb->len);
  55                 kunmap(page);
  56         }
  57 #endif
  58         page->index = lnb->offset >> PAGE_SHIFT;
  59
  60         RETURN(0);
  61 }
  62
  63 static void filter_free_dio_pages(int objcount, struct obd_ioobj *obj,
  64                            int niocount, struct niobuf_local *res)
  65 {
  66         int i, j;
  67
  68         for (i = 0; i < objcount; i++, obj++) {
  69                 for (j = 0 ; j < obj->ioo_bufcnt ; j++, res++)
  70                                 res->page = NULL;
  71         }
  72 }
  73
  74 /* Grab the dirty and seen grant announcements from the incoming obdo.
  75  * We will later calculate the clients new grant and return it.
  76  * Caller must hold osfs lock */
  77 static void filter_grant_incoming(struct obd_export *exp, struct obdo *oa)
  78 {
  79         struct filter_export_data *fed;
  80         struct obd_device *obd = exp->exp_obd;
  81         static unsigned long last_msg;
  82         static int last_count;
  83         int mask = D_CACHE;
  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         /* Don't print this to the console the first time it happens, since
  98          * it can happen legitimately on occasion, but only rarely. */
  99         if (time_after(jiffies, last_msg + 60 * HZ)) {
 100                 last_count = 0;
 101                 last_msg = jiffies;
 102         }
 103         if ((last_count & (-last_count)) == last_count)
 104                 mask = D_HA /* until bug 3273 is fixed D_WARNING */;
 105         last_count++;
 106
 107         /* Add some margin, since there is a small race if other RPCs arrive
 108          * out-or-order and have already consumed some grant.  We want to
 109          * leave this here in case there is a large error in accounting. */
 110         CDEBUG(oa->o_grant > fed->fed_grant + FILTER_GRANT_CHUNK ? mask:D_CACHE,
 111                "%s: cli %s/%p reports grant: "LPU64" dropped: %u, local: %lu\n",
 112                obd->obd_name, exp->exp_client_uuid.uuid, exp, oa->o_grant,
 113                oa->o_dropped, fed->fed_grant);
 114
 115         /* Update our accounting now so that statfs takes it into account.
 116          * Note that fed_dirty is only approximate and can become incorrect
 117          * if RPCs arrive out-of-order.  No important calculations depend
 118          * on fed_dirty however, but we must check sanity to not assert. */
 119         if ((long long)oa->o_dirty < 0)
 120                 oa->o_dirty = 0;
 121         else if (oa->o_dirty > fed->fed_grant + 4 * FILTER_GRANT_CHUNK)
 122                 oa->o_dirty = fed->fed_grant + 4 * FILTER_GRANT_CHUNK;
 123         obd->u.filter.fo_tot_dirty += oa->o_dirty - fed->fed_dirty;
 124         if (fed->fed_grant < oa->o_dropped) {
 125                 CDEBUG(D_HA,"%s: cli %s/%p reports %u dropped > fedgrant %lu\n",
 126                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
 127                        oa->o_dropped, fed->fed_grant);
 128                 oa->o_dropped = 0;
 129         }
 130         if (obd->u.filter.fo_tot_granted < oa->o_dropped) {
 131                 CERROR("%s: cli %s/%p reports %u dropped > tot_grant "LPU64"\n",
 132                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
 133                        oa->o_dropped, obd->u.filter.fo_tot_granted);
 134                 oa->o_dropped = 0;
 135         }
 136         obd->u.filter.fo_tot_granted -= oa->o_dropped;
 137         fed->fed_grant -= oa->o_dropped;
 138         fed->fed_dirty = oa->o_dirty;
 139         if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
 140                 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
 141                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
 142                        fed->fed_dirty, fed->fed_pending, fed->fed_grant);
 143                 spin_unlock(&obd->obd_osfs_lock);
 144                 LBUG();
 145         }
 146         EXIT;
 147 }
 148
 149 /* Figure out how much space is available between what we've granted
 150  * and what remains in the filesystem.  Compensate for ext3 indirect
 151  * block overhead when computing how much free space is left ungranted.
 152  *
 153  * Caller must hold obd_osfs_lock. */
 154 obd_size filter_grant_space_left(struct obd_export *exp)
 155 {
 156         struct obd_device *obd = exp->exp_obd;
 157         int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
 158         obd_size tot_granted = obd->u.filter.fo_tot_granted, avail, left = 0;
 159         int rc, statfs_done = 0;
 160
 161         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
 162
 163         if (cfs_time_before_64(obd->obd_osfs_age, cfs_time_current_64() - HZ)) {
 164 restat:
 165                 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb,
 166                                    cfs_time_current_64() + HZ);
 167                 if (rc) /* N.B. statfs can't really fail */
 168                         RETURN(0);
 169                 statfs_done = 1;
 170         }
 171
 172         avail = obd->obd_osfs.os_bavail;
 173         left = avail - (avail >> (blockbits - 3)); /* (d)indirect */
 174         if (left > GRANT_FOR_LLOG(obd)) {
 175                 left = (left - GRANT_FOR_LLOG(obd)) << blockbits;
 176         } else {
 177                 left = 0 /* << blockbits */;
 178         }
 179
 180         if (!statfs_done && left < 32 * FILTER_GRANT_CHUNK + tot_granted) {
 181                 CDEBUG(D_CACHE, "fs has no space left and statfs too old\n");
 182                 goto restat;
 183         }
 184
 185         if (left >= tot_granted) {
 186                 left -= tot_granted;
 187         } else {
 188                 if (left < tot_granted - obd->u.filter.fo_tot_pending) {
 189                         CERROR("%s: cli %s/%p grant "LPU64" > available "
 190                                LPU64" and pending "LPU64"\n", obd->obd_name,
 191                                exp->exp_client_uuid.uuid, exp, tot_granted,
 192                                left, obd->u.filter.fo_tot_pending);
 193                 }
 194                 left = 0;
 195         }
 196
 197         CDEBUG(D_CACHE, "%s: cli %s/%p free: "LPU64" avail: "LPU64" grant "LPU64
 198                " left: "LPU64" pending: "LPU64"\n", obd->obd_name,
 199                exp->exp_client_uuid.uuid, exp,
 200                obd->obd_osfs.os_bfree << blockbits, avail << blockbits,
 201                tot_granted, left, obd->u.filter.fo_tot_pending);
 202
 203         return left;
 204 }
 205
 206 /* Calculate how much grant space to allocate to this client, based on how
 207  * much space is currently free and how much of that is already granted.
 208  *
 209  * Caller must hold obd_osfs_lock. */
 210 long filter_grant(struct obd_export *exp, obd_size current_grant,
 211                   obd_size want, obd_size fs_space_left)
 212 {
 213         struct obd_device *obd = exp->exp_obd;
 214         struct filter_export_data *fed = &exp->exp_filter_data;
 215         int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
 216         __u64 grant = 0;
 217
 218         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
 219
 220         /* Grant some fraction of the client's requested grant space so that
 221          * they are not always waiting for write credits (not all of it to
 222          * avoid overgranting in face of multiple RPCs in flight).  This
 223          * essentially will be able to control the OSC_MAX_RIF for a client.
 224          *
 225          * If we do have a large disparity between what the client thinks it
 226          * has and what we think it has, don't grant very much and let the
 227          * client consume its grant first.  Either it just has lots of RPCs
 228          * in flight, or it was evicted and its grants will soon be used up. */
 229         if (want > 0x7fffffff) {
 230                 CERROR("%s: client %s/%p requesting > 2GB grant "LPU64"\n",
 231                        obd->obd_name, exp->exp_client_uuid.uuid, exp, want);
 232         } else if (current_grant < want &&
 233                    current_grant < fed->fed_grant + FILTER_GRANT_CHUNK) {
 234                 grant = min((want >> blockbits),
 235                             (fs_space_left >> blockbits) / 8);
 236                 grant <<= blockbits;
 237
 238                 if (grant) {
 239                         /* Allow >FILTER_GRANT_CHUNK size when clients
 240                          * reconnect due to a server reboot.
 241                          */
 242                         if ((grant > FILTER_GRANT_CHUNK) &&
 243                             (!obd->obd_recovering))
 244                                 grant = FILTER_GRANT_CHUNK;
 245
 246                         obd->u.filter.fo_tot_granted += grant;
 247                         fed->fed_grant += grant;
 248                         if (fed->fed_grant < 0) {
 249                                 CERROR("%s: cli %s/%p grant %ld want "LPU64
 250                                        "current"LPU64"\n",
 251                                        obd->obd_name, exp->exp_client_uuid.uuid,
 252                                        exp, fed->fed_grant, want,current_grant);
 253                                 spin_unlock(&obd->obd_osfs_lock);
 254                                 LBUG();
 255                         }
 256                 }
 257         }
 258
 259         CDEBUG(D_CACHE,"%s: cli %s/%p wants: "LPU64" granting: "LPU64"\n",
 260                obd->obd_name, exp->exp_client_uuid.uuid, exp, want, grant);
 261         CDEBUG(D_CACHE,
 262                "%s: cli %s/%p tot cached:"LPU64" granted:"LPU64
 263                " num_exports: %d\n", obd->obd_name, exp->exp_client_uuid.uuid,
 264                exp, obd->u.filter.fo_tot_dirty,
 265                obd->u.filter.fo_tot_granted, obd->obd_num_exports);
 266
 267         return grant;
 268 }
 269
 270 static int filter_preprw_read(int cmd, struct obd_export *exp, struct obdo *oa,
 271                               int objcount, struct obd_ioobj *obj,
 272                               int niocount, struct niobuf_remote *nb,
 273                               struct niobuf_local *res,
 274                               struct obd_trans_info *oti,
 275                               struct lustre_capa *capa)
 276 {
 277         struct obd_device *obd = exp->exp_obd;
 278         struct lvfs_run_ctxt saved;
 279         struct niobuf_remote *rnb;
 280         struct niobuf_local *lnb;
 281         struct dentry *dentry = NULL;
 282         struct inode *inode;
 283         void *iobuf = NULL;
 284         int rc = 0, i, tot_bytes = 0;
 285         unsigned long now = jiffies;
 286         ENTRY;
 287
 288         /* We are currently not supporting multi-obj BRW_READ RPCS at all.
 289          * When we do this function's dentry cleanup will need to be fixed.
 290          * These values are verified in ost_brw_write() from the wire. */
 291         LASSERTF(objcount == 1, "%d\n", objcount);
 292         LASSERTF(obj->ioo_bufcnt > 0, "%d\n", obj->ioo_bufcnt);
 293
 294         rc = filter_verify_capa(exp, NULL, obdo_mdsno(oa), capa,
 295                                 CAPA_OPC_OSS_READ);
 296         if (rc)
 297                 RETURN(rc);
 298
 299         if (oa && oa->o_valid & OBD_MD_FLGRANT) {
 300                 spin_lock(&obd->obd_osfs_lock);
 301                 filter_grant_incoming(exp, oa);
 302
 303                 oa->o_grant = 0;
 304                 spin_unlock(&obd->obd_osfs_lock);
 305         }
 306
 307         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 308         if (IS_ERR(iobuf))
 309                 RETURN(PTR_ERR(iobuf));
 310
 311         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 312         dentry = filter_oa2dentry(obd, oa);
 313         if (IS_ERR(dentry)) {
 314                 rc = PTR_ERR(dentry);
 315                 dentry = NULL;
 316                 GOTO(cleanup, rc);
 317         }
 318
 319         inode = dentry->d_inode;
 320
 321         obdo_to_inode(inode, oa, OBD_MD_FLATIME);
 322         fsfilt_check_slow(obd, now, obd_timeout, "preprw_read setup");
 323
 324         for (i = 0, lnb = res, rnb = nb; i < obj->ioo_bufcnt;
 325              i++, rnb++, lnb++) {
 326                 lnb->dentry = dentry;
 327                 lnb->offset = rnb->offset;
 328                 lnb->len    = rnb->len;
 329                 lnb->flags  = rnb->flags;
 330
 331                 /*
 332                  * ost_brw_write()->ost_nio_pages_get() already initialized
 333                  * lnb->page to point to the page from the per-thread page
 334                  * pool (bug 5137), initialize page.
 335                  */
 336                 LASSERT(lnb->page != NULL);
 337
 338                 if (inode->i_size <= rnb->offset)
 339                         /* If there's no more data, abort early.  lnb->rc == 0,
 340                          * so it's easy to detect later. */
 341                         break;
 342                 else
 343                         filter_alloc_dio_page(obd, inode, lnb);
 344
 345                 if (inode->i_size < lnb->offset + lnb->len - 1)
 346                         lnb->rc = inode->i_size - lnb->offset;
 347                 else
 348                         lnb->rc = lnb->len;
 349
 350                 tot_bytes += lnb->rc;
 351
 352                 filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 353         }
 354
 355         fsfilt_check_slow(obd, now, obd_timeout, "start_page_read");
 356
 357         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf,
 358                               exp, NULL, NULL, NULL);
 359         if (rc)
 360                 GOTO(cleanup, rc);
 361
 362         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_READ_BYTES, tot_bytes);
 363
 364         filter_tally_read(&exp->exp_obd->u.filter, res, niocount);
 365
 366         EXIT;
 367
 368  cleanup:
 369         if (rc != 0) {
 370                 filter_free_dio_pages(objcount, obj, niocount, res);
 371
 372                 if (dentry != NULL)
 373                         f_dput(dentry);
 374         }
 375
 376         filter_iobuf_put(&obd->u.filter, iobuf, oti);
 377
 378         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 379         if (rc)
 380                 CERROR("io error %d\n", rc);
 381
 382         return rc;
 383 }
 384
 385 /* When clients have dirtied as much space as they've been granted they
 386  * fall through to sync writes.  These sync writes haven't been expressed
 387  * in grants and need to error with ENOSPC when there isn't room in the
 388  * filesystem for them after grants are taken into account.  However,
 389  * writeback of the dirty data that was already granted space can write
 390  * right on through.
 391  *
 392  * Caller must hold obd_osfs_lock. */
 393 static int filter_grant_check(struct obd_export *exp, int objcount,
 394                               struct fsfilt_objinfo *fso, int niocount,
 395                               struct niobuf_remote *rnb,
 396                               struct niobuf_local *lnb, obd_size *left,
 397                               struct inode *inode)
 398 {
 399         struct filter_export_data *fed = &exp->exp_filter_data;
 400         int blocksize = exp->exp_obd->u.obt.obt_sb->s_blocksize;
 401         unsigned long used = 0, ungranted = 0, using;
 402         int i, rc = -ENOSPC, obj, n = 0, mask = D_CACHE;
 403
 404         LASSERT_SPIN_LOCKED(&exp->exp_obd->obd_osfs_lock);
 405
 406         for (obj = 0; obj < objcount; obj++) {
 407                 for (i = 0; i < fso[obj].fso_bufcnt; i++, n++) {
 408                         int tmp, bytes;
 409
 410                         /* should match the code in osc_exit_cache */
 411                         bytes = rnb[n].len;
 412                         bytes += rnb[n].offset & (blocksize - 1);
 413                         tmp = (rnb[n].offset + rnb[n].len) & (blocksize - 1);
 414                         if (tmp)
 415                                 bytes += blocksize - tmp;
 416
 417                         if (rnb[n].flags & OBD_BRW_FROM_GRANT) {
 418                                 if (fed->fed_grant < used + bytes) {
 419                                         CDEBUG(D_CACHE,
 420                                                "%s: cli %s/%p claims %ld+%d "
 421                                                "GRANT, real grant %lu idx %d\n",
 422                                                exp->exp_obd->obd_name,
 423                                                exp->exp_client_uuid.uuid, exp,
 424                                                used, bytes, fed->fed_grant, n);
 425                                         mask = D_RPCTRACE;
 426                                 } else {
 427                                         used += bytes;
 428                                         rnb[n].flags |= OBD_BRW_GRANTED;
 429                                         lnb[n].lnb_grant_used = bytes;
 430                                         CDEBUG(0, "idx %d used=%lu\n", n, used);
 431                                         rc = 0;
 432                                         continue;
 433                                 }
 434                         }
 435                         if (*left > ungranted) {
 436                                 /* if enough space, pretend it was granted */
 437                                 ungranted += bytes;
 438                                 rnb[n].flags |= OBD_BRW_GRANTED;
 439                                 CDEBUG(0, "idx %d ungranted=%lu\n",n,ungranted);
 440                                 rc = 0;
 441                                 continue;
 442                         }
 443
 444                         /* We can't check for already-mapped blocks here, as
 445                          * it requires dropping the osfs lock to do the bmap.
 446                          * Instead, we return ENOSPC and in that case we need
 447                          * to go through and verify if all of the blocks not
 448                          * marked BRW_GRANTED are already mapped and we can
 449                          * ignore this error. */
 450                         lnb[n].rc = -ENOSPC;
 451                         rnb[n].flags &= ~OBD_BRW_GRANTED;
 452                         CDEBUG(D_CACHE,"%s: cli %s/%p idx %d no space for %d\n",
 453                                exp->exp_obd->obd_name,
 454                                exp->exp_client_uuid.uuid, exp, n, bytes);
 455                 }
 456         }
 457
 458         /* Now substract what client have used already.  We don't subtract
 459          * this from the tot_granted yet, so that other client's can't grab
 460          * that space before we have actually allocated our blocks.  That
 461          * happens in filter_grant_commit() after the writes are done. */
 462         *left -= ungranted;
 463         fed->fed_grant -= used;
 464         fed->fed_pending += used;
 465         exp->exp_obd->u.filter.fo_tot_pending += used;
 466
 467         CDEBUG(mask,
 468                "%s: cli %s/%p used: %lu ungranted: %lu grant: %lu dirty: %lu\n",
 469                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp, used,
 470                ungranted, fed->fed_grant, fed->fed_dirty);
 471
 472         /* Rough calc in case we don't refresh cached statfs data */
 473         using = (used + ungranted + 1 ) >>
 474                 exp->exp_obd->u.obt.obt_sb->s_blocksize_bits;
 475         if (exp->exp_obd->obd_osfs.os_bavail > using)
 476                 exp->exp_obd->obd_osfs.os_bavail -= using;
 477         else
 478                 exp->exp_obd->obd_osfs.os_bavail = 0;
 479
 480         if (fed->fed_dirty < used) {
 481                 CERROR("%s: cli %s/%p claims used %lu > fed_dirty %lu\n",
 482                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 483                        used, fed->fed_dirty);
 484                 used = fed->fed_dirty;
 485         }
 486         exp->exp_obd->u.filter.fo_tot_dirty -= used;
 487         fed->fed_dirty -= used;
 488
 489         if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
 490                 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
 491                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 492                        fed->fed_dirty, fed->fed_pending, fed->fed_grant);
 493                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
 494                 LBUG();
 495         }
 496         return rc;
 497 }
 498
 499 /* If we ever start to support multi-object BRW RPCs, we will need to get locks
 500  * on mulitple inodes.  That isn't all, because there still exists the
 501  * possibility of a truncate starting a new transaction while holding the ext3
 502  * rwsem = write while some writes (which have started their transactions here)
 503  * blocking on the ext3 rwsem = read => lock inversion.
 504  *
 505  * The handling gets very ugly when dealing with locked pages.  It may be easier
 506  * to just get rid of the locked page code (which has problems of its own) and
 507  * either discover we do not need it anymore (i.e. it was a symptom of another
 508  * bug) or ensure we get the page locks in an appropriate order. */
 509 static int filter_preprw_write(int cmd, struct obd_export *exp, struct obdo *oa,
 510                                int objcount, struct obd_ioobj *obj,
 511                                int niocount, struct niobuf_remote *nb,
 512                                struct niobuf_local *res,
 513                                struct obd_trans_info *oti,
 514                                struct lustre_capa *capa)
 515 {
 516         struct lvfs_run_ctxt saved;
 517         struct niobuf_remote *rnb;
 518         struct niobuf_local *lnb = res;
 519         struct fsfilt_objinfo fso;
 520         struct filter_mod_data *fmd;
 521         struct dentry *dentry = NULL;
 522         void *iobuf;
 523         obd_size left;
 524         unsigned long now = jiffies;
 525         int rc = 0, i, tot_bytes = 0, cleanup_phase = 0;
 526         ENTRY;
 527         LASSERT(objcount == 1);
 528         LASSERT(obj->ioo_bufcnt > 0);
 529
 530         rc = filter_verify_capa(exp, NULL, obdo_mdsno(oa), capa,
 531                                 CAPA_OPC_OSS_WRITE);
 532         if (rc)
 533                 RETURN(rc);
 534
 535         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 536         iobuf = filter_iobuf_get(&exp->exp_obd->u.filter, oti);
 537         if (IS_ERR(iobuf))
 538                 GOTO(cleanup, rc = PTR_ERR(iobuf));
 539         cleanup_phase = 1;
 540
 541         dentry = filter_fid2dentry(exp->exp_obd, NULL, obj->ioo_gr,
 542                                    obj->ioo_id);
 543         if (IS_ERR(dentry))
 544                 GOTO(cleanup, rc = PTR_ERR(dentry));
 545         cleanup_phase = 2;
 546
 547         if (dentry->d_inode == NULL) {
 548                 CERROR("%s: trying to BRW to non-existent file "LPU64"\n",
 549                        exp->exp_obd->obd_name, obj->ioo_id);
 550                 GOTO(cleanup, rc = -ENOENT);
 551         }
 552
 553         fso.fso_dentry = dentry;
 554         fso.fso_bufcnt = obj->ioo_bufcnt;
 555
 556         fsfilt_check_slow(exp->exp_obd, now, obd_timeout, "preprw_write setup");
 557
 558         /* Don't update inode timestamps if this write is older than a
 559          * setattr which modifies the timestamps. b=10150 */
 560         /* XXX when we start having persistent reservations this needs to
 561          * be changed to filter_fmd_get() to create the fmd if it doesn't
 562          * already exist so we can store the reservation handle there. */
 563         fmd = filter_fmd_find(exp, obj->ioo_id, obj->ioo_gr);
 564
 565         spin_lock(&exp->exp_obd->obd_osfs_lock);
 566         if (oa) {
 567                 filter_grant_incoming(exp, oa);
 568                 if (fmd && fmd->fmd_mactime_xid > oti->oti_xid)
 569                         oa->o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 570                                          OBD_MD_FLATIME);
 571                 else
 572                         obdo_to_inode(dentry->d_inode, oa, OBD_MD_FLATIME |
 573                                       OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 574         }
 575         cleanup_phase = 3;
 576
 577         left = filter_grant_space_left(exp);
 578
 579         rc = filter_grant_check(exp, objcount, &fso, niocount, nb, res,
 580                                 &left, dentry->d_inode);
 581
 582         /* do not zero out oa->o_valid as it is used in filter_commitrw_write()
 583          * for setting UID/GID and fid EA in first write time. */
 584         if (oa && oa->o_valid & OBD_MD_FLGRANT) {
 585                 oa->o_grant = filter_grant(exp,oa->o_grant,oa->o_undirty,left);
 586                 oa->o_valid |= OBD_MD_FLGRANT;
 587         }
 588
 589         spin_unlock(&exp->exp_obd->obd_osfs_lock);
 590         filter_fmd_put(exp, fmd);
 591
 592         if (rc)
 593                 GOTO(cleanup, rc);
 594
 595         for (i = 0, rnb = nb, lnb = res; i < obj->ioo_bufcnt;
 596              i++, lnb++, rnb++) {
 597                 /* We still set up for ungranted pages so that granted pages
 598                  * can be written to disk as they were promised, and portals
 599                  * needs to keep the pages all aligned properly. */
 600                 lnb->dentry = dentry;
 601                 lnb->offset = rnb->offset;
 602                 lnb->len    = rnb->len;
 603                 lnb->flags  = rnb->flags;
 604
 605                 /*
 606                  * ost_brw_write()->ost_nio_pages_get() already initialized
 607                  * lnb->page to point to the page from the per-thread page
 608                  * pool (bug 5137), initialize page.
 609                  */
 610                 LASSERT(lnb->page != NULL);
 611                 if (lnb->len != PAGE_SIZE) {
 612                         memset(kmap(lnb->page) + lnb->len,
 613                                0, PAGE_SIZE - lnb->len);
 614                         kunmap(lnb->page);
 615                 }
 616                 lnb->page->index = lnb->offset >> PAGE_SHIFT;
 617
 618                 cleanup_phase = 4;
 619
 620                 /* If the filter writes a partial page, then has the file
 621                  * extended, the client will read in the whole page.  the
 622                  * filter has to be careful to zero the rest of the partial
 623                  * page on disk.  we do it by hand for partial extending
 624                  * writes, send_bio() is responsible for zeroing pages when
 625                  * asked to read unmapped blocks -- brw_kiovec() does this. */
 626                 if (lnb->len != PAGE_SIZE) {
 627                         __s64 maxidx;
 628
 629                         maxidx = ((dentry->d_inode->i_size + PAGE_SIZE - 1) >>
 630                                  PAGE_SHIFT) - 1;
 631                         if (maxidx >= lnb->page->index) {
 632                                 LL_CDEBUG_PAGE(D_PAGE, lnb->page, "write %u @ "
 633                                                LPU64" flg %x before EOF %llu\n",
 634                                                lnb->len, lnb->offset,lnb->flags,
 635                                                dentry->d_inode->i_size);
 636                                 filter_iobuf_add_page(exp->exp_obd, iobuf,
 637                                                       dentry->d_inode,
 638                                                       lnb->page);
 639                         } else {
 640                                 long off;
 641                                 char *p = kmap(lnb->page);
 642
 643                                 off = lnb->offset & ~PAGE_MASK;
 644                                 if (off)
 645                                         memset(p, 0, off);
 646                                 off = (lnb->offset + lnb->len) & ~PAGE_MASK;
 647                                 if (off)
 648                                         memset(p + off, 0, PAGE_SIZE - off);
 649                                 kunmap(lnb->page);
 650                         }
 651                 }
 652                 if (lnb->rc == 0)
 653                         tot_bytes += lnb->len;
 654         }
 655
 656         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf, exp,
 657                               NULL, NULL, NULL);
 658
 659         fsfilt_check_slow(exp->exp_obd, now, obd_timeout, "start_page_write");
 660
 661         lprocfs_counter_add(exp->exp_obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
 662                             tot_bytes);
 663         EXIT;
 664 cleanup:
 665         switch(cleanup_phase) {
 666         case 4:
 667         case 3:
 668                 filter_iobuf_put(&exp->exp_obd->u.filter, iobuf, oti);
 669         case 2:
 670                 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 671                 if (rc)
 672                         f_dput(dentry);
 673                 break;
 674         case 1:
 675                 filter_iobuf_put(&exp->exp_obd->u.filter, iobuf, oti);
 676         case 0:
 677                 spin_lock(&exp->exp_obd->obd_osfs_lock);
 678                 if (oa)
 679                         filter_grant_incoming(exp, oa);
 680                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
 681                 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 682                 break;
 683         default:;
 684         }
 685         return rc;
 686 }
 687
 688 int filter_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
 689                   int objcount, struct obd_ioobj *obj, int niocount,
 690                   struct niobuf_remote *nb, struct niobuf_local *res,
 691                   struct obd_trans_info *oti, struct lustre_capa *capa)
 692 {
 693         if (cmd == OBD_BRW_WRITE)
 694                 return filter_preprw_write(cmd, exp, oa, objcount, obj,
 695                                            niocount, nb, res, oti, capa);
 696         if (cmd == OBD_BRW_READ)
 697                 return filter_preprw_read(cmd, exp, oa, objcount, obj,
 698                                           niocount, nb, res, oti, capa);
 699         LBUG();
 700         return -EPROTO;
 701 }
 702
 703 void filter_release_read_page(struct filter_obd *filter, struct inode *inode,
 704                               struct page *page)
 705 {
 706         int drop = 0;
 707
 708         if (inode != NULL &&
 709             (inode->i_size > filter->fo_readcache_max_filesize))
 710                 drop = 1;
 711
 712         /* drop from cache like truncate_list_pages() */
 713         if (drop && !TryLockPage(page)) {
 714                 if (page->mapping)
 715                         ll_truncate_complete_page(page);
 716                 unlock_page(page);
 717         }
 718         page_cache_release(page);
 719 }
 720
 721 static int filter_commitrw_read(struct obd_export *exp, struct obdo *oa,
 722                                 int objcount, struct obd_ioobj *obj,
 723                                 int niocount, struct niobuf_local *res,
 724                                 struct obd_trans_info *oti, int rc)
 725 {
 726         struct inode *inode = NULL;
 727         struct ldlm_res_id res_id = { .name = { obj->ioo_id, 0,
 728                                                 obj->ioo_gr, 0} };
 729         struct ldlm_resource *resource = NULL;
 730         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 731         ENTRY;
 732
 733         /* If oa != NULL then filter_preprw_read updated the inode atime
 734          * and we should update the lvb so that other glimpses will also
 735          * get the updated value. bug 5972 */
 736         if (oa && ns && ns->ns_lvbo && ns->ns_lvbo->lvbo_update) {
 737                 resource = ldlm_resource_get(ns, NULL, &res_id, LDLM_EXTENT, 0);
 738
 739                 if (resource != NULL) {
 740                         ns->ns_lvbo->lvbo_update(resource, NULL, 0, 1);
 741                         ldlm_resource_putref(resource);
 742                 }
 743         }
 744
 745         if (res->dentry != NULL)
 746                 inode = res->dentry->d_inode;
 747
 748         filter_free_dio_pages(objcount, obj, niocount, res);
 749
 750         if (res->dentry != NULL)
 751                 f_dput(res->dentry);
 752         RETURN(rc);
 753 }
 754
 755 void flip_into_page_cache(struct inode *inode, struct page *new_page)
 756 {
 757         struct page *old_page;
 758         int rc;
 759
 760         do {
 761                 /* the dlm is protecting us from read/write concurrency, so we
 762                  * expect this find_lock_page to return quickly.  even if we
 763                  * race with another writer it won't be doing much work with
 764                  * the page locked.  we do this 'cause t_c_p expects a
 765                  * locked page, and it wants to grab the pagecache lock
 766                  * as well. */
 767                 old_page = find_lock_page(inode->i_mapping, new_page->index);
 768                 if (old_page) {
 769                         ll_truncate_complete_page(old_page);
 770                         unlock_page(old_page);
 771                         page_cache_release(old_page);
 772                 }
 773
 774 #if 0 /* this should be a /proc tunable someday */
 775                 /* racing o_directs (no locking ioctl) could race adding
 776                  * their pages, so we repeat the page invalidation unless
 777                  * we successfully added our new page */
 778                 rc = add_to_page_cache_unique(new_page, inode->i_mapping,
 779                                               new_page->index,
 780                                               page_hash(inode->i_mapping,
 781                                                         new_page->index));
 782                 if (rc == 0) {
 783                         /* add_to_page_cache clears uptodate|dirty and locks
 784                          * the page */
 785                         SetPageUptodate(new_page);
 786                         unlock_page(new_page);
 787                 }
 788 #else
 789                 rc = 0;
 790 #endif
 791         } while (rc != 0);
 792 }
 793
 794 void filter_grant_commit(struct obd_export *exp, int niocount,
 795                          struct niobuf_local *res)
 796 {
 797         struct filter_obd *filter = &exp->exp_obd->u.filter;
 798         struct niobuf_local *lnb = res;
 799         unsigned long pending = 0;
 800         int i;
 801
 802         spin_lock(&exp->exp_obd->obd_osfs_lock);
 803         for (i = 0, lnb = res; i < niocount; i++, lnb++)
 804                 pending += lnb->lnb_grant_used;
 805
 806         LASSERTF(exp->exp_filter_data.fed_pending >= pending,
 807                  "%s: cli %s/%p fed_pending: %lu grant_used: %lu\n",
 808                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 809                  exp->exp_filter_data.fed_pending, pending);
 810         exp->exp_filter_data.fed_pending -= pending;
 811         LASSERTF(filter->fo_tot_granted >= pending,
 812                  "%s: cli %s/%p tot_granted: "LPU64" grant_used: %lu\n",
 813                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 814                  exp->exp_obd->u.filter.fo_tot_granted, pending);
 815         filter->fo_tot_granted -= pending;
 816         LASSERTF(filter->fo_tot_pending >= pending,
 817                  "%s: cli %s/%p tot_pending: "LPU64" grant_used: %lu\n",
 818                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 819                  filter->fo_tot_pending, pending);
 820         filter->fo_tot_pending -= pending;
 821
 822         spin_unlock(&exp->exp_obd->obd_osfs_lock);
 823 }
 824
 825 int filter_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
 826                     int objcount, struct obd_ioobj *obj, int niocount,
 827                     struct niobuf_local *res, struct obd_trans_info *oti,
 828                     int rc)
 829 {
 830         if (cmd == OBD_BRW_WRITE)
 831                 return filter_commitrw_write(exp, oa, objcount, obj, niocount,
 832                                              res, oti, rc);
 833         if (cmd == OBD_BRW_READ)
 834                 return filter_commitrw_read(exp, oa, objcount, obj, niocount,
 835                                             res, oti, rc);
 836         LBUG();
 837         return -EPROTO;
 838 }
 839
 840 int filter_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
 841                obd_count oa_bufs, struct brw_page *pga,
 842                struct obd_trans_info *oti)
 843 {
 844         struct obd_ioobj ioo;
 845         struct niobuf_local *lnb;
 846         struct niobuf_remote *rnb;
 847         obd_count i;
 848         int ret = 0;
 849         ENTRY;
 850
 851         OBD_ALLOC(lnb, oa_bufs * sizeof(struct niobuf_local));
 852         OBD_ALLOC(rnb, oa_bufs * sizeof(struct niobuf_remote));
 853
 854         if (lnb == NULL || rnb == NULL)
 855                 GOTO(out, ret = -ENOMEM);
 856
 857         for (i = 0; i < oa_bufs; i++) {
 858                 lnb[i].page = pga[i].pg;
 859                 rnb[i].offset = pga[i].off;
 860                 rnb[i].len = pga[i].count;
 861         }
 862
 863         obdo_to_ioobj(oinfo->oi_oa, &ioo);
 864         ioo.ioo_bufcnt = oa_bufs;
 865
 866         ret = filter_preprw(cmd, exp, oinfo->oi_oa, 1, &ioo,
 867                             oa_bufs, rnb, lnb, oti, oinfo_capa(oinfo));
 868         if (ret != 0)
 869                 GOTO(out, ret);
 870
 871         ret = filter_commitrw(cmd, exp, oinfo->oi_oa, 1, &ioo,
 872                               oa_bufs, lnb, oti, ret);
 873
 874 out:
 875         if (lnb)
 876                 OBD_FREE(lnb, oa_bufs * sizeof(struct niobuf_local));
 877         if (rnb)
 878                 OBD_FREE(rnb, oa_bufs * sizeof(struct niobuf_remote));
 879         RETURN(ret);
 880 }