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 Lustre, http://www.lustre.org.
  12  *
  13  *   Lustre is free software; you can redistribute it and/or
  14  *   modify it under the terms of version 2 of the GNU General Public
  15  *   License as published by the Free Software Foundation.
  16  *
  17  *   Lustre is distributed in the hope that it will be useful,
  18  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  19  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  20  *   GNU General Public License for more details.
  21  *
  22  *   You should have received a copy of the GNU General Public License
  23  *   along with Lustre; if not, write to the Free Software
  24  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  25  */
  26
  27 #define DEBUG_SUBSYSTEM S_FILTER
  28
  29 #include <linux/config.h>
  30 #include <linux/module.h>
  31 #include <linux/pagemap.h> // XXX kill me soon
  32 #include <linux/version.h>
  33
  34 #include <linux/obd_class.h>
  35 #include <linux/lustre_fsfilt.h>
  36 #include <linux/lustre_smfs.h>
  37 #include <linux/lustre_snap.h>
  38 #include "filter_internal.h"
  39
  40 int *obdfilter_created_scratchpad;
  41
  42 static int filter_alloc_dio_page(struct obd_device *obd, struct inode *inode,
  43                                  struct niobuf_local *lnb)
  44
  45 {
  46         struct page *page;
  47         ENTRY;
  48
  49         page = alloc_pages(GFP_HIGHUSER, 0);
  50         if (page == NULL) {
  51                 CERROR("no memory for a temp page\n");
  52                 lnb->rc = -ENOMEM;
  53                 RETURN(-ENOMEM);
  54         }
  55
  56 #if 0
  57         POISON_PAGE(page, 0xf1);
  58         if (lnb->len != PAGE_SIZE) {
  59                 memset(kmap(page) + lnb->len, 0, PAGE_SIZE - lnb->len);
  60                 kunmap(page);
  61         }
  62 #endif
  63         page->index = lnb->offset >> PAGE_SHIFT;
  64
  65         lnb->page = page;
  66
  67         RETURN(0);
  68 }
  69
  70 void filter_free_dio_pages(int objcount, struct obd_ioobj *obj,
  71                            int niocount, struct niobuf_local *res)
  72 {
  73         int i, j;
  74
  75         for (i = 0; i < objcount; i++, obj++) {
  76                 for (j = 0 ; j < obj->ioo_bufcnt ; j++, res++) {
  77                         if (res->page != NULL) {
  78                                 __free_page(res->page);
  79                                 res->page = NULL;
  80                         }
  81                 }
  82         }
  83 }
  84
  85 /* Grab the dirty and seen grant announcements from the incoming obdo.
  86  * We will later calculate the clients new grant and return it.
  87  * Caller must hold osfs lock */
  88 static void filter_grant_incoming(struct obd_export *exp, struct obdo *oa)
  89 {
  90         struct filter_export_data *fed;
  91         struct obd_device *obd = exp->exp_obd;
  92         static unsigned long last_msg;
  93         static int last_count;
  94         int mask = D_CACHE;
  95         ENTRY;
  96
  97         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
  98
  99         if ((oa->o_valid & (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) !=
 100                                         (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) {
 101                 oa->o_valid &= ~OBD_MD_FLGRANT;
 102                 EXIT;
 103                 return;
 104         }
 105
 106         fed = &exp->exp_filter_data;
 107
 108         /* Don't print this to the console the first time it happens, since
 109          * it can happen legitimately on occasion, but only rarely. */
 110         if (time_after(jiffies, last_msg + 60 * HZ)) {
 111                 last_count = 0;
 112                 last_msg = jiffies;
 113         }
 114         if ((last_count & (-last_count)) == last_count)
 115                 mask = D_WARNING;
 116         last_count++;
 117
 118         /* Add some margin, since there is a small race if other RPCs arrive
 119          * out-or-order and have already consumed some grant.  We want to
 120          * leave this here in case there is a large error in accounting. */
 121         CDEBUG(oa->o_grant > fed->fed_grant + FILTER_GRANT_CHUNK ? mask:D_CACHE,
 122                "%s: cli %s/%p reports grant: "LPU64" dropped: %u, local: %lu\n",
 123                obd->obd_name, exp->exp_client_uuid.uuid, exp, oa->o_grant,
 124                oa->o_dropped, fed->fed_grant);
 125
 126         /* Update our accounting now so that statfs takes it into account.
 127          * Note that fed_dirty is only approximate and can become incorrect
 128          * if RPCs arrive out-of-order.  No important calculations depend
 129          * on fed_dirty however. */
 130         obd->u.filter.fo_tot_dirty += oa->o_dirty - fed->fed_dirty;
 131         if (fed->fed_grant < oa->o_dropped) {
 132                 CERROR("%s: cli %s/%p reports %u dropped > fed_grant %lu\n",
 133                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
 134                        oa->o_dropped, fed->fed_grant);
 135                 oa->o_dropped = 0;
 136         }
 137         if (obd->u.filter.fo_tot_granted < oa->o_dropped) {
 138                 CERROR("%s: cli %s/%p reports %u dropped > tot_grant "LPU64"\n",
 139                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
 140                        oa->o_dropped, obd->u.filter.fo_tot_granted);
 141                 oa->o_dropped = 0;
 142         }
 143         obd->u.filter.fo_tot_granted -= oa->o_dropped;
 144         fed->fed_grant -= oa->o_dropped;
 145         fed->fed_dirty = oa->o_dirty;
 146         EXIT;
 147 }
 148
 149 #define GRANT_FOR_LLOG(obd) 16
 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.filter.fo_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 (time_before(obd->obd_osfs_age, jiffies - HZ)) {
 166 restat:
 167                 rc = fsfilt_statfs(obd, obd->u.filter.fo_sb, jiffies + 1);
 168                 if (rc) /* N.B. statfs can't really fail */
 169                         RETURN(0);
 170                 statfs_done = 1;
 171         }
 172
 173         avail = obd->obd_osfs.os_bavail;
 174         left = avail - (avail >> (blockbits - 3)); /* (d)indirect */
 175         if (left > GRANT_FOR_LLOG(obd)) {
 176                 left = (left - GRANT_FOR_LLOG(obd)) << blockbits;
 177         } else {
 178                 left = 0 /* << blockbits */;
 179         }
 180
 181         if (!statfs_done && left < 32 * FILTER_GRANT_CHUNK + tot_granted) {
 182                 CDEBUG(D_CACHE, "fs has no space left and statfs too old\n");
 183                 goto restat;
 184         }
 185
 186         if (left >= tot_granted) {
 187                 left -= tot_granted;
 188         } else {
 189                 static unsigned long next;
 190                 if (left < tot_granted - obd->u.filter.fo_tot_pending &&
 191                     time_after(jiffies, next)) {
 192                         spin_unlock(&obd->obd_osfs_lock);
 193                         CERROR("%s: cli %s/%p grant "LPU64" > available "
 194                                LPU64" and pending "LPU64"\n", obd->obd_name,
 195                                exp->exp_client_uuid.uuid, exp, tot_granted,
 196                                left, obd->u.filter.fo_tot_pending);
 197                         if (next == 0)
 198                                 portals_debug_dumplog();
 199                         next = jiffies + 20 * HZ;
 200                         spin_lock(&obd->obd_osfs_lock);
 201                 }
 202                 left = 0;
 203         }
 204
 205         CDEBUG(D_CACHE, "%s: cli %s/%p free: "LPU64" avail: "LPU64" grant "LPU64
 206                " left: "LPU64" pending: "LPU64"\n", obd->obd_name,
 207                exp->exp_client_uuid.uuid, exp,
 208                obd->obd_osfs.os_bfree << blockbits, avail << blockbits,
 209                tot_granted, left, obd->u.filter.fo_tot_pending);
 210
 211         return left;
 212 }
 213
 214 /* Calculate how much grant space to allocate to this client, based on how
 215  * much space is currently free and how much of that is already granted.
 216  *
 217  * Caller must hold obd_osfs_lock. */
 218 long filter_grant(struct obd_export *exp, obd_size current_grant,
 219                   obd_size want, obd_size fs_space_left)
 220 {
 221         struct obd_device *obd = exp->exp_obd;
 222         struct filter_export_data *fed = &exp->exp_filter_data;
 223         int blockbits = obd->u.filter.fo_sb->s_blocksize_bits;
 224         __u64 grant = 0;
 225
 226         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
 227
 228         /* Grant some fraction of the client's requested grant space so that
 229          * they are not always waiting for write credits (not all of it to
 230          * avoid overgranting in face of multiple RPCs in flight).  This
 231          * essentially will be able to control the OSC_MAX_RIF for a client.
 232          *
 233          * If we do have a large disparity between what the client thinks it
 234          * has and what we think it has, don't grant very much and let the
 235          * client consume its grant first.  Either it just has lots of RPCs
 236          * in flight, or it was evicted and its grants will soon be used up. */
 237         if (current_grant < want &&
 238             current_grant < fed->fed_grant + FILTER_GRANT_CHUNK) {
 239                 grant = min((want >> blockbits) / 2,
 240                             (fs_space_left >> blockbits) / 8);
 241                 grant <<= blockbits;
 242
 243                 if (grant) {
 244                         if (grant > FILTER_GRANT_CHUNK)
 245                                 grant = FILTER_GRANT_CHUNK;
 246
 247                         obd->u.filter.fo_tot_granted += grant;
 248                         fed->fed_grant += grant;
 249                 }
 250         }
 251
 252         CDEBUG(D_CACHE,"%s: cli %s/%p wants: "LPU64" granting: "LPU64"\n",
 253                obd->obd_name, exp->exp_client_uuid.uuid, exp, want, grant);
 254         CDEBUG(D_CACHE,
 255                "%s: cli %s/%p tot cached:"LPU64" granted:"LPU64
 256                " num_exports: %d\n", obd->obd_name, exp->exp_client_uuid.uuid,
 257                exp, obd->u.filter.fo_tot_dirty,
 258                obd->u.filter.fo_tot_granted, obd->obd_num_exports);
 259
 260         return grant;
 261 }
 262
 263 static int filter_preprw_read(int cmd, struct obd_export *exp, struct obdo *oa,
 264                               int objcount, struct obd_ioobj *obj,
 265                               int niocount, struct niobuf_remote *nb,
 266                               struct niobuf_local *res,
 267                               struct obd_trans_info *oti)
 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         LASSERTF(objcount == 1, "%d\n", objcount);
 283         LASSERTF(obj->ioo_bufcnt > 0, "%d\n", obj->ioo_bufcnt);
 284
 285         if (oa && oa->o_valid & OBD_MD_FLGRANT) {
 286                 spin_lock(&obd->obd_osfs_lock);
 287                 filter_grant_incoming(exp, oa);
 288
 289                 oa->o_grant = 0;
 290                 spin_unlock(&obd->obd_osfs_lock);
 291         }
 292
 293         memset(res, 0, niocount * sizeof(*res));
 294
 295         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 296         rc = filter_alloc_iobuf(OBD_BRW_READ, obj->ioo_bufcnt, &iobuf);
 297         if (rc)
 298                 GOTO(cleanup, rc);
 299
 300         dentry = filter_id2dentry(obd, NULL, oa->o_gr, oa->o_id);
 301         if (IS_ERR(dentry))
 302                 GOTO(cleanup, rc = PTR_ERR(dentry));
 303
 304         inode = dentry->d_inode;
 305
 306         fsfilt_check_slow(now, obd_timeout, "preprw_read setup");
 307
 308         for (i = 0, lnb = res, rnb = nb; i < obj->ioo_bufcnt;
 309              i++, rnb++, lnb++) {
 310                 lnb->dentry = dentry;
 311                 lnb->offset = rnb->offset;
 312                 lnb->len    = rnb->len;
 313                 lnb->flags  = rnb->flags;
 314
 315                 if ((inode && inode->i_size <= rnb->offset) || inode == NULL)
 316                         /*
 317                          * if there's no more data, abort early.  lnb->page == *
 318                          * NULL and lnb->rc == 0, so it's easy to detect later.
 319                          */
 320                         break;
 321
 322                 rc = filter_alloc_dio_page(obd, inode, lnb);
 323                 if (rc) {
 324                         CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
 325                              "page err %u@"LPU64" %u/%u %p: rc %d\n",
 326                               lnb->len, lnb->offset, i, obj->ioo_bufcnt,
 327                               dentry, rc);
 328                         GOTO(cleanup, rc);
 329                 }
 330
 331                 if (inode->i_size < lnb->offset + lnb->len - 1)
 332                         lnb->rc = inode->i_size - lnb->offset;
 333                 else
 334                         lnb->rc = lnb->len;
 335
 336                 tot_bytes += lnb->rc;
 337
 338                 filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 339         }
 340
 341         fsfilt_check_slow(now, obd_timeout, "start_page_read");
 342
 343         if (inode != NULL) {
 344                 rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf,
 345                                       exp, NULL, NULL, NULL);
 346                 if (rc)
 347                         GOTO(cleanup, rc);
 348         }
 349
 350         lprocfs_counter_add(obd->obd_stats,
 351                             LPROC_FILTER_READ_BYTES, tot_bytes);
 352         filter_tally_read(&exp->exp_obd->u.filter, res, niocount);
 353
 354         EXIT;
 355 cleanup:
 356         if (rc) {
 357                 filter_free_dio_pages(objcount, obj,
 358                                       niocount, res);
 359                 /*
 360                  * in other cases (no errors) dentry is released in
 361                  * filter_commitrw_read().
 362                  */
 363                 f_dput(dentry);
 364         }
 365
 366         if (iobuf != NULL)
 367                 filter_free_iobuf(iobuf);
 368
 369         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 370
 371         if (rc)
 372                 CERROR("io error %d\n", rc);
 373
 374         return rc;
 375 }
 376
 377 /* When clients have dirtied as much space as they've been granted they
 378  * fall through to sync writes.  These sync writes haven't been expressed
 379  * in grants and need to error with ENOSPC when there isn't room in the
 380  * filesystem for them after grants are taken into account.  However,
 381  * writeback of the dirty data that was already granted space can write
 382  * right on through.
 383  *
 384  * Caller must hold obd_osfs_lock. */
 385 static int filter_grant_check(struct obd_export *exp, int objcount,
 386                               struct fsfilt_objinfo *fso, int niocount,
 387                               struct niobuf_remote *rnb,
 388                               struct niobuf_local *lnb, obd_size *left,
 389                               struct inode *inode)
 390 {
 391         struct filter_export_data *fed = &exp->exp_filter_data;
 392         int blocksize = exp->exp_obd->u.filter.fo_sb->s_blocksize;
 393         unsigned long used = 0, ungranted = 0, using;
 394         int i, rc = -ENOSPC, obj, n = 0, mask = D_CACHE;
 395
 396         LASSERT_SPIN_LOCKED(&exp->exp_obd->obd_osfs_lock);
 397
 398         for (obj = 0; obj < objcount; obj++) {
 399                 for (i = 0; i < fso[obj].fso_bufcnt; i++, n++) {
 400                         int tmp, bytes;
 401
 402                         /* FIXME: this is calculated with PAGE_SIZE on client */
 403                         bytes = rnb[n].len;
 404                         bytes += rnb[n].offset & (blocksize - 1);
 405                         tmp = (rnb[n].offset + rnb[n].len) & (blocksize - 1);
 406                         if (tmp)
 407                                 bytes += blocksize - tmp;
 408
 409                         if (rnb[n].flags & OBD_BRW_FROM_GRANT) {
 410                                 if (fed->fed_grant < used + bytes) {
 411                                         CDEBUG(D_CACHE,
 412                                                "%s: cli %s/%p claims %ld+%d "
 413                                                "GRANT, real grant %lu idx %d\n",
 414                                                exp->exp_obd->obd_name,
 415                                                exp->exp_client_uuid.uuid, exp,
 416                                                used, bytes, fed->fed_grant, n);
 417                                         mask = D_ERROR;
 418                                 } else {
 419                                         used += bytes;
 420                                         rnb[n].flags |= OBD_BRW_GRANTED;
 421                                         lnb[n].lnb_grant_used = bytes;
 422                                         CDEBUG(0, "idx %d used=%lu\n", n, used);
 423                                         rc = 0;
 424                                         continue;
 425                                 }
 426                         }
 427                         if (*left > ungranted) {
 428                                 /* if enough space, pretend it was granted */
 429                                 ungranted += bytes;
 430                                 rnb[n].flags |= OBD_BRW_GRANTED;
 431                                 CDEBUG(0, "idx %d ungranted=%lu\n",n,ungranted);
 432                                 rc = 0;
 433                                 continue;
 434                         }
 435
 436                         /* We can't check for already-mapped blocks here, as
 437                          * it requires dropping the osfs lock to do the bmap.
 438                          * Instead, we return ENOSPC and in that case we need
 439                          * to go through and verify if all of the blocks not
 440                          * marked BRW_GRANTED are already mapped and we can
 441                          * ignore this error. */
 442                         lnb[n].rc = -ENOSPC;
 443                         rnb[n].flags &= OBD_BRW_GRANTED;
 444                         CDEBUG(D_CACHE,"%s: cli %s/%p idx %d no space for %d\n",
 445                                exp->exp_obd->obd_name,
 446                                exp->exp_client_uuid.uuid, exp, n, bytes);
 447                 }
 448         }
 449
 450         /* Now substract what client have used already.  We don't subtract
 451          * this from the tot_granted yet, so that other client's can't grab
 452          * that space before we have actually allocated our blocks.  That
 453          * happens in filter_grant_commit() after the writes are done. */
 454         *left -= ungranted;
 455         fed->fed_grant -= used;
 456         fed->fed_pending += used;
 457         exp->exp_obd->u.filter.fo_tot_pending += used;
 458
 459         CDEBUG(mask,
 460                "%s: cli %s/%p used: %lu ungranted: %lu grant: %lu dirty: %lu\n",
 461                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp, used,
 462                ungranted, fed->fed_grant, fed->fed_dirty);
 463
 464         /* Rough calc in case we don't refresh cached statfs data */
 465         using = (used + ungranted + 1 ) >>
 466                 exp->exp_obd->u.filter.fo_sb->s_blocksize_bits;
 467         if (exp->exp_obd->obd_osfs.os_bavail > using)
 468                 exp->exp_obd->obd_osfs.os_bavail -= using;
 469         else
 470                 exp->exp_obd->obd_osfs.os_bavail = 0;
 471
 472         if (fed->fed_dirty < used) {
 473                 CERROR("%s: cli %s/%p claims used %lu > fed_dirty %lu\n",
 474                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 475                        used, fed->fed_dirty);
 476                 used = fed->fed_dirty;
 477         }
 478         exp->exp_obd->u.filter.fo_tot_dirty -= used;
 479         fed->fed_dirty -= used;
 480
 481         return rc;
 482 }
 483
 484 /* If we ever start to support multi-object BRW RPCs, we will need to get locks
 485  * on mulitple inodes.  That isn't all, because there still exists the
 486  * possibility of a truncate starting a new transaction while holding the ext3
 487  * rwsem = write while some writes (which have started their transactions here)
 488  * blocking on the ext3 rwsem = read => lock inversion.
 489  *
 490  * The handling gets very ugly when dealing with locked pages.  It may be easier
 491  * to just get rid of the locked page code (which has problems of its own) and
 492  * either discover we do not need it anymore (i.e. it was a symptom of another
 493  * bug) or ensure we get the page locks in an appropriate order. */
 494 static int filter_preprw_write(int cmd, struct obd_export *exp, struct obdo *oa,
 495                                int objcount, struct obd_ioobj *obj,
 496                                int niocount, struct niobuf_remote *nb,
 497                                struct niobuf_local *res,
 498                                struct obd_trans_info *oti)
 499 {
 500         int rc = 0, i, tot_bytes = 0, cleanup_phase = 0;
 501         struct niobuf_local *lnb = res;
 502         struct dentry *dentry = NULL;
 503         unsigned long now = jiffies;
 504         struct lvfs_run_ctxt saved;
 505         struct niobuf_remote *rnb;
 506         struct fsfilt_objinfo fso;
 507         struct obd_device *obd;
 508         obd_size left;
 509         obd_uid uid;
 510         obd_gid gid;
 511         void *iobuf;
 512
 513         ENTRY;
 514         LASSERT(objcount == 1);
 515         LASSERT(obj->ioo_bufcnt > 0);
 516
 517         memset(res, 0, niocount * sizeof(*res));
 518
 519         rc = filter_alloc_iobuf(OBD_BRW_READ, obj->ioo_bufcnt, &iobuf);
 520         if (rc)
 521                 GOTO(cleanup, rc);
 522         cleanup_phase = 1;
 523
 524         obd = exp->exp_obd;
 525         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 526
 527         uid = oa->o_valid & OBD_MD_FLUID ? oa->o_uid : 0;
 528         gid = oa->o_valid & OBD_MD_FLGID ? oa->o_gid : 0;
 529
 530         /* make sure that object is already allocated */
 531         dentry = filter_crow_object(obd, oa);
 532         if (IS_ERR(dentry))
 533                 GOTO(cleanup, rc = PTR_ERR(dentry));
 534
 535         cleanup_phase = 2;
 536
 537         fso.fso_dentry = dentry;
 538         fso.fso_bufcnt = obj->ioo_bufcnt;
 539
 540         fsfilt_check_slow(now, obd_timeout, "preprw_write setup");
 541
 542         spin_lock(&obd->obd_osfs_lock);
 543         if (oa)
 544                 filter_grant_incoming(exp, oa);
 545
 546         cleanup_phase = 3;
 547
 548         left = filter_grant_space_left(exp);
 549
 550         rc = filter_grant_check(exp, objcount, &fso, niocount, nb, res,
 551                                 &left, dentry->d_inode);
 552         if (oa && oa->o_valid & OBD_MD_FLGRANT)
 553                 oa->o_grant = filter_grant(exp,oa->o_grant,oa->o_undirty,left);
 554
 555         /* We're finishing using body->oa as an input variable, so reset
 556          * o_valid here. */
 557         oa->o_valid = 0;
 558
 559         spin_unlock(&obd->obd_osfs_lock);
 560
 561         if (rc)
 562                 GOTO(cleanup, rc);
 563
 564         for (i = 0, rnb = nb, lnb = res; i < obj->ioo_bufcnt;
 565              i++, lnb++, rnb++) {
 566                 /* We still set up for ungranted pages so that granted pages
 567                  * can be written to disk as they were promised, and portals
 568                  * needs to keep the pages all aligned properly. */
 569                 lnb->dentry = dentry;
 570                 lnb->offset = rnb->offset;
 571                 lnb->len    = rnb->len;
 572                 lnb->flags  = rnb->flags;
 573
 574                 rc = filter_alloc_dio_page(obd, dentry->d_inode,lnb);
 575                 if (rc) {
 576                         CERROR("page err %u@"LPU64" %u/%u %p: rc %d\n",
 577                                lnb->len, lnb->offset,
 578                                i, obj->ioo_bufcnt, dentry, rc);
 579                         GOTO(cleanup, rc);
 580                 }
 581                 cleanup_phase = 4;
 582
 583                 /* If the filter writes a partial page, then has the file
 584                  * extended, the client will read in the whole page.  the
 585                  * filter has to be careful to zero the rest of the partial
 586                  * page on disk.  we do it by hand for partial extending
 587                  * writes, send_bio() is responsible for zeroing pages when
 588                  * asked to read unmapped blocks -- brw_kiovec() does this. */
 589                 if (lnb->len != PAGE_SIZE) {
 590                         if (lnb->offset + lnb->len < dentry->d_inode->i_size) {
 591                                 filter_iobuf_add_page(obd, iobuf, dentry->d_inode,
 592                                                       lnb->page);
 593                         } else {
 594                                 memset(kmap(lnb->page) + lnb->len, 0,
 595                                        PAGE_SIZE - lnb->len);
 596                                 kunmap(lnb->page);
 597                         }
 598                 }
 599                 if (lnb->rc == 0)
 600                         tot_bytes += lnb->len;
 601         }
 602
 603         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf, exp,
 604                               NULL, NULL, NULL);
 605
 606         fsfilt_check_slow(now, obd_timeout, "start_page_write");
 607
 608         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
 609                             tot_bytes);
 610         EXIT;
 611 cleanup:
 612         switch(cleanup_phase) {
 613         case 4:
 614                 if (rc)
 615                         filter_free_dio_pages(objcount, obj, niocount, res);
 616         case 3:
 617                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 618                 filter_free_iobuf(iobuf);
 619         case 2:
 620                 if (rc && dentry && !IS_ERR(dentry))
 621                         f_dput(dentry);
 622                 break;
 623         case 1:
 624                 spin_lock(&obd->obd_osfs_lock);
 625                 if (oa)
 626                         filter_grant_incoming(exp, oa);
 627                 spin_unlock(&obd->obd_osfs_lock);
 628                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 629                 filter_free_iobuf(iobuf);
 630                 break;
 631         default:;
 632
 633         }
 634         RETURN(rc);
 635 }
 636
 637 int filter_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
 638                   int objcount, struct obd_ioobj *obj, int niocount,
 639                   struct niobuf_remote *nb, struct niobuf_local *res,
 640                   struct obd_trans_info *oti, struct lustre_capa *capa)
 641 {
 642         int rc;
 643
 644         rc = filter_verify_capa(cmd, exp, capa);
 645         if (rc)
 646                 return rc;
 647
 648         if (cmd == OBD_BRW_WRITE)
 649                 return filter_preprw_write(cmd, exp, oa, objcount, obj,
 650                                            niocount, nb, res, oti);
 651
 652         if (cmd == OBD_BRW_READ)
 653                 return filter_preprw_read(cmd, exp, oa, objcount, obj,
 654                                           niocount, nb, res, oti);
 655
 656         LBUG();
 657         return -EPROTO;
 658 }
 659
 660 void filter_release_read_page(struct filter_obd *filter, struct inode *inode,
 661                               struct page *page)
 662 {
 663         int drop = 0;
 664
 665         if (inode != NULL &&
 666             (inode->i_size > filter->fo_readcache_max_filesize))
 667                 drop = 1;
 668
 669         /* drop from cache like truncate_list_pages() */
 670         if (drop && !TryLockPage(page)) {
 671                 if (page->mapping)
 672                         ll_truncate_complete_page(page);
 673                 unlock_page(page);
 674         }
 675         page_cache_release(page);
 676 }
 677
 678 static int filter_commitrw_read(struct obd_export *exp, struct obdo *oa,
 679                                 int objcount, struct obd_ioobj *obj,
 680                                 int niocount, struct niobuf_local *res,
 681                                 struct obd_trans_info *oti, int rc)
 682 {
 683         struct inode *inode = NULL;
 684         ENTRY;
 685
 686         if (res->dentry != NULL)
 687                 inode = res->dentry->d_inode;
 688
 689         filter_free_dio_pages(objcount, obj, niocount, res);
 690
 691         if (res->dentry != NULL)
 692                 f_dput(res->dentry);
 693         RETURN(rc);
 694 }
 695
 696 void flip_into_page_cache(struct inode *inode, struct page *new_page)
 697 {
 698         struct page *old_page;
 699         int rc;
 700
 701         do {
 702                 /* the dlm is protecting us from read/write concurrency, so we
 703                  * expect this find_lock_page to return quickly.  even if we
 704                  * race with another writer it won't be doing much work with
 705                  * the page locked.  we do this 'cause t_c_p expects a
 706                  * locked page, and it wants to grab the pagecache lock
 707                  * as well. */
 708                 old_page = find_lock_page(inode->i_mapping, new_page->index);
 709                 if (old_page) {
 710                         ll_truncate_complete_page(old_page);
 711                         unlock_page(old_page);
 712                         page_cache_release(old_page);
 713                 }
 714
 715 #if 0 /* this should be a /proc tunable someday */
 716                 /* racing o_directs (no locking ioctl) could race adding
 717                  * their pages, so we repeat the page invalidation unless
 718                  * we successfully added our new page */
 719                 rc = add_to_page_cache_unique(new_page, inode->i_mapping,
 720                                               new_page->index,
 721                                               page_hash(inode->i_mapping,
 722                                                         new_page->index));
 723                 if (rc == 0) {
 724                         /* add_to_page_cache clears uptodate|dirty and locks
 725                          * the page */
 726                         SetPageUptodate(new_page);
 727                         unlock_page(new_page);
 728                 }
 729 #else
 730                 rc = 0;
 731 #endif
 732         } while (rc != 0);
 733 }
 734
 735 void filter_grant_commit(struct obd_export *exp, int niocount,
 736                          struct niobuf_local *res)
 737 {
 738         struct filter_obd *filter = &exp->exp_obd->u.filter;
 739         struct niobuf_local *lnb = res;
 740         unsigned long pending = 0;
 741         int i;
 742
 743         spin_lock(&exp->exp_obd->obd_osfs_lock);
 744         for (i = 0, lnb = res; i < niocount; i++, lnb++)
 745                 pending += lnb->lnb_grant_used;
 746
 747         LASSERTF(exp->exp_filter_data.fed_pending >= pending,
 748                  "%s: cli %s/%p fed_pending: %lu grant_used: %lu\n",
 749                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 750                  exp->exp_filter_data.fed_pending, pending);
 751         exp->exp_filter_data.fed_pending -= pending;
 752         LASSERTF(filter->fo_tot_granted >= pending,
 753                  "%s: cli %s/%p tot_granted: "LPU64" grant_used: %lu\n",
 754                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 755                  exp->exp_obd->u.filter.fo_tot_granted, pending);
 756         filter->fo_tot_granted -= pending;
 757         LASSERTF(filter->fo_tot_pending >= pending,
 758                  "%s: cli %s/%p tot_pending: "LPU64" grant_used: %lu\n",
 759                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 760                  filter->fo_tot_pending, pending);
 761         filter->fo_tot_pending -= pending;
 762
 763         spin_unlock(&exp->exp_obd->obd_osfs_lock);
 764 }
 765 int filter_do_cow(struct obd_export *exp, struct obd_ioobj *obj,
 766                   int nioo, struct niobuf_remote *rnb)
 767 {
 768         struct dentry *dentry;
 769         struct lvfs_run_ctxt saved;
 770         struct write_extents *extents = NULL;
 771         int j, rc = 0, numexts = 0, flags = 0;
 772
 773         ENTRY;
 774
 775         LASSERT(nioo == 1);
 776
 777         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 778
 779         dentry = filter_id2dentry(exp->exp_obd, NULL, obj->ioo_gr,
 780                                   obj->ioo_id);
 781         if (IS_ERR(dentry)) {
 782                 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 783                 RETURN (PTR_ERR(dentry));
 784         }
 785
 786         if (dentry->d_inode == NULL) {
 787                 CERROR("trying to write extents to non-existent file "LPU64"\n",
 788                        obj->ioo_id);
 789                 GOTO(cleanup, rc = -ENOENT);
 790         }
 791
 792         flags = fsfilt_get_fs_flags(exp->exp_obd, dentry);
 793         if (!(flags & SM_DO_COW)) {
 794                 GOTO(cleanup, rc);
 795         }
 796         OBD_ALLOC(extents, obj->ioo_bufcnt * sizeof(struct write_extents));
 797         if (!extents) {
 798                 CERROR("No Memory\n");
 799                 GOTO(cleanup, rc = -ENOMEM);
 800         }
 801         for (j = 0; j < obj->ioo_bufcnt; j++) {
 802                 if (rnb[j].len != 0) {
 803                         extents[numexts].w_count = rnb[j].len;
 804                         extents[numexts].w_pos = rnb[j].offset;
 805                         numexts++;
 806                 }
 807         }
 808         rc = fsfilt_do_write_cow(exp->exp_obd, dentry, extents, numexts);
 809         if (rc) {
 810                 CERROR("Do cow error id "LPU64" rc:%d \n",
 811                         obj->ioo_id, rc);
 812                 GOTO(cleanup, rc);
 813         }
 814
 815 cleanup:
 816         if (extents) {
 817                 OBD_FREE(extents, obj->ioo_bufcnt * sizeof(struct write_extents));
 818         }
 819         f_dput(dentry);
 820         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 821         RETURN(rc);
 822
 823 }
 824 int filter_write_extents(struct obd_export *exp, struct obd_ioobj *obj, int nobj,
 825                          int niocount, struct niobuf_local *local, int rc)
 826 {
 827         struct lvfs_run_ctxt saved;
 828         struct dentry *dentry;
 829         struct niobuf_local *lnb;
 830         __u64  offset = 0;
 831         __u32  len = 0;
 832         int    i, flags;
 833
 834         ENTRY;
 835
 836         LASSERT(nobj == 1);
 837
 838         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 839         dentry = filter_id2dentry(exp->exp_obd, NULL, obj->ioo_gr,
 840                                   obj->ioo_id);
 841         if (IS_ERR(dentry)) {
 842                 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 843                 RETURN (PTR_ERR(dentry));
 844         }
 845
 846         if (dentry->d_inode == NULL) {
 847                 CERROR("trying to write extents to non-existent file "LPU64"\n",
 848                        obj->ioo_id);
 849                 GOTO(cleanup, rc = -ENOENT);
 850         }
 851
 852         flags = fsfilt_get_fs_flags(exp->exp_obd, dentry);
 853         if (!(flags & SM_DO_REC)) {
 854                 GOTO(cleanup, rc);
 855         }
 856
 857         for (i = 0, lnb = local; i < obj->ioo_bufcnt; i++, lnb++) {
 858                 if (len == 0) {
 859                         offset = lnb->offset;
 860                         len = lnb->len;
 861                 } else if (lnb->offset == (offset + len)) {
 862                         len += lnb->len;
 863                 } else {
 864                         rc = fsfilt_write_extents(exp->exp_obd, dentry,
 865                                                   offset, len);
 866                         if (rc) {
 867                                 CERROR("write exts off "LPU64" num %u rc:%d\n",
 868                                         offset, len, rc);
 869                                 GOTO(cleanup, rc);
 870                         }
 871                         offset = lnb->offset;
 872                         len = lnb->len;
 873                 }
 874         }
 875         if (len > 0) {
 876                 rc = fsfilt_write_extents(exp->exp_obd, dentry,
 877                                           offset, len);
 878                 if (rc) {
 879                         CERROR("write exts off "LPU64" num %u rc:%d\n",
 880                                 offset, len, rc);
 881                         GOTO(cleanup, rc);
 882                 }
 883         }
 884 cleanup:
 885         f_dput(dentry);
 886         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 887         RETURN(rc);
 888 }
 889
 890 int filter_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
 891                     int objcount, struct obd_ioobj *obj, int niocount,
 892                     struct niobuf_local *res, struct obd_trans_info *oti,int ret)
 893 {
 894         int rc = -EPROTO;
 895         struct lustre_id *id = obdo_id(oa);
 896         __u32 len = sizeof(*id);
 897         struct inode * inode = res->dentry->d_inode;
 898         struct super_block * sb = res->dentry->d_sb;
 899         struct obd_device *obd = class_exp2obd(exp);
 900
 901         if (cmd == OBD_BRW_WRITE) {
 902                 rc = filter_commitrw_write(exp, oa, objcount, obj, niocount,
 903                                            res, oti, ret);
 904                 fsfilt_set_info(obd, sb, inode, 10, "file_write", len, (void*)id);
 905         }
 906         else if (cmd == OBD_BRW_READ) {
 907                 rc = filter_commitrw_read(exp, oa, objcount, obj, niocount,
 908                                           res, oti, ret);
 909                 fsfilt_set_info(obd, sb, inode, 9, "file_read", len, (void*)id);
 910         }
 911         else
 912                 LBUG();
 913
 914         return rc;
 915 }
 916
 917 int filter_brw(int cmd, struct obd_export *exp, struct obdo *oa,
 918                struct lov_stripe_md *lsm, obd_count oa_bufs,
 919                struct brw_page *pga, struct obd_trans_info *oti)
 920 {
 921         struct obd_ioobj ioo;
 922         struct niobuf_local *lnb;
 923         struct niobuf_remote *rnb;
 924         obd_count i;
 925         int ret = 0;
 926         ENTRY;
 927
 928         OBD_ALLOC(lnb, oa_bufs * sizeof(struct niobuf_local));
 929         OBD_ALLOC(rnb, oa_bufs * sizeof(struct niobuf_remote));
 930
 931         if (lnb == NULL || rnb == NULL)
 932                 GOTO(out, ret = -ENOMEM);
 933
 934         for (i = 0; i < oa_bufs; i++) {
 935                 rnb[i].offset = pga[i].disk_offset;
 936                 rnb[i].len = pga[i].count;
 937         }
 938
 939         obdo_to_ioobj(oa, &ioo);
 940         ioo.ioo_bufcnt = oa_bufs;
 941
 942         ret = filter_preprw(cmd, exp, oa, 1, &ioo, oa_bufs, rnb, lnb, oti,NULL);
 943         if (ret != 0)
 944                 GOTO(out, ret);
 945
 946         for (i = 0; i < oa_bufs; i++) {
 947                 void *virt;
 948                 obd_off off;
 949                 void *addr;
 950
 951                 if (lnb[i].page == NULL)
 952                         break;
 953
 954                 off = pga[i].disk_offset & ~PAGE_MASK;
 955                 virt = kmap(pga[i].pg);
 956                 addr = kmap(lnb[i].page);
 957
 958                 /* 2 kmaps == vanishingly small deadlock opportunity */
 959
 960                 if (cmd & OBD_BRW_WRITE)
 961                         memcpy(addr + off, virt + off, pga[i].count);
 962                 else
 963                         memcpy(virt + off, addr + off, pga[i].count);
 964
 965                 kunmap(lnb[i].page);
 966                 kunmap(pga[i].pg);
 967         }
 968
 969         ret = filter_commitrw(cmd, exp, oa, 1, &ioo, oa_bufs, lnb, oti, ret);
 970
 971 out:
 972         if (lnb)
 973                 OBD_FREE(lnb, oa_bufs * sizeof(struct niobuf_local));
 974         if (rnb)
 975                 OBD_FREE(rnb, oa_bufs * sizeof(struct niobuf_remote));
 976         RETURN(ret);
 977 }