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                               struct lustre_capa *capa)
 269 {
 270         struct obd_device *obd = exp->exp_obd;
 271         struct lvfs_run_ctxt saved;
 272         struct niobuf_remote *rnb;
 273         struct niobuf_local *lnb;
 274         struct dentry *dentry = NULL;
 275         struct inode *inode;
 276         void *iobuf = NULL;
 277         int rc = 0, i, tot_bytes = 0;
 278         unsigned long now = jiffies;
 279         ENTRY;
 280
 281         /* We are currently not supporting multi-obj BRW_READ RPCS at all.
 282          * When we do this function's dentry cleanup will need to be fixed */
 283         LASSERTF(objcount == 1, "%d\n", objcount);
 284         LASSERTF(obj->ioo_bufcnt > 0, "%d\n", obj->ioo_bufcnt);
 285
 286         if (oa && oa->o_valid & OBD_MD_FLGRANT) {
 287                 spin_lock(&obd->obd_osfs_lock);
 288                 filter_grant_incoming(exp, oa);
 289
 290                 oa->o_grant = 0;
 291                 spin_unlock(&obd->obd_osfs_lock);
 292         }
 293
 294         memset(res, 0, niocount * sizeof(*res));
 295
 296         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 297         rc = filter_alloc_iobuf(OBD_BRW_READ, obj->ioo_bufcnt, &iobuf);
 298         if (rc)
 299                 GOTO(cleanup, rc);
 300
 301         dentry = filter_id2dentry(obd, NULL, oa->o_gr, oa->o_id);
 302         if (IS_ERR(dentry))
 303                 GOTO(cleanup, rc = PTR_ERR(dentry));
 304
 305         inode = dentry->d_inode;
 306
 307         if (inode) {
 308                 rc = filter_verify_fid(exp, inode, capa);
 309                 if (rc)
 310                         GOTO(cleanup, rc);
 311         }
 312
 313         fsfilt_check_slow(now, obd_timeout, "preprw_read setup");
 314
 315         for (i = 0, lnb = res, rnb = nb; i < obj->ioo_bufcnt;
 316              i++, rnb++, lnb++) {
 317                 lnb->dentry = dentry;
 318                 lnb->offset = rnb->offset;
 319                 lnb->len    = rnb->len;
 320                 lnb->flags  = rnb->flags;
 321
 322                 if ((inode && inode->i_size <= rnb->offset) || inode == NULL)
 323                         /*
 324                          * if there's no more data, abort early.  lnb->page == *
 325                          * NULL and lnb->rc == 0, so it's easy to detect later.
 326                          */
 327                         break;
 328
 329                 rc = filter_alloc_dio_page(obd, inode, lnb);
 330                 if (rc) {
 331                         CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
 332                              "page err %u@"LPU64" %u/%u %p: rc %d\n",
 333                               lnb->len, lnb->offset, i, obj->ioo_bufcnt,
 334                               dentry, rc);
 335                         GOTO(cleanup, rc);
 336                 }
 337
 338                 if (inode->i_size < lnb->offset + lnb->len - 1)
 339                         lnb->rc = inode->i_size - lnb->offset;
 340                 else
 341                         lnb->rc = lnb->len;
 342
 343                 tot_bytes += lnb->rc;
 344
 345                 filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 346         }
 347
 348         fsfilt_check_slow(now, obd_timeout, "start_page_read");
 349
 350         if (inode != NULL) {
 351                 rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf,
 352                                       exp, NULL, NULL, NULL);
 353                 if (rc)
 354                         GOTO(cleanup, rc);
 355         }
 356
 357         lprocfs_counter_add(obd->obd_stats,
 358                             LPROC_FILTER_READ_BYTES, tot_bytes);
 359         filter_tally_read(&exp->exp_obd->u.filter, res, niocount);
 360
 361         EXIT;
 362 cleanup:
 363         if (rc) {
 364                 filter_free_dio_pages(objcount, obj,
 365                                       niocount, res);
 366                 /*
 367                  * in other cases (no errors) dentry is released in
 368                  * filter_commitrw_read().
 369                  */
 370                 f_dput(dentry);
 371         }
 372
 373         if (iobuf != NULL)
 374                 filter_free_iobuf(iobuf);
 375
 376         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 377
 378         if (rc)
 379                 CERROR("io error %d\n", rc);
 380
 381         return rc;
 382 }
 383
 384 /* When clients have dirtied as much space as they've been granted they
 385  * fall through to sync writes.  These sync writes haven't been expressed
 386  * in grants and need to error with ENOSPC when there isn't room in the
 387  * filesystem for them after grants are taken into account.  However,
 388  * writeback of the dirty data that was already granted space can write
 389  * right on through.
 390  *
 391  * Caller must hold obd_osfs_lock. */
 392 static int filter_grant_check(struct obd_export *exp, int objcount,
 393                               struct fsfilt_objinfo *fso, int niocount,
 394                               struct niobuf_remote *rnb,
 395                               struct niobuf_local *lnb, obd_size *left,
 396                               struct inode *inode)
 397 {
 398         struct filter_export_data *fed = &exp->exp_filter_data;
 399         int blocksize = exp->exp_obd->u.filter.fo_sb->s_blocksize;
 400         unsigned long used = 0, ungranted = 0, using;
 401         int i, rc = -ENOSPC, obj, n = 0, mask = D_CACHE;
 402
 403         LASSERT_SPIN_LOCKED(&exp->exp_obd->obd_osfs_lock);
 404
 405         for (obj = 0; obj < objcount; obj++) {
 406                 for (i = 0; i < fso[obj].fso_bufcnt; i++, n++) {
 407                         int tmp, bytes;
 408
 409                         /* should match code in osc_exit_cache() */
 410                         bytes = rnb[n].len;
 411                         bytes += rnb[n].offset & (blocksize - 1);
 412                         tmp = (rnb[n].offset + rnb[n].len) & (blocksize - 1);
 413                         if (tmp)
 414                                 bytes += blocksize - tmp;
 415
 416                         if (rnb[n].flags & OBD_BRW_FROM_GRANT) {
 417                                 if (fed->fed_grant < used + bytes) {
 418                                         CDEBUG(D_CACHE,
 419                                                "%s: cli %s/%p claims %ld+%d "
 420                                                "GRANT, real grant %lu idx %d\n",
 421                                                exp->exp_obd->obd_name,
 422                                                exp->exp_client_uuid.uuid, exp,
 423                                                used, bytes, fed->fed_grant, n);
 424                                         mask = D_ERROR;
 425                                 } else {
 426                                         used += bytes;
 427                                         rnb[n].flags |= OBD_BRW_GRANTED;
 428                                         lnb[n].lnb_grant_used = bytes;
 429                                         CDEBUG(0, "idx %d used=%lu\n", n, used);
 430                                         rc = 0;
 431                                         continue;
 432                                 }
 433                         }
 434                         if (*left > ungranted) {
 435                                 /* if enough space, pretend it was granted */
 436                                 ungranted += bytes;
 437                                 rnb[n].flags |= OBD_BRW_GRANTED;
 438                                 CDEBUG(0, "idx %d ungranted=%lu\n",n,ungranted);
 439                                 rc = 0;
 440                                 continue;
 441                         }
 442
 443                         /* We can't check for already-mapped blocks here, as
 444                          * it requires dropping the osfs lock to do the bmap.
 445                          * Instead, we return ENOSPC and in that case we need
 446                          * to go through and verify if all of the blocks not
 447                          * marked BRW_GRANTED are already mapped and we can
 448                          * ignore this error. */
 449                         lnb[n].rc = -ENOSPC;
 450                         rnb[n].flags &= OBD_BRW_GRANTED;
 451                         CDEBUG(D_CACHE,"%s: cli %s/%p idx %d no space for %d\n",
 452                                exp->exp_obd->obd_name,
 453                                exp->exp_client_uuid.uuid, exp, n, bytes);
 454                 }
 455         }
 456
 457         /* Now substract what client have used already.  We don't subtract
 458          * this from the tot_granted yet, so that other client's can't grab
 459          * that space before we have actually allocated our blocks.  That
 460          * happens in filter_grant_commit() after the writes are done. */
 461         *left -= ungranted;
 462         fed->fed_grant -= used;
 463         fed->fed_pending += used;
 464         exp->exp_obd->u.filter.fo_tot_pending += used;
 465
 466         CDEBUG(mask,
 467                "%s: cli %s/%p used: %lu ungranted: %lu grant: %lu dirty: %lu\n",
 468                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp, used,
 469                ungranted, fed->fed_grant, fed->fed_dirty);
 470
 471         /* Rough calc in case we don't refresh cached statfs data */
 472         using = (used + ungranted + 1 ) >>
 473                 exp->exp_obd->u.filter.fo_sb->s_blocksize_bits;
 474         if (exp->exp_obd->obd_osfs.os_bavail > using)
 475                 exp->exp_obd->obd_osfs.os_bavail -= using;
 476         else
 477                 exp->exp_obd->obd_osfs.os_bavail = 0;
 478
 479         if (fed->fed_dirty < used) {
 480                 CERROR("%s: cli %s/%p claims used %lu > fed_dirty %lu\n",
 481                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 482                        used, fed->fed_dirty);
 483                 used = fed->fed_dirty;
 484         }
 485         exp->exp_obd->u.filter.fo_tot_dirty -= used;
 486         fed->fed_dirty -= used;
 487
 488         if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
 489                 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
 490                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 491                        fed->fed_dirty, fed->fed_pending, fed->fed_grant);
 492                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
 493                 LBUG();
 494         }
 495         return rc;
 496 }
 497
 498 /* If we ever start to support multi-object BRW RPCs, we will need to get locks
 499  * on mulitple inodes.  That isn't all, because there still exists the
 500  * possibility of a truncate starting a new transaction while holding the ext3
 501  * rwsem = write while some writes (which have started their transactions here)
 502  * blocking on the ext3 rwsem = read => lock inversion.
 503  *
 504  * The handling gets very ugly when dealing with locked pages.  It may be easier
 505  * to just get rid of the locked page code (which has problems of its own) and
 506  * either discover we do not need it anymore (i.e. it was a symptom of another
 507  * bug) or ensure we get the page locks in an appropriate order. */
 508 static int filter_preprw_write(int cmd, struct obd_export *exp, struct obdo *oa,
 509                                int objcount, struct obd_ioobj *obj,
 510                                int niocount, struct niobuf_remote *nb,
 511                                struct niobuf_local *res,
 512                                struct obd_trans_info *oti,
 513                                struct lustre_capa *capa)
 514 {
 515         int rc = 0, i, tot_bytes = 0, cleanup_phase = 0;
 516         struct obd_device *obd = exp->exp_obd;
 517         struct niobuf_local *lnb = res;
 518         struct dentry *dentry = NULL;
 519         unsigned long now = jiffies;
 520         struct lvfs_run_ctxt saved;
 521         struct niobuf_remote *rnb;
 522         struct fsfilt_objinfo fso;
 523         obd_size left;
 524         obd_uid uid;
 525         obd_gid gid;
 526         void *iobuf;
 527
 528         ENTRY;
 529         LASSERT(objcount == 1);
 530         LASSERT(obj->ioo_bufcnt > 0);
 531
 532         memset(res, 0, niocount * sizeof(*res));
 533
 534         rc = filter_alloc_iobuf(OBD_BRW_READ, obj->ioo_bufcnt, &iobuf);
 535         if (rc)
 536                 GOTO(cleanup, rc);
 537         cleanup_phase = 1;
 538
 539         obd = exp->exp_obd;
 540         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 541
 542         uid = oa->o_valid & OBD_MD_FLUID ? oa->o_uid : 0;
 543         gid = oa->o_valid & OBD_MD_FLGID ? oa->o_gid : 0;
 544
 545         /* make sure that object is already allocated */
 546         dentry = filter_crow_object(obd, oa);
 547         if (IS_ERR(dentry))
 548                 GOTO(cleanup, rc = PTR_ERR(dentry));
 549
 550         rc = filter_verify_fid(exp, dentry->d_inode, capa);
 551         if (rc)
 552                 GOTO(cleanup, rc);
 553
 554         cleanup_phase = 2;
 555
 556         fso.fso_dentry = dentry;
 557         fso.fso_bufcnt = obj->ioo_bufcnt;
 558
 559         fsfilt_check_slow(now, obd_timeout, "preprw_write setup");
 560
 561         spin_lock(&obd->obd_osfs_lock);
 562         if (oa)
 563                 filter_grant_incoming(exp, oa);
 564
 565         cleanup_phase = 3;
 566
 567         left = filter_grant_space_left(exp);
 568
 569         rc = filter_grant_check(exp, objcount, &fso, niocount, nb, res,
 570                                 &left, dentry->d_inode);
 571         if (oa && oa->o_valid & OBD_MD_FLGRANT)
 572                 oa->o_grant = filter_grant(exp,oa->o_grant,oa->o_undirty,left);
 573
 574         /* We're finishing using body->oa as an input variable, so reset
 575          * o_valid here. */
 576         oa->o_valid = 0;
 577
 578         spin_unlock(&obd->obd_osfs_lock);
 579
 580         if (rc)
 581                 GOTO(cleanup, rc);
 582
 583         for (i = 0, rnb = nb, lnb = res; i < obj->ioo_bufcnt;
 584              i++, lnb++, rnb++) {
 585                 /* We still set up for ungranted pages so that granted pages
 586                  * can be written to disk as they were promised, and portals
 587                  * needs to keep the pages all aligned properly. */
 588                 lnb->dentry = dentry;
 589                 lnb->offset = rnb->offset;
 590                 lnb->len    = rnb->len;
 591                 lnb->flags  = rnb->flags;
 592
 593                 rc = filter_alloc_dio_page(obd, dentry->d_inode,lnb);
 594                 if (rc) {
 595                         CERROR("page err %u@"LPU64" %u/%u %p: rc %d\n",
 596                                lnb->len, lnb->offset,
 597                                i, obj->ioo_bufcnt, dentry, rc);
 598                         GOTO(cleanup, rc);
 599                 }
 600                 cleanup_phase = 4;
 601
 602                 /* If the filter writes a partial page, then has the file
 603                  * extended, the client will read in the whole page.  the
 604                  * filter has to be careful to zero the rest of the partial
 605                  * page on disk.  we do it by hand for partial extending
 606                  * writes, send_bio() is responsible for zeroing pages when
 607                  * asked to read unmapped blocks -- brw_kiovec() does this. */
 608                 if (lnb->len != PAGE_SIZE) {
 609                         if (lnb->offset + lnb->len < dentry->d_inode->i_size) {
 610                                 filter_iobuf_add_page(obd, iobuf, dentry->d_inode,
 611                                                       lnb->page);
 612                         } else {
 613                                 memset(kmap(lnb->page) + lnb->len, 0,
 614                                        PAGE_SIZE - lnb->len);
 615                                 kunmap(lnb->page);
 616                         }
 617                 }
 618                 if (lnb->rc == 0)
 619                         tot_bytes += lnb->len;
 620         }
 621
 622         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf, exp,
 623                               NULL, NULL, NULL);
 624
 625         fsfilt_check_slow(now, obd_timeout, "start_page_write");
 626
 627         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
 628                             tot_bytes);
 629         EXIT;
 630 cleanup:
 631         switch(cleanup_phase) {
 632         case 4:
 633                 if (rc)
 634                         filter_free_dio_pages(objcount, obj, niocount, res);
 635         case 3:
 636                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 637                 filter_free_iobuf(iobuf);
 638         case 2:
 639                 if (rc && dentry && !IS_ERR(dentry))
 640                         f_dput(dentry);
 641                 break;
 642         case 1:
 643                 spin_lock(&obd->obd_osfs_lock);
 644                 if (oa)
 645                         filter_grant_incoming(exp, oa);
 646                 spin_unlock(&obd->obd_osfs_lock);
 647                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 648                 filter_free_iobuf(iobuf);
 649                 break;
 650         default:;
 651
 652         }
 653         RETURN(rc);
 654 }
 655
 656 int filter_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
 657                   int objcount, struct obd_ioobj *obj, int niocount,
 658                   struct niobuf_remote *nb, struct niobuf_local *res,
 659                   struct obd_trans_info *oti, struct lustre_capa *capa)
 660 {
 661         int rc;
 662
 663         rc = filter_verify_capa(cmd, exp, capa);
 664         if (rc)
 665                 return rc;
 666
 667         if (cmd == OBD_BRW_WRITE)
 668                 return filter_preprw_write(cmd, exp, oa, objcount, obj,
 669                                            niocount, nb, res, oti, capa);
 670
 671         if (cmd == OBD_BRW_READ)
 672                 return filter_preprw_read(cmd, exp, oa, objcount, obj,
 673                                           niocount, nb, res, oti, capa);
 674
 675         LBUG();
 676         return -EPROTO;
 677 }
 678
 679 void filter_release_read_page(struct filter_obd *filter, struct inode *inode,
 680                               struct page *page)
 681 {
 682         int drop = 0;
 683
 684         if (inode != NULL &&
 685             (inode->i_size > filter->fo_readcache_max_filesize))
 686                 drop = 1;
 687
 688         /* drop from cache like truncate_list_pages() */
 689         if (drop && !TryLockPage(page)) {
 690                 if (page->mapping)
 691                         ll_truncate_complete_page(page);
 692                 unlock_page(page);
 693         }
 694         page_cache_release(page);
 695 }
 696
 697 static int filter_commitrw_read(struct obd_export *exp, struct obdo *oa,
 698                                 int objcount, struct obd_ioobj *obj,
 699                                 int niocount, struct niobuf_local *res,
 700                                 struct obd_trans_info *oti, int rc)
 701 {
 702         struct inode *inode = NULL;
 703         ENTRY;
 704
 705         if (res->dentry != NULL)
 706                 inode = res->dentry->d_inode;
 707
 708         filter_free_dio_pages(objcount, obj, niocount, res);
 709
 710         if (res->dentry != NULL)
 711                 f_dput(res->dentry);
 712         RETURN(rc);
 713 }
 714
 715 void flip_into_page_cache(struct inode *inode, struct page *new_page)
 716 {
 717         struct page *old_page;
 718         int rc;
 719
 720         do {
 721                 /* the dlm is protecting us from read/write concurrency, so we
 722                  * expect this find_lock_page to return quickly.  even if we
 723                  * race with another writer it won't be doing much work with
 724                  * the page locked.  we do this 'cause t_c_p expects a
 725                  * locked page, and it wants to grab the pagecache lock
 726                  * as well. */
 727                 old_page = find_lock_page(inode->i_mapping, new_page->index);
 728                 if (old_page) {
 729                         ll_truncate_complete_page(old_page);
 730                         unlock_page(old_page);
 731                         page_cache_release(old_page);
 732                 }
 733
 734 #if 0 /* this should be a /proc tunable someday */
 735                 /* racing o_directs (no locking ioctl) could race adding
 736                  * their pages, so we repeat the page invalidation unless
 737                  * we successfully added our new page */
 738                 rc = add_to_page_cache_unique(new_page, inode->i_mapping,
 739                                               new_page->index,
 740                                               page_hash(inode->i_mapping,
 741                                                         new_page->index));
 742                 if (rc == 0) {
 743                         /* add_to_page_cache clears uptodate|dirty and locks
 744                          * the page */
 745                         SetPageUptodate(new_page);
 746                         unlock_page(new_page);
 747                 }
 748 #else
 749                 rc = 0;
 750 #endif
 751         } while (rc != 0);
 752 }
 753
 754 void filter_grant_commit(struct obd_export *exp, int niocount,
 755                          struct niobuf_local *res)
 756 {
 757         struct filter_obd *filter = &exp->exp_obd->u.filter;
 758         struct niobuf_local *lnb = res;
 759         unsigned long pending = 0;
 760         int i;
 761
 762         spin_lock(&exp->exp_obd->obd_osfs_lock);
 763         for (i = 0, lnb = res; i < niocount; i++, lnb++)
 764                 pending += lnb->lnb_grant_used;
 765
 766         LASSERTF(exp->exp_filter_data.fed_pending >= pending,
 767                  "%s: cli %s/%p fed_pending: %lu grant_used: %lu\n",
 768                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 769                  exp->exp_filter_data.fed_pending, pending);
 770         exp->exp_filter_data.fed_pending -= pending;
 771         LASSERTF(filter->fo_tot_granted >= pending,
 772                  "%s: cli %s/%p tot_granted: "LPU64" grant_used: %lu\n",
 773                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 774                  exp->exp_obd->u.filter.fo_tot_granted, pending);
 775         filter->fo_tot_granted -= pending;
 776         LASSERTF(filter->fo_tot_pending >= pending,
 777                  "%s: cli %s/%p tot_pending: "LPU64" grant_used: %lu\n",
 778                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 779                  filter->fo_tot_pending, pending);
 780         filter->fo_tot_pending -= pending;
 781
 782         spin_unlock(&exp->exp_obd->obd_osfs_lock);
 783 }
 784 int filter_do_cow(struct obd_export *exp, struct obd_ioobj *obj,
 785                   int nioo, struct niobuf_remote *rnb)
 786 {
 787         struct dentry *dentry;
 788         struct lvfs_run_ctxt saved;
 789         struct write_extents *extents = NULL;
 790         int j, rc = 0, numexts = 0, flags = 0;
 791
 792         ENTRY;
 793
 794         LASSERT(nioo == 1);
 795
 796         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 797
 798         dentry = filter_id2dentry(exp->exp_obd, NULL, obj->ioo_gr,
 799                                   obj->ioo_id);
 800         if (IS_ERR(dentry)) {
 801                 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 802                 RETURN (PTR_ERR(dentry));
 803         }
 804
 805         if (dentry->d_inode == NULL) {
 806                 CERROR("trying to write extents to non-existent file "LPU64"\n",
 807                        obj->ioo_id);
 808                 GOTO(cleanup, rc = -ENOENT);
 809         }
 810
 811         flags = fsfilt_get_fs_flags(exp->exp_obd, dentry);
 812         if (!(flags & SM_DO_COW))
 813                 GOTO(cleanup, rc);
 814
 815         OBD_ALLOC(extents, obj->ioo_bufcnt * sizeof(struct write_extents));
 816         if (!extents) {
 817                 CERROR("No Memory\n");
 818                 GOTO(cleanup, rc = -ENOMEM);
 819         }
 820         for (j = 0; j < obj->ioo_bufcnt; j++) {
 821                 if (rnb[j].len != 0) {
 822                         extents[numexts].w_count = rnb[j].len;
 823                         extents[numexts].w_pos = rnb[j].offset;
 824                         numexts++;
 825                 }
 826         }
 827         rc = fsfilt_do_write_cow(exp->exp_obd, dentry, extents, numexts);
 828         if (rc) {
 829                 CERROR("Do cow error id "LPU64" rc:%d \n",
 830                         obj->ioo_id, rc);
 831                 GOTO(cleanup, rc);
 832         }
 833
 834 cleanup:
 835         if (extents) {
 836                 OBD_FREE(extents, obj->ioo_bufcnt * sizeof(struct write_extents));
 837         }
 838         f_dput(dentry);
 839         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 840         RETURN(rc);
 841
 842 }
 843
 844 int filter_write_extents(struct obd_export *exp, struct obd_ioobj *obj, int nobj,
 845                          int niocount, struct niobuf_local *local, int rc)
 846 {
 847         struct lvfs_run_ctxt saved;
 848         struct dentry *dentry;
 849         struct niobuf_local *lnb;
 850         __u64  offset = 0;
 851         __u32  len = 0;
 852         int    i, flags;
 853
 854         ENTRY;
 855
 856         LASSERT(nobj == 1);
 857
 858         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 859         dentry = filter_id2dentry(exp->exp_obd, NULL, obj->ioo_gr,
 860                                   obj->ioo_id);
 861         if (IS_ERR(dentry)) {
 862                 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 863                 RETURN (PTR_ERR(dentry));
 864         }
 865
 866         if (dentry->d_inode == NULL) {
 867                 CERROR("trying to write extents to non-existent file "LPU64"\n",
 868                        obj->ioo_id);
 869                 GOTO(cleanup, rc = -ENOENT);
 870         }
 871
 872         flags = fsfilt_get_fs_flags(exp->exp_obd, dentry);
 873         if (!(flags & SM_DO_REC))
 874                 GOTO(cleanup, rc);
 875
 876         for (i = 0, lnb = local; i < obj->ioo_bufcnt; i++, lnb++) {
 877                 if (len == 0) {
 878                         offset = lnb->offset;
 879                         len = lnb->len;
 880                 } else if (lnb->offset == (offset + len)) {
 881                         len += lnb->len;
 882                 } else {
 883                         rc = fsfilt_write_extents(exp->exp_obd, dentry,
 884                                                   offset, len);
 885                         if (rc) {
 886                                 CERROR("write exts off "LPU64" num %u rc:%d\n",
 887                                         offset, len, rc);
 888                                 GOTO(cleanup, rc);
 889                         }
 890                         offset = lnb->offset;
 891                         len = lnb->len;
 892                 }
 893         }
 894         if (len > 0) {
 895                 rc = fsfilt_write_extents(exp->exp_obd, dentry,
 896                                           offset, len);
 897                 if (rc) {
 898                         CERROR("write exts off "LPU64" num %u rc:%d\n",
 899                                 offset, len, rc);
 900                         GOTO(cleanup, rc);
 901                 }
 902         }
 903 cleanup:
 904         f_dput(dentry);
 905         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 906         RETURN(rc);
 907 }
 908
 909 int filter_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
 910                     int objcount, struct obd_ioobj *obj, int niocount,
 911                     struct niobuf_local *res, struct obd_trans_info *oti,int ret)
 912 {
 913         int rc = -EPROTO;
 914         /* remove that audit handlers due to fsfilt_inode_map_pages hooks
 915         //struct lustre_id *id = obdo_id(oa);
 916         //__u32 len = sizeof(*id);
 917         //struct inode * inode = res->dentry->d_inode;
 918         //struct super_block * sb = res->dentry->d_sb;
 919         //struct obd_device *obd = class_exp2obd(exp);
 920         */
 921         current->user->nid = oti->oti_nid;
 922
 923         if (cmd == OBD_BRW_WRITE) {
 924                 rc = filter_commitrw_write(exp, oa, objcount, obj, niocount,
 925                                            res, oti, ret);
 926                 /* fsfilt_set_info(obd, sb, inode, 10, "file_write", len, (void*)id); */
 927         }
 928         else if (cmd == OBD_BRW_READ) {
 929                 rc = filter_commitrw_read(exp, oa, objcount, obj, niocount,
 930                                           res, oti, ret);
 931                 /* fsfilt_set_info(obd, sb, inode, 9, "file_read", len, (void*)id); */
 932         }
 933         else
 934                 LBUG();
 935
 936         return rc;
 937 }
 938
 939 int filter_brw(int cmd, struct obd_export *exp, struct obdo *oa,
 940                struct lov_stripe_md *lsm, obd_count oa_bufs,
 941                struct brw_page *pga, struct obd_trans_info *oti)
 942 {
 943         struct obd_ioobj ioo;
 944         struct niobuf_local *lnb;
 945         struct niobuf_remote *rnb;
 946         obd_count i;
 947         int ret = 0;
 948         ENTRY;
 949
 950         OBD_ALLOC(lnb, oa_bufs * sizeof(struct niobuf_local));
 951         OBD_ALLOC(rnb, oa_bufs * sizeof(struct niobuf_remote));
 952
 953         if (lnb == NULL || rnb == NULL)
 954                 GOTO(out, ret = -ENOMEM);
 955
 956         for (i = 0; i < oa_bufs; i++) {
 957                 rnb[i].offset = pga[i].disk_offset;
 958                 rnb[i].len = pga[i].count;
 959         }
 960
 961         obdo_to_ioobj(oa, &ioo);
 962         ioo.ioo_bufcnt = oa_bufs;
 963
 964         ret = filter_preprw(cmd, exp, oa, 1, &ioo, oa_bufs, rnb,
 965                             lnb, oti, NULL);
 966         if (ret != 0)
 967                 GOTO(out, ret);
 968
 969         for (i = 0; i < oa_bufs; i++) {
 970                 void *virt;
 971                 obd_off off;
 972                 void *addr;
 973
 974                 if (lnb[i].page == NULL)
 975                         break;
 976
 977                 off = pga[i].disk_offset & ~PAGE_MASK;
 978                 virt = kmap(pga[i].pg);
 979                 addr = kmap(lnb[i].page);
 980
 981                 /* 2 kmaps == vanishingly small deadlock opportunity */
 982
 983                 if (cmd & OBD_BRW_WRITE)
 984                         memcpy(addr + off, virt + off, pga[i].count);
 985                 else
 986                         memcpy(virt + off, addr + off, pga[i].count);
 987
 988                 kunmap(lnb[i].page);
 989                 kunmap(pga[i].pg);
 990         }
 991
 992         ret = filter_commitrw(cmd, exp, oa, 1, &ioo, oa_bufs, lnb, oti, ret);
 993
 994 out:
 995         if (lnb)
 996                 OBD_FREE(lnb, oa_bufs * sizeof(struct niobuf_local));
 997         if (rnb)
 998                 OBD_FREE(rnb, oa_bufs * sizeof(struct niobuf_remote));
 999         RETURN(ret);
1000 }