lustre/obdfilter/filter_io.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  */
  30 /*
  31  * This file is part of Lustre, http://www.lustre.org/
  32  * Lustre is a trademark of Sun Microsystems, Inc.
  33  *
  34  * lustre/obdfilter/filter_io.c
  35  *
  36  * Author: Peter Braam <braam@clusterfs.com>
  37  * Author: Andreas Dilger <adilger@clusterfs.com>
  38  * Author: Phil Schwan <phil@clusterfs.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_FILTER
  42
  43 #include <linux/module.h>
  44 #include <linux/pagemap.h> // XXX kill me soon
  45 #include <linux/version.h>
  46
  47 #include <obd_class.h>
  48 #include <obd_ost.h>
  49 #include <lustre_fsfilt.h>
  50 #include "filter_internal.h"
  51
  52 int *obdfilter_created_scratchpad;
  53
  54 /* Grab the dirty and seen grant announcements from the incoming obdo.
  55  * We will later calculate the clients new grant and return it.
  56  * Caller must hold osfs lock */
  57 void filter_grant_incoming(struct obd_export *exp, struct obdo *oa)
  58 {
  59         struct filter_export_data *fed;
  60         struct obd_device *obd = exp->exp_obd;
  61         ENTRY;
  62
  63         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
  64
  65         if ((oa->o_valid & (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) !=
  66                                         (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) {
  67                 oa->o_valid &= ~OBD_MD_FLGRANT;
  68                 EXIT;
  69                 return;
  70         }
  71
  72         fed = &exp->exp_filter_data;
  73
  74         /* Add some margin, since there is a small race if other RPCs arrive
  75          * out-or-order and have already consumed some grant.  We want to
  76          * leave this here in case there is a large error in accounting. */
  77         CDEBUG(D_CACHE,
  78                "%s: cli %s/%p reports grant: "LPU64" dropped: %u, local: %lu\n",
  79                obd->obd_name, exp->exp_client_uuid.uuid, exp, oa->o_grant,
  80                oa->o_dropped, fed->fed_grant);
  81
  82         /* Update our accounting now so that statfs takes it into account.
  83          * Note that fed_dirty is only approximate and can become incorrect
  84          * if RPCs arrive out-of-order.  No important calculations depend
  85          * on fed_dirty however, but we must check sanity to not assert. */
  86         if ((long long)oa->o_dirty < 0)
  87                 oa->o_dirty = 0;
  88         else if (oa->o_dirty > fed->fed_grant + 4 * FILTER_GRANT_CHUNK)
  89                 oa->o_dirty = fed->fed_grant + 4 * FILTER_GRANT_CHUNK;
  90         obd->u.filter.fo_tot_dirty += oa->o_dirty - fed->fed_dirty;
  91         if (fed->fed_grant < oa->o_dropped) {
  92                 CDEBUG(D_CACHE,"%s: cli %s/%p reports %u dropped > grant %lu\n",
  93                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
  94                        oa->o_dropped, fed->fed_grant);
  95                 oa->o_dropped = 0;
  96         }
  97         if (obd->u.filter.fo_tot_granted < oa->o_dropped) {
  98                 CERROR("%s: cli %s/%p reports %u dropped > tot_grant "LPU64"\n",
  99                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
 100                        oa->o_dropped, obd->u.filter.fo_tot_granted);
 101                 oa->o_dropped = 0;
 102         }
 103         obd->u.filter.fo_tot_granted -= oa->o_dropped;
 104         fed->fed_grant -= oa->o_dropped;
 105         fed->fed_dirty = oa->o_dirty;
 106
 107         if (oa->o_valid & OBD_MD_FLFLAGS && oa->o_flags & OBD_FL_SHRINK_GRANT) {
 108                 obd_size left_space = filter_grant_space_left(exp);
 109                 struct filter_obd *filter = &exp->exp_obd->u.filter;
 110
 111                 /*Only if left_space < fo_tot_clients * 32M,
 112                  *then the grant space could be shrinked */
 113                 if (left_space < filter->fo_tot_granted_clients *
 114                                  FILTER_GRANT_SHRINK_LIMIT) {
 115                         fed->fed_grant -= oa->o_grant;
 116                         filter->fo_tot_granted -= oa->o_grant;
 117                         CDEBUG(D_CACHE, "%s: cli %s/%p shrink "LPU64
 118                                "fed_grant %ld total "LPU64"\n",
 119                                obd->obd_name, exp->exp_client_uuid.uuid,
 120                                exp, oa->o_grant, fed->fed_grant,
 121                                filter->fo_tot_granted);
 122                         oa->o_grant = 0;
 123                 }
 124         }
 125
 126         if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
 127                 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
 128                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
 129                        fed->fed_dirty, fed->fed_pending, fed->fed_grant);
 130                 cfs_spin_unlock(&obd->obd_osfs_lock);
 131                 LBUG();
 132         }
 133         EXIT;
 134 }
 135
 136 /* Figure out how much space is available between what we've granted
 137  * and what remains in the filesystem.  Compensate for ext3 indirect
 138  * block overhead when computing how much free space is left ungranted.
 139  *
 140  * Caller must hold obd_osfs_lock. */
 141 obd_size filter_grant_space_left(struct obd_export *exp)
 142 {
 143         struct obd_device *obd = exp->exp_obd;
 144         int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
 145         obd_size tot_granted = obd->u.filter.fo_tot_granted, avail, left = 0;
 146         int rc, statfs_done = 0;
 147
 148         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
 149
 150         if (cfs_time_before_64(obd->obd_osfs_age,
 151                                cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS))) {
 152 restat:
 153                 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb,
 154                                    cfs_time_shift_64(OBD_STATFS_CACHE_SECONDS));
 155                 if (rc) /* N.B. statfs can't really fail */
 156                         RETURN(0);
 157                 statfs_done = 1;
 158         }
 159
 160         avail = obd->obd_osfs.os_bavail;
 161         left = avail - (avail >> (blockbits - 3)); /* (d)indirect */
 162         if (left > GRANT_FOR_LLOG(obd)) {
 163                 left = (left - GRANT_FOR_LLOG(obd)) << blockbits;
 164         } else {
 165                 left = 0 /* << blockbits */;
 166         }
 167
 168         if (!statfs_done && left < 32 * FILTER_GRANT_CHUNK + tot_granted) {
 169                 CDEBUG(D_CACHE, "fs has no space left and statfs too old\n");
 170                 goto restat;
 171         }
 172
 173         if (left >= tot_granted) {
 174                 left -= tot_granted;
 175         } else {
 176                 if (left < tot_granted - obd->u.filter.fo_tot_pending) {
 177                         CERROR("%s: cli %s/%p grant "LPU64" > available "
 178                                LPU64" and pending "LPU64"\n", obd->obd_name,
 179                                exp->exp_client_uuid.uuid, exp, tot_granted,
 180                                left, obd->u.filter.fo_tot_pending);
 181                 }
 182                 left = 0;
 183         }
 184
 185         CDEBUG(D_CACHE, "%s: cli %s/%p free: "LPU64" avail: "LPU64" grant "LPU64
 186                " left: "LPU64" pending: "LPU64"\n", obd->obd_name,
 187                exp->exp_client_uuid.uuid, exp,
 188                obd->obd_osfs.os_bfree << blockbits, avail << blockbits,
 189                tot_granted, left, obd->u.filter.fo_tot_pending);
 190
 191         return left;
 192 }
 193
 194 /* Calculate how much grant space to allocate to this client, based on how
 195  * much space is currently free and how much of that is already granted.
 196  *
 197  * if @conservative != 0, we limit the maximum grant to FILTER_GRANT_CHUNK;
 198  * otherwise we'll satisfy the requested amount as possible as we can, this
 199  * is usually due to client reconnect.
 200  *
 201  * Caller must hold obd_osfs_lock. */
 202 long filter_grant(struct obd_export *exp, obd_size current_grant,
 203                   obd_size want, obd_size fs_space_left, int conservative)
 204 {
 205         struct obd_device *obd = exp->exp_obd;
 206         struct filter_export_data *fed = &exp->exp_filter_data;
 207         int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
 208         __u64 grant = 0;
 209
 210         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
 211
 212         /* Grant some fraction of the client's requested grant space so that
 213          * they are not always waiting for write credits (not all of it to
 214          * avoid overgranting in face of multiple RPCs in flight).  This
 215          * essentially will be able to control the OSC_MAX_RIF for a client.
 216          *
 217          * If we do have a large disparity between what the client thinks it
 218          * has and what we think it has, don't grant very much and let the
 219          * client consume its grant first.  Either it just has lots of RPCs
 220          * in flight, or it was evicted and its grants will soon be used up. */
 221         if (want > 0x7fffffff) {
 222                 CERROR("%s: client %s/%p requesting > 2GB grant "LPU64"\n",
 223                        obd->obd_name, exp->exp_client_uuid.uuid, exp, want);
 224         } else if (current_grant < want &&
 225                    current_grant < fed->fed_grant + FILTER_GRANT_CHUNK) {
 226                 grant = min(want + (1 << blockbits) - 1, fs_space_left / 8);
 227                 grant &= ~((1ULL << blockbits) - 1);
 228
 229                 if (grant) {
 230                         if (grant > FILTER_GRANT_CHUNK && conservative)
 231                                 grant = FILTER_GRANT_CHUNK;
 232
 233                         obd->u.filter.fo_tot_granted += grant;
 234                         fed->fed_grant += grant;
 235                         if (fed->fed_grant < 0) {
 236                                 CERROR("%s: cli %s/%p grant %ld want "LPU64
 237                                        "current"LPU64"\n",
 238                                        obd->obd_name, exp->exp_client_uuid.uuid,
 239                                        exp, fed->fed_grant, want,current_grant);
 240                                 cfs_spin_unlock(&obd->obd_osfs_lock);
 241                                 LBUG();
 242                         }
 243                 }
 244         }
 245
 246         CDEBUG(D_CACHE,
 247                "%s: cli %s/%p wants: "LPU64" current grant "LPU64
 248                " granting: "LPU64"\n", obd->obd_name, exp->exp_client_uuid.uuid,
 249                exp, want, current_grant, grant);
 250         CDEBUG(D_CACHE,
 251                "%s: cli %s/%p tot cached:"LPU64" granted:"LPU64
 252                " num_exports: %d\n", obd->obd_name, exp->exp_client_uuid.uuid,
 253                exp, obd->u.filter.fo_tot_dirty,
 254                obd->u.filter.fo_tot_granted, obd->obd_num_exports);
 255
 256         return grant;
 257 }
 258
 259 /*
 260  * the routine is used to request pages from pagecache
 261  *
 262  * use GFP_NOFS for requests from a local client not allowing to enter FS
 263  * as we might end up waiting on a page he sent in the request we're serving.
 264  * use __GFP_HIGHMEM so that the pages can use all of the available memory
 265  * on 32-bit machines
 266  * use more aggressive GFP_HIGHUSER flags from non-local clients to be able to
 267  * generate more memory pressure.
 268  *
 269  * See Bug 19529 and Bug 19917 for details.
 270  */
 271 static struct page *filter_get_page(struct obd_device *obd,
 272                                     struct inode *inode,
 273                                     obd_off offset,
 274                                     int localreq)
 275 {
 276         struct page *page;
 277
 278         page = find_or_create_page(inode->i_mapping, offset >> CFS_PAGE_SHIFT,
 279                                    (localreq ? (GFP_NOFS | __GFP_HIGHMEM)
 280                                              : GFP_HIGHUSER));
 281         if (unlikely(page == NULL))
 282                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_NO_PAGE, 1);
 283
 284         return page;
 285 }
 286
 287 /*
 288  * the routine initializes array of local_niobuf from remote_niobuf
 289  */
 290 static int filter_map_remote_to_local(int objcount, struct obd_ioobj *obj,
 291                                       struct niobuf_remote *nb,
 292                                       int *nrpages, struct niobuf_local *res)
 293 {
 294         struct niobuf_remote *rnb;
 295         struct niobuf_local *lnb;
 296         int i, max;
 297         ENTRY;
 298
 299         /* we don't support multiobject RPC yet
 300          * ost_brw_read() and ost_brw_write() check this */
 301         LASSERT(objcount == 1);
 302
 303         max = *nrpages;
 304         *nrpages = 0;
 305         for (i = 0, rnb = nb, lnb = res; i < obj->ioo_bufcnt; i++, rnb++) {
 306                 obd_off offset = rnb->offset;
 307                 unsigned int len = rnb->len;
 308
 309                 while (len > 0) {
 310                         int poff = offset & (CFS_PAGE_SIZE - 1);
 311                         int plen = CFS_PAGE_SIZE - poff;
 312
 313                         if (*nrpages >= max) {
 314                                 CERROR("small array of local bufs: %d\n", max);
 315                                 RETURN(-EINVAL);
 316                         }
 317
 318                         if (plen > len)
 319                                 plen = len;
 320                         lnb->offset = offset;
 321                         lnb->len = plen;
 322                         lnb->flags = rnb->flags;
 323                         lnb->page = NULL;
 324                         lnb->rc = 0;
 325                         lnb->lnb_grant_used = 0;
 326
 327                         LASSERTF(plen <= len, "plen %u, len %u\n", plen, len);
 328                         offset += plen;
 329                         len -= plen;
 330                         lnb++;
 331                         (*nrpages)++;
 332                 }
 333         }
 334         RETURN(0);
 335 }
 336
 337 /*
 338  * the invalidate above doesn't work during read because lnet pins pages.
 339  * The truncate is used here instead to drop pages from cache
 340  */
 341 void filter_release_cache(struct obd_device *obd, struct obd_ioobj *obj,
 342                           struct niobuf_remote *rnb, struct inode *inode)
 343 {
 344         int i;
 345
 346         LASSERT(inode != NULL);
 347         for (i = 0; i < obj->ioo_bufcnt; i++, rnb++) {
 348                 /* remove pages in which range is fit */
 349                 truncate_inode_pages_range(inode->i_mapping,
 350                                            rnb->offset & CFS_PAGE_MASK,
 351                                            (rnb->offset + rnb->len - 1) |
 352                                            ~CFS_PAGE_MASK);
 353         }
 354 }
 355
 356 static int filter_preprw_read(int cmd, struct obd_export *exp, struct obdo *oa,
 357                               int objcount, struct obd_ioobj *obj,
 358                               struct niobuf_remote *nb,
 359                               int *npages, struct niobuf_local *res,
 360                               struct obd_trans_info *oti,
 361                               struct lustre_capa *capa)
 362 {
 363         struct obd_device *obd = exp->exp_obd;
 364         struct timeval start, end;
 365         struct lvfs_run_ctxt saved;
 366         struct niobuf_local *lnb;
 367         struct dentry *dentry = NULL;
 368         struct inode *inode = NULL;
 369         void *iobuf = NULL;
 370         int rc = 0, i, tot_bytes = 0;
 371         unsigned long now = jiffies;
 372         long timediff;
 373         loff_t isize;
 374         ENTRY;
 375
 376         /* We are currently not supporting multi-obj BRW_READ RPCS at all.
 377          * When we do this function's dentry cleanup will need to be fixed.
 378          * These values are verified in ost_brw_write() from the wire. */
 379         LASSERTF(objcount == 1, "%d\n", objcount);
 380         LASSERTF(obj->ioo_bufcnt > 0, "%d\n", obj->ioo_bufcnt);
 381
 382         rc = filter_auth_capa(exp, NULL, oa->o_seq, capa,
 383                               CAPA_OPC_OSS_READ);
 384         if (rc)
 385                 RETURN(rc);
 386
 387         if (oa && oa->o_valid & OBD_MD_FLGRANT) {
 388                 cfs_spin_lock(&obd->obd_osfs_lock);
 389                 filter_grant_incoming(exp, oa);
 390
 391                 if (!(oa->o_valid & OBD_MD_FLFLAGS) ||
 392                     !(oa->o_flags & OBD_FL_SHRINK_GRANT))
 393                         oa->o_grant = 0;
 394                 cfs_spin_unlock(&obd->obd_osfs_lock);
 395         }
 396
 397         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 398         if (IS_ERR(iobuf))
 399                 RETURN(PTR_ERR(iobuf));
 400
 401         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 402         dentry = filter_oa2dentry(obd, &oa->o_oi);
 403         if (IS_ERR(dentry)) {
 404                 rc = PTR_ERR(dentry);
 405                 dentry = NULL;
 406                 GOTO(cleanup, rc);
 407         }
 408
 409         inode = dentry->d_inode;
 410         isize = i_size_read(inode);
 411
 412         obdo_to_inode(inode, oa, OBD_MD_FLATIME);
 413
 414         rc = filter_map_remote_to_local(objcount, obj, nb, npages, res);
 415         if (rc)
 416                 GOTO(cleanup, rc);
 417
 418         fsfilt_check_slow(obd, now, "preprw_read setup");
 419
 420         /* find pages for all segments, fill array with them */
 421         cfs_gettimeofday(&start);
 422         for (i = 0, lnb = res; i < *npages; i++, lnb++) {
 423
 424                 lnb->dentry = dentry;
 425
 426                 if (isize <= lnb->offset)
 427                         /* If there's no more data, abort early.  lnb->rc == 0,
 428                          * so it's easy to detect later. */
 429                         break;
 430
 431                 lnb->page = filter_get_page(obd, inode, lnb->offset, 0);
 432                 if (lnb->page == NULL)
 433                         GOTO(cleanup, rc = -ENOMEM);
 434
 435                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS, 1);
 436
 437                 if (isize < lnb->offset + lnb->len - 1)
 438                         lnb->rc = isize - lnb->offset;
 439                 else
 440                         lnb->rc = lnb->len;
 441
 442                 tot_bytes += lnb->rc;
 443
 444                 if (PageUptodate(lnb->page)) {
 445                         lprocfs_counter_add(obd->obd_stats,
 446                                             LPROC_FILTER_CACHE_HIT, 1);
 447                         continue;
 448                 }
 449
 450                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_MISS, 1);
 451                 filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 452         }
 453         cfs_gettimeofday(&end);
 454         timediff = cfs_timeval_sub(&end, &start, NULL);
 455         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
 456
 457         if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
 458                 GOTO(cleanup, rc = -ENOMEM);
 459
 460         fsfilt_check_slow(obd, now, "start_page_read");
 461
 462         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf,
 463                               exp, NULL, NULL, NULL);
 464         if (rc)
 465                 GOTO(cleanup, rc);
 466
 467         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_READ_BYTES, tot_bytes);
 468
 469         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
 470                 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
 471                                     LPROC_FILTER_READ_BYTES, tot_bytes);
 472         filter_counter_incr(exp, LPROC_FILTER_STATS_READ,
 473                             oti ? oti->oti_jobid : NULL, tot_bytes);
 474         EXIT;
 475
 476  cleanup:
 477         /* unlock pages to allow access from concurrent OST_READ */
 478         for (i = 0, lnb = res; i < *npages; i++, lnb++) {
 479                 if (lnb->page) {
 480                         LASSERT(PageLocked(lnb->page));
 481                         unlock_page(lnb->page);
 482
 483                         if (rc) {
 484                                 page_cache_release(lnb->page);
 485                                 lnb->page = NULL;
 486                         }
 487                 }
 488         }
 489
 490         if (rc != 0) {
 491                 if (dentry != NULL)
 492                         f_dput(dentry);
 493         }
 494
 495         filter_iobuf_put(&obd->u.filter, iobuf, oti);
 496
 497         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 498         if (rc)
 499                 CERROR("io error %d\n", rc);
 500
 501         return rc;
 502 }
 503
 504 /* When clients have dirtied as much space as they've been granted they
 505  * fall through to sync writes.  These sync writes haven't been expressed
 506  * in grants and need to error with ENOSPC when there isn't room in the
 507  * filesystem for them after grants are taken into account.  However,
 508  * writeback of the dirty data that was already granted space can write
 509  * right on through.
 510  *
 511  * Caller must hold obd_osfs_lock. */
 512 static int filter_grant_check(struct obd_export *exp, struct obdo *oa,
 513                               int objcount, struct fsfilt_objinfo *fso,
 514                               int niocount, struct niobuf_local *lnb,
 515                               obd_size *left, struct inode *inode)
 516 {
 517         struct filter_export_data *fed = &exp->exp_filter_data;
 518         int blocksize = exp->exp_obd->u.obt.obt_sb->s_blocksize;
 519         unsigned long used = 0, ungranted = 0, using;
 520         int i, rc = -ENOSPC, obj, n = 0;
 521         int resend = 0;
 522
 523         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 524             (oa->o_flags & OBD_FL_RECOV_RESEND)) {
 525                 resend = 1;
 526                 CDEBUG(D_CACHE, "Recoverable resend arrived, skipping "
 527                        "accounting\n");
 528         }
 529
 530         LASSERT_SPIN_LOCKED(&exp->exp_obd->obd_osfs_lock);
 531
 532         for (obj = 0; obj < objcount; obj++) {
 533                 for (i = 0; i < fso[obj].fso_bufcnt; i++, n++) {
 534                         int tmp, bytes;
 535
 536                         /* should match the code in osc_exit_cache */
 537                         bytes = lnb[n].len;
 538                         bytes += lnb[n].offset & (blocksize - 1);
 539                         tmp = (lnb[n].offset + lnb[n].len) & (blocksize - 1);
 540                         if (tmp)
 541                                 bytes += blocksize - tmp;
 542
 543                         lnb[n].rc = 0;
 544                         if ((lnb[n].flags & OBD_BRW_FROM_GRANT) &&
 545                             (oa->o_valid & OBD_MD_FLGRANT)) {
 546                                 if (resend) {
 547                                         /* this is a recoverable resent */
 548                                         lnb[n].flags |= OBD_BRW_GRANTED;
 549                                         rc = 0;
 550                                         continue;
 551                                 } else if (fed->fed_grant < used + bytes) {
 552                                         CDEBUG(D_CACHE,
 553                                                "%s: cli %s/%p claims %ld+%d "
 554                                                "GRANT, real grant %lu idx %d\n",
 555                                                exp->exp_obd->obd_name,
 556                                                exp->exp_client_uuid.uuid, exp,
 557                                                used, bytes, fed->fed_grant, n);
 558                                 } else {
 559                                         used += bytes;
 560                                         lnb[n].flags |= OBD_BRW_GRANTED;
 561                                         lnb[n].lnb_grant_used = bytes;
 562                                         CDEBUG(0, "idx %d used=%lu\n", n, used);
 563                                         rc = 0;
 564                                         continue;
 565                                 }
 566                         }
 567                         if (*left > ungranted + bytes) {
 568                                 /* if enough space, pretend it was granted */
 569                                 ungranted += bytes;
 570                                 lnb[n].flags |= OBD_BRW_GRANTED;
 571                                 lnb[n].lnb_grant_used = bytes;
 572                                 CDEBUG(0, "idx %d ungranted=%lu\n",n,ungranted);
 573                                 rc = 0;
 574                                 continue;
 575                         }
 576
 577                         /* We can't check for already-mapped blocks here, as
 578                          * it requires dropping the osfs lock to do the bmap.
 579                          * Instead, we return ENOSPC and in that case we need
 580                          * to go through and verify if all of the blocks not
 581                          * marked BRW_GRANTED are already mapped and we can
 582                          * ignore this error. */
 583                         lnb[n].rc = -ENOSPC;
 584                         lnb[n].flags &= ~OBD_BRW_GRANTED;
 585                         CDEBUG(D_CACHE,"%s: cli %s/%p idx %d no space for %d\n",
 586                                exp->exp_obd->obd_name,
 587                                exp->exp_client_uuid.uuid, exp, n, bytes);
 588                 }
 589         }
 590
 591         /* Now substract what client have used already.  We don't subtract
 592          * this from the tot_granted yet, so that other client's can't grab
 593          * that space before we have actually allocated our blocks.  That
 594          * happens in filter_grant_commit() after the writes are done. */
 595         *left -= ungranted;
 596         fed->fed_grant -= used;
 597         fed->fed_pending += used + ungranted;
 598         exp->exp_obd->u.filter.fo_tot_granted += ungranted;
 599         exp->exp_obd->u.filter.fo_tot_pending += used + ungranted;
 600
 601         CDEBUG(D_CACHE,
 602                "%s: cli %s/%p used: %lu ungranted: %lu grant: %lu dirty: %lu\n",
 603                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp, used,
 604                ungranted, fed->fed_grant, fed->fed_dirty);
 605
 606         /* Rough calc in case we don't refresh cached statfs data */
 607         using = (used + ungranted + 1 ) >>
 608                 exp->exp_obd->u.obt.obt_sb->s_blocksize_bits;
 609         if (exp->exp_obd->obd_osfs.os_bavail > using)
 610                 exp->exp_obd->obd_osfs.os_bavail -= using;
 611         else
 612                 exp->exp_obd->obd_osfs.os_bavail = 0;
 613
 614         if (fed->fed_dirty < used) {
 615                 CERROR("%s: cli %s/%p claims used %lu > fed_dirty %lu\n",
 616                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 617                        used, fed->fed_dirty);
 618                 used = fed->fed_dirty;
 619         }
 620         exp->exp_obd->u.filter.fo_tot_dirty -= used;
 621         fed->fed_dirty -= used;
 622
 623         if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
 624                 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
 625                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 626                        fed->fed_dirty, fed->fed_pending, fed->fed_grant);
 627                 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
 628                 LBUG();
 629         }
 630         return rc;
 631 }
 632
 633 /* If we ever start to support multi-object BRW RPCs, we will need to get locks
 634  * on mulitple inodes.  That isn't all, because there still exists the
 635  * possibility of a truncate starting a new transaction while holding the ext3
 636  * rwsem = write while some writes (which have started their transactions here)
 637  * blocking on the ext3 rwsem = read => lock inversion.
 638  *
 639  * The handling gets very ugly when dealing with locked pages.  It may be easier
 640  * to just get rid of the locked page code (which has problems of its own) and
 641  * either discover we do not need it anymore (i.e. it was a symptom of another
 642  * bug) or ensure we get the page locks in an appropriate order. */
 643 static int filter_preprw_write(int cmd, struct obd_export *exp, struct obdo *oa,
 644                                int objcount, struct obd_ioobj *obj,
 645                                struct niobuf_remote *nb, int *npages,
 646                                struct niobuf_local *res,
 647                                struct obd_trans_info *oti,
 648                                struct lustre_capa *capa)
 649 {
 650         struct obd_device *obd = exp->exp_obd;
 651         struct timeval start, end;
 652         struct lvfs_run_ctxt saved;
 653         struct niobuf_local *lnb = res;
 654         struct fsfilt_objinfo fso;
 655         struct filter_mod_data *fmd;
 656         struct dentry *dentry = NULL;
 657         void *iobuf;
 658         obd_size left;
 659         unsigned long now = jiffies, timediff;
 660         int rc = 0, i, tot_bytes = 0, cleanup_phase = 0, localreq = 0;
 661         int retries = 0;
 662         ENTRY;
 663         LASSERT(objcount == 1);
 664         LASSERT(obj->ioo_bufcnt > 0);
 665
 666         rc = filter_auth_capa(exp, NULL, oa->o_seq, capa,
 667                               CAPA_OPC_OSS_WRITE);
 668         if (rc)
 669                 RETURN(rc);
 670
 671         if (exp->exp_connection &&
 672             exp->exp_connection->c_peer.nid == exp->exp_connection->c_self)
 673                 localreq = 1;
 674
 675         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 676         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 677         if (IS_ERR(iobuf))
 678                 GOTO(cleanup, rc = PTR_ERR(iobuf));
 679         cleanup_phase = 1;
 680
 681         dentry = filter_fid2dentry(obd, NULL, obj->ioo_seq,
 682                                    obj->ioo_id);
 683         if (IS_ERR(dentry))
 684                 GOTO(cleanup, rc = PTR_ERR(dentry));
 685         cleanup_phase = 2;
 686
 687         if (dentry->d_inode == NULL) {
 688                 if (exp->exp_obd->obd_recovering) {
 689                         struct obdo *noa = oa;
 690
 691                         if (oa == NULL) {
 692                                 OBDO_ALLOC(noa);
 693                                 if (noa == NULL)
 694                                         GOTO(recreate_out, rc = -ENOMEM);
 695                                 noa->o_id = obj->ioo_id;
 696                                 noa->o_valid = OBD_MD_FLID;
 697                         }
 698
 699                         if (filter_create(NULL, exp, noa, NULL, oti) == 0) {
 700                                 f_dput(dentry);
 701                                 dentry = filter_fid2dentry(exp->exp_obd, NULL,
 702                                                            obj->ioo_seq,
 703                                                            obj->ioo_id);
 704                         }
 705                         if (oa == NULL)
 706                                 OBDO_FREE(noa);
 707                 }
 708     recreate_out:
 709                 if (IS_ERR(dentry) || dentry->d_inode == NULL) {
 710                         CERROR("%s: BRW to missing obj "LPU64"/"LPU64":rc %d\n",
 711                                exp->exp_obd->obd_name,
 712                                obj->ioo_id, obj->ioo_seq,
 713                                IS_ERR(dentry) ? (int)PTR_ERR(dentry) : -ENOENT);
 714                         if (IS_ERR(dentry))
 715                                 cleanup_phase = 1;
 716                         GOTO(cleanup, rc = -ENOENT);
 717                 }
 718         }
 719
 720         if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID) &&
 721             dentry->d_inode->i_mode & (S_ISUID | S_ISGID)) {
 722                 rc = filter_capa_fixoa(exp, oa, oa->o_seq, capa);
 723                 if (rc)
 724                         GOTO(cleanup, rc);
 725         }
 726
 727         rc = filter_map_remote_to_local(objcount, obj, nb, npages, res);
 728         if (rc)
 729                 GOTO(cleanup, rc);
 730
 731         fsfilt_check_slow(obd, now, "preprw_write setup");
 732
 733         /* Filter truncate first locks i_mutex then partially truncated
 734          * page, filter write code first locks pages then take
 735          * i_mutex.  To avoid a deadlock in case of concurrent
 736          * punch/write requests from one client, filter writes and
 737          * filter truncates are serialized by i_alloc_sem, allowing
 738          * multiple writes or single truncate. */
 739         down_read(&dentry->d_inode->i_alloc_sem);
 740         fsfilt_check_slow(obd, now, "i_alloc_sem");
 741
 742         /* Don't update inode timestamps if this write is older than a
 743          * setattr which modifies the timestamps. b=10150 */
 744         /* XXX when we start having persistent reservations this needs to
 745          * be changed to filter_fmd_get() to create the fmd if it doesn't
 746          * already exist so we can store the reservation handle there. */
 747         fmd = filter_fmd_find(exp, obj->ioo_id, obj->ioo_seq);
 748
 749         LASSERT(oa != NULL);
 750 retry:
 751         cfs_spin_lock(&obd->obd_osfs_lock);
 752         if (retries == 0)
 753                 filter_grant_incoming(exp, oa);
 754         if (fmd && fmd->fmd_mactime_xid > oti->oti_xid)
 755                 oa->o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 756                                  OBD_MD_FLATIME);
 757         else
 758                 obdo_to_inode(dentry->d_inode, oa, OBD_MD_FLATIME |
 759                               OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 760         cleanup_phase = 3;
 761
 762         left = filter_grant_space_left(exp);
 763
 764         fso.fso_dentry = dentry;
 765         fso.fso_bufcnt = *npages;
 766
 767         rc = filter_grant_check(exp, oa, objcount, &fso, *npages, res,
 768                                 &left, dentry->d_inode);
 769
 770         /* do not zero out oa->o_valid as it is used in filter_commitrw_write()
 771          * for setting UID/GID and fid EA in first write time. */
 772         /* If OBD_FL_SHRINK_GRANT is set, the client just returned us some grant
 773          * so no sense in allocating it some more. We either return the grant
 774          * back to the client if we have plenty of space or we don't return
 775          * anything if we are short. This was decided in filter_grant_incoming*/
 776         if ((retries == 0) && (oa->o_valid & OBD_MD_FLGRANT) &&
 777             (!(oa->o_valid & OBD_MD_FLFLAGS) ||
 778              !(oa->o_flags & OBD_FL_SHRINK_GRANT)))
 779                 oa->o_grant = filter_grant(exp, oa->o_grant, oa->o_undirty,
 780                                            left, 1);
 781
 782         cfs_spin_unlock(&obd->obd_osfs_lock);
 783
 784         if (rc == -ENOSPC && retries == 0) {
 785                 void *handle = NULL;
 786
 787                 CDEBUG(D_INODE, "retry after commit pending journals");
 788
 789                 retries = 1;
 790                 handle = fsfilt_start(obd, dentry->d_inode,
 791                                       FSFILT_OP_SETATTR, NULL);
 792                 if (handle != NULL &&
 793                     fsfilt_commit(obd, dentry->d_inode, handle, 1) == 0)
 794                         goto retry;
 795         }
 796
 797         filter_fmd_put(exp, fmd);
 798
 799         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK2, (obd_timeout + 1) / 4);
 800
 801         if (rc)
 802                 GOTO(cleanup, rc);
 803         cleanup_phase = 4;
 804
 805         cfs_gettimeofday(&start);
 806         for (i = 0, lnb = res; i < *npages; i++, lnb++) {
 807
 808                 /* We still set up for ungranted pages so that granted pages
 809                  * can be written to disk as they were promised, and portals
 810                  * needs to keep the pages all aligned properly. */
 811                 lnb->dentry = dentry;
 812
 813                 lnb->page = filter_get_page(obd, dentry->d_inode, lnb->offset,
 814                                             localreq);
 815                 if (lnb->page == NULL)
 816                         GOTO(cleanup, rc = -ENOMEM);
 817
 818                 /* DLM locking protects us from write and truncate competing
 819                  * for same region, but truncate can leave dirty page in the
 820                  * cache. it's possible the writeout on a such a page is in
 821                  * progress when we access it. it's also possible that during
 822                  * this writeout we put new (partial) data, but then won't
 823                  * be able to proceed in filter_commitrw_write(). thus let's
 824                  * just wait for writeout completion, should be rare enough.
 825                  * -bzzz */
 826                 wait_on_page_writeback(lnb->page);
 827                 BUG_ON(PageWriteback(lnb->page));
 828
 829                 /* If the filter writes a partial page, then has the file
 830                  * extended, the client will read in the whole page.  the
 831                  * filter has to be careful to zero the rest of the partial
 832                  * page on disk.  we do it by hand for partial extending
 833                  * writes, send_bio() is responsible for zeroing pages when
 834                  * asked to read unmapped blocks -- brw_kiovec() does this. */
 835                 if (lnb->len != CFS_PAGE_SIZE) {
 836                         __s64 maxidx;
 837
 838                         maxidx = ((i_size_read(dentry->d_inode) +
 839                                    CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT) - 1;
 840                         if (maxidx >= lnb->page->index) {
 841                                 LL_CDEBUG_PAGE(D_PAGE, lnb->page, "write %u @ "
 842                                                LPU64" flg %x before EOF %llu\n",
 843                                                lnb->len, lnb->offset,lnb->flags,
 844                                                i_size_read(dentry->d_inode));
 845                                 filter_iobuf_add_page(obd, iobuf,
 846                                                       dentry->d_inode,
 847                                                       lnb->page);
 848                         } else {
 849                                 long off;
 850                                 char *p = kmap(lnb->page);
 851
 852                                 off = lnb->offset & ~CFS_PAGE_MASK;
 853                                 if (off)
 854                                         memset(p, 0, off);
 855                                 off = (lnb->offset + lnb->len) & ~CFS_PAGE_MASK;
 856                                 if (off)
 857                                         memset(p + off, 0, CFS_PAGE_SIZE - off);
 858                                 kunmap(lnb->page);
 859                         }
 860                 }
 861                 if (lnb->rc == 0)
 862                         tot_bytes += lnb->len;
 863         }
 864         cfs_gettimeofday(&end);
 865         timediff = cfs_timeval_sub(&end, &start, NULL);
 866         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
 867
 868         if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
 869                 GOTO(cleanup, rc = -ENOMEM);
 870
 871         /* don't unlock pages to prevent any access */
 872         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf, exp,
 873                               NULL, NULL, NULL);
 874
 875         fsfilt_check_slow(obd, now, "start_page_write");
 876
 877         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
 878                             tot_bytes);
 879
 880         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
 881                 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
 882                                     LPROC_FILTER_WRITE_BYTES, tot_bytes);
 883         filter_counter_incr(exp, LPROC_FILTER_STATS_WRITE,
 884                             oti ? oti->oti_jobid : NULL, tot_bytes);
 885         EXIT;
 886 cleanup:
 887         switch(cleanup_phase) {
 888         case 4:
 889                 if (rc) {
 890                         for (i = 0, lnb = res; i < *npages; i++, lnb++) {
 891                                 if (lnb->page != NULL) {
 892                                         unlock_page(lnb->page);
 893                                         page_cache_release(lnb->page);
 894                                         lnb->page = NULL;
 895                                 }
 896                         }
 897                 }
 898         case 3:
 899                 if (rc)
 900                         up_read(&dentry->d_inode->i_alloc_sem);
 901
 902                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 903         case 2:
 904                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 905                 if (rc)
 906                         f_dput(dentry);
 907                 break;
 908         case 1:
 909                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 910         case 0:
 911                 cfs_spin_lock(&obd->obd_osfs_lock);
 912                 if (oa)
 913                         filter_grant_incoming(exp, oa);
 914                 cfs_spin_unlock(&obd->obd_osfs_lock);
 915                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 916                 break;
 917         default:;
 918         }
 919         return rc;
 920 }
 921
 922 int filter_preprw(const struct lu_env *env, int cmd, struct obd_export *exp,
 923                   struct obdo *oa, int objcount, struct obd_ioobj *obj,
 924                   struct niobuf_remote *nb, int *npages,
 925                   struct niobuf_local *res, struct obd_trans_info *oti,
 926                   struct lustre_capa *capa)
 927 {
 928         if (cmd == OBD_BRW_WRITE)
 929                 return filter_preprw_write(cmd, exp, oa, objcount, obj,
 930                                            nb, npages, res, oti, capa);
 931         if (cmd == OBD_BRW_READ)
 932                 return filter_preprw_read(cmd, exp, oa, objcount, obj,
 933                                           nb, npages, res, oti, capa);
 934         LBUG();
 935         return -EPROTO;
 936 }
 937
 938 static int filter_commitrw_read(struct obd_export *exp, struct obdo *oa,
 939                                 int objcount, struct obd_ioobj *obj,
 940                                 struct niobuf_remote *rnb,
 941                                 int npages, struct niobuf_local *res,
 942                                 struct obd_trans_info *oti, int rc)
 943 {
 944         struct filter_obd *fo = &exp->exp_obd->u.filter;
 945         struct inode *inode = NULL;
 946         struct ldlm_res_id res_id;
 947         struct ldlm_resource *resource = NULL;
 948         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 949         struct niobuf_local *lnb;
 950         int i;
 951         ENTRY;
 952
 953         osc_build_res_name(obj->ioo_id, obj->ioo_seq, &res_id);
 954         /* If oa != NULL then filter_preprw_read updated the inode atime
 955          * and we should update the lvb so that other glimpses will also
 956          * get the updated value. bug 5972 */
 957         if (oa && ns && ns->ns_lvbo && ns->ns_lvbo->lvbo_update) {
 958                 resource = ldlm_resource_get(ns, NULL, &res_id, LDLM_EXTENT, 0);
 959
 960                 if (resource != NULL) {
 961                         LDLM_RESOURCE_ADDREF(resource);
 962                         ns->ns_lvbo->lvbo_update(resource, NULL, 1);
 963                         LDLM_RESOURCE_DELREF(resource);
 964                         ldlm_resource_putref(resource);
 965                 }
 966         }
 967
 968         if (res->dentry != NULL)
 969                 inode = res->dentry->d_inode;
 970
 971         for (i = 0, lnb = res; i < npages; i++, lnb++) {
 972                 if (lnb->page != NULL) {
 973                         page_cache_release(lnb->page);
 974                         lnb->page = NULL;
 975                 }
 976         }
 977         if (inode && (fo->fo_read_cache == 0 ||
 978                       i_size_read(inode) > fo->fo_readcache_max_filesize))
 979                 filter_release_cache(exp->exp_obd, obj, rnb, inode);
 980
 981         if (res->dentry != NULL)
 982                 f_dput(res->dentry);
 983         RETURN(rc);
 984 }
 985
 986 void filter_grant_commit(struct obd_export *exp, int niocount,
 987                          struct niobuf_local *res)
 988 {
 989         struct filter_obd *filter = &exp->exp_obd->u.filter;
 990         struct niobuf_local *lnb = res;
 991         unsigned long pending = 0;
 992         int i;
 993
 994         cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
 995         for (i = 0, lnb = res; i < niocount; i++, lnb++)
 996                 pending += lnb->lnb_grant_used;
 997
 998         LASSERTF(exp->exp_filter_data.fed_pending >= pending,
 999                  "%s: cli %s/%p fed_pending: %lu grant_used: %lu\n",
1000                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
1001                  exp->exp_filter_data.fed_pending, pending);
1002         exp->exp_filter_data.fed_pending -= pending;
1003         LASSERTF(filter->fo_tot_granted >= pending,
1004                  "%s: cli %s/%p tot_granted: "LPU64" grant_used: %lu\n",
1005                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
1006                  exp->exp_obd->u.filter.fo_tot_granted, pending);
1007         filter->fo_tot_granted -= pending;
1008         LASSERTF(filter->fo_tot_pending >= pending,
1009                  "%s: cli %s/%p tot_pending: "LPU64" grant_used: %lu\n",
1010                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
1011                  filter->fo_tot_pending, pending);
1012         filter->fo_tot_pending -= pending;
1013
1014         cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
1015 }
1016
1017 int filter_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
1018                     struct obdo *oa, int objcount, struct obd_ioobj *obj,
1019                     struct niobuf_remote *nb, int npages,
1020                     struct niobuf_local *res, struct obd_trans_info *oti,
1021                     int rc)
1022 {
1023         if (cmd == OBD_BRW_WRITE)
1024                 return filter_commitrw_write(exp, oa, objcount, obj,
1025                                              nb, npages, res, oti, rc);
1026         if (cmd == OBD_BRW_READ)
1027                 return filter_commitrw_read(exp, oa, objcount, obj,
1028                                             nb, npages, res, oti, rc);
1029         LBUG();
1030         return -EPROTO;
1031 }