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  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/obdfilter/filter_io.c
  37  *
  38  * Author: Peter Braam <braam@clusterfs.com>
  39  * Author: Andreas Dilger <adilger@clusterfs.com>
  40  * Author: Phil Schwan <phil@clusterfs.com>
  41  */
  42
  43 #define DEBUG_SUBSYSTEM S_FILTER
  44
  45 #ifndef AUTOCONF_INCLUDED
  46 #include <linux/config.h>
  47 #endif
  48 #include <linux/module.h>
  49 #include <linux/pagemap.h> // XXX kill me soon
  50 #include <linux/version.h>
  51
  52 #include <obd_class.h>
  53 #include <obd_ost.h>
  54 #include <lustre_fsfilt.h>
  55 #include "filter_internal.h"
  56
  57 int *obdfilter_created_scratchpad;
  58
  59 /* Grab the dirty and seen grant announcements from the incoming obdo.
  60  * We will later calculate the clients new grant and return it.
  61  * Caller must hold osfs lock */
  62 void filter_grant_incoming(struct obd_export *exp, struct obdo *oa)
  63 {
  64         struct filter_export_data *fed;
  65         struct obd_device *obd = exp->exp_obd;
  66         ENTRY;
  67
  68         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
  69
  70         if ((oa->o_valid & (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) !=
  71                                         (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) {
  72                 oa->o_valid &= ~OBD_MD_FLGRANT;
  73                 EXIT;
  74                 return;
  75         }
  76
  77         fed = &exp->exp_filter_data;
  78
  79         /* Add some margin, since there is a small race if other RPCs arrive
  80          * out-or-order and have already consumed some grant.  We want to
  81          * leave this here in case there is a large error in accounting. */
  82         CDEBUG(D_CACHE,
  83                "%s: cli %s/%p reports grant: "LPU64" dropped: %u, local: %lu\n",
  84                obd->obd_name, exp->exp_client_uuid.uuid, exp, oa->o_grant,
  85                oa->o_dropped, fed->fed_grant);
  86
  87         /* Update our accounting now so that statfs takes it into account.
  88          * Note that fed_dirty is only approximate and can become incorrect
  89          * if RPCs arrive out-of-order.  No important calculations depend
  90          * on fed_dirty however, but we must check sanity to not assert. */
  91         if ((long long)oa->o_dirty < 0)
  92                 oa->o_dirty = 0;
  93         else if (oa->o_dirty > fed->fed_grant + 4 * FILTER_GRANT_CHUNK)
  94                 oa->o_dirty = fed->fed_grant + 4 * FILTER_GRANT_CHUNK;
  95         obd->u.filter.fo_tot_dirty += oa->o_dirty - fed->fed_dirty;
  96         if (fed->fed_grant < oa->o_dropped) {
  97                 CDEBUG(D_CACHE,"%s: cli %s/%p reports %u dropped > grant %lu\n",
  98                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
  99                        oa->o_dropped, fed->fed_grant);
 100                 oa->o_dropped = 0;
 101         }
 102         if (obd->u.filter.fo_tot_granted < oa->o_dropped) {
 103                 CERROR("%s: cli %s/%p reports %u dropped > tot_grant "LPU64"\n",
 104                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
 105                        oa->o_dropped, obd->u.filter.fo_tot_granted);
 106                 oa->o_dropped = 0;
 107         }
 108         obd->u.filter.fo_tot_granted -= oa->o_dropped;
 109         fed->fed_grant -= oa->o_dropped;
 110         fed->fed_dirty = oa->o_dirty;
 111
 112         if (oa->o_flags & OBD_FL_SHRINK_GRANT) {
 113                 obd_size left_space = filter_grant_space_left(exp);
 114                 struct filter_obd *filter = &exp->exp_obd->u.filter;
 115
 116                 /*Only if left_space < fo_tot_clients * 32M,
 117                  *then the grant space could be shrinked */
 118                 if (left_space < filter->fo_tot_granted_clients *
 119                                  FILTER_GRANT_SHRINK_LIMIT) {
 120                         fed->fed_grant -= oa->o_grant;
 121                         filter->fo_tot_granted -= oa->o_grant;
 122                         CDEBUG(D_CACHE, "%s: cli %s/%p shrink "LPU64
 123                                "fed_grant %ld total "LPU64"\n",
 124                                obd->obd_name, exp->exp_client_uuid.uuid,
 125                                exp, oa->o_grant, fed->fed_grant,
 126                                filter->fo_tot_granted);
 127                         oa->o_grant = 0;
 128                 }
 129         }
 130
 131         if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
 132                 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
 133                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
 134                        fed->fed_dirty, fed->fed_pending, fed->fed_grant);
 135                 spin_unlock(&obd->obd_osfs_lock);
 136                 LBUG();
 137         }
 138         EXIT;
 139 }
 140
 141 /* Figure out how much space is available between what we've granted
 142  * and what remains in the filesystem.  Compensate for ext3 indirect
 143  * block overhead when computing how much free space is left ungranted.
 144  *
 145  * Caller must hold obd_osfs_lock. */
 146 obd_size filter_grant_space_left(struct obd_export *exp)
 147 {
 148         struct obd_device *obd = exp->exp_obd;
 149         int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
 150         obd_size tot_granted = obd->u.filter.fo_tot_granted, avail, left = 0;
 151         int rc, statfs_done = 0;
 152
 153         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
 154
 155         if (cfs_time_before_64(obd->obd_osfs_age, cfs_time_current_64() - HZ)) {
 156 restat:
 157                 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb,
 158                                    cfs_time_current_64() + HZ);
 159                 if (rc) /* N.B. statfs can't really fail */
 160                         RETURN(0);
 161                 statfs_done = 1;
 162         }
 163
 164         avail = obd->obd_osfs.os_bavail;
 165         left = avail - (avail >> (blockbits - 3)); /* (d)indirect */
 166         if (left > GRANT_FOR_LLOG(obd)) {
 167                 left = (left - GRANT_FOR_LLOG(obd)) << blockbits;
 168         } else {
 169                 left = 0 /* << blockbits */;
 170         }
 171
 172         if (!statfs_done && left < 32 * FILTER_GRANT_CHUNK + tot_granted) {
 173                 CDEBUG(D_CACHE, "fs has no space left and statfs too old\n");
 174                 goto restat;
 175         }
 176
 177         if (left >= tot_granted) {
 178                 left -= tot_granted;
 179         } else {
 180                 if (left < tot_granted - obd->u.filter.fo_tot_pending) {
 181                         CERROR("%s: cli %s/%p grant "LPU64" > available "
 182                                LPU64" and pending "LPU64"\n", obd->obd_name,
 183                                exp->exp_client_uuid.uuid, exp, tot_granted,
 184                                left, obd->u.filter.fo_tot_pending);
 185                 }
 186                 left = 0;
 187         }
 188
 189         CDEBUG(D_CACHE, "%s: cli %s/%p free: "LPU64" avail: "LPU64" grant "LPU64
 190                " left: "LPU64" pending: "LPU64"\n", obd->obd_name,
 191                exp->exp_client_uuid.uuid, exp,
 192                obd->obd_osfs.os_bfree << blockbits, avail << blockbits,
 193                tot_granted, left, obd->u.filter.fo_tot_pending);
 194
 195         return left;
 196 }
 197
 198 /* Calculate how much grant space to allocate to this client, based on how
 199  * much space is currently free and how much of that is already granted.
 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)
 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 >> blockbits),
 227                             (fs_space_left >> blockbits) / 8);
 228                 grant <<= blockbits;
 229
 230                 if (grant) {
 231                         /* Allow >FILTER_GRANT_CHUNK size when clients
 232                          * reconnect due to a server reboot.
 233                          */
 234                         if ((grant > FILTER_GRANT_CHUNK) &&
 235                             (!obd->obd_recovering))
 236                                 grant = FILTER_GRANT_CHUNK;
 237
 238                         obd->u.filter.fo_tot_granted += grant;
 239                         fed->fed_grant += grant;
 240                         if (fed->fed_grant < 0) {
 241                                 CERROR("%s: cli %s/%p grant %ld want "LPU64
 242                                        "current"LPU64"\n",
 243                                        obd->obd_name, exp->exp_client_uuid.uuid,
 244                                        exp, fed->fed_grant, want,current_grant);
 245                                 spin_unlock(&obd->obd_osfs_lock);
 246                                 LBUG();
 247                         }
 248                 }
 249         }
 250
 251         CDEBUG(D_CACHE,
 252                "%s: cli %s/%p wants: "LPU64" current grant "LPU64
 253                " granting: "LPU64"\n", obd->obd_name, exp->exp_client_uuid.uuid,
 254                exp, want, current_grant, grant);
 255         CDEBUG(D_CACHE,
 256                "%s: cli %s/%p tot cached:"LPU64" granted:"LPU64
 257                " num_exports: %d\n", obd->obd_name, exp->exp_client_uuid.uuid,
 258                exp, obd->u.filter.fo_tot_dirty,
 259                obd->u.filter.fo_tot_granted, obd->obd_num_exports);
 260
 261         return grant;
 262 }
 263
 264 /*
 265  * the routine is used to request pages from pagecache
 266  *
 267  * use GFP_NOFS not allowing to enter FS as the client can run on this node
 268  * and we might end waiting on a page he sent in the request we're serving.
 269  *
 270  * use __GFP_HIGHMEM so that the pages can use all of the available memory
 271  * on 32-bit machines
 272  */
 273 static struct page *filter_get_page(struct obd_device *obd,
 274                                     struct inode *inode,
 275                                     obd_off offset)
 276 {
 277         struct page *page;
 278
 279         page = find_or_create_page(inode->i_mapping, offset >> CFS_PAGE_SHIFT,
 280                                    GFP_NOFS | __GFP_HIGHMEM);
 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 function is used to free all pages used for request
 339  * just to mimic cacheless OSS which don't occupy much memory
 340  */
 341 void filter_invalidate_cache(struct obd_device *obd, struct obd_ioobj *obj,
 342                              struct niobuf_remote *nb, struct inode *inode)
 343 {
 344         struct niobuf_remote *rnb;
 345         int i;
 346
 347         LASSERT(inode != NULL);
 348
 349         for (i = 0, rnb = nb; i < obj->ioo_bufcnt; i++, rnb++) {
 350                 obd_off start;
 351                 obd_off end;
 352
 353                 start = rnb->offset >> CFS_PAGE_SHIFT;
 354                 end = (rnb->offset + rnb->len) >> CFS_PAGE_SHIFT;
 355                 invalidate_mapping_pages(inode->i_mapping, start, end);
 356                 /* just to avoid warnings */
 357                 start = 0;
 358                 end = 0;
 359         }
 360 }
 361
 362 /*
 363  * the invalidate above doesn't work during read because lnet pins pages.
 364  * The truncate is used here instead to drop pages from cache
 365  */
 366 void filter_truncate_cache(struct obd_device *obd, struct obd_ioobj *obj,
 367                            struct niobuf_remote *nb, int pages,
 368                            struct niobuf_local *res, struct inode *inode)
 369 {
 370         struct niobuf_remote *rnb;
 371         int i;
 372
 373         LASSERT(inode != NULL);
 374 #ifdef HAVE_TRUNCATE_RANGE
 375         for (i = 0, rnb = nb; i < obj->ioo_bufcnt; i++, rnb++) {
 376                 /* remove pages in which range is fit */
 377                 truncate_inode_pages_range(inode->i_mapping,
 378                                            rnb->offset & CFS_PAGE_MASK,
 379                                            (rnb->offset + rnb->len - 1) |
 380                                            ~CFS_PAGE_MASK);
 381         }
 382 #elif (defined HAVE_TRUNCATE_COMPLETE)
 383         for (i = 0, lnb = res; i < pages; i++, lnb++)
 384                 truncate_complete_page(inode->i_mapping, lnb->page);
 385 #else
 386 #error "Nor truncate_inode_pages_range or truncate_complete_page are supported"
 387 #endif
 388 }
 389
 390 static int filter_preprw_read(int cmd, struct obd_export *exp, struct obdo *oa,
 391                               int objcount, struct obd_ioobj *obj,
 392                               struct niobuf_remote *nb,
 393                               int *npages, struct niobuf_local *res,
 394                               struct obd_trans_info *oti,
 395                               struct lustre_capa *capa)
 396 {
 397         struct obd_device *obd = exp->exp_obd;
 398         struct timeval start, end;
 399         struct lvfs_run_ctxt saved;
 400         struct niobuf_local *lnb;
 401         struct dentry *dentry = NULL;
 402         struct inode *inode = NULL;
 403         void *iobuf = NULL;
 404         int rc = 0, i, tot_bytes = 0;
 405         unsigned long now = jiffies;
 406         long timediff;
 407         ENTRY;
 408
 409         /* We are currently not supporting multi-obj BRW_READ RPCS at all.
 410          * When we do this function's dentry cleanup will need to be fixed.
 411          * These values are verified in ost_brw_write() from the wire. */
 412         LASSERTF(objcount == 1, "%d\n", objcount);
 413         LASSERTF(obj->ioo_bufcnt > 0, "%d\n", obj->ioo_bufcnt);
 414
 415         rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa), capa,
 416                               CAPA_OPC_OSS_READ);
 417         if (rc)
 418                 RETURN(rc);
 419
 420         if (oa && oa->o_valid & OBD_MD_FLGRANT) {
 421                 spin_lock(&obd->obd_osfs_lock);
 422                 filter_grant_incoming(exp, oa);
 423
 424                 if (!(oa->o_flags & OBD_FL_SHRINK_GRANT))
 425                         oa->o_grant = 0;
 426                 spin_unlock(&obd->obd_osfs_lock);
 427         }
 428
 429         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 430         if (IS_ERR(iobuf))
 431                 RETURN(PTR_ERR(iobuf));
 432
 433         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 434         dentry = filter_oa2dentry(obd, oa);
 435         if (IS_ERR(dentry)) {
 436                 rc = PTR_ERR(dentry);
 437                 dentry = NULL;
 438                 GOTO(cleanup, rc);
 439         }
 440
 441         inode = dentry->d_inode;
 442
 443         obdo_to_inode(inode, oa, OBD_MD_FLATIME);
 444
 445         rc = filter_map_remote_to_local(objcount, obj, nb, npages, res);
 446         if (rc)
 447                 GOTO(cleanup, rc);
 448
 449         fsfilt_check_slow(obd, now, "preprw_read setup");
 450
 451         /* find pages for all segments, fill array with them */
 452         do_gettimeofday(&start);
 453         for (i = 0, lnb = res; i < *npages; i++, lnb++) {
 454
 455                 lnb->dentry = dentry;
 456
 457                 if (i_size_read(inode) <= lnb->offset)
 458                         /* If there's no more data, abort early.  lnb->rc == 0,
 459                          * so it's easy to detect later. */
 460                         break;
 461
 462                 lnb->page = filter_get_page(obd, inode, lnb->offset);
 463                 if (lnb->page == NULL)
 464                         GOTO(cleanup, rc = -ENOMEM);
 465
 466                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS, 1);
 467
 468                 if (i_size_read(inode) < lnb->offset + lnb->len - 1)
 469                         lnb->rc = i_size_read(inode) - lnb->offset;
 470                 else
 471                         lnb->rc = lnb->len;
 472
 473                 tot_bytes += lnb->rc;
 474
 475                 if (PageUptodate(lnb->page)) {
 476                         lprocfs_counter_add(obd->obd_stats,
 477                                             LPROC_FILTER_CACHE_HIT, 1);
 478                         continue;
 479                 }
 480
 481                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_MISS, 1);
 482                 filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 483         }
 484         do_gettimeofday(&end);
 485         timediff = cfs_timeval_sub(&end, &start, NULL);
 486         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
 487
 488         if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
 489                 GOTO(cleanup, rc = -ENOMEM);
 490
 491         fsfilt_check_slow(obd, now, "start_page_read");
 492
 493         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf,
 494                               exp, NULL, NULL, NULL);
 495         if (rc)
 496                 GOTO(cleanup, rc);
 497
 498         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_READ_BYTES, tot_bytes);
 499
 500         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
 501                 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
 502                                     LPROC_FILTER_READ_BYTES, tot_bytes);
 503
 504         EXIT;
 505
 506  cleanup:
 507         /* unlock pages to allow access from concurrent OST_READ */
 508         for (i = 0, lnb = res; i < *npages; i++, lnb++) {
 509                 if (lnb->page) {
 510                         LASSERT(PageLocked(lnb->page));
 511                         unlock_page(lnb->page);
 512
 513                         if (rc) {
 514                                 page_cache_release(lnb->page);
 515                                 lnb->page = NULL;
 516                         }
 517                 }
 518         }
 519
 520         if (rc != 0) {
 521                 if (dentry != NULL)
 522                         f_dput(dentry);
 523         }
 524
 525         filter_iobuf_put(&obd->u.filter, iobuf, oti);
 526
 527         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 528         if (rc)
 529                 CERROR("io error %d\n", rc);
 530
 531         return rc;
 532 }
 533
 534 /* When clients have dirtied as much space as they've been granted they
 535  * fall through to sync writes.  These sync writes haven't been expressed
 536  * in grants and need to error with ENOSPC when there isn't room in the
 537  * filesystem for them after grants are taken into account.  However,
 538  * writeback of the dirty data that was already granted space can write
 539  * right on through.
 540  *
 541  * Caller must hold obd_osfs_lock. */
 542 static int filter_grant_check(struct obd_export *exp, struct obdo *oa,
 543                               int objcount, struct fsfilt_objinfo *fso,
 544                               int niocount, struct niobuf_local *lnb,
 545                               obd_size *left, struct inode *inode)
 546 {
 547         struct filter_export_data *fed = &exp->exp_filter_data;
 548         int blocksize = exp->exp_obd->u.obt.obt_sb->s_blocksize;
 549         unsigned long used = 0, ungranted = 0, using;
 550         int i, rc = -ENOSPC, obj, n = 0;
 551
 552         LASSERT_SPIN_LOCKED(&exp->exp_obd->obd_osfs_lock);
 553
 554         for (obj = 0; obj < objcount; obj++) {
 555                 for (i = 0; i < fso[obj].fso_bufcnt; i++, n++) {
 556                         int tmp, bytes;
 557
 558                         /* should match the code in osc_exit_cache */
 559                         bytes = lnb[n].len;
 560                         bytes += lnb[n].offset & (blocksize - 1);
 561                         tmp = (lnb[n].offset + lnb[n].len) & (blocksize - 1);
 562                         if (tmp)
 563                                 bytes += blocksize - tmp;
 564
 565                         if ((lnb[n].flags & OBD_BRW_FROM_GRANT) &&
 566                             (oa->o_valid & OBD_MD_FLGRANT)) {
 567                                 if (fed->fed_grant < used + bytes) {
 568                                         CDEBUG(D_CACHE,
 569                                                "%s: cli %s/%p claims %ld+%d "
 570                                                "GRANT, real grant %lu idx %d\n",
 571                                                exp->exp_obd->obd_name,
 572                                                exp->exp_client_uuid.uuid, exp,
 573                                                used, bytes, fed->fed_grant, n);
 574                                 } else {
 575                                         used += bytes;
 576                                         lnb[n].flags |= OBD_BRW_GRANTED;
 577                                         lnb[n].lnb_grant_used = bytes;
 578                                         CDEBUG(0, "idx %d used=%lu\n", n, used);
 579                                         rc = 0;
 580                                         continue;
 581                                 }
 582                         }
 583                         if (*left > ungranted + bytes) {
 584                                 /* if enough space, pretend it was granted */
 585                                 ungranted += bytes;
 586                                 lnb[n].flags |= OBD_BRW_GRANTED;
 587                                 lnb[n].lnb_grant_used = bytes;
 588                                 CDEBUG(0, "idx %d ungranted=%lu\n",n,ungranted);
 589                                 rc = 0;
 590                                 continue;
 591                         }
 592
 593                         /* We can't check for already-mapped blocks here, as
 594                          * it requires dropping the osfs lock to do the bmap.
 595                          * Instead, we return ENOSPC and in that case we need
 596                          * to go through and verify if all of the blocks not
 597                          * marked BRW_GRANTED are already mapped and we can
 598                          * ignore this error. */
 599                         lnb[n].rc = -ENOSPC;
 600                         lnb[n].flags &= ~OBD_BRW_GRANTED;
 601                         CDEBUG(D_CACHE,"%s: cli %s/%p idx %d no space for %d\n",
 602                                exp->exp_obd->obd_name,
 603                                exp->exp_client_uuid.uuid, exp, n, bytes);
 604                 }
 605         }
 606
 607         /* Now substract what client have used already.  We don't subtract
 608          * this from the tot_granted yet, so that other client's can't grab
 609          * that space before we have actually allocated our blocks.  That
 610          * happens in filter_grant_commit() after the writes are done. */
 611         *left -= ungranted;
 612         fed->fed_grant -= used;
 613         fed->fed_pending += used + ungranted;
 614         exp->exp_obd->u.filter.fo_tot_granted += ungranted;
 615         exp->exp_obd->u.filter.fo_tot_pending += used + ungranted;
 616
 617         CDEBUG(D_CACHE,
 618                "%s: cli %s/%p used: %lu ungranted: %lu grant: %lu dirty: %lu\n",
 619                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp, used,
 620                ungranted, fed->fed_grant, fed->fed_dirty);
 621
 622         /* Rough calc in case we don't refresh cached statfs data */
 623         using = (used + ungranted + 1 ) >>
 624                 exp->exp_obd->u.obt.obt_sb->s_blocksize_bits;
 625         if (exp->exp_obd->obd_osfs.os_bavail > using)
 626                 exp->exp_obd->obd_osfs.os_bavail -= using;
 627         else
 628                 exp->exp_obd->obd_osfs.os_bavail = 0;
 629
 630         if (fed->fed_dirty < used) {
 631                 CERROR("%s: cli %s/%p claims used %lu > fed_dirty %lu\n",
 632                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 633                        used, fed->fed_dirty);
 634                 used = fed->fed_dirty;
 635         }
 636         exp->exp_obd->u.filter.fo_tot_dirty -= used;
 637         fed->fed_dirty -= used;
 638
 639         if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
 640                 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
 641                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 642                        fed->fed_dirty, fed->fed_pending, fed->fed_grant);
 643                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
 644                 LBUG();
 645         }
 646         return rc;
 647 }
 648
 649 /* If we ever start to support multi-object BRW RPCs, we will need to get locks
 650  * on mulitple inodes.  That isn't all, because there still exists the
 651  * possibility of a truncate starting a new transaction while holding the ext3
 652  * rwsem = write while some writes (which have started their transactions here)
 653  * blocking on the ext3 rwsem = read => lock inversion.
 654  *
 655  * The handling gets very ugly when dealing with locked pages.  It may be easier
 656  * to just get rid of the locked page code (which has problems of its own) and
 657  * either discover we do not need it anymore (i.e. it was a symptom of another
 658  * bug) or ensure we get the page locks in an appropriate order. */
 659 static int filter_preprw_write(int cmd, struct obd_export *exp, struct obdo *oa,
 660                                int objcount, struct obd_ioobj *obj,
 661                                struct niobuf_remote *nb, int *npages,
 662                                struct niobuf_local *res,
 663                                struct obd_trans_info *oti,
 664                                struct lustre_capa *capa)
 665 {
 666         struct obd_device *obd = exp->exp_obd;
 667         struct timeval start, end;
 668         struct lvfs_run_ctxt saved;
 669         struct niobuf_local *lnb = res;
 670         struct fsfilt_objinfo fso;
 671         struct filter_mod_data *fmd;
 672         struct dentry *dentry = NULL;
 673         void *iobuf;
 674         obd_size left;
 675         unsigned long now = jiffies, timediff;
 676         int rc = 0, i, tot_bytes = 0, cleanup_phase = 0;
 677         ENTRY;
 678         LASSERT(objcount == 1);
 679         LASSERT(obj->ioo_bufcnt > 0);
 680
 681         rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa), capa,
 682                               CAPA_OPC_OSS_WRITE);
 683         if (rc)
 684                 RETURN(rc);
 685
 686         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 687         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 688         if (IS_ERR(iobuf))
 689                 GOTO(cleanup, rc = PTR_ERR(iobuf));
 690         cleanup_phase = 1;
 691
 692         dentry = filter_fid2dentry(obd, NULL, obj->ioo_gr,
 693                                    obj->ioo_id);
 694         if (IS_ERR(dentry))
 695                 GOTO(cleanup, rc = PTR_ERR(dentry));
 696         cleanup_phase = 2;
 697
 698         if (dentry->d_inode == NULL) {
 699                 CERROR("%s: trying to BRW to non-existent file "LPU64"\n",
 700                        obd->obd_name, obj->ioo_id);
 701                 GOTO(cleanup, rc = -ENOENT);
 702         }
 703
 704         if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID) &&
 705             dentry->d_inode->i_mode & (S_ISUID | S_ISGID)) {
 706                 rc = filter_capa_fixoa(exp, oa, obdo_mdsno(oa), capa);
 707                 if (rc)
 708                         GOTO(cleanup, rc);
 709         }
 710
 711         rc = filter_map_remote_to_local(objcount, obj, nb, npages, res);
 712         if (rc)
 713                 GOTO(cleanup, rc);
 714
 715         fsfilt_check_slow(obd, now, "preprw_write setup");
 716
 717         /* Don't update inode timestamps if this write is older than a
 718          * setattr which modifies the timestamps. b=10150 */
 719         /* XXX when we start having persistent reservations this needs to
 720          * be changed to filter_fmd_get() to create the fmd if it doesn't
 721          * already exist so we can store the reservation handle there. */
 722         fmd = filter_fmd_find(exp, obj->ioo_id, obj->ioo_gr);
 723
 724         LASSERT(oa != NULL);
 725         spin_lock(&obd->obd_osfs_lock);
 726         filter_grant_incoming(exp, oa);
 727         if (fmd && fmd->fmd_mactime_xid > oti->oti_xid)
 728                 oa->o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 729                                  OBD_MD_FLATIME);
 730         else
 731                 obdo_to_inode(dentry->d_inode, oa, OBD_MD_FLATIME |
 732                               OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 733         cleanup_phase = 3;
 734
 735         left = filter_grant_space_left(exp);
 736
 737         fso.fso_dentry = dentry;
 738         fso.fso_bufcnt = *npages;
 739
 740         rc = filter_grant_check(exp, oa, objcount, &fso, *npages, res,
 741                                 &left, dentry->d_inode);
 742
 743         /* do not zero out oa->o_valid as it is used in filter_commitrw_write()
 744          * for setting UID/GID and fid EA in first write time. */
 745         if (oa->o_valid & OBD_MD_FLGRANT)
 746                 oa->o_grant = filter_grant(exp,oa->o_grant,oa->o_undirty,left);
 747
 748         spin_unlock(&obd->obd_osfs_lock);
 749         filter_fmd_put(exp, fmd);
 750
 751         if (rc)
 752                 GOTO(cleanup, rc);
 753         cleanup_phase = 4;
 754
 755         /* Filter truncate first locks i_mutex then partally truncated
 756          * page, filter write code first locks pages then take
 757          * i_mutex.  To avoid a deadlock in case of concurrent
 758          * punch/write requests from one client, filter writes and
 759          * filter truncates are serialized by i_alloc_sem, allowing
 760          * multiple writes or single truncate. */
 761         down_read(&dentry->d_inode->i_alloc_sem);
 762
 763         do_gettimeofday(&start);
 764         for (i = 0, lnb = res; i < *npages; i++, lnb++) {
 765
 766                 /* We still set up for ungranted pages so that granted pages
 767                  * can be written to disk as they were promised, and portals
 768                  * needs to keep the pages all aligned properly. */
 769                 lnb->dentry = dentry;
 770
 771                 lnb->page = filter_get_page(obd, dentry->d_inode, lnb->offset);
 772                 if (lnb->page == NULL)
 773                         GOTO(cleanup, rc = -ENOMEM);
 774
 775                 /* DLM locking protects us from write and truncate competing
 776                  * for same region, but truncate can leave dirty page in the
 777                  * cache. it's possible the writeout on a such a page is in
 778                  * progress when we access it. it's also possible that during
 779                  * this writeout we put new (partial) data, but then won't
 780                  * be able to proceed in filter_commitrw_write(). thus let's
 781                  * just wait for writeout completion, should be rare enough.
 782                  * -bzzz */
 783                 if (obd->u.filter.fo_writethrough_cache)
 784                         wait_on_page_writeback(lnb->page);
 785                 BUG_ON(PageWriteback(lnb->page));
 786
 787                 /* If the filter writes a partial page, then has the file
 788                  * extended, the client will read in the whole page.  the
 789                  * filter has to be careful to zero the rest of the partial
 790                  * page on disk.  we do it by hand for partial extending
 791                  * writes, send_bio() is responsible for zeroing pages when
 792                  * asked to read unmapped blocks -- brw_kiovec() does this. */
 793                 if (lnb->len != CFS_PAGE_SIZE) {
 794                         __s64 maxidx;
 795
 796                         maxidx = ((i_size_read(dentry->d_inode) +
 797                                    CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT) - 1;
 798                         if (maxidx >= lnb->page->index) {
 799                                 LL_CDEBUG_PAGE(D_PAGE, lnb->page, "write %u @ "
 800                                                LPU64" flg %x before EOF %llu\n",
 801                                                lnb->len, lnb->offset,lnb->flags,
 802                                                i_size_read(dentry->d_inode));
 803                                 filter_iobuf_add_page(obd, iobuf,
 804                                                       dentry->d_inode,
 805                                                       lnb->page);
 806                         } else {
 807                                 long off;
 808                                 char *p = kmap(lnb->page);
 809
 810                                 off = lnb->offset & ~CFS_PAGE_MASK;
 811                                 if (off)
 812                                         memset(p, 0, off);
 813                                 off = (lnb->offset + lnb->len) & ~CFS_PAGE_MASK;
 814                                 if (off)
 815                                         memset(p + off, 0, CFS_PAGE_SIZE - off);
 816                                 kunmap(lnb->page);
 817                         }
 818                 }
 819                 if (lnb->rc == 0)
 820                         tot_bytes += lnb->len;
 821         }
 822         do_gettimeofday(&end);
 823         timediff = cfs_timeval_sub(&end, &start, NULL);
 824         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
 825
 826         if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
 827                 GOTO(cleanup, rc = -ENOMEM);
 828
 829         /* don't unlock pages to prevent any access */
 830         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf, exp,
 831                               NULL, NULL, NULL);
 832
 833         fsfilt_check_slow(obd, now, "start_page_write");
 834
 835         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
 836                 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
 837                                     LPROC_FILTER_WRITE_BYTES, tot_bytes);
 838         EXIT;
 839 cleanup:
 840         switch(cleanup_phase) {
 841         case 4:
 842                 if (rc) {
 843                         for (i = 0, lnb = res; i < *npages; i++, lnb++) {
 844                                 if (lnb->page != NULL) {
 845                                         unlock_page(lnb->page);
 846                                         page_cache_release(lnb->page);
 847                                         lnb->page = NULL;
 848                                 }
 849                         }
 850                         up_read(&dentry->d_inode->i_alloc_sem);
 851                 }
 852         case 3:
 853                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 854         case 2:
 855                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 856                 if (rc)
 857                         f_dput(dentry);
 858                 break;
 859         case 1:
 860                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 861         case 0:
 862                 spin_lock(&obd->obd_osfs_lock);
 863                 if (oa)
 864                         filter_grant_incoming(exp, oa);
 865                 spin_unlock(&obd->obd_osfs_lock);
 866                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 867                 break;
 868         default:;
 869         }
 870         return rc;
 871 }
 872
 873 int filter_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
 874                   int objcount, struct obd_ioobj *obj,
 875                   struct niobuf_remote *nb, int *npages,
 876                   struct niobuf_local *res, struct obd_trans_info *oti,
 877                   struct lustre_capa *capa)
 878 {
 879         if (cmd == OBD_BRW_WRITE)
 880                 return filter_preprw_write(cmd, exp, oa, objcount, obj,
 881                                            nb, npages, res, oti, capa);
 882         if (cmd == OBD_BRW_READ)
 883                 return filter_preprw_read(cmd, exp, oa, objcount, obj,
 884                                           nb, npages, res, oti, capa);
 885         LBUG();
 886         return -EPROTO;
 887 }
 888
 889 static int filter_commitrw_read(struct obd_export *exp, struct obdo *oa,
 890                                 int objcount, struct obd_ioobj *obj,
 891                                 struct niobuf_remote *rnb,
 892                                 int npages, struct niobuf_local *res,
 893                                 struct obd_trans_info *oti, int rc)
 894 {
 895         struct filter_obd *fo = &exp->exp_obd->u.filter;
 896         struct inode *inode = NULL;
 897         struct ldlm_res_id res_id;
 898         struct ldlm_resource *resource = NULL;
 899         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 900         struct niobuf_local *lnb;
 901         int i;
 902         ENTRY;
 903
 904         osc_build_res_name(obj->ioo_id, obj->ioo_gr, &res_id);
 905         /* If oa != NULL then filter_preprw_read updated the inode atime
 906          * and we should update the lvb so that other glimpses will also
 907          * get the updated value. bug 5972 */
 908         if (oa && ns && ns->ns_lvbo && ns->ns_lvbo->lvbo_update) {
 909                 resource = ldlm_resource_get(ns, NULL, &res_id, LDLM_EXTENT, 0);
 910
 911                 if (resource != NULL) {
 912                         LDLM_RESOURCE_ADDREF(resource);
 913                         ns->ns_lvbo->lvbo_update(resource, NULL, 0, 1);
 914                         LDLM_RESOURCE_DELREF(resource);
 915                         ldlm_resource_putref(resource);
 916                 }
 917         }
 918
 919         if (res->dentry != NULL)
 920                 inode = res->dentry->d_inode;
 921
 922         for (i = 0, lnb = res; i < npages; i++, lnb++)
 923                 if (lnb->page != NULL)
 924                         page_cache_release(lnb->page);
 925
 926         if (inode && (fo->fo_read_cache == 0 ||
 927                       i_size_read(inode) > fo->fo_readcache_max_filesize))
 928                 filter_truncate_cache(exp->exp_obd, obj, rnb, npages, res,
 929                                       inode);
 930
 931         for (i = 0, lnb = res; i < npages; i++, lnb++)
 932                 lnb->page = NULL;
 933
 934         if (res->dentry != NULL)
 935                 f_dput(res->dentry);
 936         RETURN(rc);
 937 }
 938
 939 void filter_grant_commit(struct obd_export *exp, int niocount,
 940                          struct niobuf_local *res)
 941 {
 942         struct filter_obd *filter = &exp->exp_obd->u.filter;
 943         struct niobuf_local *lnb = res;
 944         unsigned long pending = 0;
 945         int i;
 946
 947         spin_lock(&exp->exp_obd->obd_osfs_lock);
 948         for (i = 0, lnb = res; i < niocount; i++, lnb++)
 949                 pending += lnb->lnb_grant_used;
 950
 951         LASSERTF(exp->exp_filter_data.fed_pending >= pending,
 952                  "%s: cli %s/%p fed_pending: %lu grant_used: %lu\n",
 953                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 954                  exp->exp_filter_data.fed_pending, pending);
 955         exp->exp_filter_data.fed_pending -= pending;
 956         LASSERTF(filter->fo_tot_granted >= pending,
 957                  "%s: cli %s/%p tot_granted: "LPU64" grant_used: %lu\n",
 958                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 959                  exp->exp_obd->u.filter.fo_tot_granted, pending);
 960         filter->fo_tot_granted -= pending;
 961         LASSERTF(filter->fo_tot_pending >= pending,
 962                  "%s: cli %s/%p tot_pending: "LPU64" grant_used: %lu\n",
 963                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 964                  filter->fo_tot_pending, pending);
 965         filter->fo_tot_pending -= pending;
 966
 967         spin_unlock(&exp->exp_obd->obd_osfs_lock);
 968 }
 969
 970 int filter_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
 971                     int objcount, struct obd_ioobj *obj,
 972                     struct niobuf_remote *nb, int npages,
 973                     struct niobuf_local *res, struct obd_trans_info *oti,
 974                     int rc)
 975 {
 976         if (cmd == OBD_BRW_WRITE)
 977                 return filter_commitrw_write(exp, oa, objcount, obj,
 978                                              nb, npages, res, oti, rc);
 979         if (cmd == OBD_BRW_READ)
 980                 return filter_commitrw_read(exp, oa, objcount, obj,
 981                                             nb, npages, res, oti, rc);
 982         LBUG();
 983         return -EPROTO;
 984 }
 985
 986 int filter_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
 987                obd_count oa_bufs, struct brw_page *pga,
 988                struct obd_trans_info *oti)
 989 {
 990         struct obd_ioobj ioo;
 991         struct niobuf_local *lnb;
 992         struct niobuf_remote *rnb;
 993         obd_count i;
 994         int ret = 0, npages;
 995         ENTRY;
 996
 997         OBD_ALLOC(lnb, oa_bufs * sizeof(struct niobuf_local));
 998         OBD_ALLOC(rnb, oa_bufs * sizeof(struct niobuf_remote));
 999
1000         if (lnb == NULL || rnb == NULL)
1001                 GOTO(out, ret = -ENOMEM);
1002
1003         for (i = 0; i < oa_bufs; i++) {
1004                 lnb[i].page = pga[i].pg;
1005                 rnb[i].offset = pga[i].off;
1006                 rnb[i].len = pga[i].count;
1007         }
1008
1009         obdo_to_ioobj(oinfo->oi_oa, &ioo);
1010         ioo.ioo_bufcnt = oa_bufs;
1011
1012         npages = oa_bufs;
1013         ret = filter_preprw(cmd, exp, oinfo->oi_oa, 1, &ioo,
1014                             rnb, &npages, lnb, oti, oinfo_capa(oinfo));
1015         if (ret != 0)
1016                 GOTO(out, ret);
1017         LASSERTF(oa_bufs == npages, "%u != %u\n", oa_bufs, npages);
1018
1019         ret = filter_commitrw(cmd, exp, oinfo->oi_oa, 1, &ioo, rnb,
1020                               npages, lnb, oti, ret);
1021
1022 out:
1023         if (lnb)
1024                 OBD_FREE(lnb, oa_bufs * sizeof(struct niobuf_local));
1025         if (rnb)
1026                 OBD_FREE(rnb, oa_bufs * sizeof(struct niobuf_remote));
1027         RETURN(ret);
1028 }