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 <lustre_fsfilt.h>
  54 #include "filter_internal.h"
  55
  56 int *obdfilter_created_scratchpad;
  57
  58 /* Grab the dirty and seen grant announcements from the incoming obdo.
  59  * We will later calculate the clients new grant and return it.
  60  * Caller must hold osfs lock */
  61 static void filter_grant_incoming(struct obd_export *exp, struct obdo *oa)
  62 {
  63         struct filter_export_data *fed;
  64         struct obd_device *obd = exp->exp_obd;
  65         ENTRY;
  66
  67         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
  68
  69         if ((oa->o_valid & (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) !=
  70                                         (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) {
  71                 oa->o_valid &= ~OBD_MD_FLGRANT;
  72                 EXIT;
  73                 return;
  74         }
  75
  76         fed = &exp->exp_filter_data;
  77
  78         /* Add some margin, since there is a small race if other RPCs arrive
  79          * out-or-order and have already consumed some grant.  We want to
  80          * leave this here in case there is a large error in accounting. */
  81         CDEBUG(D_CACHE,
  82                "%s: cli %s/%p reports grant: "LPU64" dropped: %u, local: %lu\n",
  83                obd->obd_name, exp->exp_client_uuid.uuid, exp, oa->o_grant,
  84                oa->o_dropped, fed->fed_grant);
  85
  86         /* Update our accounting now so that statfs takes it into account.
  87          * Note that fed_dirty is only approximate and can become incorrect
  88          * if RPCs arrive out-of-order.  No important calculations depend
  89          * on fed_dirty however, but we must check sanity to not assert. */
  90         if ((long long)oa->o_dirty < 0)
  91                 oa->o_dirty = 0;
  92         else if (oa->o_dirty > fed->fed_grant + 4 * FILTER_GRANT_CHUNK)
  93                 oa->o_dirty = fed->fed_grant + 4 * FILTER_GRANT_CHUNK;
  94         obd->u.filter.fo_tot_dirty += oa->o_dirty - fed->fed_dirty;
  95         if (fed->fed_grant < oa->o_dropped) {
  96                 CDEBUG(D_CACHE,"%s: cli %s/%p reports %u dropped > grant %lu\n",
  97                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
  98                        oa->o_dropped, fed->fed_grant);
  99                 oa->o_dropped = 0;
 100         }
 101         if (obd->u.filter.fo_tot_granted < oa->o_dropped) {
 102                 CERROR("%s: cli %s/%p reports %u dropped > tot_grant "LPU64"\n",
 103                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
 104                        oa->o_dropped, obd->u.filter.fo_tot_granted);
 105                 oa->o_dropped = 0;
 106         }
 107         obd->u.filter.fo_tot_granted -= oa->o_dropped;
 108         fed->fed_grant -= oa->o_dropped;
 109         fed->fed_dirty = oa->o_dirty;
 110         if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
 111                 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
 112                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
 113                        fed->fed_dirty, fed->fed_pending, fed->fed_grant);
 114                 spin_unlock(&obd->obd_osfs_lock);
 115                 LBUG();
 116         }
 117         EXIT;
 118 }
 119
 120 /* Figure out how much space is available between what we've granted
 121  * and what remains in the filesystem.  Compensate for ext3 indirect
 122  * block overhead when computing how much free space is left ungranted.
 123  *
 124  * Caller must hold obd_osfs_lock. */
 125 obd_size filter_grant_space_left(struct obd_export *exp)
 126 {
 127         struct obd_device *obd = exp->exp_obd;
 128         int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
 129         obd_size tot_granted = obd->u.filter.fo_tot_granted, avail, left = 0;
 130         int rc, statfs_done = 0;
 131
 132         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
 133
 134         if (cfs_time_before_64(obd->obd_osfs_age, cfs_time_current_64() - HZ)) {
 135 restat:
 136                 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb,
 137                                    cfs_time_current_64() + HZ);
 138                 if (rc) /* N.B. statfs can't really fail */
 139                         RETURN(0);
 140                 statfs_done = 1;
 141         }
 142
 143         avail = obd->obd_osfs.os_bavail;
 144         left = avail - (avail >> (blockbits - 3)); /* (d)indirect */
 145         if (left > GRANT_FOR_LLOG(obd)) {
 146                 left = (left - GRANT_FOR_LLOG(obd)) << blockbits;
 147         } else {
 148                 left = 0 /* << blockbits */;
 149         }
 150
 151         if (!statfs_done && left < 32 * FILTER_GRANT_CHUNK + tot_granted) {
 152                 CDEBUG(D_CACHE, "fs has no space left and statfs too old\n");
 153                 goto restat;
 154         }
 155
 156         if (left >= tot_granted) {
 157                 left -= tot_granted;
 158         } else {
 159                 if (left < tot_granted - obd->u.filter.fo_tot_pending) {
 160                         CERROR("%s: cli %s/%p grant "LPU64" > available "
 161                                LPU64" and pending "LPU64"\n", obd->obd_name,
 162                                exp->exp_client_uuid.uuid, exp, tot_granted,
 163                                left, obd->u.filter.fo_tot_pending);
 164                 }
 165                 left = 0;
 166         }
 167
 168         CDEBUG(D_CACHE, "%s: cli %s/%p free: "LPU64" avail: "LPU64" grant "LPU64
 169                " left: "LPU64" pending: "LPU64"\n", obd->obd_name,
 170                exp->exp_client_uuid.uuid, exp,
 171                obd->obd_osfs.os_bfree << blockbits, avail << blockbits,
 172                tot_granted, left, obd->u.filter.fo_tot_pending);
 173
 174         return left;
 175 }
 176
 177 /* Calculate how much grant space to allocate to this client, based on how
 178  * much space is currently free and how much of that is already granted.
 179  *
 180  * Caller must hold obd_osfs_lock. */
 181 long filter_grant(struct obd_export *exp, obd_size current_grant,
 182                   obd_size want, obd_size fs_space_left)
 183 {
 184         struct obd_device *obd = exp->exp_obd;
 185         struct filter_export_data *fed = &exp->exp_filter_data;
 186         int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
 187         __u64 grant = 0;
 188
 189         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
 190
 191         /* Grant some fraction of the client's requested grant space so that
 192          * they are not always waiting for write credits (not all of it to
 193          * avoid overgranting in face of multiple RPCs in flight).  This
 194          * essentially will be able to control the OSC_MAX_RIF for a client.
 195          *
 196          * If we do have a large disparity between what the client thinks it
 197          * has and what we think it has, don't grant very much and let the
 198          * client consume its grant first.  Either it just has lots of RPCs
 199          * in flight, or it was evicted and its grants will soon be used up. */
 200         if (want > 0x7fffffff) {
 201                 CERROR("%s: client %s/%p requesting > 2GB grant "LPU64"\n",
 202                        obd->obd_name, exp->exp_client_uuid.uuid, exp, want);
 203         } else if (current_grant < want &&
 204                    current_grant < fed->fed_grant + FILTER_GRANT_CHUNK) {
 205                 grant = min((want >> blockbits),
 206                             (fs_space_left >> blockbits) / 8);
 207                 grant <<= blockbits;
 208
 209                 if (grant) {
 210                         /* Allow >FILTER_GRANT_CHUNK size when clients
 211                          * reconnect due to a server reboot.
 212                          */
 213                         if ((grant > FILTER_GRANT_CHUNK) &&
 214                             (!obd->obd_recovering))
 215                                 grant = FILTER_GRANT_CHUNK;
 216
 217                         obd->u.filter.fo_tot_granted += grant;
 218                         fed->fed_grant += grant;
 219                         if (fed->fed_grant < 0) {
 220                                 CERROR("%s: cli %s/%p grant %ld want "LPU64
 221                                        "current"LPU64"\n",
 222                                        obd->obd_name, exp->exp_client_uuid.uuid,
 223                                        exp, fed->fed_grant, want,current_grant);
 224                                 spin_unlock(&obd->obd_osfs_lock);
 225                                 LBUG();
 226                         }
 227                 }
 228         }
 229
 230         CDEBUG(D_CACHE,
 231                "%s: cli %s/%p wants: "LPU64" current grant "LPU64
 232                " granting: "LPU64"\n", obd->obd_name, exp->exp_client_uuid.uuid,
 233                exp, want, current_grant, grant);
 234         CDEBUG(D_CACHE,
 235                "%s: cli %s/%p tot cached:"LPU64" granted:"LPU64
 236                " num_exports: %d\n", obd->obd_name, exp->exp_client_uuid.uuid,
 237                exp, obd->u.filter.fo_tot_dirty,
 238                obd->u.filter.fo_tot_granted, obd->obd_num_exports);
 239
 240         return grant;
 241 }
 242
 243 /*
 244  * the routine is used to request pages from pagecache
 245  *
 246  * use GFP_NOFS not allowing to enter FS as the client can run on this node
 247  * and we might end waiting on a page he sent in the request we're serving.
 248  *
 249  * use __GFP_HIGHMEM so that the pages can use all of the available memory
 250  * on 32-bit machines
 251  *
 252  */
 253 static struct page * filter_get_page(struct obd_device *obd,
 254                                      struct inode *inode,
 255                                      obd_off offset)
 256 {
 257         struct page *page;
 258
 259         page = find_or_create_page(inode->i_mapping, offset >> CFS_PAGE_SHIFT,
 260                                    GFP_NOFS | __GFP_HIGHMEM);
 261         if (unlikely(page == NULL))
 262                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_NO_PAGE, 1);
 263
 264         return page;
 265 }
 266
 267 /*
 268  * the routine initializes array of local_niobuf from remote_niobuf
 269  */
 270 static int filter_map_remote_to_local(int objcount, struct obd_ioobj *obj,
 271                                       struct niobuf_remote *nb,
 272                                       int *nrpages, struct niobuf_local *res)
 273 {
 274         struct niobuf_remote *rnb;
 275         struct niobuf_local *lnb;
 276         int i, max;
 277         ENTRY;
 278
 279         /* we don't support multiobject RPC yet
 280          * ost_brw_read() and ost_brw_write() check this */
 281         LASSERT(objcount == 1);
 282
 283         max = *nrpages;
 284         *nrpages = 0;
 285         for (i = 0, rnb = nb, lnb = res; i < obj->ioo_bufcnt; i++, rnb++) {
 286                 obd_off offset = rnb->offset;
 287                 unsigned int len = rnb->len;
 288
 289                 while (len > 0) {
 290                         int poff = offset & (CFS_PAGE_SIZE - 1);
 291                         int plen = CFS_PAGE_SIZE - poff;
 292
 293                         if (*nrpages >= max) {
 294                                 CERROR("small array of local bufs: %d\n", max);
 295                                 RETURN(-EINVAL);
 296                         }
 297
 298                         if (plen > len)
 299                                 plen = len;
 300                         lnb->offset = offset;
 301                         lnb->len = plen;
 302                         lnb->flags = rnb->flags;
 303                         lnb->page = NULL;
 304                         lnb->rc = 0;
 305                         lnb->lnb_grant_used = 0;
 306
 307                         LASSERTF(plen <= len, "plen %u, len %u\n", plen, len);
 308                         offset += plen;
 309                         len -= plen;
 310                         lnb++;
 311                         (*nrpages)++;
 312                 }
 313         }
 314         RETURN(0);
 315 }
 316
 317 /*
 318  * the function is used to free all pages used for request
 319  * just to mimic cacheless OSS which don't occupy much memory
 320  */
 321 void filter_invalidate_cache(struct obd_device *obd, struct obd_ioobj *obj,
 322                              struct niobuf_remote *nb, struct inode *inode)
 323 {
 324         struct niobuf_remote *rnb;
 325         int i;
 326
 327         LASSERT(inode != NULL);
 328
 329         for (i = 0, rnb = nb; i < obj->ioo_bufcnt; i++, rnb++) {
 330                 invalidate_mapping_pages(inode->i_mapping,
 331                                          rnb->offset >> CFS_PAGE_SHIFT,
 332                                          (rnb->offset + rnb->len) >>
 333                                          CFS_PAGE_SHIFT);
 334         }
 335
 336 }
 337
 338 static int filter_preprw_read(int cmd, struct obd_export *exp, struct obdo *oa,
 339                               int objcount, struct obd_ioobj *obj,
 340                               struct niobuf_remote *nb,
 341                               int *pages, struct niobuf_local *res,
 342                               struct obd_trans_info *oti)
 343 {
 344         struct obd_device *obd = exp->exp_obd;
 345         struct timeval start, end;
 346         struct lvfs_run_ctxt saved;
 347         struct niobuf_local *lnb;
 348         struct dentry *dentry = NULL;
 349         struct inode *inode = NULL;
 350         void *iobuf = NULL;
 351         int rc = 0, i, tot_bytes = 0;
 352         unsigned long now = jiffies;
 353         long timediff;
 354         ENTRY;
 355
 356         /* We are currently not supporting multi-obj BRW_READ RPCS at all.
 357          * When we do this function's dentry cleanup will need to be fixed.
 358          * These values are verified in ost_brw_write() from the wire. */
 359         LASSERTF(objcount == 1, "%d\n", objcount);
 360         LASSERTF(obj->ioo_bufcnt > 0, "%d\n", obj->ioo_bufcnt);
 361
 362         if (oa->o_valid & OBD_MD_FLGRANT) {
 363                 spin_lock(&obd->obd_osfs_lock);
 364                 filter_grant_incoming(exp, oa);
 365
 366                 oa->o_grant = 0;
 367                 spin_unlock(&obd->obd_osfs_lock);
 368         }
 369
 370         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 371         if (IS_ERR(iobuf))
 372                 RETURN(PTR_ERR(iobuf));
 373
 374         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 375         dentry = filter_oa2dentry(obd, oa);
 376         if (IS_ERR(dentry)) {
 377                 rc = PTR_ERR(dentry);
 378                 dentry = NULL;
 379                 GOTO(cleanup, rc);
 380         }
 381
 382         inode = dentry->d_inode;
 383
 384         obdo_to_inode(inode, oa, OBD_MD_FLATIME);
 385
 386         rc = filter_map_remote_to_local(objcount, obj, nb, pages, res);
 387         if (rc)
 388                 GOTO(cleanup, rc);
 389
 390         fsfilt_check_slow(obd, now, "preprw_read setup");
 391
 392         /* find pages for all segments, fill array with them */
 393         do_gettimeofday(&start);
 394         for (i = 0, lnb = res; i < *pages; i++, lnb++) {
 395
 396                 lnb->dentry = dentry;
 397
 398                 if (i_size_read(inode) <= lnb->offset)
 399                         /* If there's no more data, abort early.  lnb->rc == 0,
 400                          * so it's easy to detect later. */
 401                         break;
 402
 403                 lnb->page = filter_get_page(obd, inode, lnb->offset);
 404                 if (lnb->page == NULL)
 405                         GOTO(cleanup, rc = -ENOMEM);
 406
 407                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS, 1);
 408
 409                 if (i_size_read(inode) < lnb->offset + lnb->len - 1)
 410                         lnb->rc = i_size_read(inode) - lnb->offset;
 411                 else
 412                         lnb->rc = lnb->len;
 413
 414                 tot_bytes += lnb->rc;
 415
 416                 if (PageUptodate(lnb->page)) {
 417                         lprocfs_counter_add(obd->obd_stats,
 418                                             LPROC_FILTER_CACHE_HIT, 1);
 419                         continue;
 420                 }
 421
 422                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_MISS, 1);
 423                 filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 424         }
 425         do_gettimeofday(&end);
 426         timediff = cfs_timeval_sub(&end, &start, NULL);
 427         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
 428
 429         if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
 430                 GOTO(cleanup, rc = -ENOMEM);
 431
 432         fsfilt_check_slow(obd, now, "start_page_read");
 433
 434         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf,
 435                               exp, NULL, NULL, NULL);
 436         if (rc)
 437                 GOTO(cleanup, rc);
 438
 439         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_READ_BYTES, tot_bytes);
 440         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
 441                 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
 442                                     LPROC_FILTER_READ_BYTES, tot_bytes);
 443
 444         EXIT;
 445
 446  cleanup:
 447         /* unlock pages to allow access from concurrent OST_READ */
 448         for (i = 0, lnb = res; i < *pages; i++, lnb++) {
 449                 if (lnb->page) {
 450                         LASSERT(PageLocked(lnb->page));
 451                         unlock_page(lnb->page);
 452
 453                         if (rc) {
 454                                 page_cache_release(lnb->page);
 455                                 lnb->page = NULL;
 456                         }
 457                 }
 458         }
 459
 460         if (rc != 0) {
 461                 if (dentry != NULL)
 462                         f_dput(dentry);
 463         }
 464
 465         filter_iobuf_put(&obd->u.filter, iobuf, oti);
 466
 467         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 468         if (rc)
 469                 CERROR("io error %d\n", rc);
 470
 471         return rc;
 472 }
 473
 474 /* When clients have dirtied as much space as they've been granted they
 475  * fall through to sync writes.  These sync writes haven't been expressed
 476  * in grants and need to error with ENOSPC when there isn't room in the
 477  * filesystem for them after grants are taken into account.  However,
 478  * writeback of the dirty data that was already granted space can write
 479  * right on through.
 480  *
 481  * Caller must hold obd_osfs_lock. */
 482 static int filter_grant_check(struct obd_export *exp, struct obdo *oa,
 483                               int objcount, struct fsfilt_objinfo *fso,
 484                               int niocount, struct niobuf_local *lnb,
 485                               obd_size *left, struct inode *inode)
 486 {
 487         struct filter_export_data *fed = &exp->exp_filter_data;
 488         int blocksize = exp->exp_obd->u.obt.obt_sb->s_blocksize;
 489         unsigned long used = 0, ungranted = 0, using;
 490         int i, rc = -ENOSPC, obj, n = 0;
 491
 492         LASSERT_SPIN_LOCKED(&exp->exp_obd->obd_osfs_lock);
 493
 494         for (obj = 0; obj < objcount; obj++) {
 495                 for (i = 0; i < fso[obj].fso_bufcnt; i++, n++) {
 496                         int tmp, bytes;
 497
 498                         /* should match the code in osc_exit_cache */
 499                         bytes = lnb[n].len;
 500                         bytes += lnb[n].offset & (blocksize - 1);
 501                         tmp = (lnb[n].offset + lnb[n].len) & (blocksize - 1);
 502                         if (tmp)
 503                                 bytes += blocksize - tmp;
 504
 505                         if ((lnb[n].flags & OBD_BRW_FROM_GRANT) &&
 506                             (oa->o_valid & OBD_MD_FLGRANT)) {
 507                                 if (fed->fed_grant < used + bytes) {
 508                                         CDEBUG(D_CACHE,
 509                                                "%s: cli %s/%p claims %ld+%d "
 510                                                "GRANT, real grant %lu idx %d\n",
 511                                                exp->exp_obd->obd_name,
 512                                                exp->exp_client_uuid.uuid, exp,
 513                                                used, bytes, fed->fed_grant, n);
 514                                 } else {
 515                                         used += bytes;
 516                                         lnb[n].flags |= OBD_BRW_GRANTED;
 517                                         lnb[n].lnb_grant_used = bytes;
 518                                         CDEBUG(0, "idx %d used=%lu\n", n, used);
 519                                         rc = 0;
 520                                         continue;
 521                                 }
 522                         }
 523                         if (*left > ungranted + bytes) {
 524                                 /* if enough space, pretend it was granted */
 525                                 ungranted += bytes;
 526                                 lnb[n].flags |= OBD_BRW_GRANTED;
 527                                 lnb[n].lnb_grant_used = bytes;
 528                                 CDEBUG(0, "idx %d ungranted=%lu\n",n,ungranted);
 529                                 rc = 0;
 530                                 continue;
 531                         }
 532
 533                         /* We can't check for already-mapped blocks here, as
 534                          * it requires dropping the osfs lock to do the bmap.
 535                          * Instead, we return ENOSPC and in that case we need
 536                          * to go through and verify if all of the blocks not
 537                          * marked BRW_GRANTED are already mapped and we can
 538                          * ignore this error. */
 539                         lnb[n].rc = -ENOSPC;
 540                         lnb[n].flags &= ~OBD_BRW_GRANTED;
 541                         CDEBUG(D_CACHE,"%s: cli %s/%p idx %d no space for %d\n",
 542                                exp->exp_obd->obd_name,
 543                                exp->exp_client_uuid.uuid, exp, n, bytes);
 544                 }
 545         }
 546
 547         /* Now substract what client have used already.  We don't subtract
 548          * this from the tot_granted yet, so that other client's can't grab
 549          * that space before we have actually allocated our blocks.  That
 550          * happens in filter_grant_commit() after the writes are done. */
 551         *left -= ungranted;
 552         fed->fed_grant -= used;
 553         fed->fed_pending += used + ungranted;
 554         exp->exp_obd->u.filter.fo_tot_granted += ungranted;
 555         exp->exp_obd->u.filter.fo_tot_pending += used + ungranted;
 556
 557         CDEBUG(D_CACHE,
 558                "%s: cli %s/%p used: %lu ungranted: %lu grant: %lu dirty: %lu\n",
 559                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp, used,
 560                ungranted, fed->fed_grant, fed->fed_dirty);
 561
 562         /* Rough calc in case we don't refresh cached statfs data */
 563         using = (used + ungranted + 1 ) >>
 564                 exp->exp_obd->u.obt.obt_sb->s_blocksize_bits;
 565         if (exp->exp_obd->obd_osfs.os_bavail > using)
 566                 exp->exp_obd->obd_osfs.os_bavail -= using;
 567         else
 568                 exp->exp_obd->obd_osfs.os_bavail = 0;
 569
 570         if (fed->fed_dirty < used) {
 571                 CERROR("%s: cli %s/%p claims used %lu > fed_dirty %lu\n",
 572                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 573                        used, fed->fed_dirty);
 574                 used = fed->fed_dirty;
 575         }
 576         exp->exp_obd->u.filter.fo_tot_dirty -= used;
 577         fed->fed_dirty -= used;
 578
 579         if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
 580                 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
 581                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 582                        fed->fed_dirty, fed->fed_pending, fed->fed_grant);
 583                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
 584                 LBUG();
 585         }
 586         return rc;
 587 }
 588
 589 /* If we ever start to support multi-object BRW RPCs, we will need to get locks
 590  * on mulitple inodes.  That isn't all, because there still exists the
 591  * possibility of a truncate starting a new transaction while holding the ext3
 592  * rwsem = write while some writes (which have started their transactions here)
 593  * blocking on the ext3 rwsem = read => lock inversion.
 594  *
 595  * The handling gets very ugly when dealing with locked pages.  It may be easier
 596  * to just get rid of the locked page code (which has problems of its own) and
 597  * either discover we do not need it anymore (i.e. it was a symptom of another
 598  * bug) or ensure we get the page locks in an appropriate order. */
 599 static int filter_preprw_write(int cmd, struct obd_export *exp, struct obdo *oa,
 600                                int objcount, struct obd_ioobj *obj,
 601                                struct niobuf_remote *nb, int *pages,
 602                                struct niobuf_local *res,
 603                                struct obd_trans_info *oti)
 604 {
 605         struct obd_device *obd = exp->exp_obd;
 606         struct timeval start, end;
 607         struct lvfs_run_ctxt saved;
 608         struct niobuf_local *lnb = res;
 609         struct fsfilt_objinfo fso;
 610         struct filter_mod_data *fmd;
 611         struct dentry *dentry = NULL;
 612         void *iobuf;
 613         obd_size left;
 614         unsigned long now = jiffies, timediff;
 615         int rc = 0, i, tot_bytes = 0, cleanup_phase = 0;
 616         ENTRY;
 617         LASSERT(objcount == 1);
 618         LASSERT(obj->ioo_bufcnt > 0);
 619
 620         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 621         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 622         if (IS_ERR(iobuf))
 623                 GOTO(cleanup, rc = PTR_ERR(iobuf));
 624         cleanup_phase = 1;
 625
 626         dentry = filter_fid2dentry(obd, NULL, obj->ioo_gr, obj->ioo_id);
 627         if (IS_ERR(dentry))
 628                 GOTO(cleanup, rc = PTR_ERR(dentry));
 629         cleanup_phase = 2;
 630
 631         if (dentry->d_inode == NULL) {
 632                 if (exp->exp_obd->obd_recovering) {
 633                         struct obdo *noa = oa;
 634
 635                         if (oa == NULL) {
 636                                 OBDO_ALLOC(noa);
 637                                 if (noa == NULL)
 638                                         GOTO(recreate_out, rc = -ENOMEM);
 639                                 noa->o_id = obj->ioo_id;
 640                                 noa->o_valid = OBD_MD_FLID;
 641                         }
 642
 643                         if (filter_recreate(exp->exp_obd, noa) == 0) {
 644                                 f_dput(dentry);
 645                                 dentry = filter_fid2dentry(exp->exp_obd, NULL,
 646                                                            obj->ioo_gr,
 647                                                            obj->ioo_id);
 648                         }
 649                         if (oa == NULL)
 650                                 OBDO_FREE(noa);
 651                 }
 652     recreate_out:
 653                 if (IS_ERR(dentry) || dentry->d_inode == NULL) {
 654                         CERROR("%s: BRW to missing obj "LPU64"/"LPU64":rc %d\n",
 655                                exp->exp_obd->obd_name,
 656                                obj->ioo_id, obj->ioo_gr,
 657                                IS_ERR(dentry) ? (int)PTR_ERR(dentry) : -ENOENT);
 658                         if (IS_ERR(dentry))
 659                                 cleanup_phase = 1;
 660                         GOTO(cleanup, rc = -ENOENT);
 661                 }
 662         }
 663
 664         rc = filter_map_remote_to_local(objcount, obj, nb, pages, res);
 665         if (rc)
 666                 GOTO(cleanup, rc);
 667
 668         fsfilt_check_slow(obd, now, "preprw_write setup");
 669
 670         /* Don't update inode timestamps if this write is older than a
 671          * setattr which modifies the timestamps. b=10150 */
 672         /* XXX when we start having persistent reservations this needs to
 673          * be changed to filter_fmd_get() to create the fmd if it doesn't
 674          * already exist so we can store the reservation handle there. */
 675         fmd = filter_fmd_find(exp, obj->ioo_id, obj->ioo_gr);
 676
 677         LASSERT(oa != NULL);
 678         spin_lock(&obd->obd_osfs_lock);
 679
 680         filter_grant_incoming(exp, oa);
 681         if (fmd && fmd->fmd_mactime_xid > oti->oti_xid)
 682                 oa->o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 683                                  OBD_MD_FLATIME);
 684         else
 685                 obdo_to_inode(dentry->d_inode, oa, OBD_MD_FLATIME |
 686                               OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 687         cleanup_phase = 3;
 688
 689         left = filter_grant_space_left(exp);
 690
 691         fso.fso_dentry = dentry;
 692         fso.fso_bufcnt = *pages;
 693
 694         rc = filter_grant_check(exp, oa, objcount, &fso, *pages, res,
 695                                 &left, dentry->d_inode);
 696
 697         /* do not zero out oa->o_valid as it is used in filter_commitrw_write()
 698          * for setting UID/GID and fid EA in first write time. */
 699         if (oa->o_valid & OBD_MD_FLGRANT)
 700                 oa->o_grant = filter_grant(exp,oa->o_grant,oa->o_undirty,left);
 701
 702         spin_unlock(&obd->obd_osfs_lock);
 703         filter_fmd_put(exp, fmd);
 704
 705         if (rc)
 706                 GOTO(cleanup, rc);
 707         cleanup_phase = 4;
 708
 709         do_gettimeofday(&start);
 710         for (i = 0, lnb = res; i < *pages; i++, lnb++) {
 711
 712                 /* We still set up for ungranted pages so that granted pages
 713                  * can be written to disk as they were promised, and portals
 714                  * needs to keep the pages all aligned properly. */
 715                 lnb->dentry = dentry;
 716
 717                 lnb->page = filter_get_page(obd, dentry->d_inode, lnb->offset);
 718                 if (lnb->page == NULL)
 719                         GOTO(cleanup, rc = -ENOMEM);
 720
 721                 /* DLM locking protects us from write and truncate competing
 722                  * for same region, but truncate can leave dirty page in the
 723                  * cache. it's possible the writeout on a such a page is in
 724                  * progress when we access it. it's also possible that during
 725                  * this writeout we put new (partial) data, but then won't
 726                  * be able to proceed in filter_commitrw_write(). thus let's
 727                  * just wait for writeout completion, should be rare enough.
 728                  * -bzzz */
 729                 if (obd->u.filter.fo_writethrough_cache)
 730                         wait_on_page_writeback(lnb->page);
 731                 BUG_ON(PageWriteback(lnb->page));
 732
 733                 /* If the filter writes a partial page, then has the file
 734                  * extended, the client will read in the whole page.  the
 735                  * filter has to be careful to zero the rest of the partial
 736                  * page on disk.  we do it by hand for partial extending
 737                  * writes, send_bio() is responsible for zeroing pages when
 738                  * asked to read unmapped blocks -- brw_kiovec() does this. */
 739                 if (lnb->len != CFS_PAGE_SIZE) {
 740                         __s64 maxidx;
 741
 742                         maxidx = ((i_size_read(dentry->d_inode) +
 743                                    CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT) - 1;
 744                         if (maxidx >= lnb->page->index) {
 745                                 LL_CDEBUG_PAGE(D_PAGE, lnb->page, "write %u @ "
 746                                                LPU64" flg %x before EOF %llu\n",
 747                                                lnb->len, lnb->offset,lnb->flags,
 748                                                i_size_read(dentry->d_inode));
 749                                 filter_iobuf_add_page(obd, iobuf,
 750                                                       dentry->d_inode,
 751                                                       lnb->page);
 752                         } else {
 753                                 long off;
 754                                 char *p = kmap(lnb->page);
 755
 756                                 off = lnb->offset & ~CFS_PAGE_MASK;
 757                                 if (off)
 758                                         memset(p, 0, off);
 759                                 off = (lnb->offset + lnb->len) & ~CFS_PAGE_MASK;
 760                                 if (off)
 761                                         memset(p + off, 0, CFS_PAGE_SIZE - off);
 762                                 kunmap(lnb->page);
 763                         }
 764                 }
 765                 if (lnb->rc == 0)
 766                         tot_bytes += lnb->len;
 767         }
 768         do_gettimeofday(&end);
 769         timediff = cfs_timeval_sub(&end, &start, NULL);
 770         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
 771
 772         if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
 773                 GOTO(cleanup, rc = -ENOMEM);
 774
 775         /* don't unlock pages to prevent any access */
 776         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf, exp,
 777                               NULL, NULL, NULL);
 778
 779         fsfilt_check_slow(obd, now, "start_page_write");
 780
 781         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
 782                             tot_bytes);
 783         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
 784                 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
 785                                     LPROC_FILTER_WRITE_BYTES, tot_bytes);
 786         EXIT;
 787 cleanup:
 788         switch(cleanup_phase) {
 789         case 4:
 790                 if (rc) {
 791                         for (i = 0, lnb = res; i < *pages; i++, lnb++) {
 792                                 if (lnb->page != NULL) {
 793                                         unlock_page(lnb->page);
 794                                         page_cache_release(lnb->page);
 795                                         lnb->page = NULL;
 796                                 }
 797                         }
 798                         filter_grant_commit(exp, *pages, res);
 799                 }
 800         case 3:
 801                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 802         case 2:
 803                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 804                 if (rc)
 805                         f_dput(dentry);
 806                 break;
 807         case 1:
 808                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 809         case 0:
 810                 spin_lock(&obd->obd_osfs_lock);
 811                 if (oa)
 812                         filter_grant_incoming(exp, oa);
 813                 spin_unlock(&obd->obd_osfs_lock);
 814                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 815                 break;
 816         default:;
 817         }
 818         return rc;
 819 }
 820
 821 int filter_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
 822                   int objcount, struct obd_ioobj *obj,
 823                   struct niobuf_remote *nb, int *pages,
 824                   struct niobuf_local *res, struct obd_trans_info *oti)
 825 {
 826         if (cmd == OBD_BRW_WRITE)
 827                 return filter_preprw_write(cmd, exp, oa, objcount, obj,
 828                                            nb, pages, res, oti);
 829         if (cmd == OBD_BRW_READ)
 830                 return filter_preprw_read(cmd, exp, oa, objcount, obj,
 831                                           nb, pages, res, oti);
 832         LBUG();
 833         return -EPROTO;
 834 }
 835
 836 static int filter_commitrw_read(struct obd_export *exp, struct obdo *oa,
 837                                 int objcount, struct obd_ioobj *obj,
 838                                 struct niobuf_remote *rnb,
 839                                 int pages, struct niobuf_local *res,
 840                                 struct obd_trans_info *oti, int rc)
 841 {
 842         struct filter_obd *fo = &exp->exp_obd->u.filter;
 843         struct inode *inode = NULL;
 844         struct ldlm_res_id res_id = { .name = { obj->ioo_id } };
 845         struct ldlm_resource *resource = NULL;
 846         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 847         struct niobuf_local *lnb;
 848         int i;
 849         ENTRY;
 850
 851         /* If oa != NULL then filter_preprw_read updated the inode atime
 852          * and we should update the lvb so that other glimpses will also
 853          * get the updated value. bug 5972 */
 854         if (oa && ns && ns->ns_lvbo && ns->ns_lvbo->lvbo_update) {
 855                 resource = ldlm_resource_get(ns, NULL, res_id, LDLM_EXTENT, 0);
 856
 857                 if (resource != NULL) {
 858                         ns->ns_lvbo->lvbo_update(resource, NULL, 0, 1);
 859                         ldlm_resource_putref(resource);
 860                 }
 861         }
 862
 863         if (res->dentry != NULL)
 864                 inode = res->dentry->d_inode;
 865
 866         for (i = 0, lnb = res; i < pages; i++, lnb++) {
 867                 if (lnb->page != NULL) {
 868                         page_cache_release(lnb->page);
 869                         lnb->page = NULL;
 870                 }
 871         }
 872
 873         if (inode && (fo->fo_read_cache == 0 ||
 874                         i_size_read(inode) > fo->fo_readcache_max_filesize))
 875                 filter_invalidate_cache(exp->exp_obd, obj, rnb, inode);
 876
 877         if (res->dentry != NULL)
 878                 f_dput(res->dentry);
 879         RETURN(rc);
 880 }
 881
 882 void filter_grant_commit(struct obd_export *exp, int niocount,
 883                          struct niobuf_local *res)
 884 {
 885         struct filter_obd *filter = &exp->exp_obd->u.filter;
 886         struct niobuf_local *lnb = res;
 887         unsigned long pending = 0;
 888         int i;
 889
 890         spin_lock(&exp->exp_obd->obd_osfs_lock);
 891         for (i = 0, lnb = res; i < niocount; i++, lnb++)
 892                 pending += lnb->lnb_grant_used;
 893
 894         LASSERTF(exp->exp_filter_data.fed_pending >= pending,
 895                  "%s: cli %s/%p fed_pending: %lu grant_used: %lu\n",
 896                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 897                  exp->exp_filter_data.fed_pending, pending);
 898         exp->exp_filter_data.fed_pending -= pending;
 899         LASSERTF(filter->fo_tot_granted >= pending,
 900                  "%s: cli %s/%p tot_granted: "LPU64" grant_used: %lu\n",
 901                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 902                  exp->exp_obd->u.filter.fo_tot_granted, pending);
 903         filter->fo_tot_granted -= pending;
 904         LASSERTF(filter->fo_tot_pending >= pending,
 905                  "%s: cli %s/%p tot_pending: "LPU64" grant_used: %lu\n",
 906                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 907                  filter->fo_tot_pending, pending);
 908         filter->fo_tot_pending -= pending;
 909
 910         spin_unlock(&exp->exp_obd->obd_osfs_lock);
 911 }
 912
 913 int filter_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
 914                     int objcount, struct obd_ioobj *obj,
 915                     struct niobuf_remote *nb, int pages,
 916                     struct niobuf_local *res, struct obd_trans_info *oti,
 917                     int rc)
 918 {
 919         if (cmd == OBD_BRW_WRITE)
 920                 return filter_commitrw_write(exp, oa, objcount, obj,
 921                                              nb, pages, res, oti, rc);
 922         if (cmd == OBD_BRW_READ)
 923                 return filter_commitrw_read(exp, oa, objcount, obj,
 924                                             nb, pages, res, oti, rc);
 925         LBUG();
 926         return -EPROTO;
 927 }
 928
 929 int filter_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
 930                obd_count oa_bufs, struct brw_page *pga,
 931                struct obd_trans_info *oti)
 932 {
 933         struct obd_ioobj ioo;
 934         struct niobuf_local *lnb;
 935         struct niobuf_remote *rnb;
 936         obd_count i;
 937         int ret = 0, npages;
 938         ENTRY;
 939
 940         OBD_ALLOC(lnb, oa_bufs * sizeof(struct niobuf_local));
 941         OBD_ALLOC(rnb, oa_bufs * sizeof(struct niobuf_remote));
 942
 943         if (lnb == NULL || rnb == NULL)
 944                 GOTO(out, ret = -ENOMEM);
 945
 946         for (i = 0; i < oa_bufs; i++) {
 947                 lnb[i].page = pga[i].pg;
 948                 rnb[i].offset = pga[i].off;
 949                 rnb[i].len = pga[i].count;
 950         }
 951
 952         obdo_to_ioobj(oinfo->oi_oa, &ioo);
 953         ioo.ioo_bufcnt = oa_bufs;
 954
 955         npages = oa_bufs;
 956         ret = filter_preprw(cmd, exp, oinfo->oi_oa, 1, &ioo,
 957                             rnb, &npages, lnb, oti);
 958         if (ret != 0)
 959                 GOTO(out, ret);
 960         LASSERTF(oa_bufs == npages, "%u != %u\n", oa_bufs, npages);
 961
 962         ret = filter_commitrw(cmd, exp, oinfo->oi_oa, 1, &ioo, rnb,
 963                               npages, lnb, oti, ret);
 964
 965 out:
 966         if (lnb)
 967                 OBD_FREE(lnb, oa_bufs * sizeof(struct niobuf_local));
 968         if (rnb)
 969                 OBD_FREE(rnb, oa_bufs * sizeof(struct niobuf_remote));
 970         RETURN(ret);
 971 }