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 (c) 2003, 2010, Oracle and/or its affiliates. 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                 cfs_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,
 156                                cfs_time_current_64() - CFS_HZ)) {
 157 restat:
 158                 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb,
 159                                    cfs_time_current_64() + CFS_HZ);
 160                 if (rc) /* N.B. statfs can't really fail */
 161                         RETURN(0);
 162                 statfs_done = 1;
 163         }
 164
 165         avail = obd->obd_osfs.os_bavail;
 166         left = avail - (avail >> (blockbits - 3)); /* (d)indirect */
 167         if (left > GRANT_FOR_LLOG(obd)) {
 168                 left = (left - GRANT_FOR_LLOG(obd)) << blockbits;
 169         } else {
 170                 left = 0 /* << blockbits */;
 171         }
 172
 173         if (!statfs_done && left < 32 * FILTER_GRANT_CHUNK + tot_granted) {
 174                 CDEBUG(D_CACHE, "fs has no space left and statfs too old\n");
 175                 goto restat;
 176         }
 177
 178         if (left >= tot_granted) {
 179                 left -= tot_granted;
 180         } else {
 181                 if (left < tot_granted - obd->u.filter.fo_tot_pending) {
 182                         CERROR("%s: cli %s/%p grant "LPU64" > available "
 183                                LPU64" and pending "LPU64"\n", obd->obd_name,
 184                                exp->exp_client_uuid.uuid, exp, tot_granted,
 185                                left, obd->u.filter.fo_tot_pending);
 186                 }
 187                 left = 0;
 188         }
 189
 190         CDEBUG(D_CACHE, "%s: cli %s/%p free: "LPU64" avail: "LPU64" grant "LPU64
 191                " left: "LPU64" pending: "LPU64"\n", obd->obd_name,
 192                exp->exp_client_uuid.uuid, exp,
 193                obd->obd_osfs.os_bfree << blockbits, avail << blockbits,
 194                tot_granted, left, obd->u.filter.fo_tot_pending);
 195
 196         return left;
 197 }
 198
 199 /* Calculate how much grant space to allocate to this client, based on how
 200  * much space is currently free and how much of that is already granted.
 201  *
 202  * if @conservative != 0, we limit the maximum grant to FILTER_GRANT_CHUNK;
 203  * otherwise we'll satisfy the requested amount as possible as we can, this
 204  * is usually due to client reconnect.
 205  *
 206  * Caller must hold obd_osfs_lock. */
 207 long filter_grant(struct obd_export *exp, obd_size current_grant,
 208                   obd_size want, obd_size fs_space_left, int conservative)
 209 {
 210         struct obd_device *obd = exp->exp_obd;
 211         struct filter_export_data *fed = &exp->exp_filter_data;
 212         int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
 213         __u64 grant = 0;
 214
 215         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
 216
 217         /* Grant some fraction of the client's requested grant space so that
 218          * they are not always waiting for write credits (not all of it to
 219          * avoid overgranting in face of multiple RPCs in flight).  This
 220          * essentially will be able to control the OSC_MAX_RIF for a client.
 221          *
 222          * If we do have a large disparity between what the client thinks it
 223          * has and what we think it has, don't grant very much and let the
 224          * client consume its grant first.  Either it just has lots of RPCs
 225          * in flight, or it was evicted and its grants will soon be used up. */
 226         if (want > 0x7fffffff) {
 227                 CERROR("%s: client %s/%p requesting > 2GB grant "LPU64"\n",
 228                        obd->obd_name, exp->exp_client_uuid.uuid, exp, want);
 229         } else if (current_grant < want &&
 230                    current_grant < fed->fed_grant + FILTER_GRANT_CHUNK) {
 231                 grant = min(want + (1 << blockbits) - 1, fs_space_left / 8);
 232                 grant &= ~((1ULL << blockbits) - 1);
 233
 234                 if (grant) {
 235                         if (grant > FILTER_GRANT_CHUNK && conservative)
 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                                 cfs_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 for requests from a local client not allowing to enter FS
 268  * as we might end up waiting on a page he sent in the request we're serving.
 269  * use __GFP_HIGHMEM so that the pages can use all of the available memory
 270  * on 32-bit machines
 271  * use more agressive GFP_HIGHUSER flags from non-local clients to be able to
 272  * generate more memory pressure, but at the same time use __GFP_NOMEMALLOC
 273  * in order not to exhaust emergency reserves.
 274  *
 275  * See Bug 19529 and Bug 19917 for details.
 276  */
 277 static struct page *filter_get_page(struct obd_device *obd,
 278                                     struct inode *inode,
 279                                     obd_off offset,
 280                                     int localreq)
 281 {
 282         struct page *page;
 283
 284         page = find_or_create_page(inode->i_mapping, offset >> CFS_PAGE_SHIFT,
 285                                    (localreq ? (GFP_NOFS | __GFP_HIGHMEM)
 286                                              : GFP_HIGHUSER));
 287         if (unlikely(page == NULL))
 288                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_NO_PAGE, 1);
 289
 290         return page;
 291 }
 292
 293 /*
 294  * the routine initializes array of local_niobuf from remote_niobuf
 295  */
 296 static int filter_map_remote_to_local(int objcount, struct obd_ioobj *obj,
 297                                       struct niobuf_remote *nb,
 298                                       int *nrpages, struct niobuf_local *res)
 299 {
 300         struct niobuf_remote *rnb;
 301         struct niobuf_local *lnb;
 302         int i, max;
 303         ENTRY;
 304
 305         /* we don't support multiobject RPC yet
 306          * ost_brw_read() and ost_brw_write() check this */
 307         LASSERT(objcount == 1);
 308
 309         max = *nrpages;
 310         *nrpages = 0;
 311         for (i = 0, rnb = nb, lnb = res; i < obj->ioo_bufcnt; i++, rnb++) {
 312                 obd_off offset = rnb->offset;
 313                 unsigned int len = rnb->len;
 314
 315                 while (len > 0) {
 316                         int poff = offset & (CFS_PAGE_SIZE - 1);
 317                         int plen = CFS_PAGE_SIZE - poff;
 318
 319                         if (*nrpages >= max) {
 320                                 CERROR("small array of local bufs: %d\n", max);
 321                                 RETURN(-EINVAL);
 322                         }
 323
 324                         if (plen > len)
 325                                 plen = len;
 326                         lnb->offset = offset;
 327                         lnb->len = plen;
 328                         lnb->flags = rnb->flags;
 329                         lnb->page = NULL;
 330                         lnb->rc = 0;
 331                         lnb->lnb_grant_used = 0;
 332
 333                         LASSERTF(plen <= len, "plen %u, len %u\n", plen, len);
 334                         offset += plen;
 335                         len -= plen;
 336                         lnb++;
 337                         (*nrpages)++;
 338                 }
 339         }
 340         RETURN(0);
 341 }
 342
 343 /*
 344  * the invalidate above doesn't work during read because lnet pins pages.
 345  * The truncate is used here instead to drop pages from cache
 346  */
 347 void filter_release_cache(struct obd_device *obd, struct obd_ioobj *obj,
 348                           struct niobuf_remote *rnb, struct inode *inode)
 349 {
 350         int i;
 351
 352         LASSERT(inode != NULL);
 353         for (i = 0; i < obj->ioo_bufcnt; i++, rnb++) {
 354 #ifdef HAVE_TRUNCATE_RANGE
 355                 /* remove pages in which range is fit */
 356                 truncate_inode_pages_range(inode->i_mapping,
 357                                            rnb->offset & CFS_PAGE_MASK,
 358                                            (rnb->offset + rnb->len - 1) |
 359                                            ~CFS_PAGE_MASK);
 360 #else
 361                 /* use invalidate for old kernels */
 362                 invalidate_mapping_pages(inode->i_mapping,
 363                                          rnb->offset >> CFS_PAGE_SHIFT,
 364                                          (rnb->offset + rnb->len) >>
 365                                          CFS_PAGE_SHIFT);
 366 #endif
 367         }
 368 }
 369
 370 static int filter_preprw_read(int cmd, struct obd_export *exp, struct obdo *oa,
 371                               int objcount, struct obd_ioobj *obj,
 372                               struct niobuf_remote *nb,
 373                               int *npages, struct niobuf_local *res,
 374                               struct obd_trans_info *oti,
 375                               struct lustre_capa *capa)
 376 {
 377         struct obd_device *obd = exp->exp_obd;
 378         struct timeval start, end;
 379         struct lvfs_run_ctxt saved;
 380         struct niobuf_local *lnb;
 381         struct dentry *dentry = NULL;
 382         struct inode *inode = NULL;
 383         void *iobuf = NULL;
 384         int rc = 0, i, tot_bytes = 0;
 385         unsigned long now = jiffies;
 386         long timediff;
 387         loff_t isize;
 388         ENTRY;
 389
 390         /* We are currently not supporting multi-obj BRW_READ RPCS at all.
 391          * When we do this function's dentry cleanup will need to be fixed.
 392          * These values are verified in ost_brw_write() from the wire. */
 393         LASSERTF(objcount == 1, "%d\n", objcount);
 394         LASSERTF(obj->ioo_bufcnt > 0, "%d\n", obj->ioo_bufcnt);
 395
 396         rc = filter_auth_capa(exp, NULL, oa->o_seq, capa,
 397                               CAPA_OPC_OSS_READ);
 398         if (rc)
 399                 RETURN(rc);
 400
 401         if (oa && oa->o_valid & OBD_MD_FLGRANT) {
 402                 cfs_spin_lock(&obd->obd_osfs_lock);
 403                 filter_grant_incoming(exp, oa);
 404
 405                 if (!(oa->o_flags & OBD_FL_SHRINK_GRANT))
 406                         oa->o_grant = 0;
 407                 cfs_spin_unlock(&obd->obd_osfs_lock);
 408         }
 409
 410         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 411         if (IS_ERR(iobuf))
 412                 RETURN(PTR_ERR(iobuf));
 413
 414         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 415         dentry = filter_oa2dentry(obd, &oa->o_oi);
 416         if (IS_ERR(dentry)) {
 417                 rc = PTR_ERR(dentry);
 418                 dentry = NULL;
 419                 GOTO(cleanup, rc);
 420         }
 421
 422         inode = dentry->d_inode;
 423         isize = i_size_read(inode);
 424
 425         obdo_to_inode(inode, oa, OBD_MD_FLATIME);
 426
 427         rc = filter_map_remote_to_local(objcount, obj, nb, npages, res);
 428         if (rc)
 429                 GOTO(cleanup, rc);
 430
 431         fsfilt_check_slow(obd, now, "preprw_read setup");
 432
 433         /* find pages for all segments, fill array with them */
 434         cfs_gettimeofday(&start);
 435         for (i = 0, lnb = res; i < *npages; i++, lnb++) {
 436
 437                 lnb->dentry = dentry;
 438
 439                 if (isize <= lnb->offset)
 440                         /* If there's no more data, abort early.  lnb->rc == 0,
 441                          * so it's easy to detect later. */
 442                         break;
 443
 444                 lnb->page = filter_get_page(obd, inode, lnb->offset, 0);
 445                 if (lnb->page == NULL)
 446                         GOTO(cleanup, rc = -ENOMEM);
 447
 448                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS, 1);
 449
 450                 if (isize < lnb->offset + lnb->len - 1)
 451                         lnb->rc = isize - lnb->offset;
 452                 else
 453                         lnb->rc = lnb->len;
 454
 455                 tot_bytes += lnb->rc;
 456
 457                 if (PageUptodate(lnb->page)) {
 458                         lprocfs_counter_add(obd->obd_stats,
 459                                             LPROC_FILTER_CACHE_HIT, 1);
 460                         continue;
 461                 }
 462
 463                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_MISS, 1);
 464                 filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 465         }
 466         cfs_gettimeofday(&end);
 467         timediff = cfs_timeval_sub(&end, &start, NULL);
 468         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
 469
 470         if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
 471                 GOTO(cleanup, rc = -ENOMEM);
 472
 473         fsfilt_check_slow(obd, now, "start_page_read");
 474
 475         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf,
 476                               exp, NULL, NULL, NULL);
 477         if (rc)
 478                 GOTO(cleanup, rc);
 479
 480         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_READ_BYTES, tot_bytes);
 481
 482         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
 483                 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
 484                                     LPROC_FILTER_READ_BYTES, tot_bytes);
 485
 486         EXIT;
 487
 488  cleanup:
 489         /* unlock pages to allow access from concurrent OST_READ */
 490         for (i = 0, lnb = res; i < *npages; i++, lnb++) {
 491                 if (lnb->page) {
 492                         LASSERT(PageLocked(lnb->page));
 493                         unlock_page(lnb->page);
 494
 495                         if (rc) {
 496                                 page_cache_release(lnb->page);
 497                                 lnb->page = NULL;
 498                         }
 499                 }
 500         }
 501
 502         if (rc != 0) {
 503                 if (dentry != NULL)
 504                         f_dput(dentry);
 505         }
 506
 507         filter_iobuf_put(&obd->u.filter, iobuf, oti);
 508
 509         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 510         if (rc)
 511                 CERROR("io error %d\n", rc);
 512
 513         return rc;
 514 }
 515
 516 /* When clients have dirtied as much space as they've been granted they
 517  * fall through to sync writes.  These sync writes haven't been expressed
 518  * in grants and need to error with ENOSPC when there isn't room in the
 519  * filesystem for them after grants are taken into account.  However,
 520  * writeback of the dirty data that was already granted space can write
 521  * right on through.
 522  *
 523  * Caller must hold obd_osfs_lock. */
 524 static int filter_grant_check(struct obd_export *exp, struct obdo *oa,
 525                               int objcount, struct fsfilt_objinfo *fso,
 526                               int niocount, struct niobuf_local *lnb,
 527                               obd_size *left, struct inode *inode)
 528 {
 529         struct filter_export_data *fed = &exp->exp_filter_data;
 530         int blocksize = exp->exp_obd->u.obt.obt_sb->s_blocksize;
 531         unsigned long used = 0, ungranted = 0, using;
 532         int i, rc = -ENOSPC, obj, n = 0;
 533
 534         LASSERT_SPIN_LOCKED(&exp->exp_obd->obd_osfs_lock);
 535
 536         for (obj = 0; obj < objcount; obj++) {
 537                 for (i = 0; i < fso[obj].fso_bufcnt; i++, n++) {
 538                         int tmp, bytes;
 539
 540                         /* should match the code in osc_exit_cache */
 541                         bytes = lnb[n].len;
 542                         bytes += lnb[n].offset & (blocksize - 1);
 543                         tmp = (lnb[n].offset + lnb[n].len) & (blocksize - 1);
 544                         if (tmp)
 545                                 bytes += blocksize - tmp;
 546
 547                         if ((lnb[n].flags & OBD_BRW_FROM_GRANT) &&
 548                             (oa->o_valid & OBD_MD_FLGRANT)) {
 549                                 if (fed->fed_grant < used + bytes) {
 550                                         CDEBUG(D_CACHE,
 551                                                "%s: cli %s/%p claims %ld+%d "
 552                                                "GRANT, real grant %lu idx %d\n",
 553                                                exp->exp_obd->obd_name,
 554                                                exp->exp_client_uuid.uuid, exp,
 555                                                used, bytes, fed->fed_grant, n);
 556                                 } else {
 557                                         used += bytes;
 558                                         lnb[n].flags |= OBD_BRW_GRANTED;
 559                                         lnb[n].lnb_grant_used = bytes;
 560                                         CDEBUG(0, "idx %d used=%lu\n", n, used);
 561                                         rc = 0;
 562                                         continue;
 563                                 }
 564                         }
 565                         if (*left > ungranted + bytes) {
 566                                 /* if enough space, pretend it was granted */
 567                                 ungranted += bytes;
 568                                 lnb[n].flags |= OBD_BRW_GRANTED;
 569                                 lnb[n].lnb_grant_used = bytes;
 570                                 CDEBUG(0, "idx %d ungranted=%lu\n",n,ungranted);
 571                                 rc = 0;
 572                                 continue;
 573                         }
 574
 575                         /* We can't check for already-mapped blocks here, as
 576                          * it requires dropping the osfs lock to do the bmap.
 577                          * Instead, we return ENOSPC and in that case we need
 578                          * to go through and verify if all of the blocks not
 579                          * marked BRW_GRANTED are already mapped and we can
 580                          * ignore this error. */
 581                         lnb[n].rc = -ENOSPC;
 582                         lnb[n].flags &= ~OBD_BRW_GRANTED;
 583                         CDEBUG(D_CACHE,"%s: cli %s/%p idx %d no space for %d\n",
 584                                exp->exp_obd->obd_name,
 585                                exp->exp_client_uuid.uuid, exp, n, bytes);
 586                 }
 587         }
 588
 589         /* Now substract what client have used already.  We don't subtract
 590          * this from the tot_granted yet, so that other client's can't grab
 591          * that space before we have actually allocated our blocks.  That
 592          * happens in filter_grant_commit() after the writes are done. */
 593         *left -= ungranted;
 594         fed->fed_grant -= used;
 595         fed->fed_pending += used + ungranted;
 596         exp->exp_obd->u.filter.fo_tot_granted += ungranted;
 597         exp->exp_obd->u.filter.fo_tot_pending += used + ungranted;
 598
 599         CDEBUG(D_CACHE,
 600                "%s: cli %s/%p used: %lu ungranted: %lu grant: %lu dirty: %lu\n",
 601                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp, used,
 602                ungranted, fed->fed_grant, fed->fed_dirty);
 603
 604         /* Rough calc in case we don't refresh cached statfs data */
 605         using = (used + ungranted + 1 ) >>
 606                 exp->exp_obd->u.obt.obt_sb->s_blocksize_bits;
 607         if (exp->exp_obd->obd_osfs.os_bavail > using)
 608                 exp->exp_obd->obd_osfs.os_bavail -= using;
 609         else
 610                 exp->exp_obd->obd_osfs.os_bavail = 0;
 611
 612         if (fed->fed_dirty < used) {
 613                 CERROR("%s: cli %s/%p claims used %lu > fed_dirty %lu\n",
 614                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 615                        used, fed->fed_dirty);
 616                 used = fed->fed_dirty;
 617         }
 618         exp->exp_obd->u.filter.fo_tot_dirty -= used;
 619         fed->fed_dirty -= used;
 620
 621         if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
 622                 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
 623                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 624                        fed->fed_dirty, fed->fed_pending, fed->fed_grant);
 625                 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
 626                 LBUG();
 627         }
 628         return rc;
 629 }
 630
 631 /* If we ever start to support multi-object BRW RPCs, we will need to get locks
 632  * on mulitple inodes.  That isn't all, because there still exists the
 633  * possibility of a truncate starting a new transaction while holding the ext3
 634  * rwsem = write while some writes (which have started their transactions here)
 635  * blocking on the ext3 rwsem = read => lock inversion.
 636  *
 637  * The handling gets very ugly when dealing with locked pages.  It may be easier
 638  * to just get rid of the locked page code (which has problems of its own) and
 639  * either discover we do not need it anymore (i.e. it was a symptom of another
 640  * bug) or ensure we get the page locks in an appropriate order. */
 641 static int filter_preprw_write(int cmd, struct obd_export *exp, struct obdo *oa,
 642                                int objcount, struct obd_ioobj *obj,
 643                                struct niobuf_remote *nb, int *npages,
 644                                struct niobuf_local *res,
 645                                struct obd_trans_info *oti,
 646                                struct lustre_capa *capa)
 647 {
 648         struct obd_device *obd = exp->exp_obd;
 649         struct timeval start, end;
 650         struct lvfs_run_ctxt saved;
 651         struct niobuf_local *lnb = res;
 652         struct fsfilt_objinfo fso;
 653         struct filter_mod_data *fmd;
 654         struct dentry *dentry = NULL;
 655         void *iobuf;
 656         obd_size left;
 657         unsigned long now = jiffies, timediff;
 658         int rc = 0, i, tot_bytes = 0, cleanup_phase = 0, localreq = 0;
 659         ENTRY;
 660         LASSERT(objcount == 1);
 661         LASSERT(obj->ioo_bufcnt > 0);
 662
 663         rc = filter_auth_capa(exp, NULL, oa->o_seq, capa,
 664                               CAPA_OPC_OSS_WRITE);
 665         if (rc)
 666                 RETURN(rc);
 667
 668         if (exp->exp_connection &&
 669             exp->exp_connection->c_peer.nid == exp->exp_connection->c_self)
 670                 localreq = 1;
 671
 672         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 673         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 674         if (IS_ERR(iobuf))
 675                 GOTO(cleanup, rc = PTR_ERR(iobuf));
 676         cleanup_phase = 1;
 677
 678         dentry = filter_fid2dentry(obd, NULL, obj->ioo_seq,
 679                                    obj->ioo_id);
 680         if (IS_ERR(dentry))
 681                 GOTO(cleanup, rc = PTR_ERR(dentry));
 682         cleanup_phase = 2;
 683
 684         if (dentry->d_inode == NULL) {
 685                 CERROR("%s: trying to BRW to non-existent file "LPU64"\n",
 686                        obd->obd_name, obj->ioo_id);
 687                 GOTO(cleanup, rc = -ENOENT);
 688         }
 689
 690         if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID) &&
 691             dentry->d_inode->i_mode & (S_ISUID | S_ISGID)) {
 692                 rc = filter_capa_fixoa(exp, oa, oa->o_seq, capa);
 693                 if (rc)
 694                         GOTO(cleanup, rc);
 695         }
 696
 697         rc = filter_map_remote_to_local(objcount, obj, nb, npages, res);
 698         if (rc)
 699                 GOTO(cleanup, rc);
 700
 701         fsfilt_check_slow(obd, now, "preprw_write setup");
 702
 703         /* Filter truncate first locks i_mutex then partially truncated
 704          * page, filter write code first locks pages then take
 705          * i_mutex.  To avoid a deadlock in case of concurrent
 706          * punch/write requests from one client, filter writes and
 707          * filter truncates are serialized by i_alloc_sem, allowing
 708          * multiple writes or single truncate. */
 709         down_read(&dentry->d_inode->i_alloc_sem);
 710
 711         /* Don't update inode timestamps if this write is older than a
 712          * setattr which modifies the timestamps. b=10150 */
 713         /* XXX when we start having persistent reservations this needs to
 714          * be changed to filter_fmd_get() to create the fmd if it doesn't
 715          * already exist so we can store the reservation handle there. */
 716         fmd = filter_fmd_find(exp, obj->ioo_id, obj->ioo_seq);
 717
 718         LASSERT(oa != NULL);
 719         cfs_spin_lock(&obd->obd_osfs_lock);
 720         filter_grant_incoming(exp, oa);
 721         if (fmd && fmd->fmd_mactime_xid > oti->oti_xid)
 722                 oa->o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 723                                  OBD_MD_FLATIME);
 724         else
 725                 obdo_to_inode(dentry->d_inode, oa, OBD_MD_FLATIME |
 726                               OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 727         cleanup_phase = 3;
 728
 729         left = filter_grant_space_left(exp);
 730
 731         fso.fso_dentry = dentry;
 732         fso.fso_bufcnt = *npages;
 733
 734         rc = filter_grant_check(exp, oa, objcount, &fso, *npages, res,
 735                                 &left, dentry->d_inode);
 736
 737         /* do not zero out oa->o_valid as it is used in filter_commitrw_write()
 738          * for setting UID/GID and fid EA in first write time. */
 739         if (oa->o_valid & OBD_MD_FLGRANT)
 740                 oa->o_grant = filter_grant(exp, oa->o_grant, oa->o_undirty,
 741                                            left, 1);
 742
 743         cfs_spin_unlock(&obd->obd_osfs_lock);
 744         filter_fmd_put(exp, fmd);
 745
 746         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK2, (obd_timeout + 1) / 4);
 747
 748         if (rc)
 749                 GOTO(cleanup, rc);
 750         cleanup_phase = 4;
 751
 752         cfs_gettimeofday(&start);
 753         for (i = 0, lnb = res; i < *npages; i++, lnb++) {
 754
 755                 /* We still set up for ungranted pages so that granted pages
 756                  * can be written to disk as they were promised, and portals
 757                  * needs to keep the pages all aligned properly. */
 758                 lnb->dentry = dentry;
 759
 760                 lnb->page = filter_get_page(obd, dentry->d_inode, lnb->offset,
 761                                             localreq);
 762                 if (lnb->page == NULL)
 763                         GOTO(cleanup, rc = -ENOMEM);
 764
 765                 /* DLM locking protects us from write and truncate competing
 766                  * for same region, but truncate can leave dirty page in the
 767                  * cache. it's possible the writeout on a such a page is in
 768                  * progress when we access it. it's also possible that during
 769                  * this writeout we put new (partial) data, but then won't
 770                  * be able to proceed in filter_commitrw_write(). thus let's
 771                  * just wait for writeout completion, should be rare enough.
 772                  * -bzzz */
 773                 wait_on_page_writeback(lnb->page);
 774                 BUG_ON(PageWriteback(lnb->page));
 775
 776                 /* If the filter writes a partial page, then has the file
 777                  * extended, the client will read in the whole page.  the
 778                  * filter has to be careful to zero the rest of the partial
 779                  * page on disk.  we do it by hand for partial extending
 780                  * writes, send_bio() is responsible for zeroing pages when
 781                  * asked to read unmapped blocks -- brw_kiovec() does this. */
 782                 if (lnb->len != CFS_PAGE_SIZE) {
 783                         __s64 maxidx;
 784
 785                         maxidx = ((i_size_read(dentry->d_inode) +
 786                                    CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT) - 1;
 787                         if (maxidx >= lnb->page->index) {
 788                                 LL_CDEBUG_PAGE(D_PAGE, lnb->page, "write %u @ "
 789                                                LPU64" flg %x before EOF %llu\n",
 790                                                lnb->len, lnb->offset,lnb->flags,
 791                                                i_size_read(dentry->d_inode));
 792                                 filter_iobuf_add_page(obd, iobuf,
 793                                                       dentry->d_inode,
 794                                                       lnb->page);
 795                         } else {
 796                                 long off;
 797                                 char *p = kmap(lnb->page);
 798
 799                                 off = lnb->offset & ~CFS_PAGE_MASK;
 800                                 if (off)
 801                                         memset(p, 0, off);
 802                                 off = (lnb->offset + lnb->len) & ~CFS_PAGE_MASK;
 803                                 if (off)
 804                                         memset(p + off, 0, CFS_PAGE_SIZE - off);
 805                                 kunmap(lnb->page);
 806                         }
 807                 }
 808                 if (lnb->rc == 0)
 809                         tot_bytes += lnb->len;
 810         }
 811         cfs_gettimeofday(&end);
 812         timediff = cfs_timeval_sub(&end, &start, NULL);
 813         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
 814
 815         if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
 816                 GOTO(cleanup, rc = -ENOMEM);
 817
 818         /* don't unlock pages to prevent any access */
 819         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf, exp,
 820                               NULL, NULL, NULL);
 821
 822         fsfilt_check_slow(obd, now, "start_page_write");
 823
 824         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
 825                             tot_bytes);
 826
 827         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
 828                 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
 829                                     LPROC_FILTER_WRITE_BYTES, tot_bytes);
 830         EXIT;
 831 cleanup:
 832         switch(cleanup_phase) {
 833         case 4:
 834                 if (rc) {
 835                         for (i = 0, lnb = res; i < *npages; i++, lnb++) {
 836                                 if (lnb->page != NULL) {
 837                                         unlock_page(lnb->page);
 838                                         page_cache_release(lnb->page);
 839                                         lnb->page = NULL;
 840                                 }
 841                         }
 842                 }
 843         case 3:
 844                 if (rc)
 845                         up_read(&dentry->d_inode->i_alloc_sem);
 846
 847                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 848         case 2:
 849                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 850                 if (rc)
 851                         f_dput(dentry);
 852                 break;
 853         case 1:
 854                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 855         case 0:
 856                 cfs_spin_lock(&obd->obd_osfs_lock);
 857                 if (oa)
 858                         filter_grant_incoming(exp, oa);
 859                 cfs_spin_unlock(&obd->obd_osfs_lock);
 860                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 861                 break;
 862         default:;
 863         }
 864         return rc;
 865 }
 866
 867 int filter_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
 868                   int objcount, struct obd_ioobj *obj,
 869                   struct niobuf_remote *nb, int *npages,
 870                   struct niobuf_local *res, struct obd_trans_info *oti,
 871                   struct lustre_capa *capa)
 872 {
 873         if (cmd == OBD_BRW_WRITE)
 874                 return filter_preprw_write(cmd, exp, oa, objcount, obj,
 875                                            nb, npages, res, oti, capa);
 876         if (cmd == OBD_BRW_READ)
 877                 return filter_preprw_read(cmd, exp, oa, objcount, obj,
 878                                           nb, npages, res, oti, capa);
 879         LBUG();
 880         return -EPROTO;
 881 }
 882
 883 static int filter_commitrw_read(struct obd_export *exp, struct obdo *oa,
 884                                 int objcount, struct obd_ioobj *obj,
 885                                 struct niobuf_remote *rnb,
 886                                 int npages, struct niobuf_local *res,
 887                                 struct obd_trans_info *oti, int rc)
 888 {
 889         struct filter_obd *fo = &exp->exp_obd->u.filter;
 890         struct inode *inode = NULL;
 891         struct ldlm_res_id res_id;
 892         struct ldlm_resource *resource = NULL;
 893         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 894         struct niobuf_local *lnb;
 895         int i;
 896         ENTRY;
 897
 898         osc_build_res_name(obj->ioo_id, obj->ioo_seq, &res_id);
 899         /* If oa != NULL then filter_preprw_read updated the inode atime
 900          * and we should update the lvb so that other glimpses will also
 901          * get the updated value. bug 5972 */
 902         if (oa && ns && ns->ns_lvbo && ns->ns_lvbo->lvbo_update) {
 903                 resource = ldlm_resource_get(ns, NULL, &res_id, LDLM_EXTENT, 0);
 904
 905                 if (resource != NULL) {
 906                         LDLM_RESOURCE_ADDREF(resource);
 907                         ns->ns_lvbo->lvbo_update(resource, NULL, 1);
 908                         LDLM_RESOURCE_DELREF(resource);
 909                         ldlm_resource_putref(resource);
 910                 }
 911         }
 912
 913         if (res->dentry != NULL)
 914                 inode = res->dentry->d_inode;
 915
 916         for (i = 0, lnb = res; i < npages; i++, lnb++) {
 917                 if (lnb->page != NULL) {
 918                         page_cache_release(lnb->page);
 919                         lnb->page = NULL;
 920                 }
 921         }
 922         if (inode && (fo->fo_read_cache == 0 ||
 923                       i_size_read(inode) > fo->fo_readcache_max_filesize))
 924                 filter_release_cache(exp->exp_obd, obj, rnb, inode);
 925
 926         if (res->dentry != NULL)
 927                 f_dput(res->dentry);
 928         RETURN(rc);
 929 }
 930
 931 void filter_grant_commit(struct obd_export *exp, int niocount,
 932                          struct niobuf_local *res)
 933 {
 934         struct filter_obd *filter = &exp->exp_obd->u.filter;
 935         struct niobuf_local *lnb = res;
 936         unsigned long pending = 0;
 937         int i;
 938
 939         cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
 940         for (i = 0, lnb = res; i < niocount; i++, lnb++)
 941                 pending += lnb->lnb_grant_used;
 942
 943         LASSERTF(exp->exp_filter_data.fed_pending >= pending,
 944                  "%s: cli %s/%p fed_pending: %lu grant_used: %lu\n",
 945                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 946                  exp->exp_filter_data.fed_pending, pending);
 947         exp->exp_filter_data.fed_pending -= pending;
 948         LASSERTF(filter->fo_tot_granted >= pending,
 949                  "%s: cli %s/%p tot_granted: "LPU64" grant_used: %lu\n",
 950                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 951                  exp->exp_obd->u.filter.fo_tot_granted, pending);
 952         filter->fo_tot_granted -= pending;
 953         LASSERTF(filter->fo_tot_pending >= pending,
 954                  "%s: cli %s/%p tot_pending: "LPU64" grant_used: %lu\n",
 955                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 956                  filter->fo_tot_pending, pending);
 957         filter->fo_tot_pending -= pending;
 958
 959         cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
 960 }
 961
 962 int filter_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
 963                     int objcount, struct obd_ioobj *obj,
 964                     struct niobuf_remote *nb, int npages,
 965                     struct niobuf_local *res, struct obd_trans_info *oti,
 966                     int rc)
 967 {
 968         if (cmd == OBD_BRW_WRITE)
 969                 return filter_commitrw_write(exp, oa, objcount, obj,
 970                                              nb, npages, res, oti, rc);
 971         if (cmd == OBD_BRW_READ)
 972                 return filter_commitrw_read(exp, oa, objcount, obj,
 973                                             nb, npages, res, oti, rc);
 974         LBUG();
 975         return -EPROTO;
 976 }
 977
 978 int filter_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
 979                obd_count oa_bufs, struct brw_page *pga,
 980                struct obd_trans_info *oti)
 981 {
 982         struct obd_ioobj ioo;
 983         struct niobuf_local *lnb;
 984         struct niobuf_remote *rnb;
 985         obd_count i;
 986         int ret = 0, npages;
 987         ENTRY;
 988
 989         OBD_ALLOC(lnb, oa_bufs * sizeof(struct niobuf_local));
 990         OBD_ALLOC(rnb, oa_bufs * sizeof(struct niobuf_remote));
 991
 992         if (lnb == NULL || rnb == NULL)
 993                 GOTO(out, ret = -ENOMEM);
 994
 995         for (i = 0; i < oa_bufs; i++) {
 996                 lnb[i].page = pga[i].pg;
 997                 rnb[i].offset = pga[i].off;
 998                 rnb[i].len = pga[i].count;
 999         }
1000
1001         obdo_to_ioobj(oinfo->oi_oa, &ioo);
1002         ioo.ioo_bufcnt = oa_bufs;
1003
1004         npages = oa_bufs;
1005         ret = filter_preprw(cmd, exp, oinfo->oi_oa, 1, &ioo,
1006                             rnb, &npages, lnb, oti, oinfo_capa(oinfo));
1007         if (ret != 0)
1008                 GOTO(out, ret);
1009         LASSERTF(oa_bufs == npages, "%u != %u\n", oa_bufs, npages);
1010
1011         ret = filter_commitrw(cmd, exp, oinfo->oi_oa, 1, &ioo, rnb,
1012                               npages, lnb, oti, ret);
1013
1014 out:
1015         if (lnb)
1016                 OBD_FREE(lnb, oa_bufs * sizeof(struct niobuf_local));
1017         if (rnb)
1018                 OBD_FREE(rnb, oa_bufs * sizeof(struct niobuf_remote));
1019         RETURN(ret);
1020 }