lustre/obdfilter/filter_io.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/obdfilter/filter_io.c
  37  *
  38  * Author: Peter Braam <braam@clusterfs.com>
  39  * Author: Andreas Dilger <adilger@clusterfs.com>
  40  * Author: Phil Schwan <phil@clusterfs.com>
  41  */
  42
  43 #define DEBUG_SUBSYSTEM S_FILTER
  44
  45 #ifndef AUTOCONF_INCLUDED
  46 #include <linux/config.h>
  47 #endif
  48 #include <linux/module.h>
  49 #include <linux/pagemap.h> // XXX kill me soon
  50 #include <linux/version.h>
  51
  52 #include <obd_class.h>
  53 #include <obd_ost.h>
  54 #include <lustre_fsfilt.h>
  55 #include "filter_internal.h"
  56
  57 int *obdfilter_created_scratchpad;
  58
  59 /* Grab the dirty and seen grant announcements from the incoming obdo.
  60  * We will later calculate the clients new grant and return it.
  61  * Caller must hold osfs lock */
  62 static 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         if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
 112                 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
 113                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
 114                        fed->fed_dirty, fed->fed_pending, fed->fed_grant);
 115                 spin_unlock(&obd->obd_osfs_lock);
 116                 LBUG();
 117         }
 118         EXIT;
 119 }
 120
 121 /* Figure out how much space is available between what we've granted
 122  * and what remains in the filesystem.  Compensate for ext3 indirect
 123  * block overhead when computing how much free space is left ungranted.
 124  *
 125  * Caller must hold obd_osfs_lock. */
 126 obd_size filter_grant_space_left(struct obd_export *exp)
 127 {
 128         struct obd_device *obd = exp->exp_obd;
 129         int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
 130         obd_size tot_granted = obd->u.filter.fo_tot_granted, avail, left = 0;
 131         int rc, statfs_done = 0;
 132
 133         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
 134
 135         if (cfs_time_before_64(obd->obd_osfs_age, cfs_time_current_64() - HZ)) {
 136 restat:
 137                 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb,
 138                                    cfs_time_current_64() + HZ);
 139                 if (rc) /* N.B. statfs can't really fail */
 140                         RETURN(0);
 141                 statfs_done = 1;
 142         }
 143
 144         avail = obd->obd_osfs.os_bavail;
 145         left = avail - (avail >> (blockbits - 3)); /* (d)indirect */
 146         if (left > GRANT_FOR_LLOG(obd)) {
 147                 left = (left - GRANT_FOR_LLOG(obd)) << blockbits;
 148         } else {
 149                 left = 0 /* << blockbits */;
 150         }
 151
 152         if (!statfs_done && left < 32 * FILTER_GRANT_CHUNK + tot_granted) {
 153                 CDEBUG(D_CACHE, "fs has no space left and statfs too old\n");
 154                 goto restat;
 155         }
 156
 157         if (left >= tot_granted) {
 158                 left -= tot_granted;
 159         } else {
 160                 if (left < tot_granted - obd->u.filter.fo_tot_pending) {
 161                         CERROR("%s: cli %s/%p grant "LPU64" > available "
 162                                LPU64" and pending "LPU64"\n", obd->obd_name,
 163                                exp->exp_client_uuid.uuid, exp, tot_granted,
 164                                left, obd->u.filter.fo_tot_pending);
 165                 }
 166                 left = 0;
 167         }
 168
 169         CDEBUG(D_CACHE, "%s: cli %s/%p free: "LPU64" avail: "LPU64" grant "LPU64
 170                " left: "LPU64" pending: "LPU64"\n", obd->obd_name,
 171                exp->exp_client_uuid.uuid, exp,
 172                obd->obd_osfs.os_bfree << blockbits, avail << blockbits,
 173                tot_granted, left, obd->u.filter.fo_tot_pending);
 174
 175         return left;
 176 }
 177
 178 /* Calculate how much grant space to allocate to this client, based on how
 179  * much space is currently free and how much of that is already granted.
 180  *
 181  * Caller must hold obd_osfs_lock. */
 182 long filter_grant(struct obd_export *exp, obd_size current_grant,
 183                   obd_size want, obd_size fs_space_left)
 184 {
 185         struct obd_device *obd = exp->exp_obd;
 186         struct filter_export_data *fed = &exp->exp_filter_data;
 187         int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
 188         __u64 grant = 0;
 189
 190         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
 191
 192         /* Grant some fraction of the client's requested grant space so that
 193          * they are not always waiting for write credits (not all of it to
 194          * avoid overgranting in face of multiple RPCs in flight).  This
 195          * essentially will be able to control the OSC_MAX_RIF for a client.
 196          *
 197          * If we do have a large disparity between what the client thinks it
 198          * has and what we think it has, don't grant very much and let the
 199          * client consume its grant first.  Either it just has lots of RPCs
 200          * in flight, or it was evicted and its grants will soon be used up. */
 201         if (want > 0x7fffffff) {
 202                 CERROR("%s: client %s/%p requesting > 2GB grant "LPU64"\n",
 203                        obd->obd_name, exp->exp_client_uuid.uuid, exp, want);
 204         } else if (current_grant < want &&
 205                    current_grant < fed->fed_grant + FILTER_GRANT_CHUNK) {
 206                 grant = min((want >> blockbits),
 207                             (fs_space_left >> blockbits) / 8);
 208                 grant <<= blockbits;
 209
 210                 if (grant) {
 211                         /* Allow >FILTER_GRANT_CHUNK size when clients
 212                          * reconnect due to a server reboot.
 213                          */
 214                         if ((grant > FILTER_GRANT_CHUNK) &&
 215                             (!obd->obd_recovering))
 216                                 grant = FILTER_GRANT_CHUNK;
 217
 218                         obd->u.filter.fo_tot_granted += grant;
 219                         fed->fed_grant += grant;
 220                         if (fed->fed_grant < 0) {
 221                                 CERROR("%s: cli %s/%p grant %ld want "LPU64
 222                                        "current"LPU64"\n",
 223                                        obd->obd_name, exp->exp_client_uuid.uuid,
 224                                        exp, fed->fed_grant, want,current_grant);
 225                                 spin_unlock(&obd->obd_osfs_lock);
 226                                 LBUG();
 227                         }
 228                 }
 229         }
 230
 231         CDEBUG(D_CACHE,
 232                "%s: cli %s/%p wants: "LPU64" current grant "LPU64
 233                " granting: "LPU64"\n", obd->obd_name, exp->exp_client_uuid.uuid,
 234                exp, want, current_grant, grant);
 235         CDEBUG(D_CACHE,
 236                "%s: cli %s/%p tot cached:"LPU64" granted:"LPU64
 237                " num_exports: %d\n", obd->obd_name, exp->exp_client_uuid.uuid,
 238                exp, obd->u.filter.fo_tot_dirty,
 239                obd->u.filter.fo_tot_granted, obd->obd_num_exports);
 240
 241         return grant;
 242 }
 243
 244 /*
 245  * the routine is used to request pages from pagecache
 246  *
 247  * use GFP_NOFS not allowing to enter FS as the client can run on this node
 248  * and we might end waiting on a page he sent in the request we're serving.
 249  *
 250  * use NORETRY so that the allocator doesn't go crazy: chance to more lucky
 251  * thread have enough memory to complete his request. for our request client
 252  * will do resend hopefully -bzzz
 253  */
 254 static struct page * filter_get_page(struct obd_device *obd,
 255                                      struct inode *inode,
 256                                      obd_off offset)
 257 {
 258         struct page *page;
 259
 260         page = find_or_create_page(inode->i_mapping, offset >> CFS_PAGE_SHIFT,
 261                                    GFP_NOFS | __GFP_NORETRY);
 262         if (unlikely(page == NULL))
 263                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_NO_PAGE, 1);
 264
 265         return page;
 266 }
 267
 268 /*
 269  * the routine initializes array of local_niobuf from remote_niobuf
 270  */
 271 static int filter_map_remote_to_local(int objcount, struct obd_ioobj *obj,
 272                                       struct niobuf_remote *nb,
 273                                       int *nrpages, struct niobuf_local *res)
 274 {
 275         struct niobuf_remote *rnb;
 276         struct niobuf_local *lnb;
 277         int i, max;
 278         ENTRY;
 279
 280         /* we don't support multiobject RPC yet
 281          * ost_brw_read() and ost_brw_write() check this */
 282         LASSERT(objcount == 1);
 283
 284         max = *nrpages;
 285         *nrpages = 0;
 286         for (i = 0, rnb = nb, lnb = res; i < obj->ioo_bufcnt; i++, rnb++) {
 287                 obd_off offset = rnb->offset;
 288                 unsigned int len = rnb->len;
 289
 290                 while (len > 0) {
 291                         int poff = offset & (CFS_PAGE_SIZE - 1);
 292                         int plen = CFS_PAGE_SIZE - poff;
 293
 294                         if (*nrpages >= max) {
 295                                 CERROR("small array of local bufs: %d\n", max);
 296                                 RETURN(-EINVAL);
 297                         }
 298
 299                         if (plen > len)
 300                                 plen = len;
 301                         lnb->offset = offset;
 302                         lnb->len = plen;
 303                         lnb->flags = rnb->flags;
 304                         lnb->page = NULL;
 305                         lnb->rc = 0;
 306                         lnb->lnb_grant_used = 0;
 307
 308                         LASSERTF(plen <= len, "plen %u, len %u\n", plen, len);
 309                         offset += plen;
 310                         len -= plen;
 311                         lnb++;
 312                         (*nrpages)++;
 313                 }
 314         }
 315         RETURN(0);
 316 }
 317
 318 /*
 319  * the function is used to free all pages used for request
 320  * just to mimic cacheless OSS which don't occupy much memory
 321  */
 322 void filter_invalidate_cache(struct obd_device *obd, struct obd_ioobj *obj,
 323                              struct niobuf_remote *nb, struct inode *inode)
 324 {
 325         struct niobuf_remote *rnb;
 326         int i;
 327
 328         LASSERT(inode != NULL);
 329
 330         for (i = 0, rnb = nb; i < obj->ioo_bufcnt; i++, rnb++) {
 331                 obd_off start = rnb->offset >> CFS_PAGE_SHIFT;
 332                 obd_off end = (rnb->offset + rnb->len) >> CFS_PAGE_SHIFT;
 333                 invalidate_mapping_pages(inode->i_mapping, start, end);
 334                 /* just to avoid warnings */
 335                 start = 0;
 336                 end = 0;
 337         }
 338 }
 339
 340 static int filter_preprw_read(int cmd, struct obd_export *exp, struct obdo *oa,
 341                               int objcount, struct obd_ioobj *obj,
 342                               struct niobuf_remote *nb,
 343                               int *npages, struct niobuf_local *res,
 344                               struct obd_trans_info *oti,
 345                               struct lustre_capa *capa)
 346 {
 347         struct obd_device *obd = exp->exp_obd;
 348         struct filter_obd *fo = &obd->u.filter;
 349         struct timeval start, end;
 350         struct lvfs_run_ctxt saved;
 351         struct niobuf_local *lnb;
 352         struct dentry *dentry = NULL;
 353         struct inode *inode = NULL;
 354         void *iobuf = NULL;
 355         int rc = 0, i, tot_bytes = 0;
 356         unsigned long now = jiffies;
 357         long timediff;
 358         ENTRY;
 359
 360         /* We are currently not supporting multi-obj BRW_READ RPCS at all.
 361          * When we do this function's dentry cleanup will need to be fixed.
 362          * These values are verified in ost_brw_write() from the wire. */
 363         LASSERTF(objcount == 1, "%d\n", objcount);
 364         LASSERTF(obj->ioo_bufcnt > 0, "%d\n", obj->ioo_bufcnt);
 365
 366         rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa), capa,
 367                               CAPA_OPC_OSS_READ);
 368         if (rc)
 369                 RETURN(rc);
 370
 371         if (oa && oa->o_valid & OBD_MD_FLGRANT) {
 372                 spin_lock(&obd->obd_osfs_lock);
 373                 filter_grant_incoming(exp, oa);
 374
 375                 oa->o_grant = 0;
 376                 spin_unlock(&obd->obd_osfs_lock);
 377         }
 378
 379         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 380         if (IS_ERR(iobuf))
 381                 RETURN(PTR_ERR(iobuf));
 382
 383         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 384         dentry = filter_oa2dentry(obd, oa);
 385         if (IS_ERR(dentry)) {
 386                 rc = PTR_ERR(dentry);
 387                 dentry = NULL;
 388                 GOTO(cleanup, rc);
 389         }
 390
 391         inode = dentry->d_inode;
 392
 393         obdo_to_inode(inode, oa, OBD_MD_FLATIME);
 394
 395         rc = filter_map_remote_to_local(objcount, obj, nb, npages, res);
 396         if (rc)
 397                 GOTO(cleanup, rc);
 398
 399         fsfilt_check_slow(obd, now, "preprw_read setup");
 400
 401         /* find pages for all segments, fill array with them */
 402         do_gettimeofday(&start);
 403         for (i = 0, lnb = res; i < *npages; i++, lnb++) {
 404
 405                 lnb->dentry = dentry;
 406
 407                 if (i_size_read(inode) <= lnb->offset)
 408                         /* If there's no more data, abort early.  lnb->rc == 0,
 409                          * so it's easy to detect later. */
 410                         break;
 411
 412                 lnb->page = filter_get_page(obd, inode, lnb->offset);
 413                 if (lnb->page == NULL)
 414                         GOTO(cleanup, rc = -ENOMEM);
 415
 416                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS, 1);
 417
 418                 if (i_size_read(inode) < lnb->offset + lnb->len - 1)
 419                         lnb->rc = i_size_read(inode) - lnb->offset;
 420                 else
 421                         lnb->rc = lnb->len;
 422
 423                 tot_bytes += lnb->rc;
 424
 425                 if (PageUptodate(lnb->page)) {
 426                         lprocfs_counter_add(obd->obd_stats,
 427                                             LPROC_FILTER_CACHE_HIT, 1);
 428                         continue;
 429                 }
 430
 431                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_MISS, 1);
 432                 filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 433         }
 434         do_gettimeofday(&end);
 435         timediff = cfs_timeval_sub(&end, &start, NULL);
 436         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
 437
 438         if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
 439                 GOTO(cleanup, rc = -ENOMEM);
 440
 441         fsfilt_check_slow(obd, now, "start_page_read");
 442
 443         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf,
 444                               exp, NULL, NULL, NULL);
 445         if (rc)
 446                 GOTO(cleanup, rc);
 447
 448         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_READ_BYTES, tot_bytes);
 449
 450         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
 451                 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
 452                                     LPROC_FILTER_READ_BYTES, tot_bytes);
 453
 454         EXIT;
 455
 456  cleanup:
 457         /* unlock pages to allow access from concurrent OST_READ */
 458         for (i = 0, lnb = res; i < *npages; i++, lnb++) {
 459                 if (lnb->page) {
 460                         LASSERT(PageLocked(lnb->page));
 461                         unlock_page(lnb->page);
 462
 463                         if (rc) {
 464                                 page_cache_release(lnb->page);
 465                                 lnb->page = NULL;
 466                         }
 467                 }
 468         }
 469
 470         if (inode && (fo->fo_read_cache == 0 ||
 471                         i_size_read(inode) > fo->fo_readcache_max_filesize))
 472                 filter_invalidate_cache(obd, obj, nb, inode);
 473
 474         if (rc != 0) {
 475                 if (dentry != NULL)
 476                         f_dput(dentry);
 477         }
 478
 479         filter_iobuf_put(&obd->u.filter, iobuf, oti);
 480
 481         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 482         if (rc)
 483                 CERROR("io error %d\n", rc);
 484
 485         return rc;
 486 }
 487
 488 /* When clients have dirtied as much space as they've been granted they
 489  * fall through to sync writes.  These sync writes haven't been expressed
 490  * in grants and need to error with ENOSPC when there isn't room in the
 491  * filesystem for them after grants are taken into account.  However,
 492  * writeback of the dirty data that was already granted space can write
 493  * right on through.
 494  *
 495  * Caller must hold obd_osfs_lock. */
 496 static int filter_grant_check(struct obd_export *exp, struct obdo *oa,
 497                               int objcount, struct fsfilt_objinfo *fso,
 498                               int niocount, struct niobuf_local *lnb,
 499                               obd_size *left, struct inode *inode)
 500 {
 501         struct filter_export_data *fed = &exp->exp_filter_data;
 502         int blocksize = exp->exp_obd->u.obt.obt_sb->s_blocksize;
 503         unsigned long used = 0, ungranted = 0, using;
 504         int i, rc = -ENOSPC, obj, n = 0;
 505
 506         LASSERT_SPIN_LOCKED(&exp->exp_obd->obd_osfs_lock);
 507
 508         for (obj = 0; obj < objcount; obj++) {
 509                 for (i = 0; i < fso[obj].fso_bufcnt; i++, n++) {
 510                         int tmp, bytes;
 511
 512                         /* should match the code in osc_exit_cache */
 513                         bytes = lnb[n].len;
 514                         bytes += lnb[n].offset & (blocksize - 1);
 515                         tmp = (lnb[n].offset + lnb[n].len) & (blocksize - 1);
 516                         if (tmp)
 517                                 bytes += blocksize - tmp;
 518
 519                         if ((lnb[n].flags & OBD_BRW_FROM_GRANT) &&
 520                             (oa->o_valid & OBD_MD_FLGRANT)) {
 521                                 if (fed->fed_grant < used + bytes) {
 522                                         CDEBUG(D_CACHE,
 523                                                "%s: cli %s/%p claims %ld+%d "
 524                                                "GRANT, real grant %lu idx %d\n",
 525                                                exp->exp_obd->obd_name,
 526                                                exp->exp_client_uuid.uuid, exp,
 527                                                used, bytes, fed->fed_grant, n);
 528                                 } else {
 529                                         used += bytes;
 530                                         lnb[n].flags |= OBD_BRW_GRANTED;
 531                                         lnb[n].lnb_grant_used = bytes;
 532                                         CDEBUG(0, "idx %d used=%lu\n", n, used);
 533                                         rc = 0;
 534                                         continue;
 535                                 }
 536                         }
 537                         if (*left > ungranted + bytes) {
 538                                 /* if enough space, pretend it was granted */
 539                                 ungranted += bytes;
 540                                 lnb[n].flags |= OBD_BRW_GRANTED;
 541                                 lnb[n].lnb_grant_used = bytes;
 542                                 CDEBUG(0, "idx %d ungranted=%lu\n",n,ungranted);
 543                                 rc = 0;
 544                                 continue;
 545                         }
 546
 547                         /* We can't check for already-mapped blocks here, as
 548                          * it requires dropping the osfs lock to do the bmap.
 549                          * Instead, we return ENOSPC and in that case we need
 550                          * to go through and verify if all of the blocks not
 551                          * marked BRW_GRANTED are already mapped and we can
 552                          * ignore this error. */
 553                         lnb[n].rc = -ENOSPC;
 554                         lnb[n].flags &= ~OBD_BRW_GRANTED;
 555                         CDEBUG(D_CACHE,"%s: cli %s/%p idx %d no space for %d\n",
 556                                exp->exp_obd->obd_name,
 557                                exp->exp_client_uuid.uuid, exp, n, bytes);
 558                 }
 559         }
 560
 561         /* Now substract what client have used already.  We don't subtract
 562          * this from the tot_granted yet, so that other client's can't grab
 563          * that space before we have actually allocated our blocks.  That
 564          * happens in filter_grant_commit() after the writes are done. */
 565         *left -= ungranted;
 566         fed->fed_grant -= used;
 567         fed->fed_pending += used + ungranted;
 568         exp->exp_obd->u.filter.fo_tot_granted += ungranted;
 569         exp->exp_obd->u.filter.fo_tot_pending += used + ungranted;
 570
 571         CDEBUG(D_CACHE,
 572                "%s: cli %s/%p used: %lu ungranted: %lu grant: %lu dirty: %lu\n",
 573                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp, used,
 574                ungranted, fed->fed_grant, fed->fed_dirty);
 575
 576         /* Rough calc in case we don't refresh cached statfs data */
 577         using = (used + ungranted + 1 ) >>
 578                 exp->exp_obd->u.obt.obt_sb->s_blocksize_bits;
 579         if (exp->exp_obd->obd_osfs.os_bavail > using)
 580                 exp->exp_obd->obd_osfs.os_bavail -= using;
 581         else
 582                 exp->exp_obd->obd_osfs.os_bavail = 0;
 583
 584         if (fed->fed_dirty < used) {
 585                 CERROR("%s: cli %s/%p claims used %lu > fed_dirty %lu\n",
 586                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 587                        used, fed->fed_dirty);
 588                 used = fed->fed_dirty;
 589         }
 590         exp->exp_obd->u.filter.fo_tot_dirty -= used;
 591         fed->fed_dirty -= used;
 592
 593         if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
 594                 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
 595                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 596                        fed->fed_dirty, fed->fed_pending, fed->fed_grant);
 597                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
 598                 LBUG();
 599         }
 600         return rc;
 601 }
 602
 603 /* If we ever start to support multi-object BRW RPCs, we will need to get locks
 604  * on mulitple inodes.  That isn't all, because there still exists the
 605  * possibility of a truncate starting a new transaction while holding the ext3
 606  * rwsem = write while some writes (which have started their transactions here)
 607  * blocking on the ext3 rwsem = read => lock inversion.
 608  *
 609  * The handling gets very ugly when dealing with locked pages.  It may be easier
 610  * to just get rid of the locked page code (which has problems of its own) and
 611  * either discover we do not need it anymore (i.e. it was a symptom of another
 612  * bug) or ensure we get the page locks in an appropriate order. */
 613 static int filter_preprw_write(int cmd, struct obd_export *exp, struct obdo *oa,
 614                                int objcount, struct obd_ioobj *obj,
 615                                struct niobuf_remote *nb, int *npages,
 616                                struct niobuf_local *res,
 617                                struct obd_trans_info *oti,
 618                                struct lustre_capa *capa)
 619 {
 620         struct obd_device *obd = exp->exp_obd;
 621         struct timeval start, end;
 622         struct lvfs_run_ctxt saved;
 623         struct niobuf_local *lnb = res;
 624         struct fsfilt_objinfo fso;
 625         struct filter_mod_data *fmd;
 626         struct dentry *dentry = NULL;
 627         void *iobuf;
 628         obd_size left;
 629         unsigned long now = jiffies, timediff;
 630         int rc = 0, i, tot_bytes = 0, cleanup_phase = 0;
 631         ENTRY;
 632         LASSERT(objcount == 1);
 633         LASSERT(obj->ioo_bufcnt > 0);
 634
 635         rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa), capa,
 636                               CAPA_OPC_OSS_WRITE);
 637         if (rc)
 638                 RETURN(rc);
 639
 640         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 641         iobuf = filter_iobuf_get(&exp->exp_obd->u.filter, oti);
 642         if (IS_ERR(iobuf))
 643                 GOTO(cleanup, rc = PTR_ERR(iobuf));
 644         cleanup_phase = 1;
 645
 646         dentry = filter_fid2dentry(exp->exp_obd, NULL, obj->ioo_gr,
 647                                    obj->ioo_id);
 648         if (IS_ERR(dentry))
 649                 GOTO(cleanup, rc = PTR_ERR(dentry));
 650         cleanup_phase = 2;
 651
 652         if (dentry->d_inode == NULL) {
 653                 CERROR("%s: trying to BRW to non-existent file "LPU64"\n",
 654                        exp->exp_obd->obd_name, obj->ioo_id);
 655                 GOTO(cleanup, rc = -ENOENT);
 656         }
 657
 658         rc = filter_map_remote_to_local(objcount, obj, nb, npages, res);
 659         if (rc)
 660                 GOTO(cleanup, rc);
 661
 662         fsfilt_check_slow(exp->exp_obd, now, "preprw_write setup");
 663
 664         /* Don't update inode timestamps if this write is older than a
 665          * setattr which modifies the timestamps. b=10150 */
 666         /* XXX when we start having persistent reservations this needs to
 667          * be changed to filter_fmd_get() to create the fmd if it doesn't
 668          * already exist so we can store the reservation handle there. */
 669         fmd = filter_fmd_find(exp, obj->ioo_id, obj->ioo_gr);
 670
 671         LASSERT(oa != NULL);
 672         spin_lock(&exp->exp_obd->obd_osfs_lock);
 673         filter_grant_incoming(exp, oa);
 674         if (fmd && fmd->fmd_mactime_xid > oti->oti_xid)
 675                 oa->o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 676                                  OBD_MD_FLATIME);
 677         else
 678                 obdo_to_inode(dentry->d_inode, oa, OBD_MD_FLATIME |
 679                               OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 680         cleanup_phase = 3;
 681
 682         left = filter_grant_space_left(exp);
 683
 684         fso.fso_dentry = dentry;
 685         fso.fso_bufcnt = *npages;
 686
 687         rc = filter_grant_check(exp, oa, objcount, &fso, *npages, res,
 688                                 &left, dentry->d_inode);
 689
 690         /* do not zero out oa->o_valid as it is used in filter_commitrw_write()
 691          * for setting UID/GID and fid EA in first write time. */
 692         if (oa->o_valid & OBD_MD_FLGRANT)
 693                 oa->o_grant = filter_grant(exp,oa->o_grant,oa->o_undirty,left);
 694
 695         spin_unlock(&exp->exp_obd->obd_osfs_lock);
 696         filter_fmd_put(exp, fmd);
 697
 698         if (rc)
 699                 GOTO(cleanup, rc);
 700
 701         do_gettimeofday(&start);
 702         for (i = 0, lnb = res; i < *npages; i++, lnb++) {
 703
 704                 /* We still set up for ungranted pages so that granted pages
 705                  * can be written to disk as they were promised, and portals
 706                  * needs to keep the pages all aligned properly. */
 707                 lnb->dentry = dentry;
 708
 709                 lnb->page = filter_get_page(obd, dentry->d_inode, lnb->offset);
 710                 if (lnb->page == NULL)
 711                         GOTO(cleanup, rc = -ENOMEM);
 712                 cleanup_phase = 4;
 713
 714                 /* DLM locking protects us from write and truncate competing
 715                  * for same region, but truncate can leave dirty page in the
 716                  * cache. it's possible the writeout on a such a page is in
 717                  * progress when we access it. it's also possible that during
 718                  * this writeout we put new (partial) data, but then won't
 719                  * be able to proceed in filter_commitrw_write(). thus let's
 720                  * just wait for writeout completion, should be rare enough.
 721                  * -bzzz */
 722                 wait_on_page_writeback(lnb->page);
 723
 724                 /* If the filter writes a partial page, then has the file
 725                  * extended, the client will read in the whole page.  the
 726                  * filter has to be careful to zero the rest of the partial
 727                  * page on disk.  we do it by hand for partial extending
 728                  * writes, send_bio() is responsible for zeroing pages when
 729                  * asked to read unmapped blocks -- brw_kiovec() does this. */
 730                 if (lnb->len != CFS_PAGE_SIZE) {
 731                         __s64 maxidx;
 732
 733                         maxidx = ((i_size_read(dentry->d_inode) +
 734                                    CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT) - 1;
 735                         if (maxidx >= lnb->page->index) {
 736                                 LL_CDEBUG_PAGE(D_PAGE, lnb->page, "write %u @ "
 737                                                LPU64" flg %x before EOF %llu\n",
 738                                                lnb->len, lnb->offset,lnb->flags,
 739                                                i_size_read(dentry->d_inode));
 740                                 filter_iobuf_add_page(exp->exp_obd, iobuf,
 741                                                       dentry->d_inode,
 742                                                       lnb->page);
 743                         } else {
 744                                 long off;
 745                                 char *p = kmap(lnb->page);
 746
 747                                 off = lnb->offset & ~CFS_PAGE_MASK;
 748                                 if (off)
 749                                         memset(p, 0, off);
 750                                 off = (lnb->offset + lnb->len) & ~CFS_PAGE_MASK;
 751                                 if (off)
 752                                         memset(p + off, 0, CFS_PAGE_SIZE - off);
 753                                 kunmap(lnb->page);
 754                         }
 755                 }
 756                 if (lnb->rc == 0)
 757                         tot_bytes += lnb->len;
 758         }
 759         do_gettimeofday(&end);
 760         timediff = cfs_timeval_sub(&end, &start, NULL);
 761         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
 762
 763         if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
 764                 GOTO(cleanup, rc = -ENOMEM);
 765
 766         /* don't unlock pages to prevent any access */
 767         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf, exp,
 768                               NULL, NULL, NULL);
 769
 770         fsfilt_check_slow(exp->exp_obd, now, "start_page_write");
 771
 772         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
 773                 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
 774                                     LPROC_FILTER_WRITE_BYTES, tot_bytes);
 775         EXIT;
 776 cleanup:
 777         switch(cleanup_phase) {
 778         case 4:
 779                 if (rc) {
 780                         for (i = 0, lnb = res; i < *npages; i++, lnb++) {
 781                                 if (lnb->page != NULL) {
 782                                         unlock_page(lnb->page);
 783                                         page_cache_release(lnb->page);
 784                                         lnb->page = NULL;
 785                                 }
 786                         }
 787                 }
 788         case 3:
 789                 filter_iobuf_put(&exp->exp_obd->u.filter, iobuf, oti);
 790         case 2:
 791                 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 792                 if (rc)
 793                         f_dput(dentry);
 794                 break;
 795         case 1:
 796                 filter_iobuf_put(&exp->exp_obd->u.filter, iobuf, oti);
 797         case 0:
 798                 spin_lock(&exp->exp_obd->obd_osfs_lock);
 799                 if (oa)
 800                         filter_grant_incoming(exp, oa);
 801                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
 802                 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 803                 break;
 804         default:;
 805         }
 806         return rc;
 807 }
 808
 809 int filter_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
 810                   int objcount, struct obd_ioobj *obj,
 811                   struct niobuf_remote *nb, int *npages,
 812                   struct niobuf_local *res, struct obd_trans_info *oti,
 813                   struct lustre_capa *capa)
 814 {
 815         if (cmd == OBD_BRW_WRITE)
 816                 return filter_preprw_write(cmd, exp, oa, objcount, obj,
 817                                            nb, npages, res, oti, capa);
 818         if (cmd == OBD_BRW_READ)
 819                 return filter_preprw_read(cmd, exp, oa, objcount, obj,
 820                                           nb, npages, res, oti, capa);
 821         LBUG();
 822         return -EPROTO;
 823 }
 824
 825 static int filter_commitrw_read(struct obd_export *exp, struct obdo *oa,
 826                                 int objcount, struct obd_ioobj *obj,
 827                                 struct niobuf_remote *rnb,
 828                                 int npages, struct niobuf_local *res,
 829                                 struct obd_trans_info *oti, int rc)
 830 {
 831         struct inode *inode = NULL;
 832         struct ldlm_res_id res_id;
 833         struct ldlm_resource *resource = NULL;
 834         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 835         struct niobuf_local *lnb;
 836         int i;
 837         ENTRY;
 838
 839         osc_build_res_name(obj->ioo_id, obj->ioo_gr, &res_id);
 840         /* If oa != NULL then filter_preprw_read updated the inode atime
 841          * and we should update the lvb so that other glimpses will also
 842          * get the updated value. bug 5972 */
 843         if (oa && ns && ns->ns_lvbo && ns->ns_lvbo->lvbo_update) {
 844                 resource = ldlm_resource_get(ns, NULL, &res_id, LDLM_EXTENT, 0);
 845
 846                 if (resource != NULL) {
 847                         ns->ns_lvbo->lvbo_update(resource, NULL, 0, 1);
 848                         ldlm_resource_putref(resource);
 849                 }
 850         }
 851
 852         if (res->dentry != NULL)
 853                 inode = res->dentry->d_inode;
 854
 855         for (i = 0, lnb = res; i < npages; i++, lnb++) {
 856                 if (lnb->page != NULL) {
 857                         page_cache_release(lnb->page);
 858                         lnb->page = NULL;
 859                 }
 860         }
 861
 862         if (res->dentry != NULL)
 863                 f_dput(res->dentry);
 864         RETURN(rc);
 865 }
 866
 867 void filter_grant_commit(struct obd_export *exp, int niocount,
 868                          struct niobuf_local *res)
 869 {
 870         struct filter_obd *filter = &exp->exp_obd->u.filter;
 871         struct niobuf_local *lnb = res;
 872         unsigned long pending = 0;
 873         int i;
 874
 875         spin_lock(&exp->exp_obd->obd_osfs_lock);
 876         for (i = 0, lnb = res; i < niocount; i++, lnb++)
 877                 pending += lnb->lnb_grant_used;
 878
 879         LASSERTF(exp->exp_filter_data.fed_pending >= pending,
 880                  "%s: cli %s/%p fed_pending: %lu grant_used: %lu\n",
 881                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 882                  exp->exp_filter_data.fed_pending, pending);
 883         exp->exp_filter_data.fed_pending -= pending;
 884         LASSERTF(filter->fo_tot_granted >= pending,
 885                  "%s: cli %s/%p tot_granted: "LPU64" grant_used: %lu\n",
 886                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 887                  exp->exp_obd->u.filter.fo_tot_granted, pending);
 888         filter->fo_tot_granted -= pending;
 889         LASSERTF(filter->fo_tot_pending >= pending,
 890                  "%s: cli %s/%p tot_pending: "LPU64" grant_used: %lu\n",
 891                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 892                  filter->fo_tot_pending, pending);
 893         filter->fo_tot_pending -= pending;
 894
 895         spin_unlock(&exp->exp_obd->obd_osfs_lock);
 896 }
 897
 898 int filter_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
 899                     int objcount, struct obd_ioobj *obj,
 900                     struct niobuf_remote *nb, int npages,
 901                     struct niobuf_local *res, struct obd_trans_info *oti,
 902                     int rc)
 903 {
 904         if (cmd == OBD_BRW_WRITE)
 905                 return filter_commitrw_write(exp, oa, objcount, obj,
 906                                              nb, npages, res, oti, rc);
 907         if (cmd == OBD_BRW_READ)
 908                 return filter_commitrw_read(exp, oa, objcount, obj,
 909                                             nb, npages, res, oti, rc);
 910         LBUG();
 911         return -EPROTO;
 912 }
 913
 914 int filter_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
 915                obd_count oa_bufs, struct brw_page *pga,
 916                struct obd_trans_info *oti)
 917 {
 918         struct obd_ioobj ioo;
 919         struct niobuf_local *lnb;
 920         struct niobuf_remote *rnb;
 921         obd_count i;
 922         int ret = 0, npages;
 923         ENTRY;
 924
 925         OBD_ALLOC(lnb, oa_bufs * sizeof(struct niobuf_local));
 926         OBD_ALLOC(rnb, oa_bufs * sizeof(struct niobuf_remote));
 927
 928         if (lnb == NULL || rnb == NULL)
 929                 GOTO(out, ret = -ENOMEM);
 930
 931         for (i = 0; i < oa_bufs; i++) {
 932                 lnb[i].page = pga[i].pg;
 933                 rnb[i].offset = pga[i].off;
 934                 rnb[i].len = pga[i].count;
 935         }
 936
 937         obdo_to_ioobj(oinfo->oi_oa, &ioo);
 938         ioo.ioo_bufcnt = oa_bufs;
 939
 940         npages = oa_bufs;
 941         ret = filter_preprw(cmd, exp, oinfo->oi_oa, 1, &ioo,
 942                             rnb, &npages, lnb, oti, oinfo_capa(oinfo));
 943         if (ret != 0)
 944                 GOTO(out, ret);
 945         LASSERTF(oa_bufs == npages, "%u != %u\n", oa_bufs, npages);
 946
 947         ret = filter_commitrw(cmd, exp, oinfo->oi_oa, 1, &ioo, rnb,
 948                               npages, lnb, oti, ret);
 949
 950 out:
 951         if (lnb)
 952                 OBD_FREE(lnb, oa_bufs * sizeof(struct niobuf_local));
 953         if (rnb)
 954                 OBD_FREE(rnb, oa_bufs * sizeof(struct niobuf_remote));
 955         RETURN(ret);
 956 }