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         }
 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                               struct lustre_capa *capa)
 344 {
 345         struct obd_device *obd = exp->exp_obd;
 346         struct filter_obd *fo = &obd->u.filter;
 347         struct timeval start, end;
 348         struct lvfs_run_ctxt saved;
 349         struct niobuf_local *lnb;
 350         struct dentry *dentry = NULL;
 351         struct inode *inode = NULL;
 352         void *iobuf = NULL;
 353         int rc = 0, i, tot_bytes = 0;
 354         unsigned long now = jiffies;
 355         long timediff;
 356         ENTRY;
 357
 358         /* We are currently not supporting multi-obj BRW_READ RPCS at all.
 359          * When we do this function's dentry cleanup will need to be fixed.
 360          * These values are verified in ost_brw_write() from the wire. */
 361         LASSERTF(objcount == 1, "%d\n", objcount);
 362         LASSERTF(obj->ioo_bufcnt > 0, "%d\n", obj->ioo_bufcnt);
 363
 364         rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa), capa,
 365                               CAPA_OPC_OSS_READ);
 366         if (rc)
 367                 RETURN(rc);
 368
 369         if (oa && oa->o_valid & OBD_MD_FLGRANT) {
 370                 spin_lock(&obd->obd_osfs_lock);
 371                 filter_grant_incoming(exp, oa);
 372
 373                 oa->o_grant = 0;
 374                 spin_unlock(&obd->obd_osfs_lock);
 375         }
 376
 377         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 378         if (IS_ERR(iobuf))
 379                 RETURN(PTR_ERR(iobuf));
 380
 381         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 382         dentry = filter_oa2dentry(obd, oa);
 383         if (IS_ERR(dentry)) {
 384                 rc = PTR_ERR(dentry);
 385                 dentry = NULL;
 386                 GOTO(cleanup, rc);
 387         }
 388
 389         inode = dentry->d_inode;
 390
 391         obdo_to_inode(inode, oa, OBD_MD_FLATIME);
 392
 393         rc = filter_map_remote_to_local(objcount, obj, nb, pages, res);
 394         if (rc)
 395                 GOTO(cleanup, rc);
 396
 397         fsfilt_check_slow(obd, now, "preprw_read setup");
 398
 399         /* find pages for all segments, fill array with them */
 400         do_gettimeofday(&start);
 401         for (i = 0, lnb = res; i < *pages; i++, lnb++) {
 402
 403                 lnb->dentry = dentry;
 404
 405                 if (i_size_read(inode) <= lnb->offset)
 406                         /* If there's no more data, abort early.  lnb->rc == 0,
 407                          * so it's easy to detect later. */
 408                         break;
 409
 410                 lnb->page = filter_get_page(obd, inode, lnb->offset);
 411                 if (lnb->page == NULL)
 412                         GOTO(cleanup, rc = -ENOMEM);
 413
 414                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS, 1);
 415
 416                 if (i_size_read(inode) < lnb->offset + lnb->len - 1)
 417                         lnb->rc = i_size_read(inode) - lnb->offset;
 418                 else
 419                         lnb->rc = lnb->len;
 420
 421                 tot_bytes += lnb->rc;
 422
 423                 if (PageUptodate(lnb->page)) {
 424                         lprocfs_counter_add(obd->obd_stats,
 425                                             LPROC_FILTER_CACHE_HIT, 1);
 426                         continue;
 427                 }
 428
 429                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_MISS, 1);
 430                 filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 431         }
 432         do_gettimeofday(&end);
 433         timediff = cfs_timeval_sub(&end, &start, NULL);
 434         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
 435
 436         if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
 437                 GOTO(cleanup, rc = -ENOMEM);
 438
 439         fsfilt_check_slow(obd, now, "start_page_read");
 440
 441         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf,
 442                               exp, NULL, NULL, NULL);
 443         if (rc)
 444                 GOTO(cleanup, rc);
 445
 446         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_READ_BYTES, tot_bytes);
 447
 448         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
 449                 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
 450                                     LPROC_FILTER_READ_BYTES, tot_bytes);
 451
 452         EXIT;
 453
 454  cleanup:
 455         /* unlock pages to allow access from concurrent OST_READ */
 456         for (i = 0, lnb = res; i < *pages; i++, lnb++) {
 457                 if (lnb->page) {
 458                         LASSERT(PageLocked(lnb->page));
 459                         unlock_page(lnb->page);
 460
 461                         if (rc) {
 462                                 page_cache_release(lnb->page);
 463                                 lnb->page = NULL;
 464                         }
 465                 }
 466         }
 467
 468         if (inode && (fo->fo_read_cache == 0 ||
 469                         i_size_read(inode) > fo->fo_readcache_max_filesize))
 470                 filter_invalidate_cache(obd, obj, nb, inode);
 471
 472         if (rc != 0) {
 473                 if (dentry != NULL)
 474                         f_dput(dentry);
 475         }
 476
 477         filter_iobuf_put(&obd->u.filter, iobuf, oti);
 478
 479         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 480         if (rc)
 481                 CERROR("io error %d\n", rc);
 482
 483         return rc;
 484 }
 485
 486 /* When clients have dirtied as much space as they've been granted they
 487  * fall through to sync writes.  These sync writes haven't been expressed
 488  * in grants and need to error with ENOSPC when there isn't room in the
 489  * filesystem for them after grants are taken into account.  However,
 490  * writeback of the dirty data that was already granted space can write
 491  * right on through.
 492  *
 493  * Caller must hold obd_osfs_lock. */
 494 static int filter_grant_check(struct obd_export *exp, struct obdo *oa,
 495                               int objcount, struct fsfilt_objinfo *fso,
 496                               int niocount, struct niobuf_local *lnb,
 497                               obd_size *left, struct inode *inode)
 498 {
 499         struct filter_export_data *fed = &exp->exp_filter_data;
 500         int blocksize = exp->exp_obd->u.obt.obt_sb->s_blocksize;
 501         unsigned long used = 0, ungranted = 0, using;
 502         int i, rc = -ENOSPC, obj, n = 0;
 503
 504         LASSERT_SPIN_LOCKED(&exp->exp_obd->obd_osfs_lock);
 505
 506         for (obj = 0; obj < objcount; obj++) {
 507                 for (i = 0; i < fso[obj].fso_bufcnt; i++, n++) {
 508                         int tmp, bytes;
 509
 510                         /* should match the code in osc_exit_cache */
 511                         bytes = lnb[n].len;
 512                         bytes += lnb[n].offset & (blocksize - 1);
 513                         tmp = (lnb[n].offset + lnb[n].len) & (blocksize - 1);
 514                         if (tmp)
 515                                 bytes += blocksize - tmp;
 516
 517                         if ((lnb[n].flags & OBD_BRW_FROM_GRANT) &&
 518                             (oa->o_valid & OBD_MD_FLGRANT)) {
 519                                 if (fed->fed_grant < used + bytes) {
 520                                         CDEBUG(D_CACHE,
 521                                                "%s: cli %s/%p claims %ld+%d "
 522                                                "GRANT, real grant %lu idx %d\n",
 523                                                exp->exp_obd->obd_name,
 524                                                exp->exp_client_uuid.uuid, exp,
 525                                                used, bytes, fed->fed_grant, n);
 526                                 } else {
 527                                         used += bytes;
 528                                         lnb[n].flags |= OBD_BRW_GRANTED;
 529                                         lnb[n].lnb_grant_used = bytes;
 530                                         CDEBUG(0, "idx %d used=%lu\n", n, used);
 531                                         rc = 0;
 532                                         continue;
 533                                 }
 534                         }
 535                         if (*left > ungranted + bytes) {
 536                                 /* if enough space, pretend it was granted */
 537                                 ungranted += bytes;
 538                                 lnb[n].flags |= OBD_BRW_GRANTED;
 539                                 lnb[n].lnb_grant_used = bytes;
 540                                 CDEBUG(0, "idx %d ungranted=%lu\n",n,ungranted);
 541                                 rc = 0;
 542                                 continue;
 543                         }
 544
 545                         /* We can't check for already-mapped blocks here, as
 546                          * it requires dropping the osfs lock to do the bmap.
 547                          * Instead, we return ENOSPC and in that case we need
 548                          * to go through and verify if all of the blocks not
 549                          * marked BRW_GRANTED are already mapped and we can
 550                          * ignore this error. */
 551                         lnb[n].rc = -ENOSPC;
 552                         lnb[n].flags &= ~OBD_BRW_GRANTED;
 553                         CDEBUG(D_CACHE,"%s: cli %s/%p idx %d no space for %d\n",
 554                                exp->exp_obd->obd_name,
 555                                exp->exp_client_uuid.uuid, exp, n, bytes);
 556                 }
 557         }
 558
 559         /* Now substract what client have used already.  We don't subtract
 560          * this from the tot_granted yet, so that other client's can't grab
 561          * that space before we have actually allocated our blocks.  That
 562          * happens in filter_grant_commit() after the writes are done. */
 563         *left -= ungranted;
 564         fed->fed_grant -= used;
 565         fed->fed_pending += used + ungranted;
 566         exp->exp_obd->u.filter.fo_tot_granted += ungranted;
 567         exp->exp_obd->u.filter.fo_tot_pending += used + ungranted;
 568
 569         CDEBUG(D_CACHE,
 570                "%s: cli %s/%p used: %lu ungranted: %lu grant: %lu dirty: %lu\n",
 571                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp, used,
 572                ungranted, fed->fed_grant, fed->fed_dirty);
 573
 574         /* Rough calc in case we don't refresh cached statfs data */
 575         using = (used + ungranted + 1 ) >>
 576                 exp->exp_obd->u.obt.obt_sb->s_blocksize_bits;
 577         if (exp->exp_obd->obd_osfs.os_bavail > using)
 578                 exp->exp_obd->obd_osfs.os_bavail -= using;
 579         else
 580                 exp->exp_obd->obd_osfs.os_bavail = 0;
 581
 582         if (fed->fed_dirty < used) {
 583                 CERROR("%s: cli %s/%p claims used %lu > fed_dirty %lu\n",
 584                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 585                        used, fed->fed_dirty);
 586                 used = fed->fed_dirty;
 587         }
 588         exp->exp_obd->u.filter.fo_tot_dirty -= used;
 589         fed->fed_dirty -= used;
 590
 591         if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
 592                 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
 593                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 594                        fed->fed_dirty, fed->fed_pending, fed->fed_grant);
 595                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
 596                 LBUG();
 597         }
 598         return rc;
 599 }
 600
 601 /* If we ever start to support multi-object BRW RPCs, we will need to get locks
 602  * on mulitple inodes.  That isn't all, because there still exists the
 603  * possibility of a truncate starting a new transaction while holding the ext3
 604  * rwsem = write while some writes (which have started their transactions here)
 605  * blocking on the ext3 rwsem = read => lock inversion.
 606  *
 607  * The handling gets very ugly when dealing with locked pages.  It may be easier
 608  * to just get rid of the locked page code (which has problems of its own) and
 609  * either discover we do not need it anymore (i.e. it was a symptom of another
 610  * bug) or ensure we get the page locks in an appropriate order. */
 611 static int filter_preprw_write(int cmd, struct obd_export *exp, struct obdo *oa,
 612                                int objcount, struct obd_ioobj *obj,
 613                                struct niobuf_remote *nb, int *pages,
 614                                struct niobuf_local *res,
 615                                struct obd_trans_info *oti,
 616                                struct lustre_capa *capa)
 617 {
 618         struct obd_device *obd = exp->exp_obd;
 619         struct timeval start, end;
 620         struct lvfs_run_ctxt saved;
 621         struct niobuf_local *lnb = res;
 622         struct fsfilt_objinfo fso;
 623         struct filter_mod_data *fmd;
 624         struct dentry *dentry = NULL;
 625         void *iobuf;
 626         obd_size left;
 627         unsigned long now = jiffies, timediff;
 628         int rc = 0, i, tot_bytes = 0, cleanup_phase = 0;
 629         ENTRY;
 630         LASSERT(objcount == 1);
 631         LASSERT(obj->ioo_bufcnt > 0);
 632
 633         rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa), capa,
 634                               CAPA_OPC_OSS_WRITE);
 635         if (rc)
 636                 RETURN(rc);
 637
 638         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 639         iobuf = filter_iobuf_get(&exp->exp_obd->u.filter, oti);
 640         if (IS_ERR(iobuf))
 641                 GOTO(cleanup, rc = PTR_ERR(iobuf));
 642         cleanup_phase = 1;
 643
 644         dentry = filter_fid2dentry(exp->exp_obd, NULL, obj->ioo_gr,
 645                                    obj->ioo_id);
 646         if (IS_ERR(dentry))
 647                 GOTO(cleanup, rc = PTR_ERR(dentry));
 648         cleanup_phase = 2;
 649
 650         if (dentry->d_inode == NULL) {
 651                 CERROR("%s: trying to BRW to non-existent file "LPU64"\n",
 652                        exp->exp_obd->obd_name, obj->ioo_id);
 653                 GOTO(cleanup, rc = -ENOENT);
 654         }
 655
 656         rc = filter_map_remote_to_local(objcount, obj, nb, pages, res);
 657         if (rc)
 658                 GOTO(cleanup, rc);
 659
 660         fsfilt_check_slow(exp->exp_obd, now, "preprw_write setup");
 661
 662         /* Don't update inode timestamps if this write is older than a
 663          * setattr which modifies the timestamps. b=10150 */
 664         /* XXX when we start having persistent reservations this needs to
 665          * be changed to filter_fmd_get() to create the fmd if it doesn't
 666          * already exist so we can store the reservation handle there. */
 667         fmd = filter_fmd_find(exp, obj->ioo_id, obj->ioo_gr);
 668
 669         LASSERT(oa != NULL);
 670         spin_lock(&exp->exp_obd->obd_osfs_lock);
 671         filter_grant_incoming(exp, oa);
 672         if (fmd && fmd->fmd_mactime_xid > oti->oti_xid)
 673                 oa->o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 674                                  OBD_MD_FLATIME);
 675         else
 676                 obdo_to_inode(dentry->d_inode, oa, OBD_MD_FLATIME |
 677                               OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 678         cleanup_phase = 3;
 679
 680         left = filter_grant_space_left(exp);
 681
 682         fso.fso_dentry = dentry;
 683         fso.fso_bufcnt = *pages;
 684
 685         rc = filter_grant_check(exp, oa, objcount, &fso, *pages, res,
 686                                 &left, dentry->d_inode);
 687
 688         /* do not zero out oa->o_valid as it is used in filter_commitrw_write()
 689          * for setting UID/GID and fid EA in first write time. */
 690         if (oa->o_valid & OBD_MD_FLGRANT)
 691                 oa->o_grant = filter_grant(exp,oa->o_grant,oa->o_undirty,left);
 692
 693         spin_unlock(&exp->exp_obd->obd_osfs_lock);
 694         filter_fmd_put(exp, fmd);
 695
 696         if (rc)
 697                 GOTO(cleanup, rc);
 698
 699         do_gettimeofday(&start);
 700         for (i = 0, lnb = res; i < *pages; i++, lnb++) {
 701
 702                 /* We still set up for ungranted pages so that granted pages
 703                  * can be written to disk as they were promised, and portals
 704                  * needs to keep the pages all aligned properly. */
 705                 lnb->dentry = dentry;
 706
 707                 lnb->page = filter_get_page(obd, dentry->d_inode, lnb->offset);
 708                 if (lnb->page == NULL)
 709                         GOTO(cleanup, rc = -ENOMEM);
 710                 cleanup_phase = 4;
 711
 712                 /* DLM locking protects us from write and truncate competing
 713                  * for same region, but truncate can leave dirty page in the
 714                  * cache. it's possible the writeout on a such a page is in
 715                  * progress when we access it. it's also possible that during
 716                  * this writeout we put new (partial) data, but then won't
 717                  * be able to proceed in filter_commitrw_write(). thus let's
 718                  * just wait for writeout completion, should be rare enough.
 719                  * -bzzz */
 720                 wait_on_page_writeback(lnb->page);
 721
 722                 /* If the filter writes a partial page, then has the file
 723                  * extended, the client will read in the whole page.  the
 724                  * filter has to be careful to zero the rest of the partial
 725                  * page on disk.  we do it by hand for partial extending
 726                  * writes, send_bio() is responsible for zeroing pages when
 727                  * asked to read unmapped blocks -- brw_kiovec() does this. */
 728                 if (lnb->len != CFS_PAGE_SIZE) {
 729                         __s64 maxidx;
 730
 731                         maxidx = ((i_size_read(dentry->d_inode) +
 732                                    CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT) - 1;
 733                         if (maxidx >= lnb->page->index) {
 734                                 LL_CDEBUG_PAGE(D_PAGE, lnb->page, "write %u @ "
 735                                                LPU64" flg %x before EOF %llu\n",
 736                                                lnb->len, lnb->offset,lnb->flags,
 737                                                i_size_read(dentry->d_inode));
 738                                 filter_iobuf_add_page(exp->exp_obd, iobuf,
 739                                                       dentry->d_inode,
 740                                                       lnb->page);
 741                         } else {
 742                                 long off;
 743                                 char *p = kmap(lnb->page);
 744
 745                                 off = lnb->offset & ~CFS_PAGE_MASK;
 746                                 if (off)
 747                                         memset(p, 0, off);
 748                                 off = (lnb->offset + lnb->len) & ~CFS_PAGE_MASK;
 749                                 if (off)
 750                                         memset(p + off, 0, CFS_PAGE_SIZE - off);
 751                                 kunmap(lnb->page);
 752                         }
 753                 }
 754                 if (lnb->rc == 0)
 755                         tot_bytes += lnb->len;
 756         }
 757         do_gettimeofday(&end);
 758         timediff = cfs_timeval_sub(&end, &start, NULL);
 759         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
 760
 761         if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
 762                 GOTO(cleanup, rc = -ENOMEM);
 763
 764         /* don't unlock pages to prevent any access */
 765         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf, exp,
 766                               NULL, NULL, NULL);
 767
 768         fsfilt_check_slow(exp->exp_obd, now, "start_page_write");
 769
 770         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
 771                 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
 772                                     LPROC_FILTER_WRITE_BYTES, tot_bytes);
 773         EXIT;
 774 cleanup:
 775         switch(cleanup_phase) {
 776         case 4:
 777                 if (rc) {
 778                         for (i = 0, lnb = res; i < *pages; i++, lnb++) {
 779                                 if (lnb->page != NULL) {
 780                                         unlock_page(lnb->page);
 781                                         page_cache_release(lnb->page);
 782                                         lnb->page = NULL;
 783                                 }
 784                         }
 785                 }
 786         case 3:
 787                 filter_iobuf_put(&exp->exp_obd->u.filter, iobuf, oti);
 788         case 2:
 789                 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 790                 if (rc)
 791                         f_dput(dentry);
 792                 break;
 793         case 1:
 794                 filter_iobuf_put(&exp->exp_obd->u.filter, iobuf, oti);
 795         case 0:
 796                 spin_lock(&exp->exp_obd->obd_osfs_lock);
 797                 if (oa)
 798                         filter_grant_incoming(exp, oa);
 799                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
 800                 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 801                 break;
 802         default:;
 803         }
 804         return rc;
 805 }
 806
 807 int filter_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
 808                   int objcount, struct obd_ioobj *obj,
 809                   struct niobuf_remote *nb, int *pages,
 810                   struct niobuf_local *res, struct obd_trans_info *oti,
 811                   struct lustre_capa *capa)
 812 {
 813         if (cmd == OBD_BRW_WRITE)
 814                 return filter_preprw_write(cmd, exp, oa, objcount, obj,
 815                                            nb, pages, res, oti, capa);
 816         if (cmd == OBD_BRW_READ)
 817                 return filter_preprw_read(cmd, exp, oa, objcount, obj,
 818                                           nb, pages, res, oti, capa);
 819         LBUG();
 820         return -EPROTO;
 821 }
 822
 823 static int filter_commitrw_read(struct obd_export *exp, struct obdo *oa,
 824                                 int objcount, struct obd_ioobj *obj,
 825                                 struct niobuf_remote *rnb,
 826                                 int pages, struct niobuf_local *res,
 827                                 struct obd_trans_info *oti, int rc)
 828 {
 829         struct inode *inode = NULL;
 830         struct ldlm_res_id res_id;
 831         struct ldlm_resource *resource = NULL;
 832         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 833         struct niobuf_local *lnb;
 834         int i;
 835         ENTRY;
 836
 837         osc_build_res_name(obj->ioo_id, obj->ioo_gr, &res_id);
 838         /* If oa != NULL then filter_preprw_read updated the inode atime
 839          * and we should update the lvb so that other glimpses will also
 840          * get the updated value. bug 5972 */
 841         if (oa && ns && ns->ns_lvbo && ns->ns_lvbo->lvbo_update) {
 842                 resource = ldlm_resource_get(ns, NULL, &res_id, LDLM_EXTENT, 0);
 843
 844                 if (resource != NULL) {
 845                         ns->ns_lvbo->lvbo_update(resource, NULL, 0, 1);
 846                         ldlm_resource_putref(resource);
 847                 }
 848         }
 849
 850         if (res->dentry != NULL)
 851                 inode = res->dentry->d_inode;
 852
 853         for (i = 0, lnb = res; i < pages; i++, lnb++) {
 854                 if (lnb->page != NULL) {
 855                         page_cache_release(lnb->page);
 856                         lnb->page = NULL;
 857                 }
 858         }
 859
 860         if (res->dentry != NULL)
 861                 f_dput(res->dentry);
 862         RETURN(rc);
 863 }
 864
 865 void filter_grant_commit(struct obd_export *exp, int niocount,
 866                          struct niobuf_local *res)
 867 {
 868         struct filter_obd *filter = &exp->exp_obd->u.filter;
 869         struct niobuf_local *lnb = res;
 870         unsigned long pending = 0;
 871         int i;
 872
 873         spin_lock(&exp->exp_obd->obd_osfs_lock);
 874         for (i = 0, lnb = res; i < niocount; i++, lnb++)
 875                 pending += lnb->lnb_grant_used;
 876
 877         LASSERTF(exp->exp_filter_data.fed_pending >= pending,
 878                  "%s: cli %s/%p fed_pending: %lu grant_used: %lu\n",
 879                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 880                  exp->exp_filter_data.fed_pending, pending);
 881         exp->exp_filter_data.fed_pending -= pending;
 882         LASSERTF(filter->fo_tot_granted >= pending,
 883                  "%s: cli %s/%p tot_granted: "LPU64" grant_used: %lu\n",
 884                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 885                  exp->exp_obd->u.filter.fo_tot_granted, pending);
 886         filter->fo_tot_granted -= pending;
 887         LASSERTF(filter->fo_tot_pending >= pending,
 888                  "%s: cli %s/%p tot_pending: "LPU64" grant_used: %lu\n",
 889                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 890                  filter->fo_tot_pending, pending);
 891         filter->fo_tot_pending -= pending;
 892
 893         spin_unlock(&exp->exp_obd->obd_osfs_lock);
 894 }
 895
 896 int filter_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
 897                     int objcount, struct obd_ioobj *obj,
 898                     struct niobuf_remote *nb, int pages,
 899                     struct niobuf_local *res, struct obd_trans_info *oti,
 900                     int rc)
 901 {
 902         if (cmd == OBD_BRW_WRITE)
 903                 return filter_commitrw_write(exp, oa, objcount, obj,
 904                                              nb, pages, res, oti, rc);
 905         if (cmd == OBD_BRW_READ)
 906                 return filter_commitrw_read(exp, oa, objcount, obj,
 907                                             nb, pages, res, oti, rc);
 908         LBUG();
 909         return -EPROTO;
 910 }
 911
 912 int filter_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
 913                obd_count oa_bufs, struct brw_page *pga,
 914                struct obd_trans_info *oti)
 915 {
 916         struct obd_ioobj ioo;
 917         struct niobuf_local *lnb;
 918         struct niobuf_remote *rnb;
 919         obd_count i;
 920         int ret = 0, npages;
 921         ENTRY;
 922
 923         OBD_ALLOC(lnb, oa_bufs * sizeof(struct niobuf_local));
 924         OBD_ALLOC(rnb, oa_bufs * sizeof(struct niobuf_remote));
 925
 926         if (lnb == NULL || rnb == NULL)
 927                 GOTO(out, ret = -ENOMEM);
 928
 929         for (i = 0; i < oa_bufs; i++) {
 930                 lnb[i].page = pga[i].pg;
 931                 rnb[i].offset = pga[i].off;
 932                 rnb[i].len = pga[i].count;
 933         }
 934
 935         obdo_to_ioobj(oinfo->oi_oa, &ioo);
 936         ioo.ioo_bufcnt = oa_bufs;
 937
 938         npages = oa_bufs;
 939         ret = filter_preprw(cmd, exp, oinfo->oi_oa, 1, &ioo,
 940                             rnb, &npages, lnb, oti, oinfo_capa(oinfo));
 941         if (ret != 0)
 942                 GOTO(out, ret);
 943         LASSERTF(oa_bufs == npages, "%u != %u\n", oa_bufs, npages);
 944
 945         ret = filter_commitrw(cmd, exp, oinfo->oi_oa, 1, &ioo, rnb,
 946                               npages, lnb, oti, ret);
 947
 948 out:
 949         if (lnb)
 950                 OBD_FREE(lnb, oa_bufs * sizeof(struct niobuf_local));
 951         if (rnb)
 952                 OBD_FREE(rnb, oa_bufs * sizeof(struct niobuf_remote));
 953         RETURN(ret);
 954 }