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
 341 static int filter_preprw_read(int cmd, struct obd_export *exp, struct obdo *oa,
 342                               int objcount, struct obd_ioobj *obj,
 343                               struct niobuf_remote *nb,
 344                               int *pages, struct niobuf_local *res,
 345                               struct obd_trans_info *oti,
 346                               struct lustre_capa *capa)
 347 {
 348         struct obd_device *obd = exp->exp_obd;
 349         struct filter_obd *fo = &obd->u.filter;
 350         struct timeval start, end;
 351         struct lvfs_run_ctxt saved;
 352         struct niobuf_local *lnb;
 353         struct dentry *dentry = NULL;
 354         struct inode *inode = NULL;
 355         void *iobuf = NULL;
 356         int rc = 0, i, tot_bytes = 0;
 357         unsigned long now = jiffies;
 358         long timediff;
 359         ENTRY;
 360
 361         /* We are currently not supporting multi-obj BRW_READ RPCS at all.
 362          * When we do this function's dentry cleanup will need to be fixed.
 363          * These values are verified in ost_brw_write() from the wire. */
 364         LASSERTF(objcount == 1, "%d\n", objcount);
 365         LASSERTF(obj->ioo_bufcnt > 0, "%d\n", obj->ioo_bufcnt);
 366
 367         rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa), capa,
 368                               CAPA_OPC_OSS_READ);
 369         if (rc)
 370                 RETURN(rc);
 371
 372         if (oa && oa->o_valid & OBD_MD_FLGRANT) {
 373                 spin_lock(&obd->obd_osfs_lock);
 374                 filter_grant_incoming(exp, oa);
 375
 376                 oa->o_grant = 0;
 377                 spin_unlock(&obd->obd_osfs_lock);
 378         }
 379
 380         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 381         if (IS_ERR(iobuf))
 382                 RETURN(PTR_ERR(iobuf));
 383
 384         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 385         dentry = filter_oa2dentry(obd, oa);
 386         if (IS_ERR(dentry)) {
 387                 rc = PTR_ERR(dentry);
 388                 dentry = NULL;
 389                 GOTO(cleanup, rc);
 390         }
 391
 392         inode = dentry->d_inode;
 393
 394         obdo_to_inode(inode, oa, OBD_MD_FLATIME);
 395
 396         rc = filter_map_remote_to_local(objcount, obj, nb, pages, res);
 397         if (rc)
 398                 GOTO(cleanup, rc);
 399
 400         fsfilt_check_slow(obd, now, "preprw_read setup");
 401
 402         /* find pages for all segments, fill array with them */
 403         do_gettimeofday(&start);
 404         for (i = 0, lnb = res; i < *pages; i++, lnb++) {
 405
 406                 lnb->dentry = dentry;
 407
 408                 if (i_size_read(inode) <= lnb->offset)
 409                         /* If there's no more data, abort early.  lnb->rc == 0,
 410                          * so it's easy to detect later. */
 411                         break;
 412
 413                 lnb->page = filter_get_page(obd, inode, lnb->offset);
 414                 if (lnb->page == NULL)
 415                         GOTO(cleanup, rc = -ENOMEM);
 416
 417                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS, 1);
 418
 419                 if (i_size_read(inode) < lnb->offset + lnb->len - 1)
 420                         lnb->rc = i_size_read(inode) - lnb->offset;
 421                 else
 422                         lnb->rc = lnb->len;
 423
 424                 tot_bytes += lnb->rc;
 425
 426                 if (PageUptodate(lnb->page)) {
 427                         lprocfs_counter_add(obd->obd_stats,
 428                                             LPROC_FILTER_CACHE_HIT, 1);
 429                         continue;
 430                 }
 431
 432                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_MISS, 1);
 433                 filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 434         }
 435         do_gettimeofday(&end);
 436         timediff = cfs_timeval_sub(&end, &start, NULL);
 437         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
 438
 439         if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
 440                 GOTO(cleanup, rc = -ENOMEM);
 441
 442         fsfilt_check_slow(obd, now, "start_page_read");
 443
 444         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf,
 445                               exp, NULL, NULL, NULL);
 446         if (rc)
 447                 GOTO(cleanup, rc);
 448
 449         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_READ_BYTES, tot_bytes);
 450
 451         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
 452                 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
 453                                     LPROC_FILTER_READ_BYTES, tot_bytes);
 454
 455         EXIT;
 456
 457  cleanup:
 458         /* unlock pages to allow access from concurrent OST_READ */
 459         for (i = 0, lnb = res; i < *pages; i++, lnb++) {
 460                 if (lnb->page) {
 461                         LASSERT(PageLocked(lnb->page));
 462                         unlock_page(lnb->page);
 463
 464                         if (rc) {
 465                                 page_cache_release(lnb->page);
 466                                 lnb->page = NULL;
 467                         }
 468                 }
 469         }
 470
 471         if (inode && (fo->fo_read_cache == 0 ||
 472                         i_size_read(inode) > fo->fo_readcache_max_filesize))
 473                 filter_invalidate_cache(obd, obj, nb, inode);
 474
 475         if (rc != 0) {
 476                 if (dentry != NULL)
 477                         f_dput(dentry);
 478         }
 479
 480         filter_iobuf_put(&obd->u.filter, iobuf, oti);
 481
 482         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 483         if (rc)
 484                 CERROR("io error %d\n", rc);
 485
 486         return rc;
 487 }
 488
 489 /* When clients have dirtied as much space as they've been granted they
 490  * fall through to sync writes.  These sync writes haven't been expressed
 491  * in grants and need to error with ENOSPC when there isn't room in the
 492  * filesystem for them after grants are taken into account.  However,
 493  * writeback of the dirty data that was already granted space can write
 494  * right on through.
 495  *
 496  * Caller must hold obd_osfs_lock. */
 497 static int filter_grant_check(struct obd_export *exp, struct obdo *oa,
 498                               int objcount, struct fsfilt_objinfo *fso,
 499                               int niocount, struct niobuf_local *lnb,
 500                               obd_size *left, struct inode *inode)
 501 {
 502         struct filter_export_data *fed = &exp->exp_filter_data;
 503         int blocksize = exp->exp_obd->u.obt.obt_sb->s_blocksize;
 504         unsigned long used = 0, ungranted = 0, using;
 505         int i, rc = -ENOSPC, obj, n = 0;
 506
 507         LASSERT_SPIN_LOCKED(&exp->exp_obd->obd_osfs_lock);
 508
 509         for (obj = 0; obj < objcount; obj++) {
 510                 for (i = 0; i < fso[obj].fso_bufcnt; i++, n++) {
 511                         int tmp, bytes;
 512
 513                         /* should match the code in osc_exit_cache */
 514                         bytes = lnb[n].len;
 515                         bytes += lnb[n].offset & (blocksize - 1);
 516                         tmp = (lnb[n].offset + lnb[n].len) & (blocksize - 1);
 517                         if (tmp)
 518                                 bytes += blocksize - tmp;
 519
 520                         if ((lnb[n].flags & OBD_BRW_FROM_GRANT) &&
 521                             (oa->o_valid & OBD_MD_FLGRANT)) {
 522                                 if (fed->fed_grant < used + bytes) {
 523                                         CDEBUG(D_CACHE,
 524                                                "%s: cli %s/%p claims %ld+%d "
 525                                                "GRANT, real grant %lu idx %d\n",
 526                                                exp->exp_obd->obd_name,
 527                                                exp->exp_client_uuid.uuid, exp,
 528                                                used, bytes, fed->fed_grant, n);
 529                                 } else {
 530                                         used += bytes;
 531                                         lnb[n].flags |= OBD_BRW_GRANTED;
 532                                         lnb[n].lnb_grant_used = bytes;
 533                                         CDEBUG(0, "idx %d used=%lu\n", n, used);
 534                                         rc = 0;
 535                                         continue;
 536                                 }
 537                         }
 538                         if (*left > ungranted + bytes) {
 539                                 /* if enough space, pretend it was granted */
 540                                 ungranted += bytes;
 541                                 lnb[n].flags |= OBD_BRW_GRANTED;
 542                                 lnb[n].lnb_grant_used = bytes;
 543                                 CDEBUG(0, "idx %d ungranted=%lu\n",n,ungranted);
 544                                 rc = 0;
 545                                 continue;
 546                         }
 547
 548                         /* We can't check for already-mapped blocks here, as
 549                          * it requires dropping the osfs lock to do the bmap.
 550                          * Instead, we return ENOSPC and in that case we need
 551                          * to go through and verify if all of the blocks not
 552                          * marked BRW_GRANTED are already mapped and we can
 553                          * ignore this error. */
 554                         lnb[n].rc = -ENOSPC;
 555                         lnb[n].flags &= ~OBD_BRW_GRANTED;
 556                         CDEBUG(D_CACHE,"%s: cli %s/%p idx %d no space for %d\n",
 557                                exp->exp_obd->obd_name,
 558                                exp->exp_client_uuid.uuid, exp, n, bytes);
 559                 }
 560         }
 561
 562         /* Now substract what client have used already.  We don't subtract
 563          * this from the tot_granted yet, so that other client's can't grab
 564          * that space before we have actually allocated our blocks.  That
 565          * happens in filter_grant_commit() after the writes are done. */
 566         *left -= ungranted;
 567         fed->fed_grant -= used;
 568         fed->fed_pending += used + ungranted;
 569         exp->exp_obd->u.filter.fo_tot_granted += ungranted;
 570         exp->exp_obd->u.filter.fo_tot_pending += used + ungranted;
 571
 572         CDEBUG(D_CACHE,
 573                "%s: cli %s/%p used: %lu ungranted: %lu grant: %lu dirty: %lu\n",
 574                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp, used,
 575                ungranted, fed->fed_grant, fed->fed_dirty);
 576
 577         /* Rough calc in case we don't refresh cached statfs data */
 578         using = (used + ungranted + 1 ) >>
 579                 exp->exp_obd->u.obt.obt_sb->s_blocksize_bits;
 580         if (exp->exp_obd->obd_osfs.os_bavail > using)
 581                 exp->exp_obd->obd_osfs.os_bavail -= using;
 582         else
 583                 exp->exp_obd->obd_osfs.os_bavail = 0;
 584
 585         if (fed->fed_dirty < used) {
 586                 CERROR("%s: cli %s/%p claims used %lu > fed_dirty %lu\n",
 587                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 588                        used, fed->fed_dirty);
 589                 used = fed->fed_dirty;
 590         }
 591         exp->exp_obd->u.filter.fo_tot_dirty -= used;
 592         fed->fed_dirty -= used;
 593
 594         if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
 595                 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
 596                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 597                        fed->fed_dirty, fed->fed_pending, fed->fed_grant);
 598                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
 599                 LBUG();
 600         }
 601         return rc;
 602 }
 603
 604 /* If we ever start to support multi-object BRW RPCs, we will need to get locks
 605  * on mulitple inodes.  That isn't all, because there still exists the
 606  * possibility of a truncate starting a new transaction while holding the ext3
 607  * rwsem = write while some writes (which have started their transactions here)
 608  * blocking on the ext3 rwsem = read => lock inversion.
 609  *
 610  * The handling gets very ugly when dealing with locked pages.  It may be easier
 611  * to just get rid of the locked page code (which has problems of its own) and
 612  * either discover we do not need it anymore (i.e. it was a symptom of another
 613  * bug) or ensure we get the page locks in an appropriate order. */
 614 static int filter_preprw_write(int cmd, struct obd_export *exp, struct obdo *oa,
 615                                int objcount, struct obd_ioobj *obj,
 616                                struct niobuf_remote *nb, int *pages,
 617                                struct niobuf_local *res,
 618                                struct obd_trans_info *oti,
 619                                struct lustre_capa *capa)
 620 {
 621         struct obd_device *obd = exp->exp_obd;
 622         struct timeval start, end;
 623         struct lvfs_run_ctxt saved;
 624         struct niobuf_local *lnb = res;
 625         struct fsfilt_objinfo fso;
 626         struct filter_mod_data *fmd;
 627         struct dentry *dentry = NULL;
 628         void *iobuf;
 629         obd_size left;
 630         unsigned long now = jiffies, timediff;
 631         int rc = 0, i, tot_bytes = 0, cleanup_phase = 0;
 632         ENTRY;
 633         LASSERT(objcount == 1);
 634         LASSERT(obj->ioo_bufcnt > 0);
 635
 636         rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa), capa,
 637                               CAPA_OPC_OSS_WRITE);
 638         if (rc)
 639                 RETURN(rc);
 640
 641         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 642         iobuf = filter_iobuf_get(&exp->exp_obd->u.filter, oti);
 643         if (IS_ERR(iobuf))
 644                 GOTO(cleanup, rc = PTR_ERR(iobuf));
 645         cleanup_phase = 1;
 646
 647         dentry = filter_fid2dentry(exp->exp_obd, NULL, obj->ioo_gr,
 648                                    obj->ioo_id);
 649         if (IS_ERR(dentry))
 650                 GOTO(cleanup, rc = PTR_ERR(dentry));
 651         cleanup_phase = 2;
 652
 653         if (dentry->d_inode == NULL) {
 654                 CERROR("%s: trying to BRW to non-existent file "LPU64"\n",
 655                        exp->exp_obd->obd_name, obj->ioo_id);
 656                 GOTO(cleanup, rc = -ENOENT);
 657         }
 658
 659         rc = filter_map_remote_to_local(objcount, obj, nb, pages, res);
 660         if (rc)
 661                 GOTO(cleanup, rc);
 662
 663         fsfilt_check_slow(exp->exp_obd, now, "preprw_write setup");
 664
 665         /* Don't update inode timestamps if this write is older than a
 666          * setattr which modifies the timestamps. b=10150 */
 667         /* XXX when we start having persistent reservations this needs to
 668          * be changed to filter_fmd_get() to create the fmd if it doesn't
 669          * already exist so we can store the reservation handle there. */
 670         fmd = filter_fmd_find(exp, obj->ioo_id, obj->ioo_gr);
 671
 672         LASSERT(oa != NULL);
 673         spin_lock(&exp->exp_obd->obd_osfs_lock);
 674         filter_grant_incoming(exp, oa);
 675         if (fmd && fmd->fmd_mactime_xid > oti->oti_xid)
 676                 oa->o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 677                                  OBD_MD_FLATIME);
 678         else
 679                 obdo_to_inode(dentry->d_inode, oa, OBD_MD_FLATIME |
 680                               OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 681         cleanup_phase = 3;
 682
 683         left = filter_grant_space_left(exp);
 684
 685         fso.fso_dentry = dentry;
 686         fso.fso_bufcnt = *pages;
 687
 688         rc = filter_grant_check(exp, oa, objcount, &fso, *pages, res,
 689                                 &left, dentry->d_inode);
 690
 691         /* do not zero out oa->o_valid as it is used in filter_commitrw_write()
 692          * for setting UID/GID and fid EA in first write time. */
 693         if (oa->o_valid & OBD_MD_FLGRANT)
 694                 oa->o_grant = filter_grant(exp,oa->o_grant,oa->o_undirty,left);
 695
 696         spin_unlock(&exp->exp_obd->obd_osfs_lock);
 697         filter_fmd_put(exp, fmd);
 698
 699         if (rc)
 700                 GOTO(cleanup, rc);
 701
 702         do_gettimeofday(&start);
 703         for (i = 0, lnb = res; i < *pages; i++, lnb++) {
 704
 705                 /* We still set up for ungranted pages so that granted pages
 706                  * can be written to disk as they were promised, and portals
 707                  * needs to keep the pages all aligned properly. */
 708                 lnb->dentry = dentry;
 709
 710                 lnb->page = filter_get_page(obd, dentry->d_inode, lnb->offset);
 711                 if (lnb->page == NULL)
 712                         GOTO(cleanup, rc = -ENOMEM);
 713                 cleanup_phase = 4;
 714
 715                 /* DLM locking protects us from write and truncate competing
 716                  * for same region, but truncate can leave dirty page in the
 717                  * cache. it's possible the writeout on a such a page is in
 718                  * progress when we access it. it's also possible that during
 719                  * this writeout we put new (partial) data, but then won't
 720                  * be able to proceed in filter_commitrw_write(). thus let's
 721                  * just wait for writeout completion, should be rare enough.
 722                  * -bzzz */
 723                 wait_on_page_writeback(lnb->page);
 724
 725                 /* If the filter writes a partial page, then has the file
 726                  * extended, the client will read in the whole page.  the
 727                  * filter has to be careful to zero the rest of the partial
 728                  * page on disk.  we do it by hand for partial extending
 729                  * writes, send_bio() is responsible for zeroing pages when
 730                  * asked to read unmapped blocks -- brw_kiovec() does this. */
 731                 if (lnb->len != CFS_PAGE_SIZE) {
 732                         __s64 maxidx;
 733
 734                         maxidx = ((i_size_read(dentry->d_inode) +
 735                                    CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT) - 1;
 736                         if (maxidx >= lnb->page->index) {
 737                                 LL_CDEBUG_PAGE(D_PAGE, lnb->page, "write %u @ "
 738                                                LPU64" flg %x before EOF %llu\n",
 739                                                lnb->len, lnb->offset,lnb->flags,
 740                                                i_size_read(dentry->d_inode));
 741                                 filter_iobuf_add_page(exp->exp_obd, iobuf,
 742                                                       dentry->d_inode,
 743                                                       lnb->page);
 744                         } else {
 745                                 long off;
 746                                 char *p = kmap(lnb->page);
 747
 748                                 off = lnb->offset & ~CFS_PAGE_MASK;
 749                                 if (off)
 750                                         memset(p, 0, off);
 751                                 off = (lnb->offset + lnb->len) & ~CFS_PAGE_MASK;
 752                                 if (off)
 753                                         memset(p + off, 0, CFS_PAGE_SIZE - off);
 754                                 kunmap(lnb->page);
 755                         }
 756                 }
 757                 if (lnb->rc == 0)
 758                         tot_bytes += lnb->len;
 759         }
 760         do_gettimeofday(&end);
 761         timediff = cfs_timeval_sub(&end, &start, NULL);
 762         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
 763
 764         if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
 765                 GOTO(cleanup, rc = -ENOMEM);
 766
 767         /* don't unlock pages to prevent any access */
 768         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf, exp,
 769                               NULL, NULL, NULL);
 770
 771         fsfilt_check_slow(exp->exp_obd, now, "start_page_write");
 772
 773         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
 774                 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
 775                                     LPROC_FILTER_WRITE_BYTES, tot_bytes);
 776         EXIT;
 777 cleanup:
 778         switch(cleanup_phase) {
 779         case 4:
 780                 if (rc) {
 781                         for (i = 0, lnb = res; i < *pages; i++, lnb++) {
 782                                 if (lnb->page != NULL) {
 783                                         unlock_page(lnb->page);
 784                                         page_cache_release(lnb->page);
 785                                         lnb->page = NULL;
 786                                 }
 787                         }
 788                 }
 789         case 3:
 790                 filter_iobuf_put(&exp->exp_obd->u.filter, iobuf, oti);
 791         case 2:
 792                 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 793                 if (rc)
 794                         f_dput(dentry);
 795                 break;
 796         case 1:
 797                 filter_iobuf_put(&exp->exp_obd->u.filter, iobuf, oti);
 798         case 0:
 799                 spin_lock(&exp->exp_obd->obd_osfs_lock);
 800                 if (oa)
 801                         filter_grant_incoming(exp, oa);
 802                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
 803                 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 804                 break;
 805         default:;
 806         }
 807         return rc;
 808 }
 809
 810 int filter_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
 811                   int objcount, struct obd_ioobj *obj,
 812                   struct niobuf_remote *nb, int *pages,
 813                   struct niobuf_local *res, struct obd_trans_info *oti,
 814                   struct lustre_capa *capa)
 815 {
 816         if (cmd == OBD_BRW_WRITE)
 817                 return filter_preprw_write(cmd, exp, oa, objcount, obj,
 818                                            nb, pages, res, oti, capa);
 819         if (cmd == OBD_BRW_READ)
 820                 return filter_preprw_read(cmd, exp, oa, objcount, obj,
 821                                           nb, pages, res, oti, capa);
 822         LBUG();
 823         return -EPROTO;
 824 }
 825
 826 static int filter_commitrw_read(struct obd_export *exp, struct obdo *oa,
 827                                 int objcount, struct obd_ioobj *obj,
 828                                 struct niobuf_remote *rnb,
 829                                 int pages, struct niobuf_local *res,
 830                                 struct obd_trans_info *oti, int rc)
 831 {
 832         struct inode *inode = NULL;
 833         struct ldlm_res_id res_id;
 834         struct ldlm_resource *resource = NULL;
 835         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 836         struct niobuf_local *lnb;
 837         int i;
 838         ENTRY;
 839
 840         osc_build_res_name(obj->ioo_id, obj->ioo_gr, &res_id);
 841         /* If oa != NULL then filter_preprw_read updated the inode atime
 842          * and we should update the lvb so that other glimpses will also
 843          * get the updated value. bug 5972 */
 844         if (oa && ns && ns->ns_lvbo && ns->ns_lvbo->lvbo_update) {
 845                 resource = ldlm_resource_get(ns, NULL, &res_id, LDLM_EXTENT, 0);
 846
 847                 if (resource != NULL) {
 848                         ns->ns_lvbo->lvbo_update(resource, NULL, 0, 1);
 849                         ldlm_resource_putref(resource);
 850                 }
 851         }
 852
 853         if (res->dentry != NULL)
 854                 inode = res->dentry->d_inode;
 855
 856         for (i = 0, lnb = res; i < pages; i++, lnb++) {
 857                 if (lnb->page != NULL) {
 858                         page_cache_release(lnb->page);
 859                         lnb->page = NULL;
 860                 }
 861         }
 862
 863         if (res->dentry != NULL)
 864                 f_dput(res->dentry);
 865         RETURN(rc);
 866 }
 867
 868 void filter_grant_commit(struct obd_export *exp, int niocount,
 869                          struct niobuf_local *res)
 870 {
 871         struct filter_obd *filter = &exp->exp_obd->u.filter;
 872         struct niobuf_local *lnb = res;
 873         unsigned long pending = 0;
 874         int i;
 875
 876         spin_lock(&exp->exp_obd->obd_osfs_lock);
 877         for (i = 0, lnb = res; i < niocount; i++, lnb++)
 878                 pending += lnb->lnb_grant_used;
 879
 880         LASSERTF(exp->exp_filter_data.fed_pending >= pending,
 881                  "%s: cli %s/%p fed_pending: %lu grant_used: %lu\n",
 882                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 883                  exp->exp_filter_data.fed_pending, pending);
 884         exp->exp_filter_data.fed_pending -= pending;
 885         LASSERTF(filter->fo_tot_granted >= pending,
 886                  "%s: cli %s/%p tot_granted: "LPU64" grant_used: %lu\n",
 887                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 888                  exp->exp_obd->u.filter.fo_tot_granted, pending);
 889         filter->fo_tot_granted -= pending;
 890         LASSERTF(filter->fo_tot_pending >= pending,
 891                  "%s: cli %s/%p tot_pending: "LPU64" grant_used: %lu\n",
 892                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 893                  filter->fo_tot_pending, pending);
 894         filter->fo_tot_pending -= pending;
 895
 896         spin_unlock(&exp->exp_obd->obd_osfs_lock);
 897 }
 898
 899 int filter_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
 900                     int objcount, struct obd_ioobj *obj,
 901                     struct niobuf_remote *nb, int pages,
 902                     struct niobuf_local *res, struct obd_trans_info *oti,
 903                     int rc)
 904 {
 905         if (cmd == OBD_BRW_WRITE)
 906                 return filter_commitrw_write(exp, oa, objcount, obj,
 907                                              nb, pages, res, oti, rc);
 908         if (cmd == OBD_BRW_READ)
 909                 return filter_commitrw_read(exp, oa, objcount, obj,
 910                                             nb, pages, res, oti, rc);
 911         LBUG();
 912         return -EPROTO;
 913 }
 914
 915 int filter_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
 916                obd_count oa_bufs, struct brw_page *pga,
 917                struct obd_trans_info *oti)
 918 {
 919         struct obd_ioobj ioo;
 920         struct niobuf_local *lnb;
 921         struct niobuf_remote *rnb;
 922         obd_count i;
 923         int ret = 0, npages;
 924         ENTRY;
 925
 926         OBD_ALLOC(lnb, oa_bufs * sizeof(struct niobuf_local));
 927         OBD_ALLOC(rnb, oa_bufs * sizeof(struct niobuf_remote));
 928
 929         if (lnb == NULL || rnb == NULL)
 930                 GOTO(out, ret = -ENOMEM);
 931
 932         for (i = 0; i < oa_bufs; i++) {
 933                 lnb[i].page = pga[i].pg;
 934                 rnb[i].offset = pga[i].off;
 935                 rnb[i].len = pga[i].count;
 936         }
 937
 938         obdo_to_ioobj(oinfo->oi_oa, &ioo);
 939         ioo.ioo_bufcnt = oa_bufs;
 940
 941         npages = oa_bufs;
 942         ret = filter_preprw(cmd, exp, oinfo->oi_oa, 1, &ioo,
 943                             rnb, &npages, lnb, oti, oinfo_capa(oinfo));
 944         if (ret != 0)
 945                 GOTO(out, ret);
 946         LASSERTF(oa_bufs == npages, "%u != %u\n", oa_bufs, npages);
 947
 948         ret = filter_commitrw(cmd, exp, oinfo->oi_oa, 1, &ioo, rnb,
 949                               npages, lnb, oti, ret);
 950
 951 out:
 952         if (lnb)
 953                 OBD_FREE(lnb, oa_bufs * sizeof(struct niobuf_local));
 954         if (rnb)
 955                 OBD_FREE(rnb, oa_bufs * sizeof(struct niobuf_remote));
 956         RETURN(ret);
 957 }