lustre/obdfilter/filter_io.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  */
  30 /*
  31  * This file is part of Lustre, http://www.lustre.org/
  32  * Lustre is a trademark of Sun Microsystems, Inc.
  33  *
  34  * lustre/obdfilter/filter_io.c
  35  *
  36  * Author: Peter Braam <braam@clusterfs.com>
  37  * Author: Andreas Dilger <adilger@clusterfs.com>
  38  * Author: Phil Schwan <phil@clusterfs.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_FILTER
  42
  43 #ifndef AUTOCONF_INCLUDED
  44 #include <linux/config.h>
  45 #endif
  46 #include <linux/module.h>
  47 #include <linux/pagemap.h> // XXX kill me soon
  48 #include <linux/version.h>
  49
  50 #include <obd_class.h>
  51 #include <obd_ost.h>
  52 #include <lustre_fsfilt.h>
  53 #include "filter_internal.h"
  54
  55 int *obdfilter_created_scratchpad;
  56
  57 /* Grab the dirty and seen grant announcements from the incoming obdo.
  58  * We will later calculate the clients new grant and return it.
  59  * Caller must hold osfs lock */
  60 void filter_grant_incoming(struct obd_export *exp, struct obdo *oa)
  61 {
  62         struct filter_export_data *fed;
  63         struct obd_device *obd = exp->exp_obd;
  64         ENTRY;
  65
  66         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
  67
  68         if ((oa->o_valid & (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) !=
  69                                         (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) {
  70                 oa->o_valid &= ~OBD_MD_FLGRANT;
  71                 EXIT;
  72                 return;
  73         }
  74
  75         fed = &exp->exp_filter_data;
  76
  77         /* Add some margin, since there is a small race if other RPCs arrive
  78          * out-or-order and have already consumed some grant.  We want to
  79          * leave this here in case there is a large error in accounting. */
  80         CDEBUG(D_CACHE,
  81                "%s: cli %s/%p reports grant: "LPU64" dropped: %u, local: %lu\n",
  82                obd->obd_name, exp->exp_client_uuid.uuid, exp, oa->o_grant,
  83                oa->o_dropped, fed->fed_grant);
  84
  85         /* Update our accounting now so that statfs takes it into account.
  86          * Note that fed_dirty is only approximate and can become incorrect
  87          * if RPCs arrive out-of-order.  No important calculations depend
  88          * on fed_dirty however, but we must check sanity to not assert. */
  89         if ((long long)oa->o_dirty < 0)
  90                 oa->o_dirty = 0;
  91         else if (oa->o_dirty > fed->fed_grant + 4 * FILTER_GRANT_CHUNK)
  92                 oa->o_dirty = fed->fed_grant + 4 * FILTER_GRANT_CHUNK;
  93         obd->u.filter.fo_tot_dirty += oa->o_dirty - fed->fed_dirty;
  94         if (fed->fed_grant < oa->o_dropped) {
  95                 CDEBUG(D_CACHE,"%s: cli %s/%p reports %u dropped > grant %lu\n",
  96                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
  97                        oa->o_dropped, fed->fed_grant);
  98                 oa->o_dropped = 0;
  99         }
 100         if (obd->u.filter.fo_tot_granted < oa->o_dropped) {
 101                 CERROR("%s: cli %s/%p reports %u dropped > tot_grant "LPU64"\n",
 102                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
 103                        oa->o_dropped, obd->u.filter.fo_tot_granted);
 104                 oa->o_dropped = 0;
 105         }
 106         obd->u.filter.fo_tot_granted -= oa->o_dropped;
 107         fed->fed_grant -= oa->o_dropped;
 108         fed->fed_dirty = oa->o_dirty;
 109
 110         if (oa->o_valid & OBD_MD_FLFLAGS && oa->o_flags & OBD_FL_SHRINK_GRANT) {
 111                 obd_size left_space = filter_grant_space_left(exp);
 112                 struct filter_obd *filter = &exp->exp_obd->u.filter;
 113
 114                 /*Only if left_space < fo_tot_clients * 32M,
 115                  *then the grant space could be shrinked */
 116                 if (left_space < filter->fo_tot_granted_clients *
 117                                  FILTER_GRANT_SHRINK_LIMIT) {
 118                         fed->fed_grant -= oa->o_grant;
 119                         filter->fo_tot_granted -= oa->o_grant;
 120                         CDEBUG(D_CACHE, "%s: cli %s/%p shrink "LPU64
 121                                "fed_grant %ld total "LPU64"\n",
 122                                obd->obd_name, exp->exp_client_uuid.uuid,
 123                                exp, oa->o_grant, fed->fed_grant,
 124                                filter->fo_tot_granted);
 125                         oa->o_grant = 0;
 126                 }
 127         }
 128
 129         if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
 130                 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
 131                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
 132                        fed->fed_dirty, fed->fed_pending, fed->fed_grant);
 133                 cfs_spin_unlock(&obd->obd_osfs_lock);
 134                 LBUG();
 135         }
 136         EXIT;
 137 }
 138
 139 /* Figure out how much space is available between what we've granted
 140  * and what remains in the filesystem.  Compensate for ext3 indirect
 141  * block overhead when computing how much free space is left ungranted.
 142  *
 143  * Caller must hold obd_osfs_lock. */
 144 obd_size filter_grant_space_left(struct obd_export *exp)
 145 {
 146         struct obd_device *obd = exp->exp_obd;
 147         int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
 148         obd_size tot_granted = obd->u.filter.fo_tot_granted, avail, left = 0;
 149         int rc, statfs_done = 0;
 150
 151         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
 152
 153         if (cfs_time_before_64(obd->obd_osfs_age,
 154                                cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS))) {
 155 restat:
 156                 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb,
 157                                    cfs_time_shift_64(OBD_STATFS_CACHE_SECONDS));
 158                 if (rc) /* N.B. statfs can't really fail */
 159                         RETURN(0);
 160                 statfs_done = 1;
 161         }
 162
 163         avail = obd->obd_osfs.os_bavail;
 164         left = avail - (avail >> (blockbits - 3)); /* (d)indirect */
 165         if (left > GRANT_FOR_LLOG(obd)) {
 166                 left = (left - GRANT_FOR_LLOG(obd)) << blockbits;
 167         } else {
 168                 left = 0 /* << blockbits */;
 169         }
 170
 171         if (!statfs_done && left < 32 * FILTER_GRANT_CHUNK + tot_granted) {
 172                 CDEBUG(D_CACHE, "fs has no space left and statfs too old\n");
 173                 goto restat;
 174         }
 175
 176         if (left >= tot_granted) {
 177                 left -= tot_granted;
 178         } else {
 179                 if (left < tot_granted - obd->u.filter.fo_tot_pending) {
 180                         CERROR("%s: cli %s/%p grant "LPU64" > available "
 181                                LPU64" and pending "LPU64"\n", obd->obd_name,
 182                                exp->exp_client_uuid.uuid, exp, tot_granted,
 183                                left, obd->u.filter.fo_tot_pending);
 184                 }
 185                 left = 0;
 186         }
 187
 188         CDEBUG(D_CACHE, "%s: cli %s/%p free: "LPU64" avail: "LPU64" grant "LPU64
 189                " left: "LPU64" pending: "LPU64"\n", obd->obd_name,
 190                exp->exp_client_uuid.uuid, exp,
 191                obd->obd_osfs.os_bfree << blockbits, avail << blockbits,
 192                tot_granted, left, obd->u.filter.fo_tot_pending);
 193
 194         return left;
 195 }
 196
 197 /* Calculate how much grant space to allocate to this client, based on how
 198  * much space is currently free and how much of that is already granted.
 199  *
 200  * if @conservative != 0, we limit the maximum grant to FILTER_GRANT_CHUNK;
 201  * otherwise we'll satisfy the requested amount as possible as we can, this
 202  * is usually due to client reconnect.
 203  *
 204  * Caller must hold obd_osfs_lock. */
 205 long filter_grant(struct obd_export *exp, obd_size current_grant,
 206                   obd_size want, obd_size fs_space_left, int conservative)
 207 {
 208         struct obd_device *obd = exp->exp_obd;
 209         struct filter_export_data *fed = &exp->exp_filter_data;
 210         int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
 211         __u64 grant = 0;
 212
 213         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
 214
 215         /* Grant some fraction of the client's requested grant space so that
 216          * they are not always waiting for write credits (not all of it to
 217          * avoid overgranting in face of multiple RPCs in flight).  This
 218          * essentially will be able to control the OSC_MAX_RIF for a client.
 219          *
 220          * If we do have a large disparity between what the client thinks it
 221          * has and what we think it has, don't grant very much and let the
 222          * client consume its grant first.  Either it just has lots of RPCs
 223          * in flight, or it was evicted and its grants will soon be used up. */
 224         if (want > 0x7fffffff) {
 225                 CERROR("%s: client %s/%p requesting > 2GB grant "LPU64"\n",
 226                        obd->obd_name, exp->exp_client_uuid.uuid, exp, want);
 227         } else if (current_grant < want &&
 228                    current_grant < fed->fed_grant + FILTER_GRANT_CHUNK) {
 229                 grant = min(want + (1 << blockbits) - 1, fs_space_left / 8);
 230                 grant &= ~((1ULL << blockbits) - 1);
 231
 232                 if (grant) {
 233                         if (grant > FILTER_GRANT_CHUNK && conservative)
 234                                 grant = FILTER_GRANT_CHUNK;
 235
 236                         obd->u.filter.fo_tot_granted += grant;
 237                         fed->fed_grant += grant;
 238                         if (fed->fed_grant < 0) {
 239                                 CERROR("%s: cli %s/%p grant %ld want "LPU64
 240                                        "current"LPU64"\n",
 241                                        obd->obd_name, exp->exp_client_uuid.uuid,
 242                                        exp, fed->fed_grant, want,current_grant);
 243                                 cfs_spin_unlock(&obd->obd_osfs_lock);
 244                                 LBUG();
 245                         }
 246                 }
 247         }
 248
 249         CDEBUG(D_CACHE,
 250                "%s: cli %s/%p wants: "LPU64" current grant "LPU64
 251                " granting: "LPU64"\n", obd->obd_name, exp->exp_client_uuid.uuid,
 252                exp, want, current_grant, grant);
 253         CDEBUG(D_CACHE,
 254                "%s: cli %s/%p tot cached:"LPU64" granted:"LPU64
 255                " num_exports: %d\n", obd->obd_name, exp->exp_client_uuid.uuid,
 256                exp, obd->u.filter.fo_tot_dirty,
 257                obd->u.filter.fo_tot_granted, obd->obd_num_exports);
 258
 259         return grant;
 260 }
 261
 262 /*
 263  * the routine is used to request pages from pagecache
 264  *
 265  * use GFP_NOFS for requests from a local client not allowing to enter FS
 266  * as we might end up waiting on a page he sent in the request we're serving.
 267  * use __GFP_HIGHMEM so that the pages can use all of the available memory
 268  * on 32-bit machines
 269  * use more aggressive GFP_HIGHUSER flags from non-local clients to be able to
 270  * generate more memory pressure.
 271  *
 272  * See Bug 19529 and Bug 19917 for details.
 273  */
 274 static struct page *filter_get_page(struct obd_device *obd,
 275                                     struct inode *inode,
 276                                     obd_off offset,
 277                                     int localreq)
 278 {
 279         struct page *page;
 280
 281         page = find_or_create_page(inode->i_mapping, offset >> CFS_PAGE_SHIFT,
 282                                    (localreq ? (GFP_NOFS | __GFP_HIGHMEM)
 283                                              : GFP_HIGHUSER));
 284         if (unlikely(page == NULL))
 285                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_NO_PAGE, 1);
 286
 287         return page;
 288 }
 289
 290 /*
 291  * the routine initializes array of local_niobuf from remote_niobuf
 292  */
 293 static int filter_map_remote_to_local(int objcount, struct obd_ioobj *obj,
 294                                       struct niobuf_remote *nb,
 295                                       int *nrpages, struct niobuf_local *res)
 296 {
 297         struct niobuf_remote *rnb;
 298         struct niobuf_local *lnb;
 299         int i, max;
 300         ENTRY;
 301
 302         /* we don't support multiobject RPC yet
 303          * ost_brw_read() and ost_brw_write() check this */
 304         LASSERT(objcount == 1);
 305
 306         max = *nrpages;
 307         *nrpages = 0;
 308         for (i = 0, rnb = nb, lnb = res; i < obj->ioo_bufcnt; i++, rnb++) {
 309                 obd_off offset = rnb->offset;
 310                 unsigned int len = rnb->len;
 311
 312                 while (len > 0) {
 313                         int poff = offset & (CFS_PAGE_SIZE - 1);
 314                         int plen = CFS_PAGE_SIZE - poff;
 315
 316                         if (*nrpages >= max) {
 317                                 CERROR("small array of local bufs: %d\n", max);
 318                                 RETURN(-EINVAL);
 319                         }
 320
 321                         if (plen > len)
 322                                 plen = len;
 323                         lnb->offset = offset;
 324                         lnb->len = plen;
 325                         lnb->flags = rnb->flags;
 326                         lnb->page = NULL;
 327                         lnb->rc = 0;
 328                         lnb->lnb_grant_used = 0;
 329
 330                         LASSERTF(plen <= len, "plen %u, len %u\n", plen, len);
 331                         offset += plen;
 332                         len -= plen;
 333                         lnb++;
 334                         (*nrpages)++;
 335                 }
 336         }
 337         RETURN(0);
 338 }
 339
 340 /*
 341  * the invalidate above doesn't work during read because lnet pins pages.
 342  * The truncate is used here instead to drop pages from cache
 343  */
 344 void filter_release_cache(struct obd_device *obd, struct obd_ioobj *obj,
 345                           struct niobuf_remote *rnb, struct inode *inode)
 346 {
 347         int i;
 348
 349         LASSERT(inode != NULL);
 350         for (i = 0; i < obj->ioo_bufcnt; i++, rnb++) {
 351                 /* remove pages in which range is fit */
 352                 truncate_inode_pages_range(inode->i_mapping,
 353                                            rnb->offset & CFS_PAGE_MASK,
 354                                            (rnb->offset + rnb->len - 1) |
 355                                            ~CFS_PAGE_MASK);
 356         }
 357 }
 358
 359 static int filter_preprw_read(int cmd, struct obd_export *exp, struct obdo *oa,
 360                               int objcount, struct obd_ioobj *obj,
 361                               struct niobuf_remote *nb,
 362                               int *npages, struct niobuf_local *res,
 363                               struct obd_trans_info *oti,
 364                               struct lustre_capa *capa)
 365 {
 366         struct obd_device *obd = exp->exp_obd;
 367         struct timeval start, end;
 368         struct lvfs_run_ctxt saved;
 369         struct niobuf_local *lnb;
 370         struct dentry *dentry = NULL;
 371         struct inode *inode = NULL;
 372         void *iobuf = NULL;
 373         int rc = 0, i, tot_bytes = 0;
 374         unsigned long now = jiffies;
 375         long timediff;
 376         loff_t isize;
 377         ENTRY;
 378
 379         /* We are currently not supporting multi-obj BRW_READ RPCS at all.
 380          * When we do this function's dentry cleanup will need to be fixed.
 381          * These values are verified in ost_brw_write() from the wire. */
 382         LASSERTF(objcount == 1, "%d\n", objcount);
 383         LASSERTF(obj->ioo_bufcnt > 0, "%d\n", obj->ioo_bufcnt);
 384
 385         rc = filter_auth_capa(exp, NULL, oa->o_seq, capa,
 386                               CAPA_OPC_OSS_READ);
 387         if (rc)
 388                 RETURN(rc);
 389
 390         if (oa && oa->o_valid & OBD_MD_FLGRANT) {
 391                 cfs_spin_lock(&obd->obd_osfs_lock);
 392                 filter_grant_incoming(exp, oa);
 393
 394                 if (!(oa->o_valid & OBD_MD_FLFLAGS) ||
 395                     !(oa->o_flags & OBD_FL_SHRINK_GRANT))
 396                         oa->o_grant = 0;
 397                 cfs_spin_unlock(&obd->obd_osfs_lock);
 398         }
 399
 400         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 401         if (IS_ERR(iobuf))
 402                 RETURN(PTR_ERR(iobuf));
 403
 404         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 405         dentry = filter_oa2dentry(obd, &oa->o_oi);
 406         if (IS_ERR(dentry)) {
 407                 rc = PTR_ERR(dentry);
 408                 dentry = NULL;
 409                 GOTO(cleanup, rc);
 410         }
 411
 412         inode = dentry->d_inode;
 413         isize = i_size_read(inode);
 414
 415         obdo_to_inode(inode, oa, OBD_MD_FLATIME);
 416
 417         rc = filter_map_remote_to_local(objcount, obj, nb, npages, res);
 418         if (rc)
 419                 GOTO(cleanup, rc);
 420
 421         fsfilt_check_slow(obd, now, "preprw_read setup");
 422
 423         /* find pages for all segments, fill array with them */
 424         cfs_gettimeofday(&start);
 425         for (i = 0, lnb = res; i < *npages; i++, lnb++) {
 426
 427                 lnb->dentry = dentry;
 428
 429                 if (isize <= lnb->offset)
 430                         /* If there's no more data, abort early.  lnb->rc == 0,
 431                          * so it's easy to detect later. */
 432                         break;
 433
 434                 lnb->page = filter_get_page(obd, inode, lnb->offset, 0);
 435                 if (lnb->page == NULL)
 436                         GOTO(cleanup, rc = -ENOMEM);
 437
 438                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS, 1);
 439
 440                 if (isize < lnb->offset + lnb->len - 1)
 441                         lnb->rc = isize - lnb->offset;
 442                 else
 443                         lnb->rc = lnb->len;
 444
 445                 tot_bytes += lnb->rc;
 446
 447                 if (PageUptodate(lnb->page)) {
 448                         lprocfs_counter_add(obd->obd_stats,
 449                                             LPROC_FILTER_CACHE_HIT, 1);
 450                         continue;
 451                 }
 452
 453                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_MISS, 1);
 454                 filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 455         }
 456         cfs_gettimeofday(&end);
 457         timediff = cfs_timeval_sub(&end, &start, NULL);
 458         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
 459
 460         if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
 461                 GOTO(cleanup, rc = -ENOMEM);
 462
 463         fsfilt_check_slow(obd, now, "start_page_read");
 464
 465         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf,
 466                               exp, NULL, NULL, NULL);
 467         if (rc)
 468                 GOTO(cleanup, rc);
 469
 470         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_READ_BYTES, tot_bytes);
 471
 472         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
 473                 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
 474                                     LPROC_FILTER_READ_BYTES, tot_bytes);
 475
 476         EXIT;
 477
 478  cleanup:
 479         /* unlock pages to allow access from concurrent OST_READ */
 480         for (i = 0, lnb = res; i < *npages; i++, lnb++) {
 481                 if (lnb->page) {
 482                         LASSERT(PageLocked(lnb->page));
 483                         unlock_page(lnb->page);
 484
 485                         if (rc) {
 486                                 page_cache_release(lnb->page);
 487                                 lnb->page = NULL;
 488                         }
 489                 }
 490         }
 491
 492         if (rc != 0) {
 493                 if (dentry != NULL)
 494                         f_dput(dentry);
 495         }
 496
 497         filter_iobuf_put(&obd->u.filter, iobuf, oti);
 498
 499         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 500         if (rc)
 501                 CERROR("io error %d\n", rc);
 502
 503         return rc;
 504 }
 505
 506 /* When clients have dirtied as much space as they've been granted they
 507  * fall through to sync writes.  These sync writes haven't been expressed
 508  * in grants and need to error with ENOSPC when there isn't room in the
 509  * filesystem for them after grants are taken into account.  However,
 510  * writeback of the dirty data that was already granted space can write
 511  * right on through.
 512  *
 513  * Caller must hold obd_osfs_lock. */
 514 static int filter_grant_check(struct obd_export *exp, struct obdo *oa,
 515                               int objcount, struct fsfilt_objinfo *fso,
 516                               int niocount, struct niobuf_local *lnb,
 517                               obd_size *left, struct inode *inode)
 518 {
 519         struct filter_export_data *fed = &exp->exp_filter_data;
 520         int blocksize = exp->exp_obd->u.obt.obt_sb->s_blocksize;
 521         unsigned long used = 0, ungranted = 0, using;
 522         int i, rc = -ENOSPC, obj, n = 0;
 523         int resend = 0;
 524
 525         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 526             (oa->o_flags & OBD_FL_RECOV_RESEND)) {
 527                 resend = 1;
 528                 CDEBUG(D_CACHE, "Recoverable resend arrived, skipping "
 529                        "accounting\n");
 530         }
 531
 532         LASSERT_SPIN_LOCKED(&exp->exp_obd->obd_osfs_lock);
 533
 534         for (obj = 0; obj < objcount; obj++) {
 535                 for (i = 0; i < fso[obj].fso_bufcnt; i++, n++) {
 536                         int tmp, bytes;
 537
 538                         /* should match the code in osc_exit_cache */
 539                         bytes = lnb[n].len;
 540                         bytes += lnb[n].offset & (blocksize - 1);
 541                         tmp = (lnb[n].offset + lnb[n].len) & (blocksize - 1);
 542                         if (tmp)
 543                                 bytes += blocksize - tmp;
 544
 545                         if ((lnb[n].flags & OBD_BRW_FROM_GRANT) &&
 546                             (oa->o_valid & OBD_MD_FLGRANT)) {
 547                                 if (resend) {
 548                                         /* this is a recoverable resent */
 549                                         lnb[n].flags |= OBD_BRW_GRANTED;
 550                                         rc = 0;
 551                                         continue;
 552                                 } else if (fed->fed_grant < used + bytes) {
 553                                         CDEBUG(D_CACHE,
 554                                                "%s: cli %s/%p claims %ld+%d "
 555                                                "GRANT, real grant %lu idx %d\n",
 556                                                exp->exp_obd->obd_name,
 557                                                exp->exp_client_uuid.uuid, exp,
 558                                                used, bytes, fed->fed_grant, n);
 559                                 } else {
 560                                         used += bytes;
 561                                         lnb[n].flags |= OBD_BRW_GRANTED;
 562                                         lnb[n].lnb_grant_used = bytes;
 563                                         CDEBUG(0, "idx %d used=%lu\n", n, used);
 564                                         rc = 0;
 565                                         continue;
 566                                 }
 567                         }
 568                         if (*left > ungranted + bytes) {
 569                                 /* if enough space, pretend it was granted */
 570                                 ungranted += bytes;
 571                                 lnb[n].flags |= OBD_BRW_GRANTED;
 572                                 lnb[n].lnb_grant_used = bytes;
 573                                 CDEBUG(0, "idx %d ungranted=%lu\n",n,ungranted);
 574                                 rc = 0;
 575                                 continue;
 576                         }
 577
 578                         /* We can't check for already-mapped blocks here, as
 579                          * it requires dropping the osfs lock to do the bmap.
 580                          * Instead, we return ENOSPC and in that case we need
 581                          * to go through and verify if all of the blocks not
 582                          * marked BRW_GRANTED are already mapped and we can
 583                          * ignore this error. */
 584                         lnb[n].rc = -ENOSPC;
 585                         lnb[n].flags &= ~OBD_BRW_GRANTED;
 586                         CDEBUG(D_CACHE,"%s: cli %s/%p idx %d no space for %d\n",
 587                                exp->exp_obd->obd_name,
 588                                exp->exp_client_uuid.uuid, exp, n, bytes);
 589                 }
 590         }
 591
 592         /* Now substract what client have used already.  We don't subtract
 593          * this from the tot_granted yet, so that other client's can't grab
 594          * that space before we have actually allocated our blocks.  That
 595          * happens in filter_grant_commit() after the writes are done. */
 596         *left -= ungranted;
 597         fed->fed_grant -= used;
 598         fed->fed_pending += used + ungranted;
 599         exp->exp_obd->u.filter.fo_tot_granted += ungranted;
 600         exp->exp_obd->u.filter.fo_tot_pending += used + ungranted;
 601
 602         CDEBUG(D_CACHE,
 603                "%s: cli %s/%p used: %lu ungranted: %lu grant: %lu dirty: %lu\n",
 604                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp, used,
 605                ungranted, fed->fed_grant, fed->fed_dirty);
 606
 607         /* Rough calc in case we don't refresh cached statfs data */
 608         using = (used + ungranted + 1 ) >>
 609                 exp->exp_obd->u.obt.obt_sb->s_blocksize_bits;
 610         if (exp->exp_obd->obd_osfs.os_bavail > using)
 611                 exp->exp_obd->obd_osfs.os_bavail -= using;
 612         else
 613                 exp->exp_obd->obd_osfs.os_bavail = 0;
 614
 615         if (fed->fed_dirty < used) {
 616                 CERROR("%s: cli %s/%p claims used %lu > fed_dirty %lu\n",
 617                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 618                        used, fed->fed_dirty);
 619                 used = fed->fed_dirty;
 620         }
 621         exp->exp_obd->u.filter.fo_tot_dirty -= used;
 622         fed->fed_dirty -= used;
 623
 624         if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
 625                 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
 626                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 627                        fed->fed_dirty, fed->fed_pending, fed->fed_grant);
 628                 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
 629                 LBUG();
 630         }
 631         return rc;
 632 }
 633
 634 /* If we ever start to support multi-object BRW RPCs, we will need to get locks
 635  * on mulitple inodes.  That isn't all, because there still exists the
 636  * possibility of a truncate starting a new transaction while holding the ext3
 637  * rwsem = write while some writes (which have started their transactions here)
 638  * blocking on the ext3 rwsem = read => lock inversion.
 639  *
 640  * The handling gets very ugly when dealing with locked pages.  It may be easier
 641  * to just get rid of the locked page code (which has problems of its own) and
 642  * either discover we do not need it anymore (i.e. it was a symptom of another
 643  * bug) or ensure we get the page locks in an appropriate order. */
 644 static int filter_preprw_write(int cmd, struct obd_export *exp, struct obdo *oa,
 645                                int objcount, struct obd_ioobj *obj,
 646                                struct niobuf_remote *nb, int *npages,
 647                                struct niobuf_local *res,
 648                                struct obd_trans_info *oti,
 649                                struct lustre_capa *capa)
 650 {
 651         struct obd_device *obd = exp->exp_obd;
 652         struct timeval start, end;
 653         struct lvfs_run_ctxt saved;
 654         struct niobuf_local *lnb = res;
 655         struct fsfilt_objinfo fso;
 656         struct filter_mod_data *fmd;
 657         struct dentry *dentry = NULL;
 658         void *iobuf;
 659         obd_size left;
 660         unsigned long now = jiffies, timediff;
 661         int rc = 0, i, tot_bytes = 0, cleanup_phase = 0, localreq = 0;
 662         ENTRY;
 663         LASSERT(objcount == 1);
 664         LASSERT(obj->ioo_bufcnt > 0);
 665
 666         rc = filter_auth_capa(exp, NULL, oa->o_seq, capa,
 667                               CAPA_OPC_OSS_WRITE);
 668         if (rc)
 669                 RETURN(rc);
 670
 671         if (exp->exp_connection &&
 672             exp->exp_connection->c_peer.nid == exp->exp_connection->c_self)
 673                 localreq = 1;
 674
 675         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 676         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 677         if (IS_ERR(iobuf))
 678                 GOTO(cleanup, rc = PTR_ERR(iobuf));
 679         cleanup_phase = 1;
 680
 681         dentry = filter_fid2dentry(obd, NULL, obj->ioo_seq,
 682                                    obj->ioo_id);
 683         if (IS_ERR(dentry))
 684                 GOTO(cleanup, rc = PTR_ERR(dentry));
 685         cleanup_phase = 2;
 686
 687         if (dentry->d_inode == NULL) {
 688                 if (exp->exp_obd->obd_recovering) {
 689                         struct obdo *noa = oa;
 690
 691                         if (oa == NULL) {
 692                                 OBDO_ALLOC(noa);
 693                                 if (noa == NULL)
 694                                         GOTO(recreate_out, rc = -ENOMEM);
 695                                 noa->o_id = obj->ioo_id;
 696                                 noa->o_valid = OBD_MD_FLID;
 697                         }
 698
 699                         if (filter_create(exp, noa, NULL, oti) == 0) {
 700                                 f_dput(dentry);
 701                                 dentry = filter_fid2dentry(exp->exp_obd, NULL,
 702                                                            obj->ioo_seq,
 703                                                            obj->ioo_id);
 704                         }
 705                         if (oa == NULL)
 706                                 OBDO_FREE(noa);
 707                 }
 708     recreate_out:
 709                 if (IS_ERR(dentry) || dentry->d_inode == NULL) {
 710                         CERROR("%s: BRW to missing obj "LPU64"/"LPU64":rc %d\n",
 711                                exp->exp_obd->obd_name,
 712                                obj->ioo_id, obj->ioo_seq,
 713                                IS_ERR(dentry) ? (int)PTR_ERR(dentry) : -ENOENT);
 714                         if (IS_ERR(dentry))
 715                                 cleanup_phase = 1;
 716                         GOTO(cleanup, rc = -ENOENT);
 717                 }
 718         }
 719
 720         if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID) &&
 721             dentry->d_inode->i_mode & (S_ISUID | S_ISGID)) {
 722                 rc = filter_capa_fixoa(exp, oa, oa->o_seq, capa);
 723                 if (rc)
 724                         GOTO(cleanup, rc);
 725         }
 726
 727         rc = filter_map_remote_to_local(objcount, obj, nb, npages, res);
 728         if (rc)
 729                 GOTO(cleanup, rc);
 730
 731         fsfilt_check_slow(obd, now, "preprw_write setup");
 732
 733         /* Filter truncate first locks i_mutex then partially truncated
 734          * page, filter write code first locks pages then take
 735          * i_mutex.  To avoid a deadlock in case of concurrent
 736          * punch/write requests from one client, filter writes and
 737          * filter truncates are serialized by i_alloc_sem, allowing
 738          * multiple writes or single truncate. */
 739         down_read(&dentry->d_inode->i_alloc_sem);
 740         fsfilt_check_slow(obd, now, "i_alloc_sem");
 741
 742         /* Don't update inode timestamps if this write is older than a
 743          * setattr which modifies the timestamps. b=10150 */
 744         /* XXX when we start having persistent reservations this needs to
 745          * be changed to filter_fmd_get() to create the fmd if it doesn't
 746          * already exist so we can store the reservation handle there. */
 747         fmd = filter_fmd_find(exp, obj->ioo_id, obj->ioo_seq);
 748
 749         LASSERT(oa != NULL);
 750         cfs_spin_lock(&obd->obd_osfs_lock);
 751         filter_grant_incoming(exp, oa);
 752         if (fmd && fmd->fmd_mactime_xid > oti->oti_xid)
 753                 oa->o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 754                                  OBD_MD_FLATIME);
 755         else
 756                 obdo_to_inode(dentry->d_inode, oa, OBD_MD_FLATIME |
 757                               OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 758         cleanup_phase = 3;
 759
 760         left = filter_grant_space_left(exp);
 761
 762         fso.fso_dentry = dentry;
 763         fso.fso_bufcnt = *npages;
 764
 765         rc = filter_grant_check(exp, oa, objcount, &fso, *npages, res,
 766                                 &left, dentry->d_inode);
 767
 768         /* do not zero out oa->o_valid as it is used in filter_commitrw_write()
 769          * for setting UID/GID and fid EA in first write time. */
 770         /* If OBD_FL_SHRINK_GRANT is set, the client just returned us some grant
 771          * so no sense in allocating it some more. We either return the grant
 772          * back to the client if we have plenty of space or we don't return
 773          * anything if we are short. This was decided in filter_grant_incoming*/
 774         if ((oa->o_valid & OBD_MD_FLGRANT) &&
 775             (!(oa->o_valid & OBD_MD_FLFLAGS) ||
 776              !(oa->o_flags & OBD_FL_SHRINK_GRANT)))
 777                 oa->o_grant = filter_grant(exp, oa->o_grant, oa->o_undirty,
 778                                            left, 1);
 779
 780         cfs_spin_unlock(&obd->obd_osfs_lock);
 781         filter_fmd_put(exp, fmd);
 782
 783         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK2, (obd_timeout + 1) / 4);
 784
 785         if (rc)
 786                 GOTO(cleanup, rc);
 787         cleanup_phase = 4;
 788
 789         cfs_gettimeofday(&start);
 790         for (i = 0, lnb = res; i < *npages; i++, lnb++) {
 791
 792                 /* We still set up for ungranted pages so that granted pages
 793                  * can be written to disk as they were promised, and portals
 794                  * needs to keep the pages all aligned properly. */
 795                 lnb->dentry = dentry;
 796
 797                 lnb->page = filter_get_page(obd, dentry->d_inode, lnb->offset,
 798                                             localreq);
 799                 if (lnb->page == NULL)
 800                         GOTO(cleanup, rc = -ENOMEM);
 801
 802                 /* DLM locking protects us from write and truncate competing
 803                  * for same region, but truncate can leave dirty page in the
 804                  * cache. it's possible the writeout on a such a page is in
 805                  * progress when we access it. it's also possible that during
 806                  * this writeout we put new (partial) data, but then won't
 807                  * be able to proceed in filter_commitrw_write(). thus let's
 808                  * just wait for writeout completion, should be rare enough.
 809                  * -bzzz */
 810                 wait_on_page_writeback(lnb->page);
 811                 BUG_ON(PageWriteback(lnb->page));
 812
 813                 /* If the filter writes a partial page, then has the file
 814                  * extended, the client will read in the whole page.  the
 815                  * filter has to be careful to zero the rest of the partial
 816                  * page on disk.  we do it by hand for partial extending
 817                  * writes, send_bio() is responsible for zeroing pages when
 818                  * asked to read unmapped blocks -- brw_kiovec() does this. */
 819                 if (lnb->len != CFS_PAGE_SIZE) {
 820                         __s64 maxidx;
 821
 822                         maxidx = ((i_size_read(dentry->d_inode) +
 823                                    CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT) - 1;
 824                         if (maxidx >= lnb->page->index) {
 825                                 LL_CDEBUG_PAGE(D_PAGE, lnb->page, "write %u @ "
 826                                                LPU64" flg %x before EOF %llu\n",
 827                                                lnb->len, lnb->offset,lnb->flags,
 828                                                i_size_read(dentry->d_inode));
 829                                 filter_iobuf_add_page(obd, iobuf,
 830                                                       dentry->d_inode,
 831                                                       lnb->page);
 832                         } else {
 833                                 long off;
 834                                 char *p = kmap(lnb->page);
 835
 836                                 off = lnb->offset & ~CFS_PAGE_MASK;
 837                                 if (off)
 838                                         memset(p, 0, off);
 839                                 off = (lnb->offset + lnb->len) & ~CFS_PAGE_MASK;
 840                                 if (off)
 841                                         memset(p + off, 0, CFS_PAGE_SIZE - off);
 842                                 kunmap(lnb->page);
 843                         }
 844                 }
 845                 if (lnb->rc == 0)
 846                         tot_bytes += lnb->len;
 847         }
 848         cfs_gettimeofday(&end);
 849         timediff = cfs_timeval_sub(&end, &start, NULL);
 850         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
 851
 852         if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
 853                 GOTO(cleanup, rc = -ENOMEM);
 854
 855         /* don't unlock pages to prevent any access */
 856         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf, exp,
 857                               NULL, NULL, NULL);
 858
 859         fsfilt_check_slow(obd, now, "start_page_write");
 860
 861         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
 862                             tot_bytes);
 863
 864         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
 865                 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
 866                                     LPROC_FILTER_WRITE_BYTES, tot_bytes);
 867         EXIT;
 868 cleanup:
 869         switch(cleanup_phase) {
 870         case 4:
 871                 if (rc) {
 872                         for (i = 0, lnb = res; i < *npages; i++, lnb++) {
 873                                 if (lnb->page != NULL) {
 874                                         unlock_page(lnb->page);
 875                                         page_cache_release(lnb->page);
 876                                         lnb->page = NULL;
 877                                 }
 878                         }
 879                 }
 880         case 3:
 881                 if (rc)
 882                         up_read(&dentry->d_inode->i_alloc_sem);
 883
 884                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 885         case 2:
 886                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 887                 if (rc)
 888                         f_dput(dentry);
 889                 break;
 890         case 1:
 891                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 892         case 0:
 893                 cfs_spin_lock(&obd->obd_osfs_lock);
 894                 if (oa)
 895                         filter_grant_incoming(exp, oa);
 896                 cfs_spin_unlock(&obd->obd_osfs_lock);
 897                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 898                 break;
 899         default:;
 900         }
 901         return rc;
 902 }
 903
 904 int filter_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
 905                   int objcount, struct obd_ioobj *obj,
 906                   struct niobuf_remote *nb, int *npages,
 907                   struct niobuf_local *res, struct obd_trans_info *oti,
 908                   struct lustre_capa *capa)
 909 {
 910         if (cmd == OBD_BRW_WRITE)
 911                 return filter_preprw_write(cmd, exp, oa, objcount, obj,
 912                                            nb, npages, res, oti, capa);
 913         if (cmd == OBD_BRW_READ)
 914                 return filter_preprw_read(cmd, exp, oa, objcount, obj,
 915                                           nb, npages, res, oti, capa);
 916         LBUG();
 917         return -EPROTO;
 918 }
 919
 920 static int filter_commitrw_read(struct obd_export *exp, struct obdo *oa,
 921                                 int objcount, struct obd_ioobj *obj,
 922                                 struct niobuf_remote *rnb,
 923                                 int npages, struct niobuf_local *res,
 924                                 struct obd_trans_info *oti, int rc)
 925 {
 926         struct filter_obd *fo = &exp->exp_obd->u.filter;
 927         struct inode *inode = NULL;
 928         struct ldlm_res_id res_id;
 929         struct ldlm_resource *resource = NULL;
 930         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 931         struct niobuf_local *lnb;
 932         int i;
 933         ENTRY;
 934
 935         osc_build_res_name(obj->ioo_id, obj->ioo_seq, &res_id);
 936         /* If oa != NULL then filter_preprw_read updated the inode atime
 937          * and we should update the lvb so that other glimpses will also
 938          * get the updated value. bug 5972 */
 939         if (oa && ns && ns->ns_lvbo && ns->ns_lvbo->lvbo_update) {
 940                 resource = ldlm_resource_get(ns, NULL, &res_id, LDLM_EXTENT, 0);
 941
 942                 if (resource != NULL) {
 943                         LDLM_RESOURCE_ADDREF(resource);
 944                         ns->ns_lvbo->lvbo_update(resource, NULL, 1);
 945                         LDLM_RESOURCE_DELREF(resource);
 946                         ldlm_resource_putref(resource);
 947                 }
 948         }
 949
 950         if (res->dentry != NULL)
 951                 inode = res->dentry->d_inode;
 952
 953         for (i = 0, lnb = res; i < npages; i++, lnb++) {
 954                 if (lnb->page != NULL) {
 955                         page_cache_release(lnb->page);
 956                         lnb->page = NULL;
 957                 }
 958         }
 959         if (inode && (fo->fo_read_cache == 0 ||
 960                       i_size_read(inode) > fo->fo_readcache_max_filesize))
 961                 filter_release_cache(exp->exp_obd, obj, rnb, inode);
 962
 963         if (res->dentry != NULL)
 964                 f_dput(res->dentry);
 965         RETURN(rc);
 966 }
 967
 968 void filter_grant_commit(struct obd_export *exp, int niocount,
 969                          struct niobuf_local *res)
 970 {
 971         struct filter_obd *filter = &exp->exp_obd->u.filter;
 972         struct niobuf_local *lnb = res;
 973         unsigned long pending = 0;
 974         int i;
 975
 976         cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
 977         for (i = 0, lnb = res; i < niocount; i++, lnb++)
 978                 pending += lnb->lnb_grant_used;
 979
 980         LASSERTF(exp->exp_filter_data.fed_pending >= pending,
 981                  "%s: cli %s/%p fed_pending: %lu grant_used: %lu\n",
 982                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 983                  exp->exp_filter_data.fed_pending, pending);
 984         exp->exp_filter_data.fed_pending -= pending;
 985         LASSERTF(filter->fo_tot_granted >= pending,
 986                  "%s: cli %s/%p tot_granted: "LPU64" grant_used: %lu\n",
 987                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 988                  exp->exp_obd->u.filter.fo_tot_granted, pending);
 989         filter->fo_tot_granted -= pending;
 990         LASSERTF(filter->fo_tot_pending >= pending,
 991                  "%s: cli %s/%p tot_pending: "LPU64" grant_used: %lu\n",
 992                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 993                  filter->fo_tot_pending, pending);
 994         filter->fo_tot_pending -= pending;
 995
 996         cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
 997 }
 998
 999 int filter_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
1000                     int objcount, struct obd_ioobj *obj,
1001                     struct niobuf_remote *nb, int npages,
1002                     struct niobuf_local *res, struct obd_trans_info *oti,
1003                     int rc)
1004 {
1005         if (cmd == OBD_BRW_WRITE)
1006                 return filter_commitrw_write(exp, oa, objcount, obj,
1007                                              nb, npages, res, oti, rc);
1008         if (cmd == OBD_BRW_READ)
1009                 return filter_commitrw_read(exp, oa, objcount, obj,
1010                                             nb, npages, res, oti, rc);
1011         LBUG();
1012         return -EPROTO;
1013 }
1014