lustre/obdfilter/filter_io.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/obdfilter/filter_io.c
  37  *
  38  * Author: Peter Braam <braam@clusterfs.com>
  39  * Author: Andreas Dilger <adilger@clusterfs.com>
  40  * Author: Phil Schwan <phil@clusterfs.com>
  41  */
  42
  43 #define DEBUG_SUBSYSTEM S_FILTER
  44
  45 #ifndef AUTOCONF_INCLUDED
  46 #include <linux/config.h>
  47 #endif
  48 #include <linux/module.h>
  49 #include <linux/pagemap.h> // XXX kill me soon
  50 #include <linux/version.h>
  51
  52 #include <obd_class.h>
  53 #include <lustre_fsfilt.h>
  54 #include "filter_internal.h"
  55
  56 int *obdfilter_created_scratchpad;
  57
  58 /* Grab the dirty and seen grant announcements from the incoming obdo.
  59  * We will later calculate the clients new grant and return it.
  60  * Caller must hold osfs lock */
  61 void filter_grant_incoming(struct obd_export *exp, struct obdo *oa)
  62 {
  63         struct filter_export_data *fed;
  64         struct obd_device *obd = exp->exp_obd;
  65         ENTRY;
  66
  67         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
  68
  69         if ((oa->o_valid & (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) !=
  70                                         (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) {
  71                 oa->o_valid &= ~OBD_MD_FLGRANT;
  72                 EXIT;
  73                 return;
  74         }
  75
  76         fed = &exp->exp_filter_data;
  77
  78         /* Add some margin, since there is a small race if other RPCs arrive
  79          * out-or-order and have already consumed some grant.  We want to
  80          * leave this here in case there is a large error in accounting. */
  81         CDEBUG(D_CACHE,
  82                "%s: cli %s/%p reports grant: "LPU64" dropped: %u, local: %lu\n",
  83                obd->obd_name, exp->exp_client_uuid.uuid, exp, oa->o_grant,
  84                oa->o_dropped, fed->fed_grant);
  85
  86         /* Update our accounting now so that statfs takes it into account.
  87          * Note that fed_dirty is only approximate and can become incorrect
  88          * if RPCs arrive out-of-order.  No important calculations depend
  89          * on fed_dirty however, but we must check sanity to not assert. */
  90         if ((long long)oa->o_dirty < 0)
  91                 oa->o_dirty = 0;
  92         else if (oa->o_dirty > fed->fed_grant + 4 * FILTER_GRANT_CHUNK)
  93                 oa->o_dirty = fed->fed_grant + 4 * FILTER_GRANT_CHUNK;
  94         obd->u.filter.fo_tot_dirty += oa->o_dirty - fed->fed_dirty;
  95         if (fed->fed_grant < oa->o_dropped) {
  96                 CDEBUG(D_CACHE,"%s: cli %s/%p reports %u dropped > grant %lu\n",
  97                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
  98                        oa->o_dropped, fed->fed_grant);
  99                 oa->o_dropped = 0;
 100         }
 101         if (obd->u.filter.fo_tot_granted < oa->o_dropped) {
 102                 CERROR("%s: cli %s/%p reports %u dropped > tot_grant "LPU64"\n",
 103                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
 104                        oa->o_dropped, obd->u.filter.fo_tot_granted);
 105                 oa->o_dropped = 0;
 106         }
 107         obd->u.filter.fo_tot_granted -= oa->o_dropped;
 108         fed->fed_grant -= oa->o_dropped;
 109         fed->fed_dirty = oa->o_dirty;
 110
 111         if (oa->o_valid & OBD_MD_FLFLAGS && oa->o_flags & OBD_FL_SHRINK_GRANT) {
 112                 obd_size left_space = filter_grant_space_left(exp);
 113                 struct filter_obd *filter = &exp->exp_obd->u.filter;
 114
 115                 /*Only if left_space < fo_tot_clients * 32M,
 116                  *then the grant space could be shrinked */
 117                 if (left_space < filter->fo_tot_granted_clients *
 118                                  FILTER_GRANT_SHRINK_LIMIT) {
 119                         fed->fed_grant -= oa->o_grant;
 120                         filter->fo_tot_granted -= oa->o_grant;
 121                         CDEBUG(D_CACHE, "%s: cli %s/%p shrink "LPU64
 122                                "fed_grant %ld total "LPU64"\n",
 123                                obd->obd_name, exp->exp_client_uuid.uuid,
 124                                exp, oa->o_grant, fed->fed_grant,
 125                                filter->fo_tot_granted);
 126                         oa->o_grant = 0;
 127                 }
 128         }
 129
 130         if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
 131                 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
 132                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
 133                        fed->fed_dirty, fed->fed_pending, fed->fed_grant);
 134                 spin_unlock(&obd->obd_osfs_lock);
 135                 LBUG();
 136         }
 137         EXIT;
 138 }
 139
 140 /* Figure out how much space is available between what we've granted
 141  * and what remains in the filesystem.  Compensate for ext3 indirect
 142  * block overhead when computing how much free space is left ungranted.
 143  *
 144  * Caller must hold obd_osfs_lock. */
 145 obd_size filter_grant_space_left(struct obd_export *exp)
 146 {
 147         struct obd_device *obd = exp->exp_obd;
 148         int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
 149         obd_size tot_granted = obd->u.filter.fo_tot_granted, avail, left = 0;
 150         int rc, statfs_done = 0;
 151
 152         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
 153
 154         if (cfs_time_before_64(obd->obd_osfs_age, cfs_time_current_64() - HZ)) {
 155 restat:
 156                 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb,
 157                                    cfs_time_current_64() + HZ);
 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  * 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                                 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  * Invalidating the pages to get them out of cache doesn't work because
 342  * LNET pins the pages.  Instead (on newer kernels) the pages are truncated
 343  * from the cache, while older kernels (RHEL4 and SLES9) just leave them in
 344  * the cache.  b=18718/
 345  */
 346 void filter_release_cache(struct obd_device *obd, struct obd_ioobj *obj,
 347                           struct niobuf_remote *rnb, struct inode *inode)
 348 {
 349         int i;
 350
 351         LASSERT(inode != NULL);
 352         for (i = 0; i < obj->ioo_bufcnt; i++, rnb++) {
 353 #ifdef HAVE_TRUNCATE_RANGE
 354                 /* remove pages in which range is fit */
 355                 truncate_inode_pages_range(inode->i_mapping,
 356                                            rnb->offset & CFS_PAGE_MASK,
 357                                            (rnb->offset + rnb->len - 1) |
 358                                            ~CFS_PAGE_MASK);
 359 #else
 360                 /* use invalidate for old kernels */
 361                 invalidate_mapping_pages(inode->i_mapping,
 362                                          rnb->offset >> CFS_PAGE_SHIFT,
 363                                          (rnb->offset + rnb->len) >>
 364                                          CFS_PAGE_SHIFT);
 365 #endif
 366         }
 367 }
 368
 369 static int filter_preprw_read(int cmd, struct obd_export *exp, struct obdo *oa,
 370                               int objcount, struct obd_ioobj *obj,
 371                               struct niobuf_remote *nb,
 372                               int *pages, struct niobuf_local *res,
 373                               struct obd_trans_info *oti)
 374 {
 375         struct obd_device *obd = exp->exp_obd;
 376         struct timeval start, end;
 377         struct lvfs_run_ctxt saved;
 378         struct niobuf_local *lnb;
 379         struct dentry *dentry = NULL;
 380         struct inode *inode = NULL;
 381         void *iobuf = NULL;
 382         int rc = 0, i, tot_bytes = 0;
 383         unsigned long now = jiffies;
 384         long timediff;
 385         loff_t isize;
 386         ENTRY;
 387
 388         /* We are currently not supporting multi-obj BRW_READ RPCS at all.
 389          * When we do this function's dentry cleanup will need to be fixed.
 390          * These values are verified in ost_brw_write() from the wire. */
 391         LASSERTF(objcount == 1, "%d\n", objcount);
 392         LASSERTF(obj->ioo_bufcnt > 0, "%d\n", obj->ioo_bufcnt);
 393
 394         if (oa->o_valid & OBD_MD_FLGRANT) {
 395                 spin_lock(&obd->obd_osfs_lock);
 396                 filter_grant_incoming(exp, oa);
 397
 398                 if (!(oa->o_valid & OBD_MD_FLFLAGS) ||
 399                     !(oa->o_flags & OBD_FL_SHRINK_GRANT))
 400                         oa->o_grant = 0;
 401                 spin_unlock(&obd->obd_osfs_lock);
 402         }
 403
 404         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 405         if (IS_ERR(iobuf))
 406                 RETURN(PTR_ERR(iobuf));
 407
 408         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 409         dentry = filter_oa2dentry(obd, oa);
 410         if (IS_ERR(dentry)) {
 411                 rc = PTR_ERR(dentry);
 412                 dentry = NULL;
 413                 GOTO(cleanup, rc);
 414         }
 415
 416         inode = dentry->d_inode;
 417         /* While we are reading i_size only once, it might change after that
 418          * while we are still reading, but this is perfectly fine race that
 419          * we do not need to care about (bug 20142).                       */
 420         isize = i_size_read(inode);
 421
 422         obdo_to_inode(inode, oa, OBD_MD_FLATIME);
 423
 424         rc = filter_map_remote_to_local(objcount, obj, nb, pages, res);
 425         if (rc)
 426                 GOTO(cleanup, rc);
 427
 428         fsfilt_check_slow(obd, now, "preprw_read setup");
 429
 430         /* find pages for all segments, fill array with them */
 431         do_gettimeofday(&start);
 432         for (i = 0, lnb = res; i < *pages; i++, lnb++) {
 433
 434                 lnb->dentry = dentry;
 435
 436                 if (isize <= lnb->offset)
 437                         /* If there's no more data, abort early.  lnb->rc == 0,
 438                          * so it's easy to detect later. */
 439                         break;
 440
 441                 lnb->page = filter_get_page(obd, inode, lnb->offset, 0);
 442                 if (lnb->page == NULL)
 443                         GOTO(cleanup, rc = -ENOMEM);
 444
 445                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS, 1);
 446
 447                 if (isize < lnb->offset + lnb->len - 1)
 448                         lnb->rc = isize - lnb->offset;
 449                 else
 450                         lnb->rc = lnb->len;
 451
 452                 tot_bytes += lnb->rc;
 453
 454                 if (PageUptodate(lnb->page)) {
 455                         lprocfs_counter_add(obd->obd_stats,
 456                                             LPROC_FILTER_CACHE_HIT, 1);
 457                         continue;
 458                 }
 459
 460                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_MISS, 1);
 461                 filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 462         }
 463         do_gettimeofday(&end);
 464         timediff = cfs_timeval_sub(&end, &start, NULL);
 465         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
 466
 467         if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
 468                 GOTO(cleanup, rc = -ENOMEM);
 469
 470         fsfilt_check_slow(obd, now, "start_page_read");
 471
 472         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf,
 473                               exp, NULL, NULL, NULL);
 474         if (rc)
 475                 GOTO(cleanup, rc);
 476
 477         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_READ_BYTES, tot_bytes);
 478         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
 479                 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
 480                                     LPROC_FILTER_READ_BYTES, tot_bytes);
 481
 482         EXIT;
 483
 484  cleanup:
 485         /* unlock pages to allow access from concurrent OST_READ */
 486         for (i = 0, lnb = res; i < *pages; i++, lnb++) {
 487                 if (lnb->page) {
 488                         LASSERT(PageLocked(lnb->page));
 489                         unlock_page(lnb->page);
 490
 491                         if (rc) {
 492                                 page_cache_release(lnb->page);
 493                                 lnb->page = NULL;
 494                         }
 495                 }
 496         }
 497
 498         if (rc != 0) {
 499                 if (dentry != NULL)
 500                         f_dput(dentry);
 501         }
 502
 503         filter_iobuf_put(&obd->u.filter, iobuf, oti);
 504
 505         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 506         if (rc)
 507                 CERROR("io error %d\n", rc);
 508
 509         return rc;
 510 }
 511
 512 /* When clients have dirtied as much space as they've been granted they
 513  * fall through to sync writes.  These sync writes haven't been expressed
 514  * in grants and need to error with ENOSPC when there isn't room in the
 515  * filesystem for them after grants are taken into account.  However,
 516  * writeback of the dirty data that was already granted space can write
 517  * right on through.
 518  *
 519  * Caller must hold obd_osfs_lock. */
 520 static int filter_grant_check(struct obd_export *exp, struct obdo *oa,
 521                               int objcount, struct fsfilt_objinfo *fso,
 522                               int niocount, struct niobuf_local *lnb,
 523                               obd_size *left, struct inode *inode)
 524 {
 525         struct filter_export_data *fed = &exp->exp_filter_data;
 526         int blocksize = exp->exp_obd->u.obt.obt_sb->s_blocksize;
 527         unsigned long used = 0, ungranted = 0, using;
 528         int i, rc = -ENOSPC, obj, n = 0;
 529         int resend = 0;
 530
 531         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 532             (oa->o_flags & OBD_FL_RECOV_RESEND)) {
 533                 resend = 1;
 534                 CDEBUG(D_CACHE, "Recoverable resend arrived, skipping "
 535                        "accounting\n");
 536         }
 537
 538         LASSERT_SPIN_LOCKED(&exp->exp_obd->obd_osfs_lock);
 539
 540         for (obj = 0; obj < objcount; obj++) {
 541                 for (i = 0; i < fso[obj].fso_bufcnt; i++, n++) {
 542                         int tmp, bytes;
 543
 544                         /* should match the code in osc_exit_cache */
 545                         bytes = lnb[n].len;
 546                         bytes += lnb[n].offset & (blocksize - 1);
 547                         tmp = (lnb[n].offset + lnb[n].len) & (blocksize - 1);
 548                         if (tmp)
 549                                 bytes += blocksize - tmp;
 550
 551                         if ((lnb[n].flags & OBD_BRW_FROM_GRANT) &&
 552                             (oa->o_valid & OBD_MD_FLGRANT)) {
 553                                 if (resend) {
 554                                         /* this is a recoverable resent */
 555                                         lnb[n].flags |= OBD_BRW_GRANTED;
 556                                         lnb[n].lnb_grant_used = 0;
 557                                         rc = 0;
 558                                         continue;
 559                                 } else if (fed->fed_grant < used + bytes) {
 560                                         CDEBUG(D_CACHE,
 561                                                "%s: cli %s/%p claims %ld+%d "
 562                                                "GRANT, real grant %lu idx %d\n",
 563                                                exp->exp_obd->obd_name,
 564                                                exp->exp_client_uuid.uuid, exp,
 565                                                used, bytes, fed->fed_grant, n);
 566                                 } else {
 567                                         used += bytes;
 568                                         lnb[n].flags |= OBD_BRW_GRANTED;
 569                                         lnb[n].lnb_grant_used = bytes;
 570                                         CDEBUG(0, "idx %d used=%lu\n", n, used);
 571                                         rc = 0;
 572                                         continue;
 573                                 }
 574                         }
 575                         if (*left > ungranted + bytes) {
 576                                 /* if enough space, pretend it was granted */
 577                                 ungranted += bytes;
 578                                 lnb[n].flags |= OBD_BRW_GRANTED;
 579                                 lnb[n].lnb_grant_used = bytes;
 580                                 CDEBUG(0, "idx %d ungranted=%lu\n",n,ungranted);
 581                                 rc = 0;
 582                                 continue;
 583                         }
 584
 585                         /* We can't check for already-mapped blocks here, as
 586                          * it requires dropping the osfs lock to do the bmap.
 587                          * Instead, we return ENOSPC and in that case we need
 588                          * to go through and verify if all of the blocks not
 589                          * marked BRW_GRANTED are already mapped and we can
 590                          * ignore this error. */
 591                         lnb[n].rc = -ENOSPC;
 592                         lnb[n].flags &= ~OBD_BRW_GRANTED;
 593                         CDEBUG(D_CACHE,"%s: cli %s/%p idx %d no space for %d\n",
 594                                exp->exp_obd->obd_name,
 595                                exp->exp_client_uuid.uuid, exp, n, bytes);
 596                 }
 597         }
 598
 599         /* Now substract what client have used already.  We don't subtract
 600          * this from the tot_granted yet, so that other client's can't grab
 601          * that space before we have actually allocated our blocks.  That
 602          * happens in filter_grant_commit() after the writes are done. */
 603         *left -= ungranted;
 604         fed->fed_grant -= used;
 605         fed->fed_pending += used + ungranted;
 606         exp->exp_obd->u.filter.fo_tot_granted += ungranted;
 607         exp->exp_obd->u.filter.fo_tot_pending += used + ungranted;
 608
 609         CDEBUG(D_CACHE,
 610                "%s: cli %s/%p used: %lu ungranted: %lu grant: %lu dirty: %lu\n",
 611                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp, used,
 612                ungranted, fed->fed_grant, fed->fed_dirty);
 613
 614         /* Rough calc in case we don't refresh cached statfs data */
 615         using = (used + ungranted + 1 ) >>
 616                 exp->exp_obd->u.obt.obt_sb->s_blocksize_bits;
 617         if (exp->exp_obd->obd_osfs.os_bavail > using)
 618                 exp->exp_obd->obd_osfs.os_bavail -= using;
 619         else
 620                 exp->exp_obd->obd_osfs.os_bavail = 0;
 621
 622         if (fed->fed_dirty < used) {
 623                 CERROR("%s: cli %s/%p claims used %lu > fed_dirty %lu\n",
 624                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 625                        used, fed->fed_dirty);
 626                 used = fed->fed_dirty;
 627         }
 628         exp->exp_obd->u.filter.fo_tot_dirty -= used;
 629         fed->fed_dirty -= used;
 630
 631         if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
 632                 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
 633                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 634                        fed->fed_dirty, fed->fed_pending, fed->fed_grant);
 635                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
 636                 LBUG();
 637         }
 638         return rc;
 639 }
 640
 641 /* If we ever start to support multi-object BRW RPCs, we will need to get locks
 642  * on mulitple inodes.  That isn't all, because there still exists the
 643  * possibility of a truncate starting a new transaction while holding the ext3
 644  * rwsem = write while some writes (which have started their transactions here)
 645  * blocking on the ext3 rwsem = read => lock inversion.
 646  *
 647  * The handling gets very ugly when dealing with locked pages.  It may be easier
 648  * to just get rid of the locked page code (which has problems of its own) and
 649  * either discover we do not need it anymore (i.e. it was a symptom of another
 650  * bug) or ensure we get the page locks in an appropriate order. */
 651 static int filter_preprw_write(int cmd, struct obd_export *exp, struct obdo *oa,
 652                                int objcount, struct obd_ioobj *obj,
 653                                struct niobuf_remote *nb, int *pages,
 654                                struct niobuf_local *res,
 655                                struct obd_trans_info *oti)
 656 {
 657         struct obd_device *obd = exp->exp_obd;
 658         struct timeval start, end;
 659         struct lvfs_run_ctxt saved;
 660         struct niobuf_local *lnb = res;
 661         struct fsfilt_objinfo fso;
 662         struct filter_mod_data *fmd;
 663         struct dentry *dentry = NULL;
 664         void *iobuf;
 665         obd_size left;
 666         unsigned long now = jiffies, timediff;
 667         int rc = 0, i, tot_bytes = 0, cleanup_phase = 0, localreq = 0;
 668         ENTRY;
 669         LASSERT(objcount == 1);
 670         LASSERT(obj->ioo_bufcnt > 0);
 671
 672         if (exp->exp_connection &&
 673             exp->exp_connection->c_peer.nid == exp->exp_connection->c_self)
 674                 localreq = 1;
 675
 676         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 677         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 678         if (IS_ERR(iobuf))
 679                 GOTO(cleanup, rc = PTR_ERR(iobuf));
 680         cleanup_phase = 1;
 681
 682         dentry = filter_fid2dentry(obd, NULL, obj->ioo_gr, 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_recreate(exp->exp_obd, noa) == 0) {
 700                                 f_dput(dentry);
 701                                 dentry = filter_fid2dentry(exp->exp_obd, NULL,
 702                                                            obj->ioo_gr,
 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_gr,
 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         rc = filter_map_remote_to_local(objcount, obj, nb, pages, res);
 721         if (rc)
 722                 GOTO(cleanup, rc);
 723
 724         fsfilt_check_slow(obd, now, "preprw_write setup");
 725
 726         /* Filter truncate first locks i_mutex then partially truncated
 727          * page, filter write code first locks pages then take
 728          * i_mutex.  To avoid a deadlock in case of concurrent
 729          * punch/write requests from one client, filter writes and
 730          * filter truncates are serialized by i_alloc_sem, allowing
 731          * multiple writes or single truncate. */
 732         down_read(&dentry->d_inode->i_alloc_sem);
 733
 734         /* Don't update inode timestamps if this write is older than a
 735          * setattr which modifies the timestamps. b=10150 */
 736         /* XXX when we start having persistent reservations this needs to
 737          * be changed to filter_fmd_get() to create the fmd if it doesn't
 738          * already exist so we can store the reservation handle there. */
 739         fmd = filter_fmd_find(exp, obj->ioo_id, obj->ioo_gr);
 740
 741         LASSERT(oa != NULL);
 742         spin_lock(&obd->obd_osfs_lock);
 743
 744         filter_grant_incoming(exp, oa);
 745         if (fmd && fmd->fmd_mactime_xid > oti->oti_xid)
 746                 oa->o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 747                                  OBD_MD_FLATIME);
 748         else
 749                 obdo_to_inode(dentry->d_inode, oa, OBD_MD_FLATIME |
 750                               OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 751         cleanup_phase = 3;
 752
 753         left = filter_grant_space_left(exp);
 754
 755         fso.fso_dentry = dentry;
 756         fso.fso_bufcnt = *pages;
 757
 758         rc = filter_grant_check(exp, oa, objcount, &fso, *pages, res,
 759                                 &left, dentry->d_inode);
 760
 761         /* do not zero out oa->o_valid as it is used in filter_commitrw_write()
 762          * for setting UID/GID and fid EA in first write time. */
 763         /* If OBD_FL_SHRINK_GRANT is set, the client just returned us some grant
 764          * so no sense in allocating it some more. We either return the grant
 765          * back to the client if we have plenty of space or we don't return
 766          * anything if we are short. This was decided in filter_grant_incoming*/
 767         if ((oa->o_valid & OBD_MD_FLGRANT) &&
 768             (!(oa->o_valid & OBD_MD_FLFLAGS) ||
 769              !(oa->o_flags & OBD_FL_SHRINK_GRANT)))
 770                 oa->o_grant = filter_grant(exp, oa->o_grant, oa->o_undirty,
 771                                            left, 1);
 772
 773         spin_unlock(&obd->obd_osfs_lock);
 774         filter_fmd_put(exp, fmd);
 775
 776         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK2, (obd_timeout + 1) / 4);
 777
 778         if (rc)
 779                 GOTO(cleanup, rc);
 780         cleanup_phase = 4;
 781
 782         do_gettimeofday(&start);
 783         for (i = 0, lnb = res; i < *pages; i++, lnb++) {
 784
 785                 /* We still set up for ungranted pages so that granted pages
 786                  * can be written to disk as they were promised, and portals
 787                  * needs to keep the pages all aligned properly. */
 788                 lnb->dentry = dentry;
 789
 790                 lnb->page = filter_get_page(obd, dentry->d_inode, lnb->offset,
 791                                             localreq);
 792                 if (lnb->page == NULL)
 793                         GOTO(cleanup, rc = -ENOMEM);
 794
 795                 /* DLM locking protects us from write and truncate competing
 796                  * for same region, but truncate can leave dirty page in the
 797                  * cache. it's possible the writeout on a such a page is in
 798                  * progress when we access it. it's also possible that during
 799                  * this writeout we put new (partial) data, but then won't
 800                  * be able to proceed in filter_commitrw_write(). thus let's
 801                  * just wait for writeout completion, should be rare enough.
 802                  * -bzzz */
 803                 wait_on_page_writeback(lnb->page);
 804                 BUG_ON(PageWriteback(lnb->page));
 805
 806                 /* If the filter writes a partial page, then has the file
 807                  * extended, the client will read in the whole page.  the
 808                  * filter has to be careful to zero the rest of the partial
 809                  * page on disk.  we do it by hand for partial extending
 810                  * writes, send_bio() is responsible for zeroing pages when
 811                  * asked to read unmapped blocks -- brw_kiovec() does this. */
 812                 if (lnb->len != CFS_PAGE_SIZE) {
 813                         __s64 maxidx;
 814
 815                         maxidx = ((i_size_read(dentry->d_inode) +
 816                                    CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT) - 1;
 817                         if (maxidx >= lnb->page->index) {
 818                                 LL_CDEBUG_PAGE(D_PAGE, lnb->page, "write %u @ "
 819                                                LPU64" flg %x before EOF %llu\n",
 820                                                lnb->len, lnb->offset,lnb->flags,
 821                                                i_size_read(dentry->d_inode));
 822                                 filter_iobuf_add_page(obd, iobuf,
 823                                                       dentry->d_inode,
 824                                                       lnb->page);
 825                         } else {
 826                                 long off;
 827                                 char *p = kmap(lnb->page);
 828
 829                                 off = lnb->offset & ~CFS_PAGE_MASK;
 830                                 if (off)
 831                                         memset(p, 0, off);
 832                                 off = (lnb->offset + lnb->len) & ~CFS_PAGE_MASK;
 833                                 if (off)
 834                                         memset(p + off, 0, CFS_PAGE_SIZE - off);
 835                                 kunmap(lnb->page);
 836                         }
 837                 }
 838                 if (lnb->rc == 0)
 839                         tot_bytes += lnb->len;
 840         }
 841         do_gettimeofday(&end);
 842         timediff = cfs_timeval_sub(&end, &start, NULL);
 843         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
 844
 845         if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
 846                 GOTO(cleanup, rc = -ENOMEM);
 847
 848         /* don't unlock pages to prevent any access */
 849         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf, exp,
 850                               NULL, NULL, NULL);
 851
 852         fsfilt_check_slow(obd, now, "start_page_write");
 853
 854         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
 855                             tot_bytes);
 856         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
 857                 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
 858                                     LPROC_FILTER_WRITE_BYTES, tot_bytes);
 859         EXIT;
 860 cleanup:
 861         switch(cleanup_phase) {
 862         case 4:
 863                 if (rc) {
 864                         for (i = 0, lnb = res; i < *pages; i++, lnb++) {
 865                                 if (lnb->page != NULL) {
 866                                         unlock_page(lnb->page);
 867                                         page_cache_release(lnb->page);
 868                                         lnb->page = NULL;
 869                                 }
 870                         }
 871                         filter_grant_commit(exp, *pages, res);
 872                 }
 873         case 3:
 874                 if (rc)
 875                         up_read(&dentry->d_inode->i_alloc_sem);
 876
 877                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 878         case 2:
 879                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 880                 if (rc)
 881                         f_dput(dentry);
 882                 break;
 883         case 1:
 884                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 885         case 0:
 886                 spin_lock(&obd->obd_osfs_lock);
 887                 if (oa)
 888                         filter_grant_incoming(exp, oa);
 889                 spin_unlock(&obd->obd_osfs_lock);
 890                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 891                 break;
 892         default:;
 893         }
 894         return rc;
 895 }
 896
 897 int filter_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
 898                   int objcount, struct obd_ioobj *obj,
 899                   struct niobuf_remote *nb, int *pages,
 900                   struct niobuf_local *res, struct obd_trans_info *oti)
 901 {
 902         if (cmd == OBD_BRW_WRITE)
 903                 return filter_preprw_write(cmd, exp, oa, objcount, obj,
 904                                            nb, pages, res, oti);
 905         if (cmd == OBD_BRW_READ)
 906                 return filter_preprw_read(cmd, exp, oa, objcount, obj,
 907                                           nb, pages, res, oti);
 908         LBUG();
 909         return -EPROTO;
 910 }
 911
 912 static int filter_commitrw_read(struct obd_export *exp, struct obdo *oa,
 913                                 int objcount, struct obd_ioobj *obj,
 914                                 struct niobuf_remote *rnb,
 915                                 int pages, struct niobuf_local *res,
 916                                 struct obd_trans_info *oti, int rc)
 917 {
 918         struct filter_obd *fo = &exp->exp_obd->u.filter;
 919         struct inode *inode = NULL;
 920         struct ldlm_res_id res_id = { .name = { obj->ioo_id } };
 921         struct ldlm_resource *resource = NULL;
 922         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 923         struct niobuf_local *lnb;
 924         int i;
 925         ENTRY;
 926
 927         /* If oa != NULL then filter_preprw_read updated the inode atime
 928          * and we should update the lvb so that other glimpses will also
 929          * get the updated value. bug 5972 */
 930         if (oa && ns && ns->ns_lvbo && ns->ns_lvbo->lvbo_update) {
 931                 resource = ldlm_resource_get(ns, NULL, res_id, LDLM_EXTENT, 0);
 932
 933                 if (resource != NULL) {
 934                         ns->ns_lvbo->lvbo_update(resource, NULL, 0, 1);
 935                         ldlm_resource_putref(resource);
 936                 }
 937         }
 938
 939         if (res->dentry != NULL)
 940                 inode = res->dentry->d_inode;
 941
 942         for (i = 0, lnb = res; i < pages; i++, lnb++) {
 943                 if (lnb->page != NULL) {
 944                         page_cache_release(lnb->page);
 945                         lnb->page = NULL;
 946                 }
 947         }
 948
 949         if (inode && (fo->fo_read_cache == 0 ||
 950                       i_size_read(inode) > fo->fo_readcache_max_filesize))
 951                 filter_release_cache(exp->exp_obd, obj, rnb, inode);
 952
 953         if (res->dentry != NULL)
 954                 f_dput(res->dentry);
 955         RETURN(rc);
 956 }
 957
 958 void filter_grant_commit(struct obd_export *exp, int niocount,
 959                          struct niobuf_local *res)
 960 {
 961         struct filter_obd *filter = &exp->exp_obd->u.filter;
 962         struct niobuf_local *lnb = res;
 963         unsigned long pending = 0;
 964         int i;
 965
 966         spin_lock(&exp->exp_obd->obd_osfs_lock);
 967         for (i = 0, lnb = res; i < niocount; i++, lnb++)
 968                 pending += lnb->lnb_grant_used;
 969
 970         LASSERTF(exp->exp_filter_data.fed_pending >= pending,
 971                  "%s: cli %s/%p fed_pending: %lu grant_used: %lu\n",
 972                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 973                  exp->exp_filter_data.fed_pending, pending);
 974         exp->exp_filter_data.fed_pending -= pending;
 975         LASSERTF(filter->fo_tot_granted >= pending,
 976                  "%s: cli %s/%p tot_granted: "LPU64" grant_used: %lu\n",
 977                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 978                  exp->exp_obd->u.filter.fo_tot_granted, pending);
 979         filter->fo_tot_granted -= pending;
 980         LASSERTF(filter->fo_tot_pending >= pending,
 981                  "%s: cli %s/%p tot_pending: "LPU64" grant_used: %lu\n",
 982                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 983                  filter->fo_tot_pending, pending);
 984         filter->fo_tot_pending -= pending;
 985
 986         spin_unlock(&exp->exp_obd->obd_osfs_lock);
 987 }
 988
 989 int filter_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
 990                     int objcount, struct obd_ioobj *obj,
 991                     struct niobuf_remote *nb, int pages,
 992                     struct niobuf_local *res, struct obd_trans_info *oti,
 993                     int rc)
 994 {
 995         if (cmd == OBD_BRW_WRITE)
 996                 return filter_commitrw_write(exp, oa, objcount, obj,
 997                                              nb, pages, res, oti, rc);
 998         if (cmd == OBD_BRW_READ)
 999                 return filter_commitrw_read(exp, oa, objcount, obj,
1000                                             nb, pages, res, oti, rc);
1001         LBUG();
1002         return -EPROTO;
1003 }
1004
1005 int filter_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1006                obd_count oa_bufs, struct brw_page *pga,
1007                struct obd_trans_info *oti)
1008 {
1009         struct obd_ioobj ioo;
1010         struct niobuf_local *lnb;
1011         struct niobuf_remote *rnb;
1012         obd_count i;
1013         int ret = 0, npages;
1014         ENTRY;
1015
1016         OBD_ALLOC(lnb, oa_bufs * sizeof(struct niobuf_local));
1017         OBD_ALLOC(rnb, oa_bufs * sizeof(struct niobuf_remote));
1018
1019         if (lnb == NULL || rnb == NULL)
1020                 GOTO(out, ret = -ENOMEM);
1021
1022         for (i = 0; i < oa_bufs; i++) {
1023                 lnb[i].page = pga[i].pg;
1024                 rnb[i].offset = pga[i].off;
1025                 rnb[i].len = pga[i].count;
1026                 lnb[i].flags = rnb[i].flags = pga[i].flag;
1027         }
1028
1029         obdo_to_ioobj(oinfo->oi_oa, &ioo);
1030         ioo.ioo_bufcnt = oa_bufs;
1031
1032         npages = oa_bufs;
1033         ret = filter_preprw(cmd, exp, oinfo->oi_oa, 1, &ioo,
1034                             rnb, &npages, lnb, oti);
1035         if (ret != 0)
1036                 GOTO(out, ret);
1037         LASSERTF(oa_bufs == npages, "%u != %u\n", oa_bufs, npages);
1038
1039         ret = filter_commitrw(cmd, exp, oinfo->oi_oa, 1, &ioo, rnb,
1040                               npages, lnb, oti, ret);
1041
1042 out:
1043         if (lnb)
1044                 OBD_FREE(lnb, oa_bufs * sizeof(struct niobuf_local));
1045         if (rnb)
1046                 OBD_FREE(rnb, oa_bufs * sizeof(struct niobuf_remote));
1047         RETURN(ret);
1048 }