lustre/obdfilter/filter_io.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/obdfilter/filter_io.c
  37  *
  38  * Author: Peter Braam <braam@clusterfs.com>
  39  * Author: Andreas Dilger <adilger@clusterfs.com>
  40  * Author: Phil Schwan <phil@clusterfs.com>
  41  */
  42
  43 #define DEBUG_SUBSYSTEM S_FILTER
  44
  45 #ifndef AUTOCONF_INCLUDED
  46 #include <linux/config.h>
  47 #endif
  48 #include <linux/module.h>
  49 #include <linux/pagemap.h> // XXX kill me soon
  50 #include <linux/version.h>
  51
  52 #include <obd_class.h>
  53 #include <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_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  * Caller must hold obd_osfs_lock. */
 201 long filter_grant(struct obd_export *exp, obd_size current_grant,
 202                   obd_size want, obd_size fs_space_left)
 203 {
 204         struct obd_device *obd = exp->exp_obd;
 205         struct filter_export_data *fed = &exp->exp_filter_data;
 206         int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
 207         __u64 grant = 0;
 208
 209         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
 210
 211         /* Grant some fraction of the client's requested grant space so that
 212          * they are not always waiting for write credits (not all of it to
 213          * avoid overgranting in face of multiple RPCs in flight).  This
 214          * essentially will be able to control the OSC_MAX_RIF for a client.
 215          *
 216          * If we do have a large disparity between what the client thinks it
 217          * has and what we think it has, don't grant very much and let the
 218          * client consume its grant first.  Either it just has lots of RPCs
 219          * in flight, or it was evicted and its grants will soon be used up. */
 220         if (want > 0x7fffffff) {
 221                 CERROR("%s: client %s/%p requesting > 2GB grant "LPU64"\n",
 222                        obd->obd_name, exp->exp_client_uuid.uuid, exp, want);
 223         } else if (current_grant < want &&
 224                    current_grant < fed->fed_grant + FILTER_GRANT_CHUNK) {
 225                 grant = min((want >> blockbits),
 226                             (fs_space_left >> blockbits) / 8);
 227                 grant <<= blockbits;
 228
 229                 if (grant) {
 230                         /* Allow >FILTER_GRANT_CHUNK size when clients
 231                          * reconnect due to a server reboot.
 232                          */
 233                         if ((grant > FILTER_GRANT_CHUNK) &&
 234                             (!obd->obd_recovering))
 235                                 grant = FILTER_GRANT_CHUNK;
 236
 237                         obd->u.filter.fo_tot_granted += grant;
 238                         fed->fed_grant += grant;
 239                         if (fed->fed_grant < 0) {
 240                                 CERROR("%s: cli %s/%p grant %ld want "LPU64
 241                                        "current"LPU64"\n",
 242                                        obd->obd_name, exp->exp_client_uuid.uuid,
 243                                        exp, fed->fed_grant, want,current_grant);
 244                                 spin_unlock(&obd->obd_osfs_lock);
 245                                 LBUG();
 246                         }
 247                 }
 248         }
 249
 250         CDEBUG(D_CACHE,
 251                "%s: cli %s/%p wants: "LPU64" current grant "LPU64
 252                " granting: "LPU64"\n", obd->obd_name, exp->exp_client_uuid.uuid,
 253                exp, want, current_grant, grant);
 254         CDEBUG(D_CACHE,
 255                "%s: cli %s/%p tot cached:"LPU64" granted:"LPU64
 256                " num_exports: %d\n", obd->obd_name, exp->exp_client_uuid.uuid,
 257                exp, obd->u.filter.fo_tot_dirty,
 258                obd->u.filter.fo_tot_granted, obd->obd_num_exports);
 259
 260         return grant;
 261 }
 262
 263 /*
 264  * the routine is used to request pages from pagecache
 265  *
 266  * use GFP_NOFS for requests from a local client not allowing to enter FS
 267  * as we might end up waiting on a page he sent in the request we're serving.
 268  * use __GFP_HIGHMEM so that the pages can use all of the available memory
 269  * on 32-bit machines
 270  * use more agressive GFP_HIGHUSER flags from non-local clients to be able to
 271  * generate more memory pressure, but at the same time use __GFP_NOMEMALLOC
 272  * in order not to exhaust emergency reserves.
 273  *
 274  * See Bug 19529 and Bug 19917 for details.
 275  */
 276 static struct page * filter_get_page(struct obd_device *obd,
 277                                      struct inode *inode,
 278                                      obd_off offset,
 279                                      int localreq)
 280 {
 281         struct page *page;
 282
 283         page = find_or_create_page(inode->i_mapping, offset >> CFS_PAGE_SHIFT,
 284                                    (localreq ? (GFP_NOFS | __GFP_HIGHMEM)
 285                                              : (GFP_HIGHUSER | __GFP_NOMEMALLOC)));
 286         if (unlikely(page == NULL))
 287                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_NO_PAGE, 1);
 288
 289         return page;
 290 }
 291
 292 /*
 293  * the routine initializes array of local_niobuf from remote_niobuf
 294  */
 295 static int filter_map_remote_to_local(int objcount, struct obd_ioobj *obj,
 296                                       struct niobuf_remote *nb,
 297                                       int *nrpages, struct niobuf_local *res)
 298 {
 299         struct niobuf_remote *rnb;
 300         struct niobuf_local *lnb;
 301         int i, max;
 302         ENTRY;
 303
 304         /* we don't support multiobject RPC yet
 305          * ost_brw_read() and ost_brw_write() check this */
 306         LASSERT(objcount == 1);
 307
 308         max = *nrpages;
 309         *nrpages = 0;
 310         for (i = 0, rnb = nb, lnb = res; i < obj->ioo_bufcnt; i++, rnb++) {
 311                 obd_off offset = rnb->offset;
 312                 unsigned int len = rnb->len;
 313
 314                 while (len > 0) {
 315                         int poff = offset & (CFS_PAGE_SIZE - 1);
 316                         int plen = CFS_PAGE_SIZE - poff;
 317
 318                         if (*nrpages >= max) {
 319                                 CERROR("small array of local bufs: %d\n", max);
 320                                 RETURN(-EINVAL);
 321                         }
 322
 323                         if (plen > len)
 324                                 plen = len;
 325                         lnb->offset = offset;
 326                         lnb->len = plen;
 327                         lnb->flags = rnb->flags;
 328                         lnb->page = NULL;
 329                         lnb->rc = 0;
 330                         lnb->lnb_grant_used = 0;
 331
 332                         LASSERTF(plen <= len, "plen %u, len %u\n", plen, len);
 333                         offset += plen;
 334                         len -= plen;
 335                         lnb++;
 336                         (*nrpages)++;
 337                 }
 338         }
 339         RETURN(0);
 340 }
 341
 342 /*
 343  * Invalidating the pages to get them out of cache doesn't work because
 344  * LNET pins the pages.  Instead (on newer kernels) the pages are truncated
 345  * from the cache, while older kernels (RHEL4 and SLES9) just leave them in
 346  * the cache.  b=18718/
 347  */
 348 void filter_release_cache(struct obd_device *obd, struct obd_ioobj *obj,
 349                           struct niobuf_remote *rnb, struct inode *inode)
 350 {
 351         int i;
 352
 353         LASSERT(inode != NULL);
 354         for (i = 0; i < obj->ioo_bufcnt; i++, rnb++) {
 355 #ifdef HAVE_TRUNCATE_RANGE
 356                 /* remove pages in which range is fit */
 357                 truncate_inode_pages_range(inode->i_mapping,
 358                                            rnb->offset & CFS_PAGE_MASK,
 359                                            (rnb->offset + rnb->len - 1) |
 360                                            ~CFS_PAGE_MASK);
 361 #else
 362                 /* use invalidate for old kernels */
 363                 invalidate_mapping_pages(inode->i_mapping,
 364                                          rnb->offset >> CFS_PAGE_SHIFT,
 365                                          (rnb->offset + rnb->len) >>
 366                                          CFS_PAGE_SHIFT);
 367 #endif
 368         }
 369 }
 370
 371 static int filter_preprw_read(int cmd, struct obd_export *exp, struct obdo *oa,
 372                               int objcount, struct obd_ioobj *obj,
 373                               struct niobuf_remote *nb,
 374                               int *pages, struct niobuf_local *res,
 375                               struct obd_trans_info *oti)
 376 {
 377         struct obd_device *obd = exp->exp_obd;
 378         struct timeval start, end;
 379         struct lvfs_run_ctxt saved;
 380         struct niobuf_local *lnb;
 381         struct dentry *dentry = NULL;
 382         struct inode *inode = NULL;
 383         void *iobuf = NULL;
 384         int rc = 0, i, tot_bytes = 0;
 385         unsigned long now = jiffies;
 386         long timediff;
 387         ENTRY;
 388
 389         /* We are currently not supporting multi-obj BRW_READ RPCS at all.
 390          * When we do this function's dentry cleanup will need to be fixed.
 391          * These values are verified in ost_brw_write() from the wire. */
 392         LASSERTF(objcount == 1, "%d\n", objcount);
 393         LASSERTF(obj->ioo_bufcnt > 0, "%d\n", obj->ioo_bufcnt);
 394
 395         if (oa->o_valid & OBD_MD_FLGRANT) {
 396                 spin_lock(&obd->obd_osfs_lock);
 397                 filter_grant_incoming(exp, oa);
 398
 399                 if (!(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
 418         obdo_to_inode(inode, oa, OBD_MD_FLATIME);
 419
 420         rc = filter_map_remote_to_local(objcount, obj, nb, pages, res);
 421         if (rc)
 422                 GOTO(cleanup, rc);
 423
 424         fsfilt_check_slow(obd, now, "preprw_read setup");
 425
 426         /* find pages for all segments, fill array with them */
 427         do_gettimeofday(&start);
 428         for (i = 0, lnb = res; i < *pages; i++, lnb++) {
 429
 430                 lnb->dentry = dentry;
 431
 432                 if (i_size_read(inode) <= lnb->offset)
 433                         /* If there's no more data, abort early.  lnb->rc == 0,
 434                          * so it's easy to detect later. */
 435                         break;
 436
 437                 lnb->page = filter_get_page(obd, inode, lnb->offset, 0);
 438                 if (lnb->page == NULL)
 439                         GOTO(cleanup, rc = -ENOMEM);
 440
 441                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS, 1);
 442
 443                 if (i_size_read(inode) < lnb->offset + lnb->len - 1)
 444                         lnb->rc = i_size_read(inode) - lnb->offset;
 445                 else
 446                         lnb->rc = lnb->len;
 447
 448                 tot_bytes += lnb->rc;
 449
 450                 if (PageUptodate(lnb->page)) {
 451                         lprocfs_counter_add(obd->obd_stats,
 452                                             LPROC_FILTER_CACHE_HIT, 1);
 453                         continue;
 454                 }
 455
 456                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_MISS, 1);
 457                 filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 458         }
 459         do_gettimeofday(&end);
 460         timediff = cfs_timeval_sub(&end, &start, NULL);
 461         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
 462
 463         if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
 464                 GOTO(cleanup, rc = -ENOMEM);
 465
 466         fsfilt_check_slow(obd, now, "start_page_read");
 467
 468         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf,
 469                               exp, NULL, NULL, NULL);
 470         if (rc)
 471                 GOTO(cleanup, rc);
 472
 473         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_READ_BYTES, tot_bytes);
 474         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
 475                 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
 476                                     LPROC_FILTER_READ_BYTES, tot_bytes);
 477
 478         EXIT;
 479
 480  cleanup:
 481         /* unlock pages to allow access from concurrent OST_READ */
 482         for (i = 0, lnb = res; i < *pages; i++, lnb++) {
 483                 if (lnb->page) {
 484                         LASSERT(PageLocked(lnb->page));
 485                         unlock_page(lnb->page);
 486
 487                         if (rc) {
 488                                 page_cache_release(lnb->page);
 489                                 lnb->page = NULL;
 490                         }
 491                 }
 492         }
 493
 494         if (rc != 0) {
 495                 if (dentry != NULL)
 496                         f_dput(dentry);
 497         }
 498
 499         filter_iobuf_put(&obd->u.filter, iobuf, oti);
 500
 501         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 502         if (rc)
 503                 CERROR("io error %d\n", rc);
 504
 505         return rc;
 506 }
 507
 508 /* When clients have dirtied as much space as they've been granted they
 509  * fall through to sync writes.  These sync writes haven't been expressed
 510  * in grants and need to error with ENOSPC when there isn't room in the
 511  * filesystem for them after grants are taken into account.  However,
 512  * writeback of the dirty data that was already granted space can write
 513  * right on through.
 514  *
 515  * Caller must hold obd_osfs_lock. */
 516 static int filter_grant_check(struct obd_export *exp, struct obdo *oa,
 517                               int objcount, struct fsfilt_objinfo *fso,
 518                               int niocount, struct niobuf_local *lnb,
 519                               obd_size *left, struct inode *inode)
 520 {
 521         struct filter_export_data *fed = &exp->exp_filter_data;
 522         int blocksize = exp->exp_obd->u.obt.obt_sb->s_blocksize;
 523         unsigned long used = 0, ungranted = 0, using;
 524         int i, rc = -ENOSPC, obj, n = 0;
 525
 526         LASSERT_SPIN_LOCKED(&exp->exp_obd->obd_osfs_lock);
 527
 528         for (obj = 0; obj < objcount; obj++) {
 529                 for (i = 0; i < fso[obj].fso_bufcnt; i++, n++) {
 530                         int tmp, bytes;
 531
 532                         /* should match the code in osc_exit_cache */
 533                         bytes = lnb[n].len;
 534                         bytes += lnb[n].offset & (blocksize - 1);
 535                         tmp = (lnb[n].offset + lnb[n].len) & (blocksize - 1);
 536                         if (tmp)
 537                                 bytes += blocksize - tmp;
 538
 539                         if ((lnb[n].flags & OBD_BRW_FROM_GRANT) &&
 540                             (oa->o_valid & OBD_MD_FLGRANT)) {
 541                                 if (fed->fed_grant < used + bytes) {
 542                                         CDEBUG(D_CACHE,
 543                                                "%s: cli %s/%p claims %ld+%d "
 544                                                "GRANT, real grant %lu idx %d\n",
 545                                                exp->exp_obd->obd_name,
 546                                                exp->exp_client_uuid.uuid, exp,
 547                                                used, bytes, fed->fed_grant, n);
 548                                 } else {
 549                                         used += bytes;
 550                                         lnb[n].flags |= OBD_BRW_GRANTED;
 551                                         lnb[n].lnb_grant_used = bytes;
 552                                         CDEBUG(0, "idx %d used=%lu\n", n, used);
 553                                         rc = 0;
 554                                         continue;
 555                                 }
 556                         }
 557                         if (*left > ungranted + bytes) {
 558                                 /* if enough space, pretend it was granted */
 559                                 ungranted += bytes;
 560                                 lnb[n].flags |= OBD_BRW_GRANTED;
 561                                 lnb[n].lnb_grant_used = bytes;
 562                                 CDEBUG(0, "idx %d ungranted=%lu\n",n,ungranted);
 563                                 rc = 0;
 564                                 continue;
 565                         }
 566
 567                         /* We can't check for already-mapped blocks here, as
 568                          * it requires dropping the osfs lock to do the bmap.
 569                          * Instead, we return ENOSPC and in that case we need
 570                          * to go through and verify if all of the blocks not
 571                          * marked BRW_GRANTED are already mapped and we can
 572                          * ignore this error. */
 573                         lnb[n].rc = -ENOSPC;
 574                         lnb[n].flags &= ~OBD_BRW_GRANTED;
 575                         CDEBUG(D_CACHE,"%s: cli %s/%p idx %d no space for %d\n",
 576                                exp->exp_obd->obd_name,
 577                                exp->exp_client_uuid.uuid, exp, n, bytes);
 578                 }
 579         }
 580
 581         /* Now substract what client have used already.  We don't subtract
 582          * this from the tot_granted yet, so that other client's can't grab
 583          * that space before we have actually allocated our blocks.  That
 584          * happens in filter_grant_commit() after the writes are done. */
 585         *left -= ungranted;
 586         fed->fed_grant -= used;
 587         fed->fed_pending += used + ungranted;
 588         exp->exp_obd->u.filter.fo_tot_granted += ungranted;
 589         exp->exp_obd->u.filter.fo_tot_pending += used + ungranted;
 590
 591         CDEBUG(D_CACHE,
 592                "%s: cli %s/%p used: %lu ungranted: %lu grant: %lu dirty: %lu\n",
 593                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp, used,
 594                ungranted, fed->fed_grant, fed->fed_dirty);
 595
 596         /* Rough calc in case we don't refresh cached statfs data */
 597         using = (used + ungranted + 1 ) >>
 598                 exp->exp_obd->u.obt.obt_sb->s_blocksize_bits;
 599         if (exp->exp_obd->obd_osfs.os_bavail > using)
 600                 exp->exp_obd->obd_osfs.os_bavail -= using;
 601         else
 602                 exp->exp_obd->obd_osfs.os_bavail = 0;
 603
 604         if (fed->fed_dirty < used) {
 605                 CERROR("%s: cli %s/%p claims used %lu > fed_dirty %lu\n",
 606                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 607                        used, fed->fed_dirty);
 608                 used = fed->fed_dirty;
 609         }
 610         exp->exp_obd->u.filter.fo_tot_dirty -= used;
 611         fed->fed_dirty -= used;
 612
 613         if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
 614                 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
 615                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 616                        fed->fed_dirty, fed->fed_pending, fed->fed_grant);
 617                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
 618                 LBUG();
 619         }
 620         return rc;
 621 }
 622
 623 /* If we ever start to support multi-object BRW RPCs, we will need to get locks
 624  * on mulitple inodes.  That isn't all, because there still exists the
 625  * possibility of a truncate starting a new transaction while holding the ext3
 626  * rwsem = write while some writes (which have started their transactions here)
 627  * blocking on the ext3 rwsem = read => lock inversion.
 628  *
 629  * The handling gets very ugly when dealing with locked pages.  It may be easier
 630  * to just get rid of the locked page code (which has problems of its own) and
 631  * either discover we do not need it anymore (i.e. it was a symptom of another
 632  * bug) or ensure we get the page locks in an appropriate order. */
 633 static int filter_preprw_write(int cmd, struct obd_export *exp, struct obdo *oa,
 634                                int objcount, struct obd_ioobj *obj,
 635                                struct niobuf_remote *nb, int *pages,
 636                                struct niobuf_local *res,
 637                                struct obd_trans_info *oti)
 638 {
 639         struct obd_device *obd = exp->exp_obd;
 640         struct timeval start, end;
 641         struct lvfs_run_ctxt saved;
 642         struct niobuf_local *lnb = res;
 643         struct fsfilt_objinfo fso;
 644         struct filter_mod_data *fmd;
 645         struct dentry *dentry = NULL;
 646         void *iobuf;
 647         obd_size left;
 648         unsigned long now = jiffies, timediff;
 649         int rc = 0, i, tot_bytes = 0, cleanup_phase = 0, localreq = 0;
 650         ENTRY;
 651         LASSERT(objcount == 1);
 652         LASSERT(obj->ioo_bufcnt > 0);
 653
 654         if (exp->exp_connection->c_peer.nid == exp->exp_connection->c_self)
 655                 localreq = 1;
 656
 657         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 658         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 659         if (IS_ERR(iobuf))
 660                 GOTO(cleanup, rc = PTR_ERR(iobuf));
 661         cleanup_phase = 1;
 662
 663         dentry = filter_fid2dentry(obd, NULL, obj->ioo_gr, obj->ioo_id);
 664         if (IS_ERR(dentry))
 665                 GOTO(cleanup, rc = PTR_ERR(dentry));
 666         cleanup_phase = 2;
 667
 668         if (dentry->d_inode == NULL) {
 669                 if (exp->exp_obd->obd_recovering) {
 670                         struct obdo *noa = oa;
 671
 672                         if (oa == NULL) {
 673                                 OBDO_ALLOC(noa);
 674                                 if (noa == NULL)
 675                                         GOTO(recreate_out, rc = -ENOMEM);
 676                                 noa->o_id = obj->ioo_id;
 677                                 noa->o_valid = OBD_MD_FLID;
 678                         }
 679
 680                         if (filter_recreate(exp->exp_obd, noa) == 0) {
 681                                 f_dput(dentry);
 682                                 dentry = filter_fid2dentry(exp->exp_obd, NULL,
 683                                                            obj->ioo_gr,
 684                                                            obj->ioo_id);
 685                         }
 686                         if (oa == NULL)
 687                                 OBDO_FREE(noa);
 688                 }
 689     recreate_out:
 690                 if (IS_ERR(dentry) || dentry->d_inode == NULL) {
 691                         CERROR("%s: BRW to missing obj "LPU64"/"LPU64":rc %d\n",
 692                                exp->exp_obd->obd_name,
 693                                obj->ioo_id, obj->ioo_gr,
 694                                IS_ERR(dentry) ? (int)PTR_ERR(dentry) : -ENOENT);
 695                         if (IS_ERR(dentry))
 696                                 cleanup_phase = 1;
 697                         GOTO(cleanup, rc = -ENOENT);
 698                 }
 699         }
 700
 701         rc = filter_map_remote_to_local(objcount, obj, nb, pages, res);
 702         if (rc)
 703                 GOTO(cleanup, rc);
 704
 705         fsfilt_check_slow(obd, now, "preprw_write setup");
 706
 707         /* Don't update inode timestamps if this write is older than a
 708          * setattr which modifies the timestamps. b=10150 */
 709         /* XXX when we start having persistent reservations this needs to
 710          * be changed to filter_fmd_get() to create the fmd if it doesn't
 711          * already exist so we can store the reservation handle there. */
 712         fmd = filter_fmd_find(exp, obj->ioo_id, obj->ioo_gr);
 713
 714         LASSERT(oa != NULL);
 715         spin_lock(&obd->obd_osfs_lock);
 716
 717         filter_grant_incoming(exp, oa);
 718         if (fmd && fmd->fmd_mactime_xid > oti->oti_xid)
 719                 oa->o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 720                                  OBD_MD_FLATIME);
 721         else
 722                 obdo_to_inode(dentry->d_inode, oa, OBD_MD_FLATIME |
 723                               OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 724         cleanup_phase = 3;
 725
 726         left = filter_grant_space_left(exp);
 727
 728         fso.fso_dentry = dentry;
 729         fso.fso_bufcnt = *pages;
 730
 731         rc = filter_grant_check(exp, oa, objcount, &fso, *pages, res,
 732                                 &left, dentry->d_inode);
 733
 734         /* do not zero out oa->o_valid as it is used in filter_commitrw_write()
 735          * for setting UID/GID and fid EA in first write time. */
 736         if (oa->o_valid & OBD_MD_FLGRANT)
 737                 oa->o_grant = filter_grant(exp,oa->o_grant,oa->o_undirty,left);
 738
 739         spin_unlock(&obd->obd_osfs_lock);
 740         filter_fmd_put(exp, fmd);
 741
 742         if (rc)
 743                 GOTO(cleanup, rc);
 744         cleanup_phase = 4;
 745
 746         /* Filter truncate first locks i_mutex then partally truncated
 747          * page, filter write code first locks pages then take
 748          * i_mutex.  To avoid a deadlock in case of concurrent
 749          * punch/write requests from one client, filter writes and
 750          * filter truncates are serialized by i_alloc_sem, allowing
 751          * multiple writes or single truncate. */
 752         down_read(&dentry->d_inode->i_alloc_sem);
 753
 754         do_gettimeofday(&start);
 755         for (i = 0, lnb = res; i < *pages; i++, lnb++) {
 756
 757                 /* We still set up for ungranted pages so that granted pages
 758                  * can be written to disk as they were promised, and portals
 759                  * needs to keep the pages all aligned properly. */
 760                 lnb->dentry = dentry;
 761
 762                 lnb->page = filter_get_page(obd, dentry->d_inode, lnb->offset,
 763                                             localreq);
 764                 if (lnb->page == NULL)
 765                         GOTO(cleanup, rc = -ENOMEM);
 766
 767                 /* DLM locking protects us from write and truncate competing
 768                  * for same region, but truncate can leave dirty page in the
 769                  * cache. it's possible the writeout on a such a page is in
 770                  * progress when we access it. it's also possible that during
 771                  * this writeout we put new (partial) data, but then won't
 772                  * be able to proceed in filter_commitrw_write(). thus let's
 773                  * just wait for writeout completion, should be rare enough.
 774                  * -bzzz */
 775                 if (obd->u.filter.fo_writethrough_cache)
 776                         wait_on_page_writeback(lnb->page);
 777                 BUG_ON(PageWriteback(lnb->page));
 778
 779                 /* If the filter writes a partial page, then has the file
 780                  * extended, the client will read in the whole page.  the
 781                  * filter has to be careful to zero the rest of the partial
 782                  * page on disk.  we do it by hand for partial extending
 783                  * writes, send_bio() is responsible for zeroing pages when
 784                  * asked to read unmapped blocks -- brw_kiovec() does this. */
 785                 if (lnb->len != CFS_PAGE_SIZE) {
 786                         __s64 maxidx;
 787
 788                         maxidx = ((i_size_read(dentry->d_inode) +
 789                                    CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT) - 1;
 790                         if (maxidx >= lnb->page->index) {
 791                                 LL_CDEBUG_PAGE(D_PAGE, lnb->page, "write %u @ "
 792                                                LPU64" flg %x before EOF %llu\n",
 793                                                lnb->len, lnb->offset,lnb->flags,
 794                                                i_size_read(dentry->d_inode));
 795                                 filter_iobuf_add_page(obd, iobuf,
 796                                                       dentry->d_inode,
 797                                                       lnb->page);
 798                         } else {
 799                                 long off;
 800                                 char *p = kmap(lnb->page);
 801
 802                                 off = lnb->offset & ~CFS_PAGE_MASK;
 803                                 if (off)
 804                                         memset(p, 0, off);
 805                                 off = (lnb->offset + lnb->len) & ~CFS_PAGE_MASK;
 806                                 if (off)
 807                                         memset(p + off, 0, CFS_PAGE_SIZE - off);
 808                                 kunmap(lnb->page);
 809                         }
 810                 }
 811                 if (lnb->rc == 0)
 812                         tot_bytes += lnb->len;
 813         }
 814         do_gettimeofday(&end);
 815         timediff = cfs_timeval_sub(&end, &start, NULL);
 816         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
 817
 818         if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
 819                 GOTO(cleanup, rc = -ENOMEM);
 820
 821         /* don't unlock pages to prevent any access */
 822         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf, exp,
 823                               NULL, NULL, NULL);
 824
 825         fsfilt_check_slow(obd, now, "start_page_write");
 826
 827         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
 828                             tot_bytes);
 829         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
 830                 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
 831                                     LPROC_FILTER_WRITE_BYTES, tot_bytes);
 832         EXIT;
 833 cleanup:
 834         switch(cleanup_phase) {
 835         case 4:
 836                 if (rc) {
 837                         for (i = 0, lnb = res; i < *pages; i++, lnb++) {
 838                                 if (lnb->page != NULL) {
 839                                         unlock_page(lnb->page);
 840                                         page_cache_release(lnb->page);
 841                                         lnb->page = NULL;
 842                                 }
 843                         }
 844                         filter_grant_commit(exp, *pages, res);
 845                         up_read(&dentry->d_inode->i_alloc_sem);
 846                 }
 847         case 3:
 848                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 849         case 2:
 850                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 851                 if (rc)
 852                         f_dput(dentry);
 853                 break;
 854         case 1:
 855                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 856         case 0:
 857                 spin_lock(&obd->obd_osfs_lock);
 858                 if (oa)
 859                         filter_grant_incoming(exp, oa);
 860                 spin_unlock(&obd->obd_osfs_lock);
 861                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 862                 break;
 863         default:;
 864         }
 865         return rc;
 866 }
 867
 868 int filter_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
 869                   int objcount, struct obd_ioobj *obj,
 870                   struct niobuf_remote *nb, int *pages,
 871                   struct niobuf_local *res, struct obd_trans_info *oti)
 872 {
 873         if (cmd == OBD_BRW_WRITE)
 874                 return filter_preprw_write(cmd, exp, oa, objcount, obj,
 875                                            nb, pages, res, oti);
 876         if (cmd == OBD_BRW_READ)
 877                 return filter_preprw_read(cmd, exp, oa, objcount, obj,
 878                                           nb, pages, res, oti);
 879         LBUG();
 880         return -EPROTO;
 881 }
 882
 883 static int filter_commitrw_read(struct obd_export *exp, struct obdo *oa,
 884                                 int objcount, struct obd_ioobj *obj,
 885                                 struct niobuf_remote *rnb,
 886                                 int pages, struct niobuf_local *res,
 887                                 struct obd_trans_info *oti, int rc)
 888 {
 889         struct filter_obd *fo = &exp->exp_obd->u.filter;
 890         struct inode *inode = NULL;
 891         struct ldlm_res_id res_id = { .name = { obj->ioo_id } };
 892         struct ldlm_resource *resource = NULL;
 893         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 894         struct niobuf_local *lnb;
 895         int i;
 896         ENTRY;
 897
 898         /* If oa != NULL then filter_preprw_read updated the inode atime
 899          * and we should update the lvb so that other glimpses will also
 900          * get the updated value. bug 5972 */
 901         if (oa && ns && ns->ns_lvbo && ns->ns_lvbo->lvbo_update) {
 902                 resource = ldlm_resource_get(ns, NULL, res_id, LDLM_EXTENT, 0);
 903
 904                 if (resource != NULL) {
 905                         ns->ns_lvbo->lvbo_update(resource, NULL, 0, 1);
 906                         ldlm_resource_putref(resource);
 907                 }
 908         }
 909
 910         if (res->dentry != NULL)
 911                 inode = res->dentry->d_inode;
 912
 913         for (i = 0, lnb = res; i < pages; i++, lnb++) {
 914                 if (lnb->page != NULL) {
 915                         page_cache_release(lnb->page);
 916                         lnb->page = NULL;
 917                 }
 918         }
 919
 920         if (inode && (fo->fo_read_cache == 0 ||
 921                       i_size_read(inode) > fo->fo_readcache_max_filesize))
 922                 filter_release_cache(exp->exp_obd, obj, rnb, inode);
 923
 924         if (res->dentry != NULL)
 925                 f_dput(res->dentry);
 926         RETURN(rc);
 927 }
 928
 929 void filter_grant_commit(struct obd_export *exp, int niocount,
 930                          struct niobuf_local *res)
 931 {
 932         struct filter_obd *filter = &exp->exp_obd->u.filter;
 933         struct niobuf_local *lnb = res;
 934         unsigned long pending = 0;
 935         int i;
 936
 937         spin_lock(&exp->exp_obd->obd_osfs_lock);
 938         for (i = 0, lnb = res; i < niocount; i++, lnb++)
 939                 pending += lnb->lnb_grant_used;
 940
 941         LASSERTF(exp->exp_filter_data.fed_pending >= pending,
 942                  "%s: cli %s/%p fed_pending: %lu grant_used: %lu\n",
 943                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 944                  exp->exp_filter_data.fed_pending, pending);
 945         exp->exp_filter_data.fed_pending -= pending;
 946         LASSERTF(filter->fo_tot_granted >= pending,
 947                  "%s: cli %s/%p tot_granted: "LPU64" grant_used: %lu\n",
 948                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 949                  exp->exp_obd->u.filter.fo_tot_granted, pending);
 950         filter->fo_tot_granted -= pending;
 951         LASSERTF(filter->fo_tot_pending >= pending,
 952                  "%s: cli %s/%p tot_pending: "LPU64" grant_used: %lu\n",
 953                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 954                  filter->fo_tot_pending, pending);
 955         filter->fo_tot_pending -= pending;
 956
 957         spin_unlock(&exp->exp_obd->obd_osfs_lock);
 958 }
 959
 960 int filter_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
 961                     int objcount, struct obd_ioobj *obj,
 962                     struct niobuf_remote *nb, int pages,
 963                     struct niobuf_local *res, struct obd_trans_info *oti,
 964                     int rc)
 965 {
 966         if (cmd == OBD_BRW_WRITE)
 967                 return filter_commitrw_write(exp, oa, objcount, obj,
 968                                              nb, pages, res, oti, rc);
 969         if (cmd == OBD_BRW_READ)
 970                 return filter_commitrw_read(exp, oa, objcount, obj,
 971                                             nb, pages, res, oti, rc);
 972         LBUG();
 973         return -EPROTO;
 974 }
 975
 976 int filter_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
 977                obd_count oa_bufs, struct brw_page *pga,
 978                struct obd_trans_info *oti)
 979 {
 980         struct obd_ioobj ioo;
 981         struct niobuf_local *lnb;
 982         struct niobuf_remote *rnb;
 983         obd_count i;
 984         int ret = 0, npages;
 985         ENTRY;
 986
 987         OBD_ALLOC(lnb, oa_bufs * sizeof(struct niobuf_local));
 988         OBD_ALLOC(rnb, oa_bufs * sizeof(struct niobuf_remote));
 989
 990         if (lnb == NULL || rnb == NULL)
 991                 GOTO(out, ret = -ENOMEM);
 992
 993         for (i = 0; i < oa_bufs; i++) {
 994                 lnb[i].page = pga[i].pg;
 995                 rnb[i].offset = pga[i].off;
 996                 rnb[i].len = pga[i].count;
 997         }
 998
 999         obdo_to_ioobj(oinfo->oi_oa, &ioo);
1000         ioo.ioo_bufcnt = oa_bufs;
1001
1002         npages = oa_bufs;
1003         ret = filter_preprw(cmd, exp, oinfo->oi_oa, 1, &ioo,
1004                             rnb, &npages, lnb, oti);
1005         if (ret != 0)
1006                 GOTO(out, ret);
1007         LASSERTF(oa_bufs == npages, "%u != %u\n", oa_bufs, npages);
1008
1009         ret = filter_commitrw(cmd, exp, oinfo->oi_oa, 1, &ioo, rnb,
1010                               npages, lnb, oti, ret);
1011
1012 out:
1013         if (lnb)
1014                 OBD_FREE(lnb, oa_bufs * sizeof(struct niobuf_local));
1015         if (rnb)
1016                 OBD_FREE(rnb, oa_bufs * sizeof(struct niobuf_remote));
1017         RETURN(ret);
1018 }