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 not allowing to enter FS as the client can run on this node
 267  * and we might end waiting on a page he sent in the request we're serving.
 268  *
 269  * use __GFP_HIGHMEM so that the pages can use all of the available memory
 270  * on 32-bit machines
 271  *
 272  */
 273 static struct page * filter_get_page(struct obd_device *obd,
 274                                      struct inode *inode,
 275                                      obd_off offset)
 276 {
 277         struct page *page;
 278
 279         page = find_or_create_page(inode->i_mapping, offset >> CFS_PAGE_SHIFT,
 280                                    GFP_NOFS | __GFP_HIGHMEM);
 281         if (unlikely(page == NULL))
 282                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_NO_PAGE, 1);
 283
 284         return page;
 285 }
 286
 287 /*
 288  * the routine initializes array of local_niobuf from remote_niobuf
 289  */
 290 static int filter_map_remote_to_local(int objcount, struct obd_ioobj *obj,
 291                                       struct niobuf_remote *nb,
 292                                       int *nrpages, struct niobuf_local *res)
 293 {
 294         struct niobuf_remote *rnb;
 295         struct niobuf_local *lnb;
 296         int i, max;
 297         ENTRY;
 298
 299         /* we don't support multiobject RPC yet
 300          * ost_brw_read() and ost_brw_write() check this */
 301         LASSERT(objcount == 1);
 302
 303         max = *nrpages;
 304         *nrpages = 0;
 305         for (i = 0, rnb = nb, lnb = res; i < obj->ioo_bufcnt; i++, rnb++) {
 306                 obd_off offset = rnb->offset;
 307                 unsigned int len = rnb->len;
 308
 309                 while (len > 0) {
 310                         int poff = offset & (CFS_PAGE_SIZE - 1);
 311                         int plen = CFS_PAGE_SIZE - poff;
 312
 313                         if (*nrpages >= max) {
 314                                 CERROR("small array of local bufs: %d\n", max);
 315                                 RETURN(-EINVAL);
 316                         }
 317
 318                         if (plen > len)
 319                                 plen = len;
 320                         lnb->offset = offset;
 321                         lnb->len = plen;
 322                         lnb->flags = rnb->flags;
 323                         lnb->page = NULL;
 324                         lnb->rc = 0;
 325                         lnb->lnb_grant_used = 0;
 326
 327                         LASSERTF(plen <= len, "plen %u, len %u\n", plen, len);
 328                         offset += plen;
 329                         len -= plen;
 330                         lnb++;
 331                         (*nrpages)++;
 332                 }
 333         }
 334         RETURN(0);
 335 }
 336
 337 /*
 338  * the function is used to free all pages used for request
 339  * just to mimic cacheless OSS which don't occupy much memory
 340  */
 341 void filter_invalidate_cache(struct obd_device *obd, struct obd_ioobj *obj,
 342                              struct niobuf_remote *nb, struct inode *inode)
 343 {
 344         struct niobuf_remote *rnb;
 345         int i;
 346
 347         LASSERT(inode != NULL);
 348
 349         for (i = 0, rnb = nb; i < obj->ioo_bufcnt; i++, rnb++) {
 350                 invalidate_mapping_pages(inode->i_mapping,
 351                                          rnb->offset >> CFS_PAGE_SHIFT,
 352                                          (rnb->offset + rnb->len) >>
 353                                          CFS_PAGE_SHIFT);
 354         }
 355
 356 }
 357
 358 static int filter_preprw_read(int cmd, struct obd_export *exp, struct obdo *oa,
 359                               int objcount, struct obd_ioobj *obj,
 360                               struct niobuf_remote *nb,
 361                               int *pages, struct niobuf_local *res,
 362                               struct obd_trans_info *oti)
 363 {
 364         struct obd_device *obd = exp->exp_obd;
 365         struct timeval start, end;
 366         struct lvfs_run_ctxt saved;
 367         struct niobuf_local *lnb;
 368         struct dentry *dentry = NULL;
 369         struct inode *inode = NULL;
 370         void *iobuf = NULL;
 371         int rc = 0, i, tot_bytes = 0;
 372         unsigned long now = jiffies;
 373         long timediff;
 374         ENTRY;
 375
 376         /* We are currently not supporting multi-obj BRW_READ RPCS at all.
 377          * When we do this function's dentry cleanup will need to be fixed.
 378          * These values are verified in ost_brw_write() from the wire. */
 379         LASSERTF(objcount == 1, "%d\n", objcount);
 380         LASSERTF(obj->ioo_bufcnt > 0, "%d\n", obj->ioo_bufcnt);
 381
 382         if (oa->o_valid & OBD_MD_FLGRANT) {
 383                 spin_lock(&obd->obd_osfs_lock);
 384                 filter_grant_incoming(exp, oa);
 385
 386                 if (!(oa->o_flags & OBD_FL_SHRINK_GRANT))
 387                         oa->o_grant = 0;
 388                 spin_unlock(&obd->obd_osfs_lock);
 389         }
 390
 391         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 392         if (IS_ERR(iobuf))
 393                 RETURN(PTR_ERR(iobuf));
 394
 395         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 396         dentry = filter_oa2dentry(obd, oa);
 397         if (IS_ERR(dentry)) {
 398                 rc = PTR_ERR(dentry);
 399                 dentry = NULL;
 400                 GOTO(cleanup, rc);
 401         }
 402
 403         inode = dentry->d_inode;
 404
 405         obdo_to_inode(inode, oa, OBD_MD_FLATIME);
 406
 407         rc = filter_map_remote_to_local(objcount, obj, nb, pages, res);
 408         if (rc)
 409                 GOTO(cleanup, rc);
 410
 411         fsfilt_check_slow(obd, now, "preprw_read setup");
 412
 413         /* find pages for all segments, fill array with them */
 414         do_gettimeofday(&start);
 415         for (i = 0, lnb = res; i < *pages; i++, lnb++) {
 416
 417                 lnb->dentry = dentry;
 418
 419                 if (i_size_read(inode) <= lnb->offset)
 420                         /* If there's no more data, abort early.  lnb->rc == 0,
 421                          * so it's easy to detect later. */
 422                         break;
 423
 424                 lnb->page = filter_get_page(obd, inode, lnb->offset);
 425                 if (lnb->page == NULL)
 426                         GOTO(cleanup, rc = -ENOMEM);
 427
 428                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS, 1);
 429
 430                 if (i_size_read(inode) < lnb->offset + lnb->len - 1)
 431                         lnb->rc = i_size_read(inode) - lnb->offset;
 432                 else
 433                         lnb->rc = lnb->len;
 434
 435                 tot_bytes += lnb->rc;
 436
 437                 if (PageUptodate(lnb->page)) {
 438                         lprocfs_counter_add(obd->obd_stats,
 439                                             LPROC_FILTER_CACHE_HIT, 1);
 440                         continue;
 441                 }
 442
 443                 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_MISS, 1);
 444                 filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 445         }
 446         do_gettimeofday(&end);
 447         timediff = cfs_timeval_sub(&end, &start, NULL);
 448         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
 449
 450         if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
 451                 GOTO(cleanup, rc = -ENOMEM);
 452
 453         fsfilt_check_slow(obd, now, "start_page_read");
 454
 455         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf,
 456                               exp, NULL, NULL, NULL);
 457         if (rc)
 458                 GOTO(cleanup, rc);
 459
 460         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_READ_BYTES, tot_bytes);
 461         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
 462                 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
 463                                     LPROC_FILTER_READ_BYTES, tot_bytes);
 464
 465         EXIT;
 466
 467  cleanup:
 468         /* unlock pages to allow access from concurrent OST_READ */
 469         for (i = 0, lnb = res; i < *pages; i++, lnb++) {
 470                 if (lnb->page) {
 471                         LASSERT(PageLocked(lnb->page));
 472                         unlock_page(lnb->page);
 473
 474                         if (rc) {
 475                                 page_cache_release(lnb->page);
 476                                 lnb->page = NULL;
 477                         }
 478                 }
 479         }
 480
 481         if (rc != 0) {
 482                 if (dentry != NULL)
 483                         f_dput(dentry);
 484         }
 485
 486         filter_iobuf_put(&obd->u.filter, iobuf, oti);
 487
 488         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 489         if (rc)
 490                 CERROR("io error %d\n", rc);
 491
 492         return rc;
 493 }
 494
 495 /* When clients have dirtied as much space as they've been granted they
 496  * fall through to sync writes.  These sync writes haven't been expressed
 497  * in grants and need to error with ENOSPC when there isn't room in the
 498  * filesystem for them after grants are taken into account.  However,
 499  * writeback of the dirty data that was already granted space can write
 500  * right on through.
 501  *
 502  * Caller must hold obd_osfs_lock. */
 503 static int filter_grant_check(struct obd_export *exp, struct obdo *oa,
 504                               int objcount, struct fsfilt_objinfo *fso,
 505                               int niocount, struct niobuf_local *lnb,
 506                               obd_size *left, struct inode *inode)
 507 {
 508         struct filter_export_data *fed = &exp->exp_filter_data;
 509         int blocksize = exp->exp_obd->u.obt.obt_sb->s_blocksize;
 510         unsigned long used = 0, ungranted = 0, using;
 511         int i, rc = -ENOSPC, obj, n = 0;
 512
 513         LASSERT_SPIN_LOCKED(&exp->exp_obd->obd_osfs_lock);
 514
 515         for (obj = 0; obj < objcount; obj++) {
 516                 for (i = 0; i < fso[obj].fso_bufcnt; i++, n++) {
 517                         int tmp, bytes;
 518
 519                         /* should match the code in osc_exit_cache */
 520                         bytes = lnb[n].len;
 521                         bytes += lnb[n].offset & (blocksize - 1);
 522                         tmp = (lnb[n].offset + lnb[n].len) & (blocksize - 1);
 523                         if (tmp)
 524                                 bytes += blocksize - tmp;
 525
 526                         if ((lnb[n].flags & OBD_BRW_FROM_GRANT) &&
 527                             (oa->o_valid & OBD_MD_FLGRANT)) {
 528                                 if (fed->fed_grant < used + bytes) {
 529                                         CDEBUG(D_CACHE,
 530                                                "%s: cli %s/%p claims %ld+%d "
 531                                                "GRANT, real grant %lu idx %d\n",
 532                                                exp->exp_obd->obd_name,
 533                                                exp->exp_client_uuid.uuid, exp,
 534                                                used, bytes, fed->fed_grant, n);
 535                                 } else {
 536                                         used += bytes;
 537                                         lnb[n].flags |= OBD_BRW_GRANTED;
 538                                         lnb[n].lnb_grant_used = bytes;
 539                                         CDEBUG(0, "idx %d used=%lu\n", n, used);
 540                                         rc = 0;
 541                                         continue;
 542                                 }
 543                         }
 544                         if (*left > ungranted + bytes) {
 545                                 /* if enough space, pretend it was granted */
 546                                 ungranted += bytes;
 547                                 lnb[n].flags |= OBD_BRW_GRANTED;
 548                                 lnb[n].lnb_grant_used = bytes;
 549                                 CDEBUG(0, "idx %d ungranted=%lu\n",n,ungranted);
 550                                 rc = 0;
 551                                 continue;
 552                         }
 553
 554                         /* We can't check for already-mapped blocks here, as
 555                          * it requires dropping the osfs lock to do the bmap.
 556                          * Instead, we return ENOSPC and in that case we need
 557                          * to go through and verify if all of the blocks not
 558                          * marked BRW_GRANTED are already mapped and we can
 559                          * ignore this error. */
 560                         lnb[n].rc = -ENOSPC;
 561                         lnb[n].flags &= ~OBD_BRW_GRANTED;
 562                         CDEBUG(D_CACHE,"%s: cli %s/%p idx %d no space for %d\n",
 563                                exp->exp_obd->obd_name,
 564                                exp->exp_client_uuid.uuid, exp, n, bytes);
 565                 }
 566         }
 567
 568         /* Now substract what client have used already.  We don't subtract
 569          * this from the tot_granted yet, so that other client's can't grab
 570          * that space before we have actually allocated our blocks.  That
 571          * happens in filter_grant_commit() after the writes are done. */
 572         *left -= ungranted;
 573         fed->fed_grant -= used;
 574         fed->fed_pending += used + ungranted;
 575         exp->exp_obd->u.filter.fo_tot_granted += ungranted;
 576         exp->exp_obd->u.filter.fo_tot_pending += used + ungranted;
 577
 578         CDEBUG(D_CACHE,
 579                "%s: cli %s/%p used: %lu ungranted: %lu grant: %lu dirty: %lu\n",
 580                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp, used,
 581                ungranted, fed->fed_grant, fed->fed_dirty);
 582
 583         /* Rough calc in case we don't refresh cached statfs data */
 584         using = (used + ungranted + 1 ) >>
 585                 exp->exp_obd->u.obt.obt_sb->s_blocksize_bits;
 586         if (exp->exp_obd->obd_osfs.os_bavail > using)
 587                 exp->exp_obd->obd_osfs.os_bavail -= using;
 588         else
 589                 exp->exp_obd->obd_osfs.os_bavail = 0;
 590
 591         if (fed->fed_dirty < used) {
 592                 CERROR("%s: cli %s/%p claims used %lu > fed_dirty %lu\n",
 593                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 594                        used, fed->fed_dirty);
 595                 used = fed->fed_dirty;
 596         }
 597         exp->exp_obd->u.filter.fo_tot_dirty -= used;
 598         fed->fed_dirty -= used;
 599
 600         if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
 601                 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
 602                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 603                        fed->fed_dirty, fed->fed_pending, fed->fed_grant);
 604                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
 605                 LBUG();
 606         }
 607         return rc;
 608 }
 609
 610 /* If we ever start to support multi-object BRW RPCs, we will need to get locks
 611  * on mulitple inodes.  That isn't all, because there still exists the
 612  * possibility of a truncate starting a new transaction while holding the ext3
 613  * rwsem = write while some writes (which have started their transactions here)
 614  * blocking on the ext3 rwsem = read => lock inversion.
 615  *
 616  * The handling gets very ugly when dealing with locked pages.  It may be easier
 617  * to just get rid of the locked page code (which has problems of its own) and
 618  * either discover we do not need it anymore (i.e. it was a symptom of another
 619  * bug) or ensure we get the page locks in an appropriate order. */
 620 static int filter_preprw_write(int cmd, struct obd_export *exp, struct obdo *oa,
 621                                int objcount, struct obd_ioobj *obj,
 622                                struct niobuf_remote *nb, int *pages,
 623                                struct niobuf_local *res,
 624                                struct obd_trans_info *oti)
 625 {
 626         struct obd_device *obd = exp->exp_obd;
 627         struct timeval start, end;
 628         struct lvfs_run_ctxt saved;
 629         struct niobuf_local *lnb = res;
 630         struct fsfilt_objinfo fso;
 631         struct filter_mod_data *fmd;
 632         struct dentry *dentry = NULL;
 633         void *iobuf;
 634         obd_size left;
 635         unsigned long now = jiffies, timediff;
 636         int rc = 0, i, tot_bytes = 0, cleanup_phase = 0;
 637         ENTRY;
 638         LASSERT(objcount == 1);
 639         LASSERT(obj->ioo_bufcnt > 0);
 640
 641         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 642         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 643         if (IS_ERR(iobuf))
 644                 GOTO(cleanup, rc = PTR_ERR(iobuf));
 645         cleanup_phase = 1;
 646
 647         dentry = filter_fid2dentry(obd, NULL, obj->ioo_gr, obj->ioo_id);
 648         if (IS_ERR(dentry))
 649                 GOTO(cleanup, rc = PTR_ERR(dentry));
 650         cleanup_phase = 2;
 651
 652         if (dentry->d_inode == NULL) {
 653                 if (exp->exp_obd->obd_recovering) {
 654                         struct obdo *noa = oa;
 655
 656                         if (oa == NULL) {
 657                                 OBDO_ALLOC(noa);
 658                                 if (noa == NULL)
 659                                         GOTO(recreate_out, rc = -ENOMEM);
 660                                 noa->o_id = obj->ioo_id;
 661                                 noa->o_valid = OBD_MD_FLID;
 662                         }
 663
 664                         if (filter_recreate(exp->exp_obd, noa) == 0) {
 665                                 f_dput(dentry);
 666                                 dentry = filter_fid2dentry(exp->exp_obd, NULL,
 667                                                            obj->ioo_gr,
 668                                                            obj->ioo_id);
 669                         }
 670                         if (oa == NULL)
 671                                 OBDO_FREE(noa);
 672                 }
 673     recreate_out:
 674                 if (IS_ERR(dentry) || dentry->d_inode == NULL) {
 675                         CERROR("%s: BRW to missing obj "LPU64"/"LPU64":rc %d\n",
 676                                exp->exp_obd->obd_name,
 677                                obj->ioo_id, obj->ioo_gr,
 678                                IS_ERR(dentry) ? (int)PTR_ERR(dentry) : -ENOENT);
 679                         if (IS_ERR(dentry))
 680                                 cleanup_phase = 1;
 681                         GOTO(cleanup, rc = -ENOENT);
 682                 }
 683         }
 684
 685         rc = filter_map_remote_to_local(objcount, obj, nb, pages, res);
 686         if (rc)
 687                 GOTO(cleanup, rc);
 688
 689         fsfilt_check_slow(obd, now, "preprw_write setup");
 690
 691         /* Don't update inode timestamps if this write is older than a
 692          * setattr which modifies the timestamps. b=10150 */
 693         /* XXX when we start having persistent reservations this needs to
 694          * be changed to filter_fmd_get() to create the fmd if it doesn't
 695          * already exist so we can store the reservation handle there. */
 696         fmd = filter_fmd_find(exp, obj->ioo_id, obj->ioo_gr);
 697
 698         LASSERT(oa != NULL);
 699         spin_lock(&obd->obd_osfs_lock);
 700
 701         filter_grant_incoming(exp, oa);
 702         if (fmd && fmd->fmd_mactime_xid > oti->oti_xid)
 703                 oa->o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 704                                  OBD_MD_FLATIME);
 705         else
 706                 obdo_to_inode(dentry->d_inode, oa, OBD_MD_FLATIME |
 707                               OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 708         cleanup_phase = 3;
 709
 710         left = filter_grant_space_left(exp);
 711
 712         fso.fso_dentry = dentry;
 713         fso.fso_bufcnt = *pages;
 714
 715         rc = filter_grant_check(exp, oa, objcount, &fso, *pages, res,
 716                                 &left, dentry->d_inode);
 717
 718         /* do not zero out oa->o_valid as it is used in filter_commitrw_write()
 719          * for setting UID/GID and fid EA in first write time. */
 720         if (oa->o_valid & OBD_MD_FLGRANT)
 721                 oa->o_grant = filter_grant(exp,oa->o_grant,oa->o_undirty,left);
 722
 723         spin_unlock(&obd->obd_osfs_lock);
 724         filter_fmd_put(exp, fmd);
 725
 726         if (rc)
 727                 GOTO(cleanup, rc);
 728         cleanup_phase = 4;
 729
 730         do_gettimeofday(&start);
 731         for (i = 0, lnb = res; i < *pages; i++, lnb++) {
 732
 733                 /* We still set up for ungranted pages so that granted pages
 734                  * can be written to disk as they were promised, and portals
 735                  * needs to keep the pages all aligned properly. */
 736                 lnb->dentry = dentry;
 737
 738                 lnb->page = filter_get_page(obd, dentry->d_inode, lnb->offset);
 739                 if (lnb->page == NULL)
 740                         GOTO(cleanup, rc = -ENOMEM);
 741
 742                 /* DLM locking protects us from write and truncate competing
 743                  * for same region, but truncate can leave dirty page in the
 744                  * cache. it's possible the writeout on a such a page is in
 745                  * progress when we access it. it's also possible that during
 746                  * this writeout we put new (partial) data, but then won't
 747                  * be able to proceed in filter_commitrw_write(). thus let's
 748                  * just wait for writeout completion, should be rare enough.
 749                  * -bzzz */
 750                 if (obd->u.filter.fo_writethrough_cache)
 751                         wait_on_page_writeback(lnb->page);
 752                 BUG_ON(PageWriteback(lnb->page));
 753
 754                 /* If the filter writes a partial page, then has the file
 755                  * extended, the client will read in the whole page.  the
 756                  * filter has to be careful to zero the rest of the partial
 757                  * page on disk.  we do it by hand for partial extending
 758                  * writes, send_bio() is responsible for zeroing pages when
 759                  * asked to read unmapped blocks -- brw_kiovec() does this. */
 760                 if (lnb->len != CFS_PAGE_SIZE) {
 761                         __s64 maxidx;
 762
 763                         maxidx = ((i_size_read(dentry->d_inode) +
 764                                    CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT) - 1;
 765                         if (maxidx >= lnb->page->index) {
 766                                 LL_CDEBUG_PAGE(D_PAGE, lnb->page, "write %u @ "
 767                                                LPU64" flg %x before EOF %llu\n",
 768                                                lnb->len, lnb->offset,lnb->flags,
 769                                                i_size_read(dentry->d_inode));
 770                                 filter_iobuf_add_page(obd, iobuf,
 771                                                       dentry->d_inode,
 772                                                       lnb->page);
 773                         } else {
 774                                 long off;
 775                                 char *p = kmap(lnb->page);
 776
 777                                 off = lnb->offset & ~CFS_PAGE_MASK;
 778                                 if (off)
 779                                         memset(p, 0, off);
 780                                 off = (lnb->offset + lnb->len) & ~CFS_PAGE_MASK;
 781                                 if (off)
 782                                         memset(p + off, 0, CFS_PAGE_SIZE - off);
 783                                 kunmap(lnb->page);
 784                         }
 785                 }
 786                 if (lnb->rc == 0)
 787                         tot_bytes += lnb->len;
 788         }
 789         do_gettimeofday(&end);
 790         timediff = cfs_timeval_sub(&end, &start, NULL);
 791         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
 792
 793         if (OBD_FAIL_CHECK(OBD_FAIL_OST_NOMEM))
 794                 GOTO(cleanup, rc = -ENOMEM);
 795
 796         /* don't unlock pages to prevent any access */
 797         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf, exp,
 798                               NULL, NULL, NULL);
 799
 800         fsfilt_check_slow(obd, now, "start_page_write");
 801
 802         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
 803                             tot_bytes);
 804         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
 805                 lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
 806                                     LPROC_FILTER_WRITE_BYTES, tot_bytes);
 807         EXIT;
 808 cleanup:
 809         switch(cleanup_phase) {
 810         case 4:
 811                 if (rc) {
 812                         for (i = 0, lnb = res; i < *pages; i++, lnb++) {
 813                                 if (lnb->page != NULL) {
 814                                         unlock_page(lnb->page);
 815                                         page_cache_release(lnb->page);
 816                                         lnb->page = NULL;
 817                                 }
 818                         }
 819                         filter_grant_commit(exp, *pages, res);
 820                 }
 821         case 3:
 822                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 823         case 2:
 824                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 825                 if (rc)
 826                         f_dput(dentry);
 827                 break;
 828         case 1:
 829                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 830         case 0:
 831                 spin_lock(&obd->obd_osfs_lock);
 832                 if (oa)
 833                         filter_grant_incoming(exp, oa);
 834                 spin_unlock(&obd->obd_osfs_lock);
 835                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 836                 break;
 837         default:;
 838         }
 839         return rc;
 840 }
 841
 842 int filter_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
 843                   int objcount, struct obd_ioobj *obj,
 844                   struct niobuf_remote *nb, int *pages,
 845                   struct niobuf_local *res, struct obd_trans_info *oti)
 846 {
 847         if (cmd == OBD_BRW_WRITE)
 848                 return filter_preprw_write(cmd, exp, oa, objcount, obj,
 849                                            nb, pages, res, oti);
 850         if (cmd == OBD_BRW_READ)
 851                 return filter_preprw_read(cmd, exp, oa, objcount, obj,
 852                                           nb, pages, res, oti);
 853         LBUG();
 854         return -EPROTO;
 855 }
 856
 857 static int filter_commitrw_read(struct obd_export *exp, struct obdo *oa,
 858                                 int objcount, struct obd_ioobj *obj,
 859                                 struct niobuf_remote *rnb,
 860                                 int pages, struct niobuf_local *res,
 861                                 struct obd_trans_info *oti, int rc)
 862 {
 863         struct filter_obd *fo = &exp->exp_obd->u.filter;
 864         struct inode *inode = NULL;
 865         struct ldlm_res_id res_id = { .name = { obj->ioo_id } };
 866         struct ldlm_resource *resource = NULL;
 867         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 868         struct niobuf_local *lnb;
 869         int i;
 870         ENTRY;
 871
 872         /* If oa != NULL then filter_preprw_read updated the inode atime
 873          * and we should update the lvb so that other glimpses will also
 874          * get the updated value. bug 5972 */
 875         if (oa && ns && ns->ns_lvbo && ns->ns_lvbo->lvbo_update) {
 876                 resource = ldlm_resource_get(ns, NULL, res_id, LDLM_EXTENT, 0);
 877
 878                 if (resource != NULL) {
 879                         ns->ns_lvbo->lvbo_update(resource, NULL, 0, 1);
 880                         ldlm_resource_putref(resource);
 881                 }
 882         }
 883
 884         if (res->dentry != NULL)
 885                 inode = res->dentry->d_inode;
 886
 887         for (i = 0, lnb = res; i < pages; i++, lnb++) {
 888                 if (lnb->page != NULL) {
 889                         page_cache_release(lnb->page);
 890                         lnb->page = NULL;
 891                 }
 892         }
 893
 894         if (inode && (fo->fo_read_cache == 0 ||
 895                         i_size_read(inode) > fo->fo_readcache_max_filesize))
 896                 filter_invalidate_cache(exp->exp_obd, obj, rnb, inode);
 897
 898         if (res->dentry != NULL)
 899                 f_dput(res->dentry);
 900         RETURN(rc);
 901 }
 902
 903 void filter_grant_commit(struct obd_export *exp, int niocount,
 904                          struct niobuf_local *res)
 905 {
 906         struct filter_obd *filter = &exp->exp_obd->u.filter;
 907         struct niobuf_local *lnb = res;
 908         unsigned long pending = 0;
 909         int i;
 910
 911         spin_lock(&exp->exp_obd->obd_osfs_lock);
 912         for (i = 0, lnb = res; i < niocount; i++, lnb++)
 913                 pending += lnb->lnb_grant_used;
 914
 915         LASSERTF(exp->exp_filter_data.fed_pending >= pending,
 916                  "%s: cli %s/%p fed_pending: %lu grant_used: %lu\n",
 917                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 918                  exp->exp_filter_data.fed_pending, pending);
 919         exp->exp_filter_data.fed_pending -= pending;
 920         LASSERTF(filter->fo_tot_granted >= pending,
 921                  "%s: cli %s/%p tot_granted: "LPU64" grant_used: %lu\n",
 922                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 923                  exp->exp_obd->u.filter.fo_tot_granted, pending);
 924         filter->fo_tot_granted -= pending;
 925         LASSERTF(filter->fo_tot_pending >= pending,
 926                  "%s: cli %s/%p tot_pending: "LPU64" grant_used: %lu\n",
 927                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 928                  filter->fo_tot_pending, pending);
 929         filter->fo_tot_pending -= pending;
 930
 931         spin_unlock(&exp->exp_obd->obd_osfs_lock);
 932 }
 933
 934 int filter_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
 935                     int objcount, struct obd_ioobj *obj,
 936                     struct niobuf_remote *nb, int pages,
 937                     struct niobuf_local *res, struct obd_trans_info *oti,
 938                     int rc)
 939 {
 940         if (cmd == OBD_BRW_WRITE)
 941                 return filter_commitrw_write(exp, oa, objcount, obj,
 942                                              nb, pages, res, oti, rc);
 943         if (cmd == OBD_BRW_READ)
 944                 return filter_commitrw_read(exp, oa, objcount, obj,
 945                                             nb, pages, res, oti, rc);
 946         LBUG();
 947         return -EPROTO;
 948 }
 949
 950 int filter_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
 951                obd_count oa_bufs, struct brw_page *pga,
 952                struct obd_trans_info *oti)
 953 {
 954         struct obd_ioobj ioo;
 955         struct niobuf_local *lnb;
 956         struct niobuf_remote *rnb;
 957         obd_count i;
 958         int ret = 0, npages;
 959         ENTRY;
 960
 961         OBD_ALLOC(lnb, oa_bufs * sizeof(struct niobuf_local));
 962         OBD_ALLOC(rnb, oa_bufs * sizeof(struct niobuf_remote));
 963
 964         if (lnb == NULL || rnb == NULL)
 965                 GOTO(out, ret = -ENOMEM);
 966
 967         for (i = 0; i < oa_bufs; i++) {
 968                 lnb[i].page = pga[i].pg;
 969                 rnb[i].offset = pga[i].off;
 970                 rnb[i].len = pga[i].count;
 971         }
 972
 973         obdo_to_ioobj(oinfo->oi_oa, &ioo);
 974         ioo.ioo_bufcnt = oa_bufs;
 975
 976         npages = oa_bufs;
 977         ret = filter_preprw(cmd, exp, oinfo->oi_oa, 1, &ioo,
 978                             rnb, &npages, lnb, oti);
 979         if (ret != 0)
 980                 GOTO(out, ret);
 981         LASSERTF(oa_bufs == npages, "%u != %u\n", oa_bufs, npages);
 982
 983         ret = filter_commitrw(cmd, exp, oinfo->oi_oa, 1, &ioo, rnb,
 984                               npages, lnb, oti, ret);
 985
 986 out:
 987         if (lnb)
 988                 OBD_FREE(lnb, oa_bufs * sizeof(struct niobuf_local));
 989         if (rnb)
 990                 OBD_FREE(rnb, oa_bufs * sizeof(struct niobuf_remote));
 991         RETURN(ret);
 992 }