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  *  linux/fs/obdfilter/filter_io.c
   5  *
   6  *  Copyright (c) 2001-2003 Cluster File Systems, Inc.
   7  *   Author: Peter Braam <braam@clusterfs.com>
   8  *   Author: Andreas Dilger <adilger@clusterfs.com>
   9  *   Author: Phil Schwan <phil@clusterfs.com>
  10  *
  11  *   This file is part of Lustre, http://www.lustre.org.
  12  *
  13  *   Lustre is free software; you can redistribute it and/or
  14  *   modify it under the terms of version 2 of the GNU General Public
  15  *   License as published by the Free Software Foundation.
  16  *
  17  *   Lustre is distributed in the hope that it will be useful,
  18  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  19  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  20  *   GNU General Public License for more details.
  21  *
  22  *   You should have received a copy of the GNU General Public License
  23  *   along with Lustre; if not, write to the Free Software
  24  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  25  */
  26
  27 #define DEBUG_SUBSYSTEM S_FILTER
  28
  29 #include <linux/config.h>
  30 #include <linux/module.h>
  31 #include <linux/pagemap.h> // XXX kill me soon
  32 #include <linux/version.h>
  33
  34 #include <linux/obd_class.h>
  35 #include <linux/lustre_fsfilt.h>
  36 #include "filter_internal.h"
  37
  38 static int filter_start_page_read(struct inode *inode, struct niobuf_local *lnb)
  39 {
  40         struct address_space *mapping = inode->i_mapping;
  41         struct page *page;
  42         unsigned long index = lnb->offset >> PAGE_SHIFT;
  43         int rc;
  44
  45         page = grab_cache_page(mapping, index); /* locked page */
  46         if (page == NULL)
  47                 return lnb->rc = -ENOMEM;
  48
  49         LASSERT(page->mapping == mapping);
  50
  51         lnb->page = page;
  52
  53         if (inode->i_size < lnb->offset + lnb->len - 1)
  54                 lnb->rc = inode->i_size - lnb->offset;
  55         else
  56                 lnb->rc = lnb->len;
  57
  58         if (PageUptodate(page)) {
  59                 unlock_page(page);
  60                 return 0;
  61         }
  62
  63         rc = mapping->a_ops->readpage(NULL, page);
  64         if (rc < 0) {
  65                 CERROR("page index %lu, rc = %d\n", index, rc);
  66                 lnb->page = NULL;
  67                 page_cache_release(page);
  68                 return lnb->rc = rc;
  69         }
  70
  71         return 0;
  72 }
  73
  74 static int filter_finish_page_read(struct niobuf_local *lnb)
  75 {
  76         if (lnb->page == NULL)
  77                 return 0;
  78
  79         if (PageUptodate(lnb->page))
  80                 return 0;
  81
  82         wait_on_page(lnb->page);
  83         if (!PageUptodate(lnb->page)) {
  84                 CERROR("page index %lu/offset "LPX64" not uptodate\n",
  85                        lnb->page->index, lnb->offset);
  86                 GOTO(err_page, lnb->rc = -EIO);
  87         }
  88         if (PageError(lnb->page)) {
  89                 CERROR("page index %lu/offset "LPX64" has error\n",
  90                        lnb->page->index, lnb->offset);
  91                 GOTO(err_page, lnb->rc = -EIO);
  92         }
  93
  94         return 0;
  95
  96 err_page:
  97         page_cache_release(lnb->page);
  98         lnb->page = NULL;
  99         return lnb->rc;
 100 }
 101
 102 /* Grab the dirty and seen grant announcements from the incoming obdo.
 103  * We will later calculate the clients new grant and return it.
 104  * Caller must hold osfs lock */
 105 static void filter_grant_incoming(struct obd_export *exp, struct obdo *oa)
 106 {
 107         struct filter_export_data *fed;
 108         struct obd_device *obd = exp->exp_obd;
 109         static unsigned long last_msg;
 110         static int last_count;
 111         int mask = D_CACHE;
 112         ENTRY;
 113
 114         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
 115
 116         if ((oa->o_valid & (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) !=
 117                                         (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) {
 118                 oa->o_valid &= ~OBD_MD_FLGRANT;
 119                 EXIT;
 120                 return;
 121         }
 122
 123         fed = &exp->exp_filter_data;
 124
 125         /* Don't print this to the console the first time it happens, since
 126          * it can happen legitimately on occasion, but only rarely. */
 127         if (time_after(jiffies, last_msg + 60 * HZ)) {
 128                 last_count = 0;
 129                 last_msg = jiffies;
 130         }
 131         if ((last_count & (-last_count)) == last_count)
 132                 mask = D_WARNING;
 133         last_count++;
 134
 135         /* Add some margin, since there is a small race if other RPCs arrive
 136          * out-or-order and have already consumed some grant.  We want to
 137          * leave this here in case there is a large error in accounting. */
 138         CDEBUG(oa->o_grant > fed->fed_grant + FILTER_GRANT_CHUNK ? mask:D_CACHE,
 139                "%s: cli %s/%p reports grant: "LPU64" dropped: %u, local: %lu\n",
 140                obd->obd_name, exp->exp_client_uuid.uuid, exp, oa->o_grant,
 141                oa->o_dropped, fed->fed_grant);
 142
 143         /* Update our accounting now so that statfs takes it into account.
 144          * Note that fed_dirty is only approximate and can become incorrect
 145          * if RPCs arrive out-of-order.  No important calculations depend
 146          * on fed_dirty however. */
 147         obd->u.filter.fo_tot_dirty += oa->o_dirty - fed->fed_dirty;
 148         if (fed->fed_grant < oa->o_dropped) {
 149                 CERROR("%s: cli %s/%p reports %u dropped > fed_grant %lu\n",
 150                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
 151                        oa->o_dropped, fed->fed_grant);
 152                 oa->o_dropped = 0;
 153         }
 154         if (obd->u.filter.fo_tot_granted < oa->o_dropped) {
 155                 CERROR("%s: cli %s/%p reports %u dropped > tot_grant "LPU64"\n",
 156                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
 157                        oa->o_dropped, obd->u.filter.fo_tot_granted);
 158                 oa->o_dropped = 0;
 159         }
 160         obd->u.filter.fo_tot_granted -= oa->o_dropped;
 161         fed->fed_grant -= oa->o_dropped;
 162         fed->fed_dirty = oa->o_dirty;
 163         EXIT;
 164 }
 165
 166 #define GRANT_FOR_LLOG(obd) 16
 167
 168 /* Figure out how much space is available between what we've granted
 169  * and what remains in the filesystem.  Compensate for ext3 indirect
 170  * block overhead when computing how much free space is left ungranted.
 171  *
 172  * Caller must hold obd_osfs_lock. */
 173 obd_size filter_grant_space_left(struct obd_export *exp)
 174 {
 175         struct obd_device *obd = exp->exp_obd;
 176         int blockbits = obd->u.filter.fo_sb->s_blocksize_bits;
 177         obd_size tot_granted = obd->u.filter.fo_tot_granted, avail, left = 0;
 178         int rc, statfs_done = 0;
 179
 180         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
 181
 182         if (time_before(obd->obd_osfs_age, jiffies - HZ)) {
 183 restat:
 184                 rc = fsfilt_statfs(obd, obd->u.filter.fo_sb, jiffies + 1);
 185                 if (rc) /* N.B. statfs can't really fail */
 186                         RETURN(0);
 187                 statfs_done = 1;
 188         }
 189
 190         avail = obd->obd_osfs.os_bavail;
 191         left = avail - (avail >> (blockbits - 3)); /* (d)indirect */
 192         if (left > GRANT_FOR_LLOG(obd)) {
 193                 left = (left - GRANT_FOR_LLOG(obd)) << blockbits;
 194         } else {
 195                 left = 0 /* << blockbits */;
 196         }
 197
 198         if (!statfs_done && left < 32 * FILTER_GRANT_CHUNK + tot_granted) {
 199                 CDEBUG(D_CACHE, "fs has no space left and statfs too old\n");
 200                 goto restat;
 201         }
 202
 203         if (left >= tot_granted) {
 204                 left -= tot_granted;
 205         } else {
 206                 static unsigned long next;
 207                 if (left < tot_granted - obd->u.filter.fo_tot_pending &&
 208                     time_after(jiffies, next)) {
 209                         spin_unlock(&obd->obd_osfs_lock);
 210                         CERROR("%s: cli %s/%p grant "LPU64" > available "
 211                                LPU64" and pending "LPU64"\n", obd->obd_name,
 212                                exp->exp_client_uuid.uuid, exp, tot_granted,
 213                                left, obd->u.filter.fo_tot_pending);
 214                         if (next == 0)
 215                                 portals_debug_dumplog();
 216                         next = jiffies + 20 * HZ;
 217                         spin_lock(&obd->obd_osfs_lock);
 218                 }
 219                 left = 0;
 220         }
 221
 222         CDEBUG(D_CACHE, "%s: cli %s/%p free: "LPU64" avail: "LPU64" grant "LPU64
 223                " left: "LPU64" pending: "LPU64"\n", obd->obd_name,
 224                exp->exp_client_uuid.uuid, exp,
 225                obd->obd_osfs.os_bfree << blockbits, avail << blockbits,
 226                tot_granted, left, obd->u.filter.fo_tot_pending);
 227
 228         return left;
 229 }
 230
 231 /* Calculate how much grant space to allocate to this client, based on how
 232  * much space is currently free and how much of that is already granted.
 233  *
 234  * Caller must hold obd_osfs_lock. */
 235 long filter_grant(struct obd_export *exp, obd_size current_grant,
 236                   obd_size want, obd_size fs_space_left)
 237 {
 238         struct obd_device *obd = exp->exp_obd;
 239         struct filter_export_data *fed = &exp->exp_filter_data;
 240         int blockbits = obd->u.filter.fo_sb->s_blocksize_bits;
 241         __u64 grant = 0;
 242
 243         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
 244
 245         /* Grant some fraction of the client's requested grant space so that
 246          * they are not always waiting for write credits (not all of it to
 247          * avoid overgranting in face of multiple RPCs in flight).  This
 248          * essentially will be able to control the OSC_MAX_RIF for a client.
 249          *
 250          * If we do have a large disparity between what the client thinks it
 251          * has and what we think it has, don't grant very much and let the
 252          * client consume its grant first.  Either it just has lots of RPCs
 253          * in flight, or it was evicted and its grants will soon be used up. */
 254         if (current_grant < want &&
 255             current_grant < fed->fed_grant + FILTER_GRANT_CHUNK) {
 256                 grant = min((want >> blockbits) / 2,
 257                             (fs_space_left >> blockbits) / 8);
 258                 grant <<= blockbits;
 259
 260                 if (grant) {
 261                         if (grant > FILTER_GRANT_CHUNK)
 262                                 grant = FILTER_GRANT_CHUNK;
 263
 264                         obd->u.filter.fo_tot_granted += grant;
 265                         fed->fed_grant += grant;
 266                 }
 267         }
 268
 269         CDEBUG(D_CACHE,"%s: cli %s/%p wants: "LPU64" granting: "LPU64"\n",
 270                obd->obd_name, exp->exp_client_uuid.uuid, exp, want, grant);
 271         CDEBUG(D_CACHE,
 272                "%s: cli %s/%p tot cached:"LPU64" granted:"LPU64
 273                " num_exports: %d\n", obd->obd_name, exp->exp_client_uuid.uuid,
 274                exp, obd->u.filter.fo_tot_dirty,
 275                obd->u.filter.fo_tot_granted, obd->obd_num_exports);
 276
 277         return grant;
 278 }
 279
 280 static int filter_preprw_read(int cmd, struct obd_export *exp, struct obdo *oa,
 281                               int objcount, struct obd_ioobj *obj,
 282                               int niocount, struct niobuf_remote *nb,
 283                               struct niobuf_local *res,
 284                               struct obd_trans_info *oti)
 285 {
 286         struct obd_device *obd = exp->exp_obd;
 287         struct obd_run_ctxt saved;
 288         struct obd_ioobj *o;
 289         struct niobuf_remote *rnb;
 290         struct niobuf_local *lnb = NULL;
 291         struct fsfilt_objinfo *fso;
 292         struct dentry *dentry;
 293         struct inode *inode;
 294         int rc = 0, i, j, tot_bytes = 0, cleanup_phase = 0;
 295         unsigned long now = jiffies;
 296         ENTRY;
 297
 298         /* We are currently not supporting multi-obj BRW_READ RPCS at all.
 299          * When we do this function's dentry cleanup will need to be fixed */
 300         LASSERT(objcount == 1);
 301         LASSERT(obj->ioo_bufcnt > 0);
 302
 303         if (oa && oa->o_valid & OBD_MD_FLGRANT) {
 304                 spin_lock(&obd->obd_osfs_lock);
 305                 filter_grant_incoming(exp, oa);
 306
 307 #if 0
 308                 /* Reads do not increase grants */
 309                 oa->o_grant = filter_grant(exp, oa->o_grant, oa->o_undirty,
 310                                            filter_grant_space_left(exp));
 311 #else
 312                 oa->o_grant = 0;
 313 #endif
 314                 spin_unlock(&obd->obd_osfs_lock);
 315         }
 316
 317         OBD_ALLOC(fso, objcount * sizeof(*fso));
 318         if (fso == NULL)
 319                 RETURN(-ENOMEM);
 320
 321         memset(res, 0, niocount * sizeof(*res));
 322
 323         push_ctxt(&saved, &exp->exp_obd->obd_ctxt, NULL);
 324         for (i = 0, o = obj; i < objcount; i++, o++) {
 325                 LASSERT(o->ioo_bufcnt);
 326
 327                 dentry = filter_oa2dentry(obd, oa);
 328                 if (IS_ERR(dentry))
 329                         GOTO(cleanup, rc = PTR_ERR(dentry));
 330
 331                 if (dentry->d_inode == NULL) {
 332                         CERROR("trying to BRW to non-existent file "LPU64"\n",
 333                                o->ioo_id);
 334                         f_dput(dentry);
 335                         GOTO(cleanup, rc = -ENOENT);
 336                 }
 337
 338                 fso[i].fso_dentry = dentry;
 339                 fso[i].fso_bufcnt = o->ioo_bufcnt;
 340         }
 341
 342         if (time_after(jiffies, now + 15 * HZ))
 343                 CERROR("slow preprw_read setup %lus\n", (jiffies - now) / HZ);
 344         else
 345                 CDEBUG(D_INFO, "preprw_read setup: %lu jiffies\n",
 346                        (jiffies - now));
 347
 348         for (i = 0, o = obj, rnb = nb, lnb = res; i < objcount; i++, o++) {
 349                 dentry = fso[i].fso_dentry;
 350                 inode = dentry->d_inode;
 351
 352                 for (j = 0; j < o->ioo_bufcnt; j++, rnb++, lnb++) {
 353                         lnb->dentry = dentry;
 354                         lnb->offset = rnb->offset;
 355                         lnb->len    = rnb->len;
 356                         lnb->flags  = rnb->flags;
 357
 358                         if (inode->i_size <= rnb->offset) {
 359                                 /* If there's no more data, abort early.
 360                                  * lnb->page == NULL and lnb->rc == 0, so it's
 361                                  * easy to detect later. */
 362                                 break;
 363                         } else {
 364                                 rc = filter_start_page_read(inode, lnb);
 365                         }
 366
 367                         if (rc) {
 368                                 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
 369                                        "page err %u@"LPU64" %u/%u %p: rc %d\n",
 370                                        lnb->len, lnb->offset, j, o->ioo_bufcnt,
 371                                        dentry, rc);
 372                                 cleanup_phase = 1;
 373                                 GOTO(cleanup, rc);
 374                         }
 375
 376                         tot_bytes += lnb->rc;
 377                         if (lnb->rc < lnb->len) {
 378                                 /* short read, be sure to wait on it */
 379                                 lnb++;
 380                                 break;
 381                         }
 382                 }
 383         }
 384
 385         if (time_after(jiffies, now + 15 * HZ))
 386                 CERROR("slow start_page_read %lus\n", (jiffies - now) / HZ);
 387         else
 388                 CDEBUG(D_INFO, "start_page_read: %lu jiffies\n",
 389                        (jiffies - now));
 390
 391         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_READ_BYTES, tot_bytes);
 392         while (lnb-- > res) {
 393                 rc = filter_finish_page_read(lnb);
 394                 if (rc) {
 395                         CERROR("error page %u@"LPU64" %u %p: rc %d\n", lnb->len,
 396                                lnb->offset, (int)(lnb - res), lnb->dentry, rc);
 397                         cleanup_phase = 1;
 398                         GOTO(cleanup, rc);
 399                 }
 400         }
 401
 402         if (time_after(jiffies, now + 15 * HZ))
 403                 CERROR("slow finish_page_read %lus\n", (jiffies - now) / HZ);
 404         else
 405                 CDEBUG(D_INFO, "finish_page_read: %lu jiffies\n",
 406                        (jiffies - now));
 407
 408         filter_tally_read(&exp->exp_obd->u.filter, res, niocount);
 409
 410         EXIT;
 411
 412  cleanup:
 413         switch (cleanup_phase) {
 414         case 1:
 415                 for (lnb = res; lnb < (res + niocount); lnb++) {
 416                         if (lnb->page)
 417                                 page_cache_release(lnb->page);
 418                 }
 419                 if (res->dentry != NULL)
 420                         f_dput(res->dentry);
 421                 else
 422                         CERROR("NULL dentry in cleanup -- tell CFS\n");
 423         case 0:
 424                 OBD_FREE(fso, objcount * sizeof(*fso));
 425                 pop_ctxt(&saved, &exp->exp_obd->obd_ctxt, NULL);
 426         }
 427         return rc;
 428 }
 429
 430 /* When clients have dirtied as much space as they've been granted they
 431  * fall through to sync writes.  These sync writes haven't been expressed
 432  * in grants and need to error with ENOSPC when there isn't room in the
 433  * filesystem for them after grants are taken into account.  However,
 434  * writeback of the dirty data that was already granted space can write
 435  * right on through.
 436  *
 437  * Caller must hold obd_osfs_lock. */
 438 static int filter_grant_check(struct obd_export *exp, int objcount,
 439                               struct fsfilt_objinfo *fso, int niocount,
 440                               struct niobuf_remote *rnb,
 441                               struct niobuf_local *lnb, obd_size *left,
 442                               struct inode *inode)
 443 {
 444         struct filter_export_data *fed = &exp->exp_filter_data;
 445         int blocksize = exp->exp_obd->u.filter.fo_sb->s_blocksize;
 446         unsigned long used = 0, ungranted = 0, using;
 447         int i, rc = -ENOSPC, obj, n = 0, mask = D_CACHE;
 448
 449         LASSERT_SPIN_LOCKED(&exp->exp_obd->obd_osfs_lock);
 450
 451         for (obj = 0; obj < objcount; obj++) {
 452                 for (i = 0; i < fso[obj].fso_bufcnt; i++, n++) {
 453                         int tmp, bytes;
 454
 455                         /* FIXME: this is calculated with PAGE_SIZE on client */
 456                         bytes = rnb[n].len;
 457                         bytes += rnb[n].offset & (blocksize - 1);
 458                         tmp = (rnb[n].offset + rnb[n].len) & (blocksize - 1);
 459                         if (tmp)
 460                                 bytes += blocksize - tmp;
 461
 462                         if (rnb[n].flags & OBD_BRW_FROM_GRANT) {
 463                                 if (fed->fed_grant < used + bytes) {
 464                                         CDEBUG(D_CACHE,
 465                                                "%s: cli %s/%p claims %ld+%d "
 466                                                "GRANT, real grant %lu idx %d\n",
 467                                                exp->exp_obd->obd_name,
 468                                                exp->exp_client_uuid.uuid, exp,
 469                                                used, bytes, fed->fed_grant, n);
 470                                         mask = D_ERROR;
 471                                 } else {
 472                                         used += bytes;
 473                                         rnb[n].flags |= OBD_BRW_GRANTED;
 474                                         lnb[n].lnb_grant_used = bytes;
 475                                         CDEBUG(0, "idx %d used=%lu\n", n, used);
 476                                         rc = 0;
 477                                         continue;
 478                                 }
 479                         }
 480                         if (*left > ungranted) {
 481                                 /* if enough space, pretend it was granted */
 482                                 ungranted += bytes;
 483                                 rnb[n].flags |= OBD_BRW_GRANTED;
 484                                 CDEBUG(0, "idx %d ungranted=%lu\n",n,ungranted);
 485                                 rc = 0;
 486                                 continue;
 487                         }
 488
 489                         /* We can't check for already-mapped blocks here, as
 490                          * it requires dropping the osfs lock to do the bmap.
 491                          * Instead, we return ENOSPC and in that case we need
 492                          * to go through and verify if all of the blocks not
 493                          * marked BRW_GRANTED are already mapped and we can
 494                          * ignore this error. */
 495                         lnb[n].rc = -ENOSPC;
 496                         rnb[n].flags &= OBD_BRW_GRANTED;
 497                         CDEBUG(D_CACHE,"%s: cli %s/%p idx %d no space for %d\n",
 498                                exp->exp_obd->obd_name,
 499                                exp->exp_client_uuid.uuid, exp, n, bytes);
 500                 }
 501         }
 502
 503         /* Now substract what client have used already.  We don't subtract
 504          * this from the tot_granted yet, so that other client's can't grab
 505          * that space before we have actually allocated our blocks.  That
 506          * happens in filter_grant_commit() after the writes are done. */
 507         *left -= ungranted;
 508         fed->fed_grant -= used;
 509         fed->fed_pending += used;
 510         exp->exp_obd->u.filter.fo_tot_pending += used;
 511
 512         CDEBUG(mask,
 513                "%s: cli %s/%p used: %lu ungranted: %lu grant: %lu dirty: %lu\n",
 514                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp, used,
 515                ungranted, fed->fed_grant, fed->fed_dirty);
 516
 517         /* Rough calc in case we don't refresh cached statfs data */
 518         using = (used + ungranted + 1 ) >>
 519                 exp->exp_obd->u.filter.fo_sb->s_blocksize_bits;
 520         if (exp->exp_obd->obd_osfs.os_bavail > using)
 521                 exp->exp_obd->obd_osfs.os_bavail -= using;
 522         else
 523                 exp->exp_obd->obd_osfs.os_bavail = 0;
 524
 525         if (fed->fed_dirty < used) {
 526                 CERROR("%s: cli %s/%p claims used %lu > fed_dirty %lu\n",
 527                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 528                        used, fed->fed_dirty);
 529                 used = fed->fed_dirty;
 530         }
 531         exp->exp_obd->u.filter.fo_tot_dirty -= used;
 532         fed->fed_dirty -= used;
 533
 534         return rc;
 535 }
 536
 537 static int filter_start_page_write(struct inode *inode,
 538                                    struct niobuf_local *lnb)
 539 {
 540         struct page *page = alloc_pages(GFP_HIGHUSER, 0);
 541         if (page == NULL) {
 542                 CERROR("no memory for a temp page\n");
 543                 RETURN(lnb->rc = -ENOMEM);
 544         }
 545         POISON_PAGE(page, 0xf1);
 546         if (lnb->len != PAGE_SIZE) {
 547                 memset(kmap(page) + lnb->len, 0, PAGE_SIZE - lnb->len);
 548                 kunmap(page);
 549         }
 550         page->index = lnb->offset >> PAGE_SHIFT;
 551         lnb->page = page;
 552
 553         return 0;
 554 }
 555
 556 /* If we ever start to support multi-object BRW RPCs, we will need to get locks
 557  * on mulitple inodes.  That isn't all, because there still exists the
 558  * possibility of a truncate starting a new transaction while holding the ext3
 559  * rwsem = write while some writes (which have started their transactions here)
 560  * blocking on the ext3 rwsem = read => lock inversion.
 561  *
 562  * The handling gets very ugly when dealing with locked pages.  It may be easier
 563  * to just get rid of the locked page code (which has problems of its own) and
 564  * either discover we do not need it anymore (i.e. it was a symptom of another
 565  * bug) or ensure we get the page locks in an appropriate order. */
 566 static int filter_preprw_write(int cmd, struct obd_export *exp, struct obdo *oa,
 567                                int objcount, struct obd_ioobj *obj,
 568                                int niocount, struct niobuf_remote *nb,
 569                                struct niobuf_local *res,
 570                                struct obd_trans_info *oti)
 571 {
 572         struct obd_run_ctxt saved;
 573         struct niobuf_remote *rnb;
 574         struct niobuf_local *lnb;
 575         struct fsfilt_objinfo fso;
 576         struct dentry *dentry;
 577         obd_size left;
 578         unsigned long now = jiffies;
 579         int rc = 0, i, tot_bytes = 0, cleanup_phase = 1;
 580         ENTRY;
 581         LASSERT(objcount == 1);
 582         LASSERT(obj->ioo_bufcnt > 0);
 583
 584         memset(res, 0, niocount * sizeof(*res));
 585
 586         push_ctxt(&saved, &exp->exp_obd->obd_ctxt, NULL);
 587         dentry = filter_fid2dentry(exp->exp_obd, NULL, obj->ioo_gr,
 588                                    obj->ioo_id);
 589         if (IS_ERR(dentry))
 590                 GOTO(cleanup, rc = PTR_ERR(dentry));
 591
 592         if (dentry->d_inode == NULL) {
 593                 CERROR("trying to BRW to non-existent file "LPU64"\n",
 594                        obj->ioo_id);
 595                 f_dput(dentry);
 596                 GOTO(cleanup, rc = -ENOENT);
 597         }
 598
 599         fso.fso_dentry = dentry;
 600         fso.fso_bufcnt = obj->ioo_bufcnt;
 601
 602         if (time_after(jiffies, now + 15 * HZ))
 603                 CERROR("slow preprw_write setup %lus\n", (jiffies - now) / HZ);
 604         else
 605                 CDEBUG(D_INFO, "preprw_write setup: %lu jiffies\n",
 606                        (jiffies - now));
 607
 608         spin_lock(&exp->exp_obd->obd_osfs_lock);
 609         if (oa)
 610                 filter_grant_incoming(exp, oa);
 611         cleanup_phase = 0;
 612
 613         left = filter_grant_space_left(exp);
 614
 615         rc = filter_grant_check(exp, objcount, &fso, niocount, nb, res,
 616                                 &left, dentry->d_inode);
 617         if (oa && oa->o_valid & OBD_MD_FLGRANT)
 618                 oa->o_grant = filter_grant(exp,oa->o_grant,oa->o_undirty,left);
 619
 620         spin_unlock(&exp->exp_obd->obd_osfs_lock);
 621
 622         if (rc) {
 623                 f_dput(dentry);
 624                 GOTO(cleanup, rc);
 625         }
 626
 627         for (i = 0, rnb = nb, lnb = res; i < obj->ioo_bufcnt;
 628              i++, lnb++, rnb++) {
 629                 /* We still set up for ungranted pages so that granted pages
 630                  * can be written to disk as they were promised, and portals
 631                  * needs to keep the pages all aligned properly. */
 632                 lnb->dentry = dentry;
 633                 lnb->offset = rnb->offset;
 634                 lnb->len    = rnb->len;
 635                 lnb->flags  = rnb->flags;
 636
 637                 rc = filter_start_page_write(dentry->d_inode, lnb);
 638                 if (rc) {
 639                         CERROR("page err %u@"LPU64" %u/%u %p: rc %d\n",
 640                                lnb->len, lnb->offset,
 641                                i, obj->ioo_bufcnt, dentry, rc);
 642                         while (lnb-- > res)
 643                                 __free_pages(lnb->page, 0);
 644                         f_dput(dentry);
 645                         GOTO(cleanup, rc);
 646                 }
 647                 if (lnb->rc == 0)
 648                         tot_bytes += lnb->len;
 649         }
 650
 651         if (time_after(jiffies, now + 15 * HZ))
 652                 CERROR("slow start_page_write %lus\n", (jiffies - now) / HZ);
 653         else
 654                 CDEBUG(D_INFO, "start_page_write: %lu jiffies\n",
 655                        (jiffies - now));
 656
 657         lprocfs_counter_add(exp->exp_obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
 658                             tot_bytes);
 659         EXIT;
 660 cleanup:
 661         switch(cleanup_phase) {
 662         case 1:
 663                 spin_lock(&exp->exp_obd->obd_osfs_lock);
 664                 if (oa)
 665                         filter_grant_incoming(exp, oa);
 666                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
 667         default: ;
 668         }
 669         pop_ctxt(&saved, &exp->exp_obd->obd_ctxt, NULL);
 670         return rc;
 671 }
 672
 673 int filter_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
 674                   int objcount, struct obd_ioobj *obj, int niocount,
 675                   struct niobuf_remote *nb, struct niobuf_local *res,
 676                   struct obd_trans_info *oti)
 677 {
 678         if (cmd == OBD_BRW_WRITE)
 679                 return filter_preprw_write(cmd, exp, oa, objcount, obj,
 680                                            niocount, nb, res, oti);
 681
 682         if (cmd == OBD_BRW_READ)
 683                 return filter_preprw_read(cmd, exp, oa, objcount, obj,
 684                                           niocount, nb, res, oti);
 685
 686         LBUG();
 687         return -EPROTO;
 688 }
 689
 690 static int filter_commitrw_read(struct obd_export *exp, struct obdo *oa,
 691                                 int objcount, struct obd_ioobj *obj,
 692                                 int niocount, struct niobuf_local *res,
 693                                 struct obd_trans_info *oti, int rc)
 694 {
 695         struct obd_ioobj *o;
 696         struct niobuf_local *lnb;
 697         int i, j, drop = 0;
 698         ENTRY;
 699
 700         if (res->dentry != NULL)
 701                 drop = (res->dentry->d_inode->i_size >
 702                         exp->exp_obd->u.filter.fo_readcache_max_filesize);
 703
 704         for (i = 0, o = obj, lnb = res; i < objcount; i++, o++) {
 705                 for (j = 0 ; j < o->ioo_bufcnt ; j++, lnb++) {
 706                         if (lnb->page == NULL)
 707                                 continue;
 708                         /* drop from cache like truncate_list_pages() */
 709                         if (drop && !TryLockPage(lnb->page)) {
 710                                 if (lnb->page->mapping)
 711                                         ll_truncate_complete_page(lnb->page);
 712                                 unlock_page(lnb->page);
 713                         }
 714                         page_cache_release(lnb->page);
 715                 }
 716         }
 717
 718         if (res->dentry != NULL)
 719                 f_dput(res->dentry);
 720         RETURN(rc);
 721 }
 722
 723 void flip_into_page_cache(struct inode *inode, struct page *new_page)
 724 {
 725         struct page *old_page;
 726         int rc;
 727
 728         do {
 729                 /* the dlm is protecting us from read/write concurrency, so we
 730                  * expect this find_lock_page to return quickly.  even if we
 731                  * race with another writer it won't be doing much work with
 732                  * the page locked.  we do this 'cause t_c_p expects a
 733                  * locked page, and it wants to grab the pagecache lock
 734                  * as well. */
 735                 old_page = find_lock_page(inode->i_mapping, new_page->index);
 736                 if (old_page) {
 737                         ll_truncate_complete_page(old_page);
 738                         unlock_page(old_page);
 739                         page_cache_release(old_page);
 740                 }
 741
 742 #if 0 /* this should be a /proc tunable someday */
 743                 /* racing o_directs (no locking ioctl) could race adding
 744                  * their pages, so we repeat the page invalidation unless
 745                  * we successfully added our new page */
 746                 rc = add_to_page_cache_unique(new_page, inode->i_mapping,
 747                                               new_page->index,
 748                                               page_hash(inode->i_mapping,
 749                                                         new_page->index));
 750                 if (rc == 0) {
 751                         /* add_to_page_cache clears uptodate|dirty and locks
 752                          * the page */
 753                         SetPageUptodate(new_page);
 754                         unlock_page(new_page);
 755                 }
 756 #else
 757                 rc = 0;
 758 #endif
 759         } while (rc != 0);
 760 }
 761
 762 void filter_grant_commit(struct obd_export *exp, int niocount,
 763                          struct niobuf_local *res)
 764 {
 765         struct filter_obd *filter = &exp->exp_obd->u.filter;
 766         struct niobuf_local *lnb = res;
 767         unsigned long pending = 0;
 768         int i;
 769
 770         spin_lock(&exp->exp_obd->obd_osfs_lock);
 771         for (i = 0, lnb = res; i < niocount; i++, lnb++)
 772                 pending += lnb->lnb_grant_used;
 773
 774         LASSERTF(exp->exp_filter_data.fed_pending >= pending,
 775                  "%s: cli %s/%p fed_pending: %lu grant_used: %lu\n",
 776                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 777                  exp->exp_filter_data.fed_pending, pending);
 778         exp->exp_filter_data.fed_pending -= pending;
 779         LASSERTF(filter->fo_tot_granted >= pending,
 780                  "%s: cli %s/%p tot_granted: "LPU64" grant_used: %lu\n",
 781                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 782                  exp->exp_obd->u.filter.fo_tot_granted, pending);
 783         filter->fo_tot_granted -= pending;
 784         LASSERTF(filter->fo_tot_pending >= pending,
 785                  "%s: cli %s/%p tot_pending: "LPU64" grant_used: %lu\n",
 786                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 787                  filter->fo_tot_pending, pending);
 788         filter->fo_tot_pending -= pending;
 789
 790         spin_unlock(&exp->exp_obd->obd_osfs_lock);
 791 }
 792
 793 int filter_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
 794                     int objcount, struct obd_ioobj *obj, int niocount,
 795                     struct niobuf_local *res, struct obd_trans_info *oti,int rc)
 796 {
 797         if (cmd == OBD_BRW_WRITE)
 798                 return filter_commitrw_write(exp, oa, objcount, obj, niocount,
 799                                              res, oti, rc);
 800         if (cmd == OBD_BRW_READ)
 801                 return filter_commitrw_read(exp, oa, objcount, obj, niocount,
 802                                             res, oti, rc);
 803         LBUG();
 804         return -EPROTO;
 805 }
 806
 807 int filter_brw(int cmd, struct obd_export *exp, struct obdo *oa,
 808                struct lov_stripe_md *lsm, obd_count oa_bufs,
 809                struct brw_page *pga, struct obd_trans_info *oti)
 810 {
 811         struct obd_ioobj ioo;
 812         struct niobuf_local *lnb;
 813         struct niobuf_remote *rnb;
 814         obd_count i;
 815         int ret = 0;
 816         ENTRY;
 817
 818         OBD_ALLOC(lnb, oa_bufs * sizeof(struct niobuf_local));
 819         OBD_ALLOC(rnb, oa_bufs * sizeof(struct niobuf_remote));
 820
 821         if (lnb == NULL || rnb == NULL)
 822                 GOTO(out, ret = -ENOMEM);
 823
 824         for (i = 0; i < oa_bufs; i++) {
 825                 rnb[i].offset = pga[i].off;
 826                 rnb[i].len = pga[i].count;
 827         }
 828
 829         obdo_to_ioobj(oa, &ioo);
 830         ioo.ioo_bufcnt = oa_bufs;
 831
 832         ret = filter_preprw(cmd, exp, oa, 1, &ioo, oa_bufs, rnb, lnb, oti);
 833         if (ret != 0)
 834                 GOTO(out, ret);
 835
 836         for (i = 0; i < oa_bufs; i++) {
 837                 void *virt = kmap(pga[i].pg);
 838                 obd_off off = pga[i].off & ~PAGE_MASK;
 839                 void *addr = kmap(lnb[i].page);
 840
 841                 /* 2 kmaps == vanishingly small deadlock opportunity */
 842
 843                 if (cmd & OBD_BRW_WRITE)
 844                         memcpy(addr + off, virt + off, pga[i].count);
 845                 else
 846                         memcpy(virt + off, addr + off, pga[i].count);
 847
 848                 kunmap(lnb[i].page);
 849                 kunmap(pga[i].pg);
 850         }
 851
 852         ret = filter_commitrw(cmd, exp, oa, 1, &ioo, oa_bufs, lnb, oti, ret);
 853
 854 out:
 855         if (lnb)
 856                 OBD_FREE(lnb, oa_bufs * sizeof(struct niobuf_local));
 857         if (rnb)
 858                 OBD_FREE(rnb, oa_bufs * sizeof(struct niobuf_remote));
 859         RETURN(ret);
 860 }