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