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