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 <linux/lustre_smfs.h>
  37 #include <linux/lustre_snap.h>
  38 #include "filter_internal.h"
  39
  40 static int filter_alloc_dio_page(struct obd_device *obd, struct inode *inode,
  41                                  struct niobuf_local *lnb)
  42
  43 {
  44         struct page *page;
  45         ENTRY;
  46
  47         page = alloc_pages(GFP_HIGHUSER, 0);
  48         if (page == NULL) {
  49                 CERROR("no memory for a temp page\n");
  50                 lnb->rc = -ENOMEM;
  51                 RETURN(-ENOMEM);
  52         }
  53
  54 #if 0
  55         POISON_PAGE(page, 0xf1);
  56         if (lnb->len != PAGE_SIZE) {
  57                 memset(kmap(page) + lnb->len, 0, PAGE_SIZE - lnb->len);
  58                 kunmap(page);
  59         }
  60 #endif
  61         page->index = lnb->offset >> PAGE_SHIFT;
  62
  63         lnb->page = page;
  64
  65         RETURN(0);
  66 }
  67
  68 void filter_free_dio_pages(int objcount, struct obd_ioobj *obj,
  69                            int niocount, struct niobuf_local *res)
  70 {
  71         int i, j;
  72
  73         for (i = 0; i < objcount; i++, obj++) {
  74                 for (j = 0 ; j < obj->ioo_bufcnt ; j++, res++) {
  75                         if (res->page != NULL) {
  76                                 __free_page(res->page);
  77                                 res->page = NULL;
  78                         }
  79                 }
  80         }
  81 }
  82
  83 /* Grab the dirty and seen grant announcements from the incoming obdo.
  84  * We will later calculate the clients new grant and return it.
  85  * Caller must hold osfs lock */
  86 static void filter_grant_incoming(struct obd_export *exp, struct obdo *oa)
  87 {
  88         struct filter_export_data *fed;
  89         struct obd_device *obd = exp->exp_obd;
  90         static unsigned long last_msg;
  91         static int last_count;
  92         int mask = D_CACHE;
  93         ENTRY;
  94
  95         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
  96
  97         if ((oa->o_valid & (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) !=
  98                                         (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) {
  99                 oa->o_valid &= ~OBD_MD_FLGRANT;
 100                 EXIT;
 101                 return;
 102         }
 103
 104         fed = &exp->exp_filter_data;
 105
 106         /* Don't print this to the console the first time it happens, since
 107          * it can happen legitimately on occasion, but only rarely. */
 108         if (time_after(jiffies, last_msg + 60 * HZ)) {
 109                 last_count = 0;
 110                 last_msg = jiffies;
 111         }
 112         if ((last_count & (-last_count)) == last_count)
 113                 mask = D_WARNING;
 114         last_count++;
 115
 116         /* Add some margin, since there is a small race if other RPCs arrive
 117          * out-or-order and have already consumed some grant.  We want to
 118          * leave this here in case there is a large error in accounting. */
 119         CDEBUG(oa->o_grant > fed->fed_grant + FILTER_GRANT_CHUNK ? mask:D_CACHE,
 120                "%s: cli %s/%p reports grant: "LPU64" dropped: %u, local: %lu\n",
 121                obd->obd_name, exp->exp_client_uuid.uuid, exp, oa->o_grant,
 122                oa->o_dropped, fed->fed_grant);
 123
 124         /* Update our accounting now so that statfs takes it into account.
 125          * Note that fed_dirty is only approximate and can become incorrect
 126          * if RPCs arrive out-of-order.  No important calculations depend
 127          * on fed_dirty however. */
 128         obd->u.filter.fo_tot_dirty += oa->o_dirty - fed->fed_dirty;
 129         if (fed->fed_grant < oa->o_dropped) {
 130                 CERROR("%s: cli %s/%p reports %u dropped > fed_grant %lu\n",
 131                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
 132                        oa->o_dropped, fed->fed_grant);
 133                 oa->o_dropped = 0;
 134         }
 135         if (obd->u.filter.fo_tot_granted < oa->o_dropped) {
 136                 CERROR("%s: cli %s/%p reports %u dropped > tot_grant "LPU64"\n",
 137                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
 138                        oa->o_dropped, obd->u.filter.fo_tot_granted);
 139                 oa->o_dropped = 0;
 140         }
 141         obd->u.filter.fo_tot_granted -= oa->o_dropped;
 142         fed->fed_grant -= oa->o_dropped;
 143         fed->fed_dirty = oa->o_dirty;
 144         EXIT;
 145 }
 146
 147 #define GRANT_FOR_LLOG(obd) 16
 148
 149 /* Figure out how much space is available between what we've granted
 150  * and what remains in the filesystem.  Compensate for ext3 indirect
 151  * block overhead when computing how much free space is left ungranted.
 152  *
 153  * Caller must hold obd_osfs_lock. */
 154 obd_size filter_grant_space_left(struct obd_export *exp)
 155 {
 156         struct obd_device *obd = exp->exp_obd;
 157         int blockbits = obd->u.filter.fo_sb->s_blocksize_bits;
 158         obd_size tot_granted = obd->u.filter.fo_tot_granted, avail, left = 0;
 159         int rc, statfs_done = 0;
 160
 161         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
 162
 163         if (time_before(obd->obd_osfs_age, jiffies - HZ)) {
 164 restat:
 165                 rc = fsfilt_statfs(obd, obd->u.filter.fo_sb, jiffies + 1);
 166                 if (rc) /* N.B. statfs can't really fail */
 167                         RETURN(0);
 168                 statfs_done = 1;
 169         }
 170
 171         avail = obd->obd_osfs.os_bavail;
 172         left = avail - (avail >> (blockbits - 3)); /* (d)indirect */
 173         if (left > GRANT_FOR_LLOG(obd)) {
 174                 left = (left - GRANT_FOR_LLOG(obd)) << blockbits;
 175         } else {
 176                 left = 0 /* << blockbits */;
 177         }
 178
 179         if (!statfs_done && left < 32 * FILTER_GRANT_CHUNK + tot_granted) {
 180                 CDEBUG(D_CACHE, "fs has no space left and statfs too old\n");
 181                 goto restat;
 182         }
 183
 184         if (left >= tot_granted) {
 185                 left -= tot_granted;
 186         } else {
 187                 static unsigned long next;
 188                 if (left < tot_granted - obd->u.filter.fo_tot_pending &&
 189                     time_after(jiffies, next)) {
 190                         spin_unlock(&obd->obd_osfs_lock);
 191                         CERROR("%s: cli %s/%p grant "LPU64" > available "
 192                                LPU64" and pending "LPU64"\n", obd->obd_name,
 193                                exp->exp_client_uuid.uuid, exp, tot_granted,
 194                                left, obd->u.filter.fo_tot_pending);
 195                         if (next == 0)
 196                                 portals_debug_dumplog();
 197                         next = jiffies + 20 * HZ;
 198                         spin_lock(&obd->obd_osfs_lock);
 199                 }
 200                 left = 0;
 201         }
 202
 203         CDEBUG(D_CACHE, "%s: cli %s/%p free: "LPU64" avail: "LPU64" grant "LPU64
 204                " left: "LPU64" pending: "LPU64"\n", obd->obd_name,
 205                exp->exp_client_uuid.uuid, exp,
 206                obd->obd_osfs.os_bfree << blockbits, avail << blockbits,
 207                tot_granted, left, obd->u.filter.fo_tot_pending);
 208
 209         return left;
 210 }
 211
 212 /* Calculate how much grant space to allocate to this client, based on how
 213  * much space is currently free and how much of that is already granted.
 214  *
 215  * Caller must hold obd_osfs_lock. */
 216 long filter_grant(struct obd_export *exp, obd_size current_grant,
 217                   obd_size want, obd_size fs_space_left)
 218 {
 219         struct obd_device *obd = exp->exp_obd;
 220         struct filter_export_data *fed = &exp->exp_filter_data;
 221         int blockbits = obd->u.filter.fo_sb->s_blocksize_bits;
 222         __u64 grant = 0;
 223
 224         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
 225
 226         /* Grant some fraction of the client's requested grant space so that
 227          * they are not always waiting for write credits (not all of it to
 228          * avoid overgranting in face of multiple RPCs in flight).  This
 229          * essentially will be able to control the OSC_MAX_RIF for a client.
 230          *
 231          * If we do have a large disparity between what the client thinks it
 232          * has and what we think it has, don't grant very much and let the
 233          * client consume its grant first.  Either it just has lots of RPCs
 234          * in flight, or it was evicted and its grants will soon be used up. */
 235         if (current_grant < want &&
 236             current_grant < fed->fed_grant + FILTER_GRANT_CHUNK) {
 237                 grant = min((want >> blockbits) / 2,
 238                             (fs_space_left >> blockbits) / 8);
 239                 grant <<= blockbits;
 240
 241                 if (grant) {
 242                         if (grant > FILTER_GRANT_CHUNK)
 243                                 grant = FILTER_GRANT_CHUNK;
 244
 245                         obd->u.filter.fo_tot_granted += grant;
 246                         fed->fed_grant += grant;
 247                 }
 248         }
 249
 250         CDEBUG(D_CACHE,"%s: cli %s/%p wants: "LPU64" granting: "LPU64"\n",
 251                obd->obd_name, exp->exp_client_uuid.uuid, exp, want, grant);
 252         CDEBUG(D_CACHE,
 253                "%s: cli %s/%p tot cached:"LPU64" granted:"LPU64
 254                " num_exports: %d\n", obd->obd_name, exp->exp_client_uuid.uuid,
 255                exp, obd->u.filter.fo_tot_dirty,
 256                obd->u.filter.fo_tot_granted, obd->obd_num_exports);
 257
 258         return grant;
 259 }
 260
 261
 262 static int filter_preprw_read(int cmd, struct obd_export *exp, struct obdo *oa,
 263                               int objcount, struct obd_ioobj *obj,
 264                               int niocount, struct niobuf_remote *nb,
 265                               struct niobuf_local *res,
 266                               struct obd_trans_info *oti)
 267 {
 268         struct obd_device *obd = exp->exp_obd;
 269         struct lvfs_run_ctxt saved;
 270         struct niobuf_remote *rnb;
 271         struct niobuf_local *lnb;
 272         struct dentry *dentry = NULL;
 273         struct inode *inode;
 274         void *iobuf = NULL;
 275         int rc = 0, i, tot_bytes = 0;
 276         unsigned long now = jiffies;
 277         ENTRY;
 278
 279         /* We are currently not supporting multi-obj BRW_READ RPCS at all.
 280          * When we do this function's dentry cleanup will need to be fixed */
 281         LASSERTF(objcount == 1, "%d\n", objcount);
 282         LASSERTF(obj->ioo_bufcnt > 0, "%d\n", obj->ioo_bufcnt);
 283
 284         if (oa && oa->o_valid & OBD_MD_FLGRANT) {
 285                 spin_lock(&obd->obd_osfs_lock);
 286                 filter_grant_incoming(exp, oa);
 287
 288                 oa->o_grant = 0;
 289
 290                 spin_unlock(&obd->obd_osfs_lock);
 291         }
 292
 293         memset(res, 0, niocount * sizeof(*res));
 294
 295         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 296         rc = filter_alloc_iobuf(OBD_BRW_READ, obj->ioo_bufcnt, &iobuf);
 297         if (rc)
 298                 GOTO(cleanup, rc);
 299
 300         dentry = filter_oa2dentry(obd, oa);
 301         if (IS_ERR(dentry))
 302                 GOTO(cleanup, rc = PTR_ERR(dentry));
 303
 304         if (dentry->d_inode == NULL) {
 305                 CERROR("trying to BRW to non-existent file "LPU64"\n",
 306                                obj->ioo_id);
 307                 GOTO(cleanup, rc = -ENOENT);
 308         }
 309
 310         inode = dentry->d_inode;
 311
 312         if (time_after(jiffies, now + 15 * HZ))
 313                 CERROR("slow preprw_read setup %lus\n", (jiffies - now) / HZ);
 314         else
 315                 CDEBUG(D_INFO, "preprw_read setup: %lu jiffies\n",
 316                        (jiffies - now));
 317
 318         for (i = 0, lnb = res, rnb = nb; i < obj->ioo_bufcnt;
 319              i++, rnb++, lnb++) {
 320                 lnb->dentry = dentry;
 321                 lnb->offset = rnb->offset;
 322                 lnb->len    = rnb->len;
 323                 lnb->flags  = rnb->flags;
 324
 325                 if (inode->i_size <= rnb->offset)
 326                       /* If there's no more data, abort early.
 327                       * lnb->page == NULL and lnb->rc == 0, so it's
 328                       * easy to detect later. */
 329                         break;
 330                 else
 331                         rc = filter_alloc_dio_page(obd, inode, lnb);
 332                 if (rc) {
 333                         CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
 334                              "page err %u@"LPU64" %u/%u %p: rc %d\n",
 335                               lnb->len, lnb->offset, i, obj->ioo_bufcnt,
 336                               dentry, rc);
 337                         GOTO(cleanup, rc);
 338                 }
 339
 340                 if (inode->i_size < lnb->offset + lnb->len - 1)
 341                         lnb->rc = inode->i_size - lnb->offset;
 342                 else
 343                         lnb->rc = lnb->len;
 344
 345                 tot_bytes += lnb->rc;
 346
 347                 filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 348         }
 349
 350         if (time_after(jiffies, now + 15 * HZ))
 351                 CERROR("slow start_page_read %lus\n", (jiffies - now) / HZ);
 352         else
 353                 CDEBUG(D_INFO, "start_page_read: %lu jiffies\n",
 354                        (jiffies - now));
 355
 356         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf, exp,
 357                               NULL, NULL, NULL);
 358         if (rc)
 359                 GOTO(cleanup, rc);
 360
 361         lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_READ_BYTES, tot_bytes);
 362
 363         filter_tally_read(&exp->exp_obd->u.filter, res, niocount);
 364
 365         EXIT;
 366
 367 cleanup:
 368         if (rc != 0) {
 369                 filter_free_dio_pages(objcount, obj, niocount, res);
 370
 371                 if (dentry != NULL)
 372                         f_dput(dentry);
 373                 else
 374                         CERROR("NULL dentry in cleanup -- tell CFS\n");
 375         }
 376
 377         if (iobuf != NULL)
 378                 filter_free_iobuf(iobuf);
 379
 380         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 381         if (rc)
 382                 CERROR("io error %d\n", rc);
 383         return rc;
 384 }
 385
 386 /* When clients have dirtied as much space as they've been granted they
 387  * fall through to sync writes.  These sync writes haven't been expressed
 388  * in grants and need to error with ENOSPC when there isn't room in the
 389  * filesystem for them after grants are taken into account.  However,
 390  * writeback of the dirty data that was already granted space can write
 391  * right on through.
 392  *
 393  * Caller must hold obd_osfs_lock. */
 394 static int filter_grant_check(struct obd_export *exp, int objcount,
 395                               struct fsfilt_objinfo *fso, int niocount,
 396                               struct niobuf_remote *rnb,
 397                               struct niobuf_local *lnb, obd_size *left,
 398                               struct inode *inode)
 399 {
 400         struct filter_export_data *fed = &exp->exp_filter_data;
 401         int blocksize = exp->exp_obd->u.filter.fo_sb->s_blocksize;
 402         unsigned long used = 0, ungranted = 0, using;
 403         int i, rc = -ENOSPC, obj, n = 0, mask = D_CACHE;
 404
 405         LASSERT_SPIN_LOCKED(&exp->exp_obd->obd_osfs_lock);
 406
 407         for (obj = 0; obj < objcount; obj++) {
 408                 for (i = 0; i < fso[obj].fso_bufcnt; i++, n++) {
 409                         int tmp, bytes;
 410
 411                         /* FIXME: this is calculated with PAGE_SIZE on client */
 412                         bytes = rnb[n].len;
 413                         bytes += rnb[n].offset & (blocksize - 1);
 414                         tmp = (rnb[n].offset + rnb[n].len) & (blocksize - 1);
 415                         if (tmp)
 416                                 bytes += blocksize - tmp;
 417
 418                         if (rnb[n].flags & OBD_BRW_FROM_GRANT) {
 419                                 if (fed->fed_grant < used + bytes) {
 420                                         CDEBUG(D_CACHE,
 421                                                "%s: cli %s/%p claims %ld+%d "
 422                                                "GRANT, real grant %lu idx %d\n",
 423                                                exp->exp_obd->obd_name,
 424                                                exp->exp_client_uuid.uuid, exp,
 425                                                used, bytes, fed->fed_grant, n);
 426                                         mask = D_ERROR;
 427                                 } else {
 428                                         used += bytes;
 429                                         rnb[n].flags |= OBD_BRW_GRANTED;
 430                                         lnb[n].lnb_grant_used = bytes;
 431                                         CDEBUG(0, "idx %d used=%lu\n", n, used);
 432                                         rc = 0;
 433                                         continue;
 434                                 }
 435                         }
 436                         if (*left > ungranted) {
 437                                 /* if enough space, pretend it was granted */
 438                                 ungranted += bytes;
 439                                 rnb[n].flags |= OBD_BRW_GRANTED;
 440                                 CDEBUG(0, "idx %d ungranted=%lu\n",n,ungranted);
 441                                 rc = 0;
 442                                 continue;
 443                         }
 444
 445                         /* We can't check for already-mapped blocks here, as
 446                          * it requires dropping the osfs lock to do the bmap.
 447                          * Instead, we return ENOSPC and in that case we need
 448                          * to go through and verify if all of the blocks not
 449                          * marked BRW_GRANTED are already mapped and we can
 450                          * ignore this error. */
 451                         lnb[n].rc = -ENOSPC;
 452                         rnb[n].flags &= OBD_BRW_GRANTED;
 453                         CDEBUG(D_CACHE,"%s: cli %s/%p idx %d no space for %d\n",
 454                                exp->exp_obd->obd_name,
 455                                exp->exp_client_uuid.uuid, exp, n, bytes);
 456                 }
 457         }
 458
 459         /* Now substract what client have used already.  We don't subtract
 460          * this from the tot_granted yet, so that other client's can't grab
 461          * that space before we have actually allocated our blocks.  That
 462          * happens in filter_grant_commit() after the writes are done. */
 463         *left -= ungranted;
 464         fed->fed_grant -= used;
 465         fed->fed_pending += used;
 466         exp->exp_obd->u.filter.fo_tot_pending += used;
 467
 468         CDEBUG(mask,
 469                "%s: cli %s/%p used: %lu ungranted: %lu grant: %lu dirty: %lu\n",
 470                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp, used,
 471                ungranted, fed->fed_grant, fed->fed_dirty);
 472
 473         /* Rough calc in case we don't refresh cached statfs data */
 474         using = (used + ungranted + 1 ) >>
 475                 exp->exp_obd->u.filter.fo_sb->s_blocksize_bits;
 476         if (exp->exp_obd->obd_osfs.os_bavail > using)
 477                 exp->exp_obd->obd_osfs.os_bavail -= using;
 478         else
 479                 exp->exp_obd->obd_osfs.os_bavail = 0;
 480
 481         if (fed->fed_dirty < used) {
 482                 CERROR("%s: cli %s/%p claims used %lu > fed_dirty %lu\n",
 483                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 484                        used, fed->fed_dirty);
 485                 used = fed->fed_dirty;
 486         }
 487         exp->exp_obd->u.filter.fo_tot_dirty -= used;
 488         fed->fed_dirty -= used;
 489
 490         return rc;
 491 }
 492
 493 /* If we ever start to support multi-object BRW RPCs, we will need to get locks
 494  * on mulitple inodes.  That isn't all, because there still exists the
 495  * possibility of a truncate starting a new transaction while holding the ext3
 496  * rwsem = write while some writes (which have started their transactions here)
 497  * blocking on the ext3 rwsem = read => lock inversion.
 498  *
 499  * The handling gets very ugly when dealing with locked pages.  It may be easier
 500  * to just get rid of the locked page code (which has problems of its own) and
 501  * either discover we do not need it anymore (i.e. it was a symptom of another
 502  * bug) or ensure we get the page locks in an appropriate order. */
 503 static int filter_preprw_write(int cmd, struct obd_export *exp, struct obdo *oa,
 504                                int objcount, struct obd_ioobj *obj,
 505                                int niocount, struct niobuf_remote *nb,
 506                                struct niobuf_local *res,
 507                                struct obd_trans_info *oti)
 508 {
 509         struct lvfs_run_ctxt saved;
 510         struct niobuf_remote *rnb;
 511         struct niobuf_local *lnb = res;
 512         struct fsfilt_objinfo fso;
 513         struct dentry *dentry = NULL;
 514         void *iobuf;
 515         obd_size left;
 516         unsigned long now = jiffies;
 517         int rc = 0, i, tot_bytes = 0, cleanup_phase = 0;
 518         ENTRY;
 519         LASSERT(objcount == 1);
 520         LASSERT(obj->ioo_bufcnt > 0);
 521
 522         memset(res, 0, niocount * sizeof(*res));
 523
 524         rc = filter_alloc_iobuf(OBD_BRW_READ, obj->ioo_bufcnt, &iobuf);
 525         if (rc)
 526                 GOTO(cleanup, rc);
 527         cleanup_phase = 1;
 528
 529         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 530         dentry = filter_fid2dentry(exp->exp_obd, NULL, obj->ioo_gr,
 531                                    obj->ioo_id);
 532         if (IS_ERR(dentry))
 533                 GOTO(cleanup, rc = PTR_ERR(dentry));
 534
 535         cleanup_phase = 2;
 536
 537         if (dentry->d_inode == NULL) {
 538                 CERROR("trying to BRW to non-existent file "LPU64"\n",
 539                        obj->ioo_id);
 540                 f_dput(dentry);
 541                 GOTO(cleanup, rc = -ENOENT);
 542         }
 543
 544         fso.fso_dentry = dentry;
 545         fso.fso_bufcnt = obj->ioo_bufcnt;
 546
 547         if (time_after(jiffies, now + 15 * HZ))
 548                 CERROR("slow preprw_write setup %lus\n", (jiffies - now) / HZ);
 549         else
 550                 CDEBUG(D_INFO, "preprw_write setup: %lu jiffies\n",
 551                        (jiffies - now));
 552
 553         spin_lock(&exp->exp_obd->obd_osfs_lock);
 554         if (oa)
 555                 filter_grant_incoming(exp, oa);
 556
 557         cleanup_phase = 3;
 558
 559         left = filter_grant_space_left(exp);
 560
 561         rc = filter_grant_check(exp, objcount, &fso, niocount, nb, res,
 562                                 &left, dentry->d_inode);
 563         if (oa && oa->o_valid & OBD_MD_FLGRANT)
 564                 oa->o_grant = filter_grant(exp,oa->o_grant,oa->o_undirty,left);
 565
 566         spin_unlock(&exp->exp_obd->obd_osfs_lock);
 567
 568         if (rc)
 569                 GOTO(cleanup, rc);
 570
 571         for (i = 0, rnb = nb, lnb = res; i < obj->ioo_bufcnt;
 572              i++, lnb++, rnb++) {
 573                 /* We still set up for ungranted pages so that granted pages
 574                  * can be written to disk as they were promised, and portals
 575                  * needs to keep the pages all aligned properly. */
 576                 lnb->dentry = dentry;
 577                 lnb->offset = rnb->offset;
 578                 lnb->len    = rnb->len;
 579                 lnb->flags  = rnb->flags;
 580
 581                 rc = filter_alloc_dio_page(exp->exp_obd, dentry->d_inode,lnb);
 582                 if (rc) {
 583                         CERROR("page err %u@"LPU64" %u/%u %p: rc %d\n",
 584                                lnb->len, lnb->offset,
 585                                i, obj->ioo_bufcnt, dentry, rc);
 586                         GOTO(cleanup, rc);
 587                 }
 588                 cleanup_phase = 4;
 589
 590                 /* If the filter writes a partial page, then has the file
 591                  * extended, the client will read in the whole page.  the
 592                  * filter has to be careful to zero the rest of the partial
 593                  * page on disk.  we do it by hand for partial extending
 594                  * writes, send_bio() is responsible for zeroing pages when
 595                  * asked to read unmapped blocks -- brw_kiovec() does this. */
 596                 if (lnb->len != PAGE_SIZE) {
 597                         if (lnb->offset + lnb->len < dentry->d_inode->i_size) {
 598                                 filter_iobuf_add_page(exp->exp_obd, iobuf,
 599                                                       dentry->d_inode,
 600                                                       lnb->page);
 601                         } else {
 602                                 memset(kmap(lnb->page) + lnb->len, 0,
 603                                        PAGE_SIZE - lnb->len);
 604                                 kunmap(lnb->page);
 605                         }
 606                 }
 607                 if (lnb->rc == 0)
 608                         tot_bytes += lnb->len;
 609         }
 610
 611         rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf, exp,
 612                               NULL, NULL, NULL);
 613
 614         if (time_after(jiffies, now + 15 * HZ))
 615                 CERROR("slow start_page_write %lus\n", (jiffies - now) / HZ);
 616         else
 617                 CDEBUG(D_INFO, "start_page_write: %lu jiffies\n",
 618                        (jiffies - now));
 619
 620         lprocfs_counter_add(exp->exp_obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
 621                             tot_bytes);
 622         EXIT;
 623 cleanup:
 624         switch(cleanup_phase) {
 625         case 4:
 626                 if (rc)
 627                         filter_free_dio_pages(objcount, obj, niocount, res);
 628         case 3:
 629                 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 630                 filter_free_iobuf(iobuf);
 631         case 2:
 632                 if (rc)
 633                         f_dput(dentry);
 634                 break;
 635         case 1:
 636                 spin_lock(&exp->exp_obd->obd_osfs_lock);
 637                 if (oa)
 638                         filter_grant_incoming(exp, oa);
 639                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
 640                 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 641                 filter_free_iobuf(iobuf);
 642                 break;
 643         default:;
 644
 645         }
 646         RETURN(rc);
 647 }
 648
 649 int filter_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
 650                   int objcount, struct obd_ioobj *obj, int niocount,
 651                   struct niobuf_remote *nb, struct niobuf_local *res,
 652                   struct obd_trans_info *oti)
 653 {
 654         if (cmd == OBD_BRW_WRITE)
 655                 return filter_preprw_write(cmd, exp, oa, objcount, obj,
 656                                            niocount, nb, res, oti);
 657
 658         if (cmd == OBD_BRW_READ)
 659                 return filter_preprw_read(cmd, exp, oa, objcount, obj,
 660                                           niocount, nb, res, oti);
 661
 662         LBUG();
 663         return -EPROTO;
 664 }
 665
 666 void filter_release_read_page(struct filter_obd *filter, struct inode *inode,
 667                               struct page *page)
 668 {
 669         int drop = 0;
 670
 671         if (inode != NULL &&
 672             (inode->i_size > filter->fo_readcache_max_filesize))
 673                 drop = 1;
 674
 675         /* drop from cache like truncate_list_pages() */
 676         if (drop && !TryLockPage(page)) {
 677                 if (page->mapping)
 678                         ll_truncate_complete_page(page);
 679                 unlock_page(page);
 680         }
 681         page_cache_release(page);
 682 }
 683
 684 static int filter_commitrw_read(struct obd_export *exp, struct obdo *oa,
 685                                 int objcount, struct obd_ioobj *obj,
 686                                 int niocount, struct niobuf_local *res,
 687                                 struct obd_trans_info *oti, int rc)
 688 {
 689         struct inode *inode = NULL;
 690         ENTRY;
 691
 692         if (res->dentry != NULL)
 693                 inode = res->dentry->d_inode;
 694
 695         filter_free_dio_pages(objcount, obj, niocount, res);
 696
 697         if (res->dentry != NULL)
 698                 f_dput(res->dentry);
 699         RETURN(rc);
 700 }
 701
 702 void flip_into_page_cache(struct inode *inode, struct page *new_page)
 703 {
 704         struct page *old_page;
 705         int rc;
 706
 707         do {
 708                 /* the dlm is protecting us from read/write concurrency, so we
 709                  * expect this find_lock_page to return quickly.  even if we
 710                  * race with another writer it won't be doing much work with
 711                  * the page locked.  we do this 'cause t_c_p expects a
 712                  * locked page, and it wants to grab the pagecache lock
 713                  * as well. */
 714                 old_page = find_lock_page(inode->i_mapping, new_page->index);
 715                 if (old_page) {
 716                         ll_truncate_complete_page(old_page);
 717                         unlock_page(old_page);
 718                         page_cache_release(old_page);
 719                 }
 720
 721 #if 0 /* this should be a /proc tunable someday */
 722                 /* racing o_directs (no locking ioctl) could race adding
 723                  * their pages, so we repeat the page invalidation unless
 724                  * we successfully added our new page */
 725                 rc = add_to_page_cache_unique(new_page, inode->i_mapping,
 726                                               new_page->index,
 727                                               page_hash(inode->i_mapping,
 728                                                         new_page->index));
 729                 if (rc == 0) {
 730                         /* add_to_page_cache clears uptodate|dirty and locks
 731                          * the page */
 732                         SetPageUptodate(new_page);
 733                         unlock_page(new_page);
 734                 }
 735 #else
 736                 rc = 0;
 737 #endif
 738         } while (rc != 0);
 739 }
 740
 741 void filter_grant_commit(struct obd_export *exp, int niocount,
 742                          struct niobuf_local *res)
 743 {
 744         struct filter_obd *filter = &exp->exp_obd->u.filter;
 745         struct niobuf_local *lnb = res;
 746         unsigned long pending = 0;
 747         int i;
 748
 749         spin_lock(&exp->exp_obd->obd_osfs_lock);
 750         for (i = 0, lnb = res; i < niocount; i++, lnb++)
 751                 pending += lnb->lnb_grant_used;
 752
 753         LASSERTF(exp->exp_filter_data.fed_pending >= pending,
 754                  "%s: cli %s/%p fed_pending: %lu grant_used: %lu\n",
 755                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 756                  exp->exp_filter_data.fed_pending, pending);
 757         exp->exp_filter_data.fed_pending -= pending;
 758         LASSERTF(filter->fo_tot_granted >= pending,
 759                  "%s: cli %s/%p tot_granted: "LPU64" grant_used: %lu\n",
 760                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 761                  exp->exp_obd->u.filter.fo_tot_granted, pending);
 762         filter->fo_tot_granted -= pending;
 763         LASSERTF(filter->fo_tot_pending >= pending,
 764                  "%s: cli %s/%p tot_pending: "LPU64" grant_used: %lu\n",
 765                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 766                  filter->fo_tot_pending, pending);
 767         filter->fo_tot_pending -= pending;
 768
 769         spin_unlock(&exp->exp_obd->obd_osfs_lock);
 770 }
 771 int filter_do_cow(struct obd_export *exp, struct obd_ioobj *obj,
 772                   int nioo, struct niobuf_remote *rnb)
 773 {
 774         struct dentry *dentry;
 775         struct lvfs_run_ctxt saved;
 776         struct write_extents *extents = NULL;
 777         int j, rc = 0, numexts = 0, flags = 0;
 778
 779         ENTRY;
 780
 781         LASSERT(nioo == 1);
 782
 783         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 784
 785         dentry = filter_fid2dentry(exp->exp_obd, NULL, obj->ioo_gr,
 786                                    obj->ioo_id);
 787         if (IS_ERR(dentry)) {
 788                 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 789                 RETURN (PTR_ERR(dentry));
 790         }
 791
 792         if (dentry->d_inode == NULL) {
 793                 CERROR("trying to write extents to non-existent file "LPU64"\n",
 794                        obj->ioo_id);
 795                 GOTO(cleanup, rc = -ENOENT);
 796         }
 797
 798         flags = fsfilt_get_fs_flags(exp->exp_obd, dentry);
 799         if (!(flags & SM_DO_COW)) {
 800                 GOTO(cleanup, rc);
 801         }
 802         OBD_ALLOC(extents, obj->ioo_bufcnt * sizeof(struct write_extents));
 803         if (!extents) {
 804                 CERROR("No Memory\n");
 805                 GOTO(cleanup, rc = -ENOMEM);
 806         }
 807         for (j = 0; j < obj->ioo_bufcnt; j++) {
 808                 if (rnb[j].len != 0) {
 809                         extents[numexts].w_count = rnb[j].len;
 810                         extents[numexts].w_pos = rnb[j].offset;
 811                         numexts++;
 812                 }
 813         }
 814         rc = fsfilt_do_write_cow(exp->exp_obd, dentry, extents, numexts);
 815         if (rc) {
 816                 CERROR("Do cow error id "LPU64" rc:%d \n",
 817                         obj->ioo_id, rc);
 818                 GOTO(cleanup, rc);
 819         }
 820
 821 cleanup:
 822         if (extents) {
 823                 OBD_FREE(extents, obj->ioo_bufcnt * sizeof(struct write_extents));
 824         }
 825         f_dput(dentry);
 826         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 827         RETURN(rc);
 828
 829 }
 830 int filter_write_extents(struct obd_export *exp, struct obd_ioobj *obj, int nobj,
 831                          int niocount, struct niobuf_local *local, int rc)
 832 {
 833         struct lvfs_run_ctxt saved;
 834         struct dentry *dentry;
 835         struct niobuf_local *lnb;
 836         __u64  offset = 0;
 837         __u32  len = 0;
 838         int    i, flags;
 839
 840         ENTRY;
 841
 842         LASSERT(nobj == 1);
 843
 844         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 845
 846         dentry = filter_fid2dentry(exp->exp_obd, NULL, obj->ioo_gr,
 847                                    obj->ioo_id);
 848         if (IS_ERR(dentry)) {
 849                 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 850                 RETURN (PTR_ERR(dentry));
 851         }
 852
 853         if (dentry->d_inode == NULL) {
 854                 CERROR("trying to write extents to non-existent file "LPU64"\n",
 855                        obj->ioo_id);
 856                 GOTO(cleanup, rc = -ENOENT);
 857         }
 858
 859         flags = fsfilt_get_fs_flags(exp->exp_obd, dentry);
 860         if (!(flags & SM_DO_REC)) {
 861                 GOTO(cleanup, rc);
 862         }
 863
 864         for (i = 0, lnb = local; i < obj->ioo_bufcnt; i++, lnb++) {
 865                 if (len == 0) {
 866                         offset = lnb->offset;
 867                         len = lnb->len;
 868                 } else if (lnb->offset == (offset + len)) {
 869                         len += lnb->len;
 870                 } else {
 871                         rc = fsfilt_write_extents(exp->exp_obd, dentry,
 872                                                   offset, len);
 873                         if (rc) {
 874                                 CERROR("write exts off "LPU64" num %u rc:%d\n",
 875                                         offset, len, rc);
 876                                 GOTO(cleanup, rc);
 877                         }
 878                         offset = lnb->offset;
 879                         len = lnb->len;
 880                 }
 881         }
 882         if (len > 0) {
 883                 rc = fsfilt_write_extents(exp->exp_obd, dentry,
 884                                           offset, len);
 885                 if (rc) {
 886                         CERROR("write exts off "LPU64" num %u rc:%d\n",
 887                                 offset, len, rc);
 888                         GOTO(cleanup, rc);
 889                 }
 890         }
 891 cleanup:
 892         f_dput(dentry);
 893         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 894         RETURN(rc);
 895 }
 896
 897 int filter_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
 898                     int objcount, struct obd_ioobj *obj, int niocount,
 899                     struct niobuf_local *res, struct obd_trans_info *oti,int rc)
 900 {
 901         if (cmd == OBD_BRW_WRITE)
 902                 return filter_commitrw_write(exp, oa, objcount, obj, niocount,
 903                                              res, oti, rc);
 904         if (cmd == OBD_BRW_READ)
 905                 return filter_commitrw_read(exp, oa, objcount, obj, niocount,
 906                                             res, oti, rc);
 907         LBUG();
 908         return -EPROTO;
 909 }
 910
 911 int filter_brw(int cmd, struct obd_export *exp, struct obdo *oa,
 912                struct lov_stripe_md *lsm, obd_count oa_bufs,
 913                struct brw_page *pga, struct obd_trans_info *oti)
 914 {
 915         struct obd_ioobj ioo;
 916         struct niobuf_local *lnb;
 917         struct niobuf_remote *rnb;
 918         obd_count i;
 919         int ret = 0;
 920         ENTRY;
 921
 922         OBD_ALLOC(lnb, oa_bufs * sizeof(struct niobuf_local));
 923         OBD_ALLOC(rnb, oa_bufs * sizeof(struct niobuf_remote));
 924
 925         if (lnb == NULL || rnb == NULL)
 926                 GOTO(out, ret = -ENOMEM);
 927
 928         for (i = 0; i < oa_bufs; i++) {
 929                 rnb[i].offset = pga[i].disk_offset;
 930                 rnb[i].len = pga[i].count;
 931         }
 932
 933         obdo_to_ioobj(oa, &ioo);
 934         ioo.ioo_bufcnt = oa_bufs;
 935
 936         ret = filter_preprw(cmd, exp, oa, 1, &ioo, oa_bufs, rnb, lnb, oti);
 937         if (ret != 0)
 938                 GOTO(out, ret);
 939
 940         for (i = 0; i < oa_bufs; i++) {
 941                 void *virt;
 942                 obd_off off;
 943                 void *addr;
 944
 945                 if (lnb[i].page == NULL)
 946                         break;
 947
 948                 off = pga[i].disk_offset & ~PAGE_MASK;
 949                 virt = kmap(pga[i].pg);
 950                 addr = kmap(lnb[i].page);
 951
 952                 /* 2 kmaps == vanishingly small deadlock opportunity */
 953
 954                 if (cmd & OBD_BRW_WRITE)
 955                         memcpy(addr + off, virt + off, pga[i].count);
 956                 else
 957                         memcpy(virt + off, addr + off, pga[i].count);
 958
 959                 kunmap(lnb[i].page);
 960                 kunmap(pga[i].pg);
 961         }
 962
 963         ret = filter_commitrw(cmd, exp, oa, 1, &ioo, oa_bufs, lnb, oti, ret);
 964
 965 out:
 966         if (lnb)
 967                 OBD_FREE(lnb, oa_bufs * sizeof(struct niobuf_local));
 968         if (rnb)
 969                 OBD_FREE(rnb, oa_bufs * sizeof(struct niobuf_remote));
 970         RETURN(ret);
 971 }