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         return 0;
  56 }
  57
  58 static int filter_finish_page_read(struct niobuf_local *lnb)
  59 {
  60         if (lnb->page == NULL)
  61                 return 0;
  62
  63         if (PageUptodate(lnb->page))
  64                 return 0;
  65
  66         wait_on_page(lnb->page);
  67         if (!PageUptodate(lnb->page)) {
  68                 CERROR("page index %lu/offset "LPX64" not uptodate\n",
  69                        lnb->page->index, lnb->offset);
  70                 GOTO(err_page, lnb->rc = -EIO);
  71         }
  72         if (PageError(lnb->page)) {
  73                 CERROR("page index %lu/offset "LPX64" has error\n",
  74                        lnb->page->index, lnb->offset);
  75                 GOTO(err_page, lnb->rc = -EIO);
  76         }
  77
  78         return 0;
  79
  80 err_page:
  81         page_cache_release(lnb->page);
  82         lnb->page = NULL;
  83         return lnb->rc;
  84 }
  85
  86 /* Grab the dirty and seen grant announcements from the incoming obdo.
  87  * We will later calculate the clients new grant and return it.
  88  * Caller must hold osfs lock */
  89 static void filter_grant_incoming(struct obd_export *exp, struct obdo *oa)
  90 {
  91         struct filter_export_data *fed;
  92         struct obd_device *obd = exp->exp_obd;
  93         static unsigned long last_msg;
  94         static int last_count;
  95         int mask = D_CACHE;
  96         ENTRY;
  97
  98         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
  99
 100         if ((oa->o_valid & (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) !=
 101                                         (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) {
 102                 oa->o_valid &= ~OBD_MD_FLGRANT;
 103                 EXIT;
 104                 return;
 105         }
 106
 107         fed = &exp->exp_filter_data;
 108
 109         /* Don't print this to the console the first time it happens, since
 110          * it can happen legitimately on occasion, but only rarely. */
 111         if (time_after(jiffies, last_msg + 60 * HZ)) {
 112                 last_count = 0;
 113                 last_msg = jiffies;
 114         }
 115         if ((last_count & (-last_count)) == last_count)
 116                 mask = D_WARNING;
 117         last_count++;
 118
 119         /* Add some margin, since there is a small race if other RPCs arrive
 120          * out-or-order and have already consumed some grant.  We want to
 121          * leave this here in case there is a large error in accounting. */
 122         CDEBUG(oa->o_grant > fed->fed_grant + FILTER_GRANT_CHUNK ? mask:D_CACHE,
 123                "%s: cli %s/%p reports grant: "LPU64" dropped: %u, local: %lu\n",
 124                obd->obd_name, exp->exp_client_uuid.uuid, exp, oa->o_grant,
 125                oa->o_dropped, fed->fed_grant);
 126
 127         /* Update our accounting now so that statfs takes it into account.
 128          * Note that fed_dirty is only approximate and can become incorrect
 129          * if RPCs arrive out-of-order.  No important calculations depend
 130          * on fed_dirty however. */
 131         obd->u.filter.fo_tot_dirty += oa->o_dirty - fed->fed_dirty;
 132         if (fed->fed_grant < oa->o_dropped) {
 133                 CERROR("%s: cli %s/%p reports %u dropped > fed_grant %lu\n",
 134                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
 135                        oa->o_dropped, fed->fed_grant);
 136                 oa->o_dropped = 0;
 137         }
 138         if (obd->u.filter.fo_tot_granted < oa->o_dropped) {
 139                 CERROR("%s: cli %s/%p reports %u dropped > tot_grant "LPU64"\n",
 140                        obd->obd_name, exp->exp_client_uuid.uuid, exp,
 141                        oa->o_dropped, obd->u.filter.fo_tot_granted);
 142                 oa->o_dropped = 0;
 143         }
 144         obd->u.filter.fo_tot_granted -= oa->o_dropped;
 145         fed->fed_grant -= oa->o_dropped;
 146         fed->fed_dirty = oa->o_dirty;
 147         EXIT;
 148 }
 149
 150 #define GRANT_FOR_LLOG(obd) 16
 151
 152 /* Figure out how much space is available between what we've granted
 153  * and what remains in the filesystem.  Compensate for ext3 indirect
 154  * block overhead when computing how much free space is left ungranted.
 155  *
 156  * Caller must hold obd_osfs_lock. */
 157 obd_size filter_grant_space_left(struct obd_export *exp)
 158 {
 159         struct obd_device *obd = exp->exp_obd;
 160         int blockbits = obd->u.filter.fo_sb->s_blocksize_bits;
 161         obd_size tot_granted = obd->u.filter.fo_tot_granted, avail, left = 0;
 162         int rc, statfs_done = 0;
 163
 164         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
 165
 166         if (time_before(obd->obd_osfs_age, jiffies - HZ)) {
 167 restat:
 168                 rc = fsfilt_statfs(obd, obd->u.filter.fo_sb, jiffies + 1);
 169                 if (rc) /* N.B. statfs can't really fail */
 170                         RETURN(0);
 171                 statfs_done = 1;
 172         }
 173
 174         avail = obd->obd_osfs.os_bavail;
 175         left = avail - (avail >> (blockbits - 3)); /* (d)indirect */
 176         if (left > GRANT_FOR_LLOG(obd)) {
 177                 left = (left - GRANT_FOR_LLOG(obd)) << blockbits;
 178         } else {
 179                 left = 0 /* << blockbits */;
 180         }
 181
 182         if (!statfs_done && left < 32 * FILTER_GRANT_CHUNK + tot_granted) {
 183                 CDEBUG(D_CACHE, "fs has no space left and statfs too old\n");
 184                 goto restat;
 185         }
 186
 187         if (left >= tot_granted) {
 188                 left -= tot_granted;
 189         } else {
 190                 static unsigned long next;
 191                 if (left < tot_granted - obd->u.filter.fo_tot_pending &&
 192                     time_after(jiffies, next)) {
 193                         spin_unlock(&obd->obd_osfs_lock);
 194                         CERROR("%s: cli %s/%p grant "LPU64" > available "
 195                                LPU64" and pending "LPU64"\n", obd->obd_name,
 196                                exp->exp_client_uuid.uuid, exp, tot_granted,
 197                                left, obd->u.filter.fo_tot_pending);
 198                         if (next == 0)
 199                                 portals_debug_dumplog();
 200                         next = jiffies + 20 * HZ;
 201                         spin_lock(&obd->obd_osfs_lock);
 202                 }
 203                 left = 0;
 204         }
 205
 206         CDEBUG(D_CACHE, "%s: cli %s/%p free: "LPU64" avail: "LPU64" grant "LPU64
 207                " left: "LPU64" pending: "LPU64"\n", obd->obd_name,
 208                exp->exp_client_uuid.uuid, exp,
 209                obd->obd_osfs.os_bfree << blockbits, avail << blockbits,
 210                tot_granted, left, obd->u.filter.fo_tot_pending);
 211
 212         return left;
 213 }
 214
 215 /* Calculate how much grant space to allocate to this client, based on how
 216  * much space is currently free and how much of that is already granted.
 217  *
 218  * Caller must hold obd_osfs_lock. */
 219 long filter_grant(struct obd_export *exp, obd_size current_grant,
 220                   obd_size want, obd_size fs_space_left)
 221 {
 222         struct obd_device *obd = exp->exp_obd;
 223         struct filter_export_data *fed = &exp->exp_filter_data;
 224         int blockbits = obd->u.filter.fo_sb->s_blocksize_bits;
 225         __u64 grant = 0;
 226
 227         LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
 228
 229         /* Grant some fraction of the client's requested grant space so that
 230          * they are not always waiting for write credits (not all of it to
 231          * avoid overgranting in face of multiple RPCs in flight).  This
 232          * essentially will be able to control the OSC_MAX_RIF for a client.
 233          *
 234          * If we do have a large disparity between what the client thinks it
 235          * has and what we think it has, don't grant very much and let the
 236          * client consume its grant first.  Either it just has lots of RPCs
 237          * in flight, or it was evicted and its grants will soon be used up. */
 238         if (current_grant < want &&
 239             current_grant < fed->fed_grant + FILTER_GRANT_CHUNK) {
 240                 grant = min((want >> blockbits) / 2,
 241                             (fs_space_left >> blockbits) / 8);
 242                 grant <<= blockbits;
 243
 244                 if (grant) {
 245                         if (grant > FILTER_GRANT_CHUNK)
 246                                 grant = FILTER_GRANT_CHUNK;
 247
 248                         obd->u.filter.fo_tot_granted += grant;
 249                         fed->fed_grant += grant;
 250                 }
 251         }
 252
 253         CDEBUG(D_CACHE,"%s: cli %s/%p wants: "LPU64" granting: "LPU64"\n",
 254                obd->obd_name, exp->exp_client_uuid.uuid, exp, want, grant);
 255         CDEBUG(D_CACHE,
 256                "%s: cli %s/%p tot cached:"LPU64" granted:"LPU64
 257                " num_exports: %d\n", obd->obd_name, exp->exp_client_uuid.uuid,
 258                exp, obd->u.filter.fo_tot_dirty,
 259                obd->u.filter.fo_tot_granted, obd->obd_num_exports);
 260
 261         return grant;
 262 }
 263
 264 static int filter_preprw_read(int cmd, struct obd_export *exp, struct obdo *oa,
 265                               int objcount, struct obd_ioobj *obj,
 266                               int niocount, struct niobuf_remote *nb,
 267                               struct niobuf_local *res,
 268                               struct obd_trans_info *oti)
 269 {
 270         struct obd_device *obd = exp->exp_obd;
 271         struct lvfs_run_ctxt saved;
 272         struct obd_ioobj *o;
 273         struct niobuf_remote *rnb;
 274         struct niobuf_local *lnb = NULL;
 275         struct fsfilt_objinfo *fso;
 276         struct dentry *dentry;
 277         struct inode *inode;
 278         int rc = 0, i, j, tot_bytes = 0, cleanup_phase = 0;
 279         unsigned long now = jiffies;
 280         ENTRY;
 281
 282         /* We are currently not supporting multi-obj BRW_READ RPCS at all.
 283          * When we do this function's dentry cleanup will need to be fixed */
 284         LASSERT(objcount == 1);
 285         LASSERT(obj->ioo_bufcnt > 0);
 286
 287         if (oa && oa->o_valid & OBD_MD_FLGRANT) {
 288                 spin_lock(&obd->obd_osfs_lock);
 289                 filter_grant_incoming(exp, oa);
 290
 291 #if 0
 292                 /* Reads do not increase grants */
 293                 oa->o_grant = filter_grant(exp, oa->o_grant, oa->o_undirty,
 294                                            filter_grant_space_left(exp));
 295 #else
 296                 oa->o_grant = 0;
 297 #endif
 298                 spin_unlock(&obd->obd_osfs_lock);
 299         }
 300
 301         OBD_ALLOC(fso, objcount * sizeof(*fso));
 302         if (fso == NULL)
 303                 RETURN(-ENOMEM);
 304
 305         memset(res, 0, niocount * sizeof(*res));
 306
 307         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 308         for (i = 0, o = obj; i < objcount; i++, o++) {
 309                 LASSERT(o->ioo_bufcnt);
 310
 311                 dentry = filter_oa2dentry(obd, oa);
 312                 if (IS_ERR(dentry))
 313                         GOTO(cleanup, rc = PTR_ERR(dentry));
 314
 315                 if (dentry->d_inode == NULL) {
 316                         CERROR("trying to BRW to non-existent file "LPU64"\n",
 317                                o->ioo_id);
 318                         f_dput(dentry);
 319                         GOTO(cleanup, rc = -ENOENT);
 320                 }
 321
 322                 fso[i].fso_dentry = dentry;
 323                 fso[i].fso_bufcnt = o->ioo_bufcnt;
 324         }
 325
 326         if (time_after(jiffies, now + 15 * HZ))
 327                 CERROR("slow preprw_read setup %lus\n", (jiffies - now) / HZ);
 328         else
 329                 CDEBUG(D_INFO, "preprw_read setup: %lu jiffies\n",
 330                        (jiffies - now));
 331
 332         for (i = 0, o = obj, rnb = nb, lnb = res; i < objcount; i++, o++) {
 333                 dentry = fso[i].fso_dentry;
 334                 inode = dentry->d_inode;
 335
 336                 for (j = 0; j < o->ioo_bufcnt; j++, rnb++, lnb++) {
 337                         lnb->dentry = dentry;
 338                         lnb->offset = rnb->offset;
 339                         lnb->len    = rnb->len;
 340                         lnb->flags  = rnb->flags;
 341
 342                         if (inode->i_size <= rnb->offset) {
 343                                 /* If there's no more data, abort early.
 344                                  * lnb->page == NULL and lnb->rc == 0, so it's
 345                                  * easy to detect later. */
 346                                 break;
 347                         } else {
 348                                 rc = filter_start_page_read(obd, inode, lnb);
 349                         }
 350
 351                         if (rc) {
 352                                 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
 353                                        "page err %u@"LPU64" %u/%u %p: rc %d\n",
 354                                        lnb->len, lnb->offset, j, o->ioo_bufcnt,
 355                                        dentry, rc);
 356                                 cleanup_phase = 1;
 357                                 GOTO(cleanup, rc);
 358                         }
 359
 360                         if (inode->i_size < lnb->offset + lnb->len - 1)
 361                                 lnb->rc = inode->i_size - lnb->offset;
 362                         else
 363                                 lnb->rc = lnb->len;
 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 obd_device *obd, struct inode *inode,
 527                                    struct niobuf_local *lnb)
 528 {
 529         struct page *page;
 530
 531         if (lnb->len != PAGE_SIZE)
 532                 return filter_start_page_read(obd, inode, lnb);
 533
 534         page = alloc_pages(GFP_HIGHUSER, 0);
 535         if (page == NULL) {
 536                 CERROR("no memory for a temp page\n");
 537                 RETURN(lnb->rc = -ENOMEM);
 538         }
 539 #if 0
 540         POISON_PAGE(page, 0xf1);
 541         if (lnb->len != PAGE_SIZE) {
 542                 memset(kmap(page) + lnb->len, 0, PAGE_SIZE - lnb->len);
 543                 kunmap(page);
 544         }
 545 #endif
 546         page->index = lnb->offset >> PAGE_SHIFT;
 547         lnb->page = page;
 548
 549         return 0;
 550 }
 551
 552 static void filter_abort_page_write(struct niobuf_local *lnb)
 553 {
 554         LASSERT(lnb->page != NULL);
 555
 556         if (lnb->len != PAGE_SIZE)
 557                 page_cache_release(lnb->page);
 558         else
 559                 __free_pages(lnb->page, 0);
 560 }
 561
 562 /* a helper for both the 2.4 and 2.6 commitrw paths which are both built
 563  * up by our shared filter_preprw_write() */
 564 void filter_release_write_page(struct filter_obd *filter, struct inode *inode,
 565                                struct niobuf_local *lnb, int rc)
 566 {
 567         if (lnb->len != PAGE_SIZE)
 568                 return filter_release_read_page(filter, inode, lnb->page);
 569
 570         if (rc == 0)
 571                 flip_into_page_cache(inode, lnb->page);
 572         __free_page(lnb->page);
 573 }
 574
 575 /* If we ever start to support multi-object BRW RPCs, we will need to get locks
 576  * on mulitple inodes.  That isn't all, because there still exists the
 577  * possibility of a truncate starting a new transaction while holding the ext3
 578  * rwsem = write while some writes (which have started their transactions here)
 579  * blocking on the ext3 rwsem = read => lock inversion.
 580  *
 581  * The handling gets very ugly when dealing with locked pages.  It may be easier
 582  * to just get rid of the locked page code (which has problems of its own) and
 583  * either discover we do not need it anymore (i.e. it was a symptom of another
 584  * bug) or ensure we get the page locks in an appropriate order. */
 585 static int filter_preprw_write(int cmd, struct obd_export *exp, struct obdo *oa,
 586                                int objcount, struct obd_ioobj *obj,
 587                                int niocount, struct niobuf_remote *nb,
 588                                struct niobuf_local *res,
 589                                struct obd_trans_info *oti)
 590 {
 591         struct lvfs_run_ctxt saved;
 592         struct niobuf_remote *rnb;
 593         struct niobuf_local *lnb;
 594         struct fsfilt_objinfo fso;
 595         struct dentry *dentry;
 596         obd_size left;
 597         unsigned long now = jiffies;
 598         int rc = 0, i, tot_bytes = 0, cleanup_phase = 1;
 599         ENTRY;
 600         LASSERT(objcount == 1);
 601         LASSERT(obj->ioo_bufcnt > 0);
 602
 603         memset(res, 0, niocount * sizeof(*res));
 604
 605         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 606         dentry = filter_fid2dentry(exp->exp_obd, NULL, obj->ioo_gr,
 607                                    obj->ioo_id);
 608         if (IS_ERR(dentry))
 609                 GOTO(cleanup, rc = PTR_ERR(dentry));
 610
 611         if (dentry->d_inode == NULL) {
 612                 CERROR("trying to BRW to non-existent file "LPU64"\n",
 613                        obj->ioo_id);
 614                 f_dput(dentry);
 615                 GOTO(cleanup, rc = -ENOENT);
 616         }
 617
 618         fso.fso_dentry = dentry;
 619         fso.fso_bufcnt = obj->ioo_bufcnt;
 620
 621         if (time_after(jiffies, now + 15 * HZ))
 622                 CERROR("slow preprw_write setup %lus\n", (jiffies - now) / HZ);
 623         else
 624                 CDEBUG(D_INFO, "preprw_write setup: %lu jiffies\n",
 625                        (jiffies - now));
 626
 627         spin_lock(&exp->exp_obd->obd_osfs_lock);
 628         if (oa)
 629                 filter_grant_incoming(exp, oa);
 630         cleanup_phase = 0;
 631
 632         left = filter_grant_space_left(exp);
 633
 634         rc = filter_grant_check(exp, objcount, &fso, niocount, nb, res,
 635                                 &left, dentry->d_inode);
 636         if (oa && oa->o_valid & OBD_MD_FLGRANT)
 637                 oa->o_grant = filter_grant(exp,oa->o_grant,oa->o_undirty,left);
 638
 639         spin_unlock(&exp->exp_obd->obd_osfs_lock);
 640
 641         if (rc) {
 642                 f_dput(dentry);
 643                 GOTO(cleanup, rc);
 644         }
 645
 646         for (i = 0, rnb = nb, lnb = res; i < obj->ioo_bufcnt;
 647              i++, lnb++, rnb++) {
 648                 /* We still set up for ungranted pages so that granted pages
 649                  * can be written to disk as they were promised, and portals
 650                  * needs to keep the pages all aligned properly. */
 651                 lnb->dentry = dentry;
 652                 lnb->offset = rnb->offset;
 653                 lnb->len    = rnb->len;
 654                 lnb->flags  = rnb->flags;
 655
 656                 rc = filter_start_page_write(exp->exp_obd, dentry->d_inode,lnb);
 657                 if (rc) {
 658                         CERROR("page err %u@"LPU64" %u/%u %p: rc %d\n",
 659                                lnb->len, lnb->offset,
 660                                i, obj->ioo_bufcnt, dentry, rc);
 661                         while (lnb-- > res)
 662                                 filter_abort_page_write(lnb);
 663                         f_dput(dentry);
 664                         GOTO(cleanup, rc);
 665                 }
 666                 if (lnb->rc == 0)
 667                         tot_bytes += lnb->len;
 668         }
 669
 670         while (lnb-- > res) {
 671                 if (lnb->len == PAGE_SIZE)
 672                         continue;
 673                 rc = filter_finish_page_read(lnb);
 674                 if (rc) {
 675                         CERROR("error page %u@"LPU64" %u %p: rc %d\n", lnb->len,
 676                                lnb->offset, (int)(lnb - res), lnb->dentry, rc);
 677                         GOTO(cleanup, rc);
 678                 }
 679         }
 680
 681         if (time_after(jiffies, now + 15 * HZ))
 682                 CERROR("slow start_page_write %lus\n", (jiffies - now) / HZ);
 683         else
 684                 CDEBUG(D_INFO, "start_page_write: %lu jiffies\n",
 685                        (jiffies - now));
 686
 687         lprocfs_counter_add(exp->exp_obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
 688                             tot_bytes);
 689         EXIT;
 690 cleanup:
 691         switch(cleanup_phase) {
 692         case 1:
 693                 spin_lock(&exp->exp_obd->obd_osfs_lock);
 694                 if (oa)
 695                         filter_grant_incoming(exp, oa);
 696                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
 697         default: ;
 698         }
 699         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 700         return rc;
 701 }
 702
 703 int filter_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
 704                   int objcount, struct obd_ioobj *obj, int niocount,
 705                   struct niobuf_remote *nb, struct niobuf_local *res,
 706                   struct obd_trans_info *oti)
 707 {
 708         if (cmd == OBD_BRW_WRITE)
 709                 return filter_preprw_write(cmd, exp, oa, objcount, obj,
 710                                            niocount, nb, res, oti);
 711
 712         if (cmd == OBD_BRW_READ)
 713                 return filter_preprw_read(cmd, exp, oa, objcount, obj,
 714                                           niocount, nb, res, oti);
 715
 716         LBUG();
 717         return -EPROTO;
 718 }
 719
 720 void filter_release_read_page(struct filter_obd *filter, struct inode *inode,
 721                               struct page *page)
 722 {
 723         int drop = 0;
 724
 725         if (inode != NULL &&
 726             (inode->i_size > filter->fo_readcache_max_filesize))
 727                 drop = 1;
 728
 729         /* drop from cache like truncate_list_pages() */
 730         if (drop && !TryLockPage(page)) {
 731                 if (page->mapping)
 732                         ll_truncate_complete_page(page);
 733                 unlock_page(page);
 734         }
 735         page_cache_release(page);
 736 }
 737
 738 static int filter_commitrw_read(struct obd_export *exp, struct obdo *oa,
 739                                 int objcount, struct obd_ioobj *obj,
 740                                 int niocount, struct niobuf_local *res,
 741                                 struct obd_trans_info *oti, int rc)
 742 {
 743         struct obd_ioobj *o;
 744         struct niobuf_local *lnb;
 745         int i, j;
 746         struct inode *inode = NULL;
 747         ENTRY;
 748
 749         if (res->dentry != NULL)
 750                 inode = res->dentry->d_inode;
 751
 752         for (i = 0, o = obj, lnb = res; i < objcount; i++, o++) {
 753                 for (j = 0 ; j < o->ioo_bufcnt ; j++, lnb++) {
 754                         if (lnb->page == NULL)
 755                                 continue;
 756                         filter_release_read_page(&exp->exp_obd->u.filter,
 757                                                  inode, lnb->page);
 758                 }
 759         }
 760
 761         if (res->dentry != NULL)
 762                 f_dput(res->dentry);
 763         RETURN(rc);
 764 }
 765
 766 void flip_into_page_cache(struct inode *inode, struct page *new_page)
 767 {
 768         struct page *old_page;
 769         int rc;
 770
 771         do {
 772                 /* the dlm is protecting us from read/write concurrency, so we
 773                  * expect this find_lock_page to return quickly.  even if we
 774                  * race with another writer it won't be doing much work with
 775                  * the page locked.  we do this 'cause t_c_p expects a
 776                  * locked page, and it wants to grab the pagecache lock
 777                  * as well. */
 778                 old_page = find_lock_page(inode->i_mapping, new_page->index);
 779                 if (old_page) {
 780                         ll_truncate_complete_page(old_page);
 781                         unlock_page(old_page);
 782                         page_cache_release(old_page);
 783                 }
 784
 785 #if 0 /* this should be a /proc tunable someday */
 786                 /* racing o_directs (no locking ioctl) could race adding
 787                  * their pages, so we repeat the page invalidation unless
 788                  * we successfully added our new page */
 789                 rc = add_to_page_cache_unique(new_page, inode->i_mapping,
 790                                               new_page->index,
 791                                               page_hash(inode->i_mapping,
 792                                                         new_page->index));
 793                 if (rc == 0) {
 794                         /* add_to_page_cache clears uptodate|dirty and locks
 795                          * the page */
 796                         SetPageUptodate(new_page);
 797                         unlock_page(new_page);
 798                 }
 799 #else
 800                 rc = 0;
 801 #endif
 802         } while (rc != 0);
 803 }
 804
 805 void filter_grant_commit(struct obd_export *exp, int niocount,
 806                          struct niobuf_local *res)
 807 {
 808         struct filter_obd *filter = &exp->exp_obd->u.filter;
 809         struct niobuf_local *lnb = res;
 810         unsigned long pending = 0;
 811         int i;
 812
 813         spin_lock(&exp->exp_obd->obd_osfs_lock);
 814         for (i = 0, lnb = res; i < niocount; i++, lnb++)
 815                 pending += lnb->lnb_grant_used;
 816
 817         LASSERTF(exp->exp_filter_data.fed_pending >= pending,
 818                  "%s: cli %s/%p fed_pending: %lu grant_used: %lu\n",
 819                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 820                  exp->exp_filter_data.fed_pending, pending);
 821         exp->exp_filter_data.fed_pending -= pending;
 822         LASSERTF(filter->fo_tot_granted >= pending,
 823                  "%s: cli %s/%p tot_granted: "LPU64" grant_used: %lu\n",
 824                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 825                  exp->exp_obd->u.filter.fo_tot_granted, pending);
 826         filter->fo_tot_granted -= pending;
 827         LASSERTF(filter->fo_tot_pending >= pending,
 828                  "%s: cli %s/%p tot_pending: "LPU64" grant_used: %lu\n",
 829                  exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
 830                  filter->fo_tot_pending, pending);
 831         filter->fo_tot_pending -= pending;
 832
 833         spin_unlock(&exp->exp_obd->obd_osfs_lock);
 834 }
 835
 836 int filter_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
 837                     int objcount, struct obd_ioobj *obj, int niocount,
 838                     struct niobuf_local *res, struct obd_trans_info *oti,int rc)
 839 {
 840         if (cmd == OBD_BRW_WRITE)
 841                 return filter_commitrw_write(exp, oa, objcount, obj, niocount,
 842                                              res, oti, rc);
 843         if (cmd == OBD_BRW_READ)
 844                 return filter_commitrw_read(exp, oa, objcount, obj, niocount,
 845                                             res, oti, rc);
 846         LBUG();
 847         return -EPROTO;
 848 }
 849
 850 int filter_brw(int cmd, struct obd_export *exp, struct obdo *oa,
 851                struct lov_stripe_md *lsm, obd_count oa_bufs,
 852                struct brw_page *pga, struct obd_trans_info *oti)
 853 {
 854         struct obd_ioobj ioo;
 855         struct niobuf_local *lnb;
 856         struct niobuf_remote *rnb;
 857         obd_count i;
 858         int ret = 0;
 859         ENTRY;
 860
 861         OBD_ALLOC(lnb, oa_bufs * sizeof(struct niobuf_local));
 862         OBD_ALLOC(rnb, oa_bufs * sizeof(struct niobuf_remote));
 863
 864         if (lnb == NULL || rnb == NULL)
 865                 GOTO(out, ret = -ENOMEM);
 866
 867         for (i = 0; i < oa_bufs; i++) {
 868                 rnb[i].offset = pga[i].disk_offset;
 869                 rnb[i].len = pga[i].count;
 870         }
 871
 872         obdo_to_ioobj(oa, &ioo);
 873         ioo.ioo_bufcnt = oa_bufs;
 874
 875         ret = filter_preprw(cmd, exp, oa, 1, &ioo, oa_bufs, rnb, lnb, oti);
 876         if (ret != 0)
 877                 GOTO(out, ret);
 878
 879         for (i = 0; i < oa_bufs; i++) {
 880                 void *virt = kmap(pga[i].pg);
 881                 obd_off off = pga[i].disk_offset & ~PAGE_MASK;
 882                 void *addr = kmap(lnb[i].page);
 883
 884                 /* 2 kmaps == vanishingly small deadlock opportunity */
 885
 886                 if (cmd & OBD_BRW_WRITE)
 887                         memcpy(addr + off, virt + off, pga[i].count);
 888                 else
 889                         memcpy(virt + off, addr + off, pga[i].count);
 890
 891                 kunmap(lnb[i].page);
 892                 kunmap(pga[i].pg);
 893         }
 894
 895         ret = filter_commitrw(cmd, exp, oa, 1, &ioo, oa_bufs, lnb, oti, ret);
 896
 897 out:
 898         if (lnb)
 899                 OBD_FREE(lnb, oa_bufs * sizeof(struct niobuf_local));
 900         if (rnb)
 901                 OBD_FREE(rnb, oa_bufs * sizeof(struct niobuf_remote));
 902         RETURN(ret);
 903 }