sys/vfs/hammer/hammer_vnops.c

   1 /*
   2  * Copyright (c) 2007-2008 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  *
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in
  15  *    the documentation and/or other materials provided with the
  16  *    distribution.
  17  * 3. Neither the name of The DragonFly Project nor the names of its
  18  *    contributors may be used to endorse or promote products derived
  19  *    from this software without specific, prior written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  32  * SUCH DAMAGE.
  33  *
  34  * $DragonFly: src/sys/vfs/hammer/hammer_vnops.c,v 1.102 2008/10/16 17:24:16 dillon Exp $
  35  */
  36
  37 #include <sys/param.h>
  38 #include <sys/systm.h>
  39 #include <sys/kernel.h>
  40 #include <sys/fcntl.h>
  41 #include <sys/namecache.h>
  42 #include <sys/vnode.h>
  43 #include <sys/lockf.h>
  44 #include <sys/event.h>
  45 #include <sys/stat.h>
  46 #include <sys/dirent.h>
  47 #include <sys/file.h>
  48 #include <vm/vm_extern.h>
  49 #include <vm/swap_pager.h>
  50 #include <vfs/fifofs/fifo.h>
  51
  52 #include "hammer.h"
  53
  54 /*
  55  * USERFS VNOPS
  56  */
  57 /*static int hammer_vop_vnoperate(struct vop_generic_args *);*/
  58 static int hammer_vop_fsync(struct vop_fsync_args *);
  59 static int hammer_vop_read(struct vop_read_args *);
  60 static int hammer_vop_write(struct vop_write_args *);
  61 static int hammer_vop_access(struct vop_access_args *);
  62 static int hammer_vop_advlock(struct vop_advlock_args *);
  63 static int hammer_vop_close(struct vop_close_args *);
  64 static int hammer_vop_ncreate(struct vop_ncreate_args *);
  65 static int hammer_vop_getattr(struct vop_getattr_args *);
  66 static int hammer_vop_nresolve(struct vop_nresolve_args *);
  67 static int hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *);
  68 static int hammer_vop_nlink(struct vop_nlink_args *);
  69 static int hammer_vop_nmkdir(struct vop_nmkdir_args *);
  70 static int hammer_vop_nmknod(struct vop_nmknod_args *);
  71 static int hammer_vop_open(struct vop_open_args *);
  72 static int hammer_vop_print(struct vop_print_args *);
  73 static int hammer_vop_readdir(struct vop_readdir_args *);
  74 static int hammer_vop_readlink(struct vop_readlink_args *);
  75 static int hammer_vop_nremove(struct vop_nremove_args *);
  76 static int hammer_vop_nrename(struct vop_nrename_args *);
  77 static int hammer_vop_nrmdir(struct vop_nrmdir_args *);
  78 static int hammer_vop_markatime(struct vop_markatime_args *);
  79 static int hammer_vop_setattr(struct vop_setattr_args *);
  80 static int hammer_vop_strategy(struct vop_strategy_args *);
  81 static int hammer_vop_bmap(struct vop_bmap_args *ap);
  82 static int hammer_vop_nsymlink(struct vop_nsymlink_args *);
  83 static int hammer_vop_nwhiteout(struct vop_nwhiteout_args *);
  84 static int hammer_vop_ioctl(struct vop_ioctl_args *);
  85 static int hammer_vop_mountctl(struct vop_mountctl_args *);
  86 static int hammer_vop_kqfilter (struct vop_kqfilter_args *);
  87
  88 static int hammer_vop_fifoclose (struct vop_close_args *);
  89 static int hammer_vop_fiforead (struct vop_read_args *);
  90 static int hammer_vop_fifowrite (struct vop_write_args *);
  91 static int hammer_vop_fifokqfilter (struct vop_kqfilter_args *);
  92
  93 struct vop_ops hammer_vnode_vops = {
  94         .vop_default =          vop_defaultop,
  95         .vop_fsync =            hammer_vop_fsync,
  96         .vop_getpages =         vop_stdgetpages,
  97         .vop_putpages =         vop_stdputpages,
  98         .vop_read =             hammer_vop_read,
  99         .vop_write =            hammer_vop_write,
 100         .vop_access =           hammer_vop_access,
 101         .vop_advlock =          hammer_vop_advlock,
 102         .vop_close =            hammer_vop_close,
 103         .vop_ncreate =          hammer_vop_ncreate,
 104         .vop_getattr =          hammer_vop_getattr,
 105         .vop_inactive =         hammer_vop_inactive,
 106         .vop_reclaim =          hammer_vop_reclaim,
 107         .vop_nresolve =         hammer_vop_nresolve,
 108         .vop_nlookupdotdot =    hammer_vop_nlookupdotdot,
 109         .vop_nlink =            hammer_vop_nlink,
 110         .vop_nmkdir =           hammer_vop_nmkdir,
 111         .vop_nmknod =           hammer_vop_nmknod,
 112         .vop_open =             hammer_vop_open,
 113         .vop_pathconf =         vop_stdpathconf,
 114         .vop_print =            hammer_vop_print,
 115         .vop_readdir =          hammer_vop_readdir,
 116         .vop_readlink =         hammer_vop_readlink,
 117         .vop_nremove =          hammer_vop_nremove,
 118         .vop_nrename =          hammer_vop_nrename,
 119         .vop_nrmdir =           hammer_vop_nrmdir,
 120         .vop_markatime =        hammer_vop_markatime,
 121         .vop_setattr =          hammer_vop_setattr,
 122         .vop_bmap =             hammer_vop_bmap,
 123         .vop_strategy =         hammer_vop_strategy,
 124         .vop_nsymlink =         hammer_vop_nsymlink,
 125         .vop_nwhiteout =        hammer_vop_nwhiteout,
 126         .vop_ioctl =            hammer_vop_ioctl,
 127         .vop_mountctl =         hammer_vop_mountctl,
 128         .vop_kqfilter =         hammer_vop_kqfilter
 129 };
 130
 131 struct vop_ops hammer_spec_vops = {
 132         .vop_default =          vop_defaultop,
 133         .vop_fsync =            hammer_vop_fsync,
 134         .vop_read =             vop_stdnoread,
 135         .vop_write =            vop_stdnowrite,
 136         .vop_access =           hammer_vop_access,
 137         .vop_close =            hammer_vop_close,
 138         .vop_markatime =        hammer_vop_markatime,
 139         .vop_getattr =          hammer_vop_getattr,
 140         .vop_inactive =         hammer_vop_inactive,
 141         .vop_reclaim =          hammer_vop_reclaim,
 142         .vop_setattr =          hammer_vop_setattr
 143 };
 144
 145 struct vop_ops hammer_fifo_vops = {
 146         .vop_default =          fifo_vnoperate,
 147         .vop_fsync =            hammer_vop_fsync,
 148         .vop_read =             hammer_vop_fiforead,
 149         .vop_write =            hammer_vop_fifowrite,
 150         .vop_access =           hammer_vop_access,
 151         .vop_close =            hammer_vop_fifoclose,
 152         .vop_markatime =        hammer_vop_markatime,
 153         .vop_getattr =          hammer_vop_getattr,
 154         .vop_inactive =         hammer_vop_inactive,
 155         .vop_reclaim =          hammer_vop_reclaim,
 156         .vop_setattr =          hammer_vop_setattr,
 157         .vop_kqfilter =         hammer_vop_fifokqfilter
 158 };
 159
 160 static __inline
 161 void
 162 hammer_knote(struct vnode *vp, int flags)
 163 {
 164         if (flags)
 165                 KNOTE(&vp->v_pollinfo.vpi_kqinfo.ki_note, flags);
 166 }
 167
 168 #ifdef DEBUG_TRUNCATE
 169 struct hammer_inode *HammerTruncIp;
 170 #endif
 171
 172 static int hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
 173                            struct vnode *dvp, struct ucred *cred,
 174                            int flags, int isdir);
 175 static int hammer_vop_strategy_read(struct vop_strategy_args *ap);
 176 static int hammer_vop_strategy_write(struct vop_strategy_args *ap);
 177
 178 #if 0
 179 static
 180 int
 181 hammer_vop_vnoperate(struct vop_generic_args *)
 182 {
 183         return (VOCALL(&hammer_vnode_vops, ap));
 184 }
 185 #endif
 186
 187 /*
 188  * hammer_vop_fsync { vp, waitfor }
 189  *
 190  * fsync() an inode to disk and wait for it to be completely committed
 191  * such that the information would not be undone if a crash occured after
 192  * return.
 193  *
 194  * NOTE: HAMMER's fsync()'s are going to remain expensive until we implement
 195  *       a REDO log.  A sysctl is provided to relax HAMMER's fsync()
 196  *       operation.
 197  *
 198  *       Ultimately the combination of a REDO log and use of fast storage
 199  *       to front-end cluster caches will make fsync fast, but it aint
 200  *       here yet.  And, in anycase, we need real transactional
 201  *       all-or-nothing features which are not restricted to a single file.
 202  */
 203 static
 204 int
 205 hammer_vop_fsync(struct vop_fsync_args *ap)
 206 {
 207         hammer_inode_t ip = VTOI(ap->a_vp);
 208         hammer_mount_t hmp = ip->hmp;
 209         int waitfor = ap->a_waitfor;
 210         int mode;
 211
 212         lwkt_gettoken(&hmp->fs_token);
 213
 214         /*
 215          * Fsync rule relaxation (default is either full synchronous flush
 216          * or REDO semantics with synchronous flush).
 217          */
 218         if (ap->a_flags & VOP_FSYNC_SYSCALL) {
 219                 switch(hammer_fsync_mode) {
 220                 case 0:
 221 mode0:
 222                         /* no REDO, full synchronous flush */
 223                         goto skip;
 224                 case 1:
 225 mode1:
 226                         /* no REDO, full asynchronous flush */
 227                         if (waitfor == MNT_WAIT)
 228                                 waitfor = MNT_NOWAIT;
 229                         goto skip;
 230                 case 2:
 231                         /* REDO semantics, synchronous flush */
 232                         if (hmp->version < HAMMER_VOL_VERSION_FOUR)
 233                                 goto mode0;
 234                         mode = HAMMER_FLUSH_UNDOS_AUTO;
 235                         break;
 236                 case 3:
 237                         /* REDO semantics, relaxed asynchronous flush */
 238                         if (hmp->version < HAMMER_VOL_VERSION_FOUR)
 239                                 goto mode1;
 240                         mode = HAMMER_FLUSH_UNDOS_RELAXED;
 241                         if (waitfor == MNT_WAIT)
 242                                 waitfor = MNT_NOWAIT;
 243                         break;
 244                 case 4:
 245                         /* ignore the fsync() system call */
 246                         lwkt_reltoken(&hmp->fs_token);
 247                         return(0);
 248                 default:
 249                         /* we have to do something */
 250                         mode = HAMMER_FLUSH_UNDOS_RELAXED;
 251                         if (waitfor == MNT_WAIT)
 252                                 waitfor = MNT_NOWAIT;
 253                         break;
 254                 }
 255
 256                 /*
 257                  * Fast fsync only needs to flush the UNDO/REDO fifo if
 258                  * HAMMER_INODE_REDO is non-zero and the only modifications
 259                  * made to the file are write or write-extends.
 260                  */
 261                 if ((ip->flags & HAMMER_INODE_REDO) &&
 262                     (ip->flags & HAMMER_INODE_MODMASK_NOREDO) == 0
 263                 ) {
 264                         ++hammer_count_fsyncs;
 265                         hammer_flusher_flush_undos(hmp, mode);
 266                         ip->redo_count = 0;
 267                         lwkt_reltoken(&hmp->fs_token);
 268                         return(0);
 269                 }
 270
 271                 /*
 272                  * REDO is enabled by fsync(), the idea being we really only
 273                  * want to lay down REDO records when programs are using
 274                  * fsync() heavily.  The first fsync() on the file starts
 275                  * the gravy train going and later fsync()s keep it hot by
 276                  * resetting the redo_count.
 277                  *
 278                  * We weren't running REDOs before now so we have to fall
 279                  * through and do a full fsync of what we have.
 280                  */
 281                 if (hmp->version >= HAMMER_VOL_VERSION_FOUR &&
 282                     (hmp->flags & HAMMER_MOUNT_REDO_RECOVERY_RUN) == 0) {
 283                         ip->flags |= HAMMER_INODE_REDO;
 284                         ip->redo_count = 0;
 285                 }
 286         }
 287 skip:
 288
 289         /*
 290          * Do a full flush sequence.
 291          *
 292          * Attempt to release the vnode while waiting for the inode to
 293          * finish flushing.  This can really mess up inactive->reclaim
 294          * sequences so only do it if the vnode is active.
 295          */
 296         ++hammer_count_fsyncs;
 297         vfsync(ap->a_vp, waitfor, 1, NULL, NULL);
 298         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 299         if (waitfor == MNT_WAIT) {
 300                 if ((ap->a_vp->v_flag & VINACTIVE) == 0)
 301                         vn_unlock(ap->a_vp);
 302                 hammer_wait_inode(ip);
 303                 if ((ap->a_vp->v_flag & VINACTIVE) == 0)
 304                         vn_lock(ap->a_vp, LK_EXCLUSIVE | LK_RETRY);
 305         }
 306         lwkt_reltoken(&hmp->fs_token);
 307         return (ip->error);
 308 }
 309
 310 /*
 311  * hammer_vop_read { vp, uio, ioflag, cred }
 312  *
 313  * MPSAFE (for the cache safe does not require fs_token)
 314  */
 315 static
 316 int
 317 hammer_vop_read(struct vop_read_args *ap)
 318 {
 319         struct hammer_transaction trans;
 320         hammer_inode_t ip;
 321         hammer_mount_t hmp;
 322         off_t offset;
 323         struct buf *bp;
 324         struct uio *uio;
 325         int error;
 326         int n;
 327         int seqcount;
 328         int ioseqcount;
 329         int blksize;
 330         int bigread;
 331         int got_fstoken;
 332         size_t resid;
 333
 334         if (ap->a_vp->v_type != VREG)
 335                 return (EINVAL);
 336         ip = VTOI(ap->a_vp);
 337         hmp = ip->hmp;
 338         error = 0;
 339         got_fstoken = 0;
 340         uio = ap->a_uio;
 341
 342         /*
 343          * Attempt to shortcut directly to the VM object using lwbufs.
 344          * This is much faster than instantiating buffer cache buffers.
 345          */
 346         resid = uio->uio_resid;
 347         error = vop_helper_read_shortcut(ap);
 348         hammer_stats_file_read += resid - uio->uio_resid;
 349         if (error)
 350                 return (error);
 351         if (uio->uio_resid == 0)
 352                 goto finished;
 353
 354         /*
 355          * Allow the UIO's size to override the sequential heuristic.
 356          */
 357         blksize = hammer_blocksize(uio->uio_offset);
 358         seqcount = (uio->uio_resid + (BKVASIZE - 1)) / BKVASIZE;
 359         ioseqcount = (ap->a_ioflag >> 16);
 360         if (seqcount < ioseqcount)
 361                 seqcount = ioseqcount;
 362
 363         /*
 364          * If reading or writing a huge amount of data we have to break
 365          * atomicy and allow the operation to be interrupted by a signal
 366          * or it can DOS the machine.
 367          */
 368         bigread = (uio->uio_resid > 100 * 1024 * 1024);
 369
 370         /*
 371          * Access the data typically in HAMMER_BUFSIZE blocks via the
 372          * buffer cache, but HAMMER may use a variable block size based
 373          * on the offset.
 374          *
 375          * XXX Temporary hack, delay the start transaction while we remain
 376          *     MPSAFE.  NOTE: ino_data.size cannot change while vnode is
 377          *     locked-shared.
 378          */
 379         while (uio->uio_resid > 0 && uio->uio_offset < ip->ino_data.size) {
 380                 int64_t base_offset;
 381                 int64_t file_limit;
 382
 383                 blksize = hammer_blocksize(uio->uio_offset);
 384                 offset = (int)uio->uio_offset & (blksize - 1);
 385                 base_offset = uio->uio_offset - offset;
 386
 387                 if (bigread && (error = hammer_signal_check(ip->hmp)) != 0)
 388                         break;
 389
 390                 /*
 391                  * MPSAFE
 392                  */
 393                 bp = getblk(ap->a_vp, base_offset, blksize, 0, 0);
 394                 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) == B_CACHE) {
 395                         bp->b_flags &= ~B_AGE;
 396                         error = 0;
 397                         goto skip;
 398                 }
 399                 if (ap->a_ioflag & IO_NRDELAY) {
 400                         bqrelse(bp);
 401                         return (EWOULDBLOCK);
 402                 }
 403
 404                 /*
 405                  * MPUNSAFE
 406                  */
 407                 if (got_fstoken == 0) {
 408                         lwkt_gettoken(&hmp->fs_token);
 409                         got_fstoken = 1;
 410                         hammer_start_transaction(&trans, ip->hmp);
 411                 }
 412
 413                 /*
 414                  * NOTE: A valid bp has already been acquired, but was not
 415                  *       B_CACHE.
 416                  */
 417                 if (hammer_cluster_enable) {
 418                         /*
 419                          * Use file_limit to prevent cluster_read() from
 420                          * creating buffers of the wrong block size past
 421                          * the demarc.
 422                          */
 423                         file_limit = ip->ino_data.size;
 424                         if (base_offset < HAMMER_XDEMARC &&
 425                             file_limit > HAMMER_XDEMARC) {
 426                                 file_limit = HAMMER_XDEMARC;
 427                         }
 428                         error = cluster_readx(ap->a_vp,
 429                                              file_limit, base_offset,
 430                                              blksize, uio->uio_resid,
 431                                              seqcount * BKVASIZE, &bp);
 432                 } else {
 433                         error = breadnx(ap->a_vp, base_offset, blksize,
 434                                         NULL, NULL, 0, &bp);
 435                 }
 436                 if (error) {
 437                         brelse(bp);
 438                         break;
 439                 }
 440 skip:
 441                 if ((hammer_debug_io & 0x0001) && (bp->b_flags & B_IODEBUG)) {
 442                         kprintf("doff %016jx read file %016jx@%016jx\n",
 443                                 (intmax_t)bp->b_bio2.bio_offset,
 444                                 (intmax_t)ip->obj_id,
 445                                 (intmax_t)bp->b_loffset);
 446                 }
 447                 bp->b_flags &= ~B_IODEBUG;
 448                 if (blksize == HAMMER_XBUFSIZE)
 449                         bp->b_flags |= B_CLUSTEROK;
 450
 451                 n = blksize - offset;
 452                 if (n > uio->uio_resid)
 453                         n = uio->uio_resid;
 454                 if (n > ip->ino_data.size - uio->uio_offset)
 455                         n = (int)(ip->ino_data.size - uio->uio_offset);
 456                 if (got_fstoken)
 457                         lwkt_reltoken(&hmp->fs_token);
 458
 459                 /*
 460                  * Set B_AGE, data has a lower priority than meta-data.
 461                  *
 462                  * Use a hold/unlock/drop sequence to run the uiomove
 463                  * with the buffer unlocked, avoiding deadlocks against
 464                  * read()s on mmap()'d spaces.
 465                  */
 466                 bp->b_flags |= B_AGE;
 467                 error = uiomovebp(bp, (char *)bp->b_data + offset, n, uio);
 468                 bqrelse(bp);
 469
 470                 if (got_fstoken)
 471                         lwkt_gettoken(&hmp->fs_token);
 472
 473                 if (error)
 474                         break;
 475                 hammer_stats_file_read += n;
 476         }
 477
 478 finished:
 479
 480         /*
 481          * Try to update the atime with just the inode lock for maximum
 482          * concurrency.  If we can't shortcut it we have to get the full
 483          * blown transaction.
 484          */
 485         if (got_fstoken == 0 && hammer_update_atime_quick(ip) < 0) {
 486                 lwkt_gettoken(&hmp->fs_token);
 487                 got_fstoken = 1;
 488                 hammer_start_transaction(&trans, ip->hmp);
 489         }
 490
 491         if (got_fstoken) {
 492                 if ((ip->flags & HAMMER_INODE_RO) == 0 &&
 493                     (ip->hmp->mp->mnt_flag & MNT_NOATIME) == 0) {
 494                         ip->ino_data.atime = trans.time;
 495                         hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
 496                 }
 497                 hammer_done_transaction(&trans);
 498                 lwkt_reltoken(&hmp->fs_token);
 499         }
 500         return (error);
 501 }
 502
 503 /*
 504  * hammer_vop_write { vp, uio, ioflag, cred }
 505  */
 506 static
 507 int
 508 hammer_vop_write(struct vop_write_args *ap)
 509 {
 510         struct hammer_transaction trans;
 511         struct hammer_inode *ip;
 512         hammer_mount_t hmp;
 513         thread_t td;
 514         struct uio *uio;
 515         int offset;
 516         off_t base_offset;
 517         int64_t cluster_eof;
 518         struct buf *bp;
 519         int kflags;
 520         int error;
 521         int n;
 522         int flags;
 523         int seqcount;
 524         int bigwrite;
 525
 526         if (ap->a_vp->v_type != VREG)
 527                 return (EINVAL);
 528         ip = VTOI(ap->a_vp);
 529         hmp = ip->hmp;
 530         error = 0;
 531         kflags = 0;
 532         seqcount = ap->a_ioflag >> 16;
 533
 534         if (ip->flags & HAMMER_INODE_RO)
 535                 return (EROFS);
 536
 537         /*
 538          * Create a transaction to cover the operations we perform.
 539          */
 540         lwkt_gettoken(&hmp->fs_token);
 541         hammer_start_transaction(&trans, hmp);
 542         uio = ap->a_uio;
 543
 544         /*
 545          * Check append mode
 546          */
 547         if (ap->a_ioflag & IO_APPEND)
 548                 uio->uio_offset = ip->ino_data.size;
 549
 550         /*
 551          * Check for illegal write offsets.  Valid range is 0...2^63-1.
 552          *
 553          * NOTE: the base_off assignment is required to work around what
 554          * I consider to be a GCC-4 optimization bug.
 555          */
 556         if (uio->uio_offset < 0) {
 557                 hammer_done_transaction(&trans);
 558                 lwkt_reltoken(&hmp->fs_token);
 559                 return (EFBIG);
 560         }
 561         base_offset = uio->uio_offset + uio->uio_resid; /* work around gcc-4 */
 562         if (uio->uio_resid > 0 && base_offset <= uio->uio_offset) {
 563                 hammer_done_transaction(&trans);
 564                 lwkt_reltoken(&hmp->fs_token);
 565                 return (EFBIG);
 566         }
 567
 568         if (uio->uio_resid > 0 && (td = uio->uio_td) != NULL && td->td_proc &&
 569             base_offset > td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
 570                 hammer_done_transaction(&trans);
 571                 lwkt_reltoken(&hmp->fs_token);
 572                 lwpsignal(td->td_proc, td->td_lwp, SIGXFSZ);
 573                 return (EFBIG);
 574         }
 575
 576         /*
 577          * If reading or writing a huge amount of data we have to break
 578          * atomicy and allow the operation to be interrupted by a signal
 579          * or it can DOS the machine.
 580          *
 581          * Preset redo_count so we stop generating REDOs earlier if the
 582          * limit is exceeded.
 583          */
 584         bigwrite = (uio->uio_resid > 100 * 1024 * 1024);
 585         if ((ip->flags & HAMMER_INODE_REDO) &&
 586             ip->redo_count < hammer_limit_redo) {
 587                 ip->redo_count += uio->uio_resid;
 588         }
 589
 590         /*
 591          * Access the data typically in HAMMER_BUFSIZE blocks via the
 592          * buffer cache, but HAMMER may use a variable block size based
 593          * on the offset.
 594          */
 595         while (uio->uio_resid > 0) {
 596                 int fixsize = 0;
 597                 int blksize;
 598                 int blkmask;
 599                 int trivial;
 600                 int endofblk;
 601                 off_t nsize;
 602
 603                 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_WRITE)) != 0)
 604                         break;
 605                 if (bigwrite && (error = hammer_signal_check(hmp)) != 0)
 606                         break;
 607
 608                 blksize = hammer_blocksize(uio->uio_offset);
 609
 610                 /*
 611                  * Do not allow HAMMER to blow out the buffer cache.  Very
 612                  * large UIOs can lockout other processes due to bwillwrite()
 613                  * mechanics.
 614                  *
 615                  * The hammer inode is not locked during these operations.
 616                  * The vnode is locked which can interfere with the pageout
 617                  * daemon for non-UIO_NOCOPY writes but should not interfere
 618                  * with the buffer cache.  Even so, we cannot afford to
 619                  * allow the pageout daemon to build up too many dirty buffer
 620                  * cache buffers.
 621                  *
 622                  * Only call this if we aren't being recursively called from
 623                  * a virtual disk device (vn), else we may deadlock.
 624                  */
 625                 if ((ap->a_ioflag & IO_RECURSE) == 0)
 626                         bwillwrite(blksize);
 627
 628                 /*
 629                  * Control the number of pending records associated with
 630                  * this inode.  If too many have accumulated start a
 631                  * flush.  Try to maintain a pipeline with the flusher.
 632                  *
 633                  * NOTE: It is possible for other sources to grow the
 634                  *       records but not necessarily issue another flush,
 635                  *       so use a timeout and ensure that a re-flush occurs.
 636                  */
 637                 if (ip->rsv_recs >= hammer_limit_inode_recs) {
 638                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 639                         while (ip->rsv_recs >= hammer_limit_inode_recs * 2) {
 640                                 ip->flags |= HAMMER_INODE_RECSW;
 641                                 tsleep(&ip->rsv_recs, 0, "hmrwww", hz);
 642                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 643                         }
 644                 }
 645
 646 #if 0
 647                 /*
 648                  * Do not allow HAMMER to blow out system memory by
 649                  * accumulating too many records.   Records are so well
 650                  * decoupled from the buffer cache that it is possible
 651                  * for userland to push data out to the media via
 652                  * direct-write, but build up the records queued to the
 653                  * backend faster then the backend can flush them out.
 654                  * HAMMER has hit its write limit but the frontend has
 655                  * no pushback to slow it down.
 656                  */
 657                 if (hmp->rsv_recs > hammer_limit_recs / 2) {
 658                         /*
 659                          * Get the inode on the flush list
 660                          */
 661                         if (ip->rsv_recs >= 64)
 662                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 663                         else if (ip->rsv_recs >= 16)
 664                                 hammer_flush_inode(ip, 0);
 665
 666                         /*
 667                          * Keep the flusher going if the system keeps
 668                          * queueing records.
 669                          */
 670                         delta = hmp->count_newrecords -
 671                                 hmp->last_newrecords;
 672                         if (delta < 0 || delta > hammer_limit_recs / 2) {
 673                                 hmp->last_newrecords = hmp->count_newrecords;
 674                                 hammer_sync_hmp(hmp, MNT_NOWAIT);
 675                         }
 676
 677                         /*
 678                          * If we have gotten behind start slowing
 679                          * down the writers.
 680                          */
 681                         delta = (hmp->rsv_recs - hammer_limit_recs) *
 682                                 hz / hammer_limit_recs;
 683                         if (delta > 0)
 684                                 tsleep(&trans, 0, "hmrslo", delta);
 685                 }
 686 #endif
 687
 688                 /*
 689                  * Calculate the blocksize at the current offset and figure
 690                  * out how much we can actually write.
 691                  */
 692                 blkmask = blksize - 1;
 693                 offset = (int)uio->uio_offset & blkmask;
 694                 base_offset = uio->uio_offset & ~(int64_t)blkmask;
 695                 n = blksize - offset;
 696                 if (n > uio->uio_resid) {
 697                         n = uio->uio_resid;
 698                         endofblk = 0;
 699                 } else {
 700                         endofblk = 1;
 701                 }
 702                 nsize = uio->uio_offset + n;
 703                 if (nsize > ip->ino_data.size) {
 704                         if (uio->uio_offset > ip->ino_data.size)
 705                                 trivial = 0;
 706                         else
 707                                 trivial = 1;
 708                         nvextendbuf(ap->a_vp,
 709                                     ip->ino_data.size,
 710                                     nsize,
 711                                     hammer_blocksize(ip->ino_data.size),
 712                                     hammer_blocksize(nsize),
 713                                     hammer_blockoff(ip->ino_data.size),
 714                                     hammer_blockoff(nsize),
 715                                     trivial);
 716                         fixsize = 1;
 717                         kflags |= NOTE_EXTEND;
 718                 }
 719
 720                 if (uio->uio_segflg == UIO_NOCOPY) {
 721                         /*
 722                          * Issuing a write with the same data backing the
 723                          * buffer.  Instantiate the buffer to collect the
 724                          * backing vm pages, then read-in any missing bits.
 725                          *
 726                          * This case is used by vop_stdputpages().
 727                          */
 728                         bp = getblk(ap->a_vp, base_offset,
 729                                     blksize, GETBLK_BHEAVY, 0);
 730                         if ((bp->b_flags & B_CACHE) == 0) {
 731                                 bqrelse(bp);
 732                                 error = bread(ap->a_vp, base_offset,
 733                                               blksize, &bp);
 734                         }
 735                 } else if (offset == 0 && uio->uio_resid >= blksize) {
 736                         /*
 737                          * Even though we are entirely overwriting the buffer
 738                          * we may still have to zero it out to avoid a
 739                          * mmap/write visibility issue.
 740                          */
 741                         bp = getblk(ap->a_vp, base_offset, blksize, GETBLK_BHEAVY, 0);
 742                         if ((bp->b_flags & B_CACHE) == 0)
 743                                 vfs_bio_clrbuf(bp);
 744                 } else if (base_offset >= ip->ino_data.size) {
 745                         /*
 746                          * If the base offset of the buffer is beyond the
 747                          * file EOF, we don't have to issue a read.
 748                          */
 749                         bp = getblk(ap->a_vp, base_offset,
 750                                     blksize, GETBLK_BHEAVY, 0);
 751                         vfs_bio_clrbuf(bp);
 752                 } else {
 753                         /*
 754                          * Partial overwrite, read in any missing bits then
 755                          * replace the portion being written.
 756                          */
 757                         error = bread(ap->a_vp, base_offset, blksize, &bp);
 758                         if (error == 0)
 759                                 bheavy(bp);
 760                 }
 761                 if (error == 0) {
 762                         lwkt_reltoken(&hmp->fs_token);
 763                         error = uiomovebp(bp, bp->b_data + offset, n, uio);
 764                         lwkt_gettoken(&hmp->fs_token);
 765                 }
 766
 767                 /*
 768                  * Generate REDO records if enabled and redo_count will not
 769                  * exceeded the limit.
 770                  *
 771                  * If redo_count exceeds the limit we stop generating records
 772                  * and clear HAMMER_INODE_REDO.  This will cause the next
 773                  * fsync() to do a full meta-data sync instead of just an
 774                  * UNDO/REDO fifo update.
 775                  *
 776                  * When clearing HAMMER_INODE_REDO any pre-existing REDOs
 777                  * will still be tracked.  The tracks will be terminated
 778                  * when the related meta-data (including possible data
 779                  * modifications which are not tracked via REDO) is
 780                  * flushed.
 781                  */
 782                 if ((ip->flags & HAMMER_INODE_REDO) && error == 0) {
 783                         if (ip->redo_count < hammer_limit_redo) {
 784                                 bp->b_flags |= B_VFSFLAG1;
 785                                 error = hammer_generate_redo(&trans, ip,
 786                                                      base_offset + offset,
 787                                                      HAMMER_REDO_WRITE,
 788                                                      bp->b_data + offset,
 789                                                      (size_t)n);
 790                         } else {
 791                                 ip->flags &= ~HAMMER_INODE_REDO;
 792                         }
 793                 }
 794
 795                 /*
 796                  * If we screwed up we have to undo any VM size changes we
 797                  * made.
 798                  */
 799                 if (error) {
 800                         brelse(bp);
 801                         if (fixsize) {
 802                                 nvtruncbuf(ap->a_vp, ip->ino_data.size,
 803                                           hammer_blocksize(ip->ino_data.size),
 804                                           hammer_blockoff(ip->ino_data.size),
 805                                           0);
 806                         }
 807                         break;
 808                 }
 809                 kflags |= NOTE_WRITE;
 810                 hammer_stats_file_write += n;
 811                 if (blksize == HAMMER_XBUFSIZE)
 812                         bp->b_flags |= B_CLUSTEROK;
 813                 if (ip->ino_data.size < uio->uio_offset) {
 814                         ip->ino_data.size = uio->uio_offset;
 815                         flags = HAMMER_INODE_SDIRTY;
 816                 } else {
 817                         flags = 0;
 818                 }
 819                 ip->ino_data.mtime = trans.time;
 820                 flags |= HAMMER_INODE_MTIME | HAMMER_INODE_BUFS;
 821                 hammer_modify_inode(&trans, ip, flags);
 822
 823                 /*
 824                  * Once we dirty the buffer any cached zone-X offset
 825                  * becomes invalid.  HAMMER NOTE: no-history mode cannot
 826                  * allow overwriting over the same data sector unless
 827                  * we provide UNDOs for the old data, which we don't.
 828                  */
 829                 bp->b_bio2.bio_offset = NOOFFSET;
 830
 831                 /*
 832                  * Final buffer disposition.
 833                  *
 834                  * Because meta-data updates are deferred, HAMMER is
 835                  * especially sensitive to excessive bdwrite()s because
 836                  * the I/O stream is not broken up by disk reads.  So the
 837                  * buffer cache simply cannot keep up.
 838                  *
 839                  * WARNING!  blksize is variable.  cluster_write() is
 840                  *           expected to not blow up if it encounters
 841                  *           buffers that do not match the passed blksize.
 842                  *
 843                  * NOTE!  Hammer shouldn't need to bawrite()/cluster_write().
 844                  *        The ip->rsv_recs check should burst-flush the data.
 845                  *        If we queue it immediately the buf could be left
 846                  *        locked on the device queue for a very long time.
 847                  *
 848                  *        However, failing to flush a dirty buffer out when
 849                  *        issued from the pageout daemon can result in a low
 850                  *        memory deadlock against bio_page_alloc(), so we
 851                  *        have to bawrite() on IO_ASYNC as well.
 852                  *
 853                  * NOTE!  To avoid degenerate stalls due to mismatched block
 854                  *        sizes we only honor IO_DIRECT on the write which
 855                  *        abuts the end of the buffer.  However, we must
 856                  *        honor IO_SYNC in case someone is silly enough to
 857                  *        configure a HAMMER file as swap, or when HAMMER
 858                  *        is serving NFS (for commits).  Ick ick.
 859                  */
 860                 bp->b_flags |= B_AGE;
 861                 if (blksize == HAMMER_XBUFSIZE)
 862                         bp->b_flags |= B_CLUSTEROK;
 863
 864                 if (ap->a_ioflag & IO_SYNC) {
 865                         bwrite(bp);
 866                 } else if ((ap->a_ioflag & IO_DIRECT) && endofblk) {
 867                         bawrite(bp);
 868                 } else if (ap->a_ioflag & IO_ASYNC) {
 869                         bawrite(bp);
 870                 } else if (hammer_cluster_enable &&
 871                            !(ap->a_vp->v_mount->mnt_flag & MNT_NOCLUSTERW)) {
 872                         if (base_offset < HAMMER_XDEMARC)
 873                                 cluster_eof = hammer_blockdemarc(base_offset,
 874                                                          ip->ino_data.size);
 875                         else
 876                                 cluster_eof = ip->ino_data.size;
 877                         cluster_write(bp, cluster_eof, blksize, seqcount);
 878                 } else {
 879                         bdwrite(bp);
 880                 }
 881         }
 882         hammer_done_transaction(&trans);
 883         hammer_knote(ap->a_vp, kflags);
 884         lwkt_reltoken(&hmp->fs_token);
 885         return (error);
 886 }
 887
 888 /*
 889  * hammer_vop_access { vp, mode, cred }
 890  *
 891  * MPSAFE - does not require fs_token
 892  */
 893 static
 894 int
 895 hammer_vop_access(struct vop_access_args *ap)
 896 {
 897         struct hammer_inode *ip = VTOI(ap->a_vp);
 898         uid_t uid;
 899         gid_t gid;
 900         int error;
 901
 902         ++hammer_stats_file_iopsr;
 903         uid = hammer_to_unix_xid(&ip->ino_data.uid);
 904         gid = hammer_to_unix_xid(&ip->ino_data.gid);
 905
 906         error = vop_helper_access(ap, uid, gid, ip->ino_data.mode,
 907                                   ip->ino_data.uflags);
 908         return (error);
 909 }
 910
 911 /*
 912  * hammer_vop_advlock { vp, id, op, fl, flags }
 913  *
 914  * MPSAFE - does not require fs_token
 915  */
 916 static
 917 int
 918 hammer_vop_advlock(struct vop_advlock_args *ap)
 919 {
 920         hammer_inode_t ip = VTOI(ap->a_vp);
 921
 922         return (lf_advlock(ap, &ip->advlock, ip->ino_data.size));
 923 }
 924
 925 /*
 926  * hammer_vop_close { vp, fflag }
 927  *
 928  * We can only sync-on-close for normal closes.  XXX disabled for now.
 929  */
 930 static
 931 int
 932 hammer_vop_close(struct vop_close_args *ap)
 933 {
 934 #if 0
 935         struct vnode *vp = ap->a_vp;
 936         hammer_inode_t ip = VTOI(vp);
 937         int waitfor;
 938         if (ip->flags & (HAMMER_INODE_CLOSESYNC|HAMMER_INODE_CLOSEASYNC)) {
 939                 if (vn_islocked(vp) == LK_EXCLUSIVE &&
 940                     (vp->v_flag & (VINACTIVE|VRECLAIMED)) == 0) {
 941                         if (ip->flags & HAMMER_INODE_CLOSESYNC)
 942                                 waitfor = MNT_WAIT;
 943                         else
 944                                 waitfor = MNT_NOWAIT;
 945                         ip->flags &= ~(HAMMER_INODE_CLOSESYNC |
 946                                        HAMMER_INODE_CLOSEASYNC);
 947                         VOP_FSYNC(vp, MNT_NOWAIT, waitfor);
 948                 }
 949         }
 950 #endif
 951         return (vop_stdclose(ap));
 952 }
 953
 954 /*
 955  * hammer_vop_ncreate { nch, dvp, vpp, cred, vap }
 956  *
 957  * The operating system has already ensured that the directory entry
 958  * does not exist and done all appropriate namespace locking.
 959  */
 960 static
 961 int
 962 hammer_vop_ncreate(struct vop_ncreate_args *ap)
 963 {
 964         struct hammer_transaction trans;
 965         struct hammer_inode *dip;
 966         struct hammer_inode *nip;
 967         struct nchandle *nch;
 968         hammer_mount_t hmp;
 969         int error;
 970
 971         nch = ap->a_nch;
 972         dip = VTOI(ap->a_dvp);
 973         hmp = dip->hmp;
 974
 975         if (dip->flags & HAMMER_INODE_RO)
 976                 return (EROFS);
 977         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
 978                 return (error);
 979
 980         /*
 981          * Create a transaction to cover the operations we perform.
 982          */
 983         lwkt_gettoken(&hmp->fs_token);
 984         hammer_start_transaction(&trans, hmp);
 985         ++hammer_stats_file_iopsw;
 986
 987         /*
 988          * Create a new filesystem object of the requested type.  The
 989          * returned inode will be referenced and shared-locked to prevent
 990          * it from being moved to the flusher.
 991          */
 992         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
 993                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
 994                                     NULL, &nip);
 995         if (error) {
 996                 hkprintf("hammer_create_inode error %d\n", error);
 997                 hammer_done_transaction(&trans);
 998                 *ap->a_vpp = NULL;
 999                 lwkt_reltoken(&hmp->fs_token);
1000                 return (error);
1001         }
1002
1003         /*
1004          * Add the new filesystem object to the directory.  This will also
1005          * bump the inode's link count.
1006          */
1007         error = hammer_ip_add_directory(&trans, dip,
1008                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1009                                         nip);
1010         if (error)
1011                 hkprintf("hammer_ip_add_directory error %d\n", error);
1012
1013         /*
1014          * Finish up.
1015          */
1016         if (error) {
1017                 hammer_rel_inode(nip, 0);
1018                 hammer_done_transaction(&trans);
1019                 *ap->a_vpp = NULL;
1020         } else {
1021                 error = hammer_get_vnode(nip, ap->a_vpp);
1022                 hammer_done_transaction(&trans);
1023                 hammer_rel_inode(nip, 0);
1024                 if (error == 0) {
1025                         cache_setunresolved(ap->a_nch);
1026                         cache_setvp(ap->a_nch, *ap->a_vpp);
1027                 }
1028                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1029         }
1030         lwkt_reltoken(&hmp->fs_token);
1031         return (error);
1032 }
1033
1034 /*
1035  * hammer_vop_getattr { vp, vap }
1036  *
1037  * Retrieve an inode's attribute information.  When accessing inodes
1038  * historically we fake the atime field to ensure consistent results.
1039  * The atime field is stored in the B-Tree element and allowed to be
1040  * updated without cycling the element.
1041  *
1042  * MPSAFE - does not require fs_token
1043  */
1044 static
1045 int
1046 hammer_vop_getattr(struct vop_getattr_args *ap)
1047 {
1048         struct hammer_inode *ip = VTOI(ap->a_vp);
1049         struct vattr *vap = ap->a_vap;
1050
1051         /*
1052          * We want the fsid to be different when accessing a filesystem
1053          * with different as-of's so programs like diff don't think
1054          * the files are the same.
1055          *
1056          * We also want the fsid to be the same when comparing snapshots,
1057          * or when comparing mirrors (which might be backed by different
1058          * physical devices).  HAMMER fsids are based on the PFS's
1059          * shared_uuid field.
1060          *
1061          * XXX there is a chance of collision here.  The va_fsid reported
1062          * by stat is different from the more involved fsid used in the
1063          * mount structure.
1064          */
1065         ++hammer_stats_file_iopsr;
1066         hammer_lock_sh(&ip->lock);
1067         vap->va_fsid = ip->pfsm->fsid_udev ^ (u_int32_t)ip->obj_asof ^
1068                        (u_int32_t)(ip->obj_asof >> 32);
1069
1070         vap->va_fileid = ip->ino_leaf.base.obj_id;
1071         vap->va_mode = ip->ino_data.mode;
1072         vap->va_nlink = ip->ino_data.nlinks;
1073         vap->va_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1074         vap->va_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1075         vap->va_rmajor = 0;
1076         vap->va_rminor = 0;
1077         vap->va_size = ip->ino_data.size;
1078
1079         /*
1080          * Special case for @@PFS softlinks.  The actual size of the
1081          * expanded softlink is "@@0x%016llx:%05d" == 26 bytes.
1082          * or for MAX_TID is    "@@-1:%05d" == 10 bytes.
1083          */
1084         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_SOFTLINK &&
1085             ip->ino_data.size == 10 &&
1086             ip->obj_asof == HAMMER_MAX_TID &&
1087             ip->obj_localization == 0 &&
1088             strncmp(ip->ino_data.ext.symlink, "@@PFS", 5) == 0) {
1089                     if (ip->pfsm->pfsd.mirror_flags & HAMMER_PFSD_SLAVE)
1090                             vap->va_size = 26;
1091                     else
1092                             vap->va_size = 10;
1093         }
1094
1095         /*
1096          * We must provide a consistent atime and mtime for snapshots
1097          * so people can do a 'tar cf - ... | md5' on them and get
1098          * consistent results.
1099          */
1100         if (ip->flags & HAMMER_INODE_RO) {
1101                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_atime);
1102                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_mtime);
1103         } else {
1104                 hammer_time_to_timespec(ip->ino_data.atime, &vap->va_atime);
1105                 hammer_time_to_timespec(ip->ino_data.mtime, &vap->va_mtime);
1106         }
1107         hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_ctime);
1108         vap->va_flags = ip->ino_data.uflags;
1109         vap->va_gen = 1;        /* hammer inums are unique for all time */
1110         vap->va_blocksize = HAMMER_BUFSIZE;
1111         if (ip->ino_data.size >= HAMMER_XDEMARC) {
1112                 vap->va_bytes = (ip->ino_data.size + HAMMER_XBUFMASK64) &
1113                                 ~HAMMER_XBUFMASK64;
1114         } else if (ip->ino_data.size > HAMMER_BUFSIZE / 2) {
1115                 vap->va_bytes = (ip->ino_data.size + HAMMER_BUFMASK64) &
1116                                 ~HAMMER_BUFMASK64;
1117         } else {
1118                 vap->va_bytes = (ip->ino_data.size + 15) & ~15;
1119         }
1120
1121         vap->va_type = hammer_get_vnode_type(ip->ino_data.obj_type);
1122         vap->va_filerev = 0;    /* XXX */
1123         vap->va_uid_uuid = ip->ino_data.uid;
1124         vap->va_gid_uuid = ip->ino_data.gid;
1125         vap->va_fsid_uuid = ip->hmp->fsid;
1126         vap->va_vaflags = VA_UID_UUID_VALID | VA_GID_UUID_VALID |
1127                           VA_FSID_UUID_VALID;
1128
1129         switch (ip->ino_data.obj_type) {
1130         case HAMMER_OBJTYPE_CDEV:
1131         case HAMMER_OBJTYPE_BDEV:
1132                 vap->va_rmajor = ip->ino_data.rmajor;
1133                 vap->va_rminor = ip->ino_data.rminor;
1134                 break;
1135         default:
1136                 break;
1137         }
1138         hammer_unlock(&ip->lock);
1139         return(0);
1140 }
1141
1142 /*
1143  * hammer_vop_nresolve { nch, dvp, cred }
1144  *
1145  * Locate the requested directory entry.
1146  */
1147 static
1148 int
1149 hammer_vop_nresolve(struct vop_nresolve_args *ap)
1150 {
1151         struct hammer_transaction trans;
1152         struct namecache *ncp;
1153         hammer_mount_t hmp;
1154         hammer_inode_t dip;
1155         hammer_inode_t ip;
1156         hammer_tid_t asof;
1157         struct hammer_cursor cursor;
1158         struct vnode *vp;
1159         int64_t namekey;
1160         int error;
1161         int i;
1162         int nlen;
1163         int flags;
1164         int ispfs;
1165         int64_t obj_id;
1166         u_int32_t localization;
1167         u_int32_t max_iterations;
1168
1169         /*
1170          * Misc initialization, plus handle as-of name extensions.  Look for
1171          * the '@@' extension.  Note that as-of files and directories cannot
1172          * be modified.
1173          */
1174         dip = VTOI(ap->a_dvp);
1175         ncp = ap->a_nch->ncp;
1176         asof = dip->obj_asof;
1177         localization = dip->obj_localization;   /* for code consistency */
1178         nlen = ncp->nc_nlen;
1179         flags = dip->flags & HAMMER_INODE_RO;
1180         ispfs = 0;
1181         hmp = dip->hmp;
1182
1183         lwkt_gettoken(&hmp->fs_token);
1184         hammer_simple_transaction(&trans, hmp);
1185         ++hammer_stats_file_iopsr;
1186
1187         for (i = 0; i < nlen; ++i) {
1188                 if (ncp->nc_name[i] == '@' && ncp->nc_name[i+1] == '@') {
1189                         error = hammer_str_to_tid(ncp->nc_name + i + 2,
1190                                                   &ispfs, &asof, &localization);
1191                         if (error != 0) {
1192                                 i = nlen;
1193                                 break;
1194                         }
1195                         if (asof != HAMMER_MAX_TID)
1196                                 flags |= HAMMER_INODE_RO;
1197                         break;
1198                 }
1199         }
1200         nlen = i;
1201
1202         /*
1203          * If this is a PFS softlink we dive into the PFS
1204          */
1205         if (ispfs && nlen == 0) {
1206                 ip = hammer_get_inode(&trans, dip, HAMMER_OBJID_ROOT,
1207                                       asof, localization,
1208                                       flags, &error);
1209                 if (error == 0) {
1210                         error = hammer_get_vnode(ip, &vp);
1211                         hammer_rel_inode(ip, 0);
1212                 } else {
1213                         vp = NULL;
1214                 }
1215                 if (error == 0) {
1216                         vn_unlock(vp);
1217                         cache_setvp(ap->a_nch, vp);
1218                         vrele(vp);
1219                 }
1220                 goto done;
1221         }
1222
1223         /*
1224          * If there is no path component the time extension is relative to dip.
1225          * e.g. "fubar/@@<snapshot>"
1226          *
1227          * "." is handled by the kernel, but ".@@<snapshot>" is not.
1228          * e.g. "fubar/.@@<snapshot>"
1229          *
1230          * ".." is handled by the kernel.  We do not currently handle
1231          * "..@<snapshot>".
1232          */
1233         if (nlen == 0 || (nlen == 1 && ncp->nc_name[0] == '.')) {
1234                 ip = hammer_get_inode(&trans, dip, dip->obj_id,
1235                                       asof, dip->obj_localization,
1236                                       flags, &error);
1237                 if (error == 0) {
1238                         error = hammer_get_vnode(ip, &vp);
1239                         hammer_rel_inode(ip, 0);
1240                 } else {
1241                         vp = NULL;
1242                 }
1243                 if (error == 0) {
1244                         vn_unlock(vp);
1245                         cache_setvp(ap->a_nch, vp);
1246                         vrele(vp);
1247                 }
1248                 goto done;
1249         }
1250
1251         /*
1252          * Calculate the namekey and setup the key range for the scan.  This
1253          * works kinda like a chained hash table where the lower 32 bits
1254          * of the namekey synthesize the chain.
1255          *
1256          * The key range is inclusive of both key_beg and key_end.
1257          */
1258         namekey = hammer_directory_namekey(dip, ncp->nc_name, nlen,
1259                                            &max_iterations);
1260
1261         error = hammer_init_cursor(&trans, &cursor, &dip->cache[1], dip);
1262         cursor.key_beg.localization = dip->obj_localization +
1263                                       hammer_dir_localization(dip);
1264         cursor.key_beg.obj_id = dip->obj_id;
1265         cursor.key_beg.key = namekey;
1266         cursor.key_beg.create_tid = 0;
1267         cursor.key_beg.delete_tid = 0;
1268         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1269         cursor.key_beg.obj_type = 0;
1270
1271         cursor.key_end = cursor.key_beg;
1272         cursor.key_end.key += max_iterations;
1273         cursor.asof = asof;
1274         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1275
1276         /*
1277          * Scan all matching records (the chain), locate the one matching
1278          * the requested path component.
1279          *
1280          * The hammer_ip_*() functions merge in-memory records with on-disk
1281          * records for the purposes of the search.
1282          */
1283         obj_id = 0;
1284         localization = HAMMER_DEF_LOCALIZATION;
1285
1286         if (error == 0) {
1287                 error = hammer_ip_first(&cursor);
1288                 while (error == 0) {
1289                         error = hammer_ip_resolve_data(&cursor);
1290                         if (error)
1291                                 break;
1292                         if (nlen == cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF &&
1293                             bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1294                                 obj_id = cursor.data->entry.obj_id;
1295                                 localization = cursor.data->entry.localization;
1296                                 break;
1297                         }
1298                         error = hammer_ip_next(&cursor);
1299                 }
1300         }
1301         hammer_done_cursor(&cursor);
1302
1303         /*
1304          * Lookup the obj_id.  This should always succeed.  If it does not
1305          * the filesystem may be damaged and we return a dummy inode.
1306          */
1307         if (error == 0) {
1308                 ip = hammer_get_inode(&trans, dip, obj_id,
1309                                       asof, localization,
1310                                       flags, &error);
1311                 if (error == ENOENT) {
1312                         kprintf("HAMMER: WARNING: Missing "
1313                                 "inode for dirent \"%s\"\n"
1314                                 "\tobj_id = %016llx, asof=%016llx, lo=%08x\n",
1315                                 ncp->nc_name,
1316                                 (long long)obj_id, (long long)asof,
1317                                 localization);
1318                         error = 0;
1319                         ip = hammer_get_dummy_inode(&trans, dip, obj_id,
1320                                                     asof, localization,
1321                                                     flags, &error);
1322                 }
1323                 if (error == 0) {
1324                         error = hammer_get_vnode(ip, &vp);
1325                         hammer_rel_inode(ip, 0);
1326                 } else {
1327                         vp = NULL;
1328                 }
1329                 if (error == 0) {
1330                         vn_unlock(vp);
1331                         cache_setvp(ap->a_nch, vp);
1332                         vrele(vp);
1333                 }
1334         } else if (error == ENOENT) {
1335                 cache_setvp(ap->a_nch, NULL);
1336         }
1337 done:
1338         hammer_done_transaction(&trans);
1339         lwkt_reltoken(&hmp->fs_token);
1340         return (error);
1341 }
1342
1343 /*
1344  * hammer_vop_nlookupdotdot { dvp, vpp, cred }
1345  *
1346  * Locate the parent directory of a directory vnode.
1347  *
1348  * dvp is referenced but not locked.  *vpp must be returned referenced and
1349  * locked.  A parent_obj_id of 0 does not necessarily indicate that we are
1350  * at the root, instead it could indicate that the directory we were in was
1351  * removed.
1352  *
1353  * NOTE: as-of sequences are not linked into the directory structure.  If
1354  * we are at the root with a different asof then the mount point, reload
1355  * the same directory with the mount point's asof.   I'm not sure what this
1356  * will do to NFS.  We encode ASOF stamps in NFS file handles so it might not
1357  * get confused, but it hasn't been tested.
1358  */
1359 static
1360 int
1361 hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *ap)
1362 {
1363         struct hammer_transaction trans;
1364         struct hammer_inode *dip;
1365         struct hammer_inode *ip;
1366         hammer_mount_t hmp;
1367         int64_t parent_obj_id;
1368         u_int32_t parent_obj_localization;
1369         hammer_tid_t asof;
1370         int error;
1371
1372         dip = VTOI(ap->a_dvp);
1373         asof = dip->obj_asof;
1374         hmp = dip->hmp;
1375
1376         /*
1377          * Whos are parent?  This could be the root of a pseudo-filesystem
1378          * whos parent is in another localization domain.
1379          */
1380         lwkt_gettoken(&hmp->fs_token);
1381         parent_obj_id = dip->ino_data.parent_obj_id;
1382         if (dip->obj_id == HAMMER_OBJID_ROOT)
1383                 parent_obj_localization = dip->ino_data.ext.obj.parent_obj_localization;
1384         else
1385                 parent_obj_localization = dip->obj_localization;
1386
1387         if (parent_obj_id == 0) {
1388                 if (dip->obj_id == HAMMER_OBJID_ROOT &&
1389                    asof != hmp->asof) {
1390                         parent_obj_id = dip->obj_id;
1391                         asof = hmp->asof;
1392                         *ap->a_fakename = kmalloc(19, M_TEMP, M_WAITOK);
1393                         ksnprintf(*ap->a_fakename, 19, "0x%016llx",
1394                                   (long long)dip->obj_asof);
1395                 } else {
1396                         *ap->a_vpp = NULL;
1397                         lwkt_reltoken(&hmp->fs_token);
1398                         return ENOENT;
1399                 }
1400         }
1401
1402         hammer_simple_transaction(&trans, hmp);
1403         ++hammer_stats_file_iopsr;
1404
1405         ip = hammer_get_inode(&trans, dip, parent_obj_id,
1406                               asof, parent_obj_localization,
1407                               dip->flags, &error);
1408         if (ip) {
1409                 error = hammer_get_vnode(ip, ap->a_vpp);
1410                 hammer_rel_inode(ip, 0);
1411         } else {
1412                 *ap->a_vpp = NULL;
1413         }
1414         hammer_done_transaction(&trans);
1415         lwkt_reltoken(&hmp->fs_token);
1416         return (error);
1417 }
1418
1419 /*
1420  * hammer_vop_nlink { nch, dvp, vp, cred }
1421  */
1422 static
1423 int
1424 hammer_vop_nlink(struct vop_nlink_args *ap)
1425 {
1426         struct hammer_transaction trans;
1427         struct hammer_inode *dip;
1428         struct hammer_inode *ip;
1429         struct nchandle *nch;
1430         hammer_mount_t hmp;
1431         int error;
1432
1433         if (ap->a_dvp->v_mount != ap->a_vp->v_mount)
1434                 return(EXDEV);
1435
1436         nch = ap->a_nch;
1437         dip = VTOI(ap->a_dvp);
1438         ip = VTOI(ap->a_vp);
1439         hmp = dip->hmp;
1440
1441         if (dip->obj_localization != ip->obj_localization)
1442                 return(EXDEV);
1443
1444         if (dip->flags & HAMMER_INODE_RO)
1445                 return (EROFS);
1446         if (ip->flags & HAMMER_INODE_RO)
1447                 return (EROFS);
1448         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1449                 return (error);
1450
1451         /*
1452          * Create a transaction to cover the operations we perform.
1453          */
1454         lwkt_gettoken(&hmp->fs_token);
1455         hammer_start_transaction(&trans, hmp);
1456         ++hammer_stats_file_iopsw;
1457
1458         /*
1459          * Add the filesystem object to the directory.  Note that neither
1460          * dip nor ip are referenced or locked, but their vnodes are
1461          * referenced.  This function will bump the inode's link count.
1462          */
1463         error = hammer_ip_add_directory(&trans, dip,
1464                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1465                                         ip);
1466
1467         /*
1468          * Finish up.
1469          */
1470         if (error == 0) {
1471                 cache_setunresolved(nch);
1472                 cache_setvp(nch, ap->a_vp);
1473         }
1474         hammer_done_transaction(&trans);
1475         hammer_knote(ap->a_vp, NOTE_LINK);
1476         hammer_knote(ap->a_dvp, NOTE_WRITE);
1477         lwkt_reltoken(&hmp->fs_token);
1478         return (error);
1479 }
1480
1481 /*
1482  * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
1483  *
1484  * The operating system has already ensured that the directory entry
1485  * does not exist and done all appropriate namespace locking.
1486  */
1487 static
1488 int
1489 hammer_vop_nmkdir(struct vop_nmkdir_args *ap)
1490 {
1491         struct hammer_transaction trans;
1492         struct hammer_inode *dip;
1493         struct hammer_inode *nip;
1494         struct nchandle *nch;
1495         hammer_mount_t hmp;
1496         int error;
1497
1498         nch = ap->a_nch;
1499         dip = VTOI(ap->a_dvp);
1500         hmp = dip->hmp;
1501
1502         if (dip->flags & HAMMER_INODE_RO)
1503                 return (EROFS);
1504         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1505                 return (error);
1506
1507         /*
1508          * Create a transaction to cover the operations we perform.
1509          */
1510         lwkt_gettoken(&hmp->fs_token);
1511         hammer_start_transaction(&trans, hmp);
1512         ++hammer_stats_file_iopsw;
1513
1514         /*
1515          * Create a new filesystem object of the requested type.  The
1516          * returned inode will be referenced but not locked.
1517          */
1518         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1519                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1520                                     NULL, &nip);
1521         if (error) {
1522                 hkprintf("hammer_mkdir error %d\n", error);
1523                 hammer_done_transaction(&trans);
1524                 *ap->a_vpp = NULL;
1525                 lwkt_reltoken(&hmp->fs_token);
1526                 return (error);
1527         }
1528         /*
1529          * Add the new filesystem object to the directory.  This will also
1530          * bump the inode's link count.
1531          */
1532         error = hammer_ip_add_directory(&trans, dip,
1533                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1534                                         nip);
1535         if (error)
1536                 hkprintf("hammer_mkdir (add) error %d\n", error);
1537
1538         /*
1539          * Finish up.
1540          */
1541         if (error) {
1542                 hammer_rel_inode(nip, 0);
1543                 *ap->a_vpp = NULL;
1544         } else {
1545                 error = hammer_get_vnode(nip, ap->a_vpp);
1546                 hammer_rel_inode(nip, 0);
1547                 if (error == 0) {
1548                         cache_setunresolved(ap->a_nch);
1549                         cache_setvp(ap->a_nch, *ap->a_vpp);
1550                 }
1551         }
1552         hammer_done_transaction(&trans);
1553         if (error == 0)
1554                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
1555         lwkt_reltoken(&hmp->fs_token);
1556         return (error);
1557 }
1558
1559 /*
1560  * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1561  *
1562  * The operating system has already ensured that the directory entry
1563  * does not exist and done all appropriate namespace locking.
1564  */
1565 static
1566 int
1567 hammer_vop_nmknod(struct vop_nmknod_args *ap)
1568 {
1569         struct hammer_transaction trans;
1570         struct hammer_inode *dip;
1571         struct hammer_inode *nip;
1572         struct nchandle *nch;
1573         hammer_mount_t hmp;
1574         int error;
1575
1576         nch = ap->a_nch;
1577         dip = VTOI(ap->a_dvp);
1578         hmp = dip->hmp;
1579
1580         if (dip->flags & HAMMER_INODE_RO)
1581                 return (EROFS);
1582         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1583                 return (error);
1584
1585         /*
1586          * Create a transaction to cover the operations we perform.
1587          */
1588         lwkt_gettoken(&hmp->fs_token);
1589         hammer_start_transaction(&trans, hmp);
1590         ++hammer_stats_file_iopsw;
1591
1592         /*
1593          * Create a new filesystem object of the requested type.  The
1594          * returned inode will be referenced but not locked.
1595          *
1596          * If mknod specifies a directory a pseudo-fs is created.
1597          */
1598         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1599                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1600                                     NULL, &nip);
1601         if (error) {
1602                 hammer_done_transaction(&trans);
1603                 *ap->a_vpp = NULL;
1604                 lwkt_reltoken(&hmp->fs_token);
1605                 return (error);
1606         }
1607
1608         /*
1609          * Add the new filesystem object to the directory.  This will also
1610          * bump the inode's link count.
1611          */
1612         error = hammer_ip_add_directory(&trans, dip,
1613                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1614                                         nip);
1615
1616         /*
1617          * Finish up.
1618          */
1619         if (error) {
1620                 hammer_rel_inode(nip, 0);
1621                 *ap->a_vpp = NULL;
1622         } else {
1623                 error = hammer_get_vnode(nip, ap->a_vpp);
1624                 hammer_rel_inode(nip, 0);
1625                 if (error == 0) {
1626                         cache_setunresolved(ap->a_nch);
1627                         cache_setvp(ap->a_nch, *ap->a_vpp);
1628                 }
1629         }
1630         hammer_done_transaction(&trans);
1631         if (error == 0)
1632                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1633         lwkt_reltoken(&hmp->fs_token);
1634         return (error);
1635 }
1636
1637 /*
1638  * hammer_vop_open { vp, mode, cred, fp }
1639  *
1640  * MPSAFE (does not require fs_token)
1641  */
1642 static
1643 int
1644 hammer_vop_open(struct vop_open_args *ap)
1645 {
1646         hammer_inode_t ip;
1647
1648         ++hammer_stats_file_iopsr;
1649         ip = VTOI(ap->a_vp);
1650
1651         if ((ap->a_mode & FWRITE) && (ip->flags & HAMMER_INODE_RO))
1652                 return (EROFS);
1653         return(vop_stdopen(ap));
1654 }
1655
1656 /*
1657  * hammer_vop_print { vp }
1658  */
1659 static
1660 int
1661 hammer_vop_print(struct vop_print_args *ap)
1662 {
1663         return EOPNOTSUPP;
1664 }
1665
1666 /*
1667  * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1668  */
1669 static
1670 int
1671 hammer_vop_readdir(struct vop_readdir_args *ap)
1672 {
1673         struct hammer_transaction trans;
1674         struct hammer_cursor cursor;
1675         struct hammer_inode *ip;
1676         hammer_mount_t hmp;
1677         struct uio *uio;
1678         hammer_base_elm_t base;
1679         int error;
1680         int cookie_index;
1681         int ncookies;
1682         off_t *cookies;
1683         off_t saveoff;
1684         int r;
1685         int dtype;
1686
1687         ++hammer_stats_file_iopsr;
1688         ip = VTOI(ap->a_vp);
1689         uio = ap->a_uio;
1690         saveoff = uio->uio_offset;
1691         hmp = ip->hmp;
1692
1693         if (ap->a_ncookies) {
1694                 ncookies = uio->uio_resid / 16 + 1;
1695                 if (ncookies > 1024)
1696                         ncookies = 1024;
1697                 cookies = kmalloc(ncookies * sizeof(off_t), M_TEMP, M_WAITOK);
1698                 cookie_index = 0;
1699         } else {
1700                 ncookies = -1;
1701                 cookies = NULL;
1702                 cookie_index = 0;
1703         }
1704
1705         lwkt_gettoken(&hmp->fs_token);
1706         hammer_simple_transaction(&trans, hmp);
1707
1708         /*
1709          * Handle artificial entries
1710          *
1711          * It should be noted that the minimum value for a directory
1712          * hash key on-media is 0x0000000100000000, so we can use anything
1713          * less then that to represent our 'special' key space.
1714          */
1715         error = 0;
1716         if (saveoff == 0) {
1717                 r = vop_write_dirent(&error, uio, ip->obj_id, DT_DIR, 1, ".");
1718                 if (r)
1719                         goto done;
1720                 if (cookies)
1721                         cookies[cookie_index] = saveoff;
1722                 ++saveoff;
1723                 ++cookie_index;
1724                 if (cookie_index == ncookies)
1725                         goto done;
1726         }
1727         if (saveoff == 1) {
1728                 if (ip->ino_data.parent_obj_id) {
1729                         r = vop_write_dirent(&error, uio,
1730                                              ip->ino_data.parent_obj_id,
1731                                              DT_DIR, 2, "..");
1732                 } else {
1733                         r = vop_write_dirent(&error, uio,
1734                                              ip->obj_id, DT_DIR, 2, "..");
1735                 }
1736                 if (r)
1737                         goto done;
1738                 if (cookies)
1739                         cookies[cookie_index] = saveoff;
1740                 ++saveoff;
1741                 ++cookie_index;
1742                 if (cookie_index == ncookies)
1743                         goto done;
1744         }
1745
1746         /*
1747          * Key range (begin and end inclusive) to scan.  Directory keys
1748          * directly translate to a 64 bit 'seek' position.
1749          */
1750         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1751         cursor.key_beg.localization = ip->obj_localization +
1752                                       hammer_dir_localization(ip);
1753         cursor.key_beg.obj_id = ip->obj_id;
1754         cursor.key_beg.create_tid = 0;
1755         cursor.key_beg.delete_tid = 0;
1756         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1757         cursor.key_beg.obj_type = 0;
1758         cursor.key_beg.key = saveoff;
1759
1760         cursor.key_end = cursor.key_beg;
1761         cursor.key_end.key = HAMMER_MAX_KEY;
1762         cursor.asof = ip->obj_asof;
1763         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1764
1765         error = hammer_ip_first(&cursor);
1766
1767         while (error == 0) {
1768                 error = hammer_ip_resolve_data(&cursor);
1769                 if (error)
1770                         break;
1771                 base = &cursor.leaf->base;
1772                 saveoff = base->key;
1773                 KKASSERT(cursor.leaf->data_len > HAMMER_ENTRY_NAME_OFF);
1774
1775                 if (base->obj_id != ip->obj_id)
1776                         panic("readdir: bad record at %p", cursor.node);
1777
1778                 /*
1779                  * Convert pseudo-filesystems into softlinks
1780                  */
1781                 dtype = hammer_get_dtype(cursor.leaf->base.obj_type);
1782                 r = vop_write_dirent(
1783                              &error, uio, cursor.data->entry.obj_id,
1784                              dtype,
1785                              cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF ,
1786                              (void *)cursor.data->entry.name);
1787                 if (r)
1788                         break;
1789                 ++saveoff;
1790                 if (cookies)
1791                         cookies[cookie_index] = base->key;
1792                 ++cookie_index;
1793                 if (cookie_index == ncookies)
1794                         break;
1795                 error = hammer_ip_next(&cursor);
1796         }
1797         hammer_done_cursor(&cursor);
1798
1799 done:
1800         hammer_done_transaction(&trans);
1801
1802         if (ap->a_eofflag)
1803                 *ap->a_eofflag = (error == ENOENT);
1804         uio->uio_offset = saveoff;
1805         if (error && cookie_index == 0) {
1806                 if (error == ENOENT)
1807                         error = 0;
1808                 if (cookies) {
1809                         kfree(cookies, M_TEMP);
1810                         *ap->a_ncookies = 0;
1811                         *ap->a_cookies = NULL;
1812                 }
1813         } else {
1814                 if (error == ENOENT)
1815                         error = 0;
1816                 if (cookies) {
1817                         *ap->a_ncookies = cookie_index;
1818                         *ap->a_cookies = cookies;
1819                 }
1820         }
1821         lwkt_reltoken(&hmp->fs_token);
1822         return(error);
1823 }
1824
1825 /*
1826  * hammer_vop_readlink { vp, uio, cred }
1827  */
1828 static
1829 int
1830 hammer_vop_readlink(struct vop_readlink_args *ap)
1831 {
1832         struct hammer_transaction trans;
1833         struct hammer_cursor cursor;
1834         struct hammer_inode *ip;
1835         hammer_mount_t hmp;
1836         char buf[32];
1837         u_int32_t localization;
1838         hammer_pseudofs_inmem_t pfsm;
1839         int error;
1840
1841         ip = VTOI(ap->a_vp);
1842         hmp = ip->hmp;
1843
1844         lwkt_gettoken(&hmp->fs_token);
1845
1846         /*
1847          * Shortcut if the symlink data was stuffed into ino_data.
1848          *
1849          * Also expand special "@@PFS%05d" softlinks (expansion only
1850          * occurs for non-historical (current) accesses made from the
1851          * primary filesystem).
1852          */
1853         if (ip->ino_data.size <= HAMMER_INODE_BASESYMLEN) {
1854                 char *ptr;
1855                 int bytes;
1856
1857                 ptr = ip->ino_data.ext.symlink;
1858                 bytes = (int)ip->ino_data.size;
1859                 if (bytes == 10 &&
1860                     ip->obj_asof == HAMMER_MAX_TID &&
1861                     ip->obj_localization == 0 &&
1862                     strncmp(ptr, "@@PFS", 5) == 0) {
1863                         hammer_simple_transaction(&trans, hmp);
1864                         bcopy(ptr + 5, buf, 5);
1865                         buf[5] = 0;
1866                         localization = strtoul(buf, NULL, 10) << 16;
1867                         pfsm = hammer_load_pseudofs(&trans, localization,
1868                                                     &error);
1869                         if (error == 0) {
1870                                 if (pfsm->pfsd.mirror_flags &
1871                                     HAMMER_PFSD_SLAVE) {
1872                                         /* vap->va_size == 26 */
1873                                         ksnprintf(buf, sizeof(buf),
1874                                                   "@@0x%016llx:%05d",
1875                                                   (long long)pfsm->pfsd.sync_end_tid,
1876                                                   localization >> 16);
1877                                 } else {
1878                                         /* vap->va_size == 10 */
1879                                         ksnprintf(buf, sizeof(buf),
1880                                                   "@@-1:%05d",
1881                                                   localization >> 16);
1882 #if 0
1883                                         ksnprintf(buf, sizeof(buf),
1884                                                   "@@0x%016llx:%05d",
1885                                                   (long long)HAMMER_MAX_TID,
1886                                                   localization >> 16);
1887 #endif
1888                                 }
1889                                 ptr = buf;
1890                                 bytes = strlen(buf);
1891                         }
1892                         if (pfsm)
1893                                 hammer_rel_pseudofs(hmp, pfsm);
1894                         hammer_done_transaction(&trans);
1895                 }
1896                 error = uiomove(ptr, bytes, ap->a_uio);
1897                 lwkt_reltoken(&hmp->fs_token);
1898                 return(error);
1899         }
1900
1901         /*
1902          * Long version
1903          */
1904         hammer_simple_transaction(&trans, hmp);
1905         ++hammer_stats_file_iopsr;
1906         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1907
1908         /*
1909          * Key range (begin and end inclusive) to scan.  Directory keys
1910          * directly translate to a 64 bit 'seek' position.
1911          */
1912         cursor.key_beg.localization = ip->obj_localization +
1913                                       HAMMER_LOCALIZE_MISC;
1914         cursor.key_beg.obj_id = ip->obj_id;
1915         cursor.key_beg.create_tid = 0;
1916         cursor.key_beg.delete_tid = 0;
1917         cursor.key_beg.rec_type = HAMMER_RECTYPE_FIX;
1918         cursor.key_beg.obj_type = 0;
1919         cursor.key_beg.key = HAMMER_FIXKEY_SYMLINK;
1920         cursor.asof = ip->obj_asof;
1921         cursor.flags |= HAMMER_CURSOR_ASOF;
1922
1923         error = hammer_ip_lookup(&cursor);
1924         if (error == 0) {
1925                 error = hammer_ip_resolve_data(&cursor);
1926                 if (error == 0) {
1927                         KKASSERT(cursor.leaf->data_len >=
1928                                  HAMMER_SYMLINK_NAME_OFF);
1929                         error = uiomove(cursor.data->symlink.name,
1930                                         cursor.leaf->data_len -
1931                                                 HAMMER_SYMLINK_NAME_OFF,
1932                                         ap->a_uio);
1933                 }
1934         }
1935         hammer_done_cursor(&cursor);
1936         hammer_done_transaction(&trans);
1937         lwkt_reltoken(&hmp->fs_token);
1938         return(error);
1939 }
1940
1941 /*
1942  * hammer_vop_nremove { nch, dvp, cred }
1943  */
1944 static
1945 int
1946 hammer_vop_nremove(struct vop_nremove_args *ap)
1947 {
1948         struct hammer_transaction trans;
1949         struct hammer_inode *dip;
1950         hammer_mount_t hmp;
1951         int error;
1952
1953         dip = VTOI(ap->a_dvp);
1954         hmp = dip->hmp;
1955
1956         if (hammer_nohistory(dip) == 0 &&
1957             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1958                 return (error);
1959         }
1960
1961         lwkt_gettoken(&hmp->fs_token);
1962         hammer_start_transaction(&trans, hmp);
1963         ++hammer_stats_file_iopsw;
1964         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 0);
1965         hammer_done_transaction(&trans);
1966         if (error == 0)
1967                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1968         lwkt_reltoken(&hmp->fs_token);
1969         return (error);
1970 }
1971
1972 /*
1973  * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1974  */
1975 static
1976 int
1977 hammer_vop_nrename(struct vop_nrename_args *ap)
1978 {
1979         struct hammer_transaction trans;
1980         struct namecache *fncp;
1981         struct namecache *tncp;
1982         struct hammer_inode *fdip;
1983         struct hammer_inode *tdip;
1984         struct hammer_inode *ip;
1985         hammer_mount_t hmp;
1986         struct hammer_cursor cursor;
1987         int64_t namekey;
1988         u_int32_t max_iterations;
1989         int nlen, error;
1990
1991         if (ap->a_fdvp->v_mount != ap->a_tdvp->v_mount)
1992                 return(EXDEV);
1993         if (ap->a_fdvp->v_mount != ap->a_fnch->ncp->nc_vp->v_mount)
1994                 return(EXDEV);
1995
1996         fdip = VTOI(ap->a_fdvp);
1997         tdip = VTOI(ap->a_tdvp);
1998         fncp = ap->a_fnch->ncp;
1999         tncp = ap->a_tnch->ncp;
2000         ip = VTOI(fncp->nc_vp);
2001         KKASSERT(ip != NULL);
2002
2003         hmp = ip->hmp;
2004
2005         if (fdip->obj_localization != tdip->obj_localization)
2006                 return(EXDEV);
2007         if (fdip->obj_localization != ip->obj_localization)
2008                 return(EXDEV);
2009
2010         if (fdip->flags & HAMMER_INODE_RO)
2011                 return (EROFS);
2012         if (tdip->flags & HAMMER_INODE_RO)
2013                 return (EROFS);
2014         if (ip->flags & HAMMER_INODE_RO)
2015                 return (EROFS);
2016         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
2017                 return (error);
2018
2019         lwkt_gettoken(&hmp->fs_token);
2020         hammer_start_transaction(&trans, hmp);
2021         ++hammer_stats_file_iopsw;
2022
2023         /*
2024          * Remove tncp from the target directory and then link ip as
2025          * tncp. XXX pass trans to dounlink
2026          *
2027          * Force the inode sync-time to match the transaction so it is
2028          * in-sync with the creation of the target directory entry.
2029          */
2030         error = hammer_dounlink(&trans, ap->a_tnch, ap->a_tdvp,
2031                                 ap->a_cred, 0, -1);
2032         if (error == 0 || error == ENOENT) {
2033                 error = hammer_ip_add_directory(&trans, tdip,
2034                                                 tncp->nc_name, tncp->nc_nlen,
2035                                                 ip);
2036                 if (error == 0) {
2037                         ip->ino_data.parent_obj_id = tdip->obj_id;
2038                         ip->ino_data.ctime = trans.time;
2039                         hammer_modify_inode(&trans, ip, HAMMER_INODE_DDIRTY);
2040                 }
2041         }
2042         if (error)
2043                 goto failed; /* XXX */
2044
2045         /*
2046          * Locate the record in the originating directory and remove it.
2047          *
2048          * Calculate the namekey and setup the key range for the scan.  This
2049          * works kinda like a chained hash table where the lower 32 bits
2050          * of the namekey synthesize the chain.
2051          *
2052          * The key range is inclusive of both key_beg and key_end.
2053          */
2054         namekey = hammer_directory_namekey(fdip, fncp->nc_name, fncp->nc_nlen,
2055                                            &max_iterations);
2056 retry:
2057         hammer_init_cursor(&trans, &cursor, &fdip->cache[1], fdip);
2058         cursor.key_beg.localization = fdip->obj_localization +
2059                                       hammer_dir_localization(fdip);
2060         cursor.key_beg.obj_id = fdip->obj_id;
2061         cursor.key_beg.key = namekey;
2062         cursor.key_beg.create_tid = 0;
2063         cursor.key_beg.delete_tid = 0;
2064         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
2065         cursor.key_beg.obj_type = 0;
2066
2067         cursor.key_end = cursor.key_beg;
2068         cursor.key_end.key += max_iterations;
2069         cursor.asof = fdip->obj_asof;
2070         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2071
2072         /*
2073          * Scan all matching records (the chain), locate the one matching
2074          * the requested path component.
2075          *
2076          * The hammer_ip_*() functions merge in-memory records with on-disk
2077          * records for the purposes of the search.
2078          */
2079         error = hammer_ip_first(&cursor);
2080         while (error == 0) {
2081                 if (hammer_ip_resolve_data(&cursor) != 0)
2082                         break;
2083                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
2084                 KKASSERT(nlen > 0);
2085                 if (fncp->nc_nlen == nlen &&
2086                     bcmp(fncp->nc_name, cursor.data->entry.name, nlen) == 0) {
2087                         break;
2088                 }
2089                 error = hammer_ip_next(&cursor);
2090         }
2091
2092         /*
2093          * If all is ok we have to get the inode so we can adjust nlinks.
2094          *
2095          * WARNING: hammer_ip_del_directory() may have to terminate the
2096          * cursor to avoid a recursion.  It's ok to call hammer_done_cursor()
2097          * twice.
2098          */
2099         if (error == 0)
2100                 error = hammer_ip_del_directory(&trans, &cursor, fdip, ip);
2101
2102         /*
2103          * XXX A deadlock here will break rename's atomicy for the purposes
2104          * of crash recovery.
2105          */
2106         if (error == EDEADLK) {
2107                 hammer_done_cursor(&cursor);
2108                 goto retry;
2109         }
2110
2111         /*
2112          * Cleanup and tell the kernel that the rename succeeded.
2113          *
2114          * NOTE: ip->vp, if non-NULL, cannot be directly referenced
2115          *       without formally acquiring the vp since the vp might
2116          *       have zero refs on it, or in the middle of a reclaim,
2117          *       etc.
2118          */
2119         hammer_done_cursor(&cursor);
2120         if (error == 0) {
2121                 cache_rename(ap->a_fnch, ap->a_tnch);
2122                 hammer_knote(ap->a_fdvp, NOTE_WRITE);
2123                 hammer_knote(ap->a_tdvp, NOTE_WRITE);
2124                 while (ip->vp) {
2125                         struct vnode *vp;
2126
2127                         error = hammer_get_vnode(ip, &vp);
2128                         if (error == 0 && vp) {
2129                                 vn_unlock(vp);
2130                                 hammer_knote(ip->vp, NOTE_RENAME);
2131                                 vrele(vp);
2132                                 break;
2133                         }
2134                         kprintf("Debug: HAMMER ip/vp race2 avoided\n");
2135                 }
2136         }
2137
2138 failed:
2139         hammer_done_transaction(&trans);
2140         lwkt_reltoken(&hmp->fs_token);
2141         return (error);
2142 }
2143
2144 /*
2145  * hammer_vop_nrmdir { nch, dvp, cred }
2146  */
2147 static
2148 int
2149 hammer_vop_nrmdir(struct vop_nrmdir_args *ap)
2150 {
2151         struct hammer_transaction trans;
2152         struct hammer_inode *dip;
2153         hammer_mount_t hmp;
2154         int error;
2155
2156         dip = VTOI(ap->a_dvp);
2157         hmp = dip->hmp;
2158
2159         if (hammer_nohistory(dip) == 0 &&
2160             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2161                 return (error);
2162         }
2163
2164         lwkt_gettoken(&hmp->fs_token);
2165         hammer_start_transaction(&trans, hmp);
2166         ++hammer_stats_file_iopsw;
2167         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 1);
2168         hammer_done_transaction(&trans);
2169         if (error == 0)
2170                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
2171         lwkt_reltoken(&hmp->fs_token);
2172         return (error);
2173 }
2174
2175 /*
2176  * hammer_vop_markatime { vp, cred }
2177  */
2178 static
2179 int
2180 hammer_vop_markatime(struct vop_markatime_args *ap)
2181 {
2182         struct hammer_transaction trans;
2183         struct hammer_inode *ip;
2184         hammer_mount_t hmp;
2185
2186         ip = VTOI(ap->a_vp);
2187         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2188                 return (EROFS);
2189         if (ip->flags & HAMMER_INODE_RO)
2190                 return (EROFS);
2191         hmp = ip->hmp;
2192         if (hmp->mp->mnt_flag & MNT_NOATIME)
2193                 return (0);
2194         lwkt_gettoken(&hmp->fs_token);
2195         hammer_start_transaction(&trans, hmp);
2196         ++hammer_stats_file_iopsw;
2197
2198         ip->ino_data.atime = trans.time;
2199         hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
2200         hammer_done_transaction(&trans);
2201         hammer_knote(ap->a_vp, NOTE_ATTRIB);
2202         lwkt_reltoken(&hmp->fs_token);
2203         return (0);
2204 }
2205
2206 /*
2207  * hammer_vop_setattr { vp, vap, cred }
2208  */
2209 static
2210 int
2211 hammer_vop_setattr(struct vop_setattr_args *ap)
2212 {
2213         struct hammer_transaction trans;
2214         struct hammer_inode *ip;
2215         struct vattr *vap;
2216         hammer_mount_t hmp;
2217         int modflags;
2218         int error;
2219         int truncating;
2220         int blksize;
2221         int kflags;
2222 #if 0
2223         int64_t aligned_size;
2224 #endif
2225         u_int32_t flags;
2226
2227         vap = ap->a_vap;
2228         ip = ap->a_vp->v_data;
2229         modflags = 0;
2230         kflags = 0;
2231         hmp = ip->hmp;
2232
2233         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2234                 return(EROFS);
2235         if (ip->flags & HAMMER_INODE_RO)
2236                 return (EROFS);
2237         if (hammer_nohistory(ip) == 0 &&
2238             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2239                 return (error);
2240         }
2241
2242         lwkt_gettoken(&hmp->fs_token);
2243         hammer_start_transaction(&trans, hmp);
2244         ++hammer_stats_file_iopsw;
2245         error = 0;
2246
2247         if (vap->va_flags != VNOVAL) {
2248                 flags = ip->ino_data.uflags;
2249                 error = vop_helper_setattr_flags(&flags, vap->va_flags,
2250                                          hammer_to_unix_xid(&ip->ino_data.uid),
2251                                          ap->a_cred);
2252                 if (error == 0) {
2253                         if (ip->ino_data.uflags != flags) {
2254                                 ip->ino_data.uflags = flags;
2255                                 ip->ino_data.ctime = trans.time;
2256                                 modflags |= HAMMER_INODE_DDIRTY;
2257                                 kflags |= NOTE_ATTRIB;
2258                         }
2259                         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2260                                 error = 0;
2261                                 goto done;
2262                         }
2263                 }
2264                 goto done;
2265         }
2266         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2267                 error = EPERM;
2268                 goto done;
2269         }
2270         if (vap->va_uid != (uid_t)VNOVAL || vap->va_gid != (gid_t)VNOVAL) {
2271                 mode_t cur_mode = ip->ino_data.mode;
2272                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2273                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2274                 uuid_t uuid_uid;
2275                 uuid_t uuid_gid;
2276
2277                 error = vop_helper_chown(ap->a_vp, vap->va_uid, vap->va_gid,
2278                                          ap->a_cred,
2279                                          &cur_uid, &cur_gid, &cur_mode);
2280                 if (error == 0) {
2281                         hammer_guid_to_uuid(&uuid_uid, cur_uid);
2282                         hammer_guid_to_uuid(&uuid_gid, cur_gid);
2283                         if (bcmp(&uuid_uid, &ip->ino_data.uid,
2284                                  sizeof(uuid_uid)) ||
2285                             bcmp(&uuid_gid, &ip->ino_data.gid,
2286                                  sizeof(uuid_gid)) ||
2287                             ip->ino_data.mode != cur_mode
2288                         ) {
2289                                 ip->ino_data.uid = uuid_uid;
2290                                 ip->ino_data.gid = uuid_gid;
2291                                 ip->ino_data.mode = cur_mode;
2292                                 ip->ino_data.ctime = trans.time;
2293                                 modflags |= HAMMER_INODE_DDIRTY;
2294                         }
2295                         kflags |= NOTE_ATTRIB;
2296                 }
2297         }
2298         while (vap->va_size != VNOVAL && ip->ino_data.size != vap->va_size) {
2299                 switch(ap->a_vp->v_type) {
2300                 case VREG:
2301                         if (vap->va_size == ip->ino_data.size)
2302                                 break;
2303
2304                         /*
2305                          * Log the operation if in fast-fsync mode or if
2306                          * there are unterminated redo write records present.
2307                          *
2308                          * The second check is needed so the recovery code
2309                          * properly truncates write redos even if nominal
2310                          * REDO operations is turned off due to excessive
2311                          * writes, because the related records might be
2312                          * destroyed and never lay down a TERM_WRITE.
2313                          */
2314                         if ((ip->flags & HAMMER_INODE_REDO) ||
2315                             (ip->flags & HAMMER_INODE_RDIRTY)) {
2316                                 error = hammer_generate_redo(&trans, ip,
2317                                                              vap->va_size,
2318                                                              HAMMER_REDO_TRUNC,
2319                                                              NULL, 0);
2320                         }
2321                         blksize = hammer_blocksize(vap->va_size);
2322
2323                         /*
2324                          * XXX break atomicy, we can deadlock the backend
2325                          * if we do not release the lock.  Probably not a
2326                          * big deal here.
2327                          */
2328                         if (vap->va_size < ip->ino_data.size) {
2329                                 nvtruncbuf(ap->a_vp, vap->va_size,
2330                                            blksize,
2331                                            hammer_blockoff(vap->va_size),
2332                                            0);
2333                                 truncating = 1;
2334                                 kflags |= NOTE_WRITE;
2335                         } else {
2336                                 nvextendbuf(ap->a_vp,
2337                                             ip->ino_data.size,
2338                                             vap->va_size,
2339                                             hammer_blocksize(ip->ino_data.size),
2340                                             hammer_blocksize(vap->va_size),
2341                                             hammer_blockoff(ip->ino_data.size),
2342                                             hammer_blockoff(vap->va_size),
2343                                             0);
2344                                 truncating = 0;
2345                                 kflags |= NOTE_WRITE | NOTE_EXTEND;
2346                         }
2347                         ip->ino_data.size = vap->va_size;
2348                         ip->ino_data.mtime = trans.time;
2349                         /* XXX safe to use SDIRTY instead of DDIRTY here? */
2350                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2351
2352                         /*
2353                          * On-media truncation is cached in the inode until
2354                          * the inode is synchronized.  We must immediately
2355                          * handle any frontend records.
2356                          */
2357                         if (truncating) {
2358                                 hammer_ip_frontend_trunc(ip, vap->va_size);
2359 #ifdef DEBUG_TRUNCATE
2360                                 if (HammerTruncIp == NULL)
2361                                         HammerTruncIp = ip;
2362 #endif
2363                                 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2364                                         ip->flags |= HAMMER_INODE_TRUNCATED;
2365                                         ip->trunc_off = vap->va_size;
2366 #ifdef DEBUG_TRUNCATE
2367                                         if (ip == HammerTruncIp)
2368                                         kprintf("truncate1 %016llx\n",
2369                                                 (long long)ip->trunc_off);
2370 #endif
2371                                 } else if (ip->trunc_off > vap->va_size) {
2372                                         ip->trunc_off = vap->va_size;
2373 #ifdef DEBUG_TRUNCATE
2374                                         if (ip == HammerTruncIp)
2375                                         kprintf("truncate2 %016llx\n",
2376                                                 (long long)ip->trunc_off);
2377 #endif
2378                                 } else {
2379 #ifdef DEBUG_TRUNCATE
2380                                         if (ip == HammerTruncIp)
2381                                         kprintf("truncate3 %016llx (ignored)\n",
2382                                                 (long long)vap->va_size);
2383 #endif
2384                                 }
2385                         }
2386
2387 #if 0
2388                         /*
2389                          * When truncating, nvtruncbuf() may have cleaned out
2390                          * a portion of the last block on-disk in the buffer
2391                          * cache.  We must clean out any frontend records
2392                          * for blocks beyond the new last block.
2393                          */
2394                         aligned_size = (vap->va_size + (blksize - 1)) &
2395                                        ~(int64_t)(blksize - 1);
2396                         if (truncating && vap->va_size < aligned_size) {
2397                                 aligned_size -= blksize;
2398                                 hammer_ip_frontend_trunc(ip, aligned_size);
2399                         }
2400 #endif
2401                         break;
2402                 case VDATABASE:
2403                         if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2404                                 ip->flags |= HAMMER_INODE_TRUNCATED;
2405                                 ip->trunc_off = vap->va_size;
2406                         } else if (ip->trunc_off > vap->va_size) {
2407                                 ip->trunc_off = vap->va_size;
2408                         }
2409                         hammer_ip_frontend_trunc(ip, vap->va_size);
2410                         ip->ino_data.size = vap->va_size;
2411                         ip->ino_data.mtime = trans.time;
2412                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2413                         kflags |= NOTE_ATTRIB;
2414                         break;
2415                 default:
2416                         error = EINVAL;
2417                         goto done;
2418                 }
2419                 break;
2420         }
2421         if (vap->va_atime.tv_sec != VNOVAL) {
2422                 ip->ino_data.atime = hammer_timespec_to_time(&vap->va_atime);
2423                 modflags |= HAMMER_INODE_ATIME;
2424                 kflags |= NOTE_ATTRIB;
2425         }
2426         if (vap->va_mtime.tv_sec != VNOVAL) {
2427                 ip->ino_data.mtime = hammer_timespec_to_time(&vap->va_mtime);
2428                 modflags |= HAMMER_INODE_MTIME;
2429                 kflags |= NOTE_ATTRIB;
2430         }
2431         if (vap->va_mode != (mode_t)VNOVAL) {
2432                 mode_t   cur_mode = ip->ino_data.mode;
2433                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2434                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2435
2436                 error = vop_helper_chmod(ap->a_vp, vap->va_mode, ap->a_cred,
2437                                          cur_uid, cur_gid, &cur_mode);
2438                 if (error == 0 && ip->ino_data.mode != cur_mode) {
2439                         ip->ino_data.mode = cur_mode;
2440                         ip->ino_data.ctime = trans.time;
2441                         modflags |= HAMMER_INODE_DDIRTY;
2442                         kflags |= NOTE_ATTRIB;
2443                 }
2444         }
2445 done:
2446         if (error == 0)
2447                 hammer_modify_inode(&trans, ip, modflags);
2448         hammer_done_transaction(&trans);
2449         hammer_knote(ap->a_vp, kflags);
2450         lwkt_reltoken(&hmp->fs_token);
2451         return (error);
2452 }
2453
2454 /*
2455  * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
2456  */
2457 static
2458 int
2459 hammer_vop_nsymlink(struct vop_nsymlink_args *ap)
2460 {
2461         struct hammer_transaction trans;
2462         struct hammer_inode *dip;
2463         struct hammer_inode *nip;
2464         hammer_record_t record;
2465         struct nchandle *nch;
2466         hammer_mount_t hmp;
2467         int error;
2468         int bytes;
2469
2470         ap->a_vap->va_type = VLNK;
2471
2472         nch = ap->a_nch;
2473         dip = VTOI(ap->a_dvp);
2474         hmp = dip->hmp;
2475
2476         if (dip->flags & HAMMER_INODE_RO)
2477                 return (EROFS);
2478         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
2479                 return (error);
2480
2481         /*
2482          * Create a transaction to cover the operations we perform.
2483          */
2484         lwkt_gettoken(&hmp->fs_token);
2485         hammer_start_transaction(&trans, hmp);
2486         ++hammer_stats_file_iopsw;
2487
2488         /*
2489          * Create a new filesystem object of the requested type.  The
2490          * returned inode will be referenced but not locked.
2491          */
2492
2493         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
2494                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
2495                                     NULL, &nip);
2496         if (error) {
2497                 hammer_done_transaction(&trans);
2498                 *ap->a_vpp = NULL;
2499                 lwkt_reltoken(&hmp->fs_token);
2500                 return (error);
2501         }
2502
2503         /*
2504          * Add a record representing the symlink.  symlink stores the link
2505          * as pure data, not a string, and is no \0 terminated.
2506          */
2507         if (error == 0) {
2508                 bytes = strlen(ap->a_target);
2509
2510                 if (bytes <= HAMMER_INODE_BASESYMLEN) {
2511                         bcopy(ap->a_target, nip->ino_data.ext.symlink, bytes);
2512                 } else {
2513                         record = hammer_alloc_mem_record(nip, bytes);
2514                         record->type = HAMMER_MEM_RECORD_GENERAL;
2515
2516                         record->leaf.base.localization = nip->obj_localization +
2517                                                          HAMMER_LOCALIZE_MISC;
2518                         record->leaf.base.key = HAMMER_FIXKEY_SYMLINK;
2519                         record->leaf.base.rec_type = HAMMER_RECTYPE_FIX;
2520                         record->leaf.data_len = bytes;
2521                         KKASSERT(HAMMER_SYMLINK_NAME_OFF == 0);
2522                         bcopy(ap->a_target, record->data->symlink.name, bytes);
2523                         error = hammer_ip_add_record(&trans, record);
2524                 }
2525
2526                 /*
2527                  * Set the file size to the length of the link.
2528                  */
2529                 if (error == 0) {
2530                         nip->ino_data.size = bytes;
2531                         hammer_modify_inode(&trans, nip, HAMMER_INODE_DDIRTY);
2532                 }
2533         }
2534         if (error == 0)
2535                 error = hammer_ip_add_directory(&trans, dip, nch->ncp->nc_name,
2536                                                 nch->ncp->nc_nlen, nip);
2537
2538         /*
2539          * Finish up.
2540          */
2541         if (error) {
2542                 hammer_rel_inode(nip, 0);
2543                 *ap->a_vpp = NULL;
2544         } else {
2545                 error = hammer_get_vnode(nip, ap->a_vpp);
2546                 hammer_rel_inode(nip, 0);
2547                 if (error == 0) {
2548                         cache_setunresolved(ap->a_nch);
2549                         cache_setvp(ap->a_nch, *ap->a_vpp);
2550                         hammer_knote(ap->a_dvp, NOTE_WRITE);
2551                 }
2552         }
2553         hammer_done_transaction(&trans);
2554         lwkt_reltoken(&hmp->fs_token);
2555         return (error);
2556 }
2557
2558 /*
2559  * hammer_vop_nwhiteout { nch, dvp, cred, flags }
2560  */
2561 static
2562 int
2563 hammer_vop_nwhiteout(struct vop_nwhiteout_args *ap)
2564 {
2565         struct hammer_transaction trans;
2566         struct hammer_inode *dip;
2567         hammer_mount_t hmp;
2568         int error;
2569
2570         dip = VTOI(ap->a_dvp);
2571         hmp = dip->hmp;
2572
2573         if (hammer_nohistory(dip) == 0 &&
2574             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0) {
2575                 return (error);
2576         }
2577
2578         lwkt_gettoken(&hmp->fs_token);
2579         hammer_start_transaction(&trans, hmp);
2580         ++hammer_stats_file_iopsw;
2581         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp,
2582                                 ap->a_cred, ap->a_flags, -1);
2583         hammer_done_transaction(&trans);
2584         lwkt_reltoken(&hmp->fs_token);
2585
2586         return (error);
2587 }
2588
2589 /*
2590  * hammer_vop_ioctl { vp, command, data, fflag, cred }
2591  */
2592 static
2593 int
2594 hammer_vop_ioctl(struct vop_ioctl_args *ap)
2595 {
2596         struct hammer_inode *ip = ap->a_vp->v_data;
2597         hammer_mount_t hmp = ip->hmp;
2598         int error;
2599
2600         ++hammer_stats_file_iopsr;
2601         lwkt_gettoken(&hmp->fs_token);
2602         error = hammer_ioctl(ip, ap->a_command, ap->a_data,
2603                              ap->a_fflag, ap->a_cred);
2604         lwkt_reltoken(&hmp->fs_token);
2605         return (error);
2606 }
2607
2608 static
2609 int
2610 hammer_vop_mountctl(struct vop_mountctl_args *ap)
2611 {
2612         static const struct mountctl_opt extraopt[] = {
2613                 { HMNT_NOHISTORY,       "nohistory" },
2614                 { HMNT_MASTERID,        "master" },
2615                 { 0, NULL}
2616
2617         };
2618         struct hammer_mount *hmp;
2619         struct mount *mp;
2620         int usedbytes;
2621         int error;
2622
2623         error = 0;
2624         usedbytes = 0;
2625         mp = ap->a_head.a_ops->head.vv_mount;
2626         KKASSERT(mp->mnt_data != NULL);
2627         hmp = (struct hammer_mount *)mp->mnt_data;
2628
2629         lwkt_gettoken(&hmp->fs_token);
2630
2631         switch(ap->a_op) {
2632         case MOUNTCTL_SET_EXPORT:
2633                 if (ap->a_ctllen != sizeof(struct export_args))
2634                         error = EINVAL;
2635                 else
2636                         error = hammer_vfs_export(mp, ap->a_op,
2637                                       (const struct export_args *)ap->a_ctl);
2638                 break;
2639         case MOUNTCTL_MOUNTFLAGS:
2640         {
2641                 /*
2642                  * Call standard mountctl VOP function
2643                  * so we get user mount flags.
2644                  */
2645                 error = vop_stdmountctl(ap);
2646                 if (error)
2647                         break;
2648
2649                 usedbytes = *ap->a_res;
2650
2651                 if (usedbytes > 0 && usedbytes < ap->a_buflen) {
2652                         usedbytes += vfs_flagstostr(hmp->hflags, extraopt,
2653                                                     ap->a_buf,
2654                                                     ap->a_buflen - usedbytes,
2655                                                     &error);
2656                 }
2657
2658                 *ap->a_res += usedbytes;
2659                 break;
2660         }
2661         default:
2662                 error = vop_stdmountctl(ap);
2663                 break;
2664         }
2665         lwkt_reltoken(&hmp->fs_token);
2666         return(error);
2667 }
2668
2669 /*
2670  * hammer_vop_strategy { vp, bio }
2671  *
2672  * Strategy call, used for regular file read & write only.  Note that the
2673  * bp may represent a cluster.
2674  *
2675  * To simplify operation and allow better optimizations in the future,
2676  * this code does not make any assumptions with regards to buffer alignment
2677  * or size.
2678  */
2679 static
2680 int
2681 hammer_vop_strategy(struct vop_strategy_args *ap)
2682 {
2683         struct buf *bp;
2684         int error;
2685
2686         bp = ap->a_bio->bio_buf;
2687
2688         switch(bp->b_cmd) {
2689         case BUF_CMD_READ:
2690                 error = hammer_vop_strategy_read(ap);
2691                 break;
2692         case BUF_CMD_WRITE:
2693                 error = hammer_vop_strategy_write(ap);
2694                 break;
2695         default:
2696                 bp->b_error = error = EINVAL;
2697                 bp->b_flags |= B_ERROR;
2698                 biodone(ap->a_bio);
2699                 break;
2700         }
2701
2702         /* hammer_dump_dedup_cache(((hammer_inode_t)ap->a_vp->v_data)->hmp); */
2703
2704         return (error);
2705 }
2706
2707 /*
2708  * Read from a regular file.  Iterate the related records and fill in the
2709  * BIO/BUF.  Gaps are zero-filled.
2710  *
2711  * The support code in hammer_object.c should be used to deal with mixed
2712  * in-memory and on-disk records.
2713  *
2714  * NOTE: Can be called from the cluster code with an oversized buf.
2715  *
2716  * XXX atime update
2717  */
2718 static
2719 int
2720 hammer_vop_strategy_read(struct vop_strategy_args *ap)
2721 {
2722         struct hammer_transaction trans;
2723         struct hammer_inode *ip;
2724         struct hammer_inode *dip;
2725         hammer_mount_t hmp;
2726         struct hammer_cursor cursor;
2727         hammer_base_elm_t base;
2728         hammer_off_t disk_offset;
2729         struct bio *bio;
2730         struct bio *nbio;
2731         struct buf *bp;
2732         int64_t rec_offset;
2733         int64_t ran_end;
2734         int64_t tmp64;
2735         int error;
2736         int boff;
2737         int roff;
2738         int n;
2739         int isdedupable;
2740
2741         bio = ap->a_bio;
2742         bp = bio->bio_buf;
2743         ip = ap->a_vp->v_data;
2744         hmp = ip->hmp;
2745
2746         /*
2747          * The zone-2 disk offset may have been set by the cluster code via
2748          * a BMAP operation, or else should be NOOFFSET.
2749          *
2750          * Checking the high bits for a match against zone-2 should suffice.
2751          *
2752          * In cases where a lot of data duplication is present it may be
2753          * more beneficial to drop through and doubule-buffer through the
2754          * device.
2755          */
2756         nbio = push_bio(bio);
2757         if ((nbio->bio_offset & HAMMER_OFF_ZONE_MASK) ==
2758             HAMMER_ZONE_LARGE_DATA) {
2759                 if (hammer_double_buffer == 0) {
2760                         lwkt_gettoken(&hmp->fs_token);
2761                         error = hammer_io_direct_read(hmp, nbio, NULL);
2762                         lwkt_reltoken(&hmp->fs_token);
2763                         return (error);
2764                 }
2765
2766                 /*
2767                  * Try to shortcut requests for double_buffer mode too.
2768                  * Since this mode runs through the device buffer cache
2769                  * only compatible buffer sizes (meaning those generated
2770                  * by normal filesystem buffers) are legal.
2771                  */
2772                 if (hammer_live_dedup == 0 && (bp->b_flags & B_PAGING) == 0) {
2773                         error = hammer_io_indirect_read(hmp, nbio, NULL);
2774                         return (error);
2775                 }
2776         }
2777
2778         /*
2779          * Well, that sucked.  Do it the hard way.  If all the stars are
2780          * aligned we may still be able to issue a direct-read.
2781          */
2782         lwkt_gettoken(&hmp->fs_token);
2783         hammer_simple_transaction(&trans, hmp);
2784         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2785
2786         /*
2787          * Key range (begin and end inclusive) to scan.  Note that the key's
2788          * stored in the actual records represent BASE+LEN, not BASE.  The
2789          * first record containing bio_offset will have a key > bio_offset.
2790          */
2791         cursor.key_beg.localization = ip->obj_localization +
2792                                       HAMMER_LOCALIZE_MISC;
2793         cursor.key_beg.obj_id = ip->obj_id;
2794         cursor.key_beg.create_tid = 0;
2795         cursor.key_beg.delete_tid = 0;
2796         cursor.key_beg.obj_type = 0;
2797         cursor.key_beg.key = bio->bio_offset + 1;
2798         cursor.asof = ip->obj_asof;
2799         cursor.flags |= HAMMER_CURSOR_ASOF;
2800
2801         cursor.key_end = cursor.key_beg;
2802         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2803 #if 0
2804         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
2805                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
2806                 cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
2807                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2808         } else
2809 #endif
2810         {
2811                 ran_end = bio->bio_offset + bp->b_bufsize;
2812                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2813                 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2814                 tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2815                 if (tmp64 < ran_end)
2816                         cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2817                 else
2818                         cursor.key_end.key = ran_end + MAXPHYS + 1;
2819         }
2820         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2821
2822         /*
2823          * Set NOSWAPCACHE for cursor data extraction if double buffering
2824          * is disabled or (if the file is not marked cacheable via chflags
2825          * and vm.swapcache_use_chflags is enabled).
2826          */
2827         if (hammer_double_buffer == 0 ||
2828             ((ap->a_vp->v_flag & VSWAPCACHE) == 0 &&
2829              vm_swapcache_use_chflags)) {
2830                 cursor.flags |= HAMMER_CURSOR_NOSWAPCACHE;
2831         }
2832
2833         error = hammer_ip_first(&cursor);
2834         boff = 0;
2835
2836         while (error == 0) {
2837                 /*
2838                  * Get the base file offset of the record.  The key for
2839                  * data records is (base + bytes) rather then (base).
2840                  */
2841                 base = &cursor.leaf->base;
2842                 rec_offset = base->key - cursor.leaf->data_len;
2843
2844                 /*
2845                  * Calculate the gap, if any, and zero-fill it.
2846                  *
2847                  * n is the offset of the start of the record verses our
2848                  * current seek offset in the bio.
2849                  */
2850                 n = (int)(rec_offset - (bio->bio_offset + boff));
2851                 if (n > 0) {
2852                         if (n > bp->b_bufsize - boff)
2853                                 n = bp->b_bufsize - boff;
2854                         bzero((char *)bp->b_data + boff, n);
2855                         boff += n;
2856                         n = 0;
2857                 }
2858
2859                 /*
2860                  * Calculate the data offset in the record and the number
2861                  * of bytes we can copy.
2862                  *
2863                  * There are two degenerate cases.  First, boff may already
2864                  * be at bp->b_bufsize.  Secondly, the data offset within
2865                  * the record may exceed the record's size.
2866                  */
2867                 roff = -n;
2868                 rec_offset += roff;
2869                 n = cursor.leaf->data_len - roff;
2870                 if (n <= 0) {
2871                         kprintf("strategy_read: bad n=%d roff=%d\n", n, roff);
2872                         n = 0;
2873                 } else if (n > bp->b_bufsize - boff) {
2874                         n = bp->b_bufsize - boff;
2875                 }
2876
2877                 /*
2878                  * Deal with cached truncations.  This cool bit of code
2879                  * allows truncate()/ftruncate() to avoid having to sync
2880                  * the file.
2881                  *
2882                  * If the frontend is truncated then all backend records are
2883                  * subject to the frontend's truncation.
2884                  *
2885                  * If the backend is truncated then backend records on-disk
2886                  * (but not in-memory) are subject to the backend's
2887                  * truncation.  In-memory records owned by the backend
2888                  * represent data written after the truncation point on the
2889                  * backend and must not be truncated.
2890                  *
2891                  * Truncate operations deal with frontend buffer cache
2892                  * buffers and frontend-owned in-memory records synchronously.
2893                  */
2894                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2895                         if (hammer_cursor_ondisk(&cursor)/* ||
2896                             cursor.iprec->flush_state == HAMMER_FST_FLUSH*/) {
2897                                 if (ip->trunc_off <= rec_offset)
2898                                         n = 0;
2899                                 else if (ip->trunc_off < rec_offset + n)
2900                                         n = (int)(ip->trunc_off - rec_offset);
2901                         }
2902                 }
2903                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2904                         if (hammer_cursor_ondisk(&cursor)) {
2905                                 if (ip->sync_trunc_off <= rec_offset)
2906                                         n = 0;
2907                                 else if (ip->sync_trunc_off < rec_offset + n)
2908                                         n = (int)(ip->sync_trunc_off - rec_offset);
2909                         }
2910                 }
2911
2912                 /*
2913                  * Try to issue a direct read into our bio if possible,
2914                  * otherwise resolve the element data into a hammer_buffer
2915                  * and copy.
2916                  *
2917                  * The buffer on-disk should be zerod past any real
2918                  * truncation point, but may not be for any synthesized
2919                  * truncation point from above.
2920                  *
2921                  * NOTE: disk_offset is only valid if the cursor data is
2922                  *       on-disk.
2923                  */
2924                 disk_offset = cursor.leaf->data_offset + roff;
2925                 isdedupable = (boff == 0 && n == bp->b_bufsize &&
2926                                hammer_cursor_ondisk(&cursor) &&
2927                                ((int)disk_offset & HAMMER_BUFMASK) == 0);
2928
2929                 if (isdedupable && hammer_double_buffer == 0) {
2930                         /*
2931                          * Direct read case
2932                          */
2933                         KKASSERT((disk_offset & HAMMER_OFF_ZONE_MASK) ==
2934                                  HAMMER_ZONE_LARGE_DATA);
2935                         nbio->bio_offset = disk_offset;
2936                         error = hammer_io_direct_read(hmp, nbio, cursor.leaf);
2937                         if (hammer_live_dedup && error == 0)
2938                                 hammer_dedup_cache_add(ip, cursor.leaf);
2939                         goto done;
2940                 } else if (isdedupable) {
2941                         /*
2942                          * Async I/O case for reading from backing store
2943                          * and copying the data to the filesystem buffer.
2944                          * live-dedup has to verify the data anyway if it
2945                          * gets a hit later so we can just add the entry
2946                          * now.
2947                          */
2948                         KKASSERT((disk_offset & HAMMER_OFF_ZONE_MASK) ==
2949                                  HAMMER_ZONE_LARGE_DATA);
2950                         nbio->bio_offset = disk_offset;
2951                         if (hammer_live_dedup)
2952                                 hammer_dedup_cache_add(ip, cursor.leaf);
2953                         error = hammer_io_indirect_read(hmp, nbio, cursor.leaf);
2954                         goto done;
2955                 } else if (n) {
2956                         error = hammer_ip_resolve_data(&cursor);
2957                         if (error == 0) {
2958                                 if (hammer_live_dedup && isdedupable)
2959                                         hammer_dedup_cache_add(ip, cursor.leaf);
2960                                 bcopy((char *)cursor.data + roff,
2961                                       (char *)bp->b_data + boff, n);
2962                         }
2963                 }
2964                 if (error)
2965                         break;
2966
2967                 /*
2968                  * We have to be sure that the only elements added to the
2969                  * dedup cache are those which are already on-media.
2970                  */
2971                 if (hammer_live_dedup && hammer_cursor_ondisk(&cursor))
2972                         hammer_dedup_cache_add(ip, cursor.leaf);
2973
2974                 /*
2975                  * Iterate until we have filled the request.
2976                  */
2977                 boff += n;
2978                 if (boff == bp->b_bufsize)
2979                         break;
2980                 error = hammer_ip_next(&cursor);
2981         }
2982
2983         /*
2984          * There may have been a gap after the last record
2985          */
2986         if (error == ENOENT)
2987                 error = 0;
2988         if (error == 0 && boff != bp->b_bufsize) {
2989                 KKASSERT(boff < bp->b_bufsize);
2990                 bzero((char *)bp->b_data + boff, bp->b_bufsize - boff);
2991                 /* boff = bp->b_bufsize; */
2992         }
2993
2994         /*
2995          * Disallow swapcache operation on the vnode buffer if double
2996          * buffering is enabled, the swapcache will get the data via
2997          * the block device buffer.
2998          */
2999         if (hammer_double_buffer)
3000                 bp->b_flags |= B_NOTMETA;
3001
3002         /*
3003          * Cleanup
3004          */
3005         bp->b_resid = 0;
3006         bp->b_error = error;
3007         if (error)
3008                 bp->b_flags |= B_ERROR;
3009         biodone(ap->a_bio);
3010
3011 done:
3012         /*
3013          * Cache the b-tree node for the last data read in cache[1].
3014          *
3015          * If we hit the file EOF then also cache the node in the
3016          * governing director's cache[3], it will be used to initialize
3017          * the inode's cache[1] for any inodes looked up via the directory.
3018          *
3019          * This doesn't reduce disk accesses since the B-Tree chain is
3020          * likely cached, but it does reduce cpu overhead when looking
3021          * up file offsets for cpdup/tar/cpio style iterations.
3022          */
3023         if (cursor.node)
3024                 hammer_cache_node(&ip->cache[1], cursor.node);
3025         if (ran_end >= ip->ino_data.size) {
3026                 dip = hammer_find_inode(&trans, ip->ino_data.parent_obj_id,
3027                                         ip->obj_asof, ip->obj_localization);
3028                 if (dip) {
3029                         hammer_cache_node(&dip->cache[3], cursor.node);
3030                         hammer_rel_inode(dip, 0);
3031                 }
3032         }
3033         hammer_done_cursor(&cursor);
3034         hammer_done_transaction(&trans);
3035         lwkt_reltoken(&hmp->fs_token);
3036         return(error);
3037 }
3038
3039 /*
3040  * BMAP operation - used to support cluster_read() only.
3041  *
3042  * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
3043  *
3044  * This routine may return EOPNOTSUPP if the opration is not supported for
3045  * the specified offset.  The contents of the pointer arguments do not
3046  * need to be initialized in that case.
3047  *
3048  * If a disk address is available and properly aligned return 0 with
3049  * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
3050  * to the run-length relative to that offset.  Callers may assume that
3051  * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
3052  * large, so return EOPNOTSUPP if it is not sufficiently large.
3053  */
3054 static
3055 int
3056 hammer_vop_bmap(struct vop_bmap_args *ap)
3057 {
3058         struct hammer_transaction trans;
3059         struct hammer_inode *ip;
3060         hammer_mount_t hmp;
3061         struct hammer_cursor cursor;
3062         hammer_base_elm_t base;
3063         int64_t rec_offset;
3064         int64_t ran_end;
3065         int64_t tmp64;
3066         int64_t base_offset;
3067         int64_t base_disk_offset;
3068         int64_t last_offset;
3069         hammer_off_t last_disk_offset;
3070         hammer_off_t disk_offset;
3071         int     rec_len;
3072         int     error;
3073         int     blksize;
3074
3075         ++hammer_stats_file_iopsr;
3076         ip = ap->a_vp->v_data;
3077         hmp = ip->hmp;
3078
3079         /*
3080          * We can only BMAP regular files.  We can't BMAP database files,
3081          * directories, etc.
3082          */
3083         if (ip->ino_data.obj_type != HAMMER_OBJTYPE_REGFILE)
3084                 return(EOPNOTSUPP);
3085
3086         /*
3087          * bmap is typically called with runp/runb both NULL when used
3088          * for writing.  We do not support BMAP for writing atm.
3089          */
3090         if (ap->a_cmd != BUF_CMD_READ)
3091                 return(EOPNOTSUPP);
3092
3093         /*
3094          * Scan the B-Tree to acquire blockmap addresses, then translate
3095          * to raw addresses.
3096          */
3097         lwkt_gettoken(&hmp->fs_token);
3098         hammer_simple_transaction(&trans, hmp);
3099 #if 0
3100         kprintf("bmap_beg %016llx ip->cache %p\n",
3101                 (long long)ap->a_loffset, ip->cache[1]);
3102 #endif
3103         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
3104
3105         /*
3106          * Key range (begin and end inclusive) to scan.  Note that the key's
3107          * stored in the actual records represent BASE+LEN, not BASE.  The
3108          * first record containing bio_offset will have a key > bio_offset.
3109          */
3110         cursor.key_beg.localization = ip->obj_localization +
3111                                       HAMMER_LOCALIZE_MISC;
3112         cursor.key_beg.obj_id = ip->obj_id;
3113         cursor.key_beg.create_tid = 0;
3114         cursor.key_beg.delete_tid = 0;
3115         cursor.key_beg.obj_type = 0;
3116         if (ap->a_runb)
3117                 cursor.key_beg.key = ap->a_loffset - MAXPHYS + 1;
3118         else
3119                 cursor.key_beg.key = ap->a_loffset + 1;
3120         if (cursor.key_beg.key < 0)
3121                 cursor.key_beg.key = 0;
3122         cursor.asof = ip->obj_asof;
3123         cursor.flags |= HAMMER_CURSOR_ASOF;
3124
3125         cursor.key_end = cursor.key_beg;
3126         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
3127
3128         ran_end = ap->a_loffset + MAXPHYS;
3129         cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
3130         cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
3131         tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
3132         if (tmp64 < ran_end)
3133                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
3134         else
3135                 cursor.key_end.key = ran_end + MAXPHYS + 1;
3136
3137         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
3138
3139         error = hammer_ip_first(&cursor);
3140         base_offset = last_offset = 0;
3141         base_disk_offset = last_disk_offset = 0;
3142
3143         while (error == 0) {
3144                 /*
3145                  * Get the base file offset of the record.  The key for
3146                  * data records is (base + bytes) rather then (base).
3147                  *
3148                  * NOTE: rec_offset + rec_len may exceed the end-of-file.
3149                  * The extra bytes should be zero on-disk and the BMAP op
3150                  * should still be ok.
3151                  */
3152                 base = &cursor.leaf->base;
3153                 rec_offset = base->key - cursor.leaf->data_len;
3154                 rec_len    = cursor.leaf->data_len;
3155
3156                 /*
3157                  * Incorporate any cached truncation.
3158                  *
3159                  * NOTE: Modifications to rec_len based on synthesized
3160                  * truncation points remove the guarantee that any extended
3161                  * data on disk is zero (since the truncations may not have
3162                  * taken place on-media yet).
3163                  */
3164                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
3165                         if (hammer_cursor_ondisk(&cursor) ||
3166                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
3167                                 if (ip->trunc_off <= rec_offset)
3168                                         rec_len = 0;
3169                                 else if (ip->trunc_off < rec_offset + rec_len)
3170                                         rec_len = (int)(ip->trunc_off - rec_offset);
3171                         }
3172                 }
3173                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
3174                         if (hammer_cursor_ondisk(&cursor)) {
3175                                 if (ip->sync_trunc_off <= rec_offset)
3176                                         rec_len = 0;
3177                                 else if (ip->sync_trunc_off < rec_offset + rec_len)
3178                                         rec_len = (int)(ip->sync_trunc_off - rec_offset);
3179                         }
3180                 }
3181
3182                 /*
3183                  * Accumulate information.  If we have hit a discontiguous
3184                  * block reset base_offset unless we are already beyond the
3185                  * requested offset.  If we are, that's it, we stop.
3186                  */
3187                 if (error)
3188                         break;
3189                 if (hammer_cursor_ondisk(&cursor)) {
3190                         disk_offset = cursor.leaf->data_offset;
3191                         if (rec_offset != last_offset ||
3192                             disk_offset != last_disk_offset) {
3193                                 if (rec_offset > ap->a_loffset)
3194                                         break;
3195                                 base_offset = rec_offset;
3196                                 base_disk_offset = disk_offset;
3197                         }
3198                         last_offset = rec_offset + rec_len;
3199                         last_disk_offset = disk_offset + rec_len;
3200
3201                         if (hammer_live_dedup)
3202                                 hammer_dedup_cache_add(ip, cursor.leaf);
3203                 }
3204
3205                 error = hammer_ip_next(&cursor);
3206         }
3207
3208 #if 0
3209         kprintf("BMAP %016llx:  %016llx - %016llx\n",
3210                 (long long)ap->a_loffset,
3211                 (long long)base_offset,
3212                 (long long)last_offset);
3213         kprintf("BMAP %16s:  %016llx - %016llx\n", "",
3214                 (long long)base_disk_offset,
3215                 (long long)last_disk_offset);
3216 #endif
3217
3218         if (cursor.node) {
3219                 hammer_cache_node(&ip->cache[1], cursor.node);
3220 #if 0
3221                 kprintf("bmap_end2 %016llx ip->cache %p\n",
3222                         (long long)ap->a_loffset, ip->cache[1]);
3223 #endif
3224         }
3225         hammer_done_cursor(&cursor);
3226         hammer_done_transaction(&trans);
3227         lwkt_reltoken(&hmp->fs_token);
3228
3229         /*
3230          * If we couldn't find any records or the records we did find were
3231          * all behind the requested offset, return failure.  A forward
3232          * truncation can leave a hole w/ no on-disk records.
3233          */
3234         if (last_offset == 0 || last_offset < ap->a_loffset)
3235                 return (EOPNOTSUPP);
3236
3237         /*
3238          * Figure out the block size at the requested offset and adjust
3239          * our limits so the cluster_read() does not create inappropriately
3240          * sized buffer cache buffers.
3241          */
3242         blksize = hammer_blocksize(ap->a_loffset);
3243         if (hammer_blocksize(base_offset) != blksize) {
3244                 base_offset = hammer_blockdemarc(base_offset, ap->a_loffset);
3245         }
3246         if (last_offset != ap->a_loffset &&
3247             hammer_blocksize(last_offset - 1) != blksize) {
3248                 last_offset = hammer_blockdemarc(ap->a_loffset,
3249                                                  last_offset - 1);
3250         }
3251
3252         /*
3253          * Returning EOPNOTSUPP simply prevents the direct-IO optimization
3254          * from occuring.
3255          */
3256         disk_offset = base_disk_offset + (ap->a_loffset - base_offset);
3257
3258         if ((disk_offset & HAMMER_OFF_ZONE_MASK) != HAMMER_ZONE_LARGE_DATA) {
3259                 /*
3260                  * Only large-data zones can be direct-IOd
3261                  */
3262                 error = EOPNOTSUPP;
3263         } else if ((disk_offset & HAMMER_BUFMASK) ||
3264                    (last_offset - ap->a_loffset) < blksize) {
3265                 /*
3266                  * doffsetp is not aligned or the forward run size does
3267                  * not cover a whole buffer, disallow the direct I/O.
3268                  */
3269                 error = EOPNOTSUPP;
3270         } else {
3271                 /*
3272                  * We're good.
3273                  */
3274                 *ap->a_doffsetp = disk_offset;
3275                 if (ap->a_runb) {
3276                         *ap->a_runb = ap->a_loffset - base_offset;
3277                         KKASSERT(*ap->a_runb >= 0);
3278                 }
3279                 if (ap->a_runp) {
3280                         *ap->a_runp = last_offset - ap->a_loffset;
3281                         KKASSERT(*ap->a_runp >= 0);
3282                 }
3283                 error = 0;
3284         }
3285         return(error);
3286 }
3287
3288 /*
3289  * Write to a regular file.   Because this is a strategy call the OS is
3290  * trying to actually get data onto the media.
3291  */
3292 static
3293 int
3294 hammer_vop_strategy_write(struct vop_strategy_args *ap)
3295 {
3296         hammer_record_t record;
3297         hammer_mount_t hmp;
3298         hammer_inode_t ip;
3299         struct bio *bio;
3300         struct buf *bp;
3301         int blksize;
3302         int bytes;
3303         int error;
3304
3305         bio = ap->a_bio;
3306         bp = bio->bio_buf;
3307         ip = ap->a_vp->v_data;
3308         hmp = ip->hmp;
3309
3310         blksize = hammer_blocksize(bio->bio_offset);
3311         KKASSERT(bp->b_bufsize == blksize);
3312
3313         if (ip->flags & HAMMER_INODE_RO) {
3314                 bp->b_error = EROFS;
3315                 bp->b_flags |= B_ERROR;
3316                 biodone(ap->a_bio);
3317                 return(EROFS);
3318         }
3319
3320         lwkt_gettoken(&hmp->fs_token);
3321
3322         /*
3323          * Disallow swapcache operation on the vnode buffer if double
3324          * buffering is enabled, the swapcache will get the data via
3325          * the block device buffer.
3326          */
3327         if (hammer_double_buffer)
3328                 bp->b_flags |= B_NOTMETA;
3329
3330         /*
3331          * Interlock with inode destruction (no in-kernel or directory
3332          * topology visibility).  If we queue new IO while trying to
3333          * destroy the inode we can deadlock the vtrunc call in
3334          * hammer_inode_unloadable_check().
3335          *
3336          * Besides, there's no point flushing a bp associated with an
3337          * inode that is being destroyed on-media and has no kernel
3338          * references.
3339          */
3340         if ((ip->flags | ip->sync_flags) &
3341             (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) {
3342                 bp->b_resid = 0;
3343                 biodone(ap->a_bio);
3344                 lwkt_reltoken(&hmp->fs_token);
3345                 return(0);
3346         }
3347
3348         /*
3349          * Reserve space and issue a direct-write from the front-end.
3350          * NOTE: The direct_io code will hammer_bread/bcopy smaller
3351          * allocations.
3352          *
3353          * An in-memory record will be installed to reference the storage
3354          * until the flusher can get to it.
3355          *
3356          * Since we own the high level bio the front-end will not try to
3357          * do a direct-read until the write completes.
3358          *
3359          * NOTE: The only time we do not reserve a full-sized buffers
3360          * worth of data is if the file is small.  We do not try to
3361          * allocate a fragment (from the small-data zone) at the end of
3362          * an otherwise large file as this can lead to wildly separated
3363          * data.
3364          */
3365         KKASSERT((bio->bio_offset & HAMMER_BUFMASK) == 0);
3366         KKASSERT(bio->bio_offset < ip->ino_data.size);
3367         if (bio->bio_offset || ip->ino_data.size > HAMMER_BUFSIZE / 2)
3368                 bytes = bp->b_bufsize;
3369         else
3370                 bytes = ((int)ip->ino_data.size + 15) & ~15;
3371
3372         record = hammer_ip_add_bulk(ip, bio->bio_offset, bp->b_data,
3373                                     bytes, &error);
3374
3375         /*
3376          * B_VFSFLAG1 indicates that a REDO_WRITE entry was generated
3377          * in hammer_vop_write().  We must flag the record so the proper
3378          * REDO_TERM_WRITE entry is generated during the flush.
3379          */
3380         if (record) {
3381                 if (bp->b_flags & B_VFSFLAG1) {
3382                         record->flags |= HAMMER_RECF_REDO;
3383                         bp->b_flags &= ~B_VFSFLAG1;
3384                 }
3385                 if (record->flags & HAMMER_RECF_DEDUPED) {
3386                         bp->b_resid = 0;
3387                         hammer_ip_replace_bulk(hmp, record);
3388                         biodone(ap->a_bio);
3389                 } else {
3390                         hammer_io_direct_write(hmp, bio, record);
3391                 }
3392                 if (ip->rsv_recs > 1 && hmp->rsv_recs > hammer_limit_recs)
3393                         hammer_flush_inode(ip, 0);
3394         } else {
3395                 bp->b_bio2.bio_offset = NOOFFSET;
3396                 bp->b_error = error;
3397                 bp->b_flags |= B_ERROR;
3398                 biodone(ap->a_bio);
3399         }
3400         lwkt_reltoken(&hmp->fs_token);
3401         return(error);
3402 }
3403
3404 /*
3405  * dounlink - disconnect a directory entry
3406  *
3407  * XXX whiteout support not really in yet
3408  */
3409 static int
3410 hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
3411                 struct vnode *dvp, struct ucred *cred,
3412                 int flags, int isdir)
3413 {
3414         struct namecache *ncp;
3415         hammer_inode_t dip;
3416         hammer_inode_t ip;
3417         hammer_mount_t hmp;
3418         struct hammer_cursor cursor;
3419         int64_t namekey;
3420         u_int32_t max_iterations;
3421         int nlen, error;
3422
3423         /*
3424          * Calculate the namekey and setup the key range for the scan.  This
3425          * works kinda like a chained hash table where the lower 32 bits
3426          * of the namekey synthesize the chain.
3427          *
3428          * The key range is inclusive of both key_beg and key_end.
3429          */
3430         dip = VTOI(dvp);
3431         ncp = nch->ncp;
3432         hmp = dip->hmp;
3433
3434         if (dip->flags & HAMMER_INODE_RO)
3435                 return (EROFS);
3436
3437         namekey = hammer_directory_namekey(dip, ncp->nc_name, ncp->nc_nlen,
3438                                            &max_iterations);
3439 retry:
3440         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
3441         cursor.key_beg.localization = dip->obj_localization +
3442                                       hammer_dir_localization(dip);
3443         cursor.key_beg.obj_id = dip->obj_id;
3444         cursor.key_beg.key = namekey;
3445         cursor.key_beg.create_tid = 0;
3446         cursor.key_beg.delete_tid = 0;
3447         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
3448         cursor.key_beg.obj_type = 0;
3449
3450         cursor.key_end = cursor.key_beg;
3451         cursor.key_end.key += max_iterations;
3452         cursor.asof = dip->obj_asof;
3453         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
3454
3455         /*
3456          * Scan all matching records (the chain), locate the one matching
3457          * the requested path component.  info->last_error contains the
3458          * error code on search termination and could be 0, ENOENT, or
3459          * something else.
3460          *
3461          * The hammer_ip_*() functions merge in-memory records with on-disk
3462          * records for the purposes of the search.
3463          */
3464         error = hammer_ip_first(&cursor);
3465
3466         while (error == 0) {
3467                 error = hammer_ip_resolve_data(&cursor);
3468                 if (error)
3469                         break;
3470                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
3471                 KKASSERT(nlen > 0);
3472                 if (ncp->nc_nlen == nlen &&
3473                     bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
3474                         break;
3475                 }
3476                 error = hammer_ip_next(&cursor);
3477         }
3478
3479         /*
3480          * If all is ok we have to get the inode so we can adjust nlinks.
3481          * To avoid a deadlock with the flusher we must release the inode
3482          * lock on the directory when acquiring the inode for the entry.
3483          *
3484          * If the target is a directory, it must be empty.
3485          */
3486         if (error == 0) {
3487                 hammer_unlock(&cursor.ip->lock);
3488                 ip = hammer_get_inode(trans, dip, cursor.data->entry.obj_id,
3489                                       hmp->asof,
3490                                       cursor.data->entry.localization,
3491                                       0, &error);
3492                 hammer_lock_sh(&cursor.ip->lock);
3493                 if (error == ENOENT) {
3494                         kprintf("HAMMER: WARNING: Removing "
3495                                 "dirent w/missing inode \"%s\"\n"
3496                                 "\tobj_id = %016llx\n",
3497                                 ncp->nc_name,
3498                                 (long long)cursor.data->entry.obj_id);
3499                         error = 0;
3500                 }
3501
3502                 /*
3503                  * If isdir >= 0 we validate that the entry is or is not a
3504                  * directory.  If isdir < 0 we don't care.
3505                  */
3506                 if (error == 0 && isdir >= 0 && ip) {
3507                         if (isdir &&
3508                             ip->ino_data.obj_type != HAMMER_OBJTYPE_DIRECTORY) {
3509                                 error = ENOTDIR;
3510                         } else if (isdir == 0 &&
3511                             ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
3512                                 error = EISDIR;
3513                         }
3514                 }
3515
3516                 /*
3517                  * If we are trying to remove a directory the directory must
3518                  * be empty.
3519                  *
3520                  * The check directory code can loop and deadlock/retry.  Our
3521                  * own cursor's node locks must be released to avoid a 3-way
3522                  * deadlock with the flusher if the check directory code
3523                  * blocks.
3524                  *
3525                  * If any changes whatsoever have been made to the cursor
3526                  * set EDEADLK and retry.
3527                  *
3528                  * WARNING: See warnings in hammer_unlock_cursor()
3529                  *          function.
3530                  */
3531                 if (error == 0 && ip && ip->ino_data.obj_type ==
3532                                         HAMMER_OBJTYPE_DIRECTORY) {
3533                         hammer_unlock_cursor(&cursor);
3534                         error = hammer_ip_check_directory_empty(trans, ip);
3535                         hammer_lock_cursor(&cursor);
3536                         if (cursor.flags & HAMMER_CURSOR_RETEST) {
3537                                 kprintf("HAMMER: Warning: avoided deadlock "
3538                                         "on rmdir '%s'\n",
3539                                         ncp->nc_name);
3540                                 error = EDEADLK;
3541                         }
3542                 }
3543
3544                 /*
3545                  * Delete the directory entry.
3546                  *
3547                  * WARNING: hammer_ip_del_directory() may have to terminate
3548                  * the cursor to avoid a deadlock.  It is ok to call
3549                  * hammer_done_cursor() twice.
3550                  */
3551                 if (error == 0) {
3552                         error = hammer_ip_del_directory(trans, &cursor,
3553                                                         dip, ip);
3554                 }
3555                 hammer_done_cursor(&cursor);
3556                 if (error == 0) {
3557                         /*
3558                          * Tell the namecache that we are now unlinked.
3559                          */
3560                         cache_unlink(nch);
3561
3562                         /*
3563                          * NOTE: ip->vp, if non-NULL, cannot be directly
3564                          *       referenced without formally acquiring the
3565                          *       vp since the vp might have zero refs on it,
3566                          *       or in the middle of a reclaim, etc.
3567                          *
3568                          * NOTE: The cache_setunresolved() can rip the vp
3569                          *       out from under us since the vp may not have
3570                          *       any refs, in which case ip->vp will be NULL
3571                          *       from the outset.
3572                          */
3573                         while (ip && ip->vp) {
3574                                 struct vnode *vp;
3575
3576                                 error = hammer_get_vnode(ip, &vp);
3577                                 if (error == 0 && vp) {
3578                                         vn_unlock(vp);
3579                                         hammer_knote(ip->vp, NOTE_DELETE);
3580                                         cache_inval_vp(ip->vp, CINV_DESTROY);
3581                                         vrele(vp);
3582                                         break;
3583                                 }
3584                                 kprintf("Debug: HAMMER ip/vp race1 avoided\n");
3585                         }
3586                 }
3587                 if (ip)
3588                         hammer_rel_inode(ip, 0);
3589         } else {
3590                 hammer_done_cursor(&cursor);
3591         }
3592         if (error == EDEADLK)
3593                 goto retry;
3594
3595         return (error);
3596 }
3597
3598 /************************************************************************
3599  *                          FIFO AND SPECFS OPS                         *
3600  ************************************************************************
3601  *
3602  */
3603 static int
3604 hammer_vop_fifoclose (struct vop_close_args *ap)
3605 {
3606         /* XXX update itimes */
3607         return (VOCALL(&fifo_vnode_vops, &ap->a_head));
3608 }
3609
3610 static int
3611 hammer_vop_fiforead (struct vop_read_args *ap)
3612 {
3613         int error;
3614
3615         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3616         /* XXX update access time */
3617         return (error);
3618 }
3619
3620 static int
3621 hammer_vop_fifowrite (struct vop_write_args *ap)
3622 {
3623         int error;
3624
3625         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3626         /* XXX update access time */
3627         return (error);
3628 }
3629
3630 static
3631 int
3632 hammer_vop_fifokqfilter(struct vop_kqfilter_args *ap)
3633 {
3634         int error;
3635
3636         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3637         if (error)
3638                 error = hammer_vop_kqfilter(ap);
3639         return(error);
3640 }
3641
3642 /************************************************************************
3643  *                          KQFILTER OPS                                *
3644  ************************************************************************
3645  *
3646  */
3647 static void filt_hammerdetach(struct knote *kn);
3648 static int filt_hammerread(struct knote *kn, long hint);
3649 static int filt_hammerwrite(struct knote *kn, long hint);
3650 static int filt_hammervnode(struct knote *kn, long hint);
3651
3652 static struct filterops hammerread_filtops =
3653         { FILTEROP_ISFD, NULL, filt_hammerdetach, filt_hammerread };
3654 static struct filterops hammerwrite_filtops =
3655         { FILTEROP_ISFD, NULL, filt_hammerdetach, filt_hammerwrite };
3656 static struct filterops hammervnode_filtops =
3657         { FILTEROP_ISFD, NULL, filt_hammerdetach, filt_hammervnode };
3658
3659 static
3660 int
3661 hammer_vop_kqfilter(struct vop_kqfilter_args *ap)
3662 {
3663         struct vnode *vp = ap->a_vp;
3664         struct knote *kn = ap->a_kn;
3665
3666         switch (kn->kn_filter) {
3667         case EVFILT_READ:
3668                 kn->kn_fop = &hammerread_filtops;
3669                 break;
3670         case EVFILT_WRITE:
3671                 kn->kn_fop = &hammerwrite_filtops;
3672                 break;
3673         case EVFILT_VNODE:
3674                 kn->kn_fop = &hammervnode_filtops;
3675                 break;
3676         default:
3677                 return (EOPNOTSUPP);
3678         }
3679
3680         kn->kn_hook = (caddr_t)vp;
3681
3682         knote_insert(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3683
3684         return(0);
3685 }
3686
3687 static void
3688 filt_hammerdetach(struct knote *kn)
3689 {
3690         struct vnode *vp = (void *)kn->kn_hook;
3691
3692         knote_remove(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3693 }
3694
3695 static int
3696 filt_hammerread(struct knote *kn, long hint)
3697 {
3698         struct vnode *vp = (void *)kn->kn_hook;
3699         hammer_inode_t ip = VTOI(vp);
3700         hammer_mount_t hmp = ip->hmp;
3701         off_t off;
3702
3703         if (hint == NOTE_REVOKE) {
3704                 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
3705                 return(1);
3706         }
3707         lwkt_gettoken(&hmp->fs_token);  /* XXX use per-ip-token */
3708         off = ip->ino_data.size - kn->kn_fp->f_offset;
3709         kn->kn_data = (off < INTPTR_MAX) ? off : INTPTR_MAX;
3710         lwkt_reltoken(&hmp->fs_token);
3711         if (kn->kn_sfflags & NOTE_OLDAPI)
3712                 return(1);
3713         return (kn->kn_data != 0);
3714 }
3715
3716 static int
3717 filt_hammerwrite(struct knote *kn, long hint)
3718 {
3719         if (hint == NOTE_REVOKE)
3720                 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
3721         kn->kn_data = 0;
3722         return (1);
3723 }
3724
3725 static int
3726 filt_hammervnode(struct knote *kn, long hint)
3727 {
3728         if (kn->kn_sfflags & hint)
3729                 kn->kn_fflags |= hint;
3730         if (hint == NOTE_REVOKE) {
3731                 kn->kn_flags |= (EV_EOF | EV_NODATA);
3732                 return (1);
3733         }
3734         return (kn->kn_fflags != 0);
3735 }
3736