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
  35 #include <sys/mountctl.h>
  36 #include <sys/namecache.h>
  37 #include <sys/buf2.h>
  38 #include <vfs/fifofs/fifo.h>
  39
  40 #include "hammer.h"
  41
  42 /*
  43  * USERFS VNOPS
  44  */
  45 static int hammer_vop_fsync(struct vop_fsync_args *);
  46 static int hammer_vop_read(struct vop_read_args *);
  47 static int hammer_vop_write(struct vop_write_args *);
  48 static int hammer_vop_access(struct vop_access_args *);
  49 static int hammer_vop_advlock(struct vop_advlock_args *);
  50 static int hammer_vop_close(struct vop_close_args *);
  51 static int hammer_vop_ncreate(struct vop_ncreate_args *);
  52 static int hammer_vop_getattr(struct vop_getattr_args *);
  53 static int hammer_vop_nresolve(struct vop_nresolve_args *);
  54 static int hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *);
  55 static int hammer_vop_nlink(struct vop_nlink_args *);
  56 static int hammer_vop_nmkdir(struct vop_nmkdir_args *);
  57 static int hammer_vop_nmknod(struct vop_nmknod_args *);
  58 static int hammer_vop_open(struct vop_open_args *);
  59 static int hammer_vop_print(struct vop_print_args *);
  60 static int hammer_vop_readdir(struct vop_readdir_args *);
  61 static int hammer_vop_readlink(struct vop_readlink_args *);
  62 static int hammer_vop_nremove(struct vop_nremove_args *);
  63 static int hammer_vop_nrename(struct vop_nrename_args *);
  64 static int hammer_vop_nrmdir(struct vop_nrmdir_args *);
  65 static int hammer_vop_markatime(struct vop_markatime_args *);
  66 static int hammer_vop_setattr(struct vop_setattr_args *);
  67 static int hammer_vop_strategy(struct vop_strategy_args *);
  68 static int hammer_vop_bmap(struct vop_bmap_args *ap);
  69 static int hammer_vop_nsymlink(struct vop_nsymlink_args *);
  70 static int hammer_vop_nwhiteout(struct vop_nwhiteout_args *);
  71 static int hammer_vop_ioctl(struct vop_ioctl_args *);
  72 static int hammer_vop_mountctl(struct vop_mountctl_args *);
  73 static int hammer_vop_kqfilter (struct vop_kqfilter_args *);
  74
  75 static int hammer_vop_fifoclose (struct vop_close_args *);
  76 static int hammer_vop_fiforead (struct vop_read_args *);
  77 static int hammer_vop_fifowrite (struct vop_write_args *);
  78 static int hammer_vop_fifokqfilter (struct vop_kqfilter_args *);
  79
  80 struct vop_ops hammer_vnode_vops = {
  81         .vop_default =          vop_defaultop,
  82         .vop_fsync =            hammer_vop_fsync,
  83         .vop_getpages =         vop_stdgetpages,
  84         .vop_putpages =         vop_stdputpages,
  85         .vop_read =             hammer_vop_read,
  86         .vop_write =            hammer_vop_write,
  87         .vop_access =           hammer_vop_access,
  88         .vop_advlock =          hammer_vop_advlock,
  89         .vop_close =            hammer_vop_close,
  90         .vop_ncreate =          hammer_vop_ncreate,
  91         .vop_getattr =          hammer_vop_getattr,
  92         .vop_inactive =         hammer_vop_inactive,
  93         .vop_reclaim =          hammer_vop_reclaim,
  94         .vop_nresolve =         hammer_vop_nresolve,
  95         .vop_nlookupdotdot =    hammer_vop_nlookupdotdot,
  96         .vop_nlink =            hammer_vop_nlink,
  97         .vop_nmkdir =           hammer_vop_nmkdir,
  98         .vop_nmknod =           hammer_vop_nmknod,
  99         .vop_open =             hammer_vop_open,
 100         .vop_pathconf =         vop_stdpathconf,
 101         .vop_print =            hammer_vop_print,
 102         .vop_readdir =          hammer_vop_readdir,
 103         .vop_readlink =         hammer_vop_readlink,
 104         .vop_nremove =          hammer_vop_nremove,
 105         .vop_nrename =          hammer_vop_nrename,
 106         .vop_nrmdir =           hammer_vop_nrmdir,
 107         .vop_markatime =        hammer_vop_markatime,
 108         .vop_setattr =          hammer_vop_setattr,
 109         .vop_bmap =             hammer_vop_bmap,
 110         .vop_strategy =         hammer_vop_strategy,
 111         .vop_nsymlink =         hammer_vop_nsymlink,
 112         .vop_nwhiteout =        hammer_vop_nwhiteout,
 113         .vop_ioctl =            hammer_vop_ioctl,
 114         .vop_mountctl =         hammer_vop_mountctl,
 115         .vop_kqfilter =         hammer_vop_kqfilter
 116 };
 117
 118 struct vop_ops hammer_spec_vops = {
 119         .vop_default =          vop_defaultop,
 120         .vop_fsync =            hammer_vop_fsync,
 121         .vop_read =             vop_stdnoread,
 122         .vop_write =            vop_stdnowrite,
 123         .vop_access =           hammer_vop_access,
 124         .vop_close =            hammer_vop_close,
 125         .vop_markatime =        hammer_vop_markatime,
 126         .vop_getattr =          hammer_vop_getattr,
 127         .vop_inactive =         hammer_vop_inactive,
 128         .vop_reclaim =          hammer_vop_reclaim,
 129         .vop_setattr =          hammer_vop_setattr
 130 };
 131
 132 struct vop_ops hammer_fifo_vops = {
 133         .vop_default =          fifo_vnoperate,
 134         .vop_fsync =            hammer_vop_fsync,
 135         .vop_read =             hammer_vop_fiforead,
 136         .vop_write =            hammer_vop_fifowrite,
 137         .vop_access =           hammer_vop_access,
 138         .vop_close =            hammer_vop_fifoclose,
 139         .vop_markatime =        hammer_vop_markatime,
 140         .vop_getattr =          hammer_vop_getattr,
 141         .vop_inactive =         hammer_vop_inactive,
 142         .vop_reclaim =          hammer_vop_reclaim,
 143         .vop_setattr =          hammer_vop_setattr,
 144         .vop_kqfilter =         hammer_vop_fifokqfilter
 145 };
 146
 147 static __inline
 148 void
 149 hammer_knote(struct vnode *vp, int flags)
 150 {
 151         if (flags)
 152                 KNOTE(&vp->v_pollinfo.vpi_kqinfo.ki_note, flags);
 153 }
 154
 155 #ifdef DEBUG_TRUNCATE
 156 struct hammer_inode *HammerTruncIp;
 157 #endif
 158
 159 static int hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
 160                            struct vnode *dvp, struct ucred *cred,
 161                            int flags, int isdir);
 162 static int hammer_vop_strategy_read(struct vop_strategy_args *ap);
 163 static int hammer_vop_strategy_write(struct vop_strategy_args *ap);
 164
 165 /*
 166  * hammer_vop_fsync { vp, waitfor }
 167  *
 168  * fsync() an inode to disk and wait for it to be completely committed
 169  * such that the information would not be undone if a crash occured after
 170  * return.
 171  *
 172  * NOTE: HAMMER's fsync()'s are going to remain expensive until we implement
 173  *       a REDO log.  A sysctl is provided to relax HAMMER's fsync()
 174  *       operation.
 175  *
 176  *       Ultimately the combination of a REDO log and use of fast storage
 177  *       to front-end cluster caches will make fsync fast, but it aint
 178  *       here yet.  And, in anycase, we need real transactional
 179  *       all-or-nothing features which are not restricted to a single file.
 180  */
 181 static
 182 int
 183 hammer_vop_fsync(struct vop_fsync_args *ap)
 184 {
 185         hammer_inode_t ip = VTOI(ap->a_vp);
 186         hammer_mount_t hmp = ip->hmp;
 187         int waitfor = ap->a_waitfor;
 188         int mode;
 189
 190         lwkt_gettoken(&hmp->fs_token);
 191
 192         /*
 193          * Fsync rule relaxation (default is either full synchronous flush
 194          * or REDO semantics with synchronous flush).
 195          */
 196         if (ap->a_flags & VOP_FSYNC_SYSCALL) {
 197                 switch(hammer_fsync_mode) {
 198                 case 0:
 199 mode0:
 200                         /* no REDO, full synchronous flush */
 201                         goto skip;
 202                 case 1:
 203 mode1:
 204                         /* no REDO, full asynchronous flush */
 205                         if (waitfor == MNT_WAIT)
 206                                 waitfor = MNT_NOWAIT;
 207                         goto skip;
 208                 case 2:
 209                         /* REDO semantics, synchronous flush */
 210                         if (hmp->version < HAMMER_VOL_VERSION_FOUR)
 211                                 goto mode0;
 212                         mode = HAMMER_FLUSH_UNDOS_AUTO;
 213                         break;
 214                 case 3:
 215                         /* REDO semantics, relaxed asynchronous flush */
 216                         if (hmp->version < HAMMER_VOL_VERSION_FOUR)
 217                                 goto mode1;
 218                         mode = HAMMER_FLUSH_UNDOS_RELAXED;
 219                         if (waitfor == MNT_WAIT)
 220                                 waitfor = MNT_NOWAIT;
 221                         break;
 222                 case 4:
 223                         /* ignore the fsync() system call */
 224                         lwkt_reltoken(&hmp->fs_token);
 225                         return(0);
 226                 default:
 227                         /* we have to do something */
 228                         mode = HAMMER_FLUSH_UNDOS_RELAXED;
 229                         if (waitfor == MNT_WAIT)
 230                                 waitfor = MNT_NOWAIT;
 231                         break;
 232                 }
 233
 234                 /*
 235                  * Fast fsync only needs to flush the UNDO/REDO fifo if
 236                  * HAMMER_INODE_REDO is non-zero and the only modifications
 237                  * made to the file are write or write-extends.
 238                  */
 239                 if ((ip->flags & HAMMER_INODE_REDO) &&
 240                     (ip->flags & HAMMER_INODE_MODMASK_NOREDO) == 0) {
 241                         ++hammer_count_fsyncs;
 242                         hammer_flusher_flush_undos(hmp, mode);
 243                         ip->redo_count = 0;
 244                         if (ip->vp && (ip->flags & HAMMER_INODE_MODMASK) == 0)
 245                                 vclrisdirty(ip->vp);
 246                         lwkt_reltoken(&hmp->fs_token);
 247                         return(0);
 248                 }
 249
 250                 /*
 251                  * REDO is enabled by fsync(), the idea being we really only
 252                  * want to lay down REDO records when programs are using
 253                  * fsync() heavily.  The first fsync() on the file starts
 254                  * the gravy train going and later fsync()s keep it hot by
 255                  * resetting the redo_count.
 256                  *
 257                  * We weren't running REDOs before now so we have to fall
 258                  * through and do a full fsync of what we have.
 259                  */
 260                 if (hmp->version >= HAMMER_VOL_VERSION_FOUR &&
 261                     (hmp->flags & HAMMER_MOUNT_REDO_RECOVERY_RUN) == 0) {
 262                         ip->flags |= HAMMER_INODE_REDO;
 263                         ip->redo_count = 0;
 264                 }
 265         }
 266 skip:
 267
 268         /*
 269          * Do a full flush sequence.
 270          *
 271          * Attempt to release the vnode while waiting for the inode to
 272          * finish flushing.  This can really mess up inactive->reclaim
 273          * sequences so only do it if the vnode is active.
 274          *
 275          * WARNING! The VX lock functions must be used.  vn_lock() will
 276          *          fail when this is part of a VOP_RECLAIM sequence.
 277          */
 278         ++hammer_count_fsyncs;
 279         vfsync(ap->a_vp, waitfor, 1, NULL, NULL);
 280         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 281         if (waitfor == MNT_WAIT) {
 282                 int dorelock;
 283
 284                 if ((ap->a_vp->v_flag & VRECLAIMED) == 0) {
 285                         vx_unlock(ap->a_vp);
 286                         dorelock = 1;
 287                 } else {
 288                         dorelock = 0;
 289                 }
 290                 hammer_wait_inode(ip);
 291                 if (dorelock)
 292                         vx_lock(ap->a_vp);
 293         }
 294         if (ip->vp && (ip->flags & HAMMER_INODE_MODMASK) == 0)
 295                 vclrisdirty(ip->vp);
 296         lwkt_reltoken(&hmp->fs_token);
 297         return (ip->error);
 298 }
 299
 300 /*
 301  * hammer_vop_read { vp, uio, ioflag, cred }
 302  *
 303  * MPSAFE (for the cache safe does not require fs_token)
 304  */
 305 static
 306 int
 307 hammer_vop_read(struct vop_read_args *ap)
 308 {
 309         struct hammer_transaction trans;
 310         hammer_inode_t ip;
 311         hammer_mount_t hmp;
 312         off_t offset;
 313         struct buf *bp;
 314         struct uio *uio;
 315         int error;
 316         int n;
 317         int seqcount;
 318         int ioseqcount;
 319         int blksize;
 320         int bigread;
 321         int got_trans;
 322         size_t resid;
 323
 324         if (ap->a_vp->v_type != VREG)
 325                 return (EINVAL);
 326         ip = VTOI(ap->a_vp);
 327         hmp = ip->hmp;
 328         error = 0;
 329         got_trans = 0;
 330         uio = ap->a_uio;
 331
 332         /*
 333          * Attempt to shortcut directly to the VM object using lwbufs.
 334          * This is much faster than instantiating buffer cache buffers.
 335          */
 336         resid = uio->uio_resid;
 337         error = vop_helper_read_shortcut(ap);
 338         hammer_stats_file_read += resid - uio->uio_resid;
 339         if (error)
 340                 return (error);
 341         if (uio->uio_resid == 0)
 342                 goto finished;
 343
 344         /*
 345          * Allow the UIO's size to override the sequential heuristic.
 346          */
 347         blksize = hammer_blocksize(uio->uio_offset);
 348         seqcount = (uio->uio_resid + (BKVASIZE - 1)) / BKVASIZE;
 349         ioseqcount = (ap->a_ioflag >> 16);
 350         if (seqcount < ioseqcount)
 351                 seqcount = ioseqcount;
 352
 353         /*
 354          * If reading or writing a huge amount of data we have to break
 355          * atomicy and allow the operation to be interrupted by a signal
 356          * or it can DOS the machine.
 357          */
 358         bigread = (uio->uio_resid > 100 * 1024 * 1024);
 359
 360         /*
 361          * Access the data typically in HAMMER_BUFSIZE blocks via the
 362          * buffer cache, but HAMMER may use a variable block size based
 363          * on the offset.
 364          *
 365          * XXX Temporary hack, delay the start transaction while we remain
 366          *     MPSAFE.  NOTE: ino_data.size cannot change while vnode is
 367          *     locked-shared.
 368          */
 369         while (uio->uio_resid > 0 && uio->uio_offset < ip->ino_data.size) {
 370                 int64_t base_offset;
 371                 int64_t file_limit;
 372
 373                 blksize = hammer_blocksize(uio->uio_offset);
 374                 offset = (int)uio->uio_offset & (blksize - 1);
 375                 base_offset = uio->uio_offset - offset;
 376
 377                 if (bigread && (error = hammer_signal_check(ip->hmp)) != 0)
 378                         break;
 379
 380                 /*
 381                  * MPSAFE
 382                  */
 383                 bp = getblk(ap->a_vp, base_offset, blksize, 0, 0);
 384                 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) == B_CACHE) {
 385                         bp->b_flags &= ~B_AGE;
 386                         error = 0;
 387                         goto skip;
 388                 }
 389                 if (ap->a_ioflag & IO_NRDELAY) {
 390                         bqrelse(bp);
 391                         return (EWOULDBLOCK);
 392                 }
 393
 394                 /*
 395                  * MPUNSAFE
 396                  */
 397                 if (got_trans == 0) {
 398                         hammer_start_transaction(&trans, ip->hmp);
 399                         got_trans = 1;
 400                 }
 401
 402                 /*
 403                  * NOTE: A valid bp has already been acquired, but was not
 404                  *       B_CACHE.
 405                  */
 406                 if (hammer_cluster_enable) {
 407                         /*
 408                          * Use file_limit to prevent cluster_read() from
 409                          * creating buffers of the wrong block size past
 410                          * the demarc.
 411                          */
 412                         file_limit = ip->ino_data.size;
 413                         if (base_offset < HAMMER_XDEMARC &&
 414                             file_limit > HAMMER_XDEMARC) {
 415                                 file_limit = HAMMER_XDEMARC;
 416                         }
 417                         error = cluster_readx(ap->a_vp,
 418                                              file_limit, base_offset,
 419                                              blksize, uio->uio_resid,
 420                                              seqcount * BKVASIZE, &bp);
 421                 } else {
 422                         error = breadnx(ap->a_vp, base_offset, blksize,
 423                                         NULL, NULL, 0, &bp);
 424                 }
 425                 if (error) {
 426                         brelse(bp);
 427                         break;
 428                 }
 429 skip:
 430                 if ((hammer_debug_io & 0x0001) && (bp->b_flags & B_IODEBUG)) {
 431                         hdkprintf("doff %016jx read file %016jx@%016jx\n",
 432                                 (intmax_t)bp->b_bio2.bio_offset,
 433                                 (intmax_t)ip->obj_id,
 434                                 (intmax_t)bp->b_loffset);
 435                 }
 436                 bp->b_flags &= ~B_IODEBUG;
 437                 if (blksize == HAMMER_XBUFSIZE)
 438                         bp->b_flags |= B_CLUSTEROK;
 439
 440                 n = blksize - offset;
 441                 if (n > uio->uio_resid)
 442                         n = uio->uio_resid;
 443                 if (n > ip->ino_data.size - uio->uio_offset)
 444                         n = (int)(ip->ino_data.size - uio->uio_offset);
 445
 446                 /*
 447                  * Set B_AGE, data has a lower priority than meta-data.
 448                  *
 449                  * Use a hold/unlock/drop sequence to run the uiomove
 450                  * with the buffer unlocked, avoiding deadlocks against
 451                  * read()s on mmap()'d spaces.
 452                  */
 453                 bp->b_flags |= B_AGE;
 454                 error = uiomovebp(bp, (char *)bp->b_data + offset, n, uio);
 455                 bqrelse(bp);
 456
 457                 if (error)
 458                         break;
 459                 hammer_stats_file_read += n;
 460         }
 461
 462 finished:
 463
 464         /*
 465          * Try to update the atime with just the inode lock for maximum
 466          * concurrency.  If we can't shortcut it we have to get the full
 467          * blown transaction.
 468          */
 469         if (got_trans == 0 && hammer_update_atime_quick(ip) < 0) {
 470                 hammer_start_transaction(&trans, ip->hmp);
 471                 got_trans = 1;
 472         }
 473
 474         if (got_trans) {
 475                 if ((ip->flags & HAMMER_INODE_RO) == 0 &&
 476                     (ip->hmp->mp->mnt_flag & MNT_NOATIME) == 0) {
 477                         lwkt_gettoken(&hmp->fs_token);
 478                         ip->ino_data.atime = trans.time;
 479                         hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
 480                         hammer_done_transaction(&trans);
 481                         lwkt_reltoken(&hmp->fs_token);
 482                 } else {
 483                         hammer_done_transaction(&trans);
 484                 }
 485         }
 486         return (error);
 487 }
 488
 489 /*
 490  * hammer_vop_write { vp, uio, ioflag, cred }
 491  */
 492 static
 493 int
 494 hammer_vop_write(struct vop_write_args *ap)
 495 {
 496         struct hammer_transaction trans;
 497         struct hammer_inode *ip;
 498         hammer_mount_t hmp;
 499         thread_t td;
 500         struct uio *uio;
 501         int offset;
 502         off_t base_offset;
 503         int64_t cluster_eof;
 504         struct buf *bp;
 505         int kflags;
 506         int error;
 507         int n;
 508         int flags;
 509         int seqcount;
 510         int bigwrite;
 511
 512         if (ap->a_vp->v_type != VREG)
 513                 return (EINVAL);
 514         ip = VTOI(ap->a_vp);
 515         hmp = ip->hmp;
 516         error = 0;
 517         kflags = 0;
 518         seqcount = ap->a_ioflag >> 16;
 519
 520         if (ip->flags & HAMMER_INODE_RO)
 521                 return (EROFS);
 522
 523         /*
 524          * Create a transaction to cover the operations we perform.
 525          */
 526         hammer_start_transaction(&trans, hmp);
 527         uio = ap->a_uio;
 528
 529         /*
 530          * Check append mode
 531          */
 532         if (ap->a_ioflag & IO_APPEND)
 533                 uio->uio_offset = ip->ino_data.size;
 534
 535         /*
 536          * Check for illegal write offsets.  Valid range is 0...2^63-1.
 537          *
 538          * NOTE: the base_off assignment is required to work around what
 539          * I consider to be a GCC-4 optimization bug.
 540          */
 541         if (uio->uio_offset < 0) {
 542                 hammer_done_transaction(&trans);
 543                 return (EFBIG);
 544         }
 545         base_offset = uio->uio_offset + uio->uio_resid; /* work around gcc-4 */
 546         if (uio->uio_resid > 0 && base_offset <= uio->uio_offset) {
 547                 hammer_done_transaction(&trans);
 548                 return (EFBIG);
 549         }
 550
 551         if (uio->uio_resid > 0 && (td = uio->uio_td) != NULL && td->td_proc &&
 552             base_offset > td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
 553                 hammer_done_transaction(&trans);
 554                 lwpsignal(td->td_proc, td->td_lwp, SIGXFSZ);
 555                 return (EFBIG);
 556         }
 557
 558         /*
 559          * If reading or writing a huge amount of data we have to break
 560          * atomicy and allow the operation to be interrupted by a signal
 561          * or it can DOS the machine.
 562          *
 563          * Preset redo_count so we stop generating REDOs earlier if the
 564          * limit is exceeded.
 565          *
 566          * redo_count is heuristical, SMP races are ok
 567          */
 568         bigwrite = (uio->uio_resid > 100 * 1024 * 1024);
 569         if ((ip->flags & HAMMER_INODE_REDO) &&
 570             ip->redo_count < hammer_limit_redo) {
 571                 ip->redo_count += uio->uio_resid;
 572         }
 573
 574         /*
 575          * Access the data typically in HAMMER_BUFSIZE blocks via the
 576          * buffer cache, but HAMMER may use a variable block size based
 577          * on the offset.
 578          */
 579         while (uio->uio_resid > 0) {
 580                 int fixsize = 0;
 581                 int blksize;
 582                 int blkmask;
 583                 int trivial;
 584                 int endofblk;
 585                 off_t nsize;
 586
 587                 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_WRITE)) != 0)
 588                         break;
 589                 if (bigwrite && (error = hammer_signal_check(hmp)) != 0)
 590                         break;
 591
 592                 blksize = hammer_blocksize(uio->uio_offset);
 593
 594                 /*
 595                  * Control the number of pending records associated with
 596                  * this inode.  If too many have accumulated start a
 597                  * flush.  Try to maintain a pipeline with the flusher.
 598                  *
 599                  * NOTE: It is possible for other sources to grow the
 600                  *       records but not necessarily issue another flush,
 601                  *       so use a timeout and ensure that a re-flush occurs.
 602                  */
 603                 if (ip->rsv_recs >= hammer_limit_inode_recs) {
 604                         lwkt_gettoken(&hmp->fs_token);
 605                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 606                         while (ip->rsv_recs >= hammer_limit_inode_recs * 2) {
 607                                 ip->flags |= HAMMER_INODE_RECSW;
 608                                 tsleep(&ip->rsv_recs, 0, "hmrwww", hz);
 609                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 610                         }
 611                         lwkt_reltoken(&hmp->fs_token);
 612                 }
 613
 614                 /*
 615                  * Do not allow HAMMER to blow out the buffer cache.  Very
 616                  * large UIOs can lockout other processes due to bwillwrite()
 617                  * mechanics.
 618                  *
 619                  * The hammer inode is not locked during these operations.
 620                  * The vnode is locked which can interfere with the pageout
 621                  * daemon for non-UIO_NOCOPY writes but should not interfere
 622                  * with the buffer cache.  Even so, we cannot afford to
 623                  * allow the pageout daemon to build up too many dirty buffer
 624                  * cache buffers.
 625                  *
 626                  * Only call this if we aren't being recursively called from
 627                  * a virtual disk device (vn), else we may deadlock.
 628                  */
 629                 if ((ap->a_ioflag & IO_RECURSE) == 0)
 630                         bwillwrite(blksize);
 631
 632                 /*
 633                  * Calculate the blocksize at the current offset and figure
 634                  * out how much we can actually write.
 635                  */
 636                 blkmask = blksize - 1;
 637                 offset = (int)uio->uio_offset & blkmask;
 638                 base_offset = uio->uio_offset & ~(int64_t)blkmask;
 639                 n = blksize - offset;
 640                 if (n > uio->uio_resid) {
 641                         n = uio->uio_resid;
 642                         endofblk = 0;
 643                 } else {
 644                         endofblk = 1;
 645                 }
 646                 nsize = uio->uio_offset + n;
 647                 if (nsize > ip->ino_data.size) {
 648                         if (uio->uio_offset > ip->ino_data.size)
 649                                 trivial = 0;
 650                         else
 651                                 trivial = 1;
 652                         nvextendbuf(ap->a_vp,
 653                                     ip->ino_data.size,
 654                                     nsize,
 655                                     hammer_blocksize(ip->ino_data.size),
 656                                     hammer_blocksize(nsize),
 657                                     hammer_blockoff(ip->ino_data.size),
 658                                     hammer_blockoff(nsize),
 659                                     trivial);
 660                         fixsize = 1;
 661                         kflags |= NOTE_EXTEND;
 662                 }
 663
 664                 if (uio->uio_segflg == UIO_NOCOPY) {
 665                         /*
 666                          * Issuing a write with the same data backing the
 667                          * buffer.  Instantiate the buffer to collect the
 668                          * backing vm pages, then read-in any missing bits.
 669                          *
 670                          * This case is used by vop_stdputpages().
 671                          */
 672                         bp = getblk(ap->a_vp, base_offset,
 673                                     blksize, GETBLK_BHEAVY, 0);
 674                         if ((bp->b_flags & B_CACHE) == 0) {
 675                                 bqrelse(bp);
 676                                 error = bread(ap->a_vp, base_offset,
 677                                               blksize, &bp);
 678                         }
 679                 } else if (offset == 0 && uio->uio_resid >= blksize) {
 680                         /*
 681                          * Even though we are entirely overwriting the buffer
 682                          * we may still have to zero it out to avoid a
 683                          * mmap/write visibility issue.
 684                          */
 685                         bp = getblk(ap->a_vp, base_offset, blksize, GETBLK_BHEAVY, 0);
 686                         if ((bp->b_flags & B_CACHE) == 0)
 687                                 vfs_bio_clrbuf(bp);
 688                 } else if (base_offset >= ip->ino_data.size) {
 689                         /*
 690                          * If the base offset of the buffer is beyond the
 691                          * file EOF, we don't have to issue a read.
 692                          */
 693                         bp = getblk(ap->a_vp, base_offset,
 694                                     blksize, GETBLK_BHEAVY, 0);
 695                         vfs_bio_clrbuf(bp);
 696                 } else {
 697                         /*
 698                          * Partial overwrite, read in any missing bits then
 699                          * replace the portion being written.
 700                          */
 701                         error = bread(ap->a_vp, base_offset, blksize, &bp);
 702                         if (error == 0)
 703                                 bheavy(bp);
 704                 }
 705                 if (error == 0)
 706                         error = uiomovebp(bp, bp->b_data + offset, n, uio);
 707
 708                 lwkt_gettoken(&hmp->fs_token);
 709
 710                 /*
 711                  * Generate REDO records if enabled and redo_count will not
 712                  * exceeded the limit.
 713                  *
 714                  * If redo_count exceeds the limit we stop generating records
 715                  * and clear HAMMER_INODE_REDO.  This will cause the next
 716                  * fsync() to do a full meta-data sync instead of just an
 717                  * UNDO/REDO fifo update.
 718                  *
 719                  * When clearing HAMMER_INODE_REDO any pre-existing REDOs
 720                  * will still be tracked.  The tracks will be terminated
 721                  * when the related meta-data (including possible data
 722                  * modifications which are not tracked via REDO) is
 723                  * flushed.
 724                  */
 725                 if ((ip->flags & HAMMER_INODE_REDO) && error == 0) {
 726                         if (ip->redo_count < hammer_limit_redo) {
 727                                 bp->b_flags |= B_VFSFLAG1;
 728                                 error = hammer_generate_redo(&trans, ip,
 729                                                      base_offset + offset,
 730                                                      HAMMER_REDO_WRITE,
 731                                                      bp->b_data + offset,
 732                                                      (size_t)n);
 733                         } else {
 734                                 ip->flags &= ~HAMMER_INODE_REDO;
 735                         }
 736                 }
 737
 738                 /*
 739                  * If we screwed up we have to undo any VM size changes we
 740                  * made.
 741                  */
 742                 if (error) {
 743                         brelse(bp);
 744                         if (fixsize) {
 745                                 nvtruncbuf(ap->a_vp, ip->ino_data.size,
 746                                           hammer_blocksize(ip->ino_data.size),
 747                                           hammer_blockoff(ip->ino_data.size),
 748                                           0);
 749                         }
 750                         lwkt_reltoken(&hmp->fs_token);
 751                         break;
 752                 }
 753                 kflags |= NOTE_WRITE;
 754                 hammer_stats_file_write += n;
 755                 if (blksize == HAMMER_XBUFSIZE)
 756                         bp->b_flags |= B_CLUSTEROK;
 757                 if (ip->ino_data.size < uio->uio_offset) {
 758                         ip->ino_data.size = uio->uio_offset;
 759                         flags = HAMMER_INODE_SDIRTY;
 760                 } else {
 761                         flags = 0;
 762                 }
 763                 ip->ino_data.mtime = trans.time;
 764                 flags |= HAMMER_INODE_MTIME | HAMMER_INODE_BUFS;
 765                 hammer_modify_inode(&trans, ip, flags);
 766
 767                 /*
 768                  * Once we dirty the buffer any cached zone-X offset
 769                  * becomes invalid.  HAMMER NOTE: no-history mode cannot
 770                  * allow overwriting over the same data sector unless
 771                  * we provide UNDOs for the old data, which we don't.
 772                  */
 773                 bp->b_bio2.bio_offset = NOOFFSET;
 774
 775                 lwkt_reltoken(&hmp->fs_token);
 776
 777                 /*
 778                  * Final buffer disposition.
 779                  *
 780                  * Because meta-data updates are deferred, HAMMER is
 781                  * especially sensitive to excessive bdwrite()s because
 782                  * the I/O stream is not broken up by disk reads.  So the
 783                  * buffer cache simply cannot keep up.
 784                  *
 785                  * WARNING!  blksize is variable.  cluster_write() is
 786                  *           expected to not blow up if it encounters
 787                  *           buffers that do not match the passed blksize.
 788                  *
 789                  * NOTE!  Hammer shouldn't need to bawrite()/cluster_write().
 790                  *        The ip->rsv_recs check should burst-flush the data.
 791                  *        If we queue it immediately the buf could be left
 792                  *        locked on the device queue for a very long time.
 793                  *
 794                  *        However, failing to flush a dirty buffer out when
 795                  *        issued from the pageout daemon can result in a low
 796                  *        memory deadlock against bio_page_alloc(), so we
 797                  *        have to bawrite() on IO_ASYNC as well.
 798                  *
 799                  * NOTE!  To avoid degenerate stalls due to mismatched block
 800                  *        sizes we only honor IO_DIRECT on the write which
 801                  *        abuts the end of the buffer.  However, we must
 802                  *        honor IO_SYNC in case someone is silly enough to
 803                  *        configure a HAMMER file as swap, or when HAMMER
 804                  *        is serving NFS (for commits).  Ick ick.
 805                  */
 806                 bp->b_flags |= B_AGE;
 807                 if (blksize == HAMMER_XBUFSIZE)
 808                         bp->b_flags |= B_CLUSTEROK;
 809
 810                 if (ap->a_ioflag & IO_SYNC) {
 811                         bwrite(bp);
 812                 } else if ((ap->a_ioflag & IO_DIRECT) && endofblk) {
 813                         bawrite(bp);
 814                 } else if (ap->a_ioflag & IO_ASYNC) {
 815                         bawrite(bp);
 816                 } else if (hammer_cluster_enable &&
 817                            !(ap->a_vp->v_mount->mnt_flag & MNT_NOCLUSTERW)) {
 818                         if (base_offset < HAMMER_XDEMARC)
 819                                 cluster_eof = hammer_blockdemarc(base_offset,
 820                                                          ip->ino_data.size);
 821                         else
 822                                 cluster_eof = ip->ino_data.size;
 823                         cluster_write(bp, cluster_eof, blksize, seqcount);
 824                 } else {
 825                         bdwrite(bp);
 826                 }
 827         }
 828         hammer_done_transaction(&trans);
 829         hammer_knote(ap->a_vp, kflags);
 830
 831         return (error);
 832 }
 833
 834 /*
 835  * hammer_vop_access { vp, mode, cred }
 836  *
 837  * MPSAFE - does not require fs_token
 838  */
 839 static
 840 int
 841 hammer_vop_access(struct vop_access_args *ap)
 842 {
 843         struct hammer_inode *ip = VTOI(ap->a_vp);
 844         uid_t uid;
 845         gid_t gid;
 846         int error;
 847
 848         ++hammer_stats_file_iopsr;
 849         uid = hammer_to_unix_xid(&ip->ino_data.uid);
 850         gid = hammer_to_unix_xid(&ip->ino_data.gid);
 851
 852         error = vop_helper_access(ap, uid, gid, ip->ino_data.mode,
 853                                   ip->ino_data.uflags);
 854         return (error);
 855 }
 856
 857 /*
 858  * hammer_vop_advlock { vp, id, op, fl, flags }
 859  *
 860  * MPSAFE - does not require fs_token
 861  */
 862 static
 863 int
 864 hammer_vop_advlock(struct vop_advlock_args *ap)
 865 {
 866         hammer_inode_t ip = VTOI(ap->a_vp);
 867
 868         return (lf_advlock(ap, &ip->advlock, ip->ino_data.size));
 869 }
 870
 871 /*
 872  * hammer_vop_close { vp, fflag }
 873  *
 874  * We can only sync-on-close for normal closes.  XXX disabled for now.
 875  */
 876 static
 877 int
 878 hammer_vop_close(struct vop_close_args *ap)
 879 {
 880 #if 0
 881         struct vnode *vp = ap->a_vp;
 882         hammer_inode_t ip = VTOI(vp);
 883         int waitfor;
 884         if (ip->flags & (HAMMER_INODE_CLOSESYNC|HAMMER_INODE_CLOSEASYNC)) {
 885                 if (vn_islocked(vp) == LK_EXCLUSIVE &&
 886                     (vp->v_flag & (VINACTIVE|VRECLAIMED)) == 0) {
 887                         if (ip->flags & HAMMER_INODE_CLOSESYNC)
 888                                 waitfor = MNT_WAIT;
 889                         else
 890                                 waitfor = MNT_NOWAIT;
 891                         ip->flags &= ~(HAMMER_INODE_CLOSESYNC |
 892                                        HAMMER_INODE_CLOSEASYNC);
 893                         VOP_FSYNC(vp, MNT_NOWAIT, waitfor);
 894                 }
 895         }
 896 #endif
 897         return (vop_stdclose(ap));
 898 }
 899
 900 /*
 901  * hammer_vop_ncreate { nch, dvp, vpp, cred, vap }
 902  *
 903  * The operating system has already ensured that the directory entry
 904  * does not exist and done all appropriate namespace locking.
 905  */
 906 static
 907 int
 908 hammer_vop_ncreate(struct vop_ncreate_args *ap)
 909 {
 910         struct hammer_transaction trans;
 911         struct hammer_inode *dip;
 912         struct hammer_inode *nip;
 913         struct nchandle *nch;
 914         hammer_mount_t hmp;
 915         int error;
 916
 917         nch = ap->a_nch;
 918         dip = VTOI(ap->a_dvp);
 919         hmp = dip->hmp;
 920
 921         if (dip->flags & HAMMER_INODE_RO)
 922                 return (EROFS);
 923         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
 924                 return (error);
 925
 926         /*
 927          * Create a transaction to cover the operations we perform.
 928          */
 929         lwkt_gettoken(&hmp->fs_token);
 930         hammer_start_transaction(&trans, hmp);
 931         ++hammer_stats_file_iopsw;
 932
 933         /*
 934          * Create a new filesystem object of the requested type.  The
 935          * returned inode will be referenced and shared-locked to prevent
 936          * it from being moved to the flusher.
 937          */
 938         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
 939                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
 940                                     NULL, &nip);
 941         if (error) {
 942                 hkprintf("hammer_create_inode error %d\n", error);
 943                 hammer_done_transaction(&trans);
 944                 *ap->a_vpp = NULL;
 945                 lwkt_reltoken(&hmp->fs_token);
 946                 return (error);
 947         }
 948
 949         /*
 950          * Add the new filesystem object to the directory.  This will also
 951          * bump the inode's link count.
 952          */
 953         error = hammer_ip_add_directory(&trans, dip,
 954                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
 955                                         nip);
 956         if (error)
 957                 hkprintf("hammer_ip_add_directory error %d\n", error);
 958
 959         /*
 960          * Finish up.
 961          */
 962         if (error) {
 963                 hammer_rel_inode(nip, 0);
 964                 hammer_done_transaction(&trans);
 965                 *ap->a_vpp = NULL;
 966         } else {
 967                 error = hammer_get_vnode(nip, ap->a_vpp);
 968                 hammer_done_transaction(&trans);
 969                 hammer_rel_inode(nip, 0);
 970                 if (error == 0) {
 971                         cache_setunresolved(ap->a_nch);
 972                         cache_setvp(ap->a_nch, *ap->a_vpp);
 973                 }
 974                 hammer_knote(ap->a_dvp, NOTE_WRITE);
 975         }
 976         lwkt_reltoken(&hmp->fs_token);
 977         return (error);
 978 }
 979
 980 /*
 981  * hammer_vop_getattr { vp, vap }
 982  *
 983  * Retrieve an inode's attribute information.  When accessing inodes
 984  * historically we fake the atime field to ensure consistent results.
 985  * The atime field is stored in the B-Tree element and allowed to be
 986  * updated without cycling the element.
 987  *
 988  * MPSAFE - does not require fs_token
 989  */
 990 static
 991 int
 992 hammer_vop_getattr(struct vop_getattr_args *ap)
 993 {
 994         struct hammer_inode *ip = VTOI(ap->a_vp);
 995         struct vattr *vap = ap->a_vap;
 996
 997         /*
 998          * We want the fsid to be different when accessing a filesystem
 999          * with different as-of's so programs like diff don't think
1000          * the files are the same.
1001          *
1002          * We also want the fsid to be the same when comparing snapshots,
1003          * or when comparing mirrors (which might be backed by different
1004          * physical devices).  HAMMER fsids are based on the PFS's
1005          * shared_uuid field.
1006          *
1007          * XXX there is a chance of collision here.  The va_fsid reported
1008          * by stat is different from the more involved fsid used in the
1009          * mount structure.
1010          */
1011         ++hammer_stats_file_iopsr;
1012         hammer_lock_sh(&ip->lock);
1013         vap->va_fsid = ip->pfsm->fsid_udev ^ (u_int32_t)ip->obj_asof ^
1014                        (u_int32_t)(ip->obj_asof >> 32);
1015
1016         vap->va_fileid = ip->ino_leaf.base.obj_id;
1017         vap->va_mode = ip->ino_data.mode;
1018         vap->va_nlink = ip->ino_data.nlinks;
1019         vap->va_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1020         vap->va_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1021         vap->va_rmajor = 0;
1022         vap->va_rminor = 0;
1023         vap->va_size = ip->ino_data.size;
1024
1025         /*
1026          * Special case for @@PFS softlinks.  The actual size of the
1027          * expanded softlink is "@@0x%016llx:%05d" == 26 bytes.
1028          * or for MAX_TID is    "@@-1:%05d" == 10 bytes.
1029          *
1030          * Note that userspace hammer command does not allow users to
1031          * create a @@PFS softlink under an existing other PFS (id!=0)
1032          * so the ip localization here for @@PFS softlink is always 0.
1033          */
1034         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_SOFTLINK &&
1035             ip->ino_data.size == 10 &&
1036             ip->obj_asof == HAMMER_MAX_TID &&
1037             ip->obj_localization == 0 &&
1038             strncmp(ip->ino_data.ext.symlink, "@@PFS", 5) == 0) {
1039                     if (ip->pfsm->pfsd.mirror_flags & HAMMER_PFSD_SLAVE)
1040                             vap->va_size = 26;
1041                     else
1042                             vap->va_size = 10;
1043         }
1044
1045         /*
1046          * We must provide a consistent atime and mtime for snapshots
1047          * so people can do a 'tar cf - ... | md5' on them and get
1048          * consistent results.
1049          */
1050         if (ip->flags & HAMMER_INODE_RO) {
1051                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_atime);
1052                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_mtime);
1053         } else {
1054                 hammer_time_to_timespec(ip->ino_data.atime, &vap->va_atime);
1055                 hammer_time_to_timespec(ip->ino_data.mtime, &vap->va_mtime);
1056         }
1057         hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_ctime);
1058         vap->va_flags = ip->ino_data.uflags;
1059         vap->va_gen = 1;        /* hammer inums are unique for all time */
1060         vap->va_blocksize = HAMMER_BUFSIZE;
1061         if (ip->ino_data.size >= HAMMER_XDEMARC) {
1062                 vap->va_bytes = (ip->ino_data.size + HAMMER_XBUFMASK64) &
1063                                 ~HAMMER_XBUFMASK64;
1064         } else if (ip->ino_data.size > HAMMER_HBUFSIZE) {
1065                 vap->va_bytes = (ip->ino_data.size + HAMMER_BUFMASK64) &
1066                                 ~HAMMER_BUFMASK64;
1067         } else {
1068                 vap->va_bytes = (ip->ino_data.size + 15) & ~15;
1069         }
1070
1071         vap->va_type = hammer_get_vnode_type(ip->ino_data.obj_type);
1072         vap->va_filerev = 0;    /* XXX */
1073         vap->va_uid_uuid = ip->ino_data.uid;
1074         vap->va_gid_uuid = ip->ino_data.gid;
1075         vap->va_fsid_uuid = ip->hmp->fsid;
1076         vap->va_vaflags = VA_UID_UUID_VALID | VA_GID_UUID_VALID |
1077                           VA_FSID_UUID_VALID;
1078
1079         switch (ip->ino_data.obj_type) {
1080         case HAMMER_OBJTYPE_CDEV:
1081         case HAMMER_OBJTYPE_BDEV:
1082                 vap->va_rmajor = ip->ino_data.rmajor;
1083                 vap->va_rminor = ip->ino_data.rminor;
1084                 break;
1085         default:
1086                 break;
1087         }
1088         hammer_unlock(&ip->lock);
1089         return(0);
1090 }
1091
1092 /*
1093  * hammer_vop_nresolve { nch, dvp, cred }
1094  *
1095  * Locate the requested directory entry.
1096  */
1097 static
1098 int
1099 hammer_vop_nresolve(struct vop_nresolve_args *ap)
1100 {
1101         struct hammer_transaction trans;
1102         struct namecache *ncp;
1103         hammer_mount_t hmp;
1104         hammer_inode_t dip;
1105         hammer_inode_t ip;
1106         hammer_tid_t asof;
1107         struct hammer_cursor cursor;
1108         struct vnode *vp;
1109         int64_t namekey;
1110         int error;
1111         int i;
1112         int nlen;
1113         int flags;
1114         int ispfs;
1115         int64_t obj_id;
1116         u_int32_t localization;
1117         u_int32_t max_iterations;
1118
1119         /*
1120          * Misc initialization, plus handle as-of name extensions.  Look for
1121          * the '@@' extension.  Note that as-of files and directories cannot
1122          * be modified.
1123          */
1124         dip = VTOI(ap->a_dvp);
1125         ncp = ap->a_nch->ncp;
1126         asof = dip->obj_asof;
1127         localization = dip->obj_localization;   /* for code consistency */
1128         nlen = ncp->nc_nlen;
1129         flags = dip->flags & HAMMER_INODE_RO;
1130         ispfs = 0;
1131         hmp = dip->hmp;
1132
1133         lwkt_gettoken(&hmp->fs_token);
1134         hammer_simple_transaction(&trans, hmp);
1135         ++hammer_stats_file_iopsr;
1136
1137         for (i = 0; i < nlen; ++i) {
1138                 if (ncp->nc_name[i] == '@' && ncp->nc_name[i+1] == '@') {
1139                         error = hammer_str_to_tid(ncp->nc_name + i + 2,
1140                                                   &ispfs, &asof, &localization);
1141                         if (error != 0) {
1142                                 i = nlen;
1143                                 break;
1144                         }
1145                         if (asof != HAMMER_MAX_TID)
1146                                 flags |= HAMMER_INODE_RO;
1147                         break;
1148                 }
1149         }
1150         nlen = i;
1151
1152         /*
1153          * If this is a PFS softlink we dive into the PFS
1154          */
1155         if (ispfs && nlen == 0) {
1156                 ip = hammer_get_inode(&trans, dip, HAMMER_OBJID_ROOT,
1157                                       asof, localization,
1158                                       flags, &error);
1159                 if (error == 0) {
1160                         error = hammer_get_vnode(ip, &vp);
1161                         hammer_rel_inode(ip, 0);
1162                 } else {
1163                         vp = NULL;
1164                 }
1165                 if (error == 0) {
1166                         vn_unlock(vp);
1167                         cache_setvp(ap->a_nch, vp);
1168                         vrele(vp);
1169                 }
1170                 goto done;
1171         }
1172
1173         /*
1174          * If there is no path component the time extension is relative to dip.
1175          * e.g. "fubar/@@<snapshot>"
1176          *
1177          * "." is handled by the kernel, but ".@@<snapshot>" is not.
1178          * e.g. "fubar/.@@<snapshot>"
1179          *
1180          * ".." is handled by the kernel.  We do not currently handle
1181          * "..@<snapshot>".
1182          */
1183         if (nlen == 0 || (nlen == 1 && ncp->nc_name[0] == '.')) {
1184                 ip = hammer_get_inode(&trans, dip, dip->obj_id,
1185                                       asof, dip->obj_localization,
1186                                       flags, &error);
1187                 if (error == 0) {
1188                         error = hammer_get_vnode(ip, &vp);
1189                         hammer_rel_inode(ip, 0);
1190                 } else {
1191                         vp = NULL;
1192                 }
1193                 if (error == 0) {
1194                         vn_unlock(vp);
1195                         cache_setvp(ap->a_nch, vp);
1196                         vrele(vp);
1197                 }
1198                 goto done;
1199         }
1200
1201         /*
1202          * Calculate the namekey and setup the key range for the scan.  This
1203          * works kinda like a chained hash table where the lower 32 bits
1204          * of the namekey synthesize the chain.
1205          *
1206          * The key range is inclusive of both key_beg and key_end.
1207          */
1208         namekey = hammer_directory_namekey(dip, ncp->nc_name, nlen,
1209                                            &max_iterations);
1210
1211         error = hammer_init_cursor(&trans, &cursor, &dip->cache[1], dip);
1212         cursor.key_beg.localization = dip->obj_localization +
1213                                       hammer_dir_localization(dip);
1214         cursor.key_beg.obj_id = dip->obj_id;
1215         cursor.key_beg.key = namekey;
1216         cursor.key_beg.create_tid = 0;
1217         cursor.key_beg.delete_tid = 0;
1218         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1219         cursor.key_beg.obj_type = 0;
1220
1221         cursor.key_end = cursor.key_beg;
1222         cursor.key_end.key += max_iterations;
1223         cursor.asof = asof;
1224         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1225
1226         /*
1227          * Scan all matching records (the chain), locate the one matching
1228          * the requested path component.
1229          *
1230          * The hammer_ip_*() functions merge in-memory records with on-disk
1231          * records for the purposes of the search.
1232          */
1233         obj_id = 0;
1234         localization = HAMMER_DEF_LOCALIZATION;
1235
1236         if (error == 0) {
1237                 error = hammer_ip_first(&cursor);
1238                 while (error == 0) {
1239                         error = hammer_ip_resolve_data(&cursor);
1240                         if (error)
1241                                 break;
1242                         if (nlen == cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF &&
1243                             bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1244                                 obj_id = cursor.data->entry.obj_id;
1245                                 localization = cursor.data->entry.localization;
1246                                 break;
1247                         }
1248                         error = hammer_ip_next(&cursor);
1249                 }
1250         }
1251         hammer_done_cursor(&cursor);
1252
1253         /*
1254          * Lookup the obj_id.  This should always succeed.  If it does not
1255          * the filesystem may be damaged and we return a dummy inode.
1256          */
1257         if (error == 0) {
1258                 ip = hammer_get_inode(&trans, dip, obj_id,
1259                                       asof, localization,
1260                                       flags, &error);
1261                 if (error == ENOENT) {
1262                         hkprintf("WARNING: Missing inode for dirent \"%s\"\n"
1263                                 "\tobj_id = %016llx, asof=%016llx, lo=%08x\n",
1264                                 ncp->nc_name,
1265                                 (long long)obj_id, (long long)asof,
1266                                 localization);
1267                         error = 0;
1268                         ip = hammer_get_dummy_inode(&trans, dip, obj_id,
1269                                                     asof, localization,
1270                                                     flags, &error);
1271                 }
1272                 if (error == 0) {
1273                         error = hammer_get_vnode(ip, &vp);
1274                         hammer_rel_inode(ip, 0);
1275                 } else {
1276                         vp = NULL;
1277                 }
1278                 if (error == 0) {
1279                         vn_unlock(vp);
1280                         cache_setvp(ap->a_nch, vp);
1281                         vrele(vp);
1282                 }
1283         } else if (error == ENOENT) {
1284                 cache_setvp(ap->a_nch, NULL);
1285         }
1286 done:
1287         hammer_done_transaction(&trans);
1288         lwkt_reltoken(&hmp->fs_token);
1289         return (error);
1290 }
1291
1292 /*
1293  * hammer_vop_nlookupdotdot { dvp, vpp, cred }
1294  *
1295  * Locate the parent directory of a directory vnode.
1296  *
1297  * dvp is referenced but not locked.  *vpp must be returned referenced and
1298  * locked.  A parent_obj_id of 0 does not necessarily indicate that we are
1299  * at the root, instead it could indicate that the directory we were in was
1300  * removed.
1301  *
1302  * NOTE: as-of sequences are not linked into the directory structure.  If
1303  * we are at the root with a different asof then the mount point, reload
1304  * the same directory with the mount point's asof.   I'm not sure what this
1305  * will do to NFS.  We encode ASOF stamps in NFS file handles so it might not
1306  * get confused, but it hasn't been tested.
1307  */
1308 static
1309 int
1310 hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *ap)
1311 {
1312         struct hammer_transaction trans;
1313         struct hammer_inode *dip;
1314         struct hammer_inode *ip;
1315         hammer_mount_t hmp;
1316         int64_t parent_obj_id;
1317         u_int32_t parent_obj_localization;
1318         hammer_tid_t asof;
1319         int error;
1320
1321         dip = VTOI(ap->a_dvp);
1322         asof = dip->obj_asof;
1323         hmp = dip->hmp;
1324
1325         /*
1326          * Whos are parent?  This could be the root of a pseudo-filesystem
1327          * whos parent is in another localization domain.
1328          */
1329         lwkt_gettoken(&hmp->fs_token);
1330         parent_obj_id = dip->ino_data.parent_obj_id;
1331         if (dip->obj_id == HAMMER_OBJID_ROOT)
1332                 parent_obj_localization = dip->ino_data.ext.obj.parent_obj_localization;
1333         else
1334                 parent_obj_localization = dip->obj_localization;
1335
1336         /*
1337          * It's probably a PFS root when dip->ino_data.parent_obj_id is 0.
1338          */
1339         if (parent_obj_id == 0) {
1340                 if (dip->obj_id == HAMMER_OBJID_ROOT &&
1341                    asof != hmp->asof) {
1342                         parent_obj_id = dip->obj_id;
1343                         asof = hmp->asof;
1344                         *ap->a_fakename = kmalloc(19, M_TEMP, M_WAITOK);
1345                         ksnprintf(*ap->a_fakename, 19, "0x%016llx",
1346                                   (long long)dip->obj_asof);
1347                 } else {
1348                         *ap->a_vpp = NULL;
1349                         lwkt_reltoken(&hmp->fs_token);
1350                         return ENOENT;
1351                 }
1352         }
1353
1354         hammer_simple_transaction(&trans, hmp);
1355         ++hammer_stats_file_iopsr;
1356
1357         ip = hammer_get_inode(&trans, dip, parent_obj_id,
1358                               asof, parent_obj_localization,
1359                               dip->flags, &error);
1360         if (ip) {
1361                 error = hammer_get_vnode(ip, ap->a_vpp);
1362                 hammer_rel_inode(ip, 0);
1363         } else {
1364                 *ap->a_vpp = NULL;
1365         }
1366         hammer_done_transaction(&trans);
1367         lwkt_reltoken(&hmp->fs_token);
1368         return (error);
1369 }
1370
1371 /*
1372  * hammer_vop_nlink { nch, dvp, vp, cred }
1373  */
1374 static
1375 int
1376 hammer_vop_nlink(struct vop_nlink_args *ap)
1377 {
1378         struct hammer_transaction trans;
1379         struct hammer_inode *dip;
1380         struct hammer_inode *ip;
1381         struct nchandle *nch;
1382         hammer_mount_t hmp;
1383         int error;
1384
1385         if (ap->a_dvp->v_mount != ap->a_vp->v_mount)
1386                 return(EXDEV);
1387
1388         nch = ap->a_nch;
1389         dip = VTOI(ap->a_dvp);
1390         ip = VTOI(ap->a_vp);
1391         hmp = dip->hmp;
1392
1393         if (dip->obj_localization != ip->obj_localization)
1394                 return(EXDEV);
1395
1396         if (dip->flags & HAMMER_INODE_RO)
1397                 return (EROFS);
1398         if (ip->flags & HAMMER_INODE_RO)
1399                 return (EROFS);
1400         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1401                 return (error);
1402
1403         /*
1404          * Create a transaction to cover the operations we perform.
1405          */
1406         lwkt_gettoken(&hmp->fs_token);
1407         hammer_start_transaction(&trans, hmp);
1408         ++hammer_stats_file_iopsw;
1409
1410         /*
1411          * Add the filesystem object to the directory.  Note that neither
1412          * dip nor ip are referenced or locked, but their vnodes are
1413          * referenced.  This function will bump the inode's link count.
1414          */
1415         error = hammer_ip_add_directory(&trans, dip,
1416                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1417                                         ip);
1418
1419         /*
1420          * Finish up.
1421          */
1422         if (error == 0) {
1423                 cache_setunresolved(nch);
1424                 cache_setvp(nch, ap->a_vp);
1425         }
1426         hammer_done_transaction(&trans);
1427         hammer_knote(ap->a_vp, NOTE_LINK);
1428         hammer_knote(ap->a_dvp, NOTE_WRITE);
1429         lwkt_reltoken(&hmp->fs_token);
1430         return (error);
1431 }
1432
1433 /*
1434  * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
1435  *
1436  * The operating system has already ensured that the directory entry
1437  * does not exist and done all appropriate namespace locking.
1438  */
1439 static
1440 int
1441 hammer_vop_nmkdir(struct vop_nmkdir_args *ap)
1442 {
1443         struct hammer_transaction trans;
1444         struct hammer_inode *dip;
1445         struct hammer_inode *nip;
1446         struct nchandle *nch;
1447         hammer_mount_t hmp;
1448         int error;
1449
1450         nch = ap->a_nch;
1451         dip = VTOI(ap->a_dvp);
1452         hmp = dip->hmp;
1453
1454         if (dip->flags & HAMMER_INODE_RO)
1455                 return (EROFS);
1456         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1457                 return (error);
1458
1459         /*
1460          * Create a transaction to cover the operations we perform.
1461          */
1462         lwkt_gettoken(&hmp->fs_token);
1463         hammer_start_transaction(&trans, hmp);
1464         ++hammer_stats_file_iopsw;
1465
1466         /*
1467          * Create a new filesystem object of the requested type.  The
1468          * returned inode will be referenced but not locked.
1469          */
1470         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1471                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1472                                     NULL, &nip);
1473         if (error) {
1474                 hkprintf("hammer_mkdir error %d\n", error);
1475                 hammer_done_transaction(&trans);
1476                 *ap->a_vpp = NULL;
1477                 lwkt_reltoken(&hmp->fs_token);
1478                 return (error);
1479         }
1480         /*
1481          * Add the new filesystem object to the directory.  This will also
1482          * bump the inode's link count.
1483          */
1484         error = hammer_ip_add_directory(&trans, dip,
1485                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1486                                         nip);
1487         if (error)
1488                 hkprintf("hammer_mkdir (add) error %d\n", error);
1489
1490         /*
1491          * Finish up.
1492          */
1493         if (error) {
1494                 hammer_rel_inode(nip, 0);
1495                 *ap->a_vpp = NULL;
1496         } else {
1497                 error = hammer_get_vnode(nip, ap->a_vpp);
1498                 hammer_rel_inode(nip, 0);
1499                 if (error == 0) {
1500                         cache_setunresolved(ap->a_nch);
1501                         cache_setvp(ap->a_nch, *ap->a_vpp);
1502                 }
1503         }
1504         hammer_done_transaction(&trans);
1505         if (error == 0)
1506                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
1507         lwkt_reltoken(&hmp->fs_token);
1508         return (error);
1509 }
1510
1511 /*
1512  * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1513  *
1514  * The operating system has already ensured that the directory entry
1515  * does not exist and done all appropriate namespace locking.
1516  */
1517 static
1518 int
1519 hammer_vop_nmknod(struct vop_nmknod_args *ap)
1520 {
1521         struct hammer_transaction trans;
1522         struct hammer_inode *dip;
1523         struct hammer_inode *nip;
1524         struct nchandle *nch;
1525         hammer_mount_t hmp;
1526         int error;
1527
1528         nch = ap->a_nch;
1529         dip = VTOI(ap->a_dvp);
1530         hmp = dip->hmp;
1531
1532         if (dip->flags & HAMMER_INODE_RO)
1533                 return (EROFS);
1534         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1535                 return (error);
1536
1537         /*
1538          * Create a transaction to cover the operations we perform.
1539          */
1540         lwkt_gettoken(&hmp->fs_token);
1541         hammer_start_transaction(&trans, hmp);
1542         ++hammer_stats_file_iopsw;
1543
1544         /*
1545          * Create a new filesystem object of the requested type.  The
1546          * returned inode will be referenced but not locked.
1547          *
1548          * If mknod specifies a directory a pseudo-fs is created.
1549          */
1550         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1551                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1552                                     NULL, &nip);
1553         if (error) {
1554                 hammer_done_transaction(&trans);
1555                 *ap->a_vpp = NULL;
1556                 lwkt_reltoken(&hmp->fs_token);
1557                 return (error);
1558         }
1559
1560         /*
1561          * Add the new filesystem object to the directory.  This will also
1562          * bump the inode's link count.
1563          */
1564         error = hammer_ip_add_directory(&trans, dip,
1565                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1566                                         nip);
1567
1568         /*
1569          * Finish up.
1570          */
1571         if (error) {
1572                 hammer_rel_inode(nip, 0);
1573                 *ap->a_vpp = NULL;
1574         } else {
1575                 error = hammer_get_vnode(nip, ap->a_vpp);
1576                 hammer_rel_inode(nip, 0);
1577                 if (error == 0) {
1578                         cache_setunresolved(ap->a_nch);
1579                         cache_setvp(ap->a_nch, *ap->a_vpp);
1580                 }
1581         }
1582         hammer_done_transaction(&trans);
1583         if (error == 0)
1584                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1585         lwkt_reltoken(&hmp->fs_token);
1586         return (error);
1587 }
1588
1589 /*
1590  * hammer_vop_open { vp, mode, cred, fp }
1591  *
1592  * MPSAFE (does not require fs_token)
1593  */
1594 static
1595 int
1596 hammer_vop_open(struct vop_open_args *ap)
1597 {
1598         hammer_inode_t ip;
1599
1600         ++hammer_stats_file_iopsr;
1601         ip = VTOI(ap->a_vp);
1602
1603         if ((ap->a_mode & FWRITE) && (ip->flags & HAMMER_INODE_RO))
1604                 return (EROFS);
1605         return(vop_stdopen(ap));
1606 }
1607
1608 /*
1609  * hammer_vop_print { vp }
1610  */
1611 static
1612 int
1613 hammer_vop_print(struct vop_print_args *ap)
1614 {
1615         return EOPNOTSUPP;
1616 }
1617
1618 /*
1619  * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1620  */
1621 static
1622 int
1623 hammer_vop_readdir(struct vop_readdir_args *ap)
1624 {
1625         struct hammer_transaction trans;
1626         struct hammer_cursor cursor;
1627         struct hammer_inode *ip;
1628         hammer_mount_t hmp;
1629         struct uio *uio;
1630         hammer_base_elm_t base;
1631         int error;
1632         int cookie_index;
1633         int ncookies;
1634         off_t *cookies;
1635         off_t saveoff;
1636         int r;
1637         int dtype;
1638
1639         ++hammer_stats_file_iopsr;
1640         ip = VTOI(ap->a_vp);
1641         uio = ap->a_uio;
1642         saveoff = uio->uio_offset;
1643         hmp = ip->hmp;
1644
1645         if (ap->a_ncookies) {
1646                 ncookies = uio->uio_resid / 16 + 1;
1647                 if (ncookies > 1024)
1648                         ncookies = 1024;
1649                 cookies = kmalloc(ncookies * sizeof(off_t), M_TEMP, M_WAITOK);
1650                 cookie_index = 0;
1651         } else {
1652                 ncookies = -1;
1653                 cookies = NULL;
1654                 cookie_index = 0;
1655         }
1656
1657         lwkt_gettoken(&hmp->fs_token);
1658         hammer_simple_transaction(&trans, hmp);
1659
1660         /*
1661          * Handle artificial entries
1662          *
1663          * It should be noted that the minimum value for a directory
1664          * hash key on-media is 0x0000000100000000, so we can use anything
1665          * less then that to represent our 'special' key space.
1666          */
1667         error = 0;
1668         if (saveoff == 0) {
1669                 r = vop_write_dirent(&error, uio, ip->obj_id, DT_DIR, 1, ".");
1670                 if (r)
1671                         goto done;
1672                 if (cookies)
1673                         cookies[cookie_index] = saveoff;
1674                 ++saveoff;
1675                 ++cookie_index;
1676                 if (cookie_index == ncookies)
1677                         goto done;
1678         }
1679         if (saveoff == 1) {
1680                 if (ip->ino_data.parent_obj_id) {
1681                         r = vop_write_dirent(&error, uio,
1682                                              ip->ino_data.parent_obj_id,
1683                                              DT_DIR, 2, "..");
1684                 } else {
1685                         r = vop_write_dirent(&error, uio,
1686                                              ip->obj_id, DT_DIR, 2, "..");
1687                 }
1688                 if (r)
1689                         goto done;
1690                 if (cookies)
1691                         cookies[cookie_index] = saveoff;
1692                 ++saveoff;
1693                 ++cookie_index;
1694                 if (cookie_index == ncookies)
1695                         goto done;
1696         }
1697
1698         /*
1699          * Key range (begin and end inclusive) to scan.  Directory keys
1700          * directly translate to a 64 bit 'seek' position.
1701          */
1702         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1703         cursor.key_beg.localization = ip->obj_localization +
1704                                       hammer_dir_localization(ip);
1705         cursor.key_beg.obj_id = ip->obj_id;
1706         cursor.key_beg.create_tid = 0;
1707         cursor.key_beg.delete_tid = 0;
1708         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1709         cursor.key_beg.obj_type = 0;
1710         cursor.key_beg.key = saveoff;
1711
1712         cursor.key_end = cursor.key_beg;
1713         cursor.key_end.key = HAMMER_MAX_KEY;
1714         cursor.asof = ip->obj_asof;
1715         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1716
1717         error = hammer_ip_first(&cursor);
1718
1719         while (error == 0) {
1720                 error = hammer_ip_resolve_data(&cursor);
1721                 if (error)
1722                         break;
1723                 base = &cursor.leaf->base;
1724                 saveoff = base->key;
1725                 KKASSERT(cursor.leaf->data_len > HAMMER_ENTRY_NAME_OFF);
1726
1727                 if (base->obj_id != ip->obj_id)
1728                         hpanic("bad record at %p", cursor.node);
1729
1730                 /*
1731                  * Convert pseudo-filesystems into softlinks
1732                  */
1733                 dtype = hammer_get_dtype(cursor.leaf->base.obj_type);
1734                 r = vop_write_dirent(
1735                              &error, uio, cursor.data->entry.obj_id,
1736                              dtype,
1737                              cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF ,
1738                              (void *)cursor.data->entry.name);
1739                 if (r)
1740                         break;
1741                 ++saveoff;
1742                 if (cookies)
1743                         cookies[cookie_index] = base->key;
1744                 ++cookie_index;
1745                 if (cookie_index == ncookies)
1746                         break;
1747                 error = hammer_ip_next(&cursor);
1748         }
1749         hammer_done_cursor(&cursor);
1750
1751 done:
1752         hammer_done_transaction(&trans);
1753
1754         if (ap->a_eofflag)
1755                 *ap->a_eofflag = (error == ENOENT);
1756         uio->uio_offset = saveoff;
1757         if (error && cookie_index == 0) {
1758                 if (error == ENOENT)
1759                         error = 0;
1760                 if (cookies) {
1761                         kfree(cookies, M_TEMP);
1762                         *ap->a_ncookies = 0;
1763                         *ap->a_cookies = NULL;
1764                 }
1765         } else {
1766                 if (error == ENOENT)
1767                         error = 0;
1768                 if (cookies) {
1769                         *ap->a_ncookies = cookie_index;
1770                         *ap->a_cookies = cookies;
1771                 }
1772         }
1773         lwkt_reltoken(&hmp->fs_token);
1774         return(error);
1775 }
1776
1777 /*
1778  * hammer_vop_readlink { vp, uio, cred }
1779  */
1780 static
1781 int
1782 hammer_vop_readlink(struct vop_readlink_args *ap)
1783 {
1784         struct hammer_transaction trans;
1785         struct hammer_cursor cursor;
1786         struct hammer_inode *ip;
1787         hammer_mount_t hmp;
1788         char buf[32];
1789         u_int32_t localization;
1790         hammer_pseudofs_inmem_t pfsm;
1791         int error;
1792
1793         ip = VTOI(ap->a_vp);
1794         hmp = ip->hmp;
1795
1796         lwkt_gettoken(&hmp->fs_token);
1797
1798         /*
1799          * Shortcut if the symlink data was stuffed into ino_data.
1800          *
1801          * Also expand special "@@PFS%05d" softlinks (expansion only
1802          * occurs for non-historical (current) accesses made from the
1803          * primary filesystem).
1804          *
1805          * Note that userspace hammer command does not allow users to
1806          * create a @@PFS softlink under an existing other PFS (id!=0)
1807          * so the ip localization here for @@PFS softlink is always 0.
1808          */
1809         if (ip->ino_data.size <= HAMMER_INODE_BASESYMLEN) {
1810                 char *ptr;
1811                 int bytes;
1812
1813                 ptr = ip->ino_data.ext.symlink;
1814                 bytes = (int)ip->ino_data.size;
1815                 if (bytes == 10 &&
1816                     ip->obj_asof == HAMMER_MAX_TID &&
1817                     ip->obj_localization == 0 &&
1818                     strncmp(ptr, "@@PFS", 5) == 0) {
1819                         hammer_simple_transaction(&trans, hmp);
1820                         bcopy(ptr + 5, buf, 5);
1821                         buf[5] = 0;
1822                         localization = strtoul(buf, NULL, 10) << 16;
1823                         pfsm = hammer_load_pseudofs(&trans, localization,
1824                                                     &error);
1825                         if (error == 0) {
1826                                 if (pfsm->pfsd.mirror_flags &
1827                                     HAMMER_PFSD_SLAVE) {
1828                                         /* vap->va_size == 26 */
1829                                         ksnprintf(buf, sizeof(buf),
1830                                                   "@@0x%016llx:%05d",
1831                                                   (long long)pfsm->pfsd.sync_end_tid,
1832                                                   localization >> 16);
1833                                 } else {
1834                                         /* vap->va_size == 10 */
1835                                         ksnprintf(buf, sizeof(buf),
1836                                                   "@@-1:%05d",
1837                                                   localization >> 16);
1838 #if 0
1839                                         ksnprintf(buf, sizeof(buf),
1840                                                   "@@0x%016llx:%05d",
1841                                                   (long long)HAMMER_MAX_TID,
1842                                                   localization >> 16);
1843 #endif
1844                                 }
1845                                 ptr = buf;
1846                                 bytes = strlen(buf);
1847                         }
1848                         if (pfsm)
1849                                 hammer_rel_pseudofs(hmp, pfsm);
1850                         hammer_done_transaction(&trans);
1851                 }
1852                 error = uiomove(ptr, bytes, ap->a_uio);
1853                 lwkt_reltoken(&hmp->fs_token);
1854                 return(error);
1855         }
1856
1857         /*
1858          * Long version
1859          */
1860         hammer_simple_transaction(&trans, hmp);
1861         ++hammer_stats_file_iopsr;
1862         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1863
1864         /*
1865          * Key range (begin and end inclusive) to scan.  Directory keys
1866          * directly translate to a 64 bit 'seek' position.
1867          */
1868         cursor.key_beg.localization = ip->obj_localization +
1869                                       HAMMER_LOCALIZE_MISC;
1870         cursor.key_beg.obj_id = ip->obj_id;
1871         cursor.key_beg.create_tid = 0;
1872         cursor.key_beg.delete_tid = 0;
1873         cursor.key_beg.rec_type = HAMMER_RECTYPE_FIX;
1874         cursor.key_beg.obj_type = 0;
1875         cursor.key_beg.key = HAMMER_FIXKEY_SYMLINK;
1876         cursor.asof = ip->obj_asof;
1877         cursor.flags |= HAMMER_CURSOR_ASOF;
1878
1879         error = hammer_ip_lookup(&cursor);
1880         if (error == 0) {
1881                 error = hammer_ip_resolve_data(&cursor);
1882                 if (error == 0) {
1883                         KKASSERT(cursor.leaf->data_len >=
1884                                  HAMMER_SYMLINK_NAME_OFF);
1885                         error = uiomove(cursor.data->symlink.name,
1886                                         cursor.leaf->data_len -
1887                                                 HAMMER_SYMLINK_NAME_OFF,
1888                                         ap->a_uio);
1889                 }
1890         }
1891         hammer_done_cursor(&cursor);
1892         hammer_done_transaction(&trans);
1893         lwkt_reltoken(&hmp->fs_token);
1894         return(error);
1895 }
1896
1897 /*
1898  * hammer_vop_nremove { nch, dvp, cred }
1899  */
1900 static
1901 int
1902 hammer_vop_nremove(struct vop_nremove_args *ap)
1903 {
1904         struct hammer_transaction trans;
1905         struct hammer_inode *dip;
1906         hammer_mount_t hmp;
1907         int error;
1908
1909         dip = VTOI(ap->a_dvp);
1910         hmp = dip->hmp;
1911
1912         if (hammer_nohistory(dip) == 0 &&
1913             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1914                 return (error);
1915         }
1916
1917         lwkt_gettoken(&hmp->fs_token);
1918         hammer_start_transaction(&trans, hmp);
1919         ++hammer_stats_file_iopsw;
1920         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 0);
1921         hammer_done_transaction(&trans);
1922         if (error == 0)
1923                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1924         lwkt_reltoken(&hmp->fs_token);
1925         return (error);
1926 }
1927
1928 /*
1929  * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1930  */
1931 static
1932 int
1933 hammer_vop_nrename(struct vop_nrename_args *ap)
1934 {
1935         struct hammer_transaction trans;
1936         struct namecache *fncp;
1937         struct namecache *tncp;
1938         struct hammer_inode *fdip;
1939         struct hammer_inode *tdip;
1940         struct hammer_inode *ip;
1941         hammer_mount_t hmp;
1942         struct hammer_cursor cursor;
1943         int64_t namekey;
1944         u_int32_t max_iterations;
1945         int nlen, error;
1946
1947         if (ap->a_fdvp->v_mount != ap->a_tdvp->v_mount)
1948                 return(EXDEV);
1949         if (ap->a_fdvp->v_mount != ap->a_fnch->ncp->nc_vp->v_mount)
1950                 return(EXDEV);
1951
1952         fdip = VTOI(ap->a_fdvp);
1953         tdip = VTOI(ap->a_tdvp);
1954         fncp = ap->a_fnch->ncp;
1955         tncp = ap->a_tnch->ncp;
1956         ip = VTOI(fncp->nc_vp);
1957         KKASSERT(ip != NULL);
1958
1959         hmp = ip->hmp;
1960
1961         if (fdip->obj_localization != tdip->obj_localization)
1962                 return(EXDEV);
1963         if (fdip->obj_localization != ip->obj_localization)
1964                 return(EXDEV);
1965
1966         if (fdip->flags & HAMMER_INODE_RO)
1967                 return (EROFS);
1968         if (tdip->flags & HAMMER_INODE_RO)
1969                 return (EROFS);
1970         if (ip->flags & HAMMER_INODE_RO)
1971                 return (EROFS);
1972         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1973                 return (error);
1974
1975         lwkt_gettoken(&hmp->fs_token);
1976         hammer_start_transaction(&trans, hmp);
1977         ++hammer_stats_file_iopsw;
1978
1979         /*
1980          * Remove tncp from the target directory and then link ip as
1981          * tncp. XXX pass trans to dounlink
1982          *
1983          * Force the inode sync-time to match the transaction so it is
1984          * in-sync with the creation of the target directory entry.
1985          */
1986         error = hammer_dounlink(&trans, ap->a_tnch, ap->a_tdvp,
1987                                 ap->a_cred, 0, -1);
1988         if (error == 0 || error == ENOENT) {
1989                 error = hammer_ip_add_directory(&trans, tdip,
1990                                                 tncp->nc_name, tncp->nc_nlen,
1991                                                 ip);
1992                 if (error == 0) {
1993                         ip->ino_data.parent_obj_id = tdip->obj_id;
1994                         ip->ino_data.ctime = trans.time;
1995                         hammer_modify_inode(&trans, ip, HAMMER_INODE_DDIRTY);
1996                 }
1997         }
1998         if (error)
1999                 goto failed; /* XXX */
2000
2001         /*
2002          * Locate the record in the originating directory and remove it.
2003          *
2004          * Calculate the namekey and setup the key range for the scan.  This
2005          * works kinda like a chained hash table where the lower 32 bits
2006          * of the namekey synthesize the chain.
2007          *
2008          * The key range is inclusive of both key_beg and key_end.
2009          */
2010         namekey = hammer_directory_namekey(fdip, fncp->nc_name, fncp->nc_nlen,
2011                                            &max_iterations);
2012 retry:
2013         hammer_init_cursor(&trans, &cursor, &fdip->cache[1], fdip);
2014         cursor.key_beg.localization = fdip->obj_localization +
2015                                       hammer_dir_localization(fdip);
2016         cursor.key_beg.obj_id = fdip->obj_id;
2017         cursor.key_beg.key = namekey;
2018         cursor.key_beg.create_tid = 0;
2019         cursor.key_beg.delete_tid = 0;
2020         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
2021         cursor.key_beg.obj_type = 0;
2022
2023         cursor.key_end = cursor.key_beg;
2024         cursor.key_end.key += max_iterations;
2025         cursor.asof = fdip->obj_asof;
2026         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2027
2028         /*
2029          * Scan all matching records (the chain), locate the one matching
2030          * the requested path component.
2031          *
2032          * The hammer_ip_*() functions merge in-memory records with on-disk
2033          * records for the purposes of the search.
2034          */
2035         error = hammer_ip_first(&cursor);
2036         while (error == 0) {
2037                 if (hammer_ip_resolve_data(&cursor) != 0)
2038                         break;
2039                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
2040                 KKASSERT(nlen > 0);
2041                 if (fncp->nc_nlen == nlen &&
2042                     bcmp(fncp->nc_name, cursor.data->entry.name, nlen) == 0) {
2043                         break;
2044                 }
2045                 error = hammer_ip_next(&cursor);
2046         }
2047
2048         /*
2049          * If all is ok we have to get the inode so we can adjust nlinks.
2050          *
2051          * WARNING: hammer_ip_del_directory() may have to terminate the
2052          * cursor to avoid a recursion.  It's ok to call hammer_done_cursor()
2053          * twice.
2054          */
2055         if (error == 0)
2056                 error = hammer_ip_del_directory(&trans, &cursor, fdip, ip);
2057
2058         /*
2059          * XXX A deadlock here will break rename's atomicy for the purposes
2060          * of crash recovery.
2061          */
2062         if (error == EDEADLK) {
2063                 hammer_done_cursor(&cursor);
2064                 goto retry;
2065         }
2066
2067         /*
2068          * Cleanup and tell the kernel that the rename succeeded.
2069          *
2070          * NOTE: ip->vp, if non-NULL, cannot be directly referenced
2071          *       without formally acquiring the vp since the vp might
2072          *       have zero refs on it, or in the middle of a reclaim,
2073          *       etc.
2074          */
2075         hammer_done_cursor(&cursor);
2076         if (error == 0) {
2077                 cache_rename(ap->a_fnch, ap->a_tnch);
2078                 hammer_knote(ap->a_fdvp, NOTE_WRITE);
2079                 hammer_knote(ap->a_tdvp, NOTE_WRITE);
2080                 while (ip->vp) {
2081                         struct vnode *vp;
2082
2083                         error = hammer_get_vnode(ip, &vp);
2084                         if (error == 0 && vp) {
2085                                 vn_unlock(vp);
2086                                 hammer_knote(ip->vp, NOTE_RENAME);
2087                                 vrele(vp);
2088                                 break;
2089                         }
2090                         hdkprintf("ip/vp race2 avoided\n");
2091                 }
2092         }
2093
2094 failed:
2095         hammer_done_transaction(&trans);
2096         lwkt_reltoken(&hmp->fs_token);
2097         return (error);
2098 }
2099
2100 /*
2101  * hammer_vop_nrmdir { nch, dvp, cred }
2102  */
2103 static
2104 int
2105 hammer_vop_nrmdir(struct vop_nrmdir_args *ap)
2106 {
2107         struct hammer_transaction trans;
2108         struct hammer_inode *dip;
2109         hammer_mount_t hmp;
2110         int error;
2111
2112         dip = VTOI(ap->a_dvp);
2113         hmp = dip->hmp;
2114
2115         if (hammer_nohistory(dip) == 0 &&
2116             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2117                 return (error);
2118         }
2119
2120         lwkt_gettoken(&hmp->fs_token);
2121         hammer_start_transaction(&trans, hmp);
2122         ++hammer_stats_file_iopsw;
2123         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 1);
2124         hammer_done_transaction(&trans);
2125         if (error == 0)
2126                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
2127         lwkt_reltoken(&hmp->fs_token);
2128         return (error);
2129 }
2130
2131 /*
2132  * hammer_vop_markatime { vp, cred }
2133  */
2134 static
2135 int
2136 hammer_vop_markatime(struct vop_markatime_args *ap)
2137 {
2138         struct hammer_transaction trans;
2139         struct hammer_inode *ip;
2140         hammer_mount_t hmp;
2141
2142         ip = VTOI(ap->a_vp);
2143         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2144                 return (EROFS);
2145         if (ip->flags & HAMMER_INODE_RO)
2146                 return (EROFS);
2147         hmp = ip->hmp;
2148         if (hmp->mp->mnt_flag & MNT_NOATIME)
2149                 return (0);
2150         lwkt_gettoken(&hmp->fs_token);
2151         hammer_start_transaction(&trans, hmp);
2152         ++hammer_stats_file_iopsw;
2153
2154         ip->ino_data.atime = trans.time;
2155         hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
2156         hammer_done_transaction(&trans);
2157         hammer_knote(ap->a_vp, NOTE_ATTRIB);
2158         lwkt_reltoken(&hmp->fs_token);
2159         return (0);
2160 }
2161
2162 /*
2163  * hammer_vop_setattr { vp, vap, cred }
2164  */
2165 static
2166 int
2167 hammer_vop_setattr(struct vop_setattr_args *ap)
2168 {
2169         struct hammer_transaction trans;
2170         struct hammer_inode *ip;
2171         struct vattr *vap;
2172         hammer_mount_t hmp;
2173         int modflags;
2174         int error;
2175         int truncating;
2176         int blksize;
2177         int kflags;
2178 #if 0
2179         int64_t aligned_size;
2180 #endif
2181         u_int32_t flags;
2182
2183         vap = ap->a_vap;
2184         ip = ap->a_vp->v_data;
2185         modflags = 0;
2186         kflags = 0;
2187         hmp = ip->hmp;
2188
2189         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2190                 return(EROFS);
2191         if (ip->flags & HAMMER_INODE_RO)
2192                 return (EROFS);
2193         if (hammer_nohistory(ip) == 0 &&
2194             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2195                 return (error);
2196         }
2197
2198         lwkt_gettoken(&hmp->fs_token);
2199         hammer_start_transaction(&trans, hmp);
2200         ++hammer_stats_file_iopsw;
2201         error = 0;
2202
2203         if (vap->va_flags != VNOVAL) {
2204                 flags = ip->ino_data.uflags;
2205                 error = vop_helper_setattr_flags(&flags, vap->va_flags,
2206                                          hammer_to_unix_xid(&ip->ino_data.uid),
2207                                          ap->a_cred);
2208                 if (error == 0) {
2209                         if (ip->ino_data.uflags != flags) {
2210                                 ip->ino_data.uflags = flags;
2211                                 ip->ino_data.ctime = trans.time;
2212                                 modflags |= HAMMER_INODE_DDIRTY;
2213                                 kflags |= NOTE_ATTRIB;
2214                         }
2215                         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2216                                 error = 0;
2217                                 goto done;
2218                         }
2219                 }
2220                 goto done;
2221         }
2222         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2223                 error = EPERM;
2224                 goto done;
2225         }
2226         if (vap->va_uid != (uid_t)VNOVAL || vap->va_gid != (gid_t)VNOVAL) {
2227                 mode_t cur_mode = ip->ino_data.mode;
2228                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2229                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2230                 uuid_t uuid_uid;
2231                 uuid_t uuid_gid;
2232
2233                 error = vop_helper_chown(ap->a_vp, vap->va_uid, vap->va_gid,
2234                                          ap->a_cred,
2235                                          &cur_uid, &cur_gid, &cur_mode);
2236                 if (error == 0) {
2237                         hammer_guid_to_uuid(&uuid_uid, cur_uid);
2238                         hammer_guid_to_uuid(&uuid_gid, cur_gid);
2239                         if (bcmp(&uuid_uid, &ip->ino_data.uid,
2240                                  sizeof(uuid_uid)) ||
2241                             bcmp(&uuid_gid, &ip->ino_data.gid,
2242                                  sizeof(uuid_gid)) ||
2243                             ip->ino_data.mode != cur_mode) {
2244                                 ip->ino_data.uid = uuid_uid;
2245                                 ip->ino_data.gid = uuid_gid;
2246                                 ip->ino_data.mode = cur_mode;
2247                                 ip->ino_data.ctime = trans.time;
2248                                 modflags |= HAMMER_INODE_DDIRTY;
2249                         }
2250                         kflags |= NOTE_ATTRIB;
2251                 }
2252         }
2253         while (vap->va_size != VNOVAL && ip->ino_data.size != vap->va_size) {
2254                 switch(ap->a_vp->v_type) {
2255                 case VREG:
2256                         if (vap->va_size == ip->ino_data.size)
2257                                 break;
2258
2259                         /*
2260                          * Log the operation if in fast-fsync mode or if
2261                          * there are unterminated redo write records present.
2262                          *
2263                          * The second check is needed so the recovery code
2264                          * properly truncates write redos even if nominal
2265                          * REDO operations is turned off due to excessive
2266                          * writes, because the related records might be
2267                          * destroyed and never lay down a TERM_WRITE.
2268                          */
2269                         if ((ip->flags & HAMMER_INODE_REDO) ||
2270                             (ip->flags & HAMMER_INODE_RDIRTY)) {
2271                                 error = hammer_generate_redo(&trans, ip,
2272                                                              vap->va_size,
2273                                                              HAMMER_REDO_TRUNC,
2274                                                              NULL, 0);
2275                         }
2276                         blksize = hammer_blocksize(vap->va_size);
2277
2278                         /*
2279                          * XXX break atomicy, we can deadlock the backend
2280                          * if we do not release the lock.  Probably not a
2281                          * big deal here.
2282                          */
2283                         if (vap->va_size < ip->ino_data.size) {
2284                                 nvtruncbuf(ap->a_vp, vap->va_size,
2285                                            blksize,
2286                                            hammer_blockoff(vap->va_size),
2287                                            0);
2288                                 truncating = 1;
2289                                 kflags |= NOTE_WRITE;
2290                         } else {
2291                                 nvextendbuf(ap->a_vp,
2292                                             ip->ino_data.size,
2293                                             vap->va_size,
2294                                             hammer_blocksize(ip->ino_data.size),
2295                                             hammer_blocksize(vap->va_size),
2296                                             hammer_blockoff(ip->ino_data.size),
2297                                             hammer_blockoff(vap->va_size),
2298                                             0);
2299                                 truncating = 0;
2300                                 kflags |= NOTE_WRITE | NOTE_EXTEND;
2301                         }
2302                         ip->ino_data.size = vap->va_size;
2303                         ip->ino_data.mtime = trans.time;
2304                         /* XXX safe to use SDIRTY instead of DDIRTY here? */
2305                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2306
2307                         /*
2308                          * On-media truncation is cached in the inode until
2309                          * the inode is synchronized.  We must immediately
2310                          * handle any frontend records.
2311                          */
2312                         if (truncating) {
2313                                 hammer_ip_frontend_trunc(ip, vap->va_size);
2314 #ifdef DEBUG_TRUNCATE
2315                                 if (HammerTruncIp == NULL)
2316                                         HammerTruncIp = ip;
2317 #endif
2318                                 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2319                                         ip->flags |= HAMMER_INODE_TRUNCATED;
2320                                         ip->trunc_off = vap->va_size;
2321                                         hammer_inode_dirty(ip);
2322 #ifdef DEBUG_TRUNCATE
2323                                         if (ip == HammerTruncIp)
2324                                                 hdkprintf("truncate1 %016llx\n",
2325                                                         (long long)ip->trunc_off);
2326 #endif
2327                                 } else if (ip->trunc_off > vap->va_size) {
2328                                         ip->trunc_off = vap->va_size;
2329 #ifdef DEBUG_TRUNCATE
2330                                         if (ip == HammerTruncIp)
2331                                                 hdkprintf("truncate2 %016llx\n",
2332                                                         (long long)ip->trunc_off);
2333 #endif
2334                                 } else {
2335 #ifdef DEBUG_TRUNCATE
2336                                         if (ip == HammerTruncIp)
2337                                                 hdkprintf("truncate3 %016llx (ignored)\n",
2338                                                         (long long)vap->va_size);
2339 #endif
2340                                 }
2341                         }
2342
2343 #if 0
2344                         /*
2345                          * When truncating, nvtruncbuf() may have cleaned out
2346                          * a portion of the last block on-disk in the buffer
2347                          * cache.  We must clean out any frontend records
2348                          * for blocks beyond the new last block.
2349                          */
2350                         aligned_size = (vap->va_size + (blksize - 1)) &
2351                                        ~(int64_t)(blksize - 1);
2352                         if (truncating && vap->va_size < aligned_size) {
2353                                 aligned_size -= blksize;
2354                                 hammer_ip_frontend_trunc(ip, aligned_size);
2355                         }
2356 #endif
2357                         break;
2358                 case VDATABASE:
2359                         if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2360                                 ip->flags |= HAMMER_INODE_TRUNCATED;
2361                                 ip->trunc_off = vap->va_size;
2362                                 hammer_inode_dirty(ip);
2363                         } else if (ip->trunc_off > vap->va_size) {
2364                                 ip->trunc_off = vap->va_size;
2365                         }
2366                         hammer_ip_frontend_trunc(ip, vap->va_size);
2367                         ip->ino_data.size = vap->va_size;
2368                         ip->ino_data.mtime = trans.time;
2369                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2370                         kflags |= NOTE_ATTRIB;
2371                         break;
2372                 default:
2373                         error = EINVAL;
2374                         goto done;
2375                 }
2376                 break;
2377         }
2378         if (vap->va_atime.tv_sec != VNOVAL) {
2379                 ip->ino_data.atime = hammer_timespec_to_time(&vap->va_atime);
2380                 modflags |= HAMMER_INODE_ATIME;
2381                 kflags |= NOTE_ATTRIB;
2382         }
2383         if (vap->va_mtime.tv_sec != VNOVAL) {
2384                 ip->ino_data.mtime = hammer_timespec_to_time(&vap->va_mtime);
2385                 modflags |= HAMMER_INODE_MTIME;
2386                 kflags |= NOTE_ATTRIB;
2387         }
2388         if (vap->va_mode != (mode_t)VNOVAL) {
2389                 mode_t   cur_mode = ip->ino_data.mode;
2390                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2391                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2392
2393                 error = vop_helper_chmod(ap->a_vp, vap->va_mode, ap->a_cred,
2394                                          cur_uid, cur_gid, &cur_mode);
2395                 if (error == 0 && ip->ino_data.mode != cur_mode) {
2396                         ip->ino_data.mode = cur_mode;
2397                         ip->ino_data.ctime = trans.time;
2398                         modflags |= HAMMER_INODE_DDIRTY;
2399                         kflags |= NOTE_ATTRIB;
2400                 }
2401         }
2402 done:
2403         if (error == 0)
2404                 hammer_modify_inode(&trans, ip, modflags);
2405         hammer_done_transaction(&trans);
2406         hammer_knote(ap->a_vp, kflags);
2407         lwkt_reltoken(&hmp->fs_token);
2408         return (error);
2409 }
2410
2411 /*
2412  * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
2413  */
2414 static
2415 int
2416 hammer_vop_nsymlink(struct vop_nsymlink_args *ap)
2417 {
2418         struct hammer_transaction trans;
2419         struct hammer_inode *dip;
2420         struct hammer_inode *nip;
2421         hammer_record_t record;
2422         struct nchandle *nch;
2423         hammer_mount_t hmp;
2424         int error;
2425         int bytes;
2426
2427         ap->a_vap->va_type = VLNK;
2428
2429         nch = ap->a_nch;
2430         dip = VTOI(ap->a_dvp);
2431         hmp = dip->hmp;
2432
2433         if (dip->flags & HAMMER_INODE_RO)
2434                 return (EROFS);
2435         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
2436                 return (error);
2437
2438         /*
2439          * Create a transaction to cover the operations we perform.
2440          */
2441         lwkt_gettoken(&hmp->fs_token);
2442         hammer_start_transaction(&trans, hmp);
2443         ++hammer_stats_file_iopsw;
2444
2445         /*
2446          * Create a new filesystem object of the requested type.  The
2447          * returned inode will be referenced but not locked.
2448          */
2449
2450         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
2451                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
2452                                     NULL, &nip);
2453         if (error) {
2454                 hammer_done_transaction(&trans);
2455                 *ap->a_vpp = NULL;
2456                 lwkt_reltoken(&hmp->fs_token);
2457                 return (error);
2458         }
2459
2460         /*
2461          * Add a record representing the symlink.  symlink stores the link
2462          * as pure data, not a string, and is no \0 terminated.
2463          */
2464         if (error == 0) {
2465                 bytes = strlen(ap->a_target);
2466
2467                 if (bytes <= HAMMER_INODE_BASESYMLEN) {
2468                         bcopy(ap->a_target, nip->ino_data.ext.symlink, bytes);
2469                 } else {
2470                         record = hammer_alloc_mem_record(nip, bytes);
2471                         record->type = HAMMER_MEM_RECORD_GENERAL;
2472
2473                         record->leaf.base.localization = nip->obj_localization +
2474                                                          HAMMER_LOCALIZE_MISC;
2475                         record->leaf.base.key = HAMMER_FIXKEY_SYMLINK;
2476                         record->leaf.base.rec_type = HAMMER_RECTYPE_FIX;
2477                         record->leaf.data_len = bytes;
2478                         KKASSERT(HAMMER_SYMLINK_NAME_OFF == 0);
2479                         bcopy(ap->a_target, record->data->symlink.name, bytes);
2480                         error = hammer_ip_add_record(&trans, record);
2481                 }
2482
2483                 /*
2484                  * Set the file size to the length of the link.
2485                  */
2486                 if (error == 0) {
2487                         nip->ino_data.size = bytes;
2488                         hammer_modify_inode(&trans, nip, HAMMER_INODE_DDIRTY);
2489                 }
2490         }
2491         if (error == 0)
2492                 error = hammer_ip_add_directory(&trans, dip, nch->ncp->nc_name,
2493                                                 nch->ncp->nc_nlen, nip);
2494
2495         /*
2496          * Finish up.
2497          */
2498         if (error) {
2499                 hammer_rel_inode(nip, 0);
2500                 *ap->a_vpp = NULL;
2501         } else {
2502                 error = hammer_get_vnode(nip, ap->a_vpp);
2503                 hammer_rel_inode(nip, 0);
2504                 if (error == 0) {
2505                         cache_setunresolved(ap->a_nch);
2506                         cache_setvp(ap->a_nch, *ap->a_vpp);
2507                         hammer_knote(ap->a_dvp, NOTE_WRITE);
2508                 }
2509         }
2510         hammer_done_transaction(&trans);
2511         lwkt_reltoken(&hmp->fs_token);
2512         return (error);
2513 }
2514
2515 /*
2516  * hammer_vop_nwhiteout { nch, dvp, cred, flags }
2517  */
2518 static
2519 int
2520 hammer_vop_nwhiteout(struct vop_nwhiteout_args *ap)
2521 {
2522         struct hammer_transaction trans;
2523         struct hammer_inode *dip;
2524         hammer_mount_t hmp;
2525         int error;
2526
2527         dip = VTOI(ap->a_dvp);
2528         hmp = dip->hmp;
2529
2530         if (hammer_nohistory(dip) == 0 &&
2531             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0) {
2532                 return (error);
2533         }
2534
2535         lwkt_gettoken(&hmp->fs_token);
2536         hammer_start_transaction(&trans, hmp);
2537         ++hammer_stats_file_iopsw;
2538         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp,
2539                                 ap->a_cred, ap->a_flags, -1);
2540         hammer_done_transaction(&trans);
2541         lwkt_reltoken(&hmp->fs_token);
2542
2543         return (error);
2544 }
2545
2546 /*
2547  * hammer_vop_ioctl { vp, command, data, fflag, cred }
2548  */
2549 static
2550 int
2551 hammer_vop_ioctl(struct vop_ioctl_args *ap)
2552 {
2553         struct hammer_inode *ip = ap->a_vp->v_data;
2554         hammer_mount_t hmp = ip->hmp;
2555         int error;
2556
2557         ++hammer_stats_file_iopsr;
2558         lwkt_gettoken(&hmp->fs_token);
2559         error = hammer_ioctl(ip, ap->a_command, ap->a_data,
2560                              ap->a_fflag, ap->a_cred);
2561         lwkt_reltoken(&hmp->fs_token);
2562         return (error);
2563 }
2564
2565 static
2566 int
2567 hammer_vop_mountctl(struct vop_mountctl_args *ap)
2568 {
2569         static const struct mountctl_opt extraopt[] = {
2570                 { HMNT_NOHISTORY,       "nohistory" },
2571                 { HMNT_MASTERID,        "master" },
2572                 { HMNT_NOMIRROR,        "nomirror" },
2573                 { 0, NULL}
2574
2575         };
2576         struct hammer_mount *hmp;
2577         struct mount *mp;
2578         int usedbytes;
2579         int error;
2580
2581         error = 0;
2582         usedbytes = 0;
2583         mp = ap->a_head.a_ops->head.vv_mount;
2584         KKASSERT(mp->mnt_data != NULL);
2585         hmp = (struct hammer_mount *)mp->mnt_data;
2586
2587         lwkt_gettoken(&hmp->fs_token);
2588
2589         switch(ap->a_op) {
2590         case MOUNTCTL_SET_EXPORT:
2591                 if (ap->a_ctllen != sizeof(struct export_args))
2592                         error = EINVAL;
2593                 else
2594                         error = hammer_vfs_export(mp, ap->a_op,
2595                                       (const struct export_args *)ap->a_ctl);
2596                 break;
2597         case MOUNTCTL_MOUNTFLAGS:
2598                 /*
2599                  * Call standard mountctl VOP function
2600                  * so we get user mount flags.
2601                  */
2602                 error = vop_stdmountctl(ap);
2603                 if (error)
2604                         break;
2605
2606                 usedbytes = *ap->a_res;
2607
2608                 if (usedbytes > 0 && usedbytes < ap->a_buflen) {
2609                         usedbytes += vfs_flagstostr(hmp->hflags, extraopt,
2610                                                     ap->a_buf,
2611                                                     ap->a_buflen - usedbytes,
2612                                                     &error);
2613                 }
2614
2615                 *ap->a_res += usedbytes;
2616                 break;
2617         default:
2618                 error = vop_stdmountctl(ap);
2619                 break;
2620         }
2621         lwkt_reltoken(&hmp->fs_token);
2622         return(error);
2623 }
2624
2625 /*
2626  * hammer_vop_strategy { vp, bio }
2627  *
2628  * Strategy call, used for regular file read & write only.  Note that the
2629  * bp may represent a cluster.
2630  *
2631  * To simplify operation and allow better optimizations in the future,
2632  * this code does not make any assumptions with regards to buffer alignment
2633  * or size.
2634  */
2635 static
2636 int
2637 hammer_vop_strategy(struct vop_strategy_args *ap)
2638 {
2639         struct buf *bp;
2640         int error;
2641
2642         bp = ap->a_bio->bio_buf;
2643
2644         switch(bp->b_cmd) {
2645         case BUF_CMD_READ:
2646                 error = hammer_vop_strategy_read(ap);
2647                 break;
2648         case BUF_CMD_WRITE:
2649                 error = hammer_vop_strategy_write(ap);
2650                 break;
2651         default:
2652                 bp->b_error = error = EINVAL;
2653                 bp->b_flags |= B_ERROR;
2654                 biodone(ap->a_bio);
2655                 break;
2656         }
2657
2658         /* hammer_dump_dedup_cache(((hammer_inode_t)ap->a_vp->v_data)->hmp); */
2659
2660         return (error);
2661 }
2662
2663 /*
2664  * Read from a regular file.  Iterate the related records and fill in the
2665  * BIO/BUF.  Gaps are zero-filled.
2666  *
2667  * The support code in hammer_object.c should be used to deal with mixed
2668  * in-memory and on-disk records.
2669  *
2670  * NOTE: Can be called from the cluster code with an oversized buf.
2671  *
2672  * XXX atime update
2673  */
2674 static
2675 int
2676 hammer_vop_strategy_read(struct vop_strategy_args *ap)
2677 {
2678         struct hammer_transaction trans;
2679         struct hammer_inode *ip;
2680         struct hammer_inode *dip;
2681         hammer_mount_t hmp;
2682         struct hammer_cursor cursor;
2683         hammer_base_elm_t base;
2684         hammer_off_t disk_offset;
2685         struct bio *bio;
2686         struct bio *nbio;
2687         struct buf *bp;
2688         int64_t rec_offset;
2689         int64_t ran_end;
2690         int64_t tmp64;
2691         int error;
2692         int boff;
2693         int roff;
2694         int n;
2695         int isdedupable;
2696
2697         bio = ap->a_bio;
2698         bp = bio->bio_buf;
2699         ip = ap->a_vp->v_data;
2700         hmp = ip->hmp;
2701
2702         /*
2703          * The zone-2 disk offset may have been set by the cluster code via
2704          * a BMAP operation, or else should be NOOFFSET.
2705          *
2706          * Checking the high bits for a match against zone-2 should suffice.
2707          *
2708          * In cases where a lot of data duplication is present it may be
2709          * more beneficial to drop through and doubule-buffer through the
2710          * device.
2711          */
2712         nbio = push_bio(bio);
2713         if ((nbio->bio_offset & HAMMER_OFF_ZONE_MASK) ==
2714             HAMMER_ZONE_LARGE_DATA) {
2715                 if (hammer_double_buffer == 0) {
2716                         lwkt_gettoken(&hmp->fs_token);
2717                         error = hammer_io_direct_read(hmp, nbio, NULL);
2718                         lwkt_reltoken(&hmp->fs_token);
2719                         return (error);
2720                 }
2721
2722                 /*
2723                  * Try to shortcut requests for double_buffer mode too.
2724                  * Since this mode runs through the device buffer cache
2725                  * only compatible buffer sizes (meaning those generated
2726                  * by normal filesystem buffers) are legal.
2727                  */
2728                 if (hammer_live_dedup == 0 && (bp->b_flags & B_PAGING) == 0) {
2729                         lwkt_gettoken(&hmp->fs_token);
2730                         error = hammer_io_indirect_read(hmp, nbio, NULL);
2731                         lwkt_reltoken(&hmp->fs_token);
2732                         return (error);
2733                 }
2734         }
2735
2736         /*
2737          * Well, that sucked.  Do it the hard way.  If all the stars are
2738          * aligned we may still be able to issue a direct-read.
2739          */
2740         lwkt_gettoken(&hmp->fs_token);
2741         hammer_simple_transaction(&trans, hmp);
2742         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2743
2744         /*
2745          * Key range (begin and end inclusive) to scan.  Note that the key's
2746          * stored in the actual records represent BASE+LEN, not BASE.  The
2747          * first record containing bio_offset will have a key > bio_offset.
2748          */
2749         cursor.key_beg.localization = ip->obj_localization +
2750                                       HAMMER_LOCALIZE_MISC;
2751         cursor.key_beg.obj_id = ip->obj_id;
2752         cursor.key_beg.create_tid = 0;
2753         cursor.key_beg.delete_tid = 0;
2754         cursor.key_beg.obj_type = 0;
2755         cursor.key_beg.key = bio->bio_offset + 1;
2756         cursor.asof = ip->obj_asof;
2757         cursor.flags |= HAMMER_CURSOR_ASOF;
2758
2759         cursor.key_end = cursor.key_beg;
2760         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2761 #if 0
2762         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
2763                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
2764                 cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
2765                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2766         } else
2767 #endif
2768         {
2769                 ran_end = bio->bio_offset + bp->b_bufsize;
2770                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2771                 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2772                 tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2773                 if (tmp64 < ran_end)
2774                         cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2775                 else
2776                         cursor.key_end.key = ran_end + MAXPHYS + 1;
2777         }
2778         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2779
2780         /*
2781          * Set NOSWAPCACHE for cursor data extraction if double buffering
2782          * is disabled or (if the file is not marked cacheable via chflags
2783          * and vm.swapcache_use_chflags is enabled).
2784          */
2785         if (hammer_double_buffer == 0 ||
2786             ((ap->a_vp->v_flag & VSWAPCACHE) == 0 &&
2787              vm_swapcache_use_chflags)) {
2788                 cursor.flags |= HAMMER_CURSOR_NOSWAPCACHE;
2789         }
2790
2791         error = hammer_ip_first(&cursor);
2792         boff = 0;
2793
2794         while (error == 0) {
2795                 /*
2796                  * Get the base file offset of the record.  The key for
2797                  * data records is (base + bytes) rather then (base).
2798                  */
2799                 base = &cursor.leaf->base;
2800                 rec_offset = base->key - cursor.leaf->data_len;
2801
2802                 /*
2803                  * Calculate the gap, if any, and zero-fill it.
2804                  *
2805                  * n is the offset of the start of the record verses our
2806                  * current seek offset in the bio.
2807                  */
2808                 n = (int)(rec_offset - (bio->bio_offset + boff));
2809                 if (n > 0) {
2810                         if (n > bp->b_bufsize - boff)
2811                                 n = bp->b_bufsize - boff;
2812                         bzero((char *)bp->b_data + boff, n);
2813                         boff += n;
2814                         n = 0;
2815                 }
2816
2817                 /*
2818                  * Calculate the data offset in the record and the number
2819                  * of bytes we can copy.
2820                  *
2821                  * There are two degenerate cases.  First, boff may already
2822                  * be at bp->b_bufsize.  Secondly, the data offset within
2823                  * the record may exceed the record's size.
2824                  */
2825                 roff = -n;
2826                 rec_offset += roff;
2827                 n = cursor.leaf->data_len - roff;
2828                 if (n <= 0) {
2829                         hdkprintf("bad n=%d roff=%d\n", n, roff);
2830                         n = 0;
2831                 } else if (n > bp->b_bufsize - boff) {
2832                         n = bp->b_bufsize - boff;
2833                 }
2834
2835                 /*
2836                  * Deal with cached truncations.  This cool bit of code
2837                  * allows truncate()/ftruncate() to avoid having to sync
2838                  * the file.
2839                  *
2840                  * If the frontend is truncated then all backend records are
2841                  * subject to the frontend's truncation.
2842                  *
2843                  * If the backend is truncated then backend records on-disk
2844                  * (but not in-memory) are subject to the backend's
2845                  * truncation.  In-memory records owned by the backend
2846                  * represent data written after the truncation point on the
2847                  * backend and must not be truncated.
2848                  *
2849                  * Truncate operations deal with frontend buffer cache
2850                  * buffers and frontend-owned in-memory records synchronously.
2851                  */
2852                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2853                         if (hammer_cursor_ondisk(&cursor)/* ||
2854                             cursor.iprec->flush_state == HAMMER_FST_FLUSH*/) {
2855                                 if (ip->trunc_off <= rec_offset)
2856                                         n = 0;
2857                                 else if (ip->trunc_off < rec_offset + n)
2858                                         n = (int)(ip->trunc_off - rec_offset);
2859                         }
2860                 }
2861                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2862                         if (hammer_cursor_ondisk(&cursor)) {
2863                                 if (ip->sync_trunc_off <= rec_offset)
2864                                         n = 0;
2865                                 else if (ip->sync_trunc_off < rec_offset + n)
2866                                         n = (int)(ip->sync_trunc_off - rec_offset);
2867                         }
2868                 }
2869
2870                 /*
2871                  * Try to issue a direct read into our bio if possible,
2872                  * otherwise resolve the element data into a hammer_buffer
2873                  * and copy.
2874                  *
2875                  * The buffer on-disk should be zerod past any real
2876                  * truncation point, but may not be for any synthesized
2877                  * truncation point from above.
2878                  *
2879                  * NOTE: disk_offset is only valid if the cursor data is
2880                  *       on-disk.
2881                  */
2882                 disk_offset = cursor.leaf->data_offset + roff;
2883                 isdedupable = (boff == 0 && n == bp->b_bufsize &&
2884                                hammer_cursor_ondisk(&cursor) &&
2885                                ((int)disk_offset & HAMMER_BUFMASK) == 0);
2886
2887                 if (isdedupable && hammer_double_buffer == 0) {
2888                         /*
2889                          * Direct read case
2890                          */
2891                         KKASSERT((disk_offset & HAMMER_OFF_ZONE_MASK) ==
2892                                  HAMMER_ZONE_LARGE_DATA);
2893                         nbio->bio_offset = disk_offset;
2894                         error = hammer_io_direct_read(hmp, nbio, cursor.leaf);
2895                         if (hammer_live_dedup && error == 0)
2896                                 hammer_dedup_cache_add(ip, cursor.leaf);
2897                         goto done;
2898                 } else if (isdedupable) {
2899                         /*
2900                          * Async I/O case for reading from backing store
2901                          * and copying the data to the filesystem buffer.
2902                          * live-dedup has to verify the data anyway if it
2903                          * gets a hit later so we can just add the entry
2904                          * now.
2905                          */
2906                         KKASSERT((disk_offset & HAMMER_OFF_ZONE_MASK) ==
2907                                  HAMMER_ZONE_LARGE_DATA);
2908                         nbio->bio_offset = disk_offset;
2909                         if (hammer_live_dedup)
2910                                 hammer_dedup_cache_add(ip, cursor.leaf);
2911                         error = hammer_io_indirect_read(hmp, nbio, cursor.leaf);
2912                         goto done;
2913                 } else if (n) {
2914                         error = hammer_ip_resolve_data(&cursor);
2915                         if (error == 0) {
2916                                 if (hammer_live_dedup && isdedupable)
2917                                         hammer_dedup_cache_add(ip, cursor.leaf);
2918                                 bcopy((char *)cursor.data + roff,
2919                                       (char *)bp->b_data + boff, n);
2920                         }
2921                 }
2922                 if (error)
2923                         break;
2924
2925                 /*
2926                  * We have to be sure that the only elements added to the
2927                  * dedup cache are those which are already on-media.
2928                  */
2929                 if (hammer_live_dedup && hammer_cursor_ondisk(&cursor))
2930                         hammer_dedup_cache_add(ip, cursor.leaf);
2931
2932                 /*
2933                  * Iterate until we have filled the request.
2934                  */
2935                 boff += n;
2936                 if (boff == bp->b_bufsize)
2937                         break;
2938                 error = hammer_ip_next(&cursor);
2939         }
2940
2941         /*
2942          * There may have been a gap after the last record
2943          */
2944         if (error == ENOENT)
2945                 error = 0;
2946         if (error == 0 && boff != bp->b_bufsize) {
2947                 KKASSERT(boff < bp->b_bufsize);
2948                 bzero((char *)bp->b_data + boff, bp->b_bufsize - boff);
2949                 /* boff = bp->b_bufsize; */
2950         }
2951
2952         /*
2953          * Disallow swapcache operation on the vnode buffer if double
2954          * buffering is enabled, the swapcache will get the data via
2955          * the block device buffer.
2956          */
2957         if (hammer_double_buffer)
2958                 bp->b_flags |= B_NOTMETA;
2959
2960         /*
2961          * Cleanup
2962          */
2963         bp->b_resid = 0;
2964         bp->b_error = error;
2965         if (error)
2966                 bp->b_flags |= B_ERROR;
2967         biodone(ap->a_bio);
2968
2969 done:
2970         /*
2971          * Cache the b-tree node for the last data read in cache[1].
2972          *
2973          * If we hit the file EOF then also cache the node in the
2974          * governing directory's cache[3], it will be used to initialize
2975          * the new inode's cache[1] for any inodes looked up via the directory.
2976          *
2977          * This doesn't reduce disk accesses since the B-Tree chain is
2978          * likely cached, but it does reduce cpu overhead when looking
2979          * up file offsets for cpdup/tar/cpio style iterations.
2980          */
2981         if (cursor.node)
2982                 hammer_cache_node(&ip->cache[1], cursor.node);
2983         if (ran_end >= ip->ino_data.size) {
2984                 dip = hammer_find_inode(&trans, ip->ino_data.parent_obj_id,
2985                                         ip->obj_asof, ip->obj_localization);
2986                 if (dip) {
2987                         hammer_cache_node(&dip->cache[3], cursor.node);
2988                         hammer_rel_inode(dip, 0);
2989                 }
2990         }
2991         hammer_done_cursor(&cursor);
2992         hammer_done_transaction(&trans);
2993         lwkt_reltoken(&hmp->fs_token);
2994         return(error);
2995 }
2996
2997 /*
2998  * BMAP operation - used to support cluster_read() only.
2999  *
3000  * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
3001  *
3002  * This routine may return EOPNOTSUPP if the opration is not supported for
3003  * the specified offset.  The contents of the pointer arguments do not
3004  * need to be initialized in that case.
3005  *
3006  * If a disk address is available and properly aligned return 0 with
3007  * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
3008  * to the run-length relative to that offset.  Callers may assume that
3009  * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
3010  * large, so return EOPNOTSUPP if it is not sufficiently large.
3011  */
3012 static
3013 int
3014 hammer_vop_bmap(struct vop_bmap_args *ap)
3015 {
3016         struct hammer_transaction trans;
3017         struct hammer_inode *ip;
3018         hammer_mount_t hmp;
3019         struct hammer_cursor cursor;
3020         hammer_base_elm_t base;
3021         int64_t rec_offset;
3022         int64_t ran_end;
3023         int64_t tmp64;
3024         int64_t base_offset;
3025         int64_t base_disk_offset;
3026         int64_t last_offset;
3027         hammer_off_t last_disk_offset;
3028         hammer_off_t disk_offset;
3029         int     rec_len;
3030         int     error;
3031         int     blksize;
3032
3033         ++hammer_stats_file_iopsr;
3034         ip = ap->a_vp->v_data;
3035         hmp = ip->hmp;
3036
3037         /*
3038          * We can only BMAP regular files.  We can't BMAP database files,
3039          * directories, etc.
3040          */
3041         if (ip->ino_data.obj_type != HAMMER_OBJTYPE_REGFILE)
3042                 return(EOPNOTSUPP);
3043
3044         /*
3045          * bmap is typically called with runp/runb both NULL when used
3046          * for writing.  We do not support BMAP for writing atm.
3047          */
3048         if (ap->a_cmd != BUF_CMD_READ)
3049                 return(EOPNOTSUPP);
3050
3051         /*
3052          * Scan the B-Tree to acquire blockmap addresses, then translate
3053          * to raw addresses.
3054          */
3055         lwkt_gettoken(&hmp->fs_token);
3056         hammer_simple_transaction(&trans, hmp);
3057 #if 0
3058         hkprintf("bmap_beg %016llx ip->cache %p\n",
3059                 (long long)ap->a_loffset, ip->cache[1]);
3060 #endif
3061         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
3062
3063         /*
3064          * Key range (begin and end inclusive) to scan.  Note that the key's
3065          * stored in the actual records represent BASE+LEN, not BASE.  The
3066          * first record containing bio_offset will have a key > bio_offset.
3067          */
3068         cursor.key_beg.localization = ip->obj_localization +
3069                                       HAMMER_LOCALIZE_MISC;
3070         cursor.key_beg.obj_id = ip->obj_id;
3071         cursor.key_beg.create_tid = 0;
3072         cursor.key_beg.delete_tid = 0;
3073         cursor.key_beg.obj_type = 0;
3074         if (ap->a_runb)
3075                 cursor.key_beg.key = ap->a_loffset - MAXPHYS + 1;
3076         else
3077                 cursor.key_beg.key = ap->a_loffset + 1;
3078         if (cursor.key_beg.key < 0)
3079                 cursor.key_beg.key = 0;
3080         cursor.asof = ip->obj_asof;
3081         cursor.flags |= HAMMER_CURSOR_ASOF;
3082
3083         cursor.key_end = cursor.key_beg;
3084         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
3085
3086         ran_end = ap->a_loffset + MAXPHYS;
3087         cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
3088         cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
3089         tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
3090         if (tmp64 < ran_end)
3091                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
3092         else
3093                 cursor.key_end.key = ran_end + MAXPHYS + 1;
3094
3095         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
3096
3097         error = hammer_ip_first(&cursor);
3098         base_offset = last_offset = 0;
3099         base_disk_offset = last_disk_offset = 0;
3100
3101         while (error == 0) {
3102                 /*
3103                  * Get the base file offset of the record.  The key for
3104                  * data records is (base + bytes) rather then (base).
3105                  *
3106                  * NOTE: rec_offset + rec_len may exceed the end-of-file.
3107                  * The extra bytes should be zero on-disk and the BMAP op
3108                  * should still be ok.
3109                  */
3110                 base = &cursor.leaf->base;
3111                 rec_offset = base->key - cursor.leaf->data_len;
3112                 rec_len    = cursor.leaf->data_len;
3113
3114                 /*
3115                  * Incorporate any cached truncation.
3116                  *
3117                  * NOTE: Modifications to rec_len based on synthesized
3118                  * truncation points remove the guarantee that any extended
3119                  * data on disk is zero (since the truncations may not have
3120                  * taken place on-media yet).
3121                  */
3122                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
3123                         if (hammer_cursor_ondisk(&cursor) ||
3124                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
3125                                 if (ip->trunc_off <= rec_offset)
3126                                         rec_len = 0;
3127                                 else if (ip->trunc_off < rec_offset + rec_len)
3128                                         rec_len = (int)(ip->trunc_off - rec_offset);
3129                         }
3130                 }
3131                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
3132                         if (hammer_cursor_ondisk(&cursor)) {
3133                                 if (ip->sync_trunc_off <= rec_offset)
3134                                         rec_len = 0;
3135                                 else if (ip->sync_trunc_off < rec_offset + rec_len)
3136                                         rec_len = (int)(ip->sync_trunc_off - rec_offset);
3137                         }
3138                 }
3139
3140                 /*
3141                  * Accumulate information.  If we have hit a discontiguous
3142                  * block reset base_offset unless we are already beyond the
3143                  * requested offset.  If we are, that's it, we stop.
3144                  */
3145                 if (error)
3146                         break;
3147                 if (hammer_cursor_ondisk(&cursor)) {
3148                         disk_offset = cursor.leaf->data_offset;
3149                         if (rec_offset != last_offset ||
3150                             disk_offset != last_disk_offset) {
3151                                 if (rec_offset > ap->a_loffset)
3152                                         break;
3153                                 base_offset = rec_offset;
3154                                 base_disk_offset = disk_offset;
3155                         }
3156                         last_offset = rec_offset + rec_len;
3157                         last_disk_offset = disk_offset + rec_len;
3158
3159                         if (hammer_live_dedup)
3160                                 hammer_dedup_cache_add(ip, cursor.leaf);
3161                 }
3162
3163                 error = hammer_ip_next(&cursor);
3164         }
3165
3166 #if 0
3167         hkprintf("BMAP %016llx:  %016llx - %016llx\n",
3168                 (long long)ap->a_loffset,
3169                 (long long)base_offset,
3170                 (long long)last_offset);
3171         hkprintf("BMAP %16s:  %016llx - %016llx\n", "",
3172                 (long long)base_disk_offset,
3173                 (long long)last_disk_offset);
3174 #endif
3175
3176         if (cursor.node) {
3177                 hammer_cache_node(&ip->cache[1], cursor.node);
3178 #if 0
3179                 hkprintf("bmap_end2 %016llx ip->cache %p\n",
3180                         (long long)ap->a_loffset, ip->cache[1]);
3181 #endif
3182         }
3183         hammer_done_cursor(&cursor);
3184         hammer_done_transaction(&trans);
3185         lwkt_reltoken(&hmp->fs_token);
3186
3187         /*
3188          * If we couldn't find any records or the records we did find were
3189          * all behind the requested offset, return failure.  A forward
3190          * truncation can leave a hole w/ no on-disk records.
3191          */
3192         if (last_offset == 0 || last_offset < ap->a_loffset)
3193                 return (EOPNOTSUPP);
3194
3195         /*
3196          * Figure out the block size at the requested offset and adjust
3197          * our limits so the cluster_read() does not create inappropriately
3198          * sized buffer cache buffers.
3199          */
3200         blksize = hammer_blocksize(ap->a_loffset);
3201         if (hammer_blocksize(base_offset) != blksize) {
3202                 base_offset = hammer_blockdemarc(base_offset, ap->a_loffset);
3203         }
3204         if (last_offset != ap->a_loffset &&
3205             hammer_blocksize(last_offset - 1) != blksize) {
3206                 last_offset = hammer_blockdemarc(ap->a_loffset,
3207                                                  last_offset - 1);
3208         }
3209
3210         /*
3211          * Returning EOPNOTSUPP simply prevents the direct-IO optimization
3212          * from occuring.
3213          */
3214         disk_offset = base_disk_offset + (ap->a_loffset - base_offset);
3215
3216         if ((disk_offset & HAMMER_OFF_ZONE_MASK) != HAMMER_ZONE_LARGE_DATA) {
3217                 /*
3218                  * Only large-data zones can be direct-IOd
3219                  */
3220                 error = EOPNOTSUPP;
3221         } else if ((disk_offset & HAMMER_BUFMASK) ||
3222                    (last_offset - ap->a_loffset) < blksize) {
3223                 /*
3224                  * doffsetp is not aligned or the forward run size does
3225                  * not cover a whole buffer, disallow the direct I/O.
3226                  */
3227                 error = EOPNOTSUPP;
3228         } else {
3229                 /*
3230                  * We're good.
3231                  */
3232                 *ap->a_doffsetp = disk_offset;
3233                 if (ap->a_runb) {
3234                         *ap->a_runb = ap->a_loffset - base_offset;
3235                         KKASSERT(*ap->a_runb >= 0);
3236                 }
3237                 if (ap->a_runp) {
3238                         *ap->a_runp = last_offset - ap->a_loffset;
3239                         KKASSERT(*ap->a_runp >= 0);
3240                 }
3241                 error = 0;
3242         }
3243         return(error);
3244 }
3245
3246 /*
3247  * Write to a regular file.   Because this is a strategy call the OS is
3248  * trying to actually get data onto the media.
3249  */
3250 static
3251 int
3252 hammer_vop_strategy_write(struct vop_strategy_args *ap)
3253 {
3254         hammer_record_t record;
3255         hammer_mount_t hmp;
3256         hammer_inode_t ip;
3257         struct bio *bio;
3258         struct buf *bp;
3259         int blksize __debugvar;
3260         int bytes;
3261         int error;
3262
3263         bio = ap->a_bio;
3264         bp = bio->bio_buf;
3265         ip = ap->a_vp->v_data;
3266         hmp = ip->hmp;
3267
3268         blksize = hammer_blocksize(bio->bio_offset);
3269         KKASSERT(bp->b_bufsize == blksize);
3270
3271         if (ip->flags & HAMMER_INODE_RO) {
3272                 bp->b_error = EROFS;
3273                 bp->b_flags |= B_ERROR;
3274                 biodone(ap->a_bio);
3275                 return(EROFS);
3276         }
3277
3278         lwkt_gettoken(&hmp->fs_token);
3279
3280         /*
3281          * Disallow swapcache operation on the vnode buffer if double
3282          * buffering is enabled, the swapcache will get the data via
3283          * the block device buffer.
3284          */
3285         if (hammer_double_buffer)
3286                 bp->b_flags |= B_NOTMETA;
3287
3288         /*
3289          * Interlock with inode destruction (no in-kernel or directory
3290          * topology visibility).  If we queue new IO while trying to
3291          * destroy the inode we can deadlock the vtrunc call in
3292          * hammer_inode_unloadable_check().
3293          *
3294          * Besides, there's no point flushing a bp associated with an
3295          * inode that is being destroyed on-media and has no kernel
3296          * references.
3297          */
3298         if ((ip->flags | ip->sync_flags) &
3299             (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) {
3300                 bp->b_resid = 0;
3301                 biodone(ap->a_bio);
3302                 lwkt_reltoken(&hmp->fs_token);
3303                 return(0);
3304         }
3305
3306         /*
3307          * Reserve space and issue a direct-write from the front-end.
3308          * NOTE: The direct_io code will hammer_bread/bcopy smaller
3309          * allocations.
3310          *
3311          * An in-memory record will be installed to reference the storage
3312          * until the flusher can get to it.
3313          *
3314          * Since we own the high level bio the front-end will not try to
3315          * do a direct-read until the write completes.
3316          *
3317          * NOTE: The only time we do not reserve a full-sized buffers
3318          * worth of data is if the file is small.  We do not try to
3319          * allocate a fragment (from the small-data zone) at the end of
3320          * an otherwise large file as this can lead to wildly separated
3321          * data.
3322          */
3323         KKASSERT((bio->bio_offset & HAMMER_BUFMASK) == 0);
3324         KKASSERT(bio->bio_offset < ip->ino_data.size);
3325         if (bio->bio_offset || ip->ino_data.size > HAMMER_HBUFSIZE)
3326                 bytes = bp->b_bufsize;
3327         else
3328                 bytes = ((int)ip->ino_data.size + 15) & ~15;
3329
3330         record = hammer_ip_add_bulk(ip, bio->bio_offset, bp->b_data,
3331                                     bytes, &error);
3332
3333         /*
3334          * B_VFSFLAG1 indicates that a REDO_WRITE entry was generated
3335          * in hammer_vop_write().  We must flag the record so the proper
3336          * REDO_TERM_WRITE entry is generated during the flush.
3337          */
3338         if (record) {
3339                 if (bp->b_flags & B_VFSFLAG1) {
3340                         record->flags |= HAMMER_RECF_REDO;
3341                         bp->b_flags &= ~B_VFSFLAG1;
3342                 }
3343                 if (record->flags & HAMMER_RECF_DEDUPED) {
3344                         bp->b_resid = 0;
3345                         hammer_ip_replace_bulk(hmp, record);
3346                         biodone(ap->a_bio);
3347                 } else {
3348                         hammer_io_direct_write(hmp, bio, record);
3349                 }
3350                 if (ip->rsv_recs > 1 && hmp->rsv_recs > hammer_limit_recs)
3351                         hammer_flush_inode(ip, 0);
3352         } else {
3353                 bp->b_bio2.bio_offset = NOOFFSET;
3354                 bp->b_error = error;
3355                 bp->b_flags |= B_ERROR;
3356                 biodone(ap->a_bio);
3357         }
3358         lwkt_reltoken(&hmp->fs_token);
3359         return(error);
3360 }
3361
3362 /*
3363  * dounlink - disconnect a directory entry
3364  *
3365  * XXX whiteout support not really in yet
3366  */
3367 static int
3368 hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
3369                 struct vnode *dvp, struct ucred *cred,
3370                 int flags, int isdir)
3371 {
3372         struct namecache *ncp;
3373         hammer_inode_t dip;
3374         hammer_inode_t ip;
3375         hammer_mount_t hmp;
3376         struct hammer_cursor cursor;
3377         int64_t namekey;
3378         u_int32_t max_iterations;
3379         int nlen, error;
3380
3381         /*
3382          * Calculate the namekey and setup the key range for the scan.  This
3383          * works kinda like a chained hash table where the lower 32 bits
3384          * of the namekey synthesize the chain.
3385          *
3386          * The key range is inclusive of both key_beg and key_end.
3387          */
3388         dip = VTOI(dvp);
3389         ncp = nch->ncp;
3390         hmp = dip->hmp;
3391
3392         if (dip->flags & HAMMER_INODE_RO)
3393                 return (EROFS);
3394
3395         namekey = hammer_directory_namekey(dip, ncp->nc_name, ncp->nc_nlen,
3396                                            &max_iterations);
3397 retry:
3398         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
3399         cursor.key_beg.localization = dip->obj_localization +
3400                                       hammer_dir_localization(dip);
3401         cursor.key_beg.obj_id = dip->obj_id;
3402         cursor.key_beg.key = namekey;
3403         cursor.key_beg.create_tid = 0;
3404         cursor.key_beg.delete_tid = 0;
3405         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
3406         cursor.key_beg.obj_type = 0;
3407
3408         cursor.key_end = cursor.key_beg;
3409         cursor.key_end.key += max_iterations;
3410         cursor.asof = dip->obj_asof;
3411         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
3412
3413         /*
3414          * Scan all matching records (the chain), locate the one matching
3415          * the requested path component.  info->last_error contains the
3416          * error code on search termination and could be 0, ENOENT, or
3417          * something else.
3418          *
3419          * The hammer_ip_*() functions merge in-memory records with on-disk
3420          * records for the purposes of the search.
3421          */
3422         error = hammer_ip_first(&cursor);
3423
3424         while (error == 0) {
3425                 error = hammer_ip_resolve_data(&cursor);
3426                 if (error)
3427                         break;
3428                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
3429                 KKASSERT(nlen > 0);
3430                 if (ncp->nc_nlen == nlen &&
3431                     bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
3432                         break;
3433                 }
3434                 error = hammer_ip_next(&cursor);
3435         }
3436
3437         /*
3438          * If all is ok we have to get the inode so we can adjust nlinks.
3439          * To avoid a deadlock with the flusher we must release the inode
3440          * lock on the directory when acquiring the inode for the entry.
3441          *
3442          * If the target is a directory, it must be empty.
3443          */
3444         if (error == 0) {
3445                 hammer_unlock(&cursor.ip->lock);
3446                 ip = hammer_get_inode(trans, dip, cursor.data->entry.obj_id,
3447                                       hmp->asof,
3448                                       cursor.data->entry.localization,
3449                                       0, &error);
3450                 hammer_lock_sh(&cursor.ip->lock);
3451                 if (error == ENOENT) {
3452                         hkprintf("WARNING: Removing dirent w/missing inode "
3453                                 "\"%s\"\n"
3454                                 "\tobj_id = %016llx\n",
3455                                 ncp->nc_name,
3456                                 (long long)cursor.data->entry.obj_id);
3457                         error = 0;
3458                 }
3459
3460                 /*
3461                  * If isdir >= 0 we validate that the entry is or is not a
3462                  * directory.  If isdir < 0 we don't care.
3463                  */
3464                 if (error == 0 && isdir >= 0 && ip) {
3465                         if (isdir &&
3466                             ip->ino_data.obj_type != HAMMER_OBJTYPE_DIRECTORY) {
3467                                 error = ENOTDIR;
3468                         } else if (isdir == 0 &&
3469                             ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
3470                                 error = EISDIR;
3471                         }
3472                 }
3473
3474                 /*
3475                  * If we are trying to remove a directory the directory must
3476                  * be empty.
3477                  *
3478                  * The check directory code can loop and deadlock/retry.  Our
3479                  * own cursor's node locks must be released to avoid a 3-way
3480                  * deadlock with the flusher if the check directory code
3481                  * blocks.
3482                  *
3483                  * If any changes whatsoever have been made to the cursor
3484                  * set EDEADLK and retry.
3485                  *
3486                  * WARNING: See warnings in hammer_unlock_cursor()
3487                  *          function.
3488                  */
3489                 if (error == 0 && ip && ip->ino_data.obj_type ==
3490                                         HAMMER_OBJTYPE_DIRECTORY) {
3491                         hammer_unlock_cursor(&cursor);
3492                         error = hammer_ip_check_directory_empty(trans, ip);
3493                         hammer_lock_cursor(&cursor);
3494                         if (cursor.flags & HAMMER_CURSOR_RETEST) {
3495                                 hkprintf("Warning: avoided deadlock "
3496                                         "on rmdir '%s'\n",
3497                                         ncp->nc_name);
3498                                 error = EDEADLK;
3499                         }
3500                 }
3501
3502                 /*
3503                  * Delete the directory entry.
3504                  *
3505                  * WARNING: hammer_ip_del_directory() may have to terminate
3506                  * the cursor to avoid a deadlock.  It is ok to call
3507                  * hammer_done_cursor() twice.
3508                  */
3509                 if (error == 0) {
3510                         error = hammer_ip_del_directory(trans, &cursor,
3511                                                         dip, ip);
3512                 }
3513                 hammer_done_cursor(&cursor);
3514                 if (error == 0) {
3515                         /*
3516                          * Tell the namecache that we are now unlinked.
3517                          */
3518                         cache_unlink(nch);
3519
3520                         /*
3521                          * NOTE: ip->vp, if non-NULL, cannot be directly
3522                          *       referenced without formally acquiring the
3523                          *       vp since the vp might have zero refs on it,
3524                          *       or in the middle of a reclaim, etc.
3525                          *
3526                          * NOTE: The cache_setunresolved() can rip the vp
3527                          *       out from under us since the vp may not have
3528                          *       any refs, in which case ip->vp will be NULL
3529                          *       from the outset.
3530                          */
3531                         while (ip && ip->vp) {
3532                                 struct vnode *vp;
3533
3534                                 error = hammer_get_vnode(ip, &vp);
3535                                 if (error == 0 && vp) {
3536                                         vn_unlock(vp);
3537                                         hammer_knote(ip->vp, NOTE_DELETE);
3538 #if 0
3539                                         /*
3540                                          * Don't do this, it can deadlock
3541                                          * on concurrent rm's of hardlinks.
3542                                          * Shouldn't be needed any more.
3543                                          */
3544                                         cache_inval_vp(ip->vp, CINV_DESTROY);
3545 #endif
3546                                         vrele(vp);
3547                                         break;
3548                                 }
3549                                 hdkprintf("ip/vp race1 avoided\n");
3550                         }
3551                 }
3552                 if (ip)
3553                         hammer_rel_inode(ip, 0);
3554         } else {
3555                 hammer_done_cursor(&cursor);
3556         }
3557         if (error == EDEADLK)
3558                 goto retry;
3559
3560         return (error);
3561 }
3562
3563 /************************************************************************
3564  *                          FIFO AND SPECFS OPS                         *
3565  ************************************************************************
3566  *
3567  */
3568 static int
3569 hammer_vop_fifoclose (struct vop_close_args *ap)
3570 {
3571         /* XXX update itimes */
3572         return (VOCALL(&fifo_vnode_vops, &ap->a_head));
3573 }
3574
3575 static int
3576 hammer_vop_fiforead (struct vop_read_args *ap)
3577 {
3578         int error;
3579
3580         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3581         /* XXX update access time */
3582         return (error);
3583 }
3584
3585 static int
3586 hammer_vop_fifowrite (struct vop_write_args *ap)
3587 {
3588         int error;
3589
3590         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3591         /* XXX update access time */
3592         return (error);
3593 }
3594
3595 static
3596 int
3597 hammer_vop_fifokqfilter(struct vop_kqfilter_args *ap)
3598 {
3599         int error;
3600
3601         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3602         if (error)
3603                 error = hammer_vop_kqfilter(ap);
3604         return(error);
3605 }
3606
3607 /************************************************************************
3608  *                          KQFILTER OPS                                *
3609  ************************************************************************
3610  *
3611  */
3612 static void filt_hammerdetach(struct knote *kn);
3613 static int filt_hammerread(struct knote *kn, long hint);
3614 static int filt_hammerwrite(struct knote *kn, long hint);
3615 static int filt_hammervnode(struct knote *kn, long hint);
3616
3617 static struct filterops hammerread_filtops =
3618         { FILTEROP_ISFD | FILTEROP_MPSAFE,
3619           NULL, filt_hammerdetach, filt_hammerread };
3620 static struct filterops hammerwrite_filtops =
3621         { FILTEROP_ISFD | FILTEROP_MPSAFE,
3622           NULL, filt_hammerdetach, filt_hammerwrite };
3623 static struct filterops hammervnode_filtops =
3624         { FILTEROP_ISFD | FILTEROP_MPSAFE,
3625           NULL, filt_hammerdetach, filt_hammervnode };
3626
3627 static
3628 int
3629 hammer_vop_kqfilter(struct vop_kqfilter_args *ap)
3630 {
3631         struct vnode *vp = ap->a_vp;
3632         struct knote *kn = ap->a_kn;
3633
3634         switch (kn->kn_filter) {
3635         case EVFILT_READ:
3636                 kn->kn_fop = &hammerread_filtops;
3637                 break;
3638         case EVFILT_WRITE:
3639                 kn->kn_fop = &hammerwrite_filtops;
3640                 break;
3641         case EVFILT_VNODE:
3642                 kn->kn_fop = &hammervnode_filtops;
3643                 break;
3644         default:
3645                 return (EOPNOTSUPP);
3646         }
3647
3648         kn->kn_hook = (caddr_t)vp;
3649
3650         knote_insert(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3651
3652         return(0);
3653 }
3654
3655 static void
3656 filt_hammerdetach(struct knote *kn)
3657 {
3658         struct vnode *vp = (void *)kn->kn_hook;
3659
3660         knote_remove(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3661 }
3662
3663 static int
3664 filt_hammerread(struct knote *kn, long hint)
3665 {
3666         struct vnode *vp = (void *)kn->kn_hook;
3667         hammer_inode_t ip = VTOI(vp);
3668         hammer_mount_t hmp = ip->hmp;
3669         off_t off;
3670
3671         if (hint == NOTE_REVOKE) {
3672                 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
3673                 return(1);
3674         }
3675         lwkt_gettoken(&hmp->fs_token);  /* XXX use per-ip-token */
3676         off = ip->ino_data.size - kn->kn_fp->f_offset;
3677         kn->kn_data = (off < INTPTR_MAX) ? off : INTPTR_MAX;
3678         lwkt_reltoken(&hmp->fs_token);
3679         if (kn->kn_sfflags & NOTE_OLDAPI)
3680                 return(1);
3681         return (kn->kn_data != 0);
3682 }
3683
3684 static int
3685 filt_hammerwrite(struct knote *kn, long hint)
3686 {
3687         if (hint == NOTE_REVOKE)
3688                 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
3689         kn->kn_data = 0;
3690         return (1);
3691 }
3692
3693 static int
3694 filt_hammervnode(struct knote *kn, long hint)
3695 {
3696         if (kn->kn_sfflags & hint)
3697                 kn->kn_fflags |= hint;
3698         if (hint == NOTE_REVOKE) {
3699                 kn->kn_flags |= (EV_EOF | EV_NODATA);
3700                 return (1);
3701         }
3702         return (kn->kn_fflags != 0);
3703 }
3704