sys/vfs/hammer/hammer_vnops.c

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