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                         kprintf("error %d\n", error);
 366                         brelse(bp);
 367                         break;
 368                 }
 369 skip:
 370
 371                 /* bp->b_flags |= B_CLUSTEROK; temporarily disabled */
 372                 n = blksize - offset;
 373                 if (n > uio->uio_resid)
 374                         n = uio->uio_resid;
 375                 if (n > ip->ino_data.size - uio->uio_offset)
 376                         n = (int)(ip->ino_data.size - uio->uio_offset);
 377                 error = uiomove((char *)bp->b_data + offset, n, uio);
 378
 379                 /* data has a lower priority then meta-data */
 380                 bp->b_flags |= B_AGE;
 381                 bqrelse(bp);
 382                 if (error)
 383                         break;
 384                 hammer_stats_file_read += n;
 385         }
 386
 387         /*
 388          * XXX only update the atime if we had to get the MP lock.
 389          * XXX hack hack hack, fixme.
 390          */
 391         if (got_mplock) {
 392                 if ((ip->flags & HAMMER_INODE_RO) == 0 &&
 393                     (ip->hmp->mp->mnt_flag & MNT_NOATIME) == 0) {
 394                         ip->ino_data.atime = trans.time;
 395                         hammer_modify_inode(ip, HAMMER_INODE_ATIME);
 396                 }
 397                 hammer_done_transaction(&trans);
 398                 if (got_mplock > 0)
 399                         rel_mplock();
 400         }
 401         return (error);
 402 }
 403
 404 /*
 405  * hammer_vop_write { vp, uio, ioflag, cred }
 406  */
 407 static
 408 int
 409 hammer_vop_write(struct vop_write_args *ap)
 410 {
 411         struct hammer_transaction trans;
 412         struct hammer_inode *ip;
 413         hammer_mount_t hmp;
 414         struct uio *uio;
 415         int offset;
 416         off_t base_offset;
 417         struct buf *bp;
 418         int kflags;
 419         int error;
 420         int n;
 421         int flags;
 422         int seqcount;
 423         int bigwrite;
 424
 425         if (ap->a_vp->v_type != VREG)
 426                 return (EINVAL);
 427         ip = VTOI(ap->a_vp);
 428         hmp = ip->hmp;
 429         error = 0;
 430         kflags = 0;
 431         seqcount = ap->a_ioflag >> 16;
 432
 433         if (ip->flags & HAMMER_INODE_RO)
 434                 return (EROFS);
 435
 436         /*
 437          * Create a transaction to cover the operations we perform.
 438          */
 439         hammer_start_transaction(&trans, hmp);
 440         uio = ap->a_uio;
 441
 442         /*
 443          * Check append mode
 444          */
 445         if (ap->a_ioflag & IO_APPEND)
 446                 uio->uio_offset = ip->ino_data.size;
 447
 448         /*
 449          * Check for illegal write offsets.  Valid range is 0...2^63-1.
 450          *
 451          * NOTE: the base_off assignment is required to work around what
 452          * I consider to be a GCC-4 optimization bug.
 453          */
 454         if (uio->uio_offset < 0) {
 455                 hammer_done_transaction(&trans);
 456                 return (EFBIG);
 457         }
 458         base_offset = uio->uio_offset + uio->uio_resid; /* work around gcc-4 */
 459         if (uio->uio_resid > 0 && base_offset <= uio->uio_offset) {
 460                 hammer_done_transaction(&trans);
 461                 return (EFBIG);
 462         }
 463
 464         /*
 465          * If reading or writing a huge amount of data we have to break
 466          * atomicy and allow the operation to be interrupted by a signal
 467          * or it can DOS the machine.
 468          */
 469         bigwrite = (uio->uio_resid > 100 * 1024 * 1024);
 470
 471         /*
 472          * Access the data typically in HAMMER_BUFSIZE blocks via the
 473          * buffer cache, but HAMMER may use a variable block size based
 474          * on the offset.
 475          */
 476         while (uio->uio_resid > 0) {
 477                 int fixsize = 0;
 478                 int blksize;
 479                 int blkmask;
 480
 481                 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_WRITE)) != 0)
 482                         break;
 483                 if (bigwrite && (error = hammer_signal_check(hmp)) != 0)
 484                         break;
 485
 486                 blksize = hammer_blocksize(uio->uio_offset);
 487
 488                 /*
 489                  * Do not allow HAMMER to blow out the buffer cache.  Very
 490                  * large UIOs can lockout other processes due to bwillwrite()
 491                  * mechanics.
 492                  *
 493                  * The hammer inode is not locked during these operations.
 494                  * The vnode is locked which can interfere with the pageout
 495                  * daemon for non-UIO_NOCOPY writes but should not interfere
 496                  * with the buffer cache.  Even so, we cannot afford to
 497                  * allow the pageout daemon to build up too many dirty buffer
 498                  * cache buffers.
 499                  *
 500                  * Only call this if we aren't being recursively called from
 501                  * a virtual disk device (vn), else we may deadlock.
 502                  */
 503                 if ((ap->a_ioflag & IO_RECURSE) == 0)
 504                         bwillwrite(blksize);
 505
 506                 /*
 507                  * Control the number of pending records associated with
 508                  * this inode.  If too many have accumulated start a
 509                  * flush.  Try to maintain a pipeline with the flusher.
 510                  */
 511                 if (ip->rsv_recs >= hammer_limit_inode_recs) {
 512                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 513                 }
 514                 if (ip->rsv_recs >= hammer_limit_inode_recs * 2) {
 515                         while (ip->rsv_recs >= hammer_limit_inode_recs) {
 516                                 tsleep(&ip->rsv_recs, 0, "hmrwww", hz);
 517                         }
 518                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 519                 }
 520
 521 #if 0
 522                 /*
 523                  * Do not allow HAMMER to blow out system memory by
 524                  * accumulating too many records.   Records are so well
 525                  * decoupled from the buffer cache that it is possible
 526                  * for userland to push data out to the media via
 527                  * direct-write, but build up the records queued to the
 528                  * backend faster then the backend can flush them out.
 529                  * HAMMER has hit its write limit but the frontend has
 530                  * no pushback to slow it down.
 531                  */
 532                 if (hmp->rsv_recs > hammer_limit_recs / 2) {
 533                         /*
 534                          * Get the inode on the flush list
 535                          */
 536                         if (ip->rsv_recs >= 64)
 537                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 538                         else if (ip->rsv_recs >= 16)
 539                                 hammer_flush_inode(ip, 0);
 540
 541                         /*
 542                          * Keep the flusher going if the system keeps
 543                          * queueing records.
 544                          */
 545                         delta = hmp->count_newrecords -
 546                                 hmp->last_newrecords;
 547                         if (delta < 0 || delta > hammer_limit_recs / 2) {
 548                                 hmp->last_newrecords = hmp->count_newrecords;
 549                                 hammer_sync_hmp(hmp, MNT_NOWAIT);
 550                         }
 551
 552                         /*
 553                          * If we have gotten behind start slowing
 554                          * down the writers.
 555                          */
 556                         delta = (hmp->rsv_recs - hammer_limit_recs) *
 557                                 hz / hammer_limit_recs;
 558                         if (delta > 0)
 559                                 tsleep(&trans, 0, "hmrslo", delta);
 560                 }
 561 #endif
 562
 563                 /*
 564                  * Calculate the blocksize at the current offset and figure
 565                  * out how much we can actually write.
 566                  */
 567                 blkmask = blksize - 1;
 568                 offset = (int)uio->uio_offset & blkmask;
 569                 base_offset = uio->uio_offset & ~(int64_t)blkmask;
 570                 n = blksize - offset;
 571                 if (n > uio->uio_resid)
 572                         n = uio->uio_resid;
 573                 if (uio->uio_offset + n > ip->ino_data.size) {
 574                         vnode_pager_setsize(ap->a_vp, uio->uio_offset + n);
 575                         fixsize = 1;
 576                         kflags |= NOTE_EXTEND;
 577                 }
 578
 579                 if (uio->uio_segflg == UIO_NOCOPY) {
 580                         /*
 581                          * Issuing a write with the same data backing the
 582                          * buffer.  Instantiate the buffer to collect the
 583                          * backing vm pages, then read-in any missing bits.
 584                          *
 585                          * This case is used by vop_stdputpages().
 586                          */
 587                         bp = getblk(ap->a_vp, base_offset,
 588                                     blksize, GETBLK_BHEAVY, 0);
 589                         if ((bp->b_flags & B_CACHE) == 0) {
 590                                 bqrelse(bp);
 591                                 error = bread(ap->a_vp, base_offset,
 592                                               blksize, &bp);
 593                         }
 594                 } else if (offset == 0 && uio->uio_resid >= blksize) {
 595                         /*
 596                          * Even though we are entirely overwriting the buffer
 597                          * we may still have to zero it out to avoid a
 598                          * mmap/write visibility issue.
 599                          */
 600                         bp = getblk(ap->a_vp, base_offset, blksize, GETBLK_BHEAVY, 0);
 601                         if ((bp->b_flags & B_CACHE) == 0)
 602                                 vfs_bio_clrbuf(bp);
 603                 } else if (base_offset >= ip->ino_data.size) {
 604                         /*
 605                          * If the base offset of the buffer is beyond the
 606                          * file EOF, we don't have to issue a read.
 607                          */
 608                         bp = getblk(ap->a_vp, base_offset,
 609                                     blksize, GETBLK_BHEAVY, 0);
 610                         vfs_bio_clrbuf(bp);
 611                 } else {
 612                         /*
 613                          * Partial overwrite, read in any missing bits then
 614                          * replace the portion being written.
 615                          */
 616                         error = bread(ap->a_vp, base_offset, blksize, &bp);
 617                         if (error == 0)
 618                                 bheavy(bp);
 619                 }
 620                 if (error == 0) {
 621                         error = uiomove((char *)bp->b_data + offset,
 622                                         n, uio);
 623                 }
 624
 625                 /*
 626                  * If we screwed up we have to undo any VM size changes we
 627                  * made.
 628                  */
 629                 if (error) {
 630                         brelse(bp);
 631                         if (fixsize) {
 632                                 vtruncbuf(ap->a_vp, ip->ino_data.size,
 633                                           hammer_blocksize(ip->ino_data.size));
 634                         }
 635                         break;
 636                 }
 637                 kflags |= NOTE_WRITE;
 638                 hammer_stats_file_write += n;
 639                 /* bp->b_flags |= B_CLUSTEROK; temporarily disabled */
 640                 if (ip->ino_data.size < uio->uio_offset) {
 641                         ip->ino_data.size = uio->uio_offset;
 642                         flags = HAMMER_INODE_DDIRTY;
 643                         vnode_pager_setsize(ap->a_vp, ip->ino_data.size);
 644                 } else {
 645                         flags = 0;
 646                 }
 647                 ip->ino_data.mtime = trans.time;
 648                 flags |= HAMMER_INODE_MTIME | HAMMER_INODE_BUFS;
 649                 hammer_modify_inode(ip, flags);
 650
 651                 /*
 652                  * Once we dirty the buffer any cached zone-X offset
 653                  * becomes invalid.  HAMMER NOTE: no-history mode cannot
 654                  * allow overwriting over the same data sector unless
 655                  * we provide UNDOs for the old data, which we don't.
 656                  */
 657                 bp->b_bio2.bio_offset = NOOFFSET;
 658
 659                 /*
 660                  * Final buffer disposition.
 661                  *
 662                  * Because meta-data updates are deferred, HAMMER is
 663                  * especially sensitive to excessive bdwrite()s because
 664                  * the I/O stream is not broken up by disk reads.  So the
 665                  * buffer cache simply cannot keep up.
 666                  *
 667                  * WARNING!  blksize is variable.  cluster_write() is
 668                  * expected to not blow up if it encounters buffers that
 669                  * do not match the passed blksize.
 670                  *
 671                  * NOTE!  Hammer shouldn't need to bawrite()/cluster_write().
 672                  *        The ip->rsv_recs check should burst-flush the data.
 673                  *        If we queue it immediately the buf could be left
 674                  *        locked on the device queue for a very long time.
 675                  */
 676                 bp->b_flags |= B_AGE;
 677                 if (ap->a_ioflag & IO_SYNC) {
 678                         bwrite(bp);
 679                 } else if (ap->a_ioflag & IO_DIRECT) {
 680                         bawrite(bp);
 681                 } else {
 682 #if 0
 683                 if (offset + n == blksize) {
 684                         if (hammer_cluster_enable == 0 ||
 685                             (ap->a_vp->v_mount->mnt_flag & MNT_NOCLUSTERW)) {
 686                                 bawrite(bp);
 687                         } else {
 688                                 cluster_write(bp, ip->ino_data.size,
 689                                               blksize, seqcount);
 690                         }
 691                 } else {
 692 #endif
 693                         bdwrite(bp);
 694                 }
 695         }
 696         hammer_done_transaction(&trans);
 697         hammer_knote(ap->a_vp, kflags);
 698         return (error);
 699 }
 700
 701 /*
 702  * hammer_vop_access { vp, mode, cred }
 703  */
 704 static
 705 int
 706 hammer_vop_access(struct vop_access_args *ap)
 707 {
 708         struct hammer_inode *ip = VTOI(ap->a_vp);
 709         uid_t uid;
 710         gid_t gid;
 711         int error;
 712
 713         ++hammer_stats_file_iopsr;
 714         uid = hammer_to_unix_xid(&ip->ino_data.uid);
 715         gid = hammer_to_unix_xid(&ip->ino_data.gid);
 716
 717         error = vop_helper_access(ap, uid, gid, ip->ino_data.mode,
 718                                   ip->ino_data.uflags);
 719         return (error);
 720 }
 721
 722 /*
 723  * hammer_vop_advlock { vp, id, op, fl, flags }
 724  */
 725 static
 726 int
 727 hammer_vop_advlock(struct vop_advlock_args *ap)
 728 {
 729         hammer_inode_t ip = VTOI(ap->a_vp);
 730
 731         return (lf_advlock(ap, &ip->advlock, ip->ino_data.size));
 732 }
 733
 734 /*
 735  * hammer_vop_close { vp, fflag }
 736  *
 737  * We can only sync-on-close for normal closes.
 738  */
 739 static
 740 int
 741 hammer_vop_close(struct vop_close_args *ap)
 742 {
 743         struct vnode *vp = ap->a_vp;
 744         hammer_inode_t ip = VTOI(vp);
 745         int waitfor;
 746
 747         if (ip->flags & (HAMMER_INODE_CLOSESYNC|HAMMER_INODE_CLOSEASYNC)) {
 748                 if (vn_islocked(vp) == LK_EXCLUSIVE &&
 749                     (vp->v_flag & (VINACTIVE|VRECLAIMED)) == 0) {
 750                         if (ip->flags & HAMMER_INODE_CLOSESYNC)
 751                                 waitfor = MNT_WAIT;
 752                         else
 753                                 waitfor = MNT_NOWAIT;
 754                         ip->flags &= ~(HAMMER_INODE_CLOSESYNC |
 755                                        HAMMER_INODE_CLOSEASYNC);
 756                         VOP_FSYNC(vp, MNT_NOWAIT, waitfor);
 757                 }
 758         }
 759         return (vop_stdclose(ap));
 760 }
 761
 762 /*
 763  * hammer_vop_ncreate { nch, dvp, vpp, cred, vap }
 764  *
 765  * The operating system has already ensured that the directory entry
 766  * does not exist and done all appropriate namespace locking.
 767  */
 768 static
 769 int
 770 hammer_vop_ncreate(struct vop_ncreate_args *ap)
 771 {
 772         struct hammer_transaction trans;
 773         struct hammer_inode *dip;
 774         struct hammer_inode *nip;
 775         struct nchandle *nch;
 776         int error;
 777
 778         nch = ap->a_nch;
 779         dip = VTOI(ap->a_dvp);
 780
 781         if (dip->flags & HAMMER_INODE_RO)
 782                 return (EROFS);
 783         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
 784                 return (error);
 785
 786         /*
 787          * Create a transaction to cover the operations we perform.
 788          */
 789         hammer_start_transaction(&trans, dip->hmp);
 790         ++hammer_stats_file_iopsw;
 791
 792         /*
 793          * Create a new filesystem object of the requested type.  The
 794          * returned inode will be referenced and shared-locked to prevent
 795          * it from being moved to the flusher.
 796          */
 797         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
 798                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
 799                                     NULL, &nip);
 800         if (error) {
 801                 hkprintf("hammer_create_inode error %d\n", error);
 802                 hammer_done_transaction(&trans);
 803                 *ap->a_vpp = NULL;
 804                 return (error);
 805         }
 806
 807         /*
 808          * Add the new filesystem object to the directory.  This will also
 809          * bump the inode's link count.
 810          */
 811         error = hammer_ip_add_directory(&trans, dip,
 812                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
 813                                         nip);
 814         if (error)
 815                 hkprintf("hammer_ip_add_directory error %d\n", error);
 816
 817         /*
 818          * Finish up.
 819          */
 820         if (error) {
 821                 hammer_rel_inode(nip, 0);
 822                 hammer_done_transaction(&trans);
 823                 *ap->a_vpp = NULL;
 824         } else {
 825                 error = hammer_get_vnode(nip, ap->a_vpp);
 826                 hammer_done_transaction(&trans);
 827                 hammer_rel_inode(nip, 0);
 828                 if (error == 0) {
 829                         cache_setunresolved(ap->a_nch);
 830                         cache_setvp(ap->a_nch, *ap->a_vpp);
 831                 }
 832                 hammer_knote(ap->a_dvp, NOTE_WRITE);
 833         }
 834         return (error);
 835 }
 836
 837 /*
 838  * hammer_vop_getattr { vp, vap }
 839  *
 840  * Retrieve an inode's attribute information.  When accessing inodes
 841  * historically we fake the atime field to ensure consistent results.
 842  * The atime field is stored in the B-Tree element and allowed to be
 843  * updated without cycling the element.
 844  *
 845  * MPSAFE
 846  */
 847 static
 848 int
 849 hammer_vop_getattr(struct vop_getattr_args *ap)
 850 {
 851         struct hammer_inode *ip = VTOI(ap->a_vp);
 852         struct vattr *vap = ap->a_vap;
 853
 854         /*
 855          * We want the fsid to be different when accessing a filesystem
 856          * with different as-of's so programs like diff don't think
 857          * the files are the same.
 858          *
 859          * We also want the fsid to be the same when comparing snapshots,
 860          * or when comparing mirrors (which might be backed by different
 861          * physical devices).  HAMMER fsids are based on the PFS's
 862          * shared_uuid field.
 863          *
 864          * XXX there is a chance of collision here.  The va_fsid reported
 865          * by stat is different from the more involved fsid used in the
 866          * mount structure.
 867          */
 868         ++hammer_stats_file_iopsr;
 869         hammer_lock_sh(&ip->lock);
 870         vap->va_fsid = ip->pfsm->fsid_udev ^ (u_int32_t)ip->obj_asof ^
 871                        (u_int32_t)(ip->obj_asof >> 32);
 872
 873         vap->va_fileid = ip->ino_leaf.base.obj_id;
 874         vap->va_mode = ip->ino_data.mode;
 875         vap->va_nlink = ip->ino_data.nlinks;
 876         vap->va_uid = hammer_to_unix_xid(&ip->ino_data.uid);
 877         vap->va_gid = hammer_to_unix_xid(&ip->ino_data.gid);
 878         vap->va_rmajor = 0;
 879         vap->va_rminor = 0;
 880         vap->va_size = ip->ino_data.size;
 881
 882         /*
 883          * Special case for @@PFS softlinks.  The actual size of the
 884          * expanded softlink is "@@0x%016llx:%05d" == 26 bytes.
 885          * or for MAX_TID is    "@@-1:%05d" == 10 bytes.
 886          */
 887         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_SOFTLINK &&
 888             ip->ino_data.size == 10 &&
 889             ip->obj_asof == HAMMER_MAX_TID &&
 890             ip->obj_localization == 0 &&
 891             strncmp(ip->ino_data.ext.symlink, "@@PFS", 5) == 0) {
 892                     if (ip->pfsm->pfsd.mirror_flags & HAMMER_PFSD_SLAVE)
 893                             vap->va_size = 26;
 894                     else
 895                             vap->va_size = 10;
 896         }
 897
 898         /*
 899          * We must provide a consistent atime and mtime for snapshots
 900          * so people can do a 'tar cf - ... | md5' on them and get
 901          * consistent results.
 902          */
 903         if (ip->flags & HAMMER_INODE_RO) {
 904                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_atime);
 905                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_mtime);
 906         } else {
 907                 hammer_time_to_timespec(ip->ino_data.atime, &vap->va_atime);
 908                 hammer_time_to_timespec(ip->ino_data.mtime, &vap->va_mtime);
 909         }
 910         hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_ctime);
 911         vap->va_flags = ip->ino_data.uflags;
 912         vap->va_gen = 1;        /* hammer inums are unique for all time */
 913         vap->va_blocksize = HAMMER_BUFSIZE;
 914         if (ip->ino_data.size >= HAMMER_XDEMARC) {
 915                 vap->va_bytes = (ip->ino_data.size + HAMMER_XBUFMASK64) &
 916                                 ~HAMMER_XBUFMASK64;
 917         } else if (ip->ino_data.size > HAMMER_BUFSIZE / 2) {
 918                 vap->va_bytes = (ip->ino_data.size + HAMMER_BUFMASK64) &
 919                                 ~HAMMER_BUFMASK64;
 920         } else {
 921                 vap->va_bytes = (ip->ino_data.size + 15) & ~15;
 922         }
 923
 924         vap->va_type = hammer_get_vnode_type(ip->ino_data.obj_type);
 925         vap->va_filerev = 0;    /* XXX */
 926         /* mtime uniquely identifies any adjustments made to the file XXX */
 927         vap->va_fsmid = ip->ino_data.mtime;
 928         vap->va_uid_uuid = ip->ino_data.uid;
 929         vap->va_gid_uuid = ip->ino_data.gid;
 930         vap->va_fsid_uuid = ip->hmp->fsid;
 931         vap->va_vaflags = VA_UID_UUID_VALID | VA_GID_UUID_VALID |
 932                           VA_FSID_UUID_VALID;
 933
 934         switch (ip->ino_data.obj_type) {
 935         case HAMMER_OBJTYPE_CDEV:
 936         case HAMMER_OBJTYPE_BDEV:
 937                 vap->va_rmajor = ip->ino_data.rmajor;
 938                 vap->va_rminor = ip->ino_data.rminor;
 939                 break;
 940         default:
 941                 break;
 942         }
 943         hammer_unlock(&ip->lock);
 944         return(0);
 945 }
 946
 947 /*
 948  * hammer_vop_nresolve { nch, dvp, cred }
 949  *
 950  * Locate the requested directory entry.
 951  */
 952 static
 953 int
 954 hammer_vop_nresolve(struct vop_nresolve_args *ap)
 955 {
 956         struct hammer_transaction trans;
 957         struct namecache *ncp;
 958         hammer_inode_t dip;
 959         hammer_inode_t ip;
 960         hammer_tid_t asof;
 961         struct hammer_cursor cursor;
 962         struct vnode *vp;
 963         int64_t namekey;
 964         int error;
 965         int i;
 966         int nlen;
 967         int flags;
 968         int ispfs;
 969         int64_t obj_id;
 970         u_int32_t localization;
 971         u_int32_t max_iterations;
 972
 973         /*
 974          * Misc initialization, plus handle as-of name extensions.  Look for
 975          * the '@@' extension.  Note that as-of files and directories cannot
 976          * be modified.
 977          */
 978         dip = VTOI(ap->a_dvp);
 979         ncp = ap->a_nch->ncp;
 980         asof = dip->obj_asof;
 981         localization = dip->obj_localization;   /* for code consistency */
 982         nlen = ncp->nc_nlen;
 983         flags = dip->flags & HAMMER_INODE_RO;
 984         ispfs = 0;
 985
 986         hammer_simple_transaction(&trans, dip->hmp);
 987         ++hammer_stats_file_iopsr;
 988
 989         for (i = 0; i < nlen; ++i) {
 990                 if (ncp->nc_name[i] == '@' && ncp->nc_name[i+1] == '@') {
 991                         error = hammer_str_to_tid(ncp->nc_name + i + 2,
 992                                                   &ispfs, &asof, &localization);
 993                         if (error != 0) {
 994                                 i = nlen;
 995                                 break;
 996                         }
 997                         if (asof != HAMMER_MAX_TID)
 998                                 flags |= HAMMER_INODE_RO;
 999                         break;
1000                 }
1001         }
1002         nlen = i;
1003
1004         /*
1005          * If this is a PFS softlink we dive into the PFS
1006          */
1007         if (ispfs && nlen == 0) {
1008                 ip = hammer_get_inode(&trans, dip, HAMMER_OBJID_ROOT,
1009                                       asof, localization,
1010                                       flags, &error);
1011                 if (error == 0) {
1012                         error = hammer_get_vnode(ip, &vp);
1013                         hammer_rel_inode(ip, 0);
1014                 } else {
1015                         vp = NULL;
1016                 }
1017                 if (error == 0) {
1018                         vn_unlock(vp);
1019                         cache_setvp(ap->a_nch, vp);
1020                         vrele(vp);
1021                 }
1022                 goto done;
1023         }
1024
1025         /*
1026          * If there is no path component the time extension is relative to dip.
1027          * e.g. "fubar/@@<snapshot>"
1028          *
1029          * "." is handled by the kernel, but ".@@<snapshot>" is not.
1030          * e.g. "fubar/.@@<snapshot>"
1031          *
1032          * ".." is handled by the kernel.  We do not currently handle
1033          * "..@<snapshot>".
1034          */
1035         if (nlen == 0 || (nlen == 1 && ncp->nc_name[0] == '.')) {
1036                 ip = hammer_get_inode(&trans, dip, dip->obj_id,
1037                                       asof, dip->obj_localization,
1038                                       flags, &error);
1039                 if (error == 0) {
1040                         error = hammer_get_vnode(ip, &vp);
1041                         hammer_rel_inode(ip, 0);
1042                 } else {
1043                         vp = NULL;
1044                 }
1045                 if (error == 0) {
1046                         vn_unlock(vp);
1047                         cache_setvp(ap->a_nch, vp);
1048                         vrele(vp);
1049                 }
1050                 goto done;
1051         }
1052
1053         /*
1054          * Calculate the namekey and setup the key range for the scan.  This
1055          * works kinda like a chained hash table where the lower 32 bits
1056          * of the namekey synthesize the chain.
1057          *
1058          * The key range is inclusive of both key_beg and key_end.
1059          */
1060         namekey = hammer_directory_namekey(dip, ncp->nc_name, nlen,
1061                                            &max_iterations);
1062
1063         error = hammer_init_cursor(&trans, &cursor, &dip->cache[1], dip);
1064         cursor.key_beg.localization = dip->obj_localization +
1065                                       hammer_dir_localization(dip);
1066         cursor.key_beg.obj_id = dip->obj_id;
1067         cursor.key_beg.key = namekey;
1068         cursor.key_beg.create_tid = 0;
1069         cursor.key_beg.delete_tid = 0;
1070         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1071         cursor.key_beg.obj_type = 0;
1072
1073         cursor.key_end = cursor.key_beg;
1074         cursor.key_end.key += max_iterations;
1075         cursor.asof = asof;
1076         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1077
1078         /*
1079          * Scan all matching records (the chain), locate the one matching
1080          * the requested path component.
1081          *
1082          * The hammer_ip_*() functions merge in-memory records with on-disk
1083          * records for the purposes of the search.
1084          */
1085         obj_id = 0;
1086         localization = HAMMER_DEF_LOCALIZATION;
1087
1088         if (error == 0) {
1089                 error = hammer_ip_first(&cursor);
1090                 while (error == 0) {
1091                         error = hammer_ip_resolve_data(&cursor);
1092                         if (error)
1093                                 break;
1094                         if (nlen == cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF &&
1095                             bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1096                                 obj_id = cursor.data->entry.obj_id;
1097                                 localization = cursor.data->entry.localization;
1098                                 break;
1099                         }
1100                         error = hammer_ip_next(&cursor);
1101                 }
1102         }
1103         hammer_done_cursor(&cursor);
1104
1105         /*
1106          * Lookup the obj_id.  This should always succeed.  If it does not
1107          * the filesystem may be damaged and we return a dummy inode.
1108          */
1109         if (error == 0) {
1110                 ip = hammer_get_inode(&trans, dip, obj_id,
1111                                       asof, localization,
1112                                       flags, &error);
1113                 if (error == ENOENT) {
1114                         kprintf("HAMMER: WARNING: Missing "
1115                                 "inode for dirent \"%s\"\n"
1116                                 "\tobj_id = %016llx, asof=%016llx, lo=%08x\n",
1117                                 ncp->nc_name,
1118                                 (long long)obj_id, (long long)asof,
1119                                 localization);
1120                         error = 0;
1121                         ip = hammer_get_dummy_inode(&trans, dip, obj_id,
1122                                                     asof, localization,
1123                                                     flags, &error);
1124                 }
1125                 if (error == 0) {
1126                         error = hammer_get_vnode(ip, &vp);
1127                         hammer_rel_inode(ip, 0);
1128                 } else {
1129                         vp = NULL;
1130                 }
1131                 if (error == 0) {
1132                         vn_unlock(vp);
1133                         cache_setvp(ap->a_nch, vp);
1134                         vrele(vp);
1135                 }
1136         } else if (error == ENOENT) {
1137                 cache_setvp(ap->a_nch, NULL);
1138         }
1139 done:
1140         hammer_done_transaction(&trans);
1141         return (error);
1142 }
1143
1144 /*
1145  * hammer_vop_nlookupdotdot { dvp, vpp, cred }
1146  *
1147  * Locate the parent directory of a directory vnode.
1148  *
1149  * dvp is referenced but not locked.  *vpp must be returned referenced and
1150  * locked.  A parent_obj_id of 0 does not necessarily indicate that we are
1151  * at the root, instead it could indicate that the directory we were in was
1152  * removed.
1153  *
1154  * NOTE: as-of sequences are not linked into the directory structure.  If
1155  * we are at the root with a different asof then the mount point, reload
1156  * the same directory with the mount point's asof.   I'm not sure what this
1157  * will do to NFS.  We encode ASOF stamps in NFS file handles so it might not
1158  * get confused, but it hasn't been tested.
1159  */
1160 static
1161 int
1162 hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *ap)
1163 {
1164         struct hammer_transaction trans;
1165         struct hammer_inode *dip;
1166         struct hammer_inode *ip;
1167         int64_t parent_obj_id;
1168         u_int32_t parent_obj_localization;
1169         hammer_tid_t asof;
1170         int error;
1171
1172         dip = VTOI(ap->a_dvp);
1173         asof = dip->obj_asof;
1174
1175         /*
1176          * Whos are parent?  This could be the root of a pseudo-filesystem
1177          * whos parent is in another localization domain.
1178          */
1179         parent_obj_id = dip->ino_data.parent_obj_id;
1180         if (dip->obj_id == HAMMER_OBJID_ROOT)
1181                 parent_obj_localization = dip->ino_data.ext.obj.parent_obj_localization;
1182         else
1183                 parent_obj_localization = dip->obj_localization;
1184
1185         if (parent_obj_id == 0) {
1186                 if (dip->obj_id == HAMMER_OBJID_ROOT &&
1187                    asof != dip->hmp->asof) {
1188                         parent_obj_id = dip->obj_id;
1189                         asof = dip->hmp->asof;
1190                         *ap->a_fakename = kmalloc(19, M_TEMP, M_WAITOK);
1191                         ksnprintf(*ap->a_fakename, 19, "0x%016llx",
1192                                   (long long)dip->obj_asof);
1193                 } else {
1194                         *ap->a_vpp = NULL;
1195                         return ENOENT;
1196                 }
1197         }
1198
1199         hammer_simple_transaction(&trans, dip->hmp);
1200         ++hammer_stats_file_iopsr;
1201
1202         ip = hammer_get_inode(&trans, dip, parent_obj_id,
1203                               asof, parent_obj_localization,
1204                               dip->flags, &error);
1205         if (ip) {
1206                 error = hammer_get_vnode(ip, ap->a_vpp);
1207                 hammer_rel_inode(ip, 0);
1208         } else {
1209                 *ap->a_vpp = NULL;
1210         }
1211         hammer_done_transaction(&trans);
1212         return (error);
1213 }
1214
1215 /*
1216  * hammer_vop_nlink { nch, dvp, vp, cred }
1217  */
1218 static
1219 int
1220 hammer_vop_nlink(struct vop_nlink_args *ap)
1221 {
1222         struct hammer_transaction trans;
1223         struct hammer_inode *dip;
1224         struct hammer_inode *ip;
1225         struct nchandle *nch;
1226         int error;
1227
1228         if (ap->a_dvp->v_mount != ap->a_vp->v_mount)
1229                 return(EXDEV);
1230
1231         nch = ap->a_nch;
1232         dip = VTOI(ap->a_dvp);
1233         ip = VTOI(ap->a_vp);
1234
1235         if (dip->obj_localization != ip->obj_localization)
1236                 return(EXDEV);
1237
1238         if (dip->flags & HAMMER_INODE_RO)
1239                 return (EROFS);
1240         if (ip->flags & HAMMER_INODE_RO)
1241                 return (EROFS);
1242         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1243                 return (error);
1244
1245         /*
1246          * Create a transaction to cover the operations we perform.
1247          */
1248         hammer_start_transaction(&trans, dip->hmp);
1249         ++hammer_stats_file_iopsw;
1250
1251         /*
1252          * Add the filesystem object to the directory.  Note that neither
1253          * dip nor ip are referenced or locked, but their vnodes are
1254          * referenced.  This function will bump the inode's link count.
1255          */
1256         error = hammer_ip_add_directory(&trans, dip,
1257                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1258                                         ip);
1259
1260         /*
1261          * Finish up.
1262          */
1263         if (error == 0) {
1264                 cache_setunresolved(nch);
1265                 cache_setvp(nch, ap->a_vp);
1266         }
1267         hammer_done_transaction(&trans);
1268         hammer_knote(ap->a_vp, NOTE_LINK);
1269         hammer_knote(ap->a_dvp, NOTE_WRITE);
1270         return (error);
1271 }
1272
1273 /*
1274  * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
1275  *
1276  * The operating system has already ensured that the directory entry
1277  * does not exist and done all appropriate namespace locking.
1278  */
1279 static
1280 int
1281 hammer_vop_nmkdir(struct vop_nmkdir_args *ap)
1282 {
1283         struct hammer_transaction trans;
1284         struct hammer_inode *dip;
1285         struct hammer_inode *nip;
1286         struct nchandle *nch;
1287         int error;
1288
1289         nch = ap->a_nch;
1290         dip = VTOI(ap->a_dvp);
1291
1292         if (dip->flags & HAMMER_INODE_RO)
1293                 return (EROFS);
1294         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1295                 return (error);
1296
1297         /*
1298          * Create a transaction to cover the operations we perform.
1299          */
1300         hammer_start_transaction(&trans, dip->hmp);
1301         ++hammer_stats_file_iopsw;
1302
1303         /*
1304          * Create a new filesystem object of the requested type.  The
1305          * returned inode will be referenced but not locked.
1306          */
1307         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1308                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1309                                     NULL, &nip);
1310         if (error) {
1311                 hkprintf("hammer_mkdir error %d\n", error);
1312                 hammer_done_transaction(&trans);
1313                 *ap->a_vpp = NULL;
1314                 return (error);
1315         }
1316         /*
1317          * Add the new filesystem object to the directory.  This will also
1318          * bump the inode's link count.
1319          */
1320         error = hammer_ip_add_directory(&trans, dip,
1321                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1322                                         nip);
1323         if (error)
1324                 hkprintf("hammer_mkdir (add) error %d\n", error);
1325
1326         /*
1327          * Finish up.
1328          */
1329         if (error) {
1330                 hammer_rel_inode(nip, 0);
1331                 *ap->a_vpp = NULL;
1332         } else {
1333                 error = hammer_get_vnode(nip, ap->a_vpp);
1334                 hammer_rel_inode(nip, 0);
1335                 if (error == 0) {
1336                         cache_setunresolved(ap->a_nch);
1337                         cache_setvp(ap->a_nch, *ap->a_vpp);
1338                 }
1339         }
1340         hammer_done_transaction(&trans);
1341         if (error == 0)
1342                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
1343         return (error);
1344 }
1345
1346 /*
1347  * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1348  *
1349  * The operating system has already ensured that the directory entry
1350  * does not exist and done all appropriate namespace locking.
1351  */
1352 static
1353 int
1354 hammer_vop_nmknod(struct vop_nmknod_args *ap)
1355 {
1356         struct hammer_transaction trans;
1357         struct hammer_inode *dip;
1358         struct hammer_inode *nip;
1359         struct nchandle *nch;
1360         int error;
1361
1362         nch = ap->a_nch;
1363         dip = VTOI(ap->a_dvp);
1364
1365         if (dip->flags & HAMMER_INODE_RO)
1366                 return (EROFS);
1367         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1368                 return (error);
1369
1370         /*
1371          * Create a transaction to cover the operations we perform.
1372          */
1373         hammer_start_transaction(&trans, dip->hmp);
1374         ++hammer_stats_file_iopsw;
1375
1376         /*
1377          * Create a new filesystem object of the requested type.  The
1378          * returned inode will be referenced but not locked.
1379          *
1380          * If mknod specifies a directory a pseudo-fs is created.
1381          */
1382         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1383                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1384                                     NULL, &nip);
1385         if (error) {
1386                 hammer_done_transaction(&trans);
1387                 *ap->a_vpp = NULL;
1388                 return (error);
1389         }
1390
1391         /*
1392          * Add the new filesystem object to the directory.  This will also
1393          * bump the inode's link count.
1394          */
1395         error = hammer_ip_add_directory(&trans, dip,
1396                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1397                                         nip);
1398
1399         /*
1400          * Finish up.
1401          */
1402         if (error) {
1403                 hammer_rel_inode(nip, 0);
1404                 *ap->a_vpp = NULL;
1405         } else {
1406                 error = hammer_get_vnode(nip, ap->a_vpp);
1407                 hammer_rel_inode(nip, 0);
1408                 if (error == 0) {
1409                         cache_setunresolved(ap->a_nch);
1410                         cache_setvp(ap->a_nch, *ap->a_vpp);
1411                 }
1412         }
1413         hammer_done_transaction(&trans);
1414         if (error == 0)
1415                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1416         return (error);
1417 }
1418
1419 /*
1420  * hammer_vop_open { vp, mode, cred, fp }
1421  */
1422 static
1423 int
1424 hammer_vop_open(struct vop_open_args *ap)
1425 {
1426         hammer_inode_t ip;
1427
1428         ++hammer_stats_file_iopsr;
1429         ip = VTOI(ap->a_vp);
1430
1431         if ((ap->a_mode & FWRITE) && (ip->flags & HAMMER_INODE_RO))
1432                 return (EROFS);
1433         return(vop_stdopen(ap));
1434 }
1435
1436 /*
1437  * hammer_vop_print { vp }
1438  */
1439 static
1440 int
1441 hammer_vop_print(struct vop_print_args *ap)
1442 {
1443         return EOPNOTSUPP;
1444 }
1445
1446 /*
1447  * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1448  */
1449 static
1450 int
1451 hammer_vop_readdir(struct vop_readdir_args *ap)
1452 {
1453         struct hammer_transaction trans;
1454         struct hammer_cursor cursor;
1455         struct hammer_inode *ip;
1456         struct uio *uio;
1457         hammer_base_elm_t base;
1458         int error;
1459         int cookie_index;
1460         int ncookies;
1461         off_t *cookies;
1462         off_t saveoff;
1463         int r;
1464         int dtype;
1465
1466         ++hammer_stats_file_iopsr;
1467         ip = VTOI(ap->a_vp);
1468         uio = ap->a_uio;
1469         saveoff = uio->uio_offset;
1470
1471         if (ap->a_ncookies) {
1472                 ncookies = uio->uio_resid / 16 + 1;
1473                 if (ncookies > 1024)
1474                         ncookies = 1024;
1475                 cookies = kmalloc(ncookies * sizeof(off_t), M_TEMP, M_WAITOK);
1476                 cookie_index = 0;
1477         } else {
1478                 ncookies = -1;
1479                 cookies = NULL;
1480                 cookie_index = 0;
1481         }
1482
1483         hammer_simple_transaction(&trans, ip->hmp);
1484
1485         /*
1486          * Handle artificial entries
1487          *
1488          * It should be noted that the minimum value for a directory
1489          * hash key on-media is 0x0000000100000000, so we can use anything
1490          * less then that to represent our 'special' key space.
1491          */
1492         error = 0;
1493         if (saveoff == 0) {
1494                 r = vop_write_dirent(&error, uio, ip->obj_id, DT_DIR, 1, ".");
1495                 if (r)
1496                         goto done;
1497                 if (cookies)
1498                         cookies[cookie_index] = saveoff;
1499                 ++saveoff;
1500                 ++cookie_index;
1501                 if (cookie_index == ncookies)
1502                         goto done;
1503         }
1504         if (saveoff == 1) {
1505                 if (ip->ino_data.parent_obj_id) {
1506                         r = vop_write_dirent(&error, uio,
1507                                              ip->ino_data.parent_obj_id,
1508                                              DT_DIR, 2, "..");
1509                 } else {
1510                         r = vop_write_dirent(&error, uio,
1511                                              ip->obj_id, DT_DIR, 2, "..");
1512                 }
1513                 if (r)
1514                         goto done;
1515                 if (cookies)
1516                         cookies[cookie_index] = saveoff;
1517                 ++saveoff;
1518                 ++cookie_index;
1519                 if (cookie_index == ncookies)
1520                         goto done;
1521         }
1522
1523         /*
1524          * Key range (begin and end inclusive) to scan.  Directory keys
1525          * directly translate to a 64 bit 'seek' position.
1526          */
1527         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1528         cursor.key_beg.localization = ip->obj_localization +
1529                                       hammer_dir_localization(ip);
1530         cursor.key_beg.obj_id = ip->obj_id;
1531         cursor.key_beg.create_tid = 0;
1532         cursor.key_beg.delete_tid = 0;
1533         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1534         cursor.key_beg.obj_type = 0;
1535         cursor.key_beg.key = saveoff;
1536
1537         cursor.key_end = cursor.key_beg;
1538         cursor.key_end.key = HAMMER_MAX_KEY;
1539         cursor.asof = ip->obj_asof;
1540         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1541
1542         error = hammer_ip_first(&cursor);
1543
1544         while (error == 0) {
1545                 error = hammer_ip_resolve_data(&cursor);
1546                 if (error)
1547                         break;
1548                 base = &cursor.leaf->base;
1549                 saveoff = base->key;
1550                 KKASSERT(cursor.leaf->data_len > HAMMER_ENTRY_NAME_OFF);
1551
1552                 if (base->obj_id != ip->obj_id)
1553                         panic("readdir: bad record at %p", cursor.node);
1554
1555                 /*
1556                  * Convert pseudo-filesystems into softlinks
1557                  */
1558                 dtype = hammer_get_dtype(cursor.leaf->base.obj_type);
1559                 r = vop_write_dirent(
1560                              &error, uio, cursor.data->entry.obj_id,
1561                              dtype,
1562                              cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF ,
1563                              (void *)cursor.data->entry.name);
1564                 if (r)
1565                         break;
1566                 ++saveoff;
1567                 if (cookies)
1568                         cookies[cookie_index] = base->key;
1569                 ++cookie_index;
1570                 if (cookie_index == ncookies)
1571                         break;
1572                 error = hammer_ip_next(&cursor);
1573         }
1574         hammer_done_cursor(&cursor);
1575
1576 done:
1577         hammer_done_transaction(&trans);
1578
1579         if (ap->a_eofflag)
1580                 *ap->a_eofflag = (error == ENOENT);
1581         uio->uio_offset = saveoff;
1582         if (error && cookie_index == 0) {
1583                 if (error == ENOENT)
1584                         error = 0;
1585                 if (cookies) {
1586                         kfree(cookies, M_TEMP);
1587                         *ap->a_ncookies = 0;
1588                         *ap->a_cookies = NULL;
1589                 }
1590         } else {
1591                 if (error == ENOENT)
1592                         error = 0;
1593                 if (cookies) {
1594                         *ap->a_ncookies = cookie_index;
1595                         *ap->a_cookies = cookies;
1596                 }
1597         }
1598         return(error);
1599 }
1600
1601 /*
1602  * hammer_vop_readlink { vp, uio, cred }
1603  */
1604 static
1605 int
1606 hammer_vop_readlink(struct vop_readlink_args *ap)
1607 {
1608         struct hammer_transaction trans;
1609         struct hammer_cursor cursor;
1610         struct hammer_inode *ip;
1611         char buf[32];
1612         u_int32_t localization;
1613         hammer_pseudofs_inmem_t pfsm;
1614         int error;
1615
1616         ip = VTOI(ap->a_vp);
1617
1618         /*
1619          * Shortcut if the symlink data was stuffed into ino_data.
1620          *
1621          * Also expand special "@@PFS%05d" softlinks (expansion only
1622          * occurs for non-historical (current) accesses made from the
1623          * primary filesystem).
1624          */
1625         if (ip->ino_data.size <= HAMMER_INODE_BASESYMLEN) {
1626                 char *ptr;
1627                 int bytes;
1628
1629                 ptr = ip->ino_data.ext.symlink;
1630                 bytes = (int)ip->ino_data.size;
1631                 if (bytes == 10 &&
1632                     ip->obj_asof == HAMMER_MAX_TID &&
1633                     ip->obj_localization == 0 &&
1634                     strncmp(ptr, "@@PFS", 5) == 0) {
1635                         hammer_simple_transaction(&trans, ip->hmp);
1636                         bcopy(ptr + 5, buf, 5);
1637                         buf[5] = 0;
1638                         localization = strtoul(buf, NULL, 10) << 16;
1639                         pfsm = hammer_load_pseudofs(&trans, localization,
1640                                                     &error);
1641                         if (error == 0) {
1642                                 if (pfsm->pfsd.mirror_flags &
1643                                     HAMMER_PFSD_SLAVE) {
1644                                         /* vap->va_size == 26 */
1645                                         ksnprintf(buf, sizeof(buf),
1646                                                   "@@0x%016llx:%05d",
1647                                                   (long long)pfsm->pfsd.sync_end_tid,
1648                                                   localization >> 16);
1649                                 } else {
1650                                         /* vap->va_size == 10 */
1651                                         ksnprintf(buf, sizeof(buf),
1652                                                   "@@-1:%05d",
1653                                                   localization >> 16);
1654 #if 0
1655                                         ksnprintf(buf, sizeof(buf),
1656                                                   "@@0x%016llx:%05d",
1657                                                   (long long)HAMMER_MAX_TID,
1658                                                   localization >> 16);
1659 #endif
1660                                 }
1661                                 ptr = buf;
1662                                 bytes = strlen(buf);
1663                         }
1664                         if (pfsm)
1665                                 hammer_rel_pseudofs(trans.hmp, pfsm);
1666                         hammer_done_transaction(&trans);
1667                 }
1668                 error = uiomove(ptr, bytes, ap->a_uio);
1669                 return(error);
1670         }
1671
1672         /*
1673          * Long version
1674          */
1675         hammer_simple_transaction(&trans, ip->hmp);
1676         ++hammer_stats_file_iopsr;
1677         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1678
1679         /*
1680          * Key range (begin and end inclusive) to scan.  Directory keys
1681          * directly translate to a 64 bit 'seek' position.
1682          */
1683         cursor.key_beg.localization = ip->obj_localization +
1684                                       HAMMER_LOCALIZE_MISC;
1685         cursor.key_beg.obj_id = ip->obj_id;
1686         cursor.key_beg.create_tid = 0;
1687         cursor.key_beg.delete_tid = 0;
1688         cursor.key_beg.rec_type = HAMMER_RECTYPE_FIX;
1689         cursor.key_beg.obj_type = 0;
1690         cursor.key_beg.key = HAMMER_FIXKEY_SYMLINK;
1691         cursor.asof = ip->obj_asof;
1692         cursor.flags |= HAMMER_CURSOR_ASOF;
1693
1694         error = hammer_ip_lookup(&cursor);
1695         if (error == 0) {
1696                 error = hammer_ip_resolve_data(&cursor);
1697                 if (error == 0) {
1698                         KKASSERT(cursor.leaf->data_len >=
1699                                  HAMMER_SYMLINK_NAME_OFF);
1700                         error = uiomove(cursor.data->symlink.name,
1701                                         cursor.leaf->data_len -
1702                                                 HAMMER_SYMLINK_NAME_OFF,
1703                                         ap->a_uio);
1704                 }
1705         }
1706         hammer_done_cursor(&cursor);
1707         hammer_done_transaction(&trans);
1708         return(error);
1709 }
1710
1711 /*
1712  * hammer_vop_nremove { nch, dvp, cred }
1713  */
1714 static
1715 int
1716 hammer_vop_nremove(struct vop_nremove_args *ap)
1717 {
1718         struct hammer_transaction trans;
1719         struct hammer_inode *dip;
1720         int error;
1721
1722         dip = VTOI(ap->a_dvp);
1723
1724         if (hammer_nohistory(dip) == 0 &&
1725             (error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1726                 return (error);
1727         }
1728
1729         hammer_start_transaction(&trans, dip->hmp);
1730         ++hammer_stats_file_iopsw;
1731         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 0);
1732         hammer_done_transaction(&trans);
1733         if (error == 0)
1734                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1735         return (error);
1736 }
1737
1738 /*
1739  * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1740  */
1741 static
1742 int
1743 hammer_vop_nrename(struct vop_nrename_args *ap)
1744 {
1745         struct hammer_transaction trans;
1746         struct namecache *fncp;
1747         struct namecache *tncp;
1748         struct hammer_inode *fdip;
1749         struct hammer_inode *tdip;
1750         struct hammer_inode *ip;
1751         struct hammer_cursor cursor;
1752         int64_t namekey;
1753         u_int32_t max_iterations;
1754         int nlen, error;
1755
1756         if (ap->a_fdvp->v_mount != ap->a_tdvp->v_mount)
1757                 return(EXDEV);
1758         if (ap->a_fdvp->v_mount != ap->a_fnch->ncp->nc_vp->v_mount)
1759                 return(EXDEV);
1760
1761         fdip = VTOI(ap->a_fdvp);
1762         tdip = VTOI(ap->a_tdvp);
1763         fncp = ap->a_fnch->ncp;
1764         tncp = ap->a_tnch->ncp;
1765         ip = VTOI(fncp->nc_vp);
1766         KKASSERT(ip != NULL);
1767
1768         if (fdip->obj_localization != tdip->obj_localization)
1769                 return(EXDEV);
1770         if (fdip->obj_localization != ip->obj_localization)
1771                 return(EXDEV);
1772
1773         if (fdip->flags & HAMMER_INODE_RO)
1774                 return (EROFS);
1775         if (tdip->flags & HAMMER_INODE_RO)
1776                 return (EROFS);
1777         if (ip->flags & HAMMER_INODE_RO)
1778                 return (EROFS);
1779         if ((error = hammer_checkspace(fdip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1780                 return (error);
1781
1782         hammer_start_transaction(&trans, fdip->hmp);
1783         ++hammer_stats_file_iopsw;
1784
1785         /*
1786          * Remove tncp from the target directory and then link ip as
1787          * tncp. XXX pass trans to dounlink
1788          *
1789          * Force the inode sync-time to match the transaction so it is
1790          * in-sync with the creation of the target directory entry.
1791          */
1792         error = hammer_dounlink(&trans, ap->a_tnch, ap->a_tdvp,
1793                                 ap->a_cred, 0, -1);
1794         if (error == 0 || error == ENOENT) {
1795                 error = hammer_ip_add_directory(&trans, tdip,
1796                                                 tncp->nc_name, tncp->nc_nlen,
1797                                                 ip);
1798                 if (error == 0) {
1799                         ip->ino_data.parent_obj_id = tdip->obj_id;
1800                         ip->ino_data.ctime = trans.time;
1801                         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
1802                 }
1803         }
1804         if (error)
1805                 goto failed; /* XXX */
1806
1807         /*
1808          * Locate the record in the originating directory and remove it.
1809          *
1810          * Calculate the namekey and setup the key range for the scan.  This
1811          * works kinda like a chained hash table where the lower 32 bits
1812          * of the namekey synthesize the chain.
1813          *
1814          * The key range is inclusive of both key_beg and key_end.
1815          */
1816         namekey = hammer_directory_namekey(fdip, fncp->nc_name, fncp->nc_nlen,
1817                                            &max_iterations);
1818 retry:
1819         hammer_init_cursor(&trans, &cursor, &fdip->cache[1], fdip);
1820         cursor.key_beg.localization = fdip->obj_localization +
1821                                       hammer_dir_localization(fdip);
1822         cursor.key_beg.obj_id = fdip->obj_id;
1823         cursor.key_beg.key = namekey;
1824         cursor.key_beg.create_tid = 0;
1825         cursor.key_beg.delete_tid = 0;
1826         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1827         cursor.key_beg.obj_type = 0;
1828
1829         cursor.key_end = cursor.key_beg;
1830         cursor.key_end.key += max_iterations;
1831         cursor.asof = fdip->obj_asof;
1832         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1833
1834         /*
1835          * Scan all matching records (the chain), locate the one matching
1836          * the requested path component.
1837          *
1838          * The hammer_ip_*() functions merge in-memory records with on-disk
1839          * records for the purposes of the search.
1840          */
1841         error = hammer_ip_first(&cursor);
1842         while (error == 0) {
1843                 if (hammer_ip_resolve_data(&cursor) != 0)
1844                         break;
1845                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
1846                 KKASSERT(nlen > 0);
1847                 if (fncp->nc_nlen == nlen &&
1848                     bcmp(fncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1849                         break;
1850                 }
1851                 error = hammer_ip_next(&cursor);
1852         }
1853
1854         /*
1855          * If all is ok we have to get the inode so we can adjust nlinks.
1856          *
1857          * WARNING: hammer_ip_del_directory() may have to terminate the
1858          * cursor to avoid a recursion.  It's ok to call hammer_done_cursor()
1859          * twice.
1860          */
1861         if (error == 0)
1862                 error = hammer_ip_del_directory(&trans, &cursor, fdip, ip);
1863
1864         /*
1865          * XXX A deadlock here will break rename's atomicy for the purposes
1866          * of crash recovery.
1867          */
1868         if (error == EDEADLK) {
1869                 hammer_done_cursor(&cursor);
1870                 goto retry;
1871         }
1872
1873         /*
1874          * Cleanup and tell the kernel that the rename succeeded.
1875          */
1876         hammer_done_cursor(&cursor);
1877         if (error == 0) {
1878                 cache_rename(ap->a_fnch, ap->a_tnch);
1879                 hammer_knote(ap->a_fdvp, NOTE_WRITE);
1880                 hammer_knote(ap->a_tdvp, NOTE_WRITE);
1881                 if (ip->vp)
1882                         hammer_knote(ip->vp, NOTE_RENAME);
1883         }
1884
1885 failed:
1886         hammer_done_transaction(&trans);
1887         return (error);
1888 }
1889
1890 /*
1891  * hammer_vop_nrmdir { nch, dvp, cred }
1892  */
1893 static
1894 int
1895 hammer_vop_nrmdir(struct vop_nrmdir_args *ap)
1896 {
1897         struct hammer_transaction trans;
1898         struct hammer_inode *dip;
1899         int error;
1900
1901         dip = VTOI(ap->a_dvp);
1902
1903         if (hammer_nohistory(dip) == 0 &&
1904             (error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1905                 return (error);
1906         }
1907
1908         hammer_start_transaction(&trans, dip->hmp);
1909         ++hammer_stats_file_iopsw;
1910         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 1);
1911         hammer_done_transaction(&trans);
1912         if (error == 0)
1913                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
1914         return (error);
1915 }
1916
1917 /*
1918  * hammer_vop_markatime { vp, cred }
1919  */
1920 static
1921 int
1922 hammer_vop_markatime(struct vop_markatime_args *ap)
1923 {
1924         struct hammer_transaction trans;
1925         struct hammer_inode *ip;
1926
1927         ip = VTOI(ap->a_vp);
1928         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
1929                 return (EROFS);
1930         if (ip->flags & HAMMER_INODE_RO)
1931                 return (EROFS);
1932         if (ip->hmp->mp->mnt_flag & MNT_NOATIME)
1933                 return (0);
1934         hammer_start_transaction(&trans, ip->hmp);
1935         ++hammer_stats_file_iopsw;
1936
1937         ip->ino_data.atime = trans.time;
1938         hammer_modify_inode(ip, HAMMER_INODE_ATIME);
1939         hammer_done_transaction(&trans);
1940         hammer_knote(ap->a_vp, NOTE_ATTRIB);
1941         return (0);
1942 }
1943
1944 /*
1945  * hammer_vop_setattr { vp, vap, cred }
1946  */
1947 static
1948 int
1949 hammer_vop_setattr(struct vop_setattr_args *ap)
1950 {
1951         struct hammer_transaction trans;
1952         struct vattr *vap;
1953         struct hammer_inode *ip;
1954         int modflags;
1955         int error;
1956         int truncating;
1957         int blksize;
1958         int kflags;
1959         int64_t aligned_size;
1960         u_int32_t flags;
1961
1962         vap = ap->a_vap;
1963         ip = ap->a_vp->v_data;
1964         modflags = 0;
1965         kflags = 0;
1966
1967         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
1968                 return(EROFS);
1969         if (ip->flags & HAMMER_INODE_RO)
1970                 return (EROFS);
1971         if (hammer_nohistory(ip) == 0 &&
1972             (error = hammer_checkspace(ip->hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1973                 return (error);
1974         }
1975
1976         hammer_start_transaction(&trans, ip->hmp);
1977         ++hammer_stats_file_iopsw;
1978         error = 0;
1979
1980         if (vap->va_flags != VNOVAL) {
1981                 flags = ip->ino_data.uflags;
1982                 error = vop_helper_setattr_flags(&flags, vap->va_flags,
1983                                          hammer_to_unix_xid(&ip->ino_data.uid),
1984                                          ap->a_cred);
1985                 if (error == 0) {
1986                         if (ip->ino_data.uflags != flags) {
1987                                 ip->ino_data.uflags = flags;
1988                                 ip->ino_data.ctime = trans.time;
1989                                 modflags |= HAMMER_INODE_DDIRTY;
1990                                 kflags |= NOTE_ATTRIB;
1991                         }
1992                         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
1993                                 error = 0;
1994                                 goto done;
1995                         }
1996                 }
1997                 goto done;
1998         }
1999         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2000                 error = EPERM;
2001                 goto done;
2002         }
2003         if (vap->va_uid != (uid_t)VNOVAL || vap->va_gid != (gid_t)VNOVAL) {
2004                 mode_t cur_mode = ip->ino_data.mode;
2005                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2006                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2007                 uuid_t uuid_uid;
2008                 uuid_t uuid_gid;
2009
2010                 error = vop_helper_chown(ap->a_vp, vap->va_uid, vap->va_gid,
2011                                          ap->a_cred,
2012                                          &cur_uid, &cur_gid, &cur_mode);
2013                 if (error == 0) {
2014                         hammer_guid_to_uuid(&uuid_uid, cur_uid);
2015                         hammer_guid_to_uuid(&uuid_gid, cur_gid);
2016                         if (bcmp(&uuid_uid, &ip->ino_data.uid,
2017                                  sizeof(uuid_uid)) ||
2018                             bcmp(&uuid_gid, &ip->ino_data.gid,
2019                                  sizeof(uuid_gid)) ||
2020                             ip->ino_data.mode != cur_mode
2021                         ) {
2022                                 ip->ino_data.uid = uuid_uid;
2023                                 ip->ino_data.gid = uuid_gid;
2024                                 ip->ino_data.mode = cur_mode;
2025                                 ip->ino_data.ctime = trans.time;
2026                                 modflags |= HAMMER_INODE_DDIRTY;
2027                         }
2028                         kflags |= NOTE_ATTRIB;
2029                 }
2030         }
2031         while (vap->va_size != VNOVAL && ip->ino_data.size != vap->va_size) {
2032                 switch(ap->a_vp->v_type) {
2033                 case VREG:
2034                         if (vap->va_size == ip->ino_data.size)
2035                                 break;
2036                         /*
2037                          * XXX break atomicy, we can deadlock the backend
2038                          * if we do not release the lock.  Probably not a
2039                          * big deal here.
2040                          */
2041                         blksize = hammer_blocksize(vap->va_size);
2042                         if (vap->va_size < ip->ino_data.size) {
2043                                 vtruncbuf(ap->a_vp, vap->va_size, blksize);
2044                                 truncating = 1;
2045                                 kflags |= NOTE_WRITE;
2046                         } else {
2047                                 vnode_pager_setsize(ap->a_vp, vap->va_size);
2048                                 truncating = 0;
2049                                 kflags |= NOTE_WRITE | NOTE_EXTEND;
2050                         }
2051                         ip->ino_data.size = vap->va_size;
2052                         ip->ino_data.mtime = trans.time;
2053                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2054
2055                         /*
2056                          * on-media truncation is cached in the inode until
2057                          * the inode is synchronized.
2058                          */
2059                         if (truncating) {
2060                                 hammer_ip_frontend_trunc(ip, vap->va_size);
2061 #ifdef DEBUG_TRUNCATE
2062                                 if (HammerTruncIp == NULL)
2063                                         HammerTruncIp = ip;
2064 #endif
2065                                 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2066                                         ip->flags |= HAMMER_INODE_TRUNCATED;
2067                                         ip->trunc_off = vap->va_size;
2068 #ifdef DEBUG_TRUNCATE
2069                                         if (ip == HammerTruncIp)
2070                                         kprintf("truncate1 %016llx\n",
2071                                                 (long long)ip->trunc_off);
2072 #endif
2073                                 } else if (ip->trunc_off > vap->va_size) {
2074                                         ip->trunc_off = vap->va_size;
2075 #ifdef DEBUG_TRUNCATE
2076                                         if (ip == HammerTruncIp)
2077                                         kprintf("truncate2 %016llx\n",
2078                                                 (long long)ip->trunc_off);
2079 #endif
2080                                 } else {
2081 #ifdef DEBUG_TRUNCATE
2082                                         if (ip == HammerTruncIp)
2083                                         kprintf("truncate3 %016llx (ignored)\n",
2084                                                 (long long)vap->va_size);
2085 #endif
2086                                 }
2087                         }
2088
2089                         /*
2090                          * If truncating we have to clean out a portion of
2091                          * the last block on-disk.  We do this in the
2092                          * front-end buffer cache.
2093                          */
2094                         aligned_size = (vap->va_size + (blksize - 1)) &
2095                                        ~(int64_t)(blksize - 1);
2096                         if (truncating && vap->va_size < aligned_size) {
2097                                 struct buf *bp;
2098                                 int offset;
2099
2100                                 aligned_size -= blksize;
2101
2102                                 offset = (int)vap->va_size & (blksize - 1);
2103                                 error = bread(ap->a_vp, aligned_size,
2104                                               blksize, &bp);
2105                                 hammer_ip_frontend_trunc(ip, aligned_size);
2106                                 if (error == 0) {
2107                                         bzero(bp->b_data + offset,
2108                                               blksize - offset);
2109                                         /* must de-cache direct-io offset */
2110                                         bp->b_bio2.bio_offset = NOOFFSET;
2111                                         bdwrite(bp);
2112                                 } else {
2113                                         kprintf("ERROR %d\n", error);
2114                                         brelse(bp);
2115                                 }
2116                         }
2117                         break;
2118                 case VDATABASE:
2119                         if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2120                                 ip->flags |= HAMMER_INODE_TRUNCATED;
2121                                 ip->trunc_off = vap->va_size;
2122                         } else if (ip->trunc_off > vap->va_size) {
2123                                 ip->trunc_off = vap->va_size;
2124                         }
2125                         hammer_ip_frontend_trunc(ip, vap->va_size);
2126                         ip->ino_data.size = vap->va_size;
2127                         ip->ino_data.mtime = trans.time;
2128                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2129                         kflags |= NOTE_ATTRIB;
2130                         break;
2131                 default:
2132                         error = EINVAL;
2133                         goto done;
2134                 }
2135                 break;
2136         }
2137         if (vap->va_atime.tv_sec != VNOVAL) {
2138                 ip->ino_data.atime = hammer_timespec_to_time(&vap->va_atime);
2139                 modflags |= HAMMER_INODE_ATIME;
2140                 kflags |= NOTE_ATTRIB;
2141         }
2142         if (vap->va_mtime.tv_sec != VNOVAL) {
2143                 ip->ino_data.mtime = hammer_timespec_to_time(&vap->va_mtime);
2144                 modflags |= HAMMER_INODE_MTIME;
2145                 kflags |= NOTE_ATTRIB;
2146         }
2147         if (vap->va_mode != (mode_t)VNOVAL) {
2148                 mode_t   cur_mode = ip->ino_data.mode;
2149                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2150                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2151
2152                 error = vop_helper_chmod(ap->a_vp, vap->va_mode, ap->a_cred,
2153                                          cur_uid, cur_gid, &cur_mode);
2154                 if (error == 0 && ip->ino_data.mode != cur_mode) {
2155                         ip->ino_data.mode = cur_mode;
2156                         ip->ino_data.ctime = trans.time;
2157                         modflags |= HAMMER_INODE_DDIRTY;
2158                         kflags |= NOTE_ATTRIB;
2159                 }
2160         }
2161 done:
2162         if (error == 0)
2163                 hammer_modify_inode(ip, modflags);
2164         hammer_done_transaction(&trans);
2165         hammer_knote(ap->a_vp, kflags);
2166         return (error);
2167 }
2168
2169 /*
2170  * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
2171  */
2172 static
2173 int
2174 hammer_vop_nsymlink(struct vop_nsymlink_args *ap)
2175 {
2176         struct hammer_transaction trans;
2177         struct hammer_inode *dip;
2178         struct hammer_inode *nip;
2179         struct nchandle *nch;
2180         hammer_record_t record;
2181         int error;
2182         int bytes;
2183
2184         ap->a_vap->va_type = VLNK;
2185
2186         nch = ap->a_nch;
2187         dip = VTOI(ap->a_dvp);
2188
2189         if (dip->flags & HAMMER_INODE_RO)
2190                 return (EROFS);
2191         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
2192                 return (error);
2193
2194         /*
2195          * Create a transaction to cover the operations we perform.
2196          */
2197         hammer_start_transaction(&trans, dip->hmp);
2198         ++hammer_stats_file_iopsw;
2199
2200         /*
2201          * Create a new filesystem object of the requested type.  The
2202          * returned inode will be referenced but not locked.
2203          */
2204
2205         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
2206                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
2207                                     NULL, &nip);
2208         if (error) {
2209                 hammer_done_transaction(&trans);
2210                 *ap->a_vpp = NULL;
2211                 return (error);
2212         }
2213
2214         /*
2215          * Add a record representing the symlink.  symlink stores the link
2216          * as pure data, not a string, and is no \0 terminated.
2217          */
2218         if (error == 0) {
2219                 bytes = strlen(ap->a_target);
2220
2221                 if (bytes <= HAMMER_INODE_BASESYMLEN) {
2222                         bcopy(ap->a_target, nip->ino_data.ext.symlink, bytes);
2223                 } else {
2224                         record = hammer_alloc_mem_record(nip, bytes);
2225                         record->type = HAMMER_MEM_RECORD_GENERAL;
2226
2227                         record->leaf.base.localization = nip->obj_localization +
2228                                                          HAMMER_LOCALIZE_MISC;
2229                         record->leaf.base.key = HAMMER_FIXKEY_SYMLINK;
2230                         record->leaf.base.rec_type = HAMMER_RECTYPE_FIX;
2231                         record->leaf.data_len = bytes;
2232                         KKASSERT(HAMMER_SYMLINK_NAME_OFF == 0);
2233                         bcopy(ap->a_target, record->data->symlink.name, bytes);
2234                         error = hammer_ip_add_record(&trans, record);
2235                 }
2236
2237                 /*
2238                  * Set the file size to the length of the link.
2239                  */
2240                 if (error == 0) {
2241                         nip->ino_data.size = bytes;
2242                         hammer_modify_inode(nip, HAMMER_INODE_DDIRTY);
2243                 }
2244         }
2245         if (error == 0)
2246                 error = hammer_ip_add_directory(&trans, dip, nch->ncp->nc_name,
2247                                                 nch->ncp->nc_nlen, nip);
2248
2249         /*
2250          * Finish up.
2251          */
2252         if (error) {
2253                 hammer_rel_inode(nip, 0);
2254                 *ap->a_vpp = NULL;
2255         } else {
2256                 error = hammer_get_vnode(nip, ap->a_vpp);
2257                 hammer_rel_inode(nip, 0);
2258                 if (error == 0) {
2259                         cache_setunresolved(ap->a_nch);
2260                         cache_setvp(ap->a_nch, *ap->a_vpp);
2261                         hammer_knote(ap->a_dvp, NOTE_WRITE);
2262                 }
2263         }
2264         hammer_done_transaction(&trans);
2265         return (error);
2266 }
2267
2268 /*
2269  * hammer_vop_nwhiteout { nch, dvp, cred, flags }
2270  */
2271 static
2272 int
2273 hammer_vop_nwhiteout(struct vop_nwhiteout_args *ap)
2274 {
2275         struct hammer_transaction trans;
2276         struct hammer_inode *dip;
2277         int error;
2278
2279         dip = VTOI(ap->a_dvp);
2280
2281         if (hammer_nohistory(dip) == 0 &&
2282             (error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0) {
2283                 return (error);
2284         }
2285
2286         hammer_start_transaction(&trans, dip->hmp);
2287         ++hammer_stats_file_iopsw;
2288         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp,
2289                                 ap->a_cred, ap->a_flags, -1);
2290         hammer_done_transaction(&trans);
2291
2292         return (error);
2293 }
2294
2295 /*
2296  * hammer_vop_ioctl { vp, command, data, fflag, cred }
2297  */
2298 static
2299 int
2300 hammer_vop_ioctl(struct vop_ioctl_args *ap)
2301 {
2302         struct hammer_inode *ip = ap->a_vp->v_data;
2303
2304         ++hammer_stats_file_iopsr;
2305         return(hammer_ioctl(ip, ap->a_command, ap->a_data,
2306                             ap->a_fflag, ap->a_cred));
2307 }
2308
2309 static
2310 int
2311 hammer_vop_mountctl(struct vop_mountctl_args *ap)
2312 {
2313         static const struct mountctl_opt extraopt[] = {
2314                 { HMNT_NOHISTORY,       "nohistory" },
2315                 { HMNT_MASTERID,        "master" },
2316                 { 0, NULL}
2317
2318         };
2319         struct hammer_mount *hmp;
2320         struct mount *mp;
2321         int usedbytes;
2322         int error;
2323
2324         error = 0;
2325         usedbytes = 0;
2326         mp = ap->a_head.a_ops->head.vv_mount;
2327         KKASSERT(mp->mnt_data != NULL);
2328         hmp = (struct hammer_mount *)mp->mnt_data;
2329
2330         switch(ap->a_op) {
2331
2332         case MOUNTCTL_SET_EXPORT:
2333                 if (ap->a_ctllen != sizeof(struct export_args))
2334                         error = EINVAL;
2335                 else
2336                         error = hammer_vfs_export(mp, ap->a_op,
2337                                       (const struct export_args *)ap->a_ctl);
2338                 break;
2339         case MOUNTCTL_MOUNTFLAGS:
2340         {
2341                 /*
2342                  * Call standard mountctl VOP function
2343                  * so we get user mount flags.
2344                  */
2345                 error = vop_stdmountctl(ap);
2346                 if (error)
2347                         break;
2348
2349                 usedbytes = *ap->a_res;
2350
2351                 if (usedbytes > 0 && usedbytes < ap->a_buflen) {
2352                         usedbytes += vfs_flagstostr(hmp->hflags, extraopt, ap->a_buf,
2353                                                     ap->a_buflen - usedbytes,
2354                                                     &error);
2355                 }
2356
2357                 *ap->a_res += usedbytes;
2358                 break;
2359         }
2360         default:
2361                 error = vop_stdmountctl(ap);
2362                 break;
2363         }
2364         return(error);
2365 }
2366
2367 /*
2368  * hammer_vop_strategy { vp, bio }
2369  *
2370  * Strategy call, used for regular file read & write only.  Note that the
2371  * bp may represent a cluster.
2372  *
2373  * To simplify operation and allow better optimizations in the future,
2374  * this code does not make any assumptions with regards to buffer alignment
2375  * or size.
2376  */
2377 static
2378 int
2379 hammer_vop_strategy(struct vop_strategy_args *ap)
2380 {
2381         struct buf *bp;
2382         int error;
2383
2384         bp = ap->a_bio->bio_buf;
2385
2386         switch(bp->b_cmd) {
2387         case BUF_CMD_READ:
2388                 error = hammer_vop_strategy_read(ap);
2389                 break;
2390         case BUF_CMD_WRITE:
2391                 error = hammer_vop_strategy_write(ap);
2392                 break;
2393         default:
2394                 bp->b_error = error = EINVAL;
2395                 bp->b_flags |= B_ERROR;
2396                 biodone(ap->a_bio);
2397                 break;
2398         }
2399         return (error);
2400 }
2401
2402 /*
2403  * Read from a regular file.  Iterate the related records and fill in the
2404  * BIO/BUF.  Gaps are zero-filled.
2405  *
2406  * The support code in hammer_object.c should be used to deal with mixed
2407  * in-memory and on-disk records.
2408  *
2409  * NOTE: Can be called from the cluster code with an oversized buf.
2410  *
2411  * XXX atime update
2412  */
2413 static
2414 int
2415 hammer_vop_strategy_read(struct vop_strategy_args *ap)
2416 {
2417         struct hammer_transaction trans;
2418         struct hammer_inode *ip;
2419         struct hammer_inode *dip;
2420         struct hammer_cursor cursor;
2421         hammer_base_elm_t base;
2422         hammer_off_t disk_offset;
2423         struct bio *bio;
2424         struct bio *nbio;
2425         struct buf *bp;
2426         int64_t rec_offset;
2427         int64_t ran_end;
2428         int64_t tmp64;
2429         int error;
2430         int boff;
2431         int roff;
2432         int n;
2433
2434         bio = ap->a_bio;
2435         bp = bio->bio_buf;
2436         ip = ap->a_vp->v_data;
2437
2438         /*
2439          * The zone-2 disk offset may have been set by the cluster code via
2440          * a BMAP operation, or else should be NOOFFSET.
2441          *
2442          * Checking the high bits for a match against zone-2 should suffice.
2443          */
2444         nbio = push_bio(bio);
2445         if ((nbio->bio_offset & HAMMER_OFF_ZONE_MASK) ==
2446             HAMMER_ZONE_LARGE_DATA) {
2447                 error = hammer_io_direct_read(ip->hmp, nbio, NULL);
2448                 return (error);
2449         }
2450
2451         /*
2452          * Well, that sucked.  Do it the hard way.  If all the stars are
2453          * aligned we may still be able to issue a direct-read.
2454          */
2455         hammer_simple_transaction(&trans, ip->hmp);
2456         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2457
2458         /*
2459          * Key range (begin and end inclusive) to scan.  Note that the key's
2460          * stored in the actual records represent BASE+LEN, not BASE.  The
2461          * first record containing bio_offset will have a key > bio_offset.
2462          */
2463         cursor.key_beg.localization = ip->obj_localization +
2464                                       HAMMER_LOCALIZE_MISC;
2465         cursor.key_beg.obj_id = ip->obj_id;
2466         cursor.key_beg.create_tid = 0;
2467         cursor.key_beg.delete_tid = 0;
2468         cursor.key_beg.obj_type = 0;
2469         cursor.key_beg.key = bio->bio_offset + 1;
2470         cursor.asof = ip->obj_asof;
2471         cursor.flags |= HAMMER_CURSOR_ASOF;
2472
2473         cursor.key_end = cursor.key_beg;
2474         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2475 #if 0
2476         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
2477                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
2478                 cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
2479                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2480         } else
2481 #endif
2482         {
2483                 ran_end = bio->bio_offset + bp->b_bufsize;
2484                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2485                 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2486                 tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2487                 if (tmp64 < ran_end)
2488                         cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2489                 else
2490                         cursor.key_end.key = ran_end + MAXPHYS + 1;
2491         }
2492         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2493
2494         error = hammer_ip_first(&cursor);
2495         boff = 0;
2496
2497         while (error == 0) {
2498                 /*
2499                  * Get the base file offset of the record.  The key for
2500                  * data records is (base + bytes) rather then (base).
2501                  */
2502                 base = &cursor.leaf->base;
2503                 rec_offset = base->key - cursor.leaf->data_len;
2504
2505                 /*
2506                  * Calculate the gap, if any, and zero-fill it.
2507                  *
2508                  * n is the offset of the start of the record verses our
2509                  * current seek offset in the bio.
2510                  */
2511                 n = (int)(rec_offset - (bio->bio_offset + boff));
2512                 if (n > 0) {
2513                         if (n > bp->b_bufsize - boff)
2514                                 n = bp->b_bufsize - boff;
2515                         bzero((char *)bp->b_data + boff, n);
2516                         boff += n;
2517                         n = 0;
2518                 }
2519
2520                 /*
2521                  * Calculate the data offset in the record and the number
2522                  * of bytes we can copy.
2523                  *
2524                  * There are two degenerate cases.  First, boff may already
2525                  * be at bp->b_bufsize.  Secondly, the data offset within
2526                  * the record may exceed the record's size.
2527                  */
2528                 roff = -n;
2529                 rec_offset += roff;
2530                 n = cursor.leaf->data_len - roff;
2531                 if (n <= 0) {
2532                         kprintf("strategy_read: bad n=%d roff=%d\n", n, roff);
2533                         n = 0;
2534                 } else if (n > bp->b_bufsize - boff) {
2535                         n = bp->b_bufsize - boff;
2536                 }
2537
2538                 /*
2539                  * Deal with cached truncations.  This cool bit of code
2540                  * allows truncate()/ftruncate() to avoid having to sync
2541                  * the file.
2542                  *
2543                  * If the frontend is truncated then all backend records are
2544                  * subject to the frontend's truncation.
2545                  *
2546                  * If the backend is truncated then backend records on-disk
2547                  * (but not in-memory) are subject to the backend's
2548                  * truncation.  In-memory records owned by the backend
2549                  * represent data written after the truncation point on the
2550                  * backend and must not be truncated.
2551                  *
2552                  * Truncate operations deal with frontend buffer cache
2553                  * buffers and frontend-owned in-memory records synchronously.
2554                  */
2555                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2556                         if (hammer_cursor_ondisk(&cursor) ||
2557                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
2558                                 if (ip->trunc_off <= rec_offset)
2559                                         n = 0;
2560                                 else if (ip->trunc_off < rec_offset + n)
2561                                         n = (int)(ip->trunc_off - rec_offset);
2562                         }
2563                 }
2564                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2565                         if (hammer_cursor_ondisk(&cursor)) {
2566                                 if (ip->sync_trunc_off <= rec_offset)
2567                                         n = 0;
2568                                 else if (ip->sync_trunc_off < rec_offset + n)
2569                                         n = (int)(ip->sync_trunc_off - rec_offset);
2570                         }
2571                 }
2572
2573                 /*
2574                  * Try to issue a direct read into our bio if possible,
2575                  * otherwise resolve the element data into a hammer_buffer
2576                  * and copy.
2577                  *
2578                  * The buffer on-disk should be zerod past any real
2579                  * truncation point, but may not be for any synthesized
2580                  * truncation point from above.
2581                  */
2582                 disk_offset = cursor.leaf->data_offset + roff;
2583                 if (boff == 0 && n == bp->b_bufsize &&
2584                     hammer_cursor_ondisk(&cursor) &&
2585                     (disk_offset & HAMMER_BUFMASK) == 0) {
2586                         KKASSERT((disk_offset & HAMMER_OFF_ZONE_MASK) ==
2587                                  HAMMER_ZONE_LARGE_DATA);
2588                         nbio->bio_offset = disk_offset;
2589                         error = hammer_io_direct_read(trans.hmp, nbio,
2590                                                       cursor.leaf);
2591                         goto done;
2592                 } else if (n) {
2593                         error = hammer_ip_resolve_data(&cursor);
2594                         if (error == 0) {
2595                                 bcopy((char *)cursor.data + roff,
2596                                       (char *)bp->b_data + boff, n);
2597                         }
2598                 }
2599                 if (error)
2600                         break;
2601
2602                 /*
2603                  * Iterate until we have filled the request.
2604                  */
2605                 boff += n;
2606                 if (boff == bp->b_bufsize)
2607                         break;
2608                 error = hammer_ip_next(&cursor);
2609         }
2610
2611         /*
2612          * There may have been a gap after the last record
2613          */
2614         if (error == ENOENT)
2615                 error = 0;
2616         if (error == 0 && boff != bp->b_bufsize) {
2617                 KKASSERT(boff < bp->b_bufsize);
2618                 bzero((char *)bp->b_data + boff, bp->b_bufsize - boff);
2619                 /* boff = bp->b_bufsize; */
2620         }
2621         bp->b_resid = 0;
2622         bp->b_error = error;
2623         if (error)
2624                 bp->b_flags |= B_ERROR;
2625         biodone(ap->a_bio);
2626
2627 done:
2628         /*
2629          * Cache the b-tree node for the last data read in cache[1].
2630          *
2631          * If we hit the file EOF then also cache the node in the
2632          * governing director's cache[3], it will be used to initialize
2633          * the inode's cache[1] for any inodes looked up via the directory.
2634          *
2635          * This doesn't reduce disk accesses since the B-Tree chain is
2636          * likely cached, but it does reduce cpu overhead when looking
2637          * up file offsets for cpdup/tar/cpio style iterations.
2638          */
2639         if (cursor.node)
2640                 hammer_cache_node(&ip->cache[1], cursor.node);
2641         if (ran_end >= ip->ino_data.size) {
2642                 dip = hammer_find_inode(&trans, ip->ino_data.parent_obj_id,
2643                                         ip->obj_asof, ip->obj_localization);
2644                 if (dip) {
2645                         hammer_cache_node(&dip->cache[3], cursor.node);
2646                         hammer_rel_inode(dip, 0);
2647                 }
2648         }
2649         hammer_done_cursor(&cursor);
2650         hammer_done_transaction(&trans);
2651         return(error);
2652 }
2653
2654 /*
2655  * BMAP operation - used to support cluster_read() only.
2656  *
2657  * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
2658  *
2659  * This routine may return EOPNOTSUPP if the opration is not supported for
2660  * the specified offset.  The contents of the pointer arguments do not
2661  * need to be initialized in that case.
2662  *
2663  * If a disk address is available and properly aligned return 0 with
2664  * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
2665  * to the run-length relative to that offset.  Callers may assume that
2666  * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
2667  * large, so return EOPNOTSUPP if it is not sufficiently large.
2668  */
2669 static
2670 int
2671 hammer_vop_bmap(struct vop_bmap_args *ap)
2672 {
2673         struct hammer_transaction trans;
2674         struct hammer_inode *ip;
2675         struct hammer_cursor cursor;
2676         hammer_base_elm_t base;
2677         int64_t rec_offset;
2678         int64_t ran_end;
2679         int64_t tmp64;
2680         int64_t base_offset;
2681         int64_t base_disk_offset;
2682         int64_t last_offset;
2683         hammer_off_t last_disk_offset;
2684         hammer_off_t disk_offset;
2685         int     rec_len;
2686         int     error;
2687         int     blksize;
2688
2689         ++hammer_stats_file_iopsr;
2690         ip = ap->a_vp->v_data;
2691
2692         /*
2693          * We can only BMAP regular files.  We can't BMAP database files,
2694          * directories, etc.
2695          */
2696         if (ip->ino_data.obj_type != HAMMER_OBJTYPE_REGFILE)
2697                 return(EOPNOTSUPP);
2698
2699         /*
2700          * bmap is typically called with runp/runb both NULL when used
2701          * for writing.  We do not support BMAP for writing atm.
2702          */
2703         if (ap->a_cmd != BUF_CMD_READ)
2704                 return(EOPNOTSUPP);
2705
2706         /*
2707          * Scan the B-Tree to acquire blockmap addresses, then translate
2708          * to raw addresses.
2709          */
2710         hammer_simple_transaction(&trans, ip->hmp);
2711 #if 0
2712         kprintf("bmap_beg %016llx ip->cache %p\n",
2713                 (long long)ap->a_loffset, ip->cache[1]);
2714 #endif
2715         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2716
2717         /*
2718          * Key range (begin and end inclusive) to scan.  Note that the key's
2719          * stored in the actual records represent BASE+LEN, not BASE.  The
2720          * first record containing bio_offset will have a key > bio_offset.
2721          */
2722         cursor.key_beg.localization = ip->obj_localization +
2723                                       HAMMER_LOCALIZE_MISC;
2724         cursor.key_beg.obj_id = ip->obj_id;
2725         cursor.key_beg.create_tid = 0;
2726         cursor.key_beg.delete_tid = 0;
2727         cursor.key_beg.obj_type = 0;
2728         if (ap->a_runb)
2729                 cursor.key_beg.key = ap->a_loffset - MAXPHYS + 1;
2730         else
2731                 cursor.key_beg.key = ap->a_loffset + 1;
2732         if (cursor.key_beg.key < 0)
2733                 cursor.key_beg.key = 0;
2734         cursor.asof = ip->obj_asof;
2735         cursor.flags |= HAMMER_CURSOR_ASOF;
2736
2737         cursor.key_end = cursor.key_beg;
2738         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2739
2740         ran_end = ap->a_loffset + MAXPHYS;
2741         cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2742         cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2743         tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2744         if (tmp64 < ran_end)
2745                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2746         else
2747                 cursor.key_end.key = ran_end + MAXPHYS + 1;
2748
2749         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2750
2751         error = hammer_ip_first(&cursor);
2752         base_offset = last_offset = 0;
2753         base_disk_offset = last_disk_offset = 0;
2754
2755         while (error == 0) {
2756                 /*
2757                  * Get the base file offset of the record.  The key for
2758                  * data records is (base + bytes) rather then (base).
2759                  *
2760                  * NOTE: rec_offset + rec_len may exceed the end-of-file.
2761                  * The extra bytes should be zero on-disk and the BMAP op
2762                  * should still be ok.
2763                  */
2764                 base = &cursor.leaf->base;
2765                 rec_offset = base->key - cursor.leaf->data_len;
2766                 rec_len    = cursor.leaf->data_len;
2767
2768                 /*
2769                  * Incorporate any cached truncation.
2770                  *
2771                  * NOTE: Modifications to rec_len based on synthesized
2772                  * truncation points remove the guarantee that any extended
2773                  * data on disk is zero (since the truncations may not have
2774                  * taken place on-media yet).
2775                  */
2776                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2777                         if (hammer_cursor_ondisk(&cursor) ||
2778                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
2779                                 if (ip->trunc_off <= rec_offset)
2780                                         rec_len = 0;
2781                                 else if (ip->trunc_off < rec_offset + rec_len)
2782                                         rec_len = (int)(ip->trunc_off - rec_offset);
2783                         }
2784                 }
2785                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2786                         if (hammer_cursor_ondisk(&cursor)) {
2787                                 if (ip->sync_trunc_off <= rec_offset)
2788                                         rec_len = 0;
2789                                 else if (ip->sync_trunc_off < rec_offset + rec_len)
2790                                         rec_len = (int)(ip->sync_trunc_off - rec_offset);
2791                         }
2792                 }
2793
2794                 /*
2795                  * Accumulate information.  If we have hit a discontiguous
2796                  * block reset base_offset unless we are already beyond the
2797                  * requested offset.  If we are, that's it, we stop.
2798                  */
2799                 if (error)
2800                         break;
2801                 if (hammer_cursor_ondisk(&cursor)) {
2802                         disk_offset = cursor.leaf->data_offset;
2803                         if (rec_offset != last_offset ||
2804                             disk_offset != last_disk_offset) {
2805                                 if (rec_offset > ap->a_loffset)
2806                                         break;
2807                                 base_offset = rec_offset;
2808                                 base_disk_offset = disk_offset;
2809                         }
2810                         last_offset = rec_offset + rec_len;
2811                         last_disk_offset = disk_offset + rec_len;
2812                 }
2813                 error = hammer_ip_next(&cursor);
2814         }
2815
2816 #if 0
2817         kprintf("BMAP %016llx:  %016llx - %016llx\n",
2818                 (long long)ap->a_loffset,
2819                 (long long)base_offset,
2820                 (long long)last_offset);
2821         kprintf("BMAP %16s:  %016llx - %016llx\n", "",
2822                 (long long)base_disk_offset,
2823                 (long long)last_disk_offset);
2824 #endif
2825
2826         if (cursor.node) {
2827                 hammer_cache_node(&ip->cache[1], cursor.node);
2828 #if 0
2829                 kprintf("bmap_end2 %016llx ip->cache %p\n",
2830                         (long long)ap->a_loffset, ip->cache[1]);
2831 #endif
2832         }
2833         hammer_done_cursor(&cursor);
2834         hammer_done_transaction(&trans);
2835
2836         /*
2837          * If we couldn't find any records or the records we did find were
2838          * all behind the requested offset, return failure.  A forward
2839          * truncation can leave a hole w/ no on-disk records.
2840          */
2841         if (last_offset == 0 || last_offset < ap->a_loffset)
2842                 return (EOPNOTSUPP);
2843
2844         /*
2845          * Figure out the block size at the requested offset and adjust
2846          * our limits so the cluster_read() does not create inappropriately
2847          * sized buffer cache buffers.
2848          */
2849         blksize = hammer_blocksize(ap->a_loffset);
2850         if (hammer_blocksize(base_offset) != blksize) {
2851                 base_offset = hammer_blockdemarc(base_offset, ap->a_loffset);
2852         }
2853         if (last_offset != ap->a_loffset &&
2854             hammer_blocksize(last_offset - 1) != blksize) {
2855                 last_offset = hammer_blockdemarc(ap->a_loffset,
2856                                                  last_offset - 1);
2857         }
2858
2859         /*
2860          * Returning EOPNOTSUPP simply prevents the direct-IO optimization
2861          * from occuring.
2862          */
2863         disk_offset = base_disk_offset + (ap->a_loffset - base_offset);
2864
2865         if ((disk_offset & HAMMER_OFF_ZONE_MASK) != HAMMER_ZONE_LARGE_DATA) {
2866                 /*
2867                  * Only large-data zones can be direct-IOd
2868                  */
2869                 error = EOPNOTSUPP;
2870         } else if ((disk_offset & HAMMER_BUFMASK) ||
2871                    (last_offset - ap->a_loffset) < blksize) {
2872                 /*
2873                  * doffsetp is not aligned or the forward run size does
2874                  * not cover a whole buffer, disallow the direct I/O.
2875                  */
2876                 error = EOPNOTSUPP;
2877         } else {
2878                 /*
2879                  * We're good.
2880                  */
2881                 *ap->a_doffsetp = disk_offset;
2882                 if (ap->a_runb) {
2883                         *ap->a_runb = ap->a_loffset - base_offset;
2884                         KKASSERT(*ap->a_runb >= 0);
2885                 }
2886                 if (ap->a_runp) {
2887                         *ap->a_runp = last_offset - ap->a_loffset;
2888                         KKASSERT(*ap->a_runp >= 0);
2889                 }
2890                 error = 0;
2891         }
2892         return(error);
2893 }
2894
2895 /*
2896  * Write to a regular file.   Because this is a strategy call the OS is
2897  * trying to actually get data onto the media.
2898  */
2899 static
2900 int
2901 hammer_vop_strategy_write(struct vop_strategy_args *ap)
2902 {
2903         hammer_record_t record;
2904         hammer_mount_t hmp;
2905         hammer_inode_t ip;
2906         struct bio *bio;
2907         struct buf *bp;
2908         int blksize;
2909         int bytes;
2910         int error;
2911
2912         bio = ap->a_bio;
2913         bp = bio->bio_buf;
2914         ip = ap->a_vp->v_data;
2915         hmp = ip->hmp;
2916
2917         blksize = hammer_blocksize(bio->bio_offset);
2918         KKASSERT(bp->b_bufsize == blksize);
2919
2920         if (ip->flags & HAMMER_INODE_RO) {
2921                 bp->b_error = EROFS;
2922                 bp->b_flags |= B_ERROR;
2923                 biodone(ap->a_bio);
2924                 return(EROFS);
2925         }
2926
2927         /*
2928          * Interlock with inode destruction (no in-kernel or directory
2929          * topology visibility).  If we queue new IO while trying to
2930          * destroy the inode we can deadlock the vtrunc call in
2931          * hammer_inode_unloadable_check().
2932          *
2933          * Besides, there's no point flushing a bp associated with an
2934          * inode that is being destroyed on-media and has no kernel
2935          * references.
2936          */
2937         if ((ip->flags | ip->sync_flags) &
2938             (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) {
2939                 bp->b_resid = 0;
2940                 biodone(ap->a_bio);
2941                 return(0);
2942         }
2943
2944         /*
2945          * Reserve space and issue a direct-write from the front-end.
2946          * NOTE: The direct_io code will hammer_bread/bcopy smaller
2947          * allocations.
2948          *
2949          * An in-memory record will be installed to reference the storage
2950          * until the flusher can get to it.
2951          *
2952          * Since we own the high level bio the front-end will not try to
2953          * do a direct-read until the write completes.
2954          *
2955          * NOTE: The only time we do not reserve a full-sized buffers
2956          * worth of data is if the file is small.  We do not try to
2957          * allocate a fragment (from the small-data zone) at the end of
2958          * an otherwise large file as this can lead to wildly separated
2959          * data.
2960          */
2961         KKASSERT((bio->bio_offset & HAMMER_BUFMASK) == 0);
2962         KKASSERT(bio->bio_offset < ip->ino_data.size);
2963         if (bio->bio_offset || ip->ino_data.size > HAMMER_BUFSIZE / 2)
2964                 bytes = bp->b_bufsize;
2965         else
2966                 bytes = ((int)ip->ino_data.size + 15) & ~15;
2967
2968         record = hammer_ip_add_bulk(ip, bio->bio_offset, bp->b_data,
2969                                     bytes, &error);
2970         if (record) {
2971                 hammer_io_direct_write(hmp, record, bio);
2972                 if (ip->rsv_recs > 1 && hmp->rsv_recs > hammer_limit_recs)
2973                         hammer_flush_inode(ip, 0);
2974         } else {
2975                 bp->b_bio2.bio_offset = NOOFFSET;
2976                 bp->b_error = error;
2977                 bp->b_flags |= B_ERROR;
2978                 biodone(ap->a_bio);
2979         }
2980         return(error);
2981 }
2982
2983 /*
2984  * dounlink - disconnect a directory entry
2985  *
2986  * XXX whiteout support not really in yet
2987  */
2988 static int
2989 hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
2990                 struct vnode *dvp, struct ucred *cred,
2991                 int flags, int isdir)
2992 {
2993         struct namecache *ncp;
2994         hammer_inode_t dip;
2995         hammer_inode_t ip;
2996         struct hammer_cursor cursor;
2997         int64_t namekey;
2998         u_int32_t max_iterations;
2999         int nlen, error;
3000
3001         /*
3002          * Calculate the namekey and setup the key range for the scan.  This
3003          * works kinda like a chained hash table where the lower 32 bits
3004          * of the namekey synthesize the chain.
3005          *
3006          * The key range is inclusive of both key_beg and key_end.
3007          */
3008         dip = VTOI(dvp);
3009         ncp = nch->ncp;
3010
3011         if (dip->flags & HAMMER_INODE_RO)
3012                 return (EROFS);
3013
3014         namekey = hammer_directory_namekey(dip, ncp->nc_name, ncp->nc_nlen,
3015                                            &max_iterations);
3016 retry:
3017         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
3018         cursor.key_beg.localization = dip->obj_localization +
3019                                       hammer_dir_localization(dip);
3020         cursor.key_beg.obj_id = dip->obj_id;
3021         cursor.key_beg.key = namekey;
3022         cursor.key_beg.create_tid = 0;
3023         cursor.key_beg.delete_tid = 0;
3024         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
3025         cursor.key_beg.obj_type = 0;
3026
3027         cursor.key_end = cursor.key_beg;
3028         cursor.key_end.key += max_iterations;
3029         cursor.asof = dip->obj_asof;
3030         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
3031
3032         /*
3033          * Scan all matching records (the chain), locate the one matching
3034          * the requested path component.  info->last_error contains the
3035          * error code on search termination and could be 0, ENOENT, or
3036          * something else.
3037          *
3038          * The hammer_ip_*() functions merge in-memory records with on-disk
3039          * records for the purposes of the search.
3040          */
3041         error = hammer_ip_first(&cursor);
3042
3043         while (error == 0) {
3044                 error = hammer_ip_resolve_data(&cursor);
3045                 if (error)
3046                         break;
3047                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
3048                 KKASSERT(nlen > 0);
3049                 if (ncp->nc_nlen == nlen &&
3050                     bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
3051                         break;
3052                 }
3053                 error = hammer_ip_next(&cursor);
3054         }
3055
3056         /*
3057          * If all is ok we have to get the inode so we can adjust nlinks.
3058          * To avoid a deadlock with the flusher we must release the inode
3059          * lock on the directory when acquiring the inode for the entry.
3060          *
3061          * If the target is a directory, it must be empty.
3062          */
3063         if (error == 0) {
3064                 hammer_unlock(&cursor.ip->lock);
3065                 ip = hammer_get_inode(trans, dip, cursor.data->entry.obj_id,
3066                                       dip->hmp->asof,
3067                                       cursor.data->entry.localization,
3068                                       0, &error);
3069                 hammer_lock_sh(&cursor.ip->lock);
3070                 if (error == ENOENT) {
3071                         kprintf("HAMMER: WARNING: Removing "
3072                                 "dirent w/missing inode \"%s\"\n"
3073                                 "\tobj_id = %016llx\n",
3074                                 ncp->nc_name,
3075                                 (long long)cursor.data->entry.obj_id);
3076                         error = 0;
3077                 }
3078
3079                 /*
3080                  * If isdir >= 0 we validate that the entry is or is not a
3081                  * directory.  If isdir < 0 we don't care.
3082                  */
3083                 if (error == 0 && isdir >= 0 && ip) {
3084                         if (isdir &&
3085                             ip->ino_data.obj_type != HAMMER_OBJTYPE_DIRECTORY) {
3086                                 error = ENOTDIR;
3087                         } else if (isdir == 0 &&
3088                             ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
3089                                 error = EISDIR;
3090                         }
3091                 }
3092
3093                 /*
3094                  * If we are trying to remove a directory the directory must
3095                  * be empty.
3096                  *
3097                  * The check directory code can loop and deadlock/retry.  Our
3098                  * own cursor's node locks must be released to avoid a 3-way
3099                  * deadlock with the flusher if the check directory code
3100                  * blocks.
3101                  *
3102                  * If any changes whatsoever have been made to the cursor
3103                  * set EDEADLK and retry.
3104                  *
3105                  * WARNING: See warnings in hammer_unlock_cursor()
3106                  *          function.
3107                  */
3108                 if (error == 0 && ip && ip->ino_data.obj_type ==
3109                                         HAMMER_OBJTYPE_DIRECTORY) {
3110                         hammer_unlock_cursor(&cursor);
3111                         error = hammer_ip_check_directory_empty(trans, ip);
3112                         hammer_lock_cursor(&cursor);
3113                         if (cursor.flags & HAMMER_CURSOR_RETEST) {
3114                                 kprintf("HAMMER: Warning: avoided deadlock "
3115                                         "on rmdir '%s'\n",
3116                                         ncp->nc_name);
3117                                 error = EDEADLK;
3118                         }
3119                 }
3120
3121                 /*
3122                  * Delete the directory entry.
3123                  *
3124                  * WARNING: hammer_ip_del_directory() may have to terminate
3125                  * the cursor to avoid a deadlock.  It is ok to call
3126                  * hammer_done_cursor() twice.
3127                  */
3128                 if (error == 0) {
3129                         error = hammer_ip_del_directory(trans, &cursor,
3130                                                         dip, ip);
3131                 }
3132                 hammer_done_cursor(&cursor);
3133                 if (error == 0) {
3134                         cache_setunresolved(nch);
3135                         cache_setvp(nch, NULL);
3136                         /* XXX locking */
3137                         if (ip && ip->vp) {
3138                                 hammer_knote(ip->vp, NOTE_DELETE);
3139                                 cache_inval_vp(ip->vp, CINV_DESTROY);
3140                         }
3141                 }
3142                 if (ip)
3143                         hammer_rel_inode(ip, 0);
3144         } else {
3145                 hammer_done_cursor(&cursor);
3146         }
3147         if (error == EDEADLK)
3148                 goto retry;
3149
3150         return (error);
3151 }
3152
3153 /************************************************************************
3154  *                          FIFO AND SPECFS OPS                         *
3155  ************************************************************************
3156  *
3157  */
3158
3159 static int
3160 hammer_vop_fifoclose (struct vop_close_args *ap)
3161 {
3162         /* XXX update itimes */
3163         return (VOCALL(&fifo_vnode_vops, &ap->a_head));
3164 }
3165
3166 static int
3167 hammer_vop_fiforead (struct vop_read_args *ap)
3168 {
3169         int error;
3170
3171         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3172         /* XXX update access time */
3173         return (error);
3174 }
3175
3176 static int
3177 hammer_vop_fifowrite (struct vop_write_args *ap)
3178 {
3179         int error;
3180
3181         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3182         /* XXX update access time */
3183         return (error);
3184 }
3185
3186 static
3187 int
3188 hammer_vop_fifokqfilter(struct vop_kqfilter_args *ap)
3189 {
3190         int error;
3191
3192         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3193         if (error)
3194                 error = hammer_vop_kqfilter(ap);
3195         return(error);
3196 }
3197
3198 /************************************************************************
3199  *                          KQFILTER OPS                                *
3200  ************************************************************************
3201  *
3202  */
3203 static void filt_hammerdetach(struct knote *kn);
3204 static int filt_hammerread(struct knote *kn, long hint);
3205 static int filt_hammerwrite(struct knote *kn, long hint);
3206 static int filt_hammervnode(struct knote *kn, long hint);
3207
3208 static struct filterops hammerread_filtops =
3209         { 1, NULL, filt_hammerdetach, filt_hammerread };
3210 static struct filterops hammerwrite_filtops =
3211         { 1, NULL, filt_hammerdetach, filt_hammerwrite };
3212 static struct filterops hammervnode_filtops =
3213         { 1, NULL, filt_hammerdetach, filt_hammervnode };
3214
3215 static
3216 int
3217 hammer_vop_kqfilter(struct vop_kqfilter_args *ap)
3218 {
3219         struct vnode *vp = ap->a_vp;
3220         struct knote *kn = ap->a_kn;
3221         lwkt_tokref vlock;
3222
3223         switch (kn->kn_filter) {
3224         case EVFILT_READ:
3225                 kn->kn_fop = &hammerread_filtops;
3226                 break;
3227         case EVFILT_WRITE:
3228                 kn->kn_fop = &hammerwrite_filtops;
3229                 break;
3230         case EVFILT_VNODE:
3231                 kn->kn_fop = &hammervnode_filtops;
3232                 break;
3233         default:
3234                 return (1);
3235         }
3236
3237         kn->kn_hook = (caddr_t)vp;
3238
3239         lwkt_gettoken(&vlock, &vp->v_token);
3240         SLIST_INSERT_HEAD(&vp->v_pollinfo.vpi_selinfo.si_note, kn, kn_selnext);
3241         lwkt_reltoken(&vlock);
3242
3243         return(0);
3244 }
3245
3246 static void
3247 filt_hammerdetach(struct knote *kn)
3248 {
3249         struct vnode *vp = (void *)kn->kn_hook;
3250         lwkt_tokref vlock;
3251
3252         lwkt_gettoken(&vlock, &vp->v_token);
3253         SLIST_REMOVE(&vp->v_pollinfo.vpi_selinfo.si_note,
3254                      kn, knote, kn_selnext);
3255         lwkt_reltoken(&vlock);
3256 }
3257
3258 static int
3259 filt_hammerread(struct knote *kn, long hint)
3260 {
3261         struct vnode *vp = (void *)kn->kn_hook;
3262         hammer_inode_t ip = VTOI(vp);
3263
3264         if (hint == NOTE_REVOKE) {
3265                 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
3266                 return(1);
3267         }
3268         kn->kn_data = ip->ino_data.size - kn->kn_fp->f_offset;
3269         return (kn->kn_data != 0);
3270 }
3271
3272 static int
3273 filt_hammerwrite(struct knote *kn, long hint)
3274 {
3275         if (hint == NOTE_REVOKE)
3276                 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
3277         kn->kn_data = 0;
3278         return (1);
3279 }
3280
3281 static int
3282 filt_hammervnode(struct knote *kn, long hint)
3283 {
3284         if (kn->kn_sfflags & hint)
3285                 kn->kn_fflags |= hint;
3286         if (hint == NOTE_REVOKE) {
3287                 kn->kn_flags |= EV_EOF;
3288                 return (1);
3289         }
3290         return (kn->kn_fflags != 0);
3291 }
3292