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 1.82 2008/07/07 03:49:51 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 <vm/vm_extern.h>
  48 #include <vfs/fifofs/fifo.h>
  49 #include "hammer.h"
  50
  51 /*
  52  * USERFS VNOPS
  53  */
  54 /*static int hammer_vop_vnoperate(struct vop_generic_args *);*/
  55 static int hammer_vop_fsync(struct vop_fsync_args *);
  56 static int hammer_vop_read(struct vop_read_args *);
  57 static int hammer_vop_write(struct vop_write_args *);
  58 static int hammer_vop_access(struct vop_access_args *);
  59 static int hammer_vop_advlock(struct vop_advlock_args *);
  60 static int hammer_vop_close(struct vop_close_args *);
  61 static int hammer_vop_ncreate(struct vop_ncreate_args *);
  62 static int hammer_vop_getattr(struct vop_getattr_args *);
  63 static int hammer_vop_nresolve(struct vop_nresolve_args *);
  64 static int hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *);
  65 static int hammer_vop_nlink(struct vop_nlink_args *);
  66 static int hammer_vop_nmkdir(struct vop_nmkdir_args *);
  67 static int hammer_vop_nmknod(struct vop_nmknod_args *);
  68 static int hammer_vop_open(struct vop_open_args *);
  69 static int hammer_vop_pathconf(struct vop_pathconf_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_setattr(struct vop_setattr_args *);
  77 static int hammer_vop_strategy(struct vop_strategy_args *);
  78 static int hammer_vop_bmap(struct vop_bmap_args *ap);
  79 static int hammer_vop_nsymlink(struct vop_nsymlink_args *);
  80 static int hammer_vop_nwhiteout(struct vop_nwhiteout_args *);
  81 static int hammer_vop_ioctl(struct vop_ioctl_args *);
  82 static int hammer_vop_mountctl(struct vop_mountctl_args *);
  83
  84 static int hammer_vop_fifoclose (struct vop_close_args *);
  85 static int hammer_vop_fiforead (struct vop_read_args *);
  86 static int hammer_vop_fifowrite (struct vop_write_args *);
  87
  88 static int hammer_vop_specclose (struct vop_close_args *);
  89 static int hammer_vop_specread (struct vop_read_args *);
  90 static int hammer_vop_specwrite (struct vop_write_args *);
  91
  92 struct vop_ops hammer_vnode_vops = {
  93         .vop_default =          vop_defaultop,
  94         .vop_fsync =            hammer_vop_fsync,
  95         .vop_getpages =         vop_stdgetpages,
  96         .vop_putpages =         vop_stdputpages,
  97         .vop_read =             hammer_vop_read,
  98         .vop_write =            hammer_vop_write,
  99         .vop_access =           hammer_vop_access,
 100         .vop_advlock =          hammer_vop_advlock,
 101         .vop_close =            hammer_vop_close,
 102         .vop_ncreate =          hammer_vop_ncreate,
 103         .vop_getattr =          hammer_vop_getattr,
 104         .vop_inactive =         hammer_vop_inactive,
 105         .vop_reclaim =          hammer_vop_reclaim,
 106         .vop_nresolve =         hammer_vop_nresolve,
 107         .vop_nlookupdotdot =    hammer_vop_nlookupdotdot,
 108         .vop_nlink =            hammer_vop_nlink,
 109         .vop_nmkdir =           hammer_vop_nmkdir,
 110         .vop_nmknod =           hammer_vop_nmknod,
 111         .vop_open =             hammer_vop_open,
 112         .vop_pathconf =         hammer_vop_pathconf,
 113         .vop_print =            hammer_vop_print,
 114         .vop_readdir =          hammer_vop_readdir,
 115         .vop_readlink =         hammer_vop_readlink,
 116         .vop_nremove =          hammer_vop_nremove,
 117         .vop_nrename =          hammer_vop_nrename,
 118         .vop_nrmdir =           hammer_vop_nrmdir,
 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 };
 127
 128 struct vop_ops hammer_spec_vops = {
 129         .vop_default =          spec_vnoperate,
 130         .vop_fsync =            hammer_vop_fsync,
 131         .vop_read =             hammer_vop_specread,
 132         .vop_write =            hammer_vop_specwrite,
 133         .vop_access =           hammer_vop_access,
 134         .vop_close =            hammer_vop_specclose,
 135         .vop_getattr =          hammer_vop_getattr,
 136         .vop_inactive =         hammer_vop_inactive,
 137         .vop_reclaim =          hammer_vop_reclaim,
 138         .vop_setattr =          hammer_vop_setattr
 139 };
 140
 141 struct vop_ops hammer_fifo_vops = {
 142         .vop_default =          fifo_vnoperate,
 143         .vop_fsync =            hammer_vop_fsync,
 144         .vop_read =             hammer_vop_fiforead,
 145         .vop_write =            hammer_vop_fifowrite,
 146         .vop_access =           hammer_vop_access,
 147         .vop_close =            hammer_vop_fifoclose,
 148         .vop_getattr =          hammer_vop_getattr,
 149         .vop_inactive =         hammer_vop_inactive,
 150         .vop_reclaim =          hammer_vop_reclaim,
 151         .vop_setattr =          hammer_vop_setattr
 152 };
 153
 154 #ifdef DEBUG_TRUNCATE
 155 struct hammer_inode *HammerTruncIp;
 156 #endif
 157
 158 static int hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
 159                            struct vnode *dvp, struct ucred *cred, int flags);
 160 static int hammer_vop_strategy_read(struct vop_strategy_args *ap);
 161 static int hammer_vop_strategy_write(struct vop_strategy_args *ap);
 162
 163 #if 0
 164 static
 165 int
 166 hammer_vop_vnoperate(struct vop_generic_args *)
 167 {
 168         return (VOCALL(&hammer_vnode_vops, ap));
 169 }
 170 #endif
 171
 172 /*
 173  * hammer_vop_fsync { vp, waitfor }
 174  *
 175  * fsync() an inode to disk and wait for it to be completely committed
 176  * such that the information would not be undone if a crash occured after
 177  * return.
 178  */
 179 static
 180 int
 181 hammer_vop_fsync(struct vop_fsync_args *ap)
 182 {
 183         hammer_inode_t ip = VTOI(ap->a_vp);
 184
 185         vfsync(ap->a_vp, ap->a_waitfor, 1, NULL, NULL);
 186         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 187         if (ap->a_waitfor == MNT_WAIT)
 188                 hammer_wait_inode(ip);
 189         return (ip->error);
 190 }
 191
 192 /*
 193  * hammer_vop_read { vp, uio, ioflag, cred }
 194  */
 195 static
 196 int
 197 hammer_vop_read(struct vop_read_args *ap)
 198 {
 199         struct hammer_transaction trans;
 200         hammer_inode_t ip;
 201         off_t offset;
 202         struct buf *bp;
 203         struct uio *uio;
 204         int error;
 205         int n;
 206         int seqcount;
 207         int ioseqcount;
 208         int blksize;
 209
 210         if (ap->a_vp->v_type != VREG)
 211                 return (EINVAL);
 212         ip = VTOI(ap->a_vp);
 213         error = 0;
 214         uio = ap->a_uio;
 215
 216         /*
 217          * Allow the UIO's size to override the sequential heuristic.
 218          */
 219         blksize = hammer_blocksize(uio->uio_offset);
 220         seqcount = (uio->uio_resid + (blksize - 1)) / blksize;
 221         ioseqcount = ap->a_ioflag >> 16;
 222         if (seqcount < ioseqcount)
 223                 seqcount = ioseqcount;
 224
 225         hammer_start_transaction(&trans, ip->hmp);
 226
 227         /*
 228          * Access the data typically in HAMMER_BUFSIZE blocks via the
 229          * buffer cache, but HAMMER may use a variable block size based
 230          * on the offset.
 231          */
 232         while (uio->uio_resid > 0 && uio->uio_offset < ip->ino_data.size) {
 233                 int64_t base_offset;
 234                 int64_t file_limit;
 235
 236                 blksize = hammer_blocksize(uio->uio_offset);
 237                 offset = (int)uio->uio_offset & (blksize - 1);
 238                 base_offset = uio->uio_offset - offset;
 239
 240                 if (hammer_debug_cluster_enable) {
 241                         /*
 242                          * Use file_limit to prevent cluster_read() from
 243                          * creating buffers of the wrong block size past
 244                          * the demarc.
 245                          */
 246                         file_limit = ip->ino_data.size;
 247                         if (base_offset < HAMMER_XDEMARC &&
 248                             file_limit > HAMMER_XDEMARC) {
 249                                 file_limit = HAMMER_XDEMARC;
 250                         }
 251                         error = cluster_read(ap->a_vp,
 252                                              file_limit, base_offset,
 253                                              blksize, MAXPHYS,
 254                                              seqcount, &bp);
 255                 } else {
 256                         error = bread(ap->a_vp, base_offset, blksize, &bp);
 257                 }
 258                 if (error) {
 259                         kprintf("error %d\n", error);
 260                         brelse(bp);
 261                         break;
 262                 }
 263
 264                 /* bp->b_flags |= B_CLUSTEROK; temporarily disabled */
 265                 n = blksize - offset;
 266                 if (n > uio->uio_resid)
 267                         n = uio->uio_resid;
 268                 if (n > ip->ino_data.size - uio->uio_offset)
 269                         n = (int)(ip->ino_data.size - uio->uio_offset);
 270                 error = uiomove((char *)bp->b_data + offset, n, uio);
 271
 272                 /* data has a lower priority then meta-data */
 273                 bp->b_flags |= B_AGE;
 274                 bqrelse(bp);
 275                 if (error)
 276                         break;
 277         }
 278         if ((ip->flags & HAMMER_INODE_RO) == 0 &&
 279             (ip->hmp->mp->mnt_flag & MNT_NOATIME) == 0) {
 280                 ip->ino_data.atime = trans.time;
 281                 hammer_modify_inode(ip, HAMMER_INODE_ATIME);
 282         }
 283         hammer_done_transaction(&trans);
 284         return (error);
 285 }
 286
 287 /*
 288  * hammer_vop_write { vp, uio, ioflag, cred }
 289  */
 290 static
 291 int
 292 hammer_vop_write(struct vop_write_args *ap)
 293 {
 294         struct hammer_transaction trans;
 295         struct hammer_inode *ip;
 296         hammer_mount_t hmp;
 297         struct uio *uio;
 298         int offset;
 299         off_t base_offset;
 300         struct buf *bp;
 301         int error;
 302         int n;
 303         int flags;
 304         int delta;
 305         int seqcount;
 306
 307         if (ap->a_vp->v_type != VREG)
 308                 return (EINVAL);
 309         ip = VTOI(ap->a_vp);
 310         hmp = ip->hmp;
 311         error = 0;
 312         seqcount = ap->a_ioflag >> 16;
 313
 314         if (ip->flags & HAMMER_INODE_RO)
 315                 return (EROFS);
 316
 317         /*
 318          * Create a transaction to cover the operations we perform.
 319          */
 320         hammer_start_transaction(&trans, hmp);
 321         uio = ap->a_uio;
 322
 323         /*
 324          * Check append mode
 325          */
 326         if (ap->a_ioflag & IO_APPEND)
 327                 uio->uio_offset = ip->ino_data.size;
 328
 329         /*
 330          * Check for illegal write offsets.  Valid range is 0...2^63-1.
 331          *
 332          * NOTE: the base_off assignment is required to work around what
 333          * I consider to be a GCC-4 optimization bug.
 334          */
 335         if (uio->uio_offset < 0) {
 336                 hammer_done_transaction(&trans);
 337                 return (EFBIG);
 338         }
 339         base_offset = uio->uio_offset + uio->uio_resid; /* work around gcc-4 */
 340         if (uio->uio_resid > 0 && base_offset <= 0) {
 341                 hammer_done_transaction(&trans);
 342                 return (EFBIG);
 343         }
 344
 345         /*
 346          * Access the data typically in HAMMER_BUFSIZE blocks via the
 347          * buffer cache, but HAMMER may use a variable block size based
 348          * on the offset.
 349          */
 350         while (uio->uio_resid > 0) {
 351                 int fixsize = 0;
 352                 int blksize;
 353                 int blkmask;
 354
 355                 if ((error = hammer_checkspace(hmp, HAMMER_CHECKSPACE_SLOP_WRITE)) != 0)
 356                         break;
 357
 358                 blksize = hammer_blocksize(uio->uio_offset);
 359
 360                 /*
 361                  * Do not allow HAMMER to blow out the buffer cache.  Very
 362                  * large UIOs can lockout other processes due to bwillwrite()
 363                  * mechanics.
 364                  *
 365                  * Do not allow HAMMER to blow out system memory by
 366                  * accumulating too many records.   Records are so well
 367                  * decoupled from the buffer cache that it is possible
 368                  * for userland to push data out to the media via
 369                  * direct-write, but build up the records queued to the
 370                  * backend faster then the backend can flush them out.
 371                  * HAMMER has hit its write limit but the frontend has
 372                  * no pushback to slow it down.
 373                  *
 374                  * The hammer inode is not locked during these operations.
 375                  * The vnode is locked which can interfere with the pageout
 376                  * daemon for non-UIO_NOCOPY writes but should not interfere
 377                  * with the buffer cache.  Even so, we cannot afford to
 378                  * allow the pageout daemon to build up too many dirty buffer
 379                  * cache buffers.
 380                  */
 381                 bwillwrite(blksize);
 382
 383                 /*
 384                  * Pending record flush check.
 385                  */
 386                 if (hmp->rsv_recs > hammer_limit_recs / 2) {
 387                         /*
 388                          * Get the inode on the flush list
 389                          */
 390                         if (ip->rsv_recs >= 64)
 391                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 392                         else if (ip->rsv_recs >= 16)
 393                                 hammer_flush_inode(ip, 0);
 394
 395                         /*
 396                          * Keep the flusher going if the system keeps
 397                          * queueing records.
 398                          */
 399                         delta = hmp->count_newrecords -
 400                                 hmp->last_newrecords;
 401                         if (delta < 0 || delta > hammer_limit_recs / 2) {
 402                                 hmp->last_newrecords = hmp->count_newrecords;
 403                                 hammer_sync_hmp(hmp, MNT_NOWAIT);
 404                         }
 405
 406                         /*
 407                          * If we have gotten behind start slowing
 408                          * down the writers.
 409                          */
 410                         delta = (hmp->rsv_recs - hammer_limit_recs) *
 411                                 hz / hammer_limit_recs;
 412                         if (delta > 0)
 413                                 tsleep(&trans, 0, "hmrslo", delta);
 414                 }
 415
 416                 /*
 417                  * Calculate the blocksize at the current offset and figure
 418                  * out how much we can actually write.
 419                  */
 420                 blkmask = blksize - 1;
 421                 offset = (int)uio->uio_offset & blkmask;
 422                 base_offset = uio->uio_offset & ~(int64_t)blkmask;
 423                 n = blksize - offset;
 424                 if (n > uio->uio_resid)
 425                         n = uio->uio_resid;
 426                 if (uio->uio_offset + n > ip->ino_data.size) {
 427                         vnode_pager_setsize(ap->a_vp, uio->uio_offset + n);
 428                         fixsize = 1;
 429                 }
 430
 431                 if (uio->uio_segflg == UIO_NOCOPY) {
 432                         /*
 433                          * Issuing a write with the same data backing the
 434                          * buffer.  Instantiate the buffer to collect the
 435                          * backing vm pages, then read-in any missing bits.
 436                          *
 437                          * This case is used by vop_stdputpages().
 438                          */
 439                         bp = getblk(ap->a_vp, base_offset,
 440                                     blksize, GETBLK_BHEAVY, 0);
 441                         if ((bp->b_flags & B_CACHE) == 0) {
 442                                 bqrelse(bp);
 443                                 error = bread(ap->a_vp, base_offset,
 444                                               blksize, &bp);
 445                         }
 446                 } else if (offset == 0 && uio->uio_resid >= blksize) {
 447                         /*
 448                          * Even though we are entirely overwriting the buffer
 449                          * we may still have to zero it out to avoid a
 450                          * mmap/write visibility issue.
 451                          */
 452                         bp = getblk(ap->a_vp, base_offset, blksize, GETBLK_BHEAVY, 0);
 453                         if ((bp->b_flags & B_CACHE) == 0)
 454                                 vfs_bio_clrbuf(bp);
 455                 } else if (base_offset >= ip->ino_data.size) {
 456                         /*
 457                          * If the base offset of the buffer is beyond the
 458                          * file EOF, we don't have to issue a read.
 459                          */
 460                         bp = getblk(ap->a_vp, base_offset,
 461                                     blksize, GETBLK_BHEAVY, 0);
 462                         vfs_bio_clrbuf(bp);
 463                 } else {
 464                         /*
 465                          * Partial overwrite, read in any missing bits then
 466                          * replace the portion being written.
 467                          */
 468                         error = bread(ap->a_vp, base_offset, blksize, &bp);
 469                         if (error == 0)
 470                                 bheavy(bp);
 471                 }
 472                 if (error == 0) {
 473                         error = uiomove((char *)bp->b_data + offset,
 474                                         n, uio);
 475                 }
 476
 477                 /*
 478                  * If we screwed up we have to undo any VM size changes we
 479                  * made.
 480                  */
 481                 if (error) {
 482                         brelse(bp);
 483                         if (fixsize) {
 484                                 vtruncbuf(ap->a_vp, ip->ino_data.size,
 485                                           hammer_blocksize(ip->ino_data.size));
 486                         }
 487                         break;
 488                 }
 489                 /* bp->b_flags |= B_CLUSTEROK; temporarily disabled */
 490                 if (ip->ino_data.size < uio->uio_offset) {
 491                         ip->ino_data.size = uio->uio_offset;
 492                         flags = HAMMER_INODE_DDIRTY;
 493                         vnode_pager_setsize(ap->a_vp, ip->ino_data.size);
 494                 } else {
 495                         flags = 0;
 496                 }
 497                 ip->ino_data.mtime = trans.time;
 498                 flags |= HAMMER_INODE_MTIME | HAMMER_INODE_BUFS;
 499                 hammer_modify_inode(ip, flags);
 500
 501                 /*
 502                  * Final buffer disposition.
 503                  */
 504                 bp->b_flags |= B_AGE;
 505                 if (ap->a_ioflag & IO_SYNC) {
 506                         bwrite(bp);
 507                 } else if (ap->a_ioflag & IO_DIRECT) {
 508                         bawrite(bp);
 509                 } else {
 510                         bdwrite(bp);
 511                 }
 512         }
 513         hammer_done_transaction(&trans);
 514         return (error);
 515 }
 516
 517 /*
 518  * hammer_vop_access { vp, mode, cred }
 519  */
 520 static
 521 int
 522 hammer_vop_access(struct vop_access_args *ap)
 523 {
 524         struct hammer_inode *ip = VTOI(ap->a_vp);
 525         uid_t uid;
 526         gid_t gid;
 527         int error;
 528
 529         uid = hammer_to_unix_xid(&ip->ino_data.uid);
 530         gid = hammer_to_unix_xid(&ip->ino_data.gid);
 531
 532         error = vop_helper_access(ap, uid, gid, ip->ino_data.mode,
 533                                   ip->ino_data.uflags);
 534         return (error);
 535 }
 536
 537 /*
 538  * hammer_vop_advlock { vp, id, op, fl, flags }
 539  */
 540 static
 541 int
 542 hammer_vop_advlock(struct vop_advlock_args *ap)
 543 {
 544         hammer_inode_t ip = VTOI(ap->a_vp);
 545
 546         return (lf_advlock(ap, &ip->advlock, ip->ino_data.size));
 547 }
 548
 549 /*
 550  * hammer_vop_close { vp, fflag }
 551  */
 552 static
 553 int
 554 hammer_vop_close(struct vop_close_args *ap)
 555 {
 556         hammer_inode_t ip = VTOI(ap->a_vp);
 557
 558         if ((ip->flags | ip->sync_flags) & HAMMER_INODE_MODMASK)
 559                 hammer_inode_waitreclaims(ip->hmp);
 560         return (vop_stdclose(ap));
 561 }
 562
 563 /*
 564  * hammer_vop_ncreate { nch, dvp, vpp, cred, vap }
 565  *
 566  * The operating system has already ensured that the directory entry
 567  * does not exist and done all appropriate namespace locking.
 568  */
 569 static
 570 int
 571 hammer_vop_ncreate(struct vop_ncreate_args *ap)
 572 {
 573         struct hammer_transaction trans;
 574         struct hammer_inode *dip;
 575         struct hammer_inode *nip;
 576         struct nchandle *nch;
 577         int error;
 578
 579         nch = ap->a_nch;
 580         dip = VTOI(ap->a_dvp);
 581
 582         if (dip->flags & HAMMER_INODE_RO)
 583                 return (EROFS);
 584         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHECKSPACE_SLOP_CREATE)) != 0)
 585                 return (error);
 586
 587         /*
 588          * Create a transaction to cover the operations we perform.
 589          */
 590         hammer_start_transaction(&trans, dip->hmp);
 591
 592         /*
 593          * Create a new filesystem object of the requested type.  The
 594          * returned inode will be referenced and shared-locked to prevent
 595          * it from being moved to the flusher.
 596          */
 597
 598         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
 599                                     dip, 0, &nip);
 600         if (error) {
 601                 hkprintf("hammer_create_inode error %d\n", error);
 602                 hammer_done_transaction(&trans);
 603                 *ap->a_vpp = NULL;
 604                 return (error);
 605         }
 606
 607         /*
 608          * Add the new filesystem object to the directory.  This will also
 609          * bump the inode's link count.
 610          */
 611         error = hammer_ip_add_directory(&trans, dip,
 612                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
 613                                         nip);
 614         if (error)
 615                 hkprintf("hammer_ip_add_directory error %d\n", error);
 616
 617         /*
 618          * Finish up.
 619          */
 620         if (error) {
 621                 hammer_rel_inode(nip, 0);
 622                 hammer_done_transaction(&trans);
 623                 *ap->a_vpp = NULL;
 624         } else {
 625                 error = hammer_get_vnode(nip, ap->a_vpp);
 626                 hammer_done_transaction(&trans);
 627                 hammer_rel_inode(nip, 0);
 628                 if (error == 0) {
 629                         cache_setunresolved(ap->a_nch);
 630                         cache_setvp(ap->a_nch, *ap->a_vpp);
 631                 }
 632         }
 633         return (error);
 634 }
 635
 636 /*
 637  * hammer_vop_getattr { vp, vap }
 638  *
 639  * Retrieve an inode's attribute information.  When accessing inodes
 640  * historically we fake the atime field to ensure consistent results.
 641  * The atime field is stored in the B-Tree element and allowed to be
 642  * updated without cycling the element.
 643  */
 644 static
 645 int
 646 hammer_vop_getattr(struct vop_getattr_args *ap)
 647 {
 648         struct hammer_inode *ip = VTOI(ap->a_vp);
 649         struct vattr *vap = ap->a_vap;
 650
 651         /*
 652          * We want the fsid to be different when accessing a filesystem
 653          * with different as-of's so programs like diff don't think
 654          * the files are the same.
 655          *
 656          * We also want the fsid to be the same when comparing snapshots,
 657          * or when comparing mirrors (which might be backed by different
 658          * physical devices).  HAMMER fsids are based on the PFS's
 659          * shared_uuid field.
 660          *
 661          * XXX there is a chance of collision here.  The va_fsid reported
 662          * by stat is different from the more involved fsid used in the
 663          * mount structure.
 664          */
 665         vap->va_fsid = ip->pfsm->fsid_udev ^ (u_int32_t)ip->obj_asof ^
 666                        (u_int32_t)(ip->obj_asof >> 32);
 667
 668         vap->va_fileid = ip->ino_leaf.base.obj_id;
 669         vap->va_mode = ip->ino_data.mode;
 670         vap->va_nlink = ip->ino_data.nlinks;
 671         vap->va_uid = hammer_to_unix_xid(&ip->ino_data.uid);
 672         vap->va_gid = hammer_to_unix_xid(&ip->ino_data.gid);
 673         vap->va_rmajor = 0;
 674         vap->va_rminor = 0;
 675         vap->va_size = ip->ino_data.size;
 676
 677         /*
 678          * We must provide a consistent atime and mtime for snapshots
 679          * so people can do a 'tar cf - ... | md5' on them and get
 680          * consistent results.
 681          */
 682         if (ip->flags & HAMMER_INODE_RO) {
 683                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_atime);
 684                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_mtime);
 685         } else {
 686                 hammer_time_to_timespec(ip->ino_data.atime, &vap->va_atime);
 687                 hammer_time_to_timespec(ip->ino_data.mtime, &vap->va_mtime);
 688         }
 689         hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_ctime);
 690         vap->va_flags = ip->ino_data.uflags;
 691         vap->va_gen = 1;        /* hammer inums are unique for all time */
 692         vap->va_blocksize = HAMMER_BUFSIZE;
 693         if (ip->ino_data.size >= HAMMER_XDEMARC) {
 694                 vap->va_bytes = (ip->ino_data.size + HAMMER_XBUFMASK64) &
 695                                 ~HAMMER_XBUFMASK64;
 696         } else if (ip->ino_data.size > HAMMER_BUFSIZE / 2) {
 697                 vap->va_bytes = (ip->ino_data.size + HAMMER_BUFMASK64) &
 698                                 ~HAMMER_BUFMASK64;
 699         } else {
 700                 vap->va_bytes = (ip->ino_data.size + 15) & ~15;
 701         }
 702         vap->va_type = hammer_get_vnode_type(ip->ino_data.obj_type);
 703         vap->va_filerev = 0;    /* XXX */
 704         /* mtime uniquely identifies any adjustments made to the file XXX */
 705         vap->va_fsmid = ip->ino_data.mtime;
 706         vap->va_uid_uuid = ip->ino_data.uid;
 707         vap->va_gid_uuid = ip->ino_data.gid;
 708         vap->va_fsid_uuid = ip->hmp->fsid;
 709         vap->va_vaflags = VA_UID_UUID_VALID | VA_GID_UUID_VALID |
 710                           VA_FSID_UUID_VALID;
 711
 712         switch (ip->ino_data.obj_type) {
 713         case HAMMER_OBJTYPE_CDEV:
 714         case HAMMER_OBJTYPE_BDEV:
 715                 vap->va_rmajor = ip->ino_data.rmajor;
 716                 vap->va_rminor = ip->ino_data.rminor;
 717                 break;
 718         default:
 719                 break;
 720         }
 721
 722         return(0);
 723 }
 724
 725 /*
 726  * hammer_vop_nresolve { nch, dvp, cred }
 727  *
 728  * Locate the requested directory entry.
 729  */
 730 static
 731 int
 732 hammer_vop_nresolve(struct vop_nresolve_args *ap)
 733 {
 734         struct hammer_transaction trans;
 735         struct namecache *ncp;
 736         hammer_inode_t dip;
 737         hammer_inode_t ip;
 738         hammer_tid_t asof;
 739         struct hammer_cursor cursor;
 740         struct vnode *vp;
 741         int64_t namekey;
 742         int error;
 743         int i;
 744         int nlen;
 745         int flags;
 746         int ispfs;
 747         int64_t obj_id;
 748         u_int32_t localization;
 749
 750         /*
 751          * Misc initialization, plus handle as-of name extensions.  Look for
 752          * the '@@' extension.  Note that as-of files and directories cannot
 753          * be modified.
 754          */
 755         dip = VTOI(ap->a_dvp);
 756         ncp = ap->a_nch->ncp;
 757         asof = dip->obj_asof;
 758         nlen = ncp->nc_nlen;
 759         flags = dip->flags;
 760         ispfs = 0;
 761
 762         hammer_simple_transaction(&trans, dip->hmp);
 763
 764         for (i = 0; i < nlen; ++i) {
 765                 if (ncp->nc_name[i] == '@' && ncp->nc_name[i+1] == '@') {
 766                         asof = hammer_str_to_tid(ncp->nc_name + i + 2);
 767                         flags |= HAMMER_INODE_RO;
 768                         break;
 769                 }
 770         }
 771         nlen = i;
 772
 773         /*
 774          * If there is no path component the time extension is relative to
 775          * dip.
 776          */
 777         if (nlen == 0) {
 778                 ip = hammer_get_inode(&trans, dip, dip->obj_id,
 779                                       asof, dip->obj_localization,
 780                                       flags, &error);
 781                 if (error == 0) {
 782                         error = hammer_get_vnode(ip, &vp);
 783                         hammer_rel_inode(ip, 0);
 784                 } else {
 785                         vp = NULL;
 786                 }
 787                 if (error == 0) {
 788                         vn_unlock(vp);
 789                         cache_setvp(ap->a_nch, vp);
 790                         vrele(vp);
 791                 }
 792                 goto done;
 793         }
 794
 795         /*
 796          * Calculate the namekey and setup the key range for the scan.  This
 797          * works kinda like a chained hash table where the lower 32 bits
 798          * of the namekey synthesize the chain.
 799          *
 800          * The key range is inclusive of both key_beg and key_end.
 801          */
 802         namekey = hammer_directory_namekey(ncp->nc_name, nlen);
 803
 804         error = hammer_init_cursor(&trans, &cursor, &dip->cache[1], dip);
 805         cursor.key_beg.localization = dip->obj_localization +
 806                                       HAMMER_LOCALIZE_MISC;
 807         cursor.key_beg.obj_id = dip->obj_id;
 808         cursor.key_beg.key = namekey;
 809         cursor.key_beg.create_tid = 0;
 810         cursor.key_beg.delete_tid = 0;
 811         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
 812         cursor.key_beg.obj_type = 0;
 813
 814         cursor.key_end = cursor.key_beg;
 815         cursor.key_end.key |= 0xFFFFFFFFULL;
 816         cursor.asof = asof;
 817         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
 818
 819         /*
 820          * Scan all matching records (the chain), locate the one matching
 821          * the requested path component.
 822          *
 823          * The hammer_ip_*() functions merge in-memory records with on-disk
 824          * records for the purposes of the search.
 825          */
 826         obj_id = 0;
 827         localization = HAMMER_DEF_LOCALIZATION;
 828
 829         if (error == 0) {
 830                 error = hammer_ip_first(&cursor);
 831                 while (error == 0) {
 832                         error = hammer_ip_resolve_data(&cursor);
 833                         if (error)
 834                                 break;
 835                         if (nlen == cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF &&
 836                             bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
 837                                 obj_id = cursor.data->entry.obj_id;
 838
 839                                 /*
 840                                  * Force relookups whenever a PFS root is
 841                                  * accessed.
 842                                  */
 843                                 if (obj_id == HAMMER_OBJID_ROOT)
 844                                         ispfs = 1;
 845                                 localization = cursor.data->entry.localization;
 846                                 break;
 847                         }
 848                         error = hammer_ip_next(&cursor);
 849                 }
 850         }
 851         hammer_done_cursor(&cursor);
 852         if (error == 0) {
 853                 ip = hammer_get_inode(&trans, dip, obj_id,
 854                                       asof, localization,
 855                                       flags, &error);
 856                 if (ispfs && asof > ip->pfsm->pfsd.sync_end_tid) {
 857                         asof = ip->pfsm->pfsd.sync_end_tid;
 858                         hammer_rel_inode(ip, 0);
 859                         ip = hammer_get_inode(&trans, dip, obj_id,
 860                                               asof, localization,
 861                                               flags, &error);
 862                 }
 863
 864
 865                 if (error == 0) {
 866                         error = hammer_get_vnode(ip, &vp);
 867                         hammer_rel_inode(ip, 0);
 868                 } else {
 869                         vp = NULL;
 870                 }
 871                 if (error == 0) {
 872                         vn_unlock(vp);
 873                         cache_setvp(ap->a_nch, vp);
 874                         if (ispfs)
 875                                 cache_settimeout(ap->a_nch, 0);
 876                         vrele(vp);
 877                 }
 878         } else if (error == ENOENT) {
 879                 cache_setvp(ap->a_nch, NULL);
 880         }
 881 done:
 882         hammer_done_transaction(&trans);
 883         return (error);
 884 }
 885
 886 /*
 887  * hammer_vop_nlookupdotdot { dvp, vpp, cred }
 888  *
 889  * Locate the parent directory of a directory vnode.
 890  *
 891  * dvp is referenced but not locked.  *vpp must be returned referenced and
 892  * locked.  A parent_obj_id of 0 does not necessarily indicate that we are
 893  * at the root, instead it could indicate that the directory we were in was
 894  * removed.
 895  *
 896  * NOTE: as-of sequences are not linked into the directory structure.  If
 897  * we are at the root with a different asof then the mount point, reload
 898  * the same directory with the mount point's asof.   I'm not sure what this
 899  * will do to NFS.  We encode ASOF stamps in NFS file handles so it might not
 900  * get confused, but it hasn't been tested.
 901  */
 902 static
 903 int
 904 hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *ap)
 905 {
 906         struct hammer_transaction trans;
 907         struct hammer_inode *dip;
 908         struct hammer_inode *ip;
 909         int64_t parent_obj_id;
 910         u_int32_t parent_obj_localization;
 911         hammer_tid_t asof;
 912         int error;
 913
 914         dip = VTOI(ap->a_dvp);
 915         asof = dip->obj_asof;
 916
 917         /*
 918          * Whos are parent?  This could be the root of a pseudo-filesystem
 919          * whos parent is in another localization domain.
 920          */
 921         parent_obj_id = dip->ino_data.parent_obj_id;
 922         if (dip->obj_id == HAMMER_OBJID_ROOT)
 923                 parent_obj_localization = dip->ino_data.ext.obj.parent_obj_localization;
 924         else
 925                 parent_obj_localization = dip->obj_localization;
 926
 927         if (parent_obj_id == 0) {
 928                 if (dip->obj_id == HAMMER_OBJID_ROOT &&
 929                    asof != dip->hmp->asof) {
 930                         parent_obj_id = dip->obj_id;
 931                         asof = dip->hmp->asof;
 932                         *ap->a_fakename = kmalloc(19, M_TEMP, M_WAITOK);
 933                         ksnprintf(*ap->a_fakename, 19, "0x%016llx",
 934                                    dip->obj_asof);
 935                 } else {
 936                         *ap->a_vpp = NULL;
 937                         return ENOENT;
 938                 }
 939         }
 940
 941         hammer_simple_transaction(&trans, dip->hmp);
 942
 943         ip = hammer_get_inode(&trans, dip, parent_obj_id,
 944                               asof, parent_obj_localization,
 945                               dip->flags, &error);
 946         if (ip) {
 947                 error = hammer_get_vnode(ip, ap->a_vpp);
 948                 hammer_rel_inode(ip, 0);
 949         } else {
 950                 *ap->a_vpp = NULL;
 951         }
 952         hammer_done_transaction(&trans);
 953         return (error);
 954 }
 955
 956 /*
 957  * hammer_vop_nlink { nch, dvp, vp, cred }
 958  */
 959 static
 960 int
 961 hammer_vop_nlink(struct vop_nlink_args *ap)
 962 {
 963         struct hammer_transaction trans;
 964         struct hammer_inode *dip;
 965         struct hammer_inode *ip;
 966         struct nchandle *nch;
 967         int error;
 968
 969         nch = ap->a_nch;
 970         dip = VTOI(ap->a_dvp);
 971         ip = VTOI(ap->a_vp);
 972
 973         if (dip->flags & HAMMER_INODE_RO)
 974                 return (EROFS);
 975         if (ip->flags & HAMMER_INODE_RO)
 976                 return (EROFS);
 977         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHECKSPACE_SLOP_CREATE)) != 0)
 978                 return (error);
 979
 980         /*
 981          * Create a transaction to cover the operations we perform.
 982          */
 983         hammer_start_transaction(&trans, dip->hmp);
 984
 985         /*
 986          * Add the filesystem object to the directory.  Note that neither
 987          * dip nor ip are referenced or locked, but their vnodes are
 988          * referenced.  This function will bump the inode's link count.
 989          */
 990         error = hammer_ip_add_directory(&trans, dip,
 991                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
 992                                         ip);
 993
 994         /*
 995          * Finish up.
 996          */
 997         if (error == 0) {
 998                 cache_setunresolved(nch);
 999                 cache_setvp(nch, ap->a_vp);
1000         }
1001         hammer_done_transaction(&trans);
1002         return (error);
1003 }
1004
1005 /*
1006  * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
1007  *
1008  * The operating system has already ensured that the directory entry
1009  * does not exist and done all appropriate namespace locking.
1010  */
1011 static
1012 int
1013 hammer_vop_nmkdir(struct vop_nmkdir_args *ap)
1014 {
1015         struct hammer_transaction trans;
1016         struct hammer_inode *dip;
1017         struct hammer_inode *nip;
1018         struct nchandle *nch;
1019         int error;
1020
1021         nch = ap->a_nch;
1022         dip = VTOI(ap->a_dvp);
1023
1024         if (dip->flags & HAMMER_INODE_RO)
1025                 return (EROFS);
1026         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHECKSPACE_SLOP_CREATE)) != 0)
1027                 return (error);
1028
1029         /*
1030          * Create a transaction to cover the operations we perform.
1031          */
1032         hammer_start_transaction(&trans, dip->hmp);
1033
1034         /*
1035          * Create a new filesystem object of the requested type.  The
1036          * returned inode will be referenced but not locked.
1037          */
1038         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1039                                     dip, 0, &nip);
1040         if (error) {
1041                 hkprintf("hammer_mkdir error %d\n", error);
1042                 hammer_done_transaction(&trans);
1043                 *ap->a_vpp = NULL;
1044                 return (error);
1045         }
1046         /*
1047          * Add the new filesystem object to the directory.  This will also
1048          * bump the inode's link count.
1049          */
1050         error = hammer_ip_add_directory(&trans, dip,
1051                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1052                                         nip);
1053         if (error)
1054                 hkprintf("hammer_mkdir (add) error %d\n", error);
1055
1056         /*
1057          * Finish up.
1058          */
1059         if (error) {
1060                 hammer_rel_inode(nip, 0);
1061                 *ap->a_vpp = NULL;
1062         } else {
1063                 error = hammer_get_vnode(nip, ap->a_vpp);
1064                 hammer_rel_inode(nip, 0);
1065                 if (error == 0) {
1066                         cache_setunresolved(ap->a_nch);
1067                         cache_setvp(ap->a_nch, *ap->a_vpp);
1068                 }
1069         }
1070         hammer_done_transaction(&trans);
1071         return (error);
1072 }
1073
1074 /*
1075  * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1076  *
1077  * The operating system has already ensured that the directory entry
1078  * does not exist and done all appropriate namespace locking.
1079  */
1080 static
1081 int
1082 hammer_vop_nmknod(struct vop_nmknod_args *ap)
1083 {
1084         struct hammer_transaction trans;
1085         struct hammer_inode *dip;
1086         struct hammer_inode *nip;
1087         struct nchandle *nch;
1088         int error;
1089         int pseudofs;
1090
1091         nch = ap->a_nch;
1092         dip = VTOI(ap->a_dvp);
1093
1094         if (dip->flags & HAMMER_INODE_RO)
1095                 return (EROFS);
1096         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHECKSPACE_SLOP_CREATE)) != 0)
1097                 return (error);
1098
1099         /*
1100          * Create a transaction to cover the operations we perform.
1101          */
1102         hammer_start_transaction(&trans, dip->hmp);
1103
1104         /*
1105          * Create a new filesystem object of the requested type.  The
1106          * returned inode will be referenced but not locked.
1107          *
1108          * If mknod specifies a directory a pseudo-fs is created.
1109          */
1110         pseudofs = (ap->a_vap->va_type == VDIR);
1111         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1112                                     dip, pseudofs, &nip);
1113         if (error) {
1114                 hammer_done_transaction(&trans);
1115                 *ap->a_vpp = NULL;
1116                 return (error);
1117         }
1118
1119         /*
1120          * Add the new filesystem object to the directory.  This will also
1121          * bump the inode's link count.
1122          */
1123         error = hammer_ip_add_directory(&trans, dip,
1124                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1125                                         nip);
1126
1127         /*
1128          * Finish up.
1129          */
1130         if (error) {
1131                 hammer_rel_inode(nip, 0);
1132                 *ap->a_vpp = NULL;
1133         } else {
1134                 error = hammer_get_vnode(nip, ap->a_vpp);
1135                 hammer_rel_inode(nip, 0);
1136                 if (error == 0) {
1137                         cache_setunresolved(ap->a_nch);
1138                         cache_setvp(ap->a_nch, *ap->a_vpp);
1139                 }
1140         }
1141         hammer_done_transaction(&trans);
1142         return (error);
1143 }
1144
1145 /*
1146  * hammer_vop_open { vp, mode, cred, fp }
1147  */
1148 static
1149 int
1150 hammer_vop_open(struct vop_open_args *ap)
1151 {
1152         hammer_inode_t ip;
1153
1154         ip = VTOI(ap->a_vp);
1155
1156         if ((ap->a_mode & FWRITE) && (ip->flags & HAMMER_INODE_RO))
1157                 return (EROFS);
1158         return(vop_stdopen(ap));
1159 }
1160
1161 /*
1162  * hammer_vop_pathconf { vp, name, retval }
1163  */
1164 static
1165 int
1166 hammer_vop_pathconf(struct vop_pathconf_args *ap)
1167 {
1168         return EOPNOTSUPP;
1169 }
1170
1171 /*
1172  * hammer_vop_print { vp }
1173  */
1174 static
1175 int
1176 hammer_vop_print(struct vop_print_args *ap)
1177 {
1178         return EOPNOTSUPP;
1179 }
1180
1181 /*
1182  * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1183  */
1184 static
1185 int
1186 hammer_vop_readdir(struct vop_readdir_args *ap)
1187 {
1188         struct hammer_transaction trans;
1189         struct hammer_cursor cursor;
1190         struct hammer_inode *ip;
1191         struct uio *uio;
1192         hammer_base_elm_t base;
1193         int error;
1194         int cookie_index;
1195         int ncookies;
1196         off_t *cookies;
1197         off_t saveoff;
1198         int r;
1199
1200         ip = VTOI(ap->a_vp);
1201         uio = ap->a_uio;
1202         saveoff = uio->uio_offset;
1203
1204         if (ap->a_ncookies) {
1205                 ncookies = uio->uio_resid / 16 + 1;
1206                 if (ncookies > 1024)
1207                         ncookies = 1024;
1208                 cookies = kmalloc(ncookies * sizeof(off_t), M_TEMP, M_WAITOK);
1209                 cookie_index = 0;
1210         } else {
1211                 ncookies = -1;
1212                 cookies = NULL;
1213                 cookie_index = 0;
1214         }
1215
1216         hammer_simple_transaction(&trans, ip->hmp);
1217
1218         /*
1219          * Handle artificial entries
1220          */
1221         error = 0;
1222         if (saveoff == 0) {
1223                 r = vop_write_dirent(&error, uio, ip->obj_id, DT_DIR, 1, ".");
1224                 if (r)
1225                         goto done;
1226                 if (cookies)
1227                         cookies[cookie_index] = saveoff;
1228                 ++saveoff;
1229                 ++cookie_index;
1230                 if (cookie_index == ncookies)
1231                         goto done;
1232         }
1233         if (saveoff == 1) {
1234                 if (ip->ino_data.parent_obj_id) {
1235                         r = vop_write_dirent(&error, uio,
1236                                              ip->ino_data.parent_obj_id,
1237                                              DT_DIR, 2, "..");
1238                 } else {
1239                         r = vop_write_dirent(&error, uio,
1240                                              ip->obj_id, DT_DIR, 2, "..");
1241                 }
1242                 if (r)
1243                         goto done;
1244                 if (cookies)
1245                         cookies[cookie_index] = saveoff;
1246                 ++saveoff;
1247                 ++cookie_index;
1248                 if (cookie_index == ncookies)
1249                         goto done;
1250         }
1251
1252         /*
1253          * Key range (begin and end inclusive) to scan.  Directory keys
1254          * directly translate to a 64 bit 'seek' position.
1255          */
1256         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1257         cursor.key_beg.localization = ip->obj_localization +
1258                                       HAMMER_LOCALIZE_MISC;
1259         cursor.key_beg.obj_id = ip->obj_id;
1260         cursor.key_beg.create_tid = 0;
1261         cursor.key_beg.delete_tid = 0;
1262         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1263         cursor.key_beg.obj_type = 0;
1264         cursor.key_beg.key = saveoff;
1265
1266         cursor.key_end = cursor.key_beg;
1267         cursor.key_end.key = HAMMER_MAX_KEY;
1268         cursor.asof = ip->obj_asof;
1269         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1270
1271         error = hammer_ip_first(&cursor);
1272
1273         while (error == 0) {
1274                 error = hammer_ip_resolve_data(&cursor);
1275                 if (error)
1276                         break;
1277                 base = &cursor.leaf->base;
1278                 saveoff = base->key;
1279                 KKASSERT(cursor.leaf->data_len > HAMMER_ENTRY_NAME_OFF);
1280
1281                 if (base->obj_id != ip->obj_id)
1282                         panic("readdir: bad record at %p", cursor.node);
1283
1284                 r = vop_write_dirent(
1285                              &error, uio, cursor.data->entry.obj_id,
1286                              hammer_get_dtype(cursor.leaf->base.obj_type),
1287                              cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF ,
1288                              (void *)cursor.data->entry.name);
1289                 if (r)
1290                         break;
1291                 ++saveoff;
1292                 if (cookies)
1293                         cookies[cookie_index] = base->key;
1294                 ++cookie_index;
1295                 if (cookie_index == ncookies)
1296                         break;
1297                 error = hammer_ip_next(&cursor);
1298         }
1299         hammer_done_cursor(&cursor);
1300
1301 done:
1302         hammer_done_transaction(&trans);
1303
1304         if (ap->a_eofflag)
1305                 *ap->a_eofflag = (error == ENOENT);
1306         uio->uio_offset = saveoff;
1307         if (error && cookie_index == 0) {
1308                 if (error == ENOENT)
1309                         error = 0;
1310                 if (cookies) {
1311                         kfree(cookies, M_TEMP);
1312                         *ap->a_ncookies = 0;
1313                         *ap->a_cookies = NULL;
1314                 }
1315         } else {
1316                 if (error == ENOENT)
1317                         error = 0;
1318                 if (cookies) {
1319                         *ap->a_ncookies = cookie_index;
1320                         *ap->a_cookies = cookies;
1321                 }
1322         }
1323         return(error);
1324 }
1325
1326 /*
1327  * hammer_vop_readlink { vp, uio, cred }
1328  */
1329 static
1330 int
1331 hammer_vop_readlink(struct vop_readlink_args *ap)
1332 {
1333         struct hammer_transaction trans;
1334         struct hammer_cursor cursor;
1335         struct hammer_inode *ip;
1336         int error;
1337
1338         ip = VTOI(ap->a_vp);
1339
1340         /*
1341          * Shortcut if the symlink data was stuffed into ino_data.
1342          */
1343         if (ip->ino_data.size <= HAMMER_INODE_BASESYMLEN) {
1344                 error = uiomove(ip->ino_data.ext.symlink,
1345                                 ip->ino_data.size, ap->a_uio);
1346                 return(error);
1347         }
1348
1349         /*
1350          * Long version
1351          */
1352         hammer_simple_transaction(&trans, ip->hmp);
1353         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1354
1355         /*
1356          * Key range (begin and end inclusive) to scan.  Directory keys
1357          * directly translate to a 64 bit 'seek' position.
1358          */
1359         cursor.key_beg.localization = ip->obj_localization +
1360                                       HAMMER_LOCALIZE_MISC;
1361         cursor.key_beg.obj_id = ip->obj_id;
1362         cursor.key_beg.create_tid = 0;
1363         cursor.key_beg.delete_tid = 0;
1364         cursor.key_beg.rec_type = HAMMER_RECTYPE_FIX;
1365         cursor.key_beg.obj_type = 0;
1366         cursor.key_beg.key = HAMMER_FIXKEY_SYMLINK;
1367         cursor.asof = ip->obj_asof;
1368         cursor.flags |= HAMMER_CURSOR_ASOF;
1369
1370         error = hammer_ip_lookup(&cursor);
1371         if (error == 0) {
1372                 error = hammer_ip_resolve_data(&cursor);
1373                 if (error == 0) {
1374                         KKASSERT(cursor.leaf->data_len >=
1375                                  HAMMER_SYMLINK_NAME_OFF);
1376                         error = uiomove(cursor.data->symlink.name,
1377                                         cursor.leaf->data_len -
1378                                                 HAMMER_SYMLINK_NAME_OFF,
1379                                         ap->a_uio);
1380                 }
1381         }
1382         hammer_done_cursor(&cursor);
1383         hammer_done_transaction(&trans);
1384         return(error);
1385 }
1386
1387 /*
1388  * hammer_vop_nremove { nch, dvp, cred }
1389  */
1390 static
1391 int
1392 hammer_vop_nremove(struct vop_nremove_args *ap)
1393 {
1394         struct hammer_transaction trans;
1395         struct hammer_inode *dip;
1396         int error;
1397
1398         dip = VTOI(ap->a_dvp);
1399
1400         if (hammer_nohistory(dip) == 0 &&
1401             (error = hammer_checkspace(dip->hmp, HAMMER_CHECKSPACE_SLOP_REMOVE)) != 0) {
1402                 return (error);
1403         }
1404
1405         hammer_start_transaction(&trans, dip->hmp);
1406         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0);
1407         hammer_done_transaction(&trans);
1408
1409         return (error);
1410 }
1411
1412 /*
1413  * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1414  */
1415 static
1416 int
1417 hammer_vop_nrename(struct vop_nrename_args *ap)
1418 {
1419         struct hammer_transaction trans;
1420         struct namecache *fncp;
1421         struct namecache *tncp;
1422         struct hammer_inode *fdip;
1423         struct hammer_inode *tdip;
1424         struct hammer_inode *ip;
1425         struct hammer_cursor cursor;
1426         int64_t namekey;
1427         int nlen, error;
1428
1429         fdip = VTOI(ap->a_fdvp);
1430         tdip = VTOI(ap->a_tdvp);
1431         fncp = ap->a_fnch->ncp;
1432         tncp = ap->a_tnch->ncp;
1433         ip = VTOI(fncp->nc_vp);
1434         KKASSERT(ip != NULL);
1435
1436         if (fdip->flags & HAMMER_INODE_RO)
1437                 return (EROFS);
1438         if (tdip->flags & HAMMER_INODE_RO)
1439                 return (EROFS);
1440         if (ip->flags & HAMMER_INODE_RO)
1441                 return (EROFS);
1442         if ((error = hammer_checkspace(fdip->hmp, HAMMER_CHECKSPACE_SLOP_CREATE)) != 0)
1443                 return (error);
1444
1445         hammer_start_transaction(&trans, fdip->hmp);
1446
1447         /*
1448          * Remove tncp from the target directory and then link ip as
1449          * tncp. XXX pass trans to dounlink
1450          *
1451          * Force the inode sync-time to match the transaction so it is
1452          * in-sync with the creation of the target directory entry.
1453          */
1454         error = hammer_dounlink(&trans, ap->a_tnch, ap->a_tdvp, ap->a_cred, 0);
1455         if (error == 0 || error == ENOENT) {
1456                 error = hammer_ip_add_directory(&trans, tdip,
1457                                                 tncp->nc_name, tncp->nc_nlen,
1458                                                 ip);
1459                 if (error == 0) {
1460                         ip->ino_data.parent_obj_id = tdip->obj_id;
1461                         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
1462                 }
1463         }
1464         if (error)
1465                 goto failed; /* XXX */
1466
1467         /*
1468          * Locate the record in the originating directory and remove it.
1469          *
1470          * Calculate the namekey and setup the key range for the scan.  This
1471          * works kinda like a chained hash table where the lower 32 bits
1472          * of the namekey synthesize the chain.
1473          *
1474          * The key range is inclusive of both key_beg and key_end.
1475          */
1476         namekey = hammer_directory_namekey(fncp->nc_name, fncp->nc_nlen);
1477 retry:
1478         hammer_init_cursor(&trans, &cursor, &fdip->cache[1], fdip);
1479         cursor.key_beg.localization = fdip->obj_localization +
1480                                       HAMMER_LOCALIZE_MISC;
1481         cursor.key_beg.obj_id = fdip->obj_id;
1482         cursor.key_beg.key = namekey;
1483         cursor.key_beg.create_tid = 0;
1484         cursor.key_beg.delete_tid = 0;
1485         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1486         cursor.key_beg.obj_type = 0;
1487
1488         cursor.key_end = cursor.key_beg;
1489         cursor.key_end.key |= 0xFFFFFFFFULL;
1490         cursor.asof = fdip->obj_asof;
1491         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1492
1493         /*
1494          * Scan all matching records (the chain), locate the one matching
1495          * the requested path component.
1496          *
1497          * The hammer_ip_*() functions merge in-memory records with on-disk
1498          * records for the purposes of the search.
1499          */
1500         error = hammer_ip_first(&cursor);
1501         while (error == 0) {
1502                 if (hammer_ip_resolve_data(&cursor) != 0)
1503                         break;
1504                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
1505                 KKASSERT(nlen > 0);
1506                 if (fncp->nc_nlen == nlen &&
1507                     bcmp(fncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1508                         break;
1509                 }
1510                 error = hammer_ip_next(&cursor);
1511         }
1512
1513         /*
1514          * If all is ok we have to get the inode so we can adjust nlinks.
1515          *
1516          * WARNING: hammer_ip_del_directory() may have to terminate the
1517          * cursor to avoid a recursion.  It's ok to call hammer_done_cursor()
1518          * twice.
1519          */
1520         if (error == 0)
1521                 error = hammer_ip_del_directory(&trans, &cursor, fdip, ip);
1522
1523         /*
1524          * XXX A deadlock here will break rename's atomicy for the purposes
1525          * of crash recovery.
1526          */
1527         if (error == EDEADLK) {
1528                 hammer_done_cursor(&cursor);
1529                 goto retry;
1530         }
1531
1532         /*
1533          * Cleanup and tell the kernel that the rename succeeded.
1534          */
1535         hammer_done_cursor(&cursor);
1536         if (error == 0)
1537                 cache_rename(ap->a_fnch, ap->a_tnch);
1538
1539 failed:
1540         hammer_done_transaction(&trans);
1541         return (error);
1542 }
1543
1544 /*
1545  * hammer_vop_nrmdir { nch, dvp, cred }
1546  */
1547 static
1548 int
1549 hammer_vop_nrmdir(struct vop_nrmdir_args *ap)
1550 {
1551         struct hammer_transaction trans;
1552         struct hammer_inode *dip;
1553         int error;
1554
1555         dip = VTOI(ap->a_dvp);
1556
1557         if (hammer_nohistory(dip) == 0 &&
1558             (error = hammer_checkspace(dip->hmp, HAMMER_CHECKSPACE_SLOP_REMOVE)) != 0) {
1559                 return (error);
1560         }
1561
1562         hammer_start_transaction(&trans, dip->hmp);
1563         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0);
1564         hammer_done_transaction(&trans);
1565
1566         return (error);
1567 }
1568
1569 /*
1570  * hammer_vop_setattr { vp, vap, cred }
1571  */
1572 static
1573 int
1574 hammer_vop_setattr(struct vop_setattr_args *ap)
1575 {
1576         struct hammer_transaction trans;
1577         struct vattr *vap;
1578         struct hammer_inode *ip;
1579         int modflags;
1580         int error;
1581         int truncating;
1582         int blksize;
1583         int64_t aligned_size;
1584         u_int32_t flags;
1585
1586         vap = ap->a_vap;
1587         ip = ap->a_vp->v_data;
1588         modflags = 0;
1589
1590         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
1591                 return(EROFS);
1592         if (ip->flags & HAMMER_INODE_RO)
1593                 return (EROFS);
1594         if (hammer_nohistory(ip) == 0 &&
1595             (error = hammer_checkspace(ip->hmp, HAMMER_CHECKSPACE_SLOP_REMOVE)) != 0) {
1596                 return (error);
1597         }
1598
1599         hammer_start_transaction(&trans, ip->hmp);
1600         error = 0;
1601
1602         if (vap->va_flags != VNOVAL) {
1603                 flags = ip->ino_data.uflags;
1604                 error = vop_helper_setattr_flags(&flags, vap->va_flags,
1605                                          hammer_to_unix_xid(&ip->ino_data.uid),
1606                                          ap->a_cred);
1607                 if (error == 0) {
1608                         if (ip->ino_data.uflags != flags) {
1609                                 ip->ino_data.uflags = flags;
1610                                 modflags |= HAMMER_INODE_DDIRTY;
1611                         }
1612                         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
1613                                 error = 0;
1614                                 goto done;
1615                         }
1616                 }
1617                 goto done;
1618         }
1619         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
1620                 error = EPERM;
1621                 goto done;
1622         }
1623         if (vap->va_uid != (uid_t)VNOVAL || vap->va_gid != (gid_t)VNOVAL) {
1624                 mode_t cur_mode = ip->ino_data.mode;
1625                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1626                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1627                 uuid_t uuid_uid;
1628                 uuid_t uuid_gid;
1629
1630                 error = vop_helper_chown(ap->a_vp, vap->va_uid, vap->va_gid,
1631                                          ap->a_cred,
1632                                          &cur_uid, &cur_gid, &cur_mode);
1633                 if (error == 0) {
1634                         hammer_guid_to_uuid(&uuid_uid, cur_uid);
1635                         hammer_guid_to_uuid(&uuid_gid, cur_gid);
1636                         if (bcmp(&uuid_uid, &ip->ino_data.uid,
1637                                  sizeof(uuid_uid)) ||
1638                             bcmp(&uuid_gid, &ip->ino_data.gid,
1639                                  sizeof(uuid_gid)) ||
1640                             ip->ino_data.mode != cur_mode
1641                         ) {
1642                                 ip->ino_data.uid = uuid_uid;
1643                                 ip->ino_data.gid = uuid_gid;
1644                                 ip->ino_data.mode = cur_mode;
1645                         }
1646                         modflags |= HAMMER_INODE_DDIRTY;
1647                 }
1648         }
1649         while (vap->va_size != VNOVAL && ip->ino_data.size != vap->va_size) {
1650                 switch(ap->a_vp->v_type) {
1651                 case VREG:
1652                         if (vap->va_size == ip->ino_data.size)
1653                                 break;
1654                         /*
1655                          * XXX break atomicy, we can deadlock the backend
1656                          * if we do not release the lock.  Probably not a
1657                          * big deal here.
1658                          */
1659                         blksize = hammer_blocksize(vap->va_size);
1660                         if (vap->va_size < ip->ino_data.size) {
1661                                 vtruncbuf(ap->a_vp, vap->va_size, blksize);
1662                                 truncating = 1;
1663                         } else {
1664                                 vnode_pager_setsize(ap->a_vp, vap->va_size);
1665                                 truncating = 0;
1666                         }
1667                         ip->ino_data.size = vap->va_size;
1668                         modflags |= HAMMER_INODE_DDIRTY;
1669
1670                         /*
1671                          * on-media truncation is cached in the inode until
1672                          * the inode is synchronized.
1673                          */
1674                         if (truncating) {
1675                                 hammer_ip_frontend_trunc(ip, vap->va_size);
1676 #ifdef DEBUG_TRUNCATE
1677                                 if (HammerTruncIp == NULL)
1678                                         HammerTruncIp = ip;
1679 #endif
1680                                 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
1681                                         ip->flags |= HAMMER_INODE_TRUNCATED;
1682                                         ip->trunc_off = vap->va_size;
1683 #ifdef DEBUG_TRUNCATE
1684                                         if (ip == HammerTruncIp)
1685                                         kprintf("truncate1 %016llx\n", ip->trunc_off);
1686 #endif
1687                                 } else if (ip->trunc_off > vap->va_size) {
1688                                         ip->trunc_off = vap->va_size;
1689 #ifdef DEBUG_TRUNCATE
1690                                         if (ip == HammerTruncIp)
1691                                         kprintf("truncate2 %016llx\n", ip->trunc_off);
1692 #endif
1693                                 } else {
1694 #ifdef DEBUG_TRUNCATE
1695                                         if (ip == HammerTruncIp)
1696                                         kprintf("truncate3 %016llx (ignored)\n", vap->va_size);
1697 #endif
1698                                 }
1699                         }
1700
1701                         /*
1702                          * If truncating we have to clean out a portion of
1703                          * the last block on-disk.  We do this in the
1704                          * front-end buffer cache.
1705                          */
1706                         aligned_size = (vap->va_size + (blksize - 1)) &
1707                                        ~(int64_t)(blksize - 1);
1708                         if (truncating && vap->va_size < aligned_size) {
1709                                 struct buf *bp;
1710                                 int offset;
1711
1712                                 aligned_size -= blksize;
1713
1714                                 offset = (int)vap->va_size & (blksize - 1);
1715                                 error = bread(ap->a_vp, aligned_size,
1716                                               blksize, &bp);
1717                                 hammer_ip_frontend_trunc(ip, aligned_size);
1718                                 if (error == 0) {
1719                                         bzero(bp->b_data + offset,
1720                                               blksize - offset);
1721                                         bdwrite(bp);
1722                                 } else {
1723                                         kprintf("ERROR %d\n", error);
1724                                         brelse(bp);
1725                                 }
1726                         }
1727                         break;
1728                 case VDATABASE:
1729                         if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
1730                                 ip->flags |= HAMMER_INODE_TRUNCATED;
1731                                 ip->trunc_off = vap->va_size;
1732                         } else if (ip->trunc_off > vap->va_size) {
1733                                 ip->trunc_off = vap->va_size;
1734                         }
1735                         hammer_ip_frontend_trunc(ip, vap->va_size);
1736                         ip->ino_data.size = vap->va_size;
1737                         modflags |= HAMMER_INODE_DDIRTY;
1738                         break;
1739                 default:
1740                         error = EINVAL;
1741                         goto done;
1742                 }
1743                 break;
1744         }
1745         if (vap->va_atime.tv_sec != VNOVAL) {
1746                 ip->ino_data.atime =
1747                         hammer_timespec_to_time(&vap->va_atime);
1748                 modflags |= HAMMER_INODE_ATIME;
1749         }
1750         if (vap->va_mtime.tv_sec != VNOVAL) {
1751                 ip->ino_data.mtime =
1752                         hammer_timespec_to_time(&vap->va_mtime);
1753                 modflags |= HAMMER_INODE_MTIME;
1754         }
1755         if (vap->va_mode != (mode_t)VNOVAL) {
1756                 mode_t   cur_mode = ip->ino_data.mode;
1757                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1758                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1759
1760                 error = vop_helper_chmod(ap->a_vp, vap->va_mode, ap->a_cred,
1761                                          cur_uid, cur_gid, &cur_mode);
1762                 if (error == 0 && ip->ino_data.mode != cur_mode) {
1763                         ip->ino_data.mode = cur_mode;
1764                         modflags |= HAMMER_INODE_DDIRTY;
1765                 }
1766         }
1767 done:
1768         if (error == 0)
1769                 hammer_modify_inode(ip, modflags);
1770         hammer_done_transaction(&trans);
1771         return (error);
1772 }
1773
1774 /*
1775  * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
1776  */
1777 static
1778 int
1779 hammer_vop_nsymlink(struct vop_nsymlink_args *ap)
1780 {
1781         struct hammer_transaction trans;
1782         struct hammer_inode *dip;
1783         struct hammer_inode *nip;
1784         struct nchandle *nch;
1785         hammer_record_t record;
1786         int error;
1787         int bytes;
1788
1789         ap->a_vap->va_type = VLNK;
1790
1791         nch = ap->a_nch;
1792         dip = VTOI(ap->a_dvp);
1793
1794         if (dip->flags & HAMMER_INODE_RO)
1795                 return (EROFS);
1796         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHECKSPACE_SLOP_CREATE)) != 0)
1797                 return (error);
1798
1799         /*
1800          * Create a transaction to cover the operations we perform.
1801          */
1802         hammer_start_transaction(&trans, dip->hmp);
1803
1804         /*
1805          * Create a new filesystem object of the requested type.  The
1806          * returned inode will be referenced but not locked.
1807          */
1808
1809         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1810                                     dip, 0, &nip);
1811         if (error) {
1812                 hammer_done_transaction(&trans);
1813                 *ap->a_vpp = NULL;
1814                 return (error);
1815         }
1816
1817         /*
1818          * Add a record representing the symlink.  symlink stores the link
1819          * as pure data, not a string, and is no \0 terminated.
1820          */
1821         if (error == 0) {
1822                 bytes = strlen(ap->a_target);
1823
1824                 if (bytes <= HAMMER_INODE_BASESYMLEN) {
1825                         bcopy(ap->a_target, nip->ino_data.ext.symlink, bytes);
1826                 } else {
1827                         record = hammer_alloc_mem_record(nip, bytes);
1828                         record->type = HAMMER_MEM_RECORD_GENERAL;
1829
1830                         record->leaf.base.localization = nip->obj_localization +
1831                                                          HAMMER_LOCALIZE_MISC;
1832                         record->leaf.base.key = HAMMER_FIXKEY_SYMLINK;
1833                         record->leaf.base.rec_type = HAMMER_RECTYPE_FIX;
1834                         record->leaf.data_len = bytes;
1835                         KKASSERT(HAMMER_SYMLINK_NAME_OFF == 0);
1836                         bcopy(ap->a_target, record->data->symlink.name, bytes);
1837                         error = hammer_ip_add_record(&trans, record);
1838                 }
1839
1840                 /*
1841                  * Set the file size to the length of the link.
1842                  */
1843                 if (error == 0) {
1844                         nip->ino_data.size = bytes;
1845                         hammer_modify_inode(nip, HAMMER_INODE_DDIRTY);
1846                 }
1847         }
1848         if (error == 0)
1849                 error = hammer_ip_add_directory(&trans, dip, nch->ncp->nc_name,
1850                                                 nch->ncp->nc_nlen, nip);
1851
1852         /*
1853          * Finish up.
1854          */
1855         if (error) {
1856                 hammer_rel_inode(nip, 0);
1857                 *ap->a_vpp = NULL;
1858         } else {
1859                 error = hammer_get_vnode(nip, ap->a_vpp);
1860                 hammer_rel_inode(nip, 0);
1861                 if (error == 0) {
1862                         cache_setunresolved(ap->a_nch);
1863                         cache_setvp(ap->a_nch, *ap->a_vpp);
1864                 }
1865         }
1866         hammer_done_transaction(&trans);
1867         return (error);
1868 }
1869
1870 /*
1871  * hammer_vop_nwhiteout { nch, dvp, cred, flags }
1872  */
1873 static
1874 int
1875 hammer_vop_nwhiteout(struct vop_nwhiteout_args *ap)
1876 {
1877         struct hammer_transaction trans;
1878         struct hammer_inode *dip;
1879         int error;
1880
1881         dip = VTOI(ap->a_dvp);
1882
1883         if (hammer_nohistory(dip) == 0 &&
1884             (error = hammer_checkspace(dip->hmp, HAMMER_CHECKSPACE_SLOP_CREATE)) != 0) {
1885                 return (error);
1886         }
1887
1888         hammer_start_transaction(&trans, dip->hmp);
1889         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp,
1890                                 ap->a_cred, ap->a_flags);
1891         hammer_done_transaction(&trans);
1892
1893         return (error);
1894 }
1895
1896 /*
1897  * hammer_vop_ioctl { vp, command, data, fflag, cred }
1898  */
1899 static
1900 int
1901 hammer_vop_ioctl(struct vop_ioctl_args *ap)
1902 {
1903         struct hammer_inode *ip = ap->a_vp->v_data;
1904
1905         return(hammer_ioctl(ip, ap->a_command, ap->a_data,
1906                             ap->a_fflag, ap->a_cred));
1907 }
1908
1909 static
1910 int
1911 hammer_vop_mountctl(struct vop_mountctl_args *ap)
1912 {
1913         struct mount *mp;
1914         int error;
1915
1916         mp = ap->a_head.a_ops->head.vv_mount;
1917
1918         switch(ap->a_op) {
1919         case MOUNTCTL_SET_EXPORT:
1920                 if (ap->a_ctllen != sizeof(struct export_args))
1921                         error = EINVAL;
1922                 error = hammer_vfs_export(mp, ap->a_op,
1923                                       (const struct export_args *)ap->a_ctl);
1924                 break;
1925         default:
1926                 error = journal_mountctl(ap);
1927                 break;
1928         }
1929         return(error);
1930 }
1931
1932 /*
1933  * hammer_vop_strategy { vp, bio }
1934  *
1935  * Strategy call, used for regular file read & write only.  Note that the
1936  * bp may represent a cluster.
1937  *
1938  * To simplify operation and allow better optimizations in the future,
1939  * this code does not make any assumptions with regards to buffer alignment
1940  * or size.
1941  */
1942 static
1943 int
1944 hammer_vop_strategy(struct vop_strategy_args *ap)
1945 {
1946         struct buf *bp;
1947         int error;
1948
1949         bp = ap->a_bio->bio_buf;
1950
1951         switch(bp->b_cmd) {
1952         case BUF_CMD_READ:
1953                 error = hammer_vop_strategy_read(ap);
1954                 break;
1955         case BUF_CMD_WRITE:
1956                 error = hammer_vop_strategy_write(ap);
1957                 break;
1958         default:
1959                 bp->b_error = error = EINVAL;
1960                 bp->b_flags |= B_ERROR;
1961                 biodone(ap->a_bio);
1962                 break;
1963         }
1964         return (error);
1965 }
1966
1967 /*
1968  * Read from a regular file.  Iterate the related records and fill in the
1969  * BIO/BUF.  Gaps are zero-filled.
1970  *
1971  * The support code in hammer_object.c should be used to deal with mixed
1972  * in-memory and on-disk records.
1973  *
1974  * NOTE: Can be called from the cluster code with an oversized buf.
1975  *
1976  * XXX atime update
1977  */
1978 static
1979 int
1980 hammer_vop_strategy_read(struct vop_strategy_args *ap)
1981 {
1982         struct hammer_transaction trans;
1983         struct hammer_inode *ip;
1984         struct hammer_cursor cursor;
1985         hammer_base_elm_t base;
1986         hammer_off_t disk_offset;
1987         struct bio *bio;
1988         struct bio *nbio;
1989         struct buf *bp;
1990         int64_t rec_offset;
1991         int64_t ran_end;
1992         int64_t tmp64;
1993         int error;
1994         int boff;
1995         int roff;
1996         int n;
1997
1998         bio = ap->a_bio;
1999         bp = bio->bio_buf;
2000         ip = ap->a_vp->v_data;
2001
2002         /*
2003          * The zone-2 disk offset may have been set by the cluster code via
2004          * a BMAP operation, or else should be NOOFFSET.
2005          *
2006          * Checking the high bits for a match against zone-2 should suffice.
2007          */
2008         nbio = push_bio(bio);
2009         if ((nbio->bio_offset & HAMMER_OFF_ZONE_MASK) ==
2010             HAMMER_ZONE_RAW_BUFFER) {
2011                 error = hammer_io_direct_read(ip->hmp, nbio);
2012                 return (error);
2013         }
2014
2015         /*
2016          * Well, that sucked.  Do it the hard way.  If all the stars are
2017          * aligned we may still be able to issue a direct-read.
2018          */
2019         hammer_simple_transaction(&trans, ip->hmp);
2020         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2021
2022         /*
2023          * Key range (begin and end inclusive) to scan.  Note that the key's
2024          * stored in the actual records represent BASE+LEN, not BASE.  The
2025          * first record containing bio_offset will have a key > bio_offset.
2026          */
2027         cursor.key_beg.localization = ip->obj_localization +
2028                                       HAMMER_LOCALIZE_MISC;
2029         cursor.key_beg.obj_id = ip->obj_id;
2030         cursor.key_beg.create_tid = 0;
2031         cursor.key_beg.delete_tid = 0;
2032         cursor.key_beg.obj_type = 0;
2033         cursor.key_beg.key = bio->bio_offset + 1;
2034         cursor.asof = ip->obj_asof;
2035         cursor.flags |= HAMMER_CURSOR_ASOF;
2036
2037         cursor.key_end = cursor.key_beg;
2038         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2039 #if 0
2040         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
2041                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
2042                 cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
2043                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2044         } else
2045 #endif
2046         {
2047                 ran_end = bio->bio_offset + bp->b_bufsize;
2048                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2049                 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2050                 tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2051                 if (tmp64 < ran_end)
2052                         cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2053                 else
2054                         cursor.key_end.key = ran_end + MAXPHYS + 1;
2055         }
2056         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2057
2058         error = hammer_ip_first(&cursor);
2059         boff = 0;
2060
2061         while (error == 0) {
2062                 /*
2063                  * Get the base file offset of the record.  The key for
2064                  * data records is (base + bytes) rather then (base).
2065                  */
2066                 base = &cursor.leaf->base;
2067                 rec_offset = base->key - cursor.leaf->data_len;
2068
2069                 /*
2070                  * Calculate the gap, if any, and zero-fill it.
2071                  *
2072                  * n is the offset of the start of the record verses our
2073                  * current seek offset in the bio.
2074                  */
2075                 n = (int)(rec_offset - (bio->bio_offset + boff));
2076                 if (n > 0) {
2077                         if (n > bp->b_bufsize - boff)
2078                                 n = bp->b_bufsize - boff;
2079                         bzero((char *)bp->b_data + boff, n);
2080                         boff += n;
2081                         n = 0;
2082                 }
2083
2084                 /*
2085                  * Calculate the data offset in the record and the number
2086                  * of bytes we can copy.
2087                  *
2088                  * There are two degenerate cases.  First, boff may already
2089                  * be at bp->b_bufsize.  Secondly, the data offset within
2090                  * the record may exceed the record's size.
2091                  */
2092                 roff = -n;
2093                 rec_offset += roff;
2094                 n = cursor.leaf->data_len - roff;
2095                 if (n <= 0) {
2096                         kprintf("strategy_read: bad n=%d roff=%d\n", n, roff);
2097                         n = 0;
2098                 } else if (n > bp->b_bufsize - boff) {
2099                         n = bp->b_bufsize - boff;
2100                 }
2101
2102                 /*
2103                  * Deal with cached truncations.  This cool bit of code
2104                  * allows truncate()/ftruncate() to avoid having to sync
2105                  * the file.
2106                  *
2107                  * If the frontend is truncated then all backend records are
2108                  * subject to the frontend's truncation.
2109                  *
2110                  * If the backend is truncated then backend records on-disk
2111                  * (but not in-memory) are subject to the backend's
2112                  * truncation.  In-memory records owned by the backend
2113                  * represent data written after the truncation point on the
2114                  * backend and must not be truncated.
2115                  *
2116                  * Truncate operations deal with frontend buffer cache
2117                  * buffers and frontend-owned in-memory records synchronously.
2118                  */
2119                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2120                         if (hammer_cursor_ondisk(&cursor) ||
2121                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
2122                                 if (ip->trunc_off <= rec_offset)
2123                                         n = 0;
2124                                 else if (ip->trunc_off < rec_offset + n)
2125                                         n = (int)(ip->trunc_off - rec_offset);
2126                         }
2127                 }
2128                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2129                         if (hammer_cursor_ondisk(&cursor)) {
2130                                 if (ip->sync_trunc_off <= rec_offset)
2131                                         n = 0;
2132                                 else if (ip->sync_trunc_off < rec_offset + n)
2133                                         n = (int)(ip->sync_trunc_off - rec_offset);
2134                         }
2135                 }
2136
2137                 /*
2138                  * Try to issue a direct read into our bio if possible,
2139                  * otherwise resolve the element data into a hammer_buffer
2140                  * and copy.
2141                  *
2142                  * The buffer on-disk should be zerod past any real
2143                  * truncation point, but may not be for any synthesized
2144                  * truncation point from above.
2145                  */
2146                 if (boff == 0 && n == bp->b_bufsize &&
2147                     ((cursor.leaf->data_offset + roff) & HAMMER_BUFMASK) == 0) {
2148                         disk_offset = hammer_blockmap_lookup(
2149                                                 trans.hmp,
2150                                                 cursor.leaf->data_offset + roff,
2151                                                 &error);
2152                         if (error)
2153                                 break;
2154                         nbio->bio_offset = disk_offset;
2155                         error = hammer_io_direct_read(trans.hmp, nbio);
2156                         goto done;
2157                 } else if (n) {
2158                         error = hammer_ip_resolve_data(&cursor);
2159                         if (error == 0) {
2160                                 bcopy((char *)cursor.data + roff,
2161                                       (char *)bp->b_data + boff, n);
2162                         }
2163                 }
2164                 if (error)
2165                         break;
2166
2167                 /*
2168                  * Iterate until we have filled the request.
2169                  */
2170                 boff += n;
2171                 if (boff == bp->b_bufsize)
2172                         break;
2173                 error = hammer_ip_next(&cursor);
2174         }
2175
2176         /*
2177          * There may have been a gap after the last record
2178          */
2179         if (error == ENOENT)
2180                 error = 0;
2181         if (error == 0 && boff != bp->b_bufsize) {
2182                 KKASSERT(boff < bp->b_bufsize);
2183                 bzero((char *)bp->b_data + boff, bp->b_bufsize - boff);
2184                 /* boff = bp->b_bufsize; */
2185         }
2186         bp->b_resid = 0;
2187         bp->b_error = error;
2188         if (error)
2189                 bp->b_flags |= B_ERROR;
2190         biodone(ap->a_bio);
2191
2192 done:
2193         if (cursor.node)
2194                 hammer_cache_node(&ip->cache[1], cursor.node);
2195         hammer_done_cursor(&cursor);
2196         hammer_done_transaction(&trans);
2197         return(error);
2198 }
2199
2200 /*
2201  * BMAP operation - used to support cluster_read() only.
2202  *
2203  * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
2204  *
2205  * This routine may return EOPNOTSUPP if the opration is not supported for
2206  * the specified offset.  The contents of the pointer arguments do not
2207  * need to be initialized in that case.
2208  *
2209  * If a disk address is available and properly aligned return 0 with
2210  * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
2211  * to the run-length relative to that offset.  Callers may assume that
2212  * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
2213  * large, so return EOPNOTSUPP if it is not sufficiently large.
2214  */
2215 static
2216 int
2217 hammer_vop_bmap(struct vop_bmap_args *ap)
2218 {
2219         struct hammer_transaction trans;
2220         struct hammer_inode *ip;
2221         struct hammer_cursor cursor;
2222         hammer_base_elm_t base;
2223         int64_t rec_offset;
2224         int64_t ran_end;
2225         int64_t tmp64;
2226         int64_t base_offset;
2227         int64_t base_disk_offset;
2228         int64_t last_offset;
2229         hammer_off_t last_disk_offset;
2230         hammer_off_t disk_offset;
2231         int     rec_len;
2232         int     error;
2233         int     blksize;
2234
2235         ip = ap->a_vp->v_data;
2236
2237         /*
2238          * We can only BMAP regular files.  We can't BMAP database files,
2239          * directories, etc.
2240          */
2241         if (ip->ino_data.obj_type != HAMMER_OBJTYPE_REGFILE)
2242                 return(EOPNOTSUPP);
2243
2244         /*
2245          * bmap is typically called with runp/runb both NULL when used
2246          * for writing.  We do not support BMAP for writing atm.
2247          */
2248         if (ap->a_cmd != BUF_CMD_READ)
2249                 return(EOPNOTSUPP);
2250
2251         /*
2252          * Scan the B-Tree to acquire blockmap addresses, then translate
2253          * to raw addresses.
2254          */
2255         hammer_simple_transaction(&trans, ip->hmp);
2256 #if 0
2257         kprintf("bmap_beg %016llx ip->cache %p\n", ap->a_loffset, ip->cache[1]);
2258 #endif
2259         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2260
2261         /*
2262          * Key range (begin and end inclusive) to scan.  Note that the key's
2263          * stored in the actual records represent BASE+LEN, not BASE.  The
2264          * first record containing bio_offset will have a key > bio_offset.
2265          */
2266         cursor.key_beg.localization = ip->obj_localization +
2267                                       HAMMER_LOCALIZE_MISC;
2268         cursor.key_beg.obj_id = ip->obj_id;
2269         cursor.key_beg.create_tid = 0;
2270         cursor.key_beg.delete_tid = 0;
2271         cursor.key_beg.obj_type = 0;
2272         if (ap->a_runb)
2273                 cursor.key_beg.key = ap->a_loffset - MAXPHYS + 1;
2274         else
2275                 cursor.key_beg.key = ap->a_loffset + 1;
2276         if (cursor.key_beg.key < 0)
2277                 cursor.key_beg.key = 0;
2278         cursor.asof = ip->obj_asof;
2279         cursor.flags |= HAMMER_CURSOR_ASOF;
2280
2281         cursor.key_end = cursor.key_beg;
2282         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2283
2284         ran_end = ap->a_loffset + MAXPHYS;
2285         cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2286         cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2287         tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2288         if (tmp64 < ran_end)
2289                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2290         else
2291                 cursor.key_end.key = ran_end + MAXPHYS + 1;
2292
2293         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2294
2295         error = hammer_ip_first(&cursor);
2296         base_offset = last_offset = 0;
2297         base_disk_offset = last_disk_offset = 0;
2298
2299         while (error == 0) {
2300                 /*
2301                  * Get the base file offset of the record.  The key for
2302                  * data records is (base + bytes) rather then (base).
2303                  *
2304                  * NOTE: rec_offset + rec_len may exceed the end-of-file.
2305                  * The extra bytes should be zero on-disk and the BMAP op
2306                  * should still be ok.
2307                  */
2308                 base = &cursor.leaf->base;
2309                 rec_offset = base->key - cursor.leaf->data_len;
2310                 rec_len    = cursor.leaf->data_len;
2311
2312                 /*
2313                  * Incorporate any cached truncation.
2314                  *
2315                  * NOTE: Modifications to rec_len based on synthesized
2316                  * truncation points remove the guarantee that any extended
2317                  * data on disk is zero (since the truncations may not have
2318                  * taken place on-media yet).
2319                  */
2320                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2321                         if (hammer_cursor_ondisk(&cursor) ||
2322                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
2323                                 if (ip->trunc_off <= rec_offset)
2324                                         rec_len = 0;
2325                                 else if (ip->trunc_off < rec_offset + rec_len)
2326                                         rec_len = (int)(ip->trunc_off - rec_offset);
2327                         }
2328                 }
2329                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2330                         if (hammer_cursor_ondisk(&cursor)) {
2331                                 if (ip->sync_trunc_off <= rec_offset)
2332                                         rec_len = 0;
2333                                 else if (ip->sync_trunc_off < rec_offset + rec_len)
2334                                         rec_len = (int)(ip->sync_trunc_off - rec_offset);
2335                         }
2336                 }
2337
2338                 /*
2339                  * Accumulate information.  If we have hit a discontiguous
2340                  * block reset base_offset unless we are already beyond the
2341                  * requested offset.  If we are, that's it, we stop.
2342                  */
2343                 disk_offset = hammer_blockmap_lookup(trans.hmp,
2344                                                      cursor.leaf->data_offset,
2345                                                      &error);
2346                 if (error)
2347                         break;
2348                 if (rec_offset != last_offset ||
2349                     disk_offset != last_disk_offset) {
2350                         if (rec_offset > ap->a_loffset)
2351                                 break;
2352                         base_offset = rec_offset;
2353                         base_disk_offset = disk_offset;
2354                 }
2355                 last_offset = rec_offset + rec_len;
2356                 last_disk_offset = disk_offset + rec_len;
2357
2358                 error = hammer_ip_next(&cursor);
2359         }
2360
2361 #if 0
2362         kprintf("BMAP %016llx:  %016llx - %016llx\n",
2363                 ap->a_loffset, base_offset, last_offset);
2364         kprintf("BMAP %16s:  %016llx - %016llx\n",
2365                 "", base_disk_offset, last_disk_offset);
2366 #endif
2367
2368         if (cursor.node) {
2369                 hammer_cache_node(&ip->cache[1], cursor.node);
2370 #if 0
2371                 kprintf("bmap_end2 %016llx ip->cache %p\n", ap->a_loffset, ip->cache[1]);
2372 #endif
2373         }
2374         hammer_done_cursor(&cursor);
2375         hammer_done_transaction(&trans);
2376
2377         /*
2378          * If we couldn't find any records or the records we did find were
2379          * all behind the requested offset, return failure.  A forward
2380          * truncation can leave a hole w/ no on-disk records.
2381          */
2382         if (last_offset == 0 || last_offset < ap->a_loffset)
2383                 return (EOPNOTSUPP);
2384
2385         /*
2386          * Figure out the block size at the requested offset and adjust
2387          * our limits so the cluster_read() does not create inappropriately
2388          * sized buffer cache buffers.
2389          */
2390         blksize = hammer_blocksize(ap->a_loffset);
2391         if (hammer_blocksize(base_offset) != blksize) {
2392                 base_offset = hammer_blockdemarc(base_offset, ap->a_loffset);
2393         }
2394         if (last_offset != ap->a_loffset &&
2395             hammer_blocksize(last_offset - 1) != blksize) {
2396                 last_offset = hammer_blockdemarc(ap->a_loffset,
2397                                                  last_offset - 1);
2398         }
2399
2400         /*
2401          * Returning EOPNOTSUPP simply prevents the direct-IO optimization
2402          * from occuring.
2403          */
2404         disk_offset = base_disk_offset + (ap->a_loffset - base_offset);
2405
2406         /*
2407          * If doffsetp is not aligned or the forward run size does
2408          * not cover a whole buffer, disallow the direct I/O.
2409          */
2410         if ((disk_offset & HAMMER_BUFMASK) ||
2411             (last_offset - ap->a_loffset) < blksize) {
2412                 error = EOPNOTSUPP;
2413         } else {
2414                 *ap->a_doffsetp = disk_offset;
2415                 if (ap->a_runb) {
2416                         *ap->a_runb = ap->a_loffset - base_offset;
2417                         KKASSERT(*ap->a_runb >= 0);
2418                 }
2419                 if (ap->a_runp) {
2420                         *ap->a_runp = last_offset - ap->a_loffset;
2421                         KKASSERT(*ap->a_runp >= 0);
2422                 }
2423                 error = 0;
2424         }
2425         return(error);
2426 }
2427
2428 /*
2429  * Write to a regular file.   Because this is a strategy call the OS is
2430  * trying to actually get data onto the media.
2431  */
2432 static
2433 int
2434 hammer_vop_strategy_write(struct vop_strategy_args *ap)
2435 {
2436         hammer_record_t record;
2437         hammer_mount_t hmp;
2438         hammer_inode_t ip;
2439         struct bio *bio;
2440         struct buf *bp;
2441         int blksize;
2442         int bytes;
2443         int error;
2444
2445         bio = ap->a_bio;
2446         bp = bio->bio_buf;
2447         ip = ap->a_vp->v_data;
2448         hmp = ip->hmp;
2449
2450         blksize = hammer_blocksize(bio->bio_offset);
2451         KKASSERT(bp->b_bufsize == blksize);
2452
2453         if (ip->flags & HAMMER_INODE_RO) {
2454                 bp->b_error = EROFS;
2455                 bp->b_flags |= B_ERROR;
2456                 biodone(ap->a_bio);
2457                 return(EROFS);
2458         }
2459
2460         /*
2461          * Interlock with inode destruction (no in-kernel or directory
2462          * topology visibility).  If we queue new IO while trying to
2463          * destroy the inode we can deadlock the vtrunc call in
2464          * hammer_inode_unloadable_check().
2465          */
2466         if (ip->flags & (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) {
2467                 bp->b_resid = 0;
2468                 biodone(ap->a_bio);
2469                 return(0);
2470         }
2471
2472         /*
2473          * Reserve space and issue a direct-write from the front-end.
2474          * NOTE: The direct_io code will hammer_bread/bcopy smaller
2475          * allocations.
2476          *
2477          * An in-memory record will be installed to reference the storage
2478          * until the flusher can get to it.
2479          *
2480          * Since we own the high level bio the front-end will not try to
2481          * do a direct-read until the write completes.
2482          *
2483          * NOTE: The only time we do not reserve a full-sized buffers
2484          * worth of data is if the file is small.  We do not try to
2485          * allocate a fragment (from the small-data zone) at the end of
2486          * an otherwise large file as this can lead to wildly separated
2487          * data.
2488          */
2489         KKASSERT((bio->bio_offset & HAMMER_BUFMASK) == 0);
2490         KKASSERT(bio->bio_offset < ip->ino_data.size);
2491         if (bio->bio_offset || ip->ino_data.size > HAMMER_BUFSIZE / 2)
2492                 bytes = bp->b_bufsize;
2493         else
2494                 bytes = ((int)ip->ino_data.size + 15) & ~15;
2495
2496         record = hammer_ip_add_bulk(ip, bio->bio_offset, bp->b_data,
2497                                     bytes, &error);
2498         if (record) {
2499                 hammer_io_direct_write(hmp, &record->leaf, bio);
2500                 hammer_rel_mem_record(record);
2501                 if (ip->rsv_recs > 1 && hmp->rsv_recs > hammer_limit_recs)
2502                         hammer_flush_inode(ip, 0);
2503         } else {
2504                 bp->b_bio2.bio_offset = NOOFFSET;
2505                 bp->b_error = error;
2506                 bp->b_flags |= B_ERROR;
2507                 biodone(ap->a_bio);
2508         }
2509         return(error);
2510 }
2511
2512 /*
2513  * dounlink - disconnect a directory entry
2514  *
2515  * XXX whiteout support not really in yet
2516  */
2517 static int
2518 hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
2519                 struct vnode *dvp, struct ucred *cred, int flags)
2520 {
2521         struct namecache *ncp;
2522         hammer_inode_t dip;
2523         hammer_inode_t ip;
2524         struct hammer_cursor cursor;
2525         int64_t namekey;
2526         int nlen, error;
2527
2528         /*
2529          * Calculate the namekey and setup the key range for the scan.  This
2530          * works kinda like a chained hash table where the lower 32 bits
2531          * of the namekey synthesize the chain.
2532          *
2533          * The key range is inclusive of both key_beg and key_end.
2534          */
2535         dip = VTOI(dvp);
2536         ncp = nch->ncp;
2537
2538         if (dip->flags & HAMMER_INODE_RO)
2539                 return (EROFS);
2540
2541         namekey = hammer_directory_namekey(ncp->nc_name, ncp->nc_nlen);
2542 retry:
2543         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
2544         cursor.key_beg.localization = dip->obj_localization +
2545                                       HAMMER_LOCALIZE_MISC;
2546         cursor.key_beg.obj_id = dip->obj_id;
2547         cursor.key_beg.key = namekey;
2548         cursor.key_beg.create_tid = 0;
2549         cursor.key_beg.delete_tid = 0;
2550         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
2551         cursor.key_beg.obj_type = 0;
2552
2553         cursor.key_end = cursor.key_beg;
2554         cursor.key_end.key |= 0xFFFFFFFFULL;
2555         cursor.asof = dip->obj_asof;
2556         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2557
2558         /*
2559          * Scan all matching records (the chain), locate the one matching
2560          * the requested path component.  info->last_error contains the
2561          * error code on search termination and could be 0, ENOENT, or
2562          * something else.
2563          *
2564          * The hammer_ip_*() functions merge in-memory records with on-disk
2565          * records for the purposes of the search.
2566          */
2567         error = hammer_ip_first(&cursor);
2568
2569         while (error == 0) {
2570                 error = hammer_ip_resolve_data(&cursor);
2571                 if (error)
2572                         break;
2573                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
2574                 KKASSERT(nlen > 0);
2575                 if (ncp->nc_nlen == nlen &&
2576                     bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
2577                         break;
2578                 }
2579                 error = hammer_ip_next(&cursor);
2580         }
2581
2582         /*
2583          * If all is ok we have to get the inode so we can adjust nlinks.
2584          * To avoid a deadlock with the flusher we must release the inode
2585          * lock on the directory when acquiring the inode for the entry.
2586          *
2587          * If the target is a directory, it must be empty.
2588          */
2589         if (error == 0) {
2590                 hammer_unlock(&cursor.ip->lock);
2591                 ip = hammer_get_inode(trans, dip, cursor.data->entry.obj_id,
2592                                       dip->hmp->asof,
2593                                       cursor.data->entry.localization,
2594                                       0, &error);
2595                 hammer_lock_sh(&cursor.ip->lock);
2596                 if (error == ENOENT) {
2597                         kprintf("obj_id %016llx\n", cursor.data->entry.obj_id);
2598                         Debugger("ENOENT unlinking object that should exist");
2599                 }
2600
2601                 /*
2602                  * If we are trying to remove a directory the directory must
2603                  * be empty.
2604                  *
2605                  * WARNING: hammer_ip_check_directory_empty() may have to
2606                  * terminate the cursor to avoid a deadlock.  It is ok to
2607                  * call hammer_done_cursor() twice.
2608                  */
2609                 if (error == 0 && ip->ino_data.obj_type ==
2610                                   HAMMER_OBJTYPE_DIRECTORY) {
2611                         error = hammer_ip_check_directory_empty(trans, ip);
2612                 }
2613
2614                 /*
2615                  * Delete the directory entry.
2616                  *
2617                  * WARNING: hammer_ip_del_directory() may have to terminate
2618                  * the cursor to avoid a deadlock.  It is ok to call
2619                  * hammer_done_cursor() twice.
2620                  */
2621                 if (error == 0) {
2622                         error = hammer_ip_del_directory(trans, &cursor,
2623                                                         dip, ip);
2624                 }
2625                 hammer_done_cursor(&cursor);
2626                 if (error == 0) {
2627                         cache_setunresolved(nch);
2628                         cache_setvp(nch, NULL);
2629                         /* XXX locking */
2630                         if (ip->vp)
2631                                 cache_inval_vp(ip->vp, CINV_DESTROY);
2632                 }
2633                 if (ip)
2634                         hammer_rel_inode(ip, 0);
2635         } else {
2636                 hammer_done_cursor(&cursor);
2637         }
2638         if (error == EDEADLK)
2639                 goto retry;
2640
2641         return (error);
2642 }
2643
2644 /************************************************************************
2645  *                          FIFO AND SPECFS OPS                         *
2646  ************************************************************************
2647  *
2648  */
2649
2650 static int
2651 hammer_vop_fifoclose (struct vop_close_args *ap)
2652 {
2653         /* XXX update itimes */
2654         return (VOCALL(&fifo_vnode_vops, &ap->a_head));
2655 }
2656
2657 static int
2658 hammer_vop_fiforead (struct vop_read_args *ap)
2659 {
2660         int error;
2661
2662         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
2663         /* XXX update access time */
2664         return (error);
2665 }
2666
2667 static int
2668 hammer_vop_fifowrite (struct vop_write_args *ap)
2669 {
2670         int error;
2671
2672         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
2673         /* XXX update access time */
2674         return (error);
2675 }
2676
2677 static int
2678 hammer_vop_specclose (struct vop_close_args *ap)
2679 {
2680         /* XXX update itimes */
2681         return (VOCALL(&spec_vnode_vops, &ap->a_head));
2682 }
2683
2684 static int
2685 hammer_vop_specread (struct vop_read_args *ap)
2686 {
2687         /* XXX update access time */
2688         return (VOCALL(&spec_vnode_vops, &ap->a_head));
2689 }
2690
2691 static int
2692 hammer_vop_specwrite (struct vop_write_args *ap)
2693 {
2694         /* XXX update last change time */
2695         return (VOCALL(&spec_vnode_vops, &ap->a_head));
2696 }
2697