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 1.88 2008/07/12 23:04:50 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         ++hammer_count_fsyncs;
 186         vfsync(ap->a_vp, ap->a_waitfor, 1, NULL, NULL);
 187         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 188         if (ap->a_waitfor == MNT_WAIT)
 189                 hammer_wait_inode(ip);
 190         return (ip->error);
 191 }
 192
 193 /*
 194  * hammer_vop_read { vp, uio, ioflag, cred }
 195  */
 196 static
 197 int
 198 hammer_vop_read(struct vop_read_args *ap)
 199 {
 200         struct hammer_transaction trans;
 201         hammer_inode_t ip;
 202         off_t offset;
 203         struct buf *bp;
 204         struct uio *uio;
 205         int error;
 206         int n;
 207         int seqcount;
 208         int ioseqcount;
 209         int blksize;
 210
 211         if (ap->a_vp->v_type != VREG)
 212                 return (EINVAL);
 213         ip = VTOI(ap->a_vp);
 214         error = 0;
 215         uio = ap->a_uio;
 216
 217         /*
 218          * Allow the UIO's size to override the sequential heuristic.
 219          */
 220         blksize = hammer_blocksize(uio->uio_offset);
 221         seqcount = (uio->uio_resid + (blksize - 1)) / blksize;
 222         ioseqcount = ap->a_ioflag >> 16;
 223         if (seqcount < ioseqcount)
 224                 seqcount = ioseqcount;
 225
 226         hammer_start_transaction(&trans, ip->hmp);
 227
 228         /*
 229          * Access the data typically in HAMMER_BUFSIZE blocks via the
 230          * buffer cache, but HAMMER may use a variable block size based
 231          * on the offset.
 232          */
 233         while (uio->uio_resid > 0 && uio->uio_offset < ip->ino_data.size) {
 234                 int64_t base_offset;
 235                 int64_t file_limit;
 236
 237                 blksize = hammer_blocksize(uio->uio_offset);
 238                 offset = (int)uio->uio_offset & (blksize - 1);
 239                 base_offset = uio->uio_offset - offset;
 240
 241                 if (hammer_debug_cluster_enable) {
 242                         /*
 243                          * Use file_limit to prevent cluster_read() from
 244                          * creating buffers of the wrong block size past
 245                          * the demarc.
 246                          */
 247                         file_limit = ip->ino_data.size;
 248                         if (base_offset < HAMMER_XDEMARC &&
 249                             file_limit > HAMMER_XDEMARC) {
 250                                 file_limit = HAMMER_XDEMARC;
 251                         }
 252                         error = cluster_read(ap->a_vp,
 253                                              file_limit, base_offset,
 254                                              blksize, MAXPHYS,
 255                                              seqcount, &bp);
 256                 } else {
 257                         error = bread(ap->a_vp, base_offset, blksize, &bp);
 258                 }
 259                 if (error) {
 260                         kprintf("error %d\n", error);
 261                         brelse(bp);
 262                         break;
 263                 }
 264
 265                 /* bp->b_flags |= B_CLUSTEROK; temporarily disabled */
 266                 n = blksize - offset;
 267                 if (n > uio->uio_resid)
 268                         n = uio->uio_resid;
 269                 if (n > ip->ino_data.size - uio->uio_offset)
 270                         n = (int)(ip->ino_data.size - uio->uio_offset);
 271                 error = uiomove((char *)bp->b_data + offset, n, uio);
 272
 273                 /* data has a lower priority then meta-data */
 274                 bp->b_flags |= B_AGE;
 275                 bqrelse(bp);
 276                 if (error)
 277                         break;
 278         }
 279         if ((ip->flags & HAMMER_INODE_RO) == 0 &&
 280             (ip->hmp->mp->mnt_flag & MNT_NOATIME) == 0) {
 281                 ip->ino_data.atime = trans.time;
 282                 hammer_modify_inode(ip, HAMMER_INODE_ATIME);
 283         }
 284         hammer_done_transaction(&trans);
 285         return (error);
 286 }
 287
 288 /*
 289  * hammer_vop_write { vp, uio, ioflag, cred }
 290  */
 291 static
 292 int
 293 hammer_vop_write(struct vop_write_args *ap)
 294 {
 295         struct hammer_transaction trans;
 296         struct hammer_inode *ip;
 297         hammer_mount_t hmp;
 298         struct uio *uio;
 299         int offset;
 300         off_t base_offset;
 301         struct buf *bp;
 302         int error;
 303         int n;
 304         int flags;
 305         int delta;
 306         int seqcount;
 307
 308         if (ap->a_vp->v_type != VREG)
 309                 return (EINVAL);
 310         ip = VTOI(ap->a_vp);
 311         hmp = ip->hmp;
 312         error = 0;
 313         seqcount = ap->a_ioflag >> 16;
 314
 315         if (ip->flags & HAMMER_INODE_RO)
 316                 return (EROFS);
 317
 318         /*
 319          * Create a transaction to cover the operations we perform.
 320          */
 321         hammer_start_transaction(&trans, hmp);
 322         uio = ap->a_uio;
 323
 324         /*
 325          * Check append mode
 326          */
 327         if (ap->a_ioflag & IO_APPEND)
 328                 uio->uio_offset = ip->ino_data.size;
 329
 330         /*
 331          * Check for illegal write offsets.  Valid range is 0...2^63-1.
 332          *
 333          * NOTE: the base_off assignment is required to work around what
 334          * I consider to be a GCC-4 optimization bug.
 335          */
 336         if (uio->uio_offset < 0) {
 337                 hammer_done_transaction(&trans);
 338                 return (EFBIG);
 339         }
 340         base_offset = uio->uio_offset + uio->uio_resid; /* work around gcc-4 */
 341         if (uio->uio_resid > 0 && base_offset <= 0) {
 342                 hammer_done_transaction(&trans);
 343                 return (EFBIG);
 344         }
 345
 346         /*
 347          * Access the data typically in HAMMER_BUFSIZE blocks via the
 348          * buffer cache, but HAMMER may use a variable block size based
 349          * on the offset.
 350          */
 351         while (uio->uio_resid > 0) {
 352                 int fixsize = 0;
 353                 int blksize;
 354                 int blkmask;
 355
 356                 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_WRITE)) != 0)
 357                         break;
 358
 359                 blksize = hammer_blocksize(uio->uio_offset);
 360
 361                 /*
 362                  * Do not allow HAMMER to blow out the buffer cache.  Very
 363                  * large UIOs can lockout other processes due to bwillwrite()
 364                  * mechanics.
 365                  *
 366                  * The hammer inode is not locked during these operations.
 367                  * The vnode is locked which can interfere with the pageout
 368                  * daemon for non-UIO_NOCOPY writes but should not interfere
 369                  * with the buffer cache.  Even so, we cannot afford to
 370                  * allow the pageout daemon to build up too many dirty buffer
 371                  * cache buffers.
 372                  */
 373                 /*if (((int)uio->uio_offset & (blksize - 1)) == 0)*/
 374                 bwillwrite(blksize);
 375
 376                 /*
 377                  * Do not allow HAMMER to blow out system memory by
 378                  * accumulating too many records.   Records are so well
 379                  * decoupled from the buffer cache that it is possible
 380                  * for userland to push data out to the media via
 381                  * direct-write, but build up the records queued to the
 382                  * backend faster then the backend can flush them out.
 383                  * HAMMER has hit its write limit but the frontend has
 384                  * no pushback to slow it down.
 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_CHKSPC_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, NULL, &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         return(0);
 722 }
 723
 724 /*
 725  * hammer_vop_nresolve { nch, dvp, cred }
 726  *
 727  * Locate the requested directory entry.
 728  */
 729 static
 730 int
 731 hammer_vop_nresolve(struct vop_nresolve_args *ap)
 732 {
 733         struct hammer_transaction trans;
 734         struct namecache *ncp;
 735         hammer_inode_t dip;
 736         hammer_inode_t ip;
 737         hammer_tid_t asof;
 738         struct hammer_cursor cursor;
 739         struct vnode *vp;
 740         int64_t namekey;
 741         int error;
 742         int i;
 743         int nlen;
 744         int flags;
 745         int ispfs;
 746         int64_t obj_id;
 747         u_int32_t localization;
 748
 749         /*
 750          * Misc initialization, plus handle as-of name extensions.  Look for
 751          * the '@@' extension.  Note that as-of files and directories cannot
 752          * be modified.
 753          */
 754         dip = VTOI(ap->a_dvp);
 755         ncp = ap->a_nch->ncp;
 756         asof = dip->obj_asof;
 757         nlen = ncp->nc_nlen;
 758         flags = dip->flags & HAMMER_INODE_RO;
 759         ispfs = 0;
 760
 761         hammer_simple_transaction(&trans, dip->hmp);
 762
 763         for (i = 0; i < nlen; ++i) {
 764                 if (ncp->nc_name[i] == '@' && ncp->nc_name[i+1] == '@') {
 765                         asof = hammer_str_to_tid(ncp->nc_name + i + 2,
 766                                                  &ispfs, &localization);
 767                         if (asof != HAMMER_MAX_TID)
 768                                 flags |= HAMMER_INODE_RO;
 769                         break;
 770                 }
 771         }
 772         nlen = i;
 773
 774         /*
 775          * If this is a PFS softlink we dive into the PFS
 776          */
 777         if (ispfs && nlen == 0) {
 778                 ip = hammer_get_inode(&trans, dip, HAMMER_OBJID_ROOT,
 779                                       asof, 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          * If there is no path component the time extension is relative to
 797          * dip.
 798          */
 799         if (nlen == 0) {
 800                 ip = hammer_get_inode(&trans, dip, dip->obj_id,
 801                                       asof, dip->obj_localization,
 802                                       flags, &error);
 803                 if (error == 0) {
 804                         error = hammer_get_vnode(ip, &vp);
 805                         hammer_rel_inode(ip, 0);
 806                 } else {
 807                         vp = NULL;
 808                 }
 809                 if (error == 0) {
 810                         vn_unlock(vp);
 811                         cache_setvp(ap->a_nch, vp);
 812                         vrele(vp);
 813                 }
 814                 goto done;
 815         }
 816
 817         /*
 818          * Calculate the namekey and setup the key range for the scan.  This
 819          * works kinda like a chained hash table where the lower 32 bits
 820          * of the namekey synthesize the chain.
 821          *
 822          * The key range is inclusive of both key_beg and key_end.
 823          */
 824         namekey = hammer_directory_namekey(ncp->nc_name, nlen);
 825
 826         error = hammer_init_cursor(&trans, &cursor, &dip->cache[1], dip);
 827         cursor.key_beg.localization = dip->obj_localization +
 828                                       HAMMER_LOCALIZE_MISC;
 829         cursor.key_beg.obj_id = dip->obj_id;
 830         cursor.key_beg.key = namekey;
 831         cursor.key_beg.create_tid = 0;
 832         cursor.key_beg.delete_tid = 0;
 833         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
 834         cursor.key_beg.obj_type = 0;
 835
 836         cursor.key_end = cursor.key_beg;
 837         cursor.key_end.key |= 0xFFFFFFFFULL;
 838         cursor.asof = asof;
 839         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
 840
 841         /*
 842          * Scan all matching records (the chain), locate the one matching
 843          * the requested path component.
 844          *
 845          * The hammer_ip_*() functions merge in-memory records with on-disk
 846          * records for the purposes of the search.
 847          */
 848         obj_id = 0;
 849         localization = HAMMER_DEF_LOCALIZATION;
 850
 851         if (error == 0) {
 852                 error = hammer_ip_first(&cursor);
 853                 while (error == 0) {
 854                         error = hammer_ip_resolve_data(&cursor);
 855                         if (error)
 856                                 break;
 857                         if (nlen == cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF &&
 858                             bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
 859                                 obj_id = cursor.data->entry.obj_id;
 860                                 localization = cursor.data->entry.localization;
 861                                 break;
 862                         }
 863                         error = hammer_ip_next(&cursor);
 864                 }
 865         }
 866         hammer_done_cursor(&cursor);
 867         if (error == 0) {
 868                 ip = hammer_get_inode(&trans, dip, obj_id,
 869                                       asof, localization,
 870                                       flags, &error);
 871                 if (error == 0) {
 872                         error = hammer_get_vnode(ip, &vp);
 873                         hammer_rel_inode(ip, 0);
 874                 } else {
 875                         vp = NULL;
 876                 }
 877                 if (error == 0) {
 878                         vn_unlock(vp);
 879                         cache_setvp(ap->a_nch, vp);
 880                         vrele(vp);
 881                 }
 882         } else if (error == ENOENT) {
 883                 cache_setvp(ap->a_nch, NULL);
 884         }
 885 done:
 886         hammer_done_transaction(&trans);
 887         return (error);
 888 }
 889
 890 /*
 891  * hammer_vop_nlookupdotdot { dvp, vpp, cred }
 892  *
 893  * Locate the parent directory of a directory vnode.
 894  *
 895  * dvp is referenced but not locked.  *vpp must be returned referenced and
 896  * locked.  A parent_obj_id of 0 does not necessarily indicate that we are
 897  * at the root, instead it could indicate that the directory we were in was
 898  * removed.
 899  *
 900  * NOTE: as-of sequences are not linked into the directory structure.  If
 901  * we are at the root with a different asof then the mount point, reload
 902  * the same directory with the mount point's asof.   I'm not sure what this
 903  * will do to NFS.  We encode ASOF stamps in NFS file handles so it might not
 904  * get confused, but it hasn't been tested.
 905  */
 906 static
 907 int
 908 hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *ap)
 909 {
 910         struct hammer_transaction trans;
 911         struct hammer_inode *dip;
 912         struct hammer_inode *ip;
 913         int64_t parent_obj_id;
 914         u_int32_t parent_obj_localization;
 915         hammer_tid_t asof;
 916         int error;
 917
 918         dip = VTOI(ap->a_dvp);
 919         asof = dip->obj_asof;
 920
 921         /*
 922          * Whos are parent?  This could be the root of a pseudo-filesystem
 923          * whos parent is in another localization domain.
 924          */
 925         parent_obj_id = dip->ino_data.parent_obj_id;
 926         if (dip->obj_id == HAMMER_OBJID_ROOT)
 927                 parent_obj_localization = dip->ino_data.ext.obj.parent_obj_localization;
 928         else
 929                 parent_obj_localization = dip->obj_localization;
 930
 931         if (parent_obj_id == 0) {
 932                 if (dip->obj_id == HAMMER_OBJID_ROOT &&
 933                    asof != dip->hmp->asof) {
 934                         parent_obj_id = dip->obj_id;
 935                         asof = dip->hmp->asof;
 936                         *ap->a_fakename = kmalloc(19, M_TEMP, M_WAITOK);
 937                         ksnprintf(*ap->a_fakename, 19, "0x%016llx",
 938                                    dip->obj_asof);
 939                 } else {
 940                         *ap->a_vpp = NULL;
 941                         return ENOENT;
 942                 }
 943         }
 944
 945         hammer_simple_transaction(&trans, dip->hmp);
 946
 947         ip = hammer_get_inode(&trans, dip, parent_obj_id,
 948                               asof, parent_obj_localization,
 949                               dip->flags, &error);
 950         if (ip) {
 951                 error = hammer_get_vnode(ip, ap->a_vpp);
 952                 hammer_rel_inode(ip, 0);
 953         } else {
 954                 *ap->a_vpp = NULL;
 955         }
 956         hammer_done_transaction(&trans);
 957         return (error);
 958 }
 959
 960 /*
 961  * hammer_vop_nlink { nch, dvp, vp, cred }
 962  */
 963 static
 964 int
 965 hammer_vop_nlink(struct vop_nlink_args *ap)
 966 {
 967         struct hammer_transaction trans;
 968         struct hammer_inode *dip;
 969         struct hammer_inode *ip;
 970         struct nchandle *nch;
 971         int error;
 972
 973         nch = ap->a_nch;
 974         dip = VTOI(ap->a_dvp);
 975         ip = VTOI(ap->a_vp);
 976
 977         if (dip->flags & HAMMER_INODE_RO)
 978                 return (EROFS);
 979         if (ip->flags & HAMMER_INODE_RO)
 980                 return (EROFS);
 981         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
 982                 return (error);
 983
 984         /*
 985          * Create a transaction to cover the operations we perform.
 986          */
 987         hammer_start_transaction(&trans, dip->hmp);
 988
 989         /*
 990          * Add the filesystem object to the directory.  Note that neither
 991          * dip nor ip are referenced or locked, but their vnodes are
 992          * referenced.  This function will bump the inode's link count.
 993          */
 994         error = hammer_ip_add_directory(&trans, dip,
 995                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
 996                                         ip);
 997
 998         /*
 999          * Finish up.
1000          */
1001         if (error == 0) {
1002                 cache_setunresolved(nch);
1003                 cache_setvp(nch, ap->a_vp);
1004         }
1005         hammer_done_transaction(&trans);
1006         return (error);
1007 }
1008
1009 /*
1010  * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
1011  *
1012  * The operating system has already ensured that the directory entry
1013  * does not exist and done all appropriate namespace locking.
1014  */
1015 static
1016 int
1017 hammer_vop_nmkdir(struct vop_nmkdir_args *ap)
1018 {
1019         struct hammer_transaction trans;
1020         struct hammer_inode *dip;
1021         struct hammer_inode *nip;
1022         struct nchandle *nch;
1023         int error;
1024
1025         nch = ap->a_nch;
1026         dip = VTOI(ap->a_dvp);
1027
1028         if (dip->flags & HAMMER_INODE_RO)
1029                 return (EROFS);
1030         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1031                 return (error);
1032
1033         /*
1034          * Create a transaction to cover the operations we perform.
1035          */
1036         hammer_start_transaction(&trans, dip->hmp);
1037
1038         /*
1039          * Create a new filesystem object of the requested type.  The
1040          * returned inode will be referenced but not locked.
1041          */
1042         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1043                                     dip, NULL, &nip);
1044         if (error) {
1045                 hkprintf("hammer_mkdir error %d\n", error);
1046                 hammer_done_transaction(&trans);
1047                 *ap->a_vpp = NULL;
1048                 return (error);
1049         }
1050         /*
1051          * Add the new filesystem object to the directory.  This will also
1052          * bump the inode's link count.
1053          */
1054         error = hammer_ip_add_directory(&trans, dip,
1055                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1056                                         nip);
1057         if (error)
1058                 hkprintf("hammer_mkdir (add) error %d\n", error);
1059
1060         /*
1061          * Finish up.
1062          */
1063         if (error) {
1064                 hammer_rel_inode(nip, 0);
1065                 *ap->a_vpp = NULL;
1066         } else {
1067                 error = hammer_get_vnode(nip, ap->a_vpp);
1068                 hammer_rel_inode(nip, 0);
1069                 if (error == 0) {
1070                         cache_setunresolved(ap->a_nch);
1071                         cache_setvp(ap->a_nch, *ap->a_vpp);
1072                 }
1073         }
1074         hammer_done_transaction(&trans);
1075         return (error);
1076 }
1077
1078 /*
1079  * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1080  *
1081  * The operating system has already ensured that the directory entry
1082  * does not exist and done all appropriate namespace locking.
1083  */
1084 static
1085 int
1086 hammer_vop_nmknod(struct vop_nmknod_args *ap)
1087 {
1088         struct hammer_transaction trans;
1089         struct hammer_inode *dip;
1090         struct hammer_inode *nip;
1091         struct nchandle *nch;
1092         int error;
1093
1094         nch = ap->a_nch;
1095         dip = VTOI(ap->a_dvp);
1096
1097         if (dip->flags & HAMMER_INODE_RO)
1098                 return (EROFS);
1099         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1100                 return (error);
1101
1102         /*
1103          * Create a transaction to cover the operations we perform.
1104          */
1105         hammer_start_transaction(&trans, dip->hmp);
1106
1107         /*
1108          * Create a new filesystem object of the requested type.  The
1109          * returned inode will be referenced but not locked.
1110          *
1111          * If mknod specifies a directory a pseudo-fs is created.
1112          */
1113         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1114                                     dip, NULL, &nip);
1115         if (error) {
1116                 hammer_done_transaction(&trans);
1117                 *ap->a_vpp = NULL;
1118                 return (error);
1119         }
1120
1121         /*
1122          * Add the new filesystem object to the directory.  This will also
1123          * bump the inode's link count.
1124          */
1125         error = hammer_ip_add_directory(&trans, dip,
1126                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1127                                         nip);
1128
1129         /*
1130          * Finish up.
1131          */
1132         if (error) {
1133                 hammer_rel_inode(nip, 0);
1134                 *ap->a_vpp = NULL;
1135         } else {
1136                 error = hammer_get_vnode(nip, ap->a_vpp);
1137                 hammer_rel_inode(nip, 0);
1138                 if (error == 0) {
1139                         cache_setunresolved(ap->a_nch);
1140                         cache_setvp(ap->a_nch, *ap->a_vpp);
1141                 }
1142         }
1143         hammer_done_transaction(&trans);
1144         return (error);
1145 }
1146
1147 /*
1148  * hammer_vop_open { vp, mode, cred, fp }
1149  */
1150 static
1151 int
1152 hammer_vop_open(struct vop_open_args *ap)
1153 {
1154         hammer_inode_t ip;
1155
1156         ip = VTOI(ap->a_vp);
1157
1158         if ((ap->a_mode & FWRITE) && (ip->flags & HAMMER_INODE_RO))
1159                 return (EROFS);
1160         return(vop_stdopen(ap));
1161 }
1162
1163 /*
1164  * hammer_vop_pathconf { vp, name, retval }
1165  */
1166 static
1167 int
1168 hammer_vop_pathconf(struct vop_pathconf_args *ap)
1169 {
1170         return EOPNOTSUPP;
1171 }
1172
1173 /*
1174  * hammer_vop_print { vp }
1175  */
1176 static
1177 int
1178 hammer_vop_print(struct vop_print_args *ap)
1179 {
1180         return EOPNOTSUPP;
1181 }
1182
1183 /*
1184  * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1185  */
1186 static
1187 int
1188 hammer_vop_readdir(struct vop_readdir_args *ap)
1189 {
1190         struct hammer_transaction trans;
1191         struct hammer_cursor cursor;
1192         struct hammer_inode *ip;
1193         struct uio *uio;
1194         hammer_base_elm_t base;
1195         int error;
1196         int cookie_index;
1197         int ncookies;
1198         off_t *cookies;
1199         off_t saveoff;
1200         int r;
1201         int dtype;
1202
1203         ip = VTOI(ap->a_vp);
1204         uio = ap->a_uio;
1205         saveoff = uio->uio_offset;
1206
1207         if (ap->a_ncookies) {
1208                 ncookies = uio->uio_resid / 16 + 1;
1209                 if (ncookies > 1024)
1210                         ncookies = 1024;
1211                 cookies = kmalloc(ncookies * sizeof(off_t), M_TEMP, M_WAITOK);
1212                 cookie_index = 0;
1213         } else {
1214                 ncookies = -1;
1215                 cookies = NULL;
1216                 cookie_index = 0;
1217         }
1218
1219         hammer_simple_transaction(&trans, ip->hmp);
1220
1221         /*
1222          * Handle artificial entries
1223          */
1224         error = 0;
1225         if (saveoff == 0) {
1226                 r = vop_write_dirent(&error, uio, ip->obj_id, DT_DIR, 1, ".");
1227                 if (r)
1228                         goto done;
1229                 if (cookies)
1230                         cookies[cookie_index] = saveoff;
1231                 ++saveoff;
1232                 ++cookie_index;
1233                 if (cookie_index == ncookies)
1234                         goto done;
1235         }
1236         if (saveoff == 1) {
1237                 if (ip->ino_data.parent_obj_id) {
1238                         r = vop_write_dirent(&error, uio,
1239                                              ip->ino_data.parent_obj_id,
1240                                              DT_DIR, 2, "..");
1241                 } else {
1242                         r = vop_write_dirent(&error, uio,
1243                                              ip->obj_id, DT_DIR, 2, "..");
1244                 }
1245                 if (r)
1246                         goto done;
1247                 if (cookies)
1248                         cookies[cookie_index] = saveoff;
1249                 ++saveoff;
1250                 ++cookie_index;
1251                 if (cookie_index == ncookies)
1252                         goto done;
1253         }
1254
1255         /*
1256          * Key range (begin and end inclusive) to scan.  Directory keys
1257          * directly translate to a 64 bit 'seek' position.
1258          */
1259         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1260         cursor.key_beg.localization = ip->obj_localization +
1261                                       HAMMER_LOCALIZE_MISC;
1262         cursor.key_beg.obj_id = ip->obj_id;
1263         cursor.key_beg.create_tid = 0;
1264         cursor.key_beg.delete_tid = 0;
1265         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1266         cursor.key_beg.obj_type = 0;
1267         cursor.key_beg.key = saveoff;
1268
1269         cursor.key_end = cursor.key_beg;
1270         cursor.key_end.key = HAMMER_MAX_KEY;
1271         cursor.asof = ip->obj_asof;
1272         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1273
1274         error = hammer_ip_first(&cursor);
1275
1276         while (error == 0) {
1277                 error = hammer_ip_resolve_data(&cursor);
1278                 if (error)
1279                         break;
1280                 base = &cursor.leaf->base;
1281                 saveoff = base->key;
1282                 KKASSERT(cursor.leaf->data_len > HAMMER_ENTRY_NAME_OFF);
1283
1284                 if (base->obj_id != ip->obj_id)
1285                         panic("readdir: bad record at %p", cursor.node);
1286
1287                 /*
1288                  * Convert pseudo-filesystems into softlinks
1289                  */
1290                 dtype = hammer_get_dtype(cursor.leaf->base.obj_type);
1291                 r = vop_write_dirent(
1292                              &error, uio, cursor.data->entry.obj_id,
1293                              dtype,
1294                              cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF ,
1295                              (void *)cursor.data->entry.name);
1296                 if (r)
1297                         break;
1298                 ++saveoff;
1299                 if (cookies)
1300                         cookies[cookie_index] = base->key;
1301                 ++cookie_index;
1302                 if (cookie_index == ncookies)
1303                         break;
1304                 error = hammer_ip_next(&cursor);
1305         }
1306         hammer_done_cursor(&cursor);
1307
1308 done:
1309         hammer_done_transaction(&trans);
1310
1311         if (ap->a_eofflag)
1312                 *ap->a_eofflag = (error == ENOENT);
1313         uio->uio_offset = saveoff;
1314         if (error && cookie_index == 0) {
1315                 if (error == ENOENT)
1316                         error = 0;
1317                 if (cookies) {
1318                         kfree(cookies, M_TEMP);
1319                         *ap->a_ncookies = 0;
1320                         *ap->a_cookies = NULL;
1321                 }
1322         } else {
1323                 if (error == ENOENT)
1324                         error = 0;
1325                 if (cookies) {
1326                         *ap->a_ncookies = cookie_index;
1327                         *ap->a_cookies = cookies;
1328                 }
1329         }
1330         return(error);
1331 }
1332
1333 /*
1334  * hammer_vop_readlink { vp, uio, cred }
1335  */
1336 static
1337 int
1338 hammer_vop_readlink(struct vop_readlink_args *ap)
1339 {
1340         struct hammer_transaction trans;
1341         struct hammer_cursor cursor;
1342         struct hammer_inode *ip;
1343         char buf[32];
1344         u_int32_t localization;
1345         hammer_pseudofs_inmem_t pfsm;
1346         int error;
1347
1348         ip = VTOI(ap->a_vp);
1349
1350         /*
1351          * Shortcut if the symlink data was stuffed into ino_data.
1352          *
1353          * Also expand special "@@PFS%05d" softlinks (expansion only
1354          * occurs for non-historical (current) accesses made from the
1355          * primary filesystem).
1356          */
1357         if (ip->ino_data.size <= HAMMER_INODE_BASESYMLEN) {
1358                 char *ptr;
1359                 int bytes;
1360
1361                 ptr = ip->ino_data.ext.symlink;
1362                 bytes = (int)ip->ino_data.size;
1363                 if (bytes == 10 &&
1364                     ip->obj_asof == HAMMER_MAX_TID &&
1365                     ip->obj_localization == 0 &&
1366                     strncmp(ptr, "@@PFS", 5) == 0) {
1367                         hammer_simple_transaction(&trans, ip->hmp);
1368                         bcopy(ptr + 5, buf, 5);
1369                         buf[5] = 0;
1370                         localization = strtoul(buf, NULL, 10) << 16;
1371                         pfsm = hammer_load_pseudofs(&trans, localization,
1372                                                     &error);
1373                         if (error == 0) {
1374                                 if (pfsm->pfsd.mirror_flags &
1375                                     HAMMER_PFSD_SLAVE) {
1376                                         ksnprintf(buf, sizeof(buf),
1377                                                   "@@0x%016llx:%05d",
1378                                                   pfsm->pfsd.sync_end_tid,
1379                                                   localization >> 16);
1380                                 } else {
1381                                         ksnprintf(buf, sizeof(buf),
1382                                                   "@@0x%016llx:%05d",
1383                                                   HAMMER_MAX_TID,
1384                                                   localization >> 16);
1385                                 }
1386                                 ptr = buf;
1387                                 bytes = strlen(buf);
1388                         }
1389                         if (pfsm)
1390                                 hammer_rel_pseudofs(trans.hmp, pfsm);
1391                         hammer_done_transaction(&trans);
1392                 }
1393                 error = uiomove(ptr, bytes, ap->a_uio);
1394                 return(error);
1395         }
1396
1397         /*
1398          * Long version
1399          */
1400         hammer_simple_transaction(&trans, ip->hmp);
1401         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1402
1403         /*
1404          * Key range (begin and end inclusive) to scan.  Directory keys
1405          * directly translate to a 64 bit 'seek' position.
1406          */
1407         cursor.key_beg.localization = ip->obj_localization +
1408                                       HAMMER_LOCALIZE_MISC;
1409         cursor.key_beg.obj_id = ip->obj_id;
1410         cursor.key_beg.create_tid = 0;
1411         cursor.key_beg.delete_tid = 0;
1412         cursor.key_beg.rec_type = HAMMER_RECTYPE_FIX;
1413         cursor.key_beg.obj_type = 0;
1414         cursor.key_beg.key = HAMMER_FIXKEY_SYMLINK;
1415         cursor.asof = ip->obj_asof;
1416         cursor.flags |= HAMMER_CURSOR_ASOF;
1417
1418         error = hammer_ip_lookup(&cursor);
1419         if (error == 0) {
1420                 error = hammer_ip_resolve_data(&cursor);
1421                 if (error == 0) {
1422                         KKASSERT(cursor.leaf->data_len >=
1423                                  HAMMER_SYMLINK_NAME_OFF);
1424                         error = uiomove(cursor.data->symlink.name,
1425                                         cursor.leaf->data_len -
1426                                                 HAMMER_SYMLINK_NAME_OFF,
1427                                         ap->a_uio);
1428                 }
1429         }
1430         hammer_done_cursor(&cursor);
1431         hammer_done_transaction(&trans);
1432         return(error);
1433 }
1434
1435 /*
1436  * hammer_vop_nremove { nch, dvp, cred }
1437  */
1438 static
1439 int
1440 hammer_vop_nremove(struct vop_nremove_args *ap)
1441 {
1442         struct hammer_transaction trans;
1443         struct hammer_inode *dip;
1444         int error;
1445
1446         dip = VTOI(ap->a_dvp);
1447
1448         if (hammer_nohistory(dip) == 0 &&
1449             (error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1450                 return (error);
1451         }
1452
1453         hammer_start_transaction(&trans, dip->hmp);
1454         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0);
1455         hammer_done_transaction(&trans);
1456
1457         return (error);
1458 }
1459
1460 /*
1461  * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1462  */
1463 static
1464 int
1465 hammer_vop_nrename(struct vop_nrename_args *ap)
1466 {
1467         struct hammer_transaction trans;
1468         struct namecache *fncp;
1469         struct namecache *tncp;
1470         struct hammer_inode *fdip;
1471         struct hammer_inode *tdip;
1472         struct hammer_inode *ip;
1473         struct hammer_cursor cursor;
1474         int64_t namekey;
1475         int nlen, error;
1476
1477         fdip = VTOI(ap->a_fdvp);
1478         tdip = VTOI(ap->a_tdvp);
1479         fncp = ap->a_fnch->ncp;
1480         tncp = ap->a_tnch->ncp;
1481         ip = VTOI(fncp->nc_vp);
1482         KKASSERT(ip != NULL);
1483
1484         if (fdip->flags & HAMMER_INODE_RO)
1485                 return (EROFS);
1486         if (tdip->flags & HAMMER_INODE_RO)
1487                 return (EROFS);
1488         if (ip->flags & HAMMER_INODE_RO)
1489                 return (EROFS);
1490         if ((error = hammer_checkspace(fdip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1491                 return (error);
1492
1493         hammer_start_transaction(&trans, fdip->hmp);
1494
1495         /*
1496          * Remove tncp from the target directory and then link ip as
1497          * tncp. XXX pass trans to dounlink
1498          *
1499          * Force the inode sync-time to match the transaction so it is
1500          * in-sync with the creation of the target directory entry.
1501          */
1502         error = hammer_dounlink(&trans, ap->a_tnch, ap->a_tdvp, ap->a_cred, 0);
1503         if (error == 0 || error == ENOENT) {
1504                 error = hammer_ip_add_directory(&trans, tdip,
1505                                                 tncp->nc_name, tncp->nc_nlen,
1506                                                 ip);
1507                 if (error == 0) {
1508                         ip->ino_data.parent_obj_id = tdip->obj_id;
1509                         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
1510                 }
1511         }
1512         if (error)
1513                 goto failed; /* XXX */
1514
1515         /*
1516          * Locate the record in the originating directory and remove it.
1517          *
1518          * Calculate the namekey and setup the key range for the scan.  This
1519          * works kinda like a chained hash table where the lower 32 bits
1520          * of the namekey synthesize the chain.
1521          *
1522          * The key range is inclusive of both key_beg and key_end.
1523          */
1524         namekey = hammer_directory_namekey(fncp->nc_name, fncp->nc_nlen);
1525 retry:
1526         hammer_init_cursor(&trans, &cursor, &fdip->cache[1], fdip);
1527         cursor.key_beg.localization = fdip->obj_localization +
1528                                       HAMMER_LOCALIZE_MISC;
1529         cursor.key_beg.obj_id = fdip->obj_id;
1530         cursor.key_beg.key = namekey;
1531         cursor.key_beg.create_tid = 0;
1532         cursor.key_beg.delete_tid = 0;
1533         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1534         cursor.key_beg.obj_type = 0;
1535
1536         cursor.key_end = cursor.key_beg;
1537         cursor.key_end.key |= 0xFFFFFFFFULL;
1538         cursor.asof = fdip->obj_asof;
1539         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1540
1541         /*
1542          * Scan all matching records (the chain), locate the one matching
1543          * the requested path component.
1544          *
1545          * The hammer_ip_*() functions merge in-memory records with on-disk
1546          * records for the purposes of the search.
1547          */
1548         error = hammer_ip_first(&cursor);
1549         while (error == 0) {
1550                 if (hammer_ip_resolve_data(&cursor) != 0)
1551                         break;
1552                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
1553                 KKASSERT(nlen > 0);
1554                 if (fncp->nc_nlen == nlen &&
1555                     bcmp(fncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1556                         break;
1557                 }
1558                 error = hammer_ip_next(&cursor);
1559         }
1560
1561         /*
1562          * If all is ok we have to get the inode so we can adjust nlinks.
1563          *
1564          * WARNING: hammer_ip_del_directory() may have to terminate the
1565          * cursor to avoid a recursion.  It's ok to call hammer_done_cursor()
1566          * twice.
1567          */
1568         if (error == 0)
1569                 error = hammer_ip_del_directory(&trans, &cursor, fdip, ip);
1570
1571         /*
1572          * XXX A deadlock here will break rename's atomicy for the purposes
1573          * of crash recovery.
1574          */
1575         if (error == EDEADLK) {
1576                 hammer_done_cursor(&cursor);
1577                 goto retry;
1578         }
1579
1580         /*
1581          * Cleanup and tell the kernel that the rename succeeded.
1582          */
1583         hammer_done_cursor(&cursor);
1584         if (error == 0)
1585                 cache_rename(ap->a_fnch, ap->a_tnch);
1586
1587 failed:
1588         hammer_done_transaction(&trans);
1589         return (error);
1590 }
1591
1592 /*
1593  * hammer_vop_nrmdir { nch, dvp, cred }
1594  */
1595 static
1596 int
1597 hammer_vop_nrmdir(struct vop_nrmdir_args *ap)
1598 {
1599         struct hammer_transaction trans;
1600         struct hammer_inode *dip;
1601         int error;
1602
1603         dip = VTOI(ap->a_dvp);
1604
1605         if (hammer_nohistory(dip) == 0 &&
1606             (error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1607                 return (error);
1608         }
1609
1610         hammer_start_transaction(&trans, dip->hmp);
1611         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0);
1612         hammer_done_transaction(&trans);
1613
1614         return (error);
1615 }
1616
1617 /*
1618  * hammer_vop_setattr { vp, vap, cred }
1619  */
1620 static
1621 int
1622 hammer_vop_setattr(struct vop_setattr_args *ap)
1623 {
1624         struct hammer_transaction trans;
1625         struct vattr *vap;
1626         struct hammer_inode *ip;
1627         int modflags;
1628         int error;
1629         int truncating;
1630         int blksize;
1631         int64_t aligned_size;
1632         u_int32_t flags;
1633
1634         vap = ap->a_vap;
1635         ip = ap->a_vp->v_data;
1636         modflags = 0;
1637
1638         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
1639                 return(EROFS);
1640         if (ip->flags & HAMMER_INODE_RO)
1641                 return (EROFS);
1642         if (hammer_nohistory(ip) == 0 &&
1643             (error = hammer_checkspace(ip->hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1644                 return (error);
1645         }
1646
1647         hammer_start_transaction(&trans, ip->hmp);
1648         error = 0;
1649
1650         if (vap->va_flags != VNOVAL) {
1651                 flags = ip->ino_data.uflags;
1652                 error = vop_helper_setattr_flags(&flags, vap->va_flags,
1653                                          hammer_to_unix_xid(&ip->ino_data.uid),
1654                                          ap->a_cred);
1655                 if (error == 0) {
1656                         if (ip->ino_data.uflags != flags) {
1657                                 ip->ino_data.uflags = flags;
1658                                 modflags |= HAMMER_INODE_DDIRTY;
1659                         }
1660                         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
1661                                 error = 0;
1662                                 goto done;
1663                         }
1664                 }
1665                 goto done;
1666         }
1667         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
1668                 error = EPERM;
1669                 goto done;
1670         }
1671         if (vap->va_uid != (uid_t)VNOVAL || vap->va_gid != (gid_t)VNOVAL) {
1672                 mode_t cur_mode = ip->ino_data.mode;
1673                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1674                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1675                 uuid_t uuid_uid;
1676                 uuid_t uuid_gid;
1677
1678                 error = vop_helper_chown(ap->a_vp, vap->va_uid, vap->va_gid,
1679                                          ap->a_cred,
1680                                          &cur_uid, &cur_gid, &cur_mode);
1681                 if (error == 0) {
1682                         hammer_guid_to_uuid(&uuid_uid, cur_uid);
1683                         hammer_guid_to_uuid(&uuid_gid, cur_gid);
1684                         if (bcmp(&uuid_uid, &ip->ino_data.uid,
1685                                  sizeof(uuid_uid)) ||
1686                             bcmp(&uuid_gid, &ip->ino_data.gid,
1687                                  sizeof(uuid_gid)) ||
1688                             ip->ino_data.mode != cur_mode
1689                         ) {
1690                                 ip->ino_data.uid = uuid_uid;
1691                                 ip->ino_data.gid = uuid_gid;
1692                                 ip->ino_data.mode = cur_mode;
1693                         }
1694                         modflags |= HAMMER_INODE_DDIRTY;
1695                 }
1696         }
1697         while (vap->va_size != VNOVAL && ip->ino_data.size != vap->va_size) {
1698                 switch(ap->a_vp->v_type) {
1699                 case VREG:
1700                         if (vap->va_size == ip->ino_data.size)
1701                                 break;
1702                         /*
1703                          * XXX break atomicy, we can deadlock the backend
1704                          * if we do not release the lock.  Probably not a
1705                          * big deal here.
1706                          */
1707                         blksize = hammer_blocksize(vap->va_size);
1708                         if (vap->va_size < ip->ino_data.size) {
1709                                 vtruncbuf(ap->a_vp, vap->va_size, blksize);
1710                                 truncating = 1;
1711                         } else {
1712                                 vnode_pager_setsize(ap->a_vp, vap->va_size);
1713                                 truncating = 0;
1714                         }
1715                         ip->ino_data.size = vap->va_size;
1716                         modflags |= HAMMER_INODE_DDIRTY;
1717
1718                         /*
1719                          * on-media truncation is cached in the inode until
1720                          * the inode is synchronized.
1721                          */
1722                         if (truncating) {
1723                                 hammer_ip_frontend_trunc(ip, vap->va_size);
1724 #ifdef DEBUG_TRUNCATE
1725                                 if (HammerTruncIp == NULL)
1726                                         HammerTruncIp = ip;
1727 #endif
1728                                 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
1729                                         ip->flags |= HAMMER_INODE_TRUNCATED;
1730                                         ip->trunc_off = vap->va_size;
1731 #ifdef DEBUG_TRUNCATE
1732                                         if (ip == HammerTruncIp)
1733                                         kprintf("truncate1 %016llx\n", ip->trunc_off);
1734 #endif
1735                                 } else if (ip->trunc_off > vap->va_size) {
1736                                         ip->trunc_off = vap->va_size;
1737 #ifdef DEBUG_TRUNCATE
1738                                         if (ip == HammerTruncIp)
1739                                         kprintf("truncate2 %016llx\n", ip->trunc_off);
1740 #endif
1741                                 } else {
1742 #ifdef DEBUG_TRUNCATE
1743                                         if (ip == HammerTruncIp)
1744                                         kprintf("truncate3 %016llx (ignored)\n", vap->va_size);
1745 #endif
1746                                 }
1747                         }
1748
1749                         /*
1750                          * If truncating we have to clean out a portion of
1751                          * the last block on-disk.  We do this in the
1752                          * front-end buffer cache.
1753                          */
1754                         aligned_size = (vap->va_size + (blksize - 1)) &
1755                                        ~(int64_t)(blksize - 1);
1756                         if (truncating && vap->va_size < aligned_size) {
1757                                 struct buf *bp;
1758                                 int offset;
1759
1760                                 aligned_size -= blksize;
1761
1762                                 offset = (int)vap->va_size & (blksize - 1);
1763                                 error = bread(ap->a_vp, aligned_size,
1764                                               blksize, &bp);
1765                                 hammer_ip_frontend_trunc(ip, aligned_size);
1766                                 if (error == 0) {
1767                                         bzero(bp->b_data + offset,
1768                                               blksize - offset);
1769                                         bdwrite(bp);
1770                                 } else {
1771                                         kprintf("ERROR %d\n", error);
1772                                         brelse(bp);
1773                                 }
1774                         }
1775                         break;
1776                 case VDATABASE:
1777                         if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
1778                                 ip->flags |= HAMMER_INODE_TRUNCATED;
1779                                 ip->trunc_off = vap->va_size;
1780                         } else if (ip->trunc_off > vap->va_size) {
1781                                 ip->trunc_off = vap->va_size;
1782                         }
1783                         hammer_ip_frontend_trunc(ip, vap->va_size);
1784                         ip->ino_data.size = vap->va_size;
1785                         modflags |= HAMMER_INODE_DDIRTY;
1786                         break;
1787                 default:
1788                         error = EINVAL;
1789                         goto done;
1790                 }
1791                 break;
1792         }
1793         if (vap->va_atime.tv_sec != VNOVAL) {
1794                 ip->ino_data.atime =
1795                         hammer_timespec_to_time(&vap->va_atime);
1796                 modflags |= HAMMER_INODE_ATIME;
1797         }
1798         if (vap->va_mtime.tv_sec != VNOVAL) {
1799                 ip->ino_data.mtime =
1800                         hammer_timespec_to_time(&vap->va_mtime);
1801                 modflags |= HAMMER_INODE_MTIME;
1802         }
1803         if (vap->va_mode != (mode_t)VNOVAL) {
1804                 mode_t   cur_mode = ip->ino_data.mode;
1805                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1806                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1807
1808                 error = vop_helper_chmod(ap->a_vp, vap->va_mode, ap->a_cred,
1809                                          cur_uid, cur_gid, &cur_mode);
1810                 if (error == 0 && ip->ino_data.mode != cur_mode) {
1811                         ip->ino_data.mode = cur_mode;
1812                         modflags |= HAMMER_INODE_DDIRTY;
1813                 }
1814         }
1815 done:
1816         if (error == 0)
1817                 hammer_modify_inode(ip, modflags);
1818         hammer_done_transaction(&trans);
1819         return (error);
1820 }
1821
1822 /*
1823  * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
1824  */
1825 static
1826 int
1827 hammer_vop_nsymlink(struct vop_nsymlink_args *ap)
1828 {
1829         struct hammer_transaction trans;
1830         struct hammer_inode *dip;
1831         struct hammer_inode *nip;
1832         struct nchandle *nch;
1833         hammer_record_t record;
1834         int error;
1835         int bytes;
1836
1837         ap->a_vap->va_type = VLNK;
1838
1839         nch = ap->a_nch;
1840         dip = VTOI(ap->a_dvp);
1841
1842         if (dip->flags & HAMMER_INODE_RO)
1843                 return (EROFS);
1844         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1845                 return (error);
1846
1847         /*
1848          * Create a transaction to cover the operations we perform.
1849          */
1850         hammer_start_transaction(&trans, dip->hmp);
1851
1852         /*
1853          * Create a new filesystem object of the requested type.  The
1854          * returned inode will be referenced but not locked.
1855          */
1856
1857         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1858                                     dip, NULL, &nip);
1859         if (error) {
1860                 hammer_done_transaction(&trans);
1861                 *ap->a_vpp = NULL;
1862                 return (error);
1863         }
1864
1865         /*
1866          * Add a record representing the symlink.  symlink stores the link
1867          * as pure data, not a string, and is no \0 terminated.
1868          */
1869         if (error == 0) {
1870                 bytes = strlen(ap->a_target);
1871
1872                 if (bytes <= HAMMER_INODE_BASESYMLEN) {
1873                         bcopy(ap->a_target, nip->ino_data.ext.symlink, bytes);
1874                 } else {
1875                         record = hammer_alloc_mem_record(nip, bytes);
1876                         record->type = HAMMER_MEM_RECORD_GENERAL;
1877
1878                         record->leaf.base.localization = nip->obj_localization +
1879                                                          HAMMER_LOCALIZE_MISC;
1880                         record->leaf.base.key = HAMMER_FIXKEY_SYMLINK;
1881                         record->leaf.base.rec_type = HAMMER_RECTYPE_FIX;
1882                         record->leaf.data_len = bytes;
1883                         KKASSERT(HAMMER_SYMLINK_NAME_OFF == 0);
1884                         bcopy(ap->a_target, record->data->symlink.name, bytes);
1885                         error = hammer_ip_add_record(&trans, record);
1886                 }
1887
1888                 /*
1889                  * Set the file size to the length of the link.
1890                  */
1891                 if (error == 0) {
1892                         nip->ino_data.size = bytes;
1893                         hammer_modify_inode(nip, HAMMER_INODE_DDIRTY);
1894                 }
1895         }
1896         if (error == 0)
1897                 error = hammer_ip_add_directory(&trans, dip, nch->ncp->nc_name,
1898                                                 nch->ncp->nc_nlen, nip);
1899
1900         /*
1901          * Finish up.
1902          */
1903         if (error) {
1904                 hammer_rel_inode(nip, 0);
1905                 *ap->a_vpp = NULL;
1906         } else {
1907                 error = hammer_get_vnode(nip, ap->a_vpp);
1908                 hammer_rel_inode(nip, 0);
1909                 if (error == 0) {
1910                         cache_setunresolved(ap->a_nch);
1911                         cache_setvp(ap->a_nch, *ap->a_vpp);
1912                 }
1913         }
1914         hammer_done_transaction(&trans);
1915         return (error);
1916 }
1917
1918 /*
1919  * hammer_vop_nwhiteout { nch, dvp, cred, flags }
1920  */
1921 static
1922 int
1923 hammer_vop_nwhiteout(struct vop_nwhiteout_args *ap)
1924 {
1925         struct hammer_transaction trans;
1926         struct hammer_inode *dip;
1927         int error;
1928
1929         dip = VTOI(ap->a_dvp);
1930
1931         if (hammer_nohistory(dip) == 0 &&
1932             (error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0) {
1933                 return (error);
1934         }
1935
1936         hammer_start_transaction(&trans, dip->hmp);
1937         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp,
1938                                 ap->a_cred, ap->a_flags);
1939         hammer_done_transaction(&trans);
1940
1941         return (error);
1942 }
1943
1944 /*
1945  * hammer_vop_ioctl { vp, command, data, fflag, cred }
1946  */
1947 static
1948 int
1949 hammer_vop_ioctl(struct vop_ioctl_args *ap)
1950 {
1951         struct hammer_inode *ip = ap->a_vp->v_data;
1952
1953         return(hammer_ioctl(ip, ap->a_command, ap->a_data,
1954                             ap->a_fflag, ap->a_cred));
1955 }
1956
1957 static
1958 int
1959 hammer_vop_mountctl(struct vop_mountctl_args *ap)
1960 {
1961         struct mount *mp;
1962         int error;
1963
1964         mp = ap->a_head.a_ops->head.vv_mount;
1965
1966         switch(ap->a_op) {
1967         case MOUNTCTL_SET_EXPORT:
1968                 if (ap->a_ctllen != sizeof(struct export_args))
1969                         error = EINVAL;
1970                 error = hammer_vfs_export(mp, ap->a_op,
1971                                       (const struct export_args *)ap->a_ctl);
1972                 break;
1973         default:
1974                 error = journal_mountctl(ap);
1975                 break;
1976         }
1977         return(error);
1978 }
1979
1980 /*
1981  * hammer_vop_strategy { vp, bio }
1982  *
1983  * Strategy call, used for regular file read & write only.  Note that the
1984  * bp may represent a cluster.
1985  *
1986  * To simplify operation and allow better optimizations in the future,
1987  * this code does not make any assumptions with regards to buffer alignment
1988  * or size.
1989  */
1990 static
1991 int
1992 hammer_vop_strategy(struct vop_strategy_args *ap)
1993 {
1994         struct buf *bp;
1995         int error;
1996
1997         bp = ap->a_bio->bio_buf;
1998
1999         switch(bp->b_cmd) {
2000         case BUF_CMD_READ:
2001                 error = hammer_vop_strategy_read(ap);
2002                 break;
2003         case BUF_CMD_WRITE:
2004                 error = hammer_vop_strategy_write(ap);
2005                 break;
2006         default:
2007                 bp->b_error = error = EINVAL;
2008                 bp->b_flags |= B_ERROR;
2009                 biodone(ap->a_bio);
2010                 break;
2011         }
2012         return (error);
2013 }
2014
2015 /*
2016  * Read from a regular file.  Iterate the related records and fill in the
2017  * BIO/BUF.  Gaps are zero-filled.
2018  *
2019  * The support code in hammer_object.c should be used to deal with mixed
2020  * in-memory and on-disk records.
2021  *
2022  * NOTE: Can be called from the cluster code with an oversized buf.
2023  *
2024  * XXX atime update
2025  */
2026 static
2027 int
2028 hammer_vop_strategy_read(struct vop_strategy_args *ap)
2029 {
2030         struct hammer_transaction trans;
2031         struct hammer_inode *ip;
2032         struct hammer_cursor cursor;
2033         hammer_base_elm_t base;
2034         hammer_off_t disk_offset;
2035         struct bio *bio;
2036         struct bio *nbio;
2037         struct buf *bp;
2038         int64_t rec_offset;
2039         int64_t ran_end;
2040         int64_t tmp64;
2041         int error;
2042         int boff;
2043         int roff;
2044         int n;
2045
2046         bio = ap->a_bio;
2047         bp = bio->bio_buf;
2048         ip = ap->a_vp->v_data;
2049
2050         /*
2051          * The zone-2 disk offset may have been set by the cluster code via
2052          * a BMAP operation, or else should be NOOFFSET.
2053          *
2054          * Checking the high bits for a match against zone-2 should suffice.
2055          */
2056         nbio = push_bio(bio);
2057         if ((nbio->bio_offset & HAMMER_OFF_ZONE_MASK) ==
2058             HAMMER_ZONE_RAW_BUFFER) {
2059                 error = hammer_io_direct_read(ip->hmp, nbio);
2060                 return (error);
2061         }
2062
2063         /*
2064          * Well, that sucked.  Do it the hard way.  If all the stars are
2065          * aligned we may still be able to issue a direct-read.
2066          */
2067         hammer_simple_transaction(&trans, ip->hmp);
2068         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2069
2070         /*
2071          * Key range (begin and end inclusive) to scan.  Note that the key's
2072          * stored in the actual records represent BASE+LEN, not BASE.  The
2073          * first record containing bio_offset will have a key > bio_offset.
2074          */
2075         cursor.key_beg.localization = ip->obj_localization +
2076                                       HAMMER_LOCALIZE_MISC;
2077         cursor.key_beg.obj_id = ip->obj_id;
2078         cursor.key_beg.create_tid = 0;
2079         cursor.key_beg.delete_tid = 0;
2080         cursor.key_beg.obj_type = 0;
2081         cursor.key_beg.key = bio->bio_offset + 1;
2082         cursor.asof = ip->obj_asof;
2083         cursor.flags |= HAMMER_CURSOR_ASOF;
2084
2085         cursor.key_end = cursor.key_beg;
2086         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2087 #if 0
2088         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
2089                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
2090                 cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
2091                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2092         } else
2093 #endif
2094         {
2095                 ran_end = bio->bio_offset + bp->b_bufsize;
2096                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2097                 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2098                 tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2099                 if (tmp64 < ran_end)
2100                         cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2101                 else
2102                         cursor.key_end.key = ran_end + MAXPHYS + 1;
2103         }
2104         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2105
2106         error = hammer_ip_first(&cursor);
2107         boff = 0;
2108
2109         while (error == 0) {
2110                 /*
2111                  * Get the base file offset of the record.  The key for
2112                  * data records is (base + bytes) rather then (base).
2113                  */
2114                 base = &cursor.leaf->base;
2115                 rec_offset = base->key - cursor.leaf->data_len;
2116
2117                 /*
2118                  * Calculate the gap, if any, and zero-fill it.
2119                  *
2120                  * n is the offset of the start of the record verses our
2121                  * current seek offset in the bio.
2122                  */
2123                 n = (int)(rec_offset - (bio->bio_offset + boff));
2124                 if (n > 0) {
2125                         if (n > bp->b_bufsize - boff)
2126                                 n = bp->b_bufsize - boff;
2127                         bzero((char *)bp->b_data + boff, n);
2128                         boff += n;
2129                         n = 0;
2130                 }
2131
2132                 /*
2133                  * Calculate the data offset in the record and the number
2134                  * of bytes we can copy.
2135                  *
2136                  * There are two degenerate cases.  First, boff may already
2137                  * be at bp->b_bufsize.  Secondly, the data offset within
2138                  * the record may exceed the record's size.
2139                  */
2140                 roff = -n;
2141                 rec_offset += roff;
2142                 n = cursor.leaf->data_len - roff;
2143                 if (n <= 0) {
2144                         kprintf("strategy_read: bad n=%d roff=%d\n", n, roff);
2145                         n = 0;
2146                 } else if (n > bp->b_bufsize - boff) {
2147                         n = bp->b_bufsize - boff;
2148                 }
2149
2150                 /*
2151                  * Deal with cached truncations.  This cool bit of code
2152                  * allows truncate()/ftruncate() to avoid having to sync
2153                  * the file.
2154                  *
2155                  * If the frontend is truncated then all backend records are
2156                  * subject to the frontend's truncation.
2157                  *
2158                  * If the backend is truncated then backend records on-disk
2159                  * (but not in-memory) are subject to the backend's
2160                  * truncation.  In-memory records owned by the backend
2161                  * represent data written after the truncation point on the
2162                  * backend and must not be truncated.
2163                  *
2164                  * Truncate operations deal with frontend buffer cache
2165                  * buffers and frontend-owned in-memory records synchronously.
2166                  */
2167                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2168                         if (hammer_cursor_ondisk(&cursor) ||
2169                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
2170                                 if (ip->trunc_off <= rec_offset)
2171                                         n = 0;
2172                                 else if (ip->trunc_off < rec_offset + n)
2173                                         n = (int)(ip->trunc_off - rec_offset);
2174                         }
2175                 }
2176                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2177                         if (hammer_cursor_ondisk(&cursor)) {
2178                                 if (ip->sync_trunc_off <= rec_offset)
2179                                         n = 0;
2180                                 else if (ip->sync_trunc_off < rec_offset + n)
2181                                         n = (int)(ip->sync_trunc_off - rec_offset);
2182                         }
2183                 }
2184
2185                 /*
2186                  * Try to issue a direct read into our bio if possible,
2187                  * otherwise resolve the element data into a hammer_buffer
2188                  * and copy.
2189                  *
2190                  * The buffer on-disk should be zerod past any real
2191                  * truncation point, but may not be for any synthesized
2192                  * truncation point from above.
2193                  */
2194                 if (boff == 0 && n == bp->b_bufsize &&
2195                     ((cursor.leaf->data_offset + roff) & HAMMER_BUFMASK) == 0) {
2196                         disk_offset = hammer_blockmap_lookup(
2197                                                 trans.hmp,
2198                                                 cursor.leaf->data_offset + roff,
2199                                                 &error);
2200                         if (error)
2201                                 break;
2202                         nbio->bio_offset = disk_offset;
2203                         error = hammer_io_direct_read(trans.hmp, nbio);
2204                         goto done;
2205                 } else if (n) {
2206                         error = hammer_ip_resolve_data(&cursor);
2207                         if (error == 0) {
2208                                 bcopy((char *)cursor.data + roff,
2209                                       (char *)bp->b_data + boff, n);
2210                         }
2211                 }
2212                 if (error)
2213                         break;
2214
2215                 /*
2216                  * Iterate until we have filled the request.
2217                  */
2218                 boff += n;
2219                 if (boff == bp->b_bufsize)
2220                         break;
2221                 error = hammer_ip_next(&cursor);
2222         }
2223
2224         /*
2225          * There may have been a gap after the last record
2226          */
2227         if (error == ENOENT)
2228                 error = 0;
2229         if (error == 0 && boff != bp->b_bufsize) {
2230                 KKASSERT(boff < bp->b_bufsize);
2231                 bzero((char *)bp->b_data + boff, bp->b_bufsize - boff);
2232                 /* boff = bp->b_bufsize; */
2233         }
2234         bp->b_resid = 0;
2235         bp->b_error = error;
2236         if (error)
2237                 bp->b_flags |= B_ERROR;
2238         biodone(ap->a_bio);
2239
2240 done:
2241         if (cursor.node)
2242                 hammer_cache_node(&ip->cache[1], cursor.node);
2243         hammer_done_cursor(&cursor);
2244         hammer_done_transaction(&trans);
2245         return(error);
2246 }
2247
2248 /*
2249  * BMAP operation - used to support cluster_read() only.
2250  *
2251  * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
2252  *
2253  * This routine may return EOPNOTSUPP if the opration is not supported for
2254  * the specified offset.  The contents of the pointer arguments do not
2255  * need to be initialized in that case.
2256  *
2257  * If a disk address is available and properly aligned return 0 with
2258  * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
2259  * to the run-length relative to that offset.  Callers may assume that
2260  * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
2261  * large, so return EOPNOTSUPP if it is not sufficiently large.
2262  */
2263 static
2264 int
2265 hammer_vop_bmap(struct vop_bmap_args *ap)
2266 {
2267         struct hammer_transaction trans;
2268         struct hammer_inode *ip;
2269         struct hammer_cursor cursor;
2270         hammer_base_elm_t base;
2271         int64_t rec_offset;
2272         int64_t ran_end;
2273         int64_t tmp64;
2274         int64_t base_offset;
2275         int64_t base_disk_offset;
2276         int64_t last_offset;
2277         hammer_off_t last_disk_offset;
2278         hammer_off_t disk_offset;
2279         int     rec_len;
2280         int     error;
2281         int     blksize;
2282
2283         ip = ap->a_vp->v_data;
2284
2285         /*
2286          * We can only BMAP regular files.  We can't BMAP database files,
2287          * directories, etc.
2288          */
2289         if (ip->ino_data.obj_type != HAMMER_OBJTYPE_REGFILE)
2290                 return(EOPNOTSUPP);
2291
2292         /*
2293          * bmap is typically called with runp/runb both NULL when used
2294          * for writing.  We do not support BMAP for writing atm.
2295          */
2296         if (ap->a_cmd != BUF_CMD_READ)
2297                 return(EOPNOTSUPP);
2298
2299         /*
2300          * Scan the B-Tree to acquire blockmap addresses, then translate
2301          * to raw addresses.
2302          */
2303         hammer_simple_transaction(&trans, ip->hmp);
2304 #if 0
2305         kprintf("bmap_beg %016llx ip->cache %p\n", ap->a_loffset, ip->cache[1]);
2306 #endif
2307         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2308
2309         /*
2310          * Key range (begin and end inclusive) to scan.  Note that the key's
2311          * stored in the actual records represent BASE+LEN, not BASE.  The
2312          * first record containing bio_offset will have a key > bio_offset.
2313          */
2314         cursor.key_beg.localization = ip->obj_localization +
2315                                       HAMMER_LOCALIZE_MISC;
2316         cursor.key_beg.obj_id = ip->obj_id;
2317         cursor.key_beg.create_tid = 0;
2318         cursor.key_beg.delete_tid = 0;
2319         cursor.key_beg.obj_type = 0;
2320         if (ap->a_runb)
2321                 cursor.key_beg.key = ap->a_loffset - MAXPHYS + 1;
2322         else
2323                 cursor.key_beg.key = ap->a_loffset + 1;
2324         if (cursor.key_beg.key < 0)
2325                 cursor.key_beg.key = 0;
2326         cursor.asof = ip->obj_asof;
2327         cursor.flags |= HAMMER_CURSOR_ASOF;
2328
2329         cursor.key_end = cursor.key_beg;
2330         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2331
2332         ran_end = ap->a_loffset + MAXPHYS;
2333         cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2334         cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2335         tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2336         if (tmp64 < ran_end)
2337                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2338         else
2339                 cursor.key_end.key = ran_end + MAXPHYS + 1;
2340
2341         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2342
2343         error = hammer_ip_first(&cursor);
2344         base_offset = last_offset = 0;
2345         base_disk_offset = last_disk_offset = 0;
2346
2347         while (error == 0) {
2348                 /*
2349                  * Get the base file offset of the record.  The key for
2350                  * data records is (base + bytes) rather then (base).
2351                  *
2352                  * NOTE: rec_offset + rec_len may exceed the end-of-file.
2353                  * The extra bytes should be zero on-disk and the BMAP op
2354                  * should still be ok.
2355                  */
2356                 base = &cursor.leaf->base;
2357                 rec_offset = base->key - cursor.leaf->data_len;
2358                 rec_len    = cursor.leaf->data_len;
2359
2360                 /*
2361                  * Incorporate any cached truncation.
2362                  *
2363                  * NOTE: Modifications to rec_len based on synthesized
2364                  * truncation points remove the guarantee that any extended
2365                  * data on disk is zero (since the truncations may not have
2366                  * taken place on-media yet).
2367                  */
2368                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2369                         if (hammer_cursor_ondisk(&cursor) ||
2370                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
2371                                 if (ip->trunc_off <= rec_offset)
2372                                         rec_len = 0;
2373                                 else if (ip->trunc_off < rec_offset + rec_len)
2374                                         rec_len = (int)(ip->trunc_off - rec_offset);
2375                         }
2376                 }
2377                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2378                         if (hammer_cursor_ondisk(&cursor)) {
2379                                 if (ip->sync_trunc_off <= rec_offset)
2380                                         rec_len = 0;
2381                                 else if (ip->sync_trunc_off < rec_offset + rec_len)
2382                                         rec_len = (int)(ip->sync_trunc_off - rec_offset);
2383                         }
2384                 }
2385
2386                 /*
2387                  * Accumulate information.  If we have hit a discontiguous
2388                  * block reset base_offset unless we are already beyond the
2389                  * requested offset.  If we are, that's it, we stop.
2390                  */
2391                 disk_offset = hammer_blockmap_lookup(trans.hmp,
2392                                                      cursor.leaf->data_offset,
2393                                                      &error);
2394                 if (error)
2395                         break;
2396                 if (rec_offset != last_offset ||
2397                     disk_offset != last_disk_offset) {
2398                         if (rec_offset > ap->a_loffset)
2399                                 break;
2400                         base_offset = rec_offset;
2401                         base_disk_offset = disk_offset;
2402                 }
2403                 last_offset = rec_offset + rec_len;
2404                 last_disk_offset = disk_offset + rec_len;
2405
2406                 error = hammer_ip_next(&cursor);
2407         }
2408
2409 #if 0
2410         kprintf("BMAP %016llx:  %016llx - %016llx\n",
2411                 ap->a_loffset, base_offset, last_offset);
2412         kprintf("BMAP %16s:  %016llx - %016llx\n",
2413                 "", base_disk_offset, last_disk_offset);
2414 #endif
2415
2416         if (cursor.node) {
2417                 hammer_cache_node(&ip->cache[1], cursor.node);
2418 #if 0
2419                 kprintf("bmap_end2 %016llx ip->cache %p\n", ap->a_loffset, ip->cache[1]);
2420 #endif
2421         }
2422         hammer_done_cursor(&cursor);
2423         hammer_done_transaction(&trans);
2424
2425         /*
2426          * If we couldn't find any records or the records we did find were
2427          * all behind the requested offset, return failure.  A forward
2428          * truncation can leave a hole w/ no on-disk records.
2429          */
2430         if (last_offset == 0 || last_offset < ap->a_loffset)
2431                 return (EOPNOTSUPP);
2432
2433         /*
2434          * Figure out the block size at the requested offset and adjust
2435          * our limits so the cluster_read() does not create inappropriately
2436          * sized buffer cache buffers.
2437          */
2438         blksize = hammer_blocksize(ap->a_loffset);
2439         if (hammer_blocksize(base_offset) != blksize) {
2440                 base_offset = hammer_blockdemarc(base_offset, ap->a_loffset);
2441         }
2442         if (last_offset != ap->a_loffset &&
2443             hammer_blocksize(last_offset - 1) != blksize) {
2444                 last_offset = hammer_blockdemarc(ap->a_loffset,
2445                                                  last_offset - 1);
2446         }
2447
2448         /*
2449          * Returning EOPNOTSUPP simply prevents the direct-IO optimization
2450          * from occuring.
2451          */
2452         disk_offset = base_disk_offset + (ap->a_loffset - base_offset);
2453
2454         /*
2455          * If doffsetp is not aligned or the forward run size does
2456          * not cover a whole buffer, disallow the direct I/O.
2457          */
2458         if ((disk_offset & HAMMER_BUFMASK) ||
2459             (last_offset - ap->a_loffset) < blksize) {
2460                 error = EOPNOTSUPP;
2461         } else {
2462                 *ap->a_doffsetp = disk_offset;
2463                 if (ap->a_runb) {
2464                         *ap->a_runb = ap->a_loffset - base_offset;
2465                         KKASSERT(*ap->a_runb >= 0);
2466                 }
2467                 if (ap->a_runp) {
2468                         *ap->a_runp = last_offset - ap->a_loffset;
2469                         KKASSERT(*ap->a_runp >= 0);
2470                 }
2471                 error = 0;
2472         }
2473         return(error);
2474 }
2475
2476 /*
2477  * Write to a regular file.   Because this is a strategy call the OS is
2478  * trying to actually get data onto the media.
2479  */
2480 static
2481 int
2482 hammer_vop_strategy_write(struct vop_strategy_args *ap)
2483 {
2484         hammer_record_t record;
2485         hammer_mount_t hmp;
2486         hammer_inode_t ip;
2487         struct bio *bio;
2488         struct buf *bp;
2489         int blksize;
2490         int bytes;
2491         int error;
2492
2493         bio = ap->a_bio;
2494         bp = bio->bio_buf;
2495         ip = ap->a_vp->v_data;
2496         hmp = ip->hmp;
2497
2498         blksize = hammer_blocksize(bio->bio_offset);
2499         KKASSERT(bp->b_bufsize == blksize);
2500
2501         if (ip->flags & HAMMER_INODE_RO) {
2502                 bp->b_error = EROFS;
2503                 bp->b_flags |= B_ERROR;
2504                 biodone(ap->a_bio);
2505                 return(EROFS);
2506         }
2507
2508         /*
2509          * Interlock with inode destruction (no in-kernel or directory
2510          * topology visibility).  If we queue new IO while trying to
2511          * destroy the inode we can deadlock the vtrunc call in
2512          * hammer_inode_unloadable_check().
2513          */
2514         if (ip->flags & (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) {
2515                 bp->b_resid = 0;
2516                 biodone(ap->a_bio);
2517                 return(0);
2518         }
2519
2520         /*
2521          * Reserve space and issue a direct-write from the front-end.
2522          * NOTE: The direct_io code will hammer_bread/bcopy smaller
2523          * allocations.
2524          *
2525          * An in-memory record will be installed to reference the storage
2526          * until the flusher can get to it.
2527          *
2528          * Since we own the high level bio the front-end will not try to
2529          * do a direct-read until the write completes.
2530          *
2531          * NOTE: The only time we do not reserve a full-sized buffers
2532          * worth of data is if the file is small.  We do not try to
2533          * allocate a fragment (from the small-data zone) at the end of
2534          * an otherwise large file as this can lead to wildly separated
2535          * data.
2536          */
2537         KKASSERT((bio->bio_offset & HAMMER_BUFMASK) == 0);
2538         KKASSERT(bio->bio_offset < ip->ino_data.size);
2539         if (bio->bio_offset || ip->ino_data.size > HAMMER_BUFSIZE / 2)
2540                 bytes = bp->b_bufsize;
2541         else
2542                 bytes = ((int)ip->ino_data.size + 15) & ~15;
2543
2544         record = hammer_ip_add_bulk(ip, bio->bio_offset, bp->b_data,
2545                                     bytes, &error);
2546         if (record) {
2547                 hammer_io_direct_write(hmp, &record->leaf, bio);
2548                 hammer_rel_mem_record(record);
2549                 if (ip->rsv_recs > 1 && hmp->rsv_recs > hammer_limit_recs)
2550                         hammer_flush_inode(ip, 0);
2551         } else {
2552                 bp->b_bio2.bio_offset = NOOFFSET;
2553                 bp->b_error = error;
2554                 bp->b_flags |= B_ERROR;
2555                 biodone(ap->a_bio);
2556         }
2557         return(error);
2558 }
2559
2560 /*
2561  * dounlink - disconnect a directory entry
2562  *
2563  * XXX whiteout support not really in yet
2564  */
2565 static int
2566 hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
2567                 struct vnode *dvp, struct ucred *cred, int flags)
2568 {
2569         struct namecache *ncp;
2570         hammer_inode_t dip;
2571         hammer_inode_t ip;
2572         struct hammer_cursor cursor;
2573         int64_t namekey;
2574         int nlen, error;
2575
2576         /*
2577          * Calculate the namekey and setup the key range for the scan.  This
2578          * works kinda like a chained hash table where the lower 32 bits
2579          * of the namekey synthesize the chain.
2580          *
2581          * The key range is inclusive of both key_beg and key_end.
2582          */
2583         dip = VTOI(dvp);
2584         ncp = nch->ncp;
2585
2586         if (dip->flags & HAMMER_INODE_RO)
2587                 return (EROFS);
2588
2589         namekey = hammer_directory_namekey(ncp->nc_name, ncp->nc_nlen);
2590 retry:
2591         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
2592         cursor.key_beg.localization = dip->obj_localization +
2593                                       HAMMER_LOCALIZE_MISC;
2594         cursor.key_beg.obj_id = dip->obj_id;
2595         cursor.key_beg.key = namekey;
2596         cursor.key_beg.create_tid = 0;
2597         cursor.key_beg.delete_tid = 0;
2598         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
2599         cursor.key_beg.obj_type = 0;
2600
2601         cursor.key_end = cursor.key_beg;
2602         cursor.key_end.key |= 0xFFFFFFFFULL;
2603         cursor.asof = dip->obj_asof;
2604         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2605
2606         /*
2607          * Scan all matching records (the chain), locate the one matching
2608          * the requested path component.  info->last_error contains the
2609          * error code on search termination and could be 0, ENOENT, or
2610          * something else.
2611          *
2612          * The hammer_ip_*() functions merge in-memory records with on-disk
2613          * records for the purposes of the search.
2614          */
2615         error = hammer_ip_first(&cursor);
2616
2617         while (error == 0) {
2618                 error = hammer_ip_resolve_data(&cursor);
2619                 if (error)
2620                         break;
2621                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
2622                 KKASSERT(nlen > 0);
2623                 if (ncp->nc_nlen == nlen &&
2624                     bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
2625                         break;
2626                 }
2627                 error = hammer_ip_next(&cursor);
2628         }
2629
2630         /*
2631          * If all is ok we have to get the inode so we can adjust nlinks.
2632          * To avoid a deadlock with the flusher we must release the inode
2633          * lock on the directory when acquiring the inode for the entry.
2634          *
2635          * If the target is a directory, it must be empty.
2636          */
2637         if (error == 0) {
2638                 hammer_unlock(&cursor.ip->lock);
2639                 ip = hammer_get_inode(trans, dip, cursor.data->entry.obj_id,
2640                                       dip->hmp->asof,
2641                                       cursor.data->entry.localization,
2642                                       0, &error);
2643                 hammer_lock_sh(&cursor.ip->lock);
2644                 if (error == ENOENT) {
2645                         kprintf("obj_id %016llx\n", cursor.data->entry.obj_id);
2646                         Debugger("ENOENT unlinking object that should exist");
2647                 }
2648
2649                 /*
2650                  * If we are trying to remove a directory the directory must
2651                  * be empty.
2652                  *
2653                  * WARNING: hammer_ip_check_directory_empty() may have to
2654                  * terminate the cursor to avoid a deadlock.  It is ok to
2655                  * call hammer_done_cursor() twice.
2656                  */
2657                 if (error == 0 && ip->ino_data.obj_type ==
2658                                   HAMMER_OBJTYPE_DIRECTORY) {
2659                         error = hammer_ip_check_directory_empty(trans, ip);
2660                 }
2661
2662                 /*
2663                  * Delete the directory entry.
2664                  *
2665                  * WARNING: hammer_ip_del_directory() may have to terminate
2666                  * the cursor to avoid a deadlock.  It is ok to call
2667                  * hammer_done_cursor() twice.
2668                  */
2669                 if (error == 0) {
2670                         error = hammer_ip_del_directory(trans, &cursor,
2671                                                         dip, ip);
2672                 }
2673                 hammer_done_cursor(&cursor);
2674                 if (error == 0) {
2675                         cache_setunresolved(nch);
2676                         cache_setvp(nch, NULL);
2677                         /* XXX locking */
2678                         if (ip->vp)
2679                                 cache_inval_vp(ip->vp, CINV_DESTROY);
2680                 }
2681                 if (ip)
2682                         hammer_rel_inode(ip, 0);
2683         } else {
2684                 hammer_done_cursor(&cursor);
2685         }
2686         if (error == EDEADLK)
2687                 goto retry;
2688
2689         return (error);
2690 }
2691
2692 /************************************************************************
2693  *                          FIFO AND SPECFS OPS                         *
2694  ************************************************************************
2695  *
2696  */
2697
2698 static int
2699 hammer_vop_fifoclose (struct vop_close_args *ap)
2700 {
2701         /* XXX update itimes */
2702         return (VOCALL(&fifo_vnode_vops, &ap->a_head));
2703 }
2704
2705 static int
2706 hammer_vop_fiforead (struct vop_read_args *ap)
2707 {
2708         int error;
2709
2710         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
2711         /* XXX update access time */
2712         return (error);
2713 }
2714
2715 static int
2716 hammer_vop_fifowrite (struct vop_write_args *ap)
2717 {
2718         int error;
2719
2720         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
2721         /* XXX update access time */
2722         return (error);
2723 }
2724
2725 static int
2726 hammer_vop_specclose (struct vop_close_args *ap)
2727 {
2728         /* XXX update itimes */
2729         return (VOCALL(&spec_vnode_vops, &ap->a_head));
2730 }
2731
2732 static int
2733 hammer_vop_specread (struct vop_read_args *ap)
2734 {
2735         /* XXX update access time */
2736         return (VOCALL(&spec_vnode_vops, &ap->a_head));
2737 }
2738
2739 static int
2740 hammer_vop_specwrite (struct vop_write_args *ap)
2741 {
2742         /* XXX update last change time */
2743         return (VOCALL(&spec_vnode_vops, &ap->a_head));
2744 }
2745