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 1.79 2008/07/01 02:08:58 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 static void hammer_cleanup_write_io(hammer_inode_t ip);
 163 static void hammer_update_rsv_databufs(hammer_inode_t ip);
 164
 165 #if 0
 166 static
 167 int
 168 hammer_vop_vnoperate(struct vop_generic_args *)
 169 {
 170         return (VOCALL(&hammer_vnode_vops, ap));
 171 }
 172 #endif
 173
 174 /*
 175  * hammer_vop_fsync { vp, waitfor }
 176  *
 177  * fsync() an inode to disk and wait for it to be completely committed
 178  * such that the information would not be undone if a crash occured after
 179  * return.
 180  */
 181 static
 182 int
 183 hammer_vop_fsync(struct vop_fsync_args *ap)
 184 {
 185         hammer_inode_t ip = VTOI(ap->a_vp);
 186
 187         vfsync(ap->a_vp, ap->a_waitfor, 1, NULL, NULL);
 188         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 189         if (ap->a_waitfor == MNT_WAIT)
 190                 hammer_wait_inode(ip);
 191         return (ip->error);
 192 }
 193
 194 /*
 195  * hammer_vop_read { vp, uio, ioflag, cred }
 196  */
 197 static
 198 int
 199 hammer_vop_read(struct vop_read_args *ap)
 200 {
 201         struct hammer_transaction trans;
 202         hammer_inode_t ip;
 203         off_t offset;
 204         struct buf *bp;
 205         struct uio *uio;
 206         int error;
 207         int n;
 208         int seqcount;
 209         int ioseqcount;
 210         int blksize;
 211
 212         if (ap->a_vp->v_type != VREG)
 213                 return (EINVAL);
 214         ip = VTOI(ap->a_vp);
 215         error = 0;
 216         uio = ap->a_uio;
 217
 218         /*
 219          * Allow the UIO's size to override the sequential heuristic.
 220          */
 221         blksize = hammer_blocksize(uio->uio_offset);
 222         seqcount = (uio->uio_resid + (blksize - 1)) / blksize;
 223         ioseqcount = ap->a_ioflag >> 16;
 224         if (seqcount < ioseqcount)
 225                 seqcount = ioseqcount;
 226
 227         hammer_start_transaction(&trans, ip->hmp);
 228
 229         /*
 230          * Access the data typically in HAMMER_BUFSIZE blocks via the
 231          * buffer cache, but HAMMER may use a variable block size based
 232          * on the offset.
 233          */
 234         while (uio->uio_resid > 0 && uio->uio_offset < ip->ino_data.size) {
 235                 int64_t base_offset;
 236                 int64_t file_limit;
 237
 238                 blksize = hammer_blocksize(uio->uio_offset);
 239                 offset = (int)uio->uio_offset & (blksize - 1);
 240                 base_offset = uio->uio_offset - offset;
 241
 242                 if (hammer_debug_cluster_enable) {
 243                         /*
 244                          * Use file_limit to prevent cluster_read() from
 245                          * creating buffers of the wrong block size past
 246                          * the demarc.
 247                          */
 248                         file_limit = ip->ino_data.size;
 249                         if (base_offset < HAMMER_XDEMARC &&
 250                             file_limit > HAMMER_XDEMARC) {
 251                                 file_limit = HAMMER_XDEMARC;
 252                         }
 253                         error = cluster_read(ap->a_vp,
 254                                              file_limit, base_offset,
 255                                              blksize, MAXPHYS,
 256                                              seqcount, &bp);
 257                 } else {
 258                         error = bread(ap->a_vp, base_offset, blksize, &bp);
 259                 }
 260                 if (error) {
 261                         kprintf("error %d\n", error);
 262                         brelse(bp);
 263                         break;
 264                 }
 265
 266                 /* bp->b_flags |= B_CLUSTEROK; temporarily disabled */
 267                 n = blksize - offset;
 268                 if (n > uio->uio_resid)
 269                         n = uio->uio_resid;
 270                 if (n > ip->ino_data.size - uio->uio_offset)
 271                         n = (int)(ip->ino_data.size - uio->uio_offset);
 272                 error = uiomove((char *)bp->b_data + offset, n, uio);
 273
 274                 /* data has a lower priority then meta-data */
 275                 bp->b_flags |= B_AGE;
 276                 bqrelse(bp);
 277                 if (error)
 278                         break;
 279         }
 280         if ((ip->flags & HAMMER_INODE_RO) == 0 &&
 281             (ip->hmp->mp->mnt_flag & MNT_NOATIME) == 0) {
 282                 ip->ino_data.atime = trans.time;
 283                 hammer_modify_inode(ip, HAMMER_INODE_ATIME);
 284         }
 285         hammer_done_transaction(&trans);
 286         return (error);
 287 }
 288
 289 /*
 290  * hammer_vop_write { vp, uio, ioflag, cred }
 291  */
 292 static
 293 int
 294 hammer_vop_write(struct vop_write_args *ap)
 295 {
 296         struct hammer_transaction trans;
 297         struct hammer_inode *ip;
 298         hammer_mount_t hmp;
 299         struct uio *uio;
 300         int offset;
 301         off_t base_offset;
 302         struct buf *bp;
 303         int error;
 304         int n;
 305         int flags;
 306         int delta;
 307         int seqcount;
 308
 309         if (ap->a_vp->v_type != VREG)
 310                 return (EINVAL);
 311         ip = VTOI(ap->a_vp);
 312         hmp = ip->hmp;
 313         error = 0;
 314         seqcount = ap->a_ioflag >> 16;
 315
 316         if (ip->flags & HAMMER_INODE_RO)
 317                 return (EROFS);
 318
 319         /*
 320          * Create a transaction to cover the operations we perform.
 321          */
 322         hammer_start_transaction(&trans, hmp);
 323         uio = ap->a_uio;
 324
 325         /*
 326          * Check append mode
 327          */
 328         if (ap->a_ioflag & IO_APPEND)
 329                 uio->uio_offset = ip->ino_data.size;
 330
 331         /*
 332          * Check for illegal write offsets.  Valid range is 0...2^63-1.
 333          *
 334          * NOTE: the base_off assignment is required to work around what
 335          * I consider to be a GCC-4 optimization bug.
 336          */
 337         if (uio->uio_offset < 0) {
 338                 hammer_done_transaction(&trans);
 339                 return (EFBIG);
 340         }
 341         base_offset = uio->uio_offset + uio->uio_resid; /* work around gcc-4 */
 342         if (uio->uio_resid > 0 && base_offset <= 0) {
 343                 hammer_done_transaction(&trans);
 344                 return (EFBIG);
 345         }
 346
 347         /*
 348          * Access the data typically in HAMMER_BUFSIZE blocks via the
 349          * buffer cache, but HAMMER may use a variable block size based
 350          * on the offset.
 351          */
 352         while (uio->uio_resid > 0) {
 353                 int fixsize = 0;
 354                 int blksize;
 355                 int blkmask;
 356
 357                 if ((error = hammer_checkspace(hmp)) != 0)
 358                         break;
 359
 360                 blksize = hammer_blocksize(uio->uio_offset);
 361
 362                 /*
 363                  * Do not allow HAMMER to blow out the buffer cache.  Very
 364                  * large UIOs can lockout other processes due to bwillwrite()
 365                  * mechanics.
 366                  *
 367                  * Do not allow HAMMER to blow out system memory by
 368                  * accumulating too many records.   Records are so well
 369                  * decoupled from the buffer cache that it is possible
 370                  * for userland to push data out to the media via
 371                  * direct-write, but build up the records queued to the
 372                  * backend faster then the backend can flush them out.
 373                  * HAMMER has hit its write limit but the frontend has
 374                  * no pushback to slow it down.
 375                  *
 376                  * The hammer inode is not locked during these operations.
 377                  * The vnode is locked which can interfere with the pageout
 378                  * daemon for non-UIO_NOCOPY writes but should not interfere
 379                  * with the buffer cache.  Even so, we cannot afford to
 380                  * allow the pageout daemon to build up too many dirty buffer
 381                  * cache buffers.
 382                  */
 383                 bwillwrite(blksize);
 384
 385                 /*
 386                  * Pending record flush check.
 387                  */
 388                 if (hmp->rsv_recs > hammer_limit_recs / 2) {
 389                         /*
 390                          * Get the inode on the flush list
 391                          */
 392                         if (ip->rsv_recs >= 64)
 393                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 394                         else if (ip->rsv_recs >= 16)
 395                                 hammer_flush_inode(ip, 0);
 396
 397                         /*
 398                          * Keep the flusher going if the system keeps
 399                          * queueing records.
 400                          */
 401                         delta = hmp->count_newrecords -
 402                                 hmp->last_newrecords;
 403                         if (delta < 0 || delta > hammer_limit_recs / 2) {
 404                                 hmp->last_newrecords = hmp->count_newrecords;
 405                                 hammer_sync_hmp(hmp, MNT_NOWAIT);
 406                         }
 407
 408                         /*
 409                          * If we have gotten behind start slowing
 410                          * down the writers.
 411                          */
 412                         delta = (hmp->rsv_recs - hammer_limit_recs) *
 413                                 hz / hammer_limit_recs;
 414                         if (delta > 0)
 415                                 tsleep(&trans, 0, "hmrslo", delta);
 416                 }
 417
 418                 /*
 419                  * Calculate the blocksize at the current offset and figure
 420                  * out how much we can actually write.
 421                  */
 422                 blkmask = blksize - 1;
 423                 offset = (int)uio->uio_offset & blkmask;
 424                 base_offset = uio->uio_offset & ~(int64_t)blkmask;
 425                 n = blksize - offset;
 426                 if (n > uio->uio_resid)
 427                         n = uio->uio_resid;
 428                 if (uio->uio_offset + n > ip->ino_data.size) {
 429                         vnode_pager_setsize(ap->a_vp, uio->uio_offset + n);
 430                         fixsize = 1;
 431                 }
 432
 433                 if (uio->uio_segflg == UIO_NOCOPY) {
 434                         /*
 435                          * Issuing a write with the same data backing the
 436                          * buffer.  Instantiate the buffer to collect the
 437                          * backing vm pages, then read-in any missing bits.
 438                          *
 439                          * This case is used by vop_stdputpages().
 440                          */
 441                         bp = getblk(ap->a_vp, base_offset,
 442                                     blksize, GETBLK_BHEAVY, 0);
 443                         if ((bp->b_flags & B_CACHE) == 0) {
 444                                 bqrelse(bp);
 445                                 error = bread(ap->a_vp, base_offset,
 446                                               blksize, &bp);
 447                         }
 448                 } else if (offset == 0 && uio->uio_resid >= blksize) {
 449                         /*
 450                          * Even though we are entirely overwriting the buffer
 451                          * we may still have to zero it out to avoid a
 452                          * mmap/write visibility issue.
 453                          */
 454                         bp = getblk(ap->a_vp, base_offset, blksize, GETBLK_BHEAVY, 0);
 455                         if ((bp->b_flags & B_CACHE) == 0)
 456                                 vfs_bio_clrbuf(bp);
 457                 } else if (base_offset >= ip->ino_data.size) {
 458                         /*
 459                          * If the base offset of the buffer is beyond the
 460                          * file EOF, we don't have to issue a read.
 461                          */
 462                         bp = getblk(ap->a_vp, base_offset,
 463                                     blksize, GETBLK_BHEAVY, 0);
 464                         vfs_bio_clrbuf(bp);
 465                 } else {
 466                         /*
 467                          * Partial overwrite, read in any missing bits then
 468                          * replace the portion being written.
 469                          */
 470                         error = bread(ap->a_vp, base_offset, blksize, &bp);
 471                         if (error == 0)
 472                                 bheavy(bp);
 473                 }
 474                 if (error == 0) {
 475                         error = uiomove((char *)bp->b_data + offset,
 476                                         n, uio);
 477                 }
 478
 479                 /*
 480                  * If we screwed up we have to undo any VM size changes we
 481                  * made.
 482                  */
 483                 if (error) {
 484                         brelse(bp);
 485                         if (fixsize) {
 486                                 vtruncbuf(ap->a_vp, ip->ino_data.size,
 487                                           hammer_blocksize(ip->ino_data.size));
 488                         }
 489                         break;
 490                 }
 491                 /* bp->b_flags |= B_CLUSTEROK; temporarily disabled */
 492                 if (ip->ino_data.size < uio->uio_offset) {
 493                         ip->ino_data.size = uio->uio_offset;
 494                         flags = HAMMER_INODE_DDIRTY;
 495                         vnode_pager_setsize(ap->a_vp, ip->ino_data.size);
 496                 } else {
 497                         flags = 0;
 498                 }
 499                 ip->ino_data.mtime = trans.time;
 500                 flags |= HAMMER_INODE_MTIME | HAMMER_INODE_BUFS;
 501                 hammer_modify_inode(ip, flags);
 502
 503                 /*
 504                  * Try to keep track of cached dirty data.
 505                  */
 506                 if ((bp->b_flags & B_DIRTY) == 0) {
 507                         ++ip->rsv_databufs;
 508                         ++hmp->rsv_databufs;
 509                 }
 510
 511                 /*
 512                  * Final buffer disposition.
 513                  *
 514                  */
 515                 bp->b_flags |= B_AGE;
 516                 if (ap->a_ioflag & IO_SYNC) {
 517                         bwrite(bp);
 518                 } else if (ap->a_ioflag & IO_DIRECT) {
 519                         bawrite(bp);
 520                 } else {
 521                         bdwrite(bp);
 522                 }
 523         }
 524         hammer_done_transaction(&trans);
 525         return (error);
 526 }
 527
 528 /*
 529  * hammer_vop_access { vp, mode, cred }
 530  */
 531 static
 532 int
 533 hammer_vop_access(struct vop_access_args *ap)
 534 {
 535         struct hammer_inode *ip = VTOI(ap->a_vp);
 536         uid_t uid;
 537         gid_t gid;
 538         int error;
 539
 540         uid = hammer_to_unix_xid(&ip->ino_data.uid);
 541         gid = hammer_to_unix_xid(&ip->ino_data.gid);
 542
 543         error = vop_helper_access(ap, uid, gid, ip->ino_data.mode,
 544                                   ip->ino_data.uflags);
 545         return (error);
 546 }
 547
 548 /*
 549  * hammer_vop_advlock { vp, id, op, fl, flags }
 550  */
 551 static
 552 int
 553 hammer_vop_advlock(struct vop_advlock_args *ap)
 554 {
 555         hammer_inode_t ip = VTOI(ap->a_vp);
 556
 557         return (lf_advlock(ap, &ip->advlock, ip->ino_data.size));
 558 }
 559
 560 /*
 561  * hammer_vop_close { vp, fflag }
 562  */
 563 static
 564 int
 565 hammer_vop_close(struct vop_close_args *ap)
 566 {
 567         hammer_inode_t ip = VTOI(ap->a_vp);
 568
 569         if ((ip->flags | ip->sync_flags) & HAMMER_INODE_MODMASK)
 570                 hammer_inode_waitreclaims(ip->hmp);
 571         return (vop_stdclose(ap));
 572 }
 573
 574 /*
 575  * hammer_vop_ncreate { nch, dvp, vpp, cred, vap }
 576  *
 577  * The operating system has already ensured that the directory entry
 578  * does not exist and done all appropriate namespace locking.
 579  */
 580 static
 581 int
 582 hammer_vop_ncreate(struct vop_ncreate_args *ap)
 583 {
 584         struct hammer_transaction trans;
 585         struct hammer_inode *dip;
 586         struct hammer_inode *nip;
 587         struct nchandle *nch;
 588         int error;
 589
 590         nch = ap->a_nch;
 591         dip = VTOI(ap->a_dvp);
 592
 593         if (dip->flags & HAMMER_INODE_RO)
 594                 return (EROFS);
 595         if ((error = hammer_checkspace(dip->hmp)) != 0)
 596                 return (error);
 597
 598         /*
 599          * Create a transaction to cover the operations we perform.
 600          */
 601         hammer_start_transaction(&trans, dip->hmp);
 602
 603         /*
 604          * Create a new filesystem object of the requested type.  The
 605          * returned inode will be referenced and shared-locked to prevent
 606          * it from being moved to the flusher.
 607          */
 608
 609         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
 610                                     dip, 0, &nip);
 611         if (error) {
 612                 hkprintf("hammer_create_inode error %d\n", error);
 613                 hammer_done_transaction(&trans);
 614                 *ap->a_vpp = NULL;
 615                 return (error);
 616         }
 617
 618         /*
 619          * Add the new filesystem object to the directory.  This will also
 620          * bump the inode's link count.
 621          */
 622         error = hammer_ip_add_directory(&trans, dip,
 623                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
 624                                         nip);
 625         if (error)
 626                 hkprintf("hammer_ip_add_directory error %d\n", error);
 627
 628         /*
 629          * Finish up.
 630          */
 631         if (error) {
 632                 hammer_rel_inode(nip, 0);
 633                 hammer_done_transaction(&trans);
 634                 *ap->a_vpp = NULL;
 635         } else {
 636                 error = hammer_get_vnode(nip, ap->a_vpp);
 637                 hammer_done_transaction(&trans);
 638                 hammer_rel_inode(nip, 0);
 639                 if (error == 0) {
 640                         cache_setunresolved(ap->a_nch);
 641                         cache_setvp(ap->a_nch, *ap->a_vpp);
 642                 }
 643         }
 644         return (error);
 645 }
 646
 647 /*
 648  * hammer_vop_getattr { vp, vap }
 649  *
 650  * Retrieve an inode's attribute information.  When accessing inodes
 651  * historically we fake the atime field to ensure consistent results.
 652  * The atime field is stored in the B-Tree element and allowed to be
 653  * updated without cycling the element.
 654  */
 655 static
 656 int
 657 hammer_vop_getattr(struct vop_getattr_args *ap)
 658 {
 659         struct hammer_inode *ip = VTOI(ap->a_vp);
 660         struct vattr *vap = ap->a_vap;
 661
 662         vap->va_fsid = ip->hmp->fsid_udev;
 663         /*
 664          * XXX munge the device if we are in a pseudo-fs, so user utilities
 665          * do not think its the same 'filesystem'.
 666          */
 667         if (ip->obj_localization)
 668                 vap->va_fsid += ip->obj_localization;
 669         vap->va_fileid = ip->ino_leaf.base.obj_id;
 670         vap->va_mode = ip->ino_data.mode;
 671         vap->va_nlink = ip->ino_data.nlinks;
 672         vap->va_uid = hammer_to_unix_xid(&ip->ino_data.uid);
 673         vap->va_gid = hammer_to_unix_xid(&ip->ino_data.gid);
 674         vap->va_rmajor = 0;
 675         vap->va_rminor = 0;
 676         vap->va_size = ip->ino_data.size;
 677
 678         /*
 679          * We must provide a consistent atime and mtime for snapshots
 680          * so people can do a 'tar cf - ... | md5' on them and get
 681          * consistent results.
 682          */
 683         if (ip->flags & HAMMER_INODE_RO) {
 684                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_atime);
 685                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_mtime);
 686         } else {
 687                 hammer_time_to_timespec(ip->ino_data.atime, &vap->va_atime);
 688                 hammer_time_to_timespec(ip->ino_data.mtime, &vap->va_mtime);
 689         }
 690         hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_ctime);
 691         vap->va_flags = ip->ino_data.uflags;
 692         vap->va_gen = 1;        /* hammer inums are unique for all time */
 693         vap->va_blocksize = HAMMER_BUFSIZE;
 694         if (ip->ino_data.size >= HAMMER_XDEMARC) {
 695                 vap->va_bytes = (ip->ino_data.size + HAMMER_XBUFMASK64) &
 696                                 ~HAMMER_XBUFMASK64;
 697         } else if (ip->ino_data.size > HAMMER_BUFSIZE / 2) {
 698                 vap->va_bytes = (ip->ino_data.size + HAMMER_BUFMASK64) &
 699                                 ~HAMMER_BUFMASK64;
 700         } else {
 701                 vap->va_bytes = (ip->ino_data.size + 15) & ~15;
 702         }
 703         vap->va_type = hammer_get_vnode_type(ip->ino_data.obj_type);
 704         vap->va_filerev = 0;    /* XXX */
 705         /* mtime uniquely identifies any adjustments made to the file XXX */
 706         vap->va_fsmid = ip->ino_data.mtime;
 707         vap->va_uid_uuid = ip->ino_data.uid;
 708         vap->va_gid_uuid = ip->ino_data.gid;
 709         vap->va_fsid_uuid = ip->hmp->fsid;
 710         vap->va_vaflags = VA_UID_UUID_VALID | VA_GID_UUID_VALID |
 711                           VA_FSID_UUID_VALID;
 712
 713         switch (ip->ino_data.obj_type) {
 714         case HAMMER_OBJTYPE_CDEV:
 715         case HAMMER_OBJTYPE_BDEV:
 716                 vap->va_rmajor = ip->ino_data.rmajor;
 717                 vap->va_rminor = ip->ino_data.rminor;
 718                 break;
 719         default:
 720                 break;
 721         }
 722
 723         return(0);
 724 }
 725
 726 /*
 727  * hammer_vop_nresolve { nch, dvp, cred }
 728  *
 729  * Locate the requested directory entry.
 730  */
 731 static
 732 int
 733 hammer_vop_nresolve(struct vop_nresolve_args *ap)
 734 {
 735         struct hammer_transaction trans;
 736         struct namecache *ncp;
 737         hammer_inode_t dip;
 738         hammer_inode_t ip;
 739         hammer_tid_t asof;
 740         struct hammer_cursor cursor;
 741         struct vnode *vp;
 742         int64_t namekey;
 743         int error;
 744         int i;
 745         int nlen;
 746         int flags;
 747         u_int64_t obj_id;
 748         u_int32_t localization;
 749
 750         /*
 751          * Misc initialization, plus handle as-of name extensions.  Look for
 752          * the '@@' extension.  Note that as-of files and directories cannot
 753          * be modified.
 754          */
 755         dip = VTOI(ap->a_dvp);
 756         ncp = ap->a_nch->ncp;
 757         asof = dip->obj_asof;
 758         nlen = ncp->nc_nlen;
 759         flags = dip->flags;
 760
 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                         flags |= HAMMER_INODE_RO;
 767                         break;
 768                 }
 769         }
 770         nlen = i;
 771
 772         /*
 773          * If there is no path component the time extension is relative to
 774          * dip.
 775          */
 776         if (nlen == 0) {
 777                 ip = hammer_get_inode(&trans, dip, dip->obj_id,
 778                                       asof, dip->obj_localization,
 779                                       flags, &error);
 780                 if (error == 0) {
 781                         error = hammer_get_vnode(ip, &vp);
 782                         hammer_rel_inode(ip, 0);
 783                 } else {
 784                         vp = NULL;
 785                 }
 786                 if (error == 0) {
 787                         vn_unlock(vp);
 788                         cache_setvp(ap->a_nch, vp);
 789                         vrele(vp);
 790                 }
 791                 goto done;
 792         }
 793
 794         /*
 795          * Calculate the namekey and setup the key range for the scan.  This
 796          * works kinda like a chained hash table where the lower 32 bits
 797          * of the namekey synthesize the chain.
 798          *
 799          * The key range is inclusive of both key_beg and key_end.
 800          */
 801         namekey = hammer_directory_namekey(ncp->nc_name, nlen);
 802
 803         error = hammer_init_cursor(&trans, &cursor, &dip->cache[1], dip);
 804         cursor.key_beg.localization = dip->obj_localization +
 805                                       HAMMER_LOCALIZE_MISC;
 806         cursor.key_beg.obj_id = dip->obj_id;
 807         cursor.key_beg.key = namekey;
 808         cursor.key_beg.create_tid = 0;
 809         cursor.key_beg.delete_tid = 0;
 810         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
 811         cursor.key_beg.obj_type = 0;
 812
 813         cursor.key_end = cursor.key_beg;
 814         cursor.key_end.key |= 0xFFFFFFFFULL;
 815         cursor.asof = asof;
 816         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
 817
 818         /*
 819          * Scan all matching records (the chain), locate the one matching
 820          * the requested path component.
 821          *
 822          * The hammer_ip_*() functions merge in-memory records with on-disk
 823          * records for the purposes of the search.
 824          */
 825         obj_id = 0;
 826         localization = HAMMER_DEF_LOCALIZATION;
 827
 828         if (error == 0) {
 829                 error = hammer_ip_first(&cursor);
 830                 while (error == 0) {
 831                         error = hammer_ip_resolve_data(&cursor);
 832                         if (error)
 833                                 break;
 834                         if (nlen == cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF &&
 835                             bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
 836                                 obj_id = cursor.data->entry.obj_id;
 837                                 localization = cursor.data->entry.localization;
 838                                 break;
 839                         }
 840                         error = hammer_ip_next(&cursor);
 841                 }
 842         }
 843         hammer_done_cursor(&cursor);
 844         if (error == 0) {
 845                 ip = hammer_get_inode(&trans, dip, obj_id,
 846                                       asof, localization,
 847                                       flags, &error);
 848                 if (error == 0) {
 849                         error = hammer_get_vnode(ip, &vp);
 850                         hammer_rel_inode(ip, 0);
 851                 } else {
 852                         vp = NULL;
 853                 }
 854                 if (error == 0) {
 855                         vn_unlock(vp);
 856                         cache_setvp(ap->a_nch, vp);
 857                         vrele(vp);
 858                 }
 859         } else if (error == ENOENT) {
 860                 cache_setvp(ap->a_nch, NULL);
 861         }
 862 done:
 863         hammer_done_transaction(&trans);
 864         return (error);
 865 }
 866
 867 /*
 868  * hammer_vop_nlookupdotdot { dvp, vpp, cred }
 869  *
 870  * Locate the parent directory of a directory vnode.
 871  *
 872  * dvp is referenced but not locked.  *vpp must be returned referenced and
 873  * locked.  A parent_obj_id of 0 does not necessarily indicate that we are
 874  * at the root, instead it could indicate that the directory we were in was
 875  * removed.
 876  *
 877  * NOTE: as-of sequences are not linked into the directory structure.  If
 878  * we are at the root with a different asof then the mount point, reload
 879  * the same directory with the mount point's asof.   I'm not sure what this
 880  * will do to NFS.  We encode ASOF stamps in NFS file handles so it might not
 881  * get confused, but it hasn't been tested.
 882  */
 883 static
 884 int
 885 hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *ap)
 886 {
 887         struct hammer_transaction trans;
 888         struct hammer_inode *dip;
 889         struct hammer_inode *ip;
 890         int64_t parent_obj_id;
 891         u_int32_t parent_obj_localization;
 892         hammer_tid_t asof;
 893         int error;
 894
 895         dip = VTOI(ap->a_dvp);
 896         asof = dip->obj_asof;
 897
 898         /*
 899          * Whos are parent?  This could be the root of a pseudo-filesystem
 900          * whos parent is in another localization domain.
 901          */
 902         parent_obj_id = dip->ino_data.parent_obj_id;
 903         if (dip->obj_id == HAMMER_OBJID_ROOT)
 904                 parent_obj_localization = dip->ino_data.ext.obj.parent_obj_localization;
 905         else
 906                 parent_obj_localization = dip->obj_localization;
 907
 908         if (parent_obj_id == 0) {
 909                 if (dip->obj_id == HAMMER_OBJID_ROOT &&
 910                    asof != dip->hmp->asof) {
 911                         parent_obj_id = dip->obj_id;
 912                         asof = dip->hmp->asof;
 913                         *ap->a_fakename = kmalloc(19, M_TEMP, M_WAITOK);
 914                         ksnprintf(*ap->a_fakename, 19, "0x%016llx",
 915                                    dip->obj_asof);
 916                 } else {
 917                         *ap->a_vpp = NULL;
 918                         return ENOENT;
 919                 }
 920         }
 921
 922         hammer_simple_transaction(&trans, dip->hmp);
 923
 924         ip = hammer_get_inode(&trans, dip, parent_obj_id,
 925                               asof, parent_obj_localization,
 926                               dip->flags, &error);
 927         if (ip) {
 928                 error = hammer_get_vnode(ip, ap->a_vpp);
 929                 hammer_rel_inode(ip, 0);
 930         } else {
 931                 *ap->a_vpp = NULL;
 932         }
 933         hammer_done_transaction(&trans);
 934         return (error);
 935 }
 936
 937 /*
 938  * hammer_vop_nlink { nch, dvp, vp, cred }
 939  */
 940 static
 941 int
 942 hammer_vop_nlink(struct vop_nlink_args *ap)
 943 {
 944         struct hammer_transaction trans;
 945         struct hammer_inode *dip;
 946         struct hammer_inode *ip;
 947         struct nchandle *nch;
 948         int error;
 949
 950         nch = ap->a_nch;
 951         dip = VTOI(ap->a_dvp);
 952         ip = VTOI(ap->a_vp);
 953
 954         if (dip->flags & HAMMER_INODE_RO)
 955                 return (EROFS);
 956         if (ip->flags & HAMMER_INODE_RO)
 957                 return (EROFS);
 958         if ((error = hammer_checkspace(dip->hmp)) != 0)
 959                 return (error);
 960
 961         /*
 962          * Create a transaction to cover the operations we perform.
 963          */
 964         hammer_start_transaction(&trans, dip->hmp);
 965
 966         /*
 967          * Add the filesystem object to the directory.  Note that neither
 968          * dip nor ip are referenced or locked, but their vnodes are
 969          * referenced.  This function will bump the inode's link count.
 970          */
 971         error = hammer_ip_add_directory(&trans, dip,
 972                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
 973                                         ip);
 974
 975         /*
 976          * Finish up.
 977          */
 978         if (error == 0) {
 979                 cache_setunresolved(nch);
 980                 cache_setvp(nch, ap->a_vp);
 981         }
 982         hammer_done_transaction(&trans);
 983         return (error);
 984 }
 985
 986 /*
 987  * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
 988  *
 989  * The operating system has already ensured that the directory entry
 990  * does not exist and done all appropriate namespace locking.
 991  */
 992 static
 993 int
 994 hammer_vop_nmkdir(struct vop_nmkdir_args *ap)
 995 {
 996         struct hammer_transaction trans;
 997         struct hammer_inode *dip;
 998         struct hammer_inode *nip;
 999         struct nchandle *nch;
1000         int error;
1001
1002         nch = ap->a_nch;
1003         dip = VTOI(ap->a_dvp);
1004
1005         if (dip->flags & HAMMER_INODE_RO)
1006                 return (EROFS);
1007         if ((error = hammer_checkspace(dip->hmp)) != 0)
1008                 return (error);
1009
1010         /*
1011          * Create a transaction to cover the operations we perform.
1012          */
1013         hammer_start_transaction(&trans, dip->hmp);
1014
1015         /*
1016          * Create a new filesystem object of the requested type.  The
1017          * returned inode will be referenced but not locked.
1018          */
1019         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1020                                     dip, 0, &nip);
1021         if (error) {
1022                 hkprintf("hammer_mkdir error %d\n", error);
1023                 hammer_done_transaction(&trans);
1024                 *ap->a_vpp = NULL;
1025                 return (error);
1026         }
1027         /*
1028          * Add the new filesystem object to the directory.  This will also
1029          * bump the inode's link count.
1030          */
1031         error = hammer_ip_add_directory(&trans, dip,
1032                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1033                                         nip);
1034         if (error)
1035                 hkprintf("hammer_mkdir (add) error %d\n", error);
1036
1037         /*
1038          * Finish up.
1039          */
1040         if (error) {
1041                 hammer_rel_inode(nip, 0);
1042                 *ap->a_vpp = NULL;
1043         } else {
1044                 error = hammer_get_vnode(nip, ap->a_vpp);
1045                 hammer_rel_inode(nip, 0);
1046                 if (error == 0) {
1047                         cache_setunresolved(ap->a_nch);
1048                         cache_setvp(ap->a_nch, *ap->a_vpp);
1049                 }
1050         }
1051         hammer_done_transaction(&trans);
1052         return (error);
1053 }
1054
1055 /*
1056  * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1057  *
1058  * The operating system has already ensured that the directory entry
1059  * does not exist and done all appropriate namespace locking.
1060  */
1061 static
1062 int
1063 hammer_vop_nmknod(struct vop_nmknod_args *ap)
1064 {
1065         struct hammer_transaction trans;
1066         struct hammer_inode *dip;
1067         struct hammer_inode *nip;
1068         struct nchandle *nch;
1069         int error;
1070         int pseudofs;
1071
1072         nch = ap->a_nch;
1073         dip = VTOI(ap->a_dvp);
1074
1075         if (dip->flags & HAMMER_INODE_RO)
1076                 return (EROFS);
1077         if ((error = hammer_checkspace(dip->hmp)) != 0)
1078                 return (error);
1079
1080         /*
1081          * Create a transaction to cover the operations we perform.
1082          */
1083         hammer_start_transaction(&trans, dip->hmp);
1084
1085         /*
1086          * Create a new filesystem object of the requested type.  The
1087          * returned inode will be referenced but not locked.
1088          *
1089          * If mknod specifies a directory a pseudo-fs is created.
1090          */
1091         pseudofs = (ap->a_vap->va_type == VDIR);
1092         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1093                                     dip, pseudofs, &nip);
1094         if (error) {
1095                 hammer_done_transaction(&trans);
1096                 *ap->a_vpp = NULL;
1097                 return (error);
1098         }
1099
1100         /*
1101          * Add the new filesystem object to the directory.  This will also
1102          * bump the inode's link count.
1103          */
1104         error = hammer_ip_add_directory(&trans, dip,
1105                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1106                                         nip);
1107
1108         /*
1109          * Finish up.
1110          */
1111         if (error) {
1112                 hammer_rel_inode(nip, 0);
1113                 *ap->a_vpp = NULL;
1114         } else {
1115                 error = hammer_get_vnode(nip, ap->a_vpp);
1116                 hammer_rel_inode(nip, 0);
1117                 if (error == 0) {
1118                         cache_setunresolved(ap->a_nch);
1119                         cache_setvp(ap->a_nch, *ap->a_vpp);
1120                 }
1121         }
1122         hammer_done_transaction(&trans);
1123         return (error);
1124 }
1125
1126 /*
1127  * hammer_vop_open { vp, mode, cred, fp }
1128  */
1129 static
1130 int
1131 hammer_vop_open(struct vop_open_args *ap)
1132 {
1133         hammer_inode_t ip;
1134
1135         ip = VTOI(ap->a_vp);
1136
1137         if ((ap->a_mode & FWRITE) && (ip->flags & HAMMER_INODE_RO))
1138                 return (EROFS);
1139         return(vop_stdopen(ap));
1140 }
1141
1142 /*
1143  * hammer_vop_pathconf { vp, name, retval }
1144  */
1145 static
1146 int
1147 hammer_vop_pathconf(struct vop_pathconf_args *ap)
1148 {
1149         return EOPNOTSUPP;
1150 }
1151
1152 /*
1153  * hammer_vop_print { vp }
1154  */
1155 static
1156 int
1157 hammer_vop_print(struct vop_print_args *ap)
1158 {
1159         return EOPNOTSUPP;
1160 }
1161
1162 /*
1163  * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1164  */
1165 static
1166 int
1167 hammer_vop_readdir(struct vop_readdir_args *ap)
1168 {
1169         struct hammer_transaction trans;
1170         struct hammer_cursor cursor;
1171         struct hammer_inode *ip;
1172         struct uio *uio;
1173         hammer_base_elm_t base;
1174         int error;
1175         int cookie_index;
1176         int ncookies;
1177         off_t *cookies;
1178         off_t saveoff;
1179         int r;
1180
1181         ip = VTOI(ap->a_vp);
1182         uio = ap->a_uio;
1183         saveoff = uio->uio_offset;
1184
1185         if (ap->a_ncookies) {
1186                 ncookies = uio->uio_resid / 16 + 1;
1187                 if (ncookies > 1024)
1188                         ncookies = 1024;
1189                 cookies = kmalloc(ncookies * sizeof(off_t), M_TEMP, M_WAITOK);
1190                 cookie_index = 0;
1191         } else {
1192                 ncookies = -1;
1193                 cookies = NULL;
1194                 cookie_index = 0;
1195         }
1196
1197         hammer_simple_transaction(&trans, ip->hmp);
1198
1199         /*
1200          * Handle artificial entries
1201          */
1202         error = 0;
1203         if (saveoff == 0) {
1204                 r = vop_write_dirent(&error, uio, ip->obj_id, DT_DIR, 1, ".");
1205                 if (r)
1206                         goto done;
1207                 if (cookies)
1208                         cookies[cookie_index] = saveoff;
1209                 ++saveoff;
1210                 ++cookie_index;
1211                 if (cookie_index == ncookies)
1212                         goto done;
1213         }
1214         if (saveoff == 1) {
1215                 if (ip->ino_data.parent_obj_id) {
1216                         r = vop_write_dirent(&error, uio,
1217                                              ip->ino_data.parent_obj_id,
1218                                              DT_DIR, 2, "..");
1219                 } else {
1220                         r = vop_write_dirent(&error, uio,
1221                                              ip->obj_id, DT_DIR, 2, "..");
1222                 }
1223                 if (r)
1224                         goto done;
1225                 if (cookies)
1226                         cookies[cookie_index] = saveoff;
1227                 ++saveoff;
1228                 ++cookie_index;
1229                 if (cookie_index == ncookies)
1230                         goto done;
1231         }
1232
1233         /*
1234          * Key range (begin and end inclusive) to scan.  Directory keys
1235          * directly translate to a 64 bit 'seek' position.
1236          */
1237         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1238         cursor.key_beg.localization = ip->obj_localization +
1239                                       HAMMER_LOCALIZE_MISC;
1240         cursor.key_beg.obj_id = ip->obj_id;
1241         cursor.key_beg.create_tid = 0;
1242         cursor.key_beg.delete_tid = 0;
1243         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1244         cursor.key_beg.obj_type = 0;
1245         cursor.key_beg.key = saveoff;
1246
1247         cursor.key_end = cursor.key_beg;
1248         cursor.key_end.key = HAMMER_MAX_KEY;
1249         cursor.asof = ip->obj_asof;
1250         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1251
1252         error = hammer_ip_first(&cursor);
1253
1254         while (error == 0) {
1255                 error = hammer_ip_resolve_data(&cursor);
1256                 if (error)
1257                         break;
1258                 base = &cursor.leaf->base;
1259                 saveoff = base->key;
1260                 KKASSERT(cursor.leaf->data_len > HAMMER_ENTRY_NAME_OFF);
1261
1262                 if (base->obj_id != ip->obj_id)
1263                         panic("readdir: bad record at %p", cursor.node);
1264
1265                 r = vop_write_dirent(
1266                              &error, uio, cursor.data->entry.obj_id,
1267                              hammer_get_dtype(cursor.leaf->base.obj_type),
1268                              cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF ,
1269                              (void *)cursor.data->entry.name);
1270                 if (r)
1271                         break;
1272                 ++saveoff;
1273                 if (cookies)
1274                         cookies[cookie_index] = base->key;
1275                 ++cookie_index;
1276                 if (cookie_index == ncookies)
1277                         break;
1278                 error = hammer_ip_next(&cursor);
1279         }
1280         hammer_done_cursor(&cursor);
1281
1282 done:
1283         hammer_done_transaction(&trans);
1284
1285         if (ap->a_eofflag)
1286                 *ap->a_eofflag = (error == ENOENT);
1287         uio->uio_offset = saveoff;
1288         if (error && cookie_index == 0) {
1289                 if (error == ENOENT)
1290                         error = 0;
1291                 if (cookies) {
1292                         kfree(cookies, M_TEMP);
1293                         *ap->a_ncookies = 0;
1294                         *ap->a_cookies = NULL;
1295                 }
1296         } else {
1297                 if (error == ENOENT)
1298                         error = 0;
1299                 if (cookies) {
1300                         *ap->a_ncookies = cookie_index;
1301                         *ap->a_cookies = cookies;
1302                 }
1303         }
1304         return(error);
1305 }
1306
1307 /*
1308  * hammer_vop_readlink { vp, uio, cred }
1309  */
1310 static
1311 int
1312 hammer_vop_readlink(struct vop_readlink_args *ap)
1313 {
1314         struct hammer_transaction trans;
1315         struct hammer_cursor cursor;
1316         struct hammer_inode *ip;
1317         int error;
1318
1319         ip = VTOI(ap->a_vp);
1320
1321         /*
1322          * Shortcut if the symlink data was stuffed into ino_data.
1323          */
1324         if (ip->ino_data.size <= HAMMER_INODE_BASESYMLEN) {
1325                 error = uiomove(ip->ino_data.ext.symlink,
1326                                 ip->ino_data.size, ap->a_uio);
1327                 return(error);
1328         }
1329
1330         /*
1331          * Long version
1332          */
1333         hammer_simple_transaction(&trans, ip->hmp);
1334         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1335
1336         /*
1337          * Key range (begin and end inclusive) to scan.  Directory keys
1338          * directly translate to a 64 bit 'seek' position.
1339          */
1340         cursor.key_beg.localization = ip->obj_localization +
1341                                       HAMMER_LOCALIZE_MISC;
1342         cursor.key_beg.obj_id = ip->obj_id;
1343         cursor.key_beg.create_tid = 0;
1344         cursor.key_beg.delete_tid = 0;
1345         cursor.key_beg.rec_type = HAMMER_RECTYPE_FIX;
1346         cursor.key_beg.obj_type = 0;
1347         cursor.key_beg.key = HAMMER_FIXKEY_SYMLINK;
1348         cursor.asof = ip->obj_asof;
1349         cursor.flags |= HAMMER_CURSOR_ASOF;
1350
1351         error = hammer_ip_lookup(&cursor);
1352         if (error == 0) {
1353                 error = hammer_ip_resolve_data(&cursor);
1354                 if (error == 0) {
1355                         KKASSERT(cursor.leaf->data_len >=
1356                                  HAMMER_SYMLINK_NAME_OFF);
1357                         error = uiomove(cursor.data->symlink.name,
1358                                         cursor.leaf->data_len -
1359                                                 HAMMER_SYMLINK_NAME_OFF,
1360                                         ap->a_uio);
1361                 }
1362         }
1363         hammer_done_cursor(&cursor);
1364         hammer_done_transaction(&trans);
1365         return(error);
1366 }
1367
1368 /*
1369  * hammer_vop_nremove { nch, dvp, cred }
1370  */
1371 static
1372 int
1373 hammer_vop_nremove(struct vop_nremove_args *ap)
1374 {
1375         struct hammer_transaction trans;
1376         struct hammer_inode *dip;
1377         int error;
1378
1379         dip = VTOI(ap->a_dvp);
1380
1381         if (hammer_nohistory(dip) == 0 &&
1382             (error = hammer_checkspace(dip->hmp)) != 0) {
1383                 return (error);
1384         }
1385
1386         hammer_start_transaction(&trans, dip->hmp);
1387         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0);
1388         hammer_done_transaction(&trans);
1389
1390         return (error);
1391 }
1392
1393 /*
1394  * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1395  */
1396 static
1397 int
1398 hammer_vop_nrename(struct vop_nrename_args *ap)
1399 {
1400         struct hammer_transaction trans;
1401         struct namecache *fncp;
1402         struct namecache *tncp;
1403         struct hammer_inode *fdip;
1404         struct hammer_inode *tdip;
1405         struct hammer_inode *ip;
1406         struct hammer_cursor cursor;
1407         int64_t namekey;
1408         int nlen, error;
1409
1410         fdip = VTOI(ap->a_fdvp);
1411         tdip = VTOI(ap->a_tdvp);
1412         fncp = ap->a_fnch->ncp;
1413         tncp = ap->a_tnch->ncp;
1414         ip = VTOI(fncp->nc_vp);
1415         KKASSERT(ip != NULL);
1416
1417         if (fdip->flags & HAMMER_INODE_RO)
1418                 return (EROFS);
1419         if (tdip->flags & HAMMER_INODE_RO)
1420                 return (EROFS);
1421         if (ip->flags & HAMMER_INODE_RO)
1422                 return (EROFS);
1423         if ((error = hammer_checkspace(fdip->hmp)) != 0)
1424                 return (error);
1425
1426         hammer_start_transaction(&trans, fdip->hmp);
1427
1428         /*
1429          * Remove tncp from the target directory and then link ip as
1430          * tncp. XXX pass trans to dounlink
1431          *
1432          * Force the inode sync-time to match the transaction so it is
1433          * in-sync with the creation of the target directory entry.
1434          */
1435         error = hammer_dounlink(&trans, ap->a_tnch, ap->a_tdvp, ap->a_cred, 0);
1436         if (error == 0 || error == ENOENT) {
1437                 error = hammer_ip_add_directory(&trans, tdip,
1438                                                 tncp->nc_name, tncp->nc_nlen,
1439                                                 ip);
1440                 if (error == 0) {
1441                         ip->ino_data.parent_obj_id = tdip->obj_id;
1442                         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
1443                 }
1444         }
1445         if (error)
1446                 goto failed; /* XXX */
1447
1448         /*
1449          * Locate the record in the originating directory and remove it.
1450          *
1451          * Calculate the namekey and setup the key range for the scan.  This
1452          * works kinda like a chained hash table where the lower 32 bits
1453          * of the namekey synthesize the chain.
1454          *
1455          * The key range is inclusive of both key_beg and key_end.
1456          */
1457         namekey = hammer_directory_namekey(fncp->nc_name, fncp->nc_nlen);
1458 retry:
1459         hammer_init_cursor(&trans, &cursor, &fdip->cache[1], fdip);
1460         cursor.key_beg.localization = fdip->obj_localization +
1461                                       HAMMER_LOCALIZE_MISC;
1462         cursor.key_beg.obj_id = fdip->obj_id;
1463         cursor.key_beg.key = namekey;
1464         cursor.key_beg.create_tid = 0;
1465         cursor.key_beg.delete_tid = 0;
1466         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1467         cursor.key_beg.obj_type = 0;
1468
1469         cursor.key_end = cursor.key_beg;
1470         cursor.key_end.key |= 0xFFFFFFFFULL;
1471         cursor.asof = fdip->obj_asof;
1472         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1473
1474         /*
1475          * Scan all matching records (the chain), locate the one matching
1476          * the requested path component.
1477          *
1478          * The hammer_ip_*() functions merge in-memory records with on-disk
1479          * records for the purposes of the search.
1480          */
1481         error = hammer_ip_first(&cursor);
1482         while (error == 0) {
1483                 if (hammer_ip_resolve_data(&cursor) != 0)
1484                         break;
1485                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
1486                 KKASSERT(nlen > 0);
1487                 if (fncp->nc_nlen == nlen &&
1488                     bcmp(fncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1489                         break;
1490                 }
1491                 error = hammer_ip_next(&cursor);
1492         }
1493
1494         /*
1495          * If all is ok we have to get the inode so we can adjust nlinks.
1496          *
1497          * WARNING: hammer_ip_del_directory() may have to terminate the
1498          * cursor to avoid a recursion.  It's ok to call hammer_done_cursor()
1499          * twice.
1500          */
1501         if (error == 0)
1502                 error = hammer_ip_del_directory(&trans, &cursor, fdip, ip);
1503
1504         /*
1505          * XXX A deadlock here will break rename's atomicy for the purposes
1506          * of crash recovery.
1507          */
1508         if (error == EDEADLK) {
1509                 hammer_done_cursor(&cursor);
1510                 goto retry;
1511         }
1512
1513         /*
1514          * Cleanup and tell the kernel that the rename succeeded.
1515          */
1516         hammer_done_cursor(&cursor);
1517         if (error == 0)
1518                 cache_rename(ap->a_fnch, ap->a_tnch);
1519
1520 failed:
1521         hammer_done_transaction(&trans);
1522         return (error);
1523 }
1524
1525 /*
1526  * hammer_vop_nrmdir { nch, dvp, cred }
1527  */
1528 static
1529 int
1530 hammer_vop_nrmdir(struct vop_nrmdir_args *ap)
1531 {
1532         struct hammer_transaction trans;
1533         struct hammer_inode *dip;
1534         int error;
1535
1536         dip = VTOI(ap->a_dvp);
1537
1538         if (hammer_nohistory(dip) == 0 &&
1539             (error = hammer_checkspace(dip->hmp)) != 0) {
1540                 return (error);
1541         }
1542
1543         hammer_start_transaction(&trans, dip->hmp);
1544         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0);
1545         hammer_done_transaction(&trans);
1546
1547         return (error);
1548 }
1549
1550 /*
1551  * hammer_vop_setattr { vp, vap, cred }
1552  */
1553 static
1554 int
1555 hammer_vop_setattr(struct vop_setattr_args *ap)
1556 {
1557         struct hammer_transaction trans;
1558         struct vattr *vap;
1559         struct hammer_inode *ip;
1560         int modflags;
1561         int error;
1562         int truncating;
1563         int blksize;
1564         int64_t aligned_size;
1565         u_int32_t flags;
1566
1567         vap = ap->a_vap;
1568         ip = ap->a_vp->v_data;
1569         modflags = 0;
1570
1571         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
1572                 return(EROFS);
1573         if (ip->flags & HAMMER_INODE_RO)
1574                 return (EROFS);
1575         if (hammer_nohistory(ip) == 0 &&
1576             (error = hammer_checkspace(ip->hmp)) != 0) {
1577                 return (error);
1578         }
1579
1580         hammer_start_transaction(&trans, ip->hmp);
1581         error = 0;
1582
1583         if (vap->va_flags != VNOVAL) {
1584                 flags = ip->ino_data.uflags;
1585                 error = vop_helper_setattr_flags(&flags, vap->va_flags,
1586                                          hammer_to_unix_xid(&ip->ino_data.uid),
1587                                          ap->a_cred);
1588                 if (error == 0) {
1589                         if (ip->ino_data.uflags != flags) {
1590                                 ip->ino_data.uflags = flags;
1591                                 modflags |= HAMMER_INODE_DDIRTY;
1592                         }
1593                         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
1594                                 error = 0;
1595                                 goto done;
1596                         }
1597                 }
1598                 goto done;
1599         }
1600         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
1601                 error = EPERM;
1602                 goto done;
1603         }
1604         if (vap->va_uid != (uid_t)VNOVAL || vap->va_gid != (gid_t)VNOVAL) {
1605                 mode_t cur_mode = ip->ino_data.mode;
1606                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1607                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1608                 uuid_t uuid_uid;
1609                 uuid_t uuid_gid;
1610
1611                 error = vop_helper_chown(ap->a_vp, vap->va_uid, vap->va_gid,
1612                                          ap->a_cred,
1613                                          &cur_uid, &cur_gid, &cur_mode);
1614                 if (error == 0) {
1615                         hammer_guid_to_uuid(&uuid_uid, cur_uid);
1616                         hammer_guid_to_uuid(&uuid_gid, cur_gid);
1617                         if (bcmp(&uuid_uid, &ip->ino_data.uid,
1618                                  sizeof(uuid_uid)) ||
1619                             bcmp(&uuid_gid, &ip->ino_data.gid,
1620                                  sizeof(uuid_gid)) ||
1621                             ip->ino_data.mode != cur_mode
1622                         ) {
1623                                 ip->ino_data.uid = uuid_uid;
1624                                 ip->ino_data.gid = uuid_gid;
1625                                 ip->ino_data.mode = cur_mode;
1626                         }
1627                         modflags |= HAMMER_INODE_DDIRTY;
1628                 }
1629         }
1630         while (vap->va_size != VNOVAL && ip->ino_data.size != vap->va_size) {
1631                 switch(ap->a_vp->v_type) {
1632                 case VREG:
1633                         if (vap->va_size == ip->ino_data.size)
1634                                 break;
1635                         /*
1636                          * XXX break atomicy, we can deadlock the backend
1637                          * if we do not release the lock.  Probably not a
1638                          * big deal here.
1639                          */
1640                         blksize = hammer_blocksize(vap->va_size);
1641                         if (vap->va_size < ip->ino_data.size) {
1642                                 vtruncbuf(ap->a_vp, vap->va_size, blksize);
1643                                 truncating = 1;
1644                         } else {
1645                                 vnode_pager_setsize(ap->a_vp, vap->va_size);
1646                                 truncating = 0;
1647                         }
1648                         ip->ino_data.size = vap->va_size;
1649                         modflags |= HAMMER_INODE_DDIRTY;
1650
1651                         /*
1652                          * on-media truncation is cached in the inode until
1653                          * the inode is synchronized.
1654                          */
1655                         if (truncating) {
1656                                 hammer_ip_frontend_trunc(ip, vap->va_size);
1657                                 hammer_update_rsv_databufs(ip);
1658 #ifdef DEBUG_TRUNCATE
1659                                 if (HammerTruncIp == NULL)
1660                                         HammerTruncIp = ip;
1661 #endif
1662                                 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
1663                                         ip->flags |= HAMMER_INODE_TRUNCATED;
1664                                         ip->trunc_off = vap->va_size;
1665 #ifdef DEBUG_TRUNCATE
1666                                         if (ip == HammerTruncIp)
1667                                         kprintf("truncate1 %016llx\n", ip->trunc_off);
1668 #endif
1669                                 } else if (ip->trunc_off > vap->va_size) {
1670                                         ip->trunc_off = vap->va_size;
1671 #ifdef DEBUG_TRUNCATE
1672                                         if (ip == HammerTruncIp)
1673                                         kprintf("truncate2 %016llx\n", ip->trunc_off);
1674 #endif
1675                                 } else {
1676 #ifdef DEBUG_TRUNCATE
1677                                         if (ip == HammerTruncIp)
1678                                         kprintf("truncate3 %016llx (ignored)\n", vap->va_size);
1679 #endif
1680                                 }
1681                         }
1682
1683                         /*
1684                          * If truncating we have to clean out a portion of
1685                          * the last block on-disk.  We do this in the
1686                          * front-end buffer cache.
1687                          */
1688                         aligned_size = (vap->va_size + (blksize - 1)) &
1689                                        ~(int64_t)(blksize - 1);
1690                         if (truncating && vap->va_size < aligned_size) {
1691                                 struct buf *bp;
1692                                 int offset;
1693
1694                                 aligned_size -= blksize;
1695
1696                                 offset = (int)vap->va_size & (blksize - 1);
1697                                 error = bread(ap->a_vp, aligned_size,
1698                                               blksize, &bp);
1699                                 hammer_ip_frontend_trunc(ip, aligned_size);
1700                                 if (error == 0) {
1701                                         bzero(bp->b_data + offset,
1702                                               blksize - offset);
1703                                         bdwrite(bp);
1704                                 } else {
1705                                         kprintf("ERROR %d\n", error);
1706                                         brelse(bp);
1707                                 }
1708                         }
1709                         break;
1710                 case VDATABASE:
1711                         if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
1712                                 ip->flags |= HAMMER_INODE_TRUNCATED;
1713                                 ip->trunc_off = vap->va_size;
1714                         } else if (ip->trunc_off > vap->va_size) {
1715                                 ip->trunc_off = vap->va_size;
1716                         }
1717                         hammer_ip_frontend_trunc(ip, vap->va_size);
1718                         ip->ino_data.size = vap->va_size;
1719                         modflags |= HAMMER_INODE_DDIRTY;
1720                         break;
1721                 default:
1722                         error = EINVAL;
1723                         goto done;
1724                 }
1725                 break;
1726         }
1727         if (vap->va_atime.tv_sec != VNOVAL) {
1728                 ip->ino_data.atime =
1729                         hammer_timespec_to_time(&vap->va_atime);
1730                 modflags |= HAMMER_INODE_ATIME;
1731         }
1732         if (vap->va_mtime.tv_sec != VNOVAL) {
1733                 ip->ino_data.mtime =
1734                         hammer_timespec_to_time(&vap->va_mtime);
1735                 modflags |= HAMMER_INODE_MTIME;
1736         }
1737         if (vap->va_mode != (mode_t)VNOVAL) {
1738                 mode_t   cur_mode = ip->ino_data.mode;
1739                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1740                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1741
1742                 error = vop_helper_chmod(ap->a_vp, vap->va_mode, ap->a_cred,
1743                                          cur_uid, cur_gid, &cur_mode);
1744                 if (error == 0 && ip->ino_data.mode != cur_mode) {
1745                         ip->ino_data.mode = cur_mode;
1746                         modflags |= HAMMER_INODE_DDIRTY;
1747                 }
1748         }
1749 done:
1750         if (error == 0)
1751                 hammer_modify_inode(ip, modflags);
1752         hammer_done_transaction(&trans);
1753         return (error);
1754 }
1755
1756 /*
1757  * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
1758  */
1759 static
1760 int
1761 hammer_vop_nsymlink(struct vop_nsymlink_args *ap)
1762 {
1763         struct hammer_transaction trans;
1764         struct hammer_inode *dip;
1765         struct hammer_inode *nip;
1766         struct nchandle *nch;
1767         hammer_record_t record;
1768         int error;
1769         int bytes;
1770
1771         ap->a_vap->va_type = VLNK;
1772
1773         nch = ap->a_nch;
1774         dip = VTOI(ap->a_dvp);
1775
1776         if (dip->flags & HAMMER_INODE_RO)
1777                 return (EROFS);
1778         if ((error = hammer_checkspace(dip->hmp)) != 0)
1779                 return (error);
1780
1781         /*
1782          * Create a transaction to cover the operations we perform.
1783          */
1784         hammer_start_transaction(&trans, dip->hmp);
1785
1786         /*
1787          * Create a new filesystem object of the requested type.  The
1788          * returned inode will be referenced but not locked.
1789          */
1790
1791         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1792                                     dip, 0, &nip);
1793         if (error) {
1794                 hammer_done_transaction(&trans);
1795                 *ap->a_vpp = NULL;
1796                 return (error);
1797         }
1798
1799         /*
1800          * Add a record representing the symlink.  symlink stores the link
1801          * as pure data, not a string, and is no \0 terminated.
1802          */
1803         if (error == 0) {
1804                 bytes = strlen(ap->a_target);
1805
1806                 if (bytes <= HAMMER_INODE_BASESYMLEN) {
1807                         bcopy(ap->a_target, nip->ino_data.ext.symlink, bytes);
1808                 } else {
1809                         record = hammer_alloc_mem_record(nip, bytes);
1810                         record->type = HAMMER_MEM_RECORD_GENERAL;
1811
1812                         record->leaf.base.localization = nip->obj_localization +
1813                                                          HAMMER_LOCALIZE_MISC;
1814                         record->leaf.base.key = HAMMER_FIXKEY_SYMLINK;
1815                         record->leaf.base.rec_type = HAMMER_RECTYPE_FIX;
1816                         record->leaf.data_len = bytes;
1817                         KKASSERT(HAMMER_SYMLINK_NAME_OFF == 0);
1818                         bcopy(ap->a_target, record->data->symlink.name, bytes);
1819                         error = hammer_ip_add_record(&trans, record);
1820                 }
1821
1822                 /*
1823                  * Set the file size to the length of the link.
1824                  */
1825                 if (error == 0) {
1826                         nip->ino_data.size = bytes;
1827                         hammer_modify_inode(nip, HAMMER_INODE_DDIRTY);
1828                 }
1829         }
1830         if (error == 0)
1831                 error = hammer_ip_add_directory(&trans, dip, nch->ncp->nc_name,
1832                                                 nch->ncp->nc_nlen, nip);
1833
1834         /*
1835          * Finish up.
1836          */
1837         if (error) {
1838                 hammer_rel_inode(nip, 0);
1839                 *ap->a_vpp = NULL;
1840         } else {
1841                 error = hammer_get_vnode(nip, ap->a_vpp);
1842                 hammer_rel_inode(nip, 0);
1843                 if (error == 0) {
1844                         cache_setunresolved(ap->a_nch);
1845                         cache_setvp(ap->a_nch, *ap->a_vpp);
1846                 }
1847         }
1848         hammer_done_transaction(&trans);
1849         return (error);
1850 }
1851
1852 /*
1853  * hammer_vop_nwhiteout { nch, dvp, cred, flags }
1854  */
1855 static
1856 int
1857 hammer_vop_nwhiteout(struct vop_nwhiteout_args *ap)
1858 {
1859         struct hammer_transaction trans;
1860         struct hammer_inode *dip;
1861         int error;
1862
1863         dip = VTOI(ap->a_dvp);
1864
1865         if (hammer_nohistory(dip) == 0 &&
1866             (error = hammer_checkspace(dip->hmp)) != 0) {
1867                 return (error);
1868         }
1869
1870         hammer_start_transaction(&trans, dip->hmp);
1871         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp,
1872                                 ap->a_cred, ap->a_flags);
1873         hammer_done_transaction(&trans);
1874
1875         return (error);
1876 }
1877
1878 /*
1879  * hammer_vop_ioctl { vp, command, data, fflag, cred }
1880  */
1881 static
1882 int
1883 hammer_vop_ioctl(struct vop_ioctl_args *ap)
1884 {
1885         struct hammer_inode *ip = ap->a_vp->v_data;
1886
1887         return(hammer_ioctl(ip, ap->a_command, ap->a_data,
1888                             ap->a_fflag, ap->a_cred));
1889 }
1890
1891 static
1892 int
1893 hammer_vop_mountctl(struct vop_mountctl_args *ap)
1894 {
1895         struct mount *mp;
1896         int error;
1897
1898         mp = ap->a_head.a_ops->head.vv_mount;
1899
1900         switch(ap->a_op) {
1901         case MOUNTCTL_SET_EXPORT:
1902                 if (ap->a_ctllen != sizeof(struct export_args))
1903                         error = EINVAL;
1904                 error = hammer_vfs_export(mp, ap->a_op,
1905                                       (const struct export_args *)ap->a_ctl);
1906                 break;
1907         default:
1908                 error = journal_mountctl(ap);
1909                 break;
1910         }
1911         return(error);
1912 }
1913
1914 /*
1915  * hammer_vop_strategy { vp, bio }
1916  *
1917  * Strategy call, used for regular file read & write only.  Note that the
1918  * bp may represent a cluster.
1919  *
1920  * To simplify operation and allow better optimizations in the future,
1921  * this code does not make any assumptions with regards to buffer alignment
1922  * or size.
1923  */
1924 static
1925 int
1926 hammer_vop_strategy(struct vop_strategy_args *ap)
1927 {
1928         struct buf *bp;
1929         int error;
1930
1931         bp = ap->a_bio->bio_buf;
1932
1933         switch(bp->b_cmd) {
1934         case BUF_CMD_READ:
1935                 error = hammer_vop_strategy_read(ap);
1936                 break;
1937         case BUF_CMD_WRITE:
1938                 error = hammer_vop_strategy_write(ap);
1939                 break;
1940         default:
1941                 bp->b_error = error = EINVAL;
1942                 bp->b_flags |= B_ERROR;
1943                 biodone(ap->a_bio);
1944                 break;
1945         }
1946         return (error);
1947 }
1948
1949 /*
1950  * Read from a regular file.  Iterate the related records and fill in the
1951  * BIO/BUF.  Gaps are zero-filled.
1952  *
1953  * The support code in hammer_object.c should be used to deal with mixed
1954  * in-memory and on-disk records.
1955  *
1956  * NOTE: Can be called from the cluster code with an oversized buf.
1957  *
1958  * XXX atime update
1959  */
1960 static
1961 int
1962 hammer_vop_strategy_read(struct vop_strategy_args *ap)
1963 {
1964         struct hammer_transaction trans;
1965         struct hammer_inode *ip;
1966         struct hammer_cursor cursor;
1967         hammer_base_elm_t base;
1968         hammer_off_t disk_offset;
1969         struct bio *bio;
1970         struct bio *nbio;
1971         struct buf *bp;
1972         int64_t rec_offset;
1973         int64_t ran_end;
1974         int64_t tmp64;
1975         int error;
1976         int boff;
1977         int roff;
1978         int n;
1979
1980         bio = ap->a_bio;
1981         bp = bio->bio_buf;
1982         ip = ap->a_vp->v_data;
1983
1984         /*
1985          * The zone-2 disk offset may have been set by the cluster code via
1986          * a BMAP operation, or else should be NOOFFSET.
1987          *
1988          * Checking the high bits for a match against zone-2 should suffice.
1989          */
1990         nbio = push_bio(bio);
1991         if ((nbio->bio_offset & HAMMER_OFF_ZONE_MASK) ==
1992             HAMMER_ZONE_RAW_BUFFER) {
1993                 error = hammer_io_direct_read(ip->hmp, nbio);
1994                 return (error);
1995         }
1996
1997         /*
1998          * Well, that sucked.  Do it the hard way.  If all the stars are
1999          * aligned we may still be able to issue a direct-read.
2000          */
2001         hammer_simple_transaction(&trans, ip->hmp);
2002         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2003
2004         /*
2005          * Key range (begin and end inclusive) to scan.  Note that the key's
2006          * stored in the actual records represent BASE+LEN, not BASE.  The
2007          * first record containing bio_offset will have a key > bio_offset.
2008          */
2009         cursor.key_beg.localization = ip->obj_localization +
2010                                       HAMMER_LOCALIZE_MISC;
2011         cursor.key_beg.obj_id = ip->obj_id;
2012         cursor.key_beg.create_tid = 0;
2013         cursor.key_beg.delete_tid = 0;
2014         cursor.key_beg.obj_type = 0;
2015         cursor.key_beg.key = bio->bio_offset + 1;
2016         cursor.asof = ip->obj_asof;
2017         cursor.flags |= HAMMER_CURSOR_ASOF;
2018
2019         cursor.key_end = cursor.key_beg;
2020         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2021 #if 0
2022         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
2023                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
2024                 cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
2025                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2026         } else
2027 #endif
2028         {
2029                 ran_end = bio->bio_offset + bp->b_bufsize;
2030                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2031                 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2032                 tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2033                 if (tmp64 < ran_end)
2034                         cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2035                 else
2036                         cursor.key_end.key = ran_end + MAXPHYS + 1;
2037         }
2038         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2039
2040         error = hammer_ip_first(&cursor);
2041         boff = 0;
2042
2043         while (error == 0) {
2044                 /*
2045                  * Get the base file offset of the record.  The key for
2046                  * data records is (base + bytes) rather then (base).
2047                  */
2048                 base = &cursor.leaf->base;
2049                 rec_offset = base->key - cursor.leaf->data_len;
2050
2051                 /*
2052                  * Calculate the gap, if any, and zero-fill it.
2053                  *
2054                  * n is the offset of the start of the record verses our
2055                  * current seek offset in the bio.
2056                  */
2057                 n = (int)(rec_offset - (bio->bio_offset + boff));
2058                 if (n > 0) {
2059                         if (n > bp->b_bufsize - boff)
2060                                 n = bp->b_bufsize - boff;
2061                         bzero((char *)bp->b_data + boff, n);
2062                         boff += n;
2063                         n = 0;
2064                 }
2065
2066                 /*
2067                  * Calculate the data offset in the record and the number
2068                  * of bytes we can copy.
2069                  *
2070                  * There are two degenerate cases.  First, boff may already
2071                  * be at bp->b_bufsize.  Secondly, the data offset within
2072                  * the record may exceed the record's size.
2073                  */
2074                 roff = -n;
2075                 rec_offset += roff;
2076                 n = cursor.leaf->data_len - roff;
2077                 if (n <= 0) {
2078                         kprintf("strategy_read: bad n=%d roff=%d\n", n, roff);
2079                         n = 0;
2080                 } else if (n > bp->b_bufsize - boff) {
2081                         n = bp->b_bufsize - boff;
2082                 }
2083
2084                 /*
2085                  * Deal with cached truncations.  This cool bit of code
2086                  * allows truncate()/ftruncate() to avoid having to sync
2087                  * the file.
2088                  *
2089                  * If the frontend is truncated then all backend records are
2090                  * subject to the frontend's truncation.
2091                  *
2092                  * If the backend is truncated then backend records on-disk
2093                  * (but not in-memory) are subject to the backend's
2094                  * truncation.  In-memory records owned by the backend
2095                  * represent data written after the truncation point on the
2096                  * backend and must not be truncated.
2097                  *
2098                  * Truncate operations deal with frontend buffer cache
2099                  * buffers and frontend-owned in-memory records synchronously.
2100                  */
2101                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2102                         if (hammer_cursor_ondisk(&cursor) ||
2103                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
2104                                 if (ip->trunc_off <= rec_offset)
2105                                         n = 0;
2106                                 else if (ip->trunc_off < rec_offset + n)
2107                                         n = (int)(ip->trunc_off - rec_offset);
2108                         }
2109                 }
2110                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2111                         if (hammer_cursor_ondisk(&cursor)) {
2112                                 if (ip->sync_trunc_off <= rec_offset)
2113                                         n = 0;
2114                                 else if (ip->sync_trunc_off < rec_offset + n)
2115                                         n = (int)(ip->sync_trunc_off - rec_offset);
2116                         }
2117                 }
2118
2119                 /*
2120                  * Try to issue a direct read into our bio if possible,
2121                  * otherwise resolve the element data into a hammer_buffer
2122                  * and copy.
2123                  *
2124                  * The buffer on-disk should be zerod past any real
2125                  * truncation point, but may not be for any synthesized
2126                  * truncation point from above.
2127                  */
2128                 if (boff == 0 && n == bp->b_bufsize &&
2129                     ((cursor.leaf->data_offset + roff) & HAMMER_BUFMASK) == 0) {
2130                         disk_offset = hammer_blockmap_lookup(
2131                                                 trans.hmp,
2132                                                 cursor.leaf->data_offset + roff,
2133                                                 &error);
2134                         if (error)
2135                                 break;
2136                         nbio->bio_offset = disk_offset;
2137                         error = hammer_io_direct_read(trans.hmp, nbio);
2138                         goto done;
2139                 } else if (n) {
2140                         error = hammer_ip_resolve_data(&cursor);
2141                         if (error == 0) {
2142                                 bcopy((char *)cursor.data + roff,
2143                                       (char *)bp->b_data + boff, n);
2144                         }
2145                 }
2146                 if (error)
2147                         break;
2148
2149                 /*
2150                  * Iterate until we have filled the request.
2151                  */
2152                 boff += n;
2153                 if (boff == bp->b_bufsize)
2154                         break;
2155                 error = hammer_ip_next(&cursor);
2156         }
2157
2158         /*
2159          * There may have been a gap after the last record
2160          */
2161         if (error == ENOENT)
2162                 error = 0;
2163         if (error == 0 && boff != bp->b_bufsize) {
2164                 KKASSERT(boff < bp->b_bufsize);
2165                 bzero((char *)bp->b_data + boff, bp->b_bufsize - boff);
2166                 /* boff = bp->b_bufsize; */
2167         }
2168         bp->b_resid = 0;
2169         bp->b_error = error;
2170         if (error)
2171                 bp->b_flags |= B_ERROR;
2172         biodone(ap->a_bio);
2173
2174 done:
2175         if (cursor.node)
2176                 hammer_cache_node(&ip->cache[1], cursor.node);
2177         hammer_done_cursor(&cursor);
2178         hammer_done_transaction(&trans);
2179         return(error);
2180 }
2181
2182 /*
2183  * BMAP operation - used to support cluster_read() only.
2184  *
2185  * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
2186  *
2187  * This routine may return EOPNOTSUPP if the opration is not supported for
2188  * the specified offset.  The contents of the pointer arguments do not
2189  * need to be initialized in that case.
2190  *
2191  * If a disk address is available and properly aligned return 0 with
2192  * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
2193  * to the run-length relative to that offset.  Callers may assume that
2194  * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
2195  * large, so return EOPNOTSUPP if it is not sufficiently large.
2196  */
2197 static
2198 int
2199 hammer_vop_bmap(struct vop_bmap_args *ap)
2200 {
2201         struct hammer_transaction trans;
2202         struct hammer_inode *ip;
2203         struct hammer_cursor cursor;
2204         hammer_base_elm_t base;
2205         int64_t rec_offset;
2206         int64_t ran_end;
2207         int64_t tmp64;
2208         int64_t base_offset;
2209         int64_t base_disk_offset;
2210         int64_t last_offset;
2211         hammer_off_t last_disk_offset;
2212         hammer_off_t disk_offset;
2213         int     rec_len;
2214         int     error;
2215         int     blksize;
2216
2217         ip = ap->a_vp->v_data;
2218
2219         /*
2220          * We can only BMAP regular files.  We can't BMAP database files,
2221          * directories, etc.
2222          */
2223         if (ip->ino_data.obj_type != HAMMER_OBJTYPE_REGFILE)
2224                 return(EOPNOTSUPP);
2225
2226         /*
2227          * bmap is typically called with runp/runb both NULL when used
2228          * for writing.  We do not support BMAP for writing atm.
2229          */
2230         if (ap->a_cmd != BUF_CMD_READ)
2231                 return(EOPNOTSUPP);
2232
2233         /*
2234          * Scan the B-Tree to acquire blockmap addresses, then translate
2235          * to raw addresses.
2236          */
2237         hammer_simple_transaction(&trans, ip->hmp);
2238 #if 0
2239         kprintf("bmap_beg %016llx ip->cache %p\n", ap->a_loffset, ip->cache[1]);
2240 #endif
2241         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2242
2243         /*
2244          * Key range (begin and end inclusive) to scan.  Note that the key's
2245          * stored in the actual records represent BASE+LEN, not BASE.  The
2246          * first record containing bio_offset will have a key > bio_offset.
2247          */
2248         cursor.key_beg.localization = ip->obj_localization +
2249                                       HAMMER_LOCALIZE_MISC;
2250         cursor.key_beg.obj_id = ip->obj_id;
2251         cursor.key_beg.create_tid = 0;
2252         cursor.key_beg.delete_tid = 0;
2253         cursor.key_beg.obj_type = 0;
2254         if (ap->a_runb)
2255                 cursor.key_beg.key = ap->a_loffset - MAXPHYS + 1;
2256         else
2257                 cursor.key_beg.key = ap->a_loffset + 1;
2258         if (cursor.key_beg.key < 0)
2259                 cursor.key_beg.key = 0;
2260         cursor.asof = ip->obj_asof;
2261         cursor.flags |= HAMMER_CURSOR_ASOF;
2262
2263         cursor.key_end = cursor.key_beg;
2264         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2265
2266         ran_end = ap->a_loffset + MAXPHYS;
2267         cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2268         cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2269         tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2270         if (tmp64 < ran_end)
2271                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2272         else
2273                 cursor.key_end.key = ran_end + MAXPHYS + 1;
2274
2275         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2276
2277         error = hammer_ip_first(&cursor);
2278         base_offset = last_offset = 0;
2279         base_disk_offset = last_disk_offset = 0;
2280
2281         while (error == 0) {
2282                 /*
2283                  * Get the base file offset of the record.  The key for
2284                  * data records is (base + bytes) rather then (base).
2285                  *
2286                  * NOTE: rec_offset + rec_len may exceed the end-of-file.
2287                  * The extra bytes should be zero on-disk and the BMAP op
2288                  * should still be ok.
2289                  */
2290                 base = &cursor.leaf->base;
2291                 rec_offset = base->key - cursor.leaf->data_len;
2292                 rec_len    = cursor.leaf->data_len;
2293
2294                 /*
2295                  * Incorporate any cached truncation.
2296                  *
2297                  * NOTE: Modifications to rec_len based on synthesized
2298                  * truncation points remove the guarantee that any extended
2299                  * data on disk is zero (since the truncations may not have
2300                  * taken place on-media yet).
2301                  */
2302                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2303                         if (hammer_cursor_ondisk(&cursor) ||
2304                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
2305                                 if (ip->trunc_off <= rec_offset)
2306                                         rec_len = 0;
2307                                 else if (ip->trunc_off < rec_offset + rec_len)
2308                                         rec_len = (int)(ip->trunc_off - rec_offset);
2309                         }
2310                 }
2311                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2312                         if (hammer_cursor_ondisk(&cursor)) {
2313                                 if (ip->sync_trunc_off <= rec_offset)
2314                                         rec_len = 0;
2315                                 else if (ip->sync_trunc_off < rec_offset + rec_len)
2316                                         rec_len = (int)(ip->sync_trunc_off - rec_offset);
2317                         }
2318                 }
2319
2320                 /*
2321                  * Accumulate information.  If we have hit a discontiguous
2322                  * block reset base_offset unless we are already beyond the
2323                  * requested offset.  If we are, that's it, we stop.
2324                  */
2325                 disk_offset = hammer_blockmap_lookup(trans.hmp,
2326                                                      cursor.leaf->data_offset,
2327                                                      &error);
2328                 if (error)
2329                         break;
2330                 if (rec_offset != last_offset ||
2331                     disk_offset != last_disk_offset) {
2332                         if (rec_offset > ap->a_loffset)
2333                                 break;
2334                         base_offset = rec_offset;
2335                         base_disk_offset = disk_offset;
2336                 }
2337                 last_offset = rec_offset + rec_len;
2338                 last_disk_offset = disk_offset + rec_len;
2339
2340                 error = hammer_ip_next(&cursor);
2341         }
2342
2343 #if 0
2344         kprintf("BMAP %016llx:  %016llx - %016llx\n",
2345                 ap->a_loffset, base_offset, last_offset);
2346         kprintf("BMAP %16s:  %016llx - %016llx\n",
2347                 "", base_disk_offset, last_disk_offset);
2348 #endif
2349
2350         if (cursor.node) {
2351                 hammer_cache_node(&ip->cache[1], cursor.node);
2352 #if 0
2353                 kprintf("bmap_end2 %016llx ip->cache %p\n", ap->a_loffset, ip->cache[1]);
2354 #endif
2355         }
2356         hammer_done_cursor(&cursor);
2357         hammer_done_transaction(&trans);
2358
2359         /*
2360          * If we couldn't find any records or the records we did find were
2361          * all behind the requested offset, return failure.  A forward
2362          * truncation can leave a hole w/ no on-disk records.
2363          */
2364         if (last_offset == 0 || last_offset < ap->a_loffset)
2365                 return (EOPNOTSUPP);
2366
2367         /*
2368          * Figure out the block size at the requested offset and adjust
2369          * our limits so the cluster_read() does not create inappropriately
2370          * sized buffer cache buffers.
2371          */
2372         blksize = hammer_blocksize(ap->a_loffset);
2373         if (hammer_blocksize(base_offset) != blksize) {
2374                 base_offset = hammer_blockdemarc(base_offset, ap->a_loffset);
2375         }
2376         if (last_offset != ap->a_loffset &&
2377             hammer_blocksize(last_offset - 1) != blksize) {
2378                 last_offset = hammer_blockdemarc(ap->a_loffset,
2379                                                  last_offset - 1);
2380         }
2381
2382         /*
2383          * Returning EOPNOTSUPP simply prevents the direct-IO optimization
2384          * from occuring.
2385          */
2386         disk_offset = base_disk_offset + (ap->a_loffset - base_offset);
2387
2388         /*
2389          * If doffsetp is not aligned or the forward run size does
2390          * not cover a whole buffer, disallow the direct I/O.
2391          */
2392         if ((disk_offset & HAMMER_BUFMASK) ||
2393             (last_offset - ap->a_loffset) < blksize) {
2394                 error = EOPNOTSUPP;
2395         } else {
2396                 *ap->a_doffsetp = disk_offset;
2397                 if (ap->a_runb) {
2398                         *ap->a_runb = ap->a_loffset - base_offset;
2399                         KKASSERT(*ap->a_runb >= 0);
2400                 }
2401                 if (ap->a_runp) {
2402                         *ap->a_runp = last_offset - ap->a_loffset;
2403                         KKASSERT(*ap->a_runp >= 0);
2404                 }
2405                 error = 0;
2406         }
2407         return(error);
2408 }
2409
2410 /*
2411  * Write to a regular file.   Because this is a strategy call the OS is
2412  * trying to actually get data onto the media.
2413  */
2414 static
2415 int
2416 hammer_vop_strategy_write(struct vop_strategy_args *ap)
2417 {
2418         hammer_record_t record;
2419         hammer_mount_t hmp;
2420         hammer_inode_t ip;
2421         struct bio *bio;
2422         struct buf *bp;
2423         int bytes;
2424         int error;
2425
2426         bio = ap->a_bio;
2427         bp = bio->bio_buf;
2428         ip = ap->a_vp->v_data;
2429         hmp = ip->hmp;
2430
2431         KKASSERT(bp->b_bufsize == hammer_blocksize(bio->bio_offset));
2432
2433         if (ip->flags & HAMMER_INODE_RO) {
2434                 bp->b_error = EROFS;
2435                 bp->b_flags |= B_ERROR;
2436                 biodone(ap->a_bio);
2437                 hammer_cleanup_write_io(ip);
2438                 return(EROFS);
2439         }
2440
2441         /*
2442          * Interlock with inode destruction (no in-kernel or directory
2443          * topology visibility).  If we queue new IO while trying to
2444          * destroy the inode we can deadlock the vtrunc call in
2445          * hammer_inode_unloadable_check().
2446          */
2447         if (ip->flags & (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) {
2448                 bp->b_resid = 0;
2449                 biodone(ap->a_bio);
2450                 hammer_cleanup_write_io(ip);
2451                 return(0);
2452         }
2453
2454         /*
2455          * Reserve space and issue a direct-write from the front-end.
2456          * NOTE: The direct_io code will hammer_bread/bcopy smaller
2457          * allocations.
2458          *
2459          * An in-memory record will be installed to reference the storage
2460          * until the flusher can get to it.
2461          *
2462          * Since we own the high level bio the front-end will not try to
2463          * do a direct-read until the write completes.
2464          *
2465          * NOTE: The only time we do not reserve a full-sized buffers
2466          * worth of data is if the file is small.  We do not try to
2467          * allocate a fragment (from the small-data zone) at the end of
2468          * an otherwise large file as this can lead to wildly separated
2469          * data.
2470          */
2471         KKASSERT((bio->bio_offset & HAMMER_BUFMASK) == 0);
2472         KKASSERT(bio->bio_offset < ip->ino_data.size);
2473         if (bio->bio_offset || ip->ino_data.size > HAMMER_BUFSIZE / 2)
2474                 bytes = bp->b_bufsize;
2475         else
2476                 bytes = ((int)ip->ino_data.size + 15) & ~15;
2477
2478         record = hammer_ip_add_bulk(ip, bio->bio_offset, bp->b_data,
2479                                     bytes, &error);
2480         if (record) {
2481                 hammer_io_direct_write(hmp, &record->leaf, bio);
2482                 hammer_rel_mem_record(record);
2483                 if (ip->rsv_recs > 1 && hmp->rsv_recs > hammer_limit_recs)
2484                         hammer_flush_inode(ip, 0);
2485         } else {
2486                 bp->b_bio2.bio_offset = NOOFFSET;
2487                 bp->b_error = error;
2488                 bp->b_flags |= B_ERROR;
2489                 biodone(ap->a_bio);
2490         }
2491         hammer_cleanup_write_io(ip);
2492         return(error);
2493 }
2494
2495 /*
2496  * Clean-up after disposing of a dirty frontend buffer's data.
2497  * This is somewhat heuristical so try to be robust.
2498  */
2499 static void
2500 hammer_cleanup_write_io(hammer_inode_t ip)
2501 {
2502         if (ip->rsv_databufs) {
2503                 --ip->rsv_databufs;
2504                 --ip->hmp->rsv_databufs;
2505         }
2506 }
2507
2508 /*
2509  * We can lose track of dirty buffer cache buffers if we truncate, this
2510  * routine will resynchronize the count.
2511  */
2512 static
2513 void
2514 hammer_update_rsv_databufs(hammer_inode_t ip)
2515 {
2516         struct buf *bp;
2517         int delta;
2518         int n;
2519
2520         if (ip->vp) {
2521                 n = 0;
2522                 RB_FOREACH(bp, buf_rb_tree, &ip->vp->v_rbdirty_tree) {
2523                         ++n;
2524                 }
2525         } else {
2526                 n = 0;
2527         }
2528         delta = n - ip->rsv_databufs;
2529         ip->rsv_databufs += delta;
2530         ip->hmp->rsv_databufs += delta;
2531 }
2532
2533 /*
2534  * dounlink - disconnect a directory entry
2535  *
2536  * XXX whiteout support not really in yet
2537  */
2538 static int
2539 hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
2540                 struct vnode *dvp, struct ucred *cred, int flags)
2541 {
2542         struct namecache *ncp;
2543         hammer_inode_t dip;
2544         hammer_inode_t ip;
2545         struct hammer_cursor cursor;
2546         int64_t namekey;
2547         int nlen, error;
2548
2549         /*
2550          * Calculate the namekey and setup the key range for the scan.  This
2551          * works kinda like a chained hash table where the lower 32 bits
2552          * of the namekey synthesize the chain.
2553          *
2554          * The key range is inclusive of both key_beg and key_end.
2555          */
2556         dip = VTOI(dvp);
2557         ncp = nch->ncp;
2558
2559         if (dip->flags & HAMMER_INODE_RO)
2560                 return (EROFS);
2561
2562         namekey = hammer_directory_namekey(ncp->nc_name, ncp->nc_nlen);
2563 retry:
2564         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
2565         cursor.key_beg.localization = dip->obj_localization +
2566                                       HAMMER_LOCALIZE_MISC;
2567         cursor.key_beg.obj_id = dip->obj_id;
2568         cursor.key_beg.key = namekey;
2569         cursor.key_beg.create_tid = 0;
2570         cursor.key_beg.delete_tid = 0;
2571         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
2572         cursor.key_beg.obj_type = 0;
2573
2574         cursor.key_end = cursor.key_beg;
2575         cursor.key_end.key |= 0xFFFFFFFFULL;
2576         cursor.asof = dip->obj_asof;
2577         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2578
2579         /*
2580          * Scan all matching records (the chain), locate the one matching
2581          * the requested path component.  info->last_error contains the
2582          * error code on search termination and could be 0, ENOENT, or
2583          * something else.
2584          *
2585          * The hammer_ip_*() functions merge in-memory records with on-disk
2586          * records for the purposes of the search.
2587          */
2588         error = hammer_ip_first(&cursor);
2589
2590         while (error == 0) {
2591                 error = hammer_ip_resolve_data(&cursor);
2592                 if (error)
2593                         break;
2594                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
2595                 KKASSERT(nlen > 0);
2596                 if (ncp->nc_nlen == nlen &&
2597                     bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
2598                         break;
2599                 }
2600                 error = hammer_ip_next(&cursor);
2601         }
2602
2603         /*
2604          * If all is ok we have to get the inode so we can adjust nlinks.
2605          * To avoid a deadlock with the flusher we must release the inode
2606          * lock on the directory when acquiring the inode for the entry.
2607          *
2608          * If the target is a directory, it must be empty.
2609          */
2610         if (error == 0) {
2611                 hammer_unlock(&cursor.ip->lock);
2612                 ip = hammer_get_inode(trans, dip, cursor.data->entry.obj_id,
2613                                       dip->hmp->asof,
2614                                       cursor.data->entry.localization,
2615                                       0, &error);
2616                 hammer_lock_sh(&cursor.ip->lock);
2617                 if (error == ENOENT) {
2618                         kprintf("obj_id %016llx\n", cursor.data->entry.obj_id);
2619                         Debugger("ENOENT unlinking object that should exist");
2620                 }
2621
2622                 /*
2623                  * If we are trying to remove a directory the directory must
2624                  * be empty.
2625                  *
2626                  * WARNING: hammer_ip_check_directory_empty() may have to
2627                  * terminate the cursor to avoid a deadlock.  It is ok to
2628                  * call hammer_done_cursor() twice.
2629                  */
2630                 if (error == 0 && ip->ino_data.obj_type ==
2631                                   HAMMER_OBJTYPE_DIRECTORY) {
2632                         error = hammer_ip_check_directory_empty(trans, ip);
2633                 }
2634
2635                 /*
2636                  * Delete the directory entry.
2637                  *
2638                  * WARNING: hammer_ip_del_directory() may have to terminate
2639                  * the cursor to avoid a deadlock.  It is ok to call
2640                  * hammer_done_cursor() twice.
2641                  */
2642                 if (error == 0) {
2643                         error = hammer_ip_del_directory(trans, &cursor,
2644                                                         dip, ip);
2645                 }
2646                 hammer_done_cursor(&cursor);
2647                 if (error == 0) {
2648                         cache_setunresolved(nch);
2649                         cache_setvp(nch, NULL);
2650                         /* XXX locking */
2651                         if (ip->vp)
2652                                 cache_inval_vp(ip->vp, CINV_DESTROY);
2653                 }
2654                 if (ip)
2655                         hammer_rel_inode(ip, 0);
2656         } else {
2657                 hammer_done_cursor(&cursor);
2658         }
2659         if (error == EDEADLK)
2660                 goto retry;
2661
2662         return (error);
2663 }
2664
2665 /************************************************************************
2666  *                          FIFO AND SPECFS OPS                         *
2667  ************************************************************************
2668  *
2669  */
2670
2671 static int
2672 hammer_vop_fifoclose (struct vop_close_args *ap)
2673 {
2674         /* XXX update itimes */
2675         return (VOCALL(&fifo_vnode_vops, &ap->a_head));
2676 }
2677
2678 static int
2679 hammer_vop_fiforead (struct vop_read_args *ap)
2680 {
2681         int error;
2682
2683         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
2684         /* XXX update access time */
2685         return (error);
2686 }
2687
2688 static int
2689 hammer_vop_fifowrite (struct vop_write_args *ap)
2690 {
2691         int error;
2692
2693         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
2694         /* XXX update access time */
2695         return (error);
2696 }
2697
2698 static int
2699 hammer_vop_specclose (struct vop_close_args *ap)
2700 {
2701         /* XXX update itimes */
2702         return (VOCALL(&spec_vnode_vops, &ap->a_head));
2703 }
2704
2705 static int
2706 hammer_vop_specread (struct vop_read_args *ap)
2707 {
2708         /* XXX update access time */
2709         return (VOCALL(&spec_vnode_vops, &ap->a_head));
2710 }
2711
2712 static int
2713 hammer_vop_specwrite (struct vop_write_args *ap)
2714 {
2715         /* XXX update last change time */
2716         return (VOCALL(&spec_vnode_vops, &ap->a_head));
2717 }
2718