sys/vfs/hammer/hammer_vnops.c

   1 /*
   2  * Copyright (c) 2007-2008 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  *
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in
  15  *    the documentation and/or other materials provided with the
  16  *    distribution.
  17  * 3. Neither the name of The DragonFly Project nor the names of its
  18  *    contributors may be used to endorse or promote products derived
  19  *    from this software without specific, prior written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  32  * SUCH DAMAGE.
  33  *
  34  * $DragonFly: src/sys/vfs/hammer/hammer_vnops.c,v 1.102 2008/10/16 17:24:16 dillon Exp $
  35  */
  36
  37 #include <sys/param.h>
  38 #include <sys/systm.h>
  39 #include <sys/kernel.h>
  40 #include <sys/fcntl.h>
  41 #include <sys/namecache.h>
  42 #include <sys/vnode.h>
  43 #include <sys/lockf.h>
  44 #include <sys/event.h>
  45 #include <sys/stat.h>
  46 #include <sys/dirent.h>
  47 #include <sys/file.h>
  48 #include <vm/vm_extern.h>
  49 #include <vm/swap_pager.h>
  50 #include <vfs/fifofs/fifo.h>
  51
  52 #include "hammer.h"
  53
  54 /*
  55  * USERFS VNOPS
  56  */
  57 /*static int hammer_vop_vnoperate(struct vop_generic_args *);*/
  58 static int hammer_vop_fsync(struct vop_fsync_args *);
  59 static int hammer_vop_read(struct vop_read_args *);
  60 static int hammer_vop_write(struct vop_write_args *);
  61 static int hammer_vop_access(struct vop_access_args *);
  62 static int hammer_vop_advlock(struct vop_advlock_args *);
  63 static int hammer_vop_close(struct vop_close_args *);
  64 static int hammer_vop_ncreate(struct vop_ncreate_args *);
  65 static int hammer_vop_getattr(struct vop_getattr_args *);
  66 static int hammer_vop_nresolve(struct vop_nresolve_args *);
  67 static int hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *);
  68 static int hammer_vop_nlink(struct vop_nlink_args *);
  69 static int hammer_vop_nmkdir(struct vop_nmkdir_args *);
  70 static int hammer_vop_nmknod(struct vop_nmknod_args *);
  71 static int hammer_vop_open(struct vop_open_args *);
  72 static int hammer_vop_print(struct vop_print_args *);
  73 static int hammer_vop_readdir(struct vop_readdir_args *);
  74 static int hammer_vop_readlink(struct vop_readlink_args *);
  75 static int hammer_vop_nremove(struct vop_nremove_args *);
  76 static int hammer_vop_nrename(struct vop_nrename_args *);
  77 static int hammer_vop_nrmdir(struct vop_nrmdir_args *);
  78 static int hammer_vop_markatime(struct vop_markatime_args *);
  79 static int hammer_vop_setattr(struct vop_setattr_args *);
  80 static int hammer_vop_strategy(struct vop_strategy_args *);
  81 static int hammer_vop_bmap(struct vop_bmap_args *ap);
  82 static int hammer_vop_nsymlink(struct vop_nsymlink_args *);
  83 static int hammer_vop_nwhiteout(struct vop_nwhiteout_args *);
  84 static int hammer_vop_ioctl(struct vop_ioctl_args *);
  85 static int hammer_vop_mountctl(struct vop_mountctl_args *);
  86 static int hammer_vop_kqfilter (struct vop_kqfilter_args *);
  87
  88 static int hammer_vop_fifoclose (struct vop_close_args *);
  89 static int hammer_vop_fiforead (struct vop_read_args *);
  90 static int hammer_vop_fifowrite (struct vop_write_args *);
  91 static int hammer_vop_fifokqfilter (struct vop_kqfilter_args *);
  92
  93 struct vop_ops hammer_vnode_vops = {
  94         .vop_default =          vop_defaultop,
  95         .vop_fsync =            hammer_vop_fsync,
  96         .vop_getpages =         vop_stdgetpages,
  97         .vop_putpages =         vop_stdputpages,
  98         .vop_read =             hammer_vop_read,
  99         .vop_write =            hammer_vop_write,
 100         .vop_access =           hammer_vop_access,
 101         .vop_advlock =          hammer_vop_advlock,
 102         .vop_close =            hammer_vop_close,
 103         .vop_ncreate =          hammer_vop_ncreate,
 104         .vop_getattr =          hammer_vop_getattr,
 105         .vop_inactive =         hammer_vop_inactive,
 106         .vop_reclaim =          hammer_vop_reclaim,
 107         .vop_nresolve =         hammer_vop_nresolve,
 108         .vop_nlookupdotdot =    hammer_vop_nlookupdotdot,
 109         .vop_nlink =            hammer_vop_nlink,
 110         .vop_nmkdir =           hammer_vop_nmkdir,
 111         .vop_nmknod =           hammer_vop_nmknod,
 112         .vop_open =             hammer_vop_open,
 113         .vop_pathconf =         vop_stdpathconf,
 114         .vop_print =            hammer_vop_print,
 115         .vop_readdir =          hammer_vop_readdir,
 116         .vop_readlink =         hammer_vop_readlink,
 117         .vop_nremove =          hammer_vop_nremove,
 118         .vop_nrename =          hammer_vop_nrename,
 119         .vop_nrmdir =           hammer_vop_nrmdir,
 120         .vop_markatime =        hammer_vop_markatime,
 121         .vop_setattr =          hammer_vop_setattr,
 122         .vop_bmap =             hammer_vop_bmap,
 123         .vop_strategy =         hammer_vop_strategy,
 124         .vop_nsymlink =         hammer_vop_nsymlink,
 125         .vop_nwhiteout =        hammer_vop_nwhiteout,
 126         .vop_ioctl =            hammer_vop_ioctl,
 127         .vop_mountctl =         hammer_vop_mountctl,
 128         .vop_kqfilter =         hammer_vop_kqfilter
 129 };
 130
 131 struct vop_ops hammer_spec_vops = {
 132         .vop_default =          vop_defaultop,
 133         .vop_fsync =            hammer_vop_fsync,
 134         .vop_read =             vop_stdnoread,
 135         .vop_write =            vop_stdnowrite,
 136         .vop_access =           hammer_vop_access,
 137         .vop_close =            hammer_vop_close,
 138         .vop_markatime =        hammer_vop_markatime,
 139         .vop_getattr =          hammer_vop_getattr,
 140         .vop_inactive =         hammer_vop_inactive,
 141         .vop_reclaim =          hammer_vop_reclaim,
 142         .vop_setattr =          hammer_vop_setattr
 143 };
 144
 145 struct vop_ops hammer_fifo_vops = {
 146         .vop_default =          fifo_vnoperate,
 147         .vop_fsync =            hammer_vop_fsync,
 148         .vop_read =             hammer_vop_fiforead,
 149         .vop_write =            hammer_vop_fifowrite,
 150         .vop_access =           hammer_vop_access,
 151         .vop_close =            hammer_vop_fifoclose,
 152         .vop_markatime =        hammer_vop_markatime,
 153         .vop_getattr =          hammer_vop_getattr,
 154         .vop_inactive =         hammer_vop_inactive,
 155         .vop_reclaim =          hammer_vop_reclaim,
 156         .vop_setattr =          hammer_vop_setattr,
 157         .vop_kqfilter =         hammer_vop_fifokqfilter
 158 };
 159
 160 static __inline
 161 void
 162 hammer_knote(struct vnode *vp, int flags)
 163 {
 164         if (flags)
 165                 KNOTE(&vp->v_pollinfo.vpi_kqinfo.ki_note, flags);
 166 }
 167
 168 #ifdef DEBUG_TRUNCATE
 169 struct hammer_inode *HammerTruncIp;
 170 #endif
 171
 172 static int hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
 173                            struct vnode *dvp, struct ucred *cred,
 174                            int flags, int isdir);
 175 static int hammer_vop_strategy_read(struct vop_strategy_args *ap);
 176 static int hammer_vop_strategy_write(struct vop_strategy_args *ap);
 177
 178 #if 0
 179 static
 180 int
 181 hammer_vop_vnoperate(struct vop_generic_args *)
 182 {
 183         return (VOCALL(&hammer_vnode_vops, ap));
 184 }
 185 #endif
 186
 187 /*
 188  * hammer_vop_fsync { vp, waitfor }
 189  *
 190  * fsync() an inode to disk and wait for it to be completely committed
 191  * such that the information would not be undone if a crash occured after
 192  * return.
 193  *
 194  * NOTE: HAMMER's fsync()'s are going to remain expensive until we implement
 195  *       a REDO log.  A sysctl is provided to relax HAMMER's fsync()
 196  *       operation.
 197  *
 198  *       Ultimately the combination of a REDO log and use of fast storage
 199  *       to front-end cluster caches will make fsync fast, but it aint
 200  *       here yet.  And, in anycase, we need real transactional
 201  *       all-or-nothing features which are not restricted to a single file.
 202  */
 203 static
 204 int
 205 hammer_vop_fsync(struct vop_fsync_args *ap)
 206 {
 207         hammer_inode_t ip = VTOI(ap->a_vp);
 208         hammer_mount_t hmp = ip->hmp;
 209         int waitfor = ap->a_waitfor;
 210         int mode;
 211
 212         lwkt_gettoken(&hmp->fs_token);
 213
 214         /*
 215          * Fsync rule relaxation (default is either full synchronous flush
 216          * or REDO semantics with synchronous flush).
 217          */
 218         if (ap->a_flags & VOP_FSYNC_SYSCALL) {
 219                 switch(hammer_fsync_mode) {
 220                 case 0:
 221 mode0:
 222                         /* no REDO, full synchronous flush */
 223                         goto skip;
 224                 case 1:
 225 mode1:
 226                         /* no REDO, full asynchronous flush */
 227                         if (waitfor == MNT_WAIT)
 228                                 waitfor = MNT_NOWAIT;
 229                         goto skip;
 230                 case 2:
 231                         /* REDO semantics, synchronous flush */
 232                         if (hmp->version < HAMMER_VOL_VERSION_FOUR)
 233                                 goto mode0;
 234                         mode = HAMMER_FLUSH_UNDOS_AUTO;
 235                         break;
 236                 case 3:
 237                         /* REDO semantics, relaxed asynchronous flush */
 238                         if (hmp->version < HAMMER_VOL_VERSION_FOUR)
 239                                 goto mode1;
 240                         mode = HAMMER_FLUSH_UNDOS_RELAXED;
 241                         if (waitfor == MNT_WAIT)
 242                                 waitfor = MNT_NOWAIT;
 243                         break;
 244                 case 4:
 245                         /* ignore the fsync() system call */
 246                         lwkt_reltoken(&hmp->fs_token);
 247                         return(0);
 248                 default:
 249                         /* we have to do something */
 250                         mode = HAMMER_FLUSH_UNDOS_RELAXED;
 251                         if (waitfor == MNT_WAIT)
 252                                 waitfor = MNT_NOWAIT;
 253                         break;
 254                 }
 255
 256                 /*
 257                  * Fast fsync only needs to flush the UNDO/REDO fifo if
 258                  * HAMMER_INODE_REDO is non-zero and the only modifications
 259                  * made to the file are write or write-extends.
 260                  */
 261                 if ((ip->flags & HAMMER_INODE_REDO) &&
 262                     (ip->flags & HAMMER_INODE_MODMASK_NOREDO) == 0
 263                 ) {
 264                         ++hammer_count_fsyncs;
 265                         hammer_flusher_flush_undos(hmp, mode);
 266                         ip->redo_count = 0;
 267                         lwkt_reltoken(&hmp->fs_token);
 268                         return(0);
 269                 }
 270
 271                 /*
 272                  * REDO is enabled by fsync(), the idea being we really only
 273                  * want to lay down REDO records when programs are using
 274                  * fsync() heavily.  The first fsync() on the file starts
 275                  * the gravy train going and later fsync()s keep it hot by
 276                  * resetting the redo_count.
 277                  *
 278                  * We weren't running REDOs before now so we have to fall
 279                  * through and do a full fsync of what we have.
 280                  */
 281                 if (hmp->version >= HAMMER_VOL_VERSION_FOUR &&
 282                     (hmp->flags & HAMMER_MOUNT_REDO_RECOVERY_RUN) == 0) {
 283                         ip->flags |= HAMMER_INODE_REDO;
 284                         ip->redo_count = 0;
 285                 }
 286         }
 287 skip:
 288
 289         /*
 290          * Do a full flush sequence.
 291          */
 292         ++hammer_count_fsyncs;
 293         vfsync(ap->a_vp, waitfor, 1, NULL, NULL);
 294         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 295         if (waitfor == MNT_WAIT) {
 296                 vn_unlock(ap->a_vp);
 297                 hammer_wait_inode(ip);
 298                 vn_lock(ap->a_vp, LK_EXCLUSIVE | LK_RETRY);
 299         }
 300         lwkt_reltoken(&hmp->fs_token);
 301         return (ip->error);
 302 }
 303
 304 /*
 305  * hammer_vop_read { vp, uio, ioflag, cred }
 306  *
 307  * MPSAFE (for the cache safe does not require fs_token)
 308  */
 309 static
 310 int
 311 hammer_vop_read(struct vop_read_args *ap)
 312 {
 313         struct hammer_transaction trans;
 314         hammer_inode_t ip;
 315         hammer_mount_t hmp;
 316         off_t offset;
 317         struct buf *bp;
 318         struct uio *uio;
 319         int error;
 320         int n;
 321         int seqcount;
 322         int ioseqcount;
 323         int blksize;
 324         int bigread;
 325         int got_fstoken;
 326
 327         if (ap->a_vp->v_type != VREG)
 328                 return (EINVAL);
 329         ip = VTOI(ap->a_vp);
 330         hmp = ip->hmp;
 331         error = 0;
 332         uio = ap->a_uio;
 333
 334         /*
 335          * Allow the UIO's size to override the sequential heuristic.
 336          */
 337         blksize = hammer_blocksize(uio->uio_offset);
 338         seqcount = (uio->uio_resid + (BKVASIZE - 1)) / BKVASIZE;
 339         ioseqcount = (ap->a_ioflag >> 16);
 340         if (seqcount < ioseqcount)
 341                 seqcount = ioseqcount;
 342
 343         /*
 344          * If reading or writing a huge amount of data we have to break
 345          * atomicy and allow the operation to be interrupted by a signal
 346          * or it can DOS the machine.
 347          */
 348         bigread = (uio->uio_resid > 100 * 1024 * 1024);
 349         got_fstoken = 0;
 350
 351         /*
 352          * Access the data typically in HAMMER_BUFSIZE blocks via the
 353          * buffer cache, but HAMMER may use a variable block size based
 354          * on the offset.
 355          *
 356          * XXX Temporary hack, delay the start transaction while we remain
 357          *     MPSAFE.  NOTE: ino_data.size cannot change while vnode is
 358          *     locked-shared.
 359          */
 360         while (uio->uio_resid > 0 && uio->uio_offset < ip->ino_data.size) {
 361                 int64_t base_offset;
 362                 int64_t file_limit;
 363
 364                 blksize = hammer_blocksize(uio->uio_offset);
 365                 offset = (int)uio->uio_offset & (blksize - 1);
 366                 base_offset = uio->uio_offset - offset;
 367
 368                 if (bigread && (error = hammer_signal_check(ip->hmp)) != 0)
 369                         break;
 370
 371                 /*
 372                  * MPSAFE
 373                  */
 374                 bp = getcacheblk(ap->a_vp, base_offset, blksize);
 375                 if (bp) {
 376                         error = 0;
 377                         goto skip;
 378                 } else {
 379                         if (ap->a_ioflag & IO_NRDELAY)
 380                                 return (EWOULDBLOCK);
 381                 }
 382
 383                 /*
 384                  * MPUNSAFE
 385                  */
 386                 if (got_fstoken == 0) {
 387                         lwkt_gettoken(&hmp->fs_token);
 388                         got_fstoken = 1;
 389                         hammer_start_transaction(&trans, ip->hmp);
 390                 }
 391
 392                 if (hammer_cluster_enable) {
 393                         /*
 394                          * Use file_limit to prevent cluster_read() from
 395                          * creating buffers of the wrong block size past
 396                          * the demarc.
 397                          */
 398                         file_limit = ip->ino_data.size;
 399                         if (base_offset < HAMMER_XDEMARC &&
 400                             file_limit > HAMMER_XDEMARC) {
 401                                 file_limit = HAMMER_XDEMARC;
 402                         }
 403                         error = cluster_read(ap->a_vp,
 404                                              file_limit, base_offset,
 405                                              blksize, uio->uio_resid,
 406                                              seqcount * BKVASIZE, &bp);
 407                 } else {
 408                         error = bread(ap->a_vp, base_offset, blksize, &bp);
 409                 }
 410                 if (error) {
 411                         brelse(bp);
 412                         break;
 413                 }
 414 skip:
 415                 if ((hammer_debug_io & 0x0001) && (bp->b_flags & B_IODEBUG)) {
 416                         kprintf("doff %016jx read file %016jx@%016jx\n",
 417                                 (intmax_t)bp->b_bio2.bio_offset,
 418                                 (intmax_t)ip->obj_id,
 419                                 (intmax_t)bp->b_loffset);
 420                 }
 421                 bp->b_flags &= ~B_IODEBUG;
 422
 423                 /* bp->b_flags |= B_CLUSTEROK; temporarily disabled */
 424                 n = blksize - offset;
 425                 if (n > uio->uio_resid)
 426                         n = uio->uio_resid;
 427                 if (n > ip->ino_data.size - uio->uio_offset)
 428                         n = (int)(ip->ino_data.size - uio->uio_offset);
 429                 if (got_fstoken)
 430                         lwkt_reltoken(&hmp->fs_token);
 431
 432                 /*
 433                  * Set B_AGE, data has a lower priority than meta-data.
 434                  *
 435                  * Use a hold/unlock/drop sequence to run the uiomove
 436                  * with the buffer unlocked, avoiding deadlocks against
 437                  * read()s on mmap()'d spaces.
 438                  */
 439                 bp->b_flags |= B_AGE;
 440                 bqhold(bp);
 441                 bqrelse(bp);
 442                 error = uiomove((char *)bp->b_data + offset, n, uio);
 443                 bqdrop(bp);
 444
 445                 if (got_fstoken)
 446                         lwkt_gettoken(&hmp->fs_token);
 447
 448                 if (error)
 449                         break;
 450                 hammer_stats_file_read += n;
 451         }
 452
 453         /*
 454          * XXX only update the atime if we had to get the MP lock.
 455          * XXX hack hack hack, fixme.
 456          */
 457         if (got_fstoken) {
 458                 if ((ip->flags & HAMMER_INODE_RO) == 0 &&
 459                     (ip->hmp->mp->mnt_flag & MNT_NOATIME) == 0) {
 460                         ip->ino_data.atime = trans.time;
 461                         hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
 462                 }
 463                 hammer_done_transaction(&trans);
 464                 lwkt_reltoken(&hmp->fs_token);
 465         }
 466         return (error);
 467 }
 468
 469 /*
 470  * hammer_vop_write { vp, uio, ioflag, cred }
 471  */
 472 static
 473 int
 474 hammer_vop_write(struct vop_write_args *ap)
 475 {
 476         struct hammer_transaction trans;
 477         struct hammer_inode *ip;
 478         hammer_mount_t hmp;
 479         struct uio *uio;
 480         int offset;
 481         off_t base_offset;
 482         struct buf *bp;
 483         int kflags;
 484         int error;
 485         int n;
 486         int flags;
 487         int seqcount;
 488         int bigwrite;
 489
 490         if (ap->a_vp->v_type != VREG)
 491                 return (EINVAL);
 492         ip = VTOI(ap->a_vp);
 493         hmp = ip->hmp;
 494         error = 0;
 495         kflags = 0;
 496         seqcount = ap->a_ioflag >> 16;
 497
 498         if (ip->flags & HAMMER_INODE_RO)
 499                 return (EROFS);
 500
 501         /*
 502          * Create a transaction to cover the operations we perform.
 503          */
 504         lwkt_gettoken(&hmp->fs_token);
 505         hammer_start_transaction(&trans, hmp);
 506         uio = ap->a_uio;
 507
 508         /*
 509          * Check append mode
 510          */
 511         if (ap->a_ioflag & IO_APPEND)
 512                 uio->uio_offset = ip->ino_data.size;
 513
 514         /*
 515          * Check for illegal write offsets.  Valid range is 0...2^63-1.
 516          *
 517          * NOTE: the base_off assignment is required to work around what
 518          * I consider to be a GCC-4 optimization bug.
 519          */
 520         if (uio->uio_offset < 0) {
 521                 hammer_done_transaction(&trans);
 522                 lwkt_reltoken(&hmp->fs_token);
 523                 return (EFBIG);
 524         }
 525         base_offset = uio->uio_offset + uio->uio_resid; /* work around gcc-4 */
 526         if (uio->uio_resid > 0 && base_offset <= uio->uio_offset) {
 527                 hammer_done_transaction(&trans);
 528                 lwkt_reltoken(&hmp->fs_token);
 529                 return (EFBIG);
 530         }
 531
 532         /*
 533          * If reading or writing a huge amount of data we have to break
 534          * atomicy and allow the operation to be interrupted by a signal
 535          * or it can DOS the machine.
 536          *
 537          * Preset redo_count so we stop generating REDOs earlier if the
 538          * limit is exceeded.
 539          */
 540         bigwrite = (uio->uio_resid > 100 * 1024 * 1024);
 541         if ((ip->flags & HAMMER_INODE_REDO) &&
 542             ip->redo_count < hammer_limit_redo) {
 543                 ip->redo_count += uio->uio_resid;
 544         }
 545
 546         /*
 547          * Access the data typically in HAMMER_BUFSIZE blocks via the
 548          * buffer cache, but HAMMER may use a variable block size based
 549          * on the offset.
 550          */
 551         while (uio->uio_resid > 0) {
 552                 int fixsize = 0;
 553                 int blksize;
 554                 int blkmask;
 555                 int trivial;
 556                 int endofblk;
 557                 off_t nsize;
 558
 559                 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_WRITE)) != 0)
 560                         break;
 561                 if (bigwrite && (error = hammer_signal_check(hmp)) != 0)
 562                         break;
 563
 564                 blksize = hammer_blocksize(uio->uio_offset);
 565
 566                 /*
 567                  * Do not allow HAMMER to blow out the buffer cache.  Very
 568                  * large UIOs can lockout other processes due to bwillwrite()
 569                  * mechanics.
 570                  *
 571                  * The hammer inode is not locked during these operations.
 572                  * The vnode is locked which can interfere with the pageout
 573                  * daemon for non-UIO_NOCOPY writes but should not interfere
 574                  * with the buffer cache.  Even so, we cannot afford to
 575                  * allow the pageout daemon to build up too many dirty buffer
 576                  * cache buffers.
 577                  *
 578                  * Only call this if we aren't being recursively called from
 579                  * a virtual disk device (vn), else we may deadlock.
 580                  */
 581                 if ((ap->a_ioflag & IO_RECURSE) == 0)
 582                         bwillwrite(blksize);
 583
 584                 /*
 585                  * Control the number of pending records associated with
 586                  * this inode.  If too many have accumulated start a
 587                  * flush.  Try to maintain a pipeline with the flusher.
 588                  *
 589                  * NOTE: It is possible for other sources to grow the
 590                  *       records but not necessarily issue another flush,
 591                  *       so use a timeout and ensure that a re-flush occurs.
 592                  */
 593                 if (ip->rsv_recs >= hammer_limit_inode_recs) {
 594                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 595                         while (ip->rsv_recs >= hammer_limit_inode_recs * 2) {
 596                                 ip->flags |= HAMMER_INODE_RECSW;
 597                                 tsleep(&ip->rsv_recs, 0, "hmrwww", hz);
 598                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 599                         }
 600                 }
 601
 602 #if 0
 603                 /*
 604                  * Do not allow HAMMER to blow out system memory by
 605                  * accumulating too many records.   Records are so well
 606                  * decoupled from the buffer cache that it is possible
 607                  * for userland to push data out to the media via
 608                  * direct-write, but build up the records queued to the
 609                  * backend faster then the backend can flush them out.
 610                  * HAMMER has hit its write limit but the frontend has
 611                  * no pushback to slow it down.
 612                  */
 613                 if (hmp->rsv_recs > hammer_limit_recs / 2) {
 614                         /*
 615                          * Get the inode on the flush list
 616                          */
 617                         if (ip->rsv_recs >= 64)
 618                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 619                         else if (ip->rsv_recs >= 16)
 620                                 hammer_flush_inode(ip, 0);
 621
 622                         /*
 623                          * Keep the flusher going if the system keeps
 624                          * queueing records.
 625                          */
 626                         delta = hmp->count_newrecords -
 627                                 hmp->last_newrecords;
 628                         if (delta < 0 || delta > hammer_limit_recs / 2) {
 629                                 hmp->last_newrecords = hmp->count_newrecords;
 630                                 hammer_sync_hmp(hmp, MNT_NOWAIT);
 631                         }
 632
 633                         /*
 634                          * If we have gotten behind start slowing
 635                          * down the writers.
 636                          */
 637                         delta = (hmp->rsv_recs - hammer_limit_recs) *
 638                                 hz / hammer_limit_recs;
 639                         if (delta > 0)
 640                                 tsleep(&trans, 0, "hmrslo", delta);
 641                 }
 642 #endif
 643
 644                 /*
 645                  * Calculate the blocksize at the current offset and figure
 646                  * out how much we can actually write.
 647                  */
 648                 blkmask = blksize - 1;
 649                 offset = (int)uio->uio_offset & blkmask;
 650                 base_offset = uio->uio_offset & ~(int64_t)blkmask;
 651                 n = blksize - offset;
 652                 if (n > uio->uio_resid) {
 653                         n = uio->uio_resid;
 654                         endofblk = 0;
 655                 } else {
 656                         endofblk = 1;
 657                 }
 658                 nsize = uio->uio_offset + n;
 659                 if (nsize > ip->ino_data.size) {
 660                         if (uio->uio_offset > ip->ino_data.size)
 661                                 trivial = 0;
 662                         else
 663                                 trivial = 1;
 664                         nvextendbuf(ap->a_vp,
 665                                     ip->ino_data.size,
 666                                     nsize,
 667                                     hammer_blocksize(ip->ino_data.size),
 668                                     hammer_blocksize(nsize),
 669                                     hammer_blockoff(ip->ino_data.size),
 670                                     hammer_blockoff(nsize),
 671                                     trivial);
 672                         fixsize = 1;
 673                         kflags |= NOTE_EXTEND;
 674                 }
 675
 676                 if (uio->uio_segflg == UIO_NOCOPY) {
 677                         /*
 678                          * Issuing a write with the same data backing the
 679                          * buffer.  Instantiate the buffer to collect the
 680                          * backing vm pages, then read-in any missing bits.
 681                          *
 682                          * This case is used by vop_stdputpages().
 683                          */
 684                         bp = getblk(ap->a_vp, base_offset,
 685                                     blksize, GETBLK_BHEAVY, 0);
 686                         if ((bp->b_flags & B_CACHE) == 0) {
 687                                 bqrelse(bp);
 688                                 error = bread(ap->a_vp, base_offset,
 689                                               blksize, &bp);
 690                         }
 691                 } else if (offset == 0 && uio->uio_resid >= blksize) {
 692                         /*
 693                          * Even though we are entirely overwriting the buffer
 694                          * we may still have to zero it out to avoid a
 695                          * mmap/write visibility issue.
 696                          */
 697                         bp = getblk(ap->a_vp, base_offset, blksize, GETBLK_BHEAVY, 0);
 698                         if ((bp->b_flags & B_CACHE) == 0)
 699                                 vfs_bio_clrbuf(bp);
 700                 } else if (base_offset >= ip->ino_data.size) {
 701                         /*
 702                          * If the base offset of the buffer is beyond the
 703                          * file EOF, we don't have to issue a read.
 704                          */
 705                         bp = getblk(ap->a_vp, base_offset,
 706                                     blksize, GETBLK_BHEAVY, 0);
 707                         vfs_bio_clrbuf(bp);
 708                 } else {
 709                         /*
 710                          * Partial overwrite, read in any missing bits then
 711                          * replace the portion being written.
 712                          */
 713                         error = bread(ap->a_vp, base_offset, blksize, &bp);
 714                         if (error == 0)
 715                                 bheavy(bp);
 716                 }
 717                 if (error == 0) {
 718                         lwkt_reltoken(&hmp->fs_token);
 719                         error = uiomove(bp->b_data + offset, n, uio);
 720                         lwkt_gettoken(&hmp->fs_token);
 721                 }
 722
 723                 /*
 724                  * Generate REDO records if enabled and redo_count will not
 725                  * exceeded the limit.
 726                  *
 727                  * If redo_count exceeds the limit we stop generating records
 728                  * and clear HAMMER_INODE_REDO.  This will cause the next
 729                  * fsync() to do a full meta-data sync instead of just an
 730                  * UNDO/REDO fifo update.
 731                  *
 732                  * When clearing HAMMER_INODE_REDO any pre-existing REDOs
 733                  * will still be tracked.  The tracks will be terminated
 734                  * when the related meta-data (including possible data
 735                  * modifications which are not tracked via REDO) is
 736                  * flushed.
 737                  */
 738                 if ((ip->flags & HAMMER_INODE_REDO) && error == 0) {
 739                         if (ip->redo_count < hammer_limit_redo) {
 740                                 bp->b_flags |= B_VFSFLAG1;
 741                                 error = hammer_generate_redo(&trans, ip,
 742                                                      base_offset + offset,
 743                                                      HAMMER_REDO_WRITE,
 744                                                      bp->b_data + offset,
 745                                                      (size_t)n);
 746                         } else {
 747                                 ip->flags &= ~HAMMER_INODE_REDO;
 748                         }
 749                 }
 750
 751                 /*
 752                  * If we screwed up we have to undo any VM size changes we
 753                  * made.
 754                  */
 755                 if (error) {
 756                         brelse(bp);
 757                         if (fixsize) {
 758                                 nvtruncbuf(ap->a_vp, ip->ino_data.size,
 759                                           hammer_blocksize(ip->ino_data.size),
 760                                           hammer_blockoff(ip->ino_data.size));
 761                         }
 762                         break;
 763                 }
 764                 kflags |= NOTE_WRITE;
 765                 hammer_stats_file_write += n;
 766                 /* bp->b_flags |= B_CLUSTEROK; temporarily disabled */
 767                 if (ip->ino_data.size < uio->uio_offset) {
 768                         ip->ino_data.size = uio->uio_offset;
 769                         flags = HAMMER_INODE_SDIRTY;
 770                 } else {
 771                         flags = 0;
 772                 }
 773                 ip->ino_data.mtime = trans.time;
 774                 flags |= HAMMER_INODE_MTIME | HAMMER_INODE_BUFS;
 775                 hammer_modify_inode(&trans, ip, flags);
 776
 777                 /*
 778                  * Once we dirty the buffer any cached zone-X offset
 779                  * becomes invalid.  HAMMER NOTE: no-history mode cannot
 780                  * allow overwriting over the same data sector unless
 781                  * we provide UNDOs for the old data, which we don't.
 782                  */
 783                 bp->b_bio2.bio_offset = NOOFFSET;
 784
 785                 /*
 786                  * Final buffer disposition.
 787                  *
 788                  * Because meta-data updates are deferred, HAMMER is
 789                  * especially sensitive to excessive bdwrite()s because
 790                  * the I/O stream is not broken up by disk reads.  So the
 791                  * buffer cache simply cannot keep up.
 792                  *
 793                  * WARNING!  blksize is variable.  cluster_write() is
 794                  *           expected to not blow up if it encounters
 795                  *           buffers that do not match the passed blksize.
 796                  *
 797                  * NOTE!  Hammer shouldn't need to bawrite()/cluster_write().
 798                  *        The ip->rsv_recs check should burst-flush the data.
 799                  *        If we queue it immediately the buf could be left
 800                  *        locked on the device queue for a very long time.
 801                  *
 802                  * NOTE!  To avoid degenerate stalls due to mismatched block
 803                  *        sizes we only honor IO_DIRECT on the write which
 804                  *        abuts the end of the buffer.  However, we must
 805                  *        honor IO_SYNC in case someone is silly enough to
 806                  *        configure a HAMMER file as swap, or when HAMMER
 807                  *        is serving NFS (for commits).  Ick ick.
 808                  */
 809                 bp->b_flags |= B_AGE;
 810                 if (ap->a_ioflag & IO_SYNC) {
 811                         bwrite(bp);
 812                 } else if ((ap->a_ioflag & IO_DIRECT) && endofblk) {
 813                         bawrite(bp);
 814                 } else {
 815 #if 0
 816                 if (offset + n == blksize) {
 817                         if (hammer_cluster_enable == 0 ||
 818                             (ap->a_vp->v_mount->mnt_flag & MNT_NOCLUSTERW)) {
 819                                 bawrite(bp);
 820                         } else {
 821                                 cluster_write(bp, ip->ino_data.size,
 822                                               blksize, seqcount);
 823                         }
 824                 } else {
 825 #endif
 826                         bdwrite(bp);
 827                 }
 828         }
 829         hammer_done_transaction(&trans);
 830         hammer_knote(ap->a_vp, kflags);
 831         lwkt_reltoken(&hmp->fs_token);
 832         return (error);
 833 }
 834
 835 /*
 836  * hammer_vop_access { vp, mode, cred }
 837  *
 838  * MPSAFE - does not require fs_token
 839  */
 840 static
 841 int
 842 hammer_vop_access(struct vop_access_args *ap)
 843 {
 844         struct hammer_inode *ip = VTOI(ap->a_vp);
 845         uid_t uid;
 846         gid_t gid;
 847         int error;
 848
 849         ++hammer_stats_file_iopsr;
 850         uid = hammer_to_unix_xid(&ip->ino_data.uid);
 851         gid = hammer_to_unix_xid(&ip->ino_data.gid);
 852
 853         error = vop_helper_access(ap, uid, gid, ip->ino_data.mode,
 854                                   ip->ino_data.uflags);
 855         return (error);
 856 }
 857
 858 /*
 859  * hammer_vop_advlock { vp, id, op, fl, flags }
 860  *
 861  * MPSAFE - does not require fs_token
 862  */
 863 static
 864 int
 865 hammer_vop_advlock(struct vop_advlock_args *ap)
 866 {
 867         hammer_inode_t ip = VTOI(ap->a_vp);
 868
 869         return (lf_advlock(ap, &ip->advlock, ip->ino_data.size));
 870 }
 871
 872 /*
 873  * hammer_vop_close { vp, fflag }
 874  *
 875  * We can only sync-on-close for normal closes.  XXX disabled for now.
 876  */
 877 static
 878 int
 879 hammer_vop_close(struct vop_close_args *ap)
 880 {
 881 #if 0
 882         struct vnode *vp = ap->a_vp;
 883         hammer_inode_t ip = VTOI(vp);
 884         int waitfor;
 885         if (ip->flags & (HAMMER_INODE_CLOSESYNC|HAMMER_INODE_CLOSEASYNC)) {
 886                 if (vn_islocked(vp) == LK_EXCLUSIVE &&
 887                     (vp->v_flag & (VINACTIVE|VRECLAIMED)) == 0) {
 888                         if (ip->flags & HAMMER_INODE_CLOSESYNC)
 889                                 waitfor = MNT_WAIT;
 890                         else
 891                                 waitfor = MNT_NOWAIT;
 892                         ip->flags &= ~(HAMMER_INODE_CLOSESYNC |
 893                                        HAMMER_INODE_CLOSEASYNC);
 894                         VOP_FSYNC(vp, MNT_NOWAIT, waitfor);
 895                 }
 896         }
 897 #endif
 898         return (vop_stdclose(ap));
 899 }
 900
 901 /*
 902  * hammer_vop_ncreate { nch, dvp, vpp, cred, vap }
 903  *
 904  * The operating system has already ensured that the directory entry
 905  * does not exist and done all appropriate namespace locking.
 906  */
 907 static
 908 int
 909 hammer_vop_ncreate(struct vop_ncreate_args *ap)
 910 {
 911         struct hammer_transaction trans;
 912         struct hammer_inode *dip;
 913         struct hammer_inode *nip;
 914         struct nchandle *nch;
 915         hammer_mount_t hmp;
 916         int error;
 917
 918         nch = ap->a_nch;
 919         dip = VTOI(ap->a_dvp);
 920         hmp = dip->hmp;
 921
 922         if (dip->flags & HAMMER_INODE_RO)
 923                 return (EROFS);
 924         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
 925                 return (error);
 926
 927         /*
 928          * Create a transaction to cover the operations we perform.
 929          */
 930         lwkt_gettoken(&hmp->fs_token);
 931         hammer_start_transaction(&trans, hmp);
 932         ++hammer_stats_file_iopsw;
 933
 934         /*
 935          * Create a new filesystem object of the requested type.  The
 936          * returned inode will be referenced and shared-locked to prevent
 937          * it from being moved to the flusher.
 938          */
 939         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
 940                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
 941                                     NULL, &nip);
 942         if (error) {
 943                 hkprintf("hammer_create_inode error %d\n", error);
 944                 hammer_done_transaction(&trans);
 945                 *ap->a_vpp = NULL;
 946                 lwkt_reltoken(&hmp->fs_token);
 947                 return (error);
 948         }
 949
 950         /*
 951          * Add the new filesystem object to the directory.  This will also
 952          * bump the inode's link count.
 953          */
 954         error = hammer_ip_add_directory(&trans, dip,
 955                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
 956                                         nip);
 957         if (error)
 958                 hkprintf("hammer_ip_add_directory error %d\n", error);
 959
 960         /*
 961          * Finish up.
 962          */
 963         if (error) {
 964                 hammer_rel_inode(nip, 0);
 965                 hammer_done_transaction(&trans);
 966                 *ap->a_vpp = NULL;
 967         } else {
 968                 error = hammer_get_vnode(nip, ap->a_vpp);
 969                 hammer_done_transaction(&trans);
 970                 hammer_rel_inode(nip, 0);
 971                 if (error == 0) {
 972                         cache_setunresolved(ap->a_nch);
 973                         cache_setvp(ap->a_nch, *ap->a_vpp);
 974                 }
 975                 hammer_knote(ap->a_dvp, NOTE_WRITE);
 976         }
 977         lwkt_reltoken(&hmp->fs_token);
 978         return (error);
 979 }
 980
 981 /*
 982  * hammer_vop_getattr { vp, vap }
 983  *
 984  * Retrieve an inode's attribute information.  When accessing inodes
 985  * historically we fake the atime field to ensure consistent results.
 986  * The atime field is stored in the B-Tree element and allowed to be
 987  * updated without cycling the element.
 988  *
 989  * MPSAFE - does not require fs_token
 990  */
 991 static
 992 int
 993 hammer_vop_getattr(struct vop_getattr_args *ap)
 994 {
 995         struct hammer_inode *ip = VTOI(ap->a_vp);
 996         struct vattr *vap = ap->a_vap;
 997
 998         /*
 999          * We want the fsid to be different when accessing a filesystem
1000          * with different as-of's so programs like diff don't think
1001          * the files are the same.
1002          *
1003          * We also want the fsid to be the same when comparing snapshots,
1004          * or when comparing mirrors (which might be backed by different
1005          * physical devices).  HAMMER fsids are based on the PFS's
1006          * shared_uuid field.
1007          *
1008          * XXX there is a chance of collision here.  The va_fsid reported
1009          * by stat is different from the more involved fsid used in the
1010          * mount structure.
1011          */
1012         ++hammer_stats_file_iopsr;
1013         hammer_lock_sh(&ip->lock);
1014         vap->va_fsid = ip->pfsm->fsid_udev ^ (u_int32_t)ip->obj_asof ^
1015                        (u_int32_t)(ip->obj_asof >> 32);
1016
1017         vap->va_fileid = ip->ino_leaf.base.obj_id;
1018         vap->va_mode = ip->ino_data.mode;
1019         vap->va_nlink = ip->ino_data.nlinks;
1020         vap->va_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1021         vap->va_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1022         vap->va_rmajor = 0;
1023         vap->va_rminor = 0;
1024         vap->va_size = ip->ino_data.size;
1025
1026         /*
1027          * Special case for @@PFS softlinks.  The actual size of the
1028          * expanded softlink is "@@0x%016llx:%05d" == 26 bytes.
1029          * or for MAX_TID is    "@@-1:%05d" == 10 bytes.
1030          */
1031         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_SOFTLINK &&
1032             ip->ino_data.size == 10 &&
1033             ip->obj_asof == HAMMER_MAX_TID &&
1034             ip->obj_localization == 0 &&
1035             strncmp(ip->ino_data.ext.symlink, "@@PFS", 5) == 0) {
1036                     if (ip->pfsm->pfsd.mirror_flags & HAMMER_PFSD_SLAVE)
1037                             vap->va_size = 26;
1038                     else
1039                             vap->va_size = 10;
1040         }
1041
1042         /*
1043          * We must provide a consistent atime and mtime for snapshots
1044          * so people can do a 'tar cf - ... | md5' on them and get
1045          * consistent results.
1046          */
1047         if (ip->flags & HAMMER_INODE_RO) {
1048                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_atime);
1049                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_mtime);
1050         } else {
1051                 hammer_time_to_timespec(ip->ino_data.atime, &vap->va_atime);
1052                 hammer_time_to_timespec(ip->ino_data.mtime, &vap->va_mtime);
1053         }
1054         hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_ctime);
1055         vap->va_flags = ip->ino_data.uflags;
1056         vap->va_gen = 1;        /* hammer inums are unique for all time */
1057         vap->va_blocksize = HAMMER_BUFSIZE;
1058         if (ip->ino_data.size >= HAMMER_XDEMARC) {
1059                 vap->va_bytes = (ip->ino_data.size + HAMMER_XBUFMASK64) &
1060                                 ~HAMMER_XBUFMASK64;
1061         } else if (ip->ino_data.size > HAMMER_BUFSIZE / 2) {
1062                 vap->va_bytes = (ip->ino_data.size + HAMMER_BUFMASK64) &
1063                                 ~HAMMER_BUFMASK64;
1064         } else {
1065                 vap->va_bytes = (ip->ino_data.size + 15) & ~15;
1066         }
1067
1068         vap->va_type = hammer_get_vnode_type(ip->ino_data.obj_type);
1069         vap->va_filerev = 0;    /* XXX */
1070         vap->va_uid_uuid = ip->ino_data.uid;
1071         vap->va_gid_uuid = ip->ino_data.gid;
1072         vap->va_fsid_uuid = ip->hmp->fsid;
1073         vap->va_vaflags = VA_UID_UUID_VALID | VA_GID_UUID_VALID |
1074                           VA_FSID_UUID_VALID;
1075
1076         switch (ip->ino_data.obj_type) {
1077         case HAMMER_OBJTYPE_CDEV:
1078         case HAMMER_OBJTYPE_BDEV:
1079                 vap->va_rmajor = ip->ino_data.rmajor;
1080                 vap->va_rminor = ip->ino_data.rminor;
1081                 break;
1082         default:
1083                 break;
1084         }
1085         hammer_unlock(&ip->lock);
1086         return(0);
1087 }
1088
1089 /*
1090  * hammer_vop_nresolve { nch, dvp, cred }
1091  *
1092  * Locate the requested directory entry.
1093  */
1094 static
1095 int
1096 hammer_vop_nresolve(struct vop_nresolve_args *ap)
1097 {
1098         struct hammer_transaction trans;
1099         struct namecache *ncp;
1100         hammer_mount_t hmp;
1101         hammer_inode_t dip;
1102         hammer_inode_t ip;
1103         hammer_tid_t asof;
1104         struct hammer_cursor cursor;
1105         struct vnode *vp;
1106         int64_t namekey;
1107         int error;
1108         int i;
1109         int nlen;
1110         int flags;
1111         int ispfs;
1112         int64_t obj_id;
1113         u_int32_t localization;
1114         u_int32_t max_iterations;
1115
1116         /*
1117          * Misc initialization, plus handle as-of name extensions.  Look for
1118          * the '@@' extension.  Note that as-of files and directories cannot
1119          * be modified.
1120          */
1121         dip = VTOI(ap->a_dvp);
1122         ncp = ap->a_nch->ncp;
1123         asof = dip->obj_asof;
1124         localization = dip->obj_localization;   /* for code consistency */
1125         nlen = ncp->nc_nlen;
1126         flags = dip->flags & HAMMER_INODE_RO;
1127         ispfs = 0;
1128         hmp = dip->hmp;
1129
1130         lwkt_gettoken(&hmp->fs_token);
1131         hammer_simple_transaction(&trans, hmp);
1132         ++hammer_stats_file_iopsr;
1133
1134         for (i = 0; i < nlen; ++i) {
1135                 if (ncp->nc_name[i] == '@' && ncp->nc_name[i+1] == '@') {
1136                         error = hammer_str_to_tid(ncp->nc_name + i + 2,
1137                                                   &ispfs, &asof, &localization);
1138                         if (error != 0) {
1139                                 i = nlen;
1140                                 break;
1141                         }
1142                         if (asof != HAMMER_MAX_TID)
1143                                 flags |= HAMMER_INODE_RO;
1144                         break;
1145                 }
1146         }
1147         nlen = i;
1148
1149         /*
1150          * If this is a PFS softlink we dive into the PFS
1151          */
1152         if (ispfs && nlen == 0) {
1153                 ip = hammer_get_inode(&trans, dip, HAMMER_OBJID_ROOT,
1154                                       asof, localization,
1155                                       flags, &error);
1156                 if (error == 0) {
1157                         error = hammer_get_vnode(ip, &vp);
1158                         hammer_rel_inode(ip, 0);
1159                 } else {
1160                         vp = NULL;
1161                 }
1162                 if (error == 0) {
1163                         vn_unlock(vp);
1164                         cache_setvp(ap->a_nch, vp);
1165                         vrele(vp);
1166                 }
1167                 goto done;
1168         }
1169
1170         /*
1171          * If there is no path component the time extension is relative to dip.
1172          * e.g. "fubar/@@<snapshot>"
1173          *
1174          * "." is handled by the kernel, but ".@@<snapshot>" is not.
1175          * e.g. "fubar/.@@<snapshot>"
1176          *
1177          * ".." is handled by the kernel.  We do not currently handle
1178          * "..@<snapshot>".
1179          */
1180         if (nlen == 0 || (nlen == 1 && ncp->nc_name[0] == '.')) {
1181                 ip = hammer_get_inode(&trans, dip, dip->obj_id,
1182                                       asof, dip->obj_localization,
1183                                       flags, &error);
1184                 if (error == 0) {
1185                         error = hammer_get_vnode(ip, &vp);
1186                         hammer_rel_inode(ip, 0);
1187                 } else {
1188                         vp = NULL;
1189                 }
1190                 if (error == 0) {
1191                         vn_unlock(vp);
1192                         cache_setvp(ap->a_nch, vp);
1193                         vrele(vp);
1194                 }
1195                 goto done;
1196         }
1197
1198         /*
1199          * Calculate the namekey and setup the key range for the scan.  This
1200          * works kinda like a chained hash table where the lower 32 bits
1201          * of the namekey synthesize the chain.
1202          *
1203          * The key range is inclusive of both key_beg and key_end.
1204          */
1205         namekey = hammer_directory_namekey(dip, ncp->nc_name, nlen,
1206                                            &max_iterations);
1207
1208         error = hammer_init_cursor(&trans, &cursor, &dip->cache[1], dip);
1209         cursor.key_beg.localization = dip->obj_localization +
1210                                       hammer_dir_localization(dip);
1211         cursor.key_beg.obj_id = dip->obj_id;
1212         cursor.key_beg.key = namekey;
1213         cursor.key_beg.create_tid = 0;
1214         cursor.key_beg.delete_tid = 0;
1215         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1216         cursor.key_beg.obj_type = 0;
1217
1218         cursor.key_end = cursor.key_beg;
1219         cursor.key_end.key += max_iterations;
1220         cursor.asof = asof;
1221         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1222
1223         /*
1224          * Scan all matching records (the chain), locate the one matching
1225          * the requested path component.
1226          *
1227          * The hammer_ip_*() functions merge in-memory records with on-disk
1228          * records for the purposes of the search.
1229          */
1230         obj_id = 0;
1231         localization = HAMMER_DEF_LOCALIZATION;
1232
1233         if (error == 0) {
1234                 error = hammer_ip_first(&cursor);
1235                 while (error == 0) {
1236                         error = hammer_ip_resolve_data(&cursor);
1237                         if (error)
1238                                 break;
1239                         if (nlen == cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF &&
1240                             bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1241                                 obj_id = cursor.data->entry.obj_id;
1242                                 localization = cursor.data->entry.localization;
1243                                 break;
1244                         }
1245                         error = hammer_ip_next(&cursor);
1246                 }
1247         }
1248         hammer_done_cursor(&cursor);
1249
1250         /*
1251          * Lookup the obj_id.  This should always succeed.  If it does not
1252          * the filesystem may be damaged and we return a dummy inode.
1253          */
1254         if (error == 0) {
1255                 ip = hammer_get_inode(&trans, dip, obj_id,
1256                                       asof, localization,
1257                                       flags, &error);
1258                 if (error == ENOENT) {
1259                         kprintf("HAMMER: WARNING: Missing "
1260                                 "inode for dirent \"%s\"\n"
1261                                 "\tobj_id = %016llx, asof=%016llx, lo=%08x\n",
1262                                 ncp->nc_name,
1263                                 (long long)obj_id, (long long)asof,
1264                                 localization);
1265                         error = 0;
1266                         ip = hammer_get_dummy_inode(&trans, dip, obj_id,
1267                                                     asof, localization,
1268                                                     flags, &error);
1269                 }
1270                 if (error == 0) {
1271                         error = hammer_get_vnode(ip, &vp);
1272                         hammer_rel_inode(ip, 0);
1273                 } else {
1274                         vp = NULL;
1275                 }
1276                 if (error == 0) {
1277                         vn_unlock(vp);
1278                         cache_setvp(ap->a_nch, vp);
1279                         vrele(vp);
1280                 }
1281         } else if (error == ENOENT) {
1282                 cache_setvp(ap->a_nch, NULL);
1283         }
1284 done:
1285         hammer_done_transaction(&trans);
1286         lwkt_reltoken(&hmp->fs_token);
1287         return (error);
1288 }
1289
1290 /*
1291  * hammer_vop_nlookupdotdot { dvp, vpp, cred }
1292  *
1293  * Locate the parent directory of a directory vnode.
1294  *
1295  * dvp is referenced but not locked.  *vpp must be returned referenced and
1296  * locked.  A parent_obj_id of 0 does not necessarily indicate that we are
1297  * at the root, instead it could indicate that the directory we were in was
1298  * removed.
1299  *
1300  * NOTE: as-of sequences are not linked into the directory structure.  If
1301  * we are at the root with a different asof then the mount point, reload
1302  * the same directory with the mount point's asof.   I'm not sure what this
1303  * will do to NFS.  We encode ASOF stamps in NFS file handles so it might not
1304  * get confused, but it hasn't been tested.
1305  */
1306 static
1307 int
1308 hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *ap)
1309 {
1310         struct hammer_transaction trans;
1311         struct hammer_inode *dip;
1312         struct hammer_inode *ip;
1313         hammer_mount_t hmp;
1314         int64_t parent_obj_id;
1315         u_int32_t parent_obj_localization;
1316         hammer_tid_t asof;
1317         int error;
1318
1319         dip = VTOI(ap->a_dvp);
1320         asof = dip->obj_asof;
1321         hmp = dip->hmp;
1322
1323         /*
1324          * Whos are parent?  This could be the root of a pseudo-filesystem
1325          * whos parent is in another localization domain.
1326          */
1327         lwkt_gettoken(&hmp->fs_token);
1328         parent_obj_id = dip->ino_data.parent_obj_id;
1329         if (dip->obj_id == HAMMER_OBJID_ROOT)
1330                 parent_obj_localization = dip->ino_data.ext.obj.parent_obj_localization;
1331         else
1332                 parent_obj_localization = dip->obj_localization;
1333
1334         if (parent_obj_id == 0) {
1335                 if (dip->obj_id == HAMMER_OBJID_ROOT &&
1336                    asof != hmp->asof) {
1337                         parent_obj_id = dip->obj_id;
1338                         asof = hmp->asof;
1339                         *ap->a_fakename = kmalloc(19, M_TEMP, M_WAITOK);
1340                         ksnprintf(*ap->a_fakename, 19, "0x%016llx",
1341                                   (long long)dip->obj_asof);
1342                 } else {
1343                         *ap->a_vpp = NULL;
1344                         lwkt_reltoken(&hmp->fs_token);
1345                         return ENOENT;
1346                 }
1347         }
1348
1349         hammer_simple_transaction(&trans, hmp);
1350         ++hammer_stats_file_iopsr;
1351
1352         ip = hammer_get_inode(&trans, dip, parent_obj_id,
1353                               asof, parent_obj_localization,
1354                               dip->flags, &error);
1355         if (ip) {
1356                 error = hammer_get_vnode(ip, ap->a_vpp);
1357                 hammer_rel_inode(ip, 0);
1358         } else {
1359                 *ap->a_vpp = NULL;
1360         }
1361         hammer_done_transaction(&trans);
1362         lwkt_reltoken(&hmp->fs_token);
1363         return (error);
1364 }
1365
1366 /*
1367  * hammer_vop_nlink { nch, dvp, vp, cred }
1368  */
1369 static
1370 int
1371 hammer_vop_nlink(struct vop_nlink_args *ap)
1372 {
1373         struct hammer_transaction trans;
1374         struct hammer_inode *dip;
1375         struct hammer_inode *ip;
1376         struct nchandle *nch;
1377         hammer_mount_t hmp;
1378         int error;
1379
1380         if (ap->a_dvp->v_mount != ap->a_vp->v_mount)
1381                 return(EXDEV);
1382
1383         nch = ap->a_nch;
1384         dip = VTOI(ap->a_dvp);
1385         ip = VTOI(ap->a_vp);
1386         hmp = dip->hmp;
1387
1388         if (dip->obj_localization != ip->obj_localization)
1389                 return(EXDEV);
1390
1391         if (dip->flags & HAMMER_INODE_RO)
1392                 return (EROFS);
1393         if (ip->flags & HAMMER_INODE_RO)
1394                 return (EROFS);
1395         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1396                 return (error);
1397
1398         /*
1399          * Create a transaction to cover the operations we perform.
1400          */
1401         lwkt_gettoken(&hmp->fs_token);
1402         hammer_start_transaction(&trans, hmp);
1403         ++hammer_stats_file_iopsw;
1404
1405         /*
1406          * Add the filesystem object to the directory.  Note that neither
1407          * dip nor ip are referenced or locked, but their vnodes are
1408          * referenced.  This function will bump the inode's link count.
1409          */
1410         error = hammer_ip_add_directory(&trans, dip,
1411                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1412                                         ip);
1413
1414         /*
1415          * Finish up.
1416          */
1417         if (error == 0) {
1418                 cache_setunresolved(nch);
1419                 cache_setvp(nch, ap->a_vp);
1420         }
1421         hammer_done_transaction(&trans);
1422         hammer_knote(ap->a_vp, NOTE_LINK);
1423         hammer_knote(ap->a_dvp, NOTE_WRITE);
1424         lwkt_reltoken(&hmp->fs_token);
1425         return (error);
1426 }
1427
1428 /*
1429  * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
1430  *
1431  * The operating system has already ensured that the directory entry
1432  * does not exist and done all appropriate namespace locking.
1433  */
1434 static
1435 int
1436 hammer_vop_nmkdir(struct vop_nmkdir_args *ap)
1437 {
1438         struct hammer_transaction trans;
1439         struct hammer_inode *dip;
1440         struct hammer_inode *nip;
1441         struct nchandle *nch;
1442         hammer_mount_t hmp;
1443         int error;
1444
1445         nch = ap->a_nch;
1446         dip = VTOI(ap->a_dvp);
1447         hmp = dip->hmp;
1448
1449         if (dip->flags & HAMMER_INODE_RO)
1450                 return (EROFS);
1451         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1452                 return (error);
1453
1454         /*
1455          * Create a transaction to cover the operations we perform.
1456          */
1457         lwkt_gettoken(&hmp->fs_token);
1458         hammer_start_transaction(&trans, hmp);
1459         ++hammer_stats_file_iopsw;
1460
1461         /*
1462          * Create a new filesystem object of the requested type.  The
1463          * returned inode will be referenced but not locked.
1464          */
1465         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1466                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1467                                     NULL, &nip);
1468         if (error) {
1469                 hkprintf("hammer_mkdir error %d\n", error);
1470                 hammer_done_transaction(&trans);
1471                 *ap->a_vpp = NULL;
1472                 lwkt_reltoken(&hmp->fs_token);
1473                 return (error);
1474         }
1475         /*
1476          * Add the new filesystem object to the directory.  This will also
1477          * bump the inode's link count.
1478          */
1479         error = hammer_ip_add_directory(&trans, dip,
1480                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1481                                         nip);
1482         if (error)
1483                 hkprintf("hammer_mkdir (add) error %d\n", error);
1484
1485         /*
1486          * Finish up.
1487          */
1488         if (error) {
1489                 hammer_rel_inode(nip, 0);
1490                 *ap->a_vpp = NULL;
1491         } else {
1492                 error = hammer_get_vnode(nip, ap->a_vpp);
1493                 hammer_rel_inode(nip, 0);
1494                 if (error == 0) {
1495                         cache_setunresolved(ap->a_nch);
1496                         cache_setvp(ap->a_nch, *ap->a_vpp);
1497                 }
1498         }
1499         hammer_done_transaction(&trans);
1500         if (error == 0)
1501                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
1502         lwkt_reltoken(&hmp->fs_token);
1503         return (error);
1504 }
1505
1506 /*
1507  * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1508  *
1509  * The operating system has already ensured that the directory entry
1510  * does not exist and done all appropriate namespace locking.
1511  */
1512 static
1513 int
1514 hammer_vop_nmknod(struct vop_nmknod_args *ap)
1515 {
1516         struct hammer_transaction trans;
1517         struct hammer_inode *dip;
1518         struct hammer_inode *nip;
1519         struct nchandle *nch;
1520         hammer_mount_t hmp;
1521         int error;
1522
1523         nch = ap->a_nch;
1524         dip = VTOI(ap->a_dvp);
1525         hmp = dip->hmp;
1526
1527         if (dip->flags & HAMMER_INODE_RO)
1528                 return (EROFS);
1529         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1530                 return (error);
1531
1532         /*
1533          * Create a transaction to cover the operations we perform.
1534          */
1535         lwkt_gettoken(&hmp->fs_token);
1536         hammer_start_transaction(&trans, hmp);
1537         ++hammer_stats_file_iopsw;
1538
1539         /*
1540          * Create a new filesystem object of the requested type.  The
1541          * returned inode will be referenced but not locked.
1542          *
1543          * If mknod specifies a directory a pseudo-fs is created.
1544          */
1545         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1546                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1547                                     NULL, &nip);
1548         if (error) {
1549                 hammer_done_transaction(&trans);
1550                 *ap->a_vpp = NULL;
1551                 lwkt_reltoken(&hmp->fs_token);
1552                 return (error);
1553         }
1554
1555         /*
1556          * Add the new filesystem object to the directory.  This will also
1557          * bump the inode's link count.
1558          */
1559         error = hammer_ip_add_directory(&trans, dip,
1560                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1561                                         nip);
1562
1563         /*
1564          * Finish up.
1565          */
1566         if (error) {
1567                 hammer_rel_inode(nip, 0);
1568                 *ap->a_vpp = NULL;
1569         } else {
1570                 error = hammer_get_vnode(nip, ap->a_vpp);
1571                 hammer_rel_inode(nip, 0);
1572                 if (error == 0) {
1573                         cache_setunresolved(ap->a_nch);
1574                         cache_setvp(ap->a_nch, *ap->a_vpp);
1575                 }
1576         }
1577         hammer_done_transaction(&trans);
1578         if (error == 0)
1579                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1580         lwkt_reltoken(&hmp->fs_token);
1581         return (error);
1582 }
1583
1584 /*
1585  * hammer_vop_open { vp, mode, cred, fp }
1586  *
1587  * MPSAFE (does not require fs_token)
1588  */
1589 static
1590 int
1591 hammer_vop_open(struct vop_open_args *ap)
1592 {
1593         hammer_inode_t ip;
1594
1595         ++hammer_stats_file_iopsr;
1596         ip = VTOI(ap->a_vp);
1597
1598         if ((ap->a_mode & FWRITE) && (ip->flags & HAMMER_INODE_RO))
1599                 return (EROFS);
1600         return(vop_stdopen(ap));
1601 }
1602
1603 /*
1604  * hammer_vop_print { vp }
1605  */
1606 static
1607 int
1608 hammer_vop_print(struct vop_print_args *ap)
1609 {
1610         return EOPNOTSUPP;
1611 }
1612
1613 /*
1614  * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1615  */
1616 static
1617 int
1618 hammer_vop_readdir(struct vop_readdir_args *ap)
1619 {
1620         struct hammer_transaction trans;
1621         struct hammer_cursor cursor;
1622         struct hammer_inode *ip;
1623         hammer_mount_t hmp;
1624         struct uio *uio;
1625         hammer_base_elm_t base;
1626         int error;
1627         int cookie_index;
1628         int ncookies;
1629         off_t *cookies;
1630         off_t saveoff;
1631         int r;
1632         int dtype;
1633
1634         ++hammer_stats_file_iopsr;
1635         ip = VTOI(ap->a_vp);
1636         uio = ap->a_uio;
1637         saveoff = uio->uio_offset;
1638         hmp = ip->hmp;
1639
1640         if (ap->a_ncookies) {
1641                 ncookies = uio->uio_resid / 16 + 1;
1642                 if (ncookies > 1024)
1643                         ncookies = 1024;
1644                 cookies = kmalloc(ncookies * sizeof(off_t), M_TEMP, M_WAITOK);
1645                 cookie_index = 0;
1646         } else {
1647                 ncookies = -1;
1648                 cookies = NULL;
1649                 cookie_index = 0;
1650         }
1651
1652         lwkt_gettoken(&hmp->fs_token);
1653         hammer_simple_transaction(&trans, hmp);
1654
1655         /*
1656          * Handle artificial entries
1657          *
1658          * It should be noted that the minimum value for a directory
1659          * hash key on-media is 0x0000000100000000, so we can use anything
1660          * less then that to represent our 'special' key space.
1661          */
1662         error = 0;
1663         if (saveoff == 0) {
1664                 r = vop_write_dirent(&error, uio, ip->obj_id, DT_DIR, 1, ".");
1665                 if (r)
1666                         goto done;
1667                 if (cookies)
1668                         cookies[cookie_index] = saveoff;
1669                 ++saveoff;
1670                 ++cookie_index;
1671                 if (cookie_index == ncookies)
1672                         goto done;
1673         }
1674         if (saveoff == 1) {
1675                 if (ip->ino_data.parent_obj_id) {
1676                         r = vop_write_dirent(&error, uio,
1677                                              ip->ino_data.parent_obj_id,
1678                                              DT_DIR, 2, "..");
1679                 } else {
1680                         r = vop_write_dirent(&error, uio,
1681                                              ip->obj_id, DT_DIR, 2, "..");
1682                 }
1683                 if (r)
1684                         goto done;
1685                 if (cookies)
1686                         cookies[cookie_index] = saveoff;
1687                 ++saveoff;
1688                 ++cookie_index;
1689                 if (cookie_index == ncookies)
1690                         goto done;
1691         }
1692
1693         /*
1694          * Key range (begin and end inclusive) to scan.  Directory keys
1695          * directly translate to a 64 bit 'seek' position.
1696          */
1697         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1698         cursor.key_beg.localization = ip->obj_localization +
1699                                       hammer_dir_localization(ip);
1700         cursor.key_beg.obj_id = ip->obj_id;
1701         cursor.key_beg.create_tid = 0;
1702         cursor.key_beg.delete_tid = 0;
1703         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1704         cursor.key_beg.obj_type = 0;
1705         cursor.key_beg.key = saveoff;
1706
1707         cursor.key_end = cursor.key_beg;
1708         cursor.key_end.key = HAMMER_MAX_KEY;
1709         cursor.asof = ip->obj_asof;
1710         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1711
1712         error = hammer_ip_first(&cursor);
1713
1714         while (error == 0) {
1715                 error = hammer_ip_resolve_data(&cursor);
1716                 if (error)
1717                         break;
1718                 base = &cursor.leaf->base;
1719                 saveoff = base->key;
1720                 KKASSERT(cursor.leaf->data_len > HAMMER_ENTRY_NAME_OFF);
1721
1722                 if (base->obj_id != ip->obj_id)
1723                         panic("readdir: bad record at %p", cursor.node);
1724
1725                 /*
1726                  * Convert pseudo-filesystems into softlinks
1727                  */
1728                 dtype = hammer_get_dtype(cursor.leaf->base.obj_type);
1729                 r = vop_write_dirent(
1730                              &error, uio, cursor.data->entry.obj_id,
1731                              dtype,
1732                              cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF ,
1733                              (void *)cursor.data->entry.name);
1734                 if (r)
1735                         break;
1736                 ++saveoff;
1737                 if (cookies)
1738                         cookies[cookie_index] = base->key;
1739                 ++cookie_index;
1740                 if (cookie_index == ncookies)
1741                         break;
1742                 error = hammer_ip_next(&cursor);
1743         }
1744         hammer_done_cursor(&cursor);
1745
1746 done:
1747         hammer_done_transaction(&trans);
1748
1749         if (ap->a_eofflag)
1750                 *ap->a_eofflag = (error == ENOENT);
1751         uio->uio_offset = saveoff;
1752         if (error && cookie_index == 0) {
1753                 if (error == ENOENT)
1754                         error = 0;
1755                 if (cookies) {
1756                         kfree(cookies, M_TEMP);
1757                         *ap->a_ncookies = 0;
1758                         *ap->a_cookies = NULL;
1759                 }
1760         } else {
1761                 if (error == ENOENT)
1762                         error = 0;
1763                 if (cookies) {
1764                         *ap->a_ncookies = cookie_index;
1765                         *ap->a_cookies = cookies;
1766                 }
1767         }
1768         lwkt_reltoken(&hmp->fs_token);
1769         return(error);
1770 }
1771
1772 /*
1773  * hammer_vop_readlink { vp, uio, cred }
1774  */
1775 static
1776 int
1777 hammer_vop_readlink(struct vop_readlink_args *ap)
1778 {
1779         struct hammer_transaction trans;
1780         struct hammer_cursor cursor;
1781         struct hammer_inode *ip;
1782         hammer_mount_t hmp;
1783         char buf[32];
1784         u_int32_t localization;
1785         hammer_pseudofs_inmem_t pfsm;
1786         int error;
1787
1788         ip = VTOI(ap->a_vp);
1789         hmp = ip->hmp;
1790
1791         lwkt_gettoken(&hmp->fs_token);
1792
1793         /*
1794          * Shortcut if the symlink data was stuffed into ino_data.
1795          *
1796          * Also expand special "@@PFS%05d" softlinks (expansion only
1797          * occurs for non-historical (current) accesses made from the
1798          * primary filesystem).
1799          */
1800         if (ip->ino_data.size <= HAMMER_INODE_BASESYMLEN) {
1801                 char *ptr;
1802                 int bytes;
1803
1804                 ptr = ip->ino_data.ext.symlink;
1805                 bytes = (int)ip->ino_data.size;
1806                 if (bytes == 10 &&
1807                     ip->obj_asof == HAMMER_MAX_TID &&
1808                     ip->obj_localization == 0 &&
1809                     strncmp(ptr, "@@PFS", 5) == 0) {
1810                         hammer_simple_transaction(&trans, hmp);
1811                         bcopy(ptr + 5, buf, 5);
1812                         buf[5] = 0;
1813                         localization = strtoul(buf, NULL, 10) << 16;
1814                         pfsm = hammer_load_pseudofs(&trans, localization,
1815                                                     &error);
1816                         if (error == 0) {
1817                                 if (pfsm->pfsd.mirror_flags &
1818                                     HAMMER_PFSD_SLAVE) {
1819                                         /* vap->va_size == 26 */
1820                                         ksnprintf(buf, sizeof(buf),
1821                                                   "@@0x%016llx:%05d",
1822                                                   (long long)pfsm->pfsd.sync_end_tid,
1823                                                   localization >> 16);
1824                                 } else {
1825                                         /* vap->va_size == 10 */
1826                                         ksnprintf(buf, sizeof(buf),
1827                                                   "@@-1:%05d",
1828                                                   localization >> 16);
1829 #if 0
1830                                         ksnprintf(buf, sizeof(buf),
1831                                                   "@@0x%016llx:%05d",
1832                                                   (long long)HAMMER_MAX_TID,
1833                                                   localization >> 16);
1834 #endif
1835                                 }
1836                                 ptr = buf;
1837                                 bytes = strlen(buf);
1838                         }
1839                         if (pfsm)
1840                                 hammer_rel_pseudofs(hmp, pfsm);
1841                         hammer_done_transaction(&trans);
1842                 }
1843                 error = uiomove(ptr, bytes, ap->a_uio);
1844                 lwkt_reltoken(&hmp->fs_token);
1845                 return(error);
1846         }
1847
1848         /*
1849          * Long version
1850          */
1851         hammer_simple_transaction(&trans, hmp);
1852         ++hammer_stats_file_iopsr;
1853         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1854
1855         /*
1856          * Key range (begin and end inclusive) to scan.  Directory keys
1857          * directly translate to a 64 bit 'seek' position.
1858          */
1859         cursor.key_beg.localization = ip->obj_localization +
1860                                       HAMMER_LOCALIZE_MISC;
1861         cursor.key_beg.obj_id = ip->obj_id;
1862         cursor.key_beg.create_tid = 0;
1863         cursor.key_beg.delete_tid = 0;
1864         cursor.key_beg.rec_type = HAMMER_RECTYPE_FIX;
1865         cursor.key_beg.obj_type = 0;
1866         cursor.key_beg.key = HAMMER_FIXKEY_SYMLINK;
1867         cursor.asof = ip->obj_asof;
1868         cursor.flags |= HAMMER_CURSOR_ASOF;
1869
1870         error = hammer_ip_lookup(&cursor);
1871         if (error == 0) {
1872                 error = hammer_ip_resolve_data(&cursor);
1873                 if (error == 0) {
1874                         KKASSERT(cursor.leaf->data_len >=
1875                                  HAMMER_SYMLINK_NAME_OFF);
1876                         error = uiomove(cursor.data->symlink.name,
1877                                         cursor.leaf->data_len -
1878                                                 HAMMER_SYMLINK_NAME_OFF,
1879                                         ap->a_uio);
1880                 }
1881         }
1882         hammer_done_cursor(&cursor);
1883         hammer_done_transaction(&trans);
1884         lwkt_reltoken(&hmp->fs_token);
1885         return(error);
1886 }
1887
1888 /*
1889  * hammer_vop_nremove { nch, dvp, cred }
1890  */
1891 static
1892 int
1893 hammer_vop_nremove(struct vop_nremove_args *ap)
1894 {
1895         struct hammer_transaction trans;
1896         struct hammer_inode *dip;
1897         hammer_mount_t hmp;
1898         int error;
1899
1900         dip = VTOI(ap->a_dvp);
1901         hmp = dip->hmp;
1902
1903         if (hammer_nohistory(dip) == 0 &&
1904             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1905                 return (error);
1906         }
1907
1908         lwkt_gettoken(&hmp->fs_token);
1909         hammer_start_transaction(&trans, hmp);
1910         ++hammer_stats_file_iopsw;
1911         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 0);
1912         hammer_done_transaction(&trans);
1913         if (error == 0)
1914                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1915         lwkt_reltoken(&hmp->fs_token);
1916         return (error);
1917 }
1918
1919 /*
1920  * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1921  */
1922 static
1923 int
1924 hammer_vop_nrename(struct vop_nrename_args *ap)
1925 {
1926         struct hammer_transaction trans;
1927         struct namecache *fncp;
1928         struct namecache *tncp;
1929         struct hammer_inode *fdip;
1930         struct hammer_inode *tdip;
1931         struct hammer_inode *ip;
1932         hammer_mount_t hmp;
1933         struct hammer_cursor cursor;
1934         int64_t namekey;
1935         u_int32_t max_iterations;
1936         int nlen, error;
1937
1938         if (ap->a_fdvp->v_mount != ap->a_tdvp->v_mount)
1939                 return(EXDEV);
1940         if (ap->a_fdvp->v_mount != ap->a_fnch->ncp->nc_vp->v_mount)
1941                 return(EXDEV);
1942
1943         fdip = VTOI(ap->a_fdvp);
1944         tdip = VTOI(ap->a_tdvp);
1945         fncp = ap->a_fnch->ncp;
1946         tncp = ap->a_tnch->ncp;
1947         ip = VTOI(fncp->nc_vp);
1948         KKASSERT(ip != NULL);
1949
1950         hmp = ip->hmp;
1951
1952         if (fdip->obj_localization != tdip->obj_localization)
1953                 return(EXDEV);
1954         if (fdip->obj_localization != ip->obj_localization)
1955                 return(EXDEV);
1956
1957         if (fdip->flags & HAMMER_INODE_RO)
1958                 return (EROFS);
1959         if (tdip->flags & HAMMER_INODE_RO)
1960                 return (EROFS);
1961         if (ip->flags & HAMMER_INODE_RO)
1962                 return (EROFS);
1963         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1964                 return (error);
1965
1966         lwkt_gettoken(&hmp->fs_token);
1967         hammer_start_transaction(&trans, hmp);
1968         ++hammer_stats_file_iopsw;
1969
1970         /*
1971          * Remove tncp from the target directory and then link ip as
1972          * tncp. XXX pass trans to dounlink
1973          *
1974          * Force the inode sync-time to match the transaction so it is
1975          * in-sync with the creation of the target directory entry.
1976          */
1977         error = hammer_dounlink(&trans, ap->a_tnch, ap->a_tdvp,
1978                                 ap->a_cred, 0, -1);
1979         if (error == 0 || error == ENOENT) {
1980                 error = hammer_ip_add_directory(&trans, tdip,
1981                                                 tncp->nc_name, tncp->nc_nlen,
1982                                                 ip);
1983                 if (error == 0) {
1984                         ip->ino_data.parent_obj_id = tdip->obj_id;
1985                         ip->ino_data.ctime = trans.time;
1986                         hammer_modify_inode(&trans, ip, HAMMER_INODE_DDIRTY);
1987                 }
1988         }
1989         if (error)
1990                 goto failed; /* XXX */
1991
1992         /*
1993          * Locate the record in the originating directory and remove it.
1994          *
1995          * Calculate the namekey and setup the key range for the scan.  This
1996          * works kinda like a chained hash table where the lower 32 bits
1997          * of the namekey synthesize the chain.
1998          *
1999          * The key range is inclusive of both key_beg and key_end.
2000          */
2001         namekey = hammer_directory_namekey(fdip, fncp->nc_name, fncp->nc_nlen,
2002                                            &max_iterations);
2003 retry:
2004         hammer_init_cursor(&trans, &cursor, &fdip->cache[1], fdip);
2005         cursor.key_beg.localization = fdip->obj_localization +
2006                                       hammer_dir_localization(fdip);
2007         cursor.key_beg.obj_id = fdip->obj_id;
2008         cursor.key_beg.key = namekey;
2009         cursor.key_beg.create_tid = 0;
2010         cursor.key_beg.delete_tid = 0;
2011         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
2012         cursor.key_beg.obj_type = 0;
2013
2014         cursor.key_end = cursor.key_beg;
2015         cursor.key_end.key += max_iterations;
2016         cursor.asof = fdip->obj_asof;
2017         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2018
2019         /*
2020          * Scan all matching records (the chain), locate the one matching
2021          * the requested path component.
2022          *
2023          * The hammer_ip_*() functions merge in-memory records with on-disk
2024          * records for the purposes of the search.
2025          */
2026         error = hammer_ip_first(&cursor);
2027         while (error == 0) {
2028                 if (hammer_ip_resolve_data(&cursor) != 0)
2029                         break;
2030                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
2031                 KKASSERT(nlen > 0);
2032                 if (fncp->nc_nlen == nlen &&
2033                     bcmp(fncp->nc_name, cursor.data->entry.name, nlen) == 0) {
2034                         break;
2035                 }
2036                 error = hammer_ip_next(&cursor);
2037         }
2038
2039         /*
2040          * If all is ok we have to get the inode so we can adjust nlinks.
2041          *
2042          * WARNING: hammer_ip_del_directory() may have to terminate the
2043          * cursor to avoid a recursion.  It's ok to call hammer_done_cursor()
2044          * twice.
2045          */
2046         if (error == 0)
2047                 error = hammer_ip_del_directory(&trans, &cursor, fdip, ip);
2048
2049         /*
2050          * XXX A deadlock here will break rename's atomicy for the purposes
2051          * of crash recovery.
2052          */
2053         if (error == EDEADLK) {
2054                 hammer_done_cursor(&cursor);
2055                 goto retry;
2056         }
2057
2058         /*
2059          * Cleanup and tell the kernel that the rename succeeded.
2060          *
2061          * NOTE: ip->vp, if non-NULL, cannot be directly referenced
2062          *       without formally acquiring the vp since the vp might
2063          *       have zero refs on it, or in the middle of a reclaim,
2064          *       etc.
2065          */
2066         hammer_done_cursor(&cursor);
2067         if (error == 0) {
2068                 cache_rename(ap->a_fnch, ap->a_tnch);
2069                 hammer_knote(ap->a_fdvp, NOTE_WRITE);
2070                 hammer_knote(ap->a_tdvp, NOTE_WRITE);
2071                 while (ip->vp) {
2072                         struct vnode *vp;
2073
2074                         error = hammer_get_vnode(ip, &vp);
2075                         if (error == 0 && vp) {
2076                                 vn_unlock(vp);
2077                                 hammer_knote(ip->vp, NOTE_RENAME);
2078                                 vrele(vp);
2079                                 break;
2080                         }
2081                         kprintf("Debug: HAMMER ip/vp race2 avoided\n");
2082                 }
2083         }
2084
2085 failed:
2086         hammer_done_transaction(&trans);
2087         lwkt_reltoken(&hmp->fs_token);
2088         return (error);
2089 }
2090
2091 /*
2092  * hammer_vop_nrmdir { nch, dvp, cred }
2093  */
2094 static
2095 int
2096 hammer_vop_nrmdir(struct vop_nrmdir_args *ap)
2097 {
2098         struct hammer_transaction trans;
2099         struct hammer_inode *dip;
2100         hammer_mount_t hmp;
2101         int error;
2102
2103         dip = VTOI(ap->a_dvp);
2104         hmp = dip->hmp;
2105
2106         if (hammer_nohistory(dip) == 0 &&
2107             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2108                 return (error);
2109         }
2110
2111         lwkt_gettoken(&hmp->fs_token);
2112         hammer_start_transaction(&trans, hmp);
2113         ++hammer_stats_file_iopsw;
2114         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 1);
2115         hammer_done_transaction(&trans);
2116         if (error == 0)
2117                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
2118         lwkt_reltoken(&hmp->fs_token);
2119         return (error);
2120 }
2121
2122 /*
2123  * hammer_vop_markatime { vp, cred }
2124  */
2125 static
2126 int
2127 hammer_vop_markatime(struct vop_markatime_args *ap)
2128 {
2129         struct hammer_transaction trans;
2130         struct hammer_inode *ip;
2131         hammer_mount_t hmp;
2132
2133         ip = VTOI(ap->a_vp);
2134         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2135                 return (EROFS);
2136         if (ip->flags & HAMMER_INODE_RO)
2137                 return (EROFS);
2138         hmp = ip->hmp;
2139         if (hmp->mp->mnt_flag & MNT_NOATIME)
2140                 return (0);
2141         lwkt_gettoken(&hmp->fs_token);
2142         hammer_start_transaction(&trans, hmp);
2143         ++hammer_stats_file_iopsw;
2144
2145         ip->ino_data.atime = trans.time;
2146         hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
2147         hammer_done_transaction(&trans);
2148         hammer_knote(ap->a_vp, NOTE_ATTRIB);
2149         lwkt_reltoken(&hmp->fs_token);
2150         return (0);
2151 }
2152
2153 /*
2154  * hammer_vop_setattr { vp, vap, cred }
2155  */
2156 static
2157 int
2158 hammer_vop_setattr(struct vop_setattr_args *ap)
2159 {
2160         struct hammer_transaction trans;
2161         struct hammer_inode *ip;
2162         struct vattr *vap;
2163         hammer_mount_t hmp;
2164         int modflags;
2165         int error;
2166         int truncating;
2167         int blksize;
2168         int kflags;
2169 #if 0
2170         int64_t aligned_size;
2171 #endif
2172         u_int32_t flags;
2173
2174         vap = ap->a_vap;
2175         ip = ap->a_vp->v_data;
2176         modflags = 0;
2177         kflags = 0;
2178         hmp = ip->hmp;
2179
2180         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2181                 return(EROFS);
2182         if (ip->flags & HAMMER_INODE_RO)
2183                 return (EROFS);
2184         if (hammer_nohistory(ip) == 0 &&
2185             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2186                 return (error);
2187         }
2188
2189         lwkt_gettoken(&hmp->fs_token);
2190         hammer_start_transaction(&trans, hmp);
2191         ++hammer_stats_file_iopsw;
2192         error = 0;
2193
2194         if (vap->va_flags != VNOVAL) {
2195                 flags = ip->ino_data.uflags;
2196                 error = vop_helper_setattr_flags(&flags, vap->va_flags,
2197                                          hammer_to_unix_xid(&ip->ino_data.uid),
2198                                          ap->a_cred);
2199                 if (error == 0) {
2200                         if (ip->ino_data.uflags != flags) {
2201                                 ip->ino_data.uflags = flags;
2202                                 ip->ino_data.ctime = trans.time;
2203                                 modflags |= HAMMER_INODE_DDIRTY;
2204                                 kflags |= NOTE_ATTRIB;
2205                         }
2206                         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2207                                 error = 0;
2208                                 goto done;
2209                         }
2210                 }
2211                 goto done;
2212         }
2213         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2214                 error = EPERM;
2215                 goto done;
2216         }
2217         if (vap->va_uid != (uid_t)VNOVAL || vap->va_gid != (gid_t)VNOVAL) {
2218                 mode_t cur_mode = ip->ino_data.mode;
2219                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2220                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2221                 uuid_t uuid_uid;
2222                 uuid_t uuid_gid;
2223
2224                 error = vop_helper_chown(ap->a_vp, vap->va_uid, vap->va_gid,
2225                                          ap->a_cred,
2226                                          &cur_uid, &cur_gid, &cur_mode);
2227                 if (error == 0) {
2228                         hammer_guid_to_uuid(&uuid_uid, cur_uid);
2229                         hammer_guid_to_uuid(&uuid_gid, cur_gid);
2230                         if (bcmp(&uuid_uid, &ip->ino_data.uid,
2231                                  sizeof(uuid_uid)) ||
2232                             bcmp(&uuid_gid, &ip->ino_data.gid,
2233                                  sizeof(uuid_gid)) ||
2234                             ip->ino_data.mode != cur_mode
2235                         ) {
2236                                 ip->ino_data.uid = uuid_uid;
2237                                 ip->ino_data.gid = uuid_gid;
2238                                 ip->ino_data.mode = cur_mode;
2239                                 ip->ino_data.ctime = trans.time;
2240                                 modflags |= HAMMER_INODE_DDIRTY;
2241                         }
2242                         kflags |= NOTE_ATTRIB;
2243                 }
2244         }
2245         while (vap->va_size != VNOVAL && ip->ino_data.size != vap->va_size) {
2246                 switch(ap->a_vp->v_type) {
2247                 case VREG:
2248                         if (vap->va_size == ip->ino_data.size)
2249                                 break;
2250
2251                         /*
2252                          * Log the operation if in fast-fsync mode or if
2253                          * there are unterminated redo write records present.
2254                          *
2255                          * The second check is needed so the recovery code
2256                          * properly truncates write redos even if nominal
2257                          * REDO operations is turned off due to excessive
2258                          * writes, because the related records might be
2259                          * destroyed and never lay down a TERM_WRITE.
2260                          */
2261                         if ((ip->flags & HAMMER_INODE_REDO) ||
2262                             (ip->flags & HAMMER_INODE_RDIRTY)) {
2263                                 error = hammer_generate_redo(&trans, ip,
2264                                                              vap->va_size,
2265                                                              HAMMER_REDO_TRUNC,
2266                                                              NULL, 0);
2267                         }
2268                         blksize = hammer_blocksize(vap->va_size);
2269
2270                         /*
2271                          * XXX break atomicy, we can deadlock the backend
2272                          * if we do not release the lock.  Probably not a
2273                          * big deal here.
2274                          */
2275                         if (vap->va_size < ip->ino_data.size) {
2276                                 nvtruncbuf(ap->a_vp, vap->va_size,
2277                                            blksize,
2278                                            hammer_blockoff(vap->va_size));
2279                                 truncating = 1;
2280                                 kflags |= NOTE_WRITE;
2281                         } else {
2282                                 nvextendbuf(ap->a_vp,
2283                                             ip->ino_data.size,
2284                                             vap->va_size,
2285                                             hammer_blocksize(ip->ino_data.size),
2286                                             hammer_blocksize(vap->va_size),
2287                                             hammer_blockoff(ip->ino_data.size),
2288                                             hammer_blockoff(vap->va_size),
2289                                             0);
2290                                 truncating = 0;
2291                                 kflags |= NOTE_WRITE | NOTE_EXTEND;
2292                         }
2293                         ip->ino_data.size = vap->va_size;
2294                         ip->ino_data.mtime = trans.time;
2295                         /* XXX safe to use SDIRTY instead of DDIRTY here? */
2296                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2297
2298                         /*
2299                          * On-media truncation is cached in the inode until
2300                          * the inode is synchronized.  We must immediately
2301                          * handle any frontend records.
2302                          */
2303                         if (truncating) {
2304                                 hammer_ip_frontend_trunc(ip, vap->va_size);
2305 #ifdef DEBUG_TRUNCATE
2306                                 if (HammerTruncIp == NULL)
2307                                         HammerTruncIp = ip;
2308 #endif
2309                                 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2310                                         ip->flags |= HAMMER_INODE_TRUNCATED;
2311                                         ip->trunc_off = vap->va_size;
2312 #ifdef DEBUG_TRUNCATE
2313                                         if (ip == HammerTruncIp)
2314                                         kprintf("truncate1 %016llx\n",
2315                                                 (long long)ip->trunc_off);
2316 #endif
2317                                 } else if (ip->trunc_off > vap->va_size) {
2318                                         ip->trunc_off = vap->va_size;
2319 #ifdef DEBUG_TRUNCATE
2320                                         if (ip == HammerTruncIp)
2321                                         kprintf("truncate2 %016llx\n",
2322                                                 (long long)ip->trunc_off);
2323 #endif
2324                                 } else {
2325 #ifdef DEBUG_TRUNCATE
2326                                         if (ip == HammerTruncIp)
2327                                         kprintf("truncate3 %016llx (ignored)\n",
2328                                                 (long long)vap->va_size);
2329 #endif
2330                                 }
2331                         }
2332
2333 #if 0
2334                         /*
2335                          * When truncating, nvtruncbuf() may have cleaned out
2336                          * a portion of the last block on-disk in the buffer
2337                          * cache.  We must clean out any frontend records
2338                          * for blocks beyond the new last block.
2339                          */
2340                         aligned_size = (vap->va_size + (blksize - 1)) &
2341                                        ~(int64_t)(blksize - 1);
2342                         if (truncating && vap->va_size < aligned_size) {
2343                                 aligned_size -= blksize;
2344                                 hammer_ip_frontend_trunc(ip, aligned_size);
2345                         }
2346 #endif
2347                         break;
2348                 case VDATABASE:
2349                         if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2350                                 ip->flags |= HAMMER_INODE_TRUNCATED;
2351                                 ip->trunc_off = vap->va_size;
2352                         } else if (ip->trunc_off > vap->va_size) {
2353                                 ip->trunc_off = vap->va_size;
2354                         }
2355                         hammer_ip_frontend_trunc(ip, vap->va_size);
2356                         ip->ino_data.size = vap->va_size;
2357                         ip->ino_data.mtime = trans.time;
2358                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2359                         kflags |= NOTE_ATTRIB;
2360                         break;
2361                 default:
2362                         error = EINVAL;
2363                         goto done;
2364                 }
2365                 break;
2366         }
2367         if (vap->va_atime.tv_sec != VNOVAL) {
2368                 ip->ino_data.atime = hammer_timespec_to_time(&vap->va_atime);
2369                 modflags |= HAMMER_INODE_ATIME;
2370                 kflags |= NOTE_ATTRIB;
2371         }
2372         if (vap->va_mtime.tv_sec != VNOVAL) {
2373                 ip->ino_data.mtime = hammer_timespec_to_time(&vap->va_mtime);
2374                 modflags |= HAMMER_INODE_MTIME;
2375                 kflags |= NOTE_ATTRIB;
2376         }
2377         if (vap->va_mode != (mode_t)VNOVAL) {
2378                 mode_t   cur_mode = ip->ino_data.mode;
2379                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2380                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2381
2382                 error = vop_helper_chmod(ap->a_vp, vap->va_mode, ap->a_cred,
2383                                          cur_uid, cur_gid, &cur_mode);
2384                 if (error == 0 && ip->ino_data.mode != cur_mode) {
2385                         ip->ino_data.mode = cur_mode;
2386                         ip->ino_data.ctime = trans.time;
2387                         modflags |= HAMMER_INODE_DDIRTY;
2388                         kflags |= NOTE_ATTRIB;
2389                 }
2390         }
2391 done:
2392         if (error == 0)
2393                 hammer_modify_inode(&trans, ip, modflags);
2394         hammer_done_transaction(&trans);
2395         hammer_knote(ap->a_vp, kflags);
2396         lwkt_reltoken(&hmp->fs_token);
2397         return (error);
2398 }
2399
2400 /*
2401  * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
2402  */
2403 static
2404 int
2405 hammer_vop_nsymlink(struct vop_nsymlink_args *ap)
2406 {
2407         struct hammer_transaction trans;
2408         struct hammer_inode *dip;
2409         struct hammer_inode *nip;
2410         hammer_record_t record;
2411         struct nchandle *nch;
2412         hammer_mount_t hmp;
2413         int error;
2414         int bytes;
2415
2416         ap->a_vap->va_type = VLNK;
2417
2418         nch = ap->a_nch;
2419         dip = VTOI(ap->a_dvp);
2420         hmp = dip->hmp;
2421
2422         if (dip->flags & HAMMER_INODE_RO)
2423                 return (EROFS);
2424         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
2425                 return (error);
2426
2427         /*
2428          * Create a transaction to cover the operations we perform.
2429          */
2430         lwkt_gettoken(&hmp->fs_token);
2431         hammer_start_transaction(&trans, hmp);
2432         ++hammer_stats_file_iopsw;
2433
2434         /*
2435          * Create a new filesystem object of the requested type.  The
2436          * returned inode will be referenced but not locked.
2437          */
2438
2439         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
2440                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
2441                                     NULL, &nip);
2442         if (error) {
2443                 hammer_done_transaction(&trans);
2444                 *ap->a_vpp = NULL;
2445                 lwkt_reltoken(&hmp->fs_token);
2446                 return (error);
2447         }
2448
2449         /*
2450          * Add a record representing the symlink.  symlink stores the link
2451          * as pure data, not a string, and is no \0 terminated.
2452          */
2453         if (error == 0) {
2454                 bytes = strlen(ap->a_target);
2455
2456                 if (bytes <= HAMMER_INODE_BASESYMLEN) {
2457                         bcopy(ap->a_target, nip->ino_data.ext.symlink, bytes);
2458                 } else {
2459                         record = hammer_alloc_mem_record(nip, bytes);
2460                         record->type = HAMMER_MEM_RECORD_GENERAL;
2461
2462                         record->leaf.base.localization = nip->obj_localization +
2463                                                          HAMMER_LOCALIZE_MISC;
2464                         record->leaf.base.key = HAMMER_FIXKEY_SYMLINK;
2465                         record->leaf.base.rec_type = HAMMER_RECTYPE_FIX;
2466                         record->leaf.data_len = bytes;
2467                         KKASSERT(HAMMER_SYMLINK_NAME_OFF == 0);
2468                         bcopy(ap->a_target, record->data->symlink.name, bytes);
2469                         error = hammer_ip_add_record(&trans, record);
2470                 }
2471
2472                 /*
2473                  * Set the file size to the length of the link.
2474                  */
2475                 if (error == 0) {
2476                         nip->ino_data.size = bytes;
2477                         hammer_modify_inode(&trans, nip, HAMMER_INODE_DDIRTY);
2478                 }
2479         }
2480         if (error == 0)
2481                 error = hammer_ip_add_directory(&trans, dip, nch->ncp->nc_name,
2482                                                 nch->ncp->nc_nlen, nip);
2483
2484         /*
2485          * Finish up.
2486          */
2487         if (error) {
2488                 hammer_rel_inode(nip, 0);
2489                 *ap->a_vpp = NULL;
2490         } else {
2491                 error = hammer_get_vnode(nip, ap->a_vpp);
2492                 hammer_rel_inode(nip, 0);
2493                 if (error == 0) {
2494                         cache_setunresolved(ap->a_nch);
2495                         cache_setvp(ap->a_nch, *ap->a_vpp);
2496                         hammer_knote(ap->a_dvp, NOTE_WRITE);
2497                 }
2498         }
2499         hammer_done_transaction(&trans);
2500         lwkt_reltoken(&hmp->fs_token);
2501         return (error);
2502 }
2503
2504 /*
2505  * hammer_vop_nwhiteout { nch, dvp, cred, flags }
2506  */
2507 static
2508 int
2509 hammer_vop_nwhiteout(struct vop_nwhiteout_args *ap)
2510 {
2511         struct hammer_transaction trans;
2512         struct hammer_inode *dip;
2513         hammer_mount_t hmp;
2514         int error;
2515
2516         dip = VTOI(ap->a_dvp);
2517         hmp = dip->hmp;
2518
2519         if (hammer_nohistory(dip) == 0 &&
2520             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0) {
2521                 return (error);
2522         }
2523
2524         lwkt_gettoken(&hmp->fs_token);
2525         hammer_start_transaction(&trans, hmp);
2526         ++hammer_stats_file_iopsw;
2527         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp,
2528                                 ap->a_cred, ap->a_flags, -1);
2529         hammer_done_transaction(&trans);
2530         lwkt_reltoken(&hmp->fs_token);
2531
2532         return (error);
2533 }
2534
2535 /*
2536  * hammer_vop_ioctl { vp, command, data, fflag, cred }
2537  */
2538 static
2539 int
2540 hammer_vop_ioctl(struct vop_ioctl_args *ap)
2541 {
2542         struct hammer_inode *ip = ap->a_vp->v_data;
2543         hammer_mount_t hmp = ip->hmp;
2544         int error;
2545
2546         ++hammer_stats_file_iopsr;
2547         lwkt_gettoken(&hmp->fs_token);
2548         error = hammer_ioctl(ip, ap->a_command, ap->a_data,
2549                              ap->a_fflag, ap->a_cred);
2550         lwkt_reltoken(&hmp->fs_token);
2551         return (error);
2552 }
2553
2554 static
2555 int
2556 hammer_vop_mountctl(struct vop_mountctl_args *ap)
2557 {
2558         static const struct mountctl_opt extraopt[] = {
2559                 { HMNT_NOHISTORY,       "nohistory" },
2560                 { HMNT_MASTERID,        "master" },
2561                 { 0, NULL}
2562
2563         };
2564         struct hammer_mount *hmp;
2565         struct mount *mp;
2566         int usedbytes;
2567         int error;
2568
2569         error = 0;
2570         usedbytes = 0;
2571         mp = ap->a_head.a_ops->head.vv_mount;
2572         KKASSERT(mp->mnt_data != NULL);
2573         hmp = (struct hammer_mount *)mp->mnt_data;
2574
2575         lwkt_gettoken(&hmp->fs_token);
2576
2577         switch(ap->a_op) {
2578         case MOUNTCTL_SET_EXPORT:
2579                 if (ap->a_ctllen != sizeof(struct export_args))
2580                         error = EINVAL;
2581                 else
2582                         error = hammer_vfs_export(mp, ap->a_op,
2583                                       (const struct export_args *)ap->a_ctl);
2584                 break;
2585         case MOUNTCTL_MOUNTFLAGS:
2586         {
2587                 /*
2588                  * Call standard mountctl VOP function
2589                  * so we get user mount flags.
2590                  */
2591                 error = vop_stdmountctl(ap);
2592                 if (error)
2593                         break;
2594
2595                 usedbytes = *ap->a_res;
2596
2597                 if (usedbytes > 0 && usedbytes < ap->a_buflen) {
2598                         usedbytes += vfs_flagstostr(hmp->hflags, extraopt,
2599                                                     ap->a_buf,
2600                                                     ap->a_buflen - usedbytes,
2601                                                     &error);
2602                 }
2603
2604                 *ap->a_res += usedbytes;
2605                 break;
2606         }
2607         default:
2608                 error = vop_stdmountctl(ap);
2609                 break;
2610         }
2611         lwkt_reltoken(&hmp->fs_token);
2612         return(error);
2613 }
2614
2615 /*
2616  * hammer_vop_strategy { vp, bio }
2617  *
2618  * Strategy call, used for regular file read & write only.  Note that the
2619  * bp may represent a cluster.
2620  *
2621  * To simplify operation and allow better optimizations in the future,
2622  * this code does not make any assumptions with regards to buffer alignment
2623  * or size.
2624  */
2625 static
2626 int
2627 hammer_vop_strategy(struct vop_strategy_args *ap)
2628 {
2629         struct buf *bp;
2630         int error;
2631
2632         bp = ap->a_bio->bio_buf;
2633
2634         switch(bp->b_cmd) {
2635         case BUF_CMD_READ:
2636                 error = hammer_vop_strategy_read(ap);
2637                 break;
2638         case BUF_CMD_WRITE:
2639                 error = hammer_vop_strategy_write(ap);
2640                 break;
2641         default:
2642                 bp->b_error = error = EINVAL;
2643                 bp->b_flags |= B_ERROR;
2644                 biodone(ap->a_bio);
2645                 break;
2646         }
2647
2648         /* hammer_dump_dedup_cache(((hammer_inode_t)ap->a_vp->v_data)->hmp); */
2649
2650         return (error);
2651 }
2652
2653 /*
2654  * Read from a regular file.  Iterate the related records and fill in the
2655  * BIO/BUF.  Gaps are zero-filled.
2656  *
2657  * The support code in hammer_object.c should be used to deal with mixed
2658  * in-memory and on-disk records.
2659  *
2660  * NOTE: Can be called from the cluster code with an oversized buf.
2661  *
2662  * XXX atime update
2663  */
2664 static
2665 int
2666 hammer_vop_strategy_read(struct vop_strategy_args *ap)
2667 {
2668         struct hammer_transaction trans;
2669         struct hammer_inode *ip;
2670         struct hammer_inode *dip;
2671         hammer_mount_t hmp;
2672         struct hammer_cursor cursor;
2673         hammer_base_elm_t base;
2674         hammer_off_t disk_offset;
2675         struct bio *bio;
2676         struct bio *nbio;
2677         struct buf *bp;
2678         int64_t rec_offset;
2679         int64_t ran_end;
2680         int64_t tmp64;
2681         int error;
2682         int boff;
2683         int roff;
2684         int n;
2685         int isdedupable;
2686
2687         bio = ap->a_bio;
2688         bp = bio->bio_buf;
2689         ip = ap->a_vp->v_data;
2690         hmp = ip->hmp;
2691
2692         /*
2693          * The zone-2 disk offset may have been set by the cluster code via
2694          * a BMAP operation, or else should be NOOFFSET.
2695          *
2696          * Checking the high bits for a match against zone-2 should suffice.
2697          *
2698          * In cases where a lot of data duplication is present it may be
2699          * more beneficial to drop through and doubule-buffer through the
2700          * device.
2701          */
2702         nbio = push_bio(bio);
2703         if (hammer_double_buffer == 0 &&
2704             (nbio->bio_offset & HAMMER_OFF_ZONE_MASK) ==
2705             HAMMER_ZONE_LARGE_DATA) {
2706                 lwkt_gettoken(&hmp->fs_token);
2707                 error = hammer_io_direct_read(hmp, nbio, NULL);
2708                 lwkt_reltoken(&hmp->fs_token);
2709                 return (error);
2710         }
2711
2712         /*
2713          * Well, that sucked.  Do it the hard way.  If all the stars are
2714          * aligned we may still be able to issue a direct-read.
2715          */
2716         lwkt_gettoken(&hmp->fs_token);
2717         hammer_simple_transaction(&trans, hmp);
2718         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2719
2720         /*
2721          * Key range (begin and end inclusive) to scan.  Note that the key's
2722          * stored in the actual records represent BASE+LEN, not BASE.  The
2723          * first record containing bio_offset will have a key > bio_offset.
2724          */
2725         cursor.key_beg.localization = ip->obj_localization +
2726                                       HAMMER_LOCALIZE_MISC;
2727         cursor.key_beg.obj_id = ip->obj_id;
2728         cursor.key_beg.create_tid = 0;
2729         cursor.key_beg.delete_tid = 0;
2730         cursor.key_beg.obj_type = 0;
2731         cursor.key_beg.key = bio->bio_offset + 1;
2732         cursor.asof = ip->obj_asof;
2733         cursor.flags |= HAMMER_CURSOR_ASOF;
2734
2735         cursor.key_end = cursor.key_beg;
2736         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2737 #if 0
2738         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
2739                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
2740                 cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
2741                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2742         } else
2743 #endif
2744         {
2745                 ran_end = bio->bio_offset + bp->b_bufsize;
2746                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2747                 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2748                 tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2749                 if (tmp64 < ran_end)
2750                         cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2751                 else
2752                         cursor.key_end.key = ran_end + MAXPHYS + 1;
2753         }
2754         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2755
2756         /*
2757          * Set NOSWAPCACHE for cursor data extraction if double buffering
2758          * is disabled or (if the file is not marked cacheable via chflags
2759          * and vm.swapcache_use_chflags is enabled).
2760          */
2761         if (hammer_double_buffer == 0 ||
2762             ((ap->a_vp->v_flag & VSWAPCACHE) == 0 &&
2763              vm_swapcache_use_chflags)) {
2764                 cursor.flags |= HAMMER_CURSOR_NOSWAPCACHE;
2765         }
2766
2767         error = hammer_ip_first(&cursor);
2768         boff = 0;
2769
2770         while (error == 0) {
2771                 /*
2772                  * Get the base file offset of the record.  The key for
2773                  * data records is (base + bytes) rather then (base).
2774                  */
2775                 base = &cursor.leaf->base;
2776                 rec_offset = base->key - cursor.leaf->data_len;
2777
2778                 /*
2779                  * Calculate the gap, if any, and zero-fill it.
2780                  *
2781                  * n is the offset of the start of the record verses our
2782                  * current seek offset in the bio.
2783                  */
2784                 n = (int)(rec_offset - (bio->bio_offset + boff));
2785                 if (n > 0) {
2786                         if (n > bp->b_bufsize - boff)
2787                                 n = bp->b_bufsize - boff;
2788                         bzero((char *)bp->b_data + boff, n);
2789                         boff += n;
2790                         n = 0;
2791                 }
2792
2793                 /*
2794                  * Calculate the data offset in the record and the number
2795                  * of bytes we can copy.
2796                  *
2797                  * There are two degenerate cases.  First, boff may already
2798                  * be at bp->b_bufsize.  Secondly, the data offset within
2799                  * the record may exceed the record's size.
2800                  */
2801                 roff = -n;
2802                 rec_offset += roff;
2803                 n = cursor.leaf->data_len - roff;
2804                 if (n <= 0) {
2805                         kprintf("strategy_read: bad n=%d roff=%d\n", n, roff);
2806                         n = 0;
2807                 } else if (n > bp->b_bufsize - boff) {
2808                         n = bp->b_bufsize - boff;
2809                 }
2810
2811                 /*
2812                  * Deal with cached truncations.  This cool bit of code
2813                  * allows truncate()/ftruncate() to avoid having to sync
2814                  * the file.
2815                  *
2816                  * If the frontend is truncated then all backend records are
2817                  * subject to the frontend's truncation.
2818                  *
2819                  * If the backend is truncated then backend records on-disk
2820                  * (but not in-memory) are subject to the backend's
2821                  * truncation.  In-memory records owned by the backend
2822                  * represent data written after the truncation point on the
2823                  * backend and must not be truncated.
2824                  *
2825                  * Truncate operations deal with frontend buffer cache
2826                  * buffers and frontend-owned in-memory records synchronously.
2827                  */
2828                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2829                         if (hammer_cursor_ondisk(&cursor)/* ||
2830                             cursor.iprec->flush_state == HAMMER_FST_FLUSH*/) {
2831                                 if (ip->trunc_off <= rec_offset)
2832                                         n = 0;
2833                                 else if (ip->trunc_off < rec_offset + n)
2834                                         n = (int)(ip->trunc_off - rec_offset);
2835                         }
2836                 }
2837                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2838                         if (hammer_cursor_ondisk(&cursor)) {
2839                                 if (ip->sync_trunc_off <= rec_offset)
2840                                         n = 0;
2841                                 else if (ip->sync_trunc_off < rec_offset + n)
2842                                         n = (int)(ip->sync_trunc_off - rec_offset);
2843                         }
2844                 }
2845
2846                 /*
2847                  * Try to issue a direct read into our bio if possible,
2848                  * otherwise resolve the element data into a hammer_buffer
2849                  * and copy.
2850                  *
2851                  * The buffer on-disk should be zerod past any real
2852                  * truncation point, but may not be for any synthesized
2853                  * truncation point from above.
2854                  */
2855                 disk_offset = cursor.leaf->data_offset + roff;
2856                 isdedupable = (boff == 0 && n == bp->b_bufsize &&
2857                                hammer_cursor_ondisk(&cursor) &&
2858                                ((int)disk_offset & HAMMER_BUFMASK) == 0);
2859
2860                 if (isdedupable && hammer_double_buffer == 0) {
2861                         KKASSERT((disk_offset & HAMMER_OFF_ZONE_MASK) ==
2862                                  HAMMER_ZONE_LARGE_DATA);
2863                         nbio->bio_offset = disk_offset;
2864                         error = hammer_io_direct_read(hmp, nbio, cursor.leaf);
2865                         if (hammer_live_dedup && error == 0)
2866                                 hammer_dedup_cache_add(ip, cursor.leaf);
2867                         goto done;
2868                 } else if (n) {
2869                         error = hammer_ip_resolve_data(&cursor);
2870                         if (error == 0) {
2871                                 if (hammer_live_dedup && isdedupable)
2872                                         hammer_dedup_cache_add(ip, cursor.leaf);
2873                                 bcopy((char *)cursor.data + roff,
2874                                       (char *)bp->b_data + boff, n);
2875                         }
2876                 }
2877                 if (error)
2878                         break;
2879
2880                 /*
2881                  * We have to be sure that the only elements added to the
2882                  * dedup cache are those which are already on-media.
2883                  */
2884                 if (hammer_live_dedup && hammer_cursor_ondisk(&cursor))
2885                         hammer_dedup_cache_add(ip, cursor.leaf);
2886
2887                 /*
2888                  * Iterate until we have filled the request.
2889                  */
2890                 boff += n;
2891                 if (boff == bp->b_bufsize)
2892                         break;
2893                 error = hammer_ip_next(&cursor);
2894         }
2895
2896         /*
2897          * There may have been a gap after the last record
2898          */
2899         if (error == ENOENT)
2900                 error = 0;
2901         if (error == 0 && boff != bp->b_bufsize) {
2902                 KKASSERT(boff < bp->b_bufsize);
2903                 bzero((char *)bp->b_data + boff, bp->b_bufsize - boff);
2904                 /* boff = bp->b_bufsize; */
2905         }
2906
2907         /*
2908          * Disallow swapcache operation on the vnode buffer if double
2909          * buffering is enabled, the swapcache will get the data via
2910          * the block device buffer.
2911          */
2912         if (hammer_double_buffer)
2913                 bp->b_flags |= B_NOTMETA;
2914
2915         /*
2916          * Cleanup
2917          */
2918         bp->b_resid = 0;
2919         bp->b_error = error;
2920         if (error)
2921                 bp->b_flags |= B_ERROR;
2922         biodone(ap->a_bio);
2923
2924 done:
2925         /*
2926          * Cache the b-tree node for the last data read in cache[1].
2927          *
2928          * If we hit the file EOF then also cache the node in the
2929          * governing director's cache[3], it will be used to initialize
2930          * the inode's cache[1] for any inodes looked up via the directory.
2931          *
2932          * This doesn't reduce disk accesses since the B-Tree chain is
2933          * likely cached, but it does reduce cpu overhead when looking
2934          * up file offsets for cpdup/tar/cpio style iterations.
2935          */
2936         if (cursor.node)
2937                 hammer_cache_node(&ip->cache[1], cursor.node);
2938         if (ran_end >= ip->ino_data.size) {
2939                 dip = hammer_find_inode(&trans, ip->ino_data.parent_obj_id,
2940                                         ip->obj_asof, ip->obj_localization);
2941                 if (dip) {
2942                         hammer_cache_node(&dip->cache[3], cursor.node);
2943                         hammer_rel_inode(dip, 0);
2944                 }
2945         }
2946         hammer_done_cursor(&cursor);
2947         hammer_done_transaction(&trans);
2948         lwkt_reltoken(&hmp->fs_token);
2949         return(error);
2950 }
2951
2952 /*
2953  * BMAP operation - used to support cluster_read() only.
2954  *
2955  * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
2956  *
2957  * This routine may return EOPNOTSUPP if the opration is not supported for
2958  * the specified offset.  The contents of the pointer arguments do not
2959  * need to be initialized in that case.
2960  *
2961  * If a disk address is available and properly aligned return 0 with
2962  * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
2963  * to the run-length relative to that offset.  Callers may assume that
2964  * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
2965  * large, so return EOPNOTSUPP if it is not sufficiently large.
2966  */
2967 static
2968 int
2969 hammer_vop_bmap(struct vop_bmap_args *ap)
2970 {
2971         struct hammer_transaction trans;
2972         struct hammer_inode *ip;
2973         hammer_mount_t hmp;
2974         struct hammer_cursor cursor;
2975         hammer_base_elm_t base;
2976         int64_t rec_offset;
2977         int64_t ran_end;
2978         int64_t tmp64;
2979         int64_t base_offset;
2980         int64_t base_disk_offset;
2981         int64_t last_offset;
2982         hammer_off_t last_disk_offset;
2983         hammer_off_t disk_offset;
2984         int     rec_len;
2985         int     error;
2986         int     blksize;
2987
2988         ++hammer_stats_file_iopsr;
2989         ip = ap->a_vp->v_data;
2990         hmp = ip->hmp;
2991
2992         /*
2993          * We can only BMAP regular files.  We can't BMAP database files,
2994          * directories, etc.
2995          */
2996         if (ip->ino_data.obj_type != HAMMER_OBJTYPE_REGFILE)
2997                 return(EOPNOTSUPP);
2998
2999         /*
3000          * bmap is typically called with runp/runb both NULL when used
3001          * for writing.  We do not support BMAP for writing atm.
3002          */
3003         if (ap->a_cmd != BUF_CMD_READ)
3004                 return(EOPNOTSUPP);
3005
3006         /*
3007          * Scan the B-Tree to acquire blockmap addresses, then translate
3008          * to raw addresses.
3009          */
3010         lwkt_gettoken(&hmp->fs_token);
3011         hammer_simple_transaction(&trans, hmp);
3012 #if 0
3013         kprintf("bmap_beg %016llx ip->cache %p\n",
3014                 (long long)ap->a_loffset, ip->cache[1]);
3015 #endif
3016         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
3017
3018         /*
3019          * Key range (begin and end inclusive) to scan.  Note that the key's
3020          * stored in the actual records represent BASE+LEN, not BASE.  The
3021          * first record containing bio_offset will have a key > bio_offset.
3022          */
3023         cursor.key_beg.localization = ip->obj_localization +
3024                                       HAMMER_LOCALIZE_MISC;
3025         cursor.key_beg.obj_id = ip->obj_id;
3026         cursor.key_beg.create_tid = 0;
3027         cursor.key_beg.delete_tid = 0;
3028         cursor.key_beg.obj_type = 0;
3029         if (ap->a_runb)
3030                 cursor.key_beg.key = ap->a_loffset - MAXPHYS + 1;
3031         else
3032                 cursor.key_beg.key = ap->a_loffset + 1;
3033         if (cursor.key_beg.key < 0)
3034                 cursor.key_beg.key = 0;
3035         cursor.asof = ip->obj_asof;
3036         cursor.flags |= HAMMER_CURSOR_ASOF;
3037
3038         cursor.key_end = cursor.key_beg;
3039         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
3040
3041         ran_end = ap->a_loffset + MAXPHYS;
3042         cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
3043         cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
3044         tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
3045         if (tmp64 < ran_end)
3046                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
3047         else
3048                 cursor.key_end.key = ran_end + MAXPHYS + 1;
3049
3050         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
3051
3052         error = hammer_ip_first(&cursor);
3053         base_offset = last_offset = 0;
3054         base_disk_offset = last_disk_offset = 0;
3055
3056         while (error == 0) {
3057                 /*
3058                  * Get the base file offset of the record.  The key for
3059                  * data records is (base + bytes) rather then (base).
3060                  *
3061                  * NOTE: rec_offset + rec_len may exceed the end-of-file.
3062                  * The extra bytes should be zero on-disk and the BMAP op
3063                  * should still be ok.
3064                  */
3065                 base = &cursor.leaf->base;
3066                 rec_offset = base->key - cursor.leaf->data_len;
3067                 rec_len    = cursor.leaf->data_len;
3068
3069                 /*
3070                  * Incorporate any cached truncation.
3071                  *
3072                  * NOTE: Modifications to rec_len based on synthesized
3073                  * truncation points remove the guarantee that any extended
3074                  * data on disk is zero (since the truncations may not have
3075                  * taken place on-media yet).
3076                  */
3077                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
3078                         if (hammer_cursor_ondisk(&cursor) ||
3079                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
3080                                 if (ip->trunc_off <= rec_offset)
3081                                         rec_len = 0;
3082                                 else if (ip->trunc_off < rec_offset + rec_len)
3083                                         rec_len = (int)(ip->trunc_off - rec_offset);
3084                         }
3085                 }
3086                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
3087                         if (hammer_cursor_ondisk(&cursor)) {
3088                                 if (ip->sync_trunc_off <= rec_offset)
3089                                         rec_len = 0;
3090                                 else if (ip->sync_trunc_off < rec_offset + rec_len)
3091                                         rec_len = (int)(ip->sync_trunc_off - rec_offset);
3092                         }
3093                 }
3094
3095                 /*
3096                  * Accumulate information.  If we have hit a discontiguous
3097                  * block reset base_offset unless we are already beyond the
3098                  * requested offset.  If we are, that's it, we stop.
3099                  */
3100                 if (error)
3101                         break;
3102                 if (hammer_cursor_ondisk(&cursor)) {
3103                         disk_offset = cursor.leaf->data_offset;
3104                         if (rec_offset != last_offset ||
3105                             disk_offset != last_disk_offset) {
3106                                 if (rec_offset > ap->a_loffset)
3107                                         break;
3108                                 base_offset = rec_offset;
3109                                 base_disk_offset = disk_offset;
3110                         }
3111                         last_offset = rec_offset + rec_len;
3112                         last_disk_offset = disk_offset + rec_len;
3113
3114                         if (hammer_live_dedup)
3115                                 hammer_dedup_cache_add(ip, cursor.leaf);
3116                 }
3117
3118                 error = hammer_ip_next(&cursor);
3119         }
3120
3121 #if 0
3122         kprintf("BMAP %016llx:  %016llx - %016llx\n",
3123                 (long long)ap->a_loffset,
3124                 (long long)base_offset,
3125                 (long long)last_offset);
3126         kprintf("BMAP %16s:  %016llx - %016llx\n", "",
3127                 (long long)base_disk_offset,
3128                 (long long)last_disk_offset);
3129 #endif
3130
3131         if (cursor.node) {
3132                 hammer_cache_node(&ip->cache[1], cursor.node);
3133 #if 0
3134                 kprintf("bmap_end2 %016llx ip->cache %p\n",
3135                         (long long)ap->a_loffset, ip->cache[1]);
3136 #endif
3137         }
3138         hammer_done_cursor(&cursor);
3139         hammer_done_transaction(&trans);
3140         lwkt_reltoken(&hmp->fs_token);
3141
3142         /*
3143          * If we couldn't find any records or the records we did find were
3144          * all behind the requested offset, return failure.  A forward
3145          * truncation can leave a hole w/ no on-disk records.
3146          */
3147         if (last_offset == 0 || last_offset < ap->a_loffset)
3148                 return (EOPNOTSUPP);
3149
3150         /*
3151          * Figure out the block size at the requested offset and adjust
3152          * our limits so the cluster_read() does not create inappropriately
3153          * sized buffer cache buffers.
3154          */
3155         blksize = hammer_blocksize(ap->a_loffset);
3156         if (hammer_blocksize(base_offset) != blksize) {
3157                 base_offset = hammer_blockdemarc(base_offset, ap->a_loffset);
3158         }
3159         if (last_offset != ap->a_loffset &&
3160             hammer_blocksize(last_offset - 1) != blksize) {
3161                 last_offset = hammer_blockdemarc(ap->a_loffset,
3162                                                  last_offset - 1);
3163         }
3164
3165         /*
3166          * Returning EOPNOTSUPP simply prevents the direct-IO optimization
3167          * from occuring.
3168          */
3169         disk_offset = base_disk_offset + (ap->a_loffset - base_offset);
3170
3171         if ((disk_offset & HAMMER_OFF_ZONE_MASK) != HAMMER_ZONE_LARGE_DATA) {
3172                 /*
3173                  * Only large-data zones can be direct-IOd
3174                  */
3175                 error = EOPNOTSUPP;
3176         } else if ((disk_offset & HAMMER_BUFMASK) ||
3177                    (last_offset - ap->a_loffset) < blksize) {
3178                 /*
3179                  * doffsetp is not aligned or the forward run size does
3180                  * not cover a whole buffer, disallow the direct I/O.
3181                  */
3182                 error = EOPNOTSUPP;
3183         } else {
3184                 /*
3185                  * We're good.
3186                  */
3187                 *ap->a_doffsetp = disk_offset;
3188                 if (ap->a_runb) {
3189                         *ap->a_runb = ap->a_loffset - base_offset;
3190                         KKASSERT(*ap->a_runb >= 0);
3191                 }
3192                 if (ap->a_runp) {
3193                         *ap->a_runp = last_offset - ap->a_loffset;
3194                         KKASSERT(*ap->a_runp >= 0);
3195                 }
3196                 error = 0;
3197         }
3198         return(error);
3199 }
3200
3201 /*
3202  * Write to a regular file.   Because this is a strategy call the OS is
3203  * trying to actually get data onto the media.
3204  */
3205 static
3206 int
3207 hammer_vop_strategy_write(struct vop_strategy_args *ap)
3208 {
3209         hammer_record_t record;
3210         hammer_mount_t hmp;
3211         hammer_inode_t ip;
3212         struct bio *bio;
3213         struct buf *bp;
3214         int blksize;
3215         int bytes;
3216         int error;
3217
3218         bio = ap->a_bio;
3219         bp = bio->bio_buf;
3220         ip = ap->a_vp->v_data;
3221         hmp = ip->hmp;
3222
3223         blksize = hammer_blocksize(bio->bio_offset);
3224         KKASSERT(bp->b_bufsize == blksize);
3225
3226         if (ip->flags & HAMMER_INODE_RO) {
3227                 bp->b_error = EROFS;
3228                 bp->b_flags |= B_ERROR;
3229                 biodone(ap->a_bio);
3230                 return(EROFS);
3231         }
3232
3233         lwkt_gettoken(&hmp->fs_token);
3234
3235         /*
3236          * Interlock with inode destruction (no in-kernel or directory
3237          * topology visibility).  If we queue new IO while trying to
3238          * destroy the inode we can deadlock the vtrunc call in
3239          * hammer_inode_unloadable_check().
3240          *
3241          * Besides, there's no point flushing a bp associated with an
3242          * inode that is being destroyed on-media and has no kernel
3243          * references.
3244          */
3245         if ((ip->flags | ip->sync_flags) &
3246             (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) {
3247                 bp->b_resid = 0;
3248                 biodone(ap->a_bio);
3249                 lwkt_reltoken(&hmp->fs_token);
3250                 return(0);
3251         }
3252
3253         /*
3254          * Reserve space and issue a direct-write from the front-end.
3255          * NOTE: The direct_io code will hammer_bread/bcopy smaller
3256          * allocations.
3257          *
3258          * An in-memory record will be installed to reference the storage
3259          * until the flusher can get to it.
3260          *
3261          * Since we own the high level bio the front-end will not try to
3262          * do a direct-read until the write completes.
3263          *
3264          * NOTE: The only time we do not reserve a full-sized buffers
3265          * worth of data is if the file is small.  We do not try to
3266          * allocate a fragment (from the small-data zone) at the end of
3267          * an otherwise large file as this can lead to wildly separated
3268          * data.
3269          */
3270         KKASSERT((bio->bio_offset & HAMMER_BUFMASK) == 0);
3271         KKASSERT(bio->bio_offset < ip->ino_data.size);
3272         if (bio->bio_offset || ip->ino_data.size > HAMMER_BUFSIZE / 2)
3273                 bytes = bp->b_bufsize;
3274         else
3275                 bytes = ((int)ip->ino_data.size + 15) & ~15;
3276
3277         record = hammer_ip_add_bulk(ip, bio->bio_offset, bp->b_data,
3278                                     bytes, &error);
3279
3280         /*
3281          * B_VFSFLAG1 indicates that a REDO_WRITE entry was generated
3282          * in hammer_vop_write().  We must flag the record so the proper
3283          * REDO_TERM_WRITE entry is generated during the flush.
3284          */
3285         if (record) {
3286                 if (bp->b_flags & B_VFSFLAG1) {
3287                         record->flags |= HAMMER_RECF_REDO;
3288                         bp->b_flags &= ~B_VFSFLAG1;
3289                 }
3290                 if (record->flags & HAMMER_RECF_DEDUPED) {
3291                         bp->b_resid = 0;
3292                         hammer_ip_replace_bulk(hmp, record);
3293                         biodone(ap->a_bio);
3294                 } else {
3295                         hammer_io_direct_write(hmp, bio, record);
3296                 }
3297                 if (ip->rsv_recs > 1 && hmp->rsv_recs > hammer_limit_recs)
3298                         hammer_flush_inode(ip, 0);
3299         } else {
3300                 bp->b_bio2.bio_offset = NOOFFSET;
3301                 bp->b_error = error;
3302                 bp->b_flags |= B_ERROR;
3303                 biodone(ap->a_bio);
3304         }
3305         lwkt_reltoken(&hmp->fs_token);
3306         return(error);
3307 }
3308
3309 /*
3310  * dounlink - disconnect a directory entry
3311  *
3312  * XXX whiteout support not really in yet
3313  */
3314 static int
3315 hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
3316                 struct vnode *dvp, struct ucred *cred,
3317                 int flags, int isdir)
3318 {
3319         struct namecache *ncp;
3320         hammer_inode_t dip;
3321         hammer_inode_t ip;
3322         hammer_mount_t hmp;
3323         struct hammer_cursor cursor;
3324         int64_t namekey;
3325         u_int32_t max_iterations;
3326         int nlen, error;
3327
3328         /*
3329          * Calculate the namekey and setup the key range for the scan.  This
3330          * works kinda like a chained hash table where the lower 32 bits
3331          * of the namekey synthesize the chain.
3332          *
3333          * The key range is inclusive of both key_beg and key_end.
3334          */
3335         dip = VTOI(dvp);
3336         ncp = nch->ncp;
3337         hmp = dip->hmp;
3338
3339         if (dip->flags & HAMMER_INODE_RO)
3340                 return (EROFS);
3341
3342         namekey = hammer_directory_namekey(dip, ncp->nc_name, ncp->nc_nlen,
3343                                            &max_iterations);
3344 retry:
3345         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
3346         cursor.key_beg.localization = dip->obj_localization +
3347                                       hammer_dir_localization(dip);
3348         cursor.key_beg.obj_id = dip->obj_id;
3349         cursor.key_beg.key = namekey;
3350         cursor.key_beg.create_tid = 0;
3351         cursor.key_beg.delete_tid = 0;
3352         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
3353         cursor.key_beg.obj_type = 0;
3354
3355         cursor.key_end = cursor.key_beg;
3356         cursor.key_end.key += max_iterations;
3357         cursor.asof = dip->obj_asof;
3358         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
3359
3360         /*
3361          * Scan all matching records (the chain), locate the one matching
3362          * the requested path component.  info->last_error contains the
3363          * error code on search termination and could be 0, ENOENT, or
3364          * something else.
3365          *
3366          * The hammer_ip_*() functions merge in-memory records with on-disk
3367          * records for the purposes of the search.
3368          */
3369         error = hammer_ip_first(&cursor);
3370
3371         while (error == 0) {
3372                 error = hammer_ip_resolve_data(&cursor);
3373                 if (error)
3374                         break;
3375                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
3376                 KKASSERT(nlen > 0);
3377                 if (ncp->nc_nlen == nlen &&
3378                     bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
3379                         break;
3380                 }
3381                 error = hammer_ip_next(&cursor);
3382         }
3383
3384         /*
3385          * If all is ok we have to get the inode so we can adjust nlinks.
3386          * To avoid a deadlock with the flusher we must release the inode
3387          * lock on the directory when acquiring the inode for the entry.
3388          *
3389          * If the target is a directory, it must be empty.
3390          */
3391         if (error == 0) {
3392                 hammer_unlock(&cursor.ip->lock);
3393                 ip = hammer_get_inode(trans, dip, cursor.data->entry.obj_id,
3394                                       hmp->asof,
3395                                       cursor.data->entry.localization,
3396                                       0, &error);
3397                 hammer_lock_sh(&cursor.ip->lock);
3398                 if (error == ENOENT) {
3399                         kprintf("HAMMER: WARNING: Removing "
3400                                 "dirent w/missing inode \"%s\"\n"
3401                                 "\tobj_id = %016llx\n",
3402                                 ncp->nc_name,
3403                                 (long long)cursor.data->entry.obj_id);
3404                         error = 0;
3405                 }
3406
3407                 /*
3408                  * If isdir >= 0 we validate that the entry is or is not a
3409                  * directory.  If isdir < 0 we don't care.
3410                  */
3411                 if (error == 0 && isdir >= 0 && ip) {
3412                         if (isdir &&
3413                             ip->ino_data.obj_type != HAMMER_OBJTYPE_DIRECTORY) {
3414                                 error = ENOTDIR;
3415                         } else if (isdir == 0 &&
3416                             ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
3417                                 error = EISDIR;
3418                         }
3419                 }
3420
3421                 /*
3422                  * If we are trying to remove a directory the directory must
3423                  * be empty.
3424                  *
3425                  * The check directory code can loop and deadlock/retry.  Our
3426                  * own cursor's node locks must be released to avoid a 3-way
3427                  * deadlock with the flusher if the check directory code
3428                  * blocks.
3429                  *
3430                  * If any changes whatsoever have been made to the cursor
3431                  * set EDEADLK and retry.
3432                  *
3433                  * WARNING: See warnings in hammer_unlock_cursor()
3434                  *          function.
3435                  */
3436                 if (error == 0 && ip && ip->ino_data.obj_type ==
3437                                         HAMMER_OBJTYPE_DIRECTORY) {
3438                         hammer_unlock_cursor(&cursor);
3439                         error = hammer_ip_check_directory_empty(trans, ip);
3440                         hammer_lock_cursor(&cursor);
3441                         if (cursor.flags & HAMMER_CURSOR_RETEST) {
3442                                 kprintf("HAMMER: Warning: avoided deadlock "
3443                                         "on rmdir '%s'\n",
3444                                         ncp->nc_name);
3445                                 error = EDEADLK;
3446                         }
3447                 }
3448
3449                 /*
3450                  * Delete the directory entry.
3451                  *
3452                  * WARNING: hammer_ip_del_directory() may have to terminate
3453                  * the cursor to avoid a deadlock.  It is ok to call
3454                  * hammer_done_cursor() twice.
3455                  */
3456                 if (error == 0) {
3457                         error = hammer_ip_del_directory(trans, &cursor,
3458                                                         dip, ip);
3459                 }
3460                 hammer_done_cursor(&cursor);
3461                 if (error == 0) {
3462                         cache_setunresolved(nch);
3463                         cache_setvp(nch, NULL);
3464
3465                         /*
3466                          * NOTE: ip->vp, if non-NULL, cannot be directly
3467                          *       referenced without formally acquiring the
3468                          *       vp since the vp might have zero refs on it,
3469                          *       or in the middle of a reclaim, etc.
3470                          *
3471                          * NOTE: The cache_setunresolved() can rip the vp
3472                          *       out from under us since the vp may not have
3473                          *       any refs, in which case ip->vp will be NULL
3474                          *       from the outset.
3475                          */
3476                         while (ip && ip->vp) {
3477                                 struct vnode *vp;
3478
3479                                 error = hammer_get_vnode(ip, &vp);
3480                                 if (error == 0 && vp) {
3481                                         vn_unlock(vp);
3482                                         hammer_knote(ip->vp, NOTE_DELETE);
3483                                         cache_inval_vp(ip->vp, CINV_DESTROY);
3484                                         vrele(vp);
3485                                         break;
3486                                 }
3487                                 kprintf("Debug: HAMMER ip/vp race1 avoided\n");
3488                         }
3489                 }
3490                 if (ip)
3491                         hammer_rel_inode(ip, 0);
3492         } else {
3493                 hammer_done_cursor(&cursor);
3494         }
3495         if (error == EDEADLK)
3496                 goto retry;
3497
3498         return (error);
3499 }
3500
3501 /************************************************************************
3502  *                          FIFO AND SPECFS OPS                         *
3503  ************************************************************************
3504  *
3505  */
3506 static int
3507 hammer_vop_fifoclose (struct vop_close_args *ap)
3508 {
3509         /* XXX update itimes */
3510         return (VOCALL(&fifo_vnode_vops, &ap->a_head));
3511 }
3512
3513 static int
3514 hammer_vop_fiforead (struct vop_read_args *ap)
3515 {
3516         int error;
3517
3518         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3519         /* XXX update access time */
3520         return (error);
3521 }
3522
3523 static int
3524 hammer_vop_fifowrite (struct vop_write_args *ap)
3525 {
3526         int error;
3527
3528         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3529         /* XXX update access time */
3530         return (error);
3531 }
3532
3533 static
3534 int
3535 hammer_vop_fifokqfilter(struct vop_kqfilter_args *ap)
3536 {
3537         int error;
3538
3539         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3540         if (error)
3541                 error = hammer_vop_kqfilter(ap);
3542         return(error);
3543 }
3544
3545 /************************************************************************
3546  *                          KQFILTER OPS                                *
3547  ************************************************************************
3548  *
3549  */
3550 static void filt_hammerdetach(struct knote *kn);
3551 static int filt_hammerread(struct knote *kn, long hint);
3552 static int filt_hammerwrite(struct knote *kn, long hint);
3553 static int filt_hammervnode(struct knote *kn, long hint);
3554
3555 static struct filterops hammerread_filtops =
3556         { FILTEROP_ISFD, NULL, filt_hammerdetach, filt_hammerread };
3557 static struct filterops hammerwrite_filtops =
3558         { FILTEROP_ISFD, NULL, filt_hammerdetach, filt_hammerwrite };
3559 static struct filterops hammervnode_filtops =
3560         { FILTEROP_ISFD, NULL, filt_hammerdetach, filt_hammervnode };
3561
3562 static
3563 int
3564 hammer_vop_kqfilter(struct vop_kqfilter_args *ap)
3565 {
3566         struct vnode *vp = ap->a_vp;
3567         struct knote *kn = ap->a_kn;
3568
3569         switch (kn->kn_filter) {
3570         case EVFILT_READ:
3571                 kn->kn_fop = &hammerread_filtops;
3572                 break;
3573         case EVFILT_WRITE:
3574                 kn->kn_fop = &hammerwrite_filtops;
3575                 break;
3576         case EVFILT_VNODE:
3577                 kn->kn_fop = &hammervnode_filtops;
3578                 break;
3579         default:
3580                 return (EOPNOTSUPP);
3581         }
3582
3583         kn->kn_hook = (caddr_t)vp;
3584
3585         knote_insert(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3586
3587         return(0);
3588 }
3589
3590 static void
3591 filt_hammerdetach(struct knote *kn)
3592 {
3593         struct vnode *vp = (void *)kn->kn_hook;
3594
3595         knote_remove(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3596 }
3597
3598 static int
3599 filt_hammerread(struct knote *kn, long hint)
3600 {
3601         struct vnode *vp = (void *)kn->kn_hook;
3602         hammer_inode_t ip = VTOI(vp);
3603         hammer_mount_t hmp = ip->hmp;
3604         off_t off;
3605
3606         if (hint == NOTE_REVOKE) {
3607                 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
3608                 return(1);
3609         }
3610         lwkt_gettoken(&hmp->fs_token);  /* XXX use per-ip-token */
3611         off = ip->ino_data.size - kn->kn_fp->f_offset;
3612         kn->kn_data = (off < INTPTR_MAX) ? off : INTPTR_MAX;
3613         lwkt_reltoken(&hmp->fs_token);
3614         if (kn->kn_sfflags & NOTE_OLDAPI)
3615                 return(1);
3616         return (kn->kn_data != 0);
3617 }
3618
3619 static int
3620 filt_hammerwrite(struct knote *kn, long hint)
3621 {
3622         if (hint == NOTE_REVOKE)
3623                 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
3624         kn->kn_data = 0;
3625         return (1);
3626 }
3627
3628 static int
3629 filt_hammervnode(struct knote *kn, long hint)
3630 {
3631         if (kn->kn_sfflags & hint)
3632                 kn->kn_fflags |= hint;
3633         if (hint == NOTE_REVOKE) {
3634                 kn->kn_flags |= EV_EOF;
3635                 return (1);
3636         }
3637         return (kn->kn_fflags != 0);
3638 }
3639