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
  35 #include <sys/mountctl.h>
  36 #include <sys/namecache.h>
  37 #include <sys/buf2.h>
  38 #include <vfs/fifofs/fifo.h>
  39
  40 #include "hammer.h"
  41
  42 /*
  43  * USERFS VNOPS
  44  */
  45 static int hammer_vop_fsync(struct vop_fsync_args *);
  46 static int hammer_vop_read(struct vop_read_args *);
  47 static int hammer_vop_write(struct vop_write_args *);
  48 static int hammer_vop_access(struct vop_access_args *);
  49 static int hammer_vop_advlock(struct vop_advlock_args *);
  50 static int hammer_vop_close(struct vop_close_args *);
  51 static int hammer_vop_ncreate(struct vop_ncreate_args *);
  52 static int hammer_vop_getattr(struct vop_getattr_args *);
  53 static int hammer_vop_nresolve(struct vop_nresolve_args *);
  54 static int hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *);
  55 static int hammer_vop_nlink(struct vop_nlink_args *);
  56 static int hammer_vop_nmkdir(struct vop_nmkdir_args *);
  57 static int hammer_vop_nmknod(struct vop_nmknod_args *);
  58 static int hammer_vop_open(struct vop_open_args *);
  59 static int hammer_vop_print(struct vop_print_args *);
  60 static int hammer_vop_readdir(struct vop_readdir_args *);
  61 static int hammer_vop_readlink(struct vop_readlink_args *);
  62 static int hammer_vop_nremove(struct vop_nremove_args *);
  63 static int hammer_vop_nrename(struct vop_nrename_args *);
  64 static int hammer_vop_nrmdir(struct vop_nrmdir_args *);
  65 static int hammer_vop_markatime(struct vop_markatime_args *);
  66 static int hammer_vop_setattr(struct vop_setattr_args *);
  67 static int hammer_vop_strategy(struct vop_strategy_args *);
  68 static int hammer_vop_bmap(struct vop_bmap_args *ap);
  69 static int hammer_vop_nsymlink(struct vop_nsymlink_args *);
  70 static int hammer_vop_nwhiteout(struct vop_nwhiteout_args *);
  71 static int hammer_vop_ioctl(struct vop_ioctl_args *);
  72 static int hammer_vop_mountctl(struct vop_mountctl_args *);
  73 static int hammer_vop_kqfilter (struct vop_kqfilter_args *);
  74
  75 static int hammer_vop_fifoclose (struct vop_close_args *);
  76 static int hammer_vop_fiforead (struct vop_read_args *);
  77 static int hammer_vop_fifowrite (struct vop_write_args *);
  78 static int hammer_vop_fifokqfilter (struct vop_kqfilter_args *);
  79
  80 struct vop_ops hammer_vnode_vops = {
  81         .vop_default =          vop_defaultop,
  82         .vop_fsync =            hammer_vop_fsync,
  83         .vop_getpages =         vop_stdgetpages,
  84         .vop_putpages =         vop_stdputpages,
  85         .vop_read =             hammer_vop_read,
  86         .vop_write =            hammer_vop_write,
  87         .vop_access =           hammer_vop_access,
  88         .vop_advlock =          hammer_vop_advlock,
  89         .vop_close =            hammer_vop_close,
  90         .vop_ncreate =          hammer_vop_ncreate,
  91         .vop_getattr =          hammer_vop_getattr,
  92         .vop_inactive =         hammer_vop_inactive,
  93         .vop_reclaim =          hammer_vop_reclaim,
  94         .vop_nresolve =         hammer_vop_nresolve,
  95         .vop_nlookupdotdot =    hammer_vop_nlookupdotdot,
  96         .vop_nlink =            hammer_vop_nlink,
  97         .vop_nmkdir =           hammer_vop_nmkdir,
  98         .vop_nmknod =           hammer_vop_nmknod,
  99         .vop_open =             hammer_vop_open,
 100         .vop_pathconf =         vop_stdpathconf,
 101         .vop_print =            hammer_vop_print,
 102         .vop_readdir =          hammer_vop_readdir,
 103         .vop_readlink =         hammer_vop_readlink,
 104         .vop_nremove =          hammer_vop_nremove,
 105         .vop_nrename =          hammer_vop_nrename,
 106         .vop_nrmdir =           hammer_vop_nrmdir,
 107         .vop_markatime =        hammer_vop_markatime,
 108         .vop_setattr =          hammer_vop_setattr,
 109         .vop_bmap =             hammer_vop_bmap,
 110         .vop_strategy =         hammer_vop_strategy,
 111         .vop_nsymlink =         hammer_vop_nsymlink,
 112         .vop_nwhiteout =        hammer_vop_nwhiteout,
 113         .vop_ioctl =            hammer_vop_ioctl,
 114         .vop_mountctl =         hammer_vop_mountctl,
 115         .vop_kqfilter =         hammer_vop_kqfilter
 116 };
 117
 118 struct vop_ops hammer_spec_vops = {
 119         .vop_default =          vop_defaultop,
 120         .vop_fsync =            hammer_vop_fsync,
 121         .vop_read =             vop_stdnoread,
 122         .vop_write =            vop_stdnowrite,
 123         .vop_access =           hammer_vop_access,
 124         .vop_close =            hammer_vop_close,
 125         .vop_markatime =        hammer_vop_markatime,
 126         .vop_getattr =          hammer_vop_getattr,
 127         .vop_inactive =         hammer_vop_inactive,
 128         .vop_reclaim =          hammer_vop_reclaim,
 129         .vop_setattr =          hammer_vop_setattr
 130 };
 131
 132 struct vop_ops hammer_fifo_vops = {
 133         .vop_default =          fifo_vnoperate,
 134         .vop_fsync =            hammer_vop_fsync,
 135         .vop_read =             hammer_vop_fiforead,
 136         .vop_write =            hammer_vop_fifowrite,
 137         .vop_access =           hammer_vop_access,
 138         .vop_close =            hammer_vop_fifoclose,
 139         .vop_markatime =        hammer_vop_markatime,
 140         .vop_getattr =          hammer_vop_getattr,
 141         .vop_inactive =         hammer_vop_inactive,
 142         .vop_reclaim =          hammer_vop_reclaim,
 143         .vop_setattr =          hammer_vop_setattr,
 144         .vop_kqfilter =         hammer_vop_fifokqfilter
 145 };
 146
 147 static __inline
 148 void
 149 hammer_knote(struct vnode *vp, int flags)
 150 {
 151         if (flags)
 152                 KNOTE(&vp->v_pollinfo.vpi_kqinfo.ki_note, flags);
 153 }
 154
 155 static int hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
 156                            struct vnode *dvp, struct ucred *cred,
 157                            int flags, int isdir);
 158 static int hammer_vop_strategy_read(struct vop_strategy_args *ap);
 159 static int hammer_vop_strategy_write(struct vop_strategy_args *ap);
 160
 161 /*
 162  * hammer_vop_fsync { vp, waitfor }
 163  *
 164  * fsync() an inode to disk and wait for it to be completely committed
 165  * such that the information would not be undone if a crash occured after
 166  * return.
 167  *
 168  * NOTE: HAMMER's fsync()'s are going to remain expensive until we implement
 169  *       a REDO log.  A sysctl is provided to relax HAMMER's fsync()
 170  *       operation.
 171  *
 172  *       Ultimately the combination of a REDO log and use of fast storage
 173  *       to front-end cluster caches will make fsync fast, but it aint
 174  *       here yet.  And, in anycase, we need real transactional
 175  *       all-or-nothing features which are not restricted to a single file.
 176  */
 177 static
 178 int
 179 hammer_vop_fsync(struct vop_fsync_args *ap)
 180 {
 181         hammer_inode_t ip = VTOI(ap->a_vp);
 182         hammer_mount_t hmp = ip->hmp;
 183         int waitfor = ap->a_waitfor;
 184         int mode;
 185
 186         lwkt_gettoken(&hmp->fs_token);
 187
 188         /*
 189          * Fsync rule relaxation (default is either full synchronous flush
 190          * or REDO semantics with synchronous flush).
 191          */
 192         if (ap->a_flags & VOP_FSYNC_SYSCALL) {
 193                 switch(hammer_fsync_mode) {
 194                 case 0:
 195 mode0:
 196                         /* no REDO, full synchronous flush */
 197                         goto skip;
 198                 case 1:
 199 mode1:
 200                         /* no REDO, full asynchronous flush */
 201                         if (waitfor == MNT_WAIT)
 202                                 waitfor = MNT_NOWAIT;
 203                         goto skip;
 204                 case 2:
 205                         /* REDO semantics, synchronous flush */
 206                         if (hmp->version < HAMMER_VOL_VERSION_FOUR)
 207                                 goto mode0;
 208                         mode = HAMMER_FLUSH_UNDOS_AUTO;
 209                         break;
 210                 case 3:
 211                         /* REDO semantics, relaxed asynchronous flush */
 212                         if (hmp->version < HAMMER_VOL_VERSION_FOUR)
 213                                 goto mode1;
 214                         mode = HAMMER_FLUSH_UNDOS_RELAXED;
 215                         if (waitfor == MNT_WAIT)
 216                                 waitfor = MNT_NOWAIT;
 217                         break;
 218                 case 4:
 219                         /* ignore the fsync() system call */
 220                         lwkt_reltoken(&hmp->fs_token);
 221                         return(0);
 222                 default:
 223                         /* we have to do something */
 224                         mode = HAMMER_FLUSH_UNDOS_RELAXED;
 225                         if (waitfor == MNT_WAIT)
 226                                 waitfor = MNT_NOWAIT;
 227                         break;
 228                 }
 229
 230                 /*
 231                  * Fast fsync only needs to flush the UNDO/REDO fifo if
 232                  * HAMMER_INODE_REDO is non-zero and the only modifications
 233                  * made to the file are write or write-extends.
 234                  */
 235                 if ((ip->flags & HAMMER_INODE_REDO) &&
 236                     (ip->flags & HAMMER_INODE_MODMASK_NOREDO) == 0) {
 237                         ++hammer_count_fsyncs;
 238                         hammer_flusher_flush_undos(hmp, mode);
 239                         ip->redo_count = 0;
 240                         if (ip->vp && (ip->flags & HAMMER_INODE_MODMASK) == 0)
 241                                 vclrisdirty(ip->vp);
 242                         lwkt_reltoken(&hmp->fs_token);
 243                         return(0);
 244                 }
 245
 246                 /*
 247                  * REDO is enabled by fsync(), the idea being we really only
 248                  * want to lay down REDO records when programs are using
 249                  * fsync() heavily.  The first fsync() on the file starts
 250                  * the gravy train going and later fsync()s keep it hot by
 251                  * resetting the redo_count.
 252                  *
 253                  * We weren't running REDOs before now so we have to fall
 254                  * through and do a full fsync of what we have.
 255                  */
 256                 if (hmp->version >= HAMMER_VOL_VERSION_FOUR &&
 257                     (hmp->flags & HAMMER_MOUNT_REDO_RECOVERY_RUN) == 0) {
 258                         ip->flags |= HAMMER_INODE_REDO;
 259                         ip->redo_count = 0;
 260                 }
 261         }
 262 skip:
 263
 264         /*
 265          * Do a full flush sequence.
 266          *
 267          * Attempt to release the vnode while waiting for the inode to
 268          * finish flushing.  This can really mess up inactive->reclaim
 269          * sequences so only do it if the vnode is active.
 270          *
 271          * WARNING! The VX lock functions must be used.  vn_lock() will
 272          *          fail when this is part of a VOP_RECLAIM sequence.
 273          */
 274         ++hammer_count_fsyncs;
 275         vfsync(ap->a_vp, waitfor, 1, NULL, NULL);
 276         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 277         if (waitfor == MNT_WAIT) {
 278                 int dorelock;
 279
 280                 if ((ap->a_vp->v_flag & VRECLAIMED) == 0) {
 281                         vx_unlock(ap->a_vp);
 282                         dorelock = 1;
 283                 } else {
 284                         dorelock = 0;
 285                 }
 286                 hammer_wait_inode(ip);
 287                 if (dorelock)
 288                         vx_lock(ap->a_vp);
 289         }
 290         if (ip->vp && (ip->flags & HAMMER_INODE_MODMASK) == 0)
 291                 vclrisdirty(ip->vp);
 292         lwkt_reltoken(&hmp->fs_token);
 293         return (ip->error);
 294 }
 295
 296 /*
 297  * hammer_vop_read { vp, uio, ioflag, cred }
 298  *
 299  * MPSAFE (for the cache safe does not require fs_token)
 300  */
 301 static
 302 int
 303 hammer_vop_read(struct vop_read_args *ap)
 304 {
 305         struct hammer_transaction trans;
 306         hammer_inode_t ip;
 307         hammer_mount_t hmp;
 308         off_t offset;
 309         struct buf *bp;
 310         struct uio *uio;
 311         int error;
 312         int n;
 313         int seqcount;
 314         int ioseqcount;
 315         int blksize;
 316         int bigread;
 317         int got_trans;
 318         size_t resid;
 319
 320         if (ap->a_vp->v_type != VREG)
 321                 return (EINVAL);
 322         ip = VTOI(ap->a_vp);
 323         hmp = ip->hmp;
 324         error = 0;
 325         got_trans = 0;
 326         uio = ap->a_uio;
 327
 328         /*
 329          * Attempt to shortcut directly to the VM object using lwbufs.
 330          * This is much faster than instantiating buffer cache buffers.
 331          */
 332         resid = uio->uio_resid;
 333         error = vop_helper_read_shortcut(ap);
 334         hammer_stats_file_read += resid - uio->uio_resid;
 335         if (error)
 336                 return (error);
 337         if (uio->uio_resid == 0)
 338                 goto finished;
 339
 340         /*
 341          * Allow the UIO's size to override the sequential heuristic.
 342          */
 343         blksize = hammer_blocksize(uio->uio_offset);
 344         seqcount = (uio->uio_resid + (MAXBSIZE - 1)) / MAXBSIZE;
 345         ioseqcount = (ap->a_ioflag >> 16);
 346         if (seqcount < ioseqcount)
 347                 seqcount = ioseqcount;
 348
 349         /*
 350          * If reading or writing a huge amount of data we have to break
 351          * atomicy and allow the operation to be interrupted by a signal
 352          * or it can DOS the machine.
 353          */
 354         bigread = (uio->uio_resid > 100 * 1024 * 1024);
 355
 356         /*
 357          * Access the data typically in HAMMER_BUFSIZE blocks via the
 358          * buffer cache, but HAMMER may use a variable block size based
 359          * on the offset.
 360          *
 361          * XXX Temporary hack, delay the start transaction while we remain
 362          *     MPSAFE.  NOTE: ino_data.size cannot change while vnode is
 363          *     locked-shared.
 364          */
 365         while (uio->uio_resid > 0 && uio->uio_offset < ip->ino_data.size) {
 366                 int64_t base_offset;
 367                 int64_t file_limit;
 368
 369                 blksize = hammer_blocksize(uio->uio_offset);
 370                 offset = (int)uio->uio_offset & (blksize - 1);
 371                 base_offset = uio->uio_offset - offset;
 372
 373                 if (bigread && (error = hammer_signal_check(ip->hmp)) != 0)
 374                         break;
 375
 376                 /*
 377                  * MPSAFE
 378                  */
 379                 bp = getblk(ap->a_vp, base_offset, blksize, 0, 0);
 380                 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) == B_CACHE) {
 381                         bp->b_flags &= ~B_AGE;
 382                         error = 0;
 383                         goto skip;
 384                 }
 385                 if (ap->a_ioflag & IO_NRDELAY) {
 386                         bqrelse(bp);
 387                         return (EWOULDBLOCK);
 388                 }
 389
 390                 /*
 391                  * MPUNSAFE
 392                  */
 393                 if (got_trans == 0) {
 394                         hammer_start_transaction(&trans, ip->hmp);
 395                         got_trans = 1;
 396                 }
 397
 398                 /*
 399                  * NOTE: A valid bp has already been acquired, but was not
 400                  *       B_CACHE.
 401                  */
 402                 if (hammer_cluster_enable) {
 403                         /*
 404                          * Use file_limit to prevent cluster_read() from
 405                          * creating buffers of the wrong block size past
 406                          * the demarc.
 407                          */
 408                         file_limit = ip->ino_data.size;
 409                         if (base_offset < HAMMER_XDEMARC &&
 410                             file_limit > HAMMER_XDEMARC) {
 411                                 file_limit = HAMMER_XDEMARC;
 412                         }
 413                         error = cluster_readx(ap->a_vp,
 414                                              file_limit, base_offset,
 415                                              blksize, B_NOTMETA,
 416                                              uio->uio_resid,
 417                                              seqcount * MAXBSIZE,
 418                                              &bp);
 419                 } else {
 420                         error = breadnx(ap->a_vp, base_offset,
 421                                         blksize, B_NOTMETA,
 422                                         NULL, NULL, 0, &bp);
 423                 }
 424                 if (error) {
 425                         brelse(bp);
 426                         break;
 427                 }
 428 skip:
 429                 if ((hammer_debug_io & 0x0001) && (bp->b_flags & B_IOISSUED)) {
 430                         hdkprintf("zone2_offset %016jx read file %016jx@%016jx\n",
 431                                 (intmax_t)bp->b_bio2.bio_offset,
 432                                 (intmax_t)ip->obj_id,
 433                                 (intmax_t)bp->b_loffset);
 434                 }
 435                 bp->b_flags &= ~B_IOISSUED;
 436                 if (blksize == HAMMER_XBUFSIZE)
 437                         bp->b_flags |= B_CLUSTEROK;
 438
 439                 n = blksize - offset;
 440                 if (n > uio->uio_resid)
 441                         n = uio->uio_resid;
 442                 if (n > ip->ino_data.size - uio->uio_offset)
 443                         n = (int)(ip->ino_data.size - uio->uio_offset);
 444
 445                 /*
 446                  * Set B_AGE, data has a lower priority than meta-data.
 447                  *
 448                  * Use a hold/unlock/drop sequence to run the uiomove
 449                  * with the buffer unlocked, avoiding deadlocks against
 450                  * read()s on mmap()'d spaces.
 451                  */
 452                 bp->b_flags |= B_AGE;
 453                 error = uiomovebp(bp, (char *)bp->b_data + offset, n, uio);
 454                 bqrelse(bp);
 455
 456                 if (error)
 457                         break;
 458                 hammer_stats_file_read += n;
 459         }
 460
 461 finished:
 462
 463         /*
 464          * Try to update the atime with just the inode lock for maximum
 465          * concurrency.  If we can't shortcut it we have to get the full
 466          * blown transaction.
 467          */
 468         if (got_trans == 0 && hammer_update_atime_quick(ip) < 0) {
 469                 hammer_start_transaction(&trans, ip->hmp);
 470                 got_trans = 1;
 471         }
 472
 473         if (got_trans) {
 474                 if ((ip->flags & HAMMER_INODE_RO) == 0 &&
 475                     (ip->hmp->mp->mnt_flag & MNT_NOATIME) == 0) {
 476                         lwkt_gettoken(&hmp->fs_token);
 477                         ip->ino_data.atime = trans.time;
 478                         hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
 479                         hammer_done_transaction(&trans);
 480                         lwkt_reltoken(&hmp->fs_token);
 481                 } else {
 482                         hammer_done_transaction(&trans);
 483                 }
 484         }
 485         return (error);
 486 }
 487
 488 /*
 489  * hammer_vop_write { vp, uio, ioflag, cred }
 490  */
 491 static
 492 int
 493 hammer_vop_write(struct vop_write_args *ap)
 494 {
 495         struct hammer_transaction trans;
 496         hammer_inode_t ip;
 497         hammer_mount_t hmp;
 498         thread_t td;
 499         struct vnode *vp;
 500         struct uio *uio;
 501         int offset;
 502         off_t base_offset;
 503         int64_t cluster_eof;
 504         struct buf *bp;
 505         int kflags;
 506         int error;
 507         int n;
 508         int flags;
 509         int seqcount;
 510         int bigwrite;
 511
 512         vp = ap->a_vp;
 513         if (vp->v_type != VREG)
 514                 return (EINVAL);
 515         ip = VTOI(ap->a_vp);
 516         hmp = ip->hmp;
 517         error = 0;
 518         kflags = 0;
 519         seqcount = ap->a_ioflag >> 16;
 520
 521         if (ip->flags & HAMMER_INODE_RO)
 522                 return (EROFS);
 523
 524         /*
 525          * Create a transaction to cover the operations we perform.
 526          */
 527         hammer_start_transaction(&trans, hmp);
 528         uio = ap->a_uio;
 529
 530         /*
 531          * Use v_lastwrite_ts if file not open for writing
 532          * (i.e. a late msync)
 533          */
 534         if (vp->v_writecount == 0) {
 535                 trans.time = vp->v_lastwrite_ts.tv_sec * 1000000 +
 536                              vp->v_lastwrite_ts.tv_nsec / 1000;
 537         }
 538
 539
 540         /*
 541          * Check append mode
 542          */
 543         if (ap->a_ioflag & IO_APPEND)
 544                 uio->uio_offset = ip->ino_data.size;
 545
 546         /*
 547          * Check for illegal write offsets.  Valid range is 0...2^63-1.
 548          *
 549          * NOTE: the base_off assignment is required to work around what
 550          * I consider to be a GCC-4 optimization bug.
 551          */
 552         if (uio->uio_offset < 0) {
 553                 hammer_done_transaction(&trans);
 554                 return (EFBIG);
 555         }
 556         base_offset = uio->uio_offset + uio->uio_resid; /* work around gcc-4 */
 557         if (uio->uio_resid > 0 && base_offset <= uio->uio_offset) {
 558                 hammer_done_transaction(&trans);
 559                 return (EFBIG);
 560         }
 561
 562         if (uio->uio_resid > 0 && (td = uio->uio_td) != NULL && td->td_proc &&
 563             base_offset > td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
 564                 hammer_done_transaction(&trans);
 565                 lwpsignal(td->td_proc, td->td_lwp, SIGXFSZ);
 566                 return (EFBIG);
 567         }
 568
 569         /*
 570          * If reading or writing a huge amount of data we have to break
 571          * atomicy and allow the operation to be interrupted by a signal
 572          * or it can DOS the machine.
 573          *
 574          * Preset redo_count so we stop generating REDOs earlier if the
 575          * limit is exceeded.
 576          *
 577          * redo_count is heuristical, SMP races are ok
 578          */
 579         bigwrite = (uio->uio_resid > 100 * 1024 * 1024);
 580         if ((ip->flags & HAMMER_INODE_REDO) &&
 581             ip->redo_count < hammer_limit_redo) {
 582                 ip->redo_count += uio->uio_resid;
 583         }
 584
 585         /*
 586          * Access the data typically in HAMMER_BUFSIZE blocks via the
 587          * buffer cache, but HAMMER may use a variable block size based
 588          * on the offset.
 589          */
 590         while (uio->uio_resid > 0) {
 591                 int fixsize = 0;
 592                 int blksize;
 593                 int blkmask;
 594                 int trivial;
 595                 int endofblk;
 596                 off_t nsize;
 597
 598                 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_WRITE)) != 0)
 599                         break;
 600                 if (bigwrite && (error = hammer_signal_check(hmp)) != 0)
 601                         break;
 602
 603                 blksize = hammer_blocksize(uio->uio_offset);
 604
 605                 /*
 606                  * Control the number of pending records associated with
 607                  * this inode.  If too many have accumulated start a
 608                  * flush.  Try to maintain a pipeline with the flusher.
 609                  *
 610                  * NOTE: It is possible for other sources to grow the
 611                  *       records but not necessarily issue another flush,
 612                  *       so use a timeout and ensure that a re-flush occurs.
 613                  */
 614                 if (ip->rsv_recs >= hammer_limit_inode_recs) {
 615                         lwkt_gettoken(&hmp->fs_token);
 616                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 617                         while (ip->rsv_recs >= hammer_limit_inode_recs * 2) {
 618                                 ip->flags |= HAMMER_INODE_RECSW;
 619                                 tsleep(&ip->rsv_recs, 0, "hmrwww", hz);
 620                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 621                         }
 622                         lwkt_reltoken(&hmp->fs_token);
 623                 }
 624
 625                 /*
 626                  * Do not allow HAMMER to blow out the buffer cache.  Very
 627                  * large UIOs can lockout other processes due to bwillwrite()
 628                  * mechanics.
 629                  *
 630                  * The hammer inode is not locked during these operations.
 631                  * The vnode is locked which can interfere with the pageout
 632                  * daemon for non-UIO_NOCOPY writes but should not interfere
 633                  * with the buffer cache.  Even so, we cannot afford to
 634                  * allow the pageout daemon to build up too many dirty buffer
 635                  * cache buffers.
 636                  *
 637                  * Only call this if we aren't being recursively called from
 638                  * a virtual disk device (vn), else we may deadlock.
 639                  */
 640                 if ((ap->a_ioflag & IO_RECURSE) == 0)
 641                         bwillwrite(blksize);
 642
 643                 /*
 644                  * Calculate the blocksize at the current offset and figure
 645                  * out how much we can actually write.
 646                  */
 647                 blkmask = blksize - 1;
 648                 offset = (int)uio->uio_offset & blkmask;
 649                 base_offset = uio->uio_offset & ~(int64_t)blkmask;
 650                 n = blksize - offset;
 651                 if (n > uio->uio_resid) {
 652                         n = uio->uio_resid;
 653                         endofblk = 0;
 654                 } else {
 655                         endofblk = 1;
 656                 }
 657                 nsize = uio->uio_offset + n;
 658                 if (nsize > ip->ino_data.size) {
 659                         if (uio->uio_offset > ip->ino_data.size)
 660                                 trivial = 0;
 661                         else
 662                                 trivial = 1;
 663                         nvextendbuf(ap->a_vp,
 664                                     ip->ino_data.size,
 665                                     nsize,
 666                                     hammer_blocksize(ip->ino_data.size),
 667                                     hammer_blocksize(nsize),
 668                                     hammer_blockoff(ip->ino_data.size),
 669                                     hammer_blockoff(nsize),
 670                                     trivial);
 671                         fixsize = 1;
 672                         kflags |= NOTE_EXTEND;
 673                 }
 674
 675                 if (uio->uio_segflg == UIO_NOCOPY) {
 676                         /*
 677                          * Issuing a write with the same data backing the
 678                          * buffer.  Instantiate the buffer to collect the
 679                          * backing vm pages, then read-in any missing bits.
 680                          *
 681                          * This case is used by vop_stdputpages().
 682                          */
 683                         bp = getblk(ap->a_vp, base_offset,
 684                                     blksize, GETBLK_BHEAVY, 0);
 685                         if ((bp->b_flags & B_CACHE) == 0) {
 686                                 bqrelse(bp);
 687                                 error = bread(ap->a_vp, base_offset,
 688                                               blksize, &bp);
 689                         }
 690                 } else if (offset == 0 && uio->uio_resid >= blksize) {
 691                         /*
 692                          * Even though we are entirely overwriting the buffer
 693                          * we may still have to zero it out to avoid a
 694                          * mmap/write visibility issue.
 695                          */
 696                         bp = getblk(ap->a_vp, base_offset, blksize, GETBLK_BHEAVY, 0);
 697                         if ((bp->b_flags & B_CACHE) == 0)
 698                                 vfs_bio_clrbuf(bp);
 699                 } else if (base_offset >= ip->ino_data.size) {
 700                         /*
 701                          * If the base offset of the buffer is beyond the
 702                          * file EOF, we don't have to issue a read.
 703                          */
 704                         bp = getblk(ap->a_vp, base_offset,
 705                                     blksize, GETBLK_BHEAVY, 0);
 706                         vfs_bio_clrbuf(bp);
 707                 } else {
 708                         /*
 709                          * Partial overwrite, read in any missing bits then
 710                          * replace the portion being written.
 711                          */
 712                         error = bread(ap->a_vp, base_offset, blksize, &bp);
 713                         if (error == 0)
 714                                 bheavy(bp);
 715                 }
 716                 if (error == 0)
 717                         error = uiomovebp(bp, bp->b_data + offset, n, uio);
 718
 719                 lwkt_gettoken(&hmp->fs_token);
 720
 721                 /*
 722                  * Generate REDO records if enabled and redo_count will not
 723                  * exceeded the limit.
 724                  *
 725                  * If redo_count exceeds the limit we stop generating records
 726                  * and clear HAMMER_INODE_REDO.  This will cause the next
 727                  * fsync() to do a full meta-data sync instead of just an
 728                  * UNDO/REDO fifo update.
 729                  *
 730                  * When clearing HAMMER_INODE_REDO any pre-existing REDOs
 731                  * will still be tracked.  The tracks will be terminated
 732                  * when the related meta-data (including possible data
 733                  * modifications which are not tracked via REDO) is
 734                  * flushed.
 735                  */
 736                 if ((ip->flags & HAMMER_INODE_REDO) && error == 0) {
 737                         if (ip->redo_count < hammer_limit_redo) {
 738                                 bp->b_flags |= B_VFSFLAG1;
 739                                 error = hammer_generate_redo(&trans, ip,
 740                                                      base_offset + offset,
 741                                                      HAMMER_REDO_WRITE,
 742                                                      bp->b_data + offset,
 743                                                      (size_t)n);
 744                         } else {
 745                                 ip->flags &= ~HAMMER_INODE_REDO;
 746                         }
 747                 }
 748
 749                 /*
 750                  * If we screwed up we have to undo any VM size changes we
 751                  * made.
 752                  */
 753                 if (error) {
 754                         brelse(bp);
 755                         if (fixsize) {
 756                                 nvtruncbuf(ap->a_vp, ip->ino_data.size,
 757                                           hammer_blocksize(ip->ino_data.size),
 758                                           hammer_blockoff(ip->ino_data.size),
 759                                           0);
 760                         }
 761                         lwkt_reltoken(&hmp->fs_token);
 762                         break;
 763                 }
 764                 kflags |= NOTE_WRITE;
 765                 hammer_stats_file_write += n;
 766                 if (blksize == HAMMER_XBUFSIZE)
 767                         bp->b_flags |= B_CLUSTEROK;
 768                 if (ip->ino_data.size < uio->uio_offset) {
 769                         ip->ino_data.size = uio->uio_offset;
 770                         flags = HAMMER_INODE_SDIRTY;
 771                 } else {
 772                         flags = 0;
 773                 }
 774                 ip->ino_data.mtime = trans.time;
 775                 flags |= HAMMER_INODE_MTIME | HAMMER_INODE_BUFS;
 776                 hammer_modify_inode(&trans, ip, flags);
 777
 778                 /*
 779                  * Once we dirty the buffer any cached zone-X offset
 780                  * becomes invalid.  HAMMER NOTE: no-history mode cannot
 781                  * allow overwriting over the same data sector unless
 782                  * we provide UNDOs for the old data, which we don't.
 783                  */
 784                 bp->b_bio2.bio_offset = NOOFFSET;
 785
 786                 lwkt_reltoken(&hmp->fs_token);
 787
 788                 /*
 789                  * Final buffer disposition.
 790                  *
 791                  * Because meta-data updates are deferred, HAMMER is
 792                  * especially sensitive to excessive bdwrite()s because
 793                  * the I/O stream is not broken up by disk reads.  So the
 794                  * buffer cache simply cannot keep up.
 795                  *
 796                  * WARNING!  blksize is variable.  cluster_write() is
 797                  *           expected to not blow up if it encounters
 798                  *           buffers that do not match the passed blksize.
 799                  *
 800                  * NOTE!  Hammer shouldn't need to bawrite()/cluster_write().
 801                  *        The ip->rsv_recs check should burst-flush the data.
 802                  *        If we queue it immediately the buf could be left
 803                  *        locked on the device queue for a very long time.
 804                  *
 805                  *        However, failing to flush a dirty buffer out when
 806                  *        issued from the pageout daemon can result in a low
 807                  *        memory deadlock against bio_page_alloc(), so we
 808                  *        have to bawrite() on IO_ASYNC as well.
 809                  *
 810                  * NOTE!  To avoid degenerate stalls due to mismatched block
 811                  *        sizes we only honor IO_DIRECT on the write which
 812                  *        abuts the end of the buffer.  However, we must
 813                  *        honor IO_SYNC in case someone is silly enough to
 814                  *        configure a HAMMER file as swap, or when HAMMER
 815                  *        is serving NFS (for commits).  Ick ick.
 816                  */
 817                 bp->b_flags |= B_AGE;
 818                 if (blksize == HAMMER_XBUFSIZE)
 819                         bp->b_flags |= B_CLUSTEROK;
 820
 821                 if (ap->a_ioflag & IO_SYNC) {
 822                         bwrite(bp);
 823                 } else if ((ap->a_ioflag & IO_DIRECT) && endofblk) {
 824                         bawrite(bp);
 825                 } else if (ap->a_ioflag & IO_ASYNC) {
 826                         bawrite(bp);
 827                 } else if (hammer_cluster_enable &&
 828                            !(ap->a_vp->v_mount->mnt_flag & MNT_NOCLUSTERW)) {
 829                         if (base_offset < HAMMER_XDEMARC)
 830                                 cluster_eof = hammer_blockdemarc(base_offset,
 831                                                          ip->ino_data.size);
 832                         else
 833                                 cluster_eof = ip->ino_data.size;
 834                         cluster_write(bp, cluster_eof, blksize, seqcount);
 835                 } else {
 836                         bdwrite(bp);
 837                 }
 838         }
 839         hammer_done_transaction(&trans);
 840         hammer_knote(ap->a_vp, kflags);
 841
 842         return (error);
 843 }
 844
 845 /*
 846  * hammer_vop_access { vp, mode, cred }
 847  *
 848  * MPSAFE - does not require fs_token
 849  */
 850 static
 851 int
 852 hammer_vop_access(struct vop_access_args *ap)
 853 {
 854         hammer_inode_t ip = VTOI(ap->a_vp);
 855         uid_t uid;
 856         gid_t gid;
 857         int error;
 858
 859         uid = hammer_to_unix_xid(&ip->ino_data.uid);
 860         gid = hammer_to_unix_xid(&ip->ino_data.gid);
 861
 862         error = vop_helper_access(ap, uid, gid, ip->ino_data.mode,
 863                                   ip->ino_data.uflags);
 864         return (error);
 865 }
 866
 867 /*
 868  * hammer_vop_advlock { vp, id, op, fl, flags }
 869  *
 870  * MPSAFE - does not require fs_token
 871  */
 872 static
 873 int
 874 hammer_vop_advlock(struct vop_advlock_args *ap)
 875 {
 876         hammer_inode_t ip = VTOI(ap->a_vp);
 877
 878         return (lf_advlock(ap, &ip->advlock, ip->ino_data.size));
 879 }
 880
 881 /*
 882  * hammer_vop_close { vp, fflag }
 883  *
 884  * We can only sync-on-close for normal closes.  XXX disabled for now.
 885  */
 886 static
 887 int
 888 hammer_vop_close(struct vop_close_args *ap)
 889 {
 890 #if 0
 891         struct vnode *vp = ap->a_vp;
 892         hammer_inode_t ip = VTOI(vp);
 893         int waitfor;
 894         if (ip->flags & (HAMMER_INODE_CLOSESYNC|HAMMER_INODE_CLOSEASYNC)) {
 895                 if (vn_islocked(vp) == LK_EXCLUSIVE &&
 896                     (vp->v_flag & (VINACTIVE|VRECLAIMED)) == 0) {
 897                         if (ip->flags & HAMMER_INODE_CLOSESYNC)
 898                                 waitfor = MNT_WAIT;
 899                         else
 900                                 waitfor = MNT_NOWAIT;
 901                         ip->flags &= ~(HAMMER_INODE_CLOSESYNC |
 902                                        HAMMER_INODE_CLOSEASYNC);
 903                         VOP_FSYNC(vp, MNT_NOWAIT, waitfor);
 904                 }
 905         }
 906 #endif
 907         return (vop_stdclose(ap));
 908 }
 909
 910 /*
 911  * hammer_vop_ncreate { nch, dvp, vpp, cred, vap }
 912  *
 913  * The operating system has already ensured that the directory entry
 914  * does not exist and done all appropriate namespace locking.
 915  */
 916 static
 917 int
 918 hammer_vop_ncreate(struct vop_ncreate_args *ap)
 919 {
 920         struct hammer_transaction trans;
 921         hammer_inode_t dip;
 922         hammer_inode_t nip;
 923         struct nchandle *nch;
 924         hammer_mount_t hmp;
 925         int error;
 926
 927         nch = ap->a_nch;
 928         dip = VTOI(ap->a_dvp);
 929         hmp = dip->hmp;
 930
 931         if (dip->flags & HAMMER_INODE_RO)
 932                 return (EROFS);
 933         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
 934                 return (error);
 935
 936         /*
 937          * Create a transaction to cover the operations we perform.
 938          */
 939         lwkt_gettoken(&hmp->fs_token);
 940         hammer_start_transaction(&trans, hmp);
 941
 942         /*
 943          * Create a new filesystem object of the requested type.  The
 944          * returned inode will be referenced and shared-locked to prevent
 945          * it from being moved to the flusher.
 946          */
 947         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
 948                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
 949                                     NULL, &nip);
 950         if (error) {
 951                 hkprintf("hammer_create_inode error %d\n", error);
 952                 hammer_done_transaction(&trans);
 953                 *ap->a_vpp = NULL;
 954                 lwkt_reltoken(&hmp->fs_token);
 955                 return (error);
 956         }
 957
 958         /*
 959          * Add the new filesystem object to the directory.  This will also
 960          * bump the inode's link count.
 961          */
 962         error = hammer_ip_add_direntry(&trans, dip,
 963                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
 964                                         nip);
 965         if (error)
 966                 hkprintf("hammer_ip_add_direntry error %d\n", error);
 967
 968         /*
 969          * Finish up.
 970          */
 971         if (error) {
 972                 hammer_rel_inode(nip, 0);
 973                 hammer_done_transaction(&trans);
 974                 *ap->a_vpp = NULL;
 975         } else {
 976                 error = hammer_get_vnode(nip, ap->a_vpp);
 977                 hammer_done_transaction(&trans);
 978                 hammer_rel_inode(nip, 0);
 979                 if (error == 0) {
 980                         cache_setunresolved(ap->a_nch);
 981                         cache_setvp(ap->a_nch, *ap->a_vpp);
 982                 }
 983                 hammer_knote(ap->a_dvp, NOTE_WRITE);
 984         }
 985         lwkt_reltoken(&hmp->fs_token);
 986         return (error);
 987 }
 988
 989 /*
 990  * hammer_vop_getattr { vp, vap }
 991  *
 992  * Retrieve an inode's attribute information.  When accessing inodes
 993  * historically we fake the atime field to ensure consistent results.
 994  * The atime field is stored in the B-Tree element and allowed to be
 995  * updated without cycling the element.
 996  *
 997  * MPSAFE - does not require fs_token
 998  */
 999 static
1000 int
1001 hammer_vop_getattr(struct vop_getattr_args *ap)
1002 {
1003         hammer_inode_t ip = VTOI(ap->a_vp);
1004         struct vattr *vap = ap->a_vap;
1005
1006         /*
1007          * We want the fsid to be different when accessing a filesystem
1008          * with different as-of's so programs like diff don't think
1009          * the files are the same.
1010          *
1011          * We also want the fsid to be the same when comparing snapshots,
1012          * or when comparing mirrors (which might be backed by different
1013          * physical devices).  HAMMER fsids are based on the PFS's
1014          * shared_uuid field.
1015          *
1016          * XXX there is a chance of collision here.  The va_fsid reported
1017          * by stat is different from the more involved fsid used in the
1018          * mount structure.
1019          */
1020         hammer_lock_sh(&ip->lock);
1021         vap->va_fsid = ip->pfsm->fsid_udev ^ (uint32_t)ip->obj_asof ^
1022                        (uint32_t)(ip->obj_asof >> 32);
1023
1024         vap->va_fileid = ip->ino_leaf.base.obj_id;
1025         vap->va_mode = ip->ino_data.mode;
1026         vap->va_nlink = ip->ino_data.nlinks;
1027         vap->va_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1028         vap->va_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1029         vap->va_rmajor = 0;
1030         vap->va_rminor = 0;
1031         vap->va_size = ip->ino_data.size;
1032
1033         /*
1034          * Special case for @@PFS softlinks.  The actual size of the
1035          * expanded softlink is "@@0x%016llx:%05d" == 26 bytes.
1036          * or for MAX_TID is    "@@-1:%05d" == 10 bytes.
1037          *
1038          * Note that userspace hammer command does not allow users to
1039          * create a @@PFS softlink under an existing other PFS (id!=0)
1040          * so the ip localization here for @@PFS softlink is always 0.
1041          */
1042         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_SOFTLINK &&
1043             ip->ino_data.size == 10 &&
1044             ip->obj_asof == HAMMER_MAX_TID &&
1045             ip->obj_localization == HAMMER_DEF_LOCALIZATION &&
1046             strncmp(ip->ino_data.ext.symlink, "@@PFS", 5) == 0) {
1047                 if (hammer_is_pfs_slave(&ip->pfsm->pfsd))
1048                         vap->va_size = 26;
1049                 else
1050                         vap->va_size = 10;
1051         }
1052
1053         /*
1054          * We must provide a consistent atime and mtime for snapshots
1055          * so people can do a 'tar cf - ... | md5' on them and get
1056          * consistent results.
1057          */
1058         if (ip->flags & HAMMER_INODE_RO) {
1059                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_atime);
1060                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_mtime);
1061         } else {
1062                 hammer_time_to_timespec(ip->ino_data.atime, &vap->va_atime);
1063                 hammer_time_to_timespec(ip->ino_data.mtime, &vap->va_mtime);
1064         }
1065         hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_ctime);
1066         vap->va_flags = ip->ino_data.uflags;
1067         vap->va_gen = 1;        /* hammer inums are unique for all time */
1068         vap->va_blocksize = HAMMER_BUFSIZE;
1069         if (ip->ino_data.size >= HAMMER_XDEMARC) {
1070                 vap->va_bytes = HAMMER_XBUFSIZE64_DOALIGN(ip->ino_data.size);
1071         } else if (ip->ino_data.size > HAMMER_HBUFSIZE) {
1072                 vap->va_bytes = HAMMER_BUFSIZE64_DOALIGN(ip->ino_data.size);
1073         } else {
1074                 vap->va_bytes = HAMMER_DATA_DOALIGN(ip->ino_data.size);
1075         }
1076
1077         vap->va_type = hammer_get_vnode_type(ip->ino_data.obj_type);
1078         vap->va_filerev = 0;    /* XXX */
1079         vap->va_uid_uuid = ip->ino_data.uid;
1080         vap->va_gid_uuid = ip->ino_data.gid;
1081         vap->va_fsid_uuid = ip->hmp->fsid;
1082         vap->va_vaflags = VA_UID_UUID_VALID | VA_GID_UUID_VALID |
1083                           VA_FSID_UUID_VALID;
1084
1085         switch (ip->ino_data.obj_type) {
1086         case HAMMER_OBJTYPE_CDEV:
1087         case HAMMER_OBJTYPE_BDEV:
1088                 vap->va_rmajor = ip->ino_data.rmajor;
1089                 vap->va_rminor = ip->ino_data.rminor;
1090                 break;
1091         default:
1092                 break;
1093         }
1094         hammer_unlock(&ip->lock);
1095         return(0);
1096 }
1097
1098 /*
1099  * hammer_vop_nresolve { nch, dvp, cred }
1100  *
1101  * Locate the requested directory entry.
1102  */
1103 static
1104 int
1105 hammer_vop_nresolve(struct vop_nresolve_args *ap)
1106 {
1107         struct hammer_transaction trans;
1108         struct namecache *ncp;
1109         hammer_mount_t hmp;
1110         hammer_inode_t dip;
1111         hammer_inode_t ip;
1112         hammer_tid_t asof;
1113         struct hammer_cursor cursor;
1114         struct vnode *vp;
1115         int64_t namekey;
1116         int error;
1117         int i;
1118         int nlen;
1119         int flags;
1120         int ispfs;
1121         int64_t obj_id;
1122         uint32_t localization;
1123         uint32_t max_iterations;
1124
1125         /*
1126          * Misc initialization, plus handle as-of name extensions.  Look for
1127          * the '@@' extension.  Note that as-of files and directories cannot
1128          * be modified.
1129          */
1130         dip = VTOI(ap->a_dvp);
1131         ncp = ap->a_nch->ncp;
1132         asof = dip->obj_asof;
1133         localization = dip->obj_localization;   /* for code consistency */
1134         nlen = ncp->nc_nlen;
1135         flags = dip->flags & HAMMER_INODE_RO;
1136         ispfs = 0;
1137         hmp = dip->hmp;
1138
1139         lwkt_gettoken(&hmp->fs_token);
1140         hammer_simple_transaction(&trans, hmp);
1141
1142         for (i = 0; i < nlen; ++i) {
1143                 if (ncp->nc_name[i] == '@' && ncp->nc_name[i+1] == '@') {
1144                         error = hammer_str_to_tid(ncp->nc_name + i + 2,
1145                                                   &ispfs, &asof, &localization);
1146                         if (error != 0) {
1147                                 i = nlen;
1148                                 break;
1149                         }
1150                         if (asof != HAMMER_MAX_TID)
1151                                 flags |= HAMMER_INODE_RO;
1152                         break;
1153                 }
1154         }
1155         nlen = i;
1156
1157         /*
1158          * If this is a PFS we dive into the PFS root inode
1159          */
1160         if (ispfs && nlen == 0) {
1161                 ip = hammer_get_inode(&trans, dip, HAMMER_OBJID_ROOT,
1162                                       asof, localization,
1163                                       flags, &error);
1164                 if (error == 0) {
1165                         error = hammer_get_vnode(ip, &vp);
1166                         hammer_rel_inode(ip, 0);
1167                 } else {
1168                         vp = NULL;
1169                 }
1170                 if (error == 0) {
1171                         vn_unlock(vp);
1172                         cache_setvp(ap->a_nch, vp);
1173                         vrele(vp);
1174                 }
1175                 goto done;
1176         }
1177
1178         /*
1179          * If there is no path component the time extension is relative to dip.
1180          * e.g. "fubar/@@<snapshot>"
1181          *
1182          * "." is handled by the kernel, but ".@@<snapshot>" is not.
1183          * e.g. "fubar/.@@<snapshot>"
1184          *
1185          * ".." is handled by the kernel.  We do not currently handle
1186          * "..@<snapshot>".
1187          */
1188         if (nlen == 0 || (nlen == 1 && ncp->nc_name[0] == '.')) {
1189                 ip = hammer_get_inode(&trans, dip, dip->obj_id,
1190                                       asof, dip->obj_localization,
1191                                       flags, &error);
1192                 if (error == 0) {
1193                         error = hammer_get_vnode(ip, &vp);
1194                         hammer_rel_inode(ip, 0);
1195                 } else {
1196                         vp = NULL;
1197                 }
1198                 if (error == 0) {
1199                         vn_unlock(vp);
1200                         cache_setvp(ap->a_nch, vp);
1201                         vrele(vp);
1202                 }
1203                 goto done;
1204         }
1205
1206         /*
1207          * Calculate the namekey and setup the key range for the scan.  This
1208          * works kinda like a chained hash table where the lower 32 bits
1209          * of the namekey synthesize the chain.
1210          *
1211          * The key range is inclusive of both key_beg and key_end.
1212          */
1213         namekey = hammer_direntry_namekey(dip, ncp->nc_name, nlen,
1214                                            &max_iterations);
1215
1216         error = hammer_init_cursor(&trans, &cursor, &dip->cache[1], dip);
1217         cursor.key_beg.localization = dip->obj_localization |
1218                                       hammer_dir_localization(dip);
1219         cursor.key_beg.obj_id = dip->obj_id;
1220         cursor.key_beg.key = namekey;
1221         cursor.key_beg.create_tid = 0;
1222         cursor.key_beg.delete_tid = 0;
1223         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1224         cursor.key_beg.obj_type = 0;
1225
1226         cursor.key_end = cursor.key_beg;
1227         cursor.key_end.key += max_iterations;
1228         cursor.asof = asof;
1229         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1230
1231         /*
1232          * Scan all matching records (the chain), locate the one matching
1233          * the requested path component.
1234          *
1235          * The hammer_ip_*() functions merge in-memory records with on-disk
1236          * records for the purposes of the search.
1237          */
1238         obj_id = 0;
1239         localization = HAMMER_DEF_LOCALIZATION;
1240
1241         if (error == 0) {
1242                 error = hammer_ip_first(&cursor);
1243                 while (error == 0) {
1244                         error = hammer_ip_resolve_data(&cursor);
1245                         if (error)
1246                                 break;
1247                         if (nlen == cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF &&
1248                             bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1249                                 obj_id = cursor.data->entry.obj_id;
1250                                 localization = cursor.data->entry.localization;
1251                                 break;
1252                         }
1253                         error = hammer_ip_next(&cursor);
1254                 }
1255         }
1256         hammer_done_cursor(&cursor);
1257
1258         /*
1259          * Lookup the obj_id.  This should always succeed.  If it does not
1260          * the filesystem may be damaged and we return a dummy inode.
1261          */
1262         if (error == 0) {
1263                 ip = hammer_get_inode(&trans, dip, obj_id,
1264                                       asof, localization,
1265                                       flags, &error);
1266                 if (error == ENOENT) {
1267                         hkprintf("WARNING: Missing inode for dirent \"%s\"\n"
1268                                 "\tobj_id = %016jx, asof=%016jx, lo=%08x\n",
1269                                 ncp->nc_name,
1270                                 (intmax_t)obj_id, (intmax_t)asof,
1271                                 localization);
1272                         error = 0;
1273                         ip = hammer_get_dummy_inode(&trans, dip, obj_id,
1274                                                     asof, localization,
1275                                                     flags, &error);
1276                 }
1277                 if (error == 0) {
1278                         error = hammer_get_vnode(ip, &vp);
1279                         hammer_rel_inode(ip, 0);
1280                 } else {
1281                         vp = NULL;
1282                 }
1283                 if (error == 0) {
1284                         vn_unlock(vp);
1285                         cache_setvp(ap->a_nch, vp);
1286                         vrele(vp);
1287                 }
1288         } else if (error == ENOENT) {
1289                 cache_setvp(ap->a_nch, NULL);
1290         }
1291 done:
1292         hammer_done_transaction(&trans);
1293         lwkt_reltoken(&hmp->fs_token);
1294         return (error);
1295 }
1296
1297 /*
1298  * hammer_vop_nlookupdotdot { dvp, vpp, cred }
1299  *
1300  * Locate the parent directory of a directory vnode.
1301  *
1302  * dvp is referenced but not locked.  *vpp must be returned referenced and
1303  * locked.  A parent_obj_id of 0 indicates that we are at the root.
1304  *
1305  * NOTE: as-of sequences are not linked into the directory structure.  If
1306  * we are at the root with a different asof then the mount point, reload
1307  * the same directory with the mount point's asof.   I'm not sure what this
1308  * will do to NFS.  We encode ASOF stamps in NFS file handles so it might not
1309  * get confused, but it hasn't been tested.
1310  */
1311 static
1312 int
1313 hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *ap)
1314 {
1315         struct hammer_transaction trans;
1316         hammer_inode_t dip;
1317         hammer_inode_t ip;
1318         hammer_mount_t hmp;
1319         int64_t parent_obj_id;
1320         uint32_t parent_obj_localization;
1321         hammer_tid_t asof;
1322         int error;
1323
1324         dip = VTOI(ap->a_dvp);
1325         asof = dip->obj_asof;
1326         hmp = dip->hmp;
1327
1328         /*
1329          * Whos are parent?  This could be the root of a pseudo-filesystem
1330          * whos parent is in another localization domain.
1331          */
1332         lwkt_gettoken(&hmp->fs_token);
1333         parent_obj_id = dip->ino_data.parent_obj_id;
1334         if (dip->obj_id == HAMMER_OBJID_ROOT)
1335                 parent_obj_localization = HAMMER_DEF_LOCALIZATION;
1336         else
1337                 parent_obj_localization = dip->obj_localization;
1338
1339         /*
1340          * It's probably a PFS root when dip->ino_data.parent_obj_id is 0.
1341          */
1342         if (parent_obj_id == 0) {
1343                 if (dip->obj_id == HAMMER_OBJID_ROOT &&
1344                    asof != hmp->asof) {
1345                         parent_obj_id = dip->obj_id;
1346                         asof = hmp->asof;
1347                         *ap->a_fakename = kmalloc(19, M_TEMP, M_WAITOK);
1348                         ksnprintf(*ap->a_fakename, 19, "0x%016jx",
1349                                   (intmax_t)dip->obj_asof);
1350                 } else {
1351                         *ap->a_vpp = NULL;
1352                         lwkt_reltoken(&hmp->fs_token);
1353                         return ENOENT;
1354                 }
1355         }
1356
1357         hammer_simple_transaction(&trans, hmp);
1358
1359         ip = hammer_get_inode(&trans, dip, parent_obj_id,
1360                               asof, parent_obj_localization,
1361                               dip->flags, &error);
1362         if (ip) {
1363                 error = hammer_get_vnode(ip, ap->a_vpp);
1364                 hammer_rel_inode(ip, 0);
1365         } else {
1366                 *ap->a_vpp = NULL;
1367         }
1368         hammer_done_transaction(&trans);
1369         lwkt_reltoken(&hmp->fs_token);
1370         return (error);
1371 }
1372
1373 /*
1374  * hammer_vop_nlink { nch, dvp, vp, cred }
1375  */
1376 static
1377 int
1378 hammer_vop_nlink(struct vop_nlink_args *ap)
1379 {
1380         struct hammer_transaction trans;
1381         hammer_inode_t dip;
1382         hammer_inode_t ip;
1383         struct nchandle *nch;
1384         hammer_mount_t hmp;
1385         int error;
1386
1387         if (ap->a_dvp->v_mount != ap->a_vp->v_mount)
1388                 return(EXDEV);
1389
1390         nch = ap->a_nch;
1391         dip = VTOI(ap->a_dvp);
1392         ip = VTOI(ap->a_vp);
1393         hmp = dip->hmp;
1394
1395         if (dip->obj_localization != ip->obj_localization)
1396                 return(EXDEV);
1397
1398         if (dip->flags & HAMMER_INODE_RO)
1399                 return (EROFS);
1400         if (ip->flags & HAMMER_INODE_RO)
1401                 return (EROFS);
1402         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1403                 return (error);
1404
1405         /*
1406          * Create a transaction to cover the operations we perform.
1407          */
1408         lwkt_gettoken(&hmp->fs_token);
1409         hammer_start_transaction(&trans, hmp);
1410
1411         /*
1412          * Add the filesystem object to the directory.  Note that neither
1413          * dip nor ip are referenced or locked, but their vnodes are
1414          * referenced.  This function will bump the inode's link count.
1415          */
1416         error = hammer_ip_add_direntry(&trans, dip,
1417                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1418                                         ip);
1419
1420         /*
1421          * Finish up.
1422          */
1423         if (error == 0) {
1424                 cache_setunresolved(nch);
1425                 cache_setvp(nch, ap->a_vp);
1426         }
1427         hammer_done_transaction(&trans);
1428         hammer_knote(ap->a_vp, NOTE_LINK);
1429         hammer_knote(ap->a_dvp, NOTE_WRITE);
1430         lwkt_reltoken(&hmp->fs_token);
1431         return (error);
1432 }
1433
1434 /*
1435  * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
1436  *
1437  * The operating system has already ensured that the directory entry
1438  * does not exist and done all appropriate namespace locking.
1439  */
1440 static
1441 int
1442 hammer_vop_nmkdir(struct vop_nmkdir_args *ap)
1443 {
1444         struct hammer_transaction trans;
1445         hammer_inode_t dip;
1446         hammer_inode_t nip;
1447         struct nchandle *nch;
1448         hammer_mount_t hmp;
1449         int error;
1450
1451         nch = ap->a_nch;
1452         dip = VTOI(ap->a_dvp);
1453         hmp = dip->hmp;
1454
1455         if (dip->flags & HAMMER_INODE_RO)
1456                 return (EROFS);
1457         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1458                 return (error);
1459
1460         /*
1461          * Create a transaction to cover the operations we perform.
1462          */
1463         lwkt_gettoken(&hmp->fs_token);
1464         hammer_start_transaction(&trans, hmp);
1465
1466         /*
1467          * Create a new filesystem object of the requested type.  The
1468          * returned inode will be referenced but not locked.
1469          */
1470         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1471                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1472                                     NULL, &nip);
1473         if (error) {
1474                 hammer_done_transaction(&trans);
1475                 *ap->a_vpp = NULL;
1476                 lwkt_reltoken(&hmp->fs_token);
1477                 return (error);
1478         }
1479         /*
1480          * Add the new filesystem object to the directory.  This will also
1481          * bump the inode's link count.
1482          */
1483         error = hammer_ip_add_direntry(&trans, dip,
1484                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1485                                         nip);
1486         if (error)
1487                 hkprintf("hammer_mkdir (add) error %d\n", error);
1488
1489         /*
1490          * Finish up.
1491          */
1492         if (error) {
1493                 hammer_rel_inode(nip, 0);
1494                 *ap->a_vpp = NULL;
1495         } else {
1496                 error = hammer_get_vnode(nip, ap->a_vpp);
1497                 hammer_rel_inode(nip, 0);
1498                 if (error == 0) {
1499                         cache_setunresolved(ap->a_nch);
1500                         cache_setvp(ap->a_nch, *ap->a_vpp);
1501                 }
1502         }
1503         hammer_done_transaction(&trans);
1504         if (error == 0)
1505                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
1506         lwkt_reltoken(&hmp->fs_token);
1507         return (error);
1508 }
1509
1510 /*
1511  * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1512  *
1513  * The operating system has already ensured that the directory entry
1514  * does not exist and done all appropriate namespace locking.
1515  */
1516 static
1517 int
1518 hammer_vop_nmknod(struct vop_nmknod_args *ap)
1519 {
1520         struct hammer_transaction trans;
1521         hammer_inode_t dip;
1522         hammer_inode_t nip;
1523         struct nchandle *nch;
1524         hammer_mount_t hmp;
1525         int error;
1526
1527         nch = ap->a_nch;
1528         dip = VTOI(ap->a_dvp);
1529         hmp = dip->hmp;
1530
1531         if (dip->flags & HAMMER_INODE_RO)
1532                 return (EROFS);
1533         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1534                 return (error);
1535
1536         /*
1537          * Create a transaction to cover the operations we perform.
1538          */
1539         lwkt_gettoken(&hmp->fs_token);
1540         hammer_start_transaction(&trans, hmp);
1541
1542         /*
1543          * Create a new filesystem object of the requested type.  The
1544          * returned inode will be referenced but not locked.
1545          *
1546          * If mknod specifies a directory a pseudo-fs is created.
1547          */
1548         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1549                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1550                                     NULL, &nip);
1551         if (error) {
1552                 hammer_done_transaction(&trans);
1553                 *ap->a_vpp = NULL;
1554                 lwkt_reltoken(&hmp->fs_token);
1555                 return (error);
1556         }
1557
1558         /*
1559          * Add the new filesystem object to the directory.  This will also
1560          * bump the inode's link count.
1561          */
1562         error = hammer_ip_add_direntry(&trans, dip,
1563                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1564                                         nip);
1565
1566         /*
1567          * Finish up.
1568          */
1569         if (error) {
1570                 hammer_rel_inode(nip, 0);
1571                 *ap->a_vpp = NULL;
1572         } else {
1573                 error = hammer_get_vnode(nip, ap->a_vpp);
1574                 hammer_rel_inode(nip, 0);
1575                 if (error == 0) {
1576                         cache_setunresolved(ap->a_nch);
1577                         cache_setvp(ap->a_nch, *ap->a_vpp);
1578                 }
1579         }
1580         hammer_done_transaction(&trans);
1581         if (error == 0)
1582                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1583         lwkt_reltoken(&hmp->fs_token);
1584         return (error);
1585 }
1586
1587 /*
1588  * hammer_vop_open { vp, mode, cred, fp }
1589  *
1590  * MPSAFE (does not require fs_token)
1591  */
1592 static
1593 int
1594 hammer_vop_open(struct vop_open_args *ap)
1595 {
1596         hammer_inode_t ip;
1597
1598         ip = VTOI(ap->a_vp);
1599
1600         if ((ap->a_mode & FWRITE) && (ip->flags & HAMMER_INODE_RO))
1601                 return (EROFS);
1602         return(vop_stdopen(ap));
1603 }
1604
1605 /*
1606  * hammer_vop_print { vp }
1607  */
1608 static
1609 int
1610 hammer_vop_print(struct vop_print_args *ap)
1611 {
1612         return EOPNOTSUPP;
1613 }
1614
1615 /*
1616  * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1617  */
1618 static
1619 int
1620 hammer_vop_readdir(struct vop_readdir_args *ap)
1621 {
1622         struct hammer_transaction trans;
1623         struct hammer_cursor cursor;
1624         hammer_inode_t ip;
1625         hammer_mount_t hmp;
1626         struct uio *uio;
1627         hammer_base_elm_t base;
1628         int error;
1629         int cookie_index;
1630         int ncookies;
1631         off_t *cookies;
1632         off_t saveoff;
1633         int r;
1634         int dtype;
1635
1636         ip = VTOI(ap->a_vp);
1637         uio = ap->a_uio;
1638         saveoff = uio->uio_offset;
1639         hmp = ip->hmp;
1640
1641         if (ap->a_ncookies) {
1642                 ncookies = uio->uio_resid / 16 + 1;
1643                 if (ncookies > 1024)
1644                         ncookies = 1024;
1645                 cookies = kmalloc(ncookies * sizeof(off_t), M_TEMP, M_WAITOK);
1646                 cookie_index = 0;
1647         } else {
1648                 ncookies = -1;
1649                 cookies = NULL;
1650                 cookie_index = 0;
1651         }
1652
1653         lwkt_gettoken(&hmp->fs_token);
1654         hammer_simple_transaction(&trans, hmp);
1655
1656         /*
1657          * Handle artificial entries
1658          *
1659          * It should be noted that the minimum value for a directory
1660          * hash key on-media is 0x0000000100000000, so we can use anything
1661          * less then that to represent our 'special' key space.
1662          */
1663         error = 0;
1664         if (saveoff == 0) {
1665                 r = vop_write_dirent(&error, uio, ip->obj_id, DT_DIR, 1, ".");
1666                 if (r)
1667                         goto done;
1668                 if (cookies)
1669                         cookies[cookie_index] = saveoff;
1670                 ++saveoff;
1671                 ++cookie_index;
1672                 if (cookie_index == ncookies)
1673                         goto done;
1674         }
1675         if (saveoff == 1) {
1676                 if (ip->ino_data.parent_obj_id) {
1677                         r = vop_write_dirent(&error, uio,
1678                                              ip->ino_data.parent_obj_id,
1679                                              DT_DIR, 2, "..");
1680                 } else {
1681                         r = vop_write_dirent(&error, uio,
1682                                              ip->obj_id, DT_DIR, 2, "..");
1683                 }
1684                 if (r)
1685                         goto done;
1686                 if (cookies)
1687                         cookies[cookie_index] = saveoff;
1688                 ++saveoff;
1689                 ++cookie_index;
1690                 if (cookie_index == ncookies)
1691                         goto done;
1692         }
1693
1694         /*
1695          * Key range (begin and end inclusive) to scan.  Directory keys
1696          * directly translate to a 64 bit 'seek' position.
1697          */
1698         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1699         cursor.key_beg.localization = ip->obj_localization |
1700                                       hammer_dir_localization(ip);
1701         cursor.key_beg.obj_id = ip->obj_id;
1702         cursor.key_beg.create_tid = 0;
1703         cursor.key_beg.delete_tid = 0;
1704         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1705         cursor.key_beg.obj_type = 0;
1706         cursor.key_beg.key = saveoff;
1707
1708         cursor.key_end = cursor.key_beg;
1709         cursor.key_end.key = HAMMER_MAX_KEY;
1710         cursor.asof = ip->obj_asof;
1711         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1712
1713         error = hammer_ip_first(&cursor);
1714
1715         while (error == 0) {
1716                 error = hammer_ip_resolve_data(&cursor);
1717                 if (error)
1718                         break;
1719                 base = &cursor.leaf->base;
1720                 saveoff = base->key;
1721                 KKASSERT(cursor.leaf->data_len > HAMMER_ENTRY_NAME_OFF);
1722
1723                 if (base->obj_id != ip->obj_id)
1724                         hpanic("bad record at %p", cursor.node);
1725
1726                 dtype = hammer_get_dtype(cursor.leaf->base.obj_type);
1727                 r = vop_write_dirent(
1728                              &error, uio, cursor.data->entry.obj_id,
1729                              dtype,
1730                              cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF ,
1731                              (void *)cursor.data->entry.name);
1732                 if (r)
1733                         break;
1734                 ++saveoff;
1735                 if (cookies)
1736                         cookies[cookie_index] = base->key;
1737                 ++cookie_index;
1738                 if (cookie_index == ncookies)
1739                         break;
1740                 error = hammer_ip_next(&cursor);
1741         }
1742         hammer_done_cursor(&cursor);
1743
1744 done:
1745         hammer_done_transaction(&trans);
1746
1747         if (ap->a_eofflag)
1748                 *ap->a_eofflag = (error == ENOENT);
1749         uio->uio_offset = saveoff;
1750         if (error && cookie_index == 0) {
1751                 if (error == ENOENT)
1752                         error = 0;
1753                 if (cookies) {
1754                         kfree(cookies, M_TEMP);
1755                         *ap->a_ncookies = 0;
1756                         *ap->a_cookies = NULL;
1757                 }
1758         } else {
1759                 if (error == ENOENT)
1760                         error = 0;
1761                 if (cookies) {
1762                         *ap->a_ncookies = cookie_index;
1763                         *ap->a_cookies = cookies;
1764                 }
1765         }
1766         lwkt_reltoken(&hmp->fs_token);
1767         return(error);
1768 }
1769
1770 /*
1771  * hammer_vop_readlink { vp, uio, cred }
1772  */
1773 static
1774 int
1775 hammer_vop_readlink(struct vop_readlink_args *ap)
1776 {
1777         struct hammer_transaction trans;
1778         struct hammer_cursor cursor;
1779         hammer_inode_t ip;
1780         hammer_mount_t hmp;
1781         char buf[32];
1782         uint32_t localization;
1783         hammer_pseudofs_inmem_t pfsm;
1784         int error;
1785
1786         ip = VTOI(ap->a_vp);
1787         hmp = ip->hmp;
1788
1789         lwkt_gettoken(&hmp->fs_token);
1790
1791         /*
1792          * Shortcut if the symlink data was stuffed into ino_data.
1793          *
1794          * Also expand special "@@PFS%05d" softlinks (expansion only
1795          * occurs for non-historical (current) accesses made from the
1796          * primary filesystem).
1797          *
1798          * Note that userspace hammer command does not allow users to
1799          * create a @@PFS softlink under an existing other PFS (id!=0)
1800          * so the ip localization here for @@PFS softlink is always 0.
1801          */
1802         if (ip->ino_data.size <= HAMMER_INODE_BASESYMLEN) {
1803                 char *ptr;
1804                 int bytes;
1805
1806                 ptr = ip->ino_data.ext.symlink;
1807                 bytes = (int)ip->ino_data.size;
1808                 if (bytes == 10 &&
1809                     ip->obj_asof == HAMMER_MAX_TID &&
1810                     ip->obj_localization == HAMMER_DEF_LOCALIZATION &&
1811                     strncmp(ptr, "@@PFS", 5) == 0) {
1812                         hammer_simple_transaction(&trans, hmp);
1813                         bcopy(ptr + 5, buf, 5);
1814                         buf[5] = 0;
1815                         localization = pfs_to_lo(strtoul(buf, NULL, 10));
1816                         pfsm = hammer_load_pseudofs(&trans, localization,
1817                                                     &error);
1818                         if (error == 0) {
1819                                 if (hammer_is_pfs_slave(&pfsm->pfsd)) {
1820                                         /* vap->va_size == 26 */
1821                                         ksnprintf(buf, sizeof(buf),
1822                                                   "@@0x%016jx:%05d",
1823                                                   (intmax_t)pfsm->pfsd.sync_end_tid,
1824                                                   lo_to_pfs(localization));
1825                                 } else {
1826                                         /* vap->va_size == 10 */
1827                                         ksnprintf(buf, sizeof(buf),
1828                                                   "@@-1:%05d",
1829                                                   lo_to_pfs(localization));
1830                                 }
1831                                 ptr = buf;
1832                                 bytes = strlen(buf);
1833                         }
1834                         if (pfsm)
1835                                 hammer_rel_pseudofs(hmp, pfsm);
1836                         hammer_done_transaction(&trans);
1837                 }
1838                 error = uiomove(ptr, bytes, ap->a_uio);
1839                 lwkt_reltoken(&hmp->fs_token);
1840                 return(error);
1841         }
1842
1843         /*
1844          * Long version
1845          */
1846         hammer_simple_transaction(&trans, hmp);
1847         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1848
1849         /*
1850          * Key range (begin and end inclusive) to scan.  Directory keys
1851          * directly translate to a 64 bit 'seek' position.
1852          */
1853         cursor.key_beg.localization = ip->obj_localization |
1854                                       HAMMER_LOCALIZE_MISC;
1855         cursor.key_beg.obj_id = ip->obj_id;
1856         cursor.key_beg.create_tid = 0;
1857         cursor.key_beg.delete_tid = 0;
1858         cursor.key_beg.rec_type = HAMMER_RECTYPE_FIX;
1859         cursor.key_beg.obj_type = 0;
1860         cursor.key_beg.key = HAMMER_FIXKEY_SYMLINK;
1861         cursor.asof = ip->obj_asof;
1862         cursor.flags |= HAMMER_CURSOR_ASOF;
1863
1864         error = hammer_ip_lookup(&cursor);
1865         if (error == 0) {
1866                 error = hammer_ip_resolve_data(&cursor);
1867                 if (error == 0) {
1868                         KKASSERT(cursor.leaf->data_len >=
1869                                  HAMMER_SYMLINK_NAME_OFF);
1870                         error = uiomove(cursor.data->symlink.name,
1871                                         cursor.leaf->data_len -
1872                                                 HAMMER_SYMLINK_NAME_OFF,
1873                                         ap->a_uio);
1874                 }
1875         }
1876         hammer_done_cursor(&cursor);
1877         hammer_done_transaction(&trans);
1878         lwkt_reltoken(&hmp->fs_token);
1879         return(error);
1880 }
1881
1882 /*
1883  * hammer_vop_nremove { nch, dvp, cred }
1884  */
1885 static
1886 int
1887 hammer_vop_nremove(struct vop_nremove_args *ap)
1888 {
1889         struct hammer_transaction trans;
1890         hammer_inode_t dip;
1891         hammer_mount_t hmp;
1892         int error;
1893
1894         dip = VTOI(ap->a_dvp);
1895         hmp = dip->hmp;
1896
1897         if (hammer_nohistory(dip) == 0 &&
1898             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1899                 return (error);
1900         }
1901
1902         lwkt_gettoken(&hmp->fs_token);
1903         hammer_start_transaction(&trans, hmp);
1904         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 0);
1905         hammer_done_transaction(&trans);
1906         if (error == 0)
1907                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1908         lwkt_reltoken(&hmp->fs_token);
1909         return (error);
1910 }
1911
1912 /*
1913  * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1914  */
1915 static
1916 int
1917 hammer_vop_nrename(struct vop_nrename_args *ap)
1918 {
1919         struct hammer_transaction trans;
1920         struct namecache *fncp;
1921         struct namecache *tncp;
1922         hammer_inode_t fdip;
1923         hammer_inode_t tdip;
1924         hammer_inode_t ip;
1925         hammer_mount_t hmp;
1926         struct hammer_cursor cursor;
1927         int64_t namekey;
1928         uint32_t max_iterations;
1929         int nlen, error;
1930
1931         if (ap->a_fdvp->v_mount != ap->a_tdvp->v_mount)
1932                 return(EXDEV);
1933         if (ap->a_fdvp->v_mount != ap->a_fnch->ncp->nc_vp->v_mount)
1934                 return(EXDEV);
1935
1936         fdip = VTOI(ap->a_fdvp);
1937         tdip = VTOI(ap->a_tdvp);
1938         fncp = ap->a_fnch->ncp;
1939         tncp = ap->a_tnch->ncp;
1940         ip = VTOI(fncp->nc_vp);
1941         KKASSERT(ip != NULL);
1942
1943         hmp = ip->hmp;
1944
1945         if (fdip->obj_localization != tdip->obj_localization)
1946                 return(EXDEV);
1947         if (fdip->obj_localization != ip->obj_localization)
1948                 return(EXDEV);
1949
1950         if (fdip->flags & HAMMER_INODE_RO)
1951                 return (EROFS);
1952         if (tdip->flags & HAMMER_INODE_RO)
1953                 return (EROFS);
1954         if (ip->flags & HAMMER_INODE_RO)
1955                 return (EROFS);
1956         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1957                 return (error);
1958
1959         lwkt_gettoken(&hmp->fs_token);
1960         hammer_start_transaction(&trans, hmp);
1961
1962         /*
1963          * Remove tncp from the target directory and then link ip as
1964          * tncp. XXX pass trans to dounlink
1965          *
1966          * Force the inode sync-time to match the transaction so it is
1967          * in-sync with the creation of the target directory entry.
1968          */
1969         error = hammer_dounlink(&trans, ap->a_tnch, ap->a_tdvp,
1970                                 ap->a_cred, 0, -1);
1971         if (error == 0 || error == ENOENT) {
1972                 error = hammer_ip_add_direntry(&trans, tdip,
1973                                                 tncp->nc_name, tncp->nc_nlen,
1974                                                 ip);
1975                 if (error == 0) {
1976                         ip->ino_data.parent_obj_id = tdip->obj_id;
1977                         ip->ino_data.ctime = trans.time;
1978                         hammer_modify_inode(&trans, ip, HAMMER_INODE_DDIRTY);
1979                 }
1980         }
1981         if (error)
1982                 goto failed; /* XXX */
1983
1984         /*
1985          * Locate the record in the originating directory and remove it.
1986          *
1987          * Calculate the namekey and setup the key range for the scan.  This
1988          * works kinda like a chained hash table where the lower 32 bits
1989          * of the namekey synthesize the chain.
1990          *
1991          * The key range is inclusive of both key_beg and key_end.
1992          */
1993         namekey = hammer_direntry_namekey(fdip, fncp->nc_name, fncp->nc_nlen,
1994                                            &max_iterations);
1995 retry:
1996         hammer_init_cursor(&trans, &cursor, &fdip->cache[1], fdip);
1997         cursor.key_beg.localization = fdip->obj_localization |
1998                                       hammer_dir_localization(fdip);
1999         cursor.key_beg.obj_id = fdip->obj_id;
2000         cursor.key_beg.key = namekey;
2001         cursor.key_beg.create_tid = 0;
2002         cursor.key_beg.delete_tid = 0;
2003         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
2004         cursor.key_beg.obj_type = 0;
2005
2006         cursor.key_end = cursor.key_beg;
2007         cursor.key_end.key += max_iterations;
2008         cursor.asof = fdip->obj_asof;
2009         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2010
2011         /*
2012          * Scan all matching records (the chain), locate the one matching
2013          * the requested path component.
2014          *
2015          * The hammer_ip_*() functions merge in-memory records with on-disk
2016          * records for the purposes of the search.
2017          */
2018         error = hammer_ip_first(&cursor);
2019         while (error == 0) {
2020                 if (hammer_ip_resolve_data(&cursor) != 0)
2021                         break;
2022                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
2023                 KKASSERT(nlen > 0);
2024                 if (fncp->nc_nlen == nlen &&
2025                     bcmp(fncp->nc_name, cursor.data->entry.name, nlen) == 0) {
2026                         break;
2027                 }
2028                 error = hammer_ip_next(&cursor);
2029         }
2030
2031         /*
2032          * If all is ok we have to get the inode so we can adjust nlinks.
2033          *
2034          * WARNING: hammer_ip_del_direntry() may have to terminate the
2035          * cursor to avoid a recursion.  It's ok to call hammer_done_cursor()
2036          * twice.
2037          */
2038         if (error == 0)
2039                 error = hammer_ip_del_direntry(&trans, &cursor, fdip, ip);
2040
2041         /*
2042          * XXX A deadlock here will break rename's atomicy for the purposes
2043          * of crash recovery.
2044          */
2045         if (error == EDEADLK) {
2046                 hammer_done_cursor(&cursor);
2047                 goto retry;
2048         }
2049
2050         /*
2051          * Cleanup and tell the kernel that the rename succeeded.
2052          *
2053          * NOTE: ip->vp, if non-NULL, cannot be directly referenced
2054          *       without formally acquiring the vp since the vp might
2055          *       have zero refs on it, or in the middle of a reclaim,
2056          *       etc.
2057          */
2058         hammer_done_cursor(&cursor);
2059         if (error == 0) {
2060                 cache_rename(ap->a_fnch, ap->a_tnch);
2061                 hammer_knote(ap->a_fdvp, NOTE_WRITE);
2062                 hammer_knote(ap->a_tdvp, NOTE_WRITE);
2063                 while (ip->vp) {
2064                         struct vnode *vp;
2065
2066                         error = hammer_get_vnode(ip, &vp);
2067                         if (error == 0 && vp) {
2068                                 vn_unlock(vp);
2069                                 hammer_knote(ip->vp, NOTE_RENAME);
2070                                 vrele(vp);
2071                                 break;
2072                         }
2073                         hdkprintf("ip/vp race2 avoided\n");
2074                 }
2075         }
2076
2077 failed:
2078         hammer_done_transaction(&trans);
2079         lwkt_reltoken(&hmp->fs_token);
2080         return (error);
2081 }
2082
2083 /*
2084  * hammer_vop_nrmdir { nch, dvp, cred }
2085  */
2086 static
2087 int
2088 hammer_vop_nrmdir(struct vop_nrmdir_args *ap)
2089 {
2090         struct hammer_transaction trans;
2091         hammer_inode_t dip;
2092         hammer_mount_t hmp;
2093         int error;
2094
2095         dip = VTOI(ap->a_dvp);
2096         hmp = dip->hmp;
2097
2098         if (hammer_nohistory(dip) == 0 &&
2099             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2100                 return (error);
2101         }
2102
2103         lwkt_gettoken(&hmp->fs_token);
2104         hammer_start_transaction(&trans, hmp);
2105         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 1);
2106         hammer_done_transaction(&trans);
2107         if (error == 0)
2108                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
2109         lwkt_reltoken(&hmp->fs_token);
2110         return (error);
2111 }
2112
2113 /*
2114  * hammer_vop_markatime { vp, cred }
2115  */
2116 static
2117 int
2118 hammer_vop_markatime(struct vop_markatime_args *ap)
2119 {
2120         struct hammer_transaction trans;
2121         hammer_inode_t ip;
2122         hammer_mount_t hmp;
2123
2124         ip = VTOI(ap->a_vp);
2125         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2126                 return (EROFS);
2127         if (ip->flags & HAMMER_INODE_RO)
2128                 return (EROFS);
2129         hmp = ip->hmp;
2130         if (hmp->mp->mnt_flag & MNT_NOATIME)
2131                 return (0);
2132         lwkt_gettoken(&hmp->fs_token);
2133         hammer_start_transaction(&trans, hmp);
2134
2135         ip->ino_data.atime = trans.time;
2136         hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
2137         hammer_done_transaction(&trans);
2138         hammer_knote(ap->a_vp, NOTE_ATTRIB);
2139         lwkt_reltoken(&hmp->fs_token);
2140         return (0);
2141 }
2142
2143 /*
2144  * hammer_vop_setattr { vp, vap, cred }
2145  */
2146 static
2147 int
2148 hammer_vop_setattr(struct vop_setattr_args *ap)
2149 {
2150         struct hammer_transaction trans;
2151         hammer_inode_t ip;
2152         struct vattr *vap;
2153         hammer_mount_t hmp;
2154         int modflags;
2155         int error;
2156         int truncating;
2157         int blksize;
2158         int kflags;
2159 #if 0
2160         int64_t aligned_size;
2161 #endif
2162         uint32_t flags;
2163
2164         vap = ap->a_vap;
2165         ip = ap->a_vp->v_data;
2166         modflags = 0;
2167         kflags = 0;
2168         hmp = ip->hmp;
2169
2170         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2171                 return(EROFS);
2172         if (ip->flags & HAMMER_INODE_RO)
2173                 return (EROFS);
2174         if (hammer_nohistory(ip) == 0 &&
2175             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2176                 return (error);
2177         }
2178
2179         lwkt_gettoken(&hmp->fs_token);
2180         hammer_start_transaction(&trans, hmp);
2181         error = 0;
2182
2183         if (vap->va_flags != VNOVAL) {
2184                 flags = ip->ino_data.uflags;
2185                 error = vop_helper_setattr_flags(&flags, vap->va_flags,
2186                                          hammer_to_unix_xid(&ip->ino_data.uid),
2187                                          ap->a_cred);
2188                 if (error == 0) {
2189                         if (ip->ino_data.uflags != flags) {
2190                                 ip->ino_data.uflags = flags;
2191                                 ip->ino_data.ctime = trans.time;
2192                                 modflags |= HAMMER_INODE_DDIRTY;
2193                                 kflags |= NOTE_ATTRIB;
2194                         }
2195                         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2196                                 error = 0;
2197                                 goto done;
2198                         }
2199                 }
2200                 goto done;
2201         }
2202         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2203                 error = EPERM;
2204                 goto done;
2205         }
2206         if (vap->va_uid != (uid_t)VNOVAL || vap->va_gid != (gid_t)VNOVAL) {
2207                 mode_t cur_mode = ip->ino_data.mode;
2208                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2209                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2210                 hammer_uuid_t uuid_uid;
2211                 hammer_uuid_t uuid_gid;
2212
2213                 error = vop_helper_chown(ap->a_vp, vap->va_uid, vap->va_gid,
2214                                          ap->a_cred,
2215                                          &cur_uid, &cur_gid, &cur_mode);
2216                 if (error == 0) {
2217                         hammer_guid_to_uuid(&uuid_uid, cur_uid);
2218                         hammer_guid_to_uuid(&uuid_gid, cur_gid);
2219                         if (kuuid_compare(&uuid_uid, &ip->ino_data.uid) ||
2220                             kuuid_compare(&uuid_gid, &ip->ino_data.gid) ||
2221                             ip->ino_data.mode != cur_mode) {
2222                                 ip->ino_data.uid = uuid_uid;
2223                                 ip->ino_data.gid = uuid_gid;
2224                                 ip->ino_data.mode = cur_mode;
2225                                 ip->ino_data.ctime = trans.time;
2226                                 modflags |= HAMMER_INODE_DDIRTY;
2227                         }
2228                         kflags |= NOTE_ATTRIB;
2229                 }
2230         }
2231         while (vap->va_size != VNOVAL && ip->ino_data.size != vap->va_size) {
2232                 switch(ap->a_vp->v_type) {
2233                 case VREG:
2234                         if (vap->va_size == ip->ino_data.size)
2235                                 break;
2236
2237                         /*
2238                          * Log the operation if in fast-fsync mode or if
2239                          * there are unterminated redo write records present.
2240                          *
2241                          * The second check is needed so the recovery code
2242                          * properly truncates write redos even if nominal
2243                          * REDO operations is turned off due to excessive
2244                          * writes, because the related records might be
2245                          * destroyed and never lay down a TERM_WRITE.
2246                          */
2247                         if ((ip->flags & HAMMER_INODE_REDO) ||
2248                             (ip->flags & HAMMER_INODE_RDIRTY)) {
2249                                 error = hammer_generate_redo(&trans, ip,
2250                                                              vap->va_size,
2251                                                              HAMMER_REDO_TRUNC,
2252                                                              NULL, 0);
2253                         }
2254                         blksize = hammer_blocksize(vap->va_size);
2255
2256                         /*
2257                          * XXX break atomicy, we can deadlock the backend
2258                          * if we do not release the lock.  Probably not a
2259                          * big deal here.
2260                          */
2261                         if (vap->va_size < ip->ino_data.size) {
2262                                 nvtruncbuf(ap->a_vp, vap->va_size,
2263                                            blksize,
2264                                            hammer_blockoff(vap->va_size),
2265                                            0);
2266                                 truncating = 1;
2267                                 kflags |= NOTE_WRITE;
2268                         } else {
2269                                 nvextendbuf(ap->a_vp,
2270                                             ip->ino_data.size,
2271                                             vap->va_size,
2272                                             hammer_blocksize(ip->ino_data.size),
2273                                             hammer_blocksize(vap->va_size),
2274                                             hammer_blockoff(ip->ino_data.size),
2275                                             hammer_blockoff(vap->va_size),
2276                                             0);
2277                                 truncating = 0;
2278                                 kflags |= NOTE_WRITE | NOTE_EXTEND;
2279                         }
2280                         ip->ino_data.size = vap->va_size;
2281                         ip->ino_data.mtime = trans.time;
2282                         /* XXX safe to use SDIRTY instead of DDIRTY here? */
2283                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2284
2285                         /*
2286                          * On-media truncation is cached in the inode until
2287                          * the inode is synchronized.  We must immediately
2288                          * handle any frontend records.
2289                          */
2290                         if (truncating) {
2291                                 hammer_ip_frontend_trunc(ip, vap->va_size);
2292                                 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2293                                         ip->flags |= HAMMER_INODE_TRUNCATED;
2294                                         ip->trunc_off = vap->va_size;
2295                                         hammer_inode_dirty(ip);
2296                                 } else if (ip->trunc_off > vap->va_size) {
2297                                         ip->trunc_off = vap->va_size;
2298                                 }
2299                         }
2300
2301 #if 0
2302                         /*
2303                          * When truncating, nvtruncbuf() may have cleaned out
2304                          * a portion of the last block on-disk in the buffer
2305                          * cache.  We must clean out any frontend records
2306                          * for blocks beyond the new last block.
2307                          */
2308                         aligned_size = (vap->va_size + (blksize - 1)) &
2309                                        ~(int64_t)(blksize - 1);
2310                         if (truncating && vap->va_size < aligned_size) {
2311                                 aligned_size -= blksize;
2312                                 hammer_ip_frontend_trunc(ip, aligned_size);
2313                         }
2314 #endif
2315                         break;
2316                 case VDATABASE:
2317                         if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2318                                 ip->flags |= HAMMER_INODE_TRUNCATED;
2319                                 ip->trunc_off = vap->va_size;
2320                                 hammer_inode_dirty(ip);
2321                         } else if (ip->trunc_off > vap->va_size) {
2322                                 ip->trunc_off = vap->va_size;
2323                         }
2324                         hammer_ip_frontend_trunc(ip, vap->va_size);
2325                         ip->ino_data.size = vap->va_size;
2326                         ip->ino_data.mtime = trans.time;
2327                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2328                         kflags |= NOTE_ATTRIB;
2329                         break;
2330                 default:
2331                         error = EINVAL;
2332                         goto done;
2333                 }
2334                 break;
2335         }
2336         if (vap->va_atime.tv_sec != VNOVAL) {
2337                 ip->ino_data.atime = hammer_timespec_to_time(&vap->va_atime);
2338                 modflags |= HAMMER_INODE_ATIME;
2339                 kflags |= NOTE_ATTRIB;
2340         }
2341         if (vap->va_mtime.tv_sec != VNOVAL) {
2342                 ip->ino_data.mtime = hammer_timespec_to_time(&vap->va_mtime);
2343                 modflags |= HAMMER_INODE_MTIME;
2344                 kflags |= NOTE_ATTRIB;
2345         }
2346         if (vap->va_mode != (mode_t)VNOVAL) {
2347                 mode_t   cur_mode = ip->ino_data.mode;
2348                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2349                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2350
2351                 error = vop_helper_chmod(ap->a_vp, vap->va_mode, ap->a_cred,
2352                                          cur_uid, cur_gid, &cur_mode);
2353                 if (error == 0 && ip->ino_data.mode != cur_mode) {
2354                         ip->ino_data.mode = cur_mode;
2355                         ip->ino_data.ctime = trans.time;
2356                         modflags |= HAMMER_INODE_DDIRTY;
2357                         kflags |= NOTE_ATTRIB;
2358                 }
2359         }
2360 done:
2361         if (error == 0)
2362                 hammer_modify_inode(&trans, ip, modflags);
2363         hammer_done_transaction(&trans);
2364         hammer_knote(ap->a_vp, kflags);
2365         lwkt_reltoken(&hmp->fs_token);
2366         return (error);
2367 }
2368
2369 /*
2370  * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
2371  */
2372 static
2373 int
2374 hammer_vop_nsymlink(struct vop_nsymlink_args *ap)
2375 {
2376         struct hammer_transaction trans;
2377         hammer_inode_t dip;
2378         hammer_inode_t nip;
2379         hammer_record_t record;
2380         struct nchandle *nch;
2381         hammer_mount_t hmp;
2382         int error;
2383         int bytes;
2384
2385         ap->a_vap->va_type = VLNK;
2386
2387         nch = ap->a_nch;
2388         dip = VTOI(ap->a_dvp);
2389         hmp = dip->hmp;
2390
2391         if (dip->flags & HAMMER_INODE_RO)
2392                 return (EROFS);
2393         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
2394                 return (error);
2395
2396         /*
2397          * Create a transaction to cover the operations we perform.
2398          */
2399         lwkt_gettoken(&hmp->fs_token);
2400         hammer_start_transaction(&trans, hmp);
2401
2402         /*
2403          * Create a new filesystem object of the requested type.  The
2404          * returned inode will be referenced but not locked.
2405          */
2406
2407         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
2408                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
2409                                     NULL, &nip);
2410         if (error) {
2411                 hammer_done_transaction(&trans);
2412                 *ap->a_vpp = NULL;
2413                 lwkt_reltoken(&hmp->fs_token);
2414                 return (error);
2415         }
2416
2417         /*
2418          * Add a record representing the symlink.  symlink stores the link
2419          * as pure data, not a string, and is no \0 terminated.
2420          */
2421         if (error == 0) {
2422                 bytes = strlen(ap->a_target);
2423
2424                 if (bytes <= HAMMER_INODE_BASESYMLEN) {
2425                         bcopy(ap->a_target, nip->ino_data.ext.symlink, bytes);
2426                 } else {
2427                         record = hammer_alloc_mem_record(nip, bytes);
2428                         record->type = HAMMER_MEM_RECORD_GENERAL;
2429
2430                         record->leaf.base.localization = nip->obj_localization |
2431                                                          HAMMER_LOCALIZE_MISC;
2432                         record->leaf.base.key = HAMMER_FIXKEY_SYMLINK;
2433                         record->leaf.base.rec_type = HAMMER_RECTYPE_FIX;
2434                         record->leaf.data_len = bytes;
2435                         KKASSERT(HAMMER_SYMLINK_NAME_OFF == 0);
2436                         bcopy(ap->a_target, record->data->symlink.name, bytes);
2437                         error = hammer_ip_add_record(&trans, record);
2438                 }
2439
2440                 /*
2441                  * Set the file size to the length of the link.
2442                  */
2443                 if (error == 0) {
2444                         nip->ino_data.size = bytes;
2445                         hammer_modify_inode(&trans, nip, HAMMER_INODE_DDIRTY);
2446                 }
2447         }
2448         if (error == 0)
2449                 error = hammer_ip_add_direntry(&trans, dip, nch->ncp->nc_name,
2450                                                 nch->ncp->nc_nlen, nip);
2451
2452         /*
2453          * Finish up.
2454          */
2455         if (error) {
2456                 hammer_rel_inode(nip, 0);
2457                 *ap->a_vpp = NULL;
2458         } else {
2459                 error = hammer_get_vnode(nip, ap->a_vpp);
2460                 hammer_rel_inode(nip, 0);
2461                 if (error == 0) {
2462                         cache_setunresolved(ap->a_nch);
2463                         cache_setvp(ap->a_nch, *ap->a_vpp);
2464                         hammer_knote(ap->a_dvp, NOTE_WRITE);
2465                 }
2466         }
2467         hammer_done_transaction(&trans);
2468         lwkt_reltoken(&hmp->fs_token);
2469         return (error);
2470 }
2471
2472 /*
2473  * hammer_vop_nwhiteout { nch, dvp, cred, flags }
2474  */
2475 static
2476 int
2477 hammer_vop_nwhiteout(struct vop_nwhiteout_args *ap)
2478 {
2479         struct hammer_transaction trans;
2480         hammer_inode_t dip;
2481         hammer_mount_t hmp;
2482         int error;
2483
2484         dip = VTOI(ap->a_dvp);
2485         hmp = dip->hmp;
2486
2487         if (hammer_nohistory(dip) == 0 &&
2488             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0) {
2489                 return (error);
2490         }
2491
2492         lwkt_gettoken(&hmp->fs_token);
2493         hammer_start_transaction(&trans, hmp);
2494         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp,
2495                                 ap->a_cred, ap->a_flags, -1);
2496         hammer_done_transaction(&trans);
2497         lwkt_reltoken(&hmp->fs_token);
2498
2499         return (error);
2500 }
2501
2502 /*
2503  * hammer_vop_ioctl { vp, command, data, fflag, cred }
2504  */
2505 static
2506 int
2507 hammer_vop_ioctl(struct vop_ioctl_args *ap)
2508 {
2509         hammer_inode_t ip = ap->a_vp->v_data;
2510         hammer_mount_t hmp = ip->hmp;
2511         int error;
2512
2513         lwkt_gettoken(&hmp->fs_token);
2514         error = hammer_ioctl(ip, ap->a_command, ap->a_data,
2515                              ap->a_fflag, ap->a_cred);
2516         lwkt_reltoken(&hmp->fs_token);
2517         return (error);
2518 }
2519
2520 static
2521 int
2522 hammer_vop_mountctl(struct vop_mountctl_args *ap)
2523 {
2524         static const struct mountctl_opt extraopt[] = {
2525                 { HMNT_NOHISTORY,       "nohistory" },
2526                 { HMNT_MASTERID,        "master" },
2527                 { HMNT_NOMIRROR,        "nomirror" },
2528                 { 0, NULL}
2529
2530         };
2531         hammer_mount_t hmp;
2532         struct mount *mp;
2533         int usedbytes;
2534         int error;
2535
2536         error = 0;
2537         usedbytes = 0;
2538         mp = ap->a_head.a_ops->head.vv_mount;
2539         KKASSERT(mp->mnt_data != NULL);
2540         hmp = (hammer_mount_t)mp->mnt_data;
2541
2542         lwkt_gettoken(&hmp->fs_token);
2543
2544         switch(ap->a_op) {
2545         case MOUNTCTL_SET_EXPORT:
2546                 if (ap->a_ctllen != sizeof(struct export_args))
2547                         error = EINVAL;
2548                 else
2549                         error = hammer_vfs_export(mp, ap->a_op,
2550                                       (const struct export_args *)ap->a_ctl);
2551                 break;
2552         case MOUNTCTL_MOUNTFLAGS:
2553                 /*
2554                  * Call standard mountctl VOP function
2555                  * so we get user mount flags.
2556                  */
2557                 error = vop_stdmountctl(ap);
2558                 if (error)
2559                         break;
2560
2561                 usedbytes = *ap->a_res;
2562
2563                 if (usedbytes > 0 && usedbytes < ap->a_buflen) {
2564                         usedbytes += vfs_flagstostr(hmp->hflags, extraopt,
2565                                                     ap->a_buf,
2566                                                     ap->a_buflen - usedbytes,
2567                                                     &error);
2568                 }
2569
2570                 *ap->a_res += usedbytes;
2571                 break;
2572         default:
2573                 error = vop_stdmountctl(ap);
2574                 break;
2575         }
2576         lwkt_reltoken(&hmp->fs_token);
2577         return(error);
2578 }
2579
2580 /*
2581  * hammer_vop_strategy { vp, bio }
2582  *
2583  * Strategy call, used for regular file read & write only.  Note that the
2584  * bp may represent a cluster.
2585  *
2586  * To simplify operation and allow better optimizations in the future,
2587  * this code does not make any assumptions with regards to buffer alignment
2588  * or size.
2589  */
2590 static
2591 int
2592 hammer_vop_strategy(struct vop_strategy_args *ap)
2593 {
2594         struct buf *bp;
2595         int error;
2596
2597         bp = ap->a_bio->bio_buf;
2598
2599         switch(bp->b_cmd) {
2600         case BUF_CMD_READ:
2601                 error = hammer_vop_strategy_read(ap);
2602                 break;
2603         case BUF_CMD_WRITE:
2604                 error = hammer_vop_strategy_write(ap);
2605                 break;
2606         default:
2607                 bp->b_error = error = EINVAL;
2608                 bp->b_flags |= B_ERROR;
2609                 biodone(ap->a_bio);
2610                 break;
2611         }
2612
2613         /* hammer_dump_dedup_cache(((hammer_inode_t)ap->a_vp->v_data)->hmp); */
2614
2615         return (error);
2616 }
2617
2618 /*
2619  * Read from a regular file.  Iterate the related records and fill in the
2620  * BIO/BUF.  Gaps are zero-filled.
2621  *
2622  * The support code in hammer_object.c should be used to deal with mixed
2623  * in-memory and on-disk records.
2624  *
2625  * NOTE: Can be called from the cluster code with an oversized buf.
2626  *
2627  * XXX atime update
2628  */
2629 static
2630 int
2631 hammer_vop_strategy_read(struct vop_strategy_args *ap)
2632 {
2633         struct hammer_transaction trans;
2634         hammer_inode_t ip;
2635         hammer_inode_t dip;
2636         hammer_mount_t hmp;
2637         struct hammer_cursor cursor;
2638         hammer_base_elm_t base;
2639         hammer_off_t disk_offset;
2640         struct bio *bio;
2641         struct bio *nbio;
2642         struct buf *bp;
2643         int64_t rec_offset;
2644         int64_t ran_end;
2645         int64_t tmp64;
2646         int error;
2647         int boff;
2648         int roff;
2649         int n;
2650         int isdedupable;
2651
2652         bio = ap->a_bio;
2653         bp = bio->bio_buf;
2654         ip = ap->a_vp->v_data;
2655         hmp = ip->hmp;
2656
2657         /*
2658          * The zone-2 disk offset may have been set by the cluster code via
2659          * a BMAP operation, or else should be NOOFFSET.
2660          *
2661          * Checking the high bits for a match against zone-2 should suffice.
2662          *
2663          * In cases where a lot of data duplication is present it may be
2664          * more beneficial to drop through and doubule-buffer through the
2665          * device.
2666          */
2667         nbio = push_bio(bio);
2668         if (hammer_is_zone_large_data(nbio->bio_offset)) {
2669                 if (hammer_double_buffer == 0) {
2670                         lwkt_gettoken(&hmp->fs_token);
2671                         error = hammer_io_direct_read(hmp, nbio, NULL);
2672                         lwkt_reltoken(&hmp->fs_token);
2673                         return (error);
2674                 }
2675
2676                 /*
2677                  * Try to shortcut requests for double_buffer mode too.
2678                  * Since this mode runs through the device buffer cache
2679                  * only compatible buffer sizes (meaning those generated
2680                  * by normal filesystem buffers) are legal.
2681                  */
2682                 if (hammer_live_dedup == 0 && (bp->b_flags & B_PAGING) == 0) {
2683                         lwkt_gettoken(&hmp->fs_token);
2684                         error = hammer_io_indirect_read(hmp, nbio, NULL);
2685                         lwkt_reltoken(&hmp->fs_token);
2686                         return (error);
2687                 }
2688         }
2689
2690         /*
2691          * Well, that sucked.  Do it the hard way.  If all the stars are
2692          * aligned we may still be able to issue a direct-read.
2693          */
2694         lwkt_gettoken(&hmp->fs_token);
2695         hammer_simple_transaction(&trans, hmp);
2696         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2697
2698         /*
2699          * Key range (begin and end inclusive) to scan.  Note that the key's
2700          * stored in the actual records represent BASE+LEN, not BASE.  The
2701          * first record containing bio_offset will have a key > bio_offset.
2702          */
2703         cursor.key_beg.localization = ip->obj_localization |
2704                                       HAMMER_LOCALIZE_MISC;
2705         cursor.key_beg.obj_id = ip->obj_id;
2706         cursor.key_beg.create_tid = 0;
2707         cursor.key_beg.delete_tid = 0;
2708         cursor.key_beg.obj_type = 0;
2709         cursor.key_beg.key = bio->bio_offset + 1;
2710         cursor.asof = ip->obj_asof;
2711         cursor.flags |= HAMMER_CURSOR_ASOF;
2712
2713         cursor.key_end = cursor.key_beg;
2714         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2715 #if 0
2716         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
2717                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
2718                 cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
2719                 cursor.key_end.key = HAMMER_MAX_KEY;
2720         } else
2721 #endif
2722         {
2723                 ran_end = bio->bio_offset + bp->b_bufsize;
2724                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2725                 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2726                 tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2727                 if (tmp64 < ran_end)
2728                         cursor.key_end.key = HAMMER_MAX_KEY;
2729                 else
2730                         cursor.key_end.key = ran_end + MAXPHYS + 1;
2731         }
2732         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2733
2734         /*
2735          * Set NOSWAPCACHE for cursor data extraction if double buffering
2736          * is disabled or (if the file is not marked cacheable via chflags
2737          * and vm.swapcache_use_chflags is enabled).
2738          */
2739         if (hammer_double_buffer == 0 ||
2740             ((ap->a_vp->v_flag & VSWAPCACHE) == 0 &&
2741              vm_swapcache_use_chflags)) {
2742                 cursor.flags |= HAMMER_CURSOR_NOSWAPCACHE;
2743         }
2744
2745         error = hammer_ip_first(&cursor);
2746         boff = 0;
2747
2748         while (error == 0) {
2749                 /*
2750                  * Get the base file offset of the record.  The key for
2751                  * data records is (base + bytes) rather then (base).
2752                  */
2753                 base = &cursor.leaf->base;
2754                 rec_offset = base->key - cursor.leaf->data_len;
2755
2756                 /*
2757                  * Calculate the gap, if any, and zero-fill it.
2758                  *
2759                  * n is the offset of the start of the record verses our
2760                  * current seek offset in the bio.
2761                  */
2762                 n = (int)(rec_offset - (bio->bio_offset + boff));
2763                 if (n > 0) {
2764                         if (n > bp->b_bufsize - boff)
2765                                 n = bp->b_bufsize - boff;
2766                         bzero((char *)bp->b_data + boff, n);
2767                         boff += n;
2768                         n = 0;
2769                 }
2770
2771                 /*
2772                  * Calculate the data offset in the record and the number
2773                  * of bytes we can copy.
2774                  *
2775                  * There are two degenerate cases.  First, boff may already
2776                  * be at bp->b_bufsize.  Secondly, the data offset within
2777                  * the record may exceed the record's size.
2778                  */
2779                 roff = -n;
2780                 rec_offset += roff;
2781                 n = cursor.leaf->data_len - roff;
2782                 if (n <= 0) {
2783                         hdkprintf("bad n=%d roff=%d\n", n, roff);
2784                         n = 0;
2785                 } else if (n > bp->b_bufsize - boff) {
2786                         n = bp->b_bufsize - boff;
2787                 }
2788
2789                 /*
2790                  * Deal with cached truncations.  This cool bit of code
2791                  * allows truncate()/ftruncate() to avoid having to sync
2792                  * the file.
2793                  *
2794                  * If the frontend is truncated then all backend records are
2795                  * subject to the frontend's truncation.
2796                  *
2797                  * If the backend is truncated then backend records on-disk
2798                  * (but not in-memory) are subject to the backend's
2799                  * truncation.  In-memory records owned by the backend
2800                  * represent data written after the truncation point on the
2801                  * backend and must not be truncated.
2802                  *
2803                  * Truncate operations deal with frontend buffer cache
2804                  * buffers and frontend-owned in-memory records synchronously.
2805                  */
2806                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2807                         if (hammer_cursor_ondisk(&cursor)/* ||
2808                             cursor.iprec->flush_state == HAMMER_FST_FLUSH*/) {
2809                                 if (ip->trunc_off <= rec_offset)
2810                                         n = 0;
2811                                 else if (ip->trunc_off < rec_offset + n)
2812                                         n = (int)(ip->trunc_off - rec_offset);
2813                         }
2814                 }
2815                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2816                         if (hammer_cursor_ondisk(&cursor)) {
2817                                 if (ip->sync_trunc_off <= rec_offset)
2818                                         n = 0;
2819                                 else if (ip->sync_trunc_off < rec_offset + n)
2820                                         n = (int)(ip->sync_trunc_off - rec_offset);
2821                         }
2822                 }
2823
2824                 /*
2825                  * Try to issue a direct read into our bio if possible,
2826                  * otherwise resolve the element data into a hammer_buffer
2827                  * and copy.
2828                  *
2829                  * The buffer on-disk should be zerod past any real
2830                  * truncation point, but may not be for any synthesized
2831                  * truncation point from above.
2832                  *
2833                  * NOTE: disk_offset is only valid if the cursor data is
2834                  *       on-disk.
2835                  */
2836                 disk_offset = cursor.leaf->data_offset + roff;
2837                 isdedupable = (boff == 0 && n == bp->b_bufsize &&
2838                                hammer_cursor_ondisk(&cursor) &&
2839                                ((int)disk_offset & HAMMER_BUFMASK) == 0);
2840
2841                 if (isdedupable && hammer_double_buffer == 0) {
2842                         /*
2843                          * Direct read case
2844                          */
2845                         KKASSERT(hammer_is_zone_large_data(disk_offset));
2846                         nbio->bio_offset = disk_offset;
2847                         error = hammer_io_direct_read(hmp, nbio, cursor.leaf);
2848                         if (hammer_live_dedup && error == 0)
2849                                 hammer_dedup_cache_add(ip, cursor.leaf);
2850                         goto done;
2851                 } else if (isdedupable) {
2852                         /*
2853                          * Async I/O case for reading from backing store
2854                          * and copying the data to the filesystem buffer.
2855                          * live-dedup has to verify the data anyway if it
2856                          * gets a hit later so we can just add the entry
2857                          * now.
2858                          */
2859                         KKASSERT(hammer_is_zone_large_data(disk_offset));
2860                         nbio->bio_offset = disk_offset;
2861                         if (hammer_live_dedup)
2862                                 hammer_dedup_cache_add(ip, cursor.leaf);
2863                         error = hammer_io_indirect_read(hmp, nbio, cursor.leaf);
2864                         goto done;
2865                 } else if (n) {
2866                         error = hammer_ip_resolve_data(&cursor);
2867                         if (error == 0) {
2868                                 if (hammer_live_dedup && isdedupable)
2869                                         hammer_dedup_cache_add(ip, cursor.leaf);
2870                                 bcopy((char *)cursor.data + roff,
2871                                       (char *)bp->b_data + boff, n);
2872                         }
2873                 }
2874                 if (error)
2875                         break;
2876
2877                 /*
2878                  * We have to be sure that the only elements added to the
2879                  * dedup cache are those which are already on-media.
2880                  */
2881                 if (hammer_live_dedup && hammer_cursor_ondisk(&cursor))
2882                         hammer_dedup_cache_add(ip, cursor.leaf);
2883
2884                 /*
2885                  * Iterate until we have filled the request.
2886                  */
2887                 boff += n;
2888                 if (boff == bp->b_bufsize)
2889                         break;
2890                 error = hammer_ip_next(&cursor);
2891         }
2892
2893         /*
2894          * There may have been a gap after the last record
2895          */
2896         if (error == ENOENT)
2897                 error = 0;
2898         if (error == 0 && boff != bp->b_bufsize) {
2899                 KKASSERT(boff < bp->b_bufsize);
2900                 bzero((char *)bp->b_data + boff, bp->b_bufsize - boff);
2901                 /* boff = bp->b_bufsize; */
2902         }
2903
2904         /*
2905          * Disallow swapcache operation on the vnode buffer if double
2906          * buffering is enabled, the swapcache will get the data via
2907          * the block device buffer.
2908          */
2909         if (hammer_double_buffer)
2910                 bp->b_flags |= B_NOTMETA;
2911
2912         /*
2913          * Cleanup
2914          */
2915         bp->b_resid = 0;
2916         bp->b_error = error;
2917         if (error)
2918                 bp->b_flags |= B_ERROR;
2919         biodone(ap->a_bio);
2920
2921 done:
2922         /*
2923          * Cache the b-tree node for the last data read in cache[1].
2924          *
2925          * If we hit the file EOF then also cache the node in the
2926          * governing directory's cache[3], it will be used to initialize
2927          * the new inode's cache[1] for any inodes looked up via the directory.
2928          *
2929          * This doesn't reduce disk accesses since the B-Tree chain is
2930          * likely cached, but it does reduce cpu overhead when looking
2931          * up file offsets for cpdup/tar/cpio style iterations.
2932          */
2933         if (cursor.node)
2934                 hammer_cache_node(&ip->cache[1], cursor.node);
2935         if (ran_end >= ip->ino_data.size) {
2936                 dip = hammer_find_inode(&trans, ip->ino_data.parent_obj_id,
2937                                         ip->obj_asof, ip->obj_localization);
2938                 if (dip) {
2939                         hammer_cache_node(&dip->cache[3], cursor.node);
2940                         hammer_rel_inode(dip, 0);
2941                 }
2942         }
2943         hammer_done_cursor(&cursor);
2944         hammer_done_transaction(&trans);
2945         lwkt_reltoken(&hmp->fs_token);
2946         return(error);
2947 }
2948
2949 /*
2950  * BMAP operation - used to support cluster_read() only.
2951  *
2952  * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
2953  *
2954  * This routine may return EOPNOTSUPP if the opration is not supported for
2955  * the specified offset.  The contents of the pointer arguments do not
2956  * need to be initialized in that case.
2957  *
2958  * If a disk address is available and properly aligned return 0 with
2959  * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
2960  * to the run-length relative to that offset.  Callers may assume that
2961  * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
2962  * large, so return EOPNOTSUPP if it is not sufficiently large.
2963  */
2964 static
2965 int
2966 hammer_vop_bmap(struct vop_bmap_args *ap)
2967 {
2968         struct hammer_transaction trans;
2969         hammer_inode_t ip;
2970         hammer_mount_t hmp;
2971         struct hammer_cursor cursor;
2972         hammer_base_elm_t base;
2973         int64_t rec_offset;
2974         int64_t ran_end;
2975         int64_t tmp64;
2976         int64_t base_offset;
2977         int64_t base_disk_offset;
2978         int64_t last_offset;
2979         hammer_off_t last_disk_offset;
2980         hammer_off_t disk_offset;
2981         int     rec_len;
2982         int     error;
2983         int     blksize;
2984
2985         ip = ap->a_vp->v_data;
2986         hmp = ip->hmp;
2987
2988         /*
2989          * We can only BMAP regular files.  We can't BMAP database files,
2990          * directories, etc.
2991          */
2992         if (ip->ino_data.obj_type != HAMMER_OBJTYPE_REGFILE)
2993                 return(EOPNOTSUPP);
2994
2995         /*
2996          * bmap is typically called with runp/runb both NULL when used
2997          * for writing.  We do not support BMAP for writing atm.
2998          */
2999         if (ap->a_cmd != BUF_CMD_READ)
3000                 return(EOPNOTSUPP);
3001
3002         /*
3003          * Scan the B-Tree to acquire blockmap addresses, then translate
3004          * to raw addresses.
3005          */
3006         lwkt_gettoken(&hmp->fs_token);
3007         hammer_simple_transaction(&trans, hmp);
3008
3009         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
3010
3011         /*
3012          * Key range (begin and end inclusive) to scan.  Note that the key's
3013          * stored in the actual records represent BASE+LEN, not BASE.  The
3014          * first record containing bio_offset will have a key > bio_offset.
3015          */
3016         cursor.key_beg.localization = ip->obj_localization |
3017                                       HAMMER_LOCALIZE_MISC;
3018         cursor.key_beg.obj_id = ip->obj_id;
3019         cursor.key_beg.create_tid = 0;
3020         cursor.key_beg.delete_tid = 0;
3021         cursor.key_beg.obj_type = 0;
3022         if (ap->a_runb)
3023                 cursor.key_beg.key = ap->a_loffset - MAXPHYS + 1;
3024         else
3025                 cursor.key_beg.key = ap->a_loffset + 1;
3026         if (cursor.key_beg.key < 0)
3027                 cursor.key_beg.key = 0;
3028         cursor.asof = ip->obj_asof;
3029         cursor.flags |= HAMMER_CURSOR_ASOF;
3030
3031         cursor.key_end = cursor.key_beg;
3032         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
3033
3034         ran_end = ap->a_loffset + MAXPHYS;
3035         cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
3036         cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
3037         tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
3038         if (tmp64 < ran_end)
3039                 cursor.key_end.key = HAMMER_MAX_KEY;
3040         else
3041                 cursor.key_end.key = ran_end + MAXPHYS + 1;
3042
3043         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
3044
3045         error = hammer_ip_first(&cursor);
3046         base_offset = last_offset = 0;
3047         base_disk_offset = last_disk_offset = 0;
3048
3049         while (error == 0) {
3050                 /*
3051                  * Get the base file offset of the record.  The key for
3052                  * data records is (base + bytes) rather then (base).
3053                  *
3054                  * NOTE: rec_offset + rec_len may exceed the end-of-file.
3055                  * The extra bytes should be zero on-disk and the BMAP op
3056                  * should still be ok.
3057                  */
3058                 base = &cursor.leaf->base;
3059                 rec_offset = base->key - cursor.leaf->data_len;
3060                 rec_len    = cursor.leaf->data_len;
3061
3062                 /*
3063                  * Incorporate any cached truncation.
3064                  *
3065                  * NOTE: Modifications to rec_len based on synthesized
3066                  * truncation points remove the guarantee that any extended
3067                  * data on disk is zero (since the truncations may not have
3068                  * taken place on-media yet).
3069                  */
3070                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
3071                         if (hammer_cursor_ondisk(&cursor) ||
3072                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
3073                                 if (ip->trunc_off <= rec_offset)
3074                                         rec_len = 0;
3075                                 else if (ip->trunc_off < rec_offset + rec_len)
3076                                         rec_len = (int)(ip->trunc_off - rec_offset);
3077                         }
3078                 }
3079                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
3080                         if (hammer_cursor_ondisk(&cursor)) {
3081                                 if (ip->sync_trunc_off <= rec_offset)
3082                                         rec_len = 0;
3083                                 else if (ip->sync_trunc_off < rec_offset + rec_len)
3084                                         rec_len = (int)(ip->sync_trunc_off - rec_offset);
3085                         }
3086                 }
3087
3088                 /*
3089                  * Accumulate information.  If we have hit a discontiguous
3090                  * block reset base_offset unless we are already beyond the
3091                  * requested offset.  If we are, that's it, we stop.
3092                  */
3093                 if (error)
3094                         break;
3095                 if (hammer_cursor_ondisk(&cursor)) {
3096                         disk_offset = cursor.leaf->data_offset;
3097                         if (rec_offset != last_offset ||
3098                             disk_offset != last_disk_offset) {
3099                                 if (rec_offset > ap->a_loffset)
3100                                         break;
3101                                 base_offset = rec_offset;
3102                                 base_disk_offset = disk_offset;
3103                         }
3104                         last_offset = rec_offset + rec_len;
3105                         last_disk_offset = disk_offset + rec_len;
3106
3107                         if (hammer_live_dedup)
3108                                 hammer_dedup_cache_add(ip, cursor.leaf);
3109                 }
3110
3111                 error = hammer_ip_next(&cursor);
3112         }
3113
3114         if (cursor.node)
3115                 hammer_cache_node(&ip->cache[1], cursor.node);
3116
3117         hammer_done_cursor(&cursor);
3118         hammer_done_transaction(&trans);
3119         lwkt_reltoken(&hmp->fs_token);
3120
3121         /*
3122          * If we couldn't find any records or the records we did find were
3123          * all behind the requested offset, return failure.  A forward
3124          * truncation can leave a hole w/ no on-disk records.
3125          */
3126         if (last_offset == 0 || last_offset < ap->a_loffset)
3127                 return (EOPNOTSUPP);
3128
3129         /*
3130          * Figure out the block size at the requested offset and adjust
3131          * our limits so the cluster_read() does not create inappropriately
3132          * sized buffer cache buffers.
3133          */
3134         blksize = hammer_blocksize(ap->a_loffset);
3135         if (hammer_blocksize(base_offset) != blksize) {
3136                 base_offset = hammer_blockdemarc(base_offset, ap->a_loffset);
3137         }
3138         if (last_offset != ap->a_loffset &&
3139             hammer_blocksize(last_offset - 1) != blksize) {
3140                 last_offset = hammer_blockdemarc(ap->a_loffset,
3141                                                  last_offset - 1);
3142         }
3143
3144         /*
3145          * Returning EOPNOTSUPP simply prevents the direct-IO optimization
3146          * from occuring.
3147          */
3148         disk_offset = base_disk_offset + (ap->a_loffset - base_offset);
3149
3150         if (!hammer_is_zone_large_data(disk_offset)) {
3151                 /*
3152                  * Only large-data zones can be direct-IOd
3153                  */
3154                 error = EOPNOTSUPP;
3155         } else if ((disk_offset & HAMMER_BUFMASK) ||
3156                    (last_offset - ap->a_loffset) < blksize) {
3157                 /*
3158                  * doffsetp is not aligned or the forward run size does
3159                  * not cover a whole buffer, disallow the direct I/O.
3160                  */
3161                 error = EOPNOTSUPP;
3162         } else {
3163                 /*
3164                  * We're good.
3165                  */
3166                 *ap->a_doffsetp = disk_offset;
3167                 if (ap->a_runb) {
3168                         *ap->a_runb = ap->a_loffset - base_offset;
3169                         KKASSERT(*ap->a_runb >= 0);
3170                 }
3171                 if (ap->a_runp) {
3172                         *ap->a_runp = last_offset - ap->a_loffset;
3173                         KKASSERT(*ap->a_runp >= 0);
3174                 }
3175                 error = 0;
3176         }
3177         return(error);
3178 }
3179
3180 /*
3181  * Write to a regular file.   Because this is a strategy call the OS is
3182  * trying to actually get data onto the media.
3183  */
3184 static
3185 int
3186 hammer_vop_strategy_write(struct vop_strategy_args *ap)
3187 {
3188         hammer_record_t record;
3189         hammer_mount_t hmp;
3190         hammer_inode_t ip;
3191         struct bio *bio;
3192         struct buf *bp;
3193         int blksize __debugvar;
3194         int bytes;
3195         int error;
3196
3197         bio = ap->a_bio;
3198         bp = bio->bio_buf;
3199         ip = ap->a_vp->v_data;
3200         hmp = ip->hmp;
3201
3202         blksize = hammer_blocksize(bio->bio_offset);
3203         KKASSERT(bp->b_bufsize == blksize);
3204
3205         if (ip->flags & HAMMER_INODE_RO) {
3206                 bp->b_error = EROFS;
3207                 bp->b_flags |= B_ERROR;
3208                 biodone(ap->a_bio);
3209                 return(EROFS);
3210         }
3211
3212         lwkt_gettoken(&hmp->fs_token);
3213
3214         /*
3215          * Disallow swapcache operation on the vnode buffer if double
3216          * buffering is enabled, the swapcache will get the data via
3217          * the block device buffer.
3218          */
3219         if (hammer_double_buffer)
3220                 bp->b_flags |= B_NOTMETA;
3221
3222         /*
3223          * Interlock with inode destruction (no in-kernel or directory
3224          * topology visibility).  If we queue new IO while trying to
3225          * destroy the inode we can deadlock the vtrunc call in
3226          * hammer_inode_unloadable_check().
3227          *
3228          * Besides, there's no point flushing a bp associated with an
3229          * inode that is being destroyed on-media and has no kernel
3230          * references.
3231          */
3232         if ((ip->flags | ip->sync_flags) &
3233             (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) {
3234                 bp->b_resid = 0;
3235                 biodone(ap->a_bio);
3236                 lwkt_reltoken(&hmp->fs_token);
3237                 return(0);
3238         }
3239
3240         /*
3241          * Reserve space and issue a direct-write from the front-end.
3242          * NOTE: The direct_io code will hammer_bread/bcopy smaller
3243          * allocations.
3244          *
3245          * An in-memory record will be installed to reference the storage
3246          * until the flusher can get to it.
3247          *
3248          * Since we own the high level bio the front-end will not try to
3249          * do a direct-read until the write completes.
3250          *
3251          * NOTE: The only time we do not reserve a full-sized buffers
3252          * worth of data is if the file is small.  We do not try to
3253          * allocate a fragment (from the small-data zone) at the end of
3254          * an otherwise large file as this can lead to wildly separated
3255          * data.
3256          */
3257         KKASSERT((bio->bio_offset & HAMMER_BUFMASK) == 0);
3258         KKASSERT(bio->bio_offset < ip->ino_data.size);
3259         if (bio->bio_offset || ip->ino_data.size > HAMMER_HBUFSIZE)
3260                 bytes = bp->b_bufsize;
3261         else
3262                 bytes = HAMMER_DATA_DOALIGN_WITH(int, ip->ino_data.size);
3263
3264         record = hammer_ip_add_bulk(ip, bio->bio_offset, bp->b_data,
3265                                     bytes, &error);
3266
3267         /*
3268          * B_VFSFLAG1 indicates that a REDO_WRITE entry was generated
3269          * in hammer_vop_write().  We must flag the record so the proper
3270          * REDO_TERM_WRITE entry is generated during the flush.
3271          */
3272         if (record) {
3273                 if (bp->b_flags & B_VFSFLAG1) {
3274                         record->flags |= HAMMER_RECF_REDO;
3275                         bp->b_flags &= ~B_VFSFLAG1;
3276                 }
3277                 if (record->flags & HAMMER_RECF_DEDUPED) {
3278                         bp->b_resid = 0;
3279                         hammer_ip_replace_bulk(hmp, record);
3280                         biodone(ap->a_bio);
3281                 } else {
3282                         hammer_io_direct_write(hmp, bio, record);
3283                 }
3284                 if (ip->rsv_recs > 1 && hmp->rsv_recs > hammer_limit_recs)
3285                         hammer_flush_inode(ip, 0);
3286         } else {
3287                 bp->b_bio2.bio_offset = NOOFFSET;
3288                 bp->b_error = error;
3289                 bp->b_flags |= B_ERROR;
3290                 biodone(ap->a_bio);
3291         }
3292         lwkt_reltoken(&hmp->fs_token);
3293         return(error);
3294 }
3295
3296 /*
3297  * dounlink - disconnect a directory entry
3298  *
3299  * XXX whiteout support not really in yet
3300  */
3301 static int
3302 hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
3303                 struct vnode *dvp, struct ucred *cred,
3304                 int flags, int isdir)
3305 {
3306         struct namecache *ncp;
3307         hammer_inode_t dip;
3308         hammer_inode_t ip;
3309         hammer_mount_t hmp;
3310         struct hammer_cursor cursor;
3311         int64_t namekey;
3312         uint32_t max_iterations;
3313         int nlen, error;
3314
3315         /*
3316          * Calculate the namekey and setup the key range for the scan.  This
3317          * works kinda like a chained hash table where the lower 32 bits
3318          * of the namekey synthesize the chain.
3319          *
3320          * The key range is inclusive of both key_beg and key_end.
3321          */
3322         dip = VTOI(dvp);
3323         ncp = nch->ncp;
3324         hmp = dip->hmp;
3325
3326         if (dip->flags & HAMMER_INODE_RO)
3327                 return (EROFS);
3328
3329         namekey = hammer_direntry_namekey(dip, ncp->nc_name, ncp->nc_nlen,
3330                                            &max_iterations);
3331 retry:
3332         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
3333         cursor.key_beg.localization = dip->obj_localization |
3334                                       hammer_dir_localization(dip);
3335         cursor.key_beg.obj_id = dip->obj_id;
3336         cursor.key_beg.key = namekey;
3337         cursor.key_beg.create_tid = 0;
3338         cursor.key_beg.delete_tid = 0;
3339         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
3340         cursor.key_beg.obj_type = 0;
3341
3342         cursor.key_end = cursor.key_beg;
3343         cursor.key_end.key += max_iterations;
3344         cursor.asof = dip->obj_asof;
3345         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
3346
3347         /*
3348          * Scan all matching records (the chain), locate the one matching
3349          * the requested path component.  info->last_error contains the
3350          * error code on search termination and could be 0, ENOENT, or
3351          * something else.
3352          *
3353          * The hammer_ip_*() functions merge in-memory records with on-disk
3354          * records for the purposes of the search.
3355          */
3356         error = hammer_ip_first(&cursor);
3357
3358         while (error == 0) {
3359                 error = hammer_ip_resolve_data(&cursor);
3360                 if (error)
3361                         break;
3362                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
3363                 KKASSERT(nlen > 0);
3364                 if (ncp->nc_nlen == nlen &&
3365                     bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
3366                         break;
3367                 }
3368                 error = hammer_ip_next(&cursor);
3369         }
3370
3371         /*
3372          * If all is ok we have to get the inode so we can adjust nlinks.
3373          * To avoid a deadlock with the flusher we must release the inode
3374          * lock on the directory when acquiring the inode for the entry.
3375          *
3376          * If the target is a directory, it must be empty.
3377          */
3378         if (error == 0) {
3379                 hammer_unlock(&cursor.ip->lock);
3380                 ip = hammer_get_inode(trans, dip, cursor.data->entry.obj_id,
3381                                       hmp->asof,
3382                                       cursor.data->entry.localization,
3383                                       0, &error);
3384                 hammer_lock_sh(&cursor.ip->lock);
3385                 if (error == ENOENT) {
3386                         hkprintf("WARNING: Removing dirent w/missing inode "
3387                                 "\"%s\"\n"
3388                                 "\tobj_id = %016jx\n",
3389                                 ncp->nc_name,
3390                                 (intmax_t)cursor.data->entry.obj_id);
3391                         error = 0;
3392                 }
3393
3394                 /*
3395                  * If isdir >= 0 we validate that the entry is or is not a
3396                  * directory.  If isdir < 0 we don't care.
3397                  */
3398                 if (error == 0 && isdir >= 0 && ip) {
3399                         if (isdir &&
3400                             ip->ino_data.obj_type != HAMMER_OBJTYPE_DIRECTORY) {
3401                                 error = ENOTDIR;
3402                         } else if (isdir == 0 &&
3403                             ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
3404                                 error = EISDIR;
3405                         }
3406                 }
3407
3408                 /*
3409                  * If we are trying to remove a directory the directory must
3410                  * be empty.
3411                  *
3412                  * The check directory code can loop and deadlock/retry.  Our
3413                  * own cursor's node locks must be released to avoid a 3-way
3414                  * deadlock with the flusher if the check directory code
3415                  * blocks.
3416                  *
3417                  * If any changes whatsoever have been made to the cursor
3418                  * set EDEADLK and retry.
3419                  *
3420                  * WARNING: See warnings in hammer_unlock_cursor()
3421                  *          function.
3422                  */
3423                 if (error == 0 && ip && ip->ino_data.obj_type ==
3424                                         HAMMER_OBJTYPE_DIRECTORY) {
3425                         hammer_unlock_cursor(&cursor);
3426                         error = hammer_ip_check_directory_empty(trans, ip);
3427                         hammer_lock_cursor(&cursor);
3428                         if (cursor.flags & HAMMER_CURSOR_RETEST) {
3429                                 hkprintf("Warning: avoided deadlock "
3430                                         "on rmdir '%s'\n",
3431                                         ncp->nc_name);
3432                                 error = EDEADLK;
3433                         }
3434                 }
3435
3436                 /*
3437                  * Delete the directory entry.
3438                  *
3439                  * WARNING: hammer_ip_del_direntry() may have to terminate
3440                  * the cursor to avoid a deadlock.  It is ok to call
3441                  * hammer_done_cursor() twice.
3442                  */
3443                 if (error == 0) {
3444                         error = hammer_ip_del_direntry(trans, &cursor,
3445                                                         dip, ip);
3446                 }
3447                 hammer_done_cursor(&cursor);
3448                 if (error == 0) {
3449                         /*
3450                          * Tell the namecache that we are now unlinked.
3451                          */
3452                         cache_unlink(nch);
3453
3454                         /*
3455                          * NOTE: ip->vp, if non-NULL, cannot be directly
3456                          *       referenced without formally acquiring the
3457                          *       vp since the vp might have zero refs on it,
3458                          *       or in the middle of a reclaim, etc.
3459                          *
3460                          * NOTE: The cache_setunresolved() can rip the vp
3461                          *       out from under us since the vp may not have
3462                          *       any refs, in which case ip->vp will be NULL
3463                          *       from the outset.
3464                          */
3465                         while (ip && ip->vp) {
3466                                 struct vnode *vp;
3467
3468                                 error = hammer_get_vnode(ip, &vp);
3469                                 if (error == 0 && vp) {
3470                                         vn_unlock(vp);
3471                                         hammer_knote(ip->vp, NOTE_DELETE);
3472 #if 0
3473                                         /*
3474                                          * Don't do this, it can deadlock
3475                                          * on concurrent rm's of hardlinks.
3476                                          * Shouldn't be needed any more.
3477                                          */
3478                                         cache_inval_vp(ip->vp, CINV_DESTROY);
3479 #endif
3480                                         vrele(vp);
3481                                         break;
3482                                 }
3483                                 hdkprintf("ip/vp race1 avoided\n");
3484                         }
3485                 }
3486                 if (ip)
3487                         hammer_rel_inode(ip, 0);
3488         } else {
3489                 hammer_done_cursor(&cursor);
3490         }
3491         if (error == EDEADLK)
3492                 goto retry;
3493
3494         return (error);
3495 }
3496
3497 /************************************************************************
3498  *                          FIFO AND SPECFS OPS                         *
3499  ************************************************************************
3500  *
3501  */
3502 static int
3503 hammer_vop_fifoclose (struct vop_close_args *ap)
3504 {
3505         /* XXX update itimes */
3506         return (VOCALL(&fifo_vnode_vops, &ap->a_head));
3507 }
3508
3509 static int
3510 hammer_vop_fiforead (struct vop_read_args *ap)
3511 {
3512         int error;
3513
3514         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3515         /* XXX update access time */
3516         return (error);
3517 }
3518
3519 static int
3520 hammer_vop_fifowrite (struct vop_write_args *ap)
3521 {
3522         int error;
3523
3524         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3525         /* XXX update access time */
3526         return (error);
3527 }
3528
3529 static
3530 int
3531 hammer_vop_fifokqfilter(struct vop_kqfilter_args *ap)
3532 {
3533         int error;
3534
3535         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3536         if (error)
3537                 error = hammer_vop_kqfilter(ap);
3538         return(error);
3539 }
3540
3541 /************************************************************************
3542  *                          KQFILTER OPS                                *
3543  ************************************************************************
3544  *
3545  */
3546 static void filt_hammerdetach(struct knote *kn);
3547 static int filt_hammerread(struct knote *kn, long hint);
3548 static int filt_hammerwrite(struct knote *kn, long hint);
3549 static int filt_hammervnode(struct knote *kn, long hint);
3550
3551 static struct filterops hammerread_filtops =
3552         { FILTEROP_ISFD | FILTEROP_MPSAFE,
3553           NULL, filt_hammerdetach, filt_hammerread };
3554 static struct filterops hammerwrite_filtops =
3555         { FILTEROP_ISFD | FILTEROP_MPSAFE,
3556           NULL, filt_hammerdetach, filt_hammerwrite };
3557 static struct filterops hammervnode_filtops =
3558         { FILTEROP_ISFD | FILTEROP_MPSAFE,
3559           NULL, filt_hammerdetach, filt_hammervnode };
3560
3561 static
3562 int
3563 hammer_vop_kqfilter(struct vop_kqfilter_args *ap)
3564 {
3565         struct vnode *vp = ap->a_vp;
3566         struct knote *kn = ap->a_kn;
3567
3568         switch (kn->kn_filter) {
3569         case EVFILT_READ:
3570                 kn->kn_fop = &hammerread_filtops;
3571                 break;
3572         case EVFILT_WRITE:
3573                 kn->kn_fop = &hammerwrite_filtops;
3574                 break;
3575         case EVFILT_VNODE:
3576                 kn->kn_fop = &hammervnode_filtops;
3577                 break;
3578         default:
3579                 return (EOPNOTSUPP);
3580         }
3581
3582         kn->kn_hook = (caddr_t)vp;
3583
3584         knote_insert(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3585
3586         return(0);
3587 }
3588
3589 static void
3590 filt_hammerdetach(struct knote *kn)
3591 {
3592         struct vnode *vp = (void *)kn->kn_hook;
3593
3594         knote_remove(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3595 }
3596
3597 static int
3598 filt_hammerread(struct knote *kn, long hint)
3599 {
3600         struct vnode *vp = (void *)kn->kn_hook;
3601         hammer_inode_t ip = VTOI(vp);
3602         hammer_mount_t hmp = ip->hmp;
3603         off_t off;
3604
3605         if (hint == NOTE_REVOKE) {
3606                 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
3607                 return(1);
3608         }
3609         lwkt_gettoken(&hmp->fs_token);  /* XXX use per-ip-token */
3610         off = ip->ino_data.size - kn->kn_fp->f_offset;
3611         kn->kn_data = (off < INTPTR_MAX) ? off : INTPTR_MAX;
3612         lwkt_reltoken(&hmp->fs_token);
3613         if (kn->kn_sfflags & NOTE_OLDAPI)
3614                 return(1);
3615         return (kn->kn_data != 0);
3616 }
3617
3618 static int
3619 filt_hammerwrite(struct knote *kn, long hint)
3620 {
3621         if (hint == NOTE_REVOKE)
3622                 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
3623         kn->kn_data = 0;
3624         return (1);
3625 }
3626
3627 static int
3628 filt_hammervnode(struct knote *kn, long hint)
3629 {
3630         if (kn->kn_sfflags & hint)
3631                 kn->kn_fflags |= hint;
3632         if (hint == NOTE_REVOKE) {
3633                 kn->kn_flags |= (EV_EOF | EV_NODATA);
3634                 return (1);
3635         }
3636         return (kn->kn_fflags != 0);
3637 }
3638