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 + (BKVASIZE - 1)) / BKVASIZE;
 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, uio->uio_resid,
 416                                              seqcount * BKVASIZE, &bp);
 417                 } else {
 418                         error = breadnx(ap->a_vp, base_offset, blksize,
 419                                         NULL, NULL, 0, &bp);
 420                 }
 421                 if (error) {
 422                         brelse(bp);
 423                         break;
 424                 }
 425 skip:
 426                 if ((hammer_debug_io & 0x0001) && (bp->b_flags & B_IODEBUG)) {
 427                         hdkprintf("zone2_offset %016jx read file %016jx@%016jx\n",
 428                                 (intmax_t)bp->b_bio2.bio_offset,
 429                                 (intmax_t)ip->obj_id,
 430                                 (intmax_t)bp->b_loffset);
 431                 }
 432                 bp->b_flags &= ~B_IODEBUG;
 433                 if (blksize == HAMMER_XBUFSIZE)
 434                         bp->b_flags |= B_CLUSTEROK;
 435
 436                 n = blksize - offset;
 437                 if (n > uio->uio_resid)
 438                         n = uio->uio_resid;
 439                 if (n > ip->ino_data.size - uio->uio_offset)
 440                         n = (int)(ip->ino_data.size - uio->uio_offset);
 441
 442                 /*
 443                  * Set B_AGE, data has a lower priority than meta-data.
 444                  *
 445                  * Use a hold/unlock/drop sequence to run the uiomove
 446                  * with the buffer unlocked, avoiding deadlocks against
 447                  * read()s on mmap()'d spaces.
 448                  */
 449                 bp->b_flags |= B_AGE;
 450                 error = uiomovebp(bp, (char *)bp->b_data + offset, n, uio);
 451                 bqrelse(bp);
 452
 453                 if (error)
 454                         break;
 455                 hammer_stats_file_read += n;
 456         }
 457
 458 finished:
 459
 460         /*
 461          * Try to update the atime with just the inode lock for maximum
 462          * concurrency.  If we can't shortcut it we have to get the full
 463          * blown transaction.
 464          */
 465         if (got_trans == 0 && hammer_update_atime_quick(ip) < 0) {
 466                 hammer_start_transaction(&trans, ip->hmp);
 467                 got_trans = 1;
 468         }
 469
 470         if (got_trans) {
 471                 if ((ip->flags & HAMMER_INODE_RO) == 0 &&
 472                     (ip->hmp->mp->mnt_flag & MNT_NOATIME) == 0) {
 473                         lwkt_gettoken(&hmp->fs_token);
 474                         ip->ino_data.atime = trans.time;
 475                         hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
 476                         hammer_done_transaction(&trans);
 477                         lwkt_reltoken(&hmp->fs_token);
 478                 } else {
 479                         hammer_done_transaction(&trans);
 480                 }
 481         }
 482         return (error);
 483 }
 484
 485 /*
 486  * hammer_vop_write { vp, uio, ioflag, cred }
 487  */
 488 static
 489 int
 490 hammer_vop_write(struct vop_write_args *ap)
 491 {
 492         struct hammer_transaction trans;
 493         struct hammer_inode *ip;
 494         hammer_mount_t hmp;
 495         thread_t td;
 496         struct uio *uio;
 497         int offset;
 498         off_t base_offset;
 499         int64_t cluster_eof;
 500         struct buf *bp;
 501         int kflags;
 502         int error;
 503         int n;
 504         int flags;
 505         int seqcount;
 506         int bigwrite;
 507
 508         if (ap->a_vp->v_type != VREG)
 509                 return (EINVAL);
 510         ip = VTOI(ap->a_vp);
 511         hmp = ip->hmp;
 512         error = 0;
 513         kflags = 0;
 514         seqcount = ap->a_ioflag >> 16;
 515
 516         if (ip->flags & HAMMER_INODE_RO)
 517                 return (EROFS);
 518
 519         /*
 520          * Create a transaction to cover the operations we perform.
 521          */
 522         hammer_start_transaction(&trans, hmp);
 523         uio = ap->a_uio;
 524
 525         /*
 526          * Check append mode
 527          */
 528         if (ap->a_ioflag & IO_APPEND)
 529                 uio->uio_offset = ip->ino_data.size;
 530
 531         /*
 532          * Check for illegal write offsets.  Valid range is 0...2^63-1.
 533          *
 534          * NOTE: the base_off assignment is required to work around what
 535          * I consider to be a GCC-4 optimization bug.
 536          */
 537         if (uio->uio_offset < 0) {
 538                 hammer_done_transaction(&trans);
 539                 return (EFBIG);
 540         }
 541         base_offset = uio->uio_offset + uio->uio_resid; /* work around gcc-4 */
 542         if (uio->uio_resid > 0 && base_offset <= uio->uio_offset) {
 543                 hammer_done_transaction(&trans);
 544                 return (EFBIG);
 545         }
 546
 547         if (uio->uio_resid > 0 && (td = uio->uio_td) != NULL && td->td_proc &&
 548             base_offset > td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
 549                 hammer_done_transaction(&trans);
 550                 lwpsignal(td->td_proc, td->td_lwp, SIGXFSZ);
 551                 return (EFBIG);
 552         }
 553
 554         /*
 555          * If reading or writing a huge amount of data we have to break
 556          * atomicy and allow the operation to be interrupted by a signal
 557          * or it can DOS the machine.
 558          *
 559          * Preset redo_count so we stop generating REDOs earlier if the
 560          * limit is exceeded.
 561          *
 562          * redo_count is heuristical, SMP races are ok
 563          */
 564         bigwrite = (uio->uio_resid > 100 * 1024 * 1024);
 565         if ((ip->flags & HAMMER_INODE_REDO) &&
 566             ip->redo_count < hammer_limit_redo) {
 567                 ip->redo_count += uio->uio_resid;
 568         }
 569
 570         /*
 571          * Access the data typically in HAMMER_BUFSIZE blocks via the
 572          * buffer cache, but HAMMER may use a variable block size based
 573          * on the offset.
 574          */
 575         while (uio->uio_resid > 0) {
 576                 int fixsize = 0;
 577                 int blksize;
 578                 int blkmask;
 579                 int trivial;
 580                 int endofblk;
 581                 off_t nsize;
 582
 583                 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_WRITE)) != 0)
 584                         break;
 585                 if (bigwrite && (error = hammer_signal_check(hmp)) != 0)
 586                         break;
 587
 588                 blksize = hammer_blocksize(uio->uio_offset);
 589
 590                 /*
 591                  * Control the number of pending records associated with
 592                  * this inode.  If too many have accumulated start a
 593                  * flush.  Try to maintain a pipeline with the flusher.
 594                  *
 595                  * NOTE: It is possible for other sources to grow the
 596                  *       records but not necessarily issue another flush,
 597                  *       so use a timeout and ensure that a re-flush occurs.
 598                  */
 599                 if (ip->rsv_recs >= hammer_limit_inode_recs) {
 600                         lwkt_gettoken(&hmp->fs_token);
 601                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 602                         while (ip->rsv_recs >= hammer_limit_inode_recs * 2) {
 603                                 ip->flags |= HAMMER_INODE_RECSW;
 604                                 tsleep(&ip->rsv_recs, 0, "hmrwww", hz);
 605                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 606                         }
 607                         lwkt_reltoken(&hmp->fs_token);
 608                 }
 609
 610                 /*
 611                  * Do not allow HAMMER to blow out the buffer cache.  Very
 612                  * large UIOs can lockout other processes due to bwillwrite()
 613                  * mechanics.
 614                  *
 615                  * The hammer inode is not locked during these operations.
 616                  * The vnode is locked which can interfere with the pageout
 617                  * daemon for non-UIO_NOCOPY writes but should not interfere
 618                  * with the buffer cache.  Even so, we cannot afford to
 619                  * allow the pageout daemon to build up too many dirty buffer
 620                  * cache buffers.
 621                  *
 622                  * Only call this if we aren't being recursively called from
 623                  * a virtual disk device (vn), else we may deadlock.
 624                  */
 625                 if ((ap->a_ioflag & IO_RECURSE) == 0)
 626                         bwillwrite(blksize);
 627
 628                 /*
 629                  * Calculate the blocksize at the current offset and figure
 630                  * out how much we can actually write.
 631                  */
 632                 blkmask = blksize - 1;
 633                 offset = (int)uio->uio_offset & blkmask;
 634                 base_offset = uio->uio_offset & ~(int64_t)blkmask;
 635                 n = blksize - offset;
 636                 if (n > uio->uio_resid) {
 637                         n = uio->uio_resid;
 638                         endofblk = 0;
 639                 } else {
 640                         endofblk = 1;
 641                 }
 642                 nsize = uio->uio_offset + n;
 643                 if (nsize > ip->ino_data.size) {
 644                         if (uio->uio_offset > ip->ino_data.size)
 645                                 trivial = 0;
 646                         else
 647                                 trivial = 1;
 648                         nvextendbuf(ap->a_vp,
 649                                     ip->ino_data.size,
 650                                     nsize,
 651                                     hammer_blocksize(ip->ino_data.size),
 652                                     hammer_blocksize(nsize),
 653                                     hammer_blockoff(ip->ino_data.size),
 654                                     hammer_blockoff(nsize),
 655                                     trivial);
 656                         fixsize = 1;
 657                         kflags |= NOTE_EXTEND;
 658                 }
 659
 660                 if (uio->uio_segflg == UIO_NOCOPY) {
 661                         /*
 662                          * Issuing a write with the same data backing the
 663                          * buffer.  Instantiate the buffer to collect the
 664                          * backing vm pages, then read-in any missing bits.
 665                          *
 666                          * This case is used by vop_stdputpages().
 667                          */
 668                         bp = getblk(ap->a_vp, base_offset,
 669                                     blksize, GETBLK_BHEAVY, 0);
 670                         if ((bp->b_flags & B_CACHE) == 0) {
 671                                 bqrelse(bp);
 672                                 error = bread(ap->a_vp, base_offset,
 673                                               blksize, &bp);
 674                         }
 675                 } else if (offset == 0 && uio->uio_resid >= blksize) {
 676                         /*
 677                          * Even though we are entirely overwriting the buffer
 678                          * we may still have to zero it out to avoid a
 679                          * mmap/write visibility issue.
 680                          */
 681                         bp = getblk(ap->a_vp, base_offset, blksize, GETBLK_BHEAVY, 0);
 682                         if ((bp->b_flags & B_CACHE) == 0)
 683                                 vfs_bio_clrbuf(bp);
 684                 } else if (base_offset >= ip->ino_data.size) {
 685                         /*
 686                          * If the base offset of the buffer is beyond the
 687                          * file EOF, we don't have to issue a read.
 688                          */
 689                         bp = getblk(ap->a_vp, base_offset,
 690                                     blksize, GETBLK_BHEAVY, 0);
 691                         vfs_bio_clrbuf(bp);
 692                 } else {
 693                         /*
 694                          * Partial overwrite, read in any missing bits then
 695                          * replace the portion being written.
 696                          */
 697                         error = bread(ap->a_vp, base_offset, blksize, &bp);
 698                         if (error == 0)
 699                                 bheavy(bp);
 700                 }
 701                 if (error == 0)
 702                         error = uiomovebp(bp, bp->b_data + offset, n, uio);
 703
 704                 lwkt_gettoken(&hmp->fs_token);
 705
 706                 /*
 707                  * Generate REDO records if enabled and redo_count will not
 708                  * exceeded the limit.
 709                  *
 710                  * If redo_count exceeds the limit we stop generating records
 711                  * and clear HAMMER_INODE_REDO.  This will cause the next
 712                  * fsync() to do a full meta-data sync instead of just an
 713                  * UNDO/REDO fifo update.
 714                  *
 715                  * When clearing HAMMER_INODE_REDO any pre-existing REDOs
 716                  * will still be tracked.  The tracks will be terminated
 717                  * when the related meta-data (including possible data
 718                  * modifications which are not tracked via REDO) is
 719                  * flushed.
 720                  */
 721                 if ((ip->flags & HAMMER_INODE_REDO) && error == 0) {
 722                         if (ip->redo_count < hammer_limit_redo) {
 723                                 bp->b_flags |= B_VFSFLAG1;
 724                                 error = hammer_generate_redo(&trans, ip,
 725                                                      base_offset + offset,
 726                                                      HAMMER_REDO_WRITE,
 727                                                      bp->b_data + offset,
 728                                                      (size_t)n);
 729                         } else {
 730                                 ip->flags &= ~HAMMER_INODE_REDO;
 731                         }
 732                 }
 733
 734                 /*
 735                  * If we screwed up we have to undo any VM size changes we
 736                  * made.
 737                  */
 738                 if (error) {
 739                         brelse(bp);
 740                         if (fixsize) {
 741                                 nvtruncbuf(ap->a_vp, ip->ino_data.size,
 742                                           hammer_blocksize(ip->ino_data.size),
 743                                           hammer_blockoff(ip->ino_data.size),
 744                                           0);
 745                         }
 746                         lwkt_reltoken(&hmp->fs_token);
 747                         break;
 748                 }
 749                 kflags |= NOTE_WRITE;
 750                 hammer_stats_file_write += n;
 751                 if (blksize == HAMMER_XBUFSIZE)
 752                         bp->b_flags |= B_CLUSTEROK;
 753                 if (ip->ino_data.size < uio->uio_offset) {
 754                         ip->ino_data.size = uio->uio_offset;
 755                         flags = HAMMER_INODE_SDIRTY;
 756                 } else {
 757                         flags = 0;
 758                 }
 759                 ip->ino_data.mtime = trans.time;
 760                 flags |= HAMMER_INODE_MTIME | HAMMER_INODE_BUFS;
 761                 hammer_modify_inode(&trans, ip, flags);
 762
 763                 /*
 764                  * Once we dirty the buffer any cached zone-X offset
 765                  * becomes invalid.  HAMMER NOTE: no-history mode cannot
 766                  * allow overwriting over the same data sector unless
 767                  * we provide UNDOs for the old data, which we don't.
 768                  */
 769                 bp->b_bio2.bio_offset = NOOFFSET;
 770
 771                 lwkt_reltoken(&hmp->fs_token);
 772
 773                 /*
 774                  * Final buffer disposition.
 775                  *
 776                  * Because meta-data updates are deferred, HAMMER is
 777                  * especially sensitive to excessive bdwrite()s because
 778                  * the I/O stream is not broken up by disk reads.  So the
 779                  * buffer cache simply cannot keep up.
 780                  *
 781                  * WARNING!  blksize is variable.  cluster_write() is
 782                  *           expected to not blow up if it encounters
 783                  *           buffers that do not match the passed blksize.
 784                  *
 785                  * NOTE!  Hammer shouldn't need to bawrite()/cluster_write().
 786                  *        The ip->rsv_recs check should burst-flush the data.
 787                  *        If we queue it immediately the buf could be left
 788                  *        locked on the device queue for a very long time.
 789                  *
 790                  *        However, failing to flush a dirty buffer out when
 791                  *        issued from the pageout daemon can result in a low
 792                  *        memory deadlock against bio_page_alloc(), so we
 793                  *        have to bawrite() on IO_ASYNC as well.
 794                  *
 795                  * NOTE!  To avoid degenerate stalls due to mismatched block
 796                  *        sizes we only honor IO_DIRECT on the write which
 797                  *        abuts the end of the buffer.  However, we must
 798                  *        honor IO_SYNC in case someone is silly enough to
 799                  *        configure a HAMMER file as swap, or when HAMMER
 800                  *        is serving NFS (for commits).  Ick ick.
 801                  */
 802                 bp->b_flags |= B_AGE;
 803                 if (blksize == HAMMER_XBUFSIZE)
 804                         bp->b_flags |= B_CLUSTEROK;
 805
 806                 if (ap->a_ioflag & IO_SYNC) {
 807                         bwrite(bp);
 808                 } else if ((ap->a_ioflag & IO_DIRECT) && endofblk) {
 809                         bawrite(bp);
 810                 } else if (ap->a_ioflag & IO_ASYNC) {
 811                         bawrite(bp);
 812                 } else if (hammer_cluster_enable &&
 813                            !(ap->a_vp->v_mount->mnt_flag & MNT_NOCLUSTERW)) {
 814                         if (base_offset < HAMMER_XDEMARC)
 815                                 cluster_eof = hammer_blockdemarc(base_offset,
 816                                                          ip->ino_data.size);
 817                         else
 818                                 cluster_eof = ip->ino_data.size;
 819                         cluster_write(bp, cluster_eof, blksize, seqcount);
 820                 } else {
 821                         bdwrite(bp);
 822                 }
 823         }
 824         hammer_done_transaction(&trans);
 825         hammer_knote(ap->a_vp, kflags);
 826
 827         return (error);
 828 }
 829
 830 /*
 831  * hammer_vop_access { vp, mode, cred }
 832  *
 833  * MPSAFE - does not require fs_token
 834  */
 835 static
 836 int
 837 hammer_vop_access(struct vop_access_args *ap)
 838 {
 839         struct hammer_inode *ip = VTOI(ap->a_vp);
 840         uid_t uid;
 841         gid_t gid;
 842         int error;
 843
 844         ++hammer_stats_file_iopsr;
 845         uid = hammer_to_unix_xid(&ip->ino_data.uid);
 846         gid = hammer_to_unix_xid(&ip->ino_data.gid);
 847
 848         error = vop_helper_access(ap, uid, gid, ip->ino_data.mode,
 849                                   ip->ino_data.uflags);
 850         return (error);
 851 }
 852
 853 /*
 854  * hammer_vop_advlock { vp, id, op, fl, flags }
 855  *
 856  * MPSAFE - does not require fs_token
 857  */
 858 static
 859 int
 860 hammer_vop_advlock(struct vop_advlock_args *ap)
 861 {
 862         hammer_inode_t ip = VTOI(ap->a_vp);
 863
 864         return (lf_advlock(ap, &ip->advlock, ip->ino_data.size));
 865 }
 866
 867 /*
 868  * hammer_vop_close { vp, fflag }
 869  *
 870  * We can only sync-on-close for normal closes.  XXX disabled for now.
 871  */
 872 static
 873 int
 874 hammer_vop_close(struct vop_close_args *ap)
 875 {
 876 #if 0
 877         struct vnode *vp = ap->a_vp;
 878         hammer_inode_t ip = VTOI(vp);
 879         int waitfor;
 880         if (ip->flags & (HAMMER_INODE_CLOSESYNC|HAMMER_INODE_CLOSEASYNC)) {
 881                 if (vn_islocked(vp) == LK_EXCLUSIVE &&
 882                     (vp->v_flag & (VINACTIVE|VRECLAIMED)) == 0) {
 883                         if (ip->flags & HAMMER_INODE_CLOSESYNC)
 884                                 waitfor = MNT_WAIT;
 885                         else
 886                                 waitfor = MNT_NOWAIT;
 887                         ip->flags &= ~(HAMMER_INODE_CLOSESYNC |
 888                                        HAMMER_INODE_CLOSEASYNC);
 889                         VOP_FSYNC(vp, MNT_NOWAIT, waitfor);
 890                 }
 891         }
 892 #endif
 893         return (vop_stdclose(ap));
 894 }
 895
 896 /*
 897  * hammer_vop_ncreate { nch, dvp, vpp, cred, vap }
 898  *
 899  * The operating system has already ensured that the directory entry
 900  * does not exist and done all appropriate namespace locking.
 901  */
 902 static
 903 int
 904 hammer_vop_ncreate(struct vop_ncreate_args *ap)
 905 {
 906         struct hammer_transaction trans;
 907         struct hammer_inode *dip;
 908         struct hammer_inode *nip;
 909         struct nchandle *nch;
 910         hammer_mount_t hmp;
 911         int error;
 912
 913         nch = ap->a_nch;
 914         dip = VTOI(ap->a_dvp);
 915         hmp = dip->hmp;
 916
 917         if (dip->flags & HAMMER_INODE_RO)
 918                 return (EROFS);
 919         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
 920                 return (error);
 921
 922         /*
 923          * Create a transaction to cover the operations we perform.
 924          */
 925         lwkt_gettoken(&hmp->fs_token);
 926         hammer_start_transaction(&trans, hmp);
 927         ++hammer_stats_file_iopsw;
 928
 929         /*
 930          * Create a new filesystem object of the requested type.  The
 931          * returned inode will be referenced and shared-locked to prevent
 932          * it from being moved to the flusher.
 933          */
 934         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
 935                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
 936                                     NULL, &nip);
 937         if (error) {
 938                 hkprintf("hammer_create_inode error %d\n", error);
 939                 hammer_done_transaction(&trans);
 940                 *ap->a_vpp = NULL;
 941                 lwkt_reltoken(&hmp->fs_token);
 942                 return (error);
 943         }
 944
 945         /*
 946          * Add the new filesystem object to the directory.  This will also
 947          * bump the inode's link count.
 948          */
 949         error = hammer_ip_add_directory(&trans, dip,
 950                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
 951                                         nip);
 952         if (error)
 953                 hkprintf("hammer_ip_add_directory error %d\n", error);
 954
 955         /*
 956          * Finish up.
 957          */
 958         if (error) {
 959                 hammer_rel_inode(nip, 0);
 960                 hammer_done_transaction(&trans);
 961                 *ap->a_vpp = NULL;
 962         } else {
 963                 error = hammer_get_vnode(nip, ap->a_vpp);
 964                 hammer_done_transaction(&trans);
 965                 hammer_rel_inode(nip, 0);
 966                 if (error == 0) {
 967                         cache_setunresolved(ap->a_nch);
 968                         cache_setvp(ap->a_nch, *ap->a_vpp);
 969                 }
 970                 hammer_knote(ap->a_dvp, NOTE_WRITE);
 971         }
 972         lwkt_reltoken(&hmp->fs_token);
 973         return (error);
 974 }
 975
 976 /*
 977  * hammer_vop_getattr { vp, vap }
 978  *
 979  * Retrieve an inode's attribute information.  When accessing inodes
 980  * historically we fake the atime field to ensure consistent results.
 981  * The atime field is stored in the B-Tree element and allowed to be
 982  * updated without cycling the element.
 983  *
 984  * MPSAFE - does not require fs_token
 985  */
 986 static
 987 int
 988 hammer_vop_getattr(struct vop_getattr_args *ap)
 989 {
 990         struct hammer_inode *ip = VTOI(ap->a_vp);
 991         struct vattr *vap = ap->a_vap;
 992
 993         /*
 994          * We want the fsid to be different when accessing a filesystem
 995          * with different as-of's so programs like diff don't think
 996          * the files are the same.
 997          *
 998          * We also want the fsid to be the same when comparing snapshots,
 999          * or when comparing mirrors (which might be backed by different
1000          * physical devices).  HAMMER fsids are based on the PFS's
1001          * shared_uuid field.
1002          *
1003          * XXX there is a chance of collision here.  The va_fsid reported
1004          * by stat is different from the more involved fsid used in the
1005          * mount structure.
1006          */
1007         ++hammer_stats_file_iopsr;
1008         hammer_lock_sh(&ip->lock);
1009         vap->va_fsid = ip->pfsm->fsid_udev ^ (uint32_t)ip->obj_asof ^
1010                        (uint32_t)(ip->obj_asof >> 32);
1011
1012         vap->va_fileid = ip->ino_leaf.base.obj_id;
1013         vap->va_mode = ip->ino_data.mode;
1014         vap->va_nlink = ip->ino_data.nlinks;
1015         vap->va_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1016         vap->va_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1017         vap->va_rmajor = 0;
1018         vap->va_rminor = 0;
1019         vap->va_size = ip->ino_data.size;
1020
1021         /*
1022          * Special case for @@PFS softlinks.  The actual size of the
1023          * expanded softlink is "@@0x%016llx:%05d" == 26 bytes.
1024          * or for MAX_TID is    "@@-1:%05d" == 10 bytes.
1025          *
1026          * Note that userspace hammer command does not allow users to
1027          * create a @@PFS softlink under an existing other PFS (id!=0)
1028          * so the ip localization here for @@PFS softlink is always 0.
1029          */
1030         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_SOFTLINK &&
1031             ip->ino_data.size == 10 &&
1032             ip->obj_asof == HAMMER_MAX_TID &&
1033             ip->obj_localization == HAMMER_DEF_LOCALIZATION &&
1034             strncmp(ip->ino_data.ext.symlink, "@@PFS", 5) == 0) {
1035                     if (ip->pfsm->pfsd.mirror_flags & HAMMER_PFSD_SLAVE)
1036                             vap->va_size = 26;
1037                     else
1038                             vap->va_size = 10;
1039         }
1040
1041         /*
1042          * We must provide a consistent atime and mtime for snapshots
1043          * so people can do a 'tar cf - ... | md5' on them and get
1044          * consistent results.
1045          */
1046         if (ip->flags & HAMMER_INODE_RO) {
1047                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_atime);
1048                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_mtime);
1049         } else {
1050                 hammer_time_to_timespec(ip->ino_data.atime, &vap->va_atime);
1051                 hammer_time_to_timespec(ip->ino_data.mtime, &vap->va_mtime);
1052         }
1053         hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_ctime);
1054         vap->va_flags = ip->ino_data.uflags;
1055         vap->va_gen = 1;        /* hammer inums are unique for all time */
1056         vap->va_blocksize = HAMMER_BUFSIZE;
1057         if (ip->ino_data.size >= HAMMER_XDEMARC) {
1058                 vap->va_bytes = (ip->ino_data.size + HAMMER_XBUFMASK64) &
1059                                 ~HAMMER_XBUFMASK64;
1060         } else if (ip->ino_data.size > HAMMER_HBUFSIZE) {
1061                 vap->va_bytes = (ip->ino_data.size + HAMMER_BUFMASK64) &
1062                                 ~HAMMER_BUFMASK64;
1063         } else {
1064                 vap->va_bytes = (ip->ino_data.size + 15) & ~15;
1065         }
1066
1067         vap->va_type = hammer_get_vnode_type(ip->ino_data.obj_type);
1068         vap->va_filerev = 0;    /* XXX */
1069         vap->va_uid_uuid = ip->ino_data.uid;
1070         vap->va_gid_uuid = ip->ino_data.gid;
1071         vap->va_fsid_uuid = ip->hmp->fsid;
1072         vap->va_vaflags = VA_UID_UUID_VALID | VA_GID_UUID_VALID |
1073                           VA_FSID_UUID_VALID;
1074
1075         switch (ip->ino_data.obj_type) {
1076         case HAMMER_OBJTYPE_CDEV:
1077         case HAMMER_OBJTYPE_BDEV:
1078                 vap->va_rmajor = ip->ino_data.rmajor;
1079                 vap->va_rminor = ip->ino_data.rminor;
1080                 break;
1081         default:
1082                 break;
1083         }
1084         hammer_unlock(&ip->lock);
1085         return(0);
1086 }
1087
1088 /*
1089  * hammer_vop_nresolve { nch, dvp, cred }
1090  *
1091  * Locate the requested directory entry.
1092  */
1093 static
1094 int
1095 hammer_vop_nresolve(struct vop_nresolve_args *ap)
1096 {
1097         struct hammer_transaction trans;
1098         struct namecache *ncp;
1099         hammer_mount_t hmp;
1100         hammer_inode_t dip;
1101         hammer_inode_t ip;
1102         hammer_tid_t asof;
1103         struct hammer_cursor cursor;
1104         struct vnode *vp;
1105         int64_t namekey;
1106         int error;
1107         int i;
1108         int nlen;
1109         int flags;
1110         int ispfs;
1111         int64_t obj_id;
1112         uint32_t localization;
1113         uint32_t max_iterations;
1114
1115         /*
1116          * Misc initialization, plus handle as-of name extensions.  Look for
1117          * the '@@' extension.  Note that as-of files and directories cannot
1118          * be modified.
1119          */
1120         dip = VTOI(ap->a_dvp);
1121         ncp = ap->a_nch->ncp;
1122         asof = dip->obj_asof;
1123         localization = dip->obj_localization;   /* for code consistency */
1124         nlen = ncp->nc_nlen;
1125         flags = dip->flags & HAMMER_INODE_RO;
1126         ispfs = 0;
1127         hmp = dip->hmp;
1128
1129         lwkt_gettoken(&hmp->fs_token);
1130         hammer_simple_transaction(&trans, hmp);
1131         ++hammer_stats_file_iopsr;
1132
1133         for (i = 0; i < nlen; ++i) {
1134                 if (ncp->nc_name[i] == '@' && ncp->nc_name[i+1] == '@') {
1135                         error = hammer_str_to_tid(ncp->nc_name + i + 2,
1136                                                   &ispfs, &asof, &localization);
1137                         if (error != 0) {
1138                                 i = nlen;
1139                                 break;
1140                         }
1141                         if (asof != HAMMER_MAX_TID)
1142                                 flags |= HAMMER_INODE_RO;
1143                         break;
1144                 }
1145         }
1146         nlen = i;
1147
1148         /*
1149          * If this is a PFS softlink we dive into the PFS
1150          */
1151         if (ispfs && nlen == 0) {
1152                 ip = hammer_get_inode(&trans, dip, HAMMER_OBJID_ROOT,
1153                                       asof, localization,
1154                                       flags, &error);
1155                 if (error == 0) {
1156                         error = hammer_get_vnode(ip, &vp);
1157                         hammer_rel_inode(ip, 0);
1158                 } else {
1159                         vp = NULL;
1160                 }
1161                 if (error == 0) {
1162                         vn_unlock(vp);
1163                         cache_setvp(ap->a_nch, vp);
1164                         vrele(vp);
1165                 }
1166                 goto done;
1167         }
1168
1169         /*
1170          * If there is no path component the time extension is relative to dip.
1171          * e.g. "fubar/@@<snapshot>"
1172          *
1173          * "." is handled by the kernel, but ".@@<snapshot>" is not.
1174          * e.g. "fubar/.@@<snapshot>"
1175          *
1176          * ".." is handled by the kernel.  We do not currently handle
1177          * "..@<snapshot>".
1178          */
1179         if (nlen == 0 || (nlen == 1 && ncp->nc_name[0] == '.')) {
1180                 ip = hammer_get_inode(&trans, dip, dip->obj_id,
1181                                       asof, dip->obj_localization,
1182                                       flags, &error);
1183                 if (error == 0) {
1184                         error = hammer_get_vnode(ip, &vp);
1185                         hammer_rel_inode(ip, 0);
1186                 } else {
1187                         vp = NULL;
1188                 }
1189                 if (error == 0) {
1190                         vn_unlock(vp);
1191                         cache_setvp(ap->a_nch, vp);
1192                         vrele(vp);
1193                 }
1194                 goto done;
1195         }
1196
1197         /*
1198          * Calculate the namekey and setup the key range for the scan.  This
1199          * works kinda like a chained hash table where the lower 32 bits
1200          * of the namekey synthesize the chain.
1201          *
1202          * The key range is inclusive of both key_beg and key_end.
1203          */
1204         namekey = hammer_directory_namekey(dip, ncp->nc_name, nlen,
1205                                            &max_iterations);
1206
1207         error = hammer_init_cursor(&trans, &cursor, &dip->cache[1], dip);
1208         cursor.key_beg.localization = dip->obj_localization |
1209                                       hammer_dir_localization(dip);
1210         cursor.key_beg.obj_id = dip->obj_id;
1211         cursor.key_beg.key = namekey;
1212         cursor.key_beg.create_tid = 0;
1213         cursor.key_beg.delete_tid = 0;
1214         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1215         cursor.key_beg.obj_type = 0;
1216
1217         cursor.key_end = cursor.key_beg;
1218         cursor.key_end.key += max_iterations;
1219         cursor.asof = asof;
1220         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1221
1222         /*
1223          * Scan all matching records (the chain), locate the one matching
1224          * the requested path component.
1225          *
1226          * The hammer_ip_*() functions merge in-memory records with on-disk
1227          * records for the purposes of the search.
1228          */
1229         obj_id = 0;
1230         localization = HAMMER_DEF_LOCALIZATION;
1231
1232         if (error == 0) {
1233                 error = hammer_ip_first(&cursor);
1234                 while (error == 0) {
1235                         error = hammer_ip_resolve_data(&cursor);
1236                         if (error)
1237                                 break;
1238                         if (nlen == cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF &&
1239                             bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1240                                 obj_id = cursor.data->entry.obj_id;
1241                                 localization = cursor.data->entry.localization;
1242                                 break;
1243                         }
1244                         error = hammer_ip_next(&cursor);
1245                 }
1246         }
1247         hammer_done_cursor(&cursor);
1248
1249         /*
1250          * Lookup the obj_id.  This should always succeed.  If it does not
1251          * the filesystem may be damaged and we return a dummy inode.
1252          */
1253         if (error == 0) {
1254                 ip = hammer_get_inode(&trans, dip, obj_id,
1255                                       asof, localization,
1256                                       flags, &error);
1257                 if (error == ENOENT) {
1258                         hkprintf("WARNING: Missing inode for dirent \"%s\"\n"
1259                                 "\tobj_id = %016llx, asof=%016llx, lo=%08x\n",
1260                                 ncp->nc_name,
1261                                 (long long)obj_id, (long long)asof,
1262                                 localization);
1263                         error = 0;
1264                         ip = hammer_get_dummy_inode(&trans, dip, obj_id,
1265                                                     asof, localization,
1266                                                     flags, &error);
1267                 }
1268                 if (error == 0) {
1269                         error = hammer_get_vnode(ip, &vp);
1270                         hammer_rel_inode(ip, 0);
1271                 } else {
1272                         vp = NULL;
1273                 }
1274                 if (error == 0) {
1275                         vn_unlock(vp);
1276                         cache_setvp(ap->a_nch, vp);
1277                         vrele(vp);
1278                 }
1279         } else if (error == ENOENT) {
1280                 cache_setvp(ap->a_nch, NULL);
1281         }
1282 done:
1283         hammer_done_transaction(&trans);
1284         lwkt_reltoken(&hmp->fs_token);
1285         return (error);
1286 }
1287
1288 /*
1289  * hammer_vop_nlookupdotdot { dvp, vpp, cred }
1290  *
1291  * Locate the parent directory of a directory vnode.
1292  *
1293  * dvp is referenced but not locked.  *vpp must be returned referenced and
1294  * locked.  A parent_obj_id of 0 does not necessarily indicate that we are
1295  * at the root, instead it could indicate that the directory we were in was
1296  * removed.
1297  *
1298  * NOTE: as-of sequences are not linked into the directory structure.  If
1299  * we are at the root with a different asof then the mount point, reload
1300  * the same directory with the mount point's asof.   I'm not sure what this
1301  * will do to NFS.  We encode ASOF stamps in NFS file handles so it might not
1302  * get confused, but it hasn't been tested.
1303  */
1304 static
1305 int
1306 hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *ap)
1307 {
1308         struct hammer_transaction trans;
1309         struct hammer_inode *dip;
1310         struct hammer_inode *ip;
1311         hammer_mount_t hmp;
1312         int64_t parent_obj_id;
1313         uint32_t parent_obj_localization;
1314         hammer_tid_t asof;
1315         int error;
1316
1317         dip = VTOI(ap->a_dvp);
1318         asof = dip->obj_asof;
1319         hmp = dip->hmp;
1320
1321         /*
1322          * Whos are parent?  This could be the root of a pseudo-filesystem
1323          * whos parent is in another localization domain.
1324          */
1325         lwkt_gettoken(&hmp->fs_token);
1326         parent_obj_id = dip->ino_data.parent_obj_id;
1327         if (dip->obj_id == HAMMER_OBJID_ROOT)
1328                 parent_obj_localization = HAMMER_DEF_LOCALIZATION;
1329         else
1330                 parent_obj_localization = dip->obj_localization;
1331
1332         /*
1333          * It's probably a PFS root when dip->ino_data.parent_obj_id is 0.
1334          */
1335         if (parent_obj_id == 0) {
1336                 if (dip->obj_id == HAMMER_OBJID_ROOT &&
1337                    asof != hmp->asof) {
1338                         parent_obj_id = dip->obj_id;
1339                         asof = hmp->asof;
1340                         *ap->a_fakename = kmalloc(19, M_TEMP, M_WAITOK);
1341                         ksnprintf(*ap->a_fakename, 19, "0x%016llx",
1342                                   (long long)dip->obj_asof);
1343                 } else {
1344                         *ap->a_vpp = NULL;
1345                         lwkt_reltoken(&hmp->fs_token);
1346                         return ENOENT;
1347                 }
1348         }
1349
1350         hammer_simple_transaction(&trans, hmp);
1351         ++hammer_stats_file_iopsr;
1352
1353         ip = hammer_get_inode(&trans, dip, parent_obj_id,
1354                               asof, parent_obj_localization,
1355                               dip->flags, &error);
1356         if (ip) {
1357                 error = hammer_get_vnode(ip, ap->a_vpp);
1358                 hammer_rel_inode(ip, 0);
1359         } else {
1360                 *ap->a_vpp = NULL;
1361         }
1362         hammer_done_transaction(&trans);
1363         lwkt_reltoken(&hmp->fs_token);
1364         return (error);
1365 }
1366
1367 /*
1368  * hammer_vop_nlink { nch, dvp, vp, cred }
1369  */
1370 static
1371 int
1372 hammer_vop_nlink(struct vop_nlink_args *ap)
1373 {
1374         struct hammer_transaction trans;
1375         struct hammer_inode *dip;
1376         struct hammer_inode *ip;
1377         struct nchandle *nch;
1378         hammer_mount_t hmp;
1379         int error;
1380
1381         if (ap->a_dvp->v_mount != ap->a_vp->v_mount)
1382                 return(EXDEV);
1383
1384         nch = ap->a_nch;
1385         dip = VTOI(ap->a_dvp);
1386         ip = VTOI(ap->a_vp);
1387         hmp = dip->hmp;
1388
1389         if (dip->obj_localization != ip->obj_localization)
1390                 return(EXDEV);
1391
1392         if (dip->flags & HAMMER_INODE_RO)
1393                 return (EROFS);
1394         if (ip->flags & HAMMER_INODE_RO)
1395                 return (EROFS);
1396         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1397                 return (error);
1398
1399         /*
1400          * Create a transaction to cover the operations we perform.
1401          */
1402         lwkt_gettoken(&hmp->fs_token);
1403         hammer_start_transaction(&trans, hmp);
1404         ++hammer_stats_file_iopsw;
1405
1406         /*
1407          * Add the filesystem object to the directory.  Note that neither
1408          * dip nor ip are referenced or locked, but their vnodes are
1409          * referenced.  This function will bump the inode's link count.
1410          */
1411         error = hammer_ip_add_directory(&trans, dip,
1412                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1413                                         ip);
1414
1415         /*
1416          * Finish up.
1417          */
1418         if (error == 0) {
1419                 cache_setunresolved(nch);
1420                 cache_setvp(nch, ap->a_vp);
1421         }
1422         hammer_done_transaction(&trans);
1423         hammer_knote(ap->a_vp, NOTE_LINK);
1424         hammer_knote(ap->a_dvp, NOTE_WRITE);
1425         lwkt_reltoken(&hmp->fs_token);
1426         return (error);
1427 }
1428
1429 /*
1430  * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
1431  *
1432  * The operating system has already ensured that the directory entry
1433  * does not exist and done all appropriate namespace locking.
1434  */
1435 static
1436 int
1437 hammer_vop_nmkdir(struct vop_nmkdir_args *ap)
1438 {
1439         struct hammer_transaction trans;
1440         struct hammer_inode *dip;
1441         struct hammer_inode *nip;
1442         struct nchandle *nch;
1443         hammer_mount_t hmp;
1444         int error;
1445
1446         nch = ap->a_nch;
1447         dip = VTOI(ap->a_dvp);
1448         hmp = dip->hmp;
1449
1450         if (dip->flags & HAMMER_INODE_RO)
1451                 return (EROFS);
1452         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1453                 return (error);
1454
1455         /*
1456          * Create a transaction to cover the operations we perform.
1457          */
1458         lwkt_gettoken(&hmp->fs_token);
1459         hammer_start_transaction(&trans, hmp);
1460         ++hammer_stats_file_iopsw;
1461
1462         /*
1463          * Create a new filesystem object of the requested type.  The
1464          * returned inode will be referenced but not locked.
1465          */
1466         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1467                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1468                                     NULL, &nip);
1469         if (error) {
1470                 hammer_done_transaction(&trans);
1471                 *ap->a_vpp = NULL;
1472                 lwkt_reltoken(&hmp->fs_token);
1473                 return (error);
1474         }
1475         /*
1476          * Add the new filesystem object to the directory.  This will also
1477          * bump the inode's link count.
1478          */
1479         error = hammer_ip_add_directory(&trans, dip,
1480                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1481                                         nip);
1482         if (error)
1483                 hkprintf("hammer_mkdir (add) error %d\n", error);
1484
1485         /*
1486          * Finish up.
1487          */
1488         if (error) {
1489                 hammer_rel_inode(nip, 0);
1490                 *ap->a_vpp = NULL;
1491         } else {
1492                 error = hammer_get_vnode(nip, ap->a_vpp);
1493                 hammer_rel_inode(nip, 0);
1494                 if (error == 0) {
1495                         cache_setunresolved(ap->a_nch);
1496                         cache_setvp(ap->a_nch, *ap->a_vpp);
1497                 }
1498         }
1499         hammer_done_transaction(&trans);
1500         if (error == 0)
1501                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
1502         lwkt_reltoken(&hmp->fs_token);
1503         return (error);
1504 }
1505
1506 /*
1507  * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1508  *
1509  * The operating system has already ensured that the directory entry
1510  * does not exist and done all appropriate namespace locking.
1511  */
1512 static
1513 int
1514 hammer_vop_nmknod(struct vop_nmknod_args *ap)
1515 {
1516         struct hammer_transaction trans;
1517         struct hammer_inode *dip;
1518         struct hammer_inode *nip;
1519         struct nchandle *nch;
1520         hammer_mount_t hmp;
1521         int error;
1522
1523         nch = ap->a_nch;
1524         dip = VTOI(ap->a_dvp);
1525         hmp = dip->hmp;
1526
1527         if (dip->flags & HAMMER_INODE_RO)
1528                 return (EROFS);
1529         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1530                 return (error);
1531
1532         /*
1533          * Create a transaction to cover the operations we perform.
1534          */
1535         lwkt_gettoken(&hmp->fs_token);
1536         hammer_start_transaction(&trans, hmp);
1537         ++hammer_stats_file_iopsw;
1538
1539         /*
1540          * Create a new filesystem object of the requested type.  The
1541          * returned inode will be referenced but not locked.
1542          *
1543          * If mknod specifies a directory a pseudo-fs is created.
1544          */
1545         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1546                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1547                                     NULL, &nip);
1548         if (error) {
1549                 hammer_done_transaction(&trans);
1550                 *ap->a_vpp = NULL;
1551                 lwkt_reltoken(&hmp->fs_token);
1552                 return (error);
1553         }
1554
1555         /*
1556          * Add the new filesystem object to the directory.  This will also
1557          * bump the inode's link count.
1558          */
1559         error = hammer_ip_add_directory(&trans, dip,
1560                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1561                                         nip);
1562
1563         /*
1564          * Finish up.
1565          */
1566         if (error) {
1567                 hammer_rel_inode(nip, 0);
1568                 *ap->a_vpp = NULL;
1569         } else {
1570                 error = hammer_get_vnode(nip, ap->a_vpp);
1571                 hammer_rel_inode(nip, 0);
1572                 if (error == 0) {
1573                         cache_setunresolved(ap->a_nch);
1574                         cache_setvp(ap->a_nch, *ap->a_vpp);
1575                 }
1576         }
1577         hammer_done_transaction(&trans);
1578         if (error == 0)
1579                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1580         lwkt_reltoken(&hmp->fs_token);
1581         return (error);
1582 }
1583
1584 /*
1585  * hammer_vop_open { vp, mode, cred, fp }
1586  *
1587  * MPSAFE (does not require fs_token)
1588  */
1589 static
1590 int
1591 hammer_vop_open(struct vop_open_args *ap)
1592 {
1593         hammer_inode_t ip;
1594
1595         ++hammer_stats_file_iopsr;
1596         ip = VTOI(ap->a_vp);
1597
1598         if ((ap->a_mode & FWRITE) && (ip->flags & HAMMER_INODE_RO))
1599                 return (EROFS);
1600         return(vop_stdopen(ap));
1601 }
1602
1603 /*
1604  * hammer_vop_print { vp }
1605  */
1606 static
1607 int
1608 hammer_vop_print(struct vop_print_args *ap)
1609 {
1610         return EOPNOTSUPP;
1611 }
1612
1613 /*
1614  * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1615  */
1616 static
1617 int
1618 hammer_vop_readdir(struct vop_readdir_args *ap)
1619 {
1620         struct hammer_transaction trans;
1621         struct hammer_cursor cursor;
1622         struct hammer_inode *ip;
1623         hammer_mount_t hmp;
1624         struct uio *uio;
1625         hammer_base_elm_t base;
1626         int error;
1627         int cookie_index;
1628         int ncookies;
1629         off_t *cookies;
1630         off_t saveoff;
1631         int r;
1632         int dtype;
1633
1634         ++hammer_stats_file_iopsr;
1635         ip = VTOI(ap->a_vp);
1636         uio = ap->a_uio;
1637         saveoff = uio->uio_offset;
1638         hmp = ip->hmp;
1639
1640         if (ap->a_ncookies) {
1641                 ncookies = uio->uio_resid / 16 + 1;
1642                 if (ncookies > 1024)
1643                         ncookies = 1024;
1644                 cookies = kmalloc(ncookies * sizeof(off_t), M_TEMP, M_WAITOK);
1645                 cookie_index = 0;
1646         } else {
1647                 ncookies = -1;
1648                 cookies = NULL;
1649                 cookie_index = 0;
1650         }
1651
1652         lwkt_gettoken(&hmp->fs_token);
1653         hammer_simple_transaction(&trans, hmp);
1654
1655         /*
1656          * Handle artificial entries
1657          *
1658          * It should be noted that the minimum value for a directory
1659          * hash key on-media is 0x0000000100000000, so we can use anything
1660          * less then that to represent our 'special' key space.
1661          */
1662         error = 0;
1663         if (saveoff == 0) {
1664                 r = vop_write_dirent(&error, uio, ip->obj_id, DT_DIR, 1, ".");
1665                 if (r)
1666                         goto done;
1667                 if (cookies)
1668                         cookies[cookie_index] = saveoff;
1669                 ++saveoff;
1670                 ++cookie_index;
1671                 if (cookie_index == ncookies)
1672                         goto done;
1673         }
1674         if (saveoff == 1) {
1675                 if (ip->ino_data.parent_obj_id) {
1676                         r = vop_write_dirent(&error, uio,
1677                                              ip->ino_data.parent_obj_id,
1678                                              DT_DIR, 2, "..");
1679                 } else {
1680                         r = vop_write_dirent(&error, uio,
1681                                              ip->obj_id, DT_DIR, 2, "..");
1682                 }
1683                 if (r)
1684                         goto done;
1685                 if (cookies)
1686                         cookies[cookie_index] = saveoff;
1687                 ++saveoff;
1688                 ++cookie_index;
1689                 if (cookie_index == ncookies)
1690                         goto done;
1691         }
1692
1693         /*
1694          * Key range (begin and end inclusive) to scan.  Directory keys
1695          * directly translate to a 64 bit 'seek' position.
1696          */
1697         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1698         cursor.key_beg.localization = ip->obj_localization |
1699                                       hammer_dir_localization(ip);
1700         cursor.key_beg.obj_id = ip->obj_id;
1701         cursor.key_beg.create_tid = 0;
1702         cursor.key_beg.delete_tid = 0;
1703         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1704         cursor.key_beg.obj_type = 0;
1705         cursor.key_beg.key = saveoff;
1706
1707         cursor.key_end = cursor.key_beg;
1708         cursor.key_end.key = HAMMER_MAX_KEY;
1709         cursor.asof = ip->obj_asof;
1710         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1711
1712         error = hammer_ip_first(&cursor);
1713
1714         while (error == 0) {
1715                 error = hammer_ip_resolve_data(&cursor);
1716                 if (error)
1717                         break;
1718                 base = &cursor.leaf->base;
1719                 saveoff = base->key;
1720                 KKASSERT(cursor.leaf->data_len > HAMMER_ENTRY_NAME_OFF);
1721
1722                 if (base->obj_id != ip->obj_id)
1723                         hpanic("bad record at %p", cursor.node);
1724
1725                 /*
1726                  * Convert pseudo-filesystems into softlinks
1727                  */
1728                 dtype = hammer_get_dtype(cursor.leaf->base.obj_type);
1729                 r = vop_write_dirent(
1730                              &error, uio, cursor.data->entry.obj_id,
1731                              dtype,
1732                              cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF ,
1733                              (void *)cursor.data->entry.name);
1734                 if (r)
1735                         break;
1736                 ++saveoff;
1737                 if (cookies)
1738                         cookies[cookie_index] = base->key;
1739                 ++cookie_index;
1740                 if (cookie_index == ncookies)
1741                         break;
1742                 error = hammer_ip_next(&cursor);
1743         }
1744         hammer_done_cursor(&cursor);
1745
1746 done:
1747         hammer_done_transaction(&trans);
1748
1749         if (ap->a_eofflag)
1750                 *ap->a_eofflag = (error == ENOENT);
1751         uio->uio_offset = saveoff;
1752         if (error && cookie_index == 0) {
1753                 if (error == ENOENT)
1754                         error = 0;
1755                 if (cookies) {
1756                         kfree(cookies, M_TEMP);
1757                         *ap->a_ncookies = 0;
1758                         *ap->a_cookies = NULL;
1759                 }
1760         } else {
1761                 if (error == ENOENT)
1762                         error = 0;
1763                 if (cookies) {
1764                         *ap->a_ncookies = cookie_index;
1765                         *ap->a_cookies = cookies;
1766                 }
1767         }
1768         lwkt_reltoken(&hmp->fs_token);
1769         return(error);
1770 }
1771
1772 /*
1773  * hammer_vop_readlink { vp, uio, cred }
1774  */
1775 static
1776 int
1777 hammer_vop_readlink(struct vop_readlink_args *ap)
1778 {
1779         struct hammer_transaction trans;
1780         struct hammer_cursor cursor;
1781         struct hammer_inode *ip;
1782         hammer_mount_t hmp;
1783         char buf[32];
1784         uint32_t localization;
1785         hammer_pseudofs_inmem_t pfsm;
1786         int error;
1787
1788         ip = VTOI(ap->a_vp);
1789         hmp = ip->hmp;
1790
1791         lwkt_gettoken(&hmp->fs_token);
1792
1793         /*
1794          * Shortcut if the symlink data was stuffed into ino_data.
1795          *
1796          * Also expand special "@@PFS%05d" softlinks (expansion only
1797          * occurs for non-historical (current) accesses made from the
1798          * primary filesystem).
1799          *
1800          * Note that userspace hammer command does not allow users to
1801          * create a @@PFS softlink under an existing other PFS (id!=0)
1802          * so the ip localization here for @@PFS softlink is always 0.
1803          */
1804         if (ip->ino_data.size <= HAMMER_INODE_BASESYMLEN) {
1805                 char *ptr;
1806                 int bytes;
1807
1808                 ptr = ip->ino_data.ext.symlink;
1809                 bytes = (int)ip->ino_data.size;
1810                 if (bytes == 10 &&
1811                     ip->obj_asof == HAMMER_MAX_TID &&
1812                     ip->obj_localization == HAMMER_DEF_LOCALIZATION &&
1813                     strncmp(ptr, "@@PFS", 5) == 0) {
1814                         hammer_simple_transaction(&trans, hmp);
1815                         bcopy(ptr + 5, buf, 5);
1816                         buf[5] = 0;
1817                         localization = pfs_to_lo(strtoul(buf, NULL, 10));
1818                         pfsm = hammer_load_pseudofs(&trans, localization,
1819                                                     &error);
1820                         if (error == 0) {
1821                                 if (pfsm->pfsd.mirror_flags &
1822                                     HAMMER_PFSD_SLAVE) {
1823                                         /* vap->va_size == 26 */
1824                                         ksnprintf(buf, sizeof(buf),
1825                                                   "@@0x%016llx:%05d",
1826                                                   (long long)pfsm->pfsd.sync_end_tid,
1827                                                   lo_to_pfs(localization));
1828                                 } else {
1829                                         /* vap->va_size == 10 */
1830                                         ksnprintf(buf, sizeof(buf),
1831                                                   "@@-1:%05d",
1832                                                   lo_to_pfs(localization));
1833                                 }
1834                                 ptr = buf;
1835                                 bytes = strlen(buf);
1836                         }
1837                         if (pfsm)
1838                                 hammer_rel_pseudofs(hmp, pfsm);
1839                         hammer_done_transaction(&trans);
1840                 }
1841                 error = uiomove(ptr, bytes, ap->a_uio);
1842                 lwkt_reltoken(&hmp->fs_token);
1843                 return(error);
1844         }
1845
1846         /*
1847          * Long version
1848          */
1849         hammer_simple_transaction(&trans, hmp);
1850         ++hammer_stats_file_iopsr;
1851         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1852
1853         /*
1854          * Key range (begin and end inclusive) to scan.  Directory keys
1855          * directly translate to a 64 bit 'seek' position.
1856          */
1857         cursor.key_beg.localization = ip->obj_localization |
1858                                       HAMMER_LOCALIZE_MISC;
1859         cursor.key_beg.obj_id = ip->obj_id;
1860         cursor.key_beg.create_tid = 0;
1861         cursor.key_beg.delete_tid = 0;
1862         cursor.key_beg.rec_type = HAMMER_RECTYPE_FIX;
1863         cursor.key_beg.obj_type = 0;
1864         cursor.key_beg.key = HAMMER_FIXKEY_SYMLINK;
1865         cursor.asof = ip->obj_asof;
1866         cursor.flags |= HAMMER_CURSOR_ASOF;
1867
1868         error = hammer_ip_lookup(&cursor);
1869         if (error == 0) {
1870                 error = hammer_ip_resolve_data(&cursor);
1871                 if (error == 0) {
1872                         KKASSERT(cursor.leaf->data_len >=
1873                                  HAMMER_SYMLINK_NAME_OFF);
1874                         error = uiomove(cursor.data->symlink.name,
1875                                         cursor.leaf->data_len -
1876                                                 HAMMER_SYMLINK_NAME_OFF,
1877                                         ap->a_uio);
1878                 }
1879         }
1880         hammer_done_cursor(&cursor);
1881         hammer_done_transaction(&trans);
1882         lwkt_reltoken(&hmp->fs_token);
1883         return(error);
1884 }
1885
1886 /*
1887  * hammer_vop_nremove { nch, dvp, cred }
1888  */
1889 static
1890 int
1891 hammer_vop_nremove(struct vop_nremove_args *ap)
1892 {
1893         struct hammer_transaction trans;
1894         struct hammer_inode *dip;
1895         hammer_mount_t hmp;
1896         int error;
1897
1898         dip = VTOI(ap->a_dvp);
1899         hmp = dip->hmp;
1900
1901         if (hammer_nohistory(dip) == 0 &&
1902             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1903                 return (error);
1904         }
1905
1906         lwkt_gettoken(&hmp->fs_token);
1907         hammer_start_transaction(&trans, hmp);
1908         ++hammer_stats_file_iopsw;
1909         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 0);
1910         hammer_done_transaction(&trans);
1911         if (error == 0)
1912                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1913         lwkt_reltoken(&hmp->fs_token);
1914         return (error);
1915 }
1916
1917 /*
1918  * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1919  */
1920 static
1921 int
1922 hammer_vop_nrename(struct vop_nrename_args *ap)
1923 {
1924         struct hammer_transaction trans;
1925         struct namecache *fncp;
1926         struct namecache *tncp;
1927         struct hammer_inode *fdip;
1928         struct hammer_inode *tdip;
1929         struct hammer_inode *ip;
1930         hammer_mount_t hmp;
1931         struct hammer_cursor cursor;
1932         int64_t namekey;
1933         uint32_t max_iterations;
1934         int nlen, error;
1935
1936         if (ap->a_fdvp->v_mount != ap->a_tdvp->v_mount)
1937                 return(EXDEV);
1938         if (ap->a_fdvp->v_mount != ap->a_fnch->ncp->nc_vp->v_mount)
1939                 return(EXDEV);
1940
1941         fdip = VTOI(ap->a_fdvp);
1942         tdip = VTOI(ap->a_tdvp);
1943         fncp = ap->a_fnch->ncp;
1944         tncp = ap->a_tnch->ncp;
1945         ip = VTOI(fncp->nc_vp);
1946         KKASSERT(ip != NULL);
1947
1948         hmp = ip->hmp;
1949
1950         if (fdip->obj_localization != tdip->obj_localization)
1951                 return(EXDEV);
1952         if (fdip->obj_localization != ip->obj_localization)
1953                 return(EXDEV);
1954
1955         if (fdip->flags & HAMMER_INODE_RO)
1956                 return (EROFS);
1957         if (tdip->flags & HAMMER_INODE_RO)
1958                 return (EROFS);
1959         if (ip->flags & HAMMER_INODE_RO)
1960                 return (EROFS);
1961         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1962                 return (error);
1963
1964         lwkt_gettoken(&hmp->fs_token);
1965         hammer_start_transaction(&trans, hmp);
1966         ++hammer_stats_file_iopsw;
1967
1968         /*
1969          * Remove tncp from the target directory and then link ip as
1970          * tncp. XXX pass trans to dounlink
1971          *
1972          * Force the inode sync-time to match the transaction so it is
1973          * in-sync with the creation of the target directory entry.
1974          */
1975         error = hammer_dounlink(&trans, ap->a_tnch, ap->a_tdvp,
1976                                 ap->a_cred, 0, -1);
1977         if (error == 0 || error == ENOENT) {
1978                 error = hammer_ip_add_directory(&trans, tdip,
1979                                                 tncp->nc_name, tncp->nc_nlen,
1980                                                 ip);
1981                 if (error == 0) {
1982                         ip->ino_data.parent_obj_id = tdip->obj_id;
1983                         ip->ino_data.ctime = trans.time;
1984                         hammer_modify_inode(&trans, ip, HAMMER_INODE_DDIRTY);
1985                 }
1986         }
1987         if (error)
1988                 goto failed; /* XXX */
1989
1990         /*
1991          * Locate the record in the originating directory and remove it.
1992          *
1993          * Calculate the namekey and setup the key range for the scan.  This
1994          * works kinda like a chained hash table where the lower 32 bits
1995          * of the namekey synthesize the chain.
1996          *
1997          * The key range is inclusive of both key_beg and key_end.
1998          */
1999         namekey = hammer_directory_namekey(fdip, fncp->nc_name, fncp->nc_nlen,
2000                                            &max_iterations);
2001 retry:
2002         hammer_init_cursor(&trans, &cursor, &fdip->cache[1], fdip);
2003         cursor.key_beg.localization = fdip->obj_localization |
2004                                       hammer_dir_localization(fdip);
2005         cursor.key_beg.obj_id = fdip->obj_id;
2006         cursor.key_beg.key = namekey;
2007         cursor.key_beg.create_tid = 0;
2008         cursor.key_beg.delete_tid = 0;
2009         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
2010         cursor.key_beg.obj_type = 0;
2011
2012         cursor.key_end = cursor.key_beg;
2013         cursor.key_end.key += max_iterations;
2014         cursor.asof = fdip->obj_asof;
2015         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2016
2017         /*
2018          * Scan all matching records (the chain), locate the one matching
2019          * the requested path component.
2020          *
2021          * The hammer_ip_*() functions merge in-memory records with on-disk
2022          * records for the purposes of the search.
2023          */
2024         error = hammer_ip_first(&cursor);
2025         while (error == 0) {
2026                 if (hammer_ip_resolve_data(&cursor) != 0)
2027                         break;
2028                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
2029                 KKASSERT(nlen > 0);
2030                 if (fncp->nc_nlen == nlen &&
2031                     bcmp(fncp->nc_name, cursor.data->entry.name, nlen) == 0) {
2032                         break;
2033                 }
2034                 error = hammer_ip_next(&cursor);
2035         }
2036
2037         /*
2038          * If all is ok we have to get the inode so we can adjust nlinks.
2039          *
2040          * WARNING: hammer_ip_del_directory() may have to terminate the
2041          * cursor to avoid a recursion.  It's ok to call hammer_done_cursor()
2042          * twice.
2043          */
2044         if (error == 0)
2045                 error = hammer_ip_del_directory(&trans, &cursor, fdip, ip);
2046
2047         /*
2048          * XXX A deadlock here will break rename's atomicy for the purposes
2049          * of crash recovery.
2050          */
2051         if (error == EDEADLK) {
2052                 hammer_done_cursor(&cursor);
2053                 goto retry;
2054         }
2055
2056         /*
2057          * Cleanup and tell the kernel that the rename succeeded.
2058          *
2059          * NOTE: ip->vp, if non-NULL, cannot be directly referenced
2060          *       without formally acquiring the vp since the vp might
2061          *       have zero refs on it, or in the middle of a reclaim,
2062          *       etc.
2063          */
2064         hammer_done_cursor(&cursor);
2065         if (error == 0) {
2066                 cache_rename(ap->a_fnch, ap->a_tnch);
2067                 hammer_knote(ap->a_fdvp, NOTE_WRITE);
2068                 hammer_knote(ap->a_tdvp, NOTE_WRITE);
2069                 while (ip->vp) {
2070                         struct vnode *vp;
2071
2072                         error = hammer_get_vnode(ip, &vp);
2073                         if (error == 0 && vp) {
2074                                 vn_unlock(vp);
2075                                 hammer_knote(ip->vp, NOTE_RENAME);
2076                                 vrele(vp);
2077                                 break;
2078                         }
2079                         hdkprintf("ip/vp race2 avoided\n");
2080                 }
2081         }
2082
2083 failed:
2084         hammer_done_transaction(&trans);
2085         lwkt_reltoken(&hmp->fs_token);
2086         return (error);
2087 }
2088
2089 /*
2090  * hammer_vop_nrmdir { nch, dvp, cred }
2091  */
2092 static
2093 int
2094 hammer_vop_nrmdir(struct vop_nrmdir_args *ap)
2095 {
2096         struct hammer_transaction trans;
2097         struct hammer_inode *dip;
2098         hammer_mount_t hmp;
2099         int error;
2100
2101         dip = VTOI(ap->a_dvp);
2102         hmp = dip->hmp;
2103
2104         if (hammer_nohistory(dip) == 0 &&
2105             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2106                 return (error);
2107         }
2108
2109         lwkt_gettoken(&hmp->fs_token);
2110         hammer_start_transaction(&trans, hmp);
2111         ++hammer_stats_file_iopsw;
2112         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 1);
2113         hammer_done_transaction(&trans);
2114         if (error == 0)
2115                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
2116         lwkt_reltoken(&hmp->fs_token);
2117         return (error);
2118 }
2119
2120 /*
2121  * hammer_vop_markatime { vp, cred }
2122  */
2123 static
2124 int
2125 hammer_vop_markatime(struct vop_markatime_args *ap)
2126 {
2127         struct hammer_transaction trans;
2128         struct hammer_inode *ip;
2129         hammer_mount_t hmp;
2130
2131         ip = VTOI(ap->a_vp);
2132         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2133                 return (EROFS);
2134         if (ip->flags & HAMMER_INODE_RO)
2135                 return (EROFS);
2136         hmp = ip->hmp;
2137         if (hmp->mp->mnt_flag & MNT_NOATIME)
2138                 return (0);
2139         lwkt_gettoken(&hmp->fs_token);
2140         hammer_start_transaction(&trans, hmp);
2141         ++hammer_stats_file_iopsw;
2142
2143         ip->ino_data.atime = trans.time;
2144         hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
2145         hammer_done_transaction(&trans);
2146         hammer_knote(ap->a_vp, NOTE_ATTRIB);
2147         lwkt_reltoken(&hmp->fs_token);
2148         return (0);
2149 }
2150
2151 /*
2152  * hammer_vop_setattr { vp, vap, cred }
2153  */
2154 static
2155 int
2156 hammer_vop_setattr(struct vop_setattr_args *ap)
2157 {
2158         struct hammer_transaction trans;
2159         struct hammer_inode *ip;
2160         struct vattr *vap;
2161         hammer_mount_t hmp;
2162         int modflags;
2163         int error;
2164         int truncating;
2165         int blksize;
2166         int kflags;
2167 #if 0
2168         int64_t aligned_size;
2169 #endif
2170         uint32_t flags;
2171
2172         vap = ap->a_vap;
2173         ip = ap->a_vp->v_data;
2174         modflags = 0;
2175         kflags = 0;
2176         hmp = ip->hmp;
2177
2178         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2179                 return(EROFS);
2180         if (ip->flags & HAMMER_INODE_RO)
2181                 return (EROFS);
2182         if (hammer_nohistory(ip) == 0 &&
2183             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2184                 return (error);
2185         }
2186
2187         lwkt_gettoken(&hmp->fs_token);
2188         hammer_start_transaction(&trans, hmp);
2189         ++hammer_stats_file_iopsw;
2190         error = 0;
2191
2192         if (vap->va_flags != VNOVAL) {
2193                 flags = ip->ino_data.uflags;
2194                 error = vop_helper_setattr_flags(&flags, vap->va_flags,
2195                                          hammer_to_unix_xid(&ip->ino_data.uid),
2196                                          ap->a_cred);
2197                 if (error == 0) {
2198                         if (ip->ino_data.uflags != flags) {
2199                                 ip->ino_data.uflags = flags;
2200                                 ip->ino_data.ctime = trans.time;
2201                                 modflags |= HAMMER_INODE_DDIRTY;
2202                                 kflags |= NOTE_ATTRIB;
2203                         }
2204                         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2205                                 error = 0;
2206                                 goto done;
2207                         }
2208                 }
2209                 goto done;
2210         }
2211         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2212                 error = EPERM;
2213                 goto done;
2214         }
2215         if (vap->va_uid != (uid_t)VNOVAL || vap->va_gid != (gid_t)VNOVAL) {
2216                 mode_t cur_mode = ip->ino_data.mode;
2217                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2218                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2219                 uuid_t uuid_uid;
2220                 uuid_t uuid_gid;
2221
2222                 error = vop_helper_chown(ap->a_vp, vap->va_uid, vap->va_gid,
2223                                          ap->a_cred,
2224                                          &cur_uid, &cur_gid, &cur_mode);
2225                 if (error == 0) {
2226                         hammer_guid_to_uuid(&uuid_uid, cur_uid);
2227                         hammer_guid_to_uuid(&uuid_gid, cur_gid);
2228                         if (bcmp(&uuid_uid, &ip->ino_data.uid,
2229                                  sizeof(uuid_uid)) ||
2230                             bcmp(&uuid_gid, &ip->ino_data.gid,
2231                                  sizeof(uuid_gid)) ||
2232                             ip->ino_data.mode != cur_mode) {
2233                                 ip->ino_data.uid = uuid_uid;
2234                                 ip->ino_data.gid = uuid_gid;
2235                                 ip->ino_data.mode = cur_mode;
2236                                 ip->ino_data.ctime = trans.time;
2237                                 modflags |= HAMMER_INODE_DDIRTY;
2238                         }
2239                         kflags |= NOTE_ATTRIB;
2240                 }
2241         }
2242         while (vap->va_size != VNOVAL && ip->ino_data.size != vap->va_size) {
2243                 switch(ap->a_vp->v_type) {
2244                 case VREG:
2245                         if (vap->va_size == ip->ino_data.size)
2246                                 break;
2247
2248                         /*
2249                          * Log the operation if in fast-fsync mode or if
2250                          * there are unterminated redo write records present.
2251                          *
2252                          * The second check is needed so the recovery code
2253                          * properly truncates write redos even if nominal
2254                          * REDO operations is turned off due to excessive
2255                          * writes, because the related records might be
2256                          * destroyed and never lay down a TERM_WRITE.
2257                          */
2258                         if ((ip->flags & HAMMER_INODE_REDO) ||
2259                             (ip->flags & HAMMER_INODE_RDIRTY)) {
2260                                 error = hammer_generate_redo(&trans, ip,
2261                                                              vap->va_size,
2262                                                              HAMMER_REDO_TRUNC,
2263                                                              NULL, 0);
2264                         }
2265                         blksize = hammer_blocksize(vap->va_size);
2266
2267                         /*
2268                          * XXX break atomicy, we can deadlock the backend
2269                          * if we do not release the lock.  Probably not a
2270                          * big deal here.
2271                          */
2272                         if (vap->va_size < ip->ino_data.size) {
2273                                 nvtruncbuf(ap->a_vp, vap->va_size,
2274                                            blksize,
2275                                            hammer_blockoff(vap->va_size),
2276                                            0);
2277                                 truncating = 1;
2278                                 kflags |= NOTE_WRITE;
2279                         } else {
2280                                 nvextendbuf(ap->a_vp,
2281                                             ip->ino_data.size,
2282                                             vap->va_size,
2283                                             hammer_blocksize(ip->ino_data.size),
2284                                             hammer_blocksize(vap->va_size),
2285                                             hammer_blockoff(ip->ino_data.size),
2286                                             hammer_blockoff(vap->va_size),
2287                                             0);
2288                                 truncating = 0;
2289                                 kflags |= NOTE_WRITE | NOTE_EXTEND;
2290                         }
2291                         ip->ino_data.size = vap->va_size;
2292                         ip->ino_data.mtime = trans.time;
2293                         /* XXX safe to use SDIRTY instead of DDIRTY here? */
2294                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2295
2296                         /*
2297                          * On-media truncation is cached in the inode until
2298                          * the inode is synchronized.  We must immediately
2299                          * handle any frontend records.
2300                          */
2301                         if (truncating) {
2302                                 hammer_ip_frontend_trunc(ip, vap->va_size);
2303                                 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2304                                         ip->flags |= HAMMER_INODE_TRUNCATED;
2305                                         ip->trunc_off = vap->va_size;
2306                                         hammer_inode_dirty(ip);
2307                                 } else if (ip->trunc_off > vap->va_size) {
2308                                         ip->trunc_off = vap->va_size;
2309                                 }
2310                         }
2311
2312 #if 0
2313                         /*
2314                          * When truncating, nvtruncbuf() may have cleaned out
2315                          * a portion of the last block on-disk in the buffer
2316                          * cache.  We must clean out any frontend records
2317                          * for blocks beyond the new last block.
2318                          */
2319                         aligned_size = (vap->va_size + (blksize - 1)) &
2320                                        ~(int64_t)(blksize - 1);
2321                         if (truncating && vap->va_size < aligned_size) {
2322                                 aligned_size -= blksize;
2323                                 hammer_ip_frontend_trunc(ip, aligned_size);
2324                         }
2325 #endif
2326                         break;
2327                 case VDATABASE:
2328                         if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2329                                 ip->flags |= HAMMER_INODE_TRUNCATED;
2330                                 ip->trunc_off = vap->va_size;
2331                                 hammer_inode_dirty(ip);
2332                         } else if (ip->trunc_off > vap->va_size) {
2333                                 ip->trunc_off = vap->va_size;
2334                         }
2335                         hammer_ip_frontend_trunc(ip, vap->va_size);
2336                         ip->ino_data.size = vap->va_size;
2337                         ip->ino_data.mtime = trans.time;
2338                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2339                         kflags |= NOTE_ATTRIB;
2340                         break;
2341                 default:
2342                         error = EINVAL;
2343                         goto done;
2344                 }
2345                 break;
2346         }
2347         if (vap->va_atime.tv_sec != VNOVAL) {
2348                 ip->ino_data.atime = hammer_timespec_to_time(&vap->va_atime);
2349                 modflags |= HAMMER_INODE_ATIME;
2350                 kflags |= NOTE_ATTRIB;
2351         }
2352         if (vap->va_mtime.tv_sec != VNOVAL) {
2353                 ip->ino_data.mtime = hammer_timespec_to_time(&vap->va_mtime);
2354                 modflags |= HAMMER_INODE_MTIME;
2355                 kflags |= NOTE_ATTRIB;
2356         }
2357         if (vap->va_mode != (mode_t)VNOVAL) {
2358                 mode_t   cur_mode = ip->ino_data.mode;
2359                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2360                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2361
2362                 error = vop_helper_chmod(ap->a_vp, vap->va_mode, ap->a_cred,
2363                                          cur_uid, cur_gid, &cur_mode);
2364                 if (error == 0 && ip->ino_data.mode != cur_mode) {
2365                         ip->ino_data.mode = cur_mode;
2366                         ip->ino_data.ctime = trans.time;
2367                         modflags |= HAMMER_INODE_DDIRTY;
2368                         kflags |= NOTE_ATTRIB;
2369                 }
2370         }
2371 done:
2372         if (error == 0)
2373                 hammer_modify_inode(&trans, ip, modflags);
2374         hammer_done_transaction(&trans);
2375         hammer_knote(ap->a_vp, kflags);
2376         lwkt_reltoken(&hmp->fs_token);
2377         return (error);
2378 }
2379
2380 /*
2381  * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
2382  */
2383 static
2384 int
2385 hammer_vop_nsymlink(struct vop_nsymlink_args *ap)
2386 {
2387         struct hammer_transaction trans;
2388         struct hammer_inode *dip;
2389         struct hammer_inode *nip;
2390         hammer_record_t record;
2391         struct nchandle *nch;
2392         hammer_mount_t hmp;
2393         int error;
2394         int bytes;
2395
2396         ap->a_vap->va_type = VLNK;
2397
2398         nch = ap->a_nch;
2399         dip = VTOI(ap->a_dvp);
2400         hmp = dip->hmp;
2401
2402         if (dip->flags & HAMMER_INODE_RO)
2403                 return (EROFS);
2404         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
2405                 return (error);
2406
2407         /*
2408          * Create a transaction to cover the operations we perform.
2409          */
2410         lwkt_gettoken(&hmp->fs_token);
2411         hammer_start_transaction(&trans, hmp);
2412         ++hammer_stats_file_iopsw;
2413
2414         /*
2415          * Create a new filesystem object of the requested type.  The
2416          * returned inode will be referenced but not locked.
2417          */
2418
2419         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
2420                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
2421                                     NULL, &nip);
2422         if (error) {
2423                 hammer_done_transaction(&trans);
2424                 *ap->a_vpp = NULL;
2425                 lwkt_reltoken(&hmp->fs_token);
2426                 return (error);
2427         }
2428
2429         /*
2430          * Add a record representing the symlink.  symlink stores the link
2431          * as pure data, not a string, and is no \0 terminated.
2432          */
2433         if (error == 0) {
2434                 bytes = strlen(ap->a_target);
2435
2436                 if (bytes <= HAMMER_INODE_BASESYMLEN) {
2437                         bcopy(ap->a_target, nip->ino_data.ext.symlink, bytes);
2438                 } else {
2439                         record = hammer_alloc_mem_record(nip, bytes);
2440                         record->type = HAMMER_MEM_RECORD_GENERAL;
2441
2442                         record->leaf.base.localization = nip->obj_localization |
2443                                                          HAMMER_LOCALIZE_MISC;
2444                         record->leaf.base.key = HAMMER_FIXKEY_SYMLINK;
2445                         record->leaf.base.rec_type = HAMMER_RECTYPE_FIX;
2446                         record->leaf.data_len = bytes;
2447                         KKASSERT(HAMMER_SYMLINK_NAME_OFF == 0);
2448                         bcopy(ap->a_target, record->data->symlink.name, bytes);
2449                         error = hammer_ip_add_record(&trans, record);
2450                 }
2451
2452                 /*
2453                  * Set the file size to the length of the link.
2454                  */
2455                 if (error == 0) {
2456                         nip->ino_data.size = bytes;
2457                         hammer_modify_inode(&trans, nip, HAMMER_INODE_DDIRTY);
2458                 }
2459         }
2460         if (error == 0)
2461                 error = hammer_ip_add_directory(&trans, dip, nch->ncp->nc_name,
2462                                                 nch->ncp->nc_nlen, nip);
2463
2464         /*
2465          * Finish up.
2466          */
2467         if (error) {
2468                 hammer_rel_inode(nip, 0);
2469                 *ap->a_vpp = NULL;
2470         } else {
2471                 error = hammer_get_vnode(nip, ap->a_vpp);
2472                 hammer_rel_inode(nip, 0);
2473                 if (error == 0) {
2474                         cache_setunresolved(ap->a_nch);
2475                         cache_setvp(ap->a_nch, *ap->a_vpp);
2476                         hammer_knote(ap->a_dvp, NOTE_WRITE);
2477                 }
2478         }
2479         hammer_done_transaction(&trans);
2480         lwkt_reltoken(&hmp->fs_token);
2481         return (error);
2482 }
2483
2484 /*
2485  * hammer_vop_nwhiteout { nch, dvp, cred, flags }
2486  */
2487 static
2488 int
2489 hammer_vop_nwhiteout(struct vop_nwhiteout_args *ap)
2490 {
2491         struct hammer_transaction trans;
2492         struct hammer_inode *dip;
2493         hammer_mount_t hmp;
2494         int error;
2495
2496         dip = VTOI(ap->a_dvp);
2497         hmp = dip->hmp;
2498
2499         if (hammer_nohistory(dip) == 0 &&
2500             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0) {
2501                 return (error);
2502         }
2503
2504         lwkt_gettoken(&hmp->fs_token);
2505         hammer_start_transaction(&trans, hmp);
2506         ++hammer_stats_file_iopsw;
2507         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp,
2508                                 ap->a_cred, ap->a_flags, -1);
2509         hammer_done_transaction(&trans);
2510         lwkt_reltoken(&hmp->fs_token);
2511
2512         return (error);
2513 }
2514
2515 /*
2516  * hammer_vop_ioctl { vp, command, data, fflag, cred }
2517  */
2518 static
2519 int
2520 hammer_vop_ioctl(struct vop_ioctl_args *ap)
2521 {
2522         struct hammer_inode *ip = ap->a_vp->v_data;
2523         hammer_mount_t hmp = ip->hmp;
2524         int error;
2525
2526         ++hammer_stats_file_iopsr;
2527         lwkt_gettoken(&hmp->fs_token);
2528         error = hammer_ioctl(ip, ap->a_command, ap->a_data,
2529                              ap->a_fflag, ap->a_cred);
2530         lwkt_reltoken(&hmp->fs_token);
2531         return (error);
2532 }
2533
2534 static
2535 int
2536 hammer_vop_mountctl(struct vop_mountctl_args *ap)
2537 {
2538         static const struct mountctl_opt extraopt[] = {
2539                 { HMNT_NOHISTORY,       "nohistory" },
2540                 { HMNT_MASTERID,        "master" },
2541                 { HMNT_NOMIRROR,        "nomirror" },
2542                 { 0, NULL}
2543
2544         };
2545         struct hammer_mount *hmp;
2546         struct mount *mp;
2547         int usedbytes;
2548         int error;
2549
2550         error = 0;
2551         usedbytes = 0;
2552         mp = ap->a_head.a_ops->head.vv_mount;
2553         KKASSERT(mp->mnt_data != NULL);
2554         hmp = (struct hammer_mount *)mp->mnt_data;
2555
2556         lwkt_gettoken(&hmp->fs_token);
2557
2558         switch(ap->a_op) {
2559         case MOUNTCTL_SET_EXPORT:
2560                 if (ap->a_ctllen != sizeof(struct export_args))
2561                         error = EINVAL;
2562                 else
2563                         error = hammer_vfs_export(mp, ap->a_op,
2564                                       (const struct export_args *)ap->a_ctl);
2565                 break;
2566         case MOUNTCTL_MOUNTFLAGS:
2567                 /*
2568                  * Call standard mountctl VOP function
2569                  * so we get user mount flags.
2570                  */
2571                 error = vop_stdmountctl(ap);
2572                 if (error)
2573                         break;
2574
2575                 usedbytes = *ap->a_res;
2576
2577                 if (usedbytes > 0 && usedbytes < ap->a_buflen) {
2578                         usedbytes += vfs_flagstostr(hmp->hflags, extraopt,
2579                                                     ap->a_buf,
2580                                                     ap->a_buflen - usedbytes,
2581                                                     &error);
2582                 }
2583
2584                 *ap->a_res += usedbytes;
2585                 break;
2586         default:
2587                 error = vop_stdmountctl(ap);
2588                 break;
2589         }
2590         lwkt_reltoken(&hmp->fs_token);
2591         return(error);
2592 }
2593
2594 /*
2595  * hammer_vop_strategy { vp, bio }
2596  *
2597  * Strategy call, used for regular file read & write only.  Note that the
2598  * bp may represent a cluster.
2599  *
2600  * To simplify operation and allow better optimizations in the future,
2601  * this code does not make any assumptions with regards to buffer alignment
2602  * or size.
2603  */
2604 static
2605 int
2606 hammer_vop_strategy(struct vop_strategy_args *ap)
2607 {
2608         struct buf *bp;
2609         int error;
2610
2611         bp = ap->a_bio->bio_buf;
2612
2613         switch(bp->b_cmd) {
2614         case BUF_CMD_READ:
2615                 error = hammer_vop_strategy_read(ap);
2616                 break;
2617         case BUF_CMD_WRITE:
2618                 error = hammer_vop_strategy_write(ap);
2619                 break;
2620         default:
2621                 bp->b_error = error = EINVAL;
2622                 bp->b_flags |= B_ERROR;
2623                 biodone(ap->a_bio);
2624                 break;
2625         }
2626
2627         /* hammer_dump_dedup_cache(((hammer_inode_t)ap->a_vp->v_data)->hmp); */
2628
2629         return (error);
2630 }
2631
2632 /*
2633  * Read from a regular file.  Iterate the related records and fill in the
2634  * BIO/BUF.  Gaps are zero-filled.
2635  *
2636  * The support code in hammer_object.c should be used to deal with mixed
2637  * in-memory and on-disk records.
2638  *
2639  * NOTE: Can be called from the cluster code with an oversized buf.
2640  *
2641  * XXX atime update
2642  */
2643 static
2644 int
2645 hammer_vop_strategy_read(struct vop_strategy_args *ap)
2646 {
2647         struct hammer_transaction trans;
2648         struct hammer_inode *ip;
2649         struct hammer_inode *dip;
2650         hammer_mount_t hmp;
2651         struct hammer_cursor cursor;
2652         hammer_base_elm_t base;
2653         hammer_off_t disk_offset;
2654         struct bio *bio;
2655         struct bio *nbio;
2656         struct buf *bp;
2657         int64_t rec_offset;
2658         int64_t ran_end;
2659         int64_t tmp64;
2660         int error;
2661         int boff;
2662         int roff;
2663         int n;
2664         int isdedupable;
2665
2666         bio = ap->a_bio;
2667         bp = bio->bio_buf;
2668         ip = ap->a_vp->v_data;
2669         hmp = ip->hmp;
2670
2671         /*
2672          * The zone-2 disk offset may have been set by the cluster code via
2673          * a BMAP operation, or else should be NOOFFSET.
2674          *
2675          * Checking the high bits for a match against zone-2 should suffice.
2676          *
2677          * In cases where a lot of data duplication is present it may be
2678          * more beneficial to drop through and doubule-buffer through the
2679          * device.
2680          */
2681         nbio = push_bio(bio);
2682         if ((nbio->bio_offset & HAMMER_OFF_ZONE_MASK) ==
2683             HAMMER_ZONE_LARGE_DATA) {
2684                 if (hammer_double_buffer == 0) {
2685                         lwkt_gettoken(&hmp->fs_token);
2686                         error = hammer_io_direct_read(hmp, nbio, NULL);
2687                         lwkt_reltoken(&hmp->fs_token);
2688                         return (error);
2689                 }
2690
2691                 /*
2692                  * Try to shortcut requests for double_buffer mode too.
2693                  * Since this mode runs through the device buffer cache
2694                  * only compatible buffer sizes (meaning those generated
2695                  * by normal filesystem buffers) are legal.
2696                  */
2697                 if (hammer_live_dedup == 0 && (bp->b_flags & B_PAGING) == 0) {
2698                         lwkt_gettoken(&hmp->fs_token);
2699                         error = hammer_io_indirect_read(hmp, nbio, NULL);
2700                         lwkt_reltoken(&hmp->fs_token);
2701                         return (error);
2702                 }
2703         }
2704
2705         /*
2706          * Well, that sucked.  Do it the hard way.  If all the stars are
2707          * aligned we may still be able to issue a direct-read.
2708          */
2709         lwkt_gettoken(&hmp->fs_token);
2710         hammer_simple_transaction(&trans, hmp);
2711         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2712
2713         /*
2714          * Key range (begin and end inclusive) to scan.  Note that the key's
2715          * stored in the actual records represent BASE+LEN, not BASE.  The
2716          * first record containing bio_offset will have a key > bio_offset.
2717          */
2718         cursor.key_beg.localization = ip->obj_localization |
2719                                       HAMMER_LOCALIZE_MISC;
2720         cursor.key_beg.obj_id = ip->obj_id;
2721         cursor.key_beg.create_tid = 0;
2722         cursor.key_beg.delete_tid = 0;
2723         cursor.key_beg.obj_type = 0;
2724         cursor.key_beg.key = bio->bio_offset + 1;
2725         cursor.asof = ip->obj_asof;
2726         cursor.flags |= HAMMER_CURSOR_ASOF;
2727
2728         cursor.key_end = cursor.key_beg;
2729         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2730 #if 0
2731         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
2732                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
2733                 cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
2734                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2735         } else
2736 #endif
2737         {
2738                 ran_end = bio->bio_offset + bp->b_bufsize;
2739                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2740                 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2741                 tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2742                 if (tmp64 < ran_end)
2743                         cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2744                 else
2745                         cursor.key_end.key = ran_end + MAXPHYS + 1;
2746         }
2747         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2748
2749         /*
2750          * Set NOSWAPCACHE for cursor data extraction if double buffering
2751          * is disabled or (if the file is not marked cacheable via chflags
2752          * and vm.swapcache_use_chflags is enabled).
2753          */
2754         if (hammer_double_buffer == 0 ||
2755             ((ap->a_vp->v_flag & VSWAPCACHE) == 0 &&
2756              vm_swapcache_use_chflags)) {
2757                 cursor.flags |= HAMMER_CURSOR_NOSWAPCACHE;
2758         }
2759
2760         error = hammer_ip_first(&cursor);
2761         boff = 0;
2762
2763         while (error == 0) {
2764                 /*
2765                  * Get the base file offset of the record.  The key for
2766                  * data records is (base + bytes) rather then (base).
2767                  */
2768                 base = &cursor.leaf->base;
2769                 rec_offset = base->key - cursor.leaf->data_len;
2770
2771                 /*
2772                  * Calculate the gap, if any, and zero-fill it.
2773                  *
2774                  * n is the offset of the start of the record verses our
2775                  * current seek offset in the bio.
2776                  */
2777                 n = (int)(rec_offset - (bio->bio_offset + boff));
2778                 if (n > 0) {
2779                         if (n > bp->b_bufsize - boff)
2780                                 n = bp->b_bufsize - boff;
2781                         bzero((char *)bp->b_data + boff, n);
2782                         boff += n;
2783                         n = 0;
2784                 }
2785
2786                 /*
2787                  * Calculate the data offset in the record and the number
2788                  * of bytes we can copy.
2789                  *
2790                  * There are two degenerate cases.  First, boff may already
2791                  * be at bp->b_bufsize.  Secondly, the data offset within
2792                  * the record may exceed the record's size.
2793                  */
2794                 roff = -n;
2795                 rec_offset += roff;
2796                 n = cursor.leaf->data_len - roff;
2797                 if (n <= 0) {
2798                         hdkprintf("bad n=%d roff=%d\n", n, roff);
2799                         n = 0;
2800                 } else if (n > bp->b_bufsize - boff) {
2801                         n = bp->b_bufsize - boff;
2802                 }
2803
2804                 /*
2805                  * Deal with cached truncations.  This cool bit of code
2806                  * allows truncate()/ftruncate() to avoid having to sync
2807                  * the file.
2808                  *
2809                  * If the frontend is truncated then all backend records are
2810                  * subject to the frontend's truncation.
2811                  *
2812                  * If the backend is truncated then backend records on-disk
2813                  * (but not in-memory) are subject to the backend's
2814                  * truncation.  In-memory records owned by the backend
2815                  * represent data written after the truncation point on the
2816                  * backend and must not be truncated.
2817                  *
2818                  * Truncate operations deal with frontend buffer cache
2819                  * buffers and frontend-owned in-memory records synchronously.
2820                  */
2821                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2822                         if (hammer_cursor_ondisk(&cursor)/* ||
2823                             cursor.iprec->flush_state == HAMMER_FST_FLUSH*/) {
2824                                 if (ip->trunc_off <= rec_offset)
2825                                         n = 0;
2826                                 else if (ip->trunc_off < rec_offset + n)
2827                                         n = (int)(ip->trunc_off - rec_offset);
2828                         }
2829                 }
2830                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2831                         if (hammer_cursor_ondisk(&cursor)) {
2832                                 if (ip->sync_trunc_off <= rec_offset)
2833                                         n = 0;
2834                                 else if (ip->sync_trunc_off < rec_offset + n)
2835                                         n = (int)(ip->sync_trunc_off - rec_offset);
2836                         }
2837                 }
2838
2839                 /*
2840                  * Try to issue a direct read into our bio if possible,
2841                  * otherwise resolve the element data into a hammer_buffer
2842                  * and copy.
2843                  *
2844                  * The buffer on-disk should be zerod past any real
2845                  * truncation point, but may not be for any synthesized
2846                  * truncation point from above.
2847                  *
2848                  * NOTE: disk_offset is only valid if the cursor data is
2849                  *       on-disk.
2850                  */
2851                 disk_offset = cursor.leaf->data_offset + roff;
2852                 isdedupable = (boff == 0 && n == bp->b_bufsize &&
2853                                hammer_cursor_ondisk(&cursor) &&
2854                                ((int)disk_offset & HAMMER_BUFMASK) == 0);
2855
2856                 if (isdedupable && hammer_double_buffer == 0) {
2857                         /*
2858                          * Direct read case
2859                          */
2860                         KKASSERT((disk_offset & HAMMER_OFF_ZONE_MASK) ==
2861                                  HAMMER_ZONE_LARGE_DATA);
2862                         nbio->bio_offset = disk_offset;
2863                         error = hammer_io_direct_read(hmp, nbio, cursor.leaf);
2864                         if (hammer_live_dedup && error == 0)
2865                                 hammer_dedup_cache_add(ip, cursor.leaf);
2866                         goto done;
2867                 } else if (isdedupable) {
2868                         /*
2869                          * Async I/O case for reading from backing store
2870                          * and copying the data to the filesystem buffer.
2871                          * live-dedup has to verify the data anyway if it
2872                          * gets a hit later so we can just add the entry
2873                          * now.
2874                          */
2875                         KKASSERT((disk_offset & HAMMER_OFF_ZONE_MASK) ==
2876                                  HAMMER_ZONE_LARGE_DATA);
2877                         nbio->bio_offset = disk_offset;
2878                         if (hammer_live_dedup)
2879                                 hammer_dedup_cache_add(ip, cursor.leaf);
2880                         error = hammer_io_indirect_read(hmp, nbio, cursor.leaf);
2881                         goto done;
2882                 } else if (n) {
2883                         error = hammer_ip_resolve_data(&cursor);
2884                         if (error == 0) {
2885                                 if (hammer_live_dedup && isdedupable)
2886                                         hammer_dedup_cache_add(ip, cursor.leaf);
2887                                 bcopy((char *)cursor.data + roff,
2888                                       (char *)bp->b_data + boff, n);
2889                         }
2890                 }
2891                 if (error)
2892                         break;
2893
2894                 /*
2895                  * We have to be sure that the only elements added to the
2896                  * dedup cache are those which are already on-media.
2897                  */
2898                 if (hammer_live_dedup && hammer_cursor_ondisk(&cursor))
2899                         hammer_dedup_cache_add(ip, cursor.leaf);
2900
2901                 /*
2902                  * Iterate until we have filled the request.
2903                  */
2904                 boff += n;
2905                 if (boff == bp->b_bufsize)
2906                         break;
2907                 error = hammer_ip_next(&cursor);
2908         }
2909
2910         /*
2911          * There may have been a gap after the last record
2912          */
2913         if (error == ENOENT)
2914                 error = 0;
2915         if (error == 0 && boff != bp->b_bufsize) {
2916                 KKASSERT(boff < bp->b_bufsize);
2917                 bzero((char *)bp->b_data + boff, bp->b_bufsize - boff);
2918                 /* boff = bp->b_bufsize; */
2919         }
2920
2921         /*
2922          * Disallow swapcache operation on the vnode buffer if double
2923          * buffering is enabled, the swapcache will get the data via
2924          * the block device buffer.
2925          */
2926         if (hammer_double_buffer)
2927                 bp->b_flags |= B_NOTMETA;
2928
2929         /*
2930          * Cleanup
2931          */
2932         bp->b_resid = 0;
2933         bp->b_error = error;
2934         if (error)
2935                 bp->b_flags |= B_ERROR;
2936         biodone(ap->a_bio);
2937
2938 done:
2939         /*
2940          * Cache the b-tree node for the last data read in cache[1].
2941          *
2942          * If we hit the file EOF then also cache the node in the
2943          * governing directory's cache[3], it will be used to initialize
2944          * the new inode's cache[1] for any inodes looked up via the directory.
2945          *
2946          * This doesn't reduce disk accesses since the B-Tree chain is
2947          * likely cached, but it does reduce cpu overhead when looking
2948          * up file offsets for cpdup/tar/cpio style iterations.
2949          */
2950         if (cursor.node)
2951                 hammer_cache_node(&ip->cache[1], cursor.node);
2952         if (ran_end >= ip->ino_data.size) {
2953                 dip = hammer_find_inode(&trans, ip->ino_data.parent_obj_id,
2954                                         ip->obj_asof, ip->obj_localization);
2955                 if (dip) {
2956                         hammer_cache_node(&dip->cache[3], cursor.node);
2957                         hammer_rel_inode(dip, 0);
2958                 }
2959         }
2960         hammer_done_cursor(&cursor);
2961         hammer_done_transaction(&trans);
2962         lwkt_reltoken(&hmp->fs_token);
2963         return(error);
2964 }
2965
2966 /*
2967  * BMAP operation - used to support cluster_read() only.
2968  *
2969  * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
2970  *
2971  * This routine may return EOPNOTSUPP if the opration is not supported for
2972  * the specified offset.  The contents of the pointer arguments do not
2973  * need to be initialized in that case.
2974  *
2975  * If a disk address is available and properly aligned return 0 with
2976  * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
2977  * to the run-length relative to that offset.  Callers may assume that
2978  * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
2979  * large, so return EOPNOTSUPP if it is not sufficiently large.
2980  */
2981 static
2982 int
2983 hammer_vop_bmap(struct vop_bmap_args *ap)
2984 {
2985         struct hammer_transaction trans;
2986         struct hammer_inode *ip;
2987         hammer_mount_t hmp;
2988         struct hammer_cursor cursor;
2989         hammer_base_elm_t base;
2990         int64_t rec_offset;
2991         int64_t ran_end;
2992         int64_t tmp64;
2993         int64_t base_offset;
2994         int64_t base_disk_offset;
2995         int64_t last_offset;
2996         hammer_off_t last_disk_offset;
2997         hammer_off_t disk_offset;
2998         int     rec_len;
2999         int     error;
3000         int     blksize;
3001
3002         ++hammer_stats_file_iopsr;
3003         ip = ap->a_vp->v_data;
3004         hmp = ip->hmp;
3005
3006         /*
3007          * We can only BMAP regular files.  We can't BMAP database files,
3008          * directories, etc.
3009          */
3010         if (ip->ino_data.obj_type != HAMMER_OBJTYPE_REGFILE)
3011                 return(EOPNOTSUPP);
3012
3013         /*
3014          * bmap is typically called with runp/runb both NULL when used
3015          * for writing.  We do not support BMAP for writing atm.
3016          */
3017         if (ap->a_cmd != BUF_CMD_READ)
3018                 return(EOPNOTSUPP);
3019
3020         /*
3021          * Scan the B-Tree to acquire blockmap addresses, then translate
3022          * to raw addresses.
3023          */
3024         lwkt_gettoken(&hmp->fs_token);
3025         hammer_simple_transaction(&trans, hmp);
3026
3027         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
3028
3029         /*
3030          * Key range (begin and end inclusive) to scan.  Note that the key's
3031          * stored in the actual records represent BASE+LEN, not BASE.  The
3032          * first record containing bio_offset will have a key > bio_offset.
3033          */
3034         cursor.key_beg.localization = ip->obj_localization |
3035                                       HAMMER_LOCALIZE_MISC;
3036         cursor.key_beg.obj_id = ip->obj_id;
3037         cursor.key_beg.create_tid = 0;
3038         cursor.key_beg.delete_tid = 0;
3039         cursor.key_beg.obj_type = 0;
3040         if (ap->a_runb)
3041                 cursor.key_beg.key = ap->a_loffset - MAXPHYS + 1;
3042         else
3043                 cursor.key_beg.key = ap->a_loffset + 1;
3044         if (cursor.key_beg.key < 0)
3045                 cursor.key_beg.key = 0;
3046         cursor.asof = ip->obj_asof;
3047         cursor.flags |= HAMMER_CURSOR_ASOF;
3048
3049         cursor.key_end = cursor.key_beg;
3050         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
3051
3052         ran_end = ap->a_loffset + MAXPHYS;
3053         cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
3054         cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
3055         tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
3056         if (tmp64 < ran_end)
3057                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
3058         else
3059                 cursor.key_end.key = ran_end + MAXPHYS + 1;
3060
3061         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
3062
3063         error = hammer_ip_first(&cursor);
3064         base_offset = last_offset = 0;
3065         base_disk_offset = last_disk_offset = 0;
3066
3067         while (error == 0) {
3068                 /*
3069                  * Get the base file offset of the record.  The key for
3070                  * data records is (base + bytes) rather then (base).
3071                  *
3072                  * NOTE: rec_offset + rec_len may exceed the end-of-file.
3073                  * The extra bytes should be zero on-disk and the BMAP op
3074                  * should still be ok.
3075                  */
3076                 base = &cursor.leaf->base;
3077                 rec_offset = base->key - cursor.leaf->data_len;
3078                 rec_len    = cursor.leaf->data_len;
3079
3080                 /*
3081                  * Incorporate any cached truncation.
3082                  *
3083                  * NOTE: Modifications to rec_len based on synthesized
3084                  * truncation points remove the guarantee that any extended
3085                  * data on disk is zero (since the truncations may not have
3086                  * taken place on-media yet).
3087                  */
3088                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
3089                         if (hammer_cursor_ondisk(&cursor) ||
3090                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
3091                                 if (ip->trunc_off <= rec_offset)
3092                                         rec_len = 0;
3093                                 else if (ip->trunc_off < rec_offset + rec_len)
3094                                         rec_len = (int)(ip->trunc_off - rec_offset);
3095                         }
3096                 }
3097                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
3098                         if (hammer_cursor_ondisk(&cursor)) {
3099                                 if (ip->sync_trunc_off <= rec_offset)
3100                                         rec_len = 0;
3101                                 else if (ip->sync_trunc_off < rec_offset + rec_len)
3102                                         rec_len = (int)(ip->sync_trunc_off - rec_offset);
3103                         }
3104                 }
3105
3106                 /*
3107                  * Accumulate information.  If we have hit a discontiguous
3108                  * block reset base_offset unless we are already beyond the
3109                  * requested offset.  If we are, that's it, we stop.
3110                  */
3111                 if (error)
3112                         break;
3113                 if (hammer_cursor_ondisk(&cursor)) {
3114                         disk_offset = cursor.leaf->data_offset;
3115                         if (rec_offset != last_offset ||
3116                             disk_offset != last_disk_offset) {
3117                                 if (rec_offset > ap->a_loffset)
3118                                         break;
3119                                 base_offset = rec_offset;
3120                                 base_disk_offset = disk_offset;
3121                         }
3122                         last_offset = rec_offset + rec_len;
3123                         last_disk_offset = disk_offset + rec_len;
3124
3125                         if (hammer_live_dedup)
3126                                 hammer_dedup_cache_add(ip, cursor.leaf);
3127                 }
3128
3129                 error = hammer_ip_next(&cursor);
3130         }
3131
3132         if (cursor.node)
3133                 hammer_cache_node(&ip->cache[1], cursor.node);
3134
3135         hammer_done_cursor(&cursor);
3136         hammer_done_transaction(&trans);
3137         lwkt_reltoken(&hmp->fs_token);
3138
3139         /*
3140          * If we couldn't find any records or the records we did find were
3141          * all behind the requested offset, return failure.  A forward
3142          * truncation can leave a hole w/ no on-disk records.
3143          */
3144         if (last_offset == 0 || last_offset < ap->a_loffset)
3145                 return (EOPNOTSUPP);
3146
3147         /*
3148          * Figure out the block size at the requested offset and adjust
3149          * our limits so the cluster_read() does not create inappropriately
3150          * sized buffer cache buffers.
3151          */
3152         blksize = hammer_blocksize(ap->a_loffset);
3153         if (hammer_blocksize(base_offset) != blksize) {
3154                 base_offset = hammer_blockdemarc(base_offset, ap->a_loffset);
3155         }
3156         if (last_offset != ap->a_loffset &&
3157             hammer_blocksize(last_offset - 1) != blksize) {
3158                 last_offset = hammer_blockdemarc(ap->a_loffset,
3159                                                  last_offset - 1);
3160         }
3161
3162         /*
3163          * Returning EOPNOTSUPP simply prevents the direct-IO optimization
3164          * from occuring.
3165          */
3166         disk_offset = base_disk_offset + (ap->a_loffset - base_offset);
3167
3168         if ((disk_offset & HAMMER_OFF_ZONE_MASK) != HAMMER_ZONE_LARGE_DATA) {
3169                 /*
3170                  * Only large-data zones can be direct-IOd
3171                  */
3172                 error = EOPNOTSUPP;
3173         } else if ((disk_offset & HAMMER_BUFMASK) ||
3174                    (last_offset - ap->a_loffset) < blksize) {
3175                 /*
3176                  * doffsetp is not aligned or the forward run size does
3177                  * not cover a whole buffer, disallow the direct I/O.
3178                  */
3179                 error = EOPNOTSUPP;
3180         } else {
3181                 /*
3182                  * We're good.
3183                  */
3184                 *ap->a_doffsetp = disk_offset;
3185                 if (ap->a_runb) {
3186                         *ap->a_runb = ap->a_loffset - base_offset;
3187                         KKASSERT(*ap->a_runb >= 0);
3188                 }
3189                 if (ap->a_runp) {
3190                         *ap->a_runp = last_offset - ap->a_loffset;
3191                         KKASSERT(*ap->a_runp >= 0);
3192                 }
3193                 error = 0;
3194         }
3195         return(error);
3196 }
3197
3198 /*
3199  * Write to a regular file.   Because this is a strategy call the OS is
3200  * trying to actually get data onto the media.
3201  */
3202 static
3203 int
3204 hammer_vop_strategy_write(struct vop_strategy_args *ap)
3205 {
3206         hammer_record_t record;
3207         hammer_mount_t hmp;
3208         hammer_inode_t ip;
3209         struct bio *bio;
3210         struct buf *bp;
3211         int blksize __debugvar;
3212         int bytes;
3213         int error;
3214
3215         bio = ap->a_bio;
3216         bp = bio->bio_buf;
3217         ip = ap->a_vp->v_data;
3218         hmp = ip->hmp;
3219
3220         blksize = hammer_blocksize(bio->bio_offset);
3221         KKASSERT(bp->b_bufsize == blksize);
3222
3223         if (ip->flags & HAMMER_INODE_RO) {
3224                 bp->b_error = EROFS;
3225                 bp->b_flags |= B_ERROR;
3226                 biodone(ap->a_bio);
3227                 return(EROFS);
3228         }
3229
3230         lwkt_gettoken(&hmp->fs_token);
3231
3232         /*
3233          * Disallow swapcache operation on the vnode buffer if double
3234          * buffering is enabled, the swapcache will get the data via
3235          * the block device buffer.
3236          */
3237         if (hammer_double_buffer)
3238                 bp->b_flags |= B_NOTMETA;
3239
3240         /*
3241          * Interlock with inode destruction (no in-kernel or directory
3242          * topology visibility).  If we queue new IO while trying to
3243          * destroy the inode we can deadlock the vtrunc call in
3244          * hammer_inode_unloadable_check().
3245          *
3246          * Besides, there's no point flushing a bp associated with an
3247          * inode that is being destroyed on-media and has no kernel
3248          * references.
3249          */
3250         if ((ip->flags | ip->sync_flags) &
3251             (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) {
3252                 bp->b_resid = 0;
3253                 biodone(ap->a_bio);
3254                 lwkt_reltoken(&hmp->fs_token);
3255                 return(0);
3256         }
3257
3258         /*
3259          * Reserve space and issue a direct-write from the front-end.
3260          * NOTE: The direct_io code will hammer_bread/bcopy smaller
3261          * allocations.
3262          *
3263          * An in-memory record will be installed to reference the storage
3264          * until the flusher can get to it.
3265          *
3266          * Since we own the high level bio the front-end will not try to
3267          * do a direct-read until the write completes.
3268          *
3269          * NOTE: The only time we do not reserve a full-sized buffers
3270          * worth of data is if the file is small.  We do not try to
3271          * allocate a fragment (from the small-data zone) at the end of
3272          * an otherwise large file as this can lead to wildly separated
3273          * data.
3274          */
3275         KKASSERT((bio->bio_offset & HAMMER_BUFMASK) == 0);
3276         KKASSERT(bio->bio_offset < ip->ino_data.size);
3277         if (bio->bio_offset || ip->ino_data.size > HAMMER_HBUFSIZE)
3278                 bytes = bp->b_bufsize;
3279         else
3280                 bytes = ((int)ip->ino_data.size + 15) & ~15;
3281
3282         record = hammer_ip_add_bulk(ip, bio->bio_offset, bp->b_data,
3283                                     bytes, &error);
3284
3285         /*
3286          * B_VFSFLAG1 indicates that a REDO_WRITE entry was generated
3287          * in hammer_vop_write().  We must flag the record so the proper
3288          * REDO_TERM_WRITE entry is generated during the flush.
3289          */
3290         if (record) {
3291                 if (bp->b_flags & B_VFSFLAG1) {
3292                         record->flags |= HAMMER_RECF_REDO;
3293                         bp->b_flags &= ~B_VFSFLAG1;
3294                 }
3295                 if (record->flags & HAMMER_RECF_DEDUPED) {
3296                         bp->b_resid = 0;
3297                         hammer_ip_replace_bulk(hmp, record);
3298                         biodone(ap->a_bio);
3299                 } else {
3300                         hammer_io_direct_write(hmp, bio, record);
3301                 }
3302                 if (ip->rsv_recs > 1 && hmp->rsv_recs > hammer_limit_recs)
3303                         hammer_flush_inode(ip, 0);
3304         } else {
3305                 bp->b_bio2.bio_offset = NOOFFSET;
3306                 bp->b_error = error;
3307                 bp->b_flags |= B_ERROR;
3308                 biodone(ap->a_bio);
3309         }
3310         lwkt_reltoken(&hmp->fs_token);
3311         return(error);
3312 }
3313
3314 /*
3315  * dounlink - disconnect a directory entry
3316  *
3317  * XXX whiteout support not really in yet
3318  */
3319 static int
3320 hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
3321                 struct vnode *dvp, struct ucred *cred,
3322                 int flags, int isdir)
3323 {
3324         struct namecache *ncp;
3325         hammer_inode_t dip;
3326         hammer_inode_t ip;
3327         hammer_mount_t hmp;
3328         struct hammer_cursor cursor;
3329         int64_t namekey;
3330         uint32_t max_iterations;
3331         int nlen, error;
3332
3333         /*
3334          * Calculate the namekey and setup the key range for the scan.  This
3335          * works kinda like a chained hash table where the lower 32 bits
3336          * of the namekey synthesize the chain.
3337          *
3338          * The key range is inclusive of both key_beg and key_end.
3339          */
3340         dip = VTOI(dvp);
3341         ncp = nch->ncp;
3342         hmp = dip->hmp;
3343
3344         if (dip->flags & HAMMER_INODE_RO)
3345                 return (EROFS);
3346
3347         namekey = hammer_directory_namekey(dip, ncp->nc_name, ncp->nc_nlen,
3348                                            &max_iterations);
3349 retry:
3350         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
3351         cursor.key_beg.localization = dip->obj_localization |
3352                                       hammer_dir_localization(dip);
3353         cursor.key_beg.obj_id = dip->obj_id;
3354         cursor.key_beg.key = namekey;
3355         cursor.key_beg.create_tid = 0;
3356         cursor.key_beg.delete_tid = 0;
3357         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
3358         cursor.key_beg.obj_type = 0;
3359
3360         cursor.key_end = cursor.key_beg;
3361         cursor.key_end.key += max_iterations;
3362         cursor.asof = dip->obj_asof;
3363         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
3364
3365         /*
3366          * Scan all matching records (the chain), locate the one matching
3367          * the requested path component.  info->last_error contains the
3368          * error code on search termination and could be 0, ENOENT, or
3369          * something else.
3370          *
3371          * The hammer_ip_*() functions merge in-memory records with on-disk
3372          * records for the purposes of the search.
3373          */
3374         error = hammer_ip_first(&cursor);
3375
3376         while (error == 0) {
3377                 error = hammer_ip_resolve_data(&cursor);
3378                 if (error)
3379                         break;
3380                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
3381                 KKASSERT(nlen > 0);
3382                 if (ncp->nc_nlen == nlen &&
3383                     bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
3384                         break;
3385                 }
3386                 error = hammer_ip_next(&cursor);
3387         }
3388
3389         /*
3390          * If all is ok we have to get the inode so we can adjust nlinks.
3391          * To avoid a deadlock with the flusher we must release the inode
3392          * lock on the directory when acquiring the inode for the entry.
3393          *
3394          * If the target is a directory, it must be empty.
3395          */
3396         if (error == 0) {
3397                 hammer_unlock(&cursor.ip->lock);
3398                 ip = hammer_get_inode(trans, dip, cursor.data->entry.obj_id,
3399                                       hmp->asof,
3400                                       cursor.data->entry.localization,
3401                                       0, &error);
3402                 hammer_lock_sh(&cursor.ip->lock);
3403                 if (error == ENOENT) {
3404                         hkprintf("WARNING: Removing dirent w/missing inode "
3405                                 "\"%s\"\n"
3406                                 "\tobj_id = %016llx\n",
3407                                 ncp->nc_name,
3408                                 (long long)cursor.data->entry.obj_id);
3409                         error = 0;
3410                 }
3411
3412                 /*
3413                  * If isdir >= 0 we validate that the entry is or is not a
3414                  * directory.  If isdir < 0 we don't care.
3415                  */
3416                 if (error == 0 && isdir >= 0 && ip) {
3417                         if (isdir &&
3418                             ip->ino_data.obj_type != HAMMER_OBJTYPE_DIRECTORY) {
3419                                 error = ENOTDIR;
3420                         } else if (isdir == 0 &&
3421                             ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
3422                                 error = EISDIR;
3423                         }
3424                 }
3425
3426                 /*
3427                  * If we are trying to remove a directory the directory must
3428                  * be empty.
3429                  *
3430                  * The check directory code can loop and deadlock/retry.  Our
3431                  * own cursor's node locks must be released to avoid a 3-way
3432                  * deadlock with the flusher if the check directory code
3433                  * blocks.
3434                  *
3435                  * If any changes whatsoever have been made to the cursor
3436                  * set EDEADLK and retry.
3437                  *
3438                  * WARNING: See warnings in hammer_unlock_cursor()
3439                  *          function.
3440                  */
3441                 if (error == 0 && ip && ip->ino_data.obj_type ==
3442                                         HAMMER_OBJTYPE_DIRECTORY) {
3443                         hammer_unlock_cursor(&cursor);
3444                         error = hammer_ip_check_directory_empty(trans, ip);
3445                         hammer_lock_cursor(&cursor);
3446                         if (cursor.flags & HAMMER_CURSOR_RETEST) {
3447                                 hkprintf("Warning: avoided deadlock "
3448                                         "on rmdir '%s'\n",
3449                                         ncp->nc_name);
3450                                 error = EDEADLK;
3451                         }
3452                 }
3453
3454                 /*
3455                  * Delete the directory entry.
3456                  *
3457                  * WARNING: hammer_ip_del_directory() may have to terminate
3458                  * the cursor to avoid a deadlock.  It is ok to call
3459                  * hammer_done_cursor() twice.
3460                  */
3461                 if (error == 0) {
3462                         error = hammer_ip_del_directory(trans, &cursor,
3463                                                         dip, ip);
3464                 }
3465                 hammer_done_cursor(&cursor);
3466                 if (error == 0) {
3467                         /*
3468                          * Tell the namecache that we are now unlinked.
3469                          */
3470                         cache_unlink(nch);
3471
3472                         /*
3473                          * NOTE: ip->vp, if non-NULL, cannot be directly
3474                          *       referenced without formally acquiring the
3475                          *       vp since the vp might have zero refs on it,
3476                          *       or in the middle of a reclaim, etc.
3477                          *
3478                          * NOTE: The cache_setunresolved() can rip the vp
3479                          *       out from under us since the vp may not have
3480                          *       any refs, in which case ip->vp will be NULL
3481                          *       from the outset.
3482                          */
3483                         while (ip && ip->vp) {
3484                                 struct vnode *vp;
3485
3486                                 error = hammer_get_vnode(ip, &vp);
3487                                 if (error == 0 && vp) {
3488                                         vn_unlock(vp);
3489                                         hammer_knote(ip->vp, NOTE_DELETE);
3490 #if 0
3491                                         /*
3492                                          * Don't do this, it can deadlock
3493                                          * on concurrent rm's of hardlinks.
3494                                          * Shouldn't be needed any more.
3495                                          */
3496                                         cache_inval_vp(ip->vp, CINV_DESTROY);
3497 #endif
3498                                         vrele(vp);
3499                                         break;
3500                                 }
3501                                 hdkprintf("ip/vp race1 avoided\n");
3502                         }
3503                 }
3504                 if (ip)
3505                         hammer_rel_inode(ip, 0);
3506         } else {
3507                 hammer_done_cursor(&cursor);
3508         }
3509         if (error == EDEADLK)
3510                 goto retry;
3511
3512         return (error);
3513 }
3514
3515 /************************************************************************
3516  *                          FIFO AND SPECFS OPS                         *
3517  ************************************************************************
3518  *
3519  */
3520 static int
3521 hammer_vop_fifoclose (struct vop_close_args *ap)
3522 {
3523         /* XXX update itimes */
3524         return (VOCALL(&fifo_vnode_vops, &ap->a_head));
3525 }
3526
3527 static int
3528 hammer_vop_fiforead (struct vop_read_args *ap)
3529 {
3530         int error;
3531
3532         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3533         /* XXX update access time */
3534         return (error);
3535 }
3536
3537 static int
3538 hammer_vop_fifowrite (struct vop_write_args *ap)
3539 {
3540         int error;
3541
3542         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3543         /* XXX update access time */
3544         return (error);
3545 }
3546
3547 static
3548 int
3549 hammer_vop_fifokqfilter(struct vop_kqfilter_args *ap)
3550 {
3551         int error;
3552
3553         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3554         if (error)
3555                 error = hammer_vop_kqfilter(ap);
3556         return(error);
3557 }
3558
3559 /************************************************************************
3560  *                          KQFILTER OPS                                *
3561  ************************************************************************
3562  *
3563  */
3564 static void filt_hammerdetach(struct knote *kn);
3565 static int filt_hammerread(struct knote *kn, long hint);
3566 static int filt_hammerwrite(struct knote *kn, long hint);
3567 static int filt_hammervnode(struct knote *kn, long hint);
3568
3569 static struct filterops hammerread_filtops =
3570         { FILTEROP_ISFD | FILTEROP_MPSAFE,
3571           NULL, filt_hammerdetach, filt_hammerread };
3572 static struct filterops hammerwrite_filtops =
3573         { FILTEROP_ISFD | FILTEROP_MPSAFE,
3574           NULL, filt_hammerdetach, filt_hammerwrite };
3575 static struct filterops hammervnode_filtops =
3576         { FILTEROP_ISFD | FILTEROP_MPSAFE,
3577           NULL, filt_hammerdetach, filt_hammervnode };
3578
3579 static
3580 int
3581 hammer_vop_kqfilter(struct vop_kqfilter_args *ap)
3582 {
3583         struct vnode *vp = ap->a_vp;
3584         struct knote *kn = ap->a_kn;
3585
3586         switch (kn->kn_filter) {
3587         case EVFILT_READ:
3588                 kn->kn_fop = &hammerread_filtops;
3589                 break;
3590         case EVFILT_WRITE:
3591                 kn->kn_fop = &hammerwrite_filtops;
3592                 break;
3593         case EVFILT_VNODE:
3594                 kn->kn_fop = &hammervnode_filtops;
3595                 break;
3596         default:
3597                 return (EOPNOTSUPP);
3598         }
3599
3600         kn->kn_hook = (caddr_t)vp;
3601
3602         knote_insert(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3603
3604         return(0);
3605 }
3606
3607 static void
3608 filt_hammerdetach(struct knote *kn)
3609 {
3610         struct vnode *vp = (void *)kn->kn_hook;
3611
3612         knote_remove(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3613 }
3614
3615 static int
3616 filt_hammerread(struct knote *kn, long hint)
3617 {
3618         struct vnode *vp = (void *)kn->kn_hook;
3619         hammer_inode_t ip = VTOI(vp);
3620         hammer_mount_t hmp = ip->hmp;
3621         off_t off;
3622
3623         if (hint == NOTE_REVOKE) {
3624                 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
3625                 return(1);
3626         }
3627         lwkt_gettoken(&hmp->fs_token);  /* XXX use per-ip-token */
3628         off = ip->ino_data.size - kn->kn_fp->f_offset;
3629         kn->kn_data = (off < INTPTR_MAX) ? off : INTPTR_MAX;
3630         lwkt_reltoken(&hmp->fs_token);
3631         if (kn->kn_sfflags & NOTE_OLDAPI)
3632                 return(1);
3633         return (kn->kn_data != 0);
3634 }
3635
3636 static int
3637 filt_hammerwrite(struct knote *kn, long hint)
3638 {
3639         if (hint == NOTE_REVOKE)
3640                 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
3641         kn->kn_data = 0;
3642         return (1);
3643 }
3644
3645 static int
3646 filt_hammervnode(struct knote *kn, long hint)
3647 {
3648         if (kn->kn_sfflags & hint)
3649                 kn->kn_fflags |= hint;
3650         if (hint == NOTE_REVOKE) {
3651                 kn->kn_flags |= (EV_EOF | EV_NODATA);
3652                 return (1);
3653         }
3654         return (kn->kn_fflags != 0);
3655 }
3656