sys/vfs/hammer/hammer_vnops.c

   1 /*
   2  * Copyright (c) 2007-2008 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  *
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in
  15  *    the documentation and/or other materials provided with the
  16  *    distribution.
  17  * 3. Neither the name of The DragonFly Project nor the names of its
  18  *    contributors may be used to endorse or promote products derived
  19  *    from this software without specific, prior written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  32  * SUCH DAMAGE.
  33  *
  34  * $DragonFly: src/sys/vfs/hammer/hammer_vnops.c,v 1.102 2008/10/16 17:24:16 dillon Exp $
  35  */
  36
  37 #include <sys/param.h>
  38 #include <sys/systm.h>
  39 #include <sys/kernel.h>
  40 #include <sys/fcntl.h>
  41 #include <sys/namecache.h>
  42 #include <sys/vnode.h>
  43 #include <sys/lockf.h>
  44 #include <sys/event.h>
  45 #include <sys/stat.h>
  46 #include <sys/dirent.h>
  47 #include <sys/file.h>
  48 #include <vm/vm_extern.h>
  49 #include <vfs/fifofs/fifo.h>
  50
  51 #include <sys/mplock2.h>
  52
  53 #include "hammer.h"
  54
  55 /*
  56  * USERFS VNOPS
  57  */
  58 /*static int hammer_vop_vnoperate(struct vop_generic_args *);*/
  59 static int hammer_vop_fsync(struct vop_fsync_args *);
  60 static int hammer_vop_read(struct vop_read_args *);
  61 static int hammer_vop_write(struct vop_write_args *);
  62 static int hammer_vop_access(struct vop_access_args *);
  63 static int hammer_vop_advlock(struct vop_advlock_args *);
  64 static int hammer_vop_close(struct vop_close_args *);
  65 static int hammer_vop_ncreate(struct vop_ncreate_args *);
  66 static int hammer_vop_getattr(struct vop_getattr_args *);
  67 static int hammer_vop_nresolve(struct vop_nresolve_args *);
  68 static int hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *);
  69 static int hammer_vop_nlink(struct vop_nlink_args *);
  70 static int hammer_vop_nmkdir(struct vop_nmkdir_args *);
  71 static int hammer_vop_nmknod(struct vop_nmknod_args *);
  72 static int hammer_vop_open(struct vop_open_args *);
  73 static int hammer_vop_print(struct vop_print_args *);
  74 static int hammer_vop_readdir(struct vop_readdir_args *);
  75 static int hammer_vop_readlink(struct vop_readlink_args *);
  76 static int hammer_vop_nremove(struct vop_nremove_args *);
  77 static int hammer_vop_nrename(struct vop_nrename_args *);
  78 static int hammer_vop_nrmdir(struct vop_nrmdir_args *);
  79 static int hammer_vop_markatime(struct vop_markatime_args *);
  80 static int hammer_vop_setattr(struct vop_setattr_args *);
  81 static int hammer_vop_strategy(struct vop_strategy_args *);
  82 static int hammer_vop_bmap(struct vop_bmap_args *ap);
  83 static int hammer_vop_nsymlink(struct vop_nsymlink_args *);
  84 static int hammer_vop_nwhiteout(struct vop_nwhiteout_args *);
  85 static int hammer_vop_ioctl(struct vop_ioctl_args *);
  86 static int hammer_vop_mountctl(struct vop_mountctl_args *);
  87 static int hammer_vop_kqfilter (struct vop_kqfilter_args *);
  88
  89 static int hammer_vop_fifoclose (struct vop_close_args *);
  90 static int hammer_vop_fiforead (struct vop_read_args *);
  91 static int hammer_vop_fifowrite (struct vop_write_args *);
  92 static int hammer_vop_fifokqfilter (struct vop_kqfilter_args *);
  93
  94 struct vop_ops hammer_vnode_vops = {
  95         .vop_default =          vop_defaultop,
  96         .vop_fsync =            hammer_vop_fsync,
  97         .vop_getpages =         vop_stdgetpages,
  98         .vop_putpages =         vop_stdputpages,
  99         .vop_read =             hammer_vop_read,
 100         .vop_write =            hammer_vop_write,
 101         .vop_access =           hammer_vop_access,
 102         .vop_advlock =          hammer_vop_advlock,
 103         .vop_close =            hammer_vop_close,
 104         .vop_ncreate =          hammer_vop_ncreate,
 105         .vop_getattr =          hammer_vop_getattr,
 106         .vop_inactive =         hammer_vop_inactive,
 107         .vop_reclaim =          hammer_vop_reclaim,
 108         .vop_nresolve =         hammer_vop_nresolve,
 109         .vop_nlookupdotdot =    hammer_vop_nlookupdotdot,
 110         .vop_nlink =            hammer_vop_nlink,
 111         .vop_nmkdir =           hammer_vop_nmkdir,
 112         .vop_nmknod =           hammer_vop_nmknod,
 113         .vop_open =             hammer_vop_open,
 114         .vop_pathconf =         vop_stdpathconf,
 115         .vop_print =            hammer_vop_print,
 116         .vop_readdir =          hammer_vop_readdir,
 117         .vop_readlink =         hammer_vop_readlink,
 118         .vop_nremove =          hammer_vop_nremove,
 119         .vop_nrename =          hammer_vop_nrename,
 120         .vop_nrmdir =           hammer_vop_nrmdir,
 121         .vop_markatime =        hammer_vop_markatime,
 122         .vop_setattr =          hammer_vop_setattr,
 123         .vop_bmap =             hammer_vop_bmap,
 124         .vop_strategy =         hammer_vop_strategy,
 125         .vop_nsymlink =         hammer_vop_nsymlink,
 126         .vop_nwhiteout =        hammer_vop_nwhiteout,
 127         .vop_ioctl =            hammer_vop_ioctl,
 128         .vop_mountctl =         hammer_vop_mountctl,
 129         .vop_kqfilter =         hammer_vop_kqfilter
 130 };
 131
 132 struct vop_ops hammer_spec_vops = {
 133         .vop_default =          vop_defaultop,
 134         .vop_fsync =            hammer_vop_fsync,
 135         .vop_read =             vop_stdnoread,
 136         .vop_write =            vop_stdnowrite,
 137         .vop_access =           hammer_vop_access,
 138         .vop_close =            hammer_vop_close,
 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 };
 145
 146 struct vop_ops hammer_fifo_vops = {
 147         .vop_default =          fifo_vnoperate,
 148         .vop_fsync =            hammer_vop_fsync,
 149         .vop_read =             hammer_vop_fiforead,
 150         .vop_write =            hammer_vop_fifowrite,
 151         .vop_access =           hammer_vop_access,
 152         .vop_close =            hammer_vop_fifoclose,
 153         .vop_markatime =        hammer_vop_markatime,
 154         .vop_getattr =          hammer_vop_getattr,
 155         .vop_inactive =         hammer_vop_inactive,
 156         .vop_reclaim =          hammer_vop_reclaim,
 157         .vop_setattr =          hammer_vop_setattr,
 158         .vop_kqfilter =         hammer_vop_fifokqfilter
 159 };
 160
 161 static __inline
 162 void
 163 hammer_knote(struct vnode *vp, int flags)
 164 {
 165         if (flags)
 166                 KNOTE(&vp->v_pollinfo.vpi_kqinfo.ki_note, flags);
 167 }
 168
 169 #ifdef DEBUG_TRUNCATE
 170 struct hammer_inode *HammerTruncIp;
 171 #endif
 172
 173 static int hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
 174                            struct vnode *dvp, struct ucred *cred,
 175                            int flags, int isdir);
 176 static int hammer_vop_strategy_read(struct vop_strategy_args *ap);
 177 static int hammer_vop_strategy_write(struct vop_strategy_args *ap);
 178
 179 #if 0
 180 static
 181 int
 182 hammer_vop_vnoperate(struct vop_generic_args *)
 183 {
 184         return (VOCALL(&hammer_vnode_vops, ap));
 185 }
 186 #endif
 187
 188 /*
 189  * hammer_vop_fsync { vp, waitfor }
 190  *
 191  * fsync() an inode to disk and wait for it to be completely committed
 192  * such that the information would not be undone if a crash occured after
 193  * return.
 194  *
 195  * NOTE: HAMMER's fsync()'s are going to remain expensive until we implement
 196  *       a REDO log.  A sysctl is provided to relax HAMMER's fsync()
 197  *       operation.
 198  *
 199  *       Ultimately the combination of a REDO log and use of fast storage
 200  *       to front-end cluster caches will make fsync fast, but it aint
 201  *       here yet.  And, in anycase, we need real transactional
 202  *       all-or-nothing features which are not restricted to a single file.
 203  */
 204 static
 205 int
 206 hammer_vop_fsync(struct vop_fsync_args *ap)
 207 {
 208         hammer_inode_t ip = VTOI(ap->a_vp);
 209         hammer_mount_t hmp = ip->hmp;
 210         int waitfor = ap->a_waitfor;
 211         int mode;
 212
 213         lwkt_gettoken(&hmp->fs_token);
 214
 215         /*
 216          * Fsync rule relaxation (default is either full synchronous flush
 217          * or REDO semantics with synchronous flush).
 218          */
 219         if (ap->a_flags & VOP_FSYNC_SYSCALL) {
 220                 switch(hammer_fsync_mode) {
 221                 case 0:
 222 mode0:
 223                         /* no REDO, full synchronous flush */
 224                         goto skip;
 225                 case 1:
 226 mode1:
 227                         /* no REDO, full asynchronous flush */
 228                         if (waitfor == MNT_WAIT)
 229                                 waitfor = MNT_NOWAIT;
 230                         goto skip;
 231                 case 2:
 232                         /* REDO semantics, synchronous flush */
 233                         if (hmp->version < HAMMER_VOL_VERSION_FOUR)
 234                                 goto mode0;
 235                         mode = HAMMER_FLUSH_UNDOS_AUTO;
 236                         break;
 237                 case 3:
 238                         /* REDO semantics, relaxed asynchronous flush */
 239                         if (hmp->version < HAMMER_VOL_VERSION_FOUR)
 240                                 goto mode1;
 241                         mode = HAMMER_FLUSH_UNDOS_RELAXED;
 242                         if (waitfor == MNT_WAIT)
 243                                 waitfor = MNT_NOWAIT;
 244                         break;
 245                 case 4:
 246                         /* ignore the fsync() system call */
 247                         lwkt_reltoken(&hmp->fs_token);
 248                         return(0);
 249                 default:
 250                         /* we have to do something */
 251                         mode = HAMMER_FLUSH_UNDOS_RELAXED;
 252                         if (waitfor == MNT_WAIT)
 253                                 waitfor = MNT_NOWAIT;
 254                         break;
 255                 }
 256
 257                 /*
 258                  * Fast fsync only needs to flush the UNDO/REDO fifo if
 259                  * HAMMER_INODE_REDO is non-zero and the only modifications
 260                  * made to the file are write or write-extends.
 261                  */
 262                 if ((ip->flags & HAMMER_INODE_REDO) &&
 263                     (ip->flags & HAMMER_INODE_MODMASK_NOREDO) == 0
 264                 ) {
 265                         ++hammer_count_fsyncs;
 266                         hammer_flusher_flush_undos(hmp, mode);
 267                         ip->redo_count = 0;
 268                         lwkt_reltoken(&hmp->fs_token);
 269                         return(0);
 270                 }
 271
 272                 /*
 273                  * REDO is enabled by fsync(), the idea being we really only
 274                  * want to lay down REDO records when programs are using
 275                  * fsync() heavily.  The first fsync() on the file starts
 276                  * the gravy train going and later fsync()s keep it hot by
 277                  * resetting the redo_count.
 278                  *
 279                  * We weren't running REDOs before now so we have to fall
 280                  * through and do a full fsync of what we have.
 281                  */
 282                 if (hmp->version >= HAMMER_VOL_VERSION_FOUR &&
 283                     (hmp->flags & HAMMER_MOUNT_REDO_RECOVERY_RUN) == 0) {
 284                         ip->flags |= HAMMER_INODE_REDO;
 285                         ip->redo_count = 0;
 286                 }
 287         }
 288 skip:
 289
 290         /*
 291          * Do a full flush sequence.
 292          */
 293         ++hammer_count_fsyncs;
 294         vfsync(ap->a_vp, waitfor, 1, NULL, NULL);
 295         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 296         if (waitfor == MNT_WAIT) {
 297                 vn_unlock(ap->a_vp);
 298                 hammer_wait_inode(ip);
 299                 vn_lock(ap->a_vp, LK_EXCLUSIVE | LK_RETRY);
 300         }
 301         lwkt_reltoken(&hmp->fs_token);
 302         return (ip->error);
 303 }
 304
 305 /*
 306  * hammer_vop_read { vp, uio, ioflag, cred }
 307  *
 308  * MPSAFE (for the cache safe does not require fs_token)
 309  */
 310 static
 311 int
 312 hammer_vop_read(struct vop_read_args *ap)
 313 {
 314         struct hammer_transaction trans;
 315         hammer_inode_t ip;
 316         hammer_mount_t hmp;
 317         off_t offset;
 318         struct buf *bp;
 319         struct uio *uio;
 320         int error;
 321         int n;
 322         int seqcount;
 323         int ioseqcount;
 324         int blksize;
 325         int bigread;
 326         int got_fstoken;
 327
 328         if (ap->a_vp->v_type != VREG)
 329                 return (EINVAL);
 330         ip = VTOI(ap->a_vp);
 331         hmp = ip->hmp;
 332         error = 0;
 333         uio = ap->a_uio;
 334
 335         /*
 336          * Allow the UIO's size to override the sequential heuristic.
 337          */
 338         blksize = hammer_blocksize(uio->uio_offset);
 339         seqcount = (uio->uio_resid + (BKVASIZE - 1)) / BKVASIZE;
 340         ioseqcount = (ap->a_ioflag >> 16);
 341         if (seqcount < ioseqcount)
 342                 seqcount = ioseqcount;
 343
 344         /*
 345          * If reading or writing a huge amount of data we have to break
 346          * atomicy and allow the operation to be interrupted by a signal
 347          * or it can DOS the machine.
 348          */
 349         bigread = (uio->uio_resid > 100 * 1024 * 1024);
 350         got_fstoken = 0;
 351
 352         /*
 353          * Access the data typically in HAMMER_BUFSIZE blocks via the
 354          * buffer cache, but HAMMER may use a variable block size based
 355          * on the offset.
 356          *
 357          * XXX Temporary hack, delay the start transaction while we remain
 358          *     MPSAFE.  NOTE: ino_data.size cannot change while vnode is
 359          *     locked-shared.
 360          */
 361         while (uio->uio_resid > 0 && uio->uio_offset < ip->ino_data.size) {
 362                 int64_t base_offset;
 363                 int64_t file_limit;
 364
 365                 blksize = hammer_blocksize(uio->uio_offset);
 366                 offset = (int)uio->uio_offset & (blksize - 1);
 367                 base_offset = uio->uio_offset - offset;
 368
 369                 if (bigread && (error = hammer_signal_check(ip->hmp)) != 0)
 370                         break;
 371
 372                 /*
 373                  * MPSAFE
 374                  */
 375                 bp = getcacheblk(ap->a_vp, base_offset);
 376                 if (bp) {
 377                         error = 0;
 378                         goto skip;
 379                 }
 380
 381                 /*
 382                  * MPUNSAFE
 383                  */
 384                 if (got_fstoken == 0) {
 385                         lwkt_gettoken(&hmp->fs_token);
 386                         got_fstoken = 1;
 387                         hammer_start_transaction(&trans, ip->hmp);
 388                 }
 389
 390                 if (hammer_cluster_enable) {
 391                         /*
 392                          * Use file_limit to prevent cluster_read() from
 393                          * creating buffers of the wrong block size past
 394                          * the demarc.
 395                          */
 396                         file_limit = ip->ino_data.size;
 397                         if (base_offset < HAMMER_XDEMARC &&
 398                             file_limit > HAMMER_XDEMARC) {
 399                                 file_limit = HAMMER_XDEMARC;
 400                         }
 401                         error = cluster_read(ap->a_vp,
 402                                              file_limit, base_offset,
 403                                              blksize, uio->uio_resid,
 404                                              seqcount * BKVASIZE, &bp);
 405                 } else {
 406                         error = bread(ap->a_vp, base_offset, blksize, &bp);
 407                 }
 408                 if (error) {
 409                         brelse(bp);
 410                         break;
 411                 }
 412 skip:
 413                 if ((hammer_debug_io & 0x0001) && (bp->b_flags & B_IODEBUG)) {
 414                         kprintf("doff %016jx read file %016jx@%016jx\n",
 415                                 (intmax_t)bp->b_bio2.bio_offset,
 416                                 (intmax_t)ip->obj_id,
 417                                 (intmax_t)bp->b_loffset);
 418                 }
 419                 bp->b_flags &= ~B_IODEBUG;
 420
 421                 /* bp->b_flags |= B_CLUSTEROK; temporarily disabled */
 422                 n = blksize - offset;
 423                 if (n > uio->uio_resid)
 424                         n = uio->uio_resid;
 425                 if (n > ip->ino_data.size - uio->uio_offset)
 426                         n = (int)(ip->ino_data.size - uio->uio_offset);
 427                 error = uiomove((char *)bp->b_data + offset, n, uio);
 428
 429                 /* data has a lower priority then meta-data */
 430                 bp->b_flags |= B_AGE;
 431                 bqrelse(bp);
 432                 if (error)
 433                         break;
 434                 hammer_stats_file_read += n;
 435         }
 436
 437         /*
 438          * XXX only update the atime if we had to get the MP lock.
 439          * XXX hack hack hack, fixme.
 440          */
 441         if (got_fstoken) {
 442                 if ((ip->flags & HAMMER_INODE_RO) == 0 &&
 443                     (ip->hmp->mp->mnt_flag & MNT_NOATIME) == 0) {
 444                         ip->ino_data.atime = trans.time;
 445                         hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
 446                 }
 447                 hammer_done_transaction(&trans);
 448                 lwkt_reltoken(&hmp->fs_token);
 449         }
 450         return (error);
 451 }
 452
 453 /*
 454  * hammer_vop_write { vp, uio, ioflag, cred }
 455  */
 456 static
 457 int
 458 hammer_vop_write(struct vop_write_args *ap)
 459 {
 460         struct hammer_transaction trans;
 461         struct hammer_inode *ip;
 462         hammer_mount_t hmp;
 463         struct uio *uio;
 464         int offset;
 465         off_t base_offset;
 466         struct buf *bp;
 467         int kflags;
 468         int error;
 469         int n;
 470         int flags;
 471         int seqcount;
 472         int bigwrite;
 473
 474         if (ap->a_vp->v_type != VREG)
 475                 return (EINVAL);
 476         ip = VTOI(ap->a_vp);
 477         hmp = ip->hmp;
 478         error = 0;
 479         kflags = 0;
 480         seqcount = ap->a_ioflag >> 16;
 481
 482         if (ip->flags & HAMMER_INODE_RO)
 483                 return (EROFS);
 484
 485         /*
 486          * Create a transaction to cover the operations we perform.
 487          */
 488         lwkt_gettoken(&hmp->fs_token);
 489         hammer_start_transaction(&trans, hmp);
 490         uio = ap->a_uio;
 491
 492         /*
 493          * Check append mode
 494          */
 495         if (ap->a_ioflag & IO_APPEND)
 496                 uio->uio_offset = ip->ino_data.size;
 497
 498         /*
 499          * Check for illegal write offsets.  Valid range is 0...2^63-1.
 500          *
 501          * NOTE: the base_off assignment is required to work around what
 502          * I consider to be a GCC-4 optimization bug.
 503          */
 504         if (uio->uio_offset < 0) {
 505                 hammer_done_transaction(&trans);
 506                 lwkt_reltoken(&hmp->fs_token);
 507                 return (EFBIG);
 508         }
 509         base_offset = uio->uio_offset + uio->uio_resid; /* work around gcc-4 */
 510         if (uio->uio_resid > 0 && base_offset <= uio->uio_offset) {
 511                 hammer_done_transaction(&trans);
 512                 lwkt_reltoken(&hmp->fs_token);
 513                 return (EFBIG);
 514         }
 515
 516         /*
 517          * If reading or writing a huge amount of data we have to break
 518          * atomicy and allow the operation to be interrupted by a signal
 519          * or it can DOS the machine.
 520          *
 521          * Preset redo_count so we stop generating REDOs earlier if the
 522          * limit is exceeded.
 523          */
 524         bigwrite = (uio->uio_resid > 100 * 1024 * 1024);
 525         if ((ip->flags & HAMMER_INODE_REDO) &&
 526             ip->redo_count < hammer_limit_redo) {
 527                 ip->redo_count += uio->uio_resid;
 528         }
 529
 530         /*
 531          * Access the data typically in HAMMER_BUFSIZE blocks via the
 532          * buffer cache, but HAMMER may use a variable block size based
 533          * on the offset.
 534          */
 535         while (uio->uio_resid > 0) {
 536                 int fixsize = 0;
 537                 int blksize;
 538                 int blkmask;
 539                 int trivial;
 540                 int endofblk;
 541                 off_t nsize;
 542
 543                 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_WRITE)) != 0)
 544                         break;
 545                 if (bigwrite && (error = hammer_signal_check(hmp)) != 0)
 546                         break;
 547
 548                 blksize = hammer_blocksize(uio->uio_offset);
 549
 550                 /*
 551                  * Do not allow HAMMER to blow out the buffer cache.  Very
 552                  * large UIOs can lockout other processes due to bwillwrite()
 553                  * mechanics.
 554                  *
 555                  * The hammer inode is not locked during these operations.
 556                  * The vnode is locked which can interfere with the pageout
 557                  * daemon for non-UIO_NOCOPY writes but should not interfere
 558                  * with the buffer cache.  Even so, we cannot afford to
 559                  * allow the pageout daemon to build up too many dirty buffer
 560                  * cache buffers.
 561                  *
 562                  * Only call this if we aren't being recursively called from
 563                  * a virtual disk device (vn), else we may deadlock.
 564                  */
 565                 if ((ap->a_ioflag & IO_RECURSE) == 0)
 566                         bwillwrite(blksize);
 567
 568                 /*
 569                  * Control the number of pending records associated with
 570                  * this inode.  If too many have accumulated start a
 571                  * flush.  Try to maintain a pipeline with the flusher.
 572                  */
 573                 if (ip->rsv_recs >= hammer_limit_inode_recs) {
 574                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 575                 }
 576                 if (ip->rsv_recs >= hammer_limit_inode_recs * 2) {
 577                         while (ip->rsv_recs >= hammer_limit_inode_recs) {
 578                                 tsleep(&ip->rsv_recs, 0, "hmrwww", hz);
 579                         }
 580                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 581                 }
 582
 583 #if 0
 584                 /*
 585                  * Do not allow HAMMER to blow out system memory by
 586                  * accumulating too many records.   Records are so well
 587                  * decoupled from the buffer cache that it is possible
 588                  * for userland to push data out to the media via
 589                  * direct-write, but build up the records queued to the
 590                  * backend faster then the backend can flush them out.
 591                  * HAMMER has hit its write limit but the frontend has
 592                  * no pushback to slow it down.
 593                  */
 594                 if (hmp->rsv_recs > hammer_limit_recs / 2) {
 595                         /*
 596                          * Get the inode on the flush list
 597                          */
 598                         if (ip->rsv_recs >= 64)
 599                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 600                         else if (ip->rsv_recs >= 16)
 601                                 hammer_flush_inode(ip, 0);
 602
 603                         /*
 604                          * Keep the flusher going if the system keeps
 605                          * queueing records.
 606                          */
 607                         delta = hmp->count_newrecords -
 608                                 hmp->last_newrecords;
 609                         if (delta < 0 || delta > hammer_limit_recs / 2) {
 610                                 hmp->last_newrecords = hmp->count_newrecords;
 611                                 hammer_sync_hmp(hmp, MNT_NOWAIT);
 612                         }
 613
 614                         /*
 615                          * If we have gotten behind start slowing
 616                          * down the writers.
 617                          */
 618                         delta = (hmp->rsv_recs - hammer_limit_recs) *
 619                                 hz / hammer_limit_recs;
 620                         if (delta > 0)
 621                                 tsleep(&trans, 0, "hmrslo", delta);
 622                 }
 623 #endif
 624
 625                 /*
 626                  * Calculate the blocksize at the current offset and figure
 627                  * out how much we can actually write.
 628                  */
 629                 blkmask = blksize - 1;
 630                 offset = (int)uio->uio_offset & blkmask;
 631                 base_offset = uio->uio_offset & ~(int64_t)blkmask;
 632                 n = blksize - offset;
 633                 if (n > uio->uio_resid) {
 634                         n = uio->uio_resid;
 635                         endofblk = 0;
 636                 } else {
 637                         endofblk = 1;
 638                 }
 639                 nsize = uio->uio_offset + n;
 640                 if (nsize > ip->ino_data.size) {
 641                         if (uio->uio_offset > ip->ino_data.size)
 642                                 trivial = 0;
 643                         else
 644                                 trivial = 1;
 645                         nvextendbuf(ap->a_vp,
 646                                     ip->ino_data.size,
 647                                     nsize,
 648                                     hammer_blocksize(ip->ino_data.size),
 649                                     hammer_blocksize(nsize),
 650                                     hammer_blockoff(ip->ino_data.size),
 651                                     hammer_blockoff(nsize),
 652                                     trivial);
 653                         fixsize = 1;
 654                         kflags |= NOTE_EXTEND;
 655                 }
 656
 657                 if (uio->uio_segflg == UIO_NOCOPY) {
 658                         /*
 659                          * Issuing a write with the same data backing the
 660                          * buffer.  Instantiate the buffer to collect the
 661                          * backing vm pages, then read-in any missing bits.
 662                          *
 663                          * This case is used by vop_stdputpages().
 664                          */
 665                         bp = getblk(ap->a_vp, base_offset,
 666                                     blksize, GETBLK_BHEAVY, 0);
 667                         if ((bp->b_flags & B_CACHE) == 0) {
 668                                 bqrelse(bp);
 669                                 error = bread(ap->a_vp, base_offset,
 670                                               blksize, &bp);
 671                         }
 672                 } else if (offset == 0 && uio->uio_resid >= blksize) {
 673                         /*
 674                          * Even though we are entirely overwriting the buffer
 675                          * we may still have to zero it out to avoid a
 676                          * mmap/write visibility issue.
 677                          */
 678                         bp = getblk(ap->a_vp, base_offset, blksize, GETBLK_BHEAVY, 0);
 679                         if ((bp->b_flags & B_CACHE) == 0)
 680                                 vfs_bio_clrbuf(bp);
 681                 } else if (base_offset >= ip->ino_data.size) {
 682                         /*
 683                          * If the base offset of the buffer is beyond the
 684                          * file EOF, we don't have to issue a read.
 685                          */
 686                         bp = getblk(ap->a_vp, base_offset,
 687                                     blksize, GETBLK_BHEAVY, 0);
 688                         vfs_bio_clrbuf(bp);
 689                 } else {
 690                         /*
 691                          * Partial overwrite, read in any missing bits then
 692                          * replace the portion being written.
 693                          */
 694                         error = bread(ap->a_vp, base_offset, blksize, &bp);
 695                         if (error == 0)
 696                                 bheavy(bp);
 697                 }
 698                 if (error == 0)
 699                         error = uiomove(bp->b_data + offset, n, uio);
 700
 701                 /*
 702                  * Generate REDO records if enabled and redo_count will not
 703                  * exceeded the limit.
 704                  *
 705                  * If redo_count exceeds the limit we stop generating records
 706                  * and clear HAMMER_INODE_REDO.  This will cause the next
 707                  * fsync() to do a full meta-data sync instead of just an
 708                  * UNDO/REDO fifo update.
 709                  *
 710                  * When clearing HAMMER_INODE_REDO any pre-existing REDOs
 711                  * will still be tracked.  The tracks will be terminated
 712                  * when the related meta-data (including possible data
 713                  * modifications which are not tracked via REDO) is
 714                  * flushed.
 715                  */
 716                 if ((ip->flags & HAMMER_INODE_REDO) && error == 0) {
 717                         if (ip->redo_count < hammer_limit_redo) {
 718                                 bp->b_flags |= B_VFSFLAG1;
 719                                 error = hammer_generate_redo(&trans, ip,
 720                                                      base_offset + offset,
 721                                                      HAMMER_REDO_WRITE,
 722                                                      bp->b_data + offset,
 723                                                      (size_t)n);
 724                         } else {
 725                                 ip->flags &= ~HAMMER_INODE_REDO;
 726                         }
 727                 }
 728
 729                 /*
 730                  * If we screwed up we have to undo any VM size changes we
 731                  * made.
 732                  */
 733                 if (error) {
 734                         brelse(bp);
 735                         if (fixsize) {
 736                                 nvtruncbuf(ap->a_vp, ip->ino_data.size,
 737                                           hammer_blocksize(ip->ino_data.size),
 738                                           hammer_blockoff(ip->ino_data.size));
 739                         }
 740                         break;
 741                 }
 742                 kflags |= NOTE_WRITE;
 743                 hammer_stats_file_write += n;
 744                 /* bp->b_flags |= B_CLUSTEROK; temporarily disabled */
 745                 if (ip->ino_data.size < uio->uio_offset) {
 746                         ip->ino_data.size = uio->uio_offset;
 747                         flags = HAMMER_INODE_SDIRTY;
 748                 } else {
 749                         flags = 0;
 750                 }
 751                 ip->ino_data.mtime = trans.time;
 752                 flags |= HAMMER_INODE_MTIME | HAMMER_INODE_BUFS;
 753                 hammer_modify_inode(&trans, ip, flags);
 754
 755                 /*
 756                  * Once we dirty the buffer any cached zone-X offset
 757                  * becomes invalid.  HAMMER NOTE: no-history mode cannot
 758                  * allow overwriting over the same data sector unless
 759                  * we provide UNDOs for the old data, which we don't.
 760                  */
 761                 bp->b_bio2.bio_offset = NOOFFSET;
 762
 763                 /*
 764                  * Final buffer disposition.
 765                  *
 766                  * Because meta-data updates are deferred, HAMMER is
 767                  * especially sensitive to excessive bdwrite()s because
 768                  * the I/O stream is not broken up by disk reads.  So the
 769                  * buffer cache simply cannot keep up.
 770                  *
 771                  * WARNING!  blksize is variable.  cluster_write() is
 772                  *           expected to not blow up if it encounters
 773                  *           buffers that do not match the passed blksize.
 774                  *
 775                  * NOTE!  Hammer shouldn't need to bawrite()/cluster_write().
 776                  *        The ip->rsv_recs check should burst-flush the data.
 777                  *        If we queue it immediately the buf could be left
 778                  *        locked on the device queue for a very long time.
 779                  *
 780                  * NOTE!  To avoid degenerate stalls due to mismatched block
 781                  *        sizes we only honor IO_DIRECT on the write which
 782                  *        abuts the end of the buffer.  However, we must
 783                  *        honor IO_SYNC in case someone is silly enough to
 784                  *        configure a HAMMER file as swap, or when HAMMER
 785                  *        is serving NFS (for commits).  Ick ick.
 786                  */
 787                 bp->b_flags |= B_AGE;
 788                 if (ap->a_ioflag & IO_SYNC) {
 789                         bwrite(bp);
 790                 } else if ((ap->a_ioflag & IO_DIRECT) && endofblk) {
 791                         bawrite(bp);
 792                 } else {
 793 #if 0
 794                 if (offset + n == blksize) {
 795                         if (hammer_cluster_enable == 0 ||
 796                             (ap->a_vp->v_mount->mnt_flag & MNT_NOCLUSTERW)) {
 797                                 bawrite(bp);
 798                         } else {
 799                                 cluster_write(bp, ip->ino_data.size,
 800                                               blksize, seqcount);
 801                         }
 802                 } else {
 803 #endif
 804                         bdwrite(bp);
 805                 }
 806         }
 807         hammer_done_transaction(&trans);
 808         hammer_knote(ap->a_vp, kflags);
 809         lwkt_reltoken(&hmp->fs_token);
 810         return (error);
 811 }
 812
 813 /*
 814  * hammer_vop_access { vp, mode, cred }
 815  *
 816  * MPSAFE - does not require fs_token
 817  */
 818 static
 819 int
 820 hammer_vop_access(struct vop_access_args *ap)
 821 {
 822         struct hammer_inode *ip = VTOI(ap->a_vp);
 823         uid_t uid;
 824         gid_t gid;
 825         int error;
 826
 827         ++hammer_stats_file_iopsr;
 828         uid = hammer_to_unix_xid(&ip->ino_data.uid);
 829         gid = hammer_to_unix_xid(&ip->ino_data.gid);
 830
 831         error = vop_helper_access(ap, uid, gid, ip->ino_data.mode,
 832                                   ip->ino_data.uflags);
 833         return (error);
 834 }
 835
 836 /*
 837  * hammer_vop_advlock { vp, id, op, fl, flags }
 838  *
 839  * MPSAFE - does not require fs_token
 840  */
 841 static
 842 int
 843 hammer_vop_advlock(struct vop_advlock_args *ap)
 844 {
 845         hammer_inode_t ip = VTOI(ap->a_vp);
 846
 847         return (lf_advlock(ap, &ip->advlock, ip->ino_data.size));
 848 }
 849
 850 /*
 851  * hammer_vop_close { vp, fflag }
 852  *
 853  * We can only sync-on-close for normal closes.  XXX disabled for now.
 854  */
 855 static
 856 int
 857 hammer_vop_close(struct vop_close_args *ap)
 858 {
 859 #if 0
 860         struct vnode *vp = ap->a_vp;
 861         hammer_inode_t ip = VTOI(vp);
 862         int waitfor;
 863         if (ip->flags & (HAMMER_INODE_CLOSESYNC|HAMMER_INODE_CLOSEASYNC)) {
 864                 if (vn_islocked(vp) == LK_EXCLUSIVE &&
 865                     (vp->v_flag & (VINACTIVE|VRECLAIMED)) == 0) {
 866                         if (ip->flags & HAMMER_INODE_CLOSESYNC)
 867                                 waitfor = MNT_WAIT;
 868                         else
 869                                 waitfor = MNT_NOWAIT;
 870                         ip->flags &= ~(HAMMER_INODE_CLOSESYNC |
 871                                        HAMMER_INODE_CLOSEASYNC);
 872                         VOP_FSYNC(vp, MNT_NOWAIT, waitfor);
 873                 }
 874         }
 875 #endif
 876         return (vop_stdclose(ap));
 877 }
 878
 879 /*
 880  * hammer_vop_ncreate { nch, dvp, vpp, cred, vap }
 881  *
 882  * The operating system has already ensured that the directory entry
 883  * does not exist and done all appropriate namespace locking.
 884  */
 885 static
 886 int
 887 hammer_vop_ncreate(struct vop_ncreate_args *ap)
 888 {
 889         struct hammer_transaction trans;
 890         struct hammer_inode *dip;
 891         struct hammer_inode *nip;
 892         struct nchandle *nch;
 893         hammer_mount_t hmp;
 894         int error;
 895
 896         nch = ap->a_nch;
 897         dip = VTOI(ap->a_dvp);
 898         hmp = dip->hmp;
 899
 900         if (dip->flags & HAMMER_INODE_RO)
 901                 return (EROFS);
 902         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
 903                 return (error);
 904
 905         /*
 906          * Create a transaction to cover the operations we perform.
 907          */
 908         lwkt_gettoken(&hmp->fs_token);
 909         hammer_start_transaction(&trans, hmp);
 910         ++hammer_stats_file_iopsw;
 911
 912         /*
 913          * Create a new filesystem object of the requested type.  The
 914          * returned inode will be referenced and shared-locked to prevent
 915          * it from being moved to the flusher.
 916          */
 917         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
 918                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
 919                                     NULL, &nip);
 920         if (error) {
 921                 hkprintf("hammer_create_inode error %d\n", error);
 922                 hammer_done_transaction(&trans);
 923                 *ap->a_vpp = NULL;
 924                 lwkt_reltoken(&hmp->fs_token);
 925                 return (error);
 926         }
 927
 928         /*
 929          * Add the new filesystem object to the directory.  This will also
 930          * bump the inode's link count.
 931          */
 932         error = hammer_ip_add_directory(&trans, dip,
 933                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
 934                                         nip);
 935         if (error)
 936                 hkprintf("hammer_ip_add_directory error %d\n", error);
 937
 938         /*
 939          * Finish up.
 940          */
 941         if (error) {
 942                 hammer_rel_inode(nip, 0);
 943                 hammer_done_transaction(&trans);
 944                 *ap->a_vpp = NULL;
 945         } else {
 946                 error = hammer_get_vnode(nip, ap->a_vpp);
 947                 hammer_done_transaction(&trans);
 948                 hammer_rel_inode(nip, 0);
 949                 if (error == 0) {
 950                         cache_setunresolved(ap->a_nch);
 951                         cache_setvp(ap->a_nch, *ap->a_vpp);
 952                 }
 953                 hammer_knote(ap->a_dvp, NOTE_WRITE);
 954         }
 955         lwkt_reltoken(&hmp->fs_token);
 956         return (error);
 957 }
 958
 959 /*
 960  * hammer_vop_getattr { vp, vap }
 961  *
 962  * Retrieve an inode's attribute information.  When accessing inodes
 963  * historically we fake the atime field to ensure consistent results.
 964  * The atime field is stored in the B-Tree element and allowed to be
 965  * updated without cycling the element.
 966  *
 967  * MPSAFE - does not require fs_token
 968  */
 969 static
 970 int
 971 hammer_vop_getattr(struct vop_getattr_args *ap)
 972 {
 973         struct hammer_inode *ip = VTOI(ap->a_vp);
 974         struct vattr *vap = ap->a_vap;
 975
 976         /*
 977          * We want the fsid to be different when accessing a filesystem
 978          * with different as-of's so programs like diff don't think
 979          * the files are the same.
 980          *
 981          * We also want the fsid to be the same when comparing snapshots,
 982          * or when comparing mirrors (which might be backed by different
 983          * physical devices).  HAMMER fsids are based on the PFS's
 984          * shared_uuid field.
 985          *
 986          * XXX there is a chance of collision here.  The va_fsid reported
 987          * by stat is different from the more involved fsid used in the
 988          * mount structure.
 989          */
 990         ++hammer_stats_file_iopsr;
 991         hammer_lock_sh(&ip->lock);
 992         vap->va_fsid = ip->pfsm->fsid_udev ^ (u_int32_t)ip->obj_asof ^
 993                        (u_int32_t)(ip->obj_asof >> 32);
 994
 995         vap->va_fileid = ip->ino_leaf.base.obj_id;
 996         vap->va_mode = ip->ino_data.mode;
 997         vap->va_nlink = ip->ino_data.nlinks;
 998         vap->va_uid = hammer_to_unix_xid(&ip->ino_data.uid);
 999         vap->va_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1000         vap->va_rmajor = 0;
1001         vap->va_rminor = 0;
1002         vap->va_size = ip->ino_data.size;
1003
1004         /*
1005          * Special case for @@PFS softlinks.  The actual size of the
1006          * expanded softlink is "@@0x%016llx:%05d" == 26 bytes.
1007          * or for MAX_TID is    "@@-1:%05d" == 10 bytes.
1008          */
1009         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_SOFTLINK &&
1010             ip->ino_data.size == 10 &&
1011             ip->obj_asof == HAMMER_MAX_TID &&
1012             ip->obj_localization == 0 &&
1013             strncmp(ip->ino_data.ext.symlink, "@@PFS", 5) == 0) {
1014                     if (ip->pfsm->pfsd.mirror_flags & HAMMER_PFSD_SLAVE)
1015                             vap->va_size = 26;
1016                     else
1017                             vap->va_size = 10;
1018         }
1019
1020         /*
1021          * We must provide a consistent atime and mtime for snapshots
1022          * so people can do a 'tar cf - ... | md5' on them and get
1023          * consistent results.
1024          */
1025         if (ip->flags & HAMMER_INODE_RO) {
1026                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_atime);
1027                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_mtime);
1028         } else {
1029                 hammer_time_to_timespec(ip->ino_data.atime, &vap->va_atime);
1030                 hammer_time_to_timespec(ip->ino_data.mtime, &vap->va_mtime);
1031         }
1032         hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_ctime);
1033         vap->va_flags = ip->ino_data.uflags;
1034         vap->va_gen = 1;        /* hammer inums are unique for all time */
1035         vap->va_blocksize = HAMMER_BUFSIZE;
1036         if (ip->ino_data.size >= HAMMER_XDEMARC) {
1037                 vap->va_bytes = (ip->ino_data.size + HAMMER_XBUFMASK64) &
1038                                 ~HAMMER_XBUFMASK64;
1039         } else if (ip->ino_data.size > HAMMER_BUFSIZE / 2) {
1040                 vap->va_bytes = (ip->ino_data.size + HAMMER_BUFMASK64) &
1041                                 ~HAMMER_BUFMASK64;
1042         } else {
1043                 vap->va_bytes = (ip->ino_data.size + 15) & ~15;
1044         }
1045
1046         vap->va_type = hammer_get_vnode_type(ip->ino_data.obj_type);
1047         vap->va_filerev = 0;    /* XXX */
1048         vap->va_uid_uuid = ip->ino_data.uid;
1049         vap->va_gid_uuid = ip->ino_data.gid;
1050         vap->va_fsid_uuid = ip->hmp->fsid;
1051         vap->va_vaflags = VA_UID_UUID_VALID | VA_GID_UUID_VALID |
1052                           VA_FSID_UUID_VALID;
1053
1054         switch (ip->ino_data.obj_type) {
1055         case HAMMER_OBJTYPE_CDEV:
1056         case HAMMER_OBJTYPE_BDEV:
1057                 vap->va_rmajor = ip->ino_data.rmajor;
1058                 vap->va_rminor = ip->ino_data.rminor;
1059                 break;
1060         default:
1061                 break;
1062         }
1063         hammer_unlock(&ip->lock);
1064         return(0);
1065 }
1066
1067 /*
1068  * hammer_vop_nresolve { nch, dvp, cred }
1069  *
1070  * Locate the requested directory entry.
1071  */
1072 static
1073 int
1074 hammer_vop_nresolve(struct vop_nresolve_args *ap)
1075 {
1076         struct hammer_transaction trans;
1077         struct namecache *ncp;
1078         hammer_mount_t hmp;
1079         hammer_inode_t dip;
1080         hammer_inode_t ip;
1081         hammer_tid_t asof;
1082         struct hammer_cursor cursor;
1083         struct vnode *vp;
1084         int64_t namekey;
1085         int error;
1086         int i;
1087         int nlen;
1088         int flags;
1089         int ispfs;
1090         int64_t obj_id;
1091         u_int32_t localization;
1092         u_int32_t max_iterations;
1093
1094         /*
1095          * Misc initialization, plus handle as-of name extensions.  Look for
1096          * the '@@' extension.  Note that as-of files and directories cannot
1097          * be modified.
1098          */
1099         dip = VTOI(ap->a_dvp);
1100         ncp = ap->a_nch->ncp;
1101         asof = dip->obj_asof;
1102         localization = dip->obj_localization;   /* for code consistency */
1103         nlen = ncp->nc_nlen;
1104         flags = dip->flags & HAMMER_INODE_RO;
1105         ispfs = 0;
1106         hmp = dip->hmp;
1107
1108         lwkt_gettoken(&hmp->fs_token);
1109         hammer_simple_transaction(&trans, hmp);
1110         ++hammer_stats_file_iopsr;
1111
1112         for (i = 0; i < nlen; ++i) {
1113                 if (ncp->nc_name[i] == '@' && ncp->nc_name[i+1] == '@') {
1114                         error = hammer_str_to_tid(ncp->nc_name + i + 2,
1115                                                   &ispfs, &asof, &localization);
1116                         if (error != 0) {
1117                                 i = nlen;
1118                                 break;
1119                         }
1120                         if (asof != HAMMER_MAX_TID)
1121                                 flags |= HAMMER_INODE_RO;
1122                         break;
1123                 }
1124         }
1125         nlen = i;
1126
1127         /*
1128          * If this is a PFS softlink we dive into the PFS
1129          */
1130         if (ispfs && nlen == 0) {
1131                 ip = hammer_get_inode(&trans, dip, HAMMER_OBJID_ROOT,
1132                                       asof, localization,
1133                                       flags, &error);
1134                 if (error == 0) {
1135                         error = hammer_get_vnode(ip, &vp);
1136                         hammer_rel_inode(ip, 0);
1137                 } else {
1138                         vp = NULL;
1139                 }
1140                 if (error == 0) {
1141                         vn_unlock(vp);
1142                         cache_setvp(ap->a_nch, vp);
1143                         vrele(vp);
1144                 }
1145                 goto done;
1146         }
1147
1148         /*
1149          * If there is no path component the time extension is relative to dip.
1150          * e.g. "fubar/@@<snapshot>"
1151          *
1152          * "." is handled by the kernel, but ".@@<snapshot>" is not.
1153          * e.g. "fubar/.@@<snapshot>"
1154          *
1155          * ".." is handled by the kernel.  We do not currently handle
1156          * "..@<snapshot>".
1157          */
1158         if (nlen == 0 || (nlen == 1 && ncp->nc_name[0] == '.')) {
1159                 ip = hammer_get_inode(&trans, dip, dip->obj_id,
1160                                       asof, dip->obj_localization,
1161                                       flags, &error);
1162                 if (error == 0) {
1163                         error = hammer_get_vnode(ip, &vp);
1164                         hammer_rel_inode(ip, 0);
1165                 } else {
1166                         vp = NULL;
1167                 }
1168                 if (error == 0) {
1169                         vn_unlock(vp);
1170                         cache_setvp(ap->a_nch, vp);
1171                         vrele(vp);
1172                 }
1173                 goto done;
1174         }
1175
1176         /*
1177          * Calculate the namekey and setup the key range for the scan.  This
1178          * works kinda like a chained hash table where the lower 32 bits
1179          * of the namekey synthesize the chain.
1180          *
1181          * The key range is inclusive of both key_beg and key_end.
1182          */
1183         namekey = hammer_directory_namekey(dip, ncp->nc_name, nlen,
1184                                            &max_iterations);
1185
1186         error = hammer_init_cursor(&trans, &cursor, &dip->cache[1], dip);
1187         cursor.key_beg.localization = dip->obj_localization +
1188                                       hammer_dir_localization(dip);
1189         cursor.key_beg.obj_id = dip->obj_id;
1190         cursor.key_beg.key = namekey;
1191         cursor.key_beg.create_tid = 0;
1192         cursor.key_beg.delete_tid = 0;
1193         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1194         cursor.key_beg.obj_type = 0;
1195
1196         cursor.key_end = cursor.key_beg;
1197         cursor.key_end.key += max_iterations;
1198         cursor.asof = asof;
1199         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1200
1201         /*
1202          * Scan all matching records (the chain), locate the one matching
1203          * the requested path component.
1204          *
1205          * The hammer_ip_*() functions merge in-memory records with on-disk
1206          * records for the purposes of the search.
1207          */
1208         obj_id = 0;
1209         localization = HAMMER_DEF_LOCALIZATION;
1210
1211         if (error == 0) {
1212                 error = hammer_ip_first(&cursor);
1213                 while (error == 0) {
1214                         error = hammer_ip_resolve_data(&cursor);
1215                         if (error)
1216                                 break;
1217                         if (nlen == cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF &&
1218                             bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1219                                 obj_id = cursor.data->entry.obj_id;
1220                                 localization = cursor.data->entry.localization;
1221                                 break;
1222                         }
1223                         error = hammer_ip_next(&cursor);
1224                 }
1225         }
1226         hammer_done_cursor(&cursor);
1227
1228         /*
1229          * Lookup the obj_id.  This should always succeed.  If it does not
1230          * the filesystem may be damaged and we return a dummy inode.
1231          */
1232         if (error == 0) {
1233                 ip = hammer_get_inode(&trans, dip, obj_id,
1234                                       asof, localization,
1235                                       flags, &error);
1236                 if (error == ENOENT) {
1237                         kprintf("HAMMER: WARNING: Missing "
1238                                 "inode for dirent \"%s\"\n"
1239                                 "\tobj_id = %016llx, asof=%016llx, lo=%08x\n",
1240                                 ncp->nc_name,
1241                                 (long long)obj_id, (long long)asof,
1242                                 localization);
1243                         error = 0;
1244                         ip = hammer_get_dummy_inode(&trans, dip, obj_id,
1245                                                     asof, localization,
1246                                                     flags, &error);
1247                 }
1248                 if (error == 0) {
1249                         error = hammer_get_vnode(ip, &vp);
1250                         hammer_rel_inode(ip, 0);
1251                 } else {
1252                         vp = NULL;
1253                 }
1254                 if (error == 0) {
1255                         vn_unlock(vp);
1256                         cache_setvp(ap->a_nch, vp);
1257                         vrele(vp);
1258                 }
1259         } else if (error == ENOENT) {
1260                 cache_setvp(ap->a_nch, NULL);
1261         }
1262 done:
1263         hammer_done_transaction(&trans);
1264         lwkt_reltoken(&hmp->fs_token);
1265         return (error);
1266 }
1267
1268 /*
1269  * hammer_vop_nlookupdotdot { dvp, vpp, cred }
1270  *
1271  * Locate the parent directory of a directory vnode.
1272  *
1273  * dvp is referenced but not locked.  *vpp must be returned referenced and
1274  * locked.  A parent_obj_id of 0 does not necessarily indicate that we are
1275  * at the root, instead it could indicate that the directory we were in was
1276  * removed.
1277  *
1278  * NOTE: as-of sequences are not linked into the directory structure.  If
1279  * we are at the root with a different asof then the mount point, reload
1280  * the same directory with the mount point's asof.   I'm not sure what this
1281  * will do to NFS.  We encode ASOF stamps in NFS file handles so it might not
1282  * get confused, but it hasn't been tested.
1283  */
1284 static
1285 int
1286 hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *ap)
1287 {
1288         struct hammer_transaction trans;
1289         struct hammer_inode *dip;
1290         struct hammer_inode *ip;
1291         hammer_mount_t hmp;
1292         int64_t parent_obj_id;
1293         u_int32_t parent_obj_localization;
1294         hammer_tid_t asof;
1295         int error;
1296
1297         dip = VTOI(ap->a_dvp);
1298         asof = dip->obj_asof;
1299         hmp = dip->hmp;
1300
1301         /*
1302          * Whos are parent?  This could be the root of a pseudo-filesystem
1303          * whos parent is in another localization domain.
1304          */
1305         lwkt_gettoken(&hmp->fs_token);
1306         parent_obj_id = dip->ino_data.parent_obj_id;
1307         if (dip->obj_id == HAMMER_OBJID_ROOT)
1308                 parent_obj_localization = dip->ino_data.ext.obj.parent_obj_localization;
1309         else
1310                 parent_obj_localization = dip->obj_localization;
1311
1312         if (parent_obj_id == 0) {
1313                 if (dip->obj_id == HAMMER_OBJID_ROOT &&
1314                    asof != hmp->asof) {
1315                         parent_obj_id = dip->obj_id;
1316                         asof = hmp->asof;
1317                         *ap->a_fakename = kmalloc(19, M_TEMP, M_WAITOK);
1318                         ksnprintf(*ap->a_fakename, 19, "0x%016llx",
1319                                   (long long)dip->obj_asof);
1320                 } else {
1321                         *ap->a_vpp = NULL;
1322                         lwkt_reltoken(&hmp->fs_token);
1323                         return ENOENT;
1324                 }
1325         }
1326
1327         hammer_simple_transaction(&trans, hmp);
1328         ++hammer_stats_file_iopsr;
1329
1330         ip = hammer_get_inode(&trans, dip, parent_obj_id,
1331                               asof, parent_obj_localization,
1332                               dip->flags, &error);
1333         if (ip) {
1334                 error = hammer_get_vnode(ip, ap->a_vpp);
1335                 hammer_rel_inode(ip, 0);
1336         } else {
1337                 *ap->a_vpp = NULL;
1338         }
1339         hammer_done_transaction(&trans);
1340         lwkt_reltoken(&hmp->fs_token);
1341         return (error);
1342 }
1343
1344 /*
1345  * hammer_vop_nlink { nch, dvp, vp, cred }
1346  */
1347 static
1348 int
1349 hammer_vop_nlink(struct vop_nlink_args *ap)
1350 {
1351         struct hammer_transaction trans;
1352         struct hammer_inode *dip;
1353         struct hammer_inode *ip;
1354         struct nchandle *nch;
1355         hammer_mount_t hmp;
1356         int error;
1357
1358         if (ap->a_dvp->v_mount != ap->a_vp->v_mount)
1359                 return(EXDEV);
1360
1361         nch = ap->a_nch;
1362         dip = VTOI(ap->a_dvp);
1363         ip = VTOI(ap->a_vp);
1364         hmp = dip->hmp;
1365
1366         if (dip->obj_localization != ip->obj_localization)
1367                 return(EXDEV);
1368
1369         if (dip->flags & HAMMER_INODE_RO)
1370                 return (EROFS);
1371         if (ip->flags & HAMMER_INODE_RO)
1372                 return (EROFS);
1373         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1374                 return (error);
1375
1376         /*
1377          * Create a transaction to cover the operations we perform.
1378          */
1379         lwkt_gettoken(&hmp->fs_token);
1380         hammer_start_transaction(&trans, hmp);
1381         ++hammer_stats_file_iopsw;
1382
1383         /*
1384          * Add the filesystem object to the directory.  Note that neither
1385          * dip nor ip are referenced or locked, but their vnodes are
1386          * referenced.  This function will bump the inode's link count.
1387          */
1388         error = hammer_ip_add_directory(&trans, dip,
1389                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1390                                         ip);
1391
1392         /*
1393          * Finish up.
1394          */
1395         if (error == 0) {
1396                 cache_setunresolved(nch);
1397                 cache_setvp(nch, ap->a_vp);
1398         }
1399         hammer_done_transaction(&trans);
1400         hammer_knote(ap->a_vp, NOTE_LINK);
1401         hammer_knote(ap->a_dvp, NOTE_WRITE);
1402         lwkt_reltoken(&hmp->fs_token);
1403         return (error);
1404 }
1405
1406 /*
1407  * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
1408  *
1409  * The operating system has already ensured that the directory entry
1410  * does not exist and done all appropriate namespace locking.
1411  */
1412 static
1413 int
1414 hammer_vop_nmkdir(struct vop_nmkdir_args *ap)
1415 {
1416         struct hammer_transaction trans;
1417         struct hammer_inode *dip;
1418         struct hammer_inode *nip;
1419         struct nchandle *nch;
1420         hammer_mount_t hmp;
1421         int error;
1422
1423         nch = ap->a_nch;
1424         dip = VTOI(ap->a_dvp);
1425         hmp = dip->hmp;
1426
1427         if (dip->flags & HAMMER_INODE_RO)
1428                 return (EROFS);
1429         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1430                 return (error);
1431
1432         /*
1433          * Create a transaction to cover the operations we perform.
1434          */
1435         lwkt_gettoken(&hmp->fs_token);
1436         hammer_start_transaction(&trans, hmp);
1437         ++hammer_stats_file_iopsw;
1438
1439         /*
1440          * Create a new filesystem object of the requested type.  The
1441          * returned inode will be referenced but not locked.
1442          */
1443         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1444                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1445                                     NULL, &nip);
1446         if (error) {
1447                 hkprintf("hammer_mkdir error %d\n", error);
1448                 hammer_done_transaction(&trans);
1449                 *ap->a_vpp = NULL;
1450                 lwkt_reltoken(&hmp->fs_token);
1451                 return (error);
1452         }
1453         /*
1454          * Add the new filesystem object to the directory.  This will also
1455          * bump the inode's link count.
1456          */
1457         error = hammer_ip_add_directory(&trans, dip,
1458                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1459                                         nip);
1460         if (error)
1461                 hkprintf("hammer_mkdir (add) error %d\n", error);
1462
1463         /*
1464          * Finish up.
1465          */
1466         if (error) {
1467                 hammer_rel_inode(nip, 0);
1468                 *ap->a_vpp = NULL;
1469         } else {
1470                 error = hammer_get_vnode(nip, ap->a_vpp);
1471                 hammer_rel_inode(nip, 0);
1472                 if (error == 0) {
1473                         cache_setunresolved(ap->a_nch);
1474                         cache_setvp(ap->a_nch, *ap->a_vpp);
1475                 }
1476         }
1477         hammer_done_transaction(&trans);
1478         if (error == 0)
1479                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
1480         lwkt_reltoken(&hmp->fs_token);
1481         return (error);
1482 }
1483
1484 /*
1485  * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1486  *
1487  * The operating system has already ensured that the directory entry
1488  * does not exist and done all appropriate namespace locking.
1489  */
1490 static
1491 int
1492 hammer_vop_nmknod(struct vop_nmknod_args *ap)
1493 {
1494         struct hammer_transaction trans;
1495         struct hammer_inode *dip;
1496         struct hammer_inode *nip;
1497         struct nchandle *nch;
1498         hammer_mount_t hmp;
1499         int error;
1500
1501         nch = ap->a_nch;
1502         dip = VTOI(ap->a_dvp);
1503         hmp = dip->hmp;
1504
1505         if (dip->flags & HAMMER_INODE_RO)
1506                 return (EROFS);
1507         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1508                 return (error);
1509
1510         /*
1511          * Create a transaction to cover the operations we perform.
1512          */
1513         lwkt_gettoken(&hmp->fs_token);
1514         hammer_start_transaction(&trans, hmp);
1515         ++hammer_stats_file_iopsw;
1516
1517         /*
1518          * Create a new filesystem object of the requested type.  The
1519          * returned inode will be referenced but not locked.
1520          *
1521          * If mknod specifies a directory a pseudo-fs is created.
1522          */
1523         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1524                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1525                                     NULL, &nip);
1526         if (error) {
1527                 hammer_done_transaction(&trans);
1528                 *ap->a_vpp = NULL;
1529                 lwkt_reltoken(&hmp->fs_token);
1530                 return (error);
1531         }
1532
1533         /*
1534          * Add the new filesystem object to the directory.  This will also
1535          * bump the inode's link count.
1536          */
1537         error = hammer_ip_add_directory(&trans, dip,
1538                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1539                                         nip);
1540
1541         /*
1542          * Finish up.
1543          */
1544         if (error) {
1545                 hammer_rel_inode(nip, 0);
1546                 *ap->a_vpp = NULL;
1547         } else {
1548                 error = hammer_get_vnode(nip, ap->a_vpp);
1549                 hammer_rel_inode(nip, 0);
1550                 if (error == 0) {
1551                         cache_setunresolved(ap->a_nch);
1552                         cache_setvp(ap->a_nch, *ap->a_vpp);
1553                 }
1554         }
1555         hammer_done_transaction(&trans);
1556         if (error == 0)
1557                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1558         lwkt_reltoken(&hmp->fs_token);
1559         return (error);
1560 }
1561
1562 /*
1563  * hammer_vop_open { vp, mode, cred, fp }
1564  *
1565  * MPSAFE (does not require fs_token)
1566  */
1567 static
1568 int
1569 hammer_vop_open(struct vop_open_args *ap)
1570 {
1571         hammer_inode_t ip;
1572
1573         ++hammer_stats_file_iopsr;
1574         ip = VTOI(ap->a_vp);
1575
1576         if ((ap->a_mode & FWRITE) && (ip->flags & HAMMER_INODE_RO))
1577                 return (EROFS);
1578         return(vop_stdopen(ap));
1579 }
1580
1581 /*
1582  * hammer_vop_print { vp }
1583  */
1584 static
1585 int
1586 hammer_vop_print(struct vop_print_args *ap)
1587 {
1588         return EOPNOTSUPP;
1589 }
1590
1591 /*
1592  * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1593  */
1594 static
1595 int
1596 hammer_vop_readdir(struct vop_readdir_args *ap)
1597 {
1598         struct hammer_transaction trans;
1599         struct hammer_cursor cursor;
1600         struct hammer_inode *ip;
1601         hammer_mount_t hmp;
1602         struct uio *uio;
1603         hammer_base_elm_t base;
1604         int error;
1605         int cookie_index;
1606         int ncookies;
1607         off_t *cookies;
1608         off_t saveoff;
1609         int r;
1610         int dtype;
1611
1612         ++hammer_stats_file_iopsr;
1613         ip = VTOI(ap->a_vp);
1614         uio = ap->a_uio;
1615         saveoff = uio->uio_offset;
1616         hmp = ip->hmp;
1617
1618         if (ap->a_ncookies) {
1619                 ncookies = uio->uio_resid / 16 + 1;
1620                 if (ncookies > 1024)
1621                         ncookies = 1024;
1622                 cookies = kmalloc(ncookies * sizeof(off_t), M_TEMP, M_WAITOK);
1623                 cookie_index = 0;
1624         } else {
1625                 ncookies = -1;
1626                 cookies = NULL;
1627                 cookie_index = 0;
1628         }
1629
1630         lwkt_gettoken(&hmp->fs_token);
1631         hammer_simple_transaction(&trans, hmp);
1632
1633         /*
1634          * Handle artificial entries
1635          *
1636          * It should be noted that the minimum value for a directory
1637          * hash key on-media is 0x0000000100000000, so we can use anything
1638          * less then that to represent our 'special' key space.
1639          */
1640         error = 0;
1641         if (saveoff == 0) {
1642                 r = vop_write_dirent(&error, uio, ip->obj_id, DT_DIR, 1, ".");
1643                 if (r)
1644                         goto done;
1645                 if (cookies)
1646                         cookies[cookie_index] = saveoff;
1647                 ++saveoff;
1648                 ++cookie_index;
1649                 if (cookie_index == ncookies)
1650                         goto done;
1651         }
1652         if (saveoff == 1) {
1653                 if (ip->ino_data.parent_obj_id) {
1654                         r = vop_write_dirent(&error, uio,
1655                                              ip->ino_data.parent_obj_id,
1656                                              DT_DIR, 2, "..");
1657                 } else {
1658                         r = vop_write_dirent(&error, uio,
1659                                              ip->obj_id, DT_DIR, 2, "..");
1660                 }
1661                 if (r)
1662                         goto done;
1663                 if (cookies)
1664                         cookies[cookie_index] = saveoff;
1665                 ++saveoff;
1666                 ++cookie_index;
1667                 if (cookie_index == ncookies)
1668                         goto done;
1669         }
1670
1671         /*
1672          * Key range (begin and end inclusive) to scan.  Directory keys
1673          * directly translate to a 64 bit 'seek' position.
1674          */
1675         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1676         cursor.key_beg.localization = ip->obj_localization +
1677                                       hammer_dir_localization(ip);
1678         cursor.key_beg.obj_id = ip->obj_id;
1679         cursor.key_beg.create_tid = 0;
1680         cursor.key_beg.delete_tid = 0;
1681         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1682         cursor.key_beg.obj_type = 0;
1683         cursor.key_beg.key = saveoff;
1684
1685         cursor.key_end = cursor.key_beg;
1686         cursor.key_end.key = HAMMER_MAX_KEY;
1687         cursor.asof = ip->obj_asof;
1688         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1689
1690         error = hammer_ip_first(&cursor);
1691
1692         while (error == 0) {
1693                 error = hammer_ip_resolve_data(&cursor);
1694                 if (error)
1695                         break;
1696                 base = &cursor.leaf->base;
1697                 saveoff = base->key;
1698                 KKASSERT(cursor.leaf->data_len > HAMMER_ENTRY_NAME_OFF);
1699
1700                 if (base->obj_id != ip->obj_id)
1701                         panic("readdir: bad record at %p", cursor.node);
1702
1703                 /*
1704                  * Convert pseudo-filesystems into softlinks
1705                  */
1706                 dtype = hammer_get_dtype(cursor.leaf->base.obj_type);
1707                 r = vop_write_dirent(
1708                              &error, uio, cursor.data->entry.obj_id,
1709                              dtype,
1710                              cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF ,
1711                              (void *)cursor.data->entry.name);
1712                 if (r)
1713                         break;
1714                 ++saveoff;
1715                 if (cookies)
1716                         cookies[cookie_index] = base->key;
1717                 ++cookie_index;
1718                 if (cookie_index == ncookies)
1719                         break;
1720                 error = hammer_ip_next(&cursor);
1721         }
1722         hammer_done_cursor(&cursor);
1723
1724 done:
1725         hammer_done_transaction(&trans);
1726
1727         if (ap->a_eofflag)
1728                 *ap->a_eofflag = (error == ENOENT);
1729         uio->uio_offset = saveoff;
1730         if (error && cookie_index == 0) {
1731                 if (error == ENOENT)
1732                         error = 0;
1733                 if (cookies) {
1734                         kfree(cookies, M_TEMP);
1735                         *ap->a_ncookies = 0;
1736                         *ap->a_cookies = NULL;
1737                 }
1738         } else {
1739                 if (error == ENOENT)
1740                         error = 0;
1741                 if (cookies) {
1742                         *ap->a_ncookies = cookie_index;
1743                         *ap->a_cookies = cookies;
1744                 }
1745         }
1746         lwkt_reltoken(&hmp->fs_token);
1747         return(error);
1748 }
1749
1750 /*
1751  * hammer_vop_readlink { vp, uio, cred }
1752  */
1753 static
1754 int
1755 hammer_vop_readlink(struct vop_readlink_args *ap)
1756 {
1757         struct hammer_transaction trans;
1758         struct hammer_cursor cursor;
1759         struct hammer_inode *ip;
1760         hammer_mount_t hmp;
1761         char buf[32];
1762         u_int32_t localization;
1763         hammer_pseudofs_inmem_t pfsm;
1764         int error;
1765
1766         ip = VTOI(ap->a_vp);
1767         hmp = ip->hmp;
1768
1769         lwkt_gettoken(&hmp->fs_token);
1770
1771         /*
1772          * Shortcut if the symlink data was stuffed into ino_data.
1773          *
1774          * Also expand special "@@PFS%05d" softlinks (expansion only
1775          * occurs for non-historical (current) accesses made from the
1776          * primary filesystem).
1777          */
1778         if (ip->ino_data.size <= HAMMER_INODE_BASESYMLEN) {
1779                 char *ptr;
1780                 int bytes;
1781
1782                 ptr = ip->ino_data.ext.symlink;
1783                 bytes = (int)ip->ino_data.size;
1784                 if (bytes == 10 &&
1785                     ip->obj_asof == HAMMER_MAX_TID &&
1786                     ip->obj_localization == 0 &&
1787                     strncmp(ptr, "@@PFS", 5) == 0) {
1788                         hammer_simple_transaction(&trans, hmp);
1789                         bcopy(ptr + 5, buf, 5);
1790                         buf[5] = 0;
1791                         localization = strtoul(buf, NULL, 10) << 16;
1792                         pfsm = hammer_load_pseudofs(&trans, localization,
1793                                                     &error);
1794                         if (error == 0) {
1795                                 if (pfsm->pfsd.mirror_flags &
1796                                     HAMMER_PFSD_SLAVE) {
1797                                         /* vap->va_size == 26 */
1798                                         ksnprintf(buf, sizeof(buf),
1799                                                   "@@0x%016llx:%05d",
1800                                                   (long long)pfsm->pfsd.sync_end_tid,
1801                                                   localization >> 16);
1802                                 } else {
1803                                         /* vap->va_size == 10 */
1804                                         ksnprintf(buf, sizeof(buf),
1805                                                   "@@-1:%05d",
1806                                                   localization >> 16);
1807 #if 0
1808                                         ksnprintf(buf, sizeof(buf),
1809                                                   "@@0x%016llx:%05d",
1810                                                   (long long)HAMMER_MAX_TID,
1811                                                   localization >> 16);
1812 #endif
1813                                 }
1814                                 ptr = buf;
1815                                 bytes = strlen(buf);
1816                         }
1817                         if (pfsm)
1818                                 hammer_rel_pseudofs(hmp, pfsm);
1819                         hammer_done_transaction(&trans);
1820                 }
1821                 error = uiomove(ptr, bytes, ap->a_uio);
1822                 lwkt_reltoken(&hmp->fs_token);
1823                 return(error);
1824         }
1825
1826         /*
1827          * Long version
1828          */
1829         hammer_simple_transaction(&trans, hmp);
1830         ++hammer_stats_file_iopsr;
1831         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1832
1833         /*
1834          * Key range (begin and end inclusive) to scan.  Directory keys
1835          * directly translate to a 64 bit 'seek' position.
1836          */
1837         cursor.key_beg.localization = ip->obj_localization +
1838                                       HAMMER_LOCALIZE_MISC;
1839         cursor.key_beg.obj_id = ip->obj_id;
1840         cursor.key_beg.create_tid = 0;
1841         cursor.key_beg.delete_tid = 0;
1842         cursor.key_beg.rec_type = HAMMER_RECTYPE_FIX;
1843         cursor.key_beg.obj_type = 0;
1844         cursor.key_beg.key = HAMMER_FIXKEY_SYMLINK;
1845         cursor.asof = ip->obj_asof;
1846         cursor.flags |= HAMMER_CURSOR_ASOF;
1847
1848         error = hammer_ip_lookup(&cursor);
1849         if (error == 0) {
1850                 error = hammer_ip_resolve_data(&cursor);
1851                 if (error == 0) {
1852                         KKASSERT(cursor.leaf->data_len >=
1853                                  HAMMER_SYMLINK_NAME_OFF);
1854                         error = uiomove(cursor.data->symlink.name,
1855                                         cursor.leaf->data_len -
1856                                                 HAMMER_SYMLINK_NAME_OFF,
1857                                         ap->a_uio);
1858                 }
1859         }
1860         hammer_done_cursor(&cursor);
1861         hammer_done_transaction(&trans);
1862         lwkt_reltoken(&hmp->fs_token);
1863         return(error);
1864 }
1865
1866 /*
1867  * hammer_vop_nremove { nch, dvp, cred }
1868  */
1869 static
1870 int
1871 hammer_vop_nremove(struct vop_nremove_args *ap)
1872 {
1873         struct hammer_transaction trans;
1874         struct hammer_inode *dip;
1875         hammer_mount_t hmp;
1876         int error;
1877
1878         dip = VTOI(ap->a_dvp);
1879         hmp = dip->hmp;
1880
1881         if (hammer_nohistory(dip) == 0 &&
1882             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1883                 return (error);
1884         }
1885
1886         lwkt_gettoken(&hmp->fs_token);
1887         hammer_start_transaction(&trans, hmp);
1888         ++hammer_stats_file_iopsw;
1889         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 0);
1890         hammer_done_transaction(&trans);
1891         if (error == 0)
1892                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1893         lwkt_reltoken(&hmp->fs_token);
1894         return (error);
1895 }
1896
1897 /*
1898  * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1899  */
1900 static
1901 int
1902 hammer_vop_nrename(struct vop_nrename_args *ap)
1903 {
1904         struct hammer_transaction trans;
1905         struct namecache *fncp;
1906         struct namecache *tncp;
1907         struct hammer_inode *fdip;
1908         struct hammer_inode *tdip;
1909         struct hammer_inode *ip;
1910         hammer_mount_t hmp;
1911         struct hammer_cursor cursor;
1912         int64_t namekey;
1913         u_int32_t max_iterations;
1914         int nlen, error;
1915
1916         if (ap->a_fdvp->v_mount != ap->a_tdvp->v_mount)
1917                 return(EXDEV);
1918         if (ap->a_fdvp->v_mount != ap->a_fnch->ncp->nc_vp->v_mount)
1919                 return(EXDEV);
1920
1921         fdip = VTOI(ap->a_fdvp);
1922         tdip = VTOI(ap->a_tdvp);
1923         fncp = ap->a_fnch->ncp;
1924         tncp = ap->a_tnch->ncp;
1925         ip = VTOI(fncp->nc_vp);
1926         KKASSERT(ip != NULL);
1927
1928         hmp = ip->hmp;
1929
1930         if (fdip->obj_localization != tdip->obj_localization)
1931                 return(EXDEV);
1932         if (fdip->obj_localization != ip->obj_localization)
1933                 return(EXDEV);
1934
1935         if (fdip->flags & HAMMER_INODE_RO)
1936                 return (EROFS);
1937         if (tdip->flags & HAMMER_INODE_RO)
1938                 return (EROFS);
1939         if (ip->flags & HAMMER_INODE_RO)
1940                 return (EROFS);
1941         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1942                 return (error);
1943
1944         lwkt_gettoken(&hmp->fs_token);
1945         hammer_start_transaction(&trans, hmp);
1946         ++hammer_stats_file_iopsw;
1947
1948         /*
1949          * Remove tncp from the target directory and then link ip as
1950          * tncp. XXX pass trans to dounlink
1951          *
1952          * Force the inode sync-time to match the transaction so it is
1953          * in-sync with the creation of the target directory entry.
1954          */
1955         error = hammer_dounlink(&trans, ap->a_tnch, ap->a_tdvp,
1956                                 ap->a_cred, 0, -1);
1957         if (error == 0 || error == ENOENT) {
1958                 error = hammer_ip_add_directory(&trans, tdip,
1959                                                 tncp->nc_name, tncp->nc_nlen,
1960                                                 ip);
1961                 if (error == 0) {
1962                         ip->ino_data.parent_obj_id = tdip->obj_id;
1963                         ip->ino_data.ctime = trans.time;
1964                         hammer_modify_inode(&trans, ip, HAMMER_INODE_DDIRTY);
1965                 }
1966         }
1967         if (error)
1968                 goto failed; /* XXX */
1969
1970         /*
1971          * Locate the record in the originating directory and remove it.
1972          *
1973          * Calculate the namekey and setup the key range for the scan.  This
1974          * works kinda like a chained hash table where the lower 32 bits
1975          * of the namekey synthesize the chain.
1976          *
1977          * The key range is inclusive of both key_beg and key_end.
1978          */
1979         namekey = hammer_directory_namekey(fdip, fncp->nc_name, fncp->nc_nlen,
1980                                            &max_iterations);
1981 retry:
1982         hammer_init_cursor(&trans, &cursor, &fdip->cache[1], fdip);
1983         cursor.key_beg.localization = fdip->obj_localization +
1984                                       hammer_dir_localization(fdip);
1985         cursor.key_beg.obj_id = fdip->obj_id;
1986         cursor.key_beg.key = namekey;
1987         cursor.key_beg.create_tid = 0;
1988         cursor.key_beg.delete_tid = 0;
1989         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1990         cursor.key_beg.obj_type = 0;
1991
1992         cursor.key_end = cursor.key_beg;
1993         cursor.key_end.key += max_iterations;
1994         cursor.asof = fdip->obj_asof;
1995         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1996
1997         /*
1998          * Scan all matching records (the chain), locate the one matching
1999          * the requested path component.
2000          *
2001          * The hammer_ip_*() functions merge in-memory records with on-disk
2002          * records for the purposes of the search.
2003          */
2004         error = hammer_ip_first(&cursor);
2005         while (error == 0) {
2006                 if (hammer_ip_resolve_data(&cursor) != 0)
2007                         break;
2008                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
2009                 KKASSERT(nlen > 0);
2010                 if (fncp->nc_nlen == nlen &&
2011                     bcmp(fncp->nc_name, cursor.data->entry.name, nlen) == 0) {
2012                         break;
2013                 }
2014                 error = hammer_ip_next(&cursor);
2015         }
2016
2017         /*
2018          * If all is ok we have to get the inode so we can adjust nlinks.
2019          *
2020          * WARNING: hammer_ip_del_directory() may have to terminate the
2021          * cursor to avoid a recursion.  It's ok to call hammer_done_cursor()
2022          * twice.
2023          */
2024         if (error == 0)
2025                 error = hammer_ip_del_directory(&trans, &cursor, fdip, ip);
2026
2027         /*
2028          * XXX A deadlock here will break rename's atomicy for the purposes
2029          * of crash recovery.
2030          */
2031         if (error == EDEADLK) {
2032                 hammer_done_cursor(&cursor);
2033                 goto retry;
2034         }
2035
2036         /*
2037          * Cleanup and tell the kernel that the rename succeeded.
2038          *
2039          * NOTE: ip->vp, if non-NULL, cannot be directly referenced
2040          *       without formally acquiring the vp since the vp might
2041          *       have zero refs on it, or in the middle of a reclaim,
2042          *       etc.
2043          */
2044         hammer_done_cursor(&cursor);
2045         if (error == 0) {
2046                 cache_rename(ap->a_fnch, ap->a_tnch);
2047                 hammer_knote(ap->a_fdvp, NOTE_WRITE);
2048                 hammer_knote(ap->a_tdvp, NOTE_WRITE);
2049                 while (ip->vp) {
2050                         struct vnode *vp;
2051
2052                         error = hammer_get_vnode(ip, &vp);
2053                         if (error == 0 && vp) {
2054                                 vn_unlock(vp);
2055                                 hammer_knote(ip->vp, NOTE_RENAME);
2056                                 vrele(vp);
2057                                 break;
2058                         }
2059                         kprintf("Debug: HAMMER ip/vp race2 avoided\n");
2060                 }
2061         }
2062
2063 failed:
2064         hammer_done_transaction(&trans);
2065         lwkt_reltoken(&hmp->fs_token);
2066         return (error);
2067 }
2068
2069 /*
2070  * hammer_vop_nrmdir { nch, dvp, cred }
2071  */
2072 static
2073 int
2074 hammer_vop_nrmdir(struct vop_nrmdir_args *ap)
2075 {
2076         struct hammer_transaction trans;
2077         struct hammer_inode *dip;
2078         hammer_mount_t hmp;
2079         int error;
2080
2081         dip = VTOI(ap->a_dvp);
2082         hmp = dip->hmp;
2083
2084         if (hammer_nohistory(dip) == 0 &&
2085             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2086                 return (error);
2087         }
2088
2089         lwkt_gettoken(&hmp->fs_token);
2090         hammer_start_transaction(&trans, hmp);
2091         ++hammer_stats_file_iopsw;
2092         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 1);
2093         hammer_done_transaction(&trans);
2094         if (error == 0)
2095                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
2096         lwkt_reltoken(&hmp->fs_token);
2097         return (error);
2098 }
2099
2100 /*
2101  * hammer_vop_markatime { vp, cred }
2102  */
2103 static
2104 int
2105 hammer_vop_markatime(struct vop_markatime_args *ap)
2106 {
2107         struct hammer_transaction trans;
2108         struct hammer_inode *ip;
2109         hammer_mount_t hmp;
2110
2111         ip = VTOI(ap->a_vp);
2112         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2113                 return (EROFS);
2114         if (ip->flags & HAMMER_INODE_RO)
2115                 return (EROFS);
2116         hmp = ip->hmp;
2117         if (hmp->mp->mnt_flag & MNT_NOATIME)
2118                 return (0);
2119         lwkt_gettoken(&hmp->fs_token);
2120         hammer_start_transaction(&trans, hmp);
2121         ++hammer_stats_file_iopsw;
2122
2123         ip->ino_data.atime = trans.time;
2124         hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
2125         hammer_done_transaction(&trans);
2126         hammer_knote(ap->a_vp, NOTE_ATTRIB);
2127         lwkt_reltoken(&hmp->fs_token);
2128         return (0);
2129 }
2130
2131 /*
2132  * hammer_vop_setattr { vp, vap, cred }
2133  */
2134 static
2135 int
2136 hammer_vop_setattr(struct vop_setattr_args *ap)
2137 {
2138         struct hammer_transaction trans;
2139         struct hammer_inode *ip;
2140         struct vattr *vap;
2141         hammer_mount_t hmp;
2142         int modflags;
2143         int error;
2144         int truncating;
2145         int blksize;
2146         int kflags;
2147 #if 0
2148         int64_t aligned_size;
2149 #endif
2150         u_int32_t flags;
2151
2152         vap = ap->a_vap;
2153         ip = ap->a_vp->v_data;
2154         modflags = 0;
2155         kflags = 0;
2156         hmp = ip->hmp;
2157
2158         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2159                 return(EROFS);
2160         if (ip->flags & HAMMER_INODE_RO)
2161                 return (EROFS);
2162         if (hammer_nohistory(ip) == 0 &&
2163             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2164                 return (error);
2165         }
2166
2167         lwkt_gettoken(&hmp->fs_token);
2168         hammer_start_transaction(&trans, hmp);
2169         ++hammer_stats_file_iopsw;
2170         error = 0;
2171
2172         if (vap->va_flags != VNOVAL) {
2173                 flags = ip->ino_data.uflags;
2174                 error = vop_helper_setattr_flags(&flags, vap->va_flags,
2175                                          hammer_to_unix_xid(&ip->ino_data.uid),
2176                                          ap->a_cred);
2177                 if (error == 0) {
2178                         if (ip->ino_data.uflags != flags) {
2179                                 ip->ino_data.uflags = flags;
2180                                 ip->ino_data.ctime = trans.time;
2181                                 modflags |= HAMMER_INODE_DDIRTY;
2182                                 kflags |= NOTE_ATTRIB;
2183                         }
2184                         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2185                                 error = 0;
2186                                 goto done;
2187                         }
2188                 }
2189                 goto done;
2190         }
2191         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2192                 error = EPERM;
2193                 goto done;
2194         }
2195         if (vap->va_uid != (uid_t)VNOVAL || vap->va_gid != (gid_t)VNOVAL) {
2196                 mode_t cur_mode = ip->ino_data.mode;
2197                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2198                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2199                 uuid_t uuid_uid;
2200                 uuid_t uuid_gid;
2201
2202                 error = vop_helper_chown(ap->a_vp, vap->va_uid, vap->va_gid,
2203                                          ap->a_cred,
2204                                          &cur_uid, &cur_gid, &cur_mode);
2205                 if (error == 0) {
2206                         hammer_guid_to_uuid(&uuid_uid, cur_uid);
2207                         hammer_guid_to_uuid(&uuid_gid, cur_gid);
2208                         if (bcmp(&uuid_uid, &ip->ino_data.uid,
2209                                  sizeof(uuid_uid)) ||
2210                             bcmp(&uuid_gid, &ip->ino_data.gid,
2211                                  sizeof(uuid_gid)) ||
2212                             ip->ino_data.mode != cur_mode
2213                         ) {
2214                                 ip->ino_data.uid = uuid_uid;
2215                                 ip->ino_data.gid = uuid_gid;
2216                                 ip->ino_data.mode = cur_mode;
2217                                 ip->ino_data.ctime = trans.time;
2218                                 modflags |= HAMMER_INODE_DDIRTY;
2219                         }
2220                         kflags |= NOTE_ATTRIB;
2221                 }
2222         }
2223         while (vap->va_size != VNOVAL && ip->ino_data.size != vap->va_size) {
2224                 switch(ap->a_vp->v_type) {
2225                 case VREG:
2226                         if (vap->va_size == ip->ino_data.size)
2227                                 break;
2228
2229                         /*
2230                          * Log the operation if in fast-fsync mode or if
2231                          * there are unterminated redo write records present.
2232                          *
2233                          * The second check is needed so the recovery code
2234                          * properly truncates write redos even if nominal
2235                          * REDO operations is turned off due to excessive
2236                          * writes, because the related records might be
2237                          * destroyed and never lay down a TERM_WRITE.
2238                          */
2239                         if ((ip->flags & HAMMER_INODE_REDO) ||
2240                             (ip->flags & HAMMER_INODE_RDIRTY)) {
2241                                 error = hammer_generate_redo(&trans, ip,
2242                                                              vap->va_size,
2243                                                              HAMMER_REDO_TRUNC,
2244                                                              NULL, 0);
2245                         }
2246                         blksize = hammer_blocksize(vap->va_size);
2247
2248                         /*
2249                          * XXX break atomicy, we can deadlock the backend
2250                          * if we do not release the lock.  Probably not a
2251                          * big deal here.
2252                          */
2253                         if (vap->va_size < ip->ino_data.size) {
2254                                 nvtruncbuf(ap->a_vp, vap->va_size,
2255                                            blksize,
2256                                            hammer_blockoff(vap->va_size));
2257                                 truncating = 1;
2258                                 kflags |= NOTE_WRITE;
2259                         } else {
2260                                 nvextendbuf(ap->a_vp,
2261                                             ip->ino_data.size,
2262                                             vap->va_size,
2263                                             hammer_blocksize(ip->ino_data.size),
2264                                             hammer_blocksize(vap->va_size),
2265                                             hammer_blockoff(ip->ino_data.size),
2266                                             hammer_blockoff(vap->va_size),
2267                                             0);
2268                                 truncating = 0;
2269                                 kflags |= NOTE_WRITE | NOTE_EXTEND;
2270                         }
2271                         ip->ino_data.size = vap->va_size;
2272                         ip->ino_data.mtime = trans.time;
2273                         /* XXX safe to use SDIRTY instead of DDIRTY here? */
2274                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2275
2276                         /*
2277                          * On-media truncation is cached in the inode until
2278                          * the inode is synchronized.  We must immediately
2279                          * handle any frontend records.
2280                          */
2281                         if (truncating) {
2282                                 hammer_ip_frontend_trunc(ip, vap->va_size);
2283 #ifdef DEBUG_TRUNCATE
2284                                 if (HammerTruncIp == NULL)
2285                                         HammerTruncIp = ip;
2286 #endif
2287                                 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2288                                         ip->flags |= HAMMER_INODE_TRUNCATED;
2289                                         ip->trunc_off = vap->va_size;
2290 #ifdef DEBUG_TRUNCATE
2291                                         if (ip == HammerTruncIp)
2292                                         kprintf("truncate1 %016llx\n",
2293                                                 (long long)ip->trunc_off);
2294 #endif
2295                                 } else if (ip->trunc_off > vap->va_size) {
2296                                         ip->trunc_off = vap->va_size;
2297 #ifdef DEBUG_TRUNCATE
2298                                         if (ip == HammerTruncIp)
2299                                         kprintf("truncate2 %016llx\n",
2300                                                 (long long)ip->trunc_off);
2301 #endif
2302                                 } else {
2303 #ifdef DEBUG_TRUNCATE
2304                                         if (ip == HammerTruncIp)
2305                                         kprintf("truncate3 %016llx (ignored)\n",
2306                                                 (long long)vap->va_size);
2307 #endif
2308                                 }
2309                         }
2310
2311 #if 0
2312                         /*
2313                          * When truncating, nvtruncbuf() may have cleaned out
2314                          * a portion of the last block on-disk in the buffer
2315                          * cache.  We must clean out any frontend records
2316                          * for blocks beyond the new last block.
2317                          */
2318                         aligned_size = (vap->va_size + (blksize - 1)) &
2319                                        ~(int64_t)(blksize - 1);
2320                         if (truncating && vap->va_size < aligned_size) {
2321                                 aligned_size -= blksize;
2322                                 hammer_ip_frontend_trunc(ip, aligned_size);
2323                         }
2324 #endif
2325                         break;
2326                 case VDATABASE:
2327                         if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2328                                 ip->flags |= HAMMER_INODE_TRUNCATED;
2329                                 ip->trunc_off = vap->va_size;
2330                         } else if (ip->trunc_off > vap->va_size) {
2331                                 ip->trunc_off = vap->va_size;
2332                         }
2333                         hammer_ip_frontend_trunc(ip, vap->va_size);
2334                         ip->ino_data.size = vap->va_size;
2335                         ip->ino_data.mtime = trans.time;
2336                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2337                         kflags |= NOTE_ATTRIB;
2338                         break;
2339                 default:
2340                         error = EINVAL;
2341                         goto done;
2342                 }
2343                 break;
2344         }
2345         if (vap->va_atime.tv_sec != VNOVAL) {
2346                 ip->ino_data.atime = hammer_timespec_to_time(&vap->va_atime);
2347                 modflags |= HAMMER_INODE_ATIME;
2348                 kflags |= NOTE_ATTRIB;
2349         }
2350         if (vap->va_mtime.tv_sec != VNOVAL) {
2351                 ip->ino_data.mtime = hammer_timespec_to_time(&vap->va_mtime);
2352                 modflags |= HAMMER_INODE_MTIME;
2353                 kflags |= NOTE_ATTRIB;
2354         }
2355         if (vap->va_mode != (mode_t)VNOVAL) {
2356                 mode_t   cur_mode = ip->ino_data.mode;
2357                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2358                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2359
2360                 error = vop_helper_chmod(ap->a_vp, vap->va_mode, ap->a_cred,
2361                                          cur_uid, cur_gid, &cur_mode);
2362                 if (error == 0 && ip->ino_data.mode != cur_mode) {
2363                         ip->ino_data.mode = cur_mode;
2364                         ip->ino_data.ctime = trans.time;
2365                         modflags |= HAMMER_INODE_DDIRTY;
2366                         kflags |= NOTE_ATTRIB;
2367                 }
2368         }
2369 done:
2370         if (error == 0)
2371                 hammer_modify_inode(&trans, ip, modflags);
2372         hammer_done_transaction(&trans);
2373         hammer_knote(ap->a_vp, kflags);
2374         lwkt_reltoken(&hmp->fs_token);
2375         return (error);
2376 }
2377
2378 /*
2379  * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
2380  */
2381 static
2382 int
2383 hammer_vop_nsymlink(struct vop_nsymlink_args *ap)
2384 {
2385         struct hammer_transaction trans;
2386         struct hammer_inode *dip;
2387         struct hammer_inode *nip;
2388         hammer_record_t record;
2389         struct nchandle *nch;
2390         hammer_mount_t hmp;
2391         int error;
2392         int bytes;
2393
2394         ap->a_vap->va_type = VLNK;
2395
2396         nch = ap->a_nch;
2397         dip = VTOI(ap->a_dvp);
2398         hmp = dip->hmp;
2399
2400         if (dip->flags & HAMMER_INODE_RO)
2401                 return (EROFS);
2402         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
2403                 return (error);
2404
2405         /*
2406          * Create a transaction to cover the operations we perform.
2407          */
2408         lwkt_gettoken(&hmp->fs_token);
2409         hammer_start_transaction(&trans, hmp);
2410         ++hammer_stats_file_iopsw;
2411
2412         /*
2413          * Create a new filesystem object of the requested type.  The
2414          * returned inode will be referenced but not locked.
2415          */
2416
2417         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
2418                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
2419                                     NULL, &nip);
2420         if (error) {
2421                 hammer_done_transaction(&trans);
2422                 *ap->a_vpp = NULL;
2423                 lwkt_reltoken(&hmp->fs_token);
2424                 return (error);
2425         }
2426
2427         /*
2428          * Add a record representing the symlink.  symlink stores the link
2429          * as pure data, not a string, and is no \0 terminated.
2430          */
2431         if (error == 0) {
2432                 bytes = strlen(ap->a_target);
2433
2434                 if (bytes <= HAMMER_INODE_BASESYMLEN) {
2435                         bcopy(ap->a_target, nip->ino_data.ext.symlink, bytes);
2436                 } else {
2437                         record = hammer_alloc_mem_record(nip, bytes);
2438                         record->type = HAMMER_MEM_RECORD_GENERAL;
2439
2440                         record->leaf.base.localization = nip->obj_localization +
2441                                                          HAMMER_LOCALIZE_MISC;
2442                         record->leaf.base.key = HAMMER_FIXKEY_SYMLINK;
2443                         record->leaf.base.rec_type = HAMMER_RECTYPE_FIX;
2444                         record->leaf.data_len = bytes;
2445                         KKASSERT(HAMMER_SYMLINK_NAME_OFF == 0);
2446                         bcopy(ap->a_target, record->data->symlink.name, bytes);
2447                         error = hammer_ip_add_record(&trans, record);
2448                 }
2449
2450                 /*
2451                  * Set the file size to the length of the link.
2452                  */
2453                 if (error == 0) {
2454                         nip->ino_data.size = bytes;
2455                         hammer_modify_inode(&trans, nip, HAMMER_INODE_DDIRTY);
2456                 }
2457         }
2458         if (error == 0)
2459                 error = hammer_ip_add_directory(&trans, dip, nch->ncp->nc_name,
2460                                                 nch->ncp->nc_nlen, nip);
2461
2462         /*
2463          * Finish up.
2464          */
2465         if (error) {
2466                 hammer_rel_inode(nip, 0);
2467                 *ap->a_vpp = NULL;
2468         } else {
2469                 error = hammer_get_vnode(nip, ap->a_vpp);
2470                 hammer_rel_inode(nip, 0);
2471                 if (error == 0) {
2472                         cache_setunresolved(ap->a_nch);
2473                         cache_setvp(ap->a_nch, *ap->a_vpp);
2474                         hammer_knote(ap->a_dvp, NOTE_WRITE);
2475                 }
2476         }
2477         hammer_done_transaction(&trans);
2478         lwkt_reltoken(&hmp->fs_token);
2479         return (error);
2480 }
2481
2482 /*
2483  * hammer_vop_nwhiteout { nch, dvp, cred, flags }
2484  */
2485 static
2486 int
2487 hammer_vop_nwhiteout(struct vop_nwhiteout_args *ap)
2488 {
2489         struct hammer_transaction trans;
2490         struct hammer_inode *dip;
2491         hammer_mount_t hmp;
2492         int error;
2493
2494         dip = VTOI(ap->a_dvp);
2495         hmp = dip->hmp;
2496
2497         if (hammer_nohistory(dip) == 0 &&
2498             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0) {
2499                 return (error);
2500         }
2501
2502         lwkt_gettoken(&hmp->fs_token);
2503         hammer_start_transaction(&trans, hmp);
2504         ++hammer_stats_file_iopsw;
2505         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp,
2506                                 ap->a_cred, ap->a_flags, -1);
2507         hammer_done_transaction(&trans);
2508         lwkt_reltoken(&hmp->fs_token);
2509
2510         return (error);
2511 }
2512
2513 /*
2514  * hammer_vop_ioctl { vp, command, data, fflag, cred }
2515  */
2516 static
2517 int
2518 hammer_vop_ioctl(struct vop_ioctl_args *ap)
2519 {
2520         struct hammer_inode *ip = ap->a_vp->v_data;
2521         hammer_mount_t hmp = ip->hmp;
2522         int error;
2523
2524         ++hammer_stats_file_iopsr;
2525         lwkt_gettoken(&hmp->fs_token);
2526         error = hammer_ioctl(ip, ap->a_command, ap->a_data,
2527                              ap->a_fflag, ap->a_cred);
2528         lwkt_reltoken(&hmp->fs_token);
2529         return (error);
2530 }
2531
2532 static
2533 int
2534 hammer_vop_mountctl(struct vop_mountctl_args *ap)
2535 {
2536         static const struct mountctl_opt extraopt[] = {
2537                 { HMNT_NOHISTORY,       "nohistory" },
2538                 { HMNT_MASTERID,        "master" },
2539                 { 0, NULL}
2540
2541         };
2542         struct hammer_mount *hmp;
2543         struct mount *mp;
2544         int usedbytes;
2545         int error;
2546
2547         error = 0;
2548         usedbytes = 0;
2549         mp = ap->a_head.a_ops->head.vv_mount;
2550         KKASSERT(mp->mnt_data != NULL);
2551         hmp = (struct hammer_mount *)mp->mnt_data;
2552
2553         lwkt_gettoken(&hmp->fs_token);
2554
2555         switch(ap->a_op) {
2556         case MOUNTCTL_SET_EXPORT:
2557                 if (ap->a_ctllen != sizeof(struct export_args))
2558                         error = EINVAL;
2559                 else
2560                         error = hammer_vfs_export(mp, ap->a_op,
2561                                       (const struct export_args *)ap->a_ctl);
2562                 break;
2563         case MOUNTCTL_MOUNTFLAGS:
2564         {
2565                 /*
2566                  * Call standard mountctl VOP function
2567                  * so we get user mount flags.
2568                  */
2569                 error = vop_stdmountctl(ap);
2570                 if (error)
2571                         break;
2572
2573                 usedbytes = *ap->a_res;
2574
2575                 if (usedbytes > 0 && usedbytes < ap->a_buflen) {
2576                         usedbytes += vfs_flagstostr(hmp->hflags, extraopt,
2577                                                     ap->a_buf,
2578                                                     ap->a_buflen - usedbytes,
2579                                                     &error);
2580                 }
2581
2582                 *ap->a_res += usedbytes;
2583                 break;
2584         }
2585         default:
2586                 error = vop_stdmountctl(ap);
2587                 break;
2588         }
2589         lwkt_reltoken(&hmp->fs_token);
2590         return(error);
2591 }
2592
2593 /*
2594  * hammer_vop_strategy { vp, bio }
2595  *
2596  * Strategy call, used for regular file read & write only.  Note that the
2597  * bp may represent a cluster.
2598  *
2599  * To simplify operation and allow better optimizations in the future,
2600  * this code does not make any assumptions with regards to buffer alignment
2601  * or size.
2602  */
2603 static
2604 int
2605 hammer_vop_strategy(struct vop_strategy_args *ap)
2606 {
2607         struct buf *bp;
2608         int error;
2609
2610         bp = ap->a_bio->bio_buf;
2611
2612         switch(bp->b_cmd) {
2613         case BUF_CMD_READ:
2614                 error = hammer_vop_strategy_read(ap);
2615                 break;
2616         case BUF_CMD_WRITE:
2617                 error = hammer_vop_strategy_write(ap);
2618                 break;
2619         default:
2620                 bp->b_error = error = EINVAL;
2621                 bp->b_flags |= B_ERROR;
2622                 biodone(ap->a_bio);
2623                 break;
2624         }
2625         return (error);
2626 }
2627
2628 /*
2629  * Read from a regular file.  Iterate the related records and fill in the
2630  * BIO/BUF.  Gaps are zero-filled.
2631  *
2632  * The support code in hammer_object.c should be used to deal with mixed
2633  * in-memory and on-disk records.
2634  *
2635  * NOTE: Can be called from the cluster code with an oversized buf.
2636  *
2637  * XXX atime update
2638  */
2639 static
2640 int
2641 hammer_vop_strategy_read(struct vop_strategy_args *ap)
2642 {
2643         struct hammer_transaction trans;
2644         struct hammer_inode *ip;
2645         struct hammer_inode *dip;
2646         hammer_mount_t hmp;
2647         struct hammer_cursor cursor;
2648         hammer_base_elm_t base;
2649         hammer_off_t disk_offset;
2650         struct bio *bio;
2651         struct bio *nbio;
2652         struct buf *bp;
2653         int64_t rec_offset;
2654         int64_t ran_end;
2655         int64_t tmp64;
2656         int error;
2657         int boff;
2658         int roff;
2659         int n;
2660
2661         bio = ap->a_bio;
2662         bp = bio->bio_buf;
2663         ip = ap->a_vp->v_data;
2664         hmp = ip->hmp;
2665
2666         /*
2667          * The zone-2 disk offset may have been set by the cluster code via
2668          * a BMAP operation, or else should be NOOFFSET.
2669          *
2670          * Checking the high bits for a match against zone-2 should suffice.
2671          */
2672         nbio = push_bio(bio);
2673         if ((nbio->bio_offset & HAMMER_OFF_ZONE_MASK) ==
2674             HAMMER_ZONE_LARGE_DATA) {
2675                 lwkt_gettoken(&hmp->fs_token);
2676                 error = hammer_io_direct_read(hmp, nbio, NULL);
2677                 lwkt_reltoken(&hmp->fs_token);
2678                 return (error);
2679         }
2680
2681         /*
2682          * Well, that sucked.  Do it the hard way.  If all the stars are
2683          * aligned we may still be able to issue a direct-read.
2684          */
2685         lwkt_gettoken(&hmp->fs_token);
2686         hammer_simple_transaction(&trans, hmp);
2687         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2688
2689         /*
2690          * Key range (begin and end inclusive) to scan.  Note that the key's
2691          * stored in the actual records represent BASE+LEN, not BASE.  The
2692          * first record containing bio_offset will have a key > bio_offset.
2693          */
2694         cursor.key_beg.localization = ip->obj_localization +
2695                                       HAMMER_LOCALIZE_MISC;
2696         cursor.key_beg.obj_id = ip->obj_id;
2697         cursor.key_beg.create_tid = 0;
2698         cursor.key_beg.delete_tid = 0;
2699         cursor.key_beg.obj_type = 0;
2700         cursor.key_beg.key = bio->bio_offset + 1;
2701         cursor.asof = ip->obj_asof;
2702         cursor.flags |= HAMMER_CURSOR_ASOF;
2703
2704         cursor.key_end = cursor.key_beg;
2705         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2706 #if 0
2707         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
2708                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
2709                 cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
2710                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2711         } else
2712 #endif
2713         {
2714                 ran_end = bio->bio_offset + bp->b_bufsize;
2715                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2716                 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2717                 tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2718                 if (tmp64 < ran_end)
2719                         cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2720                 else
2721                         cursor.key_end.key = ran_end + MAXPHYS + 1;
2722         }
2723         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2724
2725         error = hammer_ip_first(&cursor);
2726         boff = 0;
2727
2728         while (error == 0) {
2729                 /*
2730                  * Get the base file offset of the record.  The key for
2731                  * data records is (base + bytes) rather then (base).
2732                  */
2733                 base = &cursor.leaf->base;
2734                 rec_offset = base->key - cursor.leaf->data_len;
2735
2736                 /*
2737                  * Calculate the gap, if any, and zero-fill it.
2738                  *
2739                  * n is the offset of the start of the record verses our
2740                  * current seek offset in the bio.
2741                  */
2742                 n = (int)(rec_offset - (bio->bio_offset + boff));
2743                 if (n > 0) {
2744                         if (n > bp->b_bufsize - boff)
2745                                 n = bp->b_bufsize - boff;
2746                         bzero((char *)bp->b_data + boff, n);
2747                         boff += n;
2748                         n = 0;
2749                 }
2750
2751                 /*
2752                  * Calculate the data offset in the record and the number
2753                  * of bytes we can copy.
2754                  *
2755                  * There are two degenerate cases.  First, boff may already
2756                  * be at bp->b_bufsize.  Secondly, the data offset within
2757                  * the record may exceed the record's size.
2758                  */
2759                 roff = -n;
2760                 rec_offset += roff;
2761                 n = cursor.leaf->data_len - roff;
2762                 if (n <= 0) {
2763                         kprintf("strategy_read: bad n=%d roff=%d\n", n, roff);
2764                         n = 0;
2765                 } else if (n > bp->b_bufsize - boff) {
2766                         n = bp->b_bufsize - boff;
2767                 }
2768
2769                 /*
2770                  * Deal with cached truncations.  This cool bit of code
2771                  * allows truncate()/ftruncate() to avoid having to sync
2772                  * the file.
2773                  *
2774                  * If the frontend is truncated then all backend records are
2775                  * subject to the frontend's truncation.
2776                  *
2777                  * If the backend is truncated then backend records on-disk
2778                  * (but not in-memory) are subject to the backend's
2779                  * truncation.  In-memory records owned by the backend
2780                  * represent data written after the truncation point on the
2781                  * backend and must not be truncated.
2782                  *
2783                  * Truncate operations deal with frontend buffer cache
2784                  * buffers and frontend-owned in-memory records synchronously.
2785                  */
2786                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2787                         if (hammer_cursor_ondisk(&cursor)/* ||
2788                             cursor.iprec->flush_state == HAMMER_FST_FLUSH*/) {
2789                                 if (ip->trunc_off <= rec_offset)
2790                                         n = 0;
2791                                 else if (ip->trunc_off < rec_offset + n)
2792                                         n = (int)(ip->trunc_off - rec_offset);
2793                         }
2794                 }
2795                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2796                         if (hammer_cursor_ondisk(&cursor)) {
2797                                 if (ip->sync_trunc_off <= rec_offset)
2798                                         n = 0;
2799                                 else if (ip->sync_trunc_off < rec_offset + n)
2800                                         n = (int)(ip->sync_trunc_off - rec_offset);
2801                         }
2802                 }
2803
2804                 /*
2805                  * Try to issue a direct read into our bio if possible,
2806                  * otherwise resolve the element data into a hammer_buffer
2807                  * and copy.
2808                  *
2809                  * The buffer on-disk should be zerod past any real
2810                  * truncation point, but may not be for any synthesized
2811                  * truncation point from above.
2812                  */
2813                 disk_offset = cursor.leaf->data_offset + roff;
2814                 if (boff == 0 && n == bp->b_bufsize &&
2815                     hammer_cursor_ondisk(&cursor) &&
2816                     (disk_offset & HAMMER_BUFMASK) == 0) {
2817                         KKASSERT((disk_offset & HAMMER_OFF_ZONE_MASK) ==
2818                                  HAMMER_ZONE_LARGE_DATA);
2819                         nbio->bio_offset = disk_offset;
2820                         error = hammer_io_direct_read(hmp, nbio, cursor.leaf);
2821                         goto done;
2822                 } else if (n) {
2823                         error = hammer_ip_resolve_data(&cursor);
2824                         if (error == 0) {
2825                                 bcopy((char *)cursor.data + roff,
2826                                       (char *)bp->b_data + boff, n);
2827                         }
2828                 }
2829                 if (error)
2830                         break;
2831
2832                 /*
2833                  * Iterate until we have filled the request.
2834                  */
2835                 boff += n;
2836                 if (boff == bp->b_bufsize)
2837                         break;
2838                 error = hammer_ip_next(&cursor);
2839         }
2840
2841         /*
2842          * There may have been a gap after the last record
2843          */
2844         if (error == ENOENT)
2845                 error = 0;
2846         if (error == 0 && boff != bp->b_bufsize) {
2847                 KKASSERT(boff < bp->b_bufsize);
2848                 bzero((char *)bp->b_data + boff, bp->b_bufsize - boff);
2849                 /* boff = bp->b_bufsize; */
2850         }
2851         bp->b_resid = 0;
2852         bp->b_error = error;
2853         if (error)
2854                 bp->b_flags |= B_ERROR;
2855         biodone(ap->a_bio);
2856
2857 done:
2858         /*
2859          * Cache the b-tree node for the last data read in cache[1].
2860          *
2861          * If we hit the file EOF then also cache the node in the
2862          * governing director's cache[3], it will be used to initialize
2863          * the inode's cache[1] for any inodes looked up via the directory.
2864          *
2865          * This doesn't reduce disk accesses since the B-Tree chain is
2866          * likely cached, but it does reduce cpu overhead when looking
2867          * up file offsets for cpdup/tar/cpio style iterations.
2868          */
2869         if (cursor.node)
2870                 hammer_cache_node(&ip->cache[1], cursor.node);
2871         if (ran_end >= ip->ino_data.size) {
2872                 dip = hammer_find_inode(&trans, ip->ino_data.parent_obj_id,
2873                                         ip->obj_asof, ip->obj_localization);
2874                 if (dip) {
2875                         hammer_cache_node(&dip->cache[3], cursor.node);
2876                         hammer_rel_inode(dip, 0);
2877                 }
2878         }
2879         hammer_done_cursor(&cursor);
2880         hammer_done_transaction(&trans);
2881         lwkt_reltoken(&hmp->fs_token);
2882         return(error);
2883 }
2884
2885 /*
2886  * BMAP operation - used to support cluster_read() only.
2887  *
2888  * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
2889  *
2890  * This routine may return EOPNOTSUPP if the opration is not supported for
2891  * the specified offset.  The contents of the pointer arguments do not
2892  * need to be initialized in that case.
2893  *
2894  * If a disk address is available and properly aligned return 0 with
2895  * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
2896  * to the run-length relative to that offset.  Callers may assume that
2897  * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
2898  * large, so return EOPNOTSUPP if it is not sufficiently large.
2899  */
2900 static
2901 int
2902 hammer_vop_bmap(struct vop_bmap_args *ap)
2903 {
2904         struct hammer_transaction trans;
2905         struct hammer_inode *ip;
2906         hammer_mount_t hmp;
2907         struct hammer_cursor cursor;
2908         hammer_base_elm_t base;
2909         int64_t rec_offset;
2910         int64_t ran_end;
2911         int64_t tmp64;
2912         int64_t base_offset;
2913         int64_t base_disk_offset;
2914         int64_t last_offset;
2915         hammer_off_t last_disk_offset;
2916         hammer_off_t disk_offset;
2917         int     rec_len;
2918         int     error;
2919         int     blksize;
2920
2921         ++hammer_stats_file_iopsr;
2922         ip = ap->a_vp->v_data;
2923         hmp = ip->hmp;
2924
2925         /*
2926          * We can only BMAP regular files.  We can't BMAP database files,
2927          * directories, etc.
2928          */
2929         if (ip->ino_data.obj_type != HAMMER_OBJTYPE_REGFILE)
2930                 return(EOPNOTSUPP);
2931
2932         /*
2933          * bmap is typically called with runp/runb both NULL when used
2934          * for writing.  We do not support BMAP for writing atm.
2935          */
2936         if (ap->a_cmd != BUF_CMD_READ)
2937                 return(EOPNOTSUPP);
2938
2939         /*
2940          * Scan the B-Tree to acquire blockmap addresses, then translate
2941          * to raw addresses.
2942          */
2943         lwkt_gettoken(&hmp->fs_token);
2944         hammer_simple_transaction(&trans, hmp);
2945 #if 0
2946         kprintf("bmap_beg %016llx ip->cache %p\n",
2947                 (long long)ap->a_loffset, ip->cache[1]);
2948 #endif
2949         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2950
2951         /*
2952          * Key range (begin and end inclusive) to scan.  Note that the key's
2953          * stored in the actual records represent BASE+LEN, not BASE.  The
2954          * first record containing bio_offset will have a key > bio_offset.
2955          */
2956         cursor.key_beg.localization = ip->obj_localization +
2957                                       HAMMER_LOCALIZE_MISC;
2958         cursor.key_beg.obj_id = ip->obj_id;
2959         cursor.key_beg.create_tid = 0;
2960         cursor.key_beg.delete_tid = 0;
2961         cursor.key_beg.obj_type = 0;
2962         if (ap->a_runb)
2963                 cursor.key_beg.key = ap->a_loffset - MAXPHYS + 1;
2964         else
2965                 cursor.key_beg.key = ap->a_loffset + 1;
2966         if (cursor.key_beg.key < 0)
2967                 cursor.key_beg.key = 0;
2968         cursor.asof = ip->obj_asof;
2969         cursor.flags |= HAMMER_CURSOR_ASOF;
2970
2971         cursor.key_end = cursor.key_beg;
2972         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2973
2974         ran_end = ap->a_loffset + MAXPHYS;
2975         cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2976         cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2977         tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2978         if (tmp64 < ran_end)
2979                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2980         else
2981                 cursor.key_end.key = ran_end + MAXPHYS + 1;
2982
2983         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2984
2985         error = hammer_ip_first(&cursor);
2986         base_offset = last_offset = 0;
2987         base_disk_offset = last_disk_offset = 0;
2988
2989         while (error == 0) {
2990                 /*
2991                  * Get the base file offset of the record.  The key for
2992                  * data records is (base + bytes) rather then (base).
2993                  *
2994                  * NOTE: rec_offset + rec_len may exceed the end-of-file.
2995                  * The extra bytes should be zero on-disk and the BMAP op
2996                  * should still be ok.
2997                  */
2998                 base = &cursor.leaf->base;
2999                 rec_offset = base->key - cursor.leaf->data_len;
3000                 rec_len    = cursor.leaf->data_len;
3001
3002                 /*
3003                  * Incorporate any cached truncation.
3004                  *
3005                  * NOTE: Modifications to rec_len based on synthesized
3006                  * truncation points remove the guarantee that any extended
3007                  * data on disk is zero (since the truncations may not have
3008                  * taken place on-media yet).
3009                  */
3010                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
3011                         if (hammer_cursor_ondisk(&cursor) ||
3012                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
3013                                 if (ip->trunc_off <= rec_offset)
3014                                         rec_len = 0;
3015                                 else if (ip->trunc_off < rec_offset + rec_len)
3016                                         rec_len = (int)(ip->trunc_off - rec_offset);
3017                         }
3018                 }
3019                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
3020                         if (hammer_cursor_ondisk(&cursor)) {
3021                                 if (ip->sync_trunc_off <= rec_offset)
3022                                         rec_len = 0;
3023                                 else if (ip->sync_trunc_off < rec_offset + rec_len)
3024                                         rec_len = (int)(ip->sync_trunc_off - rec_offset);
3025                         }
3026                 }
3027
3028                 /*
3029                  * Accumulate information.  If we have hit a discontiguous
3030                  * block reset base_offset unless we are already beyond the
3031                  * requested offset.  If we are, that's it, we stop.
3032                  */
3033                 if (error)
3034                         break;
3035                 if (hammer_cursor_ondisk(&cursor)) {
3036                         disk_offset = cursor.leaf->data_offset;
3037                         if (rec_offset != last_offset ||
3038                             disk_offset != last_disk_offset) {
3039                                 if (rec_offset > ap->a_loffset)
3040                                         break;
3041                                 base_offset = rec_offset;
3042                                 base_disk_offset = disk_offset;
3043                         }
3044                         last_offset = rec_offset + rec_len;
3045                         last_disk_offset = disk_offset + rec_len;
3046                 }
3047                 error = hammer_ip_next(&cursor);
3048         }
3049
3050 #if 0
3051         kprintf("BMAP %016llx:  %016llx - %016llx\n",
3052                 (long long)ap->a_loffset,
3053                 (long long)base_offset,
3054                 (long long)last_offset);
3055         kprintf("BMAP %16s:  %016llx - %016llx\n", "",
3056                 (long long)base_disk_offset,
3057                 (long long)last_disk_offset);
3058 #endif
3059
3060         if (cursor.node) {
3061                 hammer_cache_node(&ip->cache[1], cursor.node);
3062 #if 0
3063                 kprintf("bmap_end2 %016llx ip->cache %p\n",
3064                         (long long)ap->a_loffset, ip->cache[1]);
3065 #endif
3066         }
3067         hammer_done_cursor(&cursor);
3068         hammer_done_transaction(&trans);
3069         lwkt_reltoken(&hmp->fs_token);
3070
3071         /*
3072          * If we couldn't find any records or the records we did find were
3073          * all behind the requested offset, return failure.  A forward
3074          * truncation can leave a hole w/ no on-disk records.
3075          */
3076         if (last_offset == 0 || last_offset < ap->a_loffset)
3077                 return (EOPNOTSUPP);
3078
3079         /*
3080          * Figure out the block size at the requested offset and adjust
3081          * our limits so the cluster_read() does not create inappropriately
3082          * sized buffer cache buffers.
3083          */
3084         blksize = hammer_blocksize(ap->a_loffset);
3085         if (hammer_blocksize(base_offset) != blksize) {
3086                 base_offset = hammer_blockdemarc(base_offset, ap->a_loffset);
3087         }
3088         if (last_offset != ap->a_loffset &&
3089             hammer_blocksize(last_offset - 1) != blksize) {
3090                 last_offset = hammer_blockdemarc(ap->a_loffset,
3091                                                  last_offset - 1);
3092         }
3093
3094         /*
3095          * Returning EOPNOTSUPP simply prevents the direct-IO optimization
3096          * from occuring.
3097          */
3098         disk_offset = base_disk_offset + (ap->a_loffset - base_offset);
3099
3100         if ((disk_offset & HAMMER_OFF_ZONE_MASK) != HAMMER_ZONE_LARGE_DATA) {
3101                 /*
3102                  * Only large-data zones can be direct-IOd
3103                  */
3104                 error = EOPNOTSUPP;
3105         } else if ((disk_offset & HAMMER_BUFMASK) ||
3106                    (last_offset - ap->a_loffset) < blksize) {
3107                 /*
3108                  * doffsetp is not aligned or the forward run size does
3109                  * not cover a whole buffer, disallow the direct I/O.
3110                  */
3111                 error = EOPNOTSUPP;
3112         } else {
3113                 /*
3114                  * We're good.
3115                  */
3116                 *ap->a_doffsetp = disk_offset;
3117                 if (ap->a_runb) {
3118                         *ap->a_runb = ap->a_loffset - base_offset;
3119                         KKASSERT(*ap->a_runb >= 0);
3120                 }
3121                 if (ap->a_runp) {
3122                         *ap->a_runp = last_offset - ap->a_loffset;
3123                         KKASSERT(*ap->a_runp >= 0);
3124                 }
3125                 error = 0;
3126         }
3127         return(error);
3128 }
3129
3130 /*
3131  * Write to a regular file.   Because this is a strategy call the OS is
3132  * trying to actually get data onto the media.
3133  */
3134 static
3135 int
3136 hammer_vop_strategy_write(struct vop_strategy_args *ap)
3137 {
3138         hammer_record_t record;
3139         hammer_mount_t hmp;
3140         hammer_inode_t ip;
3141         struct bio *bio;
3142         struct buf *bp;
3143         int blksize;
3144         int bytes;
3145         int error;
3146
3147         bio = ap->a_bio;
3148         bp = bio->bio_buf;
3149         ip = ap->a_vp->v_data;
3150         hmp = ip->hmp;
3151
3152         blksize = hammer_blocksize(bio->bio_offset);
3153         KKASSERT(bp->b_bufsize == blksize);
3154
3155         if (ip->flags & HAMMER_INODE_RO) {
3156                 bp->b_error = EROFS;
3157                 bp->b_flags |= B_ERROR;
3158                 biodone(ap->a_bio);
3159                 return(EROFS);
3160         }
3161
3162         lwkt_gettoken(&hmp->fs_token);
3163
3164         /*
3165          * Interlock with inode destruction (no in-kernel or directory
3166          * topology visibility).  If we queue new IO while trying to
3167          * destroy the inode we can deadlock the vtrunc call in
3168          * hammer_inode_unloadable_check().
3169          *
3170          * Besides, there's no point flushing a bp associated with an
3171          * inode that is being destroyed on-media and has no kernel
3172          * references.
3173          */
3174         if ((ip->flags | ip->sync_flags) &
3175             (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) {
3176                 bp->b_resid = 0;
3177                 biodone(ap->a_bio);
3178                 lwkt_reltoken(&hmp->fs_token);
3179                 return(0);
3180         }
3181
3182         /*
3183          * Reserve space and issue a direct-write from the front-end.
3184          * NOTE: The direct_io code will hammer_bread/bcopy smaller
3185          * allocations.
3186          *
3187          * An in-memory record will be installed to reference the storage
3188          * until the flusher can get to it.
3189          *
3190          * Since we own the high level bio the front-end will not try to
3191          * do a direct-read until the write completes.
3192          *
3193          * NOTE: The only time we do not reserve a full-sized buffers
3194          * worth of data is if the file is small.  We do not try to
3195          * allocate a fragment (from the small-data zone) at the end of
3196          * an otherwise large file as this can lead to wildly separated
3197          * data.
3198          */
3199         KKASSERT((bio->bio_offset & HAMMER_BUFMASK) == 0);
3200         KKASSERT(bio->bio_offset < ip->ino_data.size);
3201         if (bio->bio_offset || ip->ino_data.size > HAMMER_BUFSIZE / 2)
3202                 bytes = bp->b_bufsize;
3203         else
3204                 bytes = ((int)ip->ino_data.size + 15) & ~15;
3205
3206         record = hammer_ip_add_bulk(ip, bio->bio_offset, bp->b_data,
3207                                     bytes, &error);
3208
3209         /*
3210          * B_VFSFLAG1 indicates that a REDO_WRITE entry was generated
3211          * in hammer_vop_write().  We must flag the record so the proper
3212          * REDO_TERM_WRITE entry is generated during the flush.
3213          */
3214         if (record) {
3215                 if (bp->b_flags & B_VFSFLAG1) {
3216                         record->flags |= HAMMER_RECF_REDO;
3217                         bp->b_flags &= ~B_VFSFLAG1;
3218                 }
3219                 hammer_io_direct_write(hmp, bio, record);
3220                 if (ip->rsv_recs > 1 && hmp->rsv_recs > hammer_limit_recs)
3221                         hammer_flush_inode(ip, 0);
3222         } else {
3223                 bp->b_bio2.bio_offset = NOOFFSET;
3224                 bp->b_error = error;
3225                 bp->b_flags |= B_ERROR;
3226                 biodone(ap->a_bio);
3227         }
3228         lwkt_reltoken(&hmp->fs_token);
3229         return(error);
3230 }
3231
3232 /*
3233  * dounlink - disconnect a directory entry
3234  *
3235  * XXX whiteout support not really in yet
3236  */
3237 static int
3238 hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
3239                 struct vnode *dvp, struct ucred *cred,
3240                 int flags, int isdir)
3241 {
3242         struct namecache *ncp;
3243         hammer_inode_t dip;
3244         hammer_inode_t ip;
3245         hammer_mount_t hmp;
3246         struct hammer_cursor cursor;
3247         int64_t namekey;
3248         u_int32_t max_iterations;
3249         int nlen, error;
3250
3251         /*
3252          * Calculate the namekey and setup the key range for the scan.  This
3253          * works kinda like a chained hash table where the lower 32 bits
3254          * of the namekey synthesize the chain.
3255          *
3256          * The key range is inclusive of both key_beg and key_end.
3257          */
3258         dip = VTOI(dvp);
3259         ncp = nch->ncp;
3260         hmp = dip->hmp;
3261
3262         if (dip->flags & HAMMER_INODE_RO)
3263                 return (EROFS);
3264
3265         namekey = hammer_directory_namekey(dip, ncp->nc_name, ncp->nc_nlen,
3266                                            &max_iterations);
3267 retry:
3268         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
3269         cursor.key_beg.localization = dip->obj_localization +
3270                                       hammer_dir_localization(dip);
3271         cursor.key_beg.obj_id = dip->obj_id;
3272         cursor.key_beg.key = namekey;
3273         cursor.key_beg.create_tid = 0;
3274         cursor.key_beg.delete_tid = 0;
3275         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
3276         cursor.key_beg.obj_type = 0;
3277
3278         cursor.key_end = cursor.key_beg;
3279         cursor.key_end.key += max_iterations;
3280         cursor.asof = dip->obj_asof;
3281         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
3282
3283         /*
3284          * Scan all matching records (the chain), locate the one matching
3285          * the requested path component.  info->last_error contains the
3286          * error code on search termination and could be 0, ENOENT, or
3287          * something else.
3288          *
3289          * The hammer_ip_*() functions merge in-memory records with on-disk
3290          * records for the purposes of the search.
3291          */
3292         error = hammer_ip_first(&cursor);
3293
3294         while (error == 0) {
3295                 error = hammer_ip_resolve_data(&cursor);
3296                 if (error)
3297                         break;
3298                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
3299                 KKASSERT(nlen > 0);
3300                 if (ncp->nc_nlen == nlen &&
3301                     bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
3302                         break;
3303                 }
3304                 error = hammer_ip_next(&cursor);
3305         }
3306
3307         /*
3308          * If all is ok we have to get the inode so we can adjust nlinks.
3309          * To avoid a deadlock with the flusher we must release the inode
3310          * lock on the directory when acquiring the inode for the entry.
3311          *
3312          * If the target is a directory, it must be empty.
3313          */
3314         if (error == 0) {
3315                 hammer_unlock(&cursor.ip->lock);
3316                 ip = hammer_get_inode(trans, dip, cursor.data->entry.obj_id,
3317                                       hmp->asof,
3318                                       cursor.data->entry.localization,
3319                                       0, &error);
3320                 hammer_lock_sh(&cursor.ip->lock);
3321                 if (error == ENOENT) {
3322                         kprintf("HAMMER: WARNING: Removing "
3323                                 "dirent w/missing inode \"%s\"\n"
3324                                 "\tobj_id = %016llx\n",
3325                                 ncp->nc_name,
3326                                 (long long)cursor.data->entry.obj_id);
3327                         error = 0;
3328                 }
3329
3330                 /*
3331                  * If isdir >= 0 we validate that the entry is or is not a
3332                  * directory.  If isdir < 0 we don't care.
3333                  */
3334                 if (error == 0 && isdir >= 0 && ip) {
3335                         if (isdir &&
3336                             ip->ino_data.obj_type != HAMMER_OBJTYPE_DIRECTORY) {
3337                                 error = ENOTDIR;
3338                         } else if (isdir == 0 &&
3339                             ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
3340                                 error = EISDIR;
3341                         }
3342                 }
3343
3344                 /*
3345                  * If we are trying to remove a directory the directory must
3346                  * be empty.
3347                  *
3348                  * The check directory code can loop and deadlock/retry.  Our
3349                  * own cursor's node locks must be released to avoid a 3-way
3350                  * deadlock with the flusher if the check directory code
3351                  * blocks.
3352                  *
3353                  * If any changes whatsoever have been made to the cursor
3354                  * set EDEADLK and retry.
3355                  *
3356                  * WARNING: See warnings in hammer_unlock_cursor()
3357                  *          function.
3358                  */
3359                 if (error == 0 && ip && ip->ino_data.obj_type ==
3360                                         HAMMER_OBJTYPE_DIRECTORY) {
3361                         hammer_unlock_cursor(&cursor);
3362                         error = hammer_ip_check_directory_empty(trans, ip);
3363                         hammer_lock_cursor(&cursor);
3364                         if (cursor.flags & HAMMER_CURSOR_RETEST) {
3365                                 kprintf("HAMMER: Warning: avoided deadlock "
3366                                         "on rmdir '%s'\n",
3367                                         ncp->nc_name);
3368                                 error = EDEADLK;
3369                         }
3370                 }
3371
3372                 /*
3373                  * Delete the directory entry.
3374                  *
3375                  * WARNING: hammer_ip_del_directory() may have to terminate
3376                  * the cursor to avoid a deadlock.  It is ok to call
3377                  * hammer_done_cursor() twice.
3378                  */
3379                 if (error == 0) {
3380                         error = hammer_ip_del_directory(trans, &cursor,
3381                                                         dip, ip);
3382                 }
3383                 hammer_done_cursor(&cursor);
3384                 if (error == 0) {
3385                         cache_setunresolved(nch);
3386                         cache_setvp(nch, NULL);
3387
3388                         /*
3389                          * NOTE: ip->vp, if non-NULL, cannot be directly
3390                          *       referenced without formally acquiring the
3391                          *       vp since the vp might have zero refs on it,
3392                          *       or in the middle of a reclaim, etc.
3393                          *
3394                          * NOTE: The cache_setunresolved() can rip the vp
3395                          *       out from under us since the vp may not have
3396                          *       any refs, in which case ip->vp will be NULL
3397                          *       from the outset.
3398                          */
3399                         while (ip && ip->vp) {
3400                                 struct vnode *vp;
3401
3402                                 error = hammer_get_vnode(ip, &vp);
3403                                 if (error == 0 && vp) {
3404                                         vn_unlock(vp);
3405                                         hammer_knote(ip->vp, NOTE_DELETE);
3406                                         cache_inval_vp(ip->vp, CINV_DESTROY);
3407                                         vrele(vp);
3408                                         break;
3409                                 }
3410                                 kprintf("Debug: HAMMER ip/vp race1 avoided\n");
3411                         }
3412                 }
3413                 if (ip)
3414                         hammer_rel_inode(ip, 0);
3415         } else {
3416                 hammer_done_cursor(&cursor);
3417         }
3418         if (error == EDEADLK)
3419                 goto retry;
3420
3421         return (error);
3422 }
3423
3424 /************************************************************************
3425  *                          FIFO AND SPECFS OPS                         *
3426  ************************************************************************
3427  *
3428  */
3429 static int
3430 hammer_vop_fifoclose (struct vop_close_args *ap)
3431 {
3432         /* XXX update itimes */
3433         return (VOCALL(&fifo_vnode_vops, &ap->a_head));
3434 }
3435
3436 static int
3437 hammer_vop_fiforead (struct vop_read_args *ap)
3438 {
3439         int error;
3440
3441         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3442         /* XXX update access time */
3443         return (error);
3444 }
3445
3446 static int
3447 hammer_vop_fifowrite (struct vop_write_args *ap)
3448 {
3449         int error;
3450
3451         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3452         /* XXX update access time */
3453         return (error);
3454 }
3455
3456 static
3457 int
3458 hammer_vop_fifokqfilter(struct vop_kqfilter_args *ap)
3459 {
3460         int error;
3461
3462         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3463         if (error)
3464                 error = hammer_vop_kqfilter(ap);
3465         return(error);
3466 }
3467
3468 /************************************************************************
3469  *                          KQFILTER OPS                                *
3470  ************************************************************************
3471  *
3472  */
3473 static void filt_hammerdetach(struct knote *kn);
3474 static int filt_hammerread(struct knote *kn, long hint);
3475 static int filt_hammerwrite(struct knote *kn, long hint);
3476 static int filt_hammervnode(struct knote *kn, long hint);
3477
3478 static struct filterops hammerread_filtops =
3479         { FILTEROP_ISFD, NULL, filt_hammerdetach, filt_hammerread };
3480 static struct filterops hammerwrite_filtops =
3481         { FILTEROP_ISFD, NULL, filt_hammerdetach, filt_hammerwrite };
3482 static struct filterops hammervnode_filtops =
3483         { FILTEROP_ISFD, NULL, filt_hammerdetach, filt_hammervnode };
3484
3485 static
3486 int
3487 hammer_vop_kqfilter(struct vop_kqfilter_args *ap)
3488 {
3489         struct vnode *vp = ap->a_vp;
3490         struct knote *kn = ap->a_kn;
3491
3492         switch (kn->kn_filter) {
3493         case EVFILT_READ:
3494                 kn->kn_fop = &hammerread_filtops;
3495                 break;
3496         case EVFILT_WRITE:
3497                 kn->kn_fop = &hammerwrite_filtops;
3498                 break;
3499         case EVFILT_VNODE:
3500                 kn->kn_fop = &hammervnode_filtops;
3501                 break;
3502         default:
3503                 return (EOPNOTSUPP);
3504         }
3505
3506         kn->kn_hook = (caddr_t)vp;
3507
3508         knote_insert(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3509
3510         return(0);
3511 }
3512
3513 static void
3514 filt_hammerdetach(struct knote *kn)
3515 {
3516         struct vnode *vp = (void *)kn->kn_hook;
3517
3518         knote_remove(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3519 }
3520
3521 static int
3522 filt_hammerread(struct knote *kn, long hint)
3523 {
3524         struct vnode *vp = (void *)kn->kn_hook;
3525         hammer_inode_t ip = VTOI(vp);
3526         hammer_mount_t hmp = ip->hmp;
3527         off_t off;
3528
3529         if (hint == NOTE_REVOKE) {
3530                 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
3531                 return(1);
3532         }
3533         lwkt_gettoken(&hmp->fs_token);  /* XXX use per-ip-token */
3534         off = ip->ino_data.size - kn->kn_fp->f_offset;
3535         kn->kn_data = (off < INTPTR_MAX) ? off : INTPTR_MAX;
3536         lwkt_reltoken(&hmp->fs_token);
3537         if (kn->kn_sfflags & NOTE_OLDAPI)
3538                 return(1);
3539         return (kn->kn_data != 0);
3540 }
3541
3542 static int
3543 filt_hammerwrite(struct knote *kn, long hint)
3544 {
3545         if (hint == NOTE_REVOKE)
3546                 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
3547         kn->kn_data = 0;
3548         return (1);
3549 }
3550
3551 static int
3552 filt_hammervnode(struct knote *kn, long hint)
3553 {
3554         if (kn->kn_sfflags & hint)
3555                 kn->kn_fflags |= hint;
3556         if (hint == NOTE_REVOKE) {
3557                 kn->kn_flags |= EV_EOF;
3558                 return (1);
3559         }
3560         return (kn->kn_fflags != 0);
3561 }
3562