sys/vfs/hammer/hammer_vnops.c

   1 /*
   2  * Copyright (c) 2007-2008 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  *
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in
  15  *    the documentation and/or other materials provided with the
  16  *    distribution.
  17  * 3. Neither the name of The DragonFly Project nor the names of its
  18  *    contributors may be used to endorse or promote products derived
  19  *    from this software without specific, prior written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  32  * SUCH DAMAGE.
  33  *
  34  * $DragonFly: src/sys/vfs/hammer/hammer_vnops.c,v 1.102 2008/10/16 17:24:16 dillon Exp $
  35  */
  36
  37 #include <sys/param.h>
  38 #include <sys/systm.h>
  39 #include <sys/kernel.h>
  40 #include <sys/fcntl.h>
  41 #include <sys/namecache.h>
  42 #include <sys/vnode.h>
  43 #include <sys/lockf.h>
  44 #include <sys/event.h>
  45 #include <sys/stat.h>
  46 #include <sys/dirent.h>
  47 #include <sys/file.h>
  48 #include <vm/vm_extern.h>
  49 #include <vm/swap_pager.h>
  50 #include <vfs/fifofs/fifo.h>
  51
  52 #include "hammer.h"
  53
  54 /*
  55  * USERFS VNOPS
  56  */
  57 /*static int hammer_vop_vnoperate(struct vop_generic_args *);*/
  58 static int hammer_vop_fsync(struct vop_fsync_args *);
  59 static int hammer_vop_read(struct vop_read_args *);
  60 static int hammer_vop_write(struct vop_write_args *);
  61 static int hammer_vop_access(struct vop_access_args *);
  62 static int hammer_vop_advlock(struct vop_advlock_args *);
  63 static int hammer_vop_close(struct vop_close_args *);
  64 static int hammer_vop_ncreate(struct vop_ncreate_args *);
  65 static int hammer_vop_getattr(struct vop_getattr_args *);
  66 static int hammer_vop_nresolve(struct vop_nresolve_args *);
  67 static int hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *);
  68 static int hammer_vop_nlink(struct vop_nlink_args *);
  69 static int hammer_vop_nmkdir(struct vop_nmkdir_args *);
  70 static int hammer_vop_nmknod(struct vop_nmknod_args *);
  71 static int hammer_vop_open(struct vop_open_args *);
  72 static int hammer_vop_print(struct vop_print_args *);
  73 static int hammer_vop_readdir(struct vop_readdir_args *);
  74 static int hammer_vop_readlink(struct vop_readlink_args *);
  75 static int hammer_vop_nremove(struct vop_nremove_args *);
  76 static int hammer_vop_nrename(struct vop_nrename_args *);
  77 static int hammer_vop_nrmdir(struct vop_nrmdir_args *);
  78 static int hammer_vop_markatime(struct vop_markatime_args *);
  79 static int hammer_vop_setattr(struct vop_setattr_args *);
  80 static int hammer_vop_strategy(struct vop_strategy_args *);
  81 static int hammer_vop_bmap(struct vop_bmap_args *ap);
  82 static int hammer_vop_nsymlink(struct vop_nsymlink_args *);
  83 static int hammer_vop_nwhiteout(struct vop_nwhiteout_args *);
  84 static int hammer_vop_ioctl(struct vop_ioctl_args *);
  85 static int hammer_vop_mountctl(struct vop_mountctl_args *);
  86 static int hammer_vop_kqfilter (struct vop_kqfilter_args *);
  87
  88 static int hammer_vop_fifoclose (struct vop_close_args *);
  89 static int hammer_vop_fiforead (struct vop_read_args *);
  90 static int hammer_vop_fifowrite (struct vop_write_args *);
  91 static int hammer_vop_fifokqfilter (struct vop_kqfilter_args *);
  92
  93 struct vop_ops hammer_vnode_vops = {
  94         .vop_default =          vop_defaultop,
  95         .vop_fsync =            hammer_vop_fsync,
  96         .vop_getpages =         vop_stdgetpages,
  97         .vop_putpages =         vop_stdputpages,
  98         .vop_read =             hammer_vop_read,
  99         .vop_write =            hammer_vop_write,
 100         .vop_access =           hammer_vop_access,
 101         .vop_advlock =          hammer_vop_advlock,
 102         .vop_close =            hammer_vop_close,
 103         .vop_ncreate =          hammer_vop_ncreate,
 104         .vop_getattr =          hammer_vop_getattr,
 105         .vop_inactive =         hammer_vop_inactive,
 106         .vop_reclaim =          hammer_vop_reclaim,
 107         .vop_nresolve =         hammer_vop_nresolve,
 108         .vop_nlookupdotdot =    hammer_vop_nlookupdotdot,
 109         .vop_nlink =            hammer_vop_nlink,
 110         .vop_nmkdir =           hammer_vop_nmkdir,
 111         .vop_nmknod =           hammer_vop_nmknod,
 112         .vop_open =             hammer_vop_open,
 113         .vop_pathconf =         vop_stdpathconf,
 114         .vop_print =            hammer_vop_print,
 115         .vop_readdir =          hammer_vop_readdir,
 116         .vop_readlink =         hammer_vop_readlink,
 117         .vop_nremove =          hammer_vop_nremove,
 118         .vop_nrename =          hammer_vop_nrename,
 119         .vop_nrmdir =           hammer_vop_nrmdir,
 120         .vop_markatime =        hammer_vop_markatime,
 121         .vop_setattr =          hammer_vop_setattr,
 122         .vop_bmap =             hammer_vop_bmap,
 123         .vop_strategy =         hammer_vop_strategy,
 124         .vop_nsymlink =         hammer_vop_nsymlink,
 125         .vop_nwhiteout =        hammer_vop_nwhiteout,
 126         .vop_ioctl =            hammer_vop_ioctl,
 127         .vop_mountctl =         hammer_vop_mountctl,
 128         .vop_kqfilter =         hammer_vop_kqfilter
 129 };
 130
 131 struct vop_ops hammer_spec_vops = {
 132         .vop_default =          vop_defaultop,
 133         .vop_fsync =            hammer_vop_fsync,
 134         .vop_read =             vop_stdnoread,
 135         .vop_write =            vop_stdnowrite,
 136         .vop_access =           hammer_vop_access,
 137         .vop_close =            hammer_vop_close,
 138         .vop_markatime =        hammer_vop_markatime,
 139         .vop_getattr =          hammer_vop_getattr,
 140         .vop_inactive =         hammer_vop_inactive,
 141         .vop_reclaim =          hammer_vop_reclaim,
 142         .vop_setattr =          hammer_vop_setattr
 143 };
 144
 145 struct vop_ops hammer_fifo_vops = {
 146         .vop_default =          fifo_vnoperate,
 147         .vop_fsync =            hammer_vop_fsync,
 148         .vop_read =             hammer_vop_fiforead,
 149         .vop_write =            hammer_vop_fifowrite,
 150         .vop_access =           hammer_vop_access,
 151         .vop_close =            hammer_vop_fifoclose,
 152         .vop_markatime =        hammer_vop_markatime,
 153         .vop_getattr =          hammer_vop_getattr,
 154         .vop_inactive =         hammer_vop_inactive,
 155         .vop_reclaim =          hammer_vop_reclaim,
 156         .vop_setattr =          hammer_vop_setattr,
 157         .vop_kqfilter =         hammer_vop_fifokqfilter
 158 };
 159
 160 static __inline
 161 void
 162 hammer_knote(struct vnode *vp, int flags)
 163 {
 164         if (flags)
 165                 KNOTE(&vp->v_pollinfo.vpi_kqinfo.ki_note, flags);
 166 }
 167
 168 #ifdef DEBUG_TRUNCATE
 169 struct hammer_inode *HammerTruncIp;
 170 #endif
 171
 172 static int hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
 173                            struct vnode *dvp, struct ucred *cred,
 174                            int flags, int isdir);
 175 static int hammer_vop_strategy_read(struct vop_strategy_args *ap);
 176 static int hammer_vop_strategy_write(struct vop_strategy_args *ap);
 177
 178 #if 0
 179 static
 180 int
 181 hammer_vop_vnoperate(struct vop_generic_args *)
 182 {
 183         return (VOCALL(&hammer_vnode_vops, ap));
 184 }
 185 #endif
 186
 187 /*
 188  * hammer_vop_fsync { vp, waitfor }
 189  *
 190  * fsync() an inode to disk and wait for it to be completely committed
 191  * such that the information would not be undone if a crash occured after
 192  * return.
 193  *
 194  * NOTE: HAMMER's fsync()'s are going to remain expensive until we implement
 195  *       a REDO log.  A sysctl is provided to relax HAMMER's fsync()
 196  *       operation.
 197  *
 198  *       Ultimately the combination of a REDO log and use of fast storage
 199  *       to front-end cluster caches will make fsync fast, but it aint
 200  *       here yet.  And, in anycase, we need real transactional
 201  *       all-or-nothing features which are not restricted to a single file.
 202  */
 203 static
 204 int
 205 hammer_vop_fsync(struct vop_fsync_args *ap)
 206 {
 207         hammer_inode_t ip = VTOI(ap->a_vp);
 208         hammer_mount_t hmp = ip->hmp;
 209         int waitfor = ap->a_waitfor;
 210         int mode;
 211
 212         lwkt_gettoken(&hmp->fs_token);
 213
 214         /*
 215          * Fsync rule relaxation (default is either full synchronous flush
 216          * or REDO semantics with synchronous flush).
 217          */
 218         if (ap->a_flags & VOP_FSYNC_SYSCALL) {
 219                 switch(hammer_fsync_mode) {
 220                 case 0:
 221 mode0:
 222                         /* no REDO, full synchronous flush */
 223                         goto skip;
 224                 case 1:
 225 mode1:
 226                         /* no REDO, full asynchronous flush */
 227                         if (waitfor == MNT_WAIT)
 228                                 waitfor = MNT_NOWAIT;
 229                         goto skip;
 230                 case 2:
 231                         /* REDO semantics, synchronous flush */
 232                         if (hmp->version < HAMMER_VOL_VERSION_FOUR)
 233                                 goto mode0;
 234                         mode = HAMMER_FLUSH_UNDOS_AUTO;
 235                         break;
 236                 case 3:
 237                         /* REDO semantics, relaxed asynchronous flush */
 238                         if (hmp->version < HAMMER_VOL_VERSION_FOUR)
 239                                 goto mode1;
 240                         mode = HAMMER_FLUSH_UNDOS_RELAXED;
 241                         if (waitfor == MNT_WAIT)
 242                                 waitfor = MNT_NOWAIT;
 243                         break;
 244                 case 4:
 245                         /* ignore the fsync() system call */
 246                         lwkt_reltoken(&hmp->fs_token);
 247                         return(0);
 248                 default:
 249                         /* we have to do something */
 250                         mode = HAMMER_FLUSH_UNDOS_RELAXED;
 251                         if (waitfor == MNT_WAIT)
 252                                 waitfor = MNT_NOWAIT;
 253                         break;
 254                 }
 255
 256                 /*
 257                  * Fast fsync only needs to flush the UNDO/REDO fifo if
 258                  * HAMMER_INODE_REDO is non-zero and the only modifications
 259                  * made to the file are write or write-extends.
 260                  */
 261                 if ((ip->flags & HAMMER_INODE_REDO) &&
 262                     (ip->flags & HAMMER_INODE_MODMASK_NOREDO) == 0
 263                 ) {
 264                         ++hammer_count_fsyncs;
 265                         hammer_flusher_flush_undos(hmp, mode);
 266                         ip->redo_count = 0;
 267                         lwkt_reltoken(&hmp->fs_token);
 268                         return(0);
 269                 }
 270
 271                 /*
 272                  * REDO is enabled by fsync(), the idea being we really only
 273                  * want to lay down REDO records when programs are using
 274                  * fsync() heavily.  The first fsync() on the file starts
 275                  * the gravy train going and later fsync()s keep it hot by
 276                  * resetting the redo_count.
 277                  *
 278                  * We weren't running REDOs before now so we have to fall
 279                  * through and do a full fsync of what we have.
 280                  */
 281                 if (hmp->version >= HAMMER_VOL_VERSION_FOUR &&
 282                     (hmp->flags & HAMMER_MOUNT_REDO_RECOVERY_RUN) == 0) {
 283                         ip->flags |= HAMMER_INODE_REDO;
 284                         ip->redo_count = 0;
 285                 }
 286         }
 287 skip:
 288
 289         /*
 290          * Do a full flush sequence.
 291          */
 292         ++hammer_count_fsyncs;
 293         vfsync(ap->a_vp, waitfor, 1, NULL, NULL);
 294         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 295         if (waitfor == MNT_WAIT) {
 296                 vn_unlock(ap->a_vp);
 297                 hammer_wait_inode(ip);
 298                 vn_lock(ap->a_vp, LK_EXCLUSIVE | LK_RETRY);
 299         }
 300         lwkt_reltoken(&hmp->fs_token);
 301         return (ip->error);
 302 }
 303
 304 /*
 305  * hammer_vop_read { vp, uio, ioflag, cred }
 306  *
 307  * MPSAFE (for the cache safe does not require fs_token)
 308  */
 309 static
 310 int
 311 hammer_vop_read(struct vop_read_args *ap)
 312 {
 313         struct hammer_transaction trans;
 314         hammer_inode_t ip;
 315         hammer_mount_t hmp;
 316         off_t offset;
 317         struct buf *bp;
 318         struct uio *uio;
 319         int error;
 320         int n;
 321         int seqcount;
 322         int ioseqcount;
 323         int blksize;
 324         int bigread;
 325         int got_fstoken;
 326
 327         if (ap->a_vp->v_type != VREG)
 328                 return (EINVAL);
 329         ip = VTOI(ap->a_vp);
 330         hmp = ip->hmp;
 331         error = 0;
 332         uio = ap->a_uio;
 333
 334         /*
 335          * Allow the UIO's size to override the sequential heuristic.
 336          */
 337         blksize = hammer_blocksize(uio->uio_offset);
 338         seqcount = (uio->uio_resid + (BKVASIZE - 1)) / BKVASIZE;
 339         ioseqcount = (ap->a_ioflag >> 16);
 340         if (seqcount < ioseqcount)
 341                 seqcount = ioseqcount;
 342
 343         /*
 344          * If reading or writing a huge amount of data we have to break
 345          * atomicy and allow the operation to be interrupted by a signal
 346          * or it can DOS the machine.
 347          */
 348         bigread = (uio->uio_resid > 100 * 1024 * 1024);
 349         got_fstoken = 0;
 350
 351         /*
 352          * Access the data typically in HAMMER_BUFSIZE blocks via the
 353          * buffer cache, but HAMMER may use a variable block size based
 354          * on the offset.
 355          *
 356          * XXX Temporary hack, delay the start transaction while we remain
 357          *     MPSAFE.  NOTE: ino_data.size cannot change while vnode is
 358          *     locked-shared.
 359          */
 360         while (uio->uio_resid > 0 && uio->uio_offset < ip->ino_data.size) {
 361                 int64_t base_offset;
 362                 int64_t file_limit;
 363
 364                 blksize = hammer_blocksize(uio->uio_offset);
 365                 offset = (int)uio->uio_offset & (blksize - 1);
 366                 base_offset = uio->uio_offset - offset;
 367
 368                 if (bigread && (error = hammer_signal_check(ip->hmp)) != 0)
 369                         break;
 370
 371                 /*
 372                  * MPSAFE
 373                  */
 374                 bp = getblk(ap->a_vp, base_offset, blksize, 0, 0);
 375                 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) == B_CACHE) {
 376                         bp->b_flags &= ~B_AGE;
 377                         error = 0;
 378                         goto skip;
 379                 }
 380                 if (ap->a_ioflag & IO_NRDELAY) {
 381                         bqrelse(bp);
 382                         return (EWOULDBLOCK);
 383                 }
 384
 385                 /*
 386                  * MPUNSAFE
 387                  */
 388                 if (got_fstoken == 0) {
 389                         lwkt_gettoken(&hmp->fs_token);
 390                         got_fstoken = 1;
 391                         hammer_start_transaction(&trans, ip->hmp);
 392                 }
 393
 394                 /*
 395                  * NOTE: A valid bp has already been acquired, but was not
 396                  *       B_CACHE.
 397                  */
 398                 if (hammer_cluster_enable) {
 399                         /*
 400                          * Use file_limit to prevent cluster_read() from
 401                          * creating buffers of the wrong block size past
 402                          * the demarc.
 403                          */
 404                         file_limit = ip->ino_data.size;
 405                         if (base_offset < HAMMER_XDEMARC &&
 406                             file_limit > HAMMER_XDEMARC) {
 407                                 file_limit = HAMMER_XDEMARC;
 408                         }
 409                         error = cluster_readx(ap->a_vp,
 410                                              file_limit, base_offset,
 411                                              blksize, uio->uio_resid,
 412                                              seqcount * BKVASIZE, &bp);
 413                 } else {
 414                         error = breadnx(ap->a_vp, base_offset, blksize,
 415                                         NULL, NULL, 0, &bp);
 416                 }
 417                 if (error) {
 418                         brelse(bp);
 419                         break;
 420                 }
 421 skip:
 422                 if ((hammer_debug_io & 0x0001) && (bp->b_flags & B_IODEBUG)) {
 423                         kprintf("doff %016jx read file %016jx@%016jx\n",
 424                                 (intmax_t)bp->b_bio2.bio_offset,
 425                                 (intmax_t)ip->obj_id,
 426                                 (intmax_t)bp->b_loffset);
 427                 }
 428                 bp->b_flags &= ~B_IODEBUG;
 429
 430                 /* bp->b_flags |= B_CLUSTEROK; temporarily disabled */
 431                 n = blksize - offset;
 432                 if (n > uio->uio_resid)
 433                         n = uio->uio_resid;
 434                 if (n > ip->ino_data.size - uio->uio_offset)
 435                         n = (int)(ip->ino_data.size - uio->uio_offset);
 436                 if (got_fstoken)
 437                         lwkt_reltoken(&hmp->fs_token);
 438
 439                 /*
 440                  * Set B_AGE, data has a lower priority than meta-data.
 441                  *
 442                  * Use a hold/unlock/drop sequence to run the uiomove
 443                  * with the buffer unlocked, avoiding deadlocks against
 444                  * read()s on mmap()'d spaces.
 445                  */
 446                 bp->b_flags |= B_AGE;
 447                 bqhold(bp);
 448                 bqrelse(bp);
 449                 error = uiomove((char *)bp->b_data + offset, n, uio);
 450                 bqdrop(bp);
 451
 452                 if (got_fstoken)
 453                         lwkt_gettoken(&hmp->fs_token);
 454
 455                 if (error)
 456                         break;
 457                 hammer_stats_file_read += n;
 458         }
 459
 460         /*
 461          * Try to update the atime with just the inode lock for maximum
 462          * concurrency.  If we can't shortcut it we have to get the full
 463          * blown transaction.
 464          */
 465         if (got_fstoken == 0 && hammer_update_atime_quick(ip) < 0) {
 466                 lwkt_gettoken(&hmp->fs_token);
 467                 got_fstoken = 1;
 468                 hammer_start_transaction(&trans, ip->hmp);
 469         }
 470
 471         if (got_fstoken) {
 472                 if ((ip->flags & HAMMER_INODE_RO) == 0 &&
 473                     (ip->hmp->mp->mnt_flag & MNT_NOATIME) == 0) {
 474                         ip->ino_data.atime = trans.time;
 475                         hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
 476                 }
 477                 hammer_done_transaction(&trans);
 478                 lwkt_reltoken(&hmp->fs_token);
 479         }
 480         return (error);
 481 }
 482
 483 /*
 484  * hammer_vop_write { vp, uio, ioflag, cred }
 485  */
 486 static
 487 int
 488 hammer_vop_write(struct vop_write_args *ap)
 489 {
 490         struct hammer_transaction trans;
 491         struct hammer_inode *ip;
 492         hammer_mount_t hmp;
 493         thread_t td;
 494         struct uio *uio;
 495         int offset;
 496         off_t base_offset;
 497         struct buf *bp;
 498         int kflags;
 499         int error;
 500         int n;
 501         int flags;
 502         int seqcount;
 503         int bigwrite;
 504
 505         if (ap->a_vp->v_type != VREG)
 506                 return (EINVAL);
 507         ip = VTOI(ap->a_vp);
 508         hmp = ip->hmp;
 509         error = 0;
 510         kflags = 0;
 511         seqcount = ap->a_ioflag >> 16;
 512
 513         if (ip->flags & HAMMER_INODE_RO)
 514                 return (EROFS);
 515
 516         /*
 517          * Create a transaction to cover the operations we perform.
 518          */
 519         lwkt_gettoken(&hmp->fs_token);
 520         hammer_start_transaction(&trans, hmp);
 521         uio = ap->a_uio;
 522
 523         /*
 524          * Check append mode
 525          */
 526         if (ap->a_ioflag & IO_APPEND)
 527                 uio->uio_offset = ip->ino_data.size;
 528
 529         /*
 530          * Check for illegal write offsets.  Valid range is 0...2^63-1.
 531          *
 532          * NOTE: the base_off assignment is required to work around what
 533          * I consider to be a GCC-4 optimization bug.
 534          */
 535         if (uio->uio_offset < 0) {
 536                 hammer_done_transaction(&trans);
 537                 lwkt_reltoken(&hmp->fs_token);
 538                 return (EFBIG);
 539         }
 540         base_offset = uio->uio_offset + uio->uio_resid; /* work around gcc-4 */
 541         if (uio->uio_resid > 0 && base_offset <= uio->uio_offset) {
 542                 hammer_done_transaction(&trans);
 543                 lwkt_reltoken(&hmp->fs_token);
 544                 return (EFBIG);
 545         }
 546
 547         if (uio->uio_resid > 0 && (td = uio->uio_td) != NULL && td->td_proc &&
 548             base_offset > td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
 549                 hammer_done_transaction(&trans);
 550                 lwkt_reltoken(&hmp->fs_token);
 551                 lwpsignal(td->td_proc, td->td_lwp, SIGXFSZ);
 552                 return (EFBIG);
 553         }
 554
 555         /*
 556          * If reading or writing a huge amount of data we have to break
 557          * atomicy and allow the operation to be interrupted by a signal
 558          * or it can DOS the machine.
 559          *
 560          * Preset redo_count so we stop generating REDOs earlier if the
 561          * limit is exceeded.
 562          */
 563         bigwrite = (uio->uio_resid > 100 * 1024 * 1024);
 564         if ((ip->flags & HAMMER_INODE_REDO) &&
 565             ip->redo_count < hammer_limit_redo) {
 566                 ip->redo_count += uio->uio_resid;
 567         }
 568
 569         /*
 570          * Access the data typically in HAMMER_BUFSIZE blocks via the
 571          * buffer cache, but HAMMER may use a variable block size based
 572          * on the offset.
 573          */
 574         while (uio->uio_resid > 0) {
 575                 int fixsize = 0;
 576                 int blksize;
 577                 int blkmask;
 578                 int trivial;
 579                 int endofblk;
 580                 off_t nsize;
 581
 582                 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_WRITE)) != 0)
 583                         break;
 584                 if (bigwrite && (error = hammer_signal_check(hmp)) != 0)
 585                         break;
 586
 587                 blksize = hammer_blocksize(uio->uio_offset);
 588
 589                 /*
 590                  * Do not allow HAMMER to blow out the buffer cache.  Very
 591                  * large UIOs can lockout other processes due to bwillwrite()
 592                  * mechanics.
 593                  *
 594                  * The hammer inode is not locked during these operations.
 595                  * The vnode is locked which can interfere with the pageout
 596                  * daemon for non-UIO_NOCOPY writes but should not interfere
 597                  * with the buffer cache.  Even so, we cannot afford to
 598                  * allow the pageout daemon to build up too many dirty buffer
 599                  * cache buffers.
 600                  *
 601                  * Only call this if we aren't being recursively called from
 602                  * a virtual disk device (vn), else we may deadlock.
 603                  */
 604                 if ((ap->a_ioflag & IO_RECURSE) == 0)
 605                         bwillwrite(blksize);
 606
 607                 /*
 608                  * Control the number of pending records associated with
 609                  * this inode.  If too many have accumulated start a
 610                  * flush.  Try to maintain a pipeline with the flusher.
 611                  *
 612                  * NOTE: It is possible for other sources to grow the
 613                  *       records but not necessarily issue another flush,
 614                  *       so use a timeout and ensure that a re-flush occurs.
 615                  */
 616                 if (ip->rsv_recs >= hammer_limit_inode_recs) {
 617                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 618                         while (ip->rsv_recs >= hammer_limit_inode_recs * 2) {
 619                                 ip->flags |= HAMMER_INODE_RECSW;
 620                                 tsleep(&ip->rsv_recs, 0, "hmrwww", hz);
 621                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 622                         }
 623                 }
 624
 625 #if 0
 626                 /*
 627                  * Do not allow HAMMER to blow out system memory by
 628                  * accumulating too many records.   Records are so well
 629                  * decoupled from the buffer cache that it is possible
 630                  * for userland to push data out to the media via
 631                  * direct-write, but build up the records queued to the
 632                  * backend faster then the backend can flush them out.
 633                  * HAMMER has hit its write limit but the frontend has
 634                  * no pushback to slow it down.
 635                  */
 636                 if (hmp->rsv_recs > hammer_limit_recs / 2) {
 637                         /*
 638                          * Get the inode on the flush list
 639                          */
 640                         if (ip->rsv_recs >= 64)
 641                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 642                         else if (ip->rsv_recs >= 16)
 643                                 hammer_flush_inode(ip, 0);
 644
 645                         /*
 646                          * Keep the flusher going if the system keeps
 647                          * queueing records.
 648                          */
 649                         delta = hmp->count_newrecords -
 650                                 hmp->last_newrecords;
 651                         if (delta < 0 || delta > hammer_limit_recs / 2) {
 652                                 hmp->last_newrecords = hmp->count_newrecords;
 653                                 hammer_sync_hmp(hmp, MNT_NOWAIT);
 654                         }
 655
 656                         /*
 657                          * If we have gotten behind start slowing
 658                          * down the writers.
 659                          */
 660                         delta = (hmp->rsv_recs - hammer_limit_recs) *
 661                                 hz / hammer_limit_recs;
 662                         if (delta > 0)
 663                                 tsleep(&trans, 0, "hmrslo", delta);
 664                 }
 665 #endif
 666
 667                 /*
 668                  * Calculate the blocksize at the current offset and figure
 669                  * out how much we can actually write.
 670                  */
 671                 blkmask = blksize - 1;
 672                 offset = (int)uio->uio_offset & blkmask;
 673                 base_offset = uio->uio_offset & ~(int64_t)blkmask;
 674                 n = blksize - offset;
 675                 if (n > uio->uio_resid) {
 676                         n = uio->uio_resid;
 677                         endofblk = 0;
 678                 } else {
 679                         endofblk = 1;
 680                 }
 681                 nsize = uio->uio_offset + n;
 682                 if (nsize > ip->ino_data.size) {
 683                         if (uio->uio_offset > ip->ino_data.size)
 684                                 trivial = 0;
 685                         else
 686                                 trivial = 1;
 687                         nvextendbuf(ap->a_vp,
 688                                     ip->ino_data.size,
 689                                     nsize,
 690                                     hammer_blocksize(ip->ino_data.size),
 691                                     hammer_blocksize(nsize),
 692                                     hammer_blockoff(ip->ino_data.size),
 693                                     hammer_blockoff(nsize),
 694                                     trivial);
 695                         fixsize = 1;
 696                         kflags |= NOTE_EXTEND;
 697                 }
 698
 699                 if (uio->uio_segflg == UIO_NOCOPY) {
 700                         /*
 701                          * Issuing a write with the same data backing the
 702                          * buffer.  Instantiate the buffer to collect the
 703                          * backing vm pages, then read-in any missing bits.
 704                          *
 705                          * This case is used by vop_stdputpages().
 706                          */
 707                         bp = getblk(ap->a_vp, base_offset,
 708                                     blksize, GETBLK_BHEAVY, 0);
 709                         if ((bp->b_flags & B_CACHE) == 0) {
 710                                 bqrelse(bp);
 711                                 error = bread(ap->a_vp, base_offset,
 712                                               blksize, &bp);
 713                         }
 714                 } else if (offset == 0 && uio->uio_resid >= blksize) {
 715                         /*
 716                          * Even though we are entirely overwriting the buffer
 717                          * we may still have to zero it out to avoid a
 718                          * mmap/write visibility issue.
 719                          */
 720                         bp = getblk(ap->a_vp, base_offset, blksize, GETBLK_BHEAVY, 0);
 721                         if ((bp->b_flags & B_CACHE) == 0)
 722                                 vfs_bio_clrbuf(bp);
 723                 } else if (base_offset >= ip->ino_data.size) {
 724                         /*
 725                          * If the base offset of the buffer is beyond the
 726                          * file EOF, we don't have to issue a read.
 727                          */
 728                         bp = getblk(ap->a_vp, base_offset,
 729                                     blksize, GETBLK_BHEAVY, 0);
 730                         vfs_bio_clrbuf(bp);
 731                 } else {
 732                         /*
 733                          * Partial overwrite, read in any missing bits then
 734                          * replace the portion being written.
 735                          */
 736                         error = bread(ap->a_vp, base_offset, blksize, &bp);
 737                         if (error == 0)
 738                                 bheavy(bp);
 739                 }
 740                 if (error == 0) {
 741                         lwkt_reltoken(&hmp->fs_token);
 742                         error = uiomove(bp->b_data + offset, n, uio);
 743                         lwkt_gettoken(&hmp->fs_token);
 744                 }
 745
 746                 /*
 747                  * Generate REDO records if enabled and redo_count will not
 748                  * exceeded the limit.
 749                  *
 750                  * If redo_count exceeds the limit we stop generating records
 751                  * and clear HAMMER_INODE_REDO.  This will cause the next
 752                  * fsync() to do a full meta-data sync instead of just an
 753                  * UNDO/REDO fifo update.
 754                  *
 755                  * When clearing HAMMER_INODE_REDO any pre-existing REDOs
 756                  * will still be tracked.  The tracks will be terminated
 757                  * when the related meta-data (including possible data
 758                  * modifications which are not tracked via REDO) is
 759                  * flushed.
 760                  */
 761                 if ((ip->flags & HAMMER_INODE_REDO) && error == 0) {
 762                         if (ip->redo_count < hammer_limit_redo) {
 763                                 bp->b_flags |= B_VFSFLAG1;
 764                                 error = hammer_generate_redo(&trans, ip,
 765                                                      base_offset + offset,
 766                                                      HAMMER_REDO_WRITE,
 767                                                      bp->b_data + offset,
 768                                                      (size_t)n);
 769                         } else {
 770                                 ip->flags &= ~HAMMER_INODE_REDO;
 771                         }
 772                 }
 773
 774                 /*
 775                  * If we screwed up we have to undo any VM size changes we
 776                  * made.
 777                  */
 778                 if (error) {
 779                         brelse(bp);
 780                         if (fixsize) {
 781                                 nvtruncbuf(ap->a_vp, ip->ino_data.size,
 782                                           hammer_blocksize(ip->ino_data.size),
 783                                           hammer_blockoff(ip->ino_data.size));
 784                         }
 785                         break;
 786                 }
 787                 kflags |= NOTE_WRITE;
 788                 hammer_stats_file_write += n;
 789                 /* bp->b_flags |= B_CLUSTEROK; temporarily disabled */
 790                 if (ip->ino_data.size < uio->uio_offset) {
 791                         ip->ino_data.size = uio->uio_offset;
 792                         flags = HAMMER_INODE_SDIRTY;
 793                 } else {
 794                         flags = 0;
 795                 }
 796                 ip->ino_data.mtime = trans.time;
 797                 flags |= HAMMER_INODE_MTIME | HAMMER_INODE_BUFS;
 798                 hammer_modify_inode(&trans, ip, flags);
 799
 800                 /*
 801                  * Once we dirty the buffer any cached zone-X offset
 802                  * becomes invalid.  HAMMER NOTE: no-history mode cannot
 803                  * allow overwriting over the same data sector unless
 804                  * we provide UNDOs for the old data, which we don't.
 805                  */
 806                 bp->b_bio2.bio_offset = NOOFFSET;
 807
 808                 /*
 809                  * Final buffer disposition.
 810                  *
 811                  * Because meta-data updates are deferred, HAMMER is
 812                  * especially sensitive to excessive bdwrite()s because
 813                  * the I/O stream is not broken up by disk reads.  So the
 814                  * buffer cache simply cannot keep up.
 815                  *
 816                  * WARNING!  blksize is variable.  cluster_write() is
 817                  *           expected to not blow up if it encounters
 818                  *           buffers that do not match the passed blksize.
 819                  *
 820                  * NOTE!  Hammer shouldn't need to bawrite()/cluster_write().
 821                  *        The ip->rsv_recs check should burst-flush the data.
 822                  *        If we queue it immediately the buf could be left
 823                  *        locked on the device queue for a very long time.
 824                  *
 825                  * NOTE!  To avoid degenerate stalls due to mismatched block
 826                  *        sizes we only honor IO_DIRECT on the write which
 827                  *        abuts the end of the buffer.  However, we must
 828                  *        honor IO_SYNC in case someone is silly enough to
 829                  *        configure a HAMMER file as swap, or when HAMMER
 830                  *        is serving NFS (for commits).  Ick ick.
 831                  */
 832                 bp->b_flags |= B_AGE;
 833                 if (ap->a_ioflag & IO_SYNC) {
 834                         bwrite(bp);
 835                 } else if ((ap->a_ioflag & IO_DIRECT) && endofblk) {
 836                         bawrite(bp);
 837                 } else {
 838 #if 0
 839                 if (offset + n == blksize) {
 840                         if (hammer_cluster_enable == 0 ||
 841                             (ap->a_vp->v_mount->mnt_flag & MNT_NOCLUSTERW)) {
 842                                 bawrite(bp);
 843                         } else {
 844                                 cluster_write(bp, ip->ino_data.size,
 845                                               blksize, seqcount);
 846                         }
 847                 } else {
 848 #endif
 849                         bdwrite(bp);
 850                 }
 851         }
 852         hammer_done_transaction(&trans);
 853         hammer_knote(ap->a_vp, kflags);
 854         lwkt_reltoken(&hmp->fs_token);
 855         return (error);
 856 }
 857
 858 /*
 859  * hammer_vop_access { vp, mode, cred }
 860  *
 861  * MPSAFE - does not require fs_token
 862  */
 863 static
 864 int
 865 hammer_vop_access(struct vop_access_args *ap)
 866 {
 867         struct hammer_inode *ip = VTOI(ap->a_vp);
 868         uid_t uid;
 869         gid_t gid;
 870         int error;
 871
 872         ++hammer_stats_file_iopsr;
 873         uid = hammer_to_unix_xid(&ip->ino_data.uid);
 874         gid = hammer_to_unix_xid(&ip->ino_data.gid);
 875
 876         error = vop_helper_access(ap, uid, gid, ip->ino_data.mode,
 877                                   ip->ino_data.uflags);
 878         return (error);
 879 }
 880
 881 /*
 882  * hammer_vop_advlock { vp, id, op, fl, flags }
 883  *
 884  * MPSAFE - does not require fs_token
 885  */
 886 static
 887 int
 888 hammer_vop_advlock(struct vop_advlock_args *ap)
 889 {
 890         hammer_inode_t ip = VTOI(ap->a_vp);
 891
 892         return (lf_advlock(ap, &ip->advlock, ip->ino_data.size));
 893 }
 894
 895 /*
 896  * hammer_vop_close { vp, fflag }
 897  *
 898  * We can only sync-on-close for normal closes.  XXX disabled for now.
 899  */
 900 static
 901 int
 902 hammer_vop_close(struct vop_close_args *ap)
 903 {
 904 #if 0
 905         struct vnode *vp = ap->a_vp;
 906         hammer_inode_t ip = VTOI(vp);
 907         int waitfor;
 908         if (ip->flags & (HAMMER_INODE_CLOSESYNC|HAMMER_INODE_CLOSEASYNC)) {
 909                 if (vn_islocked(vp) == LK_EXCLUSIVE &&
 910                     (vp->v_flag & (VINACTIVE|VRECLAIMED)) == 0) {
 911                         if (ip->flags & HAMMER_INODE_CLOSESYNC)
 912                                 waitfor = MNT_WAIT;
 913                         else
 914                                 waitfor = MNT_NOWAIT;
 915                         ip->flags &= ~(HAMMER_INODE_CLOSESYNC |
 916                                        HAMMER_INODE_CLOSEASYNC);
 917                         VOP_FSYNC(vp, MNT_NOWAIT, waitfor);
 918                 }
 919         }
 920 #endif
 921         return (vop_stdclose(ap));
 922 }
 923
 924 /*
 925  * hammer_vop_ncreate { nch, dvp, vpp, cred, vap }
 926  *
 927  * The operating system has already ensured that the directory entry
 928  * does not exist and done all appropriate namespace locking.
 929  */
 930 static
 931 int
 932 hammer_vop_ncreate(struct vop_ncreate_args *ap)
 933 {
 934         struct hammer_transaction trans;
 935         struct hammer_inode *dip;
 936         struct hammer_inode *nip;
 937         struct nchandle *nch;
 938         hammer_mount_t hmp;
 939         int error;
 940
 941         nch = ap->a_nch;
 942         dip = VTOI(ap->a_dvp);
 943         hmp = dip->hmp;
 944
 945         if (dip->flags & HAMMER_INODE_RO)
 946                 return (EROFS);
 947         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
 948                 return (error);
 949
 950         /*
 951          * Create a transaction to cover the operations we perform.
 952          */
 953         lwkt_gettoken(&hmp->fs_token);
 954         hammer_start_transaction(&trans, hmp);
 955         ++hammer_stats_file_iopsw;
 956
 957         /*
 958          * Create a new filesystem object of the requested type.  The
 959          * returned inode will be referenced and shared-locked to prevent
 960          * it from being moved to the flusher.
 961          */
 962         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
 963                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
 964                                     NULL, &nip);
 965         if (error) {
 966                 hkprintf("hammer_create_inode error %d\n", error);
 967                 hammer_done_transaction(&trans);
 968                 *ap->a_vpp = NULL;
 969                 lwkt_reltoken(&hmp->fs_token);
 970                 return (error);
 971         }
 972
 973         /*
 974          * Add the new filesystem object to the directory.  This will also
 975          * bump the inode's link count.
 976          */
 977         error = hammer_ip_add_directory(&trans, dip,
 978                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
 979                                         nip);
 980         if (error)
 981                 hkprintf("hammer_ip_add_directory error %d\n", error);
 982
 983         /*
 984          * Finish up.
 985          */
 986         if (error) {
 987                 hammer_rel_inode(nip, 0);
 988                 hammer_done_transaction(&trans);
 989                 *ap->a_vpp = NULL;
 990         } else {
 991                 error = hammer_get_vnode(nip, ap->a_vpp);
 992                 hammer_done_transaction(&trans);
 993                 hammer_rel_inode(nip, 0);
 994                 if (error == 0) {
 995                         cache_setunresolved(ap->a_nch);
 996                         cache_setvp(ap->a_nch, *ap->a_vpp);
 997                 }
 998                 hammer_knote(ap->a_dvp, NOTE_WRITE);
 999         }
1000         lwkt_reltoken(&hmp->fs_token);
1001         return (error);
1002 }
1003
1004 /*
1005  * hammer_vop_getattr { vp, vap }
1006  *
1007  * Retrieve an inode's attribute information.  When accessing inodes
1008  * historically we fake the atime field to ensure consistent results.
1009  * The atime field is stored in the B-Tree element and allowed to be
1010  * updated without cycling the element.
1011  *
1012  * MPSAFE - does not require fs_token
1013  */
1014 static
1015 int
1016 hammer_vop_getattr(struct vop_getattr_args *ap)
1017 {
1018         struct hammer_inode *ip = VTOI(ap->a_vp);
1019         struct vattr *vap = ap->a_vap;
1020
1021         /*
1022          * We want the fsid to be different when accessing a filesystem
1023          * with different as-of's so programs like diff don't think
1024          * the files are the same.
1025          *
1026          * We also want the fsid to be the same when comparing snapshots,
1027          * or when comparing mirrors (which might be backed by different
1028          * physical devices).  HAMMER fsids are based on the PFS's
1029          * shared_uuid field.
1030          *
1031          * XXX there is a chance of collision here.  The va_fsid reported
1032          * by stat is different from the more involved fsid used in the
1033          * mount structure.
1034          */
1035         ++hammer_stats_file_iopsr;
1036         hammer_lock_sh(&ip->lock);
1037         vap->va_fsid = ip->pfsm->fsid_udev ^ (u_int32_t)ip->obj_asof ^
1038                        (u_int32_t)(ip->obj_asof >> 32);
1039
1040         vap->va_fileid = ip->ino_leaf.base.obj_id;
1041         vap->va_mode = ip->ino_data.mode;
1042         vap->va_nlink = ip->ino_data.nlinks;
1043         vap->va_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1044         vap->va_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1045         vap->va_rmajor = 0;
1046         vap->va_rminor = 0;
1047         vap->va_size = ip->ino_data.size;
1048
1049         /*
1050          * Special case for @@PFS softlinks.  The actual size of the
1051          * expanded softlink is "@@0x%016llx:%05d" == 26 bytes.
1052          * or for MAX_TID is    "@@-1:%05d" == 10 bytes.
1053          */
1054         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_SOFTLINK &&
1055             ip->ino_data.size == 10 &&
1056             ip->obj_asof == HAMMER_MAX_TID &&
1057             ip->obj_localization == 0 &&
1058             strncmp(ip->ino_data.ext.symlink, "@@PFS", 5) == 0) {
1059                     if (ip->pfsm->pfsd.mirror_flags & HAMMER_PFSD_SLAVE)
1060                             vap->va_size = 26;
1061                     else
1062                             vap->va_size = 10;
1063         }
1064
1065         /*
1066          * We must provide a consistent atime and mtime for snapshots
1067          * so people can do a 'tar cf - ... | md5' on them and get
1068          * consistent results.
1069          */
1070         if (ip->flags & HAMMER_INODE_RO) {
1071                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_atime);
1072                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_mtime);
1073         } else {
1074                 hammer_time_to_timespec(ip->ino_data.atime, &vap->va_atime);
1075                 hammer_time_to_timespec(ip->ino_data.mtime, &vap->va_mtime);
1076         }
1077         hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_ctime);
1078         vap->va_flags = ip->ino_data.uflags;
1079         vap->va_gen = 1;        /* hammer inums are unique for all time */
1080         vap->va_blocksize = HAMMER_BUFSIZE;
1081         if (ip->ino_data.size >= HAMMER_XDEMARC) {
1082                 vap->va_bytes = (ip->ino_data.size + HAMMER_XBUFMASK64) &
1083                                 ~HAMMER_XBUFMASK64;
1084         } else if (ip->ino_data.size > HAMMER_BUFSIZE / 2) {
1085                 vap->va_bytes = (ip->ino_data.size + HAMMER_BUFMASK64) &
1086                                 ~HAMMER_BUFMASK64;
1087         } else {
1088                 vap->va_bytes = (ip->ino_data.size + 15) & ~15;
1089         }
1090
1091         vap->va_type = hammer_get_vnode_type(ip->ino_data.obj_type);
1092         vap->va_filerev = 0;    /* XXX */
1093         vap->va_uid_uuid = ip->ino_data.uid;
1094         vap->va_gid_uuid = ip->ino_data.gid;
1095         vap->va_fsid_uuid = ip->hmp->fsid;
1096         vap->va_vaflags = VA_UID_UUID_VALID | VA_GID_UUID_VALID |
1097                           VA_FSID_UUID_VALID;
1098
1099         switch (ip->ino_data.obj_type) {
1100         case HAMMER_OBJTYPE_CDEV:
1101         case HAMMER_OBJTYPE_BDEV:
1102                 vap->va_rmajor = ip->ino_data.rmajor;
1103                 vap->va_rminor = ip->ino_data.rminor;
1104                 break;
1105         default:
1106                 break;
1107         }
1108         hammer_unlock(&ip->lock);
1109         return(0);
1110 }
1111
1112 /*
1113  * hammer_vop_nresolve { nch, dvp, cred }
1114  *
1115  * Locate the requested directory entry.
1116  */
1117 static
1118 int
1119 hammer_vop_nresolve(struct vop_nresolve_args *ap)
1120 {
1121         struct hammer_transaction trans;
1122         struct namecache *ncp;
1123         hammer_mount_t hmp;
1124         hammer_inode_t dip;
1125         hammer_inode_t ip;
1126         hammer_tid_t asof;
1127         struct hammer_cursor cursor;
1128         struct vnode *vp;
1129         int64_t namekey;
1130         int error;
1131         int i;
1132         int nlen;
1133         int flags;
1134         int ispfs;
1135         int64_t obj_id;
1136         u_int32_t localization;
1137         u_int32_t max_iterations;
1138
1139         /*
1140          * Misc initialization, plus handle as-of name extensions.  Look for
1141          * the '@@' extension.  Note that as-of files and directories cannot
1142          * be modified.
1143          */
1144         dip = VTOI(ap->a_dvp);
1145         ncp = ap->a_nch->ncp;
1146         asof = dip->obj_asof;
1147         localization = dip->obj_localization;   /* for code consistency */
1148         nlen = ncp->nc_nlen;
1149         flags = dip->flags & HAMMER_INODE_RO;
1150         ispfs = 0;
1151         hmp = dip->hmp;
1152
1153         lwkt_gettoken(&hmp->fs_token);
1154         hammer_simple_transaction(&trans, hmp);
1155         ++hammer_stats_file_iopsr;
1156
1157         for (i = 0; i < nlen; ++i) {
1158                 if (ncp->nc_name[i] == '@' && ncp->nc_name[i+1] == '@') {
1159                         error = hammer_str_to_tid(ncp->nc_name + i + 2,
1160                                                   &ispfs, &asof, &localization);
1161                         if (error != 0) {
1162                                 i = nlen;
1163                                 break;
1164                         }
1165                         if (asof != HAMMER_MAX_TID)
1166                                 flags |= HAMMER_INODE_RO;
1167                         break;
1168                 }
1169         }
1170         nlen = i;
1171
1172         /*
1173          * If this is a PFS softlink we dive into the PFS
1174          */
1175         if (ispfs && nlen == 0) {
1176                 ip = hammer_get_inode(&trans, dip, HAMMER_OBJID_ROOT,
1177                                       asof, localization,
1178                                       flags, &error);
1179                 if (error == 0) {
1180                         error = hammer_get_vnode(ip, &vp);
1181                         hammer_rel_inode(ip, 0);
1182                 } else {
1183                         vp = NULL;
1184                 }
1185                 if (error == 0) {
1186                         vn_unlock(vp);
1187                         cache_setvp(ap->a_nch, vp);
1188                         vrele(vp);
1189                 }
1190                 goto done;
1191         }
1192
1193         /*
1194          * If there is no path component the time extension is relative to dip.
1195          * e.g. "fubar/@@<snapshot>"
1196          *
1197          * "." is handled by the kernel, but ".@@<snapshot>" is not.
1198          * e.g. "fubar/.@@<snapshot>"
1199          *
1200          * ".." is handled by the kernel.  We do not currently handle
1201          * "..@<snapshot>".
1202          */
1203         if (nlen == 0 || (nlen == 1 && ncp->nc_name[0] == '.')) {
1204                 ip = hammer_get_inode(&trans, dip, dip->obj_id,
1205                                       asof, dip->obj_localization,
1206                                       flags, &error);
1207                 if (error == 0) {
1208                         error = hammer_get_vnode(ip, &vp);
1209                         hammer_rel_inode(ip, 0);
1210                 } else {
1211                         vp = NULL;
1212                 }
1213                 if (error == 0) {
1214                         vn_unlock(vp);
1215                         cache_setvp(ap->a_nch, vp);
1216                         vrele(vp);
1217                 }
1218                 goto done;
1219         }
1220
1221         /*
1222          * Calculate the namekey and setup the key range for the scan.  This
1223          * works kinda like a chained hash table where the lower 32 bits
1224          * of the namekey synthesize the chain.
1225          *
1226          * The key range is inclusive of both key_beg and key_end.
1227          */
1228         namekey = hammer_directory_namekey(dip, ncp->nc_name, nlen,
1229                                            &max_iterations);
1230
1231         error = hammer_init_cursor(&trans, &cursor, &dip->cache[1], dip);
1232         cursor.key_beg.localization = dip->obj_localization +
1233                                       hammer_dir_localization(dip);
1234         cursor.key_beg.obj_id = dip->obj_id;
1235         cursor.key_beg.key = namekey;
1236         cursor.key_beg.create_tid = 0;
1237         cursor.key_beg.delete_tid = 0;
1238         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1239         cursor.key_beg.obj_type = 0;
1240
1241         cursor.key_end = cursor.key_beg;
1242         cursor.key_end.key += max_iterations;
1243         cursor.asof = asof;
1244         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1245
1246         /*
1247          * Scan all matching records (the chain), locate the one matching
1248          * the requested path component.
1249          *
1250          * The hammer_ip_*() functions merge in-memory records with on-disk
1251          * records for the purposes of the search.
1252          */
1253         obj_id = 0;
1254         localization = HAMMER_DEF_LOCALIZATION;
1255
1256         if (error == 0) {
1257                 error = hammer_ip_first(&cursor);
1258                 while (error == 0) {
1259                         error = hammer_ip_resolve_data(&cursor);
1260                         if (error)
1261                                 break;
1262                         if (nlen == cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF &&
1263                             bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1264                                 obj_id = cursor.data->entry.obj_id;
1265                                 localization = cursor.data->entry.localization;
1266                                 break;
1267                         }
1268                         error = hammer_ip_next(&cursor);
1269                 }
1270         }
1271         hammer_done_cursor(&cursor);
1272
1273         /*
1274          * Lookup the obj_id.  This should always succeed.  If it does not
1275          * the filesystem may be damaged and we return a dummy inode.
1276          */
1277         if (error == 0) {
1278                 ip = hammer_get_inode(&trans, dip, obj_id,
1279                                       asof, localization,
1280                                       flags, &error);
1281                 if (error == ENOENT) {
1282                         kprintf("HAMMER: WARNING: Missing "
1283                                 "inode for dirent \"%s\"\n"
1284                                 "\tobj_id = %016llx, asof=%016llx, lo=%08x\n",
1285                                 ncp->nc_name,
1286                                 (long long)obj_id, (long long)asof,
1287                                 localization);
1288                         error = 0;
1289                         ip = hammer_get_dummy_inode(&trans, dip, obj_id,
1290                                                     asof, localization,
1291                                                     flags, &error);
1292                 }
1293                 if (error == 0) {
1294                         error = hammer_get_vnode(ip, &vp);
1295                         hammer_rel_inode(ip, 0);
1296                 } else {
1297                         vp = NULL;
1298                 }
1299                 if (error == 0) {
1300                         vn_unlock(vp);
1301                         cache_setvp(ap->a_nch, vp);
1302                         vrele(vp);
1303                 }
1304         } else if (error == ENOENT) {
1305                 cache_setvp(ap->a_nch, NULL);
1306         }
1307 done:
1308         hammer_done_transaction(&trans);
1309         lwkt_reltoken(&hmp->fs_token);
1310         return (error);
1311 }
1312
1313 /*
1314  * hammer_vop_nlookupdotdot { dvp, vpp, cred }
1315  *
1316  * Locate the parent directory of a directory vnode.
1317  *
1318  * dvp is referenced but not locked.  *vpp must be returned referenced and
1319  * locked.  A parent_obj_id of 0 does not necessarily indicate that we are
1320  * at the root, instead it could indicate that the directory we were in was
1321  * removed.
1322  *
1323  * NOTE: as-of sequences are not linked into the directory structure.  If
1324  * we are at the root with a different asof then the mount point, reload
1325  * the same directory with the mount point's asof.   I'm not sure what this
1326  * will do to NFS.  We encode ASOF stamps in NFS file handles so it might not
1327  * get confused, but it hasn't been tested.
1328  */
1329 static
1330 int
1331 hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *ap)
1332 {
1333         struct hammer_transaction trans;
1334         struct hammer_inode *dip;
1335         struct hammer_inode *ip;
1336         hammer_mount_t hmp;
1337         int64_t parent_obj_id;
1338         u_int32_t parent_obj_localization;
1339         hammer_tid_t asof;
1340         int error;
1341
1342         dip = VTOI(ap->a_dvp);
1343         asof = dip->obj_asof;
1344         hmp = dip->hmp;
1345
1346         /*
1347          * Whos are parent?  This could be the root of a pseudo-filesystem
1348          * whos parent is in another localization domain.
1349          */
1350         lwkt_gettoken(&hmp->fs_token);
1351         parent_obj_id = dip->ino_data.parent_obj_id;
1352         if (dip->obj_id == HAMMER_OBJID_ROOT)
1353                 parent_obj_localization = dip->ino_data.ext.obj.parent_obj_localization;
1354         else
1355                 parent_obj_localization = dip->obj_localization;
1356
1357         if (parent_obj_id == 0) {
1358                 if (dip->obj_id == HAMMER_OBJID_ROOT &&
1359                    asof != hmp->asof) {
1360                         parent_obj_id = dip->obj_id;
1361                         asof = hmp->asof;
1362                         *ap->a_fakename = kmalloc(19, M_TEMP, M_WAITOK);
1363                         ksnprintf(*ap->a_fakename, 19, "0x%016llx",
1364                                   (long long)dip->obj_asof);
1365                 } else {
1366                         *ap->a_vpp = NULL;
1367                         lwkt_reltoken(&hmp->fs_token);
1368                         return ENOENT;
1369                 }
1370         }
1371
1372         hammer_simple_transaction(&trans, hmp);
1373         ++hammer_stats_file_iopsr;
1374
1375         ip = hammer_get_inode(&trans, dip, parent_obj_id,
1376                               asof, parent_obj_localization,
1377                               dip->flags, &error);
1378         if (ip) {
1379                 error = hammer_get_vnode(ip, ap->a_vpp);
1380                 hammer_rel_inode(ip, 0);
1381         } else {
1382                 *ap->a_vpp = NULL;
1383         }
1384         hammer_done_transaction(&trans);
1385         lwkt_reltoken(&hmp->fs_token);
1386         return (error);
1387 }
1388
1389 /*
1390  * hammer_vop_nlink { nch, dvp, vp, cred }
1391  */
1392 static
1393 int
1394 hammer_vop_nlink(struct vop_nlink_args *ap)
1395 {
1396         struct hammer_transaction trans;
1397         struct hammer_inode *dip;
1398         struct hammer_inode *ip;
1399         struct nchandle *nch;
1400         hammer_mount_t hmp;
1401         int error;
1402
1403         if (ap->a_dvp->v_mount != ap->a_vp->v_mount)
1404                 return(EXDEV);
1405
1406         nch = ap->a_nch;
1407         dip = VTOI(ap->a_dvp);
1408         ip = VTOI(ap->a_vp);
1409         hmp = dip->hmp;
1410
1411         if (dip->obj_localization != ip->obj_localization)
1412                 return(EXDEV);
1413
1414         if (dip->flags & HAMMER_INODE_RO)
1415                 return (EROFS);
1416         if (ip->flags & HAMMER_INODE_RO)
1417                 return (EROFS);
1418         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1419                 return (error);
1420
1421         /*
1422          * Create a transaction to cover the operations we perform.
1423          */
1424         lwkt_gettoken(&hmp->fs_token);
1425         hammer_start_transaction(&trans, hmp);
1426         ++hammer_stats_file_iopsw;
1427
1428         /*
1429          * Add the filesystem object to the directory.  Note that neither
1430          * dip nor ip are referenced or locked, but their vnodes are
1431          * referenced.  This function will bump the inode's link count.
1432          */
1433         error = hammer_ip_add_directory(&trans, dip,
1434                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1435                                         ip);
1436
1437         /*
1438          * Finish up.
1439          */
1440         if (error == 0) {
1441                 cache_setunresolved(nch);
1442                 cache_setvp(nch, ap->a_vp);
1443         }
1444         hammer_done_transaction(&trans);
1445         hammer_knote(ap->a_vp, NOTE_LINK);
1446         hammer_knote(ap->a_dvp, NOTE_WRITE);
1447         lwkt_reltoken(&hmp->fs_token);
1448         return (error);
1449 }
1450
1451 /*
1452  * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
1453  *
1454  * The operating system has already ensured that the directory entry
1455  * does not exist and done all appropriate namespace locking.
1456  */
1457 static
1458 int
1459 hammer_vop_nmkdir(struct vop_nmkdir_args *ap)
1460 {
1461         struct hammer_transaction trans;
1462         struct hammer_inode *dip;
1463         struct hammer_inode *nip;
1464         struct nchandle *nch;
1465         hammer_mount_t hmp;
1466         int error;
1467
1468         nch = ap->a_nch;
1469         dip = VTOI(ap->a_dvp);
1470         hmp = dip->hmp;
1471
1472         if (dip->flags & HAMMER_INODE_RO)
1473                 return (EROFS);
1474         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1475                 return (error);
1476
1477         /*
1478          * Create a transaction to cover the operations we perform.
1479          */
1480         lwkt_gettoken(&hmp->fs_token);
1481         hammer_start_transaction(&trans, hmp);
1482         ++hammer_stats_file_iopsw;
1483
1484         /*
1485          * Create a new filesystem object of the requested type.  The
1486          * returned inode will be referenced but not locked.
1487          */
1488         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1489                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1490                                     NULL, &nip);
1491         if (error) {
1492                 hkprintf("hammer_mkdir error %d\n", error);
1493                 hammer_done_transaction(&trans);
1494                 *ap->a_vpp = NULL;
1495                 lwkt_reltoken(&hmp->fs_token);
1496                 return (error);
1497         }
1498         /*
1499          * Add the new filesystem object to the directory.  This will also
1500          * bump the inode's link count.
1501          */
1502         error = hammer_ip_add_directory(&trans, dip,
1503                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1504                                         nip);
1505         if (error)
1506                 hkprintf("hammer_mkdir (add) error %d\n", error);
1507
1508         /*
1509          * Finish up.
1510          */
1511         if (error) {
1512                 hammer_rel_inode(nip, 0);
1513                 *ap->a_vpp = NULL;
1514         } else {
1515                 error = hammer_get_vnode(nip, ap->a_vpp);
1516                 hammer_rel_inode(nip, 0);
1517                 if (error == 0) {
1518                         cache_setunresolved(ap->a_nch);
1519                         cache_setvp(ap->a_nch, *ap->a_vpp);
1520                 }
1521         }
1522         hammer_done_transaction(&trans);
1523         if (error == 0)
1524                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
1525         lwkt_reltoken(&hmp->fs_token);
1526         return (error);
1527 }
1528
1529 /*
1530  * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1531  *
1532  * The operating system has already ensured that the directory entry
1533  * does not exist and done all appropriate namespace locking.
1534  */
1535 static
1536 int
1537 hammer_vop_nmknod(struct vop_nmknod_args *ap)
1538 {
1539         struct hammer_transaction trans;
1540         struct hammer_inode *dip;
1541         struct hammer_inode *nip;
1542         struct nchandle *nch;
1543         hammer_mount_t hmp;
1544         int error;
1545
1546         nch = ap->a_nch;
1547         dip = VTOI(ap->a_dvp);
1548         hmp = dip->hmp;
1549
1550         if (dip->flags & HAMMER_INODE_RO)
1551                 return (EROFS);
1552         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1553                 return (error);
1554
1555         /*
1556          * Create a transaction to cover the operations we perform.
1557          */
1558         lwkt_gettoken(&hmp->fs_token);
1559         hammer_start_transaction(&trans, hmp);
1560         ++hammer_stats_file_iopsw;
1561
1562         /*
1563          * Create a new filesystem object of the requested type.  The
1564          * returned inode will be referenced but not locked.
1565          *
1566          * If mknod specifies a directory a pseudo-fs is created.
1567          */
1568         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1569                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1570                                     NULL, &nip);
1571         if (error) {
1572                 hammer_done_transaction(&trans);
1573                 *ap->a_vpp = NULL;
1574                 lwkt_reltoken(&hmp->fs_token);
1575                 return (error);
1576         }
1577
1578         /*
1579          * Add the new filesystem object to the directory.  This will also
1580          * bump the inode's link count.
1581          */
1582         error = hammer_ip_add_directory(&trans, dip,
1583                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1584                                         nip);
1585
1586         /*
1587          * Finish up.
1588          */
1589         if (error) {
1590                 hammer_rel_inode(nip, 0);
1591                 *ap->a_vpp = NULL;
1592         } else {
1593                 error = hammer_get_vnode(nip, ap->a_vpp);
1594                 hammer_rel_inode(nip, 0);
1595                 if (error == 0) {
1596                         cache_setunresolved(ap->a_nch);
1597                         cache_setvp(ap->a_nch, *ap->a_vpp);
1598                 }
1599         }
1600         hammer_done_transaction(&trans);
1601         if (error == 0)
1602                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1603         lwkt_reltoken(&hmp->fs_token);
1604         return (error);
1605 }
1606
1607 /*
1608  * hammer_vop_open { vp, mode, cred, fp }
1609  *
1610  * MPSAFE (does not require fs_token)
1611  */
1612 static
1613 int
1614 hammer_vop_open(struct vop_open_args *ap)
1615 {
1616         hammer_inode_t ip;
1617
1618         ++hammer_stats_file_iopsr;
1619         ip = VTOI(ap->a_vp);
1620
1621         if ((ap->a_mode & FWRITE) && (ip->flags & HAMMER_INODE_RO))
1622                 return (EROFS);
1623         return(vop_stdopen(ap));
1624 }
1625
1626 /*
1627  * hammer_vop_print { vp }
1628  */
1629 static
1630 int
1631 hammer_vop_print(struct vop_print_args *ap)
1632 {
1633         return EOPNOTSUPP;
1634 }
1635
1636 /*
1637  * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1638  */
1639 static
1640 int
1641 hammer_vop_readdir(struct vop_readdir_args *ap)
1642 {
1643         struct hammer_transaction trans;
1644         struct hammer_cursor cursor;
1645         struct hammer_inode *ip;
1646         hammer_mount_t hmp;
1647         struct uio *uio;
1648         hammer_base_elm_t base;
1649         int error;
1650         int cookie_index;
1651         int ncookies;
1652         off_t *cookies;
1653         off_t saveoff;
1654         int r;
1655         int dtype;
1656
1657         ++hammer_stats_file_iopsr;
1658         ip = VTOI(ap->a_vp);
1659         uio = ap->a_uio;
1660         saveoff = uio->uio_offset;
1661         hmp = ip->hmp;
1662
1663         if (ap->a_ncookies) {
1664                 ncookies = uio->uio_resid / 16 + 1;
1665                 if (ncookies > 1024)
1666                         ncookies = 1024;
1667                 cookies = kmalloc(ncookies * sizeof(off_t), M_TEMP, M_WAITOK);
1668                 cookie_index = 0;
1669         } else {
1670                 ncookies = -1;
1671                 cookies = NULL;
1672                 cookie_index = 0;
1673         }
1674
1675         lwkt_gettoken(&hmp->fs_token);
1676         hammer_simple_transaction(&trans, hmp);
1677
1678         /*
1679          * Handle artificial entries
1680          *
1681          * It should be noted that the minimum value for a directory
1682          * hash key on-media is 0x0000000100000000, so we can use anything
1683          * less then that to represent our 'special' key space.
1684          */
1685         error = 0;
1686         if (saveoff == 0) {
1687                 r = vop_write_dirent(&error, uio, ip->obj_id, DT_DIR, 1, ".");
1688                 if (r)
1689                         goto done;
1690                 if (cookies)
1691                         cookies[cookie_index] = saveoff;
1692                 ++saveoff;
1693                 ++cookie_index;
1694                 if (cookie_index == ncookies)
1695                         goto done;
1696         }
1697         if (saveoff == 1) {
1698                 if (ip->ino_data.parent_obj_id) {
1699                         r = vop_write_dirent(&error, uio,
1700                                              ip->ino_data.parent_obj_id,
1701                                              DT_DIR, 2, "..");
1702                 } else {
1703                         r = vop_write_dirent(&error, uio,
1704                                              ip->obj_id, DT_DIR, 2, "..");
1705                 }
1706                 if (r)
1707                         goto done;
1708                 if (cookies)
1709                         cookies[cookie_index] = saveoff;
1710                 ++saveoff;
1711                 ++cookie_index;
1712                 if (cookie_index == ncookies)
1713                         goto done;
1714         }
1715
1716         /*
1717          * Key range (begin and end inclusive) to scan.  Directory keys
1718          * directly translate to a 64 bit 'seek' position.
1719          */
1720         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1721         cursor.key_beg.localization = ip->obj_localization +
1722                                       hammer_dir_localization(ip);
1723         cursor.key_beg.obj_id = ip->obj_id;
1724         cursor.key_beg.create_tid = 0;
1725         cursor.key_beg.delete_tid = 0;
1726         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1727         cursor.key_beg.obj_type = 0;
1728         cursor.key_beg.key = saveoff;
1729
1730         cursor.key_end = cursor.key_beg;
1731         cursor.key_end.key = HAMMER_MAX_KEY;
1732         cursor.asof = ip->obj_asof;
1733         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1734
1735         error = hammer_ip_first(&cursor);
1736
1737         while (error == 0) {
1738                 error = hammer_ip_resolve_data(&cursor);
1739                 if (error)
1740                         break;
1741                 base = &cursor.leaf->base;
1742                 saveoff = base->key;
1743                 KKASSERT(cursor.leaf->data_len > HAMMER_ENTRY_NAME_OFF);
1744
1745                 if (base->obj_id != ip->obj_id)
1746                         panic("readdir: bad record at %p", cursor.node);
1747
1748                 /*
1749                  * Convert pseudo-filesystems into softlinks
1750                  */
1751                 dtype = hammer_get_dtype(cursor.leaf->base.obj_type);
1752                 r = vop_write_dirent(
1753                              &error, uio, cursor.data->entry.obj_id,
1754                              dtype,
1755                              cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF ,
1756                              (void *)cursor.data->entry.name);
1757                 if (r)
1758                         break;
1759                 ++saveoff;
1760                 if (cookies)
1761                         cookies[cookie_index] = base->key;
1762                 ++cookie_index;
1763                 if (cookie_index == ncookies)
1764                         break;
1765                 error = hammer_ip_next(&cursor);
1766         }
1767         hammer_done_cursor(&cursor);
1768
1769 done:
1770         hammer_done_transaction(&trans);
1771
1772         if (ap->a_eofflag)
1773                 *ap->a_eofflag = (error == ENOENT);
1774         uio->uio_offset = saveoff;
1775         if (error && cookie_index == 0) {
1776                 if (error == ENOENT)
1777                         error = 0;
1778                 if (cookies) {
1779                         kfree(cookies, M_TEMP);
1780                         *ap->a_ncookies = 0;
1781                         *ap->a_cookies = NULL;
1782                 }
1783         } else {
1784                 if (error == ENOENT)
1785                         error = 0;
1786                 if (cookies) {
1787                         *ap->a_ncookies = cookie_index;
1788                         *ap->a_cookies = cookies;
1789                 }
1790         }
1791         lwkt_reltoken(&hmp->fs_token);
1792         return(error);
1793 }
1794
1795 /*
1796  * hammer_vop_readlink { vp, uio, cred }
1797  */
1798 static
1799 int
1800 hammer_vop_readlink(struct vop_readlink_args *ap)
1801 {
1802         struct hammer_transaction trans;
1803         struct hammer_cursor cursor;
1804         struct hammer_inode *ip;
1805         hammer_mount_t hmp;
1806         char buf[32];
1807         u_int32_t localization;
1808         hammer_pseudofs_inmem_t pfsm;
1809         int error;
1810
1811         ip = VTOI(ap->a_vp);
1812         hmp = ip->hmp;
1813
1814         lwkt_gettoken(&hmp->fs_token);
1815
1816         /*
1817          * Shortcut if the symlink data was stuffed into ino_data.
1818          *
1819          * Also expand special "@@PFS%05d" softlinks (expansion only
1820          * occurs for non-historical (current) accesses made from the
1821          * primary filesystem).
1822          */
1823         if (ip->ino_data.size <= HAMMER_INODE_BASESYMLEN) {
1824                 char *ptr;
1825                 int bytes;
1826
1827                 ptr = ip->ino_data.ext.symlink;
1828                 bytes = (int)ip->ino_data.size;
1829                 if (bytes == 10 &&
1830                     ip->obj_asof == HAMMER_MAX_TID &&
1831                     ip->obj_localization == 0 &&
1832                     strncmp(ptr, "@@PFS", 5) == 0) {
1833                         hammer_simple_transaction(&trans, hmp);
1834                         bcopy(ptr + 5, buf, 5);
1835                         buf[5] = 0;
1836                         localization = strtoul(buf, NULL, 10) << 16;
1837                         pfsm = hammer_load_pseudofs(&trans, localization,
1838                                                     &error);
1839                         if (error == 0) {
1840                                 if (pfsm->pfsd.mirror_flags &
1841                                     HAMMER_PFSD_SLAVE) {
1842                                         /* vap->va_size == 26 */
1843                                         ksnprintf(buf, sizeof(buf),
1844                                                   "@@0x%016llx:%05d",
1845                                                   (long long)pfsm->pfsd.sync_end_tid,
1846                                                   localization >> 16);
1847                                 } else {
1848                                         /* vap->va_size == 10 */
1849                                         ksnprintf(buf, sizeof(buf),
1850                                                   "@@-1:%05d",
1851                                                   localization >> 16);
1852 #if 0
1853                                         ksnprintf(buf, sizeof(buf),
1854                                                   "@@0x%016llx:%05d",
1855                                                   (long long)HAMMER_MAX_TID,
1856                                                   localization >> 16);
1857 #endif
1858                                 }
1859                                 ptr = buf;
1860                                 bytes = strlen(buf);
1861                         }
1862                         if (pfsm)
1863                                 hammer_rel_pseudofs(hmp, pfsm);
1864                         hammer_done_transaction(&trans);
1865                 }
1866                 error = uiomove(ptr, bytes, ap->a_uio);
1867                 lwkt_reltoken(&hmp->fs_token);
1868                 return(error);
1869         }
1870
1871         /*
1872          * Long version
1873          */
1874         hammer_simple_transaction(&trans, hmp);
1875         ++hammer_stats_file_iopsr;
1876         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1877
1878         /*
1879          * Key range (begin and end inclusive) to scan.  Directory keys
1880          * directly translate to a 64 bit 'seek' position.
1881          */
1882         cursor.key_beg.localization = ip->obj_localization +
1883                                       HAMMER_LOCALIZE_MISC;
1884         cursor.key_beg.obj_id = ip->obj_id;
1885         cursor.key_beg.create_tid = 0;
1886         cursor.key_beg.delete_tid = 0;
1887         cursor.key_beg.rec_type = HAMMER_RECTYPE_FIX;
1888         cursor.key_beg.obj_type = 0;
1889         cursor.key_beg.key = HAMMER_FIXKEY_SYMLINK;
1890         cursor.asof = ip->obj_asof;
1891         cursor.flags |= HAMMER_CURSOR_ASOF;
1892
1893         error = hammer_ip_lookup(&cursor);
1894         if (error == 0) {
1895                 error = hammer_ip_resolve_data(&cursor);
1896                 if (error == 0) {
1897                         KKASSERT(cursor.leaf->data_len >=
1898                                  HAMMER_SYMLINK_NAME_OFF);
1899                         error = uiomove(cursor.data->symlink.name,
1900                                         cursor.leaf->data_len -
1901                                                 HAMMER_SYMLINK_NAME_OFF,
1902                                         ap->a_uio);
1903                 }
1904         }
1905         hammer_done_cursor(&cursor);
1906         hammer_done_transaction(&trans);
1907         lwkt_reltoken(&hmp->fs_token);
1908         return(error);
1909 }
1910
1911 /*
1912  * hammer_vop_nremove { nch, dvp, cred }
1913  */
1914 static
1915 int
1916 hammer_vop_nremove(struct vop_nremove_args *ap)
1917 {
1918         struct hammer_transaction trans;
1919         struct hammer_inode *dip;
1920         hammer_mount_t hmp;
1921         int error;
1922
1923         dip = VTOI(ap->a_dvp);
1924         hmp = dip->hmp;
1925
1926         if (hammer_nohistory(dip) == 0 &&
1927             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1928                 return (error);
1929         }
1930
1931         lwkt_gettoken(&hmp->fs_token);
1932         hammer_start_transaction(&trans, hmp);
1933         ++hammer_stats_file_iopsw;
1934         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 0);
1935         hammer_done_transaction(&trans);
1936         if (error == 0)
1937                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1938         lwkt_reltoken(&hmp->fs_token);
1939         return (error);
1940 }
1941
1942 /*
1943  * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1944  */
1945 static
1946 int
1947 hammer_vop_nrename(struct vop_nrename_args *ap)
1948 {
1949         struct hammer_transaction trans;
1950         struct namecache *fncp;
1951         struct namecache *tncp;
1952         struct hammer_inode *fdip;
1953         struct hammer_inode *tdip;
1954         struct hammer_inode *ip;
1955         hammer_mount_t hmp;
1956         struct hammer_cursor cursor;
1957         int64_t namekey;
1958         u_int32_t max_iterations;
1959         int nlen, error;
1960
1961         if (ap->a_fdvp->v_mount != ap->a_tdvp->v_mount)
1962                 return(EXDEV);
1963         if (ap->a_fdvp->v_mount != ap->a_fnch->ncp->nc_vp->v_mount)
1964                 return(EXDEV);
1965
1966         fdip = VTOI(ap->a_fdvp);
1967         tdip = VTOI(ap->a_tdvp);
1968         fncp = ap->a_fnch->ncp;
1969         tncp = ap->a_tnch->ncp;
1970         ip = VTOI(fncp->nc_vp);
1971         KKASSERT(ip != NULL);
1972
1973         hmp = ip->hmp;
1974
1975         if (fdip->obj_localization != tdip->obj_localization)
1976                 return(EXDEV);
1977         if (fdip->obj_localization != ip->obj_localization)
1978                 return(EXDEV);
1979
1980         if (fdip->flags & HAMMER_INODE_RO)
1981                 return (EROFS);
1982         if (tdip->flags & HAMMER_INODE_RO)
1983                 return (EROFS);
1984         if (ip->flags & HAMMER_INODE_RO)
1985                 return (EROFS);
1986         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1987                 return (error);
1988
1989         lwkt_gettoken(&hmp->fs_token);
1990         hammer_start_transaction(&trans, hmp);
1991         ++hammer_stats_file_iopsw;
1992
1993         /*
1994          * Remove tncp from the target directory and then link ip as
1995          * tncp. XXX pass trans to dounlink
1996          *
1997          * Force the inode sync-time to match the transaction so it is
1998          * in-sync with the creation of the target directory entry.
1999          */
2000         error = hammer_dounlink(&trans, ap->a_tnch, ap->a_tdvp,
2001                                 ap->a_cred, 0, -1);
2002         if (error == 0 || error == ENOENT) {
2003                 error = hammer_ip_add_directory(&trans, tdip,
2004                                                 tncp->nc_name, tncp->nc_nlen,
2005                                                 ip);
2006                 if (error == 0) {
2007                         ip->ino_data.parent_obj_id = tdip->obj_id;
2008                         ip->ino_data.ctime = trans.time;
2009                         hammer_modify_inode(&trans, ip, HAMMER_INODE_DDIRTY);
2010                 }
2011         }
2012         if (error)
2013                 goto failed; /* XXX */
2014
2015         /*
2016          * Locate the record in the originating directory and remove it.
2017          *
2018          * Calculate the namekey and setup the key range for the scan.  This
2019          * works kinda like a chained hash table where the lower 32 bits
2020          * of the namekey synthesize the chain.
2021          *
2022          * The key range is inclusive of both key_beg and key_end.
2023          */
2024         namekey = hammer_directory_namekey(fdip, fncp->nc_name, fncp->nc_nlen,
2025                                            &max_iterations);
2026 retry:
2027         hammer_init_cursor(&trans, &cursor, &fdip->cache[1], fdip);
2028         cursor.key_beg.localization = fdip->obj_localization +
2029                                       hammer_dir_localization(fdip);
2030         cursor.key_beg.obj_id = fdip->obj_id;
2031         cursor.key_beg.key = namekey;
2032         cursor.key_beg.create_tid = 0;
2033         cursor.key_beg.delete_tid = 0;
2034         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
2035         cursor.key_beg.obj_type = 0;
2036
2037         cursor.key_end = cursor.key_beg;
2038         cursor.key_end.key += max_iterations;
2039         cursor.asof = fdip->obj_asof;
2040         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2041
2042         /*
2043          * Scan all matching records (the chain), locate the one matching
2044          * the requested path component.
2045          *
2046          * The hammer_ip_*() functions merge in-memory records with on-disk
2047          * records for the purposes of the search.
2048          */
2049         error = hammer_ip_first(&cursor);
2050         while (error == 0) {
2051                 if (hammer_ip_resolve_data(&cursor) != 0)
2052                         break;
2053                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
2054                 KKASSERT(nlen > 0);
2055                 if (fncp->nc_nlen == nlen &&
2056                     bcmp(fncp->nc_name, cursor.data->entry.name, nlen) == 0) {
2057                         break;
2058                 }
2059                 error = hammer_ip_next(&cursor);
2060         }
2061
2062         /*
2063          * If all is ok we have to get the inode so we can adjust nlinks.
2064          *
2065          * WARNING: hammer_ip_del_directory() may have to terminate the
2066          * cursor to avoid a recursion.  It's ok to call hammer_done_cursor()
2067          * twice.
2068          */
2069         if (error == 0)
2070                 error = hammer_ip_del_directory(&trans, &cursor, fdip, ip);
2071
2072         /*
2073          * XXX A deadlock here will break rename's atomicy for the purposes
2074          * of crash recovery.
2075          */
2076         if (error == EDEADLK) {
2077                 hammer_done_cursor(&cursor);
2078                 goto retry;
2079         }
2080
2081         /*
2082          * Cleanup and tell the kernel that the rename succeeded.
2083          *
2084          * NOTE: ip->vp, if non-NULL, cannot be directly referenced
2085          *       without formally acquiring the vp since the vp might
2086          *       have zero refs on it, or in the middle of a reclaim,
2087          *       etc.
2088          */
2089         hammer_done_cursor(&cursor);
2090         if (error == 0) {
2091                 cache_rename(ap->a_fnch, ap->a_tnch);
2092                 hammer_knote(ap->a_fdvp, NOTE_WRITE);
2093                 hammer_knote(ap->a_tdvp, NOTE_WRITE);
2094                 while (ip->vp) {
2095                         struct vnode *vp;
2096
2097                         error = hammer_get_vnode(ip, &vp);
2098                         if (error == 0 && vp) {
2099                                 vn_unlock(vp);
2100                                 hammer_knote(ip->vp, NOTE_RENAME);
2101                                 vrele(vp);
2102                                 break;
2103                         }
2104                         kprintf("Debug: HAMMER ip/vp race2 avoided\n");
2105                 }
2106         }
2107
2108 failed:
2109         hammer_done_transaction(&trans);
2110         lwkt_reltoken(&hmp->fs_token);
2111         return (error);
2112 }
2113
2114 /*
2115  * hammer_vop_nrmdir { nch, dvp, cred }
2116  */
2117 static
2118 int
2119 hammer_vop_nrmdir(struct vop_nrmdir_args *ap)
2120 {
2121         struct hammer_transaction trans;
2122         struct hammer_inode *dip;
2123         hammer_mount_t hmp;
2124         int error;
2125
2126         dip = VTOI(ap->a_dvp);
2127         hmp = dip->hmp;
2128
2129         if (hammer_nohistory(dip) == 0 &&
2130             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2131                 return (error);
2132         }
2133
2134         lwkt_gettoken(&hmp->fs_token);
2135         hammer_start_transaction(&trans, hmp);
2136         ++hammer_stats_file_iopsw;
2137         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 1);
2138         hammer_done_transaction(&trans);
2139         if (error == 0)
2140                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
2141         lwkt_reltoken(&hmp->fs_token);
2142         return (error);
2143 }
2144
2145 /*
2146  * hammer_vop_markatime { vp, cred }
2147  */
2148 static
2149 int
2150 hammer_vop_markatime(struct vop_markatime_args *ap)
2151 {
2152         struct hammer_transaction trans;
2153         struct hammer_inode *ip;
2154         hammer_mount_t hmp;
2155
2156         ip = VTOI(ap->a_vp);
2157         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2158                 return (EROFS);
2159         if (ip->flags & HAMMER_INODE_RO)
2160                 return (EROFS);
2161         hmp = ip->hmp;
2162         if (hmp->mp->mnt_flag & MNT_NOATIME)
2163                 return (0);
2164         lwkt_gettoken(&hmp->fs_token);
2165         hammer_start_transaction(&trans, hmp);
2166         ++hammer_stats_file_iopsw;
2167
2168         ip->ino_data.atime = trans.time;
2169         hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
2170         hammer_done_transaction(&trans);
2171         hammer_knote(ap->a_vp, NOTE_ATTRIB);
2172         lwkt_reltoken(&hmp->fs_token);
2173         return (0);
2174 }
2175
2176 /*
2177  * hammer_vop_setattr { vp, vap, cred }
2178  */
2179 static
2180 int
2181 hammer_vop_setattr(struct vop_setattr_args *ap)
2182 {
2183         struct hammer_transaction trans;
2184         struct hammer_inode *ip;
2185         struct vattr *vap;
2186         hammer_mount_t hmp;
2187         int modflags;
2188         int error;
2189         int truncating;
2190         int blksize;
2191         int kflags;
2192 #if 0
2193         int64_t aligned_size;
2194 #endif
2195         u_int32_t flags;
2196
2197         vap = ap->a_vap;
2198         ip = ap->a_vp->v_data;
2199         modflags = 0;
2200         kflags = 0;
2201         hmp = ip->hmp;
2202
2203         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2204                 return(EROFS);
2205         if (ip->flags & HAMMER_INODE_RO)
2206                 return (EROFS);
2207         if (hammer_nohistory(ip) == 0 &&
2208             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2209                 return (error);
2210         }
2211
2212         lwkt_gettoken(&hmp->fs_token);
2213         hammer_start_transaction(&trans, hmp);
2214         ++hammer_stats_file_iopsw;
2215         error = 0;
2216
2217         if (vap->va_flags != VNOVAL) {
2218                 flags = ip->ino_data.uflags;
2219                 error = vop_helper_setattr_flags(&flags, vap->va_flags,
2220                                          hammer_to_unix_xid(&ip->ino_data.uid),
2221                                          ap->a_cred);
2222                 if (error == 0) {
2223                         if (ip->ino_data.uflags != flags) {
2224                                 ip->ino_data.uflags = flags;
2225                                 ip->ino_data.ctime = trans.time;
2226                                 modflags |= HAMMER_INODE_DDIRTY;
2227                                 kflags |= NOTE_ATTRIB;
2228                         }
2229                         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2230                                 error = 0;
2231                                 goto done;
2232                         }
2233                 }
2234                 goto done;
2235         }
2236         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2237                 error = EPERM;
2238                 goto done;
2239         }
2240         if (vap->va_uid != (uid_t)VNOVAL || vap->va_gid != (gid_t)VNOVAL) {
2241                 mode_t cur_mode = ip->ino_data.mode;
2242                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2243                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2244                 uuid_t uuid_uid;
2245                 uuid_t uuid_gid;
2246
2247                 error = vop_helper_chown(ap->a_vp, vap->va_uid, vap->va_gid,
2248                                          ap->a_cred,
2249                                          &cur_uid, &cur_gid, &cur_mode);
2250                 if (error == 0) {
2251                         hammer_guid_to_uuid(&uuid_uid, cur_uid);
2252                         hammer_guid_to_uuid(&uuid_gid, cur_gid);
2253                         if (bcmp(&uuid_uid, &ip->ino_data.uid,
2254                                  sizeof(uuid_uid)) ||
2255                             bcmp(&uuid_gid, &ip->ino_data.gid,
2256                                  sizeof(uuid_gid)) ||
2257                             ip->ino_data.mode != cur_mode
2258                         ) {
2259                                 ip->ino_data.uid = uuid_uid;
2260                                 ip->ino_data.gid = uuid_gid;
2261                                 ip->ino_data.mode = cur_mode;
2262                                 ip->ino_data.ctime = trans.time;
2263                                 modflags |= HAMMER_INODE_DDIRTY;
2264                         }
2265                         kflags |= NOTE_ATTRIB;
2266                 }
2267         }
2268         while (vap->va_size != VNOVAL && ip->ino_data.size != vap->va_size) {
2269                 switch(ap->a_vp->v_type) {
2270                 case VREG:
2271                         if (vap->va_size == ip->ino_data.size)
2272                                 break;
2273
2274                         /*
2275                          * Log the operation if in fast-fsync mode or if
2276                          * there are unterminated redo write records present.
2277                          *
2278                          * The second check is needed so the recovery code
2279                          * properly truncates write redos even if nominal
2280                          * REDO operations is turned off due to excessive
2281                          * writes, because the related records might be
2282                          * destroyed and never lay down a TERM_WRITE.
2283                          */
2284                         if ((ip->flags & HAMMER_INODE_REDO) ||
2285                             (ip->flags & HAMMER_INODE_RDIRTY)) {
2286                                 error = hammer_generate_redo(&trans, ip,
2287                                                              vap->va_size,
2288                                                              HAMMER_REDO_TRUNC,
2289                                                              NULL, 0);
2290                         }
2291                         blksize = hammer_blocksize(vap->va_size);
2292
2293                         /*
2294                          * XXX break atomicy, we can deadlock the backend
2295                          * if we do not release the lock.  Probably not a
2296                          * big deal here.
2297                          */
2298                         if (vap->va_size < ip->ino_data.size) {
2299                                 nvtruncbuf(ap->a_vp, vap->va_size,
2300                                            blksize,
2301                                            hammer_blockoff(vap->va_size));
2302                                 truncating = 1;
2303                                 kflags |= NOTE_WRITE;
2304                         } else {
2305                                 nvextendbuf(ap->a_vp,
2306                                             ip->ino_data.size,
2307                                             vap->va_size,
2308                                             hammer_blocksize(ip->ino_data.size),
2309                                             hammer_blocksize(vap->va_size),
2310                                             hammer_blockoff(ip->ino_data.size),
2311                                             hammer_blockoff(vap->va_size),
2312                                             0);
2313                                 truncating = 0;
2314                                 kflags |= NOTE_WRITE | NOTE_EXTEND;
2315                         }
2316                         ip->ino_data.size = vap->va_size;
2317                         ip->ino_data.mtime = trans.time;
2318                         /* XXX safe to use SDIRTY instead of DDIRTY here? */
2319                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2320
2321                         /*
2322                          * On-media truncation is cached in the inode until
2323                          * the inode is synchronized.  We must immediately
2324                          * handle any frontend records.
2325                          */
2326                         if (truncating) {
2327                                 hammer_ip_frontend_trunc(ip, vap->va_size);
2328 #ifdef DEBUG_TRUNCATE
2329                                 if (HammerTruncIp == NULL)
2330                                         HammerTruncIp = ip;
2331 #endif
2332                                 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2333                                         ip->flags |= HAMMER_INODE_TRUNCATED;
2334                                         ip->trunc_off = vap->va_size;
2335 #ifdef DEBUG_TRUNCATE
2336                                         if (ip == HammerTruncIp)
2337                                         kprintf("truncate1 %016llx\n",
2338                                                 (long long)ip->trunc_off);
2339 #endif
2340                                 } else if (ip->trunc_off > vap->va_size) {
2341                                         ip->trunc_off = vap->va_size;
2342 #ifdef DEBUG_TRUNCATE
2343                                         if (ip == HammerTruncIp)
2344                                         kprintf("truncate2 %016llx\n",
2345                                                 (long long)ip->trunc_off);
2346 #endif
2347                                 } else {
2348 #ifdef DEBUG_TRUNCATE
2349                                         if (ip == HammerTruncIp)
2350                                         kprintf("truncate3 %016llx (ignored)\n",
2351                                                 (long long)vap->va_size);
2352 #endif
2353                                 }
2354                         }
2355
2356 #if 0
2357                         /*
2358                          * When truncating, nvtruncbuf() may have cleaned out
2359                          * a portion of the last block on-disk in the buffer
2360                          * cache.  We must clean out any frontend records
2361                          * for blocks beyond the new last block.
2362                          */
2363                         aligned_size = (vap->va_size + (blksize - 1)) &
2364                                        ~(int64_t)(blksize - 1);
2365                         if (truncating && vap->va_size < aligned_size) {
2366                                 aligned_size -= blksize;
2367                                 hammer_ip_frontend_trunc(ip, aligned_size);
2368                         }
2369 #endif
2370                         break;
2371                 case VDATABASE:
2372                         if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2373                                 ip->flags |= HAMMER_INODE_TRUNCATED;
2374                                 ip->trunc_off = vap->va_size;
2375                         } else if (ip->trunc_off > vap->va_size) {
2376                                 ip->trunc_off = vap->va_size;
2377                         }
2378                         hammer_ip_frontend_trunc(ip, vap->va_size);
2379                         ip->ino_data.size = vap->va_size;
2380                         ip->ino_data.mtime = trans.time;
2381                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2382                         kflags |= NOTE_ATTRIB;
2383                         break;
2384                 default:
2385                         error = EINVAL;
2386                         goto done;
2387                 }
2388                 break;
2389         }
2390         if (vap->va_atime.tv_sec != VNOVAL) {
2391                 ip->ino_data.atime = hammer_timespec_to_time(&vap->va_atime);
2392                 modflags |= HAMMER_INODE_ATIME;
2393                 kflags |= NOTE_ATTRIB;
2394         }
2395         if (vap->va_mtime.tv_sec != VNOVAL) {
2396                 ip->ino_data.mtime = hammer_timespec_to_time(&vap->va_mtime);
2397                 modflags |= HAMMER_INODE_MTIME;
2398                 kflags |= NOTE_ATTRIB;
2399         }
2400         if (vap->va_mode != (mode_t)VNOVAL) {
2401                 mode_t   cur_mode = ip->ino_data.mode;
2402                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2403                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2404
2405                 error = vop_helper_chmod(ap->a_vp, vap->va_mode, ap->a_cred,
2406                                          cur_uid, cur_gid, &cur_mode);
2407                 if (error == 0 && ip->ino_data.mode != cur_mode) {
2408                         ip->ino_data.mode = cur_mode;
2409                         ip->ino_data.ctime = trans.time;
2410                         modflags |= HAMMER_INODE_DDIRTY;
2411                         kflags |= NOTE_ATTRIB;
2412                 }
2413         }
2414 done:
2415         if (error == 0)
2416                 hammer_modify_inode(&trans, ip, modflags);
2417         hammer_done_transaction(&trans);
2418         hammer_knote(ap->a_vp, kflags);
2419         lwkt_reltoken(&hmp->fs_token);
2420         return (error);
2421 }
2422
2423 /*
2424  * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
2425  */
2426 static
2427 int
2428 hammer_vop_nsymlink(struct vop_nsymlink_args *ap)
2429 {
2430         struct hammer_transaction trans;
2431         struct hammer_inode *dip;
2432         struct hammer_inode *nip;
2433         hammer_record_t record;
2434         struct nchandle *nch;
2435         hammer_mount_t hmp;
2436         int error;
2437         int bytes;
2438
2439         ap->a_vap->va_type = VLNK;
2440
2441         nch = ap->a_nch;
2442         dip = VTOI(ap->a_dvp);
2443         hmp = dip->hmp;
2444
2445         if (dip->flags & HAMMER_INODE_RO)
2446                 return (EROFS);
2447         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
2448                 return (error);
2449
2450         /*
2451          * Create a transaction to cover the operations we perform.
2452          */
2453         lwkt_gettoken(&hmp->fs_token);
2454         hammer_start_transaction(&trans, hmp);
2455         ++hammer_stats_file_iopsw;
2456
2457         /*
2458          * Create a new filesystem object of the requested type.  The
2459          * returned inode will be referenced but not locked.
2460          */
2461
2462         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
2463                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
2464                                     NULL, &nip);
2465         if (error) {
2466                 hammer_done_transaction(&trans);
2467                 *ap->a_vpp = NULL;
2468                 lwkt_reltoken(&hmp->fs_token);
2469                 return (error);
2470         }
2471
2472         /*
2473          * Add a record representing the symlink.  symlink stores the link
2474          * as pure data, not a string, and is no \0 terminated.
2475          */
2476         if (error == 0) {
2477                 bytes = strlen(ap->a_target);
2478
2479                 if (bytes <= HAMMER_INODE_BASESYMLEN) {
2480                         bcopy(ap->a_target, nip->ino_data.ext.symlink, bytes);
2481                 } else {
2482                         record = hammer_alloc_mem_record(nip, bytes);
2483                         record->type = HAMMER_MEM_RECORD_GENERAL;
2484
2485                         record->leaf.base.localization = nip->obj_localization +
2486                                                          HAMMER_LOCALIZE_MISC;
2487                         record->leaf.base.key = HAMMER_FIXKEY_SYMLINK;
2488                         record->leaf.base.rec_type = HAMMER_RECTYPE_FIX;
2489                         record->leaf.data_len = bytes;
2490                         KKASSERT(HAMMER_SYMLINK_NAME_OFF == 0);
2491                         bcopy(ap->a_target, record->data->symlink.name, bytes);
2492                         error = hammer_ip_add_record(&trans, record);
2493                 }
2494
2495                 /*
2496                  * Set the file size to the length of the link.
2497                  */
2498                 if (error == 0) {
2499                         nip->ino_data.size = bytes;
2500                         hammer_modify_inode(&trans, nip, HAMMER_INODE_DDIRTY);
2501                 }
2502         }
2503         if (error == 0)
2504                 error = hammer_ip_add_directory(&trans, dip, nch->ncp->nc_name,
2505                                                 nch->ncp->nc_nlen, nip);
2506
2507         /*
2508          * Finish up.
2509          */
2510         if (error) {
2511                 hammer_rel_inode(nip, 0);
2512                 *ap->a_vpp = NULL;
2513         } else {
2514                 error = hammer_get_vnode(nip, ap->a_vpp);
2515                 hammer_rel_inode(nip, 0);
2516                 if (error == 0) {
2517                         cache_setunresolved(ap->a_nch);
2518                         cache_setvp(ap->a_nch, *ap->a_vpp);
2519                         hammer_knote(ap->a_dvp, NOTE_WRITE);
2520                 }
2521         }
2522         hammer_done_transaction(&trans);
2523         lwkt_reltoken(&hmp->fs_token);
2524         return (error);
2525 }
2526
2527 /*
2528  * hammer_vop_nwhiteout { nch, dvp, cred, flags }
2529  */
2530 static
2531 int
2532 hammer_vop_nwhiteout(struct vop_nwhiteout_args *ap)
2533 {
2534         struct hammer_transaction trans;
2535         struct hammer_inode *dip;
2536         hammer_mount_t hmp;
2537         int error;
2538
2539         dip = VTOI(ap->a_dvp);
2540         hmp = dip->hmp;
2541
2542         if (hammer_nohistory(dip) == 0 &&
2543             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0) {
2544                 return (error);
2545         }
2546
2547         lwkt_gettoken(&hmp->fs_token);
2548         hammer_start_transaction(&trans, hmp);
2549         ++hammer_stats_file_iopsw;
2550         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp,
2551                                 ap->a_cred, ap->a_flags, -1);
2552         hammer_done_transaction(&trans);
2553         lwkt_reltoken(&hmp->fs_token);
2554
2555         return (error);
2556 }
2557
2558 /*
2559  * hammer_vop_ioctl { vp, command, data, fflag, cred }
2560  */
2561 static
2562 int
2563 hammer_vop_ioctl(struct vop_ioctl_args *ap)
2564 {
2565         struct hammer_inode *ip = ap->a_vp->v_data;
2566         hammer_mount_t hmp = ip->hmp;
2567         int error;
2568
2569         ++hammer_stats_file_iopsr;
2570         lwkt_gettoken(&hmp->fs_token);
2571         error = hammer_ioctl(ip, ap->a_command, ap->a_data,
2572                              ap->a_fflag, ap->a_cred);
2573         lwkt_reltoken(&hmp->fs_token);
2574         return (error);
2575 }
2576
2577 static
2578 int
2579 hammer_vop_mountctl(struct vop_mountctl_args *ap)
2580 {
2581         static const struct mountctl_opt extraopt[] = {
2582                 { HMNT_NOHISTORY,       "nohistory" },
2583                 { HMNT_MASTERID,        "master" },
2584                 { 0, NULL}
2585
2586         };
2587         struct hammer_mount *hmp;
2588         struct mount *mp;
2589         int usedbytes;
2590         int error;
2591
2592         error = 0;
2593         usedbytes = 0;
2594         mp = ap->a_head.a_ops->head.vv_mount;
2595         KKASSERT(mp->mnt_data != NULL);
2596         hmp = (struct hammer_mount *)mp->mnt_data;
2597
2598         lwkt_gettoken(&hmp->fs_token);
2599
2600         switch(ap->a_op) {
2601         case MOUNTCTL_SET_EXPORT:
2602                 if (ap->a_ctllen != sizeof(struct export_args))
2603                         error = EINVAL;
2604                 else
2605                         error = hammer_vfs_export(mp, ap->a_op,
2606                                       (const struct export_args *)ap->a_ctl);
2607                 break;
2608         case MOUNTCTL_MOUNTFLAGS:
2609         {
2610                 /*
2611                  * Call standard mountctl VOP function
2612                  * so we get user mount flags.
2613                  */
2614                 error = vop_stdmountctl(ap);
2615                 if (error)
2616                         break;
2617
2618                 usedbytes = *ap->a_res;
2619
2620                 if (usedbytes > 0 && usedbytes < ap->a_buflen) {
2621                         usedbytes += vfs_flagstostr(hmp->hflags, extraopt,
2622                                                     ap->a_buf,
2623                                                     ap->a_buflen - usedbytes,
2624                                                     &error);
2625                 }
2626
2627                 *ap->a_res += usedbytes;
2628                 break;
2629         }
2630         default:
2631                 error = vop_stdmountctl(ap);
2632                 break;
2633         }
2634         lwkt_reltoken(&hmp->fs_token);
2635         return(error);
2636 }
2637
2638 /*
2639  * hammer_vop_strategy { vp, bio }
2640  *
2641  * Strategy call, used for regular file read & write only.  Note that the
2642  * bp may represent a cluster.
2643  *
2644  * To simplify operation and allow better optimizations in the future,
2645  * this code does not make any assumptions with regards to buffer alignment
2646  * or size.
2647  */
2648 static
2649 int
2650 hammer_vop_strategy(struct vop_strategy_args *ap)
2651 {
2652         struct buf *bp;
2653         int error;
2654
2655         bp = ap->a_bio->bio_buf;
2656
2657         switch(bp->b_cmd) {
2658         case BUF_CMD_READ:
2659                 error = hammer_vop_strategy_read(ap);
2660                 break;
2661         case BUF_CMD_WRITE:
2662                 error = hammer_vop_strategy_write(ap);
2663                 break;
2664         default:
2665                 bp->b_error = error = EINVAL;
2666                 bp->b_flags |= B_ERROR;
2667                 biodone(ap->a_bio);
2668                 break;
2669         }
2670
2671         /* hammer_dump_dedup_cache(((hammer_inode_t)ap->a_vp->v_data)->hmp); */
2672
2673         return (error);
2674 }
2675
2676 /*
2677  * Read from a regular file.  Iterate the related records and fill in the
2678  * BIO/BUF.  Gaps are zero-filled.
2679  *
2680  * The support code in hammer_object.c should be used to deal with mixed
2681  * in-memory and on-disk records.
2682  *
2683  * NOTE: Can be called from the cluster code with an oversized buf.
2684  *
2685  * XXX atime update
2686  */
2687 static
2688 int
2689 hammer_vop_strategy_read(struct vop_strategy_args *ap)
2690 {
2691         struct hammer_transaction trans;
2692         struct hammer_inode *ip;
2693         struct hammer_inode *dip;
2694         hammer_mount_t hmp;
2695         struct hammer_cursor cursor;
2696         hammer_base_elm_t base;
2697         hammer_off_t disk_offset;
2698         struct bio *bio;
2699         struct bio *nbio;
2700         struct buf *bp;
2701         int64_t rec_offset;
2702         int64_t ran_end;
2703         int64_t tmp64;
2704         int error;
2705         int boff;
2706         int roff;
2707         int n;
2708         int isdedupable;
2709
2710         bio = ap->a_bio;
2711         bp = bio->bio_buf;
2712         ip = ap->a_vp->v_data;
2713         hmp = ip->hmp;
2714
2715         /*
2716          * The zone-2 disk offset may have been set by the cluster code via
2717          * a BMAP operation, or else should be NOOFFSET.
2718          *
2719          * Checking the high bits for a match against zone-2 should suffice.
2720          *
2721          * In cases where a lot of data duplication is present it may be
2722          * more beneficial to drop through and doubule-buffer through the
2723          * device.
2724          */
2725         nbio = push_bio(bio);
2726         if ((nbio->bio_offset & HAMMER_OFF_ZONE_MASK) ==
2727             HAMMER_ZONE_LARGE_DATA) {
2728                 if (hammer_double_buffer == 0) {
2729                         lwkt_gettoken(&hmp->fs_token);
2730                         error = hammer_io_direct_read(hmp, nbio, NULL);
2731                         lwkt_reltoken(&hmp->fs_token);
2732                         return (error);
2733                 }
2734
2735                 /*
2736                  * Try to shortcut requests for double_buffer mode too.
2737                  * Since this mode runs through the device buffer cache
2738                  * only compatible buffer sizes (meaning those generated
2739                  * by normal filesystem buffers) are legal.
2740                  */
2741                 if (hammer_live_dedup == 0 && (bp->b_flags & B_PAGING) == 0) {
2742                         error = hammer_io_indirect_read(hmp, nbio, NULL);
2743                         return (error);
2744                 }
2745         }
2746
2747         /*
2748          * Well, that sucked.  Do it the hard way.  If all the stars are
2749          * aligned we may still be able to issue a direct-read.
2750          */
2751         lwkt_gettoken(&hmp->fs_token);
2752         hammer_simple_transaction(&trans, hmp);
2753         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2754
2755         /*
2756          * Key range (begin and end inclusive) to scan.  Note that the key's
2757          * stored in the actual records represent BASE+LEN, not BASE.  The
2758          * first record containing bio_offset will have a key > bio_offset.
2759          */
2760         cursor.key_beg.localization = ip->obj_localization +
2761                                       HAMMER_LOCALIZE_MISC;
2762         cursor.key_beg.obj_id = ip->obj_id;
2763         cursor.key_beg.create_tid = 0;
2764         cursor.key_beg.delete_tid = 0;
2765         cursor.key_beg.obj_type = 0;
2766         cursor.key_beg.key = bio->bio_offset + 1;
2767         cursor.asof = ip->obj_asof;
2768         cursor.flags |= HAMMER_CURSOR_ASOF;
2769
2770         cursor.key_end = cursor.key_beg;
2771         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2772 #if 0
2773         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
2774                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
2775                 cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
2776                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2777         } else
2778 #endif
2779         {
2780                 ran_end = bio->bio_offset + bp->b_bufsize;
2781                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2782                 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2783                 tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2784                 if (tmp64 < ran_end)
2785                         cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2786                 else
2787                         cursor.key_end.key = ran_end + MAXPHYS + 1;
2788         }
2789         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2790
2791         /*
2792          * Set NOSWAPCACHE for cursor data extraction if double buffering
2793          * is disabled or (if the file is not marked cacheable via chflags
2794          * and vm.swapcache_use_chflags is enabled).
2795          */
2796         if (hammer_double_buffer == 0 ||
2797             ((ap->a_vp->v_flag & VSWAPCACHE) == 0 &&
2798              vm_swapcache_use_chflags)) {
2799                 cursor.flags |= HAMMER_CURSOR_NOSWAPCACHE;
2800         }
2801
2802         error = hammer_ip_first(&cursor);
2803         boff = 0;
2804
2805         while (error == 0) {
2806                 /*
2807                  * Get the base file offset of the record.  The key for
2808                  * data records is (base + bytes) rather then (base).
2809                  */
2810                 base = &cursor.leaf->base;
2811                 rec_offset = base->key - cursor.leaf->data_len;
2812
2813                 /*
2814                  * Calculate the gap, if any, and zero-fill it.
2815                  *
2816                  * n is the offset of the start of the record verses our
2817                  * current seek offset in the bio.
2818                  */
2819                 n = (int)(rec_offset - (bio->bio_offset + boff));
2820                 if (n > 0) {
2821                         if (n > bp->b_bufsize - boff)
2822                                 n = bp->b_bufsize - boff;
2823                         bzero((char *)bp->b_data + boff, n);
2824                         boff += n;
2825                         n = 0;
2826                 }
2827
2828                 /*
2829                  * Calculate the data offset in the record and the number
2830                  * of bytes we can copy.
2831                  *
2832                  * There are two degenerate cases.  First, boff may already
2833                  * be at bp->b_bufsize.  Secondly, the data offset within
2834                  * the record may exceed the record's size.
2835                  */
2836                 roff = -n;
2837                 rec_offset += roff;
2838                 n = cursor.leaf->data_len - roff;
2839                 if (n <= 0) {
2840                         kprintf("strategy_read: bad n=%d roff=%d\n", n, roff);
2841                         n = 0;
2842                 } else if (n > bp->b_bufsize - boff) {
2843                         n = bp->b_bufsize - boff;
2844                 }
2845
2846                 /*
2847                  * Deal with cached truncations.  This cool bit of code
2848                  * allows truncate()/ftruncate() to avoid having to sync
2849                  * the file.
2850                  *
2851                  * If the frontend is truncated then all backend records are
2852                  * subject to the frontend's truncation.
2853                  *
2854                  * If the backend is truncated then backend records on-disk
2855                  * (but not in-memory) are subject to the backend's
2856                  * truncation.  In-memory records owned by the backend
2857                  * represent data written after the truncation point on the
2858                  * backend and must not be truncated.
2859                  *
2860                  * Truncate operations deal with frontend buffer cache
2861                  * buffers and frontend-owned in-memory records synchronously.
2862                  */
2863                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2864                         if (hammer_cursor_ondisk(&cursor)/* ||
2865                             cursor.iprec->flush_state == HAMMER_FST_FLUSH*/) {
2866                                 if (ip->trunc_off <= rec_offset)
2867                                         n = 0;
2868                                 else if (ip->trunc_off < rec_offset + n)
2869                                         n = (int)(ip->trunc_off - rec_offset);
2870                         }
2871                 }
2872                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2873                         if (hammer_cursor_ondisk(&cursor)) {
2874                                 if (ip->sync_trunc_off <= rec_offset)
2875                                         n = 0;
2876                                 else if (ip->sync_trunc_off < rec_offset + n)
2877                                         n = (int)(ip->sync_trunc_off - rec_offset);
2878                         }
2879                 }
2880
2881                 /*
2882                  * Try to issue a direct read into our bio if possible,
2883                  * otherwise resolve the element data into a hammer_buffer
2884                  * and copy.
2885                  *
2886                  * The buffer on-disk should be zerod past any real
2887                  * truncation point, but may not be for any synthesized
2888                  * truncation point from above.
2889                  *
2890                  * NOTE: disk_offset is only valid if the cursor data is
2891                  *       on-disk.
2892                  */
2893                 disk_offset = cursor.leaf->data_offset + roff;
2894                 isdedupable = (boff == 0 && n == bp->b_bufsize &&
2895                                hammer_cursor_ondisk(&cursor) &&
2896                                ((int)disk_offset & HAMMER_BUFMASK) == 0);
2897
2898                 if (isdedupable && hammer_double_buffer == 0) {
2899                         /*
2900                          * Direct read case
2901                          */
2902                         KKASSERT((disk_offset & HAMMER_OFF_ZONE_MASK) ==
2903                                  HAMMER_ZONE_LARGE_DATA);
2904                         nbio->bio_offset = disk_offset;
2905                         error = hammer_io_direct_read(hmp, nbio, cursor.leaf);
2906                         if (hammer_live_dedup && error == 0)
2907                                 hammer_dedup_cache_add(ip, cursor.leaf);
2908                         goto done;
2909                 } else if (isdedupable) {
2910                         /*
2911                          * Async I/O case for reading from backing store
2912                          * and copying the data to the filesystem buffer.
2913                          * live-dedup has to verify the data anyway if it
2914                          * gets a hit later so we can just add the entry
2915                          * now.
2916                          */
2917                         KKASSERT((disk_offset & HAMMER_OFF_ZONE_MASK) ==
2918                                  HAMMER_ZONE_LARGE_DATA);
2919                         nbio->bio_offset = disk_offset;
2920                         if (hammer_live_dedup)
2921                                 hammer_dedup_cache_add(ip, cursor.leaf);
2922                         error = hammer_io_indirect_read(hmp, nbio, cursor.leaf);
2923                         goto done;
2924                 } else if (n) {
2925                         error = hammer_ip_resolve_data(&cursor);
2926                         if (error == 0) {
2927                                 if (hammer_live_dedup && isdedupable)
2928                                         hammer_dedup_cache_add(ip, cursor.leaf);
2929                                 bcopy((char *)cursor.data + roff,
2930                                       (char *)bp->b_data + boff, n);
2931                         }
2932                 }
2933                 if (error)
2934                         break;
2935
2936                 /*
2937                  * We have to be sure that the only elements added to the
2938                  * dedup cache are those which are already on-media.
2939                  */
2940                 if (hammer_live_dedup && hammer_cursor_ondisk(&cursor))
2941                         hammer_dedup_cache_add(ip, cursor.leaf);
2942
2943                 /*
2944                  * Iterate until we have filled the request.
2945                  */
2946                 boff += n;
2947                 if (boff == bp->b_bufsize)
2948                         break;
2949                 error = hammer_ip_next(&cursor);
2950         }
2951
2952         /*
2953          * There may have been a gap after the last record
2954          */
2955         if (error == ENOENT)
2956                 error = 0;
2957         if (error == 0 && boff != bp->b_bufsize) {
2958                 KKASSERT(boff < bp->b_bufsize);
2959                 bzero((char *)bp->b_data + boff, bp->b_bufsize - boff);
2960                 /* boff = bp->b_bufsize; */
2961         }
2962
2963         /*
2964          * Disallow swapcache operation on the vnode buffer if double
2965          * buffering is enabled, the swapcache will get the data via
2966          * the block device buffer.
2967          */
2968         if (hammer_double_buffer)
2969                 bp->b_flags |= B_NOTMETA;
2970
2971         /*
2972          * Cleanup
2973          */
2974         bp->b_resid = 0;
2975         bp->b_error = error;
2976         if (error)
2977                 bp->b_flags |= B_ERROR;
2978         biodone(ap->a_bio);
2979
2980 done:
2981         /*
2982          * Cache the b-tree node for the last data read in cache[1].
2983          *
2984          * If we hit the file EOF then also cache the node in the
2985          * governing director's cache[3], it will be used to initialize
2986          * the inode's cache[1] for any inodes looked up via the directory.
2987          *
2988          * This doesn't reduce disk accesses since the B-Tree chain is
2989          * likely cached, but it does reduce cpu overhead when looking
2990          * up file offsets for cpdup/tar/cpio style iterations.
2991          */
2992         if (cursor.node)
2993                 hammer_cache_node(&ip->cache[1], cursor.node);
2994         if (ran_end >= ip->ino_data.size) {
2995                 dip = hammer_find_inode(&trans, ip->ino_data.parent_obj_id,
2996                                         ip->obj_asof, ip->obj_localization);
2997                 if (dip) {
2998                         hammer_cache_node(&dip->cache[3], cursor.node);
2999                         hammer_rel_inode(dip, 0);
3000                 }
3001         }
3002         hammer_done_cursor(&cursor);
3003         hammer_done_transaction(&trans);
3004         lwkt_reltoken(&hmp->fs_token);
3005         return(error);
3006 }
3007
3008 /*
3009  * BMAP operation - used to support cluster_read() only.
3010  *
3011  * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
3012  *
3013  * This routine may return EOPNOTSUPP if the opration is not supported for
3014  * the specified offset.  The contents of the pointer arguments do not
3015  * need to be initialized in that case.
3016  *
3017  * If a disk address is available and properly aligned return 0 with
3018  * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
3019  * to the run-length relative to that offset.  Callers may assume that
3020  * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
3021  * large, so return EOPNOTSUPP if it is not sufficiently large.
3022  */
3023 static
3024 int
3025 hammer_vop_bmap(struct vop_bmap_args *ap)
3026 {
3027         struct hammer_transaction trans;
3028         struct hammer_inode *ip;
3029         hammer_mount_t hmp;
3030         struct hammer_cursor cursor;
3031         hammer_base_elm_t base;
3032         int64_t rec_offset;
3033         int64_t ran_end;
3034         int64_t tmp64;
3035         int64_t base_offset;
3036         int64_t base_disk_offset;
3037         int64_t last_offset;
3038         hammer_off_t last_disk_offset;
3039         hammer_off_t disk_offset;
3040         int     rec_len;
3041         int     error;
3042         int     blksize;
3043
3044         ++hammer_stats_file_iopsr;
3045         ip = ap->a_vp->v_data;
3046         hmp = ip->hmp;
3047
3048         /*
3049          * We can only BMAP regular files.  We can't BMAP database files,
3050          * directories, etc.
3051          */
3052         if (ip->ino_data.obj_type != HAMMER_OBJTYPE_REGFILE)
3053                 return(EOPNOTSUPP);
3054
3055         /*
3056          * bmap is typically called with runp/runb both NULL when used
3057          * for writing.  We do not support BMAP for writing atm.
3058          */
3059         if (ap->a_cmd != BUF_CMD_READ)
3060                 return(EOPNOTSUPP);
3061
3062         /*
3063          * Scan the B-Tree to acquire blockmap addresses, then translate
3064          * to raw addresses.
3065          */
3066         lwkt_gettoken(&hmp->fs_token);
3067         hammer_simple_transaction(&trans, hmp);
3068 #if 0
3069         kprintf("bmap_beg %016llx ip->cache %p\n",
3070                 (long long)ap->a_loffset, ip->cache[1]);
3071 #endif
3072         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
3073
3074         /*
3075          * Key range (begin and end inclusive) to scan.  Note that the key's
3076          * stored in the actual records represent BASE+LEN, not BASE.  The
3077          * first record containing bio_offset will have a key > bio_offset.
3078          */
3079         cursor.key_beg.localization = ip->obj_localization +
3080                                       HAMMER_LOCALIZE_MISC;
3081         cursor.key_beg.obj_id = ip->obj_id;
3082         cursor.key_beg.create_tid = 0;
3083         cursor.key_beg.delete_tid = 0;
3084         cursor.key_beg.obj_type = 0;
3085         if (ap->a_runb)
3086                 cursor.key_beg.key = ap->a_loffset - MAXPHYS + 1;
3087         else
3088                 cursor.key_beg.key = ap->a_loffset + 1;
3089         if (cursor.key_beg.key < 0)
3090                 cursor.key_beg.key = 0;
3091         cursor.asof = ip->obj_asof;
3092         cursor.flags |= HAMMER_CURSOR_ASOF;
3093
3094         cursor.key_end = cursor.key_beg;
3095         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
3096
3097         ran_end = ap->a_loffset + MAXPHYS;
3098         cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
3099         cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
3100         tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
3101         if (tmp64 < ran_end)
3102                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
3103         else
3104                 cursor.key_end.key = ran_end + MAXPHYS + 1;
3105
3106         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
3107
3108         error = hammer_ip_first(&cursor);
3109         base_offset = last_offset = 0;
3110         base_disk_offset = last_disk_offset = 0;
3111
3112         while (error == 0) {
3113                 /*
3114                  * Get the base file offset of the record.  The key for
3115                  * data records is (base + bytes) rather then (base).
3116                  *
3117                  * NOTE: rec_offset + rec_len may exceed the end-of-file.
3118                  * The extra bytes should be zero on-disk and the BMAP op
3119                  * should still be ok.
3120                  */
3121                 base = &cursor.leaf->base;
3122                 rec_offset = base->key - cursor.leaf->data_len;
3123                 rec_len    = cursor.leaf->data_len;
3124
3125                 /*
3126                  * Incorporate any cached truncation.
3127                  *
3128                  * NOTE: Modifications to rec_len based on synthesized
3129                  * truncation points remove the guarantee that any extended
3130                  * data on disk is zero (since the truncations may not have
3131                  * taken place on-media yet).
3132                  */
3133                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
3134                         if (hammer_cursor_ondisk(&cursor) ||
3135                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
3136                                 if (ip->trunc_off <= rec_offset)
3137                                         rec_len = 0;
3138                                 else if (ip->trunc_off < rec_offset + rec_len)
3139                                         rec_len = (int)(ip->trunc_off - rec_offset);
3140                         }
3141                 }
3142                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
3143                         if (hammer_cursor_ondisk(&cursor)) {
3144                                 if (ip->sync_trunc_off <= rec_offset)
3145                                         rec_len = 0;
3146                                 else if (ip->sync_trunc_off < rec_offset + rec_len)
3147                                         rec_len = (int)(ip->sync_trunc_off - rec_offset);
3148                         }
3149                 }
3150
3151                 /*
3152                  * Accumulate information.  If we have hit a discontiguous
3153                  * block reset base_offset unless we are already beyond the
3154                  * requested offset.  If we are, that's it, we stop.
3155                  */
3156                 if (error)
3157                         break;
3158                 if (hammer_cursor_ondisk(&cursor)) {
3159                         disk_offset = cursor.leaf->data_offset;
3160                         if (rec_offset != last_offset ||
3161                             disk_offset != last_disk_offset) {
3162                                 if (rec_offset > ap->a_loffset)
3163                                         break;
3164                                 base_offset = rec_offset;
3165                                 base_disk_offset = disk_offset;
3166                         }
3167                         last_offset = rec_offset + rec_len;
3168                         last_disk_offset = disk_offset + rec_len;
3169
3170                         if (hammer_live_dedup)
3171                                 hammer_dedup_cache_add(ip, cursor.leaf);
3172                 }
3173
3174                 error = hammer_ip_next(&cursor);
3175         }
3176
3177 #if 0
3178         kprintf("BMAP %016llx:  %016llx - %016llx\n",
3179                 (long long)ap->a_loffset,
3180                 (long long)base_offset,
3181                 (long long)last_offset);
3182         kprintf("BMAP %16s:  %016llx - %016llx\n", "",
3183                 (long long)base_disk_offset,
3184                 (long long)last_disk_offset);
3185 #endif
3186
3187         if (cursor.node) {
3188                 hammer_cache_node(&ip->cache[1], cursor.node);
3189 #if 0
3190                 kprintf("bmap_end2 %016llx ip->cache %p\n",
3191                         (long long)ap->a_loffset, ip->cache[1]);
3192 #endif
3193         }
3194         hammer_done_cursor(&cursor);
3195         hammer_done_transaction(&trans);
3196         lwkt_reltoken(&hmp->fs_token);
3197
3198         /*
3199          * If we couldn't find any records or the records we did find were
3200          * all behind the requested offset, return failure.  A forward
3201          * truncation can leave a hole w/ no on-disk records.
3202          */
3203         if (last_offset == 0 || last_offset < ap->a_loffset)
3204                 return (EOPNOTSUPP);
3205
3206         /*
3207          * Figure out the block size at the requested offset and adjust
3208          * our limits so the cluster_read() does not create inappropriately
3209          * sized buffer cache buffers.
3210          */
3211         blksize = hammer_blocksize(ap->a_loffset);
3212         if (hammer_blocksize(base_offset) != blksize) {
3213                 base_offset = hammer_blockdemarc(base_offset, ap->a_loffset);
3214         }
3215         if (last_offset != ap->a_loffset &&
3216             hammer_blocksize(last_offset - 1) != blksize) {
3217                 last_offset = hammer_blockdemarc(ap->a_loffset,
3218                                                  last_offset - 1);
3219         }
3220
3221         /*
3222          * Returning EOPNOTSUPP simply prevents the direct-IO optimization
3223          * from occuring.
3224          */
3225         disk_offset = base_disk_offset + (ap->a_loffset - base_offset);
3226
3227         if ((disk_offset & HAMMER_OFF_ZONE_MASK) != HAMMER_ZONE_LARGE_DATA) {
3228                 /*
3229                  * Only large-data zones can be direct-IOd
3230                  */
3231                 error = EOPNOTSUPP;
3232         } else if ((disk_offset & HAMMER_BUFMASK) ||
3233                    (last_offset - ap->a_loffset) < blksize) {
3234                 /*
3235                  * doffsetp is not aligned or the forward run size does
3236                  * not cover a whole buffer, disallow the direct I/O.
3237                  */
3238                 error = EOPNOTSUPP;
3239         } else {
3240                 /*
3241                  * We're good.
3242                  */
3243                 *ap->a_doffsetp = disk_offset;
3244                 if (ap->a_runb) {
3245                         *ap->a_runb = ap->a_loffset - base_offset;
3246                         KKASSERT(*ap->a_runb >= 0);
3247                 }
3248                 if (ap->a_runp) {
3249                         *ap->a_runp = last_offset - ap->a_loffset;
3250                         KKASSERT(*ap->a_runp >= 0);
3251                 }
3252                 error = 0;
3253         }
3254         return(error);
3255 }
3256
3257 /*
3258  * Write to a regular file.   Because this is a strategy call the OS is
3259  * trying to actually get data onto the media.
3260  */
3261 static
3262 int
3263 hammer_vop_strategy_write(struct vop_strategy_args *ap)
3264 {
3265         hammer_record_t record;
3266         hammer_mount_t hmp;
3267         hammer_inode_t ip;
3268         struct bio *bio;
3269         struct buf *bp;
3270         int blksize;
3271         int bytes;
3272         int error;
3273
3274         bio = ap->a_bio;
3275         bp = bio->bio_buf;
3276         ip = ap->a_vp->v_data;
3277         hmp = ip->hmp;
3278
3279         blksize = hammer_blocksize(bio->bio_offset);
3280         KKASSERT(bp->b_bufsize == blksize);
3281
3282         if (ip->flags & HAMMER_INODE_RO) {
3283                 bp->b_error = EROFS;
3284                 bp->b_flags |= B_ERROR;
3285                 biodone(ap->a_bio);
3286                 return(EROFS);
3287         }
3288
3289         lwkt_gettoken(&hmp->fs_token);
3290
3291         /*
3292          * Disallow swapcache operation on the vnode buffer if double
3293          * buffering is enabled, the swapcache will get the data via
3294          * the block device buffer.
3295          */
3296         if (hammer_double_buffer)
3297                 bp->b_flags |= B_NOTMETA;
3298
3299         /*
3300          * Interlock with inode destruction (no in-kernel or directory
3301          * topology visibility).  If we queue new IO while trying to
3302          * destroy the inode we can deadlock the vtrunc call in
3303          * hammer_inode_unloadable_check().
3304          *
3305          * Besides, there's no point flushing a bp associated with an
3306          * inode that is being destroyed on-media and has no kernel
3307          * references.
3308          */
3309         if ((ip->flags | ip->sync_flags) &
3310             (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) {
3311                 bp->b_resid = 0;
3312                 biodone(ap->a_bio);
3313                 lwkt_reltoken(&hmp->fs_token);
3314                 return(0);
3315         }
3316
3317         /*
3318          * Reserve space and issue a direct-write from the front-end.
3319          * NOTE: The direct_io code will hammer_bread/bcopy smaller
3320          * allocations.
3321          *
3322          * An in-memory record will be installed to reference the storage
3323          * until the flusher can get to it.
3324          *
3325          * Since we own the high level bio the front-end will not try to
3326          * do a direct-read until the write completes.
3327          *
3328          * NOTE: The only time we do not reserve a full-sized buffers
3329          * worth of data is if the file is small.  We do not try to
3330          * allocate a fragment (from the small-data zone) at the end of
3331          * an otherwise large file as this can lead to wildly separated
3332          * data.
3333          */
3334         KKASSERT((bio->bio_offset & HAMMER_BUFMASK) == 0);
3335         KKASSERT(bio->bio_offset < ip->ino_data.size);
3336         if (bio->bio_offset || ip->ino_data.size > HAMMER_BUFSIZE / 2)
3337                 bytes = bp->b_bufsize;
3338         else
3339                 bytes = ((int)ip->ino_data.size + 15) & ~15;
3340
3341         record = hammer_ip_add_bulk(ip, bio->bio_offset, bp->b_data,
3342                                     bytes, &error);
3343
3344         /*
3345          * B_VFSFLAG1 indicates that a REDO_WRITE entry was generated
3346          * in hammer_vop_write().  We must flag the record so the proper
3347          * REDO_TERM_WRITE entry is generated during the flush.
3348          */
3349         if (record) {
3350                 if (bp->b_flags & B_VFSFLAG1) {
3351                         record->flags |= HAMMER_RECF_REDO;
3352                         bp->b_flags &= ~B_VFSFLAG1;
3353                 }
3354                 if (record->flags & HAMMER_RECF_DEDUPED) {
3355                         bp->b_resid = 0;
3356                         hammer_ip_replace_bulk(hmp, record);
3357                         biodone(ap->a_bio);
3358                 } else {
3359                         hammer_io_direct_write(hmp, bio, record);
3360                 }
3361                 if (ip->rsv_recs > 1 && hmp->rsv_recs > hammer_limit_recs)
3362                         hammer_flush_inode(ip, 0);
3363         } else {
3364                 bp->b_bio2.bio_offset = NOOFFSET;
3365                 bp->b_error = error;
3366                 bp->b_flags |= B_ERROR;
3367                 biodone(ap->a_bio);
3368         }
3369         lwkt_reltoken(&hmp->fs_token);
3370         return(error);
3371 }
3372
3373 /*
3374  * dounlink - disconnect a directory entry
3375  *
3376  * XXX whiteout support not really in yet
3377  */
3378 static int
3379 hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
3380                 struct vnode *dvp, struct ucred *cred,
3381                 int flags, int isdir)
3382 {
3383         struct namecache *ncp;
3384         hammer_inode_t dip;
3385         hammer_inode_t ip;
3386         hammer_mount_t hmp;
3387         struct hammer_cursor cursor;
3388         int64_t namekey;
3389         u_int32_t max_iterations;
3390         int nlen, error;
3391
3392         /*
3393          * Calculate the namekey and setup the key range for the scan.  This
3394          * works kinda like a chained hash table where the lower 32 bits
3395          * of the namekey synthesize the chain.
3396          *
3397          * The key range is inclusive of both key_beg and key_end.
3398          */
3399         dip = VTOI(dvp);
3400         ncp = nch->ncp;
3401         hmp = dip->hmp;
3402
3403         if (dip->flags & HAMMER_INODE_RO)
3404                 return (EROFS);
3405
3406         namekey = hammer_directory_namekey(dip, ncp->nc_name, ncp->nc_nlen,
3407                                            &max_iterations);
3408 retry:
3409         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
3410         cursor.key_beg.localization = dip->obj_localization +
3411                                       hammer_dir_localization(dip);
3412         cursor.key_beg.obj_id = dip->obj_id;
3413         cursor.key_beg.key = namekey;
3414         cursor.key_beg.create_tid = 0;
3415         cursor.key_beg.delete_tid = 0;
3416         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
3417         cursor.key_beg.obj_type = 0;
3418
3419         cursor.key_end = cursor.key_beg;
3420         cursor.key_end.key += max_iterations;
3421         cursor.asof = dip->obj_asof;
3422         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
3423
3424         /*
3425          * Scan all matching records (the chain), locate the one matching
3426          * the requested path component.  info->last_error contains the
3427          * error code on search termination and could be 0, ENOENT, or
3428          * something else.
3429          *
3430          * The hammer_ip_*() functions merge in-memory records with on-disk
3431          * records for the purposes of the search.
3432          */
3433         error = hammer_ip_first(&cursor);
3434
3435         while (error == 0) {
3436                 error = hammer_ip_resolve_data(&cursor);
3437                 if (error)
3438                         break;
3439                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
3440                 KKASSERT(nlen > 0);
3441                 if (ncp->nc_nlen == nlen &&
3442                     bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
3443                         break;
3444                 }
3445                 error = hammer_ip_next(&cursor);
3446         }
3447
3448         /*
3449          * If all is ok we have to get the inode so we can adjust nlinks.
3450          * To avoid a deadlock with the flusher we must release the inode
3451          * lock on the directory when acquiring the inode for the entry.
3452          *
3453          * If the target is a directory, it must be empty.
3454          */
3455         if (error == 0) {
3456                 hammer_unlock(&cursor.ip->lock);
3457                 ip = hammer_get_inode(trans, dip, cursor.data->entry.obj_id,
3458                                       hmp->asof,
3459                                       cursor.data->entry.localization,
3460                                       0, &error);
3461                 hammer_lock_sh(&cursor.ip->lock);
3462                 if (error == ENOENT) {
3463                         kprintf("HAMMER: WARNING: Removing "
3464                                 "dirent w/missing inode \"%s\"\n"
3465                                 "\tobj_id = %016llx\n",
3466                                 ncp->nc_name,
3467                                 (long long)cursor.data->entry.obj_id);
3468                         error = 0;
3469                 }
3470
3471                 /*
3472                  * If isdir >= 0 we validate that the entry is or is not a
3473                  * directory.  If isdir < 0 we don't care.
3474                  */
3475                 if (error == 0 && isdir >= 0 && ip) {
3476                         if (isdir &&
3477                             ip->ino_data.obj_type != HAMMER_OBJTYPE_DIRECTORY) {
3478                                 error = ENOTDIR;
3479                         } else if (isdir == 0 &&
3480                             ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
3481                                 error = EISDIR;
3482                         }
3483                 }
3484
3485                 /*
3486                  * If we are trying to remove a directory the directory must
3487                  * be empty.
3488                  *
3489                  * The check directory code can loop and deadlock/retry.  Our
3490                  * own cursor's node locks must be released to avoid a 3-way
3491                  * deadlock with the flusher if the check directory code
3492                  * blocks.
3493                  *
3494                  * If any changes whatsoever have been made to the cursor
3495                  * set EDEADLK and retry.
3496                  *
3497                  * WARNING: See warnings in hammer_unlock_cursor()
3498                  *          function.
3499                  */
3500                 if (error == 0 && ip && ip->ino_data.obj_type ==
3501                                         HAMMER_OBJTYPE_DIRECTORY) {
3502                         hammer_unlock_cursor(&cursor);
3503                         error = hammer_ip_check_directory_empty(trans, ip);
3504                         hammer_lock_cursor(&cursor);
3505                         if (cursor.flags & HAMMER_CURSOR_RETEST) {
3506                                 kprintf("HAMMER: Warning: avoided deadlock "
3507                                         "on rmdir '%s'\n",
3508                                         ncp->nc_name);
3509                                 error = EDEADLK;
3510                         }
3511                 }
3512
3513                 /*
3514                  * Delete the directory entry.
3515                  *
3516                  * WARNING: hammer_ip_del_directory() may have to terminate
3517                  * the cursor to avoid a deadlock.  It is ok to call
3518                  * hammer_done_cursor() twice.
3519                  */
3520                 if (error == 0) {
3521                         error = hammer_ip_del_directory(trans, &cursor,
3522                                                         dip, ip);
3523                 }
3524                 hammer_done_cursor(&cursor);
3525                 if (error == 0) {
3526                         cache_setunresolved(nch);
3527                         cache_setvp(nch, NULL);
3528
3529                         /*
3530                          * NOTE: ip->vp, if non-NULL, cannot be directly
3531                          *       referenced without formally acquiring the
3532                          *       vp since the vp might have zero refs on it,
3533                          *       or in the middle of a reclaim, etc.
3534                          *
3535                          * NOTE: The cache_setunresolved() can rip the vp
3536                          *       out from under us since the vp may not have
3537                          *       any refs, in which case ip->vp will be NULL
3538                          *       from the outset.
3539                          */
3540                         while (ip && ip->vp) {
3541                                 struct vnode *vp;
3542
3543                                 error = hammer_get_vnode(ip, &vp);
3544                                 if (error == 0 && vp) {
3545                                         vn_unlock(vp);
3546                                         hammer_knote(ip->vp, NOTE_DELETE);
3547                                         cache_inval_vp(ip->vp, CINV_DESTROY);
3548                                         vrele(vp);
3549                                         break;
3550                                 }
3551                                 kprintf("Debug: HAMMER ip/vp race1 avoided\n");
3552                         }
3553                 }
3554                 if (ip)
3555                         hammer_rel_inode(ip, 0);
3556         } else {
3557                 hammer_done_cursor(&cursor);
3558         }
3559         if (error == EDEADLK)
3560                 goto retry;
3561
3562         return (error);
3563 }
3564
3565 /************************************************************************
3566  *                          FIFO AND SPECFS OPS                         *
3567  ************************************************************************
3568  *
3569  */
3570 static int
3571 hammer_vop_fifoclose (struct vop_close_args *ap)
3572 {
3573         /* XXX update itimes */
3574         return (VOCALL(&fifo_vnode_vops, &ap->a_head));
3575 }
3576
3577 static int
3578 hammer_vop_fiforead (struct vop_read_args *ap)
3579 {
3580         int error;
3581
3582         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3583         /* XXX update access time */
3584         return (error);
3585 }
3586
3587 static int
3588 hammer_vop_fifowrite (struct vop_write_args *ap)
3589 {
3590         int error;
3591
3592         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3593         /* XXX update access time */
3594         return (error);
3595 }
3596
3597 static
3598 int
3599 hammer_vop_fifokqfilter(struct vop_kqfilter_args *ap)
3600 {
3601         int error;
3602
3603         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3604         if (error)
3605                 error = hammer_vop_kqfilter(ap);
3606         return(error);
3607 }
3608
3609 /************************************************************************
3610  *                          KQFILTER OPS                                *
3611  ************************************************************************
3612  *
3613  */
3614 static void filt_hammerdetach(struct knote *kn);
3615 static int filt_hammerread(struct knote *kn, long hint);
3616 static int filt_hammerwrite(struct knote *kn, long hint);
3617 static int filt_hammervnode(struct knote *kn, long hint);
3618
3619 static struct filterops hammerread_filtops =
3620         { FILTEROP_ISFD, NULL, filt_hammerdetach, filt_hammerread };
3621 static struct filterops hammerwrite_filtops =
3622         { FILTEROP_ISFD, NULL, filt_hammerdetach, filt_hammerwrite };
3623 static struct filterops hammervnode_filtops =
3624         { FILTEROP_ISFD, NULL, filt_hammerdetach, filt_hammervnode };
3625
3626 static
3627 int
3628 hammer_vop_kqfilter(struct vop_kqfilter_args *ap)
3629 {
3630         struct vnode *vp = ap->a_vp;
3631         struct knote *kn = ap->a_kn;
3632
3633         switch (kn->kn_filter) {
3634         case EVFILT_READ:
3635                 kn->kn_fop = &hammerread_filtops;
3636                 break;
3637         case EVFILT_WRITE:
3638                 kn->kn_fop = &hammerwrite_filtops;
3639                 break;
3640         case EVFILT_VNODE:
3641                 kn->kn_fop = &hammervnode_filtops;
3642                 break;
3643         default:
3644                 return (EOPNOTSUPP);
3645         }
3646
3647         kn->kn_hook = (caddr_t)vp;
3648
3649         knote_insert(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3650
3651         return(0);
3652 }
3653
3654 static void
3655 filt_hammerdetach(struct knote *kn)
3656 {
3657         struct vnode *vp = (void *)kn->kn_hook;
3658
3659         knote_remove(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3660 }
3661
3662 static int
3663 filt_hammerread(struct knote *kn, long hint)
3664 {
3665         struct vnode *vp = (void *)kn->kn_hook;
3666         hammer_inode_t ip = VTOI(vp);
3667         hammer_mount_t hmp = ip->hmp;
3668         off_t off;
3669
3670         if (hint == NOTE_REVOKE) {
3671                 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
3672                 return(1);
3673         }
3674         lwkt_gettoken(&hmp->fs_token);  /* XXX use per-ip-token */
3675         off = ip->ino_data.size - kn->kn_fp->f_offset;
3676         kn->kn_data = (off < INTPTR_MAX) ? off : INTPTR_MAX;
3677         lwkt_reltoken(&hmp->fs_token);
3678         if (kn->kn_sfflags & NOTE_OLDAPI)
3679                 return(1);
3680         return (kn->kn_data != 0);
3681 }
3682
3683 static int
3684 filt_hammerwrite(struct knote *kn, long hint)
3685 {
3686         if (hint == NOTE_REVOKE)
3687                 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
3688         kn->kn_data = 0;
3689         return (1);
3690 }
3691
3692 static int
3693 filt_hammervnode(struct knote *kn, long hint)
3694 {
3695         if (kn->kn_sfflags & hint)
3696                 kn->kn_fflags |= hint;
3697         if (hint == NOTE_REVOKE) {
3698                 kn->kn_flags |= (EV_EOF | EV_NODATA);
3699                 return (1);
3700         }
3701         return (kn->kn_fflags != 0);
3702 }
3703