sys/vfs/hammer/hammer_vnops.c

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