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 1.101 2008/10/15 22:38:37 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 <vm/vm_extern.h>
  48 #include <vfs/fifofs/fifo.h>
  49 #include "hammer.h"
  50
  51 /*
  52  * USERFS VNOPS
  53  */
  54 /*static int hammer_vop_vnoperate(struct vop_generic_args *);*/
  55 static int hammer_vop_fsync(struct vop_fsync_args *);
  56 static int hammer_vop_read(struct vop_read_args *);
  57 static int hammer_vop_write(struct vop_write_args *);
  58 static int hammer_vop_access(struct vop_access_args *);
  59 static int hammer_vop_advlock(struct vop_advlock_args *);
  60 static int hammer_vop_close(struct vop_close_args *);
  61 static int hammer_vop_ncreate(struct vop_ncreate_args *);
  62 static int hammer_vop_getattr(struct vop_getattr_args *);
  63 static int hammer_vop_nresolve(struct vop_nresolve_args *);
  64 static int hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *);
  65 static int hammer_vop_nlink(struct vop_nlink_args *);
  66 static int hammer_vop_nmkdir(struct vop_nmkdir_args *);
  67 static int hammer_vop_nmknod(struct vop_nmknod_args *);
  68 static int hammer_vop_open(struct vop_open_args *);
  69 static int hammer_vop_print(struct vop_print_args *);
  70 static int hammer_vop_readdir(struct vop_readdir_args *);
  71 static int hammer_vop_readlink(struct vop_readlink_args *);
  72 static int hammer_vop_nremove(struct vop_nremove_args *);
  73 static int hammer_vop_nrename(struct vop_nrename_args *);
  74 static int hammer_vop_nrmdir(struct vop_nrmdir_args *);
  75 static int hammer_vop_setattr(struct vop_setattr_args *);
  76 static int hammer_vop_strategy(struct vop_strategy_args *);
  77 static int hammer_vop_bmap(struct vop_bmap_args *ap);
  78 static int hammer_vop_nsymlink(struct vop_nsymlink_args *);
  79 static int hammer_vop_nwhiteout(struct vop_nwhiteout_args *);
  80 static int hammer_vop_ioctl(struct vop_ioctl_args *);
  81 static int hammer_vop_mountctl(struct vop_mountctl_args *);
  82
  83 static int hammer_vop_fifoclose (struct vop_close_args *);
  84 static int hammer_vop_fiforead (struct vop_read_args *);
  85 static int hammer_vop_fifowrite (struct vop_write_args *);
  86
  87 static int hammer_vop_specclose (struct vop_close_args *);
  88 static int hammer_vop_specread (struct vop_read_args *);
  89 static int hammer_vop_specwrite (struct vop_write_args *);
  90
  91 struct vop_ops hammer_vnode_vops = {
  92         .vop_default =          vop_defaultop,
  93         .vop_fsync =            hammer_vop_fsync,
  94         .vop_getpages =         vop_stdgetpages,
  95         .vop_putpages =         vop_stdputpages,
  96         .vop_read =             hammer_vop_read,
  97         .vop_write =            hammer_vop_write,
  98         .vop_access =           hammer_vop_access,
  99         .vop_advlock =          hammer_vop_advlock,
 100         .vop_close =            hammer_vop_close,
 101         .vop_ncreate =          hammer_vop_ncreate,
 102         .vop_getattr =          hammer_vop_getattr,
 103         .vop_inactive =         hammer_vop_inactive,
 104         .vop_reclaim =          hammer_vop_reclaim,
 105         .vop_nresolve =         hammer_vop_nresolve,
 106         .vop_nlookupdotdot =    hammer_vop_nlookupdotdot,
 107         .vop_nlink =            hammer_vop_nlink,
 108         .vop_nmkdir =           hammer_vop_nmkdir,
 109         .vop_nmknod =           hammer_vop_nmknod,
 110         .vop_open =             hammer_vop_open,
 111         .vop_pathconf =         vop_stdpathconf,
 112         .vop_print =            hammer_vop_print,
 113         .vop_readdir =          hammer_vop_readdir,
 114         .vop_readlink =         hammer_vop_readlink,
 115         .vop_nremove =          hammer_vop_nremove,
 116         .vop_nrename =          hammer_vop_nrename,
 117         .vop_nrmdir =           hammer_vop_nrmdir,
 118         .vop_setattr =          hammer_vop_setattr,
 119         .vop_bmap =             hammer_vop_bmap,
 120         .vop_strategy =         hammer_vop_strategy,
 121         .vop_nsymlink =         hammer_vop_nsymlink,
 122         .vop_nwhiteout =        hammer_vop_nwhiteout,
 123         .vop_ioctl =            hammer_vop_ioctl,
 124         .vop_mountctl =         hammer_vop_mountctl
 125 };
 126
 127 struct vop_ops hammer_spec_vops = {
 128         .vop_default =          spec_vnoperate,
 129         .vop_fsync =            hammer_vop_fsync,
 130         .vop_read =             hammer_vop_specread,
 131         .vop_write =            hammer_vop_specwrite,
 132         .vop_access =           hammer_vop_access,
 133         .vop_close =            hammer_vop_specclose,
 134         .vop_getattr =          hammer_vop_getattr,
 135         .vop_inactive =         hammer_vop_inactive,
 136         .vop_reclaim =          hammer_vop_reclaim,
 137         .vop_setattr =          hammer_vop_setattr
 138 };
 139
 140 struct vop_ops hammer_fifo_vops = {
 141         .vop_default =          fifo_vnoperate,
 142         .vop_fsync =            hammer_vop_fsync,
 143         .vop_read =             hammer_vop_fiforead,
 144         .vop_write =            hammer_vop_fifowrite,
 145         .vop_access =           hammer_vop_access,
 146         .vop_close =            hammer_vop_fifoclose,
 147         .vop_getattr =          hammer_vop_getattr,
 148         .vop_inactive =         hammer_vop_inactive,
 149         .vop_reclaim =          hammer_vop_reclaim,
 150         .vop_setattr =          hammer_vop_setattr
 151 };
 152
 153 #ifdef DEBUG_TRUNCATE
 154 struct hammer_inode *HammerTruncIp;
 155 #endif
 156
 157 static int hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
 158                            struct vnode *dvp, struct ucred *cred,
 159                            int flags, int isdir);
 160 static int hammer_vop_strategy_read(struct vop_strategy_args *ap);
 161 static int hammer_vop_strategy_write(struct vop_strategy_args *ap);
 162
 163 #if 0
 164 static
 165 int
 166 hammer_vop_vnoperate(struct vop_generic_args *)
 167 {
 168         return (VOCALL(&hammer_vnode_vops, ap));
 169 }
 170 #endif
 171
 172 /*
 173  * hammer_vop_fsync { vp, waitfor }
 174  *
 175  * fsync() an inode to disk and wait for it to be completely committed
 176  * such that the information would not be undone if a crash occured after
 177  * return.
 178  */
 179 static
 180 int
 181 hammer_vop_fsync(struct vop_fsync_args *ap)
 182 {
 183         hammer_inode_t ip = VTOI(ap->a_vp);
 184
 185         ++hammer_count_fsyncs;
 186         vfsync(ap->a_vp, ap->a_waitfor, 1, NULL, NULL);
 187         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 188         if (ap->a_waitfor == MNT_WAIT) {
 189                 vn_unlock(ap->a_vp);
 190                 hammer_wait_inode(ip);
 191                 vn_lock(ap->a_vp, LK_EXCLUSIVE | LK_RETRY);
 192         }
 193         return (ip->error);
 194 }
 195
 196 /*
 197  * hammer_vop_read { vp, uio, ioflag, cred }
 198  */
 199 static
 200 int
 201 hammer_vop_read(struct vop_read_args *ap)
 202 {
 203         struct hammer_transaction trans;
 204         hammer_inode_t ip;
 205         off_t offset;
 206         struct buf *bp;
 207         struct uio *uio;
 208         int error;
 209         int n;
 210         int seqcount;
 211         int ioseqcount;
 212         int blksize;
 213
 214         if (ap->a_vp->v_type != VREG)
 215                 return (EINVAL);
 216         ip = VTOI(ap->a_vp);
 217         error = 0;
 218         uio = ap->a_uio;
 219
 220         /*
 221          * Allow the UIO's size to override the sequential heuristic.
 222          */
 223         blksize = hammer_blocksize(uio->uio_offset);
 224         seqcount = (uio->uio_resid + (blksize - 1)) / blksize;
 225         ioseqcount = ap->a_ioflag >> 16;
 226         if (seqcount < ioseqcount)
 227                 seqcount = ioseqcount;
 228
 229         hammer_start_transaction(&trans, ip->hmp);
 230
 231         /*
 232          * Access the data typically in HAMMER_BUFSIZE blocks via the
 233          * buffer cache, but HAMMER may use a variable block size based
 234          * on the offset.
 235          */
 236         while (uio->uio_resid > 0 && uio->uio_offset < ip->ino_data.size) {
 237                 int64_t base_offset;
 238                 int64_t file_limit;
 239
 240                 blksize = hammer_blocksize(uio->uio_offset);
 241                 offset = (int)uio->uio_offset & (blksize - 1);
 242                 base_offset = uio->uio_offset - offset;
 243
 244                 if (hammer_cluster_enable) {
 245                         /*
 246                          * Use file_limit to prevent cluster_read() from
 247                          * creating buffers of the wrong block size past
 248                          * the demarc.
 249                          */
 250                         file_limit = ip->ino_data.size;
 251                         if (base_offset < HAMMER_XDEMARC &&
 252                             file_limit > HAMMER_XDEMARC) {
 253                                 file_limit = HAMMER_XDEMARC;
 254                         }
 255                         error = cluster_read(ap->a_vp,
 256                                              file_limit, base_offset,
 257                                              blksize, MAXPHYS,
 258                                              seqcount, &bp);
 259                 } else {
 260                         error = bread(ap->a_vp, base_offset, blksize, &bp);
 261                 }
 262                 if (error) {
 263                         kprintf("error %d\n", error);
 264                         brelse(bp);
 265                         break;
 266                 }
 267
 268                 /* bp->b_flags |= B_CLUSTEROK; temporarily disabled */
 269                 n = blksize - offset;
 270                 if (n > uio->uio_resid)
 271                         n = uio->uio_resid;
 272                 if (n > ip->ino_data.size - uio->uio_offset)
 273                         n = (int)(ip->ino_data.size - uio->uio_offset);
 274                 error = uiomove((char *)bp->b_data + offset, n, uio);
 275
 276                 /* data has a lower priority then meta-data */
 277                 bp->b_flags |= B_AGE;
 278                 bqrelse(bp);
 279                 if (error)
 280                         break;
 281                 hammer_stats_file_read += n;
 282         }
 283         if ((ip->flags & HAMMER_INODE_RO) == 0 &&
 284             (ip->hmp->mp->mnt_flag & MNT_NOATIME) == 0) {
 285                 ip->ino_data.atime = trans.time;
 286                 hammer_modify_inode(ip, HAMMER_INODE_ATIME);
 287         }
 288         hammer_done_transaction(&trans);
 289         return (error);
 290 }
 291
 292 /*
 293  * hammer_vop_write { vp, uio, ioflag, cred }
 294  */
 295 static
 296 int
 297 hammer_vop_write(struct vop_write_args *ap)
 298 {
 299         struct hammer_transaction trans;
 300         struct hammer_inode *ip;
 301         hammer_mount_t hmp;
 302         struct uio *uio;
 303         int offset;
 304         off_t base_offset;
 305         struct buf *bp;
 306         int error;
 307         int n;
 308         int flags;
 309         int delta;
 310         int seqcount;
 311
 312         if (ap->a_vp->v_type != VREG)
 313                 return (EINVAL);
 314         ip = VTOI(ap->a_vp);
 315         hmp = ip->hmp;
 316         error = 0;
 317         seqcount = ap->a_ioflag >> 16;
 318
 319         if (ip->flags & HAMMER_INODE_RO)
 320                 return (EROFS);
 321
 322         /*
 323          * Create a transaction to cover the operations we perform.
 324          */
 325         hammer_start_transaction(&trans, hmp);
 326         uio = ap->a_uio;
 327
 328         /*
 329          * Check append mode
 330          */
 331         if (ap->a_ioflag & IO_APPEND)
 332                 uio->uio_offset = ip->ino_data.size;
 333
 334         /*
 335          * Check for illegal write offsets.  Valid range is 0...2^63-1.
 336          *
 337          * NOTE: the base_off assignment is required to work around what
 338          * I consider to be a GCC-4 optimization bug.
 339          */
 340         if (uio->uio_offset < 0) {
 341                 hammer_done_transaction(&trans);
 342                 return (EFBIG);
 343         }
 344         base_offset = uio->uio_offset + uio->uio_resid; /* work around gcc-4 */
 345         if (uio->uio_resid > 0 && base_offset <= 0) {
 346                 hammer_done_transaction(&trans);
 347                 return (EFBIG);
 348         }
 349
 350         /*
 351          * Access the data typically in HAMMER_BUFSIZE blocks via the
 352          * buffer cache, but HAMMER may use a variable block size based
 353          * on the offset.
 354          */
 355         while (uio->uio_resid > 0) {
 356                 int fixsize = 0;
 357                 int blksize;
 358                 int blkmask;
 359
 360                 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_WRITE)) != 0)
 361                         break;
 362
 363                 blksize = hammer_blocksize(uio->uio_offset);
 364
 365                 /*
 366                  * Do not allow HAMMER to blow out the buffer cache.  Very
 367                  * large UIOs can lockout other processes due to bwillwrite()
 368                  * mechanics.
 369                  *
 370                  * The hammer inode is not locked during these operations.
 371                  * The vnode is locked which can interfere with the pageout
 372                  * daemon for non-UIO_NOCOPY writes but should not interfere
 373                  * with the buffer cache.  Even so, we cannot afford to
 374                  * allow the pageout daemon to build up too many dirty buffer
 375                  * cache buffers.
 376                  */
 377                 /*if (((int)uio->uio_offset & (blksize - 1)) == 0)*/
 378                 bwillwrite(blksize);
 379
 380                 /*
 381                  * Do not allow HAMMER to blow out system memory by
 382                  * accumulating too many records.   Records are so well
 383                  * decoupled from the buffer cache that it is possible
 384                  * for userland to push data out to the media via
 385                  * direct-write, but build up the records queued to the
 386                  * backend faster then the backend can flush them out.
 387                  * HAMMER has hit its write limit but the frontend has
 388                  * no pushback to slow it down.
 389                  */
 390                 if (hmp->rsv_recs > hammer_limit_recs / 2) {
 391                         /*
 392                          * Get the inode on the flush list
 393                          */
 394                         if (ip->rsv_recs >= 64)
 395                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 396                         else if (ip->rsv_recs >= 16)
 397                                 hammer_flush_inode(ip, 0);
 398
 399                         /*
 400                          * Keep the flusher going if the system keeps
 401                          * queueing records.
 402                          */
 403                         delta = hmp->count_newrecords -
 404                                 hmp->last_newrecords;
 405                         if (delta < 0 || delta > hammer_limit_recs / 2) {
 406                                 hmp->last_newrecords = hmp->count_newrecords;
 407                                 hammer_sync_hmp(hmp, MNT_NOWAIT);
 408                         }
 409
 410                         /*
 411                          * If we have gotten behind start slowing
 412                          * down the writers.
 413                          */
 414                         delta = (hmp->rsv_recs - hammer_limit_recs) *
 415                                 hz / hammer_limit_recs;
 416                         if (delta > 0)
 417                                 tsleep(&trans, 0, "hmrslo", delta);
 418                 }
 419
 420                 /*
 421                  * Calculate the blocksize at the current offset and figure
 422                  * out how much we can actually write.
 423                  */
 424                 blkmask = blksize - 1;
 425                 offset = (int)uio->uio_offset & blkmask;
 426                 base_offset = uio->uio_offset & ~(int64_t)blkmask;
 427                 n = blksize - offset;
 428                 if (n > uio->uio_resid)
 429                         n = uio->uio_resid;
 430                 if (uio->uio_offset + n > ip->ino_data.size) {
 431                         vnode_pager_setsize(ap->a_vp, uio->uio_offset + n);
 432                         fixsize = 1;
 433                 }
 434
 435                 if (uio->uio_segflg == UIO_NOCOPY) {
 436                         /*
 437                          * Issuing a write with the same data backing the
 438                          * buffer.  Instantiate the buffer to collect the
 439                          * backing vm pages, then read-in any missing bits.
 440                          *
 441                          * This case is used by vop_stdputpages().
 442                          */
 443                         bp = getblk(ap->a_vp, base_offset,
 444                                     blksize, GETBLK_BHEAVY, 0);
 445                         if ((bp->b_flags & B_CACHE) == 0) {
 446                                 bqrelse(bp);
 447                                 error = bread(ap->a_vp, base_offset,
 448                                               blksize, &bp);
 449                         }
 450                 } else if (offset == 0 && uio->uio_resid >= blksize) {
 451                         /*
 452                          * Even though we are entirely overwriting the buffer
 453                          * we may still have to zero it out to avoid a
 454                          * mmap/write visibility issue.
 455                          */
 456                         bp = getblk(ap->a_vp, base_offset, blksize, GETBLK_BHEAVY, 0);
 457                         if ((bp->b_flags & B_CACHE) == 0)
 458                                 vfs_bio_clrbuf(bp);
 459                 } else if (base_offset >= ip->ino_data.size) {
 460                         /*
 461                          * If the base offset of the buffer is beyond the
 462                          * file EOF, we don't have to issue a read.
 463                          */
 464                         bp = getblk(ap->a_vp, base_offset,
 465                                     blksize, GETBLK_BHEAVY, 0);
 466                         vfs_bio_clrbuf(bp);
 467                 } else {
 468                         /*
 469                          * Partial overwrite, read in any missing bits then
 470                          * replace the portion being written.
 471                          */
 472                         error = bread(ap->a_vp, base_offset, blksize, &bp);
 473                         if (error == 0)
 474                                 bheavy(bp);
 475                 }
 476                 if (error == 0) {
 477                         error = uiomove((char *)bp->b_data + offset,
 478                                         n, uio);
 479                 }
 480
 481                 /*
 482                  * If we screwed up we have to undo any VM size changes we
 483                  * made.
 484                  */
 485                 if (error) {
 486                         brelse(bp);
 487                         if (fixsize) {
 488                                 vtruncbuf(ap->a_vp, ip->ino_data.size,
 489                                           hammer_blocksize(ip->ino_data.size));
 490                         }
 491                         break;
 492                 }
 493                 hammer_stats_file_write += n;
 494                 /* bp->b_flags |= B_CLUSTEROK; temporarily disabled */
 495                 if (ip->ino_data.size < uio->uio_offset) {
 496                         ip->ino_data.size = uio->uio_offset;
 497                         flags = HAMMER_INODE_DDIRTY;
 498                         vnode_pager_setsize(ap->a_vp, ip->ino_data.size);
 499                 } else {
 500                         flags = 0;
 501                 }
 502                 ip->ino_data.mtime = trans.time;
 503                 flags |= HAMMER_INODE_MTIME | HAMMER_INODE_BUFS;
 504                 hammer_modify_inode(ip, flags);
 505
 506                 /*
 507                  * Once we dirty the buffer any cached zone-X offset
 508                  * becomes invalid.  HAMMER NOTE: no-history mode cannot
 509                  * allow overwriting over the same data sector unless
 510                  * we provide UNDOs for the old data, which we don't.
 511                  */
 512                 bp->b_bio2.bio_offset = NOOFFSET;
 513
 514                 /*
 515                  * Final buffer disposition.
 516                  */
 517                 bp->b_flags |= B_AGE;
 518                 if (ap->a_ioflag & IO_SYNC) {
 519                         bwrite(bp);
 520                 } else if (ap->a_ioflag & IO_DIRECT) {
 521                         bawrite(bp);
 522                 } else {
 523                         bdwrite(bp);
 524                 }
 525         }
 526         hammer_done_transaction(&trans);
 527         return (error);
 528 }
 529
 530 /*
 531  * hammer_vop_access { vp, mode, cred }
 532  */
 533 static
 534 int
 535 hammer_vop_access(struct vop_access_args *ap)
 536 {
 537         struct hammer_inode *ip = VTOI(ap->a_vp);
 538         uid_t uid;
 539         gid_t gid;
 540         int error;
 541
 542         ++hammer_stats_file_iopsr;
 543         uid = hammer_to_unix_xid(&ip->ino_data.uid);
 544         gid = hammer_to_unix_xid(&ip->ino_data.gid);
 545
 546         error = vop_helper_access(ap, uid, gid, ip->ino_data.mode,
 547                                   ip->ino_data.uflags);
 548         return (error);
 549 }
 550
 551 /*
 552  * hammer_vop_advlock { vp, id, op, fl, flags }
 553  */
 554 static
 555 int
 556 hammer_vop_advlock(struct vop_advlock_args *ap)
 557 {
 558         hammer_inode_t ip = VTOI(ap->a_vp);
 559
 560         return (lf_advlock(ap, &ip->advlock, ip->ino_data.size));
 561 }
 562
 563 /*
 564  * hammer_vop_close { vp, fflag }
 565  */
 566 static
 567 int
 568 hammer_vop_close(struct vop_close_args *ap)
 569 {
 570         /*hammer_inode_t ip = VTOI(ap->a_vp);*/
 571         return (vop_stdclose(ap));
 572 }
 573
 574 /*
 575  * hammer_vop_ncreate { nch, dvp, vpp, cred, vap }
 576  *
 577  * The operating system has already ensured that the directory entry
 578  * does not exist and done all appropriate namespace locking.
 579  */
 580 static
 581 int
 582 hammer_vop_ncreate(struct vop_ncreate_args *ap)
 583 {
 584         struct hammer_transaction trans;
 585         struct hammer_inode *dip;
 586         struct hammer_inode *nip;
 587         struct nchandle *nch;
 588         int error;
 589
 590         nch = ap->a_nch;
 591         dip = VTOI(ap->a_dvp);
 592
 593         if (dip->flags & HAMMER_INODE_RO)
 594                 return (EROFS);
 595         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
 596                 return (error);
 597
 598         /*
 599          * Create a transaction to cover the operations we perform.
 600          */
 601         hammer_start_transaction(&trans, dip->hmp);
 602         ++hammer_stats_file_iopsw;
 603
 604         /*
 605          * Create a new filesystem object of the requested type.  The
 606          * returned inode will be referenced and shared-locked to prevent
 607          * it from being moved to the flusher.
 608          */
 609
 610         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
 611                                     dip, NULL, &nip);
 612         if (error) {
 613                 hkprintf("hammer_create_inode error %d\n", error);
 614                 hammer_done_transaction(&trans);
 615                 *ap->a_vpp = NULL;
 616                 return (error);
 617         }
 618
 619         /*
 620          * Add the new filesystem object to the directory.  This will also
 621          * bump the inode's link count.
 622          */
 623         error = hammer_ip_add_directory(&trans, dip,
 624                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
 625                                         nip);
 626         if (error)
 627                 hkprintf("hammer_ip_add_directory error %d\n", error);
 628
 629         /*
 630          * Finish up.
 631          */
 632         if (error) {
 633                 hammer_rel_inode(nip, 0);
 634                 hammer_done_transaction(&trans);
 635                 *ap->a_vpp = NULL;
 636         } else {
 637                 error = hammer_get_vnode(nip, ap->a_vpp);
 638                 hammer_done_transaction(&trans);
 639                 hammer_rel_inode(nip, 0);
 640                 if (error == 0) {
 641                         cache_setunresolved(ap->a_nch);
 642                         cache_setvp(ap->a_nch, *ap->a_vpp);
 643                 }
 644         }
 645         return (error);
 646 }
 647
 648 /*
 649  * hammer_vop_getattr { vp, vap }
 650  *
 651  * Retrieve an inode's attribute information.  When accessing inodes
 652  * historically we fake the atime field to ensure consistent results.
 653  * The atime field is stored in the B-Tree element and allowed to be
 654  * updated without cycling the element.
 655  */
 656 static
 657 int
 658 hammer_vop_getattr(struct vop_getattr_args *ap)
 659 {
 660         struct hammer_inode *ip = VTOI(ap->a_vp);
 661         struct vattr *vap = ap->a_vap;
 662
 663         /*
 664          * We want the fsid to be different when accessing a filesystem
 665          * with different as-of's so programs like diff don't think
 666          * the files are the same.
 667          *
 668          * We also want the fsid to be the same when comparing snapshots,
 669          * or when comparing mirrors (which might be backed by different
 670          * physical devices).  HAMMER fsids are based on the PFS's
 671          * shared_uuid field.
 672          *
 673          * XXX there is a chance of collision here.  The va_fsid reported
 674          * by stat is different from the more involved fsid used in the
 675          * mount structure.
 676          */
 677         ++hammer_stats_file_iopsr;
 678         vap->va_fsid = ip->pfsm->fsid_udev ^ (u_int32_t)ip->obj_asof ^
 679                        (u_int32_t)(ip->obj_asof >> 32);
 680
 681         vap->va_fileid = ip->ino_leaf.base.obj_id;
 682         vap->va_mode = ip->ino_data.mode;
 683         vap->va_nlink = ip->ino_data.nlinks;
 684         vap->va_uid = hammer_to_unix_xid(&ip->ino_data.uid);
 685         vap->va_gid = hammer_to_unix_xid(&ip->ino_data.gid);
 686         vap->va_rmajor = 0;
 687         vap->va_rminor = 0;
 688         vap->va_size = ip->ino_data.size;
 689
 690         /*
 691          * Special case for @@PFS softlinks.  The actual size of the
 692          * expanded softlink is "@@0x%016llx:%05d" == 26 bytes.
 693          */
 694         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_SOFTLINK &&
 695             ip->ino_data.size == 10 &&
 696             ip->obj_asof == HAMMER_MAX_TID &&
 697             ip->obj_localization == 0 &&
 698             strncmp(ip->ino_data.ext.symlink, "@@PFS", 5) == 0) {
 699                     vap->va_size = 26;
 700         }
 701
 702         /*
 703          * We must provide a consistent atime and mtime for snapshots
 704          * so people can do a 'tar cf - ... | md5' on them and get
 705          * consistent results.
 706          */
 707         if (ip->flags & HAMMER_INODE_RO) {
 708                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_atime);
 709                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_mtime);
 710         } else {
 711                 hammer_time_to_timespec(ip->ino_data.atime, &vap->va_atime);
 712                 hammer_time_to_timespec(ip->ino_data.mtime, &vap->va_mtime);
 713         }
 714         hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_ctime);
 715         vap->va_flags = ip->ino_data.uflags;
 716         vap->va_gen = 1;        /* hammer inums are unique for all time */
 717         vap->va_blocksize = HAMMER_BUFSIZE;
 718         if (ip->ino_data.size >= HAMMER_XDEMARC) {
 719                 vap->va_bytes = (ip->ino_data.size + HAMMER_XBUFMASK64) &
 720                                 ~HAMMER_XBUFMASK64;
 721         } else if (ip->ino_data.size > HAMMER_BUFSIZE / 2) {
 722                 vap->va_bytes = (ip->ino_data.size + HAMMER_BUFMASK64) &
 723                                 ~HAMMER_BUFMASK64;
 724         } else {
 725                 vap->va_bytes = (ip->ino_data.size + 15) & ~15;
 726         }
 727
 728         vap->va_type = hammer_get_vnode_type(ip->ino_data.obj_type);
 729         vap->va_filerev = 0;    /* XXX */
 730         /* mtime uniquely identifies any adjustments made to the file XXX */
 731         vap->va_fsmid = ip->ino_data.mtime;
 732         vap->va_uid_uuid = ip->ino_data.uid;
 733         vap->va_gid_uuid = ip->ino_data.gid;
 734         vap->va_fsid_uuid = ip->hmp->fsid;
 735         vap->va_vaflags = VA_UID_UUID_VALID | VA_GID_UUID_VALID |
 736                           VA_FSID_UUID_VALID;
 737
 738         switch (ip->ino_data.obj_type) {
 739         case HAMMER_OBJTYPE_CDEV:
 740         case HAMMER_OBJTYPE_BDEV:
 741                 vap->va_rmajor = ip->ino_data.rmajor;
 742                 vap->va_rminor = ip->ino_data.rminor;
 743                 break;
 744         default:
 745                 break;
 746         }
 747         return(0);
 748 }
 749
 750 /*
 751  * hammer_vop_nresolve { nch, dvp, cred }
 752  *
 753  * Locate the requested directory entry.
 754  */
 755 static
 756 int
 757 hammer_vop_nresolve(struct vop_nresolve_args *ap)
 758 {
 759         struct hammer_transaction trans;
 760         struct namecache *ncp;
 761         hammer_inode_t dip;
 762         hammer_inode_t ip;
 763         hammer_tid_t asof;
 764         struct hammer_cursor cursor;
 765         struct vnode *vp;
 766         int64_t namekey;
 767         int error;
 768         int i;
 769         int nlen;
 770         int flags;
 771         int ispfs;
 772         int64_t obj_id;
 773         u_int32_t localization;
 774
 775         /*
 776          * Misc initialization, plus handle as-of name extensions.  Look for
 777          * the '@@' extension.  Note that as-of files and directories cannot
 778          * be modified.
 779          */
 780         dip = VTOI(ap->a_dvp);
 781         ncp = ap->a_nch->ncp;
 782         asof = dip->obj_asof;
 783         localization = dip->obj_localization;   /* for code consistency */
 784         nlen = ncp->nc_nlen;
 785         flags = dip->flags & HAMMER_INODE_RO;
 786         ispfs = 0;
 787
 788         hammer_simple_transaction(&trans, dip->hmp);
 789         ++hammer_stats_file_iopsr;
 790
 791         for (i = 0; i < nlen; ++i) {
 792                 if (ncp->nc_name[i] == '@' && ncp->nc_name[i+1] == '@') {
 793                         error = hammer_str_to_tid(ncp->nc_name + i + 2,
 794                                                   &ispfs, &asof, &localization);
 795                         if (error != 0) {
 796                                 i = nlen;
 797                                 break;
 798                         }
 799                         if (asof != HAMMER_MAX_TID)
 800                                 flags |= HAMMER_INODE_RO;
 801                         break;
 802                 }
 803         }
 804         nlen = i;
 805
 806         /*
 807          * If this is a PFS softlink we dive into the PFS
 808          */
 809         if (ispfs && nlen == 0) {
 810                 ip = hammer_get_inode(&trans, dip, HAMMER_OBJID_ROOT,
 811                                       asof, localization,
 812                                       flags, &error);
 813                 if (error == 0) {
 814                         error = hammer_get_vnode(ip, &vp);
 815                         hammer_rel_inode(ip, 0);
 816                 } else {
 817                         vp = NULL;
 818                 }
 819                 if (error == 0) {
 820                         vn_unlock(vp);
 821                         cache_setvp(ap->a_nch, vp);
 822                         vrele(vp);
 823                 }
 824                 goto done;
 825         }
 826
 827         /*
 828          * If there is no path component the time extension is relative to
 829          * dip.
 830          */
 831         if (nlen == 0) {
 832                 ip = hammer_get_inode(&trans, dip, dip->obj_id,
 833                                       asof, dip->obj_localization,
 834                                       flags, &error);
 835                 if (error == 0) {
 836                         error = hammer_get_vnode(ip, &vp);
 837                         hammer_rel_inode(ip, 0);
 838                 } else {
 839                         vp = NULL;
 840                 }
 841                 if (error == 0) {
 842                         vn_unlock(vp);
 843                         cache_setvp(ap->a_nch, vp);
 844                         vrele(vp);
 845                 }
 846                 goto done;
 847         }
 848
 849         /*
 850          * Calculate the namekey and setup the key range for the scan.  This
 851          * works kinda like a chained hash table where the lower 32 bits
 852          * of the namekey synthesize the chain.
 853          *
 854          * The key range is inclusive of both key_beg and key_end.
 855          */
 856         namekey = hammer_directory_namekey(ncp->nc_name, nlen);
 857
 858         error = hammer_init_cursor(&trans, &cursor, &dip->cache[1], dip);
 859         cursor.key_beg.localization = dip->obj_localization +
 860                                       HAMMER_LOCALIZE_MISC;
 861         cursor.key_beg.obj_id = dip->obj_id;
 862         cursor.key_beg.key = namekey;
 863         cursor.key_beg.create_tid = 0;
 864         cursor.key_beg.delete_tid = 0;
 865         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
 866         cursor.key_beg.obj_type = 0;
 867
 868         cursor.key_end = cursor.key_beg;
 869         cursor.key_end.key |= 0xFFFFFFFFULL;
 870         cursor.asof = asof;
 871         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
 872
 873         /*
 874          * Scan all matching records (the chain), locate the one matching
 875          * the requested path component.
 876          *
 877          * The hammer_ip_*() functions merge in-memory records with on-disk
 878          * records for the purposes of the search.
 879          */
 880         obj_id = 0;
 881         localization = HAMMER_DEF_LOCALIZATION;
 882
 883         if (error == 0) {
 884                 error = hammer_ip_first(&cursor);
 885                 while (error == 0) {
 886                         error = hammer_ip_resolve_data(&cursor);
 887                         if (error)
 888                                 break;
 889                         if (nlen == cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF &&
 890                             bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
 891                                 obj_id = cursor.data->entry.obj_id;
 892                                 localization = cursor.data->entry.localization;
 893                                 break;
 894                         }
 895                         error = hammer_ip_next(&cursor);
 896                 }
 897         }
 898         hammer_done_cursor(&cursor);
 899         if (error == 0) {
 900                 ip = hammer_get_inode(&trans, dip, obj_id,
 901                                       asof, localization,
 902                                       flags, &error);
 903                 if (error == 0) {
 904                         error = hammer_get_vnode(ip, &vp);
 905                         hammer_rel_inode(ip, 0);
 906                 } else {
 907                         vp = NULL;
 908                 }
 909                 if (error == 0) {
 910                         vn_unlock(vp);
 911                         cache_setvp(ap->a_nch, vp);
 912                         vrele(vp);
 913                 }
 914         } else if (error == ENOENT) {
 915                 cache_setvp(ap->a_nch, NULL);
 916         }
 917 done:
 918         hammer_done_transaction(&trans);
 919         return (error);
 920 }
 921
 922 /*
 923  * hammer_vop_nlookupdotdot { dvp, vpp, cred }
 924  *
 925  * Locate the parent directory of a directory vnode.
 926  *
 927  * dvp is referenced but not locked.  *vpp must be returned referenced and
 928  * locked.  A parent_obj_id of 0 does not necessarily indicate that we are
 929  * at the root, instead it could indicate that the directory we were in was
 930  * removed.
 931  *
 932  * NOTE: as-of sequences are not linked into the directory structure.  If
 933  * we are at the root with a different asof then the mount point, reload
 934  * the same directory with the mount point's asof.   I'm not sure what this
 935  * will do to NFS.  We encode ASOF stamps in NFS file handles so it might not
 936  * get confused, but it hasn't been tested.
 937  */
 938 static
 939 int
 940 hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *ap)
 941 {
 942         struct hammer_transaction trans;
 943         struct hammer_inode *dip;
 944         struct hammer_inode *ip;
 945         int64_t parent_obj_id;
 946         u_int32_t parent_obj_localization;
 947         hammer_tid_t asof;
 948         int error;
 949
 950         dip = VTOI(ap->a_dvp);
 951         asof = dip->obj_asof;
 952
 953         /*
 954          * Whos are parent?  This could be the root of a pseudo-filesystem
 955          * whos parent is in another localization domain.
 956          */
 957         parent_obj_id = dip->ino_data.parent_obj_id;
 958         if (dip->obj_id == HAMMER_OBJID_ROOT)
 959                 parent_obj_localization = dip->ino_data.ext.obj.parent_obj_localization;
 960         else
 961                 parent_obj_localization = dip->obj_localization;
 962
 963         if (parent_obj_id == 0) {
 964                 if (dip->obj_id == HAMMER_OBJID_ROOT &&
 965                    asof != dip->hmp->asof) {
 966                         parent_obj_id = dip->obj_id;
 967                         asof = dip->hmp->asof;
 968                         *ap->a_fakename = kmalloc(19, M_TEMP, M_WAITOK);
 969                         ksnprintf(*ap->a_fakename, 19, "0x%016llx",
 970                                    dip->obj_asof);
 971                 } else {
 972                         *ap->a_vpp = NULL;
 973                         return ENOENT;
 974                 }
 975         }
 976
 977         hammer_simple_transaction(&trans, dip->hmp);
 978         ++hammer_stats_file_iopsr;
 979
 980         ip = hammer_get_inode(&trans, dip, parent_obj_id,
 981                               asof, parent_obj_localization,
 982                               dip->flags, &error);
 983         if (ip) {
 984                 error = hammer_get_vnode(ip, ap->a_vpp);
 985                 hammer_rel_inode(ip, 0);
 986         } else {
 987                 *ap->a_vpp = NULL;
 988         }
 989         hammer_done_transaction(&trans);
 990         return (error);
 991 }
 992
 993 /*
 994  * hammer_vop_nlink { nch, dvp, vp, cred }
 995  */
 996 static
 997 int
 998 hammer_vop_nlink(struct vop_nlink_args *ap)
 999 {
1000         struct hammer_transaction trans;
1001         struct hammer_inode *dip;
1002         struct hammer_inode *ip;
1003         struct nchandle *nch;
1004         int error;
1005
1006         if (ap->a_dvp->v_mount != ap->a_vp->v_mount)
1007                 return(EXDEV);
1008
1009         nch = ap->a_nch;
1010         dip = VTOI(ap->a_dvp);
1011         ip = VTOI(ap->a_vp);
1012
1013         if (dip->obj_localization != ip->obj_localization)
1014                 return(EXDEV);
1015
1016         if (dip->flags & HAMMER_INODE_RO)
1017                 return (EROFS);
1018         if (ip->flags & HAMMER_INODE_RO)
1019                 return (EROFS);
1020         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1021                 return (error);
1022
1023         /*
1024          * Create a transaction to cover the operations we perform.
1025          */
1026         hammer_start_transaction(&trans, dip->hmp);
1027         ++hammer_stats_file_iopsw;
1028
1029         /*
1030          * Add the filesystem object to the directory.  Note that neither
1031          * dip nor ip are referenced or locked, but their vnodes are
1032          * referenced.  This function will bump the inode's link count.
1033          */
1034         error = hammer_ip_add_directory(&trans, dip,
1035                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1036                                         ip);
1037
1038         /*
1039          * Finish up.
1040          */
1041         if (error == 0) {
1042                 cache_setunresolved(nch);
1043                 cache_setvp(nch, ap->a_vp);
1044         }
1045         hammer_done_transaction(&trans);
1046         return (error);
1047 }
1048
1049 /*
1050  * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
1051  *
1052  * The operating system has already ensured that the directory entry
1053  * does not exist and done all appropriate namespace locking.
1054  */
1055 static
1056 int
1057 hammer_vop_nmkdir(struct vop_nmkdir_args *ap)
1058 {
1059         struct hammer_transaction trans;
1060         struct hammer_inode *dip;
1061         struct hammer_inode *nip;
1062         struct nchandle *nch;
1063         int error;
1064
1065         nch = ap->a_nch;
1066         dip = VTOI(ap->a_dvp);
1067
1068         if (dip->flags & HAMMER_INODE_RO)
1069                 return (EROFS);
1070         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1071                 return (error);
1072
1073         /*
1074          * Create a transaction to cover the operations we perform.
1075          */
1076         hammer_start_transaction(&trans, dip->hmp);
1077         ++hammer_stats_file_iopsw;
1078
1079         /*
1080          * Create a new filesystem object of the requested type.  The
1081          * returned inode will be referenced but not locked.
1082          */
1083         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1084                                     dip, NULL, &nip);
1085         if (error) {
1086                 hkprintf("hammer_mkdir error %d\n", error);
1087                 hammer_done_transaction(&trans);
1088                 *ap->a_vpp = NULL;
1089                 return (error);
1090         }
1091         /*
1092          * Add the new filesystem object to the directory.  This will also
1093          * bump the inode's link count.
1094          */
1095         error = hammer_ip_add_directory(&trans, dip,
1096                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1097                                         nip);
1098         if (error)
1099                 hkprintf("hammer_mkdir (add) error %d\n", error);
1100
1101         /*
1102          * Finish up.
1103          */
1104         if (error) {
1105                 hammer_rel_inode(nip, 0);
1106                 *ap->a_vpp = NULL;
1107         } else {
1108                 error = hammer_get_vnode(nip, ap->a_vpp);
1109                 hammer_rel_inode(nip, 0);
1110                 if (error == 0) {
1111                         cache_setunresolved(ap->a_nch);
1112                         cache_setvp(ap->a_nch, *ap->a_vpp);
1113                 }
1114         }
1115         hammer_done_transaction(&trans);
1116         return (error);
1117 }
1118
1119 /*
1120  * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1121  *
1122  * The operating system has already ensured that the directory entry
1123  * does not exist and done all appropriate namespace locking.
1124  */
1125 static
1126 int
1127 hammer_vop_nmknod(struct vop_nmknod_args *ap)
1128 {
1129         struct hammer_transaction trans;
1130         struct hammer_inode *dip;
1131         struct hammer_inode *nip;
1132         struct nchandle *nch;
1133         int error;
1134
1135         nch = ap->a_nch;
1136         dip = VTOI(ap->a_dvp);
1137
1138         if (dip->flags & HAMMER_INODE_RO)
1139                 return (EROFS);
1140         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1141                 return (error);
1142
1143         /*
1144          * Create a transaction to cover the operations we perform.
1145          */
1146         hammer_start_transaction(&trans, dip->hmp);
1147         ++hammer_stats_file_iopsw;
1148
1149         /*
1150          * Create a new filesystem object of the requested type.  The
1151          * returned inode will be referenced but not locked.
1152          *
1153          * If mknod specifies a directory a pseudo-fs is created.
1154          */
1155         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1156                                     dip, NULL, &nip);
1157         if (error) {
1158                 hammer_done_transaction(&trans);
1159                 *ap->a_vpp = NULL;
1160                 return (error);
1161         }
1162
1163         /*
1164          * Add the new filesystem object to the directory.  This will also
1165          * bump the inode's link count.
1166          */
1167         error = hammer_ip_add_directory(&trans, dip,
1168                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1169                                         nip);
1170
1171         /*
1172          * Finish up.
1173          */
1174         if (error) {
1175                 hammer_rel_inode(nip, 0);
1176                 *ap->a_vpp = NULL;
1177         } else {
1178                 error = hammer_get_vnode(nip, ap->a_vpp);
1179                 hammer_rel_inode(nip, 0);
1180                 if (error == 0) {
1181                         cache_setunresolved(ap->a_nch);
1182                         cache_setvp(ap->a_nch, *ap->a_vpp);
1183                 }
1184         }
1185         hammer_done_transaction(&trans);
1186         return (error);
1187 }
1188
1189 /*
1190  * hammer_vop_open { vp, mode, cred, fp }
1191  */
1192 static
1193 int
1194 hammer_vop_open(struct vop_open_args *ap)
1195 {
1196         hammer_inode_t ip;
1197
1198         ++hammer_stats_file_iopsr;
1199         ip = VTOI(ap->a_vp);
1200
1201         if ((ap->a_mode & FWRITE) && (ip->flags & HAMMER_INODE_RO))
1202                 return (EROFS);
1203         return(vop_stdopen(ap));
1204 }
1205
1206 /*
1207  * hammer_vop_print { vp }
1208  */
1209 static
1210 int
1211 hammer_vop_print(struct vop_print_args *ap)
1212 {
1213         return EOPNOTSUPP;
1214 }
1215
1216 /*
1217  * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1218  */
1219 static
1220 int
1221 hammer_vop_readdir(struct vop_readdir_args *ap)
1222 {
1223         struct hammer_transaction trans;
1224         struct hammer_cursor cursor;
1225         struct hammer_inode *ip;
1226         struct uio *uio;
1227         hammer_base_elm_t base;
1228         int error;
1229         int cookie_index;
1230         int ncookies;
1231         off_t *cookies;
1232         off_t saveoff;
1233         int r;
1234         int dtype;
1235
1236         ++hammer_stats_file_iopsr;
1237         ip = VTOI(ap->a_vp);
1238         uio = ap->a_uio;
1239         saveoff = uio->uio_offset;
1240
1241         if (ap->a_ncookies) {
1242                 ncookies = uio->uio_resid / 16 + 1;
1243                 if (ncookies > 1024)
1244                         ncookies = 1024;
1245                 cookies = kmalloc(ncookies * sizeof(off_t), M_TEMP, M_WAITOK);
1246                 cookie_index = 0;
1247         } else {
1248                 ncookies = -1;
1249                 cookies = NULL;
1250                 cookie_index = 0;
1251         }
1252
1253         hammer_simple_transaction(&trans, ip->hmp);
1254
1255         /*
1256          * Handle artificial entries
1257          */
1258         error = 0;
1259         if (saveoff == 0) {
1260                 r = vop_write_dirent(&error, uio, ip->obj_id, DT_DIR, 1, ".");
1261                 if (r)
1262                         goto done;
1263                 if (cookies)
1264                         cookies[cookie_index] = saveoff;
1265                 ++saveoff;
1266                 ++cookie_index;
1267                 if (cookie_index == ncookies)
1268                         goto done;
1269         }
1270         if (saveoff == 1) {
1271                 if (ip->ino_data.parent_obj_id) {
1272                         r = vop_write_dirent(&error, uio,
1273                                              ip->ino_data.parent_obj_id,
1274                                              DT_DIR, 2, "..");
1275                 } else {
1276                         r = vop_write_dirent(&error, uio,
1277                                              ip->obj_id, DT_DIR, 2, "..");
1278                 }
1279                 if (r)
1280                         goto done;
1281                 if (cookies)
1282                         cookies[cookie_index] = saveoff;
1283                 ++saveoff;
1284                 ++cookie_index;
1285                 if (cookie_index == ncookies)
1286                         goto done;
1287         }
1288
1289         /*
1290          * Key range (begin and end inclusive) to scan.  Directory keys
1291          * directly translate to a 64 bit 'seek' position.
1292          */
1293         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1294         cursor.key_beg.localization = ip->obj_localization +
1295                                       HAMMER_LOCALIZE_MISC;
1296         cursor.key_beg.obj_id = ip->obj_id;
1297         cursor.key_beg.create_tid = 0;
1298         cursor.key_beg.delete_tid = 0;
1299         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1300         cursor.key_beg.obj_type = 0;
1301         cursor.key_beg.key = saveoff;
1302
1303         cursor.key_end = cursor.key_beg;
1304         cursor.key_end.key = HAMMER_MAX_KEY;
1305         cursor.asof = ip->obj_asof;
1306         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1307
1308         error = hammer_ip_first(&cursor);
1309
1310         while (error == 0) {
1311                 error = hammer_ip_resolve_data(&cursor);
1312                 if (error)
1313                         break;
1314                 base = &cursor.leaf->base;
1315                 saveoff = base->key;
1316                 KKASSERT(cursor.leaf->data_len > HAMMER_ENTRY_NAME_OFF);
1317
1318                 if (base->obj_id != ip->obj_id)
1319                         panic("readdir: bad record at %p", cursor.node);
1320
1321                 /*
1322                  * Convert pseudo-filesystems into softlinks
1323                  */
1324                 dtype = hammer_get_dtype(cursor.leaf->base.obj_type);
1325                 r = vop_write_dirent(
1326                              &error, uio, cursor.data->entry.obj_id,
1327                              dtype,
1328                              cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF ,
1329                              (void *)cursor.data->entry.name);
1330                 if (r)
1331                         break;
1332                 ++saveoff;
1333                 if (cookies)
1334                         cookies[cookie_index] = base->key;
1335                 ++cookie_index;
1336                 if (cookie_index == ncookies)
1337                         break;
1338                 error = hammer_ip_next(&cursor);
1339         }
1340         hammer_done_cursor(&cursor);
1341
1342 done:
1343         hammer_done_transaction(&trans);
1344
1345         if (ap->a_eofflag)
1346                 *ap->a_eofflag = (error == ENOENT);
1347         uio->uio_offset = saveoff;
1348         if (error && cookie_index == 0) {
1349                 if (error == ENOENT)
1350                         error = 0;
1351                 if (cookies) {
1352                         kfree(cookies, M_TEMP);
1353                         *ap->a_ncookies = 0;
1354                         *ap->a_cookies = NULL;
1355                 }
1356         } else {
1357                 if (error == ENOENT)
1358                         error = 0;
1359                 if (cookies) {
1360                         *ap->a_ncookies = cookie_index;
1361                         *ap->a_cookies = cookies;
1362                 }
1363         }
1364         return(error);
1365 }
1366
1367 /*
1368  * hammer_vop_readlink { vp, uio, cred }
1369  */
1370 static
1371 int
1372 hammer_vop_readlink(struct vop_readlink_args *ap)
1373 {
1374         struct hammer_transaction trans;
1375         struct hammer_cursor cursor;
1376         struct hammer_inode *ip;
1377         char buf[32];
1378         u_int32_t localization;
1379         hammer_pseudofs_inmem_t pfsm;
1380         int error;
1381
1382         ip = VTOI(ap->a_vp);
1383
1384         /*
1385          * Shortcut if the symlink data was stuffed into ino_data.
1386          *
1387          * Also expand special "@@PFS%05d" softlinks (expansion only
1388          * occurs for non-historical (current) accesses made from the
1389          * primary filesystem).
1390          */
1391         if (ip->ino_data.size <= HAMMER_INODE_BASESYMLEN) {
1392                 char *ptr;
1393                 int bytes;
1394
1395                 ptr = ip->ino_data.ext.symlink;
1396                 bytes = (int)ip->ino_data.size;
1397                 if (bytes == 10 &&
1398                     ip->obj_asof == HAMMER_MAX_TID &&
1399                     ip->obj_localization == 0 &&
1400                     strncmp(ptr, "@@PFS", 5) == 0) {
1401                         hammer_simple_transaction(&trans, ip->hmp);
1402                         bcopy(ptr + 5, buf, 5);
1403                         buf[5] = 0;
1404                         localization = strtoul(buf, NULL, 10) << 16;
1405                         pfsm = hammer_load_pseudofs(&trans, localization,
1406                                                     &error);
1407                         if (error == 0) {
1408                                 if (pfsm->pfsd.mirror_flags &
1409                                     HAMMER_PFSD_SLAVE) {
1410                                         ksnprintf(buf, sizeof(buf),
1411                                                   "@@0x%016llx:%05d",
1412                                                   pfsm->pfsd.sync_end_tid,
1413                                                   localization >> 16);
1414                                 } else {
1415                                         ksnprintf(buf, sizeof(buf),
1416                                                   "@@0x%016llx:%05d",
1417                                                   HAMMER_MAX_TID,
1418                                                   localization >> 16);
1419                                 }
1420                                 ptr = buf;
1421                                 bytes = strlen(buf);
1422                         }
1423                         if (pfsm)
1424                                 hammer_rel_pseudofs(trans.hmp, pfsm);
1425                         hammer_done_transaction(&trans);
1426                 }
1427                 error = uiomove(ptr, bytes, ap->a_uio);
1428                 return(error);
1429         }
1430
1431         /*
1432          * Long version
1433          */
1434         hammer_simple_transaction(&trans, ip->hmp);
1435         ++hammer_stats_file_iopsr;
1436         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1437
1438         /*
1439          * Key range (begin and end inclusive) to scan.  Directory keys
1440          * directly translate to a 64 bit 'seek' position.
1441          */
1442         cursor.key_beg.localization = ip->obj_localization +
1443                                       HAMMER_LOCALIZE_MISC;
1444         cursor.key_beg.obj_id = ip->obj_id;
1445         cursor.key_beg.create_tid = 0;
1446         cursor.key_beg.delete_tid = 0;
1447         cursor.key_beg.rec_type = HAMMER_RECTYPE_FIX;
1448         cursor.key_beg.obj_type = 0;
1449         cursor.key_beg.key = HAMMER_FIXKEY_SYMLINK;
1450         cursor.asof = ip->obj_asof;
1451         cursor.flags |= HAMMER_CURSOR_ASOF;
1452
1453         error = hammer_ip_lookup(&cursor);
1454         if (error == 0) {
1455                 error = hammer_ip_resolve_data(&cursor);
1456                 if (error == 0) {
1457                         KKASSERT(cursor.leaf->data_len >=
1458                                  HAMMER_SYMLINK_NAME_OFF);
1459                         error = uiomove(cursor.data->symlink.name,
1460                                         cursor.leaf->data_len -
1461                                                 HAMMER_SYMLINK_NAME_OFF,
1462                                         ap->a_uio);
1463                 }
1464         }
1465         hammer_done_cursor(&cursor);
1466         hammer_done_transaction(&trans);
1467         return(error);
1468 }
1469
1470 /*
1471  * hammer_vop_nremove { nch, dvp, cred }
1472  */
1473 static
1474 int
1475 hammer_vop_nremove(struct vop_nremove_args *ap)
1476 {
1477         struct hammer_transaction trans;
1478         struct hammer_inode *dip;
1479         int error;
1480
1481         dip = VTOI(ap->a_dvp);
1482
1483         if (hammer_nohistory(dip) == 0 &&
1484             (error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1485                 return (error);
1486         }
1487
1488         hammer_start_transaction(&trans, dip->hmp);
1489         ++hammer_stats_file_iopsw;
1490         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 0);
1491         hammer_done_transaction(&trans);
1492
1493         return (error);
1494 }
1495
1496 /*
1497  * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1498  */
1499 static
1500 int
1501 hammer_vop_nrename(struct vop_nrename_args *ap)
1502 {
1503         struct hammer_transaction trans;
1504         struct namecache *fncp;
1505         struct namecache *tncp;
1506         struct hammer_inode *fdip;
1507         struct hammer_inode *tdip;
1508         struct hammer_inode *ip;
1509         struct hammer_cursor cursor;
1510         int64_t namekey;
1511         int nlen, error;
1512
1513         if (ap->a_fdvp->v_mount != ap->a_tdvp->v_mount)
1514                 return(EXDEV);
1515         if (ap->a_fdvp->v_mount != ap->a_fnch->ncp->nc_vp->v_mount)
1516                 return(EXDEV);
1517
1518         fdip = VTOI(ap->a_fdvp);
1519         tdip = VTOI(ap->a_tdvp);
1520         fncp = ap->a_fnch->ncp;
1521         tncp = ap->a_tnch->ncp;
1522         ip = VTOI(fncp->nc_vp);
1523         KKASSERT(ip != NULL);
1524
1525         if (fdip->obj_localization != tdip->obj_localization)
1526                 return(EXDEV);
1527         if (fdip->obj_localization != ip->obj_localization)
1528                 return(EXDEV);
1529
1530         if (fdip->flags & HAMMER_INODE_RO)
1531                 return (EROFS);
1532         if (tdip->flags & HAMMER_INODE_RO)
1533                 return (EROFS);
1534         if (ip->flags & HAMMER_INODE_RO)
1535                 return (EROFS);
1536         if ((error = hammer_checkspace(fdip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1537                 return (error);
1538
1539         hammer_start_transaction(&trans, fdip->hmp);
1540         ++hammer_stats_file_iopsw;
1541
1542         /*
1543          * Remove tncp from the target directory and then link ip as
1544          * tncp. XXX pass trans to dounlink
1545          *
1546          * Force the inode sync-time to match the transaction so it is
1547          * in-sync with the creation of the target directory entry.
1548          */
1549         error = hammer_dounlink(&trans, ap->a_tnch, ap->a_tdvp,
1550                                 ap->a_cred, 0, -1);
1551         if (error == 0 || error == ENOENT) {
1552                 error = hammer_ip_add_directory(&trans, tdip,
1553                                                 tncp->nc_name, tncp->nc_nlen,
1554                                                 ip);
1555                 if (error == 0) {
1556                         ip->ino_data.parent_obj_id = tdip->obj_id;
1557                         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
1558                 }
1559         }
1560         if (error)
1561                 goto failed; /* XXX */
1562
1563         /*
1564          * Locate the record in the originating directory and remove it.
1565          *
1566          * Calculate the namekey and setup the key range for the scan.  This
1567          * works kinda like a chained hash table where the lower 32 bits
1568          * of the namekey synthesize the chain.
1569          *
1570          * The key range is inclusive of both key_beg and key_end.
1571          */
1572         namekey = hammer_directory_namekey(fncp->nc_name, fncp->nc_nlen);
1573 retry:
1574         hammer_init_cursor(&trans, &cursor, &fdip->cache[1], fdip);
1575         cursor.key_beg.localization = fdip->obj_localization +
1576                                       HAMMER_LOCALIZE_MISC;
1577         cursor.key_beg.obj_id = fdip->obj_id;
1578         cursor.key_beg.key = namekey;
1579         cursor.key_beg.create_tid = 0;
1580         cursor.key_beg.delete_tid = 0;
1581         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1582         cursor.key_beg.obj_type = 0;
1583
1584         cursor.key_end = cursor.key_beg;
1585         cursor.key_end.key |= 0xFFFFFFFFULL;
1586         cursor.asof = fdip->obj_asof;
1587         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1588
1589         /*
1590          * Scan all matching records (the chain), locate the one matching
1591          * the requested path component.
1592          *
1593          * The hammer_ip_*() functions merge in-memory records with on-disk
1594          * records for the purposes of the search.
1595          */
1596         error = hammer_ip_first(&cursor);
1597         while (error == 0) {
1598                 if (hammer_ip_resolve_data(&cursor) != 0)
1599                         break;
1600                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
1601                 KKASSERT(nlen > 0);
1602                 if (fncp->nc_nlen == nlen &&
1603                     bcmp(fncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1604                         break;
1605                 }
1606                 error = hammer_ip_next(&cursor);
1607         }
1608
1609         /*
1610          * If all is ok we have to get the inode so we can adjust nlinks.
1611          *
1612          * WARNING: hammer_ip_del_directory() may have to terminate the
1613          * cursor to avoid a recursion.  It's ok to call hammer_done_cursor()
1614          * twice.
1615          */
1616         if (error == 0)
1617                 error = hammer_ip_del_directory(&trans, &cursor, fdip, ip);
1618
1619         /*
1620          * XXX A deadlock here will break rename's atomicy for the purposes
1621          * of crash recovery.
1622          */
1623         if (error == EDEADLK) {
1624                 hammer_done_cursor(&cursor);
1625                 goto retry;
1626         }
1627
1628         /*
1629          * Cleanup and tell the kernel that the rename succeeded.
1630          */
1631         hammer_done_cursor(&cursor);
1632         if (error == 0)
1633                 cache_rename(ap->a_fnch, ap->a_tnch);
1634
1635 failed:
1636         hammer_done_transaction(&trans);
1637         return (error);
1638 }
1639
1640 /*
1641  * hammer_vop_nrmdir { nch, dvp, cred }
1642  */
1643 static
1644 int
1645 hammer_vop_nrmdir(struct vop_nrmdir_args *ap)
1646 {
1647         struct hammer_transaction trans;
1648         struct hammer_inode *dip;
1649         int error;
1650
1651         dip = VTOI(ap->a_dvp);
1652
1653         if (hammer_nohistory(dip) == 0 &&
1654             (error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1655                 return (error);
1656         }
1657
1658         hammer_start_transaction(&trans, dip->hmp);
1659         ++hammer_stats_file_iopsw;
1660         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 1);
1661         hammer_done_transaction(&trans);
1662
1663         return (error);
1664 }
1665
1666 /*
1667  * hammer_vop_setattr { vp, vap, cred }
1668  */
1669 static
1670 int
1671 hammer_vop_setattr(struct vop_setattr_args *ap)
1672 {
1673         struct hammer_transaction trans;
1674         struct vattr *vap;
1675         struct hammer_inode *ip;
1676         int modflags;
1677         int error;
1678         int truncating;
1679         int blksize;
1680         int64_t aligned_size;
1681         u_int32_t flags;
1682
1683         vap = ap->a_vap;
1684         ip = ap->a_vp->v_data;
1685         modflags = 0;
1686
1687         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
1688                 return(EROFS);
1689         if (ip->flags & HAMMER_INODE_RO)
1690                 return (EROFS);
1691         if (hammer_nohistory(ip) == 0 &&
1692             (error = hammer_checkspace(ip->hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1693                 return (error);
1694         }
1695
1696         hammer_start_transaction(&trans, ip->hmp);
1697         ++hammer_stats_file_iopsw;
1698         error = 0;
1699
1700         if (vap->va_flags != VNOVAL) {
1701                 flags = ip->ino_data.uflags;
1702                 error = vop_helper_setattr_flags(&flags, vap->va_flags,
1703                                          hammer_to_unix_xid(&ip->ino_data.uid),
1704                                          ap->a_cred);
1705                 if (error == 0) {
1706                         if (ip->ino_data.uflags != flags) {
1707                                 ip->ino_data.uflags = flags;
1708                                 modflags |= HAMMER_INODE_DDIRTY;
1709                         }
1710                         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
1711                                 error = 0;
1712                                 goto done;
1713                         }
1714                 }
1715                 goto done;
1716         }
1717         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
1718                 error = EPERM;
1719                 goto done;
1720         }
1721         if (vap->va_uid != (uid_t)VNOVAL || vap->va_gid != (gid_t)VNOVAL) {
1722                 mode_t cur_mode = ip->ino_data.mode;
1723                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1724                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1725                 uuid_t uuid_uid;
1726                 uuid_t uuid_gid;
1727
1728                 error = vop_helper_chown(ap->a_vp, vap->va_uid, vap->va_gid,
1729                                          ap->a_cred,
1730                                          &cur_uid, &cur_gid, &cur_mode);
1731                 if (error == 0) {
1732                         hammer_guid_to_uuid(&uuid_uid, cur_uid);
1733                         hammer_guid_to_uuid(&uuid_gid, cur_gid);
1734                         if (bcmp(&uuid_uid, &ip->ino_data.uid,
1735                                  sizeof(uuid_uid)) ||
1736                             bcmp(&uuid_gid, &ip->ino_data.gid,
1737                                  sizeof(uuid_gid)) ||
1738                             ip->ino_data.mode != cur_mode
1739                         ) {
1740                                 ip->ino_data.uid = uuid_uid;
1741                                 ip->ino_data.gid = uuid_gid;
1742                                 ip->ino_data.mode = cur_mode;
1743                         }
1744                         modflags |= HAMMER_INODE_DDIRTY;
1745                 }
1746         }
1747         while (vap->va_size != VNOVAL && ip->ino_data.size != vap->va_size) {
1748                 switch(ap->a_vp->v_type) {
1749                 case VREG:
1750                         if (vap->va_size == ip->ino_data.size)
1751                                 break;
1752                         /*
1753                          * XXX break atomicy, we can deadlock the backend
1754                          * if we do not release the lock.  Probably not a
1755                          * big deal here.
1756                          */
1757                         blksize = hammer_blocksize(vap->va_size);
1758                         if (vap->va_size < ip->ino_data.size) {
1759                                 vtruncbuf(ap->a_vp, vap->va_size, blksize);
1760                                 truncating = 1;
1761                         } else {
1762                                 vnode_pager_setsize(ap->a_vp, vap->va_size);
1763                                 truncating = 0;
1764                         }
1765                         ip->ino_data.size = vap->va_size;
1766                         modflags |= HAMMER_INODE_DDIRTY;
1767
1768                         /*
1769                          * on-media truncation is cached in the inode until
1770                          * the inode is synchronized.
1771                          */
1772                         if (truncating) {
1773                                 hammer_ip_frontend_trunc(ip, vap->va_size);
1774 #ifdef DEBUG_TRUNCATE
1775                                 if (HammerTruncIp == NULL)
1776                                         HammerTruncIp = ip;
1777 #endif
1778                                 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
1779                                         ip->flags |= HAMMER_INODE_TRUNCATED;
1780                                         ip->trunc_off = vap->va_size;
1781 #ifdef DEBUG_TRUNCATE
1782                                         if (ip == HammerTruncIp)
1783                                         kprintf("truncate1 %016llx\n", ip->trunc_off);
1784 #endif
1785                                 } else if (ip->trunc_off > vap->va_size) {
1786                                         ip->trunc_off = vap->va_size;
1787 #ifdef DEBUG_TRUNCATE
1788                                         if (ip == HammerTruncIp)
1789                                         kprintf("truncate2 %016llx\n", ip->trunc_off);
1790 #endif
1791                                 } else {
1792 #ifdef DEBUG_TRUNCATE
1793                                         if (ip == HammerTruncIp)
1794                                         kprintf("truncate3 %016llx (ignored)\n", vap->va_size);
1795 #endif
1796                                 }
1797                         }
1798
1799                         /*
1800                          * If truncating we have to clean out a portion of
1801                          * the last block on-disk.  We do this in the
1802                          * front-end buffer cache.
1803                          */
1804                         aligned_size = (vap->va_size + (blksize - 1)) &
1805                                        ~(int64_t)(blksize - 1);
1806                         if (truncating && vap->va_size < aligned_size) {
1807                                 struct buf *bp;
1808                                 int offset;
1809
1810                                 aligned_size -= blksize;
1811
1812                                 offset = (int)vap->va_size & (blksize - 1);
1813                                 error = bread(ap->a_vp, aligned_size,
1814                                               blksize, &bp);
1815                                 hammer_ip_frontend_trunc(ip, aligned_size);
1816                                 if (error == 0) {
1817                                         bzero(bp->b_data + offset,
1818                                               blksize - offset);
1819                                         /* must de-cache direct-io offset */
1820                                         bp->b_bio2.bio_offset = NOOFFSET;
1821                                         bdwrite(bp);
1822                                 } else {
1823                                         kprintf("ERROR %d\n", error);
1824                                         brelse(bp);
1825                                 }
1826                         }
1827                         break;
1828                 case VDATABASE:
1829                         if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
1830                                 ip->flags |= HAMMER_INODE_TRUNCATED;
1831                                 ip->trunc_off = vap->va_size;
1832                         } else if (ip->trunc_off > vap->va_size) {
1833                                 ip->trunc_off = vap->va_size;
1834                         }
1835                         hammer_ip_frontend_trunc(ip, vap->va_size);
1836                         ip->ino_data.size = vap->va_size;
1837                         modflags |= HAMMER_INODE_DDIRTY;
1838                         break;
1839                 default:
1840                         error = EINVAL;
1841                         goto done;
1842                 }
1843                 break;
1844         }
1845         if (vap->va_atime.tv_sec != VNOVAL) {
1846                 ip->ino_data.atime =
1847                         hammer_timespec_to_time(&vap->va_atime);
1848                 modflags |= HAMMER_INODE_ATIME;
1849         }
1850         if (vap->va_mtime.tv_sec != VNOVAL) {
1851                 ip->ino_data.mtime =
1852                         hammer_timespec_to_time(&vap->va_mtime);
1853                 modflags |= HAMMER_INODE_MTIME;
1854         }
1855         if (vap->va_mode != (mode_t)VNOVAL) {
1856                 mode_t   cur_mode = ip->ino_data.mode;
1857                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1858                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1859
1860                 error = vop_helper_chmod(ap->a_vp, vap->va_mode, ap->a_cred,
1861                                          cur_uid, cur_gid, &cur_mode);
1862                 if (error == 0 && ip->ino_data.mode != cur_mode) {
1863                         ip->ino_data.mode = cur_mode;
1864                         modflags |= HAMMER_INODE_DDIRTY;
1865                 }
1866         }
1867 done:
1868         if (error == 0)
1869                 hammer_modify_inode(ip, modflags);
1870         hammer_done_transaction(&trans);
1871         return (error);
1872 }
1873
1874 /*
1875  * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
1876  */
1877 static
1878 int
1879 hammer_vop_nsymlink(struct vop_nsymlink_args *ap)
1880 {
1881         struct hammer_transaction trans;
1882         struct hammer_inode *dip;
1883         struct hammer_inode *nip;
1884         struct nchandle *nch;
1885         hammer_record_t record;
1886         int error;
1887         int bytes;
1888
1889         ap->a_vap->va_type = VLNK;
1890
1891         nch = ap->a_nch;
1892         dip = VTOI(ap->a_dvp);
1893
1894         if (dip->flags & HAMMER_INODE_RO)
1895                 return (EROFS);
1896         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1897                 return (error);
1898
1899         /*
1900          * Create a transaction to cover the operations we perform.
1901          */
1902         hammer_start_transaction(&trans, dip->hmp);
1903         ++hammer_stats_file_iopsw;
1904
1905         /*
1906          * Create a new filesystem object of the requested type.  The
1907          * returned inode will be referenced but not locked.
1908          */
1909
1910         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1911                                     dip, NULL, &nip);
1912         if (error) {
1913                 hammer_done_transaction(&trans);
1914                 *ap->a_vpp = NULL;
1915                 return (error);
1916         }
1917
1918         /*
1919          * Add a record representing the symlink.  symlink stores the link
1920          * as pure data, not a string, and is no \0 terminated.
1921          */
1922         if (error == 0) {
1923                 bytes = strlen(ap->a_target);
1924
1925                 if (bytes <= HAMMER_INODE_BASESYMLEN) {
1926                         bcopy(ap->a_target, nip->ino_data.ext.symlink, bytes);
1927                 } else {
1928                         record = hammer_alloc_mem_record(nip, bytes);
1929                         record->type = HAMMER_MEM_RECORD_GENERAL;
1930
1931                         record->leaf.base.localization = nip->obj_localization +
1932                                                          HAMMER_LOCALIZE_MISC;
1933                         record->leaf.base.key = HAMMER_FIXKEY_SYMLINK;
1934                         record->leaf.base.rec_type = HAMMER_RECTYPE_FIX;
1935                         record->leaf.data_len = bytes;
1936                         KKASSERT(HAMMER_SYMLINK_NAME_OFF == 0);
1937                         bcopy(ap->a_target, record->data->symlink.name, bytes);
1938                         error = hammer_ip_add_record(&trans, record);
1939                 }
1940
1941                 /*
1942                  * Set the file size to the length of the link.
1943                  */
1944                 if (error == 0) {
1945                         nip->ino_data.size = bytes;
1946                         hammer_modify_inode(nip, HAMMER_INODE_DDIRTY);
1947                 }
1948         }
1949         if (error == 0)
1950                 error = hammer_ip_add_directory(&trans, dip, nch->ncp->nc_name,
1951                                                 nch->ncp->nc_nlen, nip);
1952
1953         /*
1954          * Finish up.
1955          */
1956         if (error) {
1957                 hammer_rel_inode(nip, 0);
1958                 *ap->a_vpp = NULL;
1959         } else {
1960                 error = hammer_get_vnode(nip, ap->a_vpp);
1961                 hammer_rel_inode(nip, 0);
1962                 if (error == 0) {
1963                         cache_setunresolved(ap->a_nch);
1964                         cache_setvp(ap->a_nch, *ap->a_vpp);
1965                 }
1966         }
1967         hammer_done_transaction(&trans);
1968         return (error);
1969 }
1970
1971 /*
1972  * hammer_vop_nwhiteout { nch, dvp, cred, flags }
1973  */
1974 static
1975 int
1976 hammer_vop_nwhiteout(struct vop_nwhiteout_args *ap)
1977 {
1978         struct hammer_transaction trans;
1979         struct hammer_inode *dip;
1980         int error;
1981
1982         dip = VTOI(ap->a_dvp);
1983
1984         if (hammer_nohistory(dip) == 0 &&
1985             (error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0) {
1986                 return (error);
1987         }
1988
1989         hammer_start_transaction(&trans, dip->hmp);
1990         ++hammer_stats_file_iopsw;
1991         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp,
1992                                 ap->a_cred, ap->a_flags, -1);
1993         hammer_done_transaction(&trans);
1994
1995         return (error);
1996 }
1997
1998 /*
1999  * hammer_vop_ioctl { vp, command, data, fflag, cred }
2000  */
2001 static
2002 int
2003 hammer_vop_ioctl(struct vop_ioctl_args *ap)
2004 {
2005         struct hammer_inode *ip = ap->a_vp->v_data;
2006
2007         ++hammer_stats_file_iopsr;
2008         return(hammer_ioctl(ip, ap->a_command, ap->a_data,
2009                             ap->a_fflag, ap->a_cred));
2010 }
2011
2012 static
2013 int
2014 hammer_vop_mountctl(struct vop_mountctl_args *ap)
2015 {
2016         struct mount *mp;
2017         int error;
2018
2019         mp = ap->a_head.a_ops->head.vv_mount;
2020
2021         switch(ap->a_op) {
2022         case MOUNTCTL_SET_EXPORT:
2023                 if (ap->a_ctllen != sizeof(struct export_args))
2024                         error = EINVAL;
2025                 else
2026                         error = hammer_vfs_export(mp, ap->a_op,
2027                                       (const struct export_args *)ap->a_ctl);
2028                 break;
2029         default:
2030                 error = journal_mountctl(ap);
2031                 break;
2032         }
2033         return(error);
2034 }
2035
2036 /*
2037  * hammer_vop_strategy { vp, bio }
2038  *
2039  * Strategy call, used for regular file read & write only.  Note that the
2040  * bp may represent a cluster.
2041  *
2042  * To simplify operation and allow better optimizations in the future,
2043  * this code does not make any assumptions with regards to buffer alignment
2044  * or size.
2045  */
2046 static
2047 int
2048 hammer_vop_strategy(struct vop_strategy_args *ap)
2049 {
2050         struct buf *bp;
2051         int error;
2052
2053         bp = ap->a_bio->bio_buf;
2054
2055         switch(bp->b_cmd) {
2056         case BUF_CMD_READ:
2057                 error = hammer_vop_strategy_read(ap);
2058                 break;
2059         case BUF_CMD_WRITE:
2060                 error = hammer_vop_strategy_write(ap);
2061                 break;
2062         default:
2063                 bp->b_error = error = EINVAL;
2064                 bp->b_flags |= B_ERROR;
2065                 biodone(ap->a_bio);
2066                 break;
2067         }
2068         return (error);
2069 }
2070
2071 /*
2072  * Read from a regular file.  Iterate the related records and fill in the
2073  * BIO/BUF.  Gaps are zero-filled.
2074  *
2075  * The support code in hammer_object.c should be used to deal with mixed
2076  * in-memory and on-disk records.
2077  *
2078  * NOTE: Can be called from the cluster code with an oversized buf.
2079  *
2080  * XXX atime update
2081  */
2082 static
2083 int
2084 hammer_vop_strategy_read(struct vop_strategy_args *ap)
2085 {
2086         struct hammer_transaction trans;
2087         struct hammer_inode *ip;
2088         struct hammer_cursor cursor;
2089         hammer_base_elm_t base;
2090         hammer_off_t disk_offset;
2091         struct bio *bio;
2092         struct bio *nbio;
2093         struct buf *bp;
2094         int64_t rec_offset;
2095         int64_t ran_end;
2096         int64_t tmp64;
2097         int error;
2098         int boff;
2099         int roff;
2100         int n;
2101
2102         bio = ap->a_bio;
2103         bp = bio->bio_buf;
2104         ip = ap->a_vp->v_data;
2105
2106         /*
2107          * The zone-2 disk offset may have been set by the cluster code via
2108          * a BMAP operation, or else should be NOOFFSET.
2109          *
2110          * Checking the high bits for a match against zone-2 should suffice.
2111          */
2112         nbio = push_bio(bio);
2113         if ((nbio->bio_offset & HAMMER_OFF_ZONE_MASK) ==
2114             HAMMER_ZONE_LARGE_DATA) {
2115                 error = hammer_io_direct_read(ip->hmp, nbio, NULL);
2116                 return (error);
2117         }
2118
2119         /*
2120          * Well, that sucked.  Do it the hard way.  If all the stars are
2121          * aligned we may still be able to issue a direct-read.
2122          */
2123         hammer_simple_transaction(&trans, ip->hmp);
2124         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2125
2126         /*
2127          * Key range (begin and end inclusive) to scan.  Note that the key's
2128          * stored in the actual records represent BASE+LEN, not BASE.  The
2129          * first record containing bio_offset will have a key > bio_offset.
2130          */
2131         cursor.key_beg.localization = ip->obj_localization +
2132                                       HAMMER_LOCALIZE_MISC;
2133         cursor.key_beg.obj_id = ip->obj_id;
2134         cursor.key_beg.create_tid = 0;
2135         cursor.key_beg.delete_tid = 0;
2136         cursor.key_beg.obj_type = 0;
2137         cursor.key_beg.key = bio->bio_offset + 1;
2138         cursor.asof = ip->obj_asof;
2139         cursor.flags |= HAMMER_CURSOR_ASOF;
2140
2141         cursor.key_end = cursor.key_beg;
2142         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2143 #if 0
2144         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
2145                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
2146                 cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
2147                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2148         } else
2149 #endif
2150         {
2151                 ran_end = bio->bio_offset + bp->b_bufsize;
2152                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2153                 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2154                 tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2155                 if (tmp64 < ran_end)
2156                         cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2157                 else
2158                         cursor.key_end.key = ran_end + MAXPHYS + 1;
2159         }
2160         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2161
2162         error = hammer_ip_first(&cursor);
2163         boff = 0;
2164
2165         while (error == 0) {
2166                 /*
2167                  * Get the base file offset of the record.  The key for
2168                  * data records is (base + bytes) rather then (base).
2169                  */
2170                 base = &cursor.leaf->base;
2171                 rec_offset = base->key - cursor.leaf->data_len;
2172
2173                 /*
2174                  * Calculate the gap, if any, and zero-fill it.
2175                  *
2176                  * n is the offset of the start of the record verses our
2177                  * current seek offset in the bio.
2178                  */
2179                 n = (int)(rec_offset - (bio->bio_offset + boff));
2180                 if (n > 0) {
2181                         if (n > bp->b_bufsize - boff)
2182                                 n = bp->b_bufsize - boff;
2183                         bzero((char *)bp->b_data + boff, n);
2184                         boff += n;
2185                         n = 0;
2186                 }
2187
2188                 /*
2189                  * Calculate the data offset in the record and the number
2190                  * of bytes we can copy.
2191                  *
2192                  * There are two degenerate cases.  First, boff may already
2193                  * be at bp->b_bufsize.  Secondly, the data offset within
2194                  * the record may exceed the record's size.
2195                  */
2196                 roff = -n;
2197                 rec_offset += roff;
2198                 n = cursor.leaf->data_len - roff;
2199                 if (n <= 0) {
2200                         kprintf("strategy_read: bad n=%d roff=%d\n", n, roff);
2201                         n = 0;
2202                 } else if (n > bp->b_bufsize - boff) {
2203                         n = bp->b_bufsize - boff;
2204                 }
2205
2206                 /*
2207                  * Deal with cached truncations.  This cool bit of code
2208                  * allows truncate()/ftruncate() to avoid having to sync
2209                  * the file.
2210                  *
2211                  * If the frontend is truncated then all backend records are
2212                  * subject to the frontend's truncation.
2213                  *
2214                  * If the backend is truncated then backend records on-disk
2215                  * (but not in-memory) are subject to the backend's
2216                  * truncation.  In-memory records owned by the backend
2217                  * represent data written after the truncation point on the
2218                  * backend and must not be truncated.
2219                  *
2220                  * Truncate operations deal with frontend buffer cache
2221                  * buffers and frontend-owned in-memory records synchronously.
2222                  */
2223                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2224                         if (hammer_cursor_ondisk(&cursor) ||
2225                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
2226                                 if (ip->trunc_off <= rec_offset)
2227                                         n = 0;
2228                                 else if (ip->trunc_off < rec_offset + n)
2229                                         n = (int)(ip->trunc_off - rec_offset);
2230                         }
2231                 }
2232                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2233                         if (hammer_cursor_ondisk(&cursor)) {
2234                                 if (ip->sync_trunc_off <= rec_offset)
2235                                         n = 0;
2236                                 else if (ip->sync_trunc_off < rec_offset + n)
2237                                         n = (int)(ip->sync_trunc_off - rec_offset);
2238                         }
2239                 }
2240
2241                 /*
2242                  * Try to issue a direct read into our bio if possible,
2243                  * otherwise resolve the element data into a hammer_buffer
2244                  * and copy.
2245                  *
2246                  * The buffer on-disk should be zerod past any real
2247                  * truncation point, but may not be for any synthesized
2248                  * truncation point from above.
2249                  */
2250                 disk_offset = cursor.leaf->data_offset + roff;
2251                 if (boff == 0 && n == bp->b_bufsize &&
2252                     hammer_cursor_ondisk(&cursor) &&
2253                     (disk_offset & HAMMER_BUFMASK) == 0) {
2254                         KKASSERT((disk_offset & HAMMER_OFF_ZONE_MASK) ==
2255                                  HAMMER_ZONE_LARGE_DATA);
2256                         nbio->bio_offset = disk_offset;
2257                         error = hammer_io_direct_read(trans.hmp, nbio,
2258                                                       cursor.leaf);
2259                         goto done;
2260                 } else if (n) {
2261                         error = hammer_ip_resolve_data(&cursor);
2262                         if (error == 0) {
2263                                 bcopy((char *)cursor.data + roff,
2264                                       (char *)bp->b_data + boff, n);
2265                         }
2266                 }
2267                 if (error)
2268                         break;
2269
2270                 /*
2271                  * Iterate until we have filled the request.
2272                  */
2273                 boff += n;
2274                 if (boff == bp->b_bufsize)
2275                         break;
2276                 error = hammer_ip_next(&cursor);
2277         }
2278
2279         /*
2280          * There may have been a gap after the last record
2281          */
2282         if (error == ENOENT)
2283                 error = 0;
2284         if (error == 0 && boff != bp->b_bufsize) {
2285                 KKASSERT(boff < bp->b_bufsize);
2286                 bzero((char *)bp->b_data + boff, bp->b_bufsize - boff);
2287                 /* boff = bp->b_bufsize; */
2288         }
2289         bp->b_resid = 0;
2290         bp->b_error = error;
2291         if (error)
2292                 bp->b_flags |= B_ERROR;
2293         biodone(ap->a_bio);
2294
2295 done:
2296         if (cursor.node)
2297                 hammer_cache_node(&ip->cache[1], cursor.node);
2298         hammer_done_cursor(&cursor);
2299         hammer_done_transaction(&trans);
2300         return(error);
2301 }
2302
2303 /*
2304  * BMAP operation - used to support cluster_read() only.
2305  *
2306  * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
2307  *
2308  * This routine may return EOPNOTSUPP if the opration is not supported for
2309  * the specified offset.  The contents of the pointer arguments do not
2310  * need to be initialized in that case.
2311  *
2312  * If a disk address is available and properly aligned return 0 with
2313  * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
2314  * to the run-length relative to that offset.  Callers may assume that
2315  * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
2316  * large, so return EOPNOTSUPP if it is not sufficiently large.
2317  */
2318 static
2319 int
2320 hammer_vop_bmap(struct vop_bmap_args *ap)
2321 {
2322         struct hammer_transaction trans;
2323         struct hammer_inode *ip;
2324         struct hammer_cursor cursor;
2325         hammer_base_elm_t base;
2326         int64_t rec_offset;
2327         int64_t ran_end;
2328         int64_t tmp64;
2329         int64_t base_offset;
2330         int64_t base_disk_offset;
2331         int64_t last_offset;
2332         hammer_off_t last_disk_offset;
2333         hammer_off_t disk_offset;
2334         int     rec_len;
2335         int     error;
2336         int     blksize;
2337
2338         ++hammer_stats_file_iopsr;
2339         ip = ap->a_vp->v_data;
2340
2341         /*
2342          * We can only BMAP regular files.  We can't BMAP database files,
2343          * directories, etc.
2344          */
2345         if (ip->ino_data.obj_type != HAMMER_OBJTYPE_REGFILE)
2346                 return(EOPNOTSUPP);
2347
2348         /*
2349          * bmap is typically called with runp/runb both NULL when used
2350          * for writing.  We do not support BMAP for writing atm.
2351          */
2352         if (ap->a_cmd != BUF_CMD_READ)
2353                 return(EOPNOTSUPP);
2354
2355         /*
2356          * Scan the B-Tree to acquire blockmap addresses, then translate
2357          * to raw addresses.
2358          */
2359         hammer_simple_transaction(&trans, ip->hmp);
2360 #if 0
2361         kprintf("bmap_beg %016llx ip->cache %p\n", ap->a_loffset, ip->cache[1]);
2362 #endif
2363         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2364
2365         /*
2366          * Key range (begin and end inclusive) to scan.  Note that the key's
2367          * stored in the actual records represent BASE+LEN, not BASE.  The
2368          * first record containing bio_offset will have a key > bio_offset.
2369          */
2370         cursor.key_beg.localization = ip->obj_localization +
2371                                       HAMMER_LOCALIZE_MISC;
2372         cursor.key_beg.obj_id = ip->obj_id;
2373         cursor.key_beg.create_tid = 0;
2374         cursor.key_beg.delete_tid = 0;
2375         cursor.key_beg.obj_type = 0;
2376         if (ap->a_runb)
2377                 cursor.key_beg.key = ap->a_loffset - MAXPHYS + 1;
2378         else
2379                 cursor.key_beg.key = ap->a_loffset + 1;
2380         if (cursor.key_beg.key < 0)
2381                 cursor.key_beg.key = 0;
2382         cursor.asof = ip->obj_asof;
2383         cursor.flags |= HAMMER_CURSOR_ASOF;
2384
2385         cursor.key_end = cursor.key_beg;
2386         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2387
2388         ran_end = ap->a_loffset + MAXPHYS;
2389         cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2390         cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2391         tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2392         if (tmp64 < ran_end)
2393                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2394         else
2395                 cursor.key_end.key = ran_end + MAXPHYS + 1;
2396
2397         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2398
2399         error = hammer_ip_first(&cursor);
2400         base_offset = last_offset = 0;
2401         base_disk_offset = last_disk_offset = 0;
2402
2403         while (error == 0) {
2404                 /*
2405                  * Get the base file offset of the record.  The key for
2406                  * data records is (base + bytes) rather then (base).
2407                  *
2408                  * NOTE: rec_offset + rec_len may exceed the end-of-file.
2409                  * The extra bytes should be zero on-disk and the BMAP op
2410                  * should still be ok.
2411                  */
2412                 base = &cursor.leaf->base;
2413                 rec_offset = base->key - cursor.leaf->data_len;
2414                 rec_len    = cursor.leaf->data_len;
2415
2416                 /*
2417                  * Incorporate any cached truncation.
2418                  *
2419                  * NOTE: Modifications to rec_len based on synthesized
2420                  * truncation points remove the guarantee that any extended
2421                  * data on disk is zero (since the truncations may not have
2422                  * taken place on-media yet).
2423                  */
2424                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2425                         if (hammer_cursor_ondisk(&cursor) ||
2426                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
2427                                 if (ip->trunc_off <= rec_offset)
2428                                         rec_len = 0;
2429                                 else if (ip->trunc_off < rec_offset + rec_len)
2430                                         rec_len = (int)(ip->trunc_off - rec_offset);
2431                         }
2432                 }
2433                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2434                         if (hammer_cursor_ondisk(&cursor)) {
2435                                 if (ip->sync_trunc_off <= rec_offset)
2436                                         rec_len = 0;
2437                                 else if (ip->sync_trunc_off < rec_offset + rec_len)
2438                                         rec_len = (int)(ip->sync_trunc_off - rec_offset);
2439                         }
2440                 }
2441
2442                 /*
2443                  * Accumulate information.  If we have hit a discontiguous
2444                  * block reset base_offset unless we are already beyond the
2445                  * requested offset.  If we are, that's it, we stop.
2446                  */
2447                 if (error)
2448                         break;
2449                 if (hammer_cursor_ondisk(&cursor)) {
2450                         disk_offset = cursor.leaf->data_offset;
2451                         if (rec_offset != last_offset ||
2452                             disk_offset != last_disk_offset) {
2453                                 if (rec_offset > ap->a_loffset)
2454                                         break;
2455                                 base_offset = rec_offset;
2456                                 base_disk_offset = disk_offset;
2457                         }
2458                         last_offset = rec_offset + rec_len;
2459                         last_disk_offset = disk_offset + rec_len;
2460                 }
2461                 error = hammer_ip_next(&cursor);
2462         }
2463
2464 #if 0
2465         kprintf("BMAP %016llx:  %016llx - %016llx\n",
2466                 ap->a_loffset, base_offset, last_offset);
2467         kprintf("BMAP %16s:  %016llx - %016llx\n",
2468                 "", base_disk_offset, last_disk_offset);
2469 #endif
2470
2471         if (cursor.node) {
2472                 hammer_cache_node(&ip->cache[1], cursor.node);
2473 #if 0
2474                 kprintf("bmap_end2 %016llx ip->cache %p\n", ap->a_loffset, ip->cache[1]);
2475 #endif
2476         }
2477         hammer_done_cursor(&cursor);
2478         hammer_done_transaction(&trans);
2479
2480         /*
2481          * If we couldn't find any records or the records we did find were
2482          * all behind the requested offset, return failure.  A forward
2483          * truncation can leave a hole w/ no on-disk records.
2484          */
2485         if (last_offset == 0 || last_offset < ap->a_loffset)
2486                 return (EOPNOTSUPP);
2487
2488         /*
2489          * Figure out the block size at the requested offset and adjust
2490          * our limits so the cluster_read() does not create inappropriately
2491          * sized buffer cache buffers.
2492          */
2493         blksize = hammer_blocksize(ap->a_loffset);
2494         if (hammer_blocksize(base_offset) != blksize) {
2495                 base_offset = hammer_blockdemarc(base_offset, ap->a_loffset);
2496         }
2497         if (last_offset != ap->a_loffset &&
2498             hammer_blocksize(last_offset - 1) != blksize) {
2499                 last_offset = hammer_blockdemarc(ap->a_loffset,
2500                                                  last_offset - 1);
2501         }
2502
2503         /*
2504          * Returning EOPNOTSUPP simply prevents the direct-IO optimization
2505          * from occuring.
2506          */
2507         disk_offset = base_disk_offset + (ap->a_loffset - base_offset);
2508
2509         if ((disk_offset & HAMMER_OFF_ZONE_MASK) != HAMMER_ZONE_LARGE_DATA) {
2510                 /*
2511                  * Only large-data zones can be direct-IOd
2512                  */
2513                 error = EOPNOTSUPP;
2514         } else if ((disk_offset & HAMMER_BUFMASK) ||
2515                    (last_offset - ap->a_loffset) < blksize) {
2516                 /*
2517                  * doffsetp is not aligned or the forward run size does
2518                  * not cover a whole buffer, disallow the direct I/O.
2519                  */
2520                 error = EOPNOTSUPP;
2521         } else {
2522                 /*
2523                  * We're good.
2524                  */
2525                 *ap->a_doffsetp = disk_offset;
2526                 if (ap->a_runb) {
2527                         *ap->a_runb = ap->a_loffset - base_offset;
2528                         KKASSERT(*ap->a_runb >= 0);
2529                 }
2530                 if (ap->a_runp) {
2531                         *ap->a_runp = last_offset - ap->a_loffset;
2532                         KKASSERT(*ap->a_runp >= 0);
2533                 }
2534                 error = 0;
2535         }
2536         return(error);
2537 }
2538
2539 /*
2540  * Write to a regular file.   Because this is a strategy call the OS is
2541  * trying to actually get data onto the media.
2542  */
2543 static
2544 int
2545 hammer_vop_strategy_write(struct vop_strategy_args *ap)
2546 {
2547         hammer_record_t record;
2548         hammer_mount_t hmp;
2549         hammer_inode_t ip;
2550         struct bio *bio;
2551         struct buf *bp;
2552         int blksize;
2553         int bytes;
2554         int error;
2555
2556         bio = ap->a_bio;
2557         bp = bio->bio_buf;
2558         ip = ap->a_vp->v_data;
2559         hmp = ip->hmp;
2560
2561         blksize = hammer_blocksize(bio->bio_offset);
2562         KKASSERT(bp->b_bufsize == blksize);
2563
2564         if (ip->flags & HAMMER_INODE_RO) {
2565                 bp->b_error = EROFS;
2566                 bp->b_flags |= B_ERROR;
2567                 biodone(ap->a_bio);
2568                 return(EROFS);
2569         }
2570
2571         /*
2572          * Interlock with inode destruction (no in-kernel or directory
2573          * topology visibility).  If we queue new IO while trying to
2574          * destroy the inode we can deadlock the vtrunc call in
2575          * hammer_inode_unloadable_check().
2576          *
2577          * Besides, there's no point flushing a bp associated with an
2578          * inode that is being destroyed on-media and has no kernel
2579          * references.
2580          */
2581         if ((ip->flags | ip->sync_flags) &
2582             (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) {
2583                 bp->b_resid = 0;
2584                 biodone(ap->a_bio);
2585                 return(0);
2586         }
2587
2588         /*
2589          * Reserve space and issue a direct-write from the front-end.
2590          * NOTE: The direct_io code will hammer_bread/bcopy smaller
2591          * allocations.
2592          *
2593          * An in-memory record will be installed to reference the storage
2594          * until the flusher can get to it.
2595          *
2596          * Since we own the high level bio the front-end will not try to
2597          * do a direct-read until the write completes.
2598          *
2599          * NOTE: The only time we do not reserve a full-sized buffers
2600          * worth of data is if the file is small.  We do not try to
2601          * allocate a fragment (from the small-data zone) at the end of
2602          * an otherwise large file as this can lead to wildly separated
2603          * data.
2604          */
2605         KKASSERT((bio->bio_offset & HAMMER_BUFMASK) == 0);
2606         KKASSERT(bio->bio_offset < ip->ino_data.size);
2607         if (bio->bio_offset || ip->ino_data.size > HAMMER_BUFSIZE / 2)
2608                 bytes = bp->b_bufsize;
2609         else
2610                 bytes = ((int)ip->ino_data.size + 15) & ~15;
2611
2612         record = hammer_ip_add_bulk(ip, bio->bio_offset, bp->b_data,
2613                                     bytes, &error);
2614         if (record) {
2615                 hammer_io_direct_write(hmp, record, bio);
2616                 if (ip->rsv_recs > 1 && hmp->rsv_recs > hammer_limit_recs)
2617                         hammer_flush_inode(ip, 0);
2618         } else {
2619                 bp->b_bio2.bio_offset = NOOFFSET;
2620                 bp->b_error = error;
2621                 bp->b_flags |= B_ERROR;
2622                 biodone(ap->a_bio);
2623         }
2624         return(error);
2625 }
2626
2627 /*
2628  * dounlink - disconnect a directory entry
2629  *
2630  * XXX whiteout support not really in yet
2631  */
2632 static int
2633 hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
2634                 struct vnode *dvp, struct ucred *cred,
2635                 int flags, int isdir)
2636 {
2637         struct namecache *ncp;
2638         hammer_inode_t dip;
2639         hammer_inode_t ip;
2640         struct hammer_cursor cursor;
2641         int64_t namekey;
2642         int nlen, error;
2643
2644         /*
2645          * Calculate the namekey and setup the key range for the scan.  This
2646          * works kinda like a chained hash table where the lower 32 bits
2647          * of the namekey synthesize the chain.
2648          *
2649          * The key range is inclusive of both key_beg and key_end.
2650          */
2651         dip = VTOI(dvp);
2652         ncp = nch->ncp;
2653
2654         if (dip->flags & HAMMER_INODE_RO)
2655                 return (EROFS);
2656
2657         namekey = hammer_directory_namekey(ncp->nc_name, ncp->nc_nlen);
2658 retry:
2659         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
2660         cursor.key_beg.localization = dip->obj_localization +
2661                                       HAMMER_LOCALIZE_MISC;
2662         cursor.key_beg.obj_id = dip->obj_id;
2663         cursor.key_beg.key = namekey;
2664         cursor.key_beg.create_tid = 0;
2665         cursor.key_beg.delete_tid = 0;
2666         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
2667         cursor.key_beg.obj_type = 0;
2668
2669         cursor.key_end = cursor.key_beg;
2670         cursor.key_end.key |= 0xFFFFFFFFULL;
2671         cursor.asof = dip->obj_asof;
2672         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2673
2674         /*
2675          * Scan all matching records (the chain), locate the one matching
2676          * the requested path component.  info->last_error contains the
2677          * error code on search termination and could be 0, ENOENT, or
2678          * something else.
2679          *
2680          * The hammer_ip_*() functions merge in-memory records with on-disk
2681          * records for the purposes of the search.
2682          */
2683         error = hammer_ip_first(&cursor);
2684
2685         while (error == 0) {
2686                 error = hammer_ip_resolve_data(&cursor);
2687                 if (error)
2688                         break;
2689                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
2690                 KKASSERT(nlen > 0);
2691                 if (ncp->nc_nlen == nlen &&
2692                     bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
2693                         break;
2694                 }
2695                 error = hammer_ip_next(&cursor);
2696         }
2697
2698         /*
2699          * If all is ok we have to get the inode so we can adjust nlinks.
2700          * To avoid a deadlock with the flusher we must release the inode
2701          * lock on the directory when acquiring the inode for the entry.
2702          *
2703          * If the target is a directory, it must be empty.
2704          */
2705         if (error == 0) {
2706                 hammer_unlock(&cursor.ip->lock);
2707                 ip = hammer_get_inode(trans, dip, cursor.data->entry.obj_id,
2708                                       dip->hmp->asof,
2709                                       cursor.data->entry.localization,
2710                                       0, &error);
2711                 hammer_lock_sh(&cursor.ip->lock);
2712                 if (error == ENOENT) {
2713                         kprintf("obj_id %016llx\n", cursor.data->entry.obj_id);
2714                         Debugger("ENOENT unlinking object that should exist");
2715                 }
2716
2717                 /*
2718                  * If isdir >= 0 we validate that the entry is or is not a
2719                  * directory.  If isdir < 0 we don't care.
2720                  */
2721                 if (error == 0 && isdir >= 0) {
2722                         if (isdir &&
2723                             ip->ino_data.obj_type != HAMMER_OBJTYPE_DIRECTORY) {
2724                                 error = ENOTDIR;
2725                         } else if (isdir == 0 &&
2726                             ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
2727                                 error = EISDIR;
2728                         }
2729                 }
2730
2731                 /*
2732                  * If we are trying to remove a directory the directory must
2733                  * be empty.
2734                  *
2735                  * WARNING: hammer_ip_check_directory_empty() may have to
2736                  * terminate the cursor to avoid a deadlock.  It is ok to
2737                  * call hammer_done_cursor() twice.
2738                  */
2739                 if (error == 0 && ip->ino_data.obj_type ==
2740                                   HAMMER_OBJTYPE_DIRECTORY) {
2741                         error = hammer_ip_check_directory_empty(trans, ip);
2742                 }
2743
2744                 /*
2745                  * Delete the directory entry.
2746                  *
2747                  * WARNING: hammer_ip_del_directory() may have to terminate
2748                  * the cursor to avoid a deadlock.  It is ok to call
2749                  * hammer_done_cursor() twice.
2750                  */
2751                 if (error == 0) {
2752                         error = hammer_ip_del_directory(trans, &cursor,
2753                                                         dip, ip);
2754                 }
2755                 hammer_done_cursor(&cursor);
2756                 if (error == 0) {
2757                         cache_setunresolved(nch);
2758                         cache_setvp(nch, NULL);
2759                         /* XXX locking */
2760                         if (ip->vp)
2761                                 cache_inval_vp(ip->vp, CINV_DESTROY);
2762                 }
2763                 if (ip)
2764                         hammer_rel_inode(ip, 0);
2765         } else {
2766                 hammer_done_cursor(&cursor);
2767         }
2768         if (error == EDEADLK)
2769                 goto retry;
2770
2771         return (error);
2772 }
2773
2774 /************************************************************************
2775  *                          FIFO AND SPECFS OPS                         *
2776  ************************************************************************
2777  *
2778  */
2779
2780 static int
2781 hammer_vop_fifoclose (struct vop_close_args *ap)
2782 {
2783         /* XXX update itimes */
2784         return (VOCALL(&fifo_vnode_vops, &ap->a_head));
2785 }
2786
2787 static int
2788 hammer_vop_fiforead (struct vop_read_args *ap)
2789 {
2790         int error;
2791
2792         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
2793         /* XXX update access time */
2794         return (error);
2795 }
2796
2797 static int
2798 hammer_vop_fifowrite (struct vop_write_args *ap)
2799 {
2800         int error;
2801
2802         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
2803         /* XXX update access time */
2804         return (error);
2805 }
2806
2807 static int
2808 hammer_vop_specclose (struct vop_close_args *ap)
2809 {
2810         /* XXX update itimes */
2811         return (VOCALL(&spec_vnode_vops, &ap->a_head));
2812 }
2813
2814 static int
2815 hammer_vop_specread (struct vop_read_args *ap)
2816 {
2817         /* XXX update access time */
2818         return (VOCALL(&spec_vnode_vops, &ap->a_head));
2819 }
2820
2821 static int
2822 hammer_vop_specwrite (struct vop_write_args *ap)
2823 {
2824         /* XXX update last change time */
2825         return (VOCALL(&spec_vnode_vops, &ap->a_head));
2826 }
2827