sys/vfs/hammer/hammer_vnops.c

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