sys/vfs/hammer2/hammer2.h

   1 /*
   2  * Copyright (c) 2011-2014 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@dragonflybsd.org>
   6  * by Venkatesh Srinivas <vsrinivas@dragonflybsd.org>
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  *
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in
  16  *    the documentation and/or other materials provided with the
  17  *    distribution.
  18  * 3. Neither the name of The DragonFly Project nor the names of its
  19  *    contributors may be used to endorse or promote products derived
  20  *    from this software without specific, prior written permission.
  21  *
  22  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  23  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  24  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  25  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  26  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  27  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  28  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  29  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  30  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  31  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  32  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  33  * SUCH DAMAGE.
  34  */
  35
  36 /*
  37  * HAMMER2 IN-MEMORY CACHE OF MEDIA STRUCTURES
  38  *
  39  * This header file contains structures used internally by the HAMMER2
  40  * implementation.  See hammer2_disk.h for on-disk structures.
  41  *
  42  * There is an in-memory representation of all on-media data structure.
  43  * Almost everything is represented by a hammer2_chain structure in-memory.
  44  * Other higher-level structures typically map to chains.
  45  *
  46  * A great deal of data is accessed simply via its buffer cache buffer,
  47  * which is mapped for the duration of the chain's lock.  Hammer2 must
  48  * implement its own buffer cache layer on top of the system layer to
  49  * allow for different threads to lock different sub-block-sized buffers.
  50  *
  51  * When modifications are made to a chain a new filesystem block must be
  52  * allocated.  Multiple modifications do not typically allocate new blocks
  53  * until the current block has been flushed.  Flushes do not block the
  54  * front-end unless the front-end operation crosses the current inode being
  55  * flushed.
  56  *
  57  * The in-memory representation may remain cached (for example in order to
  58  * placemark clustering locks) even after the related data has been
  59  * detached.
  60  */
  61
  62 #ifndef _VFS_HAMMER2_HAMMER2_H_
  63 #define _VFS_HAMMER2_HAMMER2_H_
  64
  65 #include <sys/param.h>
  66 #include <sys/types.h>
  67 #include <sys/kernel.h>
  68 #include <sys/conf.h>
  69 #include <sys/systm.h>
  70 #include <sys/tree.h>
  71 #include <sys/malloc.h>
  72 #include <sys/mount.h>
  73 #include <sys/vnode.h>
  74 #include <sys/proc.h>
  75 #include <sys/mountctl.h>
  76 #include <sys/priv.h>
  77 #include <sys/stat.h>
  78 #include <sys/thread.h>
  79 #include <sys/globaldata.h>
  80 #include <sys/lockf.h>
  81 #include <sys/buf.h>
  82 #include <sys/queue.h>
  83 #include <sys/limits.h>
  84 #include <sys/dmsg.h>
  85 #include <sys/mutex.h>
  86 #include <sys/kern_syscall.h>
  87
  88 #include <sys/signal2.h>
  89 #include <sys/buf2.h>
  90 #include <sys/mutex2.h>
  91
  92 #include "hammer2_disk.h"
  93 #include "hammer2_mount.h"
  94 #include "hammer2_ioctl.h"
  95
  96 struct hammer2_io;
  97 struct hammer2_iocb;
  98 struct hammer2_chain;
  99 struct hammer2_cluster;
 100 struct hammer2_inode;
 101 struct hammer2_mount;
 102 struct hammer2_pfsmount;
 103 struct hammer2_span;
 104 struct hammer2_state;
 105 struct hammer2_msg;
 106
 107 /*
 108  * Mutex and lock shims.  Hammer2 requires support for asynchronous and
 109  * abortable locks, and both exclusive and shared spinlocks.  Normal
 110  * synchronous non-abortable locks can be substituted for spinlocks.
 111  */
 112 typedef mtx_t                           hammer2_mtx_t;
 113 typedef mtx_link_t                      hammer2_mtx_link_t;
 114 typedef mtx_state_t                     hammer2_mtx_state_t;
 115
 116 typedef struct spinlock                 hammer2_spin_t;
 117
 118 #define hammer2_mtx_ex                  mtx_lock_ex_quick
 119 #define hammer2_mtx_sh                  mtx_lock_sh_quick
 120 #define hammer2_mtx_unlock              mtx_unlock
 121 #define hammer2_mtx_owned               mtx_owned
 122 #define hammer2_mtx_init                mtx_init
 123 #define hammer2_mtx_temp_release        mtx_lock_temp_release
 124 #define hammer2_mtx_temp_restore        mtx_lock_temp_restore
 125 #define hammer2_mtx_refs                mtx_lockrefs
 126
 127 #define hammer2_spin_init               spin_init
 128 #define hammer2_spin_sh                 spin_lock_shared
 129 #define hammer2_spin_ex                 spin_lock
 130 #define hammer2_spin_unsh               spin_unlock_shared
 131 #define hammer2_spin_unex               spin_unlock
 132
 133 /*
 134  * General lock support
 135  */
 136 static __inline
 137 int
 138 hammer2_mtx_upgrade(hammer2_mtx_t *mtx)
 139 {
 140         int wasexclusive;
 141
 142         if (mtx_islocked_ex(mtx)) {
 143                 wasexclusive = 1;
 144         } else {
 145                 mtx_unlock(mtx);
 146                 mtx_lock_ex_quick(mtx);
 147                 wasexclusive = 0;
 148         }
 149         return wasexclusive;
 150 }
 151
 152 /*
 153  * Downgrade an inode lock from exclusive to shared only if the inode
 154  * lock was previously shared.  If the inode lock was previously exclusive,
 155  * this is a NOP.
 156  */
 157 static __inline
 158 void
 159 hammer2_mtx_downgrade(hammer2_mtx_t *mtx, int wasexclusive)
 160 {
 161         if (wasexclusive == 0)
 162                 mtx_downgrade(mtx);
 163 }
 164
 165 /*
 166  * The xid tracks internal transactional updates.
 167  *
 168  * XXX fix-me, really needs to be 64-bits
 169  */
 170 typedef uint32_t hammer2_xid_t;
 171
 172 #define HAMMER2_XID_MIN 0x00000000U
 173 #define HAMMER2_XID_MAX 0x7FFFFFFFU
 174
 175 /*
 176  * The chain structure tracks a portion of the media topology from the
 177  * root (volume) down.  Chains represent volumes, inodes, indirect blocks,
 178  * data blocks, and freemap nodes and leafs.
 179  *
 180  * The chain structure utilizes a simple singly-homed topology and the
 181  * chain's in-memory topology will move around as the chains do, due mainly
 182  * to renames and indirect block creation.
 183  *
 184  * Block Table Updates
 185  *
 186  *      Block table updates for insertions and updates are delayed until the
 187  *      flush.  This allows us to avoid having to modify the parent chain
 188  *      all the way to the root.
 189  *
 190  *      Block table deletions are performed immediately (modifying the parent
 191  *      in the process) because the flush code uses the chain structure to
 192  *      track delayed updates and the chain will be (likely) gone or moved to
 193  *      another location in the topology after a deletion.
 194  *
 195  *      A prior iteration of the code tried to keep the relationship intact
 196  *      on deletes by doing a delete-duplicate operation on the chain, but
 197  *      it added way too much complexity to the codebase.
 198  *
 199  * Flush Synchronization
 200  *
 201  *      The flush code must flush modified chains bottom-up.  Because chain
 202  *      structures can shift around and are NOT topologically stable,
 203  *      modified chains are independently indexed for the flush.  As the flush
 204  *      runs it modifies (or further modifies) and updates the parents,
 205  *      propagating the flush all the way to the volume root.
 206  *
 207  *      Modifying front-end operations can occur during a flush but will block
 208  *      in two cases: (1) when the front-end tries to operate on the inode
 209  *      currently in the midst of being flushed and (2) if the front-end
 210  *      crosses an inode currently being flushed (such as during a rename).
 211  *      So, for example, if you rename directory "x" to "a/b/c/d/e/f/g/x" and
 212  *      the flusher is currently working on "a/b/c", the rename will block
 213  *      temporarily in order to ensure that "x" exists in one place or the
 214  *      other.
 215  *
 216  *      Meta-data statistics are updated by the flusher.  The front-end will
 217  *      make estimates but meta-data must be fully synchronized only during a
 218  *      flush in order to ensure that it remains correct across a crash.
 219  *
 220  *      Multiple flush synchronizations can theoretically be in-flight at the
 221  *      same time but the implementation is not coded to handle the case and
 222  *      currently serializes them.
 223  *
 224  * Snapshots:
 225  *
 226  *      Snapshots currently require the subdirectory tree being snapshotted
 227  *      to be flushed.  The snapshot then creates a new super-root inode which
 228  *      copies the flushed blockdata of the directory or file that was
 229  *      snapshotted.
 230  *
 231  * RBTREE NOTES:
 232  *
 233  *      - Note that the radix tree runs in powers of 2 only so sub-trees
 234  *        cannot straddle edges.
 235  */
 236 RB_HEAD(hammer2_chain_tree, hammer2_chain);
 237 TAILQ_HEAD(h2_flush_list, hammer2_chain);
 238 TAILQ_HEAD(h2_core_list, hammer2_chain);
 239 TAILQ_HEAD(h2_iocb_list, hammer2_iocb);
 240
 241 #define CHAIN_CORE_DELETE_BMAP_ENTRIES  \
 242         (HAMMER2_PBUFSIZE / sizeof(hammer2_blockref_t) / sizeof(uint32_t))
 243
 244 struct hammer2_chain_core {
 245         hammer2_mtx_t   lock;
 246         hammer2_spin_t  spin;
 247         struct hammer2_chain_tree rbtree; /* sub-chains */
 248         int             live_zero;      /* blockref array opt */
 249         u_int           flags;
 250         u_int           live_count;     /* live (not deleted) chains in tree */
 251         u_int           chain_count;    /* live + deleted chains under core */
 252         int             generation;     /* generation number (inserts only) */
 253 };
 254
 255 typedef struct hammer2_chain_core hammer2_chain_core_t;
 256
 257 #define HAMMER2_CORE_UNUSED0001         0x0001
 258 #define HAMMER2_CORE_COUNTEDBREFS       0x0002
 259
 260 RB_HEAD(hammer2_io_tree, hammer2_io);
 261
 262 /*
 263  * IOCB - IO callback (into chain, cluster, or manual request)
 264  */
 265 struct hammer2_iocb {
 266         TAILQ_ENTRY(hammer2_iocb) entry;
 267         void (*callback)(struct hammer2_iocb *iocb);
 268         struct hammer2_io       *dio;
 269         struct hammer2_cluster  *cluster;
 270         struct hammer2_chain    *chain;
 271         void                    *ptr;
 272         off_t                   lbase;
 273         int                     lsize;
 274         uint32_t                flags;
 275         int                     error;
 276 };
 277
 278 typedef struct hammer2_iocb hammer2_iocb_t;
 279
 280 #define HAMMER2_IOCB_INTERLOCK  0x00000001
 281 #define HAMMER2_IOCB_ONQ        0x00000002
 282 #define HAMMER2_IOCB_DONE       0x00000004
 283 #define HAMMER2_IOCB_INPROG     0x00000008
 284 #define HAMMER2_IOCB_UNUSED10   0x00000010
 285 #define HAMMER2_IOCB_QUICK      0x00010000
 286 #define HAMMER2_IOCB_ZERO       0x00020000
 287 #define HAMMER2_IOCB_READ       0x00040000
 288 #define HAMMER2_IOCB_WAKEUP     0x00080000
 289
 290 /*
 291  * DIO - Management structure wrapping system buffer cache.
 292  *
 293  *       Used for multiple purposes including concurrent management
 294  *       if small requests by chains into larger DIOs.
 295  */
 296 struct hammer2_io {
 297         RB_ENTRY(hammer2_io) rbnode;    /* indexed by device offset */
 298         struct h2_iocb_list iocbq;
 299         struct spinlock spin;
 300         struct hammer2_mount *hmp;
 301         struct buf      *bp;
 302         off_t           pbase;
 303         int             psize;
 304         int             refs;
 305         int             act;                    /* activity */
 306 };
 307
 308 typedef struct hammer2_io hammer2_io_t;
 309
 310 #define HAMMER2_DIO_INPROG      0x80000000      /* bio in progress */
 311 #define HAMMER2_DIO_GOOD        0x40000000      /* dio->bp is stable */
 312 #define HAMMER2_DIO_WAITING     0x20000000      /* (old) */
 313 #define HAMMER2_DIO_DIRTY       0x10000000      /* flush on last drop */
 314
 315 #define HAMMER2_DIO_MASK        0x0FFFFFFF
 316
 317 /*
 318  * Primary chain structure keeps track of the topology in-memory.
 319  */
 320 struct hammer2_chain {
 321         hammer2_chain_core_t    core;
 322         RB_ENTRY(hammer2_chain) rbnode;         /* live chain(s) */
 323         hammer2_blockref_t      bref;
 324         struct hammer2_chain    *parent;
 325         struct hammer2_state    *state;         /* if active cache msg */
 326         struct hammer2_mount    *hmp;
 327         struct hammer2_pfsmount *pmp;           /* (pfs-cluster pmp or spmp) */
 328
 329         hammer2_xid_t   flush_xid;              /* flush sequencing */
 330         hammer2_key_t   data_count;             /* delta's to apply */
 331         hammer2_key_t   inode_count;            /* delta's to apply */
 332         hammer2_key_t   data_count_up;          /* delta's to apply */
 333         hammer2_key_t   inode_count_up;         /* delta's to apply */
 334         hammer2_io_t    *dio;                   /* physical data buffer */
 335         u_int           bytes;                  /* physical data size */
 336         u_int           flags;
 337         u_int           refs;
 338         u_int           lockcnt;
 339         hammer2_media_data_t *data;             /* data pointer shortcut */
 340         TAILQ_ENTRY(hammer2_chain) flush_node;  /* flush list */
 341 };
 342
 343 typedef struct hammer2_chain hammer2_chain_t;
 344
 345 int hammer2_chain_cmp(hammer2_chain_t *chain1, hammer2_chain_t *chain2);
 346 RB_PROTOTYPE(hammer2_chain_tree, hammer2_chain, rbnode, hammer2_chain_cmp);
 347
 348 /*
 349  * Special notes on flags:
 350  *
 351  * INITIAL - This flag allows a chain to be created and for storage to
 352  *           be allocated without having to immediately instantiate the
 353  *           related buffer.  The data is assumed to be all-zeros.  It
 354  *           is primarily used for indirect blocks.
 355  *
 356  * MODIFIED- The chain's media data has been modified.
 357  * UPDATE  - Chain might not be modified but parent blocktable needs update
 358  *
 359  * BMAPPED - Indicates that the chain is present in the parent blockmap.
 360  * BMAPUPD - Indicates that the chain is present but needs to be updated
 361  *           in the parent blockmap.
 362  */
 363 #define HAMMER2_CHAIN_MODIFIED          0x00000001      /* dirty chain data */
 364 #define HAMMER2_CHAIN_ALLOCATED         0x00000002      /* kmalloc'd chain */
 365 #define HAMMER2_CHAIN_DESTROY           0x00000004
 366 #define HAMMER2_CHAIN_UNLINKED          0x00000008      /* unlinked file */
 367 #define HAMMER2_CHAIN_DELETED           0x00000010      /* deleted chain */
 368 #define HAMMER2_CHAIN_INITIAL           0x00000020      /* initial create */
 369 #define HAMMER2_CHAIN_UPDATE            0x00000040      /* need parent update */
 370 #define HAMMER2_CHAIN_DEFERRED          0x00000080      /* flush depth defer */
 371 #define HAMMER2_CHAIN_IOFLUSH           0x00000100      /* bawrite on put */
 372 #define HAMMER2_CHAIN_ONFLUSH           0x00000200      /* on a flush list */
 373 #define HAMMER2_CHAIN_UNUSED00000400    0x00000400
 374 #define HAMMER2_CHAIN_VOLUMESYNC        0x00000800      /* needs volume sync */
 375 #define HAMMER2_CHAIN_UNUSED00001000    0x00001000
 376 #define HAMMER2_CHAIN_MOUNTED           0x00002000      /* PFS is mounted */
 377 #define HAMMER2_CHAIN_ONRBTREE          0x00004000      /* on parent RB tree */
 378 #define HAMMER2_CHAIN_SNAPSHOT          0x00008000      /* snapshot special */
 379 #define HAMMER2_CHAIN_EMBEDDED          0x00010000      /* embedded data */
 380 #define HAMMER2_CHAIN_RELEASE           0x00020000      /* don't keep around */
 381 #define HAMMER2_CHAIN_BMAPPED           0x00040000      /* present in blkmap */
 382 #define HAMMER2_CHAIN_BMAPUPD           0x00080000      /* +needs updating */
 383 #define HAMMER2_CHAIN_UNUSED00100000    0x00100000
 384 #define HAMMER2_CHAIN_UNUSED00200000    0x00200000
 385 #define HAMMER2_CHAIN_PFSBOUNDARY       0x00400000      /* super->pfs inode */
 386
 387 #define HAMMER2_CHAIN_FLUSH_MASK        (HAMMER2_CHAIN_MODIFIED |       \
 388                                          HAMMER2_CHAIN_UPDATE |         \
 389                                          HAMMER2_CHAIN_ONFLUSH)
 390
 391 /*
 392  * Flags passed to hammer2_chain_lookup() and hammer2_chain_next()
 393  *
 394  * NOTE: MATCHIND allows an indirect block / freemap node to be returned
 395  *       when the passed key range matches the radix.  Remember that key_end
 396  *       is inclusive (e.g. {0x000,0xFFF}, not {0x000,0x1000}).
 397  */
 398 #define HAMMER2_LOOKUP_NOLOCK           0x00000001      /* ref only */
 399 #define HAMMER2_LOOKUP_NODATA           0x00000002      /* data left NULL */
 400 #define HAMMER2_LOOKUP_SHARED           0x00000100
 401 #define HAMMER2_LOOKUP_MATCHIND         0x00000200      /* return all chains */
 402 #define HAMMER2_LOOKUP_UNUSED0400       0x00000400
 403 #define HAMMER2_LOOKUP_ALWAYS           0x00000800      /* resolve data */
 404
 405 /*
 406  * Flags passed to hammer2_chain_modify() and hammer2_chain_resize()
 407  *
 408  * NOTE: OPTDATA allows us to avoid instantiating buffers for INDIRECT
 409  *       blocks in the INITIAL-create state.
 410  */
 411 #define HAMMER2_MODIFY_OPTDATA          0x00000002      /* data can be NULL */
 412 #define HAMMER2_MODIFY_NO_MODIFY_TID    0x00000004
 413 #define HAMMER2_MODIFY_UNUSED0008       0x00000008
 414 #define HAMMER2_MODIFY_NOREALLOC        0x00000010
 415
 416 /*
 417  * Flags passed to hammer2_chain_lock()
 418  */
 419 #define HAMMER2_RESOLVE_NEVER           1
 420 #define HAMMER2_RESOLVE_MAYBE           2
 421 #define HAMMER2_RESOLVE_ALWAYS          3
 422 #define HAMMER2_RESOLVE_MASK            0x0F
 423
 424 #define HAMMER2_RESOLVE_SHARED          0x10    /* request shared lock */
 425 #define HAMMER2_RESOLVE_NOREF           0x20    /* already ref'd on lock */
 426
 427 /*
 428  * Flags passed to hammer2_chain_delete()
 429  */
 430 #define HAMMER2_DELETE_PERMANENT        0x0001
 431 #define HAMMER2_DELETE_NOSTATS          0x0002
 432
 433 #define HAMMER2_INSERT_NOSTATS          0x0002
 434
 435 /*
 436  * Flags passed to hammer2_chain_delete_duplicate()
 437  */
 438 #define HAMMER2_DELDUP_RECORE           0x0001
 439
 440 /*
 441  * Cluster different types of storage together for allocations
 442  */
 443 #define HAMMER2_FREECACHE_INODE         0
 444 #define HAMMER2_FREECACHE_INDIR         1
 445 #define HAMMER2_FREECACHE_DATA          2
 446 #define HAMMER2_FREECACHE_UNUSED3       3
 447 #define HAMMER2_FREECACHE_TYPES         4
 448
 449 /*
 450  * hammer2_freemap_alloc() block preference
 451  */
 452 #define HAMMER2_OFF_NOPREF              ((hammer2_off_t)-1)
 453
 454 /*
 455  * BMAP read-ahead maximum parameters
 456  */
 457 #define HAMMER2_BMAP_COUNT              16      /* max bmap read-ahead */
 458 #define HAMMER2_BMAP_BYTES              (HAMMER2_PBUFSIZE * HAMMER2_BMAP_COUNT)
 459
 460 /*
 461  * hammer2_freemap_adjust()
 462  */
 463 #define HAMMER2_FREEMAP_DORECOVER       1
 464 #define HAMMER2_FREEMAP_DOMAYFREE       2
 465 #define HAMMER2_FREEMAP_DOREALFREE      3
 466
 467 /*
 468  * HAMMER2 cluster - A set of chains representing the same entity.
 469  *
 470  * hammer2_cluster typically represents a temporary set of representitive
 471  * chains.  The one exception is that a hammer2_cluster is embedded in
 472  * hammer2_inode.  This embedded cluster is ONLY used to track the
 473  * representitive chains and cannot be directly locked.
 474  *
 475  * A cluster is temporary (and thus per-thread) for locking purposes,
 476  * allowing us to embed the asynchronous storage required for
 477  * cluster operations in the cluster itself.  That is, except for the
 478  * embeddeding hammer2_inode, the cluster structure will always represent
 479  * a 'working copy'.
 480  *
 481  * Because the cluster is a 'working copy' and is usually subject to cluster
 482  * quorum rules, it is quite possible for us to end up with an insufficient
 483  * number of live chains to execute an operation.  If an insufficient number
 484  * of chains remain in a working copy, the operation may have to be
 485  * downgraded, retried, or stall until the requisit number of chains are
 486  * available.
 487  */
 488 #define HAMMER2_MAXCLUSTER      8
 489
 490 struct hammer2_cluster_item {
 491         hammer2_mtx_link_t      async_link;
 492         hammer2_chain_t         *chain;
 493         struct hammer2_cluster  *cluster;       /* link back to cluster */
 494         int                     cache_index;
 495         int                     unused01;
 496 };
 497
 498 typedef struct hammer2_cluster_item hammer2_cluster_item_t;
 499
 500 struct hammer2_cluster {
 501         int                     status;         /* operational status */
 502         int                     refs;           /* track for deallocation */
 503         struct hammer2_pfsmount *pmp;
 504         uint32_t                flags;
 505         int                     nchains;
 506         hammer2_iocb_t          iocb;
 507         hammer2_chain_t         *focus;         /* current focus (or mod) */
 508         hammer2_cluster_item_t  array[HAMMER2_MAXCLUSTER];
 509 };
 510
 511 typedef struct hammer2_cluster  hammer2_cluster_t;
 512
 513 #define HAMMER2_CLUSTER_INODE   0x00000001      /* embedded in inode */
 514 #define HAMMER2_CLUSTER_NOSYNC  0x00000002      /* not in sync (cumulative) */
 515
 516
 517 RB_HEAD(hammer2_inode_tree, hammer2_inode);
 518
 519 /*
 520  * A hammer2 inode.
 521  *
 522  * NOTE: The inode-embedded cluster is never used directly for I/O (since
 523  *       it may be shared).  Instead it will be replicated-in and synchronized
 524  *       back out if changed.
 525  */
 526 struct hammer2_inode {
 527         RB_ENTRY(hammer2_inode) rbnode;         /* inumber lookup (HL) */
 528         hammer2_mtx_t           lock;           /* inode lock */
 529         struct hammer2_pfsmount *pmp;           /* PFS mount */
 530         struct hammer2_inode    *pip;           /* parent inode */
 531         struct vnode            *vp;
 532         hammer2_cluster_t       cluster;
 533         struct lockf            advlock;
 534         hammer2_tid_t           inum;
 535         u_int                   flags;
 536         u_int                   refs;           /* +vpref, +flushref */
 537         uint8_t                 comp_heuristic;
 538         hammer2_off_t           size;
 539         uint64_t                mtime;
 540 };
 541
 542 typedef struct hammer2_inode hammer2_inode_t;
 543
 544 #define HAMMER2_INODE_MODIFIED          0x0001
 545 #define HAMMER2_INODE_SROOT             0x0002  /* kmalloc special case */
 546 #define HAMMER2_INODE_RENAME_INPROG     0x0004
 547 #define HAMMER2_INODE_ONRBTREE          0x0008
 548 #define HAMMER2_INODE_RESIZED           0x0010
 549 #define HAMMER2_INODE_MTIME             0x0020
 550
 551 int hammer2_inode_cmp(hammer2_inode_t *ip1, hammer2_inode_t *ip2);
 552 RB_PROTOTYPE2(hammer2_inode_tree, hammer2_inode, rbnode, hammer2_inode_cmp,
 553                 hammer2_tid_t);
 554
 555 /*
 556  * inode-unlink side-structure
 557  */
 558 struct hammer2_inode_unlink {
 559         TAILQ_ENTRY(hammer2_inode_unlink) entry;
 560         hammer2_inode_t *ip;
 561 };
 562 TAILQ_HEAD(h2_unlk_list, hammer2_inode_unlink);
 563
 564 typedef struct hammer2_inode_unlink hammer2_inode_unlink_t;
 565
 566 /*
 567  * A hammer2 transaction and flush sequencing structure.
 568  *
 569  * This global structure is tied into hammer2_mount and is used
 570  * to sequence modifying operations and flushes.
 571  *
 572  * (a) Any modifying operations with sync_tid >= flush_tid will stall until
 573  *     all modifying operating with sync_tid < flush_tid complete.
 574  *
 575  *     The flush related to flush_tid stalls until all modifying operations
 576  *     with sync_tid < flush_tid complete.
 577  *
 578  * (b) Once unstalled, modifying operations with sync_tid > flush_tid are
 579  *     allowed to run.  All modifications cause modify/duplicate operations
 580  *     to occur on the related chains.  Note that most INDIRECT blocks will
 581  *     be unaffected because the modifications just overload the RBTREE
 582  *     structurally instead of actually modifying the indirect blocks.
 583  *
 584  * (c) The actual flush unstalls and RUNS CONCURRENTLY with (b), but only
 585  *     utilizes the chain structures with sync_tid <= flush_tid.  The
 586  *     flush will modify related indirect blocks and inodes in-place
 587  *     (rather than duplicate) since the adjustments are compatible with
 588  *     (b)'s RBTREE overloading
 589  *
 590  *     SPECIAL NOTE:  Inode modifications have to also propagate along any
 591  *                    modify/duplicate chains.  File writes detect the flush
 592  *                    and force out the conflicting buffer cache buffer(s)
 593  *                    before reusing them.
 594  *
 595  * (d) Snapshots can be made instantly but must be flushed and disconnected
 596  *     from their duplicative source before they can be mounted.  This is
 597  *     because while H2's on-media structure supports forks, its in-memory
 598  *     structure only supports very simple forking for background flushing
 599  *     purposes.
 600  *
 601  * TODO: Flush merging.  When fsync() is called on multiple discrete files
 602  *       concurrently there is no reason to stall the second fsync.
 603  *       The final flush that reaches to root can cover both fsync()s.
 604  *
 605  *     The chains typically terminate as they fly onto the disk.  The flush
 606  *     ultimately reaches the volume header.
 607  */
 608 struct hammer2_trans {
 609         TAILQ_ENTRY(hammer2_trans) entry;
 610         struct hammer2_pfsmount *pmp;
 611         hammer2_xid_t           sync_xid;
 612         hammer2_tid_t           inode_tid;      /* inode number assignment */
 613         thread_t                td;             /* pointer */
 614         int                     flags;
 615         int                     blocked;
 616         uint8_t                 inodes_created;
 617         uint8_t                 dummy[7];
 618 };
 619
 620 typedef struct hammer2_trans hammer2_trans_t;
 621
 622 #define HAMMER2_TRANS_ISFLUSH           0x0001  /* formal flush */
 623 #define HAMMER2_TRANS_CONCURRENT        0x0002  /* concurrent w/flush */
 624 #define HAMMER2_TRANS_BUFCACHE          0x0004  /* from bioq strategy write */
 625 #define HAMMER2_TRANS_NEWINODE          0x0008  /* caller allocating inode */
 626 #define HAMMER2_TRANS_UNUSED0010        0x0010
 627 #define HAMMER2_TRANS_PREFLUSH          0x0020  /* preflush state */
 628
 629 #define HAMMER2_FREEMAP_HEUR_NRADIX     4       /* pwr 2 PBUFRADIX-MINIORADIX */
 630 #define HAMMER2_FREEMAP_HEUR_TYPES      8
 631 #define HAMMER2_FREEMAP_HEUR            (HAMMER2_FREEMAP_HEUR_NRADIX * \
 632                                          HAMMER2_FREEMAP_HEUR_TYPES)
 633
 634 /*
 635  * Transaction Rendezvous
 636  */
 637 TAILQ_HEAD(hammer2_trans_queue, hammer2_trans);
 638
 639 struct hammer2_trans_manage {
 640         hammer2_xid_t           flush_xid;      /* last flush transaction */
 641         hammer2_xid_t           alloc_xid;
 642         struct lock             translk;        /* lockmgr lock */
 643         struct hammer2_trans_queue transq;      /* modifying transactions */
 644         int                     flushcnt;       /* track flush trans */
 645 };
 646
 647 typedef struct hammer2_trans_manage hammer2_trans_manage_t;
 648
 649 /*
 650  * Global (per device) mount structure for device (aka vp->v_mount->hmp)
 651  */
 652 struct hammer2_mount {
 653         struct vnode    *devvp;         /* device vnode */
 654         int             ronly;          /* read-only mount */
 655         int             pmp_count;      /* PFS mounts backed by us */
 656         TAILQ_ENTRY(hammer2_mount) mntentry; /* hammer2_mntlist */
 657
 658         struct malloc_type *mchain;
 659         int             nipstacks;
 660         int             maxipstacks;
 661         kdmsg_iocom_t   iocom;          /* volume-level dmsg interface */
 662         struct spinlock io_spin;        /* iotree access */
 663         struct hammer2_io_tree iotree;
 664         int             iofree_count;
 665         hammer2_chain_t vchain;         /* anchor chain (topology) */
 666         hammer2_chain_t fchain;         /* anchor chain (freemap) */
 667         struct spinlock list_spin;
 668         struct h2_flush_list    flushq; /* flush seeds */
 669         struct hammer2_pfsmount *spmp;  /* super-root pmp for transactions */
 670         struct lock     vollk;          /* lockmgr lock */
 671         hammer2_off_t   heur_freemap[HAMMER2_FREEMAP_HEUR];
 672         int             volhdrno;       /* last volhdrno written */
 673         hammer2_volume_data_t voldata;
 674         hammer2_volume_data_t volsync;  /* synchronized voldata */
 675 };
 676
 677 typedef struct hammer2_mount hammer2_mount_t;
 678
 679 /*
 680  * HAMMER2 PFS mount point structure (aka vp->v_mount->mnt_data).
 681  * This has a 1:1 correspondence to struct mount (note that the
 682  * hammer2_mount structure has a N:1 correspondence).
 683  *
 684  * This structure represents a cluster mount and not necessarily a
 685  * PFS under a specific device mount (HMP).  The distinction is important
 686  * because the elements backing a cluster mount can change on the fly.
 687  *
 688  * Usually the first element under the cluster represents the original
 689  * user-requested mount that bootstraps the whole mess.  In significant
 690  * setups the original is usually just a read-only media image (or
 691  * representitive file) that simply contains a bootstrap volume header
 692  * listing the configuration.
 693  */
 694 struct hammer2_pfsmount {
 695         struct mount            *mp;
 696         TAILQ_ENTRY(hammer2_pfsmount) mntentry; /* hammer2_pfslist */
 697         uuid_t                  pfs_clid;
 698         uuid_t                  pfs_fsid;
 699         hammer2_mount_t         *spmp_hmp;      /* (spmp only) */
 700         hammer2_inode_t         *iroot;         /* PFS root inode */
 701         hammer2_inode_t         *ihidden;       /* PFS hidden directory */
 702         struct lock             lock;           /* PFS lock for certain ops */
 703         hammer2_off_t           inode_count;    /* copy of inode_count */
 704         struct netexport        export;         /* nfs export */
 705         int                     ronly;          /* read-only mount */
 706         struct malloc_type      *minode;
 707         struct malloc_type      *mmsg;
 708         struct spinlock         inum_spin;      /* inumber lookup */
 709         struct hammer2_inode_tree inum_tree;    /* (not applicable to spmp) */
 710         hammer2_tid_t           alloc_tid;
 711         hammer2_tid_t           flush_tid;
 712         hammer2_tid_t           inode_tid;
 713         long                    inmem_inodes;
 714         uint32_t                inmem_dirty_chains;
 715         int                     count_lwinprog; /* logical write in prog */
 716         struct spinlock         list_spin;
 717         struct h2_unlk_list     unlinkq;        /* last-close unlink */
 718         thread_t                wthread_td;     /* write thread td */
 719         struct bio_queue_head   wthread_bioq;   /* logical buffer bioq */
 720         hammer2_mtx_t           wthread_mtx;    /* interlock */
 721         int                     wthread_destroy;/* termination sequencing */
 722 };
 723
 724 typedef struct hammer2_pfsmount hammer2_pfsmount_t;
 725
 726 #define HAMMER2_DIRTYCHAIN_WAITING      0x80000000
 727 #define HAMMER2_DIRTYCHAIN_MASK         0x7FFFFFFF
 728
 729 #define HAMMER2_LWINPROG_WAITING        0x80000000
 730 #define HAMMER2_LWINPROG_MASK           0x7FFFFFFF
 731
 732 /*
 733  * Bulkscan
 734  */
 735 #define HAMMER2_BULK_ABORT      0x00000001
 736
 737 /*
 738  * Misc
 739  */
 740 #if defined(_KERNEL)
 741
 742 MALLOC_DECLARE(M_HAMMER2);
 743
 744 #define VTOI(vp)        ((hammer2_inode_t *)(vp)->v_data)
 745 #define ITOV(ip)        ((ip)->vp)
 746
 747 /*
 748  * Currently locked chains retain the locked buffer cache buffer for
 749  * indirect blocks, and indirect blocks can be one of two sizes.  The
 750  * device buffer has to match the case to avoid deadlocking recursive
 751  * chains that might otherwise try to access different offsets within
 752  * the same device buffer.
 753  */
 754 static __inline
 755 int
 756 hammer2_devblkradix(int radix)
 757 {
 758 #if 0
 759         if (radix <= HAMMER2_LBUFRADIX) {
 760                 return (HAMMER2_LBUFRADIX);
 761         } else {
 762                 return (HAMMER2_PBUFRADIX);
 763         }
 764 #endif
 765         return (HAMMER2_PBUFRADIX);
 766 }
 767
 768 /*
 769  * XXX almost time to remove this.  DIO uses PBUFSIZE exclusively now.
 770  */
 771 static __inline
 772 size_t
 773 hammer2_devblksize(size_t bytes)
 774 {
 775 #if 0
 776         if (bytes <= HAMMER2_LBUFSIZE) {
 777                 return(HAMMER2_LBUFSIZE);
 778         } else {
 779                 KKASSERT(bytes <= HAMMER2_PBUFSIZE &&
 780                          (bytes ^ (bytes - 1)) == ((bytes << 1) - 1));
 781                 return (HAMMER2_PBUFSIZE);
 782         }
 783 #endif
 784         return (HAMMER2_PBUFSIZE);
 785 }
 786
 787
 788 static __inline
 789 hammer2_pfsmount_t *
 790 MPTOPMP(struct mount *mp)
 791 {
 792         return ((hammer2_pfsmount_t *)mp->mnt_data);
 793 }
 794
 795 #define LOCKSTART       int __nlocks = curthread->td_locks
 796 #define LOCKENTER       (++curthread->td_locks)
 797 #define LOCKEXIT        (--curthread->td_locks)
 798 #define LOCKSTOP        KKASSERT(curthread->td_locks == __nlocks)
 799
 800 extern struct vop_ops hammer2_vnode_vops;
 801 extern struct vop_ops hammer2_spec_vops;
 802 extern struct vop_ops hammer2_fifo_vops;
 803
 804 extern int hammer2_debug;
 805 extern int hammer2_cluster_enable;
 806 extern int hammer2_hardlink_enable;
 807 extern int hammer2_flush_pipe;
 808 extern int hammer2_synchronous_flush;
 809 extern int hammer2_dio_count;
 810 extern long hammer2_limit_dirty_chains;
 811 extern long hammer2_iod_file_read;
 812 extern long hammer2_iod_meta_read;
 813 extern long hammer2_iod_indr_read;
 814 extern long hammer2_iod_fmap_read;
 815 extern long hammer2_iod_volu_read;
 816 extern long hammer2_iod_file_write;
 817 extern long hammer2_iod_meta_write;
 818 extern long hammer2_iod_indr_write;
 819 extern long hammer2_iod_fmap_write;
 820 extern long hammer2_iod_volu_write;
 821 extern long hammer2_ioa_file_read;
 822 extern long hammer2_ioa_meta_read;
 823 extern long hammer2_ioa_indr_read;
 824 extern long hammer2_ioa_fmap_read;
 825 extern long hammer2_ioa_volu_read;
 826 extern long hammer2_ioa_file_write;
 827 extern long hammer2_ioa_meta_write;
 828 extern long hammer2_ioa_indr_write;
 829 extern long hammer2_ioa_fmap_write;
 830 extern long hammer2_ioa_volu_write;
 831
 832 extern struct objcache *cache_buffer_read;
 833 extern struct objcache *cache_buffer_write;
 834
 835 extern int destroy;
 836 extern int write_thread_wakeup;
 837
 838 /*
 839  * hammer2_subr.c
 840  */
 841 #define hammer2_icrc32(buf, size)       iscsi_crc32((buf), (size))
 842 #define hammer2_icrc32c(buf, size, crc) iscsi_crc32_ext((buf), (size), (crc))
 843
 844 int hammer2_signal_check(time_t *timep);
 845 hammer2_cluster_t *hammer2_inode_lock_ex(hammer2_inode_t *ip);
 846 hammer2_cluster_t *hammer2_inode_lock_nex(hammer2_inode_t *ip, int how);
 847 hammer2_cluster_t *hammer2_inode_lock_sh(hammer2_inode_t *ip);
 848 void hammer2_inode_unlock_ex(hammer2_inode_t *ip, hammer2_cluster_t *chain);
 849 void hammer2_inode_unlock_sh(hammer2_inode_t *ip, hammer2_cluster_t *chain);
 850 hammer2_mtx_state_t hammer2_inode_lock_temp_release(hammer2_inode_t *ip);
 851 void hammer2_inode_lock_temp_restore(hammer2_inode_t *ip,
 852                         hammer2_mtx_state_t ostate);
 853 int hammer2_inode_lock_upgrade(hammer2_inode_t *ip);
 854 void hammer2_inode_lock_downgrade(hammer2_inode_t *ip, int);
 855
 856 void hammer2_mount_exlock(hammer2_mount_t *hmp);
 857 void hammer2_mount_shlock(hammer2_mount_t *hmp);
 858 void hammer2_mount_unlock(hammer2_mount_t *hmp);
 859
 860 int hammer2_get_dtype(const hammer2_inode_data_t *ipdata);
 861 int hammer2_get_vtype(const hammer2_inode_data_t *ipdata);
 862 u_int8_t hammer2_get_obj_type(enum vtype vtype);
 863 void hammer2_time_to_timespec(u_int64_t xtime, struct timespec *ts);
 864 u_int64_t hammer2_timespec_to_time(const struct timespec *ts);
 865 u_int32_t hammer2_to_unix_xid(const uuid_t *uuid);
 866 void hammer2_guid_to_uuid(uuid_t *uuid, u_int32_t guid);
 867 hammer2_xid_t hammer2_trans_newxid(hammer2_pfsmount_t *pmp);
 868 void hammer2_trans_manage_init(void);
 869
 870 hammer2_key_t hammer2_dirhash(const unsigned char *name, size_t len);
 871 int hammer2_getradix(size_t bytes);
 872
 873 int hammer2_calc_logical(hammer2_inode_t *ip, hammer2_off_t uoff,
 874                         hammer2_key_t *lbasep, hammer2_key_t *leofp);
 875 int hammer2_calc_physical(hammer2_inode_t *ip,
 876                         const hammer2_inode_data_t *ipdata,
 877                         hammer2_key_t lbase);
 878 void hammer2_update_time(uint64_t *timep);
 879 void hammer2_adjreadcounter(hammer2_blockref_t *bref, size_t bytes);
 880
 881 /*
 882  * hammer2_inode.c
 883  */
 884 struct vnode *hammer2_igetv(hammer2_inode_t *ip, hammer2_cluster_t *cparent,
 885                         int *errorp);
 886 void hammer2_inode_lock_nlinks(hammer2_inode_t *ip);
 887 void hammer2_inode_unlock_nlinks(hammer2_inode_t *ip);
 888 hammer2_inode_t *hammer2_inode_lookup(hammer2_pfsmount_t *pmp,
 889                         hammer2_tid_t inum);
 890 hammer2_inode_t *hammer2_inode_get(hammer2_pfsmount_t *pmp,
 891                         hammer2_inode_t *dip, hammer2_cluster_t *cluster);
 892 void hammer2_inode_free(hammer2_inode_t *ip);
 893 void hammer2_inode_ref(hammer2_inode_t *ip);
 894 void hammer2_inode_drop(hammer2_inode_t *ip);
 895 void hammer2_inode_repoint(hammer2_inode_t *ip, hammer2_inode_t *pip,
 896                         hammer2_cluster_t *cluster);
 897 void hammer2_run_unlinkq(hammer2_trans_t *trans, hammer2_pfsmount_t *pmp);
 898
 899 hammer2_inode_t *hammer2_inode_create(hammer2_trans_t *trans,
 900                         hammer2_inode_t *dip,
 901                         struct vattr *vap, struct ucred *cred,
 902                         const uint8_t *name, size_t name_len,
 903                         hammer2_cluster_t **clusterp, int *errorp);
 904 int hammer2_inode_connect(hammer2_trans_t *trans,
 905                         hammer2_cluster_t **clusterp, int hlink,
 906                         hammer2_inode_t *dip, hammer2_cluster_t *dcluster,
 907                         const uint8_t *name, size_t name_len,
 908                         hammer2_key_t key);
 909 hammer2_inode_t *hammer2_inode_common_parent(hammer2_inode_t *fdip,
 910                         hammer2_inode_t *tdip);
 911 void hammer2_inode_fsync(hammer2_trans_t *trans, hammer2_inode_t *ip,
 912                         hammer2_cluster_t *cparent);
 913 int hammer2_unlink_file(hammer2_trans_t *trans, hammer2_inode_t *dip,
 914                         const uint8_t *name, size_t name_len, int isdir,
 915                         int *hlinkp, struct nchandle *nch, int nlinks);
 916 int hammer2_hardlink_consolidate(hammer2_trans_t *trans,
 917                         hammer2_inode_t *ip, hammer2_cluster_t **clusterp,
 918                         hammer2_inode_t *cdip, hammer2_cluster_t *cdcluster,
 919                         int nlinks);
 920 int hammer2_hardlink_deconsolidate(hammer2_trans_t *trans, hammer2_inode_t *dip,
 921                         hammer2_chain_t **chainp, hammer2_chain_t **ochainp);
 922 int hammer2_hardlink_find(hammer2_inode_t *dip, hammer2_cluster_t **cparentp,
 923                         hammer2_cluster_t *cluster);
 924 int hammer2_parent_find(hammer2_cluster_t **cparentp,
 925                         hammer2_cluster_t *cluster);
 926 void hammer2_inode_install_hidden(hammer2_pfsmount_t *pmp);
 927
 928 /*
 929  * hammer2_chain.c
 930  */
 931 void hammer2_voldata_lock(hammer2_mount_t *hmp);
 932 void hammer2_voldata_unlock(hammer2_mount_t *hmp);
 933 void hammer2_voldata_modify(hammer2_mount_t *hmp);
 934 hammer2_chain_t *hammer2_chain_alloc(hammer2_mount_t *hmp,
 935                                 hammer2_pfsmount_t *pmp,
 936                                 hammer2_trans_t *trans,
 937                                 hammer2_blockref_t *bref);
 938 void hammer2_chain_core_alloc(hammer2_trans_t *trans, hammer2_chain_t *chain);
 939 void hammer2_chain_ref(hammer2_chain_t *chain);
 940 void hammer2_chain_drop(hammer2_chain_t *chain);
 941 int hammer2_chain_lock(hammer2_chain_t *chain, int how);
 942 const hammer2_media_data_t *hammer2_chain_rdata(hammer2_chain_t *chain);
 943 hammer2_media_data_t *hammer2_chain_wdata(hammer2_chain_t *chain);
 944
 945 /*
 946  * hammer2_cluster.c
 947  */
 948 int hammer2_cluster_isunlinked(hammer2_cluster_t *cluster);
 949 void hammer2_cluster_load_async(hammer2_cluster_t *cluster,
 950                                 void (*callback)(hammer2_iocb_t *iocb),
 951                                 void *ptr);
 952 void hammer2_chain_moved(hammer2_chain_t *chain);
 953 void hammer2_chain_modify(hammer2_trans_t *trans,
 954                                 hammer2_chain_t *chain, int flags);
 955 void hammer2_chain_resize(hammer2_trans_t *trans, hammer2_inode_t *ip,
 956                                 hammer2_chain_t *parent,
 957                                 hammer2_chain_t *chain,
 958                                 int nradix, int flags);
 959 void hammer2_chain_unlock(hammer2_chain_t *chain);
 960 void hammer2_chain_wait(hammer2_chain_t *chain);
 961 hammer2_chain_t *hammer2_chain_get(hammer2_chain_t *parent, int generation,
 962                                 hammer2_blockref_t *bref);
 963 hammer2_chain_t *hammer2_chain_lookup_init(hammer2_chain_t *parent, int flags);
 964 void hammer2_chain_lookup_done(hammer2_chain_t *parent);
 965 hammer2_chain_t *hammer2_chain_lookup(hammer2_chain_t **parentp,
 966                                 hammer2_key_t *key_nextp,
 967                                 hammer2_key_t key_beg, hammer2_key_t key_end,
 968                                 int *cache_indexp, int flags, int *ddflagp);
 969 hammer2_chain_t *hammer2_chain_next(hammer2_chain_t **parentp,
 970                                 hammer2_chain_t *chain,
 971                                 hammer2_key_t *key_nextp,
 972                                 hammer2_key_t key_beg, hammer2_key_t key_end,
 973                                 int *cache_indexp, int flags);
 974 hammer2_chain_t *hammer2_chain_scan(hammer2_chain_t *parent,
 975                                 hammer2_chain_t *chain,
 976                                 int *cache_indexp, int flags);
 977
 978 int hammer2_chain_create(hammer2_trans_t *trans, hammer2_chain_t **parentp,
 979                                 hammer2_chain_t **chainp,
 980                                 hammer2_pfsmount_t *pmp,
 981                                 hammer2_key_t key, int keybits,
 982                                 int type, size_t bytes, int flags);
 983 void hammer2_chain_rename(hammer2_trans_t *trans, hammer2_blockref_t *bref,
 984                                 hammer2_chain_t **parentp,
 985                                 hammer2_chain_t *chain, int flags);
 986 int hammer2_chain_snapshot(hammer2_trans_t *trans, hammer2_chain_t **chainp,
 987                                 hammer2_ioc_pfs_t *pfs);
 988 void hammer2_chain_delete(hammer2_trans_t *trans, hammer2_chain_t *parent,
 989                                 hammer2_chain_t *chain, int flags);
 990 void hammer2_chain_delete_duplicate(hammer2_trans_t *trans,
 991                                 hammer2_chain_t **chainp, int flags);
 992 void hammer2_flush(hammer2_trans_t *trans, hammer2_chain_t *chain);
 993 void hammer2_chain_commit(hammer2_trans_t *trans, hammer2_chain_t *chain);
 994 void hammer2_chain_setflush(hammer2_trans_t *trans, hammer2_chain_t *chain);
 995 void hammer2_chain_countbrefs(hammer2_chain_t *chain,
 996                                 hammer2_blockref_t *base, int count);
 997
 998 void hammer2_chain_setcheck(hammer2_chain_t *chain, void *bdata);
 999 int hammer2_chain_testcheck(hammer2_chain_t *chain, void *bdata);
1000
1001
1002 void hammer2_pfs_memory_wait(hammer2_pfsmount_t *pmp);
1003 void hammer2_pfs_memory_inc(hammer2_pfsmount_t *pmp);
1004 void hammer2_pfs_memory_wakeup(hammer2_pfsmount_t *pmp);
1005
1006 void hammer2_base_delete(hammer2_trans_t *trans, hammer2_chain_t *chain,
1007                                 hammer2_blockref_t *base, int count,
1008                                 int *cache_indexp, hammer2_chain_t *child);
1009 void hammer2_base_insert(hammer2_trans_t *trans, hammer2_chain_t *chain,
1010                                 hammer2_blockref_t *base, int count,
1011                                 int *cache_indexp, hammer2_chain_t *child);
1012
1013 /*
1014  * hammer2_trans.c
1015  */
1016 void hammer2_trans_init(hammer2_trans_t *trans, hammer2_pfsmount_t *pmp,
1017                                 int flags);
1018 void hammer2_trans_spmp(hammer2_trans_t *trans, hammer2_pfsmount_t *pmp);
1019 void hammer2_trans_done(hammer2_trans_t *trans);
1020
1021 /*
1022  * hammer2_ioctl.c
1023  */
1024 int hammer2_ioctl(hammer2_inode_t *ip, u_long com, void *data,
1025                                 int fflag, struct ucred *cred);
1026
1027 /*
1028  * hammer2_io.c
1029  */
1030 void hammer2_io_putblk(hammer2_io_t **diop);
1031 void hammer2_io_cleanup(hammer2_mount_t *hmp, struct hammer2_io_tree *tree);
1032 char *hammer2_io_data(hammer2_io_t *dio, off_t lbase);
1033 void hammer2_io_getblk(hammer2_mount_t *hmp, off_t lbase, int lsize,
1034                                 hammer2_iocb_t *iocb);
1035 void hammer2_io_complete(hammer2_iocb_t *iocb);
1036 void hammer2_io_callback(struct bio *bio);
1037 void hammer2_iocb_wait(hammer2_iocb_t *iocb);
1038 int hammer2_io_new(hammer2_mount_t *hmp, off_t lbase, int lsize,
1039                                 hammer2_io_t **diop);
1040 int hammer2_io_newnz(hammer2_mount_t *hmp, off_t lbase, int lsize,
1041                                 hammer2_io_t **diop);
1042 int hammer2_io_newq(hammer2_mount_t *hmp, off_t lbase, int lsize,
1043                                 hammer2_io_t **diop);
1044 int hammer2_io_bread(hammer2_mount_t *hmp, off_t lbase, int lsize,
1045                                 hammer2_io_t **diop);
1046 void hammer2_io_bawrite(hammer2_io_t **diop);
1047 void hammer2_io_bdwrite(hammer2_io_t **diop);
1048 int hammer2_io_bwrite(hammer2_io_t **diop);
1049 int hammer2_io_isdirty(hammer2_io_t *dio);
1050 void hammer2_io_setdirty(hammer2_io_t *dio);
1051 void hammer2_io_setinval(hammer2_io_t *dio, u_int bytes);
1052 void hammer2_io_brelse(hammer2_io_t **diop);
1053 void hammer2_io_bqrelse(hammer2_io_t **diop);
1054
1055 /*
1056  * hammer2_msgops.c
1057  */
1058 int hammer2_msg_dbg_rcvmsg(kdmsg_msg_t *msg);
1059 int hammer2_msg_adhoc_input(kdmsg_msg_t *msg);
1060
1061 /*
1062  * hammer2_vfsops.c
1063  */
1064 void hammer2_clusterctl_wakeup(kdmsg_iocom_t *iocom);
1065 void hammer2_volconf_update(hammer2_mount_t *hmp, int index);
1066 void hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx);
1067 void hammer2_bioq_sync(hammer2_pfsmount_t *pmp);
1068 int hammer2_vfs_sync(struct mount *mp, int waitflags);
1069 void hammer2_lwinprog_ref(hammer2_pfsmount_t *pmp);
1070 void hammer2_lwinprog_drop(hammer2_pfsmount_t *pmp);
1071 void hammer2_lwinprog_wait(hammer2_pfsmount_t *pmp);
1072
1073 /*
1074  * hammer2_freemap.c
1075  */
1076 int hammer2_freemap_alloc(hammer2_trans_t *trans, hammer2_chain_t *chain,
1077                                 size_t bytes);
1078 void hammer2_freemap_adjust(hammer2_trans_t *trans, hammer2_mount_t *hmp,
1079                                 hammer2_blockref_t *bref, int how);
1080
1081 /*
1082  * hammer2_cluster.c
1083  */
1084 int hammer2_cluster_need_resize(hammer2_cluster_t *cluster, int bytes);
1085 uint8_t hammer2_cluster_type(hammer2_cluster_t *cluster);
1086 const hammer2_media_data_t *hammer2_cluster_rdata(hammer2_cluster_t *cluster);
1087 hammer2_media_data_t *hammer2_cluster_wdata(hammer2_cluster_t *cluster);
1088 hammer2_cluster_t *hammer2_cluster_from_chain(hammer2_chain_t *chain);
1089 int hammer2_cluster_modified(hammer2_cluster_t *cluster);
1090 int hammer2_cluster_duplicated(hammer2_cluster_t *cluster);
1091 void hammer2_cluster_set_chainflags(hammer2_cluster_t *cluster, uint32_t flags);
1092 void hammer2_cluster_clr_chainflags(hammer2_cluster_t *cluster, uint32_t flags);
1093 void hammer2_cluster_bref(hammer2_cluster_t *cluster, hammer2_blockref_t *bref);
1094 void hammer2_cluster_setflush(hammer2_trans_t *trans,
1095                         hammer2_cluster_t *cluster);
1096 void hammer2_cluster_setmethod_check(hammer2_trans_t *trans,
1097                         hammer2_cluster_t *cluster, int check_algo);
1098 hammer2_cluster_t *hammer2_cluster_alloc(hammer2_pfsmount_t *pmp,
1099                         hammer2_trans_t *trans,
1100                         hammer2_blockref_t *bref);
1101 void hammer2_cluster_ref(hammer2_cluster_t *cluster);
1102 void hammer2_cluster_drop(hammer2_cluster_t *cluster);
1103 void hammer2_cluster_wait(hammer2_cluster_t *cluster);
1104 int hammer2_cluster_lock(hammer2_cluster_t *cluster, int how);
1105 void hammer2_cluster_replace(hammer2_cluster_t *dst, hammer2_cluster_t *src);
1106 void hammer2_cluster_replace_locked(hammer2_cluster_t *dst,
1107                         hammer2_cluster_t *src);
1108 hammer2_cluster_t *hammer2_cluster_copy(hammer2_cluster_t *ocluster);
1109 void hammer2_cluster_unlock(hammer2_cluster_t *cluster);
1110 void hammer2_cluster_resize(hammer2_trans_t *trans, hammer2_inode_t *ip,
1111                         hammer2_cluster_t *cparent, hammer2_cluster_t *cluster,
1112                         int nradix, int flags);
1113 hammer2_inode_data_t *hammer2_cluster_modify_ip(hammer2_trans_t *trans,
1114                         hammer2_inode_t *ip, hammer2_cluster_t *cluster,
1115                         int flags);
1116 void hammer2_cluster_modify(hammer2_trans_t *trans, hammer2_cluster_t *cluster,
1117                         int flags);
1118 void hammer2_cluster_modsync(hammer2_cluster_t *cluster);
1119 hammer2_cluster_t *hammer2_cluster_lookup_init(hammer2_cluster_t *cparent,
1120                         int flags);
1121 void hammer2_cluster_lookup_done(hammer2_cluster_t *cparent);
1122 hammer2_cluster_t *hammer2_cluster_lookup(hammer2_cluster_t *cparent,
1123                         hammer2_key_t *key_nextp,
1124                         hammer2_key_t key_beg, hammer2_key_t key_end,
1125                         int flags, int *ddflagp);
1126 hammer2_cluster_t *hammer2_cluster_next(hammer2_cluster_t *cparent,
1127                         hammer2_cluster_t *cluster,
1128                         hammer2_key_t *key_nextp,
1129                         hammer2_key_t key_beg, hammer2_key_t key_end,
1130                         int flags);
1131 hammer2_cluster_t *hammer2_cluster_scan(hammer2_cluster_t *cparent,
1132                         hammer2_cluster_t *cluster, int flags);
1133 int hammer2_cluster_create(hammer2_trans_t *trans, hammer2_cluster_t *cparent,
1134                         hammer2_cluster_t **clusterp,
1135                         hammer2_key_t key, int keybits,
1136                         int type, size_t bytes, int flags);
1137 void hammer2_cluster_rename(hammer2_trans_t *trans, hammer2_blockref_t *bref,
1138                         hammer2_cluster_t *cparent, hammer2_cluster_t *cluster,
1139                         int flags);
1140 void hammer2_cluster_delete(hammer2_trans_t *trans, hammer2_cluster_t *pcluster,
1141                         hammer2_cluster_t *cluster, int flags);
1142 int hammer2_cluster_snapshot(hammer2_trans_t *trans,
1143                         hammer2_cluster_t *ocluster, hammer2_ioc_pfs_t *pfs);
1144 hammer2_cluster_t *hammer2_cluster_parent(hammer2_cluster_t *cluster);
1145
1146 int hammer2_bulk_scan(hammer2_trans_t *trans, hammer2_chain_t *parent,
1147                         int (*func)(hammer2_chain_t *chain, void *info),
1148                         void *info);
1149 int hammer2_bulkfree_pass(hammer2_mount_t *hmp,
1150                         struct hammer2_ioc_bulkfree *bfi);
1151
1152 /*
1153  * hammer2_iocom.c
1154  */
1155 void hammer2_iocom_init(hammer2_mount_t *hmp);
1156 void hammer2_iocom_uninit(hammer2_mount_t *hmp);
1157 void hammer2_cluster_reconnect(hammer2_mount_t *hmp, struct file *fp);
1158
1159 #endif /* !_KERNEL */
1160 #endif /* !_VFS_HAMMER2_HAMMER2_H_ */