sys/vfs/hammer2/hammer2.h

   1 /*
   2  * Copyright (c) 2011-2017 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 #ifdef _KERNEL
  66 #include <sys/param.h>
  67 #endif
  68 #include <sys/types.h>
  69 #ifdef _KERNEL
  70 #include <sys/kernel.h>
  71 #endif
  72 #include <sys/conf.h>
  73 #ifdef _KERNEL
  74 #include <sys/systm.h>
  75 #endif
  76 #include <sys/tree.h>
  77 #include <sys/malloc.h>
  78 #include <sys/mount.h>
  79 #include <sys/vnode.h>
  80 #include <sys/proc.h>
  81 #include <sys/mountctl.h>
  82 #include <sys/priv.h>
  83 #include <sys/stat.h>
  84 #include <sys/thread.h>
  85 #include <sys/globaldata.h>
  86 #include <sys/lockf.h>
  87 #include <sys/buf.h>
  88 #include <sys/queue.h>
  89 #include <sys/limits.h>
  90 #include <sys/dmsg.h>
  91 #include <sys/mutex.h>
  92 #ifdef _KERNEL
  93 #include <sys/kern_syscall.h>
  94 #endif
  95
  96 #ifdef _KERNEL
  97 #include <sys/signal2.h>
  98 #include <sys/buf2.h>
  99 #include <sys/mutex2.h>
 100 #include <sys/thread2.h>
 101 #endif
 102
 103 #include "hammer2_xxhash.h"
 104 #include "hammer2_disk.h"
 105 #include "hammer2_mount.h"
 106 #include "hammer2_ioctl.h"
 107
 108 struct hammer2_io;
 109 struct hammer2_chain;
 110 struct hammer2_cluster;
 111 struct hammer2_inode;
 112 struct hammer2_dev;
 113 struct hammer2_pfs;
 114 struct hammer2_span;
 115 struct hammer2_state;
 116 struct hammer2_msg;
 117 struct hammer2_thread;
 118 union hammer2_xop;
 119
 120 /*
 121  * Mutex and lock shims.  Hammer2 requires support for asynchronous and
 122  * abortable locks, and both exclusive and shared spinlocks.  Normal
 123  * synchronous non-abortable locks can be substituted for spinlocks.
 124  */
 125 typedef mtx_t                           hammer2_mtx_t;
 126 typedef mtx_link_t                      hammer2_mtx_link_t;
 127 typedef mtx_state_t                     hammer2_mtx_state_t;
 128
 129 typedef struct spinlock                 hammer2_spin_t;
 130
 131 #define hammer2_mtx_ex                  mtx_lock_ex_quick
 132 #define hammer2_mtx_sh                  mtx_lock_sh_quick
 133 #define hammer2_mtx_sh_again            mtx_lock_sh_again
 134 #define hammer2_mtx_unlock              mtx_unlock
 135 #define hammer2_mtx_downgrade           mtx_downgrade
 136 #define hammer2_mtx_owned               mtx_owned
 137 #define hammer2_mtx_init                mtx_init
 138 #define hammer2_mtx_temp_release        mtx_lock_temp_release
 139 #define hammer2_mtx_temp_restore        mtx_lock_temp_restore
 140 #define hammer2_mtx_refs                mtx_lockrefs
 141
 142 #define hammer2_spin_init               spin_init
 143 #define hammer2_spin_sh                 spin_lock_shared
 144 #define hammer2_spin_ex                 spin_lock
 145 #define hammer2_spin_unsh               spin_unlock_shared
 146 #define hammer2_spin_unex               spin_unlock
 147
 148 TAILQ_HEAD(hammer2_xop_list, hammer2_xop_head);
 149 TAILQ_HEAD(hammer2_chain_list, hammer2_chain);
 150
 151 typedef struct hammer2_xop_list hammer2_xop_list_t;
 152
 153 #ifdef _KERNEL
 154 /*
 155  * General lock support
 156  */
 157 static __inline
 158 int
 159 hammer2_mtx_upgrade_try(hammer2_mtx_t *mtx)
 160 {
 161         return mtx_upgrade_try(mtx);
 162 }
 163
 164 #endif
 165
 166 /*
 167  * The xid tracks internal transactional updates.
 168  *
 169  * XXX fix-me, really needs to be 64-bits
 170  */
 171 typedef uint32_t hammer2_xid_t;
 172
 173 #define HAMMER2_XID_MIN                 0x00000000U
 174 #define HAMMER2_XID_MAX                 0x7FFFFFFFU
 175
 176 #define HAMMER2_LIMIT_DIRTY_CHAINS      (65536)
 177
 178 /*
 179  * The chain structure tracks a portion of the media topology from the
 180  * root (volume) down.  Chains represent volumes, inodes, indirect blocks,
 181  * data blocks, and freemap nodes and leafs.
 182  *
 183  * The chain structure utilizes a simple singly-homed topology and the
 184  * chain's in-memory topology will move around as the chains do, due mainly
 185  * to renames and indirect block creation.
 186  *
 187  * Block Table Updates
 188  *
 189  *      Block table updates for insertions and updates are delayed until the
 190  *      flush.  This allows us to avoid having to modify the parent chain
 191  *      all the way to the root.
 192  *
 193  *      Block table deletions are performed immediately (modifying the parent
 194  *      in the process) because the flush code uses the chain structure to
 195  *      track delayed updates and the chain will be (likely) gone or moved to
 196  *      another location in the topology after a deletion.
 197  *
 198  *      A prior iteration of the code tried to keep the relationship intact
 199  *      on deletes by doing a delete-duplicate operation on the chain, but
 200  *      it added way too much complexity to the codebase.
 201  *
 202  * Flush Synchronization
 203  *
 204  *      The flush code must flush modified chains bottom-up.  Because chain
 205  *      structures can shift around and are NOT topologically stable,
 206  *      modified chains are independently indexed for the flush.  As the flush
 207  *      runs it modifies (or further modifies) and updates the parents,
 208  *      propagating the flush all the way to the volume root.
 209  *
 210  *      Modifying front-end operations can occur during a flush but will block
 211  *      in two cases: (1) when the front-end tries to operate on the inode
 212  *      currently in the midst of being flushed and (2) if the front-end
 213  *      crosses an inode currently being flushed (such as during a rename).
 214  *      So, for example, if you rename directory "x" to "a/b/c/d/e/f/g/x" and
 215  *      the flusher is currently working on "a/b/c", the rename will block
 216  *      temporarily in order to ensure that "x" exists in one place or the
 217  *      other.
 218  *
 219  *      Meta-data statistics are updated by the flusher.  The front-end will
 220  *      make estimates but meta-data must be fully synchronized only during a
 221  *      flush in order to ensure that it remains correct across a crash.
 222  *
 223  *      Multiple flush synchronizations can theoretically be in-flight at the
 224  *      same time but the implementation is not coded to handle the case and
 225  *      currently serializes them.
 226  *
 227  * Snapshots:
 228  *
 229  *      Snapshots currently require the subdirectory tree being snapshotted
 230  *      to be flushed.  The snapshot then creates a new super-root inode which
 231  *      copies the flushed blockdata of the directory or file that was
 232  *      snapshotted.
 233  *
 234  * RBTREE NOTES:
 235  *
 236  *      - Note that the radix tree runs in powers of 2 only so sub-trees
 237  *        cannot straddle edges.
 238  */
 239 RB_HEAD(hammer2_chain_tree, hammer2_chain);
 240 TAILQ_HEAD(h2_flush_list, hammer2_chain);
 241 TAILQ_HEAD(h2_core_list, hammer2_chain);
 242
 243 #define CHAIN_CORE_DELETE_BMAP_ENTRIES  \
 244         (HAMMER2_PBUFSIZE / sizeof(hammer2_blockref_t) / sizeof(uint32_t))
 245
 246 /*
 247  * Core topology for chain (embedded in chain).  Protected by a spinlock.
 248  */
 249 struct hammer2_chain_core {
 250         hammer2_spin_t  spin;
 251         struct hammer2_chain_tree rbtree; /* sub-chains */
 252         int             live_zero;      /* blockref array opt */
 253         u_int           live_count;     /* live (not deleted) chains in tree */
 254         u_int           chain_count;    /* live + deleted chains under core */
 255         int             generation;     /* generation number (inserts only) */
 256 };
 257
 258 typedef struct hammer2_chain_core hammer2_chain_core_t;
 259
 260 RB_HEAD(hammer2_io_tree, hammer2_io);
 261
 262 /*
 263  * DIO - Management structure wrapping system buffer cache.
 264  *
 265  * HAMMER2 uses an I/O abstraction that allows it to cache and manipulate
 266  * fixed-sized filesystem buffers frontend by variable-sized hammer2_chain
 267  * structures.
 268  */
 269 struct hammer2_io {
 270         RB_ENTRY(hammer2_io) rbnode;    /* indexed by device offset */
 271         struct hammer2_dev *hmp;
 272         struct buf      *bp;
 273         off_t           pbase;
 274         uint64_t        refs;
 275         int             psize;
 276         int             act;            /* activity */
 277         int             btype;          /* approximate BREF_TYPE_* */
 278         int             ticks;
 279         int             error;
 280         int             unused01;
 281         uint64_t        dedup_valid;    /* valid for dedup operation */
 282         uint64_t        dedup_alloc;    /* allocated / de-dupable */
 283 };
 284
 285 typedef struct hammer2_io hammer2_io_t;
 286
 287 #define HAMMER2_DIO_INPROG      0x8000000000000000LLU   /* bio in progress */
 288 #define HAMMER2_DIO_GOOD        0x4000000000000000LLU   /* dio->bp is stable */
 289 #define HAMMER2_DIO_WAITING     0x2000000000000000LLU   /* wait on INPROG */
 290 #define HAMMER2_DIO_DIRTY       0x1000000000000000LLU   /* flush last drop */
 291
 292 #define HAMMER2_DIO_MASK        0x00FFFFFFFFFFFFFFLLU
 293
 294 /*
 295  * Primary chain structure keeps track of the topology in-memory.
 296  */
 297 struct hammer2_chain {
 298         hammer2_mtx_t           lock;
 299         hammer2_chain_core_t    core;
 300         RB_ENTRY(hammer2_chain) rbnode;         /* live chain(s) */
 301         hammer2_blockref_t      bref;
 302         struct hammer2_chain    *parent;
 303         struct hammer2_state    *state;         /* if active cache msg */
 304         struct hammer2_dev      *hmp;
 305         struct hammer2_pfs      *pmp;           /* A PFS or super-root (spmp) */
 306
 307         hammer2_io_t    *dio;                   /* physical data buffer */
 308         u_int           bytes;                  /* physical data size */
 309         u_int           flags;
 310         u_int           refs;
 311         u_int           lockcnt;
 312         int             error;                  /* on-lock data error state */
 313         int             cache_index;            /* heur speeds up lookup */
 314
 315         hammer2_media_data_t *data;             /* data pointer shortcut */
 316         TAILQ_ENTRY(hammer2_chain) flush_node;  /* flush list */
 317         TAILQ_ENTRY(hammer2_chain) lru_node;    /* 0-refs LRU */
 318 };
 319
 320 typedef struct hammer2_chain hammer2_chain_t;
 321
 322 int hammer2_chain_cmp(hammer2_chain_t *chain1, hammer2_chain_t *chain2);
 323 RB_PROTOTYPE(hammer2_chain_tree, hammer2_chain, rbnode, hammer2_chain_cmp);
 324
 325 /*
 326  * Special notes on flags:
 327  *
 328  * INITIAL      - This flag allows a chain to be created and for storage to
 329  *                be allocated without having to immediately instantiate the
 330  *                related buffer.  The data is assumed to be all-zeros.  It
 331  *                is primarily used for indirect blocks.
 332  *
 333  * MODIFIED     - The chain's media data has been modified.  Prevents chain
 334  *                free on lastdrop if still in the topology.
 335  *
 336  * UPDATE       - Chain might not be modified but parent blocktable needs
 337  *                an update.  Prevents chain free on lastdrop if still in
 338  *                the topology.
 339  *
 340  * FICTITIOUS   - Faked chain as a placeholder for an error condition.  This
 341  *                chain is unsuitable for I/O.
 342  *
 343  * BMAPPED      - Indicates that the chain is present in the parent blockmap.
 344  *
 345  * BMAPUPD      - Indicates that the chain is present but needs to be updated
 346  *                in the parent blockmap.
 347  */
 348 #define HAMMER2_CHAIN_MODIFIED          0x00000001      /* dirty chain data */
 349 #define HAMMER2_CHAIN_ALLOCATED         0x00000002      /* kmalloc'd chain */
 350 #define HAMMER2_CHAIN_DESTROY           0x00000004
 351 #define HAMMER2_CHAIN_DEDUPABLE         0x00000008      /* registered w/dedup */
 352 #define HAMMER2_CHAIN_DELETED           0x00000010      /* deleted chain */
 353 #define HAMMER2_CHAIN_INITIAL           0x00000020      /* initial create */
 354 #define HAMMER2_CHAIN_UPDATE            0x00000040      /* need parent update */
 355 #define HAMMER2_CHAIN_DEFERRED          0x00000080      /* flush depth defer */
 356 #define HAMMER2_CHAIN_TESTEDGOOD        0x00000100      /* crc tested good */
 357 #define HAMMER2_CHAIN_ONFLUSH           0x00000200      /* on a flush list */
 358 #define HAMMER2_CHAIN_FICTITIOUS        0x00000400      /* unsuitable for I/O */
 359 #define HAMMER2_CHAIN_VOLUMESYNC        0x00000800      /* needs volume sync */
 360 #define HAMMER2_CHAIN_DELAYED           0x00001000      /* delayed flush */
 361 #define HAMMER2_CHAIN_COUNTEDBREFS      0x00002000      /* block table stats */
 362 #define HAMMER2_CHAIN_ONRBTREE          0x00004000      /* on parent RB tree */
 363 #define HAMMER2_CHAIN_ONLRU             0x00008000      /* on LRU list */
 364 #define HAMMER2_CHAIN_EMBEDDED          0x00010000      /* embedded data */
 365 #define HAMMER2_CHAIN_RELEASE           0x00020000      /* don't keep around */
 366 #define HAMMER2_CHAIN_BMAPPED           0x00040000      /* present in blkmap */
 367 #define HAMMER2_CHAIN_BMAPUPD           0x00080000      /* +needs updating */
 368 #define HAMMER2_CHAIN_IOINPROG          0x00100000      /* I/O interlock */
 369 #define HAMMER2_CHAIN_IOSIGNAL          0x00200000      /* I/O interlock */
 370 #define HAMMER2_CHAIN_PFSBOUNDARY       0x00400000      /* super->pfs inode */
 371 #define HAMMER2_CHAIN_HINT_LEAF_COUNT   0x00800000      /* redo leaf count */
 372
 373 #define HAMMER2_CHAIN_FLUSH_MASK        (HAMMER2_CHAIN_MODIFIED |       \
 374                                          HAMMER2_CHAIN_UPDATE |         \
 375                                          HAMMER2_CHAIN_ONFLUSH |        \
 376                                          HAMMER2_CHAIN_DESTROY)
 377
 378 /*
 379  * Hammer2 error codes, used by chain->error and cluster->error.  The error
 380  * code is typically set on-lock unless no I/O was requested, and set on
 381  * I/O otherwise.  If set for a cluster it generally means that the cluster
 382  * code could not find a valid copy to present.
 383  *
 384  * All H2 error codes are flags and can be accumulated by ORing them
 385  * together.
 386  *
 387  * IO           - An I/O error occurred
 388  * CHECK        - I/O succeeded but did not match the check code
 389  * INCOMPLETE   - A cluster is not complete enough to use, or
 390  *                a chain cannot be loaded because its parent has an error.
 391  *
 392  * NOTE: API allows callers to check zero/non-zero to determine if an error
 393  *       condition exists.
 394  *
 395  * NOTE: Chain's data field is usually NULL on an IO error but not necessarily
 396  *       NULL on other errors.  Check chain->error, not chain->data.
 397  */
 398 #define HAMMER2_ERROR_NONE              0       /* no error (must be 0) */
 399 #define HAMMER2_ERROR_EIO               0x00000001      /* device I/O error */
 400 #define HAMMER2_ERROR_CHECK             0x00000002      /* check code error */
 401 #define HAMMER2_ERROR_INCOMPLETE        0x00000004      /* incomplete cluster */
 402 #define HAMMER2_ERROR_DEPTH             0x00000008      /* tmp depth limit */
 403 #define HAMMER2_ERROR_BADBREF           0x00000010      /* illegal bref */
 404 #define HAMMER2_ERROR_ENOSPC            0x00000020      /* allocation failure */
 405 #define HAMMER2_ERROR_ENOENT            0x00000040      /* entry not found */
 406 #define HAMMER2_ERROR_ENOTEMPTY         0x00000080      /* dir not empty */
 407 #define HAMMER2_ERROR_EAGAIN            0x00000100      /* retry */
 408 #define HAMMER2_ERROR_ENOTDIR           0x00000200      /* not directory */
 409 #define HAMMER2_ERROR_EISDIR            0x00000400      /* is directory */
 410 #define HAMMER2_ERROR_EINPROGRESS       0x00000800      /* already running */
 411 #define HAMMER2_ERROR_ABORTED           0x00001000      /* aborted operation */
 412 #define HAMMER2_ERROR_EOF               0x00002000      /* end of scan */
 413 #define HAMMER2_ERROR_EINVAL            0x00004000      /* catch-all */
 414 #define HAMMER2_ERROR_EEXIST            0x00008000      /* entry exists */
 415 #define HAMMER2_ERROR_EDEADLK           0x00010000
 416 #define HAMMER2_ERROR_ESRCH             0x00020000
 417 #define HAMMER2_ERROR_ETIMEDOUT         0x00040000
 418
 419 /*
 420  * Flags passed to hammer2_chain_lookup() and hammer2_chain_next()
 421  *
 422  * NOTES:
 423  *      NOLOCK      - Input and output chains are referenced only and not
 424  *                    locked.  Output chain might be temporarily locked
 425  *                    internally.
 426  *
 427  *      NODATA      - Asks that the chain->data not be resolved in order
 428  *                    to avoid I/O.
 429  *
 430  *      NODIRECT    - Prevents a lookup of offset 0 in an inode from returning
 431  *                    the inode itself if the inode is in DIRECTDATA mode
 432  *                    (i.e. file is <= 512 bytes).  Used by the synchronization
 433  *                    code to prevent confusion.
 434  *
 435  *      SHARED      - The input chain is expected to be locked shared,
 436  *                    and the output chain is locked shared.
 437  *
 438  *      MATCHIND    - Allows an indirect block / freemap node to be returned
 439  *                    when the passed key range matches the radix.  Remember
 440  *                    that key_end is inclusive (e.g. {0x000,0xFFF},
 441  *                    not {0x000,0x1000}).
 442  *
 443  *                    (Cannot be used for remote or cluster ops).
 444  *
 445  *      ALLNODES    - Allows NULL focus.
 446  *
 447  *      ALWAYS      - Always resolve the data.  If ALWAYS and NODATA are both
 448  *                    missing, bulk file data is not resolved but inodes and
 449  *                    other meta-data will.
 450  *
 451  *      NOUNLOCK    - Used by hammer2_chain_next() to leave the lock on
 452  *                    the input chain intact.  The chain is still dropped.
 453  *                    This allows the caller to add a reference to the chain
 454  *                    and retain it in a locked state (used by the
 455  *                    XOP/feed/collect code).
 456  */
 457 #define HAMMER2_LOOKUP_NOLOCK           0x00000001      /* ref only */
 458 #define HAMMER2_LOOKUP_NODATA           0x00000002      /* data left NULL */
 459 #define HAMMER2_LOOKUP_NODIRECT         0x00000004      /* no offset=0 DD */
 460 #define HAMMER2_LOOKUP_SHARED           0x00000100
 461 #define HAMMER2_LOOKUP_MATCHIND         0x00000200      /* return all chains */
 462 #define HAMMER2_LOOKUP_ALLNODES         0x00000400      /* allow NULL focus */
 463 #define HAMMER2_LOOKUP_ALWAYS           0x00000800      /* resolve data */
 464 #define HAMMER2_LOOKUP_NOUNLOCK         0x00001000      /* leave lock intact */
 465
 466 /*
 467  * Flags passed to hammer2_chain_modify() and hammer2_chain_resize()
 468  *
 469  * NOTE: OPTDATA allows us to avoid instantiating buffers for INDIRECT
 470  *       blocks in the INITIAL-create state.
 471  */
 472 #define HAMMER2_MODIFY_OPTDATA          0x00000002      /* data can be NULL */
 473 #define HAMMER2_MODIFY_NO_MODIFY_TID    0x00000004
 474 #define HAMMER2_MODIFY_UNUSED0008       0x00000008
 475
 476 /*
 477  * Flags passed to hammer2_chain_lock()
 478  *
 479  * NOTE: RDONLY is set to optimize cluster operations when *no* modifications
 480  *       will be made to either the cluster being locked or any underlying
 481  *       cluster.  It allows the cluster to lock and access data for a subset
 482  *       of available nodes instead of all available nodes.
 483  */
 484 #define HAMMER2_RESOLVE_NEVER           1
 485 #define HAMMER2_RESOLVE_MAYBE           2
 486 #define HAMMER2_RESOLVE_ALWAYS          3
 487 #define HAMMER2_RESOLVE_MASK            0x0F
 488
 489 #define HAMMER2_RESOLVE_SHARED          0x10    /* request shared lock */
 490 #define HAMMER2_RESOLVE_LOCKAGAIN       0x20    /* another shared lock */
 491 #define HAMMER2_RESOLVE_RDONLY          0x40    /* higher level op flag */
 492
 493 /*
 494  * Flags passed to hammer2_chain_delete()
 495  */
 496 #define HAMMER2_DELETE_PERMANENT        0x0001
 497
 498 /*
 499  * Flags passed to hammer2_chain_insert() or hammer2_chain_rename()
 500  * or hammer2_chain_create().
 501  */
 502 #define HAMMER2_INSERT_PFSROOT          0x0004
 503 #define HAMMER2_INSERT_SAMEPARENT       0x0008
 504
 505 /*
 506  * Flags passed to hammer2_chain_delete_duplicate()
 507  */
 508 #define HAMMER2_DELDUP_RECORE           0x0001
 509
 510 /*
 511  * Cluster different types of storage together for allocations
 512  */
 513 #define HAMMER2_FREECACHE_INODE         0
 514 #define HAMMER2_FREECACHE_INDIR         1
 515 #define HAMMER2_FREECACHE_DATA          2
 516 #define HAMMER2_FREECACHE_UNUSED3       3
 517 #define HAMMER2_FREECACHE_TYPES         4
 518
 519 /*
 520  * hammer2_freemap_alloc() block preference
 521  */
 522 #define HAMMER2_OFF_NOPREF              ((hammer2_off_t)-1)
 523
 524 /*
 525  * BMAP read-ahead maximum parameters
 526  */
 527 #define HAMMER2_BMAP_COUNT              16      /* max bmap read-ahead */
 528 #define HAMMER2_BMAP_BYTES              (HAMMER2_PBUFSIZE * HAMMER2_BMAP_COUNT)
 529
 530 /*
 531  * hammer2_freemap_adjust()
 532  */
 533 #define HAMMER2_FREEMAP_DORECOVER       1
 534 #define HAMMER2_FREEMAP_DOMAYFREE       2
 535 #define HAMMER2_FREEMAP_DOREALFREE      3
 536
 537 /*
 538  * HAMMER2 cluster - A set of chains representing the same entity.
 539  *
 540  * hammer2_cluster typically represents a temporary set of representitive
 541  * chains.  The one exception is that a hammer2_cluster is embedded in
 542  * hammer2_inode.  This embedded cluster is ONLY used to track the
 543  * representitive chains and cannot be directly locked.
 544  *
 545  * A cluster is usually temporary (and thus per-thread) for locking purposes,
 546  * allowing us to embed the asynchronous storage required for cluster
 547  * operations in the cluster itself and adjust the state and status without
 548  * having to worry too much about SMP issues.
 549  *
 550  * The exception is the cluster embedded in the hammer2_inode structure.
 551  * This is used to cache the cluster state on an inode-by-inode basis.
 552  * Individual hammer2_chain structures not incorporated into clusters might
 553  * also stick around to cache miscellanious elements.
 554  *
 555  * Because the cluster is a 'working copy' and is usually subject to cluster
 556  * quorum rules, it is quite possible for us to end up with an insufficient
 557  * number of live chains to execute an operation.  If an insufficient number
 558  * of chains remain in a working copy, the operation may have to be
 559  * downgraded, retried, stall until the requisit number of chains are
 560  * available, or possibly even error out depending on the mount type.
 561  *
 562  * A cluster's focus is set when it is locked.  The focus can only be set
 563  * to a chain still part of the synchronized set.
 564  */
 565 #define HAMMER2_XOPFIFO         16
 566 #define HAMMER2_XOPFIFO_MASK    (HAMMER2_XOPFIFO - 1)
 567 #define HAMMER2_XOPGROUPS       32
 568 #define HAMMER2_XOPGROUPS_MASK  (HAMMER2_XOPGROUPS - 1)
 569
 570 #define HAMMER2_MAXCLUSTER      8
 571 #define HAMMER2_XOPMASK_CLUSTER (uint64_t)((1LLU << HAMMER2_MAXCLUSTER) - 1)
 572 #define HAMMER2_XOPMASK_VOP     (uint64_t)0x0000000080000000LLU
 573 #define HAMMER2_XOPMASK_FIFOW   (uint64_t)0x0000000040000000LLU
 574 #define HAMMER2_XOPMASK_WAIT    (uint64_t)0x0000000020000000LLU
 575 #define HAMMER2_XOPMASK_FEED    (uint64_t)0x0000000100000000LLU
 576
 577 #define HAMMER2_XOPMASK_ALLDONE (HAMMER2_XOPMASK_VOP | HAMMER2_XOPMASK_CLUSTER)
 578
 579 #define HAMMER2_SPECTHREADS     1       /* sync */
 580
 581 struct hammer2_cluster_item {
 582         hammer2_chain_t         *chain;
 583         int                     error;
 584         uint32_t                flags;
 585 };
 586
 587 typedef struct hammer2_cluster_item hammer2_cluster_item_t;
 588
 589 /*
 590  * INVALID      - Invalid for focus, i.e. not part of synchronized set.
 591  *                Once set, this bit is sticky across operations.
 592  *
 593  * FEMOD        - Indicates that front-end modifying operations can
 594  *                mess with this entry and MODSYNC will copy also
 595  *                effect it.
 596  */
 597 #define HAMMER2_CITEM_INVALID   0x00000001
 598 #define HAMMER2_CITEM_FEMOD     0x00000002
 599 #define HAMMER2_CITEM_NULL      0x00000004
 600
 601 struct hammer2_cluster {
 602         int                     refs;           /* track for deallocation */
 603         int                     ddflag;
 604         struct hammer2_pfs      *pmp;
 605         uint32_t                flags;
 606         int                     nchains;
 607         int                     error;          /* error code valid on lock */
 608         int                     focus_index;
 609         hammer2_chain_t         *focus;         /* current focus (or mod) */
 610         hammer2_cluster_item_t  array[HAMMER2_MAXCLUSTER];
 611 };
 612
 613 typedef struct hammer2_cluster  hammer2_cluster_t;
 614
 615 /*
 616  * WRHARD       - Hard mounts can write fully synchronized
 617  * RDHARD       - Hard mounts can read fully synchronized
 618  * UNHARD       - Unsynchronized masters present
 619  * NOHARD       - No masters visible
 620  * WRSOFT       - Soft mounts can write to at least the SOFT_MASTER
 621  * RDSOFT       - Soft mounts can read from at least a SOFT_SLAVE
 622  * UNSOFT       - Unsynchronized slaves present
 623  * NOSOFT       - No slaves visible
 624  * RDSLAVE      - slaves are accessible (possibly unsynchronized or remote).
 625  * MSYNCED      - All masters are fully synchronized
 626  * SSYNCED      - All known local slaves are fully synchronized to masters
 627  *
 628  * All available masters are always incorporated.  All PFSs belonging to a
 629  * cluster (master, slave, copy, whatever) always try to synchronize the
 630  * total number of known masters in the PFSs root inode.
 631  *
 632  * A cluster might have access to many slaves, copies, or caches, but we
 633  * have a limited number of cluster slots.  Any such elements which are
 634  * directly mounted from block device(s) will always be incorporated.   Note
 635  * that SSYNCED only applies to such elements which are directly mounted,
 636  * not to any remote slaves, copies, or caches that could be available.  These
 637  * bits are used to monitor and drive our synchronization threads.
 638  *
 639  * When asking the question 'is any data accessible at all', then a simple
 640  * test against (RDHARD|RDSOFT|RDSLAVE) gives you the answer.  If any of
 641  * these bits are set the object can be read with certain caveats:
 642  * RDHARD - no caveats.  RDSOFT - authoritative but might not be synchronized.
 643  * and RDSLAVE - not authoritative, has some data but it could be old or
 644  * incomplete.
 645  *
 646  * When both soft and hard mounts are available, data will be read and written
 647  * via the soft mount only.  But all might be in the cluster because
 648  * background synchronization threads still need to do their work.
 649  */
 650 #define HAMMER2_CLUSTER_INODE   0x00000001      /* embedded in inode struct */
 651 #define HAMMER2_CLUSTER_UNUSED2 0x00000002
 652 #define HAMMER2_CLUSTER_LOCKED  0x00000004      /* cluster lks not recursive */
 653 #define HAMMER2_CLUSTER_WRHARD  0x00000100      /* hard-mount can write */
 654 #define HAMMER2_CLUSTER_RDHARD  0x00000200      /* hard-mount can read */
 655 #define HAMMER2_CLUSTER_UNHARD  0x00000400      /* unsynchronized masters */
 656 #define HAMMER2_CLUSTER_NOHARD  0x00000800      /* no masters visible */
 657 #define HAMMER2_CLUSTER_WRSOFT  0x00001000      /* soft-mount can write */
 658 #define HAMMER2_CLUSTER_RDSOFT  0x00002000      /* soft-mount can read */
 659 #define HAMMER2_CLUSTER_UNSOFT  0x00004000      /* unsynchronized slaves */
 660 #define HAMMER2_CLUSTER_NOSOFT  0x00008000      /* no slaves visible */
 661 #define HAMMER2_CLUSTER_MSYNCED 0x00010000      /* all masters synchronized */
 662 #define HAMMER2_CLUSTER_SSYNCED 0x00020000      /* known slaves synchronized */
 663
 664 #define HAMMER2_CLUSTER_ANYDATA ( HAMMER2_CLUSTER_RDHARD |      \
 665                                   HAMMER2_CLUSTER_RDSOFT |      \
 666                                   HAMMER2_CLUSTER_RDSLAVE)
 667
 668 #define HAMMER2_CLUSTER_RDOK    ( HAMMER2_CLUSTER_RDHARD |      \
 669                                   HAMMER2_CLUSTER_RDSOFT)
 670
 671 #define HAMMER2_CLUSTER_WROK    ( HAMMER2_CLUSTER_WRHARD |      \
 672                                   HAMMER2_CLUSTER_WRSOFT)
 673
 674 #define HAMMER2_CLUSTER_ZFLAGS  ( HAMMER2_CLUSTER_WRHARD |      \
 675                                   HAMMER2_CLUSTER_RDHARD |      \
 676                                   HAMMER2_CLUSTER_WRSOFT |      \
 677                                   HAMMER2_CLUSTER_RDSOFT |      \
 678                                   HAMMER2_CLUSTER_MSYNCED |     \
 679                                   HAMMER2_CLUSTER_SSYNCED)
 680
 681 /*
 682  * Helper functions (cluster must be locked for flags to be valid).
 683  */
 684 static __inline
 685 int
 686 hammer2_cluster_rdok(hammer2_cluster_t *cluster)
 687 {
 688         return (cluster->flags & HAMMER2_CLUSTER_RDOK);
 689 }
 690
 691 static __inline
 692 int
 693 hammer2_cluster_wrok(hammer2_cluster_t *cluster)
 694 {
 695         return (cluster->flags & HAMMER2_CLUSTER_WROK);
 696 }
 697
 698 RB_HEAD(hammer2_inode_tree, hammer2_inode);
 699
 700 /*
 701  * A hammer2 inode.
 702  *
 703  * NOTE: The inode-embedded cluster is never used directly for I/O (since
 704  *       it may be shared).  Instead it will be replicated-in and synchronized
 705  *       back out if changed.
 706  */
 707 struct hammer2_inode {
 708         RB_ENTRY(hammer2_inode) rbnode;         /* inumber lookup (HL) */
 709         hammer2_mtx_t           lock;           /* inode lock */
 710         hammer2_mtx_t           truncate_lock;  /* prevent truncates */
 711         struct hammer2_pfs      *pmp;           /* PFS mount */
 712         struct vnode            *vp;
 713         struct spinlock         cluster_spin;   /* update cluster */
 714         hammer2_cluster_t       cluster;
 715         struct lockf            advlock;
 716         u_int                   flags;
 717         u_int                   refs;           /* +vpref, +flushref */
 718         uint8_t                 comp_heuristic;
 719         hammer2_inode_meta_t    meta;           /* copy of meta-data */
 720         hammer2_off_t           osize;
 721 };
 722
 723 typedef struct hammer2_inode hammer2_inode_t;
 724
 725 /*
 726  * MODIFIED     - Inode is in a modified state, ip->meta may have changes.
 727  * RESIZED      - Inode truncated (any) or inode extended beyond
 728  *                EMBEDDED_BYTES.
 729  */
 730 #define HAMMER2_INODE_MODIFIED          0x0001
 731 #define HAMMER2_INODE_SROOT             0x0002  /* kmalloc special case */
 732 #define HAMMER2_INODE_RENAME_INPROG     0x0004
 733 #define HAMMER2_INODE_ONRBTREE          0x0008
 734 #define HAMMER2_INODE_RESIZED           0x0010  /* requires inode_fsync */
 735 #define HAMMER2_INODE_ISDELETED         0x0020  /* deleted */
 736 #define HAMMER2_INODE_ISUNLINKED        0x0040
 737 #define HAMMER2_INODE_METAGOOD          0x0080  /* inode meta-data good */
 738 #define HAMMER2_INODE_ONSIDEQ           0x0100  /* on side processing queue */
 739 #define HAMMER2_INODE_NOSIDEQ           0x0200  /* disable sideq operation */
 740
 741 int hammer2_inode_cmp(hammer2_inode_t *ip1, hammer2_inode_t *ip2);
 742 RB_PROTOTYPE2(hammer2_inode_tree, hammer2_inode, rbnode, hammer2_inode_cmp,
 743                 hammer2_tid_t);
 744
 745 /*
 746  * inode-unlink side-structure
 747  */
 748 struct hammer2_inode_sideq {
 749         TAILQ_ENTRY(hammer2_inode_sideq) entry;
 750         hammer2_inode_t *ip;
 751 };
 752 TAILQ_HEAD(h2_sideq_list, hammer2_inode_sideq);
 753
 754 typedef struct hammer2_inode_sideq hammer2_inode_sideq_t;
 755
 756 /*
 757  * Transaction management sub-structure under hammer2_pfs
 758  */
 759 struct hammer2_trans {
 760         uint32_t                flags;
 761         uint32_t                sync_wait;
 762 };
 763
 764 typedef struct hammer2_trans hammer2_trans_t;
 765
 766 #define HAMMER2_TRANS_ISFLUSH           0x80000000      /* flush code */
 767 #define HAMMER2_TRANS_BUFCACHE          0x40000000      /* bio strategy */
 768 #define HAMMER2_TRANS_UNUSED20          0x20000000
 769 #define HAMMER2_TRANS_FPENDING          0x10000000      /* flush pending */
 770 #define HAMMER2_TRANS_WAITING           0x08000000      /* someone waiting */
 771 #define HAMMER2_TRANS_MASK              0x00FFFFFF      /* count mask */
 772
 773 #define HAMMER2_FREEMAP_HEUR_NRADIX     4       /* pwr 2 PBUFRADIX-MINIORADIX */
 774 #define HAMMER2_FREEMAP_HEUR_TYPES      8
 775 #define HAMMER2_FREEMAP_HEUR_SIZE       (HAMMER2_FREEMAP_HEUR_NRADIX * \
 776                                          HAMMER2_FREEMAP_HEUR_TYPES)
 777
 778 #define HAMMER2_DEDUP_HEUR_SIZE         (65536 * 4)
 779 #define HAMMER2_DEDUP_HEUR_MASK         (HAMMER2_DEDUP_HEUR_SIZE - 1)
 780
 781 #define HAMMER2_FLUSH_TOP               0x0001
 782 #define HAMMER2_FLUSH_ALL               0x0002
 783
 784
 785 /*
 786  * Hammer2 support thread element.
 787  *
 788  * Potentially many support threads can hang off of hammer2, primarily
 789  * off the hammer2_pfs structure.  Typically:
 790  *
 791  * td x Nodes                   A synchronization thread for each node.
 792  * td x Nodes x workers         Worker threads for frontend operations.
 793  * td x 1                       Bioq thread for logical buffer writes.
 794  *
 795  * In addition, the synchronization thread(s) associated with the
 796  * super-root PFS (spmp) for a node is responsible for automatic bulkfree
 797  * and dedup scans.
 798  */
 799 struct hammer2_thread {
 800         struct hammer2_pfs *pmp;
 801         struct hammer2_dev *hmp;
 802         hammer2_xop_list_t xopq;
 803         thread_t        td;
 804         uint32_t        flags;
 805         int             depth;
 806         int             clindex;        /* cluster element index */
 807         int             repidx;
 808         char            *scratch;       /* MAXPHYS */
 809 };
 810
 811 typedef struct hammer2_thread hammer2_thread_t;
 812
 813 #define HAMMER2_THREAD_UNMOUNTING       0x0001  /* unmount request */
 814 #define HAMMER2_THREAD_DEV              0x0002  /* related to dev, not pfs */
 815 #define HAMMER2_THREAD_WAITING          0x0004  /* thread in idle tsleep */
 816 #define HAMMER2_THREAD_REMASTER         0x0008  /* remaster request */
 817 #define HAMMER2_THREAD_STOP             0x0010  /* exit request */
 818 #define HAMMER2_THREAD_FREEZE           0x0020  /* force idle */
 819 #define HAMMER2_THREAD_FROZEN           0x0040  /* thread is frozen */
 820 #define HAMMER2_THREAD_XOPQ             0x0080  /* work pending */
 821 #define HAMMER2_THREAD_STOPPED          0x0100  /* thread has stopped */
 822 #define HAMMER2_THREAD_UNFREEZE         0x0200
 823
 824 #define HAMMER2_THREAD_WAKEUP_MASK      (HAMMER2_THREAD_UNMOUNTING |    \
 825                                          HAMMER2_THREAD_REMASTER |      \
 826                                          HAMMER2_THREAD_STOP |          \
 827                                          HAMMER2_THREAD_FREEZE |        \
 828                                          HAMMER2_THREAD_XOPQ)
 829
 830 /*
 831  * Support structure for dedup heuristic.
 832  */
 833 struct hammer2_dedup {
 834         hammer2_off_t   data_off;
 835         uint64_t        data_crc;
 836         uint32_t        ticks;
 837         uint32_t        unused03;
 838 };
 839
 840 typedef struct hammer2_dedup hammer2_dedup_t;
 841
 842 /*
 843  * hammer2_xop - container for VOP/XOP operation (allocated, not on stack).
 844  *
 845  * This structure is used to distribute a VOP operation across multiple
 846  * nodes.  It provides a rendezvous for concurrent node execution and
 847  * can be detached from the frontend operation to allow the frontend to
 848  * return early.
 849  *
 850  * This structure also sequences operations on up to three inodes.
 851  */
 852 typedef void (*hammer2_xop_func_t)(hammer2_thread_t *thr,
 853                                    union hammer2_xop *xop);
 854
 855 struct hammer2_xop_fifo {
 856         TAILQ_ENTRY(hammer2_xop_head) entry;
 857         hammer2_chain_t         *array[HAMMER2_XOPFIFO];
 858         int                     errors[HAMMER2_XOPFIFO];
 859         int                     ri;
 860         int                     wi;
 861         int                     flags;
 862         hammer2_thread_t        *thr;
 863 };
 864
 865 typedef struct hammer2_xop_fifo hammer2_xop_fifo_t;
 866
 867 #define HAMMER2_XOP_FIFO_RUN    0x0001
 868 #define HAMMER2_XOP_FIFO_STALL  0x0002
 869
 870 struct hammer2_xop_head {
 871         hammer2_xop_func_t      func;
 872         hammer2_tid_t           mtid;
 873         struct hammer2_inode    *ip1;
 874         struct hammer2_inode    *ip2;
 875         struct hammer2_inode    *ip3;
 876         uint64_t                run_mask;
 877         uint64_t                chk_mask;
 878         int                     flags;
 879         int                     state;
 880         int                     error;
 881         hammer2_key_t           collect_key;
 882         char                    *name1;
 883         size_t                  name1_len;
 884         char                    *name2;
 885         size_t                  name2_len;
 886         hammer2_xop_fifo_t      collect[HAMMER2_MAXCLUSTER];
 887         hammer2_cluster_t       cluster;        /* help collections */
 888 };
 889
 890 typedef struct hammer2_xop_head hammer2_xop_head_t;
 891
 892 struct hammer2_xop_ipcluster {
 893         hammer2_xop_head_t      head;
 894 };
 895
 896 struct hammer2_xop_strategy {
 897         hammer2_xop_head_t      head;
 898         hammer2_key_t           lbase;
 899         int                     finished;
 900         hammer2_mtx_t           lock;
 901         struct bio              *bio;
 902 };
 903
 904 struct hammer2_xop_readdir {
 905         hammer2_xop_head_t      head;
 906         hammer2_key_t           lkey;
 907 };
 908
 909 struct hammer2_xop_nresolve {
 910         hammer2_xop_head_t      head;
 911         hammer2_key_t           lhc;    /* if name is NULL used lhc */
 912 };
 913
 914 struct hammer2_xop_unlink {
 915         hammer2_xop_head_t      head;
 916         int                     isdir;
 917         int                     dopermanent;
 918 };
 919
 920 #define H2DOPERM_PERMANENT      0x01
 921 #define H2DOPERM_FORCE          0x02
 922 #define H2DOPERM_IGNINO         0x04
 923
 924 struct hammer2_xop_nrename {
 925         hammer2_xop_head_t      head;
 926         hammer2_tid_t           lhc;
 927         int                     ip_key;
 928 };
 929
 930 struct hammer2_xop_scanlhc {
 931         hammer2_xop_head_t      head;
 932         hammer2_key_t           lhc;
 933 };
 934
 935 struct hammer2_xop_scanall {
 936         hammer2_xop_head_t      head;
 937         hammer2_key_t           key_beg;        /* inclusive */
 938         hammer2_key_t           key_end;        /* inclusive */
 939         int                     resolve_flags;
 940         int                     lookup_flags;
 941 };
 942
 943 struct hammer2_xop_lookup {
 944         hammer2_xop_head_t      head;
 945         hammer2_key_t           lhc;
 946 };
 947
 948 struct hammer2_xop_mkdirent {
 949         hammer2_xop_head_t      head;
 950         hammer2_dirent_head_t   dirent;
 951         hammer2_key_t           lhc;
 952 };
 953
 954 struct hammer2_xop_create {
 955         hammer2_xop_head_t      head;
 956         hammer2_inode_meta_t    meta;           /* initial metadata */
 957         hammer2_key_t           lhc;
 958         int                     flags;
 959 };
 960
 961 struct hammer2_xop_destroy {
 962         hammer2_xop_head_t      head;
 963 };
 964
 965 struct hammer2_xop_fsync {
 966         hammer2_xop_head_t      head;
 967         hammer2_inode_meta_t    meta;
 968         hammer2_off_t           osize;
 969         u_int                   ipflags;
 970         int                     clear_directdata;
 971 };
 972
 973 struct hammer2_xop_unlinkall {
 974         hammer2_xop_head_t      head;
 975         hammer2_key_t           key_beg;
 976         hammer2_key_t           key_end;
 977 };
 978
 979 struct hammer2_xop_connect {
 980         hammer2_xop_head_t      head;
 981         hammer2_key_t           lhc;
 982 };
 983
 984 struct hammer2_xop_flush {
 985         hammer2_xop_head_t      head;
 986 };
 987
 988 typedef struct hammer2_xop_readdir hammer2_xop_readdir_t;
 989 typedef struct hammer2_xop_nresolve hammer2_xop_nresolve_t;
 990 typedef struct hammer2_xop_unlink hammer2_xop_unlink_t;
 991 typedef struct hammer2_xop_nrename hammer2_xop_nrename_t;
 992 typedef struct hammer2_xop_ipcluster hammer2_xop_ipcluster_t;
 993 typedef struct hammer2_xop_strategy hammer2_xop_strategy_t;
 994 typedef struct hammer2_xop_mkdirent hammer2_xop_mkdirent_t;
 995 typedef struct hammer2_xop_create hammer2_xop_create_t;
 996 typedef struct hammer2_xop_destroy hammer2_xop_destroy_t;
 997 typedef struct hammer2_xop_fsync hammer2_xop_fsync_t;
 998 typedef struct hammer2_xop_unlinkall hammer2_xop_unlinkall_t;
 999 typedef struct hammer2_xop_scanlhc hammer2_xop_scanlhc_t;
1000 typedef struct hammer2_xop_scanall hammer2_xop_scanall_t;
1001 typedef struct hammer2_xop_lookup hammer2_xop_lookup_t;
1002 typedef struct hammer2_xop_connect hammer2_xop_connect_t;
1003 typedef struct hammer2_xop_flush hammer2_xop_flush_t;
1004
1005 union hammer2_xop {
1006         hammer2_xop_head_t      head;
1007         hammer2_xop_ipcluster_t xop_ipcluster;
1008         hammer2_xop_readdir_t   xop_readdir;
1009         hammer2_xop_nresolve_t  xop_nresolve;
1010         hammer2_xop_unlink_t    xop_unlink;
1011         hammer2_xop_nrename_t   xop_nrename;
1012         hammer2_xop_strategy_t  xop_strategy;
1013         hammer2_xop_mkdirent_t  xop_mkdirent;
1014         hammer2_xop_create_t    xop_create;
1015         hammer2_xop_destroy_t   xop_destroy;
1016         hammer2_xop_fsync_t     xop_fsync;
1017         hammer2_xop_unlinkall_t xop_unlinkall;
1018         hammer2_xop_scanlhc_t   xop_scanlhc;
1019         hammer2_xop_scanall_t   xop_scanall;
1020         hammer2_xop_lookup_t    xop_lookup;
1021         hammer2_xop_flush_t     xop_flush;
1022         hammer2_xop_connect_t   xop_connect;
1023 };
1024
1025 typedef union hammer2_xop hammer2_xop_t;
1026
1027 /*
1028  * hammer2_xop_group - Manage XOP support threads.
1029  */
1030 struct hammer2_xop_group {
1031         hammer2_thread_t        thrs[HAMMER2_MAXCLUSTER];
1032 };
1033
1034 typedef struct hammer2_xop_group hammer2_xop_group_t;
1035
1036 /*
1037  * flags to hammer2_xop_collect()
1038  */
1039 #define HAMMER2_XOP_COLLECT_NOWAIT      0x00000001
1040 #define HAMMER2_XOP_COLLECT_WAITALL     0x00000002
1041
1042 /*
1043  * flags to hammer2_xop_alloc()
1044  *
1045  * MODIFYING    - This is a modifying transaction, allocate a mtid.
1046  */
1047 #define HAMMER2_XOP_MODIFYING           0x00000001
1048 #define HAMMER2_XOP_STRATEGY            0x00000002
1049
1050 /*
1051  * Global (per partition) management structure, represents a hard block
1052  * device.  Typically referenced by hammer2_chain structures when applicable.
1053  * Typically not used for network-managed elements.
1054  *
1055  * Note that a single hammer2_dev can be indirectly tied to multiple system
1056  * mount points.  There is no direct relationship.  System mounts are
1057  * per-cluster-id, not per-block-device, and a single hard mount might contain
1058  * many PFSs and those PFSs might combine together in various ways to form
1059  * the set of available clusters.
1060  */
1061 struct hammer2_dev {
1062         struct vnode    *devvp;         /* device vnode */
1063         int             ronly;          /* read-only mount */
1064         int             mount_count;    /* number of actively mounted PFSs */
1065         TAILQ_ENTRY(hammer2_dev) mntentry; /* hammer2_mntlist */
1066
1067         struct malloc_type *mchain;
1068         int             nipstacks;
1069         int             maxipstacks;
1070         kdmsg_iocom_t   iocom;          /* volume-level dmsg interface */
1071         struct spinlock io_spin;        /* iotree, iolruq access */
1072         struct hammer2_io_tree iotree;
1073         int             iofree_count;
1074         int             freemap_relaxed;
1075         hammer2_chain_t vchain;         /* anchor chain (topology) */
1076         hammer2_chain_t fchain;         /* anchor chain (freemap) */
1077         struct spinlock list_spin;
1078         struct h2_flush_list flushq;    /* flush seeds */
1079         struct hammer2_pfs *spmp;       /* super-root pmp for transactions */
1080         struct lock     vollk;          /* lockmgr lock */
1081         struct lock     bulklk;         /* bulkfree operation lock */
1082         struct lock     bflock;         /* bulk-free manual function lock */
1083         hammer2_off_t   heur_freemap[HAMMER2_FREEMAP_HEUR_SIZE];
1084         hammer2_dedup_t heur_dedup[HAMMER2_DEDUP_HEUR_SIZE];
1085         int             volhdrno;       /* last volhdrno written */
1086         uint32_t        hflags;         /* HMNT2 flags applicable to device */
1087         hammer2_off_t   free_reserved;  /* nominal free reserved */
1088         hammer2_thread_t bfthr;         /* bulk-free thread */
1089         char            devrepname[64]; /* for kprintf */
1090         hammer2_ioc_bulkfree_t bflast;  /* stats for last bulkfree run */
1091         hammer2_volume_data_t voldata;
1092         hammer2_volume_data_t volsync;  /* synchronized voldata */
1093 };
1094
1095 typedef struct hammer2_dev hammer2_dev_t;
1096
1097 /*
1098  * Helper functions (cluster must be locked for flags to be valid).
1099  */
1100 static __inline
1101 int
1102 hammer2_chain_rdok(hammer2_chain_t *chain)
1103 {
1104         return (chain->error == 0);
1105 }
1106
1107 static __inline
1108 int
1109 hammer2_chain_wrok(hammer2_chain_t *chain)
1110 {
1111         return (chain->error == 0 && chain->hmp->ronly == 0);
1112 }
1113
1114 /*
1115  * Per-cluster management structure.  This structure will be tied to a
1116  * system mount point if the system is mounting the PFS, but is also used
1117  * to manage clusters encountered during the super-root scan or received
1118  * via LNK_SPANs that might not be mounted.
1119  *
1120  * This structure is also used to represent the super-root that hangs off
1121  * of a hard mount point.  The super-root is not really a cluster element.
1122  * In this case the spmp_hmp field will be non-NULL.  It's just easier to do
1123  * this than to special case super-root manipulation in the hammer2_chain*
1124  * code as being only hammer2_dev-related.
1125  *
1126  * pfs_mode and pfs_nmasters are rollup fields which critically describes
1127  * how elements of the cluster act on the cluster.  pfs_mode is only applicable
1128  * when a PFS is mounted by the system.  pfs_nmasters is our best guess as to
1129  * how many masters have been configured for a cluster and is always
1130  * applicable.  pfs_types[] is an array with 1:1 correspondance to the
1131  * iroot cluster and describes the PFS types of the nodes making up the
1132  * cluster.
1133  *
1134  * WARNING! Portions of this structure have deferred initialization.  In
1135  *          particular, if not mounted there will be no wthread.
1136  *          umounted network PFSs will also be missing iroot and numerous
1137  *          other fields will not be initialized prior to mount.
1138  *
1139  *          Synchronization threads are chain-specific and only applicable
1140  *          to local hard PFS entries.  A hammer2_pfs structure may contain
1141  *          more than one when multiple hard PFSs are present on the local
1142  *          machine which require synchronization monitoring.  Most PFSs
1143  *          (such as snapshots) are 1xMASTER PFSs which do not need a
1144  *          synchronization thread.
1145  *
1146  * WARNING! The chains making up pfs->iroot's cluster are accounted for in
1147  *          hammer2_dev->mount_count when the pfs is associated with a mount
1148  *          point.
1149  */
1150 struct hammer2_pfs {
1151         struct mount            *mp;
1152         TAILQ_ENTRY(hammer2_pfs) mntentry;      /* hammer2_pfslist */
1153         uuid_t                  pfs_clid;
1154         hammer2_dev_t           *spmp_hmp;      /* only if super-root pmp */
1155         hammer2_dev_t           *force_local;   /* only if 'local' mount */
1156         hammer2_inode_t         *iroot;         /* PFS root inode */
1157         uint8_t                 pfs_types[HAMMER2_MAXCLUSTER];
1158         char                    *pfs_names[HAMMER2_MAXCLUSTER];
1159         hammer2_dev_t           *pfs_hmps[HAMMER2_MAXCLUSTER];
1160         hammer2_trans_t         trans;
1161         struct lock             lock;           /* PFS lock for certain ops */
1162         struct lock             lock_nlink;     /* rename and nlink lock */
1163         struct netexport        export;         /* nfs export */
1164         int                     ronly;          /* read-only mount */
1165         int                     hflags;         /* pfs-specific mount flags */
1166         struct malloc_type      *minode;
1167         struct malloc_type      *mmsg;
1168         struct spinlock         inum_spin;      /* inumber lookup */
1169         struct hammer2_inode_tree inum_tree;    /* (not applicable to spmp) */
1170         long                    inum_count;     /* #of inodes in inum_tree */
1171         struct spinlock         lru_spin;       /* inumber lookup */
1172         struct hammer2_chain_list lru_list;     /* chains on LRU */
1173         int                     lru_count;      /* #of chains on LRU */
1174         hammer2_tid_t           modify_tid;     /* modify transaction id */
1175         hammer2_tid_t           inode_tid;      /* inode allocator */
1176         uint8_t                 pfs_nmasters;   /* total masters */
1177         uint8_t                 pfs_mode;       /* operating mode PFSMODE */
1178         uint8_t                 unused01;
1179         uint8_t                 unused02;
1180         int                     free_ticks;     /* free_* calculations */
1181         long                    inmem_inodes;
1182         hammer2_off_t           free_reserved;
1183         hammer2_off_t           free_nominal;
1184         uint32_t                inmem_dirty_chains;
1185         int                     count_lwinprog; /* logical write in prog */
1186         struct spinlock         list_spin;
1187         struct h2_sideq_list    sideq;          /* last-close dirty/unlink */
1188         long                    sideq_count;
1189         hammer2_thread_t        sync_thrs[HAMMER2_MAXCLUSTER];
1190         uint32_t                cluster_flags;  /* cached cluster flags */
1191         int                     has_xop_threads;
1192         struct spinlock         xop_spin;       /* xop sequencer */
1193         hammer2_xop_group_t     xop_groups[HAMMER2_XOPGROUPS];
1194 };
1195
1196 typedef struct hammer2_pfs hammer2_pfs_t;
1197
1198 TAILQ_HEAD(hammer2_pfslist, hammer2_pfs);
1199
1200 #define HAMMER2_LRU_LIMIT               1024    /* per pmp lru_list */
1201
1202 #define HAMMER2_DIRTYCHAIN_WAITING      0x80000000
1203 #define HAMMER2_DIRTYCHAIN_MASK         0x7FFFFFFF
1204
1205 #define HAMMER2_LWINPROG_WAITING        0x80000000
1206 #define HAMMER2_LWINPROG_WAITING0       0x40000000
1207 #define HAMMER2_LWINPROG_MASK           0x3FFFFFFF
1208
1209 /*
1210  * hammer2_cluster_check
1211  */
1212 #define HAMMER2_CHECK_NULL      0x00000001
1213
1214 /*
1215  * Misc
1216  */
1217 #if defined(_KERNEL)
1218
1219 MALLOC_DECLARE(M_HAMMER2);
1220
1221 #define VTOI(vp)        ((hammer2_inode_t *)(vp)->v_data)
1222 #define ITOV(ip)        ((ip)->vp)
1223
1224 /*
1225  * Currently locked chains retain the locked buffer cache buffer for
1226  * indirect blocks, and indirect blocks can be one of two sizes.  The
1227  * device buffer has to match the case to avoid deadlocking recursive
1228  * chains that might otherwise try to access different offsets within
1229  * the same device buffer.
1230  */
1231 static __inline
1232 int
1233 hammer2_devblkradix(int radix)
1234 {
1235 #if 0
1236         if (radix <= HAMMER2_LBUFRADIX) {
1237                 return (HAMMER2_LBUFRADIX);
1238         } else {
1239                 return (HAMMER2_PBUFRADIX);
1240         }
1241 #endif
1242         return (HAMMER2_PBUFRADIX);
1243 }
1244
1245 /*
1246  * XXX almost time to remove this.  DIO uses PBUFSIZE exclusively now.
1247  */
1248 static __inline
1249 size_t
1250 hammer2_devblksize(size_t bytes)
1251 {
1252 #if 0
1253         if (bytes <= HAMMER2_LBUFSIZE) {
1254                 return(HAMMER2_LBUFSIZE);
1255         } else {
1256                 KKASSERT(bytes <= HAMMER2_PBUFSIZE &&
1257                          (bytes ^ (bytes - 1)) == ((bytes << 1) - 1));
1258                 return (HAMMER2_PBUFSIZE);
1259         }
1260 #endif
1261         return (HAMMER2_PBUFSIZE);
1262 }
1263
1264
1265 static __inline
1266 hammer2_pfs_t *
1267 MPTOPMP(struct mount *mp)
1268 {
1269         return ((hammer2_pfs_t *)mp->mnt_data);
1270 }
1271
1272 #define HAMMER2_DEDUP_FRAG      (HAMMER2_PBUFSIZE / 64)
1273 #define HAMMER2_DEDUP_FRAGRADIX (HAMMER2_PBUFRADIX - 6)
1274
1275 static __inline
1276 uint64_t
1277 hammer2_dedup_mask(hammer2_io_t *dio, hammer2_off_t data_off, u_int bytes)
1278 {
1279         int bbeg;
1280         int bits;
1281         uint64_t mask;
1282
1283         bbeg = (int)((data_off & ~HAMMER2_OFF_MASK_RADIX) - dio->pbase) >>
1284                HAMMER2_DEDUP_FRAGRADIX;
1285         bits = (int)((bytes + (HAMMER2_DEDUP_FRAG - 1)) >>
1286                HAMMER2_DEDUP_FRAGRADIX);
1287         mask = ((uint64_t)1 << bbeg) - 1;
1288         if (bbeg + bits == 64)
1289                 mask = (uint64_t)-1;
1290         else
1291                 mask = ((uint64_t)1 << (bbeg + bits)) - 1;
1292
1293         mask &= ~(((uint64_t)1 << bbeg) - 1);
1294
1295         return mask;
1296 }
1297
1298 static __inline
1299 int
1300 hammer2_error_to_errno(int error)
1301 {
1302         if (error) {
1303                 if (error & HAMMER2_ERROR_EIO)
1304                         error = EIO;
1305                 else if (error & HAMMER2_ERROR_CHECK)
1306                         error = EDOM;
1307                 else if (error & HAMMER2_ERROR_ABORTED)
1308                         error = EINTR;
1309                 else if (error & HAMMER2_ERROR_BADBREF)
1310                         error = EIO;
1311                 else if (error & HAMMER2_ERROR_ENOSPC)
1312                         error = ENOSPC;
1313                 else if (error & HAMMER2_ERROR_ENOENT)
1314                         error = ENOENT;
1315                 else if (error & HAMMER2_ERROR_ENOTEMPTY)
1316                         error = ENOTEMPTY;
1317                 else if (error & HAMMER2_ERROR_EAGAIN)
1318                         error = EAGAIN;
1319                 else if (error & HAMMER2_ERROR_ENOTDIR)
1320                         error = ENOTDIR;
1321                 else if (error & HAMMER2_ERROR_EISDIR)
1322                         error = EISDIR;
1323                 else if (error & HAMMER2_ERROR_EINPROGRESS)
1324                         error = EINPROGRESS;
1325                 else if (error & HAMMER2_ERROR_EEXIST)
1326                         error = EEXIST;
1327                 else
1328                         error = EDOM;
1329         }
1330         return error;
1331 }
1332
1333 static __inline
1334 int
1335 hammer2_errno_to_error(int error)
1336 {
1337         switch(error) {
1338         case 0:
1339                 return 0;
1340         case EIO:
1341                 return HAMMER2_ERROR_EIO;
1342         case EINVAL:
1343         default:
1344                 return HAMMER2_ERROR_EINVAL;
1345         }
1346 }
1347
1348 extern struct vop_ops hammer2_vnode_vops;
1349 extern struct vop_ops hammer2_spec_vops;
1350 extern struct vop_ops hammer2_fifo_vops;
1351 extern struct hammer2_pfslist hammer2_pfslist;
1352 extern struct lock hammer2_mntlk;
1353
1354
1355 extern int hammer2_debug;
1356 extern int hammer2_cluster_meta_read;
1357 extern int hammer2_cluster_data_read;
1358 extern int hammer2_dedup_enable;
1359 extern int hammer2_always_compress;
1360 extern int hammer2_inval_enable;
1361 extern int hammer2_flush_pipe;
1362 extern int hammer2_synchronous_flush;
1363 extern int hammer2_dio_count;
1364 extern int hammer2_limit_dio;
1365 extern int hammer2_bulkfree_tps;
1366 extern long hammer2_chain_allocs;
1367 extern long hammer2_chain_frees;
1368 extern long hammer2_limit_dirty_chains;
1369 extern long hammer2_count_modified_chains;
1370 extern long hammer2_iod_invals;
1371 extern long hammer2_iod_file_read;
1372 extern long hammer2_iod_meta_read;
1373 extern long hammer2_iod_indr_read;
1374 extern long hammer2_iod_fmap_read;
1375 extern long hammer2_iod_volu_read;
1376 extern long hammer2_iod_file_write;
1377 extern long hammer2_iod_file_wembed;
1378 extern long hammer2_iod_file_wzero;
1379 extern long hammer2_iod_file_wdedup;
1380 extern long hammer2_iod_meta_write;
1381 extern long hammer2_iod_indr_write;
1382 extern long hammer2_iod_fmap_write;
1383 extern long hammer2_iod_volu_write;
1384
1385 extern long hammer2_check_xxhash64;
1386 extern long hammer2_check_icrc32;
1387
1388 extern struct objcache *cache_buffer_read;
1389 extern struct objcache *cache_buffer_write;
1390 extern struct objcache *cache_xops;
1391
1392 /*
1393  * hammer2_subr.c
1394  */
1395 #define hammer2_icrc32(buf, size)       iscsi_crc32((buf), (size))
1396 #define hammer2_icrc32c(buf, size, crc) iscsi_crc32_ext((buf), (size), (crc))
1397
1398 int hammer2_signal_check(time_t *timep);
1399 const char *hammer2_error_str(int error);
1400
1401 void hammer2_inode_lock(hammer2_inode_t *ip, int how);
1402 void hammer2_inode_unlock(hammer2_inode_t *ip);
1403 hammer2_chain_t *hammer2_inode_chain(hammer2_inode_t *ip, int clindex, int how);
1404 hammer2_chain_t *hammer2_inode_chain_and_parent(hammer2_inode_t *ip,
1405                         int clindex, hammer2_chain_t **parentp, int how);
1406 hammer2_mtx_state_t hammer2_inode_lock_temp_release(hammer2_inode_t *ip);
1407 void hammer2_inode_lock_temp_restore(hammer2_inode_t *ip,
1408                         hammer2_mtx_state_t ostate);
1409 int hammer2_inode_lock_upgrade(hammer2_inode_t *ip);
1410 void hammer2_inode_lock_downgrade(hammer2_inode_t *ip, int);
1411
1412 void hammer2_dev_exlock(hammer2_dev_t *hmp);
1413 void hammer2_dev_shlock(hammer2_dev_t *hmp);
1414 void hammer2_dev_unlock(hammer2_dev_t *hmp);
1415
1416 int hammer2_get_dtype(uint8_t type);
1417 int hammer2_get_vtype(uint8_t type);
1418 uint8_t hammer2_get_obj_type(enum vtype vtype);
1419 void hammer2_time_to_timespec(uint64_t xtime, struct timespec *ts);
1420 uint64_t hammer2_timespec_to_time(const struct timespec *ts);
1421 uint32_t hammer2_to_unix_xid(const uuid_t *uuid);
1422 void hammer2_guid_to_uuid(uuid_t *uuid, uint32_t guid);
1423 void hammer2_trans_manage_init(hammer2_pfs_t *pmp);
1424
1425 hammer2_key_t hammer2_dirhash(const unsigned char *name, size_t len);
1426 int hammer2_getradix(size_t bytes);
1427
1428 int hammer2_calc_logical(hammer2_inode_t *ip, hammer2_off_t uoff,
1429                         hammer2_key_t *lbasep, hammer2_key_t *leofp);
1430 int hammer2_calc_physical(hammer2_inode_t *ip, hammer2_key_t lbase);
1431 void hammer2_update_time(uint64_t *timep);
1432 void hammer2_adjreadcounter(hammer2_blockref_t *bref, size_t bytes);
1433
1434 /*
1435  * hammer2_inode.c
1436  */
1437 struct vnode *hammer2_igetv(hammer2_inode_t *ip, int *errorp);
1438 hammer2_inode_t *hammer2_inode_lookup(hammer2_pfs_t *pmp,
1439                         hammer2_tid_t inum);
1440 hammer2_inode_t *hammer2_inode_get(hammer2_pfs_t *pmp, hammer2_inode_t *dip,
1441                         hammer2_cluster_t *cluster, int idx);
1442 void hammer2_inode_free(hammer2_inode_t *ip);
1443 void hammer2_inode_ref(hammer2_inode_t *ip);
1444 void hammer2_inode_drop(hammer2_inode_t *ip);
1445 void hammer2_inode_repoint(hammer2_inode_t *ip, hammer2_inode_t *pip,
1446                         hammer2_cluster_t *cluster);
1447 void hammer2_inode_repoint_one(hammer2_inode_t *ip, hammer2_cluster_t *cluster,
1448                         int idx);
1449 void hammer2_inode_modify(hammer2_inode_t *ip);
1450 void hammer2_inode_run_sideq(hammer2_pfs_t *pmp, int doall);
1451
1452 hammer2_inode_t *hammer2_inode_create(hammer2_inode_t *dip,
1453                         hammer2_inode_t *pip,
1454                         struct vattr *vap, struct ucred *cred,
1455                         const uint8_t *name, size_t name_len, hammer2_key_t lhc,
1456                         hammer2_key_t inum, uint8_t type, uint8_t target_type,
1457                         int flags, int *errorp);
1458 void hammer2_inode_chain_sync(hammer2_inode_t *ip);
1459 int hammer2_inode_unlink_finisher(hammer2_inode_t *ip, int isopen);
1460 int hammer2_dirent_create(hammer2_inode_t *dip, const char *name,
1461                         size_t name_len, hammer2_key_t inum, uint8_t type);
1462
1463 /*
1464  * hammer2_chain.c
1465  */
1466 void hammer2_voldata_lock(hammer2_dev_t *hmp);
1467 void hammer2_voldata_unlock(hammer2_dev_t *hmp);
1468 void hammer2_voldata_modify(hammer2_dev_t *hmp);
1469 hammer2_chain_t *hammer2_chain_alloc(hammer2_dev_t *hmp,
1470                                 hammer2_pfs_t *pmp,
1471                                 hammer2_blockref_t *bref);
1472 void hammer2_chain_core_init(hammer2_chain_t *chain);
1473 void hammer2_chain_ref(hammer2_chain_t *chain);
1474 void hammer2_chain_ref_hold(hammer2_chain_t *chain);
1475 void hammer2_chain_drop(hammer2_chain_t *chain);
1476 void hammer2_chain_drop_unhold(hammer2_chain_t *chain);
1477 void hammer2_chain_lock(hammer2_chain_t *chain, int how);
1478 void hammer2_chain_lock_unhold(hammer2_chain_t *chain, int how);
1479 void hammer2_chain_load_data(hammer2_chain_t *chain);
1480 const hammer2_media_data_t *hammer2_chain_rdata(hammer2_chain_t *chain);
1481 hammer2_media_data_t *hammer2_chain_wdata(hammer2_chain_t *chain);
1482
1483 int hammer2_chain_inode_find(hammer2_pfs_t *pmp, hammer2_key_t inum,
1484                                 int clindex, int flags,
1485                                 hammer2_chain_t **parentp,
1486                                 hammer2_chain_t **chainp);
1487 int hammer2_chain_modify(hammer2_chain_t *chain, hammer2_tid_t mtid,
1488                                 hammer2_off_t dedup_off, int flags);
1489 int hammer2_chain_modify_ip(hammer2_inode_t *ip, hammer2_chain_t *chain,
1490                                 hammer2_tid_t mtid, int flags);
1491 int hammer2_chain_resize(hammer2_chain_t *chain,
1492                                 hammer2_tid_t mtid, hammer2_off_t dedup_off,
1493                                 int nradix, int flags);
1494 void hammer2_chain_unlock(hammer2_chain_t *chain);
1495 void hammer2_chain_unlock_hold(hammer2_chain_t *chain);
1496 void hammer2_chain_wait(hammer2_chain_t *chain);
1497 hammer2_chain_t *hammer2_chain_get(hammer2_chain_t *parent, int generation,
1498                                 hammer2_blockref_t *bref);
1499 hammer2_chain_t *hammer2_chain_lookup_init(hammer2_chain_t *parent, int flags);
1500 void hammer2_chain_lookup_done(hammer2_chain_t *parent);
1501 hammer2_chain_t *hammer2_chain_getparent(hammer2_chain_t *chain, int flags);
1502 hammer2_chain_t *hammer2_chain_repparent(hammer2_chain_t **chainp, int flags);
1503 hammer2_chain_t *hammer2_chain_lookup(hammer2_chain_t **parentp,
1504                                 hammer2_key_t *key_nextp,
1505                                 hammer2_key_t key_beg, hammer2_key_t key_end,
1506                                 int *errorp, int flags);
1507 hammer2_chain_t *hammer2_chain_next(hammer2_chain_t **parentp,
1508                                 hammer2_chain_t *chain,
1509                                 hammer2_key_t *key_nextp,
1510                                 hammer2_key_t key_beg, hammer2_key_t key_end,
1511                                 int *errorp, int flags);
1512 int hammer2_chain_scan(hammer2_chain_t *parent,
1513                                 hammer2_chain_t **chainp,
1514                                 hammer2_blockref_t *bref,
1515                                 int *firstp, int flags);
1516
1517 int hammer2_chain_create(hammer2_chain_t **parentp, hammer2_chain_t **chainp,
1518                                 hammer2_pfs_t *pmp, int methods,
1519                                 hammer2_key_t key, int keybits,
1520                                 int type, size_t bytes, hammer2_tid_t mtid,
1521                                 hammer2_off_t dedup_off, int flags);
1522 void hammer2_chain_rename(hammer2_blockref_t *bref,
1523                                 hammer2_chain_t **parentp,
1524                                 hammer2_chain_t *chain,
1525                                 hammer2_tid_t mtid, int flags);
1526 int hammer2_chain_delete(hammer2_chain_t *parent, hammer2_chain_t *chain,
1527                                 hammer2_tid_t mtid, int flags);
1528 int hammer2_chain_indirect_maintenance(hammer2_chain_t *parent,
1529                                 hammer2_chain_t *chain);
1530 void hammer2_chain_setflush(hammer2_chain_t *chain);
1531 void hammer2_chain_countbrefs(hammer2_chain_t *chain,
1532                                 hammer2_blockref_t *base, int count);
1533 hammer2_chain_t *hammer2_chain_bulksnap(hammer2_dev_t *hmp);
1534 void hammer2_chain_bulkdrop(hammer2_chain_t *copy);
1535
1536 void hammer2_chain_setcheck(hammer2_chain_t *chain, void *bdata);
1537 int hammer2_chain_testcheck(hammer2_chain_t *chain, void *bdata);
1538 int hammer2_chain_dirent_test(hammer2_chain_t *chain, const char *name,
1539                                 size_t name_len);
1540
1541 void hammer2_pfs_memory_wait(hammer2_pfs_t *pmp);
1542 void hammer2_pfs_memory_inc(hammer2_pfs_t *pmp);
1543 void hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp);
1544
1545 void hammer2_base_delete(hammer2_chain_t *parent,
1546                                 hammer2_blockref_t *base, int count,
1547                                 hammer2_chain_t *chain);
1548 void hammer2_base_insert(hammer2_chain_t *parent,
1549                                 hammer2_blockref_t *base, int count,
1550                                 hammer2_chain_t *chain,
1551                                 hammer2_blockref_t *elm);
1552
1553 /*
1554  * hammer2_flush.c
1555  */
1556 int hammer2_flush(hammer2_chain_t *chain, int istop);
1557 void hammer2_delayed_flush(hammer2_chain_t *chain);
1558
1559 /*
1560  * hammer2_trans.c
1561  */
1562 void hammer2_trans_init(hammer2_pfs_t *pmp, uint32_t flags);
1563 hammer2_tid_t hammer2_trans_sub(hammer2_pfs_t *pmp);
1564 void hammer2_trans_done(hammer2_pfs_t *pmp);
1565 hammer2_tid_t hammer2_trans_newinum(hammer2_pfs_t *pmp);
1566 void hammer2_trans_assert_strategy(hammer2_pfs_t *pmp);
1567 void hammer2_dedup_record(hammer2_chain_t *chain, hammer2_io_t *dio,
1568                                 char *data);
1569
1570 /*
1571  * hammer2_ioctl.c
1572  */
1573 int hammer2_ioctl(hammer2_inode_t *ip, u_long com, void *data,
1574                                 int fflag, struct ucred *cred);
1575
1576 /*
1577  * hammer2_io.c
1578  */
1579 void hammer2_io_putblk(hammer2_io_t **diop);
1580 void hammer2_io_inval(hammer2_io_t *dio, hammer2_off_t data_off, u_int bytes);
1581 void hammer2_io_cleanup(hammer2_dev_t *hmp, struct hammer2_io_tree *tree);
1582 char *hammer2_io_data(hammer2_io_t *dio, off_t lbase);
1583 void hammer2_io_bkvasync(hammer2_io_t *dio);
1584 hammer2_io_t *hammer2_io_getblk(hammer2_dev_t *hmp, int btype, off_t lbase,
1585                                 int lsize, int op);
1586 void hammer2_io_dedup_set(hammer2_dev_t *hmp, hammer2_blockref_t *bref);
1587 void hammer2_io_dedup_delete(hammer2_dev_t *hmp, uint8_t btype,
1588                                 hammer2_off_t data_off, u_int bytes);
1589 void hammer2_io_dedup_assert(hammer2_dev_t *hmp, hammer2_off_t data_off,
1590                                 u_int bytes);
1591 void hammer2_io_callback(struct bio *bio);
1592 int hammer2_io_new(hammer2_dev_t *hmp, int btype, off_t lbase, int lsize,
1593                                 hammer2_io_t **diop);
1594 int hammer2_io_newnz(hammer2_dev_t *hmp, int btype, off_t lbase, int lsize,
1595                                 hammer2_io_t **diop);
1596 int hammer2_io_bread(hammer2_dev_t *hmp, int btype, off_t lbase, int lsize,
1597                                 hammer2_io_t **diop);
1598 hammer2_io_t *hammer2_io_getquick(hammer2_dev_t *hmp, off_t lbase, int lsize);
1599 void hammer2_io_bawrite(hammer2_io_t **diop);
1600 void hammer2_io_bdwrite(hammer2_io_t **diop);
1601 int hammer2_io_bwrite(hammer2_io_t **diop);
1602 void hammer2_io_setdirty(hammer2_io_t *dio);
1603 void hammer2_io_brelse(hammer2_io_t **diop);
1604 void hammer2_io_bqrelse(hammer2_io_t **diop);
1605
1606 /*
1607  * hammer2_thread.c
1608  */
1609 void hammer2_thr_signal(hammer2_thread_t *thr, uint32_t flags);
1610 void hammer2_thr_signal2(hammer2_thread_t *thr,
1611                         uint32_t pflags, uint32_t nflags);
1612 void hammer2_thr_wait(hammer2_thread_t *thr, uint32_t flags);
1613 void hammer2_thr_wait_neg(hammer2_thread_t *thr, uint32_t flags);
1614 int hammer2_thr_wait_any(hammer2_thread_t *thr, uint32_t flags, int timo);
1615 void hammer2_thr_create(hammer2_thread_t *thr,
1616                         hammer2_pfs_t *pmp, hammer2_dev_t *hmp,
1617                         const char *id, int clindex, int repidx,
1618                         void (*func)(void *arg));
1619 void hammer2_thr_delete(hammer2_thread_t *thr);
1620 void hammer2_thr_remaster(hammer2_thread_t *thr);
1621 void hammer2_thr_freeze_async(hammer2_thread_t *thr);
1622 void hammer2_thr_freeze(hammer2_thread_t *thr);
1623 void hammer2_thr_unfreeze(hammer2_thread_t *thr);
1624 int hammer2_thr_break(hammer2_thread_t *thr);
1625 void hammer2_primary_xops_thread(void *arg);
1626
1627 /*
1628  * hammer2_thread.c (XOP API)
1629  */
1630 void hammer2_xop_group_init(hammer2_pfs_t *pmp, hammer2_xop_group_t *xgrp);
1631 void *hammer2_xop_alloc(hammer2_inode_t *ip, int flags);
1632 void hammer2_xop_setname(hammer2_xop_head_t *xop,
1633                                 const char *name, size_t name_len);
1634 void hammer2_xop_setname2(hammer2_xop_head_t *xop,
1635                                 const char *name, size_t name_len);
1636 size_t hammer2_xop_setname_inum(hammer2_xop_head_t *xop, hammer2_key_t inum);
1637 void hammer2_xop_setip2(hammer2_xop_head_t *xop, hammer2_inode_t *ip2);
1638 void hammer2_xop_setip3(hammer2_xop_head_t *xop, hammer2_inode_t *ip3);
1639 void hammer2_xop_reinit(hammer2_xop_head_t *xop);
1640 void hammer2_xop_helper_create(hammer2_pfs_t *pmp);
1641 void hammer2_xop_helper_cleanup(hammer2_pfs_t *pmp);
1642 void hammer2_xop_start(hammer2_xop_head_t *xop, hammer2_xop_func_t func);
1643 void hammer2_xop_start_except(hammer2_xop_head_t *xop, hammer2_xop_func_t func,
1644                                 int notidx);
1645 int hammer2_xop_collect(hammer2_xop_head_t *xop, int flags);
1646 void hammer2_xop_retire(hammer2_xop_head_t *xop, uint64_t mask);
1647 int hammer2_xop_active(hammer2_xop_head_t *xop);
1648 int hammer2_xop_feed(hammer2_xop_head_t *xop, hammer2_chain_t *chain,
1649                                 int clindex, int error);
1650
1651 /*
1652  * hammer2_synchro.c
1653  */
1654 void hammer2_primary_sync_thread(void *arg);
1655
1656 /*
1657  * XOP backends in hammer2_xops.c, primarily for VNOPS.  Other XOP backends
1658  * may be integrated into other source files.
1659  */
1660 void hammer2_xop_ipcluster(hammer2_thread_t *thr, hammer2_xop_t *xop);
1661 void hammer2_xop_readdir(hammer2_thread_t *thr, hammer2_xop_t *xop);
1662 void hammer2_xop_nresolve(hammer2_thread_t *thr, hammer2_xop_t *xop);
1663 void hammer2_xop_unlink(hammer2_thread_t *thr, hammer2_xop_t *xop);
1664 void hammer2_xop_nrename(hammer2_thread_t *thr, hammer2_xop_t *xop);
1665 void hammer2_xop_scanlhc(hammer2_thread_t *thr, hammer2_xop_t *xop);
1666 void hammer2_xop_scanall(hammer2_thread_t *thr, hammer2_xop_t *xop);
1667 void hammer2_xop_lookup(hammer2_thread_t *thr, hammer2_xop_t *xop);
1668 void hammer2_inode_xop_mkdirent(hammer2_thread_t *thr, hammer2_xop_t *xop);
1669 void hammer2_inode_xop_create(hammer2_thread_t *thr, hammer2_xop_t *xop);
1670 void hammer2_inode_xop_destroy(hammer2_thread_t *thr, hammer2_xop_t *xop);
1671 void hammer2_inode_xop_chain_sync(hammer2_thread_t *thr, hammer2_xop_t *xop);
1672 void hammer2_inode_xop_unlinkall(hammer2_thread_t *thr, hammer2_xop_t *xop);
1673 void hammer2_inode_xop_connect(hammer2_thread_t *thr, hammer2_xop_t *xop);
1674 void hammer2_inode_xop_flush(hammer2_thread_t *thr, hammer2_xop_t *xop);
1675
1676 /*
1677  * hammer2_msgops.c
1678  */
1679 int hammer2_msg_dbg_rcvmsg(kdmsg_msg_t *msg);
1680 int hammer2_msg_adhoc_input(kdmsg_msg_t *msg);
1681
1682 /*
1683  * hammer2_vfsops.c
1684  */
1685 void hammer2_volconf_update(hammer2_dev_t *hmp, int index);
1686 void hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx);
1687 int hammer2_vfs_sync(struct mount *mp, int waitflags);
1688 int hammer2_vfs_enospace(hammer2_inode_t *ip, off_t bytes, struct ucred *cred);
1689
1690 hammer2_pfs_t *hammer2_pfsalloc(hammer2_chain_t *chain,
1691                                 const hammer2_inode_data_t *ripdata,
1692                                 hammer2_tid_t modify_tid,
1693                                 hammer2_dev_t *force_local);
1694 void hammer2_pfsdealloc(hammer2_pfs_t *pmp, int clindex, int destroying);
1695 int hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1696                                 ino_t ino, struct vnode **vpp);
1697
1698 void hammer2_lwinprog_ref(hammer2_pfs_t *pmp);
1699 void hammer2_lwinprog_drop(hammer2_pfs_t *pmp);
1700 void hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int pipe);
1701
1702 /*
1703  * hammer2_freemap.c
1704  */
1705 int hammer2_freemap_alloc(hammer2_chain_t *chain, size_t bytes);
1706 void hammer2_freemap_adjust(hammer2_dev_t *hmp,
1707                                 hammer2_blockref_t *bref, int how);
1708
1709 /*
1710  * hammer2_cluster.c
1711  */
1712 uint8_t hammer2_cluster_type(hammer2_cluster_t *cluster);
1713 const hammer2_media_data_t *hammer2_cluster_rdata(hammer2_cluster_t *cluster);
1714 hammer2_media_data_t *hammer2_cluster_wdata(hammer2_cluster_t *cluster);
1715 hammer2_cluster_t *hammer2_cluster_from_chain(hammer2_chain_t *chain);
1716 void hammer2_cluster_bref(hammer2_cluster_t *cluster, hammer2_blockref_t *bref);
1717 hammer2_cluster_t *hammer2_cluster_alloc(hammer2_pfs_t *pmp,
1718                                 hammer2_blockref_t *bref);
1719 void hammer2_cluster_ref(hammer2_cluster_t *cluster);
1720 void hammer2_cluster_drop(hammer2_cluster_t *cluster);
1721 void hammer2_cluster_lock(hammer2_cluster_t *cluster, int how);
1722 int hammer2_cluster_check(hammer2_cluster_t *cluster, hammer2_key_t lokey,
1723                         int flags);
1724 void hammer2_cluster_resolve(hammer2_cluster_t *cluster);
1725 void hammer2_cluster_forcegood(hammer2_cluster_t *cluster);
1726 void hammer2_cluster_unlock(hammer2_cluster_t *cluster);
1727
1728 void hammer2_bulkfree_init(hammer2_dev_t *hmp);
1729 void hammer2_bulkfree_uninit(hammer2_dev_t *hmp);
1730 int hammer2_bulkfree_pass(hammer2_dev_t *hmp, hammer2_chain_t *vchain,
1731                         struct hammer2_ioc_bulkfree *bfi);
1732
1733 /*
1734  * hammer2_iocom.c
1735  */
1736 void hammer2_iocom_init(hammer2_dev_t *hmp);
1737 void hammer2_iocom_uninit(hammer2_dev_t *hmp);
1738 void hammer2_cluster_reconnect(hammer2_dev_t *hmp, struct file *fp);
1739
1740 /*
1741  * hammer2_strategy.c
1742  */
1743 int hammer2_vop_strategy(struct vop_strategy_args *ap);
1744 int hammer2_vop_bmap(struct vop_bmap_args *ap);
1745 void hammer2_write_thread(void *arg);
1746 void hammer2_bioq_sync(hammer2_pfs_t *pmp);
1747 void hammer2_dedup_clear(hammer2_dev_t *hmp);
1748
1749 #endif /* !_KERNEL */
1750 #endif /* !_VFS_HAMMER2_HAMMER2_H_ */