sys/vfs/hammer2/hammer2.h

   1 /*
   2  * Copyright (c) 2011-2015 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 #include <sys/thread2.h>
  92
  93 #include "hammer2_disk.h"
  94 #include "hammer2_mount.h"
  95 #include "hammer2_ioctl.h"
  96
  97 struct hammer2_io;
  98 struct hammer2_iocb;
  99 struct hammer2_chain;
 100 struct hammer2_cluster;
 101 struct hammer2_inode;
 102 struct hammer2_dev;
 103 struct hammer2_pfs;
 104 struct hammer2_span;
 105 struct hammer2_state;
 106 struct hammer2_msg;
 107
 108 /*
 109  * Mutex and lock shims.  Hammer2 requires support for asynchronous and
 110  * abortable locks, and both exclusive and shared spinlocks.  Normal
 111  * synchronous non-abortable locks can be substituted for spinlocks.
 112  */
 113 typedef mtx_t                           hammer2_mtx_t;
 114 typedef mtx_link_t                      hammer2_mtx_link_t;
 115 typedef mtx_state_t                     hammer2_mtx_state_t;
 116
 117 typedef struct spinlock                 hammer2_spin_t;
 118
 119 #define hammer2_mtx_ex                  mtx_lock_ex_quick
 120 #define hammer2_mtx_sh                  mtx_lock_sh_quick
 121 #define hammer2_mtx_unlock              mtx_unlock
 122 #define hammer2_mtx_owned               mtx_owned
 123 #define hammer2_mtx_init                mtx_init
 124 #define hammer2_mtx_temp_release        mtx_lock_temp_release
 125 #define hammer2_mtx_temp_restore        mtx_lock_temp_restore
 126 #define hammer2_mtx_refs                mtx_lockrefs
 127
 128 #define hammer2_spin_init               spin_init
 129 #define hammer2_spin_sh                 spin_lock_shared
 130 #define hammer2_spin_ex                 spin_lock
 131 #define hammer2_spin_unsh               spin_unlock_shared
 132 #define hammer2_spin_unex               spin_unlock
 133
 134 /*
 135  * General lock support
 136  */
 137 static __inline
 138 int
 139 hammer2_mtx_upgrade(hammer2_mtx_t *mtx)
 140 {
 141         int wasexclusive;
 142
 143         if (mtx_islocked_ex(mtx)) {
 144                 wasexclusive = 1;
 145         } else {
 146                 mtx_unlock(mtx);
 147                 mtx_lock_ex_quick(mtx);
 148                 wasexclusive = 0;
 149         }
 150         return wasexclusive;
 151 }
 152
 153 /*
 154  * Downgrade an inode lock from exclusive to shared only if the inode
 155  * lock was previously shared.  If the inode lock was previously exclusive,
 156  * this is a NOP.
 157  */
 158 static __inline
 159 void
 160 hammer2_mtx_downgrade(hammer2_mtx_t *mtx, int wasexclusive)
 161 {
 162         if (wasexclusive == 0)
 163                 mtx_downgrade(mtx);
 164 }
 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 /*
 177  * The chain structure tracks a portion of the media topology from the
 178  * root (volume) down.  Chains represent volumes, inodes, indirect blocks,
 179  * data blocks, and freemap nodes and leafs.
 180  *
 181  * The chain structure utilizes a simple singly-homed topology and the
 182  * chain's in-memory topology will move around as the chains do, due mainly
 183  * to renames and indirect block creation.
 184  *
 185  * Block Table Updates
 186  *
 187  *      Block table updates for insertions and updates are delayed until the
 188  *      flush.  This allows us to avoid having to modify the parent chain
 189  *      all the way to the root.
 190  *
 191  *      Block table deletions are performed immediately (modifying the parent
 192  *      in the process) because the flush code uses the chain structure to
 193  *      track delayed updates and the chain will be (likely) gone or moved to
 194  *      another location in the topology after a deletion.
 195  *
 196  *      A prior iteration of the code tried to keep the relationship intact
 197  *      on deletes by doing a delete-duplicate operation on the chain, but
 198  *      it added way too much complexity to the codebase.
 199  *
 200  * Flush Synchronization
 201  *
 202  *      The flush code must flush modified chains bottom-up.  Because chain
 203  *      structures can shift around and are NOT topologically stable,
 204  *      modified chains are independently indexed for the flush.  As the flush
 205  *      runs it modifies (or further modifies) and updates the parents,
 206  *      propagating the flush all the way to the volume root.
 207  *
 208  *      Modifying front-end operations can occur during a flush but will block
 209  *      in two cases: (1) when the front-end tries to operate on the inode
 210  *      currently in the midst of being flushed and (2) if the front-end
 211  *      crosses an inode currently being flushed (such as during a rename).
 212  *      So, for example, if you rename directory "x" to "a/b/c/d/e/f/g/x" and
 213  *      the flusher is currently working on "a/b/c", the rename will block
 214  *      temporarily in order to ensure that "x" exists in one place or the
 215  *      other.
 216  *
 217  *      Meta-data statistics are updated by the flusher.  The front-end will
 218  *      make estimates but meta-data must be fully synchronized only during a
 219  *      flush in order to ensure that it remains correct across a crash.
 220  *
 221  *      Multiple flush synchronizations can theoretically be in-flight at the
 222  *      same time but the implementation is not coded to handle the case and
 223  *      currently serializes them.
 224  *
 225  * Snapshots:
 226  *
 227  *      Snapshots currently require the subdirectory tree being snapshotted
 228  *      to be flushed.  The snapshot then creates a new super-root inode which
 229  *      copies the flushed blockdata of the directory or file that was
 230  *      snapshotted.
 231  *
 232  * RBTREE NOTES:
 233  *
 234  *      - Note that the radix tree runs in powers of 2 only so sub-trees
 235  *        cannot straddle edges.
 236  */
 237 RB_HEAD(hammer2_chain_tree, hammer2_chain);
 238 TAILQ_HEAD(h2_flush_list, hammer2_chain);
 239 TAILQ_HEAD(h2_core_list, hammer2_chain);
 240 TAILQ_HEAD(h2_iocb_list, hammer2_iocb);
 241
 242 #define CHAIN_CORE_DELETE_BMAP_ENTRIES  \
 243         (HAMMER2_PBUFSIZE / sizeof(hammer2_blockref_t) / sizeof(uint32_t))
 244
 245 struct hammer2_chain_core {
 246         hammer2_mtx_t   lock;
 247         hammer2_spin_t  spin;
 248         struct hammer2_chain_tree rbtree; /* sub-chains */
 249         int             live_zero;      /* blockref array opt */
 250         u_int           flags;
 251         u_int           live_count;     /* live (not deleted) chains in tree */
 252         u_int           chain_count;    /* live + deleted chains under core */
 253         int             generation;     /* generation number (inserts only) */
 254 };
 255
 256 typedef struct hammer2_chain_core hammer2_chain_core_t;
 257
 258 #define HAMMER2_CORE_UNUSED0001         0x0001
 259 #define HAMMER2_CORE_COUNTEDBREFS       0x0002
 260
 261 RB_HEAD(hammer2_io_tree, hammer2_io);
 262
 263 /*
 264  * IOCB - IO callback (into chain, cluster, or manual request)
 265  */
 266 struct hammer2_iocb {
 267         TAILQ_ENTRY(hammer2_iocb) entry;
 268         void (*callback)(struct hammer2_iocb *iocb);
 269         struct hammer2_io       *dio;
 270         struct hammer2_cluster  *cluster;
 271         struct hammer2_chain    *chain;
 272         void                    *ptr;
 273         off_t                   lbase;
 274         int                     lsize;
 275         uint32_t                flags;
 276         int                     error;
 277 };
 278
 279 typedef struct hammer2_iocb hammer2_iocb_t;
 280
 281 #define HAMMER2_IOCB_INTERLOCK  0x00000001
 282 #define HAMMER2_IOCB_ONQ        0x00000002
 283 #define HAMMER2_IOCB_DONE       0x00000004
 284 #define HAMMER2_IOCB_INPROG     0x00000008
 285 #define HAMMER2_IOCB_UNUSED10   0x00000010
 286 #define HAMMER2_IOCB_QUICK      0x00010000
 287 #define HAMMER2_IOCB_ZERO       0x00020000
 288 #define HAMMER2_IOCB_READ       0x00040000
 289 #define HAMMER2_IOCB_WAKEUP     0x00080000
 290
 291 /*
 292  * DIO - Management structure wrapping system buffer cache.
 293  *
 294  *       Used for multiple purposes including concurrent management
 295  *       if small requests by chains into larger DIOs.
 296  */
 297 struct hammer2_io {
 298         RB_ENTRY(hammer2_io) rbnode;    /* indexed by device offset */
 299         struct h2_iocb_list iocbq;
 300         struct spinlock spin;
 301         struct hammer2_dev *hmp;
 302         struct buf      *bp;
 303         off_t           pbase;
 304         int             psize;
 305         int             refs;
 306         int             act;                    /* activity */
 307 };
 308
 309 typedef struct hammer2_io hammer2_io_t;
 310
 311 #define HAMMER2_DIO_INPROG      0x80000000      /* bio in progress */
 312 #define HAMMER2_DIO_GOOD        0x40000000      /* dio->bp is stable */
 313 #define HAMMER2_DIO_WAITING     0x20000000      /* (old) */
 314 #define HAMMER2_DIO_DIRTY       0x10000000      /* flush on last drop */
 315
 316 #define HAMMER2_DIO_MASK        0x0FFFFFFF
 317
 318 /*
 319  * Primary chain structure keeps track of the topology in-memory.
 320  */
 321 struct hammer2_chain {
 322         hammer2_chain_core_t    core;
 323         RB_ENTRY(hammer2_chain) rbnode;         /* live chain(s) */
 324         hammer2_blockref_t      bref;
 325         struct hammer2_chain    *parent;
 326         struct hammer2_state    *state;         /* if active cache msg */
 327         struct hammer2_dev      *hmp;
 328         struct hammer2_pfs      *pmp;           /* A PFS or super-root (spmp) */
 329
 330         hammer2_xid_t   flush_xid;              /* flush sequencing */
 331         hammer2_key_t   data_count;             /* delta's to apply */
 332         hammer2_key_t   inode_count;            /* delta's to apply */
 333         hammer2_key_t   data_count_up;          /* delta's to apply */
 334         hammer2_key_t   inode_count_up;         /* delta's to apply */
 335         hammer2_io_t    *dio;                   /* physical data buffer */
 336         u_int           bytes;                  /* physical data size */
 337         u_int           flags;
 338         u_int           refs;
 339         u_int           lockcnt;
 340         int             error;                  /* on-lock data error state */
 341
 342         hammer2_media_data_t *data;             /* data pointer shortcut */
 343         TAILQ_ENTRY(hammer2_chain) flush_node;  /* flush list */
 344 };
 345
 346 typedef struct hammer2_chain hammer2_chain_t;
 347
 348 int hammer2_chain_cmp(hammer2_chain_t *chain1, hammer2_chain_t *chain2);
 349 RB_PROTOTYPE(hammer2_chain_tree, hammer2_chain, rbnode, hammer2_chain_cmp);
 350
 351 /*
 352  * Special notes on flags:
 353  *
 354  * INITIAL      - This flag allows a chain to be created and for storage to
 355  *                be allocated without having to immediately instantiate the
 356  *                related buffer.  The data is assumed to be all-zeros.  It
 357  *                is primarily used for indirect blocks.
 358  *
 359  * MODIFIED     - The chain's media data has been modified.
 360  *
 361  * UPDATE       - Chain might not be modified but parent blocktable needs update
 362  *
 363  * FICTITIOUS   - Faked chain as a placeholder for an error condition.  This
 364  *                chain is unsuitable for I/O.
 365  *
 366  * BMAPPED      - Indicates that the chain is present in the parent blockmap.
 367  *
 368  * BMAPUPD      - Indicates that the chain is present but needs to be updated
 369  *                in the parent blockmap.
 370  */
 371 #define HAMMER2_CHAIN_MODIFIED          0x00000001      /* dirty chain data */
 372 #define HAMMER2_CHAIN_ALLOCATED         0x00000002      /* kmalloc'd chain */
 373 #define HAMMER2_CHAIN_DESTROY           0x00000004
 374 #define HAMMER2_CHAIN_UNLINKED          0x00000008      /* unlinked file */
 375 #define HAMMER2_CHAIN_DELETED           0x00000010      /* deleted chain */
 376 #define HAMMER2_CHAIN_INITIAL           0x00000020      /* initial create */
 377 #define HAMMER2_CHAIN_UPDATE            0x00000040      /* need parent update */
 378 #define HAMMER2_CHAIN_DEFERRED          0x00000080      /* flush depth defer */
 379 #define HAMMER2_CHAIN_IOFLUSH           0x00000100      /* bawrite on put */
 380 #define HAMMER2_CHAIN_ONFLUSH           0x00000200      /* on a flush list */
 381 #define HAMMER2_CHAIN_FICTITIOUS        0x00000400      /* unsuitable for I/O */
 382 #define HAMMER2_CHAIN_VOLUMESYNC        0x00000800      /* needs volume sync */
 383 #define HAMMER2_CHAIN_UNUSED00001000    0x00001000
 384 #define HAMMER2_CHAIN_UNUSED00002000    0x00002000
 385 #define HAMMER2_CHAIN_ONRBTREE          0x00004000      /* on parent RB tree */
 386 #define HAMMER2_CHAIN_UNUSED00008000    0x00008000
 387 #define HAMMER2_CHAIN_EMBEDDED          0x00010000      /* embedded data */
 388 #define HAMMER2_CHAIN_RELEASE           0x00020000      /* don't keep around */
 389 #define HAMMER2_CHAIN_BMAPPED           0x00040000      /* present in blkmap */
 390 #define HAMMER2_CHAIN_BMAPUPD           0x00080000      /* +needs updating */
 391 #define HAMMER2_CHAIN_UNUSED00100000    0x00100000
 392 #define HAMMER2_CHAIN_UNUSED00200000    0x00200000
 393 #define HAMMER2_CHAIN_PFSBOUNDARY       0x00400000      /* super->pfs inode */
 394
 395 #define HAMMER2_CHAIN_FLUSH_MASK        (HAMMER2_CHAIN_MODIFIED |       \
 396                                          HAMMER2_CHAIN_UPDATE |         \
 397                                          HAMMER2_CHAIN_ONFLUSH)
 398
 399 /*
 400  * Hammer2 error codes, used by chain->error and cluster->error.  The error
 401  * code is typically set on-lock unless no I/O was requested, and set on
 402  * I/O otherwise.  If set for a cluster it generally means that the cluster
 403  * code could not find a valid copy to present.
 404  *
 405  * IO           - An I/O error occurred
 406  * CHECK        - I/O succeeded but did not match the check code
 407  * INCOMPLETE   - A cluster is not complete enough to use, or
 408  *                a chain cannot be loaded because its parent has an error.
 409  *
 410  * NOTE: API allows callers to check zero/non-zero to determine if an error
 411  *       condition exists.
 412  *
 413  * NOTE: Chain's data field is usually NULL on an IO error but not necessarily
 414  *       NULL on other errors.  Check chain->error, not chain->data.
 415  */
 416 #define HAMMER2_ERROR_NONE              0
 417 #define HAMMER2_ERROR_IO                1       /* device I/O error */
 418 #define HAMMER2_ERROR_CHECK             2       /* check code mismatch */
 419 #define HAMMER2_ERROR_INCOMPLETE        3       /* incomplete cluster */
 420 #define HAMMER2_ERROR_DEPTH             4       /* temporary depth limit */
 421
 422 /*
 423  * Flags passed to hammer2_chain_lookup() and hammer2_chain_next()
 424  *
 425  * NOTE: MATCHIND allows an indirect block / freemap node to be returned
 426  *       when the passed key range matches the radix.  Remember that key_end
 427  *       is inclusive (e.g. {0x000,0xFFF}, not {0x000,0x1000}).
 428  *
 429  * NOTE: NODIRECT prevents a lookup of offset 0 in an inode from returning
 430  *       the inode itself if the inode is in DIRECTDATA mode (i.e. file is
 431  *       <= 512 bytes).
 432  */
 433 #define HAMMER2_LOOKUP_NOLOCK           0x00000001      /* ref only */
 434 #define HAMMER2_LOOKUP_NODATA           0x00000002      /* data left NULL */
 435 #define HAMMER2_LOOKUP_NODIRECT         0x00000004      /* no offset=0 DD */
 436 #define HAMMER2_LOOKUP_SHARED           0x00000100
 437 #define HAMMER2_LOOKUP_MATCHIND         0x00000200      /* return all chains */
 438 #define HAMMER2_LOOKUP_ALLNODES         0x00000400      /* allow NULL focus */
 439 #define HAMMER2_LOOKUP_ALWAYS           0x00000800      /* resolve data */
 440
 441 /*
 442  * Flags passed to hammer2_chain_modify() and hammer2_chain_resize()
 443  *
 444  * NOTE: OPTDATA allows us to avoid instantiating buffers for INDIRECT
 445  *       blocks in the INITIAL-create state.
 446  */
 447 #define HAMMER2_MODIFY_OPTDATA          0x00000002      /* data can be NULL */
 448 #define HAMMER2_MODIFY_NO_MODIFY_TID    0x00000004
 449 #define HAMMER2_MODIFY_UNUSED0008       0x00000008
 450 #define HAMMER2_MODIFY_NOREALLOC        0x00000010
 451
 452 /*
 453  * Flags passed to hammer2_chain_lock()
 454  *
 455  * NOTE: RDONLY is set to optimize cluster operations when *no* modifications
 456  *       will be made to either the cluster being locked or any underlying
 457  *       cluster.  It allows the cluster to lock and access data for a subset
 458  *       of available nodes instead of all available nodes.
 459  */
 460 #define HAMMER2_RESOLVE_NEVER           1
 461 #define HAMMER2_RESOLVE_MAYBE           2
 462 #define HAMMER2_RESOLVE_ALWAYS          3
 463 #define HAMMER2_RESOLVE_MASK            0x0F
 464
 465 #define HAMMER2_RESOLVE_SHARED          0x10    /* request shared lock */
 466 #define HAMMER2_RESOLVE_UNUSED20        0x20
 467 #define HAMMER2_RESOLVE_RDONLY          0x40    /* higher level op flag */
 468
 469 /*
 470  * Flags passed to hammer2_chain_delete()
 471  */
 472 #define HAMMER2_DELETE_PERMANENT        0x0001
 473 #define HAMMER2_DELETE_NOSTATS          0x0002
 474
 475 #define HAMMER2_INSERT_NOSTATS          0x0002
 476 #define HAMMER2_INSERT_PFSROOT          0x0004
 477
 478 /*
 479  * Flags passed to hammer2_chain_delete_duplicate()
 480  */
 481 #define HAMMER2_DELDUP_RECORE           0x0001
 482
 483 /*
 484  * Cluster different types of storage together for allocations
 485  */
 486 #define HAMMER2_FREECACHE_INODE         0
 487 #define HAMMER2_FREECACHE_INDIR         1
 488 #define HAMMER2_FREECACHE_DATA          2
 489 #define HAMMER2_FREECACHE_UNUSED3       3
 490 #define HAMMER2_FREECACHE_TYPES         4
 491
 492 /*
 493  * hammer2_freemap_alloc() block preference
 494  */
 495 #define HAMMER2_OFF_NOPREF              ((hammer2_off_t)-1)
 496
 497 /*
 498  * BMAP read-ahead maximum parameters
 499  */
 500 #define HAMMER2_BMAP_COUNT              16      /* max bmap read-ahead */
 501 #define HAMMER2_BMAP_BYTES              (HAMMER2_PBUFSIZE * HAMMER2_BMAP_COUNT)
 502
 503 /*
 504  * hammer2_freemap_adjust()
 505  */
 506 #define HAMMER2_FREEMAP_DORECOVER       1
 507 #define HAMMER2_FREEMAP_DOMAYFREE       2
 508 #define HAMMER2_FREEMAP_DOREALFREE      3
 509
 510 /*
 511  * HAMMER2 cluster - A set of chains representing the same entity.
 512  *
 513  * hammer2_cluster typically represents a temporary set of representitive
 514  * chains.  The one exception is that a hammer2_cluster is embedded in
 515  * hammer2_inode.  This embedded cluster is ONLY used to track the
 516  * representitive chains and cannot be directly locked.
 517  *
 518  * A cluster is usually temporary (and thus per-thread) for locking purposes,
 519  * allowing us to embed the asynchronous storage required for cluster
 520  * operations in the cluster itself and adjust the state and status without
 521  * having to worry too much about SMP issues.
 522  *
 523  * The exception is the cluster embedded in the hammer2_inode structure.
 524  * This is used to cache the cluster state on an inode-by-inode basis.
 525  * Individual hammer2_chain structures not incorporated into clusters might
 526  * also stick around to cache miscellanious elements.
 527  *
 528  * Because the cluster is a 'working copy' and is usually subject to cluster
 529  * quorum rules, it is quite possible for us to end up with an insufficient
 530  * number of live chains to execute an operation.  If an insufficient number
 531  * of chains remain in a working copy, the operation may have to be
 532  * downgraded, retried, stall until the requisit number of chains are
 533  * available, or possibly even error out depending on the mount type.
 534  *
 535  * A cluster's focus is set when it is locked.  The focus can only be set
 536  * to a chain still part of the synchronized set.
 537  */
 538 #define HAMMER2_MAXCLUSTER      8
 539
 540 struct hammer2_cluster_item {
 541 #if 0
 542         hammer2_mtx_link_t      async_link;
 543 #endif
 544         hammer2_chain_t         *chain;
 545 #if 0
 546         struct hammer2_cluster  *cluster;       /* link back to cluster */
 547 #endif
 548         int                     cache_index;
 549         uint32_t                flags;
 550 };
 551
 552 typedef struct hammer2_cluster_item hammer2_cluster_item_t;
 553
 554 /*
 555  * INVALID      - Invalid for focus, i.e. not part of synchronized set.
 556  *                Once set, this bit is sticky across operations.
 557  *
 558  * FEMOD        - Indicates that front-end modifying operations can
 559  *                mess with this entry and MODSYNC will copy also
 560  *                effect it.
 561  */
 562 #define HAMMER2_CITEM_INVALID   0x00000001
 563 #define HAMMER2_CITEM_FEMOD     0x00000002
 564
 565 struct hammer2_cluster {
 566         int                     refs;           /* track for deallocation */
 567         int                     ddflag;
 568         struct hammer2_pfs      *pmp;
 569         uint32_t                flags;
 570         int                     nchains;
 571         int                     error;          /* error code valid on lock */
 572         int                     focus_index;
 573         hammer2_iocb_t          iocb;
 574         hammer2_chain_t         *focus;         /* current focus (or mod) */
 575         hammer2_cluster_item_t  array[HAMMER2_MAXCLUSTER];
 576 };
 577
 578 typedef struct hammer2_cluster  hammer2_cluster_t;
 579
 580 /*
 581  * WRHARD       - Hard mounts can write fully synchronized
 582  * RDHARD       - Hard mounts can read fully synchronized
 583  * UNHARD       - Unsynchronized masters present
 584  * NOHARD       - No masters visible
 585  * WRSOFT       - Soft mounts can write to at least the SOFT_MASTER
 586  * RDSOFT       - Soft mounts can read from at least a SOFT_SLAVE
 587  * UNSOFT       - Unsynchronized slaves present
 588  * NOSOFT       - No slaves visible
 589  * RDSLAVE      - slaves are accessible (possibly unsynchronized or remote).
 590  * MSYNCED      - All masters are fully synchronized
 591  * SSYNCED      - All known local slaves are fully synchronized to masters
 592  *
 593  * All available masters are always incorporated.  All PFSs belonging to a
 594  * cluster (master, slave, copy, whatever) always try to synchronize the
 595  * total number of known masters in the PFSs root inode.
 596  *
 597  * A cluster might have access to many slaves, copies, or caches, but we
 598  * have a limited number of cluster slots.  Any such elements which are
 599  * directly mounted from block device(s) will always be incorporated.   Note
 600  * that SSYNCED only applies to such elements which are directly mounted,
 601  * not to any remote slaves, copies, or caches that could be available.  These
 602  * bits are used to monitor and drive our synchronization threads.
 603  *
 604  * When asking the question 'is any data accessible at all', then a simple
 605  * test against (RDHARD|RDSOFT|RDSLAVE) gives you the answer.  If any of
 606  * these bits are set the object can be read with certain caveats:
 607  * RDHARD - no caveats.  RDSOFT - authoritative but might not be synchronized.
 608  * and RDSLAVE - not authoritative, has some data but it could be old or
 609  * incomplete.
 610  *
 611  * When both soft and hard mounts are available, data will be read and written
 612  * via the soft mount only.  But all might be in the cluster because
 613  * background synchronization threads still need to do their work.
 614  */
 615 #define HAMMER2_CLUSTER_INODE   0x00000001      /* embedded in inode struct */
 616 #define HAMMER2_CLUSTER_UNUSED2 0x00000002
 617 #define HAMMER2_CLUSTER_LOCKED  0x00000004      /* cluster lks not recursive */
 618 #define HAMMER2_CLUSTER_WRHARD  0x00000100      /* hard-mount can write */
 619 #define HAMMER2_CLUSTER_RDHARD  0x00000200      /* hard-mount can read */
 620 #define HAMMER2_CLUSTER_UNHARD  0x00000400      /* unsynchronized masters */
 621 #define HAMMER2_CLUSTER_NOHARD  0x00000800      /* no masters visible */
 622 #define HAMMER2_CLUSTER_WRSOFT  0x00001000      /* soft-mount can write */
 623 #define HAMMER2_CLUSTER_RDSOFT  0x00002000      /* soft-mount can read */
 624 #define HAMMER2_CLUSTER_UNSOFT  0x00004000      /* unsynchronized slaves */
 625 #define HAMMER2_CLUSTER_NOSOFT  0x00008000      /* no slaves visible */
 626 #define HAMMER2_CLUSTER_MSYNCED 0x00010000      /* all masters synchronized */
 627 #define HAMMER2_CLUSTER_SSYNCED 0x00020000      /* known slaves synchronized */
 628
 629 #define HAMMER2_CLUSTER_ANYDATA ( HAMMER2_CLUSTER_RDHARD |      \
 630                                   HAMMER2_CLUSTER_RDSOFT |      \
 631                                   HAMMER2_CLUSTER_RDSLAVE)
 632
 633 #define HAMMER2_CLUSTER_RDOK    ( HAMMER2_CLUSTER_RDHARD |      \
 634                                   HAMMER2_CLUSTER_RDSOFT)
 635
 636 #define HAMMER2_CLUSTER_WROK    ( HAMMER2_CLUSTER_WRHARD |      \
 637                                   HAMMER2_CLUSTER_WRSOFT)
 638
 639 #define HAMMER2_CLUSTER_ZFLAGS  ( HAMMER2_CLUSTER_WRHARD |      \
 640                                   HAMMER2_CLUSTER_RDHARD |      \
 641                                   HAMMER2_CLUSTER_WRSOFT |      \
 642                                   HAMMER2_CLUSTER_RDSOFT |      \
 643                                   HAMMER2_CLUSTER_MSYNCED |     \
 644                                   HAMMER2_CLUSTER_SSYNCED)
 645
 646 /*
 647  * Helper functions (cluster must be locked for flags to be valid).
 648  */
 649 static __inline
 650 int
 651 hammer2_cluster_rdok(hammer2_cluster_t *cluster)
 652 {
 653         return (cluster->flags & HAMMER2_CLUSTER_RDOK);
 654 }
 655
 656 static __inline
 657 int
 658 hammer2_cluster_wrok(hammer2_cluster_t *cluster)
 659 {
 660         return (cluster->flags & HAMMER2_CLUSTER_WROK);
 661 }
 662
 663 RB_HEAD(hammer2_inode_tree, hammer2_inode);
 664
 665 /*
 666  * A hammer2 inode.
 667  *
 668  * NOTE: The inode-embedded cluster is never used directly for I/O (since
 669  *       it may be shared).  Instead it will be replicated-in and synchronized
 670  *       back out if changed.
 671  */
 672 struct hammer2_inode {
 673         RB_ENTRY(hammer2_inode) rbnode;         /* inumber lookup (HL) */
 674         hammer2_mtx_t           lock;           /* inode lock */
 675         struct hammer2_pfs      *pmp;           /* PFS mount */
 676         struct hammer2_inode    *pip;           /* parent inode */
 677         struct vnode            *vp;
 678         hammer2_cluster_t       cluster;
 679         struct lockf            advlock;
 680         hammer2_tid_t           inum;
 681         u_int                   flags;
 682         u_int                   refs;           /* +vpref, +flushref */
 683         uint8_t                 comp_heuristic;
 684         hammer2_off_t           size;
 685         uint64_t                mtime;
 686 };
 687
 688 typedef struct hammer2_inode hammer2_inode_t;
 689
 690 #define HAMMER2_INODE_MODIFIED          0x0001
 691 #define HAMMER2_INODE_SROOT             0x0002  /* kmalloc special case */
 692 #define HAMMER2_INODE_RENAME_INPROG     0x0004
 693 #define HAMMER2_INODE_ONRBTREE          0x0008
 694 #define HAMMER2_INODE_RESIZED           0x0010
 695 #define HAMMER2_INODE_MTIME             0x0020
 696
 697 int hammer2_inode_cmp(hammer2_inode_t *ip1, hammer2_inode_t *ip2);
 698 RB_PROTOTYPE2(hammer2_inode_tree, hammer2_inode, rbnode, hammer2_inode_cmp,
 699                 hammer2_tid_t);
 700
 701 /*
 702  * inode-unlink side-structure
 703  */
 704 struct hammer2_inode_unlink {
 705         TAILQ_ENTRY(hammer2_inode_unlink) entry;
 706         hammer2_inode_t *ip;
 707 };
 708 TAILQ_HEAD(h2_unlk_list, hammer2_inode_unlink);
 709
 710 typedef struct hammer2_inode_unlink hammer2_inode_unlink_t;
 711
 712 /*
 713  * A hammer2 transaction and flush sequencing structure.
 714  *
 715  * This global structure is tied into hammer2_dev and is used
 716  * to sequence modifying operations and flushes.  These operations
 717  * run on whole cluster PFSs, not individual nodes (at this level),
 718  * so we do not record mirror_tid here.
 719  */
 720 struct hammer2_trans {
 721         TAILQ_ENTRY(hammer2_trans) entry;
 722         struct hammer2_pfs      *pmp;
 723         hammer2_xid_t           sync_xid;       /* transaction sequencer */
 724         hammer2_tid_t           inode_tid;      /* inode number assignment */
 725         hammer2_tid_t           modify_tid;     /* modify transaction id */
 726         thread_t                td;             /* pointer */
 727         int                     flags;
 728         int                     blocked;
 729         uint8_t                 inodes_created;
 730         uint8_t                 dummy[7];
 731 };
 732
 733 typedef struct hammer2_trans hammer2_trans_t;
 734
 735 #define HAMMER2_TRANS_ISFLUSH           0x0001  /* formal flush */
 736 #define HAMMER2_TRANS_CONCURRENT        0x0002  /* concurrent w/flush */
 737 #define HAMMER2_TRANS_BUFCACHE          0x0004  /* from bioq strategy write */
 738 #define HAMMER2_TRANS_NEWINODE          0x0008  /* caller allocating inode */
 739 #define HAMMER2_TRANS_KEEPMODIFY        0x0010  /* do not change bref.modify */
 740 #define HAMMER2_TRANS_PREFLUSH          0x0020  /* preflush state */
 741
 742 #define HAMMER2_FREEMAP_HEUR_NRADIX     4       /* pwr 2 PBUFRADIX-MINIORADIX */
 743 #define HAMMER2_FREEMAP_HEUR_TYPES      8
 744 #define HAMMER2_FREEMAP_HEUR            (HAMMER2_FREEMAP_HEUR_NRADIX * \
 745                                          HAMMER2_FREEMAP_HEUR_TYPES)
 746
 747 /*
 748  * Transaction Rendezvous
 749  */
 750 TAILQ_HEAD(hammer2_trans_queue, hammer2_trans);
 751
 752 struct hammer2_trans_manage {
 753         hammer2_xid_t           flush_xid;      /* last flush transaction */
 754         hammer2_xid_t           alloc_xid;
 755         struct lock             translk;        /* lockmgr lock */
 756         struct hammer2_trans_queue transq;      /* modifying transactions */
 757         int                     flushcnt;       /* track flush trans */
 758 };
 759
 760 typedef struct hammer2_trans_manage hammer2_trans_manage_t;
 761
 762 /*
 763  * Cluster node synchronization thread element.
 764  *
 765  * Multiple syncthr's can hang off of a hammer2_pfs structure, typically one
 766  * for each block device that is part of the PFS.  Synchronization threads
 767  * for PFSs accessed over the network are handled by their respective hosts.
 768  *
 769  * Synchronization threads are responsible for keeping a local node
 770  * synchronized to the greater cluster.
 771  *
 772  * A syncthr can also hang off each hammer2_dev's super-root PFS (spmp).
 773  * This thread is responsible for automatic bulkfree and dedup scans.
 774  */
 775 struct hammer2_syncthr {
 776         struct hammer2_pfs *pmp;
 777         kdmsg_state_t   *span;
 778         thread_t        td;
 779         uint32_t        flags;
 780         int             depth;
 781         int             clindex;        /* sync_thrs[] array index */
 782         hammer2_trans_t trans;
 783         struct lock     lk;
 784 };
 785
 786 typedef struct hammer2_syncthr hammer2_syncthr_t;
 787
 788 #define HAMMER2_SYNCTHR_UNMOUNTING      0x0001  /* unmount request */
 789 #define HAMMER2_SYNCTHR_DEV             0x0002  /* related to dev, not pfs */
 790 #define HAMMER2_SYNCTHR_SPANNED         0x0004  /* LNK_SPAN active */
 791 #define HAMMER2_SYNCTHR_REMASTER        0x0008  /* remaster request */
 792 #define HAMMER2_SYNCTHR_STOP            0x0010  /* exit request */
 793 #define HAMMER2_SYNCTHR_FREEZE          0x0020  /* force idle */
 794 #define HAMMER2_SYNCTHR_FROZEN          0x0040  /* restart */
 795
 796
 797 /*
 798  * Global (per partition) management structure, represents a hard block
 799  * device.  Typically referenced by hammer2_chain structures when applicable.
 800  * Typically not used for network-managed elements.
 801  *
 802  * Note that a single hammer2_dev can be indirectly tied to multiple system
 803  * mount points.  There is no direct relationship.  System mounts are
 804  * per-cluster-id, not per-block-device, and a single hard mount might contain
 805  * many PFSs and those PFSs might combine together in various ways to form
 806  * the set of available clusters.
 807  */
 808 struct hammer2_dev {
 809         struct vnode    *devvp;         /* device vnode */
 810         int             ronly;          /* read-only mount */
 811         int             mount_count;    /* number of actively mounted PFSs */
 812         TAILQ_ENTRY(hammer2_dev) mntentry; /* hammer2_mntlist */
 813
 814         struct malloc_type *mchain;
 815         int             nipstacks;
 816         int             maxipstacks;
 817         kdmsg_iocom_t   iocom;          /* volume-level dmsg interface */
 818         struct spinlock io_spin;        /* iotree access */
 819         struct hammer2_io_tree iotree;
 820         int             iofree_count;
 821         hammer2_chain_t vchain;         /* anchor chain (topology) */
 822         hammer2_chain_t fchain;         /* anchor chain (freemap) */
 823         struct spinlock list_spin;
 824         struct h2_flush_list    flushq; /* flush seeds */
 825         struct hammer2_pfs *spmp;       /* super-root pmp for transactions */
 826         struct lock     vollk;          /* lockmgr lock */
 827         hammer2_off_t   heur_freemap[HAMMER2_FREEMAP_HEUR];
 828         int             volhdrno;       /* last volhdrno written */
 829         char            devrepname[64]; /* for kprintf */
 830         hammer2_volume_data_t voldata;
 831         hammer2_volume_data_t volsync;  /* synchronized voldata */
 832 };
 833
 834 typedef struct hammer2_dev hammer2_dev_t;
 835
 836 /*
 837  * Helper functions (cluster must be locked for flags to be valid).
 838  */
 839 static __inline
 840 int
 841 hammer2_chain_rdok(hammer2_chain_t *chain)
 842 {
 843         return (chain->error == 0);
 844 }
 845
 846 static __inline
 847 int
 848 hammer2_chain_wrok(hammer2_chain_t *chain)
 849 {
 850         return (chain->error == 0 && chain->hmp->ronly == 0);
 851 }
 852
 853 /*
 854  * Per-cluster management structure.  This structure will be tied to a
 855  * system mount point if the system is mounting the PFS, but is also used
 856  * to manage clusters encountered during the super-root scan or received
 857  * via LNK_SPANs that might not be mounted.
 858  *
 859  * This structure is also used to represent the super-root that hangs off
 860  * of a hard mount point.  The super-root is not really a cluster element.
 861  * In this case the spmp_hmp field will be non-NULL.  It's just easier to do
 862  * this than to special case super-root manipulation in the hammer2_chain*
 863  * code as being only hammer2_dev-related.
 864  *
 865  * pfs_mode and pfs_nmasters are rollup fields which critically describes
 866  * how elements of the cluster act on the cluster.  pfs_mode is only applicable
 867  * when a PFS is mounted by the system.  pfs_nmasters is our best guess as to
 868  * how many masters have been configured for a cluster and is always
 869  * applicable.  pfs_types[] is an array with 1:1 correspondance to the
 870  * iroot cluster and describes the PFS types of the nodes making up the
 871  * cluster.
 872  *
 873  * WARNING! Portions of this structure have deferred initialization.  In
 874  *          particular, if not mounted there will be no ihidden or wthread.
 875  *          umounted network PFSs will also be missing iroot and numerous
 876  *          other fields will not be initialized prior to mount.
 877  *
 878  *          Synchronization threads are chain-specific and only applicable
 879  *          to local hard PFS entries.  A hammer2_pfs structure may contain
 880  *          more than one when multiple hard PFSs are present on the local
 881  *          machine which require synchronization monitoring.  Most PFSs
 882  *          (such as snapshots) are 1xMASTER PFSs which do not need a
 883  *          synchronization thread.
 884  *
 885  * WARNING! The chains making up pfs->iroot's cluster are accounted for in
 886  *          hammer2_dev->mount_count when the pfs is associated with a mount
 887  *          point.
 888  */
 889 struct hammer2_pfs {
 890         struct mount            *mp;
 891         TAILQ_ENTRY(hammer2_pfs) mntentry;      /* hammer2_pfslist */
 892         uuid_t                  pfs_clid;
 893         hammer2_dev_t           *spmp_hmp;      /* only if super-root pmp */
 894         hammer2_inode_t         *iroot;         /* PFS root inode */
 895         hammer2_inode_t         *ihidden;       /* PFS hidden directory */
 896         uint8_t                 pfs_types[HAMMER2_MAXCLUSTER];
 897         char                    *pfs_names[HAMMER2_MAXCLUSTER];
 898         hammer2_trans_manage_t  tmanage;        /* transaction management */
 899         struct lock             lock;           /* PFS lock for certain ops */
 900         hammer2_off_t           inode_count;    /* copy of inode_count */
 901         struct netexport        export;         /* nfs export */
 902         int                     ronly;          /* read-only mount */
 903         struct malloc_type      *minode;
 904         struct malloc_type      *mmsg;
 905         struct spinlock         inum_spin;      /* inumber lookup */
 906         struct hammer2_inode_tree inum_tree;    /* (not applicable to spmp) */
 907         hammer2_tid_t           modify_tid;     /* modify transaction id */
 908         hammer2_tid_t           inode_tid;      /* inode allocator */
 909         uint8_t                 pfs_nmasters;   /* total masters */
 910         uint8_t                 pfs_mode;       /* operating mode PFSMODE */
 911         uint8_t                 unused01;
 912         uint8_t                 unused02;
 913         uint32_t                unused03;
 914         long                    inmem_inodes;
 915         uint32_t                inmem_dirty_chains;
 916         int                     count_lwinprog; /* logical write in prog */
 917         struct spinlock         list_spin;
 918         struct h2_unlk_list     unlinkq;        /* last-close unlink */
 919         hammer2_syncthr_t       sync_thrs[HAMMER2_MAXCLUSTER];
 920         thread_t                wthread_td;     /* write thread td */
 921         struct bio_queue_head   wthread_bioq;   /* logical buffer bioq */
 922         hammer2_mtx_t           wthread_mtx;    /* interlock */
 923         int                     wthread_destroy;/* termination sequencing */
 924         uint32_t                flags;          /* cached cluster flags */
 925 };
 926
 927 typedef struct hammer2_pfs hammer2_pfs_t;
 928
 929 #define HAMMER2_DIRTYCHAIN_WAITING      0x80000000
 930 #define HAMMER2_DIRTYCHAIN_MASK         0x7FFFFFFF
 931
 932 #define HAMMER2_LWINPROG_WAITING        0x80000000
 933 #define HAMMER2_LWINPROG_MASK           0x7FFFFFFF
 934
 935 /*
 936  * Bulkscan
 937  */
 938 #define HAMMER2_BULK_ABORT      0x00000001
 939
 940 /*
 941  * Misc
 942  */
 943 #if defined(_KERNEL)
 944
 945 MALLOC_DECLARE(M_HAMMER2);
 946
 947 #define VTOI(vp)        ((hammer2_inode_t *)(vp)->v_data)
 948 #define ITOV(ip)        ((ip)->vp)
 949
 950 /*
 951  * Currently locked chains retain the locked buffer cache buffer for
 952  * indirect blocks, and indirect blocks can be one of two sizes.  The
 953  * device buffer has to match the case to avoid deadlocking recursive
 954  * chains that might otherwise try to access different offsets within
 955  * the same device buffer.
 956  */
 957 static __inline
 958 int
 959 hammer2_devblkradix(int radix)
 960 {
 961 #if 0
 962         if (radix <= HAMMER2_LBUFRADIX) {
 963                 return (HAMMER2_LBUFRADIX);
 964         } else {
 965                 return (HAMMER2_PBUFRADIX);
 966         }
 967 #endif
 968         return (HAMMER2_PBUFRADIX);
 969 }
 970
 971 /*
 972  * XXX almost time to remove this.  DIO uses PBUFSIZE exclusively now.
 973  */
 974 static __inline
 975 size_t
 976 hammer2_devblksize(size_t bytes)
 977 {
 978 #if 0
 979         if (bytes <= HAMMER2_LBUFSIZE) {
 980                 return(HAMMER2_LBUFSIZE);
 981         } else {
 982                 KKASSERT(bytes <= HAMMER2_PBUFSIZE &&
 983                          (bytes ^ (bytes - 1)) == ((bytes << 1) - 1));
 984                 return (HAMMER2_PBUFSIZE);
 985         }
 986 #endif
 987         return (HAMMER2_PBUFSIZE);
 988 }
 989
 990
 991 static __inline
 992 hammer2_pfs_t *
 993 MPTOPMP(struct mount *mp)
 994 {
 995         return ((hammer2_pfs_t *)mp->mnt_data);
 996 }
 997
 998 #define LOCKSTART       int __nlocks = curthread->td_locks
 999 #define LOCKENTER       (++curthread->td_locks)
1000 #define LOCKEXIT        (--curthread->td_locks)
1001 #define LOCKSTOP        KKASSERT(curthread->td_locks == __nlocks)
1002
1003 extern struct vop_ops hammer2_vnode_vops;
1004 extern struct vop_ops hammer2_spec_vops;
1005 extern struct vop_ops hammer2_fifo_vops;
1006
1007 extern int hammer2_debug;
1008 extern int hammer2_cluster_enable;
1009 extern int hammer2_hardlink_enable;
1010 extern int hammer2_flush_pipe;
1011 extern int hammer2_synchronous_flush;
1012 extern int hammer2_dio_count;
1013 extern long hammer2_limit_dirty_chains;
1014 extern long hammer2_iod_file_read;
1015 extern long hammer2_iod_meta_read;
1016 extern long hammer2_iod_indr_read;
1017 extern long hammer2_iod_fmap_read;
1018 extern long hammer2_iod_volu_read;
1019 extern long hammer2_iod_file_write;
1020 extern long hammer2_iod_meta_write;
1021 extern long hammer2_iod_indr_write;
1022 extern long hammer2_iod_fmap_write;
1023 extern long hammer2_iod_volu_write;
1024 extern long hammer2_ioa_file_read;
1025 extern long hammer2_ioa_meta_read;
1026 extern long hammer2_ioa_indr_read;
1027 extern long hammer2_ioa_fmap_read;
1028 extern long hammer2_ioa_volu_read;
1029 extern long hammer2_ioa_file_write;
1030 extern long hammer2_ioa_meta_write;
1031 extern long hammer2_ioa_indr_write;
1032 extern long hammer2_ioa_fmap_write;
1033 extern long hammer2_ioa_volu_write;
1034
1035 extern struct objcache *cache_buffer_read;
1036 extern struct objcache *cache_buffer_write;
1037
1038 extern int destroy;
1039 extern int write_thread_wakeup;
1040
1041 /*
1042  * hammer2_subr.c
1043  */
1044 #define hammer2_icrc32(buf, size)       iscsi_crc32((buf), (size))
1045 #define hammer2_icrc32c(buf, size, crc) iscsi_crc32_ext((buf), (size), (crc))
1046
1047 int hammer2_signal_check(time_t *timep);
1048 const char *hammer2_error_str(int error);
1049
1050 hammer2_cluster_t *hammer2_inode_lock(hammer2_inode_t *ip, int how);
1051 void hammer2_inode_unlock(hammer2_inode_t *ip, hammer2_cluster_t *cluster);
1052 hammer2_mtx_state_t hammer2_inode_lock_temp_release(hammer2_inode_t *ip);
1053 void hammer2_inode_lock_temp_restore(hammer2_inode_t *ip,
1054                         hammer2_mtx_state_t ostate);
1055 int hammer2_inode_lock_upgrade(hammer2_inode_t *ip);
1056 void hammer2_inode_lock_downgrade(hammer2_inode_t *ip, int);
1057
1058 void hammer2_dev_exlock(hammer2_dev_t *hmp);
1059 void hammer2_dev_shlock(hammer2_dev_t *hmp);
1060 void hammer2_dev_unlock(hammer2_dev_t *hmp);
1061
1062 int hammer2_get_dtype(const hammer2_inode_data_t *ipdata);
1063 int hammer2_get_vtype(const hammer2_inode_data_t *ipdata);
1064 u_int8_t hammer2_get_obj_type(enum vtype vtype);
1065 void hammer2_time_to_timespec(u_int64_t xtime, struct timespec *ts);
1066 u_int64_t hammer2_timespec_to_time(const struct timespec *ts);
1067 u_int32_t hammer2_to_unix_xid(const uuid_t *uuid);
1068 void hammer2_guid_to_uuid(uuid_t *uuid, u_int32_t guid);
1069 hammer2_xid_t hammer2_trans_newxid(hammer2_pfs_t *pmp);
1070 void hammer2_trans_manage_init(hammer2_trans_manage_t *tman);
1071
1072 hammer2_key_t hammer2_dirhash(const unsigned char *name, size_t len);
1073 int hammer2_getradix(size_t bytes);
1074
1075 int hammer2_calc_logical(hammer2_inode_t *ip, hammer2_off_t uoff,
1076                         hammer2_key_t *lbasep, hammer2_key_t *leofp);
1077 int hammer2_calc_physical(hammer2_inode_t *ip,
1078                         const hammer2_inode_data_t *ipdata,
1079                         hammer2_key_t lbase);
1080 void hammer2_update_time(uint64_t *timep);
1081 void hammer2_adjreadcounter(hammer2_blockref_t *bref, size_t bytes);
1082
1083 /*
1084  * hammer2_inode.c
1085  */
1086 struct vnode *hammer2_igetv(hammer2_inode_t *ip, hammer2_cluster_t *cparent,
1087                         int *errorp);
1088 hammer2_inode_t *hammer2_inode_lookup(hammer2_pfs_t *pmp,
1089                         hammer2_tid_t inum);
1090 hammer2_inode_t *hammer2_inode_get(hammer2_pfs_t *pmp,
1091                         hammer2_inode_t *dip, hammer2_cluster_t *cluster);
1092 void hammer2_inode_free(hammer2_inode_t *ip);
1093 void hammer2_inode_ref(hammer2_inode_t *ip);
1094 void hammer2_inode_drop(hammer2_inode_t *ip);
1095 void hammer2_inode_repoint(hammer2_inode_t *ip, hammer2_inode_t *pip,
1096                         hammer2_cluster_t *cluster);
1097 void hammer2_inode_repoint_one(hammer2_inode_t *ip, hammer2_cluster_t *cluster,
1098                         int idx);
1099
1100 void hammer2_run_unlinkq(hammer2_trans_t *trans, hammer2_pfs_t *pmp);
1101
1102 hammer2_inode_t *hammer2_inode_create(hammer2_trans_t *trans,
1103                         hammer2_inode_t *dip,
1104                         struct vattr *vap, struct ucred *cred,
1105                         const uint8_t *name, size_t name_len,
1106                         hammer2_cluster_t **clusterp,
1107                         int flags, int *errorp);
1108 int hammer2_inode_connect(hammer2_trans_t *trans,
1109                         hammer2_cluster_t **clusterp, int hlink,
1110                         hammer2_inode_t *dip, hammer2_cluster_t *dcluster,
1111                         const uint8_t *name, size_t name_len,
1112                         hammer2_key_t key);
1113 hammer2_inode_t *hammer2_inode_common_parent(hammer2_inode_t *fdip,
1114                         hammer2_inode_t *tdip);
1115 void hammer2_inode_fsync(hammer2_trans_t *trans, hammer2_inode_t *ip,
1116                         hammer2_cluster_t *cparent);
1117 int hammer2_unlink_file(hammer2_trans_t *trans, hammer2_inode_t *dip,
1118                         const uint8_t *name, size_t name_len, int isdir,
1119                         int *hlinkp, struct nchandle *nch, int nlinks);
1120 int hammer2_hardlink_consolidate(hammer2_trans_t *trans,
1121                         hammer2_inode_t *ip, hammer2_cluster_t **clusterp,
1122                         hammer2_inode_t *cdip, hammer2_cluster_t *cdcluster,
1123                         int nlinks);
1124 int hammer2_hardlink_deconsolidate(hammer2_trans_t *trans, hammer2_inode_t *dip,
1125                         hammer2_chain_t **chainp, hammer2_chain_t **ochainp);
1126 int hammer2_hardlink_find(hammer2_inode_t *dip, hammer2_cluster_t **cparentp,
1127                         hammer2_cluster_t **clusterp);
1128 int hammer2_parent_find(hammer2_cluster_t **cparentp,
1129                         hammer2_cluster_t *cluster);
1130 void hammer2_inode_install_hidden(hammer2_pfs_t *pmp);
1131
1132 /*
1133  * hammer2_chain.c
1134  */
1135 void hammer2_voldata_lock(hammer2_dev_t *hmp);
1136 void hammer2_voldata_unlock(hammer2_dev_t *hmp);
1137 void hammer2_voldata_modify(hammer2_dev_t *hmp);
1138 hammer2_chain_t *hammer2_chain_alloc(hammer2_dev_t *hmp,
1139                                 hammer2_pfs_t *pmp,
1140                                 hammer2_trans_t *trans,
1141                                 hammer2_blockref_t *bref);
1142 void hammer2_chain_core_init(hammer2_chain_t *chain);
1143 void hammer2_chain_ref(hammer2_chain_t *chain);
1144 void hammer2_chain_drop(hammer2_chain_t *chain);
1145 void hammer2_chain_lock(hammer2_chain_t *chain, int how);
1146 const hammer2_media_data_t *hammer2_chain_rdata(hammer2_chain_t *chain);
1147 hammer2_media_data_t *hammer2_chain_wdata(hammer2_chain_t *chain);
1148
1149 /*
1150  * hammer2_cluster.c
1151  */
1152 int hammer2_cluster_isunlinked(hammer2_cluster_t *cluster);
1153 void hammer2_cluster_load_async(hammer2_cluster_t *cluster,
1154                                 void (*callback)(hammer2_iocb_t *iocb),
1155                                 void *ptr);
1156 void hammer2_chain_moved(hammer2_chain_t *chain);
1157 void hammer2_chain_modify(hammer2_trans_t *trans,
1158                                 hammer2_chain_t *chain, int flags);
1159 void hammer2_chain_resize(hammer2_trans_t *trans, hammer2_inode_t *ip,
1160                                 hammer2_chain_t *parent,
1161                                 hammer2_chain_t *chain,
1162                                 int nradix, int flags);
1163 void hammer2_chain_unlock(hammer2_chain_t *chain);
1164 void hammer2_chain_wait(hammer2_chain_t *chain);
1165 hammer2_chain_t *hammer2_chain_get(hammer2_chain_t *parent, int generation,
1166                                 hammer2_blockref_t *bref);
1167 hammer2_chain_t *hammer2_chain_lookup_init(hammer2_chain_t *parent, int flags);
1168 void hammer2_chain_lookup_done(hammer2_chain_t *parent);
1169 hammer2_chain_t *hammer2_chain_lookup(hammer2_chain_t **parentp,
1170                                 hammer2_key_t *key_nextp,
1171                                 hammer2_key_t key_beg, hammer2_key_t key_end,
1172                                 int *cache_indexp, int flags);
1173 hammer2_chain_t *hammer2_chain_next(hammer2_chain_t **parentp,
1174                                 hammer2_chain_t *chain,
1175                                 hammer2_key_t *key_nextp,
1176                                 hammer2_key_t key_beg, hammer2_key_t key_end,
1177                                 int *cache_indexp, int flags);
1178 hammer2_chain_t *hammer2_chain_scan(hammer2_chain_t *parent,
1179                                 hammer2_chain_t *chain,
1180                                 int *cache_indexp, int flags);
1181
1182 int hammer2_chain_create(hammer2_trans_t *trans, hammer2_chain_t **parentp,
1183                                 hammer2_chain_t **chainp,
1184                                 hammer2_pfs_t *pmp,
1185                                 hammer2_key_t key, int keybits,
1186                                 int type, size_t bytes, int flags);
1187 void hammer2_chain_rename(hammer2_trans_t *trans, hammer2_blockref_t *bref,
1188                                 hammer2_chain_t **parentp,
1189                                 hammer2_chain_t *chain, int flags);
1190 int hammer2_chain_snapshot(hammer2_trans_t *trans, hammer2_chain_t **chainp,
1191                                 hammer2_ioc_pfs_t *pfs);
1192 void hammer2_chain_delete(hammer2_trans_t *trans, hammer2_chain_t *parent,
1193                                 hammer2_chain_t *chain, int flags);
1194 void hammer2_chain_delete_duplicate(hammer2_trans_t *trans,
1195                                 hammer2_chain_t **chainp, int flags);
1196 void hammer2_flush(hammer2_trans_t *trans, hammer2_chain_t *chain, int istop);
1197 void hammer2_chain_commit(hammer2_trans_t *trans, hammer2_chain_t *chain);
1198 void hammer2_chain_setflush(hammer2_trans_t *trans, hammer2_chain_t *chain);
1199 void hammer2_chain_countbrefs(hammer2_chain_t *chain,
1200                                 hammer2_blockref_t *base, int count);
1201
1202 void hammer2_chain_setcheck(hammer2_chain_t *chain, void *bdata);
1203 int hammer2_chain_testcheck(hammer2_chain_t *chain, void *bdata);
1204
1205
1206 void hammer2_pfs_memory_wait(hammer2_pfs_t *pmp);
1207 void hammer2_pfs_memory_inc(hammer2_pfs_t *pmp);
1208 void hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp);
1209
1210 void hammer2_base_delete(hammer2_trans_t *trans, hammer2_chain_t *chain,
1211                                 hammer2_blockref_t *base, int count,
1212                                 int *cache_indexp, hammer2_chain_t *child);
1213 void hammer2_base_insert(hammer2_trans_t *trans, hammer2_chain_t *chain,
1214                                 hammer2_blockref_t *base, int count,
1215                                 int *cache_indexp, hammer2_chain_t *child);
1216
1217 /*
1218  * hammer2_trans.c
1219  */
1220 void hammer2_trans_init(hammer2_trans_t *trans, hammer2_pfs_t *pmp,
1221                                 int flags);
1222 void hammer2_trans_done(hammer2_trans_t *trans);
1223 void hammer2_trans_assert_strategy(hammer2_pfs_t *pmp);
1224
1225 /*
1226  * hammer2_ioctl.c
1227  */
1228 int hammer2_ioctl(hammer2_inode_t *ip, u_long com, void *data,
1229                                 int fflag, struct ucred *cred);
1230
1231 /*
1232  * hammer2_io.c
1233  */
1234 void hammer2_io_putblk(hammer2_io_t **diop);
1235 void hammer2_io_cleanup(hammer2_dev_t *hmp, struct hammer2_io_tree *tree);
1236 char *hammer2_io_data(hammer2_io_t *dio, off_t lbase);
1237 void hammer2_io_getblk(hammer2_dev_t *hmp, off_t lbase, int lsize,
1238                                 hammer2_iocb_t *iocb);
1239 void hammer2_io_complete(hammer2_iocb_t *iocb);
1240 void hammer2_io_callback(struct bio *bio);
1241 void hammer2_iocb_wait(hammer2_iocb_t *iocb);
1242 int hammer2_io_new(hammer2_dev_t *hmp, off_t lbase, int lsize,
1243                                 hammer2_io_t **diop);
1244 int hammer2_io_newnz(hammer2_dev_t *hmp, off_t lbase, int lsize,
1245                                 hammer2_io_t **diop);
1246 int hammer2_io_newq(hammer2_dev_t *hmp, off_t lbase, int lsize,
1247                                 hammer2_io_t **diop);
1248 int hammer2_io_bread(hammer2_dev_t *hmp, off_t lbase, int lsize,
1249                                 hammer2_io_t **diop);
1250 void hammer2_io_bawrite(hammer2_io_t **diop);
1251 void hammer2_io_bdwrite(hammer2_io_t **diop);
1252 int hammer2_io_bwrite(hammer2_io_t **diop);
1253 int hammer2_io_isdirty(hammer2_io_t *dio);
1254 void hammer2_io_setdirty(hammer2_io_t *dio);
1255 void hammer2_io_setinval(hammer2_io_t *dio, u_int bytes);
1256 void hammer2_io_brelse(hammer2_io_t **diop);
1257 void hammer2_io_bqrelse(hammer2_io_t **diop);
1258
1259 /*
1260  * hammer2_msgops.c
1261  */
1262 int hammer2_msg_dbg_rcvmsg(kdmsg_msg_t *msg);
1263 int hammer2_msg_adhoc_input(kdmsg_msg_t *msg);
1264
1265 /*
1266  * hammer2_vfsops.c
1267  */
1268 void hammer2_clusterctl_wakeup(kdmsg_iocom_t *iocom);
1269 void hammer2_volconf_update(hammer2_dev_t *hmp, int index);
1270 void hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx);
1271 void hammer2_bioq_sync(hammer2_pfs_t *pmp);
1272 int hammer2_vfs_sync(struct mount *mp, int waitflags);
1273 hammer2_pfs_t *hammer2_pfsalloc(hammer2_cluster_t *cluster,
1274                                 const hammer2_inode_data_t *ripdata,
1275                                 hammer2_tid_t modify_tid);
1276
1277 void hammer2_lwinprog_ref(hammer2_pfs_t *pmp);
1278 void hammer2_lwinprog_drop(hammer2_pfs_t *pmp);
1279 void hammer2_lwinprog_wait(hammer2_pfs_t *pmp);
1280
1281 /*
1282  * hammer2_freemap.c
1283  */
1284 int hammer2_freemap_alloc(hammer2_trans_t *trans, hammer2_chain_t *chain,
1285                                 size_t bytes);
1286 void hammer2_freemap_adjust(hammer2_trans_t *trans, hammer2_dev_t *hmp,
1287                                 hammer2_blockref_t *bref, int how);
1288
1289 /*
1290  * hammer2_cluster.c
1291  */
1292 int hammer2_cluster_need_resize(hammer2_cluster_t *cluster, int bytes);
1293 uint8_t hammer2_cluster_type(hammer2_cluster_t *cluster);
1294 const hammer2_media_data_t *hammer2_cluster_rdata(hammer2_cluster_t *cluster);
1295 hammer2_media_data_t *hammer2_cluster_wdata(hammer2_cluster_t *cluster);
1296 hammer2_cluster_t *hammer2_cluster_from_chain(hammer2_chain_t *chain);
1297 int hammer2_cluster_modified(hammer2_cluster_t *cluster);
1298 int hammer2_cluster_duplicated(hammer2_cluster_t *cluster);
1299 void hammer2_cluster_set_chainflags(hammer2_cluster_t *cluster, uint32_t flags);
1300 void hammer2_cluster_clr_chainflags(hammer2_cluster_t *cluster, uint32_t flags);
1301 void hammer2_cluster_bref(hammer2_cluster_t *cluster, hammer2_blockref_t *bref);
1302 void hammer2_cluster_setflush(hammer2_trans_t *trans,
1303                         hammer2_cluster_t *cluster);
1304 void hammer2_cluster_setmethod_check(hammer2_trans_t *trans,
1305                         hammer2_cluster_t *cluster, int check_algo);
1306 hammer2_cluster_t *hammer2_cluster_alloc(hammer2_pfs_t *pmp,
1307                         hammer2_trans_t *trans,
1308                         hammer2_blockref_t *bref);
1309 void hammer2_cluster_ref(hammer2_cluster_t *cluster);
1310 void hammer2_cluster_drop(hammer2_cluster_t *cluster);
1311 void hammer2_cluster_wait(hammer2_cluster_t *cluster);
1312 void hammer2_cluster_lock(hammer2_cluster_t *cluster, int how);
1313 void hammer2_cluster_resolve(hammer2_cluster_t *cluster);
1314 hammer2_cluster_t *hammer2_cluster_copy(hammer2_cluster_t *ocluster);
1315 void hammer2_cluster_unlock(hammer2_cluster_t *cluster);
1316 void hammer2_cluster_resize(hammer2_trans_t *trans, hammer2_inode_t *ip,
1317                         hammer2_cluster_t *cparent, hammer2_cluster_t *cluster,
1318                         int nradix, int flags);
1319 hammer2_inode_data_t *hammer2_cluster_modify_ip(hammer2_trans_t *trans,
1320                         hammer2_inode_t *ip, hammer2_cluster_t *cluster,
1321                         int flags);
1322 void hammer2_cluster_modify(hammer2_trans_t *trans, hammer2_cluster_t *cluster,
1323                         int flags);
1324 void hammer2_cluster_modsync(hammer2_cluster_t *cluster);
1325 hammer2_cluster_t *hammer2_cluster_lookup_init(hammer2_cluster_t *cparent,
1326                         int flags);
1327 void hammer2_cluster_lookup_done(hammer2_cluster_t *cparent);
1328 hammer2_cluster_t *hammer2_cluster_lookup(hammer2_cluster_t *cparent,
1329                         hammer2_key_t *key_nextp,
1330                         hammer2_key_t key_beg, hammer2_key_t key_end,
1331                         int flags);
1332 hammer2_cluster_t *hammer2_cluster_next(hammer2_cluster_t *cparent,
1333                         hammer2_cluster_t *cluster,
1334                         hammer2_key_t *key_nextp,
1335                         hammer2_key_t key_beg, hammer2_key_t key_end,
1336                         int flags);
1337 void hammer2_cluster_next_single_chain(hammer2_cluster_t *cparent,
1338                         hammer2_cluster_t *cluster,
1339                         hammer2_key_t *key_nextp,
1340                         hammer2_key_t key_beg,
1341                         hammer2_key_t key_end,
1342                         int i, int flags);
1343 hammer2_cluster_t *hammer2_cluster_scan(hammer2_cluster_t *cparent,
1344                         hammer2_cluster_t *cluster, int flags);
1345 int hammer2_cluster_create(hammer2_trans_t *trans, hammer2_cluster_t *cparent,
1346                         hammer2_cluster_t **clusterp,
1347                         hammer2_key_t key, int keybits,
1348                         int type, size_t bytes, int flags);
1349 void hammer2_cluster_rename(hammer2_trans_t *trans, hammer2_blockref_t *bref,
1350                         hammer2_cluster_t *cparent, hammer2_cluster_t *cluster,
1351                         int flags);
1352 void hammer2_cluster_delete(hammer2_trans_t *trans, hammer2_cluster_t *pcluster,
1353                         hammer2_cluster_t *cluster, int flags);
1354 int hammer2_cluster_snapshot(hammer2_trans_t *trans,
1355                         hammer2_cluster_t *ocluster, hammer2_ioc_pfs_t *pfs);
1356 hammer2_cluster_t *hammer2_cluster_parent(hammer2_cluster_t *cluster);
1357
1358 int hammer2_bulk_scan(hammer2_trans_t *trans, hammer2_chain_t *parent,
1359                         int (*func)(hammer2_chain_t *chain, void *info),
1360                         void *info);
1361 int hammer2_bulkfree_pass(hammer2_dev_t *hmp,
1362                         struct hammer2_ioc_bulkfree *bfi);
1363
1364 /*
1365  * hammer2_iocom.c
1366  */
1367 void hammer2_iocom_init(hammer2_dev_t *hmp);
1368 void hammer2_iocom_uninit(hammer2_dev_t *hmp);
1369 void hammer2_cluster_reconnect(hammer2_dev_t *hmp, struct file *fp);
1370
1371 /*
1372  * hammer2_syncthr.c
1373  */
1374 void hammer2_syncthr_create(hammer2_syncthr_t *thr, hammer2_pfs_t *pmp,
1375                         int clindex, void (*func)(void *arg));
1376 void hammer2_syncthr_delete(hammer2_syncthr_t *thr);
1377 void hammer2_syncthr_remaster(hammer2_syncthr_t *thr);
1378 void hammer2_syncthr_freeze(hammer2_syncthr_t *thr);
1379 void hammer2_syncthr_unfreeze(hammer2_syncthr_t *thr);
1380 void hammer2_syncthr_primary(void *arg);
1381
1382 #endif /* !_KERNEL */
1383 #endif /* !_VFS_HAMMER2_HAMMER2_H_ */