sys/vfs/hammer2/hammer2.h

   1 /*
   2  * Copyright (c) 2011-2018 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_msg;
 116 struct hammer2_thread;
 117 union hammer2_xop;
 118
 119 /*
 120  * Mutex and lock shims.  Hammer2 requires support for asynchronous and
 121  * abortable locks, and both exclusive and shared spinlocks.  Normal
 122  * synchronous non-abortable locks can be substituted for spinlocks.
 123  */
 124 typedef mtx_t                           hammer2_mtx_t;
 125 typedef mtx_link_t                      hammer2_mtx_link_t;
 126 typedef mtx_state_t                     hammer2_mtx_state_t;
 127
 128 typedef struct spinlock                 hammer2_spin_t;
 129
 130 #define hammer2_mtx_ex                  mtx_lock_ex_quick
 131 #define hammer2_mtx_ex_try              mtx_lock_ex_try
 132 #define hammer2_mtx_sh                  mtx_lock_sh_quick
 133 #define hammer2_mtx_sh_again            mtx_lock_sh_again
 134 #define hammer2_mtx_sh_try              mtx_lock_sh_try
 135 #define hammer2_mtx_unlock              mtx_unlock
 136 #define hammer2_mtx_downgrade           mtx_downgrade
 137 #define hammer2_mtx_owned               mtx_owned
 138 #define hammer2_mtx_init                mtx_init
 139 #define hammer2_mtx_temp_release        mtx_lock_temp_release
 140 #define hammer2_mtx_temp_restore        mtx_lock_temp_restore
 141 #define hammer2_mtx_refs                mtx_lockrefs
 142
 143 #define hammer2_spin_init               spin_init
 144 #define hammer2_spin_sh                 spin_lock_shared
 145 #define hammer2_spin_ex                 spin_lock
 146 #define hammer2_spin_unsh               spin_unlock_shared
 147 #define hammer2_spin_unex               spin_unlock
 148
 149 TAILQ_HEAD(hammer2_xop_list, hammer2_xop_head);
 150 TAILQ_HEAD(hammer2_chain_list, hammer2_chain);
 151
 152 typedef struct hammer2_xop_list hammer2_xop_list_t;
 153
 154 #ifdef _KERNEL
 155 /*
 156  * General lock support
 157  */
 158 static __inline
 159 int
 160 hammer2_mtx_upgrade_try(hammer2_mtx_t *mtx)
 161 {
 162         return mtx_upgrade_try(mtx);
 163 }
 164
 165 #endif
 166
 167 /*
 168  * The xid tracks internal transactional updates.
 169  *
 170  * XXX fix-me, really needs to be 64-bits
 171  */
 172 typedef uint32_t hammer2_xid_t;
 173
 174 #define HAMMER2_XID_MIN                 0x00000000U
 175 #define HAMMER2_XID_MAX                 0x7FFFFFFFU
 176
 177 #define HAMMER2_LIMIT_DIRTY_CHAINS      (65536)
 178
 179 /*
 180  * The chain structure tracks a portion of the media topology from the
 181  * root (volume) down.  Chains represent volumes, inodes, indirect blocks,
 182  * data blocks, and freemap nodes and leafs.
 183  *
 184  * The chain structure utilizes a simple singly-homed topology and the
 185  * chain's in-memory topology will move around as the chains do, due mainly
 186  * to renames and indirect block creation.
 187  *
 188  * Block Table Updates
 189  *
 190  *      Block table updates for insertions and updates are delayed until the
 191  *      flush.  This allows us to avoid having to modify the parent chain
 192  *      all the way to the root.
 193  *
 194  *      Block table deletions are performed immediately (modifying the parent
 195  *      in the process) because the flush code uses the chain structure to
 196  *      track delayed updates and the chain will be (likely) gone or moved to
 197  *      another location in the topology after a deletion.
 198  *
 199  *      A prior iteration of the code tried to keep the relationship intact
 200  *      on deletes by doing a delete-duplicate operation on the chain, but
 201  *      it added way too much complexity to the codebase.
 202  *
 203  * Flush Synchronization
 204  *
 205  *      The flush code must flush modified chains bottom-up.  Because chain
 206  *      structures can shift around and are NOT topologically stable,
 207  *      modified chains are independently indexed for the flush.  As the flush
 208  *      runs it modifies (or further modifies) and updates the parents,
 209  *      propagating the flush all the way to the volume root.
 210  *
 211  *      Modifying front-end operations can occur during a flush but will block
 212  *      in two cases: (1) when the front-end tries to operate on the inode
 213  *      currently in the midst of being flushed and (2) if the front-end
 214  *      crosses an inode currently being flushed (such as during a rename).
 215  *      So, for example, if you rename directory "x" to "a/b/c/d/e/f/g/x" and
 216  *      the flusher is currently working on "a/b/c", the rename will block
 217  *      temporarily in order to ensure that "x" exists in one place or the
 218  *      other.
 219  *
 220  *      Meta-data statistics are updated by the flusher.  The front-end will
 221  *      make estimates but meta-data must be fully synchronized only during a
 222  *      flush in order to ensure that it remains correct across a crash.
 223  *
 224  *      Multiple flush synchronizations can theoretically be in-flight at the
 225  *      same time but the implementation is not coded to handle the case and
 226  *      currently serializes them.
 227  *
 228  * Snapshots:
 229  *
 230  *      Snapshots currently require the subdirectory tree being snapshotted
 231  *      to be flushed.  The snapshot then creates a new super-root inode which
 232  *      copies the flushed blockdata of the directory or file that was
 233  *      snapshotted.
 234  *
 235  * RBTREE NOTES:
 236  *
 237  *      - Note that the radix tree runs in powers of 2 only so sub-trees
 238  *        cannot straddle edges.
 239  */
 240 RB_HEAD(hammer2_chain_tree, hammer2_chain);
 241 TAILQ_HEAD(h2_flush_list, hammer2_chain);
 242 TAILQ_HEAD(h2_core_list, hammer2_chain);
 243
 244 #define CHAIN_CORE_DELETE_BMAP_ENTRIES  \
 245         (HAMMER2_PBUFSIZE / sizeof(hammer2_blockref_t) / sizeof(uint32_t))
 246
 247 struct hammer2_reptrack {
 248         hammer2_spin_t  spin;
 249         struct hammer2_reptrack *next;
 250         struct hammer2_chain    *chain;
 251 };
 252
 253 /*
 254  * Core topology for chain (embedded in chain).  Protected by a spinlock.
 255  */
 256 struct hammer2_chain_core {
 257         hammer2_spin_t  spin;
 258         struct hammer2_reptrack *reptrack;
 259         struct hammer2_chain_tree rbtree; /* sub-chains */
 260         int             live_zero;      /* blockref array opt */
 261         u_int           live_count;     /* live (not deleted) chains in tree */
 262         u_int           chain_count;    /* live + deleted chains under core */
 263         int             generation;     /* generation number (inserts only) */
 264 };
 265
 266 typedef struct hammer2_chain_core hammer2_chain_core_t;
 267
 268 RB_HEAD(hammer2_io_tree, hammer2_io);
 269
 270 /*
 271  * DIO - Management structure wrapping system buffer cache.
 272  *
 273  * HAMMER2 uses an I/O abstraction that allows it to cache and manipulate
 274  * fixed-sized filesystem buffers frontend by variable-sized hammer2_chain
 275  * structures.
 276  */
 277 struct hammer2_io {
 278         RB_ENTRY(hammer2_io) rbnode;    /* indexed by device offset */
 279         struct hammer2_dev *hmp;
 280         struct buf      *bp;
 281         off_t           pbase;
 282         uint64_t        refs;
 283         int             psize;
 284         int             act;            /* activity */
 285         int             btype;          /* approximate BREF_TYPE_* */
 286         int             ticks;
 287         int             error;
 288         int             unused01;
 289         uint64_t        dedup_valid;    /* valid for dedup operation */
 290         uint64_t        dedup_alloc;    /* allocated / de-dupable */
 291 };
 292
 293 typedef struct hammer2_io hammer2_io_t;
 294
 295 #define HAMMER2_DIO_INPROG      0x8000000000000000LLU   /* bio in progress */
 296 #define HAMMER2_DIO_GOOD        0x4000000000000000LLU   /* dio->bp is stable */
 297 #define HAMMER2_DIO_WAITING     0x2000000000000000LLU   /* wait on INPROG */
 298 #define HAMMER2_DIO_DIRTY       0x1000000000000000LLU   /* flush last drop */
 299
 300 #define HAMMER2_DIO_MASK        0x00FFFFFFFFFFFFFFLLU
 301
 302 /*
 303  * Primary chain structure keeps track of the topology in-memory.
 304  */
 305 struct hammer2_chain {
 306         hammer2_mtx_t           lock;
 307         hammer2_chain_core_t    core;
 308         RB_ENTRY(hammer2_chain) rbnode;         /* live chain(s) */
 309         hammer2_blockref_t      bref;
 310         struct hammer2_chain    *parent;
 311         struct hammer2_dev      *hmp;
 312         struct hammer2_pfs      *pmp;           /* A PFS or super-root (spmp) */
 313
 314         hammer2_io_t    *dio;                   /* physical data buffer */
 315         u_int           bytes;                  /* physical data size */
 316         u_int           flags;
 317         u_int           refs;
 318         u_int           lockcnt;
 319         int             error;                  /* on-lock data error state */
 320         int             cache_index;            /* heur speeds up lookup */
 321
 322         hammer2_media_data_t *data;             /* data pointer shortcut */
 323         TAILQ_ENTRY(hammer2_chain) flush_node;  /* flush list */
 324         TAILQ_ENTRY(hammer2_chain) lru_node;    /* 0-refs LRU */
 325 };
 326
 327 typedef struct hammer2_chain hammer2_chain_t;
 328
 329 int hammer2_chain_cmp(hammer2_chain_t *chain1, hammer2_chain_t *chain2);
 330 RB_PROTOTYPE(hammer2_chain_tree, hammer2_chain, rbnode, hammer2_chain_cmp);
 331
 332 /*
 333  * Special notes on flags:
 334  *
 335  * INITIAL      - This flag allows a chain to be created and for storage to
 336  *                be allocated without having to immediately instantiate the
 337  *                related buffer.  The data is assumed to be all-zeros.  It
 338  *                is primarily used for indirect blocks.
 339  *
 340  * MODIFIED     - The chain's media data has been modified.  Prevents chain
 341  *                free on lastdrop if still in the topology.
 342  *
 343  * UPDATE       - Chain might not be modified but parent blocktable needs
 344  *                an update.  Prevents chain free on lastdrop if still in
 345  *                the topology.
 346  *
 347  * FICTITIOUS   - Faked chain as a placeholder for an error condition.  This
 348  *                chain is unsuitable for I/O.
 349  *
 350  * BMAPPED      - Indicates that the chain is present in the parent blockmap.
 351  *
 352  * BMAPUPD      - Indicates that the chain is present but needs to be updated
 353  *                in the parent blockmap.
 354  */
 355 #define HAMMER2_CHAIN_MODIFIED          0x00000001      /* dirty chain data */
 356 #define HAMMER2_CHAIN_ALLOCATED         0x00000002      /* kmalloc'd chain */
 357 #define HAMMER2_CHAIN_DESTROY           0x00000004
 358 #define HAMMER2_CHAIN_DEDUPABLE         0x00000008      /* registered w/dedup */
 359 #define HAMMER2_CHAIN_DELETED           0x00000010      /* deleted chain */
 360 #define HAMMER2_CHAIN_INITIAL           0x00000020      /* initial create */
 361 #define HAMMER2_CHAIN_UPDATE            0x00000040      /* need parent update */
 362 #define HAMMER2_CHAIN_DEFERRED          0x00000080      /* flush depth defer */
 363 #define HAMMER2_CHAIN_TESTEDGOOD        0x00000100      /* crc tested good */
 364 #define HAMMER2_CHAIN_ONFLUSH           0x00000200      /* on a flush list */
 365 #define HAMMER2_CHAIN_FICTITIOUS        0x00000400      /* unsuitable for I/O */
 366 #define HAMMER2_CHAIN_VOLUMESYNC        0x00000800      /* needs volume sync */
 367 #define HAMMER2_CHAIN_DELAYED           0x00001000      /* delayed flush */
 368 #define HAMMER2_CHAIN_COUNTEDBREFS      0x00002000      /* block table stats */
 369 #define HAMMER2_CHAIN_ONRBTREE          0x00004000      /* on parent RB tree */
 370 #define HAMMER2_CHAIN_ONLRU             0x00008000      /* on LRU list */
 371 #define HAMMER2_CHAIN_EMBEDDED          0x00010000      /* embedded data */
 372 #define HAMMER2_CHAIN_RELEASE           0x00020000      /* don't keep around */
 373 #define HAMMER2_CHAIN_BMAPPED           0x00040000      /* present in blkmap */
 374 #define HAMMER2_CHAIN_BMAPUPD           0x00080000      /* +needs updating */
 375 #define HAMMER2_CHAIN_IOINPROG          0x00100000      /* I/O interlock */
 376 #define HAMMER2_CHAIN_IOSIGNAL          0x00200000      /* I/O interlock */
 377 #define HAMMER2_CHAIN_PFSBOUNDARY       0x00400000      /* super->pfs inode */
 378 #define HAMMER2_CHAIN_HINT_LEAF_COUNT   0x00800000      /* redo leaf count */
 379 #define HAMMER2_CHAIN_LRUHINT           0x01000000      /* was reused */
 380
 381 #define HAMMER2_CHAIN_FLUSH_MASK        (HAMMER2_CHAIN_MODIFIED |       \
 382                                          HAMMER2_CHAIN_UPDATE |         \
 383                                          HAMMER2_CHAIN_ONFLUSH |        \
 384                                          HAMMER2_CHAIN_DESTROY)
 385
 386 /*
 387  * Hammer2 error codes, used by chain->error and cluster->error.  The error
 388  * code is typically set on-lock unless no I/O was requested, and set on
 389  * I/O otherwise.  If set for a cluster it generally means that the cluster
 390  * code could not find a valid copy to present.
 391  *
 392  * All H2 error codes are flags and can be accumulated by ORing them
 393  * together.
 394  *
 395  * IO           - An I/O error occurred
 396  * CHECK        - I/O succeeded but did not match the check code
 397  * INCOMPLETE   - A cluster is not complete enough to use, or
 398  *                a chain cannot be loaded because its parent has an error.
 399  *
 400  * NOTE: API allows callers to check zero/non-zero to determine if an error
 401  *       condition exists.
 402  *
 403  * NOTE: Chain's data field is usually NULL on an IO error but not necessarily
 404  *       NULL on other errors.  Check chain->error, not chain->data.
 405  */
 406 #define HAMMER2_ERROR_NONE              0       /* no error (must be 0) */
 407 #define HAMMER2_ERROR_EIO               0x00000001      /* device I/O error */
 408 #define HAMMER2_ERROR_CHECK             0x00000002      /* check code error */
 409 #define HAMMER2_ERROR_INCOMPLETE        0x00000004      /* incomplete cluster */
 410 #define HAMMER2_ERROR_DEPTH             0x00000008      /* tmp depth limit */
 411 #define HAMMER2_ERROR_BADBREF           0x00000010      /* illegal bref */
 412 #define HAMMER2_ERROR_ENOSPC            0x00000020      /* allocation failure */
 413 #define HAMMER2_ERROR_ENOENT            0x00000040      /* entry not found */
 414 #define HAMMER2_ERROR_ENOTEMPTY         0x00000080      /* dir not empty */
 415 #define HAMMER2_ERROR_EAGAIN            0x00000100      /* retry */
 416 #define HAMMER2_ERROR_ENOTDIR           0x00000200      /* not directory */
 417 #define HAMMER2_ERROR_EISDIR            0x00000400      /* is directory */
 418 #define HAMMER2_ERROR_EINPROGRESS       0x00000800      /* already running */
 419 #define HAMMER2_ERROR_ABORTED           0x00001000      /* aborted operation */
 420 #define HAMMER2_ERROR_EOF               0x00002000      /* end of scan */
 421 #define HAMMER2_ERROR_EINVAL            0x00004000      /* catch-all */
 422 #define HAMMER2_ERROR_EEXIST            0x00008000      /* entry exists */
 423 #define HAMMER2_ERROR_EDEADLK           0x00010000
 424 #define HAMMER2_ERROR_ESRCH             0x00020000
 425 #define HAMMER2_ERROR_ETIMEDOUT         0x00040000
 426
 427 /*
 428  * Flags passed to hammer2_chain_lookup() and hammer2_chain_next()
 429  *
 430  * NOTES:
 431  *      NODATA      - Asks that the chain->data not be resolved in order
 432  *                    to avoid I/O.
 433  *
 434  *      NODIRECT    - Prevents a lookup of offset 0 in an inode from returning
 435  *                    the inode itself if the inode is in DIRECTDATA mode
 436  *                    (i.e. file is <= 512 bytes).  Used by the synchronization
 437  *                    code to prevent confusion.
 438  *
 439  *      SHARED      - The input chain is expected to be locked shared,
 440  *                    and the output chain is locked shared.
 441  *
 442  *      MATCHIND    - Allows an indirect block / freemap node to be returned
 443  *                    when the passed key range matches the radix.  Remember
 444  *                    that key_end is inclusive (e.g. {0x000,0xFFF},
 445  *                    not {0x000,0x1000}).
 446  *
 447  *                    (Cannot be used for remote or cluster ops).
 448  *
 449  *      ALLNODES    - Allows NULL focus.
 450  *
 451  *      ALWAYS      - Always resolve the data.  If ALWAYS and NODATA are both
 452  *                    missing, bulk file data is not resolved but inodes and
 453  *                    other meta-data will.
 454  */
 455 #define HAMMER2_LOOKUP_UNUSED0001       0x00000001
 456 #define HAMMER2_LOOKUP_NODATA           0x00000002      /* data left NULL */
 457 #define HAMMER2_LOOKUP_NODIRECT         0x00000004      /* no offset=0 DD */
 458 #define HAMMER2_LOOKUP_SHARED           0x00000100
 459 #define HAMMER2_LOOKUP_MATCHIND         0x00000200      /* return all chains */
 460 #define HAMMER2_LOOKUP_ALLNODES         0x00000400      /* allow NULL focus */
 461 #define HAMMER2_LOOKUP_ALWAYS           0x00000800      /* resolve data */
 462 #define HAMMER2_LOOKUP_UNUSED1000       0x00001000
 463
 464 /*
 465  * Flags passed to hammer2_chain_modify() and hammer2_chain_resize()
 466  *
 467  * NOTE: OPTDATA allows us to avoid instantiating buffers for INDIRECT
 468  *       blocks in the INITIAL-create state.
 469  */
 470 #define HAMMER2_MODIFY_OPTDATA          0x00000002      /* data can be NULL */
 471 #define HAMMER2_MODIFY_NO_MODIFY_TID    0x00000004
 472 #define HAMMER2_MODIFY_UNUSED0008       0x00000008
 473
 474 /*
 475  * Flags passed to hammer2_chain_lock()
 476  *
 477  * NOTE: RDONLY is set to optimize cluster operations when *no* modifications
 478  *       will be made to either the cluster being locked or any underlying
 479  *       cluster.  It allows the cluster to lock and access data for a subset
 480  *       of available nodes instead of all available nodes.
 481  *
 482  * NOTE: NONBLOCK is only used for hammer2_chain_repparent() and getparent(),
 483  *       other functions (e.g. hammer2_chain_lookup(), etc) can't handle its
 484  *       operation.
 485  */
 486 #define HAMMER2_RESOLVE_NEVER           1
 487 #define HAMMER2_RESOLVE_MAYBE           2
 488 #define HAMMER2_RESOLVE_ALWAYS          3
 489 #define HAMMER2_RESOLVE_MASK            0x0F
 490
 491 #define HAMMER2_RESOLVE_SHARED          0x10    /* request shared lock */
 492 #define HAMMER2_RESOLVE_LOCKAGAIN       0x20    /* another shared lock */
 493 #define HAMMER2_RESOLVE_RDONLY          0x40    /* higher level op flag */
 494 #define HAMMER2_RESOLVE_NONBLOCK        0x80    /* non-blocking */
 495
 496 /*
 497  * Flags passed to hammer2_chain_delete()
 498  */
 499 #define HAMMER2_DELETE_PERMANENT        0x0001
 500
 501 /*
 502  * Flags passed to hammer2_chain_insert() or hammer2_chain_rename()
 503  * or hammer2_chain_create().
 504  */
 505 #define HAMMER2_INSERT_PFSROOT          0x0004
 506 #define HAMMER2_INSERT_SAMEPARENT       0x0008
 507
 508 /*
 509  * Flags passed to hammer2_chain_delete_duplicate()
 510  */
 511 #define HAMMER2_DELDUP_RECORE           0x0001
 512
 513 /*
 514  * Cluster different types of storage together for allocations
 515  */
 516 #define HAMMER2_FREECACHE_INODE         0
 517 #define HAMMER2_FREECACHE_INDIR         1
 518 #define HAMMER2_FREECACHE_DATA          2
 519 #define HAMMER2_FREECACHE_UNUSED3       3
 520 #define HAMMER2_FREECACHE_TYPES         4
 521
 522 /*
 523  * hammer2_freemap_alloc() block preference
 524  */
 525 #define HAMMER2_OFF_NOPREF              ((hammer2_off_t)-1)
 526
 527 /*
 528  * BMAP read-ahead maximum parameters
 529  */
 530 #define HAMMER2_BMAP_COUNT              16      /* max bmap read-ahead */
 531 #define HAMMER2_BMAP_BYTES              (HAMMER2_PBUFSIZE * HAMMER2_BMAP_COUNT)
 532
 533 /*
 534  * hammer2_freemap_adjust()
 535  */
 536 #define HAMMER2_FREEMAP_DORECOVER       1
 537 #define HAMMER2_FREEMAP_DOMAYFREE       2
 538 #define HAMMER2_FREEMAP_DOREALFREE      3
 539
 540 /*
 541  * HAMMER2 cluster - A set of chains representing the same entity.
 542  *
 543  * hammer2_cluster typically represents a temporary set of representitive
 544  * chains.  The one exception is that a hammer2_cluster is embedded in
 545  * hammer2_inode.  This embedded cluster is ONLY used to track the
 546  * representitive chains and cannot be directly locked.
 547  *
 548  * A cluster is usually temporary (and thus per-thread) for locking purposes,
 549  * allowing us to embed the asynchronous storage required for cluster
 550  * operations in the cluster itself and adjust the state and status without
 551  * having to worry too much about SMP issues.
 552  *
 553  * The exception is the cluster embedded in the hammer2_inode structure.
 554  * This is used to cache the cluster state on an inode-by-inode basis.
 555  * Individual hammer2_chain structures not incorporated into clusters might
 556  * also stick around to cache miscellanious elements.
 557  *
 558  * Because the cluster is a 'working copy' and is usually subject to cluster
 559  * quorum rules, it is quite possible for us to end up with an insufficient
 560  * number of live chains to execute an operation.  If an insufficient number
 561  * of chains remain in a working copy, the operation may have to be
 562  * downgraded, retried, stall until the requisit number of chains are
 563  * available, or possibly even error out depending on the mount type.
 564  *
 565  * A cluster's focus is set when it is locked.  The focus can only be set
 566  * to a chain still part of the synchronized set.
 567  */
 568 #define HAMMER2_XOPFIFO         16
 569 #define HAMMER2_XOPFIFO_MASK    (HAMMER2_XOPFIFO - 1)
 570 #define HAMMER2_XOPGROUPS       32
 571 #define HAMMER2_XOPGROUPS_MASK  (HAMMER2_XOPGROUPS - 1)
 572
 573 #define HAMMER2_MAXCLUSTER      8
 574 #define HAMMER2_XOPMASK_CLUSTER (uint64_t)((1LLU << HAMMER2_MAXCLUSTER) - 1)
 575 #define HAMMER2_XOPMASK_VOP     (uint64_t)0x0000000080000000LLU
 576 #define HAMMER2_XOPMASK_FIFOW   (uint64_t)0x0000000040000000LLU
 577 #define HAMMER2_XOPMASK_WAIT    (uint64_t)0x0000000020000000LLU
 578 #define HAMMER2_XOPMASK_FEED    (uint64_t)0x0000000100000000LLU
 579
 580 #define HAMMER2_XOPMASK_ALLDONE (HAMMER2_XOPMASK_VOP | HAMMER2_XOPMASK_CLUSTER)
 581
 582 #define HAMMER2_SPECTHREADS     1       /* sync */
 583
 584 struct hammer2_cluster_item {
 585         hammer2_chain_t         *chain;
 586         int                     error;
 587         uint32_t                flags;
 588 };
 589
 590 typedef struct hammer2_cluster_item hammer2_cluster_item_t;
 591
 592 /*
 593  * INVALID      - Invalid for focus, i.e. not part of synchronized set.
 594  *                Once set, this bit is sticky across operations.
 595  *
 596  * FEMOD        - Indicates that front-end modifying operations can
 597  *                mess with this entry and MODSYNC will copy also
 598  *                effect it.
 599  */
 600 #define HAMMER2_CITEM_INVALID   0x00000001
 601 #define HAMMER2_CITEM_FEMOD     0x00000002
 602 #define HAMMER2_CITEM_NULL      0x00000004
 603
 604 struct hammer2_cluster {
 605         int                     refs;           /* track for deallocation */
 606         int                     ddflag;
 607         struct hammer2_pfs      *pmp;
 608         uint32_t                flags;
 609         int                     nchains;
 610         int                     error;          /* error code valid on lock */
 611         int                     focus_index;
 612         hammer2_chain_t         *focus;         /* current focus (or mod) */
 613         hammer2_cluster_item_t  array[HAMMER2_MAXCLUSTER];
 614 };
 615
 616 typedef struct hammer2_cluster  hammer2_cluster_t;
 617
 618 /*
 619  * WRHARD       - Hard mounts can write fully synchronized
 620  * RDHARD       - Hard mounts can read fully synchronized
 621  * UNHARD       - Unsynchronized masters present
 622  * NOHARD       - No masters visible
 623  * WRSOFT       - Soft mounts can write to at least the SOFT_MASTER
 624  * RDSOFT       - Soft mounts can read from at least a SOFT_SLAVE
 625  * UNSOFT       - Unsynchronized slaves present
 626  * NOSOFT       - No slaves visible
 627  * RDSLAVE      - slaves are accessible (possibly unsynchronized or remote).
 628  * MSYNCED      - All masters are fully synchronized
 629  * SSYNCED      - All known local slaves are fully synchronized to masters
 630  *
 631  * All available masters are always incorporated.  All PFSs belonging to a
 632  * cluster (master, slave, copy, whatever) always try to synchronize the
 633  * total number of known masters in the PFSs root inode.
 634  *
 635  * A cluster might have access to many slaves, copies, or caches, but we
 636  * have a limited number of cluster slots.  Any such elements which are
 637  * directly mounted from block device(s) will always be incorporated.   Note
 638  * that SSYNCED only applies to such elements which are directly mounted,
 639  * not to any remote slaves, copies, or caches that could be available.  These
 640  * bits are used to monitor and drive our synchronization threads.
 641  *
 642  * When asking the question 'is any data accessible at all', then a simple
 643  * test against (RDHARD|RDSOFT|RDSLAVE) gives you the answer.  If any of
 644  * these bits are set the object can be read with certain caveats:
 645  * RDHARD - no caveats.  RDSOFT - authoritative but might not be synchronized.
 646  * and RDSLAVE - not authoritative, has some data but it could be old or
 647  * incomplete.
 648  *
 649  * When both soft and hard mounts are available, data will be read and written
 650  * via the soft mount only.  But all might be in the cluster because
 651  * background synchronization threads still need to do their work.
 652  */
 653 #define HAMMER2_CLUSTER_INODE   0x00000001      /* embedded in inode struct */
 654 #define HAMMER2_CLUSTER_UNUSED2 0x00000002
 655 #define HAMMER2_CLUSTER_LOCKED  0x00000004      /* cluster lks not recursive */
 656 #define HAMMER2_CLUSTER_WRHARD  0x00000100      /* hard-mount can write */
 657 #define HAMMER2_CLUSTER_RDHARD  0x00000200      /* hard-mount can read */
 658 #define HAMMER2_CLUSTER_UNHARD  0x00000400      /* unsynchronized masters */
 659 #define HAMMER2_CLUSTER_NOHARD  0x00000800      /* no masters visible */
 660 #define HAMMER2_CLUSTER_WRSOFT  0x00001000      /* soft-mount can write */
 661 #define HAMMER2_CLUSTER_RDSOFT  0x00002000      /* soft-mount can read */
 662 #define HAMMER2_CLUSTER_UNSOFT  0x00004000      /* unsynchronized slaves */
 663 #define HAMMER2_CLUSTER_NOSOFT  0x00008000      /* no slaves visible */
 664 #define HAMMER2_CLUSTER_MSYNCED 0x00010000      /* all masters synchronized */
 665 #define HAMMER2_CLUSTER_SSYNCED 0x00020000      /* known slaves synchronized */
 666
 667 #define HAMMER2_CLUSTER_ANYDATA ( HAMMER2_CLUSTER_RDHARD |      \
 668                                   HAMMER2_CLUSTER_RDSOFT |      \
 669                                   HAMMER2_CLUSTER_RDSLAVE)
 670
 671 #define HAMMER2_CLUSTER_RDOK    ( HAMMER2_CLUSTER_RDHARD |      \
 672                                   HAMMER2_CLUSTER_RDSOFT)
 673
 674 #define HAMMER2_CLUSTER_WROK    ( HAMMER2_CLUSTER_WRHARD |      \
 675                                   HAMMER2_CLUSTER_WRSOFT)
 676
 677 #define HAMMER2_CLUSTER_ZFLAGS  ( HAMMER2_CLUSTER_WRHARD |      \
 678                                   HAMMER2_CLUSTER_RDHARD |      \
 679                                   HAMMER2_CLUSTER_WRSOFT |      \
 680                                   HAMMER2_CLUSTER_RDSOFT |      \
 681                                   HAMMER2_CLUSTER_MSYNCED |     \
 682                                   HAMMER2_CLUSTER_SSYNCED)
 683
 684 /*
 685  * Helper functions (cluster must be locked for flags to be valid).
 686  */
 687 static __inline
 688 int
 689 hammer2_cluster_rdok(hammer2_cluster_t *cluster)
 690 {
 691         return (cluster->flags & HAMMER2_CLUSTER_RDOK);
 692 }
 693
 694 static __inline
 695 int
 696 hammer2_cluster_wrok(hammer2_cluster_t *cluster)
 697 {
 698         return (cluster->flags & HAMMER2_CLUSTER_WROK);
 699 }
 700
 701 RB_HEAD(hammer2_inode_tree, hammer2_inode);
 702
 703 /*
 704  * A hammer2 inode.
 705  *
 706  * NOTE: The inode-embedded cluster is never used directly for I/O (since
 707  *       it may be shared).  Instead it will be replicated-in and synchronized
 708  *       back out if changed.
 709  */
 710 struct hammer2_inode {
 711         RB_ENTRY(hammer2_inode) rbnode;         /* inumber lookup (HL) */
 712         hammer2_mtx_t           lock;           /* inode lock */
 713         hammer2_mtx_t           truncate_lock;  /* prevent truncates */
 714         struct hammer2_pfs      *pmp;           /* PFS mount */
 715         struct vnode            *vp;
 716         struct spinlock         cluster_spin;   /* update cluster */
 717         hammer2_cluster_t       cluster;
 718         struct lockf            advlock;
 719         u_int                   flags;
 720         u_int                   refs;           /* +vpref, +flushref */
 721         uint8_t                 comp_heuristic;
 722         hammer2_inode_meta_t    meta;           /* copy of meta-data */
 723         hammer2_off_t           osize;
 724 };
 725
 726 typedef struct hammer2_inode hammer2_inode_t;
 727
 728 /*
 729  * MODIFIED     - Inode is in a modified state, ip->meta may have changes.
 730  * RESIZED      - Inode truncated (any) or inode extended beyond
 731  *                EMBEDDED_BYTES.
 732  */
 733 #define HAMMER2_INODE_MODIFIED          0x0001
 734 #define HAMMER2_INODE_SROOT             0x0002  /* kmalloc special case */
 735 #define HAMMER2_INODE_RENAME_INPROG     0x0004
 736 #define HAMMER2_INODE_ONRBTREE          0x0008
 737 #define HAMMER2_INODE_RESIZED           0x0010  /* requires inode_fsync */
 738 #define HAMMER2_INODE_ISDELETED         0x0020  /* deleted */
 739 #define HAMMER2_INODE_ISUNLINKED        0x0040
 740 #define HAMMER2_INODE_METAGOOD          0x0080  /* inode meta-data good */
 741 #define HAMMER2_INODE_ONSIDEQ           0x0100  /* on side processing queue */
 742 #define HAMMER2_INODE_NOSIDEQ           0x0200  /* disable sideq operation */
 743 #define HAMMER2_INODE_DIRTYDATA         0x0400  /* interlocks inode flush */
 744
 745 int hammer2_inode_cmp(hammer2_inode_t *ip1, hammer2_inode_t *ip2);
 746 RB_PROTOTYPE2(hammer2_inode_tree, hammer2_inode, rbnode, hammer2_inode_cmp,
 747                 hammer2_tid_t);
 748
 749 /*
 750  * inode-unlink side-structure
 751  */
 752 struct hammer2_inode_sideq {
 753         TAILQ_ENTRY(hammer2_inode_sideq) entry;
 754         hammer2_inode_t *ip;
 755 };
 756 TAILQ_HEAD(h2_sideq_list, hammer2_inode_sideq);
 757
 758 typedef struct hammer2_inode_sideq hammer2_inode_sideq_t;
 759
 760 /*
 761  * Transaction management sub-structure under hammer2_pfs
 762  */
 763 struct hammer2_trans {
 764         uint32_t                flags;
 765         uint32_t                sync_wait;
 766         int                     fticks;                 /* FPENDING start */
 767 };
 768
 769 typedef struct hammer2_trans hammer2_trans_t;
 770
 771 #define HAMMER2_TRANS_ISFLUSH           0x80000000      /* flush code */
 772 #define HAMMER2_TRANS_BUFCACHE          0x40000000      /* bio strategy */
 773 #define HAMMER2_TRANS_UNUSED20          0x20000000
 774 #define HAMMER2_TRANS_FPENDING          0x10000000      /* flush pending */
 775 #define HAMMER2_TRANS_WAITING           0x08000000      /* someone waiting */
 776 #define HAMMER2_TRANS_MASK              0x00FFFFFF      /* count mask */
 777
 778 #define HAMMER2_FREEMAP_HEUR_NRADIX     4       /* pwr 2 PBUFRADIX-MINIORADIX */
 779 #define HAMMER2_FREEMAP_HEUR_TYPES      8
 780 #define HAMMER2_FREEMAP_HEUR_SIZE       (HAMMER2_FREEMAP_HEUR_NRADIX * \
 781                                          HAMMER2_FREEMAP_HEUR_TYPES)
 782
 783 #define HAMMER2_DEDUP_HEUR_SIZE         (65536 * 4)
 784 #define HAMMER2_DEDUP_HEUR_MASK         (HAMMER2_DEDUP_HEUR_SIZE - 1)
 785
 786 #define HAMMER2_FLUSH_TOP               0x0001
 787 #define HAMMER2_FLUSH_ALL               0x0002
 788 #define HAMMER2_FLUSH_INODE_STOP        0x0004  /* stop at sub-inode */
 789
 790
 791 /*
 792  * Hammer2 support thread element.
 793  *
 794  * Potentially many support threads can hang off of hammer2, primarily
 795  * off the hammer2_pfs structure.  Typically:
 796  *
 797  * td x Nodes                   A synchronization thread for each node.
 798  * td x Nodes x workers         Worker threads for frontend operations.
 799  * td x 1                       Bioq thread for logical buffer writes.
 800  *
 801  * In addition, the synchronization thread(s) associated with the
 802  * super-root PFS (spmp) for a node is responsible for automatic bulkfree
 803  * and dedup scans.
 804  */
 805 struct hammer2_thread {
 806         struct hammer2_pfs *pmp;
 807         struct hammer2_dev *hmp;
 808         hammer2_xop_list_t xopq;
 809         thread_t        td;
 810         uint32_t        flags;
 811         int             depth;
 812         int             clindex;        /* cluster element index */
 813         int             repidx;
 814         char            *scratch;       /* MAXPHYS */
 815 };
 816
 817 typedef struct hammer2_thread hammer2_thread_t;
 818
 819 #define HAMMER2_THREAD_UNMOUNTING       0x0001  /* unmount request */
 820 #define HAMMER2_THREAD_DEV              0x0002  /* related to dev, not pfs */
 821 #define HAMMER2_THREAD_WAITING          0x0004  /* thread in idle tsleep */
 822 #define HAMMER2_THREAD_REMASTER         0x0008  /* remaster request */
 823 #define HAMMER2_THREAD_STOP             0x0010  /* exit request */
 824 #define HAMMER2_THREAD_FREEZE           0x0020  /* force idle */
 825 #define HAMMER2_THREAD_FROZEN           0x0040  /* thread is frozen */
 826 #define HAMMER2_THREAD_XOPQ             0x0080  /* work pending */
 827 #define HAMMER2_THREAD_STOPPED          0x0100  /* thread has stopped */
 828 #define HAMMER2_THREAD_UNFREEZE         0x0200
 829
 830 #define HAMMER2_THREAD_WAKEUP_MASK      (HAMMER2_THREAD_UNMOUNTING |    \
 831                                          HAMMER2_THREAD_REMASTER |      \
 832                                          HAMMER2_THREAD_STOP |          \
 833                                          HAMMER2_THREAD_FREEZE |        \
 834                                          HAMMER2_THREAD_XOPQ)
 835
 836 /*
 837  * Support structure for dedup heuristic.
 838  */
 839 struct hammer2_dedup {
 840         hammer2_off_t   data_off;
 841         uint64_t        data_crc;
 842         uint32_t        ticks;
 843         uint32_t        unused03;
 844 };
 845
 846 typedef struct hammer2_dedup hammer2_dedup_t;
 847
 848 /*
 849  * hammer2_xop - container for VOP/XOP operation (allocated, not on stack).
 850  *
 851  * This structure is used to distribute a VOP operation across multiple
 852  * nodes.  It provides a rendezvous for concurrent node execution and
 853  * can be detached from the frontend operation to allow the frontend to
 854  * return early.
 855  *
 856  * This structure also sequences operations on up to three inodes.
 857  */
 858 typedef void (*hammer2_xop_func_t)(hammer2_thread_t *thr,
 859                                    union hammer2_xop *xop);
 860
 861 struct hammer2_xop_fifo {
 862         TAILQ_ENTRY(hammer2_xop_head) entry;
 863         hammer2_chain_t         *array[HAMMER2_XOPFIFO];
 864         int                     errors[HAMMER2_XOPFIFO];
 865         int                     ri;
 866         int                     wi;
 867         int                     flags;
 868         hammer2_thread_t        *thr;
 869 };
 870
 871 typedef struct hammer2_xop_fifo hammer2_xop_fifo_t;
 872
 873 #define HAMMER2_XOP_FIFO_RUN    0x0001
 874 #define HAMMER2_XOP_FIFO_STALL  0x0002
 875
 876 struct hammer2_xop_head {
 877         hammer2_xop_func_t      func;
 878         hammer2_tid_t           mtid;
 879         struct hammer2_inode    *ip1;
 880         struct hammer2_inode    *ip2;
 881         struct hammer2_inode    *ip3;
 882         uint64_t                run_mask;
 883         uint64_t                chk_mask;
 884         int                     flags;
 885         int                     state;
 886         int                     error;
 887         hammer2_key_t           collect_key;
 888         char                    *name1;
 889         size_t                  name1_len;
 890         char                    *name2;
 891         size_t                  name2_len;
 892         hammer2_xop_fifo_t      collect[HAMMER2_MAXCLUSTER];
 893         hammer2_cluster_t       cluster;        /* help collections */
 894 };
 895
 896 typedef struct hammer2_xop_head hammer2_xop_head_t;
 897
 898 struct hammer2_xop_ipcluster {
 899         hammer2_xop_head_t      head;
 900 };
 901
 902 struct hammer2_xop_strategy {
 903         hammer2_xop_head_t      head;
 904         hammer2_key_t           lbase;
 905         int                     finished;
 906         hammer2_mtx_t           lock;
 907         struct bio              *bio;
 908 };
 909
 910 struct hammer2_xop_readdir {
 911         hammer2_xop_head_t      head;
 912         hammer2_key_t           lkey;
 913 };
 914
 915 struct hammer2_xop_nresolve {
 916         hammer2_xop_head_t      head;
 917         hammer2_key_t           lhc;    /* if name is NULL used lhc */
 918 };
 919
 920 struct hammer2_xop_unlink {
 921         hammer2_xop_head_t      head;
 922         int                     isdir;
 923         int                     dopermanent;
 924 };
 925
 926 #define H2DOPERM_PERMANENT      0x01
 927 #define H2DOPERM_FORCE          0x02
 928 #define H2DOPERM_IGNINO         0x04
 929
 930 struct hammer2_xop_nrename {
 931         hammer2_xop_head_t      head;
 932         hammer2_tid_t           lhc;
 933         int                     ip_key;
 934 };
 935
 936 struct hammer2_xop_scanlhc {
 937         hammer2_xop_head_t      head;
 938         hammer2_key_t           lhc;
 939 };
 940
 941 struct hammer2_xop_scanall {
 942         hammer2_xop_head_t      head;
 943         hammer2_key_t           key_beg;        /* inclusive */
 944         hammer2_key_t           key_end;        /* inclusive */
 945         int                     resolve_flags;
 946         int                     lookup_flags;
 947 };
 948
 949 struct hammer2_xop_lookup {
 950         hammer2_xop_head_t      head;
 951         hammer2_key_t           lhc;
 952 };
 953
 954 struct hammer2_xop_mkdirent {
 955         hammer2_xop_head_t      head;
 956         hammer2_dirent_head_t   dirent;
 957         hammer2_key_t           lhc;
 958 };
 959
 960 struct hammer2_xop_create {
 961         hammer2_xop_head_t      head;
 962         hammer2_inode_meta_t    meta;           /* initial metadata */
 963         hammer2_key_t           lhc;
 964         int                     flags;
 965 };
 966
 967 struct hammer2_xop_destroy {
 968         hammer2_xop_head_t      head;
 969 };
 970
 971 struct hammer2_xop_fsync {
 972         hammer2_xop_head_t      head;
 973         hammer2_inode_meta_t    meta;
 974         hammer2_off_t           osize;
 975         u_int                   ipflags;
 976         int                     clear_directdata;
 977 };
 978
 979 struct hammer2_xop_unlinkall {
 980         hammer2_xop_head_t      head;
 981         hammer2_key_t           key_beg;
 982         hammer2_key_t           key_end;
 983 };
 984
 985 struct hammer2_xop_connect {
 986         hammer2_xop_head_t      head;
 987         hammer2_key_t           lhc;
 988 };
 989
 990 struct hammer2_xop_flush {
 991         hammer2_xop_head_t      head;
 992 };
 993
 994 typedef struct hammer2_xop_readdir hammer2_xop_readdir_t;
 995 typedef struct hammer2_xop_nresolve hammer2_xop_nresolve_t;
 996 typedef struct hammer2_xop_unlink hammer2_xop_unlink_t;
 997 typedef struct hammer2_xop_nrename hammer2_xop_nrename_t;
 998 typedef struct hammer2_xop_ipcluster hammer2_xop_ipcluster_t;
 999 typedef struct hammer2_xop_strategy hammer2_xop_strategy_t;
1000 typedef struct hammer2_xop_mkdirent hammer2_xop_mkdirent_t;
1001 typedef struct hammer2_xop_create hammer2_xop_create_t;
1002 typedef struct hammer2_xop_destroy hammer2_xop_destroy_t;
1003 typedef struct hammer2_xop_fsync hammer2_xop_fsync_t;
1004 typedef struct hammer2_xop_unlinkall hammer2_xop_unlinkall_t;
1005 typedef struct hammer2_xop_scanlhc hammer2_xop_scanlhc_t;
1006 typedef struct hammer2_xop_scanall hammer2_xop_scanall_t;
1007 typedef struct hammer2_xop_lookup hammer2_xop_lookup_t;
1008 typedef struct hammer2_xop_connect hammer2_xop_connect_t;
1009 typedef struct hammer2_xop_flush hammer2_xop_flush_t;
1010
1011 union hammer2_xop {
1012         hammer2_xop_head_t      head;
1013         hammer2_xop_ipcluster_t xop_ipcluster;
1014         hammer2_xop_readdir_t   xop_readdir;
1015         hammer2_xop_nresolve_t  xop_nresolve;
1016         hammer2_xop_unlink_t    xop_unlink;
1017         hammer2_xop_nrename_t   xop_nrename;
1018         hammer2_xop_strategy_t  xop_strategy;
1019         hammer2_xop_mkdirent_t  xop_mkdirent;
1020         hammer2_xop_create_t    xop_create;
1021         hammer2_xop_destroy_t   xop_destroy;
1022         hammer2_xop_fsync_t     xop_fsync;
1023         hammer2_xop_unlinkall_t xop_unlinkall;
1024         hammer2_xop_scanlhc_t   xop_scanlhc;
1025         hammer2_xop_scanall_t   xop_scanall;
1026         hammer2_xop_lookup_t    xop_lookup;
1027         hammer2_xop_flush_t     xop_flush;
1028         hammer2_xop_connect_t   xop_connect;
1029 };
1030
1031 typedef union hammer2_xop hammer2_xop_t;
1032
1033 /*
1034  * hammer2_xop_group - Manage XOP support threads.
1035  */
1036 struct hammer2_xop_group {
1037         hammer2_thread_t        thrs[HAMMER2_MAXCLUSTER];
1038 };
1039
1040 typedef struct hammer2_xop_group hammer2_xop_group_t;
1041
1042 /*
1043  * flags to hammer2_xop_collect()
1044  */
1045 #define HAMMER2_XOP_COLLECT_NOWAIT      0x00000001
1046 #define HAMMER2_XOP_COLLECT_WAITALL     0x00000002
1047
1048 /*
1049  * flags to hammer2_xop_alloc()
1050  *
1051  * MODIFYING    - This is a modifying transaction, allocate a mtid.
1052  * RECURSE      - Recurse top-level inode (for root flushes)
1053  */
1054 #define HAMMER2_XOP_MODIFYING           0x00000001
1055 #define HAMMER2_XOP_STRATEGY            0x00000002
1056 #define HAMMER2_XOP_INODE_STOP          0x00000004
1057
1058 /*
1059  * Global (per partition) management structure, represents a hard block
1060  * device.  Typically referenced by hammer2_chain structures when applicable.
1061  * Typically not used for network-managed elements.
1062  *
1063  * Note that a single hammer2_dev can be indirectly tied to multiple system
1064  * mount points.  There is no direct relationship.  System mounts are
1065  * per-cluster-id, not per-block-device, and a single hard mount might contain
1066  * many PFSs and those PFSs might combine together in various ways to form
1067  * the set of available clusters.
1068  */
1069 struct hammer2_dev {
1070         struct vnode    *devvp;         /* device vnode */
1071         int             ronly;          /* read-only mount */
1072         int             mount_count;    /* number of actively mounted PFSs */
1073         TAILQ_ENTRY(hammer2_dev) mntentry; /* hammer2_mntlist */
1074
1075         struct malloc_type *mchain;
1076         int             nipstacks;
1077         int             maxipstacks;
1078         kdmsg_iocom_t   iocom;          /* volume-level dmsg interface */
1079         struct spinlock io_spin;        /* iotree, iolruq access */
1080         struct hammer2_io_tree iotree;
1081         int             iofree_count;
1082         int             freemap_relaxed;
1083         hammer2_chain_t vchain;         /* anchor chain (topology) */
1084         hammer2_chain_t fchain;         /* anchor chain (freemap) */
1085         struct spinlock list_spin;
1086         struct h2_flush_list flushq;    /* flush seeds */
1087         struct hammer2_pfs *spmp;       /* super-root pmp for transactions */
1088         struct lock     vollk;          /* lockmgr lock */
1089         struct lock     bulklk;         /* bulkfree operation lock */
1090         struct lock     bflock;         /* bulk-free manual function lock */
1091         hammer2_off_t   heur_freemap[HAMMER2_FREEMAP_HEUR_SIZE];
1092         hammer2_dedup_t heur_dedup[HAMMER2_DEDUP_HEUR_SIZE];
1093         int             volhdrno;       /* last volhdrno written */
1094         uint32_t        hflags;         /* HMNT2 flags applicable to device */
1095         hammer2_off_t   free_reserved;  /* nominal free reserved */
1096         hammer2_thread_t bfthr;         /* bulk-free thread */
1097         char            devrepname[64]; /* for kprintf */
1098         hammer2_ioc_bulkfree_t bflast;  /* stats for last bulkfree run */
1099         hammer2_volume_data_t voldata;
1100         hammer2_volume_data_t volsync;  /* synchronized voldata */
1101 };
1102
1103 typedef struct hammer2_dev hammer2_dev_t;
1104
1105 /*
1106  * Helper functions (cluster must be locked for flags to be valid).
1107  */
1108 static __inline
1109 int
1110 hammer2_chain_rdok(hammer2_chain_t *chain)
1111 {
1112         return (chain->error == 0);
1113 }
1114
1115 static __inline
1116 int
1117 hammer2_chain_wrok(hammer2_chain_t *chain)
1118 {
1119         return (chain->error == 0 && chain->hmp->ronly == 0);
1120 }
1121
1122 /*
1123  * Per-cluster management structure.  This structure will be tied to a
1124  * system mount point if the system is mounting the PFS, but is also used
1125  * to manage clusters encountered during the super-root scan or received
1126  * via LNK_SPANs that might not be mounted.
1127  *
1128  * This structure is also used to represent the super-root that hangs off
1129  * of a hard mount point.  The super-root is not really a cluster element.
1130  * In this case the spmp_hmp field will be non-NULL.  It's just easier to do
1131  * this than to special case super-root manipulation in the hammer2_chain*
1132  * code as being only hammer2_dev-related.
1133  *
1134  * pfs_mode and pfs_nmasters are rollup fields which critically describes
1135  * how elements of the cluster act on the cluster.  pfs_mode is only applicable
1136  * when a PFS is mounted by the system.  pfs_nmasters is our best guess as to
1137  * how many masters have been configured for a cluster and is always
1138  * applicable.  pfs_types[] is an array with 1:1 correspondance to the
1139  * iroot cluster and describes the PFS types of the nodes making up the
1140  * cluster.
1141  *
1142  * WARNING! Portions of this structure have deferred initialization.  In
1143  *          particular, if not mounted there will be no wthread.
1144  *          umounted network PFSs will also be missing iroot and numerous
1145  *          other fields will not be initialized prior to mount.
1146  *
1147  *          Synchronization threads are chain-specific and only applicable
1148  *          to local hard PFS entries.  A hammer2_pfs structure may contain
1149  *          more than one when multiple hard PFSs are present on the local
1150  *          machine which require synchronization monitoring.  Most PFSs
1151  *          (such as snapshots) are 1xMASTER PFSs which do not need a
1152  *          synchronization thread.
1153  *
1154  * WARNING! The chains making up pfs->iroot's cluster are accounted for in
1155  *          hammer2_dev->mount_count when the pfs is associated with a mount
1156  *          point.
1157  */
1158 struct hammer2_pfs {
1159         struct mount            *mp;
1160         TAILQ_ENTRY(hammer2_pfs) mntentry;      /* hammer2_pfslist */
1161         uuid_t                  pfs_clid;
1162         hammer2_dev_t           *spmp_hmp;      /* only if super-root pmp */
1163         hammer2_dev_t           *force_local;   /* only if 'local' mount */
1164         hammer2_inode_t         *iroot;         /* PFS root inode */
1165         uint8_t                 pfs_types[HAMMER2_MAXCLUSTER];
1166         char                    *pfs_names[HAMMER2_MAXCLUSTER];
1167         hammer2_dev_t           *pfs_hmps[HAMMER2_MAXCLUSTER];
1168         hammer2_trans_t         trans;
1169         struct lock             lock;           /* PFS lock for certain ops */
1170         struct lock             lock_nlink;     /* rename and nlink lock */
1171         struct netexport        export;         /* nfs export */
1172         int                     ronly;          /* read-only mount */
1173         int                     hflags;         /* pfs-specific mount flags */
1174         struct malloc_type      *minode;
1175         struct malloc_type      *mmsg;
1176         struct spinlock         inum_spin;      /* inumber lookup */
1177         struct hammer2_inode_tree inum_tree;    /* (not applicable to spmp) */
1178         long                    inum_count;     /* #of inodes in inum_tree */
1179         struct spinlock         lru_spin;       /* inumber lookup */
1180         struct hammer2_chain_list lru_list;     /* basis for LRU tests */
1181         int                     lru_count;      /* #of chains on LRU */
1182         hammer2_tid_t           modify_tid;     /* modify transaction id */
1183         hammer2_tid_t           inode_tid;      /* inode allocator */
1184         uint8_t                 pfs_nmasters;   /* total masters */
1185         uint8_t                 pfs_mode;       /* operating mode PFSMODE */
1186         uint8_t                 unused01;
1187         uint8_t                 unused02;
1188         int                     free_ticks;     /* free_* calculations */
1189         long                    inmem_inodes;
1190         hammer2_off_t           free_reserved;
1191         hammer2_off_t           free_nominal;
1192         uint32_t                inmem_dirty_chains;
1193         int                     count_lwinprog; /* logical write in prog */
1194         struct spinlock         list_spin;
1195         struct h2_sideq_list    sideq;          /* last-close dirty/unlink */
1196         long                    sideq_count;
1197         hammer2_thread_t        sync_thrs[HAMMER2_MAXCLUSTER];
1198         uint32_t                cluster_flags;  /* cached cluster flags */
1199         int                     has_xop_threads;
1200         struct spinlock         xop_spin;       /* xop sequencer */
1201         hammer2_xop_group_t     xop_groups[HAMMER2_XOPGROUPS];
1202 };
1203
1204 typedef struct hammer2_pfs hammer2_pfs_t;
1205
1206 TAILQ_HEAD(hammer2_pfslist, hammer2_pfs);
1207
1208 /*
1209  * NOTE: The LRU list contains at least all the chains with refs == 0
1210  *       that can be recycled, and may contain additional chains which
1211  *       cannot.
1212  */
1213 #define HAMMER2_LRU_LIMIT               4096
1214
1215 #define HAMMER2_DIRTYCHAIN_WAITING      0x80000000
1216 #define HAMMER2_DIRTYCHAIN_MASK         0x7FFFFFFF
1217
1218 #define HAMMER2_LWINPROG_WAITING        0x80000000
1219 #define HAMMER2_LWINPROG_WAITING0       0x40000000
1220 #define HAMMER2_LWINPROG_MASK           0x3FFFFFFF
1221
1222 /*
1223  * hammer2_cluster_check
1224  */
1225 #define HAMMER2_CHECK_NULL      0x00000001
1226
1227 /*
1228  * Misc
1229  */
1230 #if defined(_KERNEL) || defined(_KERNEL_STRUCTURES)
1231 #define VTOI(vp)        ((hammer2_inode_t *)(vp)->v_data)
1232 #endif
1233
1234 #if defined(_KERNEL)
1235
1236 MALLOC_DECLARE(M_HAMMER2);
1237
1238 #define ITOV(ip)        ((ip)->vp)
1239
1240 /*
1241  * Currently locked chains retain the locked buffer cache buffer for
1242  * indirect blocks, and indirect blocks can be one of two sizes.  The
1243  * device buffer has to match the case to avoid deadlocking recursive
1244  * chains that might otherwise try to access different offsets within
1245  * the same device buffer.
1246  */
1247 static __inline
1248 int
1249 hammer2_devblkradix(int radix)
1250 {
1251 #if 0
1252         if (radix <= HAMMER2_LBUFRADIX) {
1253                 return (HAMMER2_LBUFRADIX);
1254         } else {
1255                 return (HAMMER2_PBUFRADIX);
1256         }
1257 #endif
1258         return (HAMMER2_PBUFRADIX);
1259 }
1260
1261 /*
1262  * XXX almost time to remove this.  DIO uses PBUFSIZE exclusively now.
1263  */
1264 static __inline
1265 size_t
1266 hammer2_devblksize(size_t bytes)
1267 {
1268 #if 0
1269         if (bytes <= HAMMER2_LBUFSIZE) {
1270                 return(HAMMER2_LBUFSIZE);
1271         } else {
1272                 KKASSERT(bytes <= HAMMER2_PBUFSIZE &&
1273                          (bytes ^ (bytes - 1)) == ((bytes << 1) - 1));
1274                 return (HAMMER2_PBUFSIZE);
1275         }
1276 #endif
1277         return (HAMMER2_PBUFSIZE);
1278 }
1279
1280
1281 static __inline
1282 hammer2_pfs_t *
1283 MPTOPMP(struct mount *mp)
1284 {
1285         return ((hammer2_pfs_t *)mp->mnt_data);
1286 }
1287
1288 #define HAMMER2_DEDUP_FRAG      (HAMMER2_PBUFSIZE / 64)
1289 #define HAMMER2_DEDUP_FRAGRADIX (HAMMER2_PBUFRADIX - 6)
1290
1291 static __inline
1292 uint64_t
1293 hammer2_dedup_mask(hammer2_io_t *dio, hammer2_off_t data_off, u_int bytes)
1294 {
1295         int bbeg;
1296         int bits;
1297         uint64_t mask;
1298
1299         bbeg = (int)((data_off & ~HAMMER2_OFF_MASK_RADIX) - dio->pbase) >>
1300                HAMMER2_DEDUP_FRAGRADIX;
1301         bits = (int)((bytes + (HAMMER2_DEDUP_FRAG - 1)) >>
1302                HAMMER2_DEDUP_FRAGRADIX);
1303         mask = ((uint64_t)1 << bbeg) - 1;
1304         if (bbeg + bits == 64)
1305                 mask = (uint64_t)-1;
1306         else
1307                 mask = ((uint64_t)1 << (bbeg + bits)) - 1;
1308
1309         mask &= ~(((uint64_t)1 << bbeg) - 1);
1310
1311         return mask;
1312 }
1313
1314 static __inline
1315 int
1316 hammer2_error_to_errno(int error)
1317 {
1318         if (error) {
1319                 if (error & HAMMER2_ERROR_EIO)
1320                         error = EIO;
1321                 else if (error & HAMMER2_ERROR_CHECK)
1322                         error = EDOM;
1323                 else if (error & HAMMER2_ERROR_ABORTED)
1324                         error = EINTR;
1325                 else if (error & HAMMER2_ERROR_BADBREF)
1326                         error = EIO;
1327                 else if (error & HAMMER2_ERROR_ENOSPC)
1328                         error = ENOSPC;
1329                 else if (error & HAMMER2_ERROR_ENOENT)
1330                         error = ENOENT;
1331                 else if (error & HAMMER2_ERROR_ENOTEMPTY)
1332                         error = ENOTEMPTY;
1333                 else if (error & HAMMER2_ERROR_EAGAIN)
1334                         error = EAGAIN;
1335                 else if (error & HAMMER2_ERROR_ENOTDIR)
1336                         error = ENOTDIR;
1337                 else if (error & HAMMER2_ERROR_EISDIR)
1338                         error = EISDIR;
1339                 else if (error & HAMMER2_ERROR_EINPROGRESS)
1340                         error = EINPROGRESS;
1341                 else if (error & HAMMER2_ERROR_EEXIST)
1342                         error = EEXIST;
1343                 else
1344                         error = EDOM;
1345         }
1346         return error;
1347 }
1348
1349 static __inline
1350 int
1351 hammer2_errno_to_error(int error)
1352 {
1353         switch(error) {
1354         case 0:
1355                 return 0;
1356         case EIO:
1357                 return HAMMER2_ERROR_EIO;
1358         case EINVAL:
1359         default:
1360                 return HAMMER2_ERROR_EINVAL;
1361         }
1362 }
1363
1364 extern struct vop_ops hammer2_vnode_vops;
1365 extern struct vop_ops hammer2_spec_vops;
1366 extern struct vop_ops hammer2_fifo_vops;
1367 extern struct hammer2_pfslist hammer2_pfslist;
1368 extern struct lock hammer2_mntlk;
1369
1370
1371 extern int hammer2_debug;
1372 extern int hammer2_cluster_meta_read;
1373 extern int hammer2_cluster_data_read;
1374 extern int hammer2_dedup_enable;
1375 extern int hammer2_always_compress;
1376 extern int hammer2_inval_enable;
1377 extern int hammer2_flush_pipe;
1378 extern int hammer2_dio_count;
1379 extern int hammer2_dio_limit;
1380 extern int hammer2_bulkfree_tps;
1381 extern long hammer2_chain_allocs;
1382 extern long hammer2_chain_frees;
1383 extern long hammer2_limit_dirty_chains;
1384 extern long hammer2_count_modified_chains;
1385 extern long hammer2_iod_invals;
1386 extern long hammer2_iod_file_read;
1387 extern long hammer2_iod_meta_read;
1388 extern long hammer2_iod_indr_read;
1389 extern long hammer2_iod_fmap_read;
1390 extern long hammer2_iod_volu_read;
1391 extern long hammer2_iod_file_write;
1392 extern long hammer2_iod_file_wembed;
1393 extern long hammer2_iod_file_wzero;
1394 extern long hammer2_iod_file_wdedup;
1395 extern long hammer2_iod_meta_write;
1396 extern long hammer2_iod_indr_write;
1397 extern long hammer2_iod_fmap_write;
1398 extern long hammer2_iod_volu_write;
1399
1400 extern long hammer2_process_xxhash64;
1401 extern long hammer2_process_icrc32;
1402
1403 extern struct objcache *cache_buffer_read;
1404 extern struct objcache *cache_buffer_write;
1405 extern struct objcache *cache_xops;
1406
1407 /*
1408  * hammer2_subr.c
1409  */
1410 #define hammer2_icrc32(buf, size)       iscsi_crc32((buf), (size))
1411 #define hammer2_icrc32c(buf, size, crc) iscsi_crc32_ext((buf), (size), (crc))
1412
1413 int hammer2_signal_check(time_t *timep);
1414 const char *hammer2_error_str(int error);
1415
1416 void hammer2_inode_lock(hammer2_inode_t *ip, int how);
1417 void hammer2_inode_unlock(hammer2_inode_t *ip);
1418 hammer2_chain_t *hammer2_inode_chain(hammer2_inode_t *ip, int clindex, int how);
1419 hammer2_chain_t *hammer2_inode_chain_and_parent(hammer2_inode_t *ip,
1420                         int clindex, hammer2_chain_t **parentp, int how);
1421 hammer2_mtx_state_t hammer2_inode_lock_temp_release(hammer2_inode_t *ip);
1422 void hammer2_inode_lock_temp_restore(hammer2_inode_t *ip,
1423                         hammer2_mtx_state_t ostate);
1424 int hammer2_inode_lock_upgrade(hammer2_inode_t *ip);
1425 void hammer2_inode_lock_downgrade(hammer2_inode_t *ip, int);
1426
1427 void hammer2_dev_exlock(hammer2_dev_t *hmp);
1428 void hammer2_dev_shlock(hammer2_dev_t *hmp);
1429 void hammer2_dev_unlock(hammer2_dev_t *hmp);
1430
1431 int hammer2_get_dtype(uint8_t type);
1432 int hammer2_get_vtype(uint8_t type);
1433 uint8_t hammer2_get_obj_type(enum vtype vtype);
1434 void hammer2_time_to_timespec(uint64_t xtime, struct timespec *ts);
1435 uint64_t hammer2_timespec_to_time(const struct timespec *ts);
1436 uint32_t hammer2_to_unix_xid(const uuid_t *uuid);
1437 void hammer2_guid_to_uuid(uuid_t *uuid, uint32_t guid);
1438 void hammer2_trans_manage_init(hammer2_pfs_t *pmp);
1439
1440 hammer2_key_t hammer2_dirhash(const unsigned char *name, size_t len);
1441 int hammer2_getradix(size_t bytes);
1442
1443 int hammer2_calc_logical(hammer2_inode_t *ip, hammer2_off_t uoff,
1444                         hammer2_key_t *lbasep, hammer2_key_t *leofp);
1445 int hammer2_calc_physical(hammer2_inode_t *ip, hammer2_key_t lbase);
1446 void hammer2_update_time(uint64_t *timep);
1447 void hammer2_adjreadcounter(hammer2_blockref_t *bref, size_t bytes);
1448
1449 /*
1450  * hammer2_inode.c
1451  */
1452 struct vnode *hammer2_igetv(hammer2_inode_t *ip, int *errorp);
1453 hammer2_inode_t *hammer2_inode_lookup(hammer2_pfs_t *pmp,
1454                         hammer2_tid_t inum);
1455 hammer2_inode_t *hammer2_inode_get(hammer2_pfs_t *pmp, hammer2_inode_t *dip,
1456                         hammer2_cluster_t *cluster, int idx);
1457 void hammer2_inode_free(hammer2_inode_t *ip);
1458 void hammer2_inode_ref(hammer2_inode_t *ip);
1459 void hammer2_inode_drop(hammer2_inode_t *ip);
1460 void hammer2_inode_repoint(hammer2_inode_t *ip, hammer2_inode_t *pip,
1461                         hammer2_cluster_t *cluster);
1462 void hammer2_inode_repoint_one(hammer2_inode_t *ip, hammer2_cluster_t *cluster,
1463                         int idx);
1464 void hammer2_inode_modify(hammer2_inode_t *ip);
1465 void hammer2_inode_run_sideq(hammer2_pfs_t *pmp, int doall);
1466
1467 hammer2_inode_t *hammer2_inode_create(hammer2_inode_t *dip,
1468                         hammer2_inode_t *pip,
1469                         struct vattr *vap, struct ucred *cred,
1470                         const uint8_t *name, size_t name_len, hammer2_key_t lhc,
1471                         hammer2_key_t inum, uint8_t type, uint8_t target_type,
1472                         int flags, int *errorp);
1473 int hammer2_inode_chain_sync(hammer2_inode_t *ip);
1474 int hammer2_inode_chain_flush(hammer2_inode_t *ip);
1475 int hammer2_inode_unlink_finisher(hammer2_inode_t *ip, int isopen);
1476 int hammer2_dirent_create(hammer2_inode_t *dip, const char *name,
1477                         size_t name_len, hammer2_key_t inum, uint8_t type);
1478
1479 /*
1480  * hammer2_chain.c
1481  */
1482 void hammer2_voldata_lock(hammer2_dev_t *hmp);
1483 void hammer2_voldata_unlock(hammer2_dev_t *hmp);
1484 void hammer2_voldata_modify(hammer2_dev_t *hmp);
1485 hammer2_chain_t *hammer2_chain_alloc(hammer2_dev_t *hmp,
1486                                 hammer2_pfs_t *pmp,
1487                                 hammer2_blockref_t *bref);
1488 void hammer2_chain_core_init(hammer2_chain_t *chain);
1489 void hammer2_chain_ref(hammer2_chain_t *chain);
1490 void hammer2_chain_ref_hold(hammer2_chain_t *chain);
1491 void hammer2_chain_drop(hammer2_chain_t *chain);
1492 void hammer2_chain_drop_unhold(hammer2_chain_t *chain);
1493 int hammer2_chain_lock(hammer2_chain_t *chain, int how);
1494 void hammer2_chain_lock_unhold(hammer2_chain_t *chain, int how);
1495 void hammer2_chain_load_data(hammer2_chain_t *chain);
1496 const hammer2_media_data_t *hammer2_chain_rdata(hammer2_chain_t *chain);
1497 hammer2_media_data_t *hammer2_chain_wdata(hammer2_chain_t *chain);
1498
1499 int hammer2_chain_inode_find(hammer2_pfs_t *pmp, hammer2_key_t inum,
1500                                 int clindex, int flags,
1501                                 hammer2_chain_t **parentp,
1502                                 hammer2_chain_t **chainp);
1503 int hammer2_chain_modify(hammer2_chain_t *chain, hammer2_tid_t mtid,
1504                                 hammer2_off_t dedup_off, int flags);
1505 int hammer2_chain_modify_ip(hammer2_inode_t *ip, hammer2_chain_t *chain,
1506                                 hammer2_tid_t mtid, int flags);
1507 int hammer2_chain_resize(hammer2_chain_t *chain,
1508                                 hammer2_tid_t mtid, hammer2_off_t dedup_off,
1509                                 int nradix, int flags);
1510 void hammer2_chain_unlock(hammer2_chain_t *chain);
1511 void hammer2_chain_unlock_hold(hammer2_chain_t *chain);
1512 void hammer2_chain_wait(hammer2_chain_t *chain);
1513 hammer2_chain_t *hammer2_chain_get(hammer2_chain_t *parent, int generation,
1514                                 hammer2_blockref_t *bref, int how);
1515 hammer2_chain_t *hammer2_chain_lookup_init(hammer2_chain_t *parent, int flags);
1516 void hammer2_chain_lookup_done(hammer2_chain_t *parent);
1517 hammer2_chain_t *hammer2_chain_getparent(hammer2_chain_t *chain, int flags);
1518 hammer2_chain_t *hammer2_chain_repparent(hammer2_chain_t **chainp, int flags);
1519 hammer2_chain_t *hammer2_chain_lookup(hammer2_chain_t **parentp,
1520                                 hammer2_key_t *key_nextp,
1521                                 hammer2_key_t key_beg, hammer2_key_t key_end,
1522                                 int *errorp, int flags);
1523 hammer2_chain_t *hammer2_chain_next(hammer2_chain_t **parentp,
1524                                 hammer2_chain_t *chain,
1525                                 hammer2_key_t *key_nextp,
1526                                 hammer2_key_t key_beg, hammer2_key_t key_end,
1527                                 int *errorp, int flags);
1528 int hammer2_chain_scan(hammer2_chain_t *parent,
1529                                 hammer2_chain_t **chainp,
1530                                 hammer2_blockref_t *bref,
1531                                 int *firstp, int flags);
1532
1533 int hammer2_chain_create(hammer2_chain_t **parentp, hammer2_chain_t **chainp,
1534                                 hammer2_pfs_t *pmp, int methods,
1535                                 hammer2_key_t key, int keybits,
1536                                 int type, size_t bytes, hammer2_tid_t mtid,
1537                                 hammer2_off_t dedup_off, int flags);
1538 void hammer2_chain_rename(hammer2_chain_t **parentp,
1539                                 hammer2_chain_t *chain,
1540                                 hammer2_tid_t mtid, int flags);
1541 int hammer2_chain_delete(hammer2_chain_t *parent, hammer2_chain_t *chain,
1542                                 hammer2_tid_t mtid, int flags);
1543 int hammer2_chain_indirect_maintenance(hammer2_chain_t *parent,
1544                                 hammer2_chain_t *chain);
1545 void hammer2_chain_setflush(hammer2_chain_t *chain);
1546 void hammer2_chain_countbrefs(hammer2_chain_t *chain,
1547                                 hammer2_blockref_t *base, int count);
1548 hammer2_chain_t *hammer2_chain_bulksnap(hammer2_dev_t *hmp);
1549 void hammer2_chain_bulkdrop(hammer2_chain_t *copy);
1550
1551 void hammer2_chain_setcheck(hammer2_chain_t *chain, void *bdata);
1552 int hammer2_chain_testcheck(hammer2_chain_t *chain, void *bdata);
1553 int hammer2_chain_dirent_test(hammer2_chain_t *chain, const char *name,
1554                                 size_t name_len);
1555
1556 void hammer2_pfs_memory_wait(hammer2_pfs_t *pmp);
1557 void hammer2_pfs_memory_inc(hammer2_pfs_t *pmp);
1558 void hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp);
1559
1560 void hammer2_base_delete(hammer2_chain_t *parent,
1561                                 hammer2_blockref_t *base, int count,
1562                                 hammer2_chain_t *chain);
1563 void hammer2_base_insert(hammer2_chain_t *parent,
1564                                 hammer2_blockref_t *base, int count,
1565                                 hammer2_chain_t *chain,
1566                                 hammer2_blockref_t *elm);
1567
1568 /*
1569  * hammer2_flush.c
1570  */
1571 int hammer2_flush(hammer2_chain_t *chain, int istop);
1572 void hammer2_delayed_flush(hammer2_chain_t *chain);
1573
1574 /*
1575  * hammer2_trans.c
1576  */
1577 void hammer2_trans_init(hammer2_pfs_t *pmp, uint32_t flags);
1578 hammer2_tid_t hammer2_trans_sub(hammer2_pfs_t *pmp);
1579 void hammer2_trans_done(hammer2_pfs_t *pmp);
1580 hammer2_tid_t hammer2_trans_newinum(hammer2_pfs_t *pmp);
1581 void hammer2_trans_assert_strategy(hammer2_pfs_t *pmp);
1582 void hammer2_dedup_record(hammer2_chain_t *chain, hammer2_io_t *dio,
1583                                 char *data);
1584
1585 /*
1586  * hammer2_ioctl.c
1587  */
1588 int hammer2_ioctl(hammer2_inode_t *ip, u_long com, void *data,
1589                                 int fflag, struct ucred *cred);
1590
1591 /*
1592  * hammer2_io.c
1593  */
1594 void hammer2_io_putblk(hammer2_io_t **diop);
1595 void hammer2_io_inval(hammer2_io_t *dio, hammer2_off_t data_off, u_int bytes);
1596 void hammer2_io_cleanup(hammer2_dev_t *hmp, struct hammer2_io_tree *tree);
1597 char *hammer2_io_data(hammer2_io_t *dio, off_t lbase);
1598 void hammer2_io_bkvasync(hammer2_io_t *dio);
1599 hammer2_io_t *hammer2_io_getblk(hammer2_dev_t *hmp, int btype, off_t lbase,
1600                                 int lsize, int op);
1601 void hammer2_io_dedup_set(hammer2_dev_t *hmp, hammer2_blockref_t *bref);
1602 void hammer2_io_dedup_delete(hammer2_dev_t *hmp, uint8_t btype,
1603                                 hammer2_off_t data_off, u_int bytes);
1604 void hammer2_io_dedup_assert(hammer2_dev_t *hmp, hammer2_off_t data_off,
1605                                 u_int bytes);
1606 void hammer2_io_callback(struct bio *bio);
1607 int hammer2_io_new(hammer2_dev_t *hmp, int btype, off_t lbase, int lsize,
1608                                 hammer2_io_t **diop);
1609 int hammer2_io_newnz(hammer2_dev_t *hmp, int btype, off_t lbase, int lsize,
1610                                 hammer2_io_t **diop);
1611 int hammer2_io_bread(hammer2_dev_t *hmp, int btype, off_t lbase, int lsize,
1612                                 hammer2_io_t **diop);
1613 hammer2_io_t *hammer2_io_getquick(hammer2_dev_t *hmp, off_t lbase, int lsize);
1614 void hammer2_io_bawrite(hammer2_io_t **diop);
1615 void hammer2_io_bdwrite(hammer2_io_t **diop);
1616 int hammer2_io_bwrite(hammer2_io_t **diop);
1617 void hammer2_io_setdirty(hammer2_io_t *dio);
1618 void hammer2_io_brelse(hammer2_io_t **diop);
1619 void hammer2_io_bqrelse(hammer2_io_t **diop);
1620
1621 /*
1622  * hammer2_thread.c
1623  */
1624 void hammer2_thr_signal(hammer2_thread_t *thr, uint32_t flags);
1625 void hammer2_thr_signal2(hammer2_thread_t *thr,
1626                         uint32_t pflags, uint32_t nflags);
1627 void hammer2_thr_wait(hammer2_thread_t *thr, uint32_t flags);
1628 void hammer2_thr_wait_neg(hammer2_thread_t *thr, uint32_t flags);
1629 int hammer2_thr_wait_any(hammer2_thread_t *thr, uint32_t flags, int timo);
1630 void hammer2_thr_create(hammer2_thread_t *thr,
1631                         hammer2_pfs_t *pmp, hammer2_dev_t *hmp,
1632                         const char *id, int clindex, int repidx,
1633                         void (*func)(void *arg));
1634 void hammer2_thr_delete(hammer2_thread_t *thr);
1635 void hammer2_thr_remaster(hammer2_thread_t *thr);
1636 void hammer2_thr_freeze_async(hammer2_thread_t *thr);
1637 void hammer2_thr_freeze(hammer2_thread_t *thr);
1638 void hammer2_thr_unfreeze(hammer2_thread_t *thr);
1639 int hammer2_thr_break(hammer2_thread_t *thr);
1640 void hammer2_primary_xops_thread(void *arg);
1641
1642 /*
1643  * hammer2_thread.c (XOP API)
1644  */
1645 void hammer2_xop_group_init(hammer2_pfs_t *pmp, hammer2_xop_group_t *xgrp);
1646 void *hammer2_xop_alloc(hammer2_inode_t *ip, int flags);
1647 void hammer2_xop_setname(hammer2_xop_head_t *xop,
1648                                 const char *name, size_t name_len);
1649 void hammer2_xop_setname2(hammer2_xop_head_t *xop,
1650                                 const char *name, size_t name_len);
1651 size_t hammer2_xop_setname_inum(hammer2_xop_head_t *xop, hammer2_key_t inum);
1652 void hammer2_xop_setip2(hammer2_xop_head_t *xop, hammer2_inode_t *ip2);
1653 void hammer2_xop_setip3(hammer2_xop_head_t *xop, hammer2_inode_t *ip3);
1654 void hammer2_xop_reinit(hammer2_xop_head_t *xop);
1655 void hammer2_xop_helper_create(hammer2_pfs_t *pmp);
1656 void hammer2_xop_helper_cleanup(hammer2_pfs_t *pmp);
1657 void hammer2_xop_start(hammer2_xop_head_t *xop, hammer2_xop_func_t func);
1658 void hammer2_xop_start_except(hammer2_xop_head_t *xop, hammer2_xop_func_t func,
1659                                 int notidx);
1660 int hammer2_xop_collect(hammer2_xop_head_t *xop, int flags);
1661 void hammer2_xop_retire(hammer2_xop_head_t *xop, uint64_t mask);
1662 int hammer2_xop_active(hammer2_xop_head_t *xop);
1663 int hammer2_xop_feed(hammer2_xop_head_t *xop, hammer2_chain_t *chain,
1664                                 int clindex, int error);
1665
1666 /*
1667  * hammer2_synchro.c
1668  */
1669 void hammer2_primary_sync_thread(void *arg);
1670
1671 /*
1672  * XOP backends in hammer2_xops.c, primarily for VNOPS.  Other XOP backends
1673  * may be integrated into other source files.
1674  */
1675 void hammer2_xop_ipcluster(hammer2_thread_t *thr, hammer2_xop_t *xop);
1676 void hammer2_xop_readdir(hammer2_thread_t *thr, hammer2_xop_t *xop);
1677 void hammer2_xop_nresolve(hammer2_thread_t *thr, hammer2_xop_t *xop);
1678 void hammer2_xop_unlink(hammer2_thread_t *thr, hammer2_xop_t *xop);
1679 void hammer2_xop_nrename(hammer2_thread_t *thr, hammer2_xop_t *xop);
1680 void hammer2_xop_scanlhc(hammer2_thread_t *thr, hammer2_xop_t *xop);
1681 void hammer2_xop_scanall(hammer2_thread_t *thr, hammer2_xop_t *xop);
1682 void hammer2_xop_lookup(hammer2_thread_t *thr, hammer2_xop_t *xop);
1683 void hammer2_inode_xop_mkdirent(hammer2_thread_t *thr, hammer2_xop_t *xop);
1684 void hammer2_inode_xop_create(hammer2_thread_t *thr, hammer2_xop_t *xop);
1685 void hammer2_inode_xop_destroy(hammer2_thread_t *thr, hammer2_xop_t *xop);
1686 void hammer2_inode_xop_chain_sync(hammer2_thread_t *thr, hammer2_xop_t *xop);
1687 void hammer2_inode_xop_unlinkall(hammer2_thread_t *thr, hammer2_xop_t *xop);
1688 void hammer2_inode_xop_connect(hammer2_thread_t *thr, hammer2_xop_t *xop);
1689 void hammer2_inode_xop_flush(hammer2_thread_t *thr, hammer2_xop_t *xop);
1690
1691 /*
1692  * hammer2_msgops.c
1693  */
1694 int hammer2_msg_dbg_rcvmsg(kdmsg_msg_t *msg);
1695 int hammer2_msg_adhoc_input(kdmsg_msg_t *msg);
1696
1697 /*
1698  * hammer2_vfsops.c
1699  */
1700 void hammer2_volconf_update(hammer2_dev_t *hmp, int index);
1701 void hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx,
1702                                 u_int flags);
1703 int hammer2_vfs_sync(struct mount *mp, int waitflags);
1704 int hammer2_vfs_enospace(hammer2_inode_t *ip, off_t bytes, struct ucred *cred);
1705
1706 hammer2_pfs_t *hammer2_pfsalloc(hammer2_chain_t *chain,
1707                                 const hammer2_inode_data_t *ripdata,
1708                                 hammer2_tid_t modify_tid,
1709                                 hammer2_dev_t *force_local);
1710 void hammer2_pfsdealloc(hammer2_pfs_t *pmp, int clindex, int destroying);
1711 int hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1712                                 ino_t ino, struct vnode **vpp);
1713
1714 void hammer2_lwinprog_ref(hammer2_pfs_t *pmp);
1715 void hammer2_lwinprog_drop(hammer2_pfs_t *pmp);
1716 void hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int pipe);
1717
1718 /*
1719  * hammer2_freemap.c
1720  */
1721 int hammer2_freemap_alloc(hammer2_chain_t *chain, size_t bytes);
1722 void hammer2_freemap_adjust(hammer2_dev_t *hmp,
1723                                 hammer2_blockref_t *bref, int how);
1724
1725 /*
1726  * hammer2_cluster.c
1727  */
1728 uint8_t hammer2_cluster_type(hammer2_cluster_t *cluster);
1729 const hammer2_media_data_t *hammer2_cluster_rdata(hammer2_cluster_t *cluster);
1730 hammer2_media_data_t *hammer2_cluster_wdata(hammer2_cluster_t *cluster);
1731 hammer2_cluster_t *hammer2_cluster_from_chain(hammer2_chain_t *chain);
1732 void hammer2_cluster_bref(hammer2_cluster_t *cluster, hammer2_blockref_t *bref);
1733 hammer2_cluster_t *hammer2_cluster_alloc(hammer2_pfs_t *pmp,
1734                                 hammer2_blockref_t *bref);
1735 void hammer2_cluster_ref(hammer2_cluster_t *cluster);
1736 void hammer2_cluster_drop(hammer2_cluster_t *cluster);
1737 void hammer2_cluster_lock(hammer2_cluster_t *cluster, int how);
1738 int hammer2_cluster_check(hammer2_cluster_t *cluster, hammer2_key_t lokey,
1739                         int flags);
1740 void hammer2_cluster_resolve(hammer2_cluster_t *cluster);
1741 void hammer2_cluster_forcegood(hammer2_cluster_t *cluster);
1742 void hammer2_cluster_unlock(hammer2_cluster_t *cluster);
1743
1744 void hammer2_bulkfree_init(hammer2_dev_t *hmp);
1745 void hammer2_bulkfree_uninit(hammer2_dev_t *hmp);
1746 int hammer2_bulkfree_pass(hammer2_dev_t *hmp, hammer2_chain_t *vchain,
1747                         struct hammer2_ioc_bulkfree *bfi);
1748
1749 /*
1750  * hammer2_iocom.c
1751  */
1752 void hammer2_iocom_init(hammer2_dev_t *hmp);
1753 void hammer2_iocom_uninit(hammer2_dev_t *hmp);
1754 void hammer2_cluster_reconnect(hammer2_dev_t *hmp, struct file *fp);
1755
1756 /*
1757  * hammer2_strategy.c
1758  */
1759 int hammer2_vop_strategy(struct vop_strategy_args *ap);
1760 int hammer2_vop_bmap(struct vop_bmap_args *ap);
1761 void hammer2_write_thread(void *arg);
1762 void hammer2_bioq_sync(hammer2_pfs_t *pmp);
1763 void hammer2_dedup_clear(hammer2_dev_t *hmp);
1764
1765 #endif /* !_KERNEL */
1766 #endif /* !_VFS_HAMMER2_HAMMER2_H_ */