sys/vfs/hammer2/hammer2.h

   1 /*
   2  * Copyright (c) 2011-2016 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_xxhash.h"
  94 #include "hammer2_disk.h"
  95 #include "hammer2_mount.h"
  96 #include "hammer2_ioctl.h"
  97
  98 struct hammer2_io;
  99 struct hammer2_iocb;
 100 struct hammer2_chain;
 101 struct hammer2_cluster;
 102 struct hammer2_inode;
 103 struct hammer2_dev;
 104 struct hammer2_pfs;
 105 struct hammer2_span;
 106 struct hammer2_state;
 107 struct hammer2_msg;
 108 struct hammer2_thread;
 109 union hammer2_xop;
 110
 111 /*
 112  * Mutex and lock shims.  Hammer2 requires support for asynchronous and
 113  * abortable locks, and both exclusive and shared spinlocks.  Normal
 114  * synchronous non-abortable locks can be substituted for spinlocks.
 115  */
 116 typedef mtx_t                           hammer2_mtx_t;
 117 typedef mtx_link_t                      hammer2_mtx_link_t;
 118 typedef mtx_state_t                     hammer2_mtx_state_t;
 119
 120 typedef struct spinlock                 hammer2_spin_t;
 121
 122 #define hammer2_mtx_ex                  mtx_lock_ex_quick
 123 #define hammer2_mtx_sh                  mtx_lock_sh_quick
 124 #define hammer2_mtx_sh_again            mtx_lock_sh_again
 125 #define hammer2_mtx_unlock              mtx_unlock
 126 #define hammer2_mtx_downgrade           mtx_downgrade
 127 #define hammer2_mtx_owned               mtx_owned
 128 #define hammer2_mtx_init                mtx_init
 129 #define hammer2_mtx_temp_release        mtx_lock_temp_release
 130 #define hammer2_mtx_temp_restore        mtx_lock_temp_restore
 131 #define hammer2_mtx_refs                mtx_lockrefs
 132
 133 #define hammer2_spin_init               spin_init
 134 #define hammer2_spin_sh                 spin_lock_shared
 135 #define hammer2_spin_ex                 spin_lock
 136 #define hammer2_spin_unsh               spin_unlock_shared
 137 #define hammer2_spin_unex               spin_unlock
 138
 139 TAILQ_HEAD(hammer2_xop_list, hammer2_xop_head);
 140 TAILQ_HEAD(hammer2_chain_list, hammer2_chain);
 141
 142 typedef struct hammer2_xop_list hammer2_xop_list_t;
 143
 144
 145 /*
 146  * General lock support
 147  */
 148 static __inline
 149 int
 150 hammer2_mtx_upgrade(hammer2_mtx_t *mtx)
 151 {
 152         int wasexclusive;
 153
 154         if (mtx_islocked_ex(mtx)) {
 155                 wasexclusive = 1;
 156         } else {
 157                 mtx_unlock(mtx);
 158                 mtx_lock_ex_quick(mtx);
 159                 wasexclusive = 0;
 160         }
 161         return wasexclusive;
 162 }
 163
 164 /*
 165  * The xid tracks internal transactional updates.
 166  *
 167  * XXX fix-me, really needs to be 64-bits
 168  */
 169 typedef uint32_t hammer2_xid_t;
 170
 171 #define HAMMER2_XID_MIN 0x00000000U
 172 #define HAMMER2_XID_MAX 0x7FFFFFFFU
 173
 174 /*
 175  * The chain structure tracks a portion of the media topology from the
 176  * root (volume) down.  Chains represent volumes, inodes, indirect blocks,
 177  * data blocks, and freemap nodes and leafs.
 178  *
 179  * The chain structure utilizes a simple singly-homed topology and the
 180  * chain's in-memory topology will move around as the chains do, due mainly
 181  * to renames and indirect block creation.
 182  *
 183  * Block Table Updates
 184  *
 185  *      Block table updates for insertions and updates are delayed until the
 186  *      flush.  This allows us to avoid having to modify the parent chain
 187  *      all the way to the root.
 188  *
 189  *      Block table deletions are performed immediately (modifying the parent
 190  *      in the process) because the flush code uses the chain structure to
 191  *      track delayed updates and the chain will be (likely) gone or moved to
 192  *      another location in the topology after a deletion.
 193  *
 194  *      A prior iteration of the code tried to keep the relationship intact
 195  *      on deletes by doing a delete-duplicate operation on the chain, but
 196  *      it added way too much complexity to the codebase.
 197  *
 198  * Flush Synchronization
 199  *
 200  *      The flush code must flush modified chains bottom-up.  Because chain
 201  *      structures can shift around and are NOT topologically stable,
 202  *      modified chains are independently indexed for the flush.  As the flush
 203  *      runs it modifies (or further modifies) and updates the parents,
 204  *      propagating the flush all the way to the volume root.
 205  *
 206  *      Modifying front-end operations can occur during a flush but will block
 207  *      in two cases: (1) when the front-end tries to operate on the inode
 208  *      currently in the midst of being flushed and (2) if the front-end
 209  *      crosses an inode currently being flushed (such as during a rename).
 210  *      So, for example, if you rename directory "x" to "a/b/c/d/e/f/g/x" and
 211  *      the flusher is currently working on "a/b/c", the rename will block
 212  *      temporarily in order to ensure that "x" exists in one place or the
 213  *      other.
 214  *
 215  *      Meta-data statistics are updated by the flusher.  The front-end will
 216  *      make estimates but meta-data must be fully synchronized only during a
 217  *      flush in order to ensure that it remains correct across a crash.
 218  *
 219  *      Multiple flush synchronizations can theoretically be in-flight at the
 220  *      same time but the implementation is not coded to handle the case and
 221  *      currently serializes them.
 222  *
 223  * Snapshots:
 224  *
 225  *      Snapshots currently require the subdirectory tree being snapshotted
 226  *      to be flushed.  The snapshot then creates a new super-root inode which
 227  *      copies the flushed blockdata of the directory or file that was
 228  *      snapshotted.
 229  *
 230  * RBTREE NOTES:
 231  *
 232  *      - Note that the radix tree runs in powers of 2 only so sub-trees
 233  *        cannot straddle edges.
 234  */
 235 RB_HEAD(hammer2_chain_tree, hammer2_chain);
 236 TAILQ_HEAD(h2_flush_list, hammer2_chain);
 237 TAILQ_HEAD(h2_core_list, hammer2_chain);
 238 TAILQ_HEAD(h2_iocb_list, hammer2_iocb);
 239
 240 #define CHAIN_CORE_DELETE_BMAP_ENTRIES  \
 241         (HAMMER2_PBUFSIZE / sizeof(hammer2_blockref_t) / sizeof(uint32_t))
 242
 243 /*
 244  * Core topology for chain (embedded in chain).  Protected by a spinlock.
 245  */
 246 struct hammer2_chain_core {
 247         hammer2_spin_t  spin;
 248         struct hammer2_chain_tree rbtree; /* sub-chains */
 249         int             live_zero;      /* blockref array opt */
 250         u_int           live_count;     /* live (not deleted) chains in tree */
 251         u_int           chain_count;    /* live + deleted chains under core */
 252         int             generation;     /* generation number (inserts only) */
 253 };
 254
 255 typedef struct hammer2_chain_core hammer2_chain_core_t;
 256
 257 RB_HEAD(hammer2_io_tree, hammer2_io);
 258
 259 /*
 260  * IOCB - IO callback (into chain, cluster, or manual request)
 261  */
 262 struct hammer2_iocb {
 263         TAILQ_ENTRY(hammer2_iocb) entry;
 264         void (*callback)(struct hammer2_iocb *iocb);
 265         struct hammer2_io       *dio;
 266         struct hammer2_cluster  *cluster;
 267         struct hammer2_chain    *chain;
 268         void                    *ptr;
 269         off_t                   lbase;
 270         int                     lsize;
 271         uint32_t                flags;
 272         int                     error;
 273         int                     btype;
 274 };
 275
 276 typedef struct hammer2_iocb hammer2_iocb_t;
 277
 278 #define HAMMER2_IOCB_INTERLOCK  0x00000001
 279 #define HAMMER2_IOCB_ONQ        0x00000002
 280 #define HAMMER2_IOCB_DONE       0x00000004
 281 #define HAMMER2_IOCB_INPROG     0x00000008
 282 #define HAMMER2_IOCB_UNUSED10   0x00000010
 283 #define HAMMER2_IOCB_QUICK      0x00010000
 284 #define HAMMER2_IOCB_ZERO       0x00020000
 285 #define HAMMER2_IOCB_READ       0x00040000
 286 #define HAMMER2_IOCB_WAKEUP     0x00080000
 287
 288 /*
 289  * DIO - Management structure wrapping system buffer cache.
 290  *
 291  *       Used for multiple purposes including concurrent management
 292  *       if small requests by chains into larger DIOs.
 293  */
 294 struct hammer2_io {
 295         RB_ENTRY(hammer2_io) rbnode;    /* indexed by device offset */
 296         struct h2_iocb_list iocbq;
 297         struct spinlock spin;
 298         struct hammer2_dev *hmp;
 299         struct buf      *bp;
 300         uint64_t        crc_good_mask;
 301         off_t           pbase;
 302         int             psize;
 303         int             refs;
 304         int             act;            /* activity */
 305         int             btype;          /* approximate BREF_TYPE_* */
 306 };
 307
 308 typedef struct hammer2_io hammer2_io_t;
 309
 310 #define HAMMER2_DIO_INPROG      0x80000000      /* bio in progress */
 311 #define HAMMER2_DIO_GOOD        0x40000000      /* dio->bp is stable */
 312 #define HAMMER2_DIO_WAITING     0x20000000      /* wait for inprog clr */
 313 #define HAMMER2_DIO_DIRTY       0x10000000      /* flush on last drop */
 314
 315 #define HAMMER2_DIO_MASK        0x0FFFFFFF
 316
 317 /*
 318  * Primary chain structure keeps track of the topology in-memory.
 319  */
 320 struct hammer2_chain {
 321         hammer2_mtx_t           lock;
 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_io_t    *dio;                   /* physical data buffer */
 331         u_int           bytes;                  /* physical data size */
 332         u_int           flags;
 333         u_int           refs;
 334         u_int           lockcnt;
 335         int             error;                  /* on-lock data error state */
 336         int             persist_refs;           /* (aka ip->cluster) */
 337
 338         hammer2_media_data_t *data;             /* data pointer shortcut */
 339         TAILQ_ENTRY(hammer2_chain) flush_node;  /* flush list */
 340         TAILQ_ENTRY(hammer2_chain) lru_node;    /* 0-refs LRU */
 341 };
 342
 343 typedef struct hammer2_chain hammer2_chain_t;
 344
 345 int hammer2_chain_cmp(hammer2_chain_t *chain1, hammer2_chain_t *chain2);
 346 RB_PROTOTYPE(hammer2_chain_tree, hammer2_chain, rbnode, hammer2_chain_cmp);
 347
 348 /*
 349  * Special notes on flags:
 350  *
 351  * INITIAL      - This flag allows a chain to be created and for storage to
 352  *                be allocated without having to immediately instantiate the
 353  *                related buffer.  The data is assumed to be all-zeros.  It
 354  *                is primarily used for indirect blocks.
 355  *
 356  * MODIFIED     - The chain's media data has been modified.  Prevents chain
 357  *                free on lastdrop if still in the topology.
 358  *
 359  * UPDATE       - Chain might not be modified but parent blocktable needs
 360  *                an update.  Prevents chain free on lastdrop if still in
 361  *                the topology.
 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_DEDUP             0x00000008      /* used as dedup src */
 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_TESTEDGOOD        0x00000100      /* crc tested good */
 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_DELAYED           0x00001000      /* delayed flush */
 384 #define HAMMER2_CHAIN_COUNTEDBREFS      0x00002000      /* block table stats */
 385 #define HAMMER2_CHAIN_ONRBTREE          0x00004000      /* on parent RB tree */
 386 #define HAMMER2_CHAIN_ONLRU             0x00008000      /* on LRU list */
 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_IOINPROG          0x00100000      /* I/O interlock */
 392 #define HAMMER2_CHAIN_IOSIGNAL          0x00200000      /* I/O interlock */
 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  * NOTES:
 426  *      NOLOCK      - Input and output chains are referenced only and not
 427  *                    locked.  Output chain might be temporarily locked
 428  *                    internally.
 429  *
 430  *      NODATA      - Asks that the chain->data not be resolved in order
 431  *                    to avoid I/O.
 432  *
 433  *      NODIRECT    - Prevents a lookup of offset 0 in an inode from returning
 434  *                    the inode itself if the inode is in DIRECTDATA mode
 435  *                    (i.e. file is <= 512 bytes).  Used by the synchronization
 436  *                    code to prevent confusion.
 437  *
 438  *      SHARED      - The input chain is expected to be locked shared,
 439  *                    and the output chain is locked shared.
 440  *
 441  *      MATCHIND    - Allows an indirect block / freemap node to be returned
 442  *                    when the passed key range matches the radix.  Remember
 443  *                    that key_end is inclusive (e.g. {0x000,0xFFF},
 444  *                    not {0x000,0x1000}).
 445  *
 446  *                    (Cannot be used for remote or cluster ops).
 447  *
 448  *      ALLNODES    - Allows NULL focus.
 449  *
 450  *      ALWAYS      - Always resolve the data.  If ALWAYS and NODATA are both
 451  *                    missing, bulk file data is not resolved but inodes and
 452  *                    other meta-data will.
 453  *
 454  *      NOUNLOCK    - Used by hammer2_chain_next() to leave the lock on
 455  *                    the input chain intact.  The chain is still dropped.
 456  *                    This allows the caller to add a reference to the chain
 457  *                    and retain it in a locked state (used by the
 458  *                    XOP/feed/collect code).
 459  */
 460 #define HAMMER2_LOOKUP_NOLOCK           0x00000001      /* ref only */
 461 #define HAMMER2_LOOKUP_NODATA           0x00000002      /* data left NULL */
 462 #define HAMMER2_LOOKUP_NODIRECT         0x00000004      /* no offset=0 DD */
 463 #define HAMMER2_LOOKUP_SHARED           0x00000100
 464 #define HAMMER2_LOOKUP_MATCHIND         0x00000200      /* return all chains */
 465 #define HAMMER2_LOOKUP_ALLNODES         0x00000400      /* allow NULL focus */
 466 #define HAMMER2_LOOKUP_ALWAYS           0x00000800      /* resolve data */
 467 #define HAMMER2_LOOKUP_NOUNLOCK         0x00001000      /* leave lock intact */
 468
 469 /*
 470  * Flags passed to hammer2_chain_modify() and hammer2_chain_resize()
 471  *
 472  * NOTE: OPTDATA allows us to avoid instantiating buffers for INDIRECT
 473  *       blocks in the INITIAL-create state.
 474  */
 475 #define HAMMER2_MODIFY_OPTDATA          0x00000002      /* data can be NULL */
 476 #define HAMMER2_MODIFY_NO_MODIFY_TID    0x00000004
 477 #define HAMMER2_MODIFY_UNUSED0008       0x00000008
 478
 479 /*
 480  * Flags passed to hammer2_chain_lock()
 481  *
 482  * NOTE: RDONLY is set to optimize cluster operations when *no* modifications
 483  *       will be made to either the cluster being locked or any underlying
 484  *       cluster.  It allows the cluster to lock and access data for a subset
 485  *       of available nodes instead of all available nodes.
 486  */
 487 #define HAMMER2_RESOLVE_NEVER           1
 488 #define HAMMER2_RESOLVE_MAYBE           2
 489 #define HAMMER2_RESOLVE_ALWAYS          3
 490 #define HAMMER2_RESOLVE_MASK            0x0F
 491
 492 #define HAMMER2_RESOLVE_SHARED          0x10    /* request shared lock */
 493 #define HAMMER2_RESOLVE_UNUSED20        0x20
 494 #define HAMMER2_RESOLVE_RDONLY          0x40    /* higher level op flag */
 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  */
 504 #define HAMMER2_INSERT_PFSROOT          0x0004
 505
 506 /*
 507  * Flags passed to hammer2_chain_delete_duplicate()
 508  */
 509 #define HAMMER2_DELDUP_RECORE           0x0001
 510
 511 /*
 512  * Cluster different types of storage together for allocations
 513  */
 514 #define HAMMER2_FREECACHE_INODE         0
 515 #define HAMMER2_FREECACHE_INDIR         1
 516 #define HAMMER2_FREECACHE_DATA          2
 517 #define HAMMER2_FREECACHE_UNUSED3       3
 518 #define HAMMER2_FREECACHE_TYPES         4
 519
 520 /*
 521  * hammer2_freemap_alloc() block preference
 522  */
 523 #define HAMMER2_OFF_NOPREF              ((hammer2_off_t)-1)
 524
 525 /*
 526  * BMAP read-ahead maximum parameters
 527  */
 528 #define HAMMER2_BMAP_COUNT              16      /* max bmap read-ahead */
 529 #define HAMMER2_BMAP_BYTES              (HAMMER2_PBUFSIZE * HAMMER2_BMAP_COUNT)
 530
 531 /*
 532  * hammer2_freemap_adjust()
 533  */
 534 #define HAMMER2_FREEMAP_DORECOVER       1
 535 #define HAMMER2_FREEMAP_DOMAYFREE       2
 536 #define HAMMER2_FREEMAP_DOREALFREE      3
 537
 538 /*
 539  * HAMMER2 cluster - A set of chains representing the same entity.
 540  *
 541  * hammer2_cluster typically represents a temporary set of representitive
 542  * chains.  The one exception is that a hammer2_cluster is embedded in
 543  * hammer2_inode.  This embedded cluster is ONLY used to track the
 544  * representitive chains and cannot be directly locked.
 545  *
 546  * A cluster is usually temporary (and thus per-thread) for locking purposes,
 547  * allowing us to embed the asynchronous storage required for cluster
 548  * operations in the cluster itself and adjust the state and status without
 549  * having to worry too much about SMP issues.
 550  *
 551  * The exception is the cluster embedded in the hammer2_inode structure.
 552  * This is used to cache the cluster state on an inode-by-inode basis.
 553  * Individual hammer2_chain structures not incorporated into clusters might
 554  * also stick around to cache miscellanious elements.
 555  *
 556  * Because the cluster is a 'working copy' and is usually subject to cluster
 557  * quorum rules, it is quite possible for us to end up with an insufficient
 558  * number of live chains to execute an operation.  If an insufficient number
 559  * of chains remain in a working copy, the operation may have to be
 560  * downgraded, retried, stall until the requisit number of chains are
 561  * available, or possibly even error out depending on the mount type.
 562  *
 563  * A cluster's focus is set when it is locked.  The focus can only be set
 564  * to a chain still part of the synchronized set.
 565  */
 566 #define HAMMER2_MAXCLUSTER      8
 567 #define HAMMER2_XOPMASK_CLUSTER ((1U << HAMMER2_MAXCLUSTER) - 1)
 568 #define HAMMER2_XOPFIFO         16
 569 #define HAMMER2_XOPFIFO_MASK    (HAMMER2_XOPFIFO - 1)
 570 #define HAMMER2_XOPGROUPS       16
 571 #define HAMMER2_XOPGROUPS_MASK  (HAMMER2_XOPGROUPS - 1)
 572 #define HAMMER2_XOPMASK_VOP     0x80000000U
 573 #define HAMMER2_XOPMASK_FIFOW   0x40000000U
 574
 575 #define HAMMER2_XOPMASK_ALLDONE (HAMMER2_XOPMASK_VOP | HAMMER2_XOPMASK_CLUSTER)
 576
 577 #define HAMMER2_SPECTHREADS     1       /* sync */
 578
 579 struct hammer2_cluster_item {
 580         hammer2_chain_t         *chain;
 581         int                     cache_index;
 582         int                     error;
 583         uint32_t                flags;
 584 };
 585
 586 typedef struct hammer2_cluster_item hammer2_cluster_item_t;
 587
 588 /*
 589  * INVALID      - Invalid for focus, i.e. not part of synchronized set.
 590  *                Once set, this bit is sticky across operations.
 591  *
 592  * FEMOD        - Indicates that front-end modifying operations can
 593  *                mess with this entry and MODSYNC will copy also
 594  *                effect it.
 595  */
 596 #define HAMMER2_CITEM_INVALID   0x00000001
 597 #define HAMMER2_CITEM_FEMOD     0x00000002
 598 #define HAMMER2_CITEM_NULL      0x00000004
 599
 600 struct hammer2_cluster {
 601         int                     refs;           /* track for deallocation */
 602         int                     ddflag;
 603         struct hammer2_pfs      *pmp;
 604         uint32_t                flags;
 605         int                     nchains;
 606         int                     error;          /* error code valid on lock */
 607         int                     focus_index;
 608         hammer2_iocb_t          iocb;
 609         hammer2_chain_t         *focus;         /* current focus (or mod) */
 610         hammer2_cluster_item_t  array[HAMMER2_MAXCLUSTER];
 611 };
 612
 613 typedef struct hammer2_cluster  hammer2_cluster_t;
 614
 615 /*
 616  * WRHARD       - Hard mounts can write fully synchronized
 617  * RDHARD       - Hard mounts can read fully synchronized
 618  * UNHARD       - Unsynchronized masters present
 619  * NOHARD       - No masters visible
 620  * WRSOFT       - Soft mounts can write to at least the SOFT_MASTER
 621  * RDSOFT       - Soft mounts can read from at least a SOFT_SLAVE
 622  * UNSOFT       - Unsynchronized slaves present
 623  * NOSOFT       - No slaves visible
 624  * RDSLAVE      - slaves are accessible (possibly unsynchronized or remote).
 625  * MSYNCED      - All masters are fully synchronized
 626  * SSYNCED      - All known local slaves are fully synchronized to masters
 627  *
 628  * All available masters are always incorporated.  All PFSs belonging to a
 629  * cluster (master, slave, copy, whatever) always try to synchronize the
 630  * total number of known masters in the PFSs root inode.
 631  *
 632  * A cluster might have access to many slaves, copies, or caches, but we
 633  * have a limited number of cluster slots.  Any such elements which are
 634  * directly mounted from block device(s) will always be incorporated.   Note
 635  * that SSYNCED only applies to such elements which are directly mounted,
 636  * not to any remote slaves, copies, or caches that could be available.  These
 637  * bits are used to monitor and drive our synchronization threads.
 638  *
 639  * When asking the question 'is any data accessible at all', then a simple
 640  * test against (RDHARD|RDSOFT|RDSLAVE) gives you the answer.  If any of
 641  * these bits are set the object can be read with certain caveats:
 642  * RDHARD - no caveats.  RDSOFT - authoritative but might not be synchronized.
 643  * and RDSLAVE - not authoritative, has some data but it could be old or
 644  * incomplete.
 645  *
 646  * When both soft and hard mounts are available, data will be read and written
 647  * via the soft mount only.  But all might be in the cluster because
 648  * background synchronization threads still need to do their work.
 649  */
 650 #define HAMMER2_CLUSTER_INODE   0x00000001      /* embedded in inode struct */
 651 #define HAMMER2_CLUSTER_UNUSED2 0x00000002
 652 #define HAMMER2_CLUSTER_LOCKED  0x00000004      /* cluster lks not recursive */
 653 #define HAMMER2_CLUSTER_WRHARD  0x00000100      /* hard-mount can write */
 654 #define HAMMER2_CLUSTER_RDHARD  0x00000200      /* hard-mount can read */
 655 #define HAMMER2_CLUSTER_UNHARD  0x00000400      /* unsynchronized masters */
 656 #define HAMMER2_CLUSTER_NOHARD  0x00000800      /* no masters visible */
 657 #define HAMMER2_CLUSTER_WRSOFT  0x00001000      /* soft-mount can write */
 658 #define HAMMER2_CLUSTER_RDSOFT  0x00002000      /* soft-mount can read */
 659 #define HAMMER2_CLUSTER_UNSOFT  0x00004000      /* unsynchronized slaves */
 660 #define HAMMER2_CLUSTER_NOSOFT  0x00008000      /* no slaves visible */
 661 #define HAMMER2_CLUSTER_MSYNCED 0x00010000      /* all masters synchronized */
 662 #define HAMMER2_CLUSTER_SSYNCED 0x00020000      /* known slaves synchronized */
 663
 664 #define HAMMER2_CLUSTER_ANYDATA ( HAMMER2_CLUSTER_RDHARD |      \
 665                                   HAMMER2_CLUSTER_RDSOFT |      \
 666                                   HAMMER2_CLUSTER_RDSLAVE)
 667
 668 #define HAMMER2_CLUSTER_RDOK    ( HAMMER2_CLUSTER_RDHARD |      \
 669                                   HAMMER2_CLUSTER_RDSOFT)
 670
 671 #define HAMMER2_CLUSTER_WROK    ( HAMMER2_CLUSTER_WRHARD |      \
 672                                   HAMMER2_CLUSTER_WRSOFT)
 673
 674 #define HAMMER2_CLUSTER_ZFLAGS  ( HAMMER2_CLUSTER_WRHARD |      \
 675                                   HAMMER2_CLUSTER_RDHARD |      \
 676                                   HAMMER2_CLUSTER_WRSOFT |      \
 677                                   HAMMER2_CLUSTER_RDSOFT |      \
 678                                   HAMMER2_CLUSTER_MSYNCED |     \
 679                                   HAMMER2_CLUSTER_SSYNCED)
 680
 681 /*
 682  * Helper functions (cluster must be locked for flags to be valid).
 683  */
 684 static __inline
 685 int
 686 hammer2_cluster_rdok(hammer2_cluster_t *cluster)
 687 {
 688         return (cluster->flags & HAMMER2_CLUSTER_RDOK);
 689 }
 690
 691 static __inline
 692 int
 693 hammer2_cluster_wrok(hammer2_cluster_t *cluster)
 694 {
 695         return (cluster->flags & HAMMER2_CLUSTER_WROK);
 696 }
 697
 698 RB_HEAD(hammer2_inode_tree, hammer2_inode);
 699
 700 /*
 701  * A hammer2 inode.
 702  *
 703  * NOTE: The inode-embedded cluster is never used directly for I/O (since
 704  *       it may be shared).  Instead it will be replicated-in and synchronized
 705  *       back out if changed.
 706  */
 707 struct hammer2_inode {
 708         RB_ENTRY(hammer2_inode) rbnode;         /* inumber lookup (HL) */
 709         hammer2_mtx_t           lock;           /* inode lock */
 710         hammer2_mtx_t           truncate_lock;  /* prevent truncates */
 711         struct hammer2_pfs      *pmp;           /* PFS mount */
 712         struct hammer2_inode    *pip;           /* parent inode */
 713         struct vnode            *vp;
 714         struct spinlock         cluster_spin;   /* update cluster */
 715         hammer2_cluster_t       cluster;
 716         struct lockf            advlock;
 717         u_int                   flags;
 718         u_int                   refs;           /* +vpref, +flushref */
 719         uint8_t                 comp_heuristic;
 720         hammer2_inode_meta_t    meta;           /* copy of meta-data */
 721         hammer2_off_t           osize;
 722         hammer2_cluster_t       *cluster_cache;
 723 };
 724
 725 typedef struct hammer2_inode hammer2_inode_t;
 726
 727 /*
 728  * MODIFIED     - Inode is in a modified state, ip->meta may have changes.
 729  * RESIZED      - Inode truncated (any) or inode extended beyond
 730  *                EMBEDDED_BYTES.
 731  */
 732 #define HAMMER2_INODE_MODIFIED          0x0001
 733 #define HAMMER2_INODE_SROOT             0x0002  /* kmalloc special case */
 734 #define HAMMER2_INODE_RENAME_INPROG     0x0004
 735 #define HAMMER2_INODE_ONRBTREE          0x0008
 736 #define HAMMER2_INODE_RESIZED           0x0010  /* requires inode_fsync */
 737 #define HAMMER2_INODE_ISDELETED         0x0020  /* deleted, not in ihidden */
 738 #define HAMMER2_INODE_ISUNLINKED        0x0040
 739 #define HAMMER2_INODE_METAGOOD          0x0080  /* inode meta-data good */
 740 #define HAMMER2_INODE_ONSIDEQ           0x0100  /* on side processing queue */
 741
 742 int hammer2_inode_cmp(hammer2_inode_t *ip1, hammer2_inode_t *ip2);
 743 RB_PROTOTYPE2(hammer2_inode_tree, hammer2_inode, rbnode, hammer2_inode_cmp,
 744                 hammer2_tid_t);
 745
 746 /*
 747  * inode-unlink side-structure
 748  */
 749 struct hammer2_inode_sideq {
 750         TAILQ_ENTRY(hammer2_inode_sideq) entry;
 751         hammer2_inode_t *ip;
 752 };
 753 TAILQ_HEAD(h2_sideq_list, hammer2_inode_sideq);
 754
 755 typedef struct hammer2_inode_sideq hammer2_inode_sideq_t;
 756
 757 /*
 758  * Transaction management sub-structure under hammer2_pfs
 759  */
 760 struct hammer2_trans {
 761         uint32_t                flags;
 762         uint32_t                sync_wait;
 763 };
 764
 765 typedef struct hammer2_trans hammer2_trans_t;
 766
 767 #define HAMMER2_TRANS_ISFLUSH           0x80000000      /* flush code */
 768 #define HAMMER2_TRANS_BUFCACHE          0x40000000      /* bio strategy */
 769 #define HAMMER2_TRANS_PREFLUSH          0x20000000      /* preflush state */
 770 #define HAMMER2_TRANS_FPENDING          0x10000000      /* flush pending */
 771 #define HAMMER2_TRANS_WAITING           0x08000000      /* someone waiting */
 772 #define HAMMER2_TRANS_MASK              0x00FFFFFF      /* count mask */
 773
 774 #define HAMMER2_FREEMAP_HEUR_NRADIX     4       /* pwr 2 PBUFRADIX-MINIORADIX */
 775 #define HAMMER2_FREEMAP_HEUR_TYPES      8
 776 #define HAMMER2_FREEMAP_HEUR_SIZE       (HAMMER2_FREEMAP_HEUR_NRADIX * \
 777                                          HAMMER2_FREEMAP_HEUR_TYPES)
 778
 779 #define HAMMER2_DEDUP_HEUR_SIZE         65536
 780 #define HAMMER2_DEDUP_HEUR_MASK         (HAMMER2_DEDUP_HEUR_SIZE - 1)
 781
 782 #define HAMMER2_FLUSH_TOP               0x0001
 783 #define HAMMER2_FLUSH_ALL               0x0002
 784
 785
 786 /*
 787  * Hammer2 support thread element.
 788  *
 789  * Potentially many support threads can hang off of hammer2, primarily
 790  * off the hammer2_pfs structure.  Typically:
 791  *
 792  * td x Nodes                   A synchronization thread for each node.
 793  * td x Nodes x workers         Worker threads for frontend operations.
 794  * td x 1                       Bioq thread for logical buffer writes.
 795  *
 796  * In addition, the synchronization thread(s) associated with the
 797  * super-root PFS (spmp) for a node is responsible for automatic bulkfree
 798  * and dedup scans.
 799  */
 800 struct hammer2_thread {
 801         struct hammer2_pfs *pmp;
 802         hammer2_xop_list_t *xopq;       /* points into pmp->xopq[] */
 803         thread_t        td;
 804         uint32_t        flags;
 805         int             depth;
 806         int             clindex;        /* cluster element index */
 807         int             repidx;
 808 };
 809
 810 typedef struct hammer2_thread hammer2_thread_t;
 811
 812 #define HAMMER2_THREAD_UNMOUNTING       0x0001  /* unmount request */
 813 #define HAMMER2_THREAD_DEV              0x0002  /* related to dev, not pfs */
 814 #define HAMMER2_THREAD_WAITING          0x0004  /* thread in idle tsleep */
 815 #define HAMMER2_THREAD_REMASTER         0x0008  /* remaster request */
 816 #define HAMMER2_THREAD_STOP             0x0010  /* exit request */
 817 #define HAMMER2_THREAD_FREEZE           0x0020  /* force idle */
 818 #define HAMMER2_THREAD_FROZEN           0x0040  /* thread is frozen */
 819 #define HAMMER2_THREAD_XOPQ             0x0080  /* work pending */
 820 #define HAMMER2_THREAD_STOPPED          0x0100  /* thread has stopped */
 821 #define HAMMER2_THREAD_UNFREEZE         0x0200
 822 #define HAMMER2_THREAD_CLIENTWAIT       0x0400
 823
 824 #define HAMMER2_THREAD_WAKEUP_MASK      (HAMMER2_THREAD_UNMOUNTING |    \
 825                                          HAMMER2_THREAD_REMASTER |      \
 826                                          HAMMER2_THREAD_STOP |          \
 827                                          HAMMER2_THREAD_FREEZE |        \
 828                                          HAMMER2_THREAD_XOPQ)
 829
 830 /*
 831  * Support structure for dedup heuristic.
 832  */
 833 struct hammer2_dedup {
 834         hammer2_off_t   data_off;
 835         uint64_t        data_crc;
 836         uint32_t        ticks;
 837         uint32_t        unused03;
 838 };
 839
 840 typedef struct hammer2_dedup hammer2_dedup_t;
 841
 842 /*
 843  * hammer2_xop - container for VOP/XOP operation (allocated, not on stack).
 844  *
 845  * This structure is used to distribute a VOP operation across multiple
 846  * nodes.  It provides a rendezvous for concurrent node execution and
 847  * can be detached from the frontend operation to allow the frontend to
 848  * return early.
 849  *
 850  * This structure also sequences operations on up to three inodes.
 851  */
 852 typedef void (*hammer2_xop_func_t)(union hammer2_xop *xop, int clidx);
 853
 854 struct hammer2_xop_fifo {
 855         TAILQ_ENTRY(hammer2_xop_head) entry;
 856         hammer2_chain_t         *array[HAMMER2_XOPFIFO];
 857         int                     errors[HAMMER2_XOPFIFO];
 858         int                     ri;
 859         int                     wi;
 860         int                     flags;
 861 };
 862
 863 typedef struct hammer2_xop_fifo hammer2_xop_fifo_t;
 864
 865 #define HAMMER2_XOP_FIFO_RUN    0x0001
 866 #define HAMMER2_XOP_FIFO_STALL  0x0002
 867
 868 struct hammer2_xop_head {
 869         hammer2_xop_func_t      func;
 870         hammer2_tid_t           mtid;
 871         struct hammer2_inode    *ip1;
 872         struct hammer2_inode    *ip2;
 873         struct hammer2_inode    *ip3;
 874         uint32_t                check_counter;
 875         uint32_t                run_mask;
 876         uint32_t                chk_mask;
 877         int                     flags;
 878         int                     state;
 879         int                     error;
 880         hammer2_key_t           collect_key;
 881         char                    *name1;
 882         size_t                  name1_len;
 883         char                    *name2;
 884         size_t                  name2_len;
 885         hammer2_xop_fifo_t      collect[HAMMER2_MAXCLUSTER];
 886         hammer2_cluster_t       cluster;        /* help collections */
 887 };
 888
 889 typedef struct hammer2_xop_head hammer2_xop_head_t;
 890
 891 struct hammer2_xop_ipcluster {
 892         hammer2_xop_head_t      head;
 893 };
 894
 895 struct hammer2_xop_strategy {
 896         hammer2_xop_head_t      head;
 897         hammer2_key_t           lbase;
 898         int                     finished;
 899         hammer2_mtx_t           lock;
 900         struct bio              *bio;
 901 };
 902
 903 struct hammer2_xop_readdir {
 904         hammer2_xop_head_t      head;
 905         hammer2_key_t           lkey;
 906 };
 907
 908 struct hammer2_xop_nresolve {
 909         hammer2_xop_head_t      head;
 910         hammer2_key_t           lhc;    /* if name is NULL used lhc */
 911 };
 912
 913 struct hammer2_xop_nlink {
 914         hammer2_xop_head_t      head;
 915         int                     nlinks_delta;
 916 };
 917
 918 struct hammer2_xop_unlink {
 919         hammer2_xop_head_t      head;
 920         int                     isdir;
 921         int                     dopermanent;
 922 };
 923
 924 struct hammer2_xop_nrename {
 925         hammer2_xop_head_t      head;
 926         hammer2_tid_t           lhc;
 927         int                     ip_key;
 928 };
 929
 930 struct hammer2_xop_scanlhc {
 931         hammer2_xop_head_t      head;
 932         hammer2_key_t           lhc;
 933 };
 934
 935 struct hammer2_xop_scanall {
 936         hammer2_xop_head_t      head;
 937         hammer2_key_t           key_beg;        /* inclusive */
 938         hammer2_key_t           key_end;        /* inclusive */
 939         int                     resolve_flags;
 940         int                     lookup_flags;
 941 };
 942
 943 struct hammer2_xop_lookup {
 944         hammer2_xop_head_t      head;
 945         hammer2_key_t           lhc;
 946 };
 947
 948 struct hammer2_xop_create {
 949         hammer2_xop_head_t      head;
 950         hammer2_inode_meta_t    meta;           /* initial metadata */
 951         hammer2_key_t           lhc;
 952         int                     flags;
 953 };
 954
 955 struct hammer2_xop_destroy {
 956         hammer2_xop_head_t      head;
 957 };
 958
 959 struct hammer2_xop_fsync {
 960         hammer2_xop_head_t      head;
 961         hammer2_inode_meta_t    meta;
 962         hammer2_off_t           osize;
 963         u_int                   ipflags;
 964         int                     clear_directdata;
 965 };
 966
 967 struct hammer2_xop_unlinkall {
 968         hammer2_xop_head_t      head;
 969         hammer2_key_t           key_beg;
 970         hammer2_key_t           key_end;
 971 };
 972
 973 struct hammer2_xop_connect {
 974         hammer2_xop_head_t      head;
 975         hammer2_key_t           lhc;
 976 };
 977
 978 struct hammer2_xop_flush {
 979         hammer2_xop_head_t      head;
 980 };
 981
 982 typedef struct hammer2_xop_readdir hammer2_xop_readdir_t;
 983 typedef struct hammer2_xop_nresolve hammer2_xop_nresolve_t;
 984 typedef struct hammer2_xop_nlink hammer2_xop_nlink_t;
 985 typedef struct hammer2_xop_unlink hammer2_xop_unlink_t;
 986 typedef struct hammer2_xop_nrename hammer2_xop_nrename_t;
 987 typedef struct hammer2_xop_ipcluster hammer2_xop_ipcluster_t;
 988 typedef struct hammer2_xop_strategy hammer2_xop_strategy_t;
 989 typedef struct hammer2_xop_create hammer2_xop_create_t;
 990 typedef struct hammer2_xop_destroy hammer2_xop_destroy_t;
 991 typedef struct hammer2_xop_fsync hammer2_xop_fsync_t;
 992 typedef struct hammer2_xop_unlinkall hammer2_xop_unlinkall_t;
 993 typedef struct hammer2_xop_scanlhc hammer2_xop_scanlhc_t;
 994 typedef struct hammer2_xop_scanall hammer2_xop_scanall_t;
 995 typedef struct hammer2_xop_lookup hammer2_xop_lookup_t;
 996 typedef struct hammer2_xop_connect hammer2_xop_connect_t;
 997 typedef struct hammer2_xop_flush hammer2_xop_flush_t;
 998
 999 union hammer2_xop {
1000         hammer2_xop_head_t      head;
1001         hammer2_xop_ipcluster_t xop_ipcluster;
1002         hammer2_xop_readdir_t   xop_readdir;
1003         hammer2_xop_nresolve_t  xop_nresolve;
1004         hammer2_xop_nlink_t     xop_nlink;
1005         hammer2_xop_unlink_t    xop_unlink;
1006         hammer2_xop_nrename_t   xop_nrename;
1007         hammer2_xop_strategy_t  xop_strategy;
1008         hammer2_xop_create_t    xop_create;
1009         hammer2_xop_destroy_t   xop_destroy;
1010         hammer2_xop_fsync_t     xop_fsync;
1011         hammer2_xop_unlinkall_t xop_unlinkall;
1012         hammer2_xop_scanlhc_t   xop_scanlhc;
1013         hammer2_xop_scanall_t   xop_scanall;
1014         hammer2_xop_lookup_t    xop_lookup;
1015         hammer2_xop_flush_t     xop_flush;
1016         hammer2_xop_connect_t   xop_connect;
1017 };
1018
1019 typedef union hammer2_xop hammer2_xop_t;
1020
1021 /*
1022  * hammer2_xop_group - Manage XOP support threads.
1023  */
1024 struct hammer2_xop_group {
1025         hammer2_thread_t        thrs[HAMMER2_MAXCLUSTER];
1026 };
1027
1028 typedef struct hammer2_xop_group hammer2_xop_group_t;
1029
1030 /*
1031  * flags to hammer2_xop_collect()
1032  */
1033 #define HAMMER2_XOP_COLLECT_NOWAIT      0x00000001
1034 #define HAMMER2_XOP_COLLECT_WAITALL     0x00000002
1035
1036 /*
1037  * flags to hammer2_xop_alloc()
1038  *
1039  * MODIFYING    - This is a modifying transaction, allocate a mtid.
1040  */
1041 #define HAMMER2_XOP_MODIFYING           0x00000001
1042 #define HAMMER2_XOP_ITERATOR            0x00000002
1043
1044 /*
1045  * Global (per partition) management structure, represents a hard block
1046  * device.  Typically referenced by hammer2_chain structures when applicable.
1047  * Typically not used for network-managed elements.
1048  *
1049  * Note that a single hammer2_dev can be indirectly tied to multiple system
1050  * mount points.  There is no direct relationship.  System mounts are
1051  * per-cluster-id, not per-block-device, and a single hard mount might contain
1052  * many PFSs and those PFSs might combine together in various ways to form
1053  * the set of available clusters.
1054  */
1055 struct hammer2_dev {
1056         struct vnode    *devvp;         /* device vnode */
1057         int             ronly;          /* read-only mount */
1058         int             mount_count;    /* number of actively mounted PFSs */
1059         TAILQ_ENTRY(hammer2_dev) mntentry; /* hammer2_mntlist */
1060
1061         struct malloc_type *mchain;
1062         int             nipstacks;
1063         int             maxipstacks;
1064         kdmsg_iocom_t   iocom;          /* volume-level dmsg interface */
1065         struct spinlock io_spin;        /* iotree access */
1066         struct hammer2_io_tree iotree;
1067         int             iofree_count;
1068         hammer2_chain_t vchain;         /* anchor chain (topology) */
1069         hammer2_chain_t fchain;         /* anchor chain (freemap) */
1070         struct spinlock list_spin;
1071         struct h2_flush_list    flushq; /* flush seeds */
1072         struct hammer2_pfs *spmp;       /* super-root pmp for transactions */
1073         struct lock     bulklk;         /* bulkfree lock */
1074         struct lock     vollk;          /* lockmgr lock */
1075         hammer2_off_t   heur_freemap[HAMMER2_FREEMAP_HEUR_SIZE];
1076         hammer2_dedup_t heur_dedup[HAMMER2_DEDUP_HEUR_SIZE];
1077         int             volhdrno;       /* last volhdrno written */
1078         int             hflags;         /* HMNT2 flags applicable to device */
1079         char            devrepname[64]; /* for kprintf */
1080         hammer2_volume_data_t voldata;
1081         hammer2_volume_data_t volsync;  /* synchronized voldata */
1082 };
1083
1084 typedef struct hammer2_dev hammer2_dev_t;
1085
1086 /*
1087  * Helper functions (cluster must be locked for flags to be valid).
1088  */
1089 static __inline
1090 int
1091 hammer2_chain_rdok(hammer2_chain_t *chain)
1092 {
1093         return (chain->error == 0);
1094 }
1095
1096 static __inline
1097 int
1098 hammer2_chain_wrok(hammer2_chain_t *chain)
1099 {
1100         return (chain->error == 0 && chain->hmp->ronly == 0);
1101 }
1102
1103 /*
1104  * Per-cluster management structure.  This structure will be tied to a
1105  * system mount point if the system is mounting the PFS, but is also used
1106  * to manage clusters encountered during the super-root scan or received
1107  * via LNK_SPANs that might not be mounted.
1108  *
1109  * This structure is also used to represent the super-root that hangs off
1110  * of a hard mount point.  The super-root is not really a cluster element.
1111  * In this case the spmp_hmp field will be non-NULL.  It's just easier to do
1112  * this than to special case super-root manipulation in the hammer2_chain*
1113  * code as being only hammer2_dev-related.
1114  *
1115  * pfs_mode and pfs_nmasters are rollup fields which critically describes
1116  * how elements of the cluster act on the cluster.  pfs_mode is only applicable
1117  * when a PFS is mounted by the system.  pfs_nmasters is our best guess as to
1118  * how many masters have been configured for a cluster and is always
1119  * applicable.  pfs_types[] is an array with 1:1 correspondance to the
1120  * iroot cluster and describes the PFS types of the nodes making up the
1121  * cluster.
1122  *
1123  * WARNING! Portions of this structure have deferred initialization.  In
1124  *          particular, if not mounted there will be no ihidden or wthread.
1125  *          umounted network PFSs will also be missing iroot and numerous
1126  *          other fields will not be initialized prior to mount.
1127  *
1128  *          Synchronization threads are chain-specific and only applicable
1129  *          to local hard PFS entries.  A hammer2_pfs structure may contain
1130  *          more than one when multiple hard PFSs are present on the local
1131  *          machine which require synchronization monitoring.  Most PFSs
1132  *          (such as snapshots) are 1xMASTER PFSs which do not need a
1133  *          synchronization thread.
1134  *
1135  * WARNING! The chains making up pfs->iroot's cluster are accounted for in
1136  *          hammer2_dev->mount_count when the pfs is associated with a mount
1137  *          point.
1138  */
1139 struct hammer2_pfs {
1140         struct mount            *mp;
1141         TAILQ_ENTRY(hammer2_pfs) mntentry;      /* hammer2_pfslist */
1142         uuid_t                  pfs_clid;
1143         hammer2_dev_t           *spmp_hmp;      /* only if super-root pmp */
1144         hammer2_dev_t           *force_local;   /* only if 'local' mount */
1145         hammer2_inode_t         *iroot;         /* PFS root inode */
1146         hammer2_inode_t         *ihidden;       /* PFS hidden directory */
1147         uint8_t                 pfs_types[HAMMER2_MAXCLUSTER];
1148         char                    *pfs_names[HAMMER2_MAXCLUSTER];
1149         hammer2_dev_t           *pfs_hmps[HAMMER2_MAXCLUSTER];
1150         hammer2_trans_t         trans;
1151         struct lock             lock;           /* PFS lock for certain ops */
1152         struct lock             lock_nlink;     /* rename and nlink lock */
1153         struct netexport        export;         /* nfs export */
1154         int                     ronly;          /* read-only mount */
1155         int                     hflags;         /* pfs-specific mount flags */
1156         struct malloc_type      *minode;
1157         struct malloc_type      *mmsg;
1158         struct spinlock         inum_spin;      /* inumber lookup */
1159         struct hammer2_inode_tree inum_tree;    /* (not applicable to spmp) */
1160         struct spinlock         lru_spin;       /* inumber lookup */
1161         struct hammer2_chain_list lru_list;     /* chains on LRU */
1162         int                     lru_count;      /* #of chains on LRU */
1163         hammer2_tid_t           modify_tid;     /* modify transaction id */
1164         hammer2_tid_t           inode_tid;      /* inode allocator */
1165         uint8_t                 pfs_nmasters;   /* total masters */
1166         uint8_t                 pfs_mode;       /* operating mode PFSMODE */
1167         uint8_t                 unused01;
1168         uint8_t                 unused02;
1169         int                     xop_iterator;
1170         long                    inmem_inodes;
1171         uint32_t                inmem_dirty_chains;
1172         int                     count_lwinprog; /* logical write in prog */
1173         struct spinlock         list_spin;
1174         struct h2_sideq_list    sideq;          /* last-close dirty/unlink */
1175         hammer2_thread_t        sync_thrs[HAMMER2_MAXCLUSTER];
1176         uint32_t                cluster_flags;  /* cached cluster flags */
1177         int                     has_xop_threads;
1178         struct spinlock         xop_spin;       /* xop sequencer */
1179         hammer2_xop_group_t     xop_groups[HAMMER2_XOPGROUPS];
1180         hammer2_xop_list_t      xopq[HAMMER2_MAXCLUSTER][HAMMER2_XOPGROUPS+
1181                                                          HAMMER2_SPECTHREADS];
1182 };
1183
1184 typedef struct hammer2_pfs hammer2_pfs_t;
1185
1186 #define HAMMER2_LRU_LIMIT               1024    /* per pmp lru_list */
1187
1188 #define HAMMER2_DIRTYCHAIN_WAITING      0x80000000
1189 #define HAMMER2_DIRTYCHAIN_MASK         0x7FFFFFFF
1190
1191 #define HAMMER2_LWINPROG_WAITING        0x80000000
1192 #define HAMMER2_LWINPROG_WAITING0       0x40000000
1193 #define HAMMER2_LWINPROG_MASK           0x3FFFFFFF
1194
1195 /*
1196  * hammer2_cluster_check
1197  */
1198 #define HAMMER2_CHECK_NULL      0x00000001
1199
1200 /*
1201  * Bulkscan
1202  */
1203 #define HAMMER2_BULK_ABORT      0x00000001
1204
1205 /*
1206  * Misc
1207  */
1208 #if defined(_KERNEL)
1209
1210 MALLOC_DECLARE(M_HAMMER2);
1211
1212 #define VTOI(vp)        ((hammer2_inode_t *)(vp)->v_data)
1213 #define ITOV(ip)        ((ip)->vp)
1214
1215 /*
1216  * Currently locked chains retain the locked buffer cache buffer for
1217  * indirect blocks, and indirect blocks can be one of two sizes.  The
1218  * device buffer has to match the case to avoid deadlocking recursive
1219  * chains that might otherwise try to access different offsets within
1220  * the same device buffer.
1221  */
1222 static __inline
1223 int
1224 hammer2_devblkradix(int radix)
1225 {
1226 #if 0
1227         if (radix <= HAMMER2_LBUFRADIX) {
1228                 return (HAMMER2_LBUFRADIX);
1229         } else {
1230                 return (HAMMER2_PBUFRADIX);
1231         }
1232 #endif
1233         return (HAMMER2_PBUFRADIX);
1234 }
1235
1236 /*
1237  * XXX almost time to remove this.  DIO uses PBUFSIZE exclusively now.
1238  */
1239 static __inline
1240 size_t
1241 hammer2_devblksize(size_t bytes)
1242 {
1243 #if 0
1244         if (bytes <= HAMMER2_LBUFSIZE) {
1245                 return(HAMMER2_LBUFSIZE);
1246         } else {
1247                 KKASSERT(bytes <= HAMMER2_PBUFSIZE &&
1248                          (bytes ^ (bytes - 1)) == ((bytes << 1) - 1));
1249                 return (HAMMER2_PBUFSIZE);
1250         }
1251 #endif
1252         return (HAMMER2_PBUFSIZE);
1253 }
1254
1255
1256 static __inline
1257 hammer2_pfs_t *
1258 MPTOPMP(struct mount *mp)
1259 {
1260         return ((hammer2_pfs_t *)mp->mnt_data);
1261 }
1262
1263 #define LOCKSTART       int __nlocks = curthread->td_locks
1264 #define LOCKENTER       (++curthread->td_locks)
1265 #define LOCKEXIT        (--curthread->td_locks)
1266 #define LOCKSTOP        KKASSERT(curthread->td_locks == __nlocks)
1267
1268 extern struct vop_ops hammer2_vnode_vops;
1269 extern struct vop_ops hammer2_spec_vops;
1270 extern struct vop_ops hammer2_fifo_vops;
1271
1272 extern int hammer2_debug;
1273 extern int hammer2_cluster_enable;
1274 extern int hammer2_hardlink_enable;
1275 extern int hammer2_flush_pipe;
1276 extern int hammer2_synchronous_flush;
1277 extern int hammer2_dio_count;
1278 extern long hammer2_chain_allocs;
1279 extern long hammer2_chain_frees;
1280 extern long hammer2_limit_dirty_chains;
1281 extern long hammer2_count_modified_chains;
1282 extern long hammer2_iod_file_read;
1283 extern long hammer2_iod_meta_read;
1284 extern long hammer2_iod_indr_read;
1285 extern long hammer2_iod_fmap_read;
1286 extern long hammer2_iod_volu_read;
1287 extern long hammer2_iod_file_write;
1288 extern long hammer2_iod_file_wembed;
1289 extern long hammer2_iod_file_wzero;
1290 extern long hammer2_iod_file_wdedup;
1291 extern long hammer2_iod_meta_write;
1292 extern long hammer2_iod_indr_write;
1293 extern long hammer2_iod_fmap_write;
1294 extern long hammer2_iod_volu_write;
1295 extern long hammer2_ioa_file_read;
1296 extern long hammer2_ioa_meta_read;
1297 extern long hammer2_ioa_indr_read;
1298 extern long hammer2_ioa_fmap_read;
1299 extern long hammer2_ioa_volu_read;
1300 extern long hammer2_ioa_file_write;
1301 extern long hammer2_ioa_meta_write;
1302 extern long hammer2_ioa_indr_write;
1303 extern long hammer2_ioa_fmap_write;
1304 extern long hammer2_ioa_volu_write;
1305
1306 extern long hammer2_check_xxhash64;
1307 extern long hammer2_check_icrc32;
1308
1309 extern struct objcache *cache_buffer_read;
1310 extern struct objcache *cache_buffer_write;
1311 extern struct objcache *cache_xops;
1312
1313 /*
1314  * hammer2_subr.c
1315  */
1316 #define hammer2_icrc32(buf, size)       iscsi_crc32((buf), (size))
1317 #define hammer2_icrc32c(buf, size, crc) iscsi_crc32_ext((buf), (size), (crc))
1318
1319 int hammer2_signal_check(time_t *timep);
1320 const char *hammer2_error_str(int error);
1321
1322 void hammer2_inode_lock(hammer2_inode_t *ip, int how);
1323 void hammer2_inode_unlock(hammer2_inode_t *ip);
1324 hammer2_chain_t *hammer2_inode_chain(hammer2_inode_t *ip, int clindex, int how);
1325 hammer2_chain_t *hammer2_inode_chain_and_parent(hammer2_inode_t *ip,
1326                         int clindex, hammer2_chain_t **parentp, int how);
1327 hammer2_mtx_state_t hammer2_inode_lock_temp_release(hammer2_inode_t *ip);
1328 void hammer2_inode_lock_temp_restore(hammer2_inode_t *ip,
1329                         hammer2_mtx_state_t ostate);
1330 int hammer2_inode_lock_upgrade(hammer2_inode_t *ip);
1331 void hammer2_inode_lock_downgrade(hammer2_inode_t *ip, int);
1332
1333 void hammer2_dev_exlock(hammer2_dev_t *hmp);
1334 void hammer2_dev_shlock(hammer2_dev_t *hmp);
1335 void hammer2_dev_unlock(hammer2_dev_t *hmp);
1336
1337 int hammer2_get_dtype(const hammer2_inode_data_t *ipdata);
1338 int hammer2_get_vtype(uint8_t type);
1339 uint8_t hammer2_get_obj_type(enum vtype vtype);
1340 void hammer2_time_to_timespec(uint64_t xtime, struct timespec *ts);
1341 uint64_t hammer2_timespec_to_time(const struct timespec *ts);
1342 uint32_t hammer2_to_unix_xid(const uuid_t *uuid);
1343 void hammer2_guid_to_uuid(uuid_t *uuid, uint32_t guid);
1344 void hammer2_trans_manage_init(hammer2_pfs_t *pmp);
1345
1346 hammer2_key_t hammer2_dirhash(const unsigned char *name, size_t len);
1347 int hammer2_getradix(size_t bytes);
1348
1349 int hammer2_calc_logical(hammer2_inode_t *ip, hammer2_off_t uoff,
1350                         hammer2_key_t *lbasep, hammer2_key_t *leofp);
1351 int hammer2_calc_physical(hammer2_inode_t *ip, hammer2_key_t lbase);
1352 void hammer2_update_time(uint64_t *timep);
1353 void hammer2_adjreadcounter(hammer2_blockref_t *bref, size_t bytes);
1354
1355 /*
1356  * hammer2_inode.c
1357  */
1358 struct vnode *hammer2_igetv(hammer2_inode_t *ip, int *errorp);
1359 hammer2_inode_t *hammer2_inode_lookup(hammer2_pfs_t *pmp,
1360                         hammer2_tid_t inum);
1361 hammer2_inode_t *hammer2_inode_get(hammer2_pfs_t *pmp, hammer2_inode_t *dip,
1362                         hammer2_cluster_t *cluster, int idx);
1363 void hammer2_inode_free(hammer2_inode_t *ip);
1364 void hammer2_inode_ref(hammer2_inode_t *ip);
1365 void hammer2_inode_drop(hammer2_inode_t *ip);
1366 void hammer2_inode_repoint(hammer2_inode_t *ip, hammer2_inode_t *pip,
1367                         hammer2_cluster_t *cluster);
1368 void hammer2_inode_repoint_one(hammer2_inode_t *ip, hammer2_cluster_t *cluster,
1369                         int idx);
1370 void hammer2_inode_modify(hammer2_inode_t *ip);
1371 void hammer2_inode_run_sideq(hammer2_pfs_t *pmp);
1372
1373 hammer2_inode_t *hammer2_inode_create(hammer2_inode_t *dip,
1374                         struct vattr *vap, struct ucred *cred,
1375                         const uint8_t *name, size_t name_len, hammer2_key_t lhc,
1376                         hammer2_key_t inum, uint8_t type, uint8_t target_type,
1377                         int flags, int *errorp);
1378 int hammer2_inode_connect(hammer2_inode_t *dip, hammer2_inode_t *ip,
1379                         const char *name, size_t name_len,
1380                         hammer2_key_t lhc);
1381 hammer2_inode_t *hammer2_inode_common_parent(hammer2_inode_t *fdip,
1382                         hammer2_inode_t *tdip, int *errorp, int ishardlink);
1383 void hammer2_inode_chain_sync(hammer2_inode_t *ip);
1384 int hammer2_inode_unlink_finisher(hammer2_inode_t *ip, int isopen);
1385 void hammer2_inode_install_hidden(hammer2_pfs_t *pmp);
1386
1387 /*
1388  * hammer2_chain.c
1389  */
1390 void hammer2_voldata_lock(hammer2_dev_t *hmp);
1391 void hammer2_voldata_unlock(hammer2_dev_t *hmp);
1392 void hammer2_voldata_modify(hammer2_dev_t *hmp);
1393 hammer2_chain_t *hammer2_chain_alloc(hammer2_dev_t *hmp,
1394                                 hammer2_pfs_t *pmp,
1395                                 hammer2_blockref_t *bref);
1396 void hammer2_chain_core_init(hammer2_chain_t *chain);
1397 void hammer2_chain_ref(hammer2_chain_t *chain);
1398 void hammer2_chain_drop(hammer2_chain_t *chain);
1399 void hammer2_chain_lock(hammer2_chain_t *chain, int how);
1400 void hammer2_chain_lock_downgrade(hammer2_chain_t *chain);
1401 void hammer2_chain_push_shared_lock(hammer2_chain_t *chain);
1402 void hammer2_chain_pull_shared_lock(hammer2_chain_t *chain);
1403 void hammer2_chain_load_data(hammer2_chain_t *chain);
1404 const hammer2_media_data_t *hammer2_chain_rdata(hammer2_chain_t *chain);
1405 hammer2_media_data_t *hammer2_chain_wdata(hammer2_chain_t *chain);
1406 int hammer2_chain_snapshot(hammer2_chain_t *chain, hammer2_ioc_pfs_t *pmp,
1407                                 hammer2_tid_t mtid);
1408
1409 int hammer2_chain_hardlink_find(hammer2_inode_t *dip,
1410                                 hammer2_chain_t **parentp,
1411                                 hammer2_chain_t **chainp,
1412                                 int flags);
1413 void hammer2_chain_modify(hammer2_chain_t *chain, hammer2_tid_t mtid,
1414                                 hammer2_off_t dedup_off, int flags);
1415 void hammer2_chain_modify_ip(hammer2_inode_t *ip, hammer2_chain_t *chain,
1416                                 hammer2_tid_t mtid, int flags);
1417 void hammer2_chain_resize(hammer2_inode_t *ip, hammer2_chain_t *parent,
1418                                 hammer2_chain_t *chain,
1419                                 hammer2_tid_t mtid, hammer2_off_t dedup_off,
1420                                 int nradix, int flags);
1421 void hammer2_chain_unlock(hammer2_chain_t *chain);
1422 void hammer2_chain_wait(hammer2_chain_t *chain);
1423 hammer2_chain_t *hammer2_chain_get(hammer2_chain_t *parent, int generation,
1424                                 hammer2_blockref_t *bref);
1425 hammer2_chain_t *hammer2_chain_lookup_init(hammer2_chain_t *parent, int flags);
1426 void hammer2_chain_lookup_done(hammer2_chain_t *parent);
1427 hammer2_chain_t *hammer2_chain_getparent(hammer2_chain_t **parentp, int how);
1428 hammer2_chain_t *hammer2_chain_lookup(hammer2_chain_t **parentp,
1429                                 hammer2_key_t *key_nextp,
1430                                 hammer2_key_t key_beg, hammer2_key_t key_end,
1431                                 int *cache_indexp, int flags);
1432 hammer2_chain_t *hammer2_chain_next(hammer2_chain_t **parentp,
1433                                 hammer2_chain_t *chain,
1434                                 hammer2_key_t *key_nextp,
1435                                 hammer2_key_t key_beg, hammer2_key_t key_end,
1436                                 int *cache_indexp, int flags);
1437 hammer2_blockref_t *hammer2_chain_scan(hammer2_chain_t *parent,
1438                                 hammer2_chain_t **chainp,
1439                                 hammer2_blockref_t *bref,
1440                                 int *firstp, int *cache_indexp, int flags);
1441
1442 int hammer2_chain_create(hammer2_chain_t **parentp,
1443                                 hammer2_chain_t **chainp, hammer2_pfs_t *pmp,
1444                                 hammer2_key_t key, int keybits,
1445                                 int type, size_t bytes, hammer2_tid_t mtid,
1446                                 hammer2_off_t dedup_off, int flags);
1447 void hammer2_chain_rename(hammer2_blockref_t *bref,
1448                                 hammer2_chain_t **parentp,
1449                                 hammer2_chain_t *chain,
1450                                 hammer2_tid_t mtid, int flags);
1451 void hammer2_chain_delete(hammer2_chain_t *parent, hammer2_chain_t *chain,
1452                                 hammer2_tid_t mtid, int flags);
1453 void hammer2_chain_setflush(hammer2_chain_t *chain);
1454 void hammer2_chain_countbrefs(hammer2_chain_t *chain,
1455                                 hammer2_blockref_t *base, int count);
1456 hammer2_chain_t *hammer2_chain_bulksnap(hammer2_chain_t *chain);
1457 void hammer2_chain_bulkdrop(hammer2_chain_t *copy);
1458
1459 void hammer2_chain_setcheck(hammer2_chain_t *chain, void *bdata);
1460 int hammer2_chain_testcheck(hammer2_chain_t *chain, void *bdata);
1461
1462
1463 void hammer2_pfs_memory_wait(hammer2_pfs_t *pmp);
1464 void hammer2_pfs_memory_inc(hammer2_pfs_t *pmp);
1465 void hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp);
1466
1467 void hammer2_base_delete(hammer2_chain_t *chain,
1468                                 hammer2_blockref_t *base, int count,
1469                                 int *cache_indexp, hammer2_chain_t *child);
1470 void hammer2_base_insert(hammer2_chain_t *chain,
1471                                 hammer2_blockref_t *base, int count,
1472                                 int *cache_indexp, hammer2_chain_t *child);
1473
1474 /*
1475  * hammer2_flush.c
1476  */
1477 void hammer2_flush(hammer2_chain_t *chain, int istop);
1478 void hammer2_flush_quick(hammer2_dev_t *hmp);
1479 void hammer2_delayed_flush(hammer2_chain_t *chain);
1480
1481 /*
1482  * hammer2_trans.c
1483  */
1484 void hammer2_trans_init(hammer2_pfs_t *pmp, uint32_t flags);
1485 hammer2_tid_t hammer2_trans_sub(hammer2_pfs_t *pmp);
1486 void hammer2_trans_clear_preflush(hammer2_pfs_t *pmp);
1487 void hammer2_trans_done(hammer2_pfs_t *pmp);
1488 hammer2_tid_t hammer2_trans_newinum(hammer2_pfs_t *pmp);
1489 void hammer2_trans_assert_strategy(hammer2_pfs_t *pmp);
1490
1491 /*
1492  * hammer2_ioctl.c
1493  */
1494 int hammer2_ioctl(hammer2_inode_t *ip, u_long com, void *data,
1495                                 int fflag, struct ucred *cred);
1496
1497 /*
1498  * hammer2_io.c
1499  */
1500 void hammer2_io_putblk(hammer2_io_t **diop);
1501 void hammer2_io_cleanup(hammer2_dev_t *hmp, struct hammer2_io_tree *tree);
1502 char *hammer2_io_data(hammer2_io_t *dio, off_t lbase);
1503 hammer2_io_t *hammer2_io_getquick(hammer2_dev_t *hmp, off_t lbase, int lsize);
1504 void hammer2_io_getblk(hammer2_dev_t *hmp, off_t lbase, int lsize,
1505                                 hammer2_iocb_t *iocb);
1506 void hammer2_io_complete(hammer2_iocb_t *iocb);
1507 void hammer2_io_callback(struct bio *bio);
1508 void hammer2_iocb_wait(hammer2_iocb_t *iocb);
1509 int hammer2_io_new(hammer2_dev_t *hmp, int btype, off_t lbase, int lsize,
1510                                 hammer2_io_t **diop);
1511 int hammer2_io_newnz(hammer2_dev_t *hmp, int btype, off_t lbase, int lsize,
1512                                 hammer2_io_t **diop);
1513 int hammer2_io_newq(hammer2_dev_t *hmp, int btype, off_t lbase, int lsize,
1514                                 hammer2_io_t **diop);
1515 int hammer2_io_bread(hammer2_dev_t *hmp, int btype, off_t lbase, int lsize,
1516                                 hammer2_io_t **diop);
1517 void hammer2_io_bawrite(hammer2_io_t **diop);
1518 void hammer2_io_bdwrite(hammer2_io_t **diop);
1519 int hammer2_io_bwrite(hammer2_io_t **diop);
1520 int hammer2_io_isdirty(hammer2_io_t *dio);
1521 void hammer2_io_setdirty(hammer2_io_t *dio);
1522 void hammer2_io_setinval(hammer2_io_t *dio, hammer2_off_t off, u_int bytes);
1523 void hammer2_io_brelse(hammer2_io_t **diop);
1524 void hammer2_io_bqrelse(hammer2_io_t **diop);
1525 int hammer2_io_crc_good(hammer2_chain_t *chain, uint64_t *maskp);
1526 void hammer2_io_crc_setmask(hammer2_io_t *dio, uint64_t mask);
1527 void hammer2_io_crc_clrmask(hammer2_io_t *dio, uint64_t mask);
1528
1529 /*
1530  * hammer2_thread.c
1531  */
1532 void hammer2_thr_signal(hammer2_thread_t *thr, uint32_t flags);
1533 void hammer2_thr_return(hammer2_thread_t *thr, uint32_t flags);
1534 void hammer2_thr_wait(hammer2_thread_t *thr, uint32_t flags);
1535 void hammer2_thr_wait_neg(hammer2_thread_t *thr, uint32_t flags);
1536 void hammer2_thr_create(hammer2_thread_t *thr, hammer2_pfs_t *pmp,
1537                         const char *id, int clindex, int repidx,
1538                         void (*func)(void *arg));
1539 void hammer2_thr_delete(hammer2_thread_t *thr);
1540 void hammer2_thr_remaster(hammer2_thread_t *thr);
1541 void hammer2_thr_freeze_async(hammer2_thread_t *thr);
1542 void hammer2_thr_freeze(hammer2_thread_t *thr);
1543 void hammer2_thr_unfreeze(hammer2_thread_t *thr);
1544 int hammer2_thr_break(hammer2_thread_t *thr);
1545 void hammer2_primary_xops_thread(void *arg);
1546
1547 /*
1548  * hammer2_thread.c (XOP API)
1549  */
1550 void hammer2_xop_group_init(hammer2_pfs_t *pmp, hammer2_xop_group_t *xgrp);
1551 void *hammer2_xop_alloc(hammer2_inode_t *ip, int flags);
1552 void hammer2_xop_setname(hammer2_xop_head_t *xop,
1553                                 const char *name, size_t name_len);
1554 void hammer2_xop_setname2(hammer2_xop_head_t *xop,
1555                                 const char *name, size_t name_len);
1556 void hammer2_xop_setip2(hammer2_xop_head_t *xop, hammer2_inode_t *ip2);
1557 void hammer2_xop_setip3(hammer2_xop_head_t *xop, hammer2_inode_t *ip3);
1558 void hammer2_xop_reinit(hammer2_xop_head_t *xop);
1559 void hammer2_xop_helper_create(hammer2_pfs_t *pmp);
1560 void hammer2_xop_helper_cleanup(hammer2_pfs_t *pmp);
1561 void hammer2_xop_start(hammer2_xop_head_t *xop, hammer2_xop_func_t func);
1562 void hammer2_xop_start_except(hammer2_xop_head_t *xop, hammer2_xop_func_t func,
1563                                 int notidx);
1564 int hammer2_xop_collect(hammer2_xop_head_t *xop, int flags);
1565 void hammer2_xop_retire(hammer2_xop_head_t *xop, uint32_t mask);
1566 int hammer2_xop_active(hammer2_xop_head_t *xop);
1567 int hammer2_xop_feed(hammer2_xop_head_t *xop, hammer2_chain_t *chain,
1568                                 int clindex, int error);
1569
1570 /*
1571  * hammer2_synchro.c
1572  */
1573 void hammer2_primary_sync_thread(void *arg);
1574
1575 /*
1576  * XOP backends in hammer2_xops.c, primarily for VNOPS.  Other XOP backends
1577  * may be integrated into other source files.
1578  */
1579 void hammer2_xop_ipcluster(hammer2_xop_t *xop, int clidx);
1580 void hammer2_xop_readdir(hammer2_xop_t *xop, int clidx);
1581 void hammer2_xop_nresolve(hammer2_xop_t *xop, int clidx);
1582 void hammer2_xop_unlink(hammer2_xop_t *xop, int clidx);
1583 void hammer2_xop_nrename(hammer2_xop_t *xop, int clidx);
1584 void hammer2_xop_nlink(hammer2_xop_t *xop, int clidx);
1585 void hammer2_xop_scanlhc(hammer2_xop_t *xop, int clidx);
1586 void hammer2_xop_scanall(hammer2_xop_t *xop, int clidx);
1587 void hammer2_xop_lookup(hammer2_xop_t *xop, int clidx);
1588 void hammer2_inode_xop_create(hammer2_xop_t *xop, int clidx);
1589 void hammer2_inode_xop_destroy(hammer2_xop_t *xop, int clidx);
1590 void hammer2_inode_xop_chain_sync(hammer2_xop_t *xop, int clidx);
1591 void hammer2_inode_xop_unlinkall(hammer2_xop_t *xop, int clidx);
1592 void hammer2_inode_xop_connect(hammer2_xop_t *xop, int clidx);
1593 void hammer2_inode_xop_flush(hammer2_xop_t *xop, int clidx);
1594
1595 /*
1596  * hammer2_msgops.c
1597  */
1598 int hammer2_msg_dbg_rcvmsg(kdmsg_msg_t *msg);
1599 int hammer2_msg_adhoc_input(kdmsg_msg_t *msg);
1600
1601 /*
1602  * hammer2_vfsops.c
1603  */
1604 void hammer2_volconf_update(hammer2_dev_t *hmp, int index);
1605 void hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx);
1606 int hammer2_vfs_sync(struct mount *mp, int waitflags);
1607 hammer2_pfs_t *hammer2_pfsalloc(hammer2_chain_t *chain,
1608                                 const hammer2_inode_data_t *ripdata,
1609                                 hammer2_tid_t modify_tid,
1610                                 hammer2_dev_t *force_local);
1611
1612 void hammer2_lwinprog_ref(hammer2_pfs_t *pmp);
1613 void hammer2_lwinprog_drop(hammer2_pfs_t *pmp);
1614 void hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int pipe);
1615
1616 /*
1617  * hammer2_freemap.c
1618  */
1619 int hammer2_freemap_alloc(hammer2_chain_t *chain, size_t bytes);
1620 void hammer2_freemap_adjust(hammer2_dev_t *hmp,
1621                                 hammer2_blockref_t *bref, int how);
1622
1623 /*
1624  * hammer2_cluster.c
1625  */
1626 uint8_t hammer2_cluster_type(hammer2_cluster_t *cluster);
1627 const hammer2_media_data_t *hammer2_cluster_rdata(hammer2_cluster_t *cluster);
1628 hammer2_media_data_t *hammer2_cluster_wdata(hammer2_cluster_t *cluster);
1629 hammer2_cluster_t *hammer2_cluster_from_chain(hammer2_chain_t *chain);
1630 void hammer2_cluster_bref(hammer2_cluster_t *cluster, hammer2_blockref_t *bref);
1631 hammer2_cluster_t *hammer2_cluster_alloc(hammer2_pfs_t *pmp,
1632                                 hammer2_blockref_t *bref);
1633 void hammer2_cluster_ref(hammer2_cluster_t *cluster);
1634 void hammer2_cluster_drop(hammer2_cluster_t *cluster);
1635 void hammer2_cluster_lock(hammer2_cluster_t *cluster, int how);
1636 int hammer2_cluster_check(hammer2_cluster_t *cluster, hammer2_key_t lokey,
1637                         int flags);
1638 void hammer2_cluster_resolve(hammer2_cluster_t *cluster);
1639 void hammer2_cluster_forcegood(hammer2_cluster_t *cluster);
1640 hammer2_cluster_t *hammer2_cluster_copy(hammer2_cluster_t *ocluster);
1641 void hammer2_cluster_unlock(hammer2_cluster_t *cluster);
1642
1643 int hammer2_bulkfree_pass(hammer2_dev_t *hmp,
1644                         struct hammer2_ioc_bulkfree *bfi);
1645
1646 /*
1647  * hammer2_iocom.c
1648  */
1649 void hammer2_iocom_init(hammer2_dev_t *hmp);
1650 void hammer2_iocom_uninit(hammer2_dev_t *hmp);
1651 void hammer2_cluster_reconnect(hammer2_dev_t *hmp, struct file *fp);
1652
1653 /*
1654  * hammer2_strategy.c
1655  */
1656 int hammer2_vop_strategy(struct vop_strategy_args *ap);
1657 int hammer2_vop_bmap(struct vop_bmap_args *ap);
1658 void hammer2_write_thread(void *arg);
1659 void hammer2_bioq_sync(hammer2_pfs_t *pmp);
1660 void hammer2_dedup_clear(hammer2_dev_t *hmp);
1661
1662 #endif /* !_KERNEL */
1663 #endif /* !_VFS_HAMMER2_HAMMER2_H_ */